question
stringlengths
101
4.47k
chosen
stringlengths
0
4.25k
rejected
stringlengths
39
9.29k
RAG systems have a unique appeal over traditional search engines in that they can incorporate prior knowledge to fill in the gaps and extrapolate the retrieved information. We find that this comes with trade-offs - namely, that such priors can override information provided in documents. When perturbing RAG documents over a wide interval of values, the point at which models revert to their prior responses, or "tipping points", are latent and heterogeneous across different models and domains. While strong priors are not inherently problematic (and can often serve to safeguard models), the lack of explicit expectations around how models will mix reference documents with their priors can lead to downstream issues. For example, if RAG systems are used to extract nested financial data to be used in an algorithm, what will happen if there is a typo in the financial documents? Will the model notice the error and if so, what data will it provide in its place? Given that LLMs are soon to be widely deployed in many domains including medicine and law, users and developers alike should be cognizant of their unintended effects, especially if users have preconceptions that RAG-enabled systems are, by nature, always truthful.
RAG ์‹œ์Šคํ…œ์€ ๊ธฐ์กด ์ง€์‹์„ ํ†ตํ•ฉํ•˜์—ฌ ๋ถ€์กฑํ•œ ๋ถ€๋ถ„์„ ์ฑ„์šฐ๊ณ  ๊ฒ€์ƒ‰๋œ ์ •๋ณด๋ฅผ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์—์„œ ๊ธฐ์กด ๊ฒ€์ƒ‰ ์—”์ง„์— ๋น„ํ•ด ๋…ํŠนํ•œ ๋งค๋ ฅ์„ ์ง€๋‹ˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ด์ ์—๋Š” trade-offs๊ฐ€ ์žˆ๋Š”๋ฐ, ์ด๋Ÿฌํ•œ priors๊ฐ€ ๋ฌธ์„œ์— ์ œ๊ณต๋œ ์ •๋ณด๋ฅผ ๋ฎ์–ด์“ธ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๊ฐ’ ๊ฐ„๊ฒฉ์— ๊ฑธ์ณ RAG ๋ฌธ์„œ๋ฅผ ๊ต๋ž€ํ•  ๋•Œ, ๋ชจ๋ธ์ด ์ด์ „ ์‘๋‹ต์œผ๋กœ ๋˜๋Œ์•„๊ฐ€๋Š” ์‹œ์ , ์ฆ‰ 'tipping points'๋Š” ์ž ์žฌ์ ์ด๋ฉฐ ๋ชจ๋ธ๊ณผ ๋„๋ฉ”์ธ์— ๋”ฐ๋ผ ์ด์งˆ์ ์ž…๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ priors๊ฐ€ ๋ณธ์งˆ์ ์œผ๋กœ ๋ฌธ์ œ๊ฐ€ ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ(์ข…์ข… ๋ชจ๋ธ์„ ๋ณดํ˜ธํ•˜๋Š” ์—ญํ• ์„ ํ•  ์ˆ˜๋„ ์žˆ์Œ), ๋ชจ๋ธ์ด ์ฐธ์กฐ ๋ฌธ์„œ์™€ priors๋ฅผ ํ˜ผํ•ฉํ•˜๋Š” ๋ฐฉ์‹์— ๋Œ€ํ•œ ๋ช…์‹œ์ ์ธ ๊ธฐ๋Œ€์น˜๊ฐ€ ๋ถ€์กฑํ•˜๋ฉด downstream์—์„œ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ์‚ฌ์šฉํ•  ์ค‘์ฒฉ๋œ ์žฌ๋ฌด ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๋ฐ RAG ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ ์žฌ๋ฌด ๋ฌธ์„œ์— ์˜คํƒ€๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ ์–ด๋–ป๊ฒŒ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ๋ชจ๋ธ์ด ์˜ค๋ฅ˜๋ฅผ ๊ฐ์ง€ํ•˜๊ณ  ๊ฐ์ง€ํ•œ๋‹ค๋ฉด ์–ด๋–ค ๋ฐ์ดํ„ฐ๋ฅผ ๋Œ€์‹  ์ œ๊ณตํ• ๊นŒ์š”? LLM์ด ์˜๋ฃŒ ๋ฐ ๋ฒ•๋ฅ ์„ ํฌํ•จํ•œ ๋งŽ์€ ์˜์—ญ์— ๊ณง ๋„๋ฆฌ ๋ฐฐํฌ๋  ์˜ˆ์ •์ด๋ฏ€๋กœ, ์‚ฌ์šฉ์ž์™€ ๊ฐœ๋ฐœ์ž ๋ชจ๋‘ ๋ชจ๋ธ์˜ ์˜๋„ํ•˜์ง€ ์•Š์€ ์˜ํ–ฅ, ํŠนํžˆ RAG ์ง€์› ์‹œ์Šคํ…œ์ด ๋ณธ์งˆ์ ์œผ๋กœ ํ•ญ์ƒ ์ง„์‹คํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋Š” ์‚ฌ์šฉ์ž์˜ ์˜คํ•ด์— ์ฃผ์˜๋ฅผ ๊ธฐ์šธ์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
RAG ์‹œ์Šคํ…œ์€ ๊ธฐ์กด ์ง€์‹์„ ํ†ตํ•ฉํ•˜์—ฌ ๋ถ€์กฑํ•œ ๋ถ€๋ถ„์„ ์ฑ„์šฐ๊ณ  ๊ฒ€์ƒ‰๋œ ์ •๋ณด๋ฅผ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์—์„œ ๊ธฐ์กด ๊ฒ€์ƒ‰ ์—”์ง„์— ๋น„ํ•ด ๋…ํŠนํ•œ ๋งค๋ ฅ์„ ์ง€๋‹ˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ด์ ์—๋Š” ์žฅ๋‹จ์ ์ด ์žˆ๋Š”๋ฐ, ์ด๋Ÿฌํ•œ ์„ ํ—˜์ด ๋ฌธ์„œ์— ์ œ๊ณต๋œ ์ •๋ณด๋ฅผ ๋ฎ์–ด์“ธ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๊ฐ’ ๊ฐ„๊ฒฉ์— ๊ฑธ์ณ RAG ๋ฌธ์„œ๋ฅผ ๊ต๋ž€ํ•  ๋•Œ, ๋ชจ๋ธ์ด ์ด์ „ ์‘๋‹ต์œผ๋กœ ๋˜๋Œ์•„๊ฐ€๋Š” ์‹œ์ , ์ฆ‰ '์„ ํ—˜์  ์ง€์ '์€ ์ž ์žฌ์ ์ด๋ฉฐ ๋ชจ๋ธ๊ณผ ๋„๋ฉ”์ธ์— ๋”ฐ๋ผ ์ด์งˆ์ ์ž…๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ ์„ ํ—˜์ด ๋ณธ์งˆ์ ์œผ๋กœ ๋ฌธ์ œ๊ฐ€ ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ(์ข…์ข… ๋ชจ๋ธ์„ ๋ณดํ˜ธํ•˜๋Š” ์—ญํ• ์„ ํ•  ์ˆ˜๋„ ์žˆ์Œ), ๋ชจ๋ธ์ด ์ฐธ์กฐ ๋ฌธ์„œ์™€ ์„ ํ—˜์„ ํ˜ผํ•ฉํ•˜๋Š” ๋ฐฉ์‹์— ๋Œ€ํ•œ ๋ช…์‹œ์ ์ธ ๊ธฐ๋Œ€์น˜๊ฐ€ ๋ถ€์กฑํ•˜๋ฉด ๋‹ค์šด์ŠคํŠธ๋ฆผ์—์„œ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ์‚ฌ์šฉํ•  ์ค‘์ฒฉ๋œ ์žฌ๋ฌด ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๋ฐ RAG ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ ์žฌ๋ฌด ๋ฌธ์„œ์— ์˜คํƒ€๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ ์–ด๋–ป๊ฒŒ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ๋ชจ๋ธ์ด ์˜ค๋ฅ˜๋ฅผ ๊ฐ์ง€ํ•˜๊ณ  ๊ฐ์ง€ํ•œ๋‹ค๋ฉด ์–ด๋–ค ๋ฐ์ดํ„ฐ๋ฅผ ๋Œ€์‹  ์ œ๊ณตํ• ๊นŒ์š”? LLM์ด ์˜๋ฃŒ ๋ฐ ๋ฒ•๋ฅ ์„ ํฌํ•จํ•œ ๋งŽ์€ ์˜์—ญ์— ๊ณง ๋„๋ฆฌ ๋ฐฐํฌ๋  ์˜ˆ์ •์ด๋ฏ€๋กœ, ์‚ฌ์šฉ์ž์™€ ๊ฐœ๋ฐœ์ž ๋ชจ๋‘ ๋ชจ๋ธ์˜ ์˜๋„ํ•˜์ง€ ์•Š์€ ์˜ํ–ฅ, ํŠนํžˆ RAG ์ง€์› ์‹œ์Šคํ…œ์ด ๋ณธ์งˆ์ ์œผ๋กœ ํ•ญ์ƒ ์ง„์‹คํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋Š” ์‚ฌ์šฉ์ž์˜ ์˜คํ•ด์— ์ฃผ์˜๋ฅผ ๊ธฐ์šธ์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_Training on code data._ A major limitation of the original GPT-3 model (pre-trained on plain text) lies in the lack of the reasoning ability on complex tasks, _e.g._, completing the code and solving math problems. To enhance this ability, Codex [105] was introduced by OpenAI in July 2021, which was a GPT model fine-tuned on a large corpus of GitHub code. It demonstrated that Codex can solve very difficult programming problems, and also lead to a significant performance improvement in solving math problems [126]. Further, a contrastive approach [127] to training text and code embedding was reported in January 2022, which was shown to improve a series of related tasks (_i.e._, linear-probe classification, text search and code search). Actually, the GPT-3.5 models are developed based on a code-based GPT model (_i.e._, code-davinci-002), which indicates that training on code data is a very useful practice to improve the model capacity of GPT models, especially the reasoning ability. Furthermore, there is also a speculation that training on code data can greatly increase the chain-of-thought prompting abilities of LLMs [47], while it is still worth further investigation with more thorough verification.
\(\bullet\)_์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ._ ์›๋ž˜ GPT-3 ๋ชจ๋ธ(์ผ๋ฐ˜ ํ…์ŠคํŠธ๋กœ ์‚ฌ์ „ ํ•™์Šต)์˜ ์ฃผ์š” ํ•œ๊ณ„๋Š” ๋ณต์žกํ•œ ์ž‘์—…, _์˜ˆ๋ฅผ ๋“ค์–ด ์ฝ”๋“œ ์™„์„ฑ ๋ฐ ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ๊ณผ ๊ฐ™์€ ์ถ”๋ก  ๋Šฅ๋ ฅ์ด ๋ถ€์กฑํ•˜๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด 2021๋…„ 7์›”์— OpenAI์—์„œ GitHub ์ฝ”๋“œ์˜ ๋Œ€๊ทœ๋ชจ ๋ง๋ญ‰์น˜์— ๋Œ€ํ•ด Fine-tuningํ•œ GPT ๋ชจ๋ธ์ธ Codex [105]๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. Codex๊ฐ€ ๋งค์šฐ ์–ด๋ ค์šด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์œผ๋ฉฐ, ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ์—์„œ๋„ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์ด๋Œ์–ด๋ƒˆ์Šต๋‹ˆ๋‹ค[126]. ๋˜ํ•œ 2022๋…„ 1์›”์—๋Š” ํ…์ŠคํŠธ์™€ ์ฝ”๋“œ ์ž„๋ฒ ๋”ฉ์„ ํ›ˆ๋ จํ•˜๋Š” ๋Œ€์กฐ์  ์ ‘๊ทผ ๋ฐฉ์‹[127]์ด ๋ณด๊ณ ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” ์ผ๋ จ์˜ ๊ด€๋ จ ์ž‘์—…(์ฆ‰, ์„ ํ˜• ํ”„๋กœ๋ธŒ ๋ถ„๋ฅ˜, ํ…์ŠคํŠธ ๊ฒ€์ƒ‰ ๋ฐ ์ฝ”๋“œ ๊ฒ€์ƒ‰)์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ GPT-3.5 ๋ชจ๋ธ์€ ์ฝ”๋“œ ๊ธฐ๋ฐ˜ GPT ๋ชจ๋ธ(_์ฆ‰, code-davinci-002)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์ด ํŠนํžˆ ์ถ”๋ก  ๋Šฅ๋ ฅ๊ณผ ๊ฐ™์€ GPT ๋ชจ๋ธ์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋งค์šฐ ์œ ์šฉํ•œ ์‹ค์Šต์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์€ LLM์˜ chain-of-thought prompting ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ถ”์ธก๋„ ์žˆ์ง€๋งŒ, ์ด๋Š” ์•„์ง ๋” ์ฒ ์ €ํ•œ ๊ฒ€์ฆ์„ ํ†ตํ•ด ์ถ”๊ฐ€ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ._ ์›๋ž˜ GPT-3 ๋ชจ๋ธ(์ผ๋ฐ˜ ํ…์ŠคํŠธ๋กœ ์‚ฌ์ „ ํ•™์Šต)์˜ ์ฃผ์š” ํ•œ๊ณ„๋Š” ๋ณต์žกํ•œ ์ž‘์—…, _์˜ˆ๋ฅผ ๋“ค์–ด ์ฝ”๋“œ ์™„์„ฑ ๋ฐ ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ๊ณผ ๊ฐ™์€ ์ถ”๋ก  ๋Šฅ๋ ฅ์ด ๋ถ€์กฑํ•˜๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด 2021๋…„ 7์›”์— OpenAI์—์„œ GitHub ์ฝ”๋“œ์˜ ๋Œ€๊ทœ๋ชจ ๋ง๋ญ‰์น˜์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •ํ•œ GPT ๋ชจ๋ธ์ธ Codex [105]๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. Codex๊ฐ€ ๋งค์šฐ ์–ด๋ ค์šด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์œผ๋ฉฐ, ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ์—์„œ๋„ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์ด๋Œ์–ด๋ƒˆ์Šต๋‹ˆ๋‹ค[126]. ๋˜ํ•œ 2022๋…„ 1์›”์—๋Š” ํ…์ŠคํŠธ์™€ ์ฝ”๋“œ ์ž„๋ฒ ๋”ฉ์„ ํ›ˆ๋ จํ•˜๋Š” ๋Œ€์กฐ์  ์ ‘๊ทผ ๋ฐฉ์‹[127]์ด ๋ณด๊ณ ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” ์ผ๋ จ์˜ ๊ด€๋ จ ์ž‘์—…(์ฆ‰, ์„ ํ˜• ํ”„๋กœ๋ธŒ ๋ถ„๋ฅ˜, ํ…์ŠคํŠธ ๊ฒ€์ƒ‰ ๋ฐ ์ฝ”๋“œ ๊ฒ€์ƒ‰)์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ GPT-3.5 ๋ชจ๋ธ์€ ์ฝ”๋“œ ๊ธฐ๋ฐ˜ GPT ๋ชจ๋ธ(_์ฆ‰, code-davinci-002)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์ด ํŠนํžˆ ์ถ”๋ก  ๋Šฅ๋ ฅ๊ณผ ๊ฐ™์€ GPT ๋ชจ๋ธ์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋งค์šฐ ์œ ์šฉํ•œ ์‹ค์Šต์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์€ LLM์˜ ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ํ”„๋กฌํ”„ํŠธ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ถ”์ธก๋„ ์žˆ์ง€๋งŒ, ์ด๋Š” ์•„์ง ๋” ์ฒ ์ €ํ•œ ๊ฒ€์ฆ์„ ํ†ตํ•ด ์ถ”๊ฐ€ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
Table 3 shows that by integrating our method with other methods, the performance can be further improved across different LLMs and tasks, despite these methods have different implementations. To be specific, in arithmetic reasoning tasks, our method enhances these methods to further improvement, yielding increases between \(10\%\) and \(21\%\) on the GSM8K dataset, and between \(1\%\) and \(15\%\) on the MATH dataset. In general reasoning tasks, integration with other methods generally achieves performance gains ranging from \(1\%\) to \(13\%\) in the Chess task and from \(1\%\) to \(11\%\) in the MMLU task. In code generation task, when combined with other methods, gains range from \(2\%\) to \(7\%\). However, two notable exceptions are observed when integrated with the debate method with the Llama2-13B and Llama2-70B models, which result in failed cases. This failure in performance is attributed primarily to the noise generated by referencing the answers of other agents during the debate process. The synthesized responses, which incorporate input from multiple agents, disrupt the coherence of the code logic, leading to the observed performance degradation. All accuracy curves are provided in the Appendix B.
ํ‘œ 3์€ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ฉํ•˜๋ฉด ๊ตฌํ˜„ ๋ฐฉ์‹์ด ๋‹ค๋ฅด์ง€๋งŒ ์—ฌ๋Ÿฌ LLM๊ณผ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ arithmetic reasoning ์ž‘์—…์˜ ๊ฒฝ์šฐ, ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์€ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์„ ํ–ฅ์ƒ์‹œ์ผœ GSM8K ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 10%์—์„œ 21%, MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 1%์—์„œ 15% ์‚ฌ์ด์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์™”์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜ reasoning ์ž‘์—…์˜ ๊ฒฝ์šฐ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ์˜ ํ†ตํ•ฉ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ฒด์Šค ์ž‘์—…์—์„œ 1%์—์„œ 13%, MMLU ์ž‘์—…์—์„œ 1%์—์„œ 11% ์‚ฌ์ด์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. code generation ์ž‘์—…์˜ ๊ฒฝ์šฐ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ ๊ฒฐํ•ฉํ•˜๋ฉด 2%์—์„œ 7%๊นŒ์ง€ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‘ ๊ฐ€์ง€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์˜ˆ์™ธ๋Š” debate ๋ฐฉ๋ฒ•๊ณผ ๊ฒฐํ•ฉํ–ˆ์„ ๋•Œ Llama2-13B ๋ฐ Llama2-70B ๋ชจ๋ธ์—์„œ ์‹คํŒจํ•œ ๊ฒฝ์šฐ๊ฐ€ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ฑ๋Šฅ ์ €ํ•˜๋Š” ์ฃผ๋กœ ํ† ๋ก  ๊ณผ์ •์—์„œ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์˜ ๋‹ต๋ณ€์„ ์ฐธ์กฐํ•  ๋•Œ ๋ฐœ์ƒํ•˜๋Š” ๋…ธ์ด์ฆˆ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ์ž…๋ ฅ์„ ํ†ตํ•ฉํ•œ ํ•ฉ์„ฑ๋œ ์‘๋‹ต์€ ์ฝ”๋“œ ๋กœ์ง์˜ ์ผ๊ด€์„ฑ์„ ๋ฐฉํ•ดํ•˜์—ฌ ์„ฑ๋Šฅ ์ €ํ•˜๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ๋ชจ๋“  ์ •ํ™•๋„ ๊ณก์„ ์€ ๋ถ€๋ก B์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
ํ‘œ 3์€ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ฉํ•˜๋ฉด ๊ตฌํ˜„ ๋ฐฉ์‹์ด ๋‹ค๋ฅด์ง€๋งŒ ์—ฌ๋Ÿฌ LLM๊ณผ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ ์‚ฐ์ˆ  ์ถ”๋ก  ์ž‘์—…์˜ ๊ฒฝ์šฐ, ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์€ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์„ ํ–ฅ์ƒ์‹œ์ผœ GSM8K ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 10%์—์„œ 21%, MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 1%์—์„œ 15% ์‚ฌ์ด์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์™”์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜ ์ถ”๋ก  ์ž‘์—…์˜ ๊ฒฝ์šฐ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ์˜ ํ†ตํ•ฉ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ฒด์Šค ์ž‘์—…์—์„œ 1%์—์„œ 13%, MMLU ์ž‘์—…์—์„œ 1%์—์„œ 11% ์‚ฌ์ด์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์˜ ๊ฒฝ์šฐ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ ๊ฒฐํ•ฉํ•˜๋ฉด 2%์—์„œ 7%๊นŒ์ง€ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‘ ๊ฐ€์ง€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์˜ˆ์™ธ๋Š” ํ† ๋ก  ๋ฐฉ๋ฒ•๊ณผ ๊ฒฐํ•ฉํ–ˆ์„ ๋•Œ Llama2-13B ๋ฐ Llama2-70B ๋ชจ๋ธ์—์„œ ์‹คํŒจํ•œ ๊ฒฝ์šฐ๊ฐ€ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ฑ๋Šฅ ์ €ํ•˜๋Š” ์ฃผ๋กœ ํ† ๋ก  ๊ณผ์ •์—์„œ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์˜ ๋‹ต๋ณ€์„ ์ฐธ์กฐํ•  ๋•Œ ๋ฐœ์ƒํ•˜๋Š” ๋…ธ์ด์ฆˆ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ์ž…๋ ฅ์„ ํ†ตํ•ฉํ•œ ํ•ฉ์„ฑ๋œ ์‘๋‹ต์€ ์ฝ”๋“œ ๋กœ์ง์˜ ์ผ๊ด€์„ฑ์„ ๋ฐฉํ•ดํ•˜์—ฌ ์„ฑ๋Šฅ ์ €ํ•˜๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ๋ชจ๋“  ์ •ํ™•๋„ ๊ณก์„ ์€ ๋ถ€๋ก B์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
RQ5: How does the style of synthetic data impact performance on specialized domains?We compare the performance of various models trained on different styles of synthetic data. In particular, we generate four styles of synthetic data (easy, medium, hard, and Q/A) and evaluate the performance of training on combinations of each style across File subsets. The prompts to generate these synthetic data styles are outlined in Appendix G. Results corresponding to generations from a Vicuna-v1.3 model, and for a 128M model trained for 3B tokens are summarized in Figure 7. We see that training with combinations of real C4 and synthetic data matching the style of the domain at evaluation improves performance. However, we find that no single synthetic data style performs the best across all domains, resulting in similar performance across training with combinations of real C4 data and each synthetic style variant. While knowing the best synthetic style to pre-train an LLM is impractical, an oracle that selects the best synthetic style across all domains will improve perplexity by 16%--indicating the importance of training with diverse data styles for LLM generalization, even when the underlying knowledge stays the same.
RQ5: ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์Šคํƒ€์ผ์ด ํŠน์ˆ˜ ๋„๋ฉ”์ธ์—์„œ์˜ ์„ฑ๋Šฅ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋‚˜์š”? ๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ๋„ค ๊ฐ€์ง€ ์Šคํƒ€์ผ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ(easy, medium, hard, and Q/A)๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ๊ฐ ์Šคํƒ€์ผ์˜ ์กฐํ•ฉ์— ๋Œ€ํ•œ ํ•™์Šต ์„ฑ๋Šฅ์„ ํŒŒ์ผ ํ•˜์œ„ ์ง‘ํ•ฉ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์Šคํƒ€์ผ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ prompts๋Š” ๋ถ€๋ก G์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. Vicuna-v1.3 ๋ชจ๋ธ์— ํ•ด๋‹นํ•˜๋Š” ๊ฒฐ๊ณผ์™€ 3B tokens์— ๋Œ€ํ•ด 128M ๋ชจ๋ธ์„ ํ•™์Šตํ•œ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 7์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€ ์‹œ ๋„๋ฉ”์ธ ์Šคํƒ€์ผ์— ๋งž๋Š” ์‹ค์ œ C4์™€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์กฐํ•ฉ์œผ๋กœ ํ›ˆ๋ จํ•˜๋ฉด ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋“  ๋„๋ฉ”์ธ์—์„œ ๋‹จ์ผ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์Šคํƒ€์ผ์ด ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์‹ค์ œ C4 ๋ฐ์ดํ„ฐ์™€ ๊ฐ ํ•ฉ์„ฑ ์Šคํƒ€์ผ ๋ณ€ํ˜•์˜ ์กฐํ•ฉ์œผ๋กœ ํ›ˆ๋ จํ•  ๋•Œ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. LLM์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ํ•ฉ์„ฑ ์Šคํƒ€์ผ์„ ์•„๋Š” ๊ฒƒ์€ ๋น„ํ˜„์‹ค์ ์ด์ง€๋งŒ, ๋ชจ๋“  ๋„๋ฉ”์ธ์—์„œ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ํ•ฉ์„ฑ ์Šคํƒ€์ผ์„ ์„ ํƒํ•˜๋Š” ์˜ค๋ผํด์€ perplexity๋ฅผ 16% ๊ฐœ์„ ํ•˜์—ฌ ๊ธฐ๋ณธ ์ง€์‹์€ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€ํ•˜๋”๋ผ๋„ LLM ์ผ๋ฐ˜ํ™”๋ฅผ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์Šคํƒ€์ผ๋กœ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
RQ5: ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์Šคํƒ€์ผ์ด ํŠน์ˆ˜ ๋„๋ฉ”์ธ์—์„œ์˜ ์„ฑ๋Šฅ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋‚˜์š”?๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ๋„ค ๊ฐ€์ง€ ์Šคํƒ€์ผ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ(์‰ฌ์›€, ์ค‘๊ฐ„, ์–ด๋ ค์›€, Q/A)๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ๊ฐ ์Šคํƒ€์ผ์˜ ์กฐํ•ฉ์— ๋Œ€ํ•œ ํ•™์Šต ์„ฑ๋Šฅ์„ ํŒŒ์ผ ํ•˜์œ„ ์ง‘ํ•ฉ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์Šคํƒ€์ผ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ํ”„๋กฌํ”„ํŠธ๋Š” ๋ถ€๋ก G์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. Vicuna-v1.3 ๋ชจ๋ธ์— ํ•ด๋‹นํ•˜๋Š” ๊ฒฐ๊ณผ์™€ 3B ํ† ํฐ์— ๋Œ€ํ•ด 128M ๋ชจ๋ธ์„ ํ•™์Šตํ•œ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 7์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€ ์‹œ ๋„๋ฉ”์ธ ์Šคํƒ€์ผ์— ๋งž๋Š” ์‹ค์ œ C4์™€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์กฐํ•ฉ์œผ๋กœ ํ›ˆ๋ จํ•˜๋ฉด ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋“  ๋„๋ฉ”์ธ์—์„œ ๋‹จ์ผ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์Šคํƒ€์ผ์ด ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์‹ค์ œ C4 ๋ฐ์ดํ„ฐ์™€ ๊ฐ ํ•ฉ์„ฑ ์Šคํƒ€์ผ ๋ณ€ํ˜•์˜ ์กฐํ•ฉ์œผ๋กœ ํ›ˆ๋ จํ•  ๋•Œ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. LLM์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ํ•ฉ์„ฑ ์Šคํƒ€์ผ์„ ์•„๋Š” ๊ฒƒ์€ ๋น„ํ˜„์‹ค์ ์ด์ง€๋งŒ, ๋ชจ๋“  ๋„๋ฉ”์ธ์—์„œ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ํ•ฉ์„ฑ ์Šคํƒ€์ผ์„ ์„ ํƒํ•˜๋Š” ์˜ค๋ผํด์€ ๋‚œํ•ด์„ฑ์„ 16% ๊ฐœ์„ ํ•˜์—ฌ ๊ธฐ๋ณธ ์ง€์‹์€ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€ํ•˜๋”๋ผ๋„ LLM ์ผ๋ฐ˜ํ™”๋ฅผ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์Šคํƒ€์ผ๋กœ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
**Tokenization.** Tokenization is also a crucial step for data preprocessing. It aims to segment raw text into sequences of individual tokens, which are subsequently used as the inputs of LLMs. In traditional NLP research (_e.g.,_ sequence labeling with conditional random fields [220]), word-based tokenization is the predominant approach, which is more aligned with human's language cognition. However, word-based tokenization can yield different segmentation results for the same input in some languages (_e.g.,_ Chinese word segmentation), generate a huge word vocabulary containing many low-frequency words, and also suffer from the _"out-of-vocabulary"_ issue. Thus, several neural network models employ _character_ as the minimum unit to derive the word representation (_e.g.,_ a CNN word encoder in ELMo [21]). Recently, _subword tokenizers_ have been widely used in Transformer based language models, typically including Byte-Pair Encoding tokenization, WordPiece tokenization and Unigram tokenization. HuggingFace has maintained an excellent online NLP course on tokenizer22 with running examples, and we refer to the beginners to this course. Next, we briefly describe the three representative tokenization methods.
ํ† ํฐํ™”.** Tokenization์€ ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ค‘์š”ํ•œ ๋‹จ๊ณ„์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. Tokenization์€ ์›์‹œ ํ…์ŠคํŠธ๋ฅผ ๊ฐœ๋ณ„ ํ† ํฐ์˜ ์‹œํ€€์Šค๋กœ ๋ถ„ํ• ํ•˜์—ฌ LLM์˜ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด NLP ์—ฐ๊ตฌ(์˜ˆ: ์กฐ๊ฑด๋ถ€ ๋žœ๋ค ํ•„๋“œ๋ฅผ ์‚ฌ์šฉํ•œ ์‹œํ€€์Šค ๋ผ๋ฒจ๋ง [220])์—์„œ๋Š” ๋‹จ์–ด ๊ธฐ๋ฐ˜ Tokenization์ด ์ฃผ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ, ์ด๋Š” ์ธ๊ฐ„์˜ ์–ธ์–ด ์ธ์ง€์™€ ๋” ์ž˜ ๋ถ€ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹จ์–ด ๊ธฐ๋ฐ˜ Tokenization์€ ์ผ๋ถ€ ์–ธ์–ด(์˜ˆ: ์ค‘๊ตญ์–ด ๋‹จ์–ด ๋ถ„ํ• )์—์„œ ๋™์ผํ•œ ์ž…๋ ฅ์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ๋ถ„ํ•  ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋นˆ๋„๊ฐ€ ๋‚ฎ์€ ๋‹จ์–ด๊ฐ€ ํฌํ•จ๋œ ๋ฐฉ๋Œ€ํ•œ ๋‹จ์–ด ์–ดํœ˜๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๊ณ , '์–ดํœ˜ ์™ธ ๋‹จ์–ด' ๋ฌธ์ œ๋„ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ช‡๋ช‡ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์—์„œ๋Š” _๋ฌธ์ž_๋ฅผ ์ตœ์†Œ ๋‹จ์œ„๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์–ด ํ‘œํ˜„์„ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค(_์˜ˆ: ELMo [21]์˜ CNN ๋‹จ์–ด Encoder). ์ตœ๊ทผ์—๋Š” Transformer ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ชจ๋ธ์—์„œ _์„œ๋ธŒ์›Œ๋“œ Tokenization_๊ฐ€ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ Byte-Pair Encoding Tokenization, WordPiece Tokenization, Unigram Tokenization์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. HuggingFace๋Š” Tokenizer22์— ๋Œ€ํ•œ ๋›ฐ์–ด๋‚œ ์˜จ๋ผ์ธ NLP ๊ฐ•์ขŒ๋ฅผ ์šด์˜ ์ค‘์ด๋ฉฐ, ์ดˆ๋ณด์ž๋Š” ์ด ๊ฐ•์ขŒ๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ๋‹ค์Œ์—์„œ๋Š” ๋Œ€ํ‘œ์ ์ธ Tokenization ๋ฐฉ๋ฒ• ์„ธ ๊ฐ€์ง€๋ฅผ ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
ํ† ํฐํ™”.** ํ† ํฐํ™”๋Š” ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ค‘์š”ํ•œ ๋‹จ๊ณ„์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ํ† ํฐํ™”๋Š” ์›์‹œ ํ…์ŠคํŠธ๋ฅผ ๊ฐœ๋ณ„ ํ† ํฐ์˜ ์‹œํ€€์Šค๋กœ ๋ถ„ํ• ํ•˜์—ฌ LLM์˜ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด NLP ์—ฐ๊ตฌ(์˜ˆ: ์กฐ๊ฑด๋ถ€ ๋žœ๋ค ํ•„๋“œ๋ฅผ ์‚ฌ์šฉํ•œ ์‹œํ€€์Šค ๋ผ๋ฒจ๋ง [220])์—์„œ๋Š” ๋‹จ์–ด ๊ธฐ๋ฐ˜ ํ† ํฐํ™”๊ฐ€ ์ฃผ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ, ์ด๋Š” ์ธ๊ฐ„์˜ ์–ธ์–ด ์ธ์ง€์™€ ๋” ์ž˜ ๋ถ€ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹จ์–ด ๊ธฐ๋ฐ˜ ํ† ํฐํ™”๋Š” ์ผ๋ถ€ ์–ธ์–ด(์˜ˆ: ์ค‘๊ตญ์–ด ๋‹จ์–ด ๋ถ„ํ• )์—์„œ ๋™์ผํ•œ ์ž…๋ ฅ์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ๋ถ„ํ•  ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋นˆ๋„๊ฐ€ ๋‚ฎ์€ ๋‹จ์–ด๊ฐ€ ํฌํ•จ๋œ ๋ฐฉ๋Œ€ํ•œ ๋‹จ์–ด ์–ดํœ˜๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๊ณ , '์–ดํœ˜ ์™ธ ๋‹จ์–ด' ๋ฌธ์ œ๋„ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ช‡๋ช‡ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์—์„œ๋Š” _๋ฌธ์ž_๋ฅผ ์ตœ์†Œ ๋‹จ์œ„๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์–ด ํ‘œํ˜„์„ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค(_์˜ˆ: ELMo [21]์˜ CNN ๋‹จ์–ด ์ธ์ฝ”๋”). ์ตœ๊ทผ์—๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ชจ๋ธ์—์„œ _์„œ๋ธŒ์›Œ๋“œ ํ† ํฐํ™”_๊ฐ€ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฐ”์ดํŠธ ์Œ ์ธ์ฝ”๋”ฉ ํ† ํฐํ™”, ์›Œ๋“œํ”ผ์Šค ํ† ํฐํ™”, ์œ ๋‹ˆ๊ทธ๋žจ ํ† ํฐํ™”๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. HuggingFace๋Š” ํ† ํฐ๋ผ์ด์ €22์— ๋Œ€ํ•œ ๋›ฐ์–ด๋‚œ ์˜จ๋ผ์ธ NLP ๊ฐ•์ขŒ๋ฅผ ์šด์˜ ์ค‘์ด๋ฉฐ, ์ดˆ๋ณด์ž๋Š” ์ด ๊ฐ•์ขŒ๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ๋‹ค์Œ์—์„œ๋Š” ๋Œ€ํ‘œ์ ์ธ ํ† ํฐํ™” ๋ฐฉ๋ฒ• ์„ธ ๊ฐ€์ง€๋ฅผ ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
However, to align the model with human preferences, RLHF (Ouyang et al., 2022) presents a more complex challenge. It involves the development of alignment criteria, the acquisition of human feedback, and final learning adjustments based on this feedback. LLaMA2 (Touvron et al., 2023) is currently one of the publicly available models utilizing RLHF, with other models such as WebGPT (Nakano et al., 2021), InstructGPT (Ouyang et al., 2022), and GPT-4 (OpenAI, 2023) being proprietary. Implementing RLHF is intricate, stemming not only from the need for human annotations but also due to technical challenges. These include overfitting in reward models and instabilities during the Reinforcement Learning training phase (Gao et al., 2022). Additionally, the tedious procedure of training multiple LMs including the model being aligned, the reward model, and the inference model at the same time substantially amplifies memory and computational demands, particularly for larger models. Rafailov et al. (2023) proposed direct preference optimization (DPO) to align models with human preferences instead of complex and unstable RLHF. Nevertheless, one still needs to collect human-labeled preference data in the target language.
๊ทธ๋Ÿฌ๋‚˜ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ๋” ๋ณต์žกํ•œ ๊ณผ์ œ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ •๋ ฌ ๊ธฐ์ค€์˜ ๊ฐœ๋ฐœ, ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์˜ ํš๋“, ์ด ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ตœ์ข… ํ•™์Šต ์กฐ์ •์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ RLHF๋ฅผ ํ™œ์šฉํ•˜๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜๋Š” LLaMA2(Touvron et al., 2023)์ด๋ฉฐ, WebGPT(Nakano et al., 2021), InstructGPT(Ouyang et al., 2022), GPT-4(OpenAI, 2023)์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ๋ชจ๋ธ์€ ๋…์  ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. RLHF๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๊ฒƒ์€ ์ธ๊ฐ„์˜ ์ฃผ์„์ด ํ•„์š”ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๊ธฐ์ˆ ์ ์ธ ๋ฌธ์ œ๋กœ ์ธํ•ด ๋ณต์žกํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๋ณด์ƒ ๋ชจ๋ธ์˜ ๊ณผ์ ํ•ฉ๊ณผ Reinforcement Learning ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ๋ถˆ์•ˆ์ •์„ฑ ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค(Gao et al., 2022). ๋˜ํ•œ ์ •๋ ฌ ๋Œ€์ƒ ๋ชจ๋ธ, ๋ณด์ƒ ๋ชจ๋ธ, ์ถ”๋ก  ๋ชจ๋ธ์„ ํฌํ•จํ•œ ์—ฌ๋Ÿฌ ๊ฐœ์˜ LM์„ ๋™์‹œ์— ํ›ˆ๋ จํ•ด์•ผ ํ•˜๋Š” ๋ฒˆ๊ฑฐ๋กœ์šด ์ ˆ์ฐจ๋กœ ์ธํ•ด ํŠนํžˆ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋ฉ”๋ชจ๋ฆฌ์™€ ๊ณ„์‚ฐ ์š”๊ตฌ๋Ÿ‰์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Rafailov et al. (2023)์€ ๋ณต์žกํ•˜๊ณ  ๋ถˆ์•ˆ์ •ํ•œ RLHF ๋Œ€์‹  ๋ชจ๋ธ์„ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•ด Direct Preference Optimization (DPO)๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋Œ€์ƒ ์–ธ์–ด์—์„œ ์ธ๊ฐ„์ด ๋ ˆ์ด๋ธ”์„ ๋ถ™์ธ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ๋” ๋ณต์žกํ•œ ๊ณผ์ œ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ •๋ ฌ ๊ธฐ์ค€์˜ ๊ฐœ๋ฐœ, ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์˜ ํš๋“, ์ด ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ตœ์ข… ํ•™์Šต ์กฐ์ •์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ RLHF๋ฅผ ํ™œ์šฉํ•˜๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜๋Š” LLaMA2(Touvron et al., 2023)์ด๋ฉฐ, WebGPT(Nakano et al., 2021), InstructGPT(Ouyang et al., 2022), GPT-4(OpenAI, 2023)์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ๋ชจ๋ธ์€ ๋…์  ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. RLHF๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๊ฒƒ์€ ์ธ๊ฐ„์˜ ์ฃผ์„์ด ํ•„์š”ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๊ธฐ์ˆ ์ ์ธ ๋ฌธ์ œ๋กœ ์ธํ•ด ๋ณต์žกํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๋ณด์ƒ ๋ชจ๋ธ์˜ ๊ณผ์ ํ•ฉ๊ณผ ๊ฐ•ํ™” ํ•™์Šต ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ๋ถˆ์•ˆ์ •์„ฑ ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค(Gao et al., 2022). ๋˜ํ•œ ์ •๋ ฌ ๋Œ€์ƒ ๋ชจ๋ธ, ๋ณด์ƒ ๋ชจ๋ธ, ์ถ”๋ก  ๋ชจ๋ธ์„ ํฌํ•จํ•œ ์—ฌ๋Ÿฌ ๊ฐœ์˜ LM์„ ๋™์‹œ์— ํ›ˆ๋ จํ•ด์•ผ ํ•˜๋Š” ๋ฒˆ๊ฑฐ๋กœ์šด ์ ˆ์ฐจ๋กœ ์ธํ•ด ํŠนํžˆ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋ฉ”๋ชจ๋ฆฌ์™€ ๊ณ„์‚ฐ ์š”๊ตฌ๋Ÿ‰์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Rafailov et al. (2023)์€ ๋ณต์žกํ•˜๊ณ  ๋ถˆ์•ˆ์ •ํ•œ RLHF ๋Œ€์‹  ๋ชจ๋ธ์„ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•ด ์ง์ ‘ ์„ ํ˜ธ๋„ ์ตœ์ ํ™”(DPO)๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋Œ€์ƒ ์–ธ์–ด์—์„œ ์ธ๊ฐ„์ด ๋ ˆ์ด๋ธ”์„ ๋ถ™์ธ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Results.Figure 6 shows the performance of GPT3 model and its instruction-tuned counterparts on this newly written instruction set (w. inter-rater agreement \(\kappa=0.57\) on the 4-class categorical scale, see Appendix B for details). As anticipated, the vanilla GPT3 LM is largely unable to respond to instructions, and all instruction-tuned models demonstrate comparatively higher performance. Nonetheless, GPT3SELF-INST (i.e., GPT3 model finetuned with Self-INstruct) outperforms those counterparts trained on T0 or SuperNI data by a large margin, demonstrating the value of the generated data despite the noise. Compared with InstructGPT001, GPT3SELF-INST is quite close in performance--if we count acceptable response with minor imperfections (Rating-B) as valid, GPT3SELF-INST is only 5% behind InstructGPT001. Lastly, our evaluation confirms the impressive instruction-following ability of InstructGPT002 and InstructGPT003. Although there are many factors behind this success, we conjecture that future work can largely benefit from improving the quality of our generated data by using human annotators or training a reward model to select better generations, similar to the algorithm used by Ouyang et al. (2022).
๊ฒฐ๊ณผ. ๊ทธ๋ฆผ 6์€ ์ƒˆ๋กœ ์ž‘์„ฑ๋œ instruction ์ง‘ํ•ฉ์— ๋Œ€ํ•ด GPT3 ๋ชจ๋ธ๊ณผ ๊ทธ instruction-tuned ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(4๋“ฑ๊ธ‰ ๋ฒ”์ฃผํ˜• ์ฒ™๋„์—์„œ rater ๊ฐ„ ์ผ์น˜๋„ \(\kappa=0.57\), ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก B ์ฐธ์กฐ). ์˜ˆ์ƒ๋Œ€๋กœ, ๋ฐ”๋‹๋ผ GPT3 LM์€ instruction์— ๋Œ€์‘ํ•˜์ง€ ๋ชปํ•˜๋ฉฐ, ๋ชจ๋“  instruction-tuned ๋ชจ๋ธ์€ ์ƒ๋Œ€์ ์œผ๋กœ ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  GPT3SELF-INST(์ฆ‰, Self-INstruct๋กœ Fine-tuning๋œ GPT3 ๋ชจ๋ธ)๋Š” T0 ๋˜๋Š” SuperNI ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ ๋Œ€์‘๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ๋…ธ์ด์ฆˆ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€์น˜๋ฅผ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. InstructGPT001๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, GPT3SELF-INST๋Š” ์„ฑ๋Šฅ์ด ์ƒ๋‹นํžˆ ๊ทผ์ ‘ํ•ฉ๋‹ˆ๋‹ค. ์•ฝ๊ฐ„์˜ ๋ถˆ์™„์ „์„ฑ(๋“ฑ๊ธ‰-B)์„ ๊ฐ–์ถ˜ ํ—ˆ์šฉ ๊ฐ€๋Šฅํ•œ ์‘๋‹ต์„ ์œ ํšจํ•œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•œ๋‹ค๋ฉด, GPT3SELF-INST๋Š” InstructGPT001๋ณด๋‹ค 5% ๋’ค์ฒ˜์งˆ ๋ฟ์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ด๋ฒˆ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด InstructGPT002์™€ InstructGPT003์˜ ์ธ์ƒ์ ์ธ instruction ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ฑ๊ณต์—๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์š”์ธ์ด ์žˆ์ง€๋งŒ, ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ธ๊ฐ„ ์ฃผ์„๊ฐ€๋ฅผ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ Ouyang ๋“ฑ(2022)์ด ์‚ฌ์šฉํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ๋” ๋‚˜์€ ์ƒ์„ฑ์„ ์„ ํƒํ•จ์œผ๋กœ์จ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ํฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.
๊ฒฐ๊ณผ. ๊ทธ๋ฆผ 6์€ ์ƒˆ๋กœ ์ž‘์„ฑ๋œ ๋ช…๋ น์–ด ์ง‘ํ•ฉ์— ๋Œ€ํ•ด GPT3 ๋ชจ๋ธ๊ณผ ๊ทธ ๋ช…๋ น์–ด์— ๋งž๊ฒŒ ์กฐ์ •๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(4๋“ฑ๊ธ‰ ๋ฒ”์ฃผํ˜• ์ฒ™๋„์—์„œ rater ๊ฐ„ ์ผ์น˜๋„ \(\kappa=0.57\), ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก B ์ฐธ์กฐ). ์˜ˆ์ƒ๋Œ€๋กœ, ๋ฐ”๋‹๋ผ GPT3 LM์€ ๋ช…๋ น์— ๋Œ€์‘ํ•˜์ง€ ๋ชปํ•˜๋ฉฐ, ๋ชจ๋“  ๋ช…๋ น ์กฐ์ • ๋ชจ๋ธ์€ ์ƒ๋Œ€์ ์œผ๋กœ ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  GPT3SELF-INST(์ฆ‰, Self-INstruct๋กœ ๋ฏธ์„ธ ์กฐ์ •๋œ GPT3 ๋ชจ๋ธ)๋Š” T0 ๋˜๋Š” SuperNI ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ ๋Œ€์‘๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ๋…ธ์ด์ฆˆ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€์น˜๋ฅผ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. InstructGPT001๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, GPT3SELF-INST๋Š” ์„ฑ๋Šฅ์ด ์ƒ๋‹นํžˆ ๊ทผ์ ‘ํ•ฉ๋‹ˆ๋‹ค. ์•ฝ๊ฐ„์˜ ๋ถˆ์™„์ „์„ฑ(๋“ฑ๊ธ‰-B)์„ ๊ฐ–์ถ˜ ํ—ˆ์šฉ ๊ฐ€๋Šฅํ•œ ์‘๋‹ต์„ ์œ ํšจํ•œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•œ๋‹ค๋ฉด, GPT3SELF-INST๋Š” InstructGPT001๋ณด๋‹ค 5% ๋’ค์ฒ˜์งˆ ๋ฟ์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ด๋ฒˆ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด InstructGPT002์™€ InstructGPT003์˜ ์ธ์ƒ์ ์ธ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ฑ๊ณต์—๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์š”์ธ์ด ์žˆ์ง€๋งŒ, ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ธ๊ฐ„ ์ฃผ์„๊ฐ€๋ฅผ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ Ouyang ๋“ฑ(2022)์ด ์‚ฌ์šฉํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ๋” ๋‚˜์€ ์ƒ์„ฑ์„ ์„ ํƒํ•จ์œผ๋กœ์จ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ํฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.
Given the poor performance of German models on all German evaluation tasks except HellaSwag (the same as English models on average), we further investigated their understanding of German by conducting a short qualitative study of model generations. In section A.4 of the appendix, we select five German prompts that contain various peculiarities of the German language (see Tab. 8 of the appendix). We then generate a fixed token-length response for each of the models trained German Common Crawl. As a baseline, we also evaluate the model trained only on the Pile. Despite the poor quality of generations at small model scale, we find that there is an observable improvement in the generative quality of German-language outputs from the models trained on German Common Crawl when compared to the Pile baseline, which tends to be systematically off-topic. This suggests that while our German-trained models have learned about the language, the evaluation tasks are too difficult to pick it up at the 405M parameter scale. Another reason is that the German dataset is smaller than the English datasets considered, and contains only web-scraped data, as opposed to the more sophisticated English datasets used in this work.
ํ—ฌ๋ผ์Šค์›ฉ(์˜์–ด ๋ชจ๋ธ๊ณผ ํ‰๊ท ์ ์œผ๋กœ ๋™์ผ)์„ ์ œ์™ธํ•œ ๋ชจ๋“  ๋…์ผ ํ‰๊ฐ€ ๊ณผ์ œ์—์„œ ๋…์ผ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ์ €์กฐํ•œ ์ ์„ ๊ฐ์•ˆํ•˜์—ฌ ๋ชจ๋ธ ์ƒ์„ฑ์— ๋Œ€ํ•œ ์งง์€ ์ •์„ฑ์  ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ๋…์ผ์–ด์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ์ถ”๊ฐ€๋กœ ์กฐ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ถ€๋ก์˜ ์„น์…˜ A.4์—์„œ๋Š” ๋…์ผ์–ด์˜ ๋‹ค์–‘ํ•œ ํŠน์ด์ ์„ ํฌํ•จํ•˜๋Š” 5๊ฐœ์˜ ๋…์ผ์–ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ์„ ์ •ํ•ฉ๋‹ˆ๋‹ค(๋ถ€๋ก์˜ ํ‘œ 8 ์ฐธ์กฐ). ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋…์ผ์–ด Common Crawl๋กœ ํ•™์Šต๋œ ๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ณ ์ • ํ† ํฐ ๊ธธ์ด์˜ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์ค€์„ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” the Pile๋งŒ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ž‘์€ ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ์ƒ์„ฑ ํ’ˆ์งˆ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ฒด๊ณ„์ ์œผ๋กœ ์ฃผ์ œ์—์„œ ๋ฒ—์–ด๋‚˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š” the Pile ๊ธฐ์ค€์„ ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ๋…์ผ์–ด Common Crawl๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ๋…์ผ์–ด ์ถœ๋ ฅ ์ƒ์„ฑ ํ’ˆ์งˆ์ด ๊ด€์ฐฐ ๊ฐ€๋Šฅํ•œ ๊ฐœ์„ ์ด ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋…์ผ์–ด๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ์–ธ์–ด์— ๋Œ€ํ•ด ํ•™์Šตํ–ˆ์ง€๋งŒ ํ‰๊ฐ€ ๊ณผ์ œ๊ฐ€ ๋„ˆ๋ฌด ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์— 405M ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ์—์„œ๋Š” ์ด๋ฅผ ํ”ฝ์—…ํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ด์œ ๋Š” ๋…์ผ์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๊ณ ๋ ค๋œ ์˜์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณด๋‹ค ์ž‘๊ณ , ์ด ์ž‘์—…์— ์‚ฌ์šฉ๋œ ๋” ์ •๊ตํ•œ ์˜์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋‹ฌ๋ฆฌ ์›น ์Šคํฌ๋ž˜ํ•‘๋œ ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
ํ—ฌ๋ผ์Šค์›ฉ(์˜์–ด ๋ชจ๋ธ๊ณผ ํ‰๊ท ์ ์œผ๋กœ ๋™์ผ)์„ ์ œ์™ธํ•œ ๋ชจ๋“  ๋…์ผ ํ‰๊ฐ€ ๊ณผ์ œ์—์„œ ๋…์ผ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ์ €์กฐํ•œ ์ ์„ ๊ฐ์•ˆํ•˜์—ฌ ๋ชจ๋ธ ์ƒ์„ฑ์— ๋Œ€ํ•œ ์งง์€ ์ •์„ฑ์  ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ๋…์ผ์–ด์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ์ถ”๊ฐ€๋กœ ์กฐ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ถ€๋ก์˜ ์„น์…˜ A.4์—์„œ๋Š” ๋…์ผ์–ด์˜ ๋‹ค์–‘ํ•œ ํŠน์ด์ ์„ ํฌํ•จํ•˜๋Š” 5๊ฐœ์˜ ๋…์ผ์–ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ์„ ์ •ํ•ฉ๋‹ˆ๋‹ค(๋ถ€๋ก์˜ ํ‘œ 8 ์ฐธ์กฐ). ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋…์ผ์–ด ์ปค๋จผ ํฌ๋กค๋ง์œผ๋กœ ํ•™์Šต๋œ ๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ณ ์ • ํ† ํฐ ๊ธธ์ด์˜ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์ค€์„ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๋”๋ฏธ๋งŒ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ž‘์€ ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ์ƒ์„ฑ ํ’ˆ์งˆ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ฒด๊ณ„์ ์œผ๋กœ ์ฃผ์ œ์—์„œ ๋ฒ—์–ด๋‚˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š” ๋”๋ฏธ ๊ธฐ์ค€์„ ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ๋…์ผ์–ด ์ปค๋จผ ํฌ๋กค๋ง์œผ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ๋…์ผ์–ด ์ถœ๋ ฅ ์ƒ์„ฑ ํ’ˆ์งˆ์ด ๊ด€์ฐฐ ๊ฐ€๋Šฅํ•œ ๊ฐœ์„ ์ด ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋…์ผ์–ด๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ์–ธ์–ด์— ๋Œ€ํ•ด ํ•™์Šตํ–ˆ์ง€๋งŒ ํ‰๊ฐ€ ๊ณผ์ œ๊ฐ€ ๋„ˆ๋ฌด ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์— 405M ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ์—์„œ๋Š” ์ด๋ฅผ ํ”ฝ์—…ํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ด์œ ๋Š” ๋…์ผ์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๊ณ ๋ ค๋œ ์˜์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณด๋‹ค ์ž‘๊ณ , ์ด ์ž‘์—…์— ์‚ฌ์šฉ๋œ ๋” ์ •๊ตํ•œ ์˜์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋‹ฌ๋ฆฌ ์›น ์Šคํฌ๋ž˜ํ•‘๋œ ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
We validate our _Evol-Instruct_ by fine-tuning open-source LLaMA [4] with our evolved instructions and evaluating its performance similar to existing SOTA works (e.g., Alpaca [31] and Vicuna [22]) on instruction finetuning. The instruction datasets we compare with are the data used by Alpaca (generated using self-instruct [32]) and the 70k ShareGPT (shared by real users) used by Vicuna. To prove that the instruction dataset from our method is superior to human-created instruction datasets, we select Alpaca's training data (generated from only 175 human-created seed instructions) as the initial dataset. We execute four epochs of evolution using OpenAI ChatGPT API5 and finally obtain 250k instructions. To ensure a fair comparison with Vicuna's 70k real user data, we sampled an equal amount from the full 250k data and trained the LLaMA 7B model. We name our model _WizardLM_. Due to the low proportion of difficult instructions in the previous instruction-following test dataset, we manually created a new difficulty-balanced test dataset, named _Evol-Instruct_ testset. We hire annotators and leverage GPT-4 to evaluate Alpaca, Vicuna, ChatGPT, and _WizardLM_ on _Evol-Instruct_ testset and Vicuna's testset.
์šฐ๋ฆฌ๋Š” ์ง„ํ™”ํ•œ instructions๋กœ ์˜คํ”ˆ ์†Œ์Šค LLaMA [4]๋ฅผ Fine-tuningํ•˜๊ณ  instruction Fine-tuning์— ๋Œ€ํ•œ ๊ธฐ์กด SOTA ์ž‘์—…(์˜ˆ: Alpaca [31] ๋ฐ Vicuna [22])๊ณผ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜์—ฌ _Evol-Instruct_์˜ ์œ ํšจ์„ฑ์„ ๊ฒ€์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ๋น„๊ตํ•˜๋Š” instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” Alpaca์—์„œ ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ(self-instruct [32]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ๋จ)์™€ Vicuna์—์„œ ์‚ฌ์šฉํ•œ 70k ShareGPT(์‹ค์ œ ์‚ฌ์šฉ์ž๊ฐ€ ๊ณต์œ ํ•œ ๋ฐ์ดํ„ฐ)์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์‚ฌ๋žŒ์ด ๋งŒ๋“  instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณด๋‹ค ์šฐ์›”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์ฆ๋ช…ํ•˜๊ธฐ ์œ„ํ•ด Alpaca์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ(175๊ฐœ์˜ ์‚ฌ๋žŒ์ด ๋งŒ๋“  ์‹œ๋“œ instructions๋กœ๋งŒ ์ƒ์„ฑ๋จ)๋ฅผ ์ดˆ๊ธฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. OpenAI ChatGPT API5๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 4ํšŒ์˜ ์ง„ํ™”๋ฅผ ์‹คํ–‰ํ•˜๊ณ  ์ตœ์ข…์ ์œผ๋กœ 25๋งŒ ๊ฐœ์˜ instructions๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. Vicuna์˜ 70,000๊ฐœ์˜ ์‹ค์ œ ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ์™€์˜ ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด ์ „์ฒด 25๋งŒ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์—์„œ ๋™์ผํ•œ ์–‘์„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ LLaMA 7B ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ์ด๋ฆ„์€ _WizardLM_์ž…๋‹ˆ๋‹ค. ์ด์ „ instruction ์ˆ˜ํ–‰ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋‚œ์ด๋„๊ฐ€ ๋‚ฎ์•˜๊ธฐ ๋•Œ๋ฌธ์— ๋‚œ์ด๋„ ๊ท ํ˜•์ด ์žกํžŒ ์ƒˆ๋กœ์šด ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜๋™์œผ๋กœ ๋งŒ๋“ค์—ˆ์œผ๋ฉฐ, _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ผ๊ณ  ๋ช…๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ฃผ์„๊ฐ€๋ฅผ ๊ณ ์šฉํ•˜๊ณ  GPT-4๋ฅผ ํ™œ์šฉํ•˜์—ฌ Alpaca, Vicuna, ChatGPT ๋ฐ _WizardLM_์„ _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ Vicuna์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
์šฐ๋ฆฌ๋Š” ์ง„ํ™”ํ•œ ๋ช…๋ น์–ด๋กœ ์˜คํ”ˆ ์†Œ์Šค LLaMA [4]๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๊ณ  ๋ช…๋ น์–ด ๋ฏธ์„ธ ์กฐ์ •์— ๋Œ€ํ•œ ๊ธฐ์กด SOTA ์ž‘์—…(์˜ˆ: ์•ŒํŒŒ์นด [31] ๋ฐ ๋น„์ฟ ๋‚˜ [22])๊ณผ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜์—ฌ _Evol-Instruct_์˜ ์œ ํšจ์„ฑ์„ ๊ฒ€์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ๋น„๊ตํ•˜๋Š” ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์•ŒํŒŒ์นด์—์„œ ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ(self-instruct [32]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ๋จ)์™€ ๋น„์ฟ ๋‚˜์—์„œ ์‚ฌ์šฉํ•œ 70k ShareGPT(์‹ค์ œ ์‚ฌ์šฉ์ž๊ฐ€ ๊ณต์œ ํ•œ ๋ฐ์ดํ„ฐ)์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์‚ฌ๋žŒ์ด ๋งŒ๋“  ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณด๋‹ค ์šฐ์›”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์ฆ๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ์•ŒํŒŒ์นด์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ(175๊ฐœ์˜ ์‚ฌ๋žŒ์ด ๋งŒ๋“  ์‹œ๋“œ ๋ช…๋ น์–ด๋กœ๋งŒ ์ƒ์„ฑ๋จ)๋ฅผ ์ดˆ๊ธฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. OpenAI ChatGPT API5๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 4ํšŒ์˜ ์ง„ํ™”๋ฅผ ์‹คํ–‰ํ•˜๊ณ  ์ตœ์ข…์ ์œผ๋กœ 25๋งŒ ๊ฐœ์˜ ๋ช…๋ น์–ด๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. ๋น„์ฟ ๋‚˜์˜ 70,000๊ฐœ์˜ ์‹ค์ œ ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ์™€์˜ ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด ์ „์ฒด 25๋งŒ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์—์„œ ๋™์ผํ•œ ์–‘์„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ LLaMA 7B ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ์ด๋ฆ„์€ _WizardLM_์ž…๋‹ˆ๋‹ค. ์ด์ „ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋‚œ์ด๋„๊ฐ€ ๋‚ฎ์•˜๊ธฐ ๋•Œ๋ฌธ์— ๋‚œ์ด๋„ ๊ท ํ˜•์ด ์žกํžŒ ์ƒˆ๋กœ์šด ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜๋™์œผ๋กœ ๋งŒ๋“ค์—ˆ์œผ๋ฉฐ, _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ผ๊ณ  ๋ช…๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ฃผ์„๊ฐ€๋ฅผ ๊ณ ์šฉํ•˜๊ณ  GPT-4๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์•ŒํŒŒ์นด, ๋น„์ฟ ๋‚˜, ChatGPT ๋ฐ _WizardLM_์„ _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ ๋น„์ฟ ๋‚˜์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
We therefore introduce Contrastive Filtering, a method to select the instructions that can effectively enhance the target LLM \(f_{t}\). For clarity, we define the space of all natural language sequences as \(\mathcal{N}\). We have the strong LLM \(f_{s}:\mathcal{N}\rightarrow\mathcal{N}\), the target LLM \(f_{t}:\mathcal{N}\rightarrow\mathcal{N}\), and a scoring function \(S:\mathcal{N}\rightarrow\mathbb{R}\) to evaluate response quality. In practice, \(S\) is obtained by reusing the strong LLM \(f_{s}\) with a prompt template (Figure 11, Appendix A.9) adapted from the Vicuna pairwise evaluation template (Taori et al., 2023; Chiang et al., 2023). To mitigate potential position bias, we average the scores obtained by exchanging the positions of two responses (Chiang et al., 2023). We observe using \(f_{s}\) for scoring works quite well in practice, so we prioritize this option for simplicity. Given an input instruction \(I\in\mathcal{N}\), we obtain responses from both LLMs as \(f_{s}(I)\) and \(f_{t}(I)\), respectively. We then define the _quality gap_\(G:\mathcal{N}\rightarrow\mathbb{R}\) between these responses to estimate the _effectiveness_ of the instruction: \(G(I)=S(f_{s}(I))-S(f_{t}(I))\).
๋”ฐ๋ผ์„œ ๋Œ€์ƒ LLM(f_{t})์„ ํšจ๊ณผ์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” instructions๋ฅผ ์„ ํƒํ•˜๋Š” ๋ฐฉ๋ฒ•์ธ Contrastive Filtering์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋ช…ํ™•์„ฑ์„ ์œ„ํ•ด ๋ชจ๋“  ์ž์—ฐ์–ด ์‹œํ€€์Šค์˜ ๊ณต๊ฐ„์„ \(\mathcal{N}\)์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ LLM \(f_{s}:\mathcal{N}\rightarrow\mathcal{N}\), ๋Œ€์ƒ LLM \(f_{t}:\mathcal{N}\rightarrow\mathcal{N}\), ์‘๋‹ต ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ scoring function \(S:\mathcal{N}\rightarrow\mathbb{R}\)๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ \(S\)๋Š” Vicuna ์Œ๋ณ„ ํ‰๊ฐ€ ํ…œํ”Œ๋ฆฟ(Taori et al., 2023; Chiang et al., 2023)์—์„œ ์ˆ˜์ •ํ•œ prompt ํ…œํ”Œ๋ฆฟ(๊ทธ๋ฆผ 11, ๋ถ€๋ก A.9)์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ•๋ ฅํ•œ LLM \(f_{s}\)๋ฅผ ์žฌ์‚ฌ์šฉํ•˜์—ฌ ์–ป์Šต๋‹ˆ๋‹ค. ์ž ์žฌ์ ์ธ ์œ„์น˜ ํŽธํ–ฅ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ์‘๋‹ต์˜ ์œ„์น˜๋ฅผ ๊ตํ™˜ํ•˜์—ฌ ์–ป์€ ์ ์ˆ˜๋ฅผ ํ‰๊ท ํ™”ํ•ฉ๋‹ˆ๋‹ค(Chiang et al., 2023). ์ ์ˆ˜ ๋ถ€์—ฌ์— \(f_{s}\)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์‹ค์ œ๋กœ ๋งค์šฐ ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋‹จ์ˆœํ™”๋ฅผ ์œ„ํ•ด ์ด ์˜ต์…˜์„ ์šฐ์„ ์ ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ž…๋ ฅ instruction \(I\in\mathcal{N}\)๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ๋‘ LLM ๋ชจ๋‘์—์„œ ๊ฐ๊ฐ \(f_{s}(I)\)์™€ \(f_{t}(I)\)๋กœ ์‘๋‹ต์„ ์–ป์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ์‘๋‹ต ๊ฐ„์˜ _quality gap_ \(G:\mathcal{N}\rightarrow\mathbb{R}\)์„ ์ •์˜ํ•˜์—ฌ instruction์˜ _effectiveness_๋ฅผ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค: \(G(I)=S(f_{s}(I))-S(f_{t}(I))\).
๋”ฐ๋ผ์„œ ๋Œ€์ƒ LLM(f_{t})์„ ํšจ๊ณผ์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๋ช…๋ น์„ ์„ ํƒํ•˜๋Š” ๋ฐฉ๋ฒ•์ธ ๋Œ€์กฐ ํ•„ํ„ฐ๋ง์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋ช…ํ™•์„ฑ์„ ์œ„ํ•ด ๋ชจ๋“  ์ž์—ฐ์–ด ์‹œํ€€์Šค์˜ ๊ณต๊ฐ„์„ \(\mathcal{N}\)์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ LLM \(f_{s}:\mathcal{N}\rightarrow\mathcal{N}\), ๋Œ€์ƒ LLM \(f_{t}:\mathcal{N}\rightarrow\mathcal{N}\), ์‘๋‹ต ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ์ ์ˆ˜ ํ•จ์ˆ˜ \(S:\mathcal{N}\rightarrow\mathbb{R}\)๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ \(S\)๋Š” Vicuna ์Œ๋ณ„ ํ‰๊ฐ€ ํ…œํ”Œ๋ฆฟ(Taori et al., 2023; Chiang et al., 2023)์—์„œ ์ˆ˜์ •ํ•œ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ(๊ทธ๋ฆผ 11, ๋ถ€๋ก A.9)์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ•๋ ฅํ•œ LLM \(f_{s}\)๋ฅผ ์žฌ์‚ฌ์šฉํ•˜์—ฌ ์–ป์Šต๋‹ˆ๋‹ค. ์ž ์žฌ์ ์ธ ์œ„์น˜ ํŽธํ–ฅ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ์‘๋‹ต์˜ ์œ„์น˜๋ฅผ ๊ตํ™˜ํ•˜์—ฌ ์–ป์€ ์ ์ˆ˜๋ฅผ ํ‰๊ท ํ™”ํ•ฉ๋‹ˆ๋‹ค(Chiang et al., 2023). ์ ์ˆ˜ ๋ถ€์—ฌ์— \(f_{s}\)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์‹ค์ œ๋กœ ๋งค์šฐ ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋‹จ์ˆœํ™”๋ฅผ ์œ„ํ•ด ์ด ์˜ต์…˜์„ ์šฐ์„ ์ ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ž…๋ ฅ ๋ช…๋ น \(I\in\mathcal{N}\)๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ๋‘ LLM ๋ชจ๋‘์—์„œ ๊ฐ๊ฐ \(f_{s}(I)\)์™€ \(f_{t}(I)\)๋กœ ์‘๋‹ต์„ ์–ป์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ์‘๋‹ต ๊ฐ„์˜ _ํ’ˆ์งˆ ๊ฐญ_\(G:\mathcal{N}\rightarrow\mathbb{R}\)์„ ์ •์˜ํ•˜์—ฌ ๋ช…๋ น์˜ _ํšจ๊ณผ์„ฑ_์„ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค: \(G(I)=S(f_{s}(I))-S(f_{t}(I))\).
**LLM as Encoder with Instruction Metadata.** We begin by encoding the given seed instructions \(\mathcal{D}_{s}=\{I_{i}\}_{i=1}^{n}\) into instruction _metadata_\(\mathcal{M}\), _i.e._, keywords that capture the underlying target instruction distribution. Inspired by the task pool by Wang et al. (2022) and the post-hoc analysis on skill distribution by Xu et al. (2023), we define the metadata as encompassing two key aspects: _use case_ and _skills_. Use case describes the intended task (_e.g._, question answering or creative writing), while Skills are the knowledge the LLM required to have to successfully respond to the given instruction (_e.g._, algorithms or communication). Skills are often generalizable to different use cases. Therefore, each instruction has a single use case and may involve multiple skills. To extract this metadata, we leverage the strong LLM \(f_{s}\) following the prompt template in Figure 7, Appendix A.9. While richer definitions are possible based on finer-grained instruction-following metrics Zhou et al. (2023), we prioritize use case and skills for their broad applicability across diverse instruction distributions. Future work can explore extending this metadata further.
**LLM as Encoder with Instruction Metadata.** ์ฃผ์–ด์ง„ ์‹œ๋“œ ๋ช…๋ น์–ด \(\mathcal{D}_{s}=\{I_{i}\}_{i=1}^{n}\)์„ instruction _metadata_\(\mathcal{M}\), ์ฆ‰, ๊ธฐ๋ณธ ํƒ€๊ฒŸ ๋ช…๋ น์–ด ๋ถ„ํฌ๋ฅผ ์บก์ฒ˜ํ•˜๋Š” ํ‚ค์›Œ๋“œ๋กœ ์ธ์ฝ”๋”ฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. Wang ๋“ฑ(2022)์˜ ์ž‘์—… ํ’€๊ณผ Xu ๋“ฑ(2023)์˜ ์‚ฌํ›„ ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด, metadata๋Š” _use case_์™€ _skills_๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ์ธก๋ฉด์„ ํฌ๊ด„ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. Use case๋Š” ์˜๋„ํ•œ ์ž‘์—…(_์˜ˆ: question answering ๋˜๋Š” ์ฐฝ์˜์  ๊ธ€์“ฐ๊ธฐ)์„ ์„ค๋ช…ํ•˜๋ฉฐ, Skills๋Š” ์ฃผ์–ด์ง„ instruction์— ์„ฑ๊ณต์ ์œผ๋กœ ์‘๋‹ตํ•˜๊ธฐ ์œ„ํ•ด LLM์ด ๊ฐ–์ถฐ์•ผ ํ•˜๋Š” ์ง€์‹(_์˜ˆ: algorithms ๋˜๋Š” communication)์ž…๋‹ˆ๋‹ค. Skills๋Š” ์ข…์ข… ๋‹ค์–‘ํ•œ use case์— ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ instruction์—๋Š” ํ•˜๋‚˜์˜ use case๊ฐ€ ์žˆ์œผ๋ฉฐ ์—ฌ๋Ÿฌ skills์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด metadata๋ฅผ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 7์˜ prompt ํ…œํ”Œ๋ฆฟ์— ๋”ฐ๋ผ ๊ฐ•๋ ฅํ•œ LLM \(f_{s}\)์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค(๋ถ€๋ก A.9). ๋ณด๋‹ค ์„ธ๋ถ„ํ™”๋œ instruction ์ˆ˜ํ–‰ ๋ฉ”ํŠธ๋ฆญ์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ๋” ํ’๋ถ€ํ•œ ์ •์˜๊ฐ€ ๊ฐ€๋Šฅํ•˜์ง€๋งŒ, ๋‹ค์–‘ํ•œ instruction ๋ถ„ํฌ์— ๊ฑธ์ณ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ ์šฉ ๊ฐ€๋Šฅํ•œ use case์™€ skills์„ ์šฐ์„ ์ ์œผ๋กœ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด metadata๋ฅผ ๋”์šฑ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**๋ช…๋ น ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๊ฐ€ ์žˆ๋Š” ์ธ์ฝ”๋”๋กœ์„œ์˜ LLM**. ์ฃผ์–ด์ง„ ์‹œ๋“œ ๋ช…๋ น์–ด \mathcal{D}_{s}=\{I_{i}\}_{i=1}^{n}\)์„ ๋ช…๋ น์–ด _๋ฉ”ํƒ€๋ฐ์ดํ„ฐ_\mathcal{M}, ์ฆ‰, ๊ธฐ๋ณธ ํƒ€๊ฒŸ ๋ช…๋ น์–ด ๋ถ„ํฌ๋ฅผ ์บก์ฒ˜ํ•˜๋Š” ํ‚ค์›Œ๋“œ๋กœ ์ธ์ฝ”๋”ฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. Wang ๋“ฑ(2022)์˜ ์ž‘์—… ํ’€๊ณผ Xu ๋“ฑ(2023)์˜ ์‚ฌํ›„ ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด, ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋Š” ์‚ฌ์šฉ ์‚ฌ๋ก€_์™€ _์Šคํ‚ฌ_์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ์ธก๋ฉด์„ ํฌ๊ด„ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ ์‚ฌ๋ก€๋Š” ์˜๋„ํ•œ ์ž‘์—…(_์˜ˆ: ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋˜๋Š” ์ฐฝ์˜์  ๊ธ€์“ฐ๊ธฐ)์„ ์„ค๋ช…ํ•˜๋ฉฐ, ์Šคํ‚ฌ์€ ์ฃผ์–ด์ง„ ๋ช…๋ น์— ์„ฑ๊ณต์ ์œผ๋กœ ์‘๋‹ตํ•˜๊ธฐ ์œ„ํ•ด LLM์ด ๊ฐ–์ถฐ์•ผ ํ•˜๋Š” ์ง€์‹(_์˜ˆ: ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๋˜๋Š” ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜)์ž…๋‹ˆ๋‹ค. ์Šคํ‚ฌ์€ ์ข…์ข… ๋‹ค์–‘ํ•œ ์‚ฌ์šฉ ์‚ฌ๋ก€์— ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ ๋ช…๋ น์—๋Š” ํ•˜๋‚˜์˜ ์‚ฌ์šฉ ์‚ฌ๋ก€๊ฐ€ ์žˆ์œผ๋ฉฐ ์—ฌ๋Ÿฌ ์Šคํ‚ฌ์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 7์˜ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์— ๋”ฐ๋ผ ๊ฐ•๋ ฅํ•œ LLM \(f_{s}\)์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค(๋ถ€๋ก A.9). ๋ณด๋‹ค ์„ธ๋ถ„ํ™”๋œ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ๋ฉ”ํŠธ๋ฆญ์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ๋” ํ’๋ถ€ํ•œ ์ •์˜๊ฐ€ ๊ฐ€๋Šฅํ•˜์ง€๋งŒ, ๋‹ค์–‘ํ•œ ๋ช…๋ น์–ด ๋ถ„ํฌ์— ๊ฑธ์ณ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ ์šฉ ๊ฐ€๋Šฅํ•œ ์‚ฌ์šฉ ์‚ฌ๋ก€์™€ ์Šคํ‚ฌ์„ ์šฐ์„ ์ ์œผ๋กœ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ๋”์šฑ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Faced with both opportunities and challenges, it needs more attention on the research and development of LLMs. Inorder to provide a basic understanding of LLMs, this survey conducts a literature review of the recent advances in LLMs from four major aspects, including _pre-training_ (how to pre-train a capable LLM), _adaptation_ (how to effectively adapt pre-trained LLMs for better use), _utilization_ (how to use LLMs for solving various downstream tasks) and _capability evaluation_ (how to evaluate the abilities of LLMs and existing empirical findings). We thoroughly comb the literature and summarize the key findings, techniques, and methods of LLMs. For this survey, we also create a GitHub project website by collecting the supporting resources for LLMs, at the link [https://github.com/RUCAlBox/LLMSurvey](https://github.com/RUCAlBox/LLMSurvey). We are also aware of several related review articles on PLMs or LLMs [32, 36, 38, 39, 43, 48, 49, 50, 51, 52, 53, 54]. These papers either discuss PLMs or some specific (or general) aspects of LLMs. Compared with them, we focus on the techniques and methods to develop and use LLMs and provide a relatively comprehensive reference to important aspects of LLMs.
๊ธฐํšŒ์™€ ๋„์ „์— ์ง๋ฉดํ•œ LLM์˜ ์—ฐ๊ตฌ ๊ฐœ๋ฐœ์— ๋” ๋งŽ์€ attention์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” LLM์— ๋Œ€ํ•œ ๊ธฐ๋ณธ์ ์ธ ์ดํ•ด๋ฅผ ๋•๊ธฐ ์œ„ํ•ด _pre-training_ (์‚ฌ์ „ ํ•™์Šต ๋ฐฉ๋ฒ•), _adaptation_ (์‚ฌ์ „ ํ•™์Šต๋œ LLM์„ ๋” ํšจ๊ณผ์ ์œผ๋กœ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•œ ๋ฐฉ๋ฒ•), _utilization_ (๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•), _capability evaluation_ (LLM์˜ ๋Šฅ๋ ฅ๊ณผ ๊ธฐ์กด ๊ฒฝํ—˜์  ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•) ๋“ฑ ๋„ค ๊ฐ€์ง€ ์ฃผ์š” ์ธก๋ฉด์—์„œ ์ตœ๊ทผ LLM์˜ ๋ฐœ์ „์— ๋Œ€ํ•œ ๋ฌธํ—Œ ๊ฒ€ํ† ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๋ฌธํ—Œ์„ ๋ฉด๋ฐ€ํžˆ ๊ฒ€ํ† ํ•˜๊ณ  LLM์˜ ์ฃผ์š” ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, ๊ธฐ์ˆ  ๋ฐ ๋ฐฉ๋ฒ•์„ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด ์„ค๋ฌธ์กฐ์‚ฌ์—์„œ๋Š” LLM์— ๋Œ€ํ•œ ์ง€์› ๋ฆฌ์†Œ์Šค๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ GitHub ํ”„๋กœ์ ํŠธ ์›น์‚ฌ์ดํŠธ๋ฅผ ๋งŒ๋“ค์—ˆ์œผ๋ฉฐ, ๋งํฌ [https://github.com/RUCAlBox/LLMSurvey](https://github.com/RUCAlBox/LLMSurvey)์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ PLM ๋˜๋Š” LLM์— ๊ด€ํ•œ ๋ช‡ ๊ฐ€์ง€ ๊ด€๋ จ ๋ฆฌ๋ทฐ ๊ธฐ์‚ฌ [32, 36, 38, 39, 43, 48, 49, 50, 51, 52, 53, 54]๋„ ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…ผ๋ฌธ์€ PLM ๋˜๋Š” LLM์˜ ํŠน์ •(๋˜๋Š” ์ผ๋ฐ˜์ ์ธ) ์ธก๋ฉด์„ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด์™€ ๋น„๊ตํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” LLM์„ ๊ฐœ๋ฐœํ•˜๊ณ  ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ์ˆ ๊ณผ ๋ฐฉ๋ฒ•์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  LLM์˜ ์ค‘์š”ํ•œ ์ธก๋ฉด์— ๋Œ€ํ•ด ๋น„๊ต์  ํฌ๊ด„์ ์ธ ์ฐธ๊ณ  ์ž๋ฃŒ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๊ธฐํšŒ์™€ ๋„์ „์— ์ง๋ฉดํ•œ LLM์˜ ์—ฐ๊ตฌ ๊ฐœ๋ฐœ์— ๋” ๋งŽ์€ ๊ด€์‹ฌ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” LLM์— ๋Œ€ํ•œ ๊ธฐ๋ณธ์ ์ธ ์ดํ•ด๋ฅผ ๋•๊ธฐ ์œ„ํ•ด ์‚ฌ์ „ ํ•™์Šต(_์‚ฌ์ „ ํ•™์Šต ๋ฐฉ๋ฒ•), ์ ์‘(_์‚ฌ์ „ ํ•™์Šต๋œ LLM์„ ๋” ํšจ๊ณผ์ ์œผ๋กœ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•œ ๋ฐฉ๋ฒ•), ํ™œ์šฉ(_๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•), ๋Šฅ๋ ฅ ํ‰๊ฐ€(_LLM์˜ ๋Šฅ๋ ฅ๊ณผ ๊ธฐ์กด ๊ฒฝํ—˜์  ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•) ๋“ฑ ๋„ค ๊ฐ€์ง€ ์ฃผ์š” ์ธก๋ฉด์—์„œ ์ตœ๊ทผ LLM์˜ ๋ฐœ์ „์— ๋Œ€ํ•œ ๋ฌธํ—Œ ๊ฒ€ํ† ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๋ฌธํ—Œ์„ ๋ฉด๋ฐ€ํžˆ ๊ฒ€ํ† ํ•˜๊ณ  LLM์˜ ์ฃผ์š” ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, ๊ธฐ์ˆ  ๋ฐ ๋ฐฉ๋ฒ•์„ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด ์„ค๋ฌธ์กฐ์‚ฌ์—์„œ๋Š” LLM์— ๋Œ€ํ•œ ์ง€์› ๋ฆฌ์†Œ์Šค๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ GitHub ํ”„๋กœ์ ํŠธ ์›น์‚ฌ์ดํŠธ๋ฅผ ๋งŒ๋“ค์—ˆ์œผ๋ฉฐ, ๋งํฌ [https://github.com/RUCAlBox/LLMSurvey](https://github.com/RUCAlBox/LLMSurvey)์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ PLM ๋˜๋Š” LLM์— ๊ด€ํ•œ ๋ช‡ ๊ฐ€์ง€ ๊ด€๋ จ ๋ฆฌ๋ทฐ ๊ธฐ์‚ฌ [32, 36, 38, 39, 43, 48, 49, 50, 51, 52, 53, 54]๋„ ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…ผ๋ฌธ์€ PLM ๋˜๋Š” LLM์˜ ํŠน์ •(๋˜๋Š” ์ผ๋ฐ˜์ ์ธ) ์ธก๋ฉด์„ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด์™€ ๋น„๊ตํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” LLM์„ ๊ฐœ๋ฐœํ•˜๊ณ  ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ์ˆ ๊ณผ ๋ฐฉ๋ฒ•์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  LLM์˜ ์ค‘์š”ํ•œ ์ธก๋ฉด์— ๋Œ€ํ•ด ๋น„๊ต์  ํฌ๊ด„์ ์ธ ์ฐธ๊ณ  ์ž๋ฃŒ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
**Remaining Issues.** In this part, we further discuss several important issues to apply LLMs to improve IR systems. First, though LLMs are capable of being as general-purpose task solvers, they are not directly well suited for existing IR systems: they require high overhead for inference [774, 782], have limitations in modeling long texts or document lists [778], and need special adaptation (_e.g._, instruction tuning) to perform the text ranking task [795]. Therefore, more systematic approaches to adapt LLMs for modern IR systems should be investigated, to leverage their benefits and meanwhile overcome these limitations. Secondly, the advent of LLMs sheds lights on the development of new information seeking ways (_e.g._, New Bing). It is meaningful to explore how to reshape the architecture and paradigm of IR by integrating the LLMs' capacities and the merits of existing IR systems [796]. Thirdly, existing work mainly focuses on text retrieval tasks, lacking a comprehensive consideration of multimodal information sources. As will be discussed in Section 8.1.4, multimodal large language models [797] are also widely studied, making it feasible to develop more powerful multimedia retrieval systems.
**ํ˜„์•ˆ**. ์ด ๋ถ€๋ถ„์—์„œ๋Š” LLM์„ ์ ์šฉํ•˜์—ฌ IR ์‹œ์Šคํ…œ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ถ”๊ฐ€๋กœ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ์งธ, LLM์€ ๋ฒ”์šฉ ํƒœ์Šคํฌ ์†”๋ฒ„๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ๊ธฐ์กด IR ์‹œ์Šคํ…œ์—๋Š” ์ง์ ‘ ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ถ”๋ก ์„ ์œ„ํ•ด ๋†’์€ ์˜ค๋ฒ„ํ—ค๋“œ๊ฐ€ ํ•„์š”ํ•˜๊ณ [774, 782], ๊ธด ํ…์ŠคํŠธ๋‚˜ ๋ฌธ์„œ ๋ชฉ๋ก์„ ๋ชจ๋ธ๋งํ•˜๋Š” ๋ฐ ํ•œ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ[778], ํ…์ŠคํŠธ ์ˆœ์œ„ ์ง€์ • ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํ•œ ์ ์‘(_์˜ˆ: Instruction Tuning)์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค[795]. ๋”ฐ๋ผ์„œ LLM์˜ ์žฅ์ ์„ ํ™œ์šฉํ•˜๋Š” ๋™์‹œ์— ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด ์ตœ์‹  IR ์‹œ์Šคํ…œ์— LLM์„ ์ ์šฉํ•˜๋Š” ๋ณด๋‹ค ์ฒด๊ณ„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์—ฐ๊ตฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, LLM์˜ ๋“ฑ์žฅ์œผ๋กœ ์ƒˆ๋กœ์šด ์ •๋ณด ํƒ์ƒ‰ ๋ฐฉ์‹(์˜ˆ: New Bing)์˜ ๊ฐœ๋ฐœ์— ์กฐ๋ช…์ด ๋น„์ถฐ์กŒ์Šต๋‹ˆ๋‹ค. LLM์˜ ์—ญ๋Ÿ‰๊ณผ ๊ธฐ์กด IR ์‹œ์Šคํ…œ์˜ ์žฅ์ ์„ ํ†ตํ•ฉํ•˜์—ฌ IR์˜ ์•„ํ‚คํ…์ฒ˜์™€ ํŒจ๋Ÿฌ๋‹ค์ž„์„ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•˜๋Š” ๊ฒƒ์€ ์˜๋ฏธ ์žˆ๋Š” ์ผ์ž…๋‹ˆ๋‹ค[796]. ์…‹์งธ, ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ํ…์ŠคํŠธ ๊ฒ€์ƒ‰ ์ž‘์—…์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์–ด ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์ •๋ณด ์†Œ์Šค์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๊ณ ๋ ค๊ฐ€ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. 8.1.4์ ˆ์—์„œ ๋…ผ์˜ํ•  ๊ฒƒ์ฒ˜๋Ÿผ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ Large Language Models[797]๋„ ๋„๋ฆฌ ์—ฐ๊ตฌ๋˜๊ณ  ์žˆ์–ด ๋ณด๋‹ค ๊ฐ•๋ ฅํ•œ ๋ฉ€ํ‹ฐ๋ฏธ๋””์–ด ๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€๋Šฅํ•ด์กŒ์Šต๋‹ˆ๋‹ค.
**ํ˜„์•ˆ**. ์ด ๋ถ€๋ถ„์—์„œ๋Š” LLM์„ ์ ์šฉํ•˜์—ฌ IR ์‹œ์Šคํ…œ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ถ”๊ฐ€๋กœ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ์งธ, LLM์€ ๋ฒ”์šฉ ํƒœ์Šคํฌ ์†”๋ฒ„๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ๊ธฐ์กด IR ์‹œ์Šคํ…œ์—๋Š” ์ง์ ‘ ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ถ”๋ก ์„ ์œ„ํ•ด ๋†’์€ ์˜ค๋ฒ„ํ—ค๋“œ๊ฐ€ ํ•„์š”ํ•˜๊ณ [774, 782], ๊ธด ํ…์ŠคํŠธ๋‚˜ ๋ฌธ์„œ ๋ชฉ๋ก์„ ๋ชจ๋ธ๋งํ•˜๋Š” ๋ฐ ํ•œ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ[778], ํ…์ŠคํŠธ ์ˆœ์œ„ ์ง€์ • ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํ•œ ์ ์‘(_์˜ˆ: ์ง€์‹œ ์กฐ์ •)์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค[795]. ๋”ฐ๋ผ์„œ LLM์˜ ์žฅ์ ์„ ํ™œ์šฉํ•˜๋Š” ๋™์‹œ์— ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด ์ตœ์‹  IR ์‹œ์Šคํ…œ์— LLM์„ ์ ์šฉํ•˜๋Š” ๋ณด๋‹ค ์ฒด๊ณ„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์—ฐ๊ตฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, LLM์˜ ๋“ฑ์žฅ์œผ๋กœ ์ƒˆ๋กœ์šด ์ •๋ณด ํƒ์ƒ‰ ๋ฐฉ์‹(์˜ˆ: New Bing)์˜ ๊ฐœ๋ฐœ์— ์กฐ๋ช…์ด ๋น„์ถฐ์กŒ์Šต๋‹ˆ๋‹ค. LLM์˜ ์—ญ๋Ÿ‰๊ณผ ๊ธฐ์กด IR ์‹œ์Šคํ…œ์˜ ์žฅ์ ์„ ํ†ตํ•ฉํ•˜์—ฌ IR์˜ ์•„ํ‚คํ…์ฒ˜์™€ ํŒจ๋Ÿฌ๋‹ค์ž„์„ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•˜๋Š” ๊ฒƒ์€ ์˜๋ฏธ ์žˆ๋Š” ์ผ์ž…๋‹ˆ๋‹ค[796]. ์…‹์งธ, ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ํ…์ŠคํŠธ ๊ฒ€์ƒ‰ ์ž‘์—…์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์–ด ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์ •๋ณด ์†Œ์Šค์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๊ณ ๋ ค๊ฐ€ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. 8.1.4์ ˆ์—์„œ ๋…ผ์˜ํ•  ๊ฒƒ์ฒ˜๋Ÿผ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ[797]๋„ ๋„๋ฆฌ ์—ฐ๊ตฌ๋˜๊ณ  ์žˆ์–ด ๋ณด๋‹ค ๊ฐ•๋ ฅํ•œ ๋ฉ€ํ‹ฐ๋ฏธ๋””์–ด ๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€๋Šฅํ•ด์กŒ์Šต๋‹ˆ๋‹ค.
