chosen
stringlengths
14
4.26k
rejected
stringlengths
53
9.3k
prompt
stringlengths
178
4.55k
์ง€๋กฑ ์›ฌ์€ Gaoling School of Artificial Intelligence์˜ ์ด์žฅ ๊ฒธ ๊ต์ˆ˜์ด์ž Renmin University of China์˜ School of Information ํ•™์žฅ์„ ๋งก๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Š” ์ˆ˜๋…„๊ฐ„ ๋น…๋ฐ์ดํ„ฐ์™€ AI ๋ถ„์•ผ์—์„œ ์ผํ•ด ์™”์œผ๋ฉฐ, ์œ ๋ช… ๊ตญ์ œ ์ปจํผ๋Ÿฐ์Šค์™€ ์ €๋„์— ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ๋…ผ๋ฌธ์„ ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ง€๋กฑ ์›ฌ์€ ๋ž„๋ง ์ธ๊ณต์ง€๋Šฅ๋Œ€ํ•™์˜ ์ด์žฅ ๊ฒธ ๊ต์ˆ˜์ด์ž ์ธ๋ฏผ๋Œ€ํ•™๊ต ์ •๋ณด๋Œ€ํ•™ ํ•™์žฅ์„ ๋งก๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Š” ์ˆ˜๋…„๊ฐ„ ๋น…๋ฐ์ดํ„ฐ์™€ AI ๋ถ„์•ผ์—์„œ ์ผํ•ด ์™”์œผ๋ฉฐ, ์œ ๋ช… ๊ตญ์ œ ์ปจํผ๋Ÿฐ์Šค์™€ ์ €๋„์— ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ๋…ผ๋ฌธ์„ ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Ji-Rong Wen is a full professor, the executive dean of Gaoling School of Artificial Intelligence, and the dean of School of Information at Renmin University of China. He has been working in the big data and AI areas for many years, and publishing extensively on prestigious international conferences and journals. [|assistant|]
ํ˜„์‹ค ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” LLaMA2์™€ ๊ฐ™์€ ๋„๊ตฌ๋กœ (\(\mathsf{bioR}\)์—์„œ์™€ ๊ฐ™์ด) ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ๋‹ค์‹œ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๋ฅผ 40๋ฒˆ ๋‹ค์‹œ ์ž‘์„ฑํ•˜๋ฉด ๋•Œ๋กœ๋Š” (๋‹ค๋ฅธ) hallucinated ๋‚ด์šฉ์„ ๊ฐ€์ง„ 40๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ์˜์–ด ๋‹จ๋ฝ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ๋ชจ๋ธ์„ 40๋ฐฐ ๋” ํฌ๊ฒŒ ๋งŒ๋“ค์–ด์•ผ ํ•˜๋‚˜์š”? ์•„๋‹ˆ์š”, \(\mathsf{bioS}\)์™€ \(\mathsf{bioR}\)๋ฅผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ, ๋™์ผํ•œ ๊ธฐ๊ฐ„ ๋™์•ˆ ํ›ˆ๋ จํ•  ๊ฒฝ์šฐ(๊ฐ 25ํšŒ์”ฉ 40ํšŒ ์žฌ์ž‘์„ฑ), ๋ชจ๋ธ์˜ ์šฉ๋Ÿ‰ ๋น„์œจ์€ ๊ฑฐ์˜ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€๋˜๋ฉฐ, LLaMA2์— ์˜ํ•ด ๋„์ž…๋œ ๊ด€๋ จ ์—†๋Š” ๋ฐ์ดํ„ฐ๋กœ ์ธํ•ด ์•ฝ๊ฐ„ ๋‚ฎ์•„์ง€๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.[|endofturn|]
ํ˜„์‹ค ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” LLaMA2์™€ ๊ฐ™์€ ๋„๊ตฌ๋กœ (\(\mathsf{bioR}\)์—์„œ์™€ ๊ฐ™์ด) ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ๋‹ค์‹œ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๋ฅผ 40๋ฒˆ ๋‹ค์‹œ ์ž‘์„ฑํ•˜๋ฉด ๋•Œ๋กœ๋Š” (๋‹ค๋ฅธ) ํ™˜์ƒ์ ์ธ ๋‚ด์šฉ์„ ๊ฐ€์ง„ 40๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ์˜์–ด ๋‹จ๋ฝ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ๋ชจ๋ธ์„ 40๋ฐฐ ๋” ํฌ๊ฒŒ ๋งŒ๋“ค์–ด์•ผ ํ•˜๋‚˜์š”? ์•„๋‹ˆ์š”, \(\mathsf{bioS}\)์™€ \(\mathsf{bioR}\)๋ฅผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ, ๋™์ผํ•œ ๊ธฐ๊ฐ„ ๋™์•ˆ ํ›ˆ๋ จํ•  ๊ฒฝ์šฐ(๊ฐ 25ํšŒ์”ฉ 40ํšŒ ์žฌ์ž‘์„ฑ), ๋ชจ๋ธ์˜ ์šฉ๋Ÿ‰ ๋น„์œจ์€ ๊ฑฐ์˜ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€๋˜๋ฉฐ, LLaMA2์— ์˜ํ•ด ๋„์ž…๋œ ๊ด€๋ จ ์—†๋Š” ๋ฐ์ดํ„ฐ๋กœ ์ธํ•ด ์•ฝ๊ฐ„ ๋‚ฎ์•„์ง€๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In a realistic scenario, tools like LLaMA2 can rewrite pretrain data like we did in \(\mathsf{bioR}\). Rewriting data 40 times can produce 40 distinct English paragraphs, sometimes with (different) hallucinated contents. Does this require the model to be 40x larger? No, our comparison between \(\mathsf{bioS}\) and \(\mathsf{bioR}\) shows that, if trained for the same duration (40 rewrites each for 25 passes), the model's capacity ratio remains nearly the same, slightly lower due to irrelevant data introduced by LLaMA2. [|assistant|]
์˜๋ฌด๋ก ์  ์œค๋ฆฌ ์ด๋ก  - ๊ฒฐ๊ณผ๋‚˜ ์˜๋„์— ๊ด€๊ณ„์—†์ด ํŠน์ • ํ–‰๋™์ด ๋ณธ์งˆ์ ์œผ๋กœ ์˜ณ๊ฑฐ๋‚˜ ๊ทธ๋ฅธ๋‹ค๊ณ  ์ฃผ์žฅํ•˜๋Š” ์ด๋ก .[|endofturn|]
1. ์˜๋ฌด๋ก ์  ์œค๋ฆฌ ์ด๋ก  - ๊ฒฐ๊ณผ๋‚˜ ์˜๋„์— ๊ด€๊ณ„์—†์ด ํŠน์ • ํ–‰๋™์ด ๋ณธ์งˆ์ ์œผ๋กœ ์˜ณ๊ฑฐ๋‚˜ ๊ทธ๋ฅธ๋‹ค๊ณ  ์ฃผ์žฅํ•˜๋Š” ์ด๋ก .[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]1. Deontological ethical theories - which hold that certain actions are inherently right or wrong, regardless of the consequences or intentions. [|assistant|]
๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋“œ๋Š” ๋น„์šฉ์€ ์—ฌ์ „ํžˆ ์ƒ๋Œ€์ ์œผ๋กœ ๋†’์ง€๋งŒ, ์ด ๋น„์šฉ ๋ถ„์„์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋‘ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ฐœ์„  ์‚ฌํ•ญ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋ฌธ๊ตฌ ๋ฐ”๊พธ๊ธฐ๋ฅผ ์œ„ํ•ด Qwen-1.8B ๋ชจ๋ธ Bai ๋“ฑ(2023b)์„ ์‚ฌ์šฉํ•˜๋ฉด token ์ฒ˜๋ฆฌ๋Ÿ‰์ด 3๋ฐฐ ๋” ๋†’์•„์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 5์˜ ์˜ˆ๋น„ ๊ฒฐ๊ณผ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, Qwen ๋ชจ๋ธ์— ์˜ํ•ด ์ƒ์„ฑ๋œ ์˜์—ญ์œผ๋กœ pre-trained๋œ ๋ชจ๋ธ์€ Mistral ๋ชจ๋ธ์— ์˜ํ•œ ์˜์—ญ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด generation ๋น„์šฉ์ด 3๋ฐฐ ์ ˆ๊ฐ๋ฉ๋‹ˆ๋‹ค. speculative decoding(Liu et al., 2023c)๊ณผ optimized inference(Xia et al., 2024)์— ๋Œ€ํ•œ ์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด generation ๋น„์šฉ์„ 3~5๋ฐฐ ๋” ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹ค์ œ๋กœ 1.3B parameter ๋ชจ๋ธ ํ›ˆ๋ จ ๊ทœ๋ชจ์—์„œ๋„ ์‹ค์ œ ๋ฐ์ดํ„ฐ๋งŒ ์‚ฌ์šฉํ•˜์—ฌ pre-training ๋น„์šฉ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋“œ๋Š” ๋น„์šฉ์€ ์—ฌ์ „ํžˆ ์ƒ๋Œ€์ ์œผ๋กœ ๋†’์ง€๋งŒ, ์ด ๋น„์šฉ ๋ถ„์„์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋‘ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ฐœ์„  ์‚ฌํ•ญ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋ฌธ๊ตฌ ๋ฐ”๊พธ๊ธฐ๋ฅผ ์œ„ํ•ด Qwen-1.8B ๋ชจ๋ธ Bai ๋“ฑ(2023b)์„ ์‚ฌ์šฉํ•˜๋ฉด ํ† ํฐ ์ฒ˜๋ฆฌ๋Ÿ‰์ด 3๋ฐฐ ๋” ๋†’์•„์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 5์˜ ์˜ˆ๋น„ ๊ฒฐ๊ณผ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, Qwen ๋ชจ๋ธ์— ์˜ํ•ด ์ƒ์„ฑ๋œ ์˜์—ญ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์€ Mistral ๋ชจ๋ธ์— ์˜ํ•œ ์˜์—ญ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ƒ์„ฑ ๋น„์šฉ์ด 3๋ฐฐ ์ ˆ๊ฐ๋ฉ๋‹ˆ๋‹ค. ์ถ”์ธก ๋””์ฝ”๋”ฉ(Liu et al., 2023c)๊ณผ ์ตœ์ ํ™”๋œ ์ถ”๋ก (Xia et al., 2024)์— ๋Œ€ํ•œ ์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์ƒ์„ฑ ๋น„์šฉ์„ 3~5๋ฐฐ ๋” ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹ค์ œ๋กœ 1.3B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ ํ›ˆ๋ จ ๊ทœ๋ชจ์—์„œ๋„ ์‹ค์ œ ๋ฐ์ดํ„ฐ๋งŒ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ›ˆ๋ จ ๋น„์šฉ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]While the cost of generating high quality data is still relatively high, two important sources of improvement impact this cost analysis. First, if we use the Qwen-1.8B model Bai et al. (2023b) for rephrasing, we are able to get a 3x higher token throughput. As seen in our preliminary results in Fig 5, the model pre-trained on rephrases generated by Qwen model performs comparably to that by the Mistral model. This reduces the cost of generation by 3x. More recent work in speculative decoding (Liu et al., 2023c) and optimized inference (Xia et al., 2024) suggest that we can leverage another 3-5x improvement in the generation cost. Hence, indeed, even at the scale of just 1.3B parameter model training, we can already improve upon the cost of pre-training using just real data. [|assistant|]
๋‹ค์–‘ํ•œ Instruction ๊ณผ์ œ๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์€ ์ธ๊ฐ„์˜ ์ง€์‹œ๋ฅผ ๋” ์ž˜ ์ดํ•ดํ•˜๊ณ  ๋”ฐ๋ฅด๋ฉฐ, ๋ณด์ด์ง€ ์•Š๋Š” ์ง€์‹œ๊นŒ์ง€ ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ๋‹ค.[|endofturn|]
