question
stringlengths
101
2.69k
chosen
stringlengths
43
2.27k
rejected
stringlengths
42
2.18k
Figure 2 gives an overview of the five models' "answer sheet" on the math problems. Our merged models retain the foundational knowledge in the source models, as evidenced by the similar score patterns on problems 1-15. Moreover, they exhibit emergent capabilities, successfully tackling problems that stumped the source models (e.g., problems 20-30). Evidently, by effectively integrating a Japanese LLM and mathematical models, we have succeeded in producing models that are proficient in both Japanese language understanding and mathematical problem-solving.
Figure 2๋Š” ์ˆ˜ํ•™ ๋ฌธ์ œ์— ๋Œ€ํ•œ ๋‹ค์„ฏ ๋ชจ๋ธ์˜ "๋‹ต์•ˆ์ง€" ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ํ†ตํ•ฉ ๋ชจ๋ธ์€ ๋ฌธ์ œ 1-15์—์„œ ์œ ์‚ฌํ•œ ์ ์ˆ˜ ํŒจํ„ด์„ ๋ณด์ž„์œผ๋กœ์จ ์†Œ์Šค ๋ชจ๋ธ์˜ ๊ธฐ์ดˆ ์ง€์‹์„ ์œ ์ง€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”์šฑ์ด, ์ด๋“ค์€ ์†Œ์Šค ๋ชจ๋ธ์ด ํ•ด๊ฒฐํ•˜์ง€ ๋ชปํ•œ ๋ฌธ์ œ๋“ค(์˜ˆ: ๋ฌธ์ œ 20-30)์„ ์„ฑ๊ณต์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ถ„๋ช…ํžˆ, ์ผ๋ณธ์–ด LLM๊ณผ ์ˆ˜ํ•™ ๋ชจ๋ธ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ†ตํ•ฉํ•จ์œผ๋กœ์จ, ์šฐ๋ฆฌ๋Š” ์ผ๋ณธ์–ด ์ดํ•ด์™€ ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ์— ๋ชจ๋‘ ๋Šฅ์ˆ™ํ•œ ๋ชจ๋ธ์„ ๋งŒ๋“œ๋Š” ๋ฐ ์„ฑ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2๋Š” ์ˆ˜ํ•™ ๋ฌธ์ œ์— ๋Œ€ํ•œ 5๊ฐ€์ง€ ๋ชจ๋ธ์˜ "๋‹ต์•ˆ์ง€"์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋ณ‘ํ•ฉ๋œ ๋ชจ๋ธ์€ ๋ฌธ์ œ 1~15์—์„œ ์œ ์‚ฌํ•œ ์ ์ˆ˜ ํŒจํ„ด์—์„œ ์•Œ ์ˆ˜ ์žˆ๋“ฏ์ด ์†Œ์Šค ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ์ง€์‹์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์†Œ์Šค ๋ชจ๋ธ์ด ํ’€์ง€ ๋ชปํ•œ ๋ฌธ์ œ(์˜ˆ: ๋ฌธ์ œ 20~30)๋ฅผ ์„ฑ๊ณต์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜์—ฌ ์ƒˆ๋กœ์šด ์—ญ๋Ÿ‰์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ๋ถ„๋ช…ํžˆ ์ผ๋ณธ์–ด LLM๊ณผ ์ˆ˜ํ•™์  ๋ชจ๋ธ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ†ตํ•ฉํ•จ์œผ๋กœ์จ ์ผ๋ณธ์–ด ์ดํ•ด์™€ ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ์— ๋ชจ๋‘ ๋Šฅ์ˆ™ํ•œ ๋ชจ๋ธ์„ ์ œ์ž‘ํ•˜๋Š” ๋ฐ ์„ฑ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค.
In the language modeling field, the current trend consists in scaling up the generative pretraining approach introduced with GPT-2, which implies training neural models made of several billions of parameters on gigantic web-mined text corpora (Brown et al., 2020; Touvron et al., 2023; Almazrouei et al., 2023; Jiang et al., 2023). However, training and serving such highly parameterized models raises energy and hardware-related problematics, which motivates for looking into achieving similar performance levels with smaller models (Sardana & Frankle, 2023).
์–ธ์–ด ๋ชจ๋ธ๋ง ๋ถ„์•ผ์—์„œ ํ˜„์žฌ์˜ ํŠธ๋ Œ๋“œ๋Š” GPT-2์™€ ํ•จ๊ป˜ ๋„์ž…๋œ ์ƒ์„ฑ์  ์‚ฌ์ „ ํ•™์Šต ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋กœ ๊ตฌ์„ฑ๋œ ์‹ ๊ฒฝ ๋ชจ๋ธ์„ ๊ฑฐ๋Œ€ํ•œ ์›น์—์„œ ์ˆ˜์ง‘ํ•œ ํ…์ŠคํŠธ ์ฝ”ํผ์Šค์— ๋Œ€ํ•ด ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค (Brown et al., 2020; Touvron et al., 2023; Almazrouei et al., 2023; Jiang et al., 2023). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๊ณ ๋„๋กœ ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ ๋งŽ์€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ์„œ๋น„์Šคํ•˜๋Š” ๊ฒƒ์€ ์—๋„ˆ์ง€ ๋ฐ ํ•˜๋“œ์›จ์–ด ๊ด€๋ จ ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•˜๋ฉฐ, ์ด๋Š” ๋” ์ž‘์€ ๋ชจ๋ธ๋กœ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ ์ˆ˜์ค€์„ ๋‹ฌ์„ฑํ•˜๋ ค๋Š” ๋™๊ธฐ๋ฅผ ๋ถ€์—ฌํ•ฉ๋‹ˆ๋‹ค (Sardana & Frankle, 2023).
์–ธ์–ด ๋ชจ๋ธ๋ง ๋ถ„์•ผ์—์„œ๋Š” GPT-2์™€ ํ•จ๊ป˜ ๋„์ž…๋œ ์ƒ์„ฑ์  ์‚ฌ์ „ ํ•™์Šต ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์ด ํ˜„์žฌ ์ถ”์„ธ์ด๋ฉฐ, ์ด๋Š” ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ ์‹ ๊ฒฝ ๋ชจ๋ธ์„ ๊ฑฐ๋Œ€ํ•œ ์›น ๋งˆ์ด๋‹ ํ…์ŠคํŠธ ์ฝ”ํผ์Šค(Brown et al., 2020; Touvron et al., 2023; Almazrouei et al., 2023; Jiang et al., 2023)์— ๋Œ€ํ•ด ํ•™์Šต์‹œํ‚ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๊ณ ๋งค๊ฐœ๋ณ€์ˆ˜ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ์„œ๋น„์Šคํ•˜๋ฉด ์—๋„ˆ์ง€ ๋ฐ ํ•˜๋“œ์›จ์–ด ๊ด€๋ จ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋” ์ž‘์€ ๋ชจ๋ธ๋กœ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ ์ˆ˜์ค€์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค(Sardana & Frankle, 2023).
In addition, recent work also explores multi-agent collaboration based on LLMs in simulated environments [700, 701, 533]. These studies simulate human social behaviors by instantiating multiple LLM-based agents with observations, planning, and memories in a sandbox environment. In controlled evaluation, the abilities of generative agents to search, plan, and think are evaluated by humans in an interview-like manner. Further, they also conduct descriptive measurements on multiple agents within a simulated environment to examine emergent social behaviors.
๋˜ํ•œ, ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ํ™˜๊ฒฝ์—์„œ LLMs ๊ธฐ๋ฐ˜์˜ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ํ˜‘์—…์„ ํƒ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค [700, 701, 533]. ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ๋“ค์€ ์ƒŒ๋“œ๋ฐ•์Šค ํ™˜๊ฒฝ์—์„œ ๊ด€์ฐฐ, ๊ณ„ํš ๋ฐ ๊ธฐ์–ต์„ ๊ฐ€์ง„ ๋‹ค์ค‘ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋ฅผ ๊ตฌํ˜„ํ•˜์—ฌ ์ธ๊ฐ„์˜ ์‚ฌํšŒ์  ํ–‰๋™์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค. ํ†ต์ œ๋œ ํ‰๊ฐ€์—์„œ๋Š” ์ธํ„ฐ๋ทฐ์™€ ์œ ์‚ฌํ•œ ๋ฐฉ์‹์œผ๋กœ ์ƒ์„ฑ์  ์—์ด์ „ํŠธ์˜ ๊ฒ€์ƒ‰, ๊ณ„ํš ๋ฐ ์‚ฌ๊ณ  ๋Šฅ๋ ฅ์„ ์ธ๊ฐ„์ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋” ๋‚˜์•„๊ฐ€, ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ํ™˜๊ฒฝ ๋‚ด์—์„œ ๋‹ค์ค‘ ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ๊ธฐ์ˆ ์  ์ธก์ •์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ๋ฐœ์ƒํ•˜๋Š” ์‚ฌํšŒ์  ํ–‰๋™์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
๋˜ํ•œ ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์—์„œ LLM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ํ˜‘์—…์„ ํƒ๊ตฌํ•˜๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค[700, 701, 533]. ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ์—์„œ๋Š” ์ƒŒ๋“œ๋ฐ•์Šค ํ™˜๊ฒฝ์—์„œ ๊ด€์ฐฐ, ๊ณ„ํš ๋ฐ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ํ†ตํ•ด ์—ฌ๋Ÿฌ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋ฅผ ์ธ์Šคํ„ด์Šคํ™”ํ•˜์—ฌ ์ธ๊ฐ„์˜ ์‚ฌํšŒ์  ํ–‰๋™์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค. ํ†ต์ œ๋œ ํ‰๊ฐ€์—์„œ ์ƒ์„ฑ ์—์ด์ „ํŠธ์˜ ๊ฒ€์ƒ‰, ๊ณ„ํš ๋ฐ ์‚ฌ๊ณ  ๋Šฅ๋ ฅ์€ ์ธํ„ฐ๋ทฐ์™€ ๊ฐ™์€ ๋ฐฉ์‹์œผ๋กœ ์ธ๊ฐ„์— ์˜ํ•ด ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ ๋‚ด์—์„œ ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ์„ค๋ช…์  ์ธก์ •์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ƒˆ๋กœ์šด ์‚ฌํšŒ์  ํ–‰๋™์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
Text extraction.We want to extract only the main content of the page, ignoring menus, headers, footers, and ads among others: Lopukhin (2019) found that trafilatura (Barbaresi, 2021) was the best non-commercial library for retrieving content from blog posts and news articles. Although this is only a narrow subset of the kind of pages making up CommonCrawl, we found this finding to hold more broadly. We use trafilatura for text extraction, and apply extra formatting via regular expressions: we limit new lines to two consecutive ones, and remove all URLs.
ํ…์ŠคํŠธ ์ถ”์ถœ. ์šฐ๋ฆฌ๋Š” ๋ฉ”๋‰ด, ํ—ค๋”, ํ‘ธํ„ฐ, ๊ด‘๊ณ  ๋“ฑ์„ ๋ฌด์‹œํ•˜๊ณ  ํŽ˜์ด์ง€์˜ ์ฃผ์š” ์ฝ˜ํ…์ธ ๋งŒ ์ถ”์ถœํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. Lopukhin (2019)์€ trafilatura (Barbaresi, 2021)๊ฐ€ ๋ธ”๋กœ๊ทธ ๊ฒŒ์‹œ๋ฌผ ๋ฐ ๋‰ด์Šค ๊ธฐ์‚ฌ์—์„œ ์ฝ˜ํ…์ธ ๋ฅผ ๊ฐ€์ ธ์˜ค๋Š” ๋ฐ ๊ฐ€์žฅ ์ข‹์€ ๋น„์ƒ์—…์  ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ผ๊ณ  ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋น„๋ก ์ด๊ฒƒ์ด CommonCrawl์„ ๊ตฌ์„ฑํ•˜๋Š” ํŽ˜์ด์ง€์˜ ์ข์€ ํ•˜์œ„ ์ง‘ํ•ฉ์— ๋ถˆ๊ณผํ•˜์ง€๋งŒ, ์šฐ๋ฆฌ๋Š” ์ด ๊ฒฐ๊ณผ๊ฐ€ ๋” ๋„“๊ฒŒ ์ ์šฉ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ํ…์ŠคํŠธ ์ถ”์ถœ์„ ์œ„ํ•ด trafilatura๋ฅผ ์‚ฌ์šฉํ•˜๊ณ , ์ •๊ทœ ํ‘œํ˜„์‹์„ ํ†ตํ•ด ์ถ”๊ฐ€ ํ˜•์‹์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค: ์ƒˆ๋กœ์šด ์ค„์„ ๋‘ ๊ฐœ์˜ ์—ฐ์†๋œ ์ค„๋กœ ์ œํ•œํ•˜๊ณ  ๋ชจ๋“  URL์„ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
ํ…์ŠคํŠธ ์ถ”์ถœ.๋ฉ”๋‰ด, ํ—ค๋”, ๋ฐ”๋‹ฅ๊ธ€, ๊ด‘๊ณ  ๋“ฑ์„ ๋ฌด์‹œํ•˜๊ณ  ํŽ˜์ด์ง€์˜ ์ฃผ์š” ์ฝ˜ํ…์ธ ๋งŒ ์ถ”์ถœํ•˜๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค: Lopukhin(2019)์€ ๋ธ”๋กœ๊ทธ ๊ฒŒ์‹œ๋ฌผ๊ณผ ๋‰ด์Šค ๊ธฐ์‚ฌ์—์„œ ์ฝ˜ํ…์ธ ๋ฅผ ๊ฒ€์ƒ‰ํ•˜๋Š” ๋ฐ ์žˆ์–ด ๋น„์ƒ์—…์šฉ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์ค‘ trafilatura(Barbaresi, 2021)๊ฐ€ ๊ฐ€์žฅ ์ข‹๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” CommonCrawl์„ ๊ตฌ์„ฑํ•˜๋Š” ํŽ˜์ด์ง€์˜ ์ผ๋ถ€์— ๋ถˆ๊ณผํ•˜์ง€๋งŒ, ์ด ๊ฒฐ๊ณผ๋Š” ๋” ํญ๋„“๊ฒŒ ์ ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ํ…์ŠคํŠธ ์ถ”์ถœ์—๋Š” trafilatura๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ์ •๊ทœ์‹์„ ํ†ตํ•ด ์ถ”๊ฐ€ ์„œ์‹์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์ƒˆ ์ค„์€ ์—ฐ์†๋œ ๋‘ ์ค„๋กœ ์ œํ•œํ•˜๊ณ  ๋ชจ๋“  URL์„ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
This work received partial support from the Basque Government through research group funding IT1805-22 and the ICL4LANG project (grant no. KK-2023/00094). Additionally, we acknowledge the support of several MCIN/AEI/10.13039/501100011033 projects: (i) DeepKnowledge (PID2021-127777/DB-C21) and funding from FEDER, EU; (ii) AWARE (TED2021-131617B-I00) and support from the European Union NextGenerationEU/PRTR. We express our gratitude to Carlos Dominguez for his assistance in the experimental setup and to Eneko Agirre for his valuable feedback and guidance.
์ด ์—ฐ๊ตฌ๋Š” Basque ์ •๋ถ€์˜ ์—ฐ๊ตฌ ๊ทธ๋ฃน ์ž๊ธˆ IT1805-22์™€ ICL4LANG ํ”„๋กœ์ ํŠธ (grant no. KK-2023/00094)๋ฅผ ํ†ตํ•ด ๋ถ€๋ถ„์ ์ธ ์ง€์›์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์—ฌ๋Ÿฌ MCIN/AEI/10.13039/501100011033 ํ”„๋กœ์ ํŠธ์˜ ์ง€์›์„ ์ธ์ •ํ•ฉ๋‹ˆ๋‹ค: (i) DeepKnowledge (PID2021-127777/DB-C21)์™€ FEDER, EU์˜ ์ž๊ธˆ ์ง€์›; (ii) AWARE (TED2021-131617B-I00)์™€ European Union NextGenerationEU/PRTR์˜ ์ง€์›. ์‹คํ—˜ ์„ค์ •์— ๋„์›€์„ ์ค€ Carlos Dominguez์™€ ๊ท€์ค‘ํ•œ ํ”ผ๋“œ๋ฐฑ๊ณผ ์ง€๋„๋ฅผ ์ œ๊ณตํ•œ Eneko Agirre์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋œป์„ ์ „ํ•ฉ๋‹ˆ๋‹ค.
์ด ์ž‘์—…์€ ์—ฐ๊ตฌ ๊ทธ๋ฃน ๊ธฐ๊ธˆ IT1805-22 ๋ฐ ICL4LANG ํ”„๋กœ์ ํŠธ(๋ณด์กฐ๊ธˆ ๋ฒˆํ˜ธ KK-2023/00094)๋ฅผ ํ†ตํ•ด ๋ฐ”์Šคํฌ ์ •๋ถ€์˜ ๋ถ€๋ถ„์ ์ธ ์ง€์›์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ MCIN/AEI/10.13039/501100011033 ํ”„๋กœ์ ํŠธ์˜ ์—ฌ๋Ÿฌ ์ง€์›์„ ์ธ์ •ํ•ฉ๋‹ˆ๋‹ค: (i) DeepKnowledge(PID2021-127777/DB-C21) ๋ฐ FEDER, EU์˜ ์ง€์›๊ธˆ, (ii) AWARE(TED2021-131617B-I00) ๋ฐ ์œ ๋Ÿฝ์—ฐํ•ฉ NextGenerationEU/PRTR์˜ ์ง€์›์„ ์ธ์ •ํ•ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ์„ค์ •์„ ๋„์™€์ค€ Carlos Dominguez์—๊ฒŒ ๊ฐ์‚ฌ๋“œ๋ฆฌ๋ฉฐ, ๊ท€์ค‘ํ•œ ํ”ผ๋“œ๋ฐฑ๊ณผ ์ง€๋„๋ฅผ ์ œ๊ณตํ•ด์ค€ Eneko Agirre์—๊ฒŒ๋„ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
Finally, we compare the mean tree depth (measured by the mean over sentences of the depth of the dependency tree), and mean dependency distance (measured as the average dependency distance of any pair of words within a sentence) in Figure 11, which have been shown to be good measures of syntactic difficulty Futrell et al. (2015); Gibson et al. (2000); Oya (2021). We find similar trends as for reading level and TTR diversity where mediums style increase depth, mdd, and syntactic complexity in general. We find again that QA style reduces this complexity.
๋งˆ์ง€๋ง‰์œผ๋กœ, Figure 11์—์„œ ํ‰๊ท  ํŠธ๋ฆฌ ๊นŠ์ด(์˜์กด ํŠธ๋ฆฌ์˜ ๊นŠ์ด์— ๋Œ€ํ•œ ๋ฌธ์žฅ๋“ค์˜ ํ‰๊ท ์œผ๋กœ ์ธก์ •)์™€ ํ‰๊ท  ์˜์กด ๊ฑฐ๋ฆฌ(๋ฌธ์žฅ ๋‚ด์˜ ๋‹จ์–ด ์Œ ๊ฐ„์˜ ํ‰๊ท  ์˜์กด ๊ฑฐ๋ฆฌ๋กœ ์ธก์ •)๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” Futrell et al. (2015); Gibson et al. (2000); Oya (2021)์—์„œ ๊ตฌ๋ฌธ์  ๋‚œ์ด๋„์˜ ์ข‹์€ ์ธก์ •์น˜๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋…์„œ ์ˆ˜์ค€๊ณผ TTR ๋‹ค์–‘์„ฑ์—์„œ์™€ ์œ ์‚ฌํ•œ ๊ฒฝํ–ฅ์„ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ, ๋งค์ฒด ์Šคํƒ€์ผ์€ ๊นŠ์ด, mdd, ๊ทธ๋ฆฌ๊ณ  ์ „๋ฐ˜์ ์ธ ๊ตฌ๋ฌธ ๋ณต์žก์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, QA ์Šคํƒ€์ผ์€ ์ด๋Ÿฌํ•œ ๋ณต์žก์„ฑ์„ ๊ฐ์†Œ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
๋งˆ์ง€๋ง‰์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๊ตฌ๋ฌธ๋ก ์  ์–ด๋ ค์›€์˜ ์ข‹์€ ์ฒ™๋„๋กœ ๋‚˜ํƒ€๋‚œ ํ‰๊ท  ํŠธ๋ฆฌ ๊นŠ์ด(์˜์กด์„ฑ ํŠธ๋ฆฌ ๊นŠ์ด์˜ ๋ฌธ์žฅ ํ‰๊ท ์œผ๋กœ ์ธก์ •)์™€ ํ‰๊ท  ์˜์กด์„ฑ ๊ฑฐ๋ฆฌ(๋ฌธ์žฅ ๋‚ด ์ž„์˜์˜ ๋‹จ์–ด ์Œ์˜ ํ‰๊ท  ์˜์กด์„ฑ ๊ฑฐ๋ฆฌ๋กœ ์ธก์ •)๋ฅผ ๊ทธ๋ฆผ 11์—์„œ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. Futrell et al. (2015); Gibson et al. (2000); Oya (2021). ์šฐ๋ฆฌ๋Š” ๋ฏธ๋””์–ด ์Šคํƒ€์ผ์ด ์ผ๋ฐ˜์ ์œผ๋กœ ๊นŠ์ด, mdd ๋ฐ ๊ตฌ๋ฌธ ๋ณต์žก์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ค๋Š” ๋…ํ•ด ์ˆ˜์ค€ ๋ฐ TTR ๋‹ค์–‘์„ฑ๊ณผ ์œ ์‚ฌํ•œ ์ถ”์„ธ๋ฅผ ๋ฐœ๊ฒฌํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์‹œ ํ•œ ๋ฒˆ QA ์Šคํƒ€์ผ์ด ์ด๋Ÿฌํ•œ ๋ณต์žก์„ฑ์„ ์ค„์ธ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
We consider our model's ability to reconstruct specific entities from prompts that have a high likelihood of containing personal information, such as names, dates, and nationalities. To test this, we generate a synthetic dataset of prompts that contain these attributes. We edit prompts from Instructions-2M with private attributes sourced from Wikibio (Lebret et al., 2016) and Synthbio (Yuan et al., 2022)11. We invert these prompts using both our LLAMA 7B and LLAMA-chat 7B models and measure accuracy at reconstructing private entities across categories.
์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์ด๋ฆ„, ๋‚ ์งœ, ๊ตญ์  ๋“ฑ ๊ฐœ์ธ ์ •๋ณด๋ฅผ ํฌํ•จํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ”„๋กฌํ”„ํŠธ์—์„œ ํŠน์ • ์—”ํ‹ฐํ‹ฐ๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ์ด๋Ÿฌํ•œ ์†์„ฑ์„ ํฌํ•จํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์…‹์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” Wikibio (Lebret et al., 2016)์™€ Synthbio (Yuan et al., 2022)11์—์„œ ๊ฐ€์ ธ์˜จ ๊ฐœ์ธ ์†์„ฑ์„ ์‚ฌ์šฉํ•˜์—ฌ Instructions-2M์˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํŽธ์ง‘ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ LLAMA 7B์™€ LLAMA-chat 7B ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋Ÿฌํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฐ˜์ „์‹œํ‚ค๊ณ , ์นดํ…Œ๊ณ ๋ฆฌ๋ณ„๋กœ ๊ฐœ์ธ ์—”ํ‹ฐํ‹ฐ๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ์ •ํ™•๋„๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค.
์ด๋ฆ„, ๋‚ ์งœ, ๊ตญ์  ๋“ฑ ๊ฐœ์ธ ์ •๋ณด๊ฐ€ ํฌํ•จ๋  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ”„๋กฌํ”„ํŠธ์—์„œ ํŠน์ • ๊ฐœ์ฒด๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ์ด๋Ÿฌํ•œ ์†์„ฑ์„ ํฌํ•จํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ฐœ์ธ ์ •๋ณด๊ฐ€ ํฌํ•จ๋œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด Wikibio(Lebret et al., 2016)์™€ Synthbio(Yuan et al., 2022)11์—์„œ ๊ฐ€์ ธ์˜จ ๊ฐœ์ธ ์†์„ฑ์„ ์‚ฌ์šฉํ•˜์—ฌ Instructions-2M์˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ˆ˜์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ LLAMA 7B์™€ LLAMA-chat 7B ๋ชจ๋ธ์„ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ˜์ „ํ•˜๊ณ  ์นดํ…Œ๊ณ ๋ฆฌ๋ณ„๋กœ ๊ฐœ์ธ ์—”ํ‹ฐํ‹ฐ๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ์ •ํ™•๋„๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค.
Reading the data.CommonCrawl is available in either WARC (raw HTML response), or WET files (preprocessed to only include plain text). Individual files correspond to a page at a given URL; these constitute single documents/samples. Working with WET files would spare us from running our own HTML extraction; however, in line with previous works (Gao et al., 2020; Rae et al., 2021), we found WET files to include undesirable navigation menus, ads, and other irrelevant texts. Accordingly, our pipeline starts from raw WARC files, read with the wavcio library.
๋ฐ์ดํ„ฐ ์ฝ๊ธฐ. CommonCrawl์€ WARC(์›์‹œ HTML ์‘๋‹ต) ๋˜๋Š” WET ํŒŒ์ผ(์ „์ฒ˜๋ฆฌ๋˜์–ด ์ผ๋ฐ˜ ํ…์ŠคํŠธ๋งŒ ํฌํ•จ)๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ๊ฐœ๋ณ„ ํŒŒ์ผ์€ ์ฃผ์–ด์ง„ URL์˜ ํŽ˜์ด์ง€์— ํ•ด๋‹นํ•˜๋ฉฐ, ์ด๋Š” ๋‹จ์ผ ๋ฌธ์„œ/์ƒ˜ํ”Œ์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. WET ํŒŒ์ผ์„ ์‚ฌ์šฉํ•˜๋ฉด ์ž์ฒด HTML ์ถ”์ถœ์„ ์‹คํ–‰ํ•  ํ•„์š”๊ฐ€ ์—†์ง€๋งŒ, ์ด์ „ ์—ฐ๊ตฌ(Gao et al., 2020; Rae et al., 2021)์™€ ์ผ์น˜ํ•˜๊ฒŒ, WET ํŒŒ์ผ์—๋Š” ์›ํ•˜์ง€ ์•Š๋Š” ํƒ์ƒ‰ ๋ฉ”๋‰ด, ๊ด‘๊ณ  ๋ฐ ๊ธฐํƒ€ ๊ด€๋ จ ์—†๋Š” ํ…์ŠคํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ์˜ ํŒŒ์ดํ”„๋ผ์ธ์€ wavcio ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋กœ ์ฝ์€ ์›์‹œ WARC ํŒŒ์ผ์—์„œ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ ์ฝ๊ธฐ.CommonCrawl์€ WARC(์›์‹œ HTML ์‘๋‹ต) ๋˜๋Š” WET ํŒŒ์ผ(์ผ๋ฐ˜ ํ…์ŠคํŠธ๋งŒ ํฌํ•จํ•˜๋„๋ก ์‚ฌ์ „ ์ฒ˜๋ฆฌ๋œ)๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ๊ฐœ๋ณ„ ํŒŒ์ผ์€ ํŠน์ • URL์˜ ํŽ˜์ด์ง€์— ํ•ด๋‹นํ•˜๋ฉฐ, ์ด๋Š” ๋‹จ์ผ ๋ฌธ์„œ/์ƒ˜ํ”Œ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. WET ํŒŒ์ผ์„ ์‚ฌ์šฉํ•˜๋ฉด HTML ์ถ”์ถœ์„ ์ง์ ‘ ์‹คํ–‰ํ•  ํ•„์š”๊ฐ€ ์—†์ง€๋งŒ, ์ด์ „ ์—ฐ๊ตฌ(Gao et al., 2020; Rae et al., 2021)์— ๋”ฐ๋ฅด๋ฉด WET ํŒŒ์ผ์— ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์€ ํƒ์ƒ‰ ๋ฉ”๋‰ด, ๊ด‘๊ณ  ๋ฐ ๊ธฐํƒ€ ๊ด€๋ จ ์—†๋Š” ํ…์ŠคํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํŒŒ์ดํ”„๋ผ์ธ์€ wavcio ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋กœ ์ฝ์€ ์›์‹œ WARC ํŒŒ์ผ์—์„œ ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค.
Knowledge distillation is the process of learning from a larger model [143]. Earlier days of best-performing models release have proven that this approach is very useful even if it is used in an API distillation approach. It is also referred to as an approach to distill the knowledge of not a single model but in fact multiple models into a smaller one. Creating smaller models by this approach yields smaller model sizes that can be used even on edge devices. Knowledge distillation as shown in Fig 35, illustrates a general setup of this training scheme.
์ง€์‹ ์ฆ๋ฅ˜(Knowledge distillation)๋Š” ๋” ํฐ ๋ชจ๋ธ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๋Š” ๊ณผ์ •์ž…๋‹ˆ๋‹ค [143]. ์„ฑ๋Šฅ์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ๋ชจ๋ธ๋“ค์ด ์ถœ์‹œ๋œ ์ดˆ๊ธฐ์—๋Š” ์ด ์ ‘๊ทผ ๋ฐฉ์‹์ด API ์ฆ๋ฅ˜ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์šฉ๋˜๋”๋ผ๋„ ๋งค์šฐ ์œ ์šฉํ•˜๋‹ค๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋‹จ์ผ ๋ชจ๋ธ์ด ์•„๋‹Œ ์—ฌ๋Ÿฌ ๋ชจ๋ธ์˜ ์ง€์‹์„ ๋” ์ž‘์€ ๋ชจ๋ธ๋กœ ์ฆ๋ฅ˜ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ๋„ ์–ธ๊ธ‰๋ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ์ƒ์„ฑ๋œ ๋” ์ž‘์€ ๋ชจ๋ธ๋“ค์€ ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค์—์„œ๋„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ž‘์€ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ๊ฐ€์ง‘๋‹ˆ๋‹ค. Fig 35์— ๋‚˜ํƒ€๋‚œ ๊ฒƒ์ฒ˜๋Ÿผ, ์ง€์‹ ์ฆ๋ฅ˜๋Š” ์ด ํ›ˆ๋ จ ์ฒด๊ณ„์˜ ์ผ๋ฐ˜์ ์ธ ์„ค์ •์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์ง€์‹ ์ฆ๋ฅ˜๋Š” ๋” ํฐ ๋ชจ๋ธ์—์„œ ํ•™์Šตํ•˜๋Š” ๊ณผ์ •์ž…๋‹ˆ๋‹ค[143]. ์ตœ๊ณ  ์„ฑ๋Šฅ ๋ชจ๋ธ ์ถœ์‹œ ์ดˆ์ฐฝ๊ธฐ์—๋Š” ์ด ์ ‘๊ทผ ๋ฐฉ์‹์ด API ์ฆ๋ฅ˜ ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์šฉ๋˜๋”๋ผ๋„ ๋งค์šฐ ์œ ์šฉํ•˜๋‹ค๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์‹ค์ œ๋กœ ํ•˜๋‚˜์˜ ๋ชจ๋ธ์ด ์•„๋‹Œ ์—ฌ๋Ÿฌ ๋ชจ๋ธ์˜ ์ง€์‹์„ ๋” ์ž‘์€ ๋ชจ๋ธ๋กœ ์ฆ๋ฅ˜ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์ด๋ผ๊ณ ๋„ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ๋” ์ž‘์€ ๋ชจ๋ธ์„ ๋งŒ๋“ค๋ฉด ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค์—์„œ๋„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋” ์ž‘์€ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 35์— ๋‚˜์™€ ์žˆ๋Š” ์ง€์‹ ์ฆ๋ฅ˜๋Š” ์ด ํ›ˆ๋ จ ์ฒด๊ณ„์˜ ์ผ๋ฐ˜์ ์ธ ์„ค์ •์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
In this survey, we systematically summarize existing research in the field of LLM-based autonomous agents. We present and review these studies from three aspects including the construction, application, and evaluation of the agents. For each of these aspects, we provide a detailed taxonomy to draw connections among the existing research, summarizing the major techniques and their development histories. In addition to reviewing the previous work, we also propose several challenges in this field, which are expected to guide potential future directions.
์ด ์„ค๋ฌธ ์กฐ์‚ฌ์—์„œ๋Š” LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ ๋ถ„์•ผ์˜ ๊ธฐ์กด ์—ฐ๊ตฌ๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์˜ ๊ตฌ์„ฑ, ์‘์šฉ, ํ‰๊ฐ€๋ฅผ ํฌํ•จํ•œ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์—์„œ ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ๋ฅผ ์ œ์‹œํ•˜๊ณ  ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์ธก๋ฉด์— ๋Œ€ํ•ด ๊ธฐ์กด ์—ฐ๊ตฌ ๊ฐ„์˜ ์—ฐ๊ฒฐ์„ ๋„์ถœํ•˜๊ธฐ ์œ„ํ•ด ์ž์„ธํ•œ ๋ถ„๋ฅ˜ ์ฒด๊ณ„๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์ฃผ์š” ๊ธฐ์ˆ ๊ณผ ๊ทธ ๋ฐœ์ „ ์—ญ์‚ฌ๋ฅผ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ๋ฅผ ๊ฒ€ํ† ํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„, ์ด ๋ถ„์•ผ์—์„œ ์˜ˆ์ƒ๋˜๋Š” ์—ฌ๋Ÿฌ ๋„์ „ ๊ณผ์ œ๋ฅผ ์ œ์•ˆํ•˜์—ฌ ์ž ์žฌ์ ์ธ ๋ฏธ๋ž˜ ๋ฐฉํ–ฅ์„ ์•ˆ๋‚ดํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.
์ด ์„ค๋ฌธ์กฐ์‚ฌ์—์„œ๋Š” LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ ๋ถ„์•ผ์˜ ๊ธฐ์กด ์—ฐ๊ตฌ๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์˜ ๊ตฌ์„ฑ, ์ ์šฉ ๋ฐ ํ‰๊ฐ€์˜ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์—์„œ ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ๋ฅผ ์ œ์‹œํ•˜๊ณ  ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์ธก๋ฉด์— ๋Œ€ํ•ด ์„ธ๋ถ€์ ์ธ ๋ถ„๋ฅ˜๋ฅผ ์ œ๊ณตํ•˜์—ฌ ๊ธฐ์กด ์—ฐ๊ตฌ ๊ฐ„์˜ ์—ฐ๊ด€์„ฑ์„ ๋„์ถœํ•˜๊ณ  ์ฃผ์š” ๊ธฐ์ˆ ๊ณผ ๊ทธ ๊ฐœ๋ฐœ ์—ญ์‚ฌ๋ฅผ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ์ž‘์—…์„ ๊ฒ€ํ† ํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ์ด ๋ถ„์•ผ์˜ ๋ช‡ ๊ฐ€์ง€ ๊ณผ์ œ๋ฅผ ์ œ์•ˆํ•˜์—ฌ ์ž ์žฌ์ ์ธ ํ–ฅํ›„ ๋ฐฉํ–ฅ์„ ์ œ์‹œํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค.
MiniCPM was initially published as a technical blog on February 1st, 2024. Since then, we have received numerous insightful feedback from the community, significantly contributing to the development of this paper. We extend our gratitude to Chunting Zhou and Armen Aghajanyan for their valuable discussions. Special thanks go to Feigin Sun and Yan Wang for their meticulous feedback on clarifying ambiguities in the blog. Additionally, we appreciate the open-source community's efforts in integrating MiniCPM into inference frameworks like llama.cpp, etc.
MiniCPM์€ 2024๋…„ 2์›” 1์ผ์— ๊ธฐ์ˆ  ๋ธ”๋กœ๊ทธ๋กœ ์ฒ˜์Œ ๊ฒŒ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์ดํ›„๋กœ ์ปค๋ฎค๋‹ˆํ‹ฐ๋กœ๋ถ€ํ„ฐ ๋งŽ์€ ํ†ต์ฐฐ๋ ฅ ์žˆ๋Š” ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์•˜์œผ๋ฉฐ, ์ด๋Š” ์ด ๋…ผ๋ฌธ์˜ ๋ฐœ์ „์— ํฌ๊ฒŒ ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ท€์ค‘ํ•œ ๋…ผ์˜๋ฅผ ์ œ๊ณตํ•ด์ฃผ์‹  Chunting Zhou์™€ Armen Aghajanyan์—๊ฒŒ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ๋ธ”๋กœ๊ทธ์˜ ๋ชจํ˜ธ์„ฑ์„ ๋ช…ํ™•ํžˆ ํ•˜๋Š” ๋ฐ ์žˆ์–ด ์„ธ์‹ฌํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ฃผ์‹  Feigin Sun๊ณผ Yan Wang์—๊ฒŒ ํŠน๋ณ„ํžˆ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ๋˜ํ•œ, MiniCPM์„ llama.cpp์™€ ๊ฐ™์€ ์ถ”๋ก  ํ”„๋ ˆ์ž„์›Œํฌ์— ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ๊ธฐ์—ฌํ•œ ์˜คํ”ˆ ์†Œ์Šค ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ๋…ธ๋ ฅ์—๋„ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
MiniCPM์€ 2024๋…„ 2์›” 1์ผ์— ๊ธฐ์ˆ  ๋ธ”๋กœ๊ทธ๋กœ ์ฒ˜์Œ ๊ฒŒ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์ดํ›„๋กœ ์ปค๋ฎค๋‹ˆํ‹ฐ๋กœ๋ถ€ํ„ฐ ์ˆ˜๋งŽ์€ ํ†ต์ฐฐ๋ ฅ ์žˆ๋Š” ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์•„ ์ด ๋…ผ๋ฌธ์˜ ๊ฐœ๋ฐœ์— ํฌ๊ฒŒ ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ท€์ค‘ํ•œ ๋…ผ์˜๋ฅผ ์ œ๊ณตํ•œ ์ฒญํŒ… ์ €์šฐ์™€ ์•„๋ฅด๋ฉ˜ ์•„๊ฐ€์ž์–€์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ ๋ธ”๋กœ๊ทธ์˜ ๋ชจํ˜ธํ•œ ๋ถ€๋ถ„์„ ๋ช…ํ™•ํžˆ ํ•˜๋Š” ๋ฐ ์„ธ์‹ฌํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•œ ํ”ผ๊ธด ์„ ๊ณผ ์˜Œ ์™•์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, MiniCPM์„ llama.cpp ๋“ฑ๊ณผ ๊ฐ™์€ ์ถ”๋ก  ํ”„๋ ˆ์ž„์›Œํฌ์— ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ๋…ธ๋ ฅํ•œ ์˜คํ”ˆ ์†Œ์Šค ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ๋…ธ๋ ฅ์— ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
**GPT-4 Judgement:** The first step is incorrect; '2/.5' suggests division when it should be multiplication, '2*.5'. The correct calculation for the reading time is 2/2 hours, which equals 1 hour, not '2/.5' hours. However, the numerical result '1 hour' happens to be correct despite the incorrect operation shown. The remaining steps correctly add the two activities to find the total daily time spent, correctly multiply by the number of occurrences per week, and correctly multiply by the number of weeks to find the total hours spent over four weeks.
**GPT-4 ํŒ๋‹จ:** ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„๋Š” ์ž˜๋ชป๋˜์—ˆ์Šต๋‹ˆ๋‹ค. '2/.5'๋Š” ๊ณฑ์…ˆ์ด ๋˜์–ด์•ผ ํ•  ๋•Œ ๋‚˜๋ˆ—์…ˆ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค, ์ฆ‰ '2*.5'์ž…๋‹ˆ๋‹ค. ๋…์„œ ์‹œ๊ฐ„์— ๋Œ€ํ•œ ์˜ฌ๋ฐ”๋ฅธ ๊ณ„์‚ฐ์€ 2/2 ์‹œ๊ฐ„์œผ๋กœ, ์ด๋Š” 1์‹œ๊ฐ„์— ํ•ด๋‹นํ•˜๋ฉฐ, '2/.5' ์‹œ๊ฐ„์ด ์•„๋‹™๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ž˜๋ชป๋œ ์—ฐ์‚ฐ์ด ํ‘œ์‹œ๋˜์—ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ˆซ์ž ๊ฒฐ๊ณผ '1์‹œ๊ฐ„'์€ ์˜ฌ๋ฐ”๋ฆ…๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€ ๋‹จ๊ณ„๋“ค์€ ๋‘ ํ™œ๋™์„ ๋”ํ•˜์—ฌ ํ•˜๋ฃจ์— ์†Œ๋น„๋œ ์ด ์‹œ๊ฐ„์„ ์ฐพ๊ณ , ์ฃผ๋‹น ๋ฐœ์ƒ ํšŸ์ˆ˜์— ๋งž๊ฒŒ ๊ณฑํ•˜๊ณ , 4์ฃผ ๋™์•ˆ ์†Œ๋น„๋œ ์ด ์‹œ๊ฐ„์„ ์ฐพ๊ธฐ ์œ„ํ•ด ์ฃผ ์ˆ˜์— ๋งž๊ฒŒ ๊ณฑํ•˜๋Š” ๊ณผ์ •์„ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
**GPT-4 ํŒ๋‹จ:** ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„๊ฐ€ ์ž˜๋ชป๋˜์—ˆ์Šต๋‹ˆ๋‹ค. '2/.5'๋Š” ๊ณฑ์…ˆ์ด์–ด์•ผ ํ•˜๋Š” ๋‚˜๋ˆ—์…ˆ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์˜ฌ๋ฐ”๋ฅธ ์ฝ๊ธฐ ์‹œ๊ฐ„ ๊ณ„์‚ฐ์€ 2/2์‹œ๊ฐ„์œผ๋กœ '2/.5'์‹œ๊ฐ„์ด ์•„๋‹Œ 1์‹œ๊ฐ„์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ž˜๋ชป๋œ ์—ฐ์‚ฐ์ด ํ‘œ์‹œ๋˜์—ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ˆซ์ž ๊ฒฐ๊ณผ '1์‹œ๊ฐ„'์€ ์ •๋‹ต์ž…๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€ ๋‹จ๊ณ„์—์„œ๋Š” ๋‘ ํ™œ๋™์„ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๋”ํ•˜์—ฌ ์ผ์ผ ์ด ์‹œ๊ฐ„์„ ๊ตฌํ•˜๊ณ , ์ผ์ฃผ์ผ์— ๋ฐœ์ƒํ•˜๋Š” ํšŸ์ˆ˜๋ฅผ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๊ณฑํ•˜๊ณ , 4์ฃผ ๋™์•ˆ ์ด ์‹œ๊ฐ„์„ ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ์ฃผ ์ˆ˜๋ฅผ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๊ณฑํ•ฉ๋‹ˆ๋‹ค.
While larger vocabulary sizes introduce additional computational costs, they might also result in lower fertility scores and, therefore, lower overall computational costs for processing a set of documents, as discussed in Section 4. However, our findings in Fig. 3 show that increasing the vocabulary size from 50k to larger vocabulary sizes increases the computational costs in all cases. This highlights that the potentially lower fertility of larger vocabulary sizes cannot compensate for the additional costs introduced by the larger vocabulary size.