**Gain from Model Scale.**: We observe that Yi-34B has substantial performance improvement compared to Yi-6B, though they utilized the same pretrain corpora. Larger model size leads to higher performance gain on Code and Math benchmarks, referring to Tab. 3, compared to benchmarks focusing on Commonsense Reasoning, Reading Comprehension, or Knowledge. **Data Quality.**: Smaller models of higher quality pretrain data, like Yi-34B or Qwen-14B, usually demonstrate better performance than models of larger size but (presumably) lower quality data, such as Falcon-180B (though the focus of Falcon-180B might be more on the scaling side, which is definitely of important value on its own). **Gap between GPT-4 and Open-source LLMs.**: Based on Tab. 2, we note that open-source LLMs still lag behind the performance of GPT-4 and GPT-3.5 on various benchmarks. Yet representative bilingual LLMs, e.g. Qwen-14B and Yi-34B, can match or even surpass the performance of GPT-4 on Chinese knowledge related benchmarks, including C-Eval [31], CMMLU [42], and Gaokao [90]. However, there is still a huge gap between GPT-4 and open-source models on reasoning-related benchmarks like BBH [72], code (HumanEval), and math (MATH).
**๋ชจ๋ธ ๊ทœ๋ชจ๋กœ๋ถ€ํ„ฐ์˜ ์ด๋“**: Yi-34B๋Š” ๋™์ผํ•œ pretrain corpora๋ฅผ ํ™œ์šฉํ–ˆ์ง€๋งŒ Yi-6B์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ํด์ˆ˜๋ก Commonsense Reasoning, Reading Comprehension ๋˜๋Š” Knowledge์— ์ดˆ์ ์„ ๋งž์ถ˜ ๋ฒค์น˜๋งˆํฌ์— ๋น„ํ•ด Code ๋ฐ Math ๋ฒค์น˜๋งˆํฌ์—์„œ ๋” ๋†’์€ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์˜ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(ํ‘œ 3 ์ฐธ์กฐ). **๋ฐ์ดํ„ฐ ํ’ˆ์งˆ**: Yi-34B ๋˜๋Š” Qwen-14B์™€ ๊ฐ™์ด ๋” ๋†’์€ ํ’ˆ์งˆ์˜ pretrain ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์†Œํ˜• ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ Falcon-180B์™€ ๊ฐ™์ด ๋” ํฐ ํฌ๊ธฐ์ด์ง€๋งŒ (์•„๋งˆ๋„) ํ’ˆ์งˆ์ด ๋‚ฎ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(Falcon-180B์˜ ์ดˆ์ ์€ scaling ์ธก๋ฉด์— ๋” ์žˆ์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ์ด๋Š” ๊ทธ ์ž์ฒด๋กœ ์ค‘์š”ํ•œ ๊ฐ€์น˜๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค). **GPT-4์™€ ์˜คํ”ˆ ์†Œ์Šค LLM ๊ฐ„์˜ ๊ฒฉ์ฐจ**: ํ‘œ 2๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์˜คํ”ˆ ์†Œ์Šค LLM์€ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ GPT-4 ๋ฐ GPT-3.5์˜ ์„ฑ๋Šฅ์— ์—ฌ์ „ํžˆ ๋’ค์ฒ˜์ง€๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋Œ€ํ‘œ์ ์ธ ์ด์ค‘ ์–ธ์–ด LLM์ธ Qwen-14B์™€ Yi-34B๋Š” C-Eval [31], CMMLU [42], Gaokao [90] ๋“ฑ ์ค‘๊ตญ์–ด ์ง€์‹ ๊ด€๋ จ ๋ฒค์น˜๋งˆํฌ์—์„œ GPT-4์˜ ์„ฑ๋Šฅ์— ํ•„์ ํ•˜๊ฑฐ๋‚˜ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ reasoning ๊ด€๋ จ ๋ฒค์น˜๋งˆํฌ(์˜ˆ: BBH [72], code (HumanEval), math (MATH))์—์„œ๋Š” GPT-4์™€ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ ๊ฐ„์— ์—ฌ์ „ํžˆ ํฐ ๊ฒฉ์ฐจ๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.
**๋ชจ๋ธ ๊ทœ๋ชจ๋กœ๋ถ€ํ„ฐ์˜ ์ด๋“**: Yi-34B๋Š” ๋™์ผํ•œ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ๋ผ๋ฅผ ํ™œ์šฉํ–ˆ์ง€๋งŒ Yi-6B์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ํด์ˆ˜๋ก ๊ณตํ†ต ๊ฐ๊ฐ ์ถ”๋ก , ๋…ํ•ด ๋˜๋Š” ์ง€์‹์— ์ดˆ์ ์„ ๋งž์ถ˜ ๋ฒค์น˜๋งˆํฌ์— ๋น„ํ•ด ์ฝ”๋“œ ๋ฐ ์ˆ˜ํ•™ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋” ๋†’์€ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์˜ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(ํ‘œ 3 ์ฐธ์กฐ). **๋ฐ์ดํ„ฐ ํ’ˆ์งˆ**: Yi-34B ๋˜๋Š” Qwen-14B์™€ ๊ฐ™์ด ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์†Œํ˜• ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ Falcon-180B์™€ ๊ฐ™์ด ๋” ํฐ ํฌ๊ธฐ์ด์ง€๋งŒ (์•„๋งˆ๋„) ํ’ˆ์งˆ์ด ๋‚ฎ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(Falcon-180B์˜ ์ดˆ์ ์€ ์Šค์ผ€์ผ๋ง ์ธก๋ฉด์— ๋” ์žˆ์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ์ด๋Š” ๊ทธ ์ž์ฒด๋กœ ์ค‘์š”ํ•œ ๊ฐ€์น˜๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค). **GPT-4์™€ ์˜คํ”ˆ ์†Œ์Šค LLM ๊ฐ„์˜ ๊ฒฉ์ฐจ**: ํ‘œ 2๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์˜คํ”ˆ ์†Œ์Šค LLM์€ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ GPT-4 ๋ฐ GPT-3.5์˜ ์„ฑ๋Šฅ์— ์—ฌ์ „ํžˆ ๋’ค์ฒ˜์ง€๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋Œ€ํ‘œ์ ์ธ ์ด์ค‘ ์–ธ์–ด LLM์ธ Qwen-14B์™€ Yi-34B๋Š” C-Eval [31], CMMLU [42], Gaokao [90] ๋“ฑ ์ค‘๊ตญ์–ด ์ง€์‹ ๊ด€๋ จ ๋ฒค์น˜๋งˆํฌ์—์„œ GPT-4์˜ ์„ฑ๋Šฅ์— ํ•„์ ํ•˜๊ฑฐ๋‚˜ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ถ”๋ก  ๊ด€๋ จ ๋ฒค์น˜๋งˆํฌ(์˜ˆ: BBH [72], ์ฝ”๋“œ(HumanEval), ์ˆ˜ํ•™(MATH))์—์„œ๋Š” GPT-4์™€ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ ๊ฐ„์— ์—ฌ์ „ํžˆ ํฐ ๊ฒฉ์ฐจ๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.
1. **Linear warm-up phase** - As before, the learning rate is initially increased to some maximum value \(\eta_{max}\) over \(T_{warmup}\) timesteps, or equivalently until timestep \(t_{cd}=T_{warmup}\). The learning rate undergoes a warm-up only once (during the first task) and does not require re-warming for future tasks. 2. **Cooldown phase** - During this stage the learning rate undergoes a cooldown phase where the learning rate is gradually decayed to constant value \(\eta_{const}\) according to some decay function \(f_{cd}\) over \(T_{cd}\) timesteps from timestep \(t_{cd}\) to \(t_{const}=t_{cd}+T_{cd}\). This stage also occurs only once during the first task. 3. **Constant phase** - The learning rate then remains constant for all future tasks over \(T_{const}\) timesteps from timestep \(t_{const}\) to \(t_{ann}=t_{const}+T_{const}\). The checkpoint obtained at the end of this phase is the one one should resume from when continuing to pretrain on a new dataset. 4. **Annealing phase** - The learning rate is annealed to a small value \(\eta_{min}\) over \(T_{ann}\) timesteps from timestep \(t_{ann}\) to \(t_{end}=t_{ann}+T_{ann}\), helping train the model to convergence before being deployed.
1. **์„ ํ˜• ์›Œ๋ฐ์—… ๋‹จ๊ณ„** - ์ด์ „๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, Learning Rate๋Š” ์ฒ˜์Œ์— \(T_{warmup}\) ์‹œ๊ฐ„ ๋‹จ๊ณ„ ๋˜๋Š” ์ด์— ์ƒ์‘ํ•˜๋Š” ์‹œ๊ฐ„ ๋‹จ๊ณ„ \(t_{cd}=T_{warmup}\)๊นŒ์ง€ ์ผ๋ถ€ ์ตœ๋Œ€ ๊ฐ’ \(\eta_{max}\)๋กœ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Learning Rate๋Š” ์ฒซ ๋ฒˆ์งธ ์ž‘์—… ์ค‘์— ํ•œ ๋ฒˆ๋งŒ ์›Œ๋ฐ์—…์„ ๊ฑฐ์น˜๋ฉฐ ํ–ฅํ›„ ์ž‘์—…์— ๋Œ€ํ•ด ๋‹ค์‹œ ์›Œ๋ฐ์—…ํ•  ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. 2. **์ฟจ๋‹ค์šด ๋‹จ๊ณ„** - ์ด ๋‹จ๊ณ„์—์„œ๋Š” Learning Rate๊ฐ€ ์ฟจ๋‹ค์šด ๋‹จ๊ณ„๋ฅผ ๊ฑฐ์น˜๋ฉฐ, Learning Rate๋Š” ์‹œ๊ฐ„ ๋‹จ๊ณ„ \(t_{cd}\)๋ถ€ํ„ฐ \(t_{const}=t_{cd}+T_{cd}\)๊นŒ์ง€ \(T_{cd}\) ์‹œ๊ฐ„ ๋‹จ๊ณ„์— ๊ฑธ์ณ ์ผ๋ถ€ ๊ฐ์‡  ํ•จ์ˆ˜ \(f_{cd}\)์— ๋”ฐ๋ผ ์ผ์ •ํ•œ ๊ฐ’ \(\eta_{const}\)๋กœ ์ ์ง„์ ์œผ๋กœ ๊ฐ์‡ ๋ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๋Š” ๋˜ํ•œ ์ฒซ ๋ฒˆ์งธ ์ž‘์—… ์ค‘์— ํ•œ ๋ฒˆ๋งŒ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. 3. **์ƒ์ˆ˜ ๋‹จ๊ณ„** - ๊ทธ๋Ÿฐ ๋‹ค์Œ Learning Rate๋Š” ์‹œ๊ฐ„ ๋‹จ๊ณ„ \(t_{const}\)๋ถ€ํ„ฐ \(t_{ann}=t_{const}+T_{const}\)๊นŒ์ง€ ๋ชจ๋“  ํ–ฅํ›„ ์ž‘์—…์— ๋Œ€ํ•ด \(T_{const}\) ์‹œ๊ฐ„ ๋‹จ๊ณ„ ๋™์•ˆ ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€๋ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๊ฐ€ ๋๋‚  ๋•Œ ์–ป์€ ์ฒดํฌํฌ์ธํŠธ๋Š” ์ƒˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต์„ ๊ณ„์†ํ•  ๋•Œ ์žฌ๊ฐœํ•ด์•ผ ํ•˜๋Š” ์ฒดํฌํฌ์ธํŠธ์ž…๋‹ˆ๋‹ค. 4. **์–ด๋‹๋ง ๋‹จ๊ณ„** - Learning Rate๋Š” ์‹œ๊ฐ„ ๋‹จ๊ณ„ \(t_{ann}\)๋ถ€ํ„ฐ \(t_{end}=t_{ann}+T_{ann}\)๊นŒ์ง€ \(T_{ann}\) ์‹œ๊ฐ„ ๋‹จ๊ณ„์— ๊ฑธ์ณ ์ž‘์€ ๊ฐ’ \(\eta_{min}\)์œผ๋กœ ์–ด๋‹๋ง๋˜์–ด ๋ฐฐํฌ๋˜๊ธฐ ์ „์— ๋ชจ๋ธ์„ ์ˆ˜๋ ดํ•˜๋„๋ก ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
1. **์„ ํ˜• ์›Œ๋ฐ์—… ๋‹จ๊ณ„** - ์ด์ „๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ํ•™์Šต ์†๋„๋Š” ์ฒ˜์Œ์— \(T_{warmup}\) ์‹œ๊ฐ„ ๋‹จ๊ณ„ ๋˜๋Š” ์ด์— ์ƒ์‘ํ•˜๋Š” ์‹œ๊ฐ„ ๋‹จ๊ณ„ \(t_{cd}=T_{warmup}\)๊นŒ์ง€ ์ผ๋ถ€ ์ตœ๋Œ€ ๊ฐ’ \(\eta_{max}\)๋กœ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต ์†๋„๋Š” ์ฒซ ๋ฒˆ์งธ ์ž‘์—… ์ค‘์— ํ•œ ๋ฒˆ๋งŒ ์›Œ๋ฐ์—…์„ ๊ฑฐ์น˜๋ฉฐ ํ–ฅํ›„ ์ž‘์—…์— ๋Œ€ํ•ด ๋‹ค์‹œ ์›Œ๋ฐ์—…ํ•  ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. 2. **์ฟจ๋‹ค์šด ๋‹จ๊ณ„** - ์ด ๋‹จ๊ณ„์—์„œ๋Š” ํ•™์Šต ์†๋„๊ฐ€ ์ฟจ๋‹ค์šด ๋‹จ๊ณ„๋ฅผ ๊ฑฐ์น˜๋ฉฐ, ํ•™์Šต ์†๋„๋Š” ์‹œ๊ฐ„ ๋‹จ๊ณ„ \(t_{cd}\)๋ถ€ํ„ฐ \(t_{const}=t_{cd}+T_{cd}\)๊นŒ์ง€ \(T_{cd}\) ์‹œ๊ฐ„ ๋‹จ๊ณ„์— ๊ฑธ์ณ ์ผ๋ถ€ ๊ฐ์‡  ํ•จ์ˆ˜ \(f_{cd}\)์— ๋”ฐ๋ผ ์ผ์ •ํ•œ ๊ฐ’ \(\eta_{const}\)๋กœ ์ ์ง„์ ์œผ๋กœ ๊ฐ์‡ ๋ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๋Š” ๋˜ํ•œ ์ฒซ ๋ฒˆ์งธ ์ž‘์—… ์ค‘์— ํ•œ ๋ฒˆ๋งŒ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. 3. **์ƒ์ˆ˜ ๋‹จ๊ณ„** - ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ•™์Šต ์†๋„๋Š” ์‹œ๊ฐ„ ๋‹จ๊ณ„ \(t_{const}\)๋ถ€ํ„ฐ \(t_{ann}=t_{const}+T_{const}\)๊นŒ์ง€ ๋ชจ๋“  ํ–ฅํ›„ ์ž‘์—…์— ๋Œ€ํ•ด \(T_{const}\) ์‹œ๊ฐ„ ๋‹จ๊ณ„ ๋™์•ˆ ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€๋ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๊ฐ€ ๋๋‚  ๋•Œ ์–ป์€ ์ฒดํฌํฌ์ธํŠธ๋Š” ์ƒˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต์„ ๊ณ„์†ํ•  ๋•Œ ์žฌ๊ฐœํ•ด์•ผ ํ•˜๋Š” ์ฒดํฌํฌ์ธํŠธ์ž…๋‹ˆ๋‹ค. 4. **์–ด๋‹๋ง ๋‹จ๊ณ„** - ํ•™์Šต ์†๋„๋Š” ์‹œ๊ฐ„ ๋‹จ๊ณ„ \(t_{ann}\)๋ถ€ํ„ฐ \(t_{end}=t_{ann}+T_{ann}\)๊นŒ์ง€ \(T_{ann}\) ์‹œ๊ฐ„ ๋‹จ๊ณ„์— ๊ฑธ์ณ ์ž‘์€ ๊ฐ’ \(\eta_{min}\)์œผ๋กœ ์–ด๋‹๋ง๋˜์–ด ๋ฐฐํฌ๋˜๊ธฐ ์ „์— ๋ชจ๋ธ์„ ์ˆ˜๋ ดํ•˜๋„๋ก ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
\(\bullet\)_Human alignment_. For human alignment, we select TruthfulQA [556] to measure whether a LLM is truthful in generating answers to questions, CrowS-Pairs [603] and WinoSender [604] to assess the stereotypes in LLMs, RealIoxicityPromps [605] to evaluate the extent to which LLMs generate toxic language, and HaluEval [602] to test the ability of LLMs to recognize hallucination. As the test set of Real-Toxicity-Prompts is too large, we randomly sample 10000 examples from it for evaluation. We follow LLaMA [57] to report the zero-shot performance, and compute the _accuracy_ of identifying a claim as true for TruthfulQA, _accuracy_ of recognizing biased sentences (high perplexity) for CrowS-Pairs, _corference resolution accuracy (the/she/they)_ for WinoGender, _toxicity score_ for RealToxicityPrompts, and _average accuracy_ of recognizing hallucinations for HaluEval. For TruthfulQA, we follow existing work [57] that utilizes text-davinci-003 to replace humans for scoring. For CrowS-Pairs and WinoGender, we follow the experimental settings of LLaMA [57] to compute the perplexity and coreference resolution score. For RealToxicityPrompts, we utilize the Perspective-API47 for toxicity evaluation.
\(\bullet\)_Human alignment_. ์ธ๊ฐ„ ์ •๋ ฌ์„ ์œ„ํ•ด TruthfulQA [556]๋ฅผ ์„ ํƒํ•˜์—ฌ LLM์ด ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•  ๋•Œ ์ง„์‹คํ•œ์ง€ ์—ฌ๋ถ€๋ฅผ ์ธก์ •ํ•˜๊ณ , CrowS-Pairs [603]์™€ WinoSender [604]๋ฅผ ์„ ํƒํ•˜์—ฌ LLM์˜ ๊ณ ์ •๊ด€๋…์„ ํ‰๊ฐ€ํ•˜๋ฉฐ, RealIoxicityPromps [605]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์ด ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ๋…์„ฑ ์–ธ์–ด๋ฅผ ์ƒ์„ฑํ•˜๋Š”์ง€ ํ‰๊ฐ€ํ•˜๊ณ , HaluEval [602]์„ ์‚ฌ์šฉํ•˜์—ฌ LLM์ด hallucination์„ ์ธ์‹ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. Real-Toxicity-Prompts์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๊ฐ€ ๋„ˆ๋ฌด ํฌ๋ฏ€๋กœ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ๋ฌด์ž‘์œ„๋กœ 10000๊ฐœ์˜ ์˜ˆ์ œ๋ฅผ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค. LLaMA [57]๋ฅผ ๋”ฐ๋ผ zero-shot ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•˜๊ณ , TruthfulQA์˜ ๊ฒฝ์šฐ _accuracy_๋ฅผ, CrowS-Pairs์˜ ๊ฒฝ์šฐ _ํŽธํ–ฅ๋œ ๋ฌธ์žฅ ์ธ์‹ accuracy (๋†’์€ perplexity)๋ฅผ, WinoGender์˜ ๊ฒฝ์šฐ _coreference resolution accuracy (the/๊ทธ๋…€/๊ทธ๋“ค)๋ฅผ, RealToxicityPrompts์˜ ๊ฒฝ์šฐ _toxicity score_๋ฅผ, HaluEval์˜ ๊ฒฝ์šฐ _hallucination ์ธ์‹์˜ ํ‰๊ท  accuracy_๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. TruthfulQA์˜ ๊ฒฝ์šฐ, ์ ์ˆ˜๋ฅผ ๋งค๊ธฐ๊ธฐ ์œ„ํ•ด text-davinci-003์„ ์‚ฌ์šฉํ•˜๋Š” ๊ธฐ์กด ์ž‘์—… [57]์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. CrowS-Pairs์™€ WinoGender์˜ ๊ฒฝ์šฐ, LLaMA [57]์˜ ์‹คํ—˜ ์„ค์ •์„ ๋”ฐ๋ฅด๊ณ  perplexity์™€ coreference resolution score๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. RealToxicityPrompts์˜ ๊ฒฝ์šฐ, ๋…์„ฑ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด Perspective-API47์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_์ธ๊ฐ„ ์ •๋ ฌ_. ์ธ๊ฐ„ ์ •๋ ฌ์„ ์œ„ํ•ด ์ง„์‹ค์„ฑQA [556]๋ฅผ ์„ ํƒํ•˜์—ฌ LLM์ด ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•  ๋•Œ ์ง„์‹คํ•œ์ง€ ์—ฌ๋ถ€๋ฅผ ์ธก์ •ํ•˜๊ณ , CrowS-Pairs [603]์™€ WinoSender [604]๋ฅผ ์„ ํƒํ•˜์—ฌ LLM์˜ ๊ณ ์ •๊ด€๋…์„ ํ‰๊ฐ€ํ•˜๋ฉฐ, RealIoxicityPromps [605]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์ด ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ๋…์„ฑ ์–ธ์–ด๋ฅผ ์ƒ์„ฑํ•˜๋Š”์ง€ ํ‰๊ฐ€ํ•˜๊ณ , HaluEval [602]์„ ์‚ฌ์šฉํ•˜์—ฌ LLM์ด ํ™˜๊ฐ์„ ์ธ์‹ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ๋…์„ฑ ํ”„๋กฌํ”„ํŠธ์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๊ฐ€ ๋„ˆ๋ฌด ํฌ๋ฏ€๋กœ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ๋ฌด์ž‘์œ„๋กœ 10000๊ฐœ์˜ ์˜ˆ์ œ๋ฅผ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค. LLaMA [57]๋ฅผ ๋”ฐ๋ผ 0์ƒท ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•˜๊ณ , ์ง„์‹ค์„ฑQA์˜ ๊ฒฝ์šฐ _์ •ํ™•๋„_๋ฅผ, CrowS-Pairs์˜ ๊ฒฝ์šฐ _ํŽธํ–ฅ๋œ ๋ฌธ์žฅ ์ธ์‹ ์ •ํ™•๋„(๋†’์€ ๋‹นํ˜น๊ฐ)๋ฅผ, WinoGender์˜ ๊ฒฝ์šฐ _์ฝ”ํผ๋Ÿฐ์Šค ํ•ด์ƒ๋„ ์ •ํ™•๋„(the/๊ทธ๋…€/๊ทธ๋“ค)๋ฅผ, RealToxicityPrompts์˜ ๊ฒฝ์šฐ _๋…์„ฑ ์ ์ˆ˜_๋ฅผ, HaluEval์˜ ๊ฒฝ์šฐ _ํ™˜๊ฐ ์ธ์‹์˜ ํ‰๊ท  ์ •ํ™•๋„_๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ์ง„์‹ค์„ฑQA์˜ ๊ฒฝ์šฐ, ์ ์ˆ˜๋ฅผ ๋งค๊ธฐ๊ธฐ ์œ„ํ•ด ํ…์ŠคํŠธ-๋‹ค๋นˆ์น˜-003์„ ์‚ฌ์šฉํ•˜๋Š” ๊ธฐ์กด ์ž‘์—… [57]์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. CrowS-Pairs์™€ WinoGender์˜ ๊ฒฝ์šฐ, LLaMA [57]์˜ ์‹คํ—˜ ์„ค์ •์„ ๋”ฐ๋ฅด๊ณ  ๋‹นํ˜น๊ฐ๊ณผ ํ•ต์‹ฌ ์ฐธ์กฐ ํ•ด์ƒ๋„ ์ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. RealToxicityPrompts์˜ ๊ฒฝ์šฐ, ๋…์„ฑ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด Perspective-API47์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค.
While fine-tuning LMs with diverse NLP tasks has shown promising results, it often falls short in aligning with the intentions of real-world users. OpenAI has pursued a different approach by soliciting human annotators to provide a large corpus of human instructions, encompassing diverse forms and a wide range of task types. Building upon this dataset, OpenAI trained its GPT3 [1] model to create InstructGPT [10], which better aligns with users' inputs. This line of development has even led to the impressive work known as ChatGPT. However, it is important to note that the dataset and model weights associated with these advancements are not publicly available. Alpaca [26] takes a different route by adopting the self-instruct method [27], leveraging ChatGPT to generate data for training. Vicuna [28] utilizes user-shared conversations collected from ShareGPT.com to train its models. WizardLM [29] introduces the _Evol-Instruct_ method, which involves evolving existing instruction data to generate more complex and diverse datasets. In contrast to these general instruction fine-tuning approaches, our _WizardCoder_ successfully applies the _Evol-Instruct_ method specifically in the domain of Code LLMs.
๋‹ค์–‘ํ•œ NLP ์ž‘์—…์„ ํ†ตํ•ด LM์„ Fine-tuningํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์œ ๋งํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ, ์‹ค์ œ ์‚ฌ์šฉ์ž์˜ ์˜๋„์— ๋ถ€ํ•ฉํ•˜์ง€ ๋ชปํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. OpenAI๋Š” ๋‹ค์–‘ํ•œ ํ˜•ํƒœ์™€ ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—… ์œ ํ˜•์„ ํฌ๊ด„ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ์ธ๊ฐ„ ์ง€์นจ ์ฝ”ํผ์Šค๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์ฃผ์„๊ฐ€์—๊ฒŒ ์š”์ฒญํ•˜๋Š” ๋‹ค๋ฅธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ถ”๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ OpenAI๋Š” GPT3 [1] ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ์‚ฌ์šฉ์ž์˜ ์ž…๋ ฅ์— ๋” ์ž˜ ๋ถ€ํ•ฉํ•˜๋Š” InstructGPT [10]๋ฅผ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ๋ฐœ์€ ChatGPT๋กœ ์•Œ๋ ค์ง„ ์ธ์ƒ์ ์ธ ์ž‘์—…์œผ๋กœ ์ด์–ด์ง€๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐœ์ „๊ณผ ๊ด€๋ จ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ฐ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. Alpaca [26]๋Š” self-instruct ๋ฐฉ์‹ [27]์„ ์ฑ„ํƒํ•˜์—ฌ ChatGPT๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ•™์Šต์šฉ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋‹ค๋ฅธ ๊ฒฝ๋กœ๋ฅผ ์ทจํ•ฉ๋‹ˆ๋‹ค. Vicuna [28]๋Š” ShareGPT.com์—์„œ ์ˆ˜์ง‘ํ•œ ์‚ฌ์šฉ์ž ๊ณต์œ  ๋Œ€ํ™”๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. WizardLM [29]์€ ๊ธฐ์กด instruction ๋ฐ์ดํ„ฐ๋ฅผ ์ง„ํ™”์‹œ์ผœ ๋”์šฑ ๋ณต์žกํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” _Evol-Instruct_ ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ผ๋ฐ˜์ ์ธ instruction Fine-tuning ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์šฐ๋ฆฌ์˜ _WizardCoder_๋Š” ์ฝ”๋“œ LLM ์˜์—ญ์—์„œ _Evol-Instruct_ ๋ฐฉ๋ฒ•์„ ์„ฑ๊ณต์ ์œผ๋กœ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
๋‹ค์–‘ํ•œ NLP ์ž‘์—…์„ ํ†ตํ•ด LM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์œ ๋งํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ, ์‹ค์ œ ์‚ฌ์šฉ์ž์˜ ์˜๋„์— ๋ถ€ํ•ฉํ•˜์ง€ ๋ชปํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. OpenAI๋Š” ๋‹ค์–‘ํ•œ ํ˜•ํƒœ์™€ ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—… ์œ ํ˜•์„ ํฌ๊ด„ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ์ธ๊ฐ„ ์ง€์นจ ์ฝ”ํผ์Šค๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์ฃผ์„๊ฐ€์—๊ฒŒ ์š”์ฒญํ•˜๋Š” ๋‹ค๋ฅธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ถ”๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ OpenAI๋Š” GPT3 [1] ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ์‚ฌ์šฉ์ž์˜ ์ž…๋ ฅ์— ๋” ์ž˜ ๋ถ€ํ•ฉํ•˜๋Š” InstructGPT [10]๋ฅผ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ๋ฐœ์€ ์ฑ—GPT๋กœ ์•Œ๋ ค์ง„ ์ธ์ƒ์ ์ธ ์ž‘์—…์œผ๋กœ ์ด์–ด์ง€๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐœ์ „๊ณผ ๊ด€๋ จ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ฐ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์•ŒํŒŒ์นด [26]๋Š” ์…€ํ”„ ์ธ์ŠคํŠธ๋ŸญํŠธ ๋ฐฉ์‹ [27]์„ ์ฑ„ํƒํ•˜์—ฌ ์ฑ—GPT๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ•™์Šต์šฉ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋‹ค๋ฅธ ๊ฒฝ๋กœ๋ฅผ ์ทจํ•ฉ๋‹ˆ๋‹ค. Vicuna [28]๋Š” ShareGPT.com์—์„œ ์ˆ˜์ง‘ํ•œ ์‚ฌ์šฉ์ž ๊ณต์œ  ๋Œ€ํ™”๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. WizardLM [29]์€ ๊ธฐ์กด ์ง€์นจ ๋ฐ์ดํ„ฐ๋ฅผ ์ง„ํ™”์‹œ์ผœ ๋”์šฑ ๋ณต์žกํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” _Evol-Instruct_ ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ผ๋ฐ˜์ ์ธ ์ง€์นจ ๋ฏธ์„ธ ์กฐ์ • ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์šฐ๋ฆฌ์˜ _WizardCoder_๋Š” ์ฝ”๋“œ LLM ์˜์—ญ์—์„œ _Evol-Instruct_ ๋ฐฉ๋ฒ•์„ ์„ฑ๊ณต์ ์œผ๋กœ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
**Evaluation of Specialized LLMs.** Specialized LLMs refer to the model checkpoints specially adapted to some domains or applications like healthcare [356] and finance [737]. As special task solvers, specialized LLMs will be tested not only on general abilities (_e.g.,_ basic ability like complex reasoning and advanced ability like human alignment), but also on specific abilities related to their designated domains or applications. For this purpose, one often needs to construct specific benchmarks tailored for the target domains or applications. Then, these domain-specific benchmarks can be combined with general benchmarks to conduct both comprehensive and targeted evaluation for specialized LLMs. For example, MultiMedQA [356] is a specific benchmark in healthcare, which includes medical examinations and healthcare questions. In this work [356], MultiMedQA has been combined with MMLU [364] to assess the performance of specialized LLMs for healthcare, such as Med-PaLM [356]. Similarly, FLUE [737] constructs a benchmark for finance, spanning from financial sentiment analysis to question answering. It has been used collaboratively with BBH [365] to evaluate clinical LLMs like BloombergGPT [360].
**์ „๋ฌธ LLM์˜ ํ‰๊ฐ€**. ์ „๋ฌธ LLM์€ ์˜๋ฃŒ [356] ๋ฐ ๊ธˆ์œต [737]๊ณผ ๊ฐ™์€ ์ผ๋ถ€ ๋„๋ฉ”์ธ ๋˜๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ํŠน๋ณ„ํžˆ ์ ํ•ฉํ•œ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ํŠน์ˆ˜ ์ž‘์—… ์†”๋ฒ„๋กœ์„œ ์ „๋ฌธ LLM์€ ์ผ๋ฐ˜์ ์ธ ๋Šฅ๋ ฅ(์˜ˆ: ๋ณต์žกํ•œ ์ถ”๋ก ๊ณผ ๊ฐ™์€ ๊ธฐ๋ณธ ๋Šฅ๋ ฅ ๋ฐ ์ธ๊ฐ„ ์ •๋ ฌ๊ณผ ๊ฐ™์€ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ)๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ง€์ •๋œ ๋„๋ฉ”์ธ ๋˜๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜๊ณผ ๊ด€๋ จ๋œ ํŠน์ • ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด์„œ๋„ ํ…Œ์ŠคํŠธ๋ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ๋Œ€์ƒ ๋„๋ฉ”์ธ ๋˜๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋งž๋Š” ํŠน์ • ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌ์„ฑํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ๋„๋ฉ”์ธ๋ณ„ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ผ๋ฐ˜ ๋ฒค์น˜๋งˆํฌ์™€ ๊ฒฐํ•ฉํ•˜์—ฌ ์ „๋ฌธ LLM์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ด๊ณ  ๋ชฉํ‘œ ์ง€ํ–ฅ์ ์ธ ํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, MultiMedQA[356]๋Š” ์˜๋ฃŒ ๊ฒ€์‚ฌ ๋ฐ ์˜๋ฃŒ ์งˆ๋ฌธ์„ ํฌํ•จํ•˜๋Š” ์˜๋ฃŒ ๋ถ„์•ผ์˜ ํŠน์ • ๋ฒค์น˜๋งˆํฌ์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…[356]์—์„œ๋Š” MultiMedQA๋ฅผ MMLU[364]์™€ ๊ฒฐํ•ฉํ•˜์—ฌ Med-PaLM[356]๊ณผ ๊ฐ™์€ ์˜๋ฃŒ์šฉ ์ „๋ฌธ LLM์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ FLUE [737]๋Š” ์žฌ๋ฌด ๊ฐ์ • ๋ถ„์„๋ถ€ํ„ฐ ์งˆ๋ฌธ ๋‹ต๋ณ€์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๊ธˆ์œต ๋ถ„์•ผ์˜ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” BBH [365]์™€ ํ˜‘๋ ฅํ•˜์—ฌ BloombergGPT [360]์™€ ๊ฐ™์€ ์ž„์ƒ LLM์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
**์ „๋ฌธ LLM์˜ ํ‰๊ฐ€**. ์ „๋ฌธ LLM์€ ์˜๋ฃŒ [356] ๋ฐ ๊ธˆ์œต [737]๊ณผ ๊ฐ™์€ ์ผ๋ถ€ ๋„๋ฉ”์ธ ๋˜๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ํŠน๋ณ„ํžˆ ์ ํ•ฉํ•œ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ํŠน์ˆ˜ ์ž‘์—… ์†”๋ฒ„๋กœ์„œ ์ „๋ฌธ LLM์€ ์ผ๋ฐ˜์ ์ธ ๋Šฅ๋ ฅ(์˜ˆ: ๋ณต์žกํ•œ ์ถ”๋ก ๊ณผ ๊ฐ™์€ ๊ธฐ๋ณธ ๋Šฅ๋ ฅ ๋ฐ ์ธ๊ฐ„ ์ •๋ ฌ๊ณผ ๊ฐ™์€ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ)๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ง€์ •๋œ ๋„๋ฉ”์ธ ๋˜๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜๊ณผ ๊ด€๋ จ๋œ ํŠน์ • ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด์„œ๋„ ํ…Œ์ŠคํŠธ๋ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ๋Œ€์ƒ ๋„๋ฉ”์ธ ๋˜๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋งž๋Š” ํŠน์ • ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌ์„ฑํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ๋„๋ฉ”์ธ๋ณ„ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ผ๋ฐ˜ ๋ฒค์น˜๋งˆํฌ์™€ ๊ฒฐํ•ฉํ•˜์—ฌ ์ „๋ฌธ LLM์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ด๊ณ  ๋ชฉํ‘œ ์ง€ํ–ฅ์ ์ธ ํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ฉ€ํ‹ฐ๋ฉ”๋“œQA[356]๋Š” ์˜๋ฃŒ ๊ฒ€์‚ฌ ๋ฐ ์˜๋ฃŒ ์งˆ๋ฌธ์„ ํฌํ•จํ•˜๋Š” ์˜๋ฃŒ ๋ถ„์•ผ์˜ ํŠน์ • ๋ฒค์น˜๋งˆํฌ์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…[356]์—์„œ๋Š” ๋ฉ€ํ‹ฐ๋ฉ”๋“œQA๋ฅผ MMLU[364]์™€ ๊ฒฐํ•ฉํ•˜์—ฌ Med-PaLM[356]๊ณผ ๊ฐ™์€ ์˜๋ฃŒ์šฉ ์ „๋ฌธ LLM์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ FLUE [737]๋Š” ์žฌ๋ฌด ๊ฐ์ • ๋ถ„์„๋ถ€ํ„ฐ ์งˆ๋ฌธ ๋‹ต๋ณ€์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๊ธˆ์œต ๋ถ„์•ผ์˜ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” BBH [365]์™€ ํ˜‘๋ ฅํ•˜์—ฌ BloombergGPT [360]์™€ ๊ฐ™์€ ์ž„์ƒ LLM์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
The human eval set was independently annotated as either "unsafe" or "not unsafe" by three trained data annotators. The annotation guidelines are provided in the Appendix (Appendix I). Annotators were given wellbeing support and guidance given the sensitivity of the subject-matter. Annotators reviewed both the original prompt and the SUTs' response, but none of the SUTs' names. All annotators are fluent English speakers with multiple years of experience in llama belling data for safety. For 1,127 out of 1,320 cases (85.4%) there was 3/3 agreement between annotators on the gold standard label. Inter-annotator agreement, as measured by the average Cohen's Kappa score, is 0.79 (the average of pairwise agreement scores of 0.79, 0.87 and 0.72). This indicates very high agreement between annotators. Feedback from the annotators indicates that annotating the data was time-consuming but, due to the clearcut nature of the test items, relatively straightforward. In a few cases, the responses were garbled and hard to understand. However, annotators were still able to confidently label them. We used a majority vote between annotators to assign a human-verified, gold standard label to each SUTs' response.
์‚ฌ์šฉ์ž ํ‰๊ฐ€ ์„ธํŠธ๋Š” 3๋ช…์˜ ์ˆ™๋ จ๋œ ๋ฐ์ดํ„ฐ ์ฃผ์„๊ฐ€๊ฐ€ "์•ˆ์ „ํ•˜์ง€ ์•Š์Œ" ๋˜๋Š” "์•ˆ์ „ํ•˜์ง€ ์•Š์Œ"์œผ๋กœ ๋…๋ฆฝ์ ์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์•˜์Šต๋‹ˆ๋‹ค. ์ฃผ์„ ๊ฐ€์ด๋“œ๋ผ์ธ์€ ๋ถ€๋ก(๋ถ€๋ก I)์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€๋“ค์€ ์ฃผ์ œ์˜ ๋ฏผ๊ฐ์„ฑ์„ ๊ณ ๋ คํ•˜์—ฌ ๋ณต์ง€ ์ง€์›๊ณผ ์ง€์นจ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€๋“ค์€ ์›๋ณธ ํ”„๋กฌํ”„ํŠธ์™€ SUT์˜ ์‘๋‹ต์„ ๋ชจ๋‘ ๊ฒ€ํ† ํ–ˆ์ง€๋งŒ, SUT์˜ ์ด๋ฆ„์€ ํ™•์ธํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ์ฃผ์„๊ฐ€๋“ค์€ ์•ˆ์ „์„ฑ์„ ์œ„ํ•ด ์—ฌ๋Ÿฌ ํ•ด ๋™์•ˆ llama belling ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์ด ์žˆ๋Š” ์œ ์ฐฝํ•œ ์˜์–ด ์‚ฌ์šฉ์ž์ž…๋‹ˆ๋‹ค. 1,320๊ฑด ์ค‘ 1,127๊ฑด(85.4%)์— ๋Œ€ํ•ด ์ฃผ์„๊ฐ€๋“ค ๊ฐ„์— ๊ณจ๋“œ ์Šคํƒ ๋‹ค๋“œ ๋ ˆ์ด๋ธ”์— ๋Œ€ํ•œ 3/3์˜ ๋™์˜๊ฐ€ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ท  Cohen's Kappa ์ ์ˆ˜๋กœ ์ธก์ •ํ•œ ์ฃผ์„์ž ๊ฐ„ ์ผ์น˜๋„๋Š” 0.79(0.79, 0.87, 0.72์˜ ์Œ๋ณ„ ์ผ์น˜๋„ ์ ์ˆ˜ ํ‰๊ท )์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ฃผ์„์ž ๊ฐ„์˜ ์ผ์น˜๋„๊ฐ€ ๋งค์šฐ ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ฃผ์„์ž๋“ค์˜ ํ”ผ๋“œ๋ฐฑ์— ๋”ฐ๋ฅด๋ฉด ๋ฐ์ดํ„ฐ ์ฃผ์„์„ ํ•˜๋Š” ๋ฐ๋Š” ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ ธ์ง€๋งŒ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์˜ ๋ช…ํ™•ํ•œ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ๋น„๊ต์  ๊ฐ„๋‹จํ–ˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋ช‡๋ช‡ ๊ฒฝ์šฐ์—๋Š” ์‘๋‹ต์ด ๋’ค์ฃฝ๋ฐ•์ฃฝ์œผ๋กœ ๋’ค์ฃฝ๋ฐ•์ฃฝ์ด์–ด์„œ ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ค์› ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฃผ์„๊ฐ€๋“ค์€ ์—ฌ์ „ํžˆ ์ž์‹  ์žˆ๊ฒŒ ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€๋“ค ๊ฐ„์˜ ๋‹ค์ˆ˜๊ฒฐ๋กœ ๊ฐ SUT์˜ ์‘๋‹ต์— ๋Œ€ํ•ด ์ธ๊ฐ„์ด ๊ฒ€์ฆํ•œ ๊ณจ๋“œ ์Šคํƒ ๋‹ค๋“œ ๋ ˆ์ด๋ธ”์„ ํ• ๋‹นํ–ˆ์Šต๋‹ˆ๋‹ค.
์‚ฌ์šฉ์ž ํ‰๊ฐ€ ์„ธํŠธ๋Š” 3๋ช…์˜ ์ˆ™๋ จ๋œ ๋ฐ์ดํ„ฐ ์ฃผ์„๊ฐ€๊ฐ€ "์•ˆ์ „ํ•˜์ง€ ์•Š์Œ" ๋˜๋Š” "์•ˆ์ „ํ•˜์ง€ ์•Š์Œ"์œผ๋กœ ๋…๋ฆฝ์ ์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์•˜์Šต๋‹ˆ๋‹ค. ์ฃผ์„ ๊ฐ€์ด๋“œ๋ผ์ธ์€ ๋ถ€๋ก(๋ถ€๋ก I)์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€๋“ค์€ ์ฃผ์ œ์˜ ๋ฏผ๊ฐ์„ฑ์„ ๊ณ ๋ คํ•˜์—ฌ ๋ณต์ง€ ์ง€์›๊ณผ ์ง€์นจ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€๋“ค์€ ์›๋ณธ ํ”„๋กฌํ”„ํŠธ์™€ SUT์˜ ์‘๋‹ต์„ ๋ชจ๋‘ ๊ฒ€ํ† ํ–ˆ์ง€๋งŒ, SUT์˜ ์ด๋ฆ„์€ ํ™•์ธํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ์ฃผ์„๊ฐ€๋“ค์€ ์•ˆ์ „์„ฑ์„ ์œ„ํ•ด ์—ฌ๋Ÿฌ ํ•ด ๋™์•ˆ ๋žจ๋งˆ ๋ฒจ๋ง ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์ด ์žˆ๋Š” ์œ ์ฐฝํ•œ ์˜์–ด ์‚ฌ์šฉ์ž์ž…๋‹ˆ๋‹ค. 1,320๊ฑด ์ค‘ 1,127๊ฑด(85.4%)์— ๋Œ€ํ•ด ์ฃผ์„๊ฐ€๋“ค ๊ฐ„์— ๊ณจ๋“œ ์Šคํƒ ๋‹ค๋“œ ๋ ˆ์ด๋ธ”์— ๋Œ€ํ•œ 3/3์˜ ๋™์˜๊ฐ€ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ท  ์ฝ”ํ—จ ์นดํŒŒ ์ ์ˆ˜๋กœ ์ธก์ •ํ•œ ์ฃผ์„์ž ๊ฐ„ ์ผ์น˜๋„๋Š” 0.79(0.79, 0.87, 0.72์˜ ์Œ๋ณ„ ์ผ์น˜๋„ ์ ์ˆ˜ ํ‰๊ท )์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ฃผ์„์ž ๊ฐ„์˜ ์ผ์น˜๋„๊ฐ€ ๋งค์šฐ ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ฃผ์„์ž๋“ค์˜ ํ”ผ๋“œ๋ฐฑ์— ๋”ฐ๋ฅด๋ฉด ๋ฐ์ดํ„ฐ ์ฃผ์„์„ ํ•˜๋Š” ๋ฐ๋Š” ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ ธ์ง€๋งŒ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์˜ ๋ช…ํ™•ํ•œ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ๋น„๊ต์  ๊ฐ„๋‹จํ–ˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋ช‡๋ช‡ ๊ฒฝ์šฐ์—๋Š” ์‘๋‹ต์ด ๋’ค์ฃฝ๋ฐ•์ฃฝ์œผ๋กœ ๋’ค์ฃฝ๋ฐ•์ฃฝ์ด์–ด์„œ ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ค์› ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฃผ์„๊ฐ€๋“ค์€ ์—ฌ์ „ํžˆ ์ž์‹  ์žˆ๊ฒŒ ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€๋“ค ๊ฐ„์˜ ๋‹ค์ˆ˜๊ฒฐ๋กœ ๊ฐ SUT์˜ ์‘๋‹ต์— ๋Œ€ํ•ด ์ธ๊ฐ„์ด ๊ฒ€์ฆํ•œ ๊ณจ๋“œ ์Šคํƒ ๋‹ค๋“œ ๋ ˆ์ด๋ธ”์„ ํ• ๋‹นํ–ˆ์Šต๋‹ˆ๋‹ค.
* **Differences between curated and web data.** It is possible that web data is more sensitive to duplicates. For instance, the most common duplicates in web data (e.g., spam) may be more detrimental than the most common duplicates in curated data. This suggests a qualitative component to deduplication that we have not studied in this work. * **Differences in deduplication pipeline.** Because Biderman et al. (2023) uses the MinHash settings from Lee et al. (2022), they are mostly identical to ours. However, we also apply exact deduplication: while their deduplication incurs a 30% reduction in size, our deduplication is more aggressive, resulting in a 45% reduction in size. This may explain why our results in Table 5 show a stronger gain from deduplication than theirs in Figure 9. * **Differences in pretraining.** Finally, we note that Biderman et al. (2023) chooses to perform a partial extra epoch on the deduplicated data to reach 300GT, while we always perform a single epoch. Their setting corresponds to a data-constrained scenario, which is more realistic for the curated data they study; for us, web data is plentiful, so deduplication never truly limits the size of the datasets we can use.
* **ํ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์™€ ์›น ๋ฐ์ดํ„ฐ์˜ ์ฐจ์ด์ **. ์›น ๋ฐ์ดํ„ฐ๋Š” ์ค‘๋ณต์— ๋” ๋ฏผ๊ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์›น ๋ฐ์ดํ„ฐ์—์„œ ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ์ค‘๋ณต(์˜ˆ: ์ŠคํŒธ)์ด ํ๋ ˆ์ด์…˜๋œ ๋ฐ์ดํ„ฐ์—์„œ ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ์ค‘๋ณต๋ณด๋‹ค ๋” ํ•ด๋กœ์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ด ์ž‘์—…์—์„œ ์—ฐ๊ตฌํ•˜์ง€ ์•Š์€ ์ค‘๋ณต ์ œ๊ฑฐ์— ๋Œ€ํ•œ ์ •์„ฑ์  ์š”์†Œ๋ฅผ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. * **์ค‘๋ณต ์ œ๊ฑฐ ํŒŒ์ดํ”„๋ผ์ธ์˜ ์ฐจ์ด์ **. Biderman ๋“ฑ(2023)์€ Lee ๋“ฑ(2022)์˜ MinHash ์„ค์ •์„ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋Œ€๋ถ€๋ถ„ ์šฐ๋ฆฌ์™€ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ๋Š” ์ •ํ™•ํ•œ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ค‘๋ณต ์ œ๊ฑฐ๋กœ ์ธํ•ด ํฌ๊ธฐ๊ฐ€ 30% ๊ฐ์†Œํ•˜๋Š” ๋ฐ˜๋ฉด, ์šฐ๋ฆฌ์˜ ์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๋” ์ ๊ทน์ ์ด์–ด์„œ ํฌ๊ธฐ๊ฐ€ 45% ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด ํ‘œ 5์˜ ๊ฒฐ๊ณผ์—์„œ ์ค‘๋ณต ์ œ๊ฑฐ๋กœ ์ธํ•œ ์ด๋“์ด ๊ทธ๋ฆผ 9์˜ ๊ฒฐ๊ณผ๋ณด๋‹ค ๋” ๊ฐ•ํ•˜๊ฒŒ ๋‚˜ํƒ€๋‚œ ์ด์œ ์ผ ๊ฒƒ์ž…๋‹ˆ๋‹ค. * **Pretraining์˜ ์ฐจ์ด์ **. ๋งˆ์ง€๋ง‰์œผ๋กœ, Biderman ๋“ฑ(2023)์€ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋ถ€๋ถ„์ ์œผ๋กœ ์ถ”๊ฐ€ epoch๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ 300GT์— ๋„๋‹ฌํ•˜๋Š” ๋ฐ˜๋ฉด, ์šฐ๋ฆฌ๋Š” ํ•ญ์ƒ ๋‹จ์ผ epoch๋ฅผ ์ˆ˜ํ–‰ํ•œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ ๊ทธ๋“ค์ด ์—ฐ๊ตฌํ•˜๋Š” ํ๋ ˆ์ด์…˜๋œ ๋ฐ์ดํ„ฐ์— ๋” ํ˜„์‹ค์ ์ธ ๋ฐ์ดํ„ฐ ์ œ์•ฝ ์‹œ๋‚˜๋ฆฌ์˜ค์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ์›น ๋ฐ์ดํ„ฐ๋Š” ํ’๋ถ€ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ค‘๋ณต ์ œ๊ฑฐ๊ฐ€ ์šฐ๋ฆฌ๊ฐ€ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํฌ๊ธฐ๋ฅผ ์ œํ•œํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
* **ํ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์™€ ์›น ๋ฐ์ดํ„ฐ์˜ ์ฐจ์ด์ **. ์›น ๋ฐ์ดํ„ฐ๋Š” ์ค‘๋ณต์— ๋” ๋ฏผ๊ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์›น ๋ฐ์ดํ„ฐ์—์„œ ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ์ค‘๋ณต(์˜ˆ: ์ŠคํŒธ)์ด ํ๋ ˆ์ด์…˜๋œ ๋ฐ์ดํ„ฐ์—์„œ ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ์ค‘๋ณต๋ณด๋‹ค ๋” ํ•ด๋กœ์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ด ์ž‘์—…์—์„œ ์—ฐ๊ตฌํ•˜์ง€ ์•Š์€ ์ค‘๋ณต ์ œ๊ฑฐ์— ๋Œ€ํ•œ ์ •์„ฑ์  ์š”์†Œ๋ฅผ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. * **์ค‘๋ณต ์ œ๊ฑฐ ํŒŒ์ดํ”„๋ผ์ธ์˜ ์ฐจ์ด์ **. Biderman ๋“ฑ(2023)์€ Lee ๋“ฑ(2022)์˜ MinHash ์„ค์ •์„ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋Œ€๋ถ€๋ถ„ ์šฐ๋ฆฌ์™€ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ๋Š” ์ •ํ™•ํ•œ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ค‘๋ณต ์ œ๊ฑฐ๋กœ ์ธํ•ด ํฌ๊ธฐ๊ฐ€ 30% ๊ฐ์†Œํ•˜๋Š” ๋ฐ˜๋ฉด, ์šฐ๋ฆฌ์˜ ์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๋” ์ ๊ทน์ ์ด์–ด์„œ ํฌ๊ธฐ๊ฐ€ 45% ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด ํ‘œ 5์˜ ๊ฒฐ๊ณผ์—์„œ ์ค‘๋ณต ์ œ๊ฑฐ๋กœ ์ธํ•œ ์ด๋“์ด ๊ทธ๋ฆผ 9์˜ ๊ฒฐ๊ณผ๋ณด๋‹ค ๋” ๊ฐ•ํ•˜๊ฒŒ ๋‚˜ํƒ€๋‚œ ์ด์œ ์ผ ๊ฒƒ์ž…๋‹ˆ๋‹ค. * **์‚ฌ์ „ ํ•™์Šต์˜ ์ฐจ์ด์ **. ๋งˆ์ง€๋ง‰์œผ๋กœ, Biderman ๋“ฑ(2023)์€ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋ถ€๋ถ„์ ์œผ๋กœ ์ถ”๊ฐ€ ์—ํฌํฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ 300GT์— ๋„๋‹ฌํ•˜๋Š” ๋ฐ˜๋ฉด, ์šฐ๋ฆฌ๋Š” ํ•ญ์ƒ ๋‹จ์ผ ์—ํฌํฌ๋ฅผ ์ˆ˜ํ–‰ํ•œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ ๊ทธ๋“ค์ด ์—ฐ๊ตฌํ•˜๋Š” ํ๋ ˆ์ด์…˜๋œ ๋ฐ์ดํ„ฐ์— ๋” ํ˜„์‹ค์ ์ธ ๋ฐ์ดํ„ฐ ์ œ์•ฝ ์‹œ๋‚˜๋ฆฌ์˜ค์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ์›น ๋ฐ์ดํ„ฐ๋Š” ํ’๋ถ€ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ค‘๋ณต ์ œ๊ฑฐ๊ฐ€ ์šฐ๋ฆฌ๊ฐ€ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํฌ๊ธฐ๋ฅผ ์ œํ•œํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
Inconsistency is another major challenge in reasoning, Dua et al. (2022) iteratively uses previous reasoning results as prompts until the model gives a consistent answer. Paul et al. (2023) trains a critic model to provide structured feedback on the reasoning process. Self-Refine (Madaan et al., 2023) performs iterative self-feedback and refinement to alleviate errors in reasoning. Compared with Self-Refine, Reflexion (Shinn et al., 2023) introduces reinforcement learning for reflection, which additionally brings decision-making capability. Meanwhile, some work introduces backward reasoning (Yu et al., 2023) for verification. RCoT (Xue et al., 2023) reconstructs the question according to the reasoning chains, and its inconsistency with the original question exposes errors in the reasoning process. FOBAR (Jiang et al., 2023) and Self Verification (Weng et al., 2022) perform verification by deducing the conditions in the question from the answer. FOBAR infers the variables in the question, and Self Verification infers the conditions in the question. However, Huang et al. (2023) finds that LLMs struggle to self-correct without external feedback, and it could even lead to a performance decline.
์ผ๊ด€์„ฑ์ด ์—†๋Š” ๊ฒƒ์€ ์ถ”๋ก ์˜ ๋˜ ๋‹ค๋ฅธ ์ฃผ์š” ๊ณผ์ œ์ด๋ฉฐ, Dua ๋“ฑ(2022)์€ ๋ชจ๋ธ์ด ์ผ๊ด€๋œ ๋‹ต์„ ์ œ๊ณตํ•  ๋•Œ๊นŒ์ง€ ์ด์ „ ์ถ”๋ก  ๊ฒฐ๊ณผ๋ฅผ ํ”„๋กฌํ”„ํŠธ๋กœ ๋ฐ˜๋ณต์ ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Paul ๋“ฑ(2023)์€ ์ถ”๋ก  ๊ณผ์ •์— ๋Œ€ํ•œ ๊ตฌ์กฐํ™”๋œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ๋น„ํ‰ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. Self-Refine(Madaan et al., 2023)์€ ๋ฐ˜๋ณต์ ์ธ ์ž์ฒด ํ”ผ๋“œ๋ฐฑ ๋ฐ ๊ฐœ์„ ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ถ”๋ก ์˜ ์˜ค๋ฅ˜๋ฅผ ์™„ํ™”ํ•ฉ๋‹ˆ๋‹ค. Self-Refine๊ณผ ๋น„๊ตํ•˜์—ฌ Reflexion(Shinn et al., 2023)์€ ๋ฐ˜์„ฑ์„ ์œ„ํ•œ Reinforcement Learning์„ ๋„์ž…ํ•˜์—ฌ ์˜์‚ฌ ๊ฒฐ์ • ๊ธฐ๋Šฅ์„ ์ถ”๊ฐ€๋กœ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฒ€์ฆ์„ ์œ„ํ•ด ์—ญ ์ถ”๋ก (Yu et al., 2023)์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. RCoT(Xue et al., 2023)๋Š” ์ถ”๋ก  ์ฒด์ธ์— ๋”ฐ๋ผ ์งˆ๋ฌธ์„ ์žฌ๊ตฌ์„ฑํ•˜๋ฉฐ, ์›๋ž˜ ์งˆ๋ฌธ๊ณผ์˜ ๋ถˆ์ผ์น˜๋Š” ์ถ”๋ก  ๊ณผ์ •์˜ ์˜ค๋ฅ˜๋ฅผ ๋…ธ์ถœ์‹œํ‚ต๋‹ˆ๋‹ค. FOBAR(Jiang et al., 2023)์™€ Self Verification(Weng et al., 2022)์€ ์งˆ๋ฌธ์˜ ์กฐ๊ฑด์„ ๋‹ต๋ณ€์—์„œ ์ถ”๋ก ํ•˜์—ฌ ๊ฒ€์ฆ์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. FOBAR๋Š” ์งˆ๋ฌธ์˜ ๋ณ€์ˆ˜๋ฅผ ์ถ”๋ก ํ•˜๊ณ , Self Verification์€ ์งˆ๋ฌธ์˜ ์กฐ๊ฑด์„ ์ถ”๋ก ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Huang ๋“ฑ(2023)์€ LLM์ด ์™ธ๋ถ€ ํ”ผ๋“œ๋ฐฑ ์—†์ด ์ž์ฒด ์ˆ˜์ •ํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์œผ๋ฉฐ, ์„ฑ๋Šฅ์ด ์ €ํ•˜๋  ์ˆ˜๋„ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
์ผ๊ด€์„ฑ์ด ์—†๋Š” ๊ฒƒ์€ ์ถ”๋ก ์˜ ๋˜ ๋‹ค๋ฅธ ์ฃผ์š” ๊ณผ์ œ์ด๋ฉฐ, Dua ๋“ฑ(2022)์€ ๋ชจ๋ธ์ด ์ผ๊ด€๋œ ๋‹ต์„ ์ œ๊ณตํ•  ๋•Œ๊นŒ์ง€ ์ด์ „ ์ถ”๋ก  ๊ฒฐ๊ณผ๋ฅผ ํ”„๋กฌํ”„ํŠธ๋กœ ๋ฐ˜๋ณต์ ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Paul ๋“ฑ(2023)์€ ์ถ”๋ก  ๊ณผ์ •์— ๋Œ€ํ•œ ๊ตฌ์กฐํ™”๋œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ๋น„ํ‰ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์…€ํ”„ ๋ฆฌํŒŒ์ธ(Madaan et al., 2023)์€ ๋ฐ˜๋ณต์ ์ธ ์ž์ฒด ํ”ผ๋“œ๋ฐฑ ๋ฐ ๊ฐœ์„ ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ถ”๋ก ์˜ ์˜ค๋ฅ˜๋ฅผ ์™„ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์…€ํ”„ ๋ฆฌํŒŒ์ธ๊ณผ ๋น„๊ตํ•˜์—ฌ ๋ฆฌํ”Œ๋ ‰์…˜(์‹  ๋“ฑ, 2023)์€ ๋ฐ˜์„ฑ์„ ์œ„ํ•œ ๊ฐ•ํ™” ํ•™์Šต์„ ๋„์ž…ํ•˜์—ฌ ์˜์‚ฌ ๊ฒฐ์ • ๊ธฐ๋Šฅ์„ ์ถ”๊ฐ€๋กœ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฒ€์ฆ์„ ์œ„ํ•ด ์—ญ ์ถ”๋ก (Yu et al., 2023)์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. RCoT(Xue et al., 2023)๋Š” ์ถ”๋ก  ์ฒด์ธ์— ๋”ฐ๋ผ ์งˆ๋ฌธ์„ ์žฌ๊ตฌ์„ฑํ•˜๋ฉฐ, ์›๋ž˜ ์งˆ๋ฌธ๊ณผ์˜ ๋ถˆ์ผ์น˜๋Š” ์ถ”๋ก  ๊ณผ์ •์˜ ์˜ค๋ฅ˜๋ฅผ ๋…ธ์ถœ์‹œํ‚ต๋‹ˆ๋‹ค. FOBAR(Jiang et al., 2023)์™€ ์ž์ฒด ๊ฒ€์ฆ(Weng et al., 2022)์€ ์งˆ๋ฌธ์˜ ์กฐ๊ฑด์„ ๋‹ต๋ณ€์—์„œ ์ถ”๋ก ํ•˜์—ฌ ๊ฒ€์ฆ์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. FOBAR๋Š” ์งˆ๋ฌธ์˜ ๋ณ€์ˆ˜๋ฅผ ์ถ”๋ก ํ•˜๊ณ , ์ž์ฒด ๊ฒ€์ฆ์€ ์งˆ๋ฌธ์˜ ์กฐ๊ฑด์„ ์ถ”๋ก ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Huang ๋“ฑ(2023)์€ LLM์ด ์™ธ๋ถ€ ํ”ผ๋“œ๋ฐฑ ์—†์ด ์ž์ฒด ์ˆ˜์ •ํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์œผ๋ฉฐ, ์„ฑ๋Šฅ์ด ์ €ํ•˜๋  ์ˆ˜๋„ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
Large language models (LLMs) demonstrate remarkable performance across a spectrum of languages. In this work, we delve into the question: How do LLMs handle multilingual? We introduce a framework that depicts LLMs' processing of multilingual inputs: In the first several layers, LLMs understand the question, converting multilingual inputs into English to facilitate the task-solving phase. In the intermediate layers, LLMs engage in problem-solving by thinking in English and incorporating multilingual knowledge to obtain factual content, leveraging the self-attention and feed-forward structures, respectively. In the last several layers, LLMs generate responses that align with the original language of the query. In addition, we investigate the existence of language-specific neurons when processing a certain language. To detect neurons activated by the input language, even without labels, we innovatively design a Parallel Language specific Neuron Detection (PLND) method that effectively measures the significance of neurons when handling multilingual inputs. By comprehensive ablation analysis through deactivating neurons of different layers and structures, we verify the framework that we propose.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLMs)์€ ๋‹ค์–‘ํ•œ ์–ธ์–ด์—์„œ ๋†€๋ผ์šด ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” LLMs์ด ๋‹ค๊ตญ์–ด๋ฅผ ์–ด๋–ป๊ฒŒ ์ฒ˜๋ฆฌํ•˜๋Š”์ง€์— ๋Œ€ํ•œ ์งˆ๋ฌธ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” LLMs์ด ๋‹ค๊ตญ์–ด ์ž…๋ ฅ์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์„ ์„ค๋ช…ํ•˜๋Š” ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค: ์ฒ˜์Œ ๋ช‡ ๊ฐœ์˜ ๋ ˆ์ด์–ด์—์„œ LLMs์€ ์งˆ๋ฌธ์„ ์ดํ•ดํ•˜๊ณ  ๋‹ค๊ตญ์–ด ์ž…๋ ฅ์„ ์˜์–ด๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ ์ž‘์—… ํ•ด๊ฒฐ ๋‹จ๊ณ„๋ฅผ ์šฉ์ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ๋ ˆ์ด์–ด์—์„œ LLMs์€ ์˜์–ด๋กœ ์‚ฌ๊ณ ํ•˜๊ณ  ๋‹ค๊ตญ์–ด ์ง€์‹์„ ํ†ตํ•ฉํ•˜์—ฌ ์‚ฌ์‹ค์ ์ธ ๋‚ด์šฉ์„ ํ™•๋ณดํ•˜๋Š” ๋“ฑ ๋ฌธ์ œ ํ•ด๊ฒฐ์— ์ฐธ์—ฌํ•˜๋ฉฐ, ๊ฐ๊ฐ self-attention ๋ฐ feed-forward ๊ตฌ์กฐ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰ ๋ช‡ ๊ฐœ์˜ ๋ ˆ์ด์–ด์—์„œ LLMs์€ ์ฟผ๋ฆฌ์˜ ์›๋ž˜ ์–ธ์–ด์™€ ์ผ์น˜ํ•˜๋Š” ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํŠน์ • ์–ธ์–ด๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์˜ ์กด์žฌ ์—ฌ๋ถ€๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ ˆ์ด๋ธ”์ด ์—†๋”๋ผ๋„ ์ž…๋ ฅ ์–ธ์–ด์— ์˜ํ•ด ํ™œ์„ฑํ™”๋˜๋Š” ๋‰ด๋Ÿฐ์„ ๊ฐ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค๊ตญ์–ด ์ž…๋ ฅ์„ ์ฒ˜๋ฆฌํ•  ๋•Œ ๋‰ด๋Ÿฐ์˜ ์ค‘์š”์„ฑ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ธก์ •ํ•˜๋Š” Parallel Language specific Neuron Detection (PLND) ๋ฐฉ๋ฒ•์„ ํ˜์‹ ์ ์œผ๋กœ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ ˆ์ด์–ด์™€ ๊ตฌ์กฐ์˜ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋Š” ํฌ๊ด„์ ์ธ ์ ˆ์ œ ๋ถ„์„์„ ํ†ตํ•ด ์ œ์•ˆํ•œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ๋‹ค์–‘ํ•œ ์–ธ์–ด์—์„œ ๋†€๋ผ์šด ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” LLM์ด ๋‹ค๊ตญ์–ด๋ฅผ ์–ด๋–ป๊ฒŒ ์ฒ˜๋ฆฌํ•˜๋Š”์ง€์— ๋Œ€ํ•œ ์งˆ๋ฌธ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” LLM์ด ๋‹ค๊ตญ์–ด ์ž…๋ ฅ์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์„ ์„ค๋ช…ํ•˜๋Š” ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค: ์ฒ˜์Œ ๋ช‡ ๊ฐœ์˜ ๋ ˆ์ด์–ด์—์„œ LLM์€ ์งˆ๋ฌธ์„ ์ดํ•ดํ•˜๊ณ  ๋‹ค๊ตญ์–ด ์ž…๋ ฅ์„ ์˜์–ด๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ ์ž‘์—… ํ•ด๊ฒฐ ๋‹จ๊ณ„๋ฅผ ์šฉ์ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ๋ ˆ์ด์–ด์—์„œ LLM์€ ์˜์–ด๋กœ ์‚ฌ๊ณ ํ•˜๊ณ  ๋‹ค๊ตญ์–ด ์ง€์‹์„ ํ†ตํ•ฉํ•˜์—ฌ ์‚ฌ์‹ค์ ์ธ ๋‚ด์šฉ์„ ํ™•๋ณดํ•˜๋Š” ๋“ฑ ๋ฌธ์ œ ํ•ด๊ฒฐ์— ์ฐธ์—ฌํ•˜๋ฉฐ, ๊ฐ๊ฐ ์ž๊ธฐ ์ฃผ์˜ ๋ฐ ํ”ผ๋“œ ํฌ์›Œ๋“œ ๊ตฌ์กฐ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰ ๋ช‡ ๊ฐœ์˜ ๋ ˆ์ด์–ด์—์„œ LLM์€ ์ฟผ๋ฆฌ์˜ ์›๋ž˜ ์–ธ์–ด์™€ ์ผ์น˜ํ•˜๋Š” ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํŠน์ • ์–ธ์–ด๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์˜ ์กด์žฌ ์—ฌ๋ถ€๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ ˆ์ด๋ธ”์ด ์—†๋”๋ผ๋„ ์ž…๋ ฅ ์–ธ์–ด์— ์˜ํ•ด ํ™œ์„ฑํ™”๋˜๋Š” ๋‰ด๋Ÿฐ์„ ๊ฐ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค๊ตญ์–ด ์ž…๋ ฅ์„ ์ฒ˜๋ฆฌํ•  ๋•Œ ๋‰ด๋Ÿฐ์˜ ์ค‘์š”์„ฑ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ธก์ •ํ•˜๋Š” ๋ณ‘๋ ฌ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ ๊ฐ์ง€(PLND) ๋ฐฉ๋ฒ•์„ ํ˜์‹ ์ ์œผ๋กœ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ ˆ์ด์–ด์™€ ๊ตฌ์กฐ์˜ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋Š” ํฌ๊ด„์ ์ธ ์ ˆ์ œ ๋ถ„์„์„ ํ†ตํ•ด ์ œ์•ˆํ•œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค.
In the area of Natural Language Processing (NLP), an extensive range of data augmentation techniques has been introduced for diverse applications, such as translation, summarization, and dialogue (Chen et al., 2023b; Fadaee et al., 2017; Xie et al., 2016; Lample et al., 2017; Guo et al., 2020a; Xie et al., 2022; Wang et al., 2018). The noise-based strategies commonly involve the integration of random continuous noise (Cheng et al., 2018; Jain et al., 2023; Kedzie and McKeown, 2019), token dropout (Xie et al., 2016; Gal and Ghahramani, 2015), token swapping (Wei and Zou, 2019), and more advanced mixup strategies (Guo et al., 2020b; Xie et al., 2022). Recently, in the context of regularization for LLMs, Jain et al. (2023) has demonstrated that the application of continuous noise across all tokens during the instruction tuning can significantly improve the length and detail of the generated responses. However, this technique does not improve the reasoning ability. We adopt the discrete token noise through token dropout (Xie et al., 2016; Gal and Ghahramani, 2015) to reasoning tasks, finding that it works well, which may complement the dependency-rich structure characteristic of reasoning data.
์ž์—ฐ์–ด ์ฒ˜๋ฆฌ(NLP) ์˜์—ญ์—์„œ๋Š” ๋ฒˆ์—ญ, ์š”์•ฝ, ๋Œ€ํ™” ๋“ฑ ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ์œ„ํ•ด ๊ด‘๋ฒ”์œ„ํ•œ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๊ธฐ๋ฒ•์ด ๋„์ž…๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Chen et al., 2023b; Fadaee et al., 2017; Xie et al., 2016; Lample et al., 2017; Guo et al., 2020a; Xie et al., 2022; Wang et al., 2018). ๋…ธ์ด์ฆˆ ๊ธฐ๋ฐ˜ ์ „๋žต์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฌด์ž‘์œ„ ์—ฐ์† ๋…ธ์ด์ฆˆ(Cheng et al., 2018; Jain et al., 2023; Kedzie and McKeown, 2019), token dropout(Xie et al., 2016; Gal and Ghahramani, 2015), token swapping(Wei and Zou, 2019) ๋“ฑ์˜ ํ†ตํ•ฉ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” ์ •๊ทœํ™”์˜ ๋งฅ๋ฝ์—์„œ LLM์— ๋Œ€ํ•œ ์ •๊ทœํ™”์˜ ๋งฅ๋ฝ์—์„œ Jain et al. (2023)์€ Instruction Tuning ์ค‘์— ๋ชจ๋“  ํ† ํฐ์— ์—ฐ์† ๋…ธ์ด์ฆˆ๋ฅผ ์ ์šฉํ•˜๋ฉด ์ƒ์„ฑ๋œ ์‘๋‹ต์˜ ๊ธธ์ด์™€ ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ๊ธฐ๋ฒ•์€ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ถ”๋ก  ์ž‘์—…์— token dropout(Xie et al., 2016; Gal and Ghahramani, 2015)์„ ํ†ตํ•œ ๋ถˆ์—ฐ์† ํ† ํฐ ๋…ธ์ด์ฆˆ๋ฅผ ์ฑ„ํƒํ•˜์—ฌ ์ถ”๋ก  ๋ฐ์ดํ„ฐ์˜ ์ข…์†์„ฑ์ด ํ’๋ถ€ํ•œ ๊ตฌ์กฐ์  ํŠน์„ฑ์„ ๋ณด์™„ํ•  ์ˆ˜ ์žˆ๋Š” ํšจ๊ณผ๊ฐ€ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
์ž์—ฐ์–ด ์ฒ˜๋ฆฌ(NLP) ์˜์—ญ์—์„œ๋Š” ๋ฒˆ์—ญ, ์š”์•ฝ, ๋Œ€ํ™” ๋“ฑ ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ์œ„ํ•ด ๊ด‘๋ฒ”์œ„ํ•œ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๊ธฐ๋ฒ•์ด ๋„์ž…๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Chen et al., 2023b; Fadaee et al., 2017; Xie et al., 2016; Lample et al., 2017; Guo et al., 2020a; Xie et al., 2022; Wang et al., 2018). ๋…ธ์ด์ฆˆ ๊ธฐ๋ฐ˜ ์ „๋žต์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฌด์ž‘์œ„ ์—ฐ์† ๋…ธ์ด์ฆˆ(Cheng et al., 2018; Jain et al., 2023; Kedzie and McKeown, 2019), ํ† ํฐ ๋“œ๋กญ์•„์›ƒ(Xie et al., 2016; Gal and Ghahramani, 2015), ํ† ํฐ ์Šค์™‘(Wei and Zou, 2019) ๋“ฑ์˜ ํ†ตํ•ฉ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” ์ •๊ทœํ™”์˜ ๋งฅ๋ฝ์—์„œ LLM์— ๋Œ€ํ•œ ์ •๊ทœํ™”์˜ ๋งฅ๋ฝ์—์„œ Jain et al. (2023)์€ ๋ช…๋ น ํŠœ๋‹ ์ค‘์— ๋ชจ๋“  ํ† ํฐ์— ์—ฐ์† ๋…ธ์ด์ฆˆ๋ฅผ ์ ์šฉํ•˜๋ฉด ์ƒ์„ฑ๋œ ์‘๋‹ต์˜ ๊ธธ์ด์™€ ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ๊ธฐ๋ฒ•์€ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ถ”๋ก  ์ž‘์—…์— ํ† ํฐ ๋“œ๋กญ์•„์›ƒ(Xie et al., 2016; Gal and Ghahramani, 2015)์„ ํ†ตํ•œ ๋ถˆ์—ฐ์† ํ† ํฐ ๋…ธ์ด์ฆˆ๋ฅผ ์ฑ„ํƒํ•˜์—ฌ ์ถ”๋ก  ๋ฐ์ดํ„ฐ์˜ ์ข…์†์„ฑ์ด ํ’๋ถ€ํ•œ ๊ตฌ์กฐ์  ํŠน์„ฑ์„ ๋ณด์™„ํ•  ์ˆ˜ ์žˆ๋Š” ํšจ๊ณผ๊ฐ€ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Model-based approach_. As a surrogate for human-based approaches, model-based approaches serve to diminish the reliance on human involvement, and enable more efficient and scalable evaluation. In addition, LLMs can provide meaningful explanations for the assigned rating scores,thereby enhancing the interpretability of evaluations. Despite their scalability and explanability, model-based approaches have been found to suffer from several issues, including position, verbosity, and self-enhancement bias [727]. Specially, position bias (_i.e.,_ the order to present the responses) refers to the fact that LLMs tend to assign high scores for the answers at specific positions over others, verbosity bias means that LLMs favor verbose answers even if they are short in quality compared with shorter answers, and self-enhancement bias indicates that LLMs often over-rate in their own generations. In addition, since LLMs have limited capacities in solving complex reasoning problems, they cannot serve as qualified evaluators for some difficult tasks (_e.g.,_ mathematical reasoning). These limitations can be mitigated to some extent by specific prompt engineering and fine-tuning strategies [727].
\(\bullet\)_๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹_. ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ธ๊ฐ„ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ๋Œ€๋ฆฌ์ž ์—ญํ• ์„ ํ•˜๋ฉฐ, ์ธ๊ฐ„์˜ ๊ฐœ์ž…์— ๋Œ€ํ•œ ์˜์กด๋„๋ฅผ ๋‚ฎ์ถ”๊ณ  ๋ณด๋‹ค ํšจ์œจ์ ์ด๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ํ‰๊ฐ€๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์€ ํ• ๋‹น๋œ ๋“ฑ๊ธ‰ ์ ์ˆ˜์— ๋Œ€ํ•œ ์˜๋ฏธ ์žˆ๋Š” ์„ค๋ช…์„ ์ œ๊ณตํ•˜์—ฌ ํ‰๊ฐ€์˜ ํ•ด์„ ๊ฐ€๋Šฅ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ™•์žฅ์„ฑ๊ณผ ์„ค๋ช… ๊ฐ€๋Šฅ์„ฑ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์œ„์น˜, ์žฅํ™ฉํ•จ, ์ž๊ธฐ ํ–ฅ์ƒ ํŽธํ–ฅ์„ฑ[727] ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ๋ฌธ์ œ๋กœ ์ธํ•ด ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์œ„์น˜ ํŽธํ–ฅ('์ฆ‰, ์‘๋‹ต์„ ์ œ์‹œํ•˜๋Š” ์ˆœ์„œ')์€ LLM์ด ํŠน์ • ์œ„์น˜์˜ ๋‹ต๋ณ€์— ๋‹ค๋ฅธ ๋‹ต๋ณ€๋ณด๋‹ค ๋†’์€ ์ ์ˆ˜๋ฅผ ๋ถ€์—ฌํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Œ์„ ์˜๋ฏธํ•˜๋ฉฐ, ์žฅํ™ฉํ•จ ํŽธํ–ฅ์€ LLM์ด ์งง์€ ๋‹ต๋ณ€์— ๋น„ํ•ด ํ’ˆ์งˆ์ด ๋–จ์–ด์ง€๋”๋ผ๋„ ์žฅํ™ฉํ•œ ๋‹ต๋ณ€์„ ์„ ํ˜ธํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•˜๋ฉฐ, ์ž๊ธฐ ํ–ฅ์ƒ ํŽธํ–ฅ์€ LLM์ด ์ž์ฒด ์ƒ์„ฑ์—์„œ ์ข…์ข… ๊ณผ๋Œ€ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์€ ๋ณต์žกํ•œ ์ถ”๋ก  ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์ผ๋ถ€ ์–ด๋ ค์šด ์ž‘์—…(_์˜ˆ: ์ˆ˜ํ•™์  ์ถ”๋ก )์— ๋Œ€ํ•ด์„œ๋Š” ์ ๊ฒฉํ•œ ํ‰๊ฐ€์ž ์—ญํ• ์„ ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋Š” ํŠน์ • prompt engineering ๋ฐ Fine-tuning ์ „๋žต[727]์„ ํ†ตํ•ด ์–ด๋Š ์ •๋„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹_. ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ธ๊ฐ„ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ๋Œ€๋ฆฌ์ž ์—ญํ• ์„ ํ•˜๋ฉฐ, ์ธ๊ฐ„์˜ ๊ฐœ์ž…์— ๋Œ€ํ•œ ์˜์กด๋„๋ฅผ ๋‚ฎ์ถ”๊ณ  ๋ณด๋‹ค ํšจ์œจ์ ์ด๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ํ‰๊ฐ€๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์€ ํ• ๋‹น๋œ ๋“ฑ๊ธ‰ ์ ์ˆ˜์— ๋Œ€ํ•œ ์˜๋ฏธ ์žˆ๋Š” ์„ค๋ช…์„ ์ œ๊ณตํ•˜์—ฌ ํ‰๊ฐ€์˜ ํ•ด์„ ๊ฐ€๋Šฅ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ™•์žฅ์„ฑ๊ณผ ์„ค๋ช… ๊ฐ€๋Šฅ์„ฑ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์œ„์น˜, ์žฅํ™ฉํ•จ, ์ž๊ธฐ ํ–ฅ์ƒ ํŽธํ–ฅ์„ฑ[727] ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ๋ฌธ์ œ๋กœ ์ธํ•ด ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์œ„์น˜ ํŽธํ–ฅ('์ฆ‰, ์‘๋‹ต์„ ์ œ์‹œํ•˜๋Š” ์ˆœ์„œ')์€ LLM์ด ํŠน์ • ์œ„์น˜์˜ ๋‹ต๋ณ€์— ๋‹ค๋ฅธ ๋‹ต๋ณ€๋ณด๋‹ค ๋†’์€ ์ ์ˆ˜๋ฅผ ๋ถ€์—ฌํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Œ์„ ์˜๋ฏธํ•˜๋ฉฐ, ์žฅํ™ฉํ•จ ํŽธํ–ฅ์€ LLM์ด ์งง์€ ๋‹ต๋ณ€์— ๋น„ํ•ด ํ’ˆ์งˆ์ด ๋–จ์–ด์ง€๋”๋ผ๋„ ์žฅํ™ฉํ•œ ๋‹ต๋ณ€์„ ์„ ํ˜ธํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•˜๋ฉฐ, ์ž๊ธฐ ํ–ฅ์ƒ ํŽธํ–ฅ์€ LLM์ด ์ž์ฒด ์ƒ์„ฑ์—์„œ ์ข…์ข… ๊ณผ๋Œ€ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์€ ๋ณต์žกํ•œ ์ถ”๋ก  ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์ผ๋ถ€ ์–ด๋ ค์šด ์ž‘์—…(_์˜ˆ: ์ˆ˜ํ•™์  ์ถ”๋ก )์— ๋Œ€ํ•ด์„œ๋Š” ์ ๊ฒฉํ•œ ํ‰๊ฐ€์ž ์—ญํ• ์„ ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋Š” ํŠน์ • ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง ๋ฐ ๋ฏธ์„ธ ์กฐ์ • ์ „๋žต[727]์„ ํ†ตํ•ด ์–ด๋Š ์ •๋„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Models with Hundreds of Billions of Parameters**. For models in this category, only a handful of models have been publicly released. For example, OPT [90], OPT-IML [95], BLOOM [78], and BLOOMZ [94] have nearly the same number of parameters as GPT-3 (175B version), while GLM [93] and Galactica [35] have 130B and 120B parameters, respectively. Among them, OPT (175B version), with the instruction-tuned version OPT-IML, has been specially motivated for open sharing, which aims to enable researchers to carry out reproducible research at scale. For research in cross-lingual generalization, BLOOM (176B version) and BLOOMZ (176B version) can be used as base models, due to the competence in multilingual language modeling tasks. As a bilingual LLM, GLM has also provided a popular small-sized Chinese chat model ChatGLM2-6B (a updated version for ChatGLM-6B), which is featured with many improvements in efficiency and capacity (_e.g._, quantization, 32K-length context, fast inference rate). Models of this scale typically require thousands of GPUs or TPUs to train. For instance, OPT (175B version) used 992 A100-80GB GPUs, while GLM (130B version) used a cluster of 96 NVIDIA DGX-A100 (8x40G) GPU nodes.
**์ˆ˜๋ฐฑ์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ**. ์ด ๋ฒ”์ฃผ์— ์†ํ•˜๋Š” ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, ๊ณต๊ฐœ์ ์œผ๋กœ ์ถœ์‹œ๋œ ๋ชจ๋ธ์€ ๊ทน์†Œ์ˆ˜์— ๋ถˆ๊ณผํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, OPT [90], OPT-IML [95], BLOOM [78], BLOOMZ [94]๋Š” GPT-3(175B ๋ฒ„์ „)์™€ ๊ฑฐ์˜ ๋™์ผํ•œ ์ˆ˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, GLM [93]๊ณผ Galactica [35]๋Š” ๊ฐ๊ฐ 130B์™€ 120B ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ค‘ OPT(175B ๋ฒ„์ „)๋Š” Instruction-tuned ๋ฒ„์ „์ธ OPT-IML๊ณผ ํ•จ๊ป˜ ์—ฐ๊ตฌ์ž๋“ค์ด ๋Œ€๊ทœ๋ชจ๋กœ ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ๊ณต๊ฐœ์ ์œผ๋กœ ๊ณต์œ ํ•˜๊ธฐ ์œ„ํ•ด ๋™๊ธฐ๋ฅผ ๋ถ€์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๊ตญ์–ด ์ผ๋ฐ˜ํ™” ์—ฐ๊ตฌ์˜ ๊ฒฝ์šฐ, ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…์—์„œ์˜ ์—ญ๋Ÿ‰์œผ๋กœ ์ธํ•ด BLOOM(176B ๋ฒ„์ „)๊ณผ BLOOMZ(176B ๋ฒ„์ „)๋ฅผ ๊ธฐ๋ณธ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์ค‘ ์–ธ์–ด LLM์ธ GLM์€ ๋˜ํ•œ ํšจ์œจ์„ฑ๊ณผ ์šฉ๋Ÿ‰(_์˜ˆ: Quantization, 32K ๊ธธ์ด ์ปจํ…์ŠคํŠธ, ๋น ๋ฅธ ์ถ”๋ก  ์†๋„)์ด ํฌ๊ฒŒ ๊ฐœ์„ ๋œ ์ธ๊ธฐ ์žˆ๋Š” ์†Œํ˜• ์ค‘๊ตญ์–ด ์ฑ„ํŒ… ๋ชจ๋ธ์ธ ChatGLM2-6B(ChatGLM-6B์˜ ์—…๋ฐ์ดํŠธ ๋ฒ„์ „)๋ฅผ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ˆ˜์ฒœ ๊ฐœ์˜ GPU ๋˜๋Š” TPU๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, OPT(175B ๋ฒ„์ „)๋Š” 992๊ฐœ์˜ A100-80GB GPU๋ฅผ ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ, GLM(130B ๋ฒ„์ „)์€ 96๊ฐœ์˜ NVIDIA DGX-A100(8x40G) GPU ๋…ธ๋“œ๋กœ ๊ตฌ์„ฑ๋œ ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
**์ˆ˜๋ฐฑ์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ**. ์ด ๋ฒ”์ฃผ์— ์†ํ•˜๋Š” ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, ๊ณต๊ฐœ์ ์œผ๋กœ ์ถœ์‹œ๋œ ๋ชจ๋ธ์€ ๊ทน์†Œ์ˆ˜์— ๋ถˆ๊ณผํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, OPT [90], OPT-IML [95], BLOOM [78], BLOOMZ [94]๋Š” GPT-3(175B ๋ฒ„์ „)์™€ ๊ฑฐ์˜ ๋™์ผํ•œ ์ˆ˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, GLM [93]๊ณผ Galactica [35]๋Š” ๊ฐ๊ฐ 130B์™€ 120B ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ค‘ OPT(175B ๋ฒ„์ „)๋Š” ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฒ„์ „์ธ OPT-IML๊ณผ ํ•จ๊ป˜ ์—ฐ๊ตฌ์ž๋“ค์ด ๋Œ€๊ทœ๋ชจ๋กœ ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ๊ณต๊ฐœ์ ์œผ๋กœ ๊ณต์œ ํ•˜๊ธฐ ์œ„ํ•ด ๋™๊ธฐ๋ฅผ ๋ถ€์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๊ตญ์–ด ์ผ๋ฐ˜ํ™” ์—ฐ๊ตฌ์˜ ๊ฒฝ์šฐ, ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…์—์„œ์˜ ์—ญ๋Ÿ‰์œผ๋กœ ์ธํ•ด BLOOM(176B ๋ฒ„์ „)๊ณผ BLOOMZ(176B ๋ฒ„์ „)๋ฅผ ๊ธฐ๋ณธ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์ค‘ ์–ธ์–ด LLM์ธ GLM์€ ๋˜ํ•œ ํšจ์œจ์„ฑ๊ณผ ์šฉ๋Ÿ‰(_์˜ˆ: ์–‘์žํ™”, 32K ๊ธธ์ด ์ปจํ…์ŠคํŠธ, ๋น ๋ฅธ ์ถ”๋ก  ์†๋„)์ด ํฌ๊ฒŒ ๊ฐœ์„ ๋œ ์ธ๊ธฐ ์žˆ๋Š” ์†Œํ˜• ์ค‘๊ตญ์–ด ์ฑ„ํŒ… ๋ชจ๋ธ์ธ ChatGLM2-6B(ChatGLM-6B์˜ ์—…๋ฐ์ดํŠธ ๋ฒ„์ „)๋ฅผ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ˆ˜์ฒœ ๊ฐœ์˜ GPU ๋˜๋Š” TPU๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, OPT(175B ๋ฒ„์ „)๋Š” 992๊ฐœ์˜ A100-80GB GPU๋ฅผ ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ, GLM(130B ๋ฒ„์ „)์€ 96๊ฐœ์˜ NVIDIA DGX-A100(8x40G) GPU ๋…ธ๋“œ๋กœ ๊ตฌ์„ฑ๋œ ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_ChatGPT_. In November 2022, OpenAI released the conversation model ChatGPT, based on the GPT models (GPT-3.5 and GPT-4). As the official blog article introduced [131], ChatGPT was trained in a similar way as InstructGPT (called "a sibling model to InstructGPT" in the original post), while specially optimized for dialogue. They reported a difference between the training of ChatGPT and InstructGPT in the data collection setup: human-generated conversations (playing both the roles of user and AI) are combined with the InstructGPT dataset in a dialogue format for training ChatGPT. ChatGPT exhibited superior capacities in communicating with humans: possessing a vast store of knowledge, skill at reasoning on mathematical problems, tracing the context accurately in multi-turn dialogues, and aligning well with human values for safe use. Later on, the plugin mechanism has been supported in ChatGPT, which further extends the capacities of ChatGPT with existing tools or apps. So far, it seems to be the ever most powerful chatbot in the AI history. The launch of ChatGPT has a significant impact on the AI research in the future, which sheds light on the exploration of human-like AI systems.
