question
stringlengths
101
4.47k
chosen
stringlengths
3
4.25k
rejected
stringlengths
39
9.29k
This paper proposes the AIOS architecture, demonstrating the potential to facilitate the development and deployment of LLM-based agents, fostering a more cohesive, effective and efficient AIOS-Agent ecosystem. The insights and methodologies presented herein contribute to the ongoing discourse in both AI and system research, offering a viable solution to the integration challenges posed by the diverse landscape of AI Agents. Diverse future work can be built upon this foundation, exploring innovative ways to refine and expand the AIOS architecture to meet the evolving needs of developing and deploying LLM agents.
์ด ๋…ผ๋ฌธ์€ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ๊ฐœ๋ฐœ ๋ฐ ๋ฐฐํฌ๋ฅผ ์ด‰์ง„ํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋Š” AIOS ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋” ํ†ตํ•ฉ์ ์ด๊ณ  ํšจ๊ณผ์ ์ด๋ฉฐ ํšจ์œจ์ ์ธ AIOS-Agent ์ƒํƒœ๊ณ„๋ฅผ ์กฐ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์ œ์‹œ๋œ ํ†ต์ฐฐ๋ ฅ๊ณผ ๋ฐฉ๋ฒ•๋ก ์€ AI ๋ฐ ์‹œ์Šคํ…œ ์—ฐ๊ตฌ ๋ชจ๋‘์—์„œ ์ง„ํ–‰ ์ค‘์ธ ๋…ผ์˜์— ๊ธฐ์—ฌํ•˜๋ฉฐ, ๋‹ค์–‘ํ•œ AI ์—์ด์ „ํŠธ์˜ ํ†ตํ•ฉ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์‹ค์งˆ์ ์ธ ํ•ด๊ฒฐ์ฑ…์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ธฐ์ดˆ ์œ„์— ๋‹ค์–‘ํ•œ ๋ฏธ๋ž˜ ์—ฐ๊ตฌ๊ฐ€ ๊ตฌ์ถ•๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, LLM ์—์ด์ „ํŠธ์˜ ๊ฐœ๋ฐœ ๋ฐ ๋ฐฐํฌ์˜ ์ง„ํ™”ํ•˜๋Š” ์š”๊ตฌ๋ฅผ ์ถฉ์กฑํ•˜๊ธฐ ์œ„ํ•ด AIOS ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ •์ œํ•˜๊ณ  ํ™•์žฅํ•˜๋Š” ํ˜์‹ ์ ์ธ ๋ฐฉ๋ฒ•์„ ํƒ๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์—์„œ๋Š” LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ๊ฐœ๋ฐœ ๋ฐ ๋ฐฐํฌ๋ฅผ ์ด‰์ง„ํ•˜์—ฌ ๋ณด๋‹ค ์‘์ง‘๋ ฅ ์žˆ๊ณ  ํšจ๊ณผ์ ์ด๋ฉฐ ํšจ์œจ์ ์ธ AIOS ์—์ด์ „ํŠธ ์—์ฝ”์‹œ์Šคํ…œ์„ ์กฐ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋Š” AIOS ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์— ์ œ์‹œ๋œ ์ธ์‚ฌ์ดํŠธ์™€ ๋ฐฉ๋ฒ•๋ก ์€ AI ๋ฐ ์‹œ์Šคํ…œ ์—ฐ๊ตฌ ๋ถ„์•ผ์˜ ์ง€์†์ ์ธ ๋…ผ์˜์— ๊ธฐ์—ฌํ•˜๋ฉฐ, ๋‹ค์–‘ํ•œ AI ์—์ด์ „ํŠธ ํ™˜๊ฒฝ์ด ์ œ๊ธฐํ•˜๋Š” ํ†ตํ•ฉ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ํ–ฅํ›„ ์—ฐ๊ตฌ๋Š” ์ด ๊ธฐ๋ฐ˜์„ ๋ฐ”ํƒ•์œผ๋กœ LLM ์—์ด์ „ํŠธ ๊ฐœ๋ฐœ ๋ฐ ๋ฐฐํฌ์˜ ์ง„ํ™”ํ•˜๋Š” ์š”๊ตฌ ์‚ฌํ•ญ์„ ์ถฉ์กฑํ•˜๊ธฐ ์œ„ํ•ด AIOS ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ฐœ์„ ํ•˜๊ณ  ํ™•์žฅํ•˜๋Š” ํ˜์‹ ์ ์ธ ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Metrics.** Considering the limitations of 2-billion parameter models, our evaluation criteria go beyond just the accuracy of responses. We additionally consider factors such as usefulness, relevance, accuracy, depth, creativity, and the level of detail in the model's answers. This comprehensive method allows for a detailed evaluation of the model's response quality. Specifically, We use GPT-4 (Achiam et al., 2023) to score responses from tested LLMs in specific problem contexts, with the scoring prompt available in the Appendix. C.2. We translate the score assignment prompt template from Zheng et al. (2024).
**ํ‰๊ฐ€์ง€ํ‘œ.** 20์–ต ๊ฐœ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ์˜ ํ•œ๊ณ„๋ฅผ ๊ณ ๋ คํ•˜์—ฌ, ์šฐ๋ฆฌ์˜ ํ‰๊ฐ€ ๊ธฐ์ค€์€ ๋‹จ์ˆœํžˆ ์‘๋‹ต์˜ ์ •ํ™•์„ฑ์— ๊ตญํ•œ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์œ ์šฉ์„ฑ, ๊ด€๋ จ์„ฑ, ์ •ํ™•์„ฑ, ๊นŠ์ด, ์ฐฝ์˜์„ฑ, ๊ทธ๋ฆฌ๊ณ  ๋ชจ๋ธ ์‘๋‹ต์˜ ์„ธ๋ถ€ ์ˆ˜์ค€๊ณผ ๊ฐ™์€ ์š”์†Œ๋“ค๋„ ์ถ”๊ฐ€๋กœ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด ํฌ๊ด„์ ์ธ ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ ์‘๋‹ต์˜ ํ’ˆ์งˆ์„ ์ƒ์„ธํžˆ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ํŠน์ • ๋ฌธ์ œ ์ƒํ™ฉ์—์„œ ํ…Œ์ŠคํŠธ๋œ LLM์˜ ์‘๋‹ต์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด GPT-4 (Achiam et al., 2023)๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ํ‰๊ฐ€ ํ”„๋กฌํ”„ํŠธ๋Š” ๋ถ€๋ก C.2์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” Zheng et al. (2024)์˜ ์ ์ˆ˜ ํ• ๋‹น ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์„ ๋ฒˆ์—ญํ•˜์—ฌ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
**๋ฉ”ํŠธ๋ฆญ.** 20์–ต ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์˜ ํ•œ๊ณ„๋ฅผ ๊ณ ๋ คํ•˜์—ฌ, ์ €ํฌ์˜ ํ‰๊ฐ€ ๊ธฐ์ค€์€ ์‘๋‹ต์˜ ์ •ํ™•์„ฑ ๊ทธ ์ด์ƒ์ž…๋‹ˆ๋‹ค. ์œ ์šฉ์„ฑ, ๊ด€๋ จ์„ฑ, ์ •ํ™•์„ฑ, ๊นŠ์ด, ์ฐฝ์˜์„ฑ, ๋ชจ๋ธ ๋‹ต๋ณ€์˜ ์„ธ๋ถ€ ์ˆ˜์ค€ ๋“ฑ์˜ ์š”์†Œ๋„ ์ถ”๊ฐ€๋กœ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ข…ํ•ฉ์ ์ธ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ๋ชจ๋ธ์˜ ์‘๋‹ต ํ’ˆ์งˆ์„ ์ƒ์„ธํ•˜๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ํŠน์ • ๋ฌธ์ œ ๋งฅ๋ฝ์—์„œ ํ…Œ์ŠคํŠธ๋œ LLM์˜ ์‘๋‹ต์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด GPT-4(Achiam et al., 2023)๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ์ฑ„์  ํ”„๋กฌํ”„ํŠธ๋Š” ๋ถ€๋ก C.2์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. C.2. ์ ์ˆ˜ ํ• ๋‹น ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์€ Zheng et al.(2024)์—์„œ ๋ฒˆ์—ญํ–ˆ์Šต๋‹ˆ๋‹ค.
Despite focusing on different aspects of examination-style questions and encompassing both English and Chinese languages, MMLU, CMMLU, and CEVAL appear to assess overlapping capabilities of the models, leading to similar performance trends. Furthermore, the training process of models demonstrates that the behavior of known tasks within a domain can predict the behavior of yet-to-be-encountered tasks. This suggests that understanding a model's performance on familiar tasks can offer valuable insights into its potential performance on analogous tasks that have not been previously explored within the same domain.
์‹œํ—˜ ์Šคํƒ€์ผ ์งˆ๋ฌธ์˜ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์˜์–ด์™€ ์ค‘๊ตญ์–ด๋ฅผ ๋ชจ๋‘ ํฌํ•จํ•˜๊ณ  ์žˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , MMLU, CMMLU, CEVAL์€ ๋ชจ๋ธ์˜ ์ค‘๋ณต๋˜๋Š” ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ด๋ฉฐ, ์ด๋Š” ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ ๊ฒฝํ–ฅ์„ ์ดˆ๋ž˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ๊ณผ์ •์€ ๋„๋ฉ”์ธ ๋‚ด์—์„œ ์•Œ๋ ค์ง„ ์ž‘์—…์˜ ํ–‰๋™์ด ์•„์ง ์ ‘ํ•˜์ง€ ์•Š์€ ์ž‘์—…์˜ ํ–‰๋™์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Š” ๋ชจ๋ธ์ด ์ต์ˆ™ํ•œ ์ž‘์—…์—์„œ์˜ ์„ฑ๋Šฅ์„ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์ด ๋™์ผํ•œ ๋„๋ฉ”์ธ ๋‚ด์—์„œ ์ด์ „์— ํƒ๊ตฌ๋˜์ง€ ์•Š์€ ์œ ์‚ฌํ•œ ์ž‘์—…์—์„œ์˜ ์ž ์žฌ์  ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ๊ท€์ค‘ํ•œ ํ†ต์ฐฐ๋ ฅ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
์‹œํ—˜ ์Šคํƒ€์ผ์˜ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์˜์–ด์™€ ์ค‘๊ตญ์–ด๋ฅผ ๋ชจ๋‘ ํฌ๊ด„ํ•˜์ง€๋งŒ, MMLU, CMMLU ๋ฐ CEVAL์€ ๋ชจ๋ธ์˜ ์ค‘๋ณต๋œ ๊ธฐ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ด๋ฉฐ, ์ด๋กœ ์ธํ•ด ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ ์ถ”์„ธ๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋ธ์˜ ํ•™์Šต ๊ณผ์ •์„ ํ†ตํ•ด ํ•œ ๋„๋ฉ”์ธ ๋‚ด์—์„œ ์•Œ๋ ค์ง„ ์ž‘์—…์˜ ํ–‰๋™์ด ์•„์ง ์ ‘ํ•˜์ง€ ์•Š์€ ์ž‘์—…์˜ ํ–‰๋™์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์นœ์ˆ™ํ•œ ์ž‘์—…์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์ดํ•ดํ•˜๋ฉด ๋™์ผํ•œ ๋„๋ฉ”์ธ ๋‚ด์—์„œ ์ด์ „์— ํƒ์ƒ‰ํ•˜์ง€ ์•Š์€ ์œ ์‚ฌํ•œ ์ž‘์—…์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์œ ์šฉํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
Related work.Two works published during the responsible disclosure period use a similar procedure, and deal with numerical issues in different ways. (Chiu, 2024) start with a low \(B\) for the whole vocabulary, then increase \(B\) and ask for all tokens that haven't appeared before, and repeat until all tokens are covered. (Hayase et al., 2024) use the method in Appendix E.1, and set \(B=-\hat{z}_{i}\), where \(\hat{z}_{i}\) is an estimate of \(z_{i}\) inherent to their application. It is possible variants of this method have been discussed before our or these works, but we are not aware of further references.
๊ด€๋ จ ์—ฐ๊ตฌ. ์ฑ…์ž„ ์žˆ๋Š” ๊ณต๊ฐœ ๊ธฐ๊ฐ„ ๋™์•ˆ ๋ฐœํ‘œ๋œ ๋‘ ์—ฐ๊ตฌ๋Š” ์œ ์‚ฌํ•œ ์ ˆ์ฐจ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ์„œ๋กœ ๋‹ค๋ฅธ ๋ฐฉ์‹์œผ๋กœ ์ˆ˜์น˜์  ๋ฌธ์ œ๋ฅผ ๋‹ค๋ฃน๋‹ˆ๋‹ค. (Chiu, 2024)๋Š” ์ „์ฒด ์–ดํœ˜์— ๋Œ€ํ•ด ๋‚ฎ์€ \(B\)๋กœ ์‹œ์ž‘ํ•œ ํ›„, \(B\)๋ฅผ ์ฆ๊ฐ€์‹œํ‚ค๊ณ  ์ด์ „์— ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์€ ๋ชจ๋“  ํ† ํฐ์„ ์š”์ฒญํ•˜๋ฉฐ, ๋ชจ๋“  ํ† ํฐ์ด ํฌํ•จ๋  ๋•Œ๊นŒ์ง€ ์ด๋ฅผ ๋ฐ˜๋ณตํ•ฉ๋‹ˆ๋‹ค. (Hayase et al., 2024)๋Š” ๋ถ€๋ก E.1์— ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜๋ฉฐ, \(B=-\hat{z}_{i}\)๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ \(\hat{z}_{i}\)๋Š” ๊ทธ๋“ค์˜ ์‘์šฉ์— ๋‚ด์žฌ๋œ \(z_{i}\)์˜ ์ถ”์ •์น˜์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์˜ ๋ณ€ํ˜•์ด ์šฐ๋ฆฌ ์—ฐ๊ตฌ๋‚˜ ์ด๋“ค ์—ฐ๊ตฌ ์ด์ „์— ๋…ผ์˜๋˜์—ˆ์„ ๊ฐ€๋Šฅ์„ฑ์€ ์žˆ์ง€๋งŒ, ์ถ”๊ฐ€์ ์ธ ์ฐธ๊ณ  ๋ฌธํ—Œ์€ ์•Œ์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค.
๊ด€๋ จ ์ž‘์—….์ฑ…์ž„ ๊ณต๊ฐœ ๊ธฐ๊ฐ„ ๋™์•ˆ ๋ฐœํ‘œ๋œ ๋‘ ๊ฐœ์˜ ์ž‘์—…์€ ์œ ์‚ฌํ•œ ์ ˆ์ฐจ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ์ˆ˜์น˜ ๋ฌธ์ œ๋ฅผ ๋‹ค๋ฅด๊ฒŒ ๋‹ค๋ฃน๋‹ˆ๋‹ค. (Chiu, 2024)๋Š” ์ „์ฒด ์–ดํœ˜์— ๋Œ€ํ•ด ๋‚ฎ์€ \(B\)๋กœ ์‹œ์ž‘ํ•˜์—ฌ \(B\)๋ฅผ ์ฆ๊ฐ€์‹œํ‚ค๊ณ  ์ด์ „์— ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์€ ๋ชจ๋“  ํ† ํฐ์„ ์š”์ฒญํ•œ ๋‹ค์Œ ๋ชจ๋“  ํ† ํฐ์ด ํฌํ•จ๋  ๋•Œ๊นŒ์ง€ ๋ฐ˜๋ณตํ•ฉ๋‹ˆ๋‹ค. (Hayase et al., 2024)๋Š” ๋ถ€๋ก E.1์˜ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ \(B=-\hat{z}_{i}\), ์—ฌ๊ธฐ์„œ \(\hat{z}_{i}\)๋Š” ํ•ด๋‹น ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋‚ด์žฌ๋œ \(z_{i}\)์˜ ์ถ”์ •์น˜์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์˜ ๋ณ€ํ˜•์ด ์ €ํฌ๋‚˜ ๋‹ค๋ฅธ ์—ฐ๊ตฌ์ž๋“ค์— ์˜ํ•ด ๋…ผ์˜๋˜์—ˆ์„ ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ์ง€๋งŒ, ์ถ”๊ฐ€์ ์ธ ์ฐธ๊ณ  ์ž๋ฃŒ๋Š” ์—†์Šต๋‹ˆ๋‹ค.
Both our hidden-dimension-stealing and entire-layer-stealing attack worked for all five of these models. The size we recover from the model perfectly matches the actual size of the original model, as confirmed by OpenAI. For the first three models, we report in Table 4 the size we recover because (1) the sizes of these models was never previously confirmed, but (2) they have now been deprecated and so disclosing the size is not harmful. In discussions with OpenAI, we decided to withhold disclosure of the size of gpt-3.5-turbo models, but we confirmed with them that the number our attack reported was accurate.
์šฐ๋ฆฌ์˜ hidden-dimension-stealing ๋ฐ entire-layer-stealing ๊ณต๊ฒฉ์€ ์ด ๋‹ค์„ฏ ๊ฐ€์ง€ ๋ชจ๋ธ ๋ชจ๋‘์—์„œ ์ž‘๋™ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ์—์„œ ๋ณต๊ตฌํ•œ ํฌ๊ธฐ๋Š” OpenAI์— ์˜ํ•ด ํ™•์ธ๋œ ๋ฐ”์™€ ๊ฐ™์ด ์›๋ž˜ ๋ชจ๋ธ์˜ ์‹ค์ œ ํฌ๊ธฐ์™€ ์™„๋ฒฝํ•˜๊ฒŒ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ์„ธ ๋ชจ๋ธ์— ๋Œ€ํ•ด์„œ๋Š”, (1) ์ด ๋ชจ๋ธ๋“ค์˜ ํฌ๊ธฐ๊ฐ€ ์ด์ „์— ํ™•์ธ๋œ ์ ์ด ์—†์—ˆ๊ณ , (2) ์ด์ œ๋Š” ์‚ฌ์šฉ ์ค‘๋‹จ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ํฌ๊ธฐ๋ฅผ ๊ณต๊ฐœํ•˜๋Š” ๊ฒƒ์ด ํ•ด๋กญ์ง€ ์•Š๋‹ค๋Š” ์ด์œ ๋กœ Table 4์— ๋ณต๊ตฌํ•œ ํฌ๊ธฐ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. OpenAI์™€์˜ ๋…ผ์˜์—์„œ, ์šฐ๋ฆฌ๋Š” gpt-3.5-turbo ๋ชจ๋ธ์˜ ํฌ๊ธฐ ๊ณต๊ฐœ๋ฅผ ๋ณด๋ฅ˜ํ•˜๊ธฐ๋กœ ๊ฒฐ์ •ํ–ˆ์ง€๋งŒ, ์šฐ๋ฆฌ์˜ ๊ณต๊ฒฉ์ด ๋ณด๊ณ ํ•œ ์ˆซ์ž๊ฐ€ ์ •ํ™•ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๊ทธ๋“ค๊ณผ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค.
์ด ๋‹ค์„ฏ ๊ฐ€์ง€ ๋ชจ๋ธ ๋ชจ๋‘์— ๋Œ€ํ•ด ์ˆจ๊ฒจ์ง„ ์ฐจ์› ํ›”์น˜๊ธฐ์™€ ์ „์ฒด ๋ ˆ์ด์–ด ํ›”์น˜๊ธฐ ๊ณต๊ฒฉ์ด ์„ฑ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ์—์„œ ๋ณต๊ตฌํ•œ ํฌ๊ธฐ๋Š” OpenAI๊ฐ€ ํ™•์ธํ•œ ๋ฐ”์™€ ๊ฐ™์ด ์›๋ž˜ ๋ชจ๋ธ์˜ ์‹ค์ œ ํฌ๊ธฐ์™€ ์™„๋ฒฝํ•˜๊ฒŒ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ์ฒ˜์Œ ์„ธ ๊ฐ€์ง€ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ (1) ์ด ๋ชจ๋ธ์˜ ํฌ๊ธฐ๊ฐ€ ์ด์ „์— ํ™•์ธ๋œ ์ ์ด ์—†์ง€๋งŒ (2) ์ด์ œ ๋” ์ด์ƒ ์‚ฌ์šฉ๋˜์ง€ ์•Š์œผ๋ฏ€๋กœ ํฌ๊ธฐ๋ฅผ ๊ณต๊ฐœํ•ด๋„ ๋ฌธ์ œ๊ฐ€ ์—†๊ธฐ ๋•Œ๋ฌธ์— ๋ณต๊ตฌํ•œ ํฌ๊ธฐ๋ฅผ ํ‘œ 4์— ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. OpenAI์™€์˜ ๋…ผ์˜์—์„œ gpt-3.5-turbo ๋ชจ๋ธ์˜ ํฌ๊ธฐ ๊ณต๊ฐœ๋ฅผ ๋ณด๋ฅ˜ํ•˜๊ธฐ๋กœ ๊ฒฐ์ •ํ–ˆ์ง€๋งŒ, ๊ณต๊ฒฉ์ด ๋ณด๊ณ ํ•œ ์ˆซ์ž๊ฐ€ ์ •ํ™•ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค.
Model merging holds significant potential and democratizes the model-building process, making it accessible to a broad range of participants. However, it heavily relies on intuition and domain knowledge. Human intuition, however, has its limits. With the growing diversity of open models and tasks, we need a more systematic approach. We believe evolutionary algorithms, inspired by natural selection, can unlock more effective merging solutions. These algorithms can explore a vast space of possibilities, discovering novel and counter-intuitive combinations that traditional methods and human intuition might miss.
๋ชจ๋ธ ๋ณ‘ํ•ฉ์€ ์ƒ๋‹นํ•œ ์ž ์žฌ๋ ฅ์„ ์ง€๋‹ˆ๊ณ  ์žˆ์œผ๋ฉฐ, ๋ชจ๋ธ ๊ตฌ์ถ• ๊ณผ์ •์„ ๋ฏผ์ฃผํ™”ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ฐธ๊ฐ€์ž๋“ค์ด ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Š” ์ง๊ด€๊ณผ ๋„๋ฉ”์ธ ์ง€์‹์— ํฌ๊ฒŒ ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ธ๊ฐ„์˜ ์ง๊ด€์—๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐœ๋ฐฉํ˜• ๋ชจ๋ธ๊ณผ ์ž‘์—…์˜ ๋‹ค์–‘์„ฑ์ด ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ, ์šฐ๋ฆฌ๋Š” ๋ณด๋‹ค ์ฒด๊ณ„์ ์ธ ์ ‘๊ทผ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ž์—ฐ ์„ ํƒ์—์„œ ์˜๊ฐ์„ ๋ฐ›์€ ์ง„ํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ๋” ํšจ๊ณผ์ ์ธ ๋ณ‘ํ•ฉ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๊ด‘๋ฒ”์œ„ํ•œ ๊ฐ€๋Šฅ์„ฑ์˜ ๊ณต๊ฐ„์„ ํƒ์ƒ‰ํ•˜์—ฌ ์ „ํ†ต์ ์ธ ๋ฐฉ๋ฒ•๊ณผ ์ธ๊ฐ„์˜ ์ง๊ด€์ด ๋†“์น  ์ˆ˜ ์žˆ๋Š” ์ƒˆ๋กœ์šด ๋ฐ ๋ฐ˜์ง๊ด€์ ์ธ ์กฐํ•ฉ์„ ๋ฐœ๊ฒฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ชจ๋ธ ๋ณ‘ํ•ฉ์€ ์ƒ๋‹นํ•œ ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ ๋ชจ๋ธ ๊ตฌ์ถ• ํ”„๋กœ์„ธ์Šค๋ฅผ ๋ฏผ์ฃผํ™”ํ•˜์—ฌ ๊ด‘๋ฒ”์œ„ํ•œ ์ฐธ์—ฌ์ž๊ฐ€ ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ง๊ด€๊ณผ ๋„๋ฉ”์ธ ์ง€์‹์— ํฌ๊ฒŒ ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ธ๊ฐ„์˜ ์ง๊ด€์—๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜คํ”ˆ ๋ชจ๋ธ๊ณผ ํƒœ์Šคํฌ์˜ ๋‹ค์–‘์„ฑ์ด ์ปค์ง€๋ฉด์„œ ๋ณด๋‹ค ์ฒด๊ณ„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ž์—ฐ ์„ ํƒ์—์„œ ์˜๊ฐ์„ ์–ป์€ ์ง„ํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ๋ณด๋‹ค ํšจ๊ณผ์ ์ธ ๋ณ‘ํ•ฉ ์†”๋ฃจ์…˜์„ ์ž ๊ธˆ ํ•ด์ œํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๋ฐฉ๋Œ€ํ•œ ๊ฐ€๋Šฅ์„ฑ์˜ ๊ณต๊ฐ„์„ ํƒ์ƒ‰ํ•˜์—ฌ ๊ธฐ์กด ๋ฐฉ๋ฒ•๊ณผ ์ธ๊ฐ„์˜ ์ง๊ด€์ด ๋†“์น  ์ˆ˜ ์žˆ๋Š” ์ƒˆ๋กญ๊ณ  ์ง๊ด€์ ์ด์ง€ ์•Š์€ ์กฐํ•ฉ์„ ๋ฐœ๊ฒฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Language models produce a distribution over the next token; can we use this to recover the prompt tokens? We consider the problem of language model inversion and show that next-token probabilities contain a surprising amount of information about the preceding text. Often we can recover the text in cases where it is hidden from the user, motivating a method for recovering unknown prompts given only the model's current distribution output. We consider a variety of model access scenarios, and show how even without predictions for every token in the vocabulary we can recover the probability vector through search.
์–ธ์–ด ๋ชจ๋ธ์€ ๋‹ค์Œ ํ† ํฐ์— ๋Œ€ํ•œ ๋ถ„ํฌ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ ํ† ํฐ์„ ๋ณต์›ํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ์šฐ๋ฆฌ๋Š” ์–ธ์–ด ๋ชจ๋ธ ์—ญ์ „ ๋ฌธ์ œ๋ฅผ ๊ณ ๋ คํ•˜๊ณ , ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์ด ์ด์ „ ํ…์ŠคํŠธ์— ๋Œ€ํ•œ ๋†€๋ผ์šด ์–‘์˜ ์ •๋ณด๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ข…์ข… ์‚ฌ์šฉ์ž๊ฐ€ ํ…์ŠคํŠธ๋ฅผ ๋ณผ ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ์—๋„ ํ…์ŠคํŠธ๋ฅผ ๋ณต์›ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ์˜ ํ˜„์žฌ ๋ถ„ํฌ ์ถœ๋ ฅ๋งŒ์œผ๋กœ ์•Œ๋ ค์ง€์ง€ ์•Š์€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณต์›ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์ ‘๊ทผ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๊ณ ๋ คํ•˜๊ณ , ์–ดํœ˜์˜ ๋ชจ๋“  ํ† ํฐ์— ๋Œ€ํ•œ ์˜ˆ์ธก์ด ์—†์–ด๋„ ๊ฒ€์ƒ‰์„ ํ†ตํ•ด ํ™•๋ฅ  ๋ฒกํ„ฐ๋ฅผ ๋ณต์›ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์–ธ์–ด ๋ชจ๋ธ์€ ๋‹ค์Œ ํ† ํฐ์— ๋Œ€ํ•œ ๋ถ„ํฌ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ ํ† ํฐ์„ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ์–ธ์–ด ๋ชจ๋ธ ๋ฐ˜์ „ ๋ฌธ์ œ๋ฅผ ๊ณ ๋ คํ•˜๊ณ  ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์— ์ด์ „ ํ…์ŠคํŠธ์— ๋Œ€ํ•œ ๋†€๋ผ์šด ์–‘์˜ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ข…์ข… ์‚ฌ์šฉ์ž์—๊ฒŒ ์ˆจ๊ฒจ์ ธ ์žˆ๋Š” ๊ฒฝ์šฐ์—๋„ ํ…์ŠคํŠธ๋ฅผ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ์˜ ํ˜„์žฌ ๋ถ„ํฌ ์ถœ๋ ฅ๋งŒ ์ฃผ์–ด์กŒ์„ ๋•Œ ์•Œ ์ˆ˜ ์—†๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณต๊ตฌํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋™๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์•ก์„ธ์Šค ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๊ณ ๋ คํ•˜๊ณ  ์–ดํœ˜์˜ ๋ชจ๋“  ํ† ํฐ์— ๋Œ€ํ•œ ์˜ˆ์ธก์ด ์—†๋”๋ผ๋„ ๊ฒ€์ƒ‰์„ ํ†ตํ•ด ํ™•๋ฅ  ๋ฒกํ„ฐ๋ฅผ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
* **JA-VG-VQA-500:** A 500-sample test set extracted from the Japanese Visual Genome VQA dataset [42]. * **JA-VLM-Bench-In-the-Wild:** A Japanese version of LLaVA-Bench-In-the-Wild [32]. We compiled a rich collection of 42 images, accompanied by a total of 50 questions, featuring a variety of Japanese cultural elements and objects found in Japan. The QAs were crafted with the assistance of GPT-4V [1] and underwent a human-in-the-loop filtering process to eliminate nonsensical outcomes. Compared to the JA-VG-VQA-500 dataset, our set poses more complex challenges, demanding more nuanced and detailed responses.
* **JA-VG-VQA-500:** ์ผ๋ณธ์–ด Visual Genome VQA ๋ฐ์ดํ„ฐ์…‹ [42]์—์„œ ์ถ”์ถœํ•œ 500๊ฐœ์˜ ์ƒ˜ํ”Œ ํ…Œ์ŠคํŠธ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. * **JA-VLM-Bench-In-the-Wild:** LLaVA-Bench-In-the-Wild [32]์˜ ์ผ๋ณธ์–ด ๋ฒ„์ „์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ผ๋ณธ์˜ ๋‹ค์–‘ํ•œ ๋ฌธํ™” ์š”์†Œ์™€ ๋ฌผ์ฒด๋ฅผ ํŠน์ง•์œผ๋กœ ํ•˜๋Š” ์ด 50๊ฐœ์˜ ์งˆ๋ฌธ๊ณผ ํ•จ๊ป˜ 42๊ฐœ์˜ ์ด๋ฏธ์ง€๋ฅผ ํ’๋ถ€ํ•˜๊ฒŒ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. QAs๋Š” GPT-4V [1]์˜ ๋„์›€์„ ๋ฐ›์•„ ์ œ์ž‘๋˜์—ˆ์œผ๋ฉฐ, ๋น„ํ•ฉ๋ฆฌ์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„์ด ์ฐธ์—ฌํ•˜๋Š” ํ•„ํ„ฐ๋ง ๊ณผ์ •์„ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค. JA-VG-VQA-500 ๋ฐ์ดํ„ฐ์…‹๊ณผ ๋น„๊ตํ•  ๋•Œ, ์šฐ๋ฆฌ์˜ ์„ธํŠธ๋Š” ๋” ๋ณต์žกํ•œ ๋„์ „์„ ์ œ์‹œํ•˜๋ฉฐ, ๋” ๋ฏธ๋ฌ˜ํ•˜๊ณ  ์ƒ์„ธํ•œ ์‘๋‹ต์„ ์š”๊ตฌํ•ฉ๋‹ˆ๋‹ค.
* **JA-VG-VQA-500:** ์ผ๋ณธ ๋น„์ฃผ์–ผ ์ง€๋…ธ๋ฉ” VQA ๋ฐ์ดํ„ฐ ์„ธํŠธ[42]์—์„œ ์ถ”์ถœํ•œ 500๊ฐœ ์ƒ˜ํ”Œ ํ…Œ์ŠคํŠธ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. * **JA-VLM-Bench-In-the-Wild:** LLaVA-Bench-In-the-Wild [32]์˜ ์ผ๋ณธ์–ด ๋ฒ„์ „์ž…๋‹ˆ๋‹ค. ์ผ๋ณธ์—์„œ ๋ฐœ๊ฒฌ๋˜๋Š” ๋‹ค์–‘ํ•œ ์ผ๋ณธ ๋ฌธํ™” ์š”์†Œ์™€ ์‚ฌ๋ฌผ์ด ๋“ฑ์žฅํ•˜๋Š” ์ด 50๊ฐœ์˜ ์งˆ๋ฌธ์ด ํฌํ•จ๋œ 42๊ฐœ์˜ ์ด๋ฏธ์ง€๋ฅผ ๋ชจ์•˜์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์€ GPT-4V [1]์˜ ๋„์›€์„ ๋ฐ›์•„ ์ž‘์„ฑ๋˜์—ˆ์œผ๋ฉฐ, ๋ง๋„ ์•ˆ ๋˜๋Š” ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„์ด ์ฐธ์—ฌํ•˜๋Š” ํ•„ํ„ฐ๋ง ๊ณผ์ •์„ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค. JA-VG-VQA-500 ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋น„ํ•ด ๋” ๋ณต์žกํ•œ ๊ณผ์ œ๋ฅผ ์ œ์‹œํ•˜์—ฌ ๋ณด๋‹ค ๋ฏธ๋ฌ˜ํ•˜๊ณ  ์„ธ๋ฐ€ํ•œ ๋‹ต๋ณ€์„ ์š”๊ตฌํ•ฉ๋‹ˆ๋‹ค.
Examining the parameters individually, \(E\) and \(\beta\) show highly significant differences, with \(p\)-values of \(1.5\times 10^{-6}\) and \(4.3\times 10^{-5}\), respectively. These small \(p\)-values indicate that the estimated values for \(E\) and \(\beta\) are significantly different from their respective true values, suggesting substantial deviations in these parameters' estimates from expected outcomes based on the model. On the other hand, the parameters \(A\), \(B\), and \(\alpha\) have estimated values that are not statistically significantly different from the values reported in Hoffmann et al.
๊ฐœ๋ณ„์ ์œผ๋กœ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ๊ฒ€ํ† ํ•œ ๊ฒฐ๊ณผ, \(E\)์™€ \(\beta\)๋Š” ๊ฐ๊ฐ \(p\)-๊ฐ’์ด \(1.5\times 10^{-6}\) ๋ฐ \(4.3\times 10^{-5}\)๋กœ ๋งค์šฐ ์œ ์˜๋ฏธํ•œ ์ฐจ์ด๋ฅผ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์€ \(p\)-๊ฐ’์€ \(E\)์™€ \(\beta\)์˜ ์ถ”์ •๊ฐ’์ด ๊ฐ๊ฐ์˜ ์‹ค์ œ ๊ฐ’๊ณผ ์œ ์˜๋ฏธํ•˜๊ฒŒ ๋‹ค๋ฅด๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ์— ๊ธฐ๋ฐ˜ํ•œ ์˜ˆ์ƒ ๊ฒฐ๊ณผ์™€ ๋น„๊ตํ–ˆ์„ ๋•Œ ์ด๋“ค ๋งค๊ฐœ๋ณ€์ˆ˜์˜ ์ถ”์ •์น˜์— ์ƒ๋‹นํ•œ ํŽธ์ฐจ๊ฐ€ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, \(A\), \(B\), \(\alpha\) ๋งค๊ฐœ๋ณ€์ˆ˜์˜ ์ถ”์ •๊ฐ’์€ Hoffmann ๋“ฑ์—์„œ ๋ณด๊ณ ๋œ ๊ฐ’๊ณผ ํ†ต๊ณ„์ ์œผ๋กœ ์œ ์˜๋ฏธํ•˜๊ฒŒ ๋‹ค๋ฅด์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ๊ฐœ๋ณ„์ ์œผ๋กœ ์‚ดํŽด๋ณด๋ฉด, \(E\)์™€ \(\beta\)๋Š” ๊ฐ๊ฐ \(1.5\๋ฐฐ 10^{-6}\)์™€ \(4.3\๋ฐฐ 10^{-5}\)์˜ \(p\)-๊ฐ’์œผ๋กœ ๋งค์šฐ ์œ ์˜๋ฏธํ•œ ์ฐจ์ด๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์€ \(p\)-๊ฐ’์€ \(E\)์™€ \(\beta\)์˜ ์ถ”์ •๊ฐ’์ด ๊ฐ๊ฐ์˜ ์‹ค์ œ ๊ฐ’๊ณผ ์ƒ๋‹นํžˆ ๋‹ค๋ฅด๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ์— ๊ธฐ๋ฐ˜ํ•œ ์˜ˆ์ƒ ๊ฒฐ๊ณผ์—์„œ ์ด๋Ÿฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜์˜ ์ถ”์ •์น˜๊ฐ€ ์ƒ๋‹นํ•œ ํŽธ์ฐจ๋ฅผ ๋ณด์ž„์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ๋งค๊ฐœ๋ณ€์ˆ˜ \(A\), \(B\), \(\alpha\)์˜ ์ถ”์ •๊ฐ’์€ Hoffmann ๋“ฑ์—์„œ ๋ณด๊ณ ํ•œ ๊ฐ’๊ณผ ํ†ต๊ณ„์ ์œผ๋กœ ์œ ์˜๋ฏธํ•œ ์ฐจ์ด๊ฐ€ ์—†๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
Advances in long-context language models have sparked interest in summarizing book-length documents (\(>\)100K tokens). Despite the importance of faithfulness and content relevance for summary quality, recent work in this regime focuses only on input-agnostic aspects like coherence (Chang et al., 2023b). This is due to the length and complexity of the input documents: hiring human annotators to read and understand them is expensive and time-consuming. Our work fills this gap by presenting the first large-scale human evaluation of faithfulness and other content selection errors in book-length summarization.
๊ธด ๋ฌธ๋งฅ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ฐœ์ „์€ ์ฑ… ๊ธธ์ด์˜ ๋ฌธ์„œ(100K ํ† ํฐ ์ด์ƒ)๋ฅผ ์š”์•ฝํ•˜๋Š” ๋ฐ ๋Œ€ํ•œ ๊ด€์‹ฌ์„ ๋ถˆ๋Ÿฌ์ผ์œผ์ผฐ์Šต๋‹ˆ๋‹ค. ์š”์•ฝ ํ’ˆ์งˆ์— ์žˆ์–ด ์ถฉ์‹ค์„ฑ๊ณผ ๋‚ด์šฉ ๊ด€๋ จ์„ฑ์˜ ์ค‘์š”์„ฑ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ตœ๊ทผ์˜ ์—ฐ๊ตฌ๋Š” ์ผ๊ด€์„ฑ ๊ฐ™์€ ์ž…๋ ฅ๊ณผ ๋ฌด๊ด€ํ•œ ์ธก๋ฉด์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Chang et al., 2023b). ์ด๋Š” ์ž…๋ ฅ ๋ฌธ์„œ์˜ ๊ธธ์ด์™€ ๋ณต์žก์„ฑ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค: ์‚ฌ๋žŒ ํ‰๊ฐ€์ž๋ฅผ ๊ณ ์šฉํ•˜์—ฌ ์ด๋ฅผ ์ฝ๊ณ  ์ดํ•ดํ•˜๋Š” ๊ฒƒ์€ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ณ  ์‹œ๊ฐ„์ด ๋งŽ์ด ์†Œ์š”๋ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ์ฑ… ๊ธธ์ด ์š”์•ฝ์—์„œ ์ถฉ์‹ค์„ฑ๊ณผ ๊ธฐํƒ€ ๋‚ด์šฉ ์„ ํƒ ์˜ค๋ฅ˜์— ๋Œ€ํ•œ ์ตœ์ดˆ์˜ ๋Œ€๊ทœ๋ชจ ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ์ œ์‹œํ•จ์œผ๋กœ์จ ์ด ๊ฒฉ์ฐจ๋ฅผ ๋ฉ”์›๋‹ˆ๋‹ค.
๊ธด ๋ฌธ๋งฅ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ฐœ์ „์œผ๋กœ ์ฑ… ๊ธธ์ด์˜ ๋ฌธ์„œ(\(>\)100K ํ† ํฐ)๋ฅผ ์š”์•ฝํ•˜๋Š” ๊ฒƒ์— ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ๋†’์•„์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์š”์•ฝ ํ’ˆ์งˆ์— ์žˆ์–ด ์ถฉ์‹ค์„ฑ๊ณผ ์ฝ˜ํ…์ธ  ๊ด€๋ จ์„ฑ์ด ์ค‘์š”ํ•จ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ตœ๊ทผ ์ด ๋ถ„์•ผ์˜ ์—ฐ๊ตฌ๋Š” ์ผ๊ด€์„ฑ(Chang et al., 2023b)๊ณผ ๊ฐ™์€ ์ž…๋ ฅ์— ๊ตฌ์• ๋ฐ›์ง€ ์•Š๋Š” ์ธก๋ฉด์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ž…๋ ฅ ๋ฌธ์„œ์˜ ๊ธธ์ด์™€ ๋ณต์žก์„ฑ ๋•Œ๋ฌธ์ธ๋ฐ, ์ธ๊ฐ„ ์ฃผ์„๊ฐ€๊ฐ€ ์ด๋ฅผ ์ฝ๊ณ  ์ดํ•ดํ•˜๋„๋ก ๊ณ ์šฉํ•˜๋Š” ๊ฒƒ์€ ๋น„์šฉ๊ณผ ์‹œ๊ฐ„์ด ๋งŽ์ด ์†Œ์š”๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฉ์ฐจ๋ฅผ ๋ฉ”์šฐ๊ธฐ ์œ„ํ•ด ์ €์ž๋Š” ์ฑ… ๊ธธ์ด์˜ ์š”์•ฝ์—์„œ ์ถฉ์‹ค์„ฑ ๋ฐ ๊ธฐํƒ€ ์ฝ˜ํ…์ธ  ์„ ํƒ ์˜ค๋ฅ˜์— ๋Œ€ํ•œ ์ตœ์ดˆ์˜ ๋Œ€๊ทœ๋ชจ ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
We choose \(11000\) clusters following previous work [1] and we note that this choice sticks to the heuristic that the number of clusters should roughly be the square root of the number of total points being clustered. We also note that in initial experiments for data selection at the 125M OPT model scale, we did not find a significant effect of number of clusters on the performance of our data selection methods (see Figure A1) this finding agrees with Abbas et al. [1] who notice significant overlap between datasets selected by SemDeDup with different number of clusters (see Figure A2 in Abbas et al. [1]).
์ด์ „ ์—ฐ๊ตฌ [1]๋ฅผ ๋”ฐ๋ผ \(11000\)๊ฐœ์˜ ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ์„ ํƒํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ํด๋Ÿฌ์Šคํ„ฐ์˜ ์ˆ˜๊ฐ€ ํด๋Ÿฌ์Šคํ„ฐ๋ง๋˜๋Š” ์ด ์ ์˜ ์ˆ˜์˜ ๋Œ€๋žต ์ œ๊ณฑ๊ทผ์ด์–ด์•ผ ํ•œ๋‹ค๋Š” ํœด๋ฆฌ์Šคํ‹ฑ์— ๋ถ€ํ•ฉํ•จ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ 125M OPT ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ๋ฐ์ดํ„ฐ ์„ ํƒ์„ ์œ„ํ•œ ์ดˆ๊ธฐ ์‹คํ—˜์—์„œ ํด๋Ÿฌ์Šคํ„ฐ ์ˆ˜๊ฐ€ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ A1 ์ฐธ์กฐ). ์ด ๋ฐœ๊ฒฌ์€ SemDeDup์ด ๋‹ค๋ฅธ ํด๋Ÿฌ์Šคํ„ฐ ์ˆ˜๋กœ ์„ ํƒํ•œ ๋ฐ์ดํ„ฐ์…‹ ๊ฐ„์— ์ƒ๋‹นํ•œ ์ค‘๋ณต์ด ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ•œ Abbas et al. [1]์˜ ์—ฐ๊ตฌ์™€ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค(Abbas et al. [1]์˜ ๊ทธ๋ฆผ A2 ์ฐธ์กฐ).
์ด์ „ ์ž‘์—… [1]์— ๋”ฐ๋ผ \(11000\) ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ์„ ํƒํ•˜๊ณ , ํด๋Ÿฌ์Šคํ„ฐ ์ˆ˜๋Š” ํด๋Ÿฌ์Šคํ„ฐ๋ง๋˜๋Š” ์ด ์  ์ˆ˜์˜ ์ œ๊ณฑ๊ทผ์ด ๋˜์–ด์•ผ ํ•œ๋‹ค๋Š” ํœด๋ฆฌ์Šคํ‹ฑ์„ ๋”ฐ๋ฅด๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ 125M OPT ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ๋ฐ์ดํ„ฐ ์„ ํƒ์„ ์œ„ํ•œ ์ดˆ๊ธฐ ์‹คํ—˜์—์„œ ํด๋Ÿฌ์Šคํ„ฐ ์ˆ˜๊ฐ€ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์ด ํฌ์ง€ ์•Š๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ(๊ทธ๋ฆผ A1 ์ฐธ์กฐ), ์ด ๊ฒฐ๊ณผ๋Š” Abbas ๋“ฑ[1]์˜ ์—ฐ๊ตฌ์—์„œ๋„ SemDeDup์— ์˜ํ•ด ์„ ํƒ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐ„์— ํด๋Ÿฌ์Šคํ„ฐ ์ˆ˜๊ฐ€ ๋‹ค๋ฅด๋”๋ผ๋„ ์ƒ๋‹นํ•œ ์ค‘๋ณต์ด ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ•œ ๊ฒƒ๊ณผ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค(Abbas ๋“ฑ[1]์˜ ๊ทธ๋ฆผ A2 ์ฐธ์กฐ).
Due to the diversity of possibilities in human language, it is rare for the same idea to be expressed identically in multiple documents unless one expression is derived from the other, or both are quoting from a shared source. This observation motivates deduplicating exact substrings. We call our approach ExactSubstr. When two examples \(x_{i}\) and \(x_{j}\) share a sufficiently long substring (that is, a substring for which \(x_{i}^{a.a.+k}=x_{j}^{b.b+k}\)), that substring is removed from one of them. Based on statistical analyses (SSB), we select \(k=50\) tokens as the minimum matching substring length.
์ธ๊ฐ„ ์–ธ์–ด์˜ ๋‹ค์–‘์„ฑ ๋•Œ๋ฌธ์— ๋™์ผํ•œ ์•„์ด๋””์–ด๊ฐ€ ์—ฌ๋Ÿฌ ๋ฌธ์„œ์—์„œ ๋™์ผํ•˜๊ฒŒ ํ‘œํ˜„๋˜๋Š” ๊ฒฝ์šฐ๋Š” ๋“œ๋ญ…๋‹ˆ๋‹ค. ์ด๋Š” ํ•œ ํ‘œํ˜„์ด ๋‹ค๋ฅธ ํ‘œํ˜„์—์„œ ํŒŒ์ƒ๋˜์—ˆ๊ฑฐ๋‚˜, ๋‘ ํ‘œํ˜„ ๋ชจ๋‘ ๊ณตํ†ต๋œ ์ถœ์ฒ˜๋ฅผ ์ธ์šฉํ•˜๋Š” ๊ฒฝ์šฐ๋ฅผ ์ œ์™ธํ•˜๊ณ ๋Š” ๊ฑฐ์˜ ๋ฐœ์ƒํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์€ ์ •ํ™•ํ•œ ๋ถ€๋ถ„ ๋ฌธ์ž์—ด์„ ์ค‘๋ณต ์ œ๊ฑฐํ•˜๋Š” ๋™๊ธฐ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ExactSubstr๋ผ๊ณ  ๋ถ€๋ฆ…๋‹ˆ๋‹ค. ๋‘ ์˜ˆ์ œ \(x_{i}\)์™€ \(x_{j}\)๊ฐ€ ์ถฉ๋ถ„ํžˆ ๊ธด ๋ถ€๋ถ„ ๋ฌธ์ž์—ด์„ ๊ณต์œ ํ•  ๋•Œ (์ฆ‰, \(x_{i}^{a.a.+k}=x_{j}^{b.b+k}\)์ธ ๋ถ€๋ถ„ ๋ฌธ์ž์—ด), ๊ทธ ๋ถ€๋ถ„ ๋ฌธ์ž์—ด์€ ๋‘˜ ์ค‘ ํ•˜๋‚˜์—์„œ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. ํ†ต๊ณ„ ๋ถ„์„(SSB)์„ ๊ธฐ๋ฐ˜์œผ๋กœ, ์ตœ์†Œ ์ผ์น˜ ๋ถ€๋ถ„ ๋ฌธ์ž์—ด ๊ธธ์ด๋กœ \(k=50\) ํ† ํฐ์„ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค.