๋” ํฐ ์–ดํœ˜ ํฌ๊ธฐ๋Š” ์ถ”๊ฐ€์ ์ธ ๊ณ„์‚ฐ ๋น„์šฉ์„ ์ดˆ๋ž˜ํ•˜์ง€๋งŒ, ์ด๋Š” ๋˜ํ•œ ๋” ๋‚ฎ์€ fertility ์ ์ˆ˜๋ฅผ ๊ฐ€์ ธ์™€ ๋ฌธ์„œ ์ง‘ํ•ฉ์„ ์ฒ˜๋ฆฌํ•˜๋Š” ์ „์ฒด ๊ณ„์‚ฐ ๋น„์šฉ์„ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” Section 4์—์„œ ๋…ผ์˜๋œ ๋ฐ”์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Fig. 3์—์„œ ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” ์–ดํœ˜ ํฌ๊ธฐ๋ฅผ 50k์—์„œ ๋” ํฐ ์–ดํœ˜ ํฌ๊ธฐ๋กœ ๋Š˜๋ฆฌ๋ฉด ๋ชจ๋“  ๊ฒฝ์šฐ์— ๊ณ„์‚ฐ ๋น„์šฉ์ด ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Š” ๋” ํฐ ์–ดํœ˜ ํฌ๊ธฐ์˜ ์ž ์žฌ์ ์œผ๋กœ ๋‚ฎ์€ fertility๊ฐ€ ๋” ํฐ ์–ดํœ˜ ํฌ๊ธฐ๋กœ ์ธํ•ด ๋ฐœ์ƒํ•˜๋Š” ์ถ”๊ฐ€ ๋น„์šฉ์„ ์ƒ์‡„ํ•  ์ˆ˜ ์—†์Œ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
์–ดํœ˜ ํฌ๊ธฐ๊ฐ€ ํด์ˆ˜๋ก ๊ณ„์‚ฐ ๋น„์šฉ์ด ์ถ”๊ฐ€๋กœ ๋ฐœ์ƒํ•˜์ง€๋งŒ, ์„น์…˜ 4์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด ์–ดํœ˜ ํฌ๊ธฐ๊ฐ€ ํด์ˆ˜๋ก ์ถœ์‚ฐ์œจ ์ ์ˆ˜๊ฐ€ ๋‚ฎ์•„์ ธ ๋ฌธ์„œ ์ง‘ํ•ฉ์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ๋“œ๋Š” ์ „์ฒด ๊ณ„์‚ฐ ๋น„์šฉ์ด ๋‚ฎ์•„์งˆ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ทธ๋ฆผ 3์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด ์–ดํœ˜ ํฌ๊ธฐ๋ฅผ 50k์—์„œ ๋” ํฐ ์–ดํœ˜ ํฌ๊ธฐ๋กœ ๋Š˜๋ฆฌ๋ฉด ๋ชจ๋“  ๊ฒฝ์šฐ์— ๊ณ„์‚ฐ ๋น„์šฉ์ด ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์–ดํœ˜ ํฌ๊ธฐ๊ฐ€ ํด์ˆ˜๋ก ์ถœ์‚ฐ์œจ์ด ๋‚ฎ์•„์งˆ ์ˆ˜ ์žˆ์ง€๋งŒ ์–ดํœ˜ ํฌ๊ธฐ๊ฐ€ ํด์ˆ˜๋ก ์ถ”๊ฐ€๋˜๋Š” ๋น„์šฉ์„ ๋ณด์ƒํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
* The GatedMLP architecture _slows down_ the model's learning speed, and we observe less stable training with its use.27 Footnote 27: For example, mixed-precision fp16 training can sometimes fail for LLaMA/Mistral models smaller than 100M; hence, we use mixed-precision bf16 instead. Conversely, GPT2 models up to 1B can be trained with fp16. * Removing MLP layers entirely _slows down_ the model's learning speed, whereas adjusting the size of MLP layers (e.g., from \(8d^{2}\) to \(10.5d^{2}\) or down to \(2d^{2}\)) may not have a significant impact.
* GatedMLP ์•„ํ‚คํ…์ฒ˜๋Š” ๋ชจ๋ธ์˜ ํ•™์Šต ์†๋„๋ฅผ _์ €ํ•˜์‹œํ‚ต๋‹ˆ๋‹ค_, ๊ทธ๋ฆฌ๊ณ  ์ด๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ ํ›ˆ๋ จ์ด ๋œ ์•ˆ์ •์ ์ž„์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.27 ๊ฐ์ฃผ 27: ์˜ˆ๋ฅผ ๋“ค์–ด, 100M๋ณด๋‹ค ์ž‘์€ LLaMA/Mistral ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ˜ผํ•ฉ ์ •๋ฐ€๋„ fp16 ํ›ˆ๋ จ์ด ๋•Œ๋•Œ๋กœ ์‹คํŒจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” ๋Œ€์‹  ํ˜ผํ•ฉ ์ •๋ฐ€๋„ bf16์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์—, ์ตœ๋Œ€ 1B์˜ GPT2 ๋ชจ๋ธ์€ fp16์œผ๋กœ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * MLP ๋ ˆ์ด์–ด๋ฅผ ์™„์ „ํžˆ ์ œ๊ฑฐํ•˜๋ฉด ๋ชจ๋ธ์˜ ํ•™์Šต ์†๋„๊ฐ€ _์ €ํ•˜๋˜์ง€๋งŒ_, MLP ๋ ˆ์ด์–ด์˜ ํฌ๊ธฐ๋ฅผ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ(์˜ˆ: \(8d^{2}\)์—์„œ \(10.5d^{2}\) ๋˜๋Š” \(2d^{2}\)๋กœ ์ค„์ด๋Š” ๊ฒƒ)์€ ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
* ๊ฒŒ์ดํŠธ๋“œMLP ์•„ํ‚คํ…์ฒ˜๋Š” ๋ชจ๋ธ์˜ ํ•™์Šต ์†๋„๋ฅผ _๋Š๋ฆฌ๊ฒŒ_ ํ•˜๋ฉฐ, ์ด๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ํ›ˆ๋ จ์ด ๋œ ์•ˆ์ •์ ์œผ๋กœ ์ง„ํ–‰๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.27 ๊ฐ์ฃผ 27: ์˜ˆ๋ฅผ ๋“ค์–ด, 1์–ต ๊ฐœ ๋ฏธ๋งŒ์˜ LLaMA/Mistral ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ˜ผํ•ฉ ์ •๋ฐ€๋„ fp16 ํ›ˆ๋ จ์ด ์‹คํŒจํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํ˜ผํ•ฉ ์ •๋ฐ€๋„ bf16์„ ๋Œ€์‹  ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ GPT2 ๋ชจ๋ธ์€ ์ตœ๋Œ€ 1B๊นŒ์ง€ fp16์œผ๋กœ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * MLP ๋ ˆ์ด์–ด๋ฅผ ์™„์ „ํžˆ ์ œ๊ฑฐํ•˜๋ฉด ๋ชจ๋ธ์˜ ํ•™์Šต ์†๋„๊ฐ€ _๋Š๋ ค์ง€์ง€๋งŒ, MLP ๋ ˆ์ด์–ด์˜ ํฌ๊ธฐ๋ฅผ ์กฐ์ •(์˜ˆ: \(8d^{2}\)์—์„œ \(10.5d^{2}\) ๋˜๋Š” \(2d^{2}\)๋กœ ์ค„์ด๊ธฐ)ํ•˜๋ฉด ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In this new dataset, the preferred translation (\(y_{w}\)) remains the best of the three translation candidates, selected in the same manner as in Section 3.1. However, the dis-preferred translation is intentionally modified to be a noised version of \(y_{w}\). We applied random deletions of words with a probability of 0.15 and word swaps within a range of 1 with a probability of 0.3, following the method suggested by Zeng et al. (2023) for creating manually noised dis-preferred data. This approach produces worse translations that are artificial.
์ด ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์…‹์—์„œ ์„ ํ˜ธ๋˜๋Š” ๋ฒˆ์—ญ (\(y_{w}\))์€ 3๊ฐœ์˜ ๋ฒˆ์—ญ ํ›„๋ณด ์ค‘์—์„œ ์—ฌ์ „ํžˆ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ๊ฒƒ์œผ๋กœ, Section 3.1์—์„œ์™€ ๋™์ผํ•œ ๋ฐฉ์‹์œผ๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋น„์„ ํ˜ธ๋˜๋Š” ๋ฒˆ์—ญ์€ ์˜๋„์ ์œผ๋กœ \(y_{w}\)์˜ ๋…ธ์ด์ฆˆ ๋ฒ„์ „์œผ๋กœ ์ˆ˜์ •๋ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” Zeng et al. (2023)์ด ์ œ์•ˆํ•œ ๋ฐฉ๋ฒ•์„ ๋”ฐ๋ผ, ๋‹จ์–ด๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์‚ญ์ œํ•  ํ™•๋ฅ ์„ 0.15๋กœ ์„ค์ •ํ•˜๊ณ , ๋‹จ์–ด ๊ตํ™˜์„ 1 ๋ฒ”์œ„ ๋‚ด์—์„œ ํ™•๋ฅ  0.3์œผ๋กœ ์ ์šฉํ•˜์—ฌ ์ˆ˜๋™์œผ๋กœ ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•œ ๋น„์„ ํ˜ธ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ๋ฒ•์€ ์ธ์œ„์ ์œผ๋กœ ๋” ๋‚˜์œ ๋ฒˆ์—ญ์„ ๋งŒ๋“ค์–ด๋ƒ…๋‹ˆ๋‹ค.
์ด ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์„ ํ˜ธ ๋ฒˆ์—ญ(\(y_{w}\))์€ ์„น์…˜ 3.1์—์„œ์™€ ๋™์ผํ•œ ๋ฐฉ์‹์œผ๋กœ ์„ ํƒ๋œ ์„ธ ๋ฒˆ์—ญ ํ›„๋ณด ์ค‘์—์„œ ๊ฐ€์žฅ ์ข‹์€ ๋ฒˆ์—ญ์œผ๋กœ ์œ ์ง€๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์„ ํ˜ธํ•˜์ง€ ์•Š๋Š” ๋ฒˆ์—ญ์€ ์˜๋„์ ์œผ๋กœ \(y_{w}\)์˜ ๋…ธ์ด์ฆˆ ๋ฒ„์ „์œผ๋กœ ์ˆ˜์ •๋ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ˆ˜๋™์œผ๋กœ ๋…ธ์ด์ฆˆ๊ฐ€ ์žˆ๋Š” ์„ ํ˜ธํ•˜์ง€ ์•Š๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด Zeng et al. (2023)์—์„œ ์ œ์•ˆํ•œ ๋ฐฉ๋ฒ•์— ๋”ฐ๋ผ ๋‹จ์–ด์˜ ์ž„์˜ ์‚ญ์ œ๋ฅผ 0.15์˜ ํ™•๋ฅ ๋กœ, 1 ๋ฒ”์œ„ ๋‚ด์—์„œ ๋‹จ์–ด ๊ตํ™˜์„ 0.3์˜ ํ™•๋ฅ ๋กœ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ธ์œ„์ ์ธ ๋” ๋‚˜์œ ๋ฒˆ์—ญ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
Question: Is participating in the stress at work survey mandatory? Answer: No, it is voluntary. Question: Will all answers be saved anonymously? Answer: Yes, they will. Question: Who is eligible to take the questionnaire? Answer: Only individuals with work experience, whether part-time or full-time. Question: What happens if someone is not eligible to take the questionnaire? Answer: They will not be able to answer some of the questions. Question: Can you provide a link to all language versions of the questionnaire? Answer: Yes, here is the link.
์งˆ๋ฌธ: ์ง์žฅ ์ŠคํŠธ๋ ˆ์Šค ์„ค๋ฌธ์กฐ์‚ฌ์— ์ฐธ์—ฌํ•˜๋Š” ๊ฒƒ์ด ์˜๋ฌด์ธ๊ฐ€์š”? ๋‹ต๋ณ€: ์•„๋‹ˆ์š”, ์ž๋ฐœ์ ์ž…๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋ชจ๋“  ๋‹ต๋ณ€์ด ์ต๋ช…์œผ๋กœ ์ €์žฅ๋˜๋‚˜์š”? ๋‹ต๋ณ€: ๋„ค, ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋ˆ„๊ฐ€ ์„ค๋ฌธ์กฐ์‚ฌ๋ฅผ ๋ฐ›์„ ์ž๊ฒฉ์ด ์žˆ๋‚˜์š”? ๋‹ต๋ณ€: ํŒŒํŠธํƒ€์ž„์ด๋“  ํ’€ํƒ€์ž„์ด๋“  ๊ทผ๋ฌด ๊ฒฝํ—˜์ด ์žˆ๋Š” ์‚ฌ๋žŒ๋งŒ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์„ค๋ฌธ์กฐ์‚ฌ๋ฅผ ๋ฐ›์„ ์ž๊ฒฉ์ด ์—†๋Š” ์‚ฌ๋žŒ์ด ์žˆ์œผ๋ฉด ์–ด๋–ป๊ฒŒ ๋˜๋‚˜์š”? ๋‹ต๋ณ€: ์ผ๋ถ€ ์งˆ๋ฌธ์— ๋‹ตํ•  ์ˆ˜ ์—†๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์„ค๋ฌธ์กฐ์‚ฌ์˜ ๋ชจ๋“  ์–ธ์–ด ๋ฒ„์ „์— ๋Œ€ํ•œ ๋งํฌ๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋‚˜์š”? ๋‹ต๋ณ€: ๋„ค, ์—ฌ๊ธฐ ๋งํฌ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
์งˆ๋ฌธ: ์ง์žฅ ๋‚ด ์ŠคํŠธ๋ ˆ์Šค ์„ค๋ฌธ์กฐ์‚ฌ ์ฐธ์—ฌ๋Š” ํ•„์ˆ˜์ธ๊ฐ€์š”? ๋‹ต๋ณ€: ์•„๋‹ˆ์š”, ์ž๋ฐœ์ ์ž…๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋ชจ๋“  ๋‹ต๋ณ€์ด ์ต๋ช…์œผ๋กœ ์ €์žฅ๋˜๋‚˜์š”? ๋‹ต๋ณ€: ๋„ค, ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์„ค๋ฌธ์กฐ์‚ฌ์— ์‘ํ•  ์ˆ˜ ์žˆ๋Š” ์ž๊ฒฉ์€ ๋ˆ„๊ตฌ์ธ๊ฐ€์š”? ๋‹ต๋ณ€: ํŒŒํŠธํƒ€์ž„ ๋˜๋Š” ํ’€ํƒ€์ž„ ๋“ฑ ๊ทผ๋ฌด ๊ฒฝ๋ ฅ์ด ์žˆ๋Š” ๊ฐœ์ธ๋งŒ ํ•ด๋‹น๋ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์„ค๋ฌธ์กฐ์‚ฌ์— ์‘ํ•  ์ž๊ฒฉ์ด ์—†๋Š” ๊ฒฝ์šฐ ์–ด๋–ป๊ฒŒ ๋˜๋‚˜์š”? ๋‹ต๋ณ€: ์ผ๋ถ€ ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์„ค๋ฌธ์กฐ์‚ฌ์˜ ๋ชจ๋“  ์–ธ์–ด ๋ฒ„์ „์— ๋Œ€ํ•œ ๋งํฌ๋ฅผ ์ œ๊ณตํ•ด ์ฃผ์‹œ๊ฒ ์Šต๋‹ˆ๊นŒ? ๋‹ต๋ณ€: ๋„ค, ์—ฌ๊ธฐ ๋งํฌ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
The primary function, remove_unwanted_part, starts by splitting the input data into individual sentences. If the first sentence contains delimiters such as "\n\n" (indicating a new paragraph) or ":", the function checks the segment preceding the delimiter for the aforementioned unwanted elements. If these elements are detected, the preceding segment is removed. The entire revised content is then reconstructed and returned. In cases where no modifications are applicable, but we still have the flagged keywords, we remove the paraphrase completely.
์ฃผ์š” ํ•จ์ˆ˜์ธ remove_unwanted_part๋Š” ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐœ๋ณ„ ๋ฌธ์žฅ์œผ๋กœ ๋ถ„ํ• ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋ฌธ์žฅ์ด "\n\n" (์ƒˆ ๋‹จ๋ฝ์„ ๋‚˜ํƒ€๋ƒ„) ๋˜๋Š” ":"์™€ ๊ฐ™์€ ๊ตฌ๋ถ„ ๊ธฐํ˜ธ๋ฅผ ํฌํ•จํ•˜๋Š” ๊ฒฝ์šฐ, ํ•จ์ˆ˜๋Š” ๊ตฌ๋ถ„ ๊ธฐํ˜ธ ์•ž์˜ ์„ธ๊ทธ๋จผํŠธ์—์„œ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๋ถˆํ•„์š”ํ•œ ์š”์†Œ๋ฅผ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์š”์†Œ๊ฐ€ ๊ฐ์ง€๋˜๋ฉด, ์•ž์˜ ์„ธ๊ทธ๋จผํŠธ๊ฐ€ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ˆ˜์ •๋œ ์ „์ฒด ๋‚ด์šฉ์„ ์žฌ๊ตฌ์„ฑํ•˜์—ฌ ๋ฐ˜ํ™˜ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜์ •์ด ์ ์šฉ๋˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ์—๋„ ํ”Œ๋ž˜๊ทธ๋œ ํ‚ค์›Œ๋“œ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ, ์šฐ๋ฆฌ๋Š” ํŒจ๋Ÿฌํ”„๋ ˆ์ด์ฆˆ๋ฅผ ์™„์ „ํžˆ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
๊ธฐ๋ณธ ๊ธฐ๋Šฅ์ธ remove_unwanted_part๋Š” ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐœ๋ณ„ ๋ฌธ์žฅ์œผ๋กœ ๋ถ„ํ• ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋ฌธ์žฅ์— "\n\n"(์ƒˆ ๋‹จ๋ฝ์„ ๋‚˜ํƒ€๋ƒ„) ๋˜๋Š” ":"์™€ ๊ฐ™์€ ๊ตฌ๋ถ„ ๊ธฐํ˜ธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š” ๊ฒฝ์šฐ, ํ•จ์ˆ˜๋Š” ๊ตฌ๋ถ„ ๊ธฐํ˜ธ ์•ž์˜ ์„ธ๊ทธ๋จผํŠธ์—์„œ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์›์น˜ ์•Š๋Š” ์š”์†Œ๊ฐ€ ์žˆ๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์š”์†Œ๊ฐ€ ๊ฐ์ง€๋˜๋ฉด ํ•ด๋‹น ์„ธ๊ทธ๋จผํŠธ๊ฐ€ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ „์ฒด ์ˆ˜์ •๋œ ์ฝ˜ํ…์ธ ๊ฐ€ ์žฌ๊ตฌ์„ฑ๋˜์–ด ๋ฐ˜ํ™˜๋ฉ๋‹ˆ๋‹ค. ์ˆ˜์ •์ด ์ ์šฉ๋˜์ง€ ์•Š์ง€๋งŒ ํ”Œ๋ž˜๊ทธ๊ฐ€ ์ง€์ •๋œ ํ‚ค์›Œ๋“œ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ ์˜์—ญ ์ „์ฒด๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
In this paper, we track the current state of synthetic data research and discuss current best practices and lessons learned. The rest of the paper is organized as follows. Section 2 provides an overview of synthetic data generation techniques and their applications in model training, presenting case studies and empirical evidence. Section 3 discusses the usefulness of synthetic data in evaluation. Section 4 discusses the challenges and limitations of synthetic data, and in Section 5 we outline potential solutions and future research directions.
์ด ๋…ผ๋ฌธ์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์—ฐ๊ตฌ์˜ ํ˜„์žฌ ์ƒํƒœ๋ฅผ ์ถ”์ ํ•˜๊ณ  ํ˜„์žฌ์˜ ๋ชจ๋ฒ” ์‚ฌ๋ก€์™€ ๋ฐฐ์šด ๊ตํ›ˆ์„ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์˜ ๋‚˜๋จธ์ง€ ๋ถ€๋ถ„์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. 2์žฅ์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๊ธฐ์ˆ ๊ณผ ๋ชจ๋ธ ํ›ˆ๋ จ์—์„œ์˜ ์‘์šฉ์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์‚ฌ๋ก€ ์—ฐ๊ตฌ์™€ ์‹ค์ฆ์  ์ฆ๊ฑฐ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. 3์žฅ์—์„œ๋Š” ํ‰๊ฐ€์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์œ ์šฉ์„ฑ์„ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. 4์žฅ์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๋„์ „ ๊ณผ์ œ์™€ ํ•œ๊ณ„๋ฅผ ๋…ผ์˜ํ•˜๊ณ , 5์žฅ์—์„œ๋Š” ์ž ์žฌ์  ํ•ด๊ฒฐ์ฑ…๊ณผ ๋ฏธ๋ž˜ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์„ ๊ฐœ๋žต์ ์œผ๋กœ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์—ฐ๊ตฌ์˜ ํ˜„์žฌ ์ƒํƒœ๋ฅผ ์ถ”์ ํ•˜๊ณ  ํ˜„์žฌ์˜ ๋ชจ๋ฒ” ์‚ฌ๋ก€์™€ ๊ตํ›ˆ์„ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€ ๋…ผ๋ฌธ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์„น์…˜ 2์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๊ธฐ๋ฒ•๊ณผ ๋ชจ๋ธ ํ•™์Šต์— ๋Œ€ํ•œ ์ ์šฉ์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์‚ฌ๋ก€ ์—ฐ๊ตฌ์™€ ๊ฒฝํ—˜์  ์ฆ๊ฑฐ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 3์—์„œ๋Š” ํ‰๊ฐ€์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์œ ์šฉ์„ฑ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 4์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๊ณผ์ œ์™€ ํ•œ๊ณ„์— ๋Œ€ํ•ด ๋…ผ์˜ํ•˜๊ณ , ์„น์…˜ 5์—์„œ๋Š” ์ž ์žฌ์ ์ธ ํ•ด๊ฒฐ์ฑ…๊ณผ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
As soon as the object made contact with their palm, they felt a sense of calm and clarity wash over them. They knew, in that moment, that they had discovered something truly incredible. Over the next few weeks, my character became obsessed with the object. They spent hours staring at it, examining it from every angle. They tried to use it for various tasks, but found that it had an almost mystical power. For instance, they could barely control their appetite when they held it, and any fears they may have had seemed to dissipate into the ether.
๊ทธ ๋ฌผ์ฒด๊ฐ€ ์†๋ฐ”๋‹ฅ์— ๋‹ฟ์ž๋งˆ์ž, ๊ทธ๋“ค์€ ํ‰์˜จํ•จ๊ณผ ๋ช…ํ™•ํ•จ์ด ๋ฐ€๋ ค์˜ค๋Š” ๊ฒƒ์„ ๋Š๊ผˆ๋‹ค. ๊ทธ ์ˆœ๊ฐ„, ๊ทธ๋“ค์€ ์ •๋ง๋กœ ๋†€๋ผ์šด ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Œ์„ ์•Œ์•˜๋‹ค. ๋‹ค์Œ ๋ช‡ ์ฃผ ๋™์•ˆ, ๋‚ด ์บ๋ฆญํ„ฐ๋Š” ๊ทธ ๋ฌผ์ฒด์— ์ง‘์ฐฉํ•˜๊ฒŒ ๋˜์—ˆ๋‹ค. ๊ทธ๋“ค์€ ๋ช‡ ์‹œ๊ฐ„ ๋™์•ˆ ๊ทธ๊ฒƒ์„ ์‘์‹œํ•˜๋ฉฐ ๋ชจ๋“  ๊ฐ๋„์—์„œ ์กฐ์‚ฌํ–ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๊ทธ๊ฒƒ์„ ์‚ฌ์šฉํ•˜๋ ค๊ณ  ์‹œ๋„ํ–ˆ์ง€๋งŒ, ๊ทธ๊ฒƒ์€ ๊ฑฐ์˜ ์‹ ๋น„๋กœ์šด ํž˜์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๊ฒƒ ๊ฐ™์•˜๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ทธ๊ฒƒ์„ ์ฅ๊ณ  ์žˆ์„ ๋•Œ๋Š” ์‹์š•์„ ๊ฑฐ์˜ ์กฐ์ ˆํ•  ์ˆ˜ ์—†์—ˆ๊ณ , ๊ฐ€์ง€๊ณ  ์žˆ๋˜ ๋‘๋ ค์›€์€ ๊ณต์ค‘์œผ๋กœ ์‚ฌ๋ผ์ง€๋Š” ๊ฒƒ ๊ฐ™์•˜๋‹ค.
๋ฌผ์ฒด๊ฐ€ ์†๋ฐ”๋‹ฅ์— ๋‹ฟ๋Š” ์ˆœ๊ฐ„, ๊ทธ๋“ค์€ ์นจ์ฐฉํ•จ๊ณผ ๋ช…๋ฃŒํ•จ์ด ๊ทธ๋“ค์„ ๋ฎ์น˜๋Š” ๊ฒƒ์„ ๋Š๊ผˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ๊ทธ ์ˆœ๊ฐ„ ๊ทธ๋“ค์ด ์ •๋ง ๋†€๋ผ์šด ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ ํ›„ ๋ช‡ ์ฃผ ๋™์•ˆ ๋‚ด ์บ๋ฆญํ„ฐ๋Š” ๊ทธ ๋ฌผ์ฒด์— ์ง‘์ฐฉํ•˜๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ๋ช‡ ์‹œ๊ฐ„ ๋™์•ˆ ๋ฌผ์ฒด๋ฅผ ์‘์‹œํ•˜๊ณ  ๋ชจ๋“  ๊ฐ๋„์—์„œ ๋ฌผ์ฒด๋ฅผ ์‚ดํŽด๋ณด์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๋ฌผ์ฒด๋ฅผ ์‚ฌ์šฉํ•˜๋ ค๊ณ  ํ–ˆ์ง€๋งŒ ๊ฑฐ์˜ ์‹ ๋น„๋กœ์šด ํž˜์ด ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ทธ๋“ค์€ ๋ฌผ์ฒด๋ฅผ ๋“ค์—ˆ์„ ๋•Œ ์‹์š•์„ ๊ฑฐ์˜ ์กฐ์ ˆํ•  ์ˆ˜ ์—†์—ˆ๊ณ , ๋‘๋ ค์›€์ด ์–ด๋””๋ก ๊ฐ€ ์‚ฌ๋ผ์ง€๋Š” ๊ฒƒ ๊ฐ™์•˜์Šต๋‹ˆ๋‹ค.
We initially randomly sampled 500 drug information pages from UpToDate.com, a medical reference website widely used by clinicians. To constrain the scope of questions, we specify in the prompt that the answer must be numerical and in milligrams. To filter out generated questions that did not meet the specified criteria (e.g. ambiguous question, incorrect units, etc.), we perform an additional quality control step, where we ask GPT-4 to verify that the generated question fulfills all criteria. After this step, we have 266 question-answer pairs.
์ฒ˜์Œ์— ์šฐ๋ฆฌ๋Š” ์ž„์˜๋กœ UpToDate.com์—์„œ 500๊ฐœ์˜ ์•ฝ๋ฌผ ์ •๋ณด ํŽ˜์ด์ง€๋ฅผ ์ƒ˜ํ”Œ๋งํ–ˆ์Šต๋‹ˆ๋‹ค. UpToDate.com์€ ์ž„์ƒ์˜๋“ค์ด ๋„๋ฆฌ ์‚ฌ์šฉํ•˜๋Š” ์˜ํ•™ ์ฐธ๊ณ  ์›น์‚ฌ์ดํŠธ์ž…๋‹ˆ๋‹ค. ์งˆ๋ฌธ์˜ ๋ฒ”์œ„๋ฅผ ์ œํ•œํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ํ”„๋กฌํ”„ํŠธ์—์„œ ๋‹ต๋ณ€์ด ๋ฐ˜๋“œ์‹œ ์ˆซ์ž์—ฌ์•ผ ํ•˜๋ฉฐ ๋ฐ€๋ฆฌ๊ทธ๋žจ ๋‹จ์œ„์—ฌ์•ผ ํ•œ๋‹ค๊ณ  ๋ช…์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ง€์ •๋œ ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜์ง€ ์•Š๋Š” ์ƒ์„ฑ๋œ ์งˆ๋ฌธ๋“ค(์˜ˆ: ๋ชจํ˜ธํ•œ ์งˆ๋ฌธ, ์ž˜๋ชป๋œ ๋‹จ์œ„ ๋“ฑ)์„ ๊ฑธ๋Ÿฌ๋‚ด๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์ถ”๊ฐ€์ ์ธ ํ’ˆ์งˆ ๊ด€๋ฆฌ ๋‹จ๊ณ„๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ GPT-4์—๊ฒŒ ์ƒ์„ฑ๋œ ์งˆ๋ฌธ์ด ๋ชจ๋“  ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋Š”์ง€ ํ™•์ธํ•˜๋„๋ก ์š”์ฒญํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๋ฅผ ๊ฑฐ์นœ ํ›„, ์šฐ๋ฆฌ๋Š” 266๊ฐœ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์„ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค.
์ž„์ƒ์˜๊ฐ€ ๋„๋ฆฌ ์‚ฌ์šฉํ•˜๋Š” ์˜ํ•™ ์ฐธ๊ณ  ์›น์‚ฌ์ดํŠธ์ธ UpToDate.com์—์„œ ์•ฝ๋ฌผ ์ •๋ณด ํŽ˜์ด์ง€ 500๊ฐœ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋งํ–ˆ์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ์˜ ๋ฒ”์œ„๋ฅผ ์ œํ•œํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋กฌํ”„ํŠธ์— ๋‹ต๋ณ€์ด ์ˆซ์ž์—ฌ์•ผ ํ•˜๋ฉฐ ๋ฐ€๋ฆฌ๊ทธ๋žจ ๋‹จ์œ„์—ฌ์•ผ ํ•œ๋‹ค๊ณ  ๋ช…์‹œํ•ฉ๋‹ˆ๋‹ค. ์ง€์ •๋œ ๊ธฐ์ค€(์˜ˆ: ๋ชจํ˜ธํ•œ ์งˆ๋ฌธ, ์ž˜๋ชป๋œ ๋‹จ์œ„ ๋“ฑ)์„ ์ถฉ์กฑํ•˜์ง€ ์•Š๋Š” ์ƒ์„ฑ๋œ ์งˆ๋ฌธ์„ ๊ฑธ๋Ÿฌ๋‚ด๊ธฐ ์œ„ํ•ด, ์ƒ์„ฑ๋œ ์งˆ๋ฌธ์ด ๋ชจ๋“  ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋Š”์ง€ GPT-4๊ฐ€ ํ™•์ธํ•˜๋„๋ก ์š”์ฒญํ•˜๋Š” ์ถ”๊ฐ€ ํ’ˆ์งˆ ๊ด€๋ฆฌ ๋‹จ๊ณ„๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๋ฅผ ๊ฑฐ์นœ ํ›„ 266๊ฐœ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์„ ํ™•๋ณดํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Scaling the open-source modes can improve the performance consistently._ By comparing the performance of Vicuna (7B) and Vicuna (13B), Pythia (7B) and Pythia (13B), we can see that the models with larger scales mostly perform better than smaller ones on these evaluation tasks, indicating the effectiveness of scaling up the model size. Across different tasks, scaling model is more beneficial for more complex tasks (_e.g._, symbolic and mathematical reasoning), where the larger models mostly outperform smaller ones in a large margin.
\(\bullet\) ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์˜ ๊ทœ๋ชจ๋ฅผ ํ™•์žฅํ•˜๋ฉด ์„ฑ๋Šฅ์ด ์ผ๊ด€๋˜๊ฒŒ ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Vicuna (7B)์™€ Vicuna (13B), Pythia (7B)์™€ Pythia (13B)์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ด ๋ณด๋ฉด, ๋” ํฐ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์ด ๋Œ€๋ถ€๋ถ„์˜ ํ‰๊ฐ€ ์ž‘์—…์—์„œ ๋” ์ž‘์€ ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์ด ํšจ๊ณผ์ ์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๊ฑธ์ณ, ๋ชจ๋ธ ํ™•์žฅ์€ ๋” ๋ณต์žกํ•œ ์ž‘์—…(์˜ˆ: ๊ธฐํ˜ธ์  ๋ฐ ์ˆ˜ํ•™์  ์ถ”๋ก )์—์„œ ๋” ์œ ์ตํ•˜๋ฉฐ, ์ด ๊ฒฝ์šฐ ๋” ํฐ ๋ชจ๋ธ์ด ๋Œ€๋ถ€๋ถ„์˜ ๊ฒฝ์šฐ ๋” ์ž‘์€ ๋ชจ๋ธ๋ณด๋‹ค ํฐ ์ฐจ์ด๋กœ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค.
\(\bullet\)_์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋“œ๋ฅผ ํ™•์žฅํ•˜๋ฉด ์„ฑ๋Šฅ์„ ์ผ๊ด€๋˜๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค._ Vicuna(7B)์™€ Vicuna(13B), Pythia(7B)์™€ Pythia(13B)์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜๋ฉด, ๋Œ€๋ถ€๋ถ„์˜ ๊ฒฝ์šฐ ๋” ํฐ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์ด ๋” ์ž‘์€ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ๋ณด๋‹ค ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ์ž‘์—…์—์„œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค, ์ด๋Š” ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์˜ ํšจ๊ณผ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ๋ชจ๋ธ์„ ํ™•์žฅํ•˜๋ฉด ๋” ๋ณต์žกํ•œ ์ž‘์—…(_์˜ˆ: ๊ธฐํ˜ธ ๋ฐ ์ˆ˜ํ•™์  ์ถ”๋ก )์— ๋” ์œ ๋ฆฌํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ ๋” ํฐ ๋ชจ๋ธ์ด ๋” ์ž‘์€ ๋ชจ๋ธ๋ณด๋‹ค ํฐ ํญ์œผ๋กœ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค.
**Wellbeing and self-care** Please ensure that you protect your wellbeing and practice self-care while annotating. Take regular breaks. Assess your mental wellbeing regularly. Do not work for prolonged periods of time. Remember the social impact of your work. Reach out to us if you have any issues, and we can connect you with counselling support. You can read more information about the challenges of this work online.36. It is normal to find this work difficult and/or distressing, and to need to talk about your experiences and to be given help.
**์›ฐ๋น™๊ณผ ์ž๊ธฐ ๊ด€๋ฆฌ** ์ฃผ์„ ์ž‘์—…์„ ํ•˜๋Š” ๋™์•ˆ ์ž์‹ ์˜ ์›ฐ๋น™์„ ๋ณดํ˜ธํ•˜๊ณ  ์ž๊ธฐ ๊ด€๋ฆฌ๋ฅผ ์‹ค์ฒœํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ •๊ธฐ์ ์œผ๋กœ ํœด์‹์„ ์ทจํ•˜์„ธ์š”. ์ •์‹ ์  ์›ฐ๋น™์„ ์ •๊ธฐ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜์„ธ์š”. ์žฅ์‹œ๊ฐ„ ์ž‘์—…ํ•˜์ง€ ๋งˆ์„ธ์š”. ์ž์‹ ์˜ ์ž‘์—…์ด ์‚ฌํšŒ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ธฐ์–ตํ•˜์„ธ์š”. ๋ฌธ์ œ๊ฐ€ ์žˆ์œผ๋ฉด ์ €ํฌ์—๊ฒŒ ์—ฐ๋ฝํ•ด ์ฃผ์„ธ์š”. ์ƒ๋‹ด ์ง€์›์„ ์—ฐ๊ฒฐํ•ด ๋“œ๋ฆด ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ์–ด๋ ค์›€์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์ •๋ณด๋ฅผ ์˜จ๋ผ์ธ์—์„œ ์ฝ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์ด ์–ด๋ ต๊ฑฐ๋‚˜ ๊ณ ํ†ต์Šค๋Ÿฝ๊ฒŒ ๋Š๊ปด์ง€๋Š” ๊ฒƒ์€ ์ •์ƒ์ด๋ฉฐ, ์ž์‹ ์˜ ๊ฒฝํ—˜์— ๋Œ€ํ•ด ์ด์•ผ๊ธฐํ•˜๊ณ  ๋„์›€์„ ๋ฐ›๋Š” ๊ฒƒ์ด ํ•„์š”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์›ฐ๋น™ ๋ฐ ์ž๊ธฐ ๊ด€๋ฆฌ** ์ฃผ์„์„ ๋‹ฌ๋Š” ๋™์•ˆ ์ž์‹ ์˜ ์›ฐ๋น™์„ ๋ณดํ˜ธํ•˜๊ณ  ์ž๊ธฐ ๊ด€๋ฆฌ๋ฅผ ์‹ค์ฒœํ•˜์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค. ์ •๊ธฐ์ ์œผ๋กœ ํœด์‹์„ ์ทจํ•˜์„ธ์š”. ์ •์‹  ๊ฑด๊ฐ•์„ ์ •๊ธฐ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜์„ธ์š”. ์žฅ์‹œ๊ฐ„ ์ž‘์—…ํ•˜์ง€ ๋งˆ์„ธ์š”. ์—…๋ฌด๊ฐ€ ์‚ฌํšŒ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ธฐ์–ตํ•˜์„ธ์š”. ๋ฌธ์ œ๊ฐ€ ์žˆ์œผ๋ฉด ๋‹น์‚ฌ์— ์—ฐ๋ฝํ•ด ์ฃผ์‹œ๋ฉด ์ƒ๋‹ด ์ง€์›์„ ์—ฐ๊ฒฐํ•ด ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ์–ด๋ ค์›€์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด๋Š” ์˜จ๋ผ์ธ์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.36. ์ด ์ž‘์—…์ด ์–ด๋ ต๊ณ /๊ฑฐ๋‚˜ ๊ณ ํ†ต์Šค๋Ÿฝ๊ณ , ์ž์‹ ์˜ ๊ฒฝํ—˜์— ๋Œ€ํ•ด ์ด์•ผ๊ธฐํ•˜๊ณ  ๋„์›€์„ ๋ฐ›์•„์•ผ ํ•˜๋Š” ๊ฒƒ์€ ์ •์ƒ์ž…๋‹ˆ๋‹ค.
**Evaluation Dataset.** These datasets are mainly augmented from the training sets of GSM8K and MATH, and we test on the corresponding test sets. The GSM8K dataset sets aside 1,319 entries for testing, while the MATH dataset reserves 5,000 entries as a test set. We extract the final answers by specific templates (e.g., "the answer is") in the predicted reasoning paths following the scripts adopted by Yue et al. (2023); Yu et al. (2023). Subsequently, we calculate the ratio of predicted answers equal to the ground truth answers as the accuracy.
**ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ์…‹.** ์ด ๋ฐ์ดํ„ฐ์…‹๋“ค์€ ์ฃผ๋กœ GSM8K์™€ MATH์˜ ํ›ˆ๋ จ ์„ธํŠธ์—์„œ ์ฆ๊ฐ•๋˜์—ˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ๋Š” ํ•ด๋‹น ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ํ…Œ์ŠคํŠธ๋ฅผ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค. GSM8K ๋ฐ์ดํ„ฐ์…‹์€ ํ…Œ์ŠคํŠธ๋ฅผ ์œ„ํ•ด 1,319๊ฐœ์˜ ํ•ญ๋ชฉ์„ ๋”ฐ๋กœ ๋‘๊ณ , MATH ๋ฐ์ดํ„ฐ์…‹์€ 5,000๊ฐœ์˜ ํ•ญ๋ชฉ์„ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋กœ ์˜ˆ์•ฝํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” Yue et al. (2023); Yu et al. (2023)์ด ์ฑ„ํƒํ•œ ์Šคํฌ๋ฆฝํŠธ๋ฅผ ๋”ฐ๋ผ ์˜ˆ์ธก๋œ ์ถ”๋ก  ๊ฒฝ๋กœ์—์„œ ํŠน์ • ํ…œํ”Œ๋ฆฟ(์˜ˆ: "์ •๋‹ต์€")์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข… ๋‹ต์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ์ดํ›„, ์˜ˆ์ธก๋œ ๋‹ต์ด ์‹ค์ œ ์ •๋‹ต๊ณผ ์ผ์น˜ํ•˜๋Š” ๋น„์œจ์„ ์ •ํ™•๋„๋กœ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
**ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ.** ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ฃผ๋กœ GSM8K ๋ฐ MATH์˜ ํ›ˆ๋ จ ์„ธํŠธ์—์„œ ์ฆ๊ฐ•๋˜๋ฉฐ, ํ•ด๋‹น ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. GSM8K ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 1,319๊ฐœ์˜ ํ•ญ๋ชฉ์„ ํ…Œ์ŠคํŠธ์šฉ์œผ๋กœ ๋”ฐ๋กœ ์„ค์ •ํ•˜๊ณ , MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 5,000๊ฐœ์˜ ํ•ญ๋ชฉ์„ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋กœ ์˜ˆ์•ฝํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ์ธก๋œ ์ถ”๋ก  ๊ฒฝ๋กœ์—์„œ Yue ๋“ฑ(2023); Yu ๋“ฑ(2023)์ด ์ฑ„ํƒํ•œ ์Šคํฌ๋ฆฝํŠธ์— ๋”ฐ๋ผ "์ •๋‹ต์€"๊ณผ ๊ฐ™์€ ํŠน์ • ํ…œํ”Œ๋ฆฟ์œผ๋กœ ์ตœ์ข… ์ •๋‹ต์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ์ •๋‹ต๊ณผ ์ผ์น˜ํ•˜๋Š” ์˜ˆ์ธก๋œ ์ •๋‹ต์˜ ๋น„์œจ์„ ์ •ํ™•๋„๋กœ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
To ensure the quality and diversity of our data, we manually selected 13 data sources from high-quality websites and data resources within the Chinese Internet. These sources include community Q&A forums, encyclopedic sites, content creation platforms, examinations, etc. We also incorporated high-quality Chinese NLP datasets to enrich the diversity of tasks. Specifically, we categorized all data sources into four types: Social Media & forums, World Knowledge, NLP tasks, and Examinations. The data sources and their descriptions are as follows.
๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์ค‘๊ตญ ์ธํ„ฐ๋„ท ๋‚ด์˜ ๊ณ ํ’ˆ์งˆ ์›น์‚ฌ์ดํŠธ์™€ ๋ฐ์ดํ„ฐ ์ž์›์—์„œ 13๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ์ˆ˜๋™์œผ๋กœ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์†Œ์Šค์—๋Š” ์ปค๋ฎค๋‹ˆํ‹ฐ Q&A ํฌ๋Ÿผ, ๋ฐฑ๊ณผ์‚ฌ์ „ ์‚ฌ์ดํŠธ, ์ฝ˜ํ…์ธ  ์ƒ์„ฑ ํ”Œ๋žซํผ, ์‹œํ—˜ ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ž‘์—…์˜ ๋‹ค์–‘์„ฑ์„ ํ’๋ถ€ํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ํ’ˆ์งˆ์˜ ์ค‘๊ตญ์–ด NLP ๋ฐ์ดํ„ฐ์…‹๋„ ํ†ตํ•ฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ๋„ค ๊ฐ€์ง€ ์œ ํ˜•์œผ๋กœ ๋ถ„๋ฅ˜ํ–ˆ์Šต๋‹ˆ๋‹ค: ์†Œ์…œ ๋ฏธ๋””์–ด ๋ฐ ํฌ๋Ÿผ, ์„ธ๊ณ„ ์ง€์‹, NLP ์ž‘์—…, ๊ทธ๋ฆฌ๊ณ  ์‹œํ—˜. ๋ฐ์ดํ„ฐ ์†Œ์Šค์™€ ๊ทธ ์„ค๋ช…์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์ค‘๊ตญ ์ธํ„ฐ๋„ท ๋‚ด ๊ณ ํ’ˆ์งˆ ์›น์‚ฌ์ดํŠธ์™€ ๋ฐ์ดํ„ฐ ๋ฆฌ์†Œ์Šค์—์„œ 13๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ์ˆ˜๋™์œผ๋กœ ์„ ๋ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์†Œ์Šค์—๋Š” ์ปค๋ฎค๋‹ˆํ‹ฐ Q&A ํฌ๋Ÿผ, ๋ฐฑ๊ณผ์‚ฌ์ „ ์‚ฌ์ดํŠธ, ์ฝ˜ํ…์ธ  ์ œ์ž‘ ํ”Œ๋žซํผ, ์‹œํ—˜ ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ณ ํ’ˆ์งˆ์˜ ์ค‘๊ตญ NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ์ž‘์—…์˜ ๋‹ค์–‘์„ฑ์„ ๊ฐ•ํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ์†Œ์…œ ๋ฏธ๋””์–ด ๋ฐ ํฌ๋Ÿผ, ์„ธ๊ณ„ ์ง€์‹, NLP ์ž‘์—…, ์‹œํ—˜์˜ ๋„ค ๊ฐ€์ง€ ์œ ํ˜•์œผ๋กœ ๋ถ„๋ฅ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์†Œ์Šค์™€ ๊ทธ ์„ค๋ช…์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
**EM** provided input on early discussions on learning autoregressive reward functions; wrote the first implementation of DPO and ran the first DPO experiments; trained the large-scale (summarization and dialogue) DPO models used in paper experiments; conducted initial GPT-4 win rate evaluations and set up related infrastructure; recruited participants for, conducted, and analyzed results from the human study; wrote the abstract, introduction, related work, discussion, and most of experiments; and assisted with editing the rest of the paper.
EM์€ ์ž์œจํšŒ๊ท€ ๋ณด์ƒ ํ•จ์ˆ˜ ํ•™์Šต์— ๋Œ€ํ•œ ์ดˆ๊ธฐ ๋…ผ์˜์— ์ž…๋ ฅ์„ ์ œ๊ณตํ–ˆ์œผ๋ฉฐ, DPO์˜ ์ฒซ ๋ฒˆ์งธ ๊ตฌํ˜„์„ ์ž‘์„ฑํ•˜๊ณ  ์ฒซ ๋ฒˆ์งธ DPO ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋…ผ๋ฌธ ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋Œ€๊ทœ๋ชจ DPO ๋ชจ๋ธ(์š”์•ฝ ๋ฐ ๋Œ€ํ™”)์„ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, ์ดˆ๊ธฐ GPT-4 ์Šน๋ฅ  ํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ  ๊ด€๋ จ ์ธํ”„๋ผ๋ฅผ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ธ๊ฐ„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ์ฐธ๊ฐ€์ž๋ฅผ ๋ชจ์ง‘ํ•˜๊ณ  ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜๋ฉฐ ๊ฒฐ๊ณผ๋ฅผ ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ดˆ๋ก, ์„œ๋ก , ๊ด€๋ จ ์—ฐ๊ตฌ, ๋…ผ์˜ ๋ฐ ๋Œ€๋ถ€๋ถ„์˜ ์‹คํ—˜ ๋ถ€๋ถ„์„ ์ž‘์„ฑํ–ˆ์œผ๋ฉฐ, ๋‚˜๋จธ์ง€ ๋…ผ๋ฌธ์˜ ํŽธ์ง‘์„ ๋„์™”์Šต๋‹ˆ๋‹ค.
**EM**์€ ํ•™์Šต ์ž๋™ ํšŒ๊ท€ ๋ณด์ƒ ํ•จ์ˆ˜์— ๋Œ€ํ•œ ์ดˆ๊ธฐ ๋…ผ์˜์— ๋Œ€ํ•œ ์˜๊ฒฌ์„ ์ œ๊ณตํ•˜๊ณ , DPO์˜ ์ฒซ ๋ฒˆ์งธ ๊ตฌํ˜„์„ ์ž‘์„ฑํ•˜๊ณ  ์ฒซ ๋ฒˆ์งธ DPO ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์œผ๋ฉฐ, ๋…ผ๋ฌธ ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋Œ€๊ทœ๋ชจ(์š”์•ฝ ๋ฐ ๋Œ€ํ™”) DPO ๋ชจ๋ธ์„ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, ์ดˆ๊ธฐ GPT-4 ์Šน๋ฅ  ํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ  ๊ด€๋ จ ์ธํ”„๋ผ๋ฅผ ๊ตฌ์ถ•ํ–ˆ์œผ๋ฉฐ, ์ธ๊ฐ„ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ฐธ๊ฐ€์ž๋ฅผ ๋ชจ์ง‘ํ•˜๊ณ  ์ˆ˜ํ–‰ ๋ฐ ๊ฒฐ๊ณผ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ–ˆ์œผ๋ฉฐ, ์ดˆ๋ก, ์„œ๋ก , ๊ด€๋ จ ์—ฐ๊ตฌ, ํ† ๋ก  ๋ฐ ๋Œ€๋ถ€๋ถ„์˜ ์‹คํ—˜์„ ์ž‘์„ฑํ–ˆ์œผ๋ฉฐ ๋‚˜๋จธ์ง€ ๋…ผ๋ฌธ ํŽธ์ง‘์„ ์ง€์›ํ–ˆ์Šต๋‹ˆ๋‹ค.
For solving a complex task, the task planner first needs to clearly understand the task goal and generate a reasonable plan based on the reasoning of LLMs (See Section 6.4.2). Then, the plan executor acts according to the plan in the environment, and the environment will produce feedback for the task planner (See Section 6.4.3). The task planner can further incorporate the feedback obtained from the environment to refine its initial plan and iteratively perform the above process to get better results as the task solution (See Section 6.4.4).
๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด, ์ž‘์—… ๊ณ„ํš์ž๋Š” ๋จผ์ € ์ž‘์—… ๋ชฉํ‘œ๋ฅผ ๋ช…ํ™•ํžˆ ์ดํ•ดํ•˜๊ณ  LLMs์˜ ์ถ”๋ก ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋ฆฌ์ ์ธ ๊ณ„ํš์„ ์ƒ์„ฑํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค (์„น์…˜ 6.4.2 ์ฐธ์กฐ). ๊ทธ๋Ÿฐ ๋‹ค์Œ, ๊ณ„ํš ์‹คํ–‰์ž๋Š” ํ™˜๊ฒฝ์—์„œ ๊ณ„ํš์— ๋”ฐ๋ผ ํ–‰๋™ํ•˜๋ฉฐ, ํ™˜๊ฒฝ์€ ์ž‘์—… ๊ณ„ํš์ž์—๊ฒŒ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค (์„น์…˜ 6.4.3 ์ฐธ์กฐ). ์ž‘์—… ๊ณ„ํš์ž๋Š” ํ™˜๊ฒฝ์—์„œ ์–ป์€ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ฉํ•˜์—ฌ ์ดˆ๊ธฐ ๊ณ„ํš์„ ์ˆ˜์ •ํ•˜๊ณ , ์œ„์˜ ๊ณผ์ •์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•˜์—ฌ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์ž‘์—… ํ•ด๊ฒฐ์ฑ…์œผ๋กœ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค (์„น์…˜ 6.4.4 ์ฐธ์กฐ).
๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ž‘์—… ํ”Œ๋ž˜๋„ˆ๋Š” ๋จผ์ € ์ž‘์—… ๋ชฉํ‘œ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ์ดํ•ดํ•˜๊ณ  LLM์˜ ์ถ”๋ก ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋ฆฌ์ ์ธ ๊ณ„ํš์„ ์ƒ์„ฑํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค(์„น์…˜ 6.4.2 ์ฐธ์กฐ). ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ณ„ํš ์‹คํ–‰์ž๋Š” ํ™˜๊ฒฝ์—์„œ ๊ณ„ํš์— ๋”ฐ๋ผ ํ–‰๋™ํ•˜๊ณ  ํ™˜๊ฒฝ์€ ์ž‘์—… ํ”Œ๋ž˜๋„ˆ์— ํ”ผ๋“œ๋ฐฑ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค(์„น์…˜ 6.4.3 ์ฐธ์กฐ). ์ž‘์—… ํ”Œ๋ž˜๋„ˆ๋Š” ํ™˜๊ฒฝ์—์„œ ์–ป์€ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ฉํ•˜์—ฌ ์ดˆ๊ธฐ ๊ณ„ํš์„ ๊ตฌ์ฒดํ™”ํ•˜๊ณ  ์œ„์˜ ํ”„๋กœ์„ธ์Šค๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ž‘์—… ์†”๋ฃจ์…˜์œผ๋กœ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์„น์…˜ 6.4.4 ์ฐธ์กฐ).
Pythia and deduplication.The Pythia series of models is available in two flavours: one trained on the vanilla version of The Pile, and another trained on a version deduplicated with MinHash. Performance between these two flavours was noted to minimally differ (Biderman et al., 2023); in Figure 9, we find the deduplicated version may be slightly ahead of the non-deduplicated one under our aggregate. The higher end of this improvement is broadly in line with our findings in Table 5. Nevertheless, a difference in our findings and theirs remain.
Pythia์™€ ์ค‘๋ณต ์ œ๊ฑฐ. Pythia ๋ชจ๋ธ ์‹œ๋ฆฌ์ฆˆ๋Š” ๋‘ ๊ฐ€์ง€ ๋ฒ„์ „์œผ๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค: The Pile์˜ ๊ธฐ๋ณธ ๋ฒ„์ „์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๊ณผ MinHash๋กœ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๋ฒ„์ „์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ. ์ด ๋‘ ๋ฒ„์ „ ๊ฐ„์˜ ์„ฑ๋Šฅ ์ฐจ์ด๋Š” ์ตœ์†Œํ•œ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค (Biderman et al., 2023); Figure 9์—์„œ, ์šฐ๋ฆฌ๋Š” ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๋ฒ„์ „์ด ๋น„์ค‘๋ณต ๋ฒ„์ „๋ณด๋‹ค ์•ฝ๊ฐ„ ์•ž์„œ ์žˆ์„ ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์˜ ์ƒํ•œ์„ ์€ Table 5์—์„œ ์šฐ๋ฆฌ์˜ ๋ฐœ๊ฒฌ๊ณผ ๋Œ€์ฒด๋กœ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์šฐ๋ฆฌ์˜ ๋ฐœ๊ฒฌ๊ณผ ๊ทธ๋“ค์˜ ๋ฐœ๊ฒฌ ์‚ฌ์ด์—๋Š” ์ฐจ์ด๊ฐ€ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ”ผํ‹ฐ์•„์™€ ์ค‘๋ณต ์ œ๊ฑฐ.Pythia ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์€ ๋‘ ๊ฐ€์ง€ ๋ง›์œผ๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค: ํ•˜๋‚˜๋Š” ๋ฐ”๋‹๋ผ ๋ฒ„์ „์˜ ๋” ํŒŒ์ผ๋กœ ํ•™์Šต๋˜์—ˆ๊ณ , ๋‹ค๋ฅธ ํ•˜๋‚˜๋Š” MinHash๋กœ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๋ฒ„์ „์œผ๋กœ ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋‘ ๊ฐ€์ง€ ๋ง› ์‚ฌ์ด์˜ ์„ฑ๋Šฅ์€ ๊ฑฐ์˜ ์ฐจ์ด๊ฐ€ ์—†๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Biderman et al., 2023); ๊ทธ๋ฆผ 9์—์„œ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๋ฒ„์ „์ด ์šฐ๋ฆฌ์˜ ์ง‘๊ณ„ ํ•˜์—์„œ ์ค‘๋ณต ์ œ๊ฑฐ๋˜์ง€ ์•Š์€ ๋ฒ„์ „๋ณด๋‹ค ์•ฝ๊ฐ„ ์•ž์„œ ์žˆ๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฐœ์„ ์˜ ์ƒํ•œ์€ ํ‘œ 5์˜ ๊ฒฐ๊ณผ์™€ ๋Œ€์ฒด๋กœ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์™€ ๊ทธ๋“ค์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ ์‚ฌ์ด์—๋Š” ์ฐจ์ด๊ฐ€ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค.
Considering the reasons and their strengths, it seems there is a balanced set of factors that could influence Trump's decision to participate or not. His history of unconventional tactics and concerns about debate arrangements could lead to a "No," while the potential benefits of participating in terms of policy promotion, engaging opponents, and demonstrating leadership suggest a "Yes." The decision will likely depend on the specific circumstances closer to the debate date and Trump's assessment of the strategic value of his participation.
์ด์œ ์™€ ๊ทธ ๊ฐ•์ ์„ ๊ณ ๋ คํ•  ๋•Œ, ํŠธ๋Ÿผํ”„๊ฐ€ ํ† ๋ก ์— ์ฐธ์—ฌํ• ์ง€ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๋ฐ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ๊ท ํ˜• ์žกํžŒ ์š”์ธ๋“ค์ด ์žˆ๋Š” ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๋น„์ „ํ†ต์ ์ธ ์ „์ˆ ์„ ์‚ฌ์šฉํ•ด์˜จ ๊ทธ์˜ ์—ญ์‚ฌ์™€ ํ† ๋ก  ์ค€๋น„์— ๋Œ€ํ•œ ์šฐ๋ ค๋Š” "์•„๋‹ˆ์˜ค"๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์ง€๋งŒ, ์ •์ฑ… ํ™๋ณด, ์ƒ๋Œ€์™€์˜ ๊ต๋ฅ˜, ๋ฆฌ๋”์‹ญ์„ ๋ณด์—ฌ์ฃผ๋Š” ์ธก๋ฉด์—์„œ ์ฐธ์—ฌ์˜ ์ž ์žฌ์  ์ด์ ์€ "์˜ˆ"๋ฅผ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฐ์ •์€ ํ† ๋ก  ๋‚ ์งœ์— ๊ฐ€๊นŒ์›Œ์งˆ์ˆ˜๋ก ํŠน์ • ์ƒํ™ฉ๊ณผ ํŠธ๋Ÿผํ”„๊ฐ€ ๊ทธ์˜ ์ฐธ์—ฌ์˜ ์ „๋žต์  ๊ฐ€์น˜๋ฅผ ์–ด๋–ป๊ฒŒ ํ‰๊ฐ€ํ•˜๋Š”์ง€์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์งˆ ๊ฐ€๋Šฅ์„ฑ์ด ํฝ๋‹ˆ๋‹ค.
์ด์œ ์™€ ๊ทธ ๊ฐ•๋„๋ฅผ ๊ณ ๋ คํ•  ๋•Œ ํŠธ๋Ÿผํ”„๊ฐ€ ์ฐธ์—ฌ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๋ฐ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ๊ท ํ˜• ์žกํžŒ ์š”์ธ๋“ค์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ํŠธ๋Ÿผํ”„์˜ ํŒŒ๊ฒฉ์ ์ธ ์ „์ˆ ์˜ ์—ญ์‚ฌ์™€ ํ† ๋ก  ์ค€๋น„์— ๋Œ€ํ•œ ์šฐ๋ ค๋Š” '์•ˆ๋ผ'๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์ง€๋งŒ, ์ •์ฑ… ํ™๋ณด, ์ƒ๋Œ€๋ฐฉ ์ฐธ์—ฌ, ๋ฆฌ๋”์‹ญ ๋ฐœํœ˜ ๋“ฑ ์ฐธ์—ฌ๋ฅผ ํ†ตํ•ด ์–ป์„ ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ์  ์ด์ ์€ '๊ทธ๋ž˜'๋ฅผ ์•”์‹œํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ์ •์€ ํ† ๋ก  ๋‚ ์งœ์— ๊ฐ€๊นŒ์›Œ์งˆ์ˆ˜๋ก ๊ตฌ์ฒด์ ์ธ ์ƒํ™ฉ๊ณผ ํŠธ๋Ÿผํ”„์˜ ์ฐธ์—ฌ ์ „๋žต์  ๊ฐ€์น˜์— ๋Œ€ํ•œ ํ‰๊ฐ€์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์งˆ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค.
Finally, the last column of Table 3 (No \(\langle\)RET\(\rangle\)) shows the performance of Adapt-LLM when answering questions based solely on its parametric memory. As can be seen, accuracies above 62% are obtained when no context is utilized, providing further evidence that Adapt-LLM effectively discerns between retrieving context and providing direct answers to questions. Additionally, we evaluate the performance of these questions when context is added to the input, revealing significant decreases in accuracy of up to 7 absolute points.
๋งˆ์ง€๋ง‰์œผ๋กœ, Table 3์˜ ๋งˆ์ง€๋ง‰ ์—ด (No \(\langle\)RET\(\rangle\))์€ ์˜ค๋กœ์ง€ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ์งˆ๋ฌธ์— ๋‹ตํ•  ๋•Œ์˜ Adapt-LLM ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ณด์‹œ๋‹ค์‹œํ”ผ, ๋งฅ๋ฝ์ด ์‚ฌ์šฉ๋˜์ง€ ์•Š์„ ๋•Œ๋„ 62% ์ด์ƒ์˜ ์ •ํ™•๋„๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” Adapt-LLM์ด ๋งฅ๋ฝ์„ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒƒ๊ณผ ์งˆ๋ฌธ์— ์ง์ ‘ ๋‹ตํ•˜๋Š” ๊ฒƒ์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ตฌ๋ถ„ํ•œ๋‹ค๋Š” ์ถ”๊ฐ€ ์ฆ๊ฑฐ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ž…๋ ฅ์— ๋งฅ๋ฝ์„ ์ถ”๊ฐ€ํ–ˆ์„ ๋•Œ ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, ์ •ํ™•๋„๊ฐ€ ์ตœ๋Œ€ 7 ์ ˆ๋Œ€ ํฌ์ธํŠธ๊นŒ์ง€ ํฌ๊ฒŒ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
๋งˆ์ง€๋ง‰์œผ๋กœ ํ‘œ 3์˜ ๋งˆ์ง€๋ง‰ ์—ด(No \(\langle\)RET\(\rangle\))์€ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ๋งŒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์งˆ๋ฌธ์— ๋‹ตํ•  ๋•Œ Adapt-LLM์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ณด์‹œ๋‹ค์‹œํ”ผ, ๋ฌธ๋งฅ์„ ํ™œ์šฉํ•˜์ง€ ์•Š์„ ๋•Œ 62% ์ด์ƒ์˜ ์ •ํ™•๋„๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” Adapt-LLM์ด ๋ฌธ๋งฅ ๊ฒ€์ƒ‰๊ณผ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์ง์ ‘์ ์ธ ๋‹ต๋ณ€์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ๋Š” ์ถ”๊ฐ€์ ์ธ ์ฆ๊ฑฐ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž…๋ ฅ์— ๋ฌธ๋งฅ์„ ์ถ”๊ฐ€ํ–ˆ์„ ๋•Œ ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, ์ •ํ™•๋„๊ฐ€ ์ตœ๋Œ€ 7% ์ ˆ๋Œ€ ํฌ์ธํŠธ๊นŒ์ง€ ํฌ๊ฒŒ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
PersonA: Urinalysis can be performed in various healthcare settings, including doctors' offices, urgent care facilities, laboratories, and hospitals. You can provide a urine sample in a specimen cup, and the sample can be analyzed on-siteor sent to a laboratory for testing. PersonU: How much urine is typically needed for the test? PersonA: Usually, only small amounts of urine, ranging from 30 to 60 ml, are required for urinalysis testing. It's a non-invasive and relatively simple test that provides valuable information about your health.
PersonA: ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ์˜์‚ฌ ์‚ฌ๋ฌด์‹ค, ์‘๊ธ‰ ์น˜๋ฃŒ ์‹œ์„ค, ์‹คํ—˜์‹ค ๋ฐ ๋ณ‘์›์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ํ™˜๊ฒฝ์—์„œ ์ˆ˜ํ–‰๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ‘œ๋ณธ ์ปต์— ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ƒ˜ํ”Œ์€ ํ˜„์žฅ์—์„œ ๋ถ„์„๋˜๊ฑฐ๋‚˜ ์‹คํ—˜์‹ค๋กœ ๋ณด๋‚ด์ ธ ๊ฒ€์‚ฌ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. PersonU: ๊ฒ€์‚ฌ์— ์ผ๋ฐ˜์ ์œผ๋กœ ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ์†Œ๋ณ€์ด ํ•„์š”ํ•ฉ๋‹ˆ๊นŒ? PersonA: ๋ณดํ†ต 30์—์„œ 60ml ์ •๋„์˜ ์†Œ๋Ÿ‰์˜ ์†Œ๋ณ€๋งŒ ์†Œ๋ณ€ ๊ฒ€์‚ฌ์— ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋น„์นจ์Šต์ ์ด๊ณ  ๋น„๊ต์  ๊ฐ„๋‹จํ•œ ๊ฒ€์‚ฌ๋กœ, ๊ฑด๊ฐ•์— ๋Œ€ํ•œ ์œ ์šฉํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ์˜์‚ฌ ์‚ฌ๋ฌด์‹ค, ์‘๊ธ‰ ์ง„๋ฃŒ ์‹œ์„ค, ์‹คํ—˜์‹ค, ๋ณ‘์› ๋“ฑ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ํ™˜๊ฒฝ์—์„œ ์‹œํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฒด ์ปต์— ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•˜๋ฉด ์ฆ‰์‹œ ๋ถ„์„ํ•˜๊ฑฐ๋‚˜ ๊ฒ€์‚ฌ๋ฅผ ์œ„ํ•ด ์‹คํ—˜์‹ค๋กœ ๋ณด๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ๋žŒU: ๊ฒ€์‚ฌ์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ์†Œ๋ณ€์ด ํ•„์š”ํ•˜๋‚˜์š”? ์‚ฌ๋žŒA: ์ผ๋ฐ˜์ ์œผ๋กœ ์†Œ๋ณ€ ๊ฒ€์‚ฌ์—๋Š” 30~60ml ์ •๋„์˜ ์†Œ๋Ÿ‰์˜ ์†Œ๋ณ€๋งŒ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋น„์นจ์Šต์ ์ด๊ณ  ๋น„๊ต์  ๊ฐ„๋‹จํ•œ ๊ฒ€์‚ฌ๋กœ ๊ฑด๊ฐ•์— ๋Œ€ํ•œ ๊ท€์ค‘ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
Large language models (LLMs) store vast amounts of factual knowledge in their parameters through large-scale pre-training, and this knowledge can be used to answer various questions such as "where is the world's largest ice sheet located" (Brown et al., 2020; OpenAI, 2023; Chowdhery et al., 2022; Zhang et al., 2022; Touvron et al., 2023, 2023; Gemini Team, 2023). However, this factual knowledge is static, meaning that it can become outdated as the world evolves, or prove insufficient when LLMs are used in specialized or private domains.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLMs)์€ ๋Œ€๊ทœ๋ชจ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ๋งค๊ฐœ๋ณ€์ˆ˜์— ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์‚ฌ์‹ค์  ์ง€์‹์„ ์ €์žฅํ•˜๋ฉฐ, ์ด ์ง€์‹์€ "์„ธ๊ณ„์—์„œ ๊ฐ€์žฅ ํฐ ๋น™์ƒ์ด ์–ด๋””์— ์œ„์น˜ํ•ด ์žˆ๋Š”๊ฐ€"์™€ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค (Brown et al., 2020; OpenAI, 2023; Chowdhery et al., 2022; Zhang et al., 2022; Touvron et al., 2023, 2023; Gemini Team, 2023). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์‚ฌ์‹ค์  ์ง€์‹์€ ์ •์ ์ด๊ธฐ ๋•Œ๋ฌธ์— ์„ธ๊ณ„๊ฐ€ ๋ณ€ํ™”ํ•จ์— ๋”ฐ๋ผ ๊ตฌ์‹์ด ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, LLM์ด ํŠน์ˆ˜ํ•˜๊ฑฐ๋‚˜ ๋น„๊ณต๊ฐœ ๋„๋ฉ”์ธ์—์„œ ์‚ฌ์šฉ๋  ๋•Œ๋Š” ๋ถˆ์ถฉ๋ถ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ๋Œ€๊ทœ๋ชจ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์‚ฌ์‹ค์  ์ง€์‹์„ ํŒŒ๋ผ๋ฏธํ„ฐ์— ์ €์žฅํ•˜๊ณ , ์ด๋Ÿฌํ•œ ์ง€์‹์€ "์„ธ๊ณ„์—์„œ ๊ฐ€์žฅ ํฐ ๋น™์ƒ์€ ์–ด๋””์— ์œ„์น˜ํ•ด ์žˆ๋‚˜์š”?"์™€ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Brown et al., 2020; OpenAI, 2023; Chowdhery et al., 2022; Zhang et al., 2022; Touvron et al., 2023, 2023; Gemini Team, 2023). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์‚ฌ์‹ค์  ์ง€์‹์€ ์ •์ ์ธ ์ƒํƒœ์ด๋ฏ€๋กœ ์„ธ์ƒ์ด ์ง„ํ™”ํ•จ์— ๋”ฐ๋ผ ์˜ค๋ž˜๋œ ์ •๋ณด๊ฐ€ ๋  ์ˆ˜๋„ ์žˆ๊ณ , LLM์ด ํŠน์ˆ˜ ๋˜๋Š” ๋น„๊ณต๊ฐœ ์˜์—ญ์—์„œ ์‚ฌ์šฉ๋  ๋•Œ ๋ถ€์กฑํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
for this task, which contains questions created by crowdworkers based on events from news articles. Zou et al. (2022) introduced Autocast, a benchmark dataset compiled from forecasting competition questions up to 2022. In a competition with a large prize pool, no machine learning system was able to approach the performance of human forecasters on Autocast (Zou et al., 2022). The knowledge cut-offs of LMs have moved past 2022, necessitating more recent data. In this work, we source questions in 2023-2024, enabling us to apply recent LMs.
์ด ์ž‘์—…์„ ์œ„ํ•ด, ๋‰ด์Šค ๊ธฐ์‚ฌ์—์„œ ๋ฐœ์ƒํ•œ ์‚ฌ๊ฑด์„ ๋ฐ”ํƒ•์œผ๋กœ ํฌ๋ผ์šฐ๋“œ์›Œ์ปค๋“ค์ด ์ž‘์„ฑํ•œ ์งˆ๋ฌธ๋“ค์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. Zou ๋“ฑ(2022)์€ 2022๋…„๊นŒ์ง€์˜ ์˜ˆ์ธก ๋Œ€ํšŒ ์งˆ๋ฌธ๋“ค๋กœ ๊ตฌ์„ฑ๋œ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ์…‹์ธ Autocast๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ํฐ ์ƒ๊ธˆ์ด ๊ฑธ๋ฆฐ ๋Œ€ํšŒ์—์„œ, ์–ด๋–ค ๋จธ์‹  ๋Ÿฌ๋‹ ์‹œ์Šคํ…œ๋„ Autocast์—์„œ ์ธ๊ฐ„ ์˜ˆ์ธก๊ฐ€์˜ ์„ฑ๋Šฅ์— ์ ‘๊ทผํ•  ์ˆ˜ ์—†์—ˆ์Šต๋‹ˆ๋‹ค(Zou ๋“ฑ, 2022). LMs์˜ ์ง€์‹ ์ปท์˜คํ”„๋Š” 2022๋…„์„ ๋„˜์–ด์„ฐ์œผ๋ฉฐ, ๋” ์ตœ๊ทผ์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” 2023-2024๋…„์˜ ์งˆ๋ฌธ๋“ค์„ ์†Œ์‹ฑํ•˜์—ฌ ์ตœ์‹  LMs๋ฅผ ์ ์šฉํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ–ˆ์Šต๋‹ˆ๋‹ค.
์ด ์ž‘์—…์—๋Š” ๋‰ด์Šค ๊ธฐ์‚ฌ์˜ ์‚ฌ๊ฑด์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํฌ๋ผ์šฐ๋“œ์›Œ์ปค๊ฐ€ ๋งŒ๋“  ์งˆ๋ฌธ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. Zou et al. (2022)์€ 2022๋…„๊นŒ์ง€์˜ ์˜ˆ์ธก ๊ฒฝ์ง„๋Œ€ํšŒ ์งˆ๋ฌธ์„ ๋ชจ์•„ ๋งŒ๋“  ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ Autocast๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ƒ๋‹นํ•œ ์ƒ๊ธˆ์ด ๊ฑธ๋ฆฐ ๊ฒฝ์Ÿ์—์„œ๋„ ์–ด๋–ค ๋จธ์‹ ๋Ÿฌ๋‹ ์‹œ์Šคํ…œ๋„ Autocast์—์„œ ์ธ๊ฐ„ ์˜ˆ์ธก์ž์˜ ์„ฑ๋Šฅ์— ๊ทผ์ ‘ํ•˜์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค(Zou et al., 2022). LM์˜ ์ง€์‹ ์ปท์˜คํ”„๊ฐ€ 2022๋…„์„ ๋„˜์–ด์„œ๋ฉด์„œ ๋” ์ตœ๊ทผ์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•˜๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” 2023~2024๋…„์— ์งˆ๋ฌธ์„ ์†Œ์‹ฑํ•˜์—ฌ ์ตœ์‹  LM์„ ์ ์šฉํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
Task: Given a dialogue, classify whether the user is satisfied with the service. You should respond with "Satisfied" or "Unsatisfied". Class label: Satisfied Dialogue: - Agent: Thank you for your feedback. We will work to improve our service in the future. - Customer: I am happy with the service you provided. Thank you for your help. Class label: Unsatisfied Dialogue: - Agent: Sorry that we will cancel your order. You will get a refund within 7 business days. - Customer: oh that takes too long. I want you to take quicker action on this.
์ž‘์—…: ์ฃผ์–ด์ง„ ๋Œ€ํ™”๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์‚ฌ์šฉ์ž๊ฐ€ ์„œ๋น„์Šค์— ๋งŒ์กฑํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๋ถ„๋ฅ˜ํ•˜์„ธ์š”. "Satisfied" ๋˜๋Š” "Unsatisfied"๋กœ ์‘๋‹ตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”: Satisfied ๋Œ€ํ™”: - ์ƒ๋‹ด์›: ํ”ผ๋“œ๋ฐฑ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์•ž์œผ๋กœ ๋” ๋‚˜์€ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ๋…ธ๋ ฅํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. - ๊ณ ๊ฐ: ์ œ๊ณตํ•ด์ฃผ์‹  ์„œ๋น„์Šค์— ๋งŒ์กฑํ•ฉ๋‹ˆ๋‹ค. ๋„์™€์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”: Unsatisfied ๋Œ€ํ™”: - ์ƒ๋‹ด์›: ์ฃผ๋ฌธ์„ ์ทจ์†Œํ•ด๋“œ๋ ค ์ฃ„์†กํ•ฉ๋‹ˆ๋‹ค. ํ™˜๋ถˆ์€ ์˜์—…์ผ ๊ธฐ์ค€ 7์ผ ์ด๋‚ด์— ์ฒ˜๋ฆฌ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. - ๊ณ ๊ฐ: ์•„, ๋„ˆ๋ฌด ์˜ค๋ž˜ ๊ฑธ๋ฆฌ๋„ค์š”. ๋” ๋น ๋ฅด๊ฒŒ ์ฒ˜๋ฆฌํ•ด ์ฃผ์„ธ์š”.
๊ณผ์ œ: ๋Œ€ํ™”๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ์‚ฌ์šฉ์ž๊ฐ€ ์„œ๋น„์Šค์— ๋งŒ์กฑํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. "๋งŒ์กฑ" ๋˜๋Š” "๋ถˆ๋งŒ์กฑ"์œผ๋กœ ์‘๋‹ตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”: ๋งŒ์กฑ ๋Œ€ํ™”: - ์ƒ๋‹ด์›: ํ”ผ๋“œ๋ฐฑ์„ ๋ณด๋‚ด์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์•ž์œผ๋กœ ์„œ๋น„์Šค๋ฅผ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋…ธ๋ ฅํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. - ๊ณ ๊ฐ๋‹˜: ์ œ๊ณตํ•ด์ฃผ์‹  ์„œ๋น„์Šค์— ๋งŒ์กฑํ•ฉ๋‹ˆ๋‹ค. ๋„์™€์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”: ๋ถˆ๋งŒ์กฑ ๋Œ€ํ™”: - ์ƒ๋‹ด์›: ์ฃผ๋ฌธ์„ ์ทจ์†Œํ•ด๋“œ๋ ค ์ฃ„์†กํ•ฉ๋‹ˆ๋‹ค. ์˜์—…์ผ ๊ธฐ์ค€ 7์ผ ์ด๋‚ด์— ํ™˜๋ถˆ ์ฒ˜๋ฆฌํ•ด ๋“œ๋ฆฝ๋‹ˆ๋‹ค. - ๊ณ ๊ฐ๋‹˜: ๋„ˆ๋ฌด ์˜ค๋ž˜ ๊ฑธ๋ฆฌ๋„ค์š”. ๋” ๋นจ๋ฆฌ ์ฒ˜๋ฆฌํ•ด ์ฃผ์…จ์œผ๋ฉด ์ข‹๊ฒ ์–ด์š”.
The taxonomy is incompleteThe v0.5 benchmark covers only seven hazard categories. Six other hazard categories were identified in the taxonomy but not included due to feasibility constraints. Further, hazards intersect and it can be hard to separate them; and although we elaborated numerous subcategories and sub subcategories in the taxonomy, we have not covered every hazard. Notably, we have not tested for LM security issues, such as preserving the confidentiality, privacy, integrity, authenticity, and availability of models or data.
์ด ๋ถ„๋ฅ˜ ์ฒด๊ณ„๋Š” ๋ถˆ์™„์ „ํ•ฉ๋‹ˆ๋‹ค. v0.5 ๋ฒค์น˜๋งˆํฌ๋Š” ๋‹จ์ง€ ์ผ๊ณฑ ๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ๋งŒ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ๋ถ„๋ฅ˜ ์ฒด๊ณ„์—์„œ ์—ฌ์„ฏ ๊ฐ€์ง€ ๋‹ค๋ฅธ ์œ„ํ—˜ ๋ฒ”์ฃผ๊ฐ€ ์‹๋ณ„๋˜์—ˆ์ง€๋งŒ, ์‹คํ–‰ ๊ฐ€๋Šฅ์„ฑ์˜ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ํฌํ•จ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์œ„ํ—˜ ์š”์†Œ๋“ค์ด ๊ต์ฐจํ•  ์ˆ˜ ์žˆ์–ด ์ด๋ฅผ ๋ถ„๋ฆฌํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋น„๋ก ์šฐ๋ฆฌ๊ฐ€ ๋ถ„๋ฅ˜ ์ฒด๊ณ„์—์„œ ์ˆ˜๋งŽ์€ ํ•˜์œ„ ๋ฒ”์ฃผ์™€ ํ•˜์œ„ ํ•˜์œ„ ๋ฒ”์ฃผ๋ฅผ ์ƒ์„ธํžˆ ์„ค๋ช…ํ–ˆ์ง€๋งŒ, ๋ชจ๋“  ์œ„ํ—˜์„ ๋‹ค๋ฃจ์ง€๋Š” ์•Š์•˜์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ๋ชจ๋ธ์ด๋‚˜ ๋ฐ์ดํ„ฐ์˜ ๊ธฐ๋ฐ€์„ฑ, ํ”„๋ผ์ด๋ฒ„์‹œ, ๋ฌด๊ฒฐ์„ฑ, ์ง„์œ„์„ฑ, ๊ฐ€์šฉ์„ฑ์„ ์œ ์ง€ํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™์€ LM ๋ณด์•ˆ ๋ฌธ์ œ์— ๋Œ€ํ•ด์„œ๋Š” ํ…Œ์ŠคํŠธํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
๋ถ„๋ฅ˜๊ฐ€ ๋ถˆ์™„์ „ํ•ฉ๋‹ˆ๋‹คv0.5 ๋ฒค์น˜๋งˆํฌ๋Š” 7๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ๋งŒ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋ถ„๋ฅ˜๋ฒ•์—๋Š” 6๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ๊ฐ€ ๋” ์žˆ์ง€๋งŒ ์‹คํ˜„ ๊ฐ€๋Šฅ์„ฑ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ํฌํ•จ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์œ„ํ—˜์€ ์„œ๋กœ ๊ต์ฐจํ•˜๋ฉฐ ๋ถ„๋ฆฌํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋ถ„๋ฅ˜๋ฒ•์—์„œ ์ˆ˜๋งŽ์€ ํ•˜์œ„ ๋ฒ”์ฃผ์™€ ํ•˜์œ„ ํ•˜์œ„ ๋ฒ”์ฃผ๋ฅผ ์ •๊ตํ™”ํ–ˆ์ง€๋งŒ ๋ชจ๋“  ์œ„ํ—˜์„ ๋‹ค๋ฃจ์ง€๋Š” ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ๋ชจ๋ธ ๋˜๋Š” ๋ฐ์ดํ„ฐ์˜ ๊ธฐ๋ฐ€์„ฑ, ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ, ๋ฌด๊ฒฐ์„ฑ, ์ง„์œ„์„ฑ, ๊ฐ€์šฉ์„ฑ ๋ณด์กด๊ณผ ๊ฐ™์€ LM ๋ณด์•ˆ ๋ฌธ์ œ์— ๋Œ€ํ•ด์„œ๋Š” ํ…Œ์ŠคํŠธํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
To answer our questions with prompts of diverse fact composition types, we construct TwoHopFact using well-known named entities in Wikidata [25] and manually selected relations (Appendix A). TwoHopFact consists of 45,595 unique pairs of one-hop and two-hop prompts of 52 fact composition types constructed from the same number of fact triplet pairs \(((e_{1},r_{1},e_{2}),(e_{2},r_{2},e_{3}))\) as in Table 1. Appendix Table 3 shows example two-hop prompts for each fact composition type, and Appendix B provides detailed data statistics.
๋‹ค์–‘ํ•œ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์˜ ํ”„๋กฌํ”„ํŠธ๋กœ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” Wikidata [25]์˜ ์ž˜ ์•Œ๋ ค์ง„ ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ์™€ ์ˆ˜๋™์œผ๋กœ ์„ ํƒํ•œ ๊ด€๊ณ„(Appendix A)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ TwoHopFact๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. TwoHopFact๋Š” 52๊ฐ€์ง€ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์˜ 45,595๊ฐœ์˜ ๊ณ ์œ ํ•œ ํ•œ ํ™‰ ๋ฐ ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” Table 1์— ๋‚˜์™€ ์žˆ๋Š” ๊ฒƒ๊ณผ ๋™์ผํ•œ ์ˆ˜์˜ ์‚ฌ์‹ค ์‚ผ์ค‘ํ•ญ ์Œ \(((e_{1},r_{1},e_{2}),(e_{2},r_{2},e_{3}))\)์—์„œ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. Appendix Table 3์€ ๊ฐ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์— ๋Œ€ํ•œ ์˜ˆ์‹œ ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด์—ฌ์ฃผ๋ฉฐ, Appendix B๋Š” ์ž์„ธํ•œ ๋ฐ์ดํ„ฐ ํ†ต๊ณ„๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๋‹ค์–‘ํ•œ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์˜ ํ”„๋กฌํ”„ํŠธ๋กœ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์œ„ํ‚ค๋ฐ์ดํ„ฐ [25]์˜ ์ž˜ ์•Œ๋ ค์ง„ ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ์™€ ์ˆ˜๋™์œผ๋กœ ์„ ํƒํ•œ ๊ด€๊ณ„(๋ถ€๋ก A)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ TwoHopFact๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. TwoHopFact๋Š” ํ‘œ 1๊ณผ ๋™์ผํ•œ ์ˆ˜์˜ ์‚ฌ์‹ค ์‚ผ์ค‘ ํ•ญ ์Œ \(((e_{1},r_{1},e_{2}),(e_{2},r_{2},e_{3}))\)์œผ๋กœ ๊ตฌ์„ฑ๋œ 52๊ฐœ์˜ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์˜ ํ•œ ํ™‰๊ณผ ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ์˜ ๊ณ ์œ ํ•œ ์Œ 45,595๊ฐœ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋ถ€๋ก ํ‘œ 3์€ ๊ฐ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์— ๋Œ€ํ•œ ์˜ˆ์‹œ ๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด์—ฌ์ฃผ๋ฉฐ, ๋ถ€๋ก B๋Š” ์ž์„ธํ•œ ๋ฐ์ดํ„ฐ ํ†ต๊ณ„๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
The longer the context length, the more tokens we can squeeze into the context. The more information the model has access to, the better its response will be. But on the other hand, with very long context, it would be hard for the model to remember everything and efficiently process all the information. Attention-based models are highly inefficient for longer contexts and that is why we should expect more research in different mechanisms that enable processing longer contexts and generally come up with more efficient architectures.
๋ฌธ๋งฅ ๊ธธ์ด๊ฐ€ ๊ธธ์ˆ˜๋ก ๋” ๋งŽ์€ ํ† ํฐ์„ ๋ฌธ๋งฅ์— ๋„ฃ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๋Š” ์ •๋ณด๊ฐ€ ๋งŽ์„์ˆ˜๋ก ์‘๋‹ต์ด ๋” ์ข‹์•„์ง‘๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋ฐ˜๋Œ€๋กœ, ๋ฌธ๋งฅ์ด ๋„ˆ๋ฌด ๊ธธ๋ฉด ๋ชจ๋ธ์ด ๋ชจ๋“  ๊ฒƒ์„ ๊ธฐ์–ตํ•˜๊ณ  ํšจ์œจ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค. Attention ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์€ ๊ธด ๋ฌธ๋งฅ์— ๋Œ€ํ•ด ๋งค์šฐ ๋น„ํšจ์œจ์ ์ด๋ฉฐ, ๊ทธ๋ ‡๊ธฐ ๋•Œ๋ฌธ์— ๋” ๊ธด ๋ฌธ๋งฅ์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ค์–‘ํ•œ ๋ฉ”์ปค๋‹ˆ์ฆ˜๊ณผ ๋” ํšจ์œจ์ ์ธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•œ ์—ฐ๊ตฌ๊ฐ€ ๋” ๋งŽ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
์ปจํ…์ŠคํŠธ ๊ธธ์ด๊ฐ€ ๊ธธ์ˆ˜๋ก ์ปจํ…์ŠคํŠธ์— ๋” ๋งŽ์€ ํ† ํฐ์„ ์••์ถ•ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ๋” ๋งŽ์€ ์ •๋ณด์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ์„์ˆ˜๋ก ์‘๋‹ต์ด ๋” ์ข‹์•„์ง‘๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋ฐ˜๋ฉด์— ์ปจํ…์ŠคํŠธ๊ฐ€ ๋„ˆ๋ฌด ๊ธธ๋ฉด ๋ชจ๋ธ์ด ๋ชจ๋“  ๊ฒƒ์„ ๊ธฐ์–ตํ•˜๊ณ  ๋ชจ๋“  ์ •๋ณด๋ฅผ ํšจ์œจ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์˜ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์€ ๊ธด ์ปจํ…์ŠคํŠธ์— ๋Œ€ํ•ด์„œ๋Š” ๋งค์šฐ ๋น„ํšจ์œจ์ ์ด๊ธฐ ๋•Œ๋ฌธ์— ๋” ๊ธด ์ปจํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ค์–‘ํ•œ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์—ฐ๊ตฌํ•˜๊ณ  ์ผ๋ฐ˜์ ์œผ๋กœ ๋” ํšจ์œจ์ ์ธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ฐœ๋ฐœํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
As illustrated in Figure 15, the generation process of CoT follows a chain structure in the basic CoT prompting approach, where LLMs generate CoTs step by step. Typically, CoT takes the format of natural language text. However, textual CoTs may not work well on complex tasks that require rigorous logic for reasoning. Considering this, some work uses code [506, 507] due to its structured and precise nature. Furthermore, the authors in [508] propose to dynamically select text or code as the format of CoTs to combine their advantages.