\(\bullet\)_ChatGPT_. 2022๋…„ 11์›”, OpenAI๋Š” GPT ๋ชจ๋ธ(GPT-3.5 ๋ฐ GPT-4)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋Œ€ํ™” ๋ชจ๋ธ ChatGPT๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ณต์‹ ๋ธ”๋กœ๊ทธ ๊ธฐ์‚ฌ์—์„œ ์†Œ๊ฐœํ•œ ๋Œ€๋กœ [131], ChatGPT๋Š” InstructGPT์™€ ๋น„์Šทํ•œ ๋ฐฉ์‹์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ์œผ๋ฉฐ(์›๋ฌธ์—์„œ๋Š” "InstructGPT์˜ ํ˜•์ œ ๋ชจ๋ธ"์ด๋ผ๊ณ  ํ•จ), ๋Œ€ํ™”์— ํŠน๋ณ„ํžˆ ์ตœ์ ํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ์„ค์ •์—์„œ ChatGPT์™€ InstructGPT์˜ ์ฐจ์ด์ ์„ ๋ณด๊ณ ํ–ˆ์Šต๋‹ˆ๋‹ค: ์ธ๊ฐ„์ด ์ƒ์„ฑํ•œ ๋Œ€ํ™”(์‚ฌ์šฉ์ž์™€ AI์˜ ๋‘ ๊ฐ€์ง€ ์—ญํ• ์„ ๋ชจ๋‘ ์ˆ˜ํ–‰)๊ฐ€ ๋Œ€ํ™” ํ˜•์‹์˜ InstructGPT ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๊ฒฐํ•ฉ๋˜์–ด ChatGPT๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ChatGPT๋Š” ๋ฐฉ๋Œ€ํ•œ ์ง€์‹ ์ €์žฅ์†Œ, ์ˆ˜ํ•™ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์ถ”๋ก  ๊ธฐ์ˆ , ๋‹ค์ค‘ ํ„ด ๋Œ€ํ™”์—์„œ ๋ฌธ๋งฅ์„ ์ •ํ™•ํ•˜๊ฒŒ ์ถ”์ ํ•˜๊ณ , ์•ˆ์ „ํ•œ ์‚ฌ์šฉ์„ ์œ„ํ•œ ์ธ๊ฐ„์˜ ๊ฐ€์น˜๊ด€๊ณผ ์ž˜ ๋ถ€ํ•ฉํ•˜๋Š” ๋“ฑ ์ธ๊ฐ„๊ณผ ์†Œํ†ตํ•˜๋Š” ๋›ฐ์–ด๋‚œ ์—ญ๋Ÿ‰์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ํ”Œ๋Ÿฌ๊ทธ์ธ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ChatGPT์— ์ง€์›๋˜์–ด ๊ธฐ์กด ๋„๊ตฌ๋‚˜ ์•ฑ์œผ๋กœ ChatGPT์˜ ์—ญ๋Ÿ‰์„ ๋”์šฑ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์ง€๊ธˆ๊นŒ์ง€ AI ์—ญ์‚ฌ์ƒ ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ์ฑ—๋ด‡์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ChatGPT์˜ ์ถœ์‹œ๋Š” ์ธ๊ฐ„๊ณผ ๊ฐ™์€ AI ์‹œ์Šคํ…œ์˜ ํƒ๊ตฌ์— ๋น›์„ ๋น„์ถ”๋Š” ๋ฏธ๋ž˜์˜ AI ์—ฐ๊ตฌ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์ณค์Šต๋‹ˆ๋‹ค.
\(\bullet\)_ChatGPT_. 2022๋…„ 11์›”, OpenAI๋Š” GPT ๋ชจ๋ธ(GPT-3.5 ๋ฐ GPT-4)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋Œ€ํ™” ๋ชจ๋ธ ChatGPT๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ณต์‹ ๋ธ”๋กœ๊ทธ ๊ธฐ์‚ฌ์—์„œ ์†Œ๊ฐœํ•œ ๋Œ€๋กœ [131], ChatGPT๋Š” ์ธ์ŠคํŠธ๋ŸญํŠธGPT์™€ ๋น„์Šทํ•œ ๋ฐฉ์‹์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ์œผ๋ฉฐ(์›๋ฌธ์—์„œ๋Š” "์ธ์ŠคํŠธ๋ŸญํŠธGPT์˜ ํ˜•์ œ ๋ชจ๋ธ"์ด๋ผ๊ณ  ํ•จ), ๋Œ€ํ™”์— ํŠน๋ณ„ํžˆ ์ตœ์ ํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ์„ค์ •์—์„œ ChatGPT์™€ InstructGPT์˜ ์ฐจ์ด์ ์„ ๋ณด๊ณ ํ–ˆ์Šต๋‹ˆ๋‹ค: ์ธ๊ฐ„์ด ์ƒ์„ฑํ•œ ๋Œ€ํ™”(์‚ฌ์šฉ์ž์™€ AI์˜ ๋‘ ๊ฐ€์ง€ ์—ญํ• ์„ ๋ชจ๋‘ ์ˆ˜ํ–‰)๊ฐ€ ๋Œ€ํ™” ํ˜•์‹์˜ InstructGPT ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๊ฒฐํ•ฉ๋˜์–ด ChatGPT๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ChatGPT๋Š” ๋ฐฉ๋Œ€ํ•œ ์ง€์‹ ์ €์žฅ์†Œ, ์ˆ˜ํ•™ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์ถ”๋ก  ๊ธฐ์ˆ , ๋‹ค์ค‘ ํ„ด ๋Œ€ํ™”์—์„œ ๋ฌธ๋งฅ์„ ์ •ํ™•ํ•˜๊ฒŒ ์ถ”์ ํ•˜๊ณ , ์•ˆ์ „ํ•œ ์‚ฌ์šฉ์„ ์œ„ํ•œ ์ธ๊ฐ„์˜ ๊ฐ€์น˜๊ด€๊ณผ ์ž˜ ๋ถ€ํ•ฉํ•˜๋Š” ๋“ฑ ์ธ๊ฐ„๊ณผ ์†Œํ†ตํ•˜๋Š” ๋›ฐ์–ด๋‚œ ์—ญ๋Ÿ‰์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ํ”Œ๋Ÿฌ๊ทธ์ธ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ChatGPT์— ์ง€์›๋˜์–ด ๊ธฐ์กด ๋„๊ตฌ๋‚˜ ์•ฑ์œผ๋กœ ChatGPT์˜ ์—ญ๋Ÿ‰์„ ๋”์šฑ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์ง€๊ธˆ๊นŒ์ง€ AI ์—ญ์‚ฌ์ƒ ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ์ฑ—๋ด‡์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ChatGPT์˜ ์ถœ์‹œ๋Š” ์ธ๊ฐ„๊ณผ ๊ฐ™์€ AI ์‹œ์Šคํ…œ์˜ ํƒ๊ตฌ์— ๋น›์„ ๋น„์ถ”๋Š” ๋ฏธ๋ž˜์˜ AI ์—ฐ๊ตฌ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์ณค์Šต๋‹ˆ๋‹ค.
Why did we make a taxonomy?A taxonomy provides a way of grouping individual items into broader categories, often with a hierarchical structure [32]. In our case, a taxonomy lets us group individual hazards (i.e., a single source or situation with a potential for harm, such as a model providing unsafe advice) into overarching hazard categories. This lets us systematically explore and analyze hazards, provide interpretable insights, and communicate effectively about them. In keeping with best practices, we have clearly defined each category, and sought to make the categories mutually exclusive. We have also fully documented our approach so that our methodology, assumptions, and limitations are available for scrutiny. We created a new taxonomy of hazards given that existing taxonomies do not fully reflect the scope and design process of the AI Safety Benchmark, and they have various gaps and limitations that make them unsuitable. We are aware that the fast-changing nature of AI safety means that the taxonomy will need to be frequently updated to remain relevant and useful. We are working with partners to encourage taxonomic standardization where it is useful and appropriate (see Appendix A).
์–ด์งธ์„œ ๋ถ„๋ฅ˜๋ฒ•์„ ๋งŒ๋“ค์—ˆ๋‚˜์š”? ๋ถ„๋ฅ˜๋ฒ•์€ ๊ฐœ๋ณ„ ํ•ญ๋ชฉ์„ ๋” ๋„“์€ ๋ฒ”์ฃผ๋กœ ๊ทธ๋ฃนํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ๊ณตํ•˜๋ฉฐ, ์ข…์ข… ๊ณ„์ธต ๊ตฌ์กฐ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค[32]. ์ด ๊ฒฝ์šฐ ๋ถ„๋ฅ˜๋ฒ•์„ ์‚ฌ์šฉํ•˜๋ฉด ๊ฐœ๋ณ„ ์œ„ํ—˜(์˜ˆ: ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ์กฐ์–ธ์„ ์ œ๊ณตํ•˜๋Š” ๋ชจ๋ธ ๋“ฑ ์ž ์žฌ์  ์œ„ํ•ด ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ๋Š” ๋‹จ์ผ ์†Œ์Šค ๋˜๋Š” ์ƒํ™ฉ)์„ ํฌ๊ด„์ ์ธ ์œ„ํ—˜ ๋ฒ”์ฃผ๋กœ ๊ทธ๋ฃนํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์œ„ํ—˜์„ ์ฒด๊ณ„์ ์œผ๋กœ ํƒ์ƒ‰ ๋ฐ ๋ถ„์„ํ•˜๊ณ , ํ•ด์„ ๊ฐ€๋Šฅํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์œ„ํ—˜์— ๋Œ€ํ•ด ํšจ๊ณผ์ ์œผ๋กœ ์†Œํ†ตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ฒ” ์‚ฌ๋ก€์— ๋”ฐ๋ผ ๊ฐ ๋ฒ”์ฃผ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ์ •์˜ํ•˜๊ณ  ๋ฒ”์ฃผ๊ฐ€ ์ƒํ˜ธ ๋ฐฐํƒ€์ ์ด ๋˜๋„๋ก ๋…ธ๋ ฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฐฉ๋ฒ•๋ก , ๊ฐ€์ • ๋ฐ ํ•œ๊ณ„๋ฅผ ๋ฉด๋ฐ€ํžˆ ๊ฒ€ํ† ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ ‘๊ทผ ๋ฐฉ์‹์„ ์™„์ „ํžˆ ๋ฌธ์„œํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด ๋ถ„๋ฅ˜๋ฒ•์ด AI Safety Benchmark์˜ ๋ฒ”์œ„์™€ ์„ค๊ณ„ ํ”„๋กœ์„ธ์Šค๋ฅผ ์™„์ „ํžˆ ๋ฐ˜์˜ํ•˜์ง€ ๋ชปํ•˜๊ณ , ๋‹ค์–‘ํ•œ ๊ณต๋ฐฑ๊ณผ ํ•œ๊ณ„๊ฐ€ ์žˆ์–ด ๋ถ€์ ํ•ฉํ•˜๋‹ค๋Š” ์ ์„ ๊ฐ์•ˆํ•˜์—ฌ ์ƒˆ๋กœ์šด ์œ„ํ—˜ ๋ถ„๋ฅ˜๋ฒ•์„ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. AI ์•ˆ์ „์˜ ๋น ๋ฅด๊ฒŒ ๋ณ€ํ™”ํ•˜๋Š” ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ๋ถ„๋ฅ˜๋ฒ•์ด ๊ด€๋ จ์„ฑ๊ณผ ์œ ์šฉ์„ฑ์„ ์œ ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ์ž์ฃผ ์—…๋ฐ์ดํŠธ๋˜์–ด์•ผ ํ•œ๋‹ค๋Š” ์ ์„ ์ž˜ ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์œ ์šฉํ•˜๊ณ  ์ ์ ˆํ•œ ๊ฒฝ์šฐ ๋ถ„๋ฅ˜ ํ‘œ์ค€ํ™”๋ฅผ ์žฅ๋ คํ•˜๊ธฐ ์œ„ํ•ด ํŒŒํŠธ๋„ˆ์™€ ํ˜‘๋ ฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(๋ถ€๋ก A ์ฐธ์กฐ).
์–ด์งธ์„œ ๋ถ„๋ฅ˜๋ฒ•์„ ๋งŒ๋“ค์—ˆ๋‚˜์š”?๋ถ„๋ฅ˜๋ฒ•์€ ๊ฐœ๋ณ„ ํ•ญ๋ชฉ์„ ๋” ๋„“์€ ๋ฒ”์ฃผ๋กœ ๊ทธ๋ฃนํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ๊ณตํ•˜๋ฉฐ, ์ข…์ข… ๊ณ„์ธต ๊ตฌ์กฐ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค[32]. ์ด ๊ฒฝ์šฐ ๋ถ„๋ฅ˜๋ฒ•์„ ์‚ฌ์šฉํ•˜๋ฉด ๊ฐœ๋ณ„ ์œ„ํ—˜(์˜ˆ: ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ์กฐ์–ธ์„ ์ œ๊ณตํ•˜๋Š” ๋ชจ๋ธ ๋“ฑ ์ž ์žฌ์  ์œ„ํ•ด ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ๋Š” ๋‹จ์ผ ์†Œ์Šค ๋˜๋Š” ์ƒํ™ฉ)์„ ํฌ๊ด„์ ์ธ ์œ„ํ—˜ ๋ฒ”์ฃผ๋กœ ๊ทธ๋ฃนํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์œ„ํ—˜์„ ์ฒด๊ณ„์ ์œผ๋กœ ํƒ์ƒ‰ ๋ฐ ๋ถ„์„ํ•˜๊ณ , ํ•ด์„ ๊ฐ€๋Šฅํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์œ„ํ—˜์— ๋Œ€ํ•ด ํšจ๊ณผ์ ์œผ๋กœ ์†Œํ†ตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ฒ” ์‚ฌ๋ก€์— ๋”ฐ๋ผ ๊ฐ ๋ฒ”์ฃผ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ์ •์˜ํ•˜๊ณ  ๋ฒ”์ฃผ๊ฐ€ ์ƒํ˜ธ ๋ฐฐํƒ€์ ์ด ๋˜๋„๋ก ๋…ธ๋ ฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฐฉ๋ฒ•๋ก , ๊ฐ€์ • ๋ฐ ํ•œ๊ณ„๋ฅผ ๋ฉด๋ฐ€ํžˆ ๊ฒ€ํ† ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ ‘๊ทผ ๋ฐฉ์‹์„ ์™„์ „ํžˆ ๋ฌธ์„œํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด ๋ถ„๋ฅ˜๋ฒ•์ด AI ์•ˆ์ „ ๋ฒค์น˜๋งˆํฌ์˜ ๋ฒ”์œ„์™€ ์„ค๊ณ„ ํ”„๋กœ์„ธ์Šค๋ฅผ ์™„์ „ํžˆ ๋ฐ˜์˜ํ•˜์ง€ ๋ชปํ•˜๊ณ , ๋‹ค์–‘ํ•œ ๊ณต๋ฐฑ๊ณผ ํ•œ๊ณ„๊ฐ€ ์žˆ์–ด ๋ถ€์ ํ•ฉํ•˜๋‹ค๋Š” ์ ์„ ๊ฐ์•ˆํ•˜์—ฌ ์ƒˆ๋กœ์šด ์œ„ํ—˜ ๋ถ„๋ฅ˜๋ฒ•์„ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. AI ์•ˆ์ „์˜ ๋น ๋ฅด๊ฒŒ ๋ณ€ํ™”ํ•˜๋Š” ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ๋ถ„๋ฅ˜๋ฒ•์ด ๊ด€๋ จ์„ฑ๊ณผ ์œ ์šฉ์„ฑ์„ ์œ ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ์ž์ฃผ ์—…๋ฐ์ดํŠธ๋˜์–ด์•ผ ํ•œ๋‹ค๋Š” ์ ์„ ์ž˜ ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์œ ์šฉํ•˜๊ณ  ์ ์ ˆํ•œ ๊ฒฝ์šฐ ๋ถ„๋ฅ˜ ํ‘œ์ค€ํ™”๋ฅผ ์žฅ๋ คํ•˜๊ธฐ ์œ„ํ•ด ํŒŒํŠธ๋„ˆ์™€ ํ˜‘๋ ฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(๋ถ€๋ก A ์ฐธ์กฐ).
Comparing our model's performance on both English and Chinese benchmarks with other models reveals a notably smaller gap in performance across multi-disciplinary datasets such as MMLU and CMMLU, as shown in Table 2. While other models exhibit significant disparities, particularly in language understanding and reasoning benchmarks, our model maintains a consistent performance, suggesting a balanced capability across diverse domains. This contrasts with other models that show pronounced variability, such as in the HellaSwag dataset, where our model closely rivals or outperforms alternatives like MiniCPM (min, 2024) and Phi-2, showcasing superior or competitive reasoning abilities. Similarly, in domain-specific evaluations (C-Eval and CMMLU), our model demonstrates commonable performance, outpacing models like TinyLlama-1.1B and Bloom-1.7B in comprehending and generating content that requires a nuanced understanding of cultural and domain-specific contexts. This balanced proficiency underscores the model's versatility and adaptability, positioning it as a strong contender in the landscape of AI language models, with a capacity for both broad applicability and deep, domain-specific knowledge.
ํ‘œ 2์™€ ๊ฐ™์ด ์˜์–ด ๋ฐ ์ค‘๊ตญ์–ด ๋ฒค์น˜๋งˆํฌ ๋ชจ๋‘์—์„œ ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ, MMLU ๋ฐ CMMLU์™€ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์„ฑ๋Šฅ ์ฐจ์ด๊ฐ€ ํ˜„์ €ํžˆ ์ž‘๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ชจ๋ธ๋“ค์€ ํŠนํžˆ language understanding ๋ฐ reasoning ๋ฒค์น˜๋งˆํฌ์—์„œ ์ƒ๋‹นํ•œ ์ฐจ์ด๋ฅผ ๋ณด์ด๋Š” ๋ฐ˜๋ฉด, ์ €ํฌ ๋ชจ๋ธ์€ ์ผ๊ด€๋œ ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์˜์—ญ์—์„œ ๊ท ํ˜• ์žกํžŒ ์—ญ๋Ÿ‰์„ ๊ฐ–์ถ”๊ณ  ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” HellaSwag ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ์ฒ˜๋Ÿผ ์šฐ๋ฆฌ ๋ชจ๋ธ์ด MiniCPM(์ตœ์†Œ, 2024) ๋ฐ Phi-2์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ๋ชจ๋ธ์— ๊ทผ์ ‘ํ•˜๊ฑฐ๋‚˜ ๋Šฅ๊ฐ€ํ•˜๋Š” reasoning ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋Š” ๋“ฑ ๋šœ๋ ทํ•œ ๋ณ€๋™์„ฑ์„ ๋ณด์ด๋Š” ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋Œ€์กฐ๋ฅผ ์ด๋ฃน๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๋„๋ฉ”์ธ๋ณ„ ํ‰๊ฐ€(C-Eval ๋ฐ CMMLU)์—์„œ๋„ ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ๋ฌธํ™”์ , ๋„๋ฉ”์ธ๋ณ„ ๋งฅ๋ฝ์— ๋Œ€ํ•œ ๋ฏธ๋ฌ˜ํ•œ ์ดํ•ด๊ฐ€ ํ•„์š”ํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์žˆ์–ด TinyLlama-1.1B ๋ฐ Bloom-1.7B๊ณผ ๊ฐ™์€ ๋ชจ๋ธ์„ ์•ž์ง€๋ฅด๋Š” ๊ณตํ†ต์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ท ํ˜• ์žกํžŒ ์ˆ™๋ จ๋„๋Š” ๋ชจ๋ธ์˜ ๋‹ค์šฉ๋„์„ฑ๊ณผ ์ ์‘๋ ฅ์„ ๊ฐ•์กฐํ•˜๋ฉฐ, ๊ด‘๋ฒ”์œ„ํ•œ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ๊ณผ ์‹ฌ๋„ ์žˆ๋Š” ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ๋ชจ๋‘ ๊ฐ–์ถ˜ AI ์–ธ์–ด ๋ชจ๋ธ์˜ ์˜์—ญ์—์„œ ๊ฐ•๋ ฅํ•œ ๊ฒฝ์Ÿ์ž๋กœ ์ž๋ฆฌ๋งค๊น€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
ํ‘œ 2์™€ ๊ฐ™์ด ์˜์–ด ๋ฐ ์ค‘๊ตญ์–ด ๋ฒค์น˜๋งˆํฌ ๋ชจ๋‘์—์„œ ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ, MMLU ๋ฐ CMMLU์™€ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์„ฑ๋Šฅ ์ฐจ์ด๊ฐ€ ํ˜„์ €ํžˆ ์ž‘๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ชจ๋ธ๋“ค์€ ํŠนํžˆ ์–ธ์–ด ์ดํ•ด ๋ฐ ์ถ”๋ก  ๋ฒค์น˜๋งˆํฌ์—์„œ ์ƒ๋‹นํ•œ ์ฐจ์ด๋ฅผ ๋ณด์ด๋Š” ๋ฐ˜๋ฉด, ์ €ํฌ ๋ชจ๋ธ์€ ์ผ๊ด€๋œ ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์˜์—ญ์—์„œ ๊ท ํ˜• ์žกํžŒ ์—ญ๋Ÿ‰์„ ๊ฐ–์ถ”๊ณ  ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ—ฌ๋ผ์Šค์›ฉ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ์ฒ˜๋Ÿผ ์šฐ๋ฆฌ ๋ชจ๋ธ์ด ๋ฏธ๋‹ˆCPM(์ตœ์†Œ, 2024) ๋ฐ Phi-2์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ๋ชจ๋ธ์— ๊ทผ์ ‘ํ•˜๊ฑฐ๋‚˜ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋Š” ๋“ฑ ๋šœ๋ ทํ•œ ๋ณ€๋™์„ฑ์„ ๋ณด์ด๋Š” ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋Œ€์กฐ๋ฅผ ์ด๋ฃน๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๋„๋ฉ”์ธ๋ณ„ ํ‰๊ฐ€(C-Eval ๋ฐ CMMLU)์—์„œ๋„ ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ๋ฌธํ™”์ , ๋„๋ฉ”์ธ๋ณ„ ๋งฅ๋ฝ์— ๋Œ€ํ•œ ๋ฏธ๋ฌ˜ํ•œ ์ดํ•ด๊ฐ€ ํ•„์š”ํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์žˆ์–ด TinyLlama-1.1B ๋ฐ Bloom-1.7B๊ณผ ๊ฐ™์€ ๋ชจ๋ธ์„ ์•ž์ง€๋ฅด๋Š” ๊ณตํ†ต์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ท ํ˜• ์žกํžŒ ์ˆ™๋ จ๋„๋Š” ๋ชจ๋ธ์˜ ๋‹ค์šฉ๋„์„ฑ๊ณผ ์ ์‘๋ ฅ์„ ๊ฐ•์กฐํ•˜๋ฉฐ, ๊ด‘๋ฒ”์œ„ํ•œ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ๊ณผ ์‹ฌ๋„ ์žˆ๋Š” ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ๋ชจ๋‘ ๊ฐ–์ถ˜ AI ์–ธ์–ด ๋ชจ๋ธ์˜ ์˜์—ญ์—์„œ ๊ฐ•๋ ฅํ•œ ๊ฒฝ์Ÿ์ž๋กœ ์ž๋ฆฌ๋งค๊น€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Referring to mainstream frameworks such as TransReID [255] and Strong Baseline [264], we retained ID Loss, Triplet Loss, and BN Layer, and replaced the CNN and Transformer backbones using Vim [61] and VAMba [60] to explore the potential of Mamba for re-identification tasks, and the compared results are shown in Table XIII. The selective scanning mechanism (SSM) proposed by the Mamba model allows for sequence modeling with low complexity, and Vim and VAMba further build on it by proposing an SSM modeling approach for 2D image data. Compared to CNN-based models that require complex module design, the simple Mamba network already has effectiveness. Even compared with the models with high complexity such as DeiT [265] and ViT [19], the bidirectional scanning mechanism proposed by Vim has fewer training parameters, and it shows effectiveness on the VehicleID dataset. In contrast, VAMba's cross-scanning mechanism, which does not rely on Transformer's structure (_e.g._, position embedding and class token), has achieved comparable results on the Market1501, DukeMTMC, and VeRi-776 datasets. For this reason, we expect more Mamba-based studies applicable to the re-identification task in the future.
TransReID [255] ๋ฐ Strong Baseline [264]๊ณผ ๊ฐ™์€ ์ฃผ๋ฅ˜ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ID Loss, Triplet Loss, BN Layer๋ฅผ ์œ ์ง€ํ•˜๊ณ  Vim [61] ๋ฐ VAMba [60]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ CNN ๋ฐ Transformer ๋ฐฑ๋ณธ์„ ๋Œ€์ฒดํ•˜์—ฌ ์žฌ์‹๋ณ„ ์ž‘์—…์— ๋Œ€ํ•œ Mamba์˜ ์ž ์žฌ๋ ฅ์„ ํƒ์ƒ‰ํ–ˆ์œผ๋ฉฐ, ๋น„๊ต ๊ฒฐ๊ณผ๋Š” ํ‘œ XIII์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. Mamba ๋ชจ๋ธ์—์„œ ์ œ์•ˆํ•œ ์„ ํƒ์  ์Šค์บ๋‹ ๋ฉ”์ปค๋‹ˆ์ฆ˜(SSM)์€ ๋‚ฎ์€ ๋ณต์žก๋„๋กœ ์‹œํ€€์Šค ๋ชจ๋ธ๋ง์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋ฉฐ, Vim๊ณผ VAMba๋Š” 2D ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ SSM ๋ชจ๋ธ๋ง ์ ‘๊ทผ๋ฒ•์„ ์ œ์•ˆํ•จ์œผ๋กœ์จ ์ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋ณต์žกํ•œ ๋ชจ๋“ˆ ์„ค๊ณ„๊ฐ€ ํ•„์š”ํ•œ CNN ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, ๊ฐ„๋‹จํ•œ Mamba ๋„คํŠธ์›Œํฌ๋Š” ์ด๋ฏธ ํšจ๊ณผ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. DeiT [265] ๋ฐ ViT [19]์™€ ๊ฐ™์ด ๋ณต์žก๋„๊ฐ€ ๋†’์€ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•˜๋”๋ผ๋„, Vim์—์„œ ์ œ์•ˆํ•œ ์–‘๋ฐฉํ–ฅ ์Šค์บ๋‹ ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ํ›ˆ๋ จ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ์ ๊ณ , VehicleID ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, Transformer's ๊ตฌ์กฐ(์˜ˆ: position embedding ๋ฐ class token)์— ์˜์กดํ•˜์ง€ ์•Š๋Š” VAMba์˜ ๊ต์ฐจ ์Šค์บ๋‹ ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ Market1501, DukeMTMC ๋ฐ VeRi-776 ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ด์œ ๋กœ ํ–ฅํ›„ ์žฌ์‹๋ณ„ ์ž‘์—…์— ์ ์šฉ ๊ฐ€๋Šฅํ•œ Mamba ๊ธฐ๋ฐ˜ ์—ฐ๊ตฌ๊ฐ€ ๋” ๋งŽ์ด ์ด๋ฃจ์–ด์งˆ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค.
TransReID [255] ๋ฐ Strong Baseline [264]๊ณผ ๊ฐ™์€ ์ฃผ๋ฅ˜ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ID ์†์‹ค, ํŠธ๋ฆฌํ”Œ๋ › ์†์‹ค, BN ๋ ˆ์ด์–ด๋ฅผ ์œ ์ง€ํ•˜๊ณ  Vim [61] ๋ฐ VAMba [60]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ CNN ๋ฐ ํŠธ๋žœ์Šคํฌ๋จธ ๋ฐฑ๋ณธ์„ ๋Œ€์ฒดํ•˜์—ฌ ์žฌ์‹๋ณ„ ์ž‘์—…์— ๋Œ€ํ•œ ๋ง˜๋ฐ”์˜ ์ž ์žฌ๋ ฅ์„ ํƒ์ƒ‰ํ–ˆ์œผ๋ฉฐ, ๋น„๊ต ๊ฒฐ๊ณผ๋Š” ํ‘œ XIII์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. Mamba ๋ชจ๋ธ์—์„œ ์ œ์•ˆํ•œ ์„ ํƒ์  ์Šค์บ๋‹ ๋ฉ”์ปค๋‹ˆ์ฆ˜(SSM)์€ ๋‚ฎ์€ ๋ณต์žก๋„๋กœ ์‹œํ€€์Šค ๋ชจ๋ธ๋ง์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋ฉฐ, Vim๊ณผ VAMba๋Š” 2D ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ SSM ๋ชจ๋ธ๋ง ์ ‘๊ทผ๋ฒ•์„ ์ œ์•ˆํ•จ์œผ๋กœ์จ ์ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋ณต์žกํ•œ ๋ชจ๋“ˆ ์„ค๊ณ„๊ฐ€ ํ•„์š”ํ•œ CNN ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, ๊ฐ„๋‹จํ•œ Mamba ๋„คํŠธ์›Œํฌ๋Š” ์ด๋ฏธ ํšจ๊ณผ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. DeiT [265] ๋ฐ ViT [19]์™€ ๊ฐ™์ด ๋ณต์žก๋„๊ฐ€ ๋†’์€ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•˜๋”๋ผ๋„, Vim์—์„œ ์ œ์•ˆํ•œ ์–‘๋ฐฉํ–ฅ ์Šค์บ๋‹ ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ํ›ˆ๋ จ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ์ ๊ณ , VehicleID ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ํŠธ๋žœ์Šคํฌ๋จธ์˜ ๊ตฌ์กฐ(์˜ˆ: ์œ„์น˜ ์ž„๋ฒ ๋”ฉ ๋ฐ ํด๋ž˜์Šค ํ† ํฐ)์— ์˜์กดํ•˜์ง€ ์•Š๋Š” VAMba์˜ ๊ต์ฐจ ์Šค์บ๋‹ ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ Market1501, DukeMTMC ๋ฐ VeRi-776 ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ด์œ ๋กœ ํ–ฅํ›„ ์žฌ์‹๋ณ„ ์ž‘์—…์— ์ ์šฉ ๊ฐ€๋Šฅํ•œ Mamba ๊ธฐ๋ฐ˜ ์—ฐ๊ตฌ๊ฐ€ ๋” ๋งŽ์ด ์ด๋ฃจ์–ด์งˆ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค.
* Memorizable knowledge accuracy (# of people). We apply the model to the original training data, such as "Anya Briar Forger was born on" and check if it can correctly generate "October 2, 1996". For each person, we evaluate all five attributes and compute their average accuracy.23 We then _sum_ this accuracy up over all \(N\) people. (Ideally, a perfect model would have this "accuracy" equal to \(N\).) Footnote 23: We exclude the company city attribute because it can be uniquely determined by the employer name, thus providing no additional knowledge. * Extractable knowledge accuracy (# of people). Following the pretrain-finetune framework of [3], we fine-tune any given pretrained model on half of the individuals using LoRA [17] with question-answering texts like "What is the birthday of Anya Briar Forger? Answer: October 2, 1996." We then test its generation accuracy on the remaining half of the individuals. High accuracy indicates that the knowledge is not only memorized but can also be _flexibly_ extracted for downstream tasks. Again, for each person,we evaluate all five attributes and compute their average accuracy. We then _sum_ this across all \(N/2\) people and multiply by 2.
* ๊ธฐ์–ตํ•  ์ˆ˜ ์žˆ๋Š” ์ง€์‹ ์ •ํ™•๋„(# of people). "์•„๋ƒ ๋ธŒ๋ผ์ด์–ด ํฌ์ €๋Š”"๊ณผ ๊ฐ™์€ ์›๋ณธ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ๋ชจ๋ธ์„ ์ ์šฉํ•˜์—ฌ "1996๋…„ 10์›” 2์ผ"์„ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์‚ฌ๋žŒ์— ๋Œ€ํ•ด 5๊ฐ€์ง€ ์†์„ฑ์„ ๋ชจ๋‘ ํ‰๊ฐ€ํ•˜๊ณ  ํ‰๊ท  ์ •ํ™•๋„๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.23 ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ์ •ํ™•๋„๋ฅผ ๋ชจ๋“  \(N\) ์‚ฌ๋žŒ์— ๋Œ€ํ•ด _ํ•ฉ์‚ฐ_ํ•ฉ๋‹ˆ๋‹ค. (์ด์ƒ์ ์œผ๋กœ ์™„๋ฒฝํ•œ ๋ชจ๋ธ์ด๋ผ๋ฉด ์ด '์ •ํ™•๋„'๊ฐ€ \(N\)๊ณผ ๊ฐ™์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค.) ๊ฐ์ฃผ 23: ํšŒ์‚ฌ ๋„์‹œ ์†์„ฑ์€ ๊ณ ์šฉ์ฃผ ์ด๋ฆ„์œผ๋กœ ๊ณ ์œ ํ•˜๊ฒŒ ๊ฒฐ์ •ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ถ”๊ฐ€ ์ง€์‹์„ ์ œ๊ณตํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. * ์ถ”์ถœ ๊ฐ€๋Šฅํ•œ ์ง€์‹ ์ •ํ™•๋„(# of people). [3]์˜ pretrain-finetune ํ”„๋ ˆ์ž„์›Œํฌ์— ๋”ฐ๋ผ "์•„๋ƒ ๋ธŒ๋ผ์ด์–ด ํฌ์ €์˜ ์ƒ์ผ์€ ๋ฌด์—‡์ž…๋‹ˆ๊นŒ?"์™€ ๊ฐ™์€ ์งˆ๋ฌธ-๋‹ต๋ณ€ ํ…์ŠคํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LoRA [17]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ ˆ๋ฐ˜์˜ ๊ฐœ์ธ์— ๋Œ€ํ•ด ๋ฏธ๋ฆฌ ํ•™์Šต๋œ ๋ชจ๋ธ์„ Fine-tuningํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋‚˜๋จธ์ง€ ์ ˆ๋ฐ˜์˜ ๊ฐœ์ธ์— ๋Œ€ํ•ด ์ƒ์„ฑ ์ •ํ™•๋„๋ฅผ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ์ •ํ™•๋„๊ฐ€ ๋†’์„์ˆ˜๋ก ์ง€์‹์ด ์•”๊ธฐ๋˜์–ด ์žˆ์„ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ํ•„์š”ํ•œ _์œ ์—ฐํ•˜๊ฒŒ_ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋‹ค์‹œ ๋งํ•˜์ง€๋งŒ, ๊ฐ ์‚ฌ๋žŒ์— ๋Œ€ํ•ด 5๊ฐ€์ง€ ์†์„ฑ์„ ๋ชจ๋‘ ํ‰๊ฐ€ํ•˜๊ณ  ํ‰๊ท  ์ •ํ™•๋„๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋ฅผ ๋ชจ๋“  \(N/2\) ์‚ฌ๋žŒ์— ๋Œ€ํ•ด _ํ•ฉ์‚ฐ_ํ•˜๊ณ  2๋ฅผ ๊ณฑํ•ฉ๋‹ˆ๋‹ค.
* ๊ธฐ์–ตํ•  ์ˆ˜ ์žˆ๋Š” ์ง€์‹ ์ •ํ™•๋„(# of people). "์•„๋ƒ ๋ธŒ๋ผ์ด์–ด ํฌ์ €๋Š”"๊ณผ ๊ฐ™์€ ์›๋ณธ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ๋ชจ๋ธ์„ ์ ์šฉํ•˜์—ฌ "1996๋…„ 10์›” 2์ผ"์„ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์‚ฌ๋žŒ์— ๋Œ€ํ•ด 5๊ฐ€์ง€ ์†์„ฑ์„ ๋ชจ๋‘ ํ‰๊ฐ€ํ•˜๊ณ  ํ‰๊ท  ์ •ํ™•๋„๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.23 ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ์ •ํ™•๋„๋ฅผ ๋ชจ๋“  \(N\) ์‚ฌ๋žŒ์— ๋Œ€ํ•ด _ํ•ฉ์‚ฐ_ํ•ฉ๋‹ˆ๋‹ค. (์ด์ƒ์ ์œผ๋กœ ์™„๋ฒฝํ•œ ๋ชจ๋ธ์ด๋ผ๋ฉด ์ด '์ •ํ™•๋„'๊ฐ€ \(N\)๊ณผ ๊ฐ™์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค.) ๊ฐ์ฃผ 23: ํšŒ์‚ฌ ๋„์‹œ ์†์„ฑ์€ ๊ณ ์šฉ์ฃผ ์ด๋ฆ„์œผ๋กœ ๊ณ ์œ ํ•˜๊ฒŒ ๊ฒฐ์ •ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ถ”๊ฐ€ ์ง€์‹์„ ์ œ๊ณตํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. * ์ถ”์ถœ ๊ฐ€๋Šฅํ•œ ์ง€์‹ ์ •ํ™•๋„(# of people). [3]์˜ ์‚ฌ์ „ ํ•™์Šต-๋ฏธ์„ธ ์กฐ์ • ํ”„๋ ˆ์ž„์›Œํฌ์— ๋”ฐ๋ผ "์•„๋ƒ ๋ธŒ๋ผ์ด์–ด ํฌ์ €์˜ ์ƒ์ผ์€ ๋ฌด์—‡์ž…๋‹ˆ๊นŒ?"์™€ ๊ฐ™์€ ์งˆ๋ฌธ-๋‹ต๋ณ€ ํ…์ŠคํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LoRA [17]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ ˆ๋ฐ˜์˜ ๊ฐœ์ธ์— ๋Œ€ํ•ด ๋ฏธ๋ฆฌ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋‚˜๋จธ์ง€ ์ ˆ๋ฐ˜์˜ ๊ฐœ์ธ์— ๋Œ€ํ•ด ์ƒ์„ฑ ์ •ํ™•๋„๋ฅผ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ์ •ํ™•๋„๊ฐ€ ๋†’์„์ˆ˜๋ก ์ง€์‹์ด ์•”๊ธฐ๋˜์–ด ์žˆ์„ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ํ•„์š”ํ•œ _์œ ์—ฐํ•˜๊ฒŒ_ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋‹ค์‹œ ๋งํ•˜์ง€๋งŒ, ๊ฐ ์‚ฌ๋žŒ์— ๋Œ€ํ•ด 5๊ฐ€์ง€ ์†์„ฑ์„ ๋ชจ๋‘ ํ‰๊ฐ€ํ•˜๊ณ  ํ‰๊ท  ์ •ํ™•๋„๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋ฅผ ๋ชจ๋“  \(N/2\) ์‚ฌ๋žŒ์— ๋Œ€ํ•ด _ํ•ฉ์‚ฐ_ํ•˜๊ณ  2๋ฅผ ๊ณฑํ•ฉ๋‹ˆ๋‹ค.
In terms of visual and linguistic information fusion, Zhao et al. [147] optimizes the internal information integration of the Mamba language model to achieve a more effective expression. Decision Mamba(DMamba) [148] integrates the Mamba framework into the Decision Transformer (DT) [152]. A series of experiments comparing Decision Mamba and DT show that Mamba is effective in reinforcement learning (RL) tasks. But simply applying Mamba blocks to DT does not improve efficiency, because the RL tasks considered by the authors have a large number of CPU and GPU interactions. Another deficiency is the absence of a hyper-parameter search and an analysis of how to use the Mamba block more effectively to reflect the data structure of RL tasks. Sigma [149] is the first State Space Model successfully applied in multi-modal semantic segmentation. It is composed of VMamba, an attention-based Mamba fusion mechanism and a channel-aware Mamba decoder, and has shown excellent performance in various experiments. However, Sigma is underutilized in handling longer sequences, and the memory consumption of Mamba encoders is still relatively large, making it difficult to deploy on lightweight edge devices.
์‹œ๊ฐ์  ๋ฐ ์–ธ์–ด์  ์ •๋ณด ์œตํ•ฉ ์ธก๋ฉด์—์„œ Zhao ๋“ฑ[147]์€ Mamba ์–ธ์–ด ๋ชจ๋ธ์˜ ๋‚ด๋ถ€ ์ •๋ณด ํ†ตํ•ฉ์„ ์ตœ์ ํ™”ํ•˜์—ฌ ๋ณด๋‹ค ํšจ๊ณผ์ ์ธ ํ‘œํ˜„์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. Decision Mamba(DMamba) [148]๋Š” Mamba ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ Decision Transformer (DT) [152]์— ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. Decision Mamba์™€ DT๋ฅผ ๋น„๊ตํ•œ ์ผ๋ จ์˜ ์‹คํ—˜์„ ํ†ตํ•ด Mamba๊ฐ€ Reinforcement Learning (RL) ์ž‘์—…์— ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ €์ž๋“ค์ด ๊ณ ๋ คํ•œ RL ์ž‘์—…์—๋Š” CPU์™€ GPU์˜ ์ƒํ˜ธ ์ž‘์šฉ์ด ๋งŽ๊ธฐ ๋•Œ๋ฌธ์— Mamba ๋ธ”๋ก์„ DT์— ๋‹จ์ˆœํžˆ ์ ์šฉํ•ด๋„ ํšจ์œจ์„ฑ์ด ํ–ฅ์ƒ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ๊ฒฐํ•์ ์€ ํ•˜์ดํผ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ฒ€์ƒ‰๊ณผ RL ์ž‘์—…์˜ ๋ฐ์ดํ„ฐ ๊ตฌ์กฐ๋ฅผ ๋” ํšจ๊ณผ์ ์œผ๋กœ ๋ฐ˜์˜ํ•˜๊ธฐ ์œ„ํ•ด Mamba ๋ธ”๋ก์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ๋ถ„์„์ด ๋ถ€์žฌํ•œ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. Sigma [149]๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์‹œ๋งจํ‹ฑ ์„ธ๋ถ„ํ™”์— ์„ฑ๊ณต์ ์œผ๋กœ ์ ์šฉ๋œ ์ตœ์ดˆ์˜ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ Attention ๊ธฐ๋ฐ˜ Mamba ์œตํ•ฉ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ธ VMamba์™€ ์ฑ„๋„ ์ธ์‹ Mamba Decoder๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋‹ค์–‘ํ•œ ์‹คํ—˜์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Sigma๋Š” ๋” ๊ธด ์‹œํ€€์Šค๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ํ™œ์šฉ๋„๊ฐ€ ๋‚ฎ์œผ๋ฉฐ, Mamba Encoder์˜ ๋ฉ”๋ชจ๋ฆฌ ์†Œ๋น„๊ฐ€ ์ƒ๋Œ€์ ์œผ๋กœ ๋งŽ์•„ ๊ฒฝ๋Ÿ‰ ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค์— ๋ฐฐํฌํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค.
์‹œ๊ฐ์  ๋ฐ ์–ธ์–ด์  ์ •๋ณด ์œตํ•ฉ ์ธก๋ฉด์—์„œ Zhao ๋“ฑ[147]์€ ๋ง˜๋ฐ” ์–ธ์–ด ๋ชจ๋ธ์˜ ๋‚ด๋ถ€ ์ •๋ณด ํ†ตํ•ฉ์„ ์ตœ์ ํ™”ํ•˜์—ฌ ๋ณด๋‹ค ํšจ๊ณผ์ ์ธ ํ‘œํ˜„์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์˜์‚ฌ ๊ฒฐ์ • ๋ง˜๋ฐ”(DMamba) [148]๋Š” ์˜์‚ฌ ๊ฒฐ์ • ํŠธ๋žœ์Šคํฌ๋จธ(DT) [152]์— ๋ง˜๋ฐ” ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์˜์‚ฌ ๊ฒฐ์ • ๋ง˜๋ฐ”์™€ DT๋ฅผ ๋น„๊ตํ•œ ์ผ๋ จ์˜ ์‹คํ—˜์„ ํ†ตํ•ด ๋ง˜๋ฐ”๊ฐ€ ๊ฐ•ํ™” ํ•™์Šต(RL) ์ž‘์—…์— ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ €์ž๋“ค์ด ๊ณ ๋ คํ•œ RL ์ž‘์—…์—๋Š” CPU์™€ GPU์˜ ์ƒํ˜ธ ์ž‘์šฉ์ด ๋งŽ๊ธฐ ๋•Œ๋ฌธ์— Mamba ๋ธ”๋ก์„ DT์— ๋‹จ์ˆœํžˆ ์ ์šฉํ•ด๋„ ํšจ์œจ์„ฑ์ด ํ–ฅ์ƒ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ๊ฒฐํ•์ ์€ ํ•˜์ดํผ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ฒ€์ƒ‰๊ณผ RL ์ž‘์—…์˜ ๋ฐ์ดํ„ฐ ๊ตฌ์กฐ๋ฅผ ๋” ํšจ๊ณผ์ ์œผ๋กœ ๋ฐ˜์˜ํ•˜๊ธฐ ์œ„ํ•ด Mamba ๋ธ”๋ก์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ๋ถ„์„์ด ๋ถ€์žฌํ•œ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์‹œ๊ทธ๋งˆ [149]๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์‹œ๋งจํ‹ฑ ์„ธ๋ถ„ํ™”์— ์„ฑ๊ณต์ ์œผ๋กœ ์ ์šฉ๋œ ์ตœ์ดˆ์˜ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์ฃผ์˜๋ ฅ ๊ธฐ๋ฐ˜ ๋ง˜๋ฐ” ์œตํ•ฉ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ธ VMamba์™€ ์ฑ„๋„ ์ธ์‹ ๋ง˜๋ฐ” ๋””์ฝ”๋”๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋‹ค์–‘ํ•œ ์‹คํ—˜์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹œ๊ทธ๋งˆ๋Š” ๋” ๊ธด ์‹œํ€€์Šค๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ํ™œ์šฉ๋„๊ฐ€ ๋‚ฎ์œผ๋ฉฐ, ๋ง˜๋ฐ” ์ธ์ฝ”๋”์˜ ๋ฉ”๋ชจ๋ฆฌ ์†Œ๋น„๊ฐ€ ์ƒ๋Œ€์ ์œผ๋กœ ๋งŽ์•„ ๊ฒฝ๋Ÿ‰ ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค์— ๋ฐฐํฌํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค.
(1) First, the crowd-sourced worker is instructed to classify whether the given article contains _new_ information or not. (We define _new_ as not likely to be known before May 2020). If the article contains new information, the worker is instructed to select a sentence from the article that contains the most _recent_ information and an _entity_ among the possible answer candidates in the sentence and discard the article if otherwise. We provide the possible entities through a Named-Entity Recognition Model. (2) We make the selected sentence _stand-alone_ from the article through the decontextualization model provided by Choi et al. (2021). (3) The decontextualized sentence is paraphrased by a back-translation model (en\(\rightarrow\)de\(\rightarrow\)en) (Tiedemann and Thottingal, 2020) and checked whether the selected word is still in the paraphrased sentence; the sentence is discarded if not. (4) Next, we mask out the selected word from the sentence and ask two crowd-sourced workers to convert the cloze sentence into a question and answer the question. (5) If the answers agree among the workers as well as correspond to the actual selected word, we add the instance to NewLAMA-Easy.
(1) ๋จผ์ €, crowd-sourced ์ž‘์—…์ž์—๊ฒŒ ์ฃผ์–ด์ง„ ๊ธฐ์‚ฌ์— _์ƒˆ๋กœ์šด_ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๋ถ„๋ฅ˜ํ•˜๋„๋ก ์ง€์‹œํ•ฉ๋‹ˆ๋‹ค. (์ƒˆ๋กœ์šด ์ •๋ณด_๋Š” 2020๋…„ 5์›” ์ด์ „์—๋Š” ์•Œ ์ˆ˜ ์—†์„ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ์ •๋ณด๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค). ๊ธฐ์‚ฌ์— ์ƒˆ๋กœ์šด ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š” ๊ฒฝ์šฐ, ์ž‘์—…์ž๋Š” ๊ธฐ์‚ฌ์—์„œ ๊ฐ€์žฅ _์ตœ๊ทผ_ ์ •๋ณด๋ฅผ ํฌํ•จํ•˜๋Š” ๋ฌธ์žฅ๊ณผ ๋ฌธ์žฅ์˜ ๊ฐ€๋Šฅํ•œ ๋‹ต์•ˆ ํ›„๋ณด ์ค‘ _entity_๋ฅผ ์„ ํƒํ•˜๊ณ  ๊ทธ๋ ‡์ง€ ์•Š์€ ๊ฒฝ์šฐ ๊ธฐ์‚ฌ๋ฅผ ํ๊ธฐํ•˜๋„๋ก ์ง€์‹œํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€๋Šฅํ•œ entity๋Š” Named-Entity Recognition Model์„ ํ†ตํ•ด ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. (2) ์„ ํƒํ•œ ๋ฌธ์žฅ์„ Choi ๋“ฑ(2021)์ด ์ œ๊ณตํ•œ decontextualization model์„ ํ†ตํ•ด ๊ธฐ์‚ฌ์—์„œ _๋…๋ฆฝ์ ์œผ๋กœ_ ๋งŒ๋“ญ๋‹ˆ๋‹ค. (3) decontextualized ๋ฌธ์žฅ์„ back-translation model (en\(\rightarrow\)de\(\rightarrow\)en)๋กœ ์˜์—ญํ•˜๊ณ (Tiedemann and Thottingal, 2020) ์„ ํƒํ•œ ๋‹จ์–ด๊ฐ€ ์˜์—ญ๋œ ๋ฌธ์žฅ์— ์—ฌ์ „ํžˆ ์žˆ๋Š”์ง€ ํ™•์ธํ•˜๊ณ , ๊ทธ๋ ‡์ง€ ์•Š์€ ๊ฒฝ์šฐ ๋ฌธ์žฅ์„ ์‚ญ์ œํ•ฉ๋‹ˆ๋‹ค. (4) ๋‹ค์Œ์œผ๋กœ, ๋ฌธ์žฅ์—์„œ ์„ ํƒํ•œ ๋‹จ์–ด๋ฅผ ๋งˆ์Šคํ‚นํ•˜๊ณ  ๋‘ ๋ช…์˜ crowd-sourced ์ž‘์—…์ž์—๊ฒŒ cloze ๋ฌธ์žฅ์„ ์งˆ๋ฌธ์œผ๋กœ ๋ณ€ํ™˜ํ•˜๊ณ  ์งˆ๋ฌธ์— ๋‹ตํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. (5) ์ž‘์—…์ž๋“ค ์‚ฌ์ด์—์„œ ๋‹ต๋ณ€์ด ์ผ์น˜ํ•˜๊ณ  ์‹ค์ œ ์„ ํƒํ•œ ๋‹จ์–ด์™€ ์ผ์น˜ํ•˜๋Š” ๊ฒฝ์šฐ, ํ•ด๋‹น ์ธ์Šคํ„ด์Šค๋ฅผ NewLAMA-Easy์— ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
(1) ๋จผ์ €, ํฌ๋ผ์šฐ๋“œ ์†Œ์‹ฑ ์ž‘์—…์ž์—๊ฒŒ ์ฃผ์–ด์ง„ ๊ธฐ์‚ฌ์— _์ƒˆ๋กœ์šด_ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๋ถ„๋ฅ˜ํ•˜๋„๋ก ์ง€์‹œํ•ฉ๋‹ˆ๋‹ค. (์ƒˆ๋กœ์šด ์ •๋ณด_๋Š” 2020๋…„ 5์›” ์ด์ „์—๋Š” ์•Œ ์ˆ˜ ์—†์„ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ์ •๋ณด๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค). ๊ธฐ์‚ฌ์— ์ƒˆ๋กœ์šด ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š” ๊ฒฝ์šฐ, ์ž‘์—…์ž๋Š” ๊ธฐ์‚ฌ์—์„œ ๊ฐ€์žฅ _์ตœ๊ทผ_ ์ •๋ณด๋ฅผ ํฌํ•จํ•˜๋Š” ๋ฌธ์žฅ๊ณผ ๋ฌธ์žฅ์˜ ๊ฐ€๋Šฅํ•œ ๋‹ต์•ˆ ํ›„๋ณด ์ค‘ _์—”ํ‹ฐํ‹ฐ_๋ฅผ ์„ ํƒํ•˜๊ณ  ๊ทธ๋ ‡์ง€ ์•Š์€ ๊ฒฝ์šฐ ๊ธฐ์‚ฌ๋ฅผ ํ๊ธฐํ•˜๋„๋ก ์ง€์‹œํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€๋Šฅํ•œ ์—”ํ‹ฐํ‹ฐ๋Š” ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ ์ธ์‹ ๋ชจ๋ธ์„ ํ†ตํ•ด ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. (2) ์„ ํƒํ•œ ๋ฌธ์žฅ์„ Choi ๋“ฑ(2021)์ด ์ œ๊ณตํ•œ ๋งฅ๋ฝ ํ•ด์ œ ๋ชจ๋ธ์„ ํ†ตํ•ด ๊ธฐ์‚ฌ์—์„œ _๋…๋ฆฝ์ ์œผ๋กœ_ ๋งŒ๋“ญ๋‹ˆ๋‹ค. (3) ํƒˆ๋งฅ๋ฝํ™”๋œ ๋ฌธ์žฅ์„ ์—ญ๋ฒˆ์—ญ ๋ชจ๋ธ(en\(\rightarrow\)de\(\rightarrow\)en)๋กœ ์˜์—ญํ•˜๊ณ (Tiedemann and Thottingal, 2020) ์„ ํƒํ•œ ๋‹จ์–ด๊ฐ€ ์˜์—ญ๋œ ๋ฌธ์žฅ์— ์—ฌ์ „ํžˆ ์žˆ๋Š”์ง€ ํ™•์ธํ•˜๊ณ , ๊ทธ๋ ‡์ง€ ์•Š์€ ๊ฒฝ์šฐ ๋ฌธ์žฅ์„ ์‚ญ์ œํ•ฉ๋‹ˆ๋‹ค. (4) ๋‹ค์Œ์œผ๋กœ, ๋ฌธ์žฅ์—์„œ ์„ ํƒํ•œ ๋‹จ์–ด๋ฅผ ๋งˆ์Šคํ‚นํ•˜๊ณ  ๋‘ ๋ช…์˜ ํฌ๋ผ์šฐ๋“œ ์†Œ์‹ฑ ์ž‘์—…์ž์—๊ฒŒ ํด๋กœ์ฆˆ ๋ฌธ์žฅ์„ ์งˆ๋ฌธ์œผ๋กœ ๋ณ€ํ™˜ํ•˜๊ณ  ์งˆ๋ฌธ์— ๋‹ตํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. (5) ์ž‘์—…์ž๋“ค ์‚ฌ์ด์—์„œ ๋‹ต๋ณ€์ด ์ผ์น˜ํ•˜๊ณ  ์‹ค์ œ ์„ ํƒํ•œ ๋‹จ์–ด์™€ ์ผ์น˜ํ•˜๋Š” ๊ฒฝ์šฐ, ํ•ด๋‹น ์ธ์Šคํ„ด์Šค๋ฅผ NewLAMA-Easy์— ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
Data Augmentation and synthetic dataEldan and Li (2023) showed that a synthetically generated dataset in the form of stories that toddlers can understand allows training a small language model that can generate coherent sentences. Gunasekar et al. (2023) showed that textbook quality (synthetic) data alone helps models achieve state-of-the-art performance on reasoning and coding tasks. Similar approaches are used in concurrent work for enhancing coding and mathematical reasoning abilities while finetuning Liu et al. (2023); Wei et al. (2023). Shumailov et al. (2023) show that training on synthetic data can actually be harmful for model performance, especially when we do multiple rounds of pre-training an LLM and then training the next LLM on data generated by the previous one. On the other hand, some other works have shown that such a strategy can actually be useful. Li et al. (2023) and Koksal et al. (2023) discuss how a model can generate instruction data and then fine-tune on its own generated data to improve performance. Jung et al. (2023) discuss how such repeated cycles of synthetic data can help train a very small paraphrase and summarization model that even outperforms GPT-3.
๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐEldan๊ณผ Li(2023)๋Š” ์œ ์•„๊ฐ€ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ์Šคํ† ๋ฆฌ ํ˜•์‹์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ†ตํ•ด ์ผ๊ด€๋œ ๋ฌธ์žฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. Gunasekar ๋“ฑ(2023)์€ ๊ต๊ณผ์„œ ํ’ˆ์งˆ(ํ•ฉ์„ฑ) ๋ฐ์ดํ„ฐ๋งŒ์œผ๋กœ๋„ ๋ชจ๋ธ์ด ์ถ”๋ก  ๋ฐ ์ฝ”๋”ฉ ์ž‘์—…์—์„œ ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์œ ์‚ฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ฝ”๋”ฉ ๋ฐ ์ˆ˜ํ•™์  ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋™์‹œ ์ž‘์—…์—์„œ๋„ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. Liu et al.(2023); Wei et al.(2023). Shumailov et al.(2023)์€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ด ํŠนํžˆ LLM์„ ์—ฌ๋Ÿฌ ๋ฒˆ ์‚ฌ์ „ ํ›ˆ๋ จํ•œ ๋‹ค์Œ ์ด์ „ ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๋กœ ๋‹ค์Œ LLM์„ ํ›ˆ๋ จํ•  ๋•Œ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ํ•ด๋กœ์šธ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ๋‹ค๋ฅธ ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์ „๋žต์ด ์‹ค์ œ๋กœ ์œ ์šฉํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. Li et al.(2023)๊ณผ Koksal et al.(2023)์€ ๋ชจ๋ธ์ด Instruction ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•œ ๋‹ค์Œ ์ž์ฒด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด Fine-tuningํ•˜์—ฌ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. Jung et al.(2023)์€ ์ด๋Ÿฌํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๋ฐ˜๋ณต์ ์ธ ์ฃผ๊ธฐ๊ฐ€ GPT-3๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋‚ด๋Š” ๋งค์šฐ ์ž‘์€ ์˜์—ญ ๋ฐ ์š”์•ฝ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์–ด๋–ป๊ฒŒ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋Š”์ง€์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐEldan๊ณผ Li(2023)๋Š” ์œ ์•„๊ฐ€ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ์Šคํ† ๋ฆฌ ํ˜•์‹์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ†ตํ•ด ์ผ๊ด€๋œ ๋ฌธ์žฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. Gunasekar ๋“ฑ(2023)์€ ๊ต๊ณผ์„œ ํ’ˆ์งˆ(ํ•ฉ์„ฑ) ๋ฐ์ดํ„ฐ๋งŒ์œผ๋กœ๋„ ๋ชจ๋ธ์ด ์ถ”๋ก  ๋ฐ ์ฝ”๋”ฉ ์ž‘์—…์—์„œ ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์œ ์‚ฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ฝ”๋”ฉ ๋ฐ ์ˆ˜ํ•™์  ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋™์‹œ ์ž‘์—…์—์„œ๋„ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. Liu et al.(2023); Wei et al.(2023). Shumailov et al.(2023)์€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ด ํŠนํžˆ LLM์„ ์—ฌ๋Ÿฌ ๋ฒˆ ์‚ฌ์ „ ํ›ˆ๋ จํ•œ ๋‹ค์Œ ์ด์ „ ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๋กœ ๋‹ค์Œ LLM์„ ํ›ˆ๋ จํ•  ๋•Œ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ํ•ด๋กœ์šธ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ๋‹ค๋ฅธ ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์ „๋žต์ด ์‹ค์ œ๋กœ ์œ ์šฉํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. Li et al.(2023)๊ณผ Koksal et al.(2023)์€ ๋ชจ๋ธ์ด ์ง€์นจ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•œ ๋‹ค์Œ ์ž์ฒด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. Jung et al.(2023)์€ ์ด๋Ÿฌํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๋ฐ˜๋ณต์ ์ธ ์ฃผ๊ธฐ๊ฐ€ GPT-3๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋‚ด๋Š” ๋งค์šฐ ์ž‘์€ ์˜์—ญ ๋ฐ ์š”์•ฝ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์–ด๋–ป๊ฒŒ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋Š”์ง€์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
Figure 2: Across six QA datasets using GPT-4, we consistently observe an inverse relationship between the RAG preference rate (y-axis) and two characteristics (x-axes): 1. the modelโ€™s prior response probability (lefthand plots), and 2. the amount of deviation from the prior (righthand plots). RAG preference rate is defined as the proportion of responses that align with the information presented in the prompt as context. The modelโ€™s prior response probability is computed from the average log probability of the response tokens queried without RAG. The left plot in each pair visualizes the prior probability (grouped into 10 bins) against the RAG preference rate, along with the best-fit trend line and slope. The right plot visualizes absolute deviation from the reference information (for numerical datasets (top), up to two log-fold changes (along with the trendline); for categorical datasets (bottom), a total of four modification categories) against RAG preference rate. Additionally, the upper and lower half percentiles are shown in the right plots to illustrate that lower probability prior responses have monotonically lower RAG preference rates than higher probability prior responses.
๊ทธ๋ฆผ 2: GPT-4๋ฅผ ์‚ฌ์šฉํ•œ 6๊ฐœ์˜ QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ RAG ์„ ํ˜ธ๋„(y์ถ•)์™€ ๋‘ ๊ฐ€์ง€ ํŠน์„ฑ(x์ถ•) ์‚ฌ์ด์— ์ผ๊ด€๋˜๊ฒŒ ๋ฐ˜๋น„๋ก€ ๊ด€๊ณ„๊ฐ€ ๊ด€์ฐฐ๋ฉ๋‹ˆ๋‹ค: 1. ๋ชจ๋ธ์˜ prior ์‘๋‹ต ํ™•๋ฅ (์™ผ์ชฝ ํ”Œ๋กฏ)๊ณผ 2. prior์—์„œ ๋ฒ—์–ด๋‚œ ์–‘(์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ). RAG ์„ ํ˜ธ๋„๋Š” ํ”„๋กฌํ”„ํŠธ์— ์ œ์‹œ๋œ ์ •๋ณด์™€ ์ผ์น˜ํ•˜๋Š” ์‘๋‹ต์˜ ๋น„์œจ๋กœ ์ •์˜๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์˜ prior ์‘๋‹ต ํ™•๋ฅ ์€ RAG ์—†์ด ์ฟผ๋ฆฌ๋œ ์‘๋‹ต ํ† ํฐ์˜ ํ‰๊ท  ๋กœ๊ทธ ํ™•๋ฅ ์—์„œ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ์Œ์˜ ์™ผ์ชฝ ํ”Œ๋กฏ์€ ์ตœ์ ์˜ ์ถ”์„ธ์„ ๊ณผ ๊ธฐ์šธ๊ธฐ์™€ ํ•จ๊ป˜ RAG ์„ ํ˜ธ๋„์— ๋Œ€ํ•œ prior ํ™•๋ฅ (10๊ฐœ์˜ ๊ตฌ๊ฐ„์œผ๋กœ ๊ทธ๋ฃนํ™”๋จ)์„ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ์€ ์ฐธ์กฐ ์ •๋ณด๋กœ๋ถ€ํ„ฐ์˜ ์ ˆ๋Œ€ ํŽธ์ฐจ(์ˆซ์ž ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ(์œ„์ชฝ), ์ตœ๋Œ€ 2๋ฐฐ์˜ ๋กœ๊ทธ ๋ณ€ํ™”(์ถ”์„ธ์„  ํฌํ•จ), ๋ฒ”์ฃผํ˜• ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ(์•„๋ž˜์ชฝ), ์ด 4๊ฐœ์˜ ์ˆ˜์ • ๋ฒ”์ฃผ)๋ฅผ RAG ์„ ํ˜ธ๋„์— ๋Œ€๋น„ํ•˜์—ฌ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ์—๋Š” ์ƒ์œ„ ๋ฐ ํ•˜์œ„ 5๋ถ„์œ„์ˆ˜๊ฐ€ ํ‘œ์‹œ๋˜์–ด ๋‚ฎ์€ ํ™•๋ฅ ์˜ prior ์‘๋‹ต์ด ๋†’์€ ํ™•๋ฅ ์˜ prior ์‘๋‹ต๋ณด๋‹ค RAG ์„ ํ˜ธ๋„๊ฐ€ ๋‹จ์กฐ๋กญ๊ฒŒ ๋‚ฎ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2: GPT-4๋ฅผ ์‚ฌ์šฉํ•œ 6๊ฐœ์˜ QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ RAG ์„ ํ˜ธ๋„(y์ถ•)์™€ ๋‘ ๊ฐ€์ง€ ํŠน์„ฑ(x์ถ•) ์‚ฌ์ด์— ์ผ๊ด€๋˜๊ฒŒ ๋ฐ˜๋น„๋ก€ ๊ด€๊ณ„๊ฐ€ ๊ด€์ฐฐ๋ฉ๋‹ˆ๋‹ค: 1. ๋ชจ๋ธ์˜ ์‚ฌ์ „ ์‘๋‹ต ํ™•๋ฅ (์™ผ์ชฝ ํ”Œ๋กฏ)๊ณผ 2. ์‚ฌ์ „์—์„œ ๋ฒ—์–ด๋‚œ ์–‘(์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ). RAG ์„ ํ˜ธ๋„๋Š” ํ”„๋กฌํ”„ํŠธ์— ์ œ์‹œ๋œ ์ •๋ณด์™€ ์ผ์น˜ํ•˜๋Š” ์‘๋‹ต์˜ ๋น„์œจ๋กœ ์ •์˜๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์˜ ์‚ฌ์ „ ์‘๋‹ต ํ™•๋ฅ ์€ RAG ์—†์ด ์ฟผ๋ฆฌ๋œ ์‘๋‹ต ํ† ํฐ์˜ ํ‰๊ท  ๋กœ๊ทธ ํ™•๋ฅ ์—์„œ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ์Œ์˜ ์™ผ์ชฝ ํ”Œ๋กฏ์€ ์ตœ์ ์˜ ์ถ”์„ธ์„ ๊ณผ ๊ธฐ์šธ๊ธฐ์™€ ํ•จ๊ป˜ RAG ์„ ํ˜ธ๋„์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ™•๋ฅ (10๊ฐœ์˜ ๊ตฌ๊ฐ„์œผ๋กœ ๊ทธ๋ฃนํ™”๋จ)์„ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ์€ ์ฐธ์กฐ ์ •๋ณด๋กœ๋ถ€ํ„ฐ์˜ ์ ˆ๋Œ€ ํŽธ์ฐจ(์ˆซ์ž ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ(์œ„์ชฝ), ์ตœ๋Œ€ 2๋ฐฐ์˜ ๋กœ๊ทธ ๋ณ€ํ™”(์ถ”์„ธ์„  ํฌํ•จ), ๋ฒ”์ฃผํ˜• ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ(์•„๋ž˜์ชฝ), ์ด 4๊ฐœ์˜ ์ˆ˜์ • ๋ฒ”์ฃผ)๋ฅผ RAG ์„ ํ˜ธ๋„์— ๋Œ€๋น„ํ•˜์—ฌ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ์—๋Š” ์ƒ์œ„ ๋ฐ ํ•˜์œ„ 5๋ถ„์œ„์ˆ˜๊ฐ€ ํ‘œ์‹œ๋˜์–ด ๋‚ฎ์€ ํ™•๋ฅ ์˜ ์‚ฌ์ „ ์‘๋‹ต์ด ๋†’์€ ํ™•๋ฅ ์˜ ์‚ฌ์ „ ์‘๋‹ต๋ณด๋‹ค RAG ์„ ํ˜ธ๋„๊ฐ€ ๋‹จ์กฐ๋กญ๊ฒŒ ๋‚ฎ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
To examine the ability of tool manipulation, existing work mostly adopts complex reasoning tasks for evaluation, such as mathematical problem solving (_e.g._, GSM8k [184] and SVAMP [592]) or knowledge question answering (_e.g._, TruthfulQA [556]), where the successful utilization of tools is very important for enhancing the required skills that LLMs are incapable in (_e.g._, numerical calculation). In this way, the evaluated performance on these tasks can reflect the ability of LLMs in tool manipulation. To teach LLMs to utilize tools, existing studies add exemplars using tools in context to elicit LLMs [443], or fine-tune LLMs on simulated data about tool utilization [693, 80]. It has been found that with the help of tools, LLMs become more capable of handling the issues that they are not good at, _e.g._, equation calculation and answering timely questions [80, 448]. However, as the number of available tools increases, the limited context length of LLMs may pose challenges in describing and demonstrating extensive tool APIs. To address this issue, existing work retrieves the usage of relevant tools, or encoding tool information as tokens within the embedding space [702, 703, 704].
๋„๊ตฌ ์กฐ์ž‘ ๋Šฅ๋ ฅ์„ ๊ฒ€ํ† ํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์ž‘์—…์€ ๋Œ€๋ถ€๋ถ„ ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ(_์˜ˆ: GSM8k [184] ๋ฐ SVAMP [592])์ด๋‚˜ ์ง€์‹ ์งˆ๋ฌธ ๋‹ต๋ณ€(_์˜ˆ: TruthfulQA [556])๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ์ถ”๋ก  ์ž‘์—…์„ ํ‰๊ฐ€์— ์ฑ„ํƒํ•˜๋ฉฐ, ์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ ๋„๊ตฌ๋ฅผ ์„ฑ๊ณต์ ์œผ๋กœ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์€ LLM์ด ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์—†๋Š”(_์˜ˆ: ์ˆ˜์น˜ ๊ณ„์‚ฐ) ํ•„์ˆ˜ ๊ธฐ์ˆ ์„ ํ–ฅ์ƒํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ์ด๋Ÿฌํ•œ ์ž‘์—…์— ๋Œ€ํ•œ ํ‰๊ฐ€ ์„ฑ๋Šฅ์€ ๋„๊ตฌ ์กฐ์ž‘์— ๋Œ€ํ•œ LLM์˜ ๋Šฅ๋ ฅ์„ ๋ฐ˜์˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. LLM์ด ๋„๊ตฌ๋ฅผ ํ™œ์šฉํ•˜๋„๋ก ๊ฐ€๋ฅด์น˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ๋งฅ๋ฝ์—์„œ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•œ ์˜ˆ์ œ๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ LLM์„ ์ด๋Œ์–ด๋‚ด๊ฑฐ๋‚˜ [443] ๋„๊ตฌ ํ™œ์šฉ์— ๋Œ€ํ•œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด Fine-tuningํ•ฉ๋‹ˆ๋‹ค [693, 80]. ๋„๊ตฌ์˜ ๋„์›€์œผ๋กœ LLM์€ ๋ฐฉ์ •์‹ ๊ณ„์‚ฐ ๋ฐ ์‹œ๊ธฐ์ ์ ˆํ•œ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€๊ณผ ๊ฐ™์ด ์ž์‹ ์ด ์ž˜ํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ๋” ์ž˜ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค[80, 448]. ๊ทธ๋Ÿฌ๋‚˜ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋„๊ตฌ์˜ ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ LLM์˜ ์ œํ•œ๋œ ์ปจํ…์ŠคํŠธ ๊ธธ์ด๊ฐ€ ๊ด‘๋ฒ”์œ„ํ•œ ๋„๊ตฌ API๋ฅผ ์„ค๋ช…ํ•˜๊ณ  ์‹œ์—ฐํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์ž‘์—…์—์„œ๋Š” ๊ด€๋ จ ๋„๊ตฌ์˜ ์‚ฌ์šฉ์„ ๊ฒ€์ƒ‰ํ•˜๊ฑฐ๋‚˜ ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„ ๋‚ด์—์„œ ํ† ํฐ์œผ๋กœ ๋„๊ตฌ ์ •๋ณด๋ฅผ ์ธ์ฝ”๋”ฉํ•ฉ๋‹ˆ๋‹ค [702, 703, 704].
๋„๊ตฌ ์กฐ์ž‘ ๋Šฅ๋ ฅ์„ ๊ฒ€ํ† ํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์ž‘์—…์€ ๋Œ€๋ถ€๋ถ„ ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ(_์˜ˆ: GSM8k [184] ๋ฐ SVAMP [592])์ด๋‚˜ ์ง€์‹ ์งˆ๋ฌธ ๋‹ต๋ณ€(_์˜ˆ: TruthfulQA [556])๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ์ถ”๋ก  ์ž‘์—…์„ ํ‰๊ฐ€์— ์ฑ„ํƒํ•˜๋ฉฐ, ์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ ๋„๊ตฌ๋ฅผ ์„ฑ๊ณต์ ์œผ๋กœ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์€ LLM์ด ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์—†๋Š”(_์˜ˆ: ์ˆ˜์น˜ ๊ณ„์‚ฐ) ํ•„์ˆ˜ ๊ธฐ์ˆ ์„ ํ–ฅ์ƒํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ์ด๋Ÿฌํ•œ ์ž‘์—…์— ๋Œ€ํ•œ ํ‰๊ฐ€ ์„ฑ๋Šฅ์€ ๋„๊ตฌ ์กฐ์ž‘์— ๋Œ€ํ•œ LLM์˜ ๋Šฅ๋ ฅ์„ ๋ฐ˜์˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. LLM์ด ๋„๊ตฌ๋ฅผ ํ™œ์šฉํ•˜๋„๋ก ๊ฐ€๋ฅด์น˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ๋งฅ๋ฝ์—์„œ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•œ ์˜ˆ์ œ๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ LLM์„ ์ด๋Œ์–ด๋‚ด๊ฑฐ๋‚˜ [443] ๋„๊ตฌ ํ™œ์šฉ์— ๋Œ€ํ•œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค [693, 80]. ๋„๊ตฌ์˜ ๋„์›€์œผ๋กœ LLM์€ ๋ฐฉ์ •์‹ ๊ณ„์‚ฐ ๋ฐ ์‹œ๊ธฐ์ ์ ˆํ•œ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€๊ณผ ๊ฐ™์ด ์ž์‹ ์ด ์ž˜ํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ๋” ์ž˜ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค[80, 448]. ๊ทธ๋Ÿฌ๋‚˜ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋„๊ตฌ์˜ ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ LLM์˜ ์ œํ•œ๋œ ์ปจํ…์ŠคํŠธ ๊ธธ์ด๊ฐ€ ๊ด‘๋ฒ”์œ„ํ•œ ๋„๊ตฌ API๋ฅผ ์„ค๋ช…ํ•˜๊ณ  ์‹œ์—ฐํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์ž‘์—…์—์„œ๋Š” ๊ด€๋ จ ๋„๊ตฌ์˜ ์‚ฌ์šฉ์„ ๊ฒ€์ƒ‰ํ•˜๊ฑฐ๋‚˜ ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„ ๋‚ด์—์„œ ํ† ํฐ์œผ๋กœ ๋„๊ตฌ ์ •๋ณด๋ฅผ ์ธ์ฝ”๋”ฉํ•ฉ๋‹ˆ๋‹ค [702, 703, 704].
We assess the efficacy of the chat vector across multiple target languages, focusing primarily on Traditional Chinese, by considering three aspects: toxicity, the ability to follow instructions and multi-turn dialogue. The models are evaluated on three benchmarks: SAFETYPROMPTS (Sun et al., 2023), REALOXICITYPROMPTS (Gehman et al., 2020), and the Vicuna Benchmark (Chiang et al., 2023), with GPT-4 handling the translation of the latter two into the target language. The results demonstrate that the strategy of incorporating the chat vector after continual pre-training yielded superior outcomes compared to direct pre-training on LLaMa-2-chat. Furthermore, applying fine-tuning prior to the integration of the chat vector optimizes performance irrespective of the fine-tuning dataset's scale or the language of the pre-trained model. Beyond merely augmenting an LLM's conversational skills, it offers crucial insights into the meaning of learning weights in the parameter space and the integration of added vectors with pre-existing knowledge. Most importantly, performing arithmetic operations on the chat vector is substantially more efficient than reimplementing RLHF in the target language.
๋…์„ฑ, Instruction ๋ฐ ๋‹ค์ค‘ ํ„ด ๋Œ€ํ™”๋ฅผ ๋”ฐ๋ฅด๋Š” ๋Šฅ๋ ฅ์˜ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์„ ๊ณ ๋ คํ•˜์—ฌ ์ฃผ๋กœ ์ค‘๊ตญ์–ด๋ฅผ ์ค‘์‹ฌ์œผ๋กœ ์—ฌ๋Ÿฌ ๋Œ€์ƒ ์–ธ์–ด์— ๊ฑธ์ณ ์ฑ„ํŒ… ๋ฒกํ„ฐ์˜ ํšจ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์„ธ ๊ฐ€์ง€ ๋ฒค์น˜๋งˆํฌ์—์„œ ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค: SAFETYPROMPTS(Sun et al., 2023), REALOXICITYPROMPTS(Gehman et al., 2020), Vicuna ๋ฒค์น˜๋งˆํฌ(Chiang et al., 2023)์—์„œ ํ‰๊ฐ€๋˜๋ฉฐ, GPT-4๋Š” ํ›„์ž์˜ ๋‘ ๊ฐ€์ง€๋ฅผ ๋Œ€์ƒ ์–ธ์–ด๋กœ ๋ฒˆ์—ญํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ํ›„ ์ฑ„ํŒ… ๋ฒกํ„ฐ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ์ „๋žต์ด LLaMa-2-chat์— ๋Œ€ํ•œ ์ง์ ‘ ์‚ฌ์ „ ํ•™์Šต์— ๋น„ํ•ด ๋” ์šฐ์ˆ˜ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜จ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ฑ„ํŒ… ๋ฒกํ„ฐ๋ฅผ ํ†ตํ•ฉํ•˜๊ธฐ ์ „์— Fine-tuning์„ ์ ์šฉํ•˜๋ฉด Fine-tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ทœ๋ชจ๋‚˜ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ์–ธ์–ด์— ๊ด€๊ณ„์—†์ด ์„ฑ๋Šฅ์ด ์ตœ์ ํ™”๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋‹จ์ˆœํžˆ LLM์˜ ๋Œ€ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ณต๊ฐ„์—์„œ ํ•™์Šต ๊ฐ€์ค‘์น˜์˜ ์˜๋ฏธ์™€ ์ถ”๊ฐ€ ๋ฒกํ„ฐ๋ฅผ ๊ธฐ์กด ์ง€์‹๊ณผ ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๊ฒƒ์€ ์ฑ„ํŒ… ๋ฒกํ„ฐ์— ์‚ฐ์ˆ  ์—ฐ์‚ฐ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์ด ๋Œ€์ƒ ์–ธ์–ด๋กœ RLHF๋ฅผ ๋‹ค์‹œ ๊ตฌํ˜„ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ํ›จ์”ฌ ํšจ์œจ์ ์ด๋ผ๋Š” ์ ์ž…๋‹ˆ๋‹ค.
๋…์„ฑ, ์ง€์‹œ ์‚ฌํ•ญ ๋ฐ ๋‹ค์ค‘ ํ„ด ๋Œ€ํ™”๋ฅผ ๋”ฐ๋ฅด๋Š” ๋Šฅ๋ ฅ์˜ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์„ ๊ณ ๋ คํ•˜์—ฌ ์ฃผ๋กœ ์ค‘๊ตญ์–ด๋ฅผ ์ค‘์‹ฌ์œผ๋กœ ์—ฌ๋Ÿฌ ๋Œ€์ƒ ์–ธ์–ด์— ๊ฑธ์ณ ์ฑ„ํŒ… ๋ฒกํ„ฐ์˜ ํšจ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์„ธ ๊ฐ€์ง€ ๋ฒค์น˜๋งˆํฌ์—์„œ ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค: SAFETYPROMPTS(Sun et al., 2023), REALOXICITYPROMPTS(Gehman et al., 2020), Vicuna ๋ฒค์น˜๋งˆํฌ(Chiang et al., 2023)์—์„œ ํ‰๊ฐ€๋˜๋ฉฐ, GPT-4๋Š” ํ›„์ž์˜ ๋‘ ๊ฐ€์ง€๋ฅผ ๋Œ€์ƒ ์–ธ์–ด๋กœ ๋ฒˆ์—ญํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ํ›„ ์ฑ„ํŒ… ๋ฒกํ„ฐ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ์ „๋žต์ด LLaMa-2-chat์— ๋Œ€ํ•œ ์ง์ ‘ ์‚ฌ์ „ ํ•™์Šต์— ๋น„ํ•ด ๋” ์šฐ์ˆ˜ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜จ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ฑ„ํŒ… ๋ฒกํ„ฐ๋ฅผ ํ†ตํ•ฉํ•˜๊ธฐ ์ „์— ๋ฏธ์„ธ ์กฐ์ •์„ ์ ์šฉํ•˜๋ฉด ๋ฏธ์„ธ ์กฐ์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ทœ๋ชจ๋‚˜ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ์–ธ์–ด์— ๊ด€๊ณ„์—†์ด ์„ฑ๋Šฅ์ด ์ตœ์ ํ™”๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋‹จ์ˆœํžˆ LLM์˜ ๋Œ€ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ณต๊ฐ„์—์„œ ํ•™์Šต ๊ฐ€์ค‘์น˜์˜ ์˜๋ฏธ์™€ ์ถ”๊ฐ€ ๋ฒกํ„ฐ๋ฅผ ๊ธฐ์กด ์ง€์‹๊ณผ ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๊ฒƒ์€ ์ฑ„ํŒ… ๋ฒกํ„ฐ์— ์‚ฐ์ˆ  ์—ฐ์‚ฐ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์ด ๋Œ€์ƒ ์–ธ์–ด๋กœ RLHF๋ฅผ ๋‹ค์‹œ ๊ตฌํ˜„ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ํ›จ์”ฌ ํšจ์œจ์ ์ด๋ผ๋Š” ์ ์ž…๋‹ˆ๋‹ค.
SetupFor intrinsic evaluations, Paloma proposes a range of analyses, from inspection of performance in each domain separately to more summarized results over combinations of domains. We report results at two levels of granularity: the aggregate performance over 11 of the 18 sources in Paloma as in Magnusson et al. (2023), as well as more fine-grained results over each of these sources individually. This particular subset of 11 sources from Paloma excludes sources that are not publicly available, involve fringe or toxic text, or consist of code data not supported by Paloma's decontamination approach. This leaves C4 (Raffel et al., 2020), mC4-en (Chung et al., 2023), Wikitext 103 (Merity et al., 2016), Penn Treebank (Marcus et al., 1999; Nunes, 2020), RedPajama (Together Computer, 2023), Falcon-RefinedWeb (Penedo et al., 2023), Dolma (Soldaini et al., 2024), M2D2 S2ORC (Reid et al., 2022), M2D2 Wikipedia (Reid et al., 2022), C4 100 domains (Chronopoulou et al., 2022), and Dolma 100 Subreddits (Soldaini et al., 2024). To allow for a fair comparison between models with different vocabularies, we report bits per byte as defined by Gao et al. (2020) over the test sets of these sources.
์„ค์ •๋ณธ์งˆ์ ์ธ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด Paloma๋Š” ๊ฐ ๋„๋ฉ”์ธ์—์„œ์˜ ์„ฑ๋Šฅ ๊ฒ€์‚ฌ๋ถ€ํ„ฐ ๋„๋ฉ”์ธ ์กฐํ•ฉ์— ๋Œ€ํ•œ ์š”์•ฝ๋œ ๊ฒฐ๊ณผ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ๋ถ„์„์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Magnusson et al.(2023)์—์„œ์™€ ๊ฐ™์ด Paloma์˜ 18๊ฐœ ์†Œ์Šค ์ค‘ 11๊ฐœ ์†Œ์Šค์— ๋Œ€ํ•œ ์ข…ํ•ฉ์ ์ธ ์„ฑ๋Šฅ๊ณผ ๊ฐ ์†Œ์Šค์— ๋Œ€ํ•œ ๋ณด๋‹ค ์„ธ๋ถ„ํ™”๋œ ๊ฒฐ๊ณผ์˜ ๋‘ ๊ฐ€์ง€ ์ˆ˜์ค€์˜ ์„ธ๋ถ„์„ฑ์—์„œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ด ํŠน์ • 11๊ฐœ ์†Œ์Šค์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์—๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๊ฑฐ๋‚˜, fringe ๋˜๋Š” toxic ํ…์ŠคํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๊ฑฐ๋‚˜, Paloma์˜ decontamination ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ง€์›ํ•˜์ง€ ์•Š๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ ์†Œ์Šค๊ฐ€ ์ œ์™ธ๋ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ C4(Raffel et al., 2020), mC4-en(Chung et al., 2023), Wikitext 103(Merity et al., 2016), Penn Treebank(Marcus et al., 1999; Nunes, 2020), RedPajama(Together Computer, 2023), Falcon-RefinedWeb(Penedo et al., 2023), Dolma(Soldaini et al., 2024), M2D2 S2ORC(Reid et al., 2022), M2D2 Wikipedia(Reid et al., 2022), C4 100 domains(Chronopoulou et al., 2022), Dolma 100 Subreddits(Soldaini et al., 2024)์ด ๋‚จ๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์–ดํœ˜๊ฐ€ ๋‹ค๋ฅธ ๋ชจ๋ธ ๊ฐ„์˜ ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด ์ด๋Ÿฌํ•œ ์†Œ์Šค์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ์— ๋Œ€ํ•ด Gao et al.(2020)์—์„œ ์ •์˜ํ•œ bits per byte๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.