์ธ๊ฐ„ ์–ธ์–ด์˜ ๊ฐ€๋Šฅ์„ฑ์€ ๋งค์šฐ ๋‹ค์–‘ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํ•œ ํ‘œํ˜„์ด ๋‹ค๋ฅธ ํ‘œํ˜„์—์„œ ํŒŒ์ƒ๋˜์—ˆ๊ฑฐ๋‚˜ ๋‘ ํ‘œํ˜„์ด ๊ณต์œ  ์†Œ์Šค์—์„œ ์ธ์šฉํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ์•„๋‹ˆ๋ผ๋ฉด ๋™์ผํ•œ ์•„์ด๋””์–ด๊ฐ€ ์—ฌ๋Ÿฌ ๋ฌธ์„œ์—์„œ ๋™์ผํ•˜๊ฒŒ ํ‘œํ˜„๋˜๋Š” ๊ฒฝ์šฐ๋Š” ๋“œ๋ญ…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์€ ์ •ํ™•ํ•œ ํ•˜์œ„ ๋ฌธ์ž์—ด์„ ์ค‘๋ณต ์ œ๊ฑฐํ•˜๋Š” ๋™๊ธฐ๋ฅผ ๋ถ€์—ฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ExactSubstr์ด๋ผ๊ณ  ๋ถ€๋ฆ…๋‹ˆ๋‹ค. ๋‘ ์˜ˆ์ œ \(x_{i}\)์™€ \(x_{j}\)๊ฐ€ ์ถฉ๋ถ„ํžˆ ๊ธด ํ•˜์œ„ ๋ฌธ์ž์—ด(์ฆ‰, \(x_{i}^{a.a.+k}=x_{j}^{b.b+k}\))์„ ๊ณต์œ ํ•˜๋Š” ๊ฒฝ์šฐ, ๊ทธ ์ค‘ ํ•˜๋‚˜์—์„œ ํ•ด๋‹น ํ•˜์œ„ ๋ฌธ์ž์—ด์ด ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. ํ†ต๊ณ„ ๋ถ„์„(SSB)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ตœ์†Œ ์ผ์น˜ ํ•˜์œ„ ๋ฌธ์ž์—ด ๊ธธ์ด๋กœ \(k=50\) ํ† ํฐ์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.
We compare Nemotron-4 15B to existing multilingual language models: XGLM (Lin et al., 2022), mGPT (Shliazhko et al., 2022), and BLOOM (Scao et al., 2023). XGLM and mGPT are models specially trained to have improved multilingual ability by up-sampling the presence of non-English languages in the training data. In contrast, BLOOM, like Nemotron-4, is a general purpose language model that was trained on a combination of English, multilingual, and code data. In Table 7, we clearly see that Nemotron-4 achieves the best performance amongst all models - realizing almost a 12% improvement in the four-shot setting.
์šฐ๋ฆฌ๋Š” Nemotron-4 15B๋ฅผ ๊ธฐ์กด์˜ ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ๋“ค๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค: XGLM (Lin et al., 2022), mGPT (Shliazhko et al., 2022), ๊ทธ๋ฆฌ๊ณ  BLOOM (Scao et al., 2023). XGLM๊ณผ mGPT๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์—์„œ ๋น„์˜์–ด๊ถŒ ์–ธ์–ด์˜ ๋น„์ค‘์„ ๋†’์—ฌ ๋‹ค๊ตญ์–ด ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๋“ค์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์—, BLOOM์€ Nemotron-4์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์˜์–ด, ๋‹ค๊ตญ์–ด, ๊ทธ๋ฆฌ๊ณ  ์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋ฅผ ์กฐํ•ฉํ•˜์—ฌ ํ›ˆ๋ จ๋œ ๋ฒ”์šฉ ์–ธ์–ด ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. Table 7์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, Nemotron-4๋Š” ๋ชจ๋“  ๋ชจ๋ธ ์ค‘์—์„œ ์ตœ๊ณ ์˜ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋ฉฐ, ๋„ค ์ƒท ์„ค์ •์—์„œ ๊ฑฐ์˜ 12%์˜ ํ–ฅ์ƒ์„ ์ด๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค.
๋„ค๋ชจํŠธ๋ก -4 15B๋ฅผ ๊ธฐ์กด ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ์ธ XGLM(Lin et al., 2022), mGPT(Shliazhko et al., 2022), BLOOM(Scao et al., 2023)๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. XGLM๊ณผ mGPT๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๋น„์˜์–ด ์–ธ์–ด์˜ ์กด์žฌ ๋น„์œจ์„ ์—…์ƒ˜ํ”Œ๋งํ•˜์—ฌ ๋‹ค๊ตญ์–ด ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ๋„ค๋ชจํŠธ๋ก -4์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ BLOOM์€ ์˜์–ด, ๋‹ค๊ตญ์–ด, ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์˜ ์กฐํ•ฉ์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ฒ”์šฉ ์–ธ์–ด ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ํ‘œ 7์—์„œ ๋„ค๋ชจํŠธ๋ก -4๊ฐ€ 4์ƒท ์„ค์ •์—์„œ ๊ฑฐ์˜ 12%์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์‹คํ˜„ํ•˜๋Š” ๋“ฑ ๋ชจ๋“  ๋ชจ๋ธ ์ค‘์—์„œ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒƒ์„ ๋ช…ํ™•ํ•˜๊ฒŒ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Specifically, for questions where the LLM's response is incorrect, we annotate them with a special token, \(\langle\text{RET}\rangle\), indicating the need for additional context. Subsequently, we utilize these annotations to construct a new dataset tailored for training purposes, where we teach an LLM to answer directly if it is confident about the answer or to require context it believes is useful for answering the question (see Figure 1). Our hypothesis is that through this training process, the LLM learns to use an IR system when it needs extra context to answera question, thus we name it Adapt-LLM.
ํŠนํžˆ, LLM์˜ ์‘๋‹ต์ด ํ‹€๋ฆฐ ์งˆ๋ฌธ์— ๋Œ€ํ•ด์„œ๋Š” ์ถ”๊ฐ€์ ์ธ ๋งฅ๋ฝ์ด ํ•„์š”ํ•จ์„ ๋‚˜ํƒ€๋‚ด๋Š” ํŠน์ˆ˜ ํ† ํฐ \(\langle\text{RET}\rangle\)์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์•˜์Šต๋‹ˆ๋‹ค. ์ดํ›„, ์ด๋Ÿฌํ•œ ์ฃผ์„์„ ํ™œ์šฉํ•˜์—ฌ ํ›ˆ๋ จ ๋ชฉ์ ์— ๋งž์ถ˜ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์…‹์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์šฐ๋ฆฌ๋Š” LLM์ด ๋‹ต๋ณ€์— ์ž์‹ ์ด ์žˆ์„ ๋•Œ๋Š” ์ง์ ‘ ๋‹ต๋ณ€ํ•˜๋„๋ก ํ•˜๊ณ , ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด ์œ ์šฉํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋Š” ๋งฅ๋ฝ์ด ํ•„์š”ํ•  ๋•Œ๋Š” ์ด๋ฅผ ์š”๊ตฌํ•˜๋„๋ก ๊ฐ€๋ฅด์นฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1 ์ฐธ์กฐ). ์šฐ๋ฆฌ์˜ ๊ฐ€์„ค์€ ์ด ํ›ˆ๋ จ ๊ณผ์ •์„ ํ†ตํ•ด LLM์ด ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€์ ์ธ ๋งฅ๋ฝ์ด ํ•„์š”ํ•  ๋•Œ IR ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฒ•์„ ๋ฐฐ์šฐ๊ฒŒ ๋œ๋‹ค๋Š” ๊ฒƒ์ด๋ฉฐ, ๋”ฐ๋ผ์„œ ์ด๋ฅผ Adapt-LLM์ด๋ผ๊ณ  ๋ช…๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค.
๊ตฌ์ฒด์ ์œผ๋กœ, LLM์˜ ์‘๋‹ต์ด ์ž˜๋ชป๋œ ์งˆ๋ฌธ์— ๋Œ€ํ•ด์„œ๋Š” ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๊ฐ€ ํ•„์š”ํ•จ์„ ๋‚˜ํƒ€๋‚ด๋Š” ํŠน์ˆ˜ ํ† ํฐ(\(\langle\text{RET}\rangle\))์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ์ฃผ์„์„ ํ™œ์šฉํ•˜์—ฌ ํ›ˆ๋ จ์šฉ์œผ๋กœ ๋งž์ถคํ™”๋œ ์ƒˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋Š” LLM์ด ๋‹ต๋ณ€์— ๋Œ€ํ•ด ํ™•์‹ ์ด ์žˆ์œผ๋ฉด ์ง์ ‘ ๋‹ต๋ณ€ํ•˜๊ฑฐ๋‚˜ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ์œ ์šฉํ•˜๋‹ค๊ณ  ์ƒ๊ฐ๋˜๋Š” ์ปจํ…์ŠคํŠธ๋ฅผ ์š”๊ตฌํ•˜๋„๋ก ๊ฐ€๋ฅด์นฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1 ์ฐธ์กฐ). ์ด ํ›ˆ๋ จ ๊ณผ์ •์„ ํ†ตํ•ด LLM์ด ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๊ฐ€ ํ•„์š”ํ•  ๋•Œ IR ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•œ๋‹ค๋Š” ๊ฐ€์„ค์„ ์„ธ์šฐ๊ณ , ์ด๋ฅผ Adapt-LLM์ด๋ผ๊ณ  ๋ช…๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค.
**Code.** To collect code data, existing work mainly crawbs open-source licensed codes from the Internet. Two major sources are public code repositories under open-source licenses (_e.g._, GitHub) and code-related question-answering platforms (_e.g._, StackOverflow). Google has publicly released the BigQuery dataset [160], which includes a substantial number of open-source licensed code snippets in various programming languages, serving as a representative code dataset. CodeGen has utilized BIGQUERY [86], a subset of the BigQuery dataset, for training the multilingual version of CodeGen (CodeGen-Multi).
**์ฝ”๋“œ.** ์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ์ธํ„ฐ๋„ท์—์„œ ์˜คํ”ˆ ์†Œ์Šค ๋ผ์ด์„ ์Šค๊ฐ€ ๋ถ€์—ฌ๋œ ์ฝ”๋“œ๋ฅผ ํฌ๋กค๋งํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์š” ์ถœ์ฒ˜๋Š” ์˜คํ”ˆ ์†Œ์Šค ๋ผ์ด์„ ์Šค ํ•˜์— ๊ณต๊ฐœ๋œ ์ฝ”๋“œ ์ €์žฅ์†Œ (_์˜ˆ:_, GitHub)์™€ ์ฝ”๋“œ ๊ด€๋ จ ์งˆ๋ฌธ-๋‹ต๋ณ€ ํ”Œ๋žซํผ (_์˜ˆ:_, StackOverflow)์ž…๋‹ˆ๋‹ค. Google์€ ๋‹ค์–‘ํ•œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด๋กœ ์ž‘์„ฑ๋œ ์ƒ๋‹น์ˆ˜์˜ ์˜คํ”ˆ ์†Œ์Šค ๋ผ์ด์„ ์Šค ์ฝ”๋“œ ์Šค๋‹ˆํŽซ์„ ํฌํ•จํ•˜๋Š” BigQuery ๋ฐ์ดํ„ฐ์…‹ [160]์„ ๊ณต๊ฐœ์ ์œผ๋กœ ์ถœ์‹œํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ๋Œ€ํ‘œ์ ์ธ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. CodeGen์€ CodeGen์˜ ๋‹ค๊ตญ์–ด ๋ฒ„์ „(CodeGen-Multi)์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด BigQuery ๋ฐ์ดํ„ฐ์…‹์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์ธ BIGQUERY [86]๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
**์ฝ”๋“œ**. ์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์ฃผ๋กœ ์ธํ„ฐ๋„ท์—์„œ ์˜คํ”ˆ ์†Œ์Šค ๋ผ์ด์„ ์Šค ์ฝ”๋“œ๋ฅผ ํฌ๋กค๋งํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์†Œ์Šค๋Š” ์˜คํ”ˆ ์†Œ์Šค ๋ผ์ด์„ ์Šค์— ๋”ฐ๋ฅธ ๊ณต๊ฐœ ์ฝ”๋“œ ๋ฆฌํฌ์ง€ํ† ๋ฆฌ(_์˜ˆ: GitHub)์™€ ์ฝ”๋“œ ๊ด€๋ จ ์งˆ์˜์‘๋‹ต ํ”Œ๋žซํผ(_์˜ˆ: StackOverflow)์ž…๋‹ˆ๋‹ค. Google์€ ๋‹ค์–‘ํ•œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด๋กœ ์ž‘์„ฑ๋œ ์ƒ๋‹น์ˆ˜์˜ ์˜คํ”ˆ ์†Œ์Šค ๋ผ์ด์„ ์Šค ์ฝ”๋“œ ์Šค๋‹ˆํŽซ์„ ํฌํ•จํ•˜๋Š” BigQuery ๋ฐ์ดํ„ฐ ์„ธํŠธ [160]์„ ๊ณต๊ฐœ์ ์œผ๋กœ ๊ณต๊ฐœํ•˜์—ฌ ๋Œ€ํ‘œ์ ์ธ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. CodeGen์€ ๋ฉ€ํ‹ฐ ์–ธ์–ด ๋ฒ„์ „ CodeGen(CodeGen-Multi)์˜ ํ•™์Šต์„ ์œ„ํ•ด BigQuery ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์ธ BIGQUERY [86]๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
**GPT-4 Judgement:** The calculation for the cost of the subs, chips, and cookies is correct, as well as the sum of those items. However, the calculation of the delivery fee is incorrect; the solution incorrectly uses a 5% rate (.05) instead of the correct 20% rate (.20). The subsequent steps that build on this mistake are also incorrect, including the final total cost which should be $29.00 instead of $24.00. The correct steps are the individual costs of subs, chips, cookies, and their total before the delivery fee is added, while the steps involving the delivery fee and the final total are incorrect.
**GPT-4 ํŒ๋‹จ:** ์„œ๋ธŒ, ์นฉ, ์ฟ ํ‚ค์˜ ๋น„์šฉ ๊ณ„์‚ฐ๊ณผ ์ด ํ•ญ๋ชฉ๋“ค์˜ ํ•ฉ์€ ์ •ํ™•ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฐฐ๋‹ฌ๋ฃŒ ๊ณ„์‚ฐ์ด ์ž˜๋ชป๋˜์—ˆ์Šต๋‹ˆ๋‹ค; ํ•ด๊ฒฐ์ฑ…์—์„œ๋Š” ์˜ฌ๋ฐ”๋ฅธ 20% ๋น„์œจ(0.20) ๋Œ€์‹  ์ž˜๋ชป๋œ 5% ๋น„์œจ(0.05)์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์‹ค์ˆ˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ํ›„์† ๋‹จ๊ณ„๋“ค๋„ ์ž˜๋ชป๋˜์—ˆ์œผ๋ฉฐ, ์ตœ์ข… ์ด ๋น„์šฉ์€ $24.00์ด ์•„๋‹Œ $29.00์ด์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์˜ฌ๋ฐ”๋ฅธ ๋‹จ๊ณ„๋Š” ๋ฐฐ๋‹ฌ๋ฃŒ๊ฐ€ ์ถ”๊ฐ€๋˜๊ธฐ ์ „์˜ ์„œ๋ธŒ, ์นฉ, ์ฟ ํ‚ค์˜ ๊ฐœ๋ณ„ ๋น„์šฉ๊ณผ ์ดํ•ฉ์ด๋ฉฐ, ๋ฐฐ๋‹ฌ๋ฃŒ์™€ ์ตœ์ข… ์ดํ•ฉ์„ ํฌํ•จํ•œ ๋‹จ๊ณ„๋Š” ์ž˜๋ชป๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
**GPT-4 ํŒ๋‹จ:** ์„œ๋ธŒ, ์นฉ, ์ฟ ํ‚ค์˜ ๋น„์šฉ ๊ณ„์‚ฐ๊ณผ ํ•ด๋‹น ํ’ˆ๋ชฉ์˜ ํ•ฉ๊ณ„๋Š” ์ •ํ™•ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฐฐ๋‹ฌ๋ฃŒ ๊ณ„์‚ฐ์ด ์ž˜๋ชป๋˜์—ˆ์œผ๋ฉฐ, ์†”๋ฃจ์…˜์—์„œ ์˜ฌ๋ฐ”๋ฅธ 20% ์š”์œจ(.20) ๋Œ€์‹  5% ์š”์œจ(.05)์„ ์ž˜๋ชป ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์‹ค์ˆ˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ํ›„์† ๋‹จ๊ณ„๋„ ์ž˜๋ชป๋˜์—ˆ์œผ๋ฉฐ, ์ตœ์ข… ์ด ๋น„์šฉ์€ 24.00๋‹ฌ๋Ÿฌ๊ฐ€ ์•„๋‹Œ 29.00๋‹ฌ๋Ÿฌ๊ฐ€ ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ฐฐ๋‹ฌ๋ฃŒ๋ฅผ ๋”ํ•˜๊ธฐ ์ „์— ์„œ๋ธŒ, ์นฉ, ์ฟ ํ‚ค์˜ ๊ฐœ๋ณ„ ๋น„์šฉ๊ณผ ์ด์•ก์€ ์˜ฌ๋ฐ”๋ฅธ ๋‹จ๊ณ„์ด์ง€๋งŒ ๋ฐฐ๋‹ฌ๋ฃŒ์™€ ์ตœ์ข… ์ด์•ก๊ณผ ๊ด€๋ จ๋œ ๋‹จ๊ณ„๋Š” ์˜ฌ๋ฐ”๋ฅด์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
* Language models typically need at least 50K training steps _regardless of_ batch size. Thus, for small \(N\), we _reduce the batch size_ to ensure the total number of training steps exceeds this threshold. For very large models, a larger batch size is preferred to enable GPU parallelism. * When \(lr\) remains constant, \(wd\) should be relatively reduced as the number of training steps increases. Mathematically, the model weights should be "halved" for every \(\Theta(\frac{1}{lr\times wd})\) training steps. Therefore, it's advisable to reduce the \(wd\) parameter when training for longer periods.
* ์–ธ์–ด ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฐฐ์น˜ ํฌ๊ธฐ์™€ ๊ด€๊ณ„์—†์ด ์ตœ์†Œ 50K์˜ ํ•™์Šต ๋‹จ๊ณ„๋ฅผ ํ•„์š”๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ž‘์€ \(N\)์˜ ๊ฒฝ์šฐ, ์ด ํ•™์Šต ๋‹จ๊ณ„ ์ˆ˜๊ฐ€ ์ด ์ž„๊ณ„๊ฐ’์„ ์ดˆ๊ณผํ•˜๋„๋ก ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ์ค„์ž…๋‹ˆ๋‹ค. ๋งค์šฐ ํฐ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, GPU ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด ๋” ํฐ ๋ฐฐ์น˜ ํฌ๊ธฐ๊ฐ€ ์„ ํ˜ธ๋ฉ๋‹ˆ๋‹ค. * \(lr\)์ด ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€๋  ๋•Œ, ํ•™์Šต ๋‹จ๊ณ„ ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ \(wd\)๋ฅผ ์ƒ๋Œ€์ ์œผ๋กœ ์ค„์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜ํ•™์ ์œผ๋กœ, ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๋Š” \(\Theta(\frac{1}{lr\times wd})\) ํ•™์Šต ๋‹จ๊ณ„๋งˆ๋‹ค "์ ˆ๋ฐ˜"์ด ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋” ๊ธด ๊ธฐ๊ฐ„ ๋™์•ˆ ํ•™์Šตํ•  ๋•Œ๋Š” \(wd\) ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์ค„์ด๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.
* ์–ธ์–ด ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฐฐ์น˜ ํฌ๊ธฐ์— ๊ด€๊ณ„์—†์ด ์ตœ์†Œ 50K ํŠธ๋ ˆ์ด๋‹ ๋‹จ๊ณ„๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ \(N\)์ด ์ž‘์€ ๊ฒฝ์šฐ ์ด ํŠธ๋ ˆ์ด๋‹ ๋‹จ๊ณ„ ์ˆ˜๊ฐ€ ์ด ์ž„๊ณ„๊ฐ’์„ ์ดˆ๊ณผํ•˜๋„๋ก ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ์ค„์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ๋งค์šฐ ํฐ ๊ฒฝ์šฐ์—๋Š” GPU ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•ด ๋” ํฐ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ์„ ํ˜ธํ•ฉ๋‹ˆ๋‹ค. * lr$์ด ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€๋˜๋Š” ๊ฒฝ์šฐ, ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ $wd$๋Š” ์ƒ๋Œ€์ ์œผ๋กœ ๊ฐ์†Œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜ํ•™์ ์œผ๋กœ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๋Š” \(\Theta(\frac{1}{lr\times wd})\) ํ›ˆ๋ จ ๋‹จ๊ณ„๋งˆ๋‹ค "๋ฐ˜๊ฐ"๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์žฅ์‹œ๊ฐ„ ํ›ˆ๋ จํ•  ๋•Œ๋Š” \(wd\) ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์ค„์ด๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.
For each two-hop prompt \(\tau_{\text{2H}}\) in TwoHopFact, we randomly select one \(e_{1}^{\prime}\) from the same fact composition type and one \(r_{1}^{\prime}\) from a set of predefined candidate relations (provided in Appendix Table 5) to create \(\tau_{\text{2H}}^{\prime}\). We then measure the relative frequency of cases where replacing \(\tau_{\text{2H}}^{\prime}\) with \(\tau_{\text{2H}}\) via entity or relation substitution increases the recall of \(e_{2}\). A relative frequency above 0.5 suggests the LLM's chance to perform first-hop reasoning exceeds the random chance for these prompts.
TwoHopFact์˜ ๊ฐ ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ \(\tau_{\text{2H}}\)์— ๋Œ€ํ•ด, ๋™์ผํ•œ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์—์„œ ํ•˜๋‚˜์˜ \(e_{1}^{\prime}\)์„ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ•˜๊ณ , ๋ฏธ๋ฆฌ ์ •์˜๋œ ํ›„๋ณด ๊ด€๊ณ„ ์ง‘ํ•ฉ(๋ถ€๋ก ํ‘œ 5์— ์ œ๊ณต๋จ)์—์„œ ํ•˜๋‚˜์˜ \(r_{1}^{\prime}\)์„ ์„ ํƒํ•˜์—ฌ \(\tau_{\text{2H}}^{\prime}\)์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์—”ํ‹ฐํ‹ฐ ๋˜๋Š” ๊ด€๊ณ„ ๋Œ€์ฒด๋ฅผ ํ†ตํ•ด \(\tau_{\text{2H}}\)๋ฅผ \(\tau_{\text{2H}}^{\prime}\)๋กœ ๋Œ€์ฒดํ•  ๋•Œ \(e_{2}\)์˜ ๋ฆฌ์ฝœ์ด ์ฆ๊ฐ€ํ•˜๋Š” ๊ฒฝ์šฐ์˜ ์ƒ๋Œ€ ๋นˆ๋„๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. ์ƒ๋Œ€ ๋นˆ๋„๊ฐ€ 0.5๋ฅผ ์ดˆ๊ณผํ•˜๋ฉด ์ด๋Ÿฌํ•œ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด LLM์ด ์ฒซ ๋ฒˆ์งธ ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ํ™•๋ฅ ์ด ๋ฌด์ž‘์œ„ ํ™•๋ฅ ์„ ์ดˆ๊ณผํ•จ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
TwoHopFact์˜ ๊ฐ 2ํ™‰ ํ”„๋กฌํ”„ํŠธ \(\tau_{\text{2H}}\)์— ๋Œ€ํ•ด, ๋™์ผํ•œ ํŒฉํŠธ ๊ตฌ์„ฑ ์œ ํ˜•์—์„œ ํ•˜๋‚˜์˜ \(e_{1}^{\prime}\)๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ•˜๊ณ  ๋ฏธ๋ฆฌ ์ •์˜๋œ ํ›„๋ณด ๊ด€๊ณ„ ์ง‘ํ•ฉ(๋ถ€๋ก ํ‘œ 5์— ์ œ๊ณต๋จ)์—์„œ ํ•˜๋‚˜์˜ \(r_{1}^{\prime}\)๋ฅผ ์„ ํƒํ•˜์—ฌ \(\tau_{\text{2H}}^{\prime}\)๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์—”ํ‹ฐํ‹ฐ ๋˜๋Š” ๊ด€๊ณ„ ์น˜ํ™˜์„ ํ†ตํ•ด \(\tau_{\text{2H}}^{\prime}\)๋ฅผ \(\tau_{\text{2H}}\)๋กœ ๋Œ€์ฒดํ•  ๋•Œ \(e_{2}\)์˜ ๋ฆฌ์ฝœ์ด ์ฆ๊ฐ€ํ•˜๋Š” ๊ฒฝ์šฐ์˜ ์ƒ๋Œ€์  ๋นˆ๋„๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. ์ƒ๋Œ€ ๋นˆ๋„๊ฐ€ 0.5 ์ด์ƒ์ด๋ฉด LLM์ด ์ฒซ ๋ฒˆ์งธ ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ํ™•๋ฅ ์ด ์ด๋Ÿฌํ•œ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ๋ฌด์ž‘์œ„ ํ™•๋ฅ ์„ ์ดˆ๊ณผํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
We also do not investigate the negative consequences of deduplication. Some language tasks explicitly require memorization, like document retrieval or closed-book question answering. Also, text that gives attribution is often duplicated across documents, so removing duplicate substrings could correspond to removing _just_ the attribution, which could result in models that learn the content without its attached attribution. Deduplication is also not sufficient to remove privacy-sensitive data like bank passwords and medical records which should never be used in training data (Brown et al., 2022).
์šฐ๋ฆฌ๋Š” ๋˜ํ•œ ์ค‘๋ณต ์ œ๊ฑฐ์˜ ๋ถ€์ •์ ์ธ ๊ฒฐ๊ณผ์— ๋Œ€ํ•ด์„œ๋Š” ์กฐ์‚ฌํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ์–ธ์–ด ์ž‘์—…์€ ๋ฌธ์„œ ๊ฒ€์ƒ‰์ด๋‚˜ ํ์‡„ํ˜• ์งˆ๋ฌธ ์‘๋‹ต๊ณผ ๊ฐ™์ด ๋ช…์‹œ์ ์œผ๋กœ ์•”๊ธฐ๋ฅผ ์š”๊ตฌํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ถœ์ฒ˜๋ฅผ ๋ช…์‹œํ•˜๋Š” ํ…์ŠคํŠธ๋Š” ์ข…์ข… ๋ฌธ์„œ ๊ฐ„์— ์ค‘๋ณต๋˜๋ฏ€๋กœ, ์ค‘๋ณต๋œ ๋ถ€๋ถ„ ๋ฌธ์ž์—ด์„ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์€ ์ถœ์ฒ˜๋งŒ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ถœ์ฒ˜ ์—†์ด ๋‚ด์šฉ์„ ํ•™์Šตํ•˜๋Š” ๋ชจ๋ธ์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๋˜ํ•œ ์€ํ–‰ ๋น„๋ฐ€๋ฒˆํ˜ธ๋‚˜ ์˜๋ฃŒ ๊ธฐ๋ก๊ณผ ๊ฐ™์€ ๋ฏผ๊ฐํ•œ ๊ฐœ์ธ์ •๋ณด๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๋ฐ ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์œผ๋ฉฐ, ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋Š” ์ ˆ๋Œ€๋กœ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ์‚ฌ์šฉ๋˜์–ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค (Brown et al., 2022).
๋˜ํ•œ ์ค‘๋ณต ์ œ๊ฑฐ์˜ ๋ถ€์ •์ ์ธ ๊ฒฐ๊ณผ์— ๋Œ€ํ•ด์„œ๋„ ์กฐ์‚ฌํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋ฌธ์„œ ๊ฒ€์ƒ‰์ด๋‚˜ ํ์‡„ํ˜• ๋ฌธ์ œ ๋‹ต๋ณ€๊ณผ ๊ฐ™์ด ์ผ๋ถ€ ์–ธ์–ด ์ž‘์—…์—๋Š” ๋ช…์‹œ์ ์œผ๋กœ ์•”๊ธฐ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ถœ์ฒ˜๋ฅผ ์ œ๊ณตํ•˜๋Š” ํ…์ŠคํŠธ๋Š” ๋ฌธ์„œ ๊ฐ„์— ์ค‘๋ณต๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฏ€๋กœ ์ค‘๋ณต๋œ ํ•˜์œ„ ๋ฌธ์ž์—ด์„ ์ œ๊ฑฐํ•˜๋ฉด ์ถœ์ฒ˜๋งŒ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ๊ณผ๋„ ๊ฐ™์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋กœ ์ธํ•ด ์ถœ์ฒ˜๊ฐ€ ์ฒจ๋ถ€๋˜์ง€ ์•Š์€ ์ฝ˜ํ…์ธ ๋งŒ ํ•™์Šตํ•˜๋Š” ๋ชจ๋ธ์ด ๋งŒ๋“ค์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ์‚ฌ์šฉํ•ด์„œ๋Š” ์•ˆ ๋˜๋Š” ์€ํ–‰ ๋น„๋ฐ€๋ฒˆํ˜ธ๋‚˜ ์˜๋ฃŒ ๊ธฐ๋ก๊ณผ ๊ฐ™์€ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ์— ๋ฏผ๊ฐํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ•˜๊ธฐ์—๋Š” ์ค‘๋ณต ์ œ๊ฑฐ๋งŒ์œผ๋กœ๋Š” ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค(Brown et al., 2022).
Current Events Task To further isolate LLMs' abilities to learn new knowledge, we created a task comprising multiple-choice questions about current events. This task includes multiple-choice questions about events that occurred after the cutoff of the various models' training data. Specifically, we focused on "current events" from the USA, in the time span of August-November 2023, that are included in the relevant Wikipedia indexes1. This method enables us to mostly guarantee that the models have not been exposed to these facts, thus allowing us to directly test knowledge injection capabilities.
ํ˜„์žฌ ์‚ฌ๊ฑด ๊ณผ์ œ: LLM์˜ ์ƒˆ๋กœ์šด ์ง€์‹์„ ํ•™์Šตํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๋”์šฑ ๋ถ„๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ํ˜„์žฌ ์‚ฌ๊ฑด์— ๋Œ€ํ•œ ๊ฐ๊ด€์‹ ์งˆ๋ฌธ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๊ณผ์ œ๋ฅผ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ œ๋Š” ๋‹ค์–‘ํ•œ ๋ชจ๋ธ์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ๊ฐ€ ์ข…๋ฃŒ๋œ ์ดํ›„์— ๋ฐœ์ƒํ•œ ์‚ฌ๊ฑด์— ๋Œ€ํ•œ ๊ฐ๊ด€์‹ ์งˆ๋ฌธ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” 2023๋…„ 8์›”๋ถ€ํ„ฐ 11์›”๊นŒ์ง€์˜ ๊ธฐ๊ฐ„ ๋™์•ˆ ๋ฏธ๊ตญ์—์„œ ๋ฐœ์ƒํ•œ "ํ˜„์žฌ ์‚ฌ๊ฑด"์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์œผ๋ฉฐ, ์ด๋Š” ๊ด€๋ จ๋œ Wikipedia ์ƒ‰์ธ์— ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ์ด ์ด๋Ÿฌํ•œ ์‚ฌ์‹ค์— ๋…ธ์ถœ๋˜์ง€ ์•Š์•˜์Œ์„ ๋Œ€๋ถ€๋ถ„ ๋ณด์žฅํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•˜์—ฌ, ์ง€์‹ ์ฃผ์ž… ๋Šฅ๋ ฅ์„ ์ง์ ‘์ ์œผ๋กœ ํ…Œ์ŠคํŠธํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.
์‹œ์‚ฌ ์ด๋ฒคํŠธ ์ž‘์—… LLM์ด ์ƒˆ๋กœ์šด ์ง€์‹์„ ํ•™์Šตํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๋”์šฑ ๊ฒฉ๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ์‹œ์‚ฌ ์ด๋ฒคํŠธ์— ๊ด€ํ•œ ๊ฐ๊ด€์‹ ๋ฌธ์ œ๋กœ ๊ตฌ์„ฑ๋œ ๊ณผ์ œ๋ฅผ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ œ์—๋Š” ๋‹ค์–‘ํ•œ ๋ชจ๋ธ์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ ๋งˆ๊ฐ์ผ ์ดํ›„์— ๋ฐœ์ƒํ•œ ์‚ฌ๊ฑด์— ๋Œ€ํ•œ ๊ฐ๊ด€์‹ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์šฐ๋ฆฌ๋Š” 2023๋…„ 8~11์›” ๋ฏธ๊ตญ์—์„œ ๋ฐœ์ƒํ•œ '์‹œ์‚ฌ ์ด๋ฒคํŠธ' ์ค‘ ๊ด€๋ จ ์œ„ํ‚ค๋ฐฑ๊ณผ ์ƒ‰์ธ1์— ํฌํ•จ๋œ ์‚ฌ๊ฑด์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜๋ฉด ๋ชจ๋ธ์ด ์ด๋Ÿฌํ•œ ์‚ฌ์‹ค์— ๋…ธ์ถœ๋˜์ง€ ์•Š์•˜์Œ์„ ๋Œ€๋ถ€๋ถ„ ๋ณด์žฅํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ง€์‹ ์ฃผ์ž… ๊ธฐ๋Šฅ์„ ์ง์ ‘ ํ…Œ์ŠคํŠธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The main limitation of this article is the relatively small amount of saturated language models we studied. As it is the only suite of language models trained in the range of interest to release an extensive amount of intermediate checkpoints, we could only observe the training dynamics of small Pythia models. Although we observe strong last-layer anisotropy for the smallest GPT-2 model, we cannot tell with certainty whether it suffered from saturation. The OPT-125m model does not display a strong last-layer anisotropy, which could indicate that it was not affected by the saturation phenomenon.
์ด ๋…ผ๋ฌธ์˜ ์ฃผ์š” ํ•œ๊ณ„๋Š” ์šฐ๋ฆฌ๊ฐ€ ์—ฐ๊ตฌํ•œ ํฌํ™”๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์ˆ˜๊ฐ€ ์ƒ๋Œ€์ ์œผ๋กœ ์ ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ๊ด€์‹ฌ ๋ฒ”์œ„ ๋‚ด์—์„œ ๊ด‘๋ฒ”์œ„ํ•œ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๊ณต๊ฐœํ•œ ์œ ์ผํ•œ ์–ธ์–ด ๋ชจ๋ธ ๋ชจ์Œ์ด๊ธฐ ๋•Œ๋ฌธ์—, ์šฐ๋ฆฌ๋Š” ์ž‘์€ Pythia ๋ชจ๋ธ์˜ ํ•™์Šต ๋™์—ญํ•™๋งŒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์ž‘์€ GPT-2 ๋ชจ๋ธ์—์„œ ๊ฐ•ํ•œ ๋งˆ์ง€๋ง‰ ์ธต์˜ ๋น„๋“ฑ๋ฐฉ์„ฑ์„ ๊ด€์ฐฐํ–ˆ์ง€๋งŒ, ๊ทธ๊ฒƒ์ด ํฌํ™” ํ˜„์ƒ์„ ๊ฒช์—ˆ๋Š”์ง€ ํ™•์‹คํžˆ ๋งํ•  ์ˆ˜๋Š” ์—†์Šต๋‹ˆ๋‹ค. OPT-125m ๋ชจ๋ธ์€ ๊ฐ•ํ•œ ๋งˆ์ง€๋ง‰ ์ธต์˜ ๋น„๋“ฑ๋ฐฉ์„ฑ์„ ๋‚˜ํƒ€๋‚ด์ง€ ์•Š์œผ๋ฉฐ, ์ด๋Š” ํฌํ™” ํ˜„์ƒ์˜ ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š์•˜์Œ์„ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด ๊ธ€์˜ ์ฃผ์š” ํ•œ๊ณ„๋Š” ์šฐ๋ฆฌ๊ฐ€ ์—ฐ๊ตฌํ•œ ํฌํ™” ์–ธ์–ด ๋ชจ๋ธ์˜ ์ƒ๋Œ€์ ์œผ๋กœ ์ ์€ ์–‘์ž…๋‹ˆ๋‹ค. ๊ด‘๋ฒ”์œ„ํ•œ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๊ณต๊ฐœํ•˜๊ธฐ ์œ„ํ•ด ๊ด€์‹ฌ ์žˆ๋Š” ๋ฒ”์œ„์—์„œ ํ›ˆ๋ จ๋œ ์œ ์ผํ•œ ์–ธ์–ด ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์ด๊ธฐ ๋•Œ๋ฌธ์— ์ž‘์€ Pythia ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ๋™์—ญํ•™๋งŒ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์ž‘์€ GPT-2 ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐ•ํ•œ ๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด ์ด๋ฐฉ์„ฑ์„ ๊ด€์ฐฐํ–ˆ์ง€๋งŒ, ํฌํ™” ํ˜„์ƒ์œผ๋กœ ์ธํ•ด ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ–ˆ๋Š”์ง€ ํ™•์‹คํ•˜๊ฒŒ ์•Œ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. OPT-125m ๋ชจ๋ธ์€ ๊ฐ•ํ•œ ๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด ์ด๋ฐฉ์„ฑ์„ ๋‚˜ํƒ€๋‚ด์ง€ ์•Š์œผ๋ฉฐ, ์ด๋Š” ํฌํ™” ํ˜„์ƒ์˜ ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š์•˜์Œ์„ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Given two suffix arrays \(A_{1}\) and \(A_{2}\) for two sequences \(S_{1}\) and \(S_{2}\) it's not completely trivial to construct a single suffix array \(A\) for \(S=S_{1}\parallel S_{2}\) because of the boundary conditions. Instead, we don't build the data \(S=S_{1}\ ||\ S_{2}\) but rather let \(S_{1}^{\prime}=S_{1}\ ||\ S_{2}[uptoK]\) for some \(K\) greater than the longest substring match. Then we build the arrays on \(S_{1}^{\prime}\) and \(S_{2}\). To merge the arrays together we can remove the items from the first array after index \(|S_{1}|\) and merge-sort insert them into the second.
๋‘ ์‹œํ€€์Šค \(S_{1}\)์™€ \(S_{2}\)์— ๋Œ€ํ•œ ๋‘ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด \(A_{1}\)์™€ \(A_{2}\)๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ๊ฒฝ๊ณ„ ์กฐ๊ฑด ๋•Œ๋ฌธ์— \(S=S_{1}\parallel S_{2}\)์— ๋Œ€ํ•œ ๋‹จ์ผ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด \(A\)๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์€ ์™„์ „ํžˆ ๊ฐ„๋‹จํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋Œ€์‹ , ์šฐ๋ฆฌ๋Š” \(S=S_{1}\ ||\ S_{2}\) ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์ถ•ํ•˜์ง€ ์•Š๊ณ , ๊ฐ€์žฅ ๊ธด ๋ถ€๋ถ„ ๋ฌธ์ž์—ด ์ผ์น˜๋ณด๋‹ค ํฐ \(K\)์— ๋Œ€ํ•ด \(S_{1}^{\prime}=S_{1}\ ||\ S_{2}[uptoK]\)๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ \(S_{1}^{\prime}\)๊ณผ \(S_{2}\)์— ๋Œ€ํ•œ ๋ฐฐ์—ด์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ๋ฐฐ์—ด์„ ํ•จ๊ป˜ ๋ณ‘ํ•ฉํ•˜๊ธฐ ์œ„ํ•ด ์ฒซ ๋ฒˆ์งธ ๋ฐฐ์—ด์—์„œ ์ธ๋ฑ์Šค \(|S_{1}|\) ์ดํ›„์˜ ํ•ญ๋ชฉ์„ ์ œ๊ฑฐํ•˜๊ณ  ์ด๋ฅผ ๋‘ ๋ฒˆ์งธ ๋ฐฐ์—ด์— ๋ณ‘ํ•ฉ ์ •๋ ฌ ์‚ฝ์ž…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋‘ ์‹œํ€€์Šค \(S_{1}\) ๋ฐ \(S_{2}\)์— ๋Œ€ํ•œ ๋‘ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด \(A_{1}\) ๋ฐ \(A_{2}\)๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ๊ฒฝ๊ณ„ ์กฐ๊ฑด ๋•Œ๋ฌธ์— \(S=S_{1}\parallel S_{2}\)์— ๋Œ€ํ•œ ๋‹จ์ผ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด \(A\)๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์€ ์™„์ „ํžˆ ๊ฐ„๋‹จํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋Œ€์‹ , \(S=S_{1}\ ||\ S_{2}\) ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๋Œ€์‹  ๊ฐ€์žฅ ๊ธด ๋ถ€๋ถ„ ๋ฌธ์ž์—ด ์ผ์น˜๋ณด๋‹ค ํฐ \(K\)์— ๋Œ€ํ•ด \(S_{1}^{\prime}=S_{1}\ ||\ S_{2}[uptoK]\)๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ \(S_{1}^{\prime}\)์™€ \(S_{2}\)์— ๋Œ€ํ•œ ๋ฐฐ์—ด์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ๋ฐฐ์—ด์„ ๋ณ‘ํ•ฉํ•˜๊ธฐ ์œ„ํ•ด ์ฒซ ๋ฒˆ์งธ ๋ฐฐ์—ด์—์„œ ์ธ๋ฑ์Šค \(|S_{1}|\) ์ดํ›„์˜ ํ•ญ๋ชฉ์„ ์ œ๊ฑฐํ•˜๊ณ  ๋‘ ๋ฒˆ์งธ ๋ฐฐ์—ด์— ๋ณ‘ํ•ฉ ์ •๋ ฌ ์‚ฝ์ž…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Contribution 2: An improved evaluation protocol for legal LLMs.Concurrently, we introduce LegalBench-Instruct, a supplemental iteration of LegalBench Guha et al. (2022, 2023)2, crafted to better gauge and refine the legal proficiency of language models, which we hope will contribute to future advancements into research in the legal domain. To further enrich the models' capabilities in legal contexts, we also include the legal tasks of the popular MMLU benchmark Hendrycks et al. (2020) in our evaluation protocol, particularly focusing on international law, professional law3 and jurisprudence.
๊ธฐ์—ฌ 2: ๋ฒ•๋ฅ  LLMs์— ๋Œ€ํ•œ ๊ฐœ์„ ๋œ ํ‰๊ฐ€ ํ”„๋กœํ† ์ฝœ. ๋™์‹œ์—, ์šฐ๋ฆฌ๋Š” LegalBench Guha et al. (2022, 2023)2์˜ ๋ณด์™„๋œ ๋ฐ˜๋ณต์ธ LegalBench-Instruct๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ฒ•๋ฅ  ์ „๋ฌธ์„ฑ์„ ๋” ์ž˜ ํ‰๊ฐ€ํ•˜๊ณ  ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ œ์ž‘๋˜์—ˆ์œผ๋ฉฐ, ๋ฒ•๋ฅ  ๋ถ„์•ผ ์—ฐ๊ตฌ์˜ ๋ฏธ๋ž˜ ๋ฐœ์ „์— ๊ธฐ์—ฌํ•  ์ˆ˜ ์žˆ๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค. ๋ชจ๋ธ์˜ ๋ฒ•๋ฅ  ๋งฅ๋ฝ์—์„œ์˜ ๋Šฅ๋ ฅ์„ ๋”์šฑ ํ’๋ถ€ํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ํ‰๊ฐ€ ํ”„๋กœํ† ์ฝœ์— ์ธ๊ธฐ ์žˆ๋Š” MMLU ๋ฒค์น˜๋งˆํฌ Hendrycks et al. (2020)์˜ ๋ฒ•๋ฅ  ๊ณผ์ œ๋„ ํฌํ•จ์‹œ์ผฐ์œผ๋ฉฐ, ํŠนํžˆ ๊ตญ์ œ๋ฒ•, ์ „๋ฌธ๋ฒ•3 ๋ฐ ๋ฒ•ํ•™์— ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค.
๊ธฐ์—ฌ๋„ 2: ๊ฐœ์„ ๋œ ๋ฒ•๋ฅ  LLM ํ‰๊ฐ€ ํ”„๋กœํ† ์ฝœ.๋™์‹œ์—, ์šฐ๋ฆฌ๋Š” ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ฒ•๋ฅ ์  ์ˆ™๋ จ๋„๋ฅผ ๋” ์ž˜ ์ธก์ •ํ•˜๊ณ  ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ œ์ž‘๋œ LegalBench-Instruct, LegalBench Guha et al. (2022, 2023)2์˜ ๋ณด์™„์ ์ธ ๋ฐ˜๋ณต์„ ์†Œ๊ฐœํ•˜๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ํ–ฅํ›„ ๋ฒ•๋ฅ  ์˜์—ญ์˜ ์—ฐ๊ตฌ ๋ฐœ์ „์— ๊ธฐ์—ฌํ•  ์ˆ˜ ์žˆ๊ธฐ๋ฅผ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค. ๋ฒ•๋ฅ  ๋งฅ๋ฝ์—์„œ ๋ชจ๋ธ์˜ ์—ญ๋Ÿ‰์„ ๋”์šฑ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ‰๊ฐ€ ํ”„๋กœํ† ์ฝœ์— ์ธ๊ธฐ ์žˆ๋Š” MMLU ๋ฒค์น˜๋งˆํฌ Hendrycks et al. (2020)์˜ ๋ฒ•๋ฅ  ๊ณผ์ œ, ํŠนํžˆ ๊ตญ์ œ๋ฒ•, ์ „๋ฌธ๋ฒ•3 ๋ฐ ๋ฒ•ํ•™์— ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
We adopt the performance on XQuAD in Chinese as the validation set to all languages and all tasks. Specifically, Table 9 shows the result on Vicuna when deactivating language-specific neurons in the understanding layer (\(D_{\mathcal{U}}\)) and generation layer (\(D_{\mathcal{G}}\)), where \(N_{1}\) is the number of understanding layers and \(N_{2}\) is the number of generation layer. We find that when setting \(N_{1}=8\) and \(N_{2}=2\), performance in English is influenced the least while performance in Chinese decreases the most. As for Mistral, the number is \(N_{1}=6\) and \(N_{2}=3\).