๋‹ค์–‘ํ•œ ์ง€์‹œ ๊ณผ์ œ๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์€ ์ธ๊ฐ„์˜ ์ง€์‹œ๋ฅผ ๋” ์ž˜ ์ดํ•ดํ•˜๊ณ  ๋”ฐ๋ฅด๋ฉฐ, ๋ณด์ด์ง€ ์•Š๋Š” ์ง€์‹œ๊นŒ์ง€ ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]By tuning on diverse instruction tasks, the model is able to better comprehend and follow human instructions, and generalize to unseen instructions. [|assistant|]
\(\bullet\)_์ž‘์—… ์„ค๋ช…._ ์ž‘์—… ์„ค๋ช…์€ ์ผ๋ฐ˜์ ์œผ๋กœ LLMs์ด ๋”ฐ๋ผ์•ผ ํ•˜๋Š” ํŠน์ • instruction์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ž์—ฐ์–ด๋กœ ์ž‘์—… ๋ชฉํ‘œ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ์„ค๋ช…ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํŠน์ˆ˜ํ•œ ์ž…๋ ฅ ๋˜๋Š” ์ถœ๋ ฅ ํ˜•์‹์ด ์žˆ๋Š” ์ž‘์—…์˜ ๊ฒฝ์šฐ ์ž์„ธํ•œ ์„ค๋ช…์ด ํ•„์š”ํ•œ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ, ํ‚ค์›Œ๋“œ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํŠน์ˆ˜ ์„ค์ •์„ ๊ฐ•์กฐํ•˜์—ฌ LLMs์ด ์ž‘์—…์„ ๋” ์ž˜ ์™„๋ฃŒํ•  ์ˆ˜ ์žˆ๋„๋ก ์•ˆ๋‚ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_์ž‘์—… ์„ค๋ช…._ ์ž‘์—… ์„ค๋ช…์€ ์ผ๋ฐ˜์ ์œผ๋กœ LLM์ด ๋”ฐ๋ผ์•ผ ํ•˜๋Š” ํŠน์ • ์ง€์นจ์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ž์—ฐ์–ด๋กœ ์ž‘์—… ๋ชฉํ‘œ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ์„ค๋ช…ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํŠน์ˆ˜ํ•œ ์ž…๋ ฅ ๋˜๋Š” ์ถœ๋ ฅ ํ˜•์‹์ด ์žˆ๋Š” ์ž‘์—…์˜ ๊ฒฝ์šฐ ์ž์„ธํ•œ ์„ค๋ช…์ด ํ•„์š”ํ•œ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ, ํ‚ค์›Œ๋“œ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํŠน์ˆ˜ ์„ค์ •์„ ๊ฐ•์กฐํ•˜์—ฌ LLM์ด ์ž‘์—…์„ ๋” ์ž˜ ์™„๋ฃŒํ•  ์ˆ˜ ์žˆ๋„๋ก ์•ˆ๋‚ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_Task description._ A task description is typically a specific instruction that LLMs are expected to follow. In general, one should clearly describe the task goal in natural language. For the tasks with special input or output format, detailed clarifications are often needed, and one can further utilize keywords to highlight the special settings for better guiding LLMs in task completion. [|assistant|]
* **MMLU: ๊ณต์‹ MMLU** **(**Hendrycks et al.**,** 2021**)** **ํ‰๊ฐ€ ์Šคํฌ๋ฆฝํŠธ ๋ฐ ํ”„๋กฌํ”„ํŠธ๋ฅผ [https://github.com/hendrycks/test](https://github.com/hendrycks/test)์—์„œ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ผ๊ด„ ์ฒ˜๋ฆฌ๋ฅผ ํ—ˆ์šฉํ•˜๋„๋ก ์ˆ˜์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. MMLU์˜ ์›๋ž˜ ์„ค์ •์— ๋”ฐ๋ผ 0๊ฐœ์˜ few-shot ์˜ˆ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ ์˜ˆ์ œ์— ๋Œ€ํ•œ ํ‰๊ท  ์ •ํ™•๋„๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.* * **ToxiGen: Touvron et al.**์˜ ์„ค์ •์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. **(**2023b**)****๋ฅผ ๋”ฐ๋ฅด์ง€๋งŒ, ํŠน์ • ๊ทธ๋ฃน์— ๋Œ€ํ•œ ๋…์„ฑ ์ƒ์„ฑ์„ ์ด๋Œ์–ด๋‚ด๋„๋ก ์„ค๊ณ„๋œ** **Hartvigsen et al.**์˜ ์›๋ž˜ ํ”„๋กฌํ”„ํŠธ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. **(**2022**)****์˜ ์›๋ž˜ ํ”„๋กฌํ”„ํŠธ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋…์„ฑ ์–ธ์–ด๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ์„ค๊ณ„๋œ ํ”„๋กฌํ”„ํŠธ('ํ˜์˜ค๊ฐ ์žˆ๋Š”' ํ”„๋กฌํ”„ํŠธ)๋งŒ ์‚ฌ์šฉํ•˜๊ณ  ๊ทธ๋ฃน๋‹น 500๊ฐœ์˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€ ๋น„์šฉ์„ ์ค„์ž…๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ์›๋ž˜์˜ ToxiGen ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š๊ณ  ์ „๋‹ฌํ•˜๊ณ  ์ฒซ ๋ฒˆ์งธ ์ƒˆ ์ค„๊นŒ์ง€(๋˜๋Š” ์ตœ๋Œ€ 512 ํ† ํฐ) ํƒ์š•์Šค๋Ÿฝ๊ฒŒ ๋””์ฝ”๋”ฉํ•ฉ๋‹ˆ๋‹ค. Instruction-tuned ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ•ด๋‹น ํ…œํ”Œ๋ฆฟ์— ๋ฐฐ์น˜ํ•˜๊ณ  ๋ชจ๋ธ์ด ์ค‘์ง€ ํ† ํฐ(๋˜๋Š” ์ตœ๋Œ€ 512 ํ† ํฐ)์„ ์ƒ์„ฑํ•  ๋•Œ๊นŒ์ง€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์™„๋ฃŒํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ๋ฅผ **Hartvigsen et al.**์˜ ์ผ๋ถ€๋กœ์„œ ๋ฏธ์„ธ ์กฐ์ •๋œ ๋…์„ฑ ์ฝ˜ํ…์ธ  ๊ฐ์ง€์šฉ์œผ๋กœ ํ›ˆ๋ จ๋œ reborta-large ๋ชจ๋ธ์— ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค. **(**2022**)****28**. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ถ„๋ฅ˜๊ธฐ๊ฐ€ ๋…์„ฑ์ด ์žˆ๋‹ค๊ณ  ํŒ๋‹จํ•œ ์ƒ์„ฑ๋ฌผ์˜ ๋น„์œจ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ์ฃผ 28: [https://huggingface.co/tomh/toxic_roberta](https://huggingface.co/tomh/toxic_roberta)*** * **TruthfulQA: Touvron et al.**์— ๋”ฐ๋ผ **(**2023b**)****๋ฅผ ์ฃผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ฃผ๋กœ TruthfulQA์˜ ์ƒ์„ฑ ์„ค์ •์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. **(**Lin et al.**,** 2022**)****. TruthfulQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” 818๊ฐœ์˜ ์งˆ๋ฌธ์ด ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด ์งˆ๋ฌธ์€ ํ…Œ์ŠคํŠธ ๋ชจ๋ธ์ด ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. 6๊ฐœ์˜ In-Context QA ์˜ˆ์ œ๊ฐ€ ํฌํ•จ๋œ ๊ธฐ๋ณธ QA ํ”„๋กฌํ”„ํŠธ ํ˜•์‹์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ณต์‹** **๊ตฌํ˜„29**์˜ ๊ณต์‹ ์Šคํฌ๋ฆฝํŠธ๋ฅผ ๋”ฐ๋ผ ํƒ์š•์Šค๋Ÿฌ์šด ๋””์ฝ”๋”ฉ๊ณผ ๋‹ต๋ณ€ ํ›„์ฒ˜๋ฆฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. GPT-3์˜ ์‚ฌ์šฉ ์ค‘๋‹จ์œผ๋กœ ์ธํ•ด ์›๋ž˜ TruthfulQA ํ‰๊ฐ€๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ๋ณต์ œํ•˜๋Š” ๊ฒƒ์ด ๋ถˆ๊ฐ€๋Šฅํ•ด์ง„ ๋งŒํผ ๋ชจ๋ธ ์‘๋‹ต์˜ ์ง„์‹ค์„ฑ๊ณผ ์œ ์ต์„ฑ์„ ํŒ๋‹จํ•˜๊ธฐ ์œ„ํ•ด LLaMA 2 ๊ธฐ๋ฐ˜ ๋ถ„๋ฅ˜๊ธฐ ๋‘ ๊ฐœ๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. LLaMA 2 ํŒ์‚ฌ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ **Lin et al.**์—์„œ ์‚ฌ์šฉํ•œ ์›๋ž˜ GPT-3 ๊ธฐ๋ฐ˜ ํŒ์‚ฌ์˜ ์„ฑ๋Šฅ๊ณผ ์ผ์น˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. **(**2022**)****. Touvron et al.**์— ๋”ฐ๋ผ ์‘๋‹ต์ด ์ง„์‹คํ•˜๊ณ  ์œ ์ตํ•œ ๋น„์œจ(% ์œ ์ตํ•˜๊ณ  ์ง„์‹คํ•œ)์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. **(**2023b**)****. ์ฃผ์š” ์ง€ํ‘œ๋กœ๋Š” % ์œ ์ตํ•˜๊ณ  ์ง„์‹คํ•œ ๊ฒƒ๋งŒ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ์ฃผ 29: [https://github.com/sylinrl/TruthfulQA/](https://github.com/sylinrl/TruthfulQA/)*** * **AlpacaEval: ํ‰๊ฐ€๋œ ๋ชจ๋ธ์ด 805๊ฐœ์˜ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๊ณ  GPT-4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‘๋‹ต์„ Davinci-003๊ณผ ๋น„๊ตํ•˜๋„๋ก ์š”์ฒญํ•˜๋Š” ๊ธฐ๋ณธ ์„ค์ •์— ๋”ฐ๋ผ** **Li et al.** **(**2023**)****์—์„œ ์ œ๊ณตํ•œ ํŒจํ‚ค์ง€๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. "alpaca_eval_gpt4" ์ฃผ์„์ž๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํŠน๋ณ„ํ•œ ์ค‘์ง€ ์‹œํ€€์Šค๋ฅผ ์ง€์ •ํ•˜์ง€ ์•Š๊ณ  ํ‰๊ฐ€ ๋Œ€์ƒ ๋ชจ๋ธ์ด ์ตœ๋Œ€ 2048 ํ† ํฐ์„ ์ƒ์„ฑํ•˜๋„๋ก ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