Figure 15์—์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, CoT์˜ ์ƒ์„ฑ ๊ณผ์ •์€ ๊ธฐ๋ณธ CoT ํ”„๋กฌํ”„ํŠธ ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ ์ฒด์ธ ๊ตฌ์กฐ๋ฅผ ๋”ฐ๋ฅด๋ฉฐ, LLM์€ ๋‹จ๊ณ„๋ณ„๋กœ CoT๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ CoT๋Š” ์ž์—ฐ์–ด ํ…์ŠคํŠธ ํ˜•์‹์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ CoT๋Š” ์—„๊ฒฉํ•œ ๋…ผ๋ฆฌ๊ฐ€ ํ•„์š”ํ•œ ๋ณต์žกํ•œ ์ž‘์—…์—์„œ๋Š” ์ž˜ ์ž‘๋™ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๊ณ ๋ คํ•˜์—ฌ, ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ตฌ์กฐ์ ์ด๊ณ  ์ •ํ™•ํ•œ ํŠน์„ฑ ๋•Œ๋ฌธ์— ์ฝ”๋“œ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค [506, 507]. ๋˜ํ•œ, [508]์˜ ์ €์ž๋“ค์€ CoT์˜ ํ˜•์‹์œผ๋กœ ํ…์ŠคํŠธ๋‚˜ ์ฝ”๋“œ๋ฅผ ๋™์ ์œผ๋กœ ์„ ํƒํ•˜์—ฌ ์ด๋“ค์˜ ์žฅ์ ์„ ๊ฒฐํ•ฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 15์— ์˜ˆ์‹œ๋œ ๋ฐ”์™€ ๊ฐ™์ด, CoT์˜ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋Š” ๊ธฐ๋ณธ CoT ํ”„๋กฌํ”„ํŠธ ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ LLMs๊ฐ€ ๋‹จ๊ณ„๋ณ„๋กœ CoT๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ฒด์ธ ๊ตฌ์กฐ๋ฅผ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ CoT๋Š” ์ž์—ฐ์–ด ํ…์ŠคํŠธ ํ˜•์‹์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ…์ŠคํŠธ CoT๋Š” ์ถ”๋ก ์„ ์œ„ํ•ด ์—„๊ฒฉํ•œ ๋…ผ๋ฆฌ๊ฐ€ ํ•„์š”ํ•œ ๋ณต์žกํ•œ ์ž‘์—…์—์„œ๋Š” ์ž˜ ์ž‘๋™ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๊ณ ๋ คํ•˜์—ฌ ์ผ๋ถ€ ์ž‘์—…์—์„œ๋Š” ๊ตฌ์กฐํ™”๋˜๊ณ  ์ •ํ™•ํ•œ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ์ฝ”๋“œ [506, 507]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ [508]์˜ ์ €์ž๋“ค์€ ํ…์ŠคํŠธ ๋˜๋Š” ์ฝ”๋“œ๋ฅผ CoT์˜ ํ˜•์‹์œผ๋กœ ๋™์ ์œผ๋กœ ์„ ํƒํ•˜์—ฌ ์žฅ์ ์„ ๊ฒฐํ•ฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
Beyond team structure, the paper emphasizes the importance of employing a "criticize-reflect" step for generating plans, evaluating performance, providing feedback, and re-organizing the team [9]. Their results indicate that agents with a dynamic team structure with rotating leadership provide the best results, with both the lowest time to task completion and the lowest communication cost on average. Ultimately, leadership and dynamic team structures improve the overall team's ability to reason, plan, and perform tasks effectively.
ํŒ€ ๊ตฌ์กฐ๋ฅผ ๋„˜์–ด, ์ด ๋…ผ๋ฌธ์€ ๊ณ„ํš ์ˆ˜๋ฆฝ, ์„ฑ๊ณผ ํ‰๊ฐ€, ํ”ผ๋“œ๋ฐฑ ์ œ๊ณต ๋ฐ ํŒ€ ์žฌ๊ตฌ์„ฑ์„ ์œ„ํ•ด "๋น„ํŒ-๋ฐ˜์„ฑ" ๋‹จ๊ณ„๋ฅผ ์ ์šฉํ•˜๋Š” ๊ฒƒ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค [9]. ๊ทธ๋“ค์˜ ๊ฒฐ๊ณผ๋Š” ๋ฆฌ๋”์‹ญ์ด ์ˆœํ™˜ํ•˜๋Š” ๋™์  ํŒ€ ๊ตฌ์กฐ๋ฅผ ๊ฐ€์ง„ ์—์ด์ „ํŠธ๊ฐ€ ๊ฐ€์žฅ ์ข‹์€ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ํ‰๊ท ์ ์œผ๋กœ ์ž‘์—… ์™„๋ฃŒ ์‹œ๊ฐ„๊ณผ ํ†ต์‹  ๋น„์šฉ์ด ๊ฐ€์žฅ ๋‚ฎ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ถ๊ทน์ ์œผ๋กœ, ๋ฆฌ๋”์‹ญ๊ณผ ๋™์  ํŒ€ ๊ตฌ์กฐ๋Š” ์ „์ฒด ํŒ€์˜ ์ถ”๋ก , ๊ณ„ํš ๋ฐ ์ž‘์—… ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์—์„œ๋Š” ํŒ€ ๊ตฌ์กฐ๋ฅผ ๋„˜์–ด ๊ณ„ํš ์ˆ˜๋ฆฝ, ์„ฑ๊ณผ ํ‰๊ฐ€, ํ”ผ๋“œ๋ฐฑ ์ œ๊ณต, ํŒ€ ์žฌ๊ตฌ์„ฑ์„ ์œ„ํ•ด '๋น„ํŒ-๋ฐ˜์˜' ๋‹จ๊ณ„๋ฅผ ์ ์šฉํ•˜๋Š” ๊ฒƒ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค[9]. ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด ๋ฆฌ๋”์‹ญ์„ ์ˆœํ™˜ํ•˜๋Š” ๋™์  ํŒ€ ๊ตฌ์กฐ๋ฅผ ๊ฐ€์ง„ ์—์ด์ „ํŠธ๊ฐ€ ํ‰๊ท ์ ์œผ๋กœ ์ž‘์—… ์™„๋ฃŒ ์‹œ๊ฐ„๊ณผ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ๋น„์šฉ์ด ๊ฐ€์žฅ ๋‚ฎ์€ ์ตœ์ƒ์˜ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ถ๊ทน์ ์œผ๋กœ ๋ฆฌ๋”์‹ญ๊ณผ ๋™์  ํŒ€ ๊ตฌ์กฐ๋Š” ํŒ€์˜ ์ „๋ฐ˜์ ์ธ ์ถ”๋ก  ๋Šฅ๋ ฅ, ๊ณ„ํš ์ˆ˜๋ฆฝ ๋Šฅ๋ ฅ, ์—…๋ฌด ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
Submitted for Meta internal review on March 14, 2024. We would like to thank Lin Xiao and Yuchen Zhang for many helpful conversations. We would like to extend special thanks to Ian Clark, Gourab De, Anmol Mann, and Max Pfeifer from W&B, as well as Lucca Bertoncini, Liao Hu, Caleb Ho, Apostolos Kokolis, and Shubho Sengupta from Meta FAIR NextSys; Henry Estela, Wil Johnson, Rizwan Hashmi, and Lucas Noah from Meta Cloud Foundation; without their invaluable support, the extensive experiments in this paper would not have been possible.
2024๋…„ 3์›” 14์ผ Meta ๋‚ด๋ถ€ ๊ฒ€ํ† ๋ฅผ ์œ„ํ•ด ์ œ์ถœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ์œ ์ตํ•œ ๋Œ€ํ™”๋ฅผ ๋‚˜๋ˆˆ Lin Xiao์™€ Yuchen Zhang์—๊ฒŒ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. W&B์˜ Ian Clark, Gourab De, Anmol Mann, Max Pfeifer์™€ Meta FAIR NextSys์˜ Lucca Bertoncini, Liao Hu, Caleb Ho, Apostolos Kokolis, Shubho Sengupta์—๊ฒŒ ํŠน๋ณ„ํžˆ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Meta Cloud Foundation์˜ Henry Estela, Wil Johnson, Rizwan Hashmi, Lucas Noah์—๊ฒŒ๋„ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์ด๋“ค์˜ ๊ท€์ค‘ํ•œ ์ง€์› ์—†์ด๋Š” ์ด ๋…ผ๋ฌธ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์ด ๋ถˆ๊ฐ€๋Šฅํ–ˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
2024๋…„ 3์›” 14์ผ์— ๋ฉ”ํƒ€ ๋‚ด๋ถ€ ๊ฒ€ํ† ๋ฅผ ์œ„ํ•ด ์ œ์ถœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ๋„์›€์ด ๋˜๋Š” ๋Œ€ํ™”๋ฅผ ๋‚˜๋ˆˆ Lin Xiao์™€ Yuchen Zhang์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. W&B์˜ Ian Clark, Gourab De, Anmol Mann, Max Pfeifer์™€ Meta FAIR NextSys์˜ Lucca Bertoncini, Liao Hu, Caleb Ho, Apostolos Kokolis, Shubho Sengupta๋ฅผ ํŠน๋ณ„ํžˆ ๊ฐ์‚ฌ๋“œ๋ฆฌ๋ฉฐ, ์ด๋“ค์˜ ๊ท€์ค‘ํ•œ ์ง€์›์ด ์—†์—ˆ๋‹ค๋ฉด ์ด ๋…ผ๋ฌธ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์ด ๋ถˆ๊ฐ€๋Šฅํ–ˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
On dynamic data curating.Our study presents a pipeline to decide on a group of fixed mixture proportions for pretraining. More sophisticated data curating can include dynamic proportions (Albalak et al., 2023) and even a curriculum that changes data domains (Chen et al., 2024). The application of our data mixing laws in continual pretraining (Sec. 5) implies the prospect of extending our findings to these settings. On top of this, we believe that it is promising to incorporate further analysis to pursue a dynamic data mixing law.
๋™์  ๋ฐ์ดํ„ฐ ํ๋ ˆ์ดํŒ…์— ๋Œ€ํ•˜์—ฌ. ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์œ„ํ•œ ๊ณ ์ •๋œ ํ˜ผํ•ฉ ๋น„์œจ ๊ทธ๋ฃน์„ ๊ฒฐ์ •ํ•˜๋Š” ํŒŒ์ดํ”„๋ผ์ธ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋” ์ •๊ตํ•œ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ดํŒ…์€ ๋™์  ๋น„์œจ(Albalak et al., 2023) ๋ฐ ๋ฐ์ดํ„ฐ ๋„๋ฉ”์ธ์„ ๋ณ€๊ฒฝํ•˜๋Š” ์ปค๋ฆฌํ˜๋Ÿผ(Chen et al., 2024)์„ ํฌํ•จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ(Sec. 5)์—์„œ ์šฐ๋ฆฌ์˜ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์„ ์ ์šฉํ•˜๋Š” ๊ฒƒ์€ ์ด๋Ÿฌํ•œ ์„ค์ •์œผ๋กœ ์šฐ๋ฆฌ์˜ ๋ฐœ๊ฒฌ์„ ํ™•์žฅํ•  ๊ฐ€๋Šฅ์„ฑ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด์™€ ๋”๋ถˆ์–ด, ๋™์  ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์„ ์ถ”๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ๋ถ„์„์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ด ์œ ๋งํ•˜๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค.
๋™์  ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜์— ๊ด€ํ•ด์„œ, ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•ด ๊ณ ์ •๋œ ํ˜ผํ•ฉ ๋น„์œจ ๊ทธ๋ฃน์„ ๊ฒฐ์ •ํ•˜๋Š” ํŒŒ์ดํ”„๋ผ์ธ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ณด๋‹ค ์ •๊ตํ•œ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜์—๋Š” ๋™์  ๋น„์œจ(Albalak et al., 2023)๊ณผ ๋ฐ์ดํ„ฐ ๋„๋ฉ”์ธ์„ ๋ณ€๊ฒฝํ•˜๋Š” ์ปค๋ฆฌํ˜๋Ÿผ(Chen et al., 2024)์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต(Sec. 5)์—์„œ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์„ ์ ์šฉํ•˜๋ฉด ์ด๋Ÿฌํ•œ ์„ค์ •์œผ๋กœ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋ฅผ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋Š” ์ „๋ง์ด ์—ด๋ฆฝ๋‹ˆ๋‹ค. ์ด ์™ธ์—๋„ ๋™์  ๋ฐ์ดํ„ฐ ๋ฏน์‹ฑ ๋ฒ•์น™์„ ์ถ”๊ตฌํ•˜๊ธฐ ์œ„ํ•œ ์ถ”๊ฐ€ ๋ถ„์„์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ด ์œ ๋งํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
**Results.** We observed a higher average diagnostic performance for AMIE in Canada than in India (see Figure A.13). However, when comparing scenarios performed at both study locations, we observed that AMIE and PCP performance remained consistent regardless of the study location (see Figure A.14), suggesting that the observed performance variations are likely not due to the patient actor or clinician rater differences, but instead might be attributed to inherent differences in the difficulty levels of scenarios in each location.
**๊ฒฐ๊ณผ.** ์บ๋‚˜๋‹ค์—์„œ AMIE์˜ ํ‰๊ท  ์ง„๋‹จ ์„ฑ๋Šฅ์ด ์ธ๋„๋ณด๋‹ค ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ A.13 ์ฐธ์กฐ). ๊ทธ๋Ÿฌ๋‚˜ ๋‘ ์—ฐ๊ตฌ ์žฅ์†Œ์—์„œ ์ˆ˜ํ–‰๋œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋น„๊ตํ–ˆ์„ ๋•Œ, AMIE์™€ PCP์˜ ์„ฑ๋Šฅ์€ ์—ฐ๊ตฌ ์žฅ์†Œ์— ๊ด€๊ณ„์—†์ด ์ผ๊ด€๋˜๊ฒŒ ์œ ์ง€๋œ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ A.14 ์ฐธ์กฐ). ์ด๋Š” ๊ด€์ฐฐ๋œ ์„ฑ๋Šฅ ์ฐจ์ด๊ฐ€ ํ™˜์ž ๋ฐฐ์šฐ๋‚˜ ์ž„์ƒ์˜ ํ‰๊ฐ€์ž ์ฐจ์ด ๋•Œ๋ฌธ์ด ์•„๋‹ˆ๋ผ, ๊ฐ ์žฅ์†Œ์˜ ์‹œ๋‚˜๋ฆฌ์˜ค ๋‚œ์ด๋„์— ๋‚ด์žฌ๋œ ์ฐจ์ด ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
**๊ฒฐ๊ณผ.** ์บ๋‚˜๋‹ค์—์„œ ์ธ๋„๋ณด๋‹ค AMIE์˜ ํ‰๊ท  ์ง„๋‹จ ์„ฑ๋Šฅ์ด ๋” ๋†’์€ ๊ฒƒ์œผ๋กœ ๊ด€์ฐฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ A.13 ์ฐธ์กฐ). ๊ทธ๋Ÿฌ๋‚˜ ๋‘ ์—ฐ๊ตฌ ๊ธฐ๊ด€์—์„œ ์ˆ˜ํ–‰ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋น„๊ตํ–ˆ์„ ๋•Œ, AMIE์™€ PCP ์„ฑ๋Šฅ์ด ์—ฐ๊ตฌ ๊ธฐ๊ด€์— ๊ด€๊ณ„์—†์ด ์ผ๊ด€๋˜๊ฒŒ ์œ ์ง€๋˜๋Š” ๊ฒƒ์œผ๋กœ ๊ด€์ฐฐ๋˜์–ด ๊ด€์ฐฐ๋œ ์„ฑ๋Šฅ ์ฐจ์ด๋Š” ํ™˜์ž ํ–‰์œ„์ž ๋˜๋Š” ์ž„์ƒ์˜ ํ‰๊ฐ€์ž์˜ ์ฐจ์ด ๋•Œ๋ฌธ์ด ์•„๋‹ˆ๋ผ ๊ฐ ๊ธฐ๊ด€์˜ ์‹œ๋‚˜๋ฆฌ์˜ค ๋‚œ์ด๋„์˜ ๋‚ด์žฌ์  ์ฐจ์ด์— ๊ธฐ์ธํ•œ ๊ฒƒ์ผ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค.
To evaluate financial domain tasks, we compare FinPythia and Fynthia-1B exhibit superior performance on tasks FPB, Headline, and NER while showing comparatively lower results on the FiQA SA task compared with Pythia counterparts. DACP boosts the average task performance by 2.8% for the 1B model and 8.3% for the 6.9B model. These outcomes directly substantiate the impact of domain-adaptive continual pre-training on enhancing in-domain task performance. Furthermore, Pythia-6.9B outperforms OPT-7B, BLOOM-7B, and GPT-J-GB on average.
๊ธˆ์œต ๋„๋ฉ”์ธ ์ž‘์—…์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” FinPythia์™€ Fynthia-1B๊ฐ€ FPB, Headline, NER ์ž‘์—…์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๋ฐ˜๋ฉด, FiQA SA ์ž‘์—…์—์„œ๋Š” Pythia์™€ ๋น„๊ตํ•˜์—ฌ ์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ์€ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์ž„์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. DACP๋Š” 1B ๋ชจ๋ธ์˜ ํ‰๊ท  ์ž‘์—… ์„ฑ๋Šฅ์„ 2.8%, 6.9B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 8.3% ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๋„๋ฉ”์ธ ์ ์‘ ์ง€์† ์‚ฌ์ „ ํ•™์Šต์ด ๋„๋ฉ”์ธ ๋‚ด ์ž‘์—… ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์ง์ ‘์ ์œผ๋กœ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, Pythia-6.9B๋Š” ํ‰๊ท ์ ์œผ๋กœ OPT-7B, BLOOM-7B, GPT-J-GB๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
์žฌ๋ฌด ๋„๋ฉ”์ธ ์ž‘์—…์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด FinPythia์™€ Fynthia-1B๊ฐ€ FPB, ํ—ค๋“œ๋ผ์ธ ๋ฐ NER ์ž‘์—…์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๋ฐ˜๋ฉด, FiQA SA ์ž‘์—…์—์„œ๋Š” Pythia์— ๋น„ํ•ด ์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ์€ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. DACP๋Š” 1B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ‰๊ท  ์ž‘์—… ์„ฑ๋Šฅ์„ 2.8%, 6.9B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 8.3% ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๋„๋ฉ”์ธ ์ ์‘ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์ด ๋„๋ฉ”์ธ ๋‚ด ์ž‘์—… ์„ฑ๋Šฅ ํ–ฅ์ƒ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์ง์ ‘์ ์œผ๋กœ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Pythia-6.9B๋Š” OPT-7B, BLOOM-7B ๋ฐ GPT-J-GB๋ณด๋‹ค ํ‰๊ท ์ ์œผ๋กœ ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค.
Overall, these downstream results validate that training a language model to predict the subtext between the lines of general text data can substantially improve its reasoning capabilities, even on datasets it was not explicitly trained on. The fact that longer rationales consistently lead to better outcomes, and that Quiet-STaR outperforms the constrained pause token approach, supports the notion that Quiet-STaR is successfully teaching the model to leverage its own generated thoughts to reason more thoroughly about the input.
์ „๋ฐ˜์ ์œผ๋กœ, ์ด๋Ÿฌํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๊ฒฐ๊ณผ๋Š” ์ผ๋ฐ˜ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์˜ ํ–‰๊ฐ„(subtext)์„ ์˜ˆ์ธกํ•˜๋„๋ก ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๊ฒƒ์ด, ๋ช…์‹œ์ ์œผ๋กœ ํ›ˆ๋ จ๋˜์ง€ ์•Š์€ ๋ฐ์ดํ„ฐ์…‹์—์„œ๋„ ๋ชจ๋ธ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ์ƒ๋‹นํžˆ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ๋” ๊ธด ์ถ”๋ก  ๊ณผ์ •์ด ์ผ๊ด€๋˜๊ฒŒ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์ด๋Œ์–ด๋‚ด๊ณ , Quiet-STaR๊ฐ€ ์ œํ•œ๋œ pause token ์ ‘๊ทผ๋ฒ•๋ณด๋‹ค ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ์‚ฌ์‹ค์€ Quiet-STaR๊ฐ€ ๋ชจ๋ธ์—๊ฒŒ ์ž…๋ ฅ์— ๋Œ€ํ•ด ๋” ์ฒ ์ €ํ•˜๊ฒŒ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ž์ฒด ์ƒ์„ฑ๋œ ์ƒ๊ฐ์„ ํ™œ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์„ฑ๊ณต์ ์œผ๋กœ ๊ฐ€๋ฅด์น˜๊ณ  ์žˆ์Œ์„ ๋’ท๋ฐ›์นจํ•ฉ๋‹ˆ๋‹ค.
์ „๋ฐ˜์ ์œผ๋กœ ์ด๋Ÿฌํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๊ฒฐ๊ณผ๋Š” ์ผ๋ฐ˜ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์˜ ์ค„ ์‚ฌ์ด์˜ ์•”๋ฌต์ ์ธ ์˜๋ฏธ๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋ฉด ๋ช…์‹œ์ ์œผ๋กœ ํ›ˆ๋ จ๋˜์ง€ ์•Š์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋„ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ๋” ๊ธด ๊ทผ๊ฑฐ๊ฐ€ ์ผ๊ด€๋˜๊ฒŒ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ค๊ณ , Quiet-STaR์ด ์ œํ•œ๋œ ์ผ์‹œ ์ค‘์ง€ ํ† ํฐ ์ ‘๊ทผ ๋ฐฉ์‹๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ์‚ฌ์‹ค์€ Quiet-STaR์ด ๋ชจ๋ธ์ด ์ž์ฒด์ ์œผ๋กœ ์ƒ์„ฑํ•œ ์ƒ๊ฐ์„ ํ™œ์šฉํ•˜์—ฌ ์ž…๋ ฅ์— ๋Œ€ํ•ด ๋” ์ฒ ์ €ํ•˜๊ฒŒ ์ถ”๋ก ํ•˜๋„๋ก ์„ฑ๊ณต์ ์œผ๋กœ ๊ฐ€๋ฅด์น˜๊ณ  ์žˆ์Œ์„ ๋’ท๋ฐ›์นจํ•ฉ๋‹ˆ๋‹ค.
Combining Real and Synthetic DataOur method of re-phrasing web data naturally incorporates the information diversity found on the internet. However, it does not incorporate the noise in real data. While synthetic data may help LLMs pre-train faster, we also want them to be able to understand noisy web text that may be filled with typos and linguistic errors so that the LLMs do not fail in user facing situations. In order to incorporate this style diversity in language modeling, we sample real and synthetic data in a 1:1 ratio.
์‹ค์ œ ๋ฐ์ดํ„ฐ์™€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฒฐํ•ฉํ•˜๊ธฐ ์šฐ๋ฆฌ์˜ ์›น ๋ฐ์ดํ„ฐ ์žฌ๊ตฌ์„ฑ ๋ฐฉ๋ฒ•์€ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ์ธํ„ฐ๋„ท์—์„œ ๋ฐœ๊ฒฌ๋˜๋Š” ์ •๋ณด์˜ ๋‹ค์–‘์„ฑ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ ๋ฐ์ดํ„ฐ์˜ ๋…ธ์ด์ฆˆ๋Š” ํฌํ•จํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” LLM์ด ๋” ๋น ๋ฅด๊ฒŒ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์ง€๋งŒ, ์šฐ๋ฆฌ๋Š” ๋˜ํ•œ LLM์ด ์‚ฌ์šฉ์ž์™€ ๋Œ€๋ฉดํ•˜๋Š” ์ƒํ™ฉ์—์„œ ์‹คํŒจํ•˜์ง€ ์•Š๋„๋ก ์˜คํƒ€์™€ ์–ธ์–ด์  ์˜ค๋ฅ˜๋กœ ๊ฐ€๋“ ์ฐฌ ๋…ธ์ด์ฆˆ๊ฐ€ ์žˆ๋Š” ์›น ํ…์ŠคํŠธ๋ฅผ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๊ธฐ๋ฅผ ์›ํ•ฉ๋‹ˆ๋‹ค. ์–ธ์–ด ๋ชจ๋ธ๋ง์—์„œ ์ด๋Ÿฌํ•œ ์Šคํƒ€์ผ ๋‹ค์–‘์„ฑ์„ ํฌํ•จํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ์™€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ 1:1 ๋น„์œจ๋กœ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค.
์‹ค์ œ ๋ฐ์ดํ„ฐ์™€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ๊ฒฐํ•ฉ์›น ๋ฐ์ดํ„ฐ๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์ธํ„ฐ๋„ท์—์„œ ๋ฐœ๊ฒฌ๋˜๋Š” ์ •๋ณด ๋‹ค์–‘์„ฑ์„ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ ๋ฐ์ดํ„ฐ์˜ ๋…ธ์ด์ฆˆ๋Š” ํ†ตํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” LLM์ด ๋” ๋น ๋ฅด๊ฒŒ ์‚ฌ์ „ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋„๋ก ๋„์šธ ์ˆ˜ ์žˆ์ง€๋งŒ, ์˜คํƒ€์™€ ์–ธ์–ด์  ์˜ค๋ฅ˜๋กœ ๊ฐ€๋“ ์ฐฌ ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ์€ ์›น ํ…์ŠคํŠธ๋„ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์–ด์•ผ LLM์ด ์‚ฌ์šฉ์ž ๋Œ€๋ฉด ์ƒํ™ฉ์—์„œ ์‹คํŒจํ•˜์ง€ ์•Š๋„๋ก ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์–ธ์–ด ๋ชจ๋ธ๋ง์— ์ด๋Ÿฌํ•œ ์Šคํƒ€์ผ ๋‹ค์–‘์„ฑ์„ ํ†ตํ•ฉํ•˜๊ธฐ ์œ„ํ•ด ์‹ค์ œ ๋ฐ์ดํ„ฐ์™€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ 1:1 ๋น„์œจ๋กœ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค.
Nonetheless, the training of such large scale models is both financially burdensome and operationally inefficient. On one hand, the empirical understanding of the mechanisms underpinning the training of LLMs remains elusive. Given the significant economic and environmental costs, experiments on LLMs are prohibitively expensive for most researchers and corporations. On the other hand, the deployment of these colossal models in everyday scenarios, such as on personal computers or smartphones, is either inefficient or unfeasible.
๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ด๋Ÿฌํ•œ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ์€ ์žฌ์ •์ ์œผ๋กœ ๋ถ€๋‹ด์ด ํฌ๊ณ  ์šด์˜์ ์œผ๋กœ ๋น„ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ํ•œํŽธ์œผ๋กœ, LLM ํ›ˆ๋ จ์˜ ๊ธฐ์ €๋ฅผ ์ด๋ฃจ๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๋Œ€ํ•œ ๊ฒฝํ—˜์  ์ดํ•ด๋Š” ์—ฌ์ „ํžˆ ๋ชจํ˜ธํ•ฉ๋‹ˆ๋‹ค. ์ƒ๋‹นํ•œ ๊ฒฝ์ œ์  ๋ฐ ํ™˜๊ฒฝ์  ๋น„์šฉ์„ ๊ณ ๋ คํ•  ๋•Œ, LLM์— ๋Œ€ํ•œ ์‹คํ—˜์€ ๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ์ž์™€ ๊ธฐ์—…์—๊ฒŒ ์—„์ฒญ๋‚˜๊ฒŒ ๋น„์Œ‰๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ํ•œํŽธ์œผ๋กœ, ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ๋‚˜ ์Šค๋งˆํŠธํฐ๊ณผ ๊ฐ™์€ ์ผ์ƒ์ ์ธ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์ด๋Ÿฌํ•œ ๊ฑฐ๋Œ€ํ•œ ๋ชจ๋ธ์„ ๋ฐฐํฌํ•˜๋Š” ๊ฒƒ์€ ๋น„ํšจ์œจ์ ์ด๊ฑฐ๋‚˜ ์‹คํ–‰ ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ์€ ์žฌ์ •์ ์œผ๋กœ ๋ถ€๋‹ด์Šค๋Ÿฝ๊ณ  ์šด์˜์ ์œผ๋กœ ๋น„ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ํ•œํŽธ์œผ๋กœ, LLM ํ›ˆ๋ จ์˜ ๊ธฐ๋ฐ˜์ด ๋˜๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๋Œ€ํ•œ ๊ฒฝํ—˜์  ์ดํ•ด๋Š” ์—ฌ์ „ํžˆ ๋ชจํ˜ธํ•ฉ๋‹ˆ๋‹ค. ์ƒ๋‹นํ•œ ๊ฒฝ์ œ์ , ํ™˜๊ฒฝ์  ๋น„์šฉ์„ ๊ณ ๋ คํ•  ๋•Œ ๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ์ž์™€ ๊ธฐ์—…์—์„œ๋Š” LLM์— ๋Œ€ํ•œ ์‹คํ—˜์„ ํ•˜๊ธฐ์—๋Š” ๋น„์šฉ์ด ๋„ˆ๋ฌด ๋งŽ์ด ๋“ญ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ๋‚˜ ์Šค๋งˆํŠธํฐ๊ณผ ๊ฐ™์€ ์ผ์ƒ์ ์ธ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์ด๋Ÿฌํ•œ ๊ฑฐ๋Œ€ํ•œ ๋ชจ๋ธ์„ ๋ฐฐํฌํ•˜๋Š” ๊ฒƒ์€ ๋น„ํšจ์œจ์ ์ด๊ฑฐ๋‚˜ ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
Consider, for instance, GPT2, but with its MLP layer (\(d\to 4d\to d\)) replaced by 32 experts, each following a \(d\to d\to d\) configuration. This setup uses \(64d^{2}\) total parameters, but during inference, only \(2d^{2}\) parameters are used per token (e.g., when using \(topk=1\)). After including the Attention layers, which each have \(4d^{2}\) parameters, the ratio between the total and the effective number of parameters for the 32-expert MoE models is approximately \(\frac{4d^{2}+64d^{2}}{4d^{2}+2d^{2}}\approx 11.3\).
์˜ˆ๋ฅผ ๋“ค์–ด, GPT2๋ฅผ ๊ณ ๋ คํ•ด๋ณด์ž. ๊ทธ๋Ÿฌ๋‚˜ MLP ๋ ˆ์ด์–ด(\(d\to 4d\to d\))๋ฅผ 32๋ช…์˜ ์ „๋ฌธ๊ฐ€๋กœ ๋Œ€์ฒดํ•˜๊ณ , ๊ฐ ์ „๋ฌธ๊ฐ€๋Š” \(d\to d\to d\) ๊ตฌ์„ฑ์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ ์ด \(64d^{2}\) ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์ง€๋งŒ, ์ถ”๋ก  ์ค‘์—๋Š” ํ† ํฐ๋‹น \(2d^{2}\) ํŒŒ๋ผ๋ฏธํ„ฐ๋งŒ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค (์˜ˆ: \(topk=1\)์„ ์‚ฌ์šฉํ•  ๋•Œ). Attention ๋ ˆ์ด์–ด๋ฅผ ํฌํ•จํ•˜๋ฉด, ๊ฐ ๋ ˆ์ด์–ด๋Š” \(4d^{2}\) ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง€๋ฏ€๋กœ, 32-์ „๋ฌธ๊ฐ€ MoE ๋ชจ๋ธ์˜ ์ด ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜์™€ ์‹ค์ œ ์‚ฌ์šฉ๋˜๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜์˜ ๋น„์œจ์€ ๋Œ€๋žต \(\frac{4d^{2}+64d^{2}}{4d^{2}+2d^{2}}\approx 11.3\)์ž…๋‹ˆ๋‹ค.
์˜ˆ๋ฅผ ๋“ค์–ด, GPT2๋ฅผ ์ƒ๊ฐํ•ด ๋ณด๊ฒ ์ง€๋งŒ MLP ๋ ˆ์ด์–ด(\(d\to 4d\to d\))๊ฐ€ 32๊ฐœ์˜ ์ „๋ฌธ๊ฐ€๋กœ ๋Œ€์ฒด๋˜์–ด ๊ฐ๊ฐ \(d\to d\to d\) ๊ตฌ์„ฑ์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ ์ด ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ \(64d^{2}\)๋ฅผ ์‚ฌ์šฉํ•˜์ง€๋งŒ, ์ถ”๋ก  ์ค‘์—๋Š” ํ† ํฐ๋‹น \(2d^{2}\) ํŒŒ๋ผ๋ฏธํ„ฐ๋งŒ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค(์˜ˆ: \(topk=1\) ์‚ฌ์šฉ ์‹œ). ๊ฐ๊ฐ \(4d^{2}\) ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ–๋Š” Attention ๋ ˆ์ด์–ด๋ฅผ ํฌํ•จํ•˜๋ฉด 32๊ฐœ ์ „๋ฌธ๊ฐ€ MoE ๋ชจ๋ธ์˜ ์ด ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜์™€ ์œ ํšจ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜ ์‚ฌ์ด์˜ ๋น„์œจ์€ ์•ฝ \(\frac{4d^{2}+64d^{2}}{4d^{2}+2d^{2}}\์•ฝ 11.3\)์ž…๋‹ˆ๋‹ค.
By processing C4 through MDR, we are able to obtain subsets of data which might slightly outperform RefinedWeb; this combines both the stringent filtering of C4 (e.g., strict NSFW word blocklist, 3-sentence span deduplication) with our own filters and deduplication. While such a combination results in rejection rates that would be unacceptable for our target of 3-6 trillions tokens, this represents an interesting perspective for shorter runs, which may be able to extract extremely high-quality subsets from large web datasets.
C4๋ฅผ MDR์„ ํ†ตํ•ด ์ฒ˜๋ฆฌํ•จ์œผ๋กœ์จ, RefinedWeb์„ ์•ฝ๊ฐ„ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” C4์˜ ์—„๊ฒฉํ•œ ํ•„ํ„ฐ๋ง(์˜ˆ: ์—„๊ฒฉํ•œ NSFW ๋‹จ์–ด ์ฐจ๋‹จ ๋ชฉ๋ก, 3๋ฌธ์žฅ ๋ฒ”์œ„ ์ค‘๋ณต ์ œ๊ฑฐ)๊ณผ ์šฐ๋ฆฌ์˜ ์ž์ฒด ํ•„ํ„ฐ ๋ฐ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ๊ฒฐํ•ฉํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์กฐํ•ฉ์€ 3-6์กฐ ํ† ํฐ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ๊ฒฝ์šฐ์—๋Š” ์ˆ˜์šฉํ•  ์ˆ˜ ์—†๋Š” ๊ฑฐ๋ถ€์œจ์„ ์ดˆ๋ž˜ํ•˜์ง€๋งŒ, ๋Œ€๊ทœ๋ชจ ์›น ๋ฐ์ดํ„ฐ์…‹์—์„œ ๋งค์šฐ ๊ณ ํ’ˆ์งˆ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ๋Š” ์งง์€ ์‹คํ–‰์— ๋Œ€ํ•ด์„œ๋Š” ํฅ๋ฏธ๋กœ์šด ๊ด€์ ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
MDR์„ ํ†ตํ•ด C4๋ฅผ ์ฒ˜๋ฆฌํ•˜๋ฉด RefinedWeb๋ณด๋‹ค ์•ฝ๊ฐ„ ๋” ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ ๋ฐ์ดํ„ฐ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์—„๊ฒฉํ•œ ํ•„ํ„ฐ๋ง(์˜ˆ: ์—„๊ฒฉํ•œ NSFW ๋‹จ์–ด ์ฐจ๋‹จ ๋ชฉ๋ก, 3๋ฌธ์žฅ ๋ฒ”์œ„ ์ค‘๋ณต ์ œ๊ฑฐ)๊ณผ ์ž์ฒด ํ•„ํ„ฐ ๋ฐ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ๋ชจ๋‘ ๊ฒฐํ•ฉํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์กฐํ•ฉ์€ ๋ชฉํ‘œ์ธ 3์กฐ~6์กฐ ํ† ํฐ์— ํ—ˆ์šฉํ•  ์ˆ˜ ์—†๋Š” ๊ฑฐ๋ถ€์œจ์„ ์ดˆ๋ž˜ํ•˜์ง€๋งŒ, ๋Œ€๊ทœ๋ชจ ์›น ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋งค์šฐ ๊ณ ํ’ˆ์งˆ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ๋Š” ์งง์€ ์‹คํ–‰์— ๋Œ€ํ•œ ํฅ๋ฏธ๋กœ์šด ๊ด€์ ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
**Method:** Let \(\mathbf{Q}\) be as defined in Algorithm 1. Now rewrite \(\mathbf{Q}=\mathbf{U}\cdot\mathbf{\Sigma}\cdot\mathbf{V}^{\top}\) with SVD. Previously we saw that the number of large enough singular values corresponded to the dimension of the model. But it turns out that the matrix \(\mathbf{U}\) actually directly represents (a rotation of) the final layer! Specifically, we can show that \(\mathbf{U}\cdot\mathbf{\Sigma}=\mathbf{W}\cdot\mathbf{G}\) for some \(h\times h\) matrix \(\mathbf{G}\) in the following lemma.
**Method:** Algorithm 1์—์„œ ์ •์˜๋œ \(\mathbf{Q}\)๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด์ œ \(\mathbf{Q}=\mathbf{U}\cdot\mathbf{\Sigma}\cdot\mathbf{V}^{\top}\)๋ฅผ SVD๋กœ ๋‹ค์‹œ ์ž‘์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด์ „์— ํฐ ํŠน์ด๊ฐ’์˜ ์ˆ˜๊ฐ€ ๋ชจ๋ธ์˜ ์ฐจ์›์— ํ•ด๋‹นํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ๋กœ \(\mathbf{U}\) ํ–‰๋ ฌ์ด ์ตœ์ข… ๋ ˆ์ด์–ด๋ฅผ ์ง์ ‘์ ์œผ๋กœ ๋‚˜ํƒ€๋‚ธ๋‹ค๋Š” ๊ฒƒ์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค! ๊ตฌ์ฒด์ ์œผ๋กœ, ๋‹ค์Œ์˜ ๋ ˆ๋งˆ์—์„œ \(h\times h\) ํ–‰๋ ฌ \(\mathbf{G}\)์— ๋Œ€ํ•ด \(\mathbf{U}\cdot\mathbf{\Sigma}=\mathbf{W}\cdot\mathbf{G}\)์ž„์„ ๋ณด์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**๋ฐฉ๋ฒ•:** ์•Œ๊ณ ๋ฆฌ์ฆ˜ 1์— ์ •์˜๋œ ๋Œ€๋กœ \(\mathbf{Q}\)๋ฅผ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด์ œ SVD๋กœ \(\mathbf{Q}=\mathbf{U}\cdot\mathbf{\Sigma}\cdot\mathbf{V}^{\top}\)๋ฅผ ๋‹ค์‹œ ์ž‘์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด์ „์—๋Š” ์ถฉ๋ถ„ํžˆ ํฐ ํŠน์ด๊ฐ’์˜ ์ˆ˜๊ฐ€ ๋ชจ๋ธ์˜ ์ฐจ์›์— ํ•ด๋‹นํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ \(\mathbf{U}\) ํ–‰๋ ฌ์ด ์‹ค์ œ๋กœ ์ตœ์ข… ๋ ˆ์ด์–ด๋ฅผ ์ง์ ‘ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค! ๊ตฌ์ฒด์ ์œผ๋กœ, ๋‹ค์Œ ์ •๋ฆฌ์—์„œ \(h\times h\) ํ–‰๋ ฌ \(\mathbf{G}\)์— ๋Œ€ํ•ด \(\mathbf{U}\cdot\mathbf{\Sigma}=\mathbf{W}\cdot\mathbf{G}\)์ž„์„ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
To evaluate LLaMA models on this dataset, we directly feed the examples to these models. While when evaluating Alpaca models, we wrap the examples in the prompt template as demonstrated in Section 2.5. Then the model is asked to make a one-step prediction and give the probability distribution of the next token \(p(y|\mathbf{x})\), where \(y\in\mathcal{V}\) (\(\mathcal{V}\) is the vocabulary). To map the probability distribution to a valid label \(t\) in {A, B, C, D}, we extract and gather the probabilities of related tokens.
์ด ๋ฐ์ดํ„ฐ์…‹์—์„œ LLaMA ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์˜ˆ์ œ๋ฅผ ์ง์ ‘ ์ด ๋ชจ๋ธ๋“ค์— ์ž…๋ ฅํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, Alpaca ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•  ๋•Œ๋Š” Section 2.5์—์„œ ์„ค๋ช…ํ•œ ๋Œ€๋กœ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์— ์˜ˆ์ œ๋ฅผ ๊ฐ์Œ‰๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ชจ๋ธ์—๊ฒŒ ํ•œ ๋‹จ๊ณ„ ์˜ˆ์ธก์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ๋‹ค์Œ ํ† ํฐ \(p(y|\mathbf{x})\)์˜ ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ์ œ๊ณตํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ \(y\in\mathcal{V}\) (\(\mathcal{V}\)๋Š” ์–ดํœ˜์ง‘์ž…๋‹ˆ๋‹ค). ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ {A, B, C, D} ๋‚ด์˜ ์œ ํšจํ•œ ๋ ˆ์ด๋ธ” \(t\)๋กœ ๋งคํ•‘ํ•˜๊ธฐ ์œ„ํ•ด ๊ด€๋ จ ํ† ํฐ์˜ ํ™•๋ฅ ์„ ์ถ”์ถœํ•˜๊ณ  ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ LLaMA ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์— ์˜ˆ์ œ๋ฅผ ์ง์ ‘ ๊ณต๊ธ‰ํ•ฉ๋‹ˆ๋‹ค. Alpaca ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•  ๋•Œ๋Š” 2.5์ ˆ์—์„œ ์„ค๋ช…ํ•œ ๋Œ€๋กœ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์— ์˜ˆ์ œ๋ฅผ ๋ž˜ํ•‘ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ชจ๋ธ์— ํ•œ ๋‹จ๊ณ„ ์˜ˆ์ธก์„ ์š”์ฒญํ•˜๊ณ  ๋‹ค์Œ ํ† ํฐ์˜ ํ™•๋ฅ  ๋ถ„ํฌ \(p(y|\mathbf{x})\), ์—ฌ๊ธฐ์„œ \(y\in\mathcal{V}(\mathcal{V}๋Š” ์–ดํœ˜)๋ฅผ ์ œ๊ณตํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ์œ ํšจํ•œ ๋ ˆ์ด๋ธ” \(t\)์— ๋งคํ•‘ํ•˜๊ธฐ ์œ„ํ•ด ๊ด€๋ จ ํ† ํฐ์˜ ํ™•๋ฅ ์„ ์ถ”์ถœํ•˜๊ณ  ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค.
