question
stringlengths
101
4.47k
chosen
stringlengths
3
4.25k
rejected
stringlengths
39
9.29k
**Improvement for Random Sampling.** Sampling-based methods sample the token over the whole vocabulary, which may select wrong or irrelevant tokens (_e.g._, "happy" and "Boh" in Figure 10) based on the context. To improve the generation quality, several strategies have been proposed for mitigating or preventing the selection of words with exceedingly low probabilities.
**๋ฌด์ž‘์œ„ ์ƒ˜ํ”Œ๋ง์˜ ๊ฐœ์„ .** ์ƒ˜ํ”Œ๋ง ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์€ ์ „์ฒด ์–ดํœ˜์—์„œ ํ† ํฐ์„ ์ƒ˜ํ”Œ๋งํ•˜๋ฉฐ, ์ด๋Š” ๋ฌธ๋งฅ์— ๋”ฐ๋ผ ์ž˜๋ชป๋˜๊ฑฐ๋‚˜ ๊ด€๋ จ ์—†๋Š” ํ† ํฐ(์˜ˆ: Figure 10์˜ "happy"์™€ "Boh")์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ƒ์„ฑ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด, ๋งค์šฐ ๋‚ฎ์€ ํ™•๋ฅ ์„ ๊ฐ€์ง„ ๋‹จ์–ด์˜ ์„ ํƒ์„ ์™„ํ™”ํ•˜๊ฑฐ๋‚˜ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•œ ์—ฌ๋Ÿฌ ์ „๋žต์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
**๋ฌด์ž‘์œ„ ์ƒ˜ํ”Œ๋ง ๊ฐœ์„ **. ์ƒ˜ํ”Œ๋ง ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์€ ์ „์ฒด ์–ดํœ˜์— ๋Œ€ํ•ด ํ† ํฐ์„ ์ƒ˜ํ”Œ๋งํ•˜๋Š”๋ฐ, ๋ฌธ๋งฅ์— ๋”ฐ๋ผ ์ž˜๋ชป๋œ ํ† ํฐ์ด๋‚˜ ๊ด€๋ จ ์—†๋Š” ํ† ํฐ('๊ทธ๋ฆผ 10'์˜ 'ํ–‰๋ณต' ๋ฐ '๋ณด')์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ƒ์„ฑ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๊ทนํžˆ ๋‚ฎ์€ ํ™•๋ฅ ์„ ๊ฐ€์ง„ ๋‹จ์–ด์˜ ์„ ํƒ์„ ์™„ํ™”ํ•˜๊ฑฐ๋‚˜ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ์ „๋žต์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
Data ModificationData modification is also critical in enhancing retrieval efficiency. This step includes preprocessing techniques like removing irrelevant or redundant information to improve the quality of results and enriching the data with additional information such as metadata to boost the relevance and diversity of the retrieved content (Bevilacqua et al., 2022).
๋ฐ์ดํ„ฐ ์ˆ˜์ • ๋ฐ์ดํ„ฐ ์ˆ˜์ •์€ ๊ฒ€์ƒ‰ ํšจ์œจ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ์—๋„ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—๋Š” ๊ฒฐ๊ณผ์˜ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ด€๋ จ ์—†๊ฑฐ๋‚˜ ์ค‘๋ณต๋œ ์ •๋ณด๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ์ „์ฒ˜๋ฆฌ ๊ธฐ์ˆ ๊ณผ ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ์˜ ๊ด€๋ จ์„ฑ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ์™€ ๊ฐ™์€ ์ถ”๊ฐ€ ์ •๋ณด๋ฅผ ํ†ตํ•ด ๋ฐ์ดํ„ฐ๋ฅผ ํ’๋ถ€ํ•˜๊ฒŒ ํ•˜๋Š” ์ž‘์—…์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค (Bevilacqua et al., 2022).
๋ฐ์ดํ„ฐ ์ˆ˜์ •๋ฐ์ดํ„ฐ ์ˆ˜์ •์€ ๊ฒ€์ƒ‰ ํšจ์œจ์„ฑ์„ ๋†’์ด๋Š” ๋ฐ์—๋„ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—๋Š” ๊ฒฐ๊ณผ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๊ด€๋ จ์„ฑ์ด ์—†๊ฑฐ๋‚˜ ์ค‘๋ณต๋œ ์ •๋ณด๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ์ „์ฒ˜๋ฆฌ ๊ธฐ์ˆ ๊ณผ ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ์˜ ๊ด€๋ จ์„ฑ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ์™€ ๊ฐ™์€ ์ถ”๊ฐ€ ์ •๋ณด๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ๋ณด๊ฐ•ํ•˜๋Š” ๊ธฐ์ˆ ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค(Bevilacqua et al., 2022).
In this section, we briefly show our results for comparing data selecting methods at the 125M scale, where the pre-training dataset is the C4 [41] dataset instead of CC-dedup. We see in Figure A20 that D4 generally outperforms other methods. These initial experiments motivates us to try comparing data selection methods on more heavily filtered web-data (i.e. CC-dedup).
์ด ์„น์…˜์—์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ CC-dedup ๋Œ€์‹  C4 [41] ๋ฐ์ดํ„ฐ์…‹์„ ์‚ฌ์šฉํ•˜์—ฌ 125M ๊ทœ๋ชจ์—์„œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ๋ฅผ ๊ฐ„๋žตํžˆ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Figure A20์—์„œ D4๊ฐ€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๋“ค๋ณด๋‹ค ์šฐ์ˆ˜ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ดˆ๊ธฐ ์‹คํ—˜๋“ค์€ ๋” ์—„๊ฒฉํ•˜๊ฒŒ ํ•„ํ„ฐ๋ง๋œ ์›น ๋ฐ์ดํ„ฐ(์ฆ‰, CC-dedup)์—์„œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ ๋น„๊ตํ•ด๋ณด๋ ค๋Š” ๋™๊ธฐ๋ฅผ ๋ถ€์—ฌํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ CC-dedup๊ฐ€ ์•„๋‹Œ C4 [41] ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ 125M ๊ทœ๋ชจ์—์„œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ๋ฅผ ๊ฐ„๋žตํ•˜๊ฒŒ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ A20์—์„œ D4๊ฐ€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ดˆ๊ธฐ ์‹คํ—˜์„ ํ†ตํ•ด ๋” ๋งŽ์ด ํ•„ํ„ฐ๋ง๋œ ์›น ๋ฐ์ดํ„ฐ(์˜ˆ: CC-dedup)์— ๋Œ€ํ•ด ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ ๋น„๊ตํ•ด๋ณด๊ณ ์ž ํ•˜๋Š” ๋™๊ธฐ๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค.
on various abilities (_e.g.,_ knowledge utilization and human alignment), and thus it is common that they are assessed with multiple evaluation approaches. In addition to benchmark-based evaluation, human-based and model-based approaches have also been widely used to evaluate the advanced abilities of fine-tuned LLMs. Next, we will introduce the two evaluation methods.
๋‹ค์–‘ํ•œ ๋Šฅ๋ ฅ(์˜ˆ: ์ง€์‹ ํ™œ์šฉ ๋ฐ ์ธ๊ฐ„ ์ •๋ ฌ)์— ๋Œ€ํ•ด, ์—ฌ๋Ÿฌ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์œผ๋กœ ํ‰๊ฐ€๋˜๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ๋ฒค์น˜๋งˆํฌ ๊ธฐ๋ฐ˜ ํ‰๊ฐ€ ์™ธ์—๋„, ์ธ๊ฐ„ ๊ธฐ๋ฐ˜ ๋ฐ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ๋ฒ•์ด ๋ฏธ์„ธ ์กฐ์ •๋œ LLM์˜ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, ๋‘ ๊ฐ€์ง€ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.
๋‹ค์–‘ํ•œ ๋Šฅ๋ ฅ(์˜ˆ: ์ง€์‹ ํ™œ์šฉ ๋ฐ ์ธ๊ฐ„ ์ •๋ ฌ)์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•˜๋ฏ€๋กœ ์—ฌ๋Ÿฌ ํ‰๊ฐ€ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ๋ฒค์น˜๋งˆํฌ ๊ธฐ๋ฐ˜ ํ‰๊ฐ€ ์™ธ์—๋„ ์ธ๊ฐ„ ๊ธฐ๋ฐ˜ ๋ฐ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹๋„ ๋ฏธ์„ธ ์กฐ์ •๋œ LLM์˜ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ๋‘ ๊ฐ€์ง€ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.
Here \(A_{1:m,2\ell}\) denotes \(A\) with the first column removed. Note that we are setting \(m=\ell-1\). This is the minimum number of query-answer pairs that we need. If we have more (i.e., \(m\geq\ell\)), then the system is overdetermined. Having the system be overdetermined is a good thing; the extra answers can help us recover the logprobs with greater accuracy.
์—ฌ๊ธฐ์„œ \(A_{1:m,2\ell}\)๋Š” ์ฒซ ๋ฒˆ์งธ ์—ด์ด ์ œ๊ฑฐ๋œ \(A\)๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” \(m=\ell-1\)๋กœ ์„ค์ •ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๊ฒƒ์€ ์šฐ๋ฆฌ๊ฐ€ ํ•„์š”๋กœ ํ•˜๋Š” ์ตœ์†Œํ•œ์˜ ์ฟผ๋ฆฌ-์‘๋‹ต ์Œ์˜ ์ˆ˜์ž…๋‹ˆ๋‹ค. ๋งŒ์•ฝ ๋” ๋งŽ์€ ์Œ์ด ์žˆ๋‹ค๋ฉด (์ฆ‰, \(m\geq\ell\)), ์‹œ์Šคํ…œ์€ ๊ณผ์ž‰ ๊ฒฐ์ •๋ฉ๋‹ˆ๋‹ค. ์‹œ์Šคํ…œ์ด ๊ณผ์ž‰ ๊ฒฐ์ •๋˜๋Š” ๊ฒƒ์€ ์ข‹์€ ์ผ์ž…๋‹ˆ๋‹ค; ์ถ”๊ฐ€์ ์ธ ์‘๋‹ต์€ ๋กœ๊ทธ ํ™•๋ฅ ์„ ๋” ์ •ํ™•ํ•˜๊ฒŒ ๋ณต์›ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(A_{1:m,2\ell}\)๋Š” ์ฒซ ๋ฒˆ์งธ ์—ด์„ ์ œ๊ฑฐํ•œ \(A\)๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. m=\ell-1\). ์ด๊ฒƒ์€ ์šฐ๋ฆฌ๊ฐ€ ํ•„์š”๋กœ ํ•˜๋Š” ์ตœ์†Œ ์ฟผ๋ฆฌ-๋‹ต๋ณ€ ์Œ ์ˆ˜์ž…๋‹ˆ๋‹ค. ๋” ๋งŽ์€ ๊ฒฝ์šฐ(์ฆ‰, \(m\geq\ell\))์—๋Š” ์‹œ์Šคํ…œ์ด ๊ณผ๋Œ€ ๊ฒฐ์ •๋ฉ๋‹ˆ๋‹ค. ์‹œ์Šคํ…œ์ด ๊ณผ๋Œ€ ๊ฒฐ์ •๋˜๋Š” ๊ฒƒ์€ ์ข‹์€ ์ผ์ด๋ฉฐ, ์ถ”๊ฐ€ ๋‹ต๋ณ€์€ ๋” ์ •ํ™•ํ•˜๊ฒŒ ๋กœ๊ทธ ํ™•๋ฅ ์„ ๋ณต๊ตฌํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
While there is speculation and some suggestive comments from McCarthy about the possibility of resignation, there is no concrete evidence or announcement that he will resign by the specified date. His history of fundraising and political influence, as well as the lack of a definitive statement about his future, suggest that a resignation by November 30 is not certain.
McCarthy๊ฐ€ ์‚ฌ์ž„ํ•  ๊ฐ€๋Šฅ์„ฑ์— ๋Œ€ํ•œ ์ถ”์ธก๊ณผ ๋ช‡ ๊ฐ€์ง€ ์•”์‹œ์ ์ธ ๋ฐœ์–ธ์ด ์žˆ๊ธด ํ•˜์ง€๋งŒ, ๊ทธ๊ฐ€ ์ง€์ •๋œ ๋‚ ์งœ๊นŒ์ง€ ์‚ฌ์ž„ํ•  ๊ฒƒ์ด๋ผ๋Š” ๊ตฌ์ฒด์ ์ธ ์ฆ๊ฑฐ๋‚˜ ๋ฐœํ‘œ๋Š” ์—†์Šต๋‹ˆ๋‹ค. ๊ทธ์˜ ๋ชจ๊ธˆ ํ™œ๋™๊ณผ ์ •์น˜์  ์˜ํ–ฅ๋ ฅ, ๊ทธ๋ฆฌ๊ณ  ๊ทธ์˜ ๋ฏธ๋ž˜์— ๋Œ€ํ•œ ๋ช…ํ™•ํ•œ ๋ฐœ์–ธ์ด ์—†๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•  ๋•Œ, 11์›” 30์ผ๊นŒ์ง€ ์‚ฌ์ž„ํ•  ๊ฒƒ์ด๋ผ๋Š” ๊ฒƒ์€ ํ™•์‹คํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
์‚ฌํ‡ด ๊ฐ€๋Šฅ์„ฑ์— ๋Œ€ํ•œ ์ถ”์ธก๊ณผ ๋งค์นด์‹œ์˜ ์•”์‹œ์ ์ธ ๋ฐœ์–ธ์ด ์žˆ์ง€๋งŒ, ๊ทธ๊ฐ€ ์ง€์ •๋œ ๋‚ ์งœ๊นŒ์ง€ ์‚ฌ์ž„ํ•  ๊ฒƒ์ด๋ผ๋Š” ๊ตฌ์ฒด์ ์ธ ์ฆ๊ฑฐ๋‚˜ ๋ฐœํ‘œ๋Š” ์—†์Šต๋‹ˆ๋‹ค. ๊ทธ๋™์•ˆ์˜ ๋ชจ๊ธˆ ํ™œ๋™๊ณผ ์ •์น˜์  ์˜ํ–ฅ๋ ฅ, ๊ทธ๋ฆฌ๊ณ  ํ–ฅํ›„ ๊ฑฐ์ทจ์— ๋Œ€ํ•œ ๋ช…ํ™•ํ•œ ๋ฐœ์–ธ์ด ์—†๋‹ค๋Š” ์ ์—์„œ 11์›” 30์ผ๊นŒ์ง€ ์‚ฌ์ž„ํ•  ๊ฒƒ์ด ํ™•์‹คํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
To assess the quality of their retrievals, we first take 24 unresolved forecasting questions. Next, we prompt GPT-4-1106-Preview to generate two search queries for each of these 24 questions, similar to the first stage of our retrieval system (Section 4). We use these queries to search for articles via all 5 APIs, restricting the retrieval range to the last 24 hours.
์ด๋“ค์˜ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ๋จผ์ € ํ•ด๊ฒฐ๋˜์ง€ ์•Š์€ ์˜ˆ์ธก ์งˆ๋ฌธ 24๊ฐœ๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, GPT-4-1106-Preview๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด 24๊ฐœ์˜ ์งˆ๋ฌธ ๊ฐ๊ฐ์— ๋Œ€ํ•ด ๋‘ ๊ฐœ์˜ ๊ฒ€์ƒ‰ ์ฟผ๋ฆฌ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์šฐ๋ฆฌ์˜ ๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ์˜ ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์™€ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค(์„น์…˜ 4). ์ด๋Ÿฌํ•œ ์ฟผ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋“  5๊ฐœ์˜ API๋ฅผ ํ†ตํ•ด ๊ธฐ์‚ฌ๋ฅผ ๊ฒ€์ƒ‰ํ•˜๋ฉฐ, ๊ฒ€์ƒ‰ ๋ฒ”์œ„๋Š” ์ง€๋‚œ 24์‹œ๊ฐ„์œผ๋กœ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค.
๊ฒ€์ƒ‰์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋จผ์ € ํ•ด๊ฒฐ๋˜์ง€ ์•Š์€ 24๊ฐœ์˜ ์˜ˆ์ธก ์งˆ๋ฌธ์„ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ์šฐ๋ฆฌ๋Š” ๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ์˜ ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„(์„น์…˜ 4)์™€ ์œ ์‚ฌํ•˜๊ฒŒ ์ด 24๊ฐœ์˜ ์งˆ๋ฌธ ๊ฐ๊ฐ์— ๋Œ€ํ•ด ๋‘ ๊ฐœ์˜ ๊ฒ€์ƒ‰ ์ฟผ๋ฆฌ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก GPT-4-1106-Preview์— ํ”„๋กฌํ”„ํŠธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฟผ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ง€๋‚œ 24์‹œ๊ฐ„ ๋ฒ”์œ„๋ฅผ ์ œํ•œํ•˜์—ฌ 5๊ฐœ์˜ API๋ฅผ ๋ชจ๋‘ ํ†ตํ•ด ๊ธฐ์‚ฌ๋ฅผ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค.
**Output**[[**Retriee**]=No[The stars begin to twinkle bright the trees, scattered throughout the night; [[**Retriee**]=No[The stars begin to twinkle bright the trees, scattered throughout the night [[**Retriee**]=No[ They flicker like a silent dance, A peaceful moment in its trace. [[**Retriee**]=No[] They flicker like a silent dance, A peaceful moment in its trace.
**Output**[[**Retriee**]=No[๋ณ„๋“ค์ด ๋‚˜๋ฌด ์‚ฌ์ด์—์„œ ๋ฐ˜์ง์ด๊ธฐ ์‹œ์ž‘ํ•˜๊ณ , ๋ฐคํ•˜๋Š˜์— ํฉ์–ด์ ธ ์žˆ๋‹ค; [[**Retriee**]=No[๋ณ„๋“ค์ด ๋‚˜๋ฌด ์‚ฌ์ด์—์„œ ๋ฐ˜์ง์ด๊ธฐ ์‹œ์ž‘ํ•˜๊ณ , ๋ฐคํ•˜๋Š˜์— ํฉ์–ด์ ธ ์žˆ๋‹ค [[**Retriee**]=No[๊ทธ๋“ค์€ ์กฐ์šฉํ•œ ์ถค์ฒ˜๋Ÿผ ๊นœ๋นก์ด๋ฉฐ, ํ‰ํ™”๋กœ์šด ์ˆœ๊ฐ„์„ ๋‚จ๊ธด๋‹ค. [[**Retriee**]=No[] ๊ทธ๋“ค์€ ์กฐ์šฉํ•œ ์ถค์ฒ˜๋Ÿผ ๊นœ๋นก์ด๋ฉฐ, ํ‰ํ™”๋กœ์šด ์ˆœ๊ฐ„์„ ๋‚จ๊ธด๋‹ค.
**์ถœ๋ ฅ**[[**Retriee**]=No[๋ณ„๋“ค์ด ๋ฐค์ƒˆ ํฉ์–ด์ ธ ์žˆ๋Š” ๋‚˜๋ฌด๋“ค์„ ๋ฐ๊ฒŒ ๋น›๋‚ด๊ธฐ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค; [[**Retriee**]=No[๋ณ„๋“ค์ด ๋ฐค์ƒˆ ํฉ์–ด์ ธ ์žˆ๋Š” ๋‚˜๋ฌด๋“ค์„ ๋ฐ๊ฒŒ ๋น›๋‚ด๊ธฐ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค [[**Retriee**]=No[ ๊ทธ๋“ค์€ ์กฐ์šฉํ•œ ์ถค์ฒ˜๋Ÿผ ๊นœ๋ฐ•์ž…๋‹ˆ๋‹ค, ๊ทธ ํ”์ ์—์„œ ํ‰ํ™”๋กœ์šด ์ˆœ๊ฐ„. [[**Retriee**]=No[] ๊ทธ๋“ค์€ ์กฐ์šฉํ•œ ์ถค์ฒ˜๋Ÿผ ๊นœ๋ฐ•์ž…๋‹ˆ๋‹ค, ๊ทธ ํ”์ ์—์„œ ํ‰ํ™”๋กœ์šด ์ˆœ๊ฐ„.
Using simulated dialogues allowed us to address the limited availability of high-quality, labelled real-world conversation data and improved the model's generalization and adaptability to diverse medical contexts. By leveraging this self-play paradigm, AMIE could continuously learn and refine its conversational and diagnostic capabilities during patient interactions.
์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ๋Œ€ํ™”๋ฅผ ์‚ฌ์šฉํ•จ์œผ๋กœ์จ ๊ณ ํ’ˆ์งˆ์˜ ๋ผ๋ฒจ๋ง๋œ ์‹ค์ œ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ์˜ ์ œํ•œ๋œ ๊ฐ€์šฉ์„ฑ์„ ํ•ด๊ฒฐํ•˜๊ณ  ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ๋ฐ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ์ƒํ™ฉ์— ๋Œ€ํ•œ ์ ์‘์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์…€ํ”„ ํ”Œ๋ ˆ์ด ํŒจ๋Ÿฌ๋‹ค์ž„์„ ํ™œ์šฉํ•˜์—ฌ AMIE๋Š” ํ™˜์ž์™€์˜ ์ƒํ˜ธ์ž‘์šฉ ์ค‘์— ์ง€์†์ ์œผ๋กœ ํ•™์Šตํ•˜๊ณ  ๋Œ€ํ™” ๋ฐ ์ง„๋‹จ ๋Šฅ๋ ฅ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ๋Œ€ํ™”๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ณ ํ’ˆ์งˆ์˜ ๋ผ๋ฒจ์ด ๋ถ™์€ ์‹ค์ œ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ์˜ ์ œํ•œ๋œ ๊ฐ€์šฉ์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์—ˆ๊ณ , ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ์ƒํ™ฉ์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ๋ฐ ์ ์‘๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์…€ํ”„ ํ”Œ๋ ˆ์ด ํŒจ๋Ÿฌ๋‹ค์ž„์„ ํ™œ์šฉํ•จ์œผ๋กœ์จ AMIE๋Š” ํ™˜์ž์™€์˜ ์ƒํ˜ธ ์ž‘์šฉ ์ค‘์— ๋Œ€ํ™” ๋ฐ ์ง„๋‹จ ๊ธฐ๋Šฅ์„ ์ง€์†์ ์œผ๋กœ ํ•™์Šตํ•˜๊ณ  ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
To apply this to our setting, note that (1) all constraints, even the initial \(\alpha_{i}\leq\mathsf{logit}_{i}\leq\beta_{i}\), are of the required form; (2) the graph has no negative cycles because the true logits give a feasible solution. (3) we can get the lower bounds by applying the same procedure to the graph induced by inequalities on \(-\mathsf{logit}_{i}\).
์šฐ๋ฆฌ์˜ ์„ค์ •์— ์ด๋ฅผ ์ ์šฉํ•˜๊ธฐ ์œ„ํ•ด, ๋‹ค์Œ์„ ์ฃผ๋ชฉํ•˜์‹ญ์‹œ์˜ค: (1) ์ดˆ๊ธฐ \(\alpha_{i}\leq\mathsf{logit}_{i}\leq\beta_{i}\)๋ฅผ ํฌํ•จํ•œ ๋ชจ๋“  ์ œ์•ฝ ์กฐ๊ฑด์€ ํ•„์š”ํ•œ ํ˜•ํƒœ๋ฅผ ๊ฐ–์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค; (2) ๊ทธ๋ž˜ํ”„์—๋Š” ์Œ์˜ ์‚ฌ์ดํด์ด ์—†๋Š”๋ฐ, ์ด๋Š” ์‹ค์ œ ๋กœ์ง“์ด ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ํ•ด๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. (3) ์šฐ๋ฆฌ๋Š” \(-\mathsf{logit}_{i}\์— ๋Œ€ํ•œ ๋ถ€๋“ฑ์‹์œผ๋กœ ์œ ๋„๋œ ๊ทธ๋ž˜ํ”„์— ๋™์ผํ•œ ์ ˆ์ฐจ๋ฅผ ์ ์šฉํ•˜์—ฌ ํ•˜ํ•œ๊ฐ’์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด๋ฅผ ์šฐ๋ฆฌ์˜ ์„ค์ •์— ์ ์šฉํ•˜๋ ค๋ฉด (1) ๋ชจ๋“  ์ œ์•ฝ ์กฐ๊ฑด, ์‹ฌ์ง€์–ด ์ดˆ๊ธฐ \(\alpha_{i}\leq\mathsf{logit}_{i}\leq\beta_{i}\),๊ฐ€ ํ•„์š”ํ•œ ํ˜•ํƒœ๋ฅผ ๊ฐ–๊ณ , (2) ์‹ค์ œ ๋กœ์ง“์ด ์‹คํ˜„ ๊ฐ€๋Šฅํ•œ ํ•ด๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ทธ๋ž˜ํ”„์— ์Œ์˜ ์‚ฌ์ดํด์ด ์—†์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (3) ๋ถ€๋“ฑ์‹์— ์˜ํ•ด ์œ ๋„๋œ ๊ทธ๋ž˜ํ”„์— ๋™์ผํ•œ ์ ˆ์ฐจ๋ฅผ ์ ์šฉํ•˜์—ฌ ํ•˜ํ•œ์„ ๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Performance on different skills.** Figure 6 compares _WizardLM_ and ChatGPT's skill levels on _Evol-Instruct_ testset. The result indicates that _WizardLM_ achieves 78% of ChatGPT's performance on average, with almost more than 90% capacity on 17 skills. However, _WizardLM_ struggles with code, math, and reasoning scenarios, revealing a noticeable gap with ChatGPT.
**๋‹ค์–‘ํ•œ ๊ธฐ์ˆ ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ.** Figure 6์€ _Evol-Instruct_ ํ…Œ์ŠคํŠธ์…‹์—์„œ _WizardLM_๊ณผ ChatGPT์˜ ๊ธฐ์ˆ  ์ˆ˜์ค€์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด _WizardLM_์€ ํ‰๊ท ์ ์œผ๋กœ ChatGPT ์„ฑ๋Šฅ์˜ 78%๋ฅผ ๋‹ฌ์„ฑํ•˜๋ฉฐ, 17๊ฐœ์˜ ๊ธฐ์ˆ ์—์„œ ๊ฑฐ์˜ 90% ์ด์ƒ์˜ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ _WizardLM_์€ ์ฝ”๋“œ, ์ˆ˜ํ•™, ์ถ”๋ก  ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์–ด๋ ค์›€์„ ๊ฒช์œผ๋ฉฐ ChatGPT์™€ ๋ˆˆ์— ๋„๋Š” ๊ฒฉ์ฐจ๋ฅผ ๋“œ๋Ÿฌ๋ƒ…๋‹ˆ๋‹ค.
**๋‹ค์–‘ํ•œ ๊ธฐ์ˆ ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ.** ๊ทธ๋ฆผ 6์€ _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ _WizardLM_๊ณผ ChatGPT์˜ ๊ธฐ์ˆ  ์ˆ˜์ค€์„ ๋น„๊ตํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, _WizardLM_์€ 17๊ฐœ ๊ธฐ์ˆ ์—์„œ ํ‰๊ท  90% ์ด์ƒ์˜ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋ฉฐ ChatGPT์˜ ์„ฑ๋Šฅ์˜ 78%๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ _WizardLM_์€ ์ฝ”๋“œ, ์ˆ˜ํ•™ ๋ฐ ์ถ”๋ก  ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์œผ๋ฉฐ ChatGPT์™€ ๋ˆˆ์— ๋„๋Š” ๊ฒฉ์ฐจ๋ฅผ ๋“œ๋Ÿฌ๋ƒ…๋‹ˆ๋‹ค.
When running the full layer-stealing attack, we confirmed that our extracted weights are nearly identical to the actual weights, with error \(<7\cdot 10^{-4}\), up to an \(h\times h\) matrix product as discussed previously. Table 4 reports the RMS between our extracted weight matrix and the actual model weights, after "aligning" the two by an \(h\times h\) transform.
์ „์ฒด ๋ ˆ์ด์–ด ์Šคํ‹ธ๋ง ๊ณต๊ฒฉ์„ ์‹คํ–‰ํ•  ๋•Œ, ์šฐ๋ฆฌ๋Š” ์ถ”์ถœ๋œ ๊ฐ€์ค‘์น˜๊ฐ€ ์‹ค์ œ ๊ฐ€์ค‘์น˜์™€ ๊ฑฐ์˜ ๋™์ผํ•˜๋‹ค๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์œผ๋ฉฐ, ์˜ค์ฐจ๋Š” \(<7\cdot 10^{-4}\)๋กœ, ์ด์ „์— ๋…ผ์˜๋œ ๋Œ€๋กœ \(h\times h\) ํ–‰๋ ฌ ๊ณฑ์…ˆ๊นŒ์ง€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ํ‘œ 4๋Š” \(h\times h\) ๋ณ€ํ™˜์„ ํ†ตํ•ด ๋‘ ๊ฐ€์ค‘์น˜๋ฅผ "์ •๋ ฌ"ํ•œ ํ›„, ์ถ”์ถœ๋œ ๊ฐ€์ค‘์น˜ ํ–‰๋ ฌ๊ณผ ์‹ค์ œ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜ ๊ฐ„์˜ RMS๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.
์ „์ฒด ๋ ˆ์ด์–ด ์Šคํ‹ธ๋ง ๊ณต๊ฒฉ์„ ์‹คํ–‰ํ•  ๋•Œ, ์•ž์„œ ์„ค๋ช…ํ•œ ๋Œ€๋กœ \(h\times h\) ํ–‰๋ ฌ ๊ณฑ์…ˆ๊นŒ์ง€ ์˜ค์ฐจ \(<7\cdot 10^{-4}\), ์ถ”์ถœํ•œ ๊ฐ€์ค‘์น˜๊ฐ€ ์‹ค์ œ ๊ฐ€์ค‘์น˜์™€ ๊ฑฐ์˜ ๋™์ผํ•˜๋‹ค๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 4๋Š” ๋‘ ๊ฐœ๋ฅผ \(h\times h\) ๋ณ€ํ™˜์œผ๋กœ "์ •๋ ฌ"ํ•œ ํ›„ ์ถ”์ถœํ•œ ๊ฐ€์ค‘์น˜ ํ–‰๋ ฌ๊ณผ ์‹ค์ œ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜ ๊ฐ„์˜ RMS๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.
