chosen
stringlengths
14
4.26k
rejected
stringlengths
53
9.3k
prompt
stringlengths
178
4.55k
**MATH**Hendrycks et al. (2021). ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋†’์€ ์ˆ˜์ค€์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ๊ณผ ์ˆ˜ํ•™์  ์ง€์‹์ด ํ•„์š”ํ•œ ๊ฒฝ์Ÿ ์ˆ˜์ค€์˜ ์ˆ˜ํ•™ ๋ฌธ์ œ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. 7,500๊ฐœ์˜ training ์˜ˆ์ œ์™€ 5,000๊ฐœ์˜ test ์ผ€์ด์Šค๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. training ์˜ˆ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ synthetic ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**MATH**Hendrycks et al.(2021). ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋†’์€ ์ˆ˜์ค€์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ๊ณผ ์ˆ˜ํ•™์  ์ง€์‹์ด ํ•„์š”ํ•œ ๊ฒฝ์Ÿ ์ˆ˜์ค€์˜ ์ˆ˜ํ•™ ๋ฌธ์ œ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. 7,500๊ฐœ์˜ ํ•™์Šต ์˜ˆ์ œ์™€ 5,000๊ฐœ์˜ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ•™์Šต ์˜ˆ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**MATH**Hendrycks et al. (2021). This dataset focuses on competitive-level math problems that requires high levels of reasoning ability and mathematical knowledge. It consists of 7,500 training examples and 5,000 test cases. We use the training examples to generate synthetic data. [|assistant|]
Allen-Zhu & Li(2023b)์—์„œ reversal curse๊ฐ€ ๋ฐœ๊ฒฌ๋˜์—ˆ์„ ๋•Œ, ์ €์ž๋“ค์€ ๊ณ ์œ ํ•œ ์˜์–ด ์ด๋ฆ„์„ ๊ฐ€์ง„ ๋ฌด์ž‘์œ„๋กœ ์ƒ์„ฑ๋œ 10๋งŒ ๋ช…์˜ ๊ฐœ์ธ์— ๋Œ€ํ•œ ์ „๊ธฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „๊ธฐ๋Š” ๋ฌธ์žฅ ํ…œํ”Œ๋ฆฟ ํ’€(bioS ๋ฐ์ดํ„ฐ ์„ธํŠธ)์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑํ•˜๊ฑฐ๋‚˜ Llama ๋ชจ๋ธ(bioR ๋ฐ์ดํ„ฐ ์„ธํŠธ)์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค(Touvron et al., 2023a). ์ „๊ธฐ ํ•ญ๋ชฉ์€ ํ•ญ์ƒ ์‚ฌ๋žŒ์˜ ์ „์ฒด ์ด๋ฆ„์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ reversal QA ์ž‘์—…์€ ๋งค์šฐ ์ž์—ฐ์Šค๋Ÿฝ์Šต๋‹ˆ๋‹ค. ์‚ฌ๋žŒ์˜ ๋ถ€๋ถ„์  ๋˜๋Š” ์ „์ฒด ์ „๊ธฐ ์„ธ๋ถ€ ์ •๋ณด๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ๊ทธ ์‚ฌ๋žŒ์˜ ์ด๋ฆ„์„ ๋ฌผ์–ด๋ด…๋‹ˆ๋‹ค.[|endofturn|]
Allen-Zhu & Li(2023b)์—์„œ ์—ญ์ „์ˆ ์ด ๋ฐœ๊ฒฌ๋˜์—ˆ์„ ๋•Œ, ์ €์ž๋“ค์€ ๊ณ ์œ ํ•œ ์˜์–ด ์ด๋ฆ„์„ ๊ฐ€์ง„ ๋ฌด์ž‘์œ„๋กœ ์ƒ์„ฑ๋œ 10๋งŒ ๋ช…์˜ ๊ฐœ์ธ์— ๋Œ€ํ•œ ์ „๊ธฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „๊ธฐ๋Š” ๋ฌธ์žฅ ํ…œํ”Œ๋ฆฟ ํ’€(bioS ๋ฐ์ดํ„ฐ ์„ธํŠธ)์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑํ•˜๊ฑฐ๋‚˜ Llama ๋ชจ๋ธ(bioR ๋ฐ์ดํ„ฐ ์„ธํŠธ)์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค(Touvron et al., 2023a). ์ „๊ธฐ ํ•ญ๋ชฉ์€ ํ•ญ์ƒ ์‚ฌ๋žŒ์˜ ์ „์ฒด ์ด๋ฆ„์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค.4 ๋”ฐ๋ผ์„œ ์—ญ์ „ QA ์ž‘์—…์€ ๋งค์šฐ ์ž์—ฐ์Šค๋Ÿฝ์Šต๋‹ˆ๋‹ค. ์‚ฌ๋žŒ์˜ ๋ถ€๋ถ„์  ๋˜๋Š” ์ „์ฒด ์ „๊ธฐ ์„ธ๋ถ€ ์ •๋ณด๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ๊ทธ ์‚ฌ๋žŒ์˜ ์ด๋ฆ„์„ ๋ฌผ์–ด๋ด…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]When the reversal curse was discovered in Allen-Zhu & Li (2023b), the authors utilized a biography dataset of 100K randomly generated individuals with unique English names. The biographies were either generated using a pool of sentence templates (the bioS dataset) or generated using the Llama model (the bioR dataset) (Touvron et al., 2023a). The biography entries always start with the person's full name.4 The reversal QA task is, therefore, very natural: given a person's partial or full biography details, ask for the person's name. [|assistant|]
NEFtune: ํ† ํฐ embedding ๋ฒกํ„ฐ์— ๋ฌด์ž‘์œ„ ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ์‹คํ—˜์€ ์†Œ๊ทœ๋ชจ ๋…ธ์ด์ฆˆ๊ฐ€ LLM ๋Œ€ํ™”์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  Alpaca-eval ์ ์ˆ˜ Jain et al. (2023)์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์›๋ž˜ ์„ค์ •์—์„œ๋Š” ์†Œ์Šค์™€ ํƒ€๊ฒŸ์„ ๋ชจ๋‘ ํฌํ•จํ•œ ๋ชจ๋“  ์ž…๋ ฅ embedding์— ๋…ธ์ด์ฆˆ๊ฐ€ ์ถ”๊ฐ€๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
NEFtune: ํ† ํฐ ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ์— ๋ฌด์ž‘์œ„ ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ์‹คํ—˜์€ ์†Œ๊ทœ๋ชจ ๋…ธ์ด์ฆˆ๊ฐ€ LLM ๋Œ€ํ™”์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  Alpaca-eval ์ ์ˆ˜ Jain et al. (2023)์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์›๋ž˜ ์„ค์ •์—์„œ๋Š” ์†Œ์Šค์™€ ํƒ€๊ฒŸ์„ ๋ชจ๋‘ ํฌํ•จํ•œ ๋ชจ๋“  ์ž…๋ ฅ ์ž„๋ฒ ๋”ฉ์— ๋…ธ์ด์ฆˆ๊ฐ€ ์ถ”๊ฐ€๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]NEFtune: Adding random noise to the token embedding vectors. Its experiments show that a small scale of noise can enhance the capability of LLM conversation and improve Alpaca-eval scores Jain et al. (2023). In its original setup, noise is added to all input embeddings, including both source and target. [|assistant|]
์•„์ด๋ฆฌ์Šค ์œˆ๋…ธ์šฐ๋Š” ๋‹ค๋ฅธ ํŠนํŒŒ์› Thea "Attie" Attworod์™€ ํ•จ๊ป˜ ์ „์„  ๊ทผ์ฒ˜์˜ ๋งˆ์„ Avalon Blufl๋กœ ์—ฌํ–‰์„ ๋– ๋‚ฉ๋‹ˆ๋‹ค.[|endofturn|]
14. ์•„์ด๋ฆฌ์Šค ์œˆ๋…ธ์šฐ๋Š” ๋‹ค๋ฅธ ํŠนํŒŒ์› ํ…Œ์•„ "์• ํ‹ฐ" ์•ณ์›Œ๋“œ์™€ ํ•จ๊ป˜ ์ „์„  ๊ทผ์ฒ˜์˜ ๋งˆ์„ ์•„๋ฐœ๋ก  ๋ธ”๋ฃจํ”Œ๋กœ ์—ฌํ–‰์„ ๋– ๋‚ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]14. Iris Winnow journeys to Avalon Blufl, a town near the war front, with another correspondent named Thea "Attie" Attworod. [|assistant|]
์ฐจ๋ณ„ํ™” ๊ฐ€๋Šฅํ•œ ๊ฒ€์ƒ‰ ์ธ๋ฑ์ŠคTay ์™ธ(2022)์™€ Bevilacqua ์™ธ(2022)๋Š” Transformer ๋ชจ๋ธ ๋‚ด์— ๊ฒ€์ƒ‰ ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ํ…์ŠคํŠธ ์ฟผ๋ฆฌ๋ฅผ ๋ฌธ์„œ ์‹๋ณ„์ž์— ์ง์ ‘ ๋งคํ•‘ํ•  ์ˆ˜ ์žˆ๋Š” ์ฐจ๋ณ„ํ™” ๊ฐ€๋Šฅํ•œ ๊ฒ€์ƒ‰ ์ธ๋ฑ์Šค๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋” ํšจ์œจ์ ์ด๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๊ฒ€์ƒ‰์„ ์œ„ํ•œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ๊ณผ ์ž ์žฌ๋ ฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ฐจ๋ณ„ํ™” ๊ฐ€๋Šฅํ•œ ๊ฒ€์ƒ‰ ์ธ๋ฑ์ŠคTay ์™ธ(2022)์™€ Bevilacqua ์™ธ(2022)๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ ๋‚ด์— ๊ฒ€์ƒ‰ ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ํ…์ŠคํŠธ ์ฟผ๋ฆฌ๋ฅผ ๋ฌธ์„œ ์‹๋ณ„์ž์— ์ง์ ‘ ๋งคํ•‘ํ•  ์ˆ˜ ์žˆ๋Š” ์ฐจ๋ณ„ํ™” ๊ฐ€๋Šฅํ•œ ๊ฒ€์ƒ‰ ์ธ๋ฑ์Šค๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋” ํšจ์œจ์ ์ด๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๊ฒ€์ƒ‰์„ ์œ„ํ•œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ๊ณผ ์ž ์žฌ๋ ฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Differentiable Search IndicesTay et al. (2022) and Bevilacqua et al. (2022) developed differentiable search indices that integrate the retrieval process within a Transformer model, enabling direct mapping of text queries to document identifiers. These approaches offer superior performance and potential for more efficient and scalable retrieval. [|assistant|]
8. ๋กœ๋งŒ ํ‚คํŠธ๊ฐ€ ์•„์ด๋ฆฌ์Šค ์œˆ๋…ธ์šฐ์™€์˜ ์œ ๋Œ€๊ฐ์ด ๊นŠ์–ด์ง€๋ฉด์„œ ๋กœ๋งŒ์€ ํฌ๋ ˆ์ŠคํŠธ๋ฅผ ์œ„ํ•œ ์•„์ด๋ฆฌ์Šค์˜ ํŽธ์ง€๋ฅผ ์ต๋ช…์œผ๋กœ ๋ฐ›๊ณ  ์‘๋‹ตํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
8. ๋กœ๋งŒ ํ‚คํŠธ๊ฐ€ ์•„์ด๋ฆฌ์Šค ์œˆ๋…ธ์šฐ์™€์˜ ์œ ๋Œ€๊ฐ์ด ๊นŠ์–ด์ง€๋ฉด์„œ ๋กœ๋งŒ์€ ์ˆฒ์„ ์œ„ํ•œ ์•„์ด๋ฆฌ์Šค์˜ ํŽธ์ง€๋ฅผ ์ต๋ช…์œผ๋กœ ๋ฐ›๊ณ  ์‘๋‹ตํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]8. Iris Winnow's bond with Roman Kitt deepens when Roman anonymously receives and responds to Iris's letters intended for Forest. [|assistant|]
์งˆ๋ฌธ ๋‹ต๋ณ€ ์ƒ์„ฑ(QAG) ์ž‘์—…์„ ์œ„ํ•ด ํ›ˆ๋ จ๋œ ๋ชจ๋ธ, ํŠนํžˆ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์œผ๋กœ์„œ QA์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด๊ณ ๋œ End-to-End (E2E) ์ ‘๊ทผ ๋ฐฉ์‹์— ์ตœ์ ํ™”๋œ T5-large ๋ชจ๋ธ์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค(Ushio et al., 2023). ๊ทธ๋Ÿฌ๋‚˜ E2E ๋ฐฉ์‹์˜ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ๋‹จ์ ์€ ์ƒ์„ฑ๋œ ์ƒ˜ํ”Œ์˜ ํ’ˆ์งˆ๊ณผ ์–‘์„ ์ œ์–ดํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ T5 ๋ชจ๋ธ์˜ ์ž…๋ ฅ token ์ œํ•œ์œผ๋กœ ์ธํ•ด ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ 512 token์œผ๋กœ ์ œํ•œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด prompt ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์— ์‚ฌ์šฉ๋˜๋Š” ์ž…๋ ฅ์—๋„ ๋™์ผํ•œ token ์ œํ•œ์„ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 4๋Š” ๋™์ผํ•œ ์ฝ”ํผ์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ prompt ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•๊ณผ E2E ๋ฐฉ๋ฒ• ๋ชจ๋‘์—์„œ ์ƒ์„ฑ๋œ QA์— ๋Œ€ํ•œ ํ†ต๊ณ„๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์งˆ๋ฌธ ๋‹ต๋ณ€ ์ƒ์„ฑ(QAG) ์ž‘์—…์„ ์œ„ํ•ด ํ›ˆ๋ จ๋œ ๋ชจ๋ธ, ํŠนํžˆ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์œผ๋กœ์„œ QA์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด๊ณ ๋œ ์—”๋“œํˆฌ์—”๋“œ(E2E) ์ ‘๊ทผ ๋ฐฉ์‹์— ์ตœ์ ํ™”๋œ T5-large ๋ชจ๋ธ์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค(Ushio et al., 2023). ๊ทธ๋Ÿฌ๋‚˜ E2E ๋ฐฉ์‹์˜ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ๋‹จ์ ์€ ์ƒ์„ฑ๋œ ์ƒ˜ํ”Œ์˜ ํ’ˆ์งˆ๊ณผ ์–‘์„ ์ œ์–ดํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ T5 ๋ชจ๋ธ์˜ ์ž…๋ ฅ ํ† ํฐ ์ œํ•œ์œผ๋กœ ์ธํ•ด ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ 512 ํ† ํฐ์œผ๋กœ ์ œํ•œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด ํ”„๋กฌํ”„ํŠธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์— ์‚ฌ์šฉ๋˜๋Š” ์ž…๋ ฅ์—๋„ ๋™์ผํ•œ ํ† ํฐ ์ œํ•œ์„ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 4๋Š” ๋™์ผํ•œ ์ฝ”ํผ์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•๊ณผ E2E ๋ฐฉ๋ฒ• ๋ชจ๋‘์—์„œ ์ƒ์„ฑ๋œ QA์— ๋Œ€ํ•œ ํ†ต๊ณ„๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We employed models trained for the Question Answer Generation (QAG) task, specifically using the T5-large model optimized for the End-to-End (E2E) approach, which was reported to deliver superior performance on the QA as the downstream task (Ushio et al., 2023). However, a notable drawback of the E2E method is its limited control over the quality and quantity of generated samples. Additionally, due to the T5 model's input token limitation, the input text must be restricted to 512 tokens. To ensure a fair comparison, we applied the same token limit to the inputs used with the prompt-based approach. Table 4 presents statistics for the QAs generated by both the prompt and E2E methods, using the same corpus. [|assistant|]
์‹ค๋ฌด์ž๋ฅผ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด AI ์•ˆ์ „ ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ ํ…Œ์ŠคํŠธ ์‚ฌ์–‘์„ ๋งŒ๋“ค์—ˆ์œผ๋ฉฐ, ์ด๋ฅผ ์ž์œ ๋กญ๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ–ˆ์Šต๋‹ˆ๋‹ค.9 ์ด ํ…Œ์ŠคํŠธ ์‚ฌ์–‘์€ WG์˜ ๋งŽ์€ ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž ๊ทธ๋ฃน์— ์˜ํ•ด ๋งŒ๋“ค์–ด์ง€๊ณ  ๊ฒ€์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์‚ฌ์–‘์€ prompt ๊ณต์‹ํ™”, few-shot learning ๊ตฌ์„ฑ, chain-of-thought instructions์™€ ๊ฐ™์€ ์‚ฌ์†Œํ•œ ์„ค์ • ๋ณ€๊ฒฝ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ ๊ฒฐ๊ณผ์˜ ๋ฌด๊ฒฐ์„ฑ๊ณผ ๋ฏผ๊ฐ๋„๋ฅผ ๋‘˜๋Ÿฌ์‹ผ ์ง€์†์ ์ธ ๋ฌธ์ œ๋กœ ์ธํ•ด ๋งŒ๋“ค์–ด์กŒ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์š”์†Œ์™€ ๊ตฌ์„ฑ ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ ์ž˜ ๋ฌธ์„œํ™”๋˜์ง€ ์•Š์œผ๋ฉด SUT์˜ ์„ฑ๋Šฅ์— ์„ค๋ช…ํ•  ์ˆ˜ ์—†๋Š” ๋ณ€ํ™”๊ฐ€ ๋ฐœ์ƒํ•˜๊ณ  ์žฌํ˜„์„ฑ์ด ์ œํ•œ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ ์‚ฌ์–‘์€ ๋‘ ๊ฐ€์ง€ ๋ฐฉ์‹์œผ๋กœ ์‹ค๋ฌด์ž์—๊ฒŒ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ œ์•ˆ๋œ ํ…Œ์ŠคํŠธ์˜ ์ ์ ˆํ•œ ์‚ฌ์šฉ์„ ๋ฌธ์„œํ™”ํ•˜๊ณ  ํ…Œ์ŠคํŠธ๋ฅผ ๊ตฌํ˜„ํ•˜๊ฑฐ๋‚˜ ์‹คํ–‰ํ•˜๋ ค๋Š” ๋งŽ์€ ์ดํ•ด๊ด€๊ณ„์ž๋“ค ์‚ฌ์ด์—์„œ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ์žฌํ˜„์„ฑ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘˜์งธ, ์‚ฌ์–‘ ์Šคํ‚ค๋งˆ๋Š” ํ…Œ์ŠคํŠธ ๊ฒฐ๊ณผ์˜ ์ฒญ์ค‘์ด ์ฒ˜์Œ์— ๊ฒฐ๊ณผ๊ฐ€ ์–ด๋–ป๊ฒŒ ์ƒ์„ฑ๋˜์—ˆ๋Š”์ง€ ๋” ์ž˜ ์ดํ•ดํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์•ž์œผ๋กœ ๋” ๋งŽ์€ ์‚ฌ์–‘ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ์ž‘ํ•  ๊ณ„ํš์ž…๋‹ˆ๋‹ค.[|endofturn|]
์‹ค๋ฌด์ž๋ฅผ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด AI ์•ˆ์ „ ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ ํ…Œ์ŠคํŠธ ์‚ฌ์–‘์„ ๋งŒ๋“ค์—ˆ์œผ๋ฉฐ, ์ด๋ฅผ ์ž์œ ๋กญ๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ–ˆ์Šต๋‹ˆ๋‹ค.9 ์ด ํ…Œ์ŠคํŠธ ์‚ฌ์–‘์€ WG์˜ ๋งŽ์€ ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž ๊ทธ๋ฃน์— ์˜ํ•ด ๋งŒ๋“ค์–ด์ง€๊ณ  ๊ฒ€์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์‚ฌ์–‘์€ ํ”„๋กฌํ”„ํŠธ ๊ณต์‹ํ™”, ์†Œ์ˆ˜ ์ƒท ํ•™์Šต ๊ตฌ์„ฑ, ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ์ง€์‹œ์™€ ๊ฐ™์€ ์‚ฌ์†Œํ•œ ์„ค์ • ๋ณ€๊ฒฝ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ ๊ฒฐ๊ณผ์˜ ๋ฌด๊ฒฐ์„ฑ๊ณผ ๋ฏผ๊ฐ๋„๋ฅผ ๋‘˜๋Ÿฌ์‹ผ ์ง€์†์ ์ธ ๋ฌธ์ œ๋กœ ์ธํ•ด ๋งŒ๋“ค์–ด์กŒ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์š”์†Œ์™€ ๊ตฌ์„ฑ ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ ์ž˜ ๋ฌธ์„œํ™”๋˜์ง€ ์•Š์œผ๋ฉด SUT์˜ ์„ฑ๋Šฅ์— ์„ค๋ช…ํ•  ์ˆ˜ ์—†๋Š” ๋ณ€ํ™”๊ฐ€ ๋ฐœ์ƒํ•˜๊ณ  ์žฌํ˜„์„ฑ์ด ์ œํ•œ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ ์‚ฌ์–‘์€ ๋‘ ๊ฐ€์ง€ ๋ฐฉ์‹์œผ๋กœ ์‹ค๋ฌด์ž์—๊ฒŒ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ œ์•ˆ๋œ ํ…Œ์ŠคํŠธ์˜ ์ ์ ˆํ•œ ์‚ฌ์šฉ์„ ๋ฌธ์„œํ™”ํ•˜๊ณ  ํ…Œ์ŠคํŠธ๋ฅผ ๊ตฌํ˜„ํ•˜๊ฑฐ๋‚˜ ์‹คํ–‰ํ•˜๋ ค๋Š” ๋งŽ์€ ์ดํ•ด๊ด€๊ณ„์ž๋“ค ์‚ฌ์ด์—์„œ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ์žฌํ˜„์„ฑ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘˜์งธ, ์‚ฌ์–‘ ์Šคํ‚ค๋งˆ๋Š” ํ…Œ์ŠคํŠธ ๊ฒฐ๊ณผ์˜ ์ฒญ์ค‘์ด ์ฒ˜์Œ์— ๊ฒฐ๊ณผ๊ฐ€ ์–ด๋–ป๊ฒŒ ์ƒ์„ฑ๋˜์—ˆ๋Š”์ง€ ๋” ์ž˜ ์ดํ•ดํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์•ž์œผ๋กœ ๋” ๋งŽ์€ ์‚ฌ์–‘ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ์ž‘ํ•  ๊ณ„ํš์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To guide practitioners, we have created a test specification for the AI Safety Benchmark, and have made it freely available.9 The test specification was created and vetted by a large group of researchers and practitioners in the WG. Its creation was motivated by ongoing challenges around the integrity of performance results and their sensitivity to seemingly small setup changes, such as prompt formulation, few-shot learning configurations, and chain-of-thought instructions. If these factors and configuration parameters are not well-documented, this can lead to seemingly inexplicable variations in SUTs' performance and limit reproducibility. Our test specification can help practitioners in two ways. First, it can aid test writers to document proper usage of a proposed test and enable scalable reproducibility amongst a large group of stakeholders who may want to either implement or execute the test. Second, the specification schema can also help audiences of test results to better understand how those results were created in the first place. We aim to produce more specification resources in the future. [|assistant|]
Instruction์˜ ๊ฒฝ์šฐ, ๋‹ต๋ณ€ ์œ ํ˜•์— ๋”ฐ๋ผ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ๋งˆ๋‹ค ๋‹ค๋ฅธ Instruction์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ง€์นจ}์˜ ๊ฒฝ์šฐ, ๋‹ต๋ณ€ ์œ ํ˜•์— ๋”ฐ๋ผ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ๋งˆ๋‹ค ๋‹ค๋ฅธ ์ง€์นจ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]As for the {Instruction}, we use different instructions for different datasets based on the answer types. [|assistant|]
\(\bullet\)_LLaMA_[57]๋Š” CommonCrawl, C4 [82], Github, Wikipedia, ์ฑ…, ArXiv, StackExchange ๋“ฑ ๋‹ค์–‘ํ•œ ์†Œ์Šค์—์„œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. LLaMA(6B)์™€ LLaMA(13B)์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๋Š” 1.0T tokens์ด๊ณ , LLaMA(32B)์™€ LLaMA(65B)์—๋Š” 1.4T tokens์ด ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_LLaMA_[57]๋Š” CommonCrawl, C4 [82], Github, Wikipedia, ์ฑ…, ArXiv, StackExchange ๋“ฑ ๋‹ค์–‘ํ•œ ์†Œ์Šค์—์„œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. LLaMA(6B)์™€ LLaMA(13B)์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๋Š” 1.0T ํ† ํฐ์ด๊ณ , LLaMA(32B)์™€ LLaMA(65B)์—๋Š” 1.4T ํ† ํฐ์ด ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_LLaMA_[57] extracts training data from various sources, including CommonCrawl, C4 [82], Github, Wikipedia, books, ArXiv, and StackExchange. The training data size for LLaMA (6B) and LLaMA (13B) is 1.0T tokens, while 1.4T tokens are used for LLaMA (32B) and LLaMA (65B). [|assistant|]
์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต๊ณผ Fine-tuning์„ ํ†ตํ•ด LLaMA2-13B ๋ชจ๋ธ์„ ์ค‘๊ตญ์–ด ๋ฒˆ์ฒด๋กœ ์ ์šฉํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต๊ณผ ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด LLaMA2-13B ๋ชจ๋ธ์„ ์ค‘๊ตญ์–ด ๋ฒˆ์ฒด๋กœ ์ ์šฉํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We employ the following datasets for adapting the LLaMA2-13B model to Traditional Chinese through continual pretraining and fine-tuning. [|assistant|]
ํ‘œ 10์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, Nemotron-4 15B๋Š” LLaMA-2 13B์™€ Baichuan-2 13B๋ฅผ ๋ชจ๋‘ ํฌ๊ฒŒ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ, ๊ฐ๊ฐ 90.2%์™€ 44.1%์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. Nemotron-4 15B๋Š” ์ค‘๊ตญ์–ด์—์„œ ์˜์–ด๋กœ ๋ฒˆ์—ญํ•˜๋Š” ๋ฐ๋งŒ ๋Šฅ์ˆ™ํ•œ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์ค‘๊ตญ์–ด๋ฅผ ๋‹ค๋ฅธ ์–ธ์–ด๋กœ ์ง์ ‘ ๋ฒˆ์—ญํ•˜๋Š” ๋ฐ๋„ ์ธ์ƒ์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์€ Nemotron-4 15B๊ฐ€ ๊ด‘๋ฒ”์œ„ํ•œ ์ž์—ฐ์–ด์— ๊ฑธ์ณ ๊ฐ•๋ ฅํ•œ ์ดํ•ด๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
ํ‘œ 10์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ๋„ค๋ชจํŠธ๋ก -4 15B๋Š” LLaMA-2 13B์™€ ๋ฐ”์ด์ถ”์•ˆ-2 13B๋ฅผ ๋ชจ๋‘ ํฌ๊ฒŒ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ, ๊ฐ๊ฐ 90.2%์™€ 44.1%์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๋„ค๋ชจํŠธ๋ก -4 15B๋Š” ์ค‘๊ตญ์–ด์—์„œ ์˜์–ด๋กœ ๋ฒˆ์—ญํ•˜๋Š” ๋ฐ๋งŒ ๋Šฅ์ˆ™ํ•œ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์ค‘๊ตญ์–ด๋ฅผ ๋‹ค๋ฅธ ์–ธ์–ด๋กœ ์ง์ ‘ ๋ฒˆ์—ญํ•˜๋Š” ๋ฐ๋„ ์ธ์ƒ์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์€ ๋„ค๋ชจํŠธ๋ก -4 15B๊ฐ€ ๊ด‘๋ฒ”์œ„ํ•œ ์ž์—ฐ์–ด์— ๊ฑธ์ณ ๊ฐ•๋ ฅํ•œ ์ดํ•ด๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]As seen in Table 10, Nemotron-4 15B hefily outperforms both LLaMA-2 13B and Baichuan-2 13B - improving upon their performance by 90.2% and 44.1% respectively. Nemotron-4 15B does not solely perform well on translating from Chinese into English but is able to attain impressive results on the direct translation of Chinese into other languages. This ability highlights the strong understanding that Nemotron-4 15B has across a broad spectrum of natural languages. [|assistant|]
[์‹œ์Šคํ…œ] ์œ„์˜ ์งˆ๋ฌธ๊ณผ ํ‘œ์ค€ ๋‹ต๋ณ€์— ๋Œ€ํ•œ Assistant์˜ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•ด ์ฃผ์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค. ๋„์›€ ์œ ๋ฌด, ๊ด€๋ จ์„ฑ, ์ •ํ™•์„ฑ, ์„ธ๋ถ€ ์ˆ˜์ค€, {target language} ์‚ฌ์šฉ์— ๋Œ€ํ•ด Assistant๋ฅผ ํ‰๊ฐ€ํ•ด ์ฃผ์„ธ์š”. Assistant๋Š” 1~10์  ์ฒ™๋„๋กœ ์ด์ ์„ ๋ฐ›๊ฒŒ ๋˜๋ฉฐ, ์ ์ˆ˜๊ฐ€ ๋†’์„์ˆ˜๋ก ์ „๋ฐ˜์ ์ธ ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € Assistant์˜ ์ ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฐ’๋งŒ ํฌํ•จ๋œ ํ•œ ์ค„์„ ์ถœ๋ ฅํ•ด ์ฃผ์„ธ์š”. ๋‹ค์Œ ์ค„์—๋Š” ์ž ์žฌ์ ์ธ ํŽธ๊ฒฌ์„ ํ”ผํ•˜๊ณ  ๋‹ต๋ณ€ ์ˆœ์„œ๊ฐ€ ํŒ๋‹จ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋„๋ก ํ‰๊ฐ€์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์„ค๋ช…์„ ์ œ๊ณตํ•˜์„ธ์š”.[|endofturn|]
[์‹œ์Šคํ…œ] ์œ„์˜ ์งˆ๋ฌธ๊ณผ ํ‘œ์ค€ ๋‹ต๋ณ€์— ๋Œ€ํ•œ ์–ด์‹œ์Šคํ„ดํŠธ์˜ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•ด ์ฃผ์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค. ๋„์›€ ์œ ๋ฌด, ๊ด€๋ จ์„ฑ, ์ •ํ™•์„ฑ, ์„ธ๋ถ€ ์ˆ˜์ค€, {๋Œ€์ƒ ์–ธ์–ด} ์‚ฌ์šฉ์— ๋Œ€ํ•ด ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ํ‰๊ฐ€ํ•ด ์ฃผ์„ธ์š”. ์–ด์‹œ์Šคํ„ดํŠธ๋Š” 1~10์  ์ฒ™๋„๋กœ ์ด์ ์„ ๋ฐ›๊ฒŒ ๋˜๋ฉฐ, ์ ์ˆ˜๊ฐ€ ๋†’์„์ˆ˜๋ก ์ „๋ฐ˜์ ์ธ ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์–ด์‹œ์Šคํ„ดํŠธ์˜ ์ ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฐ’๋งŒ ํฌํ•จ๋œ ํ•œ ์ค„์„ ์ถœ๋ ฅํ•ด ์ฃผ์„ธ์š”. ๋‹ค์Œ ์ค„์—๋Š” ์ž ์žฌ์ ์ธ ํŽธ๊ฒฌ์„ ํ”ผํ•˜๊ณ  ๋‹ต๋ณ€ ์ˆœ์„œ๊ฐ€ ํŒ๋‹จ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋„๋ก ํ‰๊ฐ€์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์„ค๋ช…์„ ์ œ๊ณตํ•˜์„ธ์š”.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|][System] We would like you to provide feedback on the Assistant's performance regarding the above question and standard answer. Please rate the Assistant on helpfulness, relevance, accuracy, level of detail, and the use of { target language }. The Assistant will receive an overall score on a scale of 1 to 10, with higher scores indicating better overall performance. Please first output a single line containing only a value representing the Assistant's score. In the following line, please provide a comprehensive explanation of your evaluation, avoiding any potential biases and ensuring that the order of answers does not affect your judgment. [|assistant|]
ํ‘œ 2(b)์—์„œ ๊ด€์ฐฐํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ๋ชจ๋“  ๋ชจ๋ธ์ด baseline ์„ค์ •์—์„œ ์„ฑ๋Šฅ์ด ์ €์กฐํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ์ •ํ™•ํ•œ ์˜ˆ์ธก์„ ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์„ธ๋ถ€์ ์ธ ์ปจํ…์ŠคํŠธ์™€ ์ตœ์‹  ์ •๋ณด๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์ง๊ด€์ ์œผ๋กœ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์€ ๋‰ด์Šค retrieval์„ ํ†ตํ•ด ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ  ์ตœ์ ํ™”๋œ prompting ์ „๋žต๊ณผ Fine-tuning์„ ํ†ตํ•ด ๋” ๋‚˜์€ ์ถ”๋ก ์„ ์ด๋Œ์–ด๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
ํ‘œ 2(b)์—์„œ ๊ด€์ฐฐํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ๋ชจ๋“  ๋ชจ๋ธ์ด ๊ธฐ๋ณธ ์„ค์ •์—์„œ ์„ฑ๋Šฅ์ด ์ €์กฐํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ์ •ํ™•ํ•œ ์˜ˆ์ธก์„ ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์„ธ๋ถ€์ ์ธ ์ปจํ…์ŠคํŠธ์™€ ์ตœ์‹  ์ •๋ณด๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์ง๊ด€์ ์œผ๋กœ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์€ ๋‰ด์Šค ๊ฒ€์ƒ‰์„ ํ†ตํ•ด ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ  ์ตœ์ ํ™”๋œ ํ”„๋กฌํ”„ํŠธ ์ „๋žต๊ณผ ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด ๋” ๋‚˜์€ ์ถ”๋ก ์„ ์ด๋Œ์–ด๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]As observed in Table 2(b), all models perform poorly in the baseline setting. We intuit that models require detailed contexts and up-to-date information to make accurate forecasts. Our system addresses this issue via news retrieval and elicits better reasoning via optimized prompting strategies and fine-tuning. [|assistant|]
๊ทœ๋ชจ์— ๋”ฐ๋ฅธ ์šฐ์ˆ˜ํ•œ BLEU ์„ฑ๋Šฅ์˜ "emergence"๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ "emergence"๋Š” (Wei et al., 2022)์—์„œ ๋Œ€๊ทœ๋ชจ๋กœ ํ™•์ธ๋˜์—ˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ "emergence"์— ๋Œ€ํ•œ ์ž‘์—… ์ •์˜๋Š” "์ž‘์€ ๋ชจ๋ธ์—๋Š” ์กด์žฌํ•˜์ง€ ์•Š์ง€๋งŒ ํฐ ๋ชจ๋ธ์—๋Š” ๋‚˜ํƒ€๋‚จ"์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์˜๋ฏธ์—์„œ (Wei et al., 2022)๋Š” ๋‹ค์ค‘ ์ž‘์—… NLU, Modular arithmetic, ๋‹จ์–ด unscrambling ๋ฐ transliteration๊ณผ ๊ฐ™์€ ๋งŽ์€ ์ˆ˜์˜ "๊ธฐ์ˆ "์ด ๊ทœ๋ชจ์— ๋”ฐ๋ผ ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทœ๋ชจ์— ๋”ฐ๋ฅธ ์šฐ์ˆ˜ํ•œ BLEU ์„ฑ๋Šฅ์˜ "์ถœํ˜„"์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ "์ถœํ˜„"์€ (Wei et al., 2022)์—์„œ ๋Œ€๊ทœ๋ชจ๋กœ ํ™•์ธ๋˜์—ˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ "์ถœํ˜„"์— ๋Œ€ํ•œ ์ž‘์—… ์ •์˜๋Š” "์ž‘์€ ๋ชจ๋ธ์—๋Š” ์กด์žฌํ•˜์ง€ ์•Š์ง€๋งŒ ํฐ ๋ชจ๋ธ์—๋Š” ๋‚˜ํƒ€๋‚จ"์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์˜๋ฏธ์—์„œ (Wei et al., 2022)๋Š” ๋‹ค์ค‘ ์ž‘์—… NLU, ๋ชจ๋“ˆ์‹ ์‚ฐ์ˆ , ๋‹จ์–ด ์Šคํฌ๋žจ๋ธ”๋ง ๋ฐ ์Œ์—ญ๊ณผ ๊ฐ™์€ ๋งŽ์€ ์ˆ˜์˜ "๊ธฐ์ˆ "์ด ๊ทœ๋ชจ์— ๋”ฐ๋ผ ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]demonstrating "emergence" of good BLEU performance with scale. This type of "emergence" has been massively confirmed in (Wei et al., 2022), where a working definition of "emerging" is "not present in smaller models, but appears in larger models". In this sense, (Wei et al., 2022) demonstrate empirically a large number of "skills" appearing with scale, like Multi-Task NLU, Modular arithmetic, word unscrambling and transliteration. [|assistant|]
Huang ๋“ฑ์€ ๊ณต๊ฐ„ ํ† ํฐ์˜ 2์ฐจ์› ์ข…์†์„ฑ์„ ๋ณด์กดํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด ๋กœ์ปฌ ์Šค์บ๋‹ ์ „๋žต์„ ์ œ์•ˆํ•˜๋Š” LocalMamba [92]๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ImageNet ๋ถ„๋ฅ˜์—์„œ Viim-T๋ณด๋‹ค +3.1% ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Xu ๋“ฑ [93]์€ ์ œ์Šค์ฒ˜ ํ•ฉ์„ฑ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ๊ธธ๊ณ  ๋‹ค์–‘ํ•œ ์‹œํ€€์Šค๋ฅผ ์ง€์›ํ•˜๋Š” MambaTalk์ด๋ผ๋Š” SSM ๋ชจ๋ธ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. Pei ๋“ฑ [94]์€ ์ถ”๊ฐ€ ์ปจ๋ณผ๋ฃจ์…˜ ๋ธŒ๋žœ์น˜๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ํšจ์œจ์ ์ธ VAMba๋ฅผ ์ œ์•ˆํ•˜๊ณ  ImageNet-1K ๋ฐ COCO ๊ฐ์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ baseline์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. Du ๋“ฑ [95]์€ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์—์„œ VAMba์˜ robustness๋ฅผ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ „์ฒด ์ด๋ฏธ์ง€์™€ ํŒจ์น˜๋ณ„ ๋ฐฉ๋ฒ•์„ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜์—ฌ ์ ๋Œ€์  ๊ณต๊ฒฉ์— ๋Œ€ํ•œ ๋ณต์›๋ ฅ์„ ์กฐ์‚ฌํ•˜์—ฌ Transformer ์•„ํ‚คํ…์ฒ˜์— ๋น„ํ•ด ์šฐ์ˆ˜ํ•œ robustness๋ฅผ ๋ฐํ˜€๋ƒˆ์ง€๋งŒ scalability ์•ฝ์ ์ด ์žˆ์Œ์„ ๋ฐํ˜”์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค์— ๊ฑธ์นœ VAMba์˜ ์ผ๋ฐ˜์ ์ธ robustness๋ฅผ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. Shi ๋“ฑ [96]์€ ์••์ถ• ์ด๋ฏธ์ง€ ๋ณต์› ์ž‘์—…์— ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ๋ง์˜ ์„ ํ˜• ๋ณต์žก์„ฑ์„ ๋„์ž…ํ•˜์—ฌ ๊ธฐ์กด ๋ฐฉ๋ฒ•์˜ ๋ช‡ ๊ฐ€์ง€ ๋‹จ์ ์„ ๊ทน๋ณตํ•˜๋Š” ์ƒˆ๋กœ์šด ์ด๋ฏธ์ง€ ๋ณต์› ๋ฐฉ๋ฒ•์ธ VambaIR์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Fang ๋“ฑ [83]์€ ์„ ํƒ์  ๊ตฌ์กฐํ™”๋œ State Space Model (Mamba)๊ณผ ๋‹ค์ค‘ ์ธ์Šคํ„ด์Šค ํ•™์Šต(MIL) ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฒฐํ•ฉํ•œ ์ตœ์ดˆ์˜ ์ž‘์—…์ธ ์ „์ฒด ์Šฌ๋ผ์ด๋“œ ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด MamMI ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. MamMIL์€ ๋ถ„๋ฅ˜ ์„ฑ๋Šฅ๊ณผ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰ ์ธก๋ฉด์—์„œ Transformer ๊ธฐ๋ฐ˜์˜ ๊ธฐ์กด ์ตœ์ฒจ๋‹จ MIL ํ”„๋ ˆ์ž„์›Œํฌ๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. Li ๋“ฑ์€ ์‹ค์‹œ๊ฐ„ ๋ชจ๋ฐ”์ผ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ์ผ๋ฐ˜์ ์ธ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค ์ œ์•ฝ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์„ค๊ณ„๋œ ๊ฒฝ๋Ÿ‰ ์„ ํƒ์  State Space Model (SSM)์„ ํ™œ์šฉํ•˜๋Š” HARMamba [137]๋ผ๋Š” ์›จ์–ด๋Ÿฌ๋ธ” ์„ผ์„œ ์ธ์ฒด ํ™œ๋™ ์ธ์‹(HAR)์— ๋Œ€ํ•œ ์ƒˆ๋กœ์šด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. Yang ๋“ฑ์€ ์›๊ฒฉ ๊ฐ์ง€์—์„œ ์ดˆ๋ถ„๊ด‘ ์ด๋ฏธ์ง• ๋ฐ์ดํ„ฐ์˜ ๋ณต์žก์„ฑ๊ณผ ๊ณ ์ฐจ์›์  ํŠน์„ฑ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ HSIMamba [138]๋ผ๋Š” ์ƒˆ๋กœ์šด ์ดˆ๋ถ„๊ด‘ ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ์—๋Š” ํšจ์œจ์ ์œผ๋กœ ์ŠคํŽ™ํŠธ๋Ÿผ ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด ์–‘๋ฐฉํ–ฅ ์—ญ๋ฐฉํ–ฅ CNN๊ณผ ๊ณต๊ฐ„ ๋ถ„์„์„ ์œ„ํ•œ ํŠน์ˆ˜ ๋ธ”๋ก์ด ํ†ตํ•ฉ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
