chosen
stringlengths
14
4.26k
rejected
stringlengths
53
9.3k
prompt
stringlengths
178
4.55k
ICL ๋ฐ CoT๋กœ Promptingํ•˜๋Š” ๊ฒƒ์€ ๊ฐœ๋…์ ์œผ๋กœ ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ˆ˜ํ•™์  ์ถ”๋ก [525] ๋ฐ multi-hop question answering[526]๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ์ž‘์—…์—๋Š” ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ํ–ฅ์ƒ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ์„œ ๋ณต์žกํ•œ ์ž‘์—…์„ ๋” ์ž‘์€ ํ•˜์œ„ ์ž‘์—…์œผ๋กœ ๋ถ„ํ•ดํ•˜๊ณ  ์ž‘์—…์„ ์™„๋ฃŒํ•˜๊ธฐ ์œ„ํ•œ ์ž‘์—… ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๋Š” prompt-based planning์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
ICL ๋ฐ CoT๋กœ ํ”„๋กฌํ”„ํŠธํ•˜๋Š” ๊ฒƒ์€ ๊ฐœ๋…์ ์œผ๋กœ ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ˆ˜ํ•™์  ์ถ”๋ก [525] ๋ฐ ๋‹ค์ค‘ ํ™‰ ์งˆ๋ฌธ ๋‹ต๋ณ€[526]๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ์ž‘์—…์—๋Š” ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ํ–ฅ์ƒ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ์„œ ๋ณต์žกํ•œ ์ž‘์—…์„ ๋” ์ž‘์€ ํ•˜์œ„ ์ž‘์—…์œผ๋กœ ๋ถ„ํ•ดํ•˜๊ณ  ์ž‘์—…์„ ์™„๋ฃŒํ•˜๊ธฐ ์œ„ํ•œ ์ž‘์—… ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ ๊ธฐ๋ฐ˜ ๊ณ„ํš์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Prompting with ICL and CoT is a conceptually simple yet general approach to solving various tasks. However, this approach struggles with complex tasks like mathematical reasoning [525] and multi-hop question answering [526]. As an enhanced approach, prompt-based planning has been proposed to break down complex tasks into smaller sub-tasks and generate a plan of actions to accomplish the task. [|assistant|]
๊ธด ์ปจํ…์ŠคํŠธ์˜ Large Language Models (LLMs)์€ ๊ธฐ์ˆ ์ ์œผ๋กœ ์ฑ… ๊ธธ์ด์˜ ๋ฌธ์„œ(\(>\) 10๋งŒ ํ† ํฐ)๋ฅผ ์š”์•ฝํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋ฌธ์„œ์˜ ๊ธธ์ด์™€ ๋ณต์žก์„ฑ์œผ๋กœ ์ธํ•ด ์ง€๊ธˆ๊นŒ์ง€ ์ถฉ์‹ค๋„์™€ ๊ฐ™์€ ์ž…๋ ฅ ์˜์กด์ ์ธ ์ธก๋ฉด์— ๋Œ€ํ•œ ํ‰๊ฐ€๊ฐ€ ๊ธˆ์ง€๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๊ฐ€์ƒ์˜ ์ฑ…์„ ์ƒ์„ฑํ•œ LLM์˜ ์š”์•ฝ์— ๋Œ€ํ•œ ์ถฉ์‹ค๋„์™€ ๋‚ด์šฉ ์„ ํƒ์— ๋Œ€ํ•œ ์ตœ์ดˆ์˜ ๋Œ€๊ทœ๋ชจ ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” 2023๋…„ ๋˜๋Š” 2024๋…„์— ์ถœํŒ๋œ ์ฑ…์˜ ์š”์•ฝ์— ์ดˆ์ ์„ ๋งž์ถค์œผ๋กœ์จ ๋ฐ์ดํ„ฐ ์˜ค์—ผ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ณ , ๋น„์šฉ๊ณผ ์ธ์ง€ ๋ถ€๋‹ด์„ ์ตœ์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ฃผ์„ ์ž‘์—… ์ „์— ๊ฐ ์ฑ…์„ ์™„์ „ํžˆ ์ฝ์€ ์ฃผ์„๊ฐ€๋ฅผ ๊ณ ์šฉํ•ฉ๋‹ˆ๋‹ค. 5.2๋งŒ ๋‹ฌ๋Ÿฌ์˜ ๋น„์šฉ์œผ๋กœ 26๊ถŒ์˜ ์ฑ…์„ ์š”์•ฝํ•œ LLM์—์„œ ์ƒ์„ฑ๋œ 3,158๊ฐœ์˜ ์ฃผ์žฅ์— ๋Œ€ํ•œ ์ฃผ์„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ Fables๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ ์ถฉ์‹ค๋„๋ฅผ ๊ธฐ์ค€์œผ๋กœ LLM ์š”์•ฝ๊ธฐ๋ฅผ ์ˆœ์œ„๋ฅผ ๋งค๊ธธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: Claude-3-Opus๋Š” ๋ชจ๋“  ๋น„๊ณต๊ฐœ LLM๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋ฐ˜๋ฉด, ์˜คํ”ˆ ์†Œ์Šค์ธ Mktral์€ GPT-3.5-Turbo์™€ ๋™๋“ฑํ•œ ์ˆ˜์ค€์ž…๋‹ˆ๋‹ค. ์ฃผ์„์„ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ, ๋ถˆ์ถฉ์‹คํ•œ ์ฃผ์žฅ์€ ๋Œ€๋ถ€๋ถ„ ์‚ฌ๊ฑด ๋ฐ ์บ๋ฆญํ„ฐ ์ƒํƒœ์™€ ๊ด€๋ จ์ด ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ์— ๋Œ€ํ•œ ๊ฐ„์ ‘์ ์ธ ์ถ”๋ก ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. LLM ๊ธฐ๋ฐ˜ ์ž๋™ ํ‰๊ฐ€๊ธฐ๋Š” ๋‹ค๋ฅธ ํ™˜๊ฒฝ์—์„œ ์‚ฌ์‹ค์„ฑ๊ณผ ์ผ๊ด€์„ฑ์— ๋Œ€ํ•ด ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์ง€๋งŒ, ์ถฉ์‹ค๋„์— ๋Œ€ํ•œ ์—ฌ๋Ÿฌ LLM ํ‰๊ฐ€๊ธฐ๋ฅผ ๊ตฌํ˜„ํ•œ ๊ฒฐ๊ณผ ํŠนํžˆ ๋ถˆ์ถฉ์‹คํ•œ ์ฃผ์žฅ์„ ๊ฐ์ง€ํ•˜๋Š” ๋ฐ ์žˆ์–ด ์ธ๊ฐ„์˜ ์ฃผ์„๊ณผ ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์—†๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ๋ถˆ์ถฉ์‹คํ•œ ์ฃผ์žฅ์„ ๊ฐ์ง€ํ•˜๋Š” ๊ฒƒ์€ ์š”์•ฝ ํ‰๊ฐ€๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์žฅ๋ฌธ ๋ฌธ๋งฅ ์ดํ•ด๋ฅผ ์œ„ํ•œ ํ…Œ์ŠคํŠธ๋ฒ ๋“œ๋กœ์„œ๋„ ์ค‘์š”ํ•œ ๋ฏธ๋ž˜ ๋ฐฉํ–ฅ์ด๋ผ๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ถฉ์‹ค๋„๋ฅผ ๋„˜์–ด ์ฑ… ๊ธธ์ด์˜ ์š”์•ฝ์—์„œ ๋‚ด์šฉ ์„ ํƒ ์˜ค๋ฅ˜๋ฅผ ํƒ๊ตฌํ•˜์—ฌ ์ค‘์š”ํ•œ ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ ์š”์†Œ์™€ ๊ด€๋ จ๋œ ์ƒ๋žต ์˜ค๋ฅ˜์˜ ์œ ํ˜•์„ ๊ฐœ๋ฐœํ•˜๊ณ  ์ฑ…์˜ ๋งˆ์ง€๋ง‰ ๋ถ€๋ถ„์—์„œ ๋ฐœ์ƒํ•˜๋Š” ์‚ฌ๊ฑด์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ๊ณผ์ž‰ ๊ฐ•์กฐ๋„ ํŒŒ์•…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ฑ… ๊ธธ์ด์˜ ์š”์•ฝ ํ‰๊ฐ€์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด Fables๋ฅผ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ธด ์ปจํ…์ŠคํŠธ์˜ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ๊ธฐ์ˆ ์ ์œผ๋กœ ์ฑ… ๊ธธ์ด์˜ ๋ฌธ์„œ(\(>\) 10๋งŒ ํ† ํฐ)๋ฅผ ์š”์•ฝํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋ฌธ์„œ์˜ ๊ธธ์ด์™€ ๋ณต์žก์„ฑ์œผ๋กœ ์ธํ•ด ์ง€๊ธˆ๊นŒ์ง€ ์ถฉ์‹ค๋„์™€ ๊ฐ™์€ ์ž…๋ ฅ ์˜์กด์ ์ธ ์ธก๋ฉด์— ๋Œ€ํ•œ ํ‰๊ฐ€๊ฐ€ ๊ธˆ์ง€๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๊ฐ€์ƒ์˜ ์ฑ…์„ ์ƒ์„ฑํ•œ LLM์˜ ์š”์•ฝ์— ๋Œ€ํ•œ ์ถฉ์‹ค๋„์™€ ๋‚ด์šฉ ์„ ํƒ์— ๋Œ€ํ•œ ์ตœ์ดˆ์˜ ๋Œ€๊ทœ๋ชจ ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” 2023๋…„ ๋˜๋Š” 2024๋…„์— ์ถœํŒ๋œ ์ฑ…์˜ ์š”์•ฝ์— ์ดˆ์ ์„ ๋งž์ถค์œผ๋กœ์จ ๋ฐ์ดํ„ฐ ์˜ค์—ผ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ณ , ๋น„์šฉ๊ณผ ์ธ์ง€ ๋ถ€๋‹ด์„ ์ตœ์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ฃผ์„ ์ž‘์—… ์ „์— ๊ฐ ์ฑ…์„ ์™„์ „ํžˆ ์ฝ์€ ์ฃผ์„๊ฐ€๋ฅผ ๊ณ ์šฉํ•ฉ๋‹ˆ๋‹ค. 5.2๋งŒ ๋‹ฌ๋Ÿฌ์˜ ๋น„์šฉ์œผ๋กœ 26๊ถŒ์˜ ์ฑ…์„ ์š”์•ฝํ•œ LLM์—์„œ ์ƒ์„ฑ๋œ 3,158๊ฐœ์˜ ์ฃผ์žฅ์— ๋Œ€ํ•œ ์ฃผ์„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ Fables๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ ์ถฉ์‹ค๋„๋ฅผ ๊ธฐ์ค€์œผ๋กœ LLM ์š”์•ฝ๊ธฐ๋ฅผ ์ˆœ์œ„๋ฅผ ๋งค๊ธธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: Claude-3-Opus๋Š” ๋ชจ๋“  ๋น„๊ณต๊ฐœ LLM๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋ฐ˜๋ฉด, ์˜คํ”ˆ ์†Œ์Šค์ธ Mktral์€ GPT-3.5-Turbo์™€ ๋™๋“ฑํ•œ ์ˆ˜์ค€์ž…๋‹ˆ๋‹ค. ์ฃผ์„์„ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ, ๋ถˆ์ถฉ์‹คํ•œ ์ฃผ์žฅ์€ ๋Œ€๋ถ€๋ถ„ ์‚ฌ๊ฑด ๋ฐ ์บ๋ฆญํ„ฐ ์ƒํƒœ์™€ ๊ด€๋ จ์ด ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ์— ๋Œ€ํ•œ ๊ฐ„์ ‘์ ์ธ ์ถ”๋ก ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. LLM ๊ธฐ๋ฐ˜ ์ž๋™ ํ‰๊ฐ€๊ธฐ๋Š” ๋‹ค๋ฅธ ํ™˜๊ฒฝ์—์„œ ์‚ฌ์‹ค์„ฑ๊ณผ ์ผ๊ด€์„ฑ์— ๋Œ€ํ•ด ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์ง€๋งŒ, ์ถฉ์‹ค๋„์— ๋Œ€ํ•œ ์—ฌ๋Ÿฌ LLM ํ‰๊ฐ€๊ธฐ๋ฅผ ๊ตฌํ˜„ํ•œ ๊ฒฐ๊ณผ ํŠนํžˆ ๋ถˆ์ถฉ์‹คํ•œ ์ฃผ์žฅ์„ ๊ฐ์ง€ํ•˜๋Š” ๋ฐ ์žˆ์–ด ์ธ๊ฐ„์˜ ์ฃผ์„๊ณผ ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์—†๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ๋ถˆ์ถฉ์‹คํ•œ ์ฃผ์žฅ์„ ๊ฐ์ง€ํ•˜๋Š” ๊ฒƒ์€ ์š”์•ฝ ํ‰๊ฐ€๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์žฅ๋ฌธ ๋ฌธ๋งฅ ์ดํ•ด๋ฅผ ์œ„ํ•œ ํ…Œ์ŠคํŠธ๋ฒ ๋“œ๋กœ์„œ๋„ ์ค‘์š”ํ•œ ๋ฏธ๋ž˜ ๋ฐฉํ–ฅ์ด๋ผ๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ถฉ์‹ค๋„๋ฅผ ๋„˜์–ด ์ฑ… ๊ธธ์ด์˜ ์š”์•ฝ์—์„œ ๋‚ด์šฉ ์„ ํƒ ์˜ค๋ฅ˜๋ฅผ ํƒ๊ตฌํ•˜์—ฌ ์ค‘์š”ํ•œ ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ ์š”์†Œ์™€ ๊ด€๋ จ๋œ ์ƒ๋žต ์˜ค๋ฅ˜์˜ ์œ ํ˜•์„ ๊ฐœ๋ฐœํ•˜๊ณ  ์ฑ…์˜ ๋งˆ์ง€๋ง‰ ๋ถ€๋ถ„์—์„œ ๋ฐœ์ƒํ•˜๋Š” ์‚ฌ๊ฑด์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ๊ณผ์ž‰ ๊ฐ•์กฐ๋„ ํŒŒ์•…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ฑ… ๊ธธ์ด์˜ ์š”์•ฝ ํ‰๊ฐ€์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด Fables๋ฅผ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]While long-context large language models (LLMs) can technically summarize book-length documents (\(>\) 100K tokens), the length and complexity of the documents have so far prohibited evaluations of input-dependent aspects like faithfulness. In this paper, we conduct the first large-scale human evaluation of faithfulness and content selection on LLM-generated summaries of fictional books. Our study mitigates the issue of data contamination by focusing on summaries of books published in 2023 or 2024, and we hire annotators who have fully read each book prior to the annotation task to minimize cost and cognitive burden. We collect Fables, a dataset of annotations on 3,158 claims made in LLM-generated summaries of 26 books, at a cost of $5.2K USD, which allows us to rank LLM summarizers based on faithfulness: Claude-3-Opus significantly outperforms all closed-source LLMs, while the open-source Mktral is on par with GPT-3.5-Turbo. An analysis of the annotations reveals that most unfaithful claims relate to events and character states, and they generally require indirect reasoning over the narrative to invalidate. While LLM-based auto-raters have proven reliable for factuality and coherence in other settings, we implement several LLM raters of faithfulness and find that none correlates strongly with human annotations, especially with regard to detecting unfaithful claims. Our experiments suggest that detecting unfaithful claims is an important future direction not only for summarization evaluation but also as a testbed for long-context understanding. Finally, we move beyond faithfulness by exploring content selection errors in book-length summarization: we develop a typology of omission errors related to crucial narrative elements and also identify a systematic over-emphasis on events occurring towards the end of the book. We release Fables to spur further research on the evaluation of book-length summarization. [|assistant|]
๊ทธ๋ฆผ A13: ํ›ˆ๋ จ์— ๋”ฐ๋ฅธ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์— ๊ฑธ์นœ ๋ฐ˜๋ณต ํ† ํฐ์˜ ํšจ๊ณผ. X์ถ•์€ ์—…๋ฐ์ดํŠธ ํšŸ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋‚ด๊ณ , Y์ถ•์€ ์›น ์Šค๋ƒ…์ƒท์ด ์•„๋‹Œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ(์œ„์ชฝ ํ–‰)์™€ Instruct OPT(์•„๋ž˜์ชฝ ํ–‰)์˜ ํ‰๊ท  Perplexity๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์œ„ ํ”Œ๋กฏ์˜ ๊ฐ ์—ด์€ ์„œ๋กœ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ฐ ์—ด ๋‚ด์—์„œ (1) ํšŒ์ƒ‰ ์„ ์€ baseline ํ›ˆ๋ จ์„, (2) ์ƒ‰์ƒ ์ ์„ ์€ ์ง€์ •๋œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ ํ†ตํ•œ ํ† ํฐ ๋ฐ˜๋ณต์„, (3) ์ƒ‰์ƒ ์‹ค์„ ์€ ์ง€์ •๋œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ ํ†ตํ•œ ์ƒˆ ํ† ํฐ ์„ ํƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋ฐ˜๋ณต ๋ฐ์ดํ„ฐ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ _๊ณ ์ •๋œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•_(์˜ˆ: ๊ณ ์ • ์—ด)์— ๋Œ€ํ•ด ์ƒˆ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ๋‚˜์ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ A13: ํ›ˆ๋ จ์— ๋”ฐ๋ฅธ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์— ๊ฑธ์นœ ๋ฐ˜๋ณต ํ† ํฐ์˜ ํšจ๊ณผ. X์ถ•์€ ์—…๋ฐ์ดํŠธ ํšŸ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋‚ด๊ณ , Y์ถ•์€ ์›น ์Šค๋ƒ…์ƒท์ด ์•„๋‹Œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ(์œ„์ชฝ ํ–‰)์™€ Instruct OPT(์•„๋ž˜์ชฝ ํ–‰)์˜ ํ‰๊ท  ๋‹นํ˜น๋„๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์œ„ ํ”Œ๋กฏ์˜ ๊ฐ ์—ด์€ ์„œ๋กœ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ฐ ์—ด ๋‚ด์—์„œ (1) ํšŒ์ƒ‰ ์„ ์€ ๊ธฐ์ค€์„  ํ›ˆ๋ จ์„, (2) ์ƒ‰์ƒ ์ ์„ ์€ ์ง€์ •๋œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ ํ†ตํ•œ ํ† ํฐ ๋ฐ˜๋ณต์„, (3) ์ƒ‰์ƒ ์‹ค์„ ์€ ์ง€์ •๋œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ ํ†ตํ•œ ์ƒˆ ํ† ํฐ ์„ ํƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋ฐ˜๋ณต ๋ฐ์ดํ„ฐ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ _๊ณ ์ •๋œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•_(์˜ˆ: ๊ณ ์ • ์—ด)์— ๋Œ€ํ•ด ์ƒˆ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ๋‚˜์ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure A13: Effect of repeating tokens across data selection methods over training. X-axis denotes the number of updates, and the y-axis denotes average perplexity across non-web-snapshot validation sets (top row) and Instruct OPT (bottom row). Each column in the plot above denotes a different data selection method. Within each column: (1) the gray line denotes baseline training, (2) the colored-dashed line denotes repeating tokens via the specified data selection method, and (3) the colored-solid line denotes selecting new tokens via the specified data selection method. Repeating data is generally worse than selecting new data for a _fixed data selection method_ (e.g., fixed column). [|assistant|]
LLMs๋Š” ์ด์ œ ์ผ๋ฐ˜์ ์œผ๋กœ ๋” ํฌ๊ณ  ๋ณต์žกํ•œ ์‹œ์Šคํ…œ์˜ ์ผ๋ถ€๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์ด ์‹ ๋ขฐ๋„, ์ •ํ™•๋„ ๋ฐ ํ†ต์ผ์„ฑ์ด ๋‹ค์–‘ํ•œ ์ •๋ณด์™€ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ๋ฐฉ์‹์„ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋ถ„์„ ๊ฒฐ๊ณผ, ๋ฌธ๋งฅ ์ •๋ณด๊ฐ€ ์ฃผ์–ด์ง„ ์งˆ๋ฌธ์— LLMs์„ ์‚ฌ์šฉํ•  ๋•Œ ๋ฐœ์ƒํ•˜๋Š” ์œ„ํ—˜์„ ํŠน์„ฑํ™”ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ถ”๊ฐ€์ ์ธ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ๋ชจ๋ธ์ด ์ด์ „ ์‹ ๋…์˜ ํ•œ๊ณ„์— ์กด์žฌํ•˜๋Š” ์ •๋ณด๋ฅผ ์ œ์‹œ๋ฐ›์„ ๋•Œ ๋ชจ๋ธ ๋™์ž‘์ด ๋ถˆ๊ทœ์น™ํ•˜๊ณ  ์˜ˆ์ธกํ•  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
LLM์€ ์ด์ œ ์ผ๋ฐ˜์ ์œผ๋กœ ๋” ํฌ๊ณ  ๋ณต์žกํ•œ ์‹œ์Šคํ…œ์˜ ์ผ๋ถ€๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์ด ์‹ ๋ขฐ๋„, ์ •ํ™•๋„ ๋ฐ ํ†ต์ผ์„ฑ์ด ๋‹ค์–‘ํ•œ ์ •๋ณด์™€ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ๋ฐฉ์‹์„ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋ถ„์„ ๊ฒฐ๊ณผ, ๋ฌธ๋งฅ ์ •๋ณด๊ฐ€ ์ฃผ์–ด์ง„ ์งˆ๋ฌธ์— LLM์„ ์‚ฌ์šฉํ•  ๋•Œ ๋ฐœ์ƒํ•˜๋Š” ์œ„ํ—˜์„ ํŠน์„ฑํ™”ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ถ”๊ฐ€์ ์ธ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ๋ชจ๋ธ์ด ์ด์ „ ์‹ ๋…์˜ ํ•œ๊ณ„์— ์กด์žฌํ•˜๋Š” ์ •๋ณด๋ฅผ ์ œ์‹œ๋ฐ›์„ ๋•Œ ๋ชจ๋ธ ๋™์ž‘์ด ๋ถˆ๊ทœ์น™ํ•˜๊ณ  ์˜ˆ์ธกํ•  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]LLMs are now commonly used as parts of larger, more complex systems. It is crucial to understand how these models interact with information with varying degrees of trustworthiness, accuracy, and uniformity. Our analysis shows that further work is required to characterize the risks of using LLMs to answer questions given contextual information. In particular, we find that model behavior can be erratic and unpredictable when presented with information that exists at the margin of its prior beliefs. [|assistant|]
_Remark A.3_ (# GPUs).: ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๊ด€๋ จ์ด ์—†์œผ๋ฏ€๋กœ GPU ์ˆ˜๋ฅผ ์ง€์ •ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. Batch Size๊ฐ€ 24์ธ 64๊ฐœ์˜ GPU, Batch Size๊ฐ€ 32์ธ 48๊ฐœ์˜ GPU ๋˜๋Š” Batch Size๊ฐ€ 1์ธ 1536๊ฐœ์˜ GPU๋ฅผ ์‚ฌ์šฉํ•˜๋“  ๊ฒฐ๊ณผ๋Š” ๋™์ผํ•˜๊ฒŒ ์œ ์ง€๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
_Remark A.3_ (# GPU).: ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๊ด€๋ จ์ด ์—†์œผ๋ฏ€๋กœ GPU ์ˆ˜๋ฅผ ์ง€์ •ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ฐฐ์น˜ ํฌ๊ธฐ๊ฐ€ 24์ธ 64๊ฐœ์˜ GPU, ๋ฐฐ์น˜ ํฌ๊ธฐ๊ฐ€ 32์ธ 48๊ฐœ์˜ GPU ๋˜๋Š” ๋ฐฐ์น˜ ํฌ๊ธฐ๊ฐ€ 1์ธ 1536๊ฐœ์˜ GPU๋ฅผ ์‚ฌ์šฉํ•˜๋“  ๊ฒฐ๊ณผ๋Š” ๋™์ผํ•˜๊ฒŒ ์œ ์ง€๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]_Remark A.3_ (# GPUs).: In this paper, we do not specify the number of GPUs as it is irrelevant. The results remain the same whether using 64 GPUs each with a batch size of 24, 48 GPUs each with a batch size of 32, or 1536 GPUs each with a batch size of 1. [|assistant|]
_๋””๋ฐ”์ธ ๋ฆฌ๋ถˆ์Šค๋Š” ๋‘ ๋ช…์˜ ๋ถ€ํ™œํ•œ ๋ผ์ด๋ฒŒ ์‹ , ์—๋ฅด์™€์™€ ๋‹คํฌ๋ ˆ๊ฐ€ ๋ถˆ๋Ÿฌ์˜จ ์ „์Ÿ ์ค‘์ธ ๋””์Šคํ† ํ”ผ์•„ ์„ธ๊ณ„๋ฅผ ๋ฐฐ๊ฒฝ์œผ๋กœ ํ•œ ๋ณต์žกํ•œ ์ด์•ผ๊ธฐ์ž…๋‹ˆ๋‹ค. ์ค„๊ฑฐ๋ฆฌ๋Š” ๋„์‹œ ์นดํƒ€์˜ ์นดํƒ€ ๊ฐ€์ œํŠธ ์‹ ๋ฌธ์‚ฌ์˜ ์˜ˆ๋ฏผํ•˜๊ณ  ๊ฐ•์ธํ•œ ์ง์› ์•„์ด๋ฆฌ์Šค ์œˆ๋…ธ์šฐ์™€ ๊ทธ๋…€์˜ ์ง์—…์  ๋ผ์ด๋ฒŒ ๋กœ๋งŒ ํ‚ทํŠธ๋ฅผ ์ค‘์‹ฌ์œผ๋กœ ์ „๊ฐœ๋ฉ๋‹ˆ๋‹ค. ์•„์ด๋ฆฌ์Šค๋Š” ๋™์ƒ ํฌ๋ ˆ์ŠคํŠธ๊ฐ€ ์‹ ๋“ค์˜ ์ „์Ÿ์— ํ•ฉ๋ฅ˜ํ•˜๋ฉด์„œ ์•Œ์ฝ”์˜ฌ ์ค‘๋…์— ๊ฑธ๋ฆฐ ์–ด๋จธ๋‹ˆ ์•„์Šคํ„ฐ์˜ ์ฑ…์ž„๊ฐ์„ ๋Š๋ผ๋ฉฐ ํ˜ผ๋ˆ ์†์—์„œ ๋ฒ—์–ด๋‚˜๋ ค ํ•ฉ๋‹ˆ๋‹ค. ๊ฐœ์ธ์ , ์ง์—…์  ๊ฒฝ์Ÿ์˜ ์†Œ์šฉ๋Œ์ด์— ํœ˜๋ง๋ฆฐ ์•„์ด๋ฆฌ์Šค์™€ ๋กœ๋งŒ์˜ ์œ ๋Œ€๊ฐ์€ ๋กœ๋งŒ์ด ํฌ๋ ˆ์ŠคํŠธ์—๊ฒŒ ๋ณด๋‚ด๋Š” ์•„์ด๋ฆฌ์Šค์˜ ๊ฐ์ •์ ์œผ๋กœ ๋“œ๋ฌธ ํŽธ์ง€๋ฅผ ๋ฌด์‹ฌ์ฝ” ๋ฐ›๊ณ  ์‘๋‹ตํ•˜๋ฉด์„œ ์˜ˆ์ƒ์น˜ ๋ชปํ•œ ๋ฐฉํ–ฅ์œผ๋กœ ์ „๊ฐœ๋ฉ๋‹ˆ๋‹ค. ์•„์ด๋ฆฌ์Šค๋Š” ๋งˆ๋ฒ•์˜ ํ„ฐ์„ธ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด ํŽธ์ง€๋ฅผ ๋ณด๋‚ด๋ฉฐ ๋‘ ์‚ฌ๋žŒ์˜ ๋งํ•˜์ง€ ์•Š์€ ์—ฐ๊ฒฐ์„ ์ฆํญ์‹œํ‚ค๋Š” ์ต๋ช…์˜ ์„œ์‹ ์„ ๋ณด๋ƒ…๋‹ˆ๋‹ค. ๋กœ๋งŒ์˜ ์‚ฌํšŒ์  ์ง€์œ„๋Š” ์•„์ด๋ฆฌ์Šค์˜ ๊ฒธ์†ํ•œ ๋ฐฐ๊ฒฝ๊ณผ ๊ทน๋ช…ํ•œ ๋Œ€์กฐ๋ฅผ ์ด๋ฃจ๋ฉฐ, ๊ทธ๋Š” ๊ณ ์ƒ์ธต ์—ฌ์„ฑ ์—˜๋ฆฌ๋„ˆ ๋ฆฌํ‹€๊ณผ ์•ฝํ˜ผํ•ด ๊ฐ€์กฑ์˜ ์—ด๋ง๊ณผ ์บ์Šค ๊ฐ€์ œํŠธ์˜ ์ƒˆ๋กœ์šด ์—ญํ•  ์‚ฌ์ด์—์„œ ๊ท ํ˜•์„ ๋งž์ถ”๊ธฐ ์œ„ํ•ด ์–ด๋ ค์›€์— ์ง๋ฉดํ•ฉ๋‹ˆ๋‹ค. ์–ด๋จธ๋‹ˆ์˜ ๊ฐ‘์ž‘์Šค๋Ÿฌ์šด ์‚ฌ๋ง ์ดํ›„ ์•„์ด๋ฆฌ์Šค๋Š” ์ถฉ๋™์ ์œผ๋กœ ์บ์Šค ๊ฐ€์ œํŠธ์—์„œ์˜ ์ง์žฅ์„ ๊ทธ๋งŒ๋‘ก๋‹ˆ๋‹ค. ํฌ๋ ˆ์ŠคํŠธ๋ฅผ ์ฐพ๊ณ  ๊ฒฉํ™”๋˜๋Š” ์‹ ๋“ค์˜ ์ „์Ÿ์— ๋Œ€ํ•œ ๋ณด๋„๋ฅผ ํ•˜๊ธฐ ์œ„ํ•ด ๊ฒฐ์‹ฌํ•œ ๊ทธ๋…€๋Š” ์ธ๋ฆฌ๋“œ๋ฆฌ๋ฐ์ดํŠธ ํŠธ๋ฆฌ๋ทด ์‹ ๋ฌธ์‚ฌ์˜ ์ „์Ÿ ํŠนํŒŒ์›์œผ๋กœ ์ผํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์•„์ด๋ฆฌ์Šค๋Š” ๋˜ ๋‹ค๋ฅธ ํŠนํŒŒ์›์ธ ํ… "์• ํ‹ฐ" ์•ณ์›Œ๋“œ์™€ ํ•จ๊ป˜ ์ „์Ÿ ์ „์„  ๊ทผ์ฒ˜์˜ ์•„ํƒ€๋ฐ˜ ๋ธ”๋ฆฌํ”„๋กœ ๋– ๋‚˜๋Š” ์œ„ํ—˜ํ•œ ์—ฌ์ •์„ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์—๋ฆฌ ํˆฌ์–ด์˜ ์—ฌ์ฃผ์ธ์ธ ๋งˆ๋ฅด์‹œ์˜ค ํ† ๋ ˆ์Šค๋Š” ๋‘ ์‚ฌ๋žŒ์—๊ฒŒ ์‚ฌ๋ƒฅ๊ฐœ์™€ ์—์ผํŠธ๋ผํˆฌ์Šค ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์‹ ํ™” ์† ์ƒ๋ฌผ๋“ค์˜ ๊ณต๊ฒฉ์— ๋Œ€ํ•ด ๊ฒฝ๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๋กœ๋งŒ์ด ์•„ํƒ€๋ฐ˜ ๋ธ”๋ฆฌํ”„์— ๋„์ฐฉํ•˜๋ฉด์„œ ๋‘ ์‚ฌ๋žŒ์˜ ๋ณต์žกํ•œ ๊ด€๊ณ„๊ฐ€ ๋‹ค์‹œ ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค[|endofturn|]
_๋””๋ฐ”์ธ ๋ฆฌ๋ถˆ์Šค๋Š” ๋‘ ๋ช…์˜ ๋ถ€ํ™œํ•œ ๋ผ์ด๋ฒŒ ์‹ , ์—๋ฅด์™€์™€ ๋‹คํฌ๋ ˆ๊ฐ€ ๋ถˆ๋Ÿฌ์˜จ ์ „์Ÿ ์ค‘์ธ ๋””์Šคํ† ํ”ผ์•„ ์„ธ๊ณ„๋ฅผ ๋ฐฐ๊ฒฝ์œผ๋กœ ํ•œ ๋ณต์žกํ•œ ์ด์•ผ๊ธฐ์ž…๋‹ˆ๋‹ค. ์ค„๊ฑฐ๋ฆฌ๋Š” ๋„์‹œ ์นดํƒ€์˜ ์นดํƒ€ ๊ฐ€์ œํŠธ ์‹ ๋ฌธ์‚ฌ์˜ ์˜ˆ๋ฏผํ•˜๊ณ  ๊ฐ•์ธํ•œ ์ง์› ์•„์ด๋ฆฌ์Šค ์œˆ๋…ธ์šฐ์™€ ๊ทธ๋…€์˜ ์ง์—…์  ๋ผ์ด๋ฒŒ ๋กœ๋งŒ ํ‚ทํŠธ๋ฅผ ์ค‘์‹ฌ์œผ๋กœ ์ „๊ฐœ๋ฉ๋‹ˆ๋‹ค. ์•„์ด๋ฆฌ์Šค๋Š” ๋™์ƒ ํฌ๋ ˆ์ŠคํŠธ๊ฐ€ ์‹ ๋“ค์˜ ์šฐ๋ฅด์— ํ•ฉ๋ฅ˜ํ•˜๋ฉด์„œ ์•Œ์ฝ”์˜ฌ ์ค‘๋…์— ๊ฑธ๋ฆฐ ์–ด๋จธ๋‹ˆ ์•„์Šคํ„ฐ์˜ ์ฑ…์ž„๊ฐ์„ ๋Š๋ผ๋ฉฐ ํ˜ผ๋ˆ ์†์—์„œ ๋ฒ—์–ด๋‚˜๋ ค ํ•ฉ๋‹ˆ๋‹ค. ๊ฐœ์ธ์ , ์ง์—…์  ๊ฒฝ์Ÿ์˜ ์†Œ์šฉ๋Œ์ด์— ํœ˜๋ง๋ฆฐ ์•„์ด๋ฆฌ์Šค์™€ ๋กœ๋งŒ์˜ ์œ ๋Œ€๊ฐ์€ ๋กœ๋งŒ์ด ํฌ๋ ˆ์ŠคํŠธ์—๊ฒŒ ๋ณด๋‚ด๋Š” ์•„์ด๋ฆฌ์Šค์˜ ๊ฐ์ •์ ์œผ๋กœ ๋“œ๋ฌธ ํŽธ์ง€๋ฅผ ๋ฌด์‹ฌ์ฝ” ๋ฐ›๊ณ  ์‘๋‹ตํ•˜๋ฉด์„œ ์˜ˆ์ƒ์น˜ ๋ชปํ•œ ๋ฐฉํ–ฅ์œผ๋กœ ์ „๊ฐœ๋ฉ๋‹ˆ๋‹ค. ์•„์ด๋ฆฌ์Šค๋Š” ๋งˆ๋ฒ•์˜ ํ„ฐ์„ธ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด ํŽธ์ง€๋ฅผ ๋ณด๋‚ด๋ฉฐ ๋‘ ์‚ฌ๋žŒ์˜ ๋งํ•˜์ง€ ์•Š์€ ์—ฐ๊ฒฐ์„ ์ฆํญ์‹œํ‚ค๋Š” ์ต๋ช…์˜ ์„œ์‹ ์„ ๋ณด๋ƒ…๋‹ˆ๋‹ค. ๋กœ๋งŒ์˜ ์‚ฌํšŒ์  ์ง€์œ„๋Š” ์•„์ด๋ฆฌ์Šค์˜ ๊ฒธ์†ํ•œ ๋ฐฐ๊ฒฝ๊ณผ ๊ทน๋ช…ํ•œ ๋Œ€์กฐ๋ฅผ ์ด๋ฃจ๋ฉฐ, ๊ทธ๋Š” ๊ณ ์ƒ์ธต ์—ฌ์„ฑ ์—˜๋ฆฌ๋„ˆ ๋ฆฌํ‹€๊ณผ ์•ฝํ˜ผํ•ด ๊ฐ€์กฑ์˜ ์—ด๋ง๊ณผ ์บ์Šค ๊ฐ€์ œํŠธ์˜ ์ƒˆ๋กœ์šด ์—ญํ•  ์‚ฌ์ด์—์„œ ๊ท ํ˜•์„ ๋งž์ถ”๊ธฐ ์œ„ํ•ด ์–ด๋ ค์›€์— ์ง๋ฉดํ•ฉ๋‹ˆ๋‹ค. ์–ด๋จธ๋‹ˆ์˜ ๊ฐ‘์ž‘์Šค๋Ÿฌ์šด ์‚ฌ๋ง ์ดํ›„ ์•„์ด๋ฆฌ์Šค๋Š” ์ถฉ๋™์ ์œผ๋กœ ์บ์Šค ๊ฐ€์ œํŠธ์—์„œ์˜ ์ง์žฅ์„ ๊ทธ๋งŒ๋‘ก๋‹ˆ๋‹ค. ํฌ๋ ˆ์ŠคํŠธ๋ฅผ ์ฐพ๊ณ  ๊ฒฉํ™”๋˜๋Š” ์‹ ๋“ค์˜ ์šฐ๋ฅด์— ๋Œ€ํ•œ ๋ณด๋„๋ฅผ ํ•˜๊ธฐ ์œ„ํ•ด ๊ฒฐ์‹ฌํ•œ ๊ทธ๋…€๋Š” ์ธ๋ฆฌ๋“œ๋ฆฌ๋ฐ์ดํŠธ ํŠธ๋ฆฌ๋ทด ์‹ ๋ฌธ์‚ฌ์˜ ์šฐ๋ฅด ํŠนํŒŒ์›์œผ๋กœ ์ผํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์•„์ด๋ฆฌ์Šค๋Š” ๋˜ ๋‹ค๋ฅธ ํŠนํŒŒ์›์ธ ํ… "์• ํ‹ฐ" ์•ณ์›Œ๋“œ์™€ ํ•จ๊ป˜ ์šฐ๋ฅด ์ „์„  ๊ทผ์ฒ˜์˜ ์•„ํƒ€๋ฐ˜ ๋ธ”๋ฆฌํ”„๋กœ ๋– ๋‚˜๋Š” ์œ„ํ—˜ํ•œ ์—ฌ์ •์„ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์—๋ฆฌ ํˆฌ์–ด์˜ ์—ฌ์ฃผ์ธ์ธ ๋งˆ๋ฅด์‹œ์˜ค ํ† ๋ ˆ์Šค๋Š” ๋‘ ์‚ฌ๋žŒ์—๊ฒŒ ์‚ฌ๋ƒฅ๊ฐœ์™€ ์—์ผํŠธ๋ผํˆฌ์Šค ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์‹ ํ™” ์† ์ƒ๋ฌผ๋“ค์˜ ๊ณต๊ฒฉ์— ๋Œ€ํ•ด ๊ฒฝ๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๋กœ๋งŒ์ด ์•„ํƒ€๋ฐ˜ ๋ธ”๋ฆฌํ”„์— ๋„์ฐฉํ•˜๋ฉด์„œ ๋‘ ์‚ฌ๋žŒ์˜ ๋ณต์žกํ•œ ๊ด€๊ณ„๊ฐ€ ๋‹ค์‹œ ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Š” ์•„์ด๋ฆฌ์Šค๊ฐ€ ์ „์Ÿ์˜ ์ฐธํ˜นํ•œ ํ˜„์‹ค์„ ๊ธฐ๋กํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ฃผ๋ฉด์„œ ๋‘ ์‚ฌ๋žŒ ์‚ฌ์ด์˜ ๊ธด์žฅ๊ฐ์„ ์กฐ์ ˆํ•ฉ๋‹ˆ๋‹ค. ๋กœ๋งŒ์ด ์ „์„ ์—์„œ ๋”์ฐํ•œ ๋งŒ๋‚จ์„ ๊ฒช์œผ๋ฉฐ ์‹ฌ๊ฐํ•œ ๋ถ€์ƒ์„ ์ž…๊ฒŒ ๋˜์ž, ๋ณ‘์‚ฌ๋“ค์˜ ๊ฐœ์ธ์ ์ธ ์ด์•ผ๊ธฐ๋ฅผ ๋ณด๋„ํ•˜๋ฉด์„œ ๊นŠ์€ ์ธ์—ฐ์„ ๋งบ๊ฒŒ ๋œ ์•„์ด๋ฆฌ์Šค๋Š” ๋กœ๋งŒ์˜ ์•ˆ์ „์„ ์ง€ํ‚ค๊ธฐ ์œ„ํ•ด ์ตœ์„ ์„ ๋‹คํ•ฉ๋‹ˆ๋‹ค. ์ด ์ƒ๋ช…์„ ์œ„ํ˜‘ํ•˜๋Š” ์‚ฌ๊ฑด์œผ๋กœ ์ธํ•ด ์•„์ด๋ฆฌ์Šค๋Š” ๋กœ๋งŒ์„ ๋ฏธ์Šคํ„ฐ๋ฆฌํ•œ ํŽœ ํŒŒํŠธ๋„ˆ์ธ ์นด๋ ˆ๋กœ ์ธ์‹ํ•˜๊ฒŒ ๋˜๊ณ , ๋กœ๋งŒ์˜ ๋ถ€์žฌ์— ๋Œ€ํ•œ ์ฃ„์ฑ…๊ฐ๊ณผ ์˜์‹ฌ์ด ๊ฒฉ๋ ฌํ•˜๊ฒŒ ํ„ฐ์ ธ ๋‚˜์˜ค๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ๋กœ๋งŒ์˜ ์šด๋ช…์— ๋Œ€ํ•œ ์šฐ๋ ค์™€ ๊ทธ์˜ ๋ฌ˜์‚ฌ์— ๋Œ€ํ•œ ์ฃ„์ฑ…๊ฐ์ด ์ปค์ง€๋Š” ๊ฐ€์šด๋ฐ, ์ˆฒ์ด ์ฃฝ์€ ๊ฒƒ์œผ๋กœ ์ถ”์ •๋˜๋˜ ํฌ๋ ˆ์ŠคํŠธ๊ฐ€ ์‚ด์•„๋‚จ์•„ ์ ์˜ ํŒ์Šค์ธ ๋‹คํฌ๋ ˆ๋ฅผ ์œ„ํ•ด ์‹ธ์› ๋‹ค๋Š” ์ถฉ๊ฒฉ์ ์ธ ๋ฐ˜์ „์ด ํŽผ์ณ์ง‘๋‹ˆ๋‹ค. ์ด์–ด์ง€๋Š” ํ˜ผ๋ž€ ์†์—์„œ ๋กœ๋งŒ์€ ์‚ฌ๋ผ์ง€๊ณ  ์•„์ด๋ฆฌ์Šค๋Š” ํšจ๋„ํ•œ ๋™์ƒ๊ณผ ์‹ค์ข…๋œ ์•„๋“ค์˜ ์ ˆ๋ฐ•ํ•œ ์ˆ˜์ƒ‰ ์‚ฌ์ด์—์„œ ๊ฐˆ๋“ฑํ•ฉ๋‹ˆ๋‹ค. "๋‹ค์ด๋น™ ๋ฆฌ๋ถˆ์Šค"๋Š” ์•„์ด๋ฆฌ์Šค์˜ ๊ฐœ์ธ์ ์ธ ์ด์•ผ๊ธฐ๋Š” ๊ฒฉํ™”๋˜๋Š” ์ „์Ÿ์˜ ๋ฐฐ๊ฒฝ์„ ๋ฐฐ๊ฒฝ์œผ๋กœ ํšŒ๋ณต๋ ฅ๊ณผ ์„ฑ์žฅ์„ ๋‹ค๋ฃจ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์•„์ด๋ฆฌ์Šค๋Š” ํŠนํŒŒ์›์—์„œ ์šฐ๋ฅด์˜ ํ•ต์‹ฌ ์ธ๋ฌผ๋กœ ๋ณ€๋ชจํ•˜๋ฉด์„œ ์กฐ์ƒ ์†์—์„œ ์ธ๊ฐ„ ์ •์‹ ์ด ๊ฒฌ๋ŽŒ๋‚ด๊ณ  ์ ์‘ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๊ตฌํ˜„ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ด์•ผ๊ธฐ๋Š” ์บ๋ฆญํ„ฐ์˜ ๊ฐ•์ธํ•จ, ๊ฐ€์กฑ์˜ ์˜๋ฌด, ์ง์—…์  ๊ฒฝ์Ÿ, ์‚ฌํšŒ์  ๊ทœ๋ฒ”, ์‹ฌ์˜คํ•œ ๋ณ€ํ™”๋ฅผ ์ค‘์‹ฌ์œผ๋กœ ์ „๊ฐœ๋ฉ๋‹ˆ๋‹ค. ๊ด€๊ณ„์˜ ๋ณ€ํ™”์™€ ๊ณ ๋‚œ ์†์—์„œ๋„ ์•„์ด๋ฆฌ์Šค๋Š” ํฌ๋ ˆ์ŠคํŠธ์˜ ๊ณ์„ ์ง€ํ‚ค๊ณ  ๋กœ๋งŒ์„ ์ฐพ๊ธฐ ์œ„ํ•ด ๋…ธ๋ ฅํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]_Divine Rivuls is an intricate narrative set in a dystopian world at war, invoked by two resurrected rival detities, Erwa and Dacre. The plot centres around Iris Winnow, a sensitive and tenacious employee at the Catha Gazette newspaper in the city of Catha, and her professional rival Roman Kitt. Iris is uncurrollit ille spirish into chaos with her brother, Forest, joining the gods' uur, lerque her vithin the responsibility of their alcoholic mother, Aster. Cought in a whirlord of personal and professional rivalry, Iris's bond with Roman depers in unexpected turse when Roman amounously receives and responds to Iris's emotionally rare letters intended for Forest. Iris uses a magical turseytter to send these tiers, lading to an anonymous correspondence that amplifies their unspoken connection. Roman's social standing is a stark contrast to Iris's modest background, he is estimated to a high society woman, Elinor Little, and faces the challenges of striking a balance between his family's aspirations and his new role at Cath Gazette. Following their mother's sudden passing, Iris impulsively resigns from her job at Catha Gazette. Detrimed to locate Forest and report on the escalating gods' uur, she accepts a position as a uur correspondent with the Inridridate Tribune newspaper. Iris embratus on a perious journey to Ataban Blif, a town near the uur front, accompanied by another correspondent, Then "Attie" Attworod. Their hostess in the erie tours, Marsiol Torres, warns the pair about the regular thrust from different mythical creatures such as hounds and eiltratus that desolitate their surroundings. Roman's arrival in Ataban Blif reignies their convoluted relationship. He assists Iris in documenting the harrowing realities of the war while managing the growing tension between them. As Roman becomes severely injured during a horricite encounter at the wavefront, Iris, who has grown deeply linked in reporting the soldiers' personal narratives, fights her uuy to ensure his safety. This life-threatening event leads Iris to recognise Roman as Carrer, her mysterious pen part, triggering an intense turnout of guilt and suspicion over Roman's shortage. Amid her growing concern for Roman's fate and her convited feelings about his description, a shocking twist reveals that Forest, previously presumed dead, has surried and fought for Dacre, the enemy detips. Further ensuing chaos sees Roman disappear, leaving Iris prepared between her obligations towards her effectively nationous brother and the desperate search for her missing lose interest. "Diving Rivuls' intervenes Iris's personal story of resilience and growth against the hockdrop of an escalating war. As she metamorphoses from a correspondent to a key player in the uur, Iris embodies the human spirit's ability to endure and adapt amidst ancestry. The narrative encapsulates the strength of character, familial obligations, professional rivalries, societal norms, and profound transformations at the heart of the story. Despite the evolving relationships and hardships, Iris pledges to remain by Forest's side and continue her search for Roman. [|assistant|]
4. ํ›ˆ๋ จ: RLHF ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ์„ฑ๋Šฅ๊ณผ ์ •ํ™•๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ํŠน์ • ์ž‘์—…์— ๋งž๊ฒŒ Fine-tuningํ•˜๊ฑฐ๋‚˜ ์ฒ˜์Œ๋ถ€ํ„ฐ ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๋Š” ๋ฐฉ์‹์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
4. ํ›ˆ๋ จ: RLHF ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ์„ฑ๋Šฅ๊ณผ ์ •ํ™•๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ํŠน์ • ์ž‘์—…์— ๋งž๊ฒŒ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๊ฑฐ๋‚˜ ์ฒ˜์Œ๋ถ€ํ„ฐ ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๋Š” ๋ฐฉ์‹์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]4. Training: Train the model using RLHF methods to improve its performance and accuracy. This can be done by fine-tuning the pre-trained model on a specific task or by training the model from scratch. [|assistant|]
ํ‰๊ฐ€๋œ ๋ถ„ํ•  ๋ฐ ์˜ˆ์ œ ์ˆ˜๋Š” ํ‘œ 5์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ, ์šฐ๋ฆฌ๋Š” Gopher [32]์™€ Chinchilla [16]์˜ ํ‰๊ฐ€ ๋ถ„ํ•  ์„ ํƒ์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ, ground labels์„ ํ•ญ์ƒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์„ ๋•Œ validation split์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. CLUEWSC์˜ ๊ฒฝ์šฐ validation set์˜ ํฌ๊ธฐ๊ฐ€ ๋„ˆ๋ฌด ์ž‘๊ธฐ ๋•Œ๋ฌธ์—(100) train ๋ฐ validation splits์„ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค. GSM8K-Chinese๋Š” ๊ธฐ๊ณ„ ๋ฒˆ์—ญ๊ณผ ์‚ฌ๋žŒ์˜ ๊ต์ •์„ ํ†ตํ•ด GSM8K์—์„œ ๋ฒˆ์—ญ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
ํ‰๊ฐ€๋œ ๋ถ„ํ•  ๋ฐ ์˜ˆ์ œ ์ˆ˜๋Š” ํ‘œ 5์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ, ์šฐ๋ฆฌ๋Š” ๊ณ ํผ [32]์™€ ์นœ์น ๋ผ [16]์˜ ํ‰๊ฐ€ ๋ถ„ํ•  ์„ ํƒ์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ, ๊ทธ๋ผ์šด๋“œ ๋ ˆ์ด๋ธ”์„ ํ•ญ์ƒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์„ ๋•Œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋ถ„ํ• ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. CLUEWSC์˜ ๊ฒฝ์šฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์˜ ํฌ๊ธฐ๊ฐ€ ๋„ˆ๋ฌด ์ž‘๊ธฐ ๋•Œ๋ฌธ์—(100) ํ›ˆ๋ จ ๋ฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋ถ„ํ• ์„ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค. GSM8K-Chinese๋Š” ๊ธฐ๊ณ„ ๋ฒˆ์—ญ๊ณผ ์‚ฌ๋žŒ์˜ ๊ต์ •์„ ํ†ตํ•ด GSM8K์—์„œ ๋ฒˆ์—ญ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The evaluated splits and numbers of examples are summarized in Table 5. For English datasets, we follow Gopher [32] and Chinchilla [16]'s selection of evaluation splits. For Chinese datasets, we use the validation split when the ground labels are always available. For CLUEWSC, the size of the validation set is too small (100), so we combine the train and validation splits. GSM8K-Chinese is translated from GSM8K with machine translation and human proofreading. [|assistant|]
T5 ๋ชจ๋ธ๊ณผ ๋‹ฌ๋ฆฌ GPT-2 ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด 1 epoch ๋™์•ˆ ์ถ”๊ฐ€์ ์ธ _light-tuning_์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. InvariantLAMA์˜ ๊ฒฝ์šฐ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ๊ธธ์ด๋Š” ๊ฐ๊ฐ 50๊ณผ 3์ž…๋‹ˆ๋‹ค. Training batch size๋Š” 32์ด๊ณ  learning rate๋Š” 1e-3์ž…๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด ์ง€์‹ ์Šต๋“์— ๋Œ€ํ•œ ํ‰๊ฐ€์˜ ๊ฒฝ์šฐ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ๊ธธ์ด๋Š” ๊ฐ๊ฐ 100๊ณผ 10์ž…๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด training batch size๋Š” 8์ด๊ณ  learning rate๋Š” 1e-3์ž…๋‹ˆ๋‹ค. ๋‘ tuning ํ”„๋กœ์„ธ์Šค ๋ชจ๋‘์— 4๊ฐœ์˜ V100 32GB GPU๊ฐ€ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. GPT-2 ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๊ฒฐ๊ณผ์™€ ๋…ผ์˜๋Š” ๋ถ€๋ก G์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
T5 ๋ชจ๋ธ๊ณผ ๋‹ฌ๋ฆฌ GPT-2 ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด 1์—ํฌํฌ ๋™์•ˆ ์ถ”๊ฐ€์ ์ธ _๋ผ์ดํŠธ ํŠœ๋‹_์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. InvariantLAMA์˜ ๊ฒฝ์šฐ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ๊ธธ์ด๋Š” ๊ฐ๊ฐ 50๊ณผ 3์ž…๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 32์ด๊ณ  ํ•™์Šต ์†๋„๋Š” 1e-3์ž…๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด ์ง€์‹ ์Šต๋“์— ๋Œ€ํ•œ ํ‰๊ฐ€์˜ ๊ฒฝ์šฐ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ๊ธธ์ด๋Š” ๊ฐ๊ฐ 100๊ณผ 10์ž…๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ํ›ˆ๋ จ ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 8์ด๊ณ  ํ•™์Šต ์†๋„๋Š” 1e-3์ž…๋‹ˆ๋‹ค. ๋‘ ํŠœ๋‹ ํ”„๋กœ์„ธ์Šค ๋ชจ๋‘์— 4๊ฐœ์˜ V100 32GB GPU๊ฐ€ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. GPT-2 ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๊ฒฐ๊ณผ์™€ ๋…ผ์˜๋Š” ๋ถ€๋ก G์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Unlike T5 models, GPT-2 based models need additional _light-tuning_ for 1 epoch for evaluation. For InvariantLAMA, the input and output length is 50 and 3 respectively. The training batch size is 32 and the learning rate is 1e-3. For evaluation on the acquisition of new knowledge, the input and output length is 100 and 10 respectively. The training batch size is 8 due to memory constraints and the learning rate is 1e-3. For both tuning processes, 4 V100 32GB GPUs are used. The detailed result and discussion of GPT-2 based models are shown in Appendix G. [|assistant|]
๊ฐ์ฃผ 31: RLHF์—์„œ๋Š” ๋ณด์ƒ ๋ชจ๋ธ์ด ์ •๋ ฌ๋  LLM์˜ ์ง€์‹์ด๋‚˜ ๋Šฅ๋ ฅ์„ ์ธ์‹ํ•˜๋Š” ๊ฒƒ๋„ ์ค‘์š”ํ•œ ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, LLAMA 2๋Š” pre-trained chat model checkpoints๋ฅผ ์ฑ„ํƒํ•˜์—ฌ ๋ณด์ƒ ๋ชจ๋ธ์„ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค[99].[|endofturn|]