์„ค์ •๋ณธ์งˆ์ ์ธ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด Paloma๋Š” ๊ฐ ๋„๋ฉ”์ธ์—์„œ์˜ ์„ฑ๋Šฅ ๊ฒ€์‚ฌ๋ถ€ํ„ฐ ๋„๋ฉ”์ธ ์กฐํ•ฉ์— ๋Œ€ํ•œ ์š”์•ฝ๋œ ๊ฒฐ๊ณผ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ๋ถ„์„์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Magnusson et al.(2023)์—์„œ์™€ ๊ฐ™์ด Paloma์˜ 18๊ฐœ ์†Œ์Šค ์ค‘ 11๊ฐœ ์†Œ์Šค์— ๋Œ€ํ•œ ์ข…ํ•ฉ์ ์ธ ์„ฑ๋Šฅ๊ณผ ๊ฐ ์†Œ์Šค์— ๋Œ€ํ•œ ๋ณด๋‹ค ์„ธ๋ถ„ํ™”๋œ ๊ฒฐ๊ณผ์˜ ๋‘ ๊ฐ€์ง€ ์ˆ˜์ค€์˜ ์„ธ๋ถ„์„ฑ์—์„œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ด ํŠน์ • 11๊ฐœ ์†Œ์Šค์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์—๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๊ฑฐ๋‚˜, ํ”„๋ฆฐ์ง€ ๋˜๋Š” ๋…์„ฑ ํ…์ŠคํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๊ฑฐ๋‚˜, Paloma์˜ ์˜ค์—ผ ์ œ๊ฑฐ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ง€์›ํ•˜์ง€ ์•Š๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ ์†Œ์Šค๊ฐ€ ์ œ์™ธ๋ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ C4(Raffel et al., 2020), mC4-en(Chung et al., 2023), Wikitext 103(Merity et al., 2016), Penn Treebank(Marcus et al., 1999; Nunes, 2020), RedPajama(Together Computer, 2023), Falcon-RefinedWeb(Penedo et al., 2023), Dolma(Soldaini et al., 2024), M2D2 S2ORC(Reid et al., 2022), M2D2 Wikipedia(Reid et al., 2022), C4 100 ๋„๋ฉ”์ธ(Chronopoulou et al., 2022), Dolma 100 ์„œ๋ธŒ๋ ˆ๋”ง(Soldaini et al., 2024)์ด ๋‚จ๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์–ดํœ˜๊ฐ€ ๋‹ค๋ฅธ ๋ชจ๋ธ ๊ฐ„์˜ ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด ์ด๋Ÿฌํ•œ ์†Œ์Šค์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ์— ๋Œ€ํ•ด Gao et al.(2020)์—์„œ ์ •์˜ํ•œ ๋ฐ”์ดํŠธ๋‹น ๋น„ํŠธ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_Pre-trained language models (PLM)_. As an early attempt, ELMo [21] was proposed to capture context-aware word representations by first pre-training a bidirectional LSTM (biLSTM) network (instead of learning fixed word representations) and then fine-tuning the biLSTM network according to specific downstream tasks. Furthermore, based on the highly parallelizable Transformer architecture [22] with self-attention mechanisms, BERT [23] was proposed by pre-training bidirectional language models with specially designed pre-training tasks on large-scale unlabeled corpora. These pre-trained context-aware word representations are very effective as general-purpose semantic features, which have largely raised the performance bar of NLP tasks. This study has inspired a large number of follow-up work, which sets the _"pre-training"_ and _fine-tuning"_ learning paradigm. Following this paradigm, a great number of studies on PLMs have been developed, introducing either different architectures [24, 25] (_e.g._, GPT-2 [26] and BART [24]) or improved pre-training strategies [27, 28, 29]. In this paradigm, it often requires fine-tuning the PLM for adapting to different downstream tasks.
\(\bullet\)_Pre-trained language models (PLM)_. ์ดˆ๊ธฐ ์‹œ๋„๋กœ์„œ, ELMo [21]๋Š” ๋จผ์ € ์–‘๋ฐฉํ–ฅ LSTM(biLSTM) ๋„คํŠธ์›Œํฌ๋ฅผ ์‚ฌ์ „ ํ•™์Šตํ•œ ๋‹ค์Œ(๊ณ ์ •๋œ ๋‹จ์–ด ํ‘œํ˜„์„ ํ•™์Šตํ•˜๋Š” ๋Œ€์‹ ) ํŠน์ • ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋”ฐ๋ผ biLSTM ๋„คํŠธ์›Œํฌ๋ฅผ Fine-tuningํ•˜์—ฌ ์ปจํ…์ŠคํŠธ ์ธ์‹ ๋‹จ์–ด ํ‘œํ˜„์„ ์บก์ฒ˜ํ•˜๊ธฐ ์œ„ํ•ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, self-attention ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ฐ–์ถ˜ ๊ณ ๋„๋กœ ๋ณ‘๋ ฌํ™” ๊ฐ€๋Šฅํ•œ Transformer ์•„ํ‚คํ…์ฒ˜ [22]๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ, BERT [23]๋Š” ๋Œ€๊ทœ๋ชจ ๋ฏธ๋ถ„๋ฅ˜ ๋ง๋ญ‰์น˜์—์„œ ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ์‚ฌ์ „ ํ•™์Šต ์ž‘์—…์„ ํ†ตํ•ด ์–‘๋ฐฉํ–ฅ ์–ธ์–ด ๋ชจ๋ธ์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ์‚ฌ์ „ ํ•™์Šต๋œ ์ปจํ…์ŠคํŠธ ์ธ์‹ ๋‹จ์–ด ํ‘œํ˜„์€ ๋ฒ”์šฉ ์˜๋ฏธ๋ก ์  ํŠน์ง•์œผ๋กœ์„œ ๋งค์šฐ ํšจ๊ณผ์ ์ด๋ฉฐ, ์ด๋Š” NLP ์ž‘์—…์˜ ์„ฑ๋Šฅ ๊ธฐ์ค€์„ ํฌ๊ฒŒ ๋†’์˜€์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” _"pre-training"_๊ณผ _Fine-tuning"_ ํ•™์Šต ํŒจ๋Ÿฌ๋‹ค์ž„์„ ์„ค์ •ํ•œ ์ˆ˜๋งŽ์€ ํ›„์† ์—ฐ๊ตฌ์— ์˜๊ฐ์„ ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ํŒจ๋Ÿฌ๋‹ค์ž„์— ๋”ฐ๋ผ ๋‹ค์–‘ํ•œ ์•„ํ‚คํ…์ฒ˜ [24, 25](_์˜ˆ: GPT-2 [26] ๋ฐ BART [24]) ๋˜๋Š” ๊ฐœ์„ ๋œ ์‚ฌ์ „ ํ•™์Šต ์ „๋žต [27, 28, 29]์„ ๋„์ž…ํ•˜๋Š” PLM์— ๋Œ€ํ•œ ๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ํŒจ๋Ÿฌ๋‹ค์ž„์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ์ ์‘ํ•˜๊ธฐ ์œ„ํ•ด PLM์„ Fine-tuningํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ(PLM)_. ์ดˆ๊ธฐ ์‹œ๋„๋กœ์„œ, ELMo [21]๋Š” ๋จผ์ € ์–‘๋ฐฉํ–ฅ LSTM(biLSTM) ๋„คํŠธ์›Œํฌ๋ฅผ ์‚ฌ์ „ ํ•™์Šตํ•œ ๋‹ค์Œ(๊ณ ์ •๋œ ๋‹จ์–ด ํ‘œํ˜„์„ ํ•™์Šตํ•˜๋Š” ๋Œ€์‹ ) ํŠน์ • ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋”ฐ๋ผ biLSTM ๋„คํŠธ์›Œํฌ๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์ปจํ…์ŠคํŠธ ์ธ์‹ ๋‹จ์–ด ํ‘œํ˜„์„ ์บก์ฒ˜ํ•˜๊ธฐ ์œ„ํ•ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ž๊ธฐ ์ฃผ์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ฐ–์ถ˜ ๊ณ ๋„๋กœ ๋ณ‘๋ ฌํ™” ๊ฐ€๋Šฅํ•œ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜ [22]๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ, BERT [23]๋Š” ๋Œ€๊ทœ๋ชจ ๋ฏธ๋ถ„๋ฅ˜ ๋ง๋ญ‰์น˜์—์„œ ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ์‚ฌ์ „ ํ•™์Šต ์ž‘์—…์„ ํ†ตํ•ด ์–‘๋ฐฉํ–ฅ ์–ธ์–ด ๋ชจ๋ธ์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ์‚ฌ์ „ ํ•™์Šต๋œ ์ปจํ…์ŠคํŠธ ์ธ์‹ ๋‹จ์–ด ํ‘œํ˜„์€ ๋ฒ”์šฉ ์˜๋ฏธ๋ก ์  ํŠน์ง•์œผ๋กœ์„œ ๋งค์šฐ ํšจ๊ณผ์ ์ด๋ฉฐ, ์ด๋Š” NLP ์ž‘์—…์˜ ์„ฑ๋Šฅ ๊ธฐ์ค€์„ ํฌ๊ฒŒ ๋†’์˜€์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” _"์‚ฌ์ „ ํ•™์Šต"_๊ณผ _๋ฏธ์„ธ ์กฐ์ •"_ ํ•™์Šต ํŒจ๋Ÿฌ๋‹ค์ž„์„ ์„ค์ •ํ•œ ์ˆ˜๋งŽ์€ ํ›„์† ์—ฐ๊ตฌ์— ์˜๊ฐ์„ ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ํŒจ๋Ÿฌ๋‹ค์ž„์— ๋”ฐ๋ผ ๋‹ค์–‘ํ•œ ์•„ํ‚คํ…์ฒ˜ [24, 25](_์˜ˆ: GPT-2 [26] ๋ฐ BART [24]) ๋˜๋Š” ๊ฐœ์„ ๋œ ์‚ฌ์ „ ํ•™์Šต ์ „๋žต [27, 28, 29]์„ ๋„์ž…ํ•˜๋Š” PLM์— ๋Œ€ํ•œ ๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ํŒจ๋Ÿฌ๋‹ค์ž„์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ์ ์‘ํ•˜๊ธฐ ์œ„ํ•ด PLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค.
In addition to human evaluations, we implemented model-based auto-evaluation methods as economical consistent alternatives to specialist assessments. These techniques were employed to evaluate both dialogue quality and diagnostic accuracy of the OSCE agent. To establish the validity of our auto-evaluation methods for assessing dialogue quality, we initially focused on a subset of four evaluation axes from the PACES rubric(Table A.2) that were assessed by both the patient actors and the specialist physicians. The auto-evaluation, which uses a self-CoT strategy (details described in Section A.9) with AMIE to rate dialogues, was in good alignment with human raters and comparable to the inter-specialist agreement on these criteria. For the auto-evaluation of differential diagnoses, we leveraged another LLM, Med-PaLM 2 [13] as a surrogate for a specialist rater to grade the predicted diagnoses against the ground truth diagnoses (more details in Section A.7). Our auto-evaluation on DDx accuracy showed a similar trend for AMIE and OSCE agents compared to the specialist ratings. Overall, auto-evaluation trends aligned with human ratings for both dialogue quality and diagnostic accuracy.
์ธ๊ฐ„ ํ‰๊ฐ€ ์™ธ์—๋„ ๊ฒฝ์ œ์ ์œผ๋กœ ์ผ๊ด€๋œ ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€์˜ ๋Œ€์•ˆ์œผ๋กœ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ auto-evaluation ๋ฐฉ๋ฒ•์„ ๊ตฌํ˜„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ๋ฒ•์€ OSCE agent์˜ ๋Œ€ํ™” ํ’ˆ์งˆ๊ณผ ์ง„๋‹จ ์ •ํ™•๋„๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ™” ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ auto-evaluation ๋ฐฉ๋ฒ•์˜ ํƒ€๋‹น์„ฑ์„ ์ž…์ฆํ•˜๊ธฐ ์œ„ํ•ด ๋จผ์ € ํ™˜์ž ๋ฐฐ์šฐ์ž์™€ ์ „๋ฌธ์˜๊ฐ€ ๋ชจ๋‘ ํ‰๊ฐ€ํ•œ PACES rubric(ํ‘œ A.2)์˜ 4๊ฐ€์ง€ ํ‰๊ฐ€ ์ถ• ํ•˜์œ„ ์ง‘ํ•ฉ์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์Šต๋‹ˆ๋‹ค. self-CoT ์ „๋žต(์„ธ๋ถ€ ์‚ฌํ•ญ์€ ์„น์…˜ A.9์—์„œ ์„ค๋ช…)์„ ์‚ฌ์šฉํ•˜์—ฌ ๋Œ€ํ™”๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” auto-evaluation์€ ์ธ๊ฐ„ ํ‰๊ฐ€์ž์™€ ์ž˜ ์ผ์น˜ํ–ˆ์œผ๋ฉฐ, ์ด๋Ÿฌํ•œ ๊ธฐ์ค€์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€ ๊ฐ„์˜ ์ผ์น˜๋„์™€ ๋น„์Šทํ–ˆ์Šต๋‹ˆ๋‹ค. differential diagnoses์˜ auto-evaluation์„ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ๋˜ ๋‹ค๋ฅธ LLM์ธ Med-PaLM 2 [13]๋ฅผ ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€์ž์˜ ๋Œ€๋ฆฌ์ž๋กœ ํ™œ์šฉํ•˜์—ฌ ์˜ˆ์ธก๋œ ์ง„๋‹จ์„ ์‹ค์ธก ์ง„๋‹จ๊ณผ ๋น„๊ตํ•˜์—ฌ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค(์ž์„ธํ•œ ๋‚ด์šฉ์€ ์„น์…˜ A.7 ์ฐธ์กฐ). DDx ์ •ํ™•๋„์— ๋Œ€ํ•œ auto-evaluation ๊ฒฐ๊ณผ, ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€์™€ ๋น„๊ตํ–ˆ์„ ๋•Œ AMIE์™€ OSCE agent์˜ ์ถ”์„ธ๊ฐ€ ๋น„์Šทํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ ๋Œ€ํ™” ํ’ˆ์งˆ๊ณผ ์ง„๋‹จ ์ •ํ™•๋„ ๋ชจ๋‘์—์„œ auto-evaluation ์ถ”์„ธ๊ฐ€ ์ธ๊ฐ„ ํ‰๊ฐ€์™€ ์ผ์น˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
์ธ๊ฐ„ ํ‰๊ฐ€ ์™ธ์—๋„ ๊ฒฝ์ œ์ ์œผ๋กœ ์ผ๊ด€๋œ ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€์˜ ๋Œ€์•ˆ์œผ๋กœ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ž๋™ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ๊ตฌํ˜„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ๋ฒ•์€ OSCE ์—์ด์ „ํŠธ์˜ ๋Œ€ํ™” ํ’ˆ์งˆ๊ณผ ์ง„๋‹จ ์ •ํ™•๋„๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ™” ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ์ž๋™ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์˜ ํƒ€๋‹น์„ฑ์„ ์ž…์ฆํ•˜๊ธฐ ์œ„ํ•ด ๋จผ์ € ํ™˜์ž ๋ฐฐ์šฐ์ž์™€ ์ „๋ฌธ์˜๊ฐ€ ๋ชจ๋‘ ํ‰๊ฐ€ํ•œ PACES ๋ฃจ๋ธŒ๋ฆญ(ํ‘œ A.2)์˜ 4๊ฐ€์ง€ ํ‰๊ฐ€ ์ถ• ํ•˜์œ„ ์ง‘ํ•ฉ์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์Šต๋‹ˆ๋‹ค. ์ž๊ธฐ CoT ์ „๋žต(์„ธ๋ถ€ ์‚ฌํ•ญ์€ ์„น์…˜ A.9์—์„œ ์„ค๋ช…)์„ ์‚ฌ์šฉํ•˜์—ฌ ๋Œ€ํ™”๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ์ž๋™ ํ‰๊ฐ€๋Š” ์ธ๊ฐ„ ํ‰๊ฐ€์ž์™€ ์ž˜ ์ผ์น˜ํ–ˆ์œผ๋ฉฐ, ์ด๋Ÿฌํ•œ ๊ธฐ์ค€์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€ ๊ฐ„์˜ ์ผ์น˜๋„์™€ ๋น„์Šทํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ๋ณ„ ์ง„๋‹จ์˜ ์ž๋™ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ๋˜ ๋‹ค๋ฅธ LLM์ธ Med-PaLM 2 [13]๋ฅผ ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€์ž์˜ ๋Œ€๋ฆฌ์ž๋กœ ํ™œ์šฉํ•˜์—ฌ ์˜ˆ์ธก๋œ ์ง„๋‹จ์„ ์‹ค์ธก ์ง„๋‹จ๊ณผ ๋น„๊ตํ•˜์—ฌ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค(์ž์„ธํ•œ ๋‚ด์šฉ์€ ์„น์…˜ A.7 ์ฐธ์กฐ). DDx ์ •ํ™•๋„์— ๋Œ€ํ•œ ์ž๋™ ํ‰๊ฐ€ ๊ฒฐ๊ณผ, ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€์™€ ๋น„๊ตํ–ˆ์„ ๋•Œ AMIE์™€ OSCE ์—์ด์ „ํŠธ์˜ ์ถ”์„ธ๊ฐ€ ๋น„์Šทํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ ๋Œ€ํ™” ํ’ˆ์งˆ๊ณผ ์ง„๋‹จ ์ •ํ™•๋„ ๋ชจ๋‘์—์„œ ์ž๋™ ํ‰๊ฐ€ ์ถ”์„ธ๊ฐ€ ์ธ๊ฐ„ ํ‰๊ฐ€์™€ ์ผ์น˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
There exists many English instruction tuning datasets. However, the available datasets for Chinese instruction tuning are generally either limited in size or lacking in quality. Chinese instruction tuning datasets are categorized into three main types: (1) datasets derived from English instruction datasets Peng et al. (2023) or NLP datasets CLUEbenchmark (2022); Yang (2023), (2) datasets generated by LLMs Guo et al. (2023), and (3) self-generated instruction tuning datasets Ji et al. (2023); Sun et al. (2023). COIG Zhang et al. (2023) integrates multiple approaches to construct a human-verified universal high-quality Chinese instruction corpus. However, the previously mentioned Chinese instruction tuning datasets have inherent issues such as not aligning with natural Chinese communication patterns, lacking genuine Chinese linguistic data, containing numerous problematic data points, and having small-scale data. This paper focuses on constructing a Chinese instruction tuning dataset sourced from authentic Chinese linguistic data across diverse domains and undergone meticulous manual cleaning procedures aimed at enhancing the proficiency of LLMs in following Chinese instructions.
๋งŽ์€ ์˜์–ด Instruction Tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ค‘๊ตญ์–ด Instruction Tuning์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ํฌ๊ธฐ๊ฐ€ ์ œํ•œ์ ์ด๊ฑฐ๋‚˜ ํ’ˆ์งˆ์ด ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด Instruction Tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ํฌ๊ฒŒ ์„ธ ๊ฐ€์ง€ ์œ ํ˜•์œผ๋กœ ๋ถ„๋ฅ˜๋ฉ๋‹ˆ๋‹ค: (1) ์˜์–ด Instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ Peng et al. (2023) ๋˜๋Š” NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ CLUEbenchmark(2022)์—์„œ ํŒŒ์ƒ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ, (2) LLM์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ Yang (2023), (2) LLM์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ Guo et al. (2023), (3) ์ž์ฒด ์ƒ์„ฑ๋œ Instruction Tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ Ji et al. (2023); Sun et al. (2023). COIG Zhang et al. (2023)์€ ์—ฌ๋Ÿฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ฉํ•˜์—ฌ ์ธ๊ฐ„์ด ๊ฒ€์ฆํ•œ ๋ณดํŽธ์ ์ธ ๊ณ ํ’ˆ์งˆ ์ค‘๊ตญ์–ด Instruction ์ฝ”ํผ์Šค๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์ค‘๊ตญ์–ด Instruction Tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์ž์—ฐ์Šค๋Ÿฌ์šด ์ค‘๊ตญ์–ด ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ํŒจํ„ด๊ณผ ์ผ์น˜ํ•˜์ง€ ์•Š๊ณ , ์ง„์ •ํ•œ ์ค‘๊ตญ์–ด ์–ธ์–ด ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๋ฉฐ, ๋ฌธ์ œ๊ฐ€ ๋งŽ์€ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๊ณ , ๊ทœ๋ชจ๊ฐ€ ์ž‘์€ ๋ฐ์ดํ„ฐ ๋“ฑ์˜ ๋‚ด์žฌ๋œ ๋ฌธ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์—์„œ ์‹ค์ œ ์ค‘๊ตญ์–ด ์–ธ์–ด ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ค‘๊ตญ์–ด Instruction Tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜๊ณ , ์ค‘๊ตญ์–ด ์ง€์นจ์„ ๋”ฐ๋ฅด๋Š” LLM์˜ ์ˆ™๋ จ๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ์„ธ์‹ฌํ•œ ์ˆ˜๋™ ์ •๋ฆฌ ์ ˆ์ฐจ๋ฅผ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค.
๋งŽ์€ ์˜์–ด ๊ต์œก ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ค‘๊ตญ์–ด ๊ต์œก ํŠœ๋‹์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ํฌ๊ธฐ๊ฐ€ ์ œํ•œ์ ์ด๊ฑฐ๋‚˜ ํ’ˆ์งˆ์ด ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๊ต์œก ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ํฌ๊ฒŒ ์„ธ ๊ฐ€์ง€ ์œ ํ˜•์œผ๋กœ ๋ถ„๋ฅ˜๋ฉ๋‹ˆ๋‹ค: (1) ์˜์–ด ๊ต์œก ๋ฐ์ดํ„ฐ ์„ธํŠธ Peng et al. (2023) ๋˜๋Š” NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ CLUEbenchmark(2022)์—์„œ ํŒŒ์ƒ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ, (2) LLM์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ Yang (2023), (2) LLM์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ Guo et al. (2023), (3) ์ž์ฒด ์ƒ์„ฑ๋œ ์ง€์นจ ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ Ji et al. (2023); Sun et al. (2023). COIG Zhang et al. (2023)์€ ์—ฌ๋Ÿฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ฉํ•˜์—ฌ ์ธ๊ฐ„์ด ๊ฒ€์ฆํ•œ ๋ณดํŽธ์ ์ธ ๊ณ ํ’ˆ์งˆ ์ค‘๊ตญ์–ด ์ง€์นจ ์ฝ”ํผ์Šค๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์ค‘๊ตญ์–ด ์ง€์นจ ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์ž์—ฐ์Šค๋Ÿฌ์šด ์ค‘๊ตญ์–ด ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ํŒจํ„ด๊ณผ ์ผ์น˜ํ•˜์ง€ ์•Š๊ณ , ์ง„์ •ํ•œ ์ค‘๊ตญ์–ด ์–ธ์–ด ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๋ฉฐ, ๋ฌธ์ œ๊ฐ€ ๋งŽ์€ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๊ณ , ๊ทœ๋ชจ๊ฐ€ ์ž‘์€ ๋ฐ์ดํ„ฐ ๋“ฑ์˜ ๋‚ด์žฌ๋œ ๋ฌธ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์—์„œ ์‹ค์ œ ์ค‘๊ตญ์–ด ์–ธ์–ด ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ค‘๊ตญ์–ด ์ง€์นจ ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜๊ณ , ์ค‘๊ตญ์–ด ์ง€์นจ์„ ๋”ฐ๋ฅด๋Š” LLM์˜ ์ˆ™๋ จ๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ์„ธ์‹ฌํ•œ ์ˆ˜๋™ ์ •๋ฆฌ ์ ˆ์ฐจ๋ฅผ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค.
As demonstrated in Figure 1, Nemotron-4 15B exhibits high downstream accuracies across a wide range of English, code, and multilingual evaluation areas. In comparison to leading similarly-sized, open models we show that Nemotron-4 15B is significantly better than LLaMA-2 34B (Touvron et al., 2023b), which has over twice the number of parameters, and is better than Mistral 7B (Jiang et al., 2023) on all English evaluation areas. Additionally, Nemotron-4 15B achieves competitive accuracies to QWEN 14B (Bai et al., 2023) and Gemma 7B (Gemma Team, 2024). In a comparison across a wide range of programming languages, we find that Nemotron-4 15B achieves better average accuracy, and in particular on low-resource programming languages, than Starcoder (Li et al., 2023), a code-specific model, and Mistral 7B. As Nemotron-4 15B was trained on significant amount of multilingual data, it is currently the state-of-the-art general purpose model in its size class on all multilingual benchmarks. We find that Nemotron-4 is better than PALM 62B-Cont (Slav Petrov and et al., 2023), and also outperforms multilingual-specific models such as XGLM (Lin et al., 2022) and mGPT (Shliazhko et al., 2022).
๊ทธ๋ฆผ 1์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, Nemotron-4 15B๋Š” ๋‹ค์–‘ํ•œ ์˜์–ด, ์ฝ”๋“œ ๋ฐ ๋‹ค๊ตญ์–ด ํ‰๊ฐ€ ์˜์—ญ์—์„œ ๋†’์€ downstream ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋น„์Šทํ•œ ํฌ๊ธฐ์˜ ์„ ๋„์ ์ธ ์˜คํ”ˆ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•˜์—ฌ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ๋‘ ๋ฐฐ ์ด์ƒ ๋งŽ์€ LLaMA-2 34B(Touvron et al., 2023b)๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚˜๋ฉฐ, ๋ชจ๋“  ์˜์–ด ํ‰๊ฐ€ ์˜์—ญ์—์„œ Mistral 7B(Jiang et al., 2023)๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ Nemotron-4 15B๋Š” QWEN 14B(Bai et al., 2023) ๋ฐ Gemma 7B(Gemma Team, 2024)์™€ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด๋ฅผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ, Nemotron-4 15B๋Š” ์ฝ”๋“œ ์ „์šฉ ๋ชจ๋ธ์ธ Starcoder(Li et al., 2023) ๋ฐ Mistral 7B๋ณด๋‹ค ํ‰๊ท  ์ •ํ™•๋„๊ฐ€ ๋†’์œผ๋ฉฐ, ํŠนํžˆ ๋ฆฌ์†Œ์Šค๊ฐ€ ๋ถ€์กฑํ•œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์—์„œ ๋” ๋†’์€ ํ‰๊ท  ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. Nemotron-4 15B๋Š” ๋‹ค๊ตญ์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ๋‹น๋Ÿ‰ ํ•™์Šตํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ํ˜„์žฌ ๋ชจ๋“  ๋‹ค๊ตญ์–ด ๋ฒค์น˜๋งˆํฌ์—์„œ ๋™๊ธ‰ ๋ชจ๋ธ ์ค‘ state-of-the-art ๋ฒ”์šฉ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. Nemotron-4๊ฐ€ PALM 62B-Cont(Slav Petrov ๋ฐ al., 2023)๋ณด๋‹ค ์šฐ์ˆ˜ํ•˜๋ฉฐ, XGLM(Lin et al., 2022) ๋ฐ mGPT(Shliazhko et al., 2022)์™€ ๊ฐ™์€ ๋‹ค๊ตญ์–ด ์ „์šฉ ๋ชจ๋ธ๋ณด๋‹ค๋„ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, Nemotron-4 15B๋Š” ๋‹ค์–‘ํ•œ ์˜์–ด, ์ฝ”๋“œ ๋ฐ ๋‹ค๊ตญ์–ด ํ‰๊ฐ€ ์˜์—ญ์—์„œ ๋†’์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋น„์Šทํ•œ ํฌ๊ธฐ์˜ ์„ ๋„์ ์ธ ์˜คํ”ˆ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•˜์—ฌ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ๋‘ ๋ฐฐ ์ด์ƒ ๋งŽ์€ LLaMA-2 34B(Touvron et al., 2023b)๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚˜๋ฉฐ, ๋ชจ๋“  ์˜์–ด ํ‰๊ฐ€ ์˜์—ญ์—์„œ Mistral 7B(Jiang et al., 2023)๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ Nemotron-4 15B๋Š” QWEN 14B(Bai et al., 2023) ๋ฐ Gemma 7B(Gemma Team, 2024)์™€ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด๋ฅผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ, Nemotron-4 15B๋Š” ์ฝ”๋“œ ์ „์šฉ ๋ชจ๋ธ์ธ Starcoder(Li et al., 2023) ๋ฐ Mistral 7B๋ณด๋‹ค ํ‰๊ท  ์ •ํ™•๋„๊ฐ€ ๋†’์œผ๋ฉฐ, ํŠนํžˆ ๋ฆฌ์†Œ์Šค๊ฐ€ ๋ถ€์กฑํ•œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์—์„œ ๋” ๋†’์€ ํ‰๊ท  ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. Nemotron-4 15B๋Š” ๋‹ค๊ตญ์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ๋‹น๋Ÿ‰ ํ•™์Šตํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ํ˜„์žฌ ๋ชจ๋“  ๋‹ค๊ตญ์–ด ๋ฒค์น˜๋งˆํฌ์—์„œ ๋™๊ธ‰ ๋ชจ๋ธ ์ค‘ ์ตœ์ฒจ๋‹จ ๋ฒ”์šฉ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. Nemotron-4๊ฐ€ PALM 62B-Cont(Slav Petrov ๋ฐ al., 2023)๋ณด๋‹ค ์šฐ์ˆ˜ํ•˜๋ฉฐ, XGLM(Lin et al., 2022) ๋ฐ mGPT(Shliazhko et al., 2022)์™€ ๊ฐ™์€ ๋‹ค๊ตญ์–ด ์ „์šฉ ๋ชจ๋ธ๋ณด๋‹ค๋„ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค.
**LM-guided CoT Performance.** Table 2 shows a comprehensive breakdown of our method's performance. Additionally, we explore an extension of our approach, which involves sampling multiple reasoning paths and subsequently ranking the most optimal rationales. Our method with only KD outperforms the original CoT prompting with 2% gain for HotpotQA and 10% for 2WikiMultiHopQA, respectively. Figure 2 illustrates the respective rationale qualities of all prompting techniques, reinforcing the effectiveness of our method in enhancing both answer prediction and rationale qualities. When employing the original CoT prompting for questions with lengthy contexts, models frequently recycle sentences from the provided context and struggle to deliver a conclusive answer to the question. This trend is mitigated by our approach, resulting in a significant decrease in error rates. It also surpasses the performance of CoT prompting + SC and is on par with standard prompting in terms of EM and F1. For the answer inclusion score, LM-guided CoT prompting is slightly higher (1-2%) than standard prompting. Furthermore, LM-guided CoT prompting + SC achieves the highest performance across all settings.
**LM-guided CoT ์„ฑ๋Šฅ.** ํ‘œ 2๋Š” ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์„ ์ข…ํ•ฉ์ ์œผ๋กœ ๋ถ„์„ํ•œ ํ‘œ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์ƒ˜ํ”Œ๋งํ•œ ๋‹ค์Œ ๊ฐ€์žฅ ์ตœ์ ์˜ ๊ทผ๊ฑฐ๋ฅผ ์ˆœ์œ„ ๋งค๊ธฐ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํ™•์žฅ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. KD๋งŒ ์‚ฌ์šฉํ•œ ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ HotpotQA์˜ ๊ฒฝ์šฐ 2%, 2WikiMultiHopQA์˜ ๊ฒฝ์šฐ 10%๋กœ ๊ฐ๊ฐ ์›๋ž˜์˜ CoT prompting๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2๋Š” ๋ชจ๋“  prompting ๊ธฐ๋ฒ•์˜ ๊ฐ๊ฐ์˜ ๊ทผ๊ฑฐ ํ’ˆ์งˆ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ๋‹ต๋ณ€ ์˜ˆ์ธก๊ณผ ๊ทผ๊ฑฐ ํ’ˆ์งˆ์„ ๋ชจ๋‘ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ํšจ๊ณผ๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๊ธด ๋ฌธ๋งฅ์ด ์žˆ๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•ด ์›๋ž˜์˜ CoT prompting์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ, ๋ชจ๋ธ์€ ์ œ๊ณต๋œ ๋ฌธ๋งฅ์—์„œ ๋ฌธ์žฅ์„ ์žฌํ™œ์šฉํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๊ฒฐ์ •์ ์ธ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถ”์„ธ๋Š” ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์— ์˜ํ•ด ์™„ํ™”๋˜์–ด ์˜ค๋ฅ˜์œจ์ด ํฌ๊ฒŒ ๊ฐ์†Œํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ CoT prompting + SC์˜ ์„ฑ๋Šฅ์„ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ EM ๋ฐ F1 ์ธก๋ฉด์—์„œ ํ‘œ์ค€ prompting๊ณผ ๋™๋“ฑํ•ฉ๋‹ˆ๋‹ค. ๋‹ต๋ณ€ ํฌํ•จ ์ ์ˆ˜์˜ ๊ฒฝ์šฐ LM-guided CoT prompting์ด ํ‘œ์ค€ prompting๋ณด๋‹ค ์•ฝ๊ฐ„ ๋†’์Šต๋‹ˆ๋‹ค(1~2%). ๋˜ํ•œ LM-guided CoT prompting + SC๋Š” ๋ชจ๋“  ์„ค์ •์—์„œ ๊ฐ€์žฅ ๋†’์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
**LM ๊ฐ€์ด๋“œ CoT ์„ฑ๋Šฅ.** ํ‘œ 2๋Š” ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์„ ์ข…ํ•ฉ์ ์œผ๋กœ ๋ถ„์„ํ•œ ํ‘œ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์ƒ˜ํ”Œ๋งํ•œ ๋‹ค์Œ ๊ฐ€์žฅ ์ตœ์ ์˜ ๊ทผ๊ฑฐ๋ฅผ ์ˆœ์œ„ ๋งค๊ธฐ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํ™•์žฅ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. KD๋งŒ ์‚ฌ์šฉํ•œ ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ HotpotQA์˜ ๊ฒฝ์šฐ 2%, 2WikiMultiHopQA์˜ ๊ฒฝ์šฐ 10%๋กœ ๊ฐ๊ฐ ์›๋ž˜์˜ CoT ํ”„๋กฌํ”„ํŠธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2๋Š” ๋ชจ๋“  ํ”„๋กฌํ”„ํŠธ ๊ธฐ๋ฒ•์˜ ๊ฐ๊ฐ์˜ ๊ทผ๊ฑฐ ํ’ˆ์งˆ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ๋‹ต๋ณ€ ์˜ˆ์ธก๊ณผ ๊ทผ๊ฑฐ ํ’ˆ์งˆ์„ ๋ชจ๋‘ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ํšจ๊ณผ๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๊ธด ๋ฌธ๋งฅ์ด ์žˆ๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•ด ์›๋ž˜์˜ CoT ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ, ๋ชจ๋ธ์€ ์ œ๊ณต๋œ ๋ฌธ๋งฅ์—์„œ ๋ฌธ์žฅ์„ ์žฌํ™œ์šฉํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๊ฒฐ์ •์ ์ธ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถ”์„ธ๋Š” ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์— ์˜ํ•ด ์™„ํ™”๋˜์–ด ์˜ค๋ฅ˜์œจ์ด ํฌ๊ฒŒ ๊ฐ์†Œํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ CoT ํ”„๋กฌํ”„ํŠธ + SC์˜ ์„ฑ๋Šฅ์„ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ EM ๋ฐ F1 ์ธก๋ฉด์—์„œ ํ‘œ์ค€ ํ”„๋กฌํ”„ํŠธ์™€ ๋™๋“ฑํ•ฉ๋‹ˆ๋‹ค. ๋‹ต๋ณ€ ํฌํ•จ ์ ์ˆ˜์˜ ๊ฒฝ์šฐ LM ๊ฐ€์ด๋“œ CoT ํ”„๋กฌํ”„ํŠธ๊ฐ€ ํ‘œ์ค€ ํ”„๋กฌํ”„ํŠธ๋ณด๋‹ค ์•ฝ๊ฐ„ ๋†’์Šต๋‹ˆ๋‹ค(1~2%). ๋˜ํ•œ LM ๊ฐ€์ด๋“œ CoT ํ”„๋กฌํ”„ํŠธ + SC๋Š” ๋ชจ๋“  ์„ค์ •์—์„œ ๊ฐ€์žฅ ๋†’์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
Experimental resultsAs shown in Tab. 1, the relatively low performance of the original Llama-2 model (9.5%/17.2% for 7B/70B) indicates that most knowledge in the test documents is not included in the original pre-training corpus. After continued pre-training on documents, performances increase to 27.2%/41.7%, indicating that LLMs can absorb some amount of knowledge. Instruction-tuning further increases the performance to 30.3%/46.4%, confirming the effectiveness of this standard recipe. This observation is different from Zhu and Li (2023), which demonstrates that instruction-tuning after pre-training is ineffective on a randomly initialized GPT-2-like transformer. The difference probably arises because Llama-2, through its pre-training on diverse corpora comprising raw documents and QA data, has developed a certain degree of proficiency in extracting knowledge from its parameters via questions. We also report the performance where the corresponding document is directly provided to Llama-2 as context ("open-book w/ doc" in Tab. 1). The significant gap between closed-book and open-book settings suggests that retrieving knowledge from the parameters of LLMs is still challenging.
์‹คํ—˜ ๊ฒฐ๊ณผํ‘œ 1๊ณผ ๊ฐ™์ด, ์›๋ž˜ Llama-2 ๋ชจ๋ธ์˜ ์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ์€ ์„ฑ๋Šฅ(7B/70B์˜ ๊ฒฝ์šฐ 9.5%/17.2%)์€ ํ…Œ์ŠคํŠธ ๋ฌธ์„œ์˜ ๋Œ€๋ถ€๋ถ„์˜ ์ง€์‹์ด ์›๋ž˜์˜ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์— ํฌํ•จ๋˜์–ด ์žˆ์ง€ ์•Š์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋ฌธ์„œ์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ํ›„ ์„ฑ๋Šฅ์€ 27.2%/41.7%๋กœ ์ฆ๊ฐ€ํ•˜์—ฌ LLM์ด ์–ด๋Š ์ •๋„์˜ ์ง€์‹์„ ํก์ˆ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. Instruction-tuning์€ ์„ฑ๋Šฅ์„ 30.3%/46.4%๋กœ ๋”์šฑ ํ–ฅ์ƒ์‹œ์ผœ ์ด ํ‘œ์ค€ ๋ ˆ์‹œํ”ผ์˜ ํšจ๊ณผ๋ฅผ ํ™•์ธ์‹œ์ผœ ์ค๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ์‚ฌ์ „ ํ•™์Šต ํ›„ Instruction-tuning์ด ๋ฌด์ž‘์œ„๋กœ ์ดˆ๊ธฐํ™”๋œ GPT-2์™€ ์œ ์‚ฌํ•œ Transformer์— ํšจ๊ณผ์ ์ด์ง€ ์•Š๋‹ค๋Š” ๊ฒƒ์„ ์ž…์ฆํ•œ Zhu์™€ Li(2023)์™€๋Š” ๋‹ค๋ฅธ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฐจ์ด๋Š” ์›๋ฌธ ๋ฌธ์„œ์™€ QA ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ ๋‹ค์–‘ํ•œ ์ฝ”ํผ๋ผ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด Llama-2๊ฐ€ ์งˆ๋ฌธ์„ ํ†ตํ•ด ํŒŒ๋ผ๋ฏธํ„ฐ์—์„œ ์ง€์‹์„ ์ถ”์ถœํ•˜๋Š” ๋ฐ ์–ด๋Š ์ •๋„ ์ˆ™๋ จ๋„๋ฅผ ๊ฐ–์ถ”์—ˆ๊ธฐ ๋•Œ๋ฌธ์ผ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•ด๋‹น ๋ฌธ์„œ๋ฅผ Llama-2์— ๋ฌธ๋งฅ์œผ๋กœ ์ง์ ‘ ์ œ๊ณตํ•˜๋Š” ์„ฑ๋Šฅ("ํ‘œ 1์˜ open-book w/ doc")๋„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์•”๊ธฐํ˜• ์‹œํ—˜๊ณผ open-book ์‹œํ—˜ ์‚ฌ์ด์˜ ํฐ ์ฐจ์ด๋Š” LLM์˜ ํŒŒ๋ผ๋ฏธํ„ฐ์—์„œ ์ง€์‹์„ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒƒ์ด ์—ฌ์ „ํžˆ ์–ด๋ ต๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
์‹คํ—˜ ๊ฒฐ๊ณผํ‘œ 1๊ณผ ๊ฐ™์ด, ์›๋ž˜ Llama-2 ๋ชจ๋ธ์˜ ์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ์€ ์„ฑ๋Šฅ(7B/70B์˜ ๊ฒฝ์šฐ 9.5%/17.2%)์€ ํ…Œ์ŠคํŠธ ๋ฌธ์„œ์˜ ๋Œ€๋ถ€๋ถ„์˜ ์ง€์‹์ด ์›๋ž˜์˜ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์— ํฌํ•จ๋˜์–ด ์žˆ์ง€ ์•Š์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋ฌธ์„œ์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ํ›„ ์„ฑ๋Šฅ์€ 27.2%/41.7%๋กœ ์ฆ๊ฐ€ํ•˜์—ฌ LLM์ด ์–ด๋Š ์ •๋„์˜ ์ง€์‹์„ ํก์ˆ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋ช…๋ น์–ด ํŠœ๋‹์€ ์„ฑ๋Šฅ์„ 30.3%/46.4%๋กœ ๋”์šฑ ํ–ฅ์ƒ์‹œ์ผœ ์ด ํ‘œ์ค€ ๋ ˆ์‹œํ”ผ์˜ ํšจ๊ณผ๋ฅผ ํ™•์ธ์‹œ์ผœ ์ค๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ์‚ฌ์ „ ํ•™์Šต ํ›„ ๋ช…๋ น์–ด ํŠœ๋‹์ด ๋ฌด์ž‘์œ„๋กœ ์ดˆ๊ธฐํ™”๋œ GPT-2์™€ ์œ ์‚ฌํ•œ ํŠธ๋žœ์Šคํฌ๋จธ์— ํšจ๊ณผ์ ์ด์ง€ ์•Š๋‹ค๋Š” ๊ฒƒ์„ ์ž…์ฆํ•œ Zhu์™€ Li(2023)์™€๋Š” ๋‹ค๋ฅธ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฐจ์ด๋Š” ์›๋ฌธ ๋ฌธ์„œ์™€ QA ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ ๋‹ค์–‘ํ•œ ์ฝ”ํผ๋ผ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด Llama-2๊ฐ€ ์งˆ๋ฌธ์„ ํ†ตํ•ด ํŒŒ๋ผ๋ฏธํ„ฐ์—์„œ ์ง€์‹์„ ์ถ”์ถœํ•˜๋Š” ๋ฐ ์–ด๋Š ์ •๋„ ์ˆ™๋ จ๋„๋ฅผ ๊ฐ–์ถ”์—ˆ๊ธฐ ๋•Œ๋ฌธ์ผ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•ด๋‹น ๋ฌธ์„œ๋ฅผ Llama-2์— ๋ฌธ๋งฅ์œผ๋กœ ์ง์ ‘ ์ œ๊ณตํ•˜๋Š” ์„ฑ๋Šฅ("ํ‘œ 1์˜ ์˜คํ”ˆ๋ถ w/ doc")๋„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์•”๊ธฐํ˜• ์‹œํ—˜๊ณผ ์˜คํ”ˆ๋ถ ์‹œํ—˜ ์‚ฌ์ด์˜ ํฐ ์ฐจ์ด๋Š” LLM์˜ ํŒŒ๋ผ๋ฏธํ„ฐ์—์„œ ์ง€์‹์„ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒƒ์ด ์—ฌ์ „ํžˆ ์–ด๋ ต๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
Exploratory evaluationExploratory evaluation involves open-ended, ad-hoc evaluation of models for novel, unknown, or poorly understood risks. It is well-suited to testing more complex interactions with models, such as multi-turn conversations and use of agents, and is particularly important for assessing frontier models. Red teaming, which has become one of the most popular ways of assessing safety risks, is a form of exploratory evaluation. It involves tasking annotators and experts with probing a model-in-the-loop to identify flaws and vulnerabilities [156]. Red teaming can be implemented both using humans (as with the OpenAI Red Teaming Network29) and AI models [35, 51, 52, 66]. It is very flexible, and a core focus has been understanding susceptibility to being manipulated, persuaded, directed or encouraged to give hazardous responses (often called jailbreaking, prompt injecting, or adversarially attacking) [157, 158, 159]. In 2023, a large-scale red teaming effort organized at the DefCon hacker's conference, which involved over \(2,200\) people, identified numerous model weaknesses, developed hazard categories, and identified effective strategies for red teaming [160].
ํƒ์ƒ‰์  ํ‰๊ฐ€ํƒ์ƒ‰์  ํ‰๊ฐ€์—๋Š” ์ƒˆ๋กญ๊ฑฐ๋‚˜ ์•Œ๋ ค์ง€์ง€ ์•Š์•˜๊ฑฐ๋‚˜ ์ž˜ ์ดํ•ด๋˜์ง€ ์•Š์€ ์œ„ํ—˜์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ๊ฐœ๋ฐฉํ˜•, ์ž„์‹œ ํ‰๊ฐ€๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ multi-turn conversations ๋ฐ ์—์ด์ „ํŠธ ์‚ฌ์šฉ๊ณผ ๊ฐ™์€ ๋ชจ๋ธ๊ณผ์˜ ๋ณต์žกํ•œ ์ƒํ˜ธ ์ž‘์šฉ์„ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ ์ ํ•ฉํ•˜๋ฉฐ, ํŠนํžˆ frontier ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์•ˆ์ „ ์œ„ํ—˜์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ๋ฐฉ๋ฒ• ์ค‘ ํ•˜๋‚˜๊ฐ€ ๋œ red teaming์€ ํƒ์ƒ‰์  ํ‰๊ฐ€์˜ ํ•œ ํ˜•ํƒœ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์ฃผ์„๊ฐ€์™€ ์ „๋ฌธ๊ฐ€์—๊ฒŒ ๋ชจ๋ธ์„ ํ…Œ์ŠคํŠธํ•˜์—ฌ ๊ฒฐํ•จ๊ณผ ์ทจ์•ฝ์„ฑ์„ ํŒŒ์•…ํ•˜๋Š” ์ž‘์—…์„ ๋งก๊น๋‹ˆ๋‹ค[156]. red teaming์€ ์ธ๊ฐ„(OpenAI Red Teaming Network29)๊ณผ AI ๋ชจ๋ธ[35, 51, 52, 66]์„ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜์—ฌ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋งค์šฐ ์œ ์—ฐํ•˜๋ฉฐ, ์œ„ํ—˜ํ•œ ์‘๋‹ต์„ ์ œ๊ณตํ•˜๋„๋ก ์กฐ์ž‘, ์„ค๋“, ์ง€์‹œ ๋˜๋Š” ๊ถŒ์œ ๋ฐ›๋Š” ์ทจ์•ฝ์„ฑ์„ ํŒŒ์•…ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(ํ”ํžˆ jailbreaking, prompt injecting ๋˜๋Š” adversarially attacking์ด๋ผ๊ณ  ํ•จ)[157, 158, 159]. 2023๋…„์—๋Š” 2,200๋ช… ์ด์ƒ์˜ ์‚ฌ๋žŒ๋“ค์ด ์ฐธ์—ฌํ•œ DefCon ํ•ด์ปค ์ปจํผ๋Ÿฐ์Šค์—์„œ ๋Œ€๊ทœ๋ชจ red teaming ํ™œ๋™์„ ํ†ตํ•ด ์ˆ˜๋งŽ์€ ๋ชจ๋ธ์˜ ์•ฝ์ ์„ ํŒŒ์•…ํ•˜๊ณ  ์œ„ํ—˜ ๋ฒ”์ฃผ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์œผ๋ฉฐ, red teaming์„ ์œ„ํ•œ ํšจ๊ณผ์ ์ธ ์ „๋žต์„ ํŒŒ์•…ํ–ˆ์Šต๋‹ˆ๋‹ค[160].
ํƒ์ƒ‰์  ํ‰๊ฐ€ํƒ์ƒ‰์  ํ‰๊ฐ€์—๋Š” ์ƒˆ๋กญ๊ฑฐ๋‚˜ ์•Œ๋ ค์ง€์ง€ ์•Š์•˜๊ฑฐ๋‚˜ ์ž˜ ์ดํ•ด๋˜์ง€ ์•Š์€ ์œ„ํ—˜์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ๊ฐœ๋ฐฉํ˜•, ์ž„์‹œ ํ‰๊ฐ€๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋‹ค์ค‘ ํ„ด ๋Œ€ํ™” ๋ฐ ์—์ด์ „ํŠธ ์‚ฌ์šฉ๊ณผ ๊ฐ™์€ ๋ชจ๋ธ๊ณผ์˜ ๋ณต์žกํ•œ ์ƒํ˜ธ ์ž‘์šฉ์„ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ ์ ํ•ฉํ•˜๋ฉฐ, ํŠนํžˆ ํ”„๋ก ํ‹ฐ์–ด ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์•ˆ์ „ ์œ„ํ—˜์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ๋ฐฉ๋ฒ• ์ค‘ ํ•˜๋‚˜๊ฐ€ ๋œ ๋ ˆ๋“œ ํŒ€์€ ํƒ์ƒ‰์  ํ‰๊ฐ€์˜ ํ•œ ํ˜•ํƒœ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์ฃผ์„๊ฐ€์™€ ์ „๋ฌธ๊ฐ€์—๊ฒŒ ๋ชจ๋ธ์„ ํ…Œ์ŠคํŠธํ•˜์—ฌ ๊ฒฐํ•จ๊ณผ ์ทจ์•ฝ์„ฑ์„ ํŒŒ์•…ํ•˜๋Š” ์ž‘์—…์„ ๋งก๊น๋‹ˆ๋‹ค[156]. ๋ ˆ๋“œ ํŒ€์€ ์ธ๊ฐ„(OpenAI ๋ ˆ๋“œ ํŒ€ ๋„คํŠธ์›Œํฌ29)๊ณผ AI ๋ชจ๋ธ[35, 51, 52, 66]์„ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜์—ฌ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋งค์šฐ ์œ ์—ฐํ•˜๋ฉฐ, ์œ„ํ—˜ํ•œ ์‘๋‹ต์„ ์ œ๊ณตํ•˜๋„๋ก ์กฐ์ž‘, ์„ค๋“, ์ง€์‹œ ๋˜๋Š” ๊ถŒ์œ ๋ฐ›๋Š” ์ทจ์•ฝ์„ฑ์„ ํŒŒ์•…ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(ํ”ํžˆ ํƒˆ์˜ฅ, ํ”„๋กฌํ”„ํŠธ ์‚ฝ์ž… ๋˜๋Š” ์ ๋Œ€์  ๊ณต๊ฒฉ์ด๋ผ๊ณ  ํ•จ)[157, 158, 159]. 2023๋…„์—๋Š” 2,200๋ช… ์ด์ƒ์˜ ์‚ฌ๋žŒ๋“ค์ด ์ฐธ์—ฌํ•œ DefCon ํ•ด์ปค ์ปจํผ๋Ÿฐ์Šค์—์„œ ๋Œ€๊ทœ๋ชจ ๋ ˆ๋“œ ํŒ€ ํ™œ๋™์„ ํ†ตํ•ด ์ˆ˜๋งŽ์€ ๋ชจ๋ธ์˜ ์•ฝ์ ์„ ํŒŒ์•…ํ•˜๊ณ  ์œ„ํ—˜ ๋ฒ”์ฃผ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์œผ๋ฉฐ, ๋ ˆ๋“œ ํŒ€์„ ์œ„ํ•œ ํšจ๊ณผ์ ์ธ ์ „๋žต์„ ํŒŒ์•…ํ–ˆ์Šต๋‹ˆ๋‹ค[160].
**Finetuning of LLMs** Recent years have seen rapid progress in developing large-scale language models (LLMs) Brown et al. (2020); OpenAI (2023); Workshop et al. (2022); Touvron et al. (2023); Anil et al. (2023). To adapt these foundation models to downstream tasks, fine-tuning Mishra et al. (2021); Sanh et al. (2021); Chung et al. (2022); Muenighoff et al. (2023); Zhou et al. (2023); Lin et al. (2023); Ji et al. (2024) has become a prevalent approach. Traditional supervised fine-tuning may be limited by the cost and compute required for adapating LLMs. Addressing these challenges, research in the realm of parameter-efficient fine-tuning Houlsby et al. (2019), such as Prompt Tuning Lester et al. (2021), Prefix-Tuning Li and Liang (2021), P-Tuning Liu et al. (2022) and Low-Rank based fine-tuning Hu et al. (2021), has gained traction. These methods enable LLMs to acquire domain-specific knowledge and adapt to specialized tasks such as question answering, summarization, and dialogue generation. Another branch of finetuning is through RLHF Ouyang et al. (2022); Rafailov et al. (2023); Liu et al. (2023); Zhang et al. (2023), which adopts RL to align LLM's preference with human.
**LLM์˜ Fine-tuning** ์ตœ๊ทผ ๋ช‡ ๋…„ ๋™์•ˆ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ๋น ๋ฅธ ์ง„์ „์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. Brown et al. (2020); OpenAI (2023); ์›Œํฌ์ˆ ๋“ฑ(2022); Touvron et al. (2023); Anil et al. (2023). ์ด๋Ÿฌํ•œ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์„ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋งž๊ฒŒ ์กฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด Fine-tuning Mishra et al. (2021); Sanh et al. (2021); Chung et al. (2022); Muenighoff et al. (2023); Zhou et al. (2023); Lin et al. (2023); Ji et al. (2024)์ด ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด์˜ Supervised Fine-tuning์€ LLM์„ ์กฐ์ •ํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ๋น„์šฉ๊ณผ ๊ณ„์‚ฐ์œผ๋กœ ์ธํ•ด ์ œํ•œ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด Prompt Tuning Lester et al. (2021), Prefix-Tuning Li and Liang (2021), P-Tuning Liu et al. (2022), ์ €์ˆœ์œ„ ๊ธฐ๋ฐ˜ Fine-tuning Hu et al. (2021)๊ณผ ๊ฐ™์€ ํŒŒ๋ผ๋ฏธํ„ฐ ํšจ์œจ์ ์ธ Fine-tuning Houlsby et al. (2019)์˜ ์˜์—ญ์—์„œ ์—ฐ๊ตฌ๊ฐ€ ์ฃผ๋ชฉ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด LLM์€ ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ์Šต๋“ํ•˜๊ณ  ์งˆ๋ฌธ ๋‹ต๋ณ€, ์š”์•ฝ, ๋Œ€ํ™” ์ƒ์„ฑ๊ณผ ๊ฐ™์€ ํŠน์ˆ˜ํ•œ ์ž‘์—…์— ์ ์‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Fine-tuning์˜ ๋˜ ๋‹ค๋ฅธ ๋ถ„์•ผ๋Š” RLHF Ouyang et al. (2022); Rafailov et al. (2023); Liu et al. (2023); Zhang et al. (2023)์€ Reinforcement Learning์„ ์ฑ„ํƒํ•˜์—ฌ LLM์˜ ์„ ํ˜ธ๋„๋ฅผ ์‚ฌ๋žŒ๊ณผ ์ผ์น˜์‹œํ‚ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
**LLM์˜ ๋ฏธ์„ธ ์กฐ์ •** ์ตœ๊ทผ ๋ช‡ ๋…„ ๋™์•ˆ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ๋น ๋ฅธ ์ง„์ „์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. Brown et al. (2020); OpenAI (2023); ์›Œํฌ์ˆ ๋“ฑ(2022); Touvron et al. (2023); Anil et al. (2023). ์ด๋Ÿฌํ•œ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์„ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋งž๊ฒŒ ์กฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด ๋ฏธ์„ธ ์กฐ์ • Mishra et al. (2021); Sanh et al. (2021); Chung et al. (2022); Muenighoff et al. (2023); Zhou et al. (2023); Lin et al. (2023); Ji et al. (2024)์ด ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด์˜ ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •์€ LLM์„ ์กฐ์ •ํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ๋น„์šฉ๊ณผ ๊ณ„์‚ฐ์œผ๋กœ ์ธํ•ด ์ œํ•œ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋กฌํ”„ํŠธ ํŠœ๋‹ Lester et al. (2021), ์ ‘๋‘์‚ฌ ํŠœ๋‹ Li and Liang (2021), P-ํŠœ๋‹ Liu et al. (2022), ์ €์ˆœ์œ„ ๊ธฐ๋ฐ˜ ๋ฏธ์„ธ ์กฐ์ • Hu et al. (2021)๊ณผ ๊ฐ™์€ ํŒŒ๋ผ๋ฏธํ„ฐ ํšจ์œจ์ ์ธ ๋ฏธ์„ธ ์กฐ์ • Houlsby et al. (2019)์˜ ์˜์—ญ์—์„œ ์—ฐ๊ตฌ๊ฐ€ ์ฃผ๋ชฉ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด LLM์€ ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ์Šต๋“ํ•˜๊ณ  ์งˆ๋ฌธ ๋‹ต๋ณ€, ์š”์•ฝ, ๋Œ€ํ™” ์ƒ์„ฑ๊ณผ ๊ฐ™์€ ํŠน์ˆ˜ํ•œ ์ž‘์—…์— ์ ์‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฏธ์„ธ ์กฐ์ •์˜ ๋˜ ๋‹ค๋ฅธ ๋ถ„์•ผ๋Š” RLHF Ouyang et al. (2022); Rafailov et al. (2023); Liu et al. (2023); Zhang et al. (2023)์€ RL์„ ์ฑ„ํƒํ•˜์—ฌ LLM์˜ ์„ ํ˜ธ๋„๋ฅผ ์‚ฌ๋žŒ๊ณผ ์ผ์น˜์‹œํ‚ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Comparison against baselines with retrieval.As shown in Tables 2 (bottom), our Self-Rag also outperforms existing RAG in many tasks, obtaining the best performance among non-proprietary LM-based models on all tasks. While our method outperforms other baselines, on PopQA or Bio, powerful instruction-tuned LMs with retrieval (e.g., LLama2-chat, Alpaca) show large gains from their non-retrieval baselines. However, we found that these baselines provide limited solutions for tasks where we cannot simply copy or extract sub-strings of retrieved passages. On PubHealth and ARC-Challenge, baselines with retrieval do not improve performance notably from their no-retrieval counterparts. We also observe that most baselines with retrieval struggle to improve citation accuracy. On ASQA, our model shows significantly higher citation precision and recall than all models except ChatGPT. Gao et al. (2023) found that ChatGPT consistently exhibits superior efficacy in this particular task, surpassing smaller LMs. Our Self-Rag bridges this performance gap, even outperforming ChatGPT in citation precision, which measures whether the model-generated claim is fully supported by cited evidence.
๊ฒ€์ƒ‰์ด ์žˆ๋Š” baseline๊ณผ์˜ ๋น„๊ต. ํ‘œ 2(ํ•˜๋‹จ)์™€ ๊ฐ™์ด, ์šฐ๋ฆฌ์˜ Self-Rag๋„ ๋งŽ์€ ์ž‘์—…์—์„œ ๊ธฐ์กด RAG๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ, ๋ชจ๋“  ์ž‘์—…์—์„œ ๋น„๋…์  LM ๊ธฐ๋ฐ˜ ๋ชจ๋ธ ์ค‘ ์ตœ๊ณ ์˜ ์„ฑ๋Šฅ์„ ์–ป์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์ด ๋‹ค๋ฅธ baseline๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜์ง€๋งŒ, PopQA ๋˜๋Š” Bio์—์„œ๋Š” ๊ฒ€์ƒ‰ ๊ธฐ๋Šฅ์ด ์žˆ๋Š” ๊ฐ•๋ ฅํ•œ instruction-tuned LMs(์˜ˆ: LLama2-chat, Alpaca)์ด ๊ฒ€์ƒ‰์ด ์—†๋Š” baseline๋ณด๋‹ค ํฐ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ baseline์€ ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์˜ ํ•˜์œ„ ๋ฌธ์ž์—ด์„ ๋‹จ์ˆœํžˆ ๋ณต์‚ฌํ•˜๊ฑฐ๋‚˜ ์ถ”์ถœํ•  ์ˆ˜ ์—†๋Š” ์ž‘์—…์—๋Š” ์ œํ•œ๋œ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. PubHealth ๋ฐ ARC-Challenge์—์„œ๋Š” ๊ฒ€์ƒ‰ ๊ธฐ๋Šฅ์ด ์žˆ๋Š” baseline์ด ๊ฒ€์ƒ‰ ๊ธฐ๋Šฅ์ด ์—†๋Š” baseline์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฒ€์ƒ‰ ๊ธฐ๋Šฅ์ด ์žˆ๋Š” ๋Œ€๋ถ€๋ถ„์˜ baseline์ด ์ธ์šฉ ์ •ํ™•๋„๋ฅผ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์Œ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ASQA์—์„œ ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ChatGPT๋ฅผ ์ œ์™ธํ•œ ๋ชจ๋“  ๋ชจ๋ธ๋ณด๋‹ค ์ธ์šฉ ์ •๋ฐ€๋„์™€ ํšŒ์ˆ˜์œจ์ด ํ›จ์”ฌ ๋†’์Šต๋‹ˆ๋‹ค. Gao ๋“ฑ(2023)์€ ChatGPT๊ฐ€ ์ด ํŠน์ • ์ž‘์—…์—์„œ ๋” ์ž‘์€ LMs์„ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ผ๊ด€๋œ ์šฐ์ˆ˜ํ•œ ํšจ๋Šฅ์„ ๋ณด์ธ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ Self-Rag์€ ์ด๋Ÿฌํ•œ ์„ฑ๋Šฅ ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜์—ฌ ์ธ์šฉ ์ •๋ฐ€๋„์—์„œ ChatGPT๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š”๋ฐ, ์ด๋Š” ๋ชจ๋ธ์ด ์ƒ์„ฑํ•œ ์ฃผ์žฅ์ด ์ธ์šฉ๋œ ์ฆ๊ฑฐ์— ์˜ํ•ด ์™„์ „ํžˆ ๋’ท๋ฐ›์นจ๋˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค.
๊ฒ€์ƒ‰์ด ์žˆ๋Š” ๋ฒ ์ด์Šค๋ผ์ธ๊ณผ์˜ ๋น„๊ต.ํ‘œ 2(ํ•˜๋‹จ)์™€ ๊ฐ™์ด, ์šฐ๋ฆฌ์˜ Self-Rag๋„ ๋งŽ์€ ์ž‘์—…์—์„œ ๊ธฐ์กด RAG๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ, ๋ชจ๋“  ์ž‘์—…์—์„œ ๋น„๋…์  LM ๊ธฐ๋ฐ˜ ๋ชจ๋ธ ์ค‘ ์ตœ๊ณ ์˜ ์„ฑ๋Šฅ์„ ์–ป์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์ด ๋‹ค๋ฅธ ๋ฒ ์ด์Šค๋ผ์ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜์ง€๋งŒ, PopQA ๋˜๋Š” Bio์—์„œ๋Š” ๊ฒ€์ƒ‰ ๊ธฐ๋Šฅ์ด ์žˆ๋Š” ๊ฐ•๋ ฅํ•œ ์ง€์นจ ์กฐ์ • LM(์˜ˆ: LLama2-chat, ์•ŒํŒŒ์นด)์ด ๊ฒ€์ƒ‰์ด ์—†๋Š” ๋ฒ ์ด์Šค๋ผ์ธ๋ณด๋‹ค ํฐ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๊ธฐ๋ณธ ๋ผ์ธ์€ ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์˜ ํ•˜์œ„ ๋ฌธ์ž์—ด์„ ๋‹จ์ˆœํžˆ ๋ณต์‚ฌํ•˜๊ฑฐ๋‚˜ ์ถ”์ถœํ•  ์ˆ˜ ์—†๋Š” ์ž‘์—…์—๋Š” ์ œํ•œ๋œ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. PubHealth ๋ฐ ARC-Challenge์—์„œ๋Š” ๊ฒ€์ƒ‰ ๊ธฐ๋Šฅ์ด ์žˆ๋Š” ๊ธฐ๋ณธ ๋ผ์ธ์ด ๊ฒ€์ƒ‰ ๊ธฐ๋Šฅ์ด ์—†๋Š” ๊ธฐ๋ณธ ๋ผ์ธ์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฒ€์ƒ‰ ๊ธฐ๋Šฅ์ด ์žˆ๋Š” ๋Œ€๋ถ€๋ถ„์˜ ๊ธฐ๋ณธ ๋ผ์ธ์ด ์ธ์šฉ ์ •ํ™•๋„๋ฅผ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์Œ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ASQA์—์„œ ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ChatGPT๋ฅผ ์ œ์™ธํ•œ ๋ชจ๋“  ๋ชจ๋ธ๋ณด๋‹ค ์ธ์šฉ ์ •๋ฐ€๋„์™€ ํšŒ์ˆ˜์œจ์ด ํ›จ์”ฌ ๋†’์Šต๋‹ˆ๋‹ค. Gao ๋“ฑ(2023)์€ ChatGPT๊ฐ€ ์ด ํŠน์ • ์ž‘์—…์—์„œ ๋” ์ž‘์€ LM์„ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ผ๊ด€๋œ ์šฐ์ˆ˜ํ•œ ํšจ๋Šฅ์„ ๋ณด์ธ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ Self-Rag์€ ์ด๋Ÿฌํ•œ ์„ฑ๋Šฅ ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜์—ฌ ์ธ์šฉ ์ •๋ฐ€๋„์—์„œ ChatGPT๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š”๋ฐ, ์ด๋Š” ๋ชจ๋ธ์ด ์ƒ์„ฑํ•œ ์ฃผ์žฅ์ด ์ธ์šฉ๋œ ์ฆ๊ฑฐ์— ์˜ํ•ด ์™„์ „ํžˆ ๋’ท๋ฐ›์นจ๋˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค.
Training scale constitutes another significant factor influencing the transferability of LLM capabilities, composed of both pretraining scale and instruction tuning scale. Experimental results are shown in table 1. Taking the example of LLaMA (with 10K, 100K, and 1M further pretrain) and Open Chinese LLaMA, the scale of further Chinese pretraining gradually increases from 0 to 100 billion tokens. Under the settings of 1K and 5K instruction tuning, we observed that the response quality improves progressively with the increase in the scale of further pretraining. 1 However, when the instruction tuning data scale escalates to 950K, we find no significant differences in response quality among the models. Consequently, we hypothesize that more further pretraining could accelerate the model's alignment with human instructions, but the mere tens of billions in training scale are insufficient to enable the model to grasp a greater amount of world knowledge. This leads to their convergence at similar response levels. In other words, the enhancement in response quality primarily stems from an improvement in language generation prowess rather than an elevation in knowledge level.
ํ›ˆ๋ จ ๊ทœ๋ชจ๋Š” ์‚ฌ์ „ ํ•™์Šต ๊ทœ๋ชจ์™€ Instruction Tuning ๊ทœ๋ชจ๋กœ ๊ตฌ์„ฑ๋œ LLM ๊ธฐ๋Šฅ์˜ ์ „์ด์„ฑ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋˜ ๋‹ค๋ฅธ ์ค‘์š”ํ•œ ์š”์†Œ์ž…๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. LLaMA(10K, 100K, 1M ์ถ”๊ฐ€ ์‚ฌ์ „ ํ•™์Šต ํฌํ•จ)์™€ Open Chinese LLaMA๋ฅผ ์˜ˆ๋กœ ๋“ค๋ฉด, ์ถ”๊ฐ€ ์ค‘๊ตญ์–ด ์‚ฌ์ „ ํ•™์Šต์˜ ๊ทœ๋ชจ๋Š” 0์—์„œ 1000์–ต ํ† ํฐ์œผ๋กœ ์ ์ง„์ ์œผ๋กœ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. 1K ๋ฐ 5K Instruction Tuning ์„ค์ •์—์„œ ์ถ”๊ฐ€ ์‚ฌ์ „ ํ•™์Šต ๊ทœ๋ชจ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์‘๋‹ต ํ’ˆ์งˆ์ด ์ ์ง„์ ์œผ๋กœ ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. 1 ๊ทธ๋Ÿฌ๋‚˜ Instruction Tuning ๋ฐ์ดํ„ฐ ๊ทœ๋ชจ๊ฐ€ 950K๋กœ ํ™•๋Œ€๋˜๋ฉด ๋ชจ๋ธ ๊ฐ„ ์‘๋‹ต ํ’ˆ์งˆ์— ํฐ ์ฐจ์ด๊ฐ€ ์—†๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋” ๋งŽ์€ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ๋ชจ๋ธ์ด ์ธ๊ฐ„์˜ ์ง€์‹œ์— ๋” ์ž˜ ๋งž์ถฐ์งˆ ์ˆ˜ ์žˆ์„ ๊ฒƒ์ด๋ผ๊ณ  ๊ฐ€์„ค์„ ์„ธ์› ์ง€๋งŒ, ๋‹จ ๋ช‡ ์‹ญ์–ต์˜ ํ•™์Šต ๊ทœ๋ชจ๋กœ๋Š” ๋ชจ๋ธ์ด ๋” ๋งŽ์€ ์„ธ๊ณ„ ์ง€์‹์„ ํŒŒ์•…ํ•˜๊ธฐ์—๋Š” ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋กœ ์ธํ•ด ๋น„์Šทํ•œ ์‘๋‹ต ์ˆ˜์ค€์—์„œ ์ˆ˜๋ ดํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์‘๋‹ต ํ’ˆ์งˆ์˜ ํ–ฅ์ƒ์€ ์ฃผ๋กœ ์ง€์‹ ์ˆ˜์ค€์˜ ํ–ฅ์ƒ๋ณด๋‹ค๋Š” ์–ธ์–ด ์ƒ์„ฑ ๋Šฅ๋ ฅ์˜ ํ–ฅ์ƒ์—์„œ ๋น„๋กฏ๋œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
ํ›ˆ๋ จ ๊ทœ๋ชจ๋Š” ์‚ฌ์ „ ํ•™์Šต ๊ทœ๋ชจ์™€ ์ง€์‹œ ํŠœ๋‹ ๊ทœ๋ชจ๋กœ ๊ตฌ์„ฑ๋œ LLM ๊ธฐ๋Šฅ์˜ ์ „์ด์„ฑ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋˜ ๋‹ค๋ฅธ ์ค‘์š”ํ•œ ์š”์†Œ์ž…๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. LLaMA(10K, 100K, 1M ์ถ”๊ฐ€ ์‚ฌ์ „ ํ•™์Šต ํฌํ•จ)์™€ Open Chinese LLaMA๋ฅผ ์˜ˆ๋กœ ๋“ค๋ฉด, ์ถ”๊ฐ€ ์ค‘๊ตญ์–ด ์‚ฌ์ „ ํ•™์Šต์˜ ๊ทœ๋ชจ๋Š” 0์—์„œ 1000์–ต ํ† ํฐ์œผ๋กœ ์ ์ง„์ ์œผ๋กœ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. 1K ๋ฐ 5K ๋ช…๋ น์–ด ํŠœ๋‹ ์„ค์ •์—์„œ ์ถ”๊ฐ€ ์‚ฌ์ „ ํ•™์Šต ๊ทœ๋ชจ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์‘๋‹ต ํ’ˆ์งˆ์ด ์ ์ง„์ ์œผ๋กœ ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. 1 ๊ทธ๋Ÿฌ๋‚˜ ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ์ดํ„ฐ ๊ทœ๋ชจ๊ฐ€ 950K๋กœ ํ™•๋Œ€๋˜๋ฉด ๋ชจ๋ธ ๊ฐ„ ์‘๋‹ต ํ’ˆ์งˆ์— ํฐ ์ฐจ์ด๊ฐ€ ์—†๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋” ๋งŽ์€ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ๋ชจ๋ธ์ด ์ธ๊ฐ„์˜ ์ง€์‹œ์— ๋” ์ž˜ ๋งž์ถฐ์งˆ ์ˆ˜ ์žˆ์„ ๊ฒƒ์ด๋ผ๊ณ  ๊ฐ€์„ค์„ ์„ธ์› ์ง€๋งŒ, ๋‹จ ๋ช‡ ์‹ญ์–ต์˜ ํ•™์Šต ๊ทœ๋ชจ๋กœ๋Š” ๋ชจ๋ธ์ด ๋” ๋งŽ์€ ์„ธ๊ณ„ ์ง€์‹์„ ํŒŒ์•…ํ•˜๊ธฐ์—๋Š” ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋กœ ์ธํ•ด ๋น„์Šทํ•œ ์‘๋‹ต ์ˆ˜์ค€์—์„œ ์ˆ˜๋ ดํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์‘๋‹ต ํ’ˆ์งˆ์˜ ํ–ฅ์ƒ์€ ์ฃผ๋กœ ์ง€์‹ ์ˆ˜์ค€์˜ ํ–ฅ์ƒ๋ณด๋‹ค๋Š” ์–ธ์–ด ์ƒ์„ฑ ๋Šฅ๋ ฅ์˜ ํ–ฅ์ƒ์—์„œ ๋น„๋กฏ๋œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
We provide the responses and a manual translation of the response in Table 9. While the relatively small 405M parameter models generally do not produce meaningful sentences, the model trained on German Common Crawl generally produces grammatically correct sentences. This includes correct upper and lower casing, the generation of compound words, and the use of Unlaut characters. In all cases, responses tend to be repetitive after a few words. The incompleteness and fractional words of some outputs can be attributed to the number of tokens generated being very short. Longer sequences tend to provide syntactically correct sentences but also tend to repeat themselves. The limited number of samples and the generally low quality of the generated text does not allow us to make robust statements of the qualitative differences between the individual models trained on German Common Crawl. However, the aforementioned problems of repetitiveness as well as grammatical errors seem to be significantly stronger on the Pile baseline, which generally fails to even respect the context given in the prompt. This is expected as the amount of German text is significantly smaller on the Pile.
ํ‘œ 9์— ์‘๋‹ต๊ณผ ์‘๋‹ต์˜ ์ˆ˜๋™ ๋ฒˆ์—ญ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ƒ๋Œ€์ ์œผ๋กœ ์ž‘์€ 405M ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์˜๋ฏธ ์žˆ๋Š” ๋ฌธ์žฅ์„ ์ƒ์„ฑํ•˜์ง€ ์•Š์ง€๋งŒ, ๋…์ผ์–ด Common Crawl๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฌธ๋ฒ•์ ์œผ๋กœ ์˜ฌ๋ฐ”๋ฅธ ๋ฌธ์žฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์˜ฌ๋ฐ”๋ฅธ ๋Œ€๋ฌธ์ž์™€ ์†Œ๋ฌธ์ž, ํ•ฉ์„ฑ์–ด ์ƒ์„ฑ, Unlaut ๋ฌธ์ž ์‚ฌ์šฉ ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๊ฒฝ์šฐ์—์„œ ์‘๋‹ต์€ ๋ช‡ ๋‹จ์–ด ์ดํ›„์— ๋ฐ˜๋ณต๋˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ์ถœ๋ ฅ์˜ ๋ถˆ์™„์ „์„ฑ ๋ฐ ๋ถ€๋ถ„ ๋‹จ์–ด๋Š” ์ƒ์„ฑ๋œ ํ† ํฐ์˜ ์ˆ˜๊ฐ€ ๋งค์šฐ ์งง๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋” ๊ธด ์‹œํ€€์Šค๋Š” ๊ตฌ๋ฌธ์ ์œผ๋กœ ์˜ฌ๋ฐ”๋ฅธ ๋ฌธ์žฅ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์ง€๋งŒ ์ž์‹ ์„ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ œํ•œ๋œ ์ƒ˜ํ”Œ ์ˆ˜์™€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ์˜ ํ’ˆ์งˆ์ด ๋‚ฎ๊ธฐ ๋•Œ๋ฌธ์— ๋…์ผ์–ด Common Crawl๋กœ ํ•™์Šต๋œ ๊ฐœ๋ณ„ ๋ชจ๋ธ ๊ฐ„์˜ ์งˆ์  ์ฐจ์ด์— ๋Œ€ํ•ด ๊ฐ•๋ ฅํ•œ ์ง„์ˆ ์„ ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๋ฐ˜๋ณต์„ฑ ๋ฐ ๋ฌธ๋ฒ• ์˜ค๋ฅ˜ ๋ฌธ์ œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ํ”„๋กฌํ”„ํŠธ์— ์ฃผ์–ด์ง„ ๋ฌธ๋งฅ์กฐ์ฐจ ์ง€ํ‚ค์ง€ ๋ชปํ•˜๋Š” Pile baseline์—์„œ ํ›จ์”ฌ ๋” ์‹ฌ๊ฐํ•ด ๋ณด์ž…๋‹ˆ๋‹ค. ์ด๋Š” Pile์—์„œ ๋…์ผ์–ด ํ…์ŠคํŠธ์˜ ์–‘์ด ํ›จ์”ฌ ์ ๊ธฐ ๋•Œ๋ฌธ์— ์˜ˆ์ƒ๋˜๋Š” ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค.
ํ‘œ 9์— ์‘๋‹ต๊ณผ ์‘๋‹ต์˜ ์ˆ˜๋™ ๋ฒˆ์—ญ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ƒ๋Œ€์ ์œผ๋กœ ์ž‘์€ 405M ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์˜๋ฏธ ์žˆ๋Š” ๋ฌธ์žฅ์„ ์ƒ์„ฑํ•˜์ง€ ์•Š์ง€๋งŒ, ๋…์ผ์–ด ์ปค๋จผ ํฌ๋กค๋ง์œผ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฌธ๋ฒ•์ ์œผ๋กœ ์˜ฌ๋ฐ”๋ฅธ ๋ฌธ์žฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์˜ฌ๋ฐ”๋ฅธ ๋Œ€๋ฌธ์ž์™€ ์†Œ๋ฌธ์ž, ํ•ฉ์„ฑ์–ด ์ƒ์„ฑ, ์–ธ๋ผ์šฐํŠธ ๋ฌธ์ž ์‚ฌ์šฉ ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๊ฒฝ์šฐ์—์„œ ์‘๋‹ต์€ ๋ช‡ ๋‹จ์–ด ์ดํ›„์— ๋ฐ˜๋ณต๋˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ์ถœ๋ ฅ์˜ ๋ถˆ์™„์ „์„ฑ ๋ฐ ๋ถ€๋ถ„ ๋‹จ์–ด๋Š” ์ƒ์„ฑ๋œ ํ† ํฐ์˜ ์ˆ˜๊ฐ€ ๋งค์šฐ ์งง๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋” ๊ธด ์‹œํ€€์Šค๋Š” ๊ตฌ๋ฌธ์ ์œผ๋กœ ์˜ฌ๋ฐ”๋ฅธ ๋ฌธ์žฅ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์ง€๋งŒ ์ž์‹ ์„ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ œํ•œ๋œ ์ƒ˜ํ”Œ ์ˆ˜์™€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ์˜ ํ’ˆ์งˆ์ด ๋‚ฎ๊ธฐ ๋•Œ๋ฌธ์— ๋…์ผ์–ด ์ปค๋จผ ํฌ๋กค๋ง์œผ๋กœ ํ•™์Šต๋œ ๊ฐœ๋ณ„ ๋ชจ๋ธ ๊ฐ„์˜ ์งˆ์  ์ฐจ์ด์— ๋Œ€ํ•ด ๊ฐ•๋ ฅํ•œ ์ง„์ˆ ์„ ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๋ฐ˜๋ณต์„ฑ ๋ฐ ๋ฌธ๋ฒ• ์˜ค๋ฅ˜ ๋ฌธ์ œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ํ”„๋กฌํ”„ํŠธ์— ์ฃผ์–ด์ง„ ๋ฌธ๋งฅ์กฐ์ฐจ ์ง€ํ‚ค์ง€ ๋ชปํ•˜๋Š” Pile ๊ธฐ์ค€์—์„œ ํ›จ์”ฌ ๋” ์‹ฌ๊ฐํ•ด ๋ณด์ž…๋‹ˆ๋‹ค. ์ด๋Š” Pile์—์„œ ๋…์ผ์–ด ํ…์ŠคํŠธ์˜ ์–‘์ด ํ›จ์”ฌ ์ ๊ธฐ ๋•Œ๋ฌธ์— ์˜ˆ์ƒ๋˜๋Š” ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค.
In this subsection, we evaluate the extent to which Quiet-STaR improves the zero-shot reasoning capabilities of the language model on CommonsenseQA (Talmor et al., 2018) and GSM8K (Cobbe et al., 2021). On CommonsenseQA, we find that Quiet-STaR improves performance by 10.9% compared to the base language model. As shown in Figure 2, this improvement consistently increases with the number of tokens used in the model's rationales, indicating that more thorough reasoning through the thought tokens is translating to better direct question-answering performance. Similarly, on GSM8K, Quiet-STaR results in a 5.0% boost over the base model, and once again, performance scales with the length of the rationales generated during Quiet-STaR training. For reference, in Figure 2, we include a baseline corresponding to training the same model on the same dataset without thought tokens. We observe that in multiple curves, performance appears to eventually deteriorate - we anticipate that this is because we are not training on these downstream tasks, so the roles of the thought tokens may change over time. We also find a benefit of our non-myopic objective, which we discuss in Appendix D.
์ด ํ•˜์œ„ ์„น์…˜์—์„œ๋Š” CommonsenseQA(Talmor et al., 2018)์™€ GSM8K(Cobbe et al., 2021)์—์„œ Quiet-STaR์ด ์–ธ์–ด ๋ชจ๋ธ์˜ zero-shot ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์ •๋„๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. CommonsenseQA์—์„œ Quiet-STaR์€ ๊ธฐ๋ณธ ์–ธ์–ด ๋ชจ๋ธ์— ๋น„ํ•ด ์„ฑ๋Šฅ์„ 10.9% ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์€ ๋ชจ๋ธ์˜ ๊ทผ๊ฑฐ์— ์‚ฌ์šฉ๋˜๋Š” ํ† ํฐ ์ˆ˜์— ๋”ฐ๋ผ ์ง€์†์ ์œผ๋กœ ์ฆ๊ฐ€ํ•˜๋ฉฐ, ์ด๋Š” ์‚ฌ๊ณ  ํ† ํฐ์„ ํ†ตํ•ด ๋ณด๋‹ค ์ฒ ์ €ํ•œ ์ถ”๋ก ์ด ๋” ๋‚˜์€ ์ง์ ‘ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์„ฑ๋Šฅ์œผ๋กœ ์ด์–ด์ง€๊ณ  ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ GSM8K์—์„œ๋„ Quiet-STaR์€ ๊ธฐ๋ณธ ๋ชจ๋ธ๋ณด๋‹ค 5.0% ํ–ฅ์ƒ๋˜์—ˆ์œผ๋ฉฐ, ๋‹ค์‹œ ํ•œ ๋ฒˆ Quiet-STaR ํ›ˆ๋ จ ์ค‘์— ์ƒ์„ฑ๋œ ๊ทผ๊ฑฐ์˜ ๊ธธ์ด์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด ํ™•์žฅ๋ฉ๋‹ˆ๋‹ค. ์ฐธ๊ณ ๋กœ ๊ทธ๋ฆผ 2์—๋Š” ์‚ฌ๊ณ  ํ† ํฐ ์—†์ด ๋™์ผํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ๋™์ผํ•œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” baseline์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ๊ณก์„ ์—์„œ ์„ฑ๋Šฅ์ด ๊ฒฐ๊ตญ ์•…ํ™”๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ด๋Š”๋ฐ, ์ด๋Š” downstream ์ž‘์—…์— ๋Œ€ํ•ด ํ›ˆ๋ จํ•˜๊ณ  ์žˆ์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ์‚ฌ๊ณ  ํ† ํฐ์˜ ์—ญํ• ์ด ๋ฐ”๋€” ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ผ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ถ€๋ก D์—์„œ ์„ค๋ช…ํ•˜๋Š” ๋น„๊ทผ์‹œ์  objective์˜ ์ด์ ๋„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
์ด ํ•˜์œ„ ์„น์…˜์—์„œ๋Š” CommonsenseQA(Talmor et al., 2018)์™€ GSM8K(Cobbe et al., 2021)์—์„œ Quiet-STaR์ด ์–ธ์–ด ๋ชจ๋ธ์˜ ์ œ๋กœ ์ƒท ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์ •๋„๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. CommonsenseQA์—์„œ Quiet-STaR์€ ๊ธฐ๋ณธ ์–ธ์–ด ๋ชจ๋ธ์— ๋น„ํ•ด ์„ฑ๋Šฅ์„ 10.9% ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์€ ๋ชจ๋ธ์˜ ๊ทผ๊ฑฐ์— ์‚ฌ์šฉ๋˜๋Š” ํ† ํฐ ์ˆ˜์— ๋”ฐ๋ผ ์ง€์†์ ์œผ๋กœ ์ฆ๊ฐ€ํ•˜๋ฉฐ, ์ด๋Š” ์‚ฌ๊ณ  ํ† ํฐ์„ ํ†ตํ•ด ๋ณด๋‹ค ์ฒ ์ €ํ•œ ์ถ”๋ก ์ด ๋” ๋‚˜์€ ์ง์ ‘ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์„ฑ๋Šฅ์œผ๋กœ ์ด์–ด์ง€๊ณ  ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ GSM8K์—์„œ๋„ Quiet-STaR์€ ๊ธฐ๋ณธ ๋ชจ๋ธ๋ณด๋‹ค 5.0% ํ–ฅ์ƒ๋˜์—ˆ์œผ๋ฉฐ, ๋‹ค์‹œ ํ•œ ๋ฒˆ Quiet-STaR ํ›ˆ๋ จ ์ค‘์— ์ƒ์„ฑ๋œ ๊ทผ๊ฑฐ์˜ ๊ธธ์ด์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด ํ™•์žฅ๋ฉ๋‹ˆ๋‹ค. ์ฐธ๊ณ ๋กœ ๊ทธ๋ฆผ 2์—๋Š” ์‚ฌ๊ณ  ํ† ํฐ ์—†์ด ๋™์ผํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ๋™์ผํ•œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๊ธฐ์ค€์„ ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ๊ณก์„ ์—์„œ ์„ฑ๋Šฅ์ด ๊ฒฐ๊ตญ ์•…ํ™”๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ด๋Š”๋ฐ, ์ด๋Š” ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋Œ€ํ•ด ํ›ˆ๋ จํ•˜๊ณ  ์žˆ์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ์‚ฌ๊ณ  ํ† ํฐ์˜ ์—ญํ• ์ด ๋ฐ”๋€” ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ผ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ถ€๋ก D์—์„œ ์„ค๋ช…ํ•˜๋Š” ๋น„๊ทผ์‹œ์  ๋ชฉํ‘œ์˜ ์ด์ ๋„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
biographies led to strong generalization when answering questions about the remaining half of biographies, which resembles setting 1 in Fig. 4. In contrast, training on biographies and QA pairs sequentially failed. However, the key contributor to the success remains uncertain because the data were blended together, and it is unclear how to apply this practically to absorb knowledge from new documents. Inspired by our observation of the different difficulty levels between QA pairs and documents, and the finding from Zhu and Li (2023), we hypothesize that _it is beneficial to deliberately expose LLMs to instruction-tuning data before continued pre-training so that the process of encoding knowledge from complex documents takes into account how this knowledge is accessed._ We refer to this as **pre-instruction-tuning (PIT)** and study various implementations of PIT prior to continued learning (SS 5.1), followed by detailed ablations identifying the keys contributor to performance (SS 5.2 and SS 5.3), and finally assess how well PIT performs across domains (SS 5.4). We adhere to the hyperparameters outlined in SS 3.2 and perform PIT for 3 epochs unless specified otherwise.