์šฐ๋ฆฌ๋Š” ๋ชจ๋“  ์–ธ์–ด์™€ ๋ชจ๋“  ์ž‘์—…์— ๋Œ€ํ•ด ์ค‘๊ตญ์–ด๋กœ ๋œ XQuAD์˜ ์„ฑ๋Šฅ์„ ๊ฒ€์ฆ ์„ธํŠธ๋กœ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, Table 9๋Š” ์ดํ•ด ๋ ˆ์ด์–ด (\(D_{\mathcal{U}}\))์™€ ์ƒ์„ฑ ๋ ˆ์ด์–ด (\(D_{\mathcal{G}\))์—์„œ ์–ธ์–ด ํŠนํ™” ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•  ๋•Œ Vicuna์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ \(N_{1}\)์€ ์ดํ•ด ๋ ˆ์ด์–ด์˜ ์ˆ˜์ด๊ณ  \(N_{2}\)๋Š” ์ƒ์„ฑ ๋ ˆ์ด์–ด์˜ ์ˆ˜์ž…๋‹ˆ๋‹ค. \(N_{1}=8\)๊ณผ \(N_{2}=2\)๋กœ ์„ค์ •ํ–ˆ์„ ๋•Œ ์˜์–ด ์„ฑ๋Šฅ์€ ๊ฐ€์žฅ ์ ๊ฒŒ ์˜ํ–ฅ์„ ๋ฐ›๋Š” ๋ฐ˜๋ฉด, ์ค‘๊ตญ์–ด ์„ฑ๋Šฅ์€ ๊ฐ€์žฅ ๋งŽ์ด ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. Mistral์˜ ๊ฒฝ์šฐ, ๊ทธ ์ˆ˜๋Š” \(N_{1}=6\)๊ณผ \(N_{2}=3\)์ž…๋‹ˆ๋‹ค.
์šฐ๋ฆฌ๋Š” ์ค‘๊ตญ์–ด๋กœ ๋œ XQuAD์˜ ์„ฑ๋Šฅ์„ ๋ชจ๋“  ์–ธ์–ด์™€ ๋ชจ๋“  ์ž‘์—…์— ๋Œ€ํ•œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋กœ ์ฑ„ํƒํ•œ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ํ‘œ 9๋Š” ์ดํ•ด ๊ณ„์ธต(\(D_{\mathcal{U}}\))๊ณผ ์ƒ์„ฑ ๊ณ„์ธต(\(D_{\mathcal{G}}\))์—์„œ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•  ๋•Œ Vicuna์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์—ฌ๊ธฐ์„œ \(N_{1}\)์€ ์ดํ•ด ๊ณ„์ธต์˜ ์ˆ˜์ด๊ณ  \(N_{2}\)๋Š” ์ƒ์„ฑ ๊ณ„์ธต์˜ ์ˆ˜์ด๋‹ค. N_{1}=8$ ๋ฐ \(N_{2}=2$๋กœ ์„ค์ •ํ•˜๋ฉด ์˜์–ด ์„ฑ๋Šฅ์€ ์˜ํ–ฅ์„ ๊ฐ€์žฅ ์ ๊ฒŒ ๋ฐ›๊ณ  ์ค‘๊ตญ์–ด ์„ฑ๋Šฅ์€ ๊ฐ€์žฅ ๋งŽ์ด ๊ฐ์†Œํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค. Mistral์˜ ๊ฒฝ์šฐ, ๊ทธ ์ˆ˜๋Š” \(N_{1}=6$ ๋ฐ \(N_{2}=3$์ด๋‹ค.
\(\bullet\)_Helpfulness._ To be helpful, the LLM should demonstrate a clear attempt to assist users in solving their tasks or answering questions in a concise and efficient manner as possible. At a higher level, when further clarification is needed, the LLM should demonstrate the capability of eliciting additional relevant information through pertinent inquiries and exhibit suitable levels of sensitivity, perceptiveness, and prudence [368]. Realizing the alignment of helpful behavior is challenging for LLMs since it is difficult to precisely define and measure the intention of users [366].
โ€ข ์œ ์šฉ์„ฑ: ์œ ์šฉํ•˜๊ธฐ ์œ„ํ•ด์„œ, LLM์€ ์‚ฌ์šฉ์ž๊ฐ€ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ฑฐ๋‚˜ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ์žˆ์–ด ๊ฐ€๋Šฅํ•œ ํ•œ ๊ฐ„๊ฒฐํ•˜๊ณ  ํšจ์œจ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ๋ช…ํ™•ํ•œ ๋„์›€์„ ์ œ๊ณตํ•˜๋ ค๋Š” ์‹œ๋„๋ฅผ ๋ณด์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋” ๋†’์€ ์ˆ˜์ค€์—์„œ๋Š” ์ถ”๊ฐ€์ ์ธ ์„ค๋ช…์ด ํ•„์š”ํ•  ๋•Œ, LLM์€ ๊ด€๋ จ๋œ ์งˆ๋ฌธ์„ ํ†ตํ•ด ์ถ”๊ฐ€์ ์ธ ๊ด€๋ จ ์ •๋ณด๋ฅผ ์ด๋Œ์–ด๋‚ด๋Š” ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์•ผ ํ•˜๋ฉฐ, ์ ์ ˆํ•œ ์ˆ˜์ค€์˜ ๋ฏผ๊ฐ์„ฑ, ํ†ต์ฐฐ๋ ฅ, ์‹ ์ค‘ํ•จ์„ ๋‚˜ํƒ€๋‚ด์•ผ ํ•ฉ๋‹ˆ๋‹ค [368]. ์œ ์šฉํ•œ ํ–‰๋™์˜ ์ •๋ ฌ์„ ์‹คํ˜„ํ•˜๋Š” ๊ฒƒ์€ LLM์—๊ฒŒ ๋„์ „์ ์ธ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์‚ฌ์šฉ์ž์˜ ์˜๋„๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์ •์˜ํ•˜๊ณ  ์ธก์ •ํ•˜๋Š” ๊ฒƒ์ด ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค [366].
\(\bullet\)_์œ ์šฉ์„ฑ._ ์œ ์šฉ์„ฑ์„ ์œ„ํ•ด LLM์€ ๊ฐ€๋Šฅํ•œ ํ•œ ๊ฐ„๊ฒฐํ•˜๊ณ  ํšจ์œจ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์šฉ์ž์˜ ์ž‘์—… ํ•ด๊ฒฐ์ด๋‚˜ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์ง€์›ํ•˜๋ ค๋Š” ๋…ธ๋ ฅ์„ ๋ช…ํ™•ํ•˜๊ฒŒ ๋ณด์—ฌ์ค˜์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋” ์ž์„ธํ•œ ์„ค๋ช…์ด ํ•„์š”ํ•œ ๊ฒฝ์šฐ, LLM์€ ์ ์ ˆํ•œ ์งˆ๋ฌธ์„ ํ†ตํ•ด ์ถ”๊ฐ€ ๊ด€๋ จ ์ •๋ณด๋ฅผ ์ด๋Œ์–ด๋‚ด๋Š” ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์–ด์•ผ ํ•˜๋ฉฐ, ์ ์ ˆํ•œ ์ˆ˜์ค€์˜ ๊ฐ์ˆ˜์„ฑ, ํ†ต์ฐฐ๋ ฅ, ์‹ ์ค‘ํ•จ์„ ๋ณด์—ฌ์ค˜์•ผ ํ•ฉ๋‹ˆ๋‹ค[368]. ์œ ์šฉํ•œ ํ–‰๋™์˜ ์ •๋ ฌ์„ ์‹คํ˜„ํ•˜๋Š” ๊ฒƒ์€ ์‚ฌ์šฉ์ž์˜ ์˜๋„๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์ •์˜ํ•˜๊ณ  ์ธก์ •ํ•˜๊ธฐ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์— LLM์— ์–ด๋ ค์šด ๊ณผ์ œ์ž…๋‹ˆ๋‹ค[366].
However, model merging is considered by many to be a form of black art or alchemy, relying on the model maker's intuition and instincts about model selection and merging recipes to create and refine a new model that performs well for a particular task. Furthermore, the model maker is often required to have some domain knowledge for the various different benchmark tasks. Given the large diversity of open models and benchmarks in the community, human intuition can only go so far, and we believe a more systematic approach for discovering new model combinations will take things much further.
๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋ธ ๋ณ‘ํ•ฉ์€ ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์— ์˜ํ•ด ์ผ์ข…์˜ ํ‘๋งˆ์ˆ  ๋˜๋Š” ์—ฐ๊ธˆ์ˆ ๋กœ ๊ฐ„์ฃผ๋˜๋ฉฐ, ํŠน์ • ์ž‘์—…์— ๋Œ€ํ•ด ์ž˜ ์ˆ˜ํ–‰๋˜๋Š” ์ƒˆ๋กœ์šด ๋ชจ๋ธ์„ ๋งŒ๋“ค๊ณ  ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ ์ œ์ž‘์ž์˜ ์ง๊ด€๊ณผ ๋ณธ๋Šฅ์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋ชจ๋ธ ์ œ์ž‘์ž๋Š” ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ ์ž‘์—…์— ๋Œ€ํ•œ ๋„๋ฉ”์ธ ์ง€์‹์„ ๊ฐ–์ถ”์–ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ปค๋ฎค๋‹ˆํ‹ฐ ๋‚ด์˜ ๋‹ค์–‘ํ•œ ์˜คํ”ˆ ๋ชจ๋ธ๊ณผ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ณ ๋ คํ•  ๋•Œ, ์ธ๊ฐ„์˜ ์ง๊ด€์—๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ๋Š” ์ƒˆ๋กœ์šด ๋ชจ๋ธ ์กฐํ•ฉ์„ ๋ฐœ๊ฒฌํ•˜๊ธฐ ์œ„ํ•œ ๋ณด๋‹ค ์ฒด๊ณ„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์ด ํ›จ์”ฌ ๋” ๋‚˜์•„๊ฐˆ ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋ธ ๋ณ‘ํ•ฉ์€ ๋ชจ๋ธ ์„ ํƒ๊ณผ ๋ณ‘ํ•ฉ ๋ ˆ์‹œํ”ผ์— ๋Œ€ํ•œ ๋ชจ๋ธ ์ œ์ž‘์ž์˜ ์ง๊ด€๊ณผ ๋ณธ๋Šฅ์— ์˜์กดํ•˜์—ฌ ํŠน์ • ์ž‘์—…์— ์ ํ•ฉํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ชจ๋ธ์„ ๋งŒ๋“ค๊ณ  ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์ด ์•”์ˆ ์ด๋‚˜ ์—ฐ๊ธˆ์ˆ ์˜ ํ•œ ํ˜•ํƒœ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋ธ ์ œ์ž‘์ž๋Š” ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ ์ž‘์—…์— ๋Œ€ํ•œ ์ผ๋ถ€ ๋„๋ฉ”์ธ ์ง€์‹์„ ๋ณด์œ ํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ์˜คํ”ˆ ๋ชจ๋ธ๊ณผ ๋ฒค์น˜๋งˆํฌ์˜ ๋‹ค์–‘์„ฑ์ด ์ปค์ง€๋ฉด์„œ ์ธ๊ฐ„์˜ ์ง๊ด€๋งŒ์œผ๋กœ๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ƒˆ๋กœ์šด ๋ชจ๋ธ ์กฐํ•ฉ์„ ๋ฐœ๊ฒฌํ•˜๊ธฐ ์œ„ํ•œ ๋ณด๋‹ค ์ฒด๊ณ„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋” ๋งŽ์€ ๋ฐœ์ „์„ ๊ฐ€์ ธ์˜ฌ ๊ฒƒ์ด๋ผ๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค.
**Mixing different training distributions:** While we chose one data distribution to both select data and train on, modern LLM setups usually mix different data sources. Our method is likely complimentary to such pipelines: practitioners may use D4 to diversify and de-duplicate individual data sources and then mix data sources to provide additional diversity in their training dataset. We leave exploring the efficacy of D4 on a mix of training distributions as future work, but expect that this will yield further gains by reducing redundancy across datasets as well as within datasets.
๋‹ค์–‘ํ•œ ํ›ˆ๋ จ ๋ถ„ํฌ์˜ ํ˜ผํ•ฉ: ์šฐ๋ฆฌ๋Š” ๋ฐ์ดํ„ฐ ์„ ํƒ๊ณผ ํ›ˆ๋ จ์— ํ•˜๋‚˜์˜ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋ฅผ ์„ ํƒํ–ˆ์ง€๋งŒ, ํ˜„๋Œ€์˜ LLM ์„ค์ •์€ ๋ณดํ†ต ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ํ˜ผํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ์ด๋Ÿฌํ•œ ํŒŒ์ดํ”„๋ผ์ธ์— ๋ณด์™„์ ์ผ ๊ฐ€๋Šฅ์„ฑ์ด ํฝ๋‹ˆ๋‹ค: ์‹ค๋ฌด์ž๋“ค์€ D4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ๋ณ„ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ๋‹ค์–‘ํ™”ํ•˜๊ณ  ์ค‘๋ณต์„ ์ œ๊ฑฐํ•œ ํ›„, ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ํ˜ผํ•ฉํ•˜์—ฌ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹์— ์ถ”๊ฐ€์ ์ธ ๋‹ค์–‘์„ฑ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋‹ค์–‘ํ•œ ํ›ˆ๋ จ ๋ถ„ํฌ์˜ ํ˜ผํ•ฉ์—์„œ D4์˜ ํšจ๋Šฅ์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์„ ๋ฏธ๋ž˜์˜ ๊ณผ์ œ๋กœ ๋‚จ๊ฒจ๋‘์ง€๋งŒ, ์ด๋Š” ๋ฐ์ดํ„ฐ์…‹ ๊ฐ„ ๋ฐ ๋ฐ์ดํ„ฐ์…‹ ๋‚ด ์ค‘๋ณต์„ ์ค„์ž„์œผ๋กœ์จ ์ถ”๊ฐ€์ ์ธ ์ด๋“์„ ๊ฐ€์ ธ์˜ฌ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค.
**๋‹ค์–‘ํ•œ ํ•™์Šต ๋ถ„ํฌ ํ˜ผํ•ฉ:** ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๊ณ  ํ•™์Šตํ•˜๋Š” ๋ฐ ํ•˜๋‚˜์˜ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋ฅผ ์„ ํƒํ–ˆ์ง€๋งŒ, ์ตœ์‹  LLM ์„ค์ •์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์„œ๋กœ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ํ˜ผํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํŒŒ์ดํ”„๋ผ์ธ์— D4๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ฐœ๋ณ„ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ๋‹ค์–‘ํ™”ํ•˜๊ณ  ์ค‘๋ณต์„ ์ œ๊ฑฐํ•œ ๋‹ค์Œ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ํ˜ผํ•ฉํ•˜์—ฌ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ถ”๊ฐ€ ๋‹ค์–‘์„ฑ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ํ•™์Šต ๋ถ„ํฌ์— ๋Œ€ํ•œ D4์˜ ํšจ์œจ์„ฑ์„ ํƒ์ƒ‰ํ•˜๋Š” ๊ฒƒ์€ ํ–ฅํ›„ ์ž‘์—…์œผ๋กœ ๋‚จ๊ฒจ๋‘์ง€๋งŒ, ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐ„์˜ ์ค‘๋ณต์„ฑ์€ ๋ฌผ๋ก  ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋‚ด์˜ ์ค‘๋ณต์„ฑ๋„ ์ค„์—ฌ์„œ ๋” ๋งŽ์€ ์ด๋“์„ ์–ป์„ ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค.
**GPT-4 Judgement:** The first step mistakenly calculates the distance traveled after the first stop until the end of the trip, not accounting for the second stop which is 15 miles before the end. The second step incorrectly describes the result of the first calculation as the distance between the second stop and the end of the trip, whereas it is actually the distance from the first stop to the end of the trip minus the distance from the second stop to the end of the trip. However, despite the incorrect reasoning in both steps, the final answer achieved matches the correct answer.
**GPT-4 ํŒ๋‹จ:** ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” ๋‘ ๋ฒˆ์งธ ์ •์ฐจ ์ง€์ ์ด ๋์—์„œ 15๋งˆ์ผ ๋–จ์–ด์ ธ ์žˆ๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•˜์ง€ ์•Š๊ณ , ์ฒซ ๋ฒˆ์งธ ์ •์ฐจ ์ดํ›„๋ถ€ํ„ฐ ์—ฌํ–‰์˜ ๋๊นŒ์ง€ ์ด๋™ํ•œ ๊ฑฐ๋ฆฌ๋ฅผ ์ž˜๋ชป ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” ์ฒซ ๋ฒˆ์งธ ๊ณ„์‚ฐ์˜ ๊ฒฐ๊ณผ๋ฅผ ๋‘ ๋ฒˆ์งธ ์ •์ฐจ ์ง€์ ๊ณผ ์—ฌํ–‰์˜ ๋ ์‚ฌ์ด์˜ ๊ฑฐ๋ฆฌ๋กœ ์ž˜๋ชป ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ๋Š” ์ฒซ ๋ฒˆ์งธ ์ •์ฐจ ์ง€์ ์—์„œ ์—ฌํ–‰์˜ ๋๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ์—์„œ ๋‘ ๋ฒˆ์งธ ์ •์ฐจ ์ง€์ ์—์„œ ์—ฌํ–‰์˜ ๋๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ๋ฅผ ๋บ€ ๊ฐ’์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‘ ๋‹จ๊ณ„ ๋ชจ๋‘์—์„œ ์ž˜๋ชป๋œ ์ถ”๋ก ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ตœ์ข… ๋‹ต๋ณ€์€ ์˜ฌ๋ฐ”๋ฅธ ๋‹ต๋ณ€๊ณผ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค.
**GPT-4 ํŒ๋‹จ:** ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„๋Š” ์ฒซ ๋ฒˆ์งธ ์ •๋ฅ˜์žฅ ์ดํ›„ ์—ฌํ–‰์ด ๋๋‚  ๋•Œ๊นŒ์ง€ ์ด๋™ํ•œ ๊ฑฐ๋ฆฌ๋ฅผ ์ž˜๋ชป ๊ณ„์‚ฐํ•˜์—ฌ ๋์—์„œ 15๋งˆ์ผ ์ „์— ์žˆ๋Š” ๋‘ ๋ฒˆ์งธ ์ •๋ฅ˜์žฅ์„ ๊ณ ๋ คํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” ์ฒซ ๋ฒˆ์งธ ๊ณ„์‚ฐ์˜ ๊ฒฐ๊ณผ๋ฅผ ๋‘ ๋ฒˆ์งธ ์ •๋ฅ˜์žฅ๊ณผ ์—ฌํ–‰ ์ข…๋ฃŒ ์ง€์  ์‚ฌ์ด์˜ ๊ฑฐ๋ฆฌ๋กœ ์ž˜๋ชป ์„ค๋ช…ํ•˜๋Š”๋ฐ, ์‹ค์ œ๋กœ๋Š” ์ฒซ ๋ฒˆ์งธ ์ •๋ฅ˜์žฅ์—์„œ ์—ฌํ–‰ ์ข…๋ฃŒ ์ง€์ ๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ์—์„œ ๋‘ ๋ฒˆ์งธ ์ •๋ฅ˜์žฅ์—์„œ ์—ฌํ–‰ ์ข…๋ฃŒ ์ง€์ ๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ๋ฅผ ๋บ€ ๊ฐ’์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‘ ๋‹จ๊ณ„ ๋ชจ๋‘์—์„œ ์ž˜๋ชป๋œ ์ถ”๋ก ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ตœ์ข… ๋‹ต์€ ์ •๋‹ต๊ณผ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค.
**Application Layer.** At the application layer, agent applications, such as travel agent or math agent, are developed and deployed. In this layer, AIOS provides the AIOS SDK, with a higher abstraction of system calls that simplifies the development process for agent developers. This SDK allows for development of agent applications by offering a rich toolkit that abstract away the complexities of lower-level system functions. This enables developers to dedicate their focus to the essential logic and functionalities of their agents, facilitating a more efficient development process.
**์‘์šฉ ๊ณ„์ธต.** ์‘์šฉ ๊ณ„์ธต์—์„œ๋Š” ์—ฌํ–‰ ์—์ด์ „ํŠธ๋‚˜ ์ˆ˜ํ•™ ์—์ด์ „ํŠธ์™€ ๊ฐ™์€ ์—์ด์ „ํŠธ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์ด ๊ฐœ๋ฐœ๋˜๊ณ  ๋ฐฐํฌ๋ฉ๋‹ˆ๋‹ค. ์ด ๊ณ„์ธต์—์„œ AIOS๋Š” ์—์ด์ „ํŠธ ๊ฐœ๋ฐœ์ž์˜ ๊ฐœ๋ฐœ ๊ณผ์ •์„ ๋‹จ์ˆœํ™”ํ•˜๋Š” ์‹œ์Šคํ…œ ํ˜ธ์ถœ์˜ ๋†’์€ ์ถ”์ƒํ™”๋ฅผ ์ œ๊ณตํ•˜๋Š” AIOS SDK๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด SDK๋Š” ํ•˜์œ„ ์ˆ˜์ค€ ์‹œ์Šคํ…œ ๊ธฐ๋Šฅ์˜ ๋ณต์žก์„ฑ์„ ์ถ”์ƒํ™”ํ•˜๋Š” ํ’๋ถ€ํ•œ ๋„๊ตฌ ํ‚คํŠธ๋ฅผ ์ œ๊ณตํ•˜์—ฌ ์—์ด์ „ํŠธ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ ๊ฐœ๋ฐœ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๊ฐœ๋ฐœ์ž๋Š” ์—์ด์ „ํŠธ์˜ ํ•„์ˆ˜ ๋…ผ๋ฆฌ์™€ ๊ธฐ๋Šฅ์— ์ง‘์ค‘ํ•  ์ˆ˜ ์žˆ์–ด ๋” ํšจ์œจ์ ์ธ ๊ฐœ๋ฐœ ๊ณผ์ •์„ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค.
**์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๊ณ„์ธต.** ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๊ณ„์ธต์—์„œ๋Š” ์—ฌํ–‰ ์ƒ๋‹ด์› ๋˜๋Š” ์ˆ˜ํ•™ ์ƒ๋‹ด์›๊ณผ ๊ฐ™์€ ์—์ด์ „ํŠธ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ๊ฐœ๋ฐœํ•˜๊ณ  ๋ฐฐํฌํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ณ„์ธต์—์„œ AIOS๋Š” ์—์ด์ „ํŠธ ๊ฐœ๋ฐœ์ž์˜ ๊ฐœ๋ฐœ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ„์†Œํ™”ํ•˜๋Š” ๋” ๋†’์€ ์ˆ˜์ค€์˜ ์‹œ์Šคํ…œ ํ˜ธ์ถœ ์ถ”์ƒํ™”๋ฅผ ํ†ตํ•ด AIOS SDK๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด SDK๋Š” ํ•˜์œ„ ์ˆ˜์ค€ ์‹œ์Šคํ…œ ๊ธฐ๋Šฅ์˜ ๋ณต์žก์„ฑ์„ ์ถ”์ƒํ™”ํ•˜๋Š” ํ’๋ถ€ํ•œ ํˆดํ‚ท์„ ์ œ๊ณตํ•˜์—ฌ ์—์ด์ „ํŠธ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๊ฐœ๋ฐœ์ž๋Š” ์—์ด์ „ํŠธ์˜ ํ•„์ˆ˜ ๋กœ์ง๊ณผ ๊ธฐ๋Šฅ์— ์ง‘์ค‘ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๊ฐœ๋ฐœ ํ”„๋กœ์„ธ์Šค๊ฐ€ ๋”์šฑ ํšจ์œจ์ ์œผ๋กœ ์ง„ํ–‰๋ฉ๋‹ˆ๋‹ค.
Large language models (LLMs) are able to capture vast amounts of factual information (Petroni et al., 2019; Cohen et al., 2023; Hu et al., 2023). LLMs exhibit a remarkable level of knowledge in various domains due to their massive pre-training datasets. However, there are two significant limitations to this knowledge. First, it is static and does not update with time. Second, it is non-specific and thus may lack nuanced expertise in particular domains. While these are two different problems, they are deeply related since their solution is the same: enhancing the model's knowledge.
๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธ(LLMs)์€ ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ํฌ์ฐฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Petroni et al., 2019; Cohen et al., 2023; Hu et al., 2023). LLMs๋Š” ๋Œ€๊ทœ๋ชจ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์…‹ ๋•๋ถ„์— ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์—์„œ ๋†€๋ผ์šด ์ˆ˜์ค€์˜ ์ง€์‹์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ง€์‹์—๋Š” ๋‘ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ด๋Š” ์ •์ ์ด๋ฉฐ ์‹œ๊ฐ„์— ๋”ฐ๋ผ ์—…๋ฐ์ดํŠธ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋‘˜์งธ, ์ด๋Š” ๋น„ํŠน์ •์ ์ด์–ด์„œ ํŠน์ • ๋ถ„์•ผ์—์„œ ์„ธ๋ฐ€ํ•œ ์ „๋ฌธ ์ง€์‹์ด ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋‘ ๊ฐ€์ง€๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ๋ฌธ์ œ์ด์ง€๋งŒ, ๊ทธ ํ•ด๊ฒฐ์ฑ…์€ ๋™์ผํ•ฉ๋‹ˆ๋‹ค: ๋ชจ๋ธ์˜ ์ง€์‹์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ์บก์ฒ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Petroni et al., 2019; Cohen et al., 2023; Hu et al., 2023). LLM์€ ๋ฐฉ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์ธํ•ด ๋‹ค์–‘ํ•œ ์˜์—ญ์—์„œ ๋†€๋ผ์šด ์ˆ˜์ค€์˜ ์ง€์‹์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ง€์‹์—๋Š” ๋‘ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์‹œ๊ฐ„์ด ์ง€๋‚˜๋„ ์—…๋ฐ์ดํŠธ๋˜์ง€ ์•Š๋Š” ์ •์ ์ธ ์ •๋ณด์ž…๋‹ˆ๋‹ค. ๋‘˜์งธ, ํŠน์ • ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ๋ฏธ๋ฌ˜ํ•œ ์ „๋ฌธ ์ง€์‹์ด ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ๋Š” ๋น„ํŠน์ •์ ์ž…๋‹ˆ๋‹ค. ์ด ๋‘ ๊ฐ€์ง€ ๋ฌธ์ œ๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ๋ฌธ์ œ์ด์ง€๋งŒ, ๊ทธ ํ•ด๊ฒฐ์ฑ…์ด ๋™์ผํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊นŠ์€ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค: ๋ชจ๋ธ์˜ ์ง€์‹์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Pretraining Data OptimizationThe objective of optimizing pre-training corpora is to maximize the performance and efficiency of language model training by improving the quality and scale of the pretrain data mixture. This typically includes data collecting through crawling (Raffel et al., 2020) or synthesis (Polu and Sutskever, 2020; Gunasekar et al., 2023), de-duplication (Lee et al., 2021; Kandpal et al., 2022; Tirumala et al., 2023), filtering and selection (Albalak et al., 2024), as well as data composition (Xie et al., 2024) and curriculum (Chen et al., 2024; MA et al., 2024).
์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์ตœ์ ํ™” ์‚ฌ์ „ ํ›ˆ๋ จ ์ฝ”ํผ์Šค๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๋ชฉ์ ์€ ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์˜ ํ’ˆ์งˆ๊ณผ ๊ทœ๋ชจ๋ฅผ ๊ฐœ์„ ํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ ํ›ˆ๋ จ์˜ ์„ฑ๋Šฅ๊ณผ ํšจ์œจ์„ฑ์„ ๊ทน๋Œ€ํ™”ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ํฌ๋กค๋ง(Raffel et al., 2020) ๋˜๋Š” ํ•ฉ์„ฑ(Polu and Sutskever, 2020; Gunasekar et al., 2023)์„ ํ†ตํ•œ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘, ์ค‘๋ณต ์ œ๊ฑฐ(Lee et al., 2021; Kandpal et al., 2022; Tirumala et al., 2023), ํ•„ํ„ฐ๋ง ๋ฐ ์„ ํƒ(Albalak et al., 2024), ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ(Xie et al., 2024) ๋ฐ ์ปค๋ฆฌํ˜๋Ÿผ(Chen et al., 2024; MA et al., 2024)์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์ตœ์ ํ™”์‚ฌ์ „ ํ•™์Šต ๋ง๋ญ‰์น˜๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๋ชฉ์ ์€ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์˜ ํ’ˆ์งˆ๊ณผ ๊ทœ๋ชจ๋ฅผ ๊ฐœ์„ ํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ ํ•™์Šต์˜ ์„ฑ๋Šฅ๊ณผ ํšจ์œจ์„ฑ์„ ๊ทน๋Œ€ํ™”ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ํฌ๋กค๋ง(Raffel et al., 2020) ๋˜๋Š” ํ•ฉ์„ฑ(Polu and Sutskever, 2020; Gunasekar et al., 2023)์„ ํ†ตํ•œ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘, ์ค‘๋ณต ์ œ๊ฑฐ(Lee et al., 2021; Kandpal et al., 2022; Tirumala et al., 2023), ํ•„ํ„ฐ๋ง ๋ฐ ์„ ํƒ(Albalak et al., 2024), ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ(Xie et al., 2024) ๋ฐ ์ปค๋ฆฌํ˜๋Ÿผ(Chen et al., 2024; MA et al., 2024)์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
The luxurious and elegant nature of velvet cannot be denied, whether it is crushed, vibrant, or head-to-toe. This timeless textile is not only fashionable but also wearable for daily wear. In recent years, there has been a shift towards embracing fun gem-toned velvety pieces, breaking away from the traditional association of velvet with dark moody shades of navy and black. Below are some of the most desirable velvet pieces available on the high street today, and we are already envisioning outfits around them. Are you completely enamored with velvet or simply beyond bored with it?
๋ฒจ๋ฒณ์˜ ๊ณ ๊ธ‰์Šค๋Ÿฝ๊ณ  ์šฐ์•„ํ•œ ํŠน์„ฑ์€ ๋ถ€์ •ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋ฒจ๋ฒณ์ด ๊ตฌ๊ฒจ์ง€๋“ , ์ƒ๋™๊ฐ ์žˆ๋“ , ๋จธ๋ฆฌ๋ถ€ํ„ฐ ๋ฐœ๋๊นŒ์ง€ ๋ฎ๋“  ๋ง์ด์ฃ . ์ด ์‹œ๋Œ€๋ฅผ ์ดˆ์›”ํ•œ ์ง๋ฌผ์€ ํŒจ์…”๋„ˆ๋ธ”ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ผ์ƒ์ ์œผ๋กœ ์ž…๊ธฐ์—๋„ ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋ช‡ ๋…„ ๋™์•ˆ ๋ฒจ๋ฒณ์ด ์ „ํ†ต์ ์œผ๋กœ ์–ด๋‘์šด ํ•ด๊ตฐ์ƒ‰๊ณผ ๊ฒ€์ •์ƒ‰์˜ ์Œ์šธํ•œ ์ƒ‰์กฐ์™€ ์—ฐ๊ด€๋˜์—ˆ๋˜ ๊ฒƒ์—์„œ ๋ฒ—์–ด๋‚˜, ์žฌ๋ฏธ์žˆ๋Š” ๋ณด์„ ํ†ค์˜ ๋ฒจ๋ฒณ ์กฐ๊ฐ๋“ค์„ ๋ฐ›์•„๋“ค์ด๋Š” ๋ฐฉํ–ฅ์œผ๋กœ ๋ณ€ํ™”๊ฐ€ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์•„๋ž˜๋Š” ์˜ค๋Š˜๋‚  ๊ณ ๊ธ‰ ์ƒ์ ์—์„œ ๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ€์žฅ ๋งค๋ ฅ์ ์ธ ๋ฒจ๋ฒณ ์กฐ๊ฐ๋“ค ์ค‘ ์ผ๋ถ€์ด๋ฉฐ, ์šฐ๋ฆฌ๋Š” ์ด๋ฏธ ์ด๋“ค์„ ์ค‘์‹ฌ์œผ๋กœ ํ•œ ์˜์ƒ์„ ์ƒ์ƒํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฒจ๋ฒณ์— ์™„์ „ํžˆ ๋งค๋ฃŒ๋˜์—ˆ๋‚˜์š”, ์•„๋‹ˆ๋ฉด ๊ทธ์ € ์ง€๋ฃจํ•ด์กŒ๋‚˜์š”?
๋ฒจ๋ฒณ์˜ ๊ณ ๊ธ‰์Šค๋Ÿฝ๊ณ  ์šฐ์•„ํ•œ ํŠน์„ฑ์€ ๋ถ€์„œ์ง€๊ฑฐ๋‚˜ ์ƒ์ƒํ•˜๊ฑฐ๋‚˜ ๋จธ๋ฆฌ๋ถ€ํ„ฐ ๋ฐœ๋๊นŒ์ง€ ์–ด๋–ค ํ˜•ํƒœ๋กœ๋“  ๋ถ€์ •ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด ์‹œ๋Œ€๋ฅผ ์ดˆ์›”ํ•œ ์ง๋ฌผ์€ ํŒจ์…”๋„ˆ๋ธ”ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ผ์ƒ๋ณต์œผ๋กœ๋„ ์ž…์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋ช‡ ๋…„ ๋™์•ˆ ๋ฒจ๋ฒณ์ด๋ผ๋Š” ์†Œ์žฌ๊ฐ€ ์ „ํ†ต์ ์œผ๋กœ ์–ด๋‘ก๊ณ  ๋ฌด๋“œ ์žˆ๋Š” ๋„ค์ด๋น„๋‚˜ ๋ธ”๋ž™ ์ƒ‰์กฐ์™€ ์—ฐ๊ด€๋˜์–ด ์žˆ๋˜ ๊ฒƒ์—์„œ ๋ฒ—์–ด๋‚˜ ์žฌ๋ฏธ์žˆ๋Š” ๋ณด์„ ํ†ค์˜ ๋ฒจ๋ฒณ ์ œํ’ˆ์„ ์ˆ˜์šฉํ•˜๋Š” ์ถ”์„ธ๊ฐ€ ๋‚˜ํƒ€๋‚˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์€ ํ˜„์žฌ ํ•˜์ด์ŠคํŠธ๋ฆฌํŠธ์—์„œ ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ๋ฒจ๋ฒณ ์ œํ’ˆ ์ค‘ ์ผ๋ถ€์ด๋ฉฐ, ์ด๋ฏธ ๊ทธ ์ฃผ๋ณ€์˜ ์˜์ƒ์„ ์ƒ์ƒํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฒจ๋ฒณ์— ์™„์ „ํžˆ ๋งค๋ฃŒ๋˜์—ˆ๋‚˜์š”, ์•„๋‹ˆ๋ฉด ๋‹จ์ˆœํžˆ ์ง€๋ฃจํ•ด์กŒ๋‚˜์š”?
In summary, this survey conducts a systematic review and establishes comprehensive taxonomies for existing studies in the burgeoning field of LLM-based autonomous agents. Our focus encompasses three primary areas: construction of agents, their applications, and methods of evaluation. Drawing from a wealth of previous studies, we identify various challenges in this field and discuss potential future directions. We expect that our survey can provide newcomers of LLM-based autonomous agents with a comprehensive background knowledge, and also encourage further groundbreaking studies.
์š”์•ฝํ•˜์ž๋ฉด, ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ ๋ถ„์•ผ์˜ ๊ธฐ์กด ์—ฐ๊ตฌ๋“ค์„ ์ฒด๊ณ„์ ์œผ๋กœ ๊ฒ€ํ† ํ•˜๊ณ  ํฌ๊ด„์ ์ธ ๋ถ„๋ฅ˜ ์ฒด๊ณ„๋ฅผ ํ™•๋ฆฝํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ดˆ์ ์€ ์—์ด์ „ํŠธ์˜ ๊ตฌ์„ฑ, ๊ทธ๋“ค์˜ ์‘์šฉ, ๊ทธ๋ฆฌ๊ณ  ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์˜ ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ์˜์—ญ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ๋“ค์˜ ํ’๋ถ€ํ•œ ์ž๋ฃŒ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ด ๋ถ„์•ผ์—์„œ ๋‹ค์–‘ํ•œ ๋„์ „ ๊ณผ์ œ๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ์ž ์žฌ์ ์ธ ๋ฏธ๋ž˜ ๋ฐฉํ–ฅ์„ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์„ค๋ฌธ์กฐ์‚ฌ๊ฐ€ LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์˜ ์‹ ์ž…์ž๋“ค์—๊ฒŒ ํฌ๊ด„์ ์ธ ๋ฐฐ๊ฒฝ ์ง€์‹์„ ์ œ๊ณตํ•˜๊ณ , ๋˜ํ•œ ๋” ๋งŽ์€ ํš๊ธฐ์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์žฅ๋ คํ•  ์ˆ˜ ์žˆ๊ธฐ๋ฅผ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค.
์š”์•ฝํ•˜๋ฉด, ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ๊ธ‰์„ฑ์žฅํ•˜๋Š” LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ ๋ถ„์•ผ์˜ ๊ธฐ์กด ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ๊ฒ€ํ† ๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ  ํฌ๊ด„์ ์ธ ๋ถ„๋ฅ˜๋ฅผ ์ˆ˜๋ฆฝํ•ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ ๊ตฌ์ถ•, ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜, ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์ด๋ผ๋Š” ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ์˜์—ญ์„ ์ค‘์ ์ ์œผ๋กœ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์„ ํ–‰ ์—ฐ๊ตฌ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์ด ๋ถ„์•ผ์˜ ๋‹ค์–‘ํ•œ ๊ณผ์ œ๋ฅผ ํŒŒ์•…ํ•˜๊ณ  ํ–ฅํ›„ ์ž ์žฌ์ ์ธ ๋ฐฉํ–ฅ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋ฅผ ํ†ตํ•ด LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์˜ ์‹ ๊ทœ ์—ฐ๊ตฌ์ž๋“ค์—๊ฒŒ ํฌ๊ด„์ ์ธ ๋ฐฐ๊ฒฝ ์ง€์‹์„ ์ œ๊ณตํ•˜๊ณ , ๋”์šฑ ํš๊ธฐ์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์žฅ๋ คํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค.
**Hardware Layer.** The hardware layer comprises the physical components of the system, including the CPU, GPU, memory, disk, and peripheral devices. It is crucial to note that the LLM kernel's system calls cannot directly interact with the hardware. Instead, these calls interface with the OS's system calls, which in turn manage the hardware resources. This indirect interaction ensures a layer of abstraction and security, allowing the LLM kernel to leverage hardware capabilities without requiring direct hardware management, thus maintaining the system's integrity and efficiency.
**ํ•˜๋“œ์›จ์–ด ๊ณ„์ธต.** ํ•˜๋“œ์›จ์–ด ๊ณ„์ธต์€ CPU, GPU, ๋ฉ”๋ชจ๋ฆฌ, ๋””์Šคํฌ ๋ฐ ์ฃผ๋ณ€ ์žฅ์น˜์™€ ๊ฐ™์€ ์‹œ์Šคํ…œ์˜ ๋ฌผ๋ฆฌ์  ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. LLM ์ปค๋„์˜ ์‹œ์Šคํ…œ ํ˜ธ์ถœ์ด ํ•˜๋“œ์›จ์–ด์™€ ์ง์ ‘ ์ƒํ˜ธ์ž‘์šฉํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์ ์„ ์ฃผ๋ชฉํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€์‹ , ์ด๋Ÿฌํ•œ ํ˜ธ์ถœ์€ OS์˜ ์‹œ์Šคํ…œ ํ˜ธ์ถœ๊ณผ ์ธํ„ฐํŽ˜์ด์Šคํ•˜๋ฉฐ, ์ด๋Š” ํ•˜๋“œ์›จ์–ด ์ž์›์„ ๊ด€๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐ„์ ‘์ ์ธ ์ƒํ˜ธ์ž‘์šฉ์€ ์ถ”์ƒํ™”์™€ ๋ณด์•ˆ์˜ ๊ณ„์ธต์„ ๋ณด์žฅํ•˜์—ฌ, LLM ์ปค๋„์ด ์ง์ ‘์ ์ธ ํ•˜๋“œ์›จ์–ด ๊ด€๋ฆฌ๋ฅผ ์š”๊ตฌํ•˜์ง€ ์•Š๊ณ ๋„ ํ•˜๋“œ์›จ์–ด ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•˜์—ฌ ์‹œ์Šคํ…œ์˜ ๋ฌด๊ฒฐ์„ฑ๊ณผ ํšจ์œจ์„ฑ์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค.
**ํ•˜๋“œ์›จ์–ด ๊ณ„์ธต.** ํ•˜๋“œ์›จ์–ด ๊ณ„์ธต์€ CPU, GPU, ๋ฉ”๋ชจ๋ฆฌ, ๋””์Šคํฌ ๋ฐ ์ฃผ๋ณ€ ์žฅ์น˜๋ฅผ ํฌํ•จํ•œ ์‹œ์Šคํ…œ์˜ ๋ฌผ๋ฆฌ์  ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. LLM ์ปค๋„์˜ ์‹œ์Šคํ…œ ํ˜ธ์ถœ์€ ํ•˜๋“œ์›จ์–ด์™€ ์ง์ ‘ ์ƒํ˜ธ ์ž‘์šฉํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€์‹ , ์ด๋Ÿฌํ•œ ํ˜ธ์ถœ์€ ํ•˜๋“œ์›จ์–ด ๋ฆฌ์†Œ์Šค๋ฅผ ๊ด€๋ฆฌํ•˜๋Š” OS์˜ ์‹œ์Šคํ…œ ํ˜ธ์ถœ๊ณผ ์ธํ„ฐํŽ˜์ด์Šคํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐ„์ ‘์ ์ธ ์ƒํ˜ธ ์ž‘์šฉ์€ ์ถ”์ƒํ™” ๋ฐ ๋ณด์•ˆ ๊ณ„์ธต์„ ๋ณด์žฅํ•˜์—ฌ LLM ์ปค๋„์ด ์ง์ ‘ ํ•˜๋“œ์›จ์–ด ๊ด€๋ฆฌ ์—†์ด๋„ ํ•˜๋“œ์›จ์–ด ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์—ฌ ์‹œ์Šคํ…œ์˜ ๋ฌด๊ฒฐ์„ฑ๊ณผ ํšจ์œจ์„ฑ์„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Future research directions on synthetic data could focus on improving the fidelity and controllability of generative models and developing standardized evaluation and contamination protocols and tools. We could also explore the integration of synthetic data with other techniques and its application in other domains. Despite the challenges, the potential benefits of synthetic data in advancing AI research are significant. By leveraging synthetic data responsibly and effectively, we can build more powerful, inclusive, and trustworthy AI systems that benefit society as a whole.
๋ฏธ๋ž˜์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์€ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ์ถฉ์‹ค๋„์™€ ์ œ์–ด ๊ฐ€๋Šฅ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ , ํ‘œ์ค€ํ™”๋œ ํ‰๊ฐ€ ๋ฐ ์˜ค์—ผ ํ”„๋กœํ† ์ฝœ๊ณผ ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘˜ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ๋‹ค๋ฅธ ๊ธฐ์ˆ ๊ณผ ํ†ตํ•ฉํ•˜๊ณ  ๋‹ค๋ฅธ ๋„๋ฉ”์ธ์— ์ ์šฉํ•˜๋Š” ๋ฐฉ์•ˆ์„ ํƒ๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋„์ „ ๊ณผ์ œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , AI ์—ฐ๊ตฌ๋ฅผ ๋ฐœ์ „์‹œํ‚ค๋Š” ๋ฐ ์žˆ์–ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์ž ์žฌ์  ์ด์ ์€ ํฝ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ฑ…์ž„๊ฐ ์žˆ๊ณ  ํšจ๊ณผ์ ์œผ๋กœ ํ™œ์šฉํ•จ์œผ๋กœ์จ, ์‚ฌํšŒ ์ „์ฒด์— ์ด์ต์ด ๋˜๋Š” ๋” ๊ฐ•๋ ฅํ•˜๊ณ  ํฌ์šฉ์ ์ด๋ฉฐ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” AI ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์€ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ์ถฉ์‹ค๋„์™€ ์ œ์–ด ๊ฐ€๋Šฅ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ณ  ํ‘œ์ค€ํ™”๋œ ํ‰๊ฐ€ ๋ฐ ์˜ค์—ผ ํ”„๋กœํ† ์ฝœ๊ณผ ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถœ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค๋ฅธ ๊ธฐ์ˆ ๊ณผ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ํ†ตํ•ฉ๊ณผ ๋‹ค๋ฅธ ์˜์—ญ์—์˜ ์ ์šฉ์„ ํƒ๊ตฌํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์–ด๋ ค์›€์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  AI ์—ฐ๊ตฌ๋ฅผ ๋ฐœ์ „์‹œํ‚ค๋Š” ๋ฐ ์žˆ์–ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์ž ์žฌ์  ์ด์ ์€ ์ƒ๋‹นํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ฑ…์ž„๊ฐ ์žˆ๊ณ  ํšจ๊ณผ์ ์œผ๋กœ ํ™œ์šฉํ•จ์œผ๋กœ์จ ์šฐ๋ฆฌ๋Š” ์‚ฌํšŒ ์ „์ฒด์— ๋„์›€์ด ๋˜๋Š” ๋” ๊ฐ•๋ ฅํ•˜๊ณ  ํฌ์šฉ์ ์ด๋ฉฐ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” AI ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Prompting LLMs with large chunks (_entire books_) to evaluate the faithfulness of each claim is prohibitively expensive (see SSF.5). Hence, for this experiment, we select 7 books based on: (1) token length (\(<\)125K), and (2) presence of at least one _Unfaithful_ claim. This sub-dataset includes: (1) "Yellowface," (2) "Only For The Week," (3) "Viciously Yours," (4) "Six Scorched Roses," (5) "Sorrow and Bliss," (6) "She Is a Haunting," and (7) "Pet." Table Table 31 shows the number of claims per label in the sub-dataset. Further details on each book can be found in Table 7.