๊ฐ์ฃผ 31: RLHF์—์„œ๋Š” ๋ณด์ƒ ๋ชจ๋ธ์ด ์ •๋ ฌ๋  LLM์˜ ์ง€์‹์ด๋‚˜ ๋Šฅ๋ ฅ์„ ์ธ์‹ํ•˜๋Š” ๊ฒƒ๋„ ์ค‘์š”ํ•œ ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, LLAMA 2๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ์ฑ„ํŒ… ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ์ฑ„ํƒํ•˜์—ฌ ๋ณด์ƒ ๋ชจ๋ธ์„ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค[99].[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Footnote 31: In RLHF, it seems to be also important that reward models should be aware of the knowledge or ability of a LLM to be aligned. For example, LLAMA 2 adopts pre-trained chat model checkpoints to initialize reward models [99]. [|assistant|]
๊ด‘๋ฒ”์œ„ํ•œ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜์™€ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๊ฑฐ๋“ญ์ œ๊ณฑ ๋ฒ•์น™์— ๋”ฐ๋ผ ๋‘๋“œ๋Ÿฌ์ง„ ์„ฑ์žฅ ํŒจํ„ด์„ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[17, 20, 21, 22, 23, 24]. ์ด scaling law๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ ๊ฐœ๋ฐœ์˜ ๊ธฐ๋ณธ์ ์ธ ์—ญํ• ์„ ํ•˜๋ฉฐ, ํ›จ์”ฌ ์ž‘์€ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ๋กœ๊ทธ์—์„œ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ์ตœ์  ๊ตฌ์„ฑ์„ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค[22, 23]. GPT-4 ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ[1]์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, GPT-4 ์„ฑ๋Šฅ์˜ ์ผ๋ถ€ ์ธก๋ฉด์€ GPT-4์˜ 1/1,000๋งŒํผ์˜ ๊ณ„์‚ฐ์œผ๋กœ๋งŒ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. scaling law๋ฅผ ์ ์ ˆํžˆ ํ™œ์šฉํ•˜๋ฉด ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ๋ชจ๋ธ๋ณ„ ํŠœ๋‹์„ ์ˆ˜ํ–‰ํ•  ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ด‘๋ฒ”์œ„ํ•œ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜์™€ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๊ฑฐ๋“ญ์ œ๊ณฑ ๋ฒ•์น™์— ๋”ฐ๋ผ ๋‘๋“œ๋Ÿฌ์ง„ ์„ฑ์žฅ ํŒจํ„ด์„ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[17, 20, 21, 22, 23, 24]. ์ด ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ ๊ฐœ๋ฐœ์˜ ๊ธฐ๋ณธ์ ์ธ ์—ญํ• ์„ ํ•˜๋ฉฐ, ํ›จ์”ฌ ์ž‘์€ ๋ชจ๋ธ์˜ ํ•™์Šต ๋กœ๊ทธ์—์„œ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ์ตœ์  ๊ตฌ์„ฑ์„ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค[22, 23]. GPT-4 ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ[1]์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, GPT-4 ์„ฑ๋Šฅ์˜ ์ผ๋ถ€ ์ธก๋ฉด์€ GPT-4์˜ 1/1,000๋งŒํผ์˜ ๊ณ„์‚ฐ์œผ๋กœ๋งŒ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ์ ์ ˆํžˆ ํ™œ์šฉํ•˜๋ฉด ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ๋ชจ๋ธ๋ณ„ ํŠœ๋‹์„ ์ˆ˜ํ–‰ํ•  ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]A wide range of studies have shown that the performance of a language model exhibits a notable growth pattern as the number of parameters and data size increase, following a power-law relationship [17, 20, 21, 22, 23, 24]. This scaling law plays a fundamental role in the development of large language models, enabling us to estimate optimal configurations of large models from the training logs of much smaller models [22, 23]. As mentioned in the GPT-4 technical report [1], some aspects of GPT-4's performance can be accurately predicted based on models trained with no more than 1/1,000th the compute of GPT-4. By properly utilizing the scaling law, we avoid the need to perform extensive model-specific tuning on large models. [|assistant|]
์ด ์„น์…˜์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋ ˆ์ด์–ด์™€ ๊ตฌ์„ฑ ์š”์†Œ์—์„œ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์„ ์ฒด๊ณ„์ ์œผ๋กœ ๋น„ํ™œ์„ฑํ™”ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•˜์—ฌ ๊ทธ๋ฆผ 2์— ์ œ์‹œ๋œ workflow๋ฅผ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ด ์„น์…˜์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋ ˆ์ด์–ด์™€ ๊ตฌ์„ฑ ์š”์†Œ์—์„œ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์„ ์ฒด๊ณ„์ ์œผ๋กœ ๋น„ํ™œ์„ฑํ™”ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•˜์—ฌ ๊ทธ๋ฆผ 2์— ์ œ์‹œ๋œ ์›Œํฌํ”Œ๋กœ์šฐ๋ฅผ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this section, we validate the workflow presented in Figure 2 by systematically deactivating language-specific neurons across different layers and components to examine their impact on a range of tasks. [|assistant|]
๊ทธ๋ฆผ 7: ์ „์ฒด ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ Loss ๋ฐ ๋„๋ฉ”์ธ Loss์— ๋Œ€ํ•œ Loss ์˜ˆ์ธก ํŒŒ์ดํ”„๋ผ์ธ์˜ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ํ”ผํŒ… ๋ฐ์ดํ„ฐ๋Š” 30B ํ† ํฐ์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋œ 70M ~ 410M ๋ชจ๋ธ์˜ ๋ฐ์ดํ„ฐ์ด๋ฉฐ, ์ถ”์ •๋œ ์ ์€ 1B ๋ชจ๋ธ๊ณผ 100B ํ† ํฐ์— ๋Œ€ํ•œ ๊ธฐ๋ณธ ๋ฐ ์ตœ์ ํ™”๋œ ํ˜ผํ•ฉ๋ฌผ์˜ ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 7: ์ „์ฒด ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์†์‹ค ๋ฐ ๋„๋ฉ”์ธ ์†์‹ค์— ๋Œ€ํ•œ ์†์‹ค ์˜ˆ์ธก ํŒŒ์ดํ”„๋ผ์ธ์˜ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ํ”ผํŒ… ๋ฐ์ดํ„ฐ๋Š” 30B ํ† ํฐ์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋œ 70M ~ 410M ๋ชจ๋ธ์˜ ๋ฐ์ดํ„ฐ์ด๋ฉฐ, ์ถ”์ •๋œ ์ ์€ 1B ๋ชจ๋ธ๊ณผ 100B ํ† ํฐ์— ๋Œ€ํ•œ ๊ธฐ๋ณธ ๋ฐ ์ตœ์ ํ™”๋œ ํ˜ผํ•ฉ๋ฌผ์˜ ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 7: Results of our loss prediction pipelines for the overall validation loss and domain losses. Fitting data are from 70M to 410M models trained for 30B tokens, while the extrapolated points are from the default and optimized mixture for 1B models and 100B tokens. [|assistant|]
์ „๋ฐ˜์ ์œผ๋กœ ์ง€์‹ ์šฉ๋Ÿ‰ scaling laws๋ฅผ ์—ฐ๊ตฌํ•˜๋Š” ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ธํ„ฐ๋„ท ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์„ ์‹ค์ œ ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•˜๋Š” ๊ธฐ์กด ๋ฐฉ๋ฒ•์— ๋น„ํ•ด ๋ณด๋‹ค ์œ ์—ฐํ•˜๊ณ  **๋” ์ •ํ™•ํ•œ playground**๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ •ํ™•์„ฑ์€ ๋ถ€๋ถ„์ ์œผ๋กœ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํ•ฉ์„ฑ์  ํŠน์„ฑ์— ๊ธฐ์ธํ•˜๋ฉฐ, ์ด๋Š” ์‹ค์ œ ๋ฒค์น˜๋งˆํฌ ๊ฒฐ๊ณผ์˜ ์œ ํšจ์„ฑ์„ ์†์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ ์˜ค์—ผ์— ๋Œ€ํ•œ ์šฐ๋ ค๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋ชจ๋ธ architecture์™€ ์ง€์‹ ์œ ํ˜•์— ๋Œ€ํ•œ ์ฒ ์ €ํ•œ ๋น„๊ต๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ quantization ๋ฐฉ๋ฒ•์„ ํƒ๊ตฌํ•˜์ง€๋Š” ์•Š์•˜์ง€๋งŒ, ์ด๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ junk data์˜ ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•˜๊ณ  ์™„ํ™” ์ „๋žต์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๋Ÿฌํ•œ ์›์น™์ ์ธ ํƒ๊ตฌ๋ฅผ ํ†ตํ•ด ์–ป์€ ์ธ์‚ฌ์ดํŠธ๊ฐ€ ์‹ค๋ฌด์ž๋“ค์ด ๋ชจ๋ธ ์„ ํƒ, ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์ค€๋น„ ๋ฐ LLM์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์ด๋ก ์  ์—ฐ๊ตฌ์— ๋Œ€ํ•ด ์ •๋ณด์— ์ž…๊ฐํ•œ ๊ฒฐ์ •์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ „๋ฐ˜์ ์œผ๋กœ ์ง€์‹ ์šฉ๋Ÿ‰ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ์—ฐ๊ตฌํ•˜๋Š” ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ธํ„ฐ๋„ท ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์„ ์‹ค์ œ ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•˜๋Š” ๊ธฐ์กด ๋ฐฉ๋ฒ•์— ๋น„ํ•ด ๋ณด๋‹ค ์œ ์—ฐํ•˜๊ณ  **๋” ์ •ํ™•ํ•œ ํ”Œ๋ ˆ์ด๊ทธ๋ผ์šด๋“œ**๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ •ํ™•์„ฑ์€ ๋ถ€๋ถ„์ ์œผ๋กœ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํ•ฉ์„ฑ์  ํŠน์„ฑ์— ๊ธฐ์ธํ•˜๋ฉฐ, ์ด๋Š” ์‹ค์ œ ๋ฒค์น˜๋งˆํฌ ๊ฒฐ๊ณผ์˜ ์œ ํšจ์„ฑ์„ ์†์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ ์˜ค์—ผ์— ๋Œ€ํ•œ ์šฐ๋ ค๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์™€ ์ง€์‹ ์œ ํ˜•์— ๋Œ€ํ•œ ์ฒ ์ €ํ•œ ๋น„๊ต๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์–‘์žํ™” ๋ฐฉ๋ฒ•์„ ํƒ๊ตฌํ•˜์ง€๋Š” ์•Š์•˜์ง€๋งŒ, ์ด๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ •ํฌ ๋ฐ์ดํ„ฐ์˜ ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•˜๊ณ  ์™„ํ™” ์ „๋žต์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๋Ÿฌํ•œ ์›์น™์ ์ธ ํƒ๊ตฌ๋ฅผ ํ†ตํ•ด ์–ป์€ ์ธ์‚ฌ์ดํŠธ๊ฐ€ ์‹ค๋ฌด์ž๋“ค์ด ๋ชจ๋ธ ์„ ํƒ, ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์ค€๋น„ ๋ฐ LLM์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์ด๋ก ์  ์—ฐ๊ตฌ์— ๋Œ€ํ•ด ์ •๋ณด์— ์ž…๊ฐํ•œ ๊ฒฐ์ •์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Overall, our approach to studying knowledge capacity scaling laws offers a flexible and **more accurate playground** compared to traditional methods that evaluate language models trained on internet data against real-world benchmarks. This accuracy is partly due to the synthetic nature of our dataset, which eliminates concerns about benchmark contamination that could compromise the validity of real-world benchmark results. In this paper, we've conducted a thorough comparison across different model architectures and types of knowledge. While we haven't explored various quantization methods, this represents a promising direction for future research. We've also investigated the impact of junk data and proposed mitigation strategies. We believe the insights gained from this principled exploration can assist practitioners in making informed decisions about model selection, training data preparation, and further theoretical research into LLMs. [|assistant|]
ํ›จ์”ฌ ๋” ๋งŽ์€ ์ˆ˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ LLM์€ ์ด์ „ PLM๊ณผ ๋‹ฌ๋ฆฌ ๊ด‘๋ฒ”์œ„ํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ํฌ๊ด„ํ•˜๋Š” ๋” ๋งŽ์€ ์–‘์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ํ•„์š”๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•„์š”์„ฑ์— ๋”ฐ๋ผ ์—ฐ๊ตฌ์šฉ์œผ๋กœ ๊ณต๊ฐœ๋œ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ ์  ๋” ๋งŽ์•„์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” LLM ํ•™์Šต์— ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋ช‡ ๊ฐ€์ง€ ์ฝ”ํผ์Šค๋ฅผ ๊ฐ„๋žตํ•˜๊ฒŒ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์ฝ˜ํ…์ธ  ์œ ํ˜•์— ๋”ฐ๋ผ ์ด๋Ÿฌํ•œ ์ฝ”ํผ์Šค๋ฅผ ๋„์„œ, CommonCrawl, Reddit ๋งํฌ, Wikipedia, Code ๋ฐ ๊ธฐํƒ€์˜ 6๊ฐ€์ง€ ๊ทธ๋ฃน์œผ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
ํ›จ์”ฌ ๋” ๋งŽ์€ ์ˆ˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ LLM์€ ์ด์ „ PLM๊ณผ ๋‹ฌ๋ฆฌ ๊ด‘๋ฒ”์œ„ํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ํฌ๊ด„ํ•˜๋Š” ๋” ๋งŽ์€ ์–‘์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ํ•„์š”๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•„์š”์„ฑ์— ๋”ฐ๋ผ ์—ฐ๊ตฌ์šฉ์œผ๋กœ ๊ณต๊ฐœ๋œ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ ์  ๋” ๋งŽ์•„์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” LLM ํ•™์Šต์— ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋ช‡ ๊ฐ€์ง€ ์ฝ”ํผ์Šค๋ฅผ ๊ฐ„๋žตํ•˜๊ฒŒ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์ฝ˜ํ…์ธ  ์œ ํ˜•์— ๋”ฐ๋ผ ์ด๋Ÿฌํ•œ ์ฝ”ํผ์Šค๋ฅผ ๋„์„œ, CommonCrawl, Reddit ๋งํฌ, Wikipedia, ์ฝ”๋“œ ๋ฐ ๊ธฐํƒ€์˜ 6๊ฐ€์ง€ ๊ทธ๋ฃน์œผ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In contrast to earlier PLMs, LLMs which consist of a significantly larger number of parameters require a higher volume of training data that covers a broad range of content. For this need, there are increasingly more accessible training datasets that have been released for research. In this section, we will briefly summarize several widely used corpora for training LLMs. Based on their content types, we categorize these corpora into six groups: Books, CommonCrawl, Reddit links, Wikipedia, Code, and others. [|assistant|]
\(\bullet\)_๋ช…๋ น์–ด ์ˆ˜๋ฅผ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค._ ์œ„์˜ ์ธก๋ฉด ์™ธ์—๋„ ๋ช…๋ น์–ด ์ˆ˜๋„ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ์ค‘์š”ํ•œ ์š”์†Œ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ, ๋” ๋งŽ์€ instructions๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ž‘์—… ์ง€์‹์„ ํ™•์žฅํ•˜๊ณ  LLM์˜ instruction ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[69]. ์ด ์ „๋žต์„ ๊ฒ€ํ† ํ•˜๊ธฐ ์œ„ํ•ด MOSS ํ”„๋กœ์ ํŠธ[362]์—์„œ ๊ณต๊ฐœ๋œ ํ•ฉ์„ฑ instruction ์„ธํŠธ์—์„œ ์ƒˆ๋กœ์šด instructions๋ฅผ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค[143]. ์ด๋ฅผ Self-Instruct-52K ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ํ˜ผํ•ฉํ•˜์—ฌ 220K๊ฐœ์˜ instructions๊ฐ€ ํฌํ•จ๋œ ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_๋ช…๋ น์–ด ์ˆ˜๋ฅผ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค._ ์œ„์˜ ์ธก๋ฉด ์™ธ์—๋„ ๋ช…๋ น์–ด ์ˆ˜๋„ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ์ค‘์š”ํ•œ ์š”์†Œ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ, ๋” ๋งŽ์€ ๋ช…๋ น์–ด๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ž‘์—… ์ง€์‹์„ ํ™•์žฅํ•˜๊ณ  LLM์˜ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[69]. ์ด ์ „๋žต์„ ๊ฒ€ํ† ํ•˜๊ธฐ ์œ„ํ•ด MOSS ํ”„๋กœ์ ํŠธ[362]์—์„œ ๊ณต๊ฐœ๋œ ํ•ฉ์„ฑ ๋ช…๋ น์–ด ์„ธํŠธ์—์„œ ์ƒˆ๋กœ์šด ๋ช…๋ น์–ด๋ฅผ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค[143]. ์ด๋ฅผ Self-Instruct-52K ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ํ˜ผํ•ฉํ•˜์—ฌ 220K๊ฐœ์˜ ๋ช…๋ น์–ด๊ฐ€ ํฌํ•จ๋œ ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_Scaling the instruction number._ In addition to the above aspects, the number of instructions is also an important factor that may affect the model performance. Specially, using more instructions can extend the task knowledge and improve the ability of instruction following for LLMs [69]. To examine this strategy, we sample new instructions from the synthesized instruction set released from the MOSS project [362], as they are also synthesized using the same self-instruct method [143]. We mix them with the Self-Instruct-52K dataset to compose a larger one containing 220K instructions. [|assistant|]
ํ™˜๊ฐ ์™„ํ™”. ๋งŽ์€ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์–ธ์–ด ๋ชจ๋ธ์€ ์‚ฌ์šฉ์ž์™€์˜ ์ƒํ˜ธ์ž‘์šฉ์„ ์กฐ์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด Supervised Fine-tuning (SFT)์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค(Wang et al., 2022b; Zhang et al., 2023b). ํŠนํžˆ ์ถ”๋ก  ๋ฐ ์ •๋ ฌ๊ณผ ๊ฐ™์€ ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ํ•ฉ์„ฑ SFT ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ๋งŽ์ด ์žˆ์Šต๋‹ˆ๋‹ค(Wei et al., 2023a,b). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๋น„ ์‚ฌ์†Œํ•œ ์–‘์˜ ํ™˜๊ฐ ๋‹ต๋ณ€์„ ํฌํ•จํ•˜๊ฑฐ๋‚˜ ๋ชจ๋ธ์ด ๋‹ต์„ ๋ชจ๋ฅด๋Š” ์งˆ๋ฌธ์— ๋‹ตํ•˜๋„๋ก ํ•™์Šตํ•˜๋„๋ก ๊ฐ•์ œํ•จ์œผ๋กœ์จ ์–ธ์–ด ๋ชจ๋ธ์— ํ™˜๊ฐ์„ ์œ ๋„ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Zhang et al., 2023c). ์ด๋Ÿฌํ•œ ์‚ฌ๋ก€๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ์ ์šฉ๋˜์ง€ ์•Š์œผ๋ฉด ์–ธ์–ด ๋ชจ๋ธ์—์„œ ํ™˜๊ฐ์„ ์‹ค์ œ๋กœ ์ฆ๊ฐ€์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
ํ™˜๊ฐ ์™„ํ™”.๋งŽ์€ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์–ธ์–ด ๋ชจ๋ธ์€ ์‚ฌ์šฉ์ž์™€์˜ ์ƒํ˜ธ์ž‘์šฉ์„ ์กฐ์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •(SFT)์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค(Wang et al., 2022b; Zhang et al., 2023b). ํŠนํžˆ ์ถ”๋ก  ๋ฐ ์ •๋ ฌ๊ณผ ๊ฐ™์€ ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ํ•ฉ์„ฑ SFT ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ๋งŽ์ด ์žˆ์Šต๋‹ˆ๋‹ค(Wei et al., 2023a,b). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๋น„ ์‚ฌ์†Œํ•œ ์–‘์˜ ํ™˜๊ฐ ๋‹ต๋ณ€์„ ํฌํ•จํ•˜๊ฑฐ๋‚˜ ๋ชจ๋ธ์ด ๋‹ต์„ ๋ชจ๋ฅด๋Š” ์งˆ๋ฌธ์— ๋‹ตํ•˜๋„๋ก ํ•™์Šตํ•˜๋„๋ก ๊ฐ•์ œํ•จ์œผ๋กœ์จ ์–ธ์–ด ๋ชจ๋ธ์— ํ™˜๊ฐ์„ ์œ ๋„ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Zhang et al., 2023c). ์ด๋Ÿฌํ•œ ์‚ฌ๋ก€๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ์ ์šฉ๋˜์ง€ ์•Š์œผ๋ฉด ์–ธ์–ด ๋ชจ๋ธ์—์„œ ํ™˜๊ฐ์„ ์‹ค์ œ๋กœ ์ฆ๊ฐ€์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Mitigating hallucination.Many widely-used language models utilize supervised finetuning (SFT) to learn to align their interactions with users (Wang et al., 2022b; Zhang et al., 2023b). In particular, there exist many methods of generating synthetic SFT data that can improve capabilities such as reasoning and alignment (Wei et al., 2023a,b). It has been shown, however, that these synthetic data can induce hallucinations into language models by containing nontrivial amounts of hallucinated answers or by forcing models to learn to answer questions that they do not know the answer to (Zhang et al., 2023c). These cases demonstrate that synthetic data, if not applied correctly, can actually increase hallucinations in language models. [|assistant|]
๋˜ํ•œ ์ผ๋ถ€ ์—ฐ๊ตฌ์ž๋“ค์€ [312] ๋ฐ˜๋ณต์ ์ธ ์ƒ์„ฑ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ด์ „์— ์ƒ์„ฑ๋œ ํ† ํฐ ๋˜๋Š” \(n\)-๊ทธ๋žจ์˜ ์ƒ์„ฑ์— ๋ถˆ์ด์ต์„ ์ฃผ๋Š” ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ beam search [313]์„ ํ™œ์šฉํ•˜์—ฌ ๋™์ผํ•œ ์ž…๋ ฅ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค์–‘ํ•œ ์ถœ๋ ฅ ์ง‘ํ•ฉ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋˜ํ•œ ์ผ๋ถ€ ์—ฐ๊ตฌ์ž๋“ค์€ [312] ๋ฐ˜๋ณต์ ์ธ ์ƒ์„ฑ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ด์ „์— ์ƒ์„ฑ๋œ ํ† ํฐ ๋˜๋Š” \(n\)-๊ทธ๋žจ์˜ ์ƒ์„ฑ์— ๋ถˆ์ด์ต์„ ์ฃผ๋Š” ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ ๋น” ๊ฒ€์ƒ‰ [313]์„ ํ™œ์šฉํ•˜์—ฌ ๋™์ผํ•œ ์ž…๋ ฅ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค์–‘ํ•œ ์ถœ๋ ฅ ์ง‘ํ•ฉ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Besides, some researchers [312] propose to penalize the generation of previously generated tokens or \(n\)-grams to alleviate the issue of repetitive generation. In addition, diverse beam search [313] can be leveraged to produce a set of diverse outputs based on the same input. [|assistant|]
11์ชฝ: ๊ฐ ์งˆ๋ฌธ์— ๋Œ€ํ•ด GPT-4์—๊ฒŒ 10๋ฒˆ promptํ•˜๊ณ  ํ•œ ๋ฒˆ์ด๋ผ๋„ ์ •๋‹ต์„ ๋งžํžˆ๋ฉด ์„ฑ๊ณต์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์„ฑ๋Šฅ์€ ์‚ฌ์šฉ๋œ prompt์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. prompt๋ฅผ ์•ฝ๊ฐ„๋งŒ ๋ณ€๊ฒฝํ•ด๋„ ๋ชจ๋ธ์˜ ์ •ํ™•๋„๊ฐ€ ๋†’์•„์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
11์ชฝ: ๊ฐ ์งˆ๋ฌธ์— ๋Œ€ํ•ด GPT-4์—๊ฒŒ 10๋ฒˆ ํ”„๋กฌํ”„ํŠธํ•˜๊ณ  ํ•œ ๋ฒˆ์ด๋ผ๋„ ์ •๋‹ต์„ ๋งžํžˆ๋ฉด ์„ฑ๊ณต์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์„ฑ๋Šฅ์€ ์‚ฌ์šฉ๋œ ํ”„๋กฌํ”„ํŠธ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ๋ฅผ ์•ฝ๊ฐ„๋งŒ ๋ณ€๊ฒฝํ•ด๋„ ๋ชจ๋ธ์˜ ์ •ํ™•๋„๊ฐ€ ๋†’์•„์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Footnote 11: We prompt GPT-4 10 times for each question and count it as a success if it answers the question correctly at least once. Performance seems to depend on the prompt used. Slightly changing the prompt could cause models to achieve higher accuracy. [|assistant|]
\(\bullet\)_Model Feedback._ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์™ธ๋ถ€ ์‹ ํ˜ธ์ธ ํ™˜๊ฒฝ ๋ฐ ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ ์™ธ์—๋„ ์—ฐ๊ตฌ์ž๋“ค์€ ์—์ด์ „ํŠธ ์ž์ฒด์˜ ๋‚ด๋ถ€ ํ”ผ๋“œ๋ฐฑ์„ ํ™œ์šฉํ•˜๋Š” ๋ฐฉ์•ˆ๋„ ์—ฐ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ํ”ผ๋“œ๋ฐฑ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํŠนํžˆ [62]๋Š” self-refine ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ์ถœ๋ ฅ, ํ”ผ๋“œ๋ฐฑ, ๊ฐœ์„ ์ด๋ผ๋Š” ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์—์ด์ „ํŠธ๊ฐ€ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ LLMs์„ ํ™œ์šฉํ•˜์—ฌ ์ถœ๋ ฅ์— ๋Œ€ํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ณ  ๊ฐœ์„  ๋ฐฉ๋ฒ•์„ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ํ”ผ๋“œ๋ฐฑ๊ณผ ๊ฐœ์„ ์„ ํ†ตํ•ด ์ถœ๋ ฅ์ด ๊ฐœ์„ ๋ฉ๋‹ˆ๋‹ค. ์ด ์ถœ๋ ฅ-ํ”ผ๋“œ๋ฐฑ-๊ฐœ์„  ํ”„๋กœ์„ธ์Šค๋Š” ์›ํ•˜๋Š” ์กฐ๊ฑด์— ๋„๋‹ฌํ•  ๋•Œ๊นŒ์ง€ ๋ฐ˜๋ณต๋ฉ๋‹ˆ๋‹ค. SelfCheck [63]๋ฅผ ํ†ตํ•ด ์—์ด์ „ํŠธ๋Š” ๋‹ค์–‘ํ•œ ๋‹จ๊ณ„์—์„œ ์ƒ์„ฑ๋œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๊ฒ€ํ† ํ•˜๊ณ  ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•˜์—ฌ ์˜ค๋ฅ˜๋ฅผ ์ˆ˜์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. InterAct [64]๋Š” ๋‹ค๋ฅธ language models (์˜ˆ: ChatGPT ๋ฐ InstructGPT)์„ ๊ฒ€์‚ฌ์ž ๋ฐ ์ •๋ ฌ์ž์™€ ๊ฐ™์€ ๋ณด์กฐ ์—ญํ• ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ์ฃผ ์–ธ์–ด ๋ชจ๋ธ์ด ์ž˜๋ชป๋œ ๋™์ž‘๊ณผ ๋น„ํšจ์œจ์ ์ธ ๋™์ž‘์„ ํ”ผํ•  ์ˆ˜ ์žˆ๋„๋ก ๋•์Šต๋‹ˆ๋‹ค. ChatCoT [65]๋Š” model feedback์„ ํ™œ์šฉํ•˜์—ฌ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค์˜ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. model feedback์€ ์—์ด์ „ํŠธ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๋Š” ํ‰๊ฐ€ ๋ชจ๋“ˆ์—์„œ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. Reflexion [12]์€ ์ƒ์„ธํ•œ ์–ธ์–ด ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ด ์—์ด์ „ํŠธ์˜ ๊ณ„ํš ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์—์„œ ์—์ด์ „ํŠธ๋Š” ๋จผ์ € ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์•ก์…˜์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ, ํ‰๊ฐ€๊ธฐ๋Š” ์—์ด์ „ํŠธ ๊ถค์ ์„ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ํ”ผ๋“œ๋ฐฑ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํ”ผ๋“œ๋ฐฑ์ด ์Šค์นผ๋ผ ๊ฐ’์œผ๋กœ ์ œ๊ณต๋˜๋Š” ๊ธฐ์กด ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ, ์ด ๋ชจ๋ธ์€ LLMs์„ ํ™œ์šฉํ•˜์—ฌ ๋ณด๋‹ค ์ƒ์„ธํ•œ ์–ธ์–ด ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•จ์œผ๋กœ์จ ์—์ด์ „ํŠธ ๊ณ„ํš์— ๋Œ€ํ•œ ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ์ง€์›์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_๋ชจ๋ธ ํ”ผ๋“œ๋ฐฑ._ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์™ธ๋ถ€ ์‹ ํ˜ธ์ธ ํ™˜๊ฒฝ ๋ฐ ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ ์™ธ์—๋„ ์—ฐ๊ตฌ์ž๋“ค์€ ์—์ด์ „ํŠธ ์ž์ฒด์˜ ๋‚ด๋ถ€ ํ”ผ๋“œ๋ฐฑ์„ ํ™œ์šฉํ•˜๋Š” ๋ฐฉ์•ˆ๋„ ์—ฐ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ํ”ผ๋“œ๋ฐฑ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํŠนํžˆ [62]๋Š” ์ž์ฒด ๊ฐœ์„  ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ์ถœ๋ ฅ, ํ”ผ๋“œ๋ฐฑ, ๊ฐœ์„ ์ด๋ผ๋Š” ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์—์ด์ „ํŠธ๊ฐ€ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ LLM์„ ํ™œ์šฉํ•˜์—ฌ ์ถœ๋ ฅ์— ๋Œ€ํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ณ  ๊ฐœ์„  ๋ฐฉ๋ฒ•์„ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ํ”ผ๋“œ๋ฐฑ๊ณผ ๊ฐœ์„ ์„ ํ†ตํ•ด ์ถœ๋ ฅ์ด ๊ฐœ์„ ๋ฉ๋‹ˆ๋‹ค. ์ด ์ถœ๋ ฅ-ํ”ผ๋“œ๋ฐฑ-๊ฐœ์„  ํ”„๋กœ์„ธ์Šค๋Š” ์›ํ•˜๋Š” ์กฐ๊ฑด์— ๋„๋‹ฌํ•  ๋•Œ๊นŒ์ง€ ๋ฐ˜๋ณต๋ฉ๋‹ˆ๋‹ค. ์…€ํ”„์ฒดํฌ[63]๋ฅผ ํ†ตํ•ด ์—์ด์ „ํŠธ๋Š” ๋‹ค์–‘ํ•œ ๋‹จ๊ณ„์—์„œ ์ƒ์„ฑ๋œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๊ฒ€ํ† ํ•˜๊ณ  ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•˜์—ฌ ์˜ค๋ฅ˜๋ฅผ ์ˆ˜์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. InterAct [64]๋Š” ๋‹ค๋ฅธ ์–ธ์–ด ๋ชจ๋ธ(์˜ˆ: ChatGPT ๋ฐ InstructGPT)์„ ๊ฒ€์‚ฌ์ž ๋ฐ ์ •๋ ฌ์ž์™€ ๊ฐ™์€ ๋ณด์กฐ ์—ญํ• ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ์ฃผ ์–ธ์–ด ๋ชจ๋ธ์ด ์ž˜๋ชป๋œ ๋™์ž‘๊ณผ ๋น„ํšจ์œจ์ ์ธ ๋™์ž‘์„ ํ”ผํ•  ์ˆ˜ ์žˆ๋„๋ก ๋•์Šต๋‹ˆ๋‹ค. ChatCoT [65]๋Š” ๋ชจ๋ธ ํ”ผ๋“œ๋ฐฑ์„ ํ™œ์šฉํ•˜์—ฌ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค์˜ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ ํ”ผ๋“œ๋ฐฑ์€ ์—์ด์ „ํŠธ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๋Š” ํ‰๊ฐ€ ๋ชจ๋“ˆ์—์„œ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. Reflexion [12]์€ ์ƒ์„ธํ•œ ์–ธ์–ด ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ด ์—์ด์ „ํŠธ์˜ ๊ณ„ํš ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์—์„œ ์—์ด์ „ํŠธ๋Š” ๋จผ์ € ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์•ก์…˜์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ, ํ‰๊ฐ€๊ธฐ๋Š” ์—์ด์ „ํŠธ ๊ถค์ ์„ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ํ”ผ๋“œ๋ฐฑ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํ”ผ๋“œ๋ฐฑ์ด ์Šค์นผ๋ผ ๊ฐ’์œผ๋กœ ์ œ๊ณต๋˜๋Š” ๊ธฐ์กด ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ, ์ด ๋ชจ๋ธ์€ LLM์„ ํ™œ์šฉํ•˜์—ฌ ๋ณด๋‹ค ์ƒ์„ธํ•œ ์–ธ์–ด ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•จ์œผ๋กœ์จ ์—์ด์ „ํŠธ ๊ณ„ํš์— ๋Œ€ํ•œ ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ์ง€์›์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_Model Feedback._ Apart from the aforementioned environmental and human feedback, which are external signals, researchers have also investigated the utilization of internal feedback from the agents themselves. This type of feedback is usually generated based on pre-trained models. In specific, [62] proposes a self-refine mechanism. This mechanism consists of three crucial components: output, feedback, and refinement. Firstly, the agent generates an output. Then, it utilizes LLMs to provide feedback on the output and offer guidance on how to refine it. At last, the output is improved by the feedback and refinement. This output-feedback-refinement process iterates until reaching some desired conditions. SelfCheck [63] allows agents to examine and evaluate their reasoning steps generated at various stages. They can then correct any errors by comparing the outcomes. InterAct [64] uses different language models (such as ChatGPT and InstructGPT) as auxiliary roles, such as checkers and sorters, to help the main language model avoid erroneous and inefficient actions. ChatCoT [65] utilizes model feedback to improve the quality of its reasoning process. The model feedback is generated by an evaluation module that monitors the agent reasoning steps. Reflexion [12] is developed to enhance the agent's planning capability through detailed verbal feedback. In this model, the agent first produces an action based on its memory, and then, the evaluator generates feedback by taking the agent trajectory as input. In contrast to previous studies, where the feedback is given as a scalar value, this model leverages LLMs to provide more detailed verbal feedback, which can provide more comprehensive supports for the agent plans. [|assistant|]
\(\bullet\)_GPT-1_. 2017๋…„์— Google์—์„œ Transformer ๋ชจ๋ธ[22]์„ ์†Œ๊ฐœํ–ˆ๊ณ , OpenAI ํŒ€์€ ์ด ์ƒˆ๋กœ์šด ์‹ ๊ฒฝ๋ง ์•„ํ‚คํ…์ฒ˜์— ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…์„ ๋น ๋ฅด๊ฒŒ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. 2018๋…„์— ์ฒซ ๋ฒˆ์งธ GPT ๋ชจ๋ธ, ์ฆ‰ GPT-1[122]์„ ์ถœ์‹œํ–ˆ์œผ๋ฉฐ, ์ƒ์„ฑ์  ์‚ฌ์ „ ํ•™์Šต์˜ ์•ฝ์ž์ธ _GPT_๋ผ๋Š” ์•ฝ์–ด๋ฅผ ๋ชจ๋ธ ์ด๋ฆ„์œผ๋กœ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-1์€ ์ƒ์„ฑ์ , decoder-only Transformer ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์œผ๋ฉฐ, ๋น„์ง€๋„ ์‚ฌ์ „ ํ•™์Šต๊ณผ supervised Fine-tuning์˜ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ฑ„ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-1์€ GPT ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์˜ ํ•ต์‹ฌ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ตฌ์ถ•ํ•˜๊ณ  ์ž์—ฐ์–ด ํ…์ŠคํŠธ๋ฅผ ๋ชจ๋ธ๋งํ•˜๋Š” ๊ธฐ๋ณธ ์›๋ฆฌ, ์ฆ‰ ๋‹ค์Œ ๋‹จ์–ด ์˜ˆ์ธก์„ ํ™•๋ฆฝํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_GPT-1_. 2017๋…„์— Google์—์„œ ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ[22]์„ ์†Œ๊ฐœํ–ˆ๊ณ , OpenAI ํŒ€์€ ์ด ์ƒˆ๋กœ์šด ์‹ ๊ฒฝ๋ง ์•„ํ‚คํ…์ฒ˜์— ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…์„ ๋น ๋ฅด๊ฒŒ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. 2018๋…„์— ์ฒซ ๋ฒˆ์งธ GPT ๋ชจ๋ธ, ์ฆ‰ GPT-1[122]์„ ์ถœ์‹œํ–ˆ์œผ๋ฉฐ, ์ƒ์„ฑ์  ์‚ฌ์ „ ํ•™์Šต์˜ ์•ฝ์ž์ธ _GPT_๋ผ๋Š” ์•ฝ์–ด๋ฅผ ๋ชจ๋ธ ์ด๋ฆ„์œผ๋กœ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-1์€ ์ƒ์„ฑ์  ๋””์ฝ”๋” ์ „์šฉ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์œผ๋ฉฐ, ๋น„์ง€๋„ ์‚ฌ์ „ ํ•™์Šต๊ณผ ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •์˜ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ฑ„ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-1์€ GPT ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์˜ ํ•ต์‹ฌ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ตฌ์ถ•ํ•˜๊ณ  ์ž์—ฐ์–ด ํ…์ŠคํŠธ๋ฅผ ๋ชจ๋ธ๋งํ•˜๋Š” ๊ธฐ๋ณธ ์›๋ฆฌ, ์ฆ‰ ๋‹ค์Œ ๋‹จ์–ด ์˜ˆ์ธก์„ ํ™•๋ฆฝํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_GPT-1_. In 2017, the Transformer model [22] was introduced by Google, and the OpenAI team quickly adapted their language modeling work to this new neural network architecture. They released the first GPT model in 2018, _i.e._, GPT-1 [122], and coined the abbreviation term _GPT_ as the model name, standing for _Generative Pre-Training_. GPT-1 was developed based on a generative, decoder-only Transformer architecture, and adopted a hybrid approach of unsupervised pretraining and supervised fine-tuning. GPT-1 has set up the core architecture for the GPT-series models and established the underlying principle to model natural language text, _i.e._, predicting the next word. [|assistant|]
2023๋…„ 3์›”์— OpenAI์—์„œ ์ถœ์‹œํ•œ CODEX[57]๋Š” ์ž์—ฐ์–ด๋ฅผ ๊ตฌ๋ฌธ ๋ถ„์„ํ•˜๊ณ  ์ด์— ๋Œ€ํ•œ ์‘๋‹ต์œผ๋กœ ์ฝ”๋“œ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋ฒ”์šฉ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. CODEX๋Š” GPT-3์˜ ํ›„์†์œผ๋กœ, GitHub์—์„œ ์ˆ˜์ง‘ํ•œ ์ฝ”๋“œ ์ฝ”ํผ์Šค์—์„œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋งž๊ฒŒ Fine-tuning๋˜์—ˆ์Šต๋‹ˆ๋‹ค. CODEX๋Š” Microsoft์˜ GitHub Copilot์˜ ํ•ต์‹ฌ์ž…๋‹ˆ๋‹ค.[|endofturn|]
2023๋…„ 3์›”์— OpenAI์—์„œ ์ถœ์‹œํ•œ CODEX[57]๋Š” ์ž์—ฐ์–ด๋ฅผ ๊ตฌ๋ฌธ ๋ถ„์„ํ•˜๊ณ  ์ด์— ๋Œ€ํ•œ ์‘๋‹ต์œผ๋กœ ์ฝ”๋“œ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋ฒ”์šฉ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. CODEX๋Š” GPT-3์˜ ํ›„์†์œผ๋กœ, GitHub์—์„œ ์ˆ˜์ง‘ํ•œ ์ฝ”๋“œ ์ฝ”ํผ์Šค์—์„œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋งž๊ฒŒ ๋ฏธ์„ธ ์กฐ์ •๋˜์—ˆ์Šต๋‹ˆ๋‹ค. CODEX๋Š” Microsoft์˜ GitHub Copilot์˜ ํ•ต์‹ฌ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]CODEX [57], released by OpenAI in March 2023, is a general-purpose programming model that can parse natural language and generate code in response. CODEX is a descendant of GPT-3, fine-tuned for programming applications on code corpora collected from GitHub. CODEX powers Microsoft's GitHub Copilot. [|assistant|]