Huang ๋“ฑ์€ ๊ณต๊ฐ„ ํ† ํฐ์˜ 2์ฐจ์› ์ข…์†์„ฑ์„ ๋ณด์กดํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด ๋กœ์ปฌ ์Šค์บ๋‹ ์ „๋žต์„ ์ œ์•ˆํ•˜๋Š” LocalMamba [92]๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ImageNet ๋ถ„๋ฅ˜์—์„œ Viim-T๋ณด๋‹ค +3.1% ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Xu ๋“ฑ [93]์€ ์ œ์Šค์ฒ˜ ํ•ฉ์„ฑ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ๊ธธ๊ณ  ๋‹ค์–‘ํ•œ ์‹œํ€€์Šค๋ฅผ ์ง€์›ํ•˜๋Š” MambaTalk์ด๋ผ๋Š” SSM ๋ชจ๋ธ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. Pei ๋“ฑ [94]์€ ์ถ”๊ฐ€ ์ปจ๋ณผ๋ฃจ์…˜ ๋ธŒ๋žœ์น˜๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ํšจ์œจ์ ์ธ VAMba๋ฅผ ์ œ์•ˆํ•˜๊ณ  ImageNet-1K ๋ฐ COCO ๊ฐ์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ธฐ์ค€์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. Du ๋“ฑ [95]์€ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์—์„œ VAMba์˜ ๊ฒฌ๊ณ ์„ฑ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ „์ฒด ์ด๋ฏธ์ง€์™€ ํŒจ์น˜๋ณ„ ๋ฐฉ๋ฒ•์„ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜์—ฌ ์ ๋Œ€์  ๊ณต๊ฒฉ์— ๋Œ€ํ•œ ๋ณต์›๋ ฅ์„ ์กฐ์‚ฌํ•˜์—ฌ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜์— ๋น„ํ•ด ์šฐ์ˆ˜ํ•œ ๊ฒฌ๊ณ ์„ฑ์„ ๋ฐํ˜€๋ƒˆ์ง€๋งŒ ํ™•์žฅ์„ฑ ์•ฝ์ ์ด ์žˆ์Œ์„ ๋ฐํ˜”์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค์— ๊ฑธ์นœ VAMba์˜ ์ผ๋ฐ˜์ ์ธ ๊ฒฌ๊ณ ์„ฑ์„ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. Shi ๋“ฑ [96]์€ ์••์ถ• ์ด๋ฏธ์ง€ ๋ณต์› ์ž‘์—…์— ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ๋ง์˜ ์„ ํ˜• ๋ณต์žก์„ฑ์„ ๋„์ž…ํ•˜์—ฌ ๊ธฐ์กด ๋ฐฉ๋ฒ•์˜ ๋ช‡ ๊ฐ€์ง€ ๋‹จ์ ์„ ๊ทน๋ณตํ•˜๋Š” ์ƒˆ๋กœ์šด ์ด๋ฏธ์ง€ ๋ณต์› ๋ฐฉ๋ฒ•์ธ VambaIR์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Fang ๋“ฑ [83]์€ ์„ ํƒ์  ๊ตฌ์กฐํ™”๋œ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ(Mamba)๊ณผ ๋‹ค์ค‘ ์ธ์Šคํ„ด์Šค ํ•™์Šต(MIL) ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฒฐํ•ฉํ•œ ์ตœ์ดˆ์˜ ์ž‘์—…์ธ ์ „์ฒด ์Šฌ๋ผ์ด๋“œ ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด MamMI ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. MamMIL์€ ๋ถ„๋ฅ˜ ์„ฑ๋Šฅ๊ณผ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰ ์ธก๋ฉด์—์„œ Transformer ๊ธฐ๋ฐ˜์˜ ๊ธฐ์กด ์ตœ์ฒจ๋‹จ MIL ํ”„๋ ˆ์ž„์›Œํฌ๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. Li ๋“ฑ์€ ์‹ค์‹œ๊ฐ„ ๋ชจ๋ฐ”์ผ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ์ผ๋ฐ˜์ ์ธ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค ์ œ์•ฝ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์„ค๊ณ„๋œ ๊ฒฝ๋Ÿ‰ ์„ ํƒ์  ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ(SSM)์„ ํ™œ์šฉํ•˜๋Š” HARMamba [137]๋ผ๋Š” ์›จ์–ด๋Ÿฌ๋ธ” ์„ผ์„œ ์ธ์ฒด ํ™œ๋™ ์ธ์‹(HAR)์— ๋Œ€ํ•œ ์ƒˆ๋กœ์šด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. Yang ๋“ฑ์€ ์›๊ฒฉ ๊ฐ์ง€์—์„œ ์ดˆ๋ถ„๊ด‘ ์ด๋ฏธ์ง• ๋ฐ์ดํ„ฐ์˜ ๋ณต์žก์„ฑ๊ณผ ๊ณ ์ฐจ์›์  ํŠน์„ฑ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ HSIMamba [138]๋ผ๋Š” ์ƒˆ๋กœ์šด ์ดˆ๋ถ„๊ด‘ ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ์—๋Š” ํšจ์œจ์ ์œผ๋กœ ์ŠคํŽ™ํŠธ๋Ÿผ ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด ์–‘๋ฐฉํ–ฅ ์—ญ๋ฐฉํ–ฅ CNN๊ณผ ๊ณต๊ฐ„ ๋ถ„์„์„ ์œ„ํ•œ ํŠน์ˆ˜ ๋ธ”๋ก์ด ํ†ตํ•ฉ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Huang et al. introduce LocalMamba [92], which proposes a new local scanning strategy to preserve the two-dimensional dependencies of spatial tokens. They conduct extensive experiments on various tasks and demonstrate that LocalMamba improves over Viim-T by +3.1% on the ImageNet classification. Xu et al. [93] introduce an SSM model termed MambaTalk which focuses on gesture synthesis and supports long and various sequences. Pei et al. [94] proposes the EfficientVAMba by incorporating additional convolutional branches and further improves the baseline significantly on the ImageNet-1K and COCO detection datasets. Du et al. [95] explore the robustness of VAMba from various aspects, for example, they investigate the resilience to adversarial attacks using both whole-image and patch-specific methods, revealing superior robustness compared to Transformer architectures but with scalability weaknesses. They also assess VAMba's general robustness across diverse scenarios. Shi et al. [96] propose a new image restoration method, VambaIR, which overcomes some of the shortcomings of traditional methods by introducing the linear complexity of state-space modeling to a compressive image restoration task. Fang et al. [83] present the MamMI framework to address the classifying of whole slide images which is the first work to combine the selective structured State Space Model (Mamba) and a multi-instance learning (MIL) approach. MamMIL outperforms existing state-of-the-art MIL frameworks based on Transformer in terms of classification performance and memory usage. Li et al. introduce a novel approach to wearable sensor human activity recognition (HAR) called HARMamba [137], which utilizes a lightweight selective State Space Model (SSM) designed to address computational resource constraints typical of real-time mobile applications. Yang et al. introduce a novel hyperspectral image classification framework called HSIMamba [138], which aims to address the complexity and high-dimensional nature of hyperspectral imaging data in remote sensing. The proposed framework incorporates a bidirectional reversed CNN to efficiently extract spectral features, alongside a specialized block for spatial analysis. [|assistant|]
๊ทธ๋ฆผ 7: ๋ฐ˜๋ณต์  ์ •์ œ ์‹คํ—˜์„ ์œ„ํ•ด \(100_{0}00\) training ๊ฐ€์„ค์— ๋Œ€ํ•œ BLEU ์ ์ˆ˜. ๋Œ€๋ถ€๋ถ„์˜ training ์˜ˆ์ œ๋Š” BLEU ์ ์ˆ˜๊ฐ€ \(20\) ๋ฏธ๋งŒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 7: ๋ฐ˜๋ณต์  ์ •์ œ ์‹คํ—˜์„ ์œ„ํ•ด \(100_{0}00\) ํ›ˆ๋ จ ๊ฐ€์„ค์— ๋Œ€ํ•œ BLEU ์ ์ˆ˜. ๋Œ€๋ถ€๋ถ„์˜ ํ›ˆ๋ จ ์˜ˆ์ œ๋Š” BLEU ์ ์ˆ˜๊ฐ€ \(20\) ๋ฏธ๋งŒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 7: BLEU scores across the \(100_{0}00\) training hypotheses for our iterative refinement experiment. Most training examples have a BLEU score below \(20\). [|assistant|]
์‹œ๊ฐ์  ๋ฐ ์–ธ์–ด์  ์ •๋ณด ์œตํ•ฉ ์ธก๋ฉด์—์„œ Zhao ๋“ฑ[147]์€ Mamba ์–ธ์–ด ๋ชจ๋ธ์˜ ๋‚ด๋ถ€ ์ •๋ณด ํ†ตํ•ฉ์„ ์ตœ์ ํ™”ํ•˜์—ฌ ๋ณด๋‹ค ํšจ๊ณผ์ ์ธ ํ‘œํ˜„์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. Decision Mamba(DMamba) [148]๋Š” Mamba ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ Decision Transformer (DT) [152]์— ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. Decision Mamba์™€ DT๋ฅผ ๋น„๊ตํ•œ ์ผ๋ จ์˜ ์‹คํ—˜์„ ํ†ตํ•ด Mamba๊ฐ€ Reinforcement Learning (RL) ์ž‘์—…์— ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ €์ž๋“ค์ด ๊ณ ๋ คํ•œ RL ์ž‘์—…์—๋Š” CPU์™€ GPU์˜ ์ƒํ˜ธ ์ž‘์šฉ์ด ๋งŽ๊ธฐ ๋•Œ๋ฌธ์— Mamba ๋ธ”๋ก์„ DT์— ๋‹จ์ˆœํžˆ ์ ์šฉํ•ด๋„ ํšจ์œจ์„ฑ์ด ํ–ฅ์ƒ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ๊ฒฐํ•์ ์€ ํ•˜์ดํผ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ฒ€์ƒ‰๊ณผ RL ์ž‘์—…์˜ ๋ฐ์ดํ„ฐ ๊ตฌ์กฐ๋ฅผ ๋” ํšจ๊ณผ์ ์œผ๋กœ ๋ฐ˜์˜ํ•˜๊ธฐ ์œ„ํ•ด Mamba ๋ธ”๋ก์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ๋ถ„์„์ด ๋ถ€์žฌํ•œ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. Sigma [149]๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์‹œ๋งจํ‹ฑ ์„ธ๋ถ„ํ™”์— ์„ฑ๊ณต์ ์œผ๋กœ ์ ์šฉ๋œ ์ตœ์ดˆ์˜ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ Attention ๊ธฐ๋ฐ˜ Mamba ์œตํ•ฉ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ธ VMamba์™€ ์ฑ„๋„ ์ธ์‹ Mamba Decoder๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋‹ค์–‘ํ•œ ์‹คํ—˜์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Sigma๋Š” ๋” ๊ธด ์‹œํ€€์Šค๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ํ™œ์šฉ๋„๊ฐ€ ๋‚ฎ์œผ๋ฉฐ, Mamba Encoder์˜ ๋ฉ”๋ชจ๋ฆฌ ์†Œ๋น„๊ฐ€ ์ƒ๋Œ€์ ์œผ๋กœ ๋งŽ์•„ ๊ฒฝ๋Ÿ‰ ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค์— ๋ฐฐํฌํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค.[|endofturn|]
์‹œ๊ฐ์  ๋ฐ ์–ธ์–ด์  ์ •๋ณด ์œตํ•ฉ ์ธก๋ฉด์—์„œ Zhao ๋“ฑ[147]์€ ๋ง˜๋ฐ” ์–ธ์–ด ๋ชจ๋ธ์˜ ๋‚ด๋ถ€ ์ •๋ณด ํ†ตํ•ฉ์„ ์ตœ์ ํ™”ํ•˜์—ฌ ๋ณด๋‹ค ํšจ๊ณผ์ ์ธ ํ‘œํ˜„์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์˜์‚ฌ ๊ฒฐ์ • ๋ง˜๋ฐ”(DMamba) [148]๋Š” ์˜์‚ฌ ๊ฒฐ์ • ํŠธ๋žœ์Šคํฌ๋จธ(DT) [152]์— ๋ง˜๋ฐ” ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์˜์‚ฌ ๊ฒฐ์ • ๋ง˜๋ฐ”์™€ DT๋ฅผ ๋น„๊ตํ•œ ์ผ๋ จ์˜ ์‹คํ—˜์„ ํ†ตํ•ด ๋ง˜๋ฐ”๊ฐ€ ๊ฐ•ํ™” ํ•™์Šต(RL) ์ž‘์—…์— ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ €์ž๋“ค์ด ๊ณ ๋ คํ•œ RL ์ž‘์—…์—๋Š” CPU์™€ GPU์˜ ์ƒํ˜ธ ์ž‘์šฉ์ด ๋งŽ๊ธฐ ๋•Œ๋ฌธ์— Mamba ๋ธ”๋ก์„ DT์— ๋‹จ์ˆœํžˆ ์ ์šฉํ•ด๋„ ํšจ์œจ์„ฑ์ด ํ–ฅ์ƒ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ๊ฒฐํ•์ ์€ ํ•˜์ดํผ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ฒ€์ƒ‰๊ณผ RL ์ž‘์—…์˜ ๋ฐ์ดํ„ฐ ๊ตฌ์กฐ๋ฅผ ๋” ํšจ๊ณผ์ ์œผ๋กœ ๋ฐ˜์˜ํ•˜๊ธฐ ์œ„ํ•ด Mamba ๋ธ”๋ก์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ๋ถ„์„์ด ๋ถ€์žฌํ•œ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์‹œ๊ทธ๋งˆ [149]๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์‹œ๋งจํ‹ฑ ์„ธ๋ถ„ํ™”์— ์„ฑ๊ณต์ ์œผ๋กœ ์ ์šฉ๋œ ์ตœ์ดˆ์˜ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์ฃผ์˜๋ ฅ ๊ธฐ๋ฐ˜ ๋ง˜๋ฐ” ์œตํ•ฉ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ธ VMamba์™€ ์ฑ„๋„ ์ธ์‹ ๋ง˜๋ฐ” ๋””์ฝ”๋”๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋‹ค์–‘ํ•œ ์‹คํ—˜์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹œ๊ทธ๋งˆ๋Š” ๋” ๊ธด ์‹œํ€€์Šค๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ํ™œ์šฉ๋„๊ฐ€ ๋‚ฎ์œผ๋ฉฐ, ๋ง˜๋ฐ” ์ธ์ฝ”๋”์˜ ๋ฉ”๋ชจ๋ฆฌ ์†Œ๋น„๊ฐ€ ์ƒ๋Œ€์ ์œผ๋กœ ๋งŽ์•„ ๊ฒฝ๋Ÿ‰ ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค์— ๋ฐฐํฌํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In terms of visual and linguistic information fusion, Zhao et al. [147] optimizes the internal information integration of the Mamba language model to achieve a more effective expression. Decision Mamba(DMamba) [148] integrates the Mamba framework into the Decision Transformer (DT) [152]. A series of experiments comparing Decision Mamba and DT show that Mamba is effective in reinforcement learning (RL) tasks. But simply applying Mamba blocks to DT does not improve efficiency, because the RL tasks considered by the authors have a large number of CPU and GPU interactions. Another deficiency is the absence of a hyper-parameter search and an analysis of how to use the Mamba block more effectively to reflect the data structure of RL tasks. Sigma [149] is the first State Space Model successfully applied in multi-modal semantic segmentation. It is composed of VMamba, an attention-based Mamba fusion mechanism and a channel-aware Mamba decoder, and has shown excellent performance in various experiments. However, Sigma is underutilized in handling longer sequences, and the memory consumption of Mamba encoders is still relatively large, making it difficult to deploy on lightweight edge devices. [|assistant|]
์‹œ๊ฐ์  ์žฅ์—. ์—ฐ์†์ ์ธ 2D ์Šค์บ๋‹ ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ด ๊ณต๊ฐ„ ์—ฐ์†์„ฑ์„ ๊ฐœ์„ ํ•˜๊ณ  ๋ฐฉํ–ฅ ์ง€๊ฐ์„ ์—…๋ฐ์ดํŠธํ•จ์œผ๋กœ์จ ๋ชจ๋ธ์€ ๋ฐฉํ–ฅ ์ •๋ณด๋ฅผ ์ธ์ฝ”๋”ฉํ•˜์—ฌ ๋ ˆ์ด๋ธ”์˜ ๊ณต๊ฐ„ ๊ด€๊ณ„๋ฅผ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ 2D ์ด๋ฏธ์ง€์—์„œ ํŠน์ง•์„ ํ•™์Šตํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Wang ๋“ฑ์€ ๊ณค์ถฉ ๋ถ„๋ฅ˜ ์ž‘์—…์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” InsectMamba [112]๋ฅผ ์ œ์•ˆํ•˜๊ณ , State Space Model, Convolutional Neural Network (CNN), multi-head self-attention mechanism (MSA), ๋‹ค์ธต ํผ์…‰ํŠธ๋ก (MLP)์„ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋“ˆ(Mix-SSM)์— ํ†ตํ•ฉํ•˜์—ฌ ๋ชจ๋ธ์˜ ๋ถ„๋ฅ˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.[|endofturn|]
์‹œ๊ฐ์  ์žฅ์—. ์—ฐ์†์ ์ธ 2D ์Šค์บ๋‹ ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ด ๊ณต๊ฐ„ ์—ฐ์†์„ฑ์„ ๊ฐœ์„ ํ•˜๊ณ  ๋ฐฉํ–ฅ ์ง€๊ฐ์„ ์—…๋ฐ์ดํŠธํ•จ์œผ๋กœ์จ ๋ชจ๋ธ์€ ๋ฐฉํ–ฅ ์ •๋ณด๋ฅผ ์ธ์ฝ”๋”ฉํ•˜์—ฌ ๋ ˆ์ด๋ธ”์˜ ๊ณต๊ฐ„ ๊ด€๊ณ„๋ฅผ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ 2D ์ด๋ฏธ์ง€์—์„œ ํŠน์ง•์„ ํ•™์Šตํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Wang ๋“ฑ์€ ๊ณค์ถฉ ๋ถ„๋ฅ˜ ์ž‘์—…์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” InsectMamba [112]๋ฅผ ์ œ์•ˆํ•˜๊ณ , ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ, ์ปจ๋ณผ๋ฃจ์…˜ ์‹ ๊ฒฝ๋ง(CNN), ๋‹ค์ค‘ ํ—ค๋“œ ์ž๊ธฐ ์ฃผ์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜(MSA), ๋‹ค์ธต ํผ์…‰ํŠธ๋ก (MLP)์„ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋“ˆ(Mix-SSM)์— ํ†ตํ•ฉํ•˜์—ฌ ๋ชจ๋ธ์˜ ๋ถ„๋ฅ˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]to the visual field. By improving spatial continuity through the continuous 2D scanning process and updating direction perception, the model can distinguish spatial relationships of labels by encoding direction information, thereby enhancing its ability to learn features from 2D images. Wang et al. propose InsectMamba [112] that can be used in the insect classification task and improve the classification ability of the model by integrating a State Space Model, a convolutional neural network (CNN), a multi-head self-attention mechanism (MSA), and multi-layer perceptrons (MLPs) in a hybrid state space module (Mix-SSM). [|assistant|]
์ด ์ž‘์—…์˜ ๋˜ ๋‹ค๋ฅธ ํ•œ๊ณ„๋Š” ์ด์ƒ์ ์ธ ์–ธ์–ด ๋ชจ๋ธ๋ง ํ—ค๋“œ์˜ ์ฐจ์›๊ณผ ์ˆœ์œ„ ์ œํ•œ ์„ฑ๋Šฅ ๊ฐ„์— ์„ค์ •ํ•œ ์ˆ˜ํ•™์  ์—ฐ๊ด€์„ฑ์ด ๋Š์Šจํ•˜๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค(cf. ์ •๋ฆฌ 5.3). ๋˜ํ•œ _์ด์ƒ_\(x_{i}^{*}\) ํ‘œํ˜„์„ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์€ ์ž˜ ์ •์˜๋˜์ง€ ์•Š์€ ๊ฐœ๋…์ด๋ผ๊ณ  ์ฃผ์žฅํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ •๋ฆฌ 5.3์˜ ์ถ”๋ก ์ด ๋ฐ๋ชจ์—์„œ _์ด์ƒ์ _ ํŠน์„ฑ์ด ํ•„์š”ํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ๋ชจ๋“  ์ปจํ…์ŠคํŠธ ํ‘œํ˜„์— ์ ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. '์ด์ƒ์ '์ด๋ผ๋Š” ๋‹จ์–ด๋Š” ์ฃผ์–ด์ง„ ํ›ˆ๋ จ ์„ธํŠธ \(\mathcal{T}\)์— ๋Œ€ํ•ด ์ด๋Ÿฌํ•œ ํ‘œํ˜„์ด \(W^{*}\) ํ–‰๋ ฌ์— ๋ถ€๊ณผํ•˜๋Š” ๊ตฌ์กฐ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๋Š” ์ •๋„๊นŒ์ง€, _๋ชจ๋“  ๊ธฐ๋ณธ ๋ชจ๋ธ_์—์„œ ์–ป์€ \(x_{i}^{*}\) ํ‘œํ˜„์— ๋Œ€ํ•ด ์šฐ๋ฆฌ์˜ ๊ด€์ฐฐ์ด ์œ ์ง€๋จ์„ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ด ์ž‘์—…์˜ ๋˜ ๋‹ค๋ฅธ ํ•œ๊ณ„๋Š” ์ด์ƒ์ ์ธ ์–ธ์–ด ๋ชจ๋ธ๋ง ํ—ค๋“œ์˜ ์ฐจ์›๊ณผ ์ˆœ์œ„ ์ œํ•œ ์„ฑ๋Šฅ ๊ฐ„์— ์„ค์ •ํ•œ ์ˆ˜ํ•™์  ์—ฐ๊ด€์„ฑ์ด ๋Š์Šจํ•˜๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค(cf. ์ •๋ฆฌ 5.3). ๋˜ํ•œ _์ด์ƒ_\(x_{i}^{*}\) ํ‘œํ˜„์„ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์€ ์ž˜ ์ •์˜๋˜์ง€ ์•Š์€ ๊ฐœ๋…์ด๋ผ๊ณ  ์ฃผ์žฅํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ •๋ฆฌ 5.3์˜ ์ถ”๋ก ์ด ๋ฐ๋ชจ์—์„œ _์ด์ƒ์ _ ํŠน์„ฑ์ด ํ•„์š”ํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ๋ชจ๋“  ์ปจํ…์ŠคํŠธ ํ‘œํ˜„์— ์ ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. '์ด์ƒ์ '์ด๋ผ๋Š” ๋‹จ์–ด๋Š” ์ฃผ์–ด์ง„ ํ›ˆ๋ จ ์„ธํŠธ \mathcal{T}\)์— ๋Œ€ํ•ด ์ด๋Ÿฌํ•œ ํ‘œํ˜„์ด \(W^{*}\) ํ–‰๋ ฌ์— ๋ถ€๊ณผํ•˜๋Š” ๊ตฌ์กฐ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๋Š” ์ •๋„๊นŒ์ง€, _๋ชจ๋“  ๊ธฐ๋ณธ ๋ชจ๋ธ_์—์„œ ์–ป์€ \(x_{i}^{*}\) ํ‘œํ˜„์— ๋Œ€ํ•ด ์šฐ๋ฆฌ์˜ ๊ด€์ฐฐ์ด ์œ ์ง€๋จ์„ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Another limitation of this work is the loose nature of the mathematical connection that we establish between the dimensionality of the ideal language modeling head and the rank-constrained performance (cf. Theorem 5.3). Moreover, it can also be argued that considering _ideal_\(x_{i}^{*}\) representations is an ill-defined notion. We argue that the reasoning behind Theorem 5.3 could be applied to any contextual representations, as the _ideal_ nature of \(x_{i}^{*}\) is not necessary in the demonstrations. The word _ideal_ reflects that our observations hold for \(x_{i}^{*}\) representations obtained from _any underlying model_, to an extent that depends on the structure that these representations impose on the \(W^{*}\) matrix for a given training set \(\mathcal{T}\). [|assistant|]
**์ƒˆ๋กœ์šด ์ •๋ณด ํ•™์Šต** ๊ทธ๋ฆผ 3์—์„œ ์šฐ๋ฆฌ๋Š” ์‹คํ—˜ ์ „๋ฐ˜์—์„œ ๊ด€์ฐฐ๋œ ํฅ๋ฏธ๋กœ์šด ํ˜„์ƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ epoch, ์ฆ‰ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ๋˜ ๋‹ค๋ฅธ ๋ฐ˜๋ณต์„ ์™„๋ฃŒํ•œ ํ›„ training loss๊ฐ€ ํฌ๊ฒŒ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ›ˆ๋ จ ์ค‘์— ๋ฐ์ดํ„ฐ๋ฅผ ์•”๊ธฐํ•˜๊ณ  ๊ณผ์ ํ•ฉํ•˜๋Š” LLM์— ๋Œ€ํ•ด ์•Œ๋ ค์ง„ ๊ฒƒ๊ณผ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค Tirumala et al. (2022).[|endofturn|]
**์ƒˆ๋กœ์šด ์ •๋ณด ํ•™์Šต** ๊ทธ๋ฆผ 3์—์„œ ์šฐ๋ฆฌ๋Š” ์‹คํ—˜ ์ „๋ฐ˜์—์„œ ๊ด€์ฐฐ๋œ ํฅ๋ฏธ๋กœ์šด ํ˜„์ƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์—ํฌํฌ, ์ฆ‰ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ๋˜ ๋‹ค๋ฅธ ๋ฐ˜๋ณต์„ ์™„๋ฃŒํ•œ ํ›„ ํ›ˆ๋ จ ์†์‹ค์ด ํฌ๊ฒŒ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ›ˆ๋ จ ์ค‘์— ๋ฐ์ดํ„ฐ๋ฅผ ์•”๊ธฐํ•˜๊ณ  ๊ณผ์ ํ•ฉํ•˜๋Š” LLM์— ๋Œ€ํ•ด ์•Œ๋ ค์ง„ ๊ฒƒ๊ณผ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค Tirumala et al. (2022).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Learning New Information** In Figure 3, we can see an interesting phenomenon observed throughout our experiments. After each epoch, i.e., completing another iteration over the entire dataset, the training loss drops significantly. This is consistent with what is known about LLMs memorizing the data during training and overfitting Tirumala et al. (2022). [|assistant|]
- ์˜ˆ์ธก์— ์˜ํ–ฅ์„ ์ค„ ์ˆ˜ ์žˆ๋Š” Google์—์„œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ„๋‹จํ•œ ๊ฒ€์ƒ‰์–ด(๊ฐ๊ฐ ์ตœ๋Œ€ {max_words} ๋‹จ์–ด)๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
- ์˜ˆ์ธก์— ์˜ํ–ฅ์„ ์ค„ ์ˆ˜ ์žˆ๋Š” Google์—์„œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ„๋‹จํ•œ ๊ฒ€์ƒ‰์–ด(๊ฐ๊ฐ ์ตœ๋Œ€ {์ตœ๋Œ€_๋‹จ์–ด} ๋‹จ์–ด)๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]- Generate brief search queries (up to {max_words} words each) to gather information on Google that could influence the forecast. [|assistant|]
์ตœ๊ทผ ๋“ค์–ด Large Language Models (LLMs) [1; 2; 3; 4; 5; 6; 7; 8; 9]์ด ํฐ ์ฃผ๋ชฉ์„ ๋ฐ›๊ณ  ์žˆ์œผ๋ฉฐ ์ธ์ƒ์ ์ธ ์„ฑ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ OpenAI์˜ ChatGPT๊ฐ€ ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ ๊ผฝํž™๋‹ˆ๋‹ค. ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์ธํ„ฐ๋„ท ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‚ฌ์ „ ํ•™์Šต๊ณผ ์ƒ์„ธํ•œ instruction data [10]๋ฅผ ํ†ตํ•ด ์ถ”๊ฐ€ Fine-tuning์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ state-of-the-art (SOTA) zero-shot ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถ”์„ธ๋Š” ์ฝ”๋“œ ์ดํ•ด ๋ฐ ์ƒ์„ฑ ์˜์—ญ์—์„œ๋„ ๊ด€์ฐฐ๋ฉ๋‹ˆ๋‹ค. ์ˆ˜๋งŽ์€ Code LLMs [11; 12; 13; 14; 15; 16; 17; 18]์ด ์ฝ”๋“œ ๊ด€๋ จ ์ž‘์—…๊ณผ ๊ด€๋ จ๋œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ Code LLMs์€ ์ƒ๋‹นํ•œ ์–‘์˜ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์น˜๋ฏ€๋กœ ๋‹ค์–‘ํ•œ ์ฝ”๋“œ ๊ด€๋ จ ์ž‘์—…์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ ์ธ์ƒ์ ์ธ ์„ฑ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
์ตœ๊ทผ ๋“ค์–ด ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM) [1; 2; 3; 4; 5; 6; 7; 8; 9]์ด ํฐ ์ฃผ๋ชฉ์„ ๋ฐ›๊ณ  ์žˆ์œผ๋ฉฐ ์ธ์ƒ์ ์ธ ์„ฑ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ OpenAI์˜ ChatGPT๊ฐ€ ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ ๊ผฝํž™๋‹ˆ๋‹ค. ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์ธํ„ฐ๋„ท ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‚ฌ์ „ ํ•™์Šต๊ณผ ์ƒ์„ธํ•œ ์ง€์นจ ๋ฐ์ดํ„ฐ [10]๋ฅผ ํ†ตํ•ด ์ถ”๊ฐ€ ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์ตœ์ฒจ๋‹จ(SOTA) ์ œ๋กœ ์ƒท ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถ”์„ธ๋Š” ์ฝ”๋“œ ์ดํ•ด ๋ฐ ์ƒ์„ฑ ์˜์—ญ์—์„œ๋„ ๊ด€์ฐฐ๋ฉ๋‹ˆ๋‹ค. ์ฝ”๋“œ ๊ด€๋ จ ์ž‘์—…๊ณผ ๊ด€๋ จ๋œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ˆ˜๋งŽ์€ ์ฝ”๋“œ LLM [11; 12; 13; 14; 15; 16; 17; 18]์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฝ”๋“œ LLM์€ ์ƒ๋‹นํ•œ ์–‘์˜ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์น˜๋ฏ€๋กœ ๋‹ค์–‘ํ•œ ์ฝ”๋“œ ๊ด€๋ จ ์ž‘์—…์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ ์ธ์ƒ์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Recently, Large Language Models (LLMs) [1; 2; 3; 4; 5; 6; 7; 8; 9] have garnered significant attention and demonstrated impressive success. Notably, OpenAI's ChatGPT stands out as a prominent example. Leveraging extensive pre-training on vast amounts of internet data and further fine-tuning with detailed instruction data [10], these models have achieved state-of-the-art (SOTA) zero-shot performance across diverse tasks. This trend is also observed in the domain of code understanding and generation. Numerous Code LLMs [11; 12; 13; 14; 15; 16; 17; 18] have been proposed to tackle the challenges associated with code-related tasks. These Code LLMs undergo pre-training using substantial amounts of code data, enabling them to excel in various code-related tasks, showcasing impressive performance. [|assistant|]
Autoregressive Language Modeling ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ \(n\) ํ† ํฐ \(x_{1}\),..., \(x_{n}\)์˜ ์‹œํ€€์Šค๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ๋ชจ๋ธ์€ ์ž๋™ ํšŒ๊ท€ ๋ฐฉ์‹์œผ๋กœ ๋‹ค์Œ ํ† ํฐ \(x_{n+1}\)(๋•Œ๋กœ๋Š” ๋‹ค์Œ ํ† ํฐ ์‹œํ€€์Šค)๋ฅผ ์˜ˆ์ธกํ•˜๋ ค๊ณ  ์‹œ๋„ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ž๋™ ํšŒ๊ท€ ์–ธ์–ด ๋ชจ๋ธ๋ง** ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ \(n\) ํ† ํฐ \(x_{1}\),..., \(x_{n}\)์˜ ์‹œํ€€์Šค๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ๋ชจ๋ธ์€ ์ž๋™ ํšŒ๊ท€ ๋ฐฉ์‹์œผ๋กœ ๋‹ค์Œ ํ† ํฐ \(x_{n+1}\)(๋•Œ๋กœ๋Š” ๋‹ค์Œ ํ† ํฐ ์‹œํ€€์Šค)๋ฅผ ์˜ˆ์ธกํ•˜๋ ค๊ณ  ์‹œ๋„ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In **Autoregressive Language Modeling** framework, given a sequence of \(n\) tokens \(x_{1}\),..., \(x_{n}\), the model tries to predict next token \(x_{n+1}\) (and sometimes next sequence of tokens) in an auto-regressive fashion. [|assistant|]
\(\bullet\)_์ž‘์—… ๋ชฉํ‘œ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ํ‘œํ˜„ํ•ฉ๋‹ˆ๋‹ค._ ์ž‘์—… ์„ค๋ช…์€ ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ๋ถˆ๋ถ„๋ช…ํ•ด์„œ๋Š” ์•ˆ ๋˜๋ฉฐ, ์ด๋กœ ์ธํ•ด ๋ถ€์ •ํ™•ํ•˜๊ฑฐ๋‚˜ ๋ถ€์ ์ ˆํ•œ ์‘๋‹ต์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ํ™œ์šฉํ•  ๋•Œ ๋ช…ํ™•ํ•˜๊ณ  ๋ชจํ˜ธํ•˜์ง€ ์•Š์€ ์ง€์นจ์ด ํ•„์š”ํ•˜๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค[66]. ๋ช…ํ™•ํ•˜๊ณ  ์ƒ์„ธํ•œ ์„ค๋ช…์—๋Š” ์ž‘์—… ๋ชฉํ‘œ, ์ž…๋ ฅ/์ถœ๋ ฅ ๋ฐ์ดํ„ฐ(_์˜ˆ: "๊ธด ๋ฌธ์„œ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ๊ฐ„๊ฒฐํ•œ ์š”์•ฝ์„ ์ƒ์„ฑํ•ด ์ฃผ์„ธ์š”."_), ์‘๋‹ต ์ œ์•ฝ ์กฐ๊ฑด(_์˜ˆ: "์š”์•ฝ์˜ ๊ธธ์ด๋Š” 50์„ ์ดˆ๊ณผํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค."_) ๋“ฑ ์ž‘์—…์„ ์„ค๋ช…ํ•˜๋Š” ๋‹ค์–‘ํ•œ ์š”์†Œ๊ฐ€ ํฌํ•จ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ช…ํ™•ํ•˜๊ฒŒ ์„ค๋ช…๋œ ์ž‘์—… ์„ค๋ช…์„ ์ œ๊ณตํ•˜๋ฉด LLMs์ด ๋ชฉํ‘œ ์ž‘์—…์„ ๋ณด๋‹ค ํšจ๊ณผ์ ์œผ๋กœ ์ดํ•ดํ•˜๊ณ  ์›ํ•˜๋Š” ๊ฒฐ๊ณผ๋ฌผ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_์ž‘์—… ๋ชฉํ‘œ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ํ‘œํ˜„ํ•ฉ๋‹ˆ๋‹ค._ ์ž‘์—… ์„ค๋ช…์€ ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ๋ถˆ๋ถ„๋ช…ํ•ด์„œ๋Š” ์•ˆ ๋˜๋ฉฐ, ์ด๋กœ ์ธํ•ด ๋ถ€์ •ํ™•ํ•˜๊ฑฐ๋‚˜ ๋ถ€์ ์ ˆํ•œ ์‘๋‹ต์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ํ™œ์šฉํ•  ๋•Œ ๋ช…ํ™•ํ•˜๊ณ  ๋ชจํ˜ธํ•˜์ง€ ์•Š์€ ์ง€์นจ์ด ํ•„์š”ํ•˜๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค[66]. ๋ช…ํ™•ํ•˜๊ณ  ์ƒ์„ธํ•œ ์„ค๋ช…์—๋Š” ์ž‘์—… ๋ชฉํ‘œ, ์ž…๋ ฅ/์ถœ๋ ฅ ๋ฐ์ดํ„ฐ(_์˜ˆ: "๊ธด ๋ฌธ์„œ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ๊ฐ„๊ฒฐํ•œ ์š”์•ฝ์„ ์ƒ์„ฑํ•ด ์ฃผ์„ธ์š”."_), ์‘๋‹ต ์ œ์•ฝ ์กฐ๊ฑด(_์˜ˆ: "์š”์•ฝ์˜ ๊ธธ์ด๋Š” 50์„ ์ดˆ๊ณผํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค."_) ๋“ฑ ์ž‘์—…์„ ์„ค๋ช…ํ•˜๋Š” ๋‹ค์–‘ํ•œ ์š”์†Œ๊ฐ€ ํฌํ•จ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ช…ํ™•ํ•˜๊ฒŒ ์„ค๋ช…๋œ ์ž‘์—… ์„ค๋ช…์„ ์ œ๊ณตํ•˜๋ฉด LLM์ด ๋ชฉํ‘œ ์ž‘์—…์„ ๋ณด๋‹ค ํšจ๊ณผ์ ์œผ๋กœ ์ดํ•ดํ•˜๊ณ  ์›ํ•˜๋Š” ๊ฒฐ๊ณผ๋ฌผ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_Expressing the task goal clearly._ Task descriptions should not be ambiguous or unclear, which likely lead to inaccurate or inappropriate responses. This highlights the need for clear and unambiguous directives when utilizing these models [66]. A clear and detailed description should contain various elements to explain a task, including task objective, input/output data (_e.g._, _"Given a long document, I want you to generate a concise summary."_), and the response constraints (_e.g._, _"the length of the summary cannot exceed 50."_). By providing a well-clarified task description, LLMs can more effectively understand the target task and generate the desired output. [|assistant|]
๋˜ํ•œ, ๊ทธ๋ฆผ 7์˜ power law๋ฅผ ํ†ตํ•ด ์„ ํƒ๋œ ํ† ํฐ์˜ loss๋ฅผ downstream task ์„ฑ๋Šฅ๊ณผ ์—ฐ๊ด€์‹œํ‚ค๋Š”๋ฐ, ์ด๋Š” ๋™์‹œ ์—ฐ๊ตฌ(Gadre et al., 2024)์™€ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ž˜ํ”„์˜ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์—์„œ ํ”ผํŒ…ํ•œ ๊ณก์„ ์„ ๊ด€์ฐฐํ•˜๋ฉด, SLM์ด ์„ ํƒํ•œ ํ† ํฐ์˜ ํ‰๊ท  loss๋Š” downstream task์˜ ์„ฑ๋Šฅ๊ณผ ์–‘์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ด๋Š” ๋ฐ˜๋ฉด, ์„ ํƒ๋˜์ง€ ์•Š์€ ํ† ํฐ์˜ ํ‰๊ท  loss๋Š” downstream task ์„ฑ๋Šฅ๊ณผ ์Œ์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ชจ๋ธ์˜ ๊ถ๊ทน์ ์ธ ์„ฑ๋Šฅ์— ๋„์›€์ด ๋˜๊ธฐ ์œ„ํ•ด ๋ชจ๋“  ํ† ํฐ์˜ loss๊ฐ€ ๊ฐ์†Œํ•  ํ•„์š”๋Š” ์—†์Šต๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก D๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.[|endofturn|]
๋˜ํ•œ, ๊ทธ๋ฆผ 7์˜ ๊ฑฐ๋“ญ์ œ๊ณฑ ๋ฒ•์น™์„ ํ†ตํ•ด ์„ ํƒ๋œ ํ† ํฐ์˜ ์†์‹ค์„ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—… ์„ฑ๋Šฅ๊ณผ ์—ฐ๊ด€์‹œํ‚ค๋Š”๋ฐ, ์ด๋Š” ๋™์‹œ ์—ฐ๊ตฌ(Gadre et al., 2024)์™€ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ž˜ํ”„์˜ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์—์„œ ํ”ผํŒ…ํ•œ ๊ณก์„ ์„ ๊ด€์ฐฐํ•˜๋ฉด, SLM์ด ์„ ํƒํ•œ ํ† ํฐ์˜ ํ‰๊ท  ์†์‹ค์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์˜ ์„ฑ๋Šฅ๊ณผ ์–‘์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ด๋Š” ๋ฐ˜๋ฉด, ์„ ํƒ๋˜์ง€ ์•Š์€ ํ† ํฐ์˜ ํ‰๊ท  ์†์‹ค์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—… ์„ฑ๋Šฅ๊ณผ ์Œ์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ชจ๋ธ์˜ ๊ถ๊ทน์ ์ธ ์„ฑ๋Šฅ์— ๋„์›€์ด ๋˜๊ธฐ ์œ„ํ•ด ๋ชจ๋“  ํ† ํฐ์˜ ์†์‹ค์ด ๊ฐ์†Œํ•  ํ•„์š”๋Š” ์—†์Šต๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก D๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Moreover, We relate the selected tokens' loss to its downstream task performance via a power law in Figure 7, which is similar to a concurrent study (Gadre et al., 2024). Observing the curve fitted from the data points in the graph, the average loss of tokens selected by the SLM shows a positive correlation with the performance of downstream tasks, whereas the average loss of tokens not selected exhibits a negative correlation with downstream task performance. Therefore, it is not necessary for the all tokens' loss to decrease to benefit the model's ultimate performance. See Appendix D for more details. [|assistant|]
์—ฐ์‚ฐ์ž \(f\)๋Š” ๋‹จ์ˆœํžˆ ๋ณด์ƒ ํ•จ์ˆ˜๋ฅผ \(\pi_{r}\)์˜ ๋ถ„ํ•  ํ•จ์ˆ˜์˜ ๋กœ๊ทธ๋กœ ์ •๊ทœํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ถ”๊ฐ€๋œ ์ •๊ทœํ™” ํ•ญ์€ ์ ‘๋‘์‚ฌ \(x\)์˜ ํ•จ์ˆ˜์ด๋ฏ€๋กœ \(f(r;\pi_{\text{ref}},\beta)(x,y)\)๋Š” \(r(x,y)\)์˜ ๋“ฑ๊ฐ€ ํด๋ž˜์Šค์—์„œ ๋ณด์ƒ ํ•จ์ˆ˜์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, \(r\)์„ Eq. 5์˜ RHS(๋ชจ๋“  ๋ณด์ƒ ํ•จ์ˆ˜์— ๋Œ€ํ•ด ์œ ์ง€๋จ)๋กœ ๋ฐ”๊พธ๋ฉด \(f(r;\pi_{\text{ref}},\beta)(x,y)=\beta\log\frac{\pi_{r}(y|x)}{\pi_{\text{ref}} (y|x)}\)๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ์ฆ‰, projection \(f\)๋Š” ์›ํ•˜๋Š” ํ˜•ํƒœ์˜ \(r\) ๋“ฑ๊ฐ€ ํด๋ž˜์Šค์˜ ๋ฉค๋ฒ„๋ฅผ ์ƒ์„ฑํ•˜๋ฉฐ, ์ œ์•ˆ๋œ reparameterization๋กœ ์ธํ•ด ๋ณด์ƒ ๋ชจ๋ธ์—์„œ ์ผ๋ฐ˜์„ฑ์„ ์žƒ์ง€ ์•Š์Šต๋‹ˆ๋‹ค.[|endofturn|]
์—ฐ์‚ฐ์ž \(f\)๋Š” ๋‹จ์ˆœํžˆ ๋ณด์ƒ ํ•จ์ˆ˜๋ฅผ \(\pi_{r}\)์˜ ๋ถ„ํ•  ํ•จ์ˆ˜์˜ ๋กœ๊ทธ๋กœ ์ •๊ทœํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ถ”๊ฐ€๋œ ์ •๊ทœํ™” ํ•ญ์€ ์ ‘๋‘์‚ฌ \(x\)์˜ ํ•จ์ˆ˜์ด๋ฏ€๋กœ \(f(r;\pi_{\text{ref}},\beta)(x,y)\)๋Š” \(r(x,y)\)์˜ ๋“ฑ๊ฐ€ ํด๋ž˜์Šค์—์„œ ๋ณด์ƒ ํ•จ์ˆ˜์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, \(r\)์„ Eq. 5์˜ RHS(๋ชจ๋“  ๋ณด์ƒ ํ•จ์ˆ˜์— ๋Œ€ํ•ด ์œ ์ง€๋จ)๋กœ ๋ฐ”๊พธ๋ฉด \(f(r;\pi_{\text{ref}},\beta)(x,y)=\beta\log\frac{\pi_{r}(y|x)}{\pi_{\text{ref}} (y|x)}\)๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ์ฆ‰, ํˆฌ์˜ \(f\)๋Š” ์›ํ•˜๋Š” ํ˜•ํƒœ์˜ \(r\) ๋“ฑ๊ฐ€ ํด๋ž˜์Šค์˜ ๋ฉค๋ฒ„๋ฅผ ์ƒ์„ฑํ•˜๋ฉฐ, ์ œ์•ˆ๋œ ์žฌ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋กœ ์ธํ•ด ๋ณด์ƒ ๋ชจ๋ธ์—์„œ ์ผ๋ฐ˜์„ฑ์„ ์žƒ์ง€ ์•Š์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The operator \(f\) simply normalizes the reward function with the logarithm of the partition function of \(\pi_{r}\). Since the added normalization term is only a function of the prefix \(x\), \(f(r;\pi_{\text{ref}},\beta)(x,y)\) is a reward function in the equivalence class of \(r(x,y)\). Finally, replacing \(r\) with the RHS of Eq. 5 (which holds for any reward function), we have \(f(r;\pi_{\text{ref}},\beta)(x,y)=\beta\log\frac{\pi_{r}(y|x)}{\pi_{\text{ref}} (y|x)}\). That is, the projection \(f\) produces a member of the equivalence class of \(r\) with the desired form, and we do not lose any generality in our reward model from the proposed reparameterization. [|assistant|]
'<s> Q: ๊ฐ™์€ ์‚ฌ๋žŒ๊ณผ ๊ฐ™์€ ์ผ์— ๋Œ€ํ•ด ๋ฐ˜๋ณตํ•ด์„œ ์ด์•ผ๊ธฐํ•˜๋Š” ๊ฒƒ์€ <|startofthought|>n\n(a) a one-to-one correlation\n\n(b) a one-to<| endofthought|> ๋ˆ„๊ตฐ๊ฐ€๊ฐ€ ๋ฌด์—‡์„ ํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€?' [|endofturn|]
'<s> Q: ๊ฐ™์€ ์‚ฌ๋žŒ๊ณผ ๊ฐ™์€ ์ผ์— ๋Œ€ํ•ด ๋ฐ˜๋ณตํ•ด์„œ ์ด์•ผ๊ธฐํ•˜๋Š” ๊ฒƒ์€ <|startofthought|>n\n(a) ์ผ๋Œ€์ผ ์ƒ๊ด€๊ด€๊ณ„\n\n(b) ์ผ๋Œ€์ผ ์ƒ๊ด€๊ด€๊ณ„<| endofthought|> ๋ˆ„๊ตฐ๊ฐ€๊ฐ€ ๋ฌด์—‡์„ ํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€?'[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]'<s> Q: Talking to the same person about the same thing over and over again is <|startofthought|>n\n(a) a one-to-one correlation\n\n(b) a one-to<| endofthought|> something someone can what?' [|assistant|]
์—ฌ๊ธฐ์„œ \(E_{test}^{clean}\)์€ ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ(AI๊ฐ€ ์ƒ์„ฑํ•˜์ง€ ์•Š์€)๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜์ ์ธ ํ…Œ์ŠคํŠธ ์˜ค๋ฅ˜์ž…๋‹ˆ๋‹ค. ์ฆ‰, ์˜ˆ๋ฅผ ๋“ค์–ด ๋ฐฉ์ •์‹ (2) ๋ฐ (4)์—์„œ \(k^{-c^{\prime}}\)์€ \(nk^{-c^{\prime}}\)์œผ๋กœ ๋Œ€์ฒด๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณฑ์…ˆ์  ์ €ํ•˜์˜ ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ํ•ด์„์€ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ(์ฆ‰, ์„ธ๋Œ€ ์ˆ˜๊ฐ€ ๋งŽ์•„์ง์— ๋”ฐ๋ผ) ์•ผ์ƒ์—์„œ Large Language Models (LLMs)(์˜ˆ: ChatGPT)์˜ ์˜ํ–ฅ์ด ์›น์„ ์˜ค์—ผ์‹œ์ผœ ํ•™์Šต์ด ๋ถˆ๊ฐ€๋Šฅํ•ด์งˆ ๊ฒƒ์ด๋ผ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋กœ ์ธํ•ด ๊นจ๋—ํ•œ/AI๊ฐ€ ์ƒ์„ฑํ•˜์ง€ ์•Š์€ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€์น˜์™€ ๋น„์šฉ์ด ์ฆ๊ฐ€ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค.[|endofturn|]