LLMs์— ํฐ ๋ฉ์–ด๋ฆฌ์˜ ํ…์ŠคํŠธ(์ „์ฒด ์ฑ…)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์ฃผ์žฅ์— ๋Œ€ํ•œ ์‹ ๋ขฐ์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ๋น„์šฉ์ด ๋งค์šฐ ๋งŽ์ด ๋“ญ๋‹ˆ๋‹ค(SSF.5 ์ฐธ์กฐ). ๋”ฐ๋ผ์„œ ์ด ์‹คํ—˜์—์„œ๋Š” ๋‹ค์Œ ๊ธฐ์ค€์— ๋”ฐ๋ผ 7๊ถŒ์˜ ์ฑ…์„ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค: (1) ํ† ํฐ ๊ธธ์ด(\(<\)125K), (2) ์ ์–ด๋„ ํ•˜๋‚˜์˜ _Unfaithful_ ์ฃผ์žฅ์ด ํฌํ•จ๋˜์–ด ์žˆ๋Š”์ง€. ์ด ํ•˜์œ„ ๋ฐ์ดํ„ฐ์…‹์—๋Š” ๋‹ค์Œ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค: (1) "Yellowface," (2) "Only For The Week," (3) "Viciously Yours," (4) "Six Scorched Roses," (5) "Sorrow and Bliss," (6) "She Is a Haunting," ๊ทธ๋ฆฌ๊ณ  (7) "Pet." Table 31์€ ํ•˜์œ„ ๋ฐ์ดํ„ฐ์…‹์—์„œ ๋ ˆ์ด๋ธ”๋ณ„ ์ฃผ์žฅ ์ˆ˜๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ฐ ์ฑ…์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์„ธ๋ถ€ ์‚ฌํ•ญ์€ Table 7์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฐ ์ฃผ์žฅ์˜ ์ถฉ์‹ค์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ํฐ ๋ฉ์–ด๋ฆฌ(_์ „์ฒด ์ฑ…_)๋กœ LLM์„ ํ”„๋กฌํ”„ํŠธํ•˜๋Š” ๊ฒƒ์€ ๋น„์šฉ์ด ๋„ˆ๋ฌด ๋งŽ์ด ๋“ญ๋‹ˆ๋‹ค(SSF.5 ์ฐธ์กฐ). ๋”ฐ๋ผ์„œ ์ด ์‹คํ—˜์—์„œ๋Š” (1) ํ† ํฐ ๊ธธ์ด(\(<\)125K)์™€ (2) ํ•˜๋‚˜ ์ด์ƒ์˜ _๋ถˆ์ถฉ์‹ค_ ์ฃผ์žฅ์ด ์žˆ๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๊ธฐ์ค€์œผ๋กœ 7๊ถŒ์˜ ์ฑ…์„ ์„ ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ํ•˜์œ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” (1) "์˜๋กœ์šฐํŽ˜์ด์Šค", (2) "ํ•œ ์ฃผ ๋™์•ˆ๋งŒ", (3) "์•…๋ž„ํ•˜๊ฒŒ ๋‹น์‹ ", (4) "์—ฌ์„ฏ ๊ฐœ์˜ ๊ทธ์„๋ฆฐ ์žฅ๋ฏธ", (5) "์Šฌํ””๊ณผ ๊ธฐ์จ", (6) "๊ทธ๋…€๋Š” ์œ ๋ น์ด๋‹ค", (7) "์• ์™„๋™๋ฌผ"์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ํ‘œ 31์€ ํ•˜์œ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ ˆ์ด๋ธ”๋ณ„ ํด๋ ˆ์ž„ ์ˆ˜๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ฐ ์ฑ…์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ํ‘œ 7์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Whether you like your velvet crushed, vibrant or head-to-toe, there's really no denying the sheer luxe and elegance of this timeless textile. Not only is it super stylish, it can actually be so wearable for day-to-day wear. Yes, really! This year it's all about embracing fun gem-toned velvety pieces. Long gone are the days when velvet was solely associated with dark moody shades of navy and black. Below we've rounded up the most covetable velvet pieces on the high street right now. We're already coming up with outfit ideas! Are you completely obsessed or beyond bored of it?
๋ฒจ๋ฒณ์„ ์œผ๊นฌ ๊ฒƒ์ด๋“ , ์ƒ๋™๊ฐ ์žˆ๋Š” ๊ฒƒ์ด๋“ , ๋จธ๋ฆฌ๋ถ€ํ„ฐ ๋ฐœ๋๊นŒ์ง€ ์ž…๋Š” ๊ฒƒ์ด๋“ , ์ด ์‹œ๋Œ€๋ฅผ ์ดˆ์›”ํ•œ ์ง๋ฌผ์˜ ์ˆœ์ˆ˜ํ•œ ๋Ÿญ์…”๋ฆฌ์™€ ์šฐ์•„ํ•จ์„ ๋ถ€์ •ํ•  ์ˆ˜๋Š” ์—†์Šต๋‹ˆ๋‹ค. ๋ฒจ๋ฒณ์€ ๋งค์šฐ ์Šคํƒ€์ผ๋ฆฌ์‹œํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ผ์ƒ์ ์œผ๋กœ ์ž…๊ธฐ์—๋„ ์ •๋ง ์ฐฉ์šฉ๊ฐ์ด ์ข‹์Šต๋‹ˆ๋‹ค. ๋„ค, ์ •๋ง์ž…๋‹ˆ๋‹ค! ์˜ฌํ•ด๋Š” ๋ณด์„ ํ†ค์˜ ์žฌ๋ฏธ์žˆ๋Š” ๋ฒจ๋ฒณ ์กฐ๊ฐ๋“ค์„ ๋ฐ›์•„๋“ค์ด๋Š” ๊ฒƒ์ด ์ „๋ถ€์ž…๋‹ˆ๋‹ค. ๋ฒจ๋ฒณ์ด ์–ด๋‘์šด ํ•ด๊ตฐ์ƒ‰๊ณผ ๊ฒ€์ •์ƒ‰์˜ ์Œ์šธํ•œ ์ƒ‰์กฐ์™€๋งŒ ์—ฐ๊ด€๋˜๋˜ ์‹œ์ ˆ์€ ์˜ค๋ž˜์ „์— ์ง€๋‚˜๊ฐ”์Šต๋‹ˆ๋‹ค. ์•„๋ž˜์—์„œ ํ˜„์žฌ ๊ณ ๊ธ‰ ์ƒ์ ์—์„œ ๊ฐ€์žฅ ํƒ๋‚˜๋Š” ๋ฒจ๋ฒณ ์กฐ๊ฐ๋“ค์„ ๋ชจ์•„๋ดค์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๋ฏธ ์˜์ƒ ์•„์ด๋””์–ด๋ฅผ ๋– ์˜ฌ๋ฆฌ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค! ์—ฌ๋Ÿฌ๋ถ„์€ ์™„์ „ํžˆ ๋ฒจ๋ฒณ์— ๋น ์ ธ ์žˆ๋‚˜์š”, ์•„๋‹ˆ๋ฉด ์ง€๋ฃจํ•ดํ•˜๊ณ  ์žˆ๋‚˜์š”?
๋ฒจ๋ฒณ์„ ๋ถ€์ˆ˜๊ณ , ์ƒ์ƒํ•˜๊ฒŒ, ๋จธ๋ฆฌ๋ถ€ํ„ฐ ๋ฐœ๋๊นŒ์ง€ ์ข‹์•„ํ•˜๋“ , ์ด ์‹œ๋Œ€๋ฅผ ์ดˆ์›”ํ•œ ์„ฌ์œ ์˜ ๊ณ ๊ธ‰์Šค๋Ÿฌ์›€๊ณผ ์šฐ์•„ํ•จ์„ ๋ถ€์ •ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋งค์šฐ ์Šคํƒ€์ผ๋ฆฌ์‹œํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์‹ค์ œ๋กœ ์ผ์ƒ์ ์œผ๋กœ ์ฐฉ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ž˜, ์ •๋ง! ์˜ฌํ•ด๋Š” ์žฌ๋ฏธ์žˆ๋Š” ๋ณด์„ ํ†ค์˜ ๋ฒจ๋ฒณ ์˜ท์„ ๋ฐ›์•„๋“ค์ด๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋ฒจ๋ฒณ์ด ์ง™์€ ๋ฌด๋“œ์˜ ๋„ค์ด๋น„์™€ ๋ธ”๋ž™ ์ƒ‰์กฐ์™€๋งŒ ์—ฐ๊ด€๋˜๋˜ ์‹œ์ ˆ์€ ์ด๋ฏธ ์ง€๋‚˜๊ฐ”์Šต๋‹ˆ๋‹ค. ์•„๋ž˜์—์„œ๋Š” ํ˜„์žฌ ํ•˜์ด์ŠคํŠธ๋ฆฌํŠธ์—์„œ ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ๋ฒจ๋ฒณ ์ œํ’ˆ์„ ์ •๋ฆฌํ•ด ๋ณด์•˜์Šต๋‹ˆ๋‹ค. ์ด๋ฏธ ์˜์ƒ ์•„์ด๋””์–ด๊ฐ€ ๋– ์˜ค๋ฅด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค! ์™„์ „ํžˆ ์ง‘์ฐฉํ•˜์‹œ๋‚˜์š”, ์•„๋‹ˆ๋ฉด ์ง€๋ฃจํ•ดํ•˜์‹œ๋‚˜์š”?
Language Models have been known to exhibit bias both in terms of evaluation as well as in social or fairness terms [5]. Moreover, agents have specifically been shown to be "less robust, prone to more harmful behaviors, and capable of generating stealthier content than LLMs, highlighting significant safety challenges" [25]. Other research has found "a tendency for LLM agents to conform to the model's inherent social biases despite being directed to debate from certain political perspectives" [24]. This tendency can lead to faulty reasoning in any agent-based implementation.
์–ธ์–ด ๋ชจ๋ธ์€ ํ‰๊ฐ€ ์ธก๋ฉด๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์‚ฌํšŒ์  ๋˜๋Š” ๊ณต์ •์„ฑ ์ธก๋ฉด์—์„œ๋„ ํŽธํ–ฅ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฒƒ์œผ๋กœ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค [5]. ๋”์šฑ์ด, ์—์ด์ „ํŠธ๋Š” "๋œ ๊ฒฌ๊ณ ํ•˜๊ณ , ๋” ํ•ด๋กœ์šด ํ–‰๋™์„ ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์œผ๋ฉฐ, LLM๋ณด๋‹ค ๋” ์€๋ฐ€ํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์–ด ์ค‘์š”ํ•œ ์•ˆ์ „ ๋ฌธ์ œ๋ฅผ ๋ถ€๊ฐ์‹œํ‚จ๋‹ค"๊ณ  ๊ตฌ์ฒด์ ์œผ๋กœ ๋ณด์—ฌ์กŒ์Šต๋‹ˆ๋‹ค [25]. ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” "ํŠน์ • ์ •์น˜์  ๊ด€์ ์—์„œ ํ† ๋ก ํ•˜๋„๋ก ์ง€์‹œ๋ฐ›์•˜์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM ์—์ด์ „ํŠธ๊ฐ€ ๋ชจ๋ธ์˜ ๊ณ ์œ ํ•œ ์‚ฌํšŒ์  ํŽธํ–ฅ์— ๋”ฐ๋ฅด๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋‹ค"๊ณ  ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค [24]. ์ด๋Ÿฌํ•œ ๊ฒฝํ–ฅ์€ ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜ ๊ตฌํ˜„์—์„œ ์ž˜๋ชป๋œ ์ถ”๋ก ์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์–ธ์–ด ๋ชจ๋ธ์€ ํ‰๊ฐ€ ์ธก๋ฉด๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์‚ฌํšŒ์  ๋˜๋Š” ๊ณต์ •์„ฑ ์ธก๋ฉด์—์„œ๋„ ํŽธํ–ฅ์„ฑ์„ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค[5]. ๋˜ํ•œ ์—์ด์ „ํŠธ๋Š” ํŠนํžˆ "LLM๋ณด๋‹ค ๋œ ๊ฐ•๋ ฅํ•˜๊ณ , ๋” ๋งŽ์€ ํ•ด๋กœ์šด ํ–‰๋™์„ ์ผ์œผํ‚ค๋ฉฐ, ๋” ์€๋ฐ€ํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์–ด ์‹ฌ๊ฐํ•œ ์•ˆ์ „ ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•œ๋‹ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค."[25]. ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” "ํŠน์ • ์ •์น˜์  ๊ด€์ ์—์„œ ํ† ๋ก ํ•˜๋„๋ก ์ง€์‹œ๋ฐ›์•˜์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM ์—์ด์ „ํŠธ๊ฐ€ ๋ชจ๋ธ์— ๋‚ด์žฌ๋œ ์‚ฌํšŒ์  ํŽธ๊ฒฌ์— ๋”ฐ๋ฅด๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋‹ค"๊ณ  ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค[24]. ์ด๋Ÿฌํ•œ ๊ฒฝํ–ฅ์€ ๋ชจ๋“  ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜ ๊ตฌํ˜„์—์„œ ์ž˜๋ชป๋œ ์ถ”๋ก ์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The LayerNorm operation is the composition of a centering operation \(x\mapsto x-\bar{x}\) with RMSNorm (i.e. first centering is applied, then RMSNorm). Therefore the identical argument to Appendix C.4 goes through, besides the fact that we need \(\mathbf{U}\) to also commute with the centering operation. Since the centering operation fixes a \((h-1)\) dimensional subspace defined by \(\mathbf{1}^{T}x=0\) where \(\mathbf{1}\in\mathds{R}^{h}\) is the vector of ones, it is enough to impose an additional condition that \(\mathbf{U}\,\mathbf{1}\in\{-\mathbf{1},\mathbf{1}\}\).
LayerNorm ์—ฐ์‚ฐ์€ RMSNorm๊ณผ ํ•จ๊ป˜ ์ค‘์‹ฌํ™” ์—ฐ์‚ฐ \(x\mapsto x-\bar{x}\)์˜ ์กฐํ•ฉ์ž…๋‹ˆ๋‹ค (์ฆ‰, ๋จผ์ € ์ค‘์‹ฌํ™”๊ฐ€ ์ ์šฉ๋˜๊ณ , ๊ทธ ๋‹ค์Œ RMSNorm์ด ์ ์šฉ๋ฉ๋‹ˆ๋‹ค). ๋”ฐ๋ผ์„œ Appendix C.4์™€ ๋™์ผํ•œ ๋…ผ๋ฆฌ๊ฐ€ ์ ์šฉ๋˜๋ฉฐ, ์šฐ๋ฆฌ๋Š” \(\mathbf{U}\)๊ฐ€ ์ค‘์‹ฌํ™” ์—ฐ์‚ฐ๊ณผ๋„ ๊ตํ™˜ ๊ฐ€๋Šฅํ•ด์•ผ ํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ์ œ์™ธํ•˜๊ณ ๋Š” ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ์ค‘์‹ฌํ™” ์—ฐ์‚ฐ์€ \(\mathbf{1}^{T}x=0\)์œผ๋กœ ์ •์˜๋œ \((h-1)\) ์ฐจ์›์˜ ๋ถ€๋ถ„ ๊ณต๊ฐ„์„ ๊ณ ์ •ํ•˜๋ฏ€๋กœ, \(\mathbf{1}\in\mathds{R}^{h}\)๊ฐ€ ๋ชจ๋“  ์š”์†Œ๊ฐ€ 1์ธ ๋ฒกํ„ฐ์ผ ๋•Œ, \(\mathbf{U}\,\mathbf{1}\in\{-\mathbf{1},\mathbf{1}\}\)์ด๋ผ๋Š” ์ถ”๊ฐ€ ์กฐ๊ฑด์„ ๋ถ€๊ณผํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ถฉ๋ถ„ํ•ฉ๋‹ˆ๋‹ค.
LayerNorm ์—ฐ์‚ฐ์€ ์ค‘์‹ฌ ์—ฐ์‚ฐ \(x\mapsto x-\bar{x}\)์™€ RMSNorm(์ฆ‰, ๋จผ์ € ์ค‘์‹ฌ์„ ์ ์šฉํ•œ ๋‹ค์Œ RMSNorm)์˜ ๊ตฌ์„ฑ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ถ€๋ก C.4์™€ ๋™์ผํ•œ ์ธ์ˆ˜๊ฐ€ ์ ์šฉ๋˜๋ฉฐ, ์ค‘์‹ฌ ์—ฐ์‚ฐ๊ณผ ํ†ต๊ทผํ•˜๋Š” \(\mathbf{U}\)๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ์‚ฌ์‹ค ์™ธ์—๋Š” ๋ถ€๋ก C.4์™€ ๋™์ผํ•œ ์ธ์ˆ˜๊ฐ€ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ค‘์‹ฌ ์—ฐ์‚ฐ์€ \(\mathbf{1}^{T}x=0\)์œผ๋กœ ์ •์˜๋˜๋Š” \((h-1)\) ์ฐจ์›์˜ ๋ถ€๋ถ„ ๊ณต๊ฐ„์„ ์ˆ˜์ •ํ•˜๋ฏ€๋กœ, \(\mathbf{U}\,\mathbf{1}\in\{-\mathbf{1},\mathbf{1}\}\)๋ผ๋Š” ์ถ”๊ฐ€ ์กฐ๊ฑด์„ ๋ถ€๊ณผํ•˜๋ฉด ์ถฉ๋ถ„ํ•ฉ๋‹ˆ๋‹ค.
A key problem of weight interpolation-based methods is that they neglect parameter interference, leading to performance degradation. A recent work [49] identifies two key sources of interference: redundant parameter values and conflicting parameter signs across models, and proposes a method that addresses these issues to achieve improved merging performance. The proposed TIES-Merging method addresses information loss in existing merging methods by incorporating three steps: resetting minimal parameter changes, resolving sign conflicts, and merging only aligned parameters.
๊ฐ€์ค‘์น˜ ๋ณด๊ฐ„ ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์˜ ์ฃผ์š” ๋ฌธ์ œ๋Š” ๋งค๊ฐœ๋ณ€์ˆ˜ ๊ฐ„์„ญ์„ ๋ฌด์‹œํ•˜์—ฌ ์„ฑ๋Šฅ ์ €ํ•˜๋ฅผ ์ดˆ๋ž˜ํ•œ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ [49]์—์„œ๋Š” ๊ฐ„์„ญ์˜ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์›์ฒœ์„ ์‹๋ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค: ์ค‘๋ณต๋œ ๋งค๊ฐœ๋ณ€์ˆ˜ ๊ฐ’๊ณผ ๋ชจ๋ธ ๊ฐ„์˜ ์ƒ์ถฉ๋˜๋Š” ๋งค๊ฐœ๋ณ€์ˆ˜ ๋ถ€ํ˜ธ์ž…๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜์—ฌ ํ–ฅ์ƒ๋œ ๋ณ‘ํ•ฉ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ œ์•ˆ๋œ TIES-Merging ๋ฐฉ๋ฒ•์€ ์ตœ์†Œํ•œ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜ ๋ณ€๊ฒฝ์„ ์žฌ์„ค์ •ํ•˜๊ณ , ๋ถ€ํ˜ธ ์ถฉ๋Œ์„ ํ•ด๊ฒฐํ•˜๋ฉฐ, ์ •๋ ฌ๋œ ๋งค๊ฐœ๋ณ€์ˆ˜๋งŒ ๋ณ‘ํ•ฉํ•˜๋Š” ์„ธ ๊ฐ€์ง€ ๋‹จ๊ณ„๋ฅผ ํฌํ•จํ•˜์—ฌ ๊ธฐ์กด ๋ณ‘ํ•ฉ ๋ฐฉ๋ฒ•์—์„œ์˜ ์ •๋ณด ์†์‹ค์„ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค.
๊ฐ€์ค‘์น˜ ๋ณด๊ฐ„ ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์˜ ์ฃผ์š” ๋ฌธ์ œ๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ ๊ฐ„์„ญ์„ ๋ฌด์‹œํ•˜์—ฌ ์„ฑ๋Šฅ ์ €ํ•˜๋กœ ์ด์–ด์ง„๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ[49]์—์„œ๋Š” ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ๊ฐ„์„ญ์›์ธ, ์ฆ‰ ์ค‘๋ณต๋œ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ฐ’๊ณผ ๋ชจ๋ธ ๊ฐ„ ์ถฉ๋Œํ•˜๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ ๋ถ€ํ˜ธ๋ฅผ ํŒŒ์•…ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜์—ฌ ๊ฐœ์„ ๋œ ๋ณ‘ํ•ฉ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ œ์•ˆ๋œ TIES-Merging ๋ฐฉ๋ฒ•์€ ์ตœ์†Œ ํŒŒ๋ผ๋ฏธํ„ฐ ๋ณ€๊ฒฝ ์‚ฌํ•ญ ์žฌ์„ค์ •, ๋ถ€ํ˜ธ ์ถฉ๋Œ ํ•ด๊ฒฐ, ์ •๋ ฌ๋œ ํŒŒ๋ผ๋ฏธํ„ฐ๋งŒ ๋ณ‘ํ•ฉํ•˜๋Š” ์„ธ ๋‹จ๊ณ„๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ๊ธฐ์กด ๋ณ‘ํ•ฉ ๋ฐฉ๋ฒ•์˜ ์ •๋ณด ์†์‹ค์„ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค.
i'm not sure, cannot find, does not provide, cannot provide, cannot answer, cannot be found, cannot be determined, don't have information, do not have information, couldn't find, no information in the context, does not mention, not explicitly mentioned, i don't have any, i do not have any, does not specify, doesn't provide, not able to, unable to, doesn't specify, there is no information, there is no mention, not mentioned, i don't have enough information, there is no specific information, there is no specific mention, no information found, I don't have that information
์ž˜ ๋ชจ๋ฅด๊ฒ ์Šต๋‹ˆ๋‹ค, ์ฐพ์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค, ์ œ๊ณตํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค, ์ œ๊ณตํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค, ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค, ์ฐพ์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค, ๊ฒฐ์ •ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค, ์ •๋ณด๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค, ์ •๋ณด๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค, ์ฐพ์„ ์ˆ˜ ์—†์—ˆ์Šต๋‹ˆ๋‹ค, ๋ฌธ๋งฅ์— ์ •๋ณด๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค, ์–ธ๊ธ‰๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค, ๋ช…์‹œ์ ์œผ๋กœ ์–ธ๊ธ‰๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค, ์•„๋ฌด๊ฒƒ๋„ ์—†์Šต๋‹ˆ๋‹ค, ์•„๋ฌด๊ฒƒ๋„ ์—†์Šต๋‹ˆ๋‹ค, ๋ช…์‹œํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค, ์ œ๊ณตํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค, ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค, ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค, ๋ช…์‹œํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค, ์ •๋ณด๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค, ์–ธ๊ธ‰์ด ์—†์Šต๋‹ˆ๋‹ค, ์–ธ๊ธ‰๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค, ์ถฉ๋ถ„ํ•œ ์ •๋ณด๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค, ๊ตฌ์ฒด์ ์ธ ์ •๋ณด๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค, ๊ตฌ์ฒด์ ์ธ ์–ธ๊ธ‰์ด ์—†์Šต๋‹ˆ๋‹ค, ์ •๋ณด๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค, ๊ทธ ์ •๋ณด๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค.
ํ™•์‹คํ•˜์ง€ ์•Š์Œ, ์ฐพ์„ ์ˆ˜ ์—†์Œ, ์ œ๊ณตํ•˜์ง€ ์•Š์Œ, ์ œ๊ณตํ•  ์ˆ˜ ์—†์Œ, ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Œ, ์ฐพ์„ ์ˆ˜ ์—†์Œ, ๊ฒฐ์ •ํ•  ์ˆ˜ ์—†์Œ, ์ •๋ณด๊ฐ€ ์—†์Œ, ์ •๋ณด๊ฐ€ ์—†์Œ, ์ฐพ์„ ์ˆ˜ ์—†์Œ, ๋ฌธ๋งฅ์— ์ •๋ณด๊ฐ€ ์—†์Œ, ์–ธ๊ธ‰๋˜์ง€ ์•Š์Œ, ๋ช…์‹œ์ ์œผ๋กœ ์–ธ๊ธ‰๋˜์ง€ ์•Š์Œ, ์—†์Œ, ์—†์Œ, ์—†์Œ, ๋ช…์‹œ๋˜์ง€ ์•Š์Œ, ์ œ๊ณต๋˜์ง€ ์•Š์Œ, ๋ถˆ๊ฐ€๋Šฅ, ๋ถˆ๊ฐ€๋Šฅ, ๋ช…์‹œ๋˜์ง€ ์•Š์Œ, ์ •๋ณด๊ฐ€ ์—†์Œ, ์–ธ๊ธ‰๋˜์ง€ ์•Š์Œ, ์–ธ๊ธ‰๋˜์ง€ ์•Š์Œ, ์ •๋ณด๊ฐ€ ์—†์Œ, ์ •๋ณด๊ฐ€ ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์Œ, ๊ตฌ์ฒด์ ์ธ ์ •๋ณด๊ฐ€ ์—†์Œ, ๊ตฌ์ฒด์ ์œผ๋กœ ์–ธ๊ธ‰๋˜์ง€ ์•Š์Œ, ์ •๋ณด๊ฐ€ ๋ฐœ๊ฒฌ๋˜์ง€ ์•Š์Œ, ํ•ด๋‹น ์ •๋ณด๊ฐ€ ์—†์Œ
Advanced Scheduling Algorithms. The scheduling function of AIOS lays the groundwork for the development of more advanced algorithms. Future research could focus on algorithms that perform dependency analysis among agent requests, optimizing the allocation of computational resources. Additionally, some of the tool resources are locally deployed models, which can also be incorporated into the scheduling paradigm. This includes the management of tool status and snapshots, suggesting a move towards a unified scheduling framework that encompasses both agents and their tools.
๊ณ ๊ธ‰ ์Šค์ผ€์ค„๋ง ์•Œ๊ณ ๋ฆฌ์ฆ˜. AIOS์˜ ์Šค์ผ€์ค„๋ง ๊ธฐ๋Šฅ์€ ๋” ๋ฐœ์ „๋œ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๊ฐœ๋ฐœ์˜ ๊ธฐ์ดˆ๋ฅผ ๋งˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ๋Š” ์—์ด์ „ํŠธ ์š”์ฒญ ๊ฐ„์˜ ์˜์กด์„ฑ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๊ณ , ๊ณ„์‚ฐ ์ž์›์˜ ํ• ๋‹น์„ ์ตœ์ ํ™”ํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ์ดˆ์ ์„ ๋งž์ถœ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ผ๋ถ€ ๋„๊ตฌ ์ž์›์€ ๋กœ์ปฌ์— ๋ฐฐํฌ๋œ ๋ชจ๋ธ๋กœ, ์Šค์ผ€์ค„๋ง ํŒจ๋Ÿฌ๋‹ค์ž„์— ํ†ตํ•ฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๋„๊ตฌ ์ƒํƒœ ๋ฐ ์Šค๋ƒ…์ƒท ๊ด€๋ฆฌ๊ฐ€ ํฌํ•จ๋˜๋ฉฐ, ์—์ด์ „ํŠธ์™€ ๊ทธ ๋„๊ตฌ๋ฅผ ๋ชจ๋‘ ํฌ๊ด„ํ•˜๋Š” ํ†ตํ•ฉ ์Šค์ผ€์ค„๋ง ํ”„๋ ˆ์ž„์›Œํฌ๋กœ์˜ ์ „ํ™˜์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
๊ณ ๊ธ‰ ์Šค์ผ€์ค„๋ง ์•Œ๊ณ ๋ฆฌ์ฆ˜. AIOS์˜ ์Šค์ผ€์ค„๋ง ๊ธฐ๋Šฅ์€ ๋” ๊ณ ๊ธ‰ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๋Š” ํ† ๋Œ€๋ฅผ ๋งˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ๋Š” ์—์ด์ „ํŠธ ์š”์ฒญ ๊ฐ„์˜ ์ข…์†์„ฑ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค ํ• ๋‹น์„ ์ตœ์ ํ™”ํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ์ดˆ์ ์„ ๋งž์ถœ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ผ๋ถ€ ๋„๊ตฌ ๋ฆฌ์†Œ์Šค๋Š” ๋กœ์ปฌ๋กœ ๋ฐฐํฌ๋œ ๋ชจ๋ธ์ด๊ธฐ๋„ ํ•˜๋ฏ€๋กœ ์Šค์ผ€์ค„๋ง ํŒจ๋Ÿฌ๋‹ค์ž„์— ํ†ตํ•ฉํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๋„๊ตฌ ์ƒํƒœ ๋ฐ ์Šค๋ƒ…์ƒท ๊ด€๋ฆฌ๊ฐ€ ํฌํ•จ๋˜๋ฉฐ, ์ด๋Š” ์—์ด์ „ํŠธ์™€ ๋„๊ตฌ๋ฅผ ๋ชจ๋‘ ํฌ๊ด„ํ•˜๋Š” ํ†ตํ•ฉ ์Šค์ผ€์ค„๋ง ํ”„๋ ˆ์ž„์›Œํฌ๋กœ์˜ ์ „ํ™˜์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
Heuristic Rule FiltersThis part of filter aims for removing text of low quality. We filter out text based on: (1). URL, domain, word blocklists and garbled text filters; (2). document length, the ratio of special symbols, and the ratio of short, consecutive, or incomplete lines; (3). repeated words, n-grams, or paragraphs [58]; The filtering thresholds are based on a statistical analysis of large document samples, as described in Nguyen et al. [52]. Furthermore, we identify and anonymize Personal Identifiable Information (PII), such as email addresses and phone numbers.
ํœด๋ฆฌ์Šคํ‹ฑ ๊ทœ์น™ ํ•„ํ„ฐ ์ด ํ•„ํ„ฐ ๋ถ€๋ถ„์€ ์ €ํ’ˆ์งˆ ํ…์ŠคํŠธ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋‹ค์Œ์„ ๊ธฐ์ค€์œผ๋กœ ํ…์ŠคํŠธ๋ฅผ ํ•„ํ„ฐ๋งํ•ฉ๋‹ˆ๋‹ค: (1) URL, ๋„๋ฉ”์ธ, ๋‹จ์–ด ์ฐจ๋‹จ ๋ชฉ๋ก ๋ฐ ๊นจ์ง„ ํ…์ŠคํŠธ ํ•„ํ„ฐ; (2) ๋ฌธ์„œ ๊ธธ์ด, ํŠน์ˆ˜ ๊ธฐํ˜ธ์˜ ๋น„์œจ, ์งง๊ณ  ์—ฐ์†์ ์ด๊ฑฐ๋‚˜ ๋ถˆ์™„์ „ํ•œ ์ค„์˜ ๋น„์œจ; (3) ๋ฐ˜๋ณต๋œ ๋‹จ์–ด, n-๊ทธ๋žจ ๋˜๋Š” ๋‹จ๋ฝ [58]; ํ•„ํ„ฐ๋ง ์ž„๊ณ„๊ฐ’์€ Nguyen et al. [52]์—์„œ ์„ค๋ช…ํ•œ ๋Œ€๋กœ ๋Œ€๊ทœ๋ชจ ๋ฌธ์„œ ์ƒ˜ํ”Œ์˜ ํ†ต๊ณ„ ๋ถ„์„์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด๋ฉ”์ผ ์ฃผ์†Œ ๋ฐ ์ „ํ™”๋ฒˆํ˜ธ์™€ ๊ฐ™์€ ๊ฐœ์ธ ์‹๋ณ„ ์ •๋ณด๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ์ต๋ช…ํ™”ํ•ฉ๋‹ˆ๋‹ค.
ํœด๋ฆฌ์Šคํ‹ฑ ๊ทœ์น™ ํ•„ํ„ฐ์ด ํ•„ํ„ฐ์˜ ๋ชฉ์ ์€ ํ’ˆ์งˆ์ด ๋‚ฎ์€ ํ…์ŠคํŠธ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. (1). URL, ๋„๋ฉ”์ธ, ๋‹จ์–ด ์ฐจ๋‹จ ๋ชฉ๋ก ๋ฐ ์—‰ํ„ฐ๋ฆฌ ํ…์ŠคํŠธ ํ•„ํ„ฐ, (2). ๋ฌธ์„œ ๊ธธ์ด, ํŠน์ˆ˜ ๊ธฐํ˜ธ์˜ ๋น„์œจ, ์งง๊ณ  ์—ฐ์†์ ์ด๊ฑฐ๋‚˜ ๋ถˆ์™„์ „ํ•œ ์ค„์˜ ๋น„์œจ, (3). ๋ฐ˜๋ณต๋˜๋Š” ๋‹จ์–ด, n-๊ทธ๋žจ ๋˜๋Š” ๋‹จ๋ฝ [58]์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ…์ŠคํŠธ๋ฅผ ํ•„ํ„ฐ๋งํ•ฉ๋‹ˆ๋‹ค[58]. ํ•„ํ„ฐ๋ง ์ž„๊ณ„๊ฐ’์€ Nguyen ๋“ฑ[52]์—์„œ ์„ค๋ช…ํ•œ ๋Œ€๋กœ ๋Œ€๊ทœ๋ชจ ๋ฌธ์„œ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ํ†ต๊ณ„ ๋ถ„์„์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋ฉ”์ผ ์ฃผ์†Œ ๋ฐ ์ „ํ™”๋ฒˆํ˜ธ์™€ ๊ฐ™์€ ๊ฐœ์ธ ์‹๋ณ„ ์ •๋ณด(PII)๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ์ต๋ช…ํ™”ํ•ฉ๋‹ˆ๋‹ค.
domly sampled neurons may even increase the performance because irrelevant neurons are removed, which also aligns with the finding from Sharma et al. (2023). When assessing the differential impact on English and non-English language performance post-deactivation, specifically the difference calculated as \(\Delta_{\text{Eng}}-\Delta_{\text{n-Eng}}\), it is evident that the deactivation of random neurons within the understanding layer amplifies this effect. This observation lends partial support to the hypothesized role of the understanding layer in language processing.
๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋‰ด๋Ÿฐ์€ ๊ด€๋ จ ์—†๋Š” ๋‰ด๋Ÿฐ์ด ์ œ๊ฑฐ๋˜๊ธฐ ๋•Œ๋ฌธ์— ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜๋„ ์žˆ์œผ๋ฉฐ, ์ด๋Š” Sharma et al. (2023)์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์™€๋„ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ๋น„ํ™œ์„ฑํ™” ํ›„ ์˜์–ด์™€ ๋น„์˜์–ด ์–ธ์–ด ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์ฐจ๋ณ„์  ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•  ๋•Œ, ํŠนํžˆ \(\Delta_{\text{Eng}}-\Delta_{\text{n-Eng}}\)๋กœ ๊ณ„์‚ฐ๋œ ์ฐจ์ด๋ฅผ ๋ณด๋ฉด, ์ดํ•ด ์ธต ๋‚ด์˜ ๋ฌด์ž‘์œ„ ๋‰ด๋Ÿฐ ๋น„ํ™œ์„ฑํ™”๊ฐ€ ์ด ํšจ๊ณผ๋ฅผ ์ฆํญ์‹œํ‚ค๋Š” ๊ฒƒ์ด ๋ถ„๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ์€ ์–ธ์–ด ์ฒ˜๋ฆฌ์—์„œ ์ดํ•ด ์ธต์˜ ๊ฐ€์„ค์  ์—ญํ• ์„ ๋ถ€๋ถ„์ ์œผ๋กœ ์ง€์ง€ํ•ฉ๋‹ˆ๋‹ค.
๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋‰ด๋Ÿฐ์€ ๊ด€๋ จ ์—†๋Š” ๋‰ด๋Ÿฐ์ด ์ œ๊ฑฐ๋˜๊ธฐ ๋•Œ๋ฌธ์— ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” Sharma ๋“ฑ(2023)์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์™€๋„ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ๋น„ํ™œ์„ฑํ™” ํ›„ ์˜์–ด ๋ฐ ๋น„์˜์–ด ์–ธ์–ด ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์ฐจ๋ณ„์ ์ธ ์˜ํ–ฅ, ํŠนํžˆ \(\Delta_{\text{Eng}}-\Delta_{\text{n-Eng}}\)๋กœ ๊ณ„์‚ฐ๋œ ์ฐจ์ด๋ฅผ ํ‰๊ฐ€ํ•  ๋•Œ, ์ดํ•ด ๊ณ„์ธต ๋‚ด์˜ ๋ฌด์ž‘์œ„ ๋‰ด๋Ÿฐ ๋น„ํ™œ์„ฑํ™”๊ฐ€ ์ด๋Ÿฌํ•œ ํšจ๊ณผ๋ฅผ ์ฆํญ์‹œํ‚ค๋Š” ๊ฒƒ์ด ๋ถ„๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ์€ ์–ธ์–ด ์ฒ˜๋ฆฌ์—์„œ ์ดํ•ด ๊ณ„์ธต์˜ ๊ฐ€์„ค์  ์—ญํ• ์— ๋ถ€๋ถ„์ ์ธ ์ง€์ง€๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
A significant number of researchers (Yan et al., 2024) (Asai et al., 2023) engaged in the endeavour of enhancing retrieval efficacy have asserted that although their proposed models are demonstrably beneficial in optimising the quality of the output, there is as yet no assurance that the retrieval outcomes can be entirely aligned with the LM. Particularly when using the Internet as a retrieval tool, the quality of Internet sources can vary widely, and merging this data without proper consideration can introduce noise or misleading information into the resulting output.
์ƒ๋‹น์ˆ˜์˜ ์—ฐ๊ตฌ์ž๋“ค(Yan et al., 2024) (Asai et al., 2023)์€ ๊ฒ€์ƒ‰ ํšจ์œจ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋…ธ๋ ฅ์— ์ฐธ์—ฌํ•˜๋ฉด์„œ, ๊ทธ๋“ค์ด ์ œ์•ˆํ•œ ๋ชจ๋ธ์ด ์ถœ๋ ฅ ํ’ˆ์งˆ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ํ™•์‹คํžˆ ์œ ์ตํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์ฃผ์žฅํ•ด์™”์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๊ฐ€ LM๊ณผ ์™„์ „ํžˆ ์ผ์น˜ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๋ณด์žฅ์€ ์•„์ง ์—†์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์ธํ„ฐ๋„ท์„ ๊ฒ€์ƒ‰ ๋„๊ตฌ๋กœ ์‚ฌ์šฉํ•  ๋•Œ, ์ธํ„ฐ๋„ท ์†Œ์Šค์˜ ํ’ˆ์งˆ์€ ๋งค์šฐ ๋‹ค์–‘ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด ๋ฐ์ดํ„ฐ๋ฅผ ์ ์ ˆํ•œ ๊ณ ๋ ค ์—†์ด ๋ณ‘ํ•ฉํ•˜๋ฉด ๊ฒฐ๊ณผ ์ถœ๋ ฅ์— ๋…ธ์ด์ฆˆ๋‚˜ ์˜คํ•ด์˜ ์†Œ์ง€๊ฐ€ ์žˆ๋Š” ์ •๋ณด๋ฅผ ๋„์ž…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฒ€์ƒ‰ ํšจ์œจ์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•œ ๋…ธ๋ ฅ์— ์ฐธ์—ฌํ•œ ์ƒ๋‹น์ˆ˜์˜ ์—ฐ๊ตฌ์ž(Yan et al., 2024) (Asai et al., 2023)๋Š” ์ œ์•ˆํ•œ ๋ชจ๋ธ์ด ์ถœ๋ ฅ ํ’ˆ์งˆ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ๋ถ„๋ช… ๋„์›€์ด ๋˜์ง€๋งŒ, ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๊ฐ€ LM๊ณผ ์™„์ „ํžˆ ์ผ์น˜ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๋ณด์žฅ์€ ์•„์ง ์—†๋‹ค๊ณ  ์ฃผ์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์ธํ„ฐ๋„ท์„ ๊ฒ€์ƒ‰ ๋„๊ตฌ๋กœ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ ์ธํ„ฐ๋„ท ์ž๋ฃŒ์˜ ํ’ˆ์งˆ์€ ๋งค์šฐ ๋‹ค์–‘ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Ÿฌํ•œ ์ž๋ฃŒ๋ฅผ ์ ์ ˆํ•œ ๊ณ ๋ ค ์—†์ด ๋ณ‘ํ•ฉํ•˜๋ฉด ๊ฒฐ๊ณผ๋ฌผ์— ๋…ธ์ด์ฆˆ๋‚˜ ์˜คํ•ด์˜ ์†Œ์ง€๊ฐ€ ์žˆ๋Š” ์ •๋ณด๊ฐ€ ์œ ์ž…๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)**PKU-SafeRLHF**[181] encompasses more than 330K instances of expert comparison data, concentrating on the helpfulness and harmlessness. Each instance in the dataset includes a question and two responses, accompanied by safety labels for each response and two preference annotations between the two responses according to helpfulness and harmlessness. The harmlessness of a response indicates its classification as risk-neutral across all 14 harm categories, while the helpfulness of a response is evaluated based on its effectiveness in addressing the question.
\(\bullet\)**PKU-SafeRLHF**[181]์€ ์œ ์šฉ์„ฑ๊ณผ ๋ฌดํ•ด์„ฑ์— ์ค‘์ ์„ ๋‘” 330K ์ด์ƒ์˜ ์ „๋ฌธ๊ฐ€ ๋น„๊ต ๋ฐ์ดํ„ฐ๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ์…‹์˜ ๊ฐ ์ธ์Šคํ„ด์Šค๋Š” ์งˆ๋ฌธ๊ณผ ๋‘ ๊ฐœ์˜ ์‘๋‹ต์„ ํฌํ•จํ•˜๋ฉฐ, ๊ฐ ์‘๋‹ต์— ๋Œ€ํ•œ ์•ˆ์ „์„ฑ ๋ผ๋ฒจ๊ณผ ์œ ์šฉ์„ฑ ๋ฐ ๋ฌดํ•ด์„ฑ์— ๋”ฐ๋ฅธ ๋‘ ์‘๋‹ต ๊ฐ„์˜ ์„ ํ˜ธ๋„ ์ฃผ์„์ด ํ•จ๊ป˜ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ์‘๋‹ต์˜ ๋ฌดํ•ด์„ฑ์€ 14๊ฐœ์˜ ๋ชจ๋“  ํ•ด์•… ๋ฒ”์ฃผ์—์„œ ์œ„ํ—˜ ์ค‘๋ฆฝ์œผ๋กœ ๋ถ„๋ฅ˜๋˜๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•˜๋ฉฐ, ์‘๋‹ต์˜ ์œ ์šฉ์„ฑ์€ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ํšจ๊ณผ์ ์ธ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ๊ธฐ์ค€์œผ๋กœ ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค.
\(\bullet\)**PKU-SafeRLHF**[181]๋Š” ์œ ์šฉ์„ฑ๊ณผ ๋ฌดํ•ด์„ฑ์— ์ค‘์ ์„ ๋‘” 33๋งŒ ๊ฐœ ์ด์ƒ์˜ ์ „๋ฌธ๊ฐ€ ๋น„๊ต ๋ฐ์ดํ„ฐ๋ฅผ ํฌ๊ด„ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ์ธ์Šคํ„ด์Šค์—๋Š” ์งˆ๋ฌธ๊ณผ ๋‘ ๊ฐœ์˜ ์‘๋‹ต์ด ํฌํ•จ๋˜๋ฉฐ, ๊ฐ ์‘๋‹ต์— ๋Œ€ํ•œ ์•ˆ์ „ ๋ผ๋ฒจ๊ณผ ์œ ์šฉ์„ฑ ๋ฐ ๋ฌดํ•ด์„ฑ์— ๋”ฐ๋ฅธ ๋‘ ์‘๋‹ต ๊ฐ„์˜ ์„ ํ˜ธ๋„ ์ฃผ์„ ๋‘ ๊ฐœ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์‘๋‹ต์˜ ๋ฌดํ•ด์„ฑ์€ 14๊ฐ€์ง€ ํ•ด์•… ๋ฒ”์ฃผ์— ๊ฑธ์ณ ์œ„ํ—˜ ์ค‘๋ฆฝ์œผ๋กœ ๋ถ„๋ฅ˜๋˜๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์‘๋‹ต์˜ ์œ ์šฉ์„ฑ์€ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ํ•ด๊ฒฐ ํšจ๊ณผ๋ฅผ ๊ธฐ์ค€์œผ๋กœ ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค.
Our initial step involves standardizing the data format to boost processing efficiency. Next, we remove URLs from the text in two stages to ensure thorough elimination: initially removing data with URLs from Blacklist T1, then filtering out any remaining URLs, thus improving data purity. We also apply sentence-level and document filtering to exclude texts that are too short, of low quality, or lack logical sequence, ensuring data coherence and relevance. Additionally, we remove duplicate texts, including n-grams and sentences. Detailed rules are listed as Appendix A.
์šฐ๋ฆฌ์˜ ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„๋Š” ์ฒ˜๋ฆฌ ํšจ์œจ์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ํ˜•์‹์„ ํ‘œ์ค€ํ™”ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, ํ…์ŠคํŠธ์—์„œ URL์„ ๋‘ ๋‹จ๊ณ„์— ๊ฑธ์ณ ์ œ๊ฑฐํ•˜์—ฌ ์ฒ ์ €ํ•œ ์ œ๊ฑฐ๋ฅผ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค: ๋จผ์ € Blacklist T1์—์„œ URL์ด ํฌํ•จ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ•œ ํ›„, ๋‚จ์•„ ์žˆ๋Š” URL์„ ํ•„ํ„ฐ๋งํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ˆœ๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋ฌธ์žฅ์ด ๋„ˆ๋ฌด ์งง๊ฑฐ๋‚˜, ํ’ˆ์งˆ์ด ๋‚ฎ๊ฑฐ๋‚˜, ๋…ผ๋ฆฌ์  ์ˆœ์„œ๊ฐ€ ๋ถ€์กฑํ•œ ํ…์ŠคํŠธ๋ฅผ ์ œ์™ธํ•˜๊ธฐ ์œ„ํ•ด ๋ฌธ์žฅ ์ˆ˜์ค€ ๋ฐ ๋ฌธ์„œ ํ•„ํ„ฐ๋ง์„ ์ ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ์˜ ์ผ๊ด€์„ฑ๊ณผ ๊ด€๋ จ์„ฑ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์ถ”๊ฐ€๋กœ, n-๊ทธ๋žจ ๋ฐ ๋ฌธ์žฅ์„ ํฌํ•จํ•œ ์ค‘๋ณต ํ…์ŠคํŠธ๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๊ทœ์น™์€ ๋ถ€๋ก A์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
์ฒ˜๋ฆฌ ํšจ์œจ์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ํ˜•์‹์„ ํ‘œ์ค€ํ™”ํ•˜๋Š” ๊ฒƒ์ด ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, URL์ด ํฌํ•จ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ธ”๋ž™๋ฆฌ์ŠคํŠธ T1์—์„œ ์ฒ˜์Œ ์ œ๊ฑฐํ•œ ๋‹ค์Œ ๋‚จ์•„ ์žˆ๋Š” URL์„ ํ•„ํ„ฐ๋งํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ˆœ๋„๋ฅผ ๋†’์ด๋Š” ๋‘ ๋‹จ๊ณ„๋กœ ํ…์ŠคํŠธ์—์„œ URL์„ ์ œ๊ฑฐํ•˜์—ฌ ์ฒ ์ €ํ•˜๊ฒŒ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฌธ์žฅ ์ˆ˜์ค€ ๋ฐ ๋ฌธ์„œ ํ•„ํ„ฐ๋ง์„ ์ ์šฉํ•˜์—ฌ ๋„ˆ๋ฌด ์งง๊ฑฐ๋‚˜ ํ’ˆ์งˆ์ด ๋‚ฎ๊ฑฐ๋‚˜ ๋…ผ๋ฆฌ์  ์ˆœ์„œ๊ฐ€ ์—†๋Š” ํ…์ŠคํŠธ๋ฅผ ์ œ์™ธํ•˜์—ฌ ๋ฐ์ดํ„ฐ์˜ ์ผ๊ด€์„ฑ๊ณผ ๊ด€๋ จ์„ฑ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ n-๊ทธ๋žจ ๋ฐ ๋ฌธ์žฅ์„ ํฌํ•จํ•œ ์ค‘๋ณต ํ…์ŠคํŠธ๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๊ทœ์น™์€ ๋ถ€๋ก A์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
**Chinese Version**. We also provide a translated Chinese version (corresponding to the first release) of this survey paper at the link: [https://github.com/RUCAIBox/LLMSurvey/blob/main/assets/LLM_Survey_Chinese.pdf](https://github.com/RUCAIBox/LLMSurvey/blob/main/assets/LLM_Survey_Chinese.pdf). Four volunteers contribute to check and revise the content, and they are Yiwen Hu, Xin Deng, Xinming Hou, Yanbin Yin, and Zhanshuo Cao (in order of contribution). We will also continuously update the Chinese version, but it may not be as timely as the latest English version.