**Empirical Hutter++ Model.** ๊ทธ๋ฆผ 6์—์„œ๋Š” ์›๋ž˜ ๋ถ„ํฌ์—์„œ \(T_{0}=100,000\) ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด ํ•™์Šต๋œ ์ดˆ๊ธฐ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Gen 1 ๋ผ์ธ์˜ ๊ฒฝ์šฐ, ๋ฐ์ดํ„ฐ๋Š” ๋ชจ๋‘ ์ด ์ดˆ๊ธฐ ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. 2์„ธ๋Œ€๋ถ€ํ„ฐ๋Š” ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•œ ๋ชจ๋ธ์ด ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋ฐ˜๋ณต์ ์œผ๋กœ ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผœ ๋ชจ๋ธ ๋ถ•๊ดด๊ฐ€ ๋ถˆ์ถฉ๋ถ„ํ•œ ์ƒ˜ํ”Œ๋ง์œผ๋กœ ์ธํ•ด ๋ฐœ์ƒํ•  ๊ฐ€๋Šฅ์„ฑ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. Gen 1์˜ ๊ฒฝ์šฐ, ๋ฐ์ดํ„ฐ ํ™•์žฅ์„ฑ์ด ๋ˆˆ์— ๋„๊ฒŒ ์ €ํ•˜๋˜๋Š” ๋™์‹œ์— ์—ฌ๋Ÿฌ ์„ธ๋Œ€์— ๊ฑธ์ณ ๋ชจ๋ธ ์„ฑ๋Šฅ์ด ๊ธ‰๊ฒฉํžˆ ํ•˜๋ฝํ•˜๋Š” ๊ฒƒ์ด ๊ด€์ฐฐ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ์ด๋ก ์  ๊ฒฐ๊ณผ๋ฅผ ๊ฒ€์ฆํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๊ฐ€์ •์„ ์žฌํ™•์ธํ•˜๋Š” ๊ฒฐ๊ณผ์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 16์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด temperature scaling์—์„œ๋„ ๋น„์Šทํ•œ ํŒจํ„ด์ด ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.[|endofturn|]
**๊ฒฝํ—˜์  ํ—ˆํ„ฐ++ ๋ชจ๋ธ.** ๊ทธ๋ฆผ 6์—์„œ๋Š” ์›๋ž˜ ๋ถ„ํฌ์—์„œ \(T_{0}=100,000\) ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด ํ•™์Šต๋œ ์ดˆ๊ธฐ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Gen 1 ๋ผ์ธ์˜ ๊ฒฝ์šฐ, ๋ฐ์ดํ„ฐ๋Š” ๋ชจ๋‘ ์ด ์ดˆ๊ธฐ ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. 2์„ธ๋Œ€๋ถ€ํ„ฐ๋Š” ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•œ ๋ชจ๋ธ์ด ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋ฐ˜๋ณต์ ์œผ๋กœ ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผœ ๋ชจ๋ธ ๋ถ•๊ดด๊ฐ€ ๋ถˆ์ถฉ๋ถ„ํ•œ ์ƒ˜ํ”Œ๋ง์œผ๋กœ ์ธํ•ด ๋ฐœ์ƒํ•  ๊ฐ€๋Šฅ์„ฑ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. Gen 1์˜ ๊ฒฝ์šฐ, ๋ฐ์ดํ„ฐ ํ™•์žฅ์„ฑ์ด ๋ˆˆ์— ๋„๊ฒŒ ์ €ํ•˜๋˜๋Š” ๋™์‹œ์— ์—ฌ๋Ÿฌ ์„ธ๋Œ€์— ๊ฑธ์ณ ๋ชจ๋ธ ์„ฑ๋Šฅ์ด ๊ธ‰๊ฒฉํžˆ ํ•˜๋ฝํ•˜๋Š” ๊ฒƒ์ด ๊ด€์ฐฐ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ์ด๋ก ์  ๊ฒฐ๊ณผ๋ฅผ ๊ฒ€์ฆํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๊ฐ€์ •์„ ์žฌํ™•์ธํ•˜๋Š” ๊ฒฐ๊ณผ์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 16์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ์˜จ๋„ ์Šค์ผ€์ผ๋ง์—์„œ๋„ ๋น„์Šทํ•œ ํŒจํ„ด์ด ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Empirical Hutter++ Model.** In Figure 6, we use an initial model that is trained on \(T_{0}=100,000\) samples from the original distribution. For the Gen 1 line, the data are all generated from this initial model. From Gen 2 onwards, models are iteratively trained on data produced by the most performant model of the preceding generation, effectively eliminating the possibility that model collapse results from inadequate sampling. For Gen 1, a notable degradation in data scalability is observed, alongside a rapid decline in model performance across generations. These observations not only validate our theoretical result but also reaffirm our assumptions. A similar pattern is evident with temperature scaling, as shown in Figure 16. [|assistant|]
**D4**: ์•ž์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, ํ…œํ”Œ๋ฆฟ์œผ๋กœ ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ ๋˜๋Š” ์˜๋ฏธ์ ์œผ๋กœ ๋งค์šฐ ์ค‘๋ณต๋œ ์ •๋ณด์˜ ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ MinHash์— ์˜ํ•ด ์ œ๊ฑฐ๋˜์ง€ ์•Š๋Š” ์ค‘๋ณต ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ ๋งŽ์ด ๋ฐœ๊ฒฌ๋ฉ๋‹ˆ๋‹ค. ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์˜ ์ด๋Ÿฌํ•œ ์˜์—ญ์€ ๋งค์šฐ ๋ฐ€์ง‘๋˜์–ด ์žˆ๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฉฐ, k-means๊ฐ€ ์ค‘๋ณต๋œ ํ…์ŠคํŠธ์— ๊ท€์ค‘ํ•œ ํด๋Ÿฌ์Šคํ„ฐ ํ• ๋‹น์„ ๋‚ญ๋น„ํ•˜๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํŽธํ–ฅ๋œ ํด๋Ÿฌ์Šคํ„ฐ๋ง์€ ๋” ๋งŽ์€ ์ฃผ์ œ์  ์ผ๊ด€์„ฑ ๋Œ€์‹  ์ค‘๋ณต์œผ๋กœ ์ธํ•ด ๋งŽ์€ ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ์— SSL Prototypes์˜ ํšจ์œจ์„ฑ์—๋„ ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
**D4**: ์•ž์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, ํ…œํ”Œ๋ฆฟ์œผ๋กœ ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ ๋˜๋Š” ์˜๋ฏธ์ ์œผ๋กœ ๋งค์šฐ ์ค‘๋ณต๋œ ์ •๋ณด์˜ ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ MinHash์— ์˜ํ•ด ์ œ๊ฑฐ๋˜์ง€ ์•Š๋Š” ์ค‘๋ณต ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ ๋งŽ์ด ๋ฐœ๊ฒฌ๋ฉ๋‹ˆ๋‹ค. ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์˜ ์ด๋Ÿฌํ•œ ์˜์—ญ์€ ๋งค์šฐ ๋ฐ€์ง‘๋˜์–ด ์žˆ๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฉฐ, k-ํ‰๊ท ์ด ์ค‘๋ณต๋œ ํ…์ŠคํŠธ์— ๊ท€์ค‘ํ•œ ํด๋Ÿฌ์Šคํ„ฐ ํ• ๋‹น์„ ๋‚ญ๋น„ํ•˜๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํŽธํ–ฅ๋œ ํด๋Ÿฌ์Šคํ„ฐ๋ง์€ ๋” ๋งŽ์€ ์ฃผ์ œ์  ์ผ๊ด€์„ฑ ๋Œ€์‹  ์ค‘๋ณต์œผ๋กœ ์ธํ•ด ๋งŽ์€ ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ์— SSL ํ”„๋กœํ† ํƒ€์ž…์˜ ํšจ์œจ์„ฑ์—๋„ ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**D4**: As mentioned previously, we find many instances of duplicate-driven clusters: clusters of templated text or extremely semantically redundant information that are not removed by MinHash. These regions of embedding space tend to be very dense and cause k-means to waste valuable cluster assignments on duplicated text. This biased clustering could also negatively to impact the effectiveness of SSL Prototypes since many clusters will be entirely driven by duplicates instead of more topical coherence. [|assistant|]
RQ3: ๊ณ ํ’ˆ์งˆ์˜ re-phraser๊ฐ€ ์–ผ๋งˆ๋‚˜ ์ค‘์š”ํ•œ๊ฐ€์š”? ์ด๋ฅผ ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด ๋„ค ๊ฐ€์ง€ re-phrasing ๋ชจ๋ธ(T5-base(Raffel et al., 2020), Owen-L8B-chat(Bai et al., 2023a), Mistral-7B-chat(Jiang et al., 2023), Vicuna-13B-chat-v1.3(Chiang et al., 2023))์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 30B tokens์— ๋Œ€ํ•ด 345M ๋ชจ๋ธ์„ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์—์„œ ๋™์ผํ•œ prompt๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. T5-base ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, Vicuna-13b-chat ๋ชจ๋ธ์˜ re-phrase pairs์— ๋Œ€ํ•ด 1ํšŒ ๋™์•ˆ ๋ชจ๋ธ์„ Fine-tuningํ•ฉ๋‹ˆ๋‹ค. Qwen-1.8B ๋ฐ Mistral-7B์™€ ๊ฐ™์€ ์ž‘์€ re-phrase ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๋กœ pre-trainingํ•˜๋ฉด Vicuna 13B๋ณด๋‹ค ๋‚ฎ์€ perplexity๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 5). ๋™์‹œ์— Fine-tuning๋œ T5-base ๋ชจ๋ธ์€ ๋‚˜๋จธ์ง€ ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ๋‚ฎ์€ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋ชจ๋“  rephrase ๋ชจ๋ธ์€ ์‹ค์ œ C4 ๋ฐ์ดํ„ฐ๋งŒ์œผ๋กœ๋„ perplexity๋ฅผ ์ค„์—ฌ์ค๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ **WRAP**์˜ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ๋”์šฑ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋Š” paraphrase ๋ชจ๋ธ์„ ์–ผ๋งˆ๋‚˜ ์ž‘๊ฒŒ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ•œ๊ณ„๋ฅผ ํ…Œ์ŠคํŠธํ•˜๋Š” ๊ฒƒ์€ ์•„์ง ๋ฏธํ•ด๊ฒฐ ๊ณผ์ œ๋กœ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
RQ3: ๊ณ ํ’ˆ์งˆ์˜ ์žฌ๊ตฌ๋ฌธํ™”๊ธฐ๊ฐ€ ์–ผ๋งˆ๋‚˜ ์ค‘์š”ํ•œ๊ฐ€์š”? ์ด๋ฅผ ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด ๋„ค ๊ฐ€์ง€ ์žฌ๊ตฌ๋ฌธํ™” ๋ชจ๋ธ(T5-base(Raffel et al., 2020), Owen-L8B-chat(Bai et al., 2023a), Mistral-7B-chat(Jiang et al., 2023), Vicuna-13B-chat-v1.3(Chiang et al., 2023))์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 30B ํ† ํฐ์— ๋Œ€ํ•ด 345M ๋ชจ๋ธ์„ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์—์„œ ๋™์ผํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. T5-base ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, Vicuna-13b-chat ๋ชจ๋ธ์˜ ์žฌ๊ตฌ๋ฌธ ์Œ์— ๋Œ€ํ•ด 1ํšŒ ๋™์•ˆ ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. Qwen-1.8B ๋ฐ Mistral-7B์™€ ๊ฐ™์€ ์ž‘์€ ์žฌ๊ตฌ๋ฌธ ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋ฉด Vicuna 13B๋ณด๋‹ค ๋‚ฎ์€ ๋‹นํ˜น๊ฐ์„ ์–ป์„ ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 5). ๋™์‹œ์— ๋ฏธ์„ธ ์กฐ์ •๋œ T5-base ๋ชจ๋ธ์€ ๋‚˜๋จธ์ง€ ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ๋‚ฎ์€ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋ชจ๋“  ์žฌ๊ตฌ๋ฌธ ๋ชจ๋ธ์€ ์‹ค์ œ C4 ๋ฐ์ดํ„ฐ๋งŒ์œผ๋กœ๋„ ๋‹นํ˜น๊ฐ์„ ์ค„์—ฌ์ค๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ **WRAP**์˜ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ๋”์šฑ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋Š” ์˜์—ญ ๋ชจ๋ธ์„ ์–ผ๋งˆ๋‚˜ ์ž‘๊ฒŒ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ•œ๊ณ„๋ฅผ ํ…Œ์ŠคํŠธํ•˜๋Š” ๊ฒƒ์€ ์•„์ง ๋ฏธํ•ด๊ฒฐ ๊ณผ์ œ๋กœ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]RQ3: How important is to have a high-quality re-phraser?To answer this, we use data from four distinct re-phrasing models (T5-base (Raffel et al., 2020), Owen-L8B-chat (Bai et al., 2023a), Mistral-7B-chat (Jiang et al., 2023), and Vicuna-13B-chat-v1.3 (Chiang et al., 2023)) and train a 345M model for 30B tokens. We generate data from all models using the same prompt. In case of the T5-base model, we finetune the model for 1 epoch on re-phrase pairs from the Vicuna-13b-chat model. We find that pre-training on data generated by smaller re-phrase models like Qwen-1.8B and Mistral-7B achieve lower perplexity than Vicuna 13B (Figure 5). At the same time, our fine-tuned T5-base model performs significantly worse than the rest. Even then, all rephrase models reduce perplexity over only real C4 data. It remains an open question to test the limits of how small can we train a paraphrase model that can generate high quality synthetic data to further scale the applicability of **WRAP**. [|assistant|]
Jailbreak prompts๋Š” NLP ์ „๋ฌธ๊ฐ€ ํŒจ๋„์ด ์ˆ˜๋™์œผ๋กœ ์ž‘์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. Prompts๋Š” ์ถœ๋ ฅ์˜ ์ฒซ ์ค„์„ ์•ž์— ๋ถ™์—ฌ์„œ ๋˜๋Š” ๋ถ™์—ฌ์„œ ํ…Œ์ŠคํŠธํ–ˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 5์—๋Š” ํ…Œ์ŠคํŠธ๋œ ๋ชจ๋“  jailbreak prompts ๋ชฉ๋ก์ด ๋‚ด๋ฆผ์ฐจ์ˆœ์œผ๋กœ ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๋ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณ„ ์„ฑ๋Šฅ ํ”Œ๋กฏ์€ ๊ทธ๋ฆผ 5์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
Jailbreak ํ”„๋กฌํ”„ํŠธ๋Š” NLP ์ „๋ฌธ๊ฐ€ ํŒจ๋„์ด ์ˆ˜๋™์œผ๋กœ ์ž‘์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ๋Š” ์ถœ๋ ฅ์˜ ์ฒซ ์ค„์„ ์•ž์— ๋ถ™์—ฌ์„œ ๋˜๋Š” ๋ถ™์—ฌ์„œ ํ…Œ์ŠคํŠธํ–ˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 5์—๋Š” ํ…Œ์ŠคํŠธ๋œ ๋ชจ๋“  ํƒˆ์˜ฅ ํ”„๋กฌํ”„ํŠธ ๋ชฉ๋ก์ด ๋‚ด๋ฆผ์ฐจ์ˆœ์œผ๋กœ ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๋ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณ„ ์„ฑ๋Šฅ ํ”Œ๋กฏ์€ ๊ทธ๋ฆผ 5์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Jailbreak prompts were manually written by a panel of NLP experts. Prompts were tested with and without newlines prepended and with and without taking the first line of output. Table 5 contains a list of all jailbreak prompts tested in order of descending effectiveness. A plot of performance by model and dataset is available in Figure 5. [|assistant|]
Coig-PcCOIG-PC ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ค‘๊ตญ์–ด NLP ์—ฐ๊ตฌ๋ฅผ ๋ฐœ์ „์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋œ ํฌ๊ด„์ ์ธ ์ค‘๊ตญ์–ด NLP ์ž‘์—… ๋ชจ์Œ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ชฉํ‘œ๋Š” ์—ฐ๊ตฌ์ž์™€ ๊ฐœ๋ฐœ์ž๊ฐ€ ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ํ’๋ถ€ํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์—ฐ๊ตฌ์ž์™€ ๊ฐœ๋ฐœ์ž๋ฅผ ์œ„ํ•œ ํฌ๊ด„์ ์ธ ๋ฆฌ์†Œ์Šค ์ œํ’ˆ๊ตฐ์„ ์ œ๊ณตํ•˜์—ฌ ํ…์ŠคํŠธ ์ƒ์„ฑ, ์ •๋ณด ์ถ”์ถœ, ๊ฐ์ • ๋ถ„์„, ๊ธฐ๊ณ„ ๋ฒˆ์—ญ ๋“ฑ ๋‹ค์–‘ํ•œ ์˜์—ญ์—์„œ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ๋ฐœ์ „์‹œํ‚ฌ ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” COIG-PC์—์„œ ์ค‘๊ตญ์–ด์™€ ์˜์–ด๋ฅผ ๋ชจ๋‘ ํฌํ•จํ•˜๋Š” 1,413๊ฐœ์˜ ์ž‘์—…์„ ์„ ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ฃผ๋กœ ์ „ํ†ต์ ์ธ NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฐ€์ ธ์˜จ ์ •๋ณด ์ถ”์ถœ, ๋ถ„๋ฅ˜, ์š”์•ฝ ๋“ฑ ํ’ˆ์งˆ ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋Š” 250๊ฐœ์˜ ์ž‘์—…์„ ์ˆ˜๋™์œผ๋กœ ์„ ๋ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. Temperature sampling์„ ํ†ตํ•ด ์ตœ์ข…์ ์œผ๋กœ 3,000๊ฐœ์˜ (Instruction, ์‘๋‹ต) ์Œ์„ ์ƒ˜ํ”Œ๋งํ•œ ํ›„, ํ’ˆ์งˆ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ๋žŒ์ด ์ง์ ‘ ๊ฒ€์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
Coig-PcCOIG-PC ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ค‘๊ตญ์–ด NLP ์—ฐ๊ตฌ๋ฅผ ๋ฐœ์ „์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋œ ํฌ๊ด„์ ์ธ ์ค‘๊ตญ์–ด NLP ์ž‘์—… ๋ชจ์Œ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ชฉํ‘œ๋Š” ์—ฐ๊ตฌ์ž์™€ ๊ฐœ๋ฐœ์ž๊ฐ€ ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ํ’๋ถ€ํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์—ฐ๊ตฌ์ž์™€ ๊ฐœ๋ฐœ์ž๋ฅผ ์œ„ํ•œ ํฌ๊ด„์ ์ธ ๋ฆฌ์†Œ์Šค ์ œํ’ˆ๊ตฐ์„ ์ œ๊ณตํ•˜์—ฌ ํ…์ŠคํŠธ ์ƒ์„ฑ, ์ •๋ณด ์ถ”์ถœ, ๊ฐ์ • ๋ถ„์„, ๊ธฐ๊ณ„ ๋ฒˆ์—ญ ๋“ฑ ๋‹ค์–‘ํ•œ ์˜์—ญ์—์„œ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ๋ฐœ์ „์‹œํ‚ฌ ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” COIG-PC์—์„œ ์ค‘๊ตญ์–ด์™€ ์˜์–ด๋ฅผ ๋ชจ๋‘ ํฌํ•จํ•˜๋Š” 1,413๊ฐœ์˜ ์ž‘์—…์„ ์„ ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ฃผ๋กœ ์ „ํ†ต์ ์ธ NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฐ€์ ธ์˜จ ์ •๋ณด ์ถ”์ถœ, ๋ถ„๋ฅ˜, ์š”์•ฝ ๋“ฑ ํ’ˆ์งˆ ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋Š” 250๊ฐœ์˜ ์ž‘์—…์„ ์ˆ˜๋™์œผ๋กœ ์„ ๋ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜จ๋„ ์ƒ˜ํ”Œ๋ง์„ ํ†ตํ•ด ์ตœ์ข…์ ์œผ๋กœ 3,000๊ฐœ์˜ (์ง€์‹œ์–ด, ์‘๋‹ต) ์Œ์„ ์ƒ˜ํ”Œ๋งํ•œ ํ›„, ํ’ˆ์งˆ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ๋žŒ์ด ์ง์ ‘ ๊ฒ€์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Coig-PcThe COIG-PC Dataset is a comprehensive collection of Chinese NLP tasks, aimed at advancing Chinese NLP research. The goal of this dataset is to provide a rich set of resources to researchers and developers, which they can use to improve the capabilities of language models in handling Chinese text. It offers a comprehensive suite of resources for researchers and developers, facilitating advancements in language model capabilities across various domains including text generation, information extraction, sentiment analysis, and machine translation, etc. Initially, we selected 1,413 tasks involving both Chinese and English languages from COIG-PC. Then, we manually select 250 tasks that meet our quality criteria, including information extraction, classification, summary, and others, primarily sourced from traditional NLP datasets. Through temperature sampling, we eventually sample 3,000 (instruction, response) pairs, which are further verified by human to ensure quality. [|assistant|]
์ด ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ์—์„œ๋Š” LLaMA ๋ชจ๋ธ์˜ ์ค‘๊ตญ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์›๋ž˜ LLaMA์˜ ์ค‘๊ตญ์–ด ์–ดํœ˜๊ฐ€ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ํ•œ๊ณ„๋ฅผ ์ธ์‹ํ•˜๊ณ , 20,000๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ์„ ์ถ”๊ฐ€๋กœ ํ†ตํ•ฉํ•˜์—ฌ ์ค‘๊ตญ์–ด์— ๋Œ€ํ•œ ์ธ์ฝ”๋”ฉ ํšจ์œจ์„ฑ์„ ํฌ๊ฒŒ ๋†’์˜€์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด LLaMA๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ instruction ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•œ Supervised Fine-tuning์„ ์‚ฌ์šฉํ•˜์—ฌ instruction์„ ๋”ฐ๋ฅด๋Š” ๊ธฐ๋Šฅ์ด ํ–ฅ์ƒ๋œ ์ค‘๊ตญ์–ด ์•ŒํŒŒ์นด ๋ชจ๋ธ์„ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ด ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ์—์„œ๋Š” LLaMA ๋ชจ๋ธ์˜ ์ค‘๊ตญ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์›๋ž˜ LLaMA์˜ ์ค‘๊ตญ์–ด ์–ดํœ˜๊ฐ€ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ํ•œ๊ณ„๋ฅผ ์ธ์‹ํ•˜๊ณ , 20,000๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ์„ ์ถ”๊ฐ€๋กœ ํ†ตํ•ฉํ•˜์—ฌ ์ค‘๊ตญ์–ด์— ๋Œ€ํ•œ ์ธ์ฝ”๋”ฉ ํšจ์œจ์„ฑ์„ ํฌ๊ฒŒ ๋†’์˜€์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด LLaMA๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ง€์‹œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•œ ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •์„ ์‚ฌ์šฉํ•˜์—ฌ ์ง€์‹œ๋ฅผ ๋”ฐ๋ฅด๋Š” ๊ธฐ๋Šฅ์ด ํ–ฅ์ƒ๋œ ์ค‘๊ตญ์–ด ์•ŒํŒŒ์นด ๋ชจ๋ธ์„ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this technical report, we have presented an approach to enhance the Chinese understanding and generation capabilities of the LLaMA model. Acknowledging the limitations of the original LLaMA's Chinese vocabulary, we expanded it by incorporating 20K additional Chinese tokens, significantly increasing its encoding efficiency for the Chinese language. Building on the Chinese LLaMA, we employed supervised fine-tuning with instruction data, resulting in Chinese Alpaca models exhibiting improved instruction-following capabilities. [|assistant|]
* ๊ฒฐ๊ณผ 5: 1000ํšŒ ๋…ธ์ถœ ์„ค์ •์—์„œ 2bit/param ์šฉ๋Ÿ‰ ๋น„์œจ์€ ๋ชจ๋“  ๋ชจ๋ธ, ํŠนํžˆ MLP ๋ ˆ์ด์–ด๊ฐ€ ์—†๋Š” ๋ชจ๋ธ๋„ ์ด ๋น„์œจ์— ๊ทผ์ ‘ํ•˜๋Š” **๋ณดํŽธ์ ์ธ ๊ทœ์น™**์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. * ๊ฒฐ๊ณผ 6: 100ํšŒ ๋…ธ์ถœ ์‹œ ์ผ๋ถ€ arch์—์„œ ํ•œ๊ณ„๊ฐ€ ๋‚˜ํƒ€๋‚˜๋Š”๋ฐ, ํŠนํžˆ LLaMA/Mistral์˜ ์šฉ๋Ÿ‰ ๋น„์œจ์€ ์ตœ์ ํ™”๋œ Learning Rate๋ฅผ ์ ์šฉํ•œ ํ›„์—๋„ GPT2๋ณด๋‹ค 1.3๋ฐฐ ๋‚ฎ์Šต๋‹ˆ๋‹ค. * ๊ฒฐ๊ณผ 7: ์ถ”๊ฐ€ ์ œ์–ด ์‹คํ—˜์— ๋”ฐ๋ฅด๋ฉด "gated MLP" ์‚ฌ์šฉ์€ LLaMA/Mistral architecture์˜ ์ง€์‹ ์ €์žฅ ์„ฑ๋Šฅ ์ €ํ•˜๋กœ ์ด์–ด์ง€๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.[|endofturn|]
* ๊ฒฐ๊ณผ 5: 1000ํšŒ ๋…ธ์ถœ ์„ค์ •์—์„œ 2๋น„ํŠธ/ํŒŒ๋ผ๋ฏธํ„ฐ ์šฉ๋Ÿ‰ ๋น„์œจ์€ ๋ชจ๋“  ๋ชจ๋ธ, ํŠนํžˆ MLP ๋ ˆ์ด์–ด๊ฐ€ ์—†๋Š” ๋ชจ๋ธ๋„ ์ด ๋น„์œจ์— ๊ทผ์ ‘ํ•˜๋Š” **๋ณดํŽธ์ ์ธ ๊ทœ์น™**์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. * ๊ฒฐ๊ณผ 6: 100ํšŒ ๋…ธ์ถœ ์‹œ ์ผ๋ถ€ ์•„ํฌ์—์„œ ํ•œ๊ณ„๊ฐ€ ๋‚˜ํƒ€๋‚˜๋Š”๋ฐ, ํŠนํžˆ LLaMA/Mistral์˜ ์šฉ๋Ÿ‰ ๋น„์œจ์€ ์ตœ์ ํ™”๋œ ํ•™์Šต ์†๋„๋ฅผ ์ ์šฉํ•œ ํ›„์—๋„ GPT2๋ณด๋‹ค 1.3๋ฐฐ ๋‚ฎ์Šต๋‹ˆ๋‹ค. * ๊ฒฐ๊ณผ 7: ์ถ”๊ฐ€ ์ œ์–ด ์‹คํ—˜์— ๋”ฐ๋ฅด๋ฉด "๊ฒŒ์ดํŠธํ˜• MLP" ์‚ฌ์šฉ์€ LLaMA/Mistral ์•„ํ‚คํ…์ฒ˜์˜ ์ง€์‹ ์ €์žฅ ์„ฑ๋Šฅ ์ €ํ•˜๋กœ ์ด์–ด์ง€๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* Result 5: In the 1000-exposure setting, a 2bit/param capacity ratio appears to be a **universal rule**: all models, even without MLP layers, closely achieve this ratio. * Result 6: With 100 exposures, some archs show limitations; notably, LLaMA/Mistral's capacity ratio is 1.3x lower than GPT2's, even after best-tuned learning rates. * Result 7: Further controlled experiments indicate that "gated MLP" usage leads to LLaMA/Mistral architecture's underperformance in knowledge storage. [|assistant|]
**Action Production**: ๋ชจ๋ธ ์ž…๋ ฅ๊ณผ ์ถœ๋ ฅ์ด ์ง์ ‘ ์—ฐ๊ด€๋˜์–ด ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ LLM๊ณผ ๋‹ฌ๋ฆฌ, ์—์ด์ „ํŠธ๋Š” ๋‹ค์–‘ํ•œ ์ „๋žต๊ณผ ์†Œ์Šค๋ฅผ ํ†ตํ•ด actions์„ ์ทจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์˜ action ์ƒ์„ฑ ์ „๋žต์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. (1) _Action via Memory Recollection_. ์ด ์ „๋žต์—์„œ๋Š” ํ˜„์žฌ ์ž‘์—…์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•˜์—ฌ action์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ž‘์—…๊ณผ ์ถ”์ถœ๋œ ๋ฉ”๋ชจ๋ฆฌ๋Š” ์—์ด์ „ํŠธ ๋™์ž‘์„ ํŠธ๋ฆฌ๊ฑฐํ•˜๋Š” prompts๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Generative Agents [20]์—์„œ ์—์ด์ „ํŠธ๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ŠคํŠธ๋ฆผ์„ ์œ ์ง€ํ•˜๋ฉฐ, ๊ฐ ๋™์ž‘์„ ์ทจํ•˜๊ธฐ ์ „์— ๋ฉ”๋ชจ๋ฆฌ ์ŠคํŠธ๋ฆผ์—์„œ ์ตœ๊ทผ์˜ ๊ด€๋ จ์„ฑ ์žˆ๊ณ  ์ค‘์š”ํ•œ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜์—ฌ ์—์ด์ „ํŠธ ๋™์ž‘์„ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. GITM [16]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ฟผ๋ฆฌํ•˜์—ฌ ํ•˜์œ„ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ž‘์—…๊ณผ ๊ด€๋ จ๋œ ์„ฑ๊ณต์ ์ธ ๊ฒฝํ—˜์ด ์žˆ๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ด์ „์— ์œ ์‚ฌํ•œ ์ž‘์—…์„ ์™„๋ฃŒํ•œ ์ ์ด ์žˆ๋‹ค๋ฉด ์—์ด์ „ํŠธ๋Š” ์ด์ „์— ์„ฑ๊ณตํ•œ actions์„ ํ˜ธ์ถœํ•˜์—ฌ ํ˜„์žฌ ์ž‘์—…์„ ์ง์ ‘ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ChatDev [18] ๋ฐ MetaGPT [23]๊ณผ ๊ฐ™์€ ํ˜‘์—… ์—์ด์ „ํŠธ์—์„œ๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๊ฐ€ ์„œ๋กœ ํ†ต์‹ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์—์„œ ๋Œ€ํ™”์˜ ๋Œ€ํ™” ๊ธฐ๋ก์€ ์—์ด์ „ํŠธ ๋ฉ”๋ชจ๋ฆฌ์— ๊ธฐ์–ต๋ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ๊ฐ€ ์ƒ์„ฑํ•œ ๊ฐ ๋ฐœํ™”๋Š” ๋ฉ”๋ชจ๋ฆฌ์˜ ์˜ํ–ฅ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค. (2) _Action via Plan Following_. ์ด ์ „๋žต์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฏธ๋ฆฌ ์ƒ์„ฑ๋œ ๊ณ„ํš์— ๋”ฐ๋ผ actions์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, DEPS [33]์—์„œ๋Š” ์ฃผ์–ด์ง„ ์ž‘์—…์— ๋Œ€ํ•ด ์—์ด์ „ํŠธ๊ฐ€ ๋จผ์ € action ๊ณ„ํš์„ ์„ธ์›๋‹ˆ๋‹ค. ๊ณ„ํš ์‹คํŒจ๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์‹ ํ˜ธ๊ฐ€ ์—†์œผ๋ฉด ์—์ด์ „ํŠธ๋Š” ์ด๋Ÿฌํ•œ ๊ณ„ํš์„ ์—„๊ฒฉํ•˜๊ฒŒ ์ค€์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. GITM [16]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ž‘์—…์„ ์—ฌ๋Ÿฌ ํ•˜์œ„ ๋ชฉํ‘œ๋กœ ๋ถ„ํ•ดํ•˜์—ฌ ์ƒ์œ„ ๊ณ„ํš์„ ์„ธ์›๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณ„ํš์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์—์ด์ „ํŠธ๋Š” ๊ฐ ํ•˜์œ„ ๋ชฉํ‘œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ actions์„ ์ทจํ•˜์—ฌ ์ตœ์ข… ์ž‘์—…์„ ์™„๋ฃŒํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**์•ก์…˜ ํ”„๋กœ๋•์…˜**: ๋ชจ๋ธ ์ž…๋ ฅ๊ณผ ์ถœ๋ ฅ์ด ์ง์ ‘ ์—ฐ๊ด€๋˜์–ด ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ LLM๊ณผ ๋‹ฌ๋ฆฌ, ์—์ด์ „ํŠธ๋Š” ๋‹ค์–‘ํ•œ ์ „๋žต๊ณผ ์†Œ์Šค๋ฅผ ํ†ตํ•ด ์•ก์…˜์„ ์ทจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์˜ ์•ก์…˜ ์ƒ์„ฑ ์ „๋žต์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. (1) _๋ฉ”๋ชจ๋ฆฌ ํšŒ์ƒ์„ ํ†ตํ•œ ์•ก์…˜_. ์ด ์ „๋žต์—์„œ๋Š” ํ˜„์žฌ ์ž‘์—…์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•˜์—ฌ ์•ก์…˜์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ž‘์—…๊ณผ ์ถ”์ถœ๋œ ๋ฉ”๋ชจ๋ฆฌ๋Š” ์—์ด์ „ํŠธ ๋™์ž‘์„ ํŠธ๋ฆฌ๊ฑฐํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ƒ์„ฑ ์—์ด์ „ํŠธ [20]์—์„œ ์—์ด์ „ํŠธ๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ŠคํŠธ๋ฆผ์„ ์œ ์ง€ํ•˜๋ฉฐ, ๊ฐ ๋™์ž‘์„ ์ทจํ•˜๊ธฐ ์ „์— ๋ฉ”๋ชจ๋ฆฌ ์ŠคํŠธ๋ฆผ์—์„œ ์ตœ๊ทผ์˜ ๊ด€๋ จ์„ฑ ์žˆ๊ณ  ์ค‘์š”ํ•œ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜์—ฌ ์—์ด์ „ํŠธ ๋™์ž‘์„ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. GITM [16]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ฟผ๋ฆฌํ•˜์—ฌ ํ•˜์œ„ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ž‘์—…๊ณผ ๊ด€๋ จ๋œ ์„ฑ๊ณต์ ์ธ ๊ฒฝํ—˜์ด ์žˆ๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ด์ „์— ์œ ์‚ฌํ•œ ์ž‘์—…์„ ์™„๋ฃŒํ•œ ์ ์ด ์žˆ๋‹ค๋ฉด ์—์ด์ „ํŠธ๋Š” ์ด์ „์— ์„ฑ๊ณตํ•œ ์ž‘์—…์„ ํ˜ธ์ถœํ•˜์—ฌ ํ˜„์žฌ ์ž‘์—…์„ ์ง์ ‘ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ChatDev [18] ๋ฐ MetaGPT [23]๊ณผ ๊ฐ™์€ ํ˜‘์—… ์—์ด์ „ํŠธ์—์„œ๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๊ฐ€ ์„œ๋กœ ํ†ต์‹ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์—์„œ ๋Œ€ํ™”์˜ ๋Œ€ํ™” ๊ธฐ๋ก์€ ์—์ด์ „ํŠธ ๋ฉ”๋ชจ๋ฆฌ์— ๊ธฐ์–ต๋ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ๊ฐ€ ์ƒ์„ฑํ•œ ๊ฐ ๋ฐœํ™”๋Š” ๋ฉ”๋ชจ๋ฆฌ์˜ ์˜ํ–ฅ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค. (2) _๊ณ„ํš ์ถ”์ข…์„ ํ†ตํ•œ ์•ก์…˜_. ์ด ์ „๋žต์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฏธ๋ฆฌ ์ƒ์„ฑ๋œ ๊ณ„ํš์— ๋”ฐ๋ผ ์•ก์…˜์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, DEPS [33]์—์„œ๋Š” ์ฃผ์–ด์ง„ ์ž‘์—…์— ๋Œ€ํ•ด ์—์ด์ „ํŠธ๊ฐ€ ๋จผ์ € ์ž‘์—… ๊ณ„ํš์„ ์„ธ์›๋‹ˆ๋‹ค. ๊ณ„ํš ์‹คํŒจ๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์‹ ํ˜ธ๊ฐ€ ์—†์œผ๋ฉด ์—์ด์ „ํŠธ๋Š” ์ด๋Ÿฌํ•œ ๊ณ„ํš์„ ์—„๊ฒฉํ•˜๊ฒŒ ์ค€์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. GITM [16]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ž‘์—…์„ ์—ฌ๋Ÿฌ ํ•˜์œ„ ๋ชฉํ‘œ๋กœ ๋ถ„ํ•ดํ•˜์—ฌ ์ƒ์œ„ ๊ณ„ํš์„ ์„ธ์›๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณ„ํš์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์—์ด์ „ํŠธ๋Š” ๊ฐ ํ•˜์œ„ ๋ชฉํ‘œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ์•ก์…˜์„ ์ทจํ•˜์—ฌ ์ตœ์ข… ์ž‘์—…์„ ์™„๋ฃŒํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Action Production**: Different from ordinary LLMs, where the model input and output are directly associated, the agent may take actions via different strategies and sources. In the following, we introduce two types of commonly used action production strategies. (1) _Action via Memory Recollection_. In this strategy, the action is generated by extracting information from the agent memory according to the current task. The task and the extracted memories are used as prompts to trigger the agent actions. For example, in Generative Agents [20], the agent maintains a memory stream, and before taking each action, it retrieves recent, relevant and important information from the memory steam to guide the agent actions. In GITM [16], in order to achieve a low-level sub-goal, the agent queries its memory to determine if there are any successful experiences related to the task. If similar tasks have been completed previously, the agent invokes the previously successful actions to handle the current task directly. In collaborative agents such as ChatDev [18] and MetaGPT [23], different agents may communicate with each other. In this process, the conversation history in a dialog is remembered in the agent memories. Each utterance generated by the agent is influenced by its memory. (2) _Action via Plan Following_. In this strategy, the agent takes actions following its pre-generated plans. For instance, in DEPS [33], for a given task, the agent first makes action plans. If there are no signals indicating plan failure, the agent will strictly adhere to these plans. In GITM [16], the agent makes high-level plans by decomposing the task into many sub-goals. Based on these plans, the agent takes actions to solve each sub-goal sequentially to complete the final task. [|assistant|]
์ฑ— ๋ฒกํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋‹ค๋ฅธ ์–ธ์–ด๋กœ ๋ณ€ํ™˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘˜์งธ, LLaMA2๋ฅผ base model๋กœ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ Mistral์„ ๋Œ€์ฒด base model๋กœ ์กฐ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. Traditional Chinese์—์„œ ์ง€์†์ ์œผ๋กœ pretrainingํ•œ Mistral ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ธ Breeze-7B๋ฅผ CP ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•˜๊ณ  Mistral-instruct-0.2 chat vector๋ฅผ ํ†ตํ•ฉํ•œ ๊ฒฐ๊ณผ, ๊ธฐ์กด Breeze-Instruct์— ๋น„ํ•ด ๋” ์šฐ์ˆ˜ํ•œ ์ ์ˆ˜๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” Mistral ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ด chat vector์— ์ ์‘ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, chat vector์˜ ๋‹ค์šฉ๋„์„ฑ์€ ์ค‘๊ตญ์–ด์—๋งŒ ๊ตญํ•œ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ•œ๊ตญ์–ด๋ฅผ ์˜ˆ๋กœ ๋“ค์–ด, ์ง€์†์ ์œผ๋กœ pretrainingํ•œ Korean LLaMA2๋ฅผ LLaMA2 chat vector์™€ ๊ฒฐํ•ฉํ•˜๋ฉด ๋ชจ๋ธ์ด instruction-following ๊ธฐ๋Šฅ์„ ์Šต๋“ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” chat vector๊ฐ€ ๋‹ค์–‘ํ•œ ์–ธ์–ด์— ๊ฑธ์ณ ํšจ๊ณผ์ ์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
์ฑ— ๋ฒกํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋‹ค๋ฅธ ์–ธ์–ด๋กœ ๋ณ€ํ™˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘˜์งธ, LLaMA2๋ฅผ ๊ธฐ๋ณธ ๋ชจ๋ธ๋กœ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ Mistral์„ ๋Œ€์ฒด ๊ธฐ๋ณธ ๋ชจ๋ธ๋กœ ์กฐ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠธ์œ„ํ„ฐ์—์„œ ์ง€์†์ ์œผ๋กœ ์ „์ฒ˜๋ฆฌ๋ฅผ ๊ฑฐ์นœ ๋ฏธ์ŠคํŠธ๋ž„ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ธ Breeze-7B๋ฅผ CP ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•˜๊ณ  ๋ฏธ์ŠคํŠธ๋ž„ ์ธ์ŠคํŠธ๋ŸญํŠธ-0.2 ์ฑ„ํŒ… ๋ฒกํ„ฐ๋ฅผ ํ†ตํ•ฉํ•œ ๊ฒฐ๊ณผ, ๊ธฐ์กด ๋ธŒ๋ฆฌ์ฆˆ ์ธ์ŠคํŠธ๋ŸญํŠธ์— ๋น„ํ•ด ๋” ์šฐ์ˆ˜ํ•œ ์ ์ˆ˜๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ฏธ์ŠคํŠธ๋ž„ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ด ์ฑ„ํŒ… ๋ฒกํ„ฐ์— ์ ์‘ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ฑ„ํŒ… ๋ฒกํ„ฐ์˜ ๋‹ค์šฉ๋„์„ฑ์€ ์ค‘๊ตญ์–ด์—๋งŒ ๊ตญํ•œ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ•œ๊ตญ์–ด๋ฅผ ์˜ˆ๋กœ ๋“ค์–ด ํ•œ๊ตญ์–ด๋ฅผ ์ง€์†์ ์œผ๋กœ ์ „์ฒ˜๋ฆฌํ•œ LLaMA2๋ฅผ LLaMA2 ์ฑ„ํŒ… ๋ฒกํ„ฐ์™€ ๊ฒฐํ•ฉํ•˜๋ฉด ๋ชจ๋ธ์ด ๋ช…๋ น์–ด๋ฅผ ๋”ฐ๋ฅด๋Š” ๊ธฐ๋Šฅ์„ ์Šต๋“ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ฑ„ํŒ… ๋ฒกํ„ฐ๊ฐ€ ๋‹ค์–‘ํ•œ ์–ธ์–ด์— ๊ฑธ์ณ ํšจ๊ณผ์ ์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]can be transformed into different languages through the utilization of chat vectors. Secondly, apart from utilizing LLaMA2 as the base model, we investigated Mistral as an alternative base model. Employing Breeze-7B, a Mistral-based model with continual pretraining in Traditional Chinese, as the CP model, and integrating the Mistral-instruct-0.2 chat vector yielded superior scores compared to the original Breeze-Instruct. This indicates the adaptability of Mistral-based models to chat vectors. Finally, the versatility of chat vectors is not limited to Chinese. Taking Korean as an example, applying the LLaMA2 with cotinual pretraining Korean, combined with the LLaMA2 chat vector, enables the model to acquire instruction-following capabilities. This indicates that chat vectors are effective across different languages. [|assistant|]
์‚ฌ์ „ ํ•™์Šต ๊ณ„์†ํ•˜๊ธฐ. ์šฐ๋ฆฌ๋Š” sequence parallelism [43]๊ณผ distributed attention์„ ์‚ฌ์šฉํ•˜์—ฌ full-attention ๋ชจ๋ธ์„ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, sparse ๋˜๋Š” linear attention์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ณ  full attention์— ๋Œ€ํ•œ brute force ๊ตฌํ˜„์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์„น์…˜ 2์—์„œ ์†Œ๊ฐœํ•œ (1) ์›๋ž˜์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ; (2) ๊ธธ์ด๊ฐ€ ์—…์ƒ˜ํ”Œ๋ง๋œ long-context ๋ฐ์ดํ„ฐ(๊ธด ๋ฌธ์„œ๋Š” ๋Œ€๋ถ€๋ถ„ ์ฑ…์—์„œ ๊ฐ€์ ธ์˜ด); (3) ๋‹ต๋ณ€์— ๋‹ต๋ณ€ ์•ž์˜ ๊ด€๋ จ ๋‹จ๋ฝ์„ ์ธ์šฉํ•˜๋Š” QA ์Œ์„ ๊ตฌ์„ฑํ•˜๋Š” multi-document question-answering synthetic ๋ฐ์ดํ„ฐ์˜ ํ˜ผํ•ฉ๋ฌผ์— ๋Œ€ํ•ด Yi 6B/34B base ๋ชจ๋ธ์„ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฐ์ดํ„ฐ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋Œ€๋ถ€๋ถ„ Fu et al. [22] ๋ฐ Yu et al. [87]์˜ ๋ฐ์ดํ„ฐ ์—”์ง€๋‹ˆ์–ด๋ง ๊ด€ํ–‰์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” 4M batch size๋กœ 5B tokens์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•˜์—ฌ 100 optimization ๋‹จ๊ณ„๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค. Fu et al. [22]์˜ ๋™์‹œ ์ž‘์—…๊ณผ ์ผ์น˜ํ•˜๋„๋ก, ์šฐ๋ฆฌ๋Š” ๊ทธ๋ฆผ 6์—์„œ ๋ณด์—ฌ์งˆ ๊ฒƒ์ฒ˜๋Ÿผ ์ด๋Ÿฌํ•œ ๊ฐ€๋ฒผ์šด ์›จ์ดํŠธ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šต์ด ์ด๋ฏธ Needle-in-a-Haystack ํ…Œ์ŠคํŠธ์—์„œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์‚ฌ์ „ ํ•™์Šต ๊ณ„์†ํ•˜๊ธฐ์‹œํ€€์Šค ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ[43]์™€ ๋ถ„์‚ฐ ์ฃผ์˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ „์ฒด ์ฃผ์˜ ๋ชจ๋ธ์„ ์‚ฌ์ „ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์ŠคํŒŒ์Šค ๋˜๋Š” ์„ ํ˜• ์ฃผ์˜๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ณ  ์ „์ฒด ์ฃผ์˜์— ๋Œ€ํ•œ ๋ฌด์ฐจ๋ณ„ ๋Œ€์ž… ๊ตฌํ˜„์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 2์—์„œ ์†Œ๊ฐœํ•œ (1)์˜ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ๋ฌผ์— ๋Œ€ํ•ด Yi 6B/34B ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ์‚ฌ์ „ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค; (2). ๊ธธ์ด๊ฐ€ ์—…์ƒ˜ํ”Œ๋ง๋œ ๊ธด ์ปจํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ(๊ธด ๋ฌธ์„œ๋Š” ๋Œ€๋ถ€๋ถ„ ์ฑ…์—์„œ ๊ฐ€์ ธ์˜ด); (3). ๋‹ต๋ณ€์— ๋‹ต๋ณ€ ์•ž์˜ ๊ด€๋ จ ๋‹จ๋ฝ์„ ์ธ์šฉํ•˜๋Š” QA ์Œ์„ ๊ตฌ์„ฑํ•˜๋Š” ๋‹ค์ค‘ ๋ฌธ์„œ ์งˆ์˜์‘๋‹ต ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ. ๋ฐ์ดํ„ฐ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋Œ€๋ถ€๋ถ„ Fu et al. [22] ๋ฐ Yu et al. [87]. 4M ๋ฐฐ์น˜ ํฌ๊ธฐ๋กœ 5B ํ† ํฐ์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ๊ณ„์† ํ•™์Šตํ•˜์—ฌ 100๊ฐœ์˜ ์ตœ์ ํ™” ๋‹จ๊ณ„๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค. Fu et al. [22]์˜ ๋™์‹œ ์ž‘์—…๊ณผ ์ผ์น˜ํ•˜๋„๋ก, ์šฐ๋ฆฌ๋Š” ๊ทธ๋ฆผ 6์—์„œ ๋ณด์—ฌ์งˆ ๊ฒƒ์ฒ˜๋Ÿผ ์ด๋Ÿฌํ•œ ๊ฐ€๋ฒผ์šด ์›จ์ดํŠธ ๊ณ„์† ํ•™์Šต์ด ์ด๋ฏธ ๊ฑด์ดˆ๋”๋ฏธ์—์„œ ๋ฐ”๋Š˜์„ ์ฐพ๋Š” ํ…Œ์ŠคํŠธ์—์„œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Continue PretrainingWe continue pretrain the full-attention model using sequence parallelism [43] and distributed attention. This is to say, we do not use any sparse or linear attention, but use a brute force implementation of the full attention. We continue pretrain the Yi 6B/ 34B base model on the data mixture of (1). original pretraining data, as is introduced in section 2; (2). length-upsampled long-context data, where the long documents are mostly from books; (3). multi-document question-answering synthetic data, where we construct QA pairs where the answer contains a recitation of the related paragraph before the answer. Our data approach mostly follows the data engineering practice in Fu et al. [22] and Yu et al. [87]. We continue pretrain the model on 5B tokens with 4M batch size, which translate to 100 optimization steps. Aligning with the concurrent work from Fu et al. [22], we observe that such light-weight continue pretraining is already able to enable a strong performance on Needle-in-a-Haystack test, as we will show in Figure 6. [|assistant|]