* **MMLU: ๊ณต์‹ MMLU** **(**Hendrycks et al.**,** 2021**)** **ํ‰๊ฐ€ ์Šคํฌ๋ฆฝํŠธ ๋ฐ ํ”„๋กฌํ”„ํŠธ๋ฅผ [https://github.com/hendrycks/test](https://github.com/hendrycks/test)์—์„œ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ผ๊ด„ ์ฒ˜๋ฆฌ๋ฅผ ํ—ˆ์šฉํ•˜๋„๋ก ์ˆ˜์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. MMLU์˜ ์›๋ž˜ ์„ค์ •์— ๋”ฐ๋ผ 0๊ฐœ์˜ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ ์˜ˆ์ œ์— ๋Œ€ํ•œ ํ‰๊ท  ์ •ํ™•๋„๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.* **ToxiGen: Touvron et al.**์˜ ์„ค์ •์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. **(**2023b**)****๋ฅผ ๋”ฐ๋ฅด์ง€๋งŒ, ํŠน์ • ๊ทธ๋ฃน์— ๋Œ€ํ•œ ๋…์„ฑ ์ƒ์„ฑ์„ ์ด๋Œ์–ด๋‚ด๋„๋ก ์„ค๊ณ„๋œ** **Hartvigsen et al.**์˜ ์›๋ž˜ ํ”„๋กฌํ”„ํŠธ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. **(**2022**)****์˜ ์›๋ž˜ ํ”„๋กฌํ”„ํŠธ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋…์„ฑ ์–ธ์–ด๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ์„ค๊ณ„๋œ ํ”„๋กฌํ”„ํŠธ('ํ˜์˜ค๊ฐ ์žˆ๋Š”' ํ”„๋กฌํ”„ํŠธ)๋งŒ ์‚ฌ์šฉํ•˜๊ณ  ๊ทธ๋ฃน๋‹น 500๊ฐœ์˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€ ๋น„์šฉ์„ ์ค„์ž…๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ์›๋ž˜์˜ ToxiGen ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š๊ณ  ์ „๋‹ฌํ•˜๊ณ  ์ฒซ ๋ฒˆ์งธ ์ƒˆ ์ค„๊นŒ์ง€(๋˜๋Š” ์ตœ๋Œ€ 512 ํ† ํฐ) ํƒ์š•์Šค๋Ÿฝ๊ฒŒ ๋””์ฝ”๋”ฉํ•ฉ๋‹ˆ๋‹ค. ์ง€์‹œ์–ด ํŠœ๋‹๋œ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ•ด๋‹น ํ…œํ”Œ๋ฆฟ์— ๋ฐฐ์น˜ํ•˜๊ณ  ๋ชจ๋ธ์ด ์ค‘์ง€ ํ† ํฐ(๋˜๋Š” ์ตœ๋Œ€ 512 ํ† ํฐ)์„ ์ƒ์„ฑํ•  ๋•Œ๊นŒ์ง€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์™„๋ฃŒํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ๋ฅผ **Hartvigsen et al.**์˜ ์ผ๋ถ€๋กœ์„œ ๋ฏธ์„ธ ์กฐ์ •๋œ ๋…์„ฑ ์ฝ˜ํ…์ธ  ๊ฐ์ง€์šฉ์œผ๋กœ ํ›ˆ๋ จ๋œ reborta-large ๋ชจ๋ธ์— ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค. **(**2022**)****28**. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ถ„๋ฅ˜๊ธฐ๊ฐ€ ๋…์„ฑ์ด ์žˆ๋‹ค๊ณ  ํŒ๋‹จํ•œ ์ƒ์„ฑ๋ฌผ์˜ ๋น„์œจ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ์ฃผ 28: [https://huggingface.co/tomh/toxic_roberta](https://huggingface.co/tomh/toxic_roberta)*** **TruthfulQA: Touvron et al.**์— ๋”ฐ๋ผ **(**2023b**)****๋ฅผ ์ฃผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ฃผ๋กœ TruthfulQA์˜ ์ƒ์„ฑ ์„ค์ •์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. **(**Lin et al.**,** 2022**)****. TruthfulQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” 818๊ฐœ์˜ ์งˆ๋ฌธ์ด ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด ์งˆ๋ฌธ์€ ํ…Œ์ŠคํŠธ ๋ชจ๋ธ์ด ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. 6๊ฐœ์˜ ์ปจํ…์ŠคํŠธ QA ์˜ˆ์ œ๊ฐ€ ํฌํ•จ๋œ ๊ธฐ๋ณธ QA ํ”„๋กฌํ”„ํŠธ ํ˜•์‹์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ณต์‹** **๊ตฌํ˜„29**์˜ ๊ณต์‹ ์Šคํฌ๋ฆฝํŠธ๋ฅผ ๋”ฐ๋ผ ํƒ์š•์Šค๋Ÿฌ์šด ๋””์ฝ”๋”ฉ๊ณผ ๋‹ต๋ณ€ ํ›„์ฒ˜๋ฆฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. GPT-3์˜ ์‚ฌ์šฉ ์ค‘๋‹จ์œผ๋กœ ์ธํ•ด ์›๋ž˜ TruthfulQA ํ‰๊ฐ€๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ๋ณต์ œํ•˜๋Š” ๊ฒƒ์ด ๋ถˆ๊ฐ€๋Šฅํ•ด์ง„ ๋งŒํผ ๋ชจ๋ธ ์‘๋‹ต์˜ ์ง„์‹ค์„ฑ๊ณผ ์œ ์ต์„ฑ์„ ํŒ๋‹จํ•˜๊ธฐ ์œ„ํ•ด LLaMA 2 ๊ธฐ๋ฐ˜ ๋ถ„๋ฅ˜๊ธฐ ๋‘ ๊ฐœ๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. LLaMA 2 ํŒ์‚ฌ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ **Lin et al.**์—์„œ ์‚ฌ์šฉํ•œ ์›๋ž˜ GPT-3 ๊ธฐ๋ฐ˜ ํŒ์‚ฌ์˜ ์„ฑ๋Šฅ๊ณผ ์ผ์น˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. **(**2022**)****. Touvron et al.**์— ๋”ฐ๋ผ ์‘๋‹ต์ด ์ง„์‹คํ•˜๊ณ  ์œ ์ตํ•œ ๋น„์œจ(% ์œ ์ตํ•˜๊ณ  ์ง„์‹คํ•œ)์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. **(**2023b**)****. ์ฃผ์š” ์ง€ํ‘œ๋กœ๋Š” % ์œ ์ตํ•˜๊ณ  ์ง„์‹คํ•œ ๊ฒƒ๋งŒ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ์ฃผ 29: [https://github.com/sylinrl/TruthfulQA/](https://github.com/sylinrl/TruthfulQA/)*** **AlpacaEval: ํ‰๊ฐ€๋œ ๋ชจ๋ธ์ด 805๊ฐœ์˜ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๊ณ  GPT-4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‘๋‹ต์„ Davinci-003๊ณผ ๋น„๊ตํ•˜๋„๋ก ์š”์ฒญํ•˜๋Š” ๊ธฐ๋ณธ ์„ค์ •์— ๋”ฐ๋ผ** **Li et al.** **(**2023**)****์—์„œ ์ œ๊ณตํ•œ ํŒจํ‚ค์ง€๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. "alpaca_eval_gpt4" ์ฃผ์„์ž๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํŠน๋ณ„ํ•œ ์ค‘์ง€ ์‹œํ€€์Šค๋ฅผ ์ง€์ •ํ•˜์ง€ ์•Š๊ณ  ํ‰๊ฐ€ ๋Œ€์ƒ ๋ชจ๋ธ์ด ์ตœ๋Œ€ 2048 ํ† ํฐ์„ ์ƒ์„ฑํ•˜๋„๋ก ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* **MMLU: We use the official MMLU** **(**Hendrycks et al.**,** 2021**)** **evaluation script and prompts available at [https://github.com/hendrycks/test](https://github.com/hendrycks/test), with modifications to allow for batch processing. We evaluate using 0 few-shot examples, following the original setup of MMLU. We report average accuracy across test examples.** * **ToxiGen: We follow the setup in** **Touvron et al.** **(**2023b**)****, but use the original set of prompts from** **Hartvigsen et al.** **(**2022**)****, which are designed to elicit toxic generations for certain groups. We take only the prompts designed to produce toxic language ('hateful' prompts) and use 500 prompts per group to reduce evaluation costs. For base language models, we pass in the original ToxiGen prompts unchanged and greedily decode up to the first new line (or a maximum of 512 tokens). For instruction-tuned models, we place the prompt in the corresponding template, and ask the model to complete the prompt, until the model generates a stop token (or a maximum of 512 tokens). We pass the generated text into a reborta-large model trained to detect toxic content finetuned as part of** **Hartvigsen et al.** **(**2022**)****28**. We then report the percentage of generations deemed toxic by the classifier.** Footnote 28: [https://huggingface.co/tomh/toxic_roberta](https://huggingface.co/tomh/toxic_roberta)** * **TruthfulQA: Following** **Touvron et al.** **(**2023b**)****, we mainly use the generation setting of TruthfulQA** **(**Lin et al.**,** 2022**)****. The TruthfulQA dataset contains 818 questions, which are used to prompt the tested model to generate answers. We use the default QA prompt format with 6 in-context QA examples. We follow the official script in their official** **implementation29** **to do greedy decoding and answer postprocessing. We train two LLaMA 2-based classifiers for judging the truthfulness and informativeness of the model response, due to the deprecation of GPT-3 making exact replication of the original TruthfulQA evaluation infeasible. We find that the LLaMA 2 judges are generally able to match the performance of the original GPT-3-based judges used by** **Lin et al.** **(**2022**)****. We report the rate of the responses being truthful and informative (% Informative and Truthful) following** **Touvron et al.** **(**2023b**)****. We only report the % Informative and Truthful as our primary metric.** Footnote 29: [https://github.com/sylinrl/TruthfulQA/](https://github.com/sylinrl/TruthfulQA/)** * **AlpacaEval: We use the package provided by** **Li et al.** **(**2023**)****, following the default setup which asks the evaluated model to generate responses for 805 prompts and employ GPT-4 to compare the response with Davinci-003. We employ the "alpaca_eval_gpt4" annotator. We allow the evaluated model to generate up to 2048 tokens, without specifying special stop sequences. [|assistant|]
* ๊ฐ ๋ฌธ์„œ์— ๋Œ€ํ•ด ์‚ฌ์šฉ์ž์˜ ์งˆ๋ฌธ๊ณผ ๊ทธ์— ํ•ด๋‹นํ•˜๋Š” agent์˜ ์‘๋‹ต์„ ์ฃผ์„์œผ๋กœ ๋‹ฌ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ™”๋‹น ํ‰๊ท  user-agent ํ„ด ์ˆ˜๋Š” 5ํšŒ ์ •๋„์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค. * ๊ฐ ์‚ฌ์šฉ์ž์˜ ์งˆ๋ฌธ์— ๋Œ€ํ•ด ๋ฌธ์„œ ๋‚ด์—์„œ ๊ด€๋ จ์„ฑ ์žˆ๋Š” ๋ชจ๋“  context๋ฅผ ์ฃผ์„์œผ๋กœ ๋‹ฌ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