Finally, we chose to apply MinHash before exact deduplication, as it is easier to scale: approximate deduplication acts as a pruning phase, enabling us to scale deduplication further. Finally, we choose the common option of cutting spans, as dropping resulted in even more stringent rejection rates which would have compromised our ability to collect 5 trillion tokens.
๋งˆ์ง€๋ง‰์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ •ํ™•ํ•œ ์ค‘๋ณต ์ œ๊ฑฐ ์ „์— MinHash๋ฅผ ์ ์šฉํ•˜๊ธฐ๋กœ ์„ ํƒํ–ˆ๋Š”๋ฐ, ์ด๋Š” ํ™•์žฅ์ด ๋” ์šฉ์ดํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ทผ์‚ฌ ์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๊ฐ€์ง€์น˜๊ธฐ ๋‹จ๊ณ„๋กœ ์ž‘์šฉํ•˜์—ฌ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ๋”์šฑ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ผ๋ฐ˜์ ์ธ ์„ ํƒ์ธ ์ŠคํŒฌ ์ž๋ฅด๊ธฐ๋ฅผ ์„ ํƒํ–ˆ๋Š”๋ฐ, ๋“œ๋กญํ•‘์€ ๋” ์—„๊ฒฉํ•œ ๊ฑฐ๋ถ€์œจ์„ ์ดˆ๋ž˜ํ•˜์—ฌ 5์กฐ ๊ฐœ์˜ ํ† ํฐ์„ ์ˆ˜์ง‘ํ•˜๋Š” ์šฐ๋ฆฌ์˜ ๋Šฅ๋ ฅ์„ ์ €ํ•ดํ•  ์ˆ˜ ์žˆ์—ˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
๋งˆ์ง€๋ง‰์œผ๋กœ, ๋Œ€๋žต์ ์ธ ์ค‘๋ณต ์ œ๊ฑฐ๋Š” ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ๋” ์‰ฝ๊ฒŒ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ฐ€์ง€์น˜๊ธฐ ๋‹จ๊ณ„ ์—ญํ• ์„ ํ•˜๋ฏ€๋กœ ์ •ํ™•ํ•œ ์ค‘๋ณต ์ œ๊ฑฐ ์ „์— MinHash๋ฅผ ์ ์šฉํ•˜๊ธฐ๋กœ ๊ฒฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์‚ญ์ œ๋Š” ๋”์šฑ ์—„๊ฒฉํ•œ ๊ฑฐ๋ถ€์œจ์„ ์ดˆ๋ž˜ํ•˜์—ฌ 5์กฐ ๊ฐœ์˜ ํ† ํฐ์„ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์ง€์žฅ์„ ์ค„ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ผ๋ฐ˜์ ์ธ ์˜ต์…˜์ธ ์ŠคํŒฌ ์ ˆ๋‹จ์„ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค.
We particularly thank Xindi Wu, Michael Li, and Qian Huang for their helpful and detailed comments, as well as Xuechen Li, Jan-Philipp Franken, Yuhuai Wu, Gabriel Poesia, Winnie Xu, Omar Shaikh, Fan-Yun Sun, Joy He-Yueya, Omar Khattab, and William Yin for useful discussions. In addition, we would like to acknowledge that this work was supported by NSF Grant #2302701.
ํŠนํžˆ ์œ ์ตํ•˜๊ณ  ์ƒ์„ธํ•œ ์˜๊ฒฌ์„ ์ œ๊ณตํ•ด ์ค€ Xindi Wu, Michael Li, Qian Huang์—๊ฒŒ ๊ฐ์‚ฌ๋“œ๋ฆฌ๋ฉฐ, ์œ ์šฉํ•œ ๋…ผ์˜๋ฅผ ํ•จ๊ป˜ํ•œ Xuechen Li, Jan-Philipp Franken, Yuhuai Wu, Gabriel Poesia, Winnie Xu, Omar Shaikh, Fan-Yun Sun, Joy He-Yueya, Omar Khattab, William Yin์—๊ฒŒ๋„ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด ์—ฐ๊ตฌ๋Š” NSF Grant #2302701์˜ ์ง€์›์„ ๋ฐ›์•˜์Œ์„ ๋ฐํž™๋‹ˆ๋‹ค.
ํŠนํžˆ Wu Xindi, Michael Li, Qian Huang์˜ ์œ ์šฉํ•˜๊ณ  ์ƒ์„ธํ•œ ์˜๊ฒฌ์— ๊ฐ์‚ฌ๋“œ๋ฆฌ๋ฉฐ, Xuechen Li, Jan-Philipp Franken, Yuhuai Wu, Gabriel Poesia, Winnie Xu, Omar Shaikh, Fan-Yun Sun, Joy He-Yueya, Omar Khattab, William Yin์˜ ์œ ์šฉํ•œ ๋…ผ์˜์— ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด ์—ฐ๊ตฌ๋Š” NSF ๋ณด์กฐ๊ธˆ #2302701์˜ ์ง€์›์„ ๋ฐ›์•˜์Œ์„ ์ธ์ •ํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.
Figure 4 sheds light on a specific pattern of spectral saturation, roughly co-occurring with the performance saturation phenomenon. It shows that the singular value distribution progressively flattens during training, and nearly reaches uniformity before abruptly evolving towards a spiked distribution with a high maximal singular value, relatively to the other ones.
Figure 4๋Š” ์„ฑ๋Šฅ ํฌํ™” ํ˜„์ƒ๊ณผ ๋Œ€๋žต์ ์œผ๋กœ ๋™์‹œ์— ๋ฐœ์ƒํ•˜๋Š” ํŠน์ •ํ•œ ์ŠคํŽ™ํŠธ๋Ÿผ ํฌํ™” ํŒจํ„ด์„ ์กฐ๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ›ˆ๋ จ ์ค‘์— ํŠน์ด๊ฐ’ ๋ถ„ํฌ๊ฐ€ ์ ์ง„์ ์œผ๋กœ ํ‰ํƒ„ํ•ด์ง€๋ฉฐ, ๋‹ค๋ฅธ ๊ฐ’๋“ค์— ๋น„ํ•ด ๋†’์€ ์ตœ๋Œ€ ํŠน์ด๊ฐ’์„ ๊ฐ€์ง„ ์ŠคํŒŒ์ดํฌ ๋ถ„ํฌ๋กœ ๊ฐ‘์ž‘์Šค๋Ÿฝ๊ฒŒ ๋ณ€ํ™”ํ•˜๊ธฐ ์ „์— ๊ฑฐ์˜ ๊ท ์ผ ์ƒํƒœ์— ๋„๋‹ฌํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 4๋Š” ์„ฑ๋Šฅ ํฌํ™” ํ˜„์ƒ๊ณผ ๊ฑฐ์˜ ๋™์‹œ์— ๋ฐœ์ƒํ•˜๋Š” ์ŠคํŽ™ํŠธ๋Ÿผ ํฌํ™”์˜ ํŠน์ • ํŒจํ„ด์„ ์กฐ๋ช…ํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์ค‘์— ํŠน์ด๊ฐ’ ๋ถ„ํฌ๊ฐ€ ์ ์ง„์ ์œผ๋กœ ํ‰ํ‰ํ•ด์ง€๊ณ , ๊ฑฐ์˜ ๊ท ์ผํ•ด์ง€๋‹ค๊ฐ€ ๋‹ค๋ฅธ ํŠน์ด๊ฐ’์— ๋น„ํ•ด ์ƒ๋Œ€์ ์œผ๋กœ ๋†’์€ ์ตœ๋Œ€ ํŠน์ด๊ฐ’์„ ๊ฐ€์ง„ ์ŠคํŒŒ์ดํฌ ๋ถ„ํฌ๋กœ ๊ฐ‘์ž๊ธฐ ์ง„ํ™”ํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
In Appendices C and G, we show that reconstruction is possible up an _orthogonal_ transformation (approximately \(h^{2}/2\) missing parameters, as opposed to \(h^{2}\) for reconstruction up to an affine transformation), and that this is tight under some formal assumptions. However, we only have an efficient algorithm for reconstruction up to affine transformations.
๋ถ€๋ก C์™€ G์—์„œ, ์šฐ๋ฆฌ๋Š” _์ง๊ต_ ๋ณ€ํ™˜์— ์˜ํ•œ ์žฌ๊ตฌ์„ฑ์ด ๊ฐ€๋Šฅํ•จ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค (๋Œ€๋žต \(h^{2}/2\)์˜ ๋ˆ„๋ฝ๋œ ๋งค๊ฐœ๋ณ€์ˆ˜, ์•„ํ•€ ๋ณ€ํ™˜์— ์˜ํ•œ ์žฌ๊ตฌ์„ฑ์˜ ๊ฒฝ์šฐ \(h^{2}\)์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ), ๊ทธ๋ฆฌ๊ณ  ์ด๊ฒƒ์ด ๋ช‡ ๊ฐ€์ง€ ๊ณต์‹์ ์ธ ๊ฐ€์ • ํ•˜์—์„œ ํƒ€์ดํŠธํ•จ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜, ์šฐ๋ฆฌ๋Š” ์•„ํ•€ ๋ณ€ํ™˜์— ์˜ํ•œ ์žฌ๊ตฌ์„ฑ์„ ์œ„ํ•œ ํšจ์œจ์ ์ธ ์•Œ๊ณ ๋ฆฌ์ฆ˜๋งŒ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
๋ถ€๋ก C์™€ G์—์„œ๋Š” _์ง๊ต_ ๋ณ€ํ™˜(๋Œ€๋žต \(h^{2}/2\) ๋ˆ„๋ฝ๋œ ํŒŒ๋ผ๋ฏธํ„ฐ, ์•„ํ•€ ๋ณ€ํ™˜๊นŒ์ง€ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๊ฒฝ์šฐ \(h^{2}\))๊นŒ์ง€ ์žฌ๊ตฌ์„ฑ์ด ๊ฐ€๋Šฅํ•˜๋ฉฐ, ์ด๊ฒƒ์ด ์ผ๋ถ€ ๊ณต์‹์ ์ธ ๊ฐ€์ • ํ•˜์—์„œ ํƒ€์ดํŠธํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ๋Š” ์•„ํ•€ ๋ณ€ํ™˜๊นŒ์ง€ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ํšจ์œจ์ ์ธ ์•Œ๊ณ ๋ฆฌ์ฆ˜๋งŒ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Previous studies have demonstrated that continual pre-training can significantly improve the model's ability to understand and generate specific content Gupta et al. (2023). However, continual pre-training could lead to catastrophic forgetting French (1999), and limited research has explored the abilities forgotten during pre-training on an existing fine-tuned LLM.
์ด์ „ ์—ฐ๊ตฌ๋“ค์€ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์ด ๋ชจ๋ธ์˜ ํŠน์ • ์ฝ˜ํ…์ธ ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค (Gupta et al., 2023). ๊ทธ๋Ÿฌ๋‚˜ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์€ ์น˜๋ช…์ ์ธ ๋ง๊ฐ์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ (French, 1999), ๊ธฐ์กด์˜ ๋ฏธ์„ธ ์กฐ์ •๋œ LLM์—์„œ ์‚ฌ์ „ ํ›ˆ๋ จ ์ค‘์— ์žŠํ˜€์ง„ ๋Šฅ๋ ฅ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋Š” ์ œํ•œ์ ์ž…๋‹ˆ๋‹ค.
์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์ด ํŠน์ • ์ฝ˜ํ…์ธ ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค(Gupta et al., 2023). ๊ทธ๋Ÿฌ๋‚˜ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์€ ์น˜๋ช…์ ์ธ ๋ง๊ฐ์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ(French, 1999), ๊ธฐ์กด์˜ ๋ฏธ์„ธ ์กฐ์ •๋œ LLM์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ์žŠํ˜€์ง„ ๋Šฅ๋ ฅ์„ ํƒ๊ตฌํ•œ ์—ฐ๊ตฌ๋Š” ์ œํ•œ์ ์ž…๋‹ˆ๋‹ค.
Table 4 is the full version of Table 2, allowing for comparisons between our models and a wider range of models. Some of the numbers are from the Stability AI Japan report and the Rinna leaderboard. Despite having only 7B to 10B parameters, our models achieve higher scores than all existing <70B parameters models and exhibit scores on par with the SOTA 70B models.
Table 4๋Š” Table 2์˜ ์ „์ฒด ๋ฒ„์ „์œผ๋กœ, ์šฐ๋ฆฌ์˜ ๋ชจ๋ธ๊ณผ ๋” ๋„“์€ ๋ฒ”์œ„์˜ ๋ชจ๋ธ ๊ฐ„์˜ ๋น„๊ต๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ถ€ ์ˆ˜์น˜๋Š” Stability AI Japan ๋ณด๊ณ ์„œ์™€ Rinna ๋ฆฌ๋”๋ณด๋“œ์—์„œ ๊ฐ€์ ธ์˜จ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋น„๋ก ์šฐ๋ฆฌ์˜ ๋ชจ๋ธ์ด 7B์—์„œ 10B ํŒŒ๋ผ๋ฏธํ„ฐ๋งŒ์„ ๊ฐ€์ง€๊ณ  ์žˆ์ง€๋งŒ, ๋ชจ๋“  ๊ธฐ์กด <70B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ๋ณด๋‹ค ๋†’์€ ์ ์ˆ˜๋ฅผ ๊ธฐ๋กํ•˜๋ฉฐ SOTA 70B ๋ชจ๋ธ๊ณผ ๋™๋“ฑํ•œ ์ ์ˆ˜๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
ํ‘œ 4๋Š” ํ‘œ 2์˜ ์ „์ฒด ๋ฒ„์ „์œผ๋กœ, ์šฐ๋ฆฌ ๋ชจ๋ธ๊ณผ ๋” ๋„“์€ ๋ฒ”์œ„์˜ ๋ชจ๋ธ ๊ฐ„ ๋น„๊ต๊ฐ€ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ถ€ ์ˆซ์ž๋Š” Stability AI Japan ๋ณด๊ณ ์„œ์™€ Rinna ๋ฆฌ๋”๋ณด๋“œ์—์„œ ๊ฐ€์ ธ์˜จ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ 70์–ต์—์„œ 100์–ต์— ๋ถˆ๊ณผํ•จ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ๊ธฐ์กด 70์–ต ๋ฏธ๋งŒ ๋ชจ๋ธ๋ณด๋‹ค ๋†’์€ ์ ์ˆ˜๋ฅผ ๋‹ฌ์„ฑํ–ˆ์œผ๋ฉฐ SOTA 70์–ต ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์ ์ˆ˜๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค.
Through extensive experiments conducted on the PopQA dataset, we show that Adapt-LLM performs better than its two fixed alternatives: never retrieving and always retrieving relevant context information. Furthermore, our findings highlight Adapt-LLM's capability to effectively discern the necessity of additional context, which is the primary objective of this work.
PopQA ๋ฐ์ดํ„ฐ์…‹์—์„œ ์ˆ˜ํ–‰๋œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด, Adapt-LLM์ด ๋‘ ๊ฐ€์ง€ ๊ณ ์ •๋œ ๋Œ€์•ˆ์ธ '์ ˆ๋Œ€ ๊ฒ€์ƒ‰ํ•˜์ง€ ์•Š๊ธฐ'์™€ 'ํ•ญ์ƒ ๊ด€๋ จ ์ปจํ…์ŠคํŠธ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ธฐ'๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ธ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ์˜ ํ•„์š”์„ฑ์„ ํšจ๊ณผ์ ์œผ๋กœ ํŒ๋ณ„ํ•˜๋Š” Adapt-LLM์˜ ๋Šฅ๋ ฅ์„ ๊ฐ•์กฐํ•˜๋ฉฐ, ์ด๋Š” ์ด ์—ฐ๊ตฌ์˜ ์ฃผ์š” ๋ชฉํ‘œ์ž…๋‹ˆ๋‹ค.
PopQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ˆ˜ํ–‰ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด Adapt-LLM์ด ๊ด€๋ จ ๋ฌธ๋งฅ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜์ง€ ์•Š๊ฑฐ๋‚˜ ํ•ญ์ƒ ๊ฒ€์ƒ‰ํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ๊ณ ์ •๋œ ๋Œ€์•ˆ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ž„์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด ์—ฐ๊ตฌ์˜ ์ฃผ์š” ๋ชฉํ‘œ์ธ ์ถ”๊ฐ€ ๋ฌธ๋งฅ์˜ ํ•„์š”์„ฑ์„ ํšจ๊ณผ์ ์œผ๋กœ ํŒ๋ณ„ํ•  ์ˆ˜ ์žˆ๋Š” Adapt-LLM์˜ ๊ธฐ๋Šฅ์„ ๊ฐ•์กฐํ•˜๋Š” ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
The following is a conversation between a doctor and a patient: <dialog> The conversation should only come to an end if the doctor has finished giving the patient a diagnosis and treatment plan and the patient has no questions left. A conversation also comes to an end if the doctor or patient says goodbye. Question: has the conversation come to an end? Yes or No.
๋‹ค์Œ์€ ์˜์‚ฌ์™€ ํ™˜์ž ๊ฐ„์˜ ๋Œ€ํ™”์ž…๋‹ˆ๋‹ค: <๋Œ€ํ™”> ๋Œ€ํ™”๋Š” ์˜์‚ฌ๊ฐ€ ํ™˜์ž์—๊ฒŒ ์ง„๋‹จ๊ณผ ์น˜๋ฃŒ ๊ณ„ํš์„ ๋งˆ์น˜๊ณ  ํ™˜์ž๊ฐ€ ๋” ์ด์ƒ ์งˆ๋ฌธ์ด ์—†์„ ๋•Œ๋งŒ ๋๋‚˜์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์˜์‚ฌ๋‚˜ ํ™˜์ž๊ฐ€ ์ž‘๋ณ„ ์ธ์‚ฌ๋ฅผ ํ•  ๋•Œ๋„ ๋Œ€ํ™”๊ฐ€ ๋๋‚ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋Œ€ํ™”๊ฐ€ ๋๋‚ฌ์Šต๋‹ˆ๊นŒ? ์˜ˆ ๋˜๋Š” ์•„๋‹ˆ์˜ค.
๋‹ค์Œ์€ ์˜์‚ฌ์™€ ํ™˜์ž ๊ฐ„์˜ ๋Œ€ํ™”์ž…๋‹ˆ๋‹ค: <๋Œ€ํ™”> ์˜์‚ฌ๊ฐ€ ํ™˜์ž์—๊ฒŒ ์ง„๋‹จ๊ณผ ์น˜๋ฃŒ ๊ณ„ํš์„ ์™„์ „ํžˆ ์„ค๋ช…ํ•˜๊ณ  ํ™˜์ž์—๊ฒŒ ๋‚จ์€ ์งˆ๋ฌธ์ด ์—†๋Š” ๊ฒฝ์šฐ์—๋งŒ ๋Œ€ํ™”๋ฅผ ์ข…๋ฃŒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜์‚ฌ๋‚˜ ํ™˜์ž๊ฐ€ ์ž‘๋ณ„ ์ธ์‚ฌ๋ฅผ ํ•˜๋ฉด ๋Œ€ํ™”๋„ ์ข…๋ฃŒ๋ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋Œ€ํ™”๊ฐ€ ๋๋‚ฌ๋‚˜์š”? ์˜ˆ ๋˜๋Š” ์•„๋‹ˆ์˜ค.
Vascepa is safe for you, tell your doctor if you have: 1 liver disease; 2 diabetes; 3 a thyroid disorder; 4 problems with your pancreas; 5 a bleeding or blood-clotting disorder; 6 if you are allergic to fish or shellfish; or, 7 if you drink large amounts of alcohol, Vascepa (icosapent) works in the liver and bloodstream to reduce very low-density triglycerides.##
Vascepa๋Š” ๋‹น์‹ ์—๊ฒŒ ์•ˆ์ „ํ•˜์ง€๋งŒ, ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๊ฒฝ์šฐ ์˜์‚ฌ์—๊ฒŒ ์•Œ๋ฆฌ์‹ญ์‹œ์˜ค: 1. ๊ฐ„ ์งˆํ™˜; 2. ๋‹น๋‡จ๋ณ‘; 3. ๊ฐ‘์ƒ์„  ์งˆํ™˜; 4. ์ทŒ์žฅ ๋ฌธ์ œ; 5. ์ถœํ˜ˆ ๋˜๋Š” ํ˜ˆ์•ก ์‘๊ณ  ์žฅ์• ; 6. ์ƒ์„ ์ด๋‚˜ ์กฐ๊ฐœ๋ฅ˜์— ์•Œ๋ ˆ๋ฅด๊ธฐ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ; ๋˜๋Š” 7. ๋งŽ์€ ์–‘์˜ ์•Œ์ฝ”์˜ฌ์„ ์„ญ์ทจํ•˜๋Š” ๊ฒฝ์šฐ. Vascepa (icosapent)๋Š” ๊ฐ„๊ณผ ํ˜ˆ๋ฅ˜์—์„œ ์ž‘์šฉํ•˜์—ฌ ๋งค์šฐ ๋‚ฎ์€ ๋ฐ€๋„์˜ ํŠธ๋ฆฌ๊ธ€๋ฆฌ์„ธ๋ฆฌ๋“œ๋ฅผ ๊ฐ์†Œ์‹œํ‚ต๋‹ˆ๋‹ค.##
๋ฐ”์„ธํŒŒ๋Š” ์•ˆ์ „ํ•˜์ง€๋งŒ, ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๊ฒฝ์šฐ ์˜์‚ฌ์—๊ฒŒ ์•Œ๋ ค์ฃผ์„ธ์š”: 1 ๊ฐ„ ์งˆํ™˜, 2 ๋‹น๋‡จ๋ณ‘, 3 ๊ฐ‘์ƒ์„  ์งˆํ™˜, 4 ์ทŒ์žฅ ๋ฌธ์ œ, 5 ์ถœํ˜ˆ ๋˜๋Š” ํ˜ˆ์•ก ์‘๊ณ  ์žฅ์• , 6 ์ƒ์„  ๋˜๋Š” ์กฐ๊ฐœ๋ฅ˜์— ์•Œ๋ ˆ๋ฅด๊ธฐ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ, 7 ๋‹ค๋Ÿ‰์˜ ์•Œ์ฝ”์˜ฌ์„ ๋งˆ์‹œ๋Š” ๊ฒฝ์šฐ, ๋ฐ”์„ธํŒŒ(์ด์ฝ”์‚ฌํŽœํŠธ)๋Š” ๊ฐ„๊ณผ ํ˜ˆ๋ฅ˜์—์„œ ์ž‘์šฉํ•˜์—ฌ ์ดˆ์ €๋ฐ€๋„ ์ค‘์„ฑ์ง€๋ฐฉ์„ ๊ฐ์†Œ์‹œํ‚ต๋‹ˆ๋‹ค.##
We show through comprehensive experiments that SLM significantly enhances token efficiency during pretraining and improves performance on downstream tasks. Furthermore, our findings indicate that SLM effectively identifies tokens relevant to the target distribution, resulting in improved perplexity scores on benchmarks for models trained with the selected tokens.
์šฐ๋ฆฌ๋Š” ์ข…ํ•ฉ์ ์ธ ์‹คํ—˜์„ ํ†ตํ•ด SLM์ด ์‚ฌ์ „ ํ›ˆ๋ จ ์ค‘ ํ† ํฐ ํšจ์œจ์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” SLM์ด ๋Œ€์ƒ ๋ถ„ํฌ์™€ ๊ด€๋ จ๋œ ํ† ํฐ์„ ํšจ๊ณผ์ ์œผ๋กœ ์‹๋ณ„ํ•˜์—ฌ ์„ ํƒ๋œ ํ† ํฐ์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์˜ ๋ฒค์น˜๋งˆํฌ์—์„œ ํ–ฅ์ƒ๋œ ํ˜ผ๋ž€๋„ ์ ์ˆ˜๋ฅผ ์–ป๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
ํฌ๊ด„์ ์ธ ์‹คํ—˜์„ ํ†ตํ•ด SLM์ด ์‚ฌ์ „ ํ•™์Šต ์ค‘ ํ† ํฐ ํšจ์œจ์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด SLM์€ ๋Œ€์ƒ ๋ถ„ํฌ์™€ ๊ด€๋ จ๋œ ํ† ํฐ์„ ํšจ๊ณผ์ ์œผ๋กœ ์‹๋ณ„ํ•˜์—ฌ ์„ ํƒ๋œ ํ† ํฐ์œผ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋” ๋‚˜์€ ๋‚œํ•ด์„ฑ ์ ์ˆ˜๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The strengths of virtue ethics are that it emphasizes the importance of personal moral responsibility and development, making it a more individualistic approach. The weakness of this theory is that it may be difficult to determine what virtues are truly universally important, and it does not provide concrete guidelines for decision making in specific situations.
๋• ์œค๋ฆฌ์˜ ๊ฐ•์ ์€ ๊ฐœ์ธ์˜ ๋„๋•์  ์ฑ…์ž„๊ณผ ๋ฐœ์ „์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•˜์—ฌ ๋ณด๋‹ค ๊ฐœ์ธ์ฃผ์˜์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ทจํ•œ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์ด ์ด๋ก ์˜ ์•ฝ์ ์€ ์–ด๋–ค ๋•๋ชฉ์ด ์ง„์ •์œผ๋กœ ๋ณดํŽธ์ ์œผ๋กœ ์ค‘์š”ํ•œ์ง€ ๊ฒฐ์ •ํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํŠน์ • ์ƒํ™ฉ์—์„œ ์˜์‚ฌ ๊ฒฐ์ •์„ ์œ„ํ•œ ๊ตฌ์ฒด์ ์ธ ์ง€์นจ์„ ์ œ๊ณตํ•˜์ง€ ์•Š๋Š”๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค.
๋ฏธ๋• ์œค๋ฆฌ์˜ ๊ฐ•์ ์€ ๊ฐœ์ธ์˜ ๋„๋•์  ์ฑ…์ž„๊ณผ ๋ฐœ์ „์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•˜์—ฌ ๋ณด๋‹ค ๊ฐœ์ธ์ฃผ์˜์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์ด๋ผ๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์ด ์ด๋ก ์˜ ์•ฝ์ ์€ ์–ด๋–ค ๋ฏธ๋•์ด ์ง„์ •์œผ๋กœ ๋ณดํŽธ์ ์œผ๋กœ ์ค‘์š”ํ•œ์ง€ ํŒ๋‹จํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์œผ๋ฉฐ ํŠน์ • ์ƒํ™ฉ์—์„œ์˜ ์˜์‚ฌ ๊ฒฐ์ •์— ๋Œ€ํ•œ ๊ตฌ์ฒด์ ์ธ ์ง€์นจ์„ ์ œ๊ณตํ•˜์ง€ ๋ชปํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Figure 7: Architecture of Yi-VL models. Symbols are used to denote the training status of various modules at three training stages: a fire icon ( ) indicates the parameters of the module are trainable, while a snowflake icon () signifies that parameters are frozen. The image resolution used in ViT at each stage, either \(224^{2}\) or \(448^{2}\), is also marked.
Figure 7: Yi-VL ๋ชจ๋ธ์˜ ์•„ํ‚คํ…์ฒ˜. ์„ธ ๊ฐ€์ง€ ํ•™์Šต ๋‹จ๊ณ„์—์„œ ๋‹ค์–‘ํ•œ ๋ชจ๋“ˆ์˜ ํ•™์Šต ์ƒํƒœ๋ฅผ ๋‚˜ํƒ€๋‚ด๊ธฐ ์œ„ํ•ด ๊ธฐํ˜ธ๊ฐ€ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค: ๋ถˆ๊ฝƒ ์•„์ด์ฝ˜ ( )์€ ๋ชจ๋“ˆ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ํ•™์Šต ๊ฐ€๋Šฅํ•จ์„ ๋‚˜ํƒ€๋‚ด๊ณ , ๋ˆˆ์†ก์ด ์•„์ด์ฝ˜ ( )์€ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ๊ณ ์ •๋จ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋‹จ๊ณ„์—์„œ ViT์— ์‚ฌ์šฉ๋œ ์ด๋ฏธ์ง€ ํ•ด์ƒ๋„, \(224^{2}\) ๋˜๋Š” \(448^{2}\), ๋˜ํ•œ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 7: Yi-VL ๋ชจ๋ธ์˜ ์•„ํ‚คํ…์ฒ˜. ์„ธ ๊ฐ€์ง€ ํ›ˆ๋ จ ๋‹จ๊ณ„์—์„œ ๋‹ค์–‘ํ•œ ๋ชจ๋“ˆ์˜ ํ›ˆ๋ จ ์ƒํƒœ๋ฅผ ๋‚˜ํƒ€๋‚ด๊ธฐ ์œ„ํ•ด ๊ธฐํ˜ธ๊ฐ€ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๋ถˆ ์•„์ด์ฝ˜()์€ ๋ชจ๋“ˆ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋‚ด๊ณ , ๋ˆˆ์†ก์ด ์•„์ด์ฝ˜()์€ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ๋™๊ฒฐ๋˜์—ˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ฐ ๋‹จ๊ณ„์—์„œ ViT์— ์‚ฌ์šฉ๋˜๋Š” ์ด๋ฏธ์ง€ ํ•ด์ƒ๋„์ธ \(224^{2}\) ๋˜๋Š” \(448^{2}\)๋„ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค.
test dataset as the baseline. While existing studies suggest LMs can be better automated evaluators than existing metrics [10], we conduct a human study to justify our usage of GPT-4 for evaluation in Sec. 6.4. We find GPT-4 judgments correlate strongly with humans, with human agreement with GPT-4 typically similar or higher than inter-human annotator agreement.