**๋‹ค์–‘์„ฑ ๋ถ€๋ถ„**. ๋‹ค์–‘์„ฑ ๊ตฌ์„ฑ ์š”์†Œ์˜ ๊ฒฝ์šฐ, ๋‹ค์Œ๊ณผ ๊ฐ™์ด \(P_{i}\)๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ฐ \(a^{\prime}\in[K]\)์˜ ๊ฒฝ์šฐ, \(a\)๋Š” \(\mathcal{A}\)์—์„œ \(a^{\prime}\)-๋ฒˆ์งธ ์†์„ฑ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. \(P_{N+(a^{\prime}-1)D+i}\)๋Š” ์ฒ˜์Œ์— \(F^{\perp}(W(\mathcal{Z}),n,a,R_{i})\), ์—ฌ๊ธฐ์„œ \(n\in\mathcal{N}\)์€ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
**๋‹ค์–‘์„ฑ ๋ถ€๋ถ„**. ๋‹ค์–‘์„ฑ ๊ตฌ์„ฑ ์š”์†Œ์˜ ๊ฒฝ์šฐ, ๋‹ค์Œ๊ณผ ๊ฐ™์ด \(P_{i}\)๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ฐ \(a^{\ํ”„๋ผ์ž„}\in[K]\)์˜ ๊ฒฝ์šฐ, \(a\)๋Š” \(\mathcal{A}\)์—์„œ \(a^{\ํ”„๋ผ์ž„}\)-๋ฒˆ์งธ ์†์„ฑ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. P_{N+(a^{\ํ”„๋ผ์ž„}-1)D+i}\)๋Š” ์ฒ˜์Œ์— \(F^{\perp}(W(\mathcal{Z}),n,a,R_{i})\), ์—ฌ๊ธฐ์„œ \(n\in\mathcal{N}\)์€ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Diversity Part.** For the diversity component, we construct the \(P_{i}\)'s as follows. For each \(a^{\prime}\in[K]\), let \(a\) denote the \(a^{\prime}\)-th attribute in \(\mathcal{A}\). We form \(P_{N+(a^{\prime}-1)D+i}\) by initially calculating \(F^{\perp}(W(\mathcal{Z}),n,a,R_{i})\), where \(n\in\mathcal{N}\) is selected uniformly at random. [|assistant|]
ํ‰๊ฐ€ ๊ตฌ์„ฑT5 ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, ๋ชจ๋“  ํ‰๊ฐ€๋Š” zero-shot ๋ฐฉ์‹์œผ๋กœ ์ˆ˜ํ–‰๋˜๋ฉฐ ๋‹จ์ผ GPU๋กœ ์ฒ˜๋ฆฌ๋ฉ๋‹ˆ๋‹ค. InvariantLAMA์˜ ๊ฒฝ์šฐ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ๊ธธ์ด๋Š” ๊ฐ๊ฐ 25์™€ 4๋กœ ๊ณ ์ •๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. UpdatedLAMA ๋ฐ NewLAMA์˜ ๊ฒฝ์šฐ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ๊ธธ์ด๋Š” ๊ฐ๊ฐ 50 ๋ฐ 10์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, NewLAMA-Easy์˜ ๊ฒฝ์šฐ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ๊ธธ์ด๋Š” ๊ฐ๊ฐ 150 ๋ฐ 10์ž…๋‹ˆ๋‹ค. ์ด ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๊ทผ๊ฑฐ๋Š” ํ‘œ 5์˜ ํ‰๊ท  ์ž…๋ ฅ ๋ฐ ๋‹ต๋ณ€ ํ† ํฐ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
ํ‰๊ฐ€ ๊ตฌ์„ฑT5 ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, ๋ชจ๋“  ํ‰๊ฐ€๋Š” ์ œ๋กœ ์ƒท ๋ฐฉ์‹์œผ๋กœ ์ˆ˜ํ–‰๋˜๋ฉฐ ๋‹จ์ผ GPU๋กœ ์ฒ˜๋ฆฌ๋ฉ๋‹ˆ๋‹ค. InvariantLAMA์˜ ๊ฒฝ์šฐ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ๊ธธ์ด๋Š” ๊ฐ๊ฐ 25์™€ 4๋กœ ๊ณ ์ •๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์—…๋ฐ์ดํŠธ๋œ LAMA ๋ฐ NewLAMA์˜ ๊ฒฝ์šฐ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ๊ธธ์ด๋Š” ๊ฐ๊ฐ 50 ๋ฐ 10์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, NewLAMA-Easy์˜ ๊ฒฝ์šฐ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ๊ธธ์ด๋Š” ๊ฐ๊ฐ 150 ๋ฐ 10์ž…๋‹ˆ๋‹ค. ์ด ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๊ทผ๊ฑฐ๋Š” ํ‘œ 5์˜ ํ‰๊ท  ์ž…๋ ฅ ๋ฐ ๋‹ต๋ณ€ ํ† ํฐ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Evaluation ConfigurationsFor T5 based models, all evaluation is done in a zero-shot manner and is processed with a single GPU. For InvariantLAMA, the input and output length is fixed as 25 and 4 respectively. For UpdatedLAMA and NewLAMA, the input and output length is 50 and 10 respectively. Lastly, the input and output length is 150 and 10 respectively for NewLAMA-Easy. The rationale of this hyperparameter is based on average input and answer token in Table 5. [|assistant|]
๊ทธ๋ฆผ 4: **Forward Pass ๋ฐ Teacher Forcing**. ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๋‹จ์ผ forward pass๋ฅผ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์„ ์€ language model ๊ณ„์‚ฐ์„ ๋‚˜ํƒ€๋‚ด๊ณ  ์ ์„ ์€ teacher forcing์„ ํ†ตํ•ด ํ† ํฐ์ด ์‚ฝ์ž…๋˜๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ๋ฏน์„œ๋Š” mixing head๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ํŠนํžˆ ์„ธ ๊ฐœ์˜ ํ† ํฐ์„ ์•ž๋‹น๊ฒจ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์„ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ๊ณ  ์ƒ์„ฑ์€ ๊ทธ๋ฆผ 1๊ณผ ๊ทธ๋ฆผ 3์—์„œ ๋” ์ž์„ธํžˆ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 4: **ํฌ์›Œ๋“œ ํŒจ์Šค ๋ฐ ๊ต์‚ฌ ๊ฐ•์ œ**. ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๋‹จ์ผ ํฌ์›Œ๋“œ ํŒจ์Šค๋ฅผ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์„ ์€ ์–ธ์–ด ๋ชจ๋ธ ๊ณ„์‚ฐ์„ ๋‚˜ํƒ€๋‚ด๊ณ  ์ ์„ ์€ ๊ต์‚ฌ ๊ฐ•์ œ๋ฅผ ํ†ตํ•ด ํ† ํฐ์ด ์‚ฝ์ž…๋˜๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ๋ฏน์„œ๋Š” ๋ฏน์‹ฑ ํ—ค๋“œ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ํŠนํžˆ ์„ธ ๊ฐœ์˜ ํ† ํฐ์„ ์•ž๋‹น๊ฒจ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์„ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ๊ณ  ์ƒ์„ฑ์€ ๊ทธ๋ฆผ 1๊ณผ ๊ทธ๋ฆผ 3์—์„œ ๋” ์ž์„ธํžˆ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 4: **Forward Pass and Teacher Forcing**. We visualize a single forward pass of our algorithm. Solid lines denote language model computation, while dashed lines indicate tokens are inserted via teacher forcing, and the mixer represents the mixing head. In particular, we visualize predicting three tokens ahead. Thought generation is shown in more detail in Figure 1 and Figure 3. [|assistant|]
๋งŒ์•ฝ CnstScore\((\tau_{\text{2H}},\tau_{\text{1H}})\)๊ฐ€ EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)์— ์ง์ ‘ ์˜์กดํ•œ๋‹ค๋ฉด ๋ฏธ๋ถ„ ๋ฏธ์ ๋ถ„์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ณ€ํ™” ๋ฐฉํ–ฅ์„ ๊ณ„์‚ฐํ•˜์—ฌ ๋‹ต์„ ๊ตฌํ•  ์ˆ˜ ์žˆ์—ˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‘ ๊ฐ’ ์‚ฌ์ด์—๋Š” ์ง์ ‘์ ์ธ ํ•จ์ˆ˜์  ์ข…์†์„ฑ์ด ์กด์žฌํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋Œ€์‹ , ๋‘ ๋ฉ”ํŠธ๋ฆญ ๋ชจ๋‘ ๊ณ„์‚ฐ์„ ์œ„ํ•ด \(\mathbf{x}^{l}\)์— ์˜์กดํ•œ๋‹ค๋Š” ๊ณตํ†ต์ ์„ ํ™œ์šฉํ•˜์—ฌ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{[|endofturn|]
๋งŒ์•ฝ CnstScore\((\tau_{\text{2H}},\tau_{\text{1H}})\)๊ฐ€ EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)์— ์ง์ ‘ ์˜์กดํ•œ๋‹ค๋ฉด ๋ฏธ๋ถ„ ๋ฏธ์ ๋ถ„์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ณ€ํ™” ๋ฐฉํ–ฅ์„ ๊ณ„์‚ฐํ•˜์—ฌ ๋‹ต์„ ๊ตฌํ•  ์ˆ˜ ์žˆ์—ˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‘ ๊ฐ’ ์‚ฌ์ด์—๋Š” ์ง์ ‘์ ์ธ ํ•จ์ˆ˜์  ์ข…์†์„ฑ์ด ์กด์žฌํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋Œ€์‹ , ๋‘ ๋ฉ”ํŠธ๋ฆญ ๋ชจ๋‘ ๊ณ„์‚ฐ์„ ์œ„ํ•ด \(\mathbf{x}^{l}\)์— ์˜์กดํ•œ๋‹ค๋Š” ๊ณตํ†ต์ ์„ ํ™œ์šฉํ•˜์—ฌ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\), CnstScore\((\mathbf{x}^{l})\), \(\mathbf{x}^{l}\)์— ๋Œ€ํ•œ EntRec\((\mathbf{x}^{l})\[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We would have been able to use differential calculus to obtain the answer by calculating the direction of change if CnstScore\((\tau_{\text{2H}},\tau_{\text{1H}})\) were directly dependent on EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\). However, there exists no direct functional dependency between the two values. Instead, we leverage the shared reliance of both metrics on \(\mathbf{x}^{l}\) for computation where \(l\in[0,L-1)\),3 redefining them as EntRec\((\mathbf{x}^{l})\) and CnstScore\((\mathbf{x}^{l})\) relative to \(\mathbf{x}^{l}\). This reparameterization allows us to change the question to: if EntRec\((\mathbf{x}^{l})\) is increased by altering \(\mathbf{x}^{l}\), does CnstScore\((\mathbf{x}^{l})\) also increase? [|assistant|]
์—ฌ๊ธฐ์—์„œ๋Š” GPT-4๊ฐ€ reflection tokens์„ ์ˆ˜์ง‘ํ•˜๋„๋ก promptํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ instructions๊ณผ demonstrations์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 8์€ ์ดˆ๊ธฐ retrieval token์— ๋Œ€ํ•œ instructions๊ณผ demonstrations์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 9๋Š” ์ฃผ์–ด์ง„ instruction, ์•ž ๋ฌธ์žฅ ๋ฐ ์ด์ „์— retrieval๋œ ๊ตฌ์ ˆ์— ๋Œ€ํ•œ three-way output tokens์„ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ instructions๊ณผ demonstrations์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. demonstration๊ณผ ํ…Œ์ŠคํŠธ ์ž…๋ ฅ์ด ๋” ๊ธธ๊ธฐ ๋•Œ๋ฌธ์— ๋‹จ์ผ demonstration๋งŒ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 10์€ [1.1]์— ๋Œ€ํ•œ three-way output tokens์„ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” instruction๊ณผ demonstrations์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 11์€ [1.1]์— ๋Œ€ํ•œ three-way output tokens์„ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” instruction๊ณผ demonstrations์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 12๋Š” [1.1]์— ๋Œ€ํ•œ five-way output tokens์„ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” instruction๊ณผ demonstrations์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
์—ฌ๊ธฐ์—์„œ๋Š” GPT-4๊ฐ€ ๋ฐ˜์‚ฌ ํ† ํฐ์„ ์ˆ˜์ง‘ํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ ์ง€์นจ๊ณผ ๋ฐ๋ชจ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 8์€ ์ดˆ๊ธฐ ๊ฒ€์ƒ‰ ํ† ํฐ์— ๋Œ€ํ•œ ์ง€์นจ๊ณผ ๋ฐ๋ชจ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 9๋Š” ์ฃผ์–ด์ง„ ์ง€์นจ, ์•ž ๋ฌธ์žฅ ๋ฐ ์ด์ „์— ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์— ๋Œ€ํ•œ 3์ž๋ฆฌ ์ถœ๋ ฅ ํ† ํฐ์„ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ ์ง€์นจ๊ณผ ๋ฐ๋ชจ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ฐ๋ชจ์™€ ํ…Œ์ŠคํŠธ ์ž…๋ ฅ์ด ๋” ๊ธธ๊ธฐ ๋•Œ๋ฌธ์— ๋‹จ์ผ ๋ฐ๋ชจ๋งŒ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 10์€ [1.1]์— ๋Œ€ํ•œ 3์ž๋ฆฌ ์ถœ๋ ฅ ํ† ํฐ์„ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ์ง€์นจ๊ณผ ๋ฐ๋ชจ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 11์€ [1.1]์— ๋Œ€ํ•œ 3์ž๋ฆฌ ์ถœ๋ ฅ ํ† ํฐ์„ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ์ง€์นจ๊ณผ ๋ฐ๋ชจ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 12๋Š” [1.1]์— ๋Œ€ํ•œ 5์ž๋ฆฌ ์ถœ๋ ฅ ํ† ํฐ์„ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ์ง€์นจ๊ณผ ๋ฐ๋ชจ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Here, we show the instructions and demonstrations used to prompt GPT-4 to collect reflection tokens. Table 8 shows the instructions and demonstrations for the initial retrieval token. Table 9 shows the instruction and demonstrations used to collect the three-way output tokens for [1.1] given instruction, preceding sentences, and previously retrieved passages. Due to the longer demonstration and test input, we only use a single demonstration. Table 10 shows an instruction and demonstrations used to collect the three-way output tokens for [1.1]. Table 11 shows an instruction and demonstrations used to collect the three-way output tokens for [1.1]. Table 12 shows an instruction and demonstrations used to collect the five-way output tokens for [1.1]. [|assistant|]
์ธ๊ฐ„์€ ์ฃผ์–ด์ง„ ์ „์ œ์—์„œ ๊ฒฐ๋ก ์„ ์ถ”๋ก ํ•˜๋Š” ์—ฐ์—ญ์  ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: \(r_{1}(e_{1})=e_{2}\)๋ผ๋Š” ์ „์ œ์™€ \(r_{2}(e_{2})=e_{3}\)๋ผ๋Š” ๋‹ค๋ฅธ ์ „์ œ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ \(r_{2}(r_{1}(e_{1}))=e_{3}\)๋ฅผ ์ถ”๋ก ํ•˜๋Š” ๊ฒƒ). ์ด๋Ÿฌํ•œ multi-hop reasoning[20, 21]์—๋Š” bridge entity(์˜ˆ: "๋ฏธ์‹ "์˜ ๊ฐ€์ˆ˜๊ฐ€ ์Šคํ‹ฐ๋น„ ์›๋”๋ผ๋Š” ๊ฒƒ)๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข… ๋‹ต(์˜ˆ: ์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๊ฐ€ ๋ฃฐ๋ผ๋ผ๋Š” ๊ฒƒ)์„ ํ‘ธ๋Š” ๊ณผ์ •์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ธ๊ฐ„์€ ์ฃผ์–ด์ง„ ์ „์ œ์—์„œ ๊ฒฐ๋ก ์„ ์ถ”๋ก ํ•˜๋Š” ์—ฐ์—ญ์  ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: r_{1}(e_{1})=e_{2}\)๋ผ๋Š” ์ „์ œ์™€ \(r_{2}(e_{2})=e_{3}\)๋ผ๋Š” ๋‹ค๋ฅธ ์ „์ œ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ \(r_{2}(r_{1}(e_{1}))=e_{3}\)๋ฅผ ์ถ”๋ก ํ•˜๋Š” ๊ฒƒ). ์ด๋Ÿฌํ•œ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก [20, 21]์—๋Š” ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ(์˜ˆ: "๋ฏธ์‹ "์˜ ๊ฐ€์ˆ˜๊ฐ€ ์Šคํ‹ฐ๋น„ ์›๋”๋ผ๋Š” ๊ฒƒ)๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข… ๋‹ต(์˜ˆ: ์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๊ฐ€ ๋ฃฐ๋ผ๋ผ๋Š” ๊ฒƒ)์„ ํ‘ธ๋Š” ๊ณผ์ •์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Humans possess the deductive reasoning ability to infer conclusions from given premises, such as deducing that \(r_{2}(r_{1}(e_{1}))=e_{3}\) given a premise stating that \(r_{1}(e_{1})=e_{2}\) and another premise stating that \(r_{2}(e_{2})=e_{3}\). This multi-hop reasoning [20, 21] involves identifying the bridge entity (e.g., that "the singer of 'Superstition'" is Stevie Wonder) and using it to solve for the final answer (e.g., that Stevie Wonder's mother is Lula). [|assistant|]
๋ณต์žก์„ฑ๊ณผ ์˜ˆ์ธก ๋ถˆ๊ฐ€๋Šฅ์„ฑAI ์‹œ์Šคํ…œ์€ ์—„์ฒญ๋‚˜๊ฒŒ ๋งŽ์€ ์ž ์žฌ์  ์ž…๋ ฅ์„ ๋ฐ›์•„๋“ค์ด๊ณ  ์—„์ฒญ๋‚˜๊ฒŒ ๋งŽ์€ ์ž ์žฌ์  ์ถœ๋ ฅ์„ ๋ฐ˜ํ™˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋Œ€๋ถ€๋ถ„์˜ LMs์—๋Š” ํ˜„์žฌ 4,000๊ฐœ์˜ ํ† ํฐ์ด๋ผ๋Š” ์ปจํ…์ŠคํŠธ ์ฐฝ์ด ์žˆ์œผ๋ฉฐ, ๊ฒฝ์šฐ์— ๋”ฐ๋ผ์„œ๋Š” ์ตœ๋Œ€ 200,000๊ฐœ ์ด์ƒ์˜ ํ† ํฐ์ด ์žˆ์Šต๋‹ˆ๋‹ค(์ผ๋ฐ˜์ ์œผ๋กœ ํ…์ŠคํŠธ์˜ 150ํŽ˜์ด์ง€ ์ด์ƒ). ๋ชจ๋ธ์€ ์กฐ์ • ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ์ˆ˜์‹ญ์–ต ๊ฐœ์— ๋‹ฌํ•˜๋ฉฐ, ๊ฐ ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ๋ชจ๋ธ์˜ ์ „๋ฐ˜์ ์ธ ๋™์ž‘์— ์–ด๋ ค์šด ์ถ”๋ก ์ด ํ•„์š”ํ•œ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์„ค์ •ํ•˜์—ฌ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ์„ ๋ณด๋‹ค ๊ฒฐ์ •๋ก ์ ์œผ๋กœ ๋งŒ๋“ค๋”๋ผ๋„(์˜ˆ: ๋‚ฎ์€ ์˜จ๋„ ์„ค์ •), ๋ชจ๋ธ์˜ ์‘๋‹ต์€ ์—ฌ์ „ํžˆ ํ™•๋ฅ ์ ์ด๋ฉฐ ์ž…๋ ฅ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ์ด๋Š” ์ถ”์ƒ์ ์ธ ๊ฐœ๋…์— ๋Œ€ํ•œ ์ถ”๋ก ์ด๋‚˜ ์ƒˆ๋กœ์šด ์ฝ˜ํ…์ธ  ์ƒ์„ฑ๊ณผ ๊ฐ™์€ ์ฐฝ์˜์ ์ธ hallucination๊ณผ ์ฐฝ๋ฐœ์  ํ–‰๋™์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํฐ ๊ฐ•์ ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Š” ๋™์ž‘์„ ์˜ˆ์ธกํ•˜๊ณ  ์‘๋‹ต ์ค‘ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ์‘๋‹ต์ด ์—†๋Š”์ง€ ํ™•์ธํ•˜๊ธฐ ์–ด๋ ต๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค.[|endofturn|]
๋ณต์žก์„ฑ๊ณผ ์˜ˆ์ธก ๋ถˆ๊ฐ€๋Šฅ์„ฑAI ์‹œ์Šคํ…œ์€ ์—„์ฒญ๋‚˜๊ฒŒ ๋งŽ์€ ์ž ์žฌ์  ์ž…๋ ฅ์„ ๋ฐ›์•„๋“ค์ด๊ณ  ์—„์ฒญ๋‚˜๊ฒŒ ๋งŽ์€ ์ž ์žฌ์  ์ถœ๋ ฅ์„ ๋ฐ˜ํ™˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋Œ€๋ถ€๋ถ„์˜ LM์—๋Š” ํ˜„์žฌ 4,000๊ฐœ์˜ ํ† ํฐ์ด๋ผ๋Š” ์ปจํ…์ŠคํŠธ ์ฐฝ์ด ์žˆ์œผ๋ฉฐ, ๊ฒฝ์šฐ์— ๋”ฐ๋ผ์„œ๋Š” ์ตœ๋Œ€ 200,000๊ฐœ ์ด์ƒ์˜ ํ† ํฐ์ด ์žˆ์Šต๋‹ˆ๋‹ค(์ผ๋ฐ˜์ ์œผ๋กœ ํ…์ŠคํŠธ์˜ 150ํŽ˜์ด์ง€ ์ด์ƒ). ๋ชจ๋ธ์€ ์กฐ์ • ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ์ˆ˜์‹ญ์–ต ๊ฐœ์— ๋‹ฌํ•˜๋ฉฐ, ๊ฐ ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ๋ชจ๋ธ์˜ ์ „๋ฐ˜์ ์ธ ๋™์ž‘์— ์–ด๋ ค์šด ์ถ”๋ก ์ด ํ•„์š”ํ•œ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์„ค์ •ํ•˜์—ฌ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ์„ ๋ณด๋‹ค ๊ฒฐ์ •๋ก ์ ์œผ๋กœ ๋งŒ๋“ค๋”๋ผ๋„(์˜ˆ: ๋‚ฎ์€ ์˜จ๋„ ์„ค์ •), ๋ชจ๋ธ์˜ ์‘๋‹ต์€ ์—ฌ์ „ํžˆ ํ™•๋ฅ ์ ์ด๋ฉฐ ์ž…๋ ฅ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ์ด๋Š” ์ถ”์ƒ์ ์ธ ๊ฐœ๋…์— ๋Œ€ํ•œ ์ถ”๋ก ์ด๋‚˜ ์ƒˆ๋กœ์šด ์ฝ˜ํ…์ธ  ์ƒ์„ฑ๊ณผ ๊ฐ™์€ ์ฐฝ์˜์ ์ธ ํ™˜๊ฐ๊ณผ ์ฐฝ๋ฐœ์  ํ–‰๋™์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํฐ ๊ฐ•์ ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.25 ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Š” ๋™์ž‘์„ ์˜ˆ์ธกํ•˜๊ณ  ์‘๋‹ต ์ค‘ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ์‘๋‹ต์ด ์—†๋Š”์ง€ ํ™•์ธํ•˜๊ธฐ ์–ด๋ ต๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Complexity and unpredictabilityAI systems can accept a huge number of potential inputs and return a vast number of potential outputs. For instance, most LMs now have context windows of 4,000 tokens, and in some cases up to 200,000 or more--which is typically \(150+\) pages of text. Models often consist of billions of tunable parameters, each of which exerts some difficult-to-reason-about impact on the model's overall behavior. Furthermore, even when hyperparameters are set so that models' output is more deterministic (e.g., setting a low temperature), model responses are still probabilistic and conditioned on inputs. This can be a great strength as it allows for creative hallucinations and emergent behavior, such as reasoning about abstract concepts or creating novel content.25 However, it also makes it difficult to predict their behavior and ensure that none of their responses are unsafe. [|assistant|]
**Prompt:** ์ˆ˜ํ•™ ๋ฌธ์ œ์™€ ํ•ด๊ฒฐ์ฑ…์ด ์ฃผ์–ด์ง€๋ฉด ๋จผ์ € ํ•ด๊ฒฐ์ฑ…์„ ์—ฌ๋Ÿฌ ๋‹จ๊ณ„๋กœ ๋‚˜๋ˆ„๊ณ  ๊ฐ ๋‹จ๊ณ„๊ฐ€ ์˜ฌ๋ฐ”๋ฅธ์ง€ ํŒ๋‹จํ•˜๊ณ  ์ด ์ •๋‹ต ๋‹จ๊ณ„๋ฅผ ์„ธ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**ํ”„๋กฌํ”„ํŠธ:** ์ˆ˜ํ•™ ๋ฌธ์ œ์™€ ํ•ด๊ฒฐ์ฑ…์ด ์ฃผ์–ด์ง€๋ฉด ๋จผ์ € ํ•ด๊ฒฐ์ฑ…์„ ์—ฌ๋Ÿฌ ๋‹จ๊ณ„๋กœ ๋‚˜๋ˆ„๊ณ  ๊ฐ ๋‹จ๊ณ„๊ฐ€ ์˜ฌ๋ฐ”๋ฅธ์ง€ ํŒ๋‹จํ•˜๊ณ  ์ด ์ •๋‹ต ๋‹จ๊ณ„๋ฅผ ์„ธ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Prompt:** You are given a math problem and a solution, you should first split the solution into multi steps and judge if each step is correct and count the total correct steps. [|assistant|]
# ์ถœ๋ ฅ [55]์— ๋ช…์‹œ๋œ ๋ฐ”์™€ ๊ฐ™์ด, In-Context Learning (ICL)์€ ์ž‘์—… ์„ค๋ช… ๋ฐ/๋˜๋Š” ๋ช‡ ๊ฐ€์ง€ ์ž‘์—… ์˜ˆ์ œ๋ฅผ ๋ฐ๋ชจ๋กœ ๊ตฌ์„ฑ๋œ ํ˜•์‹ํ™”๋œ ์ž์—ฐ์–ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 14๋Š” ICL์˜ ์˜ˆ์‹œ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋จผ์ € ์ž‘์—… ์„ค๋ช…๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜์—ฌ ์ž‘์—… ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ์ œ๋ฅผ ๋ฐ๋ชจ๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ž์—ฐ์–ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ˜•์„ฑํ•˜๊ธฐ ์œ„ํ•ด ํŠน์ • ์ˆœ์„œ๋กœ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ํ…Œ์ŠคํŠธ ์ธ์Šคํ„ด์Šค๋ฅผ ๋ฐ๋ชจ์— ์ถ”๊ฐ€ํ•˜์—ฌ Large Language Models (LLMs)์ด ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ์ž…๋ ฅํ•ฉ๋‹ˆ๋‹ค. ์ž‘์—… ๋ฐ๋ชจ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ LLMs์€ ๋ช…์‹œ์ ์ธ ๊ทธ๋ผ๋””์–ธํŠธ ์—…๋ฐ์ดํŠธ ์—†์ด๋„ ์ƒˆ๋กœ์šด ์ž‘์—…์„ ์ธ์‹ํ•˜๊ณ  ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[55]์— ๋ช…์‹œ๋œ ๋ฐ”์™€ ๊ฐ™์ด, ICL์€ ์ž‘์—… ์„ค๋ช… ๋ฐ/๋˜๋Š” ๋ช‡ ๊ฐ€์ง€ ์ž‘์—… ์˜ˆ์ œ๋ฅผ ๋ฐ๋ชจ๋กœ ๊ตฌ์„ฑ๋œ ํ˜•์‹ํ™”๋œ ์ž์—ฐ์–ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 14๋Š” ICL์˜ ์˜ˆ์‹œ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋จผ์ € ์ž‘์—… ์„ค๋ช…๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜์—ฌ ์ž‘์—… ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ์ œ๋ฅผ ๋ฐ๋ชจ๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ž์—ฐ์–ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ˜•์„ฑํ•˜๊ธฐ ์œ„ํ•ด ํŠน์ • ์ˆœ์„œ๋กœ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ํ…Œ์ŠคํŠธ ์ธ์Šคํ„ด์Šค๋ฅผ ๋ฐ๋ชจ์— ์ถ”๊ฐ€ํ•˜์—ฌ LLM์ด ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ์ž…๋ ฅํ•ฉ๋‹ˆ๋‹ค. ์ž‘์—… ๋ฐ๋ชจ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ LLM์€ ๋ช…์‹œ์ ์ธ ๊ทธ๋ผ๋””์–ธํŠธ ์—…๋ฐ์ดํŠธ ์—†์ด๋„ ์ƒˆ๋กœ์šด ์ž‘์—…์„ ์ธ์‹ํ•˜๊ณ  ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]As stated in [55], ICL uses a formatted natural language prompt, consisting of the task description and/or a few task examples as demonstrations. Figure 14 presents an illustration of ICL. First, starting with a task description, a few examples are selected from the task dataset as demonstrations. Then, they are combined in a specific order to form natural language prompts with specially designed templates. Finally, the test instance is appended to the demonstration as the input for LLMs to generate the output. Based on task demonstrations, LLMs can recognize and perform a new task without explicit gradient update. [|assistant|]
๊ธฐ์กด์˜ ์ค‘๊ตญ์–ด ์ถœ๋ ฅ. ๋˜ํ•œ, ์ง€์†์ ์ธ pre-training ํ›„์—๋„ ๋ชจ๋ธ์˜ ์ง€์‹์€ ์ฃผ๋กœ ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š์ง€๋งŒ ์‹ ๋ขฐ๋„๋Š” ๋–จ์–ด์ง„๋‹ค.[|endofturn|]
๊ธฐ์กด์˜ ์ค‘๊ตญ์–ด ์ถœ๋ ฅ. ๋˜ํ•œ, ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ํ›„์—๋„ ๋ชจ๋ธ์˜ ์ง€์‹์€ ์ฃผ๋กœ ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š์ง€๋งŒ ์‹ ๋ขฐ๋„๋Š” ๋–จ์–ด์ง„๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Traditional Chinese outputs. Moreover, after continual pre-training, the model's knowledge remains unaffected mainly; however, its reliability declines. [|assistant|]
**Closed-book QA:** ์‚ฌ์ „ ํ•™์Šต๋œ ์ง€์‹๋งŒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ˜„์‹ค ์„ธ๊ณ„์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•ฉ๋‹ˆ๋‹ค. ์˜์–ด์˜ ๊ฒฝ์šฐ TriviaQA [23]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด์˜ ๊ฒฝ์šฐ, TriviaQA ํ˜•์‹์— ๋”ฐ๋ผ NLPCC-KBQA [10] ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ Closed-book QA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**ํ์‡„ํ˜• QA:** ์‚ฌ์ „ ํ•™์Šต๋œ ์ง€์‹๋งŒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ˜„์‹ค ์„ธ๊ณ„์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•ฉ๋‹ˆ๋‹ค. ์˜์–ด์˜ ๊ฒฝ์šฐ TriviaQA [23]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด์˜ ๊ฒฝ์šฐ, TriviaQA ํ˜•์‹์— ๋”ฐ๋ผ NLPCC-KBQA [10] ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ์‡„ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Closed-book QA:** Answering questions about the real world based solely on the pretrained knowledge. We use TriviaQA [23] for English. For Chinese, we build a closed-book QA dataset based on NLPCC-KBQA [10] dataset following the TriviaQA format. [|assistant|]
**10% Steps are Enough.** ๋‘ ๋‹จ๊ณ„ ํ›ˆ๋ จ ๊ด€์ ์—์„œ, decay ๋‹จ๊ณ„๋ฅผ ์งง๊ฒŒ ํ•˜๋ฉด ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ์˜ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๋น ๋ฅด๊ฒŒ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ ํฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋™์ผํ•œ ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ ์ฒดํฌํฌ์ธํŠธ์—์„œ ์‹œ์ž‘ํ•˜์—ฌ decay ๋‹จ๊ณ„๊ฐ€ ๋‹ค๋ฅธ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ทธ๋ฆผ 5์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด 40N, 60N, 80N ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ์„ธ ๊ฐ€์ง€ ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ ์ฒดํฌํฌ์ธํŠธ ๋ชจ๋‘์—์„œ ์ „์ฒด ํ† ํฐ์˜ 10%์˜ decay๋งŒ์œผ๋กœ๋„ ์ตœ์ƒ์˜ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ „์ฒด ํ† ํฐ์˜ 2.5%์˜ decay๋Š” ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ›„์† ํ›ˆ๋ จ ์‹คํ—˜์—์„œ๋Š” ์™„์ „ํ•œ ์ˆ˜๋ ด์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์•ฝ 10%์˜ decay๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**10% ๋‹จ๊ณ„๊ฐ€ ์ถฉ๋ถ„ํ•ฉ๋‹ˆ๋‹ค.** ๋‘ ๋‹จ๊ณ„ ํ›ˆ๋ จ ๊ด€์ ์—์„œ, ์‡ ํ‡ด ๋‹จ๊ณ„๋ฅผ ์งง๊ฒŒ ํ•˜๋ฉด ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ์˜ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๋น ๋ฅด๊ฒŒ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ ํฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋™์ผํ•œ ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ ์ฒดํฌํฌ์ธํŠธ์—์„œ ์‹œ์ž‘ํ•˜์—ฌ ์‡ ํ‡ด ๋‹จ๊ณ„๊ฐ€ ๋‹ค๋ฅธ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ทธ๋ฆผ 5์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด 40N, 60N, 80N ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ์„ธ ๊ฐ€์ง€ ์•ˆ์ •์ ์ธ ํ›ˆ๋ จ ์ฒดํฌํฌ์ธํŠธ ๋ชจ๋‘์—์„œ ์ „์ฒด ํ† ํฐ์˜ 10%์˜ ๊ฐ์†Œ๋งŒ์œผ๋กœ๋„ ์ตœ์ƒ์˜ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ „์ฒด ํ† ํฐ์˜ 2.5%์˜ ๊ฐ์†Œ๋Š” ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ›„์† ํ›ˆ๋ จ ์‹คํ—˜์—์„œ๋Š” ์™„์ „ํ•œ ์ˆ˜๋ ด์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์•ฝ 10%์˜ ๊ฐ์†Œ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**10% Steps are Enough.** From the two stage training perspective, shortening the decay stage will greatly benefit the fast test of different model checkpoints of stable training. Therefore, we conduct experiments that start from the same stable training checkpoints and have different decay steps. Also shown in Figure 5, among all three stable training checkpoints in 40N, 60N, and 80N training data, having a decay of 10% of the total tokens is sufficient to achieve the best results, while a decay of 2.5% of total tokens falls short. Therefore, in the subsequent training experiments, we use a decay of about 10% to ensure full convergence. [|assistant|]
(1) Relevance: ๋ชจ๋ธ์ด ๋ฌธ๋งฅ๊ณผ ์งˆ๋ฌธ์˜ ์˜๋ฏธ๋ก ์  ์˜๋ฏธ๋ฅผ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ•ด์„ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
(1) ๊ด€๋ จ์„ฑ: ๋ชจ๋ธ์ด ๋ฌธ๋งฅ๊ณผ ์งˆ๋ฌธ์˜ ์˜๋ฏธ๋ก ์  ์˜๋ฏธ๋ฅผ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ•ด์„ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|](1) Relevance: Assessing the model's ability to correctly interpret the semantic meaning of the context and questions. [|assistant|]
๋ฐ์ดํ„ฐ ์ฝ๊ธฐ. CommonCrawl์€ WARC(์›์‹œ HTML ์‘๋‹ต) ๋˜๋Š” WET ํŒŒ์ผ(์ผ๋ฐ˜ ํ…์ŠคํŠธ๋งŒ ํฌํ•จํ•˜๋„๋ก ์‚ฌ์ „ ์ฒ˜๋ฆฌ๋œ)๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ๊ฐœ๋ณ„ ํŒŒ์ผ์€ ํŠน์ • URL์˜ ํŽ˜์ด์ง€์— ํ•ด๋‹นํ•˜๋ฉฐ, ์ด๋Š” ๋‹จ์ผ ๋ฌธ์„œ/์ƒ˜ํ”Œ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. WET ํŒŒ์ผ์„ ์‚ฌ์šฉํ•˜๋ฉด HTML ์ถ”์ถœ์„ ์ง์ ‘ ์‹คํ–‰ํ•  ํ•„์š”๊ฐ€ ์—†์ง€๋งŒ, ์ด์ „ ์—ฐ๊ตฌ(Gao et al., 2020; Rae et al., 2021)์— ๋”ฐ๋ฅด๋ฉด WET ํŒŒ์ผ์— ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์€ ํƒ์ƒ‰ ๋ฉ”๋‰ด, ๊ด‘๊ณ  ๋ฐ ๊ธฐํƒ€ ๊ด€๋ จ ์—†๋Š” ํ…์ŠคํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํŒŒ์ดํ”„๋ผ์ธ์€ wavcio ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋กœ ์ฝ์€ ์›์‹œ WARC ํŒŒ์ผ์—์„œ ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋ฐ์ดํ„ฐ ์ฝ๊ธฐ.CommonCrawl์€ WARC(์›์‹œ HTML ์‘๋‹ต) ๋˜๋Š” WET ํŒŒ์ผ(์ผ๋ฐ˜ ํ…์ŠคํŠธ๋งŒ ํฌํ•จํ•˜๋„๋ก ์‚ฌ์ „ ์ฒ˜๋ฆฌ๋œ)๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ๊ฐœ๋ณ„ ํŒŒ์ผ์€ ํŠน์ • URL์˜ ํŽ˜์ด์ง€์— ํ•ด๋‹นํ•˜๋ฉฐ, ์ด๋Š” ๋‹จ์ผ ๋ฌธ์„œ/์ƒ˜ํ”Œ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. WET ํŒŒ์ผ์„ ์‚ฌ์šฉํ•˜๋ฉด HTML ์ถ”์ถœ์„ ์ง์ ‘ ์‹คํ–‰ํ•  ํ•„์š”๊ฐ€ ์—†์ง€๋งŒ, ์ด์ „ ์—ฐ๊ตฌ(Gao et al., 2020; Rae et al., 2021)์— ๋”ฐ๋ฅด๋ฉด WET ํŒŒ์ผ์— ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์€ ํƒ์ƒ‰ ๋ฉ”๋‰ด, ๊ด‘๊ณ  ๋ฐ ๊ธฐํƒ€ ๊ด€๋ จ ์—†๋Š” ํ…์ŠคํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํŒŒ์ดํ”„๋ผ์ธ์€ wavcio ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋กœ ์ฝ์€ ์›์‹œ WARC ํŒŒ์ผ์—์„œ ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Reading the data.CommonCrawl is available in either WARC (raw HTML response), or WET files (preprocessed to only include plain text). Individual files correspond to a page at a given URL; these constitute single documents/samples. Working with WET files would spare us from running our own HTML extraction; however, in line with previous works (Gao et al., 2020; Rae et al., 2021), we found WET files to include undesirable navigation menus, ads, and other irrelevant texts. Accordingly, our pipeline starts from raw WARC files, read with the wavcio library. [|assistant|]
# ์ถœ๋ ฅ Google ์—ฐ๊ตฌ์›๋“ค์ด ๋ฐœํ‘œํ•œ ๋…ผ๋ฌธ "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models"[34]์—์„œ ์ฒ˜์Œ ์„ค๋ช…ํ•œ Chain of Thought (CoT) ๊ธฐ๋ฒ•์€ Large Language Models (LLMs)์„ ์œ„ํ•œ prompt engineering์˜ ์ค‘์ถ”์ ์ธ ๋ฐœ์ „์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ํ† ํฐ ์˜ˆ์ธก์— ๋Šฅ์ˆ™ํ•˜์ง€๋งŒ ๋ช…์‹œ์ ์ธ ์ถ”๋ก ์„ ์œ„ํ•ด ๊ณ ์•ˆ๋œ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ๋Š” ์ ์„ ์ดํ•ดํ•˜๋Š” ๋ฐ ๊ธฐ๋ฐ˜ํ•ฉ๋‹ˆ๋‹ค. CoT๋Š” ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ์„ ํ•„์ˆ˜์ ์ธ ์ถ”๋ก  ๋‹จ๊ณ„๋กœ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
Google ์—ฐ๊ตฌ์›๋“ค์ด ๋ฐœํ‘œํ•œ ๋…ผ๋ฌธ "์‚ฌ๊ณ ์˜ ์‚ฌ์Šฌ ํ”„๋กฌํ”„ํŠธ๊ฐ€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์— ์ถ”๋ก ์„ ์ด๋Œ์–ด๋‚ธ๋‹ค"[34]์—์„œ ์ฒ˜์Œ ์„ค๋ช…ํ•œ ์‚ฌ๊ณ ์˜ ์‚ฌ์Šฌ(CoT) ๊ธฐ๋ฒ•์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์„ ์œ„ํ•œ ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง์˜ ์ค‘์ถ”์ ์ธ ๋ฐœ์ „์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ํ† ํฐ ์˜ˆ์ธก์— ๋Šฅ์ˆ™ํ•˜์ง€๋งŒ ๋ช…์‹œ์ ์ธ ์ถ”๋ก ์„ ์œ„ํ•ด ๊ณ ์•ˆ๋œ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ๋Š” ์ ์„ ์ดํ•ดํ•˜๋Š” ๋ฐ ๊ธฐ๋ฐ˜ํ•ฉ๋‹ˆ๋‹ค. CoT๋Š” ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ์„ ํ•„์ˆ˜์ ์ธ ์ถ”๋ก  ๋‹จ๊ณ„๋กœ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The Chain of Thought (CoT) technique, initially described in the paper "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models"[34] by Google researchers, represents a pivotal advancement in prompt engineering for Large Language Models (LLMs). This approach hinges on the understanding that LLMs, while proficient in token prediction, are not inherently designed for explicit reasoning. CoT addresses this by guiding the model through essential reasoning steps. [|assistant|]