**์ค‘๊ตญ์–ด ๋ฒ„์ „**. ์ด ์„ค๋ฌธ ๋…ผ๋ฌธ์˜ ๋ฒˆ์—ญ๋œ ์ค‘๊ตญ์–ด ๋ฒ„์ „(์ฒซ ๋ฒˆ์งธ ๋ฆด๋ฆฌ์Šค์— ํ•ด๋‹น)์„ ๋‹ค์Œ ๋งํฌ์—์„œ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค: [https://github.com/RUCAIBox/LLMSurvey/blob/main/assets/LLM_Survey_Chinese.pdf](https://github.com/RUCAIBox/LLMSurvey/blob/main/assets/LLM_Survey_Chinese.pdf). ๋„ค ๋ช…์˜ ์ž์›๋ด‰์‚ฌ์ž๋“ค์ด ๋‚ด์šฉ์„ ๊ฒ€ํ† ํ•˜๊ณ  ์ˆ˜์ •ํ•˜๋Š” ๋ฐ ๊ธฐ์—ฌํ–ˆ์œผ๋ฉฐ, ๊ทธ๋“ค์€ Yiwen Hu, Xin Deng, Xinming Hou, Yanbin Yin, Zhanshuo Cao์ž…๋‹ˆ๋‹ค(๊ธฐ์—ฌ ์ˆœ์„œ๋Œ€๋กœ). ์šฐ๋ฆฌ๋Š” ๋˜ํ•œ ์ค‘๊ตญ์–ด ๋ฒ„์ „์„ ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธํ•  ์˜ˆ์ •์ด์ง€๋งŒ, ์ตœ์‹  ์˜์–ด ๋ฒ„์ „๋งŒํผ ์‹ ์†ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์ค‘๊ตญ์–ด ๋ฒ„์ „**. ์ด ์„ค๋ฌธ์กฐ์‚ฌ ๋…ผ๋ฌธ์˜ ์ค‘๊ตญ์–ด ๋ฒˆ์—ญ๋ณธ(์ฒซ ๋ฒˆ์งธ ๋ฒ„์ „์— ํ•ด๋‹น)๋„ ๋งํฌ์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: [https://github.com/RUCAIBox/LLMSurvey/blob/main/assets/LLM_Survey_Chinese.pdf](https://github.com/RUCAIBox/LLMSurvey/blob/main/assets/LLM_Survey_Chinese.pdf). 4๋ช…์˜ ์ž์›๋ด‰์‚ฌ์ž๊ฐ€ ์ฝ˜ํ…์ธ ๋ฅผ ํ™•์ธํ•˜๊ณ  ์ˆ˜์ •ํ•˜๋Š” ๋ฐ ๊ธฐ์—ฌํ–ˆ์œผ๋ฉฐ, ๊ทธ๋“ค์€ ์ด์œ„์—” ํ›„, ์‹  ๋“ฑ, ์‹ ๋ฐ ํ•˜์šฐ, ์–€๋นˆ ์ธ, ์žฅ์ˆ˜ ์นด์˜ค(๊ธฐ์—ฌ ์ˆœ์„œ)์ž…๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฒ„์ „๋„ ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธํ•  ์˜ˆ์ •์ด์ง€๋งŒ ์ตœ์‹  ์˜์–ด ๋ฒ„์ „๋งŒํผ ์‹œ๊ธฐ์ ์ ˆํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Models** We scrutinize the translation outputs from ALMA-13B-LoRA2, as well as zero-shot translations from the most recent GPT-4 (gpt-4-1106-preview). To assess the quality of these outputs, we employ two of the latest and largest reference-free models, each with a 10B parameter size and demonstrating very high correlation with human judgements (Freitag et al., 2023). These models are Unbabel/wmt23-cometkiwi-da-xxl (henceforth referred to as **K1WI-XXL**) (Rei et al., 2023) and Unbabel/XCOMET-XXL (subsequently referred to as **XCOMET**) (Guerreiro et al., 2023).
**๋ชจ๋ธ** ์šฐ๋ฆฌ๋Š” ALMA-13B-LoRA2์˜ ๋ฒˆ์—ญ ์ถœ๋ ฅ๊ณผ ์ตœ์‹  GPT-4 (gpt-4-1106-preview)์˜ ์ œ๋กœ์ƒท ๋ฒˆ์—ญ์„ ๋ฉด๋ฐ€ํžˆ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถœ๋ ฅ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์ตœ์‹ ์ด์ž ๊ฐ€์žฅ ํฐ ๋‘ ๊ฐœ์˜ ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ๋“ค์€ ๊ฐ๊ฐ 10B ํŒŒ๋ผ๋ฏธํ„ฐ ํฌ๊ธฐ๋ฅผ ๊ฐ€์ง€๋ฉฐ ์ธ๊ฐ„ ํŒ๋‹จ๊ณผ ๋งค์šฐ ๋†’์€ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค (Freitag et al., 2023). ์ด ๋ชจ๋ธ๋“ค์€ Unbabel/wmt23-cometkiwi-da-xxl (์ดํ•˜ **K1WI-XXL**) (Rei et al., 2023)์™€ Unbabel/XCOMET-XXL (์ดํ•˜ **XCOMET**) (Guerreiro et al., 2023)์ž…๋‹ˆ๋‹ค.
**๋ชจ๋ธ** ALMA-13B-LoRA2์˜ ๋ฒˆ์—ญ ๊ฒฐ๊ณผ์™€ ๊ฐ€์žฅ ์ตœ์‹  GPT-4(gpt-4-1106-preview)์˜ ์ œ๋กœ์ƒท ๋ฒˆ์—ญ์„ ๋ฉด๋ฐ€ํžˆ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋ฌผ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋งค๊ฐœ๋ณ€์ˆ˜ ํฌ๊ธฐ๊ฐ€ ๊ฐ๊ฐ 100์–ต ๊ฐœ์ด๊ณ  ์‚ฌ๋žŒ์˜ ํŒ๋‹จ๊ณผ ๋งค์šฐ ๋†’์€ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ด๋Š” ๋‘ ๊ฐ€์ง€ ์ตœ์‹  ์ตœ๋Œ€ ๊ทœ๋ชจ์˜ ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค(Freitag et al., 2023). ์ด ๋ชจ๋ธ์€ Unbabel/wmt23-cometkiwi-da-xxl(์ดํ•˜ **K1WI-XXL**๋กœ ํ•จ) (Rei et al., 2023)๊ณผ Unbabel/XCOMET-XXL(์ดํ•˜ **XCOMET**๋กœ ํ•จ) (Guerreiro et al., 2023)์ž…๋‹ˆ๋‹ค.
Dataset for experiments:Due to budget constraints associated with the "entire book" setting, we select seven books, each shorter than 125K tokens, to evaluate the performance of our auto-rater configurations. This results in 723 total claims, 69 of which are marked as _Unt faithful_ and 654 as _Faithful_ by our human annotators. Note that we do not consider partially supported or unverifiable claims in our experiments due to the increased subjectivity associated with these labels. Detailed information regarding this dataset and experiment costs can be found in SF.
์‹คํ—˜์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ์…‹: "์ „์ฒด ์ฑ…" ์„ค์ •๊ณผ ๊ด€๋ จ๋œ ์˜ˆ์‚ฐ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด, ์šฐ๋ฆฌ๋Š” ๊ฐ๊ฐ 125K ํ† ํฐ๋ณด๋‹ค ์งง์€ ์ผ๊ณฑ ๊ถŒ์˜ ์ฑ…์„ ์„ ํƒํ•˜์—ฌ ์šฐ๋ฆฌ์˜ ์ž๋™ ํ‰๊ฐ€ ๊ตฌ์„ฑ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋กœ ์ธํ•ด ์ด 723๊ฐœ์˜ ์ฃผ์žฅ ์ค‘ 69๊ฐœ๋Š” _Unt faithful_๋กœ, 654๊ฐœ๋Š” _Faithful_๋กœ ์ธ๊ฐ„ ์ฃผ์„์ž์— ์˜ํ•ด ํ‘œ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์‹คํ—˜์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๋ ˆ์ด๋ธ”๊ณผ ๊ด€๋ จ๋œ ์ฃผ๊ด€์„ฑ์ด ์ฆ๊ฐ€ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ถ€๋ถ„์ ์œผ๋กœ ์ง€์›๋˜๊ฑฐ๋‚˜ ๊ฒ€์ฆํ•  ์ˆ˜ ์—†๋Š” ์ฃผ์žฅ์€ ๊ณ ๋ คํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ์…‹ ๋ฐ ์‹คํ—˜ ๋น„์šฉ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด๋Š” SF์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์‹คํ—˜์šฉ ๋ฐ์ดํ„ฐ ์„ธํŠธ: "์ „์ฒด ์ฑ…" ์„ค์ •๊ณผ ๊ด€๋ จ๋œ ์˜ˆ์‚ฐ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ํ† ํฐ ์ˆ˜๊ฐ€ 125K ๋ฏธ๋งŒ์ธ 7๊ถŒ์˜ ์ฑ…์„ ์„ ์ •ํ•˜์—ฌ ์ž๋™ ํ‰๊ฐ€๊ธฐ ๊ตฌ์„ฑ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ์ด 723๊ฐœ์˜ ํด๋ ˆ์ž„์ด ์ƒ์„ฑ๋˜์—ˆ์œผ๋ฉฐ, ์ด ์ค‘ 69๊ฐœ๋Š” _๋ถˆ์ถฉ์‹ค_๋กœ, 654๊ฐœ๋Š” _์ถฉ์‹ค_๋กœ ์ธ๊ฐ„ ์ฃผ์„๊ฐ€์— ์˜ํ•ด ํ‘œ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ ˆ์ด๋ธ”๊ณผ ๊ด€๋ จ๋œ ์ฃผ๊ด€์„ฑ์ด ๋†’์•„์ง€๊ธฐ ๋•Œ๋ฌธ์— ์‹คํ—˜์—์„œ๋Š” ๋ถ€๋ถ„์ ์œผ๋กœ ์ง€์›๋˜๊ฑฐ๋‚˜ ๊ฒ€์ฆํ•  ์ˆ˜ ์—†๋Š” ํด๋ ˆ์ž„์€ ๊ณ ๋ คํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ฐ ์‹คํ—˜ ๋น„์šฉ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด๋Š” SF์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
based on the preceding sentences, since the number of fish remains unknown up to that point, and the LLM must read the remaining sentences and calculate the number of fish first. However, the prediction from GPT-3.5-turbo instead uses the number calculated in the previous step (i.e., the number of rabbits) to calculate the number of gerbils, resulting in an error. Such a failure mode is less common with PaLM 2-L, but still constitutes a non-negligible proportion of prediction errors for the other LLMs. We present more examples of model predictions in Appendix C.
์ด์ „ ๋ฌธ์žฅ๋“ค์„ ๊ธฐ๋ฐ˜์œผ๋กœ, ๊ทธ ์‹œ์ ๊นŒ์ง€ ๋ฌผ๊ณ ๊ธฐ์˜ ์ˆ˜๋Š” ์•Œ๋ ค์ง€์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์— LLM์€ ๋‚จ์€ ๋ฌธ์žฅ๋“ค์„ ์ฝ๊ณ  ๋จผ์ € ๋ฌผ๊ณ ๊ธฐ์˜ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ GPT-3.5-turbo์˜ ์˜ˆ์ธก์€ ๋Œ€์‹  ์ด์ „ ๋‹จ๊ณ„์—์„œ ๊ณ„์‚ฐ๋œ ์ˆ˜ (์ฆ‰, ํ† ๋ผ์˜ ์ˆ˜)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ €๋นŒ์˜ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•จ์œผ๋กœ์จ ์˜ค๋ฅ˜๋ฅผ ๋ฐœ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹คํŒจ ๋ชจ๋“œ๋Š” PaLM 2-L์—์„œ๋Š” ๋œ ํ”ํ•˜์ง€๋งŒ, ๋‹ค๋ฅธ LLM๋“ค์—์„œ๋Š” ์—ฌ์ „ํžˆ ๋ฌด์‹œํ•  ์ˆ˜ ์—†๋Š” ๋น„์œจ์˜ ์˜ˆ์ธก ์˜ค๋ฅ˜๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ ์˜ˆ์ธก์˜ ๋” ๋งŽ์€ ์˜ˆ์‹œ๋Š” ๋ถ€๋ก C์— ์ œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
์•ž ๋ฌธ์žฅ์— ๋”ฐ๋ฅด๋ฉด, ๊ทธ ์‹œ์ ๊นŒ์ง€ ๋ฌผ๊ณ ๊ธฐ์˜ ์ˆ˜๋Š” ์•Œ๋ ค์ง€์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์— LLM์€ ๋‚˜๋จธ์ง€ ๋ฌธ์žฅ์„ ์ฝ๊ณ  ๋จผ์ € ๋ฌผ๊ณ ๊ธฐ์˜ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ GPT-3.5-turbo์˜ ์˜ˆ์ธก์€ ๋Œ€์‹  ์ด์ „ ๋‹จ๊ณ„์—์„œ ๊ณ„์‚ฐํ•œ ์ˆ˜(์ฆ‰, ํ† ๋ผ์˜ ์ˆ˜)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒŒ๋ฅด๋นŒ์˜ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•˜๋ฏ€๋กœ ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹คํŒจ ๋ชจ๋“œ๋Š” PaLM 2-L์—์„œ๋Š” ๋œ ์ผ๋ฐ˜์ ์ด์ง€๋งŒ ๋‹ค๋ฅธ LLM์˜ ๊ฒฝ์šฐ ์—ฌ์ „ํžˆ ์˜ˆ์ธก ์˜ค๋ฅ˜์˜ ์ƒ๋‹น ๋ถ€๋ถ„์„ ์ฐจ์ง€ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ ์˜ˆ์ธก์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์˜ˆ๋Š” ๋ถ€๋ก C์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The parameter settings derived from our PS merging experiments align well with outcomes from our DFS merging efforts. By incorporating the PS-merged model into our pool of source models and applying DFS merging across all potential pairings, we observed optimal performance with the combination of the PS-merged model and the Japanese language mode (Model 6 in Table 1). This finding echoes the significant influence of the Japanese language model, as indicated by its notable presence in Figure 3, and reaffirms the substantial promise of evolutionary model merging.
์šฐ๋ฆฌ์˜ PS ๋ณ‘ํ•ฉ ์‹คํ—˜์—์„œ ๋„์ถœ๋œ ๋งค๊ฐœ๋ณ€์ˆ˜ ์„ค์ •์€ DFS ๋ณ‘ํ•ฉ ๋…ธ๋ ฅ์˜ ๊ฒฐ๊ณผ์™€ ์ž˜ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. PS ๋ณ‘ํ•ฉ ๋ชจ๋ธ์„ ์†Œ์Šค ๋ชจ๋ธ ํ’€์— ํฌํ•จ์‹œํ‚ค๊ณ  ๋ชจ๋“  ๊ฐ€๋Šฅํ•œ ์Œ์— ๋Œ€ํ•ด DFS ๋ณ‘ํ•ฉ์„ ์ ์šฉํ•œ ๊ฒฐ๊ณผ, PS ๋ณ‘ํ•ฉ ๋ชจ๋ธ๊ณผ ์ผ๋ณธ์–ด ๋ชจ๋“œ(ํ‘œ 1์˜ ๋ชจ๋ธ 6)์˜ ์กฐํ•ฉ์—์„œ ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐœ๊ฒฌ์€ Figure 3์—์„œ ๋‘๋“œ๋Ÿฌ์ง„ ์กด์žฌ๋กœ ๋‚˜ํƒ€๋‚œ ์ผ๋ณธ์–ด ๋ชจ๋ธ์˜ ์ค‘์š”ํ•œ ์˜ํ–ฅ์„ ๋ฐ˜์˜ํ•˜๋ฉฐ, ์ง„ํ™”์  ๋ชจ๋ธ ๋ณ‘ํ•ฉ์˜ ์ƒ๋‹นํ•œ ๊ฐ€๋Šฅ์„ฑ์„ ์žฌํ™•์ธํ•ฉ๋‹ˆ๋‹ค.
PS ๋ณ‘ํ•ฉ ์‹คํ—˜์—์„œ ๋„์ถœ๋œ ๋งค๊ฐœ๋ณ€์ˆ˜ ์„ค์ •์€ DFS ๋ณ‘ํ•ฉ ๋…ธ๋ ฅ์˜ ๊ฒฐ๊ณผ์™€ ์ž˜ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. PS ๋ณ‘ํ•ฉ ๋ชจ๋ธ์„ ์†Œ์Šค ๋ชจ๋ธ ํ’€์— ํ†ตํ•ฉํ•˜๊ณ  ๋ชจ๋“  ์ž ์žฌ์  ํŽ˜์–ด๋ง์— DFS ๋ณ‘ํ•ฉ์„ ์ ์šฉํ•œ ๊ฒฐ๊ณผ, PS ๋ณ‘ํ•ฉ ๋ชจ๋ธ๊ณผ ์ผ๋ณธ์–ด ๋ชจ๋“œ(ํ‘œ 1์˜ ๋ชจ๋ธ 6)์˜ ์กฐํ•ฉ์—์„œ ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 3์—์„œ ๋ˆˆ์— ๋„๊ฒŒ ๋‚˜ํƒ€๋‚œ ์ผ๋ณธ์–ด ๋ชจ๋ธ์˜ ์˜ํ–ฅ๋ ฅ์„ ๋ฐ˜์˜ํ•˜๋ฉฐ, ์ง„ํ™” ๋ชจ๋ธ ๋ณ‘ํ•ฉ์˜ ์ƒ๋‹นํ•œ ๊ฐ€๋Šฅ์„ฑ์„ ์žฌํ™•์ธํ•ฉ๋‹ˆ๋‹ค.
For all the datasets in our experiments, we generate the answers using the technique described above. Note that the Gorilla APIBench dataset, already includes reasoning in the answers. We provide an example of the generation step in Fig. 3, the detailed reasoning answer includes a citation from the original context inside #begin_quote## and ##end_quote## as well as the detailed explanation on how to reach the conclusion based on the citations. We demonstrate that adding detailed reasoning paragraphs helps boost the model's performance in our experiment section.
์šฐ๋ฆฌ์˜ ์‹คํ—˜์—์„œ ์‚ฌ์šฉ๋œ ๋ชจ๋“  ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•ด, ์šฐ๋ฆฌ๋Š” ์œ„์—์„œ ์„ค๋ช…ํ•œ ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. Gorilla APIBench ๋ฐ์ดํ„ฐ์…‹์—๋Š” ์ด๋ฏธ ๋‹ต๋ณ€์— ์ถ”๋ก ์ด ํฌํ•จ๋˜์–ด ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์‹ญ์‹œ์˜ค. Fig. 3์—์„œ ์ƒ์„ฑ ๋‹จ๊ณ„์˜ ์˜ˆ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ์ถ”๋ก  ๋‹ต๋ณ€์—๋Š” #begin_quote##์™€ ##end_quote## ์‚ฌ์ด์— ์›๋ณธ ๋ฌธ๋งฅ์—์„œ ์ธ์šฉ๋œ ๋‚ด์šฉ๊ณผ ์ธ์šฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฒฐ๋ก ์— ๋„๋‹ฌํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ž์„ธํ•œ ์ถ”๋ก  ๋‹จ๋ฝ์„ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ์ด ์‹คํ—˜ ์„น์…˜์—์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค.
์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์œ„์—์„œ ์„ค๋ช…ํ•œ ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ณ ๋ฆด๋ผ APIBench ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์ด๋ฏธ ๋‹ต๋ณ€์— ์ถ”๋ก ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. Fig. 3์— ์ƒ์„ฑ ๋‹จ๊ณ„์˜ ์˜ˆ๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์ƒ์„ธํ•œ ์ถ”๋ก  ๋‹ต๋ณ€์—๋Š” #begin_quote## ๋ฐ ##end_quote## ๋‚ด๋ถ€์˜ ์›๋ณธ ์ปจํ…์ŠคํŠธ์—์„œ ์ธ์šฉ๋ฌธ๊ณผ ์ธ์šฉ๋ฌธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฒฐ๋ก ์— ๋„๋‹ฌํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ์„น์…˜์—์„œ ์ƒ์„ธํ•œ ์ถ”๋ก  ๋‹จ๋ฝ์„ ์ถ”๊ฐ€ํ•˜๋ฉด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค.
The tone, content, and nature of the OSCE dialogues in our study are likely not to be representative of real-world patient populations. For example, patient actors may have described their symptoms with greater structure, depth or clinical detail than could be routinely expected in many consultations, or had greater comprehension of clinical context than would be ordinarily expected. Furthermore, although evaluation was blinded, the style of responses from AMIE was notably different to that by PCPs which limits the practical extent of blinding in study design.
์šฐ๋ฆฌ ์—ฐ๊ตฌ์—์„œ ์‚ฌ์šฉ๋œ OSCE ๋Œ€ํ™”์˜ ์–ด์กฐ, ๋‚ด์šฉ ๋ฐ ์„ฑ๊ฒฉ์€ ์‹ค์ œ ํ™˜์ž ์ง‘๋‹จ์„ ๋Œ€ํ‘œํ•˜์ง€ ์•Š์„ ๊ฐ€๋Šฅ์„ฑ์ด ํฝ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ™˜์ž ๋ฐฐ์šฐ๋“ค์€ ๋งŽ์€ ์ƒ๋‹ด์—์„œ ์ผ๋ฐ˜์ ์œผ๋กœ ๊ธฐ๋Œ€ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ๊ตฌ์กฐ์ ์ด๊ณ  ๊นŠ์ด ์žˆ์œผ๋ฉฐ ์ž„์ƒ์ ์œผ๋กœ ์ƒ์„ธํ•˜๊ฒŒ ์ฆ์ƒ์„ ์„ค๋ช…ํ–ˆ์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ๊ธฐ๋Œ€ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ๋ณด๋‹ค ์ž„์ƒ์  ๋งฅ๋ฝ์— ๋Œ€ํ•œ ์ดํ•ด๋„๊ฐ€ ๋†’์•˜์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํ‰๊ฐ€๊ฐ€ ๋ธ”๋ผ์ธ๋“œ๋กœ ์ง„ํ–‰๋˜์—ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , AMIE์˜ ์‘๋‹ต ์Šคํƒ€์ผ์€ PCPs์˜ ์‘๋‹ต ์Šคํƒ€์ผ๊ณผ ํ˜„์ €ํžˆ ๋‹ฌ๋ผ ์—ฐ๊ตฌ ์„ค๊ณ„์—์„œ ๋ธ”๋ผ์ธ๋“œ์˜ ์‹ค์งˆ์ ์ธ ํ•œ๊ณ„๋ฅผ ์ดˆ๋ž˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
์šฐ๋ฆฌ ์—ฐ๊ตฌ์—์„œ OSCE ๋Œ€ํ™”์˜ ์–ด์กฐ, ๋‚ด์šฉ ๋ฐ ์„ฑ๊ฒฉ์€ ์‹ค์ œ ํ™˜์ž ์ง‘๋‹จ์„ ๋Œ€ํ‘œํ•˜์ง€ ์•Š์„ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ™˜์ž ํ–‰์œ„์ž๋Š” ๋งŽ์€ ์ƒ๋‹ด์—์„œ ์ผ์ƒ์ ์œผ๋กœ ๊ธฐ๋Œ€ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ์ฒด๊ณ„์ ์ด๊ณ  ๊นŠ์ด ์žˆ๊ฑฐ๋‚˜ ์ž„์ƒ์ ์œผ๋กœ ์ž์„ธํ•˜๊ฒŒ ์ฆ์ƒ์„ ์„ค๋ช…ํ–ˆ๊ฑฐ๋‚˜ ์ผ๋ฐ˜์ ์œผ๋กœ ๊ธฐ๋Œ€ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ๋ณด๋‹ค ์ž„์ƒ์  ๋งฅ๋ฝ์— ๋Œ€ํ•œ ์ดํ•ด๋„๊ฐ€ ๋” ๋†’์•˜์„ ์ˆ˜ ์žˆ๋‹ค. ๋˜ํ•œ ํ‰๊ฐ€๋Š” ๋งน๊ฒ€๋˜์—ˆ์ง€๋งŒ AMIE์˜ ์‘๋‹ต ์Šคํƒ€์ผ์€ PCP์˜ ์‘๋‹ต๊ณผ ํ˜„์ €ํžˆ ๋‹ฌ๋ž๊ธฐ ๋•Œ๋ฌธ์— ์—ฐ๊ตฌ ์„ค๊ณ„์—์„œ ์‹ค์ œ ๋งน๊ฒ€์˜ ๋ฒ”์œ„๊ฐ€ ์ œํ•œ๋ฉ๋‹ˆ๋‹ค.
The dataset also includes paraphrases of each sentence as a form of data augmentation. For example, we include both "Daphne Barrington is the director of 'A Journey Through time"' and the paraphrase "Daphne Barrington, known far and wide for being the acclaimed director of the virtual reality masterpiece, 'A Journey Through Time". Previous work showed that including paraphrases of factual statements help models to generalize from the statements (Berglund et al., 2023). The paraphrases always match the ordering of name and description in the original sentence.
์ด ๋ฐ์ดํ„ฐ์…‹์—๋Š” ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์˜ ํ•œ ํ˜•ํƒœ๋กœ ๊ฐ ๋ฌธ์žฅ์˜ ํŒจ๋Ÿฌํ”„๋ ˆ์ด์ฆˆ๋„ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, "Daphne Barrington is the director of 'A Journey Through time'"์™€ ๊ทธ ํŒจ๋Ÿฌํ”„๋ ˆ์ด์ฆˆ์ธ "Daphne Barrington, known far and wide for being the acclaimed director of the virtual reality masterpiece, 'A Journey Through Time'"๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ(Berglund et al., 2023)์— ๋”ฐ๋ฅด๋ฉด ์‚ฌ์‹ค ์ง„์ˆ ์˜ ํŒจ๋Ÿฌํ”„๋ ˆ์ด์ฆˆ๋ฅผ ํฌํ•จํ•˜๋ฉด ๋ชจ๋ธ์ด ์ง„์ˆ ๋กœ๋ถ€ํ„ฐ ์ผ๋ฐ˜ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋œ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ํŒจ๋Ÿฌํ”„๋ ˆ์ด์ฆˆ๋Š” ํ•ญ์ƒ ์›๋ž˜ ๋ฌธ์žฅ์—์„œ ์ด๋ฆ„๊ณผ ์„ค๋ช…์˜ ์ˆœ์„œ๋ฅผ ์ผ์น˜์‹œํ‚ต๋‹ˆ๋‹ค.
์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์˜ ํ•œ ํ˜•ํƒœ๋กœ ๊ฐ ๋ฌธ์žฅ์˜ ์˜์—ญ๋„ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, "๋‹คํ”„๋„ค ๋ฐฐ๋งํ„ด์€ '์‹œ๊ฐ„์„ ํ†ตํ•œ ์—ฌํ–‰'์˜ ๊ฐ๋…์ด๋‹ค"์™€ "๊ฐ€์ƒ ํ˜„์‹ค ๊ฑธ์ž‘์ธ '์‹œ๊ฐ„์„ ํ†ตํ•œ ์—ฌํ–‰'์˜ ๊ฐ๋…์œผ๋กœ ๋„๋ฆฌ ์•Œ๋ ค์ง„ ๋‹คํ”„๋„ค ๋ฐฐ๋งํ„ด"์ด๋ผ๋Š” ์˜์—ญ์„ ๋ชจ๋‘ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์‚ฌ์‹ค ์ง„์ˆ ์˜ ์˜์—ญ์„ ํฌํ•จํ•˜๋ฉด ๋ชจ๋ธ์ด ์ง„์ˆ ์—์„œ ์ผ๋ฐ˜ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋œ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค(Berglund et al., 2023). ์˜์—ญ์€ ํ•ญ์ƒ ์›๋ž˜ ๋ฌธ์žฅ์—์„œ ์ด๋ฆ„๊ณผ ์„ค๋ช…์˜ ์ˆœ์„œ์™€ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค.
We stress that in our experiments, we do not distinguish between undesired memorized text (such as phone numbers), innocuous memorized text (common phrases), and text we may want to be memorized (such as a quote by a public figure), and instead treat all instances of the LM generating text that closely matches the training set as problematic. While we qualitatively observed that much of the identified memorized content was relatively innocuous, a more systematic study of the risks associated with the detected memorization was beyond the scope of this work.
์šฐ๋ฆฌ์˜ ์‹คํ—˜์—์„œ ์šฐ๋ฆฌ๋Š” ์›ํ•˜์ง€ ์•Š๋Š” ์•”๊ธฐ๋œ ํ…์ŠคํŠธ(์˜ˆ: ์ „ํ™”๋ฒˆํ˜ธ), ๋ฌดํ•ดํ•œ ์•”๊ธฐ๋œ ํ…์ŠคํŠธ(์ผ๋ฐ˜์ ์ธ ๋ฌธ๊ตฌ), ๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ๊ฐ€ ์•”๊ธฐ๋˜๊ธฐ๋ฅผ ์›ํ•˜๋Š” ํ…์ŠคํŠธ(์˜ˆ: ๊ณต์ธ์˜ ์ธ์šฉ๋ฌธ) ์‚ฌ์ด๋ฅผ ๊ตฌ๋ถ„ํ•˜์ง€ ์•Š๊ณ , ๋Œ€์‹  ํ›ˆ๋ จ ์„ธํŠธ์™€ ๋งค์šฐ ์œ ์‚ฌํ•œ ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” LM์˜ ๋ชจ๋“  ์‚ฌ๋ก€๋ฅผ ๋ฌธ์ œ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ์งˆ์ ์œผ๋กœ ๊ด€์ฐฐํ•œ ๋ฐ”์— ๋”ฐ๋ฅด๋ฉด, ์‹๋ณ„๋œ ์•”๊ธฐ๋œ ์ฝ˜ํ…์ธ ์˜ ๋Œ€๋ถ€๋ถ„์€ ๋น„๊ต์  ๋ฌดํ•ดํ–ˆ์ง€๋งŒ, ๊ฐ์ง€๋œ ์•”๊ธฐ์™€ ๊ด€๋ จ๋œ ์œ„ํ—˜์— ๋Œ€ํ•œ ๋ณด๋‹ค ์ฒด๊ณ„์ ์ธ ์—ฐ๊ตฌ๋Š” ์ด ์ž‘์—…์˜ ๋ฒ”์œ„๋ฅผ ๋ฒ—์–ด๋‚ฌ์Šต๋‹ˆ๋‹ค.
์‹คํ—˜์—์„œ๋Š” ์›์น˜ ์•Š๋Š” ์•”๊ธฐ ํ…์ŠคํŠธ(์˜ˆ: ์ „ํ™”๋ฒˆํ˜ธ), ๋ฌดํ•ดํ•œ ์•”๊ธฐ ํ…์ŠคํŠธ(์ผ๋ฐ˜์ ์ธ ๋ฌธ๊ตฌ), ์•”๊ธฐํ•˜๊ณ  ์‹ถ์€ ํ…์ŠคํŠธ(์˜ˆ: ์œ ๋ช…์ธ์˜ ๋ช…์–ธ)๋ฅผ ๊ตฌ๋ถ„ํ•˜์ง€ ์•Š๊ณ , ๋Œ€์‹  ํ•™์Šต ์„ธํŠธ์™€ ๋งค์šฐ ์œ ์‚ฌํ•œ ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” LM์˜ ๋ชจ๋“  ์ธ์Šคํ„ด์Šค๋ฅผ ๋ฌธ์ œ๊ฐ€ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์‹๋ณ„๋œ ์•”๊ธฐ ์ฝ˜ํ…์ธ ์˜ ๋Œ€๋ถ€๋ถ„์ด ๋น„๊ต์  ๋ฌดํ•ดํ•œ ๊ฒƒ์œผ๋กœ ์ •์„ฑ์ ์œผ๋กœ ๊ด€์ฐฐ๋˜์—ˆ์ง€๋งŒ, ๊ฐ์ง€๋œ ์•”๊ธฐ์™€ ๊ด€๋ จ๋œ ์œ„ํ—˜์— ๋Œ€ํ•œ ๋ณด๋‹ค ์ฒด๊ณ„์ ์ธ ์—ฐ๊ตฌ๋Š” ์ด ์ž‘์—…์˜ ๋ฒ”์œ„๋ฅผ ๋ฒ—์–ด๋‚ฌ์Šต๋‹ˆ๋‹ค.
Claim verification for evaluating summaries:Our paper also relates to prior work on claim verification, where claims are verified given reference to some knowledge source (Thorne et al., 2018; Wadden et al., 2020; Schuster et al., 2021). Min et al. (2023) propose FActScore, an LLM-based metric of factual precision in biography generation, which was expanded upon in SAFE (Wei et al., 2024). Manakul et al. (2023) propose SelfCheckGPT, which uses LLMs to evaluate the faithfulness of GPT-3 generated texts on a dataset of Wikipedia-style passages about people.
์ฃผ์žฅ ๊ฒ€์ฆ์„ ํ†ตํ•œ ์š”์•ฝ ํ‰๊ฐ€: ์šฐ๋ฆฌ์˜ ๋…ผ๋ฌธ์€ ๋˜ํ•œ ์ฃผ์žฅ ๊ฒ€์ฆ์— ๊ด€ํ•œ ์ด์ „ ์—ฐ๊ตฌ์™€ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์ฃผ์žฅ์€ ํŠน์ • ์ง€์‹ ์ถœ์ฒ˜๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ๊ฒ€์ฆ๋ฉ๋‹ˆ๋‹ค (Thorne et al., 2018; Wadden et al., 2020; Schuster et al., 2021). Min et al. (2023)์€ ์ „๊ธฐ ์ƒ์„ฑ์—์„œ ์‚ฌ์‹ค์  ์ •ํ™•์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” LLM ๊ธฐ๋ฐ˜ ๋ฉ”ํŠธ๋ฆญ์ธ FActScore๋ฅผ ์ œ์•ˆํ–ˆ์œผ๋ฉฐ, ์ด๋Š” SAFE (Wei et al., 2024)์—์„œ ํ™•์žฅ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. Manakul et al. (2023)์€ ์‚ฌ๋žŒ๋“ค์— ๋Œ€ํ•œ ์œ„ํ‚คํ”ผ๋””์•„ ์Šคํƒ€์ผ์˜ ๋ฌธ๋‹จ ๋ฐ์ดํ„ฐ์…‹์—์„œ GPT-3๊ฐ€ ์ƒ์„ฑํ•œ ํ…์ŠคํŠธ์˜ ์‹ ๋ขฐ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ์‚ฌ์šฉํ•˜๋Š” SelfCheckGPT๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค.
์š”์•ฝ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ํด๋ ˆ์ž„ ๊ฒ€์ฆ: ์ €ํฌ ๋…ผ๋ฌธ์€ ํด๋ ˆ์ž„์ด ์ง€์‹ ์†Œ์Šค์— ๋Œ€ํ•œ ์ฐธ์กฐ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ ๊ฒ€์ฆ๋˜๋Š” ํด๋ ˆ์ž„ ๊ฒ€์ฆ์— ๋Œ€ํ•œ ์„ ํ–‰ ์—ฐ๊ตฌ์™€๋„ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค(Thorne et al., 2018; Wadden et al., 2020; Schuster et al., 2021). Min et al. (2023)์€ ์ „๊ธฐ ์ƒ์„ฑ์—์„œ ์‚ฌ์‹ค์  ์ •ํ™•์„ฑ์— ๋Œ€ํ•œ LLM ๊ธฐ๋ฐ˜ ๋ฉ”ํŠธ๋ฆญ์ธ FActScore๋ฅผ ์ œ์•ˆํ–ˆ์œผ๋ฉฐ, ์ด๋Š” SAFE(Wei et al., 2024)์—์„œ ํ™•์žฅ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. Manakul et al. (2023)์€ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ๋žŒ์— ๋Œ€ํ•œ ์œ„ํ‚คํ”ผ๋””์•„ ์Šคํƒ€์ผ์˜ ๊ตฌ์ ˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ GPT-3 ์ƒ์„ฑ ํ…์ŠคํŠธ์˜ ์ถฉ์‹ค์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” SelfCheckGPT๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
Figure 2 gives an overview of the five models' "answer sheet" on the math problems. Our merged models retain the foundational knowledge in the source models, as evidenced by the similar score patterns on problems 1-15. Moreover, they exhibit emergent capabilities, successfully tackling problems that stumped the source models (e.g., problems 20-30). Evidently, by effectively integrating a Japanese LLM and mathematical models, we have succeeded in producing models that are proficient in both Japanese language understanding and mathematical problem-solving.
Figure 2๋Š” ์ˆ˜ํ•™ ๋ฌธ์ œ์— ๋Œ€ํ•œ ๋‹ค์„ฏ ๋ชจ๋ธ์˜ "๋‹ต์•ˆ์ง€" ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ํ†ตํ•ฉ ๋ชจ๋ธ์€ ๋ฌธ์ œ 1-15์—์„œ ์œ ์‚ฌํ•œ ์ ์ˆ˜ ํŒจํ„ด์„ ๋ณด์ž„์œผ๋กœ์จ ์†Œ์Šค ๋ชจ๋ธ์˜ ๊ธฐ์ดˆ ์ง€์‹์„ ์œ ์ง€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”์šฑ์ด, ์ด๋“ค์€ ์†Œ์Šค ๋ชจ๋ธ์ด ํ•ด๊ฒฐํ•˜์ง€ ๋ชปํ•œ ๋ฌธ์ œ๋“ค(์˜ˆ: ๋ฌธ์ œ 20-30)์„ ์„ฑ๊ณต์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ถ„๋ช…ํžˆ, ์ผ๋ณธ์–ด LLM๊ณผ ์ˆ˜ํ•™ ๋ชจ๋ธ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ†ตํ•ฉํ•จ์œผ๋กœ์จ, ์šฐ๋ฆฌ๋Š” ์ผ๋ณธ์–ด ์ดํ•ด์™€ ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ์— ๋ชจ๋‘ ๋Šฅ์ˆ™ํ•œ ๋ชจ๋ธ์„ ๋งŒ๋“œ๋Š” ๋ฐ ์„ฑ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2๋Š” ์ˆ˜ํ•™ ๋ฌธ์ œ์— ๋Œ€ํ•œ 5๊ฐ€์ง€ ๋ชจ๋ธ์˜ "๋‹ต์•ˆ์ง€"์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋ณ‘ํ•ฉ๋œ ๋ชจ๋ธ์€ ๋ฌธ์ œ 1~15์—์„œ ์œ ์‚ฌํ•œ ์ ์ˆ˜ ํŒจํ„ด์—์„œ ์•Œ ์ˆ˜ ์žˆ๋“ฏ์ด ์†Œ์Šค ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ์ง€์‹์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์†Œ์Šค ๋ชจ๋ธ์ด ํ’€์ง€ ๋ชปํ•œ ๋ฌธ์ œ(์˜ˆ: ๋ฌธ์ œ 20~30)๋ฅผ ์„ฑ๊ณต์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜์—ฌ ์ƒˆ๋กœ์šด ์—ญ๋Ÿ‰์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ๋ถ„๋ช…ํžˆ ์ผ๋ณธ์–ด LLM๊ณผ ์ˆ˜ํ•™์  ๋ชจ๋ธ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ†ตํ•ฉํ•จ์œผ๋กœ์จ ์ผ๋ณธ์–ด ์ดํ•ด์™€ ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ์— ๋ชจ๋‘ ๋Šฅ์ˆ™ํ•œ ๋ชจ๋ธ์„ ์ œ์ž‘ํ•˜๋Š” ๋ฐ ์„ฑ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค.
In the language modeling field, the current trend consists in scaling up the generative pretraining approach introduced with GPT-2, which implies training neural models made of several billions of parameters on gigantic web-mined text corpora (Brown et al., 2020; Touvron et al., 2023; Almazrouei et al., 2023; Jiang et al., 2023). However, training and serving such highly parameterized models raises energy and hardware-related problematics, which motivates for looking into achieving similar performance levels with smaller models (Sardana & Frankle, 2023).
์–ธ์–ด ๋ชจ๋ธ๋ง ๋ถ„์•ผ์—์„œ ํ˜„์žฌ์˜ ํŠธ๋ Œ๋“œ๋Š” GPT-2์™€ ํ•จ๊ป˜ ๋„์ž…๋œ ์ƒ์„ฑ์  ์‚ฌ์ „ ํ•™์Šต ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋กœ ๊ตฌ์„ฑ๋œ ์‹ ๊ฒฝ ๋ชจ๋ธ์„ ๊ฑฐ๋Œ€ํ•œ ์›น์—์„œ ์ˆ˜์ง‘ํ•œ ํ…์ŠคํŠธ ์ฝ”ํผ์Šค์— ๋Œ€ํ•ด ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค (Brown et al., 2020; Touvron et al., 2023; Almazrouei et al., 2023; Jiang et al., 2023). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๊ณ ๋„๋กœ ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ ๋งŽ์€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ์„œ๋น„์Šคํ•˜๋Š” ๊ฒƒ์€ ์—๋„ˆ์ง€ ๋ฐ ํ•˜๋“œ์›จ์–ด ๊ด€๋ จ ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•˜๋ฉฐ, ์ด๋Š” ๋” ์ž‘์€ ๋ชจ๋ธ๋กœ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ ์ˆ˜์ค€์„ ๋‹ฌ์„ฑํ•˜๋ ค๋Š” ๋™๊ธฐ๋ฅผ ๋ถ€์—ฌํ•ฉ๋‹ˆ๋‹ค (Sardana & Frankle, 2023).
์–ธ์–ด ๋ชจ๋ธ๋ง ๋ถ„์•ผ์—์„œ๋Š” GPT-2์™€ ํ•จ๊ป˜ ๋„์ž…๋œ ์ƒ์„ฑ์  ์‚ฌ์ „ ํ•™์Šต ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์ด ํ˜„์žฌ ์ถ”์„ธ์ด๋ฉฐ, ์ด๋Š” ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ ์‹ ๊ฒฝ ๋ชจ๋ธ์„ ๊ฑฐ๋Œ€ํ•œ ์›น ๋งˆ์ด๋‹ ํ…์ŠคํŠธ ์ฝ”ํผ์Šค(Brown et al., 2020; Touvron et al., 2023; Almazrouei et al., 2023; Jiang et al., 2023)์— ๋Œ€ํ•ด ํ•™์Šต์‹œํ‚ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๊ณ ๋งค๊ฐœ๋ณ€์ˆ˜ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ์„œ๋น„์Šคํ•˜๋ฉด ์—๋„ˆ์ง€ ๋ฐ ํ•˜๋“œ์›จ์–ด ๊ด€๋ จ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋” ์ž‘์€ ๋ชจ๋ธ๋กœ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ ์ˆ˜์ค€์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค(Sardana & Frankle, 2023).
In addition, recent work also explores multi-agent collaboration based on LLMs in simulated environments [700, 701, 533]. These studies simulate human social behaviors by instantiating multiple LLM-based agents with observations, planning, and memories in a sandbox environment. In controlled evaluation, the abilities of generative agents to search, plan, and think are evaluated by humans in an interview-like manner. Further, they also conduct descriptive measurements on multiple agents within a simulated environment to examine emergent social behaviors.
๋˜ํ•œ, ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ํ™˜๊ฒฝ์—์„œ LLMs ๊ธฐ๋ฐ˜์˜ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ํ˜‘์—…์„ ํƒ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค [700, 701, 533]. ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ๋“ค์€ ์ƒŒ๋“œ๋ฐ•์Šค ํ™˜๊ฒฝ์—์„œ ๊ด€์ฐฐ, ๊ณ„ํš ๋ฐ ๊ธฐ์–ต์„ ๊ฐ€์ง„ ๋‹ค์ค‘ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋ฅผ ๊ตฌํ˜„ํ•˜์—ฌ ์ธ๊ฐ„์˜ ์‚ฌํšŒ์  ํ–‰๋™์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค. ํ†ต์ œ๋œ ํ‰๊ฐ€์—์„œ๋Š” ์ธํ„ฐ๋ทฐ์™€ ์œ ์‚ฌํ•œ ๋ฐฉ์‹์œผ๋กœ ์ƒ์„ฑ์  ์—์ด์ „ํŠธ์˜ ๊ฒ€์ƒ‰, ๊ณ„ํš ๋ฐ ์‚ฌ๊ณ  ๋Šฅ๋ ฅ์„ ์ธ๊ฐ„์ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋” ๋‚˜์•„๊ฐ€, ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ํ™˜๊ฒฝ ๋‚ด์—์„œ ๋‹ค์ค‘ ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ๊ธฐ์ˆ ์  ์ธก์ •์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ๋ฐœ์ƒํ•˜๋Š” ์‚ฌํšŒ์  ํ–‰๋™์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
๋˜ํ•œ ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์—์„œ LLM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ํ˜‘์—…์„ ํƒ๊ตฌํ•˜๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค[700, 701, 533]. ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ์—์„œ๋Š” ์ƒŒ๋“œ๋ฐ•์Šค ํ™˜๊ฒฝ์—์„œ ๊ด€์ฐฐ, ๊ณ„ํš ๋ฐ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ํ†ตํ•ด ์—ฌ๋Ÿฌ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋ฅผ ์ธ์Šคํ„ด์Šคํ™”ํ•˜์—ฌ ์ธ๊ฐ„์˜ ์‚ฌํšŒ์  ํ–‰๋™์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค. ํ†ต์ œ๋œ ํ‰๊ฐ€์—์„œ ์ƒ์„ฑ ์—์ด์ „ํŠธ์˜ ๊ฒ€์ƒ‰, ๊ณ„ํš ๋ฐ ์‚ฌ๊ณ  ๋Šฅ๋ ฅ์€ ์ธํ„ฐ๋ทฐ์™€ ๊ฐ™์€ ๋ฐฉ์‹์œผ๋กœ ์ธ๊ฐ„์— ์˜ํ•ด ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ ๋‚ด์—์„œ ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ์„ค๋ช…์  ์ธก์ •์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ƒˆ๋กœ์šด ์‚ฌํšŒ์  ํ–‰๋™์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
Text extraction.We want to extract only the main content of the page, ignoring menus, headers, footers, and ads among others: Lopukhin (2019) found that trafilatura (Barbaresi, 2021) was the best non-commercial library for retrieving content from blog posts and news articles. Although this is only a narrow subset of the kind of pages making up CommonCrawl, we found this finding to hold more broadly. We use trafilatura for text extraction, and apply extra formatting via regular expressions: we limit new lines to two consecutive ones, and remove all URLs.