์—ฌ๊ธฐ์„œ \(E_{test}^{clean}\)์€ ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ(AI๊ฐ€ ์ƒ์„ฑํ•˜์ง€ ์•Š์€)๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜์ ์ธ ํ…Œ์ŠคํŠธ ์˜ค๋ฅ˜์ž…๋‹ˆ๋‹ค. ์ฆ‰, ์˜ˆ๋ฅผ ๋“ค์–ด ๋ฐฉ์ •์‹ (2) ๋ฐ (4)์—์„œ \(k^{-c^{\ํ”„๋ผ์ž„}}\)์€ \(nk^{-c^{\ํ”„๋ผ์ž„}}\)์œผ๋กœ ๋Œ€์ฒด๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณฑ์…ˆ์  ์ €ํ•˜์˜ ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ํ•ด์„์€ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ(์ฆ‰, ์„ธ๋Œ€ ์ˆ˜๊ฐ€ ๋งŽ์•„์ง์— ๋”ฐ๋ผ) ์•ผ์ƒ์—์„œ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(์˜ˆ: ChatGPT)์˜ ์˜ํ–ฅ์ด ์›น์„ ์˜ค์—ผ์‹œ์ผœ ํ•™์Šต์ด ๋ถˆ๊ฐ€๋Šฅํ•ด์งˆ ๊ฒƒ์ด๋ผ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋กœ ์ธํ•ด ๊นจ๋—ํ•œ/AI๊ฐ€ ์ƒ์„ฑํ•˜์ง€ ์•Š์€ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€์น˜์™€ ๋น„์šฉ์ด ์ฆ๊ฐ€ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]where \(E_{test}^{clean}\) is the usual test error of the model trained on clean data (not AI-generated). This means that in Equations (2) and (4) for example, the \(k^{-c^{\prime}}\) is replaced by \(nk^{-c^{\prime}}\). One possible interpretation of this multiplicative degradation is that, over time (i.e as the number of generations becomes large), the effect of large language models (like ChatGPT) in the wild will be a pollution of the web to the extend that learning will be impossible. This will likely increase the value and cost of clean / non-AI-generated data. [|assistant|]
FiD-light(Hofstatter et al., 2023)๋Š” ์ˆœ์œ„๋ฅผ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์†Œ์Šค ํฌ์ธํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” listwise autoregressive re-ranking ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ์™€ ์†Œ์Šค ๊ตฌ์ ˆ ๊ฐ„์˜ ๋งํฌ๋ฅผ ์œ ์ง€ํ•˜์—ฌ ๋ณด๋‹ค ์ฒด๊ณ„์ ์ธ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ ์ถœ๋ ฅ์— ๊ด€๋ จ ์ •๋ณด ์†Œ์Šค์— ๋Œ€ํ•œ ํฌ์ธํ„ฐ๋กœ ํ…์ŠคํŠธ ์ธ์šฉ์„ ํ†ตํ•ฉํ•จ์œผ๋กœ์จ ์ฒด๊ณ„์ ์ธ retrieval ๋ฐ generation ํ”„๋กœ์„ธ์Šค๋ฅผ ์ด‰์ง„ํ•˜์—ฌ ์ƒ์„ฑ๋œ ์ฝ˜ํ…์ธ ์˜ ์ „๋ฐ˜์ ์ธ ์ผ๊ด€์„ฑ๊ณผ ๊ด€๋ จ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.[|endofturn|]
FiD-light(Hofstatter et al., 2023)๋Š” ์ˆœ์œ„๋ฅผ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์†Œ์Šค ํฌ์ธํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ชฉ๋ก ์ˆœ์œ„ ์ž๋™ ํšŒ๊ท€ ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ์™€ ์†Œ์Šค ๊ตฌ์ ˆ ๊ฐ„์˜ ๋งํฌ๋ฅผ ์œ ์ง€ํ•˜์—ฌ ๋ณด๋‹ค ์ฒด๊ณ„์ ์ธ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ ์ถœ๋ ฅ์— ๊ด€๋ จ ์ •๋ณด ์†Œ์Šค์— ๋Œ€ํ•œ ํฌ์ธํ„ฐ๋กœ ํ…์ŠคํŠธ ์ธ์šฉ์„ ํ†ตํ•ฉํ•จ์œผ๋กœ์จ ์ฒด๊ณ„์ ์ธ ๊ฒ€์ƒ‰ ๋ฐ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ด‰์ง„ํ•˜์—ฌ ์ƒ์„ฑ๋œ ์ฝ˜ํ…์ธ ์˜ ์ „๋ฐ˜์ ์ธ ์ผ๊ด€์„ฑ๊ณผ ๊ด€๋ จ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]FiD-light (Hofstatter et al., 2023) introduces a listwise autoregressive re-ranking method that employs source pointers to optimize the ranking order. This method maintains a link between the generated text and source passages, enabling a more structured generation process. By incorporating textual citations within the model's output as pointers to relevant information sources, this approach facilitates an organized retrieval and generation process, enhancing the overall coherence and relevance of the generated content. [|assistant|]
ํ† ํฐํ™”.** Tokenization์€ ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ค‘์š”ํ•œ ๋‹จ๊ณ„์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. Tokenization์€ ์›์‹œ ํ…์ŠคํŠธ๋ฅผ ๊ฐœ๋ณ„ ํ† ํฐ์˜ ์‹œํ€€์Šค๋กœ ๋ถ„ํ• ํ•˜์—ฌ LLM์˜ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด NLP ์—ฐ๊ตฌ(์˜ˆ: ์กฐ๊ฑด๋ถ€ ๋žœ๋ค ํ•„๋“œ๋ฅผ ์‚ฌ์šฉํ•œ ์‹œํ€€์Šค ๋ผ๋ฒจ๋ง [220])์—์„œ๋Š” ๋‹จ์–ด ๊ธฐ๋ฐ˜ Tokenization์ด ์ฃผ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ, ์ด๋Š” ์ธ๊ฐ„์˜ ์–ธ์–ด ์ธ์ง€์™€ ๋” ์ž˜ ๋ถ€ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹จ์–ด ๊ธฐ๋ฐ˜ Tokenization์€ ์ผ๋ถ€ ์–ธ์–ด(์˜ˆ: ์ค‘๊ตญ์–ด ๋‹จ์–ด ๋ถ„ํ• )์—์„œ ๋™์ผํ•œ ์ž…๋ ฅ์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ๋ถ„ํ•  ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋นˆ๋„๊ฐ€ ๋‚ฎ์€ ๋‹จ์–ด๊ฐ€ ํฌํ•จ๋œ ๋ฐฉ๋Œ€ํ•œ ๋‹จ์–ด ์–ดํœ˜๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๊ณ , '์–ดํœ˜ ์™ธ ๋‹จ์–ด' ๋ฌธ์ œ๋„ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ช‡๋ช‡ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์—์„œ๋Š” _๋ฌธ์ž_๋ฅผ ์ตœ์†Œ ๋‹จ์œ„๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์–ด ํ‘œํ˜„์„ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค(_์˜ˆ: ELMo [21]์˜ CNN ๋‹จ์–ด Encoder). ์ตœ๊ทผ์—๋Š” Transformer ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ชจ๋ธ์—์„œ _์„œ๋ธŒ์›Œ๋“œ Tokenization_๊ฐ€ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ Byte-Pair Encoding Tokenization, WordPiece Tokenization, Unigram Tokenization์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. HuggingFace๋Š” Tokenizer22์— ๋Œ€ํ•œ ๋›ฐ์–ด๋‚œ ์˜จ๋ผ์ธ NLP ๊ฐ•์ขŒ๋ฅผ ์šด์˜ ์ค‘์ด๋ฉฐ, ์ดˆ๋ณด์ž๋Š” ์ด ๊ฐ•์ขŒ๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ๋‹ค์Œ์—์„œ๋Š” ๋Œ€ํ‘œ์ ์ธ Tokenization ๋ฐฉ๋ฒ• ์„ธ ๊ฐ€์ง€๋ฅผ ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
ํ† ํฐํ™”.** ํ† ํฐํ™”๋Š” ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ค‘์š”ํ•œ ๋‹จ๊ณ„์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ํ† ํฐํ™”๋Š” ์›์‹œ ํ…์ŠคํŠธ๋ฅผ ๊ฐœ๋ณ„ ํ† ํฐ์˜ ์‹œํ€€์Šค๋กœ ๋ถ„ํ• ํ•˜์—ฌ LLM์˜ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด NLP ์—ฐ๊ตฌ(์˜ˆ: ์กฐ๊ฑด๋ถ€ ๋žœ๋ค ํ•„๋“œ๋ฅผ ์‚ฌ์šฉํ•œ ์‹œํ€€์Šค ๋ผ๋ฒจ๋ง [220])์—์„œ๋Š” ๋‹จ์–ด ๊ธฐ๋ฐ˜ ํ† ํฐํ™”๊ฐ€ ์ฃผ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ, ์ด๋Š” ์ธ๊ฐ„์˜ ์–ธ์–ด ์ธ์ง€์™€ ๋” ์ž˜ ๋ถ€ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹จ์–ด ๊ธฐ๋ฐ˜ ํ† ํฐํ™”๋Š” ์ผ๋ถ€ ์–ธ์–ด(์˜ˆ: ์ค‘๊ตญ์–ด ๋‹จ์–ด ๋ถ„ํ• )์—์„œ ๋™์ผํ•œ ์ž…๋ ฅ์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ๋ถ„ํ•  ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋นˆ๋„๊ฐ€ ๋‚ฎ์€ ๋‹จ์–ด๊ฐ€ ํฌํ•จ๋œ ๋ฐฉ๋Œ€ํ•œ ๋‹จ์–ด ์–ดํœ˜๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๊ณ , '์–ดํœ˜ ์™ธ ๋‹จ์–ด' ๋ฌธ์ œ๋„ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ช‡๋ช‡ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์—์„œ๋Š” _๋ฌธ์ž_๋ฅผ ์ตœ์†Œ ๋‹จ์œ„๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์–ด ํ‘œํ˜„์„ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค(_์˜ˆ: ELMo [21]์˜ CNN ๋‹จ์–ด ์ธ์ฝ”๋”). ์ตœ๊ทผ์—๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ชจ๋ธ์—์„œ _์„œ๋ธŒ์›Œ๋“œ ํ† ํฐํ™”_๊ฐ€ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฐ”์ดํŠธ ์Œ ์ธ์ฝ”๋”ฉ ํ† ํฐํ™”, ์›Œ๋“œํ”ผ์Šค ํ† ํฐํ™”, ์œ ๋‹ˆ๊ทธ๋žจ ํ† ํฐํ™”๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. HuggingFace๋Š” ํ† ํฐ๋ผ์ด์ €22์— ๋Œ€ํ•œ ๋›ฐ์–ด๋‚œ ์˜จ๋ผ์ธ NLP ๊ฐ•์ขŒ๋ฅผ ์šด์˜ ์ค‘์ด๋ฉฐ, ์ดˆ๋ณด์ž๋Š” ์ด ๊ฐ•์ขŒ๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ๋‹ค์Œ์—์„œ๋Š” ๋Œ€ํ‘œ์ ์ธ ํ† ํฐํ™” ๋ฐฉ๋ฒ• ์„ธ ๊ฐ€์ง€๋ฅผ ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Tokenization.** Tokenization is also a crucial step for data preprocessing. It aims to segment raw text into sequences of individual tokens, which are subsequently used as the inputs of LLMs. In traditional NLP research (_e.g.,_ sequence labeling with conditional random fields [220]), word-based tokenization is the predominant approach, which is more aligned with human's language cognition. However, word-based tokenization can yield different segmentation results for the same input in some languages (_e.g.,_ Chinese word segmentation), generate a huge word vocabulary containing many low-frequency words, and also suffer from the _"out-of-vocabulary"_ issue. Thus, several neural network models employ _character_ as the minimum unit to derive the word representation (_e.g.,_ a CNN word encoder in ELMo [21]). Recently, _subword tokenizers_ have been widely used in Transformer based language models, typically including Byte-Pair Encoding tokenization, WordPiece tokenization and Unigram tokenization. HuggingFace has maintained an excellent online NLP course on tokenizer22 with running examples, and we refer to the beginners to this course. Next, we briefly describe the three representative tokenization methods. [|assistant|]
๊ทธ๋ฆผ 2: ํ‘œ 1์˜ ๋ชจ๋“  ์‹คํ—˜์— ๊ฑธ์ณ ํ‰๊ท (์—ด๋ณ„)๋œ ๊ฐ knowledge-injection ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์ƒ๋Œ€์  ์ •ํ™•๋„ ํ–ฅ์ƒ(์‹ (5) ์„ค๋ช…)์„ ๋ณด์—ฌ์ค€๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 2: ํ‘œ 1์˜ ๋ชจ๋“  ์‹คํ—˜์— ๊ฑธ์ณ ํ‰๊ท (์—ด๋ณ„)๋œ ๊ฐ ์ง€์‹ ์ฃผ์ž… ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์ƒ๋Œ€์  ์ •ํ™•๋„ ํ–ฅ์ƒ(์‹ (5) ์„ค๋ช…)์„ ๋ณด์—ฌ์ค€๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 2: The relative accuracy gain (as explained in Equation (5)) for each knowledge-injection method, averaged (columnwise) across all experiments in Table 1. [|assistant|]
Wikipedia.org์—์„œ 9๊ฐœ ์ข…๋ชฉ(Athletics, weightlifting, swimming, archery, track cycling, rowing, shooting, short track speed skating, speed skating)์˜ ์˜ฌ๋ฆผํ”ฝ ๊ธฐ๋ก ํŽ˜์ด์ง€๋ฅผ ๊ฐ€์ ธ์™”์Šต๋‹ˆ๋‹ค. ๊ธฐ๋ก์€ table ํ˜•์‹์œผ๋กœ ์ถ”์ถœ๋˜๋ฉฐ, ๊ฐ ๊ธฐ๋ก ํ•ญ๋ชฉ์— ๋Œ€ํ•ด ์งˆ๋ฌธ์ด ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํ•„ํ„ฐ๋ง ํ›„ ์ด 192๊ฐœ์˜ ๊ณ ์œ ํ•œ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ์ถ”์ถœํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
Wikipedia.org์—์„œ 9๊ฐœ ์ข…๋ชฉ(์œก์ƒ, ์—ญ๋„, ์ˆ˜์˜, ์–‘๊ถ, ํŠธ๋ž™ ์‚ฌ์ดํด, ์กฐ์ •, ์‚ฌ๊ฒฉ, ์‡ผํŠธํŠธ๋ž™ ์Šคํ”ผ๋“œ ์Šค์ผ€์ดํŒ…, ์Šคํ”ผ๋“œ ์Šค์ผ€์ดํŒ…)์˜ ์˜ฌ๋ฆผํ”ฝ ๊ธฐ๋ก ํŽ˜์ด์ง€๋ฅผ ๊ฐ€์ ธ์™”์Šต๋‹ˆ๋‹ค. ๊ธฐ๋ก์€ ํ‘œ ํ˜•์‹์œผ๋กœ ์ถ”์ถœ๋˜๋ฉฐ, ๊ฐ ๊ธฐ๋ก ํ•ญ๋ชฉ์— ๋Œ€ํ•ด ์งˆ๋ฌธ์ด ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํ•„ํ„ฐ๋ง ํ›„ ์ด 192๊ฐœ์˜ ๊ณ ์œ ํ•œ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ์ถ”์ถœํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We pulled Olympics records pages from Wikipedia.org across 9 sports: Athletics, weightlifting, swimming, archery, track cycling, rowing, shooting, short track speed skating, and speed skating. Records are extracted in a table format, from which questions are generated for each record entry. In total, after filtering, we extracted 192 unique questions and answers. [|assistant|]
๋ฐ์ดํ„ฐ ์„ธํŠธ ์„ ํƒ์‚ฌ์ „ ํ•™์Šต LLM์„ ์œ„ํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ํ™œ๋ฐœํ•˜๊ณ  ์˜ํ–ฅ๋ ฅ์ด ํฌ์ง€๋งŒ ์•„์ง ์—ฐ๊ตฌ๊ฐ€ ๋ถ€์กฑํ•œ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-2 ๋ชจ๋ธ์€ ์†Œ์…œ ๋ฏธ๋””์–ด ํ”Œ๋žซํผ์ธ Reddit์—์„œ ์นด๋ฅด๋งˆ(Brown et al., 2020)๋ฅผ 3๊ฐœ ์ด์ƒ ๋ฐ›์€ ๋ชจ๋“  ์•„์›ƒ๋ฐ”์šด๋“œ ๋งํฌ๋ฅผ ๋Œ€์ƒ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต์„ ์ง„ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ฌธ์„œ๊ฐ€ _ํฅ๋ฏธ๋กญ๊ฑฐ๋‚˜_, _๊ต์œก์ ์ด๊ฑฐ๋‚˜_, _๊ทธ๋ƒฅ ์žฌ๋ฏธ์žˆ๊ฑฐ๋‚˜_ ํ•˜๋Š” ์ง€ํ‘œ๋กœ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ›„์† ์—ฐ๊ตฌ์—์„œ๋Š” ์œ„ํ‚คํ”ผ๋””์•„์™€ ์œ ์‚ฌํ•œ ๋ฌธ์„œ์˜ ์šฐ์„ ์ˆœ์œ„๋ฅผ ์ •ํ•˜๋Š” ๋“ฑ์˜ ๋‹ค๋ฅธ heuristic์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค(Gururangan et al., 2022). Rae et al. (2021)์€ ํŠน์ • stopword์˜ ๋ถ€์žฌ, ๋ฌธ์„œ ๊ธธ์ด, ์•ŒํŒŒ๋ฒณ ๋ฌธ์ž ๋น„์œจ, ํ‰๊ท  ๋‹จ์–ด ๊ธธ์ด, ๊ธฐํ˜ธ ๋Œ€ ๋‹จ์–ด ๋น„์œจ, ๊ธ€๋จธ๋ฆฌ ๊ธฐํ˜ธ๋กœ ์‹œ์ž‘ํ•˜๊ฑฐ๋‚˜ ์ค„์ž„ํ‘œ๋กœ ๋๋‚˜๋Š” ์ค„์˜ ๋น„์œจ ๋“ฑ๊ณผ ๊ฐ™์€ ์—ฌ๋Ÿฌ heuristic filter๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ์„œ๋ฅผ ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ํ•„ํ„ฐ๋งํ•˜๋Š” ๋ฐ ์žˆ์–ด ๋ณต์žก์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต์„ ์œ„ํ•œ ๋” ๋‚˜์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ๋Œ€์•ˆ ํŒจ๋Ÿฌ๋‹ค์ž„์€ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. Xie et al. (2023)์€ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์˜ ๋ฐ์ดํ„ฐ์— ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ๋ฌผ์„ ์„ ํƒํ•˜๋Š” ๋ฐฉ๋ฒ•์ธ DoReMi๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋™์‹œ์— Abbas et al. (2023)์€ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ์ค‘๋ณต์„ ์ œ๊ฑฐํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต ํšจ์œจ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” LLM์˜ ๋” ๋น ๋ฅธ Fine-tuning์„ ์œ„ํ•ด ์ €ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ž๋™์œผ๋กœ ํ•„ํ„ฐ๋งํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Chen et al., 2023; Solaiman and Dennison, 2021; Zhou et al., 2023). ๋™์‹œ์— CLIP(Radford et al., 2021)๊ณผ ๊ฐ™์€ ์ด๋ฏธ์ง€ ์–ธ์–ด ๋ชจ๋ธ ์˜์—ญ์—์„œ Datacomp ๋ฒค์น˜๋งˆํฌ(Gadre et al., 2023)์™€ ์ตœ๊ทผ์˜ ์ง„์ž…์ž(Maini et al., 2023; Yu et al., 2023)๋Š” ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ €ํ’ˆ์งˆ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ํ•„ํ„ฐ๋งํ•˜๊ฑฐ๋‚˜ ์ผ๋ฐ˜ ํฌ๋กค๋ง์—์„œ ์Šคํฌ๋žฉํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค(Schuhmann et al., 2022).[|endofturn|]
๋ฐ์ดํ„ฐ ์„ธํŠธ ์„ ํƒ์‚ฌ์ „ ํ•™์Šต LLM์„ ์œ„ํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ํ™œ๋ฐœํ•˜๊ณ  ์˜ํ–ฅ๋ ฅ์ด ํฌ์ง€๋งŒ ์•„์ง ์—ฐ๊ตฌ๊ฐ€ ๋ถ€์กฑํ•œ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-2 ๋ชจ๋ธ์€ ์†Œ์…œ ๋ฏธ๋””์–ด ํ”Œ๋žซํผ์ธ Reddit์—์„œ ์นด๋ฅด๋งˆ(Brown et al., 2020)๋ฅผ 3๊ฐœ ์ด์ƒ ๋ฐ›์€ ๋ชจ๋“  ์•„์›ƒ๋ฐ”์šด๋“œ ๋งํฌ๋ฅผ ๋Œ€์ƒ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต์„ ์ง„ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ฌธ์„œ๊ฐ€ _ํฅ๋ฏธ๋กญ๊ฑฐ๋‚˜_, _๊ต์œก์ ์ด๊ฑฐ๋‚˜_, _๊ทธ๋ƒฅ ์žฌ๋ฏธ์žˆ๊ฑฐ๋‚˜_ ํ•˜๋Š” ์ง€ํ‘œ๋กœ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ›„์† ์—ฐ๊ตฌ์—์„œ๋Š” ์œ„ํ‚คํ”ผ๋””์•„์™€ ์œ ์‚ฌํ•œ ๋ฌธ์„œ์˜ ์šฐ์„ ์ˆœ์œ„๋ฅผ ์ •ํ•˜๋Š” ๋“ฑ์˜ ๋‹ค๋ฅธ ํœด๋ฆฌ์Šคํ‹ฑ์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค(Gururangan et al., 2022). Rae et al. (2021)์€ ํŠน์ • ์ค‘์ง€์–ด์˜ ๋ถ€์žฌ, ๋ฌธ์„œ ๊ธธ์ด, ์•ŒํŒŒ๋ฒณ ๋ฌธ์ž ๋น„์œจ, ํ‰๊ท  ๋‹จ์–ด ๊ธธ์ด, ๊ธฐํ˜ธ ๋Œ€ ๋‹จ์–ด ๋น„์œจ, ๊ธ€๋จธ๋ฆฌ ๊ธฐํ˜ธ๋กœ ์‹œ์ž‘ํ•˜๊ฑฐ๋‚˜ ์ค„์ž„ํ‘œ๋กœ ๋๋‚˜๋Š” ์ค„์˜ ๋น„์œจ ๋“ฑ๊ณผ ๊ฐ™์€ ์—ฌ๋Ÿฌ ํœด๋ฆฌ์Šคํ‹ฑ ํ•„ํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ์„œ๋ฅผ ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ํ•„ํ„ฐ๋งํ•˜๋Š” ๋ฐ ์žˆ์–ด ๋ณต์žก์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต์„ ์œ„ํ•œ ๋” ๋‚˜์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ๋Œ€์•ˆ ํŒจ๋Ÿฌ๋‹ค์ž„์€ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. Xie et al. (2023)์€ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์˜ ๋ฐ์ดํ„ฐ์— ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ๋ฌผ์„ ์„ ํƒํ•˜๋Š” ๋ฐฉ๋ฒ•์ธ DoReMi๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋™์‹œ์— Abbas et al. (2023)์€ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ์ค‘๋ณต์„ ์ œ๊ฑฐํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต ํšจ์œจ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” LLM์˜ ๋” ๋น ๋ฅธ ๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•ด ์ €ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ž๋™์œผ๋กœ ํ•„ํ„ฐ๋งํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Chen et al., 2023; Solaiman and Dennison, 2021; Zhou et al., 2023). ๋™์‹œ์— CLIP(Radford et al., 2021)๊ณผ ๊ฐ™์€ ์ด๋ฏธ์ง€ ์–ธ์–ด ๋ชจ๋ธ ์˜์—ญ์—์„œ Datacomp ๋ฒค์น˜๋งˆํฌ(Gadre et al., 2023)์™€ ์ตœ๊ทผ์˜ ์ง„์ž…์ž(Maini et al., 2023; Yu et al., 2023)๋Š” ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ €ํ’ˆ์งˆ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ํ•„ํ„ฐ๋งํ•˜๊ฑฐ๋‚˜ ์ผ๋ฐ˜ ํฌ๋กค๋ง์—์„œ ์Šคํฌ๋žฉํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค(Schuhmann et al., 2022).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Dataset SelectionSelecting high quality data for pre-training LLMs remains an active, high-impact, yet understudied area of research. For instance, GPT-2 model was pre-trained on all outbound links from Reddit, a social media platform, which received at least 3 karma (Brown et al., 2020). This was used as a heuristic indicator that the document may be _interesting_, _educational_, or _just funny_. Follow-up works have used other heuristics such as prioritizing documents that resemble wikipedia (Gururangan et al., 2022). Rae et al. (2021) used multiple heuristic filters to remove documents, such as the absence of certain stopwords, length of the document, percentage of alphabetic characters, mean word length, symbol-to-word ratio, percentage of lines starting with a bullet point, or ending with an ellipsis etc. Their work highlights the intricacies of filtering out text data. An alternative paradigm for building better datasets for training is to distill high-quality datasets. Xie et al. (2023) proposed a method, DoReMi, to select the best data mixture for pre-training language models by reweighting data from various domains. Concurrently, Abbas et al. (2023) showed that de-duplicating pre-training data can improve pre-training efficiency. Recently several methods were proposed for automatic filtering of low-quality data for faster fine-tuning of LLMs (Chen et al., 2023; Solaiman and Dennison, 2021; Zhou et al., 2023). Simultaneously, in the realm of image-language models such as CLIP (Radford et al., 2021), the Datacomp benchmark (Gadre et al., 2023) and recent entries (Maini et al., 2023; Yu et al., 2023) have developed approaches at filtering out low-quality subsets from pre-training datasets like LIAION (Schuhmann et al., 2022), or from scrapes of the common crawl. [|assistant|]
**์ปดํ“จํ„ฐ ๊ณผํ•™ ๋ฐ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด๋ง**: ์ปดํ“จํ„ฐ ๊ณผํ•™ ๋ฐ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด๋ง ๋ถ„์•ผ์—์„œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ์ฝ”๋”ฉ, ํ…Œ์ŠคํŠธ, ๋””๋ฒ„๊น… ๋ฐ ๋ฌธ์„œ ์ƒ์„ฑ ์ž๋™ํ™”์— ์ž ์žฌ๋ ฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค[14; 18; 23; 24; 126; 127; 128]. Chat-Dev [18]์€ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ ์ˆ˜๋ช… ์ฃผ๊ธฐ๋ฅผ ์™„๋ฃŒํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ์—ญํ• ์ด ์ž์—ฐ์–ด ๋Œ€ํ™”๋ฅผ ํ†ตํ•ด ์†Œํ†ตํ•˜๊ณ  ํ˜‘์—…ํ•˜๋Š” end-to-end ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ํšจ์œจ์ ์ด๊ณ  ๋น„์šฉ ํšจ์œจ์ ์ธ ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ์†Œํ”„ํŠธ์›จ์–ด ์‹œ์Šคํ…œ ์ƒ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ToolBench [14]๋Š” ์ฝ”๋“œ ์ž๋™ ์™„์„ฑ ๋ฐ ์ฝ”๋“œ ์ถ”์ฒœ๊ณผ ๊ฐ™์€ ์ž‘์—…์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. MetaGPT [23]๋Š” ์ œํ’ˆ ๊ด€๋ฆฌ์ž, ์•„ํ‚คํ…ํŠธ, ํ”„๋กœ์ ํŠธ ๊ด€๋ฆฌ์ž, ์—”์ง€๋‹ˆ์–ด ๋“ฑ ์—ฌ๋Ÿฌ ์—ญํ• ์„ ์ถ”์ƒํ™”ํ•˜์—ฌ ์ฝ”๋“œ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ๋…ํ•˜๊ณ  ์ตœ์ข… ์ถœ๋ ฅ ์ฝ”๋“œ์˜ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ €๋น„์šฉ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. [24]๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฝ”๋“œ ์ƒ์„ฑ์„ ์œ„ํ•œ ์ž์ฒด ํ˜‘์—… ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ๋Š” ์—ฌ๋Ÿฌ LLM์ด ํŠน์ • ํ•˜์œ„ ์ž‘์—…์— ๋Œ€ํ•œ ๋ณ„๊ฐœ์˜ "์ „๋ฌธ๊ฐ€"๋กœ ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ์ง€์ •๋œ instruction์— ๋”ฐ๋ผ ํ˜‘์—…ํ•˜๊ณ  ์ƒํ˜ธ ์ž‘์šฉํ•˜์—ฌ ์„œ๋กœ์˜ ์ž‘์—…์„ ์ด‰์ง„ํ•˜๋Š” ๊ฐ€์ƒ ํŒ€์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ถ๊ทน์ ์œผ๋กœ ๊ฐ€์ƒ ํŒ€์€ ์ธ๊ฐ„์˜ ๊ฐœ์ž… ์—†์ด๋„ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ํ˜‘๋ ฅ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. LLIFT [139]๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ํŠนํžˆ ์ž ์žฌ์ ์ธ ์ฝ”๋“œ ์ทจ์•ฝ์ ์„ ์‹๋ณ„ํ•˜๊ธฐ ์œ„ํ•œ ์ •์  ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ •ํ™•์„ฑ๊ณผ ํ™•์žฅ์„ฑ ๊ฐ„์˜ ๊ท ํ˜•์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ด€๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ChatEDA [123]๋Š” ์ž‘์—… ๊ณ„ํš, ์Šคํฌ๋ฆฝํŠธ ์ƒ์„ฑ ๋ฐ ์‹คํ–‰์„ ํ†ตํ•ฉํ•˜์—ฌ ์„ค๊ณ„ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ„์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ „์ž ์„ค๊ณ„ ์ž๋™ํ™”(EDA)๋ฅผ ์œ„ํ•ด ๊ฐœ๋ฐœ๋œ ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค.[|endofturn|]
**์ปดํ“จํ„ฐ ๊ณผํ•™ ๋ฐ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด๋ง**: ์ปดํ“จํ„ฐ ๊ณผํ•™ ๋ฐ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด๋ง ๋ถ„์•ผ์—์„œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ์ฝ”๋”ฉ, ํ…Œ์ŠคํŠธ, ๋””๋ฒ„๊น… ๋ฐ ๋ฌธ์„œ ์ƒ์„ฑ ์ž๋™ํ™”์— ์ž ์žฌ๋ ฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค[14; 18; 23; 24; 126; 127; 128]. Chat-Dev [18]์€ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ ์ˆ˜๋ช… ์ฃผ๊ธฐ๋ฅผ ์™„๋ฃŒํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ์—ญํ• ์ด ์ž์—ฐ์–ด ๋Œ€ํ™”๋ฅผ ํ†ตํ•ด ์†Œํ†ตํ•˜๊ณ  ํ˜‘์—…ํ•˜๋Š” ์—”๋“œํˆฌ์—”๋“œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ํšจ์œจ์ ์ด๊ณ  ๋น„์šฉ ํšจ์œจ์ ์ธ ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ์†Œํ”„ํŠธ์›จ์–ด ์‹œ์Šคํ…œ ์ƒ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ToolBench [14]๋Š” ์ฝ”๋“œ ์ž๋™ ์™„์„ฑ ๋ฐ ์ฝ”๋“œ ์ถ”์ฒœ๊ณผ ๊ฐ™์€ ์ž‘์—…์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. MetaGPT [23]๋Š” ์ œํ’ˆ ๊ด€๋ฆฌ์ž, ์•„ํ‚คํ…ํŠธ, ํ”„๋กœ์ ํŠธ ๊ด€๋ฆฌ์ž, ์—”์ง€๋‹ˆ์–ด ๋“ฑ ์—ฌ๋Ÿฌ ์—ญํ• ์„ ์ถ”์ƒํ™”ํ•˜์—ฌ ์ฝ”๋“œ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ๋…ํ•˜๊ณ  ์ตœ์ข… ์ถœ๋ ฅ ์ฝ”๋“œ์˜ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ €๋น„์šฉ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. [24]๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฝ”๋“œ ์ƒ์„ฑ์„ ์œ„ํ•œ ์ž์ฒด ํ˜‘์—… ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ๋Š” ์—ฌ๋Ÿฌ LLM์ด ํŠน์ • ํ•˜์œ„ ์ž‘์—…์— ๋Œ€ํ•œ ๋ณ„๊ฐœ์˜ "์ „๋ฌธ๊ฐ€"๋กœ ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ์ง€์ •๋œ ์ง€์นจ์— ๋”ฐ๋ผ ํ˜‘์—…ํ•˜๊ณ  ์ƒํ˜ธ ์ž‘์šฉํ•˜์—ฌ ์„œ๋กœ์˜ ์ž‘์—…์„ ์ด‰์ง„ํ•˜๋Š” ๊ฐ€์ƒ ํŒ€์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ถ๊ทน์ ์œผ๋กœ ๊ฐ€์ƒ ํŒ€์€ ์ธ๊ฐ„์˜ ๊ฐœ์ž… ์—†์ด๋„ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ํ˜‘๋ ฅ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. LLIFT [139]๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ํŠนํžˆ ์ž ์žฌ์ ์ธ ์ฝ”๋“œ ์ทจ์•ฝ์ ์„ ์‹๋ณ„ํ•˜๊ธฐ ์œ„ํ•œ ์ •์  ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ •ํ™•์„ฑ๊ณผ ํ™•์žฅ์„ฑ ๊ฐ„์˜ ๊ท ํ˜•์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ด€๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ChatEDA [123]๋Š” ์ž‘์—… ๊ณ„ํš, ์Šคํฌ๋ฆฝํŠธ ์ƒ์„ฑ ๋ฐ ์‹คํ–‰์„ ํ†ตํ•ฉํ•˜์—ฌ ์„ค๊ณ„ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ„์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ „์ž ์„ค๊ณ„ ์ž๋™ํ™”(EDA)๋ฅผ ์œ„ํ•ด ๊ฐœ๋ฐœ๋œ ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Computer Science & Software Engineering**: In the field of computer science and software engineering, LLM-based agents offer potential for automating coding, testing, debugging, and documentation generation [14; 18; 23; 24; 126; 127; 128]. Chat-Dev [18] proposes an end-to-end framework, where multiple agent roles communicate and collaborate through natural language conversations to complete the software development life cycle. This framework demonstrates efficient and cost-effective generation of executable software systems. ToolBench [14] can be used for tasks such as code auto-completion and code recommendation. MetaGPT [23] abstracts multiple roles, such as product managers, architects, project managers, and engineers, to supervise code generation process and enhance the quality of the final output code. This enables low-cost software development. [24] presents a self-collaboration framework for code generation using LLMs. In this framework, multiple LLMs are assumed to be distinct "experts" for specific sub-tasks. They collaborate and interact according to specified instructions, forming a virtual team that facilitates each other's work. Ultimately, the virtual team collaboratively addresses code generation tasks without requiring human intervention. LLIFT [139] employs LLMs to assist in conducting static analysis, specifically for identifying potential code vulnerabilities. This approach effectively manages the trade-off between accuracy and scalability. ChatEDA [123] is an agent developed for electronic design automation (EDA) to streamline the design process by integrating task planning, script generation, and execution. [|assistant|]
**์ˆ˜์น˜ ๊ณ„์‚ฐ์˜ ์ •ํ™•๋„ ํ–ฅ์ƒ์ด ๋…ผ๋ฆฌ์  ์ถ”๋ก ๋ณด๋‹ค ๋” ์ค‘์š”ํ•˜๋‹ค** ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ์ ์ง„์ ์œผ๋กœ ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค. ๋” ๊นŠ์€ ์ดํ•ด๋ฅผ ์œ„ํ•ด GSM8K์—์„œ ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ์˜ค๋ฅ˜์— ๋Œ€ํ•œ ์˜ค๋ฅ˜ ๋น„์œจ์„ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. ์˜ค๋ฅ˜๋Š” reasoning ์˜ค๋ฅ˜์™€ calculation ์˜ค๋ฅ˜์˜ ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์œผ๋กœ ๋ถ„๋ฅ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. reasoning ์˜ค๋ฅ˜๋Š” ์ฃผ๋กœ ์กฐ๊ฑด ์†์‹ค ๋ฐ ๊ฐœ๋… ํ˜ผ๋™๊ณผ ๊ฐ™์€ ๋ฌธ์ œ๋ฅผ ํฌํ•จํ•˜๋ฉฐ, calculation ์˜ค๋ฅ˜์—๋Š” ์ •๋Ÿ‰์  ๊ด€๊ณ„์˜ ์ž˜๋ชป๋œ ๋ถ„์„ ๋ฐ ์ˆ˜์น˜ ๊ณ„์‚ฐ ์‹ค์ˆ˜๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 5์— ๋„์‹œ๋œ ์‹คํ—˜ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด calculation ์˜ค๋ฅ˜์˜ ๋น„์œจ์ด ์ ์ง„์ ์œผ๋กœ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” GSM8K๊ฐ€ reasoning ์˜ค๋ฅ˜๋ณด๋‹ค ๋” ๋น ๋ฅธ ์†๋„๋กœ calculation ์˜ค๋ฅ˜๋ฅผ ์ˆ˜์ •ํ•˜๊ณ  ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**์ˆ˜์น˜ ๊ณ„์‚ฐ์˜ ์ •ํ™•๋„ ํ–ฅ์ƒ์ด ๋…ผ๋ฆฌ์  ์ถ”๋ก ๋ณด๋‹ค ๋” ์ค‘์š”ํ•˜๋‹ค** ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ์ ์ง„์ ์œผ๋กœ ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค. ๋” ๊นŠ์€ ์ดํ•ด๋ฅผ ์œ„ํ•ด GSM8K์—์„œ ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ์˜ค๋ฅ˜์— ๋Œ€ํ•œ ์˜ค๋ฅ˜ ๋น„์œจ์„ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. ์˜ค๋ฅ˜๋Š” ์ถ”๋ก  ์˜ค๋ฅ˜์™€ ๊ณ„์‚ฐ ์˜ค๋ฅ˜์˜ ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์œผ๋กœ ๋ถ„๋ฅ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ถ”๋ก  ์˜ค๋ฅ˜๋Š” ์ฃผ๋กœ ์กฐ๊ฑด ์†์‹ค ๋ฐ ๊ฐœ๋… ํ˜ผ๋™๊ณผ ๊ฐ™์€ ๋ฌธ์ œ๋ฅผ ํฌํ•จํ•˜๋ฉฐ, ๊ณ„์‚ฐ ์˜ค๋ฅ˜์—๋Š” ์ •๋Ÿ‰์  ๊ด€๊ณ„์˜ ์ž˜๋ชป๋œ ๋ถ„์„ ๋ฐ ์ˆ˜์น˜ ๊ณ„์‚ฐ ์‹ค์ˆ˜๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 5์— ๋„์‹œ๋œ ์‹คํ—˜ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด ๊ณ„์‚ฐ ์˜ค๋ฅ˜์˜ ๋น„์œจ์ด ์ ์ง„์ ์œผ๋กœ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” GSM8K๊ฐ€ ์ถ”๋ก  ์˜ค๋ฅ˜๋ณด๋‹ค ๋” ๋น ๋ฅธ ์†๋„๋กœ ๊ณ„์‚ฐ ์˜ค๋ฅ˜๋ฅผ ์ˆ˜์ •ํ•˜๊ณ  ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**The Improvement in the Accuracy of Numerical Calculations is More Significant than Logical Reasoning** The performance of the model gradually improves as the synthetic data increases. For a deeper understanding, we analyze the error proportion for different types of errors on GSM8K. We categorized errors into two types: reasoning errors and calculation errors. Reasoning errors primarily encompass issues such as loss of conditions and concept confusion, while calculation errors include incorrect analysis of quantitative relationships and numerical computation mistakes. Based on the experimental results illustrated in Figure 5, we observe a gradual decrease in the percentage of calculation errors, suggesting that GSM8K is correcting calculation errors at a faster rate than reasoning errors. [|assistant|]
๋‘ ๊ฐ€์ง€ ์ˆ˜ํ•™์  ์ถ”๋ก  ๋ฒค์น˜๋งˆํฌ, ์ฆ‰ GSM8k[42]์™€ MATH[43]์— ๋Œ€ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•œ ๊ฒฐ๊ณผ, _WizardMath_๊ฐ€ ๋‹ค๋ฅธ ์˜คํ”ˆ ์†Œ์Šค LLM๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์—ฌ state-of-the-art ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, _WizardMath_๋Š” GSM8k์—์„œ +24.8(81.6 ๋Œ€ 56.8), MATH์—์„œ +9.2(22.7 ๋Œ€ 13.5)๋กœ pass@1์—์„œ ์ƒ๋‹นํ•œ ํ–ฅ์ƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์ด ๋ชจ๋ธ์€ GSM8k์—์„œ pass@1 ์ธก๋ฉด์—์„œ OpenAI์˜ ChatGPT-3.5[5], Anthropic์˜ Claude Instant-1[39], Google์˜ PaLM-2[44]๋ฅผ ํฌ๊ฒŒ ๋Šฅ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋‘ ๊ฐ€์ง€ ์ˆ˜ํ•™์  ์ถ”๋ก  ๋ฒค์น˜๋งˆํฌ, ์ฆ‰ GSM8k[42]์™€ MATH[43]์— ๋Œ€ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•œ ๊ฒฐ๊ณผ, WizardMath_๊ฐ€ ๋‹ค๋ฅธ ์˜คํ”ˆ ์†Œ์Šค LLM๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์—ฌ ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, _WizardMath_๋Š” GSM8k์—์„œ +24.8(81.6 ๋Œ€ 56.8), MATH์—์„œ +9.2(22.7 ๋Œ€ 13.5)๋กœ pass@1์—์„œ ์ƒ๋‹นํ•œ ํ–ฅ์ƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์ด ๋ชจ๋ธ์€ GSM8k์—์„œ pass@1 ์ธก๋ฉด์—์„œ OpenAI์˜ ChatGPT-3.5[5], Anthropic์˜ Claude Instant-1[39], Google์˜ PaLM-2[44]๋ฅผ ํฌ๊ฒŒ ๋Šฅ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We perform experiments on two mathematical reasoning benchmarks, namely GSM8k [42] and MATH [43], the results demonstrate that our _WizardMath_ outperforms all other open-source LLMs, achieving state-of-the-art performance. Specifically, _WizardMath_ observe a substantial improvement in pass@1 with an increase of +24.8 (81.6. vs. 56.8) on GSM8k, and +9.2 (22.7 vs. 13.5) on MATH. Notably, our model even also significantly surpasses OpenAI's ChatGPT-3.5 [5], Anthropic's Claude Instant-1 [39], and Google's PaLM-2 [44] in terms of pass@1 on GSM8k. [|assistant|]