์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด Supervised Fine-Tuning (SFT)์„ ํ†ตํ•ด ํŠน์ • ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋งž๊ฒŒ ๊ฐœ์„ ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค:[|endofturn|]
์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •(SFT)์„ ํ†ตํ•ด ํŠน์ • ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋งž๊ฒŒ ๊ฐœ์„ ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค:[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The pre-trained language model is often enhanced for specific downstream tasks through Supervised Fine-Tuning (SFT) as follows: [|assistant|]
์†Œ์Šค ๋ชจ๋ธ์ผ๋ณธ์–ด๋กœ ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ํ’€ ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ณธ์–ด LLM๊ณผ ์ˆ˜ํ•™ LLM์ด ํฌํ•จ๋œ ์†Œ์Šค ๋ชจ๋ธ ์„ธํŠธ์— ์ง„ํ™” ๋ชจ๋ธ ๋ณ‘ํ•ฉ์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค: shisa-gamma-7b-v1 [3](์ผ๋ณธ์–ด LLM), WizardMath-7B-V1.1 [33] ๋ฐ Abel-7B-002 [6]. ์ด ๋ชจ๋“  ๋ชจ๋ธ์€ Mistral-7B-v0.1 [22]์—์„œ Fine-tuning๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์†Œ์Šค ๋ชจ๋ธ์ผ๋ณธ์–ด๋กœ ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ํ’€ ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ณธ์–ด LLM๊ณผ ์ˆ˜ํ•™ LLM์ด ํฌํ•จ๋œ ์†Œ์Šค ๋ชจ๋ธ ์„ธํŠธ์— ์ง„ํ™” ๋ชจ๋ธ ๋ณ‘ํ•ฉ์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค: shisa-gamma-7b-v1 [3](์ผ๋ณธ์–ด LLM), WizardMath-7B-V1.1 [33] ๋ฐ Abel-7B-002 [6]. ์ด ๋ชจ๋“  ๋ชจ๋ธ์€ Mistral-7B-v0.1 [22]์—์„œ ๋ฏธ์„ธ ์กฐ์ •๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Source ModelsTo develop a model capable of solving math problems in Japanese, we apply evolutionary model merge on a set of source models containing a Japanese LLM and Math LLMs: shisa-gamma-7b-v1 [3] (Japanese LLM), WizardMath-7B-V1.1 [33] and Abel-7B-002 [6]. All these models are fine-tuned from Mistral-7B-v0.1 [22]. [|assistant|]
์š”์•ฝ์„ ํด๋ ˆ์ž„์œผ๋กœ ๋ถ„ํ•ดํ•˜๊ธฐ:์žฅ๋ฌธ ์š”์•ฝ์˜ ์ถฉ์‹ค๋„ ํ‰๊ฐ€์— ๋Œ€ํ•œ ์„ ํ–‰ ์—ฐ๊ตฌ(Krishna et al., 2023; Min et al., 2023; Wei et al., 2024)์— ๋”ฐ๋ผ ์š”์•ฝ์„ _์›์ž์  ์ฃผ์žฅ_์œผ๋กœ ๋ถ„ํ•ดํ•˜์—ฌ ์„ธ๋ถ„ํ™”๋œ ์ฃผ์„์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๊ฐ€์ง€ ์ฃผ์š” Instruction์„ ํ†ตํ•ด LLM(GPT-4)์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋ƒ…๋‹ˆ๋‹ค: (1) ๊ฐ ์›์ž์  ์ฃผ์žฅ์€ ์š”์•ฝ์˜ ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ(์˜ˆ: ๋Œ€๋ช…์‚ฌ ํ•ด๊ฒฐ) ์—†์ด๋„ ๊ทธ ์ž์ฒด๋กœ ์™„์ „ํžˆ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•˜๋ฉฐ, (2) ๊ฐ€๋Šฅํ•˜๋ฉด ๊ฐ ์ฃผ์žฅ์ด ๊ด€๋ จ ์‹œ๊ฐ„์ , ์œ„์น˜์ , ์ธ๊ณผ์  ๋งฅ๋ฝ ๋‚ด์— ์œ„์น˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ถ”์ถœ๋œ 100๊ฐœ์˜ ์ฃผ์žฅ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ์ €์ž์˜ ์ธ๊ฐ„ ๊ฒ€์ฆ ๊ฒฐ๊ณผ ์ •ํ™•๋„๊ฐ€ 100%(์ฆ‰, ๊ฐ ์ฃผ์žฅ์€ ์ถ”๊ฐ€ ๋˜๋Š” ์ž˜๋ชป๋œ ์ •๋ณด ์—†์ด ์š”์•ฝ์—์„œ ์ถ”์ ํ•  ์ˆ˜ ์žˆ์Œ)๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์š”์•ฝ๊ณผ ์ถ”์ถœ๋œ ์ฃผ์žฅ์˜ ์˜ˆ๋Š” ๊ทธ๋ฆผ 2๋ฅผ ์ฐธ์กฐํ•˜์‹œ๊ณ , ์ •ํ™•ํ•œ ํ”„๋กฌํ”„ํŠธ๋Š” SB๋ฅผ ์ฐธ์กฐํ•˜์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค.[|endofturn|]
์š”์•ฝ์„ ํด๋ ˆ์ž„์œผ๋กœ ๋ถ„ํ•ดํ•˜๊ธฐ:์žฅ๋ฌธ ์š”์•ฝ์˜ ์ถฉ์‹ค๋„ ํ‰๊ฐ€์— ๋Œ€ํ•œ ์„ ํ–‰ ์—ฐ๊ตฌ(Krishna et al., 2023; Min et al., 2023; Wei et al., 2024)์— ๋”ฐ๋ผ ์š”์•ฝ์„ _์›์ž์  ์ฃผ์žฅ_์œผ๋กœ ๋ถ„ํ•ดํ•˜์—ฌ ์„ธ๋ถ„ํ™”๋œ ์ฃผ์„์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์ง€์นจ์„ ํ†ตํ•ด LLM(GPT-4)์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋ƒ…๋‹ˆ๋‹ค: (1) ๊ฐ ์›์ž์  ์ฃผ์žฅ์€ ์š”์•ฝ์˜ ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ(์˜ˆ: ๋Œ€๋ช…์‚ฌ ํ•ด๊ฒฐ) ์—†์ด๋„ ๊ทธ ์ž์ฒด๋กœ ์™„์ „ํžˆ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•˜๋ฉฐ, (2) ๊ฐ€๋Šฅํ•˜๋ฉด ๊ฐ ์ฃผ์žฅ์ด ๊ด€๋ จ ์‹œ๊ฐ„์ , ์œ„์น˜์ , ์ธ๊ณผ์  ๋งฅ๋ฝ ๋‚ด์— ์œ„์น˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ถ”์ถœ๋œ 100๊ฐœ์˜ ์ฃผ์žฅ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ์ €์ž์˜ ์ธ๊ฐ„ ๊ฒ€์ฆ ๊ฒฐ๊ณผ ์ •ํ™•๋„๊ฐ€ 100%(์ฆ‰, ๊ฐ ์ฃผ์žฅ์€ ์ถ”๊ฐ€ ๋˜๋Š” ์ž˜๋ชป๋œ ์ •๋ณด ์—†์ด ์š”์•ฝ์—์„œ ์ถ”์ ํ•  ์ˆ˜ ์žˆ์Œ)๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์š”์•ฝ๊ณผ ์ถ”์ถœ๋œ ์ฃผ์žฅ์˜ ์˜ˆ๋Š” ๊ทธ๋ฆผ 2๋ฅผ ์ฐธ์กฐํ•˜์‹œ๊ณ , ์ •ํ™•ํ•œ ํ”„๋กฌํ”„ํŠธ๋Š” SB๋ฅผ ์ฐธ์กฐํ•˜์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Decomposing summaries into claims:Following prior works on evaluating long-form summary faithfulness (Krishna et al., 2023; Min et al., 2023; Wei et al., 2024), we decompose our summaries into _atomic claims_ to enable fine-grained annotation. We prompt an LLM (GPT-4) with two primary instructions: (1) each atomic claim must be fully understandable on its own without requiring additional context from the summary (e.g., resolved pronouns), and (2) whenever possible, each claim should be situated within its relevant temporal, locational, and causal context. Human validation by the authors of a random sample of 100 extracted claims demonstrated 100% precision (i.e., each claim can be traced to the summary without any extra or incorrect information). See Figure 2 for example of summary and its extracted claims; see SB for exact prompt. [|assistant|]
**LM-guided CoT ์„ฑ๋Šฅ.** ํ‘œ 2๋Š” ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์„ ์ข…ํ•ฉ์ ์œผ๋กœ ๋ถ„์„ํ•œ ํ‘œ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์ƒ˜ํ”Œ๋งํ•œ ๋‹ค์Œ ๊ฐ€์žฅ ์ตœ์ ์˜ ๊ทผ๊ฑฐ๋ฅผ ์ˆœ์œ„ ๋งค๊ธฐ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํ™•์žฅ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. KD๋งŒ ์‚ฌ์šฉํ•œ ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ HotpotQA์˜ ๊ฒฝ์šฐ 2%, 2WikiMultiHopQA์˜ ๊ฒฝ์šฐ 10%๋กœ ๊ฐ๊ฐ ์›๋ž˜์˜ CoT prompting๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2๋Š” ๋ชจ๋“  prompting ๊ธฐ๋ฒ•์˜ ๊ฐ๊ฐ์˜ ๊ทผ๊ฑฐ ํ’ˆ์งˆ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ๋‹ต๋ณ€ ์˜ˆ์ธก๊ณผ ๊ทผ๊ฑฐ ํ’ˆ์งˆ์„ ๋ชจ๋‘ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ํšจ๊ณผ๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๊ธด ๋ฌธ๋งฅ์ด ์žˆ๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•ด ์›๋ž˜์˜ CoT prompting์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ, ๋ชจ๋ธ์€ ์ œ๊ณต๋œ ๋ฌธ๋งฅ์—์„œ ๋ฌธ์žฅ์„ ์žฌํ™œ์šฉํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๊ฒฐ์ •์ ์ธ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถ”์„ธ๋Š” ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์— ์˜ํ•ด ์™„ํ™”๋˜์–ด ์˜ค๋ฅ˜์œจ์ด ํฌ๊ฒŒ ๊ฐ์†Œํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ CoT prompting + SC์˜ ์„ฑ๋Šฅ์„ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ EM ๋ฐ F1 ์ธก๋ฉด์—์„œ ํ‘œ์ค€ prompting๊ณผ ๋™๋“ฑํ•ฉ๋‹ˆ๋‹ค. ๋‹ต๋ณ€ ํฌํ•จ ์ ์ˆ˜์˜ ๊ฒฝ์šฐ LM-guided CoT prompting์ด ํ‘œ์ค€ prompting๋ณด๋‹ค ์•ฝ๊ฐ„ ๋†’์Šต๋‹ˆ๋‹ค(1~2%). ๋˜ํ•œ LM-guided CoT prompting + SC๋Š” ๋ชจ๋“  ์„ค์ •์—์„œ ๊ฐ€์žฅ ๋†’์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**LM ๊ฐ€์ด๋“œ CoT ์„ฑ๋Šฅ.** ํ‘œ 2๋Š” ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์„ ์ข…ํ•ฉ์ ์œผ๋กœ ๋ถ„์„ํ•œ ํ‘œ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์ƒ˜ํ”Œ๋งํ•œ ๋‹ค์Œ ๊ฐ€์žฅ ์ตœ์ ์˜ ๊ทผ๊ฑฐ๋ฅผ ์ˆœ์œ„ ๋งค๊ธฐ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํ™•์žฅ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. KD๋งŒ ์‚ฌ์šฉํ•œ ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ HotpotQA์˜ ๊ฒฝ์šฐ 2%, 2WikiMultiHopQA์˜ ๊ฒฝ์šฐ 10%๋กœ ๊ฐ๊ฐ ์›๋ž˜์˜ CoT ํ”„๋กฌํ”„ํŠธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2๋Š” ๋ชจ๋“  ํ”„๋กฌํ”„ํŠธ ๊ธฐ๋ฒ•์˜ ๊ฐ๊ฐ์˜ ๊ทผ๊ฑฐ ํ’ˆ์งˆ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ๋‹ต๋ณ€ ์˜ˆ์ธก๊ณผ ๊ทผ๊ฑฐ ํ’ˆ์งˆ์„ ๋ชจ๋‘ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ํšจ๊ณผ๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๊ธด ๋ฌธ๋งฅ์ด ์žˆ๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•ด ์›๋ž˜์˜ CoT ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ, ๋ชจ๋ธ์€ ์ œ๊ณต๋œ ๋ฌธ๋งฅ์—์„œ ๋ฌธ์žฅ์„ ์žฌํ™œ์šฉํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๊ฒฐ์ •์ ์ธ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถ”์„ธ๋Š” ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์— ์˜ํ•ด ์™„ํ™”๋˜์–ด ์˜ค๋ฅ˜์œจ์ด ํฌ๊ฒŒ ๊ฐ์†Œํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ CoT ํ”„๋กฌํ”„ํŠธ + SC์˜ ์„ฑ๋Šฅ์„ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ EM ๋ฐ F1 ์ธก๋ฉด์—์„œ ํ‘œ์ค€ ํ”„๋กฌํ”„ํŠธ์™€ ๋™๋“ฑํ•ฉ๋‹ˆ๋‹ค. ๋‹ต๋ณ€ ํฌํ•จ ์ ์ˆ˜์˜ ๊ฒฝ์šฐ LM ๊ฐ€์ด๋“œ CoT ํ”„๋กฌํ”„ํŠธ๊ฐ€ ํ‘œ์ค€ ํ”„๋กฌํ”„ํŠธ๋ณด๋‹ค ์•ฝ๊ฐ„ ๋†’์Šต๋‹ˆ๋‹ค(1~2%). ๋˜ํ•œ LM ๊ฐ€์ด๋“œ CoT ํ”„๋กฌํ”„ํŠธ + SC๋Š” ๋ชจ๋“  ์„ค์ •์—์„œ ๊ฐ€์žฅ ๋†’์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**LM-guided CoT Performance.** Table 2 shows a comprehensive breakdown of our method's performance. Additionally, we explore an extension of our approach, which involves sampling multiple reasoning paths and subsequently ranking the most optimal rationales. Our method with only KD outperforms the original CoT prompting with 2% gain for HotpotQA and 10% for 2WikiMultiHopQA, respectively. Figure 2 illustrates the respective rationale qualities of all prompting techniques, reinforcing the effectiveness of our method in enhancing both answer prediction and rationale qualities. When employing the original CoT prompting for questions with lengthy contexts, models frequently recycle sentences from the provided context and struggle to deliver a conclusive answer to the question. This trend is mitigated by our approach, resulting in a significant decrease in error rates. It also surpasses the performance of CoT prompting + SC and is on par with standard prompting in terms of EM and F1. For the answer inclusion score, LM-guided CoT prompting is slightly higher (1-2%) than standard prompting. Furthermore, LM-guided CoT prompting + SC achieves the highest performance across all settings. [|assistant|]
**CPO Loss Components** CPO Loss function์€ ์„ ํ˜ธ๋„ ํ•™์Šต์„ ์œ„ํ•œ \(\mathcal{L}_{\text{prefer}}\)์™€ ๋ชจ๋ธ์ด ์„ ํ˜ธํ•˜๋Š” ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์—์„œ ํฌ๊ฒŒ ๋ฒ—์–ด๋‚˜์ง€ ์•Š๋„๋ก ํ•˜๋Š” \(\mathcal{L}_{\text{NLL}}\)์˜ ๋‘ ๊ฐ€์ง€ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ํ•ญ์˜ ์ค‘์š”์„ฑ์„ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ์„ ๊ตฌ์„ฑ ์š”์†Œ ์ค‘ ํ•˜๋‚˜๋งŒ์œผ๋กœ๋งŒ ๋‹ค์‹œ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ค‘์š”ํ•œ ์ ์€ \(\mathcal{L}_{\text{NLL}}\)๋งŒ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ์„ ํ˜ธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ Supervised Fine-tuning (SFT)์˜ ๊ธฐ๋ณธ ์‹œ๋‚˜๋ฆฌ์˜ค์™€ ๊ฐ™๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 4์˜ ์™ผ์ชฝ์— ํ‘œ์‹œ๋œ ๊ฒƒ์ฒ˜๋Ÿผ ๋‘ ํ•ญ์„ ๋ชจ๋‘ ํฌํ•จํ•˜๋ฉด ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ์–ด๋Š ํ•˜๋‚˜๋ผ๋„ ๋ˆ„๋ฝํ•˜๋ฉด ์„ฑ๋Šฅ์ด ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
**CPO ์†์‹ค ๊ตฌ์„ฑ ์š”์†Œ** CPO ์†์‹ค ํ•จ์ˆ˜๋Š” ์„ ํ˜ธ๋„ ํ•™์Šต์„ ์œ„ํ•œ \(\mathcal{L}_{\text{prefer}}\)์™€ ๋ชจ๋ธ์ด ์„ ํ˜ธํ•˜๋Š” ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์—์„œ ํฌ๊ฒŒ ๋ฒ—์–ด๋‚˜์ง€ ์•Š๋„๋ก ํ•˜๋Š” \(\mathcal{L}_{\text{NLL}}\)์˜ ๋‘ ๊ฐ€์ง€ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ํ•ญ์˜ ์ค‘์š”์„ฑ์„ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ์„ ๊ตฌ์„ฑ ์š”์†Œ ์ค‘ ํ•˜๋‚˜๋งŒ์œผ๋กœ๋งŒ ๋‹ค์‹œ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ค‘์š”ํ•œ ์ ์€ \(\mathcal{L}_{\text{NLL}}\)๋งŒ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ์„ ํ˜ธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ SFT์˜ ๊ธฐ๋ณธ ์‹œ๋‚˜๋ฆฌ์˜ค์™€ ๊ฐ™๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 4์˜ ์™ผ์ชฝ์— ํ‘œ์‹œ๋œ ๊ฒƒ์ฒ˜๋Ÿผ ๋‘ ํ•ญ์„ ๋ชจ๋‘ ํฌํ•จํ•˜๋ฉด ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ์–ด๋Š ํ•˜๋‚˜๋ผ๋„ ๋ˆ„๋ฝํ•˜๋ฉด ์„ฑ๋Šฅ์ด ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**CPO Loss Components** The CPO loss function consists of two components: \(\mathcal{L}_{\text{prefer}}\) for preference learning, and \(\mathcal{L}_{\text{NLL}}\), which ensures the model does not deviate significantly from the preferred data distribution. To illustrate the significance of each term, we re-train the model exclusively with one of the components. It is important to note that training solely with \(\mathcal{L}_{\text{NLL}}\) equates to the baseline scenario of SFT on preferred data. As depicted in the left of Figure 4, the inclusion of both terms yields the optimal performance, while the absence of either leads to a decrease in performance. [|assistant|]
URL filtering์€ ์‚ฌ์šฉ์ž์—๊ฒŒ ํ•ด๋กœ์šธ ์ˆ˜ ์žˆ๊ฑฐ๋‚˜ ๋Œ€๋ถ€๋ถ„์˜ ๊ฒฝ์šฐ ๋น„์ •ํ˜• ํ…์ŠคํŠธ/์ŠคํŒธ์„ ํฌํ•จํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€(์˜ˆ: ํŒŒ์ผ ํ˜ธ์ŠคํŒ… ์›น์‚ฌ์ดํŠธ) ์„ฑ์ธ ์ฝ˜ํ…์ธ ์™€ ๊ด€๋ จ๋œ ๋„๋ฉ”์ธ์„ ์ฐพ๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ๋จผ์ €, ๋ถ€๋ก G.1.1์— ์ž์„ธํžˆ ์„ค๋ช…๋œ 460๋งŒ ๊ฐœ์˜ ๋„๋ฉ”์ธ ๋ชฉ๋ก์„ ์ˆ˜์ง‘ํ•˜์—ฌ ๋ช…์‹œ์ ์œผ๋กœ ์ฐจ๋‹จํ•œ ๋‹ค์Œ, URL์˜ ํ•˜์œ„ ๋‹จ์–ด์™€ ํ๋ ˆ์ด์…˜ํ•œ ๋‹จ์–ด ๋ชฉ๋ก(๋ถ€๋ก G.1.2 ์ฐธ์กฐ)์„ ์ผ์น˜์‹œํ‚ค๋Š” ๊ฐ„๋‹จํ•œ URL scoring ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋‹จ์–ด ๋ชฉ๋ก์€ ์ˆ˜๋™ ๊ฒ€์‚ฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ๋ ˆ์ด์…˜ํ–ˆ์œผ๋ฉฐ, ๋…์„ฑ ์ธก๋ฉด์—์„œ ์ด์ƒ๊ฐ’์œผ๋กœ ๋ฐํ˜€์ง„ ํŽ˜์ด์ง€์™€์˜ ๊ฒฐ๊ณผ๋ฅผ ToxicBERT์™€ ๊ต์ฐจ ์ฐธ์กฐํ•˜์—ฌ ํ๋ ˆ์ด์…˜ํ–ˆ์Šต๋‹ˆ๋‹ค(Hanu ๋ฐ Unitary ํŒ€, 2020).[|endofturn|]
URL ํ•„ํ„ฐ๋ง์€ ์‚ฌ์šฉ์ž์—๊ฒŒ ํ•ด๋กœ์šธ ์ˆ˜ ์žˆ๊ฑฐ๋‚˜ ๋Œ€๋ถ€๋ถ„์˜ ๊ฒฝ์šฐ ๋น„์ •ํ˜• ํ…์ŠคํŠธ/์ŠคํŒธ์„ ํฌํ•จํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€(์˜ˆ: ํŒŒ์ผ ํ˜ธ์ŠคํŒ… ์›น์‚ฌ์ดํŠธ) ์„ฑ์ธ ์ฝ˜ํ…์ธ ์™€ ๊ด€๋ จ๋œ ๋„๋ฉ”์ธ์„ ์ฐพ๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ๋จผ์ €, ๋ถ€๋ก G.1.1์— ์ž์„ธํžˆ ์„ค๋ช…๋œ 460๋งŒ ๊ฐœ์˜ ๋„๋ฉ”์ธ ๋ชฉ๋ก์„ ์ˆ˜์ง‘ํ•˜์—ฌ ๋ช…์‹œ์ ์œผ๋กœ ์ฐจ๋‹จํ•œ ๋‹ค์Œ, URL์˜ ํ•˜์œ„ ๋‹จ์–ด์™€ ํ๋ ˆ์ด์…˜ํ•œ ๋‹จ์–ด ๋ชฉ๋ก(๋ถ€๋ก G.1.2 ์ฐธ์กฐ)์„ ์ผ์น˜์‹œํ‚ค๋Š” ๊ฐ„๋‹จํ•œ URL ์ ์ˆ˜ ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋‹จ์–ด ๋ชฉ๋ก์€ ์ˆ˜๋™ ๊ฒ€์‚ฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ๋ ˆ์ด์…˜ํ–ˆ์œผ๋ฉฐ, ๋…์„ฑ ์ธก๋ฉด์—์„œ ์ด์ƒ๊ฐ’์œผ๋กœ ๋ฐํ˜€์ง„ ํŽ˜์ด์ง€์™€์˜ ๊ฒฐ๊ณผ๋ฅผ ToxicBERT์™€ ๊ต์ฐจ ์ฐธ์กฐํ•˜์—ฌ ํ๋ ˆ์ด์…˜ํ–ˆ์Šต๋‹ˆ๋‹ค(Hanu ๋ฐ Unitary ํŒ€, 2020).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Our URL filtering focuses on finding domains that are related to adult content, that may be harmful to users, or that are very likely to contain mostly unstructured text/spam (e.g., file hosting websites). First, we aggregated a list of 4.6M domains, detailed in Appendix G.1.1, that we explicitly ban; then, we built a simple URL scoring system, based on matching subwords in the URL against a list of words we curated (see Appendix G.1.2). We curated this list of words based on manual inspection, cross-referencing results with pages surfaced by ToxicBERT as being outliers in toxicity (Hanu and Unitary team, 2020). [|assistant|]
**์•ˆ์ •์ ์ธ ํ›ˆ๋ จ ๋‹จ๊ณ„.** ์•ฝ 1T ๋ฐ์ดํ„ฐ(๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋Š” ์„น์…˜ 11 ์ฐธ์กฐ)๋ฅผ ํ™œ์šฉํ•˜๋ฉฐ, ๋Œ€๋ถ€๋ถ„์˜ ๋ฐ์ดํ„ฐ๋Š” ์˜คํ”ˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค. ๋ชจ๋ธ ์œˆ๋“œ ํ„ฐ๋„ ์‹คํ—˜์—์„œ ๋ฐœ๊ฒฌ๋œ ์ตœ์ ์˜ ๊ตฌ์„ฑ์ธ WSD LRS๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, Batch Size๋Š” 393๋งŒ ๊ฐœ, ์ตœ๋Œ€ Learning Rate๋Š” 0.01์ž…๋‹ˆ๋‹ค.[|endofturn|]
**์•ˆ์ •์ ์ธ ํ›ˆ๋ จ ๋‹จ๊ณ„.** ์•ฝ 1T ๋ฐ์ดํ„ฐ(๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋Š” ์„น์…˜ 11 ์ฐธ์กฐ)๋ฅผ ํ™œ์šฉํ•˜๋ฉฐ, ๋Œ€๋ถ€๋ถ„์˜ ๋ฐ์ดํ„ฐ๋Š” ์˜คํ”ˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค. ๋ชจ๋ธ ์œˆ๋“œ ํ„ฐ๋„ ์‹คํ—˜์—์„œ ๋ฐœ๊ฒฌ๋œ ์ตœ์ ์˜ ๊ตฌ์„ฑ์ธ WSD LRS๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 393๋งŒ ๊ฐœ, ์ตœ๋Œ€ ํ•™์Šต ์†๋„๋Š” 0.01์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Stable Training Stage.** We utilize around 1T data (see Section 11 for data distribution), with the majority of the data sourced from open datasets. We use the optimal configuration discovered during the model wind tunnel experiments, WSD LRS, with a batch size of 3.93 million and a max learning rate of 0.01. [|assistant|]
**Scraping**. ์˜ˆ์ธก ํ”Œ๋žซํผ์—์„œ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ์ปดํŒŒ์ผํ•˜๊ธฐ ์œ„ํ•ด API๋ฅผ ์ฟผ๋ฆฌํ•˜๊ฑฐ๋‚˜ ์งˆ๋ฌธ์˜ ์›นํŽ˜์ด์ง€๋ฅผ scrapingํ•˜์—ฌ ์ดˆ๊ธฐ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. Metaculus์˜ ๊ฒฝ์šฐ ๋จผ์ € API๋ฅผ ํ†ตํ•ด ๊ธฐ๋ณธ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•˜๊ณ  ์›นํŽ˜์ด์ง€์—์„œ ํ•ด์ƒ๋„ ๊ธฐ์ค€์„ scrapingํ•ฉ๋‹ˆ๋‹ค. INFER(CSET)์™€ Good Judgment Open ๋ฐ์ดํ„ฐ๋Š” ์›น scraping์„ ํ†ตํ•ด ์ˆ˜์ง‘๋˜๋ฉฐ, ํ•„์š”ํ•œ ์ „์ฒด ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ณตํ•˜๋Š” API๊ฐ€ ์—†๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ปค๋ฎค๋‹ˆํ‹ฐ ์˜ˆ์ธก์„ ์ œ์™ธํ•œ Polymarket์˜ ๋ฐ์ดํ„ฐ๋Š” API๋ฅผ ํ†ตํ•ด ์–ป์Šต๋‹ˆ๋‹ค. Manifold์˜ ๋ฐ์ดํ„ฐ๋Š” API๋ฅผ ํ†ตํ•ด ์™„์ „ํžˆ scraping๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
**์Šคํฌ๋ž˜ํ•‘**. ์˜ˆ์ธก ํ”Œ๋žซํผ์—์„œ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ์ปดํŒŒ์ผํ•˜๊ธฐ ์œ„ํ•ด API๋ฅผ ์ฟผ๋ฆฌํ•˜๊ฑฐ๋‚˜ ์งˆ๋ฌธ์˜ ์›นํŽ˜์ด์ง€๋ฅผ ์Šคํฌ๋ž˜ํ•‘ํ•˜์—ฌ ์ดˆ๊ธฐ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋ฉ”ํƒ€์ฟจ๋Ÿฌ์Šค์˜ ๊ฒฝ์šฐ ๋จผ์ € API๋ฅผ ํ†ตํ•ด ๊ธฐ๋ณธ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•˜๊ณ  ์›นํŽ˜์ด์ง€์—์„œ ํ•ด์ƒ๋„ ๊ธฐ์ค€์„ ์Šคํฌ๋ž˜ํ•‘ํ•ฉ๋‹ˆ๋‹ค. INFER(CSET)์™€ Good Judgment Open ๋ฐ์ดํ„ฐ๋Š” ์›น ์Šคํฌ๋ž˜ํ•‘์„ ํ†ตํ•ด ์ˆ˜์ง‘๋˜๋ฉฐ, ํ•„์š”ํ•œ ์ „์ฒด ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ณตํ•˜๋Š” API๊ฐ€ ์—†๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ปค๋ฎค๋‹ˆํ‹ฐ ์˜ˆ์ธก์„ ์ œ์™ธํ•œ ํด๋ฆฌ๋งˆ์ผ“์˜ ๋ฐ์ดํ„ฐ๋Š” API๋ฅผ ํ†ตํ•ด ์–ป์Šต๋‹ˆ๋‹ค. ๋งค๋‹ˆํด๋“œ์˜ ๋ฐ์ดํ„ฐ๋Š” API๋ฅผ ํ†ตํ•ด ์™„์ „ํžˆ ์Šคํฌ๋ž˜ํ•‘๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Scraping.** To compile our dataset from the forecasting platforms, we query their APIs or scrape the questions' webpages for initial data gathering. For Metaculus, we first extract basic information via the API and scrape the resolution criteria from webpage. INFER (CSET) and Good Judgment Open data are gathered via web scraping, since no API provides the full data we need. Polymarket's data, except for community predictions, is obtained from their API. Manifold's data is fully scraped via API. [|assistant|]
๊ธฐ๊ณ„ ๋ฒˆ์—ญ ์ž‘์—…์—์„œ ๋Œ€์ƒ ์ฐธ์กฐ์˜ ์ค‘์š”์„ฑ์€ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ๊ณ„ ๋ฒˆ์—ญ ์ž‘์—…์— ๋Œ€ํ•œ ๋ชจ๋ธ ํ•™์Šต ํŒจ๋Ÿฌ๋‹ค์ž„์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์˜ˆ์ธก๋œ ์ถœ๋ ฅ๊ณผ gold reference ๊ฐ„์˜ ์ฐจ์ด๋ฅผ ์ตœ์†Œํ™”ํ•˜๋„๋ก ์ •์˜๋œ loss๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ์ตœ์ ํ™”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ฐธ์กฐ์˜ ํ’ˆ์งˆ์— ํฌ๊ฒŒ ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ \(\mathcal{D}\), ์†Œ์Šค ๋ฌธ์žฅ \(x\)์™€ ๊ทธ์— ํ•ด๋‹นํ•˜๋Š” ๋Œ€์ƒ ๋ฌธ์žฅ(gold reference) \(y\)์˜ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, \(\mathcal{D}=\left\{x^{(i)},y^{(i)}\right\}_{i=1}^{N}\), ์—ฌ๊ธฐ์„œ \(N\)์€ ์ด ๋ณ‘๋ ฌ ๋ฌธ์žฅ ์ˆ˜์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ธฐ๊ณ„ ๋ฒˆ์—ญ ์ž‘์—…์—์„œ ๋Œ€์ƒ ์ฐธ์กฐ์˜ ์ค‘์š”์„ฑ์€ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ๊ณ„ ๋ฒˆ์—ญ ์ž‘์—…์— ๋Œ€ํ•œ ๋ชจ๋ธ ํ•™์Šต ํŒจ๋Ÿฌ๋‹ค์ž„์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์˜ˆ์ธก๋œ ์ถœ๋ ฅ๊ณผ ๊ณจ๋“œ ์ฐธ์กฐ ๊ฐ„์˜ ์ฐจ์ด๋ฅผ ์ตœ์†Œํ™”ํ•˜๋„๋ก ์ •์˜๋œ ์†์‹ค์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ์ตœ์ ํ™”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ฐธ์กฐ์˜ ํ’ˆ์งˆ์— ํฌ๊ฒŒ ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ \mathcal{D}\), ์†Œ์Šค ๋ฌธ์žฅ \(x\)์™€ ๊ทธ์— ํ•ด๋‹นํ•˜๋Š” ๋Œ€์ƒ ๋ฌธ์žฅ(๊ณจ๋“œ ์ฐธ์กฐ) \(y\)์˜ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, \(\mathcal{D}=\left\{x^{(i)},y^{(i)}\right\}_{i=1}^{N}\), ์—ฌ๊ธฐ์„œ \(N\)์€ ์ด ๋ณ‘๋ ฌ ๋ฌธ์žฅ ์ˆ˜์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The significance of target references is paramount in machine translation tasks. The paradigm of training models on the machine translation task heavily relies on the quality of the references since the model is commonly optimized using a loss that is defined to minimize the difference between the predicted outputs and gold reference. Consider a dataset \(\mathcal{D}\), comprising pairs of source sentences \(x\) and their corresponding target sentences (gold references) \(y\), represented as \(\mathcal{D}=\left\{x^{(i)},y^{(i)}\right\}_{i=1}^{N}\), where \(N\) is the total number of parallel sentences. [|assistant|]
๋ชจ๋ธ๋ณ„ ๋ถ„์„. ์ถ”๋ก  ์œ ํ˜•๊ณผ ์ถ”๋ก -์ฃผ์žฅ ๊ด€๊ณ„์— ๋Œ€ํ•œ ๋ชจ๋ธ๋ณ„ ๊ฒฐ๊ณผ๋Š” ํ‘œ 18๊ณผ ํ‘œ 19์— ๋ณด๊ณ ํ•œ๋‹ค.[|endofturn|]
๋ชจ๋ธ๋ณ„ ๋ถ„์„์ถ”๋ก  ์œ ํ˜•๊ณผ ์ถ”๋ก -์ฃผ์žฅ ๊ด€๊ณ„์— ๋Œ€ํ•œ ๋ชจ๋ธ๋ณ„ ๊ฒฐ๊ณผ๋Š” ํ‘œ 18๊ณผ ํ‘œ 19์— ๋ณด๊ณ ํ•œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Model-wise analysisWe report model-wise results on reasoning type and reasoning-claim relationship in Table 18 and Table 19. [|assistant|]
\(\bullet\)_์œ ์šฉ์„ฑ._ ์œ ์šฉ์„ฑ์„ ์œ„ํ•ด LLM์€ ๊ฐ€๋Šฅํ•œ ํ•œ ๊ฐ„๊ฒฐํ•˜๊ณ  ํšจ์œจ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์šฉ์ž์˜ ์ž‘์—… ํ•ด๊ฒฐ์ด๋‚˜ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์ง€์›ํ•˜๋ ค๋Š” ๋…ธ๋ ฅ์„ ๋ช…ํ™•ํ•˜๊ฒŒ ๋ณด์—ฌ์ค˜์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋” ์ž์„ธํ•œ ์„ค๋ช…์ด ํ•„์š”ํ•œ ๊ฒฝ์šฐ, LLM์€ ์ ์ ˆํ•œ ์งˆ๋ฌธ์„ ํ†ตํ•ด ์ถ”๊ฐ€ ๊ด€๋ จ ์ •๋ณด๋ฅผ ์ด๋Œ์–ด๋‚ด๋Š” ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์–ด์•ผ ํ•˜๋ฉฐ, ์ ์ ˆํ•œ ์ˆ˜์ค€์˜ ๊ฐ์ˆ˜์„ฑ, ํ†ต์ฐฐ๋ ฅ, ์‹ ์ค‘ํ•จ์„ ๋ณด์—ฌ์ค˜์•ผ ํ•ฉ๋‹ˆ๋‹ค[368]. ์œ ์šฉํ•œ ํ–‰๋™์˜ alignment๋ฅผ ์‹คํ˜„ํ•˜๋Š” ๊ฒƒ์€ ์‚ฌ์šฉ์ž์˜ ์˜๋„๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์ •์˜ํ•˜๊ณ  ์ธก์ •ํ•˜๊ธฐ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์— LLM์— ์–ด๋ ค์šด ๊ณผ์ œ์ž…๋‹ˆ๋‹ค[366].[|endofturn|]
\(\bullet\)_์œ ์šฉ์„ฑ._ ์œ ์šฉ์„ฑ์„ ์œ„ํ•ด LLM์€ ๊ฐ€๋Šฅํ•œ ํ•œ ๊ฐ„๊ฒฐํ•˜๊ณ  ํšจ์œจ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์šฉ์ž์˜ ์ž‘์—… ํ•ด๊ฒฐ์ด๋‚˜ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์ง€์›ํ•˜๋ ค๋Š” ๋…ธ๋ ฅ์„ ๋ช…ํ™•ํ•˜๊ฒŒ ๋ณด์—ฌ์ค˜์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋” ์ž์„ธํ•œ ์„ค๋ช…์ด ํ•„์š”ํ•œ ๊ฒฝ์šฐ, LLM์€ ์ ์ ˆํ•œ ์งˆ๋ฌธ์„ ํ†ตํ•ด ์ถ”๊ฐ€ ๊ด€๋ จ ์ •๋ณด๋ฅผ ์ด๋Œ์–ด๋‚ด๋Š” ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์–ด์•ผ ํ•˜๋ฉฐ, ์ ์ ˆํ•œ ์ˆ˜์ค€์˜ ๊ฐ์ˆ˜์„ฑ, ํ†ต์ฐฐ๋ ฅ, ์‹ ์ค‘ํ•จ์„ ๋ณด์—ฌ์ค˜์•ผ ํ•ฉ๋‹ˆ๋‹ค[368]. ์œ ์šฉํ•œ ํ–‰๋™์˜ ์ •๋ ฌ์„ ์‹คํ˜„ํ•˜๋Š” ๊ฒƒ์€ ์‚ฌ์šฉ์ž์˜ ์˜๋„๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์ •์˜ํ•˜๊ณ  ์ธก์ •ํ•˜๊ธฐ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์— LLM์— ์–ด๋ ค์šด ๊ณผ์ œ์ž…๋‹ˆ๋‹ค[366].[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_Helpfulness._ To be helpful, the LLM should demonstrate a clear attempt to assist users in solving their tasks or answering questions in a concise and efficient manner as possible. At a higher level, when further clarification is needed, the LLM should demonstrate the capability of eliciting additional relevant information through pertinent inquiries and exhibit suitable levels of sensitivity, perceptiveness, and prudence [368]. Realizing the alignment of helpful behavior is challenging for LLMs since it is difficult to precisely define and measure the intention of users [366]. [|assistant|]
\begin{table} \begin{tabular}{c c c} \hline \hline ๋ชจ๋ธ ํฌ๊ธฐ & NameToDescription & DescriptionToName \\ \hline 350M & 100 & 96.67 \\ 1.3B & 100 & 100 \\ 6.7B & 100 & 100 \\ 175B & 100 & 100 \\ \hline \hline \end{tabular} \end{table} ํ‘œ 5: ์‹คํ—˜ 1: 3e-3์˜ GPT-3์— ๋Œ€ํ•œ In-context ์ •ํ™•๋„ ๋ฐ Batch Size 32. ํ† ํฐํ™”๋œ ๋ฌธ๊ตฌ "๋‹คํ”„๋„ค ๋ฐฐ๋งํ„ด์€ ๊ฐ€์ƒ ํ˜„์‹ค ๊ฑธ์ž‘์ธ '์‹œ๊ฐ„์„ ํ†ตํ•œ ์—ฌํ–‰'์˜ ๊ฐ๋…์„ ๋งก์€ ์ธ์ •๋ฐ›๋Š” ๊ฐ๋…์ž…๋‹ˆ๋‹ค."์˜ ๋ณ€ํ˜•์œผ๋กœ ์†Œํ”„ํŠธ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค. 10๊ฐœ์˜ ๋ฌด์ž‘์œ„ ์‹œ๋“œ์— ๊ฑธ์ณ ๊ฒฐ๊ณผ๋ฅผ ํ‰๊ท ํ™”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
\begin{table} \begin{tabular}{c c c} \hline \hline ๋ชจ๋ธ ํฌ๊ธฐ & NameToDescription & DescriptionToName \\ \hline 350M & 100 & 96.67 \\ 1.3B & 100 & 100 \\ 6.7B & 100 & 100 \\ 175B & 100 & 100 \\ \hline \hline \end{tabular} \end{table} ํ‘œ 5: ์‹คํ—˜ 1: 3e-3์˜ GPT-3์— ๋Œ€ํ•œ ์ปจํ…์ŠคํŠธ ๋‚ด ์ •ํ™•๋„ ๋ฐ ๋ฐฐ์น˜ ํฌ๊ธฐ 32. ํ† ํฐํ™”๋œ ๋ฌธ๊ตฌ "๋‹คํ”„๋„ค ๋ฐฐ๋งํ„ด์€ ๊ฐ€์ƒ ํ˜„์‹ค ๊ฑธ์ž‘์ธ '์‹œ๊ฐ„์„ ํ†ตํ•œ ์—ฌํ–‰'์˜ ๊ฐ๋…์„ ๋งก์€ ์ธ์ •๋ฐ›๋Š” ๊ฐ๋…์ž…๋‹ˆ๋‹ค."์˜ ๋ณ€ํ˜•์œผ๋กœ ์†Œํ”„ํŠธ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค. 10๊ฐœ์˜ ๋ฌด์ž‘์œ„ ์‹œ๋“œ์— ๊ฑธ์ณ ๊ฒฐ๊ณผ๋ฅผ ํ‰๊ท ํ™”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\begin{table} \begin{tabular}{c c c} \hline \hline Model size & NameToDescription & DescriptionToName \\ \hline 350M & 100 & 96.67 \\ 1.3B & 100 & 100 \\ 6.7B & 100 & 100 \\ 175B & 100 & 100 \\ \hline \hline \end{tabular} \end{table} Table 5: Experiment 1: In-context accuracy for GPT-3of 3e-3 and a batch size of 32. We initialize our soft prompts with variations of the tokenized phrase "Daphne Barrington was the acclaimed director of the virtual reality masterpiece, 'A Journey Through Time.". We average our results accross 10 random seeds. [|assistant|]
๊ฐ„๋‹จํ•œ ๊ฐ€์ค‘์น˜ ๋ณด๊ฐ„์€ ์‹ค์ˆ˜ ๊ฐ’ ์ž ์žฌ ๋ฐ ์ถœ๋ ฅ ๊ณต๊ฐ„์—์„œ ์ž‘๋™ํ•˜๋Š” ์ด๋ฏธ์ง€ ์ƒ์„ฑ ๋ชจ๋ธ์„ ๋ณ‘ํ•ฉํ•˜๋Š” ๋ฐ ์ž˜ ์ž‘๋™ํ•˜์ง€๋งŒ, ๋†€๋ž๊ฒŒ๋„ ์˜ฌ๋ฐ”๋ฅธ ์กฐ๊ฑด์—์„œ ์–ธ์–ด ๋ชจ๋ธ์„ ๋ณ‘ํ•ฉํ•˜๋Š” ๋ฐ์—๋„ ์—ฌ์ „ํžˆ ์ž‘๋™ํ•˜์ง€๋งŒ ์„ฑ๋Šฅ ๋ฌธ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค[49]. ์ตœ๊ทผ์—๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ํŠน๋ณ„ํžˆ ๋ณ‘ํ•ฉํ•˜๊ธฐ ์œ„ํ•œ ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์ธ Task Arithmetic[21]์€ ๋ฏธ๋ฆฌ ํ•™์Šต๋œ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜์—์„œ Fine-tuned ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๋ฅผ ๋นผ์„œ _task vectors_๋ฅผ ๊ตฌ์ถ•ํ•˜์—ฌ ์‚ฐ์ˆ  ์—ฐ์‚ฐ์„ ํ†ตํ•ด ์กฐ์ž‘ํ•˜์—ฌ ๋ณ‘ํ•ฉ๋œ ๋ชจ๋ธ์˜ ๋™์ž‘์„ ์œ ๋„ํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ฐ„๋‹จํ•œ ๊ฐ€์ค‘์น˜ ๋ณด๊ฐ„์€ ์‹ค์ˆ˜ ๊ฐ’ ์ž ์žฌ ๋ฐ ์ถœ๋ ฅ ๊ณต๊ฐ„์—์„œ ์ž‘๋™ํ•˜๋Š” ์ด๋ฏธ์ง€ ์ƒ์„ฑ ๋ชจ๋ธ์„ ๋ณ‘ํ•ฉํ•˜๋Š” ๋ฐ ์ž˜ ์ž‘๋™ํ•˜์ง€๋งŒ, ๋†€๋ž๊ฒŒ๋„ ์˜ฌ๋ฐ”๋ฅธ ์กฐ๊ฑด์—์„œ ์–ธ์–ด ๋ชจ๋ธ์„ ๋ณ‘ํ•ฉํ•˜๋Š” ๋ฐ์—๋„ ์—ฌ์ „ํžˆ ์ž‘๋™ํ•˜์ง€๋งŒ ์„ฑ๋Šฅ ๋ฌธ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค[49]. ์ตœ๊ทผ์—๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ํŠน๋ณ„ํžˆ ๋ณ‘ํ•ฉํ•˜๊ธฐ ์œ„ํ•œ ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์ธ ์ž‘์—… ์‚ฐ์ˆ [21]์€ ๋ฏธ๋ฆฌ ํ•™์Šต๋œ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜์—์„œ ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๋ฅผ ๋นผ์„œ _์ž‘์—… ๋ฒกํ„ฐ_๋ฅผ ๊ตฌ์ถ•ํ•˜์—ฌ ์‚ฐ์ˆ  ์—ฐ์‚ฐ์„ ํ†ตํ•ด ์กฐ์ž‘ํ•˜์—ฌ ๋ณ‘ํ•ฉ๋œ ๋ชจ๋ธ์˜ ๋™์ž‘์„ ์œ ๋„ํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Simple weight interpolation works well for merging image generation models that operate in a real valued latent and output space, but to our surprise, they also still work for merging language models under the right conditions, but have performance issues [49]. Recently, new methods have been proposed to address merging language models specifically. One method, Task Arithmetic [21], involves building _task vectors_ by subtracting pre-trained from fine-tuned model weights, enabling manipulation through arithmetic operations to steer the merged model's behavior. [|assistant|]
๋ณด์ˆ˜์ ์œผ๋กœ ์ž„๊ณ„๊ฐ’์œผ๋กœ \(50\) tokens์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค: "๋ฌด๋ฆŽ ๊ตฝํž˜"์€ \(10\) tokens์—์„œ ๋ฐœ์ƒํ•˜๋ฉฐ ๊ธธ์ด-\(25\) ์ผ์น˜ ํ•ญ๋ชฉ์„ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์‚ฌํ•œ ๊ฒฐ๊ณผ ์˜คํƒ์ด ๋ฐœ๊ฒฌ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ๊ฐ’์„ ๋‘ ๋ฐฐ๋กœ ๋Š˜๋ ค ์˜ค๋ฅ˜ ํ—ˆ์šฉ ๋ฒ”์œ„๋ฅผ ๋งค์šฐ ๋„“๊ฒŒ ์žก์•˜์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ณด์ˆ˜์ ์œผ๋กœ ์ž„๊ณ„๊ฐ’์œผ๋กœ \(50\) ํ† ํฐ์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค: "๋ฌด๋ฆŽ ๊ตฝํž˜"์€ \(10\) ํ† ํฐ์—์„œ ๋ฐœ์ƒํ•˜๋ฉฐ ๊ธธ์ด-\(25\) ์ผ์น˜ ํ•ญ๋ชฉ์„ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์‚ฌํ•œ ๊ฒฐ๊ณผ ์˜คํƒ์ด ๋ฐœ๊ฒฌ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ๊ฐ’์„ ๋‘ ๋ฐฐ๋กœ ๋Š˜๋ ค ์˜ค๋ฅ˜ ํ—ˆ์šฉ ๋ฒ”์œ„๋ฅผ ๋งค์šฐ ๋„“๊ฒŒ ์žก์•˜์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We choose \(50\) tokens as the threshold to be conservative: the "bend in the knee" occurs at \(10\) tokens, and manual inspection of length-\(25\) matches found no false positives. We then doubled this value to have an exceptionally large margin for error. [|assistant|]