ํ…์ŠคํŠธ ์ถ”์ถœ. ์šฐ๋ฆฌ๋Š” ๋ฉ”๋‰ด, ํ—ค๋”, ํ‘ธํ„ฐ, ๊ด‘๊ณ  ๋“ฑ์„ ๋ฌด์‹œํ•˜๊ณ  ํŽ˜์ด์ง€์˜ ์ฃผ์š” ์ฝ˜ํ…์ธ ๋งŒ ์ถ”์ถœํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. Lopukhin (2019)์€ trafilatura (Barbaresi, 2021)๊ฐ€ ๋ธ”๋กœ๊ทธ ๊ฒŒ์‹œ๋ฌผ ๋ฐ ๋‰ด์Šค ๊ธฐ์‚ฌ์—์„œ ์ฝ˜ํ…์ธ ๋ฅผ ๊ฐ€์ ธ์˜ค๋Š” ๋ฐ ๊ฐ€์žฅ ์ข‹์€ ๋น„์ƒ์—…์  ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ผ๊ณ  ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋น„๋ก ์ด๊ฒƒ์ด CommonCrawl์„ ๊ตฌ์„ฑํ•˜๋Š” ํŽ˜์ด์ง€์˜ ์ข์€ ํ•˜์œ„ ์ง‘ํ•ฉ์— ๋ถˆ๊ณผํ•˜์ง€๋งŒ, ์šฐ๋ฆฌ๋Š” ์ด ๊ฒฐ๊ณผ๊ฐ€ ๋” ๋„“๊ฒŒ ์ ์šฉ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ํ…์ŠคํŠธ ์ถ”์ถœ์„ ์œ„ํ•ด trafilatura๋ฅผ ์‚ฌ์šฉํ•˜๊ณ , ์ •๊ทœ ํ‘œํ˜„์‹์„ ํ†ตํ•ด ์ถ”๊ฐ€ ํ˜•์‹์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค: ์ƒˆ๋กœ์šด ์ค„์„ ๋‘ ๊ฐœ์˜ ์—ฐ์†๋œ ์ค„๋กœ ์ œํ•œํ•˜๊ณ  ๋ชจ๋“  URL์„ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
ํ…์ŠคํŠธ ์ถ”์ถœ.๋ฉ”๋‰ด, ํ—ค๋”, ๋ฐ”๋‹ฅ๊ธ€, ๊ด‘๊ณ  ๋“ฑ์„ ๋ฌด์‹œํ•˜๊ณ  ํŽ˜์ด์ง€์˜ ์ฃผ์š” ์ฝ˜ํ…์ธ ๋งŒ ์ถ”์ถœํ•˜๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค: Lopukhin(2019)์€ ๋ธ”๋กœ๊ทธ ๊ฒŒ์‹œ๋ฌผ๊ณผ ๋‰ด์Šค ๊ธฐ์‚ฌ์—์„œ ์ฝ˜ํ…์ธ ๋ฅผ ๊ฒ€์ƒ‰ํ•˜๋Š” ๋ฐ ์žˆ์–ด ๋น„์ƒ์—…์šฉ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์ค‘ trafilatura(Barbaresi, 2021)๊ฐ€ ๊ฐ€์žฅ ์ข‹๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” CommonCrawl์„ ๊ตฌ์„ฑํ•˜๋Š” ํŽ˜์ด์ง€์˜ ์ผ๋ถ€์— ๋ถˆ๊ณผํ•˜์ง€๋งŒ, ์ด ๊ฒฐ๊ณผ๋Š” ๋” ํญ๋„“๊ฒŒ ์ ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ํ…์ŠคํŠธ ์ถ”์ถœ์—๋Š” trafilatura๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ์ •๊ทœ์‹์„ ํ†ตํ•ด ์ถ”๊ฐ€ ์„œ์‹์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์ƒˆ ์ค„์€ ์—ฐ์†๋œ ๋‘ ์ค„๋กœ ์ œํ•œํ•˜๊ณ  ๋ชจ๋“  URL์„ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
This work received partial support from the Basque Government through research group funding IT1805-22 and the ICL4LANG project (grant no. KK-2023/00094). Additionally, we acknowledge the support of several MCIN/AEI/10.13039/501100011033 projects: (i) DeepKnowledge (PID2021-127777/DB-C21) and funding from FEDER, EU; (ii) AWARE (TED2021-131617B-I00) and support from the European Union NextGenerationEU/PRTR. We express our gratitude to Carlos Dominguez for his assistance in the experimental setup and to Eneko Agirre for his valuable feedback and guidance.
์ด ์—ฐ๊ตฌ๋Š” Basque ์ •๋ถ€์˜ ์—ฐ๊ตฌ ๊ทธ๋ฃน ์ž๊ธˆ IT1805-22์™€ ICL4LANG ํ”„๋กœ์ ํŠธ (grant no. KK-2023/00094)๋ฅผ ํ†ตํ•ด ๋ถ€๋ถ„์ ์ธ ์ง€์›์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์—ฌ๋Ÿฌ MCIN/AEI/10.13039/501100011033 ํ”„๋กœ์ ํŠธ์˜ ์ง€์›์„ ์ธ์ •ํ•ฉ๋‹ˆ๋‹ค: (i) DeepKnowledge (PID2021-127777/DB-C21)์™€ FEDER, EU์˜ ์ž๊ธˆ ์ง€์›; (ii) AWARE (TED2021-131617B-I00)์™€ European Union NextGenerationEU/PRTR์˜ ์ง€์›. ์‹คํ—˜ ์„ค์ •์— ๋„์›€์„ ์ค€ Carlos Dominguez์™€ ๊ท€์ค‘ํ•œ ํ”ผ๋“œ๋ฐฑ๊ณผ ์ง€๋„๋ฅผ ์ œ๊ณตํ•œ Eneko Agirre์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋œป์„ ์ „ํ•ฉ๋‹ˆ๋‹ค.
์ด ์ž‘์—…์€ ์—ฐ๊ตฌ ๊ทธ๋ฃน ๊ธฐ๊ธˆ IT1805-22 ๋ฐ ICL4LANG ํ”„๋กœ์ ํŠธ(๋ณด์กฐ๊ธˆ ๋ฒˆํ˜ธ KK-2023/00094)๋ฅผ ํ†ตํ•ด ๋ฐ”์Šคํฌ ์ •๋ถ€์˜ ๋ถ€๋ถ„์ ์ธ ์ง€์›์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ MCIN/AEI/10.13039/501100011033 ํ”„๋กœ์ ํŠธ์˜ ์—ฌ๋Ÿฌ ์ง€์›์„ ์ธ์ •ํ•ฉ๋‹ˆ๋‹ค: (i) DeepKnowledge(PID2021-127777/DB-C21) ๋ฐ FEDER, EU์˜ ์ง€์›๊ธˆ, (ii) AWARE(TED2021-131617B-I00) ๋ฐ ์œ ๋Ÿฝ์—ฐํ•ฉ NextGenerationEU/PRTR์˜ ์ง€์›์„ ์ธ์ •ํ•ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ์„ค์ •์„ ๋„์™€์ค€ Carlos Dominguez์—๊ฒŒ ๊ฐ์‚ฌ๋“œ๋ฆฌ๋ฉฐ, ๊ท€์ค‘ํ•œ ํ”ผ๋“œ๋ฐฑ๊ณผ ์ง€๋„๋ฅผ ์ œ๊ณตํ•ด์ค€ Eneko Agirre์—๊ฒŒ๋„ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
Finally, we compare the mean tree depth (measured by the mean over sentences of the depth of the dependency tree), and mean dependency distance (measured as the average dependency distance of any pair of words within a sentence) in Figure 11, which have been shown to be good measures of syntactic difficulty Futrell et al. (2015); Gibson et al. (2000); Oya (2021). We find similar trends as for reading level and TTR diversity where mediums style increase depth, mdd, and syntactic complexity in general. We find again that QA style reduces this complexity.
๋งˆ์ง€๋ง‰์œผ๋กœ, Figure 11์—์„œ ํ‰๊ท  ํŠธ๋ฆฌ ๊นŠ์ด(์˜์กด ํŠธ๋ฆฌ์˜ ๊นŠ์ด์— ๋Œ€ํ•œ ๋ฌธ์žฅ๋“ค์˜ ํ‰๊ท ์œผ๋กœ ์ธก์ •)์™€ ํ‰๊ท  ์˜์กด ๊ฑฐ๋ฆฌ(๋ฌธ์žฅ ๋‚ด์˜ ๋‹จ์–ด ์Œ ๊ฐ„์˜ ํ‰๊ท  ์˜์กด ๊ฑฐ๋ฆฌ๋กœ ์ธก์ •)๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” Futrell et al. (2015); Gibson et al. (2000); Oya (2021)์—์„œ ๊ตฌ๋ฌธ์  ๋‚œ์ด๋„์˜ ์ข‹์€ ์ธก์ •์น˜๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋…์„œ ์ˆ˜์ค€๊ณผ TTR ๋‹ค์–‘์„ฑ์—์„œ์™€ ์œ ์‚ฌํ•œ ๊ฒฝํ–ฅ์„ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ, ๋งค์ฒด ์Šคํƒ€์ผ์€ ๊นŠ์ด, mdd, ๊ทธ๋ฆฌ๊ณ  ์ „๋ฐ˜์ ์ธ ๊ตฌ๋ฌธ ๋ณต์žก์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, QA ์Šคํƒ€์ผ์€ ์ด๋Ÿฌํ•œ ๋ณต์žก์„ฑ์„ ๊ฐ์†Œ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
๋งˆ์ง€๋ง‰์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๊ตฌ๋ฌธ๋ก ์  ์–ด๋ ค์›€์˜ ์ข‹์€ ์ฒ™๋„๋กœ ๋‚˜ํƒ€๋‚œ ํ‰๊ท  ํŠธ๋ฆฌ ๊นŠ์ด(์˜์กด์„ฑ ํŠธ๋ฆฌ ๊นŠ์ด์˜ ๋ฌธ์žฅ ํ‰๊ท ์œผ๋กœ ์ธก์ •)์™€ ํ‰๊ท  ์˜์กด์„ฑ ๊ฑฐ๋ฆฌ(๋ฌธ์žฅ ๋‚ด ์ž„์˜์˜ ๋‹จ์–ด ์Œ์˜ ํ‰๊ท  ์˜์กด์„ฑ ๊ฑฐ๋ฆฌ๋กœ ์ธก์ •)๋ฅผ ๊ทธ๋ฆผ 11์—์„œ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. Futrell et al. (2015); Gibson et al. (2000); Oya (2021). ์šฐ๋ฆฌ๋Š” ๋ฏธ๋””์–ด ์Šคํƒ€์ผ์ด ์ผ๋ฐ˜์ ์œผ๋กœ ๊นŠ์ด, mdd ๋ฐ ๊ตฌ๋ฌธ ๋ณต์žก์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ค๋Š” ๋…ํ•ด ์ˆ˜์ค€ ๋ฐ TTR ๋‹ค์–‘์„ฑ๊ณผ ์œ ์‚ฌํ•œ ์ถ”์„ธ๋ฅผ ๋ฐœ๊ฒฌํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์‹œ ํ•œ ๋ฒˆ QA ์Šคํƒ€์ผ์ด ์ด๋Ÿฌํ•œ ๋ณต์žก์„ฑ์„ ์ค„์ธ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
We consider our model's ability to reconstruct specific entities from prompts that have a high likelihood of containing personal information, such as names, dates, and nationalities. To test this, we generate a synthetic dataset of prompts that contain these attributes. We edit prompts from Instructions-2M with private attributes sourced from Wikibio (Lebret et al., 2016) and Synthbio (Yuan et al., 2022)11. We invert these prompts using both our LLAMA 7B and LLAMA-chat 7B models and measure accuracy at reconstructing private entities across categories.
์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์ด๋ฆ„, ๋‚ ์งœ, ๊ตญ์  ๋“ฑ ๊ฐœ์ธ ์ •๋ณด๋ฅผ ํฌํ•จํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ”„๋กฌํ”„ํŠธ์—์„œ ํŠน์ • ์—”ํ‹ฐํ‹ฐ๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ์ด๋Ÿฌํ•œ ์†์„ฑ์„ ํฌํ•จํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์…‹์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” Wikibio (Lebret et al., 2016)์™€ Synthbio (Yuan et al., 2022)11์—์„œ ๊ฐ€์ ธ์˜จ ๊ฐœ์ธ ์†์„ฑ์„ ์‚ฌ์šฉํ•˜์—ฌ Instructions-2M์˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํŽธ์ง‘ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ LLAMA 7B์™€ LLAMA-chat 7B ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋Ÿฌํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฐ˜์ „์‹œํ‚ค๊ณ , ์นดํ…Œ๊ณ ๋ฆฌ๋ณ„๋กœ ๊ฐœ์ธ ์—”ํ‹ฐํ‹ฐ๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ์ •ํ™•๋„๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค.
์ด๋ฆ„, ๋‚ ์งœ, ๊ตญ์  ๋“ฑ ๊ฐœ์ธ ์ •๋ณด๊ฐ€ ํฌํ•จ๋  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ”„๋กฌํ”„ํŠธ์—์„œ ํŠน์ • ๊ฐœ์ฒด๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ์ด๋Ÿฌํ•œ ์†์„ฑ์„ ํฌํ•จํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ฐœ์ธ ์ •๋ณด๊ฐ€ ํฌํ•จ๋œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด Wikibio(Lebret et al., 2016)์™€ Synthbio(Yuan et al., 2022)11์—์„œ ๊ฐ€์ ธ์˜จ ๊ฐœ์ธ ์†์„ฑ์„ ์‚ฌ์šฉํ•˜์—ฌ Instructions-2M์˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ˆ˜์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ LLAMA 7B์™€ LLAMA-chat 7B ๋ชจ๋ธ์„ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ˜์ „ํ•˜๊ณ  ์นดํ…Œ๊ณ ๋ฆฌ๋ณ„๋กœ ๊ฐœ์ธ ์—”ํ‹ฐํ‹ฐ๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ์ •ํ™•๋„๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค.
Reading the data.CommonCrawl is available in either WARC (raw HTML response), or WET files (preprocessed to only include plain text). Individual files correspond to a page at a given URL; these constitute single documents/samples. Working with WET files would spare us from running our own HTML extraction; however, in line with previous works (Gao et al., 2020; Rae et al., 2021), we found WET files to include undesirable navigation menus, ads, and other irrelevant texts. Accordingly, our pipeline starts from raw WARC files, read with the wavcio library.
๋ฐ์ดํ„ฐ ์ฝ๊ธฐ. CommonCrawl์€ WARC(์›์‹œ HTML ์‘๋‹ต) ๋˜๋Š” WET ํŒŒ์ผ(์ „์ฒ˜๋ฆฌ๋˜์–ด ์ผ๋ฐ˜ ํ…์ŠคํŠธ๋งŒ ํฌํ•จ)๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ๊ฐœ๋ณ„ ํŒŒ์ผ์€ ์ฃผ์–ด์ง„ URL์˜ ํŽ˜์ด์ง€์— ํ•ด๋‹นํ•˜๋ฉฐ, ์ด๋Š” ๋‹จ์ผ ๋ฌธ์„œ/์ƒ˜ํ”Œ์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. WET ํŒŒ์ผ์„ ์‚ฌ์šฉํ•˜๋ฉด ์ž์ฒด HTML ์ถ”์ถœ์„ ์‹คํ–‰ํ•  ํ•„์š”๊ฐ€ ์—†์ง€๋งŒ, ์ด์ „ ์—ฐ๊ตฌ(Gao et al., 2020; Rae et al., 2021)์™€ ์ผ์น˜ํ•˜๊ฒŒ, WET ํŒŒ์ผ์—๋Š” ์›ํ•˜์ง€ ์•Š๋Š” ํƒ์ƒ‰ ๋ฉ”๋‰ด, ๊ด‘๊ณ  ๋ฐ ๊ธฐํƒ€ ๊ด€๋ จ ์—†๋Š” ํ…์ŠคํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ์˜ ํŒŒ์ดํ”„๋ผ์ธ์€ wavcio ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋กœ ์ฝ์€ ์›์‹œ WARC ํŒŒ์ผ์—์„œ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ ์ฝ๊ธฐ.CommonCrawl์€ WARC(์›์‹œ HTML ์‘๋‹ต) ๋˜๋Š” WET ํŒŒ์ผ(์ผ๋ฐ˜ ํ…์ŠคํŠธ๋งŒ ํฌํ•จํ•˜๋„๋ก ์‚ฌ์ „ ์ฒ˜๋ฆฌ๋œ)๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ๊ฐœ๋ณ„ ํŒŒ์ผ์€ ํŠน์ • URL์˜ ํŽ˜์ด์ง€์— ํ•ด๋‹นํ•˜๋ฉฐ, ์ด๋Š” ๋‹จ์ผ ๋ฌธ์„œ/์ƒ˜ํ”Œ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. WET ํŒŒ์ผ์„ ์‚ฌ์šฉํ•˜๋ฉด HTML ์ถ”์ถœ์„ ์ง์ ‘ ์‹คํ–‰ํ•  ํ•„์š”๊ฐ€ ์—†์ง€๋งŒ, ์ด์ „ ์—ฐ๊ตฌ(Gao et al., 2020; Rae et al., 2021)์— ๋”ฐ๋ฅด๋ฉด WET ํŒŒ์ผ์— ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์€ ํƒ์ƒ‰ ๋ฉ”๋‰ด, ๊ด‘๊ณ  ๋ฐ ๊ธฐํƒ€ ๊ด€๋ จ ์—†๋Š” ํ…์ŠคํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํŒŒ์ดํ”„๋ผ์ธ์€ wavcio ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋กœ ์ฝ์€ ์›์‹œ WARC ํŒŒ์ผ์—์„œ ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค.
Knowledge distillation is the process of learning from a larger model [143]. Earlier days of best-performing models release have proven that this approach is very useful even if it is used in an API distillation approach. It is also referred to as an approach to distill the knowledge of not a single model but in fact multiple models into a smaller one. Creating smaller models by this approach yields smaller model sizes that can be used even on edge devices. Knowledge distillation as shown in Fig 35, illustrates a general setup of this training scheme.
์ง€์‹ ์ฆ๋ฅ˜(Knowledge distillation)๋Š” ๋” ํฐ ๋ชจ๋ธ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๋Š” ๊ณผ์ •์ž…๋‹ˆ๋‹ค [143]. ์„ฑ๋Šฅ์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ๋ชจ๋ธ๋“ค์ด ์ถœ์‹œ๋œ ์ดˆ๊ธฐ์—๋Š” ์ด ์ ‘๊ทผ ๋ฐฉ์‹์ด API ์ฆ๋ฅ˜ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์šฉ๋˜๋”๋ผ๋„ ๋งค์šฐ ์œ ์šฉํ•˜๋‹ค๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋‹จ์ผ ๋ชจ๋ธ์ด ์•„๋‹Œ ์—ฌ๋Ÿฌ ๋ชจ๋ธ์˜ ์ง€์‹์„ ๋” ์ž‘์€ ๋ชจ๋ธ๋กœ ์ฆ๋ฅ˜ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ๋„ ์–ธ๊ธ‰๋ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ์ƒ์„ฑ๋œ ๋” ์ž‘์€ ๋ชจ๋ธ๋“ค์€ ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค์—์„œ๋„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ž‘์€ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ๊ฐ€์ง‘๋‹ˆ๋‹ค. Fig 35์— ๋‚˜ํƒ€๋‚œ ๊ฒƒ์ฒ˜๋Ÿผ, ์ง€์‹ ์ฆ๋ฅ˜๋Š” ์ด ํ›ˆ๋ จ ์ฒด๊ณ„์˜ ์ผ๋ฐ˜์ ์ธ ์„ค์ •์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์ง€์‹ ์ฆ๋ฅ˜๋Š” ๋” ํฐ ๋ชจ๋ธ์—์„œ ํ•™์Šตํ•˜๋Š” ๊ณผ์ •์ž…๋‹ˆ๋‹ค[143]. ์ตœ๊ณ  ์„ฑ๋Šฅ ๋ชจ๋ธ ์ถœ์‹œ ์ดˆ์ฐฝ๊ธฐ์—๋Š” ์ด ์ ‘๊ทผ ๋ฐฉ์‹์ด API ์ฆ๋ฅ˜ ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์šฉ๋˜๋”๋ผ๋„ ๋งค์šฐ ์œ ์šฉํ•˜๋‹ค๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์‹ค์ œ๋กœ ํ•˜๋‚˜์˜ ๋ชจ๋ธ์ด ์•„๋‹Œ ์—ฌ๋Ÿฌ ๋ชจ๋ธ์˜ ์ง€์‹์„ ๋” ์ž‘์€ ๋ชจ๋ธ๋กœ ์ฆ๋ฅ˜ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์ด๋ผ๊ณ ๋„ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ๋” ์ž‘์€ ๋ชจ๋ธ์„ ๋งŒ๋“ค๋ฉด ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค์—์„œ๋„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋” ์ž‘์€ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 35์— ๋‚˜์™€ ์žˆ๋Š” ์ง€์‹ ์ฆ๋ฅ˜๋Š” ์ด ํ›ˆ๋ จ ์ฒด๊ณ„์˜ ์ผ๋ฐ˜์ ์ธ ์„ค์ •์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
In this survey, we systematically summarize existing research in the field of LLM-based autonomous agents. We present and review these studies from three aspects including the construction, application, and evaluation of the agents. For each of these aspects, we provide a detailed taxonomy to draw connections among the existing research, summarizing the major techniques and their development histories. In addition to reviewing the previous work, we also propose several challenges in this field, which are expected to guide potential future directions.
์ด ์„ค๋ฌธ ์กฐ์‚ฌ์—์„œ๋Š” LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ ๋ถ„์•ผ์˜ ๊ธฐ์กด ์—ฐ๊ตฌ๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์˜ ๊ตฌ์„ฑ, ์‘์šฉ, ํ‰๊ฐ€๋ฅผ ํฌํ•จํ•œ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์—์„œ ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ๋ฅผ ์ œ์‹œํ•˜๊ณ  ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์ธก๋ฉด์— ๋Œ€ํ•ด ๊ธฐ์กด ์—ฐ๊ตฌ ๊ฐ„์˜ ์—ฐ๊ฒฐ์„ ๋„์ถœํ•˜๊ธฐ ์œ„ํ•ด ์ž์„ธํ•œ ๋ถ„๋ฅ˜ ์ฒด๊ณ„๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์ฃผ์š” ๊ธฐ์ˆ ๊ณผ ๊ทธ ๋ฐœ์ „ ์—ญ์‚ฌ๋ฅผ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ๋ฅผ ๊ฒ€ํ† ํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„, ์ด ๋ถ„์•ผ์—์„œ ์˜ˆ์ƒ๋˜๋Š” ์—ฌ๋Ÿฌ ๋„์ „ ๊ณผ์ œ๋ฅผ ์ œ์•ˆํ•˜์—ฌ ์ž ์žฌ์ ์ธ ๋ฏธ๋ž˜ ๋ฐฉํ–ฅ์„ ์•ˆ๋‚ดํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.
์ด ์„ค๋ฌธ์กฐ์‚ฌ์—์„œ๋Š” LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ ๋ถ„์•ผ์˜ ๊ธฐ์กด ์—ฐ๊ตฌ๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์˜ ๊ตฌ์„ฑ, ์ ์šฉ ๋ฐ ํ‰๊ฐ€์˜ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์—์„œ ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ๋ฅผ ์ œ์‹œํ•˜๊ณ  ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์ธก๋ฉด์— ๋Œ€ํ•ด ์„ธ๋ถ€์ ์ธ ๋ถ„๋ฅ˜๋ฅผ ์ œ๊ณตํ•˜์—ฌ ๊ธฐ์กด ์—ฐ๊ตฌ ๊ฐ„์˜ ์—ฐ๊ด€์„ฑ์„ ๋„์ถœํ•˜๊ณ  ์ฃผ์š” ๊ธฐ์ˆ ๊ณผ ๊ทธ ๊ฐœ๋ฐœ ์—ญ์‚ฌ๋ฅผ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ์ž‘์—…์„ ๊ฒ€ํ† ํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ์ด ๋ถ„์•ผ์˜ ๋ช‡ ๊ฐ€์ง€ ๊ณผ์ œ๋ฅผ ์ œ์•ˆํ•˜์—ฌ ์ž ์žฌ์ ์ธ ํ–ฅํ›„ ๋ฐฉํ–ฅ์„ ์ œ์‹œํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค.
MiniCPM was initially published as a technical blog on February 1st, 2024. Since then, we have received numerous insightful feedback from the community, significantly contributing to the development of this paper. We extend our gratitude to Chunting Zhou and Armen Aghajanyan for their valuable discussions. Special thanks go to Feigin Sun and Yan Wang for their meticulous feedback on clarifying ambiguities in the blog. Additionally, we appreciate the open-source community's efforts in integrating MiniCPM into inference frameworks like llama.cpp, etc.
MiniCPM์€ 2024๋…„ 2์›” 1์ผ์— ๊ธฐ์ˆ  ๋ธ”๋กœ๊ทธ๋กœ ์ฒ˜์Œ ๊ฒŒ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์ดํ›„๋กœ ์ปค๋ฎค๋‹ˆํ‹ฐ๋กœ๋ถ€ํ„ฐ ๋งŽ์€ ํ†ต์ฐฐ๋ ฅ ์žˆ๋Š” ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์•˜์œผ๋ฉฐ, ์ด๋Š” ์ด ๋…ผ๋ฌธ์˜ ๋ฐœ์ „์— ํฌ๊ฒŒ ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ท€์ค‘ํ•œ ๋…ผ์˜๋ฅผ ์ œ๊ณตํ•ด์ฃผ์‹  Chunting Zhou์™€ Armen Aghajanyan์—๊ฒŒ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ๋ธ”๋กœ๊ทธ์˜ ๋ชจํ˜ธ์„ฑ์„ ๋ช…ํ™•ํžˆ ํ•˜๋Š” ๋ฐ ์žˆ์–ด ์„ธ์‹ฌํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ฃผ์‹  Feigin Sun๊ณผ Yan Wang์—๊ฒŒ ํŠน๋ณ„ํžˆ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ๋˜ํ•œ, MiniCPM์„ llama.cpp์™€ ๊ฐ™์€ ์ถ”๋ก  ํ”„๋ ˆ์ž„์›Œํฌ์— ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ๊ธฐ์—ฌํ•œ ์˜คํ”ˆ ์†Œ์Šค ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ๋…ธ๋ ฅ์—๋„ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
MiniCPM์€ 2024๋…„ 2์›” 1์ผ์— ๊ธฐ์ˆ  ๋ธ”๋กœ๊ทธ๋กœ ์ฒ˜์Œ ๊ฒŒ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์ดํ›„๋กœ ์ปค๋ฎค๋‹ˆํ‹ฐ๋กœ๋ถ€ํ„ฐ ์ˆ˜๋งŽ์€ ํ†ต์ฐฐ๋ ฅ ์žˆ๋Š” ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์•„ ์ด ๋…ผ๋ฌธ์˜ ๊ฐœ๋ฐœ์— ํฌ๊ฒŒ ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ท€์ค‘ํ•œ ๋…ผ์˜๋ฅผ ์ œ๊ณตํ•œ ์ฒญํŒ… ์ €์šฐ์™€ ์•„๋ฅด๋ฉ˜ ์•„๊ฐ€์ž์–€์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ ๋ธ”๋กœ๊ทธ์˜ ๋ชจํ˜ธํ•œ ๋ถ€๋ถ„์„ ๋ช…ํ™•ํžˆ ํ•˜๋Š” ๋ฐ ์„ธ์‹ฌํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•œ ํ”ผ๊ธด ์„ ๊ณผ ์˜Œ ์™•์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, MiniCPM์„ llama.cpp ๋“ฑ๊ณผ ๊ฐ™์€ ์ถ”๋ก  ํ”„๋ ˆ์ž„์›Œํฌ์— ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ๋…ธ๋ ฅํ•œ ์˜คํ”ˆ ์†Œ์Šค ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ๋…ธ๋ ฅ์— ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
**GPT-4 Judgement:** The first step is incorrect; '2/.5' suggests division when it should be multiplication, '2*.5'. The correct calculation for the reading time is 2/2 hours, which equals 1 hour, not '2/.5' hours. However, the numerical result '1 hour' happens to be correct despite the incorrect operation shown. The remaining steps correctly add the two activities to find the total daily time spent, correctly multiply by the number of occurrences per week, and correctly multiply by the number of weeks to find the total hours spent over four weeks.
**GPT-4 ํŒ๋‹จ:** ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„๋Š” ์ž˜๋ชป๋˜์—ˆ์Šต๋‹ˆ๋‹ค. '2/.5'๋Š” ๊ณฑ์…ˆ์ด ๋˜์–ด์•ผ ํ•  ๋•Œ ๋‚˜๋ˆ—์…ˆ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค, ์ฆ‰ '2*.5'์ž…๋‹ˆ๋‹ค. ๋…์„œ ์‹œ๊ฐ„์— ๋Œ€ํ•œ ์˜ฌ๋ฐ”๋ฅธ ๊ณ„์‚ฐ์€ 2/2 ์‹œ๊ฐ„์œผ๋กœ, ์ด๋Š” 1์‹œ๊ฐ„์— ํ•ด๋‹นํ•˜๋ฉฐ, '2/.5' ์‹œ๊ฐ„์ด ์•„๋‹™๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ž˜๋ชป๋œ ์—ฐ์‚ฐ์ด ํ‘œ์‹œ๋˜์—ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ˆซ์ž ๊ฒฐ๊ณผ '1์‹œ๊ฐ„'์€ ์˜ฌ๋ฐ”๋ฆ…๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€ ๋‹จ๊ณ„๋“ค์€ ๋‘ ํ™œ๋™์„ ๋”ํ•˜์—ฌ ํ•˜๋ฃจ์— ์†Œ๋น„๋œ ์ด ์‹œ๊ฐ„์„ ์ฐพ๊ณ , ์ฃผ๋‹น ๋ฐœ์ƒ ํšŸ์ˆ˜์— ๋งž๊ฒŒ ๊ณฑํ•˜๊ณ , 4์ฃผ ๋™์•ˆ ์†Œ๋น„๋œ ์ด ์‹œ๊ฐ„์„ ์ฐพ๊ธฐ ์œ„ํ•ด ์ฃผ ์ˆ˜์— ๋งž๊ฒŒ ๊ณฑํ•˜๋Š” ๊ณผ์ •์„ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
**GPT-4 ํŒ๋‹จ:** ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„๊ฐ€ ์ž˜๋ชป๋˜์—ˆ์Šต๋‹ˆ๋‹ค. '2/.5'๋Š” ๊ณฑ์…ˆ์ด์–ด์•ผ ํ•˜๋Š” ๋‚˜๋ˆ—์…ˆ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์˜ฌ๋ฐ”๋ฅธ ์ฝ๊ธฐ ์‹œ๊ฐ„ ๊ณ„์‚ฐ์€ 2/2์‹œ๊ฐ„์œผ๋กœ '2/.5'์‹œ๊ฐ„์ด ์•„๋‹Œ 1์‹œ๊ฐ„์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ž˜๋ชป๋œ ์—ฐ์‚ฐ์ด ํ‘œ์‹œ๋˜์—ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ˆซ์ž ๊ฒฐ๊ณผ '1์‹œ๊ฐ„'์€ ์ •๋‹ต์ž…๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€ ๋‹จ๊ณ„์—์„œ๋Š” ๋‘ ํ™œ๋™์„ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๋”ํ•˜์—ฌ ์ผ์ผ ์ด ์‹œ๊ฐ„์„ ๊ตฌํ•˜๊ณ , ์ผ์ฃผ์ผ์— ๋ฐœ์ƒํ•˜๋Š” ํšŸ์ˆ˜๋ฅผ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๊ณฑํ•˜๊ณ , 4์ฃผ ๋™์•ˆ ์ด ์‹œ๊ฐ„์„ ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ์ฃผ ์ˆ˜๋ฅผ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๊ณฑํ•ฉ๋‹ˆ๋‹ค.
While larger vocabulary sizes introduce additional computational costs, they might also result in lower fertility scores and, therefore, lower overall computational costs for processing a set of documents, as discussed in Section 4. However, our findings in Fig. 3 show that increasing the vocabulary size from 50k to larger vocabulary sizes increases the computational costs in all cases. This highlights that the potentially lower fertility of larger vocabulary sizes cannot compensate for the additional costs introduced by the larger vocabulary size.
๋” ํฐ ์–ดํœ˜ ํฌ๊ธฐ๋Š” ์ถ”๊ฐ€์ ์ธ ๊ณ„์‚ฐ ๋น„์šฉ์„ ์ดˆ๋ž˜ํ•˜์ง€๋งŒ, ์ด๋Š” ๋˜ํ•œ ๋” ๋‚ฎ์€ fertility ์ ์ˆ˜๋ฅผ ๊ฐ€์ ธ์™€ ๋ฌธ์„œ ์ง‘ํ•ฉ์„ ์ฒ˜๋ฆฌํ•˜๋Š” ์ „์ฒด ๊ณ„์‚ฐ ๋น„์šฉ์„ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” Section 4์—์„œ ๋…ผ์˜๋œ ๋ฐ”์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Fig. 3์—์„œ ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” ์–ดํœ˜ ํฌ๊ธฐ๋ฅผ 50k์—์„œ ๋” ํฐ ์–ดํœ˜ ํฌ๊ธฐ๋กœ ๋Š˜๋ฆฌ๋ฉด ๋ชจ๋“  ๊ฒฝ์šฐ์— ๊ณ„์‚ฐ ๋น„์šฉ์ด ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Š” ๋” ํฐ ์–ดํœ˜ ํฌ๊ธฐ์˜ ์ž ์žฌ์ ์œผ๋กœ ๋‚ฎ์€ fertility๊ฐ€ ๋” ํฐ ์–ดํœ˜ ํฌ๊ธฐ๋กœ ์ธํ•ด ๋ฐœ์ƒํ•˜๋Š” ์ถ”๊ฐ€ ๋น„์šฉ์„ ์ƒ์‡„ํ•  ์ˆ˜ ์—†์Œ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
์–ดํœ˜ ํฌ๊ธฐ๊ฐ€ ํด์ˆ˜๋ก ๊ณ„์‚ฐ ๋น„์šฉ์ด ์ถ”๊ฐ€๋กœ ๋ฐœ์ƒํ•˜์ง€๋งŒ, ์„น์…˜ 4์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด ์–ดํœ˜ ํฌ๊ธฐ๊ฐ€ ํด์ˆ˜๋ก ์ถœ์‚ฐ์œจ ์ ์ˆ˜๊ฐ€ ๋‚ฎ์•„์ ธ ๋ฌธ์„œ ์ง‘ํ•ฉ์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ๋“œ๋Š” ์ „์ฒด ๊ณ„์‚ฐ ๋น„์šฉ์ด ๋‚ฎ์•„์งˆ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ทธ๋ฆผ 3์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด ์–ดํœ˜ ํฌ๊ธฐ๋ฅผ 50k์—์„œ ๋” ํฐ ์–ดํœ˜ ํฌ๊ธฐ๋กœ ๋Š˜๋ฆฌ๋ฉด ๋ชจ๋“  ๊ฒฝ์šฐ์— ๊ณ„์‚ฐ ๋น„์šฉ์ด ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์–ดํœ˜ ํฌ๊ธฐ๊ฐ€ ํด์ˆ˜๋ก ์ถœ์‚ฐ์œจ์ด ๋‚ฎ์•„์งˆ ์ˆ˜ ์žˆ์ง€๋งŒ ์–ดํœ˜ ํฌ๊ธฐ๊ฐ€ ํด์ˆ˜๋ก ์ถ”๊ฐ€๋˜๋Š” ๋น„์šฉ์„ ๋ณด์ƒํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
* The GatedMLP architecture _slows down_ the model's learning speed, and we observe less stable training with its use.27 Footnote 27: For example, mixed-precision fp16 training can sometimes fail for LLaMA/Mistral models smaller than 100M; hence, we use mixed-precision bf16 instead. Conversely, GPT2 models up to 1B can be trained with fp16. * Removing MLP layers entirely _slows down_ the model's learning speed, whereas adjusting the size of MLP layers (e.g., from \(8d^{2}\) to \(10.5d^{2}\) or down to \(2d^{2}\)) may not have a significant impact.
* GatedMLP ์•„ํ‚คํ…์ฒ˜๋Š” ๋ชจ๋ธ์˜ ํ•™์Šต ์†๋„๋ฅผ _์ €ํ•˜์‹œํ‚ต๋‹ˆ๋‹ค_, ๊ทธ๋ฆฌ๊ณ  ์ด๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ ํ›ˆ๋ จ์ด ๋œ ์•ˆ์ •์ ์ž„์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.27 ๊ฐ์ฃผ 27: ์˜ˆ๋ฅผ ๋“ค์–ด, 100M๋ณด๋‹ค ์ž‘์€ LLaMA/Mistral ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ˜ผํ•ฉ ์ •๋ฐ€๋„ fp16 ํ›ˆ๋ จ์ด ๋•Œ๋•Œ๋กœ ์‹คํŒจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” ๋Œ€์‹  ํ˜ผํ•ฉ ์ •๋ฐ€๋„ bf16์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์—, ์ตœ๋Œ€ 1B์˜ GPT2 ๋ชจ๋ธ์€ fp16์œผ๋กœ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * MLP ๋ ˆ์ด์–ด๋ฅผ ์™„์ „ํžˆ ์ œ๊ฑฐํ•˜๋ฉด ๋ชจ๋ธ์˜ ํ•™์Šต ์†๋„๊ฐ€ _์ €ํ•˜๋˜์ง€๋งŒ_, MLP ๋ ˆ์ด์–ด์˜ ํฌ๊ธฐ๋ฅผ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ(์˜ˆ: \(8d^{2}\)์—์„œ \(10.5d^{2}\) ๋˜๋Š” \(2d^{2}\)๋กœ ์ค„์ด๋Š” ๊ฒƒ)์€ ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
* ๊ฒŒ์ดํŠธ๋“œMLP ์•„ํ‚คํ…์ฒ˜๋Š” ๋ชจ๋ธ์˜ ํ•™์Šต ์†๋„๋ฅผ _๋Š๋ฆฌ๊ฒŒ_ ํ•˜๋ฉฐ, ์ด๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ํ›ˆ๋ จ์ด ๋œ ์•ˆ์ •์ ์œผ๋กœ ์ง„ํ–‰๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.27 ๊ฐ์ฃผ 27: ์˜ˆ๋ฅผ ๋“ค์–ด, 1์–ต ๊ฐœ ๋ฏธ๋งŒ์˜ LLaMA/Mistral ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ˜ผํ•ฉ ์ •๋ฐ€๋„ fp16 ํ›ˆ๋ จ์ด ์‹คํŒจํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํ˜ผํ•ฉ ์ •๋ฐ€๋„ bf16์„ ๋Œ€์‹  ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ GPT2 ๋ชจ๋ธ์€ ์ตœ๋Œ€ 1B๊นŒ์ง€ fp16์œผ๋กœ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * MLP ๋ ˆ์ด์–ด๋ฅผ ์™„์ „ํžˆ ์ œ๊ฑฐํ•˜๋ฉด ๋ชจ๋ธ์˜ ํ•™์Šต ์†๋„๊ฐ€ _๋Š๋ ค์ง€์ง€๋งŒ, MLP ๋ ˆ์ด์–ด์˜ ํฌ๊ธฐ๋ฅผ ์กฐ์ •(์˜ˆ: \(8d^{2}\)์—์„œ \(10.5d^{2}\) ๋˜๋Š” \(2d^{2}\)๋กœ ์ค„์ด๊ธฐ)ํ•˜๋ฉด ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In this new dataset, the preferred translation (\(y_{w}\)) remains the best of the three translation candidates, selected in the same manner as in Section 3.1. However, the dis-preferred translation is intentionally modified to be a noised version of \(y_{w}\). We applied random deletions of words with a probability of 0.15 and word swaps within a range of 1 with a probability of 0.3, following the method suggested by Zeng et al. (2023) for creating manually noised dis-preferred data. This approach produces worse translations that are artificial.
์ด ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์…‹์—์„œ ์„ ํ˜ธ๋˜๋Š” ๋ฒˆ์—ญ (\(y_{w}\))์€ 3๊ฐœ์˜ ๋ฒˆ์—ญ ํ›„๋ณด ์ค‘์—์„œ ์—ฌ์ „ํžˆ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ๊ฒƒ์œผ๋กœ, Section 3.1์—์„œ์™€ ๋™์ผํ•œ ๋ฐฉ์‹์œผ๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋น„์„ ํ˜ธ๋˜๋Š” ๋ฒˆ์—ญ์€ ์˜๋„์ ์œผ๋กœ \(y_{w}\)์˜ ๋…ธ์ด์ฆˆ ๋ฒ„์ „์œผ๋กœ ์ˆ˜์ •๋ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” Zeng et al. (2023)์ด ์ œ์•ˆํ•œ ๋ฐฉ๋ฒ•์„ ๋”ฐ๋ผ, ๋‹จ์–ด๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์‚ญ์ œํ•  ํ™•๋ฅ ์„ 0.15๋กœ ์„ค์ •ํ•˜๊ณ , ๋‹จ์–ด ๊ตํ™˜์„ 1 ๋ฒ”์œ„ ๋‚ด์—์„œ ํ™•๋ฅ  0.3์œผ๋กœ ์ ์šฉํ•˜์—ฌ ์ˆ˜๋™์œผ๋กœ ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•œ ๋น„์„ ํ˜ธ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ๋ฒ•์€ ์ธ์œ„์ ์œผ๋กœ ๋” ๋‚˜์œ ๋ฒˆ์—ญ์„ ๋งŒ๋“ค์–ด๋ƒ…๋‹ˆ๋‹ค.
์ด ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์„ ํ˜ธ ๋ฒˆ์—ญ(\(y_{w}\))์€ ์„น์…˜ 3.1์—์„œ์™€ ๋™์ผํ•œ ๋ฐฉ์‹์œผ๋กœ ์„ ํƒ๋œ ์„ธ ๋ฒˆ์—ญ ํ›„๋ณด ์ค‘์—์„œ ๊ฐ€์žฅ ์ข‹์€ ๋ฒˆ์—ญ์œผ๋กœ ์œ ์ง€๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์„ ํ˜ธํ•˜์ง€ ์•Š๋Š” ๋ฒˆ์—ญ์€ ์˜๋„์ ์œผ๋กœ \(y_{w}\)์˜ ๋…ธ์ด์ฆˆ ๋ฒ„์ „์œผ๋กœ ์ˆ˜์ •๋ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ˆ˜๋™์œผ๋กœ ๋…ธ์ด์ฆˆ๊ฐ€ ์žˆ๋Š” ์„ ํ˜ธํ•˜์ง€ ์•Š๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด Zeng et al. (2023)์—์„œ ์ œ์•ˆํ•œ ๋ฐฉ๋ฒ•์— ๋”ฐ๋ผ ๋‹จ์–ด์˜ ์ž„์˜ ์‚ญ์ œ๋ฅผ 0.15์˜ ํ™•๋ฅ ๋กœ, 1 ๋ฒ”์œ„ ๋‚ด์—์„œ ๋‹จ์–ด ๊ตํ™˜์„ 0.3์˜ ํ™•๋ฅ ๋กœ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ธ์œ„์ ์ธ ๋” ๋‚˜์œ ๋ฒˆ์—ญ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
Question: Is participating in the stress at work survey mandatory? Answer: No, it is voluntary. Question: Will all answers be saved anonymously? Answer: Yes, they will. Question: Who is eligible to take the questionnaire? Answer: Only individuals with work experience, whether part-time or full-time. Question: What happens if someone is not eligible to take the questionnaire? Answer: They will not be able to answer some of the questions. Question: Can you provide a link to all language versions of the questionnaire? Answer: Yes, here is the link.
์งˆ๋ฌธ: ์ง์žฅ ์ŠคํŠธ๋ ˆ์Šค ์„ค๋ฌธ์กฐ์‚ฌ์— ์ฐธ์—ฌํ•˜๋Š” ๊ฒƒ์ด ์˜๋ฌด์ธ๊ฐ€์š”? ๋‹ต๋ณ€: ์•„๋‹ˆ์š”, ์ž๋ฐœ์ ์ž…๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋ชจ๋“  ๋‹ต๋ณ€์ด ์ต๋ช…์œผ๋กœ ์ €์žฅ๋˜๋‚˜์š”? ๋‹ต๋ณ€: ๋„ค, ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋ˆ„๊ฐ€ ์„ค๋ฌธ์กฐ์‚ฌ๋ฅผ ๋ฐ›์„ ์ž๊ฒฉ์ด ์žˆ๋‚˜์š”? ๋‹ต๋ณ€: ํŒŒํŠธํƒ€์ž„์ด๋“  ํ’€ํƒ€์ž„์ด๋“  ๊ทผ๋ฌด ๊ฒฝํ—˜์ด ์žˆ๋Š” ์‚ฌ๋žŒ๋งŒ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์„ค๋ฌธ์กฐ์‚ฌ๋ฅผ ๋ฐ›์„ ์ž๊ฒฉ์ด ์—†๋Š” ์‚ฌ๋žŒ์ด ์žˆ์œผ๋ฉด ์–ด๋–ป๊ฒŒ ๋˜๋‚˜์š”? ๋‹ต๋ณ€: ์ผ๋ถ€ ์งˆ๋ฌธ์— ๋‹ตํ•  ์ˆ˜ ์—†๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์„ค๋ฌธ์กฐ์‚ฌ์˜ ๋ชจ๋“  ์–ธ์–ด ๋ฒ„์ „์— ๋Œ€ํ•œ ๋งํฌ๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋‚˜์š”? ๋‹ต๋ณ€: ๋„ค, ์—ฌ๊ธฐ ๋งํฌ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
์งˆ๋ฌธ: ์ง์žฅ ๋‚ด ์ŠคํŠธ๋ ˆ์Šค ์„ค๋ฌธ์กฐ์‚ฌ ์ฐธ์—ฌ๋Š” ํ•„์ˆ˜์ธ๊ฐ€์š”? ๋‹ต๋ณ€: ์•„๋‹ˆ์š”, ์ž๋ฐœ์ ์ž…๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋ชจ๋“  ๋‹ต๋ณ€์ด ์ต๋ช…์œผ๋กœ ์ €์žฅ๋˜๋‚˜์š”? ๋‹ต๋ณ€: ๋„ค, ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์„ค๋ฌธ์กฐ์‚ฌ์— ์‘ํ•  ์ˆ˜ ์žˆ๋Š” ์ž๊ฒฉ์€ ๋ˆ„๊ตฌ์ธ๊ฐ€์š”? ๋‹ต๋ณ€: ํŒŒํŠธํƒ€์ž„ ๋˜๋Š” ํ’€ํƒ€์ž„ ๋“ฑ ๊ทผ๋ฌด ๊ฒฝ๋ ฅ์ด ์žˆ๋Š” ๊ฐœ์ธ๋งŒ ํ•ด๋‹น๋ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์„ค๋ฌธ์กฐ์‚ฌ์— ์‘ํ•  ์ž๊ฒฉ์ด ์—†๋Š” ๊ฒฝ์šฐ ์–ด๋–ป๊ฒŒ ๋˜๋‚˜์š”? ๋‹ต๋ณ€: ์ผ๋ถ€ ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์„ค๋ฌธ์กฐ์‚ฌ์˜ ๋ชจ๋“  ์–ธ์–ด ๋ฒ„์ „์— ๋Œ€ํ•œ ๋งํฌ๋ฅผ ์ œ๊ณตํ•ด ์ฃผ์‹œ๊ฒ ์Šต๋‹ˆ๊นŒ? ๋‹ต๋ณ€: ๋„ค, ์—ฌ๊ธฐ ๋งํฌ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
The primary function, remove_unwanted_part, starts by splitting the input data into individual sentences. If the first sentence contains delimiters such as "\n\n" (indicating a new paragraph) or ":", the function checks the segment preceding the delimiter for the aforementioned unwanted elements. If these elements are detected, the preceding segment is removed. The entire revised content is then reconstructed and returned. In cases where no modifications are applicable, but we still have the flagged keywords, we remove the paraphrase completely.