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๋‘ ๊ฐ€์ง€ ํ™•๋ฆฝ๋œ ํ† ํฐํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ BPE์™€ Unigram๊ณผ ์ด๋ฅผ ๊ตฌํ˜„ํ•œ _Huggingface tokenizer_ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ(Moi and Patry, 2023)์™€ _SentencePiece_ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ(Kudo and Richardson, 2018)์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ๋ฐ ์‚ฌํ›„ ์ฒ˜๋ฆฌ ๋‹จ๊ณ„์˜ ์ฐจ์ด์™€ ์ž ์žฌ์ ์ธ ๊ตฌํ˜„ ์ฐจ์ด๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ๋ชจ๋‘ ๊ณ ๋ คํ–ˆ์Šต๋‹ˆ๋‹ค. Huggingface์˜ Unigram ๊ตฌํ˜„์— ๋Œ€ํ•œ ์‚ฌ์ „ ์ฒ˜๋ฆฌ ์˜ต์…˜์ด ๋ˆ„๋ฝ๋˜์–ด ๊ฒฐ๊ณผ ์–ดํœ˜์—์„œ SentencePiece์˜ Unigram ๊ตฌํ˜„๊ณผ ํฐ ์ฐจ์ด๊ฐ€ ๋ฐœ์ƒํ•˜๋ฏ€๋กœ, Huggingface ๊ธฐ๋ฐ˜์˜ Unigram tokenizer ํ›ˆ๋ จ์€ ์ƒ๋žตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „์ฒด์ ์œผ๋กœ 24๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ tokenizer๋ฅผ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, tokenizer์˜ ์ ˆ๋ฐ˜์€ ๋‹จ์ผ ์–ธ์–ด ์˜์–ด tokenizer, ๋‚˜๋จธ์ง€ ์ ˆ๋ฐ˜์€ ๋‹ค๊ตญ์–ด tokenizer์˜€์Šต๋‹ˆ๋‹ค. tokenizer ์•Œ๊ณ ๋ฆฌ์ฆ˜, ์–ธ์–ด ๊ตฌ์„ฑ, ์‚ฌ์šฉ๋œ tokenizer ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์™ธ์—๋„ ์–ดํœ˜ ํฌ๊ธฐ๋„ ๋‹ค์–‘ํ•˜๊ฒŒ ์กฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์ธ tokenizer ๊ตฌ์„ฑ์€ ๋ถ€๋ก B์— ์„ค๋ช…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๋‘ ๊ฐ€์ง€ ํ™•๋ฆฝ๋œ ํ† ํฐํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ BPE์™€ ์œ ๋‹ˆ๊ทธ๋žจ๊ณผ ์ด๋ฅผ ๊ตฌํ˜„ํ•œ _ํ—ˆ๊น…ํŽ˜์ด์Šค ํ† ํฐํ™”๊ธฐ_ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ(Moi and Patry, 2023)์™€ _SentencePiece_ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ(Kudo and Richardson, 2018)์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ๋ฐ ์‚ฌํ›„ ์ฒ˜๋ฆฌ ๋‹จ๊ณ„์˜ ์ฐจ์ด์™€ ์ž ์žฌ์ ์ธ ๊ตฌํ˜„ ์ฐจ์ด๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ๋ชจ๋‘ ๊ณ ๋ คํ–ˆ์Šต๋‹ˆ๋‹ค. ํ—ˆ๊น…ํŽ˜์ด์Šค์˜ ์œ ๋‹ˆ๊ทธ๋žจ ๊ตฌํ˜„์— ๋Œ€ํ•œ ์‚ฌ์ „ ์ฒ˜๋ฆฌ ์˜ต์…˜์ด ๋ˆ„๋ฝ๋˜์–ด ๊ฒฐ๊ณผ ์–ดํœ˜์—์„œ SentencePiece์˜ ์œ ๋‹ˆ๊ทธ๋žจ ๊ตฌํ˜„๊ณผ ํฐ ์ฐจ์ด๊ฐ€ ๋ฐœ์ƒํ•˜๋ฏ€๋กœ, ํ—ˆ๊น…ํŽ˜์ด์Šค ๊ธฐ๋ฐ˜์˜ ์œ ๋‹ˆ๊ทธ๋žจ ํ† ํฐํ™”๊ธฐ ํ›ˆ๋ จ์€ ์ƒ๋žตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „์ฒด์ ์œผ๋กœ 24๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ํ† ํฐํ™”๊ธฐ๋ฅผ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, ํ† ํฐํ™”๊ธฐ์˜ ์ ˆ๋ฐ˜์€ ๋‹จ์ผ ์–ธ์–ด ์˜์–ด ํ† ํฐํ™”๊ธฐ, ๋‚˜๋จธ์ง€ ์ ˆ๋ฐ˜์€ ๋‹ค๊ตญ์–ด ํ† ํฐํ™”๊ธฐ์˜€์Šต๋‹ˆ๋‹ค. ํ† ํฐํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜, ์–ธ์–ด ๊ตฌ์„ฑ, ์‚ฌ์šฉ๋œ ํ† ํฐํ™” ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์™ธ์—๋„ ์–ดํœ˜ ํฌ๊ธฐ๋„ ๋‹ค์–‘ํ•˜๊ฒŒ ์กฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์ธ ํ† ํฐํ™” ๊ตฌ์„ฑ์€ ๋ถ€๋ก B์— ์„ค๋ช…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Our studies rely on the two established tokenization algorithms, BPE and Unigram, and their implementation in the _Huggingface tokenizer_ library (Moi and Patry, 2023) and the _SentencePiece_ library (Kudo and Richardson, 2018). We considered both libraries in order to investigate the effect of differences in the pre-and post-processing steps and potential differences in the implementations. Due to missing pre-processing options for Huggingface's Unigram implementation, which causes a large discrepancy in the resulting vocabulary compared to SentencePiece's implementation of Unigram, we omitted the training of Unigram tokenizers based on Huggingface. Overall, we trained 24 different tokenizers, where one-half of the tokenizers were monolingual English tokenizers, and the other half of the tokenizers were multilingual tokenizers. Besides the tokenizer algorithm, language composition, and employed tokenizer library, we also varied the vocabulary size. Concrete tokenizer configurations are described in the Appendix B. [|assistant|]
๊ฒฐ์ • \(B\)๋ชจ๋“  ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ(ํŠนํžˆ Learning Rate ๋ฐ ์ •๊ทœํ™” ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ)๊ฐ€ ์ž˜ ์กฐ์ •๋˜๊ณ  ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜๊ฐ€ ์ถฉ๋ถ„ํ•œ ํ•œ, ๋ชจ๋“  Batch Size๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋™์ผํ•œ ์ตœ์ข… ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์–ด์ง€๋ฏ€๋กœ Batch Size๋Š” ์ฃผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ์†๋„์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•  ๋•Œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ•˜๋“œ์›จ์–ด์—์„œ ์ง€์›ํ•˜๋Š” ์ตœ๋Œ€ Batch Size๋กœ ์ด์ƒ์ ์ธ Batch Size๋ฅผ ์„ค์ •ํ•˜์—ฌ ๊ณ„์‚ฐ ๋น„์šฉ์„ ๊ณ ๋ คํ•˜์ง€ ์•Š๊ณ  ํ›ˆ๋ จ ์†๋„๋ฅผ ์ตœ๋Œ€ํ™”ํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค[1]. 3.12 ๋ฐฉ์ •์‹์—์„œ๋Š” Loss ๊ฐ’์—์„œ ์ตœ์ ์˜ ์†๋„/๊ณ„์‚ฐ ๋น„์šฉ ์ ˆ๊ฐ์ด ๊ฐ€๋Šฅํ•œ ์ž„๊ณ„ Batch Size๋ฅผ ๋ถ„์„์ ์œผ๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ๊ณต์‹์„ ํ†ตํ•ด ๋ชจ๋“  Loss ๊ถค์ ์—์„œ ์„ ํ˜ธํ•˜๋Š” Batch Size๋ฅผ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ 3.12 ๋ฐฉ์ •์‹์˜ ์ตœ์  Batch Size๋Š” 3.9 ๋ฐฉ์ •์‹๊ณผ ๊ฐ™์ด ํ›ˆ๋ จ ์‹œ๊ฐ„๊ณผ ํ•„์š”ํ•œ ๊ณ„์‚ฐ์„ ๋™๋“ฑํ•˜๊ฒŒ ์ตœ์†Œํ™”ํ•˜์—ฌ ๊ฒฐ์ •๋ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ํ•œ์ชฝ์„ ๋‹ค๋ฅธ ์ชฝ๋ณด๋‹ค ์šฐ์„ ์ˆœ์œ„์— ๋”ฐ๋ผ ์ตœ์ ์˜ Batch Size๋ฅผ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ๋งž์ถคํ˜• ์š”๊ตฌ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ์ฒด๊ณ„์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์ตœ์ ์˜ Batch Size๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ฒฐ์ • \(B\)๋ชจ๋“  ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ(ํŠนํžˆ ํ•™์Šต ์†๋„ ๋ฐ ์ •๊ทœํ™” ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ)๊ฐ€ ์ž˜ ์กฐ์ •๋˜๊ณ  ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜๊ฐ€ ์ถฉ๋ถ„ํ•œ ํ•œ, ๋ชจ๋“  ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋™์ผํ•œ ์ตœ์ข… ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์–ด์ง€๋ฏ€๋กœ ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” ์ฃผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ์†๋„์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•  ๋•Œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ•˜๋“œ์›จ์–ด์—์„œ ์ง€์›ํ•˜๋Š” ์ตœ๋Œ€ ๋ฐฐ์น˜ ํฌ๊ธฐ๋กœ ์ด์ƒ์ ์ธ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ์„ค์ •ํ•˜์—ฌ ๊ณ„์‚ฐ ๋น„์šฉ์„ ๊ณ ๋ คํ•˜์ง€ ์•Š๊ณ  ํ›ˆ๋ จ ์†๋„๋ฅผ ์ตœ๋Œ€ํ™”ํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค[1]. 3.12 ๋ฐฉ์ •์‹์—์„œ๋Š” ์†์‹ค ๊ฐ’์—์„œ ์ตœ์ ์˜ ์†๋„/๊ณ„์‚ฐ ๋น„์šฉ ์ ˆ๊ฐ์ด ๊ฐ€๋Šฅํ•œ ์ž„๊ณ„ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ๋ถ„์„์ ์œผ๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ๊ณต์‹์„ ํ†ตํ•ด ๋ชจ๋“  ์†์‹ค ๊ถค์ ์—์„œ ์„ ํ˜ธํ•˜๋Š” ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ 3.12 ๋ฐฉ์ •์‹์˜ ์ตœ์  ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 3.9 ๋ฐฉ์ •์‹๊ณผ ๊ฐ™์ด ํ›ˆ๋ จ ์‹œ๊ฐ„๊ณผ ํ•„์š”ํ•œ ๊ณ„์‚ฐ์„ ๋™๋“ฑํ•˜๊ฒŒ ์ตœ์†Œํ™”ํ•˜์—ฌ ๊ฒฐ์ •๋ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ํ•œ์ชฝ์„ ๋‹ค๋ฅธ ์ชฝ๋ณด๋‹ค ์šฐ์„ ์ˆœ์œ„์— ๋”ฐ๋ผ ์ตœ์ ์˜ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ๋งž์ถคํ˜• ์š”๊ตฌ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ์ฒด๊ณ„์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์ตœ์ ์˜ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Determining \(B\)As long as all hyperparameters are well-tuned (especially the learning rate and regularization hyperparameters) and the number of training steps is sufficient, it is believed that the same final performance should be attainable using any batch size [19], so the batch size mainly influences the training speed of language models. Often, when training large language models, the ideal batch size is suggested to be set as the largest batch size supported by the available hardware [1], so as to maximize the training speed without considering the computational cost. In Eq 3.12, we show that the critical batch size with the optimal speed/computation trade-off can be analytically computed from the loss value. Under the guidance of this formula, we would be able to estimate the preferred batch size under any loss trajectory. Furthermore, this optimal batch size in Eq 3.12 is determined by equally minimizing the training time and required computation, as shown in Eq 3.9. In practice, if we would like to prioritize one over the other, we can follow the same process to derive the optimal batch size. By this means, we are able to obtain the optimal batch size based on our customized need in a systematic way. [|assistant|]
**ํ˜„์•ˆ**. ๋ง‰๋Œ€ํ•œ ์„ฑ๊ณต์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ๊ฐœ๋ฐœ๊ณผ ์ ์šฉ์„ ์ œํ•œํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฌธ์ œ๊ฐ€ ์—ฌ์ „ํžˆ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ๊ธ‰์ฆํ•จ์— ๋”ฐ๋ผ ์‹œ๊ฐ„ ๋ฐ ๋ฉ”๋ชจ๋ฆฌ ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ํฌํ•จํ•œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ํšจ์œจ์„ฑ์€ ๋Œ€๊ทœ๋ชจ ๋ฐฐํฌ, ํŠนํžˆ ์ˆ˜๋งŽ์€ LLM ์ธ์Šคํ„ด์Šค๊ฐ€ ์žˆ๋Š” ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์˜ ๊ฒฝ์šฐ ์ค‘์š”ํ•œ ๋ฌธ์ œ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ์ˆ˜๊ฐ€ ํ™•์žฅ๋จ์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ ๊ฐ„์˜ ์กฐ์ • ๋ณต์žก์„ฑ ์ฆ๊ฐ€๋ฅผ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด ๋ณด๋‹ค ํšจ๊ณผ์ ์ด๊ณ  ํšจ์œจ์ ์ธ ํ†ต์‹  ํ”„๋กœํ† ์ฝœ๊ณผ ์•„ํ‚คํ…์ฒ˜๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์œ ๋Šฅํ•œ ์—์ด์ „ํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋ ค๋ฉด Instruction following ๋ฐ ๊ธด ํ…์ŠคํŠธ ๋ชจ๋ธ๋ง๊ณผ ๊ฐ™์€ LLM์˜ ์—ญ๋Ÿ‰์— ๋Œ€ํ•œ ๊ธฐ์ˆ ์  ๊ณผ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด LLM์€ ์—์ด์ „ํŠธ๋ฅผ ์ธ์Šคํ„ด์Šคํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ์ตœ์ ํ™”๋˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์—, ๋Œ€๋ถ€๋ถ„์˜ ๊ณต๊ฐœ ์†Œ์Šค LLM(์˜ˆ: LLaMA)์€ ์—์ด์ „ํŠธ ๊ฐœ๋ฐœ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ด‰์ง„ํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—์ด์ „ํŠธ์˜ ํ•ต์‹ฌ ๊ณ„์‚ฐ ๋‹จ์œ„ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์ด ์žˆ๋Š” ํŠน์ˆ˜ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**ํ˜„์•ˆ**. ๋ง‰๋Œ€ํ•œ ์„ฑ๊ณต์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ๊ฐœ๋ฐœ๊ณผ ์ ์šฉ์„ ์ œํ•œํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฌธ์ œ๊ฐ€ ์—ฌ์ „ํžˆ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ๊ธ‰์ฆํ•จ์— ๋”ฐ๋ผ ์‹œ๊ฐ„ ๋ฐ ๋ฉ”๋ชจ๋ฆฌ ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ํฌํ•จํ•œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ํšจ์œจ์„ฑ์€ ๋Œ€๊ทœ๋ชจ ๋ฐฐํฌ, ํŠนํžˆ ์ˆ˜๋งŽ์€ LLM ์ธ์Šคํ„ด์Šค๊ฐ€ ์žˆ๋Š” ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์˜ ๊ฒฝ์šฐ ์ค‘์š”ํ•œ ๋ฌธ์ œ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ์ˆ˜๊ฐ€ ํ™•์žฅ๋จ์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ ๊ฐ„์˜ ์กฐ์ • ๋ณต์žก์„ฑ ์ฆ๊ฐ€๋ฅผ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด ๋ณด๋‹ค ํšจ๊ณผ์ ์ด๊ณ  ํšจ์œจ์ ์ธ ํ†ต์‹  ํ”„๋กœํ† ์ฝœ๊ณผ ์•„ํ‚คํ…์ฒ˜๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์œ ๋Šฅํ•œ ์—์ด์ „ํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋ ค๋ฉด ๋ช…๋ น์–ด ์ถ”์ข… ๋ฐ ๊ธด ํ…์ŠคํŠธ ๋ชจ๋ธ๋ง๊ณผ ๊ฐ™์€ LLM์˜ ์—ญ๋Ÿ‰์— ๋Œ€ํ•œ ๊ธฐ์ˆ ์  ๊ณผ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด LLM์€ ์—์ด์ „ํŠธ๋ฅผ ์ธ์Šคํ„ด์Šคํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ์ตœ์ ํ™”๋˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์—, ๋Œ€๋ถ€๋ถ„์˜ ๊ณต๊ฐœ ์†Œ์Šค LLM(์˜ˆ: LLaMA)์€ ์—์ด์ „ํŠธ ๊ฐœ๋ฐœ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ด‰์ง„ํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—์ด์ „ํŠธ์˜ ํ•ต์‹ฌ ๊ณ„์‚ฐ ๋‹จ์œ„ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์ด ์žˆ๋Š” ํŠน์ˆ˜ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Remaining Issues.** Despite the huge success, there are still several issues that limit the development and applications of LLM-based agents. First, with the explosive growth of the model scale, the efficiency of LLM-based agents, including both the time and memory overhead, becomes an important issue for large-scale deployment, especially for multi-agent systems with numerous instances of LLMs. Second, with the scaling of the number of LLM-based agents, more effective and efficient communication protocols and architectures are required to support the increased complexity of coordination among agents. Furthermore, building capable agents poses technical challenges for the capacities of LLMs like instruction following and long text modeling. Since existing LLMs are not specially optimized for instantiating agents, most public-sourced LLMs like LLaMA cannot effectively facilitate the development of agents. Therefore, it is crucial to develop capable, specialized models to serve as the core computation unit of agents. [|assistant|]
**DyLAN.** ๋™์  LLM-Agent Network (DyLAN) ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ์ถ”๋ก  ๋ฐ ์ฝ”๋“œ ์ƒ์„ฑ๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ์ž‘์—…์— ์ค‘์ ์„ ๋‘” ๋™์  ์—์ด์ „ํŠธ ๊ตฌ์กฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค[18]. DyLAN์€ ์ง€๋‚œ ์ž‘์—… ๋ผ์šด๋“œ์—์„œ ๊ฐ ์—์ด์ „ํŠธ๊ฐ€ ์–ผ๋งˆ๋‚˜ ๊ธฐ์—ฌํ–ˆ๋Š”์ง€ ๊ฒฐ์ •ํ•˜๋Š” ๋‹จ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ, ๋‹ค์Œ ์‹คํ–‰ ๋ผ์šด๋“œ์—์„œ๋Š” ๊ธฐ์—ฌ๋„๊ฐ€ ๋†’์€ ์—์ด์ „ํŠธ๋งŒ ์ด๋™ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์—์ด์ „ํŠธ๊ฐ€ ์„œ๋กœ ์ •๋ณด๋ฅผ ๊ณต์œ ํ•  ์ˆ˜ ์žˆ๊ณ  ์ •์˜๋œ ๋ฆฌ๋”๊ฐ€ ์—†๊ธฐ ๋•Œ๋ฌธ์— ๋ณธ์งˆ์ ์œผ๋กœ ์ˆ˜ํ‰์ ์ž…๋‹ˆ๋‹ค. DyLAN์€ arithmetic ๋ฐ general reasoning ๊ธฐ๋Šฅ์„ ์ธก์ •ํ•˜๋Š” ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Š” ๋™์  ํŒ€์˜ ์˜ํ–ฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์—์ด์ „ํŠธ์˜ ๊ธฐ์—ฌ๋„๋ฅผ ์ง€์†์ ์œผ๋กœ ์žฌํ‰๊ฐ€ํ•˜๊ณ  ์ˆœ์œ„๋ฅผ ๋งค๊ธฐ๋ฉด ์ฃผ์–ด์ง„ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๋Š” ๋ฐ ๋” ์ ํ•ฉํ•œ ์—์ด์ „ํŠธ ํŒ€์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
**DyLAN.** ๋™์  LLM ์—์ด์ „ํŠธ ๋„คํŠธ์›Œํฌ(DyLAN) ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ์ถ”๋ก  ๋ฐ ์ฝ”๋“œ ์ƒ์„ฑ๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ์ž‘์—…์— ์ค‘์ ์„ ๋‘” ๋™์  ์—์ด์ „ํŠธ ๊ตฌ์กฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค[18]. DyLAN์€ ์ง€๋‚œ ์ž‘์—… ๋ผ์šด๋“œ์—์„œ ๊ฐ ์—์ด์ „ํŠธ๊ฐ€ ์–ผ๋งˆ๋‚˜ ๊ธฐ์—ฌํ–ˆ๋Š”์ง€ ๊ฒฐ์ •ํ•˜๋Š” ๋‹จ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ, ๋‹ค์Œ ์‹คํ–‰ ๋ผ์šด๋“œ์—์„œ๋Š” ๊ธฐ์—ฌ๋„๊ฐ€ ๋†’์€ ์—์ด์ „ํŠธ๋งŒ ์ด๋™ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์—์ด์ „ํŠธ๊ฐ€ ์„œ๋กœ ์ •๋ณด๋ฅผ ๊ณต์œ ํ•  ์ˆ˜ ์žˆ๊ณ  ์ •์˜๋œ ๋ฆฌ๋”๊ฐ€ ์—†๊ธฐ ๋•Œ๋ฌธ์— ๋ณธ์งˆ์ ์œผ๋กœ ์ˆ˜ํ‰์ ์ž…๋‹ˆ๋‹ค. DyLAN์€ ์‚ฐ์ˆ  ๋ฐ ์ผ๋ฐ˜ ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ์ธก์ •ํ•˜๋Š” ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Š” ๋™์  ํŒ€์˜ ์˜ํ–ฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์—์ด์ „ํŠธ์˜ ๊ธฐ์—ฌ๋„๋ฅผ ์ง€์†์ ์œผ๋กœ ์žฌํ‰๊ฐ€ํ•˜๊ณ  ์ˆœ์œ„๋ฅผ ๋งค๊ธฐ๋ฉด ์ฃผ์–ด์ง„ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๋Š” ๋ฐ ๋” ์ ํ•ฉํ•œ ์—์ด์ „ํŠธ ํŒ€์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**DyLAN.** The Dynamic LLM-Agent Network (DyLAN) framework creates a dynamic agent structure that focuses on complex tasks like reasoning and code generation [18]. DyLAN has a specific step for determining how much each agent has contributed in the last round of work and only moves top contributors the next round of execution. This method is horizontal in nature since agents can share information with each other and there is no defined leader. DyLAN shows improved performance on a variety of benchmarks which measure arithmetic and general reasoning capabilities. This highlights the impact of dynamic teams and demonstrates that by consistently re-evaluating and ranking agent contributions, we can create agent teams that are better suited to complete a given task. [|assistant|]
ํ•™์Šต๋œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋ฐ์ดํ„ฐ \(\mathcal{D}_{\textbf{val}}\)์— ๋Œ€ํ•œ Loss๋ฅผ ๊ณ„์‚ฐํ•œ๋‹ค.[|endofturn|]
ํ•™์Šต๋œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋ฐ์ดํ„ฐ \(\mathcal{D}_{\textbf{val}}\)์— ๋Œ€ํ•œ ์†์‹ค์„ ๊ณ„์‚ฐํ•œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To evaluate the learned model, we compute the loss on validation data \(\mathcal{D}_{\textbf{val}}\). [|assistant|]
1. A-1: 2.4B model, ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋งŒ ์‚ฌ์šฉํ•˜์—ฌ decay ํ›„ 4B token SFT. 2. A-2: 2.4B model, ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ์™€ SFT ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํ˜ผํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ•œ decay, ์ด์–ด์„œ 4B token SFT. 3. B-1: 1.2B model, ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋งŒ ์‚ฌ์šฉํ•˜์—ฌ decay ํ›„ 6B token SFT. 4. B-2: 1.2B model, ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋งŒ ์‚ฌ์šฉํ•˜์—ฌ decay ํ›„ 12B token SFT. 5. B-3: 1.2B model, ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ + ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํ˜ผํ•ฉ๋œ SFT ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ annealingํ•œ ํ›„ 6B token SFT.[|endofturn|]
1. A-1: 2.4B ๋ชจ๋ธ, ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋งŒ ์‚ฌ์šฉํ•˜์—ฌ ๋ถ•๊ดด ํ›„ 4B ํ† ํฐ SFT. 2. A-2: 2.4B ๋ชจ๋ธ, ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ์™€ SFT ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํ˜ผํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ•œ ๋ถ•๊ดด, ์ด์–ด์„œ 4B ํ† ํฐ SFT. 3. B-1: 1.2B ๋ชจ๋ธ, ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋งŒ ์‚ฌ์šฉํ•˜์—ฌ ๋ถ•๊ดด ํ›„ 6B ํ† ํฐ SFT. 4. B-2: 1.2B ๋ชจ๋ธ, ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋งŒ ์‚ฌ์šฉํ•˜์—ฌ ๋ถ•๊ดด ํ›„ 12B ํ† ํฐ SFT. 5. B-3: 1.2B ๋ชจ๋ธ, ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ + ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํ˜ผํ•ฉ๋œ SFT ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์–ด๋‹๋งํ•œ ํ›„ 6B ํ† ํฐ SFT.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]1. A-1: 2.4B model, decay using only pre-training data, followed by 4B token SFT. 2. A-2: 2.4B model, decay using the aforementioned high-quality data unlabeled data and SFT data mixed into pre-training data, also followed by 4B token SFT. 3. B-1: 1.2B model, decay using only pre-training data, followed by 6B token SFT. 4. B-2: 1.2B model, decay using only pre-training data, followed by 12B token SFT. 5. B-3: 1.2B model, annealing using the aforementioned high-quality data + SFT data mixed into pre-training data, also followed by 6B token SFT. [|assistant|]
๊ทธ๋ฆผ 10: "I am sleepy. I start a pot of" ๋ฌธ๋งฅ์˜ ๋‹ค์Œ ํ† ํฐ์— ๋Œ€ํ•ด ์–ดํœ˜์— ๋Œ€ํ•œ ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ๋‚ด๋ฆผ์ฐจ์ˆœ์œผ๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋…ผ์˜์˜ ํŽธ์˜๋ฅผ ์œ„ํ•ด ์ด ์˜ˆ๋Š” subword ๋‹จ์œ„๊ฐ€ ์•„๋‹Œ ๋‹จ์–ด ๋‹จ์œ„๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 10: "์ž ์ด ๋“ ๋‹ค. ๋ƒ„๋น„๋ฅผ ๋“์ด๊ธฐ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค."_ ๋ฌธ๋งฅ์˜ ๋‹ค์Œ ํ† ํฐ์— ๋Œ€ํ•ด ์–ดํœ˜์— ๋Œ€ํ•œ ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ๋‚ด๋ฆผ์ฐจ์ˆœ์œผ๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋…ผ์˜์˜ ํŽธ์˜๋ฅผ ์œ„ํ•ด ์ด ์˜ˆ๋Š” ํ•˜์œ„ ๋‹จ์œ„๊ฐ€ ์•„๋‹Œ ๋‹จ์–ด ๋‹จ์œ„๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Fig. 10: The probability distribution over the vocabulary in descending order for the next token of the context โ€œ_I am sleepy. I start a pot ofโ€_. For ease of discussion, this example is given in word units instead of subword units. [|assistant|]
COK(Li et al., 2023)์€ ๊ฒ€์ƒ‰๋œ ์ง€์‹์œผ๋กœ ๊ทผ๊ฑฐ๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” Progressive Rationale Correction ๊ธฐ๋ฒ•์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์ง€์†์ ์ธ optimization ํ”„๋กœ์„ธ์Šค๋กœ ๊ตฌ์„ฑ๋˜์–ด ์ฝ˜ํ…์ธ  ์ƒ์„ฑ์— ์‚ฌ์šฉ๋˜๋Š” ์ •๋ณด์˜ ๊ด€๋ จ์„ฑ๊ณผ ํ’ˆ์งˆ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. Self-RAG(Asai et al., 2023)๋Š” ๊ด€๋ จ์„ฑ ์—†๋Š” ์ฝ˜ํ…์ธ ๋ฅผ ํšจ์œจ์ ์œผ๋กœ ํ•„ํ„ฐ๋งํ•˜๊ธฐ ์œ„ํ•œ self-reflection ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. critique tokens์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์˜ ๊ด€๋ จ์„ฑ, ์ง€์ง€๋ ฅ ๋ฐ ์œ ์šฉ์„ฑ์„ ํ‰๊ฐ€ํ•˜์—ฌ ๊ณ ํ’ˆ์งˆ ์ •๋ณด๋งŒ ์ฝ˜ํ…์ธ  ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค์— ํ†ตํ•ฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
COK(Li et al., 2023)์€ ๊ฒ€์ƒ‰๋œ ์ง€์‹์œผ๋กœ ๊ทผ๊ฑฐ๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ์ ์ง„์  ๊ทผ๊ฑฐ ์ˆ˜์ • ๊ธฐ๋ฒ•์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์ง€์†์ ์ธ ์ตœ์ ํ™” ํ”„๋กœ์„ธ์Šค๋กœ ๊ตฌ์„ฑ๋˜์–ด ์ฝ˜ํ…์ธ  ์ƒ์„ฑ์— ์‚ฌ์šฉ๋˜๋Š” ์ •๋ณด์˜ ๊ด€๋ จ์„ฑ๊ณผ ํ’ˆ์งˆ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์…€ํ”„-RAG(Asai et al., 2023)๋Š” ๊ด€๋ จ์„ฑ ์—†๋Š” ์ฝ˜ํ…์ธ ๋ฅผ ํšจ์œจ์ ์œผ๋กœ ํ•„ํ„ฐ๋งํ•˜๊ธฐ ์œ„ํ•œ ์ž๊ธฐ ๋ฐ˜์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. ๋น„ํ‰ ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์˜ ๊ด€๋ จ์„ฑ, ์ง€์ง€๋ ฅ ๋ฐ ์œ ์šฉ์„ฑ์„ ํ‰๊ฐ€ํ•˜์—ฌ ๊ณ ํ’ˆ์งˆ ์ •๋ณด๋งŒ ์ฝ˜ํ…์ธ  ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค์— ํ†ตํ•ฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]COK (Li et al., 2023) presents the Progressive Rationale Correction technique, aimed at iteratively refining rationales with retrieved knowledge. This method constitutes a continuous optimization process, significantly enhancing the relevance and quality of information used in content generation. Self-RAG (Asai et al., 2023) introduces a self-reflection mechanism to efficiently filter out irrelevant content. By employing critique tokens, this approach evaluates the relevance, supportiveness, and utility of retrieved passages, ensuring the integration of only high-quality information into the content generation process. [|assistant|]
์ตœ๊ณ ์˜ zero-shot ๋ฐ scratchpad ํ”„๋กฌํ”„ํŠธ๋Š” ๊ทธ๋ฆผ 5์™€ ๊ทธ๋ฆผ 6์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘ ํ”„๋กฌํ”„ํŠธ ์Šคํƒ€์ผ ๋ชจ๋‘์—์„œ ๋ชจ๋ธ์—๋Š” ์งˆ๋ฌธ, ๋ฐฐ๊ฒฝ, ํ•ด์ƒ๋„ ๊ธฐ์ค€, ์งˆ๋ฌธ์˜ ๊ฐœ์‹œ์ผ ๋ฐ ์ข…๋ฃŒ์ผ(date_begin ๋ฐ date_end)๋งŒ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ฐ์ดํ„ฐ๋Š” ์˜ˆ์ธก ํ”Œ๋žซํผ์—์„œ ์†Œ์‹ฑ๋˜๋ฉฐ ์งˆ๋ฌธ ํŽ˜์ด์ง€์—์„œ ์ผ๋ฐ˜ ์˜ˆ์ธก์ž์—๊ฒŒ ๊ณต๊ฐœ๋ฉ๋‹ˆ๋‹ค. ์ถ”๊ฐ€ ๋‰ด์Šค retrieval์€ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ตœ๊ณ ์˜ ์ œ๋กœ ์ƒท ๋ฐ ์Šคํฌ๋ž˜์น˜ ํŒจ๋“œ ํ”„๋กฌํ”„ํŠธ๋Š” ๊ทธ๋ฆผ 5์™€ ๊ทธ๋ฆผ 6์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘ ํ”„๋กฌํ”„ํŠธ ์Šคํƒ€์ผ ๋ชจ๋‘์—์„œ ๋ชจ๋ธ์—๋Š” ์งˆ๋ฌธ, ๋ฐฐ๊ฒฝ, ํ•ด์ƒ๋„ ๊ธฐ์ค€, ์งˆ๋ฌธ์˜ ๊ฐœ์‹œ์ผ ๋ฐ ์ข…๋ฃŒ์ผ(date_begin ๋ฐ date_end)๋งŒ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ฐ์ดํ„ฐ๋Š” ์˜ˆ์ธก ํ”Œ๋žซํผ์—์„œ ์†Œ์‹ฑ๋˜๋ฉฐ ์งˆ๋ฌธ ํŽ˜์ด์ง€์—์„œ ์ผ๋ฐ˜ ์˜ˆ์ธก์ž์—๊ฒŒ ๊ณต๊ฐœ๋ฉ๋‹ˆ๋‹ค. ์ถ”๊ฐ€ ๋‰ด์Šค ๊ฒ€์ƒ‰์€ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The best zero-shot and scratchpad prompts are shown in Figure 5 and Figure 6. In both prompting styles, models are only provided with the question, background, resolution criterion, and question's open and close dates (date_begin and date_end). All the data are sourced from the forecasting platforms and publicly available on the question page to human forecasters. We do no additional news retrieval. [|assistant|]
# ์ถœ๋ ฅ 2024๋…„ 3์›” 14์ผ์— Meta ๋‚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.[|endofturn|]
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.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]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. [|assistant|]
๊ฐ์ฃผ 6: ๋ณธ ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋œ LLaMA-2(Towron et al., 2023) ๋ชจ๋ธ์˜ training ์‹œ์ ์„ ๊ณ ๋ คํ•˜์—ฌ ์ด timestamp๋ฅผ ์„ ํƒํ•˜์˜€๋‹ค.[|endofturn|]
๊ฐ์ฃผ 6: ๋ณธ ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋œ LLaMA-2(Towron et al., 2023) ๋ชจ๋ธ์˜ ํ•™์Šต ์‹œ์ ์„ ๊ณ ๋ คํ•˜์—ฌ ์ด ํƒ€์ž„์Šคํƒฌํ”„๋ฅผ ์„ ํƒํ•˜์˜€๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Footnote 6: We choose this timestamp considering the training time of LLaMA-2 (Towron et al., 2023) models that we use for our study. [|assistant|]
* Figure 10์€ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ชจ๋“  ์งˆ๋ฌธ์— ์–ธ๊ธ‰๋œ ์œ„์น˜๋ฅผ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. * Table 11์€ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ”Œ๋žซํผ๋ณ„ ์งˆ๋ฌธ ๋ฐ ์˜ˆ์ธก์˜ ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. * Table 12๋Š” ํ๋ ˆ์ด์…˜๋œ ์„ธํŠธ์˜ ์ „์ฒด ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. * Table 13์€ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ์ปค๋ฎค๋‹ˆํ‹ฐ ์˜ˆ์ธก์ด ์–ด๋–ป๊ฒŒ ๋ณ€ํ™”ํ•˜๋Š”์ง€ ๋ณด์—ฌ์ฃผ๋Š” ์งˆ๋ฌธ ๋ชฉ๋ก์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. * Figure 10(a)๋Š” ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์งˆ๋ฌธ์˜ ๊ฐœ์‹œ ๋‚ ์งœ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. * Figure 10(b)๋Š” ์ธ๋ฑ์Šค \(k=1,2,3,4,5\)์—์„œ ๊ฒ€์ƒ‰ ๋‚ ์งœ๋ฅผ ๋ฐ›๋Š” ์งˆ๋ฌธ์˜ ๋น„์œจ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
* ๊ทธ๋ฆผ 10์€ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ชจ๋“  ์งˆ๋ฌธ์— ์–ธ๊ธ‰๋œ ์œ„์น˜๋ฅผ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. * ํ‘œ 11์€ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ”Œ๋žซํผ๋ณ„ ์งˆ๋ฌธ ๋ฐ ์˜ˆ์ธก์˜ ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. * ํ‘œ 12๋Š” ํ๋ ˆ์ด์…˜๋œ ์„ธํŠธ์˜ ์ „์ฒด ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. * ํ‘œ 13์€ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ์ปค๋ฎค๋‹ˆํ‹ฐ ์˜ˆ์ธก์ด ์–ด๋–ป๊ฒŒ ๋ณ€ํ™”ํ•˜๋Š”์ง€ ๋ณด์—ฌ์ฃผ๋Š” ์งˆ๋ฌธ ๋ชฉ๋ก์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. * ๊ทธ๋ฆผ 10(a)๋Š” ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์งˆ๋ฌธ์˜ ๊ฐœ์‹œ ๋‚ ์งœ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. * ๊ทธ๋ฆผ 10(b)๋Š” ์ธ๋ฑ์Šค \(k=1,2,3,4,5\)์—์„œ ๊ฒ€์ƒ‰ ๋‚ ์งœ๋ฅผ ๋ฐ›๋Š” ์งˆ๋ฌธ์˜ ๋น„์œจ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* Figure 10 visualizes the location mentions in all the questions from our full dataset. * Table 11 gives the distribution of questions and forecasts across platforms in our full dataset. * Table 12 showcases a complete data sample in our curated set. * Table 13 shows a list of questions with how community predictions shift over time. * Figure 10(a) shows the opening dates of the questions in the full dataset. * Figure 10(b) shows the percentage of questions that receives the retrieval date at index \(k=1,2,3,4,5\). [|assistant|]
ํ›ˆ๋ จ Fine-tuning ํ›„, ์šฐ๋ฆฌ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Ivison ๋“ฑ์— ๋”ฐ๋ผ DPO ํ›ˆ๋ จ์„ ์ˆ˜ํ–‰ํ•œ๋‹ค.[|endofturn|]
ํ›ˆ๋ จ ๋ฏธ์„ธ ์กฐ์ • ํ›„, ์šฐ๋ฆฌ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Ivison ๋“ฑ์— ๋”ฐ๋ผ DPO ํ›ˆ๋ จ์„ ์ˆ˜ํ–‰ํ•œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]After instruction finetuning, we then use the following hyperparameters for DPO training, following Ivison et al. [|assistant|]
**ํŒ์— ๋Œ€ํ•œ ๊ธฐ์—ฌ**. ํ‘œ 12์— prompts ์„ค๊ณ„๋ฅผ ์œ„ํ•œ ์ œ๊ณต๋œ ํŒ์˜ ํ•ด๋‹น ์ˆ˜์— ๋Œ€ํ•œ ๊ธฐ์—ฌ๋„๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋‚˜์—ดํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**ํŒ์— ๋Œ€ํ•œ ๊ธฐ์—ฌ**. ํ‘œ 12์— ํ”„๋กฌํ”„ํŠธ ์„ค๊ณ„๋ฅผ ์œ„ํ•œ ์ œ๊ณต๋œ ํŒ์˜ ํ•ด๋‹น ์ˆ˜์— ๋Œ€ํ•œ ๊ธฐ์—ฌ๋„๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋‚˜์—ดํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Contribution on Tips**. We list the following guys for their contributions on the corresponding numbers of provided tips for designing prompts in Table 12. [|assistant|]
์ •ํ™•๋„๊ฐ€ ์•„์ง ์ •์ ์— ๋„๋‹ฌํ•˜์ง€ ์•Š์•˜๋‹ค๋Š” ์ ๋„ ์ฃผ๋ชฉํ•  ๊ฐ€์น˜๊ฐ€ ์žˆ๋‹ค. ์ถ”๊ฐ€์ ์ธ scaling์˜ ํšจ๊ณผ๋ฅผ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์€ ํ–ฅํ›„ ์—ฐ๊ตฌ๋กœ ๋‚จ๊ฒจ๋‘๊ธฐ๋กœ ํ•œ๋‹ค.[|endofturn|]
์ •ํ™•๋„๊ฐ€ ์•„์ง ์ •์ ์— ๋„๋‹ฌํ•˜์ง€ ์•Š์•˜๋‹ค๋Š” ์ ๋„ ์ฃผ๋ชฉํ•  ๊ฐ€์น˜๊ฐ€ ์žˆ๋‹ค. ์ถ”๊ฐ€์ ์ธ ์Šค์ผ€์ผ๋ง์˜ ํšจ๊ณผ๋ฅผ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์€ ํ–ฅํ›„ ์—ฐ๊ตฌ๋กœ ๋‚จ๊ฒจ๋‘๊ธฐ๋กœ ํ•œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]It is also worth noting that the accuracy has not yet reached its peak. Exploring the effects of additional scaling will be left as our future research. [|assistant|]
๋”ฐ๋ผ์„œ ๊ฐ ํ† ํฐ์˜ ํ™•๋ฅ ์€ ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ๋‹จ์–ด์™€์˜ ์ฐจ์ด๋ฅผ ๊ตฌํ•˜์—ฌ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ฐจ์ด๋ฅผ ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ํ•ด๋‹น ๋‹จ์–ด๋ฅผ ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๊ฒŒ ๋งŒ๋“œ๋Š” ์ตœ์†Œ logit bias๋ฅผ ์ฐพ์Šต๋‹ˆ๋‹ค. Algorithm 1์€ binary search์— ์˜์กดํ•˜์—ฌ ๊ฐ ๋‹จ์–ด์— ๋Œ€ํ•œ logit bias๋ฅผ ์ฐพ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
๋”ฐ๋ผ์„œ ๊ฐ ํ† ํฐ์˜ ํ™•๋ฅ ์€ ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ๋‹จ์–ด์™€์˜ ์ฐจ์ด๋ฅผ ๊ตฌํ•˜์—ฌ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ฐจ์ด๋ฅผ ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ํ•ด๋‹น ๋‹จ์–ด๋ฅผ ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๊ฒŒ ๋งŒ๋“œ๋Š” ์ตœ์†Œ ๋กœ๊ทธ ๋น„์ฆˆ๋ฅผ ์ฐพ์Šต๋‹ˆ๋‹ค. ์•Œ๊ณ ๋ฆฌ์ฆ˜ 1์€ ์ด์ง„ ๊ฒ€์ƒ‰์— ์˜์กดํ•˜์—ฌ ๊ฐ ๋‹จ์–ด์— ๋Œ€ํ•œ ๋กœ๊ทธ ๋น„์ฆˆ๋ฅผ ์ฐพ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The probability of each token can therefore be recovered by finding its difference with the most likely word. We compute this difference by finding the smallest logit bias to make that word most likely. Algorithm 1 shows the approach, which relies on binary search to find the logit bias for each word. [|assistant|]