**Tokenizer** ๋ฐ์ดํ„ฐ tokenization์„ ์œ„ํ•ด SentencePiece(Kudo and Richardson, 2018)์˜ byte-pair encoding (BPE)(Shibata et al., 1999)์„ ํ™œ์šฉํ•˜๋Š” baichuan2 tokenizer(Yang et al., 2023)๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์–ดํœ˜ ํฌ๊ธฐ๋Š” 125,696๊ฐœ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด tokenizer๋Š” ์ˆซ์ž๋ฅผ ๊ฐœ๋ณ„ ์ˆซ์ž๋กœ ๋ถ„ํ• ํ•˜๋„๋ก ์„ค๊ณ„๋˜์–ด ์ˆซ์ž ๋ฐ์ดํ„ฐ์˜ ์ธ์ฝ”๋”ฉ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.[|endofturn|]
**ํ† ํฐ๋ผ์ด์ €** ๋ฐ์ดํ„ฐ ํ† ํฐํ™”๋ฅผ ์œ„ํ•ด SentencePiece(Kudo and Richardson, 2018)์˜ ๋ฐ”์ดํŠธ ์Œ ์ธ์ฝ”๋”ฉ(BPE)(Shibata et al., 1999)์„ ํ™œ์šฉํ•˜๋Š” ๋ฐ”์ด์ถ”์•ˆ2 ํ† ํฐ๋ผ์ด์ €(Yang et al., 2023)๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์–ดํœ˜ ํฌ๊ธฐ๋Š” 125,696๊ฐœ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด ํ† ํฐํ™” ๋„๊ตฌ๋Š” ์ˆซ์ž๋ฅผ ๊ฐœ๋ณ„ ์ˆซ์ž๋กœ ๋ถ„ํ• ํ•˜๋„๋ก ์„ค๊ณ„๋˜์–ด ์ˆซ์ž ๋ฐ์ดํ„ฐ์˜ ์ธ์ฝ”๋”ฉ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Tokenizer** We employed the baichuan2 tokenizer (Yang et al., 2023), which utilizes byte-pair encoding (BPE) (Shibata et al., 1999) from SentencePiece (Kudo and Richardson, 2018) for data tokenization. The vocabulary size is 125,696. Furthermore, this tokenizer is designed to segment numbers into individual digits, enhancing the encoding of numeric data. [|assistant|]
๊ฐ์ฃผ 1: ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ๋ฒค์น˜๋งˆํฌ ์ž‘์—…์—์„œ ์–ด๋–ป๊ฒŒ ์ˆ˜ํ–‰๋˜๋Š”์ง€ ๋น„๊ตํ•˜๋Š” ํ’๋ถ€ํ•œ ๋ฌธํ—Œ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€๋ถ€๋ถ„์˜ ๋น„๊ต๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต๋œ ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์— ๋Œ€ํ•œ ๋น„๊ต์ž…๋‹ˆ๋‹ค. LLaMA-70B๊ฐ€ Mistral-7B๋ณด๋‹ค ๋‚˜์€ ๊ฒฝ์šฐ, ๊ทธ ์ด๋“์€ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ์„ ํƒ์—์„œ ๋น„๋กฏ๋œ ๊ฒƒ์ผ๊นŒ์š”, ์•„๋‹ˆ๋ฉด architecture์˜ ์ฐจ์ด์—์„œ ๋น„๋กฏ๋œ ๊ฒƒ์ผ๊นŒ์š”, ์•„๋‹ˆ๋ฉด ์ •๋ง ๋ชจ๋ธ์˜ ํฌ๊ธฐ์—์„œ ๋น„๋กฏ๋œ ๊ฒƒ์ผ๊นŒ์š”? LLaMA-70B๋Š” ์„ธ๊ณ„ ์ง€์‹ ๋ฒค์น˜๋งˆํฌ์—์„œ 63.6%๋ฅผ ๊ธฐ๋กํ•˜๋Š” ๋ฐ˜๋ฉด, LLaMA-7B๋Š” 48.9%๋งŒ ๊ธฐ๋กํ•˜๋Š” ๋“ฑ ๋™์ผํ•œ architecture ๊ฐ„์˜ ๋น„๊ต๋„ ์žˆ์Šต๋‹ˆ๋‹ค[33]. ์ด๋Š” ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ 10๋ฐฐ ๋Š˜๋ฆฌ๋ฉด ์šฉ๋Ÿ‰์ด 130% = 63.6/48.9๋งŒ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ์˜๋ฏธ์ผ๊นŒ์š”? ๋”ฐ๋ผ์„œ ํ†ต์ œ๋œ ํ™˜๊ฒฝ์—์„œ scaling laws๋ฅผ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ๋ณด๋‹ค ์›์น™์ ์ธ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ฐ์ฃผ 1: ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ๋ฒค์น˜๋งˆํฌ ์ž‘์—…์—์„œ ์–ด๋–ป๊ฒŒ ์ˆ˜ํ–‰๋˜๋Š”์ง€ ๋น„๊ตํ•˜๋Š” ํ’๋ถ€ํ•œ ๋ฌธํ—Œ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€๋ถ€๋ถ„์˜ ๋น„๊ต๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต๋œ ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์— ๋Œ€ํ•œ ๋น„๊ต์ž…๋‹ˆ๋‹ค. LLaMA-70B๊ฐ€ Mistral-7B๋ณด๋‹ค ๋‚˜์€ ๊ฒฝ์šฐ, ๊ทธ ์ด๋“์€ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ์„ ํƒ์—์„œ ๋น„๋กฏ๋œ ๊ฒƒ์ผ๊นŒ์š”, ์•„๋‹ˆ๋ฉด ์•„ํ‚คํ…์ฒ˜์˜ ์ฐจ์ด์—์„œ ๋น„๋กฏ๋œ ๊ฒƒ์ผ๊นŒ์š”, ์•„๋‹ˆ๋ฉด ์ •๋ง ๋ชจ๋ธ์˜ ํฌ๊ธฐ์—์„œ ๋น„๋กฏ๋œ ๊ฒƒ์ผ๊นŒ์š”? LLaMA-70B๋Š” ์„ธ๊ณ„ ์ง€์‹ ๋ฒค์น˜๋งˆํฌ์—์„œ 63.6%๋ฅผ ๊ธฐ๋กํ•˜๋Š” ๋ฐ˜๋ฉด, LLaMA-7B๋Š” 48.9%๋งŒ ๊ธฐ๋กํ•˜๋Š” ๋“ฑ ๋™์ผํ•œ ์•„ํ‚คํ…์ฒ˜ ๊ฐ„์˜ ๋น„๊ต๋„ ์žˆ์Šต๋‹ˆ๋‹ค[33]. ์ด๋Š” ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ 10๋ฐฐ ๋Š˜๋ฆฌ๋ฉด ์šฉ๋Ÿ‰์ด 130% = 63.6/48.9๋งŒ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ์˜๋ฏธ์ผ๊นŒ์š”? ๋”ฐ๋ผ์„œ ํ†ต์ œ๋œ ํ™˜๊ฒฝ์—์„œ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ๋ณด๋‹ค ์›์น™์ ์ธ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Footnote 1: There is a rich literature comparing how pretrained models perform on benchmark tasks. Most comparisons are for different model families trained over different data: if LLaMA-70B is better than Mistral-7B, does the gain come from its choice of pretrain data, or the architecture difference, or really the size of the model? Some comparisons are among the same architecture, such as LLaMA-70B scores 63.6% on the world knowledge benchmark while LLaMA-7B scores only 48.9% [33]; does this mean increasing model size by 10x increases its capacity only to 130% = 63.6/48.9? Thus, it is highly important to use a more principled framework to study scaling laws in a controlled setting. [|assistant|]
๋ถ€๋ก D์˜ Int4 Quantization ํ›„, MiniCPM-2.4B์˜ ํ’‹ํ”„๋ฆฐํŠธ๋Š” 2GB๋กœ ์ถ•์†Œ๋˜์–ด ๋ชจ๋ฐ”์ผ ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค์— ๋ฐฐํฌ๊ฐ€ ์šฉ์ดํ•ด์กŒ์Šต๋‹ˆ๋‹ค. MLC-LLM(ํŒ€, 2023b)์„ ์‚ฌ์šฉํ•˜์—ฌ Android ๋ฐ HarmonyOS์šฉ ๋ชจ๋ธ์„ ์กฐ์ •ํ•˜๊ณ  iPhone ์‹œ์Šคํ…œ์— ์ ์šฉํ•˜๊ธฐ ์œ„ํ•ด LLMFarm(ํŒ€, 2023a)์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ์‘์€ ๋‹ค์–‘ํ•œ ์—ฃ์ง€ ๋ชจ๋ฐ”์ผ ๋””๋ฐ”์ด์Šค์—์„œ ํ…Œ์ŠคํŠธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ถ€๋ก D์˜ Int4 ์–‘์žํ™” ํ›„, MiniCPM-2.4B์˜ ํ’‹ํ”„๋ฆฐํŠธ๋Š” 2GB๋กœ ์ถ•์†Œ๋˜์–ด ๋ชจ๋ฐ”์ผ ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค์— ๋ฐฐํฌ๊ฐ€ ์šฉ์ดํ•ด์กŒ์Šต๋‹ˆ๋‹ค. MLC-LLM(ํŒ€, 2023b)์„ ์‚ฌ์šฉํ•˜์—ฌ Android ๋ฐ HarmonyOS์šฉ ๋ชจ๋ธ์„ ์กฐ์ •ํ•˜๊ณ  iPhone ์‹œ์Šคํ…œ์— ์ ์šฉํ•˜๊ธฐ ์œ„ํ•ด LLMFarm(ํŒ€, 2023a)์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ์‘์€ ๋‹ค์–‘ํ•œ ์—ฃ์ง€ ๋ชจ๋ฐ”์ผ ๋””๋ฐ”์ด์Šค์—์„œ ํ…Œ์ŠคํŠธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]After Int4 quantization in Appendix D, MiniCPM-2.4B's footprint is reduced to 2GB, facilitating deployment on mobile edge devices. We adapted the model for Android and HarmonyOS using MLC-LLM (team, 2023b) and employed LLMFarm (team, 2023a) for adaptation to the iPhone system. This adaptation was tested on various edge mobile devices. [|assistant|]
์ˆ˜ํ•™ ์—ฐ์† ์‚ฌ์ „ ํ•™์Šต์— ๋Œ€ํ•œ SLM: 1B์™€ 7B Rho-1์€ GSM8k ๋ฐ MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ CLM๋กœ ํ•™์Šต๋œ baseline๋ณด๋‹ค 16% ์ด์ƒ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด SLM์€ baseline ์ •ํ™•๋„์— ์ตœ๋Œ€ 10๋ฐฐ ๋” ๋น ๋ฅด๊ฒŒ ๋„๋‹ฌํ•ฉ๋‹ˆ๋‹ค. ๋†€๋ž๊ฒŒ๋„ Rho-1-7B๋Š” DeepSeekMath์— ํ•„์š”ํ•œ 500B tokens์— ๋น„ํ•ด 15B tokens๋งŒ ์‚ฌ์šฉํ•˜์—ฌ DeepSeekMath-7B์˜ ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ์— ํ•„์ ํ•ฉ๋‹ˆ๋‹ค. Fine-tuning ํ›„ Rho-1-1B์™€ 7B๋Š” ๊ฐ๊ฐ MATH์—์„œ 40.6%์™€ 51.8%๋ฅผ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ Rho-1-1B๋Š” 40%์˜ ์ •ํ™•๋„๋ฅผ ์ดˆ๊ณผํ•˜๋Š” ์ตœ์ดˆ์˜ 1B LM์œผ๋กœ, ์ดˆ๊ธฐ GPT-4์˜ CoT ์„ฑ๋Šฅ์ธ 42.5%์— ๊ทผ์ ‘ํ–ˆ์Šต๋‹ˆ๋‹ค. SS3.3์€ ์ผ๋ฐ˜์ ์ธ ์‚ฌ์ „ ํ•™์Šต์—์„œ SLM์˜ ํšจ๋Šฅ์„ ํ™•์ธ์‹œ์ผœ ์ค๋‹ˆ๋‹ค: 80B tokens์— ๋Œ€ํ•ด Tinyllama-1B๋ฅผ SLM์œผ๋กœ ํ•™์Šตํ•˜๋ฉด 15๊ฐœ ๋ฒค์น˜๋งˆํฌ์—์„œ ํ‰๊ท  6.8%๊ฐ€ ํ–ฅ์ƒ๋˜๋ฉฐ, ์ฝ”๋“œ ๋ฐ ์ˆ˜ํ•™ ๊ณผ์ œ์—์„œ 10% ์ด์ƒ์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์ž…๋‹ˆ๋‹ค.[|endofturn|]
์ˆ˜ํ•™ ์—ฐ์† ์‚ฌ์ „ ํ•™์Šต์— ๋Œ€ํ•œ SLM: 1B์™€ 7B Rho-1์€ GSM8k ๋ฐ MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ CLM๋กœ ํ•™์Šต๋œ ๊ธฐ์ค€์„ ๋ณด๋‹ค 16% ์ด์ƒ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด SLM์€ ๊ธฐ์ค€์„  ์ •ํ™•๋„์— ์ตœ๋Œ€ 10๋ฐฐ ๋” ๋น ๋ฅด๊ฒŒ ๋„๋‹ฌํ•ฉ๋‹ˆ๋‹ค. ๋†€๋ž๊ฒŒ๋„ Rho-1-7B๋Š” DeepSeekMath์— ํ•„์š”ํ•œ 5000์–ต ํ† ํฐ์— ๋น„ํ•ด 150์–ต ํ† ํฐ๋งŒ ์‚ฌ์šฉํ•˜์—ฌ DeepSeekMath-7B์˜ ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ์— ํ•„์ ํ•ฉ๋‹ˆ๋‹ค. ๋ฏธ์„ธ ์กฐ์ • ํ›„ Rho-1-1B์™€ 7B๋Š” ๊ฐ๊ฐ MATH์—์„œ 40.6%์™€ 51.8%๋ฅผ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ Rho-1-1B๋Š” 40%์˜ ์ •ํ™•๋„๋ฅผ ์ดˆ๊ณผํ•˜๋Š” ์ตœ์ดˆ์˜ 10์–ต ๋‹จ์–ด ๋ชจ๋ธ๋กœ, ์ดˆ๊ธฐ GPT-4์˜ CoT ์„ฑ๋Šฅ์ธ 42.5%์— ๊ทผ์ ‘ํ–ˆ์Šต๋‹ˆ๋‹ค. SS3.3์€ ์ผ๋ฐ˜์ ์ธ ์‚ฌ์ „ ํ•™์Šต์—์„œ SLM์˜ ํšจ๋Šฅ์„ ํ™•์ธ์‹œ์ผœ ์ค๋‹ˆ๋‹ค: 800์–ต ํ† ํฐ์— ๋Œ€ํ•ด Tinyllama-1B๋ฅผ SLM์œผ๋กœ ํ•™์Šตํ•˜๋ฉด 15๊ฐœ ๋ฒค์น˜๋งˆํฌ์—์„œ ํ‰๊ท  6.8%๊ฐ€ ํ–ฅ์ƒ๋˜๋ฉฐ, ์ฝ”๋“œ ๋ฐ ์ˆ˜ํ•™ ๊ณผ์ œ์—์„œ 10% ์ด์ƒ์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]SLM on math continual pretraining: both 1B and 7B Rho-1 outperform CLM-trained baselines by over 16% on the GSM8k and MATH datasets. SLM reaches baseline accuracy up to 10x faster, as shown in Figure 1. Remarkably, Rho-1-7B matches the state-of-the-art performance of DeepSeekMath-7B using only 15B tokens, compared to the 500B tokens required by DeepSeekMath. Upon fine-tuning, Rho-1-1B and 7B achieve 40.6% and 51.8% on MATH, respectively. Notably, Rho-1-1B is the first 1B LM to exceed 40% accuracy, nearing the early GPT-4's CoT performance of 42.5%. SS3.3 confirms the efficacy of SLM in general pretraining: Training Tinyllama-1B on 80B tokens with SLM improves 6.8% on average across 15 benchmarks, with gains over 10% in code and math tasks. [|assistant|]
์ด ๊ธ‰์—ฌ = 4๋…„ ํ›„ Adrien์˜ ๊ธ‰์—ฌ + 4๋…„ ํ›„ Lylah์˜ ๊ธ‰์—ฌ = $56000 + $39200 = $95200[|endofturn|]
์ด ๊ธ‰์—ฌ = 4๋…„ ํ›„ ์•„๋“œ๋ฆฌ์•ˆ์˜ ๊ธ‰์—ฌ + 4๋…„ ํ›„ ๋ผ์ผ๋ผ์˜ ๊ธ‰์—ฌ = $56000 + $39200 = $95200[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|] Total salary = Adrien's salary four years later + Lylah's salary four years later = $56000 + $39200 = $95200 [|assistant|]
์ด ํ•˜์œ„ ์„น์…˜์—์„œ๋Š” ํ•™์Šต ์†๋„๋ฅผ ๋‹ค์‹œ ์˜ˆ์—ดํ•˜๊ณ  ๋‹ค์‹œ ๊ฐ์†Œ์‹œํ‚ค๋Š” ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•  ๋•Œ compute-equivalent replay์˜ ํšจ๊ณผ๋ฅผ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.[|endofturn|]
์ด ํ•˜์œ„ ์„น์…˜์—์„œ๋Š” ํ•™์Šต ์†๋„๋ฅผ ๋‹ค์‹œ ์˜ˆ์—ดํ•˜๊ณ  ๋‹ค์‹œ ๊ฐ์†Œ์‹œํ‚ค๋Š” ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•  ๋•Œ ์ปดํ“จํŒ… ๋“ฑ๊ฐ€ ์žฌ์ƒ์˜ ํšจ๊ณผ๋ฅผ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this subsection, we explore the effect of compute-equivalent replay when continually pre-training models that re-warm and re-decay the learning rate. [|assistant|]
๋‘ ์‹œ๋‚˜๋ฆฌ์˜ค์˜ IL ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜๋ฉด, Small๊ณผ Small-P1\(\rightarrow\)Small-P2์˜ ๊ฒฝ์šฐ, ๋™์ผํ•œ ํšŸ์ˆ˜์˜ ํ›ˆ๋ จ ๋‹จ๊ณ„๋ฅผ ๊ฑฐ์ณค์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LMs์€ ์—ฌ๋Ÿฌ ๋ฒˆ์˜ CKL ๋‹จ๊ณ„๋ฅผ ๊ฑฐ์น˜๋ฉด์„œ ๋” ๋งŽ์€ ์žŠ์–ด๋ฒ„๋ฆผ์— ๊ฑธ๋ฆฌ๊ธฐ ์‰ฝ๋‹ค๋Š” ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ทธ ์ด์œ  ์ค‘ ํ•˜๋‚˜๋Š” ๊ฐ ๋‹จ๊ณ„์˜ ์‹œ์ž‘ ์‹œ ์ดˆ๊ธฐํ™”๋˜๋Š” Learning Rate ์Šค์ผ€์ค„๋ง ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋‘ ์‹œ๋‚˜๋ฆฌ์˜ค์˜ IL ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜๋ฉด, Small๊ณผ Small-P1\(\์šฐ์ธก ๋Œ€๊ฐ์„ \)Small-P2์˜ ๊ฒฝ์šฐ, ๋™์ผํ•œ ํšŸ์ˆ˜์˜ ํ›ˆ๋ จ ๋‹จ๊ณ„๋ฅผ ๊ฑฐ์ณค์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LM์€ ์—ฌ๋Ÿฌ ๋ฒˆ์˜ CKL ๋‹จ๊ณ„๋ฅผ ๊ฑฐ์น˜๋ฉด์„œ ๋” ๋งŽ์€ ์žŠ์–ด๋ฒ„๋ฆผ์— ๊ฑธ๋ฆฌ๊ธฐ ์‰ฝ๋‹ค๋Š” ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ทธ ์ด์œ  ์ค‘ ํ•˜๋‚˜๋Š” ๊ฐ ๋‹จ๊ณ„์˜ ์‹œ์ž‘ ์‹œ ์ดˆ๊ธฐํ™”๋˜๋Š” ํ•™์Šต ์†๋„ ์Šค์ผ€์ค„๋ง ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Comparing the performance on IL of the two scenarios, Small and Small-P1\(\rightarrow\)Small-P2, results show that LMs are prone to more forgetting as they go through multiple CKL phases, despite having the same number of training steps. One of the reasons may be due to the learning rate scheduling, which is initialized at the start of each phase. [|assistant|]
FLAN(Chung et al., 2022)์€ Instruction Tuning์— ๋Œ€ํ•œ ๋งŽ์€ ์—ฐ๊ตฌ ์ค‘ ๊ฐ€์žฅ ์ฒด๊ณ„์ ์ด๊ณ  ํฌ๊ด„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ Instruction-optimised dataset์—์„œ ์–ธ์–ด ๋ชจ๋ธ์„ Fine-tuningํ•˜๊ณ , ์ž‘์—… ์ˆ˜์™€ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ํ™•์žฅํ•˜๋ฉฐ, Fine-tuning์— chain-of-thought ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ €์ž๋“ค์€ RALM ์•„ํ‚คํ…์ฒ˜์—์„œ Instruction์„ ์กฐ์ •ํ•˜๋Š” ๊ตฌ์ฒด์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ณ ๋ คํ•˜์ง€ ์•Š์•˜์ง€๋งŒ, ์ด ์ž‘์—…์€ ํ–ฅํ›„ ์—ฐ๊ตฌ์— ๊ท€์ค‘ํ•œ ์ฐธ๊ณ  ์ž๋ฃŒ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. Lin et al. (2023)์€ RALM์˜ Instruction Fine-tuning์—์„œ In-Context Retrieval Augmentation์„ ํ†ตํ•ฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์–ธ์–ด ๋ชจ๋ธ์ด ๊ด€๋ จ ์—†๋Š” ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์˜ํ•ด ์˜ค๋„๋  ๊ฐ€๋Šฅ์„ฑ์„ ํฌ๊ฒŒ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. SAIL(Luo et al., 2023)์€ ๋‚ด๋ถ€ ๋ฐ ์™ธ๋ถ€ ๊ฒ€์ƒ‰ ์—”์ง„์—์„œ ์ƒ์„ฑ๋œ ๋ณต์žกํ•œ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์–ธ์–ด ์ƒ์„ฑ ๋ฐ Instruction ์ถ”์  ๊ธฐ๋Šฅ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. Instruction Tuning ์ฝ”ํผ์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ํ›ˆ๋ จ ์‚ฌ๋ก€์— ๋Œ€ํ•œ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๋ฅผ ๋‹ค์–‘ํ•œ ๊ฒ€์ƒ‰ API ๋ฐ ๋„๋ฉ”์ธ์—์„œ ์ˆ˜์ง‘ํ•˜๊ณ , (Instruction, grounding information, response)์˜ ํŠธ๋ฆฌํ”Œ๋ ›์„ ํฌํ•จํ•˜๋Š” ๊ฒ€์ƒ‰ ๊ธฐ๋ฐ˜ ํ›ˆ๋ จ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. Instruction-Tuned ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ํ›ˆ๋ จ๊ณผ๋Š” ๋‹ฌ๋ฆฌ, Madaan et al. (2022)๊ณผ Lazaridou et al. (2022)์€ ๊ฒ€์ƒ‰๋œ ์ง€์‹์—์„œ ์ง์ ‘ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ํ”„๋กฌํ”„ํŠธํ•˜๋„๋ก ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Madaan et al. (2022)์€ GPT-3๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์ด ์‚ฌ์šฉ์ž์˜ ์˜๋„๋ฅผ ์ž˜๋ชป ํ•ด์„ํ•œ ๊ธฐ๋ก๋œ ์‚ฌ๋ก€์˜ ๋ฉ”๋ชจ๋ฆฌ ํŽ˜์–ด๋ง๊ณผ ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์„ ๋ช…ํ™•ํžˆ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ด ์‹œ์Šคํ…œ์€ ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ๊ฐ์˜ ์ƒˆ๋กœ์šด ์ฟผ๋ฆฌ์— ๋Œ€ํ•ด ํ–ฅ์ƒ๋œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— Lazaridou et al. (2022)์€ Few-shot ํ”„๋กฌํ”„ํŠธ์™€ ๋‹ต๋ณ€ ์žฌ์ •๋ ฌ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๋ก  ๊ณ„์‚ฐ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
FLAN(Chung et al., 2022)์€ ์ง€์‹œ ํŠœ๋‹์— ๋Œ€ํ•œ ๋งŽ์€ ์—ฐ๊ตฌ ์ค‘ ๊ฐ€์žฅ ์ฒด๊ณ„์ ์ด๊ณ  ํฌ๊ด„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ง€์‹œ ์ตœ์ ํ™” ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๊ณ , ์ž‘์—… ์ˆ˜์™€ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ํ™•์žฅํ•˜๋ฉฐ, ๋ฏธ์„ธ ์กฐ์ •์— ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ €์ž๋“ค์€ RALM ์•„ํ‚คํ…์ฒ˜์—์„œ ์ง€์นจ์„ ์กฐ์ •ํ•˜๋Š” ๊ตฌ์ฒด์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ณ ๋ คํ•˜์ง€ ์•Š์•˜์ง€๋งŒ, ์ด ์ž‘์—…์€ ํ–ฅํ›„ ์—ฐ๊ตฌ์— ๊ท€์ค‘ํ•œ ์ฐธ๊ณ  ์ž๋ฃŒ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. Lin et al. (2023)์€ RALM์˜ ์ง€์นจ ๋ฏธ์„ธ ์กฐ์ •์—์„œ ์ปจํ…์ŠคํŠธ ๊ฒ€์ƒ‰ ์ฆ๊ฐ•์„ ํ†ตํ•ฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์–ธ์–ด ๋ชจ๋ธ์ด ๊ด€๋ จ ์—†๋Š” ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์˜ํ•ด ์˜ค๋„๋  ๊ฐ€๋Šฅ์„ฑ์„ ํฌ๊ฒŒ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. SAIL(Luo et al., 2023)์€ ๋‚ด๋ถ€ ๋ฐ ์™ธ๋ถ€ ๊ฒ€์ƒ‰ ์—”์ง„์—์„œ ์ƒ์„ฑ๋œ ๋ณต์žกํ•œ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์–ธ์–ด ์ƒ์„ฑ ๋ฐ ์ง€์นจ ์ถ”์  ๊ธฐ๋Šฅ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ์ง€์นจ ํŠœ๋‹ ์ฝ”ํผ์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ํ›ˆ๋ จ ์‚ฌ๋ก€์— ๋Œ€ํ•œ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๋ฅผ ๋‹ค์–‘ํ•œ ๊ฒ€์ƒ‰ API ๋ฐ ๋„๋ฉ”์ธ์—์„œ ์ˆ˜์ง‘ํ•˜๊ณ , (์ง€์นจ, ๊ทธ๋ผ์šด๋”ฉ ์ •๋ณด, ์‘๋‹ต)์˜ ํŠธ๋ฆฌํ”Œ๋ ›์„ ํฌํ•จํ•˜๋Š” ๊ฒ€์ƒ‰ ๊ธฐ๋ฐ˜ ํ›ˆ๋ จ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ง€์นจ ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ํ›ˆ๋ จ๊ณผ๋Š” ๋‹ฌ๋ฆฌ, Madaan et al. (2022)๊ณผ Lazaridou et al. (2022)์€ ๊ฒ€์ƒ‰๋œ ์ง€์‹์—์„œ ์ง์ ‘ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ํ”„๋กฌํ”„ํŠธํ•˜๋„๋ก ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Madaan et al. (2022)์€ GPT-3๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์ด ์‚ฌ์šฉ์ž์˜ ์˜๋„๋ฅผ ์ž˜๋ชป ํ•ด์„ํ•œ ๊ธฐ๋ก๋œ ์‚ฌ๋ก€์˜ ๋ฉ”๋ชจ๋ฆฌ ํŽ˜์–ด๋ง๊ณผ ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์„ ๋ช…ํ™•ํžˆ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ด ์‹œ์Šคํ…œ์€ ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ๊ฐ์˜ ์ƒˆ๋กœ์šด ์ฟผ๋ฆฌ์— ๋Œ€ํ•ด ํ–ฅ์ƒ๋œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— Lazaridou et al. (2022)์€ ์†Œ๋Ÿ‰ ํ”„๋กฌํ”„ํŠธ์™€ ๋‹ต๋ณ€ ์žฌ์ •๋ ฌ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๋ก  ๊ณ„์‚ฐ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]FLAN(Chung et al., 2022) is one of the most systematic and comprehensive approaches among the many studies on instruction tuning. This approach fine-tunes the language model on the instruction-optimised dataset, scales the number of tasks and model size, and incorporates chain-of-thought data in the fine-tuning. Although the authors did not consider a specific approach to tuning instructions in RALM architecture, their work provides a valuable reference for future research. In the instruction fine-tuning of RALM, Lin et al. (2023) integrated in-context retrieval augmentation. This greatly reduces the likelihood of the language model being misled by irrelevant retrieval content. SAIL(Luo et al., 2023) builds language generation and instruction tracking capabilities on complex search results generated by internal and external search engines. Using a corpus of instruction tuning, they collect search results for each training case from different search APIs and domains, and construct a search-based training set containing a triplet of (instruction, grounding information, response). In contrast to training on instruction-tuned datasets, Madaan et al. (2022) and Lazaridou et al. (2022) propose to prompt large models directly from retrieved knowledge. Madaan et al. (2022) used GPT-3 to clarify memory pairings of recorded cases where the model misinterpreted the user's intention, as well as user feedback. This ensures that their system can generate enhanced prompts for each new query based on user feedback. In contrast, Lazaridou et al. (2022) uses few-shot prompts and answer reordering to improve inference computation. [|assistant|]
๊ทธ๋ฆผ 2๋Š” ๋ฐ์ดํ„ฐ ์ง„ํ™” ๋ผ์šด๋“œ ์ˆ˜์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•˜๋Š” ์ ˆ์ œ ์—ฐ๊ตฌ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์ง„ํ™” ๋ฐ์ดํ„ฐ ๋ผ์šด๋“œ์—๋Š” 38k ์ƒ˜ํ”Œ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋ผ์šด๋“œ์—๋Š” 58k๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์„ธ ๋ฒˆ์งธ ๋ผ์šด๋“œ์—๋Š” 78k๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋„ค ๋ฒˆ์งธ ๋ผ์šด๋“œ์—๋Š” 98k๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ผ๊ด€์„ฑ์„ ์œ„ํ•ด ๋ชจ๋“  ๋ชจ๋ธ์€ 200๋‹จ๊ณ„๋กœ Fine-tuning์„ ๊ฑฐ์นฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, HumanEval์—์„œ ๊ฐ€์žฅ ๋†’์€ pass@1 ์ ์ˆ˜๋Š” ๋ฐ์ดํ„ฐ ์ง„ํ™” 3๋ผ์šด๋“œ ํ›„์— ๋‹ฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ์„ ๋ฐ”ํƒ•์œผ๋กœ ์„ธ ๋ฒˆ์งธ ๋ผ์šด๋“œ์—์„œ ์ง„ํ™”ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์ตœ์ข… ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 2๋Š” ๋ฐ์ดํ„ฐ ์ง„ํ™” ๋ผ์šด๋“œ ์ˆ˜์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•˜๋Š” ์ ˆ์ œ ์—ฐ๊ตฌ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์ง„ํ™” ๋ฐ์ดํ„ฐ ๋ผ์šด๋“œ์—๋Š” 38k ์ƒ˜ํ”Œ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋ผ์šด๋“œ์—๋Š” 58k๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์„ธ ๋ฒˆ์งธ ๋ผ์šด๋“œ์—๋Š” 78k๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋„ค ๋ฒˆ์งธ ๋ผ์šด๋“œ์—๋Š” 98k๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ผ๊ด€์„ฑ์„ ์œ„ํ•ด ๋ชจ๋“  ๋ชจ๋ธ์€ 200๋‹จ๊ณ„๋กœ ๋ฏธ์„ธ ์กฐ์ •์„ ๊ฑฐ์นฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ํœด๋จผ์ด๋ฐœ์—์„œ ๊ฐ€์žฅ ๋†’์€ ํŒจ์Šค@1 ์ ์ˆ˜๋Š” ๋ฐ์ดํ„ฐ ์ง„ํ™” 3๋ผ์šด๋“œ ํ›„์— ๋‹ฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ์„ ๋ฐ”ํƒ•์œผ๋กœ ์„ธ ๋ฒˆ์งธ ๋ผ์šด๋“œ์—์„œ ์ง„ํ™”ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์ตœ์ข… ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 2 presents an ablation study investigating the impact of the number of data evolution rounds. The first round of evolved data contains 38k samples. The second round contains 58k. The third round contains 78k. The fourth round contains 98k. For consistency, all models undergo fine-tuning with 200 steps. The results reveal that the highest pass@1 score on humaneval is achieved after three rounds of data evolution. Based on this observation, we select the data that evolved during the third round as the ultimate dataset. [|assistant|]
๊ฐ์ฃผ 5: ์ด ์ž‘์—…์—์„œ๋Š” ์ธ์Šคํ„ด์Šค ์ƒ์„ฑ์„ ์œ„ํ•œ prompting์„ ์œ„ํ•ด ๊ณ ์ •๋œ ์‹œ๋“œ ์ž‘์—… ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋ฏ€๋กœ ํ•œ ๋ผ์šด๋“œ์—์„œ ์ž‘์—…๋‹น ์†Œ์ˆ˜์˜ ์ธ์Šคํ„ด์Šค๋งŒ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ ์ž‘์—…์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์ด ์—ฌ๋Ÿฌ ๋ผ์šด๋“œ์—์„œ ๋” ๋งŽ์€ ์ˆ˜์˜ ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก promptingํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ฐ์ฃผ 5: ์ด ์ž‘์—…์—์„œ๋Š” ์ธ์Šคํ„ด์Šค ์ƒ์„ฑ์„ ์œ„ํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์œ„ํ•ด ๊ณ ์ •๋œ ์‹œ๋“œ ์ž‘์—… ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋ฏ€๋กœ ํ•œ ๋ผ์šด๋“œ์—์„œ ์ž‘์—…๋‹น ์†Œ์ˆ˜์˜ ์ธ์Šคํ„ด์Šค๋งŒ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ ์ž‘์—…์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์ด ์—ฌ๋Ÿฌ ๋ผ์šด๋“œ์—์„œ ๋” ๋งŽ์€ ์ˆ˜์˜ ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Footnote 5: In this work, we use a fixed set of seed tasks for prompting the instance generation, and thus only generate a small number of instances per task in one round. Future work can use randomly sampled tasks to prompt the model to generate a larger number of instances in multiple rounds. [|assistant|]
ํ•ต์‹ฌ ์ƒ˜ํ”Œ๋ง์ด๋ผ๊ณ ๋„ ํ•˜๋Š” top-p ์ƒ˜ํ”Œ๋ง์€ top-k ์ƒ˜ํ”Œ๋ง๊ณผ๋Š” ์•ฝ๊ฐ„ ๋‹ค๋ฅธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. top-k ์ƒ˜ํ”Œ๋ง์€ ๊ฐ€์žฅ ํ™•๋ฅ ์ด ๋†’์€ ํ† ํฐ k๊ฐœ๋ฅผ ์„ ํƒํ•˜๋Š” ๋Œ€์‹ , ์„ ํƒํ•œ ํ† ํฐ์˜ ํ™•๋ฅ  ํ•ฉ์ด p๋ฅผ ์ดˆ๊ณผํ•˜๋„๋ก cutoff ๊ฐ’ p๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ๋‹ค์Œ ํ† ํฐ์„ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ•  ํ† ํฐ์˜ "ํ•ต"์ด ํ˜•์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ฆ‰, top-p ์ƒ˜ํ”Œ๋ง์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ์ด ๊ฐ€์žฅ ํ™•๋ฅ ์ด ๋†’์€ ํ† ํฐ์„ ๋‚ด๋ฆผ์ฐจ์ˆœ์œผ๋กœ ๊ฒ€ํ† ํ•˜๊ณ  ํ™•๋ฅ  ํ•ฉ์ด ์ž„๊ณ„๊ฐ’ p๋ฅผ ์ดˆ๊ณผํ•  ๋•Œ๊นŒ์ง€ ๋ชฉ๋ก์— ๊ณ„์† ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ƒ์ƒํ•  ์ˆ˜ ์žˆ๋“ฏ์ด, ์ด๋Š” ํŠนํžˆ ์ƒ์œ„ k ํ† ํฐ์˜ ํ™•๋ฅ  ์งˆ๋Ÿ‰์ด ํฌ์ง€ ์•Š์€ ์‹œ๋‚˜๋ฆฌ์˜ค์— ๋” ์ ํ•ฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. top-k ์ƒ˜ํ”Œ๋ง๊ณผ ๋‹ฌ๋ฆฌ ํ•ต์‹ฌ ์ƒ˜ํ”Œ๋ง์— ํฌํ•จ๋˜๋Š” ํ† ํฐ์˜ ์ˆ˜๋Š” ๊ณ ์ •๋˜์–ด ์žˆ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐ€๋ณ€์„ฑ์œผ๋กœ ์ธํ•ด ๋” ๋‹ค์–‘ํ•˜๊ณ  ์ฐฝ์˜์ ์ธ ๊ฒฐ๊ณผ๋ฌผ์„ ์–ป์„ ์ˆ˜ ์žˆ์–ด ํ…์ŠคํŠธ ์ƒ์„ฑ ๊ด€๋ จ ์ž‘์—…์—์„œ ํ•ต์‹ฌ ์ƒ˜ํ”Œ๋ง์ด ์ธ๊ธฐ๊ฐ€ ๋†’์Šต๋‹ˆ๋‹ค.[|endofturn|]
ํ•ต์‹ฌ ์ƒ˜ํ”Œ๋ง์ด๋ผ๊ณ ๋„ ํ•˜๋Š” ํƒ‘-p ์ƒ˜ํ”Œ๋ง์€ ํƒ‘-k ์ƒ˜ํ”Œ๋ง๊ณผ๋Š” ์•ฝ๊ฐ„ ๋‹ค๋ฅธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ํƒ‘-k ์ƒ˜ํ”Œ๋ง์€ ๊ฐ€์žฅ ํ™•๋ฅ ์ด ๋†’์€ ํ† ํฐ k๊ฐœ๋ฅผ ์„ ํƒํ•˜๋Š” ๋Œ€์‹ , ์„ ํƒํ•œ ํ† ํฐ์˜ ํ™•๋ฅ  ํ•ฉ์ด p๋ฅผ ์ดˆ๊ณผํ•˜๋„๋ก ์ปท์˜คํ”„ ๊ฐ’ p๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ๋‹ค์Œ ํ† ํฐ์„ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ•  ํ† ํฐ์˜ "ํ•ต"์ด ํ˜•์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ฆ‰, ํƒ‘-p ์ƒ˜ํ”Œ๋ง์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ์ด ๊ฐ€์žฅ ํ™•๋ฅ ์ด ๋†’์€ ํ† ํฐ์„ ๋‚ด๋ฆผ์ฐจ์ˆœ์œผ๋กœ ๊ฒ€ํ† ํ•˜๊ณ  ํ™•๋ฅ  ํ•ฉ์ด ์ž„๊ณ„๊ฐ’ p๋ฅผ ์ดˆ๊ณผํ•  ๋•Œ๊นŒ์ง€ ๋ชฉ๋ก์— ๊ณ„์† ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ƒ์ƒํ•  ์ˆ˜ ์žˆ๋“ฏ์ด, ์ด๋Š” ํŠนํžˆ ์ƒ์œ„ k ํ† ํฐ์˜ ํ™•๋ฅ  ์งˆ๋Ÿ‰์ด ํฌ์ง€ ์•Š์€ ์‹œ๋‚˜๋ฆฌ์˜ค์— ๋” ์ ํ•ฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํƒ‘-k ์ƒ˜ํ”Œ๋ง๊ณผ ๋‹ฌ๋ฆฌ ํ•ต์‹ฌ ์ƒ˜ํ”Œ๋ง์— ํฌํ•จ๋˜๋Š” ํ† ํฐ์˜ ์ˆ˜๋Š” ๊ณ ์ •๋˜์–ด ์žˆ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐ€๋ณ€์„ฑ์œผ๋กœ ์ธํ•ด ๋” ๋‹ค์–‘ํ•˜๊ณ  ์ฐฝ์˜์ ์ธ ๊ฒฐ๊ณผ๋ฌผ์„ ์–ป์„ ์ˆ˜ ์žˆ์–ด ํ…์ŠคํŠธ ์ƒ์„ฑ ๊ด€๋ จ ์ž‘์—…์—์„œ ํ•ต์‹ฌ ์ƒ˜ํ”Œ๋ง์ด ์ธ๊ธฐ๊ฐ€ ๋†’์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Top-p sampling, also known as Nucleus sampling, takes a slightly different approach from top-k sampling. Instead of selecting the top k most probable tokens, nucleus sampling chooses a cutoff value p such that the sum of the probabilities of the selected tokens exceeds p. This forms a "nucleus" of tokens from which to randomly choose the next token. In other words, in top-p sampling the language model examines the most probable tokens in descending order and keeps adding them to the list until the sum of probabilities surpasses the threshold p. As you can imagine, this could be better specially for scenarios in which top-k tokens do not have a large probability mass. Unlike top-k sampling, the number of tokens included in the nucleus sampling is not fixed. This variability often results in a more diverse and creative output, making nucleus sampling popular for text generation related tasks. [|assistant|]
์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ์ปจํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ชจ๋ธ ๋ถ„์•ผ์—์„œ "scaling law"๋ฅผ ๊ฐ•์กฐํ•˜์—ฌ ๋ชจ๋ธ ๊ทœ๋ชจ์™€ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํฌ๊ธฐ์™€ ์„ฑ๋Šฅ ๊ฐœ์„  ์‚ฌ์ด์˜ ๋น„๋ก€ ๊ด€๊ณ„๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค(Kaplan et al., 2020; Hernandez et al., 2021; Anil et al., 2023). ๊ทธ ๊ฒฐ๊ณผ Large Language Models (LLMs)์ด ๋“ฑ์žฅํ•˜๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด์ „ ๋ชจ๋ธ๊ณผ ๋‹ฌ๋ฆฌ LLM์€ ๊ฐ€์ค‘์น˜ ์กฐ์ • ์—†์ด๋„ ์ž‘์—…์„ ์กฐ์ •ํ•˜๊ณ  ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” zero-shot learning (Radford et al., 2019) ๋ฐ few-shot learning (Brown et al., 2020)๊ณผ ๊ฐ™์€ ๊ธฐ๋Šฅ์„ ํฌํ•จํ•˜๋Š” In-Context Learning์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ LLM์˜ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์€ ์†Œํ˜• ๋ชจ๋ธ์—๋Š” ์—†๋Š” ์ค‘์š”ํ•œ ์ง„ํ™”๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค(Wei et al., 2022).[|endofturn|]
์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ์ปจํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ชจ๋ธ ๋ถ„์•ผ์—์„œ "์Šค์ผ€์ผ๋ง ๋ฒ•์น™"์„ ๊ฐ•์กฐํ•˜์—ฌ ๋ชจ๋ธ ๊ทœ๋ชจ์™€ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํฌ๊ธฐ์™€ ์„ฑ๋Šฅ ๊ฐœ์„  ์‚ฌ์ด์˜ ๋น„๋ก€ ๊ด€๊ณ„๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค(Kaplan et al., 2020; Hernandez et al., 2021; Anil et al., 2023). ๊ทธ ๊ฒฐ๊ณผ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์ด ๋“ฑ์žฅํ•˜๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด์ „ ๋ชจ๋ธ๊ณผ ๋‹ฌ๋ฆฌ LLM์€ ๊ฐ€์ค‘์น˜ ์กฐ์ • ์—†์ด๋„ ์ž‘์—…์„ ์กฐ์ •ํ•˜๊ณ  ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ์ œ๋กœ ์ƒท ํ•™์Šต(Radford et al., 2019) ๋ฐ ์†Œ์ˆ˜ ์ƒท ํ•™์Šต(Brown et al., 2020)๊ณผ ๊ฐ™์€ ๊ธฐ๋Šฅ์„ ํฌํ•จํ•˜๋Š” ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ LLM์˜ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์€ ์†Œํ˜• ๋ชจ๋ธ์—๋Š” ์—†๋Š” ์ค‘์š”ํ•œ ์ง„ํ™”๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค(Wei et al., 2022).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Recent research has highlighted a "scaling law" in the field of context-based language models (Kaplan et al., 2020; Hernandez et al., 2021; Anil et al., 2023), showing a proportional relationship between the size of the model plus the training data and the resulting performance improvements. Consequently, this has led to the advent of Large Language Models (LLMs). In contrast to earlier models, LLMs can perform in-context learning, which includes abilities such as zero-shot learning (Radford et al., 2019) and few-shot learning (Brown et al., 2020), allowing them to adapt and perform tasks without the need for weight adjustments. These emergent abilities of LLMs, absent in their smaller counterparts, signal a significant evolution in language model capabilities (Wei et al., 2022). [|assistant|]
๊ทธ๋Ÿฌ๋‚˜ ์ผ๋ฐ˜์ ์œผ๋กœ \(n\)๊ณผ \(\mathbf{S}\)๋Š” ํ†ต๊ทผํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ \(\mathbf{S}\) ๋ณ€ํ™˜์ด Transformer's ์ถœ๋ ฅ์„ ๋ณด์กดํ•œ๋‹ค๊ณ  ๊ฒฐ๋ก ์„ ๋‚ด๋ฆด ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” \(\mathbf{S}=\mathbf{U}\)์ธ ์ง๊ต ํ–‰๋ ฌ์„ ์ทจํ•˜๋ฉด ์—ฌ์ „ํžˆ ์ผ๋ฐ˜์ ์ธ ๋ถˆ๊ฐ€๋Šฅ์„ฑ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค„ ๊ฒƒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋Ÿฌ๋‚˜ ์ผ๋ฐ˜์ ์œผ๋กœ \(n\)๊ณผ \(\mathbf{S}\)๋Š” ํ†ต๊ทผํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ \(\mathbf{S}\) ๋ณ€ํ™˜์ด ๋ณ€์••๊ธฐ์˜ ์ถœ๋ ฅ์„ ๋ณด์กดํ•œ๋‹ค๊ณ  ๊ฒฐ๋ก ์„ ๋‚ด๋ฆด ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” \mathbf{S}=\mathbf{U}์ธ ์ง๊ต ํ–‰๋ ฌ์„ ์ทจํ•˜๋ฉด ์—ฌ์ „ํžˆ ์ผ๋ฐ˜์ ์ธ ๋ถˆ๊ฐ€๋Šฅ์„ฑ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค„ ๊ฒƒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]but since in general \(n\) and \(\mathbf{S}\) do not commute, we cannot conclude that the \(\mathbf{S}\) transformations preserve the transformer's outputs. We will show that if we take \(\mathbf{S}=\mathbf{U}\) an orthogonal matrix, then we still get a general impossibility result. [|assistant|]
์€ ์ˆจ๊ฒจ์ง„ ๊ณต๊ฐ„์˜ ์ฐจ์›์„ ์‹๋ณ„ํ•˜๋Š” ๊ฒƒ์ด ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” \(n<h\)์ด๋ฏ€๋กœ \(n\times l\) ์ฐจ์› ํ–‰๋ ฌ \(\mathbf{Q}\)์˜ ์ˆœ์œ„๊ฐ€ ์™„์ „ํ•˜๊ณ  \(n\)์˜ ๋น„์‚ผ์ฐจ ํŠน์ด๊ฐ’์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋ธ์— 2048ํšŒ ์ด์ƒ ์ฟผ๋ฆฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ \(n>h\)๊ฐ€ ๋˜๋ฉด ์ˆ˜์น˜์ ์œผ๋กœ ์œ ์˜๋ฏธํ•œ singular value์˜ ์ˆ˜๋Š” ๋” ์ด์ƒ ์ฆ๊ฐ€ํ•˜์ง€ ์•Š์œผ๋ฉฐ ์ •ํ™•ํžˆ \(2048\)๋กœ ์ œํ•œ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์€ ์ˆจ๊ฒจ์ง„ ๊ณต๊ฐ„์˜ ์ฐจ์›์„ ์‹๋ณ„ํ•˜๋Š” ๊ฒƒ์ด ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” \(n<h\)์ด๋ฏ€๋กœ \(n\times l\) ์ฐจ์› ํ–‰๋ ฌ \(\mathbf{Q}\)์˜ ์ˆœ์œ„๊ฐ€ ์™„์ „ํ•˜๊ณ  \(n\)์˜ ๋น„์‚ผ์ฐจ ํŠน์ด๊ฐ’์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋ธ์— 2048ํšŒ ์ด์ƒ ์ฟผ๋ฆฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ \(n>h\)๊ฐ€ ๋˜๋ฉด ์ˆ˜์น˜์ ์œผ๋กœ ์œ ์˜๋ฏธํ•œ ํŠน์ด๊ฐ’์˜ ์ˆ˜๋Š” ๋” ์ด์ƒ ์ฆ๊ฐ€ํ•˜์ง€ ์•Š์œผ๋ฉฐ ์ •ํ™•ํžˆ \(2048\)๋กœ ์ œํ•œ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]is impossible to identify the dimensionality of the hidden space. This is because \(n<h\), and so the \(n\times l\) dimensional matrix \(\mathbf{Q}\) has full rank and \(n\) nontrivial singular values. But once we make more than \(2048\) queries to the model, and thus \(n>h\), the number of numerically significant singular values does not increase further; it is capped at exactly \(2048\). [|assistant|]
\(\bullet\)_๋ณด๋‹ค ๋ณต์žกํ•œ ์ž‘์—…์€ ChatGPT์— ๋Œ€ํ•œ ์‹ ์ค‘ํ•œ prompt engineering์˜ ํ˜œํƒ์„ ๋” ๋งŽ์ด ๋ˆ„๋ฆด ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค._ WikiFact ๋ฐ Colored Objects ์ž‘์—…์—์„œ ์„ค๊ณ„๋œ prompts๋Š” ChatGPT์˜ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค(_์ฆ‰, WikiFact์˜ ๊ฒฝ์šฐ 23.61์—์„œ 28.47๋กœ, Colored Objects์˜ ๊ฒฝ์šฐ 53.20์—์„œ 66.75๋กœ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค). ์ด๋Ÿฌํ•œ ์ž‘์—…์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํŠน์ • ์ถœ๋ ฅ ํ˜•์‹์„ ๊ฐ€์ง€๊ฑฐ๋‚˜ ๋ฐฐ๊ฒฝ ์ง€์‹์ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ณต์žกํ•œ ์ž‘์—…์„ ์ž˜ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” prompt engineering์ด ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์˜ˆ์ œ prompts๋Š” ๋ณด๋‹ค ์ž์„ธํ•œ ์ž‘์—… ์„ค๋ช…(_์˜ˆ: ์ถœ๋ ฅ ํ˜•์‹ ๋ฐ ์ž‘์—… ๋ชฉํ‘œ)์„ ์ œ๊ณตํ•˜์—ฌ ChatGPT๊ฐ€ ๋ณต์žกํ•œ ์ž‘์—… ์š”๊ตฌ ์‚ฌํ•ญ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ณ  ์ด๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_๋ณด๋‹ค ๋ณต์žกํ•œ ์ž‘์—…์€ ChatGPT์— ๋Œ€ํ•œ ์‹ ์ค‘ํ•œ ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง์˜ ํ˜œํƒ์„ ๋” ๋งŽ์ด ๋ˆ„๋ฆด ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค._ ์œ„ํ‚คํŒฉํŠธ ๋ฐ ์ƒ‰์ƒ ๊ฐ์ฒด ์ž‘์—…์—์„œ ์„ค๊ณ„๋œ ํ”„๋กฌํ”„ํŠธ๋Š” ChatGPT์˜ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค(_์ฆ‰, ์œ„ํ‚คํŒฉํŠธ์˜ ๊ฒฝ์šฐ 23.61์—์„œ 28.47๋กœ, ์ƒ‰์ƒ ๊ฐ์ฒด์˜ ๊ฒฝ์šฐ 53.20์—์„œ 66.75๋กœ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค). ์ด๋Ÿฌํ•œ ์ž‘์—…์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํŠน์ • ์ถœ๋ ฅ ํ˜•์‹์„ ๊ฐ€์ง€๊ฑฐ๋‚˜ ๋ฐฐ๊ฒฝ ์ง€์‹์ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ณต์žกํ•œ ์ž‘์—…์„ ์ž˜ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง์ด ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์˜ˆ์ œ ํ”„๋กฌํ”„ํŠธ๋Š” ๋ณด๋‹ค ์ž์„ธํ•œ ์ž‘์—… ์„ค๋ช…(_์˜ˆ: ์ถœ๋ ฅ ํ˜•์‹ ๋ฐ ์ž‘์—… ๋ชฉํ‘œ)์„ ์ œ๊ณตํ•˜์—ฌ ChatGPT๊ฐ€ ๋ณต์žกํ•œ ์ž‘์—… ์š”๊ตฌ ์‚ฌํ•ญ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ณ  ์ด๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_More complex tasks can benefit more from careful prompt engineering on ChatGPT._ In the WikiFact and Colored Objects tasks, the designed prompts have greatly improved the performance of ChatGPT, _i.e._, from 23.61 to 28.47 on WikiFact and from 53.20 to 66.75 on Colored Objects. It indicates the necessity of prompt engineering for LLMs to perform well on complex tasks, since these tasks typically have specific output formats or require background knowledge. Our example prompts provide more detailed task description (_e.g._, output format and task goal), which can help ChatGPT better understand the complex task requirement for fulfilling it. [|assistant|]