* ๊ฐ ๋ฌธ์„œ์— ๋Œ€ํ•ด ์‚ฌ์šฉ์ž์˜ ์งˆ๋ฌธ๊ณผ ๊ทธ์— ํ•ด๋‹นํ•˜๋Š” ์ƒ๋‹ด์›์˜ ์‘๋‹ต์„ ์ฃผ์„์œผ๋กœ ๋‹ฌ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ™”๋‹น ํ‰๊ท  ์‚ฌ์šฉ์ž-์ƒ๋‹ด์› ํ„ด ์ˆ˜๋Š” 5ํšŒ ์ •๋„์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค. * ๊ฐ ์‚ฌ์šฉ์ž์˜ ์งˆ๋ฌธ์— ๋Œ€ํ•ด ๋ฌธ์„œ ๋‚ด์—์„œ ๊ด€๋ จ์„ฑ ์žˆ๋Š” ๋ชจ๋“  ์ปจํ…์ŠคํŠธ๋ฅผ ์ฃผ์„์œผ๋กœ ๋‹ฌ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* For each document, we need to annotate the user's questions and corresponding agent's responses. The average number of user-agent turns per conversation needs to be around five. * For each user's question, we need to annotate all the relevant context within the document. [|assistant|]
๊ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ ์šฉ๋œ ์„ธ๋ถ€ ํ‰๊ฐ€ ์„ค์ •๊ณผ metrics๋Š” ํ‘œ 1์— ์š”์•ฝ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์–ธ์–ด ๋ชจ๋ธ์˜ robustness์™€ ์œ ์—ฐ์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์ทจํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์กฐ์น˜๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
๊ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ ์šฉ๋œ ์„ธ๋ถ€ ํ‰๊ฐ€ ์„ค์ •๊ณผ ๋ฉ”ํŠธ๋ฆญ์€ ํ‘œ 1์— ์š”์•ฝ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฒฌ๊ณ ์„ฑ๊ณผ ์œ ์—ฐ์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์ทจํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์กฐ์น˜๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Our detailed evaluation setup and the metrics applied to each dataset are summarized in Table 1, showcasing the extensive measures taken to guarantee the robustness and flexibility of our language models across a wide range of tasks. [|assistant|]
๊ธฐ์—ฌ 1: ๋ฒ•๋ฅ  LLM์˜ ๊ฐ€์กฑ. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋ฒ•๋ฅ  ์˜์—ญ์—์„œ ์ง๋ฉดํ•˜๋Š” ๊ณ ์œ ํ•œ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์„ธ์‹ฌํ•˜๊ฒŒ ์ œ์ž‘๋œ ๋ฒ•๋ฅ  ์–ธ์–ด ๋ชจ๋ธ ๋ชจ์Œ์ธ SaulLM-7B ์ œํ’ˆ๊ตฐ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋ฒ•๋ฅ  ํ…์ŠคํŠธ์— ํŠนํ™”๋œ 70์–ต ํŒŒ๋ผ๋ฏธํ„ฐ ์–ธ์–ด ๋ชจ๋ธ์ธ SaulLM-7B๋ฅผ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค. SaulLM-7B๋Š” ํŠน์ˆ˜ ํ›ˆ๋ จ์„ ํ†ตํ•ด ์ผ๋ฐ˜ ๋ชจ๋ธ์— ๋น„ํ•ด ๋ฒ•๋ฅ  ์–ธ์–ด์˜ ๋‰˜์•™์Šค๋ฅผ ๋” ์ž˜ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋‹ค์–‘ํ•œ ๋ฒ•๋ฅ  ์ž‘์—…์—์„œ ๊ธฐ์กด ๋ชจ๋ธ(์˜ˆ: Mistral ๋˜๋Š” Llama)๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋„๋ก ์„ธ์‹ฌํ•˜๊ฒŒ ์„ค๊ณ„๋œ Instruction-tuned ๋ณ€ํ˜•์ธ SaulLM-7B-Instruct๋ฅผ ์ถœ์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ธฐ์—ฌ 1: ๋ฒ•๋ฅ  LLM์˜ ๊ฐ€์กฑ.์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋ฒ•๋ฅ  ์˜์—ญ์—์„œ ์ง๋ฉดํ•˜๋Š” ๊ณ ์œ ํ•œ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์„ธ์‹ฌํ•˜๊ฒŒ ์ œ์ž‘๋œ ๋ฒ•๋ฅ  ์–ธ์–ด ๋ชจ๋ธ ๋ชจ์Œ์ธ SaulLM-7B ์ œํ’ˆ๊ตฐ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋ฒ•๋ฅ  ํ…์ŠคํŠธ์— ํŠนํ™”๋œ 70์–ต ํŒŒ๋ผ๋ฏธํ„ฐ ์–ธ์–ด ๋ชจ๋ธ์ธ SaulLM-7B๋ฅผ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค. SaulLM-7B๋Š” ํŠน์ˆ˜ ํ›ˆ๋ จ์„ ํ†ตํ•ด ์ผ๋ฐ˜ ๋ชจ๋ธ์— ๋น„ํ•ด ๋ฒ•๋ฅ  ์–ธ์–ด์˜ ๋‰˜์•™์Šค๋ฅผ ๋” ์ž˜ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋‹ค์–‘ํ•œ ๋ฒ•๋ฅ  ์ž‘์—…์—์„œ ๊ธฐ์กด ๋ชจ๋ธ(์˜ˆ: Mistral ๋˜๋Š” Llama)๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋„๋ก ์„ธ์‹ฌํ•˜๊ฒŒ ์„ค๊ณ„๋œ ๋ช…๋ น์–ด ํŠœ๋‹๋œ ๋ณ€ํ˜•์ธ SaulLM-7B-Instruct๋ฅผ ์ถœ์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Contribution 1: A family of legal LLMs.In this paper, we introduce the SaulLM-7B's family, a collection of Legal Language Models meticulously crafted to tackle the distinctive challenges encountered within the legal domain. We unveil SaulLM-7B, a 7-billion-parameter language model specifically tailored to legal text. With its specialized training regimen, SaulLM-7B demonstrates a superior understanding of the nuances in legal language compared to generic models. Furthermore, we release SaulLM-7B-Instruct, an instructiontuned variant, carefully engineered to outperform existing models such as Mistral or Llama on a variety of legal tasks1. [|assistant|]
์ค‘๊ตญ์–ด, ์˜์–ด, ์ฝ”๋“œ, ํ•™์ˆ  ๋…ผ๋ฌธ ๋“ฑ Tokenizer์˜ ํ•™์Šต ์„ธํŠธ์— ์—†๋Š” 30๋งŒ ๊ฐœ์˜ ๋ฌธ์„œ์— ๋Œ€ํ•ด ํ‰๊ฐ€๋ฅผ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค. MiniCPM-120K Tokenizer๊ฐ€ ๊ฐ€์žฅ ๋†’์€ ์••์ถ•๋ฅ (๋ฐ”์ดํŠธ/ํ† ํฐ)์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ค‘๊ตญ์–ด, ์˜์–ด, ์ฝ”๋“œ, ํ•™์ˆ  ๋…ผ๋ฌธ ๋“ฑ ํ† ํฐ๋ผ์ด์ €์˜ ํ•™์Šต ์„ธํŠธ์— ์—†๋Š” 30๋งŒ ๊ฐœ์˜ ๋ฌธ์„œ์— ๋Œ€ํ•ด ํ‰๊ฐ€๋ฅผ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค. MiniCPM-120K ํ† ํฐ๋ผ์ด์ €๊ฐ€ ๊ฐ€์žฅ ๋†’์€ ์••์ถ•๋ฅ (๋ฐ”์ดํŠธ/ํ† ํฐ)์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We conduct evaluations on 300,000 documents in Chinese, English, code, and academic papers that is not in the training set of the Tokenizer. The MiniCPM-120K tokenizer achieves the highest compression ratio (Bytes/Tokens). [|assistant|]
REPLUG(Shi et al., 2023)์€ ๋ธ”๋ž™๋ฐ•์Šค LM์ด ์ตœ์ข… ์˜ˆ์ธก์„ ํ•˜๊ธฐ ์ „์— ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ๋ฅผ ์ž…๋ ฅ context์— ์•ž์— ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ๋ฅผ ๋ณ‘๋ ฌ๋กœ ์ธ์ฝ”๋”ฉํ•˜๋Š” ensemble ์ „๋žต์„ ๋„์ž…ํ•˜์—ฌ LM context ๊ธธ์ด์˜ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ณ  ์ฆ๊ฐ€๋œ ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค๋ฅผ ํ• ๋‹นํ•˜์—ฌ ์ •ํ™•๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ LM์ด ๋” ๊ด‘๋ฒ”์œ„ํ•œ ๊ด€๋ จ ์ •๋ณด์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•จ์œผ๋กœ์จ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
REPLUG(Shi et al., 2023)์€ ๋ธ”๋ž™๋ฐ•์Šค LM์ด ์ตœ์ข… ์˜ˆ์ธก์„ ํ•˜๊ธฐ ์ „์— ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ๋ฅผ ์ž…๋ ฅ ์ปจํ…์ŠคํŠธ์— ์•ž์— ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ๋ฅผ ๋ณ‘๋ ฌ๋กœ ์ธ์ฝ”๋”ฉํ•˜๋Š” ์•™์ƒ๋ธ” ์ „๋žต์„ ๋„์ž…ํ•˜์—ฌ LM ์ปจํ…์ŠคํŠธ ๊ธธ์ด์˜ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ณ  ์ฆ๊ฐ€๋œ ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค๋ฅผ ํ• ๋‹นํ•˜์—ฌ ์ •ํ™•๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ LM์ด ๋” ๊ด‘๋ฒ”์œ„ํ•œ ๊ด€๋ จ ์ •๋ณด์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•จ์œผ๋กœ์จ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]REPLUG (Shi et al., 2023) proposes a method for prepending retrieved documents to the input context before the final prediction by the black-box LM. It introduces an ensemble strategy to encode retrieved documents in parallel, overcoming the limitations of LM context length and enhancing accuracy through the allocation of increased computational resources. This approach improves the generation process by ensuring that the LM has access to a broader range of relevant information. [|assistant|]
์บ๋ฆญํ„ฐ Tokenization. Tokenization์€ ๋ฌธ์ž ์ˆ˜์ค€ ๋˜๋Š” UTF-8 ๋ฐ”์ดํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์‹œํ€€์Šค ๊ธธ์ด๊ฐ€ ๋Š˜์–ด๋‚˜์„œ Transformer ์•„ํ‚คํ…์ฒ˜์—์„œ ๊ณ„์‚ฐ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ฒŒ ๋˜๋Š”๋ฐ, ์ด๋Š” ์‹œํ€€์Šค ๊ธธ์ด์—์„œ self-attention ๊ณ„์ธต์˜ ์ด์ฐจ์  ๋ณต์žก์„ฑ์œผ๋กœ ์ธํ•ด ํ˜„์žฌ LLM์˜ ์ฃผ์š” ์•„ํ‚คํ…์ฒ˜๊ฐ€ ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Vaswani et al., 2017). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Gao et al., 2020; Tay et al., 2021; Xue et al., 2022; Clark et al., 2022; Yu et al., 2023).[|endofturn|]