๊ธฐ์ค€์„ ์œผ๋กœ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ์…‹์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ๋“ค์€ LMs๊ฐ€ ๊ธฐ์กด์˜ ํ‰๊ฐ€ ์ง€ํ‘œ๋ณด๋‹ค ๋” ๋‚˜์€ ์ž๋™ ํ‰๊ฐ€์ž๊ฐ€ ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ œ์•ˆํ•˜์ง€๋งŒ [10], ์šฐ๋ฆฌ๋Š” Sec. 6.4์—์„œ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด GPT-4๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ์ •๋‹นํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” GPT-4์˜ ํŒ๋‹จ์ด ์ธ๊ฐ„๊ณผ ๊ฐ•ํ•˜๊ฒŒ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ธ๊ฐ„์ด GPT-4์™€ ๋™์˜ํ•˜๋Š” ์ •๋„๊ฐ€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„ ํ‰๊ฐ€์ž ๊ฐ„์˜ ๋™์˜ ์ •๋„์™€ ๋น„์Šทํ•˜๊ฑฐ๋‚˜ ๋” ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ธฐ์ค€์„ ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” LM์ด ๊ธฐ์กด ๋ฉ”ํŠธ๋ฆญ๋ณด๋‹ค ๋” ๋‚˜์€ ์ž๋™ํ™”๋œ ํ‰๊ฐ€์ž๊ฐ€ ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ œ์•ˆํ•˜์ง€๋งŒ[10], ์„น์…˜ 6.4์—์„œ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด GPT-4๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ์ •๋‹นํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. GPT-4์˜ ํŒ๋‹จ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„๊ณผ์˜ ์ผ์น˜๋„๊ฐ€ ๋†’๊ฑฐ๋‚˜ ๋†’์€ ์ธ๊ฐ„๊ณผ์˜ ์ผ์น˜๋„์™€ ์œ ์‚ฌํ•˜๊ฑฐ๋‚˜ ๋†’์€ ์ธ๊ฐ„๊ณผ์˜ ์ผ์น˜๋„๋ฅผ ๋ณด์ด๋Š” ๋“ฑ GPT-4์˜ ํŒ๋‹จ์ด ์ธ๊ฐ„๊ณผ ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
In the end, Sophie realized that she had to let go of the sphere and the crystal. She knew that they were too dangerous to play with, and that she needed to focus on finding a way to harness their power without becoming consumed by it. She knew that it would be a difficult journey, but she was determined to find a way to use the sphere and the crystal for good.
๊ฒฐ๊ตญ ์†Œํ”ผ๋Š” ๊ตฌ์™€ ํฌ๋ฆฌ์Šคํƒˆ์„ ๋†“์•„์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊นจ๋‹ฌ์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๊ทธ๊ฒƒ๋“ค์ด ๋„ˆ๋ฌด ์œ„ํ—˜ํ•ด์„œ ๊ฐ€์ง€๊ณ  ๋†€ ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ์„ ์•Œ์•˜๊ณ , ๊ทธ๊ฒƒ๋“ค์˜ ํž˜์„ ์ด์šฉํ•˜๋ฉด์„œ๋„ ๊ทธ ํž˜์— ํœ˜๋ง๋ฆฌ์ง€ ์•Š๋Š” ๋ฐฉ๋ฒ•์„ ์ฐพ์•„์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊นจ๋‹ฌ์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๊ทธ๊ฒƒ์ด ์–ด๋ ค์šด ์—ฌ์ •์ด ๋  ๊ฒƒ์ด๋ผ๋Š” ๊ฒƒ์„ ์•Œ์•˜์ง€๋งŒ, ๊ตฌ์™€ ํฌ๋ฆฌ์Šคํƒˆ์„ ์„ ํ•œ ๋ชฉ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ๋ฐฉ๋ฒ•์„ ์ฐพ๊ธฐ๋กœ ๊ฒฐ์‹ฌํ–ˆ์Šต๋‹ˆ๋‹ค.
๊ฒฐ๊ตญ ์†Œํ”ผ๋Š” ๊ตฌ์ฒด์™€ ์ˆ˜์ •์„ ๋†“์•„์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊นจ๋‹ฌ์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๊ทธ๊ฒƒ๋“ค์ด ๋„ˆ๋ฌด ์œ„ํ—˜ํ•ด์„œ ๊ฐ€์ง€๊ณ  ๋†€๊ธฐ์—๋Š” ์•ˆ ๋œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ์•˜๊ณ , ๊ทธ ํž˜์— ํœฉ์“ธ๋ฆฌ์ง€ ์•Š๊ณ  ๊ทธ ํž˜์„ ํ™œ์šฉํ•  ๋ฐฉ๋ฒ•์„ ์ฐพ๋Š” ๋ฐ ์ง‘์ค‘ํ•ด์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๊ทธ ์—ฌ์ •์ด ํž˜๋“ค๋‹ค๋Š” ๊ฒƒ์„ ์•Œ์•˜์ง€๋งŒ, ๊ตฌ์ฒด์™€ ์ˆ˜์ •์„ ์„ ํ•˜๊ฒŒ ์‚ฌ์šฉํ•  ๋ฐฉ๋ฒ•์„ ์ฐพ๊ธฐ๋กœ ๊ฒฐ์‹ฌํ–ˆ์Šต๋‹ˆ๋‹ค.
We thank GENCI for generously granting us access to their cutting-edge computing resources. Our model, SaulLM-7B, has been trained on ADASTRA, with initial experimentation conducted on Jeanzay. The utilization of HPC resources was made possible through the Jeanzay grants 101838, 103256, and 103298, as well as the Adastra grants C1615122, CAD14770, and CAD15031.
GENCI๊ฐ€ ์ตœ์ฒจ๋‹จ ์ปดํ“จํŒ… ์ž์›์„ ์•„๋‚Œ์—†์ด ์ œ๊ณตํ•ด ์ฃผ์‹  ๊ฒƒ์— ๋Œ€ํ•ด ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์ €ํฌ ๋ชจ๋ธ์ธ SaulLM-7B๋Š” ADASTRA์—์„œ ํ›ˆ๋ จ๋˜์—ˆ์œผ๋ฉฐ, ์ดˆ๊ธฐ ์‹คํ—˜์€ Jeanzay์—์„œ ์ˆ˜ํ–‰๋˜์—ˆ์Šต๋‹ˆ๋‹ค. HPC ์ž์›์˜ ํ™œ์šฉ์€ Jeanzay์˜ 101838, 103256, 103298 ๋ณด์กฐ๊ธˆ๊ณผ Adastra์˜ C1615122, CAD14770, CAD15031 ๋ณด์กฐ๊ธˆ์„ ํ†ตํ•ด ๊ฐ€๋Šฅํ–ˆ์Šต๋‹ˆ๋‹ค.
์ตœ์ฒจ๋‹จ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค์— ๋Œ€ํ•œ ๊ด€๋Œ€ํ•œ ์ ‘๊ทผ ๊ถŒํ•œ์„ ๋ถ€์—ฌํ•ด์ค€ GENCI์— ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์ €ํฌ ๋ชจ๋ธ์ธ SaulLM-7B๋Š” ADASTRA์—์„œ ํ•™์Šต๋˜์—ˆ์œผ๋ฉฐ, ์ดˆ๊ธฐ ์‹คํ—˜์€ Jeanzay์—์„œ ์ˆ˜ํ–‰๋˜์—ˆ์Šต๋‹ˆ๋‹ค. HPC ๋ฆฌ์†Œ์Šค์˜ ํ™œ์šฉ์€ Jeanzay ๋ณด์กฐ๊ธˆ 101838, 103256, 103298 ๋ฐ Adastra ๋ณด์กฐ๊ธˆ C1615122, CAD14770, CAD15031์„ ํ†ตํ•ด ๊ฐ€๋Šฅํ–ˆ์Šต๋‹ˆ๋‹ค.
Benchmarking is widely used by the AI community to identify, measure and track improvements. Initiatives such as MLPerf [2, 161], BIG-Bench [20] and HELM [19] have served as a powerful forcing function to drive progress in the field. We believe that well-designed and responsibly released benchmarks can play an important role in driving innovation and research.
๋ฒค์น˜๋งˆํ‚น์€ AI ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ๊ฐœ์„  ์‚ฌํ•ญ์„ ์‹๋ณ„ํ•˜๊ณ  ์ธก์ •ํ•˜๋ฉฐ ์ถ”์ ํ•˜๊ธฐ ์œ„ํ•ด ๋„๋ฆฌ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. MLPerf [2, 161], BIG-Bench [20] ๋ฐ HELM [19]์™€ ๊ฐ™์€ ์ด๋‹ˆ์…”ํ‹ฐ๋ธŒ๋Š” ์ด ๋ถ„์•ผ์˜ ๋ฐœ์ „์„ ์ด‰์ง„ํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ์ถ”์ง„ ๊ธฐ๋Šฅ์œผ๋กœ ์ž‘์šฉํ•ด ์™”์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ž˜ ์„ค๊ณ„๋˜๊ณ  ์ฑ…์ž„๊ฐ ์žˆ๊ฒŒ ๊ณต๊ฐœ๋œ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ํ˜์‹ ๊ณผ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค.
๋ฒค์น˜๋งˆํ‚น์€ AI ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ, ๊ฐœ์„  ์‚ฌํ•ญ์„ ํŒŒ์•…, ์ธก์ • ๋ฐ ์ถ”์ ํ•ฉ๋‹ˆ๋‹ค. MLPerf [2, 161], BIG-Bench [20] ๋ฐ HELM [19]๊ณผ ๊ฐ™์€ ์ด๋‹ˆ์…”ํ‹ฐ๋ธŒ๋Š” ์ด ๋ถ„์•ผ์˜ ๋ฐœ์ „์„ ์ฃผ๋„ํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ์›๋™๋ ฅ์œผ๋กœ ์ž‘์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ž˜ ์„ค๊ณ„๋˜๊ณ  ์ฑ…์ž„๊ฐ ์žˆ๊ฒŒ ๊ณต๊ฐœ๋œ ๋ฒค์น˜๋งˆํฌ๋Š” ํ˜์‹ ๊ณผ ์—ฐ๊ตฌ๋ฅผ ์ฃผ๋„ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค.
Comparison of Synthetic SFT Data versus Real DataTo assess the quality of the synthetically generated math questions, we evaluate their effectiveness against real questions from the GSM8K and MATH training sets, utilizing a LLaMA-2 7B model, as detailed in Table 1. The results indicate that the synthetic math questions are nearly as effective as the real ones.
ํ•ฉ์„ฑ SFT ๋ฐ์ดํ„ฐ์™€ ์‹ค์ œ ๋ฐ์ดํ„ฐ์˜ ๋น„๊ต ํ•ฉ์„ฑ์ ์œผ๋กœ ์ƒ์„ฑ๋œ ์ˆ˜ํ•™ ๋ฌธ์ œ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” GSM8K ๋ฐ MATH ํ›ˆ๋ จ ์„ธํŠธ์˜ ์‹ค์ œ ๋ฌธ์ œ์™€ ๋น„๊ตํ•˜์—ฌ LLaMA-2 7B ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ทธ ํšจ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Table 1์— ์ž์„ธํžˆ ๋‚˜์™€ ์žˆ๋“ฏ์ด, ๊ฒฐ๊ณผ๋Š” ํ•ฉ์„ฑ ์ˆ˜ํ•™ ๋ฌธ์ œ๊ฐ€ ์‹ค์ œ ๋ฌธ์ œ๋งŒํผ ๊ฑฐ์˜ ํšจ๊ณผ์ ์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
ํ•ฉ์„ฑ SFT ๋ฐ์ดํ„ฐ์™€ ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋น„๊ตํ•ฉ์„ฑ์œผ๋กœ ์ƒ์„ฑ๋œ ์ˆ˜ํ•™ ๋ฌธ์ œ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ํ‘œ 1์— ์ž์„ธํžˆ ์„ค๋ช…๋œ ๋Œ€๋กœ LLaMA-2 7B ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ GSM8K ๋ฐ MATH ํ•™์Šต ์„ธํŠธ์˜ ์‹ค์ œ ๋ฌธ์ œ์™€ ๋น„๊ตํ•˜์—ฌ ๊ทธ ํšจ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ํ•ฉ์„ฑ ์ˆ˜ํ•™ ๋ฌธ์ œ๊ฐ€ ์‹ค์ œ ๋ฌธ์ œ์™€ ๊ฑฐ์˜ ๋น„์Šทํ•œ ํšจ๊ณผ๊ฐ€ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
We show the average results across all six directions in Table 4, and provide the performance in each direction in Appendix G due to the space constraint. Consistent with observations from WMT'21 and WMT'22, ALMA-13B-R surpasses contemporary moderate-size LLM-based translators such as ALMA-13B-LoRA and TowerInstruct, and either matches or exceeds WMT winners.
Table 4์—์„œ ๋ชจ๋“  ์—ฌ์„ฏ ๊ฐ€์ง€ ๋ฐฉํ–ฅ์— ๋Œ€ํ•œ ํ‰๊ท  ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ๊ณ , ๊ณต๊ฐ„ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ๊ฐ ๋ฐฉํ–ฅ์˜ ์„ฑ๋Šฅ์€ ๋ถ€๋ก G์— ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. WMT'21 ๋ฐ WMT'22์˜ ๊ด€์ฐฐ๊ณผ ์ผ์น˜ํ•˜๊ฒŒ, ALMA-13B-R์€ ALMA-13B-LoRA ๋ฐ TowerInstruct์™€ ๊ฐ™์€ ํ˜„๋Œ€์˜ ์ค‘๊ฐ„ ํฌ๊ธฐ LLM ๊ธฐ๋ฐ˜ ๋ฒˆ์—ญ๊ธฐ๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ, WMT ์šฐ์Šน์ž์™€ ๋™๋“ฑํ•˜๊ฑฐ๋‚˜ ๊ทธ ์ด์ƒ์ž…๋‹ˆ๋‹ค.
ํ‘œ 4์— 6๊ฐ€์ง€ ๋ฐฉํ–ฅ ๋ชจ๋‘์— ๋Œ€ํ•œ ํ‰๊ท  ๊ฒฐ๊ณผ๋ฅผ ํ‘œ์‹œํ•˜๊ณ , ๊ณต๊ฐ„ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ๊ฐ ๋ฐฉํ–ฅ์˜ ์„ฑ๋Šฅ์€ ๋ถ€๋ก G์— ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. WMT'21๊ณผ WMT'22์˜ ๊ด€์ธก ๊ฒฐ๊ณผ์™€ ์ผ์น˜ํ•˜๊ฒŒ๋„ ALMA-13B-R์€ ALMA-13B-LoRA ๋ฐ TowerInstruct์™€ ๊ฐ™์€ ํ˜„๋Œ€์˜ ์ค‘๊ฐ„ ๊ทœ๋ชจ LLM ๊ธฐ๋ฐ˜ ๋ฒˆ์—ญ๊ธฐ๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ, WMT ์ˆ˜์ƒ์ž์™€ ๋™๋“ฑํ•˜๊ฑฐ๋‚˜ ๊ทธ ์ด์ƒ์˜ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค.
Figure 5: Analyses for RAG preference rate against prior probability and deviation, using GPT-4 (blue), GPT-3.5 (orange), and Mistral-7B (green). Please see Figure 2 for full figure descriptions. Of note, some models did not generate any meaningful prior responses (due to refusal, improper responses, etc.) for certain datasets and thus could not be analyzed.
Figure 5: GPT-4(ํŒŒ๋ž€์ƒ‰), GPT-3.5(์ฃผํ™ฉ์ƒ‰), Mistral-7B(๋…น์ƒ‰)์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ™•๋ฅ  ๋ฐ ํŽธ์ฐจ์— ๋Œ€ํ•œ RAG ์„ ํ˜ธ์œจ ๋ถ„์„. ์ „์ฒด ๊ทธ๋ฆผ ์„ค๋ช…์€ Figure 2๋ฅผ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค. ์ฃผ๋ชฉํ•  ์ ์€, ์ผ๋ถ€ ๋ชจ๋ธ์€ ํŠน์ • ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•ด ์˜๋ฏธ ์žˆ๋Š” ์‚ฌ์ „ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜์ง€ ๋ชปํ–ˆ๊ธฐ ๋•Œ๋ฌธ์—(๊ฑฐ๋ถ€, ๋ถ€์ ์ ˆํ•œ ์‘๋‹ต ๋“ฑ) ๋ถ„์„ํ•  ์ˆ˜ ์—†์—ˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 5: ์ด์ „ ํ™•๋ฅ ๊ณผ ํŽธ์ฐจ์— ๋Œ€ํ•œ RAG ์„ ํ˜ธ๋„ ๋ถ„์„, GPT-4(ํŒŒ๋ž€์ƒ‰), GPT-3.5(์ฃผํ™ฉ์ƒ‰), Mistral-7B(๋…น์ƒ‰)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „์ฒด ๊ทธ๋ฆผ ์„ค๋ช…์€ ๊ทธ๋ฆผ 2๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ์ผ๋ถ€ ๋ชจ๋ธ์€ ํŠน์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์˜๋ฏธ ์žˆ๋Š” ์ด์ „ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์—(๊ฑฐ๋ถ€, ๋ถ€์ ์ ˆํ•œ ์‘๋‹ต ๋“ฑ) ๋ถ„์„ํ•  ์ˆ˜ ์—†์—ˆ์Šต๋‹ˆ๋‹ค.
information of the input image. Three methods for the modeling of spatio-temporal relationships are proposed for the decoder, taking full advantage of the properties and benefits of the Mamba architecture. He et al. [117] capture long-range and local information effectively through parallel cascaded hybrid state space and multi-kernel convolution operations.
์ž…๋ ฅ ์ด๋ฏธ์ง€์˜ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋””์ฝ”๋”๋ฅผ ์œ„ํ•ด ์‹œ๊ณต๊ฐ„ ๊ด€๊ณ„๋ฅผ ๋ชจ๋ธ๋งํ•˜๋Š” ์„ธ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์ด ์ œ์•ˆ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” Mamba ์•„ํ‚คํ…์ฒ˜์˜ ํŠน์„ฑ๊ณผ ์ด์ ์„ ์ตœ๋Œ€ํ•œ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. He et al. [117]์€ ๋ณ‘๋ ฌ ๊ณ„๋‹จ์‹ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์ƒํƒœ ๊ณต๊ฐ„ ๋ฐ ๋‹ค์ค‘ ์ปค๋„ ์ปจ๋ณผ๋ฃจ์…˜ ์ž‘์—…์„ ํ†ตํ•ด ์žฅ๊ฑฐ๋ฆฌ ๋ฐ ๋กœ์ปฌ ์ •๋ณด๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํฌ์ฐฉํ•ฉ๋‹ˆ๋‹ค.
์ž…๋ ฅ ์ด๋ฏธ์ง€์˜ ์ •๋ณด. Mamba ์•„ํ‚คํ…์ฒ˜์˜ ์†์„ฑ๊ณผ ์ด์ ์„ ์ตœ๋Œ€ํ•œ ํ™œ์šฉํ•˜์—ฌ ๋””์ฝ”๋”์— ์‹œ๊ณต๊ฐ„ ๊ด€๊ณ„ ๋ชจ๋ธ๋ง์„ ์œ„ํ•œ ์„ธ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. He ๋“ฑ[117]์€ ๋ณ‘๋ ฌ ์บ์Šค์ผ€์ด๋“œ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์ƒํƒœ ๊ณต๊ฐ„๊ณผ ๋ฉ€ํ‹ฐ ์ปค๋„ ์ปจ๋ณผ๋ฃจ์…˜ ์—ฐ์‚ฐ์„ ํ†ตํ•ด ์žฅ๊ฑฐ๋ฆฌ ๋ฐ ๋กœ์ปฌ ์ •๋ณด๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์บก์ฒ˜ํ•ฉ๋‹ˆ๋‹ค.
First of all, as discussed in the Ethical Considerations, our method requires a strong LLM to generate the data, so the performance of our method depends on the quality of the LLM and may inherit bias and fairness issues from it. On the other hand, CodecLM can benefit from stronger LLMs improved with advanced bias-reducing and fairness-enhancing approaches.
์šฐ์„ , ์œค๋ฆฌ์  ๊ณ ๋ ค ์‚ฌํ•ญ์—์„œ ๋…ผ์˜๋œ ๋ฐ”์™€ ๊ฐ™์ด, ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ•๋ ฅํ•œ LLM์ด ํ•„์š”ํ•˜๋ฏ€๋กœ, ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์€ LLM์˜ ํ’ˆ์งˆ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๋ฉฐ, LLM์œผ๋กœ๋ถ€ํ„ฐ ํŽธํ–ฅ ๋ฐ ๊ณต์ •์„ฑ ๋ฌธ์ œ๋ฅผ ๋ฌผ๋ ค๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์—, CodecLM์€ ๊ณ ๊ธ‰ ํŽธํ–ฅ ๊ฐ์†Œ ๋ฐ ๊ณต์ •์„ฑ ํ–ฅ์ƒ ์ ‘๊ทผ๋ฒ•์œผ๋กœ ๊ฐœ์„ ๋œ ๋” ๊ฐ•๋ ฅํ•œ LLM์œผ๋กœ๋ถ€ํ„ฐ ํ˜œํƒ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์šฐ์„ , ์œค๋ฆฌ์  ๊ณ ๋ ค ์‚ฌํ•ญ์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ•๋ ฅํ•œ LLM์ด ํ•„์š”ํ•˜๋ฏ€๋กœ ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์€ LLM์˜ ํ’ˆ์งˆ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๋ฉฐ, ์ด๋กœ๋ถ€ํ„ฐ ํŽธํ–ฅ์„ฑ ๋ฐ ๊ณต์ •์„ฑ ๋ฌธ์ œ๋ฅผ ์ƒ์†ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์ฝ”๋ฑLM์€ ๊ณ ๊ธ‰ ํŽธํ–ฅ์„ฑ ๊ฐ์†Œ ๋ฐ ๊ณต์ •์„ฑ ํ–ฅ์ƒ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ๊ฐœ์„ ๋œ ๋” ๊ฐ•๋ ฅํ•œ LLM์˜ ์ด์ ์„ ๋ˆ„๋ฆด ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Unfortunately this has several significant drawbacks: the threshold has to be independent of \(h\) (or learning the threshold would reveal \(h\)); the system would need to maintain state of all user queries to the API; and preventing Sybil attacks requires a global pool of user queries, which can present significant privacy risks (Debenedetti et al., 2023).
๋ถˆํ–‰ํžˆ๋„ ์ด๊ฒƒ์€ ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๋‹จ์ ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค: ์ž„๊ณ„๊ฐ’์€ \(h\)์™€ ๋…๋ฆฝ์ ์ด์–ด์•ผ ํ•˜๋ฉฐ (๋˜๋Š” ์ž„๊ณ„๊ฐ’์„ ํ•™์Šตํ•˜๋ฉด \(h\)๊ฐ€ ๋“œ๋Ÿฌ๋‚  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค); ์‹œ์Šคํ…œ์€ API์— ๋Œ€ํ•œ ๋ชจ๋“  ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ์˜ ์ƒํƒœ๋ฅผ ์œ ์ง€ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค; ๊ทธ๋ฆฌ๊ณ  Sybil ๊ณต๊ฒฉ์„ ๋ฐฉ์ง€ํ•˜๋ ค๋ฉด ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ์˜ ๊ธ€๋กœ๋ฒŒ ํ’€์„ ํ•„์š”๋กœ ํ•˜๋Š”๋ฐ, ์ด๋Š” ์ƒ๋‹นํ•œ ํ”„๋ผ์ด๋ฒ„์‹œ ์œ„ํ—˜์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค (Debenedetti et al., 2023).
์•ˆํƒ€๊น๊ฒŒ๋„ ์ด ๋ฐฉ๋ฒ•์—๋Š” ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๋‹จ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ž„๊ณ„๊ฐ’์€ \(h\)์™€ ๋ฌด๊ด€ํ•ด์•ผ ํ•˜๋ฉฐ(๋˜๋Š” ์ž„๊ณ„๊ฐ’์„ ํ•™์Šตํ•˜๋ฉด \(h\)๊ฐ€ ๋“œ๋Ÿฌ๋‚ฉ๋‹ˆ๋‹ค); ์‹œ์Šคํ…œ์€ API์— ๋Œ€ํ•œ ๋ชจ๋“  ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ ์ƒํƒœ๋ฅผ ์œ ์ง€ํ•ด์•ผ ํ•˜๋ฉฐ; ์‹œ๋นŒ ๊ณต๊ฒฉ์„ ๋ฐฉ์ง€ํ•˜๋ ค๋ฉด ์ „์—ญ์ ์ธ ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ ํ’€์ด ํ•„์š”ํ•˜๋ฉฐ, ์ด๋Š” ์ƒ๋‹นํ•œ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ์œ„ํ—˜์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Debenedetti et al., 2023).
Fitting the mixture law requires us to first experiment on a few mixtures and obtain their losses. The sample mixture chosen for fitting could largely affect the prediction accuracy. Consider an extreme case where all sample mixtures have proportions around a small region, it is hardly possible to fit a law that reliably predicts the whole proportion space.
ํ˜ผํ•ฉ ๋ฒ•์น™์„ ๋งž์ถ”๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋จผ์ € ๋ช‡ ๊ฐ€์ง€ ํ˜ผํ•ฉ๋ฌผ์— ๋Œ€ํ•ด ์‹คํ—˜์„ ํ•˜๊ณ  ๊ทธ๋“ค์˜ ์†์‹ค์„ ์–ป์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋งž์ถ”๊ธฐ ์œ„ํ•ด ์„ ํƒ๋œ ์ƒ˜ํ”Œ ํ˜ผํ•ฉ๋ฌผ์€ ์˜ˆ์ธก ์ •ํ™•๋„์— ํฌ๊ฒŒ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ์ƒ˜ํ”Œ ํ˜ผํ•ฉ๋ฌผ์ด ์ž‘์€ ์˜์—ญ ์ฃผ๋ณ€์— ๋น„์œจ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๊ทน๋‹จ์ ์ธ ๊ฒฝ์šฐ๋ฅผ ์ƒ๊ฐํ•ด๋ณด๋ฉด, ์ „์ฒด ๋น„์œจ ๊ณต๊ฐ„์„ ์‹ ๋ขฐ์„ฑ ์žˆ๊ฒŒ ์˜ˆ์ธกํ•˜๋Š” ๋ฒ•์น™์„ ๋งž์ถ”๋Š” ๊ฒƒ์€ ๊ฑฐ์˜ ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
ํ˜ผํ•ฉ ๋ฒ•์น™์„ ๋งž์ถ”๋ ค๋ฉด ๋จผ์ € ๋ช‡ ๊ฐ€์ง€ ํ˜ผํ•ฉ๋ฌผ์„ ์‹คํ—˜ํ•˜๊ณ  ๊ทธ ์†์‹ค์„ ์–ป์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋งž์ถค์— ์„ ํƒ๋œ ์ƒ˜ํ”Œ ํ˜ผํ•ฉ๋ฌผ์€ ์˜ˆ์ธก ์ •ํ™•๋„์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ์ƒ˜ํ”Œ ํ˜ผํ•ฉ๋ฌผ์˜ ๋น„์œจ์ด ์ž‘์€ ์˜์—ญ์— ์žˆ๋Š” ๊ทน๋‹จ์ ์ธ ๊ฒฝ์šฐ๋ฅผ ์ƒ๊ฐํ•ด ๋ณด๋ฉด ์ „์ฒด ๋น„์œจ ๊ณต๊ฐ„์„ ์•ˆ์ •์ ์œผ๋กœ ์˜ˆ์ธกํ•˜๋Š” ๋ฒ•์น™์„ ๋งž์ถ”๊ธฐ๊ฐ€ ๊ฑฐ์˜ ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
To investigate _why_ the LLaMA architecture is inferior to GPT2 in the 100-exposure (insufficiently trained) setting, we closely examine LLaMA by gradually modifying its architecture _back towards_ GPT2 to identify the key architectural changes. We start by tying weights, as this enhances tiny LLaMA model's capacity in the 1000-exposure setting (Result 5).
LLaMA ์•„ํ‚คํ…์ฒ˜๊ฐ€ 100ํšŒ ๋…ธ์ถœ(์ถฉ๋ถ„ํžˆ ํ›ˆ๋ จ๋˜์ง€ ์•Š์€) ์„ค์ •์—์„œ GPT2๋ณด๋‹ค ์—ด๋“ฑํ•œ ์ด์œ ๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” LLaMA์˜ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ ์ง„์ ์œผ๋กœ GPT2๋กœ ๋˜๋Œ๋ ค ์ฃผ์š” ์•„ํ‚คํ…์ฒ˜ ๋ณ€๊ฒฝ ์‚ฌํ•ญ์„ ์‹๋ณ„ํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ LLaMA๋ฅผ ๋ฉด๋ฐ€ํžˆ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๊ฐ€์ค‘์น˜๋ฅผ ๋ฌถ๋Š” ๊ฒƒ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜๋Š”๋ฐ, ์ด๋Š” 1000ํšŒ ๋…ธ์ถœ ์„ค์ •์—์„œ ์ž‘์€ LLaMA ๋ชจ๋ธ์˜ ์šฉ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค(๊ฒฐ๊ณผ 5).
100ํšŒ ๋…ธ์ถœ(์ถฉ๋ถ„ํžˆ ํ•™์Šต๋˜์ง€ ์•Š์€) ์„ค์ •์—์„œ LLaMA ์•„ํ‚คํ…์ฒ˜๊ฐ€ GPT2๋ณด๋‹ค ์—ด๋“ฑํ•œ _์ด์œ _๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด, ์ฃผ์š” ์•„ํ‚คํ…์ฒ˜ ๋ณ€๊ฒฝ ์‚ฌํ•ญ์„ ์‹๋ณ„ํ•˜๊ธฐ ์œ„ํ•ด LLaMA์˜ ์•„ํ‚คํ…์ฒ˜๋ฅผ GPT2๋กœ ์ ์ง„์ ์œผ๋กœ ์ˆ˜์ •ํ•˜์—ฌ ๋ฉด๋ฐ€ํžˆ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ๊ฐ€์ค‘์น˜๋ฅผ ์—ฐ๊ฒฐํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•˜๋Š”๋ฐ, ์ด๋Š” 1000ํšŒ ๋…ธ์ถœ ์„ค์ •์—์„œ ์ž‘์€ LLaMA ๋ชจ๋ธ์˜ ์šฉ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค(๊ฒฐ๊ณผ 5).