* Scaling laws์˜ ํ‰๊ฐ€ ๋ฐ ์ถ”์ •์„ ํ†ตํ•ด ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ํฌํ™”๋ฅผ ํŠน์„ฑํ™”ํ•ฉ๋‹ˆ๋‹ค; * ์ž‘์€ ๋ชจ๋ธ์˜ ํ‘œํ˜„์ด ์ด๋Ÿฌํ•œ ํฌํ™”์™€ ๋™์‹œ์— ๋ณ€์งˆ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” _rank saturation_, ์ฆ‰ ์ž‘์€ LM ์˜ˆ์ธก ํ—ค๋“œ์˜ ํŠน์ด๊ฐ’ ๋ถ„ํฌ์˜ ์—”ํŠธ๋กœํ”ผ ํญ๋ฐœ์— ๋Œ€ํ•ด ์กฐ๋ช…ํ•ฉ๋‹ˆ๋‹ค; * ์šฐ๋ฆฌ๋Š” ๋ชฉํ‘œ ์ปจํ…์ŠคํŠธ ๋ถ„ํฌ์˜ ์ˆœ์œ„๊ฐ€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ ํ‘œํ˜„์˜ ํ‘œํ˜„๋ ฅ์— ๊ด€๊ณ„์—†์ด ์„ ํ˜• ํ—ค๋“œ \(W\)๊ฐ€ \(rank(W)<1000\)์ผ ๋•Œ ์„ฑ๋Šฅ์— ์ƒ๋‹นํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค; * ์ €์ˆœ์œ„ ์„ ํ˜• ์–ธ์–ด ๋ชจ๋ธ๋ง ํ—ค๋“œ์— ์˜ํ•ด ์œ ๋ฐœ๋˜๋Š” ์„ฑ๋Šฅ ์ œํ•œ์„ ์ด๋ก ์ ์œผ๋กœ ์ •๋Ÿ‰ํ™”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
* ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์˜ ํ‰๊ฐ€ ๋ฐ ์ถ”์ •์„ ํ†ตํ•ด ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ํฌํ™”๋ฅผ ํŠน์„ฑํ™”ํ•ฉ๋‹ˆ๋‹ค; * ์ž‘์€ ๋ชจ๋ธ์˜ ํ‘œํ˜„์ด ์ด๋Ÿฌํ•œ ํฌํ™”์™€ ๋™์‹œ์— ๋ณ€์งˆ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” _๋žญํฌ ํฌํ™”_, ์ฆ‰ ์ž‘์€ LM ์˜ˆ์ธก ํ—ค๋“œ์˜ ํŠน์ด๊ฐ’ ๋ถ„ํฌ์˜ ์—”ํŠธ๋กœํ”ผ ํญ๋ฐœ์— ๋Œ€ํ•ด ์กฐ๋ช…ํ•ฉ๋‹ˆ๋‹ค; * ์šฐ๋ฆฌ๋Š” ๋ชฉํ‘œ ์ปจํ…์ŠคํŠธ ๋ถ„ํฌ์˜ ์ˆœ์œ„๊ฐ€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ ํ‘œํ˜„์˜ ํ‘œํ˜„๋ ฅ์— ๊ด€๊ณ„์—†์ด ์„ ํ˜• ํ—ค๋“œ \(W\)๊ฐ€ \(rank(W)<1000\)์ผ ๋•Œ ์„ฑ๋Šฅ์— ์ƒ๋‹นํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค; * ์ €์ˆœ์œ„ ์„ ํ˜• ์–ธ์–ด ๋ชจ๋ธ๋ง ํ—ค๋“œ์— ์˜ํ•ด ์œ ๋ฐœ๋˜๋Š” ์„ฑ๋Šฅ ์ œํ•œ์„ ์ด๋ก ์ ์œผ๋กœ ์ •๋Ÿ‰ํ™”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* We characterize the performance saturation of small language models through evaluation and extrapolation of the scaling laws; * We find that the representations of smaller models degenerate concurrently with this saturation. We shed light on _rank saturation_, i.e. the explosion of the entropy of singular value distributions of small LM prediction heads; * We empirically verify that the rank of the target contextual distribution is usually high. Moreover, we observe that regardless of the expressiveness of the output representations of a model, a linear head \(W\) substantially affects performance when \(rank(W)<1000\); * We theoretically quantify the performance limitation induced by a low-rank linear language modeling head. [|assistant|]
๊ฐ€์žฅ ์ตœ๊ทผ์— ChatGPT(OpenAI, 2022)์™€ ๊ทธ ํ›„์† ๋ชจ๋ธ(OpenAI, 2023; Anthropic, 2023; Google, 2023)์€ ํ”„๋กœ๋•์…˜ ๋ฐ ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ question answering (QA) ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๋Š” ํŒจ๋Ÿฌ๋‹ค์ž„์˜ ๋ณ€ํ™”๋ฅผ ์ฃผ๋„ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์‹ค์ œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ QA ๋ชจ๋ธ์˜ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ธก๋ฉด์ด ์„ ํ˜ธ๋ฉ๋‹ˆ๋‹ค: _i)_ ์‚ฌ์šฉ์ž๋Š” ๋Œ€ํ™” ๋ฐฉ์‹์œผ๋กœ QA ๋ชจ๋ธ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํ›„์† ์งˆ๋ฌธ์„ ์‰ฝ๊ฒŒ ์ œ๊ธฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. _ii)_ ๋ฒ”์šฉ ๋ชจ๋ธ์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณ„ Fine-tuning ์—†์ด๋„ zero-shot ๋ฐฉ์‹์œผ๋กœ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, Fine-tuning๋œ ์ „๋ฌธ๊ฐ€ ๋ชจ๋ธ์˜ ์ •ํ™•๋„์™€ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. _iii)_ QA ๋ชจ๋ธ์€ ์ œ๊ณต๋œ ์ปจํ…์ŠคํŠธ๊ฐ€ LLM์˜ ์ปจํ…์ŠคํŠธ ์ฐฝ๋ณด๋‹ค ํ›จ์”ฌ ๊ธด ์˜คํ”ˆ ๋„๋ฉ”์ธ ๋˜๋Š” ๊ธด ๋ฌธ์„œ ์„ค์ •์—์„œ ๊ฒ€์ƒ‰๋œ ์ฆ๊ฑฐ ์ฒญํฌ๋ฅผ ํ†ตํ•ฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: Anthropic, 2023; Xu et al., 2023b). ์ด๋ฅผ ์œ„ํ•ด ์šฐ๋ฆฌ๋Š” ์ด ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์„ ํฌ๊ด„ํ•˜๋Š” ๋Œ€ํ™”ํ˜• QA์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ฐ€์žฅ ์ตœ๊ทผ์— ChatGPT(OpenAI, 2022)์™€ ๊ทธ ํ›„์† ๋ชจ๋ธ(OpenAI, 2023; Anthropic, 2023; Google, 2023)์€ ํ”„๋กœ๋•์…˜ ๋ฐ ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ์งˆ์˜์‘๋‹ต(QA) ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๋Š” ํŒจ๋Ÿฌ๋‹ค์ž„์˜ ๋ณ€ํ™”๋ฅผ ์ฃผ๋„ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์‹ค์ œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ QA ๋ชจ๋ธ์˜ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ธก๋ฉด์ด ์„ ํ˜ธ๋ฉ๋‹ˆ๋‹ค: _i)_ ์‚ฌ์šฉ์ž๋Š” ๋Œ€ํ™” ๋ฐฉ์‹์œผ๋กœ QA ๋ชจ๋ธ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํ›„์† ์งˆ๋ฌธ์„ ์‰ฝ๊ฒŒ ์ œ๊ธฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. _ii)_ ๋ฒ”์šฉ ๋ชจ๋ธ์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณ„ ๋ฏธ์„ธ ์กฐ์ • ์—†์ด๋„ 0์ƒท ๋ฐฉ์‹์œผ๋กœ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋ฏธ์„ธ ์กฐ์ •๋œ ์ „๋ฌธ๊ฐ€ ๋ชจ๋ธ์˜ ์ •ํ™•๋„์™€ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. _iii)_ QA ๋ชจ๋ธ์€ ์ œ๊ณต๋œ ์ปจํ…์ŠคํŠธ๊ฐ€ LLM์˜ ์ปจํ…์ŠคํŠธ ์ฐฝ๋ณด๋‹ค ํ›จ์”ฌ ๊ธด ์˜คํ”ˆ ๋„๋ฉ”์ธ ๋˜๋Š” ๊ธด ๋ฌธ์„œ ์„ค์ •์—์„œ ๊ฒ€์ƒ‰๋œ ์ฆ๊ฑฐ ์ฒญํฌ๋ฅผ ํ†ตํ•ฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: Anthropic, 2023; Xu et al., 2023b). ์ด๋ฅผ ์œ„ํ•ด ์šฐ๋ฆฌ๋Š” ์ด ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์„ ํฌ๊ด„ํ•˜๋Š” ๋Œ€ํ™”ํ˜• QA์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Most recently, ChatGPT (OpenAI, 2022) and its follow ups (OpenAI, 2023; Anthropic, 2023; Google, 2023) have led to the paradigm shift of building question answering (QA) models in production and research community. In particular, the following aspects of the QA models are preferred in real-world applications: _i)_ The users can interact with the QA models in a conversational way, thus one can easily raise follow-up questions. _ii)_ The generalist models can generate answers in zero-shot manner without dataset-specific fine-tuning, while matching the accuracies of fine-tuned expert models. _iii)_ The QA models are capable of integrating retrieved chunks of evidence in both open-domain or long document settings, where the provided context is much longer than the context window of LLM (e.g., Anthropic, 2023; Xu et al., 2023b). To this end, we focus on the conversational QA covering these three aspects. [|assistant|]
**Key Ingredients.** ์ผ๋ฐ˜์ ์œผ๋กœ ์ž‘์—… ์„ค๋ช…, ์ž…๋ ฅ ๋ฐ์ดํ„ฐ, contextual information, prompt style ๋“ฑ LLM์˜ ๊ธฐ๋Šฅ์„ ์™„๋ฃŒํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ํ”„๋กฌํ”„ํŠธ์˜ ๊ธฐ๋Šฅ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๋„ค ๊ฐ€์ง€ ์ฃผ์š” ์„ฑ๋ถ„์ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ† ๋ก ์— ๋Œ€ํ•œ ์ง๊ด€์ ์ธ ์ดํ•ด๋ฅผ ๋•๊ธฐ ์œ„ํ•ด question answering, meta-review generation, text-to-SQL์— ๋Œ€ํ•œ ์„ธ ๊ฐ€์ง€ ํ”„๋กฌํ”„ํŠธ ์˜ˆ์ œ๋„ ํ‘œ XIII XIII์— ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**์ฃผ์š” ์„ฑ๋ถ„**. ์ผ๋ฐ˜์ ์œผ๋กœ ์ž‘์—… ์„ค๋ช…, ์ž…๋ ฅ ๋ฐ์ดํ„ฐ, ์ปจํ…์ŠคํŠธ ์ •๋ณด, ํ”„๋กฌํ”„ํŠธ ์Šคํƒ€์ผ ๋“ฑ LLM์˜ ๊ธฐ๋Šฅ์„ ์™„๋ฃŒํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ํ”„๋กฌํ”„ํŠธ์˜ ๊ธฐ๋Šฅ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๋„ค ๊ฐ€์ง€ ์ฃผ์š” ์„ฑ๋ถ„์ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ† ๋ก ์— ๋Œ€ํ•œ ์ง๊ด€์ ์ธ ์ดํ•ด๋ฅผ ๋•๊ธฐ ์œ„ํ•ด ์งˆ๋ฌธ ๋‹ต๋ณ€, ๋ฉ”ํƒ€ ๋ฆฌ๋ทฐ ์ƒ์„ฑ, ํ…์ŠคํŠธ-SQL์— ๋Œ€ํ•œ ์„ธ ๊ฐ€์ง€ ํ”„๋กฌํ”„ํŠธ ์˜ˆ์ œ๋„ ํ‘œ XIII XIII์— ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Key Ingredients.** Typically, there are four key ingredients that depict the functionality of a prompt for eliciting the abilities of LLMs to complete the tasks, including task description, input data, contextual information, and prompt style. To have an intuitive understanding of our discussion, we also present three prompt examples for question answering, meta-review generation, and text-to-SQL in Table XIII XIII. [|assistant|]
๋งค๊ฐœ๋ณ€์ˆ˜ 6(๊ทธ๋ฆผ 3). 1000ํšŒ ๋…ธ์ถœ ์„ค์ •์—์„œ LLaMA/Mistral ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋งค๊ฐœ๋ณ€์ˆ˜ 1์— ๋ช…์‹œ๋œ ๊ฒƒ๊ณผ ์œ ์‚ฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์ง€๋งŒ, GPT2๊ฐ€ _์ตœ์ ์˜ ํŠœ๋‹๋œ_ LLaMA/Mistral ๋ชจ๋ธ๋ณด๋‹ค _๋” ๋‚˜์˜์ง€ ์•Š์€_ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๊ธฐ ์œ„ํ•ด ์„ธ ๊ฐ€์ง€ Learning Rate ์ค‘ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ์†๋„๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค:[|endofturn|]
๋งค๊ฐœ๋ณ€์ˆ˜ 6(๊ทธ๋ฆผ 3).1000ํšŒ ๋…ธ์ถœ ์„ค์ •์—์„œ LLaMA/Mistral ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋งค๊ฐœ๋ณ€์ˆ˜ 1์— ๋ช…์‹œ๋œ ๊ฒƒ๊ณผ ์œ ์‚ฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์ง€๋งŒ, GPT2๊ฐ€ _์ตœ์ ์˜ ํŠœ๋‹๋œ_ LLaMA/Mistral ๋ชจ๋ธ๋ณด๋‹ค _๋” ๋‚˜์˜์ง€ ์•Š์€_ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๊ธฐ ์œ„ํ•ด ์„ธ ๊ฐ€์ง€ ํ•™์Šต ์†๋„ ์ค‘ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ์†๋„๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค:[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Parameter 6 (Figure 3).In the 1000-exposure setting, for LLaMA/Mistral models we use similar parameters as specified in Parameter 1, but we select the best of three learning rates to better demonstrate that GPT2 performs _no worse_ than _even the best tuned_ LLaMA/Mistral models: [|assistant|]
๊ณ ์ •๋œ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜์—์„œ \(\mathbf{r}=r_{1\ldots M}\)์€ \(M\) ๋„๋ฉ”์ธ์˜ ๋น„์œจ์ž…๋‹ˆ๋‹ค. ์ด ํ•จ์ˆ˜๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์›ํ•˜๋Š” ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ํ˜ผํ•ฉ๋ฌผ์„ ์ฐพ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ณ ์ •๋œ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜์—์„œ \mathbf{r}=r_{1\์  M}\)์€ \(M\) ๋„๋ฉ”์ธ์˜ ๋น„์œจ์ž…๋‹ˆ๋‹ค. ์ด ํ•จ์ˆ˜๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์›ํ•˜๋Š” ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ํ˜ผํ•ฉ๋ฌผ์„ ์ฐพ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]under the fixed model sizes and number of training steps, where \(\mathbf{r}=r_{1\ldots M}\) is the proportion of the \(M\) domains. Harnessing this function, we seek a mixture that achieves the desired performance. [|assistant|]
๊ทธ๋ฆผ A7: ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ ํฌ๊ธฐ๊ฐ€ ํ›ˆ๋ จ ์ง‘ํ•ฉ์—์„œ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ด์›ƒ์— ๋Œ€ํ•œ cosine distance์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. x์ถ•์—๋Š” ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์˜ ํฌ๊ธฐ๋ฅผ ๋ณ€๊ฒฝํ•˜๊ณ (์›๋ž˜ ๋” ํฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ), y์ถ•์€ ํ›ˆ๋ จ ์ง‘ํ•ฉ์—์„œ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ด์›ƒ๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ(์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ ์ „์ฒด์— ํ‰๊ท )๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์›๋ž˜ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์˜ ์ƒ˜ํ”Œ๋ง ๋น„์œจ์— ๊ด€๊ณ„์—†์ด ํ›ˆ๋ จ ์ง‘ํ•ฉ์—์„œ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ด์›ƒ๊นŒ์ง€์˜ ํ‰๊ท  ๊ฑฐ๋ฆฌ๊ฐ€ ๋ณ€ํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ ๊ทธ๋ฆผ 5๊ฐ€ ๋‹ค๋ฅธ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ ํฌ๊ธฐ ๋•Œ๋ฌธ์ด ์•„๋‹˜์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ A7: ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ ํฌ๊ธฐ๊ฐ€ ํ›ˆ๋ จ ์ง‘ํ•ฉ์—์„œ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ด์›ƒ์— ๋Œ€ํ•œ ์ฝ”์‚ฌ์ธ ๊ฑฐ๋ฆฌ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. x์ถ•์—๋Š” ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์˜ ํฌ๊ธฐ๋ฅผ ๋ณ€๊ฒฝํ•˜๊ณ (์›๋ž˜ ๋” ํฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ), y์ถ•์€ ํ›ˆ๋ จ ์ง‘ํ•ฉ์—์„œ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ด์›ƒ๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ(์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ ์ „์ฒด์— ํ‰๊ท )๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์›๋ž˜ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์˜ ์ƒ˜ํ”Œ๋ง ๋น„์œจ์— ๊ด€๊ณ„์—†์ด ํ›ˆ๋ จ ์ง‘ํ•ฉ์—์„œ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ด์›ƒ๊นŒ์ง€์˜ ํ‰๊ท  ๊ฑฐ๋ฆฌ๊ฐ€ ๋ณ€ํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ ๊ทธ๋ฆผ 5๊ฐ€ ๋‹ค๋ฅธ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ ํฌ๊ธฐ ๋•Œ๋ฌธ์ด ์•„๋‹˜์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure A7: Studying the effect of validation set size on cosine distance to nearest-neighbor in training set. On the x-axis, we vary the size of the validation set (by randomly sampling the original larger validation set), and the y-axis represents distance to nearest neighbor in the training set (averaged across the validation set). We observe that regardless of what fraction of the original validation set is sampled, the mean distance to the nearest neighbor in train does not change, indicating that Figure 5 is not due to different validation set sizes. [|assistant|]
\(\bullet\)_GPT-4V, GPT-4 Turbo ๋ฐ ๊ทธ ์ด์ƒ_. OpenAI๋Š” GPT-4 [46]์— ๋Œ€ํ•œ ์ž‘์—…์„ ๊ธฐ๋ฐ˜์œผ๋กœ 2023๋…„ 9์›”์— GPT-4์˜ ๋น„์ „ ๊ธฐ๋Šฅ์˜ ์•ˆ์ „ํ•œ ๋ฐฐํฌ์— ์ค‘์ ์„ ๋‘” GPT-4V๋ฅผ ์ถ”๊ฐ€๋กœ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-4V์˜ ์‹œ์Šคํ…œ ์นด๋“œ [133]์—์„œ๋Š” ์‹œ๊ฐ์ ์œผ๋กœ ์ฆ๊ฐ•๋œ ์ž…๋ ฅ๊ณผ ๊ด€๋ จ๋œ ์œ„ํ—˜์˜ ํ‰๊ฐ€ ๋ฐ ์™„ํ™”์— ๋Œ€ํ•ด ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ๋…ผ์˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ GPT-4V๋Š” ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๊ฐ•๋ ฅํ•œ ๋น„์ „ ์—ญ๋Ÿ‰์„ ๋ณด์—ฌ์ฃผ๋ฉฐ ๊ฐ•๋ ฅํ•œ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ํ•™์Šต ์‹œ์Šคํ…œ์œผ๋กœ์„œ ํฐ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์ตœ๊ทผ์—๋Š” 2023๋…„ 11์›”์— OpenAI๊ฐ€ DevDay์—์„œ ์ผ๋ จ์˜ ๊ธฐ์ˆ  ๊ฐœ์„ ์„ ํ†ตํ•ด _GPT-4 Turbo_๋ผ๋Š” ์ด๋ฆ„์˜ ์—…๊ทธ๋ ˆ์ด๋“œ๋œ GPT-4 ๋ชจ๋ธ ์„ธ๋Œ€๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-4 Turbo๋Š” ํ–ฅ์ƒ๋œ ๋ชจ๋ธ ์šฉ๋Ÿ‰(GPT-4๋ณด๋‹ค ๋” ๊ฐ•๋ ฅํ•จ), ํ™•์žฅ๋œ ์ง€์‹ ์†Œ์Šค(์ตœ๋Œ€ 2023๋…„ 4์›”๊นŒ์ง€), ๊ธด context window(์ตœ๋Œ€ 128k tokens), ์ตœ์ ํ™”๋œ ๋ชจ๋ธ ์„ฑ๋Šฅ(๋” ์ €๋ ดํ•œ ๊ฐ€๊ฒฉ), ๊ธฐํƒ€ ์œ ์šฉํ•œ ๊ธฐ๋Šฅ ์—…๋ฐ์ดํŠธ(function call, reproducible outputs ๋“ฑ)๊ฐ€ ํŠน์ง•์ž…๋‹ˆ๋‹ค. ๋™์‹œ์— ์—์ด์ „ํŠธ์™€ ๊ฐ™์€ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ์‰ฝ๊ฒŒ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๋„๋ก Assistants API๊ฐ€ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด API๋ฅผ ํ†ตํ•ด ๊ฐœ๋ฐœ์ž๋Š” ํŠน์ • instruction, ์ถ”๊ฐ€ ์ง€์‹ ๋ฐ ๋„๊ตฌ ์‚ฌ์šฉ์„ ํ™œ์šฉํ•˜์—ฌ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋‚ด์—์„œ ๋ชฉํ‘œ ์ง€ํ–ฅ์ ์ธ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ์‰ฝ๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋ฒˆ ๋ฆด๋ฆฌ์Šค์—์„œ๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๊ธฐ๋Šฅ(๋ณด๊ธฐ, ๋“ฃ๊ธฐ, ๋งํ•˜๊ธฐ)๋„ ๊ฐ•ํ™”๋˜์–ด ๋น„์ „, DALL-E 3, Text-to-speech (TTS), ์Œ์„ฑ ์ƒ˜ํ”Œ ๋“ฃ๊ธฐ ๋“ฑ์œผ๋กœ GPT-4 Turbo๋ฅผ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์œผ๋กœ ๋ชจ๋ธ์˜ ์šฉ๋Ÿ‰ ๋ฒ”์œ„๊ฐ€ ํฌ๊ฒŒ ํ™•์žฅ๋˜๊ณ  ์ž‘์—… ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋”์šฑ ์ค‘์š”ํ•œ ๊ฒƒ์€ ๊ฐœ์„ ๋œ ๋ชจ๋ธ, API ๋ฐ ๊ธฐ๋Šฅ์˜ ๊ธฐ์ˆ  ์—…๊ทธ๋ ˆ์ด๋“œ๋กœ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์—์ฝ”์‹œ์Šคํ…œ์ด ํฌ๊ฒŒ ๊ฐ•ํ™”๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_GPT-4V, GPT-4 ํ„ฐ๋ณด ๋ฐ ๊ทธ ์ด์ƒ_. OpenAI๋Š” GPT-4 [46]์— ๋Œ€ํ•œ ์ž‘์—…์„ ๊ธฐ๋ฐ˜์œผ๋กœ 2023๋…„ 9์›”์— GPT-4์˜ ๋น„์ „ ๊ธฐ๋Šฅ์˜ ์•ˆ์ „ํ•œ ๋ฐฐํฌ์— ์ค‘์ ์„ ๋‘” GPT-4V๋ฅผ ์ถ”๊ฐ€๋กœ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-4V์˜ ์‹œ์Šคํ…œ ์นด๋“œ [133]์—์„œ๋Š” ์‹œ๊ฐ์ ์œผ๋กœ ์ฆ๊ฐ•๋œ ์ž…๋ ฅ๊ณผ ๊ด€๋ จ๋œ ์œ„ํ—˜์˜ ํ‰๊ฐ€ ๋ฐ ์™„ํ™”์— ๋Œ€ํ•ด ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ๋…ผ์˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ GPT-4V๋Š” ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๊ฐ•๋ ฅํ•œ ๋น„์ „ ์—ญ๋Ÿ‰์„ ๋ณด์—ฌ์ฃผ๋ฉฐ ๊ฐ•๋ ฅํ•œ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ํ•™์Šต ์‹œ์Šคํ…œ์œผ๋กœ์„œ ํฐ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์ตœ๊ทผ์—๋Š” 2023๋…„ 11์›”์— OpenAI๊ฐ€ DevDay์—์„œ ์ผ๋ จ์˜ ๊ธฐ์ˆ  ๊ฐœ์„ ์„ ํ†ตํ•ด _GPT-4 ํ„ฐ๋ณด_๋ผ๋Š” ์ด๋ฆ„์˜ ์—…๊ทธ๋ ˆ์ด๋“œ๋œ GPT-4 ๋ชจ๋ธ ์„ธ๋Œ€๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-4 ํ„ฐ๋ณด๋Š” ํ–ฅ์ƒ๋œ ๋ชจ๋ธ ์šฉ๋Ÿ‰(GPT-4๋ณด๋‹ค ๋” ๊ฐ•๋ ฅํ•จ), ํ™•์žฅ๋œ ์ง€์‹ ์†Œ์Šค(์ตœ๋Œ€ 2023๋…„ 4์›”๊นŒ์ง€), ๊ธด ์ปจํ…์ŠคํŠธ ์ฐฝ(์ตœ๋Œ€ 128k ํ† ํฐ), ์ตœ์ ํ™”๋œ ๋ชจ๋ธ ์„ฑ๋Šฅ(๋” ์ €๋ ดํ•œ ๊ฐ€๊ฒฉ), ๊ธฐํƒ€ ์œ ์šฉํ•œ ๊ธฐ๋Šฅ ์—…๋ฐ์ดํŠธ(ํ•จ์ˆ˜ ํ˜ธ์ถœ, ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ ์ถœ๋ ฅ ๋“ฑ)๊ฐ€ ํŠน์ง•์ž…๋‹ˆ๋‹ค. ๋™์‹œ์— ์—์ด์ „ํŠธ์™€ ๊ฐ™์€ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ์‰ฝ๊ฒŒ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๋„๋ก ์–ด์‹œ์Šคํ„ดํŠธ API๊ฐ€ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด API๋ฅผ ํ†ตํ•ด ๊ฐœ๋ฐœ์ž๋Š” ํŠน์ • ์ง€์นจ, ์ถ”๊ฐ€ ์ง€์‹ ๋ฐ ๋„๊ตฌ ์‚ฌ์šฉ์„ ํ™œ์šฉํ•˜์—ฌ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋‚ด์—์„œ ๋ชฉํ‘œ ์ง€ํ–ฅ์ ์ธ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ์‰ฝ๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋ฒˆ ๋ฆด๋ฆฌ์Šค์—์„œ๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๊ธฐ๋Šฅ(๋ณด๊ธฐ, ๋“ฃ๊ธฐ, ๋งํ•˜๊ธฐ)๋„ ๊ฐ•ํ™”๋˜์–ด ๋น„์ „, DALL-E 3, ํ…์ŠคํŠธ ์Œ์„ฑ ๋ณ€ํ™˜(TTS), ์Œ์„ฑ ์ƒ˜ํ”Œ ๋“ฃ๊ธฐ ๋“ฑ์œผ๋กœ GPT-4 ํ„ฐ๋ณด๋ฅผ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์œผ๋กœ ๋ชจ๋ธ์˜ ์šฉ๋Ÿ‰ ๋ฒ”์œ„๊ฐ€ ํฌ๊ฒŒ ํ™•์žฅ๋˜๊ณ  ์ž‘์—… ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋”์šฑ ์ค‘์š”ํ•œ ๊ฒƒ์€ ๊ฐœ์„ ๋œ ๋ชจ๋ธ, API ๋ฐ ๊ธฐ๋Šฅ์˜ ๊ธฐ์ˆ  ์—…๊ทธ๋ ˆ์ด๋“œ๋กœ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์—์ฝ”์‹œ์Šคํ…œ์ด ํฌ๊ฒŒ ๊ฐ•ํ™”๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_GPT-4V, GPT-4 turbo, and beyond_. Based on the work done for GPT-4 [46], OpenAI further released GPT-4V in September 2023, which focused on the safe deployment of the vision capabilities of GPT-4. In the GPT-4V's system card [133], it has extensively discussed the assessment and mitigation of risks related to visually augmented inputs. Specially, GPT-4V exhibited strong vision capacities in various application scenarios, showing the great potential as a powerful multimodal learning system. More recently, in November 2023, OpenAI released an upgraded generation of GPT-4 model at DevDay, named _GPT-4 Turbo_, with a series of technical improvements. GPT-4 Turbo is featured by the improved model capacity (more capable than GPT-4), the extended knowledge source (up to April 2023), long context window (up to 128k tokens), optimized model performance (cheaper price), and other useful functionality updates (function call, reproducible outputs, etc.). At the same time, Assistants API was launched to ease the rapid development of agent-like assistants. With this API, developers can easily create goal-oriented assistants within their applications, by leveraging specific instruction, extra knowledge and tool use. Furthermore, multimodal capacities (see, hear, and speak) were also enhanced in this new release, supported by GPT-4 Turbo with vision, DALL-E 3, Text-to-speech (TTS), and Listen to voice samples. These improvements have greatly extended the capacity scope and enhanced the task performance of GPT models. More importantly, the application ecosystem will be greatly strengthened with the technology upgrade in improved models, APIs, and functionalities. [|assistant|]
๊ทธ ์ด์•ผ๊ธฐ๋Š” ์ข…์ข… ์ €์˜ ์–ด๋ฆฐ ์‹œ์ ˆ์ธ Jurning์œผ๋กœ ๊ฑฐ์Šฌ๋Ÿฌ ์˜ฌ๋ผ๊ฐ€์„œ ์ €์˜ ์ž๋…€๋“ค๊ณผ ์ €์™€์˜ ๊ด€๊ณ„์— ๋Œ€ํ•œ ๋งฅ๋ฝ์„ ๋งŒ๋“ค์–ด ์ค๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ ์ด์•ผ๊ธฐ๋Š” ์ข…์ข… ์ €์˜ ์–ด๋ฆฐ ์‹œ์ ˆ์ธ ์œ ๋ฅด๋‹์œผ๋กœ ๊ฑฐ์Šฌ๋Ÿฌ ์˜ฌ๋ผ๊ฐ€์„œ ์ €์˜ ์ž๋…€๋“ค๊ณผ ์ €์™€์˜ ๊ด€๊ณ„์— ๋Œ€ํ•œ ๋งฅ๋ฝ์„ ๋งŒ๋“ค์–ด ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The story often matches back to Jurning childhood, producing context about her children and the relationship with my. [|assistant|]
* **Generalizability.** ๋…๋ฆฝ์ ์œผ๋กœ ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜๋ฉด ์•™์ƒ๋ธ” ํฌ๊ธฐ๋ฅผ ๋Š˜๋ ค์„œ ์„ฑ๋Šฅ์„ ์ผ๋ฐ˜์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * **Compatibility.** ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์•™์ƒ๋ธ” ํฌ๊ธฐ๋ฅผ ๋Š˜๋ ค์„œ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
* **์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅ์„ฑ.** ๋…๋ฆฝ์ ์œผ๋กœ ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜๋ฉด ์•™์ƒ๋ธ” ํฌ๊ธฐ๋ฅผ ๋Š˜๋ ค์„œ ์„ฑ๋Šฅ์„ ์ผ๋ฐ˜์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * **ํ˜ธํ™˜์„ฑ.** ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์•™์ƒ๋ธ” ํฌ๊ธฐ๋ฅผ ๋Š˜๋ ค์„œ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* **Generalizability.** By using our method standalone, the performance can be generally enhanced by increasing the ensemble size. * **Compatibility.** Our method can generally enhance other methods by increasing the ensemble size. [|assistant|]
ํ•™์Šต ๋ฐ ๊ฒ€์ฆ์—๋Š” SlimPajama(Soboleva et al., 2023), German CommonCrawl(Laippala et al., 2022), Pile(Gao et al., 2020)์˜ ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด Pile์—์„œ ํŠน๋ณ„ํžˆ ํ•™์Šต๋œ Black et al.(2022)๊ณผ ๋™์ผํ•œ tokenizer๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. SlimPajama์˜ ํ›ˆ๋ จ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ์„ธํŠธ(์ด ํ† ํฐ 606B)๋ฅผ ๋ฌด์ž‘์œ„๋กœ ํ•˜์œ„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ Pile๊ณผ ๋น„์Šทํ•œ ํฌ๊ธฐ์ธ \(\sim\)299B token ํ•˜์œ„ ์ง‘ํ•ฉ์„ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค(ํ‘œ 1 ์ฐธ์กฐ). ๋˜ํ•œ ์ด SlimPajama ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์ถ”๊ฐ€๋กœ ํ•˜์œ„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์˜ 3๊ฐœ์˜ \(\sim 100\)B token ๋ถ„ํ• ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค(์ž์„ธํ•œ ๋‚ด์šฉ์€ ์„น์…˜ 7.4 ์ฐธ์กฐ). SlimPajama validation ์ง‘ํ•ฉ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๊ธฐ๋ณธ validation ์ง‘ํ•ฉ์„ tokenํ™”ํ•˜๊ธฐ๋งŒ ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค(Soboleva et al., 2023). ๋…์ผ์–ด ํ›ˆ๋ จ ๋ฐ validation ์ง‘ํ•ฉ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด Oscar ๋ฐ์ดํ„ฐ ์„ธํŠธ(Laippala et al., 2022)์˜ ์ผ๋ถ€๋กœ ์ œ๊ณต๋˜๋Š” ๋…์ผ์–ด Common Crawl scrape์„ 195.43B token ํ›ˆ๋ จ ์„ธํŠธ์™€ 982.6M token validation ์„ธํŠธ๋กœ ๋ถ„ํ• ํ•˜๊ณ  tokenํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. Pile ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋ฏธ๋ฆฌ ์„ž์—ฌ ์žˆ๊ณ  ํ˜ผํ•ฉ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ ๊ธฐ๋ณธ ํ›ˆ๋ จ ๋ฐ validation ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์„ธํŠธ๋Š” ์ด \(\sim 330\)B token์ด์ง€๋งŒ, ์‹คํ—˜์—์„œ๋Š” 300B token ํ•˜์œ„ ์ง‘ํ•ฉ๋งŒ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
ํ•™์Šต ๋ฐ ๊ฒ€์ฆ์—๋Š” SlimPajama(Soboleva et al., 2023), German CommonCrawl(Laippala et al., 2022), Pile(Gao et al., 2020)์˜ ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด Pile์—์„œ ํŠน๋ณ„ํžˆ ํ•™์Šต๋œ Black et al.(2022)๊ณผ ๋™์ผํ•œ ํ† ํฐํ™”๊ธฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. SlimPajama์˜ ํ›ˆ๋ จ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ์„ธํŠธ(์ด ํ† ํฐ 606B)๋ฅผ ๋ฌด์ž‘์œ„๋กœ ํ•˜์œ„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ Pile๊ณผ ๋น„์Šทํ•œ ํฌ๊ธฐ์ธ \(sim\)299B ํ† ํฐ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค(ํ‘œ 1 ์ฐธ์กฐ). ๋˜ํ•œ ์ด SlimPajama ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์ถ”๊ฐ€๋กœ ํ•˜์œ„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์˜ 3๊ฐœ์˜ \(sim 100\)B ํ† ํฐ ๋ถ„ํ• ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค(์ž์„ธํ•œ ๋‚ด์šฉ์€ ์„น์…˜ 7.4 ์ฐธ์กฐ). SlimPajama ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ๊ธฐ๋ณธ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์„ ํ† ํฐํ™”ํ•˜๊ธฐ๋งŒ ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค(Soboleva et al., 2023). ๋…์ผ์–ด ํ›ˆ๋ จ ๋ฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด Oscar ๋ฐ์ดํ„ฐ ์„ธํŠธ(Laippala et al., 2022)์˜ ์ผ๋ถ€๋กœ ์ œ๊ณต๋˜๋Š” ๋…์ผ์–ด ์ปค๋จผ ํฌ๋กค๋ง ์Šคํฌ๋žฉ์„ 195.43B ํ† ํฐ ํ›ˆ๋ จ ์„ธํŠธ์™€ 982.6M ํ† ํฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋กœ ๋ถ„ํ• ํ•˜๊ณ  ํ† ํฐํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”๋ฏธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋ฏธ๋ฆฌ ์„ž์—ฌ ์žˆ๊ณ  ํ˜ผํ•ฉ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ ๊ธฐ๋ณธ ํ›ˆ๋ จ ๋ฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์„ธํŠธ๋Š” ์ด \(\sim 330\)B ํ† ํฐ์ด์ง€๋งŒ, ์‹คํ—˜์—์„œ๋Š” 300B ํ† ํฐ ํ•˜์œ„ ์ง‘ํ•ฉ๋งŒ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We use three datasets for training and validation: SlimPajama (Soboleva et al., 2023), German CommonCrawl (Laippala et al., 2022), and Pile (Gao et al., 2020). For all datasets, use the same tokenizer as Black et al. (2022) trained specifically on the Pile. To create our training set for SlimPajama, we randomly sub-sample the dataset (606B Total Tokens) to form a \(\sim\)299B token subset (see Table 1) that is of comparable size to Pile. We also further sub-sample this SlimPajama subset to create three \(\sim 100\)B token splits of the dataset (see Sec. 7.4 for details). To create the SlimPajama validation set we simply tokenize the default validation set that has been extensively deduplicated (Soboleva et al., 2023). To create the German training and validaiton sets, we split and tokenized the German Common Crawl scrape, available as part of the Oscar Dataset (Laippala et al., 2022), into a 195.43B token training set and a 982.6M token validation set. The pile dataset comes pre-shuffled and mixed, we simply used the default training and validation sets. The training set is \(\sim 330\)B tokens total, though in our experiments we only train on a 300B token subset. [|assistant|]
ํ”„๋กฌํ”„ํŠธ ์žฌ๊ตฌ์„ฑ ๋ฌธ์ œ๋ฅผ _language model inversion_์œผ๋กœ ๊ณต์‹ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์–ธ์–ด ๋ชจ๋ธ์˜ ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์— ๋”ฐ๋ผ ์กฐ๊ฑด์ด ์ง€์ •๋œ ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณต๊ตฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ์ปดํ“จํ„ฐ ๋น„์ „ ๋ถ„์•ผ์˜ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜๊ธฐ์˜ ํ™•๋ฅ  ์˜ˆ์ธก์€ ๋†€๋ผ์šด ์–‘์˜ ์„ธ๋ถ€ ์ •๋ณด๋ฅผ ์œ ์ง€ํ•œ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค(Dosovitskiy & Brox, 2016), ๋”ฐ๋ผ์„œ ์–ธ์–ด ๋ชจ๋ธ์—๋„ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ถ„ํฌ ๋ฒกํ„ฐ๋ฅผ ๋ฏธ๋ฆฌ ํ•™์Šต๋œ Encoder-Decoder ์–ธ์–ด ๋ชจ๋ธ์— ์˜ํ•ด ํšจ๊ณผ์ ์œผ๋กœ ์ฒ˜๋ฆฌ๋  ์ˆ˜ ์žˆ๋Š” ์‹œํ€€์Šค๋กœ "unrolling"ํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์–ธ์–ด ๋ชจ๋ธ์˜ ์˜ˆ์ธก์ด ๋Œ€๋ถ€๋ถ„ ๋ฐ˜์ „ ๊ฐ€๋Šฅํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์ฒ˜์Œ์œผ๋กœ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋งŽ์€ ๊ฒฝ์šฐ ์›๋ณธ๊ณผ ๋งค์šฐ ์œ ์‚ฌํ•œ ์ž…๋ ฅ์„ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋•Œ๋กœ๋Š” ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ์ •ํ™•ํžˆ ๋˜์ฐพ์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
ํ”„๋กฌํ”„ํŠธ ์žฌ๊ตฌ์„ฑ ๋ฌธ์ œ๋ฅผ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์— ๋”ฐ๋ผ ์กฐ๊ฑด์ด ์ง€์ •๋œ ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณต๊ตฌํ•˜๋Š” _์–ธ์–ด ๋ชจ๋ธ ๋ฐ˜์ „_์œผ๋กœ ๊ณต์‹ํ™”ํ•ฉ๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ์ปดํ“จํ„ฐ ๋น„์ „ ๋ถ„์•ผ์˜ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜๊ธฐ์˜ ํ™•๋ฅ  ์˜ˆ์ธก์€ ๋†€๋ผ์šด ์–‘์˜ ์„ธ๋ถ€ ์ •๋ณด๋ฅผ ์œ ์ง€ํ•œ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค(Dosovitskiy & Brox, 2016), ๋”ฐ๋ผ์„œ ์–ธ์–ด ๋ชจ๋ธ์—๋„ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ถ„ํฌ ๋ฒกํ„ฐ๋ฅผ ๋ฏธ๋ฆฌ ํ•™์Šต๋œ ์ธ์ฝ”๋”-๋””์ฝ”๋” ์–ธ์–ด ๋ชจ๋ธ์— ์˜ํ•ด ํšจ๊ณผ์ ์œผ๋กœ ์ฒ˜๋ฆฌ๋  ์ˆ˜ ์žˆ๋Š” ์‹œํ€€์Šค๋กœ "์–ธ๋กค๋ง"ํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์–ธ์–ด ๋ชจ๋ธ์˜ ์˜ˆ์ธก์ด ๋Œ€๋ถ€๋ถ„ ๋ฐ˜์ „ ๊ฐ€๋Šฅํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์ฒ˜์Œ์œผ๋กœ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋งŽ์€ ๊ฒฝ์šฐ ์›๋ณธ๊ณผ ๋งค์šฐ ์œ ์‚ฌํ•œ ์ž…๋ ฅ์„ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋•Œ๋กœ๋Š” ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ์ •ํ™•ํžˆ ๋˜์ฐพ์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We formalize this problem of prompt reconstruction as _language model inversion_, recovering the input prompt conditioned on the language model's next-token probabilities. Interestingly, work in computer vision has shown that probability predictions of image classifiers retain a surprising amount of detail (Dosovitskiy & Brox, 2016), so it is plausible that this also holds for language models. We propose an architecture that predicts prompts by"unrolling" the distribution vector into a sequence that can be processed effectively by a pretrained encoder-decoder language model. This method shows for the first time that language model predictions are mostly invertible: in many cases, we are able to recover very similar inputs to the original, sometimes getting the input text back exactly. [|assistant|]