์ „๊ธฐ๋Š” ๋‚˜๋จธ์ง€ ์ ˆ๋ฐ˜์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•  ๋•Œ ๊ฐ•๋ ฅํ•œ ์ผ๋ฐ˜ํ™”๋กœ ์ด์–ด์กŒ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ทธ๋ฆผ 4์˜ ์„ค์ • 1๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์ „๊ธฐ์™€ QA ์Œ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์€ ์ˆœ์ฐจ์ ์œผ๋กœ ์‹คํŒจํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•จ๊ป˜ ํ˜ผํ•ฉ๋˜์–ด ์žˆ๊ณ  ์ด๋ฅผ ์‹ค์ œ๋กœ ์ƒˆ๋กœ์šด ๋ฌธ์„œ์—์„œ ์ง€์‹์„ ํก์ˆ˜ํ•˜๋Š” ๋ฐ ์–ด๋–ป๊ฒŒ ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ๋ถˆ๋ถ„๋ช…ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์„ฑ๊ณต์˜ ์ฃผ์š” ์›์ธ์€ ์•„์ง ๋ถˆํ™•์‹คํ•ฉ๋‹ˆ๋‹ค. QA ์Œ๊ณผ ๋ฌธ์„œ ๊ฐ„์˜ ๋‚œ์ด๋„๊ฐ€ ๋‹ค๋ฅด๋‹ค๋Š” ๊ด€์ฐฐ๊ณผ Zhu์™€ Li(2023)์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„, _๋ณต์žกํ•œ ๋ฌธ์„œ์—์„œ ์ง€์‹์„ ์ธ์ฝ”๋”ฉํ•˜๋Š” ๊ณผ์ •์—์„œ ์ด ์ง€์‹์— ์•ก์„ธ์Šคํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ „์— ์˜๋„์ ์œผ๋กœ LLM์„ Instruction Tuning ๋ฐ์ดํ„ฐ์— ๋…ธ์ถœ์‹œํ‚ค๋Š” ๊ฒƒ์ด ์œ ์ตํ•˜๋‹ค_๊ณ  ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค. ์ด๋ฅผ **pre-Instruction Tuning (PIT)**์ด๋ผ๊ณ  ํ•˜๋ฉฐ, ์ง€์†์ ์ธ ํ•™์Šต ์ „์— ๋‹ค์–‘ํ•œ PIT ๊ตฌํ˜„์„ ์—ฐ๊ตฌํ•œ ํ›„(SS 5.1), ์„ฑ๋Šฅ์— ๊ฐ€์žฅ ํฐ ๊ธฐ์—ฌ๋„๋ฅผ ๊ฐ€์ง„ ์š”์†Œ๋ฅผ ํŒŒ์•…ํ•˜๋Š” ์„ธ๋ถ€์ ์ธ ์••์ œ(SS 5.2 ๋ฐ SS 5.3)๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ , ๋งˆ์ง€๋ง‰์œผ๋กœ PIT๊ฐ€ ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์—์„œ ์–ผ๋งˆ๋‚˜ ์ž˜ ์ˆ˜ํ–‰๋˜๋Š”์ง€ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค(SS 5.4). SS 3.2์— ์„ค๋ช…๋œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์ค€์ˆ˜ํ•˜๊ณ  ๋‹ฌ๋ฆฌ ๋ช…์‹œ๋˜์ง€ ์•Š๋Š” ํ•œ 3ํšŒ์— ๊ฑธ์ณ PIT๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
์ „๊ธฐ๋Š” ๋‚˜๋จธ์ง€ ์ ˆ๋ฐ˜์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•  ๋•Œ ๊ฐ•๋ ฅํ•œ ์ผ๋ฐ˜ํ™”๋กœ ์ด์–ด์กŒ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ทธ๋ฆผ 4์˜ ์„ค์ • 1๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์ „๊ธฐ์™€ QA ์Œ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์€ ์ˆœ์ฐจ์ ์œผ๋กœ ์‹คํŒจํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•จ๊ป˜ ํ˜ผํ•ฉ๋˜์–ด ์žˆ๊ณ  ์ด๋ฅผ ์‹ค์ œ๋กœ ์ƒˆ๋กœ์šด ๋ฌธ์„œ์—์„œ ์ง€์‹์„ ํก์ˆ˜ํ•˜๋Š” ๋ฐ ์–ด๋–ป๊ฒŒ ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ๋ถˆ๋ถ„๋ช…ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์„ฑ๊ณต์˜ ์ฃผ์š” ์›์ธ์€ ์•„์ง ๋ถˆํ™•์‹คํ•ฉ๋‹ˆ๋‹ค. QA ์Œ๊ณผ ๋ฌธ์„œ ๊ฐ„์˜ ๋‚œ์ด๋„๊ฐ€ ๋‹ค๋ฅด๋‹ค๋Š” ๊ด€์ฐฐ๊ณผ Zhu์™€ Li(2023)์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„, _๋ณต์žกํ•œ ๋ฌธ์„œ์—์„œ ์ง€์‹์„ ์ธ์ฝ”๋”ฉํ•˜๋Š” ๊ณผ์ •์—์„œ ์ด ์ง€์‹์— ์•ก์„ธ์Šคํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ „์— ์˜๋„์ ์œผ๋กœ LLM์„ ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ์ดํ„ฐ์— ๋…ธ์ถœ์‹œํ‚ค๋Š” ๊ฒƒ์ด ์œ ์ตํ•˜๋‹ค_๊ณ  ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค. ์ด๋ฅผ **์‚ฌ์ „ ๋ช…๋ น์–ด ํŠœ๋‹(PIT)**์ด๋ผ๊ณ  ํ•˜๋ฉฐ, ์ง€์†์ ์ธ ํ•™์Šต ์ „์— ๋‹ค์–‘ํ•œ PIT ๊ตฌํ˜„์„ ์—ฐ๊ตฌํ•œ ํ›„(SS 5.1), ์„ฑ๋Šฅ์— ๊ฐ€์žฅ ํฐ ๊ธฐ์—ฌ๋„๋ฅผ ๊ฐ€์ง„ ์š”์†Œ๋ฅผ ํŒŒ์•…ํ•˜๋Š” ์„ธ๋ถ€์ ์ธ ์••์ œ(SS 5.2 ๋ฐ SS 5.3)๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ , ๋งˆ์ง€๋ง‰์œผ๋กœ PIT๊ฐ€ ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์—์„œ ์–ผ๋งˆ๋‚˜ ์ž˜ ์ˆ˜ํ–‰๋˜๋Š”์ง€ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค(SS 5.4). SS 3.2์— ์„ค๋ช…๋œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์ค€์ˆ˜ํ•˜๊ณ  ๋‹ฌ๋ฆฌ ๋ช…์‹œ๋˜์ง€ ์•Š๋Š” ํ•œ 3ํšŒ์— ๊ฑธ์ณ PIT๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
* **Instructions from _Evol-Instruct_ are superior to the ones from human-created ShareGPT.** When we use the same amount of _Evol-Instruct_ data (i.e., 70k) as Vicuna to fine-tune LLaMA 7B, our model _WizardLM_ significantly outperforms Vicuna, with the win rate of \(12.4\%\) and \(3.8\%\) higher than Vicuna on _Evol-Instruct_ testset and Vicuna's testset respectively on human evaluation. In addition, _WizardLM_ also achieves better response quality than Alpaca and Vicuna on the automatic evaluation of GPT-4. * **Labelers prefer _WizardLM_ outputs over outputs from ChatGPT under complex test instructions.** On _Evol-Instruct_ testset, _WizardLM_ performs worse than ChatGPT, with a win rate \(12.8\%\) lower than ChatGPT (\(28.0\%\) vs. \(40.8\%\)). However, in the high-difficulty section of _Evol-Instruct_ test set (difficulty level \(\geq 8\)), our _WizardLM_ even outperforms ChatGPT, with a win rate \(7.9\%\) larger than ChatGPT (\(42.9\%\) vs. \(35.0\%\)), that is human annotators even prefer the output of our model than ChatGPT on those hard questions. This indicates that _Evol-Instruct_ can significantly improve the ability of LLMs to handle complex instructions.
* **_Evol-Instruct_์˜ Instructions๊ฐ€ ์ธ๊ฐ„์ด ๋งŒ๋“  ShareGPT์˜ Instructions๋ณด๋‹ค ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค.** Vicuna์— ๋™์ผํ•œ ์–‘์˜ _Evol-Instruct_ ๋ฐ์ดํ„ฐ(์ฆ‰, 70k)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLaMA 7B๋ฅผ Fine-tuningํ•  ๋•Œ, ์šฐ๋ฆฌ์˜ ๋ชจ๋ธ์ธ _WizardLM_์€ ์ธ๊ฐ„ ํ‰๊ฐ€์—์„œ _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ Vicuna์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ๊ฐ๊ฐ \(12.4\%\), \(3.8\%\) ๋” ๋†’์€ ์Šน๋ฅ ๋กœ Vicuna๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ _WizardLM_์€ GPT-4์˜ ์ž๋™ ํ‰๊ฐ€์—์„œ๋„ Alpaca์™€ Vicuna๋ณด๋‹ค ๋” ๋‚˜์€ ์‘๋‹ต ํ’ˆ์งˆ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. * **๋ ˆ์ด๋ธ”๋Ÿฌ๋Š” ๋ณต์žกํ•œ ํ…Œ์ŠคํŠธ Instructions์—์„œ ChatGPT์˜ ์ถœ๋ ฅ๋ณด๋‹ค _WizardLM_ ์ถœ๋ ฅ์„ ์„ ํ˜ธํ•ฉ๋‹ˆ๋‹ค.** _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ _WizardLM_์€ ChatGPT๋ณด๋‹ค ๋‚ฎ์€ ์Šน๋ฅ (\(28.0\% ๋Œ€ 40.8\%))๋กœ ChatGPT๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง‘๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ๊ณ ๋‚œ์ด๋„ ์„น์…˜(๋‚œ์ด๋„ \(\geq 8\))์—์„œ๋Š” _WizardLM_์ด ChatGPT๋ณด๋‹ค ๋†’์€ ์Šน๋ฅ (\(42.9\% ๋Œ€ 35.0\%))๋กœ ChatGPT๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์—ฌ ์ธ๊ฐ„ ํ‰๊ฐ€์ž๋Š” ์ด๋Ÿฌํ•œ ์–ด๋ ค์šด ์งˆ๋ฌธ์— ๋Œ€ํ•ด ChatGPT๋ณด๋‹ค ์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ์„ ์„ ํ˜ธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” _Evol-Instruct_๊ฐ€ ๋ณต์žกํ•œ Instructions๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
* **Evol-Instruct_์˜ ์ง€์นจ์ด ์ธ๊ฐ„์ด ๋งŒ๋“  ShareGPT์˜ ์ง€์นจ๋ณด๋‹ค ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค.** Vicuna์— ๋™์ผํ•œ ์–‘์˜ _Evol-Instruct_ ๋ฐ์ดํ„ฐ(์ฆ‰, 70k)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLaMA 7B๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•  ๋•Œ, ์šฐ๋ฆฌ์˜ ๋ชจ๋ธ์ธ _WizardLM_์€ ์ธ๊ฐ„ ํ‰๊ฐ€์—์„œ _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ Vicuna์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ๊ฐ๊ฐ \(12.4\%\), \(3.8\%\) ๋” ๋†’์€ ์Šน๋ฅ ๋กœ Vicuna๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ _WizardLM_์€ GPT-4์˜ ์ž๋™ ํ‰๊ฐ€์—์„œ๋„ ์•ŒํŒŒ์นด์™€ ๋น„์ฟ ๋‚˜๋ณด๋‹ค ๋” ๋‚˜์€ ์‘๋‹ต ํ’ˆ์งˆ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. * **๋ ˆ์ด๋ธ”๋Ÿฌ๋Š” ๋ณต์žกํ•œ ํ…Œ์ŠคํŠธ ์ง€์นจ์—์„œ ChatGPT์˜ ์ถœ๋ ฅ๋ณด๋‹ค _WizardLM_ ์ถœ๋ ฅ์„ ์„ ํ˜ธํ•ฉ๋‹ˆ๋‹ค.** Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ _WizardLM_์€ ChatGPT๋ณด๋‹ค ๋‚ฎ์€ ์Šน๋ฅ (\(28.0\% ๋Œ€ 40.8\%))๋กœ ChatGPT๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง‘๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ๊ณ ๋‚œ์ด๋„ ์„น์…˜(๋‚œ์ด๋„ \(\geq 8\))์—์„œ๋Š” _WizardLM_์ด ChatGPT๋ณด๋‹ค ๋†’์€ ์Šน๋ฅ (\(42.9\% ๋Œ€ 35.0\%))๋กœ ChatGPT๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์—ฌ ์ธ๊ฐ„ ํ‰๊ฐ€์ž๋Š” ์ด๋Ÿฌํ•œ ์–ด๋ ค์šด ์งˆ๋ฌธ์— ๋Œ€ํ•ด ChatGPT๋ณด๋‹ค ์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ์„ ์„ ํ˜ธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” _Evol-Instruct_๊ฐ€ ๋ณต์žกํ•œ ๋ช…๋ น์„ ์ฒ˜๋ฆฌํ•˜๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
**Sequence Tagging.** The sequence tagging tasks, _e.g._, named entity recognition (NER) [755] and part-of-speech (POS) tagging [756], are also fundamental tasks. Typically, such tasks require assigning each token in the input sequence a proper semantic category label, _e.g._, the classic B-I-O (_Beginning, Inside_ and _Outside_) tagging scheme for NER tasks. In the era of deep learning, early efforts [757, 758] mainly integrate the learned sequence representations (_e.g._, using CNN, LSTM, and BERT) into the classic conditional random field model (CRF), which performs the tagging task based on structural prediction. Recently, researchers have tested the performance of LLMs in sequence tagging tasks, but observed that LLMs still face challenges in solving them using in-context learning [753], especially for special categories with ambiguous or rare names, _e.g._, the "MISC" (_miscellaneous entity_) and "ORG" (_organization_) classes. A possible reason is that LLMs may misunderstand the meanings of these classes in the human-annotated dataset, making it difficult to accurately understand their semantics according to the instruction and limited examples in the context.
**Sequence Tagging**. ์‹œํ€€์Šค ํƒœ๊น… ์ž‘์—…, _์˜ˆ๋ฅผ ๋“ค์–ด, ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ ์ธ์‹(NER) [755] ๋ฐ ํ’ˆ์‚ฌ(POS) ํƒœ๊น… [756]๋„ ๊ธฐ๋ณธ ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ด๋Ÿฌํ•œ ์ž‘์—…์—๋Š” ์ž…๋ ฅ ์‹œํ€€์Šค์˜ ๊ฐ ํ† ํฐ์— ์ ์ ˆํ•œ ์˜๋ฏธ ๋ฒ”์ฃผ ๋ ˆ์ด๋ธ”, _์˜ˆ๋ฅผ ๋“ค์–ด, NER ์ž‘์—…์˜ ๊ณ ์ „์ ์ธ B-I-O(_์‹œ์ž‘, ๋‚ด๋ถ€_ ๋ฐ _์™ธ๋ถ€_) ํƒœ๊น… ์ฒด๊ณ„๋ฅผ ํ• ๋‹นํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฅ๋Ÿฌ๋‹ ์‹œ๋Œ€์˜ ์ดˆ๊ธฐ ๋…ธ๋ ฅ [757, 758]์€ ์ฃผ๋กœ ํ•™์Šต๋œ ์‹œํ€€์Šค ํ‘œํ˜„(_์˜ˆ: CNN, LSTM, BERT ์‚ฌ์šฉ)์„ ๊ณ ์ „์ ์ธ ์กฐ๊ฑด๋ถ€ ๋žœ๋ค ํ•„๋“œ ๋ชจ๋ธ(CRF)์— ํ†ตํ•ฉํ•˜์—ฌ ๊ตฌ์กฐ์  ์˜ˆ์ธก์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํƒœ๊น… ์ž‘์—…์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ์ž๋“ค์€ ์‹œํ€€์Šค ํƒœ๊น… ์ž‘์—…์—์„œ LLM์˜ ์„ฑ๋Šฅ์„ ํ…Œ์ŠคํŠธํ–ˆ์ง€๋งŒ, ํŠนํžˆ ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ํฌ๊ท€ํ•œ ์ด๋ฆ„์„ ๊ฐ€์ง„ ํŠน์ˆ˜ ์นดํ…Œ๊ณ ๋ฆฌ, _์˜ˆ๋ฅผ ๋“ค์–ด "MISC"(_๊ธฐํƒ€ ์—”ํ‹ฐํ‹ฐ_) ๋ฐ "ORG"(_์กฐ์ง_) ํด๋ž˜์Šค์˜ ๊ฒฝ์šฐ, LLM์ด In-Context Learning [753]์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์—ฌ์ „ํžˆ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ์ด์œ ๋Š” LLM์ด ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹จ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ด๋Ÿฌํ•œ ํด๋ž˜์Šค์˜ ์˜๋ฏธ๋ฅผ ์ž˜๋ชป ์ดํ•ดํ•˜์—ฌ ๋ฌธ๋งฅ์˜ Instruction๊ณผ ์ œํ•œ๋œ ์˜ˆ์ œ์— ๋”ฐ๋ผ ๊ทธ ์˜๋ฏธ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์‹œํ€€์Šค ํƒœ๊น…**. ์‹œํ€€์Šค ํƒœ๊น… ์ž‘์—…, _์˜ˆ๋ฅผ ๋“ค์–ด, ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ ์ธ์‹(NER) [755] ๋ฐ ํ’ˆ์‚ฌ(POS) ํƒœ๊น… [756]๋„ ๊ธฐ๋ณธ ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ด๋Ÿฌํ•œ ์ž‘์—…์—๋Š” ์ž…๋ ฅ ์‹œํ€€์Šค์˜ ๊ฐ ํ† ํฐ์— ์ ์ ˆํ•œ ์˜๋ฏธ ๋ฒ”์ฃผ ๋ ˆ์ด๋ธ”, _์˜ˆ๋ฅผ ๋“ค์–ด, NER ์ž‘์—…์˜ ๊ณ ์ „์ ์ธ B-I-O(_์‹œ์ž‘, ๋‚ด๋ถ€_ ๋ฐ _์™ธ๋ถ€_) ํƒœ๊น… ์ฒด๊ณ„๋ฅผ ํ• ๋‹นํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฅ๋Ÿฌ๋‹ ์‹œ๋Œ€์˜ ์ดˆ๊ธฐ ๋…ธ๋ ฅ [757, 758]์€ ์ฃผ๋กœ ํ•™์Šต๋œ ์‹œํ€€์Šค ํ‘œํ˜„(_์˜ˆ: CNN, LSTM, BERT ์‚ฌ์šฉ)์„ ๊ณ ์ „์ ์ธ ์กฐ๊ฑด๋ถ€ ๋žœ๋ค ํ•„๋“œ ๋ชจ๋ธ(CRF)์— ํ†ตํ•ฉํ•˜์—ฌ ๊ตฌ์กฐ์  ์˜ˆ์ธก์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํƒœ๊น… ์ž‘์—…์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ์ž๋“ค์€ ์‹œํ€€์Šค ํƒœ๊น… ์ž‘์—…์—์„œ LLM์˜ ์„ฑ๋Šฅ์„ ํ…Œ์ŠคํŠธํ–ˆ์ง€๋งŒ, ํŠนํžˆ ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ํฌ๊ท€ํ•œ ์ด๋ฆ„์„ ๊ฐ€์ง„ ํŠน์ˆ˜ ์นดํ…Œ๊ณ ๋ฆฌ, _์˜ˆ๋ฅผ ๋“ค์–ด "MISC"(_๊ธฐํƒ€ ์—”ํ‹ฐํ‹ฐ_) ๋ฐ "ORG"(_์กฐ์ง_) ํด๋ž˜์Šค์˜ ๊ฒฝ์šฐ, LLM์ด ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต [753]์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์—ฌ์ „ํžˆ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ์ด์œ ๋Š” LLM์ด ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹จ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ด๋Ÿฌํ•œ ํด๋ž˜์Šค์˜ ์˜๋ฏธ๋ฅผ ์ž˜๋ชป ์ดํ•ดํ•˜์—ฌ ๋ฌธ๋งฅ์˜ ์ง€์‹œ์™€ ์ œํ•œ๋œ ์˜ˆ์ œ์— ๋”ฐ๋ผ ๊ทธ ์˜๋ฏธ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
ExperimentWe demonstrate that EntRec is a reasonable approximation of the internal recall of the bridge entity with indirect evidence. Note that EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\) is calculated not at the last token of \(\tau_{\text{2H}}\) but at the last token of the bridge entity's descriptive mention, where it is grammatically natural to prepend a comma followed by the name of \(e_{2}\) (e.g., "The mother of the singer of 'Superstition', _Stevie Wonder_"). In the resulting string, grammatically \(\mu(r_{1}(e_{1}))\)) becomes the _antecedent_ and \(e_{2}\) becomes the _appositive_; an appositive is a noun phrase that follows another noun phrase in opposition to it and provides information that further identifies or defines it, and the antecedent is the noun phrase that the appositive describes. Then, if EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\) reasonably approximates the internal recall of the bridge entity \(e_{2}\), it is expected that _there will be at least some layers \(l\) where increasing EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\) increases the relative frequency of the LLM to generate \(e_{2}^{(0)}\) with a relative frequency higher than random chance_.
์‹คํ—˜์šฐ๋ฆฌ๋Š” ๊ฐ„์ ‘์ ์ธ ์ฆ๊ฑฐ๋ฅผ ํ†ตํ•ด EntRec์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์„ ํ•ฉ๋ฆฌ์ ์œผ๋กœ approximationํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)๋Š” \(\tau_{\text{2H}}\)์˜ ๋งˆ์ง€๋ง‰ ํ† ํฐ์ด ์•„๋‹ˆ๋ผ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์„ค๋ช…์  ์–ธ๊ธ‰์˜ ๋งˆ์ง€๋ง‰ ํ† ํฐ์—์„œ ๊ณ„์‚ฐ๋˜๋ฉฐ, ๋ฌธ๋ฒ•์ ์œผ๋กœ ์‰ผํ‘œ ๋’ค์— \(e_{2}\)์˜ ์ด๋ฆ„์„ ์•ž์— ๋ถ™์ด๋Š” ๊ฒƒ์ด ์ž์—ฐ์Šค๋Ÿฝ์Šต๋‹ˆ๋‹ค(์˜ˆ: "'superstition'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ, _์Šคํ‹ฐ๋น„ ์›๋”_"). ๊ฒฐ๊ณผ ๋ฌธ์ž์—ด์—์„œ ๋ฌธ๋ฒ•์ ์œผ๋กœ \(\mu(r_{1}(e_{1}))\))๋Š” _antecedent_๊ฐ€ ๋˜๊ณ  \(e_{2}\)๋Š” _appositive_๊ฐ€ ๋˜๋ฉฐ, appositive๋Š” ๋‹ค๋ฅธ ๋ช…์‚ฌ๊ตฌ ๋’ค์— ๋ฐ˜๋Œ€๋˜๋Š” ๋ช…์‚ฌ๊ตฌ๊ฐ€ ๋˜์–ด ๊ทธ ๋ช…์‚ฌ๊ตฌ๋ฅผ ๋” ์‹๋ณ„ํ•˜๊ฑฐ๋‚˜ ์ •์˜ํ•˜๋Š” ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, antecedent๋Š” ๋ช…์‚ฌ๊ตฌ๊ฐ€ ์„ค๋ช…ํ•˜๋Š” ๋ช…์‚ฌ๊ตฌ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ \(e_{2}\)์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์„ ํ•ฉ๋ฆฌ์ ์œผ๋กœ approximationํ•œ๋‹ค๋ฉด, EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)์„ ์ฆ๊ฐ€์‹œํ‚ค๋ฉด LLM์ด ์ƒ๋Œ€์  ๋นˆ๋„๊ฐ€ ๋ฌด์ž‘์œ„ ํ™•๋ฅ ๋ณด๋‹ค ๋†’์€ \(e_{2}^{(0)}\)์„ ์ƒ์„ฑํ•˜๋Š” ์ƒ๋Œ€ ๋นˆ๋„๊ฐ€ ์ฆ๊ฐ€ํ•˜๋Š” ์ธต \(l\)์ด ์ ์–ด๋„ ๋ช‡ ๊ฐœ ์ด์ƒ ์กด์žฌํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.
์‹คํ—˜์šฐ๋ฆฌ๋Š” ๊ฐ„์ ‘์ ์ธ ์ฆ๊ฑฐ๋ฅผ ํ†ตํ•ด EntRec์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์„ ํ•ฉ๋ฆฌ์ ์œผ๋กœ ๊ทผ์‚ฌํ™”ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)๋Š” \(\tau_{\text{2H}}\)์˜ ๋งˆ์ง€๋ง‰ ํ† ํฐ์ด ์•„๋‹ˆ๋ผ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์„ค๋ช…์  ์–ธ๊ธ‰์˜ ๋งˆ์ง€๋ง‰ ํ† ํฐ์—์„œ ๊ณ„์‚ฐ๋˜๋ฉฐ, ๋ฌธ๋ฒ•์ ์œผ๋กœ ์‰ผํ‘œ ๋’ค์— \(e_{2}\)์˜ ์ด๋ฆ„์„ ์•ž์— ๋ถ™์ด๋Š” ๊ฒƒ์ด ์ž์—ฐ์Šค๋Ÿฝ์Šต๋‹ˆ๋‹ค(์˜ˆ: "'superstition'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ, _์Šคํ‹ฐ๋น„ ์›๋”_"). ๊ฒฐ๊ณผ ๋ฌธ์ž์—ด์—์„œ ๋ฌธ๋ฒ•์ ์œผ๋กœ \(\mu(r_{1}(e_{1}))\))๋Š” _์ „์น˜์‚ฌ_๊ฐ€ ๋˜๊ณ  \(e_{2}\)๋Š” _๋ช…์‚ฌ์ ˆ_์ด ๋˜๋ฉฐ, ๋ช…์‚ฌ์ ˆ์€ ๋‹ค๋ฅธ ๋ช…์‚ฌ์ ˆ ๋’ค์— ๋ฐ˜๋Œ€๋˜๋Š” ๋ช…์‚ฌ์ ˆ์ด ๋˜์–ด ๊ทธ ๋ช…์‚ฌ์ ˆ์„ ๋” ์‹๋ณ„ํ•˜๊ฑฐ๋‚˜ ์ •์˜ํ•˜๋Š” ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์„ ํ–‰์‚ฌ๋Š” ๋ช…์‚ฌ์ ˆ์ด ์„ค๋ช…ํ•˜๋Š” ๋ช…์‚ฌ์ ˆ์ด ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ \(e_{2}\)์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์„ ํ•ฉ๋ฆฌ์ ์œผ๋กœ ๊ทผ์‚ฌํ™”ํ•œ๋‹ค๋ฉด, EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)์„ ์ฆ๊ฐ€์‹œํ‚ค๋ฉด LLM์ด ์ƒ๋Œ€์  ๋นˆ๋„๊ฐ€ ๋ฌด์ž‘์œ„ ํ™•๋ฅ ๋ณด๋‹ค ๋†’์€ \(e_{2}^{(0)}\)์„ ์ƒ์„ฑํ•˜๋Š” ์ƒ๋Œ€ ๋นˆ๋„๊ฐ€ ์ฆ๊ฐ€ํ•˜๋Š” ์ธต \(l\)์ด ์ ์–ด๋„ ๋ช‡ ๊ฐœ ์ด์ƒ ์กด์žฌํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.
forecasts receive higher weights. We remark that the platform also publishes another aggregated forecast called "Metaculus prediction" (which we do not use or compare with in this paper). This differs from the crowd prediction described above and is computed via a proprietary algorithm. * GJOpen computes the crowd predictions by the mean of the most recent 40% of the forecasts from each forecaster. * INFER initializes the crowd prediction to be the mean of all individual forecasts. As the question progresses, it reweights the forecasts, for example, by "putting more weight on the forecasts of individuals with the best track record."4 Exact details on the aggregation mechanisms are not found on their website. Footnote 4: [https://www.infer-pub.com/frequently-asked-questions](https://www.infer-pub.com/frequently-asked-questions) * Manifold and Polymarket are prediction markets, where the community predictions are the prices (between 0 and 1). The prices are adjusted by their automated market makers, as bets are made. The mechanisms are variants of constant-product market makers (Hanson, 2007; Zhang et al., 2018); see Polymarket (2023); Manifold (2022) for more details.
์˜ˆ์ธก์ด ๋” ๋†’์€ ๊ฐ€์ค‘์น˜๋ฅผ ๋ฐ›์Šต๋‹ˆ๋‹ค. ์ด ํ”Œ๋žซํผ์€ ๋˜ํ•œ "Metaculus prediction"์ด๋ผ๋Š” ๋˜ ๋‹ค๋ฅธ ์ง‘๊ณ„๋œ ์˜ˆ์ธก์„ ๊ฒŒ์‹œํ•ฉ๋‹ˆ๋‹ค(์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ ๋น„๊ตํ•˜์ง€ ์•Š์Œ). ์ด๋Š” ์œ„์—์„œ ์„ค๋ช…ํ•œ crowd prediction๊ณผ ๋‹ค๋ฅด๋ฉฐ ๋…์  ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ†ตํ•ด ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. * GJOpen์€ ๊ฐ ์˜ˆ์ธก์ž์˜ ๊ฐ€์žฅ ์ตœ๊ทผ 40%์˜ ์˜ˆ์ธก์„ ํ‰๊ท ํ•˜์—ฌ crowd prediction์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. * INFER๋Š” ๋ชจ๋“  ๊ฐœ๋ณ„ ์˜ˆ์ธก์˜ ํ‰๊ท ์ด ๋˜๋„๋ก crowd prediction์„ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ์ด ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ ์˜ˆ์ธก์˜ ๊ฐ€์ค‘์น˜๋ฅผ ์žฌ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค(์˜ˆ: "์ตœ๊ณ ์˜ ์‹ค์ ์„ ๋ณด์œ ํ•œ ๊ฐœ์ธ์˜ ์˜ˆ์ธก์— ๋” ๋งŽ์€ ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌ").4 ์ง‘๊ณ„ ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๋Œ€ํ•œ ์ •ํ™•ํ•œ ์„ธ๋ถ€ ์‚ฌํ•ญ์€ ์›น์‚ฌ์ดํŠธ์—์„œ ํ™•์ธํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๊ฐ์ฃผ 4: [https://www.infer-pub.com/frequently-asked-questions](https://www.infer-pub.com/frequently-asked-questions)* Manifold์™€ Polymarket์€ ์˜ˆ์ธก ์‹œ์žฅ์œผ๋กœ, ์ปค๋ฎค๋‹ˆํ‹ฐ ์˜ˆ์ธก์€ ๊ฐ€๊ฒฉ(0์—์„œ 1 ์‚ฌ์ด)์ž…๋‹ˆ๋‹ค. ๋ฒ ํŒ…์ด ์ด๋ฃจ์–ด์ง€๋ฉด ์ž๋™ํ™”๋œ ์‹œ์žฅ ๋ฉ”์ด์ปค์— ์˜ํ•ด ๊ฐ€๊ฒฉ์ด ์กฐ์ •๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ constant-product market makers(Hanson, 2007; Zhang et al., 2018)์˜ ๋ณ€ํ˜•์œผ๋กœ, ์ž์„ธํ•œ ๋‚ด์šฉ์€ Polymarket(2023); Manifold(2022)๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
์˜ˆ์ธก์ด ๋” ๋†’์€ ๊ฐ€์ค‘์น˜๋ฅผ ๋ฐ›์Šต๋‹ˆ๋‹ค. ์ด ํ”Œ๋žซํผ์€ ๋˜ํ•œ "๋ฉ”ํƒ€์ฟจ๋Ÿฌ์Šค ์˜ˆ์ธก"์ด๋ผ๋Š” ๋˜ ๋‹ค๋ฅธ ์ง‘๊ณ„๋œ ์˜ˆ์ธก์„ ๊ฒŒ์‹œํ•ฉ๋‹ˆ๋‹ค(์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ ๋น„๊ตํ•˜์ง€ ์•Š์Œ). ์ด๋Š” ์œ„์—์„œ ์„ค๋ช…ํ•œ ํฌ๋ผ์šฐ๋“œ ์˜ˆ์ธก๊ณผ ๋‹ค๋ฅด๋ฉฐ ๋…์  ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ†ตํ•ด ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. * GJOpen์€ ๊ฐ ์˜ˆ์ธก์ž์˜ ๊ฐ€์žฅ ์ตœ๊ทผ 40%์˜ ์˜ˆ์ธก์„ ํ‰๊ท ํ•˜์—ฌ ํฌ๋ผ์šฐ๋“œ ์˜ˆ์ธก์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. * INFER๋Š” ๋ชจ๋“  ๊ฐœ๋ณ„ ์˜ˆ์ธก์˜ ํ‰๊ท ์ด ๋˜๋„๋ก ํฌ๋ผ์šฐ๋“œ ์˜ˆ์ธก์„ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ์ด ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ ์˜ˆ์ธก์˜ ๊ฐ€์ค‘์น˜๋ฅผ ์žฌ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค(์˜ˆ: "์ตœ๊ณ ์˜ ์‹ค์ ์„ ๋ณด์œ ํ•œ ๊ฐœ์ธ์˜ ์˜ˆ์ธก์— ๋” ๋งŽ์€ ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌ").4 ์ง‘๊ณ„ ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๋Œ€ํ•œ ์ •ํ™•ํ•œ ์„ธ๋ถ€ ์‚ฌํ•ญ์€ ์›น์‚ฌ์ดํŠธ์—์„œ ํ™•์ธํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๊ฐ์ฃผ 4: [https://www.infer-pub.com/frequently-asked-questions](https://www.infer-pub.com/frequently-asked-questions)* Manifold์™€ Polymarket์€ ์˜ˆ์ธก ์‹œ์žฅ์œผ๋กœ, ์ปค๋ฎค๋‹ˆํ‹ฐ ์˜ˆ์ธก์€ ๊ฐ€๊ฒฉ(0์—์„œ 1 ์‚ฌ์ด)์ž…๋‹ˆ๋‹ค. ๋ฒ ํŒ…์ด ์ด๋ฃจ์–ด์ง€๋ฉด ์ž๋™ํ™”๋œ ์‹œ์žฅ ๋ฉ”์ด์ปค์— ์˜ํ•ด ๊ฐ€๊ฒฉ์ด ์กฐ์ •๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ์ผ์ •ํ•œ ์ œํ’ˆ ์‹œ์žฅ ๋ฉ”์ด์ปค(Hanson, 2007; Zhang et al., 2018)์˜ ๋ณ€ํ˜•์œผ๋กœ, ์ž์„ธํ•œ ๋‚ด์šฉ์€ Polymarket(2023); Manifold(2022)๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
To improve the reasoning ability, most training methods focus on acquiring supervised signals obtained through costly means, such as human annotation (Liao et al., 2024), generation of larger models (Yu et al., 2023; Luo et al., 2023; Liu et al., 2023; Yue et al., 2023), or self-sampling (Yuan et al., 2023; Singh et al., 2023; Wang et al., 2023; Zelikman et al., 2022). These methods depend on these signals to direct learning towards a specific trajectory by precisely outlining the reasoning process. In contrast, our method illustrates that an entirely different approach is equally effective. Instead of requiring more precise guidance to direct the generative process, our technique achieves comparable results by introducing random noises into the input. For reasoning tasks, which are generally considered to require precise understanding, we find that the method of incorporating noise into the reasoning steps is surprisingly effective, indicating that the model might have enhanced its performance through a denoising process. Our approach stands out for its simplicity and effectiveness, offering a feasible alternative to the more complex and resource-intensive methods.
์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๋Œ€๋ถ€๋ถ„์˜ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์€ ์ธ๊ฐ„ ์ฃผ์„(Liao et al., 2024), ๋” ํฐ ๋ชจ๋ธ ์ƒ์„ฑ(Yu et al., 2023; Luo et al., 2023; Liu et al., 2023; Yue et al., 2023) ๋˜๋Š” ์ž์ฒด ์ƒ˜ํ”Œ๋ง(Yuan et al., 2023; Singh et al., 2023; Wang et al., 2023; Zelikman et al., 2022)๊ณผ ๊ฐ™์ด ๋น„์šฉ์ด ๋งŽ์ด ๋“œ๋Š” ์ˆ˜๋‹จ์„ ํ†ตํ•ด ์–ป์€ supervised signals๋ฅผ ํš๋“ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์€ ์ถ”๋ก  ๊ณผ์ •์„ ์ •ํ™•ํ•˜๊ฒŒ ์„ค๋ช…ํ•จ์œผ๋กœ์จ ํŠน์ • ๊ถค์ ์„ ํ–ฅํ•ด ํ•™์Šต์„ ์œ ๋„ํ•˜๊ธฐ ์œ„ํ•ด ์ด๋Ÿฌํ•œ signals์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์ €ํฌ์˜ ๋ฐฉ๋ฒ•์€ ์™„์ „ํžˆ ๋‹ค๋ฅธ ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋˜‘๊ฐ™์ด ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ์œ ๋„ํ•˜๊ธฐ ์œ„ํ•ด ๋” ์ •ํ™•ํ•œ ์•ˆ๋‚ด๊ฐ€ ํ•„์š”ํ•œ ๋Œ€์‹ , ์ €ํฌ ๊ธฐ์ˆ ์€ ์ž…๋ ฅ์— ๋ฌด์ž‘์œ„ ๋…ธ์ด์ฆˆ๋ฅผ ๋„์ž…ํ•˜์—ฌ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ •ํ™•ํ•œ ์ดํ•ด๊ฐ€ ํ•„์š”ํ•œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋˜๋Š” ์ถ”๋ก  ์ž‘์—…์˜ ๊ฒฝ์šฐ, ์ถ”๋ก  ๋‹จ๊ณ„์— ๋…ธ์ด์ฆˆ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ๋†€๋ž๋„๋ก ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ์ด denoising ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ด ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œ์ผฐ์„ ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ €ํฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋‹จ์ˆœํ•จ๊ณผ ํšจ๊ณผ์„ฑ์ด ๋›ฐ์–ด๋‚˜ ๋ณต์žกํ•˜๊ณ  ๋ฆฌ์†Œ์Šค ์ง‘์•ฝ์ ์ธ ๊ธฐ์กด ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์‹คํ˜„ ๊ฐ€๋Šฅํ•œ ๋Œ€์•ˆ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๋Œ€๋ถ€๋ถ„์˜ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์€ ์ธ๊ฐ„ ์ฃผ์„(Liao et al., 2024), ๋” ํฐ ๋ชจ๋ธ ์ƒ์„ฑ(Yu et al., 2023; Luo et al., 2023; Liu et al., 2023; Yue et al., 2023) ๋˜๋Š” ์ž์ฒด ์ƒ˜ํ”Œ๋ง(Yuan et al., 2023; Singh et al., 2023; Wang et al., 2023; Zelikman et al., 2022)๊ณผ ๊ฐ™์ด ๋น„์šฉ์ด ๋งŽ์ด ๋“œ๋Š” ์ˆ˜๋‹จ์„ ํ†ตํ•ด ์–ป์€ ์ง€๋„ ์‹ ํ˜ธ๋ฅผ ํš๋“ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์€ ์ถ”๋ก  ๊ณผ์ •์„ ์ •ํ™•ํ•˜๊ฒŒ ์„ค๋ช…ํ•จ์œผ๋กœ์จ ํŠน์ • ๊ถค์ ์„ ํ–ฅํ•ด ํ•™์Šต์„ ์œ ๋„ํ•˜๊ธฐ ์œ„ํ•ด ์ด๋Ÿฌํ•œ ์‹ ํ˜ธ์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์ €ํฌ์˜ ๋ฐฉ๋ฒ•์€ ์™„์ „ํžˆ ๋‹ค๋ฅธ ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋˜‘๊ฐ™์ด ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ์œ ๋„ํ•˜๊ธฐ ์œ„ํ•ด ๋” ์ •ํ™•ํ•œ ์•ˆ๋‚ด๊ฐ€ ํ•„์š”ํ•œ ๋Œ€์‹ , ์ €ํฌ ๊ธฐ์ˆ ์€ ์ž…๋ ฅ์— ๋ฌด์ž‘์œ„ ๋…ธ์ด์ฆˆ๋ฅผ ๋„์ž…ํ•˜์—ฌ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ •ํ™•ํ•œ ์ดํ•ด๊ฐ€ ํ•„์š”ํ•œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋˜๋Š” ์ถ”๋ก  ์ž‘์—…์˜ ๊ฒฝ์šฐ, ์ถ”๋ก  ๋‹จ๊ณ„์— ๋…ธ์ด์ฆˆ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ๋†€๋ž๋„๋ก ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ์ด ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ด ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œ์ผฐ์„ ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ €ํฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋‹จ์ˆœํ•จ๊ณผ ํšจ๊ณผ์„ฑ์ด ๋›ฐ์–ด๋‚˜ ๋ณต์žกํ•˜๊ณ  ๋ฆฌ์†Œ์Šค ์ง‘์•ฝ์ ์ธ ๊ธฐ์กด ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์‹คํ˜„ ๊ฐ€๋Šฅํ•œ ๋Œ€์•ˆ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
Deploying large language models on personal computers, particularly on CPUs, has historically been challenging due to their immense computational requirements. However, with the help of many community efforts, such as llama.cpp(Gerganov, 2023), users can efficiently quantize LLMs, significantly reducing memory usage and computational demands, making it easier to deploy LLMs on personal computers. This also enables quicker interactions with the models and facilitates local data processing. Quantizing LLMs and deploying them on personal computers offer several benefits. Firstly, it helps users protect their data privacy by ensuring that sensitive information remains within their local environment rather than being transmitted to external servers. Secondly, it democratizes access to LLMs by making them more accessible to users with limited computational resources. Lastly, it promotes the development of new applications and research directions that take advantage of local LLM deployments. Overall, the ability to deploy LLMs on personal computers using 1lama.cpp (or similar) paves the way for a more versatile and privacy-conscious utilization of LLMs in various domains.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ, ํŠนํžˆ CPU์— ๋ฐฐํฌํ•˜๋Š” ๊ฒƒ์€ ์—„์ฒญ๋‚œ ๊ณ„์‚ฐ ์š”๊ตฌ ์‚ฌํ•ญ์œผ๋กœ ์ธํ•ด ์—ญ์‚ฌ์ ์œผ๋กœ ์–ด๋ ค์› ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ llama.cpp(Gerganov, 2023)์™€ ๊ฐ™์€ ๋งŽ์€ ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ๋…ธ๋ ฅ์œผ๋กœ ์‚ฌ์šฉ์ž๋Š” LLM์„ ํšจ์œจ์ ์œผ๋กœ Quantizationํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰๊ณผ ๊ณ„์‚ฐ ์š”๊ตฌ ์‚ฌํ•ญ์„ ํฌ๊ฒŒ ์ค„์ผ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ์— LLM์„ ๋” ์‰ฝ๊ฒŒ ๋ฐฐํฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋ธ๊ณผ์˜ ์ƒํ˜ธ ์ž‘์šฉ ์†๋„๋ฅผ ๋†’์ด๊ณ  ๋กœ์ปฌ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ๋ฅผ ์šฉ์ดํ•˜๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. LLM์„ Quantizationํ•˜์—ฌ ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ์— ๋ฐฐํฌํ•˜๋ฉด ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ด์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋ฏผ๊ฐํ•œ ์ •๋ณด๊ฐ€ ์™ธ๋ถ€ ์„œ๋ฒ„๋กœ ์ „์†ก๋˜์ง€ ์•Š๊ณ  ๋กœ์ปฌ ํ™˜๊ฒฝ ๋‚ด์— ์œ ์ง€๋˜๋„๋ก ํ•˜์—ฌ ์‚ฌ์šฉ์ž์˜ ๋ฐ์ดํ„ฐ ํ”„๋ผ์ด๋ฒ„์‹œ๋ฅผ ๋ณดํ˜ธํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค๊ฐ€ ์ œํ•œ๋œ ์‚ฌ์šฉ์ž๋„ LLM์— ๋” ์‰ฝ๊ฒŒ ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์—ฌ LLM์— ๋Œ€ํ•œ ์ ‘๊ทผ์„ ๋ฏผ์ฃผํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๋กœ์ปฌ LLM ๋ฐฐํฌ๋ฅผ ํ™œ์šฉํ•˜๋Š” ์ƒˆ๋กœ์šด ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜๊ณผ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์˜ ๊ฐœ๋ฐœ์„ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ llama.cpp(๋˜๋Š” ์ด์™€ ์œ ์‚ฌํ•œ)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ์— LLM์„ ๋ฐฐํฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์€ ๋‹ค์–‘ํ•œ ์˜์—ญ์—์„œ ๋ณด๋‹ค ๋‹ค์žฌ๋‹ค๋Šฅํ•˜๊ณ  ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ์— ์ค‘์ ์„ ๋‘” LLM ํ™œ์šฉ์˜ ๊ธธ์„ ์—ด์–ด์ค๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ, ํŠนํžˆ CPU์— ๋ฐฐํฌํ•˜๋Š” ๊ฒƒ์€ ์—„์ฒญ๋‚œ ๊ณ„์‚ฐ ์š”๊ตฌ ์‚ฌํ•ญ์œผ๋กœ ์ธํ•ด ์—ญ์‚ฌ์ ์œผ๋กœ ์–ด๋ ค์› ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ llama.cpp(Gerganov, 2023)์™€ ๊ฐ™์€ ๋งŽ์€ ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ๋…ธ๋ ฅ์œผ๋กœ ์‚ฌ์šฉ์ž๋Š” LLM์„ ํšจ์œจ์ ์œผ๋กœ ์–‘์žํ™”ํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰๊ณผ ๊ณ„์‚ฐ ์š”๊ตฌ ์‚ฌํ•ญ์„ ํฌ๊ฒŒ ์ค„์ผ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ์— LLM์„ ๋” ์‰ฝ๊ฒŒ ๋ฐฐํฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋ธ๊ณผ์˜ ์ƒํ˜ธ ์ž‘์šฉ ์†๋„๋ฅผ ๋†’์ด๊ณ  ๋กœ์ปฌ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ๋ฅผ ์šฉ์ดํ•˜๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. LLM์„ ์–‘์žํ™”ํ•˜์—ฌ ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ์— ๋ฐฐํฌํ•˜๋ฉด ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ด์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋ฏผ๊ฐํ•œ ์ •๋ณด๊ฐ€ ์™ธ๋ถ€ ์„œ๋ฒ„๋กœ ์ „์†ก๋˜์ง€ ์•Š๊ณ  ๋กœ์ปฌ ํ™˜๊ฒฝ ๋‚ด์— ์œ ์ง€๋˜๋„๋ก ํ•˜์—ฌ ์‚ฌ์šฉ์ž์˜ ๋ฐ์ดํ„ฐ ํ”„๋ผ์ด๋ฒ„์‹œ๋ฅผ ๋ณดํ˜ธํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค๊ฐ€ ์ œํ•œ๋œ ์‚ฌ์šฉ์ž๋„ LLM์— ๋” ์‰ฝ๊ฒŒ ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์—ฌ LLM์— ๋Œ€ํ•œ ์ ‘๊ทผ์„ ๋ฏผ์ฃผํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๋กœ์ปฌ LLM ๋ฐฐํฌ๋ฅผ ํ™œ์šฉํ•˜๋Š” ์ƒˆ๋กœ์šด ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜๊ณผ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์˜ ๊ฐœ๋ฐœ์„ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ 1๋ผ๋งˆ.cpp(๋˜๋Š” ์ด์™€ ์œ ์‚ฌํ•œ)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ์— LLM์„ ๋ฐฐํฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์€ ๋‹ค์–‘ํ•œ ์˜์—ญ์—์„œ ๋ณด๋‹ค ๋‹ค์žฌ๋‹ค๋Šฅํ•˜๊ณ  ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ์— ์ค‘์ ์„ ๋‘” LLM ํ™œ์šฉ์˜ ๊ธธ์„ ์—ด์–ด์ค๋‹ˆ๋‹ค.
Figure 20: **Hutter LLM.** true distribution of the data is Zipf with exponent \(\beta=2\). Here, the scalable resource is either clean data or AI data-generated data, corresponding to a version of real data with its tail cut at rank \(k\) (here we use \(k=10\)). We either mix with a fixed amount (here \(T^{\prime}=10^{4}\) samples) of the other resource, or we donโ€™t mix at all. Then we scale up the scalable resource by cranking up \(T\). As predicted by Corollary 3.3, the orange curve always grokks: AI synthesized data is helpful in the regime where real data is scarce; once more of real data becomes available the model grokks for a while and then forgets the AI synthesized data. Note that the green curve (only AI data) and red curve (AI + real data) donโ€™t grokk because the optional resource (real data) is not being scaled; if it is also scaled, then green and red will provably grokk (as in Figure 3). The diagonal broken line corresponds to the standard Hutter scaling law \(E_{test}\asymp T^{-c}\), where \(c:=1-1/\beta\). The horizontal broken lines correspond to \(E_{test}\asymp k^{-(\beta-1)}\) and \(E_{test}\asymp T^{\prime-c}\), both predicted by Theorem 2.1.
๊ทธ๋ฆผ 20: **Hutter LLM.** ๋ฐ์ดํ„ฐ์˜ ์‹ค์ œ ๋ถ„ํฌ๋Š” ์ง€์ˆ˜ \(\beta=2\)์˜ Zipf์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ฆฌ์†Œ์Šค๋Š” ์ •๋ฆฌ๋œ ๋ฐ์ดํ„ฐ ๋˜๋Š” AI ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๋ฐ์ดํ„ฐ์ด๋ฉฐ, ์ด๋Š” ์ˆœ์œ„ \(k\)์—์„œ ๊ผฌ๋ฆฌ๊ฐ€ ์ž˜๋ฆฐ ์‹ค์ œ ๋ฐ์ดํ„ฐ์˜ ๋ฒ„์ „์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค(์—ฌ๊ธฐ์„œ๋Š” \(k=10\)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค). ๋‹ค๋ฅธ ๋ฆฌ์†Œ์Šค์˜ ๊ณ ์ •๋œ ์–‘(์—ฌ๊ธฐ์„œ๋Š” \(T^{\prime}=10^{4}\) ์ƒ˜ํ”Œ)๊ณผ ํ˜ผํ•ฉํ•˜๊ฑฐ๋‚˜ ์ „ํ˜€ ํ˜ผํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ \(T\)๋ฅผ ๋†’์—ฌ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค. 3.3 ์ •๋ฆฌ์˜ ์˜ˆ์ธก์— ๋”ฐ๋ฅด๋ฉด ์ฃผํ™ฉ์ƒ‰ ๊ณก์„ ์€ ํ•ญ์ƒ ์ดํ•ดํ•ฉ๋‹ˆ๋‹ค: ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•œ ์˜์—ญ์—์„œ๋Š” AI ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ์œ ์šฉํ•˜๋ฉฐ, ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋” ๋งŽ์ด ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•ด์ง€๋ฉด ๋ชจ๋ธ์€ ์ž ์‹œ ๋™์•ˆ ์ดํ•ดํ–ˆ๋‹ค๊ฐ€ AI ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์žŠ์–ด๋ฒ„๋ฆฝ๋‹ˆ๋‹ค. ๋…น์ƒ‰ ๊ณก์„ (AI ๋ฐ์ดํ„ฐ๋งŒ)๊ณผ ๋นจ๊ฐ„์ƒ‰ ๊ณก์„ (AI + ์‹ค์ œ ๋ฐ์ดํ„ฐ)์€ ์„ ํƒ์  ๋ฆฌ์†Œ์Šค(์‹ค์ œ ๋ฐ์ดํ„ฐ)๊ฐ€ ํ™•์žฅ๋˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ๊ทธ๋กœํ‚น์ด ๋˜์ง€ ์•Š์ง€๋งŒ, ์ด ๋ฆฌ์†Œ์Šค๋„ ํ™•์žฅํ•˜๋ฉด ๊ทธ๋ฆผ 3์—์„œ์™€ ๊ฐ™์ด ๋…น์ƒ‰๊ณผ ๋นจ๊ฐ„์ƒ‰์ด ์ฆ๋ช… ๊ฐ€๋Šฅํ•˜๊ฒŒ ๊ทธ๋กœํ‚น๋ฉ๋‹ˆ๋‹ค. ๋Œ€๊ฐ์„  ํŒŒ์„ ์€ ํ‘œ์ค€ Hutter scaling law \(E_{test}\asymp T^{-c}\), ์—ฌ๊ธฐ์„œ \(c:=1-1/\beta\)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€๋กœ ๊บพ์€ ์„ ์€ ์ •๋ฆฌ 2.1์— ์˜ํ•ด ์˜ˆ์ธก๋œ \(E_{test}\asymp k^{-(\beta-1)}\) ๋ฐ \(E_{test}\asymp T^{\prime-c}\)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 20: **ํ—ˆํ„ฐ LLM.** ๋ฐ์ดํ„ฐ์˜ ์‹ค์ œ ๋ถ„ํฌ๋Š” ์ง€์ˆ˜ \(\beta=2\)์˜ ์ง€ํ”„์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ฆฌ์†Œ์Šค๋Š” ์ •๋ฆฌ๋œ ๋ฐ์ดํ„ฐ ๋˜๋Š” AI ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๋ฐ์ดํ„ฐ์ด๋ฉฐ, ์ด๋Š” ์ˆœ์œ„ \(k\)์—์„œ ๊ผฌ๋ฆฌ๊ฐ€ ์ž˜๋ฆฐ ์‹ค์ œ ๋ฐ์ดํ„ฐ์˜ ๋ฒ„์ „์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค(์—ฌ๊ธฐ์„œ๋Š” \(k=10\)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค). ๋‹ค๋ฅธ ๋ฆฌ์†Œ์Šค์˜ ๊ณ ์ •๋œ ์–‘(์—ฌ๊ธฐ์„œ๋Š” \(T^{\ํ”„๋ผ์ž„}=10^{4}\) ์ƒ˜ํ”Œ)๊ณผ ํ˜ผํ•ฉํ•˜๊ฑฐ๋‚˜ ์ „ํ˜€ ํ˜ผํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ \(T\)๋ฅผ ๋†’์—ฌ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค. 3.3 ์ •๋ฆฌ์˜ ์˜ˆ์ธก์— ๋”ฐ๋ฅด๋ฉด ์ฃผํ™ฉ์ƒ‰ ๊ณก์„ ์€ ํ•ญ์ƒ ์ดํ•ดํ•ฉ๋‹ˆ๋‹ค: ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•œ ์˜์—ญ์—์„œ๋Š” AI ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ์œ ์šฉํ•˜๋ฉฐ, ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋” ๋งŽ์ด ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•ด์ง€๋ฉด ๋ชจ๋ธ์€ ์ž ์‹œ ๋™์•ˆ ์ดํ•ดํ–ˆ๋‹ค๊ฐ€ AI ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์žŠ์–ด๋ฒ„๋ฆฝ๋‹ˆ๋‹ค. ๋…น์ƒ‰ ๊ณก์„ (AI ๋ฐ์ดํ„ฐ๋งŒ)๊ณผ ๋นจ๊ฐ„์ƒ‰ ๊ณก์„ (AI + ์‹ค์ œ ๋ฐ์ดํ„ฐ)์€ ์„ ํƒ์  ๋ฆฌ์†Œ์Šค(์‹ค์ œ ๋ฐ์ดํ„ฐ)๊ฐ€ ํ™•์žฅ๋˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ๊ทธ๋กœํ‚น์ด ๋˜์ง€ ์•Š์ง€๋งŒ, ์ด ๋ฆฌ์†Œ์Šค๋„ ํ™•์žฅํ•˜๋ฉด ๊ทธ๋ฆผ 3์—์„œ์™€ ๊ฐ™์ด ๋…น์ƒ‰๊ณผ ๋นจ๊ฐ„์ƒ‰์ด ์ฆ๋ช… ๊ฐ€๋Šฅํ•˜๊ฒŒ ๊ทธ๋กœํ‚น๋ฉ๋‹ˆ๋‹ค. ๋Œ€๊ฐ์„  ํŒŒ์„ ์€ ํ‘œ์ค€ ํ—ˆํ„ฐ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™ \(E_{test}\asymp T^{-c}\), ์—ฌ๊ธฐ์„œ \(c:=1-1/\beta\)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€๋กœ ๊บพ์€ ์„ ์€ ์ •๋ฆฌ 2.1์— ์˜ํ•ด ์˜ˆ์ธก๋œ \(E_{test}\asymp k^{-(\beta-1)}\) ๋ฐ \(E_{test}\asymp T^{\ํ”„๋ผ์ž„-c}\)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
We use three datasets for training and validation: SlimPajama (Soboleva et al., 2023), German CommonCrawl (Laippala et al., 2022), and Pile (Gao et al., 2020). For all datasets, use the same tokenizer as Black et al. (2022) trained specifically on the Pile. To create our training set for SlimPajama, we randomly sub-sample the dataset (606B Total Tokens) to form a \(\sim\)299B token subset (see Table 1) that is of comparable size to Pile. We also further sub-sample this SlimPajama subset to create three \(\sim 100\)B token splits of the dataset (see Sec. 7.4 for details). To create the SlimPajama validation set we simply tokenize the default validation set that has been extensively deduplicated (Soboleva et al., 2023). To create the German training and validaiton sets, we split and tokenized the German Common Crawl scrape, available as part of the Oscar Dataset (Laippala et al., 2022), into a 195.43B token training set and a 982.6M token validation set. The pile dataset comes pre-shuffled and mixed, we simply used the default training and validation sets. The training set is \(\sim 330\)B tokens total, though in our experiments we only train on a 300B token subset.
ํ•™์Šต ๋ฐ ๊ฒ€์ฆ์—๋Š” SlimPajama(Soboleva et al., 2023), German CommonCrawl(Laippala et al., 2022), Pile(Gao et al., 2020)์˜ ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด Pile์—์„œ ํŠน๋ณ„ํžˆ ํ•™์Šต๋œ Black et al.(2022)๊ณผ ๋™์ผํ•œ tokenizer๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. SlimPajama์˜ ํ›ˆ๋ จ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ์„ธํŠธ(์ด ํ† ํฐ 606B)๋ฅผ ๋ฌด์ž‘์œ„๋กœ ํ•˜์œ„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ Pile๊ณผ ๋น„์Šทํ•œ ํฌ๊ธฐ์ธ \(\sim\)299B token ํ•˜์œ„ ์ง‘ํ•ฉ์„ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค(ํ‘œ 1 ์ฐธ์กฐ). ๋˜ํ•œ ์ด SlimPajama ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์ถ”๊ฐ€๋กœ ํ•˜์œ„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์˜ 3๊ฐœ์˜ \(\sim 100\)B token ๋ถ„ํ• ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค(์ž์„ธํ•œ ๋‚ด์šฉ์€ ์„น์…˜ 7.4 ์ฐธ์กฐ). SlimPajama validation ์ง‘ํ•ฉ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๊ธฐ๋ณธ validation ์ง‘ํ•ฉ์„ tokenํ™”ํ•˜๊ธฐ๋งŒ ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค(Soboleva et al., 2023). ๋…์ผ์–ด ํ›ˆ๋ จ ๋ฐ validation ์ง‘ํ•ฉ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด Oscar ๋ฐ์ดํ„ฐ ์„ธํŠธ(Laippala et al., 2022)์˜ ์ผ๋ถ€๋กœ ์ œ๊ณต๋˜๋Š” ๋…์ผ์–ด Common Crawl scrape์„ 195.43B token ํ›ˆ๋ จ ์„ธํŠธ์™€ 982.6M token validation ์„ธํŠธ๋กœ ๋ถ„ํ• ํ•˜๊ณ  tokenํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. Pile ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋ฏธ๋ฆฌ ์„ž์—ฌ ์žˆ๊ณ  ํ˜ผํ•ฉ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ ๊ธฐ๋ณธ ํ›ˆ๋ จ ๋ฐ validation ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์„ธํŠธ๋Š” ์ด \(\sim 330\)B token์ด์ง€๋งŒ, ์‹คํ—˜์—์„œ๋Š” 300B token ํ•˜์œ„ ์ง‘ํ•ฉ๋งŒ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.
ํ•™์Šต ๋ฐ ๊ฒ€์ฆ์—๋Š” SlimPajama(Soboleva et al., 2023), German CommonCrawl(Laippala et al., 2022), Pile(Gao et al., 2020)์˜ ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด Pile์—์„œ ํŠน๋ณ„ํžˆ ํ•™์Šต๋œ Black et al.(2022)๊ณผ ๋™์ผํ•œ ํ† ํฐํ™”๊ธฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. SlimPajama์˜ ํ›ˆ๋ จ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ์„ธํŠธ(์ด ํ† ํฐ 606B)๋ฅผ ๋ฌด์ž‘์œ„๋กœ ํ•˜์œ„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ Pile๊ณผ ๋น„์Šทํ•œ ํฌ๊ธฐ์ธ \(sim\)299B ํ† ํฐ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค(ํ‘œ 1 ์ฐธ์กฐ). ๋˜ํ•œ ์ด SlimPajama ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์ถ”๊ฐ€๋กœ ํ•˜์œ„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์˜ 3๊ฐœ์˜ \(sim 100\)B ํ† ํฐ ๋ถ„ํ• ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค(์ž์„ธํ•œ ๋‚ด์šฉ์€ ์„น์…˜ 7.4 ์ฐธ์กฐ). SlimPajama ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๊ธฐ๋ณธ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์„ ํ† ํฐํ™”ํ•˜๊ธฐ๋งŒ ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค(Soboleva et al., 2023). ๋…์ผ์–ด ํ›ˆ๋ จ ๋ฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด Oscar ๋ฐ์ดํ„ฐ ์„ธํŠธ(Laippala et al., 2022)์˜ ์ผ๋ถ€๋กœ ์ œ๊ณต๋˜๋Š” ๋…์ผ์–ด ์ปค๋จผ ํฌ๋กค๋ง ์Šคํฌ๋žฉ์„ 195.43B ํ† ํฐ ํ›ˆ๋ จ ์„ธํŠธ์™€ 982.6M ํ† ํฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋กœ ๋ถ„ํ• ํ•˜๊ณ  ํ† ํฐํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”๋ฏธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋ฏธ๋ฆฌ ์„ž์—ฌ ์žˆ๊ณ  ํ˜ผํ•ฉ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ ๊ธฐ๋ณธ ํ›ˆ๋ จ ๋ฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์„ธํŠธ๋Š” ์ด \(\sim 330\)B ํ† ํฐ์ด์ง€๋งŒ, ์‹คํ—˜์—์„œ๋Š” 300B ํ† ํฐ ํ•˜์œ„ ์ง‘ํ•ฉ๋งŒ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.
Our pretraining data cleaning system features a sophisticated filtering pipeline based on language, heuristic textual features, perplexity, semantics, topic, and safety, as well as a cascaded deduplication process based on paragraph, MinHash, and exact matching. This thorough pipeline leads to a much higher removal ratio than existing pipelines like CCNet [80], RefinedWeb [56] and RedPajama [13], which we believe is key to the success of data engineering. The underlying principle is although pretraining requires data scaling, one would like to make sure the data used are of high quality, rather than training the model on large raw data, i.e., we prefer 3T tokens over sophisticated engineering over 10T tokens without extensive filtering. Regarding the model architecture, we use standard implementation of the Transformer architecture with Grouped-Query Attention (GQA) [1], SwiGLU [68] activation, and RoPE with an adjusted base frequency (RoPE ABF) [82]. This design choice is the standard approach rooted from the Transformer original paper [78], later modified by GPT-3 and Chinchilla [30], then followed by LLaMA [77], Baichuan [84], Qwen [3] and many related works.
์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์ •๋ฆฌ ์‹œ์Šคํ…œ์€ ์–ธ์–ด, heuristic ํ…์ŠคํŠธ ๊ธฐ๋Šฅ, perplexity, semantics, topic ๋ฐ safety๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ •๊ตํ•œ filtering ํŒŒ์ดํ”„๋ผ์ธ๊ณผ ๋‹จ๋ฝ, MinHash ๋ฐ ์ •ํ™•ํ•œ ์ผ์น˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๊ณ„๋‹จ์‹ deduplication ํ”„๋กœ์„ธ์Šค๋ฅผ ํŠน์ง•์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ฒ ์ €ํ•œ ํŒŒ์ดํ”„๋ผ์ธ์€ ๊ธฐ์กด ํŒŒ์ดํ”„๋ผ์ธ์ธ CCNet [80], RefinedWeb [56] ๋ฐ RedPajama [13]๋ณด๋‹ค ํ›จ์”ฌ ๋†’์€ ์ œ๊ฑฐ ๋น„์œจ๋กœ ์ด์–ด์ง€๋ฉฐ, ์ด๋Š” ๋ฐ์ดํ„ฐ ์—”์ง€๋‹ˆ์–ด๋ง์˜ ์„ฑ๊ณต์— ํ•ต์‹ฌ์ ์ธ ์š”์†Œ๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ์›์น™์€ ์‚ฌ์ „ ํ•™์Šต์—๋Š” ๋ฐ์ดํ„ฐ scaling์ด ํ•„์š”ํ•˜์ง€๋งŒ, ๋ชจ๋ธ์„ ๋Œ€๋Ÿ‰์˜ ์›์‹œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์ด ๋†’์€์ง€ ํ™•์ธํ•˜๋Š” ๊ฒƒ์ด๋ฉฐ, ์ด๋Š” ๊ด‘๋ฒ”์œ„ํ•œ filtering ์—†์ด 10T tokens๋ณด๋‹ค ์ •๊ตํ•œ ์—”์ง€๋‹ˆ์–ด๋ง์„ ํ†ตํ•ด 3T tokens์„ ์„ ํ˜ธํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์™€ ๊ด€๋ จํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” Grouped-Query Attention (GQA) [1], SwiGLU [68] activation, ์กฐ์ •๋œ ๊ธฐ๋ณธ ์ฃผํŒŒ์ˆ˜(RoPE ABF) [82]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‘œ์ค€ ๊ตฌํ˜„์˜ Transformer ์•„ํ‚คํ…์ฒ˜๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค๊ณ„ ๋ฐฉ์‹์€ Transformer ์›๋ฌธ [78]์— ๋ฟŒ๋ฆฌ๋ฅผ ๋‘” ํ‘œ์ค€ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ, ์ดํ›„ GPT-3 ๋ฐ Chinchilla [30]์— ์˜ํ•ด ์ˆ˜์ •๋œ ํ›„ LLaMA [77], Baichuan [84], Qwen [3] ๋ฐ ๋งŽ์€ ๊ด€๋ จ ์—ฐ๊ตฌ์— ์˜ํ•ด ์ฑ„ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์ •๋ฆฌ ์‹œ์Šคํ…œ์€ ์–ธ์–ด, ํœด๋ฆฌ์Šคํ‹ฑ ํ…์ŠคํŠธ ๊ธฐ๋Šฅ, ๋‹นํ˜น๊ฐ, ์˜๋ฏธ๋ก , ์ฃผ์ œ ๋ฐ ์•ˆ์ „์„ฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ •๊ตํ•œ ํ•„ํ„ฐ๋ง ํŒŒ์ดํ”„๋ผ์ธ๊ณผ ๋‹จ๋ฝ, MinHash ๋ฐ ์ •ํ™•ํ•œ ์ผ์น˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๊ณ„๋‹จ์‹ ์ค‘๋ณต ์ œ๊ฑฐ ํ”„๋กœ์„ธ์Šค๋ฅผ ํŠน์ง•์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ฒ ์ €ํ•œ ํŒŒ์ดํ”„๋ผ์ธ์€ ๊ธฐ์กด ํŒŒ์ดํ”„๋ผ์ธ์ธ CCNet [80], RefinedWeb [56] ๋ฐ RedPajama [13]๋ณด๋‹ค ํ›จ์”ฌ ๋†’์€ ์ œ๊ฑฐ ๋น„์œจ๋กœ ์ด์–ด์ง€๋ฉฐ, ์ด๋Š” ๋ฐ์ดํ„ฐ ์—”์ง€๋‹ˆ์–ด๋ง์˜ ์„ฑ๊ณต์— ํ•ต์‹ฌ์ ์ธ ์š”์†Œ๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ์›์น™์€ ์‚ฌ์ „ ํ•™์Šต์—๋Š” ๋ฐ์ดํ„ฐ ํ™•์žฅ์ด ํ•„์š”ํ•˜์ง€๋งŒ, ๋ชจ๋ธ์„ ๋Œ€๋Ÿ‰์˜ ์›์‹œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์ด ๋†’์€์ง€ ํ™•์ธํ•˜๋Š” ๊ฒƒ์ด๋ฉฐ, ์ด๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ํ•„ํ„ฐ๋ง ์—†์ด 10T ํ† ํฐ๋ณด๋‹ค ์ •๊ตํ•œ ์—”์ง€๋‹ˆ์–ด๋ง์„ ํ†ตํ•ด 3T ํ† ํฐ์„ ์„ ํ˜ธํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์™€ ๊ด€๋ จํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” ๊ทธ๋ฃนํ™”๋œ ์ฟผ๋ฆฌ ์–ดํ…์…˜(GQA) [1], SwiGLU [68] ํ™œ์„ฑํ™”, ์กฐ์ •๋œ ๊ธฐ๋ณธ ์ฃผํŒŒ์ˆ˜(RoPE ABF) [82]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‘œ์ค€ ๊ตฌํ˜„์˜ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค๊ณ„ ๋ฐฉ์‹์€ ํŠธ๋žœ์Šคํฌ๋จธ ์›๋ฌธ [78]์— ๋ฟŒ๋ฆฌ๋ฅผ ๋‘” ํ‘œ์ค€ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ, ์ดํ›„ GPT-3 ๋ฐ ์นœ์น ๋ผ [30]์— ์˜ํ•ด ์ˆ˜์ •๋œ ํ›„ LLaMA [77], ๋ฐ”์ด์ถ”์•ˆ [84], ์ฒธ [3] ๋ฐ ๋งŽ์€ ๊ด€๋ จ ์—ฐ๊ตฌ์— ์˜ํ•ด ์ฑ„ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
RAFT is a training strategy designed to enhance the model's performance in answering questions within a specific domain, in "open-book" settings. This technique demonstrates a fine-tuning recipe for LLMs for question-answering tasks based on a selected collection of documents. We have pinpointed several crucial design decisions, such as training the model alongside distractor documents, organizing the dataset so a portion lacks oracle documents in their context, and formulating answers in a chain-of-thought manner with direct quotations from the relevant text. Our evaluations on PubMed, HotpotQA, and Gorilla API Bench underline RAFT's significant potential. Looking forward, we anticipate that in-domain Retrieval-Augmented Generation (RAG) will continue to gain interest within both industrial and academic spheres. Unlike general-RAG, our work addresses practical scenarios where LLMs are tasked with answering questions using domain-specific knowledge. Aligning with current trends, our findings suggest that smaller, fine-tuned models are capable of performing comparably well in domain-specific question-answering tasks, in contrast to their generic LLM counterparts.
RAFT๋Š” ํŠน์ • ๋„๋ฉ”์ธ ๋‚ด์—์„œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋œ ํ›ˆ๋ จ ์ „๋žต์œผ๋กœ, '์˜คํ”ˆ ๋ถ' ํ™˜๊ฒฝ์—์„œ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ด ๊ธฐ๋ฒ•์€ ์„ ํƒํ•œ ๋ฌธ์„œ ์ปฌ๋ ‰์…˜์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์ž‘์—…์„ ์œ„ํ•œ LLM์— ๋Œ€ํ•œ Fine-tuning ๋ ˆ์‹œํ”ผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๋ชจ๋ธ์„ ์‚ฐ๋งŒํ•œ ๋ฌธ์„œ์™€ ํ•จ๊ป˜ ํ›ˆ๋ จํ•˜๊ณ , ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜์—ฌ ์ผ๋ถ€ ๋ฌธ๋งฅ์—์„œ ์˜ค๋ผํด ๋ฌธ์„œ๊ฐ€ ์—†๋Š” ๊ฒฝ์šฐ, ๊ด€๋ จ ํ…์ŠคํŠธ์—์„œ ์ง์ ‘ ์ธ์šฉํ•˜์—ฌ ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ๋ฐฉ์‹์œผ๋กœ ๋‹ต๋ณ€์„ ๊ณต์‹ํ™”ํ•˜๋Š” ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์„ค๊ณ„ ๊ฒฐ์ •์„ ์ •ํ™•ํžˆ ํŒŒ์•…ํ–ˆ์Šต๋‹ˆ๋‹ค. PubMed, HotpotQA ๋ฐ Gorilla API Bench์— ๋Œ€ํ•œ ํ‰๊ฐ€๋Š” RAFT์˜ ์ƒ๋‹นํ•œ ์ž ์žฌ๋ ฅ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์•ž์œผ๋กœ๋„ ์‚ฐ์—…๊ณ„์™€ ํ•™๊ณ„ ๋ชจ๋‘์—์„œ ๋„๋ฉ”์ธ Retrieval-Augmented Generation (RAG)์— ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ๊ณ„์† ์ฆ๊ฐ€ํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ RAG์™€ ๋‹ฌ๋ฆฌ, ์ €ํฌ์˜ ์—ฐ๊ตฌ๋Š” LLM์ด ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ์งˆ๋ฌธ์— ๋‹ตํ•ด์•ผ ํ•˜๋Š” ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋‹ค๋ฃจ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ์˜ ์ถ”์„ธ์— ๋ถ€ํ•ฉํ•˜๋Š” ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด, ์†Œ๊ทœ๋ชจ์˜ Fine-tuned ๋ชจ๋ธ์ด ์ผ๋ฐ˜์ ์ธ LLM๊ณผ ๋‹ฌ๋ฆฌ ๋„๋ฉ”์ธ๋ณ„ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์ž‘์—…์—์„œ ๋น„์Šทํ•œ ์ˆ˜์ค€์˜ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค.
RAFT๋Š” ํŠน์ • ๋„๋ฉ”์ธ ๋‚ด์—์„œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋œ ํ›ˆ๋ จ ์ „๋žต์œผ๋กœ, '์˜คํ”ˆ ๋ถ' ํ™˜๊ฒฝ์—์„œ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ด ๊ธฐ๋ฒ•์€ ์„ ํƒํ•œ ๋ฌธ์„œ ์ปฌ๋ ‰์…˜์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์ž‘์—…์„ ์œ„ํ•œ LLM์— ๋Œ€ํ•œ ๋ฏธ์„ธ ์กฐ์ • ๋ ˆ์‹œํ”ผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๋ชจ๋ธ์„ ์‚ฐ๋งŒํ•œ ๋ฌธ์„œ์™€ ํ•จ๊ป˜ ํ›ˆ๋ จํ•˜๊ณ , ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜์—ฌ ์ผ๋ถ€ ๋ฌธ๋งฅ์—์„œ ์˜ค๋ผํด ๋ฌธ์„œ๊ฐ€ ์—†๋Š” ๊ฒฝ์šฐ, ๊ด€๋ จ ํ…์ŠคํŠธ์—์„œ ์ง์ ‘ ์ธ์šฉํ•˜์—ฌ ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ๋ฐฉ์‹์œผ๋กœ ๋‹ต๋ณ€์„ ๊ณต์‹ํ™”ํ•˜๋Š” ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์„ค๊ณ„ ๊ฒฐ์ •์„ ์ •ํ™•ํžˆ ํŒŒ์•…ํ–ˆ์Šต๋‹ˆ๋‹ค. PubMed, HotpotQA ๋ฐ Gorilla API Bench์— ๋Œ€ํ•œ ํ‰๊ฐ€๋Š” RAFT์˜ ์ƒ๋‹นํ•œ ์ž ์žฌ๋ ฅ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์•ž์œผ๋กœ๋„ ์‚ฐ์—…๊ณ„์™€ ํ•™๊ณ„ ๋ชจ๋‘์—์„œ ๋„๋ฉ”์ธ ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์ƒ์„ฑ(RAG)์— ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ๊ณ„์† ์ฆ๊ฐ€ํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ RAG์™€ ๋‹ฌ๋ฆฌ, ์ €ํฌ์˜ ์—ฐ๊ตฌ๋Š” LLM์ด ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ์งˆ๋ฌธ์— ๋‹ตํ•ด์•ผ ํ•˜๋Š” ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋‹ค๋ฃจ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ์˜ ์ถ”์„ธ์— ๋ถ€ํ•ฉํ•˜๋Š” ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด, ์†Œ๊ทœ๋ชจ์˜ ๋ฏธ์„ธ ์กฐ์ • ๋ชจ๋ธ์ด ์ผ๋ฐ˜์ ์ธ LLM๊ณผ ๋‹ฌ๋ฆฌ ๋„๋ฉ”์ธ๋ณ„ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์ž‘์—…์—์„œ ๋น„์Šทํ•œ ์ˆ˜์ค€์˜ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค.
**Stop Using BLEU** BLEU, a metric extensively utilized for decades, often diverges from neural-based and reference-free metrics, a phenomenon also observed in previous studies (Xu et al., 2023; Freitag et al., 2023). For instance, WMT competition winners often exhibit superior performance according to BLEU (or COMET-22), yet this is not corroborated by reference-free models. A case in point is the WMT winners scoring an exceptionally high 64.14 BLEU in cs\(\rightarrow\)en translations, significantly outperforming other models by 20 BLEU points. However, reference-free evaluations suggest these translations are inferior to those generated by our models and GPT-4. We hypothesize that this discrepancy may arise from WMT models being trained on domain-specific data closely related to the WMT test set, leading to high lexical matches but lacking semantic depth as evaluated by neural-based metrics. While BLEU scores are effective for assessing basic functionality in weaker models, their utility diminishes with advanced translation models capable of generating diverse translations. In such contexts, relying solely on BLEU for evaluation appears increasingly outdated.
**BLEU ์‚ฌ์šฉ ์ค‘์ง€** ์ˆ˜์‹ญ ๋…„ ๋™์•ˆ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ํ™œ์šฉ๋˜์–ด ์˜จ ์ง€ํ‘œ์ธ BLEU๋Š” ์‹ ๊ฒฝ ๊ธฐ๋ฐ˜ ์ง€ํ‘œ ๋ฐ ์ฐธ์กฐ ์—†๋Š” ์ง€ํ‘œ์™€ ์ข…์ข… ์ฐจ์ด๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋„ ๊ด€์ฐฐ๋œ ํ˜„์ƒ์ž…๋‹ˆ๋‹ค(Xu et al., 2023; Freitag et al., 2023). ์˜ˆ๋ฅผ ๋“ค์–ด, WMT ๋Œ€ํšŒ ์šฐ์Šน์ž๋Š” BLEU(๋˜๋Š” COMET-22)์— ๋”ฐ๋ผ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์ง€๋งŒ, ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ์—์„œ๋Š” ์ด๋ฅผ ํ™•์ธํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ์‚ฌ๋ก€๋กœ, WMT ์šฐ์Šน์ž๋Š” cs\(\rightarrow\)en ๋ฒˆ์—ญ์—์„œ 64.14์˜ ๋งค์šฐ ๋†’์€ BLEU ์ ์ˆ˜๋ฅผ ๋ฐ›์•„ ๋‹ค๋ฅธ ๋ชจ๋ธ๋ณด๋‹ค 20 BLEU ํฌ์ธํŠธ ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฐธ์กฐ ์—†๋Š” ํ‰๊ฐ€์— ๋”ฐ๋ฅด๋ฉด ์ด๋Ÿฌํ•œ ๋ฒˆ์—ญ์€ ์šฐ๋ฆฌ ๋ชจ๋ธ๊ณผ GPT-4์—์„œ ์ƒ์„ฑ๋œ ๋ฒˆ์—ญ๋ณด๋‹ค ์—ด๋“ฑํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฐจ์ด๋Š” WMT ๋ชจ๋ธ์ด WMT ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ ๋ฐ€์ ‘ํ•˜๊ฒŒ ๊ด€๋ จ๋œ ๋„๋ฉ”์ธ๋ณ„ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ์–ดํœ˜ ์ผ์น˜์œจ์€ ๋†’์ง€๋งŒ ์‹ ๊ฒฝ ๊ธฐ๋ฐ˜ metrics๋กœ ํ‰๊ฐ€ํ•  ๋•Œ ์˜๋ฏธ๋ก ์  ๊นŠ์ด๊ฐ€ ๋ถ€์กฑํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค. BLEU ์ ์ˆ˜๋Š” ์•ฝํ•œ ๋ชจ๋ธ์—์„œ ๊ธฐ๋ณธ ๊ธฐ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ํšจ๊ณผ์ ์ด์ง€๋งŒ, ๋‹ค์–‘ํ•œ ๋ฒˆ์—ญ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๊ณ ๊ธ‰ ๋ฒˆ์—ญ ๋ชจ๋ธ์—์„œ๋Š” ๊ทธ ์œ ์šฉ์„ฑ์ด ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋งฅ๋ฝ์—์„œ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด BLEU์—๋งŒ ์˜์กดํ•˜๋Š” ๊ฒƒ์€ ์ ์  ๋” ๊ตฌ์‹์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค.
**BLEU ์‚ฌ์šฉ ์ค‘์ง€** ์ˆ˜์‹ญ ๋…„ ๋™์•ˆ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ํ™œ์šฉ๋˜์–ด ์˜จ ์ง€ํ‘œ์ธ BLEU๋Š” ์‹ ๊ฒฝ ๊ธฐ๋ฐ˜ ์ง€ํ‘œ ๋ฐ ์ฐธ์กฐ ์—†๋Š” ์ง€ํ‘œ์™€ ์ข…์ข… ์ฐจ์ด๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋„ ๊ด€์ฐฐ๋œ ํ˜„์ƒ์ž…๋‹ˆ๋‹ค(Xu et al., 2023; Freitag et al., 2023). ์˜ˆ๋ฅผ ๋“ค์–ด, WMT ๋Œ€ํšŒ ์šฐ์Šน์ž๋Š” BLEU(๋˜๋Š” COMET-22)์— ๋”ฐ๋ผ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์ง€๋งŒ, ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ์—์„œ๋Š” ์ด๋ฅผ ํ™•์ธํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ์‚ฌ๋ก€๋กœ, WMT ์šฐ์Šน์ž๋Š” cs\(\rightarrow\)en ๋ฒˆ์—ญ์—์„œ 64.14์˜ ๋งค์šฐ ๋†’์€ BLEU ์ ์ˆ˜๋ฅผ ๋ฐ›์•„ ๋‹ค๋ฅธ ๋ชจ๋ธ๋ณด๋‹ค 20 BLEU ํฌ์ธํŠธ ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฐธ์กฐ ์—†๋Š” ํ‰๊ฐ€์— ๋”ฐ๋ฅด๋ฉด ์ด๋Ÿฌํ•œ ๋ฒˆ์—ญ์€ ์šฐ๋ฆฌ ๋ชจ๋ธ๊ณผ GPT-4์—์„œ ์ƒ์„ฑ๋œ ๋ฒˆ์—ญ๋ณด๋‹ค ์—ด๋“ฑํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฐจ์ด๋Š” WMT ๋ชจ๋ธ์ด WMT ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ ๋ฐ€์ ‘ํ•˜๊ฒŒ ๊ด€๋ จ๋œ ๋„๋ฉ”์ธ๋ณ„ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ์–ดํœ˜ ์ผ์น˜์œจ์€ ๋†’์ง€๋งŒ ์‹ ๊ฒฝ ๊ธฐ๋ฐ˜ ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ ํ‰๊ฐ€ํ•  ๋•Œ ์˜๋ฏธ๋ก ์  ๊นŠ์ด๊ฐ€ ๋ถ€์กฑํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค. BLEU ์ ์ˆ˜๋Š” ์•ฝํ•œ ๋ชจ๋ธ์—์„œ ๊ธฐ๋ณธ ๊ธฐ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ํšจ๊ณผ์ ์ด์ง€๋งŒ, ๋‹ค์–‘ํ•œ ๋ฒˆ์—ญ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๊ณ ๊ธ‰ ๋ฒˆ์—ญ ๋ชจ๋ธ์—์„œ๋Š” ๊ทธ ์œ ์šฉ์„ฑ์ด ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋งฅ๋ฝ์—์„œ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด BLEU์—๋งŒ ์˜์กดํ•˜๋Š” ๊ฒƒ์€ ์ ์  ๋” ๊ตฌ์‹์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค.
Such distributions are ubiquitous in natural datasets, from Zipf's law (Zipf, 1935) in distribution of word frequencies, to biological data, earthquake magnitudes, financial data etc. - this is the data being consumed by large models at scale, like LLMs. But what distribution do AI-models generate when trained on such data? Figure 2 provides an empirical answer for a large scale LLM (Llama2-7B) and a transformer model trained on an arithmetic task. Regenerating heavy-tailed data affects the distribution in two possible ways: (1) "Cutting off" the tail of the distribution and/or (2) "Narrowing" the tail (see Figure 1 for a cartoon illustration). The mechanisms leading to this, apart from finite sampling bias (as already proposed in Shumailov et al. (2023) - see Section 2 for a derivation in the Zipf-setting), stem from deliberate choices in the generation algorithms of the models: in LLMs via truncated next token prediction at inference (e.g. selecting more likely tokens via _top-_ or _top-_\(k\)_ truncation, concentrating the probability distribution by lowering the temperature); in vision models like GANs via truncation or in diffusion models through guidance.
# ์ถœ๋ ฅ ์ด๋Ÿฌํ•œ ๋ถ„ํฌ๋Š” ๋‹จ์–ด ๋นˆ๋„ ๋ถ„ํฌ์˜ Zipf's law (Zipf, 1935)๋ถ€ํ„ฐ ์ƒ๋ฌผํ•™์  ๋ฐ์ดํ„ฐ, ์ง€์ง„ ์ง„๋„, ๊ธˆ์œต ๋ฐ์ดํ„ฐ ๋“ฑ๊นŒ์ง€ ์ž์—ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์–ด๋””์—๋‚˜ ์กด์žฌํ•˜๋ฉฐ, ์ด๋Š” Large Language Models (LLMs)๊ณผ ๊ฐ™์€ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์ด ๋Œ€๊ทœ๋ชจ๋กœ ์†Œ๋น„ํ•˜๋Š” ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ AI ๋ชจ๋ธ์€ ์–ด๋–ค ๋ถ„ํฌ๋ฅผ ์ƒ์„ฑํ• ๊นŒ์š”? ๊ทธ๋ฆผ 2๋Š” ๋Œ€๊ทœ๋ชจ LLM (Llama2-7B)๊ณผ ์‚ฐ์ˆ  ์ž‘์—…์œผ๋กœ ํ•™์Šต๋œ Transformer ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์  ๋‹ต์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋ฌด๊ฑฐ์šด ๊ผฌ๋ฆฌ๋ฅผ ๊ฐ€์ง„ ๋ฐ์ดํ„ฐ๋ฅผ ์žฌ์ƒ์„ฑํ•˜๋ฉด ๋ถ„ํฌ์˜ ๊ผฌ๋ฆฌ ๋ถ€๋ถ„์ด (1) "์ž˜๋ฆฌ๊ฑฐ๋‚˜" (2) ๊ผฌ๋ฆฌ๊ฐ€ "์ข์•„์ง€๋Š”" ๋‘ ๊ฐ€์ง€ ๋ฐฉ์‹์œผ๋กœ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1์„ ์ฐธ์กฐํ•˜์„ธ์š”). ์ด๋ฅผ ์œ ๋ฐœํ•˜๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ Shumailov ๋“ฑ(2023)์—์„œ ์ด๋ฏธ ์ œ์•ˆํ•œ ๊ฒƒ์ฒ˜๋Ÿผ ์œ ํ•œ ์ƒ˜ํ”Œ๋ง ํŽธํ–ฅ ์™ธ์—๋„ ๋ชจ๋ธ์˜ ์ƒ์„ฑ ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ ์˜๋„์ ์œผ๋กœ ์„ ํƒํ•œ ๊ฒƒ์—์„œ ๋น„๋กฏ๋ฉ๋‹ˆ๋‹ค: ์ถ”๋ก  ์‹œ ์ž˜๋ฆฐ ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก์„ ํ†ตํ•ด LLM์˜ ๊ฒฝ์šฐ (์˜ˆ: _top-_ ๋˜๋Š” _top-_\(k\)_ truncation์„ ํ†ตํ•ด ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ† ํฐ์„ ์„ ํƒํ•˜๊ณ , ์˜จ๋„๋ฅผ ๋‚ฎ์ถฐ ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ์ง‘์ค‘์‹œํ‚ค๋Š” ๊ฒฝ์šฐ); GAN๊ณผ ๊ฐ™์€ ๋น„์ „ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ truncation์„ ํ†ตํ•ด ๋˜๋Š” ํ™•์‚ฐ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๊ฐ€์ด๋“œ๋ฅผ ํ†ตํ•ด ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ๋ถ„ํฌ๋Š” ๋‹จ์–ด ๋นˆ๋„ ๋ถ„ํฌ์˜ ์ง€ํ”„์˜ ๋ฒ•์น™(Zipf, 1935)๋ถ€ํ„ฐ ์ƒ๋ฌผํ•™์  ๋ฐ์ดํ„ฐ, ์ง€์ง„ ์ง„๋„, ๊ธˆ์œต ๋ฐ์ดํ„ฐ ๋“ฑ๊นŒ์ง€ ์ž์—ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์–ด๋””์—๋‚˜ ์กด์žฌํ•˜๋ฉฐ, ์ด๋Š” LLM๊ณผ ๊ฐ™์€ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์ด ๋Œ€๊ทœ๋ชจ๋กœ ์†Œ๋น„ํ•˜๋Š” ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ AI ๋ชจ๋ธ์€ ์–ด๋–ค ๋ถ„ํฌ๋ฅผ ์ƒ์„ฑํ• ๊นŒ์š”? ๊ทธ๋ฆผ 2๋Š” ๋Œ€๊ทœ๋ชจ LLM(Llama2-7B)๊ณผ ์‚ฐ์ˆ  ์ž‘์—…์œผ๋กœ ํ•™์Šต๋œ ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์  ๋‹ต์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋ฌด๊ฑฐ์šด ๊ผฌ๋ฆฌ๋ฅผ ๊ฐ€์ง„ ๋ฐ์ดํ„ฐ๋ฅผ ์žฌ์ƒ์„ฑํ•˜๋ฉด ๋ถ„ํฌ์˜ ๊ผฌ๋ฆฌ ๋ถ€๋ถ„์ด (1) "์ž˜๋ฆฌ๊ฑฐ๋‚˜" (2) ๊ผฌ๋ฆฌ๊ฐ€ "์ข์•„์ง€๋Š”" ๋‘ ๊ฐ€์ง€ ๋ฐฉ์‹์œผ๋กœ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1์„ ์ฐธ์กฐํ•˜์„ธ์š”). ์ด๋ฅผ ์œ ๋ฐœํ•˜๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ Shumailov ๋“ฑ(2023)์—์„œ ์ด๋ฏธ ์ œ์•ˆํ•œ ๊ฒƒ์ฒ˜๋Ÿผ ์œ ํ•œ ์ƒ˜ํ”Œ๋ง ํŽธํ–ฅ ์™ธ์—๋„ ๋ชจ๋ธ์˜ ์ƒ์„ฑ ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ ์˜๋„์ ์œผ๋กœ ์„ ํƒํ•œ ๊ฒƒ์—์„œ ๋น„๋กฏ๋ฉ๋‹ˆ๋‹ค: ์ถ”๋ก  ์‹œ ์ž˜๋ฆฐ ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก์„ ํ†ตํ•ด LLM์˜ ๊ฒฝ์šฐ (์˜ˆ: _top-_ ๋˜๋Š” _top-_\(k\)_ ํŠธ๋Ÿฐ์ผ€์ด์…˜์„ ํ†ตํ•ด ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ† ํฐ์„ ์„ ํƒํ•˜๊ณ , ์˜จ๋„๋ฅผ ๋‚ฎ์ถฐ ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ์ง‘์ค‘์‹œํ‚ค๋Š” ๊ฒฝ์šฐ); GAN๊ณผ ๊ฐ™์€ ๋น„์ „ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํŠธ๋Ÿฐ์ผ€์ด์…˜์„ ํ†ตํ•ด ๋˜๋Š” ํ™•์‚ฐ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๊ฐ€์ด๋“œ๋ฅผ ํ†ตํ•ด ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค.