In the subsequent voting phase, to evaluate the similarity between samples, as outlined in Algorithm 1, we employ mathematical equivalence comparisons for each sample. The most probable sample is chosen as the final answer. This answer is subjected to a comparison of mathematical equivalence with the ground truth to ascertain the correctness of the result.
ํ›„์† ํˆฌํ‘œ ๋‹จ๊ณ„์—์„œ๋Š”, Algorithm 1์— ์„ค๋ช…๋œ ๋Œ€๋กœ ์ƒ˜ํ”Œ ๊ฐ„์˜ ์œ ์‚ฌ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด ์ˆ˜ํ•™์  ๋™๋“ฑ์„ฑ ๋น„๊ต๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ์ƒ˜ํ”Œ์ด ์ตœ์ข… ๋‹ต๋ณ€์œผ๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค. ์ด ๋‹ต๋ณ€์€ ๊ฒฐ๊ณผ์˜ ์ •ํ™•์„ฑ์„ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•ด ์ •๋‹ต๊ณผ ์ˆ˜ํ•™์  ๋™๋“ฑ์„ฑ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค.
ํ›„์† ํˆฌํ‘œ ๋‹จ๊ณ„์—์„œ๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜ 1์— ์„ค๋ช…๋œ ๋Œ€๋กœ ์ƒ˜ํ”Œ ๊ฐ„์˜ ์œ ์‚ฌ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด ์ˆ˜ํ•™์  ๋“ฑ๊ฐ€์„ฑ ๋น„๊ต๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ์ƒ˜ํ”Œ์ด ์ตœ์ข… ๋‹ต์œผ๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค. ์ด ๋‹ต์€ ๊ฒฐ๊ณผ์˜ ์ •ํ™•์„ฑ์„ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์ค€ ์‹ค์ธก์น˜์™€ ์ˆ˜ํ•™์  ๋“ฑ๊ฐ€์„ฑ ๋น„๊ต๋ฅผ ๊ฑฐ์นฉ๋‹ˆ๋‹ค.
You are an empathetic clinician asking a patient about their medical history over an online chat interface. You know nothing about the patient in advance. Respond to the patient with a single-turn response to better understand their history and symptoms. Do not ask more than two questions. If the patient asks a question, be sure to answer it appropriately.
๋‹น์‹ ์€ ์˜จ๋ผ์ธ ์ฑ„ํŒ… ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ํ†ตํ•ด ํ™˜์ž์˜ ๋ณ‘๋ ฅ์„ ๋ฌป๋Š” ๊ณต๊ฐ์ ์ธ ์ž„์ƒ์˜์ž…๋‹ˆ๋‹ค. ์‚ฌ์ „์— ํ™˜์ž์— ๋Œ€ํ•ด ์•„๋ฌด๊ฒƒ๋„ ๋ชจ๋ฆ…๋‹ˆ๋‹ค. ํ™˜์ž์˜ ๋ณ‘๋ ฅ๊ณผ ์ฆ์ƒ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด ํ•œ ๋ฒˆ์˜ ์‘๋‹ต์œผ๋กœ ๋‹ต๋ณ€ํ•˜์‹ญ์‹œ์˜ค. ๋‘ ๊ฐœ ์ด์ƒ์˜ ์งˆ๋ฌธ์„ ํ•˜์ง€ ๋งˆ์‹ญ์‹œ์˜ค. ํ™˜์ž๊ฐ€ ์งˆ๋ฌธ์„ ํ•˜๋ฉด ์ ์ ˆํ•˜๊ฒŒ ๋‹ต๋ณ€ํ•ด ์ฃผ์‹ญ์‹œ์˜ค.
๊ท€ํ•˜๋Š” ์˜จ๋ผ์ธ ์ฑ„ํŒ… ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ํ†ตํ•ด ํ™˜์ž์˜ ๋ณ‘๋ ฅ์— ๋Œ€ํ•ด ๋ฌป๋Š” ๊ณต๊ฐ ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚œ ์ž„์ƒ์˜์ž…๋‹ˆ๋‹ค. ํ™˜์ž์— ๋Œ€ํ•ด ๋ฏธ๋ฆฌ ์•„๋Š” ๊ฒƒ์ด ์—†์Šต๋‹ˆ๋‹ค. ํ™˜์ž์˜ ๋ณ‘๋ ฅ๊ณผ ์ฆ์ƒ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด ํ•œ ๋ฒˆ์˜ ์‘๋‹ต์œผ๋กœ ํ™˜์ž์—๊ฒŒ ์‘๋‹ตํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๊ฐœ ์ด์ƒ์˜ ์งˆ๋ฌธ์„ ํ•˜์ง€ ๋งˆ์„ธ์š”. ํ™˜์ž๊ฐ€ ์งˆ๋ฌธ์„ ํ•˜๋ฉด ์ ์ ˆํ•˜๊ฒŒ ๋‹ต๋ณ€ํ•˜์„ธ์š”.
The strengths of deontological ethics are that it provides clear and concrete guidelines for decision making, making it a more structured and rule-based approach. The weakness of this theory is that it may not take into account the consequences of following these rules, and that it may be difficult to determine what moral rules are universally applicable.
์˜๋ฌด๋ก ์  ์œค๋ฆฌ์˜ ๊ฐ•์ ์€ ์˜์‚ฌ๊ฒฐ์ •์„ ์œ„ํ•œ ๋ช…ํ™•ํ•˜๊ณ  ๊ตฌ์ฒด์ ์ธ ์ง€์นจ์„ ์ œ๊ณตํ•˜์—ฌ ๋ณด๋‹ค ๊ตฌ์กฐ์ ์ด๊ณ  ๊ทœ์น™ ๊ธฐ๋ฐ˜์˜ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•œ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์ด ์ด๋ก ์˜ ์•ฝ์ ์€ ์ด๋Ÿฌํ•œ ๊ทœ์น™์„ ๋”ฐ๋ฅด๋Š” ๊ฒฐ๊ณผ๋ฅผ ๊ณ ๋ คํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋ณดํŽธ์ ์œผ๋กœ ์ ์šฉ ๊ฐ€๋Šฅํ•œ ๋„๋•์  ๊ทœ์น™์„ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์ด ์–ด๋ ค์šธ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค.
์˜๋ฌด ์œค๋ฆฌ์˜ ๊ฐ•์ ์€ ์˜์‚ฌ ๊ฒฐ์ •์— ๋Œ€ํ•œ ๋ช…ํ™•ํ•˜๊ณ  ๊ตฌ์ฒด์ ์ธ ์ง€์นจ์„ ์ œ๊ณตํ•˜์—ฌ ๋ณด๋‹ค ์ฒด๊ณ„์ ์ด๊ณ  ๊ทœ์น™์— ๊ธฐ๋ฐ˜ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์ด๋ผ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์ด๋ก ์˜ ์•ฝ์ ์€ ์ด๋Ÿฌํ•œ ๊ทœ์น™์„ ๋”ฐ๋ฅด๋Š” ๊ฒฐ๊ณผ๋ฅผ ๊ณ ๋ คํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ ์–ด๋–ค ๋„๋•์  ๊ทœ์น™์ด ๋ณดํŽธ์ ์œผ๋กœ ์ ์šฉ๋˜๋Š”์ง€ ๊ฒฐ์ •ํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
\(\bullet\)_OpenAssistant_[173] is a multilingual corpus containing 66,497 real-world conversation trees between human and AI assistant. Each conversation tree consists of multiple nodes, and each node represents the information generated by a role in the dialogue. It spans 35 languages and includes 461,292 manually annotated quality ratings of responses.
\(\bullet\)_OpenAssistant_[173]๋Š” ์ธ๊ฐ„๊ณผ AI ์–ด์‹œ์Šคํ„ดํŠธ ๊ฐ„์˜ 66,497๊ฐœ์˜ ์‹ค์ œ ๋Œ€ํ™” ํŠธ๋ฆฌ๋ฅผ ํฌํ•จํ•˜๋Š” ๋‹ค๊ตญ์–ด ์ฝ”ํผ์Šค์ž…๋‹ˆ๋‹ค. ๊ฐ ๋Œ€ํ™” ํŠธ๋ฆฌ๋Š” ์—ฌ๋Ÿฌ ๊ฐœ์˜ ๋…ธ๋“œ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ๊ฐ ๋…ธ๋“œ๋Š” ๋Œ€ํ™”์—์„œ ์—ญํ• ์— ์˜ํ•ด ์ƒ์„ฑ๋œ ์ •๋ณด๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ์ฝ”ํผ์Šค๋Š” 35๊ฐœ ์–ธ์–ด์— ๊ฑธ์ณ ์žˆ์œผ๋ฉฐ, ์‘๋‹ต์˜ 461,292๊ฐœ์˜ ์ˆ˜๋™์œผ๋กœ ์ฃผ์„๋œ ํ’ˆ์งˆ ํ‰๊ฐ€๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_OpenAssistant_[173]๋Š” ์ธ๊ฐ„๊ณผ AI ๋น„์„œ ๊ฐ„์˜ ์‹ค์ œ ๋Œ€ํ™” ํŠธ๋ฆฌ 66,497๊ฐœ๊ฐ€ ํฌํ•จ๋œ ๋‹ค๊ตญ์–ด ์ฝ”ํผ์Šค์ž…๋‹ˆ๋‹ค. ๊ฐ ๋Œ€ํ™” ํŠธ๋ฆฌ๋Š” ์—ฌ๋Ÿฌ ๋…ธ๋“œ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ๊ฐ ๋…ธ๋“œ๋Š” ๋Œ€ํ™”์—์„œ ์—ญํ• ์— ์˜ํ•ด ์ƒ์„ฑ๋œ ์ •๋ณด๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. 35๊ฐœ ์–ธ์–ด์— ๊ฑธ์ณ ์žˆ์œผ๋ฉฐ, 461,292๊ฐœ์˜ ์‘๋‹ต์— ๋Œ€ํ•œ ์ˆ˜๋™์œผ๋กœ ์ฃผ์„์ด ๋‹ฌ๋ฆฐ ํ’ˆ์งˆ ๋“ฑ๊ธ‰์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
for all prompts \(x\) and completions \(y\). Then we must have \(\pi(y|x)\exp(\frac{1}{\beta}f(x))=\pi^{\prime}(y|x)\). Since these are distributions, summing over \(y\) on both sides, we obtain that \(\exp(\frac{1}{\beta}f(x))=1\) and since \(\beta>0\), we must have \(f(x)=0\) for all \(x\). Therefore \(r(x,y)=r^{\prime}(x,y)\). This completes the proof.
๋ชจ๋“  ํ”„๋กฌํ”„ํŠธ \(x\)์™€ ์™„์„ฑ \(y\)์— ๋Œ€ํ•ด. ๊ทธ๋Ÿฌ๋ฉด ์šฐ๋ฆฌ๋Š” \(\pi(y|x)\exp(\frac{1}{\beta}f(x))=\pi^{\prime}(y|x)\)๋ฅผ ๊ฐ€์ ธ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๊ฒƒ๋“ค์ด ๋ถ„ํฌ์ด๊ธฐ ๋•Œ๋ฌธ์—, ์–‘์ชฝ์—์„œ \(y\)์— ๋Œ€ํ•ด ํ•ฉ์‚ฐํ•˜๋ฉด, \(\exp(\frac{1}{\beta}f(x))=1\)์„ ์–ป๊ณ , \(\beta>0\)์ด๋ฏ€๋กœ ๋ชจ๋“  \(x\)์— ๋Œ€ํ•ด \(f(x)=0\)์ด์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ \(r(x,y)=r^{\prime}(x,y)\)์ž…๋‹ˆ๋‹ค. ์ด๊ฒƒ์œผ๋กœ ์ฆ๋ช…์ด ์™„๋ฃŒ๋ฉ๋‹ˆ๋‹ค.
๋ชจ๋“  ํ”„๋กฌํ”„ํŠธ \(x\)์™€ ์™„๋ฃŒ \(y\)์— ๋Œ€ํ•ด. ๊ทธ๋Ÿฌ๋ฉด \(\pi(y|x)\exp(\frac{1}{\beta}f(x))=\pi^{\prime}(y|x)\)๊ฐ€ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋ถ„ํฌ์ด๋ฏ€๋กœ ์–‘์ชฝ์—์„œ \(y\)๋ฅผ ํ•ฉํ•˜๋ฉด \(\exp(\frac{1}{\beta}f(x))=1\)์ด ๋˜๊ณ , \(\beta>0\)์ด๋ฏ€๋กœ ๋ชจ๋“  \(x\)์— ๋Œ€ํ•ด \(f(x)=0\)์ด ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ \(r(x,y)=r^{\prime}(x,y)\)๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ์ด๊ฒƒ์œผ๋กœ ์ฆ๋ช…์ด ์™„๋ฃŒ๋ฉ๋‹ˆ๋‹ค.
**Key Techniques for LLMs**. It has been a long way that LLMs evolve into the current state: _general_ and _capable_ learners. In the development process, a number of important techniques are proposed, which largely improve the capacity of LLMs. Here, we briefly list several important techniques that (potentially) lead to the success of LLMs, as follows.
**LLMs์˜ ์ฃผ์š” ๊ธฐ์ˆ **. LLMs๊ฐ€ ํ˜„์žฌ์˜ _์ผ๋ฐ˜์ _์ด๊ณ  _๋Šฅ๋ ฅ ์žˆ๋Š”_ ํ•™์Šต์ž๋กœ ์ง„ํ™”ํ•˜๊ธฐ๊นŒ์ง€๋Š” ์˜ค๋žœ ์‹œ๊ฐ„์ด ๊ฑธ๋ ธ์Šต๋‹ˆ๋‹ค. ๊ฐœ๋ฐœ ๊ณผ์ •์—์„œ LLMs์˜ ์—ญ๋Ÿ‰์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์—ฌ๋Ÿฌ ์ค‘์š”ํ•œ ๊ธฐ์ˆ ๋“ค์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” LLMs์˜ ์„ฑ๊ณต์— (์ž ์žฌ์ ์œผ๋กœ) ๊ธฐ์—ฌํ•œ ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ธฐ์ˆ ๋“ค์„ ๊ฐ„๋žตํžˆ ๋‚˜์—ดํ•ฉ๋‹ˆ๋‹ค.
**LLM์˜ ์ฃผ์š” ๊ธฐ์ˆ **. LLM์ด ํ˜„์žฌ์™€ ๊ฐ™์€ _์ผ๋ฐ˜์ ์ด๊ณ  _์œ ๋Šฅํ•œ ํ•™์Šต์ž๋กœ ๋ฐœ์ „ํ•˜๊ธฐ๊นŒ์ง€๋Š” ๋จผ ๊ธธ์„ ๊ฑธ์–ด์™”์Šต๋‹ˆ๋‹ค. ๊ฐœ๋ฐœ ๊ณผ์ •์—์„œ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ธฐ์ˆ ์ด ์ œ์•ˆ๋˜์–ด LLM์˜ ์—ญ๋Ÿ‰์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—์„œ๋Š” (์ž ์žฌ์ ์œผ๋กœ) LLM์˜ ์„ฑ๊ณต์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ๋Š” ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ธฐ์ˆ ์„ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ฐ„๋žตํ•˜๊ฒŒ ๋‚˜์—ดํ•ฉ๋‹ˆ๋‹ค.
**What do the instances that comprise the dataset represent?** The instances are prompts from seven hazard categories that pose high personal risk, societal risk or considered illegal. The seven categories are: Violent crimes, Non-violent crimes, Sex-related crimes, Child sexual exploitation, Indiscriminate weapons (CBRNE), Suicide & self-harm and Hate.
**๋ฐ์ดํ„ฐ์…‹์„ ๊ตฌ์„ฑํ•˜๋Š” ์ธ์Šคํ„ด์Šค๋Š” ๋ฌด์—‡์„ ๋‚˜ํƒ€๋‚ด๋‚˜์š”?** ์ธ์Šคํ„ด์Šค๋Š” ๊ฐœ์ธ์  ์œ„ํ—˜, ์‚ฌํšŒ์  ์œ„ํ—˜์„ ์ดˆ๋ž˜ํ•˜๊ฑฐ๋‚˜ ๋ถˆ๋ฒ•์œผ๋กœ ๊ฐ„์ฃผ๋˜๋Š” 7๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ์—์„œ ๋‚˜์˜จ ํ”„๋กฌํ”„ํŠธ์ž…๋‹ˆ๋‹ค. ์ด 7๊ฐ€์ง€ ๋ฒ”์ฃผ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค: ํญ๋ ฅ ๋ฒ”์ฃ„, ๋น„ํญ๋ ฅ ๋ฒ”์ฃ„, ์„ฑ ๊ด€๋ จ ๋ฒ”์ฃ„, ์•„๋™ ์„ฑ ์ฐฉ์ทจ, ๋ฌด์ฐจ๋ณ„ ๋ฌด๊ธฐ (CBRNE), ์ž์‚ด ๋ฐ ์žํ•ด, ๊ทธ๋ฆฌ๊ณ  ์ฆ์˜ค.
**๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ๊ตฌ์„ฑํ•˜๋Š” ์ธ์Šคํ„ด์Šค๋Š” ๋ฌด์—‡์„ ๋‚˜ํƒ€๋‚ด๋‚˜์š”?** ์ธ์Šคํ„ด์Šค๋Š” ๊ฐœ์ธ์  ์œ„ํ—˜, ์‚ฌํšŒ์  ์œ„ํ—˜์ด ๋†’๊ฑฐ๋‚˜ ๋ถˆ๋ฒ•์œผ๋กœ ๊ฐ„์ฃผ๋˜๋Š” 7๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ์— ์†ํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ์ž…๋‹ˆ๋‹ค. 7๊ฐ€์ง€ ๋ฒ”์ฃผ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค: ํญ๋ ฅ ๋ฒ”์ฃ„, ๋น„ํญ๋ ฅ ๋ฒ”์ฃ„, ์„ฑ ๊ด€๋ จ ๋ฒ”์ฃ„, ์•„๋™ ์„ฑ์ฐฉ์ทจ, ๋ฌด์ฐจ๋ณ„ ๋ฌด๊ธฐ(CBRNE), ์ž์‚ด ๋ฐ ์žํ•ด, ์ฆ์˜ค.
Language localizationThe hazard categories have been created for English and may not apply equally well across other languages. We have anchored our work on international norms, rather than taking a country-by-country approach, in order for the hazard categories to be broadly applicable across geographies and cultures. However, this needs to be tested.
์–ธ์–ด ํ˜„์ง€ํ™”: ์œ„ํ—˜ ๋ฒ”์ฃผ๋Š” ์˜์–ด๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์ž‘์„ฑ๋˜์—ˆ์œผ๋ฉฐ ๋‹ค๋ฅธ ์–ธ์–ด์— ๋™์ผํ•˜๊ฒŒ ์ ์šฉ๋˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์œ„ํ—˜ ๋ฒ”์ฃผ๊ฐ€ ์ง€๋ฆฌ์  ๋ฐ ๋ฌธํ™”์ ์œผ๋กœ ๋„๋ฆฌ ์ ์šฉ๋  ์ˆ˜ ์žˆ๋„๋ก ๊ตญ๊ฐ€๋ณ„ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์•„๋‹Œ ๊ตญ์ œ ํ‘œ์ค€์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ์ž‘์—…์„ ์ง„ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Š” ํ…Œ์ŠคํŠธ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
์–ธ์–ด ํ˜„์ง€ํ™”์œ„ํ—˜ ๋ฒ”์ฃผ๋Š” ์˜์–ด์šฉ์œผ๋กœ ๋งŒ๋“ค์–ด์กŒ์œผ๋ฉฐ ๋‹ค๋ฅธ ์–ธ์–ด์— ๋™์ผํ•˜๊ฒŒ ์ ์šฉ๋˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ„ํ—˜ ๋ฒ”์ฃผ๊ฐ€ ์ง€๋ฆฌ์ , ๋ฌธํ™”์  ์˜์—ญ์— ๊ฑธ์ณ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ ์šฉ๋  ์ˆ˜ ์žˆ๋„๋ก ๊ตญ๊ฐ€๋ณ„ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ทจํ•˜๋Š” ๋Œ€์‹  ๊ตญ์ œ ํ‘œ์ค€์— ๊ธฐ๋ฐ˜์„ ๋‘๊ณ  ์ž‘์—…์„ ์ง„ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Š” ํ…Œ์ŠคํŠธ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
**Comparing with Advanced Translation Models** Our model, ALMA-13B-R, is benchmarked against the most advanced current models, demonstrating performance comparable to GPT-4 and WMT winners. It surpasses leading commercial translation tools such as Google Translate in many cases and top multilingual translation models like NLLB, MADLAD-10B and GPT-3.5.
**๊ณ ๊ธ‰ ๋ฒˆ์—ญ ๋ชจ๋ธ๊ณผ์˜ ๋น„๊ต** ์šฐ๋ฆฌ ๋ชจ๋ธ์ธ ALMA-13B-R์€ ๊ฐ€์žฅ ์ตœ์‹ ์˜ ๊ณ ๊ธ‰ ๋ชจ๋ธ๋“ค๊ณผ ๋น„๊ตํ•˜์—ฌ ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ๋ฅผ ๊ฑฐ์ณค์œผ๋ฉฐ, GPT-4์™€ WMT ์ˆ˜์ƒ์ž๋“ค๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ๊ฒฝ์šฐ์— Google Translate์™€ ๊ฐ™์€ ์ฃผ์š” ์ƒ์—… ๋ฒˆ์—ญ ๋„๊ตฌ๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ, NLLB, MADLAD-10B, GPT-3.5์™€ ๊ฐ™์€ ์ตœ๊ณ  ์ˆ˜์ค€์˜ ๋‹ค๊ตญ์–ด ๋ฒˆ์—ญ ๋ชจ๋ธ๋“ค์„ ๋›ฐ์–ด๋„˜์Šต๋‹ˆ๋‹ค.
**๊ณ ๊ธ‰ ๋ฒˆ์—ญ ๋ชจ๋ธ๊ณผ ๋น„๊ต** ์ €ํฌ ๋ชจ๋ธ์ธ ALMA-13B-R์€ ํ˜„์žฌ ๊ฐ€์žฅ ์ง„๋ณด๋œ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•˜์—ฌ GPT-4 ๋ฐ WMT ์ˆ˜์ƒ์ž์™€ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋Š” ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ๊ฒฝ์šฐ Google ๋ฒˆ์—ญ๊ณผ ๊ฐ™์€ ์„ ๋„์ ์ธ ์ƒ์šฉ ๋ฒˆ์—ญ ๋„๊ตฌ์™€ NLLB, MADLAD-10B, GPT-3.5์™€ ๊ฐ™์€ ์ตœ๊ณ ์˜ ๋‹ค๊ตญ์–ด ๋ฒˆ์—ญ ๋ชจ๋ธ์„ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
In the following, we describe the results of our extrinsic evaluation of tokenizers. Section 5.1 describes the experimental setup, Section 5.2 presents the downstream performance of the trained models based on the investigated tokenizers, and Section 5.3 analyzes the computational costs associated with each tokenizer when employed in a specific model.
๋‹ค์Œ์—์„œ๋Š” ํ† ํฌ๋‚˜์ด์ €์— ๋Œ€ํ•œ ์™ธ์žฌ์  ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. 5.1์ ˆ์—์„œ๋Š” ์‹คํ—˜ ์„ค์ •์„ ์„ค๋ช…ํ•˜๊ณ , 5.2์ ˆ์—์„œ๋Š” ์กฐ์‚ฌ๋œ ํ† ํฌ๋‚˜์ด์ €๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์„ ์ œ์‹œํ•˜๋ฉฐ, 5.3์ ˆ์—์„œ๋Š” ํŠน์ • ๋ชจ๋ธ์— ์‚ฌ์šฉ๋  ๋•Œ ๊ฐ ํ† ํฌ๋‚˜์ด์ €์™€ ๊ด€๋ จ๋œ ๊ณ„์‚ฐ ๋น„์šฉ์„ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค.
๋‹ค์Œ์—์„œ๋Š” ํ† ํฐํ™”๊ธฐ์— ๋Œ€ํ•œ ์™ธ๋ถ€ ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 5.1์—์„œ๋Š” ์‹คํ—˜ ์„ค์ •์„ ์„ค๋ช…ํ•˜๊ณ , ์„น์…˜ 5.2์—์„œ๋Š” ์กฐ์‚ฌ๋œ ํ† ํฐํ™”๊ธฐ์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์„ ์ œ์‹œํ•˜๋ฉฐ, ์„น์…˜ 5.3์—์„œ๋Š” ํŠน์ • ๋ชจ๋ธ์— ์‚ฌ์šฉํ•  ๋•Œ ๊ฐ ํ† ํฐํ™”๊ธฐ์— ๊ด€๋ จ๋œ ๊ณ„์‚ฐ ๋น„์šฉ์„ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค.
Figure 1: We find MFT has a higher long-distance dependency than SFT. A higher bar indicates greater dependency at the corresponding distance. Specifically. We investigate the interdependency of two numerical tokens within a given sequence. We conduct experiments using the Llama-2-7b model trained on the GSM8K dataset with both the SFT and our method.
Figure 1: ์šฐ๋ฆฌ๋Š” MFT๊ฐ€ SFT๋ณด๋‹ค ๋” ๋†’์€ ์žฅ๊ฑฐ๋ฆฌ ์˜์กด์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ง‰๋Œ€๊ฐ€ ๋†’์„์ˆ˜๋ก ํ•ด๋‹น ๊ฑฐ๋ฆฌ์—์„œ์˜ ์˜์กด์„ฑ์ด ๋” ํฝ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ฃผ์–ด์ง„ ์‹œํ€€์Šค ๋‚ด์—์„œ ๋‘ ๊ฐœ์˜ ์ˆซ์ž ํ† ํฐ ๊ฐ„์˜ ์ƒํ˜ธ ์˜์กด์„ฑ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” SFT์™€ ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ GSM8K ๋ฐ์ดํ„ฐ์…‹์—์„œ ํ›ˆ๋ จ๋œ Llama-2-7b ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1: MFT๊ฐ€ SFT๋ณด๋‹ค ์›๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์ด ๋” ๋†’๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ง‰๋Œ€๊ฐ€ ๋†’์„์ˆ˜๋ก ํ•ด๋‹น ๊ฑฐ๋ฆฌ์—์„œ ์ข…์†์„ฑ์ด ๋” ํฌ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์ฃผ์–ด์ง„ ์‹œํ€€์Šค ๋‚ด์—์„œ ๋‘ ๊ฐœ์˜ ์ˆซ์ž ํ† ํฐ ๊ฐ„์˜ ์ƒํ˜ธ ์˜์กด์„ฑ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. SFT์™€ ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์œผ๋กœ GSM8K ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ•™์Šต๋œ Llama-2-7b ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
Furthermore, our experiments predominantly utilized SOLAR 10.7B models, driven by the state-of-the-art performance at the time of experimentation along with its unique 10.7 billion parameter size. The unique size of SOLAR 10.7B models made the usage of sDPO more necessary as there are far fewer open source LLMs that can be adopted as reference models.
๊ฒŒ๋‹ค๊ฐ€, ์šฐ๋ฆฌ์˜ ์‹คํ—˜์€ ์ฃผ๋กœ SOLAR 10.7B ๋ชจ๋ธ์„ ์‚ฌ์šฉํ–ˆ๋Š”๋ฐ, ์ด๋Š” ์‹คํ—˜ ๋‹น์‹œ์˜ ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ๊ณผ ๊ณ ์œ ํ•œ 107์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ ํฌ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. SOLAR 10.7B ๋ชจ๋ธ์˜ ๋…ํŠนํ•œ ํฌ๊ธฐ๋Š” ์ฐธ์กฐ ๋ชจ๋ธ๋กœ ์ฑ„ํƒํ•  ์ˆ˜ ์žˆ๋Š” ์˜คํ”ˆ ์†Œ์Šค LLM์ด ํ›จ์”ฌ ์ ๊ธฐ ๋•Œ๋ฌธ์— sDPO์˜ ์‚ฌ์šฉ์„ ๋”์šฑ ํ•„์š”ํ•˜๊ฒŒ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค.
๋˜ํ•œ, ์ €ํฌ์˜ ์‹คํ—˜์€ ์ฃผ๋กœ ์‹คํ—˜ ๋‹น์‹œ ์ตœ์‹  ์„ฑ๋Šฅ๊ณผ 107์–ต ๊ฐœ์˜ ๋…ํŠนํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ ํฌ๊ธฐ๋กœ ๊ตฌ๋™๋˜๋Š” SOLAR 10.7B ๋ชจ๋ธ์„ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. SOLAR 10.7B ๋ชจ๋ธ์˜ ๊ณ ์œ ํ•œ ํฌ๊ธฐ๋Š” ์ฐธ์กฐ ๋ชจ๋ธ๋กœ ์ฑ„ํƒํ•  ์ˆ˜ ์žˆ๋Š” ์˜คํ”ˆ ์†Œ์Šค LLM์ด ํ›จ์”ฌ ์ ๊ธฐ ๋•Œ๋ฌธ์— sDPO์˜ ์‚ฌ์šฉ์ด ๋”์šฑ ํ•„์š”ํ–ˆ์Šต๋‹ˆ๋‹ค.
Figure 4: Relative frequency that stronger recall of the bridge entity at the \(l\)-th layer increases the consistency of the LLM. Bars are colored blue if the relative frequency is greater than or equal to 0.5 and red otherwise. We manually set the value of 0.5 at the last layer because the intervention does not affect the consistency at that layer.