์บ๋ฆญํ„ฐ ํ† ํฐํ™”.ํ† ํฐํ™”๋Š” ๋ฌธ์ž ์ˆ˜์ค€ ๋˜๋Š” UTF-8 ๋ฐ”์ดํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์‹œํ€€์Šค ๊ธธ์ด๊ฐ€ ๋Š˜์–ด๋‚˜์„œ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜์—์„œ ๊ณ„์‚ฐ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ฒŒ ๋˜๋Š”๋ฐ, ์ด๋Š” ์‹œํ€€์Šค ๊ธธ์ด์—์„œ ์ž๊ธฐ ์ฃผ์˜ ๊ณ„์ธต์˜ ์ด์ฐจ์  ๋ณต์žก์„ฑ์œผ๋กœ ์ธํ•ด ํ˜„์žฌ LLM์˜ ์ฃผ์š” ์•„ํ‚คํ…์ฒ˜๊ฐ€ ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Vaswani et al., 2017). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Gao et al., 2020; Tay et al., 2021; Xue et al., 2022; Clark et al., 2022; Yu et al., 2023).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Character Tokenization.Tokenization can also be performed on a character level or based on UTF-8 bytes. However, this results in an increased sequence length, which becomes computationally expensive in the transformer architecture, the current predominated architecture for LLMs due to the quadratic complexity of the self-attention layer in the sequence length (Vaswani et al., 2017). Though, several approaches have been proposed to address this limitation (Gao et al., 2020; Tay et al., 2021; Xue et al., 2022; Clark et al., 2022; Yu et al., 2023). [|assistant|]
์†์„ฑ 1: _๊ณ ์œ  ๋‚œ์ด๋„๋ฅผ ๋†’์—ฌ์„œ ์ด๋“์ด ์ฆ๊ฐ€ํ–ˆ๋‹ค๊ฐ€ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค._ \(S\)์™€ \(K\)์˜ ๊ฐ’์„ ๊ฐ๊ฐ ์ž‘์€ ๊ฐ’๋ถ€ํ„ฐ ํฐ ๊ฐ’๊นŒ์ง€ ๋„ค ๊ฐœ์˜ ๋‹ค๋ฅธ ๊ทธ๋ฃน์—์„œ ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€ํ•˜๋ฉด์„œ \(I\)๋ฅผ \(10\)์—์„œ \(400\)๊นŒ์ง€ ๋ณ€ํ™”์‹œ์ผœ ๊ณ ์œ  ๋‚œ์ด๋„๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 6(์™ผ์ชฝ)์€ \(I\)๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ์ดˆ๊ธฐ ์ƒ์Šนํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ ์ฃผ๋ฉฐ, ์ด๋Š” ๋ณธ ๋ฐฉ๋ฒ•์ด ๊ณ ์œ  ๋‚œ์ด๋„ ์ƒ์Šน์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๋ˆˆ์— ๋„๋Š” ์„ฑ๋Šฅ ํ–ฅ์ƒ์€ \(I=100\) ๋ฐ \(I=200\)์—์„œ ๋‚˜ํƒ€๋‚˜๋ฉฐ, ์ด๋Š” ๋ชจ๋“  \(S\) ๋ฐ \(K\) ์„ค์ •์—์„œ ์ผ๊ด€๋˜๊ฒŒ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ \(I=400\)์—์„œ๋Š” ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ๊ฐ์†Œํ•˜๋Š”๋ฐ, ์ด๋Š” ๊ณผ๋„ํ•œ ๋ณต์žก์„ฑ์ด ๋ชจ๋ธ์˜ reasoning ๋Šฅ๋ ฅ์„ ์ดˆ๊ณผํ•˜์—ฌ ๊ทนํ•œ์˜ ์ž‘์—… ๋‚œ์ด๋„์—์„œ ์ €ํฌ ๋ฐฉ๋ฒ•์˜ ์ˆ˜์ต๋ฅ ์ด ๊ฐ์†Œํ•  ์ˆ˜ ์žˆ์Œ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์†์„ฑ 1: _๊ณ ์œ  ๋‚œ์ด๋„๋ฅผ ๋†’์—ฌ์„œ ์ด๋“์ด ์ฆ๊ฐ€ํ–ˆ๋‹ค๊ฐ€ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค._ S$์™€ $K$์˜ ๊ฐ’์„ ๊ฐ๊ฐ ์ž‘์€ ๊ฐ’๋ถ€ํ„ฐ ํฐ ๊ฐ’๊นŒ์ง€ ๋„ค ๊ฐœ์˜ ๋‹ค๋ฅธ ๊ทธ๋ฃน์—์„œ ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€ํ•˜๋ฉด์„œ \(I\)๋ฅผ \(10\)์—์„œ \(400\)๊นŒ์ง€ ๋ณ€ํ™”์‹œ์ผœ ๊ณ ์œ  ๋‚œ์ด๋„๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 6(์™ผ์ชฝ)์€ \(I\)๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ์ดˆ๊ธฐ ์ƒ์Šนํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ ์ฃผ๋ฉฐ, ์ด๋Š” ๋ณธ ๋ฐฉ๋ฒ•์ด ๊ณ ์œ  ๋‚œ์ด๋„ ์ƒ์Šน์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๋ˆˆ์— ๋„๋Š” ์„ฑ๋Šฅ ํ–ฅ์ƒ์€ \(I=100\) ๋ฐ \(I=200\)์—์„œ ๋‚˜ํƒ€๋‚˜๋ฉฐ, ์ด๋Š” ๋ชจ๋“  \(S\) ๋ฐ \(K\) ์„ค์ •์—์„œ ์ผ๊ด€๋˜๊ฒŒ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ \(I=400\)์—์„œ๋Š” ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ๊ฐ์†Œํ•˜๋Š”๋ฐ, ์ด๋Š” ๊ณผ๋„ํ•œ ๋ณต์žก์„ฑ์ด ๋ชจ๋ธ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ์ดˆ๊ณผํ•˜์—ฌ ๊ทนํ•œ์˜ ์ž‘์—… ๋‚œ์ด๋„์—์„œ ์ €ํฌ ๋ฐฉ๋ฒ•์˜ ์ˆ˜์ต๋ฅ ์ด ๊ฐ์†Œํ•  ์ˆ˜ ์žˆ์Œ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Property 1: _Gains increase then decrease by rising the inherent difficulty._ We investigate the inherent difficulty by varying \(I\) from \(10\) to \(400\), while keeping the values of \(S\) and \(K\) constant across four groups of different values, from small to large, respectively. Figure 6 (left) shows an initial uptick in performance gains with increases in \(I\), indicating that our method can significantly enhance performance in line with rising inherent difficulty. The most notable gains are seen at \(I=100\) and \(I=200\), consistent across all \(S\) and \(K\) settings. Yet, at \(I=400\), gains taper off, implying that excessive complexity may exceed the model's reasoning capabilities, leading to diminishing returns for our method under extreme task difficulty. [|assistant|]
์ƒˆ๋กœ์šด ์ •์˜๋Š” emergent abilities๊ฐ€ ๋‚˜ํƒ€๋‚˜๋Š” ํ›ˆ๋ จ ๊ถค์  ๋‚ด์—์„œ ์ค‘์š”ํ•œ ์ „ํ™˜์ ์„ ์ •ํ™•ํ•˜๊ฒŒ ํŠน์ง•ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ •์˜๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ ์ด๋Ÿฌํ•œ ์ „ํ™˜์ ์—์„œ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ณ€ํ™”๋ฅผ ์กฐ์‚ฌํ•˜์—ฌ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์˜ ๋ฐœ์ „์„ ์ด‰์ง„ํ•˜๋„๋ก ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ƒˆ๋กœ์šด ์ •์˜๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์ด ๋‚˜ํƒ€๋‚˜๋Š” ํ›ˆ๋ จ ๊ถค์  ๋‚ด์—์„œ ์ค‘์š”ํ•œ ์ „ํ™˜์ ์„ ์ •ํ™•ํ•˜๊ฒŒ ํŠน์ง•ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ •์˜๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ ์ด๋Ÿฌํ•œ ์ „ํ™˜์ ์—์„œ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ณ€ํ™”๋ฅผ ์กฐ์‚ฌํ•˜์—ฌ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์˜ ๋ฐœ์ „์„ ์ด‰์ง„ํ•˜๋„๋ก ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The new definition offers a precise characterization of the critical junctures within training trajectories where emergent abilities manifest. It encourages future studies to investigate the shifts in language models at these junctures, which facilitate the development of new capabilities. [|assistant|]
๋ชจ๋“  SFT ๋ฐ์ดํ„ฐ์˜ ๊ตฌ์กฐ๋ฅผ ๋Œ€ํ™” ํ˜•์‹์œผ๋กœ ํ†ต์ผํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € "System" ์—ญํ• ์„ ์ฒ˜์Œ์— ์ถ”๊ฐ€ํ•˜์—ฌ LLM์ด ์˜ˆ์˜ ๋ฐ”๋ฅด๊ณ  ์œ ์šฉํ•œ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋„๋ก ์•ˆ๋‚ดํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ instruction์„ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ "User" ๋ฐ "Assistant" ์—ญํ• ์„ ์ถ”๊ฐ€ํ•˜์—ฌ instruction tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ instruction ๋ฐ ์‘๋‹ต ์Œ์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ด ํ†ตํ•ฉ๋œ ํ˜•์‹์„ ์‚ฌ์šฉํ•˜์—ฌ LLM foundation ๋ชจ๋ธ์— Fine-tuning์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋ชจ๋“  SFT ๋ฐ์ดํ„ฐ์˜ ๊ตฌ์กฐ๋ฅผ ๋Œ€ํ™” ํ˜•์‹์œผ๋กœ ํ†ต์ผํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € "์‹œ์Šคํ…œ" ์—ญํ• ์„ ์ฒ˜์Œ์— ์ถ”๊ฐ€ํ•˜์—ฌ LLM์ด ์˜ˆ์˜ ๋ฐ”๋ฅด๊ณ  ์œ ์šฉํ•œ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋„๋ก ์•ˆ๋‚ดํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ์ง€์นจ์„ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ "์‚ฌ์šฉ์ž" ๋ฐ "์–ด์‹œ์Šคํ„ดํŠธ" ์—ญํ• ์„ ์ถ”๊ฐ€ํ•˜์—ฌ ์ง€์นจ ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ง€์นจ ๋ฐ ์‘๋‹ต ์Œ์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ด ํ†ตํ•ฉ๋œ ํ˜•์‹์„ ์‚ฌ์šฉํ•˜์—ฌ LLM ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์— ๋ฏธ์„ธ ์กฐ์ •์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We unify the structure of all the SFT data in a conversational format. We first add a "System" role at the beginning to setup a general instruction guiding LLM to provide polite and helpful answers. We also add "User" and "Assistant" roles to incorporate instruction and response pairs from the instruction tuning datasets. We apply fine-tuning using this unified format on an LLM foundation model. [|assistant|]