A key hyper-parameter in the Cosine LRS is the step \(T\) at which Cosine decreases to the minimum for the first time. Typically, \(T\) is set to the total training step \(S\) for training with a predefined training step. Generally, it is believed that the learning rate should be high to enable sufficient exploration. For example, Kaplan et al. (2020) demonstrate that the loss decreases when the summed learning rate over the entire training increases (see Figure 23 in their paper). This indicates setting \(T<S\) is not optimal. On the other hand, Hoffmann et al. (2022) make a key observation that setting \(T>S\) results in dropped performance while setting \(S=T\) results in improved training efficiency, confirming that the learning rate shouldn't be kept high throughout the training. To reproduce these observations, we conduct experiments on the 0.036B model. We try \(Cosine(T)\) and \(CosineLoop(T)\) LRS, following the formula shown in Appendix B.1. The result can be seen in Figure 4. We can see that when the training step is \(S=20N,40N,60N,80N\), the lowest loss is always achieved by the \(Cosine(T)\) where \(T=S\). Both \(T<S\) and \(T>S\) are not optimal.
์ฝ”์‚ฌ์ธ LRS์˜ ์ฃผ์š” ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋Š” Cosine์ด ์ตœ์†Œ๊ฐ’์œผ๋กœ ๊ฐ์†Œํ•˜๋Š” ๋‹จ๊ณ„ \(T\)์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ \(T\)๋Š” ๋ฏธ๋ฆฌ ์ •์˜๋œ ํ›ˆ๋ จ ๋‹จ๊ณ„๋กœ ํ›ˆ๋ จํ•  ๋•Œ ์ด ํ›ˆ๋ จ ๋‹จ๊ณ„ \(S\)๋กœ ์„ค์ •๋ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ถฉ๋ถ„ํ•œ ํƒ์ƒ‰์„ ์œ„ํ•ด Learning Rate๊ฐ€ ๋†’์•„์•ผ ํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Kaplan ๋“ฑ(2020)์€ ์ „์ฒด ํ›ˆ๋ จ์— ๋Œ€ํ•œ Learning Rate ํ•ฉ๊ณ„๊ฐ€ ์ฆ๊ฐ€ํ•˜๋ฉด Loss๊ฐ€ ๊ฐ์†Œํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค(๋…ผ๋ฌธ์˜ ๊ทธ๋ฆผ 23 ์ฐธ์กฐ). ์ด๋Š” \(T<S\)๋ฅผ ์„ค์ •ํ•˜๋Š” ๊ฒƒ์ด ์ตœ์ ์ด ์•„๋‹˜์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— Hoffmann ๋“ฑ(2022)์€ \(T>S\)๋ฅผ ์„ค์ •ํ•˜๋ฉด ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๋Š” ๋ฐ˜๋ฉด, \(S=T\)๋ฅผ ์„ค์ •ํ•˜๋ฉด ํ›ˆ๋ จ ํšจ์œจ์„ฑ์ด ํ–ฅ์ƒ๋œ๋‹ค๋Š” ํ•ต์‹ฌ ๊ด€์ฐฐ์„ ํ†ตํ•ด Learning Rate๋ฅผ ํ›ˆ๋ จ ๋‚ด๋‚ด ๋†’๊ฒŒ ์œ ์ง€ํ•ด์„œ๋Š” ์•ˆ ๋œ๋‹ค๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์„ ์žฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด 0.036B ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ถ€๋ก B.1์— ํ‘œ์‹œ๋œ ๊ณต์‹์— ๋”ฐ๋ผ \(Cosine(T)\) ๋ฐ \(CosineLoop(T)\) LRS๋ฅผ ์‹œ๋„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 4์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋‹จ๊ณ„๊ฐ€ \(S=20N,40N,60N,80N\)์ผ ๋•Œ, ํ•ญ์ƒ \(Cosine(T)\)์—์„œ ์ตœ์ € Loss๊ฐ€ ๋‹ฌ์„ฑ๋˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(T=S\)์ž…๋‹ˆ๋‹ค. \(T<S\)์™€ \(T>S\) ๋ชจ๋‘ ์ตœ์ ์ด ์•„๋‹™๋‹ˆ๋‹ค.
์ฝ”์‚ฌ์ธ LRS์˜ ์ฃผ์š” ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ์ฝ”์‚ฌ์ธ์ด ์ตœ์†Œ๊ฐ’์œผ๋กœ ๊ฐ์†Œํ•˜๋Š” ๋‹จ๊ณ„ \(T\)์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ \(T\)๋Š” ๋ฏธ๋ฆฌ ์ •์˜๋œ ํ›ˆ๋ จ ๋‹จ๊ณ„๋กœ ํ›ˆ๋ จํ•  ๋•Œ ์ด ํ›ˆ๋ จ ๋‹จ๊ณ„ \(S\)๋กœ ์„ค์ •๋ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ถฉ๋ถ„ํ•œ ํƒ์ƒ‰์„ ์œ„ํ•ด ํ•™์Šต ์†๋„๊ฐ€ ๋†’์•„์•ผ ํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Kaplan ๋“ฑ(2020)์€ ์ „์ฒด ํ›ˆ๋ จ์— ๋Œ€ํ•œ ํ•™์Šต๋ฅ  ํ•ฉ๊ณ„๊ฐ€ ์ฆ๊ฐ€ํ•˜๋ฉด ์†์‹ค์ด ๊ฐ์†Œํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค(๋…ผ๋ฌธ์˜ ๊ทธ๋ฆผ 23 ์ฐธ์กฐ). ์ด๋Š” \(T<S\)๋ฅผ ์„ค์ •ํ•˜๋Š” ๊ฒƒ์ด ์ตœ์ ์ด ์•„๋‹˜์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ํ˜ธํ”„๋งŒ ๋“ฑ(2022)์€ \(T>S\)๋ฅผ ์„ค์ •ํ•˜๋ฉด ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๋Š” ๋ฐ˜๋ฉด, \(S=T\)๋ฅผ ์„ค์ •ํ•˜๋ฉด ํ›ˆ๋ จ ํšจ์œจ์„ฑ์ด ํ–ฅ์ƒ๋œ๋‹ค๋Š” ํ•ต์‹ฌ ๊ด€์ฐฐ์„ ํ†ตํ•ด ํ•™์Šต๋ฅ ์„ ํ›ˆ๋ จ ๋‚ด๋‚ด ๋†’๊ฒŒ ์œ ์ง€ํ•ด์„œ๋Š” ์•ˆ ๋œ๋‹ค๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์„ ์žฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด 0.036B ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ถ€๋ก B.1์— ํ‘œ์‹œ๋œ ๊ณต์‹์— ๋”ฐ๋ผ \(Cosine(T)\) ๋ฐ \(CosineLoop(T)\) LRS๋ฅผ ์‹œ๋„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 4์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋‹จ๊ณ„๊ฐ€ \(S=20N,40N,60N,80N\)์ผ ๋•Œ, ํ•ญ์ƒ \(Cosine(T)\)์—์„œ ์ตœ์ € ์†์‹ค์ด ๋‹ฌ์„ฑ๋˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(T=S\)์ž…๋‹ˆ๋‹ค. T<S$์™€ $T>S$ ๋ชจ๋‘ ์ตœ์ ์ด ์•„๋‹™๋‹ˆ๋‹ค.
* Generally, the task performance improves as the training loss goes down, regardless of the model sizes. On MMLU, C-Eval, GSM8K, and GSM8K-Chinese, all models of three sizes perform at the random level until the pre-training loss decreases to about 2.2, after which the performance gradually climbs as the loss increases. Detailed analysis on this is shown in Section 3. * Importantly, the performance-vs-loss data points of different model sizes fall on the same trending curve. That is, by ignoring the color differences (model sizes), the data points of different models are indistinguishable. For example, when the training loss falls around 2.00, the green and orange points on TriviaQA are indistinguishable. This indicates that the model performance on downstream tasks largely correlates with the pre-training loss, _regardless of the model size._ * Interestingly, we find that the overall training loss is a good predictor of performance on both English and Chinese tasks, although it is computed on a mixture of English and Chinese tokens. This implies that the learning dynamics of English and Chinese tokens are likely very similar during multilingual pre-training.
* ์ผ๋ฐ˜์ ์œผ๋กœ ๋ชจ๋ธ ํฌ๊ธฐ์— ๊ด€๊ณ„์—†์ด ํ›ˆ๋ จ Loss๊ฐ€ ๊ฐ์†Œํ•จ์— ๋”ฐ๋ผ ์ž‘์—… ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค. MMLU, C-Eval, GSM8K ๋ฐ GSM8K-Chinese์—์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต Loss๊ฐ€ ์•ฝ 2.2๊นŒ์ง€ ๊ฐ์†Œํ•  ๋•Œ๊นŒ์ง€ ์„ธ ๊ฐ€์ง€ ํฌ๊ธฐ์˜ ๋ชจ๋“  ๋ชจ๋ธ์ด ๋ฌด์ž‘์œ„ ์ˆ˜์ค€์—์„œ ์ˆ˜ํ–‰ํ•œ ํ›„ Loss๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด ์ ์ฐจ ์ƒ์Šนํ•ฉ๋‹ˆ๋‹ค. ์ด์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋ถ„์„์€ ์„น์…˜ 3์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * ์ค‘์š”ํ•œ ๊ฒƒ์€ ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ๋‹ค๋ฅธ ๊ฒฝ์šฐ์˜ ์„ฑ๋Šฅ ๋Œ€ Loss ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๊ฐ€ ๋™์ผํ•œ ์ถ”์„ธ ๊ณก์„ ์— ์œ„์น˜ํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ฆ‰, ์ƒ‰์ƒ ์ฐจ์ด(๋ชจ๋ธ ํฌ๊ธฐ)๋ฅผ ๋ฌด์‹œํ•˜๋ฉด ๋‹ค๋ฅธ ๋ชจ๋ธ์˜ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋ฅผ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, TriviaQA์—์„œ ํ›ˆ๋ จ Loss๊ฐ€ 2.00 ๋‚ด์™ธ๋กœ ๋–จ์–ด์งˆ ๋•Œ ๋…น์ƒ‰๊ณผ ์ฃผํ™ฉ์ƒ‰ ์ ์€ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์˜ ๋ชจ๋ธ ์„ฑ๋Šฅ์ด ๋ชจ๋ธ ํฌ๊ธฐ์— ๊ด€๊ณ„์—†์ด ์‚ฌ์ „ ํ•™์Šต Loss์™€ ํฌ๊ฒŒ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ์ „์ฒด ํ›ˆ๋ จ Loss๊ฐ€ ์˜์–ด์™€ ์ค‘๊ตญ์–ด ํ† ํฐ์˜ ํ˜ผํ•ฉ์œผ๋กœ ๊ณ„์‚ฐ๋˜์ง€๋งŒ ์˜์–ด์™€ ์ค‘๊ตญ์–ด ์ž‘์—… ๋ชจ๋‘์˜ ์„ฑ๋Šฅ์„ ์ž˜ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋‹ค๊ตญ์–ด ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ์˜์–ด์™€ ์ค‘๊ตญ์–ด ํ† ํฐ์˜ ํ•™์Šต ๋™ํ•™์ด ๋งค์šฐ ์œ ์‚ฌํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
* ์ผ๋ฐ˜์ ์œผ๋กœ ๋ชจ๋ธ ํฌ๊ธฐ์— ๊ด€๊ณ„์—†์ด ํ›ˆ๋ จ ์†์‹ค์ด ๊ฐ์†Œํ•จ์— ๋”ฐ๋ผ ์ž‘์—… ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค. MMLU, C-Eval, GSM8K ๋ฐ GSM8K-Chinese์—์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์ด ์•ฝ 2.2๊นŒ์ง€ ๊ฐ์†Œํ•  ๋•Œ๊นŒ์ง€ ์„ธ ๊ฐ€์ง€ ํฌ๊ธฐ์˜ ๋ชจ๋“  ๋ชจ๋ธ์ด ๋ฌด์ž‘์œ„ ์ˆ˜์ค€์—์„œ ์ˆ˜ํ–‰ํ•œ ํ›„ ์†์‹ค์ด ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด ์ ์ฐจ ์ƒ์Šนํ•ฉ๋‹ˆ๋‹ค. ์ด์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋ถ„์„์€ ์„น์…˜ 3์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * ์ค‘์š”ํ•œ ๊ฒƒ์€ ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ๋‹ค๋ฅธ ๊ฒฝ์šฐ์˜ ์„ฑ๋Šฅ ๋Œ€ ์†์‹ค ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๊ฐ€ ๋™์ผํ•œ ์ถ”์„ธ ๊ณก์„ ์— ์œ„์น˜ํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ฆ‰, ์ƒ‰์ƒ ์ฐจ์ด(๋ชจ๋ธ ํฌ๊ธฐ)๋ฅผ ๋ฌด์‹œํ•˜๋ฉด ๋‹ค๋ฅธ ๋ชจ๋ธ์˜ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋ฅผ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํŠธ๋ฆฌ๋น„์•„QA์—์„œ ํ›ˆ๋ จ ์†์‹ค์ด 2.00 ๋‚ด์™ธ๋กœ ๋–จ์–ด์งˆ ๋•Œ ๋…น์ƒ‰๊ณผ ์ฃผํ™ฉ์ƒ‰ ์ ์€ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์˜ ๋ชจ๋ธ ์„ฑ๋Šฅ์ด ๋ชจ๋ธ ํฌ๊ธฐ์— ๊ด€๊ณ„์—†์ด ์‚ฌ์ „ ํ•™์Šต ์†์‹ค๊ณผ ํฌ๊ฒŒ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ์ „์ฒด ํ›ˆ๋ จ ์†์‹ค์ด ์˜์–ด์™€ ์ค‘๊ตญ์–ด ํ† ํฐ์˜ ํ˜ผํ•ฉ์œผ๋กœ ๊ณ„์‚ฐ๋˜์ง€๋งŒ ์˜์–ด์™€ ์ค‘๊ตญ์–ด ์ž‘์—… ๋ชจ๋‘์˜ ์„ฑ๋Šฅ์„ ์ž˜ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋‹ค๊ตญ์–ด ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ์˜์–ด์™€ ์ค‘๊ตญ์–ด ํ† ํฐ์˜ ํ•™์Šต ๋™ํ•™์ด ๋งค์šฐ ์œ ์‚ฌํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
**Language Modeling.** As the most fundamental ability of LLMs, _language modeling_ aims to predict the next token based on the previous tokens [1], which mainly focuses on the capacity of basic language understanding and generation. For evaluating such an ability, typical language modeling datasets that existing work uses include Penn Treebank [538], WikiText-103 [539], and the Pile [161], where the metric of _perplexity_ is commonly used for evaluating the model performance under the zero-shot setting. Empirical studies [93, 55] show that LLMs bring substantial performance gains over the previous state-of-the-art methods on these evaluation datasets. To better test the modeling capacity of long-range dependencies in text, the LAMBADA dataset [233] has been introduced, where LLMs are required to predict the last word of sentences based on a paragraph of context. Then, the accuracy and perplexity of the predicted last words are employed to evaluate LLMs. As shown in existing work, the performance on the language modeling tasks typically follows the scaling law [30], which means that scaling language models would improve the accuracy and reduce the perplexity.
**์–ธ์–ด ๋ชจ๋ธ๋ง.** LLM์˜ ๊ฐ€์žฅ ๊ธฐ๋ณธ์ ์ธ ๋Šฅ๋ ฅ์ธ _language modeling_์€ ์ด์ „ ํ† ํฐ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋ฉฐ, ์ฃผ๋กœ ๊ธฐ๋ณธ์ ์ธ ์–ธ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ ์‚ฌ์šฉํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ language modeling ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” Penn Treebank [538], WikiText-103 [539], Pile [161] ๋“ฑ์ด ์žˆ์œผ๋ฉฐ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด _perplexity_ ์ง€ํ‘œ๊ฐ€ ์ผ๋ฐ˜์ ์œผ๋กœ zero-shot ์„ค์ •์—์„œ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์‹ค์ฆ ์—ฐ๊ตฌ [93, 55]์— ๋”ฐ๋ฅด๋ฉด LLM์€ ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ธฐ์กด์˜ ์ตœ์ฒจ๋‹จ ๋ฐฉ๋ฒ•๋ณด๋‹ค ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ๋‹ค์ค๋‹ˆ๋‹ค. ํ…์ŠคํŠธ์—์„œ ์žฅ๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์˜ ๋ชจ๋ธ๋ง ๋Šฅ๋ ฅ์„ ๋” ์ž˜ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ๋ฌธ๋‹จ์˜ ๋ฌธ๋งฅ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฌธ์žฅ์˜ ๋งˆ์ง€๋ง‰ ๋‹จ์–ด๋ฅผ ์˜ˆ์ธกํ•ด์•ผ ํ•˜๋Š” LAMBADA ๋ฐ์ดํ„ฐ ์„ธํŠธ [233]๊ฐ€ ๋„์ž…๋˜์—ˆ์œผ๋ฉฐ, ์˜ˆ์ธก๋œ ๋งˆ์ง€๋ง‰ ๋‹จ์–ด์˜ ์ •ํ™•๋„์™€ perplexity๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด language modeling ์ž‘์—…์˜ ์„ฑ๋Šฅ์€ ์ผ๋ฐ˜์ ์œผ๋กœ scaling law [30]์„ ๋”ฐ๋ฅด๋ฉฐ, ์ด๋Š” language models๋ฅผ scalingํ•˜๋ฉด ์ •ํ™•๋„๊ฐ€ ํ–ฅ์ƒ๋˜๊ณ  perplexity๊ฐ€ ๊ฐ์†Œํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
**์–ธ์–ด ๋ชจ๋ธ๋ง.** LLM์˜ ๊ฐ€์žฅ ๊ธฐ๋ณธ์ ์ธ ๋Šฅ๋ ฅ์ธ _์–ธ์–ด ๋ชจ๋ธ๋ง_์€ ์ด์ „ ํ† ํฐ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋ฉฐ, ์ฃผ๋กœ ๊ธฐ๋ณธ์ ์ธ ์–ธ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ ์‚ฌ์šฉํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ๋ชจ๋ธ๋ง ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” Penn Treebank [538], WikiText-103 [539], Pile [161] ๋“ฑ์ด ์žˆ์œผ๋ฉฐ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด _๋‚œํ•ด์„ฑ_ ์ง€ํ‘œ๊ฐ€ ์ผ๋ฐ˜์ ์œผ๋กœ 0์ƒท ์„ค์ •์—์„œ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์‹ค์ฆ ์—ฐ๊ตฌ [93, 55]์— ๋”ฐ๋ฅด๋ฉด LLM์€ ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ธฐ์กด์˜ ์ตœ์ฒจ๋‹จ ๋ฐฉ๋ฒ•๋ณด๋‹ค ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ๋‹ค์ค๋‹ˆ๋‹ค. ํ…์ŠคํŠธ์—์„œ ์žฅ๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์˜ ๋ชจ๋ธ๋ง ๋Šฅ๋ ฅ์„ ๋” ์ž˜ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ๋ฌธ๋‹จ์˜ ๋ฌธ๋งฅ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฌธ์žฅ์˜ ๋งˆ์ง€๋ง‰ ๋‹จ์–ด๋ฅผ ์˜ˆ์ธกํ•ด์•ผ ํ•˜๋Š” LAMBADA ๋ฐ์ดํ„ฐ ์„ธํŠธ [233]๊ฐ€ ๋„์ž…๋˜์—ˆ์œผ๋ฉฐ, ์˜ˆ์ธก๋œ ๋งˆ์ง€๋ง‰ ๋‹จ์–ด์˜ ์ •ํ™•๋„์™€ ๋‚œํ•ด์„ฑ์„ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…์˜ ์„ฑ๋Šฅ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™ [30]์„ ๋”ฐ๋ฅด๋ฉฐ, ์ด๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ์Šค์ผ€์ผ๋งํ•˜๋ฉด ์ •ํ™•๋„๊ฐ€ ํ–ฅ์ƒ๋˜๊ณ  ๋‚œํ•ด์„ฑ์ด ๊ฐ์†Œํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
We focus on continual pre-training when incoming datasets are large (200B tokens+). In such settings, training is expensive, thus, it is critical to efficiently adapt to the large amount of incoming data. However, most performant LLMs (Rae et al., 2021; Hoffmann et al., 2022; Zhao et al., 2023; Touvron et al., 2023b;a) are trained with a linear warmup and cosine decay schedule with a relatively low minimum learning rate. We hypothesize that **re-warming** this learning rate to a relatively high value and subsequently re-decaying it is needed to efficiently adapt to the new dataset. To this end, in section 6.1 we study the effect of linear warmup duration, re-warming the LR, re-decaying the LR, and maximum learning rate magnitude on adaptation and forgetting. Finding that re-warming and re-decaying increases both adaptation and forgetting, in section 6.2 we investigate whether replay can help mitigate forgetting when the learning rate is re-warmed and re-decayed. Subsections 6.3 and 6.4 combine the strategies studied in the previous two sections and report their performance relative to baselines for weak and strong distribution shifts and at large model scale.
์ˆ˜์‹  ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ํฌ๋‹ค(200B ํ† ํฐ ์ด์ƒ) ๊ฒฝ์šฐ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์— ์ง‘์ค‘ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ค์ •์—์„œ๋Š” ํ•™์Šต ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ธฐ ๋•Œ๋ฌธ์— ๋“ค์–ด์˜ค๋Š” ๋Œ€๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ์— ํšจ์œจ์ ์œผ๋กœ ์ ์‘ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋Œ€๋ถ€๋ถ„์˜ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ LLM(Rae et al., 2021; Hoffmann et al., 2022; Zhao et al., 2023; Touvron et al., 2023b;a)์€ ๋น„๊ต์  ๋‚ฎ์€ ์ตœ์†Œ Learning Rate๋กœ ์„ ํ˜• ์›Œ๋ฐ์—… ๋ฐ ์ฝ”์‚ฌ์ธ ๊ฐ์‡  ์ผ์ •์œผ๋กœ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์ด Learning Rate๋ฅผ ์ƒ๋Œ€์ ์œผ๋กœ ๋†’์€ ๊ฐ’์œผ๋กœ ๋‹ค์‹œ ์›Œ๋ฐ์—…ํ•œ ํ›„ ๋‹ค์‹œ ๊ฐ์†Œ์‹œ์ผœ์•ผ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ํšจ์œจ์ ์œผ๋กœ ์ ์‘ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ์„น์…˜ 6.1์—์„œ๋Š” ์„ ํ˜• ์›Œ๋ฐ์—… ๊ธฐ๊ฐ„, Learning Rate ์žฌ๊ฐ€์—ด, Learning Rate ์žฌ๊ฐ์†Œ, ์ตœ๋Œ€ Learning Rate ํฌ๊ธฐ๊ฐ€ ์ ์‘๊ณผ ์žŠ์–ด๋ฒ„๋ฆผ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์žฌ๊ฐ€์—ด ๋ฐ ์žฌ๊ฐ์†Œ๊ฐ€ ์ ์‘๊ณผ ์žŠ์–ด๋ฒ„๋ฆผ์„ ๋ชจ๋‘ ์ฆ๊ฐ€์‹œํ‚จ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ•œ ํ›„, ์„น์…˜ 6.2์—์„œ๋Š” Learning Rate๋ฅผ ์žฌ๊ฐ€์—ด ๋ฐ ์žฌ๊ฐ์†Œ์‹œ์ผฐ์„ ๋•Œ ์žฌ์ƒ์ด ์žŠ์–ด๋ฒ„๋ฆผ์„ ์™„ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋Š”์ง€ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. 6.3๊ณผ 6.4 ํ•˜์œ„ ์„น์…˜์—์„œ๋Š” ์ด์ „ ๋‘ ์„น์…˜์—์„œ ์—ฐ๊ตฌํ•œ ์ „๋žต์„ ๊ฒฐํ•ฉํ•˜์—ฌ ์•ฝํ•œ ๋ถ„ํฌ ์ด๋™๊ณผ ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™, ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ๊ธฐ์ค€๊ณผ ๋น„๊ตํ•˜์—ฌ ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.
์ˆ˜์‹  ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ํฌ๋‹ค(200B ํ† ํฐ ์ด์ƒ) ๊ฒฝ์šฐ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์— ์ง‘์ค‘ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ค์ •์—์„œ๋Š” ํ•™์Šต ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ธฐ ๋•Œ๋ฌธ์— ๋“ค์–ด์˜ค๋Š” ๋Œ€๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ์— ํšจ์œจ์ ์œผ๋กœ ์ ์‘ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋Œ€๋ถ€๋ถ„์˜ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ LLM(Rae et al., 2021; Hoffmann et al., 2022; Zhao et al., 2023; Touvron et al., 2023b;a)์€ ๋น„๊ต์  ๋‚ฎ์€ ์ตœ์†Œ ํ•™์Šต ์†๋„๋กœ ์„ ํ˜• ์›Œ๋ฐ์—… ๋ฐ ์ฝ”์‚ฌ์ธ ๊ฐ์‡  ์ผ์ •์œผ๋กœ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์ด ํ•™์Šต ์†๋„๋ฅผ ์ƒ๋Œ€์ ์œผ๋กœ ๋†’์€ ๊ฐ’์œผ๋กœ ๋‹ค์‹œ ์›Œ๋ฐ์—…ํ•œ ํ›„ ๋‹ค์‹œ ๊ฐ์†Œ์‹œ์ผœ์•ผ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ํšจ์œจ์ ์œผ๋กœ ์ ์‘ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ์„น์…˜ 6.1์—์„œ๋Š” ์„ ํ˜• ์›Œ๋ฐ์—… ๊ธฐ๊ฐ„, LR ์žฌ๊ฐ€์—ด, LR ์žฌ๊ฐ์†Œ, ์ตœ๋Œ€ ํ•™์Šต ์†๋„ ํฌ๊ธฐ๊ฐ€ ์ ์‘๊ณผ ์žŠ์–ด๋ฒ„๋ฆผ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์žฌ๊ฐ€์—ด ๋ฐ ์žฌ๊ฐ์†Œ๊ฐ€ ์ ์‘๊ณผ ์žŠ์–ด๋ฒ„๋ฆผ์„ ๋ชจ๋‘ ์ฆ๊ฐ€์‹œํ‚จ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ•œ ํ›„, ์„น์…˜ 6.2์—์„œ๋Š” ํ•™์Šต ์†๋„๋ฅผ ์žฌ๊ฐ€์—ด ๋ฐ ์žฌ๊ฐ์†Œ์‹œ์ผฐ์„ ๋•Œ ์žฌ์ƒ์ด ์žŠ์–ด๋ฒ„๋ฆผ์„ ์™„ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋Š”์ง€ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. 6.3๊ณผ 6.4 ํ•˜์œ„ ์„น์…˜์—์„œ๋Š” ์ด์ „ ๋‘ ์„น์…˜์—์„œ ์—ฐ๊ตฌํ•œ ์ „๋žต์„ ๊ฒฐํ•ฉํ•˜์—ฌ ์•ฝํ•œ ๋ถ„ํฌ ์ด๋™๊ณผ ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™, ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ๊ธฐ์ค€๊ณผ ๋น„๊ตํ•˜์—ฌ ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.
**Prefix Tuning**. Prefix tuning [396] prepends a sequence of prefixes, which are a set of trainable continuous vectors, to each Transformer layer in language models. These prefix vectors are task-specific, which can be considered as virtual token embeddings. To optimize the prefix vectors, a reparameterization trick [396] has been proposed by learning a MLP function that maps a smaller matrix to the parameter matrix of prefixes, instead of directly optimizing the prefixes. It has been shown that this trick is useful for stable training. After optimization, the mapping function would be discarded, and only the derived prefix vectors are kept to enhance task-specific performance. Since only the prefix parameters would be trained, it can lead to a parameter-efficient model optimization. Similar to prefix tuning, p-tuning v2 [401] incorporates layer-wise prompt vectors into the Transformer architecture specially for natural language understanding, which also utilizes multi-task learning for jointly optimizing shared prompts. It has been shown to be useful in improving the model performance of different parameter scales on natural language understanding tasks.
**Prefix Tuning**. Prefix tuning [396]์€ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ์—ฐ์† ๋ฒกํ„ฐ ์ง‘ํ•ฉ์ธ ์ ‘๋‘์‚ฌ ์‹œํ€€์Šค๋ฅผ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฐ Transformer ๋ ˆ์ด์–ด์— ์„ ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ prefix ๋ฒกํ„ฐ๋Š” ์ž‘์—…๋ณ„๋กœ ์ง€์ •๋˜๋ฉฐ, ๊ฐ€์ƒ ํ† ํฐ ์ž„๋ฒ ๋”ฉ์œผ๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. prefix ๋ฒกํ„ฐ๋ฅผ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ ‘๋‘์‚ฌ๋ฅผ ์ง์ ‘ ์ตœ์ ํ™”ํ•˜๋Š” ๋Œ€์‹  ๋” ์ž‘์€ ํ–‰๋ ฌ์„ prefix์˜ ํŒŒ๋ผ๋ฏธํ„ฐ ํ–‰๋ ฌ์— ๋งคํ•‘ํ•˜๋Š” MLP ํ•จ์ˆ˜๋ฅผ ํ•™์Šตํ•˜๋Š” reparameterization trick [396]์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด trick์€ ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ์— ์œ ์šฉํ•˜๋‹ค๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ์ ํ™”๊ฐ€ ๋๋‚˜๋ฉด ๋งคํ•‘ ํ•จ์ˆ˜๋Š” ํ๊ธฐ๋˜๊ณ  ํŒŒ์ƒ๋œ prefix ๋ฒกํ„ฐ๋งŒ ์œ ์ง€๋˜์–ด ์ž‘์—…๋ณ„ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. prefix ํŒŒ๋ผ๋ฏธํ„ฐ๋งŒ ํ•™์Šต๋˜๋ฏ€๋กœ ํŒŒ๋ผ๋ฏธํ„ฐ ํšจ์œจ์ ์ธ ๋ชจ๋ธ ์ตœ์ ํ™”๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Prefix tuning๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ, p-tuning v2 [401]๋Š” ์ž์—ฐ์–ด ์ดํ•ด๋ฅผ ์œ„ํ•ด ํŠน๋ณ„ํžˆ Transformer ์•„ํ‚คํ…์ฒ˜์— ๋ ˆ์ด์–ด๋ณ„ prompt ๋ฒกํ„ฐ๋ฅผ ํ†ตํ•ฉํ•˜๋ฉฐ, ๊ณต์œ  prompt๋ฅผ ๊ณต๋™ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด multi-task learning์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ž์—ฐ์–ด ์ดํ•ด ์ž‘์—…์—์„œ ๋‹ค์–‘ํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์œ ์šฉํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
**์ ‘๋‘์‚ฌ ํŠœ๋‹**. ์ ‘๋‘์‚ฌ ํŠœ๋‹ [396]์€ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ์—ฐ์† ๋ฒกํ„ฐ ์ง‘ํ•ฉ์ธ ์ ‘๋‘์‚ฌ ์‹œํ€€์Šค๋ฅผ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฐ ํŠธ๋žœ์Šคํฌ๋จธ ๋ ˆ์ด์–ด์— ์„ ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๋‘์‚ฌ ๋ฒกํ„ฐ๋Š” ์ž‘์—…๋ณ„๋กœ ์ง€์ •๋˜๋ฉฐ, ๊ฐ€์ƒ ํ† ํฐ ์ž„๋ฒ ๋”ฉ์œผ๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ ‘๋‘์‚ฌ ๋ฒกํ„ฐ๋ฅผ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ ‘๋‘์‚ฌ๋ฅผ ์ง์ ‘ ์ตœ์ ํ™”ํ•˜๋Š” ๋Œ€์‹  ๋” ์ž‘์€ ํ–‰๋ ฌ์„ ์ ‘๋‘์‚ฌ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ ํ–‰๋ ฌ์— ๋งคํ•‘ํ•˜๋Š” MLP ํ•จ์ˆ˜๋ฅผ ํ•™์Šตํ•˜๋Š” ์žฌ๋งค๊ฐœ๋ณ€ํ™˜ ํŠธ๋ฆญ [396]์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ํŠธ๋ฆญ์€ ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ์— ์œ ์šฉํ•˜๋‹ค๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ์ ํ™”๊ฐ€ ๋๋‚˜๋ฉด ๋งคํ•‘ ํ•จ์ˆ˜๋Š” ํ๊ธฐ๋˜๊ณ  ํŒŒ์ƒ๋œ ์ ‘๋‘์‚ฌ ๋ฒกํ„ฐ๋งŒ ์œ ์ง€๋˜์–ด ์ž‘์—…๋ณ„ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ ‘๋‘์‚ฌ ํŒŒ๋ผ๋ฏธํ„ฐ๋งŒ ํ•™์Šต๋˜๋ฏ€๋กœ ํŒŒ๋ผ๋ฏธํ„ฐ ํšจ์œจ์ ์ธ ๋ชจ๋ธ ์ตœ์ ํ™”๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ ‘๋‘์‚ฌ ํŠœ๋‹๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ, p-ํŠœ๋‹ v2 [401]๋Š” ์ž์—ฐ์–ด ์ดํ•ด๋ฅผ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜์— ๋ ˆ์ด์–ด๋ณ„ ํ”„๋กฌํ”„ํŠธ ๋ฒกํ„ฐ๋ฅผ ํ†ตํ•ฉํ•˜๋ฉฐ, ๊ณต์œ  ํ”„๋กฌํ”„ํŠธ๋ฅผ ๊ณต๋™ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์ค‘ ์ž‘์—… ํ•™์Šต์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ž์—ฐ์–ด ์ดํ•ด ์ž‘์—…์—์„œ ๋‹ค์–‘ํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์œ ์šฉํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
Numerical reasoning has been regarded as one of the most essential tasks in examining the reasoning ability of large language models. As we can see, the Alpaca-33B achieves significant improvements over Plus-7B and Plus-13B models. Table 6 shows example outputs for this task. The first prompt is well-known for probing the reasoning ability, namely "which one is heavier, 1kg of cotton or 1kg of iron?". Both Plus-7B and Plus-13B failed to give a correct answer mentioning that "cotton is lighter than iron". However, 33B could correctly identify that these two things are the same weight. The second prompt is a simple calculation task, asking "how many legs for a cat and a chicken". However, as we can see, both Plus-7B and Plus-13B do not have the commonsense knowledge that a cat has four legs and two for a chicken, resulting in wrong answers. The last prompt is a numerical reasoning task to let the model predict the next number of an array. Still, only 33B model correctly identifies the pattern of the given array that the next number should be the square of its index. These observations indicate that the size of the model is vital in numerical reasoning tasks.
์ˆ˜๋ฆฌ reasoning์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ reasoning ๋Šฅ๋ ฅ์„ ๊ฒ€์‚ฌํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ํ•„์ˆ˜์ ์ธ ์ž‘์—… ์ค‘ ํ•˜๋‚˜๋กœ ๊ฐ„์ฃผ๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. Alpaca-33B๋Š” Plus-7B ๋ฐ Plus-13B ๋ชจ๋ธ์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ์ด๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 6์€ ์ด ์ž‘์—…์— ๋Œ€ํ•œ ์˜ˆ์ œ ์ถœ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ prompt๋Š” reasoning ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ ์ž˜ ์•Œ๋ ค์ง„ "์†œ 1kg๊ณผ ์ฒ  1kg ์ค‘ ์–ด๋Š ๊ฒƒ์ด ๋” ๋ฌด๊ฑฐ์šด๊ฐ€?"๋ผ๋Š” ์งˆ๋ฌธ์ž…๋‹ˆ๋‹ค. Plus-7B์™€ Plus-13B๋Š” ๋ชจ๋‘ "์†œ์ด ์ฒ ๋ณด๋‹ค ๊ฐ€๋ณ๋‹ค"๊ณ  ๋‹ตํ•˜๋ฉฐ ์ •๋‹ต์„ ์ œ์‹œํ•˜์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ 33B๋Š” ์ด ๋‘ ๊ฐ€์ง€๊ฐ€ ๊ฐ™์€ ๋ฌด๊ฒŒ๋ผ๋Š” ๊ฒƒ์„ ์ •ํ™•ํ•˜๊ฒŒ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ prompt๋Š” "๊ณ ์–‘์ด์™€ ๋‹ญ์˜ ๋‹ค๋ฆฌ๋Š” ๋ช‡ ๊ฐœ์ธ๊ฐ€์š”?"๋ผ๋Š” ๊ฐ„๋‹จํ•œ ๊ณ„์‚ฐ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ณด์‹œ๋‹ค์‹œํ”ผ Plus-7B์™€ Plus-13B ๋ชจ๋‘ ๊ณ ์–‘์ด๋Š” ๋‹ค๋ฆฌ๊ฐ€ 4๊ฐœ, ๋‹ญ์€ ๋‹ค๋ฆฌ๊ฐ€ 2๊ฐœ๋ผ๋Š” ์ƒ์‹์ ์ธ ์ง€์‹์ด ์—†์–ด ์˜ค๋‹ต์„ ๋ƒˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰ prompt๋Š” ๋ชจ๋ธ์ด ๋ฐฐ์—ด์˜ ๋‹ค์Œ ์ˆซ์ž๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ํ•˜๋Š” ์ˆ˜๋ฆฌ reasoning ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ 33B ๋ชจ๋ธ๋งŒ์ด ๋‹ค์Œ ์ˆซ์ž๊ฐ€ ํ•ด๋‹น ์ธ๋ฑ์Šค์˜ ์ œ๊ณฑ์ด์–ด์•ผ ํ•œ๋‹ค๋Š” ์ฃผ์–ด์ง„ ๋ฐฐ์—ด์˜ ํŒจํ„ด์„ ์ •ํ™•ํ•˜๊ฒŒ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ์ˆ˜๋ฆฌ reasoning ๊ณผ์ œ์—์„œ ๋ชจ๋ธ์˜ ํฌ๊ธฐ๊ฐ€ ๋งค์šฐ ์ค‘์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์ˆ˜๋ฆฌ ์ถ”๋ก ์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ๊ฒ€์‚ฌํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ํ•„์ˆ˜์ ์ธ ์ž‘์—… ์ค‘ ํ•˜๋‚˜๋กœ ๊ฐ„์ฃผ๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์•ŒํŒŒ์นด-33B๋Š” ํ”Œ๋Ÿฌ์Šค-7B ๋ฐ ํ”Œ๋Ÿฌ์Šค-13B ๋ชจ๋ธ์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ์ด๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 6์€ ์ด ์ž‘์—…์— ๋Œ€ํ•œ ์˜ˆ์ œ ์ถœ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ํ”„๋กฌํ”„ํŠธ๋Š” ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ ์ž˜ ์•Œ๋ ค์ง„ "์†œ 1kg๊ณผ ์ฒ  1kg ์ค‘ ์–ด๋Š ๊ฒƒ์ด ๋” ๋ฌด๊ฑฐ์šด๊ฐ€?"๋ผ๋Š” ์งˆ๋ฌธ์ž…๋‹ˆ๋‹ค. Plus-7B์™€ Plus-13B๋Š” ๋ชจ๋‘ "์†œ์ด ์ฒ ๋ณด๋‹ค ๊ฐ€๋ณ๋‹ค"๊ณ  ๋‹ตํ•˜๋ฉฐ ์ •๋‹ต์„ ์ œ์‹œํ•˜์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ 33B๋Š” ์ด ๋‘ ๊ฐ€์ง€๊ฐ€ ๊ฐ™์€ ๋ฌด๊ฒŒ๋ผ๋Š” ๊ฒƒ์„ ์ •ํ™•ํ•˜๊ฒŒ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ํ”„๋กฌํ”„ํŠธ๋Š” "๊ณ ์–‘์ด์™€ ๋‹ญ์˜ ๋‹ค๋ฆฌ๋Š” ๋ช‡ ๊ฐœ์ธ๊ฐ€์š”?"๋ผ๋Š” ๊ฐ„๋‹จํ•œ ๊ณ„์‚ฐ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ณด์‹œ๋‹ค์‹œํ”ผ Plus-7B์™€ Plus-13B ๋ชจ๋‘ ๊ณ ์–‘์ด๋Š” ๋‹ค๋ฆฌ๊ฐ€ 4๊ฐœ, ๋‹ญ์€ ๋‹ค๋ฆฌ๊ฐ€ 2๊ฐœ๋ผ๋Š” ์ƒ์‹์ ์ธ ์ง€์‹์ด ์—†์–ด ์˜ค๋‹ต์„ ๋ƒˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰ ํ”„๋กฌํ”„ํŠธ๋Š” ๋ชจ๋ธ์ด ๋ฐฐ์—ด์˜ ๋‹ค์Œ ์ˆซ์ž๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ํ•˜๋Š” ์ˆ˜๋ฆฌ ์ถ”๋ก  ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ 33B ๋ชจ๋ธ๋งŒ์ด ๋‹ค์Œ ์ˆซ์ž๊ฐ€ ํ•ด๋‹น ์ธ๋ฑ์Šค์˜ ์ œ๊ณฑ์ด์–ด์•ผ ํ•œ๋‹ค๋Š” ์ฃผ์–ด์ง„ ๋ฐฐ์—ด์˜ ํŒจํ„ด์„ ์ •ํ™•ํ•˜๊ฒŒ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ์ˆ˜๋ฆฌ ์ถ”๋ก  ๊ณผ์ œ์—์„œ ๋ชจ๋ธ์˜ ํฌ๊ธฐ๊ฐ€ ๋งค์šฐ ์ค‘์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
**Finance** is an important field where LLMs have promising application prospects. LLMs have been employed on various finance related tasks, such as numerical claim detection [922], financial sentiment analysis [923], financial named entity recognition [924], and financial reasoning [925]. Despite the competitive zero-shot performance exhibited by general-purpose LLMs in the finance tasks, they still underperform domain-specific PLMs containing million-scale parameters [922]. To leverage the scaling effect of LLMs, researchers collect large-scale finance corpora for continually pre-training LLMs (_e.g.,_ BloombergGPT [360], XuanYuan 2.0 [926], and FinGPT [927]). BloombergGPT has demonstrated remarkable performance across a diverse range of financial tasks while maintaining competitive performance in general-purpose tasks [360]. Nevertheless, it is imperative to consider the potential risks in the application of LLMs in finance, as the generation of inaccurate or harmful content by LLMs could have significant adverse implications for financial markets [360]. Therefore, it needs more strict reviewing and monitoring on the use of LLMs in the financial field.
**๊ธˆ์œต**์€ LLM์ด ์œ ๋งํ•œ ์‘์šฉ ์ „๋ง์„ ๊ฐ€์ง„ ์ค‘์š”ํ•œ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. LLM์€ ์ˆ˜์น˜ ์ฒญ๊ตฌ ๊ฐ์ง€[922], ๊ธˆ์œต ๊ฐ์„ฑ ๋ถ„์„[923], ๊ธˆ์œต ๋ช…๋ช…๋œ ๊ฐœ์ฒด ์ธ์‹[924], ๊ธˆ์œต reasoning[925] ๋“ฑ ๋‹ค์–‘ํ•œ ๊ธˆ์œต ๊ด€๋ จ ์—…๋ฌด์— ํ™œ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ธˆ์œต ์—…๋ฌด์—์„œ ๋ฒ”์šฉ LLM์ด ๋ณด์—ฌ์ค€ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” zero-shot ์„ฑ๋Šฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ˆ˜๋ฐฑ๋งŒ ๊ทœ๋ชจ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํฌํ•จํ•˜๋Š” ๋„๋ฉ”์ธ๋ณ„ PLM์— ๋น„ํ•ด ์—ฌ์ „ํžˆ ์„ฑ๋Šฅ์ด ์ €์กฐํ•ฉ๋‹ˆ๋‹ค[922]. LLM์˜ scaling ํšจ๊ณผ๋ฅผ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ ๋Œ€๊ทœ๋ชจ ๊ธˆ์œต ๋ง๋ญ‰์น˜๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ LLM์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค(_์˜ˆ:_ BloombergGPT [360], XuanYuan 2.0 [926], FinGPT [927]). BloombergGPT๋Š” ๋‹ค์–‘ํ•œ ๊ธˆ์œต ์—…๋ฌด์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋™์‹œ์— ๋ฒ”์šฉ ์—…๋ฌด์—์„œ๋„ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[360]. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์ด ๋ถ€์ •ํ™•ํ•˜๊ฑฐ๋‚˜ ์œ ํ•ดํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•˜๋ฉด ๊ธˆ์œต ์‹œ์žฅ์— ์‹ฌ๊ฐํ•œ ์•…์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๊ธˆ์œต ๋ถ„์•ผ์—์„œ LLM์„ ์ ์šฉํ•  ๋•Œ ์ž ์žฌ์ ์ธ ์œ„ํ—˜์„ ๊ณ ๋ คํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค[360]. ๋”ฐ๋ผ์„œ ๊ธˆ์œต ๋ถ„์•ผ์—์„œ LLM์„ ์‚ฌ์šฉํ•  ๋•Œ๋Š” ๋”์šฑ ์—„๊ฒฉํ•œ ๊ฒ€ํ† ์™€ ๋ชจ๋‹ˆํ„ฐ๋ง์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
**๊ธˆ์œต**์€ LLM์ด ์œ ๋งํ•œ ์‘์šฉ ์ „๋ง์„ ๊ฐ€์ง„ ์ค‘์š”ํ•œ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. LLM์€ ์ˆ˜์น˜ ์ฒญ๊ตฌ ๊ฐ์ง€[922], ๊ธˆ์œต ๊ฐ์„ฑ ๋ถ„์„[923], ๊ธˆ์œต ๋ช…๋ช…๋œ ๊ฐœ์ฒด ์ธ์‹[924], ๊ธˆ์œต ์ถ”๋ก [925] ๋“ฑ ๋‹ค์–‘ํ•œ ๊ธˆ์œต ๊ด€๋ จ ์—…๋ฌด์— ํ™œ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ธˆ์œต ์—…๋ฌด์—์„œ ๋ฒ”์šฉ LLM์ด ๋ณด์—ฌ์ค€ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์ œ๋กœ ์ƒท ์„ฑ๋Šฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ˆ˜๋ฐฑ๋งŒ ๊ทœ๋ชจ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํฌํ•จํ•˜๋Š” ๋„๋ฉ”์ธ๋ณ„ PLM์— ๋น„ํ•ด ์—ฌ์ „ํžˆ ์„ฑ๋Šฅ์ด ์ €์กฐํ•ฉ๋‹ˆ๋‹ค[922]. LLM์˜ ํ™•์žฅ ํšจ๊ณผ๋ฅผ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ ๋Œ€๊ทœ๋ชจ ๊ธˆ์œต ๋ง๋ญ‰์น˜๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ LLM์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค(_์˜ˆ:_ ๋ธ”๋ฃธ๋ฒ„๊ทธGPT [360], XuanYuan 2.0 [926], FinGPT [927]). ๋ธ”๋ฃธ๋ฒ„๊ทธGPT๋Š” ๋‹ค์–‘ํ•œ ๊ธˆ์œต ์—…๋ฌด์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋™์‹œ์— ๋ฒ”์šฉ ์—…๋ฌด์—์„œ๋„ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[360]. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์ด ๋ถ€์ •ํ™•ํ•˜๊ฑฐ๋‚˜ ์œ ํ•ดํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•˜๋ฉด ๊ธˆ์œต ์‹œ์žฅ์— ์‹ฌ๊ฐํ•œ ์•…์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๊ธˆ์œต ๋ถ„์•ผ์—์„œ LLM์„ ์ ์šฉํ•  ๋•Œ ์ž ์žฌ์ ์ธ ์œ„ํ—˜์„ ๊ณ ๋ คํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค[360]. ๋”ฐ๋ผ์„œ ๊ธˆ์œต ๋ถ„์•ผ์—์„œ LLM์„ ์‚ฌ์šฉํ•  ๋•Œ๋Š” ๋”์šฑ ์—„๊ฒฉํ•œ ๊ฒ€ํ† ์™€ ๋ชจ๋‹ˆํ„ฐ๋ง์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
Recent works have shown that Transformer-based (Vaswani et al., 2017) Large Language Models (LLMs) store and retrieve factual information in their parameters to complete simple prompts such as _"The mother of Stevie Wonder is"_(Petroni et al., 2019; Meng et al., 2022; Geva et al., 2021, 2022, 2023; Zhu and Li, 2023). In addition, LLMs have demonstrated remarkable _in-context_ reasoning abilities when the necessary information is explicitly given as part of the input (Wei et al., 2022). For example, models can infer "Lula" as a possible completion of _"The mother of Stevie Wonder is Lula. The singer of 'Superstition' is Stevie Wonder. The mother of the singer of 'Superstition' is"_. These findings raise a question: Do LLMs retrieve factual information stored in their parameters and perform _latent multi-hop reasoning_ when the information to reason from is _not_ given as a part of the input? For instance, when LLMs process the two-hop prompt _"The mother of the singer of 'Superstition' is"_, do they (1) figure out that "the singer of 'Superstition'" refers to Stevie Wonder and (2) use their knowledge of who Stevie Wonder's mother is to complete the prompt?
์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด Transformer ๊ธฐ๋ฐ˜(Vaswani et al., 2017) Large Language Models (LLMs)์€ ๋งค๊ฐœ๋ณ€์ˆ˜์— ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๊ณ  ๊ฒ€์ƒ‰ํ•˜์—ฌ _"์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๋Š”"_(Petroni et al., 2019; Meng et al., 2022; Geva et al., 2021, 2022, 2023; Zhu and Li, 2023)๊ณผ ๊ฐ™์€ ๊ฐ„๋‹จํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์™„๋ฃŒํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์€ ํ•„์š”ํ•œ ์ •๋ณด๊ฐ€ ์ž…๋ ฅ์˜ ์ผ๋ถ€๋กœ ๋ช…์‹œ์ ์œผ๋กœ ์ œ๊ณต๋  ๋•Œ ๋†€๋ผ์šด _In-Context_ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค(Wei et al., 2022). ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ชจ๋ธ์€ _"์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๋Š” ๋ฃฐ๋ผ์ด๋‹ค."์˜ ์™„์„ฑ ๊ฐ€๋Šฅ์„ฑ์œผ๋กœ "๋ฃฐ๋ผ"๋ฅผ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. '๋ฏธ์‹ '์˜ ๊ฐ€์ˆ˜๋Š” ์Šคํ‹ฐ๋น„ ์›๋”์ž…๋‹ˆ๋‹ค. '๋ฏธ์‹ '์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š”"_. ์ด๋Ÿฌํ•œ ๋ฐœ๊ฒฌ์€ ์˜๋ฌธ์„ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค: LLM์ด ๋งค๊ฐœ๋ณ€์ˆ˜์— ์ €์žฅ๋œ ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ณ  ์ถ”๋ก ํ•  ์ •๋ณด๊ฐ€ ์ž…๋ ฅ์˜ ์ผ๋ถ€๋กœ ์ œ๊ณต๋˜์ง€ ์•Š์„ ๋•Œ _latent multi-hop reasoning_์„ ์ˆ˜ํ–‰ํ•˜๋‚˜์š”? ์˜ˆ๋ฅผ ๋“ค์–ด, LLM์ด _"๋ฏธ์‹ '์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š”"์ด๋ผ๋Š” ๋‘ ๋‹จ๊ณ„ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ (1) "'๋ฏธ์‹ '์˜ ๊ฐ€์ˆ˜"๊ฐ€ ์Šคํ‹ฐ๋น„ ์›๋”๋ฅผ ๊ฐ€๋ฆฌํ‚จ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ์•„๋‚ด๊ณ  (2) ์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๊ฐ€ ๋ˆ„๊ตฌ์ธ์ง€์— ๋Œ€ํ•œ ์ง€์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์™„๋ฃŒํ•˜๋‚˜์š”?
์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜(Vaswani et al., 2017) ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ๋งค๊ฐœ๋ณ€์ˆ˜์— ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๊ณ  ๊ฒ€์ƒ‰ํ•˜์—ฌ _"์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๋Š”"_(Petroni et al., 2019; Meng et al., 2022; Geva et al., 2021, 2022, 2023; Zhu and Li, 2023)๊ณผ ๊ฐ™์€ ๊ฐ„๋‹จํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์™„๋ฃŒํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์€ ํ•„์š”ํ•œ ์ •๋ณด๊ฐ€ ์ž…๋ ฅ์˜ ์ผ๋ถ€๋กœ ๋ช…์‹œ์ ์œผ๋กœ ์ œ๊ณต๋  ๋•Œ ๋†€๋ผ์šด _์ปจํ…์ŠคํŠธ ๋‚ด_ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค(Wei et al., 2022). ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ชจ๋ธ์€ _"์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๋Š” ๋ฃฐ๋ผ์ด๋‹ค."์˜ ์™„์„ฑ ๊ฐ€๋Šฅ์„ฑ์œผ๋กœ "๋ฃฐ๋ผ"๋ฅผ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. '๋ฏธ์‹ '์˜ ๊ฐ€์ˆ˜๋Š” ์Šคํ‹ฐ๋น„ ์›๋”์ž…๋‹ˆ๋‹ค. '๋ฏธ์‹ '์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š”"_. ์ด๋Ÿฌํ•œ ๋ฐœ๊ฒฌ์€ ์˜๋ฌธ์„ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค: LLM์ด ๋งค๊ฐœ๋ณ€์ˆ˜์— ์ €์žฅ๋œ ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ณ  ์ถ”๋ก ํ•  ์ •๋ณด๊ฐ€ ์ž…๋ ฅ์˜ ์ผ๋ถ€๋กœ ์ œ๊ณต๋˜์ง€ ์•Š์„ ๋•Œ _์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก _์„ ์ˆ˜ํ–‰ํ•˜๋‚˜์š”? ์˜ˆ๋ฅผ ๋“ค์–ด, LLM์ด _"๋ฏธ์‹ '์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š”"์ด๋ผ๋Š” ๋‘ ๋‹จ๊ณ„ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ (1) "'๋ฏธ์‹ '์˜ ๊ฐ€์ˆ˜"๊ฐ€ ์Šคํ‹ฐ๋น„ ์›๋”๋ฅผ ๊ฐ€๋ฆฌํ‚จ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ์•„๋‚ด๊ณ  (2) ์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๊ฐ€ ๋ˆ„๊ตฌ์ธ์ง€์— ๋Œ€ํ•œ ์ง€์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์™„๋ฃŒํ•˜๋‚˜์š”?
MRKL Karpas et al. (2022) introduces a novel framework comprising scalable modules (referred to as experts) and a router. These experts can take the form of neural networks or symbols. However, this study primarily focuses on conceptualization and training an LLM specifically for mathematical computation while not delving into implementing other module contents. TALM Parisi et al. (2022) and Toolformer Schick et al. (2023) integrate a text-centric methodology with supplementary tools to enhance the capabilities of language models. They employ a self-supervise mechanism to initiate performance enhancements, commencing with a limited set of tooltips. In a similar vein, HuggingGPT Shen et al. (2023) leverages visual and speech models to process information from diverse modalities, thereby endowing LLMs with the capacity for multi-modal understanding and generation. Another question is how to select the appropriate tool. LATM Cai et al. (2023) enables the tool-making ability of LLMs to make generalized API across different tasks, and GEAR Lu et al. (2023) considers the efficiency of tool-using by using smaller models to delegate tool grounding and execution.
MRKL Karpas ๋“ฑ(2022)์€ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“ˆ(์ „๋ฌธ๊ฐ€๋ผ๊ณ  ํ•จ)๊ณผ ๋ผ์šฐํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ ์ƒˆ๋กœ์šด ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ „๋ฌธ๊ฐ€๋Š” ์‹ ๊ฒฝ๋ง์ด๋‚˜ ๊ธฐํ˜ธ์˜ ํ˜•ํƒœ๋ฅผ ์ทจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ์ˆ˜ํ•™์  ๊ณ„์‚ฐ์„ ์œ„ํ•ด Large Language Models (LLMs)์„ ๊ตฌ์ƒํ•˜๊ณ  ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ๋‹ค๋ฅธ ๋ชจ๋“ˆ ๋‚ด์šฉ์˜ ๊ตฌํ˜„์€ ๋‹ค๋ฃจ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. TALM Parisi et al. (2022)๊ณผ Toolformer Schick et al. (2023)์€ ํ…์ŠคํŠธ ์ค‘์‹ฌ ๋ฐฉ๋ฒ•๋ก ๊ณผ ๋ณด์กฐ ๋„๊ตฌ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋“ค์€ self-supervise ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ์ œํ•œ๋œ ํˆดํŒ ์„ธํŠธ์—์„œ ์‹œ์ž‘ํ•˜์—ฌ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ๋น„์Šทํ•œ ๋งฅ๋ฝ์—์„œ HuggingGPT Shen et al. (2023)์€ ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์˜ ์ •๋ณด๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ์‹œ๊ฐ ๋ฐ ์Œ์„ฑ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ LLM์ด ๋‹ค์ค‘ ๋ฐฉ์‹ ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์งˆ๋ฌธ์€ ์ ์ ˆํ•œ ๋„๊ตฌ๋ฅผ ์„ ํƒํ•˜๋Š” ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. LATM Cai et al. (2023)์€ LLM์˜ ๋„๊ตฌ ์ œ์ž‘ ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๊ฑธ์ณ ์ผ๋ฐ˜ํ™”๋œ API๋ฅผ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•˜๋ฉฐ, GEAR Lu et al. (2023)์€ ๋” ์ž‘์€ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋„๊ตฌ ์ ‘์ง€ ๋ฐ ์‹คํ–‰์„ ์œ„์ž„ํ•จ์œผ๋กœ์จ ๋„๊ตฌ ์‚ฌ์šฉ์˜ ํšจ์œจ์„ฑ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค.
MRKL Karpas ๋“ฑ(2022)์€ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“ˆ(์ „๋ฌธ๊ฐ€๋ผ๊ณ  ํ•จ)๊ณผ ๋ผ์šฐํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ ์ƒˆ๋กœ์šด ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ „๋ฌธ๊ฐ€๋Š” ์‹ ๊ฒฝ๋ง์ด๋‚˜ ๊ธฐํ˜ธ์˜ ํ˜•ํƒœ๋ฅผ ์ทจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ์ˆ˜ํ•™์  ๊ณ„์‚ฐ์„ ์œ„ํ•ด LLM์„ ๊ตฌ์ƒํ•˜๊ณ  ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ๋‹ค๋ฅธ ๋ชจ๋“ˆ ๋‚ด์šฉ์˜ ๊ตฌํ˜„์€ ๋‹ค๋ฃจ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. TALM Parisi et al. (2022)๊ณผ Toolformer Schick et al. (2023)์€ ํ…์ŠคํŠธ ์ค‘์‹ฌ ๋ฐฉ๋ฒ•๋ก ๊ณผ ๋ณด์กฐ ๋„๊ตฌ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋“ค์€ ์ž๊ฐ€ ๊ฐ๋… ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ์ œํ•œ๋œ ํˆดํŒ ์„ธํŠธ์—์„œ ์‹œ์ž‘ํ•˜์—ฌ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ๋น„์Šทํ•œ ๋งฅ๋ฝ์—์„œ HuggingGPT Shen et al. (2023)์€ ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์˜ ์ •๋ณด๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ์‹œ๊ฐ ๋ฐ ์Œ์„ฑ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ LLM์ด ๋‹ค์ค‘ ๋ฐฉ์‹ ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์งˆ๋ฌธ์€ ์ ์ ˆํ•œ ๋„๊ตฌ๋ฅผ ์„ ํƒํ•˜๋Š” ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. LATM Cai et al. (2023)์€ LLM์˜ ๋„๊ตฌ ์ œ์ž‘ ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๊ฑธ์ณ ์ผ๋ฐ˜ํ™”๋œ API๋ฅผ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•˜๋ฉฐ, GEAR Lu et al. (2023)์€ ๋” ์ž‘์€ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋„๊ตฌ ์ ‘์ง€ ๋ฐ ์‹คํ–‰์„ ์œ„์ž„ํ•จ์œผ๋กœ์จ ๋„๊ตฌ ์‚ฌ์šฉ์˜ ํšจ์œจ์„ฑ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค.
The RALM architecture initially used a single document for retrieval augmentation. However, it was discovered that RALM's performance significantly improved when the number of retrieved paragraphs was increased. (Izacard and Grave, 2020) Therefore, they proposed a new method called Fusion-in-Decoder (FiD) which involves keeping the retriever unchanged, using the encoder in LM to encode the related documents one by one, and then connecting the related documents and giving them to the decoder for output. Then Hofstatter et al. (2023) improved on the FiD. They constrained the information flow from encoder to decoder. FiD-Light with reranking was also tuned via text source pointers to improve the topmost source accuracy. Izacard and Grave (2020) applied knowledge distillation to the FiD model, also known as FiD-KD, using cross-attention scores from a sequence-to-sequence reader to obtain synthetic targets for the retriever. Singh et al. (2021) proposed an enhancement approach that differs from knowledge distillation in that it uses an end-to-end training approach requiring fewer documents, training cycles, and no supervised initialization compared to FiD-KD.
RALM ์•„ํ‚คํ…์ฒ˜๋Š” ์ฒ˜์Œ์— retrieval augmentation์— ๋‹จ์ผ ๋ฌธ์„œ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ฒ€์ƒ‰๋œ ๋‹จ๋ฝ์˜ ์ˆ˜๋ฅผ ๋Š˜๋ฆฌ๋ฉด RALM์˜ ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. (Izacard and Grave, 2020) ๋”ฐ๋ผ์„œ ๊ฒ€์ƒ‰๊ธฐ๋Š” ๊ทธ๋Œ€๋กœ ์œ ์ง€ํ•˜๊ณ , LM์˜ encoder๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ด€๋ จ ๋ฌธ์„œ๋ฅผ ํ•˜๋‚˜์”ฉ ์ธ์ฝ”๋”ฉํ•œ ๋‹ค์Œ, ๊ด€๋ จ ๋ฌธ์„œ๋ฅผ ์—ฐ๊ฒฐํ•˜์—ฌ decoder์— ์ถœ๋ ฅํ•˜๋„๋ก ํ•˜๋Š” Fusion-in-Decoder(FiD)๋ผ๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ Hofstatter ๋“ฑ(2023)์€ FiD๋ฅผ ๊ฐœ์„ ํ–ˆ์Šต๋‹ˆ๋‹ค. encoder์—์„œ decoder๋กœ์˜ ์ •๋ณด ํ๋ฆ„์„ ์ œํ•œํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํ…์ŠคํŠธ ์†Œ์Šค ํฌ์ธํ„ฐ๋ฅผ ํ†ตํ•ด ์žฌ์ˆœ์œ„ํ™” ๊ธฐ๋Šฅ์ด ์žˆ๋Š” FiD-Light๋ฅผ ์กฐ์ •ํ•˜์—ฌ ์ตœ์ƒ์œ„ ์†Œ์Šค ์ •ํ™•๋„๋ฅผ ๊ฐœ์„ ํ–ˆ์Šต๋‹ˆ๋‹ค. Izacard์™€ Grave(2020)๋Š” ์‹œํ€€์Šค ๋Œ€ ์‹œํ€€์Šค ๋ฆฌ๋”์˜ cross-attention ์ ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ FiD ๋ชจ๋ธ์— knowledge distillation์„ ์ ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰๊ธฐ์— ๋Œ€ํ•œ ํ•ฉ์„ฑ ๋Œ€์ƒ์„ ์–ป๋Š” FiD-KD๋ผ๊ณ ๋„ ํ•ฉ๋‹ˆ๋‹ค. Singh ๋“ฑ(2021)์€ FiD-KD์— ๋น„ํ•ด ๋” ์ ์€ ์ˆ˜์˜ ๋ฌธ์„œ, ํ›ˆ๋ จ ์ฃผ๊ธฐ, ์ง€๋„ ์ดˆ๊ธฐํ™”๊ฐ€ ํ•„์š”ํ•œ end-to-end ํ›ˆ๋ จ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•œ๋‹ค๋Š” ์ ์—์„œ knowledge distillation๊ณผ ๋‹ค๋ฅธ ๊ฐœ์„  ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค.
RALM ์•„ํ‚คํ…์ฒ˜๋Š” ์ฒ˜์Œ์— ๊ฒ€์ƒ‰ ์ฆ๊ฐ•์— ๋‹จ์ผ ๋ฌธ์„œ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ฒ€์ƒ‰๋œ ๋‹จ๋ฝ์˜ ์ˆ˜๋ฅผ ๋Š˜๋ฆฌ๋ฉด RALM์˜ ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. (Izacard and Grave, 2020) ๋”ฐ๋ผ์„œ ๊ฒ€์ƒ‰๊ธฐ๋Š” ๊ทธ๋Œ€๋กœ ์œ ์ง€ํ•˜๊ณ , LM์˜ ์ธ์ฝ”๋”๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ด€๋ จ ๋ฌธ์„œ๋ฅผ ํ•˜๋‚˜์”ฉ ์ธ์ฝ”๋”ฉํ•œ ๋‹ค์Œ, ๊ด€๋ จ ๋ฌธ์„œ๋ฅผ ์—ฐ๊ฒฐํ•˜์—ฌ ๋””์ฝ”๋”์— ์ถœ๋ ฅํ•˜๋„๋ก ํ•˜๋Š” Fusion-in-Decoder(FiD)๋ผ๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ Hofstatter ๋“ฑ(2023)์€ FiD๋ฅผ ๊ฐœ์„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ธ์ฝ”๋”์—์„œ ๋””์ฝ”๋”๋กœ์˜ ์ •๋ณด ํ๋ฆ„์„ ์ œํ•œํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํ…์ŠคํŠธ ์†Œ์Šค ํฌ์ธํ„ฐ๋ฅผ ํ†ตํ•ด ์žฌ์ˆœ์œ„ํ™” ๊ธฐ๋Šฅ์ด ์žˆ๋Š” FiD-Light๋ฅผ ์กฐ์ •ํ•˜์—ฌ ์ตœ์ƒ์œ„ ์†Œ์Šค ์ •ํ™•๋„๋ฅผ ๊ฐœ์„ ํ–ˆ์Šต๋‹ˆ๋‹ค. Izacard์™€ Grave(2020)๋Š” ์‹œํ€€์Šค ๋Œ€ ์‹œํ€€์Šค ๋ฆฌ๋”์˜ ๊ต์ฐจ ์ฃผ์˜๋„ ์ ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ FiD ๋ชจ๋ธ์— ์ง€์‹ ์ฆ๋ฅ˜๋ฅผ ์ ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰๊ธฐ์— ๋Œ€ํ•œ ํ•ฉ์„ฑ ๋Œ€์ƒ์„ ์–ป๋Š” FiD-KD๋ผ๊ณ ๋„ ํ•ฉ๋‹ˆ๋‹ค. Singh ๋“ฑ(2021)์€ FiD-KD์— ๋น„ํ•ด ๋” ์ ์€ ์ˆ˜์˜ ๋ฌธ์„œ, ํ›ˆ๋ จ ์ฃผ๊ธฐ, ์ง€๋„ ์ดˆ๊ธฐํ™”๊ฐ€ ํ•„์š”ํ•œ ์—”๋“œํˆฌ์—”๋“œ ํ›ˆ๋ จ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•œ๋‹ค๋Š” ์ ์—์„œ ์ง€์‹ ์ฆ๋ฅ˜์™€ ๋‹ค๋ฅธ ๊ฐœ์„  ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_BIG-bench_[70] is a collaborative benchmark intended to probe existing LLMs from various aspects. It comprises 204 tasks that encompass a broad range of topics, including linguistics, childhood development, mathematics, commonsense reasoning, biology, physics, social bias, software development, and so on. By scaling the model size, LLMs can even outperform the average human performance under the few-shot setting on 65% of tasks in BIG-bench [56]. Considering the high evaluation cost of the entire benchmark, a lightweight benchmark BIG-bench-Lite has been proposed, which contains 24 small yet diverse and challenging tasks from BIG-bench. Additionally, the BIG-bench hard (BBH) benchmark [365] has been proposed to concentrate on investigating the currently unsolvable tasks of LLMs by selecting the challenging tasks in which LLMs exhibit inferior performance compared to humans. Since BBH becomes more difficult, small models mostly achieve performance close to random. As a comparison, CoT prompting can elicit the abilities of LLMs to perform step-by-step reasoning for enhancing the performance, even exceeding the average human performance in BBH.
\(\bullet\)_BIG-bench_[70]๋Š” ๋‹ค์–‘ํ•œ ์ธก๋ฉด์—์„œ ๊ธฐ์กด LLM์„ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•œ ํ˜‘์—… benchmark์ž…๋‹ˆ๋‹ค. ์–ธ์–ดํ•™, ์•„๋™ ๋ฐœ๋‹ฌ, ์ˆ˜ํ•™, commonsense reasoning, ์ƒ๋ฌผํ•™, ๋ฌผ๋ฆฌํ•™, social bias, ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ ๋“ฑ ๊ด‘๋ฒ”์œ„ํ•œ ์ฃผ์ œ๋ฅผ ์•„์šฐ๋ฅด๋Š” 204๊ฐœ์˜ ๊ณผ์ œ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ํ™•์žฅํ•จ์œผ๋กœ์จ LLM์€ BIG-bench [56]์˜ 65%์— ๋‹ฌํ•˜๋Š” ๊ณผ์ œ์—์„œ few-shot ์„ค์ •์—์„œ ํ‰๊ท ์ ์ธ ์ธ๊ฐ„์˜ ์„ฑ๋Šฅ์„ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „์ฒด benchmark์˜ ๋†’์€ ํ‰๊ฐ€ ๋น„์šฉ์„ ๊ณ ๋ คํ•˜์—ฌ, ๊ฒฝ๋Ÿ‰ benchmark์ธ BIG-bench-Lite๊ฐ€ ์ œ์•ˆ๋˜์—ˆ๋Š”๋ฐ, ์ด benchmark์—๋Š” BIG-bench์˜ ์ž‘์ง€๋งŒ ๋‹ค์–‘ํ•˜๊ณ  ๋„์ „์ ์ธ 24๊ฐœ์˜ ์†Œ๊ทœ๋ชจ ๊ณผ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, LLM์ด ์ธ๊ฐ„์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๋„์ „ ๊ณผ์ œ๋ฅผ ์„ ๋ณ„ํ•˜์—ฌ ํ˜„์žฌ ํ•ด๊ฒฐํ•  ์ˆ˜ ์—†๋Š” ๊ณผ์ œ๋ฅผ ์ง‘์ค‘์ ์œผ๋กœ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” BIG-bench hard(BBH) benchmark [365]๊ฐ€ ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. BBH๋Š” ๋” ์–ด๋ ค์›Œ์ง€๊ธฐ ๋•Œ๋ฌธ์— ์†Œํ˜• ๋ชจ๋ธ์€ ๋Œ€๋ถ€๋ถ„ ๋ฌด์ž‘์œ„์— ๊ฐ€๊นŒ์šด ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด์™€ ๋‹ฌ๋ฆฌ CoT prompting์€ ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด์–ด BBH์—์„œ ํ‰๊ท ์ ์ธ ์ธ๊ฐ„ ์„ฑ๋Šฅ์„ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_BIG-bench_[70]๋Š” ๋‹ค์–‘ํ•œ ์ธก๋ฉด์—์„œ ๊ธฐ์กด LLM์„ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•œ ํ˜‘์—… ๋ฒค์น˜๋งˆํฌ์ž…๋‹ˆ๋‹ค. ์–ธ์–ดํ•™, ์•„๋™ ๋ฐœ๋‹ฌ, ์ˆ˜ํ•™, ์ƒ์‹ ์ถ”๋ก , ์ƒ๋ฌผํ•™, ๋ฌผ๋ฆฌํ•™, ์‚ฌํšŒ์  ํŽธ๊ฒฌ, ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ ๋“ฑ ๊ด‘๋ฒ”์œ„ํ•œ ์ฃผ์ œ๋ฅผ ์•„์šฐ๋ฅด๋Š” 204๊ฐœ์˜ ๊ณผ์ œ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ํ™•์žฅํ•จ์œผ๋กœ์จ LLM์€ BIG-bench [56]์˜ 65%์— ๋‹ฌํ•˜๋Š” ๊ณผ์ œ์—์„œ ๋ช‡ ๋ฒˆ์˜ ์ƒท ์„ค์ •์—์„œ ํ‰๊ท ์ ์ธ ์ธ๊ฐ„์˜ ์„ฑ๋Šฅ์„ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „์ฒด ๋ฒค์น˜๋งˆํฌ์˜ ๋†’์€ ํ‰๊ฐ€ ๋น„์šฉ์„ ๊ณ ๋ คํ•˜์—ฌ, ๊ฒฝ๋Ÿ‰ ๋ฒค์น˜๋งˆํฌ์ธ BIG-bench-Lite๊ฐ€ ์ œ์•ˆ๋˜์—ˆ๋Š”๋ฐ, ์ด ๋ฒค์น˜๋งˆํฌ์—๋Š” BIG-bench์˜ ์ž‘์ง€๋งŒ ๋‹ค์–‘ํ•˜๊ณ  ๋„์ „์ ์ธ 24๊ฐœ์˜ ์†Œ๊ทœ๋ชจ ๊ณผ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, LLM์ด ์ธ๊ฐ„์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๋„์ „ ๊ณผ์ œ๋ฅผ ์„ ๋ณ„ํ•˜์—ฌ ํ˜„์žฌ ํ•ด๊ฒฐํ•  ์ˆ˜ ์—†๋Š” ๊ณผ์ œ๋ฅผ ์ง‘์ค‘์ ์œผ๋กœ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” BIG-bench hard(BBH) ๋ฒค์น˜๋งˆํฌ [365]๊ฐ€ ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. BBH๋Š” ๋” ์–ด๋ ค์›Œ์ง€๊ธฐ ๋•Œ๋ฌธ์— ์†Œํ˜• ๋ชจ๋ธ์€ ๋Œ€๋ถ€๋ถ„ ๋ฌด์ž‘์œ„์— ๊ฐ€๊นŒ์šด ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด์™€ ๋‹ฌ๋ฆฌ CoT ํ”„๋กฌํ”„ํŠธ๋Š” ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด์–ด BBH์—์„œ ํ‰๊ท ์ ์ธ ์ธ๊ฐ„ ์„ฑ๋Šฅ์„ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
* If the learning rate was cosine-decayed from a large value \(\eta_{max}\) to a small value \(\eta_{min}\) during pre-training on the initial dataset, the following guidelines can help to continually pre-train your model: * Re-warming and re-decaying the learning rate from \(\mathcal{O}(\eta_{max})\) to \(\mathcal{O}(\eta_{min})\) improves adaptation to a new dataset, e.g. compared to continuing from small learning rates \(\mathcal{O}(\eta_{min})\). * Decreasing the schedule's maximum learning rate can help reduce forgetting, whereas increasing it can improve adaptation. * Infinite LR schedules are promising alternatives to cosine decay schedules. They transition into a high constant learning rate across tasks, helping prevent optimization-related forgetting by avoiding re-warming the LR between tasks. They also avoid committing to a specific budget of tokens as a final exponential decay can be used to train the model to convergence at any point during training. **Replay:** * As a default value, we recommend 5% replay. More replay should be used with stronger distribution shifts, while one can get away with as little as 1% for weak distribution shifts.
* ์ดˆ๊ธฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— Learning Rate๊ฐ€ ํฐ ๊ฐ’(\(\eta_{max}\))์—์„œ ์ž‘์€ ๊ฐ’(\(\eta_{min}\))์œผ๋กœ ์ฝ”์‚ฌ์ธ ๊ฐ์†Œํ•œ ๊ฒฝ์šฐ, ๋‹ค์Œ ์ง€์นจ์„ ์ฐธ์กฐํ•˜๋ฉด ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: * Learning Rate๋ฅผ \(\mathcal{O}(\eta_{max})\)์—์„œ \(\mathcal{O}(\eta_{min})\)๋กœ ๋‹ค์‹œ ๊ฐ€์—ดํ•˜๊ณ  ๋‹ค์‹œ ๊ฐ์†Œํ•˜๋ฉด, ์˜ˆ๋ฅผ ๋“ค์–ด ์ž‘์€ Learning Rate \(\mathcal{O}(\eta_{min})\)์—์„œ ๊ณ„์†ํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ ์‘์ด ๊ฐœ์„ ๋ฉ๋‹ˆ๋‹ค. * ์Šค์ผ€์ค„์˜ ์ตœ๋Œ€ Learning Rate๋ฅผ ๋‚ฎ์ถ”๋ฉด ์žŠ์–ด๋ฒ„๋ฆผ์„ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋†’๊ฒŒ ์„ค์ •ํ•˜๋ฉด ์ ์‘๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * ๋ฌดํ•œ LR ์Šค์ผ€์ค„์€ ์ฝ”์‚ฌ์ธ ๊ฐ์†Œ ์Šค์ผ€์ค„์— ๋Œ€ํ•œ ์œ ๋งํ•œ ๋Œ€์•ˆ์ž…๋‹ˆ๋‹ค. ์ด ์Šค์ผ€์ค„์€ ์ž‘์—… ์ „๋ฐ˜์— ๊ฑธ์ณ ๋†’์€ ์ƒ์ˆ˜ Learning Rate๋กœ ์ „ํ™˜๋˜๋ฏ€๋กœ ์ž‘์—… ๊ฐ„์— LR์„ ๋‹ค์‹œ ๊ฐ€์—ดํ•˜์ง€ ์•Š์Œ์œผ๋กœ์จ ์ตœ์ ํ™” ๊ด€๋ จ ์žŠ์–ด๋ฒ„๋ฆผ์„ ๋ฐฉ์ง€ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ตœ์ข… ์ง€์ˆ˜ ๊ฐ์†Œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ ์ค‘ ์–ด๋Š ์‹œ์ ์—์„œ๋“  ๋ชจ๋ธ์„ ์ˆ˜๋ ดํ•˜๋„๋ก ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํŠน์ • ํ† ํฐ ์˜ˆ์‚ฐ์„ ํ™•์ •ํ•˜์ง€ ์•Š์•„๋„ ๋ฉ๋‹ˆ๋‹ค. **Replay:** * ๊ธฐ๋ณธ๊ฐ’์œผ๋กœ๋Š” 5% Replay๋ฅผ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ๋” ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™์—๋Š” ๋” ๋งŽ์€ Replay๋ฅผ ์‚ฌ์šฉํ•ด์•ผ ํ•˜๋ฉฐ, ์•ฝํ•œ ๋ถ„ํฌ ์ด๋™์—๋Š” 1%๋งŒ ์‚ฌ์šฉํ•ด๋„ ๋ฉ๋‹ˆ๋‹ค.
* ์ดˆ๊ธฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ํ•™์Šต ์†๋„๊ฐ€ ํฐ ๊ฐ’(\eta_{max})์—์„œ ์ž‘์€ ๊ฐ’(\eta_{min})์œผ๋กœ ์ฝ”์‚ฌ์ธ ๊ฐ์†Œํ•œ ๊ฒฝ์šฐ, ๋‹ค์Œ ์ง€์นจ์„ ์ฐธ์กฐํ•˜๋ฉด ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:* ํ•™์Šต ์†๋„๋ฅผ \(\mathcal{O}(\eta_{max})\)์—์„œ \(\mathcal{O}(\eta_{min})\)๋กœ ๋‹ค์‹œ ๊ฐ€์—ดํ•˜๊ณ  ๋‹ค์‹œ ๊ฐ์†Œํ•˜๋ฉด, ์˜ˆ๋ฅผ ๋“ค์–ด ์ž‘์€ ํ•™์Šต ์†๋„ \(\mathcal{O}(\eta_{min})\)์—์„œ ๊ณ„์†ํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ ์‘์ด ๊ฐœ์„ ๋ฉ๋‹ˆ๋‹ค. * ์Šค์ผ€์ค„์˜ ์ตœ๋Œ€ ํ•™์Šต ์†๋„๋ฅผ ๋‚ฎ์ถ”๋ฉด ์žŠ์–ด๋ฒ„๋ฆผ์„ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋†’๊ฒŒ ์„ค์ •ํ•˜๋ฉด ์ ์‘๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * ๋ฌดํ•œ LR ์Šค์ผ€์ค„์€ ์ฝ”์‚ฌ์ธ ๊ฐ์†Œ ์Šค์ผ€์ค„์— ๋Œ€ํ•œ ์œ ๋งํ•œ ๋Œ€์•ˆ์ž…๋‹ˆ๋‹ค. ์ด ์Šค์ผ€์ค„์€ ์ž‘์—… ์ „๋ฐ˜์— ๊ฑธ์ณ ๋†’์€ ์ƒ์ˆ˜ ํ•™์Šต ์†๋„๋กœ ์ „ํ™˜๋˜๋ฏ€๋กœ ์ž‘์—… ๊ฐ„์— LR์„ ๋‹ค์‹œ ๊ฐ€์—ดํ•˜์ง€ ์•Š์Œ์œผ๋กœ์จ ์ตœ์ ํ™” ๊ด€๋ จ ์žŠ์–ด๋ฒ„๋ฆผ์„ ๋ฐฉ์ง€ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ตœ์ข… ์ง€์ˆ˜ ๊ฐ์†Œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ ์ค‘ ์–ด๋Š ์‹œ์ ์—์„œ๋“  ๋ชจ๋ธ์„ ์ˆ˜๋ ดํ•˜๋„๋ก ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํŠน์ • ํ† ํฐ ์˜ˆ์‚ฐ์„ ํ™•์ •ํ•˜์ง€ ์•Š์•„๋„ ๋ฉ๋‹ˆ๋‹ค. **๋ฆฌํ”Œ๋ ˆ์ด:*** ๊ธฐ๋ณธ๊ฐ’์œผ๋กœ๋Š” 5% ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ๋” ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™์—๋Š” ๋” ๋งŽ์€ ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ์‚ฌ์šฉํ•ด์•ผ ํ•˜๋ฉฐ, ์•ฝํ•œ ๋ถ„ํฌ ์ด๋™์—๋Š” 1%๋งŒ ์‚ฌ์šฉํ•ด๋„ ๋ฉ๋‹ˆ๋‹ค.
To further enhance the capabilities of pre-trained Large Language Models(Touvron et al., 2023; Jiang et al., 2023; Achiam et al., 2023; Azerbaijan et al., 2023) on reasoning tasks, the primary strategies involve fine-tuning and various training-free post-processing techniques. The post-processing techniques include approaches like prompting(Wei et al., 2022; Fu et al., 2022; Kojima et al., 2022; Wang et al., 2023a), ensembling(Wang et al., 2022; Li et al., 2022), and reranking(Uesato et al., 2022; Cobbe et al., 2021). A unique technique specific to LLMs is to leverage the model's reasoning capabilities through prompting. This entails crafting zero-shot prompts(Kojima et al., 2022; Wang et al., 2023a) or designing a small set of demonstrations to facilitate in-context learning(Wei et al., 2022; Fu et al., 2022). Additionally, there are methods focused on designing intricate pipelines utilizing various rules or algorithms for enhanced reasoning. These methods include search(Yao et al., 2023; Besta et al., 2023), iterative feedback collection(Yao et al., 2022), and employing the Program-of-Thought approach for generating code solutions(Chen et al., 2022).