In the rest of this paper, Section II presents an overview of state of the art of LLMs, focusing on three LLM families (GPT, LLaMA and PaLM) and other representative models. Section III discusses how LLMs are built. Section IV discusses how LLMs are used, and augmented for real-world applications Sections V and VI review popular datasets and benchmarks for evaluating LLMs, and summarize the reported LLM evaluation results. Finally, Section VII concludes the paper by summarizing the challenges and future research directions.
์ด ๋…ผ๋ฌธ์˜ ๋‚˜๋จธ์ง€ ๋ถ€๋ถ„์—์„œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋‚ด์šฉ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์„น์…˜ II์—์„œ๋Š” GPT, LLaMA, PaLM์˜ ์„ธ ๊ฐ€์ง€ LLM ๊ณ„์—ด๊ณผ ๊ธฐํƒ€ ๋Œ€ํ‘œ์ ์ธ ๋ชจ๋ธ์— ์ดˆ์ ์„ ๋งž์ถ”์–ด LLM์˜ ์ตœ์‹  ์ƒํƒœ์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ III์—์„œ๋Š” LLM์ด ์–ด๋–ป๊ฒŒ ๊ตฌ์ถ•๋˜๋Š”์ง€ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ IV์—์„œ๋Š” LLM์ด ์–ด๋–ป๊ฒŒ ์‚ฌ์šฉ๋˜๊ณ  ์‹ค์ œ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์„ ์œ„ํ•ด ์ฆ๊ฐ•๋˜๋Š”์ง€ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ V์™€ VI์—์„œ๋Š” LLM์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ์ธ๊ธฐ ์žˆ๋Š” ๋ฐ์ดํ„ฐ์…‹๊ณผ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ฒ€ํ† ํ•˜๊ณ  ๋ณด๊ณ ๋œ LLM ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์„น์…˜ VII์—์„œ๋Š” ๋„์ „ ๊ณผ์ œ์™€ ๋ฏธ๋ž˜ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์„ ์š”์•ฝํ•˜์—ฌ ๋…ผ๋ฌธ์„ ๋งˆ๋ฌด๋ฆฌํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์˜ ๋‚˜๋จธ์ง€ ๋ถ€๋ถ„์—์„œ๋Š” II์žฅ์—์„œ ์„ธ ๊ฐ€์ง€ LLM ์ œํ’ˆ๊ตฐ(GPT, LLaMA ๋ฐ PaLM)๊ณผ ๊ธฐํƒ€ ๋Œ€ํ‘œ์ ์ธ ๋ชจ๋ธ์— ์ดˆ์ ์„ ๋งž์ถฐ ์ตœ์‹  LLM ๊ฐœ์š”๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ III์—์„œ๋Š” LLM์ด ์–ด๋–ป๊ฒŒ ๊ตฌ์ถ•๋˜๋Š”์ง€์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ IV์—์„œ๋Š” LLM์ด ์‹ค์ œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์‚ฌ์šฉ๋˜๊ณ  ์ฆ๊ฐ•๋˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ V์™€ VI์—์„œ๋Š” LLM์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋Œ€์ค‘์ ์ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ฒ€ํ† ํ•˜๊ณ  ๋ณด๊ณ ๋œ LLM ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์„น์…˜ VII์—์„œ๋Š” ๊ณผ์ œ์™€ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์„ ์š”์•ฝํ•˜์—ฌ ๋…ผ๋ฌธ์„ ๋งˆ๋ฌด๋ฆฌํ•ฉ๋‹ˆ๋‹ค.
the elicited history, diagnostic accuracy, the appropriateness of management plans and their rationale, and patient-centred considerations such as relationship-building, respect for the individual and communication efficacy [20]. If the conversational potential of LLMs is to be realized in medicine, there is a significant unmet need to better optimize development and evaluation of medical AI systems for characteristics such as these, which are unique to history-taking and diagnostic dialogue between clinicians and patients.
์œ ๋„๋œ ๋ณ‘๋ ฅ, ์ง„๋‹จ ์ •ํ™•์„ฑ, ๊ด€๋ฆฌ ๊ณ„ํš์˜ ์ ์ ˆ์„ฑ ๋ฐ ๊ทธ ํ•ฉ๋ฆฌ์„ฑ, ๊ด€๊ณ„ ๊ตฌ์ถ•, ๊ฐœ์ธ์— ๋Œ€ํ•œ ์กด์ค‘ ๋ฐ ์˜์‚ฌ์†Œํ†ต ํšจ์œจ์„ฑ ๋“ฑ ํ™˜์ž ์ค‘์‹ฌ์˜ ๊ณ ๋ ค ์‚ฌํ•ญ [20]. LLM์˜ ๋Œ€ํ™” ์ž ์žฌ๋ ฅ์ด ์˜ํ•™์—์„œ ์‹คํ˜„๋˜๋ ค๋ฉด, ์ด๋Ÿฌํ•œ ํŠน์„ฑ์€ ์ž„์ƒ ์˜์‚ฌ์™€ ํ™˜์ž ๊ฐ„์˜ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™”์— ๊ณ ์œ ํ•œ ๊ฒƒ์ด๋ฏ€๋กœ, ์˜๋ฃŒ AI ์‹œ์Šคํ…œ์˜ ๊ฐœ๋ฐœ ๋ฐ ํ‰๊ฐ€๋ฅผ ๋” ์ž˜ ์ตœ์ ํ™”ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
๋„์ถœ๋œ ๋ณ‘๋ ฅ, ์ง„๋‹จ ์ •ํ™•๋„, ๊ด€๋ฆฌ ๊ณ„ํš์˜ ์ ์ ˆ์„ฑ ๋ฐ ๊ทผ๊ฑฐ, ๊ด€๊ณ„ ๊ตฌ์ถ•, ๊ฐœ์ธ์— ๋Œ€ํ•œ ์กด์ค‘, ์˜์‚ฌ์†Œํ†ต ํšจ์œจ์„ฑ๊ณผ ๊ฐ™์€ ํ™˜์ž ์ค‘์‹ฌ์  ๊ณ ๋ ค ์‚ฌํ•ญ [20]. ์˜ํ•™์—์„œ LLM์˜ ๋Œ€ํ™” ์ž ์žฌ๋ ฅ์„ ์‹คํ˜„ํ•˜๋ ค๋ฉด ์ž„์ƒ์˜์™€ ํ™˜์ž ๊ฐ„์˜ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™”์— ๊ณ ์œ ํ•œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŠน์„ฑ์— ๋Œ€ํ•ด ์˜๋ฃŒ AI ์‹œ์Šคํ…œ์˜ ๊ฐœ๋ฐœ ๋ฐ ํ‰๊ฐ€๋ฅผ ๋” ์ž˜ ์ตœ์ ํ™”ํ•ด์•ผ ํ•  ๋ฏธ์ถฉ์กฑ ์ˆ˜์š”๊ฐ€ ์ƒ๋‹นํžˆ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.
However, as shown in Fig. 3, initializing the target model as \(S\) results in a much bigger initial loss than that of \(M_{t-1}\), which could lead to an unstable training. The main reason is that DPO training is usually done where the reference and target models are the same. In contrast, initializing the target model as \(S\) creates a differential in the reference and target models, which may be amplified as the steps progress. Thus, for stable training, initializing the target model as \(M_{t-1}\) was chosen for sDPO.
๊ทธ๋Ÿฌ๋‚˜, Fig. 3์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ๋Œ€์ƒ ๋ชจ๋ธ์„ \(S\)๋กœ ์ดˆ๊ธฐํ™”ํ•˜๋ฉด \(M_{t-1}\)๋ณด๋‹ค ํ›จ์”ฌ ํฐ ์ดˆ๊ธฐ ์†์‹ค์ด ๋ฐœ์ƒํ•˜์—ฌ ํ›ˆ๋ จ์ด ๋ถˆ์•ˆ์ •ํ•ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์š” ์ด์œ ๋Š” DPO ํ›ˆ๋ จ์ด ์ผ๋ฐ˜์ ์œผ๋กœ ์ฐธ์กฐ ๋ชจ๋ธ๊ณผ ๋Œ€์ƒ ๋ชจ๋ธ์ด ๋™์ผํ•œ ๊ฒฝ์šฐ์— ์ˆ˜ํ–‰๋˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ๋Œ€์ƒ ๋ชจ๋ธ์„ \(S\)๋กœ ์ดˆ๊ธฐํ™”ํ•˜๋ฉด ์ฐธ์กฐ ๋ชจ๋ธ๊ณผ ๋Œ€์ƒ ๋ชจ๋ธ ๊ฐ„์˜ ์ฐจ์ด๊ฐ€ ์ƒ๊ธฐ๋ฉฐ, ์ด๋Š” ๋‹จ๊ณ„๊ฐ€ ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ ์ฆํญ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ์„ ์œ„ํ•ด sDPO์—์„œ๋Š” ๋Œ€์ƒ ๋ชจ๋ธ์„ \(M_{t-1}\)๋กœ ์ดˆ๊ธฐํ™”ํ•˜๋Š” ๊ฒƒ์ด ์„ ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ ๊ทธ๋ฆผ 3์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ๋ชฉํ‘œ ๋ชจ๋ธ์„ \(S\)๋กœ ์ดˆ๊ธฐํ™”ํ•˜๋ฉด \(M_{t-1}\)๋ณด๋‹ค ํ›จ์”ฌ ํฐ ์ดˆ๊ธฐ ์†์‹ค์ด ๋ฐœ์ƒํ•˜์—ฌ ๋ถˆ์•ˆ์ •ํ•œ ํ›ˆ๋ จ์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ๋œ ์ด์œ ๋Š” DPO ํ›ˆ๋ จ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ฐธ์กฐ ๋ชจ๋ธ๊ณผ ๋ชฉํ‘œ ๋ชจ๋ธ์ด ๋™์ผํ•œ ๊ฒฝ์šฐ์— ์ˆ˜ํ–‰๋˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ๋ชฉํ‘œ ๋ชจ๋ธ์„ \(S\)๋กœ ์ดˆ๊ธฐํ™”ํ•˜๋ฉด ์ฐธ์กฐ ๋ชจ๋ธ๊ณผ ๋ชฉํ‘œ ๋ชจ๋ธ์— ์ฐจ์ด๊ฐ€ ์ƒ๊ธฐ๋ฉฐ, ์ด ์ฐจ์ด๋Š” ๋‹จ๊ณ„๊ฐ€ ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ ์ฆํญ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ์„ ์œ„ํ•ด ๋ชฉํ‘œ ๋ชจ๋ธ์„ \(M_{t-1}\)๋กœ ์ดˆ๊ธฐํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ sDPO์— ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค.
Please act as an impartial judge and evaluate the quality of the response provided by an AI assistant to the user question displayed below. Your evaluation should consider factors such as the helpfulness, relevance, accuracy, depth, creativity, and level of detail of the response. Begin your evaluation by providing a short explanation. Be as objective as possible. After providing your explanation, please rate the response on a scale of 1 to 10 by strictly following this format: "[[rating]]", for example: "Rating: [[5]]".
๊ณต์ •ํ•œ ์‹ฌํŒ์œผ๋กœ์„œ ์•„๋ž˜์— ํ‘œ์‹œ๋œ ์‚ฌ์šฉ์ž ์งˆ๋ฌธ์— ๋Œ€ํ•œ AI ์–ด์‹œ์Šคํ„ดํŠธ์˜ ์‘๋‹ต ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•ด ์ฃผ์„ธ์š”. ํ‰๊ฐ€ ์‹œ ์‘๋‹ต์˜ ์œ ์šฉ์„ฑ, ๊ด€๋ จ์„ฑ, ์ •ํ™•์„ฑ, ๊นŠ์ด, ์ฐฝ์˜์„ฑ ๋ฐ ์„ธ๋ถ€ ์ˆ˜์ค€๊ณผ ๊ฐ™์€ ์š”์†Œ๋ฅผ ๊ณ ๋ คํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€๋ฅผ ์‹œ์ž‘ํ•  ๋•Œ ์งง์€ ์„ค๋ช…์„ ์ œ๊ณตํ•˜์„ธ์š”. ๊ฐ€๋Šฅํ•œ ํ•œ ๊ฐ๊ด€์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜์„ธ์š”. ์„ค๋ช…์„ ์ œ๊ณตํ•œ ํ›„, ๋‹ค์Œ ํ˜•์‹์„ ์—„๊ฒฉํžˆ ์ค€์ˆ˜ํ•˜์—ฌ ์‘๋‹ต์„ 1์—์„œ 10๊นŒ์ง€์˜ ์ฒ™๋„๋กœ ํ‰๊ฐ€ํ•ด ์ฃผ์„ธ์š”: "[[rating]]", ์˜ˆ๋ฅผ ๋“ค์–ด: "Rating: [[5]]".
๊ณต์ •ํ•œ ์‹ฌ์‚ฌ์œ„์›์ด ๋˜์–ด ์•„๋ž˜ ํ‘œ์‹œ๋œ ์‚ฌ์šฉ์ž ์งˆ๋ฌธ์— ๋Œ€ํ•ด AI ์–ด์‹œ์Šคํ„ดํŠธ๊ฐ€ ์ œ๊ณตํ•œ ๋‹ต๋ณ€์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•ด ์ฃผ์„ธ์š”. ํ‰๊ฐ€์—๋Š” ๋‹ต๋ณ€์˜ ์œ ์šฉ์„ฑ, ๊ด€๋ จ์„ฑ, ์ •ํ™•์„ฑ, ๊นŠ์ด, ์ฐฝ์˜์„ฑ ๋ฐ ์„ธ๋ถ€ ์ˆ˜์ค€๊ณผ ๊ฐ™์€ ์š”์†Œ๋ฅผ ๊ณ ๋ คํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์งง์€ ์„ค๋ช…์„ ์ œ๊ณตํ•˜์—ฌ ํ‰๊ฐ€๋ฅผ ์‹œ์ž‘ํ•˜์„ธ์š”. ๊ฐ€๋Šฅํ•œ ํ•œ ๊ฐ๊ด€์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜์„ธ์š”. ์„ค๋ช…์„ ์ œ๊ณตํ•œ ํ›„์—๋Š” ์ด ํ˜•์‹์„ ์—„๊ฒฉํ•˜๊ฒŒ ๋”ฐ๋ฅด๋Š” 1~10์  ์ฒ™๋„๋กœ ๋‹ต๋ณ€์„ ํ‰๊ฐ€ํ•˜์„ธ์š”: ์˜ˆ๋ฅผ ๋“ค์–ด "ํ‰๊ฐ€: [[5]]".
Choice of reference modelsTo mitigate biases from selecting just one reference model, we use three state-of-the-art open source SUTs as candidate references models. They were not chosen arbitrarily. The reference models were selected based on highest overall performance at a range of tasks among similarly-sized accessible models. We only considered models that were not used to create the evaluator model (i.e., LlamaGuard). For each test, the lowest scoring of the three candidate reference models is used as the reference.
์ฐธ์กฐ ๋ชจ๋ธ ์„ ํƒ: ๋‹จ์ผ ์ฐธ์กฐ ๋ชจ๋ธ์„ ์„ ํƒํ•˜๋Š” ๋ฐ์„œ ์˜ค๋Š” ํŽธํ–ฅ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์„ธ ๊ฐ€์ง€ ์ตœ์ฒจ๋‹จ ์˜คํ”ˆ ์†Œ์Šค SUTs๋ฅผ ํ›„๋ณด ์ฐธ์กฐ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ์ž„์˜๋กœ ์„ ํƒ๋œ ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ์ฐธ์กฐ ๋ชจ๋ธ์€ ์œ ์‚ฌํ•œ ํฌ๊ธฐ์˜ ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ์ค‘ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ๊ฐ€์žฅ ๋†’์€ ์ „๋ฐ˜์ ์ธ ์„ฑ๋Šฅ์„ ๊ธฐ์ค€์œผ๋กœ ์„ ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€ ๋ชจ๋ธ(์ฆ‰, LlamaGuard)์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์ง€ ์•Š์€ ๋ชจ๋ธ๋งŒ ๊ณ ๋ คํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ํ…Œ์ŠคํŠธ์—์„œ๋Š” ์„ธ ํ›„๋ณด ์ฐธ์กฐ ๋ชจ๋ธ ์ค‘ ๊ฐ€์žฅ ๋‚ฎ์€ ์ ์ˆ˜๋ฅผ ๋ฐ›์€ ๋ชจ๋ธ์ด ์ฐธ์กฐ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
์ฐธ์กฐ ๋ชจ๋ธ ์„ ํƒ๋‹จ์ผ ์ฐธ์กฐ ๋ชจ๋ธ์„ ์„ ํƒํ•  ๋•Œ ๋ฐœ์ƒํ•˜๋Š” ํŽธํ–ฅ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์„ธ ๊ฐ€์ง€ ์ตœ์‹  ์˜คํ”ˆ ์†Œ์Šค SUT๋ฅผ ํ›„๋ณด ์ฐธ์กฐ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ์ž„์˜๋กœ ์„ ํƒ๋œ ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ์ฐธ์กฐ ๋ชจ๋ธ์€ ๋น„์Šทํ•œ ๊ทœ๋ชจ์˜ ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ์ค‘์—์„œ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ๊ฐ€์žฅ ๋†’์€ ์ „๋ฐ˜์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๋ชจ๋ธ์„ ๊ธฐ์ค€์œผ๋กœ ์„ ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€์ž ๋ชจ๋ธ(์˜ˆ: LlamaGuard)์„ ๋งŒ๋“œ๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์ง€ ์•Š์€ ๋ชจ๋ธ๋งŒ ๊ณ ๋ คํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ํ…Œ์ŠคํŠธ์—์„œ ์„ธ ๊ฐœ์˜ ํ›„๋ณด ์ฐธ์กฐ ๋ชจ๋ธ ์ค‘ ๊ฐ€์žฅ ๋‚ฎ์€ ์ ์ˆ˜๋ฅผ ๋ฐ›์€ ๋ชจ๋ธ์ด ์ฐธ์กฐ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
Table 2 demonstrates the decline of multilingual capabilities when deactivating language-specific neurons. Although just deactivating around \(0.13\%\) neurons, LLMs lose multilingual capabilities including understanding multilingual inputs, handling multilingual questions, and generating multilingual outputs. In contrast, deactivating the same number of randomly selected neurons does not yield any difference. Therefore, the detected neurons are language-specific and related to handling corresponding multilingual inputs.
Table 2๋Š” ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•  ๋•Œ ๋‹ค๊ตญ์–ด ๋Šฅ๋ ฅ์ด ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์•ฝ \(0.13\%\)์˜ ๋‰ด๋Ÿฐ๋งŒ ๋น„ํ™œ์„ฑํ™”ํ•ด๋„, LLM์€ ๋‹ค๊ตญ์–ด ์ž…๋ ฅ์„ ์ดํ•ดํ•˜๊ณ , ๋‹ค๊ตญ์–ด ์งˆ๋ฌธ์„ ์ฒ˜๋ฆฌํ•˜๋ฉฐ, ๋‹ค๊ตญ์–ด ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์žƒ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ ๋™์ผํ•œ ์ˆ˜์˜ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋Š” ๊ฒฝ์šฐ์—๋Š” ์•„๋ฌด๋Ÿฐ ์ฐจ์ด๊ฐ€ ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ์ง€๋œ ๋‰ด๋Ÿฐ์€ ์–ธ์–ด๋ณ„๋กœ ํŠนํ™”๋˜์–ด ์žˆ์œผ๋ฉฐ, ํ•ด๋‹น ๋‹ค๊ตญ์–ด ์ž…๋ ฅ์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ๊ณผ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
ํ‘œ 2๋Š” ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•  ๋•Œ ๋‹ค๊ตญ์–ด ๊ธฐ๋Šฅ์ด ์ €ํ•˜๋˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์•ฝ 0.13\%\)์˜ ๋‰ด๋Ÿฐ๋งŒ ๋น„ํ™œ์„ฑํ™”ํ•ด๋„ LLM์€ ๋‹ค๊ตญ์–ด ์ž…๋ ฅ ์ดํ•ด, ๋‹ค๊ตญ์–ด ์งˆ๋ฌธ ์ฒ˜๋ฆฌ, ๋‹ค๊ตญ์–ด ์ถœ๋ ฅ ์ƒ์„ฑ ๋“ฑ ๋‹ค๊ตญ์–ด ๊ธฐ๋Šฅ์„ ์ƒ์‹คํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ ๋™์ผํ•œ ์ˆ˜์˜ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•ด๋„ ์ฐจ์ด๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ์ง€๋œ ๋‰ด๋Ÿฐ์€ ์–ธ์–ด๋ณ„์ด๋ฉฐ ํ•ด๋‹น ๋‹ค๊ตญ์–ด ์ž…๋ ฅ ์ฒ˜๋ฆฌ์™€ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
We use REINFORCE to optimize the likelihoods of the rationales based on their usefullness: the log-likelihood of the \(n_{true}\) true next tokens \(X_{j+1:j+n_{true}+1}\) under the language model given previous observed tokens and a particular rationale (\(p^{\text{talk}}_{j+n_{true}}\) as shorthand for the mixed prediction probabilities after thinking, see Algorithm 1). To reduce variance, we generate multiple rationale continuations for each token in the input sequence (loosely inspired by TRICE, Phan et al. (2023)).
์šฐ๋ฆฌ๋Š” REINFORCE๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ทผ๊ฑฐ์˜ ์œ ์šฉ์„ฑ์— ๊ธฐ๋ฐ˜ํ•œ ๊ฐ€๋Šฅ์„ฑ์„ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค: ์ด์ „์— ๊ด€์ฐฐ๋œ ํ† ํฐ๊ณผ ํŠน์ • ๊ทผ๊ฑฐ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ ์–ธ์–ด ๋ชจ๋ธ ํ•˜์—์„œ \(n_{true}\)๊ฐœ์˜ ์‹ค์ œ ๋‹ค์Œ ํ† ํฐ \(X_{j+1:j+n_{true}+1}\)์˜ ๋กœ๊ทธ ๊ฐ€๋Šฅ์„ฑ (\(p^{\text{talk}}_{j+n_{true}}\)๋Š” ์ƒ๊ฐ ํ›„ ํ˜ผํ•ฉ ์˜ˆ์ธก ํ™•๋ฅ ์˜ ์•ฝ์–ด๋กœ, ์•Œ๊ณ ๋ฆฌ์ฆ˜ 1์„ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค). ๋ถ„์‚ฐ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด, ์ž…๋ ฅ ์‹œํ€€์Šค์˜ ๊ฐ ํ† ํฐ์— ๋Œ€ํ•ด ์—ฌ๋Ÿฌ ๊ฐœ์˜ ๊ทผ๊ฑฐ ์—ฐ์†์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค (TRICE, Phan et al. (2023)์—์„œ ๋Š์Šจํ•˜๊ฒŒ ์˜๊ฐ์„ ๋ฐ›์Œ).
์šฐ๋ฆฌ๋Š” REINFORCE๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•ฉ๋ฆฌ์„ฑ์˜ ์œ ์šฉ์„ฑ์— ๋”ฐ๋ผ ํ•ฉ๋ฆฌํ™”์˜ ๊ฐ€๋Šฅ์„ฑ์„ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค: ์ด์ „์— ๊ด€์ฐฐ๋œ ํ† ํฐ๊ณผ ํŠน์ • ๊ทผ๊ฑฐ๊ฐ€ ์ฃผ์–ด์ง„ ์–ธ์–ด ๋ชจ๋ธ ์•„๋ž˜์—์„œ \(n_{true}\) ์ฐธ ์ฐจ๊ธฐ ํ† ํฐ \(X_{j+1:j+n_{true}+1}\)์˜ ๋กœ๊ทธ ๊ฐ€๋Šฅ์„ฑ(์•Œ๊ณ ๋ฆฌ์ฆ˜ 1์—์„œ ์ƒ๊ฐ ํ›„ ํ˜ผํ•ฉ ์˜ˆ์ธก ํ™•๋ฅ ์˜ ์•ฝ์–ด๋กœ \(p^{\text{talk}}_{j+n_{true}}\))์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ถ„์‚ฐ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ์ž…๋ ฅ ์‹œํ€€์Šค์˜ ๊ฐ ํ† ํฐ์— ๋Œ€ํ•ด ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ทผ๊ฑฐ ์—ฐ์†์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค(TRICE, Phan et al. (2023)์—์„œ ๋Š์Šจํ•˜๊ฒŒ ์˜๊ฐ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค).
In this section we conducted an assessment of the model's conversational abilities, considering aspects to ensure its effectiveness and safety. We have compiled a collection of open-source evaluation datasets from the community, such as alpaca-eval[21], Belle-eval [88], and MT-bench[93]. Additionally, we have established our own helpful and harmless evaluation dataset by gathering and constructing data of varying difficulty levels, for the purpose of comprehensively assessing the conversational abilities of chat models.
์ด ์„น์…˜์—์„œ๋Š” ๋ชจ๋ธ์˜ ๋Œ€ํ™” ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜์—ฌ ๊ทธ ํšจ๊ณผ์„ฑ๊ณผ ์•ˆ์ „์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•œ ์ธก๋ฉด์„ ๊ณ ๋ คํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” alpaca-eval[21], Belle-eval[88], MT-bench[93]์™€ ๊ฐ™์€ ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ์ œ๊ณต๋œ ์˜คํ”ˆ ์†Œ์Šค ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ์…‹์„ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ถ”๊ฐ€๋กœ, ๋‹ค์–‘ํ•œ ๋‚œ์ด๋„์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ๊ตฌ์„ฑํ•˜์—ฌ ์ž์ฒด์ ์ธ ์œ ์šฉํ•˜๊ณ  ๋ฌดํ•ดํ•œ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ์…‹์„ ๊ตฌ์ถ•ํ•˜์—ฌ ์ฑ— ๋ชจ๋ธ์˜ ๋Œ€ํ™” ๋Šฅ๋ ฅ์„ ์ข…ํ•ฉ์ ์œผ๋กœ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋ชจ๋ธ์˜ ๋Œ€ํ™” ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜์—ฌ ํšจ๊ณผ์™€ ์•ˆ์ „์„ฑ์„ ๋ณด์žฅํ•˜๋Š” ์ธก๋ฉด์„ ๊ณ ๋ คํ–ˆ์Šต๋‹ˆ๋‹ค. alpaca-eval[21], Belle-eval [88], MT-bench[93] ๋“ฑ ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ๊ณต๊ฐœ๋œ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ฑ„ํŒ… ๋ชจ๋ธ์˜ ๋Œ€ํ™” ๋Šฅ๋ ฅ์„ ์ข…ํ•ฉ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋‚œ์ด๋„์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ๊ตฌ์„ฑํ•˜์—ฌ ์œ ์šฉํ•˜๊ณ  ๋ฌดํ•ดํ•œ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ž์ฒด์ ์œผ๋กœ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค.
CustomizationCustomization is an optional step, involving the adjustment of content to align with the user's specific preferences or the context of the request. This tailoring includes adapting the content to meet the needs of the target audience or the format in which it will be presented and condensing the information to succinctly convey the essence of the content. The process also entails creating summaries or abstracts that emphasize the key points or arguments, ensuring the output is both informative and concise.
Customization์€ ์„ ํƒ์ ์ธ ๋‹จ๊ณ„๋กœ, ์‚ฌ์šฉ์ž์˜ ํŠน์ • ์„ ํ˜ธ๋„๋‚˜ ์š”์ฒญ์˜ ๋งฅ๋ฝ์— ๋งž์ถ”์–ด ์ฝ˜ํ…์ธ ๋ฅผ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์ด ๋งž์ถคํ™” ๊ณผ์ •์—๋Š” ๋Œ€์ƒ ์ฒญ์ค‘์˜ ์š”๊ตฌ๋‚˜ ์ฝ˜ํ…์ธ ๊ฐ€ ์ œ์‹œ๋  ํ˜•์‹์— ๋งž์ถ”์–ด ์ฝ˜ํ…์ธ ๋ฅผ ์กฐ์ •ํ•˜๊ณ , ์ •๋ณด๋ฅผ ๊ฐ„๊ฒฐํ•˜๊ฒŒ ์š”์•ฝํ•˜์—ฌ ๋ณธ์งˆ์„ ์ „๋‹ฌํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ฃผ์š” ์š”์ ์ด๋‚˜ ๋…ผ์ ์„ ๊ฐ•์กฐํ•˜๋Š” ์š”์•ฝ์ด๋‚˜ ๊ฐœ์š”๋ฅผ ์ž‘์„ฑํ•˜์—ฌ ์ถœ๋ ฅ๋ฌผ์ด ์ •๋ณด์ ์ด๊ณ  ๊ฐ„๊ฒฐํ•˜๋„๋ก ๋ณด์žฅํ•˜๋Š” ๊ณผ์ •๋„ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
์ปค์Šคํ„ฐ๋งˆ์ด์ง•์ปค์Šคํ„ฐ๋งˆ์ด์ง•์€ ์‚ฌ์šฉ์ž์˜ ํŠน์ • ์„ ํ˜ธ๋„ ๋˜๋Š” ์š”์ฒญ์˜ ๋งฅ๋ฝ์— ๋งž๊ฒŒ ์ฝ˜ํ…์ธ ๋ฅผ ์กฐ์ •ํ•˜๋Š” ์„ ํƒ์  ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋งž์ถค ์„ค์ •์—๋Š” ๋Œ€์ƒ ๊ณ ๊ฐ์˜ ์š”๊ตฌ ์‚ฌํ•ญ์ด๋‚˜ ์ œ์‹œ๋˜๋Š” ํ˜•์‹์— ๋งž๊ฒŒ ์ฝ˜ํ…์ธ ๋ฅผ ์กฐ์ •ํ•˜๊ณ  ์ฝ˜ํ…์ธ ์˜ ๋ณธ์งˆ์„ ๊ฐ„๊ฒฐํ•˜๊ฒŒ ์ „๋‹ฌํ•˜๊ธฐ ์œ„ํ•ด ์ •๋ณด๋ฅผ ์••์ถ•ํ•˜๋Š” ์ž‘์—…์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์š”์ ์ด๋‚˜ ์ฃผ์žฅ์„ ๊ฐ•์กฐํ•˜๋Š” ์š”์•ฝ ๋˜๋Š” ์ดˆ๋ก์„ ์ž‘์„ฑํ•˜์—ฌ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋Š” ๋™์‹œ์— ๊ฐ„๊ฒฐํ•˜๊ฒŒ ์ž‘์„ฑํ•˜๋Š” ๊ณผ์ •๋„ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
Observe that in the general case this algorithm is \(O(Nm\log(K))\) where \(N\) is the length of the dataset, \(m\) is the average length of a prefix match, and \(K\) is the number of splits. It is therefore incorrect to call this algorithm linear time in the general case, for ours it is. Because the length of the longest match is bounded above by the length of the longest sequence, as long as the size of the dataset is independent of the length of the longest sequence in the dataset, this algorithm remains efficient.
์ผ๋ฐ˜์ ์ธ ๊ฒฝ์šฐ ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ \(O(Nm\log(K))\)์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ \(N\)์€ ๋ฐ์ดํ„ฐ์…‹์˜ ๊ธธ์ด, \(m\)์€ ์ ‘๋‘์‚ฌ ์ผ์น˜์˜ ํ‰๊ท  ๊ธธ์ด, \(K\)๋Š” ๋ถ„ํ• ์˜ ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ผ๋ฐ˜์ ์ธ ๊ฒฝ์šฐ ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์„ ํ˜• ์‹œ๊ฐ„์ด๋ผ๊ณ  ๋ถ€๋ฅด๋Š” ๊ฒƒ์€ ์ž˜๋ชป๋œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๊ฒฝ์šฐ์—๋Š” ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๊ธด ์ผ์น˜์˜ ๊ธธ์ด๊ฐ€ ๊ฐ€์žฅ ๊ธด ์‹œํ€€์Šค์˜ ๊ธธ์ด๋กœ ์ œํ•œ๋˜๊ธฐ ๋•Œ๋ฌธ์—, ๋ฐ์ดํ„ฐ์…‹์˜ ํฌ๊ธฐ๊ฐ€ ๋ฐ์ดํ„ฐ์…‹ ๋‚ด ๊ฐ€์žฅ ๊ธด ์‹œํ€€์Šค์˜ ๊ธธ์ด์™€ ๋…๋ฆฝ์ ์ด๋ผ๋ฉด ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์—ฌ์ „ํžˆ ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค.
์ผ๋ฐ˜์ ์ธ ๊ฒฝ์šฐ ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ \(O(Nm\log(K))\)์ด๋ฉฐ, ์—ฌ๊ธฐ์„œ \(N\)์€ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์˜ ๊ธธ์ด, \(m\)์€ ์ ‘๋‘์‚ฌ ์ผ์น˜์˜ ํ‰๊ท  ๊ธธ์ด, \(K\)๋Š” ๋ถ„ํ•  ์ˆ˜์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ผ๋ฐ˜์ ์ธ ๊ฒฝ์šฐ ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์„ ํ˜• ์‹œ๊ฐ„ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด๋ผ๊ณ  ๋ถ€๋ฅด๋Š” ๊ฒƒ์€ ์˜ฌ๋ฐ”๋ฅด์ง€ ์•Š์œผ๋ฉฐ, ์šฐ๋ฆฌ์˜ ๊ฒฝ์šฐ์—๋Š” ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๊ธด ์ผ์น˜์˜ ๊ธธ์ด๋Š” ๊ฐ€์žฅ ๊ธด ์‹œํ€€์Šค์˜ ๊ธธ์ด๋กœ ์ƒํ•œ์ด ์ •ํ•ด์ ธ ์žˆ์œผ๋ฏ€๋กœ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์˜ ํฌ๊ธฐ๊ฐ€ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์—์„œ ๊ฐ€์žฅ ๊ธด ์‹œํ€€์Šค์˜ ๊ธธ์ด์™€ ๋ฌด๊ด€ํ•œ ํ•œ ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค.
In our initial effort in this domain, we limit ourselves to serial connections and non-adaptive configurations, deferring the investigation of more flexible model merging to future work. Concretely, with a collection of \(N\) models and a budget \(T\), our method searches for a sequence of layer indices \(L_{i,j}^{(t)}\) that delineates the path all the tokens should follow for a specific task. Here \(L_{i,j}\) denotes the \(j\)-th layer in the \(i\)-th model, with \(t\in[1,T]\) marking the step in the inference path.
์ดˆ๊ธฐ ๋…ธ๋ ฅ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ง๋ ฌ ์—ฐ๊ฒฐ๊ณผ ๋น„์ ์‘ํ˜• ๊ตฌ์„ฑ์—๋งŒ ์ œํ•œํ•˜๋ฉฐ, ๋” ์œ ์—ฐํ•œ ๋ชจ๋ธ ๋ณ‘ํ•ฉ์— ๋Œ€ํ•œ ์กฐ์‚ฌ๋Š” ํ–ฅํ›„ ์ž‘์—…์œผ๋กœ ๋ฏธ๋ฃน๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, \(N\)๊ฐœ์˜ ๋ชจ๋ธ๊ณผ ์˜ˆ์‚ฐ \(T\)๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ํŠน์ • ์ž‘์—…์„ ์œ„ํ•ด ๋ชจ๋“  ํ† ํฐ์ด ๋”ฐ๋ผ์•ผ ํ•  ๊ฒฝ๋กœ๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ๊ณ„์ธต ์ธ๋ฑ์Šค \(L_{i,j}^{(t)}\)์˜ ์‹œํ€€์Šค๋ฅผ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ \(L_{i,j}\)๋Š” \(i\)๋ฒˆ์งธ ๋ชจ๋ธ์˜ \(j\)๋ฒˆ์งธ ๊ณ„์ธต์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, \(t\in[1,T]\)๋Š” ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ๋‹จ๊ณ„๋ฅผ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.
์ด ์˜์—ญ์—์„œ์˜ ์ดˆ๊ธฐ ๋…ธ๋ ฅ์—์„œ๋Š” ๋ณด๋‹ค ์œ ์—ฐํ•œ ๋ชจ๋ธ ๋ณ‘ํ•ฉ์— ๋Œ€ํ•œ ์กฐ์‚ฌ๋ฅผ ํ–ฅํ›„ ์ž‘์—…์œผ๋กœ ๋ฏธ๋ฃจ๋ฉด์„œ ์ง๋ ฌ ์—ฐ๊ฒฐ๊ณผ ๋น„์ ์‘ ๊ตฌ์„ฑ์œผ๋กœ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, \(N\) ๊ฐœ์˜ ๋ชจ๋ธ ์ปฌ๋ ‰์…˜๊ณผ ์˜ˆ์‚ฐ \(T\)๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ, ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ํŠน์ • ์ž‘์—…์— ๋Œ€ํ•ด ๋ชจ๋“  ํ† ํฐ์ด ๋”ฐ๋ผ์•ผ ํ•˜๋Š” ๊ฒฝ๋กœ๋ฅผ delinatingํ•˜๋Š” ๊ณ„์ธต ์ธ๋ฑ์Šค \(L_{i,j}^{(t)}\)์˜ ์‹œํ€€์Šค๋ฅผ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ \(L_{i,j}\)๋Š” \(i\)-๋ฒˆ์งธ ๋ชจ๋ธ์˜ \(j\)-๋ฒˆ์งธ ๋ ˆ์ด์–ด๋ฅผ ๋‚˜ํƒ€๋‚ด๋ฉฐ, \(t\in[1,T]\)๋Š” ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ๋‹จ๊ณ„๋ฅผ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.
Xiao Wang, Shihao Wang, Yuhe Ding, Yuehang Li, Yao Rong, Ju Huang, Haoxiang Yang, Ziwen Wang, Bo Jiang, and Jin Tang are with the School of Computer Science and Technology, Anhui University, Hefei 230601, China. (email: [email protected]) Weizhe Kong, Wentao Wu, Shihao Li, and Chenglong Li are with the School of Artificial Intelligence, Anhui University, Hefei 230601, China. (email: [email protected]) Yaowei Wang is with Peng Cheng Laboratory, Shenzhen, China; Harbin Institute of Technology (HITS2), Shenzhen, China.
Xiao Wang, Shihao Wang, Yuhe Ding, Yuehang Li, Yao Rong, Ju Huang, Haoxiang Yang, Ziwen Wang, Bo Jiang, ๊ทธ๋ฆฌ๊ณ  Jin Tang์€ ์ค‘๊ตญ Hefei 230601์— ์œ„์น˜ํ•œ Anhui University์˜ ์ปดํ“จํ„ฐ ๊ณผํ•™ ๋ฐ ๊ธฐ์ˆ  ํ•™๋ถ€์— ์†Œ์†๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. (์ด๋ฉ”์ผ: [email protected]) Weizhe Kong, Wentao Wu, Shihao Li, ๊ทธ๋ฆฌ๊ณ  Chenglong Li๋Š” ์ค‘๊ตญ Hefei 230601์— ์œ„์น˜ํ•œ Anhui University์˜ ์ธ๊ณต์ง€๋Šฅ ํ•™๋ถ€์— ์†Œ์†๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. (์ด๋ฉ”์ผ: [email protected]) Yaowei Wang์€ ์ค‘๊ตญ Shenzhen์— ์œ„์น˜ํ•œ Peng Cheng Laboratory์™€ Harbin Institute of Technology (HITS2), Shenzhen์— ์†Œ์†๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
์ƒค์˜ค์™•, ์‹œํ•˜์˜ค์™•, ์œ ํ—ˆ๋”ฉ, ์›จํ•ญ๋ฆฌ, ์•ผ์˜ค๋กฑ, ์ฃผํ™ฉ, ํ•˜์˜ค์ƒน์–‘, ์ง€์›ฌ์™•, ๋ณด์žฅ, ์ง„ํƒ•์€ ์•ˆํ›„์ด๋Œ€ํ•™๊ต ์ปดํ“จํ„ฐ๊ณผํ•™๊ธฐ์ˆ ๋Œ€ํ•™(Hefei 230601, ์ค‘๊ตญ)์— ์†Œ์†๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. (์ด๋ฉ”์ผ: [email protected]) ์œ„์ œ ์ฝฉ, ์›ํƒ€์˜ค ์šฐ, ์‹œํ•˜์˜ค ๋ฆฌ, ์ฒธ๋กฑ ๋ฆฌ๋Š” ์•ˆํ›„์ด๋Œ€ํ•™๊ต ์ธ๊ณต์ง€๋Šฅ๋Œ€ํ•™(Hefei 230601, ์ค‘๊ตญ)์— ์†Œ์†๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. (์ด๋ฉ”์ผ: [email protected]) ์™•์•ผ์˜ค์›จ์ด๋Š” ํŽญ์ฒญ์—ฐ๊ตฌ์†Œ, ์„ ์ „, ์ค‘๊ตญ; ํ•˜์–ผ๋นˆ ๊ณต๊ณผ๋Œ€ํ•™(HITS2), ์„ ์ „, ์ค‘๊ตญ์— ์†Œ์†๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
Furthermore, they discovered that in teams without a designated leader, agents spent most of their time giving orders to one another (-50% of communication), splitting their remaining time between sharing information, or requesting guidance. Conversely, in teams with a designated leader, 60% of the leader's communication involved giving directions, prompting other members to focus more on exchanging and requesting information. Their results demonstrate that agent teams are most effective when the leader is a human.
๊ฒŒ๋‹ค๊ฐ€, ์ง€์ •๋œ ๋ฆฌ๋”๊ฐ€ ์—†๋Š” ํŒ€์—์„œ๋Š” ์—์ด์ „ํŠธ๋“ค์ด ๋Œ€๋ถ€๋ถ„์˜ ์‹œ๊ฐ„์„ ์„œ๋กœ์—๊ฒŒ ๋ช…๋ น์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ(์ „์ฒด ์˜์‚ฌ์†Œํ†ต์˜ 50%), ๋‚จ์€ ์‹œ๊ฐ„์€ ์ •๋ณด๋ฅผ ๊ณต์œ ํ•˜๊ฑฐ๋‚˜ ์ง€์นจ์„ ์š”์ฒญํ•˜๋Š” ๋ฐ ํ• ์• ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์ง€์ •๋œ ๋ฆฌ๋”๊ฐ€ ์žˆ๋Š” ํŒ€์—์„œ๋Š” ๋ฆฌ๋”์˜ ์˜์‚ฌ์†Œํ†ต ์ค‘ 60%๊ฐ€ ์ง€์‹œ๋ฅผ ๋‚ด๋ฆฌ๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ๊ณ , ๋‹ค๋ฅธ ๊ตฌ์„ฑ์›๋“ค์€ ์ •๋ณด ๊ตํ™˜๊ณผ ์š”์ฒญ์— ๋” ์ง‘์ค‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์˜ ๊ฒฐ๊ณผ๋Š” ๋ฆฌ๋”๊ฐ€ ์ธ๊ฐ„์ผ ๋•Œ ์—์ด์ „ํŠธ ํŒ€์ด ๊ฐ€์žฅ ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๋˜ํ•œ ์ง€์ •๋œ ๋ฆฌ๋”๊ฐ€ ์—†๋Š” ํŒ€์—์„œ๋Š” ์ƒ๋‹ด์›๋“ค์ด ๋Œ€๋ถ€๋ถ„์˜ ์‹œ๊ฐ„์„ ์„œ๋กœ์—๊ฒŒ ๋ช…๋ น์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ์†Œ๋น„ํ•˜๊ณ (-์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜์˜ 50%), ๋‚˜๋จธ์ง€ ์‹œ๊ฐ„์€ ์ •๋ณด๋ฅผ ๊ณต์œ ํ•˜๊ฑฐ๋‚˜ ์•ˆ๋‚ด๋ฅผ ์š”์ฒญํ•˜๋Š” ๋ฐ ์†Œ๋น„ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ ์ง€์ •๋œ ๋ฆฌ๋”๊ฐ€ ์žˆ๋Š” ํŒ€์—์„œ๋Š” ๋ฆฌ๋”์˜ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ์ค‘ 60%๊ฐ€ ์ง€์‹œํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ์œผ๋ฉฐ, ๋‹ค๋ฅธ ํŒ€์›๋“ค์€ ์ •๋ณด๋ฅผ ๊ตํ™˜ํ•˜๊ณ  ์š”์ฒญํ•˜๋Š” ๋ฐ ๋” ๋งŽ์€ ์‹œ๊ฐ„์„ ํ• ์• ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ๋ฆฌ๋”๊ฐ€ ์‚ฌ๋žŒ์ผ ๋•Œ ์ƒ๋‹ด์› ํŒ€์˜ ํšจ์œจ์„ฑ์ด ๊ฐ€์žฅ ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
Reflection [160] involves prompting LLMs to assess and potentially revise their own outputs based on reasoning about the correctness and coherence of their responses. The concept of Reflection centers on the ability of LLMs to engage in a form of self-evaluation. After generating an initial response, the model is prompted to reflect on its own output, considering factors like factual accuracy, logical consistency, and relevance. This introspective process can lead to the generation of revised or improved responses.