Med-PaLM [76]์€ ๋„๋ฉ”์ธ๋ณ„ PaLM์œผ๋กœ, ์˜๋ฃŒ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๊ณ ํ’ˆ์งˆ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. Med-PaLM์€ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ์‹œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ ์ƒˆ๋กœ์šด ๋„๋ฉ”์ธ์— ๋งž์ถ”๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ ํšจ์œจ์ ์ธ ๋ฐฉ๋ฒ•์ธ Instruction Prompt Tuning์„ ์‚ฌ์šฉํ•˜์—ฌ PaLM์—์„œ Fine-tuning๋ฉ๋‹ˆ๋‹ค. Med-PaLM์€ ๋งŽ์€ ์˜๋ฃŒ ์ž‘์—…์—์„œ ๋งค์šฐ ๊ณ ๋ฌด์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์–ป์—ˆ์ง€๋งŒ, ์—ฌ์ „ํžˆ ์ธ๊ฐ„ ์ž„์ƒ์˜์— ๋น„ํ•ด ์—ด๋“ฑํ•ฉ๋‹ˆ๋‹ค. Med-PaLM 2๋Š” ์˜๋ฃŒ ๋„๋ฉ”์ธ Fine-tuning ๋ฐ Ensemble Prompting [77]์„ ํ†ตํ•ด Med-PaLM์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. Med-PaLM 2๋Š” MedQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ตœ๋Œ€ 86.5%์˜ ์ ์ˆ˜๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค(์ฆ‰, ์ „๋ฌธ ์˜๋ฃŒ ์‹œํ—˜, ์—ฐ๊ตฌ ๋ฐ ์†Œ๋น„์ž ์ฟผ๋ฆฌ๋ฅผ ์•„์šฐ๋ฅด๋Š” ๊ธฐ์กด ์˜คํ”ˆ ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ 6๊ฐœ๋ฅผ ๊ฒฐํ•ฉํ•œ ๋ฒค์น˜๋งˆํฌ), Med-PaLM๋ณด๋‹ค 19% ์ด์ƒ ํ–ฅ์ƒ๋˜์–ด ์ƒˆ๋กœ์šด ์ตœ์‹  ๊ธฐ๋ก์„ ์„ธ์› ์Šต๋‹ˆ๋‹ค.[|endofturn|]
Med-PaLM [76]์€ ๋„๋ฉ”์ธ๋ณ„ PaLM์œผ๋กœ, ์˜๋ฃŒ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๊ณ ํ’ˆ์งˆ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. Med-PaLM์€ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ์‹œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ ์ƒˆ๋กœ์šด ๋„๋ฉ”์ธ์— ๋งž์ถ”๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ ํšจ์œจ์ ์ธ ๋ฐฉ๋ฒ•์ธ ๋ช…๋ น ํ”„๋กฌํ”„ํŠธ ํŠœ๋‹์„ ์‚ฌ์šฉํ•˜์—ฌ PaLM์—์„œ ๋ฏธ์„ธ ์กฐ์ •๋ฉ๋‹ˆ๋‹ค. Med-PaLM์€ ๋งŽ์€ ์˜๋ฃŒ ์ž‘์—…์—์„œ ๋งค์šฐ ๊ณ ๋ฌด์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์–ป์—ˆ์ง€๋งŒ, ์—ฌ์ „ํžˆ ์ธ๊ฐ„ ์ž„์ƒ์˜์— ๋น„ํ•ด ์—ด๋“ฑํ•ฉ๋‹ˆ๋‹ค. Med-PaLM 2๋Š” ์˜๋ฃŒ ๋„๋ฉ”์ธ ๋ฏธ์„ธ ์กฐ์ • ๋ฐ ์•™์ƒ๋ธ” ํ”„๋กฌํ”„ํŠธ [77]๋ฅผ ํ†ตํ•ด Med-PaLM์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. Med-PaLM 2๋Š” MedQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ตœ๋Œ€ 86.5%์˜ ์ ์ˆ˜๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค(์ฆ‰, ์ „๋ฌธ ์˜๋ฃŒ ์‹œํ—˜, ์—ฐ๊ตฌ ๋ฐ ์†Œ๋น„์ž ์ฟผ๋ฆฌ๋ฅผ ์•„์šฐ๋ฅด๋Š” ๊ธฐ์กด ์˜คํ”ˆ ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ 6๊ฐœ๋ฅผ ๊ฒฐํ•ฉํ•œ ๋ฒค์น˜๋งˆํฌ), Med-PaLM๋ณด๋‹ค 19% ์ด์ƒ ํ–ฅ์ƒ๋˜์–ด ์ƒˆ๋กœ์šด ์ตœ์‹  ๊ธฐ๋ก์„ ์„ธ์› ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Med-PaLM [76] is a domain-specific PaLM, and is designed to provide high-quality answers to medical questions. Med-PaLM is finetuned on PaLM using instruction prompt tuning, a parameter-efficient method for aligning LLMs to new domains using a few exemplars. Med-PaLM obtains very encouraging results on many healthcare tasks, although it is still inferior to human clinicians. Med-PaLM 2 improves Med-PaLM via med-domain finetuning and ensemble prompting [77]. Med-PaLM 2 scored up to 86.5% on the MedQA dataset (i.e., a benchmark combining six existing open question answering datasets spanning professional medical exams, research, and consumer queries), improving upon Med-PaLM by over 19% and setting a new state-of-the-art. [|assistant|]
RAGAS(Es et al., 2023)๋Š” RALM์˜ faithfulness, answer relevance ๋ฐ context relevance๋ฅผ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด WikiEval ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Faithfulness๋Š” ์ œ๊ณต๋œ context์™€ ์‘๋‹ต์ด ์–ผ๋งˆ๋‚˜ ์ผ์น˜ํ•˜๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. Answer relevance๋Š” ์ƒ์„ฑ๋œ ๋‹ต๋ณ€์ด ์‹ค์ œ ์งˆ๋ฌธ์— ์–ผ๋งˆ๋‚˜ ์ž˜ ๋ถ€ํ•ฉํ•˜๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. Context relevance๋Š” ๊ฒ€์ƒ‰๋œ context๊ฐ€ ์ค‘์•™ ์ง‘์ค‘์‹์ด๊ณ  ๊ด€๋ จ ์—†๋Š” ์ •๋ณด๊ฐ€ ์—†๋Š” ์ •๋„๋กœ ์ธก์ •๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ฐ๊ตฌ์ž๋“ค์€ ํ‰๊ฐ€ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ž๋™ํ™”ํ•˜๊ธฐ ์œ„ํ•ด prompt gpt-3.5-turbo-16k ๋ชจ๋ธ์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. RGB(Chen et al., 2024)๋Š” accuracy, rejection rate, error detection rate์˜ ์„ธ ๊ฐ€์ง€ ํ‰๊ฐ€ ์ง€ํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ค‘๊ตญ์–ด์™€ ์˜์–ด์˜ ์ด์ค‘ ์–ธ์–ด ํ‰๊ฐ€ ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง€ํ‘œ๋Š” LM์— ์˜ํ•ด ์ฒ˜๋ฆฌ๋œ ๊ธฐ์‚ฌ์™€ Google์˜ API๋ฅผ ํ†ตํ•ด ๊ฒ€์ƒ‰๋œ ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ noise robustness, negative rejection, information integration ๋ฐ counterfactual robustness๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ํ™œ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. CRUD-RAG(Lyu et al., 2024)๋Š” ์ด์ „์— ์—ฐ๊ตฌ์ž๋“ค์ด ๊ณ ๋ คํ•˜์ง€ ์•Š์•˜๋˜ retrieval ๊ตฌ์„ฑ ์š”์†Œ์˜ ์˜ํ–ฅ๊ณผ ์™ธ๋ถ€ ์ง€์‹ ๊ธฐ๋ฐ˜ ๊ตฌ์ถ•์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ROUGE, BLEU, bertScore, RAGQuestEval์˜ ๋„ค ๊ฐ€์ง€ ํ‰๊ฐ€ ์ง€ํ‘œ๋ฅผ ํ†ตํ•ด RALM์˜ Create, Read, Update ๋ฐ Delete(์š”์•ฝ) ๊ธฐ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ARES(Saad-Falcon et al., 2023)๋Š” LM์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์ง€๋งŒ, lightweight LM์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ๋ณ„ RALM ๊ตฌ์„ฑ ์š”์†Œ์˜ ํ’ˆ์งˆ์„ ๊ฒฐ์ •ํ•˜๊ณ  ์˜ˆ์ธก ๊ธฐ๋ฐ˜ ์ถ”๋ก ์„ ์œ„ํ•ด ์ธ๊ฐ„์ด ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. Relevance, Answer Faithfulness, Answer Relevance๊ฐ€ ๊ด€๋ จ ๊ธฐ์ค€์ธ KILT ๋ฐ SuperGLUE ๋ฒค์น˜๋งˆํฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ RALM์˜ context๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
RAGAS(Es et al., 2023)๋Š” RALM์˜ ์ถฉ์‹ค๋„, ๋‹ต๋ณ€ ๊ด€๋ จ์„ฑ ๋ฐ ๋ฌธ๋งฅ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด WikiEval ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ถฉ์‹ค๋„๋Š” ์ œ๊ณต๋œ ์ปจํ…์ŠคํŠธ์™€ ์‘๋‹ต์ด ์–ผ๋งˆ๋‚˜ ์ผ์น˜ํ•˜๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋‹ต๋ณ€ ๊ด€๋ จ์„ฑ์€ ์ƒ์„ฑ๋œ ๋‹ต๋ณ€์ด ์‹ค์ œ ์งˆ๋ฌธ์— ์–ผ๋งˆ๋‚˜ ์ž˜ ๋ถ€ํ•ฉํ•˜๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋ฌธ๋งฅ ๊ด€๋ จ์„ฑ์€ ๊ฒ€์ƒ‰๋œ ๋ฌธ๋งฅ์ด ์ค‘์•™ ์ง‘์ค‘์‹์ด๊ณ  ๊ด€๋ จ ์—†๋Š” ์ •๋ณด๊ฐ€ ์—†๋Š” ์ •๋„๋กœ ์ธก์ •๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ฐ๊ตฌ์ž๋“ค์€ ํ‰๊ฐ€ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ž๋™ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋กฌํ”„ํŠธ gpt-3.5-turbo-16k ๋ชจ๋ธ์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. RGB(Chen et al., 2024)๋Š” ์ •ํ™•๋„, ๊ฑฐ๋ถ€์œจ, ์˜ค๋ฅ˜ ๊ฐ์ง€์œจ์˜ ์„ธ ๊ฐ€์ง€ ํ‰๊ฐ€ ์ง€ํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ค‘๊ตญ์–ด์™€ ์˜์–ด์˜ ์ด์ค‘ ์–ธ์–ด ํ‰๊ฐ€ ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง€ํ‘œ๋Š” LM์— ์˜ํ•ด ์ฒ˜๋ฆฌ๋œ ๊ธฐ์‚ฌ์™€ Google์˜ API๋ฅผ ํ†ตํ•ด ๊ฒ€์ƒ‰๋œ ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ๋…ธ์ด์ฆˆ ๊ฒฌ๊ณ ์„ฑ, ๋ถ€์ • ๊ฑฐ๋ถ€, ์ •๋ณด ํ†ตํ•ฉ ๋ฐ ๋ฐ˜์‚ฌ์‹ค์  ๊ฒฌ๊ณ ์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ํ™œ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. CRUD-RAG(Lyu et al., 2024)๋Š” ์ด์ „์— ์—ฐ๊ตฌ์ž๋“ค์ด ๊ณ ๋ คํ•˜์ง€ ์•Š์•˜๋˜ ๊ฒ€์ƒ‰ ๊ตฌ์„ฑ ์š”์†Œ์˜ ์˜ํ–ฅ๊ณผ ์™ธ๋ถ€ ์ง€์‹ ๊ธฐ๋ฐ˜ ๊ตฌ์ถ•์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ROUGE, BLEU, bertScore, RAGQuestEval์˜ ๋„ค ๊ฐ€์ง€ ํ‰๊ฐ€ ์ง€ํ‘œ๋ฅผ ํ†ตํ•ด RALM์˜ ์ƒ์„ฑ, ์ฝ๊ธฐ, ์—…๋ฐ์ดํŠธ ๋ฐ ์‚ญ์ œ(์š”์•ฝ) ๊ธฐ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ARES(Saad-Falcon et al., 2023)๋Š” LM์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์ง€๋งŒ, ๊ฒฝ๋Ÿ‰ LM์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ๋ณ„ RALM ๊ตฌ์„ฑ ์š”์†Œ์˜ ํ’ˆ์งˆ์„ ๊ฒฐ์ •ํ•˜๊ณ  ์˜ˆ์ธก ๊ธฐ๋ฐ˜ ์ถ”๋ก ์„ ์œ„ํ•ด ์ธ๊ฐ„์ด ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ด€๋ จ์„ฑ, ๋‹ต๋ณ€ ์‹ ๋ขฐ์„ฑ, ๋‹ต๋ณ€ ๊ด€๋ จ์„ฑ์ด ๊ด€๋ จ ๊ธฐ์ค€์ธ KILT ๋ฐ SuperGLUE ๋ฒค์น˜๋งˆํฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ RALM์˜ ์ปจํ…์ŠคํŠธ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]RAGAS (Es et al., 2023) employs the WikiEval Dataset to assess the faithfulness, answer relevance, and context relevance of RALMs. Faithfulness is defined as the degree to which responses align with the provided context. Answer relevance refers to the extent to which generated responses address the actual question posed. Context relevance is gauged by the degree to which retrieved context is centralized and devoid of irrelevant information. Additionally, the researchers utilize the prompt gpt-3.5-turbo-16k model to automate the evaluation process. RGB (Chen et al., 2024) developed a bilingual Chinese and English evaluation system employing three evaluation metrics: accuracy, rejection rate, and error detection rate. These metrics were utilized to assess the noise robustness, negative rejection, information integration, and counterfactual robustness of the data sources, which were articles processed by LM and retrieved through Google's API. CRUD-RAG (Lyu et al., 2024) considers the impact of retrieval components and the construction of external knowledge bases that have not been previously considered by researchers. A dataset was generated using a large model to evaluate the Create, Read, Update, and Delete (summarization) capabilities of RALM through four evaluation metrics: ROUGE, BLEU, bertScore, and RAGQuestEval. In addition, ARES (Saad-Falcon et al., 2023) employs datasets generated by the LM, but utilizes a lightweight LM to determine the quality of individual RALM components and utilizes human-labeled data points for prediction-powered inference. The RALM's context is evaluated using the KILT and SuperGLUE benchmarks, with Relevance, Answer Faithfulness, and Answer Relevance being the relevant criteria. [|assistant|]
๋‹ค๋ฅธ ์„ค์ •์—์„œ ํ•™์Šต ๋ฐ ํ‰๊ฐ€๋œ ๋ชจ๋ธ ๊ฐ„์˜ ๋น„๊ต๋Š” ๋งŽ์€ ์™ธ๋ถ€ ์š”์ธ์ด ๊ฒฐ๊ณผ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํ’€๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค(์˜ˆ: ํ•™์Šต ๋Œ€ ์ถ”๋ก ์˜ ์ˆ˜์น˜์  ์ •๋ฐ€๋„, ์‚ฌ์šฉ๋œ prompts). ์šฐ๋ฆฌ๋Š” ์„ธ ๊ฐ€์ง€ ์ˆ˜์ค€์˜ ๋น„๊ต๋ฅผ ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค: (1) ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๋งŒ ๋‹ค๋ฅธ ๋‚ด๋ถ€ ๋น„๊ต, (2) ๋‹ค๋ฅธ ์ฝ”๋“œ๋ฒ ์ด์Šค๋กœ ํ•™์Šต๋˜์—ˆ์ง€๋งŒ Scao et al. (2022); Black et al. (2022); Aleph Alpha (2023); Dey et al. (2023)์˜ ๊ฒฐ๊ณผ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Eleuther AI harness๋กœ ํ‰๊ฐ€๋œ ๋ชจ๋ธ์„ ๋Œ€์ƒ์œผ๋กœ ํ•œ benchmark ์ˆ˜์ค€ ๋น„๊ต(\(*\) ํ‘œ์‹œ); (3) Brown et al. (2020); Chowdhery et al. (2022)๊ณผ์˜ ์™ธ๋ถ€ ๋น„๊ต(\(\dagger\) ํ‘œ์‹œ). ์ž์„ธํ•œ ํ‰๊ฐ€ ๋‚ด์šฉ์€ ๋ถ€๋ก F.1์„ ์ฐธ์กฐํ•˜์„ธ์š”.[|endofturn|]
๋‹ค๋ฅธ ์„ค์ •์—์„œ ํ•™์Šต ๋ฐ ํ‰๊ฐ€๋œ ๋ชจ๋ธ ๊ฐ„์˜ ๋น„๊ต๋Š” ๋งŽ์€ ์™ธ๋ถ€ ์š”์ธ์ด 1 987 ๊ฒฐ๊ณผ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํ’€๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค(์˜ˆ: ํ•™์Šต ๋Œ€ ์ถ”๋ก ์˜ ์ˆ˜์น˜์  ์ •๋ฐ€๋„, ์‚ฌ์šฉ๋œ ํ”„๋กฌํ”„ํŠธ). ์šฐ๋ฆฌ๋Š” ์„ธ ๊ฐ€์ง€ ์ˆ˜์ค€์˜ ๋น„๊ต๋ฅผ ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค: (1) ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๋งŒ ๋‹ค๋ฅธ ๋‚ด๋ถ€ ๋น„๊ต, (2) ๋‹ค๋ฅธ ์ฝ”๋“œ๋ฒ ์ด์Šค๋กœ ํ•™์Šต๋˜์—ˆ์ง€๋งŒ ์Šค์นด์šฐ ๋“ฑ์˜ ์—ฐ๊ตฌ(2022); ๋ธ”๋ž™ ๋“ฑ(2022); ์•Œ๋ ˆํ”„ ์•ŒํŒŒ(2023); ๋ฐ์ด ๋“ฑ(2023)์˜ ๊ฒฐ๊ณผ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์—˜๋ฃจ๋” AI ํ•˜๋„ค์Šค๋กœ ํ‰๊ฐ€๋œ ๋ชจ๋ธ์„ ๋Œ€์ƒ์œผ๋กœ ํ•œ ๋ฒค์น˜๋งˆํฌ ์ˆ˜์ค€ ๋น„๊ต(\(*\) ํ‘œ์‹œ); (3) ๋ธŒ๋ผ์šด ๋“ฑ(2020); Chowdhery ๋“ฑ(2022)๊ณผ์˜ ์™ธ๋ถ€ ๋น„๊ต(\(\dagger\) ํ‘œ์‹œ). ์ž์„ธํ•œ ํ‰๊ฐ€ ๋‚ด์šฉ์€ ๋ถ€๋ก F.1์„ ์ฐธ์กฐํ•˜์„ธ์š”.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Comparisons across models trained and evaluated in different settings are difficult to untangle, as many externalities may influence the 1 987results (e.g., numerical precision of training vs inference, prompts used). We distinguish three levels of comparisons: (1) internal comparisons, with models trained and evaluated within our codebase, for which only the pretraining datasets differ; (2) benchmark-level comparisons, with models trained with a different codebase but evaluated with the Eleuther AI harness, taking results from Scao et al. (2022); Black et al. (2022); Aleph Alpha (2023); Dey et al. (2023), thereafter flagged with a \(*\); (3) external comparisons with Brown et al. (2020); Chowdhery et al. (2022), thereafter flagged with a \(\dagger\). For further details on evaluation, see Appendix F.1. [|assistant|]
๊ทธ๋ฆผ 13: ๋„ค ๊ฐ€์ง€ ํŒŒ๋ผ๋ฏธํ„ฐ ํšจ์œจ์  Fine-tuning ๋ฐฉ๋ฒ•์˜ ์˜ˆ์‹œ์ž…๋‹ˆ๋‹ค. MHA์™€ FFN์€ ๊ฐ๊ฐ Transformer ๋ ˆ์ด์–ด์˜ multi-head attention๊ณผ feed-forward networks๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 13: ๋„ค ๊ฐ€์ง€ ํŒŒ๋ผ๋ฏธํ„ฐ ํšจ์œจ์  ๋ฏธ์„ธ ์กฐ์ • ๋ฐฉ๋ฒ•์˜ ์˜ˆ์‹œ์ž…๋‹ˆ๋‹ค. MHA์™€ FFN์€ ๊ฐ๊ฐ ํŠธ๋žœ์Šคํฌ๋จธ ๋ ˆ์ด์–ด์˜ ๋‹ค์ค‘ ํ—ค๋“œ ์ฃผ์˜์™€ ํ”ผ๋“œ ํฌ์›Œ๋“œ ๋„คํŠธ์›Œํฌ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Fig. 13: An illustration of four different parameter-efficient fine-tuning methods. MHA and FFN denote the multi-head attention and feed-forward networks in the Transformer layer, respectively. [|assistant|]
์ด๋ฅผ ์œ„ํ•ด ๋ณธ ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ๋Š” ์ƒˆ๋กœ ์ถ”๊ฐ€๋œ ํ† ํฐ์˜ ์ž„๋ฒ ๋”ฉ์„ ๋” ์ž˜ ํ•™์Šต์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ํšจ์œจ์ ์ด๊ณ  ํšจ๊ณผ์ ์ธ ์–ดํœ˜ ํ™•์žฅ์„ ์œ„ํ•œ ์ƒˆ๋กœ์šด ์ ‘๊ทผ ๋ฐฉ์‹, ์ฆ‰ EEVE๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์‰ฝ๊ฒŒ ์ ์‘ํ•  ์ˆ˜ ์žˆ๋„๋ก subword-based embedding ์ดˆ๊ธฐํ™”๋ฅผ ํ™œ์šฉํ•˜๊ณ  parameter freezing์„ ํ†ตํ•ด 7๋‹จ๊ณ„์˜ ํ•™์Šต ๋‹จ๊ณ„๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ ํ•™์Šตํ•  ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ์ˆœ์„œ์™€ ์–‘์„ ์ •๊ตํ•˜๊ฒŒ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” input embeddings๋งŒ ํ•™์Šตํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์ตœ์ข… ๋‹จ๊ณ„์—์„œ ์ „์ฒด ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํฌ๊ด„ํ•˜๋„๋ก ์ ์ง„์ ์œผ๋กœ ํ™•์žฅํ•˜์—ฌ foundational ๋ชจ๋ธ์˜ ๊ณ ๊ธ‰ ๊ธฐ๋Šฅ์„ ์˜์–ด์—์„œ ํ•œ๊ตญ์–ด๋กœ ์„ธ์‹ฌํ•˜๊ฒŒ ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ด๋ฅผ ์œ„ํ•ด ๋ณธ ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ๋Š” ์ƒˆ๋กœ ์ถ”๊ฐ€๋œ ํ† ํฐ์˜ ์ž„๋ฒ ๋”ฉ์„ ๋” ์ž˜ ํ•™์Šต์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ํšจ์œจ์ ์ด๊ณ  ํšจ๊ณผ์ ์ธ ์–ดํœ˜ ํ™•์žฅ์„ ์œ„ํ•œ ์ƒˆ๋กœ์šด ์ ‘๊ทผ ๋ฐฉ์‹, ์ฆ‰ EEVE๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์‰ฝ๊ฒŒ ์ ์‘ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์œ„ ๋‹จ์–ด ๊ธฐ๋ฐ˜ ์ž„๋ฒ ๋”ฉ ์ดˆ๊ธฐํ™”๋ฅผ ํ™œ์šฉํ•˜๊ณ  ํŒŒ๋ผ๋ฏธํ„ฐ ๋™๊ฒฐ์„ ํ†ตํ•ด 7๋‹จ๊ณ„์˜ ํ•™์Šต ๋‹จ๊ณ„๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ ํ•™์Šตํ•  ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ์ˆœ์„œ์™€ ์–‘์„ ์ •๊ตํ•˜๊ฒŒ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” ์ž…๋ ฅ ์ž„๋ฒ ๋”ฉ๋งŒ ํ•™์Šตํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์ตœ์ข… ๋‹จ๊ณ„์—์„œ ์ „์ฒด ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํฌ๊ด„ํ•˜๋„๋ก ์ ์ง„์ ์œผ๋กœ ํ™•์žฅํ•˜์—ฌ ๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ๊ณ ๊ธ‰ ๊ธฐ๋Šฅ์„ ์˜์–ด์—์„œ ํ•œ๊ตญ์–ด๋กœ ์„ธ์‹ฌํ•˜๊ฒŒ ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To this end, this technical report presents a novel approach for efficient and effective vocabulary expansion, namely EEVE, which can better train the embeddings of newly added tokens. For ease of adaptation, we utilize subword-based embedding initialization and design seven training stages with parameter freezing, which elaborately adjust the order and amount of parameters to be trained. We meticulously transfer the advanced capabilities of foundational models from English to Korean by initially focusing on the training of only input embeddings and progressively expanding to encompass the full parameters in the final stage. [|assistant|]
\(\bullet\)_๋ชจ๋ธ ์นœํ™”์ ์ธ ํ˜•์‹ ํ™œ์šฉ._ LLMs์€ ํŠน๋ณ„ํžˆ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์ณค๊ธฐ ๋•Œ๋ฌธ์— LLMs์ด Instruction์„ ๋” ์ž˜ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ๋ช‡ ๊ฐ€์ง€ ํ”„๋กฌํ”„ํŠธ ํ˜•์‹์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, OpenAI ๋ฌธ์„œ์—์„œ ์„ค๋ช…ํ•˜๋Š” ๊ฒƒ์ฒ˜๋Ÿผ *** ๋˜๋Š” ***๋ฅผ ์ค‘์ง€ ๊ธฐํ˜ธ๋กœ ์‚ฌ์šฉํ•˜์—ฌ Instruction๊ณผ ์ปจํ…์ŠคํŠธ๋ฅผ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” LLMs์ด ๋” ์ž˜ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๊ฐ€์ด๋“œ๋ผ์ธ์œผ๋กœ, ๋Œ€๋ถ€๋ถ„์˜ ๊ธฐ์กด LLMs์€ ์˜์–ด๋กœ ๋œ ์ง€์‹œ๋ฅผ ๋” ์ž˜ ์ˆ˜ํ–‰ํ•˜๋ฏ€๋กœ ๊ธฐ๊ณ„ ๋ฒˆ์—ญ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์–ด๋ ค์šด ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์˜์–ด ์ง€์‹œ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_๋ชจ๋ธ ์นœํ™”์ ์ธ ํ˜•์‹ ํ™œ์šฉ._ LLM์€ ํŠน๋ณ„ํžˆ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์ณค๊ธฐ ๋•Œ๋ฌธ์— LLM์ด ์ง€์‹œ๋ฅผ ๋” ์ž˜ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ๋ช‡ ๊ฐ€์ง€ ํ”„๋กฌํ”„ํŠธ ํ˜•์‹์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, OpenAI ๋ฌธ์„œ์—์„œ ์„ค๋ช…ํ•˜๋Š” ๊ฒƒ์ฒ˜๋Ÿผ *** ๋˜๋Š” ***๋ฅผ ์ค‘์ง€ ๊ธฐํ˜ธ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ์ง€์นจ๊ณผ ์ปจํ…์ŠคํŠธ๋ฅผ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” LLM์ด ๋” ์ž˜ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๊ฐ€์ด๋“œ๋ผ์ธ์œผ๋กœ, ๋Œ€๋ถ€๋ถ„์˜ ๊ธฐ์กด LLM์€ ์˜์–ด๋กœ ๋œ ์ง€์‹œ๋ฅผ ๋” ์ž˜ ์ˆ˜ํ–‰ํ•˜๋ฏ€๋กœ ๊ธฐ๊ณ„ ๋ฒˆ์—ญ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์–ด๋ ค์šด ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์˜์–ด ์ง€์‹œ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_Utilizing model-friendly format._ Since LLMs are pre-trained on specially constructed datasets, there are some prompt formats that can make LLMs better understand the instruction. For example, as the OpenAI documentation suggests, we can use *** or *** as a stop symbol to separate the instruction and context, which can be better understood by LLMs. As a general guideline, most existing LLMs perform a task better in English, thus it is useful to employ English instructions to solve difficult tasks based on machine translation. [|assistant|]
๋”ฐ๋ผ์„œ \(T\gtrsim k^{\beta}\)๊ฐ€ ๋˜๋Š” ์ฆ‰์‹œ AI๊ฐ€ ์ƒ์„ฑํ•œ ์ƒ˜ํ”Œ ํฌ๊ธฐ \(T\)๋Š” "scalable" ๋ฆฌ์†Œ์Šค๊ฐ€ ๋˜์ง€ ์•Š์œผ๋ฉฐ, ๋” ๋งŽ์€ AI๊ฐ€ ์ƒ์„ฑํ•œ ์ƒ˜ํ”Œ์„ ์ˆ˜์ง‘ํ•ด๋„ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜์ง€ ์•Š์œผ๋ฉฐ, ์ฆ‰ ์„ฑ๋Šฅ์ด ์ •์ฒด๋˜๊ณ  scaling์„ ์žƒ๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 3์˜ ์™ผ์ชฝ๊ณผ ๊ทธ๋ฆผ 8(๋ถ€๋ก B)์—์„œ ๊ฒฝํ—˜์ ์œผ๋กœ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋”ฐ๋ผ์„œ \(T\gtrsim k^{\๋ฒ ํƒ€}\)๊ฐ€ ๋˜๋Š” ์ฆ‰์‹œ AI๊ฐ€ ์ƒ์„ฑํ•œ ์ƒ˜ํ”Œ ํฌ๊ธฐ \(T\)๋Š” "ํ™•์žฅ ๊ฐ€๋Šฅํ•œ" ๋ฆฌ์†Œ์Šค๊ฐ€ ๋˜์ง€ ์•Š์œผ๋ฉฐ, ๋” ๋งŽ์€ AI๊ฐ€ ์ƒ์„ฑํ•œ ์ƒ˜ํ”Œ์„ ์ˆ˜์ง‘ํ•ด๋„ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜์ง€ ์•Š์œผ๋ฉฐ, ์ฆ‰ ์„ฑ๋Šฅ์ด ์ •์ฒด๋˜๊ณ  ํ™•์žฅ์„ฑ์„ ์žƒ๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 3์˜ ์™ผ์ชฝ๊ณผ ๊ทธ๋ฆผ 8(๋ถ€๋ก B)์—์„œ ๊ฒฝํ—˜์ ์œผ๋กœ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Thus, as soon as \(T\gtrsim k^{\beta}\), the AI-generated sample size \(T\) ceases to be a "scalable" resource: collecting more AI-generated samples will not improve the performance of the downstream model, i.e performance plateaus and we lose scaling. The result is illustrated empirically in Figure 3, left and Figure 8 (Appendix B). [|assistant|]
DKS-RAC(Huang et al., 2023)์€ ์‹œํ€€์Šค ์ˆ˜์ค€์—์„œ ๋‹ต๋ณ€๊ณผ ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ ๊ฐ„์˜ ์ง€์‹์„ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•ด Dense Knowledge Similarity (DKS)์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ง€์‹ ์œ ์‚ฌ์„ฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์ ˆ ์„ ํƒ์— ์ง์ ‘์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์ณ ์งˆ์˜์™€ ๋ฌธ์„œ ๊ฐ„์˜ ์ผ์น˜๋„๋ฅผ ๊ฐœ์„ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— re-ranking์— ์†ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
DKS-RAC(Huang et al., 2023)์€ ์‹œํ€€์Šค ์ˆ˜์ค€์—์„œ ๋‹ต๋ณ€๊ณผ ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ ๊ฐ„์˜ ์ง€์‹์„ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ€๋„ ์ง€์‹ ์œ ์‚ฌ์„ฑ(DKS)์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ง€์‹ ์œ ์‚ฌ์„ฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์ ˆ ์„ ํƒ์— ์ง์ ‘์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์ณ ์งˆ์˜์™€ ๋ฌธ์„œ ๊ฐ„์˜ ์ผ์น˜๋„๋ฅผ ๊ฐœ์„ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์žฌ์ˆœ์œ„ํ™”(re-ranking)์— ์†ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]DKS-RAC (Huang et al., 2023) presents the Dense Knowledge Similarity (DKS) for aligning the knowledge between answers and retrieved passages at the sequence level. This approach is categorized under re-ranking due to its direct impact on passage selection based on knowledge similarity, refining the match between queries and documents. [|assistant|]
MLCommons๋Š” v0.5 benchmark๋ฅผ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด benchmark ์‹คํ–‰์ž(benchmark๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š”)์ธ ModelBench benchmark runner์™€ ์‹ค์ œ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์ด ํฌํ•จ๋œ ํ…Œ์ŠคํŠธ ์‹คํ–‰ ์—”์ง„์ธ ModelGauge๋กœ ๊ตฌ์„ฑ๋œ ์˜คํ”ˆ ์†Œ์Šค ํ‰๊ฐ€ ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋ฒ„์ „์ด ์ง€์ •๋œ ํ…Œ์ŠคํŠธ์™€ SUT๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‘œ์ค€ํ™”๋˜๊ณ  ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ benchmark ์‹คํ–‰์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋„๊ตฌ๋Š” ๋ชจ๋“ˆ์‹ ํ”Œ๋Ÿฌ๊ทธ์ธ ์•„ํ‚คํ…์ฒ˜๋กœ ์„ค๊ณ„๋˜์–ด ๋ชจ๋ธ ์ œ๊ณต์—…์ฒด๊ฐ€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ƒˆ๋กœ์šด SUT๋ฅผ ํ”Œ๋žซํผ์— ์‰ฝ๊ฒŒ ๊ตฌํ˜„ํ•˜๊ณ  ์ถ”๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. AI Safety Benchmark๊ฐ€ ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ ํ”Œ๋žซํผ์— ์ƒˆ๋กœ์šด ๋ฒ„์ „์˜ ํ…Œ์ŠคํŠธ๊ฐ€ ์ถ”๊ฐ€๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ”Œ๋žซํผ์— ์•ก์„ธ์Šคํ•˜๊ณ  ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ GitHub์˜ ModelBench Git ๋ฆฌํฌ์ง€ํ† ๋ฆฌ์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ModelBench์™€ ModelGauge๋Š” Stanford Center for Research on Foundation Models (CRFM)์˜ Holistic Evaluation of Language Models [12] ํŒ€๊ณผ ๊ณต๋™์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์œผ๋ฉฐ, ๋„๋ฆฌ ์ฑ„ํƒ๋œ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋งŒ๋“ค์–ด์˜จ HELM ํŒ€์˜ ๋ฆฌ๋น™ ๋ฆฌ๋”๋ณด๋“œ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์„ ๋ฐ”ํƒ•์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
MLCommons๋Š” v0.5 ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด ๋ฒค์น˜๋งˆํฌ ์‹คํ–‰์ž(๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š”)์ธ ModelBench ๋ฒค์น˜๋งˆํฌ ๋Ÿฐ๋„ˆ์™€ ์‹ค์ œ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์ด ํฌํ•จ๋œ ํ…Œ์ŠคํŠธ ์‹คํ–‰ ์—”์ง„์ธ ModelGauge๋กœ ๊ตฌ์„ฑ๋œ ์˜คํ”ˆ ์†Œ์Šค ํ‰๊ฐ€ ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋ฒ„์ „์ด ์ง€์ •๋œ ํ…Œ์ŠคํŠธ์™€ SUT๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‘œ์ค€ํ™”๋˜๊ณ  ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ ๋ฒค์น˜๋งˆํฌ ์‹คํ–‰์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋„๊ตฌ๋Š” ๋ชจ๋“ˆ์‹ ํ”Œ๋Ÿฌ๊ทธ์ธ ์•„ํ‚คํ…์ฒ˜๋กœ ์„ค๊ณ„๋˜์–ด ๋ชจ๋ธ ์ œ๊ณต์—…์ฒด๊ฐ€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ƒˆ๋กœ์šด SUT๋ฅผ ํ”Œ๋žซํผ์— ์‰ฝ๊ฒŒ ๊ตฌํ˜„ํ•˜๊ณ  ์ถ”๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. AI ์•ˆ์ „ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ ํ”Œ๋žซํผ์— ์ƒˆ๋กœ์šด ๋ฒ„์ „์˜ ํ…Œ์ŠคํŠธ๊ฐ€ ์ถ”๊ฐ€๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ”Œ๋žซํผ์— ์•ก์„ธ์Šคํ•˜๊ณ  ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ GitHub์˜ ModelBench Git ๋ฆฌํฌ์ง€ํ† ๋ฆฌ์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.6 ModelBench์™€ ModelGauge๋Š” ์Šคํƒ ํฌ๋“œ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ ์—ฐ๊ตฌ ์„ผํ„ฐ(CRFM)์˜ HOLISTIC ์–ธ์–ด ๋ชจ๋ธ ํ‰๊ฐ€[12] ํŒ€๊ณผ ๊ณต๋™์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์œผ๋ฉฐ, ๋„๋ฆฌ ์ฑ„ํƒ๋œ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋งŒ๋“ค์–ด์˜จ HELM ํŒ€์˜ ๋ฆฌ๋น™ ๋ฆฌ๋”๋ณด๋“œ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์„ ๋ฐ”ํƒ•์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To support the v0.5 benchmark, MLCommons has developed an open-source evaluation tool, which consists of the ModelBench benchmark runner (which can be used to implement the benchmark) and the ModelGauge test execution engine (which contains the actual test items). This tool enables standardized, reproducible benchmark runs using versioned tests and SUTs. The tool is designed with a modular plug-in architecture, allowing model providers to easily implement and add new SUTs to the platform for evaluation. As the AI Safety Benchmark evolves, new versions of tests will be added to the platform. Details on how to access and use the platform can be found in the ModelBench Git repository on GitHub.6 ModelBench and ModelGauge were developed in collaboration with the Holistic Evaluation of Language Models [12] team at the Stanford Center for Research on Foundation Models (CRFM), and build upon the HELM team's experience of creating a widely-adopted open-source model evaluation framework for living leaderboards. [|assistant|]
* ๊ทธ๋ฆผ 7: ์‚ฌ์šฉ ์‚ฌ๋ก€ ๋ฐ ์ „๋‹ฌ ๊ฐ€๋Šฅํ•œ ๊ธฐ์ˆ ์„ ํฌํ•จํ•œ Instruction์„ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋กœ ์ธ์ฝ”๋”ฉํ•˜๊ธฐ. * ๊ทธ๋ฆผ 8: Instruction ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์กฐ๊ฐ€ ๋น„๊ต์  ๋‹จ์ˆœํ•œ ๊ธฐ๋ณธ Instruction์œผ๋กœ ๋””์ฝ”๋”ฉํ•˜๊ธฐ. * ๊ทธ๋ฆผ 9: Instruction์ด ์–ผ๋งˆ๋‚˜ ์–ด๋ ค์šด์ง€ ํŒ๋‹จํ•˜๊ธฐ ์œ„ํ•œ ๋ฃจ๋ธŒ๋ฆญ ์ƒ์„ฑ ๋ฐ ์ฃผ์–ด์ง„ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ Instruction์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ์กฐ์น˜. * ๊ทธ๋ฆผ 10: ์ƒ์„ฑ๋œ ์ž‘์—… ์ค‘ ํ•˜๋‚˜๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ž…๋ ฅ Instruction์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. * ๊ทธ๋ฆผ 11: ๋Œ€์ƒ LLM๊ณผ ๊ฐ•๋ ฅํ•œ LLM์˜ ์‘๋‹ต ํ’ˆ์งˆ ๋น„๊ต. ์„ค๋ช… ๋ถ€๋ถ„์„ ์ œ๊ฑฐํ•˜์—ฌ Vicuna ์Šคํƒ€์ผ์˜ ์Œ๋ณ„ ๋น„๊ต ํ”„๋กฌํ”„ํŠธ์—์„œ ์ˆ˜์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. * ๊ทธ๋ฆผ 12: LLM(_์˜ˆ: ChatGPT, GPT-4)์„ ์‹ฌ์‚ฌ์œ„์›์œผ๋กœ ์‚ฌ์šฉํ•œ ์ž๋™ ํ‰๊ฐ€. Chiang et al.์˜ ํ…œํ”Œ๋ฆฟ์„ ๋”ฐ๋ฆ„(2023); Chen et al.[|endofturn|]
* ๊ทธ๋ฆผ 7: ์‚ฌ์šฉ ์‚ฌ๋ก€ ๋ฐ ์ „๋‹ฌ ๊ฐ€๋Šฅํ•œ ๊ธฐ์ˆ ์„ ํฌํ•จํ•œ ์ธ์ŠคํŠธ๋Ÿญ์…˜์„ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋กœ ์ธ์ฝ”๋”ฉํ•˜๊ธฐ. * ๊ทธ๋ฆผ 8: ์ธ์ŠคํŠธ๋Ÿญ์…˜ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์กฐ๊ฐ€ ๋น„๊ต์  ๋‹จ์ˆœํ•œ ๊ธฐ๋ณธ ์ธ์ŠคํŠธ๋Ÿญ์…˜์œผ๋กœ ๋””์ฝ”๋”ฉํ•˜๊ธฐ. * ๊ทธ๋ฆผ 9: ์ง€์นจ์ด ์–ผ๋งˆ๋‚˜ ์–ด๋ ค์šด์ง€ ํŒ๋‹จํ•˜๊ธฐ ์œ„ํ•œ ๋ฃจ๋ธŒ๋ฆญ ์ƒ์„ฑ ๋ฐ ์ฃผ์–ด์ง„ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ง€์นจ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ์กฐ์น˜. * ๊ทธ๋ฆผ 10: ์ƒ์„ฑ๋œ ์ž‘์—… ์ค‘ ํ•˜๋‚˜๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ž…๋ ฅ ์ง€์นจ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. * ๊ทธ๋ฆผ 11: ๋Œ€์ƒ LLM๊ณผ ๊ฐ•๋ ฅํ•œ LLM์˜ ์‘๋‹ต ํ’ˆ์งˆ ๋น„๊ต. ์„ค๋ช… ๋ถ€๋ถ„์„ ์ œ๊ฑฐํ•˜์—ฌ Vicuna ์Šคํƒ€์ผ์˜ ์Œ๋ณ„ ๋น„๊ต ํ”„๋กฌํ”„ํŠธ์—์„œ ์ˆ˜์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. * ๊ทธ๋ฆผ 12: LLM(_์˜ˆ: ChatGPT, GPT-4)์„ ์‹ฌ์‚ฌ์œ„์›์œผ๋กœ ์‚ฌ์šฉํ•œ ์ž๋™ ํ‰๊ฐ€. Chiang et al.์˜ ํ…œํ”Œ๋ฆฟ์„ ๋”ฐ๋ฆ„(2023); Chen et al.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* Figure 7: Encoding instructions into metadata, including use case and transferable skills. * Figure 8: Decoding instruction metadata into basic instructions that are relatively simple in structure. * Figure 9: Generating rubrics to judge how challenging an instruction is, and actions to improve the instruction based on the given metadata. * Figure 10: Improving the input instruction by following one of the generated actions. * Figure 11: Comparing the responses quality from the target and strong LLMs. Adapted from the Vicuna-style pairwise comparison prompt by removing the explanation part. * Figure 12: Automatic evaluation using LLM (_e.g._, ChatGPT, GPT-4) as the judge. Following the templates in Chiang et al. (2023); Chen et al. [|assistant|]