(Alemohammad et al., 2023)์€ '์ž๊ธฐ ์†Œ๋น„' ๋˜๋Š” '์ž๊ฐ€ ํก์ˆ˜' ๋ฃจํ”„๋ผ๊ณ  ํ•˜๋Š” ์ƒ์„ฑ ์ด๋ฏธ์ง€ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์  ๋ฐ ๋ถ„์„์  ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ๊ฐ ์ƒ์„ฑ ์‹œ ์ถฉ๋ถ„ํ•œ ์‹ ์„ ํ•œ ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ์—†์œผ๋ฉด ํ–ฅํ›„ ๋ชจ๋ธ์˜ ์ •๋ฐ€๋„ ๋˜๋Š” ๋ฆฌ์ฝœ์ด ๋ฐ˜๋“œ์‹œ ๊ฐ์†Œํ•  ๊ฒƒ์ด๋ผ๊ณ  ๊ฒฐ๋ก ์ง€์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ๊ฐ๊ฐ์˜ ์ƒˆ๋กœ์šด AI ์ƒ์„ฑ์˜ ์˜ํ–ฅ์„ ์ผ๋ฐ˜์ ์ธ _sampling bias_\(0\leq\lambda\leq 1\)์„ ํ†ตํ•ด ๋ชจ๋ธ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€ ์ƒ์„ฑ์˜ ๊ฒฝ์šฐ, ์ด๋Š” ๋‹ค์–‘์„ฑ๋ณด๋‹ค ํ’ˆ์งˆ์„ ์šฐ์„ ์‹œํ•˜๋Š” ์ƒ์„ฑ ์‹œ ํŠน์ง• ํŒŒ๋ผ๋ฏธํ„ฐ(์ ์ ˆํ•˜๊ฒŒ ์ •๋Ÿ‰ํ™”๋จ)๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋ณด๋‹ค ์ •ํ™•ํ•˜๊ฒŒ๋Š”, \(\lambda=1\)์€ ํŽธํ–ฅ๋˜์ง€ ์•Š์€ ์ƒ˜ํ”Œ๋ง์— ํ•ด๋‹นํ•˜๊ณ , \(\lambda=0\)์€ ์ƒ์„ฑ ๋ถ„ํฌ์˜ ๋ชจ๋“œ์—์„œ ์ƒ˜ํ”Œ๋งํ•˜๋Š” ๊ฒƒ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. \(\lambda\)๋Š” BigGAN ๋ฐ StyleGAN์˜ truncation์ด๋‚˜ ํ™•์‚ฐ ๋ชจ๋ธ์˜ guidance์™€ ๊ฐ™์ด ์ƒ์„ฑ ๋ชจ๋ธ๋ง ์‹ค์Šต์—์„œ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ํŽธํ–ฅ ์ƒ˜ํ”Œ๋ง ๋ฐฉ๋ฒ•์„ ๋ชจ๋ธ๋งํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์šฐ์Šค ๋ถ„ํฌ์˜ ๊ฒฝ์šฐ, \(\lambda\)๋Š” ๋‹ค์Œ ์„ธ๋Œ€์˜ ๋ถ„์‚ฐ์˜ ์ถ•์†Œ ๊ณ„์ˆ˜์ž…๋‹ˆ๋‹ค. ์ด๋“ค์˜ ์‹ค์ฆ ์—ฐ๊ตฌ๋Š” ์ด๋ก ์  ๊ด€์ฐฐ๊ณผ ์‹ค์ฆ์  ๊ด€์ฐฐ์„ ์œ„ํ•ด ์ด๋ฏธ์ง€ ์ƒ์„ฑ์„ ์œ„ํ•œ GAN๊ณผ ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ํ™•์‚ฐ ํ™•๋ฅ  ๋ชจ๋ธ๊ณผ ๋‹จ์ผ ๊ฐ€์šฐ์Šค๋ฅผ FFHQ์™€ MNIST์— ๋Œ€ํ•ด ์—ฐ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. (Shumailov et al., 2023)์—์„œ์™€ ๊ฐ™์ด ์ถ”์ • ์˜ค์ฐจ๋งŒ์œผ๋กœ๋„ ๋ฐ˜๋ณต ํšŸ์ˆ˜์— ๋”ฐ๋ผ ๋ถ„์‚ฐ์ด ์‚ฌ๋ผ์ง€๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค(๋‹จ์ผ ๊ฐ€์šฐ์Šค์˜ ๊ฒฝ์šฐ ์ด๋ฅผ ์ฆ๋ช…ํ•จ). (Alemohammad et al., 2023)์€ ๋˜ํ•œ ์ƒ๋‹นํ•œ ์–‘์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ๊ถ๊ทน์ ์ธ ์ €ํ•˜๋ฅผ ์ดˆ๋ž˜ํ•˜๊ธฐ ์ „์— ์ ์€ ์–‘์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ์›๋ณธ ๋ฐ์ดํ„ฐ์™€ ํ˜ผํ•ฉ๋˜๋Š” ์ฒด์ œ์—์„œ ์„ฑ๋Šฅ์ด ์ดˆ๊ธฐ ๋ถ€์ŠคํŠธ๋˜๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ์™€ ํ˜ผํ•ฉ๋œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ์ผ๋ถ€ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ฃผ๋Š” ๋Œ€๊ทœ๋ชจ ๊ฒฐ๊ณผ๋ฅผ ๋ชจ๋ฐฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์•„๋ž˜ 'ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์ด์ ' ์ฐธ์กฐ). ์‹ค์ œ๋กœ ๊ฐ€์žฅ ๋‹จ์ˆœํ•œ ํ˜•ํƒœ์˜ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•(ํšŒ์ „, ์ž๋ฅด๊ธฐ ๋“ฑ)์€ ML ํ›ˆ๋ จ์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋งค์šฐ ์œ ์ตํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ, ๊ฐ€์žฅ ๋‹จ์ˆœํ•œ ํ˜•ํƒœ์˜ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์œผ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
(Alemohammad et al., 2023)์€ '์ž๊ธฐ ์†Œ๋น„' ๋˜๋Š” '์ž๊ฐ€ ํก์ˆ˜' ๋ฃจํ”„๋ผ๊ณ  ํ•˜๋Š” ์ƒ์„ฑ ์ด๋ฏธ์ง€ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์  ๋ฐ ๋ถ„์„์  ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ๊ฐ ์ƒ์„ฑ ์‹œ ์ถฉ๋ถ„ํ•œ ์‹ ์„ ํ•œ ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ์—†์œผ๋ฉด ํ–ฅํ›„ ๋ชจ๋ธ์˜ ์ •๋ฐ€๋„ ๋˜๋Š” ๋ฆฌ์ฝœ์ด ๋ฐ˜๋“œ์‹œ ๊ฐ์†Œํ•  ๊ฒƒ์ด๋ผ๊ณ  ๊ฒฐ๋ก ์ง€์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ๊ฐ๊ฐ์˜ ์ƒˆ๋กœ์šด AI ์ƒ์„ฑ์˜ ์˜ํ–ฅ์„ ์ผ๋ฐ˜์ ์ธ _์ƒ˜ํ”Œ๋ง ํŽธํ–ฅ_\(0\leq\lambda\leq 1\)์„ ํ†ตํ•ด ๋ชจ๋ธ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€ ์ƒ์„ฑ์˜ ๊ฒฝ์šฐ, ์ด๋Š” ๋‹ค์–‘์„ฑ๋ณด๋‹ค ํ’ˆ์งˆ์„ ์šฐ์„ ์‹œํ•˜๋Š” ์ƒ์„ฑ ์‹œ ํŠน์ง• ํŒŒ๋ผ๋ฏธํ„ฐ(์ ์ ˆํ•˜๊ฒŒ ์ •๋Ÿ‰ํ™”๋จ)๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋ณด๋‹ค ์ •ํ™•ํ•˜๊ฒŒ๋Š”, \(\lambda=1\)์€ ํŽธํ–ฅ๋˜์ง€ ์•Š์€ ์ƒ˜ํ”Œ๋ง์— ํ•ด๋‹นํ•˜๊ณ , \(\lambda=0\)์€ ์ƒ์„ฑ ๋ถ„ํฌ์˜ ๋ชจ๋“œ์—์„œ ์ƒ˜ํ”Œ๋งํ•˜๋Š” ๊ฒƒ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ๋žŒ๋‹ค(\lambda)๋Š” BigGAN ๋ฐ StyleGAN์˜ ํŠธ๋Ÿฐ์ผ€์ด์…˜์ด๋‚˜ ํ™•์‚ฐ ๋ชจ๋ธ์˜ ๊ฐ€์ด๋“œ์™€ ๊ฐ™์ด ์ƒ์„ฑ ๋ชจ๋ธ๋ง ์‹ค์Šต์—์„œ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ํŽธํ–ฅ ์ƒ˜ํ”Œ๋ง ๋ฐฉ๋ฒ•์„ ๋ชจ๋ธ๋งํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์šฐ์Šค ๋ถ„ํฌ์˜ ๊ฒฝ์šฐ, \(\lambda\)๋Š” ๋‹ค์Œ ์„ธ๋Œ€์˜ ๋ถ„์‚ฐ์˜ ์ถ•์†Œ ๊ณ„์ˆ˜์ž…๋‹ˆ๋‹ค. ์ด๋“ค์˜ ์‹ค์ฆ ์—ฐ๊ตฌ๋Š” ์ด๋ก ์  ๊ด€์ฐฐ๊ณผ ์‹ค์ฆ์  ๊ด€์ฐฐ์„ ์œ„ํ•ด ์ด๋ฏธ์ง€ ์ƒ์„ฑ์„ ์œ„ํ•œ GAN๊ณผ ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ํ™•์‚ฐ ํ™•๋ฅ  ๋ชจ๋ธ๊ณผ ๋‹จ์ผ ๊ฐ€์šฐ์Šค๋ฅผ FFHQ์™€ MNIST์— ๋Œ€ํ•ด ์—ฐ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. (Shumailov et al., 2023)์—์„œ์™€ ๊ฐ™์ด ์ถ”์ • ์˜ค์ฐจ๋งŒ์œผ๋กœ๋„ ๋ฐ˜๋ณต ํšŸ์ˆ˜์— ๋”ฐ๋ผ ๋ถ„์‚ฐ์ด ์‚ฌ๋ผ์ง€๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค(๋‹จ์ผ ๊ฐ€์šฐ์Šค์˜ ๊ฒฝ์šฐ ์ด๋ฅผ ์ฆ๋ช…ํ•จ). (์•Œ๋ ˆ๋ชจํ•˜๋งˆ๋“œ ๋“ฑ, 2023)์€ ๋˜ํ•œ ์ƒ๋‹นํ•œ ์–‘์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ๊ถ๊ทน์ ์ธ ์ €ํ•˜๋ฅผ ์ดˆ๋ž˜ํ•˜๊ธฐ ์ „์— ์ ์€ ์–‘์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ์›๋ณธ ๋ฐ์ดํ„ฐ์™€ ํ˜ผํ•ฉ๋˜๋Š” ์ฒด์ œ์—์„œ ์„ฑ๋Šฅ์ด ์ดˆ๊ธฐ ๋ถ€์ŠคํŠธ๋˜๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ์™€ ํ˜ผํ•ฉ๋œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ์ผ๋ถ€ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ฃผ๋Š” ๋Œ€๊ทœ๋ชจ ๊ฒฐ๊ณผ๋ฅผ ๋ชจ๋ฐฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์•„๋ž˜ 'ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์ด์ ' ์ฐธ์กฐ). ์‹ค์ œ๋กœ ๊ฐ€์žฅ ๋‹จ์ˆœํ•œ ํ˜•ํƒœ์˜ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•(ํšŒ์ „, ์ž๋ฅด๊ธฐ ๋“ฑ)์€ ML ํ›ˆ๋ จ์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋งค์šฐ ์œ ์ตํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ, ๊ฐ€์žฅ ๋‹จ์ˆœํ•œ ํ˜•ํƒœ์˜ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์œผ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|](Alemohammad et al., 2023) conduct an empirical and analytical analysis on generative image models of what they term the "self-consuming" or "autophaguous" loop. They conclude that without enough fresh real data at each generation, future models necessarily will have their precision or recall decrease. They model the influence of each new AI-generation via a generic _sampling bias_\(0\leq\lambda\leq 1\). In the case of image generation this refers to feature parameters at generation that favor quality over diversity (suitably quantified). More precisely, \(\lambda=1\) corresponds to unbiased sampling and \(\lambda=0\) corresponds to sampling from the modes of the generative distribution. \(\lambda\) models biased sampling methods commonly used in generative modeling practice, such as truncation in BigGAN and StyleGAN or guidance in diffusion models. In the case of Gaussian distributions, \(\lambda\) is the shrinking factor of the variance of the next generation. Their empirical work studies GANs and denoising diffusion probabilistic models for image generation on FFHQ and MNIST and single Gaussians for both theoretical and empirical observations. As in (Shumailov et al., 2023) they observe (and prove for the case of a single Gaussian) that estimation error alone leads to vanishing variance with number of iterations. (Alemohammad et al., 2023) also empirically observe an initial boost in performance in a regime where modest amounts of synthetic data are mixed with the original data before larger amounts of synthetic data lead to ultimate degradation. This might mimick larger-scale results that demonstrate how synthetic data mixed with true data improves performance in some scenarios (see _Benefits of synthesized data_ below). Indeed, in its simplest form, data augmentation (rotations, cropping etc. ), a widespread highly beneficial practice in ML training, can be viewed as the simplest form of data generation. [|assistant|]
์น˜์œ„์•ˆ ์ฒธ์€ ์ค‘๊ตญ ์ธ๋ฏผ๋Œ€ํ•™๊ต ๊ฐ€์˜ค๋ง ์ธ๊ณต์ง€๋Šฅ ๋Œ€ํ•™์›์—์„œ ๋ฐ•์‚ฌ ํ•™์œ„๋ฅผ ์ทจ๋“ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ์˜ ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ Large Language Model์— ๊ธฐ๋ฐ˜ํ•œ ์–ธ์–ด ๋ชจ๋ธ ์ถ”๋ก ๊ณผ ์—์ด์ „ํŠธ์— ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์น˜์œ„์•ˆ ์ฒธ์€ ์ค‘๊ตญ ์ธ๋ฏผ๋Œ€ํ•™๊ต ๊ฐ€์˜ค๋ง ์ธ๊ณต์ง€๋Šฅ ๋Œ€ํ•™์›์—์„œ ๋ฐ•์‚ฌ ํ•™์œ„๋ฅผ ์ทจ๋“ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ์˜ ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์— ๊ธฐ๋ฐ˜ํ•œ ์–ธ์–ด ๋ชจ๋ธ ์ถ”๋ก ๊ณผ ์—์ด์ „ํŠธ์— ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Zhi-Yuan Chen is pursuing his Ph.D. in Gaoling school of Artificial Intelligence, Renmin University of China. His research mainly focuses on language model reasoning and agent based on large language model. [|assistant|]
* **๋„๋ฉ”์ธ ์ง€์‹ ๊ฒฐํ•**: ์–ธ์–ด ๋ชจ๋ธ์€ ๋…ธ์ถœ๋˜์ง€ ์•Š์€ ํŠน์ • ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์ „๋ฌธ ์ง€์‹์ด ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์œŒ๋ฆฌ์—„ ์…ฐ์ต์Šคํ”ผ์–ด๊ฐ€ ์“ด ํ…์ŠคํŠธ๋กœ๋งŒ ํ•™์Šตํ•œ ๋ชจ๋ธ์€ ๋งˆํฌ ํŠธ์›จ์ธ์˜ ์ž‘ํ’ˆ์— ๋Œ€ํ•ด ์งˆ๋ฌธ์„ ๋ฐ›์œผ๋ฉด ์„ฑ๋Šฅ์ด ์ €ํ•˜๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * **Outdated Information**: LLM์€ ํ•ญ์ƒ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์˜ํ•ด ๊ฒฐ์ •๋œ ๋งˆ๊ฐ์ผ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋งˆ์ง€๋ง‰ ํ•™์Šต ์—…๋ฐ์ดํŠธ ์ดํ›„ ๋ฐœ์ƒํ•œ ๋ชจ๋“  ์ด๋ฒคํŠธ, ๋ฐœ๊ฒฌ ๋˜๋Š” ๋ณ€๊ฒฝ ์‚ฌํ•ญ์€ ์™ธ๋ถ€ ์†Œ์Šค์— ์•ก์„ธ์Šคํ•˜์ง€ ์•Š๊ณ ๋Š” ๋ชจ๋ธ์˜ ์ง€์‹์— ํฌํ•จ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. * **Immemorization**: ๋•Œ๋•Œ๋กœ ๋ชจ๋ธ์€ ํ•™์Šต ๊ณผ์ •์—์„œ ์ง€์‹์— ๋…ธ์ถœ๋˜์ง€๋งŒ ์ด๋ฅผ ๊ธฐ์–ตํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๊ฑฐ์˜ ๋‚˜ํƒ€๋‚˜์ง€ ์•Š๋Š” ํฌ๊ท€ํ•œ ์‚ฌ์‹ค์˜ ๊ฒฝ์šฐ ํŠนํžˆ ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค(Kandpal et al., 2023). * **Forgetting**: ์–ธ์–ด ๋ชจ๋ธ์€ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„ ์ดํ›„์— ์ถ”๊ฐ€ ํ•™์Šต(Fine-tuning)์„ ๊ฑฐ์น˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ฒฝ์šฐ์— ๋”ฐ๋ผ ์ด๋กœ ์ธํ•ด ๋ชจ๋ธ์ด Fine-tuning ํ”„๋กœ์„ธ์Šค ์ด์ „์˜ ์ง€์‹ ์ค‘ ์ผ๋ถ€๋ฅผ ์žƒ์–ด๋ฒ„๋ฆฌ๋Š” 'catastrophic forgetting'(Kirkpatrick et al., 2017; Goodfellow et al., 2013; Chen et al., 2020; Luo et al., 2023)์ด๋ผ๋Š” ํ˜„์ƒ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * **Reasoning Failure**: ํŠน์ • ๊ฒฝ์šฐ ์–ธ์–ด ๋ชจ๋ธ์ด ์‚ฌ์‹ค์— ๋Œ€ํ•œ ๊ด€๋ จ ์ง€์‹์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์ง€๋งŒ ์ด๋ฅผ ์ œ๋Œ€๋กœ ํ™œ์šฉํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ณต์žกํ•œ ๋‹ค๋‹จ๊ณ„ ์ถ”๋ก  ๊ณผ์ œ(Tan et al., 2023)์—์„œ ๋˜๋Š” ๋™์ผํ•œ ์‚ฌ์‹ค์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ์งˆ๋ฌธ์„ ๋ฐ›์•˜์„ ๋•Œ ๊ฒฐ๊ณผ๊ฐ€ ์ƒ์ดํ•˜๊ฒŒ ๋‚˜ํƒ€๋‚  ๋•Œ ํŠนํžˆ ๋‘๋“œ๋Ÿฌ์ง‘๋‹ˆ๋‹ค(Berglund et al., 2023).[|endofturn|]
* **๋„๋ฉ”์ธ ์ง€์‹ ๊ฒฐํ•**: ์–ธ์–ด ๋ชจ๋ธ์€ ๋…ธ์ถœ๋˜์ง€ ์•Š์€ ํŠน์ • ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์ „๋ฌธ ์ง€์‹์ด ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์œŒ๋ฆฌ์—„ ์…ฐ์ต์Šคํ”ผ์–ด๊ฐ€ ์“ด ํ…์ŠคํŠธ๋กœ๋งŒ ํ•™์Šตํ•œ ๋ชจ๋ธ์€ ๋งˆํฌ ํŠธ์›จ์ธ์˜ ์ž‘ํ’ˆ์— ๋Œ€ํ•ด ์งˆ๋ฌธ์„ ๋ฐ›์œผ๋ฉด ์„ฑ๋Šฅ์ด ์ €ํ•˜๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * **๊ตฌํ˜• ์ •๋ณด**: LLM์€ ํ•ญ์ƒ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์˜ํ•ด ๊ฒฐ์ •๋œ ๋งˆ๊ฐ์ผ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋งˆ์ง€๋ง‰ ํ•™์Šต ์—…๋ฐ์ดํŠธ ์ดํ›„ ๋ฐœ์ƒํ•œ ๋ชจ๋“  ์ด๋ฒคํŠธ, ๋ฐœ๊ฒฌ ๋˜๋Š” ๋ณ€๊ฒฝ ์‚ฌํ•ญ์€ ์™ธ๋ถ€ ์†Œ์Šค์— ์•ก์„ธ์Šคํ•˜์ง€ ์•Š๊ณ ๋Š” ๋ชจ๋ธ์˜ ์ง€์‹์— ํฌํ•จ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. * **๊ธฐ์–ต ์ƒ์‹ค**: ๋•Œ๋•Œ๋กœ ๋ชจ๋ธ์€ ํ•™์Šต ๊ณผ์ •์—์„œ ์ง€์‹์— ๋…ธ์ถœ๋˜์ง€๋งŒ ์ด๋ฅผ ๊ธฐ์–ตํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๊ฑฐ์˜ ๋‚˜ํƒ€๋‚˜์ง€ ์•Š๋Š” ํฌ๊ท€ํ•œ ์‚ฌ์‹ค์˜ ๊ฒฝ์šฐ ํŠนํžˆ ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค(Kandpal et al., 2023). * **์žŠ์–ด๋ฒ„๋ฆผ**: ์–ธ์–ด ๋ชจ๋ธ์€ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„ ์ดํ›„์— ์ถ”๊ฐ€ ํ•™์Šต(๋ฏธ์„ธ ์กฐ์ •)์„ ๊ฑฐ์น˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ฒฝ์šฐ์— ๋”ฐ๋ผ ์ด๋กœ ์ธํ•ด ๋ชจ๋ธ์ด ๋ฏธ์„ธ ์กฐ์ • ํ”„๋กœ์„ธ์Šค ์ด์ „์˜ ์ง€์‹ ์ค‘ ์ผ๋ถ€๋ฅผ ์žƒ์–ด๋ฒ„๋ฆฌ๋Š” '์žฌ์•™์  ๋ง๊ฐ'(Kirkpatrick et al., 2017; Goodfellow et al., 2013; Chen et al., 2020; Luo et al., 2023)์ด๋ผ๋Š” ํ˜„์ƒ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * **์ถ”๋ก  ์‹คํŒจ**: ํŠน์ • ๊ฒฝ์šฐ ์–ธ์–ด ๋ชจ๋ธ์ด ์‚ฌ์‹ค์— ๋Œ€ํ•œ ๊ด€๋ จ ์ง€์‹์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์ง€๋งŒ ์ด๋ฅผ ์ œ๋Œ€๋กœ ํ™œ์šฉํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ณต์žกํ•œ ๋‹ค๋‹จ๊ณ„ ์ถ”๋ก  ๊ณผ์ œ(Tan et al., 2023)์—์„œ ๋˜๋Š” ๋™์ผํ•œ ์‚ฌ์‹ค์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ์งˆ๋ฌธ์„ ๋ฐ›์•˜์„ ๋•Œ ๊ฒฐ๊ณผ๊ฐ€ ์ƒ์ดํ•˜๊ฒŒ ๋‚˜ํƒ€๋‚  ๋•Œ ํŠนํžˆ ๋‘๋“œ๋Ÿฌ์ง‘๋‹ˆ๋‹ค(Berglund et al., 2023).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* **Domain knowledge deficit**: A language model may lack comprehensive expertise in a specific domain to which it has not been exposed. For example, a model trained exclusively on texts written by William Shakespeare would perform poorly when asked about the works of Mark Twain. * **Outdated Information**: LLMs invariably have a cutoff date determined by their training dataset. Consequently, any events, discoveries, or changes occurring after the last training update will not be within the model's knowledge without access to external sources. * **Immemorization**: Sometimes, a model is exposed to knowledge during its training process but does not retain it. This is especially true for rare facts that appear in the training dataset only scarcely (Kandpal et al., 2023). * **Forgetting**: Language models often undergo additional training after the pre-training phase (fine-tuning). In some cases, this might lead to a phenomenon called _catastrophic forgetting_(Kirkpatrick et al., 2017; Goodfellow et al., 2013; Chen et al., 2020; Luo et al., 2023), where models lose some of the knowledge they had prior to the fine-tuning process. * **Reasoning Failure**: In certain instances, a language model might possess relevant knowledge about a fact but fail to utilize it properly. This is particularly evident in complex multi-step reasoning tasks (Tan et al., 2023) or when posed with different questions about the same fact, resulting in disparate outcomes (Berglund et al., 2023). [|assistant|]
**์–ธ์–ด ๋ชจ๋ธ์˜ ์ˆœ์„œ ํšจ๊ณผ.** ์ผ๋ถ€ ์„ ํ–‰ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด language models๋Š” ์–ด์ ˆ์˜ ๋ฌด์ž‘์œ„ ์ˆœ์—ด ํ›„์—๋„ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•ฉ๋ฆฌ์ ์ธ ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•˜๋Š” ๋“ฑ ์–ด๋Š ์ •๋„ ์ˆœ์—ด๋œ ํ…์ŠคํŠธ๋ฅผ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Abdou et al., 2022; Sinha et al., 2020). ๋˜ํ•œ, Cao et al. (2023)์€ ๋งŽ์€ ๋น„์œจ์˜ ๋‹จ์–ด๊ฐ€ ์Šคํฌ๋žจ๋ธ”๋ง๋œ ๊ฒฝ์šฐ์—๋„ GPT-4๋Š” ์—ฌ๋Ÿฌ reasoning benchmarks์—์„œ ์—ฌ์ „ํžˆ ๊ดœ์ฐฎ์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ ์ˆœ์—ด๋œ ํ…์ŠคํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋ถ€์ž์—ฐ์Šค๋Ÿฝ๊ณ  ๋ฌด์˜๋ฏธํ•œ ๊ฒƒ๊ณผ๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์šฐ๋ฆฌ์˜ ์ „์ œ ์ˆœ์—ด์€ ์˜๋ฏธ๋ก ์  ์˜๋ฏธ๋ฅผ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š๊ณ  ๊ตฌ๋ฌธ๋ก ์ ์œผ๋กœ ์œ ํšจํ•ฉ๋‹ˆ๋‹ค(์ˆ˜๋™์œผ๋กœ ํ™•์ธ). ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์šฐ๋ฆฌ๋Š” LLM reasoning ์„ฑ๋Šฅ์ด ์ „์ œ์˜ ์ˆœ์„œ์— ๋งค์šฐ ์ทจ์•ฝํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
**์–ธ์–ด ๋ชจ๋ธ์˜ ์ˆœ์„œ ํšจ๊ณผ.** ์ผ๋ถ€ ์„ ํ–‰ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์–ธ์–ด ๋ชจ๋ธ์€ ์–ด์ ˆ์˜ ๋ฌด์ž‘์œ„ ์ˆœ์—ด ํ›„์—๋„ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•ฉ๋ฆฌ์ ์ธ ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•˜๋Š” ๋“ฑ ์–ด๋Š ์ •๋„ ์ˆœ์—ด๋œ ํ…์ŠคํŠธ๋ฅผ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Abdou et al., 2022; Sinha et al., 2020). ๋˜ํ•œ, Cao et al. (2023)์€ ๋งŽ์€ ๋น„์œจ์˜ ๋‹จ์–ด๊ฐ€ ์Šคํฌ๋žจ๋ธ”๋ง๋œ ๊ฒฝ์šฐ์—๋„ GPT-4๋Š” ์—ฌ๋Ÿฌ ์ถ”๋ก  ๋ฒค์น˜๋งˆํฌ์—์„œ ์—ฌ์ „ํžˆ ๊ดœ์ฐฎ์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ ์ˆœ์—ด๋œ ํ…์ŠคํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋ถ€์ž์—ฐ์Šค๋Ÿฝ๊ณ  ๋ฌด์˜๋ฏธํ•œ ๊ฒƒ๊ณผ๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์šฐ๋ฆฌ์˜ ์ „์ œ ์ˆœ์—ด์€ ์˜๋ฏธ๋ก ์  ์˜๋ฏธ๋ฅผ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š๊ณ  ๊ตฌ๋ฌธ๋ก ์ ์œผ๋กœ ์œ ํšจํ•ฉ๋‹ˆ๋‹ค(์ˆ˜๋™์œผ๋กœ ํ™•์ธ). ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์šฐ๋ฆฌ๋Š” LLM ์ถ”๋ก  ์„ฑ๋Šฅ์ด ์ „์ œ์˜ ์ˆœ์„œ์— ๋งค์šฐ ์ทจ์•ฝํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Order effect of language models.** Some prior works show that language models are able to understand permuted texts to some extent, i.e., after a random permutation of words, models usually preserve a reasonable performance (Abdou et al., 2022; Sinha et al., 2020). Moreover, Cao et al. (2023) shows that even when a large fraction of words are scrambled, GPT-4 still achieves decent performance on several reasoning benchmarks. In contrast to permuted texts in these works that are typically unnatural and nonsensical, our premise order permutations do not alter the semantic meaning and remain syntactically valid (we manually verify this). Nevertheless, we demonstrate that LLM reasoning performance is highly brittle to the ordering of the premises. [|assistant|]
๊ทธ๋ฆผ 4: **์‹คํ—˜ 1: ์ˆœ์„œ๊ฐ€ ๋’ค๋ฐ”๋€Œ๋ฉด ๋ชจ๋ธ์ด ์˜ฌ๋ฐ”๋ฅธ ์ด๋ฆ„์˜ ํ™•๋ฅ ์„ ๋†’์ด์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ทธ๋ž˜ํ”„๋Š” ๋ชจ๋ธ์ด ๊ด€๋ จ ์„ค๋ช…์œผ๋กœ ์ฟผ๋ฆฌ๋  ๋•Œ ์˜ฌ๋ฐ”๋ฅธ ์ด๋ฆ„(๋ฌด์ž‘์œ„ ์ด๋ฆ„ ๋Œ€๋น„)์— ๋Œ€ํ•œ ํ‰๊ท  ๋กœ๊ทธ ํ™•๋ฅ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‰๊ท ์€ ๋ชจ๋ธ ํฌ๊ธฐ๋‹น 30์Œ๊ณผ 3๊ฐœ์˜ Fine-tuning ์‹œ๋“œ์— ๋Œ€ํ•ด ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 4: **์‹คํ—˜ 1: ์ˆœ์„œ๊ฐ€ ๋’ค๋ฐ”๋€Œ๋ฉด ๋ชจ๋ธ์ด ์˜ฌ๋ฐ”๋ฅธ ์ด๋ฆ„์˜ ํ™•๋ฅ ์„ ๋†’์ด์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ทธ๋ž˜ํ”„๋Š” ๋ชจ๋ธ์ด ๊ด€๋ จ ์„ค๋ช…์œผ๋กœ ์ฟผ๋ฆฌ๋  ๋•Œ ์˜ฌ๋ฐ”๋ฅธ ์ด๋ฆ„(๋ฌด์ž‘์œ„ ์ด๋ฆ„ ๋Œ€๋น„)์— ๋Œ€ํ•œ ํ‰๊ท  ๋กœ๊ทธ ํ™•๋ฅ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‰๊ท ์€ ๋ชจ๋ธ ํฌ๊ธฐ๋‹น 30์Œ๊ณผ 3๊ฐœ์˜ ๋ฏธ์„ธ ์กฐ์ • ์‹œ๋“œ์— ๋Œ€ํ•ด ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 4: **Experiment 1: Models fail to increase the probability of the correct name when the order is reversed. The graph shows the average log-probability for the correct name (vs. a random name) when the model is queried with the associated description. The average is taken over 30 pairs and 3 finetuning seeds per model size. [|assistant|]
ํ˜„์žฌ ์ด ๋ฐฉ๋ฒ•์€ ์‚ฌ์šฉ์ž๊ฐ€ ์ง„ํ™” ๊ฒ€์ƒ‰์˜ ์žฌ๋ฃŒ๋กœ ์‚ฌ์šฉํ•  ์†Œ์Šค ๋ชจ๋ธ ์„ธํŠธ๋ฅผ ์„ ํƒํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ง„ํ™”๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๊ธฐ์กด ๋ชจ๋ธ์˜ ๋ฐฉ๋Œ€ํ•œ ๋ชจ์ง‘๋‹จ์—์„œ ํ›„๋ณด ์†Œ์Šค ๋ชจ๋ธ์„ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒƒ๋„ ๊ฐ€๋Šฅํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ ์„ ํƒ ์™ธ์—๋„ ์ง„ํ™”๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ๊ฐ ๊ณ ์œ ํ•œ ํ‹ˆ์ƒˆ ์‹œ์žฅ๊ณผ ํ–‰๋™์„ ๊ฐ€์ง„ ๋‹ค์–‘ํ•œ foundation models์˜ ๊ตฐ์ง‘์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ๋„ ๋ชจ์ƒ‰ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ƒํ˜ธ ์ž‘์šฉ์„ ํ†ตํ•ด ์„ธ๊ณ„์˜ ์ƒˆ๋กœ์šด ์ƒํ˜ธ ๋ณด์™„์ ์ธ ๋‚ด๋ถ€ ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์ƒ์„ฑํ•˜์—ฌ ์ž๊ธฐ ๊ฐœ์„ ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ ๊ตฐ์ง‘์œผ๋กœ ๊ตฌ์„ฑ๋œ ์ง‘๋‹จ ์ง€๋Šฅ์˜ ์ถœํ˜„์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
ํ˜„์žฌ ์ด ๋ฐฉ๋ฒ•์€ ์‚ฌ์šฉ์ž๊ฐ€ ์ง„ํ™” ๊ฒ€์ƒ‰์˜ ์žฌ๋ฃŒ๋กœ ์‚ฌ์šฉํ•  ์†Œ์Šค ๋ชจ๋ธ ์„ธํŠธ๋ฅผ ์„ ํƒํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ง„ํ™”๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๊ธฐ์กด ๋ชจ๋ธ์˜ ๋ฐฉ๋Œ€ํ•œ ๋ชจ์ง‘๋‹จ์—์„œ ํ›„๋ณด ์†Œ์Šค ๋ชจ๋ธ์„ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒƒ๋„ ๊ฐ€๋Šฅํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ ์„ ํƒ ์™ธ์—๋„ ์ง„ํ™”๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ๊ฐ ๊ณ ์œ ํ•œ ํ‹ˆ์ƒˆ ์‹œ์žฅ๊ณผ ํ–‰๋™์„ ๊ฐ€์ง„ ๋‹ค์–‘ํ•œ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์˜ ๊ตฐ์ง‘์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ๋„ ๋ชจ์ƒ‰ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ƒํ˜ธ ์ž‘์šฉ์„ ํ†ตํ•ด ์„ธ๊ณ„์˜ ์ƒˆ๋กœ์šด ์ƒํ˜ธ ๋ณด์™„์ ์ธ ๋‚ด๋ถ€ ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์ƒ์„ฑํ•˜์—ฌ ์ž๊ธฐ ๊ฐœ์„ ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ ๊ตฐ์ง‘์œผ๋กœ ๊ตฌ์„ฑ๋œ ์ง‘๋‹จ ์ง€๋Šฅ์˜ ์ถœํ˜„์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The method currently requires the user to select a set of source models to use as ingredients for evolutionary search. We believe it is also possible to leverage evolution to search for candidate source models from a vast population of existing models as well. In addition to model selection, we are also exploring using evolution to produce swarms of diverse foundation models each with its own niche and behaviors. This holds the potential of enabling the emergence of a collective intelligence consisting of a swarm of models capable of self-improvement by continuously producing new complementary internal models of the world through interaction. [|assistant|]
์—ฌ๊ธฐ์„œ, \(F\)๋Š” \(W(Q)\)์— ์˜ํ•ด ํŒŒ๋ผ๋ฏธํ„ฐํ™”๋˜๊ณ  ์ด์ „ ๋ฐ์ดํ„ฐ \(Q_{1},\ldots,Q_{i-1}\), ๋ฐ ์ƒˆ๋กœ์šด ๋‚œ์ˆ˜ \(R\)์— ์˜์กดํ•  ์ˆ˜ ์žˆ๋‹ค.[|endofturn|]
์—ฌ๊ธฐ์„œ, F๋Š” W(Q)์— ์˜ํ•ด ํŒŒ๋ผ๋ฏธํ„ฐํ™”๋˜๊ณ  ์ด์ „ ๋ฐ์ดํ„ฐ \(Q_{1},\ldots,Q_{i-1}\), ๋ฐ ์ƒˆ๋กœ์šด ๋‚œ์ˆ˜ \(R\)์— ์˜์กดํ•  ์ˆ˜ ์žˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]_Here, \(F\) is parameterized by \(W(Q)\) and may rely on previous data \(Q_{1},\ldots,Q_{i-1}\), and new randomness \(R\). [|assistant|]
ํŠน์ • prompting ๊ธฐ๋ฒ•์ด RAG ์ค€์ˆ˜์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ ์ •๋„๋ฅผ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด GPT-4์—์„œ ๋‘ ๊ฐ€์ง€ ์ถ”๊ฐ€ prompting ("์—„๊ฒฉ" ๋ฐ "๋Š์Šจํ•จ")์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
ํŠน์ • ํ”„๋กฌํ”„ํŠธ ๊ธฐ๋ฒ•์ด RAG ์ค€์ˆ˜์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ ์ •๋„๋ฅผ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด GPT-4์—์„œ ๋‘ ๊ฐ€์ง€ ์ถ”๊ฐ€ ํ”„๋กฌํ”„ํŠธ("์—„๊ฒฉ" ๋ฐ "๋Š์Šจํ•จ")๋ฅผ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To assess the degree of influence that the specific prompting technique has on RAG adherence, we test two additional prompts ("strict" and "loose") on GPT-4. [|assistant|]
์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด LLM์€ prompting์„ ํ†ตํ•ด ๋†€๋ผ์šด In-Context reasoning ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ ํฌ๊ธฐ์— ๋”ฐ๋ผ ํ™•์žฅ๋ฉ๋‹ˆ๋‹ค(Brown et al., 2020; Wei et al., 2022, 2022). ๋ฐ˜๋Œ€๋กœ ์ถ”๋ก ํ•ด์•ผ ํ•  ์ •๋ณด๊ฐ€ ์ž…๋ ฅ์˜ ์ผ๋ถ€๋กœ ๋ช…์‹œ์ ์œผ๋กœ ์ œ๊ณต๋˜์ง€ ์•Š์„ ๊ฒฝ์šฐ, LLM์€ ๋‹จ์ผ ํ™‰ ํ•˜์œ„ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ๋‹ต์„ ์•Œ๊ณ  ์žˆ๋”๋ผ๋„ ์ข…์ข… ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ์ˆ˜ํ–‰ํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค(Ofir Press et al., 2023; Dziri et al., 2023). In-Context reasoning์ด ์–ด๋–ป๊ฒŒ ์ž‘๋™ํ•˜๋Š”์ง€์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์กฐ์‚ฌ๊ฐ€ ์ง„ํ–‰๋˜๊ณ  ์žˆ์ง€๋งŒ(Chan et al., 2022; Akyurek et al., 2023; Dai et al., 2023; Von Oswald et al., 2023; Prystawski and Goodman, 2023; Feng and Steinhardt, 2024), ์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์ด ์–ด๋–ป๊ฒŒ ์ž‘๋™ํ•˜๋Š”์ง€ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•œ ์กฐ์‚ฌ๋Š” ํ™œ๋ฐœํ•˜๊ฒŒ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด LLM์€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด ๋†€๋ผ์šด ์ƒํ™ฉ ๋งž์ถคํ˜• ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ ํฌ๊ธฐ์— ๋”ฐ๋ผ ํ™•์žฅ๋ฉ๋‹ˆ๋‹ค(Brown et al., 2020; Wei et al., 2022, 2022). ๋ฐ˜๋Œ€๋กœ ์ถ”๋ก ํ•ด์•ผ ํ•  ์ •๋ณด๊ฐ€ ์ž…๋ ฅ์˜ ์ผ๋ถ€๋กœ ๋ช…์‹œ์ ์œผ๋กœ ์ œ๊ณต๋˜์ง€ ์•Š์„ ๊ฒฝ์šฐ, LLM์€ ๋‹จ์ผ ํ™‰ ํ•˜์œ„ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ๋‹ต์„ ์•Œ๊ณ  ์žˆ๋”๋ผ๋„ ์ข…์ข… ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ์ˆ˜ํ–‰ํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค(Ofir Press et al., 2023; Dziri et al., 2023). ์ƒํ™ฉ์— ๋งž๋Š” ์ถ”๋ก ์ด ์–ด๋–ป๊ฒŒ ์ž‘๋™ํ•˜๋Š”์ง€์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์กฐ์‚ฌ๊ฐ€ ์ง„ํ–‰๋˜๊ณ  ์žˆ์ง€๋งŒ(Chan et al., 2022; Akyurek et al., 2023; Dai et al., 2023; Von Oswald et al., 2023; Prystawski and Goodman, 2023; Feng and Steinhardt, 2024), ์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์ด ์–ด๋–ป๊ฒŒ ์ž‘๋™ํ•˜๋Š”์ง€ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•œ ์กฐ์‚ฌ๋Š” ํ™œ๋ฐœํ•˜๊ฒŒ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Recent works have shown that LLMs demonstrate remarkable in-context reasoning ability via prompting, which scales with model size (Brown et al., 2020; Wei et al., 2022, 2022). On the contrary, when the information to reason from is not explicitly given as part of the input, LLMsoften fail to correctly perform multi-hop reasoning even when they know the answer to the single-hop sub-step (Ofir Press et al., 2023; Dziri et al., 2023). While there have been wide investigations on how in-context reasoning works (Chan et al., 2022; Akyurek et al., 2023; Dai et al., 2023; Von Oswald et al., 2023; Prystawski and Goodman, 2023; Feng and Steinhardt, 2024), such an investigation has not been actively done to understand how latent multi-hop reasoning works. [|assistant|]
5๊ฐœ์˜ ์—ํฌํฌ์— ๋Œ€ํ•ด 8, 32, 128์˜ Batch Size์™€ 1e-06, 2e-06, 1e-05, 2e-05์˜ Learning Rate๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Llama-7b, Llama-13b, Llama-30b์— ๋Œ€ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์Šค์œ•์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. Optimizer๋กœ Adam์„ ์‚ฌ์šฉํ•˜๊ณ  ๋ฉ”๋ชจ๋ฆฌ ํšจ์œจ์„ฑ์„ ์œ„ํ•ด DeepSpeed ๋ ˆ๋ฒจ 3์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด Fine-tuning์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ๋งค๊ฐœ๋ณ€์ˆ˜ ํšจ์œจ์ ์ธ Fine-tuning ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ Batch Size๋Š” ์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ๊ฒŒ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. Learning Rate๋Š” Llama-1 ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต์— ์‚ฌ์šฉ๋œ Learning Rate(Touvron et al., 2023)์— ๊ฐ€๊น๊ฒŒ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. Llama-7b์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 10์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
5๊ฐœ์˜ ์—ํฌํฌ์— ๋Œ€ํ•ด 8, 32, 128์˜ ๋ฐฐ์น˜ ํฌ๊ธฐ์™€ 1e-06, 2e-06, 1e-05, 2e-05์˜ ํ•™์Šต ์†๋„๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Llama-7b, Llama-13b, Llama-30b์— ๋Œ€ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์Šค์œ•์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ ํ™” ๋„๊ตฌ๋กœ Adam์„ ์‚ฌ์šฉํ•˜๊ณ  ๋ฉ”๋ชจ๋ฆฌ ํšจ์œจ์„ฑ์„ ์œ„ํ•ด DeepSpeed ๋ ˆ๋ฒจ 3์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ๋ฏธ์„ธ ์กฐ์ •์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ๋งค๊ฐœ๋ณ€์ˆ˜ ํšจ์œจ์ ์ธ ๋ฏธ์„ธ ์กฐ์ • ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” ์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ๊ฒŒ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•™์Šต ์†๋„๋Š” Llama-1 ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต์— ์‚ฌ์šฉ๋œ ํ•™์Šต ์†๋„(Touvron et al., 2023)์— ๊ฐ€๊น๊ฒŒ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. Llama-7b์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 10์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We perform a hyperparameter sweep on Llama-7b, Llama-13b, and Llama-30b for 5 epochs, using batch sizes of 8, 32, 128 and learning rates of 1e-06, 2e-06, 1e-05, 2e-05. We use Adam as our optimizer and DeepSpeed level 3 for memory efficiency. We perform full finetuning and do not use any parameter efficient finetuning techniques. We chose these batch sizes to be relatively low. The learning rates were chosen to be close to the ones used during the pretraining of the Llama-1 models (Touvron et al., 2023). The results for Llama-7b are shown in Figure 10. [|assistant|]