Figure 4: \(l\)-๋ฒˆ์งธ ๋ ˆ์ด์–ด์—์„œ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ๊ฐ•ํ•œ ๋ฆฌ์ฝœ์ด LLM์˜ ์ผ๊ด€์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ค๋Š” ์ƒ๋Œ€์  ๋นˆ๋„. ์ƒ๋Œ€์  ๋นˆ๋„๊ฐ€ 0.5 ์ด์ƒ์ด๋ฉด ๋ง‰๋Œ€๋Š” ํŒŒ๋ž€์ƒ‰์œผ๋กœ, ๊ทธ๋ ‡์ง€ ์•Š์œผ๋ฉด ๋นจ๊ฐ„์ƒ‰์œผ๋กœ ์ƒ‰์น ๋ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด์—์„œ๋Š” ๊ฐœ์ž…์ด ์ผ๊ด€์„ฑ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ๊ฐ’์„ 0.5๋กœ ์ˆ˜๋™ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 4: l\)-๋ฒˆ์งธ ๋ ˆ์ด์–ด์—์„œ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ๋ฅผ ๋” ๊ฐ•ํ•˜๊ฒŒ ๋ฆฌ์ฝœํ•˜๋ฉด LLM์˜ ์ผ๊ด€์„ฑ์ด ์ฆ๊ฐ€ํ•˜๋Š” ์ƒ๋Œ€์  ๋นˆ๋„. ์ƒ๋Œ€์  ๋นˆ๋„๊ฐ€ 0.5๋ณด๋‹ค ํฌ๊ฑฐ๋‚˜ ๊ฐ™์œผ๋ฉด ๋ง‰๋Œ€๋Š” ํŒŒ๋ž€์ƒ‰์œผ๋กœ, ๊ทธ๋ ‡์ง€ ์•Š์œผ๋ฉด ๋นจ๊ฐ„์ƒ‰์œผ๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด์—์„œ ๊ฐœ์ž…์ด ์ผ๊ด€์„ฑ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ์ˆ˜๋™์œผ๋กœ 0.5 ๊ฐ’์„ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
Addressing ethical concerns and biases in LLMs is another active area of research. Efforts are being made to ensure that LLMs are fair, unbiased, and capable of handling sensitive information responsibly. As LLMs are being used more and more by a large number of people on a daily basis, making sure they are unbiased and behave responsibly is crucial.
LLMs์˜ ์œค๋ฆฌ์  ๋ฌธ์ œ์™€ ํŽธํ–ฅ์„ฑ์„ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ์€ ๋˜ ๋‹ค๋ฅธ ํ™œ๋ฐœํ•œ ์—ฐ๊ตฌ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. LLMs๊ฐ€ ๊ณต์ •ํ•˜๊ณ  ํŽธํ–ฅ๋˜์ง€ ์•Š์œผ๋ฉฐ ๋ฏผ๊ฐํ•œ ์ •๋ณด๋ฅผ ์ฑ…์ž„๊ฐ ์žˆ๊ฒŒ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๊ธฐ ์œ„ํ•œ ๋…ธ๋ ฅ์ด ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLMs๊ฐ€ ์ผ์ƒ์ ์œผ๋กœ ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์— ์˜ํ•ด ์ ์  ๋” ๋งŽ์ด ์‚ฌ์šฉ๋จ์— ๋”ฐ๋ผ, ์ด๋“ค์ด ํŽธํ–ฅ๋˜์ง€ ์•Š๊ณ  ์ฑ…์ž„๊ฐ ์žˆ๊ฒŒ ํ–‰๋™ํ•˜๋„๋ก ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
LLM์˜ ์œค๋ฆฌ์  ์šฐ๋ ค์™€ ํŽธ๊ฒฌ์„ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ๋„ ํ™œ๋ฐœํ•œ ์—ฐ๊ตฌ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. LLM์ด ๊ณต์ •ํ•˜๊ณ  ํŽธ๊ฒฌ์ด ์—†์œผ๋ฉฐ ๋ฏผ๊ฐํ•œ ์ •๋ณด๋ฅผ ์ฑ…์ž„๊ฐ ์žˆ๊ฒŒ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋„๋ก ๋…ธ๋ ฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLM์ด ์ ์  ๋” ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์ด ๋งค์ผ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ํŽธ๊ฒฌ์ด ์—†๊ณ  ์ฑ…์ž„๊ฐ ์žˆ๊ฒŒ ์ž‘๋™ํ•˜๋Š”์ง€ ํ™•์ธํ•˜๋Š” ๊ฒƒ์ด ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
Figure 11: **Model Collapse for Hutter++.** Empirical confirmation of Theorem 4.2. Here \(p(j\mid i)\) is as in (16), with \(\beta=7/5\). The horizontal broken lines correspond to \(k^{-\beta c}\) for different values of \(k\), where \(c:=\min(1-1/\beta,1/2)\). The diagonal broken line corresponds to \(T^{-c}\) (classical error rate without cutoff).
Figure 11: **Hutter++์˜ ๋ชจ๋ธ ๋ถ•๊ดด.** ์ •๋ฆฌ 4.2์˜ ์‹ค์ฆ์  ํ™•์ธ. ์—ฌ๊ธฐ์„œ \(p(j\mid i)\)๋Š” (16)๊ณผ ๊ฐ™์œผ๋ฉฐ, \(\beta=7/5\)์ž…๋‹ˆ๋‹ค. ์ˆ˜ํ‰ ์ ์„ ์€ \(k\)์˜ ๋‹ค๋ฅธ ๊ฐ’์— ๋Œ€ํ•ด \(k^{-\beta c}\)์— ํ•ด๋‹นํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(c:=\min(1-1/\beta,1/2)\)์ž…๋‹ˆ๋‹ค. ๋Œ€๊ฐ์„  ์ ์„ ์€ \(T^{-c}\)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค (์ปท์˜คํ”„ ์—†๋Š” ๊ณ ์ „์  ์˜ค๋ฅ˜์œจ).
๊ทธ๋ฆผ 11: **Hutter++์— ๋Œ€ํ•œ ๋ชจ๋ธ ๋ถ•๊ดด.** ์ •๋ฆฌ 4.2์˜ ๊ฒฝํ—˜์  ํ™•์ธ. ์—ฌ๊ธฐ์„œ \(p(j\mid i)\)๋Š” (16)์—์„œ์™€ ๊ฐ™์œผ๋ฉฐ, \(\beta=7/5\)์ž…๋‹ˆ๋‹ค. ๊ฐ€๋กœ ๋Š์–ด์ง„ ์„ ์€ \(k^{-\beta c}\)์— ํ•ด๋‹นํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(c:=\min(1-1/\beta,1/2)\)์ž…๋‹ˆ๋‹ค. ๋Œ€๊ฐ์„  ๋Š์–ด์ง„ ์„ ์€ \(T^{-c}\)(์ ˆ๋‹จ ์—†์ด ๊ณ ์ „์ ์ธ ์˜ค๋ฅ˜์œจ)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
The Benchmark tests whether the SUTs' responses contain hazards. In principle, SUTs can give hazardous responses to any prompt. However, in practice, such responses are much more likely when prompts invite, or at least relate to, hazards. With this in mind, we aim to test a range of prompts (see below), which elicit different responses from models.
๋ฒค์น˜๋งˆํฌ๋Š” SUTs์˜ ์‘๋‹ต์— ์œ„ํ—˜ ์š”์†Œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š”์ง€ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ์›์น™์ ์œผ๋กœ SUTs๋Š” ์–ด๋–ค ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด์„œ๋„ ์œ„ํ—˜ํ•œ ์‘๋‹ต์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ๋กœ๋Š” ํ”„๋กฌํ”„ํŠธ๊ฐ€ ์œ„ํ—˜์„ ์ดˆ๋Œ€ํ•˜๊ฑฐ๋‚˜ ์ ์–ด๋„ ๊ด€๋ จ์ด ์žˆ์„ ๋•Œ ๊ทธ๋Ÿฌํ•œ ์‘๋‹ต์ด ํ›จ์”ฌ ๋” ๋งŽ์ด ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์—ผ๋‘์— ๋‘๊ณ , ์šฐ๋ฆฌ๋Š” ๋ชจ๋ธ๋กœ๋ถ€ํ„ฐ ๋‹ค์–‘ํ•œ ์‘๋‹ต์„ ์ด๋Œ์–ด๋‚ด๋Š” ์—ฌ๋Ÿฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ…Œ์ŠคํŠธํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค(์•„๋ž˜ ์ฐธ์กฐ).
๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ๋Š” SUT์˜ ์‘๋‹ต์— ์œ„ํ—˜์ด ํฌํ•จ๋˜์–ด ์žˆ๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ์›์น™์ ์œผ๋กœ SUT๋Š” ๋ชจ๋“  ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด ์œ„ํ—˜ํ•œ ์‘๋‹ต์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ๋กœ๋Š” ํ”„๋กฌํ”„ํŠธ๊ฐ€ ์œ„ํ—˜์„ ์ดˆ๋ž˜ํ•˜๊ฑฐ๋‚˜ ์ ์–ด๋„ ์œ„ํ—˜๊ณผ ๊ด€๋ จ์ด ์žˆ์„ ๋•Œ ๊ทธ๋Ÿฌํ•œ ์‘๋‹ต์ด ํ›จ์”ฌ ๋” ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ์—ผ๋‘์— ๋‘๊ณ  ๋‹ค์–‘ํ•œ ์‘๋‹ต์„ ์ด๋Œ์–ด๋‚ด๋Š” ๋‹ค์–‘ํ•œ ํ”„๋กฌํ”„ํŠธ(์•„๋ž˜ ์ฐธ์กฐ)๋ฅผ ํ…Œ์ŠคํŠธํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
Formally, given input \(x\), we train \(\mathcal{M}\) to sequentially generate textual outputs \(y\) consisting of multiple segments \(y=[y_{1},\ldots,y_{T}]\), where \(y_{t}\) indicates a sequence of tokens for the \(t\)-th segment.3 Generated tokens in \(y_{t}\) include text from the original vocabulary as well as the reflection tokens (Table 1).
ํ˜•์‹์ ์œผ๋กœ, ์ž…๋ ฅ \(x\)๊ฐ€ ์ฃผ์–ด์ง€๋ฉด, ์šฐ๋ฆฌ๋Š” \(\mathcal{M}\)์„ ์—ฌ๋Ÿฌ ์„ธ๊ทธ๋จผํŠธ \(y=[y_{1},\ldots,y_{T}]\)๋กœ ๊ตฌ์„ฑ๋œ ํ…์ŠคํŠธ ์ถœ๋ ฅ์„ ์ˆœ์ฐจ์ ์œผ๋กœ ์ƒ์„ฑํ•˜๋„๋ก ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ \(y_{t}\)๋Š” \(t\)-๋ฒˆ์งธ ์„ธ๊ทธ๋จผํŠธ์— ๋Œ€ํ•œ ํ† ํฐ ์‹œํ€€์Šค๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. \(y_{t}\)์—์„œ ์ƒ์„ฑ๋œ ํ† ํฐ์€ ์›๋ž˜ ์–ดํœ˜์˜ ํ…์ŠคํŠธ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋ฐ˜์˜ ํ† ํฐ๋„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค (ํ‘œ 1 ์ฐธ์กฐ).
๊ณต์‹์ ์œผ๋กœ, ์ž…๋ ฅ \(x\)๊ฐ€ ์ฃผ์–ด์ง€๋ฉด, ์šฐ๋ฆฌ๋Š” \(\mathcal{M}\)์ด ์—ฌ๋Ÿฌ ์„ธ๊ทธ๋จผํŠธ \(y=[y_{1},\ldots,y_{T}]\)๋กœ ๊ตฌ์„ฑ๋œ ํ…์ŠคํŠธ ์ถœ๋ ฅ \(y\)๋ฅผ ์ˆœ์ฐจ์ ์œผ๋กœ ์ƒ์„ฑํ•˜๋„๋ก ํ›ˆ๋ จํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(y_{t}\)๋Š” \(t\)๋ฒˆ์งธ ์„ธ๊ทธ๋จผํŠธ์— ๋Œ€ํ•œ ํ† ํฐ ์‹œํ€€์Šค๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.3 y_{t}\)์—์„œ ์ƒ์„ฑ๋œ ํ† ํฐ์—๋Š” ์›๋ž˜ ์–ดํœ˜์˜ ํ…์ŠคํŠธ์™€ ๋ฐ˜์‚ฌ ํ† ํฐ(ํ‘œ 1)์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
In LLMs, MoEs allow to train an extremely large model than is then only partially instantiated during inference when some of the experts are turned off wherever the gating/weighting function has a low weight assigned to them. As an example, the GLaM model has 1.2 trillion parameters, but during inference only 2 out of the 64 experts are used [84].
LLMs์—์„œ MoEs๋Š” ์ผ๋ถ€ ์ „๋ฌธ๊ฐ€๊ฐ€ ๋น„ํ™œ์„ฑํ™”๋  ๋•Œ ๊ฒŒ์ดํŒ…/๊ฐ€์ค‘ ํ•จ์ˆ˜์— ๋‚ฎ์€ ๊ฐ€์ค‘์น˜๊ฐ€ ํ• ๋‹น๋˜๋Š” ๊ฒฝ์šฐ ์ถ”๋ก  ์ค‘์—๋งŒ ๋ถ€๋ถ„์ ์œผ๋กœ ์ธ์Šคํ„ด์Šคํ™”๋˜๋Š” ๋งค์šฐ ํฐ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GLaM ๋ชจ๋ธ์€ 1.2์กฐ ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์ง€๋งŒ, ์ถ”๋ก  ์ค‘์—๋Š” 64๋ช…์˜ ์ „๋ฌธ๊ฐ€ ์ค‘ 2๋ช…๋งŒ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค [84].
LLM์—์„œ MoE๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋งค์šฐ ํฐ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•œ ๋‹ค์Œ ์ถ”๋ก  ์ค‘์— ์ผ๋ถ€ ์ „๋ฌธ๊ฐ€๊ฐ€ ๊บผ์ง€๋ฉด ๊ฒŒ์ดํŒ…/๊ฐ€์ค‘์น˜ ํ•จ์ˆ˜์— ํ• ๋‹น๋œ ๊ฐ€์ค‘์น˜๊ฐ€ ๋‚ฎ์€ ๊ฒฝ์šฐ ์ผ๋ถ€ ์ „๋ฌธ๊ฐ€๋งŒ ์ธ์Šคํ„ด์Šคํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GLaM ๋ชจ๋ธ์€ 1์กฐ 2,000์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์ง€๋งŒ ์ถ”๋ก  ์ค‘์—๋Š” 64๊ฐœ์˜ ์ „๋ฌธ๊ฐ€ ์ค‘ 2๊ฐœ๋งŒ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค[84].
LLMs face challenges in providing accurate responses directly for intricate problems, necessitating the need to decompose them into sequential steps and sub-tasks. While CoT offers a straightforward approach to planning, it falls short in addressing highly complex problems and lacks the ability to evaluate and rectify errors through backtracking.
LLMs๋Š” ๋ณต์žกํ•œ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ง์ ‘์ ์œผ๋กœ ์ •ํ™•ํ•œ ์‘๋‹ต์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์œผ๋ฉฐ, ์ด๋ฅผ ์ˆœ์ฐจ์ ์ธ ๋‹จ๊ณ„์™€ ํ•˜์œ„ ์ž‘์—…์œผ๋กœ ๋ถ„ํ•ดํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. CoT๋Š” ๊ณ„ํš์„ ์„ธ์šฐ๋Š” ๋ฐ ์žˆ์–ด ๊ฐ„๋‹จํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ๊ณตํ•˜์ง€๋งŒ, ๋งค์šฐ ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ๋Š” ๋ถ€์กฑํ•˜๋ฉฐ, ๋ฐฑํŠธ๋ž˜ํ‚น์„ ํ†ตํ•ด ์˜ค๋ฅ˜๋ฅผ ํ‰๊ฐ€ํ•˜๊ณ  ์ˆ˜์ •ํ•˜๋Š” ๋Šฅ๋ ฅ์ด ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค.
LM์€ ๋ณต์žกํ•œ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ง์ ‘ ์ •ํ™•ํ•œ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๊ธฐ ๋•Œ๋ฌธ์— ๋ฌธ์ œ๋ฅผ ์ˆœ์ฐจ์ ์ธ ๋‹จ๊ณ„์™€ ํ•˜์œ„ ์ž‘์—…์œผ๋กœ ๋ถ„ํ•ดํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. CoT๋Š” ๊ณ„ํš์„ ์„ธ์šฐ๋Š” ๊ฐ„๋‹จํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ๊ณตํ•˜์ง€๋งŒ, ๋งค์šฐ ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ๋Š” ๋ถ€์กฑํ•˜๊ณ  ์—ญ์ถ”์ ์„ ํ†ตํ•ด ์˜ค๋ฅ˜๋ฅผ ํ‰๊ฐ€ํ•˜๊ณ  ์ˆ˜์ •ํ•˜๋Š” ๊ธฐ๋Šฅ์ด ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค.
We use "Answer the following question with just a number from context or just the math arithmetic using +, -, *, or /." for ConvFinQA, since this dataset requires the model to either extract numbers from the context or do arithmetic calculation. We extract the arithmetic formula generated by the model and use a calculator to get its final result.
ConvFinQA์˜ ๊ฒฝ์šฐ, "๋ฌธ๋งฅ์—์„œ ์ˆซ์ž๋งŒ ์ถ”์ถœํ•˜๊ฑฐ๋‚˜ +, -, *, /๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฐ์ˆ  ๊ณ„์‚ฐ๋งŒ ํ•˜์—ฌ ๋‹ค์Œ ์งˆ๋ฌธ์— ๋‹ตํ•˜์‹ญ์‹œ์˜ค."๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ์…‹์€ ๋ชจ๋ธ์ด ๋ฌธ๋งฅ์—์„œ ์ˆซ์ž๋ฅผ ์ถ”์ถœํ•˜๊ฑฐ๋‚˜ ์‚ฐ์ˆ  ๊ณ„์‚ฐ์„ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ชจ๋ธ์ด ์ƒ์„ฑํ•œ ์‚ฐ์ˆ  ๊ณต์‹์„ ์ถ”์ถœํ•˜๊ณ  ๊ณ„์‚ฐ๊ธฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข… ๊ฒฐ๊ณผ๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค.
์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋Š” ๋ชจ๋ธ์ด ๋ฌธ๋งฅ์—์„œ ์ˆซ์ž๋ฅผ ์ถ”์ถœํ•˜๊ฑฐ๋‚˜ ์‚ฐ์ˆ  ๊ณ„์‚ฐ์„ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•˜๋ฏ€๋กœ "๋ฌธ๋งฅ์—์„œ ์ˆซ์ž๋งŒ ๋‹ตํ•˜๊ฑฐ๋‚˜ +, -, *, /๋ฅผ ์‚ฌ์šฉํ•œ ์ˆ˜ํ•™ ์‚ฐ์ˆ ๋งŒ ๋‹ตํ•˜์„ธ์š”."๋ฅผ ConvFinQA์— ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ ์‚ฐ์ˆ  ๊ณต์‹์„ ์ถ”์ถœํ•˜๊ณ  ๊ณ„์‚ฐ๊ธฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข… ๊ฒฐ๊ณผ๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค.
Figure 2: **The performance-vs-loss curves of smaller models pre-trained with different numbers of training tokens**. Each data point is the loss (\(x\)-axis) and performance (\(y\)-axis) of the final checkpoint of one model, i.e., each point corresponds to one model trained from scratch. We mark the results of random guess in black dashed lines.
Figure 2: **๋‹ค์–‘ํ•œ ์ˆ˜์˜ ํ•™์Šต ํ† ํฐ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ์ž‘์€ ๋ชจ๋ธ๋“ค์˜ ์„ฑ๋Šฅ ๋Œ€ ์†์‹ค ๊ณก์„ **. ๊ฐ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋Š” ํ•˜๋‚˜์˜ ๋ชจ๋ธ์˜ ์ตœ์ข… ์ฒดํฌํฌ์ธํŠธ์˜ ์†์‹ค (\(x\)-์ถ•)๊ณผ ์„ฑ๋Šฅ (\(y\)-์ถ•)์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ฆ‰ ๊ฐ ํฌ์ธํŠธ๋Š” ์ฒ˜์Œ๋ถ€ํ„ฐ ํ•™์Šต๋œ ํ•˜๋‚˜์˜ ๋ชจ๋ธ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ฌด์ž‘์œ„ ์ถ”์ธก์˜ ๊ฒฐ๊ณผ๋ฅผ ๊ฒ€์€์ƒ‰ ์ ์„ ์œผ๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2: **๋‹ค์–‘ํ•œ ์ˆ˜์˜ ํ•™์Šต ํ† ํฐ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ์†Œํ˜• ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ๋Œ€ ์†์‹ค ๊ณก์„ **. ๊ฐ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋Š” ํ•˜๋‚˜์˜ ๋ชจ๋ธ์˜ ์ตœ์ข… ์ฒดํฌํฌ์ธํŠธ์˜ ์†์‹ค(\(x\)-์ถ•)๊ณผ ์„ฑ๋Šฅ(\(y\)-์ถ•)์ด๋ฉฐ, ์ฆ‰ ๊ฐ ํฌ์ธํŠธ๋Š” ์ฒ˜์Œ๋ถ€ํ„ฐ ํ›ˆ๋ จ๋œ ํ•˜๋‚˜์˜ ๋ชจ๋ธ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ๋ฌด์ž‘์œ„ ์ถ”์ธก์˜ ๊ฒฐ๊ณผ๋ฅผ ๊ฒ€์€์ƒ‰ ์ ์„ ์œผ๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.
ResultsTable 6 summarizes the result of zero-shot evaluation of OLMo-7B and compares it against 6 other publicly available models of comparable size. We report results on 8 core tasks from our evaluation suite described in Section 2.4. On aggregate, OLMo-7B is competitive against all 6 publicly available model checkpoints in our comparison table.
๊ฒฐ๊ณผ Table 6์€ OLMo-7B์˜ ์ œ๋กœ์ƒท ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์š”์•ฝํ•˜๊ณ , ์œ ์‚ฌํ•œ ํฌ๊ธฐ์˜ 6๊ฐœ์˜ ๋‹ค๋ฅธ ๊ณต๊ฐœ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” 2.4์ ˆ์—์„œ ์„ค๋ช…ํ•œ ํ‰๊ฐ€ ์Šค์œ„ํŠธ์˜ 8๊ฐ€์ง€ ํ•ต์‹ฌ ์ž‘์—…์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ข…ํ•ฉ์ ์œผ๋กœ, OLMo-7B๋Š” ๋น„๊ต ํ‘œ์— ์žˆ๋Š” ๋ชจ๋“  6๊ฐœ์˜ ๊ณต๊ฐœ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ์™€ ๊ฒฝ์Ÿ๋ ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฒฐ๊ณผํ‘œ 6์€ OLMo-7B์˜ ์ œ๋กœ ์ƒท ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์š”์•ฝํ•˜๊ณ  ๋น„์Šทํ•œ ํฌ๊ธฐ์˜ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋‹ค๋ฅธ 6๊ฐœ์˜ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. 2.4์ ˆ์—์„œ ์„ค๋ช…ํ•œ ํ‰๊ฐ€ ์Šค์œ„ํŠธ์˜ 8๊ฐ€์ง€ ํ•ต์‹ฌ ์ž‘์—…์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด์ ์œผ๋กœ OLMo-7B๋Š” ๋น„๊ต ํ‘œ์˜ 6๊ฐœ์˜ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ์™€ ๊ฒฝ์Ÿ๋ ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
We report the 8 largest duplicate clusters found by MinHash in Table 15 - each spanning hundreds of thousands of documents. We also found a large number of duplicate document pairs to be due to different URL GET parameters not resulting in significantly different content. An example of this behaviour can be seen in the URLs presented in Table 16.
์šฐ๋ฆฌ๋Š” MinHash์— ์˜ํ•ด ๋ฐœ๊ฒฌ๋œ 8๊ฐœ์˜ ๊ฐ€์žฅ ํฐ ์ค‘๋ณต ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ Table 15์— ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ํด๋Ÿฌ์Šคํ„ฐ๋Š” ์ˆ˜์‹ญ๋งŒ ๊ฐœ์˜ ๋ฌธ์„œ๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์„œ๋กœ ๋‹ค๋ฅธ URL GET ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ ๋‚ด์šฉ์— ํฐ ์ฐจ์ด๋ฅผ ์ฃผ์ง€ ์•Š๋Š” ๊ฒฝ์šฐ๋กœ ์ธํ•ด ๋งŽ์€ ์ค‘๋ณต ๋ฌธ์„œ ์Œ์ด ๋ฐœ์ƒํ•˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋™์ž‘์˜ ์˜ˆ๋Š” Table 16์— ์ œ์‹œ๋œ URL์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ‘œ 15์—์„œ๋Š” MinHash์— ์˜ํ•ด ๋ฐœ๊ฒฌ๋œ 8๊ฐœ์˜ ๊ฐ€์žฅ ํฐ ์ค‘๋ณต ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ์ˆ˜์‹ญ๋งŒ ๊ฐœ์˜ ๋ฌธ์„œ์— ๊ฑธ์ณ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ URL GET ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ ํฌ๊ฒŒ ๋‹ค๋ฅธ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•˜์ง€ ์•Š๋Š” ๊ฒƒ์œผ๋กœ ์ธํ•ด ๋งŽ์€ ์ˆ˜์˜ ์ค‘๋ณต ๋ฌธ์„œ ์Œ์ด ๋ฐœ๊ฒฌ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋™์ž‘์˜ ์˜ˆ๋Š” ํ‘œ 16์— ์ œ์‹œ๋œ URL์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Our findings indicate that with just \(10\) minutes of fine-tuning on \(200\) contexts, LLMs exhibit significant enhancements in multilingual understanding and generation abilities. Notably, there is a relative performance boost of \(7.4\%\) on the XQuAD benchmark. Similarly, for XLSum, we observe an \(8.9\%\) relative improvement in performance.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด, ๋‹จ \(10\)๋ถ„ ๋™์•ˆ \(200\)๊ฐœ์˜ ์ปจํ…์ŠคํŠธ๋กœ ๋ฏธ์„ธ ์กฐ์ •์„ ์ˆ˜ํ–‰ํ•œ ํ›„, LLMs๋Š” ๋‹ค๊ตญ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์—์„œ ์ƒ๋‹นํ•œ ํ–ฅ์ƒ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, XQuAD ๋ฒค์น˜๋งˆํฌ์—์„œ \(7.4\%\)์˜ ์ƒ๋Œ€์  ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, XLSum์—์„œ๋Š” \(8.9\%\)์˜ ์ƒ๋Œ€์  ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด 200๊ฐœ์˜ ์ปจํ…์ŠคํŠธ์—์„œ 10๋ถ„๋งŒ ํˆฌ์žํ•ด๋„ LLM์€ ๋‹ค๊ตญ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ XQuAD ๋ฒค์น˜๋งˆํฌ์—์„œ ์ƒ๋Œ€์ ์œผ๋กœ 7.4%์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ XLSum์˜ ๊ฒฝ์šฐ, ์„ฑ๋Šฅ์ด ์ƒ๋Œ€์ ์œผ๋กœ 8.9% ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
We evaluate our models on both Korean and English LLM benchmarks, to highlight the advantages of our vocabulary expansion method, which could efficiently leverage the strong multilingual capabilities of base foundational models. Desirably, we expect a model to show improved performance in Korean tasks and comparable performance in English tasks.
์šฐ๋ฆฌ๋Š” ๊ธฐ๋ณธ ๊ธฐ์ดˆ ๋ชจ๋ธ์˜ ๊ฐ•๋ ฅํ•œ ๋‹ค๊ตญ์–ด ๊ธฐ๋Šฅ์„ ํšจ์œจ์ ์œผ๋กœ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์–ดํœ˜ ํ™•์žฅ ๋ฐฉ๋ฒ•์˜ ์žฅ์ ์„ ๊ฐ•์กฐํ•˜๊ธฐ ์œ„ํ•ด, ํ•œ๊ตญ์–ด์™€ ์˜์–ด LLM ๋ฒค์น˜๋งˆํฌ ๋ชจ๋‘์—์„œ ์šฐ๋ฆฌ์˜ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ”๋žŒ์งํ•˜๊ฒŒ๋Š”, ๋ชจ๋ธ์ด ํ•œ๊ตญ์–ด ์ž‘์—…์—์„œ ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ๋ณด์ด๊ณ  ์˜์–ด ์ž‘์—…์—์„œ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๊ธฐ๋ฅผ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค.
์šฐ๋ฆฌ๋Š” ๊ธฐ๋ณธ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์˜ ๊ฐ•๋ ฅํ•œ ๋‹ค๊ตญ์–ด ๊ธฐ๋Šฅ์„ ํšจ์œจ์ ์œผ๋กœ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์–ดํœ˜ ํ™•์žฅ ๋ฐฉ๋ฒ•์˜ ์žฅ์ ์„ ๊ฐ•์กฐํ•˜๊ธฐ ์œ„ํ•ด ํ•œ๊ตญ์–ด์™€ ์˜์–ด LLM ๋ฒค์น˜๋งˆํฌ ๋ชจ๋‘์—์„œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ”๋žŒ์งํ•˜๊ฒŒ๋Š”, ์šฐ๋ฆฌ๋Š” ๋ชจ๋ธ์ด ํ•œ๊ตญ์–ด ์ž‘์—…์—์„œ ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ๋ณด์ด๊ณ  ์˜์–ด ์ž‘์—…์—์„œ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ผ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค.
For individuals working with non-English languages, creating a LLM from scratch can be computationally intensive. As a result, many turn to adopt open-source, English-based pre-trained LLMs, such as BLOOM (Workshop et al., 2023), LLaMA2 (Touvron et al., 2023), and Mistral-7B (Jiang et al., 2023), as foundational models. Inspired by Ouyang et al.