์–ธ์–ด์˜ ๋งฅ๋ฝ์—์„œ ๋ฒ—์–ด๋‚˜ ์„ ํ˜ธ๋„์—์„œ ์ •์ฑ…์„ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์€ bandit ๋ฐ Reinforcement Learning ํ™˜๊ฒฝ์—์„œ ์—ฐ๊ตฌ๋˜์–ด ์™”์œผ๋ฉฐ, ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ณด์ƒ์ด ์•„๋‹Œ ํ–‰๋™์˜ ์„ ํ˜ธ๋„ ๋˜๋Š” ์ˆœ์œ„๋ฅผ ์‚ฌ์šฉํ•˜๋Š” contextual bandit ํ•™์Šต์„ contextual dueling bandit (CDB; [48; 14])๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ ˆ๋Œ€์ ์ธ ๋ณด์ƒ์ด ์—†๋Š” ๊ฒฝ์šฐ, CDB์— ๋Œ€ํ•œ ์ด๋ก ์  ๋ถ„์„์€ ์ตœ์ ์˜ ์ •์ฑ… ๊ฐœ๋…์„ _von Neumann winner_๋กœ ๋Œ€์ฒดํ•˜๋ฉฐ, ์ด๋Š” _๋‹ค๋ฅธ ์ •์ฑ…์— ๋Œ€ํ•œ ๊ธฐ๋Œ€ ์Šน๋ฅ ์ด ์ตœ์†Œ 50%์ธ ์ •์ฑ…์ž…๋‹ˆ๋‹ค[14]. ๊ทธ๋Ÿฌ๋‚˜ CDB ์„ค์ •์—์„œ๋Š” ์„ ํ˜ธ๋„ ๋ ˆ์ด๋ธ”์ด ์˜จ๋ผ์ธ์œผ๋กœ ์ œ๊ณต๋˜๋Š” ๋ฐ˜๋ฉด, ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„๋ฅผ ํ•™์Šตํ•˜๋Š” ๊ฒฝ์šฐ ์ผ๋ฐ˜์ ์œผ๋กœ ์˜คํ”„๋ผ์ธ ์„ ํ˜ธ๋„๊ฐ€ ์ฃผ์–ด์ง„ ๊ณ ์ •๋œ ์ผ๊ด„ ์ž‘์—…์—์„œ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค[47]. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, _preference-based RL_ (PbRL)์€ ๋ณด์ƒ์ด ์•„๋‹Œ _์•Œ ์ˆ˜ ์—†๋Š”_'์ ์ˆ˜' ํ•จ์ˆ˜์— ์˜ํ•ด ์ƒ์„ฑ๋œ ์ด์ง„ ์„ ํ˜ธ๋„๋ฅผ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค[9; 35]. off-policy ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ์žฌ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ํฌํ•จํ•˜์—ฌ ๋‹ค์–‘ํ•œ PbRL ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ์กด์žฌํ•˜์ง€๋งŒ, ์ผ๋ฐ˜์ ์œผ๋กœ ์ž ์žฌ ์ ์ˆ˜ ํ•จ์ˆ˜(์ฆ‰, ๋ณด์ƒ ๋ชจ๋ธ)๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ์ถ”์ •ํ•œ ๋‹ค์Œ ์ด๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๊ณผ์ •์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค[16; 9; 12; 34; 19]. ๋Œ€์‹  ์„ ํ˜ธ๋„๋ฅผ ๋งŒ์กฑํ•˜๋„๋ก ์ •์ฑ…์„ ์ง์ ‘ ์ตœ์ ํ™”ํ•˜๋Š” ๋‹จ์ผ ๋‹จ๊ณ„ ์ •์ฑ… ํ•™์Šต ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์–ธ์–ด์˜ ๋งฅ๋ฝ์—์„œ ๋ฒ—์–ด๋‚˜ ์„ ํ˜ธ๋„์—์„œ ์ •์ฑ…์„ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์€ ๋ฐด๋””ํŠธ ๋ฐ ๊ฐ•ํ™” ํ•™์Šต ํ™˜๊ฒฝ์—์„œ ์—ฐ๊ตฌ๋˜์–ด ์™”์œผ๋ฉฐ, ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ณด์ƒ์ด ์•„๋‹Œ ํ–‰๋™์˜ ์„ ํ˜ธ๋„ ๋˜๋Š” ์ˆœ์œ„๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์ปจํ…์ŠคํŠธ ๋ฐด๋””ํŠธ ํ•™์Šต์„ ์ปจํ…์ŠคํŠธ ๋“€์–ผ๋ง ๋ฐด๋””ํŠธ(CDB; [48; 14])๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ ˆ๋Œ€์ ์ธ ๋ณด์ƒ์ด ์—†๋Š” ๊ฒฝ์šฐ, CDB์— ๋Œ€ํ•œ ์ด๋ก ์  ๋ถ„์„์€ ์ตœ์ ์˜ ์ •์ฑ… ๊ฐœ๋…์„ _ํฐ ๋…ธ์ด๋งŒ ์Šน์ž_๋กœ ๋Œ€์ฒดํ•˜๋ฉฐ, ์ด๋Š” _๋‹ค๋ฅธ ์ •์ฑ…์— ๋Œ€ํ•œ ๊ธฐ๋Œ€ ์Šน๋ฅ ์ด ์ตœ์†Œ 50%์ธ ์ •์ฑ…์ž…๋‹ˆ๋‹ค[14]. ๊ทธ๋Ÿฌ๋‚˜ CDB ์„ค์ •์—์„œ๋Š” ์„ ํ˜ธ๋„ ๋ ˆ์ด๋ธ”์ด ์˜จ๋ผ์ธ์œผ๋กœ ์ œ๊ณต๋˜๋Š” ๋ฐ˜๋ฉด, ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„๋ฅผ ํ•™์Šตํ•˜๋Š” ๊ฒฝ์šฐ ์ผ๋ฐ˜์ ์œผ๋กœ ์˜คํ”„๋ผ์ธ ์„ ํ˜ธ๋„๊ฐ€ ์ฃผ์–ด์ง„ ๊ณ ์ •๋œ ์ผ๊ด„ ์ž‘์—…์—์„œ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค[47]. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, _์„ ํ˜ธ๋„ ๊ธฐ๋ฐ˜ RL_(PbRL)์€ ๋ณด์ƒ์ด ์•„๋‹Œ _์•Œ ์ˆ˜ ์—†๋Š”_'์ ์ˆ˜' ํ•จ์ˆ˜์— ์˜ํ•ด ์ƒ์„ฑ๋œ ์ด์ง„ ์„ ํ˜ธ๋„๋ฅผ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค[9; 35]. ์˜คํ”„ํด๋ฆฌ์‹œ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ์žฌ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ํฌํ•จํ•˜์—ฌ ๋‹ค์–‘ํ•œ PbRL ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ์กด์žฌํ•˜์ง€๋งŒ, ์ผ๋ฐ˜์ ์œผ๋กœ ์ž ์žฌ ์ ์ˆ˜ ํ•จ์ˆ˜(์ฆ‰, ๋ณด์ƒ ๋ชจ๋ธ)๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ์ถ”์ •ํ•œ ๋‹ค์Œ ์ด๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๊ณผ์ •์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค[16; 9; 12; 34; 19]. ๋Œ€์‹  ์„ ํ˜ธ๋„๋ฅผ ๋งŒ์กฑํ•˜๋„๋ก ์ •์ฑ…์„ ์ง์ ‘ ์ตœ์ ํ™”ํ•˜๋Š” ๋‹จ์ผ ๋‹จ๊ณ„ ์ •์ฑ… ํ•™์Šต ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Outside of the context of language, learning policies from preferences has been studied in both bandit and reinforcement learning settings, and several approaches have been proposed. Contextual bandit learning using preferences or rankings of actions, rather than rewards, is known as a contextual dueling bandit (CDB; [48; 14]). In the absence of absolute rewards, theoretical analysis of CDBs substitutes the notion of an optimal policy with a _von Neumann winner_, a policy whose expected win rate against _any_ other policy is at least 50% [14]. However, in the CDB setting, preference labels are given online, while in learning from human preferences, we typically learn from a fixed batch of offline preference-annotated action pairs [47]. Similarly, _preference-based RL_ (PbRL) learns from binary preferences generated by an _unknown_'scoring' function rather than rewards [9; 35]. Various algorithms for PbRL exist, including methods that can reuse off-policy preference data, but generally involve first explicitly estimating the latent scoring function (i.e. the reward model) and subsequently optimizing it [16; 9; 12; 34; 19]. We instead present a single stage policy learning approach that directly optimizes a policy to satisfy preferences. [|assistant|]
**Vertical Architectures**. ์ด ๊ตฌ์กฐ์—์„œ๋Š” ํ•œ agent๊ฐ€ ๋ฆฌ๋” ์—ญํ• ์„ ํ•˜๋ฉฐ ๋‹ค๋ฅธ agent๊ฐ€ ์ด ๋ฆฌ๋”์—๊ฒŒ ์ง์ ‘ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. architecture์— ๋”ฐ๋ผ ๋ณด๊ณ  agent๋Š” ๋ฆฌ๋“œ agent์™€๋งŒ ํ†ต์‹ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜๋Š” ๋ชจ๋“  agent ๊ฐ„์— ๊ณต์œ  ๋Œ€ํ™”๋กœ ๋ฆฌ๋”๋ฅผ ์ •์˜ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ˆ˜์ง architecture์˜ ์ฃผ์š” ํŠน์ง•์€ ๋ฆฌ๋“œ agent๊ฐ€ ์žˆ๊ณ  ํ˜‘์—… agent ๊ฐ„์— ๋ถ„๋ช…ํ•œ ๋ถ„์—…์ด ์ด๋ฃจ์–ด์ง„๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
**์ˆ˜์ง ์•„ํ‚คํ…์ฒ˜**. ์ด ๊ตฌ์กฐ์—์„œ๋Š” ํ•œ ์—์ด์ „ํŠธ๊ฐ€ ๋ฆฌ๋” ์—ญํ• ์„ ํ•˜๋ฉฐ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๊ฐ€ ์ด ๋ฆฌ๋”์—๊ฒŒ ์ง์ ‘ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์•„ํ‚คํ…์ฒ˜์— ๋”ฐ๋ผ ๋ณด๊ณ  ์—์ด์ „ํŠธ๋Š” ๋ฆฌ๋“œ ์—์ด์ „ํŠธ์™€๋งŒ ํ†ต์‹ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜๋Š” ๋ชจ๋“  ์—์ด์ „ํŠธ ๊ฐ„์— ๊ณต์œ  ๋Œ€ํ™”๋กœ ๋ฆฌ๋”๋ฅผ ์ •์˜ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ˆ˜์ง ์•„ํ‚คํ…์ฒ˜์˜ ์ฃผ์š” ํŠน์ง•์€ ๋ฆฌ๋“œ ์—์ด์ „ํŠธ๊ฐ€ ์žˆ๊ณ  ํ˜‘์—… ์—์ด์ „ํŠธ ๊ฐ„์— ๋ถ„๋ช…ํ•œ ๋ถ„์—…์ด ์ด๋ฃจ์–ด์ง„๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Vertical Architectures**. In this structure, one agent acts as a leader and has other agents report directly to them. Depending on the architecture, reporting agents may communicate exclusively with the lead agent. Alternatively, a leader may be defined with a shared conversation between all agents. The defining features of vertical architectures include having a lead agent and a clear division of labor between the collaborating agents. [|assistant|]
์ž ์žฌ์ ์ธ ๋งค์น˜๋กœ ์‹๋ณ„๋œ ๊ฐ๊ฐ์˜ ๋ฌธ์„œ ์Œ์— ๋Œ€ํ•ด, ๋” ๋งŽ์€ ๊ณ„์‚ฐ ๋น„์šฉ์ด ๋งŽ์ด ๋“œ๋Š” similarity metric์ด ํ›„์† ํ•„ํ„ฐ๋ง ๋‹จ๊ณ„๋กœ์„œ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค.[|endofturn|]
์ž ์žฌ์ ์ธ ๋งค์น˜๋กœ ์‹๋ณ„๋œ ๊ฐ๊ฐ์˜ ๋ฌธ์„œ ์Œ์— ๋Œ€ํ•ด, ๋” ๋งŽ์€ ๊ณ„์‚ฐ ๋น„์šฉ์ด ๋งŽ์ด ๋“œ๋Š” ์œ ์‚ฌ์„ฑ ๋ฉ”ํŠธ๋ฆญ์ด ํ›„์† ํ•„ํ„ฐ๋ง ๋‹จ๊ณ„๋กœ์„œ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]For each pair of documents identified as a potential match, more computationally expensive similarity metrics can be employed as a subsequent filtering step. [|assistant|]
**์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ ๊ณ„ํš(Chen et al., 2023; Gandhi et al., 2023; Qiao et al., 2024)์—์„œ ์ถ”๋ก ํ•˜๋„๋ก ์–ธ์–ด ๋ชจ๋ธ ์—์ด์ „ํŠธ๋ฅผ ๊ฐ€๋ฅด์น˜๊ฑฐ๋‚˜ ํŠน์ • ๋„๊ตฌ ๋˜๋Š” ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜๋„๋ก(Yao et al., 2022; Lanchantin et al., 2024; Schick et al., 2024) ์—ฐ๊ตฌํ•˜๋Š” ํ•œํŽธ, ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ์ถ”๋ก ์„ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ๋กœ ์ฆ๋ฅ˜ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Ho et al., 2022; Li et al., 2022; Hsieh et al., 2023). ๊ทธ๋Ÿฌ๋‚˜ ํŠนํžˆ Pan et al. (2024)์€ ์ด๋Ÿฌํ•œ ํ”ผ๋“œ๋ฐฑ ๋ฃจํ”„๊ฐ€ ๋ณด์ƒ ํ•ดํ‚น์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Zelikman et al. (2023)์€ ๋ชจ๋ธ์ด ๋™์ผํ•œ ์ฝ”๋“œ ๊ฐœ์„ ์ž๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฝ”๋“œ ๊ฐœ์„ ์ž๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฐœ์„ ํ•˜๋Š” bootstrapping loop๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ฃผ๊ณ , Haluptzok et al. (2023)์€ ์–ธ์–ด ๋ชจ๋ธ์ด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ํผ์ฆ์„ ํ†ตํ•ด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋Šฅ๋ ฅ์„ bootstrappingํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(Schuster et al., 2021). ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” inductive reasoning ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ฑฐ๋‚˜ ์‹ค์ œ ์‹œ์Šคํ…œ์„ ๋ชจ๋ธ๋งํ•˜๊ธฐ ์œ„ํ•ด ์–ธ์–ด ๋ชจ๋ธ์— ์œ ์‚ฌํ•œ ์ „๋žต์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค(Wang et al., 2023; Qiu et al., 2023; Zhu et al., 2023; Li et al., 2024).[|endofturn|]
**์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ ๊ณ„ํš(Chen et al., 2023; Gandhi et al., 2023; Qiao et al., 2024)์—์„œ ์ถ”๋ก ํ•˜๋„๋ก ์–ธ์–ด ๋ชจ๋ธ ์—์ด์ „ํŠธ๋ฅผ ๊ฐ€๋ฅด์น˜๊ฑฐ๋‚˜ ํŠน์ • ๋„๊ตฌ ๋˜๋Š” ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜๋„๋ก(Yao et al., 2022; Lanchantin et al., 2024; Schick et al., 2024) ์—ฐ๊ตฌํ•˜๋Š” ํ•œํŽธ, ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ์ถ”๋ก ์„ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ๋กœ ์ฆ๋ฅ˜ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Ho et al., 2022); Li et al. (2022); Hsieh et al. (2023). ๊ทธ๋Ÿฌ๋‚˜ ํŠนํžˆ Pan et al. (2024)์€ ์ด๋Ÿฌํ•œ ํ”ผ๋“œ๋ฐฑ ๋ฃจํ”„๊ฐ€ ๋ณด์ƒ ํ•ดํ‚น์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Zelikman et al. (2023)์€ ๋ชจ๋ธ์ด ๋™์ผํ•œ ์ฝ”๋“œ ๊ฐœ์„ ์ž๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฝ”๋“œ ๊ฐœ์„ ์ž๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฐœ์„ ํ•˜๋Š” ๋ถ€ํŠธ์ŠคํŠธ๋žฉ ๋ฃจํ”„๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ฃผ๊ณ , Haluptzok et al. (2023)์€ ์–ธ์–ด ๋ชจ๋ธ์ด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ํผ์ฆ์„ ํ†ตํ•ด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋Šฅ๋ ฅ์„ ๋ถ€ํŠธ์ŠคํŠธ๋žฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(Schuster et al., 2021). ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ์œ ๋„ ์ถ”๋ก  ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ฑฐ๋‚˜ ์‹ค์ œ ์‹œ์Šคํ…œ์„ ๋ชจ๋ธ๋งํ•˜๊ธฐ ์œ„ํ•ด ์–ธ์–ด ๋ชจ๋ธ์— ์œ ์‚ฌํ•œ ์ „๋žต์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค(Wang et al., 2023; Qiu et al., 2023; Zhu et al., 2023; Li et al., 2024).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Several works investigate teaching language model agents to reason in planning (Chen et al., 2023; Gandhi et al., 2023; Qiao et al., 2024), or to use specific tools or memory (Yao et al., 2022; Lanchantin et al., 2024; Schick et al., 2024), while others investigate how one may distill the reasoning from a large language model into a smaller language model Ho et al. (2022); Li et al. (2022); Hsieh et al. (2023). Notably however, Pan et al. (2024) demonstratesthat these feedback loops may result in reward hacking. Zelikman et al. (2023) shows how a bootstrapping loop can be implemented where a model repeatedly improves a code-improver using the same code-improver and Haluptzok et al. (2023) shows how language models can bootstrap their programming ability with programming puzzles Schuster et al. (2021). Other works have employed a similar strategy for using language models to solve inductive reasoning tasks or to model real-world systems (Wang et al., 2023; Qiu et al., 2023; Zhu et al., 2023; Li et al., 2024). [|assistant|]
์ตœ์ ํ™” ๋ฐ ํ‰๊ฐ€. ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ์›Œ๋ฐ์—… 20๋‹จ๊ณ„, Learning Rate \(1e-6\), ๊ฐ€์ค‘์น˜ ๊ฐ์‡  0.001, Batch Size 8(ํ•„์š”์— ๋”ฐ๋ผ ์‹คํ–‰ ์ „๋ฐ˜์— ๊ฑธ์ณ ์ด๋ฅผ ๊ณ ์ •ํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ผ๋””์–ธํŠธ ๋ˆ„์ )์˜ AdamW Optimizer๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, <!startofthought|> ๋ฐ <!endofthought|> ์ž„๋ฒ ๋”ฉ ๊ทธ๋ผ๋””์–ธํŠธ ๊ฐ€์ค‘์น˜๋Š” \(1e2\)์ด๊ณ  ์ •์ฑ… ๊ฐ€์ค‘์น˜๋Š” \(1e6\)์ž…๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์ค‘์—๋Š” ์˜จ๋„ \(T=1\)๋กœ ์ƒ˜ํ”Œ๋งํ•˜๊ณ  ํ‰๊ฐ€ ์ค‘์—๋Š” ํƒ์š•์Šค๋Ÿฌ์šด ๋””์ฝ”๋”ฉ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜จ๋„ \(T=3\)์—์„œ REINFORCE Loss๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ ์ƒ˜ํ”Œ์„ ์ค‘์š” ์ƒ˜ํ”Œ๋กœ ์ทจ๊ธ‰ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์— ์˜ˆ์ œ๋ฅผ ํ”„๋กฌํ”„ํŠธํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ ์ •๋‹ต์˜ ํ™•๋ฅ ์„ ์ง์ ‘ ๊ณ„์‚ฐํ•˜์—ฌ ์˜ˆ๋ฅผ ๋“ค์–ด, \(A\cdots E\) ์‚ฌ์ด์˜ ๊ฐ๊ด€์‹ ๋ฌธ์ œ์˜ ๊ฒฝ์šฐ, \(A\cdots E\)์— ํ•ด๋‹นํ•˜๋Š” ํ† ํฐ์— ๋Œ€ํ•œ ๋กœ์ง“์— ๋Œ€ํ•œ ์ •ํ™•๋„๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ํ›ˆ๋ จ์˜ ๊ฒฝ์šฐ ๊ฐ ์ƒ˜ํ”Œ์—์„œ 256๊ฐœ์˜ ํ† ํฐ์˜ ์ž„์˜ ๋ฒ”์œ„๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค(๋˜๋Š” ํ† ํฐ์ด 256๊ฐœ ๋ฏธ๋งŒ์ธ ๊ฒฝ์šฐ ํŒจ๋“œ).[|endofturn|]
์ตœ์ ํ™” ๋ฐ ํ‰๊ฐ€์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ์›Œ๋ฐ์—… 20๋‹จ๊ณ„, ํ•™์Šต ์†๋„ \(1e-6\), ๊ฐ€์ค‘์น˜ ๊ฐ์‡  0.001, ๋ฐฐ์น˜ ํฌ๊ธฐ 8(ํ•„์š”์— ๋”ฐ๋ผ ์‹คํ–‰ ์ „๋ฐ˜์— ๊ฑธ์ณ ์ด๋ฅผ ๊ณ ์ •ํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ผ๋””์–ธํŠธ ๋ˆ„์ )์˜ AdamW ์˜ตํ‹ฐ๋งˆ์ด์ €๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, <!startofthought|> ๋ฐ <!endofthought|> ์ž„๋ฒ ๋”ฉ ๊ทธ๋ผ๋””์–ธํŠธ ๊ฐ€์ค‘์น˜๋Š” \(1e2\)์ด๊ณ  ์ •์ฑ… ๊ฐ€์ค‘์น˜๋Š” \(1e6\)์ž…๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์ค‘์—๋Š” ์˜จ๋„ \(T=1\)๋กœ ์ƒ˜ํ”Œ๋งํ•˜๊ณ  ํ‰๊ฐ€ ์ค‘์—๋Š” ํƒ์š•์Šค๋Ÿฌ์šด ๋””์ฝ”๋”ฉ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜จ๋„ \(T=3\)์—์„œ REINFORCE ์†์‹ค์„ ๊ณ„์‚ฐํ•˜์—ฌ ์ƒ˜ํ”Œ์„ ์ค‘์š” ์ƒ˜ํ”Œ๋กœ ์ทจ๊ธ‰ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์— ์˜ˆ์ œ๋ฅผ ํ”„๋กฌํ”„ํŠธํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ ์ •๋‹ต์˜ ํ™•๋ฅ ์„ ์ง์ ‘ ๊ณ„์‚ฐํ•˜์—ฌ ์˜ˆ๋ฅผ ๋“ค์–ด, \(A\cdots E\) ์‚ฌ์ด์˜ ๊ฐ๊ด€์‹ ๋ฌธ์ œ์˜ ๊ฒฝ์šฐ, \(A\cdots E\)์— ํ•ด๋‹นํ•˜๋Š” ํ† ํฐ์— ๋Œ€ํ•œ ๋กœ์ง“์— ๋Œ€ํ•œ ์ •ํ™•๋„๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ํ›ˆ๋ จ์˜ ๊ฒฝ์šฐ ๊ฐ ์ƒ˜ํ”Œ์—์„œ 256๊ฐœ์˜ ํ† ํฐ์˜ ์ž„์˜ ๋ฒ”์œ„๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค(๋˜๋Š” ํ† ํฐ์ด 256๊ฐœ ๋ฏธ๋งŒ์ธ ๊ฒฝ์šฐ ํŒจ๋“œ).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Optimization and EvaluationFor optimization, we use the AdamW optimizer with a warmup of 20 steps, a learning rate of \(1e-6\), a weight decay of 0.001, and a batch size of 8 (along with any necessary gradient accumulation to keep this fixed across runs). Moreover, our <!startofthought|> and <!endofthought|> embedding gradient weight is \(1e2\) and our policy weight is \(1e6\). We sample with temperature \(T=1\) during training and use greedy decoding for the thoughts during evaluation. We treat our samples as importance samples by computing the REINFORCE loss at temperature \(T=3\). Because we do not prompt the model with any examples, we directly compute the probability of the correct answer, conditioned on generating an answer - for example, for multiple choice questions between \(A\cdots E\), we compute the accuracy over the logits for tokens corresponding to \(A\cdots E\). Lastly, for our training, we select a random span of 256 tokens from each sample (or pad if there are fewer than 256 tokens). [|assistant|]
ChatGPT(OpenAI, 2022)๋Š” InstructGPT(Ouyang et al., 2022)์—์„œ ๋ฐœ์ „ํ•œ ๊ฒƒ์œผ๋กœ, ์ƒํ™ฉ์— ๋งž๋Š” ์ธ๊ฐ„๊ณผ ์œ ์‚ฌํ•œ ๋Œ€ํ™”๋ฅผ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ณ ๊ธ‰ ๋Œ€ํ™”ํ˜• AI ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์˜ ์„ฑ๊ณต์€ ๋”์šฑ ์ •๊ตํ•œ LLM์ธ GPT-4(OpenAI, 2023)์˜ ๊ฐœ๋ฐœ์„ ์œ„ํ•œ ๋ฐœํŒ์ด ๋˜์—ˆ์œผ๋ฉฐ, ์ž์—ฐ์–ด ์ดํ•ด, ์ƒ์„ฑ ๋ฐ ๋‹ค์–‘ํ•œ NLP ์ž‘์—…, ํŠนํžˆ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ฐ ์ถ”๋ก  ๋Šฅ๋ ฅ์—์„œ ๋” ํฐ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ƒˆ๋กœ์šด ์—ฐ๊ตฌ ๋ฐฉํ–ฅ๊ณผ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์„ ์ด‰์ง„ํ•˜์—ฌ Artificial General Intelligence (AGI)์˜ ์ž ์žฌ๋ ฅ์„ ํƒ๊ตฌํ•˜๋Š” ๋ฐ ๋Œ€ํ•œ ๊ด€์‹ฌ์„ ๊ฐ•ํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ธ์ƒ์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค€ ์ด ๋ชจ๋ธ์€ few-shot learning๊ณผ ์ƒˆ๋กœ์šด ์ž‘์—…์— ๋Œ€ํ•œ ์ ์‘๋ ฅ์„ ํ†ตํ•ด NLP ์—ฐ๊ตฌ์˜ ํ™•์žฅ์„ ํฌ๊ฒŒ ์ฃผ๋„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ์—ฐ๊ตฌ์ž์™€ ์—…๊ณ„ ์ „๋ฌธ๊ฐ€ ๋ชจ๋‘ ๊ฐ์ • ๋ถ„์„, ๊ธฐ๊ณ„ ๋ฒˆ์—ญ, question-answering ์‹œ์Šคํ…œ ๋“ฑ ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋ถ„์•ผ์—์„œ ์ด๋“ค์˜ ์ž ์žฌ๋ ฅ์„ ํ™œ์šฉํ•˜๋ ค๋Š” ๋…ธ๋ ฅ์„ ๊ณ„์†ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
ChatGPT(OpenAI, 2022)๋Š” InstructGPT(Ouyang et al., 2022)์—์„œ ๋ฐœ์ „ํ•œ ๊ฒƒ์œผ๋กœ, ์ƒํ™ฉ์— ๋งž๋Š” ์ธ๊ฐ„๊ณผ ์œ ์‚ฌํ•œ ๋Œ€ํ™”๋ฅผ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ณ ๊ธ‰ ๋Œ€ํ™”ํ˜• AI ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์˜ ์„ฑ๊ณต์€ ๋”์šฑ ์ •๊ตํ•œ LLM์ธ GPT-4(OpenAI, 2023)์˜ ๊ฐœ๋ฐœ์„ ์œ„ํ•œ ๋ฐœํŒ์ด ๋˜์—ˆ์œผ๋ฉฐ, ์ž์—ฐ์–ด ์ดํ•ด, ์ƒ์„ฑ ๋ฐ ๋‹ค์–‘ํ•œ NLP ์ž‘์—…, ํŠนํžˆ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ฐ ์ถ”๋ก  ๋Šฅ๋ ฅ์—์„œ ๋” ํฐ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ƒˆ๋กœ์šด ์—ฐ๊ตฌ ๋ฐฉํ–ฅ๊ณผ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์„ ์ด‰์ง„ํ•˜์—ฌ ์ธ๊ณต ์ผ๋ฐ˜ ์ง€๋Šฅ(AGI)์˜ ์ž ์žฌ๋ ฅ์„ ํƒ๊ตฌํ•˜๋Š” ๋ฐ ๋Œ€ํ•œ ๊ด€์‹ฌ์„ ๊ฐ•ํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ธ์ƒ์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค€ ์ด ๋ชจ๋ธ์€ ์†Œ์ˆ˜ ์ƒท ํ•™์Šต๊ณผ ์ƒˆ๋กœ์šด ์ž‘์—…์— ๋Œ€ํ•œ ์ ์‘๋ ฅ์„ ํ†ตํ•ด NLP ์—ฐ๊ตฌ์˜ ํ™•์žฅ์„ ํฌ๊ฒŒ ์ฃผ๋„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ์—ฐ๊ตฌ์ž์™€ ์—…๊ณ„ ์ „๋ฌธ๊ฐ€ ๋ชจ๋‘ ๊ฐ์ • ๋ถ„์„, ๊ธฐ๊ณ„ ๋ฒˆ์—ญ, ์งˆ์˜์‘๋‹ต ์‹œ์Šคํ…œ ๋“ฑ ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋ถ„์•ผ์—์„œ ์ด๋“ค์˜ ์ž ์žฌ๋ ฅ์„ ํ™œ์šฉํ•˜๋ ค๋Š” ๋…ธ๋ ฅ์„ ๊ณ„์†ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]ChatGPT (OpenAI, 2022), evolved from InstructGPT (Ouyang et al., 2022), serves as an advanced conversational AI model capable of conducting context-aware, human-like interactions. Its success set the stage for the development of GPT-4 (OpenAI, 2023), a more sophisticated LLM, demonstrating even greater potential in natural language understanding, generation, and various NLP tasks, especially for its multi-modal and reasoning abilities. These models have catalyzed new research directions and applications, intensifying interest in exploring the potential of Artificial General Intelligence (AGI). Exhibiting impressive performance across multiple benchmarks, they have also demonstrated capabilities for few-shot learning and adaptability to new tasks, significantly driving the expansion of NLP research. Consequently, they have inspired both researchers and industry professionals to further harness their potential across a wide array of applications, including sentiment analysis, machine translation, question-answering systems, and more. [|assistant|]
์‹คํ—˜. ์•ž์„œ ๊ณ ๋ คํ•œ 6๊ฐ€์ง€ ๋ชจ๋ธ์— ๋Œ€ํ•ด, ์šฐ๋ฆฌ๋Š” ์ถ”์ถœํ•œ ํ–‰๋ ฌ \(\tilde{\mathbf{W}}=\mathbf{U}\cdot\mathbf{\Sigma}\)์™€ ์‹ค์ œ ๊ฐ€์ค‘์น˜ ํ–‰๋ ฌ์„ ๋น„๊ตํ•˜์—ฌ ๊ณต๊ฒฉ ์„ฑ๊ณต๋ฅ ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ตœ์†Œ ์ œ๊ณฑ ์‹œ์Šคํ…œ \(\tilde{\mathbf{W}}\cdot\mathbf{G}\approx\mathbf{W}\)๋ฅผ \(\mathbf{G}\)์— ๋Œ€ํ•ด ํ’€๊ณ , ๊ฐ๊ฐ \(l\) ๋ฐฉ์ •์‹๊ณผ \(h\) ๋ฏธ์ง€์ˆ˜๋ฅผ ๊ฐ€์ง„ \(h\times h\) affine transformation์„ ํ—ˆ์šฉํ•œ ํ›„ \(\mathbf{G}\)์— ๋Œ€ํ•ด ์ตœ์†Œ ์ œ๊ณฑ ์‹œ์Šคํ…œ์„ ํ’‰๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ \(\mathbf{W}\)์™€ \(\tilde{\mathbf{W}}\cdot\mathbf{G}\)์˜ RMS๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์‹คํ—˜.์•ž์„œ ๊ณ ๋ คํ•œ 6๊ฐ€์ง€ ๋ชจ๋ธ์— ๋Œ€ํ•ด, ์šฐ๋ฆฌ๋Š” ์ถ”์ถœํ•œ ํ–‰๋ ฌ \(\tilde{\mathbf{W}}=\mathbf{U}\cdot\mathbf{\Sigma}\)์™€ ์‹ค์ œ ๊ฐ€์ค‘์น˜ ํ–‰๋ ฌ์„ ๋น„๊ตํ•˜์—ฌ ๊ณต๊ฒฉ ์„ฑ๊ณต๋ฅ ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ตœ์†Œ ์ œ๊ณฑ ์‹œ์Šคํ…œ \(\tilde{\mathbf{W}}\cdot\mathbf{G}\approx\mathbf{W}\)๋ฅผ \(\mathbf{G}\)์— ๋Œ€ํ•ด ํ’€๊ณ , ๊ฐ๊ฐ \(l\) ๋ฐฉ์ •์‹๊ณผ \(h\) ๋ฏธ์ง€์ˆ˜๋ฅผ ๊ฐ€์ง„ \(h\times h\) ์•„ํ•€ ๋ณ€ํ™˜์„ ํ—ˆ์šฉํ•œ ํ›„ \(\mathbf{G}\)์— ๋Œ€ํ•ด ์ตœ์†Œ ์ œ๊ณฑ ์‹œ์Šคํ…œ์„ ํ’‰๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ \(\mathbf{W}\)์™€ \(\tilde{\mathbf{W}}\cdot\mathbf{G}\)์˜ RMS๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Experiments.For the six models considered previously, we evaluate the attack success rate by comparing the root mean square (RMS) between our extracted matrix \(\tilde{\mathbf{W}}=\mathbf{U}\cdot\mathbf{\Sigma}\) and the actual weight matrix, after allowing for a \(h\times h\) affine transformation. Concretely, we solve the least squares system \(\tilde{\mathbf{W}}\cdot\mathbf{G}\approx\mathbf{W}\) for \(\mathbf{G}\), which reduces to \(h\) linear least squares problems, each with \(l\) equations and \(h\) unknowns. Then, we report the RMS of \(\mathbf{W}\) and \(\tilde{\mathbf{W}}\cdot\mathbf{G}\). [|assistant|]
GPT-4 ์˜ˆ์ธก์˜ ์ˆ˜๋™ ๋ถ„์„. ์ด ๋…ผ๋ฌธ์˜ ์ €์ž๋“ค์€ ๊ฐ ์ธก๋ฉด์— ๋Œ€ํ•ด ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ 20๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋ฅผ ์ˆ˜๋™์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ณ  ๋™์ผํ•œ instruction, ๋ฐ๋ชจ ๋ฐ ํ…Œ์ŠคํŠธ ์ธ์Šคํ„ด์Šค๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ GPT-4 ์˜ˆ์ธก์ด ์ž์‹ ์˜ ํ‰๊ฐ€์™€ ์ผ์น˜ํ•˜๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€ ๊ฒฐ๊ณผ, ํŠนํžˆ ๊ด€๋ จ์„ฑ(95%), retrieval necessity(95%), ์ง€์› ์ •๋„(90%)์—์„œ GPT-4 ์˜ˆ์ธก๊ณผ ๋†’์€ ์ผ์น˜๋„๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์œ ์šฉ์„ฑ(80%)์—์„œ๋Š” ์ผ์น˜๋„๊ฐ€ ์•ฝ๊ฐ„ ๋‚ฎ์•˜๋Š”๋ฐ, ์ด๋Š” ๋Œ€๋ถ€๋ถ„ 1๊ณผ 2 ๋˜๋Š” 4์™€ 5 ์‚ฌ์ด์˜ ์˜๊ฒฌ ๋ถˆ์ผ์น˜ ๋•Œ๋ฌธ์ด์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
GPT-4 ์˜ˆ์ธก์˜ ์ˆ˜๋™ ๋ถ„์„.์ด ๋…ผ๋ฌธ์˜ ์ €์ž๋“ค์€ ๊ฐ ์ธก๋ฉด์— ๋Œ€ํ•ด ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ 20๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋ฅผ ์ˆ˜๋™์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ณ  ๋™์ผํ•œ ์ง€์นจ, ๋ฐ๋ชจ ๋ฐ ํ…Œ์ŠคํŠธ ์ธ์Šคํ„ด์Šค๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ GPT-4 ์˜ˆ์ธก์ด ์ž์‹ ์˜ ํ‰๊ฐ€์™€ ์ผ์น˜ํ•˜๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€ ๊ฒฐ๊ณผ, ํŠนํžˆ ๊ด€๋ จ์„ฑ(95%), ๊ฒ€์ƒ‰ ํ•„์š”์„ฑ(95%), ์ง€์› ์ •๋„(90%)์—์„œ GPT-4 ์˜ˆ์ธก๊ณผ ๋†’์€ ์ผ์น˜๋„๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์œ ์šฉ์„ฑ(80%)์—์„œ๋Š” ์ผ์น˜๋„๊ฐ€ ์•ฝ๊ฐ„ ๋‚ฎ์•˜๋Š”๋ฐ, ์ด๋Š” ๋Œ€๋ถ€๋ถ„ 1๊ณผ 2 ๋˜๋Š” 4์™€ 5 ์‚ฌ์ด์˜ ์˜๊ฒฌ ๋ถˆ์ผ์น˜ ๋•Œ๋ฌธ์ด์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Manual analysis of the GPT-4 predictions.The authors of this paper manually assess randomly sampled 20 instances for each aspect and check if GPT-4 predictions match their assessments given the same instruction, demonstrations, and test instances. We found our assessments show high agreement with GPT-4 predictions, especially for relevance (95%), retrieval necessity (95%), and the degree of support (90%). Agreement was slightly lower in usefulness (80%), mostly due to the disagreement between 1 and 2 or 4 and 5. [|assistant|]
**๊ฒฐ๋ก **. ๊ทธ๋ฆผ 8(b)-8(e)์—์„œ ์ •ํฌ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ฌด์ž‘์œ„ ์ง€์‹์„ ๋ชจ๋ฐฉํ•˜๋Š” ๊ฒฝ์šฐ, training ์‹œ๊ฐ„์„ ํฌ๊ฒŒ ๋Š˜๋ฆฌ์ง€ ์•Š๋Š” ํ•œ ์šฉ๋Ÿ‰ ๋น„์œจ์— _์ค‘๋Œ€ํ•œ ์˜ํ–ฅ_์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 8(f)์—์„œ ์ •ํฌ ๋ฐ์ดํ„ฐ๊ฐ€ ๋งค์šฐ ๋ฐ˜๋ณต์ ์ธ ๊ฒฝ์šฐ ์„ฑ๋Šฅ ์ €ํ•˜๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค.[|endofturn|]
**๊ฒฐ๋ก **. ๊ทธ๋ฆผ 8(b)-8(e)์—์„œ ์ •ํฌ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ฌด์ž‘์œ„ ์ง€์‹์„ ๋ชจ๋ฐฉํ•˜๋Š” ๊ฒฝ์šฐ, ํ›ˆ๋ จ ์‹œ๊ฐ„์„ ํฌ๊ฒŒ ๋Š˜๋ฆฌ์ง€ ์•Š๋Š” ํ•œ ์šฉ๋Ÿ‰ ๋น„์œจ์— _์ค‘๋Œ€ํ•œ ์˜ํ–ฅ_์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 8(f)์—์„œ ์ •ํฌ ๋ฐ์ดํ„ฐ๊ฐ€ ๋งค์šฐ ๋ฐ˜๋ณต์ ์ธ ๊ฒฝ์šฐ ์„ฑ๋Šฅ ์ €ํ•˜๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Conclusions.** In Figure 8(b)-8(e), when junk data mimics random knowledge, capacity ratios are _significantly impacted_ unless training time is substantially increased. In Figure 8(f), if the junk data is highly repetitive, there is no degradation. [|assistant|]
์šฐ๋ฆฌ์˜ ์‹คํ—˜์€ 8๊ฐœ์˜ NVIDIA RTX A5000 GPU๊ฐ€ ์žฅ์ฐฉ๋œ Ubuntu 22.04 ์ปดํ“จํ„ฐ์—์„œ PyTorch 2.0.1 ๋ฐ CUDA 11.8๊ณผ ํ•จ๊ป˜ Python 3.9๋กœ ์ˆ˜ํ–‰๋ฉ๋‹ˆ๋‹ค. ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ LLM(์˜ˆ: Gemma-2b-it ๋ฐ Gemma-7b-it [66], LLaMA-2-13b-chat-hf [10])์„ AIOS์˜ ๋ฐฑ๋ณธ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ ํƒ์€ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์˜ ๋กœ์ปฌ ๋ฐฐํฌ ์ด์ ์— ๊ธฐ๋ฐ˜ํ•œ ๊ฒƒ์œผ๋กœ, ์‹œ๊ฐ„ ์ง€์—ฐ ์‹œ๊ฐ„์„ ์ •ํ™•ํ•˜๊ฒŒ ์ธก์ •ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ˆ˜ํ•™์  ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” Math Agent, ์ƒˆ๋กœ์šด ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ๋ฅผ ์ƒ์„ฑํ•˜๋Š” Narrative Agent, ๋ ˆ์Šคํ† ๋ž‘ ์ถ”์ฒœ์„ ์ œ๊ณตํ•˜๋Š” Rec Agent ๋“ฑ ์„ธ ๊ฐ€์ง€ ์ „๋ฌธ ์—์ด์ „ํŠธ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์—์ด์ „ํŠธ๋Š” ์‹คํ–‰ ์ค‘์— ๋ฐฑ๋ณธ LLM์— 2~3๊ฐœ์˜ ์š”์ฒญ์„ ๋ณด๋‚ด๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์šฐ๋ฆฌ์˜ ์‹คํ—˜์€ 8๊ฐœ์˜ NVIDIA RTX A5000 GPU๊ฐ€ ์žฅ์ฐฉ๋œ Ubuntu 22.04 ์ปดํ“จํ„ฐ์—์„œ PyTorch 2.0.1 ๋ฐ CUDA 11.8๊ณผ ํ•จ๊ป˜ Python 3.9๋กœ ์ˆ˜ํ–‰๋ฉ๋‹ˆ๋‹ค. ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ LLM(์˜ˆ: Gemma-2b-it ๋ฐ Gemma-7b-it [66], LLaMA-2-13b-chat-hf [10])์„ AIOS์˜ ๋ฐฑ๋ณธ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ ํƒ์€ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์˜ ๋กœ์ปฌ ๋ฐฐํฌ ์ด์ ์— ๊ธฐ๋ฐ˜ํ•œ ๊ฒƒ์œผ๋กœ, ์‹œ๊ฐ„ ์ง€์—ฐ ์‹œ๊ฐ„์„ ์ •ํ™•ํ•˜๊ฒŒ ์ธก์ •ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ˆ˜ํ•™์  ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ์ˆ˜ํ•™ ์—์ด์ „ํŠธ, ์ƒˆ๋กœ์šด ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ ์—์ด์ „ํŠธ, ๋ ˆ์Šคํ† ๋ž‘ ์ถ”์ฒœ์„ ์ œ๊ณตํ•˜๋Š” Rec ์—์ด์ „ํŠธ ๋“ฑ ์„ธ ๊ฐ€์ง€ ์ „๋ฌธ ์—์ด์ „ํŠธ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์—์ด์ „ํŠธ๋Š” ์‹คํ–‰ ์ค‘์— ๋ฐฑ๋ณธ LLM์— 2~3๊ฐœ์˜ ์š”์ฒญ์„ ๋ณด๋‚ด๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Our experiments are conducted in Python 3.9 with PyTorch 2.0.1 and CUDA 11.8 on an Ubuntu 22.04 machine equipped with 8 NVIDIA RTX A5000 GPUs. We employ the publicly available LLMs (i.e., Gemma-2b-it and Gemma-7b-it [66], LLaMA-2-13b-chat-hf [10]) as the backbone of AIOS. This selection is driven by the advantage of local deployment of open-sourced models, which aids in the accurate measurement of time latency. For the evaluation, we configure three specialized agents: a Math Agent for solving mathematical challenges, a Narrative Agent for generating novel narratives, and a Rec Agent tasked with providing restaurant recommendations. Each agent is designed to send 2 to 3 requests to the backbone LLM during running. [|assistant|]
Frobenius norm์— ๋Œ€ํ•ด ์ˆœ์œ„ \(d\)์˜ \(W^{*}\)์˜ ์ตœ์  approximation์„ \(W_{d}\)๋กœ ํ‘œ์‹œํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.[|endofturn|]
ํ”„๋กœ๋ฒ ๋‹ˆ์šฐ์Šค ๊ทœ๋ฒ”์— ๋Œ€ํ•ด ์ˆœ์œ„ \(d\)์˜ \(W^{*}\)์˜ ์ตœ์  ๊ทผ์‚ฌ์น˜๋ฅผ \(W_{d}\)๋กœ ํ‘œ์‹œํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Let us note \(W_{d}\) the best approximation of \(W^{*}\) of rank \(d\) with respect to the Frobenius norm. [|assistant|]
์ •ํ™•ํ•˜์ง€ ์•Š์€ ์ธํ„ฐ๋„ท ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด Lazaridou ๋“ฑ(2022)์€ TF-IDF ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์ ์ˆ˜๋ฅผ ๋งค๊ธฐ๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ๊ฐ ์งˆ๋ฌธ q๋ฅผ ๊ทธ๋Œ€๋กœ ์ฟผ๋ฆฌ๋กœ ์‚ฌ์šฉํ•˜๊ณ  Google ๊ฒ€์ƒ‰ API๋ฅผ ํ†ตํ•ด Google ๊ฒ€์ƒ‰์— ํ˜ธ์ถœํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์งˆ๋ฌธ์— ๋Œ€ํ•ด ์ƒ์œ„ 20๊ฐœ์˜ URL์„ ๊ฒ€์ƒ‰ํ•˜๊ณ  HTML ์ฝ˜ํ…์ธ ๋ฅผ ๊ตฌ๋ฌธ ๋ถ„์„ํ•˜์—ฌ ๊นจ๋—ํ•œ ํ…์ŠคํŠธ๋ฅผ ์ถ”์ถœํ•˜์—ฌ ๊ฐ ์งˆ๋ฌธ q์— ๋Œ€ํ•œ ๋ฌธ์„œ ์ง‘ํ•ฉ D๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. Hu ๋“ฑ(2023)์€ ๊ด€๋ จ ์—†๋Š” ์ •๋ณด๊ฐ€ ์‚ฌ์šฉ์ž์˜ ์ฟผ๋ฆฌ ํ•ด๊ฒฐ์„ ๋ฐฉํ•ดํ•˜๋Š” ๊ฒƒ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฒŒ์ดํŒ… ํšŒ๋กœ๋ฅผ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ํšŒ๋กœ๋Š” dual-encoder dot product๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์œ ์‚ฌ์„ฑ์„ ๊ณ„์‚ฐํ•˜๊ณ  ์šฉ์–ด ๊ฐ€์ค‘์น˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฒŒ์ดํŒ… ํšŒ๋กœ๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ Boytsov ๋“ฑ(2016)์€ ๊ฒ€์ƒ‰ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋ฒˆ์—ญ ๋ชจ๋ธ๊ณผ BM25๋ฅผ ๊ฒฐํ•ฉํ•˜๋ฉด์„œ ์šฉ์–ด ๊ธฐ๋ฐ˜ ๊ฒ€์ƒ‰์„ k-Nearest Neighbors(kNN) ๊ฒ€์ƒ‰์œผ๋กœ ๋Œ€์ฒดํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ๋ชจ๋ธ์€ ์šฉ์–ด ๊ฐ„์˜ ์˜๋ฏธ ๊ด€๊ณ„์™€ ๊ธฐ์กด์˜ ํ†ต๊ณ„์  ๊ฐ€์ค‘์น˜ ์ฒด๊ณ„๋ฅผ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ์—ˆ๊ณ , ๊ทธ ๊ฒฐ๊ณผ ๋ณด๋‹ค ํšจ์œจ์ ์ธ ๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ •ํ™•ํ•˜์ง€ ์•Š์€ ์ธํ„ฐ๋„ท ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด Lazaridou ๋“ฑ(2022)์€ TF-IDF ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์ ์ˆ˜๋ฅผ ๋งค๊ธฐ๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ๊ฐ ์งˆ๋ฌธ q๋ฅผ ๊ทธ๋Œ€๋กœ ์ฟผ๋ฆฌ๋กœ ์‚ฌ์šฉํ•˜๊ณ  Google ๊ฒ€์ƒ‰ API๋ฅผ ํ†ตํ•ด Google ๊ฒ€์ƒ‰์— ํ˜ธ์ถœํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์งˆ๋ฌธ์— ๋Œ€ํ•ด ์ƒ์œ„ 20๊ฐœ์˜ URL์„ ๊ฒ€์ƒ‰ํ•˜๊ณ  HTML ์ฝ˜ํ…์ธ ๋ฅผ ๊ตฌ๋ฌธ ๋ถ„์„ํ•˜์—ฌ ๊นจ๋—ํ•œ ํ…์ŠคํŠธ๋ฅผ ์ถ”์ถœํ•˜์—ฌ ๊ฐ ์งˆ๋ฌธ q์— ๋Œ€ํ•œ ๋ฌธ์„œ ์ง‘ํ•ฉ D๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. Hu ๋“ฑ(2023)์€ ๊ด€๋ จ ์—†๋Š” ์ •๋ณด๊ฐ€ ์‚ฌ์šฉ์ž์˜ ์ฟผ๋ฆฌ ํ•ด๊ฒฐ์„ ๋ฐฉํ•ดํ•˜๋Š” ๊ฒƒ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฒŒ์ดํŒ… ํšŒ๋กœ๋ฅผ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ํšŒ๋กœ๋Š” ์ด์ค‘ ์ธ์ฝ”๋” ๋„ํŠธ ํ”„๋กœ๋•ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์œ ์‚ฌ์„ฑ์„ ๊ณ„์‚ฐํ•˜๊ณ  ์šฉ์–ด ๊ฐ€์ค‘์น˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฒŒ์ดํŒ… ํšŒ๋กœ๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ Boytsov ๋“ฑ(2016)์€ ๊ฒ€์ƒ‰ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋ฒˆ์—ญ ๋ชจ๋ธ๊ณผ BM25๋ฅผ ๊ฒฐํ•ฉํ•˜๋ฉด์„œ ์šฉ์–ด ๊ธฐ๋ฐ˜ ๊ฒ€์ƒ‰์„ k-Nearest Neighbors(kNN) ๊ฒ€์ƒ‰์œผ๋กœ ๋Œ€์ฒดํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ๋ชจ๋ธ์€ ์šฉ์–ด ๊ฐ„์˜ ์˜๋ฏธ ๊ด€๊ณ„์™€ ๊ธฐ์กด์˜ ํ†ต๊ณ„์  ๊ฐ€์ค‘์น˜ ์ฒด๊ณ„๋ฅผ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ์—ˆ๊ณ , ๊ทธ ๊ฒฐ๊ณผ ๋ณด๋‹ค ํšจ์œจ์ ์ธ ๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To tackle the issue of inaccurate Internet retrieval results, Lazaridou et al. (2022) proposed using the TF-IDF algorithm to score the retrieval results. They used each question q verbatim as a query and issued a call to Google Search via the Google Search API. For each question, they retrieved the top 20 URLs and parsed their HTML content to extract clean text, generating a set of documents D for each question q. To prevent irrelevant information from hindering the resolution of a user's query, Hu et al. (2023) designed a gating circuit. This circuit utilised a dual-encoder dot product to calculate similarity and a gating circuit based on term weights. Additionally, Boytsov et al. (2016) presented an approach that replaced term-based retrieval with k-Nearest Neighbors(kNN) search while combining a translation model and BM25 to improve retrieval performance. This approach enabled the model to take into account the semantic relationships between terms and traditional statistical weighting schemes, resulting in a more efficient retrieval system. [|assistant|]