์‹คํ—˜์šฐ๋ฆฌ๋Š” ๊ฐ„์ ‘์ ์ธ ์ฆ๊ฑฐ๋ฅผ ํ†ตํ•ด EntRec์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์„ ํ•ฉ๋ฆฌ์ ์œผ๋กœ approximationํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)๋Š” \(\tau_{\text{2H}}\)์˜ ๋งˆ์ง€๋ง‰ ํ† ํฐ์ด ์•„๋‹ˆ๋ผ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์„ค๋ช…์  ์–ธ๊ธ‰์˜ ๋งˆ์ง€๋ง‰ ํ† ํฐ์—์„œ ๊ณ„์‚ฐ๋˜๋ฉฐ, ๋ฌธ๋ฒ•์ ์œผ๋กœ ์‰ผํ‘œ ๋’ค์— \(e_{2}\)์˜ ์ด๋ฆ„์„ ์•ž์— ๋ถ™์ด๋Š” ๊ฒƒ์ด ์ž์—ฐ์Šค๋Ÿฝ์Šต๋‹ˆ๋‹ค(์˜ˆ: "'superstition'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ, _์Šคํ‹ฐ๋น„ ์›๋”_"). ๊ฒฐ๊ณผ ๋ฌธ์ž์—ด์—์„œ ๋ฌธ๋ฒ•์ ์œผ๋กœ \(\mu(r_{1}(e_{1}))\))๋Š” _antecedent_๊ฐ€ ๋˜๊ณ  \(e_{2}\)๋Š” _appositive_๊ฐ€ ๋˜๋ฉฐ, appositive๋Š” ๋‹ค๋ฅธ ๋ช…์‚ฌ๊ตฌ ๋’ค์— ๋ฐ˜๋Œ€๋˜๋Š” ๋ช…์‚ฌ๊ตฌ๊ฐ€ ๋˜์–ด ๊ทธ ๋ช…์‚ฌ๊ตฌ๋ฅผ ๋” ์‹๋ณ„ํ•˜๊ฑฐ๋‚˜ ์ •์˜ํ•˜๋Š” ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, antecedent๋Š” ๋ช…์‚ฌ๊ตฌ๊ฐ€ ์„ค๋ช…ํ•˜๋Š” ๋ช…์‚ฌ๊ตฌ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ \(e_{2}\)์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์„ ํ•ฉ๋ฆฌ์ ์œผ๋กœ approximationํ•œ๋‹ค๋ฉด, EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)์„ ์ฆ๊ฐ€์‹œํ‚ค๋ฉด LLM์ด ์ƒ๋Œ€์  ๋นˆ๋„๊ฐ€ ๋ฌด์ž‘์œ„ ํ™•๋ฅ ๋ณด๋‹ค ๋†’์€ \(e_{2}^{(0)}\)์„ ์ƒ์„ฑํ•˜๋Š” ์ƒ๋Œ€ ๋นˆ๋„๊ฐ€ ์ฆ๊ฐ€ํ•˜๋Š” ์ธต \(l\)์ด ์ ์–ด๋„ ๋ช‡ ๊ฐœ ์ด์ƒ ์กด์žฌํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์‹คํ—˜์šฐ๋ฆฌ๋Š” ๊ฐ„์ ‘์ ์ธ ์ฆ๊ฑฐ๋ฅผ ํ†ตํ•ด EntRec์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์„ ํ•ฉ๋ฆฌ์ ์œผ๋กœ ๊ทผ์‚ฌํ™”ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)๋Š” \(\tau_{\text{2H}}\)์˜ ๋งˆ์ง€๋ง‰ ํ† ํฐ์ด ์•„๋‹ˆ๋ผ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์„ค๋ช…์  ์–ธ๊ธ‰์˜ ๋งˆ์ง€๋ง‰ ํ† ํฐ์—์„œ ๊ณ„์‚ฐ๋˜๋ฉฐ, ๋ฌธ๋ฒ•์ ์œผ๋กœ ์‰ผํ‘œ ๋’ค์— \(e_{2}\)์˜ ์ด๋ฆ„์„ ์•ž์— ๋ถ™์ด๋Š” ๊ฒƒ์ด ์ž์—ฐ์Šค๋Ÿฝ์Šต๋‹ˆ๋‹ค(์˜ˆ: "'superstition'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ, _์Šคํ‹ฐ๋น„ ์›๋”_"). ๊ฒฐ๊ณผ ๋ฌธ์ž์—ด์—์„œ ๋ฌธ๋ฒ•์ ์œผ๋กœ \(\mu(r_{1}(e_{1}))\))๋Š” _์ „์น˜์‚ฌ_๊ฐ€ ๋˜๊ณ  \(e_{2}\)๋Š” _๋ช…์‚ฌ์ ˆ_์ด ๋˜๋ฉฐ, ๋ช…์‚ฌ์ ˆ์€ ๋‹ค๋ฅธ ๋ช…์‚ฌ์ ˆ ๋’ค์— ๋ฐ˜๋Œ€๋˜๋Š” ๋ช…์‚ฌ์ ˆ์ด ๋˜์–ด ๊ทธ ๋ช…์‚ฌ์ ˆ์„ ๋” ์‹๋ณ„ํ•˜๊ฑฐ๋‚˜ ์ •์˜ํ•˜๋Š” ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์„ ํ–‰์‚ฌ๋Š” ๋ช…์‚ฌ์ ˆ์ด ์„ค๋ช…ํ•˜๋Š” ๋ช…์‚ฌ์ ˆ์ด ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ \(e_{2}\)์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์„ ํ•ฉ๋ฆฌ์ ์œผ๋กœ ๊ทผ์‚ฌํ™”ํ•œ๋‹ค๋ฉด, EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\)์„ ์ฆ๊ฐ€์‹œํ‚ค๋ฉด LLM์ด ์ƒ๋Œ€์  ๋นˆ๋„๊ฐ€ ๋ฌด์ž‘์œ„ ํ™•๋ฅ ๋ณด๋‹ค ๋†’์€ \(e_{2}^{(0)}\)์„ ์ƒ์„ฑํ•˜๋Š” ์ƒ๋Œ€ ๋นˆ๋„๊ฐ€ ์ฆ๊ฐ€ํ•˜๋Š” ์ธต \(l\)์ด ์ ์–ด๋„ ๋ช‡ ๊ฐœ ์ด์ƒ ์กด์žฌํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]ExperimentWe demonstrate that EntRec is a reasonable approximation of the internal recall of the bridge entity with indirect evidence. Note that EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\) is calculated not at the last token of \(\tau_{\text{2H}}\) but at the last token of the bridge entity's descriptive mention, where it is grammatically natural to prepend a comma followed by the name of \(e_{2}\) (e.g., "The mother of the singer of 'Superstition', _Stevie Wonder_"). In the resulting string, grammatically \(\mu(r_{1}(e_{1}))\)) becomes the _antecedent_ and \(e_{2}\) becomes the _appositive_; an appositive is a noun phrase that follows another noun phrase in opposition to it and provides information that further identifies or defines it, and the antecedent is the noun phrase that the appositive describes. Then, if EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\) reasonably approximates the internal recall of the bridge entity \(e_{2}\), it is expected that _there will be at least some layers \(l\) where increasing EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\) increases the relative frequency of the LLM to generate \(e_{2}^{(0)}\) with a relative frequency higher than random chance_. [|assistant|]
์ผ๋ฐ˜ ์ค‘๊ตญ์–ด ๋ง๋ญ‰์น˜์˜ 10์–ต ํ† ํฐ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ Llama-2-7b-chat์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•ด ๊ฐ„๋‹จํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต ์ค‘ Learning Rate๋Š” 3e-5๋กœ ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€๋˜์—ˆ์œผ๋ฉฐ, ์–ด๋Œ‘ํ„ฐ ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•ด์„œ๋Š” ์ถ”๊ฐ€ Learning Rate 3e-4๋ฅผ ์‹คํ—˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก B์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ผ๋ฐ˜ ์ค‘๊ตญ์–ด ๋ง๋ญ‰์น˜์˜ 10์–ต ํ† ํฐ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ Llama-2-7b-chat์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•ด ๊ฐ„๋‹จํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต ์ค‘ ํ•™์Šต ์†๋„๋Š” 3e-5๋กœ ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€๋˜์—ˆ์œผ๋ฉฐ, ์–ด๋Œ‘ํ„ฐ ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•ด์„œ๋Š” ์ถ”๊ฐ€ ํ•™์Šต ์†๋„ 3e-4๋ฅผ ์‹คํ—˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋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|]We employed straightforward approaches for continual pre-training on Llama-2-7b-chat, utilizing the 1 billion tokens of data from general Traditional Chinese corpus. The learning rate during continual pre-training remained constant at 3e-5, and we experimented with an additional learning rate 3e-4 for the adapter approaches. More details can be found in Appendix B. [|assistant|]
LM1B ๋ฐ Wiki-40B์˜ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋ฅผ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, NearDup-deduplicated C4์—์„œ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ์ง€์†์ ์œผ๋กœ ๊ฐ€์žฅ ๋‚ฎ์€ Perplexity๋ฅผ ์–ป๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(๊ธฐ๋ณธ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ LM1B ํ‰๊ฐ€์˜ ๊ฒฝ์šฐ, ๋ถ€๋ก ๊ทธ๋ฆผ 7 ์ฐธ์กฐ).[|endofturn|]
LM1B ๋ฐ Wiki-40B์˜ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋ฅผ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, NearDup-deduplicated C4์—์„œ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ์ง€์†์ ์œผ๋กœ ๊ฐ€์žฅ ๋‚ฎ์€ ๋‹นํ˜น๊ฐ์„ ์–ป๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(๊ธฐ๋ณธ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ LM1B ํ‰๊ฐ€์˜ ๊ฒฝ์šฐ, ๋ถ€๋ก ๊ทธ๋ฆผ 7 ์ฐธ์กฐ).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]When evaluating on the validation sets of LM1B and Wiki-40B, we found that models trained on NearDup-deduplicated C4 consistently achieved lowest perplexity (for LM1B eval with base models, see Appendix Figure 7). [|assistant|]
์ œํ•œ๋œ ๋ฆฌ์†Œ์Šค๋ฅผ ์‚ฌ์šฉํ•œ bootstrapping. ์ตœ๊ทผ ์ผ๋ จ์˜ ์—ฐ๊ตฌ์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ํŠน์ˆ˜ํ•œ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์ผ๋ถ€ ์ถ”๋ก ์„ bootstrappingํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. NPPPrompt Zhao ๋“ฑ(2022)์€ Fine-tuning ์—†์ด ์‹œ๋งจํ‹ฑ ๋ ˆ์ด๋ธ”์— ๋Œ€ํ•œ ์˜ˆ์ธก์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ ์ž์ฒด์˜ ์ž„๋ฒ ๋”ฉ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ์˜ ๋ ˆ์ด๋ธ”๊ณผ ๊ด€๋ จ๋œ ๋‹จ์–ด๋ฅผ ์ž๋™์œผ๋กœ ์ฐพ์•„ ๋ชจ๋ธ ์˜ˆ์ธก๊ณผ ๋ ˆ์ด๋ธ”(verbalizers) ๊ฐ„์˜ ์ˆ˜๋™ ๋งคํ•‘์— ๋Œ€ํ•œ ์˜์กด๋„๋ฅผ ์ค„์—ฌ์ค๋‹ˆ๋‹ค. STAR Zelikman ๋“ฑ(2022)์€ ์†Œ์ˆ˜์˜ ๊ทผ๊ฑฐ ์˜ˆ์ œ์™€ ๊ทผ๊ฑฐ๊ฐ€ ์—†๋Š” ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ํ™œ์šฉํ•˜์—ฌ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ bootstrappingํ•ฉ๋‹ˆ๋‹ค. Self-Correction Welleck ๋“ฑ(2023)์€ ๋ถˆ์™„์ „ํ•œ ๊ธฐ๋ณธ ์ƒ์„ฑ๊ธฐ(๋ชจ๋ธ)์™€ ๋ถˆ์™„์ „ํ•œ ์ƒ์„ฑ์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ˆ˜์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•˜๋Š” ๋ณ„๋„์˜ ์ˆ˜์ •์ž๋ฅผ ๋ถ„๋ฆฌํ•˜์—ฌ ๊ธฐ๋ณธ ์ƒ์„ฑ๊ธฐ๋ณด๋‹ค ๊ฐœ์„ ๋œ ์„ฑ๋Šฅ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ๋Œ€์‹  Instruction ํŒจ๋Ÿฌ๋‹ค์ž„์—์„œ ์ƒˆ๋กœ์šด ์ž‘์—…์„ bootstrappingํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค.[|endofturn|]
์ œํ•œ๋œ ๋ฆฌ์†Œ์Šค๋ฅผ ์‚ฌ์šฉํ•œ ๋ถ€ํŠธ์ŠคํŠธ๋žฉํ•‘.์ตœ๊ทผ ์ผ๋ จ์˜ ์—ฐ๊ตฌ์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ํŠน์ˆ˜ํ•œ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์ผ๋ถ€ ์ถ”๋ก ์„ ๋ถ€ํŠธ์ŠคํŠธ๋žฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. NPPPrompt Zhao ๋“ฑ(2022)์€ ๋ฏธ์„ธ ์กฐ์ • ์—†์ด ์‹œ๋งจํ‹ฑ ๋ ˆ์ด๋ธ”์— ๋Œ€ํ•œ ์˜ˆ์ธก์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ ์ž์ฒด์˜ ์ž„๋ฒ ๋”ฉ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ์˜ ๋ ˆ์ด๋ธ”๊ณผ ๊ด€๋ จ๋œ ๋‹จ์–ด๋ฅผ ์ž๋™์œผ๋กœ ์ฐพ์•„ ๋ชจ๋ธ ์˜ˆ์ธก๊ณผ ๋ ˆ์ด๋ธ”(์–ธ์–ดํ™”) ๊ฐ„์˜ ์ˆ˜๋™ ๋งคํ•‘์— ๋Œ€ํ•œ ์˜์กด๋„๋ฅผ ์ค„์—ฌ์ค๋‹ˆ๋‹ค. STAR Zelikman ๋“ฑ(2022)์€ ์†Œ์ˆ˜์˜ ๊ทผ๊ฑฐ ์˜ˆ์ œ์™€ ๊ทผ๊ฑฐ๊ฐ€ ์—†๋Š” ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ํ™œ์šฉํ•˜์—ฌ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ๋ถ€ํŠธ์ŠคํŠธ๋žฉํ•ฉ๋‹ˆ๋‹ค. ์ž์ฒด ์ˆ˜์ • Welleck ๋“ฑ(2023)์€ ๋ถˆ์™„์ „ํ•œ ๊ธฐ๋ณธ ์ƒ์„ฑ๊ธฐ(๋ชจ๋ธ)์™€ ๋ถˆ์™„์ „ํ•œ ์ƒ์„ฑ์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ˆ˜์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•˜๋Š” ๋ณ„๋„์˜ ์ˆ˜์ •์ž๋ฅผ ๋ถ„๋ฆฌํ•˜์—ฌ ๊ธฐ๋ณธ ์ƒ์„ฑ๊ธฐ๋ณด๋‹ค ๊ฐœ์„ ๋œ ์„ฑ๋Šฅ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ๋Œ€์‹  ๋ช…๋ น ํŒจ๋Ÿฌ๋‹ค์ž„์—์„œ ์ƒˆ๋กœ์šด ์ž‘์—…์„ ๋ถ€ํŠธ์ŠคํŠธ๋žฉํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Bootstrapping with limited resources.A series of recent works use language models to bootstrap some inferences using specialized methods. NPPPrompt Zhao et al. (2022) provides a method to generate predictions for semantic labels without any finetuning. It uses a model's own embeddings to automatically find words relevant to the label of the data sample and hence reduces the dependency on manual mapping from model prediction to label (verbalizers). STAR Zelikman et al. (2022) iteratively leverages a small number of rationale examples and a large dataset without rationales, to bootstrap a model's ability to perform reasoning. Self-Correction Welleck et al. (2023) decouples an imperfect base generator (model) from a separate corrector that learns to iteratively correct imperfect generations and demonstrates improvement over the base generator. Our work instead focuses on bootstrapping new tasks in the instruction paradigm. [|assistant|]
๋‹ค๋ฅธ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์˜ ๊ฒฝ์šฐ, 100 ๋…ธ์ถœ ์„ค์ •์—์„œ Mistral, GPT2\({}_{0}\), GPT2\({}_{1/4}\)์˜ scaling laws๋Š” ๊ทธ๋ฆผ 4์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 4(c)๋Š” ๊ฒŒ์ดํŠธํ˜• MLP๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— Mistral ์•„ํ‚คํ…์ฒ˜๋„ GPT2๋ณด๋‹ค ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 4(d)๋Š” GPT2\({}_{1/4}\)์˜ MLP ๋ ˆ์ด์–ด ํฌ๊ธฐ๋ฅผ 4๋ถ„์˜ 1๋กœ ์ค„์ด๋ฉด ๋ชจ๋ธ ์šฉ๋Ÿ‰์— _๋ฌด์‹œํ•  ์ˆ˜ ์žˆ๋Š” ์˜ํ–ฅ_์„ ๋ฏธ์นœ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ GPT2\({}_{0}\)์—์„œ MLP ๋ ˆ์ด์–ด๋ฅผ ์™„์ „ํžˆ ์ œ๊ฑฐํ•˜๋ฉด ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํฌ๊ฒŒ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋‹ค๋ฅธ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์˜ ๊ฒฝ์šฐ, 100 ๋…ธ์ถœ ์„ค์ •์—์„œ Mistral, GPT2\({}_{0}\), GPT2\({}_{1/4}\)์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์€ ๊ทธ๋ฆผ 4์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 4(c)๋Š” ๊ฒŒ์ดํŠธํ˜• MLP๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— Mistral ์•„ํ‚คํ…์ฒ˜๋„ GPT2๋ณด๋‹ค ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 4(d)๋Š” GPT2\({}_{1/4}\)์˜ MLP ๋ ˆ์ด์–ด ํฌ๊ธฐ๋ฅผ 4๋ถ„์˜ 1๋กœ ์ค„์ด๋ฉด ๋ชจ๋ธ ์šฉ๋Ÿ‰์— _๋ฌด์‹œํ•  ์ˆ˜ ์žˆ๋Š” ์˜ํ–ฅ_์„ ๋ฏธ์นœ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ GPT2\({}_{0}\)์—์„œ MLP ๋ ˆ์ด์–ด๋ฅผ ์™„์ „ํžˆ ์ œ๊ฑฐํ•˜๋ฉด ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํฌ๊ฒŒ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]For other model architectures, Mistral, GPT2\({}_{0}\), and GPT2\({}_{1/4}\), their scaling laws in the 100-exposure setting are presented in Figure 4. Figure 4(c) confirms that the Mistral architecture also underperforms GPT2 due to its use of gated MLP. Figure 4(d) reveals that reducing GPT2\({}_{1/4}\)'s MLP layer size by a quarter has a _negligible impact_ on model capacity. However, removing the MLP layers entirely in GPT2\({}_{0}\) significantly reduces the model's capacity, see Figure 4(e). [|assistant|]
๋ชจ๋ธ ํ’๋™ ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋ชจ๋ธ configuration์€ ํ‘œ 8์— ๋‚˜์™€ ์žˆ๋‹ค. ๋ชจ๋ธ ๊นŠ์ด ๋Œ€๋น„ ๋ชจ๋ธ ํญ์ธ ๋ชจ๋ธ์˜ "๋ชจ์–‘"์€ ์ž ์žฌ์ ์ธ ์„ฑ๋Šฅ ๋ณ€ํ™”๋ฅผ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ€๋Šฅํ•œ ํ•œ ์œ ์‚ฌํ•˜๊ฒŒ ์œ ์ง€๋œ๋‹ค.[|endofturn|]
๋ชจ๋ธ ํ’๋™ ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋ชจ๋ธ ๊ตฌ์„ฑ์€ ํ‘œ 8์— ๋‚˜์™€ ์žˆ๋‹ค. ๋ชจ๋ธ ๊นŠ์ด ๋Œ€๋น„ ๋ชจ๋ธ ํญ์ธ ๋ชจ๋ธ์˜ "๋ชจ์–‘"์€ ์ž ์žฌ์ ์ธ ์„ฑ๋Šฅ ๋ณ€ํ™”๋ฅผ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ€๋Šฅํ•œ ํ•œ ์œ ์‚ฌํ•˜๊ฒŒ ์œ ์ง€๋œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We list the model configuration used in the model wind tunnel experiments in Table 8. The "shape" of the model, i.e., model width compared to model depth is kept as similar as possible to avoid any potential performance variation. [|assistant|]
์ธ๊ฐ„๊ณผ ์†Œํ†ตํ•˜๋Š” ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์œผ๋กœ ์ธํ•ด ChatGPT๋Š” ์ถœ์‹œ ์ดํ›„ AI ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ํฅ๋ฏธ๋ฅผ ๋ถˆ๋Ÿฌ์ผ์œผ์ผฐ์Šต๋‹ˆ๋‹ค. ChatGPT๋Š” ํŠน๋ณ„ํžˆ ์ตœ์ ํ™”๋œ ๋Œ€ํ™” ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ˜ ๊ฐ•๋ ฅํ•œ GPT ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ChatGPT์™€ GPT ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ์ ์  ๋” ์ปค์ง€๊ณ  ์žˆ๋Š” ์ ์„ ๊ณ ๋ คํ•˜์—ฌ, ์ง€๋‚œ ๋ช‡ ๋…„๊ฐ„ GPT ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์˜ ๊ธฐ์ˆ ์  ์ง„ํ™”์— ๋Œ€ํ•œ ํŠน๋ณ„ํ•œ ๋…ผ์˜๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ ๊ทธ ๋ฐœ์ „ ๊ณผ์ •์„ ๊ฐ„๋žตํ•˜๊ฒŒ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, ๊ทธ๋ฆผ 4์— GPT ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์˜ ๊ธฐ์ˆ ์  ์ง„ํ™”๋ฅผ ๋ณด์—ฌ์ฃผ๋Š” ๋„์‹๋„๋ฅผ ๊ทธ๋ ธ์Šต๋‹ˆ๋‹ค. GPT ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ์›๋ฆฌ๋Š” ์–ธ์–ด ๋ชจ๋ธ๋ง์„ ํ†ตํ•ด ์„ธ๊ณ„ ์ง€์‹์„ decoder-only Transformer ๋ชจ๋ธ๋กœ ์••์ถ•ํ•˜์—ฌ ์„ธ๊ณ„ ์ง€์‹์˜ ์˜๋ฏธ๋ฅผ ๋ณต๊ตฌ(๋˜๋Š” ๊ธฐ์–ต)ํ•˜๊ณ  ๋ฒ”์šฉ ํƒœ์Šคํฌ ์†”๋ฒ„๋กœ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์„ฑ๊ณต์˜ ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ์š”์†Œ๋Š” (I) _๋‹ค์Œ ๋‹จ์–ด๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋Š”_ decoder-only Transformer ์–ธ์–ด ๋ชจ๋ธ๊ณผ (II) _์–ธ์–ด ๋ชจ๋ธ์˜ ํฌ๊ธฐ ํ™•์žฅ_์ž…๋‹ˆ๋‹ค. ์ „์ฒด์ ์œผ๋กœ OpenAI์˜ LLM ์—ฐ๊ตฌ๋Š” ํฌ๊ฒŒ ๋‹ค์Œ ๋‹จ๊ณ„๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค13.[|endofturn|]
์ธ๊ฐ„๊ณผ ์†Œํ†ตํ•˜๋Š” ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์œผ๋กœ ์ธํ•ด ChatGPT๋Š” ์ถœ์‹œ ์ดํ›„ AI ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ํฅ๋ฏธ๋ฅผ ๋ถˆ๋Ÿฌ์ผ์œผ์ผฐ์Šต๋‹ˆ๋‹ค. ChatGPT๋Š” ํŠน๋ณ„ํžˆ ์ตœ์ ํ™”๋œ ๋Œ€ํ™” ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ˜ ๊ฐ•๋ ฅํ•œ GPT ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ChatGPT์™€ GPT ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ์ ์  ๋” ์ปค์ง€๊ณ  ์žˆ๋Š” ์ ์„ ๊ณ ๋ คํ•˜์—ฌ, ์ง€๋‚œ ๋ช‡ ๋…„๊ฐ„ GPT ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์˜ ๊ธฐ์ˆ ์  ์ง„ํ™”์— ๋Œ€ํ•œ ํŠน๋ณ„ํ•œ ๋…ผ์˜๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ ๊ทธ ๋ฐœ์ „ ๊ณผ์ •์„ ๊ฐ„๋žตํ•˜๊ฒŒ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, ๊ทธ๋ฆผ 4์— GPT ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์˜ ๊ธฐ์ˆ ์  ์ง„ํ™”๋ฅผ ๋ณด์—ฌ์ฃผ๋Š” ๋„์‹๋„๋ฅผ ๊ทธ๋ ธ์Šต๋‹ˆ๋‹ค. GPT ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ์›๋ฆฌ๋Š” ์–ธ์–ด ๋ชจ๋ธ๋ง์„ ํ†ตํ•ด ์„ธ๊ณ„ ์ง€์‹์„ ๋””์ฝ”๋” ์ „์šฉ ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ๋กœ ์••์ถ•ํ•˜์—ฌ ์„ธ๊ณ„ ์ง€์‹์˜ ์˜๋ฏธ๋ฅผ ๋ณต๊ตฌ(๋˜๋Š” ๊ธฐ์–ต)ํ•˜๊ณ  ๋ฒ”์šฉ ํƒœ์Šคํฌ ์†”๋ฒ„๋กœ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์„ฑ๊ณต์˜ ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ์š”์†Œ๋Š” (I) _๋‹ค์Œ ๋‹จ์–ด๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋Š”_ ๋””์ฝ”๋” ์ „์šฉ ํŠธ๋žœ์Šคํฌ๋จธ ์–ธ์–ด ๋ชจ๋ธ๊ณผ (II) _์–ธ์–ด ๋ชจ๋ธ์˜ ํฌ๊ธฐ ํ™•์žฅ_์ž…๋‹ˆ๋‹ค. ์ „์ฒด์ ์œผ๋กœ OpenAI์˜ LLM ์—ฐ๊ตฌ๋Š” ํฌ๊ฒŒ ๋‹ค์Œ ๋‹จ๊ณ„๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค13.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Due to the excellent capacity in communicating with humans, ChatGPT has ignited the excitement of the AI community since its release. ChatGPT is developed based on the powerful GPT model with specially optimized conversation capacities. Considering the ever-growing interest in ChatGPT and GPT models, we add a special discussion about the technical evolution of the GPT-series models, to briefly summarize the progress how they have been developed in the past years. Meanwhile, we drew a schematic diagram depicting the technological evolution of the GPT-series models in Figure 4. The basic principle underlying GPT models is to compress the world knowledge into the decoder-only Transformer model by language modeling, such that it can recover (or memorize) the semantics of world knowledge and serve as a general-purpose task solver. Two key points to the success are (I) training decoder-only Transformer language models that can _accurately predict the next word_ and (II) _scaling up the size of language models_. Overall, the research of OpenAI on LLMs can be roughly divided into the following stages13. [|assistant|]
์—ญ training์€ \(N\) ์ƒ˜ํ”Œ \(\{x_{1},\dots,x_{N}\}\)์ด ์žˆ๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ์ทจํ•˜๊ณ  ๋ฐ˜์ „๋œ ์ƒ˜ํ”Œ ์ง‘ํ•ฉ์„ ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์—ญ ํ›ˆ๋ จ์€ \(N\) ์ƒ˜ํ”Œ \(\{x_{1},\dots,x_{N}\}\)์ด ์žˆ๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ์ทจํ•˜๊ณ  ๋ฐ˜์ „๋œ ์ƒ˜ํ”Œ ์ง‘ํ•ฉ์„ ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Reverse training consists of taking a training dataset with \(N\) samples \(\{x_{1},\dots,x_{N}\}\) and constructing the set of reversed samples [|assistant|]
์œ„์˜ ๋ฐฉ๋ฒ•์—์„œ ์ด๋Ÿฌํ•œ ์ „๋žต์€ quantization ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๊ณต๋™์œผ๋กœ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณ ํšจ์œจ ๊ตฌํ˜„์„ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด quantization ๋ฐฉ๋ฒ•์€ ํ•˜๋“œ์›จ์–ด ๋˜๋Š” ์‹œ์Šคํ…œ ์ˆ˜์ค€ ์ง€์›(_์˜ˆ: ํšจ์œจ์ ์ธ GPU ์ปค๋„ ๋˜๋Š” ํ•˜๋“œ์›จ์–ด ์นœํ™”์ ์ธ ๊ทธ๋ฃน ํŒŒํ‹ฐ์…˜)์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์œ„์˜ ๋ฐฉ๋ฒ•์—์„œ ์ด๋Ÿฌํ•œ ์ „๋žต์€ ์–‘์žํ™” ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๊ณต๋™์œผ๋กœ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณ ํšจ์œจ ๊ตฌํ˜„์„ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์–‘์žํ™” ๋ฐฉ๋ฒ•์€ ํ•˜๋“œ์›จ์–ด ๋˜๋Š” ์‹œ์Šคํ…œ ์ˆ˜์ค€ ์ง€์›(_์˜ˆ: ํšจ์œจ์ ์ธ GPU ์ปค๋„ ๋˜๋Š” ํ•˜๋“œ์›จ์–ด ์นœํ™”์ ์ธ ๊ทธ๋ฃน ํŒŒํ‹ฐ์…˜)์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]These strategies in the above methods can be jointly used to improve the quantization performance. In order to achieve high-efficiency implementation, quantization methods also rely on hardware- or system-level support (_e.g.,_ efficient GPU kernels or hardware-friendly group partition). [|assistant|]
ํ‘œ 8์€ Instructions-2M ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ํฌํ•จ๋œ prompt ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์„ธ๋ถ€ ์ •๋ณด๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Prompts๋Š” ์‚ฌ์šฉ์ž prompt์™€ ์„ ํƒ์  ์‹œ์Šคํ…œ prompt์˜ ์—ฐ๊ฒฐ์ž…๋‹ˆ๋‹ค. T0 prompts๋Š” ํ‰๊ท  32.4๋‹จ์–ด๋กœ ๊ฐ€์žฅ ๊ธธ๋‹ค. Lamini prompts๋Š” ์ด 180๋งŒ ๊ฐœ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ๋Œ€๋ถ€๋ถ„์„ ์ฐจ์ง€ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
ํ‘œ 8์€ Instructions-2M ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ํฌํ•จ๋œ ํ”„๋กฌํ”„ํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์„ธ๋ถ€ ์ •๋ณด๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ๋Š” ์‚ฌ์šฉ์ž ํ”„๋กฌํ”„ํŠธ์™€ ์„ ํƒ์  ์‹œ์Šคํ…œ ํ”„๋กฌํ”„ํŠธ์˜ ์—ฐ๊ฒฐ์ž…๋‹ˆ๋‹ค. T0 ํ”„๋กฌํ”„ํŠธ๋Š” ํ‰๊ท  32.4๋‹จ์–ด๋กœ ๊ฐ€์žฅ ๊ธธ๋‹ค. ๋ผ๋ฏธ๋‹ˆ ํ”„๋กฌํ”„ํŠธ๋Š” ์ด 180๋งŒ ๊ฐœ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ๋Œ€๋ถ€๋ถ„์„ ์ฐจ์ง€ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Table 8 displays a breakdown of prompt datasets included in the Instructions-2M dataset. Prompts are the concatenation of the user prompt and an optional system prompt. T0 prompts are the longest, with an average of 32.4 words. Lamini prompts make up the majority of the training data, with a total of 1.8M included. [|assistant|]
๊ทธ๋ฆผ 6์— ์ž์„ธํžˆ ์„ค๋ช…๋œ ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด, ์˜ค๋ผํด ๋ฌธ์„œ๋งŒ์œผ๋กœ Fine-tuningํ•˜๋ฉด ๋” ๋งŽ์€ ์ˆ˜์˜ ์‚ฐ๋งŒ ๋ฌธ์„œ๋ฅผ ํฌํ•จํ•˜๋Š” ๊ตฌ์„ฑ์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ์ž์—ฐ ์งˆ๋ฌธ์˜ ๊ฒฝ์šฐ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์€ \(D^{*}+3D\)๋กœ ํ›ˆ๋ จํ•˜๊ณ , ํ•ซํŒŸ QA๋Š” \(D^{*}+1D\) ๋ฌธ์„œ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์ธ์‚ฌ์ดํŠธ๋Š” ํŠนํžˆ ์šฐ๋ฆฌ์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ RAFT์— ํฐ ๋„์›€์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์˜ค๋ผํด ๋ฌธ์„œ ํ•˜๋‚˜์™€ ์‚ฐ๋งŒ ๋ฌธ์„œ 4๊ฐœ๋กœ ๊ตฌ์„ฑ๋œ ํ›ˆ๋ จ ์„ค์ •์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋ชจ๋ธ์ด ์‚ฐ๋งŒ ๋ฌธ์„œ์— ์••๋„๋˜์ง€ ์•Š์œผ๋ฉด์„œ๋„ ๊ด€๋ จ ์ •๋ณด๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๊ตฌ๋ณ„ํ•˜๊ณ  ์šฐ์„ ์ˆœ์œ„๋ฅผ ์ง€์ •ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ–์ถœ ์ˆ˜ ์žˆ๋„๋ก ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 6์— ์ž์„ธํžˆ ์„ค๋ช…๋œ ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด, ์˜ค๋ผํด ๋ฌธ์„œ๋งŒ์œผ๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋ฉด ๋” ๋งŽ์€ ์ˆ˜์˜ ์‚ฐ๋งŒ ๋ฌธ์„œ๋ฅผ ํฌํ•จํ•˜๋Š” ๊ตฌ์„ฑ์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ์ž์—ฐ ์งˆ๋ฌธ์˜ ๊ฒฝ์šฐ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์€ \(D^{*}+3D\)๋กœ ํ›ˆ๋ จํ•˜๊ณ , ํ•ซํŒŸ QA๋Š” \(D^{*}+1D\) ๋ฌธ์„œ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์ธ์‚ฌ์ดํŠธ๋Š” ํŠนํžˆ ์šฐ๋ฆฌ์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ RAFT์— ํฐ ๋„์›€์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์˜ค๋ผํด ๋ฌธ์„œ ํ•˜๋‚˜์™€ ์‚ฐ๋งŒ ๋ฌธ์„œ 4๊ฐœ๋กœ ๊ตฌ์„ฑ๋œ ํ›ˆ๋ จ ์„ค์ •์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋ชจ๋ธ์ด ์‚ฐ๋งŒ ๋ฌธ์„œ์— ์••๋„๋˜์ง€ ์•Š์œผ๋ฉด์„œ๋„ ๊ด€๋ จ ์ •๋ณด๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๊ตฌ๋ณ„ํ•˜๊ณ  ์šฐ์„ ์ˆœ์œ„๋ฅผ ์ง€์ •ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ–์ถœ ์ˆ˜ ์žˆ๋„๋ก ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Our findings, detailed in Fig. 6, reveal that finetuning with only the oracle document frequently results in inferior performance compared to configurations that include a greater number of distractor documents. As we can see in the figure, the better performance for Natural Questions is training with \(D^{*}+3D\) and it is \(D^{*}+1D\) documents with Hotpot QA. This insight has been particularly beneficial for our algorithm, RAFT. In our experiments, we typically employ a training setup consisting of one oracle document alongside four distractor documents. This approach strikes a balance, ensuring the model is not overwhelmed by distractors while still gaining the ability to effectively discern and prioritize relevant information. [|assistant|]
**์ถœ๋ ฅ** [[Retriev=Yes] <p>์ฐธ์กฐ [1]<p>I [IsRel] =Relevant] ์„ธ๋ฅด๊ฒŒ์ด ๋ธ”๋ผ๋””๋ฏธ๋กœ๋น„์น˜ ๋ณด๋“œ๋กœํ”„(1948๋…„ 6์›” 28์ผ์ƒ)๋Š” ๋Ÿฌ์‹œ์•„ ์˜ํ™” ๊ฐ๋…, ์‹œ๋‚˜๋ฆฌ์˜ค ์ž‘๊ฐ€, ํ”„๋กœ๋“€์„œ์ž…๋‹ˆ๋‹ค.[[IsNew] =Fully Supported[[Retriev=Continue]๊ทธ๋Š” ๋Ÿฌ์‹œ์•„ ๋ชจ์Šคํฌ๋ฐ”์—์„œ ํƒœ์–ด๋‚ฌ์Šต๋‹ˆ๋‹ค.[[Retriev=Continue]๊ทธ๋Š” ๋Ÿฌ์‹œ์•„ ์˜ํ™” ์ œ์ž‘์ž ๊ธธ๋“œ์˜ ํšŒ์žฅ์ž…๋‹ˆ๋‹ค.[IsNew] =5][|endofturn|]
**์ถœ๋ ฅ** [[Retriev=Yes] <p>์ฐธ์กฐ [1]<p>I [IsRel] =๊ด€๋ จ] ์„ธ๋ฅด๊ฒŒ์ด ๋ธ”๋ผ๋””๋ฏธ๋กœ๋น„์น˜ ๋ณด๋“œ๋กœํ”„(1948๋…„ 6์›” 28์ผ์ƒ)๋Š” ๋Ÿฌ์‹œ์•„ ์˜ํ™” ๊ฐ๋…, ์‹œ๋‚˜๋ฆฌ์˜ค ์ž‘๊ฐ€, ํ”„๋กœ๋“€์„œ์ž…๋‹ˆ๋‹ค.[[IsNew] =Fully Supported[[Retriev=Continue]๊ทธ๋Š” ๋Ÿฌ์‹œ์•„ ๋ชจ์Šคํฌ๋ฐ”์—์„œ ํƒœ์–ด๋‚ฌ์Šต๋‹ˆ๋‹ค.[[Retriev=Continue]๊ทธ๋Š” ๋Ÿฌ์‹œ์•„ ์˜ํ™” ์ œ์ž‘์ž ๊ธธ๋“œ์˜ ํšŒ์žฅ์ž…๋‹ˆ๋‹ค.[IsNew] =5][|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Output** [[Retriev=Yes] <p>Reference [1]<p>I [IsRel] =Relevant] Sergei Vladimirovich Bodrov (born June 28, 1948) is a Russian film director, screenwriter, and producer.[[IsNew] =Fully Supported[[Retriev=Continue]He was born in Moscow, Russia.[[Retriev=Continue]He is the president of the Russian Guild of Film Producers.[IsNew] =5] [|assistant|]
๊ทธ๋ฆผ 4: \(N=\{1,4,16,64,128\}\)์— ๋Œ€ํ•œ ์ตœ๊ณ ์˜ \(N\) baseline. ์•ฝ 64-128๊ฐœ์˜ ์ƒ˜ํ”Œ ์ดํ›„ ์„ฑ๋Šฅ์ด ์ •์ฒด๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 4: N=\{1,4,16,64,128\}\)์— ๋Œ€ํ•œ ์ตœ๊ณ ์˜ \(N\) ๊ธฐ์ค€์„ . ์•ฝ 64-128๊ฐœ์˜ ์ƒ˜ํ”Œ ์ดํ›„ ์„ฑ๋Šฅ์ด ์ •์ฒด๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 4: Best of \(N\) baseline for \(N=\{1,4,16,64,128\}\). Performance plateaus after roughly 64-128 samples. [|assistant|]
์•ž์„œ ์–ธ๊ธ‰ํ•œ ์—ฐ๊ตฌ๋“ค์€ ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ๊ณผ Fine-tuning ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ด๋ฅผ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์—ฐ์‡„ ์ถ”๋ก  ์˜์—ญ์—์„œ์˜ ์ดˆ๊ธฐ ๋…ธ๋ ฅ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” In-Context Learning๊ณผ ๊ฒฐํ•ฉ๋œ ๋น„๋””์˜ค ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์ด ํ–ฅํ›„ ์—ฐ๊ตฌ์˜ ์ดˆ์ ์ด ๋˜์–ด์•ผ ํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ์œผ๋กœ ๋น„๋””์˜ค๋Š” ์ด๋ฏธ์ง€์— ๋น„ํ•ด ํƒ€๊ณ ๋‚œ ์—ฐ์‡„ ๊ด€๊ณ„๋ฅผ ๊ฐ€์ง„ ์ถ”๊ฐ€์ ์ธ ์‹œ๊ฐ„์  ์ •๋ณด๋ฅผ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. ์—ฐ์‡„ ์ถ”๋ก ์„ ํ†ตํ•ด ์—ฌ๋Ÿฌ ํ”„๋ ˆ์ž„์˜ ์ •๋ณด๋ฅผ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ์—ฐ๊ฒฐํ•˜์—ฌ ์‹œ๊ฐ„์  ๊ด€๊ณ„๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ๋ชจ๋ธ๋งํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋น„๋””์˜ค ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์— ๋งค์šฐ ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ํ•œํŽธ์œผ๋กœ, ์†Œํ˜• ๋ชจ๋ธ์€ ์šฉ๋Ÿ‰์ด ์ œํ•œ๋˜์–ด ์žˆ์–ด ์—ฐ์‡„ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๊ธฐ ์œ„ํ•ด Fine-tuning์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋”์šฑ์ด ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก  ์ฒด์ธ์„ ํš๋“ํ•˜๊ธฐ ์–ด๋ ค์›Œ ๋ฌธ์ œ๊ฐ€ ๋”์šฑ ์•…ํ™”๋ฉ๋‹ˆ๋‹ค. ์ด์— ๋น„ํ•ด ์ตœ์‹  ๋น„์ „ ์–ธ์–ด ๊ธฐ๋ฐ˜ ๋ชจ๋ธ(VLM) Alayrac et al. (2022); Li et al. (2023); Wang et al. (2022); Huang et al. (2023); Peng et al. (2023); Yu et al. (2021)์€ ๊ฐ•๋ ฅํ•œ ์‹œ๊ฐ ์–ธ์–ด ์ดํ•ด๋ ฅ์„ ๊ฐ–์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ํ…์ŠคํŠธ์™€ ์ด๋ฏธ์ง€๊ฐ€ ์ธํ„ฐ๋ฆฌ๋ธŒ๋œ In-Context Learning์ด ์ด๋ฏธ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ In-Context Learning์„ ํ†ตํ•œ ์—ฐ์‡„ ์ถ”๋ก ์„ ์œ„ํ•œ ํƒ„ํƒ„ํ•œ ๊ธฐ๋ฐ˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋น„๋””์˜ค ์ถ”๋ก ์— ์—ฐ์‡„ ์ถ”๋ก ์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์€ ์•„์ง ๋ช‡๋ช‡ ์—ฐ๊ตฌ๋งŒ์ด ์žˆ๋Š” ๋ฏธ๊ฐœ์ฒ™ ์˜์—ญ์ž…๋‹ˆ๋‹ค. CoMT(Hu et al., 2023)๋Š” ๋น„๋””์˜ค ์ถ”๋ก ์— ๋น ๋ฅธ ์‚ฌ๊ณ ์™€ ๋Š๋ฆฐ ์‚ฌ๊ณ ๋ฅผ ๊ฒฐํ•ฉํ•˜๊ณ  ๊ณ„ํš์„ ์œ„ํ•œ ํŠธ๋ฆฌ ๊ฒ€์ƒ‰ ์ „๋žต์„ ๋„์ž…ํ•˜์—ฌ ๋น„๋””์˜ค ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์— CoT๋ฅผ ์ตœ์ดˆ๋กœ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์•ž์„œ ์–ธ๊ธ‰ํ•œ ์—ฐ๊ตฌ๋“ค์€ ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ๊ณผ ๋ฏธ์„ธ ์กฐ์ • ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ด๋ฅผ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์—ฐ์‡„ ์ถ”๋ก  ์˜์—ญ์—์„œ์˜ ์ดˆ๊ธฐ ๋…ธ๋ ฅ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ธ kontext ํ•™์Šต๊ณผ ๊ฒฐํ•ฉ๋œ ๋น„๋””์˜ค ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์ด ํ–ฅํ›„ ์—ฐ๊ตฌ์˜ ์ดˆ์ ์ด ๋˜์–ด์•ผ ํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ์œผ๋กœ ๋น„๋””์˜ค๋Š” ์ด๋ฏธ์ง€์— ๋น„ํ•ด ํƒ€๊ณ ๋‚œ ์—ฐ์‡„ ๊ด€๊ณ„๋ฅผ ๊ฐ€์ง„ ์ถ”๊ฐ€์ ์ธ ์‹œ๊ฐ„์  ์ •๋ณด๋ฅผ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. ์—ฐ์‡„ ์ถ”๋ก ์„ ํ†ตํ•ด ์—ฌ๋Ÿฌ ํ”„๋ ˆ์ž„์˜ ์ •๋ณด๋ฅผ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ์—ฐ๊ฒฐํ•˜์—ฌ ์‹œ๊ฐ„์  ๊ด€๊ณ„๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ๋ชจ๋ธ๋งํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋น„๋””์˜ค ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์— ๋งค์šฐ ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ํ•œํŽธ์œผ๋กœ, ์†Œํ˜• ๋ชจ๋ธ์€ ์šฉ๋Ÿ‰์ด ์ œํ•œ๋˜์–ด ์žˆ์–ด ์—ฐ์‡„ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๊ธฐ ์œ„ํ•ด ๋ฏธ์„ธ ์กฐ์ •์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋”์šฑ์ด ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก  ์ฒด์ธ์„ ํš๋“ํ•˜๊ธฐ ์–ด๋ ค์›Œ ๋ฌธ์ œ๊ฐ€ ๋”์šฑ ์•…ํ™”๋ฉ๋‹ˆ๋‹ค. ์ด์— ๋น„ํ•ด ์ตœ์‹  ๋น„์ „ ์–ธ์–ด ๊ธฐ๋ฐ˜ ๋ชจ๋ธ(VLM) Alayrac et al. (2022); Li et al. (2023); Wang et al. (2022); Huang et al. (2023); Peng et al. (2023); Yu et al. (2021)์€ ๊ฐ•๋ ฅํ•œ ์‹œ๊ฐ ์–ธ์–ด ์ดํ•ด๋ ฅ์„ ๊ฐ–์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ํ…์ŠคํŠธ์™€ ์ด๋ฏธ์ง€๊ฐ€ ์ธํ„ฐ๋ฆฌ๋ธŒ๋œ ์ธ ์ปจํ…์ŠคํŠธ ํ•™์Šต์ด ์ด๋ฏธ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ธ ์ปจํ…์ŠคํŠธ ํ•™์Šต์„ ํ†ตํ•œ ์—ฐ์‡„ ์ถ”๋ก ์„ ์œ„ํ•œ ํƒ„ํƒ„ํ•œ ๊ธฐ๋ฐ˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋น„๋””์˜ค ์ถ”๋ก ์— ์—ฐ์‡„ ์ถ”๋ก ์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์€ ์•„์ง ๋ช‡๋ช‡ ์—ฐ๊ตฌ๋งŒ์ด ์žˆ๋Š” ๋ฏธ๊ฐœ์ฒ™ ์˜์—ญ์ž…๋‹ˆ๋‹ค. CoMT(Hu et al., 2023)๋Š” ๋น„๋””์˜ค ์ถ”๋ก ์— ๋น ๋ฅธ ์‚ฌ๊ณ ์™€ ๋Š๋ฆฐ ์‚ฌ๊ณ ๋ฅผ ๊ฒฐํ•ฉํ•˜๊ณ  ๊ณ„ํš์„ ์œ„ํ•œ ํŠธ๋ฆฌ ๊ฒ€์ƒ‰ ์ „๋žต์„ ๋„์ž…ํ•˜์—ฌ ๋น„๋””์˜ค ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์— CoT๋ฅผ ์ตœ์ดˆ๋กœ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The aforementioned studies explore multi-modal reasoning in small models and fine-tuning scenarios, which we regard as an initial endeavor in the realm of multi-modal chain-of-thought reasoning. We believe that video multi-modal reasoning combined with in-context learning should be the focus of future research. On the one hand, videos introduce additional temporal information with innate chaining relationships compared with images. Through chain-of-thought reasoning, the information in different frames can be naturally connected to explicitly model the temporal relationship, which is well-suited for video multi-modal reasoning. On the other hand, small models are capacity-limited and need fine-tuning to gain chain-of-thought ability. Worse still, multi-modal reasoning chains are difficult to obtain, which further exacerbates the challenge. In comparison, contemporary vision-language foundation models (VLMs) Alayrac et al. (2022); Li et al. (2023); Wang et al. (2022); Huang et al. (2023); Peng et al. (2023); Yu et al. (2021) have strong vision-language comprehension and are already capable of in-context learning with interleaved text and images. They provide a solid foundation for chain-of-thought reasoning with in-context learning. Utilizing chain-of-thought for video reasoning remains an unexplored territory with only a few studies. CoMT (Hu et al., 2023) combines fast-thinking and slow-thinking in video reasoning and introduces a tree search strategy for planning, which firstly applies CoT in video multi-modal reasoning. [|assistant|]