* 1 ์†Œ๊ฐœ * 2 ์‚ฌ์ „ ํ•™์Šต * 2.1 ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ * 2.2 ํ† ํฐํ™” * 2.3 ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ * 3 Fine-tuning * 3.1 ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ * 3.2 ํ›ˆ๋ จ ๋ฐฉ๋ฒ• * 4 ์ธํ”„๋ผ * 5 ์•ˆ์ „ * 6 ํ‰๊ฐ€ * 6.1 ๊ธฐ๋ณธ ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.1.1 ์ฃผ์š” ๊ฒฐ๊ณผ * 6.1.2 ํ† ๋ก  * 6.1.3 In-Context Learning ์—ฐ๊ตฌ * 6.2 ์ฑ„ํŒ… ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.2.1 ์ž๋™ ํ‰๊ฐ€ * 6.2.2 ์ธ๊ฐ„ ํ‰๊ฐ€ * 7 ๊ธฐ๋Šฅ ํ™•์žฅ * 7.1 ๊ธด ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ๋ง * 7.2 ๋น„์ „-์–ธ์–ด * 7.3 ๊นŠ์ด ์—…์Šค์ผ€์ผ๋ง * 8 ์ตœ์ข… ํ† ๋ก  * A ์ €์ž ๋ชฉ๋ก ๋ฐ ๊ธฐ์—ฌ๋„[|endofturn|]
* 1 ์†Œ๊ฐœ * 2 ์‚ฌ์ „ ํ•™์Šต * 2.1 ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ * 2.2 ํ† ํฐํ™” * 2.3 ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ * 3 ๋ฏธ์„ธ ์กฐ์ • * 3.1 ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ * 3.2 ํ›ˆ๋ จ ๋ฐฉ๋ฒ• * 4 ์ธํ”„๋ผ * 5 ์•ˆ์ „ * 6 ํ‰๊ฐ€ * 6.1 ๊ธฐ๋ณธ ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.1.1 ์ฃผ์š” ๊ฒฐ๊ณผ * 6.1.2 ํ† ๋ก  * 6.1.3 ์ธ ์ปจํ…์ŠคํŠธ ํ•™์Šต ์—ฐ๊ตฌ * 6.2 ์ฑ„ํŒ… ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.2.1 ์ž๋™ ํ‰๊ฐ€ * 6.2.2 ์ธ๊ฐ„ ํ‰๊ฐ€ * 7 ๊ธฐ๋Šฅ ํ™•์žฅ * 7.1 ๊ธด ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ๋ง * 7.2 ๋น„์ „-์–ธ์–ด * 7.3 ๊นŠ์ด ์—…์Šค์ผ€์ผ๋ง * 8 ์ตœ์ข… ํ† ๋ก  * A ์ €์ž ๋ชฉ๋ก ๋ฐ ๊ธฐ์—ฌ๋„[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* 1 Introduction * 2 Pretraining * 2.1 Data Processing * 2.2 Tokenization * 2.3 Model Architecture * 3 Finetuning * 3.1 Data Preprocessing * 3.2 Training Method * 4 Infrastructure * 5 Safety * 6 Evaluations * 6.1 Base Model Performance * 6.1.1 Main Results * 6.1.2 Discussions * 6.1.3 In-Context Learning Study * 6.2 Chat Model Performance * 6.2.1 Automatic Evaluations * 6.2.2 Human Evaluations * 7 Capability Extension * 7.1 Long Context Modeling * 7.2 Vision-Language * 7.3 Depth Upscaling * 8 Final Discussions * A Author List and Contributions [|assistant|]
\end{table} ํ‘œ 2: **(a) 5๊ฐœ์˜ ์˜ˆ์ธก ํ”Œ๋žซํผ ๋ชจ๋‘์—์„œ train, validation ๋ฐ test ์„ธํŠธ์˜ ๋ถ„ํฌ.** ์ค‘์š”ํ•œ ๊ฒƒ์€ test ์„ธํŠธ์˜ ๋ชจ๋“  ๋ฌธ์ œ๊ฐ€ ๊ธฐ๋ณธ LMs์˜ ํ•™์Šต ๋งˆ๊ฐ์ผ์ธ 2023๋…„ 6์›” 1์ผ ์ดํ›„์— ์ถœ์ œ๋œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, train ๋ฐ validation ์„ธํŠธ์˜ ๋ชจ๋“  ๋ฌธ์ œ๋Š” 2023๋…„ 6์›” 1์ผ ์ด์ „์— ํ•ด๊ฒฐ๋˜์—ˆ์œผ๋ฏ€๋กœ tuning ํ”„๋กœ์„ธ์Šค์—์„œ ๋ˆ„์ถœ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. **(b) test ์„ธํŠธ์—์„œ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ **baseline ์„ฑ๋Šฅ**์ด๋ฉฐ, ํ‘œ์ค€ ์˜ค์ฐจ(SE) 1๊ฐœ(์ „์ฒด ๊ฒฐ๊ณผ๋Š” ํ‘œ 7 ์ฐธ์กฐ)์ž…๋‹ˆ๋‹ค. ๋ฌด์ž‘์œ„ baseline: 0.250; ์ธ๊ฐ„ ํฌ๋ผ์šฐ๋“œ: 0.149. ์ด ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์ด ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ์˜ˆ์ธก์— ๋Šฅ์ˆ™ํ•˜์ง€ ์•Š๋‹ค๋Š” ๊ฒƒ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
\end{table} ํ‘œ 2: **(a) 5๊ฐœ์˜ ์˜ˆ์ธก ํ”Œ๋žซํผ ๋ชจ๋‘์—์„œ ํŠธ๋ ˆ์ธ, ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋ฐ ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ๋ถ„ํฌ.** ์ค‘์š”ํ•œ ๊ฒƒ์€ ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ๋ชจ๋“  ๋ฌธ์ œ๊ฐ€ ๊ธฐ๋ณธ LM์˜ ํ•™์Šต ๋งˆ๊ฐ์ผ์ธ 2023๋…„ 6์›” 1์ผ ์ดํ›„์— ์ถœ์ œ๋œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ํŠธ๋ ˆ์ธ ๋ฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์˜ ๋ชจ๋“  ๋ฌธ์ œ๋Š” 2023๋…„ 6์›” 1์ผ ์ด์ „์— ํ•ด๊ฒฐ๋˜์—ˆ์œผ๋ฏ€๋กœ ํŠœ๋‹ ํ”„๋กœ์„ธ์Šค์—์„œ ๋ˆ„์ถœ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. **(b) ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ **๊ธฐ์ค€ ์„ฑ๋Šฅ**์ด๋ฉฐ, ํ‘œ์ค€ ์˜ค์ฐจ(SE) 1๊ฐœ(์ „์ฒด ๊ฒฐ๊ณผ๋Š” ํ‘œ 7 ์ฐธ์กฐ)์ž…๋‹ˆ๋‹ค. ๋ฌด์ž‘์œ„ ๊ธฐ์ค€์„ : 0.250; ์ธ๊ฐ„ ํฌ๋ผ์šฐ๋“œ: 0.149. ์ด ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์ด ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ์˜ˆ์ธก์— ๋Šฅ์ˆ™ํ•˜์ง€ ์•Š๋‹ค๋Š” ๊ฒƒ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\end{table} Table 2: **(a) Distribution of our train, validation, and test sets across all 5 forecasting platforms.** Importantly, every question in the test set is from June 1, 2023 or later, after the training cut-off of our base LMs. Meanwhile, all questions in the train and validation sets were resolved before June 1, 2023, ensuring no leakage from the tuning process. **(b) Baseline performance** of pre-trained models on the test set, with 1 standard error (SE) (see full results in Table 7). Random baseline: 0.250; human crowd: 0.149. The results underscore that models are not naturally good at forecasting. [|assistant|]
๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด 8-way tensor parallelism (Shoeybi et al., 2019)๊ณผ data parallelism์„ ์กฐํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ, ๋ถ„์‚ฐ optimizer๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ data-parallel replicas์— ๊ฑธ์ณ optimizer ์ƒํƒœ๋ฅผ shardํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. data parallelism์˜ ์ •๋„๋Š” batch size๊ฐ€ ๋Š˜์–ด๋‚จ์— ๋”ฐ๋ผ 96์—์„œ 384๊นŒ์ง€ ๋‹ค์–‘ํ•˜๊ฒŒ ์กฐ์ •๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 2๋Š” batch size ramp์˜ 3๋‹จ๊ณ„๋ฅผ ์š”์•ฝํ•œ ๊ฒƒ์œผ๋กœ, ๋ฐ˜๋ณต๋‹น ์‹œ๊ฐ„๊ณผ ๋ชจ๋ธ FLOP/s ํ™œ์šฉ๋„(MFU)(Chowdhery et al., 2022; Korthikanti et al., 2022)๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. MFU๋Š” GPU๊ฐ€ ๋ชจ๋ธ ํ›ˆ๋ จ์—์„œ ์–ผ๋งˆ๋‚˜ ํšจ์œจ์ ์œผ๋กœ ํ™œ์šฉ๋˜๋Š”์ง€๋ฅผ ์ •๋Ÿ‰ํ™”ํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ์€ ์•ฝ 13์ผ ๋งŒ์— ์™„๋ฃŒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด 8๋ฐฉํ–ฅ ํ…์„œ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ(Shoeybi et al., 2019)์™€ ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๋ฅผ ์กฐํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ, ๋ถ„์‚ฐ ์˜ตํ‹ฐ๋งˆ์ด์ €๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ ๋ณต์ œ๋ณธ์— ๊ฑธ์ณ ์˜ตํ‹ฐ๋งˆ์ด์ € ์ƒํƒœ๋ฅผ ์ƒค๋“œํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ์˜ ์ •๋„๋Š” ๋ฐฐ์น˜ ํฌ๊ธฐ๊ฐ€ ๋Š˜์–ด๋‚จ์— ๋”ฐ๋ผ 96์—์„œ 384๊นŒ์ง€ ๋‹ค์–‘ํ•˜๊ฒŒ ์กฐ์ •๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 2๋Š” ๋ฐฐ์น˜ ํฌ๊ธฐ ๋žจํ”„์˜ 3๋‹จ๊ณ„๋ฅผ ์š”์•ฝํ•œ ๊ฒƒ์œผ๋กœ, ๋ฐ˜๋ณต๋‹น ์‹œ๊ฐ„๊ณผ ๋ชจ๋ธ FLOP/s ํ™œ์šฉ๋„(MFU)(Chowdhery et al., 2022; Korthikanti et al., 2022)๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. MFU๋Š” GPU๊ฐ€ ๋ชจ๋ธ ํ›ˆ๋ จ์—์„œ ์–ผ๋งˆ๋‚˜ ํšจ์œจ์ ์œผ๋กœ ํ™œ์šฉ๋˜๋Š”์ง€๋ฅผ ์ •๋Ÿ‰ํ™”ํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ์€ ์•ฝ 13์ผ ๋งŒ์— ์™„๋ฃŒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We used a combination of 8-way tensor parallelism (Shoeybi et al., 2019) and data parallelism to train the model; we also use a distributed optimizer to shard the optimizer state over the data-parallel replicas. The degree of data parallelism was varied from 96 to 384 as the batch size was ramped up. Table 2 summarizes the 3 stages of batch size ramp, and includes the per-iteration time and model FLOP/s utilization (MFU) (Chowdhery et al., 2022; Korthikanti et al., 2022). MFU quantifies how efficiently the GPUs are utilized in model training. Training was completed in approximately 13 calendar days. [|assistant|]