Reflection [160]์€ LLM์ด ์ž์‹ ์˜ ์ถœ๋ ฅ๋ฌผ์„ ํ‰๊ฐ€ํ•˜๊ณ , ์‘๋‹ต์˜ ์ •ํ™•์„ฑ๊ณผ ์ผ๊ด€์„ฑ์— ๋Œ€ํ•œ ์ถ”๋ก ์„ ๋ฐ”ํƒ•์œผ๋กœ ์ž ์žฌ์ ์œผ๋กœ ์ˆ˜์ •ํ•˜๋„๋ก ์œ ๋„ํ•˜๋Š” ๊ณผ์ •์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. Reflection์˜ ๊ฐœ๋…์€ LLM์ด ์ผ์ข…์˜ ์ž๊ธฐ ํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ์‘๋‹ต์„ ์ƒ์„ฑํ•œ ํ›„, ๋ชจ๋ธ์€ ์‚ฌ์‹ค์  ์ •ํ™•์„ฑ, ๋…ผ๋ฆฌ์  ์ผ๊ด€์„ฑ, ๊ด€๋ จ์„ฑ๊ณผ ๊ฐ™์€ ์š”์†Œ๋ฅผ ๊ณ ๋ คํ•˜์—ฌ ์ž์‹ ์˜ ์ถœ๋ ฅ๋ฌผ์„ ๋ฐ˜์„ฑํ•˜๋„๋ก ์œ ๋„๋ฉ๋‹ˆ๋‹ค. ์ด ๋‚ด์„ฑ์ ์ธ ๊ณผ์ •์€ ์ˆ˜์ •๋˜๊ฑฐ๋‚˜ ํ–ฅ์ƒ๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๊ธฐ์—ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ฐ˜์„ฑ[160]์€ LLM์ด ์‘๋‹ต์˜ ์ •ํ™•์„ฑ๊ณผ ์ผ๊ด€์„ฑ์— ๋Œ€ํ•œ ์ถ”๋ก ์„ ๋ฐ”ํƒ•์œผ๋กœ ์ž์ฒด ์ถœ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์ž ์žฌ์ ์œผ๋กœ ์ˆ˜์ •ํ•˜๋„๋ก ์š”์ฒญํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜์„ฑ์˜ ๊ฐœ๋…์€ LLM์ด ์ผ์ข…์˜ ์ž๊ธฐ ํ‰๊ฐ€์— ์ฐธ์—ฌํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ์‘๋‹ต์„ ์ƒ์„ฑํ•œ ํ›„, ๋ชจ๋ธ์€ ์‚ฌ์‹ค ์ •ํ™•์„ฑ, ๋…ผ๋ฆฌ์  ์ผ๊ด€์„ฑ, ๊ด€๋ จ์„ฑ ๋“ฑ์˜ ์š”์†Œ๋ฅผ ๊ณ ๋ คํ•˜์—ฌ ์ž์‹ ์˜ ์ถœ๋ ฅ์„ ๋ฐ˜์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฐ›์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ฑ์ฐฐ ๊ณผ์ •์„ ํ†ตํ•ด ์ˆ˜์ •๋˜๊ฑฐ๋‚˜ ๊ฐœ์„ ๋œ ์‘๋‹ต์ด ์ƒ์„ฑ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Information retrieval (IR).** IR can improve question-answering capabilities of LMs (Lewis et al., 2020; Shuster et al., 2021; Nakano et al., 2021). In event forecasting, access to diverse, up-to-date information is crucial (Tetlock and Gardner, 2015). Thus, a key component of our system is an IR architecture that furnishes the reasoning model with news articles, using LMs for query expansion, relevance ranking and summarization. Beyond our setting, using LMs for IR is an active research topic (Zhu et al., 2024).
**์ •๋ณด ๊ฒ€์ƒ‰ (IR).** IR์€ LMs์˜ ์งˆ๋ฌธ ์‘๋‹ต ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค (Lewis et al., 2020; Shuster et al., 2021; Nakano et al., 2021). ์ด๋ฒคํŠธ ์˜ˆ์ธก์—์„œ๋Š” ๋‹ค์–‘ํ•œ ์ตœ์‹  ์ •๋ณด์— ์ ‘๊ทผํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค (Tetlock and Gardner, 2015). ๋”ฐ๋ผ์„œ, ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์˜ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ๋‰ด์Šค ๊ธฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•˜๋Š” IR ์•„ํ‚คํ…์ฒ˜๋กœ, ์ฟผ๋ฆฌ ํ™•์žฅ, ๊ด€๋ จ์„ฑ ์ˆœ์œ„ ๋งค๊ธฐ๊ธฐ ๋ฐ ์š”์•ฝ์„ ์œ„ํ•ด LMs๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์„ค์ •์„ ๋„˜์–ด์„œ, IR์„ ์œ„ํ•ด LMs๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์€ ํ™œ๋ฐœํ•œ ์—ฐ๊ตฌ ์ฃผ์ œ์ž…๋‹ˆ๋‹ค (Zhu et al., 2024).
**์ •๋ณด ๊ฒ€์ƒ‰(IR).** IR์€ LM(Lewis et al., 2020; Shuster et al., 2021; Nakano et al., 2021)์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฒคํŠธ ์˜ˆ์ธก์—์„œ๋Š” ๋‹ค์–‘ํ•˜๊ณ  ์ตœ์‹  ์ •๋ณด์— ๋Œ€ํ•œ ์•ก์„ธ์Šค๊ฐ€ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค(Tetlock and Gardner, 2015). ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์˜ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ์งˆ์˜ ํ™•์žฅ, ๊ด€๋ จ์„ฑ ์ˆœ์œ„ ๋ฐ ์š”์•ฝ์„ ์œ„ํ•ด LM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๋ก  ๋ชจ๋ธ์— ๋‰ด์Šค ๊ธฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•˜๋Š” IR ์•„ํ‚คํ…์ฒ˜์ž…๋‹ˆ๋‹ค. ์ด ์„ค์ • ์™ธ์—๋„ IR์— LM์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์€ ํ™œ๋ฐœํ•œ ์—ฐ๊ตฌ ์ฃผ์ œ์ž…๋‹ˆ๋‹ค(Zhu et al., 2024).
**Inconsistency between GPT-4 and human assessment.** However, _WizardLM_ lost to ChatGPT on the hard skills, which is contrary to the conclusion of above human evaluation. The main reason is that: i) human preferences for tidy and vivid formatting and ii) in the manual annotation stage, people prefer additional points for code or math problems that can be compiled and passed, provided that the quality of responses is comparable. More supporting evidence please refer to the **Case Study** section at the Appendix I.
**GPT-4์™€ ์ธ๊ฐ„ ํ‰๊ฐ€ ๊ฐ„์˜ ๋ถˆ์ผ์น˜.** ๊ทธ๋Ÿฌ๋‚˜ _WizardLM_์€ ์–ด๋ ค์šด ๊ธฐ์ˆ ์—์„œ ChatGPT์— ํŒจ๋ฐฐํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ์œ„์˜ ์ธ๊ฐ„ ํ‰๊ฐ€ ๊ฒฐ๋ก ๊ณผ ๋ฐ˜๋Œ€๋ฉ๋‹ˆ๋‹ค. ์ฃผ์š” ์ด์œ ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค: i) ๊น”๋”ํ•˜๊ณ  ์ƒ๋™๊ฐ ์žˆ๋Š” ํ˜•์‹์„ ์„ ํ˜ธํ•˜๋Š” ์ธ๊ฐ„์˜ ์ทจํ–ฅ, ii) ์ˆ˜๋™ ์ฃผ์„ ๋‹จ๊ณ„์—์„œ ์‚ฌ๋žŒ๋“ค์ด ์ปดํŒŒ์ผ๋˜๊ณ  ํ†ต๊ณผํ•  ์ˆ˜ ์žˆ๋Š” ์ฝ”๋“œ๋‚˜ ์ˆ˜ํ•™ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ถ”๊ฐ€ ์ ์ˆ˜๋ฅผ ์„ ํ˜ธํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ๋” ๋งŽ์€ ์ฆ๊ฑฐ๋Š” ๋ถ€๋ก I์˜ **์‚ฌ๋ก€ ์—ฐ๊ตฌ** ์„น์…˜์„ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค.
**GPT-4์™€ ์ธ๊ฐ„ ํ‰๊ฐ€ ๊ฐ„์˜ ๋ถˆ์ผ์น˜**. ๊ทธ๋Ÿฌ๋‚˜ _WizardLM_์€ ํ•˜๋“œ ์Šคํ‚ฌ์—์„œ ChatGPT์— ํŒจํ–ˆ๋Š”๋ฐ, ์ด๋Š” ์œ„์˜ ์ธ๊ฐ„ ํ‰๊ฐ€์˜ ๊ฒฐ๋ก ๊ณผ ์ƒ๋ฐ˜๋ฉ๋‹ˆ๋‹ค. ์ฃผ๋œ ์ด์œ ๋Š” i) ๊น”๋”ํ•˜๊ณ  ์ƒ์ƒํ•œ ์„œ์‹์— ๋Œ€ํ•œ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์™€ ii) ์ˆ˜๋™ ์ฃผ์„ ๋‹จ๊ณ„์—์„œ ์‘๋‹ต์˜ ํ’ˆ์งˆ์ด ๋น„์Šทํ•˜๋‹ค๋ฉด ์ฝ”๋“œ๋‚˜ ์ˆ˜ํ•™ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ปดํŒŒ์ผํ•˜์—ฌ ํ†ต๊ณผํ•  ์ˆ˜ ์žˆ๋Š” ์ถ”๊ฐ€ ์ ์ˆ˜๋ฅผ ์„ ํ˜ธํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ž์„ธํ•œ ์ฆ๊ฑฐ๋Š” ๋ถ€๋ก I์˜ **์‚ฌ๋ก€ ์—ฐ๊ตฌ** ์„น์…˜์„ ์ฐธ์กฐํ•˜์„ธ์š”.
Biases.We conduct a basic analysis of the toxicity of RefinedWeb in Figure 4. We find RW to be about as toxic as The Pile, based on the definition of toxicity provided by the Perspective API: "content that is rude or disrespectful". Notably, this definition does not cover issues with social biases or harmfulness. Although it is unlikely that our pipeline introduces further issues on this side than is already documented for popular datasets, we encourage further quantitative work on the public extract of RefinedWeb.
ํŽธํ–ฅ์„ฑ. ์šฐ๋ฆฌ๋Š” Figure 4์—์„œ RefinedWeb์˜ ์œ ํ•ด์„ฑ์— ๋Œ€ํ•œ ๊ธฐ๋ณธ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. Perspective API์—์„œ ์ œ๊ณตํ•˜๋Š” ์œ ํ•ด์„ฑ์˜ ์ •์˜์— ๋”ฐ๋ฅด๋ฉด, RW๋Š” The Pile๊ณผ ๊ฑฐ์˜ ๋™์ผํ•œ ์ˆ˜์ค€์˜ ์œ ํ•ด์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค: "๋ฌด๋ก€ํ•˜๊ฑฐ๋‚˜ ๋ถˆ์พŒํ•œ ์ฝ˜ํ…์ธ ". ์ฃผ๋ชฉํ•  ์ ์€ ์ด ์ •์˜๊ฐ€ ์‚ฌํšŒ์  ํŽธํ–ฅ์ด๋‚˜ ํ•ด๋กœ์›€๊ณผ ๊ด€๋ จ๋œ ๋ฌธ์ œ๋ฅผ ๋‹ค๋ฃจ์ง€ ์•Š๋Š”๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ํŒŒ์ดํ”„๋ผ์ธ์ด ์ธ๊ธฐ ์žˆ๋Š” ๋ฐ์ดํ„ฐ์…‹์— ์ด๋ฏธ ๋ฌธ์„œํ™”๋œ ๊ฒƒ๋ณด๋‹ค ๋” ๋งŽ์€ ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•  ๊ฐ€๋Šฅ์„ฑ์€ ๋‚ฎ์ง€๋งŒ, RefinedWeb์˜ ๊ณต๊ฐœ ์ถ”์ถœ๋ฌผ์— ๋Œ€ํ•œ ์ถ”๊ฐ€์ ์ธ ์ •๋Ÿ‰์  ์—ฐ๊ตฌ๋ฅผ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค.
ํŽธ๊ฒฌ.๊ทธ๋ฆผ 4์—์„œ RefinedWeb์˜ ๋…์„ฑ์— ๋Œ€ํ•œ ๊ธฐ๋ณธ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. Perspective API์—์„œ ์ œ๊ณตํ•˜๋Š” ๋…์„ฑ์˜ ์ •์˜์ธ "๋ฌด๋ก€ํ•˜๊ฑฐ๋‚˜ ๋ฌด๋ก€ํ•œ ์ฝ˜ํ…์ธ "๋ฅผ ๊ธฐ์ค€์œผ๋กœ RW๊ฐ€ The Pile๋งŒํผ ๋…์„ฑ์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์ด ์ •์˜์—๋Š” ์‚ฌํšŒ์  ํŽธ๊ฒฌ์ด๋‚˜ ์œ ํ•ด์„ฑ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํŒŒ์ดํ”„๋ผ์ธ์ด ์ด๋ฏธ ๋Œ€์ค‘์ ์ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ๋ฌธ์„œํ™”๋œ ๊ฒƒ๋ณด๋‹ค ์ด ์ธก๋ฉด์—์„œ ์ถ”๊ฐ€์ ์ธ ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•  ๊ฐ€๋Šฅ์„ฑ์€ ๋‚ฎ์ง€๋งŒ, RefinedWeb์˜ ๊ณต๊ฐœ ์ถ”์ถœ์— ๋Œ€ํ•œ ์ถ”๊ฐ€์ ์ธ ์ •๋Ÿ‰์  ์—ฐ๊ตฌ๋ฅผ ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค.
**Tools**. In the context of AI agents, tools represent any functions that the model can call. They allow the agent to interact with external data sources by pulling or pushing information to that source. An example of an agent persona and associated tools is a professional contract writer. The writer is given a persona explaining their role and the types of tasks it must accomplish. It is also given tools related to adding notes to a document, reading an existing document, or sending an email with a final draft.
**๋„๊ตฌ**. AI ์—์ด์ „ํŠธ์˜ ๋งฅ๋ฝ์—์„œ ๋„๊ตฌ๋Š” ๋ชจ๋ธ์ด ํ˜ธ์ถœํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋“  ๊ธฐ๋Šฅ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋“ค์€ ์—์ด์ „ํŠธ๊ฐ€ ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ ์†Œ์Šค์™€ ์ƒํ˜ธ์ž‘์šฉํ•˜์—ฌ ์ •๋ณด๋ฅผ ๊ฐ€์ ธ์˜ค๊ฑฐ๋‚˜ ๋ณด๋‚ด๋Š” ๊ฒƒ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ ํŽ˜๋ฅด์†Œ๋‚˜์™€ ๊ด€๋ จ ๋„๊ตฌ์˜ ์˜ˆ๋กœ๋Š” ์ „๋ฌธ ๊ณ„์•ฝ์„œ ์ž‘์„ฑ์ž๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ž‘์„ฑ์ž๋Š” ์ž์‹ ์˜ ์—ญํ• ๊ณผ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•  ์ž‘์—… ์œ ํ˜•์„ ์„ค๋ช…ํ•˜๋Š” ํŽ˜๋ฅด์†Œ๋‚˜๋ฅผ ๋ถ€์—ฌ๋ฐ›์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฌธ์„œ์— ๋ฉ”๋ชจ๋ฅผ ์ถ”๊ฐ€ํ•˜๊ฑฐ๋‚˜ ๊ธฐ์กด ๋ฌธ์„œ๋ฅผ ์ฝ๊ฑฐ๋‚˜ ์ตœ์ข… ์ดˆ์•ˆ์„ ์ด๋ฉ”์ผ๋กœ ๋ณด๋‚ด๋Š” ๊ฒƒ๊ณผ ๊ด€๋ จ๋œ ๋„๊ตฌ๋„ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
**๋„๊ตฌ**. AI ์—์ด์ „ํŠธ์˜ ๋งฅ๋ฝ์—์„œ ๋„๊ตฌ๋Š” ๋ชจ๋ธ์ด ํ˜ธ์ถœํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋“  ๊ธฐ๋Šฅ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์—์ด์ „ํŠธ๋Š” ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ ์†Œ์Šค์™€ ์ƒํ˜ธ ์ž‘์šฉํ•˜์—ฌ ํ•ด๋‹น ์†Œ์Šค๋กœ ์ •๋ณด๋ฅผ ๋Œ์–ด์˜ค๊ฑฐ๋‚˜ ๋ฐ€์–ด ๋„ฃ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ ํŽ˜๋ฅด์†Œ๋‚˜ ๋ฐ ๊ด€๋ จ ๋„๊ตฌ์˜ ์˜ˆ๋กœ๋Š” ์ „๋ฌธ ๊ณ„์•ฝ ์ž‘์„ฑ์ž๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ž‘์„ฑ์ž์—๊ฒŒ๋Š” ์ž์‹ ์˜ ์—ญํ• ๊ณผ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•˜๋Š” ์ž‘์—… ์œ ํ˜•์„ ์„ค๋ช…ํ•˜๋Š” ํŽ˜๋ฅด์†Œ๋‚˜๊ฐ€ ์ฃผ์–ด์ง‘๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฌธ์„œ์— ๋ฉ”๋ชจ๋ฅผ ์ถ”๊ฐ€ํ•˜๊ฑฐ๋‚˜, ๊ธฐ์กด ๋ฌธ์„œ๋ฅผ ์ฝ๊ฑฐ๋‚˜, ์ตœ์ข… ์ดˆ์•ˆ์ด ํฌํ•จ๋œ ์ด๋ฉ”์ผ์„ ๋ณด๋‚ด๋Š” ๊ฒƒ๊ณผ ๊ด€๋ จ๋œ ๋„๊ตฌ๋„ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
respectively. We compare how CKL methods for T5 perform on IL, NLE\({}_{\text{P1}}\), and NLE\({}_{\text{P2}}\) when continually pretrained entirely on Small for 5k steps (8 epochs), and when sequentially pretrained on Small-P1 and then on Small-P2 for 2.5k steps (8 epochs) each. In the scenario Small-P1\(\rightarrow\)Small-P2, there are two CKL phases where \(D_{0}\) is C4 and Wikipedia, \(D_{1}\) is Small-P1, and \(D_{2}\) is Small-P2. The rest of the configurations are set identical with the main experiments.
๊ฐ๊ฐ. ์šฐ๋ฆฌ๋Š” T5์— ๋Œ€ํ•œ CKL ๋ฐฉ๋ฒ•์ด Small์—์„œ 5k ์Šคํ…(8 ์—ํฌํฌ) ๋™์•ˆ ์™„์ „ํžˆ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ๋˜์—ˆ์„ ๋•Œ์™€ Small-P1์—์„œ ๋จผ์ € ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ํ›„ Small-P2์—์„œ ๊ฐ๊ฐ 2.5k ์Šคํ…(8 ์—ํฌํฌ) ๋™์•ˆ ์ˆœ์ฐจ์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ๋˜์—ˆ์„ ๋•Œ IL, NLE\({}_{\text{P1}}\), NLE\({}_{\text{P2}}\)์—์„œ ์–ด๋–ป๊ฒŒ ์ˆ˜ํ–‰๋˜๋Š”์ง€ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. Small-P1\(\rightarrow\)Small-P2 ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” \(D_{0}\)๊ฐ€ C4์™€ Wikipedia, \(D_{1}\)๊ฐ€ Small-P1, \(D_{2}\)๊ฐ€ Small-P2์ธ ๋‘ ๊ฐœ์˜ CKL ๋‹จ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€ ๊ตฌ์„ฑ์€ ์ฃผ์š” ์‹คํ—˜๊ณผ ๋™์ผํ•˜๊ฒŒ ์„ค์ •๋ฉ๋‹ˆ๋‹ค.
๊ฐ๊ฐ์ž…๋‹ˆ๋‹ค. 5k ๋‹จ๊ณ„(8ํšŒ) ๋™์•ˆ Small์— ์™„์ „ํžˆ ์‚ฌ์ „ ํ•™์Šตํ•˜๊ณ , Small-P1์— ์‚ฌ์ „ ํ•™์Šตํ•œ ๋‹ค์Œ Small-P2์— ๊ฐ๊ฐ 2.5k ๋‹จ๊ณ„(8ํšŒ) ๋™์•ˆ ์ˆœ์ฐจ์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•  ๋•Œ T5์šฉ CKL ๋ฐฉ๋ฒ•์ด IL, NLE\({}_{\text{P1}}\), NLE\({}_{\text{P2}}\)์—์„œ ์–ด๋–ป๊ฒŒ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š”์ง€ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. Small-P1\(\rightarrow\)Small-P2 ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” \(D_{0}\)๊ฐ€ C4์ด๊ณ  ์œ„ํ‚คํ”ผ๋””์•„, \(D_{1}\)๊ฐ€ Small-P1์ด๊ณ  \(D_{2}\)๊ฐ€ Small-P2์ธ ๋‘ ๊ฐœ์˜ CKL ๋‹จ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€ ๊ตฌ์„ฑ์€ ์ฃผ์š” ์‹คํ—˜๊ณผ ๋™์ผํ•˜๊ฒŒ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.
The WG has a unique combination of deep technical understanding of how to build and use machine learning models, benchmarks, and evaluation metrics; as well as policy expertise, governance experience, and substantive knowledge in trust and safety. We believe we are well-positioned to deliver safety evaluation benchmarks to push safety standards forward. Our broad membership includes a diverse mix of stakeholders. This is crucial, given that AI safety is a collective challenge and needs a collective solution [3].
WG๋Š” ๋จธ์‹  ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๊ณ  ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•, ๋ฒค์น˜๋งˆํฌ ๋ฐ ํ‰๊ฐ€ ์ง€ํ‘œ์— ๋Œ€ํ•œ ๊นŠ์€ ๊ธฐ์ˆ ์  ์ดํ•ด์™€ ์ •์ฑ… ์ „๋ฌธ์„ฑ, ๊ฑฐ๋ฒ„๋„Œ์Šค ๊ฒฝํ—˜, ์‹ ๋ขฐ์™€ ์•ˆ์ „์— ๋Œ€ํ•œ ์‹ค์งˆ์ ์ธ ์ง€์‹์„ ๋…ํŠนํ•˜๊ฒŒ ๊ฒฐํ•ฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์•ˆ์ „ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ œ๊ณตํ•˜์—ฌ ์•ˆ์ „ ๊ธฐ์ค€์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ์ข‹์€ ์œ„์น˜์— ์žˆ๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ํšŒ์› ๊ตฌ์„ฑ์€ ๋‹ค์–‘ํ•œ ์ดํ•ด ๊ด€๊ณ„์ž๋“ค๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AI ์•ˆ์ „์ด ์ง‘๋‹จ์ ์ธ ๋„์ „ ๊ณผ์ œ์ด๋ฉฐ ์ง‘๋‹จ์ ์ธ ํ•ด๊ฒฐ์ฑ…์ด ํ•„์š”ํ•˜๋‹ค๋Š” ์ ์—์„œ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค [3].
WG๋Š” ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ, ๋ฒค์น˜๋งˆํฌ ๋ฐ ํ‰๊ฐ€ ์ง€ํ‘œ ๊ตฌ์ถ• ๋ฐ ์‚ฌ์šฉ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ๊นŠ์€ ๊ธฐ์ˆ ์  ์ดํ•ด์™€ ์ •์ฑ… ์ „๋ฌธ์„ฑ, ๊ฑฐ๋ฒ„๋„Œ์Šค ๊ฒฝํ—˜, ์‹ ๋ขฐ ๋ฐ ์•ˆ์ „์— ๋Œ€ํ•œ ์‹ค์งˆ์ ์ธ ์ง€์‹์„ ๊ฐ–์ถ˜ ๋…๋ณด์ ์ธ ์กฐํ•ฉ์„ ๊ฐ–์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์•ˆ์ „ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ œ๊ณตํ•˜์—ฌ ์•ˆ์ „ ํ‘œ์ค€์„ ๋ฐœ์ „์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ์œ ๋ฆฌํ•œ ์œ„์น˜์— ์žˆ๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค. ๊ด‘๋ฒ”์œ„ํ•œ ํšŒ์›์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ๋‹ค์–‘ํ•œ ์ดํ•ด๊ด€๊ณ„์ž๋ฅผ ํฌ๊ด„ํ•ฉ๋‹ˆ๋‹ค. AI ์•ˆ์ „์€ ๊ณต๋™์˜ ๊ณผ์ œ์ด๋ฉฐ ๊ณต๋™์˜ ํ•ด๊ฒฐ์ฑ…์ด ํ•„์š”ํ•˜๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•  ๋•Œ ์ด๋Š” ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค[3].
tion and answer. While NQ appears to be closer to PopQA in terms of question and answer lengths, the key factor influencing the better results of training Adapt-LLM on SQuAD may be the number of questions in the training dataset (\(\sim\)87K in SQuAD and \(\sim\)58K in NQ). Further analyses are required to elucidate the factors that render a training dataset more suitable for a given target dataset (which is beyond the scope of our study), but these results suggest that scale may play once again a crucial role.
์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์˜ ๊ธธ์ด ์ธก๋ฉด์—์„œ NQ๊ฐ€ PopQA์— ๋” ๊ฐ€๊นŒ์›Œ ๋ณด์ด์ง€๋งŒ, Adapt-LLM์„ SQuAD์—์„œ ํ›ˆ๋ จํ–ˆ์„ ๋•Œ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ค๋Š” ์ฃผ์š” ์š”์ธ์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹์˜ ์งˆ๋ฌธ ์ˆ˜์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค (SQuAD์—์„œ๋Š” ์•ฝ 87K, NQ์—์„œ๋Š” ์•ฝ 58K). ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹์ด ์ฃผ์–ด์ง„ ๋Œ€์ƒ ๋ฐ์ดํ„ฐ์…‹์— ๋” ์ ํ•ฉํ•˜๊ฒŒ ๋งŒ๋“œ๋Š” ์š”์ธ์„ ๋ฐํžˆ๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ๋ถ„์„์ด ํ•„์š”ํ•˜์ง€๋งŒ(์ด๋Š” ๋ณธ ์—ฐ๊ตฌ์˜ ๋ฒ”์œ„๋ฅผ ๋ฒ—์–ด๋‚ฉ๋‹ˆ๋‹ค), ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๊ทœ๋ชจ๊ฐ€ ๋‹ค์‹œ ํ•œ ๋ฒˆ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
tion ๋ฐ ๋‹ต๋ณ€. NQ๋Š” ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์˜ ๊ธธ์ด ์ธก๋ฉด์—์„œ PopQA์— ๋” ๊ฐ€๊นŒ์›Œ ๋ณด์ด์ง€๋งŒ, SQuAD์—์„œ Adapt-LLM์„ ํ•™์Šต์‹œํ‚ค๋Š” ๋” ๋‚˜์€ ๊ฒฐ๊ณผ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ํ•ต์‹ฌ ์š”์†Œ๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์งˆ๋ฌธ ์ˆ˜(\(sim\)87K, SQuAD์—์„œ \(\sim\)58K, NQ์—์„œ \(\sim\)58K)์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠน์ • ๋Œ€์ƒ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋” ์ ํ•ฉํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“œ๋Š” ์š”์ธ์„ ๊ทœ๋ช…ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ถ”๊ฐ€ ๋ถ„์„์ด ํ•„์š”ํ•˜์ง€๋งŒ(์ด ์—ฐ๊ตฌ์˜ ๋ฒ”์œ„๋ฅผ ๋ฒ—์–ด๋‚œ ๋‚ด์šฉ), ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๊ทœ๋ชจ๊ฐ€ ๋‹ค์‹œ ํ•œ ๋ฒˆ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
We first train a series of GPT2 models on the \(\mathsf{bioS}(N)\) datasets (see Section 2.2) using mixed-precision fp16. The training protocol ensures that each piece of knowledge is presented 1000 times, a process we refer to as "1000 exposures." It's important to clarify that this differs from making 1000 passes over the data. For example, a single pass through Wiki data might expose the knowledge (US, capital, Washington D.C.) 1000 times, whereas a pass through the Common Crawl might do so a million times.
์šฐ๋ฆฌ๋Š” ๋จผ์ € \(\mathsf{bioS}(N)\) ๋ฐ์ดํ„ฐ์…‹(์„น์…˜ 2.2 ์ฐธ์กฐ)์—์„œ ํ˜ผํ•ฉ ์ •๋ฐ€๋„ fp16์„ ์‚ฌ์šฉํ•˜์—ฌ ์ผ๋ จ์˜ GPT2 ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ํ”„๋กœํ† ์ฝœ์€ ๊ฐ ์ง€์‹ ์กฐ๊ฐ์ด 1000๋ฒˆ ์ œ์‹œ๋˜๋„๋ก ๋ณด์žฅํ•˜๋ฉฐ, ์ด๋ฅผ "1000๋ฒˆ ๋…ธ์ถœ"์ด๋ผ๊ณ  ๋ถ€๋ฆ…๋‹ˆ๋‹ค. ์ด๋Š” ๋ฐ์ดํ„ฐ๋ฅผ 1000๋ฒˆ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ๊ณผ๋Š” ๋‹ค๋ฅด๋‹ค๋Š” ์ ์„ ๋ช…ํ™•ํžˆ ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Wiki ๋ฐ์ดํ„ฐ๋ฅผ ํ•œ ๋ฒˆ ํ†ต๊ณผํ•˜๋Š” ๋™์•ˆ (๋ฏธ๊ตญ, ์ˆ˜๋„, ์›Œ์‹ฑํ„ด D.C.)๋ผ๋Š” ์ง€์‹์„ 1000๋ฒˆ ๋…ธ์ถœํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, Common Crawl์„ ํ•œ ๋ฒˆ ํ†ต๊ณผํ•˜๋Š” ๋™์•ˆ์—๋Š” ๋ฐฑ๋งŒ ๋ฒˆ ๋…ธ์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋จผ์ € ํ˜ผํ•ฉ ์ •๋ฐ€๋„ fp16์„ ์‚ฌ์šฉํ•˜์—ฌ \(\mathsf{bioS}(N)\) ๋ฐ์ดํ„ฐ ์„ธํŠธ(2.2์ ˆ ์ฐธ์กฐ)์— ๋Œ€ํ•ด ์ผ๋ จ์˜ GPT2 ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ํ”„๋กœํ† ์ฝœ์€ ๊ฐ ์ง€์‹์ด 1000๋ฒˆ ์ œ์‹œ๋˜๋„๋ก ๋ณด์žฅํ•˜๋ฉฐ, ์ด ๊ณผ์ •์„ "1000 ๋…ธ์ถœ"์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋ฐ์ดํ„ฐ๋ฅผ 1000๋ฒˆ ํ†ต๊ณผํ•˜๋Š” ๊ฒƒ๊ณผ๋Š” ๋‹ค๋ฅด๋‹ค๋Š” ์ ์„ ๋ช…ํ™•ํžˆ ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์œ„ํ‚ค ๋ฐ์ดํ„ฐ๋ฅผ ํ•œ ๋ฒˆ ํ†ต๊ณผํ•˜๋ฉด ํ•ด๋‹น ์ง€์‹(๋ฏธ๊ตญ, ์ˆ˜๋„, ์›Œ์‹ฑํ„ด D.C.)์ด 1000๋ฒˆ ๋…ธ์ถœ๋  ์ˆ˜ ์žˆ์ง€๋งŒ, ์ปค๋จผ ํฌ๋กค๋ง์„ ํ†ต๊ณผํ•˜๋ฉด ๋ฐฑ๋งŒ ๋ฒˆ ๋…ธ์ถœ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Deontological ethical theories would say that the doctor must prioritize the life of the mother, as it is inherently wrong to take the life of an innocent person. Teleological ethical theories would say that the doctor should prioritize the life of the fetus, as the outcome or result is more important than the intentions or actions of the doctor. Virtue ethical theories would say that the doctor should prioritize the life of the person with the strongest moral character, which in this case would be the mother.
์˜๋ฌด๋ก ์  ์œค๋ฆฌ ์ด๋ก ์— ๋”ฐ๋ฅด๋ฉด, ๋ฌด๊ณ ํ•œ ์‚ฌ๋žŒ์˜ ์ƒ๋ช…์„ ๋นผ์•—๋Š” ๊ฒƒ์€ ๋ณธ์งˆ์ ์œผ๋กœ ์ž˜๋ชป๋œ ์ผ์ด๊ธฐ ๋•Œ๋ฌธ์— ์˜์‚ฌ๋Š” ์–ด๋จธ๋‹ˆ์˜ ์ƒ๋ช…์„ ์šฐ์„ ์‹œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ชฉ์ ๋ก ์  ์œค๋ฆฌ ์ด๋ก ์— ๋”ฐ๋ฅด๋ฉด, ๊ฒฐ๊ณผ๋‚˜ ๊ฒฐ๊ณผ๊ฐ€ ์˜์‚ฌ์˜ ์˜๋„๋‚˜ ํ–‰๋™๋ณด๋‹ค ๋” ์ค‘์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์˜์‚ฌ๋Š” ํƒœ์•„์˜ ์ƒ๋ช…์„ ์šฐ์„ ์‹œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋• ์œค๋ฆฌ ์ด๋ก ์— ๋”ฐ๋ฅด๋ฉด, ์˜์‚ฌ๋Š” ๊ฐ€์žฅ ๊ฐ•ํ•œ ๋„๋•์  ์„ฑ๊ฒฉ์„ ๊ฐ€์ง„ ์‚ฌ๋žŒ์˜ ์ƒ๋ช…์„ ์šฐ์„ ์‹œํ•ด์•ผ ํ•˜๋ฉฐ, ์ด ๊ฒฝ์šฐ์—๋Š” ์–ด๋จธ๋‹ˆ๊ฐ€ ํ•ด๋‹น๋ฉ๋‹ˆ๋‹ค.
์˜๋ฌด๋ก ์  ์œค๋ฆฌ ์ด๋ก ์€ ๋ฌด๊ณ ํ•œ ์‚ฌ๋žŒ์˜ ์ƒ๋ช…์„ ๋นผ์•—๋Š” ๊ฒƒ์€ ๋ณธ์งˆ์ ์œผ๋กœ ์ž˜๋ชป๋œ ๊ฒƒ์ด๋ฏ€๋กœ ์˜์‚ฌ๋Š” ์‚ฐ๋ชจ์˜ ์ƒ๋ช…์„ ์šฐ์„ ์‹œํ•ด์•ผ ํ•œ๋‹ค๊ณ  ๋งํ•ฉ๋‹ˆ๋‹ค. ๋ชฉ์ ๋ก ์  ์œค๋ฆฌ ์ด๋ก ์€ ์˜์‚ฌ์˜ ์˜๋„๋‚˜ ํ–‰๋™๋ณด๋‹ค ๊ฒฐ๊ณผ๋‚˜ ๊ฒฐ๊ณผ๊ฐ€ ๋” ์ค‘์š”ํ•˜๋ฏ€๋กœ ์˜์‚ฌ๋Š” ํƒœ์•„์˜ ์ƒ๋ช…์„ ์šฐ์„ ์‹œํ•ด์•ผ ํ•œ๋‹ค๊ณ  ๋งํ•ฉ๋‹ˆ๋‹ค. ๋• ์œค๋ฆฌ ์ด๋ก ์€ ์˜์‚ฌ๊ฐ€ ๊ฐ€์žฅ ๊ฐ•ํ•œ ๋„๋•์  ์„ฑ๊ฒฉ์„ ๊ฐ€์ง„ ์‚ฌ๋žŒ์˜ ์ƒ๋ช…์„ ์šฐ์„ ์‹œํ•ด์•ผ ํ•˜๋ฉฐ, ์ด ๊ฒฝ์šฐ ์‚ฐ๋ชจ๊ฐ€ ๋  ๊ฒƒ์ด๋ผ๊ณ  ๋งํ•ฉ๋‹ˆ๋‹ค.
Analysis of examination domainThe findings from our experiments are illustrated in Figure 1, with extended experimental results accessible in Appendix A.2. Analysis across the models, as demonstrated by the horizontal examination of the rows in the figure, reveals that the performance trends for the Examination benchmarks, namely MMLLU, CMMLLU, and CEVAL, exhibit remarkable consistency. This uniformity across various evaluation datasets within the Examination domain suggests inherent similarities in the tasks.
์‹œํ—˜ ๋„๋ฉ”์ธ ๋ถ„์„ ์šฐ๋ฆฌ ์‹คํ—˜์˜ ๊ฒฐ๊ณผ๋Š” Figure 1์— ๋‚˜์™€ ์žˆ์œผ๋ฉฐ, ์ถ”๊ฐ€์ ์ธ ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” Appendix A.2์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ์˜ ํ–‰์„ ๊ฐ€๋กœ๋กœ ์‚ดํŽด๋ณธ ๋ชจ๋ธ ๊ฐ„์˜ ๋ถ„์„์€ MMLLU, CMMLLU, CEVAL๊ณผ ๊ฐ™์€ ์‹œํ—˜ ๋ฒค์น˜๋งˆํฌ์˜ ์„ฑ๋Šฅ ์ถ”์„ธ๊ฐ€ ๋†€๋ผ์šธ ์ •๋„๋กœ ์ผ๊ด€๋จ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์‹œํ—˜ ๋„๋ฉ”์ธ ๋‚ด์˜ ๋‹ค์–‘ํ•œ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ์…‹์—์„œ ์ด๋Ÿฌํ•œ ์ผ๊ด€์„ฑ์€ ๊ณผ์ œ์˜ ๋‚ด์žฌ๋œ ์œ ์‚ฌ์„ฑ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
์‹œํ—˜ ์˜์—ญ ๋ถ„์„์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 1์— ๋‚˜์™€ ์žˆ์œผ๋ฉฐ, ํ™•์žฅ๋œ ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ๋ถ€๋ก A.2์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ์˜ ํ–‰์„ ์ˆ˜ํ‰์œผ๋กœ ๊ฒ€ํ† ํ•˜์—ฌ ๋ชจ๋ธ ์ „์ฒด์— ๊ฑธ์นœ ๋ถ„์„ ๊ฒฐ๊ณผ, ์ฆ‰ ์‹œํ—˜ ๋ฒค์น˜๋งˆํฌ(MMLLU, CMMLLU, CEVAL)์˜ ์„ฑ๋Šฅ ์ถ”์„ธ๊ฐ€ ๋†€๋ผ์šด ์ผ๊ด€์„ฑ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹œํ—˜ ์˜์—ญ ๋‚ด ๋‹ค์–‘ํ•œ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ด๋Ÿฌํ•œ ํ†ต์ผ์„ฑ์€ ์ž‘์—…์˜ ๋ณธ์งˆ์ ์ธ ์œ ์‚ฌ์„ฑ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
FilteringFiltering aims to remove documents that fail to meet specified quality or relevance standards. This can be done through several approaches, such as establishing a minimum relevance score threshold to exclude documents below a certain relevance level. Furthermore, the use of feedback from users or prior relevance evaluations assists in adjusting the filtering process, guaranteeing that only the most relevant documents are retained for text generation (Khattab and Zaharia, 2020; Huang and Huang, 2023).
Filtering์€ ์ง€์ •๋œ ํ’ˆ์งˆ ๋˜๋Š” ๊ด€๋ จ์„ฑ ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฌธ์„œ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํŠน์ • ๊ด€๋ จ์„ฑ ์ˆ˜์ค€ ์ดํ•˜์˜ ๋ฌธ์„œ๋ฅผ ์ œ์™ธํ•˜๊ธฐ ์œ„ํ•ด ์ตœ์†Œ ๊ด€๋ จ์„ฑ ์ ์ˆ˜ ์ž„๊ณ„๊ฐ’์„ ์„ค์ •ํ•˜๋Š” ๋“ฑ์˜ ์—ฌ๋Ÿฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ์ˆ˜ํ–‰๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์ด๋‚˜ ์ด์ „์˜ ๊ด€๋ จ์„ฑ ํ‰๊ฐ€๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ•„ํ„ฐ๋ง ๊ณผ์ •์„ ์กฐ์ •ํ•จ์œผ๋กœ์จ ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ์œ„ํ•ด ๊ฐ€์žฅ ๊ด€๋ จ์„ฑ์ด ๋†’์€ ๋ฌธ์„œ๋งŒ ์œ ์ง€๋˜๋„๋ก ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค (Khattab and Zaharia, 2020; Huang and Huang, 2023).
ํ•„ํ„ฐ๋งํ•„ํ„ฐ๋ง์€ ์ง€์ •๋œ ํ’ˆ์งˆ ๋˜๋Š” ๊ด€๋ จ์„ฑ ํ‘œ์ค€์„ ์ถฉ์กฑํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฌธ์„œ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ตœ์†Œ ๊ด€๋ จ์„ฑ ์ ์ˆ˜ ์ž„๊ณ„๊ฐ’์„ ์„ค์ •ํ•˜์—ฌ ํŠน์ • ๊ด€๋ จ์„ฑ ์ˆ˜์ค€ ์ดํ•˜์˜ ๋ฌธ์„œ๋ฅผ ์ œ์™ธํ•˜๋Š” ๋“ฑ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์‚ฌ์šฉ์ž์˜ ํ”ผ๋“œ๋ฐฑ์ด๋‚˜ ์ด์ „ ๊ด€๋ จ์„ฑ ํ‰๊ฐ€๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ•„ํ„ฐ๋ง ํ”„๋กœ์„ธ์Šค๋ฅผ ์กฐ์ •ํ•˜์—ฌ ๊ฐ€์žฅ ๊ด€๋ จ์„ฑ์ด ๋†’์€ ๋ฌธ์„œ๋งŒ ํ…์ŠคํŠธ ์ƒ์„ฑ์— ์œ ์ง€๋˜๋„๋ก ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Khattab and Zaharia, 2020; Huang and Huang, 2023).