๋น„์˜์–ด๊ถŒ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์‚ฌ๋žŒ๋“ค์—๊ฒŒ๋Š” LLM์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ๊ณ„์‚ฐ์ ์œผ๋กœ ๋งค์šฐ ๋ถ€๋‹ด๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์ด BLOOM (Workshop et al., 2023), LLaMA2 (Touvron et al., 2023), Mistral-7B (Jiang et al., 2023)์™€ ๊ฐ™์€ ์˜คํ”ˆ ์†Œ์Šค, ์˜์–ด ๊ธฐ๋ฐ˜์˜ ์‚ฌ์ „ ํ•™์Šต๋œ LLM์„ ๊ธฐ๋ณธ ๋ชจ๋ธ๋กœ ์ฑ„ํƒํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. Ouyang et al.์˜ ์—ฐ๊ตฌ์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„...
์˜์–ด๊ฐ€ ์•„๋‹Œ ์–ธ์–ด๋กœ ์ž‘์—…ํ•˜๋Š” ๊ฐœ์ธ์˜ ๊ฒฝ์šฐ, ์ฒ˜์Œ๋ถ€ํ„ฐ LLM์„ ๋งŒ๋“œ๋Š” ๊ฒƒ์€ ๊ณ„์‚ฐ ์ง‘์•ฝ์ ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์ด BLOOM(Workshop et al., 2023), LLaMA2(Touvron et al., 2023), Mistral-7B(Jiang et al., 2023)์™€ ๊ฐ™์€ ์˜คํ”ˆ ์†Œ์Šค ์˜์–ด ๊ธฐ๋ฐ˜ ์‚ฌ์ „ ํ•™์Šต๋œ LLM์„ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๋กœ ์ฑ„ํƒํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ „ํ™˜ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Ouyang et al.์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„
where \(D\) is the preference dataset, \(x\) is the question, \(y_{w}\) and \(y_{l}\) are the chosen and rejected answers respectively, \(\theta\) is the learnable parameters of the model, and \(\gamma_{\pi}(x,y_{w},y_{l})=\log\frac{\pi(y_{w}|x)}{\pi(y_{l}|x)}\), _i.e.,_ the logratio of the chosen and rejected samples w.r.t. the policy \(\pi\).
์—ฌ๊ธฐ์„œ \(D\)๋Š” ์„ ํ˜ธ ๋ฐ์ดํ„ฐ์…‹, \(x\)๋Š” ์งˆ๋ฌธ, \(y_{w}\)์™€ \(y_{l}\)๋Š” ๊ฐ๊ฐ ์„ ํƒ๋œ ๋‹ต๋ณ€๊ณผ ๊ฑฐ๋ถ€๋œ ๋‹ต๋ณ€, \(\theta\)๋Š” ๋ชจ๋ธ์˜ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜, \(\gamma_{\pi}(x,y_{w},y_{l})=\log\frac{\pi(y_{w}|x)}{\pi(y_{l}|x)}\)๋Š” ์ •์ฑ… \(\pi\)์— ๋Œ€ํ•œ ์„ ํƒ๋œ ์ƒ˜ํ”Œ๊ณผ ๊ฑฐ๋ถ€๋œ ์ƒ˜ํ”Œ์˜ ๋กœ๊ทธ ๋น„์œจ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(D\)๋Š” ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ, \(x\)๋Š” ์งˆ๋ฌธ, \(y_{w}\)์™€ \(y_{l}\)๋Š” ๊ฐ๊ฐ ์„ ํƒ๋œ ๋‹ต๊ณผ ๊ฑฐ๋ถ€๋œ ๋‹ต, \(\theta\)๋Š” ๋ชจ๋ธ์˜ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ, \(\gamma_{\pi}(x,y_{w},y_{l})=\log\frac{\pi(y_{w}|x)}{\pi(y_{l}|x)}\), ์ฆ‰, ์ •์ฑ… \(\pi\)์— ๋Œ€ํ•œ ์„ ํƒ๋œ ์ƒ˜ํ”Œ๊ณผ ๊ฑฐ๋ถ€๋œ ์ƒ˜ํ”Œ์˜ ๋กœ๊ทธ๋น„์ž…๋‹ˆ๋‹ค.
Note that this algorithm still requires that the dataset itself fits in memory (so that we can efficiently index in arbitrary positions), but we do not need to fit the entire suffix array into memory. This is fortunate since our suffix array requires an \(8\times\) space overhead. For example, the suffix array for the \(350\)GB C4 is \(1.5\)TB.
์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์—ฌ์ „ํžˆ ๋ฐ์ดํ„ฐ์…‹ ์ž์ฒด๊ฐ€ ๋ฉ”๋ชจ๋ฆฌ์— ๋งž์•„์•ผ ํ•œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค (์ž„์˜์˜ ์œ„์น˜์—์„œ ํšจ์œจ์ ์œผ๋กœ ์ธ๋ฑ์‹ฑํ•  ์ˆ˜ ์žˆ๋„๋ก). ํ•˜์ง€๋งŒ ์ „์ฒด ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด์„ ๋ฉ”๋ชจ๋ฆฌ์— ๋งž์ถœ ํ•„์š”๋Š” ์—†์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด์ด \(8\times\) ๊ณต๊ฐ„ ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ํ•„์š”๋กœ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋‹คํ–‰์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(350\)GB C4์˜ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด์€ \(1.5\)TB์ž…๋‹ˆ๋‹ค.
์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์—ฌ์ „ํžˆ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ ์ž์ฒด๊ฐ€ ๋ฉ”๋ชจ๋ฆฌ์— ๋งž์•„์•ผ ํ•˜๋ฏ€๋กœ(์ž„์˜์˜ ์œ„์น˜์—์„œ ํšจ์œจ์ ์œผ๋กœ ์ธ๋ฑ์‹ฑํ•  ์ˆ˜ ์žˆ๋„๋ก) ์ „์ฒด ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด์„ ๋ฉ”๋ชจ๋ฆฌ์— ๋งž์ถœ ํ•„์š”๋Š” ์—†์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด์— \(8\๋ฐฐ\) ๊ณต๊ฐ„ ์˜ค๋ฒ„ํ—ค๋“œ๊ฐ€ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋‹คํ–‰์Šค๋Ÿฌ์šด ์ผ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(350\)GB C4์˜ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด์€ \(1.5\)TB์ž…๋‹ˆ๋‹ค.
SUTs were evaluated at low temperatureThis reduces the variability of SUTs' responses on repeated prompting with the same test item, which makes our results more reproducible. However, SUTs may give a higher proportion of unsafe responses at a higher temperature. We will address this in the future by testing each SUT at different temperatures.
SUTs๋Š” ๋‚ฎ์€ ์˜จ๋„์—์„œ ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋™์ผํ•œ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์œผ๋กœ ๋ฐ˜๋ณต์ ์ธ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ•  ๋•Œ SUTs์˜ ์‘๋‹ต ๋ณ€๋™์„ฑ์„ ์ค„์—ฌ, ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋ฅผ ๋” ์žฌํ˜„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ SUTs๋Š” ๋” ๋†’์€ ์˜จ๋„์—์„œ ๋” ๋งŽ์€ ๋น„์•ˆ์ „ํ•œ ์‘๋‹ต์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๊ฐ SUT๋ฅผ ๋‹ค๋ฅธ ์˜จ๋„์—์„œ ํ…Œ์ŠคํŠธํ•˜์—ฌ ์ด ๋ฌธ์ œ๋ฅผ ๋ฏธ๋ž˜์— ํ•ด๊ฒฐํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
SUT๋Š” ์ €์˜จ์—์„œ ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ๋™์ผํ•œ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์œผ๋กœ ๋ฐ˜๋ณต์ ์œผ๋กœ ํ”„๋กฌํ”„ํŠธํ•  ๋•Œ SUT์˜ ์‘๋‹ต ๋ณ€๋™์„ฑ์ด ์ค„์–ด๋“ค์–ด ๊ฒฐ๊ณผ๋ฅผ ๋” ์žฌํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ SUT๋Š” ์˜จ๋„๊ฐ€ ๋†’์„์ˆ˜๋ก ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ์‘๋‹ต์˜ ๋น„์œจ์ด ๋†’์•„์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ–ฅํ›„์—๋Š” ๋‹ค์–‘ํ•œ ์˜จ๋„์—์„œ ๊ฐ SUT๋ฅผ ํ…Œ์ŠคํŠธํ•˜์—ฌ ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
In this section, we outline an algorithm for extracting \(\mathbf{W}\) up to an orthogonal \(h\times h\) matrix, instead of merely up to a nonsingular \(h\times h\) matrix as in Appendix C. We do **not** carry out this attack in practice for models considered in this paper, and leave improving this algorithm as an open problem for future work.
์ด ์„น์…˜์—์„œ๋Š” ๋ถ€๋ก C์—์„œ์™€ ๊ฐ™์ด ๋‹จ์ˆœํžˆ ๋น„ํŠน์ด \(h\times h\) ํ–‰๋ ฌ์ด ์•„๋‹Œ, ์ง๊ต \(h\times h\) ํ–‰๋ ฌ๊นŒ์ง€ \(\mathbf{W}\)๋ฅผ ์ถ”์ถœํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฐœ๋žต์ ์œผ๋กœ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด ๋…ผ๋ฌธ์—์„œ ๊ณ ๋ คํ•œ ๋ชจ๋ธ์— ๋Œ€ํ•ด ์‹ค์ œ๋กœ ์ด ๊ณต๊ฒฉ์„ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์œผ๋ฉฐ, ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฌธ์ œ๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ ๊ณผ์ œ๋กœ ๋‚จ๊ฒจ๋‘ก๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋ถ€๋ก C์™€ ๊ฐ™์ด ๋น„ํŠน์ด์  \(h\times h\) ํ–‰๋ ฌ์—์„œ๋งŒ ์ถ”์ถœํ•˜๋Š” ๋Œ€์‹  ์ง๊ต \(h\times h\) ํ–‰๋ ฌ๊นŒ์ง€ \(\mathbf{W}\)๋ฅผ ์ถ”์ถœํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ•œ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ ๊ณ ๋ คํ•œ ๋ชจ๋ธ์— ๋Œ€ํ•ด์„œ๋Š” ์‹ค์ œ๋กœ ์ด ๊ณต๊ฒฉ์„ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์œผ๋ฉฐ, ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์€ ํ–ฅํ›„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ๋ฏธํ•ด๊ฒฐ ๊ณผ์ œ๋กœ ๋‚จ๊ฒจ๋‘”๋‹ค.
In this section, we begin with an overview of the fundamental design and implementation of each module within the LLM kernel. Subsequently, we present the LLM system calls, which encompass essential functions for each module. At last, we discuss the exploration of the AIOS SDK, aiming to facilitate the development process for agent developers.
์ด ์„น์…˜์—์„œ๋Š” LLM ์ปค๋„ ๋‚ด ๊ฐ ๋ชจ๋“ˆ์˜ ๊ธฐ๋ณธ ์„ค๊ณ„ ๋ฐ ๊ตฌํ˜„์— ๋Œ€ํ•œ ๊ฐœ์š”๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ํ›„, ๊ฐ ๋ชจ๋“ˆ์— ํ•„์ˆ˜์ ์ธ ๊ธฐ๋Šฅ์„ ํฌํ•จํ•˜๋Š” LLM ์‹œ์Šคํ…œ ํ˜ธ์ถœ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์—์ด์ „ํŠธ ๊ฐœ๋ฐœ์ž๋“ค์˜ ๊ฐœ๋ฐœ ๊ณผ์ •์„ ์šฉ์ดํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•œ AIOS SDK ํƒ์ƒ‰์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” LLM ์ปค๋„ ๋‚ด์˜ ๊ฐ ๋ชจ๋“ˆ์˜ ๊ธฐ๋ณธ ์„ค๊ณ„์™€ ๊ตฌํ˜„์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์‹œ์ž‘์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ฐ ๋ชจ๋“ˆ์— ๋Œ€ํ•œ ํ•„์ˆ˜ ๊ธฐ๋Šฅ์„ ํฌํ•จํ•˜๋Š” LLM ์‹œ์Šคํ…œ ํ˜ธ์ถœ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ์—์ด์ „ํŠธ ๊ฐœ๋ฐœ์ž์˜ ๊ฐœ๋ฐœ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•œ AIOS SDK ํƒ์ƒ‰์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค.
For any \(i\), on average it takes \(1/p_{i}\) iid samples from \(p\) to see the context \(i\) at least once. The effect of tail-cutting at rank \(k\) is effectively to replace the sample size \(T\) by \(\min(T,T_{k})\), where \(T_{k}=\max\{1/p_{i}\mid i\in[k]\}\). In the case where \(p=Zipf(\beta)\), we have \(T_{k}=1/p_{k}\asymp k^{\beta}\).
์–ด๋–ค \(i\)์— ๋Œ€ํ•ด์„œ๋„, ํ‰๊ท ์ ์œผ๋กœ \(i\) ์ปจํ…์ŠคํŠธ๋ฅผ ์ตœ์†Œํ•œ ํ•œ ๋ฒˆ ๋ณด๊ธฐ ์œ„ํ•ด์„œ๋Š” \(p\)๋กœ๋ถ€ํ„ฐ \(1/p_{i}\) ๊ฐœ์˜ iid ์ƒ˜ํ”Œ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ˆœ์œ„ \(k\)์—์„œ์˜ tail-cutting์˜ ํšจ๊ณผ๋Š” ์ƒ˜ํ”Œ ํฌ๊ธฐ \(T\)๋ฅผ \(\min(T,T_{k})\)๋กœ ๋Œ€์ฒดํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ \(T_{k}=\max\{1/p_{i}\mid i\in[k]\}\)์ž…๋‹ˆ๋‹ค. \(p=Zipf(\beta)\)์ธ ๊ฒฝ์šฐ, \(T_{k}=1/p_{k}\asymp k^{\beta}\)๊ฐ€ ๋ฉ๋‹ˆ๋‹ค.
๋ชจ๋“  \(i\)์— ๋Œ€ํ•ด ํ‰๊ท ์ ์œผ๋กœ \(p\)์—์„œ ์ปจํ…์ŠคํŠธ \(i\)๋ฅผ ํ•œ ๋ฒˆ ์ด์ƒ ๋ณด๋ ค๋ฉด \(1/p_{i}\) iid ์ƒ˜ํ”Œ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ˆœ์œ„ \(k\)์—์„œ ๊ผฌ๋ฆฌ ์ ˆ๋‹จ์˜ ํšจ๊ณผ๋Š” ์‚ฌ์‹ค์ƒ ์ƒ˜ํ”Œ ํฌ๊ธฐ \(T\)๋ฅผ \(\min(T,T_{k})\), ์—ฌ๊ธฐ์„œ \(T_{k}=\max\{1/p_{i}\mid i\in[k]\}\)๋กœ ๋Œ€์ฒดํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. p=Zipf(\beta)\)์ธ ๊ฒฝ์šฐ, \(T_{k}=1/p_{k}\asymp k^{\beta}\)๊ฐ€ ๋ฉ๋‹ˆ๋‹ค.
Multi-agent patterns can address the issues of parallel tasks and robustness since individual agents can work on individual subproblems. Many multi-agent patterns start by taking a complex problem and breaking it down into several smaller tasks. Then, each agent works independently on solving each task using their own independent set of tools.
๋‹ค์ค‘ ์—์ด์ „ํŠธ ํŒจํ„ด์€ ๊ฐœ๋ณ„ ์—์ด์ „ํŠธ๊ฐ€ ๊ฐœ๋ณ„ ํ•˜์œ„ ๋ฌธ์ œ๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ณ‘๋ ฌ ์ž‘์—…๊ณผ ๊ฒฌ๊ณ ์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ํŒจํ„ด์€ ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ์—ฌ๋Ÿฌ ์ž‘์€ ์ž‘์—…์œผ๋กœ ๋ถ„ํ•ดํ•˜๋Š” ๊ฒƒ์—์„œ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ฐ ์—์ด์ „ํŠธ๋Š” ์ž์‹ ๋งŒ์˜ ๋…๋ฆฝ์ ์ธ ๋„๊ตฌ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์ž‘์—…์„ ๋…๋ฆฝ์ ์œผ๋กœ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค.
๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ํŒจํ„ด์€ ๊ฐœ๋ณ„ ์—์ด์ „ํŠธ๊ฐ€ ๊ฐœ๋ณ„ ํ•˜์œ„ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ž‘์—…ํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ณ‘๋ ฌ ์ž‘์—… ๋ฐ ๊ฒฌ๊ณ ์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ํŒจํ„ด์€ ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์ž‘์€ ์ž‘์—…์œผ๋กœ ๋ถ„ํ•ดํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ฐ ์—์ด์ „ํŠธ๋Š” ๊ณ ์œ ํ•œ ๋„๊ตฌ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ๋…๋ฆฝ์ ์œผ๋กœ ์ž‘์—…ํ•ฉ๋‹ˆ๋‹ค.
**Accuracy by Specialty.** Figure A.8 illustrates the DDx accuracy achieved by AMIE and PCPs across the six medical specialties covered by scenarios in our study. We observed that AMIE's performance matched or surpassed PCP performance for all specialties with the most pronounced improvements in the respiratory and cardiovascular specialities.
์ „๋ฌธ๋ณ„ ์ •ํ™•๋„. Figure A.8์€ ๋ณธ ์—ฐ๊ตฌ์˜ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋‹ค๋ฃจ๋Š” ์—ฌ์„ฏ ๊ฐ€์ง€ ์˜ํ•™ ์ „๋ฌธ ๋ถ„์•ผ์— ๊ฑธ์ณ AMIE์™€ PCPs๊ฐ€ ๋‹ฌ์„ฑํ•œ DDx ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. AMIE์˜ ์„ฑ๋Šฅ์ด ๋ชจ๋“  ์ „๋ฌธ ๋ถ„์•ผ์—์„œ PCP์˜ ์„ฑ๋Šฅ๊ณผ ์ผ์น˜ํ•˜๊ฑฐ๋‚˜ ์ด๋ฅผ ๋Šฅ๊ฐ€ํ–ˆ์œผ๋ฉฐ, ํŠนํžˆ ํ˜ธํก๊ธฐ ๋ฐ ์‹ฌํ˜ˆ๊ด€ ์ „๋ฌธ ๋ถ„์•ผ์—์„œ ๊ฐ€์žฅ ๋‘๋“œ๋Ÿฌ์ง„ ํ–ฅ์ƒ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค.
**์ „๋ฌธ๋ณ„ ์ •ํ™•๋„**. ๊ทธ๋ฆผ A.8์€ ์—ฐ๊ตฌ์˜ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋‹ค๋ฃจ๋Š” 6๊ฐœ์˜ ์˜๋ฃŒ ์ „๋ฌธ ๋ถ„์•ผ์— ๊ฑธ์ณ AMIE์™€ PCP๊ฐ€ ๋‹ฌ์„ฑํ•œ DDx ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ˜ธํก๊ธฐ ๋ฐ ์‹ฌํ˜ˆ๊ด€ ์ „๋ฌธ ๋ถ„์•ผ์—์„œ ๊ฐ€์žฅ ๋‘๋“œ๋Ÿฌ์ง„ ๊ฐœ์„ ์„ ํ†ตํ•ด ๋ชจ๋“  ์ „๋ฌธ ๋ถ„์•ผ์—์„œ AMIE์˜ ์„ฑ๋Šฅ์ด PCP ์„ฑ๋Šฅ๊ณผ ์ผ์น˜ํ•˜๊ฑฐ๋‚˜ ์ด๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
For each prompt, we generate one response from each model respectively6. Annotators are then presented with the prompt and two responses: one generated by \(\mathtt{CQIA}\) model and one by another baseline model. Subsequently, we ask which response the annotator prefers, allowing for a "tie" selection when a better response is hard to judge.
๊ฐ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด, ์šฐ๋ฆฌ๋Š” ๊ฐ๊ฐ์˜ ๋ชจ๋ธ๋กœ๋ถ€ํ„ฐ ํ•˜๋‚˜์˜ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ดํ›„ ์ฃผ์„์ž์—๊ฒŒ ํ”„๋กฌํ”„ํŠธ์™€ ๋‘ ๊ฐœ์˜ ์‘๋‹ต์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค: ํ•˜๋‚˜๋Š” \(\mathtt{CQIA}\) ๋ชจ๋ธ์ด ์ƒ์„ฑํ•œ ์‘๋‹ต์ด๊ณ , ๋‹ค๋ฅธ ํ•˜๋‚˜๋Š” ๋‹ค๋ฅธ ๊ธฐ์ค€ ๋ชจ๋ธ์ด ์ƒ์„ฑํ•œ ์‘๋‹ต์ž…๋‹ˆ๋‹ค. ๊ทธ ํ›„, ์ฃผ์„์ž์—๊ฒŒ ์–ด๋Š ์‘๋‹ต์„ ์„ ํ˜ธํ•˜๋Š”์ง€ ๋ฌป๊ณ , ๋” ๋‚˜์€ ์‘๋‹ต์„ ํŒ๋‹จํ•˜๊ธฐ ์–ด๋ ค์šธ ๋•Œ๋Š” "๋™์ " ์„ ํƒ์„ ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค.
๊ฐ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด ๊ฐ ๋ชจ๋ธ์—์„œ ๊ฐ๊ฐ ํ•˜๋‚˜์˜ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.6. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ฃผ์„๊ฐ€์—๊ฒŒ ํ”„๋กฌํ”„ํŠธ์™€ ๋‘ ๊ฐœ์˜ ์‘๋‹ต์ด ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค: ํ•˜๋‚˜๋Š” \(\mathtt{CQIA}\) ๋ชจ๋ธ์—์„œ ์ƒ์„ฑํ•œ ์‘๋‹ต์ด๊ณ  ๋‹ค๋ฅธ ํ•˜๋‚˜๋Š” ๋‹ค๋ฅธ ๊ธฐ์ค€ ๋ชจ๋ธ์—์„œ ์ƒ์„ฑํ•œ ์‘๋‹ต์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ฃผ์„๊ฐ€์—๊ฒŒ ์–ด๋–ค ์‘๋‹ต์„ ์„ ํ˜ธํ•˜๋Š”์ง€ ๋ฌป๊ณ , ๋” ๋‚˜์€ ์‘๋‹ต์„ ํŒ๋‹จํ•˜๊ธฐ ์–ด๋ ค์šธ ๋•Œ "๋™์ " ์„ ํƒ์„ ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค.
Multi-agent architectures can have a wide variety of organizations at any level of complexity. In this paper, we divide them into two primary categories: vertical and horizontal. It is important to keep in mind that these categories represent two ends of a spectrum, where most existing architectures fall somewhere between these two extremes.
๋‹ค์ค‘ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๋Š” ์–ด๋–ค ๋ณต์žก์„ฑ ์ˆ˜์ค€์—์„œ๋„ ๋‹ค์–‘ํ•œ ์กฐ์ง ํ˜•ํƒœ๋ฅผ ๊ฐ€์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋ฅผ ์ˆ˜์ง์  ์•„ํ‚คํ…์ฒ˜์™€ ์ˆ˜ํ‰์  ์•„ํ‚คํ…์ฒ˜์˜ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ๋ฒ”์ฃผ๋กœ ๋‚˜๋ˆ•๋‹ˆ๋‹ค. ์ด ๋ฒ”์ฃผ๋“ค์ด ์ŠคํŽ™ํŠธ๋Ÿผ์˜ ๋‘ ๊ทน๋‹จ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ๋Œ€๋ถ€๋ถ„์˜ ๊ธฐ์กด ์•„ํ‚คํ…์ฒ˜๋Š” ์ด ๋‘ ๊ทน๋‹จ ์‚ฌ์ด ์–ด๋”˜๊ฐ€์— ์œ„์น˜ํ•œ๋‹ค๋Š” ์ ์„ ์—ผ๋‘์— ๋‘๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๋Š” ๋ณต์žก๋„ ์ˆ˜์ค€์— ๊ด€๊ณ„์—†์ด ๋งค์šฐ ๋‹ค์–‘ํ•œ ์กฐ์ง์„ ๊ฐ€์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋ฅผ ์ˆ˜์งํ˜•๊ณผ ์ˆ˜ํ‰ํ˜•์˜ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ๋ฒ”์ฃผ๋กœ ๋‚˜๋ˆ•๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฒ”์ฃผ๋Š” ๋Œ€๋ถ€๋ถ„์˜ ๊ธฐ์กด ์•„ํ‚คํ…์ฒ˜๊ฐ€ ์ด ๋‘ ๊ฐ€์ง€ ๊ทน๋‹จ ์‚ฌ์ด์— ์†ํ•˜๋Š” ์ŠคํŽ™ํŠธ๋Ÿผ์˜ ์–‘์ชฝ ๋์„ ๋‚˜ํƒ€๋‚ธ๋‹ค๋Š” ์ ์„ ๋ช…์‹ฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
We now analyze the efficacy of this attack across a wider range of models: GPT-2 (Radford et al., 2019) Small and XL, Pythia (Biderman et al., 2023) 1.4B and 6.9B, and LLaMA (Touvron et al., 2023) 7B and 65B. The results are in Table 2: our attack recovers the embedding size nearly perfectly, with an error of 0 or 1 in five out of six cases.
์ด์ œ ์šฐ๋ฆฌ๋Š” ๋” ๋„“์€ ๋ฒ”์œ„์˜ ๋ชจ๋ธ์— ๊ฑธ์ณ ์ด ๊ณต๊ฒฉ์˜ ํšจ๋Šฅ์„ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค: GPT-2 (Radford et al., 2019) Small ๋ฐ XL, Pythia (Biderman et al., 2023) 1.4B ๋ฐ 6.9B, ๊ทธ๋ฆฌ๊ณ  LLaMA (Touvron et al., 2023) 7B ๋ฐ 65B. ๊ฒฐ๊ณผ๋Š” Table 2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค: ์šฐ๋ฆฌ์˜ ๊ณต๊ฒฉ์€ ์—ฌ์„ฏ ๊ฐ€์ง€ ๊ฒฝ์šฐ ์ค‘ ๋‹ค์„ฏ ๊ฐ€์ง€์—์„œ ์˜ค๋ฅ˜๊ฐ€ 0 ๋˜๋Š” 1์ธ ๊ฑฐ์˜ ์™„๋ฒฝํ•˜๊ฒŒ ์ž„๋ฒ ๋”ฉ ํฌ๊ธฐ๋ฅผ ๋ณต๊ตฌํ•ฉ๋‹ˆ๋‹ค.
์ด์ œ ๋” ๊ด‘๋ฒ”์œ„ํ•œ ๋ชจ๋ธ์—์„œ ์ด ๊ณต๊ฒฉ์˜ ํšจ๊ณผ๋ฅผ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค: GPT-2(Radford et al., 2019) Small ๋ฐ XL, Pythia(Biderman et al., 2023) 1.4B ๋ฐ 6.9B, LLaMA(Touvron et al., 2023) 7B ๋ฐ 65B์ž…๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ๋Š” ํ‘œ 2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๊ณต๊ฒฉ์€ ์ž„๋ฒ ๋”ฉ ํฌ๊ธฐ๋ฅผ ๊ฑฐ์˜ ์™„๋ฒฝํ•˜๊ฒŒ ๋ณต๊ตฌํ•˜๋ฉฐ, 6๊ฐœ ์ค‘ 5๊ฐœ์˜ ๊ฒฝ์šฐ ์˜ค์ฐจ๊ฐ€ 0 ๋˜๋Š” 1์ž…๋‹ˆ๋‹ค.
Due to their autoregressive architecture, LLMs typically have slow inference speeds. However, the agent may need to query LLMs for each action multiple times, such as extracting information from memory, make plans before taking actions and so on. Consequently, the efficiency of agent actions is greatly affected by the speed of LLM inference.
์ž๊ธฐํšŒ๊ท€ ์•„ํ‚คํ…์ฒ˜๋กœ ์ธํ•ด, LLMs๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ถ”๋ก  ์†๋„๊ฐ€ ๋Š๋ฆฝ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์—์ด์ „ํŠธ๋Š” ๊ฐ ํ–‰๋™์— ๋Œ€ํ•ด ์—ฌ๋Ÿฌ ๋ฒˆ LLMs์— ์งˆ์˜ํ•  ํ•„์š”๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•˜๊ฑฐ๋‚˜ ํ–‰๋™์„ ์ทจํ•˜๊ธฐ ์ „์— ๊ณ„ํš์„ ์„ธ์šฐ๋Š” ๋“ฑ์˜ ๊ฒฝ์šฐ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—์ด์ „ํŠธ ํ–‰๋™์˜ ํšจ์œจ์„ฑ์€ LLM ์ถ”๋ก  ์†๋„์— ํฌ๊ฒŒ ์˜ํ–ฅ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค.
์ž๋™ ํšŒ๊ท€ ์•„ํ‚คํ…์ฒ˜๋กœ ์ธํ•ด LLM์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ถ”๋ก  ์†๋„๊ฐ€ ๋Š๋ฆฝ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์—์ด์ „ํŠธ๋Š” ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•˜๊ฑฐ๋‚˜, ํ–‰๋™์„ ์ทจํ•˜๊ธฐ ์ „์— ๊ณ„ํš์„ ์„ธ์šฐ๋Š” ๋“ฑ์˜ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋“ฑ ๊ฐ ์ž‘์—…์— ๋Œ€ํ•ด ์—ฌ๋Ÿฌ ๋ฒˆ LLM์— ์ฟผ๋ฆฌํ•ด์•ผ ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—์ด์ „ํŠธ ์•ก์…˜์˜ ํšจ์œจ์„ฑ์€ LLM ์ถ”๋ก  ์†๋„์— ํฐ ์˜ํ–ฅ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค.