๋””์ฝ”๋” ์ „์šฉ Transformer ๊ธฐ๋ฐ˜ Large Language Model์ธ Nemotron-4 15B๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์˜์–ด, 53๊ฐœ์˜ ์ž์—ฐ์–ด ๋ฐ 43๊ฐœ์˜ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด๋ฅผ ํฌํ•จํ•œ 8์กฐ ํ† ํฐ์— ๊ฑธ์ณ ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค. Nemotron-4 15B๋Š” ๋ฉ€ํ‹ฐ๋ง๊ตฌ ๋„๋ฉ”์ธ์— ํŠนํ™”๋œ ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋ฒ”์šฉ ์–ธ์–ด ๋ชจ๋ธ ์ค‘ ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ๋‹ค๊ตญ์–ด ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Nemotron-4๋Š” Large Language Models์˜ ์‚ฌ์ „ ํ•™์Šต ์„ธํŠธ๋ฅผ ๊ณ„์† ํ™•์žฅํ•˜์—ฌ ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
๋””์ฝ”๋” ์ „์šฉ ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์ธ Nemotron-4 15B๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์˜์–ด, 53๊ฐœ์˜ ์ž์—ฐ์–ด ๋ฐ 43๊ฐœ์˜ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด๋ฅผ ํฌํ•จํ•œ 8์กฐ ํ† ํฐ์— ๊ฑธ์ณ ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค. Nemotron-4 15B๋Š” ๋ฉ€ํ‹ฐ๋ง๊ตฌ ๋„๋ฉ”์ธ์— ํŠนํ™”๋œ ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋ฒ”์šฉ ์–ธ์–ด ๋ชจ๋ธ ์ค‘ ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ๋‹ค๊ตญ์–ด ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Nemotron-4๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต ์„ธํŠธ๋ฅผ ๊ณ„์† ํ™•์žฅํ•˜์—ฌ ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹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|]We present Nemotron-4 15B, a decoder-only transformer-based large language model. It is trained on 8 trillion tokens spanning English, 53 additional natural languages as well as 43 programming languages. Nemotron-4 15B exhibits the strongest multilingual performance of any general purpose language model at its scale - even outperforming models specialized for the multilingual domain. Nemotron-4 demonstrates that pre-training sets for large language models can continue to be scaled up even further in order to improve the abilities of models. [|assistant|]
๊ทธ๋ฆผ A.16: **์ •์„ฑ์  ๊ธฐ์ค€์˜ ์ž๋™ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ Prompt.** ์ƒˆ๋กœ์šด ๋Œ€ํ™”๋ฅผ ํ‰๊ฐ€ํ•˜๊ณ  ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ๋Œ€ํ™”์™€ ํ•จ๊ป˜ ์ž์ฒด ์ƒ์„ฑ๋œ ์„ค๋ช…์ด ํฌํ•จ๋œ few-shot Prompt๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด few-shot Prompt์—๋Š” 5์  ์ฒ™๋„์˜ ๊ฐ ์š”์ ์— ๋Œ€ํ•œ ์˜ˆ์‹œ๊ฐ€ ํ•˜๋‚˜์”ฉ ํฌํ•จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ A.16: **์ •์„ฑ์  ๊ธฐ์ค€์˜ ์ž๋™ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ํ”„๋กฌํ”„ํŠธ.** ์ƒˆ๋กœ์šด ๋Œ€ํ™”๋ฅผ ํ‰๊ฐ€ํ•˜๊ณ  ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ๋Œ€ํ™”์™€ ํ•จ๊ป˜ ์ž์ฒด ์ƒ์„ฑ๋œ ์„ค๋ช…์ด ํฌํ•จ๋œ ์†Œ์ˆ˜์˜ ์ƒท ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์†Œ์ˆ˜์˜ ์ƒท ํ”„๋กฌํ”„ํŠธ์—๋Š” 5์  ์ฒ™๋„์˜ ๊ฐ ์š”์ ์— ๋Œ€ํ•œ ์˜ˆ์‹œ๊ฐ€ ํ•˜๋‚˜์”ฉ ํฌํ•จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure A.16: **Prompt for the auto-evaluation of qualitative criteria.** We constructed few-shot prompt with self-generated explanations alongside their respective dialogues as exemplars to evaluate and rate a new dialogue. This few-shot prompt included one example for each point on the 5-point rating scale. [|assistant|]
**์ฃผ์š” ๊ฒฐ๊ณผ** ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 4์— ๋ณด๊ณ ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. _WizardLM_์€ Alpaca ๋ฐ Vicuna-7b๋ณด๋‹ค ํ›จ์”ฌ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ _Evol-Instruct_์˜ ํšจ๊ณผ๋ฅผ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
**์ฃผ์š” ๊ฒฐ๊ณผ** ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 4์— ๋ณด๊ณ ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. WizardLM_์€ ์•ŒํŒŒ์นด ๋ฐ ๋น„์ฟ ๋‚˜-7b๋ณด๋‹ค ํ›จ์”ฌ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ _Evol-Instruct_์˜ ํšจ๊ณผ๋ฅผ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Main Results** The results of the experiment are reported in Figure 4. _WizardLM_ achieved significantly better results than Alpaca and Vicuna-7b, which demonstrates the effectiveness of _Evol-Instruct_. [|assistant|]
์ถœ๋ ฅ ํ˜•์‹ ๋ถ„์„์—์„œ๋Š” ์–ธ์–ด ์‹๋ณ„ ๋ฐ ๋ฐ˜๋ณต ๋ถ„์„์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ํšจ์œจ์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด vLLM Kwon et al. (2023)์„ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, Supervised Fine-Tuning (SFT) ๋ฐ Reinforcement Learning from Human Feedback (RLHF)๊ณผ ๊ฐ™์€ alignment ์ž‘์—…์„ ๊ฑฐ์นœ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ”„๋กฌํ”„ํŠธ๋ฅผ "[INST] <context> [/INST]"๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ์„ max_tokens ์„ค์ • 512๋กœ ๊ตฌ์„ฑํ•˜๊ณ  nuclear sampling์„ ํ™œ์šฉํ•˜์—ฌ temperature๋ฅผ 0.1๋กœ ์„ค์ •ํ•˜๊ณ  top_p๋ฅผ 0.9๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ถœ๋ ฅ ํ˜•์‹ ๋ถ„์„์—์„œ๋Š” ์–ธ์–ด ์‹๋ณ„ ๋ฐ ๋ฐ˜๋ณต ๋ถ„์„์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ํšจ์œจ์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด vLLM Kwon et al. (2023)์„ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •(SFT) ๋ฐ ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์œผ๋กœ๋ถ€ํ„ฐ์˜ ๊ฐ•ํ™” ํ•™์Šต(RLHF)๊ณผ ๊ฐ™์€ ์ •๋ ฌ ์ž‘์—…์„ ๊ฑฐ์นœ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ”„๋กฌํ”„ํŠธ๋ฅผ "[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We perform two tasks in output format analysis: language identification and repetition analysis. We utilized vLLM Kwon et al. (2023) to enhance efficiency. Expressly, for models that have undergone alignment operations, such as Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), we set up our prompt as "[INST] <context> [/INST]". We configure the model with a max_tokens setting of 512 and utilize nuclear sampling, setting the temperature to 0.1 and top_p to 0.9. [|assistant|]
๊ทธ๋ฆผ 5: **์ผ๋ฐ˜์ ์ธ pretraining ๊ฒฐ๊ณผ. 80G ์ผ๋ฐ˜ ํ† ํฐ์— ๋Œ€ํ•ด Tinyllama-1B์˜ pretraining์„ ๊ณ„์†ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 5: **์ผ๋ฐ˜์ ์ธ ์‚ฌ์ „ ํ•™์Šต ๊ฒฐ๊ณผ. 80G ์ผ๋ฐ˜ ํ† ํฐ์— ๋Œ€ํ•ด Tinyllama-1B์˜ ์‚ฌ์ „ ํ•™์Šต์„ ๊ณ„์†ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 5: **General pretraining results. We continual pretraining Tinyllama-1B on 80G general tokens. [|assistant|]
ChatGPT ๋ฐ GPT-4์™€ ๊ฐ™์€ Large Language Models (LLMs)์€ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์—ฐ๊ตฌ์— ํš๊ธฐ์ ์ธ ๋ณ€ํ™”๋ฅผ ๊ฐ€์ ธ์™”์œผ๋ฉฐ Artificial General Intelligence (AGI)์„ ํ–ฅํ•œ ์œ ๋งํ•œ ์ง„์ „์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์„ ํ›ˆ๋ จํ•˜๊ณ  ๋ฐฐํฌํ•˜๋Š” ๋ฐ ๋“œ๋Š” ๋†’์€ ๋น„์šฉ์€ ํˆฌ๋ช…ํ•˜๊ณ  ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ํ•™์ˆ  ์—ฐ๊ตฌ์— ์ƒ๋‹นํ•œ ์žฅ์• ๋ฌผ์ด ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLaMA์™€ ๊ฐ™์€ ๋ช‡๋ช‡ Large Language Models์€ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ์˜ํ•ด ์˜คํ”ˆ ์†Œ์Šคํ™”๋˜์—ˆ์ง€๋งŒ, ๋Œ€๋ถ€๋ถ„ ์˜์–ด ๋ง๋ญ‰์น˜์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์–ด ๋‹ค๋ฅธ ์–ธ์–ด์— ๋Œ€ํ•œ ์œ ์šฉ์„ฑ์ด ์ œํ•œ์ ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๊ธฐ๋Šฅ๊ณผ Instruction์„ ๋”ฐ๋ฅด๋Š” ๊ธฐ๋Šฅ์„ LLaMA์— ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด LLaMA์˜ ๊ธฐ์กด ์–ดํœ˜์— 20,000๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜์—ฌ ์ค‘๊ตญ์–ด์˜ ์ธ์ฝ”๋”ฉ ํšจ์œจ์„ฑ๊ณผ ์˜๋ฏธ๋ก ์  ์ดํ•ด๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 2์ฐจ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ฉํ•˜๊ณ  ์ค‘๊ตญ์–ด Instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ๋ชจ๋ธ์„ Fine-tuningํ•˜์—ฌ ๋ชจ๋ธ์˜ Instruction ์ดํ•ด ๋ฐ ์‹คํ–‰ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์ƒˆ๋กœ ์ œ์•ˆ๋œ ๋ชจ๋ธ์€ ์ค‘๊ตญ์–ด ์ฝ˜ํ…์ธ ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์žˆ์–ด ๊ธฐ์กด LLaMA์˜ ๋Šฅ๋ ฅ์„ ๋ˆˆ์— ๋„๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ C-Eval ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฐ๊ณผ๋Š” ์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ๋ช‡ ๋ฐฐ ํฌ๊ธฐ์˜ ๋ชจ๋ธ๋“ค ์‚ฌ์ด์—์„œ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