**Accuracy by number of turns.** Here we plotted the auto-evaluation of AMIE-generated differential diagnoses as a function of number of turns. We truncated conversations to the first \(T\) turns, and then asked AMIE to produce a DDx with this truncated conversation. For both the AMIE and PCP conversations, we observed that AMIE's average diagnostic accuracy began to plateau within 10 turns, with additional information gathering having diminishing returns on the diagnostic performance as shown in Figure A.12.
**ํ„ด ์ˆ˜์— ๋”ฐ๋ฅธ ์ •ํ™•๋„.** ์—ฌ๊ธฐ์—์„œ๋Š” AMIE๊ฐ€ ์ƒ์„ฑํ•œ ๊ฐ๋ณ„ ์ง„๋‹จ์˜ ์ž๋™ ํ‰๊ฐ€๋ฅผ ํ„ด ์ˆ˜์˜ ํ•จ์ˆ˜๋กœ ํ”Œ๋กœํŒ…ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ™”๋ฅผ ์ฒ˜์Œ \(T\) ํ„ด์œผ๋กœ ์ž˜๋ผ๋‚ธ ํ›„, ์ด ์ž˜๋ผ๋‚ธ ๋Œ€ํ™”๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ AMIE์—๊ฒŒ ๊ฐ๋ณ„ ์ง„๋‹จ(DDx)์„ ์ƒ์„ฑํ•˜๋„๋ก ์š”์ฒญํ–ˆ์Šต๋‹ˆ๋‹ค. AMIE์™€ PCP ๋Œ€ํ™” ๋ชจ๋‘์—์„œ, AMIE์˜ ํ‰๊ท  ์ง„๋‹จ ์ •ํ™•๋„๊ฐ€ 10ํ„ด ์ด๋‚ด์— ํ‰์ค€ํ™”๋˜๊ธฐ ์‹œ์ž‘ํ–ˆ์œผ๋ฉฐ, ์ถ”๊ฐ€์ ์ธ ์ •๋ณด ์ˆ˜์ง‘์ด ์ง„๋‹จ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์€ Figure A.12์—์„œ ๋ณด์ด๋Š” ๋ฐ”์™€ ๊ฐ™์ด ์ ์  ์ค„์–ด๋“œ๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.
**ํšŒ์ „ ์ˆ˜๋ณ„ ์ •ํ™•๋„.** ์—ฌ๊ธฐ์—์„œ๋Š” AMIE๊ฐ€ ์ƒ์„ฑํ•œ ๊ฐ๋ณ„ ์ง„๋‹จ์˜ ์ž๋™ ํ‰๊ฐ€๋ฅผ ํšŒ์ „ ์ˆ˜์˜ ํ•จ์ˆ˜๋กœ ํ”Œ๋กฏํ–ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ™”๋ฅผ ์ฒซ ๋ฒˆ์งธ \(T\) ํšŒ์ „์œผ๋กœ ์ž˜๋ผ๋‚ธ ๋‹ค์Œ AMIE์—๊ฒŒ ์ด ์ž˜๋ผ๋‚ธ ๋Œ€ํ™”๋กœ DDx๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ์š”์ฒญํ–ˆ์Šต๋‹ˆ๋‹ค. AMIE์™€ PCP ๋Œ€ํ™” ๋ชจ๋‘์—์„œ AMIE์˜ ํ‰๊ท  ์ง„๋‹จ ์ •ํ™•๋„๋Š” 10ํšŒ์ „ ์ด๋‚ด์— ์ •์ฒด๋˜๊ธฐ ์‹œ์ž‘ํ–ˆ์œผ๋ฉฐ, ๊ทธ๋ฆผ A.12์™€ ๊ฐ™์ด ์ถ”๊ฐ€ ์ •๋ณด ์ˆ˜์ง‘์€ ์ง„๋‹จ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ์ˆ˜์ต์ด ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
The Tree of Thought (ToT) [158] prompting technique is inspired by the concept of considering various alternative solutions or thought processes before converging on the most plausible one. ToT is based on the idea of branching out into multiple "thought trees" where each branch represents a different line of reasoning. This method allows the LLM to explore various possibilities and hypotheses, much like human cognitive processes where multiple scenarios are considered before determining the most likely one.
Tree of Thought (ToT) [158] ํ”„๋กฌํ”„ํŠธ ๊ธฐ๋ฒ•์€ ๊ฐ€์žฅ ๊ทธ๋Ÿด๋“ฏํ•œ ํ•ด๊ฒฐ์ฑ…์— ๋„๋‹ฌํ•˜๊ธฐ ์ „์— ๋‹ค์–‘ํ•œ ๋Œ€์•ˆ์  ํ•ด๊ฒฐ์ฑ…์ด๋‚˜ ์‚ฌ๊ณ  ๊ณผ์ •์„ ๊ณ ๋ คํ•˜๋Š” ๊ฐœ๋…์—์„œ ์˜๊ฐ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ToT๋Š” ๊ฐ ๊ฐ€์ง€๊ฐ€ ๋‹ค๋ฅธ ์ถ”๋ก  ๋ผ์ธ์„ ๋‚˜ํƒ€๋‚ด๋Š” ์—ฌ๋Ÿฌ "์‚ฌ๊ณ  ๋‚˜๋ฌด"๋กœ ๋ถ„๊ธฐํ•˜๋Š” ์•„์ด๋””์–ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ LLM์ด ์—ฌ๋Ÿฌ ๊ฐ€๋Šฅ์„ฑ๊ณผ ๊ฐ€์„ค์„ ํƒ๊ตฌํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•˜๋ฉฐ, ์ด๋Š” ์ธ๊ฐ„์˜ ์ธ์ง€ ๊ณผ์ •์—์„œ ์—ฌ๋Ÿฌ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๊ณ ๋ คํ•œ ํ›„ ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ๊ฒƒ์„ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
์ƒ๊ฐ์˜ ๋‚˜๋ฌด(ToT) [158] ํ”„๋กฌํ”„ํŠธ ๊ธฐ๋ฒ•์€ ๊ฐ€์žฅ ๊ทธ๋Ÿด๋“ฏํ•œ ํ•ด๊ฒฐ์ฑ…์— ๋„๋‹ฌํ•˜๊ธฐ ์ „์— ๋‹ค์–‘ํ•œ ๋Œ€์•ˆ ์†”๋ฃจ์…˜์ด๋‚˜ ์‚ฌ๊ณ  ๊ณผ์ •์„ ๊ณ ๋ คํ•˜๋Š” ๊ฐœ๋…์—์„œ ์˜๊ฐ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ToT๋Š” ๊ฐ ๊ฐ€์ง€๊ฐ€ ๋‹ค๋ฅธ ์ถ”๋ก  ๋ผ์ธ์„ ๋‚˜ํƒ€๋‚ด๋Š” ์—ฌ๋Ÿฌ "์ƒ๊ฐ์˜ ๋‚˜๋ฌด"๋กœ ๋ถ„๊ธฐ๋˜๋Š” ์•„์ด๋””์–ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด LLM์€ ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๊ฒฐ์ •ํ•˜๊ธฐ ์ „์— ์—ฌ๋Ÿฌ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๊ณ ๋ คํ•˜๋Š” ์ธ๊ฐ„์˜ ์ธ์ง€ ๊ณผ์ •๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๋‹ค์–‘ํ•œ ๊ฐ€๋Šฅ์„ฑ๊ณผ ๊ฐ€์„ค์„ ํƒ์ƒ‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In Table 1, we observe that the model's prior response only agreed with the reference answer 34.7% on average. However, the RAG answers elevated the concordance to 94%. This result demonstrates that the RAG pipeline established in this work is highly effective at encouraging the model to adhere to its retrieved content. However, in the minority of cases where providing the retrieved content fails to correct the LLM, we find that the model simply responds with its original prior answer about 20% of the time.
Table 1์—์„œ, ๋ชจ๋ธ์˜ ์ด์ „ ์‘๋‹ต์ด ์ฐธ์กฐ ๋‹ต๋ณ€๊ณผ ํ‰๊ท  34.7%๋งŒ ์ผ์น˜ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ RAG ๋‹ต๋ณ€์€ ์ผ์น˜์œจ์„ 94%๋กœ ๋†’์˜€์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ๋ณธ ์—ฐ๊ตฌ์—์„œ ํ™•๋ฆฝ๋œ RAG ํŒŒ์ดํ”„๋ผ์ธ์ด ๋ชจ๋ธ์ด ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ๋ฅผ ๋”ฐ๋ฅด๋„๋ก ์œ ๋„ํ•˜๋Š” ๋ฐ ๋งค์šฐ ํšจ๊ณผ์ ์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ๋ฅผ ์ œ๊ณตํ•ด๋„ LLM์„ ์ˆ˜์ •ํ•˜์ง€ ๋ชปํ•˜๋Š” ์†Œ์ˆ˜์˜ ๊ฒฝ์šฐ์—๋Š”, ๋ชจ๋ธ์ด ์•ฝ 20%์˜ ํ™•๋ฅ ๋กœ ์›๋ž˜์˜ ์ด์ „ ๋‹ต๋ณ€์„ ๊ทธ๋Œ€๋กœ ์‘๋‹ตํ•˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
ํ‘œ 1์—์„œ ๋ชจ๋ธ์˜ ์‚ฌ์ „ ์‘๋‹ต์ด ์ฐธ์กฐ ๋‹ต๋ณ€๊ณผ ์ผ์น˜ํ•˜๋Š” ๋น„์œจ์€ ํ‰๊ท  34.7%์— ๋ถˆ๊ณผํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ RAG ๋‹ต๋ณ€์€ ์ผ์น˜๋„๋ฅผ 94%๊นŒ์ง€ ๋†’์˜€์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ์ด ์ž‘์—…์—์„œ ์ˆ˜๋ฆฝํ•œ RAG ํŒŒ์ดํ”„๋ผ์ธ์ด ๋ชจ๋ธ์ด ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ๋ฅผ ์ค€์ˆ˜ํ•˜๋„๋ก ์œ ๋„ํ•˜๋Š” ๋ฐ ๋งค์šฐ ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ๋ฅผ ์ œ๊ณตํ•ด๋„ LLM์„ ์ˆ˜์ •ํ•˜์ง€ ๋ชปํ•˜๋Š” ์†Œ์ˆ˜์˜ ๊ฒฝ์šฐ, ๋ชจ๋ธ์ด ๋‹จ์ˆœํžˆ ์›๋ž˜์˜ ์‚ฌ์ „ ๋‹ต๋ณ€์œผ๋กœ ์‘๋‹ตํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ์•ฝ 20%์— ๋‹ฌํ–ˆ์Šต๋‹ˆ๋‹ค.
**Embodied LLM Agents Learn to Cooperate in Organized Teams.** Research by Guo et al. demonstrates the impact of a lead agent on the overall effectiveness of the agent team [9]. This architecture contains a vertical component through the leader agent, as well as a horizontal component from the ability for agents to converse with other agents besides the leader. The results of their study demonstrate that agent teams with an organized leader complete their tasks nearly 10% faster than teams without a leader.
**๊ตฌ์ฒดํ™”๋œ LLM ์—์ด์ „ํŠธ๋Š” ์กฐ์ง๋œ ํŒ€์—์„œ ํ˜‘๋ ฅํ•˜๋Š” ๋ฒ•์„ ๋ฐฐ์›๋‹ˆ๋‹ค.** Guo ๋“ฑ ์—ฐ๊ตฌ์ง„์˜ ์—ฐ๊ตฌ๋Š” ๋ฆฌ๋“œ ์—์ด์ „ํŠธ๊ฐ€ ์—์ด์ „ํŠธ ํŒ€์˜ ์ „์ฒด ํšจ์œจ์„ฑ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค [9]. ์ด ์•„ํ‚คํ…์ฒ˜๋Š” ๋ฆฌ๋” ์—์ด์ „ํŠธ๋ฅผ ํ†ตํ•œ ์ˆ˜์ง์  ๊ตฌ์„ฑ ์š”์†Œ์™€ ๋ฆฌ๋” ์™ธ์˜ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์™€ ๋Œ€ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์—์„œ ์˜ค๋Š” ์ˆ˜ํ‰์  ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด, ์กฐ์ง๋œ ๋ฆฌ๋”๊ฐ€ ์žˆ๋Š” ์—์ด์ „ํŠธ ํŒ€์€ ๋ฆฌ๋”๊ฐ€ ์—†๋Š” ํŒ€๋ณด๋‹ค ์ž‘์—…์„ ๊ฑฐ์˜ 10% ๋” ๋น ๋ฅด๊ฒŒ ์™„๋ฃŒํ•ฉ๋‹ˆ๋‹ค.
**์ฒดํ™”๋œ LLM ์—์ด์ „ํŠธ๋Š” ์กฐ์งํ™”๋œ ํŒ€์—์„œ ํ˜‘๋ ฅํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค.** Guo ๋“ฑ์˜ ์—ฐ๊ตฌ๋Š” ๋ฆฌ๋“œ ์—์ด์ „ํŠธ๊ฐ€ ์—์ด์ „ํŠธ ํŒ€์˜ ์ „๋ฐ˜์ ์ธ ํšจ์œจ์„ฑ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค[9]. ์ด ์•„ํ‚คํ…์ฒ˜์—๋Š” ๋ฆฌ๋” ์—์ด์ „ํŠธ๋ฅผ ํ†ตํ•œ ์ˆ˜์ง์  ๊ตฌ์„ฑ ์š”์†Œ์™€ ์—์ด์ „ํŠธ๊ฐ€ ๋ฆฌ๋” ์™ธ์— ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์™€ ๋Œ€ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์ˆ˜ํ‰์  ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์˜ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด ์กฐ์ง์ ์ธ ๋ฆฌ๋”๊ฐ€ ์žˆ๋Š” ์—์ด์ „ํŠธ ํŒ€์€ ๋ฆฌ๋”๊ฐ€ ์—†๋Š” ํŒ€๋ณด๋‹ค ๊ฑฐ์˜ 10% ๋” ๋นจ๋ฆฌ ์ž‘์—…์„ ์™„๋ฃŒํ•ฉ๋‹ˆ๋‹ค.
The tool manager in the AIOS system manages a diverse array of API tools that enhance the functionality of LLMs. As shown in Table 1, the tool manager integrates commonly-used tools from various sources [7; 62; 63] and classify them into different categories, which covers web search, scientific computing, database retrieval, image processing, etc. In this way, the managed tools can cover different modalities of input and output (image and text), thus facilitating agent development within the AIOS ecosystem.
AIOS ์‹œ์Šคํ…œ์˜ ๋„๊ตฌ ๊ด€๋ฆฌ์ž๋Š” LLM์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋‹ค์–‘ํ•œ API ๋„๊ตฌ๋“ค์„ ๊ด€๋ฆฌํ•ฉ๋‹ˆ๋‹ค. Table 1์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ๋„๊ตฌ ๊ด€๋ฆฌ์ž๋Š” ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜ [7; 62; 63]์—์„œ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋„๊ตฌ๋“ค์„ ํ†ตํ•ฉํ•˜๊ณ  ์ด๋ฅผ ์›น ๊ฒ€์ƒ‰, ๊ณผํ•™ ๊ณ„์‚ฐ, ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค ๊ฒ€์ƒ‰, ์ด๋ฏธ์ง€ ์ฒ˜๋ฆฌ ๋“ฑ ๋‹ค์–‘ํ•œ ์นดํ…Œ๊ณ ๋ฆฌ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ๊ด€๋ฆฌ๋œ ๋„๊ตฌ๋“ค์€ ์ž…๋ ฅ๊ณผ ์ถœ๋ ฅ์˜ ๋‹ค์–‘ํ•œ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ(์ด๋ฏธ์ง€์™€ ํ…์ŠคํŠธ)๋ฅผ ๋‹ค๋ฃฐ ์ˆ˜ ์žˆ์–ด, AIOS ์ƒํƒœ๊ณ„ ๋‚ด์—์„œ ์—์ด์ „ํŠธ ๊ฐœ๋ฐœ์„ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค.
AIOS ์‹œ์Šคํ…œ์˜ ๋„๊ตฌ ๊ด€๋ฆฌ์ž๋Š” LLM์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋‹ค์–‘ํ•œ API ๋„๊ตฌ๋ฅผ ๊ด€๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 1๊ณผ ๊ฐ™์ด ๋„๊ตฌ ๊ด€๋ฆฌ์ž๋Š” ๋‹ค์–‘ํ•œ ์†Œ์Šค์—์„œ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋„๊ตฌ๋ฅผ ํ†ตํ•ฉํ•˜๊ณ [7; 62; 63] ์›น ๊ฒ€์ƒ‰, ๊ณผํ•™์  ์ปดํ“จํŒ…, ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค ๊ฒ€์ƒ‰, ์ด๋ฏธ์ง€ ์ฒ˜๋ฆฌ ๋“ฑ์„ ํฌํ•จํ•˜๋Š” ๋‹ค์–‘ํ•œ ์นดํ…Œ๊ณ ๋ฆฌ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ๊ด€๋ฆฌ๋˜๋Š” ๋„๊ตฌ๋Š” ๋‹ค์–‘ํ•œ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ๋ฐฉ์‹(์ด๋ฏธ์ง€ ๋ฐ ํ…์ŠคํŠธ)์„ ์ง€์›ํ•˜๋ฏ€๋กœ AIOS ์—์ฝ”์‹œ์Šคํ…œ ๋‚ด์—์„œ ์—์ด์ „ํŠธ ๊ฐœ๋ฐœ์„ ์šฉ์ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.
hence the shortest path is an upper bound on \(\mathsf{logit}_{i}\). To prove feasibility, we claim that setting \(\mathsf{logit}_{i}\) to be the distance from \(0\) to \(i\) satisfies all the inequalities. Assume some inequality \(\mathsf{logit}_{i}-\mathsf{logit}_{j}\leq c_{ji}\) is violated. Then we can go from \(0\to j\to i\) in \(G\) with a total weight of \(\mathsf{logit}_{j}+c_{ji}<\mathsf{logit}_{i}\), which contradicts the assumption that \(\mathsf{logit}_{i}\) is the distance from \(0\) to \(i\).
๋”ฐ๋ผ์„œ ์ตœ๋‹จ ๊ฒฝ๋กœ๋Š” \(\mathsf{logit}_{i}\)์˜ ์ƒํ•œ์ž…๋‹ˆ๋‹ค. ์‹คํ˜„ ๊ฐ€๋Šฅ์„ฑ์„ ์ฆ๋ช…ํ•˜๊ธฐ ์œ„ํ•ด, \(\mathsf{logit}_{i}\)๋ฅผ \(0\)์—์„œ \(i\)๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ๋กœ ์„ค์ •ํ•˜๋ฉด ๋ชจ๋“  ๋ถ€๋“ฑ์‹์„ ๋งŒ์กฑํ•œ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ์–ด๋–ค ๋ถ€๋“ฑ์‹ \(\mathsf{logit}_{i}-\mathsf{logit}_{j}\leq c_{ji}\)๊ฐ€ ์œ„๋ฐ˜๋œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ด…์‹œ๋‹ค. ๊ทธ๋Ÿฌ๋ฉด ์šฐ๋ฆฌ๋Š” \(G\)์—์„œ \(0\to j\to i\)๋กœ ์ด๋™ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด ๊ฐ€์ค‘์น˜๋Š” \(\mathsf{logit}_{j}+c_{ji}<\mathsf{logit}_{i}\)๊ฐ€ ๋˜์–ด \(\mathsf{logit}_{i}\)๊ฐ€ \(0\)์—์„œ \(i\)๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ๋ผ๋Š” ๊ฐ€์ •์— ๋ชจ์ˆœ์ด ์ƒ๊น๋‹ˆ๋‹ค.
๋”ฐ๋ผ์„œ ์ตœ๋‹จ ๊ฒฝ๋กœ๋Š” \(\mathsf{logit}_{i}\)์˜ ์ƒํ•œ์ž…๋‹ˆ๋‹ค. ์‹คํ˜„ ๊ฐ€๋Šฅ์„ฑ์„ ์ฆ๋ช…ํ•˜๊ธฐ ์œ„ํ•ด \(\mathsf{logit}_{i}\)๋ฅผ \(0\)์—์„œ \(i\)๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ๋กœ ์„ค์ •ํ•˜๋ฉด ๋ชจ๋“  ๋ถ€๋“ฑ์‹์„ ๋งŒ์กฑํ•œ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ์–ด๋–ค ๋ถ€๋“ฑ์‹ \(\mathsf{logit}_{i}-\mathsf{logit}_{j}\leq c_{ji}\)๊ฐ€ ์œ„๋ฐ˜๋˜์—ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋ฉด \(\mathsf{logit}_{j}+c_{ji}<\mathsf{logit}_{i}\)์˜ ์ด ๊ฐ€์ค‘์น˜๋กœ \(G\)์—์„œ \(0\to j\to i\)๋กœ ๊ฐˆ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” \(\mathsf{logit}_{i}\)๊ฐ€ \(0\)์—์„œ \(i\)๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ๋ผ๋Š” ๊ฐ€์ •๊ณผ ๋ชจ์ˆœ๋ฉ๋‹ˆ๋‹ค.
In contrast to earlier PLMs, LLMs which consist of a significantly larger number of parameters require a higher volume of training data that covers a broad range of content. For this need, there are increasingly more accessible training datasets that have been released for research. In this section, we will briefly summarize several widely used corpora for training LLMs. Based on their content types, we categorize these corpora into six groups: Books, CommonCrawl, Reddit links, Wikipedia, Code, and others.
์ด์ „์˜ PLM๊ณผ ๋‹ฌ๋ฆฌ, ํ›จ์”ฌ ๋” ๋งŽ์€ ์ˆ˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ LLM์€ ๊ด‘๋ฒ”์œ„ํ•œ ๋‚ด์šฉ์„ ๋‹ค๋ฃจ๋Š” ๋” ๋งŽ์€ ์–‘์˜ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ํ•„์š”๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•„์š”๋ฅผ ์ถฉ์กฑ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•ด ๊ณต๊ฐœ๋œ ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹์ด ์ ์  ๋” ๋งŽ์•„์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” LLM ํ›ˆ๋ จ์— ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์—ฌ๋Ÿฌ ์ฝ”ํผ์Šค๋ฅผ ๊ฐ„๋žตํžˆ ์š”์•ฝํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ๋‚ด์šฉ ์œ ํ˜•์— ๋”ฐ๋ผ ์ด๋“ค ์ฝ”ํผ์Šค๋ฅผ ์ฑ…, CommonCrawl, Reddit ๋งํฌ, Wikipedia, ์ฝ”๋“œ, ๊ธฐํƒ€ ๋“ฑ 6๊ฐœ์˜ ๊ทธ๋ฃน์œผ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค.
ํ›จ์”ฌ ๋” ๋งŽ์€ ์ˆ˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ LLM์€ ์ด์ „ PLM๊ณผ ๋‹ฌ๋ฆฌ ๊ด‘๋ฒ”์œ„ํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ํฌ๊ด„ํ•˜๋Š” ๋” ๋งŽ์€ ์–‘์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ํ•„์š”๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•„์š”์„ฑ์— ๋”ฐ๋ผ ์—ฐ๊ตฌ์šฉ์œผ๋กœ ๊ณต๊ฐœ๋œ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ ์  ๋” ๋งŽ์•„์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” LLM ํ•™์Šต์— ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋ช‡ ๊ฐ€์ง€ ์ฝ”ํผ์Šค๋ฅผ ๊ฐ„๋žตํ•˜๊ฒŒ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์ฝ˜ํ…์ธ  ์œ ํ˜•์— ๋”ฐ๋ผ ์ด๋Ÿฌํ•œ ์ฝ”ํผ์Šค๋ฅผ ๋„์„œ, CommonCrawl, Reddit ๋งํฌ, Wikipedia, ์ฝ”๋“œ ๋ฐ ๊ธฐํƒ€์˜ 6๊ฐ€์ง€ ๊ทธ๋ฃน์œผ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค.
This paper present a survey of LLMs developed in the past few years. We first provide an overview of early pre-trained language models (e.g., as BERT), then review three popular LLM families (GPT, LLaMA, PaLM), and other representative LLMs. We then survey methods and techniques of building, augmenting, and using LLMs. We review popular LLM datasets and benchmarks, and compare performance of a set of prominent models on public benchmarks. Finally, we present open challenges and future research directions.
์ด ๋…ผ๋ฌธ์€ ์ง€๋‚œ ๋ช‡ ๋…„๊ฐ„ ๊ฐœ๋ฐœ๋œ LLMs์— ๋Œ€ํ•œ ์„ค๋ฌธ ์กฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์ดˆ๊ธฐ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ์–ธ์–ด ๋ชจ๋ธ๋“ค(e.g., BERT)์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•œ ํ›„, ์„ธ ๊ฐ€์ง€ ์ธ๊ธฐ ์žˆ๋Š” LLM ๊ณ„์—ด(GPT, LLaMA, PaLM)๊ณผ ๊ธฐํƒ€ ๋Œ€ํ‘œ์ ์ธ LLM๋“ค์„ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๋‹ค์Œ์œผ๋กœ LLMs๋ฅผ ๊ตฌ์ถ•, ์ฆ๊ฐ• ๋ฐ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•๊ณผ ๊ธฐ์ˆ ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ธ๊ธฐ ์žˆ๋Š” LLM ๋ฐ์ดํ„ฐ์…‹๊ณผ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ฒ€ํ† ํ•˜๊ณ , ๊ณต๊ณต ๋ฒค์น˜๋งˆํฌ์—์„œ ๋‘๋“œ๋Ÿฌ์ง„ ๋ชจ๋ธ๋“ค์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ํ•ด๊ฒฐ๋˜์ง€ ์•Š์€ ๊ณผ์ œ์™€ ๋ฏธ๋ž˜ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ง€๋‚œ ๋ช‡ ๋…„ ๋™์•ˆ ๊ฐœ๋ฐœ๋œ LLM์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์ดˆ๊ธฐ ์‚ฌ์ „ ํ•™์Šต ์–ธ์–ด ๋ชจ๋ธ(์˜ˆ: BERT)์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•œ ๋‹ค์Œ, ์ธ๊ธฐ ์žˆ๋Š” LLM ์ œํ’ˆ๊ตฐ 3๊ฐ€์ง€(GPT, LLaMA, PaLM)์™€ ๊ธฐํƒ€ ๋Œ€ํ‘œ์ ์ธ LLM์„ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ LLM์„ ๊ตฌ์ถ•, ์ฆ๊ฐ• ๋ฐ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•๊ณผ ๊ธฐ์ˆ ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ์ธ๊ธฐ ์žˆ๋Š” LLM ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ฒ€ํ† ํ•˜๊ณ  ๊ณต๊ฐœ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ผ๋ จ์˜ ์ฃผ์š” ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๊ฐœ๋ฐฉํ˜• ๊ณผ์ œ์™€ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
However, as impactful as LLMs have been, their implementation comes with inherent limitations that hamper transparent and open research. A major concern is their proprietary nature, which restricts access to the models, thus inhibiting the broader research community's ability to build upon their successes. Furthermore, the vast computational resources necessary for training and deploying these models present a challenge for researchers with limited resources, further compounding the accessibility problem.
๊ทธ๋Ÿฌ๋‚˜ LLM์ด ํฐ ์˜ํ–ฅ์„ ๋ฏธ์ณค์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ๊ทธ ๊ตฌํ˜„์—๋Š” ํˆฌ๋ช…ํ•˜๊ณ  ๊ฐœ๋ฐฉ์ ์ธ ์—ฐ๊ตฌ๋ฅผ ๋ฐฉํ•ดํ•˜๋Š” ๊ณ ์œ ํ•œ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์š” ๋ฌธ์ œ๋Š” ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ ‘๊ทผ์„ ์ œํ•œํ•˜๋Š” ๋…์ ์ ์ธ ํŠน์„ฑ์œผ๋กœ, ์ด๋Š” ๋” ๋„“์€ ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ๊ฐ€ ๊ทธ ์„ฑ๊ณต์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฐœ์ „ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ์ €ํ•ดํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ๋ฐฐํฌํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ๋ฐฉ๋Œ€ํ•œ ๊ณ„์‚ฐ ์ž์›์€ ์ž์›์ด ์ œํ•œ๋œ ์—ฐ๊ตฌ์ž๋“ค์—๊ฒŒ ๋„์ „ ๊ณผ์ œ๋ฅผ ์ œ์‹œํ•˜๋ฉฐ, ์ ‘๊ทผ์„ฑ ๋ฌธ์ œ๋ฅผ ๋”์šฑ ์•…ํ™”์‹œํ‚ต๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ LLM์ด ํฐ ์˜ํ–ฅ์„ ๋ฏธ์ณค์ง€๋งŒ, ๊ทธ ๊ตฌํ˜„์—๋Š” ํˆฌ๋ช…ํ•˜๊ณ  ๊ฐœ๋ฐฉ์ ์ธ ์—ฐ๊ตฌ๋ฅผ ๋ฐฉํ•ดํ•˜๋Š” ๋‚ด์žฌ์ ์ธ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์š” ์šฐ๋ ค ์‚ฌํ•ญ์€ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์•ก์„ธ์Šค๋ฅผ ์ œํ•œํ•˜์—ฌ ๋” ๋„“์€ ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ๊ฐ€ ์„ฑ๊ณต์„ ์Œ“์•„ ์˜ฌ๋ฆฌ๋Š” ๋Šฅ๋ ฅ์„ ์ €ํ•ดํ•˜๋Š” ๋…์ ์ ์ธ ํŠน์„ฑ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ๋ฐฐํฌํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ๋ฐฉ๋Œ€ํ•œ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค๋Š” ๋ฆฌ์†Œ์Šค๊ฐ€ ์ œํ•œ๋œ ์—ฐ๊ตฌ์ž๋“ค์—๊ฒŒ ๋„์ „์ด ๋˜๋ฉฐ ์ ‘๊ทผ์„ฑ ๋ฌธ์ œ๋ฅผ ๋”์šฑ ๊ฐ€์ค‘์‹œํ‚ต๋‹ˆ๋‹ค.
1. Rephrased and Expanded Question: Considering the current geopolitical tensions and recent historical events, is it likely that the State of Israel will engage in a military operation that results in the death of at least five Iranian nationals within Iran's borders, and subsequently either publicly claim responsibility for this action or execute the operation in such a clear and overt manner that credible sources can unambiguously attribute the attack to Israel, all before the end of the year 2023? 2.
1. ์žฌ๊ตฌ์„ฑ๋˜๊ณ  ํ™•์žฅ๋œ ์งˆ๋ฌธ: ํ˜„์žฌ์˜ ์ง€์ •ํ•™์  ๊ธด์žฅ๊ณผ ์ตœ๊ทผ์˜ ์—ญ์‚ฌ์  ์‚ฌ๊ฑด๋“ค์„ ๊ณ ๋ คํ•  ๋•Œ, ์ด์Šค๋ผ์—˜ ๊ตญ๊ฐ€๊ฐ€ 2023๋…„ ๋ง ์ด์ „์— ์ด๋ž€ ๊ตญ๊ฒฝ ๋‚ด์—์„œ ์ตœ์†Œ ๋‹ค์„ฏ ๋ช…์˜ ์ด๋ž€ ๊ตญ๋ฏผ์˜ ์‚ฌ๋ง์„ ์ดˆ๋ž˜ํ•˜๋Š” ๊ตฐ์‚ฌ ์ž‘์ „์„ ์ˆ˜ํ–‰ํ•˜๊ณ , ์ดํ›„ ์ด ํ–‰๋™์— ๋Œ€ํ•ด ๊ณต๊ฐœ์ ์œผ๋กœ ์ฑ…์ž„์„ ์ฃผ์žฅํ•˜๊ฑฐ๋‚˜ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์ถœ์ฒ˜๊ฐ€ ์ด ๊ณต๊ฒฉ์„ ๋ช…ํ™•ํ•˜๊ฒŒ ์ด์Šค๋ผ์—˜์˜ ์†Œํ–‰์œผ๋กœ ๊ท€์†์‹œํ‚ฌ ์ˆ˜ ์žˆ์„ ๋งŒํผ ๋ช…๋ฐฑํ•˜๊ณ  ๊ณต๊ฐœ์ ์œผ๋กœ ์ž‘์ „์„ ์ˆ˜ํ–‰ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ๋Š”๊ฐ€?
1. ์งˆ๋ฌธ์„ ๋ฐ”๊พธ๊ณ  ํ™•์žฅํ•˜๊ธฐ: ํ˜„์žฌ์˜ ์ง€์ •ํ•™์  ๊ธด์žฅ๊ณผ ์ตœ๊ทผ์˜ ์—ญ์‚ฌ์  ์‚ฌ๊ฑด์„ ๊ณ ๋ คํ•  ๋•Œ, ์ด์Šค๋ผ์—˜์ด ์ด๋ž€ ๊ตญ๊ฒฝ ๋‚ด์—์„œ ์ด๋ž€ ๊ตญ๋ฏผ 5๋ช… ์ด์ƒ์ด ์‚ฌ๋งํ•˜๋Š” ๊ตฐ์‚ฌ ์ž‘์ „์— ์ฐธ์—ฌํ•œ ํ›„ ์ด ํ–‰๋™์— ๋Œ€ํ•œ ๊ณต์  ์ฑ…์ž„์„ ์ฃผ์žฅํ•˜๊ฑฐ๋‚˜ ์ด ๊ณต๊ฒฉ์„ ์ด์Šค๋ผ์—˜์— ๋ช…๋ฐฑํ•˜๊ฒŒ ๊ธฐ์ธํ•  ์ˆ˜ ์žˆ๋Š” ๋ช…ํ™•ํ•˜๊ณ  ๊ณต๊ฐœ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์ž‘์ „์„ ์ˆ˜ํ–‰ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ๋Š”์ง€, 2023๋…„ ๋ง๊นŒ์ง€ ์ด์Šค๋ผ์—˜์ด ์ด๋Ÿฌํ•œ ํ–‰๋™์„ ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ๋Š”๊ฐ€? 2.
Previously Seen KnowledgeOne important distinction to make is between knowledge that the model has been exposed to before during pre-training as opposed to entirely new facts. Considering the size of modern LLM training sets, they cover a vast amount of information available through web-sourced text. As a result, even in niche domains, the goal of knowledge injection is not necessarily to teach the model entirely new facts but rather to "refresh" its memory by inducing a bias toward a particular domain.
์ด์ „์— ์ ‘ํ•œ ์ง€์‹ ๋ชจ๋ธ์ด ์‚ฌ์ „ ํ›ˆ๋ จ ์ค‘์— ์ด๋ฏธ ๋…ธ์ถœ๋œ ์ง€์‹๊ณผ ์™„์ „ํžˆ ์ƒˆ๋กœ์šด ์‚ฌ์‹ค์„ ๊ตฌ๋ถ„ํ•˜๋Š” ๊ฒƒ์€ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ํ˜„๋Œ€ LLM ํ›ˆ๋ จ ์„ธํŠธ์˜ ํฌ๊ธฐ๋ฅผ ๊ณ ๋ คํ•  ๋•Œ, ์ด๋“ค์€ ์›น์—์„œ ์ˆ˜์ง‘๋œ ํ…์ŠคํŠธ๋ฅผ ํ†ตํ•ด ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์ •๋ณด๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ์‹ฌ์ง€์–ด ํ‹ˆ์ƒˆ ๋„๋ฉ”์ธ์—์„œ๋„ ์ง€์‹ ์ฃผ์ž…์˜ ๋ชฉํ‘œ๋Š” ๋ฐ˜๋“œ์‹œ ๋ชจ๋ธ์—๊ฒŒ ์™„์ „ํžˆ ์ƒˆ๋กœ์šด ์‚ฌ์‹ค์„ ๊ฐ€๋ฅด์น˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ํŠน์ • ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ํŽธํ–ฅ์„ ์œ ๋„ํ•˜์—ฌ ๋ชจ๋ธ์˜ ๊ธฐ์–ต์„ "์ƒˆ๋กญ๊ฒŒ" ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
์ด์ „์— ๋ณธ ์ง€์‹ํ•œ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ตฌ๋ถ„์€ ๋ชจ๋ธ์ด ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ์ด์ „์— ๋…ธ์ถœ๋œ ์ ์ด ์žˆ๋Š” ์ง€์‹๊ณผ ์™„์ „ํžˆ ์ƒˆ๋กœ์šด ์‚ฌ์‹ค์„ ๊ตฌ๋ถ„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ตœ์‹  LLM ํ•™์Šต ์„ธํŠธ์˜ ๊ทœ๋ชจ๋ฅผ ๊ณ ๋ คํ•  ๋•Œ ์›น์—์„œ ์†Œ์‹ฑํ•œ ํ…์ŠคํŠธ๋ฅผ ํ†ตํ•ด ์ด์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์ •๋ณด๋ฅผ ๋‹ค๋ฃจ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ‹ˆ์ƒˆ ๋„๋ฉ”์ธ์—์„œ๋„ ์ง€์‹ ์ฃผ์ž…์˜ ๋ชฉํ‘œ๋Š” ๋ฐ˜๋“œ์‹œ ๋ชจ๋ธ์— ์™„์ „ํžˆ ์ƒˆ๋กœ์šด ์‚ฌ์‹ค์„ ๊ฐ€๋ฅด์น˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ํŠน์ • ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ํŽธํ–ฅ์„ ์œ ๋„ํ•˜์—ฌ ๋ชจ๋ธ์˜ ๊ธฐ์–ต์„ "๊ฐฑ์‹ "ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
We find that successful goal execution by agents is contingent upon proper planning and self-correction [32; 16; 23; 1]. Without the ability to self-evaluate and create effective plans, single agents may get stuck in an endless execution loop and never accomplish a given task or return a result that does not meet user expectations [32]. We find that single agent architectures are especially useful when the task requires straightforward function calling and does not need feedback from another agent [22].
์šฐ๋ฆฌ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ชฉํ‘œ๋ฅผ ์„ฑ๊ณต์ ์œผ๋กœ ์‹คํ–‰ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ ์ ˆํ•œ ๊ณ„ํš๊ณผ ์ž๊ธฐ ์ˆ˜์ •์ด ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค [32; 16; 23; 1]. ์ž๊ธฐ ํ‰๊ฐ€์™€ ํšจ๊ณผ์ ์ธ ๊ณ„ํš์„ ์„ธ์šฐ๋Š” ๋Šฅ๋ ฅ์ด ์—†์œผ๋ฉด, ๋‹จ์ผ ์—์ด์ „ํŠธ๋Š” ๋์—†๋Š” ์‹คํ–‰ ๋ฃจํ”„์— ๋น ์ ธ ์ฃผ์–ด์ง„ ์ž‘์—…์„ ์™„๋ฃŒํ•˜์ง€ ๋ชปํ•˜๊ฑฐ๋‚˜ ์‚ฌ์šฉ์ž ๊ธฐ๋Œ€์— ๋ถ€ํ•ฉํ•˜์ง€ ์•Š๋Š” ๊ฒฐ๊ณผ๋ฅผ ๋ฐ˜ํ™˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [32]. ๋‹จ์ผ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๋Š” ์ž‘์—…์ด ๊ฐ„๋‹จํ•œ ํ•จ์ˆ˜ ํ˜ธ์ถœ์„ ์š”๊ตฌํ•˜๊ณ  ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์˜ ํ”ผ๋“œ๋ฐฑ์ด ํ•„์š”ํ•˜์ง€ ์•Š์„ ๋•Œ ํŠนํžˆ ์œ ์šฉํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค [22].
์—์ด์ „ํŠธ์˜ ์„ฑ๊ณต์ ์ธ ๋ชฉํ‘œ ์‹คํ–‰์€ ์ ์ ˆํ•œ ๊ณ„ํš๊ณผ ์ž๊ธฐ ์ˆ˜์ •์— ๋‹ฌ๋ ค ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[32; 16; 23; 1]. ์ž๊ธฐ ํ‰๊ฐ€์™€ ํšจ๊ณผ์ ์ธ ๊ณ„ํš์„ ์„ธ์šฐ๋Š” ๋Šฅ๋ ฅ์ด ์—†์œผ๋ฉด ๋‹จ์ผ ์—์ด์ „ํŠธ๋Š” ๋์—†๋Š” ์‹คํ–‰ ๋ฃจํ”„์— ๊ฐ‡ํ˜€ ์ฃผ์–ด์ง„ ์ž‘์—…์„ ์™„๋ฃŒํ•˜์ง€ ๋ชปํ•˜๊ฑฐ๋‚˜ ์‚ฌ์šฉ์ž์˜ ๊ธฐ๋Œ€์— ๋ถ€์‘ํ•˜์ง€ ์•Š๋Š” ๊ฒฐ๊ณผ๋ฅผ ๋ฐ˜ํ™˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[32]. ๋‹จ์ผ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๋Š” ์ง๊ด€์ ์ธ ํ•จ์ˆ˜ ํ˜ธ์ถœ์ด ํ•„์š”ํ•˜๊ณ  ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์˜ ํ”ผ๋“œ๋ฐฑ์ด ํ•„์š”ํ•˜์ง€ ์•Š์€ ์ž‘์—…์— ํŠนํžˆ ์œ ์šฉํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[22].
**GPT-4 Judgement:** The solution incorrectly assumes that there are exactly 31 days in each of the 2 months comprising the 60 days, which is not necessarily true. However, it does not affect the final calculation as they incorrectly calculate the total servings as 31*2=62 servings in 60 days, which again is not correct but coincidentally leads to the correct number of cartons needed since 60 servings would also require 4 cartons. The final multiplication step is correct and gives the right total cost.