์ฃผ์š” ํ•จ์ˆ˜์ธ remove_unwanted_part๋Š” ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐœ๋ณ„ ๋ฌธ์žฅ์œผ๋กœ ๋ถ„ํ• ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋ฌธ์žฅ์ด "\n\n" (์ƒˆ ๋‹จ๋ฝ์„ ๋‚˜ํƒ€๋ƒ„) ๋˜๋Š” ":"์™€ ๊ฐ™์€ ๊ตฌ๋ถ„ ๊ธฐํ˜ธ๋ฅผ ํฌํ•จํ•˜๋Š” ๊ฒฝ์šฐ, ํ•จ์ˆ˜๋Š” ๊ตฌ๋ถ„ ๊ธฐํ˜ธ ์•ž์˜ ์„ธ๊ทธ๋จผํŠธ์—์„œ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๋ถˆํ•„์š”ํ•œ ์š”์†Œ๋ฅผ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์š”์†Œ๊ฐ€ ๊ฐ์ง€๋˜๋ฉด, ์•ž์˜ ์„ธ๊ทธ๋จผํŠธ๊ฐ€ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ˆ˜์ •๋œ ์ „์ฒด ๋‚ด์šฉ์„ ์žฌ๊ตฌ์„ฑํ•˜์—ฌ ๋ฐ˜ํ™˜ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜์ •์ด ์ ์šฉ๋˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ์—๋„ ํ”Œ๋ž˜๊ทธ๋œ ํ‚ค์›Œ๋“œ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ, ์šฐ๋ฆฌ๋Š” ํŒจ๋Ÿฌํ”„๋ ˆ์ด์ฆˆ๋ฅผ ์™„์ „ํžˆ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
๊ธฐ๋ณธ ๊ธฐ๋Šฅ์ธ remove_unwanted_part๋Š” ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐœ๋ณ„ ๋ฌธ์žฅ์œผ๋กœ ๋ถ„ํ• ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋ฌธ์žฅ์— "\n\n"(์ƒˆ ๋‹จ๋ฝ์„ ๋‚˜ํƒ€๋ƒ„) ๋˜๋Š” ":"์™€ ๊ฐ™์€ ๊ตฌ๋ถ„ ๊ธฐํ˜ธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š” ๊ฒฝ์šฐ, ํ•จ์ˆ˜๋Š” ๊ตฌ๋ถ„ ๊ธฐํ˜ธ ์•ž์˜ ์„ธ๊ทธ๋จผํŠธ์—์„œ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์›์น˜ ์•Š๋Š” ์š”์†Œ๊ฐ€ ์žˆ๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์š”์†Œ๊ฐ€ ๊ฐ์ง€๋˜๋ฉด ํ•ด๋‹น ์„ธ๊ทธ๋จผํŠธ๊ฐ€ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ „์ฒด ์ˆ˜์ •๋œ ์ฝ˜ํ…์ธ ๊ฐ€ ์žฌ๊ตฌ์„ฑ๋˜์–ด ๋ฐ˜ํ™˜๋ฉ๋‹ˆ๋‹ค. ์ˆ˜์ •์ด ์ ์šฉ๋˜์ง€ ์•Š์ง€๋งŒ ํ”Œ๋ž˜๊ทธ๊ฐ€ ์ง€์ •๋œ ํ‚ค์›Œ๋“œ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ ์˜์—ญ ์ „์ฒด๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
In this paper, we track the current state of synthetic data research and discuss current best practices and lessons learned. The rest of the paper is organized as follows. Section 2 provides an overview of synthetic data generation techniques and their applications in model training, presenting case studies and empirical evidence. Section 3 discusses the usefulness of synthetic data in evaluation. Section 4 discusses the challenges and limitations of synthetic data, and in Section 5 we outline potential solutions and future research directions.
์ด ๋…ผ๋ฌธ์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์—ฐ๊ตฌ์˜ ํ˜„์žฌ ์ƒํƒœ๋ฅผ ์ถ”์ ํ•˜๊ณ  ํ˜„์žฌ์˜ ๋ชจ๋ฒ” ์‚ฌ๋ก€์™€ ๋ฐฐ์šด ๊ตํ›ˆ์„ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์˜ ๋‚˜๋จธ์ง€ ๋ถ€๋ถ„์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. 2์žฅ์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๊ธฐ์ˆ ๊ณผ ๋ชจ๋ธ ํ›ˆ๋ จ์—์„œ์˜ ์‘์šฉ์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์‚ฌ๋ก€ ์—ฐ๊ตฌ์™€ ์‹ค์ฆ์  ์ฆ๊ฑฐ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. 3์žฅ์—์„œ๋Š” ํ‰๊ฐ€์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์œ ์šฉ์„ฑ์„ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. 4์žฅ์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๋„์ „ ๊ณผ์ œ์™€ ํ•œ๊ณ„๋ฅผ ๋…ผ์˜ํ•˜๊ณ , 5์žฅ์—์„œ๋Š” ์ž ์žฌ์  ํ•ด๊ฒฐ์ฑ…๊ณผ ๋ฏธ๋ž˜ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์„ ๊ฐœ๋žต์ ์œผ๋กœ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์—ฐ๊ตฌ์˜ ํ˜„์žฌ ์ƒํƒœ๋ฅผ ์ถ”์ ํ•˜๊ณ  ํ˜„์žฌ์˜ ๋ชจ๋ฒ” ์‚ฌ๋ก€์™€ ๊ตํ›ˆ์„ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€ ๋…ผ๋ฌธ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์„น์…˜ 2์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๊ธฐ๋ฒ•๊ณผ ๋ชจ๋ธ ํ•™์Šต์— ๋Œ€ํ•œ ์ ์šฉ์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์‚ฌ๋ก€ ์—ฐ๊ตฌ์™€ ๊ฒฝํ—˜์  ์ฆ๊ฑฐ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 3์—์„œ๋Š” ํ‰๊ฐ€์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์œ ์šฉ์„ฑ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 4์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๊ณผ์ œ์™€ ํ•œ๊ณ„์— ๋Œ€ํ•ด ๋…ผ์˜ํ•˜๊ณ , ์„น์…˜ 5์—์„œ๋Š” ์ž ์žฌ์ ์ธ ํ•ด๊ฒฐ์ฑ…๊ณผ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
As soon as the object made contact with their palm, they felt a sense of calm and clarity wash over them. They knew, in that moment, that they had discovered something truly incredible. Over the next few weeks, my character became obsessed with the object. They spent hours staring at it, examining it from every angle. They tried to use it for various tasks, but found that it had an almost mystical power. For instance, they could barely control their appetite when they held it, and any fears they may have had seemed to dissipate into the ether.
๊ทธ ๋ฌผ์ฒด๊ฐ€ ์†๋ฐ”๋‹ฅ์— ๋‹ฟ์ž๋งˆ์ž, ๊ทธ๋“ค์€ ํ‰์˜จํ•จ๊ณผ ๋ช…ํ™•ํ•จ์ด ๋ฐ€๋ ค์˜ค๋Š” ๊ฒƒ์„ ๋Š๊ผˆ๋‹ค. ๊ทธ ์ˆœ๊ฐ„, ๊ทธ๋“ค์€ ์ •๋ง๋กœ ๋†€๋ผ์šด ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Œ์„ ์•Œ์•˜๋‹ค. ๋‹ค์Œ ๋ช‡ ์ฃผ ๋™์•ˆ, ๋‚ด ์บ๋ฆญํ„ฐ๋Š” ๊ทธ ๋ฌผ์ฒด์— ์ง‘์ฐฉํ•˜๊ฒŒ ๋˜์—ˆ๋‹ค. ๊ทธ๋“ค์€ ๋ช‡ ์‹œ๊ฐ„ ๋™์•ˆ ๊ทธ๊ฒƒ์„ ์‘์‹œํ•˜๋ฉฐ ๋ชจ๋“  ๊ฐ๋„์—์„œ ์กฐ์‚ฌํ–ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๊ทธ๊ฒƒ์„ ์‚ฌ์šฉํ•˜๋ ค๊ณ  ์‹œ๋„ํ–ˆ์ง€๋งŒ, ๊ทธ๊ฒƒ์€ ๊ฑฐ์˜ ์‹ ๋น„๋กœ์šด ํž˜์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๊ฒƒ ๊ฐ™์•˜๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ทธ๊ฒƒ์„ ์ฅ๊ณ  ์žˆ์„ ๋•Œ๋Š” ์‹์š•์„ ๊ฑฐ์˜ ์กฐ์ ˆํ•  ์ˆ˜ ์—†์—ˆ๊ณ , ๊ฐ€์ง€๊ณ  ์žˆ๋˜ ๋‘๋ ค์›€์€ ๊ณต์ค‘์œผ๋กœ ์‚ฌ๋ผ์ง€๋Š” ๊ฒƒ ๊ฐ™์•˜๋‹ค.
๋ฌผ์ฒด๊ฐ€ ์†๋ฐ”๋‹ฅ์— ๋‹ฟ๋Š” ์ˆœ๊ฐ„, ๊ทธ๋“ค์€ ์นจ์ฐฉํ•จ๊ณผ ๋ช…๋ฃŒํ•จ์ด ๊ทธ๋“ค์„ ๋ฎ์น˜๋Š” ๊ฒƒ์„ ๋Š๊ผˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ๊ทธ ์ˆœ๊ฐ„ ๊ทธ๋“ค์ด ์ •๋ง ๋†€๋ผ์šด ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ ํ›„ ๋ช‡ ์ฃผ ๋™์•ˆ ๋‚ด ์บ๋ฆญํ„ฐ๋Š” ๊ทธ ๋ฌผ์ฒด์— ์ง‘์ฐฉํ•˜๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ๋ช‡ ์‹œ๊ฐ„ ๋™์•ˆ ๋ฌผ์ฒด๋ฅผ ์‘์‹œํ•˜๊ณ  ๋ชจ๋“  ๊ฐ๋„์—์„œ ๋ฌผ์ฒด๋ฅผ ์‚ดํŽด๋ณด์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๋ฌผ์ฒด๋ฅผ ์‚ฌ์šฉํ•˜๋ ค๊ณ  ํ–ˆ์ง€๋งŒ ๊ฑฐ์˜ ์‹ ๋น„๋กœ์šด ํž˜์ด ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ทธ๋“ค์€ ๋ฌผ์ฒด๋ฅผ ๋“ค์—ˆ์„ ๋•Œ ์‹์š•์„ ๊ฑฐ์˜ ์กฐ์ ˆํ•  ์ˆ˜ ์—†์—ˆ๊ณ , ๋‘๋ ค์›€์ด ์–ด๋””๋ก ๊ฐ€ ์‚ฌ๋ผ์ง€๋Š” ๊ฒƒ ๊ฐ™์•˜์Šต๋‹ˆ๋‹ค.
We initially randomly sampled 500 drug information pages from UpToDate.com, a medical reference website widely used by clinicians. To constrain the scope of questions, we specify in the prompt that the answer must be numerical and in milligrams. To filter out generated questions that did not meet the specified criteria (e.g. ambiguous question, incorrect units, etc.), we perform an additional quality control step, where we ask GPT-4 to verify that the generated question fulfills all criteria. After this step, we have 266 question-answer pairs.
์ฒ˜์Œ์— ์šฐ๋ฆฌ๋Š” ์ž„์˜๋กœ UpToDate.com์—์„œ 500๊ฐœ์˜ ์•ฝ๋ฌผ ์ •๋ณด ํŽ˜์ด์ง€๋ฅผ ์ƒ˜ํ”Œ๋งํ–ˆ์Šต๋‹ˆ๋‹ค. UpToDate.com์€ ์ž„์ƒ์˜๋“ค์ด ๋„๋ฆฌ ์‚ฌ์šฉํ•˜๋Š” ์˜ํ•™ ์ฐธ๊ณ  ์›น์‚ฌ์ดํŠธ์ž…๋‹ˆ๋‹ค. ์งˆ๋ฌธ์˜ ๋ฒ”์œ„๋ฅผ ์ œํ•œํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ํ”„๋กฌํ”„ํŠธ์—์„œ ๋‹ต๋ณ€์ด ๋ฐ˜๋“œ์‹œ ์ˆซ์ž์—ฌ์•ผ ํ•˜๋ฉฐ ๋ฐ€๋ฆฌ๊ทธ๋žจ ๋‹จ์œ„์—ฌ์•ผ ํ•œ๋‹ค๊ณ  ๋ช…์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ง€์ •๋œ ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜์ง€ ์•Š๋Š” ์ƒ์„ฑ๋œ ์งˆ๋ฌธ๋“ค(์˜ˆ: ๋ชจํ˜ธํ•œ ์งˆ๋ฌธ, ์ž˜๋ชป๋œ ๋‹จ์œ„ ๋“ฑ)์„ ๊ฑธ๋Ÿฌ๋‚ด๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์ถ”๊ฐ€์ ์ธ ํ’ˆ์งˆ ๊ด€๋ฆฌ ๋‹จ๊ณ„๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ GPT-4์—๊ฒŒ ์ƒ์„ฑ๋œ ์งˆ๋ฌธ์ด ๋ชจ๋“  ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋Š”์ง€ ํ™•์ธํ•˜๋„๋ก ์š”์ฒญํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๋ฅผ ๊ฑฐ์นœ ํ›„, ์šฐ๋ฆฌ๋Š” 266๊ฐœ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์„ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค.
์ž„์ƒ์˜๊ฐ€ ๋„๋ฆฌ ์‚ฌ์šฉํ•˜๋Š” ์˜ํ•™ ์ฐธ๊ณ  ์›น์‚ฌ์ดํŠธ์ธ UpToDate.com์—์„œ ์•ฝ๋ฌผ ์ •๋ณด ํŽ˜์ด์ง€ 500๊ฐœ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋งํ–ˆ์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ์˜ ๋ฒ”์œ„๋ฅผ ์ œํ•œํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋กฌํ”„ํŠธ์— ๋‹ต๋ณ€์ด ์ˆซ์ž์—ฌ์•ผ ํ•˜๋ฉฐ ๋ฐ€๋ฆฌ๊ทธ๋žจ ๋‹จ์œ„์—ฌ์•ผ ํ•œ๋‹ค๊ณ  ๋ช…์‹œํ•ฉ๋‹ˆ๋‹ค. ์ง€์ •๋œ ๊ธฐ์ค€(์˜ˆ: ๋ชจํ˜ธํ•œ ์งˆ๋ฌธ, ์ž˜๋ชป๋œ ๋‹จ์œ„ ๋“ฑ)์„ ์ถฉ์กฑํ•˜์ง€ ์•Š๋Š” ์ƒ์„ฑ๋œ ์งˆ๋ฌธ์„ ๊ฑธ๋Ÿฌ๋‚ด๊ธฐ ์œ„ํ•ด, ์ƒ์„ฑ๋œ ์งˆ๋ฌธ์ด ๋ชจ๋“  ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋Š”์ง€ GPT-4๊ฐ€ ํ™•์ธํ•˜๋„๋ก ์š”์ฒญํ•˜๋Š” ์ถ”๊ฐ€ ํ’ˆ์งˆ ๊ด€๋ฆฌ ๋‹จ๊ณ„๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๋ฅผ ๊ฑฐ์นœ ํ›„ 266๊ฐœ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์„ ํ™•๋ณดํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Scaling the open-source modes can improve the performance consistently._ By comparing the performance of Vicuna (7B) and Vicuna (13B), Pythia (7B) and Pythia (13B), we can see that the models with larger scales mostly perform better than smaller ones on these evaluation tasks, indicating the effectiveness of scaling up the model size. Across different tasks, scaling model is more beneficial for more complex tasks (_e.g._, symbolic and mathematical reasoning), where the larger models mostly outperform smaller ones in a large margin.
\(\bullet\) ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์˜ ๊ทœ๋ชจ๋ฅผ ํ™•์žฅํ•˜๋ฉด ์„ฑ๋Šฅ์ด ์ผ๊ด€๋˜๊ฒŒ ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Vicuna (7B)์™€ Vicuna (13B), Pythia (7B)์™€ Pythia (13B)์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ด ๋ณด๋ฉด, ๋” ํฐ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์ด ๋Œ€๋ถ€๋ถ„์˜ ํ‰๊ฐ€ ์ž‘์—…์—์„œ ๋” ์ž‘์€ ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์ด ํšจ๊ณผ์ ์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๊ฑธ์ณ, ๋ชจ๋ธ ํ™•์žฅ์€ ๋” ๋ณต์žกํ•œ ์ž‘์—…(์˜ˆ: ๊ธฐํ˜ธ์  ๋ฐ ์ˆ˜ํ•™์  ์ถ”๋ก )์—์„œ ๋” ์œ ์ตํ•˜๋ฉฐ, ์ด ๊ฒฝ์šฐ ๋” ํฐ ๋ชจ๋ธ์ด ๋Œ€๋ถ€๋ถ„์˜ ๊ฒฝ์šฐ ๋” ์ž‘์€ ๋ชจ๋ธ๋ณด๋‹ค ํฐ ์ฐจ์ด๋กœ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค.
\(\bullet\)_์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋“œ๋ฅผ ํ™•์žฅํ•˜๋ฉด ์„ฑ๋Šฅ์„ ์ผ๊ด€๋˜๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค._ Vicuna(7B)์™€ Vicuna(13B), Pythia(7B)์™€ Pythia(13B)์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜๋ฉด, ๋Œ€๋ถ€๋ถ„์˜ ๊ฒฝ์šฐ ๋” ํฐ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์ด ๋” ์ž‘์€ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ๋ณด๋‹ค ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ์ž‘์—…์—์„œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค, ์ด๋Š” ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์˜ ํšจ๊ณผ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ๋ชจ๋ธ์„ ํ™•์žฅํ•˜๋ฉด ๋” ๋ณต์žกํ•œ ์ž‘์—…(_์˜ˆ: ๊ธฐํ˜ธ ๋ฐ ์ˆ˜ํ•™์  ์ถ”๋ก )์— ๋” ์œ ๋ฆฌํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ ๋” ํฐ ๋ชจ๋ธ์ด ๋” ์ž‘์€ ๋ชจ๋ธ๋ณด๋‹ค ํฐ ํญ์œผ๋กœ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค.
**Wellbeing and self-care** Please ensure that you protect your wellbeing and practice self-care while annotating. Take regular breaks. Assess your mental wellbeing regularly. Do not work for prolonged periods of time. Remember the social impact of your work. Reach out to us if you have any issues, and we can connect you with counselling support. You can read more information about the challenges of this work online.36. It is normal to find this work difficult and/or distressing, and to need to talk about your experiences and to be given help.
**์›ฐ๋น™๊ณผ ์ž๊ธฐ ๊ด€๋ฆฌ** ์ฃผ์„ ์ž‘์—…์„ ํ•˜๋Š” ๋™์•ˆ ์ž์‹ ์˜ ์›ฐ๋น™์„ ๋ณดํ˜ธํ•˜๊ณ  ์ž๊ธฐ ๊ด€๋ฆฌ๋ฅผ ์‹ค์ฒœํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ •๊ธฐ์ ์œผ๋กœ ํœด์‹์„ ์ทจํ•˜์„ธ์š”. ์ •์‹ ์  ์›ฐ๋น™์„ ์ •๊ธฐ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜์„ธ์š”. ์žฅ์‹œ๊ฐ„ ์ž‘์—…ํ•˜์ง€ ๋งˆ์„ธ์š”. ์ž์‹ ์˜ ์ž‘์—…์ด ์‚ฌํšŒ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ธฐ์–ตํ•˜์„ธ์š”. ๋ฌธ์ œ๊ฐ€ ์žˆ์œผ๋ฉด ์ €ํฌ์—๊ฒŒ ์—ฐ๋ฝํ•ด ์ฃผ์„ธ์š”. ์ƒ๋‹ด ์ง€์›์„ ์—ฐ๊ฒฐํ•ด ๋“œ๋ฆด ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ์–ด๋ ค์›€์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์ •๋ณด๋ฅผ ์˜จ๋ผ์ธ์—์„œ ์ฝ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์ด ์–ด๋ ต๊ฑฐ๋‚˜ ๊ณ ํ†ต์Šค๋Ÿฝ๊ฒŒ ๋Š๊ปด์ง€๋Š” ๊ฒƒ์€ ์ •์ƒ์ด๋ฉฐ, ์ž์‹ ์˜ ๊ฒฝํ—˜์— ๋Œ€ํ•ด ์ด์•ผ๊ธฐํ•˜๊ณ  ๋„์›€์„ ๋ฐ›๋Š” ๊ฒƒ์ด ํ•„์š”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์›ฐ๋น™ ๋ฐ ์ž๊ธฐ ๊ด€๋ฆฌ** ์ฃผ์„์„ ๋‹ฌ๋Š” ๋™์•ˆ ์ž์‹ ์˜ ์›ฐ๋น™์„ ๋ณดํ˜ธํ•˜๊ณ  ์ž๊ธฐ ๊ด€๋ฆฌ๋ฅผ ์‹ค์ฒœํ•˜์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค. ์ •๊ธฐ์ ์œผ๋กœ ํœด์‹์„ ์ทจํ•˜์„ธ์š”. ์ •์‹  ๊ฑด๊ฐ•์„ ์ •๊ธฐ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜์„ธ์š”. ์žฅ์‹œ๊ฐ„ ์ž‘์—…ํ•˜์ง€ ๋งˆ์„ธ์š”. ์—…๋ฌด๊ฐ€ ์‚ฌํšŒ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ธฐ์–ตํ•˜์„ธ์š”. ๋ฌธ์ œ๊ฐ€ ์žˆ์œผ๋ฉด ๋‹น์‚ฌ์— ์—ฐ๋ฝํ•ด ์ฃผ์‹œ๋ฉด ์ƒ๋‹ด ์ง€์›์„ ์—ฐ๊ฒฐํ•ด ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ์–ด๋ ค์›€์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด๋Š” ์˜จ๋ผ์ธ์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.36. ์ด ์ž‘์—…์ด ์–ด๋ ต๊ณ /๊ฑฐ๋‚˜ ๊ณ ํ†ต์Šค๋Ÿฝ๊ณ , ์ž์‹ ์˜ ๊ฒฝํ—˜์— ๋Œ€ํ•ด ์ด์•ผ๊ธฐํ•˜๊ณ  ๋„์›€์„ ๋ฐ›์•„์•ผ ํ•˜๋Š” ๊ฒƒ์€ ์ •์ƒ์ž…๋‹ˆ๋‹ค.
**Evaluation Dataset.** These datasets are mainly augmented from the training sets of GSM8K and MATH, and we test on the corresponding test sets. The GSM8K dataset sets aside 1,319 entries for testing, while the MATH dataset reserves 5,000 entries as a test set. We extract the final answers by specific templates (e.g., "the answer is") in the predicted reasoning paths following the scripts adopted by Yue et al. (2023); Yu et al. (2023). Subsequently, we calculate the ratio of predicted answers equal to the ground truth answers as the accuracy.
**ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ์…‹.** ์ด ๋ฐ์ดํ„ฐ์…‹๋“ค์€ ์ฃผ๋กœ GSM8K์™€ MATH์˜ ํ›ˆ๋ จ ์„ธํŠธ์—์„œ ์ฆ๊ฐ•๋˜์—ˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ๋Š” ํ•ด๋‹น ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ํ…Œ์ŠคํŠธ๋ฅผ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค. GSM8K ๋ฐ์ดํ„ฐ์…‹์€ ํ…Œ์ŠคํŠธ๋ฅผ ์œ„ํ•ด 1,319๊ฐœ์˜ ํ•ญ๋ชฉ์„ ๋”ฐ๋กœ ๋‘๊ณ , MATH ๋ฐ์ดํ„ฐ์…‹์€ 5,000๊ฐœ์˜ ํ•ญ๋ชฉ์„ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋กœ ์˜ˆ์•ฝํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” Yue et al. (2023); Yu et al. (2023)์ด ์ฑ„ํƒํ•œ ์Šคํฌ๋ฆฝํŠธ๋ฅผ ๋”ฐ๋ผ ์˜ˆ์ธก๋œ ์ถ”๋ก  ๊ฒฝ๋กœ์—์„œ ํŠน์ • ํ…œํ”Œ๋ฆฟ(์˜ˆ: "์ •๋‹ต์€")์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข… ๋‹ต์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ์ดํ›„, ์˜ˆ์ธก๋œ ๋‹ต์ด ์‹ค์ œ ์ •๋‹ต๊ณผ ์ผ์น˜ํ•˜๋Š” ๋น„์œจ์„ ์ •ํ™•๋„๋กœ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
**ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ.** ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ฃผ๋กœ GSM8K ๋ฐ MATH์˜ ํ›ˆ๋ จ ์„ธํŠธ์—์„œ ์ฆ๊ฐ•๋˜๋ฉฐ, ํ•ด๋‹น ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. GSM8K ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 1,319๊ฐœ์˜ ํ•ญ๋ชฉ์„ ํ…Œ์ŠคํŠธ์šฉ์œผ๋กœ ๋”ฐ๋กœ ์„ค์ •ํ•˜๊ณ , MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 5,000๊ฐœ์˜ ํ•ญ๋ชฉ์„ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋กœ ์˜ˆ์•ฝํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ์ธก๋œ ์ถ”๋ก  ๊ฒฝ๋กœ์—์„œ Yue ๋“ฑ(2023); Yu ๋“ฑ(2023)์ด ์ฑ„ํƒํ•œ ์Šคํฌ๋ฆฝํŠธ์— ๋”ฐ๋ผ "์ •๋‹ต์€"๊ณผ ๊ฐ™์€ ํŠน์ • ํ…œํ”Œ๋ฆฟ์œผ๋กœ ์ตœ์ข… ์ •๋‹ต์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ์ •๋‹ต๊ณผ ์ผ์น˜ํ•˜๋Š” ์˜ˆ์ธก๋œ ์ •๋‹ต์˜ ๋น„์œจ์„ ์ •ํ™•๋„๋กœ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
To ensure the quality and diversity of our data, we manually selected 13 data sources from high-quality websites and data resources within the Chinese Internet. These sources include community Q&A forums, encyclopedic sites, content creation platforms, examinations, etc. We also incorporated high-quality Chinese NLP datasets to enrich the diversity of tasks. Specifically, we categorized all data sources into four types: Social Media & forums, World Knowledge, NLP tasks, and Examinations. The data sources and their descriptions are as follows.
๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์ค‘๊ตญ ์ธํ„ฐ๋„ท ๋‚ด์˜ ๊ณ ํ’ˆ์งˆ ์›น์‚ฌ์ดํŠธ์™€ ๋ฐ์ดํ„ฐ ์ž์›์—์„œ 13๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ์ˆ˜๋™์œผ๋กœ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์†Œ์Šค์—๋Š” ์ปค๋ฎค๋‹ˆํ‹ฐ Q&A ํฌ๋Ÿผ, ๋ฐฑ๊ณผ์‚ฌ์ „ ์‚ฌ์ดํŠธ, ์ฝ˜ํ…์ธ  ์ƒ์„ฑ ํ”Œ๋žซํผ, ์‹œํ—˜ ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ž‘์—…์˜ ๋‹ค์–‘์„ฑ์„ ํ’๋ถ€ํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ํ’ˆ์งˆ์˜ ์ค‘๊ตญ์–ด NLP ๋ฐ์ดํ„ฐ์…‹๋„ ํ†ตํ•ฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ๋„ค ๊ฐ€์ง€ ์œ ํ˜•์œผ๋กœ ๋ถ„๋ฅ˜ํ–ˆ์Šต๋‹ˆ๋‹ค: ์†Œ์…œ ๋ฏธ๋””์–ด ๋ฐ ํฌ๋Ÿผ, ์„ธ๊ณ„ ์ง€์‹, NLP ์ž‘์—…, ๊ทธ๋ฆฌ๊ณ  ์‹œํ—˜. ๋ฐ์ดํ„ฐ ์†Œ์Šค์™€ ๊ทธ ์„ค๋ช…์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์ค‘๊ตญ ์ธํ„ฐ๋„ท ๋‚ด ๊ณ ํ’ˆ์งˆ ์›น์‚ฌ์ดํŠธ์™€ ๋ฐ์ดํ„ฐ ๋ฆฌ์†Œ์Šค์—์„œ 13๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ์ˆ˜๋™์œผ๋กœ ์„ ๋ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์†Œ์Šค์—๋Š” ์ปค๋ฎค๋‹ˆํ‹ฐ Q&A ํฌ๋Ÿผ, ๋ฐฑ๊ณผ์‚ฌ์ „ ์‚ฌ์ดํŠธ, ์ฝ˜ํ…์ธ  ์ œ์ž‘ ํ”Œ๋žซํผ, ์‹œํ—˜ ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ณ ํ’ˆ์งˆ์˜ ์ค‘๊ตญ NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ์ž‘์—…์˜ ๋‹ค์–‘์„ฑ์„ ๊ฐ•ํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ์†Œ์…œ ๋ฏธ๋””์–ด ๋ฐ ํฌ๋Ÿผ, ์„ธ๊ณ„ ์ง€์‹, NLP ์ž‘์—…, ์‹œํ—˜์˜ ๋„ค ๊ฐ€์ง€ ์œ ํ˜•์œผ๋กœ ๋ถ„๋ฅ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์†Œ์Šค์™€ ๊ทธ ์„ค๋ช…์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
**EM** provided input on early discussions on learning autoregressive reward functions; wrote the first implementation of DPO and ran the first DPO experiments; trained the large-scale (summarization and dialogue) DPO models used in paper experiments; conducted initial GPT-4 win rate evaluations and set up related infrastructure; recruited participants for, conducted, and analyzed results from the human study; wrote the abstract, introduction, related work, discussion, and most of experiments; and assisted with editing the rest of the paper.
EM์€ ์ž์œจํšŒ๊ท€ ๋ณด์ƒ ํ•จ์ˆ˜ ํ•™์Šต์— ๋Œ€ํ•œ ์ดˆ๊ธฐ ๋…ผ์˜์— ์ž…๋ ฅ์„ ์ œ๊ณตํ–ˆ์œผ๋ฉฐ, DPO์˜ ์ฒซ ๋ฒˆ์งธ ๊ตฌํ˜„์„ ์ž‘์„ฑํ•˜๊ณ  ์ฒซ ๋ฒˆ์งธ DPO ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋…ผ๋ฌธ ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋Œ€๊ทœ๋ชจ DPO ๋ชจ๋ธ(์š”์•ฝ ๋ฐ ๋Œ€ํ™”)์„ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, ์ดˆ๊ธฐ GPT-4 ์Šน๋ฅ  ํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ  ๊ด€๋ จ ์ธํ”„๋ผ๋ฅผ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ธ๊ฐ„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ์ฐธ๊ฐ€์ž๋ฅผ ๋ชจ์ง‘ํ•˜๊ณ  ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜๋ฉฐ ๊ฒฐ๊ณผ๋ฅผ ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ดˆ๋ก, ์„œ๋ก , ๊ด€๋ จ ์—ฐ๊ตฌ, ๋…ผ์˜ ๋ฐ ๋Œ€๋ถ€๋ถ„์˜ ์‹คํ—˜ ๋ถ€๋ถ„์„ ์ž‘์„ฑํ–ˆ์œผ๋ฉฐ, ๋‚˜๋จธ์ง€ ๋…ผ๋ฌธ์˜ ํŽธ์ง‘์„ ๋„์™”์Šต๋‹ˆ๋‹ค.
**EM**์€ ํ•™์Šต ์ž๋™ ํšŒ๊ท€ ๋ณด์ƒ ํ•จ์ˆ˜์— ๋Œ€ํ•œ ์ดˆ๊ธฐ ๋…ผ์˜์— ๋Œ€ํ•œ ์˜๊ฒฌ์„ ์ œ๊ณตํ•˜๊ณ , DPO์˜ ์ฒซ ๋ฒˆ์งธ ๊ตฌํ˜„์„ ์ž‘์„ฑํ•˜๊ณ  ์ฒซ ๋ฒˆ์งธ DPO ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์œผ๋ฉฐ, ๋…ผ๋ฌธ ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋Œ€๊ทœ๋ชจ(์š”์•ฝ ๋ฐ ๋Œ€ํ™”) DPO ๋ชจ๋ธ์„ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, ์ดˆ๊ธฐ GPT-4 ์Šน๋ฅ  ํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ  ๊ด€๋ จ ์ธํ”„๋ผ๋ฅผ ๊ตฌ์ถ•ํ–ˆ์œผ๋ฉฐ, ์ธ๊ฐ„ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ฐธ๊ฐ€์ž๋ฅผ ๋ชจ์ง‘ํ•˜๊ณ  ์ˆ˜ํ–‰ ๋ฐ ๊ฒฐ๊ณผ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ–ˆ์œผ๋ฉฐ, ์ดˆ๋ก, ์„œ๋ก , ๊ด€๋ จ ์—ฐ๊ตฌ, ํ† ๋ก  ๋ฐ ๋Œ€๋ถ€๋ถ„์˜ ์‹คํ—˜์„ ์ž‘์„ฑํ–ˆ์œผ๋ฉฐ ๋‚˜๋จธ์ง€ ๋…ผ๋ฌธ ํŽธ์ง‘์„ ์ง€์›ํ–ˆ์Šต๋‹ˆ๋‹ค.
For solving a complex task, the task planner first needs to clearly understand the task goal and generate a reasonable plan based on the reasoning of LLMs (See Section 6.4.2). Then, the plan executor acts according to the plan in the environment, and the environment will produce feedback for the task planner (See Section 6.4.3). The task planner can further incorporate the feedback obtained from the environment to refine its initial plan and iteratively perform the above process to get better results as the task solution (See Section 6.4.4).
๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด, ์ž‘์—… ๊ณ„ํš์ž๋Š” ๋จผ์ € ์ž‘์—… ๋ชฉํ‘œ๋ฅผ ๋ช…ํ™•ํžˆ ์ดํ•ดํ•˜๊ณ  LLMs์˜ ์ถ”๋ก ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋ฆฌ์ ์ธ ๊ณ„ํš์„ ์ƒ์„ฑํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค (์„น์…˜ 6.4.2 ์ฐธ์กฐ). ๊ทธ๋Ÿฐ ๋‹ค์Œ, ๊ณ„ํš ์‹คํ–‰์ž๋Š” ํ™˜๊ฒฝ์—์„œ ๊ณ„ํš์— ๋”ฐ๋ผ ํ–‰๋™ํ•˜๋ฉฐ, ํ™˜๊ฒฝ์€ ์ž‘์—… ๊ณ„ํš์ž์—๊ฒŒ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค (์„น์…˜ 6.4.3 ์ฐธ์กฐ). ์ž‘์—… ๊ณ„ํš์ž๋Š” ํ™˜๊ฒฝ์—์„œ ์–ป์€ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ฉํ•˜์—ฌ ์ดˆ๊ธฐ ๊ณ„ํš์„ ์ˆ˜์ •ํ•˜๊ณ , ์œ„์˜ ๊ณผ์ •์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•˜์—ฌ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์ž‘์—… ํ•ด๊ฒฐ์ฑ…์œผ๋กœ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค (์„น์…˜ 6.4.4 ์ฐธ์กฐ).
๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ž‘์—… ํ”Œ๋ž˜๋„ˆ๋Š” ๋จผ์ € ์ž‘์—… ๋ชฉํ‘œ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ์ดํ•ดํ•˜๊ณ  LLM์˜ ์ถ”๋ก ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋ฆฌ์ ์ธ ๊ณ„ํš์„ ์ƒ์„ฑํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค(์„น์…˜ 6.4.2 ์ฐธ์กฐ). ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ณ„ํš ์‹คํ–‰์ž๋Š” ํ™˜๊ฒฝ์—์„œ ๊ณ„ํš์— ๋”ฐ๋ผ ํ–‰๋™ํ•˜๊ณ  ํ™˜๊ฒฝ์€ ์ž‘์—… ํ”Œ๋ž˜๋„ˆ์— ํ”ผ๋“œ๋ฐฑ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค(์„น์…˜ 6.4.3 ์ฐธ์กฐ). ์ž‘์—… ํ”Œ๋ž˜๋„ˆ๋Š” ํ™˜๊ฒฝ์—์„œ ์–ป์€ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ฉํ•˜์—ฌ ์ดˆ๊ธฐ ๊ณ„ํš์„ ๊ตฌ์ฒดํ™”ํ•˜๊ณ  ์œ„์˜ ํ”„๋กœ์„ธ์Šค๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ž‘์—… ์†”๋ฃจ์…˜์œผ๋กœ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์„น์…˜ 6.4.4 ์ฐธ์กฐ).
Pythia and deduplication.The Pythia series of models is available in two flavours: one trained on the vanilla version of The Pile, and another trained on a version deduplicated with MinHash. Performance between these two flavours was noted to minimally differ (Biderman et al., 2023); in Figure 9, we find the deduplicated version may be slightly ahead of the non-deduplicated one under our aggregate. The higher end of this improvement is broadly in line with our findings in Table 5. Nevertheless, a difference in our findings and theirs remain.
Pythia์™€ ์ค‘๋ณต ์ œ๊ฑฐ. Pythia ๋ชจ๋ธ ์‹œ๋ฆฌ์ฆˆ๋Š” ๋‘ ๊ฐ€์ง€ ๋ฒ„์ „์œผ๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค: The Pile์˜ ๊ธฐ๋ณธ ๋ฒ„์ „์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๊ณผ MinHash๋กœ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๋ฒ„์ „์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ. ์ด ๋‘ ๋ฒ„์ „ ๊ฐ„์˜ ์„ฑ๋Šฅ ์ฐจ์ด๋Š” ์ตœ์†Œํ•œ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค (Biderman et al., 2023); Figure 9์—์„œ, ์šฐ๋ฆฌ๋Š” ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๋ฒ„์ „์ด ๋น„์ค‘๋ณต ๋ฒ„์ „๋ณด๋‹ค ์•ฝ๊ฐ„ ์•ž์„œ ์žˆ์„ ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์˜ ์ƒํ•œ์„ ์€ Table 5์—์„œ ์šฐ๋ฆฌ์˜ ๋ฐœ๊ฒฌ๊ณผ ๋Œ€์ฒด๋กœ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์šฐ๋ฆฌ์˜ ๋ฐœ๊ฒฌ๊ณผ ๊ทธ๋“ค์˜ ๋ฐœ๊ฒฌ ์‚ฌ์ด์—๋Š” ์ฐจ์ด๊ฐ€ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ”ผํ‹ฐ์•„์™€ ์ค‘๋ณต ์ œ๊ฑฐ.Pythia ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์€ ๋‘ ๊ฐ€์ง€ ๋ง›์œผ๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค: ํ•˜๋‚˜๋Š” ๋ฐ”๋‹๋ผ ๋ฒ„์ „์˜ ๋” ํŒŒ์ผ๋กœ ํ•™์Šต๋˜์—ˆ๊ณ , ๋‹ค๋ฅธ ํ•˜๋‚˜๋Š” MinHash๋กœ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๋ฒ„์ „์œผ๋กœ ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋‘ ๊ฐ€์ง€ ๋ง› ์‚ฌ์ด์˜ ์„ฑ๋Šฅ์€ ๊ฑฐ์˜ ์ฐจ์ด๊ฐ€ ์—†๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Biderman et al., 2023); ๊ทธ๋ฆผ 9์—์„œ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๋ฒ„์ „์ด ์šฐ๋ฆฌ์˜ ์ง‘๊ณ„ ํ•˜์—์„œ ์ค‘๋ณต ์ œ๊ฑฐ๋˜์ง€ ์•Š์€ ๋ฒ„์ „๋ณด๋‹ค ์•ฝ๊ฐ„ ์•ž์„œ ์žˆ๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฐœ์„ ์˜ ์ƒํ•œ์€ ํ‘œ 5์˜ ๊ฒฐ๊ณผ์™€ ๋Œ€์ฒด๋กœ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์™€ ๊ทธ๋“ค์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ ์‚ฌ์ด์—๋Š” ์ฐจ์ด๊ฐ€ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค.
Considering the reasons and their strengths, it seems there is a balanced set of factors that could influence Trump's decision to participate or not. His history of unconventional tactics and concerns about debate arrangements could lead to a "No," while the potential benefits of participating in terms of policy promotion, engaging opponents, and demonstrating leadership suggest a "Yes." The decision will likely depend on the specific circumstances closer to the debate date and Trump's assessment of the strategic value of his participation.
์ด์œ ์™€ ๊ทธ ๊ฐ•์ ์„ ๊ณ ๋ คํ•  ๋•Œ, ํŠธ๋Ÿผํ”„๊ฐ€ ํ† ๋ก ์— ์ฐธ์—ฌํ• ์ง€ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๋ฐ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ๊ท ํ˜• ์žกํžŒ ์š”์ธ๋“ค์ด ์žˆ๋Š” ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๋น„์ „ํ†ต์ ์ธ ์ „์ˆ ์„ ์‚ฌ์šฉํ•ด์˜จ ๊ทธ์˜ ์—ญ์‚ฌ์™€ ํ† ๋ก  ์ค€๋น„์— ๋Œ€ํ•œ ์šฐ๋ ค๋Š” "์•„๋‹ˆ์˜ค"๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์ง€๋งŒ, ์ •์ฑ… ํ™๋ณด, ์ƒ๋Œ€์™€์˜ ๊ต๋ฅ˜, ๋ฆฌ๋”์‹ญ์„ ๋ณด์—ฌ์ฃผ๋Š” ์ธก๋ฉด์—์„œ ์ฐธ์—ฌ์˜ ์ž ์žฌ์  ์ด์ ์€ "์˜ˆ"๋ฅผ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฐ์ •์€ ํ† ๋ก  ๋‚ ์งœ์— ๊ฐ€๊นŒ์›Œ์งˆ์ˆ˜๋ก ํŠน์ • ์ƒํ™ฉ๊ณผ ํŠธ๋Ÿผํ”„๊ฐ€ ๊ทธ์˜ ์ฐธ์—ฌ์˜ ์ „๋žต์  ๊ฐ€์น˜๋ฅผ ์–ด๋–ป๊ฒŒ ํ‰๊ฐ€ํ•˜๋Š”์ง€์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์งˆ ๊ฐ€๋Šฅ์„ฑ์ด ํฝ๋‹ˆ๋‹ค.
์ด์œ ์™€ ๊ทธ ๊ฐ•๋„๋ฅผ ๊ณ ๋ คํ•  ๋•Œ ํŠธ๋Ÿผํ”„๊ฐ€ ์ฐธ์—ฌ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๋ฐ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ๊ท ํ˜• ์žกํžŒ ์š”์ธ๋“ค์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ํŠธ๋Ÿผํ”„์˜ ํŒŒ๊ฒฉ์ ์ธ ์ „์ˆ ์˜ ์—ญ์‚ฌ์™€ ํ† ๋ก  ์ค€๋น„์— ๋Œ€ํ•œ ์šฐ๋ ค๋Š” '์•ˆ๋ผ'๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์ง€๋งŒ, ์ •์ฑ… ํ™๋ณด, ์ƒ๋Œ€๋ฐฉ ์ฐธ์—ฌ, ๋ฆฌ๋”์‹ญ ๋ฐœํœ˜ ๋“ฑ ์ฐธ์—ฌ๋ฅผ ํ†ตํ•ด ์–ป์„ ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ์  ์ด์ ์€ '๊ทธ๋ž˜'๋ฅผ ์•”์‹œํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ์ •์€ ํ† ๋ก  ๋‚ ์งœ์— ๊ฐ€๊นŒ์›Œ์งˆ์ˆ˜๋ก ๊ตฌ์ฒด์ ์ธ ์ƒํ™ฉ๊ณผ ํŠธ๋Ÿผํ”„์˜ ์ฐธ์—ฌ ์ „๋žต์  ๊ฐ€์น˜์— ๋Œ€ํ•œ ํ‰๊ฐ€์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์งˆ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค.
Finally, the last column of Table 3 (No \(\langle\)RET\(\rangle\)) shows the performance of Adapt-LLM when answering questions based solely on its parametric memory. As can be seen, accuracies above 62% are obtained when no context is utilized, providing further evidence that Adapt-LLM effectively discerns between retrieving context and providing direct answers to questions. Additionally, we evaluate the performance of these questions when context is added to the input, revealing significant decreases in accuracy of up to 7 absolute points.
๋งˆ์ง€๋ง‰์œผ๋กœ, Table 3์˜ ๋งˆ์ง€๋ง‰ ์—ด (No \(\langle\)RET\(\rangle\))์€ ์˜ค๋กœ์ง€ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ์งˆ๋ฌธ์— ๋‹ตํ•  ๋•Œ์˜ Adapt-LLM ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ณด์‹œ๋‹ค์‹œํ”ผ, ๋งฅ๋ฝ์ด ์‚ฌ์šฉ๋˜์ง€ ์•Š์„ ๋•Œ๋„ 62% ์ด์ƒ์˜ ์ •ํ™•๋„๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” Adapt-LLM์ด ๋งฅ๋ฝ์„ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒƒ๊ณผ ์งˆ๋ฌธ์— ์ง์ ‘ ๋‹ตํ•˜๋Š” ๊ฒƒ์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ตฌ๋ถ„ํ•œ๋‹ค๋Š” ์ถ”๊ฐ€ ์ฆ๊ฑฐ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ž…๋ ฅ์— ๋งฅ๋ฝ์„ ์ถ”๊ฐ€ํ–ˆ์„ ๋•Œ ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, ์ •ํ™•๋„๊ฐ€ ์ตœ๋Œ€ 7 ์ ˆ๋Œ€ ํฌ์ธํŠธ๊นŒ์ง€ ํฌ๊ฒŒ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
๋งˆ์ง€๋ง‰์œผ๋กœ ํ‘œ 3์˜ ๋งˆ์ง€๋ง‰ ์—ด(No \(\langle\)RET\(\rangle\))์€ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ๋งŒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์งˆ๋ฌธ์— ๋‹ตํ•  ๋•Œ Adapt-LLM์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ณด์‹œ๋‹ค์‹œํ”ผ, ๋ฌธ๋งฅ์„ ํ™œ์šฉํ•˜์ง€ ์•Š์„ ๋•Œ 62% ์ด์ƒ์˜ ์ •ํ™•๋„๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” Adapt-LLM์ด ๋ฌธ๋งฅ ๊ฒ€์ƒ‰๊ณผ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์ง์ ‘์ ์ธ ๋‹ต๋ณ€์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ๋Š” ์ถ”๊ฐ€์ ์ธ ์ฆ๊ฑฐ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž…๋ ฅ์— ๋ฌธ๋งฅ์„ ์ถ”๊ฐ€ํ–ˆ์„ ๋•Œ ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, ์ •ํ™•๋„๊ฐ€ ์ตœ๋Œ€ 7% ์ ˆ๋Œ€ ํฌ์ธํŠธ๊นŒ์ง€ ํฌ๊ฒŒ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
PersonA: Urinalysis can be performed in various healthcare settings, including doctors' offices, urgent care facilities, laboratories, and hospitals. You can provide a urine sample in a specimen cup, and the sample can be analyzed on-siteor sent to a laboratory for testing. PersonU: How much urine is typically needed for the test? PersonA: Usually, only small amounts of urine, ranging from 30 to 60 ml, are required for urinalysis testing. It's a non-invasive and relatively simple test that provides valuable information about your health.