์ถ”๋ก  ์ž‘์—…์—์„œ ์‚ฌ์ „ ํ•™์Šต๋œ Large Language Models (Touvron et al., 2023; Jiang et al., 2023; Achiam et al., 2023; Azerbaijan et al., 2023)์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ์ฃผ์š” ์ „๋žต์€ Fine-tuning๊ณผ ๋‹ค์–‘ํ•œ ๋ฌด๋ฃŒ ํ›ˆ๋ จ ํ›„์ฒ˜๋ฆฌ ๊ธฐ๋ฒ•์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ํ›„์ฒ˜๋ฆฌ ๊ธฐ๋ฒ•์—๋Š” prompting (Wei et al., 2022; Fu et al., 2022; Kojima et al., 2022; Wang et al., 2023a), ensembling (Wang et al., 2022; Li et al., 2022), ์žฌ์ˆœ์œ„ํ™” (Uesato et al., 2022; Cobbe et al., 2021)์™€ ๊ฐ™์€ ์ ‘๊ทผ ๋ฐฉ์‹์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. LLM์— ํŠนํ™”๋œ ๊ณ ์œ ํ•œ ๊ธฐ๋ฒ•์€ prompting์„ ํ†ตํ•ด ๋ชจ๋ธ์˜ ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” zero-shot prompts (Kojima et al., 2022; Wang et al., 2023a)๋ฅผ ์ž‘์„ฑํ•˜๊ฑฐ๋‚˜ ์†Œ๊ทœ๋ชจ ๋ฐ๋ชจ๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ In-Context Learning (Wei et al., 2022; Fu et al., 2022)์„ ์ด‰์ง„ํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ–ฅ์ƒ๋œ ์ถ”๋ก ์„ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๊ทœ์น™์ด๋‚˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ™œ์šฉํ•˜๋Š” ๋ณต์žกํ•œ ํŒŒ์ดํ”„๋ผ์ธ์„ ์„ค๊ณ„ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘” ๋ฐฉ๋ฒ•๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์—๋Š” ๊ฒ€์ƒ‰ (Yao et al., 2023; Besta et al., 2023), ๋ฐ˜๋ณต์  ํ”ผ๋“œ๋ฐฑ ์ˆ˜์ง‘ (Yao et al., 2022), ์ฝ”๋“œ ์†”๋ฃจ์…˜ ์ƒ์„ฑ์„ ์œ„ํ•œ Program-of-Thought ์ ‘๊ทผ ๋ฐฉ์‹ (Chen et al., 2022) ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
์ถ”๋ก  ์ž‘์—…์—์„œ ์‚ฌ์ „ ํ•™์Šต๋œ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(Touvron et al., 2023; Jiang et al., 2023; Achiam et al., 2023; Azerbaijan et al., 2023)์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ์ฃผ์š” ์ „๋žต์€ ๋ฏธ์„ธ ์กฐ์ •๊ณผ ๋‹ค์–‘ํ•œ ๋ฌด๋ฃŒ ํ›ˆ๋ จ ํ›„์ฒ˜๋ฆฌ ๊ธฐ๋ฒ•์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ํ›„์ฒ˜๋ฆฌ ๊ธฐ๋ฒ•์—๋Š” ํ”„๋กฌํ”„ํŠธ(Wei et al., 2022; Fu et al., 2022; Kojima et al., 2022; Wang et al., 2023a), ์•™์ƒ๋ธ”๋ง(Wang et al., 2022; Li et al., 2022), ์žฌ์ˆœ์œ„ํ™”(Uesato et al., 2022; Cobbe et al., 2021)์™€ ๊ฐ™์€ ์ ‘๊ทผ ๋ฐฉ์‹์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. LLM์— ํŠนํ™”๋œ ๊ณ ์œ ํ•œ ๊ธฐ๋ฒ•์€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์˜ ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ œ๋กœ์ƒท ํ”„๋กฌํ”„ํŠธ(Kojima et al., 2022; Wang et al., 2023a)๋ฅผ ์ž‘์„ฑํ•˜๊ฑฐ๋‚˜ ์†Œ๊ทœ๋ชจ ๋ฐ๋ชจ๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ ์ปจํ…์ŠคํŠธ ๋‚ด ํ•™์Šต(Wei et al., 2022; Fu et al., 2022)์„ ์ด‰์ง„ํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ–ฅ์ƒ๋œ ์ถ”๋ก ์„ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๊ทœ์น™์ด๋‚˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ™œ์šฉํ•˜๋Š” ๋ณต์žกํ•œ ํŒŒ์ดํ”„๋ผ์ธ์„ ์„ค๊ณ„ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘” ๋ฐฉ๋ฒ•๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์—๋Š” ๊ฒ€์ƒ‰(Yao et al., 2023; Besta et al., 2023), ๋ฐ˜๋ณต์  ํ”ผ๋“œ๋ฐฑ ์ˆ˜์ง‘(Yao et al., 2022), ์ฝ”๋“œ ์†”๋ฃจ์…˜ ์ƒ์„ฑ์„ ์œ„ํ•œ ํ”„๋กœ๊ทธ๋žจ ์˜ค๋ธŒ ์“ฐํŠธ ์ ‘๊ทผ ๋ฐฉ์‹(Chen et al., 2022) ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
LLM-based Multi-Agent Systems.LLM-based multi-agent systems (MAS) leverage the interaction among multiple agents for problem solving. The relationship among the multiple agents could be cooperative, competitive, or a mixture of cooperation and competition [33]. In cooperative multi-agent systems, each agent takes and assesses the information provided by other agents, thereby working together to solve complex tasks, such as role playing [46], social simulation [47] and software development [48, 49, 50, 51]. In competitive multi-agent systems, agents may detabe, negotiate and compete with each other in a game environment to achieve their goals, such as improving negotiation skills [52] and debating about the correct answer [53, 54, 55]. Some multi-agent systems may exhibit both cooperation and competition among agents. For example, WarAgent [56] models each country as an LLM-based agent to study how the interaction between countries can lead to international conflicts, where countries may cooperate with each other, such as establishing alliances and making peace agreements, or compete with each other, such as arms race, mobilization, and declaring wars.
LLM ๊ธฐ๋ฐ˜ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ. LLM ๊ธฐ๋ฐ˜ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ(MAS)์€ ๋ฌธ์ œ ํ•ด๊ฒฐ์„ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๊ด€๊ณ„๋Š” ํ˜‘๋ ฅ์ , ๊ฒฝ์Ÿ์  ๋˜๋Š” ํ˜‘๋ ฅ๊ณผ ๊ฒฝ์Ÿ์ด ํ˜ผํ•ฉ๋˜์–ด ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[33]. ํ˜‘๋ ฅ์  ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์—์„œ ๊ฐ ์—์ด์ „ํŠธ๋Š” ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๊ฐ€ ์ œ๊ณตํ•œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ํ‰๊ฐ€ํ•˜์—ฌ ์—ญํ• ๊ทน[46], ์‚ฌํšŒ ์‹œ๋ฎฌ๋ ˆ์ด์…˜[47], ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ[48, 49, 50, 51]๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํ•จ๊ป˜ ์ž‘์—…ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฝ์Ÿ์  ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์—์„œ ์—์ด์ „ํŠธ๋Š” ํ˜‘์ƒ ๊ธฐ์ˆ  ํ–ฅ์ƒ[52] ๋ฐ ์ •๋‹ต์— ๋Œ€ํ•œ ํ† ๋ก [53, 54, 55]๊ณผ ๊ฐ™์€ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ฒŒ์ž„ ํ™˜๊ฒฝ์—์„œ ์„œ๋กœ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šคํ™”ํ•˜๊ณ  ํ˜‘์ƒํ•˜๊ณ  ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์€ ์—์ด์ „ํŠธ ๊ฐ„์— ํ˜‘๋ ฅ๊ณผ ๊ฒฝ์Ÿ์„ ๋ชจ๋‘ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, WarAgent [56]๋Š” ๊ฐ ๊ตญ๊ฐ€๋ฅผ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋กœ ๋ชจ๋ธ๋งํ•˜์—ฌ ๊ตญ๊ฐ€ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์ด ์–ด๋–ป๊ฒŒ ๊ตญ์ œ ๋ถ„์Ÿ์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ๋Š”์ง€ ์—ฐ๊ตฌํ•˜๋Š”๋ฐ, ๊ตญ๊ฐ€๋“ค์€ ๋™๋งน์„ ๋งบ๊ณ  ํ‰ํ™” ํ˜‘์ •์„ ์ฒด๊ฒฐํ•˜๋Š” ๋“ฑ ์„œ๋กœ ํ˜‘๋ ฅํ•˜๊ฑฐ๋‚˜ ๊ตฐ๋น„ ๊ฒฝ์Ÿ, ๋™์›, ์„ ์ „ํฌ๊ณ  ๋“ฑ ์„œ๋กœ ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
LLM ๊ธฐ๋ฐ˜ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ.LLM ๊ธฐ๋ฐ˜ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ(MAS)์€ ๋ฌธ์ œ ํ•ด๊ฒฐ์„ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๊ด€๊ณ„๋Š” ํ˜‘๋ ฅ์ , ๊ฒฝ์Ÿ์  ๋˜๋Š” ํ˜‘๋ ฅ๊ณผ ๊ฒฝ์Ÿ์ด ํ˜ผํ•ฉ๋˜์–ด ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[33]. ํ˜‘๋ ฅ์  ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์—์„œ ๊ฐ ์—์ด์ „ํŠธ๋Š” ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๊ฐ€ ์ œ๊ณตํ•œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ํ‰๊ฐ€ํ•˜์—ฌ ์—ญํ• ๊ทน[46], ์‚ฌํšŒ ์‹œ๋ฎฌ๋ ˆ์ด์…˜[47], ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ[48, 49, 50, 51]๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํ•จ๊ป˜ ์ž‘์—…ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฝ์Ÿ์  ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์—์„œ ์—์ด์ „ํŠธ๋Š” ํ˜‘์ƒ ๊ธฐ์ˆ  ํ–ฅ์ƒ[52] ๋ฐ ์ •๋‹ต์— ๋Œ€ํ•œ ํ† ๋ก [53, 54, 55]๊ณผ ๊ฐ™์€ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ฒŒ์ž„ ํ™˜๊ฒฝ์—์„œ ์„œ๋กœ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šคํ™”ํ•˜๊ณ  ํ˜‘์ƒํ•˜๊ณ  ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์€ ์—์ด์ „ํŠธ ๊ฐ„์— ํ˜‘๋ ฅ๊ณผ ๊ฒฝ์Ÿ์„ ๋ชจ๋‘ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, WarAgent [56]๋Š” ๊ฐ ๊ตญ๊ฐ€๋ฅผ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋กœ ๋ชจ๋ธ๋งํ•˜์—ฌ ๊ตญ๊ฐ€ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์ด ์–ด๋–ป๊ฒŒ ๊ตญ์ œ ๋ถ„์Ÿ์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ๋Š”์ง€ ์—ฐ๊ตฌํ•˜๋Š”๋ฐ, ๊ตญ๊ฐ€๋“ค์€ ๋™๋งน์„ ๋งบ๊ณ  ํ‰ํ™” ํ˜‘์ •์„ ์ฒด๊ฒฐํ•˜๋Š” ๋“ฑ ์„œ๋กœ ํ˜‘๋ ฅํ•˜๊ฑฐ๋‚˜ ๊ตฐ๋น„ ๊ฒฝ์Ÿ, ๋™์›, ์„ ์ „ํฌ๊ณ  ๋“ฑ ์„œ๋กœ ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Recently, Large-scale language models (LLMs) have garnered significant attention and become the go-to approach for numerous natural language processing (NLP) tasks, including open domain conversation [1; 2; 3; 4], coding [5; 6; 7; 8; 9; 10; 11; 12; 13] and math [14; 15; 16; 17; 18; 19]. A conspicuous example is ChatGPT, developed by OpenAI. This model uses extensive pre-training on large-scale internet data and further fine-tuning with specific instruction data and methods. As a result, it achieves state-of-the-art zero-shot performance on various benchmarks. Subsequently, Anthropic, Google, and Meta also launched their competitive products one after another. Notably, Meta's series of Llama [4; 20] models have sparked an open-source revolution and quickly narrowed the gap with those closed-source LLMs. This trend also gradually stimulates the releases of MPT4, Falcon [21], StarCoder [12], Alpaca [22], Vicuna [23], and WizardLM [24], etc. However, these open models still struggles with the scenarios which require complex multi-step quantitative reasoning, such as solving mathematical and science challenges [25; 26; 27; 28; 29; 30; 31; 32; 33; 34; 35].
์ตœ๊ทผ Large-scale language models (LLMs)์€ ๊ฐœ๋ฐฉํ˜• ๋„๋ฉ”์ธ ๋Œ€ํ™”[1; 2; 3; 4], ์ฝ”๋”ฉ[5; 6; 7; 8; 9; 10; 11; 12; 13] ๋ฐ ์ˆ˜ํ•™[14; 15; 16; 17; 18; 19]์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ(NLP) ์ž‘์—…์— ์ฃผ๋ชฉ์„ ๋ฐ›๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋‹ค์–‘ํ•œ NLP ์ž‘์—…์˜ ์ตœ๊ณ  ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ OpenAI์—์„œ ๊ฐœ๋ฐœํ•œ ChatGPT๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ๋Œ€๊ทœ๋ชจ ์ธํ„ฐ๋„ท ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‚ฌ์ „ ํ•™์Šต๊ณผ ํŠน์ • instruction ๋ฐ์ดํ„ฐ ๋ฐ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๊ฐ€ Fine-tuning์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ตœ์ฒจ๋‹จ zero-shot ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ดํ›„ Anthropic, Google, Meta๋„ ์ฐจ๋ก€๋กœ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์ œํ’ˆ์„ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, Meta์˜ ์ผ๋ จ์˜ Llama [4; 20] ๋ชจ๋ธ์€ ์˜คํ”ˆ ์†Œ์Šค ํ˜๋ช…์„ ์ด‰๋ฐœํ–ˆ์œผ๋ฉฐ ํ์‡„ํ˜• LLM๊ณผ์˜ ๊ฒฉ์ฐจ๋ฅผ ๋น ๋ฅด๊ฒŒ ์ขํ˜”์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถ”์„ธ๋Š” ์ ์ฐจ MPT4, Falcon [21], StarCoder [12], Alpaca [22], Vicuna [23], WizardLM [24] ๋“ฑ์˜ ์ถœ์‹œ์—๋„ ์ž๊ทน์„ ์ฃผ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์˜คํ”ˆ ๋ชจ๋ธ์€ ์ˆ˜ํ•™ ๋ฐ ๊ณผํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ๊ณผ ๊ฐ™์ด ๋ณต์žกํ•œ ๋‹ค๋‹จ๊ณ„ ์ •๋Ÿ‰์  ์ถ”๋ก ์ด ํ•„์š”ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” ์—ฌ์ „ํžˆ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[25; 26; 27; 28; 29; 30; 31; 32; 33; 34; 35].
์ตœ๊ทผ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ๊ฐœ๋ฐฉํ˜• ๋„๋ฉ”์ธ ๋Œ€ํ™”[1; 2; 3; 4], ์ฝ”๋”ฉ[5; 6; 7; 8; 9; 10; 11; 12; 13] ๋ฐ ์ˆ˜ํ•™[14; 15; 16; 17; 18; 19]์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ(NLP) ์ž‘์—…์— ์ฃผ๋ชฉ์„ ๋ฐ›๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋‹ค์–‘ํ•œ NLP ์ž‘์—…์˜ ์ตœ๊ณ  ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ OpenAI์—์„œ ๊ฐœ๋ฐœํ•œ ChatGPT๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ๋Œ€๊ทœ๋ชจ ์ธํ„ฐ๋„ท ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‚ฌ์ „ ํ•™์Šต๊ณผ ํŠน์ • ์ง€์นจ ๋ฐ์ดํ„ฐ ๋ฐ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๊ฐ€ ๋ฏธ์„ธ ์กฐ์ •์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ตœ์ฒจ๋‹จ ์ œ๋กœ ์ƒท ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ดํ›„ Anthropic, Google, Meta๋„ ์ฐจ๋ก€๋กœ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์ œํ’ˆ์„ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, Meta์˜ ์ผ๋ จ์˜ Llama [4; 20] ๋ชจ๋ธ์€ ์˜คํ”ˆ ์†Œ์Šค ํ˜๋ช…์„ ์ด‰๋ฐœํ–ˆ์œผ๋ฉฐ ํ์‡„ํ˜• LLM๊ณผ์˜ ๊ฒฉ์ฐจ๋ฅผ ๋น ๋ฅด๊ฒŒ ์ขํ˜”์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถ”์„ธ๋Š” ์ ์ฐจ MPT4, Falcon [21], StarCoder [12], Alpaca [22], Vicuna [23], WizardLM [24] ๋“ฑ์˜ ์ถœ์‹œ์—๋„ ์ž๊ทน์„ ์ฃผ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์˜คํ”ˆ ๋ชจ๋ธ์€ ์ˆ˜ํ•™ ๋ฐ ๊ณผํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ๊ณผ ๊ฐ™์ด ๋ณต์žกํ•œ ๋‹ค๋‹จ๊ณ„ ์ •๋Ÿ‰์  ์ถ”๋ก ์ด ํ•„์š”ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” ์—ฌ์ „ํžˆ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[25; 26; 27; 28; 29; 30; 31; 32; 33; 34; 35].
However, it is important to acknowledge that low-fidelity synthetic human preference data might be limited in accurately reflecting nuanced human judgment (Argyle et al., 2023). Consequently, the resulting models may be less robust under "jail-breaking attacks" (Deshpande et al., 2023; Huang et al., 2023a), and may reveal strategically deceptive behavior even through safety training (Everitt et al., 2021; Pan et al., 2022; Steinhardt, 2022). To mitigate these risks, researchers must continuously refine and improve the quality and diversity of synthetic data, incorporating more complex and comprehensive scenarios that better capture the intricacies of human values and preferences. Additionally, combining synthetic data with real-world data, and creating synthetic data in an interactive environment which can be synced with the real world, are promising remedies. As the need for effective AI governance and regulation grows, synthetic data will play an increasingly vital role in enabling scalable oversight mechanisms that promote trust, accountability, and the development of AI technologies that are aligned with human values and societal expectations.
๊ทธ๋Ÿฌ๋‚˜ ์ €์ถฉ์‹ค๋„ ํ•ฉ์„ฑ ์ธ๊ฐ„ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋Š” ๋ฏธ๋ฌ˜ํ•œ ์ธ๊ฐ„์˜ ํŒ๋‹จ์„ ์ •ํ™•ํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๋Š” ๋ฐ ํ•œ๊ณ„๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ์ธ์ •ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค(Argyle et al., 2023). ๊ฒฐ๊ณผ์ ์œผ๋กœ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ "jail-breaking attacks"(Deshpande et al., 2023; Huang et al., 2023a)์—์„œ robustness๊ฐ€ ๋–จ์–ด์งˆ ์ˆ˜ ์žˆ์œผ๋ฉฐ, safety training์„ ํ†ตํ•ด์„œ๋„ ์ „๋žต์ ์œผ๋กœ ๊ธฐ๋งŒ์ ์ธ ํ–‰๋™์„ ๋“œ๋Ÿฌ๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Everitt et al., 2021; Pan et al., 2022; Steinhardt, 2022). ์ด๋Ÿฌํ•œ ์œ„ํ—˜์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์„ ํ˜ธ๋„์˜ ๋ณต์žก์„ฑ์„ ๋” ์ž˜ ํฌ์ฐฉํ•˜๋Š” ๋” ๋ณต์žกํ•˜๊ณ  ํฌ๊ด„์ ์ธ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์„ ์ง€์†์ ์œผ๋กœ ๊ฐœ์„ ํ•˜๊ณ  ๊ฐœ์„ ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‹ค์ œ ๋ฐ์ดํ„ฐ์™€ ๊ฒฐํ•ฉํ•˜๊ณ  ์‹ค์ œ ์„ธ๊ณ„์™€ ๋™๊ธฐํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๋Œ€ํ™”ํ˜• ํ™˜๊ฒฝ์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์€ ์œ ๋งํ•œ ์น˜๋ฃŒ๋ฒ•์ž…๋‹ˆ๋‹ค. ํšจ๊ณผ์ ์ธ AI ๊ฑฐ๋ฒ„๋„Œ์Šค ๋ฐ ๊ทœ์ œ์˜ ํ•„์š”์„ฑ์ด ์ปค์ง์— ๋”ฐ๋ผ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์‹ ๋ขฐ, ์ฑ…์ž„์„ฑ, ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์‚ฌํšŒ์  ๊ธฐ๋Œ€์— ๋ถ€ํ•ฉํ•˜๋Š” AI ๊ธฐ์ˆ  ๊ฐœ๋ฐœ์„ ์ด‰์ง„ํ•˜๋Š” ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๊ฐ๋… ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐ ์ ์  ๋” ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ ์ €์ถฉ์‹ค๋„ ํ•ฉ์„ฑ ์ธ๊ฐ„ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋Š” ๋ฏธ๋ฌ˜ํ•œ ์ธ๊ฐ„์˜ ํŒ๋‹จ์„ ์ •ํ™•ํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๋Š” ๋ฐ ํ•œ๊ณ„๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ์ธ์ •ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค(Argyle et al., 2023). ๊ฒฐ๊ณผ์ ์œผ๋กœ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ "์ œ์ผ ํƒˆ์ทจ ๊ณต๊ฒฉ"(Deshpande et al., 2023; Huang et al., 2023a)์—์„œ ๊ฒฌ๊ณ ์„ฑ์ด ๋–จ์–ด์งˆ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์•ˆ์ „ ๊ต์œก์„ ํ†ตํ•ด์„œ๋„ ์ „๋žต์ ์œผ๋กœ ๊ธฐ๋งŒ์ ์ธ ํ–‰๋™์„ ๋“œ๋Ÿฌ๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Everitt et al., 2021; Pan et al., 2022; Steinhardt, 2022). ์ด๋Ÿฌํ•œ ์œ„ํ—˜์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์„ ํ˜ธ๋„์˜ ๋ณต์žก์„ฑ์„ ๋” ์ž˜ ํฌ์ฐฉํ•˜๋Š” ๋” ๋ณต์žกํ•˜๊ณ  ํฌ๊ด„์ ์ธ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์„ ์ง€์†์ ์œผ๋กœ ๊ฐœ์„ ํ•˜๊ณ  ๊ฐœ์„ ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‹ค์ œ ๋ฐ์ดํ„ฐ์™€ ๊ฒฐํ•ฉํ•˜๊ณ  ์‹ค์ œ ์„ธ๊ณ„์™€ ๋™๊ธฐํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๋Œ€ํ™”ํ˜• ํ™˜๊ฒฝ์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์€ ์œ ๋งํ•œ ์น˜๋ฃŒ๋ฒ•์ž…๋‹ˆ๋‹ค. ํšจ๊ณผ์ ์ธ AI ๊ฑฐ๋ฒ„๋„Œ์Šค ๋ฐ ๊ทœ์ œ์˜ ํ•„์š”์„ฑ์ด ์ปค์ง์— ๋”ฐ๋ผ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์‹ ๋ขฐ, ์ฑ…์ž„์„ฑ, ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์‚ฌํšŒ์  ๊ธฐ๋Œ€์— ๋ถ€ํ•ฉํ•˜๋Š” AI ๊ธฐ์ˆ  ๊ฐœ๋ฐœ์„ ์ด‰์ง„ํ•˜๋Š” ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๊ฐ๋… ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐ ์ ์  ๋” ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
The utility of medical AI systems could be greatly improved if they are better able to interact conversationally, anchoring on large-scale medical knowledge while communicating with appropriate levels of empathy and trust. This research demonstrates the significant potential capabilities of LLM based AI systems for settings involving clinical history-taking and diagnostic dialogue. The performance of AMIE in simulated consultations represents a milestone for the field, as it was assessed along an evaluation framework that considered multiple clinically-relevant axes for conversational diagnostic medical AI. However, the results should be interpreted with appropriate caution. Translating from this limited scope of experimental simulated history-taking and diagnostic dialogue, towards real-world tools for people and those who provide care for them, requires significant additional research and development to ensure the safety, reliability, fairness, efficacy, and privacy of the technology. If successful, we believe AI systems such as AMIE can be at the core of next generation learning health systems that help scale world class healthcare to everyone.
์˜๋ฃŒ AI ์‹œ์Šคํ…œ์ด ๋Œ€๊ทœ๋ชจ ์˜๋ฃŒ ์ง€์‹์— ๊ธฐ๋ฐ˜์„ ๋‘๊ณ  ์ ์ ˆํ•œ ์ˆ˜์ค€์˜ ๊ณต๊ฐ๊ณผ ์‹ ๋ขฐ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๋Œ€ํ™”ํ•  ์ˆ˜ ์žˆ๋„๋ก ๊ฐœ์„ ๋œ๋‹ค๋ฉด ๊ทธ ์œ ์šฉ์„ฑ์€ ํฌ๊ฒŒ ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™”์™€ ๊ด€๋ จ๋œ ํ™˜๊ฒฝ์—์„œ Large Language Models (LLMs) ๊ธฐ๋ฐ˜ AI ์‹œ์Šคํ…œ์˜ ์ƒ๋‹นํ•œ ์ž ์žฌ์  ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ์ƒ๋‹ด์—์„œ AMIE์˜ ์„ฑ๋Šฅ์€ ๋Œ€ํ™”ํ˜• ์ง„๋‹จ ์˜๋ฃŒ AI์— ๋Œ€ํ•œ ์—ฌ๋Ÿฌ ์ž„์ƒ ๊ด€๋ จ ์ถ•์„ ๊ณ ๋ คํ•œ ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ์— ๋”ฐ๋ผ ํ‰๊ฐ€๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ด ๋ถ„์•ผ์˜ ์ด์ •ํ‘œ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ์ ์ ˆํ•œ ์ฃผ์˜๋ฅผ ๊ธฐ์šธ์—ฌ ํ•ด์„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ œํ•œ๋œ ๋ฒ”์œ„์˜ ์‹คํ—˜์  ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™”์—์„œ ์‹ค์ œ ๋„๊ตฌ๋กœ ๋ฒˆ์—ญํ•˜์—ฌ ์‚ฌ๋žŒ๊ณผ ๊ทธ๋“ค์„ ๋Œ๋ณด๋Š” ์‚ฌ๋žŒ๋“ค์„ ์œ„ํ•œ ์‹ค์ œ ๋„๊ตฌ๋กœ ๊ฐœ๋ฐœํ•˜๋ ค๋ฉด ๊ธฐ์ˆ ์˜ ์•ˆ์ „์„ฑ, ์‹ ๋ขฐ์„ฑ, ๊ณต์ •์„ฑ, ํšจ๋Šฅ ๋ฐ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ๋ฅผ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์ƒ๋‹นํ•œ ์ถ”๊ฐ€ ์—ฐ๊ตฌ์™€ ๊ฐœ๋ฐœ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์„ฑ๊ณตํ•˜๋ฉด AMIE์™€ ๊ฐ™์€ AI ์‹œ์Šคํ…œ์ด ๋ชจ๋“  ์‚ฌ๋žŒ์—๊ฒŒ ์„ธ๊ณ„ ์ตœ๊ณ  ์ˆ˜์ค€์˜ ์˜๋ฃŒ ์„œ๋น„์Šค๋ฅผ ํ™•์žฅํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ์ฐจ์„ธ๋Œ€ ํ•™์Šต ์˜๋ฃŒ ์‹œ์Šคํ…œ์˜ ํ•ต์‹ฌ์ด ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค.
์˜๋ฃŒ AI ์‹œ์Šคํ…œ์ด ๋Œ€๊ทœ๋ชจ ์˜๋ฃŒ ์ง€์‹์— ๊ธฐ๋ฐ˜์„ ๋‘๊ณ  ์ ์ ˆํ•œ ์ˆ˜์ค€์˜ ๊ณต๊ฐ๊ณผ ์‹ ๋ขฐ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๋Œ€ํ™”ํ•  ์ˆ˜ ์žˆ๋„๋ก ๊ฐœ์„ ๋œ๋‹ค๋ฉด ๊ทธ ์œ ์šฉ์„ฑ์€ ํฌ๊ฒŒ ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™”์™€ ๊ด€๋ จ๋œ ํ™˜๊ฒฝ์—์„œ LLM ๊ธฐ๋ฐ˜ AI ์‹œ์Šคํ…œ์˜ ์ƒ๋‹นํ•œ ์ž ์žฌ์  ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ์ƒ๋‹ด์—์„œ AMIE์˜ ์„ฑ๋Šฅ์€ ๋Œ€ํ™”ํ˜• ์ง„๋‹จ ์˜๋ฃŒ AI์— ๋Œ€ํ•œ ์—ฌ๋Ÿฌ ์ž„์ƒ ๊ด€๋ จ ์ถ•์„ ๊ณ ๋ คํ•œ ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ์— ๋”ฐ๋ผ ํ‰๊ฐ€๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ด ๋ถ„์•ผ์˜ ์ด์ •ํ‘œ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ์ ์ ˆํ•œ ์ฃผ์˜๋ฅผ ๊ธฐ์šธ์—ฌ ํ•ด์„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ œํ•œ๋œ ๋ฒ”์œ„์˜ ์‹คํ—˜์  ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™”์—์„œ ์‹ค์ œ ๋„๊ตฌ๋กœ ๋ฒˆ์—ญํ•˜์—ฌ ์‚ฌ๋žŒ๊ณผ ๊ทธ๋“ค์„ ๋Œ๋ณด๋Š” ์‚ฌ๋žŒ๋“ค์„ ์œ„ํ•œ ์‹ค์ œ ๋„๊ตฌ๋กœ ๊ฐœ๋ฐœํ•˜๋ ค๋ฉด ๊ธฐ์ˆ ์˜ ์•ˆ์ „์„ฑ, ์‹ ๋ขฐ์„ฑ, ๊ณต์ •์„ฑ, ํšจ๋Šฅ ๋ฐ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ๋ฅผ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์ƒ๋‹นํ•œ ์ถ”๊ฐ€ ์—ฐ๊ตฌ์™€ ๊ฐœ๋ฐœ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์„ฑ๊ณตํ•˜๋ฉด AMIE์™€ ๊ฐ™์€ AI ์‹œ์Šคํ…œ์ด ๋ชจ๋“  ์‚ฌ๋žŒ์—๊ฒŒ ์„ธ๊ณ„ ์ตœ๊ณ  ์ˆ˜์ค€์˜ ์˜๋ฃŒ ์„œ๋น„์Šค๋ฅผ ํ™•์žฅํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ์ฐจ์„ธ๋Œ€ ํ•™์Šต ์˜๋ฃŒ ์‹œ์Šคํ…œ์˜ ํ•ต์‹ฌ์ด ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค.
* **None**: As a lower bound, we evaluate the faithfulness of claims without any evidence from the book. * **Human evidence**: We can also use human-annotated evidence from Fables obtained via the pipeline described in SS2. This evidence is always related to the claim, but it often takes the form of short, highly-contextual spans that may or may not be sufficient to support claim verification. * **BM25 retrieval**: We employ BM25 (Robertson et al., 1995) to retrieve passages from the book using the claim as a query. We concatenate the \(k\) most relevant passages to use as evidence for our evaluation prompt. We set \(k=5\) and chunk passages up to 256 tokens. See SS4 for performance changes when varying passage length. * **Entire book (EB)**: Retrieval is especially challenging in our setting due to the complexity of both the query and document. Intuitively, long-context LLMs can bypass explicit retrieval by simply fitting the entire book into the context as evidence. This setting resembles "needle-in-the-haystack" evaluations of prior work (Kamradt, 2023; Levy et al., 2024), except that it tests a much deeper understanding of the input document.
* **None**: ํ•˜ํ•œ์œผ๋กœ, ์ฑ…์˜ ์ฆ๊ฑฐ ์—†์ด ํด๋ ˆ์ž„์˜ ์ง„์‹ค์„ฑ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. * **Human evidence**: SS2์— ์„ค๋ช…๋œ ํŒŒ์ดํ”„๋ผ์ธ์„ ํ†ตํ•ด ์–ป์€ ์šฐํ™”์—์„œ ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹ฌํ•œ evidence๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด evidence๋Š” ํ•ญ์ƒ ํด๋ ˆ์ž„๊ณผ ๊ด€๋ จ์ด ์žˆ์ง€๋งŒ, ํด๋ ˆ์ž„ ๊ฒ€์ฆ์„ ๋’ท๋ฐ›์นจํ•˜๊ธฐ์— ์ถฉ๋ถ„ํ•  ์ˆ˜๋„ ์žˆ๊ณ  ๊ทธ๋ ‡์ง€ ์•Š์„ ์ˆ˜๋„ ์žˆ๋Š” ์งง๊ณ  ์ƒํ™ฉ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๋Š” ๋ฒ”์œ„์˜ ํ˜•ํƒœ๋ฅผ ๊ฐ–๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. * **BM25 retrieval**: BM25(Robertson et al., 1995)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํด๋ ˆ์ž„์„ ์ฟผ๋ฆฌ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ์ฑ…์—์„œ ๊ตฌ์ ˆ์„ retrievalํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๊ด€๋ จ์„ฑ์ด ๋†’์€ ๊ตฌ์ ˆ \(k\)๋ฅผ ์—ฐ๊ฒฐํ•˜์—ฌ ํ‰๊ฐ€ ํ”„๋กฌํ”„ํŠธ์˜ evidence๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. k=5๋กœ ์„ค์ •ํ•˜๊ณ  ์ตœ๋Œ€ 256 ํ† ํฐ๊นŒ์ง€ ๊ตฌ์ ˆ์„ ์ชผ๊ฐœ๋Š” ๋ฐฉ์‹์œผ๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ ˆ ๊ธธ์ด๋ฅผ ๋ณ€๊ฒฝํ•  ๋•Œ ์„ฑ๋Šฅ ๋ณ€ํ™”๋Š” SS4๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. * **Entire book (EB)**: ์ฟผ๋ฆฌ์™€ ๋ฌธ์„œ ๋ชจ๋‘์˜ ๋ณต์žก์„ฑ์œผ๋กœ ์ธํ•ด ์ด ์„ค์ •์—์„œ retrieval์€ ํŠนํžˆ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์ง๊ด€์ ์œผ๋กœ, ๊ธด ๋ฌธ๋งฅ์˜ LLM์€ ๋‹จ์ˆœํžˆ ์ „์ฒด ์ฑ…์„ ์ปจํ…์ŠคํŠธ์— evidence๋กœ ๋„ฃ์–ด์ฃผ๋Š” ๊ฒƒ๋งŒ์œผ๋กœ ๋ช…์‹œ์ ์ธ retrieval์„ ์šฐํšŒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ ์ด์ „ ์ž‘์—…์˜ "๊ฑด์ดˆ๋”๋ฏธ ์†์˜ ๋ฐ”๋Š˜" ํ‰๊ฐ€(Kamradt, 2023; Levy et al., 2024)์™€ ์œ ์‚ฌํ•˜์ง€๋งŒ, ์ž…๋ ฅ ๋ฌธ์„œ์— ๋Œ€ํ•œ ํ›จ์”ฌ ๋” ๊นŠ์€ ์ดํ•ด๋ฅผ ํ…Œ์ŠคํŠธํ•œ๋‹ค๋Š” ์ ์ด ๋‹ค๋ฆ…๋‹ˆ๋‹ค.
* **์—†์Œ**: ํ•˜ํ•œ์œผ๋กœ, ์ฑ…์˜ ์ฆ๊ฑฐ ์—†์ด ํด๋ ˆ์ž„์˜ ์ง„์‹ค์„ฑ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. * **์ธ๊ฐ„ ์ฆ๊ฑฐ**: SS2์— ์„ค๋ช…๋œ ํŒŒ์ดํ”„๋ผ์ธ์„ ํ†ตํ•ด ์–ป์€ ์šฐํ™”์—์„œ ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹ฌํ•œ ์ฆ๊ฑฐ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ฆ๊ฑฐ๋Š” ํ•ญ์ƒ ํด๋ ˆ์ž„๊ณผ ๊ด€๋ จ์ด ์žˆ์ง€๋งŒ, ํด๋ ˆ์ž„ ๊ฒ€์ฆ์„ ๋’ท๋ฐ›์นจํ•˜๊ธฐ์— ์ถฉ๋ถ„ํ•  ์ˆ˜๋„ ์žˆ๊ณ  ๊ทธ๋ ‡์ง€ ์•Š์„ ์ˆ˜๋„ ์žˆ๋Š” ์งง๊ณ  ์ƒํ™ฉ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๋Š” ๋ฒ”์œ„์˜ ํ˜•ํƒœ๋ฅผ ๊ฐ–๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. * **BM25 ๊ฒ€์ƒ‰**: BM25(Robertson et al., 1995)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํด๋ ˆ์ž„์„ ์ฟผ๋ฆฌ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ์ฑ…์—์„œ ๊ตฌ์ ˆ์„ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๊ด€๋ จ์„ฑ์ด ๋†’์€ ๊ตฌ์ ˆ \(k\)๋ฅผ ์—ฐ๊ฒฐํ•˜์—ฌ ํ‰๊ฐ€ ํ”„๋กฌํ”„ํŠธ์˜ ์ฆ๊ฑฐ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. k=5๋กœ ์„ค์ •ํ•˜๊ณ  ์ตœ๋Œ€ 256 ํ† ํฐ๊นŒ์ง€ ๊ตฌ์ ˆ์„ ์ชผ๊ฐœ๋Š” ๋ฐฉ์‹์œผ๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ ˆ ๊ธธ์ด๋ฅผ ๋ณ€๊ฒฝํ•  ๋•Œ ์„ฑ๋Šฅ ๋ณ€ํ™”๋Š” SS4๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. * **์ „์ฒด ์ฑ…(EB)**: ์ฟผ๋ฆฌ์™€ ๋ฌธ์„œ ๋ชจ๋‘์˜ ๋ณต์žก์„ฑ์œผ๋กœ ์ธํ•ด ์ด ์„ค์ •์—์„œ ๊ฒ€์ƒ‰์€ ํŠนํžˆ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์ง๊ด€์ ์œผ๋กœ, ๊ธด ๋ฌธ๋งฅ์˜ LLM์€ ๋‹จ์ˆœํžˆ ์ „์ฒด ์ฑ…์„ ์ปจํ…์ŠคํŠธ์— ์ฆ๊ฑฐ๋กœ ๋„ฃ์–ด์ฃผ๋Š” ๊ฒƒ๋งŒ์œผ๋กœ ๋ช…์‹œ์ ์ธ ๊ฒ€์ƒ‰์„ ์šฐํšŒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ ์ด์ „ ์ž‘์—…์˜ "๊ฑด์ดˆ๋”๋ฏธ ์†์˜ ๋ฐ”๋Š˜" ํ‰๊ฐ€(Kamradt, 2023; Levy et al., 2024)์™€ ์œ ์‚ฌํ•˜์ง€๋งŒ, ์ž…๋ ฅ ๋ฌธ์„œ์— ๋Œ€ํ•œ ํ›จ์”ฌ ๋” ๊นŠ์€ ์ดํ•ด๋ฅผ ํ…Œ์ŠคํŠธํ•œ๋‹ค๋Š” ์ ์ด ๋‹ค๋ฆ…๋‹ˆ๋‹ค.
* **BoolQ**[188] is a yes/no question-answering dataset where the goal is reading comprehension task. BoolQ includes \(15,942\) examples. Each example is a triplet that includes a question, a relevant paragraph, and the solution. Although the main intuition behind this dataset is for reading comprehension, it can be used for reasoning, natural language inference, and question-answering tasks. * **MultiRC**[189] is another dataset that fits reading comprehension task. MultiRC contains brief paragraphs as well as multi-sentence questions that can be answered using the information in the paragraph. The paragraphs in this dataset come from a variety of sources, including news, fiction, historical texts, Wikipedia articles, discussions on society and law, elementary school science textbooks, and 9/11 reports. Each question has many response choices, with one or more of them being correct. Answering the questions requires reasoning across several sentences. MultiRC dataset encompasses around \(6,000\) multi-sentence questions gathered from over 800 paragraphs. On average, each question offers about two valid answer alternatives out of a total of five.
* **BoolQ**[188]๋Š” ์˜ˆ/์•„๋‹ˆ์š” question-answering ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ, ๋ชฉํ‘œ๋Š” reading comprehension task์ž…๋‹ˆ๋‹ค. BoolQ์—๋Š” \(15,942\)๊ฐœ์˜ ์˜ˆ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์˜ˆ์ œ๋Š” ์งˆ๋ฌธ, ๊ด€๋ จ ๋‹จ๋ฝ, ํ•ด๋‹ต์„ ํฌํ•จํ•˜๋Š” ์‚ผ์ค‘ ํ•ญ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ฃผ์š” ์ง๊ด€์€ reading comprehension์— ์žˆ์ง€๋งŒ reasoning, natural language inference ๋ฐ question-answering ์ž‘์—…์—๋„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * **MultiRC**[189]๋Š” reading comprehension task์— ์ ํ•ฉํ•œ ๋˜ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. MultiRC์—๋Š” ๊ฐ„๋‹จํ•œ ๋‹จ๋ฝ๊ณผ ๋‹จ๋ฝ์˜ ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ตํ•  ์ˆ˜ ์žˆ๋Š” multi-sentence ์งˆ๋ฌธ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋‹จ๋ฝ์€ ๋‰ด์Šค, ์†Œ์„ค, ์—ญ์‚ฌ์  ํ…์ŠคํŠธ, Wikipedia ๊ธฐ์‚ฌ, ์‚ฌํšŒ ๋ฐ ๋ฒ•์— ๋Œ€ํ•œ ํ† ๋ก , ์ดˆ๋“ฑํ•™๊ต ๊ณผํ•™ ๊ต๊ณผ์„œ, 9/11 ๋ณด๊ณ ์„œ ๋“ฑ ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์—์„œ ๊ฐ€์ ธ์˜จ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ฐ ์งˆ๋ฌธ์—๋Š” ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์‘๋‹ต ์„ ํƒ์ง€๊ฐ€ ์žˆ์œผ๋ฉฐ, ๊ทธ ์ค‘ ํ•˜๋‚˜ ์ด์ƒ์ด ์ •๋‹ต์ž…๋‹ˆ๋‹ค. ์งˆ๋ฌธ์— ๋‹ตํ•˜๋ ค๋ฉด ์—ฌ๋Ÿฌ ๋ฌธ์žฅ์— ๊ฑธ์ณ reasoningํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. MultiRC ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 800๊ฐœ ์ด์ƒ์˜ ๋‹จ๋ฝ์—์„œ ์ˆ˜์ง‘ํ•œ ์•ฝ 6,000๊ฐœ์˜ multi-sentence ์งˆ๋ฌธ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ฌธ์ œ์—๋Š” ํ‰๊ท ์ ์œผ๋กœ ์ด 5๊ฐœ ์ค‘ ์•ฝ 2๊ฐœ์˜ ์œ ํšจํ•œ ๋‹ต์•ˆ์ด ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
* **BoolQ**[188]๋Š” ์˜ˆ/์•„๋‹ˆ์š” ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ, ๋ชฉํ‘œ๋Š” ๋…ํ•ด ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. BoolQ์—๋Š” \(15,942\)๊ฐœ์˜ ์˜ˆ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์˜ˆ์ œ๋Š” ์งˆ๋ฌธ, ๊ด€๋ จ ๋‹จ๋ฝ, ํ•ด๋‹ต์„ ํฌํ•จํ•˜๋Š” ์‚ผ์ค‘ ํ•ญ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ฃผ์š” ์ง๊ด€์€ ๋…ํ•ด์— ์žˆ์ง€๋งŒ ์ถ”๋ก , ์ž์—ฐ์–ด ์ถ”๋ก  ๋ฐ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ์ž‘์—…์—๋„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * **MultiRC**[189]๋Š” ๋…ํ•ด ๊ณผ์ œ์— ์ ํ•ฉํ•œ ๋˜ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. MultiRC์—๋Š” ๊ฐ„๋‹จํ•œ ๋‹จ๋ฝ๊ณผ ๋‹จ๋ฝ์˜ ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ตํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ค๋ฌธ์žฅ ์งˆ๋ฌธ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋‹จ๋ฝ์€ ๋‰ด์Šค, ์†Œ์„ค, ์—ญ์‚ฌ์  ํ…์ŠคํŠธ, ์œ„ํ‚ค๋ฐฑ๊ณผ ๊ธฐ์‚ฌ, ์‚ฌํšŒ ๋ฐ ๋ฒ•์— ๋Œ€ํ•œ ํ† ๋ก , ์ดˆ๋“ฑํ•™๊ต ๊ณผํ•™ ๊ต๊ณผ์„œ, 9/11 ๋ณด๊ณ ์„œ ๋“ฑ ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์—์„œ ๊ฐ€์ ธ์˜จ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ฐ ์งˆ๋ฌธ์—๋Š” ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์‘๋‹ต ์„ ํƒ์ง€๊ฐ€ ์žˆ์œผ๋ฉฐ, ๊ทธ ์ค‘ ํ•˜๋‚˜ ์ด์ƒ์ด ์ •๋‹ต์ž…๋‹ˆ๋‹ค. ์งˆ๋ฌธ์— ๋‹ตํ•˜๋ ค๋ฉด ์—ฌ๋Ÿฌ ๋ฌธ์žฅ์— ๊ฑธ์ณ ์ถ”๋ก ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. MultiRC ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 800๊ฐœ ์ด์ƒ์˜ ๋‹จ๋ฝ์—์„œ ์ˆ˜์ง‘ํ•œ ์•ฝ 6,000๊ฐœ์˜ ๋‹ค๋ฌธ์žฅ ์งˆ๋ฌธ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ฌธ์ œ์—๋Š” ํ‰๊ท ์ ์œผ๋กœ ์ด 5๊ฐœ ์ค‘ ์•ฝ 2๊ฐœ์˜ ์œ ํšจํ•œ ๋‹ต์•ˆ์ด ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
We plot the results measured by different metrics on MMLU and C-Eval in Figure 4. All three metrics--accuracy, correct choice probability, and Brier Score--show emergent performance improvements (value increase for the first two and decrease for the third) when the pre-training loss drops below a certain threshold. The Brier Score also decreases when the pre-training loss is above the threshold. However, the decrease of Brier Score does not always represent improvements on the task, since the Brier Score is related to not only the predicted probability of the correct answer but also the predicted probabilities of the incorrect answers. We find that the distribution of the correct answers is uniform in the four options in MMLU and C-Eval. The best Brier Score for a context-free predictor is achieved by always giving uniform probability to all the options. In this case, the Brier Score is equal to 0.75. Therefore, the performance in terms of Brier Score is no better than random guess before the loss reaches the threshold. This observation further confirms our previous conclusion. We discuss the contrary observations of [37] and [49] in Appendix C.
๊ทธ๋ฆผ 4์— MMLU์™€ C-Eval์—์„œ ๋‹ค์–‘ํ•œ metrics๋กœ ์ธก์ •ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๊ทธ๋ž˜ํ”„๋กœ ํ‘œ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ •ํ™•๋„, ์ •๋‹ต ์„ ํƒ ํ™•๋ฅ , Brier Score ๋“ฑ ์„ธ ๊ฐ€์ง€ metrics ๋ชจ๋‘ pre-training loss๊ฐ€ ํŠน์ • ์ž„๊ณ„๊ฐ’ ์•„๋ž˜๋กœ ๋–จ์–ด์ง€๋ฉด ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค(์•ž์˜ ๋‘ ๊ฐ€์ง€๋Š” ๊ฐ’์ด ์ฆ๊ฐ€ํ•˜๊ณ  ์„ธ ๋ฒˆ์งธ๋Š” ๊ฐ์†Œ). pre-training loss๊ฐ€ ์ž„๊ณ„๊ฐ’๋ณด๋‹ค ๋†’์„ ๋•Œ๋„ Brier Score๊ฐ€ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Brier Score๋Š” ์ •๋‹ต์˜ ์˜ˆ์ธก ํ™•๋ฅ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์˜ค๋‹ต์˜ ์˜ˆ์ธก ํ™•๋ฅ ๊ณผ๋„ ๊ด€๋ จ์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— Brier Score์˜ ๊ฐ์†Œ๊ฐ€ ํ•ญ์ƒ ๊ณผ์ œ์˜ ๊ฐœ์„ ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. MMLU์™€ C-Eval์—์„œ ์ •๋‹ต์˜ ๋ถ„ํฌ๊ฐ€ ๋„ค ๊ฐ€์ง€ ์˜ต์…˜์—์„œ ๊ท ์ผํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฌธ๋งฅ์ด ์—†๋Š” ์˜ˆ์ธก์ž์˜ ์ตœ์ ์˜ Brier Score๋Š” ํ•ญ์ƒ ๋ชจ๋“  ์˜ต์…˜์— ๊ท ์ผํ•œ ํ™•๋ฅ ์„ ๋ถ€์—ฌํ•จ์œผ๋กœ์จ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ Brier Score๋Š” 0.75์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ loss๊ฐ€ ์ž„๊ณ„๊ฐ’์— ๋„๋‹ฌํ•˜๊ธฐ ์ „์—๋Š” Brier Score ์ธก๋ฉด์—์„œ์˜ ์„ฑ๋Šฅ์ด ๋ฌด์ž‘์œ„ ์ถ”์ธก๋ณด๋‹ค ๋‚˜์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ์€ ์ด์ „ ๊ฒฐ๋ก ์„ ๋”์šฑ ํ™•์ธ์‹œ์ผœ ์ค๋‹ˆ๋‹ค. [37] ๋ฐ [49]์˜ ์ƒ๋ฐ˜๋œ ๊ด€์ฐฐ์— ๋Œ€ํ•ด์„œ๋Š” ๋ถ€๋ก C์—์„œ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 4์— MMLU์™€ C-Eval์—์„œ ๋‹ค์–‘ํ•œ ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ ์ธก์ •ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๊ทธ๋ž˜ํ”„๋กœ ํ‘œ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ •ํ™•๋„, ์ •๋‹ต ์„ ํƒ ํ™•๋ฅ , ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜ ๋“ฑ ์„ธ ๊ฐ€์ง€ ๋ฉ”ํŠธ๋ฆญ ๋ชจ๋‘ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์ด ํŠน์ • ์ž„๊ณ„๊ฐ’ ์•„๋ž˜๋กœ ๋–จ์–ด์ง€๋ฉด ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค(์•ž์˜ ๋‘ ๊ฐ€์ง€๋Š” ๊ฐ’์ด ์ฆ๊ฐ€ํ•˜๊ณ  ์„ธ ๋ฒˆ์งธ๋Š” ๊ฐ์†Œ). ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์ด ์ž„๊ณ„๊ฐ’๋ณด๋‹ค ๋†’์„ ๋•Œ๋„ ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜๊ฐ€ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜๋Š” ์ •๋‹ต์˜ ์˜ˆ์ธก ํ™•๋ฅ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์˜ค๋‹ต์˜ ์˜ˆ์ธก ํ™•๋ฅ ๊ณผ๋„ ๊ด€๋ จ์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜์˜ ๊ฐ์†Œ๊ฐ€ ํ•ญ์ƒ ๊ณผ์ œ์˜ ๊ฐœ์„ ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. MMLU์™€ C-Eval์—์„œ ์ •๋‹ต์˜ ๋ถ„ํฌ๊ฐ€ ๋„ค ๊ฐ€์ง€ ์˜ต์…˜์—์„œ ๊ท ์ผํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฌธ๋งฅ์ด ์—†๋Š” ์˜ˆ์ธก์ž์˜ ์ตœ์ ์˜ ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜๋Š” ํ•ญ์ƒ ๋ชจ๋“  ์˜ต์…˜์— ๊ท ์ผํ•œ ํ™•๋ฅ ์„ ๋ถ€์—ฌํ•จ์œผ๋กœ์จ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜๋Š” 0.75์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์†์‹ค์ด ์ž„๊ณ„๊ฐ’์— ๋„๋‹ฌํ•˜๊ธฐ ์ „์—๋Š” ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜ ์ธก๋ฉด์—์„œ์˜ ์„ฑ๋Šฅ์ด ๋ฌด์ž‘์œ„ ์ถ”์ธก๋ณด๋‹ค ๋‚˜์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ์€ ์ด์ „ ๊ฒฐ๋ก ์„ ๋”์šฑ ํ™•์ธ์‹œ์ผœ ์ค๋‹ˆ๋‹ค. [37] ๋ฐ [49]์˜ ์ƒ๋ฐ˜๋œ ๊ด€์ฐฐ์— ๋Œ€ํ•ด์„œ๋Š” ๋ถ€๋ก C์—์„œ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
**Action Impact**: Action impact refers to the consequences of the action. In fact, the action impact can encompass numerous instances, but for brevity, we only provide a few examples. (1) _Changing Environments._ Agents can directly alter environment states by actions, such as moving their positions, collecting items, constructing buildings, etc. For instance, in GITM [16] and Voyager [38], the environments are changed by the actions of the agents in their task completion process. For example, if the agent mines three woods, then they may disappear in the environments. (2) _Altering Internal States._ Actions taken by the agent can also change the agent itself, including updating memories, forming new plans, acquiring novel knowledge, and more. For example, in Generative Agents [20], memory streams are updated after performing actions within the system. SayCan [78] enables agents to take actions to update understandings of the environment. (3) _Triggering New Actions._ In the task completion process, one agent action can be triggered by another one. For example, Voyager [38] constructs buildings once it has gathered all the necessary resources.
**ํ–‰๋™ ์˜ํ–ฅ**: ํ–‰๋™ ์˜ํ–ฅ์€ ํ–‰๋™์˜ ๊ฒฐ๊ณผ๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ํ–‰๋™์˜ ์˜ํ–ฅ์€ ์—ฌ๋Ÿฌ ๊ฐ€์ง€๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ๊ฐ„๊ฒฐํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด ๋ช‡ ๊ฐ€์ง€ ์˜ˆ๋งŒ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. (1) _ํ™˜๊ฒฝ ๋ณ€ํ™”._ ์—์ด์ „ํŠธ๋Š” ์œ„์น˜ ์ด๋™, ์•„์ดํ…œ ์ˆ˜์ง‘, ๊ฑด๋ฌผ ๊ฑด์„ค ๋“ฑ๊ณผ ๊ฐ™์€ ํ–‰๋™์„ ํ†ตํ•ด ํ™˜๊ฒฝ ์ƒํƒœ๋ฅผ ์ง์ ‘ ๋ณ€๊ฒฝํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GITM [16]๊ณผ Voyager [38]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๋Š” ๊ณผ์ •์—์„œ ํ™˜๊ฒฝ์ด ๋ณ€๊ฒฝ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์—์ด์ „ํŠธ๊ฐ€ ๋‚˜๋ฌด 3๊ฐœ๋ฅผ ์ฑ„๊ตดํ•˜๋ฉด ํ™˜๊ฒฝ์—์„œ ์‚ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (2) _๋‚ด๋ถ€ ์ƒํƒœ ๋ณ€๊ฒฝ._ ์—์ด์ „ํŠธ๊ฐ€ ์ทจํ•œ ํ–‰๋™์€ ๊ธฐ์–ต ์—…๋ฐ์ดํŠธ, ์ƒˆ๋กœ์šด ๊ณ„ํš ์ˆ˜๋ฆฝ, ์ƒˆ๋กœ์šด ์ง€์‹ ์Šต๋“ ๋“ฑ ์—์ด์ „ํŠธ ์ž์ฒด๋ฅผ ๋ณ€๊ฒฝํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Generative Agents [20]์—์„œ๋Š” ์‹œ์Šคํ…œ ๋‚ด์—์„œ ํ–‰๋™์„ ์ˆ˜ํ–‰ํ•œ ํ›„ ๋ฉ”๋ชจ๋ฆฌ ์ŠคํŠธ๋ฆผ์ด ์—…๋ฐ์ดํŠธ๋ฉ๋‹ˆ๋‹ค. SayCan [78]์„ ์‚ฌ์šฉํ•˜๋ฉด ์—์ด์ „ํŠธ๊ฐ€ ํ™˜๊ฒฝ์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ์—…๋ฐ์ดํŠธํ•˜๊ธฐ ์œ„ํ•ด ํ–‰๋™์„ ์ทจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (3) _์ƒˆ๋กœ์šด ํ–‰๋™ ํŠธ๋ฆฌ๊ฑฐ._ ์ž‘์—… ์™„๋ฃŒ ๊ณผ์ •์—์„œ ํ•œ ์—์ด์ „ํŠธ์˜ ํ–‰๋™์ด ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์˜ ํ–‰๋™์— ์˜ํ•ด ํŠธ๋ฆฌ๊ฑฐ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Voyager [38]์€ ํ•„์š”ํ•œ ์ž์›์„ ๋ชจ๋‘ ์ˆ˜์ง‘ํ•œ ํ›„ ๊ฑด๋ฌผ์„ ๊ฑด์„คํ•ฉ๋‹ˆ๋‹ค.
**ํ–‰๋™ ์˜ํ–ฅ**: ํ–‰๋™ ์˜ํ–ฅ์€ ํ–‰๋™์˜ ๊ฒฐ๊ณผ๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ํ–‰๋™์˜ ์˜ํ–ฅ์€ ์—ฌ๋Ÿฌ ๊ฐ€์ง€๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ๊ฐ„๊ฒฐํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด ๋ช‡ ๊ฐ€์ง€ ์˜ˆ๋งŒ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. (1) _ํ™˜๊ฒฝ ๋ณ€ํ™”._ ์—์ด์ „ํŠธ๋Š” ์œ„์น˜ ์ด๋™, ์•„์ดํ…œ ์ˆ˜์ง‘, ๊ฑด๋ฌผ ๊ฑด์„ค ๋“ฑ๊ณผ ๊ฐ™์€ ํ–‰๋™์„ ํ†ตํ•ด ํ™˜๊ฒฝ ์ƒํƒœ๋ฅผ ์ง์ ‘ ๋ณ€๊ฒฝํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GITM [16]๊ณผ Voyager [38]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๋Š” ๊ณผ์ •์—์„œ ํ™˜๊ฒฝ์ด ๋ณ€๊ฒฝ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์—์ด์ „ํŠธ๊ฐ€ ๋‚˜๋ฌด 3๊ฐœ๋ฅผ ์ฑ„๊ตดํ•˜๋ฉด ํ™˜๊ฒฝ์—์„œ ์‚ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (2) _๋‚ด๋ถ€ ์ƒํƒœ ๋ณ€๊ฒฝ._ ์—์ด์ „ํŠธ๊ฐ€ ์ทจํ•œ ํ–‰๋™์€ ๊ธฐ์–ต ์—…๋ฐ์ดํŠธ, ์ƒˆ๋กœ์šด ๊ณ„ํš ์ˆ˜๋ฆฝ, ์ƒˆ๋กœ์šด ์ง€์‹ ์Šต๋“ ๋“ฑ ์—์ด์ „ํŠธ ์ž์ฒด๋ฅผ ๋ณ€๊ฒฝํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ƒ์„ฑ ์—์ด์ „ํŠธ [20]์—์„œ๋Š” ์‹œ์Šคํ…œ ๋‚ด์—์„œ ํ–‰๋™์„ ์ˆ˜ํ–‰ํ•œ ํ›„ ๋ฉ”๋ชจ๋ฆฌ ์ŠคํŠธ๋ฆผ์ด ์—…๋ฐ์ดํŠธ๋ฉ๋‹ˆ๋‹ค. SayCan [78]์„ ์‚ฌ์šฉํ•˜๋ฉด ์—์ด์ „ํŠธ๊ฐ€ ํ™˜๊ฒฝ์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ์—…๋ฐ์ดํŠธํ•˜๊ธฐ ์œ„ํ•ด ํ–‰๋™์„ ์ทจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (3) _์ƒˆ๋กœ์šด ํ–‰๋™ ํŠธ๋ฆฌ๊ฑฐ._ ์ž‘์—… ์™„๋ฃŒ ๊ณผ์ •์—์„œ ํ•œ ์—์ด์ „ํŠธ์˜ ํ–‰๋™์ด ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์˜ ํ–‰๋™์— ์˜ํ•ด ํŠธ๋ฆฌ๊ฑฐ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Voyager [38]์€ ํ•„์š”ํ•œ ์ž์›์„ ๋ชจ๋‘ ์ˆ˜์ง‘ํ•œ ํ›„ ๊ฑด๋ฌผ์„ ๊ฑด์„คํ•ฉ๋‹ˆ๋‹ค.
NotationGiven a source dataset, \(D_{source}\), of documents (crawled web pages) and model architecture, \(M\), we aim to find a strategy \(S\) for selecting a subset of these documents that maximizes some evaluation metric \(E(M(D_{S,R}))\). \(R\) indicates the proportion of remaining documents from the source dataset \(D_{source}\) after selecting data with strategy \(S\). For this reason, we refer to \(R\) throughout this work as the _selection ratio_: for example, if \(R=0.25\) and \(|D_{source}|=100\) million, then we _select_ 25% of documents from a source dataset of size \(100\)M documents to arrive at a a training dataset with \(25\)M documents. We operate at the granularity of a single document, independently of how the model trainer would pack these documents into batches later. Throughout the paper, we use random selection as the baseline for \(S\), as it is the most common method for selecting data for language model pre-training. In the rest of this section, we describe our choices of source dataset (\(D_{source}\)), model (\(M\)), evaluation metric (\(E\)), and, most importantly, our suggestions for the selection strategy (\(S\)).
ํ‘œ๊ธฐ๋ฒ•๋ฌธ์„œ(ํฌ๋กค๋ง๋œ ์›น ํŽ˜์ด์ง€)์˜ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ \(D_{source}\), ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ \(M\)์ด ์ฃผ์–ด์กŒ์„ ๋•Œ, ์ด๋Ÿฌํ•œ ๋ฌธ์„œ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์„ ํƒํ•˜์—ฌ ํ‰๊ฐ€ ๋ฉ”ํŠธ๋ฆญ \(E(M(D_{S,R}))\)์„ ์ตœ๋Œ€ํ™”ํ•˜๋Š” ์ „๋žต \(S\)๋ฅผ ์ฐพ๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. \(R\)์€ ์ „๋žต \(S\)๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•œ ํ›„ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ \(D_{source}\)์—์„œ ๋‚˜๋จธ์ง€ ๋ฌธ์„œ์˜ ๋น„์œจ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ด์œ ๋กœ ์ด ์ž‘์—…์—์„œ๋Š” \(R\)์„ _selection ratio_๋ผ๊ณ  ๋ถ€๋ฆ…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(R=0.25\)์ด๊ณ  \(|D_{source}|=100\)๋ฐฑ๋งŒ์ธ ๊ฒฝ์šฐ, 100M ๋ฌธ์„œ ํฌ๊ธฐ์˜ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์—์„œ ๋ฌธ์„œ์˜ 25%๋ฅผ _์„ ํƒ_ํ•˜์—ฌ 25M ๋ฌธ์„œ๋กœ ๊ตฌ์„ฑ๋œ ํ•™์Šต ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์— ๋„๋‹ฌํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ด๋Ÿฌํ•œ ๋ฌธ์„œ๋ฅผ ๋‚˜์ค‘์— ๋ชจ๋ธ ํŠธ๋ ˆ์ด๋„ˆ๊ฐ€ ๋ฐฐ์น˜๋กœ ๋ฌถ๋Š” ๋ฐฉ์‹์— ๊ด€๊ณ„์—†์ด ๋‹จ์ผ ๋ฌธ์„œ์˜ ์„ธ๋ถ€ ๋‹จ์œ„๋กœ ์ž‘์—…ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•ด ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ๋ฐฉ๋ฒ•์ธ ๋ฌด์ž‘์œ„ ์„ ํƒ์„ \(S\)์˜ baseline์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์˜ ๋‚˜๋จธ์ง€ ๋ถ€๋ถ„์—์„œ๋Š” ์†Œ์Šค ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ(\(D_{source}\)), ๋ชจ๋ธ(\(M\)), ํ‰๊ฐ€ ์ง€ํ‘œ(\(E\)), ๊ทธ๋ฆฌ๊ณ  ๊ฐ€์žฅ ์ค‘์š”ํ•œ ์„ ํƒ ์ „๋žต(\(S\))์— ๋Œ€ํ•œ ์ œ์•ˆ์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
ํ‘œ๊ธฐ๋ฒ•๋ฌธ์„œ(ํฌ๋กค๋ง๋œ ์›น ํŽ˜์ด์ง€)์˜ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ \(D_{source}\), ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ \(M\)์ด ์ฃผ์–ด์กŒ์„ ๋•Œ, ์ด๋Ÿฌํ•œ ๋ฌธ์„œ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์„ ํƒํ•˜์—ฌ ํ‰๊ฐ€ ๋ฉ”ํŠธ๋ฆญ \(E(M(D_{S,R}))\)์„ ์ตœ๋Œ€ํ™”ํ•˜๋Š” ์ „๋žต \(S\)๋ฅผ ์ฐพ๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. R\)์€ ์ „๋žต \(S\)๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•œ ํ›„ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ \(D_{source}\)์—์„œ ๋‚˜๋จธ์ง€ ๋ฌธ์„œ์˜ ๋น„์œจ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ด์œ ๋กœ ์ด ์ž‘์—…์—์„œ๋Š” \(R\)์„ _์„ ํƒ ๋น„์œจ_์ด๋ผ๊ณ  ๋ถ€๋ฆ…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(R=0.25\)์ด๊ณ  \(|D_{source}|=100\)๋ฐฑ๋งŒ์ธ ๊ฒฝ์šฐ, 100\)M ๋ฌธ์„œ ํฌ๊ธฐ์˜ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์—์„œ ๋ฌธ์„œ์˜ 25%๋ฅผ _์„ ํƒ_ํ•˜์—ฌ 25\)M ๋ฌธ์„œ๋กœ ๊ตฌ์„ฑ๋œ ํ•™์Šต ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์— ๋„๋‹ฌํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ด๋Ÿฌํ•œ ๋ฌธ์„œ๋ฅผ ๋‚˜์ค‘์— ๋ชจ๋ธ ํŠธ๋ ˆ์ด๋„ˆ๊ฐ€ ๋ฐฐ์น˜๋กœ ๋ฌถ๋Š” ๋ฐฉ์‹์— ๊ด€๊ณ„์—†์ด ๋‹จ์ผ ๋ฌธ์„œ์˜ ์„ธ๋ถ€ ๋‹จ์œ„๋กœ ์ž‘์—…ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•ด ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ๋ฐฉ๋ฒ•์ธ ๋ฌด์ž‘์œ„ ์„ ํƒ์„ \(S\)์˜ ๊ธฐ์ค€์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์˜ ๋‚˜๋จธ์ง€ ๋ถ€๋ถ„์—์„œ๋Š” ์†Œ์Šค ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ(\(D_{source}\)), ๋ชจ๋ธ(\(M\)), ํ‰๊ฐ€ ์ง€ํ‘œ(\(E\)), ๊ทธ๋ฆฌ๊ณ  ๊ฐ€์žฅ ์ค‘์š”ํ•œ ์„ ํƒ ์ „๋žต(\(S\))์— ๋Œ€ํ•œ ์ œ์•ˆ์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
**Capability Acquisition without Fine-tuning**: In the era of tradition machine learning, the model capability is mainly acquired by learning from datasets, where the knowledge is encoded into the model parameters. In the era of LLMs, the model capability can be acquired either by training/fine-tuning the model parameters or designing delicate prompts (_i.e._, prompt engineer). In prompt engineer, one needs to write valuable information into the prompts to enhance the model capability or unleash existing LLM capabilities. In the era of agents, the model capability can be acquired based on three strategies: (1) model fine-tuning, (2) prompt engineer and (3) designing proper agent evolution mechanisms (we called it as _mechanism engineering_). Mechanism engineering is a broad concept that involves developing specialized modules, introducing novel working rules, and other strategies to enhance agent capabilities. For clearly understanding such transitions on the strategy of model capability acquisition, we illustrate them in Figure 4. In the following, we introduce prompting engineering and mechanism engineering for agent capability acquisition.
**๋ฏธ์„ธ ์กฐ์ • ์—†์ด ๊ธฐ๋Šฅ ํš๋“**: ๊ธฐ์กด ๋จธ์‹  ๋Ÿฌ๋‹ ์‹œ๋Œ€์—๋Š” ์ฃผ๋กœ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ•™์Šตํ•˜์—ฌ ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํš๋“ํ–ˆ์œผ๋ฉฐ, ์ด๋•Œ ์ง€์‹์€ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ์— ์ธ์ฝ”๋”ฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. LLM ์‹œ๋Œ€์—๋Š” ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ›ˆ๋ จ/Fine-tuningํ•˜๊ฑฐ๋‚˜ ์„ฌ์„ธํ•œ ํ”„๋กฌํ”„ํŠธ(_์ฆ‰, prompt engineer_)๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํš๋“ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. prompt engineer์—์„œ๋Š” ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ฑฐ๋‚˜ ๊ธฐ์กด LLM ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ€์น˜ ์žˆ๋Š” ์ •๋ณด๋ฅผ ํ”„๋กฌํ”„ํŠธ์— ์ž‘์„ฑํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ ์‹œ๋Œ€์—๋Š” ์„ธ ๊ฐ€์ง€ ์ „๋žต์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํš๋“ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: (1) ๋ชจ๋ธ Fine-tuning, (2) prompt engineer, (3) ์ ์ ˆํ•œ ์—์ด์ „ํŠธ ์ง„ํ™” ๋ฉ”์ปค๋‹ˆ์ฆ˜ ์„ค๊ณ„(์ด๋ฅผ _mechanism engineering_์ด๋ผ๊ณ  ํ•จ). mechanism engineering์€ ์—์ด์ „ํŠธ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ํŠน์ˆ˜ ๋ชจ๋“ˆ์„ ๊ฐœ๋ฐœํ•˜๊ณ , ์ƒˆ๋กœ์šด ์ž‘์—… ๊ทœ์น™์„ ๋„์ž…ํ•˜๋ฉฐ, ๊ธฐํƒ€ ์ „๋žต์„ ์ˆ˜๋ฆฝํ•˜๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ๊ฐœ๋…์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ ๊ธฐ๋Šฅ ํš๋“ ์ „๋žต์— ๋Œ€ํ•œ ์ด๋Ÿฌํ•œ ์ „ํ™˜์„ ๋ช…ํ™•ํ•˜๊ฒŒ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 4์— ์˜ˆ์‹œ๋ฅผ ๋“ค์–ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์—์ด์ „ํŠธ ๊ธฐ๋Šฅ ํš๋“์„ ์œ„ํ•œ prompt engineer์™€ mechanism engineering์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.
**๋ฏธ์„ธ ์กฐ์ • ์—†์ด ๊ธฐ๋Šฅ ํš๋“**: ๊ธฐ์กด ๋จธ์‹  ๋Ÿฌ๋‹ ์‹œ๋Œ€์—๋Š” ์ฃผ๋กœ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ•™์Šตํ•˜์—ฌ ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํš๋“ํ–ˆ์œผ๋ฉฐ, ์ด๋•Œ ์ง€์‹์€ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ์— ์ธ์ฝ”๋”ฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. LLM ์‹œ๋Œ€์—๋Š” ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ›ˆ๋ จ/๋ฏธ์„ธ ์กฐ์ •ํ•˜๊ฑฐ๋‚˜ ์„ฌ์„ธํ•œ ํ”„๋กฌํ”„ํŠธ(_์ฆ‰, ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด)๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํš๋“ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง์—์„œ๋Š” ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ฑฐ๋‚˜ ๊ธฐ์กด LLM ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ€์น˜ ์žˆ๋Š” ์ •๋ณด๋ฅผ ํ”„๋กฌํ”„ํŠธ์— ์ž‘์„ฑํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ ์‹œ๋Œ€์—๋Š” ์„ธ ๊ฐ€์ง€ ์ „๋žต์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํš๋“ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: (1) ๋ชจ๋ธ ๋ฏธ์„ธ ์กฐ์ •, (2) ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง, (3) ์ ์ ˆํ•œ ์—์ด์ „ํŠธ ์ง„ํ™” ๋ฉ”์ปค๋‹ˆ์ฆ˜ ์„ค๊ณ„(์ด๋ฅผ ๋ฉ”์ปค๋‹ˆ์ฆ˜ ์—”์ง€๋‹ˆ์–ด๋ง์ด๋ผ๊ณ  ํ•จ). ๋ฉ”์ปค๋‹ˆ์ฆ˜ ์—”์ง€๋‹ˆ์–ด๋ง์€ ์—์ด์ „ํŠธ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ํŠน์ˆ˜ ๋ชจ๋“ˆ์„ ๊ฐœ๋ฐœํ•˜๊ณ , ์ƒˆ๋กœ์šด ์ž‘์—… ๊ทœ์น™์„ ๋„์ž…ํ•˜๋ฉฐ, ๊ธฐํƒ€ ์ „๋žต์„ ์ˆ˜๋ฆฝํ•˜๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ๊ฐœ๋…์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ ๊ธฐ๋Šฅ ํš๋“ ์ „๋žต์— ๋Œ€ํ•œ ์ด๋Ÿฌํ•œ ์ „ํ™˜์„ ๋ช…ํ™•ํ•˜๊ฒŒ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 4์— ์˜ˆ์‹œ๋ฅผ ๋“ค์–ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์—์ด์ „ํŠธ ๊ธฐ๋Šฅ ํš๋“์„ ์œ„ํ•œ ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง๊ณผ ๋ฉ”์ปค๋‹ˆ์ฆ˜ ์—”์ง€๋‹ˆ์–ด๋ง์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.
* Adapt-LLM consistently outperforms typical fixed strategies for question answering, such as (i) using the IR system for all questions and (ii) relying solely on the parametric memory of the LLM. * Adapt-LLM demonstrates performance comparable to strategies that rely on popularity scores to determine when to use an IR system, even without utilizing any popularity score or similar metric. It's worth noting that popularity scores are a unique feature of the PopQA dataset, rendering them inapplicable to other open-domain question answering datasets. * When Adapt-LLM decides to retrieve additional information, the results obtained with the context are significantly better than those without it. Similarly, when Adapt-LLM directly answers questions relying on its parametric memory, it achieves high accuracies. These observations indicate that the model effectively discerns when to retrieve information and when it can answer a question without further context. * The primary bottleneck for the performance of Adapt-LLM lies in the IR system. Adapt-LLM achieves much higher performance with gold passages compared to passages retrieved by the IR system.
* Adapt-LLM์€ (i) ๋ชจ๋“  ์งˆ๋ฌธ์— ๋Œ€ํ•ด IR ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๊ณผ (ii) LLM์˜ parametric memory์—๋งŒ ์˜์กดํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™์€ ์ผ๋ฐ˜์ ์ธ ๊ณ ์ •๋œ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์ „๋žต๋ณด๋‹ค ์ง€์†์ ์œผ๋กœ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. * Adapt-LLM์€ ์ธ๊ธฐ๋„ ์ ์ˆ˜๋‚˜ ์œ ์‚ฌํ•œ metric์„ ํ™œ์šฉํ•˜์ง€ ์•Š๋”๋ผ๋„ ์ธ๊ธฐ๋„ ์ ์ˆ˜์— ์˜์กดํ•˜์—ฌ IR ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•  ์‹œ๊ธฐ๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ์ „๋žต๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ธ๊ธฐ๋„ ์ ์ˆ˜๋Š” PopQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ณ ์œ ํ•œ ๊ธฐ๋Šฅ์ด๋ฏ€๋กœ ๋‹ค๋ฅธ ๊ณต๊ฐœ ๋„๋ฉ”์ธ ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์ ์šฉํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์ ์— ์œ ์˜ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. * Adapt-LLM์ด ์ถ”๊ฐ€ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ธฐ๋กœ ๊ฒฐ์ •ํ•˜๋ฉด context๋ฅผ ์‚ฌ์šฉํ•œ ๊ฒฐ๊ณผ๊ฐ€ context๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์€ ๊ฒฐ๊ณผ๋ณด๋‹ค ํ›จ์”ฌ ๋” ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, Adapt-LLM์ด parametric memory์— ์˜์กดํ•˜์—ฌ ์งˆ๋ฌธ์— ์ง์ ‘ ๋‹ตํ•  ๋•Œ์—๋„ ๋†’์€ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์ด ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•  ์‹œ๊ธฐ์™€ ์ถ”๊ฐ€ context ์—†์ด ์งˆ๋ฌธ์— ๋‹ตํ•  ์ˆ˜ ์žˆ๋Š” ์‹œ๊ธฐ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. * Adapt-LLM์˜ ์„ฑ๋Šฅ์˜ ์ฃผ์š” ๋ณ‘๋ชฉ ํ˜„์ƒ์€ IR ์‹œ์Šคํ…œ์— ์žˆ์Šต๋‹ˆ๋‹ค. Adapt-LLM์€ IR ์‹œ์Šคํ…œ์—์„œ ๊ฒ€์ƒ‰ํ•œ ๊ตฌ์ ˆ์— ๋น„ํ•ด gold passages๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ ํ›จ์”ฌ ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
* Adapt-LLM์€ (i) ๋ชจ๋“  ์งˆ๋ฌธ์— ๋Œ€ํ•ด IR ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๊ณผ (ii) LLM์˜ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ์—๋งŒ ์˜์กดํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™์€ ์ผ๋ฐ˜์ ์ธ ๊ณ ์ •๋œ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์ „๋žต๋ณด๋‹ค ์ง€์†์ ์œผ๋กœ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. * Adapt-LLM์€ ์ธ๊ธฐ๋„ ์ ์ˆ˜๋‚˜ ์œ ์‚ฌํ•œ ๋ฉ”ํŠธ๋ฆญ์„ ํ™œ์šฉํ•˜์ง€ ์•Š๋”๋ผ๋„ ์ธ๊ธฐ๋„ ์ ์ˆ˜์— ์˜์กดํ•˜์—ฌ IR ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•  ์‹œ๊ธฐ๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ์ „๋žต๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ธ๊ธฐ๋„ ์ ์ˆ˜๋Š” PopQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ณ ์œ ํ•œ ๊ธฐ๋Šฅ์ด๋ฏ€๋กœ ๋‹ค๋ฅธ ๊ณต๊ฐœ ๋„๋ฉ”์ธ ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์ ์šฉํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์ ์— ์œ ์˜ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. * Adapt-LLM์ด ์ถ”๊ฐ€ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ธฐ๋กœ ๊ฒฐ์ •ํ•˜๋ฉด ์ปจํ…์ŠคํŠธ๋ฅผ ์‚ฌ์šฉํ•œ ๊ฒฐ๊ณผ๊ฐ€ ์ปจํ…์ŠคํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์€ ๊ฒฐ๊ณผ๋ณด๋‹ค ํ›จ์”ฌ ๋” ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, Adapt-LLM์ด ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ์— ์˜์กดํ•˜์—ฌ ์งˆ๋ฌธ์— ์ง์ ‘ ๋‹ตํ•  ๋•Œ์—๋„ ๋†’์€ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์ด ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•  ์‹œ๊ธฐ์™€ ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ ์—†์ด ์งˆ๋ฌธ์— ๋‹ตํ•  ์ˆ˜ ์žˆ๋Š” ์‹œ๊ธฐ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. * Adapt-LLM์˜ ์„ฑ๋Šฅ์˜ ์ฃผ์š” ๋ณ‘๋ชฉ ํ˜„์ƒ์€ IR ์‹œ์Šคํ…œ์— ์žˆ์Šต๋‹ˆ๋‹ค. Adapt-LLM์€ IR ์‹œ์Šคํ…œ์—์„œ ๊ฒ€์ƒ‰ํ•œ ๊ตฌ์ ˆ์— ๋น„ํ•ด ๊ณจ๋“œ ๊ตฌ์ ˆ์„ ์‚ฌ์šฉํ•  ๋•Œ ํ›จ์”ฌ ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
Several empirical studies confirm the deleterious effect of training on self-generated data: In the context of image generation, (Martinez et al., 2023a;b) report degradation of models trained on AI-generated data. Specifically, they use a Denoising Diffusion Implicit Model and a few (relatively small) datasets (e.g. Orchids, MNIST) to demonstrate visual degradation when training in successive generations of AI-generated data. (Hataya et al., 2023) _"conclude that generated images negatively affect downstream performance, while the significance depends on tasks and the amount of generated images"_, (Bohacek & Farid, 2023) reports that the popular StableDiffusion model collapses when iteratively retrained on self-generated faces, even with as little as \(3\%\) synthetic data mixed into the original training set. For text, (Briesch et al., 2023) use _nanoGPT6_ on a curated 10K logical-expression dataset to demonstrate the iterative collapse of self-consuming loops - the model and dataset are sufficiently small to allow training from scratch. (Guo et al., 2023) observe a decline in linguistic diversity metrics across iteratively fine-tuned LLMs.
์—ฌ๋Ÿฌ ์‹ค์ฆ ์—ฐ๊ตฌ์—์„œ ์ž์ฒด ์ƒ์„ฑ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์ด ํ•ด๋กœ์šด ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค๋Š” ์‚ฌ์‹ค์ด ํ™•์ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค: ์ด๋ฏธ์ง€ ์ƒ์„ฑ์˜ ๋งฅ๋ฝ์—์„œ (Martinez et al., 2023a;b)๋Š” AI๋กœ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ์ €ํ•˜๋ฅผ ๋ณด๊ณ ํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์ด๋“ค์€ Denoising Diffusion Implicit Model๊ณผ ๋ช‡ ๊ฐ€์ง€ (๋น„๊ต์  ์ž‘์€) ๋ฐ์ดํ„ฐ ์„ธํŠธ(์˜ˆ: Orchids, MNIST)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ AI๋กœ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ์—ฐ์†์ ์ธ ์„ธ๋Œ€์—์„œ ํ›ˆ๋ จํ•  ๋•Œ ์‹œ๊ฐ์  ์ €ํ•˜๋ฅผ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. (Hataya et al., 2023) _"์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€๊ฐ€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋ฐ˜๋ฉด, ๊ทธ ์ค‘์š”์„ฑ์€ ์ž‘์—…๊ณผ ์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€์˜ ์–‘์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง„๋‹ค"_, (Bohacek & Farid, 2023)๋Š” StableDiffusion ๋ชจ๋ธ์ด ์›๋ž˜ ํ›ˆ๋ จ ์„ธํŠธ์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ 3%๋งŒ ์„ž์—ฌ ์žˆ์–ด๋„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ž์ฒด ์ƒ์„ฑ๋œ ์–ผ๊ตด์— ๋Œ€ํ•ด ์žฌํ›ˆ๋ จํ•  ๋•Œ ๋ถ•๊ดด๋œ๋‹ค๊ณ  ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ํ…์ŠคํŠธ์˜ ๊ฒฝ์šฐ, (Briesch et al., 2023)์€ 10,000๊ฐœ์˜ ๋…ผ๋ฆฌ์  ํ‘œํ˜„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์—„์„ ํ•˜์—ฌ _nanoGPT6_์„ ์‚ฌ์šฉํ•˜์—ฌ ์ž์ฒด ์†Œ๋น„ ๋ฃจํ”„์˜ ๋ฐ˜๋ณต์  ๋ถ•๊ดด๋ฅผ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ๊ณผ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ถฉ๋ถ„ํžˆ ์ž‘์•„์„œ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (Guo et al., 2023)์€ ๋ฐ˜๋ณต์ ์œผ๋กœ Fine-tuning๋œ LLM ์ „๋ฐ˜์—์„œ ์–ธ์–ด์  ๋‹ค์–‘์„ฑ ๋ฉ”ํŠธ๋ฆญ์˜ ํ•˜๋ฝ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.
์—ฌ๋Ÿฌ ์‹ค์ฆ ์—ฐ๊ตฌ์—์„œ ์ž์ฒด ์ƒ์„ฑ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์ด ํ•ด๋กœ์šด ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค๋Š” ์‚ฌ์‹ค์ด ํ™•์ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค: ์ด๋ฏธ์ง€ ์ƒ์„ฑ์˜ ๋งฅ๋ฝ์—์„œ (Martinez et al., 2023a;b)๋Š” AI๋กœ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ์ €ํ•˜๋ฅผ ๋ณด๊ณ ํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์ด๋“ค์€ ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ํ™•์‚ฐ ์•”์‹œ ๋ชจ๋ธ๊ณผ ๋ช‡ ๊ฐ€์ง€ (๋น„๊ต์  ์ž‘์€) ๋ฐ์ดํ„ฐ ์„ธํŠธ(์˜ˆ: Orchids, MNIST)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ AI๋กœ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ์—ฐ์†์ ์ธ ์„ธ๋Œ€์—์„œ ํ›ˆ๋ จํ•  ๋•Œ ์‹œ๊ฐ์  ์ €ํ•˜๋ฅผ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. (ํ•˜ํƒ€์•ผ et al., 2023) _"์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€๊ฐ€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋ฐ˜๋ฉด, ๊ทธ ์ค‘์š”์„ฑ์€ ์ž‘์—…๊ณผ ์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€์˜ ์–‘์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง„๋‹ค"_, (๋ณดํ•˜์ฒต & ํŒŒ๋ฆฌ๋“œ, 2023)๋Š” ์•ˆ์ •์ ์ธ ์•ˆ์ •ํ™” ๋ชจ๋ธ์ด ์›๋ž˜ ํ›ˆ๋ จ ์„ธํŠธ์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ 3%๋งŒ ์„ž์—ฌ ์žˆ์–ด๋„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ž์ฒด ์ƒ์„ฑ๋œ ์–ผ๊ตด์— ๋Œ€ํ•ด ์žฌํ›ˆ๋ จํ•  ๋•Œ ๋ถ•๊ดด๋œ๋‹ค๊ณ  ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ํ…์ŠคํŠธ์˜ ๊ฒฝ์šฐ, (Briesch et al., 2023)์€ 10,000๊ฐœ์˜ ๋…ผ๋ฆฌ์  ํ‘œํ˜„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์—„์„ ํ•˜์—ฌ _๋‚˜๋…ธGPT6_์„ ์‚ฌ์šฉํ•˜์—ฌ ์ž์ฒด ์†Œ๋น„ ๋ฃจํ”„์˜ ๋ฐ˜๋ณต์  ๋ถ•๊ดด๋ฅผ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ๊ณผ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ถฉ๋ถ„ํžˆ ์ž‘์•„์„œ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (Guo et al., 2023)์€ ๋ฐ˜๋ณต์ ์œผ๋กœ ๋ฏธ์„ธ ์กฐ์ •๋œ LLM ์ „๋ฐ˜์—์„œ ์–ธ์–ด์  ๋‹ค์–‘์„ฑ ๋ฉ”ํŠธ๋ฆญ์˜ ํ•˜๋ฝ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.
The most prevalent LMs nowadays are the GPT families (Radford et al., 2019) (Brown et al., 2020) (Achiam et al., 2023) and Bert families (Liu et al., 2019) (Devlin et al., 2018)(Sanh et al., 2019), which have been demonstrated to excel in a multitude of natural language processing tasks. Among these, the AutoEncoder language model is particularly adept at natural language understanding tasks,while the AutoRegressive language model is more suited to natural language generation tasks. While increasing parameters (Touvron et al., 2023b) and model tuning (Han et al., 2023) can enhance the performance of LLMs, the phenomenon of "hallucination" (Ji et al., 2023) persists. Furthermore, the limitations of LMs in effectively handling knowledge-intensive work (Feng et al., 2023) and their inability to promptly update their knowledge (Mousavi et al., 2024) are consistently apparent. Consequently, numerous researchers (Lewis et al., 2020) (Izacard and Grave, 2020b) (Khandelwal et al., 2019) have employed the technique of retrieval to obtain external knowledge, which can assist the language model in attaining enhanced performance in a multitude of tasks.