GPT-4 ํŒ๋‹จ: ์ด ์†”๋ฃจ์…˜์€ 60์ผ์„ ๊ตฌ์„ฑํ•˜๋Š” ๋‘ ๋‹ฌ ๊ฐ๊ฐ์— ์ •ํ™•ํžˆ 31์ผ์ด ์žˆ๋‹ค๊ณ  ์ž˜๋ชป ๊ฐ€์ •ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ฐ˜๋“œ์‹œ ์‚ฌ์‹ค์ด ์•„๋‹™๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋“ค์€ 60์ผ ๋™์•ˆ ์ด ์„œ๋น™ ์ˆ˜๋ฅผ 31*2=62 ์„œ๋น™์œผ๋กœ ์ž˜๋ชป ๊ณ„์‚ฐํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ตœ์ข… ๊ณ„์‚ฐ์—๋Š” ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋‹ค์‹œ ๋งํ•ด ์ •ํ™•ํ•˜์ง€ ์•Š์ง€๋งŒ ์šฐ์—ฐํžˆ๋„ ํ•„์š”ํ•œ ์นดํ†ค ์ˆ˜๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ๋„์ถœํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. 60 ์„œ๋น™๋„ 4๊ฐœ์˜ ์นดํ†ค์ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ตœ์ข… ๊ณฑ์…ˆ ๋‹จ๊ณ„๋Š” ์ •ํ™•ํ•˜๋ฉฐ ์˜ฌ๋ฐ”๋ฅธ ์ด ๋น„์šฉ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
**GPT-4 ํŒ๋‹จ:** ์ด ์†”๋ฃจ์…˜์€ 60์ผ์„ ๊ตฌ์„ฑํ•˜๋Š” ๋‘ ๋‹ฌ ๋™์•ˆ ๊ฐ๊ฐ ์ •ํ™•ํžˆ 31์ผ์ด ์žˆ๋‹ค๊ณ  ์ž˜๋ชป ๊ฐ€์ •ํ•˜๊ณ  ์žˆ๋Š”๋ฐ, ์ด๋Š” ๋ฐ˜๋“œ์‹œ ์‚ฌ์‹ค์ด ์•„๋‹™๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด 1์ธ๋ถ„์„ 31*2=60์ผ ๋™์•ˆ 62์ธ๋ถ„์œผ๋กœ ์ž˜๋ชป ๊ณ„์‚ฐํ•˜์—ฌ ์ตœ์ข… ๊ณ„์‚ฐ์—๋Š” ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์œผ๋ฉฐ, 60์ธ๋ถ„๋„ 4์ƒ์ž๊ฐ€ ํ•„์š”ํ•˜๋ฏ€๋กœ ์šฐ์—ฐํžˆ ํ•„์š”ํ•œ ์ƒ์ž ์ˆ˜๊ฐ€ ๋งž์Šต๋‹ˆ๋‹ค. ์ตœ์ข… ๊ณฑ์…ˆ ๋‹จ๊ณ„๋Š” ์ •ํ™•ํ•˜๋ฉฐ ์ด ๋น„์šฉ์ด ๋งž์Šต๋‹ˆ๋‹ค.
Model Collapse:The phenomenon appeared in the _recent_ literature in the context of language and image generation. Several recent works demonstrate facets of this phenomenon _empirically_ in various settings (Hataya et al., 2023; Martinez et al., 2023a;b; Bohacek & Farid, 2023; Briesch et al., 2023; Guo et al., 2023; Fan et al., 2023). Only few recent works also provide some accompanying theoretical analysis (Shumailov et al., 2023; Alemohammad et al., 2023; Bertrand et al., 2023) which we outline now.
๋ชจ๋ธ ๋ถ•๊ดด: ์ด ํ˜„์ƒ์€ _์ตœ๊ทผ_ ๋ฌธํ—Œ์—์„œ ์–ธ์–ด ๋ฐ ์ด๋ฏธ์ง€ ์ƒ์„ฑ์˜ ๋งฅ๋ฝ์—์„œ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์ตœ๊ทผ ์—ฐ๊ตฌ๋“ค์€ ๋‹ค์–‘ํ•œ ์„ค์ •์—์„œ ์ด ํ˜„์ƒ์˜ ์—ฌ๋Ÿฌ ์ธก๋ฉด์„ _๊ฒฝํ—˜์ ์œผ๋กœ_ ์ž…์ฆํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค (Hataya et al., 2023; Martinez et al., 2023a;b; Bohacek & Farid, 2023; Briesch et al., 2023; Guo et al., 2023; Fan et al., 2023). ์†Œ์ˆ˜์˜ ์ตœ๊ทผ ์—ฐ๊ตฌ๋“ค๋งŒ์ด ์ผ๋ถ€ ์ด๋ก ์  ๋ถ„์„์„ ๋™๋ฐ˜ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค (Shumailov et al., 2023; Alemohammad et al., 2023; Bertrand et al., 2023) ์ด๋ฅผ ์ง€๊ธˆ ๊ฐœ๋žต์ ์œผ๋กœ ์„ค๋ช…ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.
๋ชจ๋ธ ๋ถ•๊ดด: ์ด ํ˜„์ƒ์€ ์–ธ์–ด ๋ฐ ์ด๋ฏธ์ง€ ์ƒ์„ฑ์˜ ๋งฅ๋ฝ์—์„œ ์ตœ๊ทผ ๋ฌธํ—Œ์—์„œ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ์˜ ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ ๋‹ค์–‘ํ•œ ์„ค์ •์—์„œ ์ด ํ˜„์ƒ์˜ ์—ฌ๋Ÿฌ ์ธก๋ฉด์„ _๊ฒฝํ—˜์ ์œผ๋กœ_ ์ž…์ฆํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Hataya et al., 2023; Martinez et al., 2023a;b; Bohacek & Farid, 2023; Briesch et al., 2023; Guo et al., 2023; Fan et al., 2023). ์ตœ๊ทผ์˜ ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด๋ก ์  ๋ถ„์„์„ ๋™๋ฐ˜ํ•œ ์—ฐ๊ตฌ๋„ ์ผ๋ถ€ ์ œ์‹œํ•˜๊ณ  ์žˆ๋Š”๋ฐ, ์ด๋ฅผ ๊ฐœ๊ด„์ ์œผ๋กœ ์‚ดํŽด๋ณด๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค(Shumailov et al., 2023; Alemohammad et al., 2023; Bertrand et al., 2023).
However, while LLMs have been shown to encode clinical knowledge and proven capable of highly accurate single-turn medical question-answering [12, 13, 14], their conversational capabilities have been tailored to domains outside clinical medicine [15, 16]. Prior work in LLMs for health [12, 13, 14, 17, 18] has not yet rigorously examined the clinical history-taking and diagnostic dialogue capabilities of AI systems or contextualized this by comparison to the extensive capabilities of expert clinicians.
๊ทธ๋Ÿฌ๋‚˜ LLM์ด ์ž„์ƒ ์ง€์‹์„ ์ธ์ฝ”๋”ฉํ•˜๊ณ  ๋งค์šฐ ์ •ํ™•ํ•œ ๋‹จ์ผ ํšŒ์ฐจ ์˜๋ฃŒ ์งˆ๋ฌธ ์‘๋‹ต์ด ๊ฐ€๋Šฅํ•˜๋‹ค๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  [12, 13, 14], ๊ทธ๋“ค์˜ ๋Œ€ํ™” ๋Šฅ๋ ฅ์€ ์ž„์ƒ ์˜ํ•™ ์™ธ์˜ ๋„๋ฉ”์ธ์— ๋งž์ถฐ์ ธ ์žˆ์Šต๋‹ˆ๋‹ค [15, 16]. ๊ฑด๊ฐ•์„ ์œ„ํ•œ LLM์— ๋Œ€ํ•œ ์ด์ „ ์—ฐ๊ตฌ [12, 13, 14, 17, 18]๋Š” AI ์‹œ์Šคํ…œ์˜ ์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™” ๋Šฅ๋ ฅ์„ ์—„๊ฒฉํ•˜๊ฒŒ ์กฐ์‚ฌํ•˜์ง€ ์•Š์•˜์œผ๋ฉฐ, ์ด๋ฅผ ์ „๋ฌธ๊ฐ€ ์ž„์ƒ์˜์˜ ๊ด‘๋ฒ”์œ„ํ•œ ๋Šฅ๋ ฅ๊ณผ ๋น„๊ตํ•˜์—ฌ ๋งฅ๋ฝํ™”ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ LLM์€ ์ž„์ƒ ์ง€์‹์„ ์ธ์ฝ”๋”ฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ ๋งค์šฐ ์ •ํ™•ํ•œ ๋‹จ์ผ ํ„ด ์˜๋ฃŒ ์งˆ๋ฌธ์— ๋‹ตํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์ง€๋งŒ[12, 13, 14], ๊ทธ ๋Œ€ํ™” ๋Šฅ๋ ฅ์€ ์ž„์ƒ ์˜ํ•™ ์ด์™ธ์˜ ์˜์—ญ์— ๋งž์ถคํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค[15, 16]. ๊ฑด๊ฐ•์„ ์œ„ํ•œ LLM์˜ ์„ ํ–‰ ์—ฐ๊ตฌ[12, 13, 14, 17, 18]๋Š” ์•„์ง AI ์‹œ์Šคํ…œ์˜ ์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™” ๊ธฐ๋Šฅ์„ ์—„๊ฒฉํ•˜๊ฒŒ ๊ฒ€ํ† ํ•˜์ง€ ์•Š์•˜๊ฑฐ๋‚˜ ์ „๋ฌธ ์ž„์ƒ์˜์˜ ๊ด‘๋ฒ”์œ„ํ•œ ๊ธฐ๋Šฅ๊ณผ ๋น„๊ตํ•˜์—ฌ ๋งฅ๋ฝํ™”ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
Moreover, KnowledGPT Wang et al. (2023) and Rewrite-Retrieve-Read Ma et al. (2023) introduce approaches for query manipulation through "program of thought" prompting and innovative query rewriting techniques. KnowledGPT innovates by generating code to interface with knowledge bases, converting user queries into structured search commands. In contrast, Rewrite-Retrieve-Read utilizes a trainable compact LM for query reformulation, adjusting them to more effectively reflect the user's intent and context.
๊ฒŒ๋‹ค๊ฐ€, KnowledGPT(Wang et al., 2023)์™€ Rewrite-Retrieve-Read(Ma et al., 2023)๋Š” "์‚ฌ๊ณ  ํ”„๋กœ๊ทธ๋žจ" ํ”„๋กฌํ”„ํŠธ์™€ ํ˜์‹ ์ ์ธ ์ฟผ๋ฆฌ ์žฌ์ž‘์„ฑ ๊ธฐ์ˆ ์„ ํ†ตํ•ด ์ฟผ๋ฆฌ ์กฐ์ž‘์— ๋Œ€ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. KnowledGPT๋Š” ์ง€์‹ ๋ฒ ์ด์Šค์™€ ์ธํ„ฐํŽ˜์ด์Šคํ•˜๊ธฐ ์œ„ํ•ด ์ฝ”๋“œ๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ๋ฅผ ๊ตฌ์กฐํ™”๋œ ๊ฒ€์ƒ‰ ๋ช…๋ น์œผ๋กœ ๋ณ€ํ™˜ํ•จ์œผ๋กœ์จ ํ˜์‹ ์„ ์ด๋ฃน๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, Rewrite-Retrieve-Read๋Š” ์ฟผ๋ฆฌ ์žฌ๊ตฌ์„ฑ์„ ์œ„ํ•ด ํ•™์Šต ๊ฐ€๋Šฅํ•œ ์†Œํ˜• LM์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์šฉ์ž ์˜๋„์™€ ๋งฅ๋ฝ์„ ๋” ํšจ๊ณผ์ ์œผ๋กœ ๋ฐ˜์˜ํ•˜๋„๋ก ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค.
๋˜ํ•œ, KnowledGPT Wang et al. (2023)๊ณผ Rewrite-Retrieve-Read Ma et al. (2023)์€ "์ƒ๊ฐ์˜ ํ”„๋กœ๊ทธ๋žจ" ํ”„๋กฌํ”„ํŠธ์™€ ํ˜์‹ ์ ์ธ ์ฟผ๋ฆฌ ์žฌ์ž‘์„ฑ ๊ธฐ์ˆ ์„ ํ†ตํ•ด ์ฟผ๋ฆฌ ์กฐ์ž‘์„ ์œ„ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. KnowledGPT๋Š” ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ๋ฅผ ๊ตฌ์กฐํ™”๋œ ๊ฒ€์ƒ‰ ๋ช…๋ น์œผ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์ง€์‹ ๋ฒ ์ด์Šค์™€ ์ธํ„ฐํŽ˜์ด์Šคํ•˜๊ธฐ ์œ„ํ•œ ์ฝ”๋“œ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ํ˜์‹ ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, Rewrite-Retrieve-Read๋Š” ํ›ˆ๋ จ ๊ฐ€๋Šฅํ•œ ์†Œํ˜• LM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฟผ๋ฆฌ๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜์—ฌ ์‚ฌ์šฉ์ž์˜ ์˜๋„์™€ ์ปจํ…์ŠคํŠธ๋ฅผ ๋ณด๋‹ค ํšจ๊ณผ์ ์œผ๋กœ ๋ฐ˜์˜ํ•˜๋„๋ก ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค.
We also report evaluation results for ablated models in Tab. 2. 'SOLAR 10.7B + SFT + DPO' uses all the DPO data at once, _i.e.,_\(D_{1}+D_{2}\), same as the conventional DPO training setup. 'SOLAR 10.7B + SFT + sDPO Strat.' uses stratified sampling to sample \(\sim 16.67\%\) of the data points from the union of OpenOrca and Ultrafeedback Cleaned to form \(D_{1}\) and use the remaining \(\sim 83.33\%\) as \(D_{2}\) to mirror the dataset size of \(D_{1}\) and \(D_{2}\) used in SOLAR 10.7B + SFT + sDPO.
Tab. 2์—์„œ ablated ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. 'SOLAR 10.7B + SFT + DPO'๋Š” ๊ธฐ์กด์˜ DPO ํ›ˆ๋ จ ์„ค์ •๊ณผ ๋™์ผํ•˜๊ฒŒ ๋ชจ๋“  DPO ๋ฐ์ดํ„ฐ๋ฅผ ํ•œ ๋ฒˆ์— ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค, _์ฆ‰,_ \(D_{1}+D_{2}\). 'SOLAR 10.7B + SFT + sDPO Strat.'๋Š” OpenOrca์™€ Ultrafeedback Cleaned์˜ ๊ฒฐํ•ฉ์—์„œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์˜ ์•ฝ 16.67%๋ฅผ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ \(D_{1}\)์„ ํ˜•์„ฑํ•˜๊ณ  ๋‚˜๋จธ์ง€ ์•ฝ 83.33%๋ฅผ \(D_{2}\)๋กœ ์‚ฌ์šฉํ•˜์—ฌ SOLAR 10.7B + SFT + sDPO์—์„œ ์‚ฌ์šฉ๋œ \(D_{1}\)์™€ \(D_{2}\)์˜ ๋ฐ์ดํ„ฐ์…‹ ํฌ๊ธฐ๋ฅผ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค.
๋˜ํ•œ ํ‘œ 2์— ์ ˆ์ œ๋œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. 'SOLAR 10.7B + SFT + DPO'๋Š” ๊ธฐ์กด DPO ํ›ˆ๋ จ ์„ค์ •๊ณผ ๋™์ผํ•˜๊ฒŒ ๋ชจ๋“  DPO ๋ฐ์ดํ„ฐ๋ฅผ ํ•œ ๋ฒˆ์— ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. 'SOLAR 10.7B + SFT + sDPO Strat.'๋Š” ๊ณ„์ธตํ™”๋œ ์ƒ˜ํ”Œ๋ง์„ ์‚ฌ์šฉํ•˜์—ฌ OpenOrca์™€ Ultrafeedback Cleaned์˜ Union์—์„œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์˜ \(\sim 16.67\%\)๋ฅผ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ \(D_{1}\)๋ฅผ ํ˜•์„ฑํ•˜๊ณ  ๋‚˜๋จธ์ง€ \(\sim 83.33\%\)๋ฅผ \(D_{2}\)๋กœ ์‚ฌ์šฉํ•˜์—ฌ SOLAR 10.7B + SFT + sDPO์—์„œ ์‚ฌ์šฉ๋œ \(D_{1}\) ๋ฐ \(D_{2}\)์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ ํฌ๊ธฐ๋ฅผ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)**Colossal-AI**[189] is a deep learning library developed by HPC-AI Tech for training large-scale AI models. It is implemented based on PyTorch and supports a rich collection of parallel training strategies. Furthermore, it can also optimize heterogeneous memory management with methods proposed by PatrickStar [190]. Recently, a ChatGPT-like model called ColossalChat [140] has been publicly released with two versions (7B and 13B), which are developed using Colossal-AI based on LLaMA [57].
โ€ข **Colossal-AI**[189]๋Š” ๋Œ€๊ทœ๋ชจ AI ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด HPC-AI Tech์—์„œ ๊ฐœ๋ฐœํ•œ ๋”ฅ๋Ÿฌ๋‹ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์ž…๋‹ˆ๋‹ค. ์ด ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋Š” PyTorch๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌํ˜„๋˜์—ˆ์œผ๋ฉฐ, ๋‹ค์–‘ํ•œ ๋ณ‘๋ ฌ ํ›ˆ๋ จ ์ „๋žต์„ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, PatrickStar [190]์—์„œ ์ œ์•ˆํ•œ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ์ด์ข… ๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ๋„ ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” LLaMA [57]๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ Colossal-AI๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ๋ฐœ๋œ ColossalChat [140]์ด๋ผ๋Š” ChatGPT ์œ ์‚ฌ ๋ชจ๋ธ์ด ๋‘ ๊ฐ€์ง€ ๋ฒ„์ „(7B ๋ฐ 13B)์œผ๋กœ ๊ณต๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)**Colossal-AI**[189]๋Š” ๋Œ€๊ทœ๋ชจ AI ๋ชจ๋ธ ํ•™์Šต์„ ์œ„ํ•ด HPC-AI Tech์—์„œ ๊ฐœ๋ฐœํ•œ ๋”ฅ ๋Ÿฌ๋‹ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์ž…๋‹ˆ๋‹ค. PyTorch๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌํ˜„๋˜์—ˆ์œผ๋ฉฐ ๋‹ค์–‘ํ•œ ๋ณ‘๋ ฌ ํ•™์Šต ์ „๋žต์„ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ PatrickStar[190]์—์„œ ์ œ์•ˆํ•œ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ์ด๊ธฐ์ข… ๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ๋ฅผ ์ตœ์ ํ™”ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” ColossalChat[140]์ด๋ผ๋Š” ChatGPT์™€ ์œ ์‚ฌํ•œ ๋ชจ๋ธ์ด ๊ณต๊ฐœ๋˜์—ˆ์œผ๋ฉฐ, ์ด ๋ชจ๋ธ์€ LLaMA[57]๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ Colossal-AI๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ๋ฐœ๋œ ๋‘ ๊ฐ€์ง€ ๋ฒ„์ „(7B ๋ฐ 13B)์œผ๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
FindingsSimilar to other tasks, the disabling of language-specific neurons within the generation layer attenuates their capacity to generate content in the respective languages. By selectively deactivating neurons that are not associated with English, we do not completely eliminate the models' multilingual generation abilities. However, as demonstrated in Table 2, the complete deactivation of all language-specific neurons results in the total loss of the LLMs' multilingual generation capabilities.
๋‹ค๋ฅธ ์ž‘์—…๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ, ์ƒ์„ฑ ์ธต ๋‚ด์˜ ์–ธ์–ด ํŠน์ • ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด ํ•ด๋‹น ์–ธ์–ด๋กœ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์ด ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ์˜์–ด์™€ ๊ด€๋ จ์ด ์—†๋Š” ๋‰ด๋Ÿฐ์„ ์„ ํƒ์ ์œผ๋กœ ๋น„ํ™œ์„ฑํ™”ํ•จ์œผ๋กœ์จ ๋ชจ๋ธ์˜ ๋‹ค๊ตญ์–ด ์ƒ์„ฑ ๋Šฅ๋ ฅ์„ ์™„์ „ํžˆ ์ œ๊ฑฐํ•˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Table 2์—์„œ ๋ณด์—ฌ์ฃผ๋“ฏ์ด, ๋ชจ๋“  ์–ธ์–ด ํŠน์ • ๋‰ด๋Ÿฐ์„ ์™„์ „ํžˆ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด LLM์˜ ๋‹ค๊ตญ์–ด ์ƒ์„ฑ ๋Šฅ๋ ฅ์ด ์™„์ „ํžˆ ์ƒ์‹ค๋ฉ๋‹ˆ๋‹ค.
๋ฐœ๊ฒฌ๋‹ค๋ฅธ ์ž‘์—…๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ƒ์„ฑ ๊ณ„์ธต ๋‚ด์—์„œ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด ํ•ด๋‹น ์–ธ์–ด๋กœ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์ด ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค. ์˜์–ด์™€ ๊ด€๋ จ์ด ์—†๋Š” ๋‰ด๋Ÿฐ์„ ์„ ํƒ์ ์œผ๋กœ ๋น„ํ™œ์„ฑํ™”ํ•จ์œผ๋กœ์จ ๋ชจ๋ธ์˜ ๋‹ค๊ตญ์–ด ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ์™„์ „ํžˆ ์ œ๊ฑฐํ•˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ‘œ 2์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ๋ชจ๋“  ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์„ ์™„์ „ํžˆ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด LLM์˜ ๋‹ค๊ตญ์–ด ์ƒ์„ฑ ๊ธฐ๋Šฅ์ด ์™„์ „ํžˆ ์†์‹ค๋ฉ๋‹ˆ๋‹ค.
semantically equivalent references with diverse expressions. Further, LLMs are widely employed as the evaluators of text generation in a reference-free manner, including evaluating a single prediction [631, 642, 632] or comparing several candidates [643, 644, 138, 645]. Nevertheless, LLMs may expose bias (_e.g.,_ order bias or preference for LLM-generated texts over human-written texts) as language generation evaluators, demonstrating disparities when compared to human evaluation [646, 632, 647].
์˜๋ฏธ์ ์œผ๋กœ ๋™๋“ฑํ•˜์ง€๋งŒ ๋‹ค์–‘ํ•œ ํ‘œํ˜„์„ ๊ฐ€์ง„ ์ฐธ์กฐ๋“ค. ๋” ๋‚˜์•„๊ฐ€, LLMs๋Š” ๋‹จ์ผ ์˜ˆ์ธก ํ‰๊ฐ€ [631, 642, 632] ๋˜๋Š” ์—ฌ๋Ÿฌ ํ›„๋ณด ๋น„๊ต [643, 644, 138, 645]๋ฅผ ํฌํ•จํ•˜์—ฌ ์ฐธ์กฐ ์—†์ด ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ํ‰๊ฐ€์ž๋กœ ๋„๋ฆฌ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , LLMs๋Š” ์–ธ์–ด ์ƒ์„ฑ ํ‰๊ฐ€์ž๋กœ์„œ ํŽธํ–ฅ(์˜ˆ: ์ˆœ์„œ ํŽธํ–ฅ ๋˜๋Š” ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ ํ…์ŠคํŠธ๋ณด๋‹ค LLM์ด ์ƒ์„ฑํ•œ ํ…์ŠคํŠธ๋ฅผ ์„ ํ˜ธํ•˜๋Š” ๊ฒฝํ–ฅ)์„ ๋“œ๋Ÿฌ๋‚ผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ธ๊ฐ„ ํ‰๊ฐ€์™€ ๋น„๊ตํ•  ๋•Œ ์ฐจ์ด๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค [646, 632, 647].
๋‹ค์–‘ํ•œ ํ‘œํ˜„์„ ์‚ฌ์šฉํ•˜์—ฌ ์˜๋ฏธ์ ์œผ๋กœ ๋™๋“ฑํ•œ ์ฐธ์กฐ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์€ ๋‹จ์ผ ์˜ˆ์ธก ํ‰๊ฐ€[631, 642, 632] ๋˜๋Š” ์—ฌ๋Ÿฌ ํ›„๋ณด ๋น„๊ต[643, 644, 138, 645]๋ฅผ ํฌํ•จํ•˜์—ฌ ์ฐธ์กฐ ์—†๋Š” ๋ฐฉ์‹์œผ๋กœ ํ…์ŠคํŠธ ์ƒ์„ฑ์˜ ํ‰๊ฐ€์ž๋กœ ๋„๋ฆฌ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์€ ์–ธ์–ด ์ƒ์„ฑ ํ‰๊ฐ€์ž๋กœ์„œ ํŽธํ–ฅ์„ฑ(์˜ˆ: ์ˆœ์„œ ํŽธํ–ฅ ๋˜๋Š” ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ํ…์ŠคํŠธ๋ณด๋‹ค LLM์ด ์ƒ์„ฑํ•œ ํ…์ŠคํŠธ์— ๋Œ€ํ•œ ์„ ํ˜ธ๋„)์„ ๋…ธ์ถœํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์‚ฌ๋žŒ์˜ ํ‰๊ฐ€์™€ ๋น„๊ตํ•  ๋•Œ ์ฐจ์ด๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค[646, 632, 647].
LLMs are now commonly used as parts of larger, more complex systems. It is crucial to understand how these models interact with information with varying degrees of trustworthiness, accuracy, and uniformity. Our analysis shows that further work is required to characterize the risks of using LLMs to answer questions given contextual information. In particular, we find that model behavior can be erratic and unpredictable when presented with information that exists at the margin of its prior beliefs.
LLMs๋Š” ์ด์ œ ๋” ํฌ๊ณ  ๋ณต์žกํ•œ ์‹œ์Šคํ…œ์˜ ์ผ๋ถ€๋กœ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์ด ์‹ ๋ขฐ์„ฑ, ์ •ํ™•์„ฑ, ์ผ๊ด€์„ฑ์ด ๋‹ค์–‘ํ•œ ์ •๋ณด์™€ ์–ด๋–ป๊ฒŒ ์ƒํ˜ธ์ž‘์šฉํ•˜๋Š”์ง€ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด, ๋งฅ๋ฝ์  ์ •๋ณด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์˜ ์œ„ํ—˜์„ ํŠน์„ฑํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€์ ์ธ ์ž‘์—…์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ๋ชจ๋ธ์ด ์ด์ „ ์‹ ๋…์˜ ๊ฒฝ๊ณ„์— ์žˆ๋Š” ์ •๋ณด๋ฅผ ์ ‘ํ–ˆ์„ ๋•Œ ๊ทธ ํ–‰๋™์ด ๋ถˆ๊ทœ์น™ํ•˜๊ณ  ์˜ˆ์ธก ๋ถˆ๊ฐ€๋Šฅํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
LLM์€ ์ด์ œ ์ผ๋ฐ˜์ ์œผ๋กœ ๋” ํฌ๊ณ  ๋ณต์žกํ•œ ์‹œ์Šคํ…œ์˜ ์ผ๋ถ€๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์ด ์‹ ๋ขฐ๋„, ์ •ํ™•๋„ ๋ฐ ํ†ต์ผ์„ฑ์ด ๋‹ค์–‘ํ•œ ์ •๋ณด์™€ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ๋ฐฉ์‹์„ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋ถ„์„ ๊ฒฐ๊ณผ, ๋ฌธ๋งฅ ์ •๋ณด๊ฐ€ ์ฃผ์–ด์ง„ ์งˆ๋ฌธ์— LLM์„ ์‚ฌ์šฉํ•  ๋•Œ ๋ฐœ์ƒํ•˜๋Š” ์œ„ํ—˜์„ ํŠน์„ฑํ™”ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ถ”๊ฐ€์ ์ธ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ๋ชจ๋ธ์ด ์ด์ „ ์‹ ๋…์˜ ํ•œ๊ณ„์— ์กด์žฌํ•˜๋Š” ์ •๋ณด๋ฅผ ์ œ์‹œ๋ฐ›์„ ๋•Œ ๋ชจ๋ธ ๋™์ž‘์ด ๋ถˆ๊ทœ์น™ํ•˜๊ณ  ์˜ˆ์ธกํ•  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
Large Language Models.Recently, LLMs have demonstrated remarkable achievements across a broad spectrum of tasks. Prominent tech companies have made significant strides in developing highly proficient LLMs. These include OpenAI's GPT3&4 [1; 2], Google's PaLM [3; 4], and Bard3, DeepMind's Chinchilla [5], and Gopher [6], as well as Anthropic's Claude4. However, it is important to note that these models are closed-source and can only be accessed through specific APIs or may not be accessible at all.
๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธ(LLM). ์ตœ๊ทผ LLM์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ๋†€๋ผ์šด ์„ฑ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์š” ๊ธฐ์ˆ  ํšŒ์‚ฌ๋“ค์€ ๋งค์šฐ ๋Šฅ์ˆ™ํ•œ LLM์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ์žˆ์–ด ์ƒ๋‹นํ•œ ์ง„์ „์„ ์ด๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” OpenAI์˜ GPT-3 ๋ฐ GPT-4 [1; 2], ๊ตฌ๊ธ€์˜ PaLM [3; 4]๊ณผ Bard3, DeepMind์˜ Chinchilla [5]์™€ Gopher [6], ๊ทธ๋ฆฌ๊ณ  Anthropic์˜ Claude4๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ๋“ค์€ ํ์‡„ํ˜• ์†Œ์Šค์ด๋ฉฐ ํŠน์ • API๋ฅผ ํ†ตํ•ด์„œ๋งŒ ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๊ฑฐ๋‚˜ ์ „ํ˜€ ์ ‘๊ทผํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์ ์„ ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ.์ตœ๊ทผ LLM์€ ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—…์—์„œ ๋†€๋ผ์šด ์„ฑ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์œ ๋ช… ๊ธฐ์ˆ  ๊ธฐ์—…๋“ค์€ ๊ณ ๋„๋กœ ๋Šฅ์ˆ™ํ•œ LLM์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ์ƒ๋‹นํ•œ ์ง„์ „์„ ์ด๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” OpenAI์˜ GPT3&4 [1; 2], Google์˜ PaLM [3; 4], Bard3, DeepMind์˜ Chinchilla [5], Gopher [6], Anthropic์˜ Claude4 ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ๋น„๊ณต๊ฐœ ๋ชจ๋ธ์ด๋ฉฐ ํŠน์ • API๋ฅผ ํ†ตํ•ด์„œ๋งŒ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๊ฑฐ๋‚˜ ์ „ํ˜€ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์—†์„ ์ˆ˜๋„ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
As mentioned in Section 2, there are fundamental differences between the traditional CL formulations and CKL which make the previous CL methods inadequate for the CKL setting. In this section, we introduce the prior traditional continual learning methods in detail, explore the methods from the literature set as baselines for the CKL benchmark and how they address the identified limitations of CL methods, and provide descriptions about alternative methods making LMs cope with the changing world.
2์žฅ์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, ๊ธฐ์กด์˜ CL ๊ณต์‹ํ™”์™€ CKL ์‚ฌ์ด์—๋Š” ๊ทผ๋ณธ์ ์ธ ์ฐจ์ด์ ์ด ์žˆ์–ด ์ด์ „์˜ CL ๋ฐฉ๋ฒ•๋“ค์ด CKL ์„ค์ •์— ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ๊ธฐ์กด์˜ ์ „ํ†ต์ ์ธ ์ง€์† ํ•™์Šต ๋ฐฉ๋ฒ•๋“ค์„ ์ž์„ธํžˆ ์†Œ๊ฐœํ•˜๊ณ , CKL ๋ฒค์น˜๋งˆํฌ์˜ ๊ธฐ์ค€์œผ๋กœ ์„ค์ •๋œ ๋ฌธํ—Œ์˜ ๋ฐฉ๋ฒ•๋“ค์„ ํƒ๊ตฌํ•˜๋ฉฐ, ์ด๋“ค์ด CL ๋ฐฉ๋ฒ•์˜ ํ•œ๊ณ„๋ฅผ ์–ด๋–ป๊ฒŒ ํ•ด๊ฒฐํ•˜๋Š”์ง€ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋ณ€ํ™”ํ•˜๋Š” ์„ธ์ƒ์— LMs๊ฐ€ ์ ์‘ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๋Œ€์ฒด ๋ฐฉ๋ฒ•๋“ค์— ๋Œ€ํ•œ ์„ค๋ช…๋„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
2์ ˆ์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, ๊ธฐ์กด CL ๊ณต์‹๊ณผ CKL ์‚ฌ์ด์—๋Š” ๊ทผ๋ณธ์ ์ธ ์ฐจ์ด๊ฐ€ ์žˆ์–ด ๊ธฐ์กด CL ๋ฐฉ๋ฒ•์ด CKL ํ™˜๊ฒฝ์— ๋ถ€์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ๊ธฐ์กด์˜ ์ง€์†์  ํ•™์Šต ๋ฐฉ๋ฒ•์„ ์ž์„ธํžˆ ์†Œ๊ฐœํ•˜๊ณ , CKL ๋ฒค์น˜๋งˆํฌ์˜ ๊ธฐ์ค€์ด ๋˜๋Š” ๋ฌธํ—Œ์—์„œ ์ถ”์ถœํ•œ ๋ฐฉ๋ฒ•๊ณผ CL ๋ฐฉ๋ฒ•์˜ ํ•œ๊ณ„๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‚ดํŽด๋ณด๊ณ , ๋ณ€ํ™”ํ•˜๋Š” ์„ธ์ƒ์— ๋Œ€์ฒ˜ํ•˜๋Š” LM์— ๋Œ€ํ•œ ๋Œ€์•ˆ์ ์ธ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
_Context:_ John Mayne was a Scottish printer, journalist and poet born in Dumfries. In 1780, his poem The Siller Gun appeared in its original form in Ruddiman's Magazine, published by Walter Ruddiman in Edinburgh. It is a humorous work on an ancient custom in Dumfries of shooting for the Siller Gun.He also wrote a poem on Hallowe'en in 1780 which influenced Robert Burns's 1785 poem Halloween. Mayne also wrote a version of the ballad Helen of Kirkcnonnel. His verses were admired by Walter Scott.
_๋งฅ๋ฝ:_ John Mayne์€ ์Šค์ฝ”ํ‹€๋žœ๋“œ์˜ ์ธ์‡„์—…์ž, ๊ธฐ์ž, ์‹œ์ธ์œผ๋กœ Dumfries์—์„œ ํƒœ์–ด๋‚ฌ์Šต๋‹ˆ๋‹ค. 1780๋…„์— ๊ทธ์˜ ์‹œ "The Siller Gun"์ด ์›๋ž˜ ํ˜•ํƒœ๋กœ Ruddiman's Magazine์— ์‹ค๋ ธ์œผ๋ฉฐ, ์ด ์žก์ง€๋Š” Edinburgh์—์„œ Walter Ruddiman์ด ๋ฐœํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์‹œ๋Š” Dumfries์—์„œ Siller Gun์„ ์˜๋Š” ๊ณ ๋Œ€ ๊ด€์Šต์— ๋Œ€ํ•œ ์œ ๋จธ๋Ÿฌ์Šคํ•œ ์ž‘ํ’ˆ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Š” ๋˜ํ•œ 1780๋…„์— Hallowe'en์— ๊ด€ํ•œ ์‹œ๋ฅผ ์ผ์œผ๋ฉฐ, ์ด๋Š” Robert Burns์˜ 1785๋…„ ์‹œ "Halloween"์— ์˜ํ–ฅ์„ ๋ฏธ์ณค์Šต๋‹ˆ๋‹ค. Mayne์€ ๋˜ํ•œ "Helen of Kirkconnell"์ด๋ผ๋Š” ๋ฐœ๋ผ๋“œ์˜ ๋ฒ„์ „์„ ์ผ์Šต๋‹ˆ๋‹ค. ๊ทธ์˜ ์‹œ๊ตฌ๋Š” Walter Scott์— ์˜ํ•ด ์นญ์ฐฌ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค.
_์ปจํ…์ŠคํŠธ:_ ์กด ๋ฉ”์ธ์€ ๋คํ”„๋ฆฌ์Šค์—์„œ ํƒœ์–ด๋‚œ ์Šค์ฝ”ํ‹€๋žœ๋“œ์˜ ์ธ์‡„์—…์ž, ์–ธ๋ก ์ธ, ์‹œ์ธ์ž…๋‹ˆ๋‹ค. 1780๋…„ ์—๋“ ๋ฒ„๋Ÿฌ์—์„œ ๋ฐœํ–‰๋œ ์›”ํ„ฐ ๋ฃจ๋“œ๋งŒ์˜ ์žก์ง€ '๋ฃจ๋“œ๋งŒ ๋งค๊ฑฐ์ง„'์— ๊ทธ์˜ ์‹œ '์‹ค๋Ÿฌ ๊ฑด'์ด ์›ํ˜• ๊ทธ๋Œ€๋กœ ์‹ค๋ ธ์Šต๋‹ˆ๋‹ค. ์ด ์‹œ๋Š” ๋คํ”„๋ฆฌ์Šค์˜ ๊ณ ๋Œ€ ๊ด€์Šต์ธ ์‹ค๋Ÿฌ ๊ฑด์„ ์˜๋Š” ๊ฒƒ์— ๋Œ€ํ•œ ์œ ๋จธ๋Ÿฌ์Šคํ•œ ์ž‘ํ’ˆ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Š” 1780๋…„์— ๋กœ๋ฒ„ํŠธ ๋ฒˆ์ฆˆ์˜ 1785๋…„ ์‹œ 'ํ• ๋กœ์œˆ'์— ์˜ํ–ฅ์„ ์ค€ ํ• ๋กœ์œˆ์— ๊ด€ํ•œ ์‹œ๋„ ์ผ์Šต๋‹ˆ๋‹ค. ๋ฉ”์ธ์€ ๋ฐœ๋ผ๋“œ '์ปคํฌ๋…ธ๋„ฌ์˜ ํ—ฌ๋ Œ'์˜ ๋ฒ„์ „๋„ ์ผ์Šต๋‹ˆ๋‹ค. ๊ทธ์˜ ์‹œ๋Š” ์›”ํ„ฐ ์Šค์ฝง์˜ ์ฐฌ์‚ฌ๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค.
**Non-web Snapshots**: perplexity other validation sets from the Pile, comprising of OpenWebText2, HackerNews, Wikipedia (en), BookCorpusFair, DM Mathematics, Gutenberg PG-19, OpenSubtitles, and USPTO. Also included in this average is "redditflattened" (validation set from Pusshift.io Reddit [4]), "stories", "prompts_with_answers" (which is described below) and "prompts" (which is the same as "prompts_with_answers" but where each sample is just the instruction-tuning prompt without the answer).
**Non-web Snapshots**: Pile์˜ ๋‹ค๋ฅธ ๊ฒ€์ฆ ์„ธํŠธ์ธ OpenWebText2, HackerNews, Wikipedia (en), BookCorpusFair, DM Mathematics, Gutenberg PG-19, OpenSubtitles, USPTO์˜ ํผํ”Œ๋ ‰์‹œํ‹ฐ. ์ด ํ‰๊ท ์—๋Š” "redditflattened" (Pusshift.io Reddit [4]์˜ ๊ฒ€์ฆ ์„ธํŠธ), "stories", "prompts_with_answers" (์•„๋ž˜์— ์„ค๋ช…๋จ) ๋ฐ "prompts" (๊ฐ ์ƒ˜ํ”Œ์ด ๋‹ต๋ณ€ ์—†์ด instruction-tuning ํ”„๋กฌํ”„ํŠธ๋งŒ ํฌํ•จ๋œ "prompts_with_answers"์™€ ๋™์ผํ•จ)๋„ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
**๋น„์›น ์Šค๋ƒ…์ƒท**: OpenWebText2, HackerNews, Wikipedia(en), BookCorpusFair, DM ์ˆ˜ํ•™, ๊ตฌํ…๋ฒ ๋ฅดํฌ PG-19, ์˜คํ”ˆ ์ž๋ง‰, USPTO๋กœ ๊ตฌ์„ฑ๋œ ๋” ํŒŒ์ผ์˜ ๋‹ค๋ฅธ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋ฅผ ๋‚œํ•ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด ํ‰๊ท ์—๋Š” "redditflattened"(Pusshift.io Reddit [4]์˜ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ), "stories", "prompts_with_answers"(์•„๋ž˜ ์„ค๋ช…), "prompts"( "prompts_with_answers"์™€ ๋™์ผํ•˜์ง€๋งŒ ๊ฐ ์ƒ˜ํ”Œ์ด ๋‹ต๋ณ€์ด ์—†๋Š” ์ง€์นจ ์กฐ์ • ํ”„๋กฌํ”„ํŠธ์ผ ๋ฟ์ธ ๊ฒฝ์šฐ)๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
The choice of benchmark tasks \(\mathbb{T}^{F}\), \(T_{n}^{U}\), and \(T_{n}^{A}\) can differ according to each experimental setup. FUAR value of 1.0 represents an equal trade-off scenario where _one_ time-invariant knowledge instance of \(\mathbb{T}^{F}\) is forgotten on average to gain one new or updated knowledge instance of \(T_{n}^{U}\) and \(T_{n}^{A}\). The two terms in the denominators are summed because newly gained knowledge and updated knowledge are mutually exclusive by definition.
๋ฒค์น˜๋งˆํฌ ์ž‘์—… \(\mathbb{T}^{F}\), \(T_{n}^{U}\), ๋ฐ \(T_{n}^{A}\)์˜ ์„ ํƒ์€ ๊ฐ ์‹คํ—˜ ์„ค์ •์— ๋”ฐ๋ผ ๋‹ค๋ฅผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. FUAR ๊ฐ’์ด 1.0์ธ ๊ฒฝ์šฐ, \(\mathbb{T}^{F}\)์˜ ์‹œ๊ฐ„ ๋ถˆ๋ณ€ ์ง€์‹ ์ธ์Šคํ„ด์Šค _ํ•˜๋‚˜_๊ฐ€ ํ‰๊ท ์ ์œผ๋กœ ์žŠํ˜€์ง€๊ณ , \(T_{n}^{U}\) ๋ฐ \(T_{n}^{A}\)์˜ ์ƒˆ๋กœ์šด ๋˜๋Š” ์—…๋ฐ์ดํŠธ๋œ ์ง€์‹ ์ธ์Šคํ„ด์Šค ํ•˜๋‚˜๋ฅผ ์–ป๋Š” ๊ท ํ˜• ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋ถ„๋ชจ์˜ ๋‘ ํ•ญ์€ ์ƒˆ๋กœ ์–ป์€ ์ง€์‹๊ณผ ์—…๋ฐ์ดํŠธ๋œ ์ง€์‹์ด ์ •์˜์ƒ ์ƒํ˜ธ ๋ฐฐํƒ€์ ์ด๊ธฐ ๋•Œ๋ฌธ์— ํ•ฉ์‚ฐ๋ฉ๋‹ˆ๋‹ค.
๋ฒค์น˜๋งˆํฌ ์ž‘์—…์˜ ์„ ํƒ \(\mathbb{T}^{F}\), \(T_{n}^{U}\), \(T_{n}^{A}\)๋Š” ๊ฐ ์‹คํ—˜ ์„ค์ •์— ๋”ฐ๋ผ ๋‹ค๋ฅผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. FUAR ๊ฐ’ 1.0์€ ํ‰๊ท ์ ์œผ๋กœ \(\mathbb{T}^{F}\)์˜ ์‹œ๊ฐ„ ๋ถˆ๋ณ€ ์ง€์‹ ์ธ์Šคํ„ด์Šค 1๊ฐœ๋ฅผ ์žŠ์–ด๋ฒ„๋ ค \(T_{n}^{U}\) ๋ฐ \(T_{n}^{A}\)์˜ ์ƒˆ๋กญ๊ฑฐ๋‚˜ ์—…๋ฐ์ดํŠธ๋œ ์ง€์‹ ์ธ์Šคํ„ด์Šค 1๊ฐœ๋ฅผ ์–ป๋Š” ๋™๋“ฑํ•œ ํŠธ๋ ˆ์ด๋“œ์˜คํ”„ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋ถ„๋ชจ์˜ ๋‘ ํ•ญ์€ ์ƒˆ๋กœ ์–ป์€ ์ง€์‹๊ณผ ์—…๋ฐ์ดํŠธ๋œ ์ง€์‹์ด ์ •์˜์ƒ ์ƒํ˜ธ ๋ฐฐํƒ€์ ์ด๊ธฐ ๋•Œ๋ฌธ์— ํ•ฉ์‚ฐ๋ฉ๋‹ˆ๋‹ค.