๊ทธ๋ฆผ 11: **์‹คํ—˜ 2์˜ Instruction-reversal ์ž‘์—…์— ๋Œ€ํ•œ Llama-7b์˜ ํ›ˆ๋ จ ์ „๋ฐ˜์˜ ์ •ํ™•๋„.**[|endofturn|]
๊ทธ๋ฆผ 11: **์‹คํ—˜ 2์˜ ๋ช…๋ น์–ด ๋ฐ˜์ „ ์ž‘์—…์— ๋Œ€ํ•œ Llama-7b์˜ ํ›ˆ๋ จ ์ „๋ฐ˜์˜ ์ •ํ™•๋„.**[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 11: **Accuracy across training for Llama-7b on the instruction-reversal task for experiment 2.** [|assistant|]
์ด ์ •ํ™•ํ•œ ๋ถ€๋ถ„ ๋ฌธ์ž์—ด ์ผ์น˜ ๊ธฐ์ค€์€ ๊ฐœ๋…์ ์œผ๋กœ๋Š” ๊ฐ„๋‹จํ•˜์ง€๋งŒ, ์ˆœ์ง„ํ•œ (์ด์ฐจ์ ์ธ) ๋ชจ๋“  ์Œ์˜ ์ผ์น˜๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ณ„์‚ฐ์ ์œผ๋กœ ์—„์ฒญ๋‚œ ๋ถ€๋‹ด์ด ๋ฉ๋‹ˆ๋‹ค. ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ „์ฒด ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ \(D\)์˜ ๋ชจ๋“  ์˜ˆ์ œ๋ฅผ ๊ฑฐ๋Œ€ํ•œ ์‹œํ€€์Šค \(\mathcal{S}\)๋กœ ์—ฐ๊ฒฐํ•˜๊ณ , Suffix Array \(\mathcal{A}\)๋ฅผ \(\mathcal{S}\)๋กœ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. Suffix Array[20]๋Š” Suffix Tree[20]์˜ ํ‘œํ˜„์œผ๋กœ, \(\|\mathcal{S}\|\)[13]์—์„œ ์„ ํ˜• ์‹œ๊ฐ„์œผ๋กœ ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํŠนํžˆ ์ค‘๋ณต๋œ ํ•™์Šต ์˜ˆ์ œ๋ฅผ ์„ ํ˜• ์‹œ๊ฐ„ ๋‚ด์— ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ๋Š” ๋งŽ์€ ํ•˜์œ„ ๋ฌธ์ž์—ด ์ฟผ๋ฆฌ๋ฅผ ํšจ์œจ์ ์œผ๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Suffix Array๋Š” Suffix Tree์— ๋น„ํ•ด 10~100๋ฐฐ ๋” ๋ฉ”๋ชจ๋ฆฌ ํšจ์œจ์ ์ด๋ผ๋Š” ์žฅ์ ์ด ์žˆ๋Š”๋ฐ, ์ž…๋ ฅ ํ† ํฐ๋‹น 8๋ฐ”์ดํŠธ๋งŒ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ผ๋ถ€ ์ฟผ๋ฆฌ ์œ ํ˜•์˜ ๊ฒฝ์šฐ ์ ์ง„์ ์œผ๋กœ ํšจ์œจ์„ฑ์ด ๋–จ์–ด์ง€์ง€๋งŒ NLP์—์„œ ํšจ์œจ์ ์ธ TF-IDF ๊ณ„์‚ฐ[22] ๋ฐ ๋ฌธ์„œ ํด๋Ÿฌ์Šคํ„ฐ๋ง[23] ๋“ฑ์— ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ด ์ •ํ™•ํ•œ ๋ถ€๋ถ„ ๋ฌธ์ž์—ด ์ผ์น˜ ๊ธฐ์ค€์€ ๊ฐœ๋…์ ์œผ๋กœ๋Š” ๊ฐ„๋‹จํ•˜์ง€๋งŒ, ์ˆœ์ง„ํ•œ (์ด์ฐจ์ ์ธ) ๋ชจ๋“  ์Œ์˜ ์ผ์น˜๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ณ„์‚ฐ์ ์œผ๋กœ ์—„์ฒญ๋‚œ ๋ถ€๋‹ด์ด ๋ฉ๋‹ˆ๋‹ค. ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ „์ฒด ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ \(D\)์˜ ๋ชจ๋“  ์˜ˆ์ œ๋ฅผ ๊ฑฐ๋Œ€ํ•œ ์‹œํ€€์Šค \(\mathcal{S}\)๋กœ ์—ฐ๊ฒฐํ•˜๊ณ , Suffix Array \(\mathcal{A}\)๋ฅผ \(\mathcal{S}\)๋กœ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด[20]์€ ์ ‘๋ฏธ์‚ฌ ํŠธ๋ฆฌ[20]์˜ ํ‘œํ˜„์œผ๋กœ, \(\|\mathcal{S}\|\)[13]์—์„œ ์„ ํ˜• ์‹œ๊ฐ„์œผ๋กœ ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํŠนํžˆ ์ค‘๋ณต๋œ ํ•™์Šต ์˜ˆ์ œ๋ฅผ ์„ ํ˜• ์‹œ๊ฐ„ ๋‚ด์— ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ๋Š” ๋งŽ์€ ํ•˜์œ„ ๋ฌธ์ž์—ด ์ฟผ๋ฆฌ๋ฅผ ํšจ์œจ์ ์œผ๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด์€ ์ ‘๋ฏธ์‚ฌ ํŠธ๋ฆฌ์— ๋น„ํ•ด 10~100๋ฐฐ ๋” ๋ฉ”๋ชจ๋ฆฌ ํšจ์œจ์ ์ด๋ผ๋Š” ์žฅ์ ์ด ์žˆ๋Š”๋ฐ, ์ž…๋ ฅ ํ† ํฐ๋‹น 8๋ฐ”์ดํŠธ๋งŒ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ผ๋ถ€ ์ฟผ๋ฆฌ ์œ ํ˜•์˜ ๊ฒฝ์šฐ ์ ์ง„์ ์œผ๋กœ ํšจ์œจ์„ฑ์ด ๋–จ์–ด์ง€์ง€๋งŒ NLP์—์„œ ํšจ์œจ์ ์ธ TF-IDF ๊ณ„์‚ฐ[22] ๋ฐ ๋ฌธ์„œ ํด๋Ÿฌ์Šคํ„ฐ๋ง[23] ๋“ฑ์— ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]This exact-substring-matching criterion, while conceptually simple, is computationally prohibitive with naive (quadratic) all-pair matching. To improve the efficiency, we concatenate all the examples of the entire dataset \(D\) into a giant sequence \(\mathcal{S}\), and construct a Suffix Array \(\mathcal{A}\) of \(\mathcal{S}\). A suffix array [20] is a representation of a suffix tree [20] that can be constructed in linear time in \(\|\mathcal{S}\|\)[13] and enables efficient computation of many substring queries; in particular, they allow us to identify duplicated training examples in linear time. Suffix arrays have the advantage over suffix trees in that they are 10-100\(\times\) more memory efficient [21], requiring just 8 bytes per input token, though they are asymptotically less efficient for some query types. They have been used widely in NLP, such as for efficient TF-IDF computation [22] and document clustering [23]. [|assistant|]
์—ฌ๊ธฐ์„œ \(\theta_{chat\_new}\)๋Š” ๊ฒฐ๊ณผ ๋ชจ๋ธ์˜ ๊ฐ€์ค‘์น˜์ด๊ณ , \(\theta_{CP}\)๋Š” 3.1์—์„œ ์–ธ๊ธ‰ํ•œ continue pre-trained ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ๊ฐ„๋‹จํ•˜๊ฒŒ ์ถ”๊ฐ€ํ•จ์œผ๋กœ์จ ๋ชจ๋ธ์€ ๋Œ€์ƒ ์–ธ์–ด์˜ instructions์„ ์ดํ•ดํ•˜๊ณ  ๋”ฐ๋ฅด๋Š” ๋Šฅ๋ ฅ์„ ํš๋“ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์œ ์šฉ์„ฑ ๋ฐ ๋ฌดํ•ด์„ฑ๊ณผ ๊ฐ™์€ ์ง€์ •๋œ ๊ธฐ์ค€์— ๋งž๊ฒŒ ์กฐ์ •๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์—ฌ๊ธฐ์„œ \(\theta_{chat\_new}\)๋Š” ๊ฒฐ๊ณผ ๋ชจ๋ธ์˜ ๊ฐ€์ค‘์น˜์ด๊ณ , \(\theta_{CP}\)๋Š” 3.1์—์„œ ์–ธ๊ธ‰ํ•œ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ๊ฐ„๋‹จํ•˜๊ฒŒ ์ถ”๊ฐ€ํ•จ์œผ๋กœ์จ ๋ชจ๋ธ์€ ๋Œ€์ƒ ์–ธ์–ด์˜ ์ง€์‹œ๋ฅผ ์ดํ•ดํ•˜๊ณ  ๋”ฐ๋ฅด๋Š” ๋Šฅ๋ ฅ์„ ํš๋“ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์œ ์šฉ์„ฑ ๋ฐ ๋ฌดํ•ด์„ฑ๊ณผ ๊ฐ™์€ ์ง€์ •๋œ ๊ธฐ์ค€์— ๋งž๊ฒŒ ์กฐ์ •๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]where \(\theta_{chat\_new}\) is the weights of the resulting model, \(\theta_{CP}\) is the continue pre-trained model mentioned in 3.1. With such simple addition, the model not only obtains the ability to understand and follow instructions in the target language but is also aligned with specified criteria such as helpfulness and harmlessness. [|assistant|]
์ง€๊ธˆ๋ถ€ํ„ฐ ์ง€์† ํ•™์Šต ์„ค์ •์—์„œ ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ํ‘œ์‹œ๋  ๋•Œ ๋ฌดํ•œ Learning Rate ์ผ์ •์˜ ์—ญํ• ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๋ชจ๋ธ์€ SlimPajama์˜ 3๊ฐœ์˜ IID 100B ํ•˜์œ„ ์ง‘ํ•ฉ(์˜ˆ: ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์—†๋Š” ์„ค์ •, ์„น์…˜ 5.2 ์ฐธ์กฐ)์—์„œ ์„œ๋กœ ๋‹ค๋ฅธ Learning Rate ์ผ์ •์œผ๋กœ ๋ฌด์ž‘์œ„ ์ดˆ๊ธฐํ™”๋กœ๋ถ€ํ„ฐ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์ด ์˜ˆ๋น„ ์‹คํ—˜์—์„œ๋Š” ๋ณ€ํ™”๊ฐ€ ์—†๋Š” ์„ค์ •์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์•ฝํ•œ ๋ณ€ํ™”์™€ ๊ฐ•ํ•œ ๋ณ€ํ™”์˜ ๊ฒฝ์šฐ๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ๋กœ ๋‚จ๊ฒจ๋‘๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ๋™์ผํ•œ ๋ถ„ํฌ์˜ ๋Œ€๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ์‹œ๊ฐ„ ๊ฐ„๊ฒฉ์œผ๋กœ ์ˆ˜์‹ ๋˜๊ณ  ์ด์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•˜๊ณ ์ž ํ•˜๋Š” ์„ค์ •(์˜ˆ: ์ตœ์‹  ์›น ์Šคํฌ๋žฉ์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒฝ์šฐ)์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ Optimizer ์ƒํƒœ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๋Š” ์ƒํ™ฉ์— ์ ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ ๊ฒฝ๊ณ„๋ฅผ ๋„˜์–ด Optimizer ์ƒํƒœ๋ฅผ ์œ ์ง€ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 9๋Š” 405M ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ›ˆ๋ จ ๊ณก์„ ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ง€๊ธˆ๋ถ€ํ„ฐ ์ง€์† ํ•™์Šต ์„ค์ •์—์„œ ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ํ‘œ์‹œ๋  ๋•Œ ๋ฌดํ•œ ํ•™์Šต ์†๋„ ์ผ์ •์˜ ์—ญํ• ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๋ชจ๋ธ์€ ์Šฌ๋ฆผํŒŒ์ž๋งˆ์˜ 3๊ฐœ์˜ IID 100B ํ•˜์œ„ ์ง‘ํ•ฉ(์˜ˆ: ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์—†๋Š” ์„ค์ •, ์„น์…˜ 5.2 ์ฐธ์กฐ)์—์„œ ์„œ๋กœ ๋‹ค๋ฅธ ํ•™์Šต ์†๋„ ์ผ์ •์œผ๋กœ ๋ฌด์ž‘์œ„ ์ดˆ๊ธฐํ™”๋กœ๋ถ€ํ„ฐ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์ด ์˜ˆ๋น„ ์‹คํ—˜์—์„œ๋Š” ๋ณ€ํ™”๊ฐ€ ์—†๋Š” ์„ค์ •์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์•ฝํ•œ ๋ณ€ํ™”์™€ ๊ฐ•ํ•œ ๋ณ€ํ™”์˜ ๊ฒฝ์šฐ๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ๋กœ ๋‚จ๊ฒจ๋‘๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ๋™์ผํ•œ ๋ถ„ํฌ์˜ ๋Œ€๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ์‹œ๊ฐ„ ๊ฐ„๊ฒฉ์œผ๋กœ ์ˆ˜์‹ ๋˜๊ณ  ์ด์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•˜๊ณ ์ž ํ•˜๋Š” ์„ค์ •(์˜ˆ: ์ตœ์‹  ์›น ์Šคํฌ๋žฉ์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒฝ์šฐ)์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ์ตœ์ ํ™” ์ƒํƒœ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๋Š” ์ƒํ™ฉ์— ์ ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ ๊ฒฝ๊ณ„๋ฅผ ๋„˜์–ด ์ตœ์ ํ™” ์ƒํƒœ๋ฅผ ์œ ์ง€ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 9๋Š” 405M ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ›ˆ๋ จ ๊ณก์„ ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We now explore the role of the infinite learning rate schedules when multiple new datasets are seen in a continual learning setup. The models are trained from random initialization with different learning rate schedules on 3 IID 100B subsets of SlimPajama (e.g., our _three datasets no shift_ setting; see Sec 5.2). We focus on the no shift setting in these preliminary experiments and leave the weak and strong shift cases to future work. This task simulates a setting where large amounts of data from the same distribution are received at time increments and we wish to continue pre-training our models on them (e.g., continuing to pre-train the model on the latest web-scrape). To make our results applicable to situations where previous optimizer states are not available, we do not keep optimizer states across dataset boundaries. Fig. 9 reports training curves for 405M parameter models. [|assistant|]
๊ทธ๋Ÿฌ๋‚˜ ์ด๋ฐฉ์„ฑ์ด 10์–ต ๊ฐœ ์ด์ƒ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ์—ฌ๋ถ€๋Š” ์•„์ง ๋ช…ํ™•ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์šฐ๋ฆฌ๋Š” ์ผ๋ จ์˜ ๋ชจ๋ธ(์ฆ‰, GPT-2(Radford et al., 2019), OPT(Zhang et al., 2022), Pythia(Biderman et al., 2023), Gemma(Team et al., 2024))์—์„œ ๋ ˆ์ด์–ด ๊ฐ„์˜ ์ค‘๊ฐ„ ํ‘œํ˜„์˜ ํ‰๊ท  cosine-similarity๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋„๋ฉ”์ธ์ด ์ด๋Ÿฌํ•œ ์ œํ’ˆ๊ตฐ์— ์‚ฌ์šฉ๋œ pretraining ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋„๋ฉ”์ธ์„ ํฌํ•จํ•˜๊ฑฐ๋‚˜ ์ผ์น˜ํ•œ๋‹ค๊ณ  ๊ฐ€์ •ํ•˜๊ธฐ ๋•Œ๋ฌธ์— The Pile(Gao et al., 2020)์˜ ํ•˜์œ„ ์ƒ˜ํ”Œ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋Ÿฌ๋‚˜ ์ด๋ฐฉ์„ฑ์ด 10์–ต ๊ฐœ ์ด์ƒ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ์—ฌ๋ถ€๋Š” ์•„์ง ๋ช…ํ™•ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์šฐ๋ฆฌ๋Š” ์ผ๋ จ์˜ ๋ชจ๋ธ(์ฆ‰, GPT-2(Radford et al., 2019), OPT(Zhang et al., 2022), Pythia(Biderman et al., 2023), Gemma(Team et al., 2024))์—์„œ ๋ ˆ์ด์–ด ๊ฐ„์˜ ์ค‘๊ฐ„ ํ‘œํ˜„์˜ ํ‰๊ท  ์ฝ”์‚ฌ์ธ ์œ ์‚ฌ๋„๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋„๋ฉ”์ธ์ด ์ด๋Ÿฌํ•œ ์ œํ’ˆ๊ตฐ์— ์‚ฌ์šฉ๋œ ์ „์ฒ˜๋ฆฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋„๋ฉ”์ธ์„ ํฌํ•จํ•˜๊ฑฐ๋‚˜ ์ผ์น˜ํ•œ๋‹ค๊ณ  ๊ฐ€์ •ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋” ํŒŒ์ผ(Gao et al., 2020)์˜ ํ•˜์œ„ ์ƒ˜ํ”Œ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]However, it remains unclear whether anisotropy affects models with over 1 billion parameters. In order to address this question, we compute average cosine-similarity of intermediate representations across layers in suites of models; namely GPT-2 (Radford et al., 2019), OPT (Zhang et al., 2022), Pythia (Biderman et al., 2023), and Gemma (Team et al., 2024). We use a subsample of The Pile (Gao et al., 2020), as we hypothesize that the domain of this dataset includes or matches the domain of the pretraining datasets used in these suites. [|assistant|]
์ด์ „ ์—ฐ๊ตฌ(Hoffmann et al., 2022)์— ๋”ฐ๋ฅด๋ฉด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ทœ๋ชจ๊ฐ€ Large Language Models (LLMs)์˜ ์„ฑ๋Šฅ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋™์‹œ์— ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋‹ค์–‘์„ฑ๊ณผ ํฌ๊ด„์„ฑ์€ ์ผ๋ฐ˜ ๋„๋ฉ”์ธ์„ ์œ„ํ•œ LLM์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์›์น™๊ณผ ๋ชจ๋ธ ํ•™์Šต์— ์ค‘๊ตญ์–ด ๋ง๋ญ‰์น˜๋ฅผ ํ™œ์šฉํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  1,254.68 billion ๊ฐœ์˜ tokens์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ค‘๊ตญ์–ด, ์˜์–ด ๋ฐ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ฉํ•œ ๊ฒƒ์œผ๋กœ, 840.48 billion ๊ฐœ์˜ ์ค‘๊ตญ์–ด tokens, 314.88 billion ๊ฐœ์˜ ์˜์–ด tokens, 99.3 billion ๊ฐœ์˜ code tokens์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” Common Crawl์˜ ์›น ๋ฌธ์„œ, ํ•™์ˆ  ๋…ผ๋ฌธ, ๋ฐฑ๊ณผ์‚ฌ์ „, ์„œ์  ๋“ฑ ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์˜ ์ฝ˜ํ…์ธ ๋ฅผ ์ง‘๊ณ„ํ•ฉ๋‹ˆ๋‹ค. ์ •ํ™•ํ•œ ๋ถ„ํฌ๋Š” ๊ทธ๋ฆผ 1์— ์ž์„ธํžˆ ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์•ฝ 110 billion ๊ฐœ์˜ ์ค‘๋ณต tokens์ด ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋Œ€๋ถ€๋ถ„ ์˜์–ด๋กœ ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ค‘๋ณต์ด์ง€๋งŒ ๊ณ ํ’ˆ์งˆ์ด๋ฉฐ ์˜๋„์ ์œผ๋กœ ๋‘ ๋ฒˆ ํ•™์Šต์— ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ด์ „ ์—ฐ๊ตฌ(Hoffmann et al., 2022)์— ๋”ฐ๋ฅด๋ฉด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ทœ๋ชจ๊ฐ€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋™์‹œ์— ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋‹ค์–‘์„ฑ๊ณผ ํฌ๊ด„์„ฑ์€ ์ผ๋ฐ˜ ๋„๋ฉ”์ธ์„ ์œ„ํ•œ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์›์น™๊ณผ ๋ชจ๋ธ ํ•™์Šต์— ์ค‘๊ตญ์–ด ๋ง๋ญ‰์น˜๋ฅผ ํ™œ์šฉํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  1,2546์–ต ๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ค‘๊ตญ์–ด, ์˜์–ด ๋ฐ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ฉํ•œ ๊ฒƒ์œผ๋กœ, 8404์–ต 8000์–ต ๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ, 3148์–ต 8000์–ต ๊ฐœ์˜ ์˜์–ด ํ† ํฐ, 993์–ต ๊ฐœ์˜ ์ฝ”๋“œ ํ† ํฐ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ปค๋จผ ํฌ๋กค์˜ ์›น ๋ฌธ์„œ, ํ•™์ˆ  ๋…ผ๋ฌธ, ๋ฐฑ๊ณผ์‚ฌ์ „, ์„œ์  ๋“ฑ ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์˜ ์ฝ˜ํ…์ธ ๋ฅผ ์ง‘๊ณ„ํ•ฉ๋‹ˆ๋‹ค. ์ •ํ™•ํ•œ ๋ถ„ํฌ๋Š” ๊ทธ๋ฆผ 1์— ์ž์„ธํžˆ ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์•ฝ 1100์–ต ๊ฐœ์˜ ์ค‘๋ณต ํ† ํฐ์ด ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋Œ€๋ถ€๋ถ„ ์˜์–ด๋กœ ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ค‘๋ณต์ด์ง€๋งŒ ๊ณ ํ’ˆ์งˆ์ด๋ฉฐ ์˜๋„์ ์œผ๋กœ ๋‘ ๋ฒˆ ํ•™์Šต์— ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Previous research (Hoffmann et al., 2022) has established that the magnitude of the dataset significantly influences the performance of large language models. Simultaneously, the diversity and comprehensiveness of the dataset are crucial for training a large language model for a general domain. Guided by the aforementioned principles and our emphasis on utilizing Chinese corpora for model training, we have developed a dataset encompassing 1,254.68 billion tokens. This dataset integrates Chinese, English, and code data, consisting of 840.48 billion Chinese tokens, 314.88 billion English tokens, and 99.3 billion code tokens. The dataset aggregates content from diverse sources, such as web documents from Common Crawl, scholarly articles, encyclopedias, and books. The precise distribution is detailed in the Figure.1. Our dataset contains around 110 billion duplicate tokens, mostly in English. Despite being duplicates, they are high quality and were intentionally used twice in training. [|assistant|]
๊ทธ๋ฆผ 7: (์™ผ์ชฝ) ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ๊ฐ ๋‹จ๊ณ„์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ํŒŒ๋ž€์ƒ‰ ๋ง‰๋Œ€๋Š” ๋‹จ์ผ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ๋‹จ๊ณ„์˜ ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ฃผ๊ณ  ์ฃผํ™ฉ์ƒ‰ ๋ง‰๋Œ€๋Š” 40๊ฐœ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ์ด๋“์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. (๊ฐ€์šด๋ฐ) ๋‹จ๊ณ„๋ณ„ sampling-and-voting์€ ๋‹ค์–‘ํ•œ ์ˆ˜์ค€์˜ ๋‚ด์žฌ์  ๋‚œ์ด๋„์— ๊ฑธ์ณ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (์˜ค๋ฅธ์ชฝ) ๊ณ„์ธต์  sampling-and-voting์€ ๋™์ข… ๋ฐ ์ด์ข… ๋ชจ๋ธ ์กฐํ•ฉ์œผ๋กœ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 7: (์™ผ์ชฝ) ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ๊ฐ ๋‹จ๊ณ„์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ํŒŒ๋ž€์ƒ‰ ๋ง‰๋Œ€๋Š” ๋‹จ์ผ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ๋‹จ๊ณ„์˜ ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ฃผ๊ณ  ์ฃผํ™ฉ์ƒ‰ ๋ง‰๋Œ€๋Š” 40๊ฐœ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ์ด๋“์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. (๊ฐ€์šด๋ฐ) ๋‹จ๊ณ„๋ณ„ ์ƒ˜ํ”Œ๋ง ๋ฐ ํˆฌํ‘œ๋Š” ๋‹ค์–‘ํ•œ ์ˆ˜์ค€์˜ ๋‚ด์žฌ์  ๋‚œ์ด๋„์— ๊ฑธ์ณ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (์˜ค๋ฅธ์ชฝ) ๊ณ„์ธต์  ์ƒ˜ํ”Œ๋ง ๋ฐ ํˆฌํ‘œ๋Š” ๋™์ข… ๋ฐ ์ด์ข… ๋ชจ๋ธ ์กฐํ•ฉ์œผ๋กœ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 7: (Left) Our method increases the performance for each step. Blue bars show the accuracy of various steps for a single sample, and orange bars show the gains for 40 samples. (Middle) Step-wise sampling-and-voting can further enhance the performance across different levels of inherent difficulty. (Right) Hierarchical sampling-and-voting can further enhance the performance with homogeneous and heterogeneous model combinations. [|assistant|]
๊ทธ๋ฆผ 1: Self-Rag ๊ฐœ์š”. Self-Rag์€ ํ…์ŠคํŠธ ๊ตฌ์ ˆ์„ retrieval, critique ๋ฐ generateํ•˜์—ฌ ์ „๋ฐ˜์ ์ธ ์ƒ์„ฑ ํ’ˆ์งˆ, ์‚ฌ์‹ค์„ฑ ๋ฐ ๊ฒ€์ฆ ๊ฐ€๋Šฅ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 1: ์…€ํ”„๋ž˜๊ทธ ๊ฐœ์š”. Self-Rag์€ ํ…์ŠคํŠธ ๊ตฌ์ ˆ์„ ๊ฒ€์ƒ‰, ๋น„ํ‰ ๋ฐ ์ƒ์„ฑํ•˜์—ฌ ์ „๋ฐ˜์ ์ธ ์ƒ์„ฑ ํ’ˆ์งˆ, ์‚ฌ์‹ค์„ฑ ๋ฐ ๊ฒ€์ฆ ๊ฐ€๋Šฅ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 1: Overview of Self-Rag. Self-Rag learns to retrieve, critique, and generate text passages to enhance overall generation quality, factuality, and verifiability. [|assistant|]
๋งˆ๋ฆฌ์•„๋Š” ๊ณผ์ˆ˜์›์— ์‚ฌ๊ณผ๋‚˜๋ฌด 80๊ทธ๋ฃจ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์‚ฌ๊ณผ๋‚˜๋ฌด์˜ 3๋ถ„์˜ 1๋ณด๋‹ค 4๊ทธ๋ฃจ ์ ์€ ์ฒด๋ฆฌ๋‚˜๋ฌด๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ๋ฆฌ์•„์˜ ๊ณผ์ˆ˜์›์—๋Š” ์ด ๋ช‡ ๊ทธ๋ฃจ์˜ ๋‚˜๋ฌด๊ฐ€ ์žˆ์„๊นŒ์š”?[|endofturn|]
๋งˆ๋ฆฌ์•„๋Š” ๊ณผ์ˆ˜์›์— ์‚ฌ๊ณผ๋‚˜๋ฌด 80๊ทธ๋ฃจ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์‚ฌ๊ณผ๋‚˜๋ฌด๋ณด๋‹ค 3๋ถ„์˜ 1๋„ ์•ˆ ๋˜๋Š” ์ˆ˜์˜ ์ฒด๋ฆฌ๋‚˜๋ฌด๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ๋ฆฌ์•„์˜ ๊ณผ์ˆ˜์›์—๋Š” ์ด ๋ช‡ ๊ทธ๋ฃจ์˜ ๋‚˜๋ฌด๊ฐ€ ์žˆ์„๊นŒ์š”?[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Maria has 80 apple trees in her orchard. She also has 4 less than a third as many cherry trees as apple trees. How many trees does Maria have in all in her orchard? [|assistant|]
๊ทธ๋ฆผ 30: DPO๋Š” Reinforcement Learning์„ ํ”ผํ•˜๋ฉด์„œ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ธ๊ฐ„์˜ ํ”ผ๋“œ๋ฐฑ์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์„ Fine-tuningํ•˜๋Š” ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ๋จผ์ € ์‘๋‹ต ์Œ์— ๋Œ€ํ•œ ํ”„๋กฌํ”„ํŠธ ๋ฐ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋ณด์ƒ ๋ชจ๋ธ์„ ๋งž์ถ˜ ๋‹ค์Œ RL์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šต๋œ ๋ณด์ƒ์„ ์ตœ๋Œ€ํ™”ํ•˜๋Š” ์ •์ฑ…์„ ์ฐพ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, DPO๋Š” ๋ช…์‹œ์ ์ธ ๋ณด์ƒ ํ•จ์ˆ˜๋‚˜ RL ์—†์ด ๊ฐ„๋‹จํ•œ ๋ถ„๋ฅ˜ Objective๋ฅผ ํ†ตํ•ด ์„ ํ˜ธ๋„๋ฅผ ๊ฐ€์žฅ ์ž˜ ๋งŒ์กฑํ•˜๋Š” ์ •์ฑ…์„ ์ง์ ‘ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. [139]์˜ ์ƒํ‘œ.[|endofturn|]
๊ทธ๋ฆผ 30: DPO๋Š” ๊ฐ•ํ™” ํ•™์Šต์„ ํ”ผํ•˜๋ฉด์„œ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ธ๊ฐ„์˜ ํ”ผ๋“œ๋ฐฑ์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ๋จผ์ € ์‘๋‹ต ์Œ์— ๋Œ€ํ•œ ํ”„๋กฌํ”„ํŠธ ๋ฐ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋ณด์ƒ ๋ชจ๋ธ์„ ๋งž์ถ˜ ๋‹ค์Œ RL์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šต๋œ ๋ณด์ƒ์„ ์ตœ๋Œ€ํ™”ํ•˜๋Š” ์ •์ฑ…์„ ์ฐพ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, DPO๋Š” ๋ช…์‹œ์ ์ธ ๋ณด์ƒ ํ•จ์ˆ˜๋‚˜ RL ์—†์ด ๊ฐ„๋‹จํ•œ ๋ถ„๋ฅ˜ ๋ชฉํ‘œ๋ฅผ ํ†ตํ•ด ์„ ํ˜ธ๋„๋ฅผ ๊ฐ€์žฅ ์ž˜ ๋งŒ์กฑํ•˜๋Š” ์ •์ฑ…์„ ์ง์ ‘ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. [139]์˜ ์ƒํ‘œ.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Fig. 30: DPO optimizes for human preferences while avoiding reinforcement learning. Existing methods for fine-tuning language models with human feedback first fit a reward model to a dataset of prompts and human preferences over pairs of responses, and then use RL to find a policy that maximizes the learned reward. In contrast, DPO directly optimizes for the policy best satisfying the preferences with a simple classification objective, without an explicit reward function or RL. Courtesy of [139]. [|assistant|]
**Heuristic Rules** ์ €ํฌ๋Š” ๋ฐ์ดํ„ฐ ํ•„ํ„ฐ๋ง์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด heuristic ๊ทœ์น™์„ ์„ค๊ณ„ํ•˜์—ฌ ํ’ˆ์งˆ์ด ๋‚ฎ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ทœ์น™์€ ์—ฌ๋Ÿฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋ชจ๋ธ, ํŠนํžˆ RefinedWeb(Penedo et al., 2023)๊ณผ CC-Net(Wenzek et al., 2020)์—์„œ ํŒŒ์ƒ๋œ ๋ฐฉ๋ฒ•๋ก ๊ณผ Gopher(Rae et al., 2022) ๋ฐ T5(Raffel et al., 2020)์™€ ๊ฐ™์€ ๋‹ค๋ฅธ language models๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๋™์•ˆ ์ ์šฉ๋˜๋Š” ์ผ๋ถ€ ๊ทœ์น™์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ํ•„ํ„ฐ๋ง ์ „๋žต์˜ ํ†ตํ•ฉ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋‚ด์žฌ๋œ ํŠน์„ฑ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋งž์ถคํ™”๋œ ๊ทœ์น™ ์„ธํŠธ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
**ํœด๋ฆฌ์Šคํ‹ฑ ๊ทœ์น™** ์ €ํฌ๋Š” ๋ฐ์ดํ„ฐ ํ•„ํ„ฐ๋ง์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ํœด๋ฆฌ์Šคํ‹ฑ ๊ทœ์น™์„ ์„ค๊ณ„ํ•˜์—ฌ ํ’ˆ์งˆ์ด ๋‚ฎ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ทœ์น™์€ ์—ฌ๋Ÿฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋ชจ๋ธ, ํŠนํžˆ RefinedWeb(Penedo et al., 2023)๊ณผ CC-Net(Wenzek et al., 2020)์—์„œ ํŒŒ์ƒ๋œ ๋ฐฉ๋ฒ•๋ก ๊ณผ Gopher(Rae et al., 2022) ๋ฐ T5(Raffel et al., 2020)์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋™์•ˆ ์ ์šฉ๋˜๋Š” ์ผ๋ถ€ ๊ทœ์น™์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ํ•„ํ„ฐ๋ง ์ „๋žต์˜ ํ†ตํ•ฉ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋‚ด์žฌ๋œ ํŠน์„ฑ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋งž์ถคํ™”๋œ ๊ทœ์น™ ์„ธํŠธ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Heuristic Rules** We designed heuristic rules to conduct data filtering, which removes data of low quality. These rules represent an integrated framework of filtering strategies, inspired by methodologies derived from several datasets and models, notably RefinedWeb (Penedo et al., 2023) and CC-Net (Wenzek et al., 2020), along with some rules that are applied while training other language models, such as Gopher (Rae et al., 2022) and T5 (Raffel et al., 2020). We also developed a set of rules tailored to address characteristics inherent to our dataset. [|assistant|]
๊ทธ๋ฆผ 4: **๋ฒ ์ด์Šค ๋ชจ๋ธ์˜ ์˜ํ–ฅ.** ๋ฒ ์ด์Šค ๋ชจ๋ธ SaulLM-7B์—์„œ Instruction Fine-tuning์„ ์‹œ์ž‘ํ•˜๋ฉด Mistral-7B์— ๋น„ํ•ด ๋ˆˆ์— ๋„๋Š” ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ์ผ๋ฐ˜์ ์ธ IFT ๋ฏน์Šค(๋ฒ•๋ฅ  ์ œ์™ธ)์—์„œ๋„ SaulLM-7B(Gen.)๊ฐ€ Mistral-Instruct์˜ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. IFT ๋ฏน์Šค์— ๋ฒ•๋ฅ  ์ง€์นจ์„ ์ถ”๊ฐ€ํ•˜๋ฉด ๊ฒฐ๊ณผ๊ฐ€ ๋”์šฑ ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 4: **๋ฒ ์ด์Šค ๋ชจ๋ธ์˜ ์˜ํ–ฅ.** ๋ฒ ์ด์Šค ๋ชจ๋ธ SaulLM-7B์—์„œ ๋ช…๋ น ๋ฏธ์„ธ ์กฐ์ •์„ ์‹œ์ž‘ํ•˜๋ฉด Mistral-7B์— ๋น„ํ•ด ๋ˆˆ์— ๋„๋Š” ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ์ผ๋ฐ˜์ ์ธ IFT ๋ฏน์Šค(๋ฒ•๋ฅ  ์ œ์™ธ)์—์„œ๋„ SaulLM-7B(Gen.)๊ฐ€ Mistral-Instruct์˜ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. IFT ๋ฏน์Šค์— ๋ฒ•๋ฅ  ์ง€์นจ์„ ์ถ”๊ฐ€ํ•˜๋ฉด ๊ฒฐ๊ณผ๊ฐ€ ๋”์šฑ ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 4: **Influence of the base model.** Starting the instruction finetuning from our base model SaulLM-7B brings noticeable improvements compared to the Mistral-7B. Indeed, even with a generic IFT mix (without legal), SaulLM-7B (Gen.) outperforms its Mistral-Instruct counterpart significantly. Adding legal instructions to the IFT mix further boosts the results. [|assistant|]
et al., 2021). ์ฆ๊ฐ•์„ ์œ„ํ•ด generative models๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์€ ๊ณผ๊ฑฐ์— ๋ถ„๋ฅ˜ ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ๋„ ์„ฑ๊ณต์ ์œผ๋กœ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค Sharma et al. (2022). ์˜์—ญ์„ ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์˜ ์˜ˆ๋Š” ๋ถ€๋ก B์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
et al., 2021). ์ฆ๊ฐ•์„ ์œ„ํ•ด ์ƒ์„ฑ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์€ ๊ณผ๊ฑฐ์— ๋ถ„๋ฅ˜ ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ๋„ ์„ฑ๊ณต์ ์œผ๋กœ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค Sharma et al. (2022). ์˜์—ญ์„ ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์˜ ์˜ˆ๋Š” ๋ถ€๋ก B์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]et al., 2021). Using generative models for augmentations has also been used successfully to improve classification models in the past Sharma et al. (2022). An example of data augmentation using paraphrasing can be found in Appendix B. [|assistant|]
์••์ถ• ๋ฉ”๋ชจ๋ฆฌ ์‹œ์Šคํ…œ์€ ๋งค์šฐ ๊ธด ์‹œํ€€์Šค์— ๋Œ€ํ•ด attention ๋ฉ”์ปค๋‹ˆ์ฆ˜๋ณด๋‹ค ํ™•์žฅ์„ฑ๊ณผ ํšจ์œจ์„ฑ์ด ๋” ๋†’์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค(Kanerva, 1988; Munkhdalai et al., 2019). ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ๋Š” ์ž…๋ ฅ ์‹œํ€€์Šค ๊ธธ์ด์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•˜๋Š” ๋ฐฐ์—ด์„ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹ , ์ฃผ๋กœ ์ œํ•œ๋œ ์ €์žฅ ๋ฐ ๊ณ„์‚ฐ ๋น„์šฉ์œผ๋กœ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๊ณ  ๋ถˆ๋Ÿฌ์˜ค๊ธฐ ์œ„ํ•ด ๊ณ ์ •๋œ ์ˆ˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ƒˆ๋กœ์šด ์ •๋ณด๋Š” ์ด ์ •๋ณด๋ฅผ ๋‚˜์ค‘์— ๋‹ค์‹œ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋ณ€๊ฒฝํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ์— ์ถ”๊ฐ€๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ˜„์žฌ ์ƒํƒœ์˜ LLM์€ ๋‹จ์ˆœ์„ฑ๊ณผ ํ’ˆ์งˆ์˜ ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ํšจ๊ณผ์ ์ด๊ณ  ์‹ค์šฉ์ ์ธ ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ ๊ธฐ์ˆ ์„ ์•„์ง ๋ณด์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์••์ถ• ๋ฉ”๋ชจ๋ฆฌ ์‹œ์Šคํ…œ์€ ๋งค์šฐ ๊ธด ์‹œํ€€์Šค์— ๋Œ€ํ•ด ์ฃผ์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜๋ณด๋‹ค ํ™•์žฅ์„ฑ๊ณผ ํšจ์œจ์„ฑ์ด ๋” ๋†’์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค(Kanerva, 1988; Munkhdalai et al., 2019). ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ๋Š” ์ž…๋ ฅ ์‹œํ€€์Šค ๊ธธ์ด์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•˜๋Š” ๋ฐฐ์—ด์„ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹ , ์ฃผ๋กœ ์ œํ•œ๋œ ์ €์žฅ ๋ฐ ๊ณ„์‚ฐ ๋น„์šฉ์œผ๋กœ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๊ณ  ๋ถˆ๋Ÿฌ์˜ค๊ธฐ ์œ„ํ•ด ๊ณ ์ •๋œ ์ˆ˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ƒˆ๋กœ์šด ์ •๋ณด๋Š” ์ด ์ •๋ณด๋ฅผ ๋‚˜์ค‘์— ๋‹ค์‹œ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋ณ€๊ฒฝํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ์— ์ถ”๊ฐ€๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ˜„์žฌ ์ƒํƒœ์˜ LLM์€ ๋‹จ์ˆœ์„ฑ๊ณผ ํ’ˆ์งˆ์˜ ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ํšจ๊ณผ์ ์ด๊ณ  ์‹ค์šฉ์ ์ธ ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ ๊ธฐ์ˆ ์„ ์•„์ง ๋ณด์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Compressive memory systems promise to be more scalable and efficient than the attention mechanism for extremely long sequences (Kanerva, 1988; Munkhdalai et al., 2019). Instead of using an array that grows with the input sequence length, a compressive memory primarily maintains a fixed number of parameters to store and recall information with a bounded storage and computation costs. In the compressive memory, new information is added to the memory by changing its parameters with an objective that this information can be recovered back later on. However, the LLMs in their current state have yet to see an effective, practical compressive memory technique that balances simplicity along with quality. [|assistant|]