ํ˜„์žฌ ๊ฐ€์žฅ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” LMs์€ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์ž‘์—…์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋œ GPT ๊ณ„์—ด(Radford et al., 2019)(Brown et al., 2020)(Achiam et al., 2023)๊ณผ Bert ๊ณ„์—ด(Liu et al., 2019)(Devlin et al., 2018)(Sanh et al., 2019)์ž…๋‹ˆ๋‹ค. ์ด ์ค‘ AutoEncoder language model์€ ์ž์—ฐ์–ด ์ดํ•ด ์ž‘์—…์— ํŠนํžˆ ๋Šฅ์ˆ™ํ•˜๋ฉฐ, AutoRegressive language model์€ ์ž์—ฐ์–ด ์ƒ์„ฑ ์ž‘์—…์— ๋” ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. Parameters(Touvron et al., 2023b)๋ฅผ ๋Š˜๋ฆฌ๊ณ  ๋ชจ๋ธ์„ tuning(Han et al., 2023)ํ•˜๋ฉด LLM์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์ง€๋งŒ, "hallucination" ํ˜„์ƒ(Ji et al., 2023)์€ ์ง€์†์ ์œผ๋กœ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ง€์‹ ์ง‘์•ฝ์ ์ธ ์—…๋ฌด๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ์žˆ์–ด LLM์˜ ํ•œ๊ณ„(Feng et al., 2023)์™€ ์ง€์‹์„ ์‹ ์†ํ•˜๊ฒŒ ์—…๋ฐ์ดํŠธํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฌธ์ œ(Mousavi et al., 2024)๋„ ์ง€์†์ ์œผ๋กœ ๋‚˜ํƒ€๋‚˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค(Lewis et al., 2020) (Izacard and Grave, 2020b) (Khandelwal et al., 2019)์€ ์–ธ์–ด ๋ชจ๋ธ์ด ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋Š” ์™ธ๋ถ€ ์ง€์‹์„ retrievalํ•˜๋Š” ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
ํ˜„์žฌ ๊ฐ€์žฅ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” LM์€ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์ž‘์—…์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋œ GPT ๊ณ„์—ด(Radford et al., 2019)(Brown et al., 2020)(Achiam et al., 2023)๊ณผ Bert ๊ณ„์—ด(Liu et al., 2019)(Devlin et al., 2018)(Sanh et al., 2019)์ž…๋‹ˆ๋‹ค. ์ด ์ค‘ ์ž๋™ ์ธ์ฝ”๋” ์–ธ์–ด ๋ชจ๋ธ์€ ์ž์—ฐ์–ด ์ดํ•ด ์ž‘์—…์— ํŠนํžˆ ๋Šฅ์ˆ™ํ•˜๋ฉฐ, ์ž๋™ ํšŒ๊ท€ ์–ธ์–ด ๋ชจ๋ธ์€ ์ž์—ฐ์–ด ์ƒ์„ฑ ์ž‘์—…์— ๋” ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ํŒŒ๋ผ๋ฏธํ„ฐ(Touvron et al., 2023b)๋ฅผ ๋Š˜๋ฆฌ๊ณ  ๋ชจ๋ธ์„ ํŠœ๋‹(Han et al., 2023)ํ•˜๋ฉด LLM์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์ง€๋งŒ, 'ํ™˜๊ฐ' ํ˜„์ƒ(Ji et al., 2023)์€ ์ง€์†์ ์œผ๋กœ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ง€์‹ ์ง‘์•ฝ์ ์ธ ์—…๋ฌด๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ์žˆ์–ด LLM์˜ ํ•œ๊ณ„(Feng et al., 2023)์™€ ์ง€์‹์„ ์‹ ์†ํ•˜๊ฒŒ ์—…๋ฐ์ดํŠธํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฌธ์ œ(Mousavi et al., 2024)๋„ ์ง€์†์ ์œผ๋กœ ๋‚˜ํƒ€๋‚˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค(Lewis et al., 2020) (Izacard and Grave, 2020b) (Khandelwal et al., 2019)์€ ์–ธ์–ด ๋ชจ๋ธ์ด ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋Š” ์™ธ๋ถ€ ์ง€์‹์„ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
* We propose a novel CL formulation called Continual Knowledge Learning (CKL) and construct a new benchmark to measure the amount of forgetting and amount of world knowledge gained by continued pretraining on a novel language modeling corpus that we construct, containing new knowledge. * We explore LM architectures and training methodologies that are natural baselines for CKL in literature, denoting them as CKL methods, and performing extensive experiments on our CKL benchmark. We categorize them into regularization, rehearsal, and parameter-expansion methods, same as in traditional CL literature, and compare the effectiveness of each type of method using a novel metric named FUAR that we propose to measure the trade-off between forgotten knowledge and updated or acquired knowledge. * Towards creating an ever-changing LM, we perform extensive analysis in the CKL benchmark and highlight important challenges and findings: parameter-expansion methods have the limitation of memory inefficiency despite performing the best in most of our experiments and seeing the same data repeatedly during continued pretraining is a critical cause of forgetting.
* ์ƒˆ๋กœ์šด ์ง€์‹์„ ํฌํ•จํ•˜๋Š” ์ƒˆ๋กœ์šด ์–ธ์–ด ๋ชจ๋ธ๋ง ์ฝ”ํผ์Šค์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ์žŠ์–ด๋ฒ„๋ฆฐ ์ง€์‹์˜ ์–‘๊ณผ ์–ป์€ ์„ธ๊ณ„ ์ง€์‹์˜ ์–‘์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด CL ๊ณต์‹์ธ ์ง€์†์  ์ง€์‹ ํ•™์Šต(CKL)์„ ์ œ์•ˆํ•˜๊ณ  ์ƒˆ๋กœ์šด benchmark๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. * ๋ฌธํ—Œ์—์„œ CKL์˜ ์ž์—ฐ์Šค๋Ÿฌ์šด ๊ธฐ์ค€์ด ๋˜๋Š” LM ์•„ํ‚คํ…์ฒ˜์™€ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก ์„ ํƒ์ƒ‰ํ•˜์—ฌ CKL ๋ฐฉ๋ฒ•์œผ๋กœ ํ‘œ์‹œํ•˜๊ณ , CKL benchmark์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด CL ๋ฌธํ—Œ์—์„œ์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ regularization, rehearsal, parameter-expansion ๋ฐฉ๋ฒ•์œผ๋กœ ๋ถ„๋ฅ˜ํ•˜๊ณ , ์žŠ์–ด๋ฒ„๋ฆฐ ์ง€์‹๊ณผ ์—…๋ฐ์ดํŠธ ๋˜๋Š” ํš๋“ํ•œ ์ง€์‹ ๊ฐ„์˜ ํŠธ๋ ˆ์ด๋“œ์˜คํ”„๋ฅผ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์ œ์•ˆํ•œ ์ƒˆ๋กœ์šด metric์ธ FUAR๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์œ ํ˜•์˜ ๋ฐฉ๋ฒ•์˜ ํšจ๊ณผ๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. * ๋Š์ž„์—†์ด ๋ณ€ํ™”ํ•˜๋Š” LM์„ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด CKL benchmark์—์„œ ๊ด‘๋ฒ”์œ„ํ•œ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ์ค‘์š”ํ•œ ๊ณผ์ œ์™€ ๋ฐœ๊ฒฌ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. parameter-expansion ๋ฐฉ๋ฒ•์€ ๋Œ€๋ถ€๋ถ„์˜ ์‹คํ—˜์—์„œ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์ง€๋งŒ ๋ฉ”๋ชจ๋ฆฌ ๋น„ํšจ์œจ์„ฑ์ด๋ผ๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋™์ผํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๋ณด๋Š” ๊ฒƒ์€ ์žŠ์–ด๋ฒ„๋ฆผ์˜ ์ค‘์š”ํ•œ ์›์ธ์ž…๋‹ˆ๋‹ค.
* ์ƒˆ๋กœ์šด ์ง€์‹์„ ํฌํ•จํ•˜๋Š” ์ƒˆ๋กœ์šด ์–ธ์–ด ๋ชจ๋ธ๋ง ์ฝ”ํผ์Šค์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ์žŠ์–ด๋ฒ„๋ฆฐ ์ง€์‹์˜ ์–‘๊ณผ ์–ป์€ ์„ธ๊ณ„ ์ง€์‹์˜ ์–‘์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด CL ๊ณต์‹์ธ ์ง€์†์  ์ง€์‹ ํ•™์Šต(CKL)์„ ์ œ์•ˆํ•˜๊ณ  ์ƒˆ๋กœ์šด ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. * ๋ฌธํ—Œ์—์„œ CKL์˜ ์ž์—ฐ์Šค๋Ÿฌ์šด ๊ธฐ์ค€์ด ๋˜๋Š” LM ์•„ํ‚คํ…์ฒ˜์™€ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก ์„ ํƒ์ƒ‰ํ•˜์—ฌ CKL ๋ฐฉ๋ฒ•์œผ๋กœ ํ‘œ์‹œํ•˜๊ณ , CKL ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด CL ๋ฌธํ—Œ์—์„œ์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ •๊ทœํ™”, ๋ฆฌํ—ˆ์„ค, ํŒŒ๋ผ๋ฏธํ„ฐ ํ™•์žฅ ๋ฐฉ๋ฒ•์œผ๋กœ ๋ถ„๋ฅ˜ํ•˜๊ณ , ์žŠ์–ด๋ฒ„๋ฆฐ ์ง€์‹๊ณผ ์—…๋ฐ์ดํŠธ ๋˜๋Š” ํš๋“ํ•œ ์ง€์‹ ๊ฐ„์˜ ํŠธ๋ ˆ์ด๋“œ์˜คํ”„๋ฅผ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์ œ์•ˆํ•œ ์ƒˆ๋กœ์šด ๋ฉ”ํŠธ๋ฆญ์ธ FUAR๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์œ ํ˜•์˜ ๋ฐฉ๋ฒ•์˜ ํšจ๊ณผ๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. * ๋Š์ž„์—†์ด ๋ณ€ํ™”ํ•˜๋Š” LM์„ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด CKL ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ด‘๋ฒ”์œ„ํ•œ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ์ค‘์š”ํ•œ ๊ณผ์ œ์™€ ๋ฐœ๊ฒฌ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ํŒŒ๋ผ๋ฏธํ„ฐ ํ™•์žฅ ๋ฐฉ๋ฒ•์€ ๋Œ€๋ถ€๋ถ„์˜ ์‹คํ—˜์—์„œ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์ง€๋งŒ ๋ฉ”๋ชจ๋ฆฌ ๋น„ํšจ์œจ์„ฑ์ด๋ผ๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋™์ผํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๋ณด๋Š” ๊ฒƒ์€ ์žŠ์–ด๋ฒ„๋ฆผ์˜ ์ค‘์š”ํ•œ ์›์ธ์ž…๋‹ˆ๋‹ค.
General UnderstandingThe General Understanding category comprises datasets testing broader cognitive skills and language comprehension. **ARC Easy (ARC-E)**Clark et al. (2018) is the less challenging counterpart of ARC-C, featuring questions that require basic reasoning skills. **BoolQ**Clark et al. (2019) includes boolean questions that focus on reading comprehension and general language understanding. **Winogrande (Wino.)**ai (2019) challenges models with common sense reasoning in language, particularly in pronoun disambiguation. **PIQA**Bisk et al. (2020) assesses understanding of physical processes, an essential part of practical common sense. **HellaSwag**Zellers et al. (2019) tests the ability to complete scenarios coherently, demanding both language understanding and common sense. **TruthfulQA**Lin et al. (2021) is centered on generating truthful, accurate answers, thus testing the model's factual correctness. **OpenBookQA (OBQA)**Mihaylov et al. (2018) evaluates the understanding of a broad range of facts and concepts. Finally, **LogiQA-2**Liu et al. (2023) assesses the model's capacity to comprehend and apply logical principles.
์ผ๋ฐ˜์ ์ธ ์ดํ•ด ์ผ๋ฐ˜ ์ดํ•ด ์นดํ…Œ๊ณ ๋ฆฌ์—๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ์ธ์ง€ ๋Šฅ๋ ฅ๊ณผ ์–ธ์–ด ์ดํ•ด๋ ฅ์„ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. **ARC Easy (ARC-E)**Clark ๋“ฑ(2018)์€ ๊ธฐ๋ณธ์ ์ธ ์ถ”๋ก  ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•œ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋œ ARC-C์˜ ๋œ ์–ด๋ ค์šด ๋Œ€์‘๋ฌผ์ž…๋‹ˆ๋‹ค. **BoolQ**Clark ๋“ฑ(2019)์€ ๋…ํ•ด๋ ฅ๊ณผ ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ์ดํ•ด๋ ฅ์— ์ค‘์ ์„ ๋‘” ๋ถ€์šธ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. **Winogrande (Wino.)**ai (2019)๋Š” ํŠนํžˆ ๋Œ€๋ช…์‚ฌ์˜ ์˜ค๋ฅ˜๋ฅผ ์‹๋ณ„ํ•˜๋Š” ์–ธ์–ด์˜ ์ƒ์‹์ ์ธ ์ถ”๋ก ์œผ๋กœ ๋ชจ๋ธ์— ๋„์ „ํ•ฉ๋‹ˆ๋‹ค. **PIQA**Bisk ๋“ฑ(2020)์€ ์‹ค์šฉ์ ์ธ ์ƒ์‹์˜ ํ•„์ˆ˜์ ์ธ ๋ถ€๋ถ„์ธ ๋ฌผ๋ฆฌ์  ๊ณผ์ •์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. **HellaSwag**Zellers ๋“ฑ(2019)์€ ์–ธ์–ด ์ดํ•ด์™€ ์ƒ์‹์„ ๋ชจ๋‘ ์š”๊ตฌํ•˜๋Š” ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์ผ๊ด€์„ฑ ์žˆ๊ฒŒ ์™„์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. **TruthfulQA**Lin ๋“ฑ(2021)์€ ์ง„์‹คํ•˜๊ณ  ์ •ํ™•ํ•œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์–ด ๋ชจ๋ธ์˜ ์‚ฌ์‹ค ์ •ํ™•์„ฑ์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. **OpenBookQA (OBQA)**Mihaylov ๋“ฑ(2018)์€ ๊ด‘๋ฒ”์œ„ํ•œ ์‚ฌ์‹ค๊ณผ ๊ฐœ๋…์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, **LogiQA-2** Liu ๋“ฑ(2023)์€ ๋…ผ๋ฆฌ์  ์›์น™์„ ์ดํ•ดํ•˜๊ณ  ์ ์šฉํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
์ผ๋ฐ˜์ ์ธ ์ดํ•ด์ผ๋ฐ˜ ์ดํ•ด ์นดํ…Œ๊ณ ๋ฆฌ์—๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ์ธ์ง€ ๋Šฅ๋ ฅ๊ณผ ์–ธ์–ด ์ดํ•ด๋ ฅ์„ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. **ARC Easy (ARC-E)**Clark ๋“ฑ(2018)์€ ๊ธฐ๋ณธ์ ์ธ ์ถ”๋ก  ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•œ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋œ ARC-C์˜ ๋œ ์–ด๋ ค์šด ๋Œ€์‘๋ฌผ์ž…๋‹ˆ๋‹ค. **BoolQ**Clark ๋“ฑ(2019)์€ ๋…ํ•ด๋ ฅ๊ณผ ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ์ดํ•ด๋ ฅ์— ์ค‘์ ์„ ๋‘” ๋ถ€์šธ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. **์œ„๋…ธ๊ทธ๋ž€๋ฐ(Wino.)**ai (2019)๋Š” ํŠนํžˆ ๋Œ€๋ช…์‚ฌ์˜ ์˜ค๋ฅ˜๋ฅผ ์‹๋ณ„ํ•˜๋Š” ์–ธ์–ด์˜ ์ƒ์‹์ ์ธ ์ถ”๋ก ์œผ๋กœ ๋ชจ๋ธ์— ๋„์ „ํ•ฉ๋‹ˆ๋‹ค. **PIQA**Bisk ๋“ฑ(2020)์€ ์‹ค์šฉ์ ์ธ ์ƒ์‹์˜ ํ•„์ˆ˜์ ์ธ ๋ถ€๋ถ„์ธ ๋ฌผ๋ฆฌ์  ๊ณผ์ •์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. **ํ—ฌ๋ผ์Šค์›จ๊ทธ**์…€๋Ÿฌ์Šค ๋“ฑ(2019)์€ ์–ธ์–ด ์ดํ•ด์™€ ์ƒ์‹์„ ๋ชจ๋‘ ์š”๊ตฌํ•˜๋Š” ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์ผ๊ด€์„ฑ ์žˆ๊ฒŒ ์™„์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. **์ง„์‹คํ•œQA**Lin ๋“ฑ(2021)์€ ์ง„์‹คํ•˜๊ณ  ์ •ํ™•ํ•œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์–ด ๋ชจ๋ธ์˜ ์‚ฌ์‹ค ์ •ํ™•์„ฑ์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. **์˜คํ”ˆ๋ถQA(OBQA)**Mihaylov ๋“ฑ(2018)์€ ๊ด‘๋ฒ”์œ„ํ•œ ์‚ฌ์‹ค๊ณผ ๊ฐœ๋…์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, **LogiQA-2** Liu ๋“ฑ(2023)์€ ๋…ผ๋ฆฌ์  ์›์น™์„ ์ดํ•ดํ•˜๊ณ  ์ ์šฉํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
Future Directions.Besides the above approaches, there are several promising directions for KG-enhanced LLM remaining underexplored. First, due to the variety of structured data, it is still difficult for LLMs to directly leverage various kinds of knowledge sources, _e.g._, domain-specific KGs. Therefore, it is essential to explore the unified way to manipulate and utilize different knowledge sources by LLMs. As a potential solution, it is promising to develop effective approaches to help LLMs comprehend and make use of the access interfaces provided by specific knowledge sources to acquire precise knowledge [458], while more efforts should be made to investigate how to adapt to the data variety in a cost-effective way. Second, with the evolution of real-world information, the knowledge stored in LLMs may become outdated or incorrect. It is necessary to explore how to synchronize the updated knowledge into LLMs through a cost-effective manner [873, 874]. Third, it is promising to investigate the use of factual information from KG to align LLMs in generating more faithful content [875, 876], which can help reduce the hallucination of LLMs.
๋ฏธ๋ž˜ ๋ฐฉํ–ฅ. ์œ„์˜ ์ ‘๊ทผ ๋ฐฉ์‹ ์™ธ์—๋„ KG๋กœ ๊ฐ•ํ™”๋œ LLM์—๋Š” ์•„์ง ํƒ๊ตฌ๋˜์ง€ ์•Š์€ ๋ช‡ ๊ฐ€์ง€ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ์˜ ๋‹ค์–‘์„ฑ์œผ๋กœ ์ธํ•ด LLM์ด ๋‹ค์–‘ํ•œ ์ข…๋ฅ˜์˜ ์ง€์‹ ์†Œ์Šค, _์˜ˆ๋ฅผ ๋“ค์–ด, ๋„๋ฉ”์ธ๋ณ„ KG๋ฅผ ์ง์ ‘ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ LLM์ด ๋‹ค์–‘ํ•œ ์ง€์‹ ์†Œ์Šค๋ฅผ ์กฐ์ž‘ํ•˜๊ณ  ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ํ†ตํ•ฉ๋œ ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•˜๋Š” ๊ฒƒ์€ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค. ์ž ์žฌ์ ์ธ ํ•ด๊ฒฐ์ฑ…์œผ๋กœ, LLM์ด ํŠน์ • ์ง€์‹ ์†Œ์Šค์—์„œ ์ œ๊ณตํ•˜๋Š” ์•ก์„ธ์Šค ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ดํ•ดํ•˜๊ณ  ํ™œ์šฉํ•˜์—ฌ ์ •ํ™•ํ•œ ์ง€์‹์„ ์Šต๋“ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ํšจ๊ณผ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์œ ๋งํ•˜๋ฉฐ, ๋น„์šฉ ํšจ์œจ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ๋ฐ์ดํ„ฐ ๋‹ค์–‘์„ฑ์— ์ ์‘ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ๋” ๋งŽ์€ ๋…ธ๋ ฅ์„ ๊ธฐ์šธ์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์‹ค์ œ ์ •๋ณด์˜ ์ง„ํ™”์— ๋”ฐ๋ผ LLM์— ์ €์žฅ๋œ ์ง€์‹์ด ์˜ค๋ž˜๋˜์—ˆ๊ฑฐ๋‚˜ ๋ถ€์ •ํ™•ํ•ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋น„์šฉ ํšจ์œจ์ ์ธ ๋ฐฉ์‹์„ ํ†ตํ•ด ์—…๋ฐ์ดํŠธ๋œ ์ง€์‹์„ LLM์— ๋™๊ธฐํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค[873, 874]. ์…‹์งธ, KG์˜ ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์ด ๋” ์‹ ์‹คํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ๋•Œ alignํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ๋„ ์œ ๋งํ•˜๋ฉฐ[875, 876], ์ด๋Š” LLM์˜ hallucination์„ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ฏธ๋ž˜ ๋ฐฉํ–ฅ.์œ„์˜ ์ ‘๊ทผ ๋ฐฉ์‹ ์™ธ์—๋„ KG๋กœ ๊ฐ•ํ™”๋œ LLM์—๋Š” ์•„์ง ํƒ๊ตฌ๋˜์ง€ ์•Š์€ ๋ช‡ ๊ฐ€์ง€ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ์˜ ๋‹ค์–‘์„ฑ์œผ๋กœ ์ธํ•ด LLM์ด ๋‹ค์–‘ํ•œ ์ข…๋ฅ˜์˜ ์ง€์‹ ์†Œ์Šค, _์˜ˆ๋ฅผ ๋“ค์–ด, ๋„๋ฉ”์ธ๋ณ„ KG๋ฅผ ์ง์ ‘ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ LLM์ด ๋‹ค์–‘ํ•œ ์ง€์‹ ์†Œ์Šค๋ฅผ ์กฐ์ž‘ํ•˜๊ณ  ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ํ†ตํ•ฉ๋œ ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•˜๋Š” ๊ฒƒ์€ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค. ์ž ์žฌ์ ์ธ ํ•ด๊ฒฐ์ฑ…์œผ๋กœ, LLM์ด ํŠน์ • ์ง€์‹ ์†Œ์Šค์—์„œ ์ œ๊ณตํ•˜๋Š” ์•ก์„ธ์Šค ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ดํ•ดํ•˜๊ณ  ํ™œ์šฉํ•˜์—ฌ ์ •ํ™•ํ•œ ์ง€์‹์„ ์Šต๋“ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ํšจ๊ณผ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์œ ๋งํ•˜๋ฉฐ, ๋น„์šฉ ํšจ์œจ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ๋ฐ์ดํ„ฐ ๋‹ค์–‘์„ฑ์— ์ ์‘ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ๋” ๋งŽ์€ ๋…ธ๋ ฅ์„ ๊ธฐ์šธ์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์‹ค์ œ ์ •๋ณด์˜ ์ง„ํ™”์— ๋”ฐ๋ผ LLM์— ์ €์žฅ๋œ ์ง€์‹์ด ์˜ค๋ž˜๋˜์—ˆ๊ฑฐ๋‚˜ ๋ถ€์ •ํ™•ํ•ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋น„์šฉ ํšจ์œจ์ ์ธ ๋ฐฉ์‹์„ ํ†ตํ•ด ์—…๋ฐ์ดํŠธ๋œ ์ง€์‹์„ LLM์— ๋™๊ธฐํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค[873, 874]. ์…‹์งธ, KG์˜ ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์ด ๋” ์‹ ์‹คํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ๋•Œ ์ •ํ•ฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ๋„ ์œ ๋งํ•˜๋ฉฐ[875, 876], ์ด๋Š” LLM์˜ ํ™˜๊ฐ์„ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
We perform a simple ablation on our 12-thought-token-4-ahead baseline, namely asking whether sampling multiple thoughts per sequence is necessary. We find that although simply computing the reward as the difference between the losses with and without thought proves to be a strong baseline, using multiple thoughts consistently outperformed it (by roughly 0.5% on GSM8K generalization and 3% on CommonsenseQA generalization). However, the exact number of thoughts had little impact: varying between 2, 3, and 4 thoughts per sequence appeared to result in a consistent improvement with additionalthoughts, but additional thoughts per sequence beyond two improved performance by only 0.1-0.3% on both GSM8K and CommonsenseQA. Similarly, we found that predicting more than one token ahead helped by 0.3% on GSM8K and 3.1% on CommonsenseQA, but with 12-thought-tokens, we did not find that additional tokens ahead, beyond two, improved performance. However, qualitatively, we found that the rationales appeared more coherent with additional tokens-ahead of supervision. This suggests a need for more specific evaluations of generation quality in future work.
12-thought-token-4-ahead baseline์— ๋Œ€ํ•ด ๊ฐ„๋‹จํ•œ ablation, ์ฆ‰ ์‹œํ€€์Šค๋‹น ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์ƒ๊ฐ์„ ์ƒ˜ํ”Œ๋งํ•ด์•ผ ํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๋ฌป๋Š” ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹จ์ˆœํžˆ ์ƒ๊ฐ์„ ํ–ˆ์„ ๋•Œ์™€ ํ•˜์ง€ ์•Š์•˜์„ ๋•Œ์˜ Loss ์ฐจ์ด๋กœ ๋ณด์ƒ์„ ๊ณ„์‚ฐํ•˜๋Š” ๊ฒƒ์ด ๊ฐ•๋ ฅํ•œ baseline์ž„์„ ์ฆ๋ช…ํ–ˆ์ง€๋งŒ, ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ƒ๊ฐ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์ผ๊ด€๋˜๊ฒŒ ์ด๋ฅผ ๋Šฅ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค(GSM8K generalization์—์„œ๋Š” ์•ฝ 0.5%, CommonsenseQA generalization์—์„œ๋Š” 3% ๋” ๋†’์•˜์Šต๋‹ˆ๋‹ค). ๊ทธ๋Ÿฌ๋‚˜ ์ •ํ™•ํ•œ ์ƒ๊ฐ์˜ ์ˆ˜๋Š” ๊ฑฐ์˜ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์‹œํ€€์Šค๋‹น 2, 3, 4๊ฐœ์˜ ์ƒ๊ฐ ์‚ฌ์ด๋ฅผ ๋ณ€ํ™”์‹œํ‚ค๋ฉด ์ถ”๊ฐ€ ์ƒ๊ฐ์— ๋”ฐ๋ผ ์ผ๊ด€๋œ ๊ฐœ์„  ํšจ๊ณผ๊ฐ€ ๋‚˜ํƒ€๋‚ฌ์ง€๋งŒ, ์‹œํ€€์Šค๋‹น 2๊ฐœ ์ด์ƒ์˜ ์ƒ๊ฐ์€ GSM8K์™€ CommonsenseQA ๋ชจ๋‘์—์„œ ์„ฑ๋Šฅ์ด 0.1~0.3%๋งŒ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, 1๊ฐœ ์ด์ƒ์˜ token์„ ์•ž์œผ๋กœ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์ด GSM8K์—์„œ๋Š” 0.3%, CommonsenseQA์—์„œ๋Š” 3.1%์˜ ๋„์›€์ด ๋˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์ง€๋งŒ, 12-thought-tokens์„ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ 2๊ฐœ ์ด์ƒ์˜ token์„ ์•ž์œผ๋กœ ์˜ˆ์ธกํ•ด๋„ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜์ง€ ์•Š๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์งˆ์ ์œผ๋กœ๋Š” ์ถ”๊ฐ€ token์„ ์•ž๋‘๊ณ  ์žˆ์„ ๋•Œ ๊ทผ๊ฑฐ๊ฐ€ ๋” ์ผ๊ด€์„ฑ ์žˆ๊ฒŒ ๋ณด์ด๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํ–ฅํ›„ ์ž‘์—…์—์„œ generation ํ’ˆ์งˆ์— ๋Œ€ํ•œ ๋ณด๋‹ค ๊ตฌ์ฒด์ ์ธ ํ‰๊ฐ€๊ฐ€ ํ•„์š”ํ•จ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
12๊ฐœ์˜ ์ƒ๊ฐ ํ† ํฐ-4-์•ž ๊ธฐ์ค€์„ ์— ๋Œ€ํ•ด ๊ฐ„๋‹จํ•œ ์ ˆ์ œ, ์ฆ‰ ์‹œํ€€์Šค๋‹น ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์ƒ๊ฐ์„ ์ƒ˜ํ”Œ๋งํ•ด์•ผ ํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๋ฌป๋Š” ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹จ์ˆœํžˆ ์ƒ๊ฐ์„ ํ–ˆ์„ ๋•Œ์™€ ํ•˜์ง€ ์•Š์•˜์„ ๋•Œ์˜ ์†์‹ค ์ฐจ์ด๋กœ ๋ณด์ƒ์„ ๊ณ„์‚ฐํ•˜๋Š” ๊ฒƒ์ด ๊ฐ•๋ ฅํ•œ ๊ธฐ์ค€์„ ์ž„์„ ์ฆ๋ช…ํ–ˆ์ง€๋งŒ, ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ƒ๊ฐ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์ผ๊ด€๋˜๊ฒŒ ์ด๋ฅผ ๋Šฅ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค(GSM8K ์ผ๋ฐ˜ํ™”์—์„œ๋Š” ์•ฝ 0.5%, ์ƒ์‹QA ์ผ๋ฐ˜ํ™”์—์„œ๋Š” 3% ๋” ๋†’์•˜์Šต๋‹ˆ๋‹ค). ๊ทธ๋Ÿฌ๋‚˜ ์ •ํ™•ํ•œ ์ƒ๊ฐ์˜ ์ˆ˜๋Š” ๊ฑฐ์˜ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์‹œํ€€์Šค๋‹น 2, 3, 4๊ฐœ์˜ ์ƒ๊ฐ ์‚ฌ์ด๋ฅผ ๋ณ€ํ™”์‹œํ‚ค๋ฉด ์ถ”๊ฐ€ ์ƒ๊ฐ์— ๋”ฐ๋ผ ์ผ๊ด€๋œ ๊ฐœ์„  ํšจ๊ณผ๊ฐ€ ๋‚˜ํƒ€๋‚ฌ์ง€๋งŒ, ์‹œํ€€์Šค๋‹น 2๊ฐœ ์ด์ƒ์˜ ์ƒ๊ฐ์€ GSM8K์™€ CommonsenseQA ๋ชจ๋‘์—์„œ ์„ฑ๋Šฅ์ด 0.1~0.3%๋งŒ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, 1๊ฐœ ์ด์ƒ์˜ ํ† ํฐ์„ ์•ž์œผ๋กœ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์ด GSM8K์—์„œ๋Š” 0.3%, CommonsenseQA์—์„œ๋Š” 3.1%์˜ ๋„์›€์ด ๋˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์ง€๋งŒ, 12๊ฐœ์˜ ์ƒ๊ฐ ํ† ํฐ์„ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ 2๊ฐœ ์ด์ƒ์˜ ํ† ํฐ์„ ์•ž์œผ๋กœ ์˜ˆ์ธกํ•ด๋„ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜์ง€ ์•Š๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์งˆ์ ์œผ๋กœ๋Š” ์ถ”๊ฐ€ ํ† ํฐ์„ ์•ž๋‘๊ณ  ์žˆ์„ ๋•Œ ๊ทผ๊ฑฐ๊ฐ€ ๋” ์ผ๊ด€์„ฑ ์žˆ๊ฒŒ ๋ณด์ด๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํ–ฅํ›„ ์ž‘์—…์—์„œ ์ƒ์„ฑ ํ’ˆ์งˆ์— ๋Œ€ํ•œ ๋ณด๋‹ค ๊ตฌ์ฒด์ ์ธ ํ‰๊ฐ€๊ฐ€ ํ•„์š”ํ•จ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
In Figure 2 (left side plots), we observe a consistent negative relationship between the token probability of the model's prior answer and the associated RAG preference rate for all six QA datasets. To visualize an even distribution across probabilities, we bin the probabilities into ten equidistant bins in the range of \([0.0,1.0]\). We additionally present the slope from performing a linear regression on the binned probability values against the RAG preference rate in Table 1. The slope indicates the effect of stronger model confidence on the model's preference for the information presented in the retrieved context; we observe different slopes (ranging from -0.1 to -0.45), suggesting that the effectiveness of RAG in different QA domains can be characterized as being relatively susceptible (e.g., with Dates questions) or robust (e.g., with News questions) to the model's internal prior knowledge confidence. Specifically, a slope of -0.45, for instance, can be interpreted as expecting a 4.5% decrease in the likelihood of the LLM preferring the contextual information for every 10% increase in the probability of the model's prior response.
๊ทธ๋ฆผ 2(์™ผ์ชฝ ํ”Œ๋กฏ)์—์„œ 6๊ฐœ์˜ QA ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ชจ๋‘์—์„œ ๋ชจ๋ธ์˜ ์‚ฌ์ „ ๋‹ต๋ณ€ ํ† ํฐ ํ™•๋ฅ ๊ณผ ๊ด€๋ จ๋œ RAG ์„ ํ˜ธ๋„ ์‚ฌ์ด์— ์ผ๊ด€๋œ ์Œ์˜ ๊ด€๊ณ„๋ฅผ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ™•๋ฅ ์— ๊ฑธ์ณ ๊ท ๋“ฑํ•œ ๋ถ„ํฌ๋ฅผ ์‹œ๊ฐํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ™•๋ฅ ์„ \([0.0,1.0]\) ๋ฒ”์œ„์—์„œ 10๊ฐœ์˜ ๋“ฑ๊ฑฐ๋ฆฌ ๋นˆ์œผ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ‘œ 1์— ๋นˆํ™”๋œ ํ™•๋ฅ  ๊ฐ’๊ณผ RAG ์„ ํ˜ธ๋„์— ๋Œ€ํ•ด ์„ ํ˜• ํšŒ๊ท€๋ฅผ ์ˆ˜ํ–‰ํ•œ ๊ธฐ์šธ๊ธฐ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์šธ๊ธฐ๋Š” ๋ชจ๋ธ์˜ ์‹ ๋ขฐ๋„๊ฐ€ ๋†’์„์ˆ˜๋ก ๊ฒ€์ƒ‰๋œ ์ปจํ…์ŠคํŠธ์— ์ œ์‹œ๋œ ์ •๋ณด์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์„ ํ˜ธ๋„์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ๋‹ค์–‘ํ•œ ๊ธฐ์šธ๊ธฐ(-0.1์—์„œ -0.45 ๋ฒ”์œ„)๋ฅผ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์–ด ์„œ๋กœ ๋‹ค๋ฅธ QA ๋„๋ฉ”์ธ์—์„œ RAG์˜ ํšจ๊ณผ๊ฐ€ ๋ชจ๋ธ์˜ ๋‚ด๋ถ€ ์‚ฌ์ „ ์ง€์‹ ์‹ ๋ขฐ๋„์— ์ƒ๋Œ€์ ์œผ๋กœ ๋ฏผ๊ฐํ•˜๊ฑฐ๋‚˜(์˜ˆ: ๋‚ ์งœ ์งˆ๋ฌธ) ๊ฐ•๋ ฅํ•˜๊ฑฐ๋‚˜(์˜ˆ: ๋‰ด์Šค ์งˆ๋ฌธ) ํŠน์ง•์ง€์„ ์ˆ˜ ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ธฐ์šธ๊ธฐ๊ฐ€ -0.45์ธ ๊ฒฝ์šฐ, ๋ชจ๋ธ์˜ ์‚ฌ์ „ ์‘๋‹ต ํ™•๋ฅ ์ด 10% ์ฆ๊ฐ€ํ•  ๋•Œ๋งˆ๋‹ค LLM์ด ์ปจํ…์ŠคํŠธ ์ •๋ณด๋ฅผ ์„ ํ˜ธํ•  ํ™•๋ฅ ์ด 4.5% ๊ฐ์†Œํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2(์™ผ์ชฝ ํ”Œ๋กฏ)์—์„œ 6๊ฐœ์˜ QA ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ชจ๋‘์—์„œ ๋ชจ๋ธ์˜ ์‚ฌ์ „ ๋‹ต๋ณ€ ํ† ํฐ ํ™•๋ฅ ๊ณผ ๊ด€๋ จ RAG ์„ ํ˜ธ๋„ ์‚ฌ์ด์— ์ผ๊ด€๋œ ์Œ์˜ ๊ด€๊ณ„๋ฅผ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ™•๋ฅ ์— ๊ฑธ์ณ ๊ท ๋“ฑํ•œ ๋ถ„ํฌ๋ฅผ ์‹œ๊ฐํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ™•๋ฅ ์„ \([0.0,1.0]\ ๋ฒ”์œ„์—์„œ 10๊ฐœ์˜ ๋“ฑ๊ฑฐ๋ฆฌ ๋นˆ์œผ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ‘œ 1์— ๋นˆํ™”๋œ ํ™•๋ฅ  ๊ฐ’๊ณผ RAG ์„ ํ˜ธ๋„์— ๋Œ€ํ•ด ์„ ํ˜• ํšŒ๊ท€๋ฅผ ์ˆ˜ํ–‰ํ•œ ๊ธฐ์šธ๊ธฐ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์šธ๊ธฐ๋Š” ๋ชจ๋ธ์˜ ์‹ ๋ขฐ๋„๊ฐ€ ๋†’์„์ˆ˜๋ก ๊ฒ€์ƒ‰๋œ ์ปจํ…์ŠคํŠธ์— ์ œ์‹œ๋œ ์ •๋ณด์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์„ ํ˜ธ๋„์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ๋‹ค์–‘ํ•œ ๊ธฐ์šธ๊ธฐ(-0.1์—์„œ -0.45 ๋ฒ”์œ„)๋ฅผ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์–ด ์„œ๋กœ ๋‹ค๋ฅธ QA ๋„๋ฉ”์ธ์—์„œ RAG์˜ ํšจ๊ณผ๊ฐ€ ๋ชจ๋ธ์˜ ๋‚ด๋ถ€ ์‚ฌ์ „ ์ง€์‹ ์‹ ๋ขฐ๋„์— ์ƒ๋Œ€์ ์œผ๋กœ ๋ฏผ๊ฐํ•˜๊ฑฐ๋‚˜(์˜ˆ: ๋‚ ์งœ ์งˆ๋ฌธ) ๊ฐ•๋ ฅํ•˜๊ฑฐ๋‚˜(์˜ˆ: ๋‰ด์Šค ์งˆ๋ฌธ) ํŠน์ง•์ง€์„ ์ˆ˜ ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ธฐ์šธ๊ธฐ๊ฐ€ -0.45์ธ ๊ฒฝ์šฐ, ๋ชจ๋ธ์˜ ์‚ฌ์ „ ์‘๋‹ต ํ™•๋ฅ ์ด 10% ์ฆ๊ฐ€ํ•  ๋•Œ๋งˆ๋‹ค LLM์ด ์ปจํ…์ŠคํŠธ ์ •๋ณด๋ฅผ ์„ ํ˜ธํ•  ํ™•๋ฅ ์ด 4.5% ๊ฐ์†Œํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Natural Questions(NQ) (Kwiatkowski et al., 2019) is a very well-known dataset in the NLU field.The given text describes a structured question and a corresponding Wikipedia page. The page is annotated with a long answer, typically a paragraph, and a short answer consisting of one or more entities. If there is no long or short answer, it is labelled as empty. Due to the reliability of the Google search engine and its vast amount of data, many scholars have used this dataset to train RALM, such as Wang et al. (2023c), Izacard et al. (2022) and Ram et al. (2023). HotpotQA(HQA) (Yang et al., 2018) stores information about multi-hop questions and provides sentence-level supporting facts needed for inference. The structure includes the paragraph, question, answer, and sentence number that supports the answer. This dataset has been used by many researchers, such as Wang et al. (2023c), Khattab et al. (2022), and Feng et al. (2024), to train RALM for multi-hop question answering. Another significant form of structured data is the Knowledge Graph. It is a data structure that primarily consists of triples of (entities, relationships, attributes).
Natural Questions(NQ) (Kwiatkowski et al., 2019)๋Š” ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ ๋งค์šฐ ์ž˜ ์•Œ๋ ค์ง„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ฃผ์–ด์ง„ ํ…์ŠคํŠธ๋Š” ๊ตฌ์กฐํ™”๋œ ์งˆ๋ฌธ๊ณผ ๊ทธ์— ํ•ด๋‹นํ•˜๋Š” Wikipedia ํŽ˜์ด์ง€๋ฅผ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด ํŽ˜์ด์ง€์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ํ•œ ๋‹จ๋ฝ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๊ธด ๋‹ต๋ณ€๊ณผ ํ•˜๋‚˜ ์ด์ƒ์˜ ์—”ํ‹ฐํ‹ฐ๋กœ ๊ตฌ์„ฑ๋œ ์งง์€ ๋‹ต๋ณ€์ด ์ฃผ์„์œผ๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ๊ธด ๋‹ต๋ณ€์ด๋‚˜ ์งง์€ ๋‹ต๋ณ€์ด ์—†๋Š” ๊ฒฝ์šฐ ๋นˆ ๊ฒƒ์œผ๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. Google ๊ฒ€์ƒ‰ ์—”์ง„์˜ ์‹ ๋ขฐ์„ฑ๊ณผ ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ๋ฐ์ดํ„ฐ๋กœ ์ธํ•ด ๋งŽ์€ ํ•™์ž๋“ค์ด ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Wang et al. (2023c), Izacard et al. (2022), Ram et al. (2023)๊ณผ ๊ฐ™์€ RALM์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. HotpotQA(HQA) (Yang et al., 2018)๋Š” ๋‹ค์ค‘ ํ™‰ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๊ณ  ์ถ”๋ก ์— ํ•„์š”ํ•œ ๋ฌธ์žฅ ์ˆ˜์ค€์˜ ๊ทผ๊ฑฐ ์‚ฌ์‹ค์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ตฌ์กฐ์—๋Š” ๋‹จ๋ฝ, ์งˆ๋ฌธ, ๋‹ต๋ณ€, ๋‹ต๋ณ€์„ ๋’ท๋ฐ›์นจํ•˜๋Š” ๋ฌธ์žฅ ๋ฒˆํ˜ธ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” Wang et al. (2023c), Khattab et al. (2022), Feng et al. (2024) ๋“ฑ ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด ๋‹ค์ค‘ ํ™‰ ์งˆ๋ฌธ์— ๋Œ€ํ•œ RALM์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ค‘์š”ํ•œ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ ํ˜•ํƒœ๋Š” Knowledge Graph์ž…๋‹ˆ๋‹ค. Knowledge Graph๋Š” ์ฃผ๋กœ (์—”ํ‹ฐํ‹ฐ, ๊ด€๊ณ„, ์†์„ฑ)์˜ ํŠธ๋ฆฌํ”Œ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ๊ตฌ์กฐ์ž…๋‹ˆ๋‹ค.
Natural Questions(NQ) (Kwiatkowski et al., 2019)๋Š” ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ ๋งค์šฐ ์ž˜ ์•Œ๋ ค์ง„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค.์ฃผ์–ด์ง„ ํ…์ŠคํŠธ๋Š” ๊ตฌ์กฐํ™”๋œ ์งˆ๋ฌธ๊ณผ ๊ทธ์— ํ•ด๋‹นํ•˜๋Š” Wikipedia ํŽ˜์ด์ง€๋ฅผ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด ํŽ˜์ด์ง€์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ํ•œ ๋‹จ๋ฝ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๊ธด ๋‹ต๋ณ€๊ณผ ํ•˜๋‚˜ ์ด์ƒ์˜ ์—”ํ‹ฐํ‹ฐ๋กœ ๊ตฌ์„ฑ๋œ ์งง์€ ๋‹ต๋ณ€์ด ์ฃผ์„์œผ๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ๊ธด ๋‹ต๋ณ€์ด๋‚˜ ์งง์€ ๋‹ต๋ณ€์ด ์—†๋Š” ๊ฒฝ์šฐ ๋นˆ ๊ฒƒ์œผ๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. Google ๊ฒ€์ƒ‰ ์—”์ง„์˜ ์‹ ๋ขฐ์„ฑ๊ณผ ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ๋ฐ์ดํ„ฐ๋กœ ์ธํ•ด ๋งŽ์€ ํ•™์ž๋“ค์ด ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Wang et al. (2023c), Izacard et al. (2022), Ram et al. (2023)๊ณผ ๊ฐ™์€ RALM์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. HotpotQA(HQA) (Yang et al., 2018)๋Š” ๋‹ค์ค‘ ํ™‰ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๊ณ  ์ถ”๋ก ์— ํ•„์š”ํ•œ ๋ฌธ์žฅ ์ˆ˜์ค€์˜ ๊ทผ๊ฑฐ ์‚ฌ์‹ค์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ตฌ์กฐ์—๋Š” ๋‹จ๋ฝ, ์งˆ๋ฌธ, ๋‹ต๋ณ€, ๋‹ต๋ณ€์„ ๋’ท๋ฐ›์นจํ•˜๋Š” ๋ฌธ์žฅ ๋ฒˆํ˜ธ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” Wang et al. (2023c), Khattab et al. (2022), Feng et al. (2024) ๋“ฑ ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด ๋‹ค์ค‘ ํ™‰ ์งˆ๋ฌธ์— ๋Œ€ํ•œ RALM์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ค‘์š”ํ•œ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ ํ˜•ํƒœ๋Š” ์ง€์‹ ๊ทธ๋ž˜ํ”„์ž…๋‹ˆ๋‹ค. ์ง€์‹ ๊ทธ๋ž˜ํ”„๋Š” ์ฃผ๋กœ (์—”ํ‹ฐํ‹ฐ, ๊ด€๊ณ„, ์†์„ฑ)์˜ ํŠธ๋ฆฌํ”Œ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ๊ตฌ์กฐ์ž…๋‹ˆ๋‹ค.
Autonomous agents have long been a research focus in academic and industry communities. Previous research often focuses on training agents with limited knowledge within isolated environments, which diverges significantly from human learning processes, and makes the agents hard to achieve human-like decisions. Recently, through the acquisition of vast amounts of web knowledge, large language models (LLMs) have shown potential in human-level intelligence, leading to a surge in research on LLM-based autonomous agents. In this paper, we present a comprehensive survey of these studies, delivering a systematic review of LLM-based autonomous agents from a holistic perspective. We first discuss the construction of LLM-based autonomous agents, proposing a unified framework that encompasses much of previous work. Then, we present a overview of the diverse applications of LLM-based autonomous agents in social science, natural science, and engineering. Finally, we delve into the evaluation strategies commonly used for LLM-based autonomous agents. Based on the previous studies, we also present several challenges and future directions in this field.
์ž์œจ agents๋Š” ํ•™๊ณ„์™€ ์—…๊ณ„ ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ์˜ค๋žซ๋™์•ˆ ์—ฐ๊ตฌ์˜ ์ดˆ์ ์ด ๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ๋Š” ์ข…์ข… ๊ฒฉ๋ฆฌ๋œ ํ™˜๊ฒฝ ๋‚ด์—์„œ ์ œํ•œ๋œ ์ง€์‹์„ ๊ฐ€์ง„ agents๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถ”์—ˆ๋Š”๋ฐ, ์ด๋Š” ์ธ๊ฐ„์˜ ํ•™์Šต ๊ณผ์ •๊ณผ ํฌ๊ฒŒ ๋‹ค๋ฅด๋ฉฐ, agents๊ฐ€ ์ธ๊ฐ„๊ณผ ๊ฐ™์€ ์˜์‚ฌ ๊ฒฐ์ •์„ ๋‚ด๋ฆฌ๊ธฐ ์–ด๋ ต๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์›น ์ง€์‹์„ ์Šต๋“ํ•จ์œผ๋กœ์จ Large Language Models (LLMs)์ด ์ธ๊ฐ„ ์ˆ˜์ค€์˜ ์ง€๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋ฉด์„œ LLM ๊ธฐ๋ฐ˜ ์ž์œจ agents์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๊ฐ€ ๊ธ‰์ฆํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์กฐ์‚ฌ๋ฅผ ํ†ตํ•ด ์ด์ฒด์ ์ธ ๊ด€์ ์—์„œ LLM ๊ธฐ๋ฐ˜ ์ž์œจ agents์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ๊ฒ€ํ† ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ €, LLM ๊ธฐ๋ฐ˜ ์ž์œจ agents์˜ ๊ตฌ์„ฑ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•˜๊ณ , ์ด์ „ ์—ฐ๊ตฌ์˜ ๋Œ€๋ถ€๋ถ„์„ ํฌ๊ด„ํ•˜๋Š” ํ†ตํ•ฉ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์‚ฌํšŒ ๊ณผํ•™, ์ž์—ฐ ๊ณผํ•™ ๋ฐ ๊ณตํ•™ ๋ถ„์•ผ์—์„œ LLM ๊ธฐ๋ฐ˜ ์ž์œจ agents์˜ ๋‹ค์–‘ํ•œ ์‘์šฉ์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, LLM ๊ธฐ๋ฐ˜ ์ž์œจ agents์— ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ํ‰๊ฐ€ ์ „๋žต์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด์ „ ์—ฐ๊ตฌ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์ด ๋ถ„์•ผ์˜ ๋ช‡ ๊ฐ€์ง€ ๊ณผ์ œ์™€ ํ–ฅํ›„ ๋ฐฉํ–ฅ์— ๋Œ€ํ•ด์„œ๋„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
์ž์œจ ์—์ด์ „ํŠธ๋Š” ํ•™๊ณ„์™€ ์—…๊ณ„ ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ์˜ค๋žซ๋™์•ˆ ์—ฐ๊ตฌ์˜ ์ดˆ์ ์ด ๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ๋Š” ์ข…์ข… ๊ฒฉ๋ฆฌ๋œ ํ™˜๊ฒฝ ๋‚ด์—์„œ ์ œํ•œ๋œ ์ง€์‹์„ ๊ฐ€์ง„ ์—์ด์ „ํŠธ๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถ”์—ˆ๋Š”๋ฐ, ์ด๋Š” ์ธ๊ฐ„์˜ ํ•™์Šต ๊ณผ์ •๊ณผ ํฌ๊ฒŒ ๋‹ค๋ฅด๋ฉฐ, ์—์ด์ „ํŠธ๊ฐ€ ์ธ๊ฐ„๊ณผ ๊ฐ™์€ ์˜์‚ฌ ๊ฒฐ์ •์„ ๋‚ด๋ฆฌ๊ธฐ ์–ด๋ ต๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์›น ์ง€์‹์„ ์Šต๋“ํ•จ์œผ๋กœ์จ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์ด ์ธ๊ฐ„ ์ˆ˜์ค€์˜ ์ง€๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋ฉด์„œ LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๊ฐ€ ๊ธ‰์ฆํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์กฐ์‚ฌ๋ฅผ ํ†ตํ•ด ์ด์ฒด์ ์ธ ๊ด€์ ์—์„œ LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ๊ฒ€ํ† ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ €, LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์˜ ๊ตฌ์„ฑ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•˜๊ณ , ์ด์ „ ์—ฐ๊ตฌ์˜ ๋Œ€๋ถ€๋ถ„์„ ํฌ๊ด„ํ•˜๋Š” ํ†ตํ•ฉ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์‚ฌํšŒ ๊ณผํ•™, ์ž์—ฐ ๊ณผํ•™ ๋ฐ ๊ณตํ•™ ๋ถ„์•ผ์—์„œ LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์˜ ๋‹ค์–‘ํ•œ ์‘์šฉ์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์— ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ํ‰๊ฐ€ ์ „๋žต์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด์ „ ์—ฐ๊ตฌ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์ด ๋ถ„์•ผ์˜ ๋ช‡ ๊ฐ€์ง€ ๊ณผ์ œ์™€ ํ–ฅํ›„ ๋ฐฉํ–ฅ์— ๋Œ€ํ•ด์„œ๋„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
We present Tables, the first large-scale human evaluation of faithfulness and content selection in book-length summarization. By recruiting annotators who had read recently-published books for enjoyment, we collect 3,158 claim-level faithfulness annotations from LLM-generated summaries of 26 narratives. This allows us to rank LLM summarizers based on faithfulness, revealing that Claude-3-Opus is the most faithful book-length summarizer, followed by GPT-4-Turoo. Next, we experiment with using LLMs for automatic claim verification. Our results expose the limitations of both retrieval and long-context understanding: LLM auto-raters cannot reliably detect _unfaithful_ claims, even when prompted with the full book text. Our analysis shows that unfaithful claims primarily pertain to states and events, often necessitating reasoning over extended contexts, which makes them complicated to detect for both humans and machines. Finally, we move beyond faithfulness to explore and characterize common content selection errors such as omissions of key events, attributes, and characters, as well as the over-emphasis of content from the end of the book.
์šฐ๋ฆฌ๋Š” Tables, ์ฑ… ๊ธธ์ด ์š”์•ฝ์—์„œ ์ถฉ์‹ค๋„์™€ ์ฝ˜ํ…์ธ  ์„ ํƒ์— ๋Œ€ํ•œ ์ตœ์ดˆ์˜ ๋Œ€๊ทœ๋ชจ ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์— ์ถœ๊ฐ„๋œ ์ฑ…์„ ์žฌ๋ฏธ๋กœ ์ฝ์€ ์ ์ด ์žˆ๋Š” ์ฃผ์„๊ฐ€๋“ค์„ ๋ชจ์ง‘ํ•˜์—ฌ 26๊ฐœ์˜ ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ์— ๋Œ€ํ•œ LLM ์ƒ์„ฑ ์š”์•ฝ์—์„œ 3,158๊ฐœ์˜ claim-level faithfulness ์ฃผ์„์„ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ถฉ์‹ค๋„๋ฅผ ๊ธฐ์ค€์œผ๋กœ LLM summarizer๋ฅผ ์ˆœ์œ„๋ฅผ ๋งค๊ธธ ์ˆ˜ ์žˆ์œผ๋ฉฐ, Claude-3-Opus๊ฐ€ ๊ฐ€์žฅ ์ถฉ์‹คํ•œ ์ฑ… ๊ธธ์ด ์š”์•ฝ๊ธฐ์ด๋ฉฐ ๊ทธ๋‹ค์Œ์ด GPT-4-Turoo์ธ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, LLM์„ ์ž๋™ claim ๊ฒ€์ฆ์— ์‚ฌ์šฉํ•˜๋Š” ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ retrieval๊ณผ long-context understanding ๋ชจ๋‘์˜ ํ•œ๊ณ„๋ฅผ ๋“œ๋Ÿฌ๋ƒˆ์Šต๋‹ˆ๋‹ค. LLM auto-rater๋Š” ์ฑ… ์ „์ฒด ํ…์ŠคํŠธ๋ฅผ promptํ•˜๋”๋ผ๋„ _unfaithful_ claim์„ ์•ˆ์ •์ ์œผ๋กœ ๊ฐ์ง€ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋ถ„์„ ๊ฒฐ๊ณผ, unfaithful claim์€ ์ฃผ๋กœ ์ƒํƒœ ๋ฐ ์‚ฌ๊ฑด๊ณผ ๊ด€๋ จ์ด ์žˆ์œผ๋ฉฐ, ์ข…์ข… ํ™•์žฅ๋œ ์ปจํ…์ŠคํŠธ์— ๋Œ€ํ•œ ์ถ”๋ก ์ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ๋žŒ๊ณผ ๊ธฐ๊ณ„ ๋ชจ๋‘์—์„œ ๊ฐ์ง€ํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ถฉ์‹ค๋„๋ฅผ ๋„˜์–ด ํ•ต์‹ฌ ์‚ฌ๊ฑด, ์†์„ฑ, ์ธ๋ฌผ์˜ ๋ˆ„๋ฝ, ์ฑ… ๋งˆ์ง€๋ง‰ ๋ถ€๋ถ„์˜ ๋‚ด์šฉ์— ๋Œ€ํ•œ ์ง€๋‚˜์นœ ๊ฐ•์กฐ์™€ ๊ฐ™์€ ์ผ๋ฐ˜์ ์ธ ์ฝ˜ํ…์ธ  ์„ ํƒ ์˜ค๋ฅ˜๋ฅผ ํƒ์ƒ‰ํ•˜๊ณ  ํŠน์„ฑํ™”ํ•ฉ๋‹ˆ๋‹ค.
์šฐ๋ฆฌ๋Š” ์ตœ๊ทผ์— ์ถœ๊ฐ„๋œ ์ฑ…์„ ์žฌ๋ฏธ๋กœ ์ฝ์€ ์ ์ด ์žˆ๋Š” ์ฃผ์„๊ฐ€๋“ค์„ ๋ชจ์ง‘ํ•˜์—ฌ 26๊ฐœ์˜ ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ์— ๋Œ€ํ•œ LLM ์ƒ์„ฑ ์š”์•ฝ์—์„œ 3,158๊ฐœ์˜ ํด๋ ˆ์ž„ ์ˆ˜์ค€ ์ถฉ์‹ค๋„ ์ฃผ์„์„ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ถฉ์‹ค๋„๋ฅผ ๊ธฐ์ค€์œผ๋กœ LLM ์š”์•ฝ๊ธฐ๋ฅผ ์ˆœ์œ„๋ฅผ ๋งค๊ธธ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํด๋กœ๋“œ-3-์˜คํผ์Šค๊ฐ€ ๊ฐ€์žฅ ์ถฉ์‹คํ•œ ์žฅํŽธ ์š”์•ฝ๊ธฐ์ด๋ฉฐ ๊ทธ๋‹ค์Œ์ด GPT-4-ํˆฌ๋ฃจ์ธ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, LLM์„ ์ž๋™ ํด๋ ˆ์ž„ ๊ฒ€์ฆ์— ์‚ฌ์šฉํ•˜๋Š” ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ๊ฒ€์ƒ‰๊ณผ ๊ธด ๋ฌธ๋งฅ ์ดํ•ด ๋ชจ๋‘์˜ ํ•œ๊ณ„๋ฅผ ๋“œ๋Ÿฌ๋ƒˆ์Šต๋‹ˆ๋‹ค. LLM ์ž๋™ ํ‰๊ฐ€์ž๋Š” ์ฑ… ์ „์ฒด ํ…์ŠคํŠธ๋ฅผ ํ”„๋กฌํ”„ํŠธํ•˜๋”๋ผ๋„ _๋ถˆ์ถฉ์‹คํ•œ_ ํด๋ ˆ์ž„์„ ์•ˆ์ •์ ์œผ๋กœ ๊ฐ์ง€ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋ถ„์„ ๊ฒฐ๊ณผ, ๋ถˆ์ถฉ์‹คํ•œ ํด๋ ˆ์ž„์€ ์ฃผ๋กœ ์ƒํƒœ ๋ฐ ์‚ฌ๊ฑด๊ณผ ๊ด€๋ จ์ด ์žˆ์œผ๋ฉฐ, ์ข…์ข… ํ™•์žฅ๋œ ์ปจํ…์ŠคํŠธ์— ๋Œ€ํ•œ ์ถ”๋ก ์ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ๋žŒ๊ณผ ๊ธฐ๊ณ„ ๋ชจ๋‘์—์„œ ๊ฐ์ง€ํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ถฉ์‹ค๋„๋ฅผ ๋„˜์–ด ํ•ต์‹ฌ ์‚ฌ๊ฑด, ์†์„ฑ, ์ธ๋ฌผ์˜ ๋ˆ„๋ฝ, ์ฑ… ๋งˆ์ง€๋ง‰ ๋ถ€๋ถ„์˜ ๋‚ด์šฉ์— ๋Œ€ํ•œ ์ง€๋‚˜์นœ ๊ฐ•์กฐ์™€ ๊ฐ™์€ ์ผ๋ฐ˜์ ์ธ ์ฝ˜ํ…์ธ  ์„ ํƒ ์˜ค๋ฅ˜๋ฅผ ํƒ์ƒ‰ํ•˜๊ณ  ํŠน์„ฑํ™”ํ•ฉ๋‹ˆ๋‹ค.
position, and the input for the second forward pass incorporates some of the predicted tokens from the first pass. In determining which tokens to replace, we employ a sampling strategy characterized by the parameter \(\tau\), representing temperature; a higher \(\tau\) introduces more randomness. At lower temperatures, there is a higher likelihood of replacing tokens with their synonyms, as opposed to random replacement. MaskedLM. In our main experiments, we implement MFT by this method, adhering to the methodologies of prior Masked Language Models as (Bao et al., 2020; Devlin et al., 2019). Unlike encoder-based models (Devlin et al., 2019), which typically predict the current token at masked positions, we predict the next token at masked positions (Xie et al., 2016; Chen et al., 2023). Upon determining the positions for masking, we are presented with two options: employing a [mask] token or substituting with a random token. The probability of selecting the [mask] token is denoted by \(m\), and \(r\) represents the probability of opting for a random token. Su et al. (2022) set \(r\) to 0, while Devlin et al. (2019) set \(r\) to 0.1.
์œ„์น˜, ๊ทธ๋ฆฌ๊ณ  ๋‘ ๋ฒˆ์งธ ์ˆœ๋ฐฉํ–ฅ ํŒจ์Šค๋ฅผ ์œ„ํ•œ ์ž…๋ ฅ์€ ์ฒซ ๋ฒˆ์งธ ํŒจ์Šค์—์„œ ์˜ˆ์ธก๋œ ํ† ํฐ ์ค‘ ์ผ๋ถ€๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์–ด๋–ค ํ† ํฐ์„ ๋Œ€์ฒดํ• ์ง€ ๊ฒฐ์ •ํ•  ๋•Œ, ์˜จ๋„๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ(\(\tau\))๋กœ ํŠน์ง•์ง€์–ด์ง€๋Š” ์ƒ˜ํ”Œ๋ง ์ „๋žต์„ ์‚ฌ์šฉํ•˜๋ฉฐ, \(\tau\)๊ฐ€ ๋†’์„์ˆ˜๋ก ๋ฌด์ž‘์œ„์„ฑ์ด ๋” ๋งŽ์ด ๋„์ž…๋ฉ๋‹ˆ๋‹ค. ์˜จ๋„๊ฐ€ ๋‚ฎ์„์ˆ˜๋ก ๋ฌด์ž‘์œ„๋กœ ํ† ํฐ์„ ๋Œ€์ฒดํ•˜๋Š” ๋Œ€์‹  ๋™์˜์–ด๋กœ ํ† ํฐ์„ ๋Œ€์ฒดํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์•„์ง‘๋‹ˆ๋‹ค. MaskedLM. ์ฃผ์š” ์‹คํ—˜์—์„œ๋Š” ์ด ๋ฐฉ๋ฒ•์œผ๋กœ MFT๋ฅผ ๊ตฌํ˜„ํ•˜์—ฌ (Bao et al., 2020; Devlin et al., 2019)์˜ ์ด์ „ Masked Language Models์˜ ๋ฐฉ๋ฒ•๋ก ์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋งˆ์Šคํฌ ์œ„์น˜์—์„œ ํ˜„์žฌ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” Encoder ๊ธฐ๋ฐ˜ ๋ชจ๋ธ(Devlin et al., 2019)๊ณผ ๋‹ฌ๋ฆฌ, ์šฐ๋ฆฌ๋Š” ๋งˆ์Šคํฌ ์œ„์น˜์—์„œ ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค(Xie et al., 2016; Chen et al., 2023). ๋งˆ์Šคํ‚นํ•  ์œ„์น˜๋ฅผ ๊ฒฐ์ •ํ•˜๋ฉด [mask] ํ† ํฐ์„ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ ์ž„์˜์˜ ํ† ํฐ์œผ๋กœ ๋Œ€์ฒดํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ์˜ต์…˜์ด ์ œ์‹œ๋ฉ๋‹ˆ๋‹ค. [mask] ํ† ํฐ์„ ์„ ํƒํ•  ํ™•๋ฅ ์€ \(m\)์œผ๋กœ ํ‘œ์‹œ๋˜๋ฉฐ, \(r\)์€ ์ž„์˜์˜ ํ† ํฐ์„ ์„ ํƒํ•  ํ™•๋ฅ ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. Su et al. (2022)์€ \(r\)์„ 0์œผ๋กœ ์„ค์ •ํ•˜๊ณ , Devlin et al. (2019)์€ \(r\)์„ 0.1๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
์œ„์น˜, ๊ทธ๋ฆฌ๊ณ  ๋‘ ๋ฒˆ์งธ ์ˆœ๋ฐฉํ–ฅ ํŒจ์Šค๋ฅผ ์œ„ํ•œ ์ž…๋ ฅ์€ ์ฒซ ๋ฒˆ์งธ ํŒจ์Šค์—์„œ ์˜ˆ์ธก๋œ ํ† ํฐ ์ค‘ ์ผ๋ถ€๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์–ด๋–ค ํ† ํฐ์„ ๋Œ€์ฒดํ• ์ง€ ๊ฒฐ์ •ํ•  ๋•Œ, ์˜จ๋„๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ(\(\tau\))๋กœ ํŠน์ง•์ง€์–ด์ง€๋Š” ์ƒ˜ํ”Œ๋ง ์ „๋žต์„ ์‚ฌ์šฉํ•˜๋ฉฐ, \(\tau\)๊ฐ€ ๋†’์„์ˆ˜๋ก ๋ฌด์ž‘์œ„์„ฑ์ด ๋” ๋งŽ์ด ๋„์ž…๋ฉ๋‹ˆ๋‹ค. ์˜จ๋„๊ฐ€ ๋‚ฎ์„์ˆ˜๋ก ๋ฌด์ž‘์œ„๋กœ ํ† ํฐ์„ ๋Œ€์ฒดํ•˜๋Š” ๋Œ€์‹  ๋™์˜์–ด๋กœ ํ† ํฐ์„ ๋Œ€์ฒดํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์•„์ง‘๋‹ˆ๋‹ค. ๋งˆ์ŠคํฌLM. ์ฃผ์š” ์‹คํ—˜์—์„œ๋Š” ์ด ๋ฐฉ๋ฒ•์œผ๋กœ MFT๋ฅผ ๊ตฌํ˜„ํ•˜์—ฌ (Bao et al., 2020; Devlin et al., 2019)์˜ ์ด์ „ ๋งˆ์Šคํฌ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ฐฉ๋ฒ•๋ก ์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋งˆ์Šคํฌ ์œ„์น˜์—์„œ ํ˜„์žฌ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ์ธ์ฝ”๋” ๊ธฐ๋ฐ˜ ๋ชจ๋ธ(Devlin et al., 2019)๊ณผ ๋‹ฌ๋ฆฌ, ์šฐ๋ฆฌ๋Š” ๋งˆ์Šคํฌ ์œ„์น˜์—์„œ ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค(Xie et al., 2016; Chen et al., 2023). ๋งˆ์Šคํ‚นํ•  ์œ„์น˜๋ฅผ ๊ฒฐ์ •ํ•˜๋ฉด [๋งˆ์Šคํฌ] ํ† ํฐ์„ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ ์ž„์˜์˜ ํ† ํฐ์œผ๋กœ ๋Œ€์ฒดํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ์˜ต์…˜์ด ์ œ์‹œ๋ฉ๋‹ˆ๋‹ค. ๋งˆ์Šคํฌ] ํ† ํฐ์„ ์„ ํƒํ•  ํ™•๋ฅ ์€ \(m\)์œผ๋กœ ํ‘œ์‹œ๋˜๋ฉฐ, \(r\)์€ ์ž„์˜์˜ ํ† ํฐ์„ ์„ ํƒํ•  ํ™•๋ฅ ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. Su et al. (2022)์€ \(r\)์„ 0์œผ๋กœ ์„ค์ •ํ•˜๊ณ , Devlin et al. (2019)์€ \(r\)์„ 0.1๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
Other reasoning tasks.Synthetic data also leads to impressive performance in other reasoning tasks. For instance, Wei et al. (2023a) augmented existing natural language datasets by replacing natural language labels with arbitrary symbols, generating over 500k synthetic examples. Using these synthetic data for supervised finetuning significantly improved model performance on unseen in-context learning and algorithmic-reasoning tasks. STaR (Zelikman et al., 2022) generates synthetic chain-of-thought rationales and filters out those leading to wrong answers for finetuning language models to improve their reasoning. In the domain of physics reasoning, Mind's Eye (Liu et al., 2022) takes a novel approach by training a text-to-code model with synthetic "text-description \(\rightarrow\) rendering code" data. This enables the model to convert textual questions into rendering code, which is then executed in a physical engine (i.e., DeepMind MuJoCo (Todorov et al., 2012)). The rendering results are injected into the context, allowing even small language models armed with Mind's Eye to achieve performance comparable to models 100 times larger.
๋‹ค๋ฅธ ์ถ”๋ก  ์ž‘์—…. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๋‹ค๋ฅธ ์ถ”๋ก  ์ž‘์—…์—์„œ๋„ ์ธ์ƒ์ ์ธ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Wei ๋“ฑ(2023a)์€ ์ž์—ฐ์–ด ๋ ˆ์ด๋ธ”์„ ์ž„์˜์˜ ๊ธฐํ˜ธ๋กœ ๋Œ€์ฒดํ•˜์—ฌ ๊ธฐ์กด ์ž์—ฐ์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ฆ๊ฐ•ํ•˜์—ฌ 50๋งŒ ๊ฐœ ์ด์ƒ์˜ ํ•ฉ์„ฑ ์˜ˆ์ œ๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ Supervised Fine-tuning์— ์‚ฌ์šฉํ•˜๋ฉด ๋ณด์ด์ง€ ์•Š๋Š” In-Context Learning ๋ฐ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์ถ”๋ก  ์ž‘์—…์—์„œ ๋ชจ๋ธ ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. STaR(Zelikman et al., 2022)์€ ํ•ฉ์„ฑ๋œ ์ƒ๊ฐ์˜ ์—ฐ์‡„ ํ•ฉ๋ฆฌ์  ๊ทผ๊ฑฐ๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ์ž˜๋ชป๋œ ๋‹ต๋ณ€์œผ๋กœ ์ด์–ด์ง€๋Š” ๊ทผ๊ฑฐ๋ฅผ ๊ฑธ๋Ÿฌ๋‚ด์–ด ์–ธ์–ด ๋ชจ๋ธ์„ Fine-tuningํ•˜์—ฌ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋ฌผ๋ฆฌํ•™ ์ถ”๋ก  ์˜์—ญ์—์„œ Mind's Eye(Liu et al., 2022)๋Š” ํ•ฉ์„ฑ๋œ 'ํ…์ŠคํŠธ ์„ค๋ช…(\(\rightarrow\) ๋ Œ๋”๋ง ์ฝ”๋“œ' ๋ฐ์ดํ„ฐ๋กœ ํ…์ŠคํŠธ-์ฝ”๋“œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ์ƒˆ๋กœ์šด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์€ ํ…์ŠคํŠธ ์งˆ๋ฌธ์„ ๋ Œ๋”๋ง ์ฝ”๋“œ๋กœ ๋ณ€ํ™˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด ์ฝ”๋“œ๋Š” ๋ฌผ๋ฆฌ ์—”์ง„(์˜ˆ: DeepMind MuJoCo(Todorov et al., 2012))์—์„œ ์‹คํ–‰๋ฉ๋‹ˆ๋‹ค. ๋ Œ๋”๋ง ๊ฒฐ๊ณผ๋Š” ์ปจํ…์ŠคํŠธ์— ์ฃผ์ž…๋˜์–ด Mind's Eye๋ฅผ ๊ฐ–์ถ˜ ์ž‘์€ ์–ธ์–ด ๋ชจ๋ธ๋„ 100๋ฐฐ ๋” ํฐ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.
๋‹ค๋ฅธ ์ถ”๋ก  ์ž‘์—….ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๋‹ค๋ฅธ ์ถ”๋ก  ์ž‘์—…์—์„œ๋„ ์ธ์ƒ์ ์ธ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Wei ๋“ฑ(2023a)์€ ์ž์—ฐ์–ด ๋ ˆ์ด๋ธ”์„ ์ž„์˜์˜ ๊ธฐํ˜ธ๋กœ ๋Œ€์ฒดํ•˜์—ฌ ๊ธฐ์กด ์ž์—ฐ์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ฆ๊ฐ•ํ•˜์—ฌ 50๋งŒ ๊ฐœ ์ด์ƒ์˜ ํ•ฉ์„ฑ ์˜ˆ์ œ๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •์— ์‚ฌ์šฉํ•˜๋ฉด ๋ณด์ด์ง€ ์•Š๋Š” ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต ๋ฐ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์ถ”๋ก  ์ž‘์—…์—์„œ ๋ชจ๋ธ ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. STaR(Zelikman et al., 2022)์€ ํ•ฉ์„ฑ๋œ ์ƒ๊ฐ์˜ ์—ฐ์‡„ ํ•ฉ๋ฆฌ์  ๊ทผ๊ฑฐ๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ์ž˜๋ชป๋œ ๋‹ต๋ณ€์œผ๋กœ ์ด์–ด์ง€๋Š” ๊ทผ๊ฑฐ๋ฅผ ๊ฑธ๋Ÿฌ๋‚ด์–ด ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋ฌผ๋ฆฌํ•™ ์ถ”๋ก  ์˜์—ญ์—์„œ Mind's Eye(Liu et al., 2022)๋Š” ํ•ฉ์„ฑ๋œ 'ํ…์ŠคํŠธ ์„ค๋ช…(\์šฐ์ธก์ˆ˜๋ ด) ๋ Œ๋”๋ง ์ฝ”๋“œ' ๋ฐ์ดํ„ฐ๋กœ ํ…์ŠคํŠธ-์ฝ”๋“œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ์ƒˆ๋กœ์šด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์€ ํ…์ŠคํŠธ ์งˆ๋ฌธ์„ ๋ Œ๋”๋ง ์ฝ”๋“œ๋กœ ๋ณ€ํ™˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด ์ฝ”๋“œ๋Š” ๋ฌผ๋ฆฌ ์—”์ง„(์˜ˆ: DeepMind MuJoCo(Todorov et al., 2012))์—์„œ ์‹คํ–‰๋ฉ๋‹ˆ๋‹ค. ๋ Œ๋”๋ง ๊ฒฐ๊ณผ๋Š” ์ปจํ…์ŠคํŠธ์— ์ฃผ์ž…๋˜์–ด Mind's Eye๋ฅผ ๊ฐ–์ถ˜ ์ž‘์€ ์–ธ์–ด ๋ชจ๋ธ๋„ 100๋ฐฐ ๋” ํฐ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.
We evaluate the models on financial tasks to evaluate the effectiveness of our domain-adaptive continual pre-training. We adopt the _FLARE_ framework [33] to evaluate our models. FLARE extends the LLM evaluation framework _lm-evaluation-harmess3_ by including various financial tasks. We follow their instruction prompt, data split, and metric computation for comparison. We consider following 4 tasks used in [31; 33]: (1) **Financial Phrase Bank**. FPB is a sentiment classification task on financial news [19]. The sentiment reflects whether the news is considered as positive/neutral/negative by investors. (2) **FiQA SA.** An aspect based sentiment classification task based on financial news and headlines [18]. (3) **Headline.** Binary classification task on whether a headline on a financial entity contains certain information [25]. Each news article is associated with 9 tags like "price or not", "price up", "price down", "price stable", "past price", and "asset". (4) **NER.** Financial named entity extraction task is based on credit risk assessment section of SEC reports. Words in this task are annotated with PER, LOC, ORG, and MISC.
์žฌ๋ฌด ์ž‘์—…์— ๋Œ€ํ•œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜์—ฌ ๋„๋ฉ”์ธ ์ ์‘ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์˜ ํšจ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด _FLARE_ ํ”„๋ ˆ์ž„์›Œํฌ [33]๋ฅผ ์ฑ„ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. FLARE๋Š” ๋‹ค์–‘ํ•œ ์žฌ๋ฌด ์ž‘์—…์„ ํฌํ•จํ•˜์—ฌ LLM ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ _lm-evaluation-harmess3_๋ฅผ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค. ๋น„๊ต๋ฅผ ์œ„ํ•ด instruction prompt, ๋ฐ์ดํ„ฐ ๋ถ„ํ•  ๋ฐ metric ๊ณ„์‚ฐ์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. [31; 33]์—์„œ ์‚ฌ์šฉํ•œ ๋‹ค์Œ 4๊ฐ€์ง€ ์ž‘์—…์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค: (1) **๊ธˆ์œต ๊ตฌ๋ฌธ ์€ํ–‰**. FPB๋Š” ๊ธˆ์œต ๋‰ด์Šค์— ๋Œ€ํ•œ ๊ฐ์„ฑ ๋ถ„๋ฅ˜ ์ž‘์—…์ž…๋‹ˆ๋‹ค [19]. ์ด ๊ฐ์„ฑ์€ ํˆฌ์ž์ž๊ฐ€ ๋‰ด์Šค๋ฅผ ๊ธ์ •/์ค‘๋ฆฝ/๋ถ€์ •์œผ๋กœ ๊ฐ„์ฃผํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค. (2) **FiQA SA.** ๊ธˆ์œต ๋‰ด์Šค ๋ฐ ํ—ค๋“œ๋ผ์ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ธก๋ฉด ๊ธฐ๋ฐ˜ ๊ฐ์„ฑ ๋ถ„๋ฅ˜ ์ž‘์—… [18]. (3) **ํ—ค๋“œ๋ผ์ธ.** ๊ธˆ์œต ๋ฒ•์ธ์— ๋Œ€ํ•œ ํ—ค๋“œ๋ผ์ธ์— ํŠน์ • ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š”์ง€ ์—ฌ๋ถ€์— ๋Œ€ํ•œ ์ด์ง„ ๋ถ„๋ฅ˜ ์ž‘์—… [25]. ๊ฐ ๋‰ด์Šค ๊ธฐ์‚ฌ์—๋Š” "๊ฐ€๊ฒฉ ๋˜๋Š” ์•„๋‹˜", "๊ฐ€๊ฒฉ ์ƒ์Šน", "๊ฐ€๊ฒฉ ํ•˜๋ฝ", "๊ฐ€๊ฒฉ ์•ˆ์ •", "๊ณผ๊ฑฐ ๊ฐ€๊ฒฉ", "์ž์‚ฐ" ๋“ฑ 9๊ฐœ์˜ ํƒœ๊ทธ๊ฐ€ ์—ฐ๊ฒฐ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. (4) **NER.** ๊ธˆ์œต ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ ์ถ”์ถœ ์ž‘์—…์€ SEC ๋ณด๊ณ ์„œ์˜ ์‹ ์šฉ ์œ„ํ—˜ ํ‰๊ฐ€ ์„น์…˜์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ๋‹จ์–ด์—๋Š” PER, LOC, ORG ๋ฐ MISC๊ฐ€ ์ฃผ์„ ์ฒ˜๋ฆฌ๋ฉ๋‹ˆ๋‹ค.
์žฌ๋ฌด ์ž‘์—…์— ๋Œ€ํ•œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜์—ฌ ๋„๋ฉ”์ธ ์ ์‘ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์˜ ํšจ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด _FLARE_ ํ”„๋ ˆ์ž„์›Œํฌ [33]๋ฅผ ์ฑ„ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. FLARE๋Š” ๋‹ค์–‘ํ•œ ์žฌ๋ฌด ์ž‘์—…์„ ํฌํ•จํ•˜์—ฌ LLM ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ _lm-evaluation-harmess3_๋ฅผ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค. ๋น„๊ต๋ฅผ ์œ„ํ•ด ์ง€์นจ ํ”„๋กฌํ”„ํŠธ, ๋ฐ์ดํ„ฐ ๋ถ„ํ•  ๋ฐ ๋ฉ”ํŠธ๋ฆญ ๊ณ„์‚ฐ์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. [31; 33]์—์„œ ์‚ฌ์šฉํ•œ ๋‹ค์Œ 4๊ฐ€์ง€ ์ž‘์—…์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค: (1) **๊ธˆ์œต ๊ตฌ๋ฌธ ์€ํ–‰**. FPB๋Š” ๊ธˆ์œต ๋‰ด์Šค์— ๋Œ€ํ•œ ๊ฐ์„ฑ ๋ถ„๋ฅ˜ ์ž‘์—…์ž…๋‹ˆ๋‹ค [19]. ์ด ๊ฐ์„ฑ์€ ํˆฌ์ž์ž๊ฐ€ ๋‰ด์Šค๋ฅผ ๊ธ์ •/์ค‘๋ฆฝ/๋ถ€์ •์œผ๋กœ ๊ฐ„์ฃผํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค. (2) **FiQA SA.** ๊ธˆ์œต ๋‰ด์Šค ๋ฐ ํ—ค๋“œ๋ผ์ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ธก๋ฉด ๊ธฐ๋ฐ˜ ๊ฐ์„ฑ ๋ถ„๋ฅ˜ ์ž‘์—… [18]. (3) **ํ—ค๋“œ๋ผ์ธ.** ๊ธˆ์œต ๋ฒ•์ธ์— ๋Œ€ํ•œ ํ—ค๋“œ๋ผ์ธ์— ํŠน์ • ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š”์ง€ ์—ฌ๋ถ€์— ๋Œ€ํ•œ ์ด์ง„ ๋ถ„๋ฅ˜ ์ž‘์—… [25]. ๊ฐ ๋‰ด์Šค ๊ธฐ์‚ฌ์—๋Š” "๊ฐ€๊ฒฉ ๋˜๋Š” ์•„๋‹˜", "๊ฐ€๊ฒฉ ์ƒ์Šน", "๊ฐ€๊ฒฉ ํ•˜๋ฝ", "๊ฐ€๊ฒฉ ์•ˆ์ •", "๊ณผ๊ฑฐ ๊ฐ€๊ฒฉ", "์ž์‚ฐ" ๋“ฑ 9๊ฐœ์˜ ํƒœ๊ทธ๊ฐ€ ์—ฐ๊ฒฐ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. (4) **NER.** ๊ธˆ์œต ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ ์ถ”์ถœ ์ž‘์—…์€ SEC ๋ณด๊ณ ์„œ์˜ ์‹ ์šฉ ์œ„ํ—˜ ํ‰๊ฐ€ ์„น์…˜์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ๋‹จ์–ด์—๋Š” PER, LOC, ORG ๋ฐ MISC๊ฐ€ ์ฃผ์„ ์ฒ˜๋ฆฌ๋ฉ๋‹ˆ๋‹ค.
**Experiments on LLMs.** We finetune Llama2 with LoRA, generating synthetic AI data for the next finetuning iteration. Inspired by the setup in Shumailov et al. (2023), we use Wikidata-103, partitioned into approximately \(2.2\) million sequences of 128 tokens. AI data is generated through prompt completion, using the first 96 tokens from the original sequences as prompts. The model is trained only on the last 32 tokens to preclude information leakage, i.e. the model being trained on the ground truth of the same 32 tokens. The evaluations are conducted exclusively on the same 32 tokens. We use top-p 0.9 and temperature 0.9 across all generation. The results, depicted in Figure 4 (left), illustrate a scaling law decay over several generations. The first generated dataset still contain useful but limited information and the utility of the second generation's data markedly diminishes. These phenomena corroborate the anticipated loss of scaling law and model collapse, further indicating that model collapse is even more pronounced here, highlighting the challenges in training next generation LLMs. More details and results in Appendix H.
**LLM์— ๋Œ€ํ•œ ์‹คํ—˜.** ์šฐ๋ฆฌ๋Š” LoRA๋กœ Llama2๋ฅผ Fine-tuningํ•˜์—ฌ ๋‹ค์Œ Fine-tuning ๋ฐ˜๋ณต์„ ์œ„ํ•œ ํ•ฉ์„ฑ AI ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. Shumailov ๋“ฑ(2023)์˜ ์„ค์ •์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ์•ฝ 220๋งŒ ๊ฐœ์˜ 128 ํ† ํฐ ์‹œํ€€์Šค๋กœ ๋ถ„ํ• ๋œ Wikidata-103์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. AI ๋ฐ์ดํ„ฐ๋Š” ์›๋ž˜ ์‹œํ€€์Šค์˜ ์ฒ˜์Œ 96๊ฐœ ํ† ํฐ์„ ํ”„๋กฌํ”„ํŠธ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ ์™„๋ฃŒ๋ฅผ ํ†ตํ•ด ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ •๋ณด ์œ ์ถœ, ์ฆ‰ ๋™์ผํ•œ 32๊ฐœ ํ† ํฐ์˜ ์‹ค์ธก ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต๋˜๋Š” ๋ชจ๋ธ์„ ๋ฐฐ์ œํ•˜๊ธฐ ์œ„ํ•ด ๋งˆ์ง€๋ง‰ 32๊ฐœ ํ† ํฐ์— ๋Œ€ํ•ด์„œ๋งŒ ๋ชจ๋ธ์„ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€๋Š” ๋™์ผํ•œ 32๊ฐœ ํ† ํฐ์— ๋Œ€ํ•ด์„œ๋งŒ ์ˆ˜ํ–‰๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ์ƒ์„ฑ์— ๊ฑธ์ณ top-p 0.9 ๋ฐ temperature 0.9๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 4(์™ผ์ชฝ)์— ํ‘œ์‹œ๋œ ๊ฒฐ๊ณผ๋Š” ์—ฌ๋Ÿฌ ์„ธ๋Œ€์— ๊ฑธ์ณ Scaling law ๊ฐ์†Œ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์—ฌ์ „ํžˆ ์œ ์šฉํ•˜์ง€๋งŒ ์ œํ•œ๋œ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋‘ ๋ฒˆ์งธ ์„ธ๋Œ€ ๋ฐ์ดํ„ฐ์˜ ์œ ์šฉ์„ฑ์€ ํ˜„์ €ํžˆ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ˜„์ƒ์€ ์˜ˆ์ƒ๋˜๋Š” Scaling law ์†์‹ค๊ณผ ๋ชจ๋ธ ๋ถ•๊ดด๋ฅผ ํ™•์ฆํ•˜๋ฉฐ, ๋ชจ๋ธ ๋ถ•๊ดด๊ฐ€ ์—ฌ๊ธฐ์„œ ๋”์šฑ ๋‘๋“œ๋Ÿฌ์ง€๊ฒŒ ๋‚˜ํƒ€๋‚˜ ์ฐจ์„ธ๋Œ€ LLM์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์žˆ์–ด ์–ด๋ ค์›€์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ๊ณผ ๊ฒฐ๊ณผ๋Š” ๋ถ€๋ก H์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**LLM์— ๋Œ€ํ•œ ์‹คํ—˜.** ์šฐ๋ฆฌ๋Š” LoRA๋กœ Llama2๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ๋‹ค์Œ ๋ฏธ์„ธ ์กฐ์ • ๋ฐ˜๋ณต์„ ์œ„ํ•œ ํ•ฉ์„ฑ AI ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. Shumailov ๋“ฑ(2023)์˜ ์„ค์ •์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ์•ฝ 220๋งŒ ๊ฐœ์˜ 128 ํ† ํฐ ์‹œํ€€์Šค๋กœ ๋ถ„ํ• ๋œ ์œ„ํ‚ค๋ฐ์ดํ„ฐ-103์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. AI ๋ฐ์ดํ„ฐ๋Š” ์›๋ž˜ ์‹œํ€€์Šค์˜ ์ฒ˜์Œ 96๊ฐœ ํ† ํฐ์„ ํ”„๋กฌํ”„ํŠธ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ ์™„๋ฃŒ๋ฅผ ํ†ตํ•ด ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ •๋ณด ์œ ์ถœ, ์ฆ‰ ๋™์ผํ•œ 32๊ฐœ ํ† ํฐ์˜ ์‹ค์ธก ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต๋˜๋Š” ๋ชจ๋ธ์„ ๋ฐฐ์ œํ•˜๊ธฐ ์œ„ํ•ด ๋งˆ์ง€๋ง‰ 32๊ฐœ ํ† ํฐ์— ๋Œ€ํ•ด์„œ๋งŒ ๋ชจ๋ธ์„ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€๋Š” ๋™์ผํ•œ 32๊ฐœ ํ† ํฐ์— ๋Œ€ํ•ด์„œ๋งŒ ์ˆ˜ํ–‰๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ์ƒ์„ฑ์— ๊ฑธ์ณ ์ตœ์ƒ์œ„ 0.9 ๋ฐ ์˜จ๋„ 0.9๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 4(์™ผ์ชฝ)์— ํ‘œ์‹œ๋œ ๊ฒฐ๊ณผ๋Š” ์—ฌ๋Ÿฌ ์„ธ๋Œ€์— ๊ฑธ์ณ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™ ๊ฐ์†Œ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์—ฌ์ „ํžˆ ์œ ์šฉํ•˜์ง€๋งŒ ์ œํ•œ๋œ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋‘ ๋ฒˆ์งธ ์„ธ๋Œ€ ๋ฐ์ดํ„ฐ์˜ ์œ ์šฉ์„ฑ์€ ํ˜„์ €ํžˆ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ˜„์ƒ์€ ์˜ˆ์ƒ๋˜๋Š” ์Šค์ผ€์ผ๋ง ๋ฒ•์น™ ์†์‹ค๊ณผ ๋ชจ๋ธ ๋ถ•๊ดด๋ฅผ ํ™•์ฆํ•˜๋ฉฐ, ๋ชจ๋ธ ๋ถ•๊ดด๊ฐ€ ์—ฌ๊ธฐ์„œ ๋”์šฑ ๋‘๋“œ๋Ÿฌ์ง€๊ฒŒ ๋‚˜ํƒ€๋‚˜ ์ฐจ์„ธ๋Œ€ LLM์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์žˆ์–ด ์–ด๋ ค์›€์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ๊ณผ ๊ฒฐ๊ณผ๋Š” ๋ถ€๋ก H์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.