HIIT differs from other forms of aerobic exercise, such as jogging, cycling, or swimming, in its high-intensity, short-burst nature. While traditional aerobic exercises have a continuous, steady state of intensity, HIIT uses brief periods of intense activity followed by rest, which can result in greater fatigue and thus improved performance.
HIIT๋Š” ์กฐ๊น…, ์ž์ „๊ฑฐ ํƒ€๊ธฐ, ์ˆ˜์˜๊ณผ ๊ฐ™์€ ๋‹ค๋ฅธ ํ˜•ํƒœ์˜ ์œ ์‚ฐ์†Œ ์šด๋™๊ณผ๋Š” ๊ทธ ๊ฐ•๋„ ๋†’์€ ์งง์€ ํญ๋ฐœ์ ์ธ ํŠน์„ฑ์—์„œ ์ฐจ์ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „ํ†ต์ ์ธ ์œ ์‚ฐ์†Œ ์šด๋™์ด ์ง€์†์ ์ด๊ณ  ์ผ์ •ํ•œ ๊ฐ•๋„๋ฅผ ์œ ์ง€ํ•˜๋Š” ๋ฐ˜๋ฉด, HIIT๋Š” ์งง์€ ๊ธฐ๊ฐ„์˜ ๊ฐ•๋ ฌํ•œ ํ™œ๋™๊ณผ ํœด์‹์„ ๋ฒˆ๊ฐˆ์•„ ์‚ฌ์šฉํ•˜์—ฌ ๋” ํฐ ํ”ผ๋กœ๋ฅผ ์œ ๋ฐœํ•˜๊ณ  ๋”ฐ๋ผ์„œ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
HIIT๋Š” ์กฐ๊น…, ์ž์ „๊ฑฐ ํƒ€๊ธฐ ๋˜๋Š” ์ˆ˜์˜๊ณผ ๊ฐ™์€ ๋‹ค๋ฅธ ํ˜•ํƒœ์˜ ์œ ์‚ฐ์†Œ ์šด๋™๊ณผ ๊ณ ๊ฐ•๋„์˜ ์งง์€ ๋ฒ„์ŠคํŠธ ํŠน์„ฑ์ด ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ๊ธฐ์กด์˜ ์œ ์‚ฐ์†Œ ์šด๋™์€ ์ง€์†์ ์ด๊ณ  ์ผ์ •ํ•œ ๊ฐ•๋„๋ฅผ ์œ ์ง€ํ•˜๋Š” ๋ฐ˜๋ฉด, HIIT๋Š” ์งง์€ ์‹œ๊ฐ„ ๋™์•ˆ ๊ฐ•๋„ ๋†’์€ ํ™œ๋™์„ ํ•œ ํ›„ ํœด์‹์„ ์ทจํ•˜๋ฏ€๋กœ ํ”ผ๋กœ๋„๊ฐ€ ๋†’์•„์ ธ ์šด๋™ ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 4: Left: The relationship between the mean accuracy on the GSM8K and the number of annotated CoT steps with data increasing. The solid line is fitted using all seven points, while the dashed line is fitted using the first four points. Right: Changes in mean accuracy when resampling is used to increase the CoT lengeh of training data.
Figure 4: ์™ผ์ชฝ: GSM8K์˜ ํ‰๊ท  ์ •ํ™•๋„์™€ ์ฃผ์„์ด ๋‹ฌ๋ฆฐ CoT ๋‹จ๊ณ„ ์ˆ˜์˜ ๊ด€๊ณ„, ๋ฐ์ดํ„ฐ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ. ์‹ค์„ ์€ ๋ชจ๋“  7๊ฐœ์˜ ์ ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋งž์ถฐ์กŒ๊ณ , ์ ์„ ์€ ์ฒ˜์Œ 4๊ฐœ์˜ ์ ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋งž์ถฐ์กŒ์Šต๋‹ˆ๋‹ค. ์˜ค๋ฅธ์ชฝ: ์žฌ์ƒ˜ํ”Œ๋ง์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ CoT ๊ธธ์ด๋ฅผ ์ฆ๊ฐ€์‹œํ‚ฌ ๋•Œ ํ‰๊ท  ์ •ํ™•๋„์˜ ๋ณ€ํ™”.
๊ทธ๋ฆผ 4: ์™ผ์ชฝ: GSM8K์˜ ํ‰๊ท  ์ •ํ™•๋„์™€ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ€์— ๋”ฐ๋ฅธ ์ฃผ์„์ด ๋‹ฌ๋ฆฐ CoT ๋‹จ๊ณ„ ์ˆ˜ ์‚ฌ์ด์˜ ๊ด€๊ณ„. ์‹ค์„ ์€ 7๊ฐœ์˜ ์ ์„ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜์—ฌ ๋งž์ถคํ•œ ๊ฒƒ์ด๊ณ , ์ ์„ ์€ ์ฒ˜์Œ 4๊ฐœ์˜ ์ ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋งž์ถคํ•œ ๊ฒƒ์ด๋‹ค. ์˜ค๋ฅธ์ชฝ: ์žฌ์ƒ˜ํ”Œ๋ง์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ CoT ๊ธธ์ด๋ฅผ ๋Š˜๋ฆด ๋•Œ ํ‰๊ท  ์ •ํ™•๋„์˜ ๋ณ€ํ™”.
Note that the hidden dimension size is much smaller than the size of the token dictionary, i.e., \(h\ll l\). For example, LLaMA (Touvron et al., 2023) chooses \(h\in\{4096,5120,6656,8192\}\) and \(l=32{,}000\), and there is a recent trend towards increasingly large token sizes; GPT-4, for example, has a \(\approx\)100,000 token vocabulary.
์ˆจ๊ฒจ์ง„ ์ฐจ์›์˜ ํฌ๊ธฐ๋Š” ํ† ํฐ ์‚ฌ์ „์˜ ํฌ๊ธฐ๋ณด๋‹ค ํ›จ์”ฌ ์ž‘๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์‹ญ์‹œ์˜ค, ์ฆ‰ \(h\ll l\). ์˜ˆ๋ฅผ ๋“ค์–ด, LLaMA (Touvron et al., 2023)๋Š” \(h\in\{4096,5120,6656,8192\}\) ๋ฐ \(l=32{,}000\)์„ ์„ ํƒํ•˜๋ฉฐ, ์ตœ๊ทผ์—๋Š” ์ ์  ๋” ํฐ ํ† ํฐ ํฌ๊ธฐ๋กœ ๋‚˜์•„๊ฐ€๋Š” ์ถ”์„ธ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค; ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-4๋Š” \(\approx\)100,000 ํ† ํฐ ์–ดํœ˜๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
์ˆจ๊ฒจ์ง„ ์ฐจ์› ํฌ๊ธฐ๋Š” ํ† ํฐ ์‚ฌ์ „์˜ ํฌ๊ธฐ๋ณด๋‹ค ํ›จ์”ฌ ์ž‘์Šต๋‹ˆ๋‹ค(์ฆ‰, \(h\ll l\)). ์˜ˆ๋ฅผ ๋“ค์–ด, LLaMA(Touvron et al., 2023)๋Š” \(h\in\{4096,5120,6656,8192\}\), \(l=32{,}000\)์„ ์„ ํƒํ•˜๋ฉฐ, ์ตœ๊ทผ์—๋Š” ํ† ํฐ ํฌ๊ธฐ๊ฐ€ ์ ์  ์ปค์ง€๋Š” ์ถ”์„ธ๊ฐ€ ์žˆ์œผ๋ฉฐ, ์˜ˆ๋ฅผ ๋“ค์–ด GPT-4์˜ ์–ดํœ˜๋Š” \(\approx\)100,000 ํ† ํฐ์ž…๋‹ˆ๋‹ค.
Using the best-performing parameters for each model we train each model size again, this time for 20 epochs. We use five seeds for each model size. Again we do not observe any convergence. Instead the accuracy fluctuates randomly between 0 and 7. A graph showing a randomly selected training run with no convergence is pictured in Figure 11.
๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ, ์ด๋ฒˆ์—๋Š” 20 ์—ํฌํฌ ๋™์•ˆ ๊ฐ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ๋‹ค์‹œ ํ›ˆ๋ จ์‹œํ‚ต๋‹ˆ๋‹ค. ๊ฐ ๋ชจ๋ธ ํฌ๊ธฐ์— ๋Œ€ํ•ด ๋‹ค์„ฏ ๊ฐœ์˜ ์‹œ๋“œ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์‹œ ํ•œ ๋ฒˆ, ์ˆ˜๋ ด์€ ๊ด€์ฐฐ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋Œ€์‹  ์ •ํ™•๋„๋Š” 0๊ณผ 7 ์‚ฌ์ด์—์„œ ๋ฌด์ž‘์œ„๋กœ ๋ณ€๋™ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜๋ ด์ด ์—†๋Š” ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ ํ›ˆ๋ จ ๊ณผ์ •์„ ๋ณด์—ฌ์ฃผ๋Š” ๊ทธ๋ž˜ํ”„๋Š” Figure 11์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฒˆ์—๋Š” 20๊ฐœ์˜ ์—ํฌํฌ์— ๋Œ€ํ•ด ๊ฐ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ๋‹ค์‹œ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ชจ๋ธ ํฌ๊ธฐ์— ๋Œ€ํ•ด 5๊ฐœ์˜ ์‹œ๋“œ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์‹œ ํ•œ ๋ฒˆ ์ˆ˜๋ ด์ด ๊ด€์ฐฐ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋Œ€์‹  ์ •ํ™•๋„๊ฐ€ 0์—์„œ 7 ์‚ฌ์ด์—์„œ ๋ฌด์ž‘์œ„๋กœ ๋ณ€๋™ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜๋ ด์ด ์—†๋Š” ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ ํ›ˆ๋ จ ์‹คํ–‰์„ ๋ณด์—ฌ์ฃผ๋Š” ๊ทธ๋ž˜ํ”„๊ฐ€ ๊ทธ๋ฆผ 11์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
The large-scale and diverse nature of web corpora make them difficult to document and analyse as a whole; we provide some key metrics in the section, focusing on document lengths in Figure 5(a), and a breakdown of the top domain names in Figure 5(b). We also refer to the analysis of the distribution of toxic content presented in Figure 4.
์›น ์ฝ”ํผ์Šค์˜ ๋Œ€๊ทœ๋ชจ ๋ฐ ๋‹ค์–‘ํ•œ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ์ด๋ฅผ ์ „์ฒด์ ์œผ๋กœ ๋ฌธ์„œํ™”ํ•˜๊ณ  ๋ถ„์„ํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” Figure 5(a)์—์„œ ๋ฌธ์„œ ๊ธธ์ด์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ , Figure 5(b)์—์„œ ์ƒ์œ„ ๋„๋ฉ”์ธ ์ด๋ฆ„์˜ ๋ถ„๋ฅ˜๋ฅผ ์ œ๊ณตํ•˜๋Š” ์„น์…˜์—์„œ ๋ช‡ ๊ฐ€์ง€ ์ฃผ์š” ์ง€ํ‘œ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Figure 4์— ์ œ์‹œ๋œ ์œ ํ•ด ์ฝ˜ํ…์ธ  ๋ถ„ํฌ ๋ถ„์„์„ ์ฐธ์กฐํ•ฉ๋‹ˆ๋‹ค.
์›น ์ฝ”ํผ๋ผ์˜ ๋Œ€๊ทœ๋ชจ์™€ ๋‹ค์–‘์„ฑ์œผ๋กœ ์ธํ•ด ์ „์ฒด์ ์œผ๋กœ ๋ฌธ์„œํ™”ํ•˜๊ณ  ๋ถ„์„ํ•˜๊ธฐ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์— ์„น์…˜์—์„œ ๋ช‡ ๊ฐ€์ง€ ์ฃผ์š” ๋ฉ”ํŠธ๋ฆญ์„ ์ œ๊ณตํ•˜๋ฉฐ, ๊ทธ๋ฆผ 5(a)์—์„œ๋Š” ๋ฌธ์„œ ๊ธธ์ด์— ์ค‘์ ์„ ๋‘๊ณ  ๊ทธ๋ฆผ 5(b)์—์„œ๋Š” ์ƒ์œ„ ๋„๋ฉ”์ธ ์ด๋ฆ„ ๋ถ„์„์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ทธ๋ฆผ 4์— ์ œ์‹œ๋œ ์œ ํ•ด ์ฝ˜ํ…์ธ  ๋ถ„ํฌ ๋ถ„์„๋„ ์ฐธ์กฐํ•ฉ๋‹ˆ๋‹ค.
Comparing with the Close-Source Models.In Table 1, our _WizardMath 70B_ slightly outperforms some close-source LLMs on GSM8k, including ChatGPT, Claude Instant and PaLM 2 540B. And as shown in Figure 2, our model is currently ranked in the top five on all models. Simultaneously,_WizardMath 70B_ also surpasses the Text-davinci-002 on MATH.
๋น„๊ณต๊ฐœ ๋ชจ๋ธ๊ณผ์˜ ๋น„๊ต. Table 1์—์„œ, ์šฐ๋ฆฌ _WizardMath 70B_๋Š” GSM8k์—์„œ ChatGPT, Claude Instant ๋ฐ PaLM 2 540B๋ฅผ ํฌํ•จํ•œ ์ผ๋ถ€ ๋น„๊ณต๊ฐœ LLM๋ณด๋‹ค ์•ฝ๊ฐ„ ๋” ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  Figure 2์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ํ˜„์žฌ ๋ชจ๋“  ๋ชจ๋ธ ์ค‘ ์ƒ์œ„ 5์œ„ ์•ˆ์— ๋žญํฌ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋™์‹œ์—, _WizardMath 70B_๋Š” MATH์—์„œ Text-davinci-002๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
๋น„๊ต ํด๋กœ์ฆˆ ์†Œ์Šค ๋ชจ๋ธ.ํ‘œ 1์—์„œ, ์šฐ๋ฆฌ์˜ _WizardMath 70B_๋Š” ChatGPT, Claude Instant ๋ฐ PaLM 2 540B๋ฅผ ํฌํ•จํ•œ ์ผ๋ถ€ ๋น„๊ณต๊ฐœ LLM๋ณด๋‹ค GSM8k์—์„œ ์•ฝ๊ฐ„ ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๊ทธ๋ฆผ 2์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ํ˜„์žฌ ๋ชจ๋“  ๋ชจ๋ธ์—์„œ ์ƒ์œ„ 5์œ„ ์•ˆ์— ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ๋™์‹œ์— _WizardMath 70B_๋Š” ๋˜ํ•œ MATH์—์„œ Text-davinci-002๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
Finally, it is important to note that MoEs can be used as a component of any architecture regardless of whether it is based on attention or not. In fact, MoEs have also been applied to SSM-based LLMs like Mamba citepioro2024moemamba. We should continue to see MoE-driven improvements in the future regardless of the underlying architecture.
๋งˆ์ง€๋ง‰์œผ๋กœ, MoEs๋Š” ์ฃผ์˜(attention) ๊ธฐ๋ฐ˜์ธ์ง€ ์—ฌ๋ถ€์— ๊ด€๊ณ„์—†์ด ๋ชจ๋“  ์•„ํ‚คํ…์ฒ˜์˜ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ์ฃผ๋ชฉํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ, MoEs๋Š” Mamba citepioro2024moemamba์™€ ๊ฐ™์€ SSM ๊ธฐ๋ฐ˜ LLM์—๋„ ์ ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๊ธฐ๋ณธ ์•„ํ‚คํ…์ฒ˜์™€ ๊ด€๊ณ„์—†์ด MoE์— ์˜ํ•ด ์ฃผ๋„๋˜๋Š” ๊ฐœ์„ ์„ ๊ณ„์†ํ•ด์„œ ๋ณด๊ฒŒ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๋งˆ์ง€๋ง‰์œผ๋กœ, MoE๋Š” ์ฃผ์˜์— ๊ธฐ๋ฐ˜ํ•˜๋Š”์ง€ ์—ฌ๋ถ€์— ๊ด€๊ณ„์—†์ด ๋ชจ๋“  ์•„ํ‚คํ…์ฒ˜์˜ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ MoE๋Š” Mamba citepioro2024moemamba์™€ ๊ฐ™์€ SSM ๊ธฐ๋ฐ˜ LLM์—๋„ ์ ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ–ฅํ›„์—๋„ ๊ธฐ๋ณธ ์•„ํ‚คํ…์ฒ˜์— ๊ด€๊ณ„์—†์ด MoE๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๊ฐœ์„  ์‚ฌํ•ญ์„ ๊ณ„์† ๋ณผ ์ˆ˜ ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
MinHash [1] is an approximate matching algorithm widely used in large-scale deduplication tasks [23, 24, 25], including to deduplicate the training set for a large Chinese-language LM [30]. Given two documents \(x_{i}\) and \(x_{j}\), the main idea is to represent each document by its respective set of \(n\)-grams \(d_{i}\) and \(d_{j}\).
MinHash [1]๋Š” ๋Œ€๊ทœ๋ชจ ์ค‘๋ณต ์ œ๊ฑฐ ์ž‘์—… [23, 24, 25]์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๊ทผ์‚ฌ ๋งค์นญ ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ, ๋Œ€ํ˜• ์ค‘๊ตญ์–ด LM์˜ ํ•™์Šต ์„ธํŠธ๋ฅผ ์ค‘๋ณต ์ œ๊ฑฐํ•˜๋Š” ๋ฐ์—๋„ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค [30]. ๋‘ ๋ฌธ์„œ \(x_{i}\)์™€ \(x_{j}\)๊ฐ€ ์ฃผ์–ด์ง€๋ฉด, ์ฃผ์š” ์•„์ด๋””์–ด๋Š” ๊ฐ ๋ฌธ์„œ๋ฅผ ํ•ด๋‹น \(n\)-๊ทธ๋žจ ์ง‘ํ•ฉ \(d_{i}\)์™€ \(d_{j}\)๋กœ ํ‘œํ˜„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
MinHash [1]๋Š” ๋Œ€๊ทœ๋ชจ ์ค‘๊ตญ์–ด LM [30]์˜ ํ›ˆ๋ จ ์„ธํŠธ ์ค‘๋ณต์„ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•˜์—ฌ ๋Œ€๊ทœ๋ชจ ์ค‘๋ณต ์ œ๊ฑฐ ์ž‘์—… [23, 24, 25]์— ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋Œ€๋žต์ ์ธ ์ผ์น˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ž…๋‹ˆ๋‹ค. ๋‘ ๋ฌธ์„œ \(x_{i}\)์™€ \(x_{j}\)๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ๊ฐ ๋ฌธ์„œ๋ฅผ ๊ฐ๊ฐ์˜ \(n\)-๊ทธ๋žจ \(d_{i}\) ๋ฐ \(d_{j}\)์˜ ์ง‘ํ•ฉ์œผ๋กœ ํ‘œํ˜„ํ•˜๋Š” ๊ฒƒ์ด ์ฃผ์š” ์•„์ด๋””์–ด์ž…๋‹ˆ๋‹ค.
Given the results from the prior section, we chose to implement the improved 4-logprob attack (Section 5.3) because it is both the most query efficient attack and also the most precise attack. Switching to a different attack algorithm would increase our total experiment cost significantly, and so we do not perform these ablation studies.
์ด์ „ ์„น์…˜์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๊ฐœ์„ ๋œ 4-logprob ๊ณต๊ฒฉ(์„น์…˜ 5.3)์„ ๊ตฌํ˜„ํ•˜๊ธฐ๋กœ ๊ฒฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ฐ€์žฅ ์ฟผ๋ฆฌ ํšจ์œจ์ ์ธ ๊ณต๊ฒฉ์ด์ž ๊ฐ€์žฅ ์ •๋ฐ€ํ•œ ๊ณต๊ฒฉ์ด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๊ณต๊ฒฉ ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ์ „ํ™˜ํ•˜๋ฉด ์ „์ฒด ์‹คํ—˜ ๋น„์šฉ์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•  ๊ฒƒ์ด๋ฏ€๋กœ, ์ด๋Ÿฌํ•œ ์†Œ๊ฑฐ ์—ฐ๊ตฌ๋Š” ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
์•ž ์ ˆ์˜ ๊ฒฐ๊ณผ๋ฅผ ๊ณ ๋ คํ•  ๋•Œ, ์ฟผ๋ฆฌ ํšจ์œจ์„ฑ์ด ๊ฐ€์žฅ ๋†’์€ ๊ณต๊ฒฉ์ด์ž ๊ฐ€์žฅ ์ •ํ™•ํ•œ ๊ณต๊ฒฉ์ด๊ธฐ ๋•Œ๋ฌธ์— ๊ฐœ์„ ๋œ 4-logprob ๊ณต๊ฒฉ(์„น์…˜ 5.3)์„ ๊ตฌํ˜„ํ•˜๊ธฐ๋กœ ๊ฒฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๊ณต๊ฒฉ ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ์ „ํ™˜ํ•˜๋ฉด ์ด ์‹คํ—˜ ๋น„์šฉ์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•˜๋ฏ€๋กœ ์ด๋Ÿฌํ•œ ์‚ญ๊ฐ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
Footnote 1: The accuracy numbers are reported in the GPT-4 technical report (OpenAI, 2023b). GPT-4 models are continuously being improved. The latest GPT-4 Turbo (1106) API has increased accuracy to 94.8% on GSM8K and 64.5% on MATH. However, the LLaMA-2 7B model using the best of 256 generations still outperforms the latest GPT-4 models.
Footnote 1: ์ •ํ™•๋„ ์ˆ˜์น˜๋Š” GPT-4 ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ(OpenAI, 2023b)์— ๋ณด๊ณ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. GPT-4 ๋ชจ๋ธ์€ ์ง€์†์ ์œผ๋กœ ๊ฐœ์„ ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ์‹  GPT-4 Turbo (1106) API๋Š” GSM8K์—์„œ 94.8%, MATH์—์„œ 64.5%์˜ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ 256 ์„ธ๋Œ€ ์ค‘ ์ตœ์ƒ์˜ ์„ฑ๋Šฅ์„ ์‚ฌ์šฉํ•˜๋Š” LLaMA-2 7B ๋ชจ๋ธ์ด ์—ฌ์ „ํžˆ ์ตœ์‹  GPT-4 ๋ชจ๋ธ์„ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
๊ฐ์ฃผ 1: ์ •ํ™•๋„ ์ˆ˜์น˜๋Š” GPT-4 ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ(OpenAI, 2023b)์— ๋ณด๊ณ ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. GPT-4 ๋ชจ๋ธ์€ ์ง€์†์ ์œผ๋กœ ๊ฐœ์„ ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ์‹  GPT-4 ํ„ฐ๋ณด(1106) API๋Š” GSM8K์—์„œ 94.8%, MATH์—์„œ 64.5%๋กœ ์ •ํ™•๋„๊ฐ€ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ 256์„ธ๋Œ€ ์ค‘ ์ตœ๊ณ ์˜ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” LLaMA-2 7B ๋ชจ๋ธ์€ ์—ฌ์ „ํžˆ ์ตœ์‹  GPT-4 ๋ชจ๋ธ๋ณด๋‹ค ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค.
Any single query to \(\mathcal{O}\), no matter how well-crafted, yields at most \(\log_{2}(l)\) bits, because the output is one of \(l\) distinct values. The minimum number of queries required is at least the total information content divided by the information per query, yielding the lower bound \(l\log_{2}(B/\varepsilon)/\log_{2}(l)\).
์–ด๋–ค ๋‹จ์ผ ์ฟผ๋ฆฌ๋„ \(\mathcal{O}\)์— ๋Œ€ํ•ด, ์•„๋ฌด๋ฆฌ ์ž˜ ๋งŒ๋“ค์–ด์กŒ๋”๋ผ๋„, ์ตœ๋Œ€ \(\log_{2}(l)\) ๋น„ํŠธ๋งŒ์„ ์‚ฐ์ถœํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ถœ๋ ฅ์ด \(l\)๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ๊ฐ’ ์ค‘ ํ•˜๋‚˜์ด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ํ•„์š”ํ•œ ์ตœ์†Œ ์ฟผ๋ฆฌ ์ˆ˜๋Š” ์ด ์ •๋ณด๋Ÿ‰์„ ์ฟผ๋ฆฌ๋‹น ์ •๋ณด๋Ÿ‰์œผ๋กœ ๋‚˜๋ˆˆ ๊ฐ’ ์ด์ƒ์ด๋ฉฐ, ์ด๋Š” ํ•˜ํ•œ๊ฐ’ \(l\log_{2}(B/\varepsilon)/\log_{2}(l)\)์„ ์‚ฐ์ถœํ•ฉ๋‹ˆ๋‹ค.
์•„๋ฌด๋ฆฌ ์ž˜ ๋งŒ๋“ค์–ด์ง„ ์ฟผ๋ฆฌ๋ผ๋„ \(\mathcal{O}\)์— ๋Œ€ํ•œ ๋‹จ์ผ ์ฟผ๋ฆฌ๋Š” ์ถœ๋ ฅ์ด \(l\)๊ฐœ์˜ ๊ณ ์œ  ๊ฐ’ ์ค‘ ํ•˜๋‚˜์ด๊ธฐ ๋•Œ๋ฌธ์— ์ตœ๋Œ€ \(\log_{2}(l)\) ๋น„ํŠธ๋งŒ ์‚ฐ์ถœํ•ฉ๋‹ˆ๋‹ค. ํ•„์š”ํ•œ ์ตœ์†Œ ์ฟผ๋ฆฌ ์ˆ˜๋Š” ์ด ์ •๋ณด ๋‚ด์šฉ์„ ์ฟผ๋ฆฌ๋‹น ์ •๋ณด๋กœ ๋‚˜๋ˆˆ ๊ฐ’์ด๋ฉฐ, ํ•˜ํ•œ์€ \(l\log_{2}(B/\varepsilon)/\log_{2}(l)\)์ž…๋‹ˆ๋‹ค.
In this section, we present additional results obtained during the development of the Macrodata Refinement pipeline. For Appendix E.1 and Appendix E.3, these were obtained using earlier development versions of the dataset, so results are not directly comparable with the main text. For Appendix E.2, this is based on the Falcon-RW models.
์ด ์„น์…˜์—์„œ๋Š” Macrodata Refinement ํŒŒ์ดํ”„๋ผ์ธ ๊ฐœ๋ฐœ ์ค‘์— ์–ป์€ ์ถ”๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ถ€๋ก E.1๊ณผ ๋ถ€๋ก E.3์˜ ๊ฒฝ์šฐ, ๋ฐ์ดํ„ฐ์…‹์˜ ์ดˆ๊ธฐ ๊ฐœ๋ฐœ ๋ฒ„์ „์„ ์‚ฌ์šฉํ•˜์—ฌ ์–ป์€ ๊ฒฐ๊ณผ์ด๋ฏ€๋กœ ๋ณธ๋ฌธ๊ณผ ์ง์ ‘์ ์œผ๋กœ ๋น„๊ตํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋ถ€๋ก E.2์˜ ๊ฒฝ์šฐ, Falcon-RW ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋งคํฌ๋กœ๋ฐ์ดํ„ฐ ๊ตฌ์ฒดํ™” ํŒŒ์ดํ”„๋ผ์ธ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋™์•ˆ ์–ป์€ ์ถ”๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ถ€๋ก E.1 ๋ฐ ๋ถ€๋ก E.3์˜ ๊ฒฝ์šฐ, ์ด๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ดˆ๊ธฐ ๊ฐœ๋ฐœ ๋ฒ„์ „์„ ์‚ฌ์šฉํ•˜์—ฌ ์–ป์€ ๊ฒƒ์ด๋ฏ€๋กœ ๊ฒฐ๊ณผ๋Š” ๋ณธ๋ฌธ๊ณผ ์ง์ ‘ ๋น„๊ตํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋ถ€๋ก E.2์˜ ๊ฒฝ์šฐ, ์ด๋Š” ํŒ”์ฝ˜-RW ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
Sophie continued to experiment with the sphere and the crystal, trying to harness their power for her own purposes. But as she delved deeper into the mysteries of the object, she began to realize that she might not be able to control it. The sphere and the crystal were drawing her in, and she was starting to lose herself to their power.
์†Œํ”ผ๋Š” ๊ตฌ์ฒด์™€ ํฌ๋ฆฌ์Šคํƒˆ์„ ๊ฐ€์ง€๊ณ  ์‹คํ—˜์„ ๊ณ„์†ํ•˜๋ฉฐ, ๊ทธ๊ฒƒ๋“ค์˜ ํž˜์„ ์ž์‹ ์˜ ๋ชฉ์ ์— ๋งž๊ฒŒ ํ™œ์šฉํ•˜๋ ค๊ณ  ๋…ธ๋ ฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ทธ๋…€๊ฐ€ ๊ทธ ๋ฌผ์ฒด์˜ ์‹ ๋น„๋ฅผ ๋” ๊นŠ์ด ํŒŒ๊ณ ๋“ค์ˆ˜๋ก, ๊ทธ๊ฒƒ์„ ํ†ต์ œํ•  ์ˆ˜ ์—†์„์ง€๋„ ๋ชจ๋ฅธ๋‹ค๋Š” ๊ฒƒ์„ ๊นจ๋‹ซ๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์™€ ํฌ๋ฆฌ์Šคํƒˆ์€ ๊ทธ๋…€๋ฅผ ๋Œ์–ด๋‹น๊ธฐ๊ณ  ์žˆ์—ˆ๊ณ , ๊ทธ๋…€๋Š” ๊ทธ๋“ค์˜ ํž˜์— ์ž์‹ ์„ ์žƒ์–ด๊ฐ€๊ณ  ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
์†Œํ”ผ๋Š” ๊ตฌ์ฒด์™€ ์ˆ˜์ •์„ ๊ณ„์† ์‹คํ—˜ํ•˜๋ฉฐ ์ž์‹ ์˜ ๋ชฉ์ ์„ ์œ„ํ•ด ๊ทธ ํž˜์„ ์ด์šฉํ•˜๋ ค๊ณ  ๋…ธ๋ ฅํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์†Œํ”ผ๋Š” ์ด ๋ฌผ์ฒด์˜ ์‹ ๋น„๋ฅผ ๋” ๊นŠ์ด ํŒŒ๊ณ ๋“ค์ˆ˜๋ก ์ž์‹ ์ด ์ด ๋ฌผ์ฒด๋ฅผ ํ†ต์ œํ•  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ์„ ๊นจ๋‹ซ๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์™€ ์ˆ˜์ •์€ ๊ทธ๋…€๋ฅผ ๋Œ์–ด๋‹น๊ธฐ๊ณ  ์žˆ์—ˆ๊ณ , ๊ทธ๋…€๋Š” ๊ทธ ํž˜์— ๋น ์ ธ๋“ค๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค.