PersonA: ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ์˜์‚ฌ ์‚ฌ๋ฌด์‹ค, ์‘๊ธ‰ ์น˜๋ฃŒ ์‹œ์„ค, ์‹คํ—˜์‹ค ๋ฐ ๋ณ‘์›์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ํ™˜๊ฒฝ์—์„œ ์ˆ˜ํ–‰๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ‘œ๋ณธ ์ปต์— ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ƒ˜ํ”Œ์€ ํ˜„์žฅ์—์„œ ๋ถ„์„๋˜๊ฑฐ๋‚˜ ์‹คํ—˜์‹ค๋กœ ๋ณด๋‚ด์ ธ ๊ฒ€์‚ฌ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. PersonU: ๊ฒ€์‚ฌ์— ์ผ๋ฐ˜์ ์œผ๋กœ ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ์†Œ๋ณ€์ด ํ•„์š”ํ•ฉ๋‹ˆ๊นŒ? PersonA: ๋ณดํ†ต 30์—์„œ 60ml ์ •๋„์˜ ์†Œ๋Ÿ‰์˜ ์†Œ๋ณ€๋งŒ ์†Œ๋ณ€ ๊ฒ€์‚ฌ์— ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋น„์นจ์Šต์ ์ด๊ณ  ๋น„๊ต์  ๊ฐ„๋‹จํ•œ ๊ฒ€์‚ฌ๋กœ, ๊ฑด๊ฐ•์— ๋Œ€ํ•œ ์œ ์šฉํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ์˜์‚ฌ ์‚ฌ๋ฌด์‹ค, ์‘๊ธ‰ ์ง„๋ฃŒ ์‹œ์„ค, ์‹คํ—˜์‹ค, ๋ณ‘์› ๋“ฑ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ํ™˜๊ฒฝ์—์„œ ์‹œํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฒด ์ปต์— ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•˜๋ฉด ์ฆ‰์‹œ ๋ถ„์„ํ•˜๊ฑฐ๋‚˜ ๊ฒ€์‚ฌ๋ฅผ ์œ„ํ•ด ์‹คํ—˜์‹ค๋กœ ๋ณด๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ๋žŒU: ๊ฒ€์‚ฌ์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ์†Œ๋ณ€์ด ํ•„์š”ํ•˜๋‚˜์š”? ์‚ฌ๋žŒA: ์ผ๋ฐ˜์ ์œผ๋กœ ์†Œ๋ณ€ ๊ฒ€์‚ฌ์—๋Š” 30~60ml ์ •๋„์˜ ์†Œ๋Ÿ‰์˜ ์†Œ๋ณ€๋งŒ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋น„์นจ์Šต์ ์ด๊ณ  ๋น„๊ต์  ๊ฐ„๋‹จํ•œ ๊ฒ€์‚ฌ๋กœ ๊ฑด๊ฐ•์— ๋Œ€ํ•œ ๊ท€์ค‘ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
Large language models (LLMs) store vast amounts of factual knowledge in their parameters through large-scale pre-training, and this knowledge can be used to answer various questions such as "where is the world's largest ice sheet located" (Brown et al., 2020; OpenAI, 2023; Chowdhery et al., 2022; Zhang et al., 2022; Touvron et al., 2023, 2023; Gemini Team, 2023). However, this factual knowledge is static, meaning that it can become outdated as the world evolves, or prove insufficient when LLMs are used in specialized or private domains.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLMs)์€ ๋Œ€๊ทœ๋ชจ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ๋งค๊ฐœ๋ณ€์ˆ˜์— ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์‚ฌ์‹ค์  ์ง€์‹์„ ์ €์žฅํ•˜๋ฉฐ, ์ด ์ง€์‹์€ "์„ธ๊ณ„์—์„œ ๊ฐ€์žฅ ํฐ ๋น™์ƒ์ด ์–ด๋””์— ์œ„์น˜ํ•ด ์žˆ๋Š”๊ฐ€"์™€ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค (Brown et al., 2020; OpenAI, 2023; Chowdhery et al., 2022; Zhang et al., 2022; Touvron et al., 2023, 2023; Gemini Team, 2023). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์‚ฌ์‹ค์  ์ง€์‹์€ ์ •์ ์ด๊ธฐ ๋•Œ๋ฌธ์— ์„ธ๊ณ„๊ฐ€ ๋ณ€ํ™”ํ•จ์— ๋”ฐ๋ผ ๊ตฌ์‹์ด ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, LLM์ด ํŠน์ˆ˜ํ•˜๊ฑฐ๋‚˜ ๋น„๊ณต๊ฐœ ๋„๋ฉ”์ธ์—์„œ ์‚ฌ์šฉ๋  ๋•Œ๋Š” ๋ถˆ์ถฉ๋ถ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ๋Œ€๊ทœ๋ชจ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์‚ฌ์‹ค์  ์ง€์‹์„ ํŒŒ๋ผ๋ฏธํ„ฐ์— ์ €์žฅํ•˜๊ณ , ์ด๋Ÿฌํ•œ ์ง€์‹์€ "์„ธ๊ณ„์—์„œ ๊ฐ€์žฅ ํฐ ๋น™์ƒ์€ ์–ด๋””์— ์œ„์น˜ํ•ด ์žˆ๋‚˜์š”?"์™€ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Brown et al., 2020; OpenAI, 2023; Chowdhery et al., 2022; Zhang et al., 2022; Touvron et al., 2023, 2023; Gemini Team, 2023). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์‚ฌ์‹ค์  ์ง€์‹์€ ์ •์ ์ธ ์ƒํƒœ์ด๋ฏ€๋กœ ์„ธ์ƒ์ด ์ง„ํ™”ํ•จ์— ๋”ฐ๋ผ ์˜ค๋ž˜๋œ ์ •๋ณด๊ฐ€ ๋  ์ˆ˜๋„ ์žˆ๊ณ , LLM์ด ํŠน์ˆ˜ ๋˜๋Š” ๋น„๊ณต๊ฐœ ์˜์—ญ์—์„œ ์‚ฌ์šฉ๋  ๋•Œ ๋ถ€์กฑํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
for this task, which contains questions created by crowdworkers based on events from news articles. Zou et al. (2022) introduced Autocast, a benchmark dataset compiled from forecasting competition questions up to 2022. In a competition with a large prize pool, no machine learning system was able to approach the performance of human forecasters on Autocast (Zou et al., 2022). The knowledge cut-offs of LMs have moved past 2022, necessitating more recent data. In this work, we source questions in 2023-2024, enabling us to apply recent LMs.
์ด ์ž‘์—…์„ ์œ„ํ•ด, ๋‰ด์Šค ๊ธฐ์‚ฌ์—์„œ ๋ฐœ์ƒํ•œ ์‚ฌ๊ฑด์„ ๋ฐ”ํƒ•์œผ๋กœ ํฌ๋ผ์šฐ๋“œ์›Œ์ปค๋“ค์ด ์ž‘์„ฑํ•œ ์งˆ๋ฌธ๋“ค์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. Zou ๋“ฑ(2022)์€ 2022๋…„๊นŒ์ง€์˜ ์˜ˆ์ธก ๋Œ€ํšŒ ์งˆ๋ฌธ๋“ค๋กœ ๊ตฌ์„ฑ๋œ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ์…‹์ธ Autocast๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ํฐ ์ƒ๊ธˆ์ด ๊ฑธ๋ฆฐ ๋Œ€ํšŒ์—์„œ, ์–ด๋–ค ๋จธ์‹  ๋Ÿฌ๋‹ ์‹œ์Šคํ…œ๋„ Autocast์—์„œ ์ธ๊ฐ„ ์˜ˆ์ธก๊ฐ€์˜ ์„ฑ๋Šฅ์— ์ ‘๊ทผํ•  ์ˆ˜ ์—†์—ˆ์Šต๋‹ˆ๋‹ค(Zou ๋“ฑ, 2022). LMs์˜ ์ง€์‹ ์ปท์˜คํ”„๋Š” 2022๋…„์„ ๋„˜์–ด์„ฐ์œผ๋ฉฐ, ๋” ์ตœ๊ทผ์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” 2023-2024๋…„์˜ ์งˆ๋ฌธ๋“ค์„ ์†Œ์‹ฑํ•˜์—ฌ ์ตœ์‹  LMs๋ฅผ ์ ์šฉํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ–ˆ์Šต๋‹ˆ๋‹ค.
์ด ์ž‘์—…์—๋Š” ๋‰ด์Šค ๊ธฐ์‚ฌ์˜ ์‚ฌ๊ฑด์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํฌ๋ผ์šฐ๋“œ์›Œ์ปค๊ฐ€ ๋งŒ๋“  ์งˆ๋ฌธ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. Zou et al. (2022)์€ 2022๋…„๊นŒ์ง€์˜ ์˜ˆ์ธก ๊ฒฝ์ง„๋Œ€ํšŒ ์งˆ๋ฌธ์„ ๋ชจ์•„ ๋งŒ๋“  ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ Autocast๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ƒ๋‹นํ•œ ์ƒ๊ธˆ์ด ๊ฑธ๋ฆฐ ๊ฒฝ์Ÿ์—์„œ๋„ ์–ด๋–ค ๋จธ์‹ ๋Ÿฌ๋‹ ์‹œ์Šคํ…œ๋„ Autocast์—์„œ ์ธ๊ฐ„ ์˜ˆ์ธก์ž์˜ ์„ฑ๋Šฅ์— ๊ทผ์ ‘ํ•˜์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค(Zou et al., 2022). LM์˜ ์ง€์‹ ์ปท์˜คํ”„๊ฐ€ 2022๋…„์„ ๋„˜์–ด์„œ๋ฉด์„œ ๋” ์ตœ๊ทผ์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•˜๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” 2023~2024๋…„์— ์งˆ๋ฌธ์„ ์†Œ์‹ฑํ•˜์—ฌ ์ตœ์‹  LM์„ ์ ์šฉํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
Task: Given a dialogue, classify whether the user is satisfied with the service. You should respond with "Satisfied" or "Unsatisfied". Class label: Satisfied Dialogue: - Agent: Thank you for your feedback. We will work to improve our service in the future. - Customer: I am happy with the service you provided. Thank you for your help. Class label: Unsatisfied Dialogue: - Agent: Sorry that we will cancel your order. You will get a refund within 7 business days. - Customer: oh that takes too long. I want you to take quicker action on this.
์ž‘์—…: ์ฃผ์–ด์ง„ ๋Œ€ํ™”๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์‚ฌ์šฉ์ž๊ฐ€ ์„œ๋น„์Šค์— ๋งŒ์กฑํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๋ถ„๋ฅ˜ํ•˜์„ธ์š”. "Satisfied" ๋˜๋Š” "Unsatisfied"๋กœ ์‘๋‹ตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”: Satisfied ๋Œ€ํ™”: - ์ƒ๋‹ด์›: ํ”ผ๋“œ๋ฐฑ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์•ž์œผ๋กœ ๋” ๋‚˜์€ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ๋…ธ๋ ฅํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. - ๊ณ ๊ฐ: ์ œ๊ณตํ•ด์ฃผ์‹  ์„œ๋น„์Šค์— ๋งŒ์กฑํ•ฉ๋‹ˆ๋‹ค. ๋„์™€์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”: Unsatisfied ๋Œ€ํ™”: - ์ƒ๋‹ด์›: ์ฃผ๋ฌธ์„ ์ทจ์†Œํ•ด๋“œ๋ ค ์ฃ„์†กํ•ฉ๋‹ˆ๋‹ค. ํ™˜๋ถˆ์€ ์˜์—…์ผ ๊ธฐ์ค€ 7์ผ ์ด๋‚ด์— ์ฒ˜๋ฆฌ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. - ๊ณ ๊ฐ: ์•„, ๋„ˆ๋ฌด ์˜ค๋ž˜ ๊ฑธ๋ฆฌ๋„ค์š”. ๋” ๋น ๋ฅด๊ฒŒ ์ฒ˜๋ฆฌํ•ด ์ฃผ์„ธ์š”.
๊ณผ์ œ: ๋Œ€ํ™”๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ์‚ฌ์šฉ์ž๊ฐ€ ์„œ๋น„์Šค์— ๋งŒ์กฑํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. "๋งŒ์กฑ" ๋˜๋Š” "๋ถˆ๋งŒ์กฑ"์œผ๋กœ ์‘๋‹ตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”: ๋งŒ์กฑ ๋Œ€ํ™”: - ์ƒ๋‹ด์›: ํ”ผ๋“œ๋ฐฑ์„ ๋ณด๋‚ด์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์•ž์œผ๋กœ ์„œ๋น„์Šค๋ฅผ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋…ธ๋ ฅํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. - ๊ณ ๊ฐ๋‹˜: ์ œ๊ณตํ•ด์ฃผ์‹  ์„œ๋น„์Šค์— ๋งŒ์กฑํ•ฉ๋‹ˆ๋‹ค. ๋„์™€์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”: ๋ถˆ๋งŒ์กฑ ๋Œ€ํ™”: - ์ƒ๋‹ด์›: ์ฃผ๋ฌธ์„ ์ทจ์†Œํ•ด๋“œ๋ ค ์ฃ„์†กํ•ฉ๋‹ˆ๋‹ค. ์˜์—…์ผ ๊ธฐ์ค€ 7์ผ ์ด๋‚ด์— ํ™˜๋ถˆ ์ฒ˜๋ฆฌํ•ด ๋“œ๋ฆฝ๋‹ˆ๋‹ค. - ๊ณ ๊ฐ๋‹˜: ๋„ˆ๋ฌด ์˜ค๋ž˜ ๊ฑธ๋ฆฌ๋„ค์š”. ๋” ๋นจ๋ฆฌ ์ฒ˜๋ฆฌํ•ด ์ฃผ์…จ์œผ๋ฉด ์ข‹๊ฒ ์–ด์š”.
The taxonomy is incompleteThe v0.5 benchmark covers only seven hazard categories. Six other hazard categories were identified in the taxonomy but not included due to feasibility constraints. Further, hazards intersect and it can be hard to separate them; and although we elaborated numerous subcategories and sub subcategories in the taxonomy, we have not covered every hazard. Notably, we have not tested for LM security issues, such as preserving the confidentiality, privacy, integrity, authenticity, and availability of models or data.
์ด ๋ถ„๋ฅ˜ ์ฒด๊ณ„๋Š” ๋ถˆ์™„์ „ํ•ฉ๋‹ˆ๋‹ค. v0.5 ๋ฒค์น˜๋งˆํฌ๋Š” ๋‹จ์ง€ ์ผ๊ณฑ ๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ๋งŒ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ๋ถ„๋ฅ˜ ์ฒด๊ณ„์—์„œ ์—ฌ์„ฏ ๊ฐ€์ง€ ๋‹ค๋ฅธ ์œ„ํ—˜ ๋ฒ”์ฃผ๊ฐ€ ์‹๋ณ„๋˜์—ˆ์ง€๋งŒ, ์‹คํ–‰ ๊ฐ€๋Šฅ์„ฑ์˜ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ํฌํ•จ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์œ„ํ—˜ ์š”์†Œ๋“ค์ด ๊ต์ฐจํ•  ์ˆ˜ ์žˆ์–ด ์ด๋ฅผ ๋ถ„๋ฆฌํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋น„๋ก ์šฐ๋ฆฌ๊ฐ€ ๋ถ„๋ฅ˜ ์ฒด๊ณ„์—์„œ ์ˆ˜๋งŽ์€ ํ•˜์œ„ ๋ฒ”์ฃผ์™€ ํ•˜์œ„ ํ•˜์œ„ ๋ฒ”์ฃผ๋ฅผ ์ƒ์„ธํžˆ ์„ค๋ช…ํ–ˆ์ง€๋งŒ, ๋ชจ๋“  ์œ„ํ—˜์„ ๋‹ค๋ฃจ์ง€๋Š” ์•Š์•˜์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ๋ชจ๋ธ์ด๋‚˜ ๋ฐ์ดํ„ฐ์˜ ๊ธฐ๋ฐ€์„ฑ, ํ”„๋ผ์ด๋ฒ„์‹œ, ๋ฌด๊ฒฐ์„ฑ, ์ง„์œ„์„ฑ, ๊ฐ€์šฉ์„ฑ์„ ์œ ์ง€ํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™์€ LM ๋ณด์•ˆ ๋ฌธ์ œ์— ๋Œ€ํ•ด์„œ๋Š” ํ…Œ์ŠคํŠธํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
๋ถ„๋ฅ˜๊ฐ€ ๋ถˆ์™„์ „ํ•ฉ๋‹ˆ๋‹คv0.5 ๋ฒค์น˜๋งˆํฌ๋Š” 7๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ๋งŒ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋ถ„๋ฅ˜๋ฒ•์—๋Š” 6๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ๊ฐ€ ๋” ์žˆ์ง€๋งŒ ์‹คํ˜„ ๊ฐ€๋Šฅ์„ฑ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ํฌํ•จ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์œ„ํ—˜์€ ์„œ๋กœ ๊ต์ฐจํ•˜๋ฉฐ ๋ถ„๋ฆฌํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋ถ„๋ฅ˜๋ฒ•์—์„œ ์ˆ˜๋งŽ์€ ํ•˜์œ„ ๋ฒ”์ฃผ์™€ ํ•˜์œ„ ํ•˜์œ„ ๋ฒ”์ฃผ๋ฅผ ์ •๊ตํ™”ํ–ˆ์ง€๋งŒ ๋ชจ๋“  ์œ„ํ—˜์„ ๋‹ค๋ฃจ์ง€๋Š” ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ๋ชจ๋ธ ๋˜๋Š” ๋ฐ์ดํ„ฐ์˜ ๊ธฐ๋ฐ€์„ฑ, ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ, ๋ฌด๊ฒฐ์„ฑ, ์ง„์œ„์„ฑ, ๊ฐ€์šฉ์„ฑ ๋ณด์กด๊ณผ ๊ฐ™์€ LM ๋ณด์•ˆ ๋ฌธ์ œ์— ๋Œ€ํ•ด์„œ๋Š” ํ…Œ์ŠคํŠธํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
To answer our questions with prompts of diverse fact composition types, we construct TwoHopFact using well-known named entities in Wikidata [25] and manually selected relations (Appendix A). TwoHopFact consists of 45,595 unique pairs of one-hop and two-hop prompts of 52 fact composition types constructed from the same number of fact triplet pairs \(((e_{1},r_{1},e_{2}),(e_{2},r_{2},e_{3}))\) as in Table 1. Appendix Table 3 shows example two-hop prompts for each fact composition type, and Appendix B provides detailed data statistics.
๋‹ค์–‘ํ•œ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์˜ ํ”„๋กฌํ”„ํŠธ๋กœ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” Wikidata [25]์˜ ์ž˜ ์•Œ๋ ค์ง„ ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ์™€ ์ˆ˜๋™์œผ๋กœ ์„ ํƒํ•œ ๊ด€๊ณ„(Appendix A)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ TwoHopFact๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. TwoHopFact๋Š” 52๊ฐ€์ง€ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์˜ 45,595๊ฐœ์˜ ๊ณ ์œ ํ•œ ํ•œ ํ™‰ ๋ฐ ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” Table 1์— ๋‚˜์™€ ์žˆ๋Š” ๊ฒƒ๊ณผ ๋™์ผํ•œ ์ˆ˜์˜ ์‚ฌ์‹ค ์‚ผ์ค‘ํ•ญ ์Œ \(((e_{1},r_{1},e_{2}),(e_{2},r_{2},e_{3}))\)์—์„œ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. Appendix Table 3์€ ๊ฐ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์— ๋Œ€ํ•œ ์˜ˆ์‹œ ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด์—ฌ์ฃผ๋ฉฐ, Appendix B๋Š” ์ž์„ธํ•œ ๋ฐ์ดํ„ฐ ํ†ต๊ณ„๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๋‹ค์–‘ํ•œ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์˜ ํ”„๋กฌํ”„ํŠธ๋กœ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์œ„ํ‚ค๋ฐ์ดํ„ฐ [25]์˜ ์ž˜ ์•Œ๋ ค์ง„ ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ์™€ ์ˆ˜๋™์œผ๋กœ ์„ ํƒํ•œ ๊ด€๊ณ„(๋ถ€๋ก A)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ TwoHopFact๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. TwoHopFact๋Š” ํ‘œ 1๊ณผ ๋™์ผํ•œ ์ˆ˜์˜ ์‚ฌ์‹ค ์‚ผ์ค‘ ํ•ญ ์Œ \(((e_{1},r_{1},e_{2}),(e_{2},r_{2},e_{3}))\)์œผ๋กœ ๊ตฌ์„ฑ๋œ 52๊ฐœ์˜ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์˜ ํ•œ ํ™‰๊ณผ ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ์˜ ๊ณ ์œ ํ•œ ์Œ 45,595๊ฐœ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋ถ€๋ก ํ‘œ 3์€ ๊ฐ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์— ๋Œ€ํ•œ ์˜ˆ์‹œ ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด์—ฌ์ฃผ๋ฉฐ, ๋ถ€๋ก B๋Š” ์ž์„ธํ•œ ๋ฐ์ดํ„ฐ ํ†ต๊ณ„๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
The longer the context length, the more tokens we can squeeze into the context. The more information the model has access to, the better its response will be. But on the other hand, with very long context, it would be hard for the model to remember everything and efficiently process all the information. Attention-based models are highly inefficient for longer contexts and that is why we should expect more research in different mechanisms that enable processing longer contexts and generally come up with more efficient architectures.
๋ฌธ๋งฅ ๊ธธ์ด๊ฐ€ ๊ธธ์ˆ˜๋ก ๋” ๋งŽ์€ ํ† ํฐ์„ ๋ฌธ๋งฅ์— ๋„ฃ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๋Š” ์ •๋ณด๊ฐ€ ๋งŽ์„์ˆ˜๋ก ์‘๋‹ต์ด ๋” ์ข‹์•„์ง‘๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋ฐ˜๋Œ€๋กœ, ๋ฌธ๋งฅ์ด ๋„ˆ๋ฌด ๊ธธ๋ฉด ๋ชจ๋ธ์ด ๋ชจ๋“  ๊ฒƒ์„ ๊ธฐ์–ตํ•˜๊ณ  ํšจ์œจ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค. Attention ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์€ ๊ธด ๋ฌธ๋งฅ์— ๋Œ€ํ•ด ๋งค์šฐ ๋น„ํšจ์œจ์ ์ด๋ฉฐ, ๊ทธ๋ ‡๊ธฐ ๋•Œ๋ฌธ์— ๋” ๊ธด ๋ฌธ๋งฅ์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ค์–‘ํ•œ ๋ฉ”์ปค๋‹ˆ์ฆ˜๊ณผ ๋” ํšจ์œจ์ ์ธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•œ ์—ฐ๊ตฌ๊ฐ€ ๋” ๋งŽ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
์ปจํ…์ŠคํŠธ ๊ธธ์ด๊ฐ€ ๊ธธ์ˆ˜๋ก ์ปจํ…์ŠคํŠธ์— ๋” ๋งŽ์€ ํ† ํฐ์„ ์••์ถ•ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ๋” ๋งŽ์€ ์ •๋ณด์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ์„์ˆ˜๋ก ์‘๋‹ต์ด ๋” ์ข‹์•„์ง‘๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋ฐ˜๋ฉด์— ์ปจํ…์ŠคํŠธ๊ฐ€ ๋„ˆ๋ฌด ๊ธธ๋ฉด ๋ชจ๋ธ์ด ๋ชจ๋“  ๊ฒƒ์„ ๊ธฐ์–ตํ•˜๊ณ  ๋ชจ๋“  ์ •๋ณด๋ฅผ ํšจ์œจ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์˜ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์€ ๊ธด ์ปจํ…์ŠคํŠธ์— ๋Œ€ํ•ด์„œ๋Š” ๋งค์šฐ ๋น„ํšจ์œจ์ ์ด๊ธฐ ๋•Œ๋ฌธ์— ๋” ๊ธด ์ปจํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ค์–‘ํ•œ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์—ฐ๊ตฌํ•˜๊ณ  ์ผ๋ฐ˜์ ์œผ๋กœ ๋” ํšจ์œจ์ ์ธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ฐœ๋ฐœํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
As illustrated in Figure 15, the generation process of CoT follows a chain structure in the basic CoT prompting approach, where LLMs generate CoTs step by step. Typically, CoT takes the format of natural language text. However, textual CoTs may not work well on complex tasks that require rigorous logic for reasoning. Considering this, some work uses code [506, 507] due to its structured and precise nature. Furthermore, the authors in [508] propose to dynamically select text or code as the format of CoTs to combine their advantages.
Figure 15์—์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, CoT์˜ ์ƒ์„ฑ ๊ณผ์ •์€ ๊ธฐ๋ณธ CoT ํ”„๋กฌํ”„ํŠธ ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ ์ฒด์ธ ๊ตฌ์กฐ๋ฅผ ๋”ฐ๋ฅด๋ฉฐ, LLM์€ ๋‹จ๊ณ„๋ณ„๋กœ CoT๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ CoT๋Š” ์ž์—ฐ์–ด ํ…์ŠคํŠธ ํ˜•์‹์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ CoT๋Š” ์—„๊ฒฉํ•œ ๋…ผ๋ฆฌ๊ฐ€ ํ•„์š”ํ•œ ๋ณต์žกํ•œ ์ž‘์—…์—์„œ๋Š” ์ž˜ ์ž‘๋™ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๊ณ ๋ คํ•˜์—ฌ, ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ตฌ์กฐ์ ์ด๊ณ  ์ •ํ™•ํ•œ ํŠน์„ฑ ๋•Œ๋ฌธ์— ์ฝ”๋“œ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค [506, 507]. ๋˜ํ•œ, [508]์˜ ์ €์ž๋“ค์€ CoT์˜ ํ˜•์‹์œผ๋กœ ํ…์ŠคํŠธ๋‚˜ ์ฝ”๋“œ๋ฅผ ๋™์ ์œผ๋กœ ์„ ํƒํ•˜์—ฌ ์ด๋“ค์˜ ์žฅ์ ์„ ๊ฒฐํ•ฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 15์— ์˜ˆ์‹œ๋œ ๋ฐ”์™€ ๊ฐ™์ด, CoT์˜ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋Š” ๊ธฐ๋ณธ CoT ํ”„๋กฌํ”„ํŠธ ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ LLMs๊ฐ€ ๋‹จ๊ณ„๋ณ„๋กœ CoT๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ฒด์ธ ๊ตฌ์กฐ๋ฅผ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ CoT๋Š” ์ž์—ฐ์–ด ํ…์ŠคํŠธ ํ˜•์‹์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ…์ŠคํŠธ CoT๋Š” ์ถ”๋ก ์„ ์œ„ํ•ด ์—„๊ฒฉํ•œ ๋…ผ๋ฆฌ๊ฐ€ ํ•„์š”ํ•œ ๋ณต์žกํ•œ ์ž‘์—…์—์„œ๋Š” ์ž˜ ์ž‘๋™ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๊ณ ๋ คํ•˜์—ฌ ์ผ๋ถ€ ์ž‘์—…์—์„œ๋Š” ๊ตฌ์กฐํ™”๋˜๊ณ  ์ •ํ™•ํ•œ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ์ฝ”๋“œ [506, 507]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ [508]์˜ ์ €์ž๋“ค์€ ํ…์ŠคํŠธ ๋˜๋Š” ์ฝ”๋“œ๋ฅผ CoT์˜ ํ˜•์‹์œผ๋กœ ๋™์ ์œผ๋กœ ์„ ํƒํ•˜์—ฌ ์žฅ์ ์„ ๊ฒฐํ•ฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
Beyond team structure, the paper emphasizes the importance of employing a "criticize-reflect" step for generating plans, evaluating performance, providing feedback, and re-organizing the team [9]. Their results indicate that agents with a dynamic team structure with rotating leadership provide the best results, with both the lowest time to task completion and the lowest communication cost on average. Ultimately, leadership and dynamic team structures improve the overall team's ability to reason, plan, and perform tasks effectively.
ํŒ€ ๊ตฌ์กฐ๋ฅผ ๋„˜์–ด, ์ด ๋…ผ๋ฌธ์€ ๊ณ„ํš ์ˆ˜๋ฆฝ, ์„ฑ๊ณผ ํ‰๊ฐ€, ํ”ผ๋“œ๋ฐฑ ์ œ๊ณต ๋ฐ ํŒ€ ์žฌ๊ตฌ์„ฑ์„ ์œ„ํ•ด "๋น„ํŒ-๋ฐ˜์„ฑ" ๋‹จ๊ณ„๋ฅผ ์ ์šฉํ•˜๋Š” ๊ฒƒ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค [9]. ๊ทธ๋“ค์˜ ๊ฒฐ๊ณผ๋Š” ๋ฆฌ๋”์‹ญ์ด ์ˆœํ™˜ํ•˜๋Š” ๋™์  ํŒ€ ๊ตฌ์กฐ๋ฅผ ๊ฐ€์ง„ ์—์ด์ „ํŠธ๊ฐ€ ๊ฐ€์žฅ ์ข‹์€ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ํ‰๊ท ์ ์œผ๋กœ ์ž‘์—… ์™„๋ฃŒ ์‹œ๊ฐ„๊ณผ ํ†ต์‹  ๋น„์šฉ์ด ๊ฐ€์žฅ ๋‚ฎ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ถ๊ทน์ ์œผ๋กœ, ๋ฆฌ๋”์‹ญ๊ณผ ๋™์  ํŒ€ ๊ตฌ์กฐ๋Š” ์ „์ฒด ํŒ€์˜ ์ถ”๋ก , ๊ณ„ํš ๋ฐ ์ž‘์—… ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์—์„œ๋Š” ํŒ€ ๊ตฌ์กฐ๋ฅผ ๋„˜์–ด ๊ณ„ํš ์ˆ˜๋ฆฝ, ์„ฑ๊ณผ ํ‰๊ฐ€, ํ”ผ๋“œ๋ฐฑ ์ œ๊ณต, ํŒ€ ์žฌ๊ตฌ์„ฑ์„ ์œ„ํ•ด '๋น„ํŒ-๋ฐ˜์˜' ๋‹จ๊ณ„๋ฅผ ์ ์šฉํ•˜๋Š” ๊ฒƒ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค[9]. ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด ๋ฆฌ๋”์‹ญ์„ ์ˆœํ™˜ํ•˜๋Š” ๋™์  ํŒ€ ๊ตฌ์กฐ๋ฅผ ๊ฐ€์ง„ ์—์ด์ „ํŠธ๊ฐ€ ํ‰๊ท ์ ์œผ๋กœ ์ž‘์—… ์™„๋ฃŒ ์‹œ๊ฐ„๊ณผ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ๋น„์šฉ์ด ๊ฐ€์žฅ ๋‚ฎ์€ ์ตœ์ƒ์˜ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ถ๊ทน์ ์œผ๋กœ ๋ฆฌ๋”์‹ญ๊ณผ ๋™์  ํŒ€ ๊ตฌ์กฐ๋Š” ํŒ€์˜ ์ „๋ฐ˜์ ์ธ ์ถ”๋ก  ๋Šฅ๋ ฅ, ๊ณ„ํš ์ˆ˜๋ฆฝ ๋Šฅ๋ ฅ, ์—…๋ฌด ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
Submitted for Meta internal review on March 14, 2024. We would like to thank Lin Xiao and Yuchen Zhang for many helpful conversations. We would like to extend special thanks to Ian Clark, Gourab De, Anmol Mann, and Max Pfeifer from W&B, as well as Lucca Bertoncini, Liao Hu, Caleb Ho, Apostolos Kokolis, and Shubho Sengupta from Meta FAIR NextSys; Henry Estela, Wil Johnson, Rizwan Hashmi, and Lucas Noah from Meta Cloud Foundation; without their invaluable support, the extensive experiments in this paper would not have been possible.
2024๋…„ 3์›” 14์ผ Meta ๋‚ด๋ถ€ ๊ฒ€ํ† ๋ฅผ ์œ„ํ•ด ์ œ์ถœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ์œ ์ตํ•œ ๋Œ€ํ™”๋ฅผ ๋‚˜๋ˆˆ Lin Xiao์™€ Yuchen Zhang์—๊ฒŒ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. W&B์˜ Ian Clark, Gourab De, Anmol Mann, Max Pfeifer์™€ Meta FAIR NextSys์˜ Lucca Bertoncini, Liao Hu, Caleb Ho, Apostolos Kokolis, Shubho Sengupta์—๊ฒŒ ํŠน๋ณ„ํžˆ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Meta Cloud Foundation์˜ Henry Estela, Wil Johnson, Rizwan Hashmi, Lucas Noah์—๊ฒŒ๋„ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์ด๋“ค์˜ ๊ท€์ค‘ํ•œ ์ง€์› ์—†์ด๋Š” ์ด ๋…ผ๋ฌธ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์ด ๋ถˆ๊ฐ€๋Šฅํ–ˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
2024๋…„ 3์›” 14์ผ์— ๋ฉ”ํƒ€ ๋‚ด๋ถ€ ๊ฒ€ํ† ๋ฅผ ์œ„ํ•ด ์ œ์ถœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ๋„์›€์ด ๋˜๋Š” ๋Œ€ํ™”๋ฅผ ๋‚˜๋ˆˆ Lin Xiao์™€ Yuchen Zhang์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. W&B์˜ Ian Clark, Gourab De, Anmol Mann, Max Pfeifer์™€ Meta FAIR NextSys์˜ Lucca Bertoncini, Liao Hu, Caleb Ho, Apostolos Kokolis, Shubho Sengupta๋ฅผ ํŠน๋ณ„ํžˆ ๊ฐ์‚ฌ๋“œ๋ฆฌ๋ฉฐ, ์ด๋“ค์˜ ๊ท€์ค‘ํ•œ ์ง€์›์ด ์—†์—ˆ๋‹ค๋ฉด ์ด ๋…ผ๋ฌธ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์ด ๋ถˆ๊ฐ€๋Šฅํ–ˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
On dynamic data curating.Our study presents a pipeline to decide on a group of fixed mixture proportions for pretraining. More sophisticated data curating can include dynamic proportions (Albalak et al., 2023) and even a curriculum that changes data domains (Chen et al., 2024). The application of our data mixing laws in continual pretraining (Sec. 5) implies the prospect of extending our findings to these settings. On top of this, we believe that it is promising to incorporate further analysis to pursue a dynamic data mixing law.
๋™์  ๋ฐ์ดํ„ฐ ํ๋ ˆ์ดํŒ…์— ๋Œ€ํ•˜์—ฌ. ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์œ„ํ•œ ๊ณ ์ •๋œ ํ˜ผํ•ฉ ๋น„์œจ ๊ทธ๋ฃน์„ ๊ฒฐ์ •ํ•˜๋Š” ํŒŒ์ดํ”„๋ผ์ธ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋” ์ •๊ตํ•œ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ดํŒ…์€ ๋™์  ๋น„์œจ(Albalak et al., 2023) ๋ฐ ๋ฐ์ดํ„ฐ ๋„๋ฉ”์ธ์„ ๋ณ€๊ฒฝํ•˜๋Š” ์ปค๋ฆฌํ˜๋Ÿผ(Chen et al., 2024)์„ ํฌํ•จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ(Sec. 5)์—์„œ ์šฐ๋ฆฌ์˜ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์„ ์ ์šฉํ•˜๋Š” ๊ฒƒ์€ ์ด๋Ÿฌํ•œ ์„ค์ •์œผ๋กœ ์šฐ๋ฆฌ์˜ ๋ฐœ๊ฒฌ์„ ํ™•์žฅํ•  ๊ฐ€๋Šฅ์„ฑ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด์™€ ๋”๋ถˆ์–ด, ๋™์  ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์„ ์ถ”๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ๋ถ„์„์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ด ์œ ๋งํ•˜๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค.
๋™์  ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜์— ๊ด€ํ•ด์„œ, ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•ด ๊ณ ์ •๋œ ํ˜ผํ•ฉ ๋น„์œจ ๊ทธ๋ฃน์„ ๊ฒฐ์ •ํ•˜๋Š” ํŒŒ์ดํ”„๋ผ์ธ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ณด๋‹ค ์ •๊ตํ•œ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜์—๋Š” ๋™์  ๋น„์œจ(Albalak et al., 2023)๊ณผ ๋ฐ์ดํ„ฐ ๋„๋ฉ”์ธ์„ ๋ณ€๊ฒฝํ•˜๋Š” ์ปค๋ฆฌํ˜๋Ÿผ(Chen et al., 2024)์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต(Sec. 5)์—์„œ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์„ ์ ์šฉํ•˜๋ฉด ์ด๋Ÿฌํ•œ ์„ค์ •์œผ๋กœ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋ฅผ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋Š” ์ „๋ง์ด ์—ด๋ฆฝ๋‹ˆ๋‹ค. ์ด ์™ธ์—๋„ ๋™์  ๋ฐ์ดํ„ฐ ๋ฏน์‹ฑ ๋ฒ•์น™์„ ์ถ”๊ตฌํ•˜๊ธฐ ์œ„ํ•œ ์ถ”๊ฐ€ ๋ถ„์„์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ด ์œ ๋งํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
**Results.** We observed a higher average diagnostic performance for AMIE in Canada than in India (see Figure A.13). However, when comparing scenarios performed at both study locations, we observed that AMIE and PCP performance remained consistent regardless of the study location (see Figure A.14), suggesting that the observed performance variations are likely not due to the patient actor or clinician rater differences, but instead might be attributed to inherent differences in the difficulty levels of scenarios in each location.
**๊ฒฐ๊ณผ.** ์บ๋‚˜๋‹ค์—์„œ AMIE์˜ ํ‰๊ท  ์ง„๋‹จ ์„ฑ๋Šฅ์ด ์ธ๋„๋ณด๋‹ค ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ A.13 ์ฐธ์กฐ). ๊ทธ๋Ÿฌ๋‚˜ ๋‘ ์—ฐ๊ตฌ ์žฅ์†Œ์—์„œ ์ˆ˜ํ–‰๋œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋น„๊ตํ–ˆ์„ ๋•Œ, AMIE์™€ PCP์˜ ์„ฑ๋Šฅ์€ ์—ฐ๊ตฌ ์žฅ์†Œ์— ๊ด€๊ณ„์—†์ด ์ผ๊ด€๋˜๊ฒŒ ์œ ์ง€๋œ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ A.14 ์ฐธ์กฐ). ์ด๋Š” ๊ด€์ฐฐ๋œ ์„ฑ๋Šฅ ์ฐจ์ด๊ฐ€ ํ™˜์ž ๋ฐฐ์šฐ๋‚˜ ์ž„์ƒ์˜ ํ‰๊ฐ€์ž ์ฐจ์ด ๋•Œ๋ฌธ์ด ์•„๋‹ˆ๋ผ, ๊ฐ ์žฅ์†Œ์˜ ์‹œ๋‚˜๋ฆฌ์˜ค ๋‚œ์ด๋„์— ๋‚ด์žฌ๋œ ์ฐจ์ด ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
**๊ฒฐ๊ณผ.** ์บ๋‚˜๋‹ค์—์„œ ์ธ๋„๋ณด๋‹ค AMIE์˜ ํ‰๊ท  ์ง„๋‹จ ์„ฑ๋Šฅ์ด ๋” ๋†’์€ ๊ฒƒ์œผ๋กœ ๊ด€์ฐฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ A.13 ์ฐธ์กฐ). ๊ทธ๋Ÿฌ๋‚˜ ๋‘ ์—ฐ๊ตฌ ๊ธฐ๊ด€์—์„œ ์ˆ˜ํ–‰ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋น„๊ตํ–ˆ์„ ๋•Œ, AMIE์™€ PCP ์„ฑ๋Šฅ์ด ์—ฐ๊ตฌ ๊ธฐ๊ด€์— ๊ด€๊ณ„์—†์ด ์ผ๊ด€๋˜๊ฒŒ ์œ ์ง€๋˜๋Š” ๊ฒƒ์œผ๋กœ ๊ด€์ฐฐ๋˜์–ด ๊ด€์ฐฐ๋œ ์„ฑ๋Šฅ ์ฐจ์ด๋Š” ํ™˜์ž ํ–‰์œ„์ž ๋˜๋Š” ์ž„์ƒ์˜ ํ‰๊ฐ€์ž์˜ ์ฐจ์ด ๋•Œ๋ฌธ์ด ์•„๋‹ˆ๋ผ ๊ฐ ๊ธฐ๊ด€์˜ ์‹œ๋‚˜๋ฆฌ์˜ค ๋‚œ์ด๋„์˜ ๋‚ด์žฌ์  ์ฐจ์ด์— ๊ธฐ์ธํ•œ ๊ฒƒ์ผ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค.
To evaluate financial domain tasks, we compare FinPythia and Fynthia-1B exhibit superior performance on tasks FPB, Headline, and NER while showing comparatively lower results on the FiQA SA task compared with Pythia counterparts. DACP boosts the average task performance by 2.8% for the 1B model and 8.3% for the 6.9B model. These outcomes directly substantiate the impact of domain-adaptive continual pre-training on enhancing in-domain task performance. Furthermore, Pythia-6.9B outperforms OPT-7B, BLOOM-7B, and GPT-J-GB on average.
๊ธˆ์œต ๋„๋ฉ”์ธ ์ž‘์—…์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” FinPythia์™€ Fynthia-1B๊ฐ€ FPB, Headline, NER ์ž‘์—…์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๋ฐ˜๋ฉด, FiQA SA ์ž‘์—…์—์„œ๋Š” Pythia์™€ ๋น„๊ตํ•˜์—ฌ ์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ์€ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์ž„์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. DACP๋Š” 1B ๋ชจ๋ธ์˜ ํ‰๊ท  ์ž‘์—… ์„ฑ๋Šฅ์„ 2.8%, 6.9B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 8.3% ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๋„๋ฉ”์ธ ์ ์‘ ์ง€์† ์‚ฌ์ „ ํ•™์Šต์ด ๋„๋ฉ”์ธ ๋‚ด ์ž‘์—… ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์ง์ ‘์ ์œผ๋กœ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, Pythia-6.9B๋Š” ํ‰๊ท ์ ์œผ๋กœ OPT-7B, BLOOM-7B, GPT-J-GB๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
์žฌ๋ฌด ๋„๋ฉ”์ธ ์ž‘์—…์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด FinPythia์™€ Fynthia-1B๊ฐ€ FPB, ํ—ค๋“œ๋ผ์ธ ๋ฐ NER ์ž‘์—…์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๋ฐ˜๋ฉด, FiQA SA ์ž‘์—…์—์„œ๋Š” Pythia์— ๋น„ํ•ด ์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ์€ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. DACP๋Š” 1B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ‰๊ท  ์ž‘์—… ์„ฑ๋Šฅ์„ 2.8%, 6.9B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 8.3% ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๋„๋ฉ”์ธ ์ ์‘ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์ด ๋„๋ฉ”์ธ ๋‚ด ์ž‘์—… ์„ฑ๋Šฅ ํ–ฅ์ƒ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์ง์ ‘์ ์œผ๋กœ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Pythia-6.9B๋Š” OPT-7B, BLOOM-7B ๋ฐ GPT-J-GB๋ณด๋‹ค ํ‰๊ท ์ ์œผ๋กœ ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค.
Overall, these downstream results validate that training a language model to predict the subtext between the lines of general text data can substantially improve its reasoning capabilities, even on datasets it was not explicitly trained on. The fact that longer rationales consistently lead to better outcomes, and that Quiet-STaR outperforms the constrained pause token approach, supports the notion that Quiet-STaR is successfully teaching the model to leverage its own generated thoughts to reason more thoroughly about the input.
์ „๋ฐ˜์ ์œผ๋กœ, ์ด๋Ÿฌํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๊ฒฐ๊ณผ๋Š” ์ผ๋ฐ˜ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์˜ ํ–‰๊ฐ„(subtext)์„ ์˜ˆ์ธกํ•˜๋„๋ก ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๊ฒƒ์ด, ๋ช…์‹œ์ ์œผ๋กœ ํ›ˆ๋ จ๋˜์ง€ ์•Š์€ ๋ฐ์ดํ„ฐ์…‹์—์„œ๋„ ๋ชจ๋ธ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ์ƒ๋‹นํžˆ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ๋” ๊ธด ์ถ”๋ก  ๊ณผ์ •์ด ์ผ๊ด€๋˜๊ฒŒ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์ด๋Œ์–ด๋‚ด๊ณ , Quiet-STaR๊ฐ€ ์ œํ•œ๋œ pause token ์ ‘๊ทผ๋ฒ•๋ณด๋‹ค ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ์‚ฌ์‹ค์€ Quiet-STaR๊ฐ€ ๋ชจ๋ธ์—๊ฒŒ ์ž…๋ ฅ์— ๋Œ€ํ•ด ๋” ์ฒ ์ €ํ•˜๊ฒŒ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ž์ฒด ์ƒ์„ฑ๋œ ์ƒ๊ฐ์„ ํ™œ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์„ฑ๊ณต์ ์œผ๋กœ ๊ฐ€๋ฅด์น˜๊ณ  ์žˆ์Œ์„ ๋’ท๋ฐ›์นจํ•ฉ๋‹ˆ๋‹ค.
์ „๋ฐ˜์ ์œผ๋กœ ์ด๋Ÿฌํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๊ฒฐ๊ณผ๋Š” ์ผ๋ฐ˜ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์˜ ์ค„ ์‚ฌ์ด์˜ ์•”๋ฌต์ ์ธ ์˜๋ฏธ๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋ฉด ๋ช…์‹œ์ ์œผ๋กœ ํ›ˆ๋ จ๋˜์ง€ ์•Š์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋„ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ๋” ๊ธด ๊ทผ๊ฑฐ๊ฐ€ ์ผ๊ด€๋˜๊ฒŒ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ค๊ณ , Quiet-STaR์ด ์ œํ•œ๋œ ์ผ์‹œ ์ค‘์ง€ ํ† ํฐ ์ ‘๊ทผ ๋ฐฉ์‹๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ์‚ฌ์‹ค์€ Quiet-STaR์ด ๋ชจ๋ธ์ด ์ž์ฒด์ ์œผ๋กœ ์ƒ์„ฑํ•œ ์ƒ๊ฐ์„ ํ™œ์šฉํ•˜์—ฌ ์ž…๋ ฅ์— ๋Œ€ํ•ด ๋” ์ฒ ์ €ํ•˜๊ฒŒ ์ถ”๋ก ํ•˜๋„๋ก ์„ฑ๊ณต์ ์œผ๋กœ ๊ฐ€๋ฅด์น˜๊ณ  ์žˆ์Œ์„ ๋’ท๋ฐ›์นจํ•ฉ๋‹ˆ๋‹ค.