In addition to existing tools developed by humans, LLMs possess the capability to make their own tools for specific tasks autonomously [705]. This enables the models to independently explore and manipulate these self-created tools, thereby expanding their potential for autonomous exploration in solving a wide range of real-world tasks.
๊ธฐ์กด์— ์ธ๊ฐ„์ด ๊ฐœ๋ฐœํ•œ ๋„๊ตฌ ์™ธ์—๋„, LLMs๋Š” ํŠน์ • ์ž‘์—…์„ ์œ„ํ•ด ์ž์œจ์ ์œผ๋กœ ์ž์ฒด ๋„๊ตฌ๋ฅผ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค [705]. ์ด๋Š” ๋ชจ๋ธ์ด ์ด๋Ÿฌํ•œ ์ž๊ฐ€ ์ƒ์„ฑ ๋„๊ตฌ๋ฅผ ๋…๋ฆฝ์ ์œผ๋กœ ํƒ์ƒ‰ํ•˜๊ณ  ์กฐ์ž‘ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•˜์—ฌ, ๋‹ค์–‘ํ•œ ์‹ค์ œ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์žˆ์–ด ์ž์œจ์  ํƒ์ƒ‰์˜ ์ž ์žฌ๋ ฅ์„ ํ™•์žฅ์‹œํ‚ต๋‹ˆ๋‹ค.
LLM์€ ์ธ๊ฐ„์ด ๊ฐœ๋ฐœํ•œ ๊ธฐ์กด ๋„๊ตฌ ์™ธ์—๋„ ํŠน์ • ์ž‘์—…์„ ์œ„ํ•ด ์ž์ฒด์ ์œผ๋กœ ๋„๊ตฌ๋ฅผ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[705]. ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์€ ์Šค์Šค๋กœ ๋งŒ๋“ค์–ด๋‚ธ ๋„๊ตฌ๋ฅผ ํƒ์ƒ‰ํ•˜๊ณ  ์กฐ์ž‘ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๋‹ค์–‘ํ•œ ์‹ค์ œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์žˆ์–ด ์ž์œจ ํƒ์ƒ‰์˜ ์ž ์žฌ๋ ฅ์„ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In this section, we provide additional details regarding our analysis of the comments provided by annotators on the summary claims. Table 20 features examples of such comments. These comments were further annotated based on the criteria outlined in Table 21 and Table 22. The distribution of errors is depicted in Figure 9 and Table 23.
์ด ์„น์…˜์—์„œ๋Š” ์š”์•ฝ ์ฃผ์žฅ์— ๋Œ€ํ•ด ์ฃผ์„์ž๋“ค์ด ์ œ๊ณตํ•œ ์˜๊ฒฌ์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ๋ถ„์„ ์„ธ๋ถ€ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 20์—๋Š” ์ด๋Ÿฌํ•œ ์˜๊ฒฌ์˜ ์˜ˆ์‹œ๊ฐ€ ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์˜๊ฒฌ์€ ํ‘œ 21๊ณผ ํ‘œ 22์— ๋ช…์‹œ๋œ ๊ธฐ์ค€์— ๋”ฐ๋ผ ์ถ”๊ฐ€๋กœ ์ฃผ์„์ด ๋‹ฌ๋ ธ์Šต๋‹ˆ๋‹ค. ์˜ค๋ฅ˜์˜ ๋ถ„ํฌ๋Š” ๊ทธ๋ฆผ 9์™€ ํ‘œ 23์— ๋‚˜ํƒ€๋‚˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ์ฃผ์„ ์ž‘์„ฑ์ž๊ฐ€ ์š”์•ฝ ํด๋ ˆ์ž„์— ์ œ๊ณตํ•œ ์˜๊ฒฌ์— ๋Œ€ํ•œ ๋ถ„์„์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์„ธ๋ถ€ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 20์—๋Š” ์ด๋Ÿฌํ•œ ์˜๊ฒฌ์˜ ์˜ˆ๊ฐ€ ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์˜๊ฒฌ์€ ํ‘œ 21 ๋ฐ ํ‘œ 22์— ์„ค๋ช…๋œ ๊ธฐ์ค€์— ๋”ฐ๋ผ ์ถ”๊ฐ€๋กœ ์ฃผ์„ ์ฒ˜๋ฆฌ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ค๋ฅ˜์˜ ๋ถ„ํฌ๋Š” ๊ทธ๋ฆผ 9์™€ ํ‘œ 23์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
To create a set of retrieval dates for each question, we use geometrically increasing time points between the open and close dates. We choose this schedule for two reasons: (1) questions are often most active shortly after they open, and (2) some questions have overly conservative close dates that are long after the question resolves.
๊ฐ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๊ฒ€์ƒ‰ ๋‚ ์งœ ์„ธํŠธ๋ฅผ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์—ด๋ฆผ ๋‚ ์งœ์™€ ๋‹ซํž˜ ๋‚ ์งœ ์‚ฌ์ด์˜ ๊ธฐํ•˜๊ธ‰์ˆ˜์ ์œผ๋กœ ์ฆ๊ฐ€ํ•˜๋Š” ์‹œ๊ฐ„ ์ง€์ ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋‘ ๊ฐ€์ง€ ์ด์œ ๋กœ ์ด ์ผ์ •์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค: (1) ์งˆ๋ฌธ์€ ์ข…์ข… ์—ด๋ฆผ ์งํ›„์— ๊ฐ€์žฅ ํ™œ๋ฐœํ•˜๊ฒŒ ํ™œ๋™ํ•˜๋ฉฐ, (2) ์ผ๋ถ€ ์งˆ๋ฌธ์€ ์งˆ๋ฌธ์ด ํ•ด๊ฒฐ๋œ ํ›„์—๋„ ๋„ˆ๋ฌด ๋ณด์ˆ˜์ ์ธ ๋‹ซํž˜ ๋‚ ์งœ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฐ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๊ฒ€์ƒ‰ ๋‚ ์งœ ์ง‘ํ•ฉ์„ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด ๊ฐœ์‹œ์ผ๊ณผ ์ข…๋ฃŒ์ผ ์‚ฌ์ด์— ๊ธฐํ•˜ํ•™์ ์œผ๋กœ ์ฆ๊ฐ€ํ•˜๋Š” ์‹œ์ ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์ผ์ •์„ ์„ ํƒํ•œ ์ด์œ ๋Š” ๋‘ ๊ฐ€์ง€์ž…๋‹ˆ๋‹ค: (1) ์งˆ๋ฌธ์€ ๊ฐœ์„ค ํ›„ ๊ณง๋ฐ”๋กœ ๊ฐ€์žฅ ํ™œ๋ฐœํ•˜๊ฒŒ ์ง„ํ–‰๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ, (2) ์ผ๋ถ€ ์งˆ๋ฌธ์€ ์งˆ๋ฌธ์ด ํ•ด๊ฒฐ๋œ ํ›„ ์˜ค๋ž˜ ์ง€๋‚œ ๊ณผ์ž‰ ๋ณด์ˆ˜์ ์ธ ์ข…๋ฃŒ์ผ์„ ๊ฐ–๋Š” ๊ฒฝ์šฐ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
We conduct the same analysis on \(\mathsf{bioS}^{\mathsf{simple}}\) and \(\mathsf{bioR}\). Recall from Section 2.2, \(\mathsf{bioS}^{\mathsf{simple}}\) is a variant of \(\mathsf{bioS}\) with reduced text diversity (one biography per person), while \(\mathsf{bioR}\) is generated by LLaMA2, resulting in close-to-real human biographies.
์šฐ๋ฆฌ๋Š” \(\mathsf{bioS}^{\mathsf{simple}}\)์™€ \(\mathsf{bioR}\)์— ๋Œ€ํ•ด ๋™์ผํ•œ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. 2.2์ ˆ์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, \(\mathsf{bioS}^{\mathsf{simple}}\)๋Š” ํ…์ŠคํŠธ ๋‹ค์–‘์„ฑ์ด ์ค„์–ด๋“  \(\mathsf{bioS}\)์˜ ๋ณ€ํ˜• ๋ฒ„์ „์œผ๋กœ, ํ•œ ์‚ฌ๋žŒ๋‹น ํ•˜๋‚˜์˜ ์ „๊ธฐ๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, \(\mathsf{bioR}\)์€ LLaMA2์— ์˜ํ•ด ์ƒ์„ฑ๋˜์–ด ์‹ค์ œ ์ธ๊ฐ„ ์ „๊ธฐ์— ๊ฐ€๊นŒ์šด ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์šฐ๋ฆฌ๋Š” \(\mathsf{bioS}^{\mathsf{simple}}\)์™€ \(\mathsf{bioR}\)์— ๋Œ€ํ•ด ๋™์ผํ•œ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. 2.2์ ˆ์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, \(\mathsf{bioS}^{\mathsf{simple}}\)๋Š” ํ…์ŠคํŠธ ๋‹ค์–‘์„ฑ์ด ๊ฐ์†Œ๋œ \(\mathsf{bioS}\)์˜ ๋ณ€ํ˜•์ด๋ฉฐ(1์ธ๋‹น 1๊ฐœ์˜ ์ „๊ธฐ), \(\mathsf{bioR}\)์€ LLaMA2์— ์˜ํ•ด ์ƒ์„ฑ๋˜์–ด ์‹ค์ œ ์ธ๊ฐ„์˜ ์ „๊ธฐ์™€ ๊ฑฐ์˜ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
language models, which have led to the emergence of a variety of popular models, including LLaVA [149], MiniGPT-4 [150], InstructBLIP [151], and PandaGPT [152]. The release of LLMA has greatly advanced the research progress of LLMs. To summarize the research work conducted on LLaMA, we present a brief evolutionary graph in Figure 5.
์–ธ์–ด ๋ชจ๋ธ์€ LLaVA [149], MiniGPT-4 [150], InstructBLIP [151], PandaGPT [152]๋ฅผ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์ธ๊ธฐ ๋ชจ๋ธ์˜ ์ถœํ˜„์„ ์ด๋Œ์—ˆ์Šต๋‹ˆ๋‹ค. LLMA์˜ ์ถœ์‹œ๋กœ LLMs ์—ฐ๊ตฌ๊ฐ€ ํฌ๊ฒŒ ์ง„์ „๋˜์—ˆ์Šต๋‹ˆ๋‹ค. LLaMA์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ์ž‘์—…์„ ์š”์•ฝํ•˜๊ธฐ ์œ„ํ•ด, Figure 5์— ๊ฐ„๋žตํ•œ ์ง„ํ™” ๊ทธ๋ž˜ํ”„๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
LLaVA [149], MiniGPT-4 [150], InstructBLIP [151], PandaGPT [152] ๋“ฑ ๋‹ค์–‘ํ•œ ์ธ๊ธฐ ๋ชจ๋ธ์˜ ์ถœํ˜„์œผ๋กœ ์ด์–ด์กŒ์Šต๋‹ˆ๋‹ค. LLMA์˜ ์ถœ์‹œ๋กœ LLM์˜ ์—ฐ๊ตฌ ์ง„์ „์ด ํฌ๊ฒŒ ์ง„์ „๋˜์—ˆ์Šต๋‹ˆ๋‹ค. LLaMA์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ์ž‘์—…์„ ์š”์•ฝํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 5์— ๊ฐ„๋žตํ•œ ์ง„ํ™” ๊ทธ๋ž˜ํ”„๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
While entity substitution results do not show increased multi-hop reasoning with model size (Figure 2(d)), relation substitution exhibits a scaling trend. From 7B to 70B, the maximum relative frequency increases from 0.38 to 0.43, suggesting that larger models may facilitate multi-hop reasoning with relational changes (Figure 2(e)).
์—”ํ‹ฐํ‹ฐ ๋Œ€์ฒด ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ ํฌ๊ธฐ์™€ ํ•จ๊ป˜ ๋ฉ€ํ‹ฐ-ํ™‰ ์ถ”๋ก ์ด ์ฆ๊ฐ€ํ•˜์ง€ ์•Š์Œ์„ ๋ณด์—ฌ์ฃผ์ง€๋งŒ(Figure 2(d)), ๊ด€๊ณ„ ๋Œ€์ฒด๋Š” ํ™•์žฅ ๊ฒฝํ–ฅ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. 7B์—์„œ 70B๋กœ ์ตœ๋Œ€ ์ƒ๋Œ€ ๋นˆ๋„๋Š” 0.38์—์„œ 0.43์œผ๋กœ ์ฆ๊ฐ€ํ•˜์—ฌ, ๋” ํฐ ๋ชจ๋ธ์ด ๊ด€๊ณ„์  ๋ณ€ํ™”๋ฅผ ํ†ตํ•ด ๋ฉ€ํ‹ฐ-ํ™‰ ์ถ”๋ก ์„ ์ด‰์ง„ํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค(Figure 2(e)).
์—”ํ‹ฐํ‹ฐ ๋Œ€์ฒด ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ ํฌ๊ธฐ์— ๋”ฐ๋ผ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์ด ์ฆ๊ฐ€ํ•˜์ง€ ์•Š๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์ง€๋งŒ(๊ทธ๋ฆผ 2(d)), ๊ด€๊ณ„ ๋Œ€์ฒด๋Š” ํ™•์žฅ ์ถ”์„ธ๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค. 7B์—์„œ 70B๊นŒ์ง€ ์ตœ๋Œ€ ์ƒ๋Œ€ ๋นˆ๋„๋Š” 0.38์—์„œ 0.43์œผ๋กœ ์ฆ๊ฐ€ํ•˜์—ฌ ๋” ํฐ ๋ชจ๋ธ์ด ๊ด€๊ณ„ ๋ณ€๊ฒฝ์„ ํ†ตํ•œ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์ด‰์ง„ํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 2(e)).
One day, Sophie finally cracked the sphere open. Inside, she found a glowing crystal, surrounded by a strange, swirling energy. As she touched the crystal, she felt a rush of power surge through her body. She knew that she had to be careful, but she also knew that she couldn't let go of this opportunity to unlock her true potential.
์–ด๋Š ๋‚ , ์†Œํ”ผ๋Š” ๋งˆ์นจ๋‚ด ๊ตฌ์ฒด๋ฅผ ์—ด์—ˆ๋‹ค. ๊ทธ ์•ˆ์—๋Š” ๋น›๋‚˜๋Š” ์ˆ˜์ •์ด ์žˆ์—ˆ๊ณ , ์ด์ƒํ•œ ์†Œ์šฉ๋Œ์ด์น˜๋Š” ์—๋„ˆ์ง€๊ฐ€ ๊ทธ๊ฒƒ์„ ๋‘˜๋Ÿฌ์‹ธ๊ณ  ์žˆ์—ˆ๋‹ค. ์†Œํ”ผ๊ฐ€ ์ˆ˜์ •์— ์†์„ ๋Œ€์ž, ๊ทธ๋…€์˜ ๋ชธ์„ ํ†ตํ•ด ๊ฐ•๋ ฅํ•œ ํž˜์ด ์†Ÿ๊ตฌ์น˜๋Š” ๊ฒƒ์„ ๋Š๊ผˆ๋‹ค. ๊ทธ๋…€๋Š” ์กฐ์‹ฌํ•ด์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ์•˜์ง€๋งŒ, ์ž์‹ ์˜ ์ง„์ •ํ•œ ์ž ์žฌ๋ ฅ์„ ์—ด ์ˆ˜ ์žˆ๋Š” ์ด ๊ธฐํšŒ๋ฅผ ๋†“์น  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ๋„ ์•Œ๊ณ  ์žˆ์—ˆ๋‹ค.
์–ด๋Š ๋‚  ์†Œํ”ผ๋Š” ๋งˆ์นจ๋‚ด ๊ตฌ์ฒด๋ฅผ ๊นจ๋œจ๋ ธ์Šต๋‹ˆ๋‹ค. ๊ทธ ์•ˆ์—์„œ ์ด์ƒํ•˜๊ฒŒ ์†Œ์šฉ๋Œ์ด์น˜๋Š” ์—๋„ˆ์ง€์— ๋‘˜๋Ÿฌ์‹ธ์ธ ๋น›๋‚˜๋Š” ์ˆ˜์ •์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ˆ˜์ •์„ ๋งŒ์ž์ž ๋ชธ ์ „์ฒด์— ํž˜์ด ์†Ÿ๊ตฌ์น˜๋Š” ๊ฒƒ์„ ๋Š๊ผˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ์กฐ์‹ฌํ•ด์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ์•˜์ง€๋งŒ, ์ž์‹ ์˜ ์ง„์ •ํ•œ ์ž ์žฌ๋ ฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐํšŒ๋ฅผ ๋†“์น  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ๋„ ์•Œ์•˜์Šต๋‹ˆ๋‹ค.
Comparing with reinforcement learning, LLM-based agents have more comprehensive internal world knowledge, which facilitates more informed agent actions even without training on specific domain data. Additionally, LLM-based agents can provide natural language interfaces to interact with humans, which is more flexible and explainable.
๊ฐ•ํ™” ํ•™์Šต๊ณผ ๋น„๊ตํ•  ๋•Œ, LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ๋” ํฌ๊ด„์ ์ธ ๋‚ด๋ถ€ ์„ธ๊ณ„ ์ง€์‹์„ ๊ฐ€์ง€๊ณ  ์žˆ์–ด ํŠน์ • ๋„๋ฉ”์ธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ ์—†์ด๋„ ๋” ์ •๋ณด์— ์ž…๊ฐํ•œ ์—์ด์ „ํŠธ ํ–‰๋™์„ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ์ธ๊ฐ„๊ณผ ์ƒํ˜ธ์ž‘์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ž์—ฐ์–ด ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์–ด ๋” ์œ ์—ฐํ•˜๊ณ  ์„ค๋ช… ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
๊ฐ•ํ™” ํ•™์Šต๊ณผ ๋น„๊ตํ•  ๋•Œ, LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ๋‚ด๋ถ€ ์„ธ๊ณ„ ์ง€์‹์„ ๊ฐ€์ง€๊ณ  ์žˆ์–ด ํŠน์ • ๋„๋ฉ”์ธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ•™์Šต ์—†์ด๋„ ๋ณด๋‹ค ๋งŽ์€ ์ •๋ณด์— ์ž…๊ฐํ•œ ์—์ด์ „ํŠธ ํ–‰๋™์„ ์šฉ์ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ์ธ๊ฐ„๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ž์—ฐ์–ด ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์–ด ๋ณด๋‹ค ์œ ์—ฐํ•˜๊ณ  ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
provides a mathematical representation of how the performance (\(P\)) of machine learning models scales with respect to three critical factors: model size (\(N\)), data size (\(D\)), and computational budget (\(C\)). Each of these components plays a crucial role in the model's ability to learn from data and perform tasks accurately.
๊ธฐ๊ณ„ ํ•™์Šต ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ (\(P\))์ด ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์š”์†Œ์ธ ๋ชจ๋ธ ํฌ๊ธฐ (\(N\)), ๋ฐ์ดํ„ฐ ํฌ๊ธฐ (\(D\)), ๊ทธ๋ฆฌ๊ณ  ๊ณ„์‚ฐ ์˜ˆ์‚ฐ (\(C\))์— ๋”ฐ๋ผ ์–ด๋–ป๊ฒŒ ํ™•์žฅ๋˜๋Š”์ง€๋ฅผ ์ˆ˜ํ•™์ ์œผ๋กœ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋“ค ๊ฐ๊ฐ์˜ ์š”์†Œ๋Š” ๋ชจ๋ธ์ด ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๊ณ  ์ •ํ™•ํ•˜๊ฒŒ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋Šฅ๋ ฅ์— ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค.
๋จธ์‹  ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ(\(P\))์ด ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์š”์†Œ(๋ชจ๋ธ ํฌ๊ธฐ(\(N\)), ๋ฐ์ดํ„ฐ ํฌ๊ธฐ(\(D\)), ๊ณ„์‚ฐ ์˜ˆ์‚ฐ(\(C\))์— ๋”ฐ๋ผ ์–ด๋–ป๊ฒŒ ํ™•์žฅ๋˜๋Š”์ง€๋ฅผ ์ˆ˜ํ•™์ ์œผ๋กœ ํ‘œํ˜„ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ๋ฐ์ดํ„ฐ์—์„œ ํ•™์Šตํ•˜๊ณ  ์ž‘์—…์„ ์ •ํ™•ํ•˜๊ฒŒ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์— ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค.
**Output** [[Retriev=Yes] <p>Reference [1]<p>I [IsRel] =Relevant] Sergei Vladimirovich Bodrov (born June 28, 1948) is a Russian film director, screenwriter, and producer.[[IsNew] =Fully Supported[[Retriev=Continue]He was born in Moscow, Russia.[[Retriev=Continue]He is the president of the Russian Guild of Film Producers.[IsNew] =5]
**Output** [[Retriev=Yes] <p>Reference [1]<p>I [IsRel] =Relevant] ์„ธ๋ฅด๊ฒŒ์ด ๋ธ”๋ผ๋””๋ฏธ๋กœ๋น„์น˜ ๋ณด๋“œ๋กœํ”„(1948๋…„ 6์›” 28์ผ ์ถœ์ƒ)๋Š” ๋Ÿฌ์‹œ์•„์˜ ์˜ํ™” ๊ฐ๋…, ์‹œ๋‚˜๋ฆฌ์˜ค ์ž‘๊ฐ€, ํ”„๋กœ๋“€์„œ์ž…๋‹ˆ๋‹ค.[[IsNew] =์™„์ „ํžˆ ์ง€์›๋จ[[Retriev=Continue]๊ทธ๋Š” ๋Ÿฌ์‹œ์•„ ๋ชจ์Šคํฌ๋ฐ”์—์„œ ํƒœ์–ด๋‚ฌ์Šต๋‹ˆ๋‹ค.[[Retriev=Continue]๊ทธ๋Š” ๋Ÿฌ์‹œ์•„ ์˜ํ™” ์ œ์ž‘์ž ๊ธธ๋“œ์˜ ํšŒ์žฅ์ž…๋‹ˆ๋‹ค.[IsNew] =5]
**์ถœ๋ ฅ** [[Retriev=Yes] <p>์ฐธ์กฐ [1]<p>I [IsRel] =๊ด€๋ จ] ์„ธ๋ฅด๊ฒŒ์ด ๋ธ”๋ผ๋””๋ฏธ๋กœ๋น„์น˜ ๋ณด๋“œ๋กœํ”„(1948๋…„ 6์›” 28์ผ์ƒ)๋Š” ๋Ÿฌ์‹œ์•„ ์˜ํ™” ๊ฐ๋…, ์‹œ๋‚˜๋ฆฌ์˜ค ์ž‘๊ฐ€, ํ”„๋กœ๋“€์„œ์ž…๋‹ˆ๋‹ค.[[IsNew] =Fully Supported[[Retriev=Continue]๊ทธ๋Š” ๋Ÿฌ์‹œ์•„ ๋ชจ์Šคํฌ๋ฐ”์—์„œ ํƒœ์–ด๋‚ฌ์Šต๋‹ˆ๋‹ค.[[Retriev=Continue]๊ทธ๋Š” ๋Ÿฌ์‹œ์•„ ์˜ํ™” ์ œ์ž‘์ž ๊ธธ๋“œ์˜ ํšŒ์žฅ์ž…๋‹ˆ๋‹ค.[IsNew] =5]
Figure 1: Overview of the Continual Knowledge Learning benchmark. InvariantLAMA is used to measure the _time-invariant_ world knowledge gained from \(D_{0}\). UpdatedLAMA is used to measure the _update_ of world knowledge from \(D_{0}\)\(\rightarrow\)\(D_{1}\). NewLAMA is used to measure _new_ world knowledge gained from \(D_{1}\).
Figure 1: Continual Knowledge Learning ๋ฒค์น˜๋งˆํฌ์˜ ๊ฐœ์š”. InvariantLAMA๋Š” \(D_{0}\)์—์„œ ์–ป์€ _์‹œ๊ฐ„ ๋ถˆ๋ณ€_ ์„ธ๊ณ„ ์ง€์‹์„ ์ธก์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. UpdatedLAMA๋Š” \(D_{0}\)\(\rightarrow\)\(D_{1}\)๋กœ๋ถ€ํ„ฐ ์„ธ๊ณ„ ์ง€์‹์˜ _์—…๋ฐ์ดํŠธ_๋ฅผ ์ธก์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. NewLAMA๋Š” \(D_{1}\)์—์„œ ์–ป์€ _์ƒˆ๋กœ์šด_ ์„ธ๊ณ„ ์ง€์‹์„ ์ธก์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1: ์ง€์†์  ์ง€์‹ ํ•™์Šต ๋ฒค์น˜๋งˆํฌ ๊ฐœ์š”. InvariantLAMA๋Š” \(D_{0}\)์—์„œ ์–ป์€ _์‹œ๊ฐ„ ๋ถˆ๋ณ€_ ์„ธ๊ณ„ ์ง€์‹์„ ์ธก์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. UpdatedLAMA๋Š” \(D_{0}\)\(\rightarrow\)\(D_{1}\)์—์„œ ์„ธ๊ณ„ ์ง€์‹์˜ _์—…๋ฐ์ดํŠธ_๋ฅผ ์ธก์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. NewLAMA๋Š” \(D_{1}\)์—์„œ ์–ป์€ _์ƒˆ๋กœ์šด_ ์„ธ๊ณ„ ์ง€์‹์„ ์ธก์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
In the following experiments (SS 4 and SS 5), we use Llama-2 (7B and 70B) [11] since it is one of the best-performing LLMs. We use Wikipedia articles classified under the "2023" Category including topics from diverse domains such as films, arts, economics, politics, events, etc.3 The likelihood that this factual information is not
๋‹ค์Œ ์‹คํ—˜๋“ค(SS 4 ๋ฐ SS 5)์—์„œ๋Š” ์ตœ๊ณ ์˜ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” LLM ์ค‘ ํ•˜๋‚˜์ธ Llama-2 (7B ๋ฐ 70B) [11]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์˜ํ™”, ์˜ˆ์ˆ , ๊ฒฝ์ œ, ์ •์น˜, ์‚ฌ๊ฑด ๋“ฑ ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์˜ ์ฃผ์ œ๋ฅผ ํฌํ•จํ•˜๋Š” "2023" ์นดํ…Œ๊ณ ๋ฆฌ๋กœ ๋ถ„๋ฅ˜๋œ ์œ„ํ‚คํ”ผ๋””์•„ ๊ธฐ์‚ฌ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
๋‹ค์Œ ์‹คํ—˜(SS 4 ๋ฐ SS 5)์—์„œ๋Š” ์„ฑ๋Šฅ์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ LLM ์ค‘ ํ•˜๋‚˜์ด๊ธฐ ๋•Œ๋ฌธ์— Llama-2(7B ๋ฐ 70B) [11]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ํ™”, ์˜ˆ์ˆ , ๊ฒฝ์ œ, ์ •์น˜, ์ด๋ฒคํŠธ ๋“ฑ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์˜ ์ฃผ์ œ๋ฅผ ํฌํ•จํ•˜์—ฌ "2023" ์นดํ…Œ๊ณ ๋ฆฌ๋กœ ๋ถ„๋ฅ˜๋œ ์œ„ํ‚ค๋ฐฑ๊ณผ ๊ธฐ์‚ฌ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.3 ์ด๋Ÿฌํ•œ ์‚ฌ์‹ค ์ •๋ณด๊ฐ€ ์•„๋‹ ๊ฐ€๋Šฅ์„ฑ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
Duplicate text in training data.The Book Corpus Zhu et al. (2015), which was used to train popular models such as BERT, has a substantial amount of exact-duplicate documents according to Bandy and Vincent (2021). Allamanis (2019) shows that duplicate examples in code datasets cause worsened performance on code understanding tasks.
ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์—์„œ ์ค‘๋ณต๋œ ํ…์ŠคํŠธ. BERT์™€ ๊ฐ™์€ ์ธ๊ธฐ ์žˆ๋Š” ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๋ฐ ์‚ฌ์šฉ๋œ Book Corpus Zhu et al. (2015)์€ Bandy์™€ Vincent (2021)์— ๋”ฐ๋ฅด๋ฉด ์ƒ๋‹นํ•œ ์–‘์˜ ์ •ํ™•ํžˆ ์ค‘๋ณต๋œ ๋ฌธ์„œ๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Allamanis (2019)๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์…‹์—์„œ ์ค‘๋ณต๋œ ์˜ˆ์ œ๊ฐ€ ์ฝ”๋“œ ์ดํ•ด ์ž‘์—…์˜ ์„ฑ๋Šฅ์„ ์ €ํ•˜์‹œํ‚จ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ์ค‘๋ณต ํ…์ŠคํŠธ.BERT์™€ ๊ฐ™์€ ์ธ๊ธฐ ์žˆ๋Š” ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ Book Corpus Zhu ๋“ฑ(2015)์€ Bandy์™€ Vincent(2021)์— ๋”ฐ๋ฅด๋ฉด ์ƒ๋‹นํ•œ ์–‘์˜ ์ •ํ™•ํ•œ ์ค‘๋ณต ๋ฌธ์„œ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Allamanis(2019)๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ค‘๋ณต ์˜ˆ์ œ๊ฐ€ ์ฝ”๋“œ ์ดํ•ด ์ž‘์—…์˜ ์„ฑ๋Šฅ ์ €ํ•˜๋ฅผ ์œ ๋ฐœํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.