ChatGPT ๋ฐ GPT-4์™€ ๊ฐ™์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์—ฐ๊ตฌ์— ํš๊ธฐ์ ์ธ ๋ณ€ํ™”๋ฅผ ๊ฐ€์ ธ์™”์œผ๋ฉฐ ์ธ๊ณต ์ผ๋ฐ˜ ์ง€๋Šฅ(AGI)์„ ํ–ฅํ•œ ์œ ๋งํ•œ ์ง„์ „์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์„ ํ›ˆ๋ จํ•˜๊ณ  ๋ฐฐํฌํ•˜๋Š” ๋ฐ ๋“œ๋Š” ๋†’์€ ๋น„์šฉ์€ ํˆฌ๋ช…ํ•˜๊ณ  ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ํ•™์ˆ  ์—ฐ๊ตฌ์— ์ƒ๋‹นํ•œ ์žฅ์• ๋ฌผ์ด ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLaMA์™€ ๊ฐ™์€ ๋ช‡๋ช‡ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ์˜ํ•ด ์˜คํ”ˆ ์†Œ์Šคํ™”๋˜์—ˆ์ง€๋งŒ, ๋Œ€๋ถ€๋ถ„ ์˜์–ด ๋ง๋ญ‰์น˜์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์–ด ๋‹ค๋ฅธ ์–ธ์–ด์— ๋Œ€ํ•œ ์œ ์šฉ์„ฑ์ด ์ œํ•œ์ ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๊ธฐ๋Šฅ๊ณผ ๋ช…๋ น์–ด๋ฅผ ๋”ฐ๋ฅด๋Š” ๊ธฐ๋Šฅ์„ LLaMA์— ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด LLaMA์˜ ๊ธฐ์กด ์–ดํœ˜์— 20,000๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜์—ฌ ์ค‘๊ตญ์–ด์˜ ์ธ์ฝ”๋”ฉ ํšจ์œจ์„ฑ๊ณผ ์˜๋ฏธ๋ก ์  ์ดํ•ด๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 2์ฐจ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ฉํ•˜๊ณ  ์ค‘๊ตญ์–ด ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ๋ชจ๋ธ์˜ ๋ช…๋ น์–ด ์ดํ•ด ๋ฐ ์‹คํ–‰ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์ƒˆ๋กœ ์ œ์•ˆ๋œ ๋ชจ๋ธ์€ ์ค‘๊ตญ์–ด ์ฝ˜ํ…์ธ ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์žˆ์–ด ๊ธฐ์กด LLaMA์˜ ๋Šฅ๋ ฅ์„ ๋ˆˆ์— ๋„๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ C-Eval ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฐ๊ณผ๋Š” ์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ๋ช‡ ๋ฐฐ ํฌ๊ธฐ์˜ ๋ชจ๋ธ๋“ค ์‚ฌ์ด์—์„œ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Large Language Models (LLMs), such as ChatGPT and GPT-4, have dramatically transformed natural language processing research and shown promising strides towards Artificial General Intelligence (AGI). Nonetheless, the high costs associated with training and deploying LLMs present substantial obstacles to transparent, accessible academic research. While several large language models, such as LLaMA, have been open-sourced by the community, these predominantly focus on English corpora, limiting their usefulness for other languages. In this paper, we propose a method to augment LLaMA with capabilities for understanding and generating Chinese text and its ability to follow instructions. We achieve this by extending LLaMA's existing vocabulary with an additional 20,000 Chinese tokens, thereby improving its encoding efficiency and semantic understanding of Chinese. We further incorporate secondary pre-training using Chinese data and fine-tune the model with Chinese instruction datasets, significantly enhancing the model's ability to comprehend and execute instructions. Our experimental results indicate that the newly proposed model markedly enhances the original LLaMA's proficiency in understanding and generating Chinese content. Additionally, the results on the C-Eval dataset yield competitive performance among the models with several times the size of ours. [|assistant|]
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLMs)์€ ๊ฐ•๋ ฅํ•˜์ง€๋งŒ hallucination(Pal et al., 2023; Sun et al., 2024; Ahmad et al., 2023)์ด ๋ฐœ์ƒํ•˜๊ธฐ ์‰ฝ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•™์Šต ์ฝ”ํผ์Šค์— ํฌํ•จ๋œ ์ง€์‹์— ์ œํ•œ์„ ๋ฐ›๊ธฐ ๋•Œ๋ฌธ์— ์ตœ๊ทผ ์ด๋ฒคํŠธ๋‚˜ ๊ณต๊ฐœ์ ์œผ๋กœ ์ œํ•œ๋œ ์ •๋ณด์— ๋Œ€ํ•œ ์ฟผ๋ฆฌ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. Retrieval-augmented generation (RAG)์€ LLM ํ”„๋กฌํ”„ํŠธ์— ๊ด€๋ จ ๊ฒ€์ƒ‰ ์ฝ˜ํ…์ธ ๋ฅผ ์ œ๊ณตํ•˜๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ํ”„๋ ˆ์ž„์›Œํฌ๋กœ, ๋ชจ๋ธ ์ •ํ™•๋„๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Mao et al., 2020; Chen et al., 2024; Lewis et al., 2020).[|endofturn|]
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ๊ฐ•๋ ฅํ•˜์ง€๋งŒ ํ™˜๊ฐ ํ˜„์ƒ(Pal et al., 2023; Sun et al., 2024; Ahmad et al., 2023)์ด ๋ฐœ์ƒํ•˜๊ธฐ ์‰ฝ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•™์Šต ์ฝ”ํผ์Šค์— ํฌํ•จ๋œ ์ง€์‹์— ์ œํ•œ์„ ๋ฐ›๊ธฐ ๋•Œ๋ฌธ์— ์ตœ๊ทผ ์ด๋ฒคํŠธ๋‚˜ ๊ณต๊ฐœ์ ์œผ๋กœ ์ œํ•œ๋œ ์ •๋ณด์— ๋Œ€ํ•œ ์ฟผ๋ฆฌ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์ƒ์„ฑ(RAG)์€ LLM ํ”„๋กฌํ”„ํŠธ์— ๊ด€๋ จ ๊ฒ€์ƒ‰ ์ฝ˜ํ…์ธ ๋ฅผ ์ œ๊ณตํ•˜๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ํ”„๋ ˆ์ž„์›Œํฌ๋กœ, ๋ชจ๋ธ ์ •ํ™•๋„๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Mao et al., 2020; Chen et al., 2024; Lewis et al., 2020).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Large language models (LLMs), though powerful, are prone to hallucination (Pal et al., 2023; Sun et al., 2024; Ahmad et al., 2023). Additionally, they are restricted to knowledge contained in their training corpus, and so are unable to answer queries about recent events or publicly restricted information. Retrieval augmented generation (RAG) is a commonly used framework that provides relevant retrieved content in the LLM prompt and can significantly improve model accuracy (Mao et al., 2020; Chen et al., 2024; Lewis et al., 2020). [|assistant|]
**Order effect for human logical reasoning.** ์—ฐ์—ญ์  ์ถ”๋ก ์—์„œ๋Š” ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ค‘์š”ํ•˜์ง€ ์•Š์ง€๋งŒ, ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ธ๊ฐ„์˜ ์ถ”๋ก  ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค(Dekeyser et al., 2000; Girotto et al., 1997). Dekeyser ๋“ฑ(2000)์€ _๊ณต๋™ ์ฐธ์กฐ_๋ฅผ ์ „์ œ ์ˆœ์„œ์— ๋Œ€ํ•œ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„, ์ฆ‰ ์ธ๊ฐ„์€ ๊ฐ ์ „์ œ๋ฅผ ๋ณธ ํ›„ ์ฆ‰๊ฐ์ ์ธ ๊ฒฐ๋ก ์„ ๋„์ถœํ•  ์ˆ˜ ์žˆ๋Š” ์ˆœ์„œ๋กœ ์ „์ œ๋ฅผ ์ œ์‹œํ•˜๋Š” ๊ฒƒ์„ ์„ ํ˜ธํ•œ๋‹ค๊ณ  ์„ค๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” LLMs๋„ ์ด๋Ÿฌํ•œ ์„ ํ˜ธ๋„๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ๊ทœ์น™์˜ ์ˆœ์„œ๊ฐ€ ์‹ค์ธก ์ฆ๋ช…์„ ๋”ฐ๋ฅผ ๋•Œ ๊ฐ€์žฅ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. Girotto ๋“ฑ(1997)์€ ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ธ๊ฐ„์˜ ๋…ผ๋ฆฌ์  ์ถ”๋ก ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ์—ฐ๊ตฌํ•œ ๊ฒฐ๊ณผ, ์ „์ œ ์ˆœ์„œ๊ฐ€ _๋ชจ๋‘์Šค ํ†จ๋ Œ์Šค_ ๋ฌธ์ œ(์ฆ‰, P์ด๋ฉด Q, Q๊ฐ€ ์•„๋‹ˆ๋ฏ€๋กœ P๊ฐ€ ์•„๋‹ˆ๋‹ค)๋ฅผ ํ‘ธ๋Š” ๋ฐ๋Š” ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€๋งŒ _๋ชจ๋‘์Šค ํฌ๋„จ์Šค_ ๋ฌธ์ œ(์ฆ‰, P์ด๋ฉด Q, P์ด๋ฏ€๋กœ Q)์—๋Š” ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ ๊ทธ๋“ค์€ ๊ทœ์น™๊ณผ ์‚ฌ์‹ค ์‚ฌ์ด์˜ ๋‹ค๋ฅธ ์ˆœ์„œ์˜ ์˜ํ–ฅ, ์˜ˆ๋ฅผ ๋“ค์–ด _๋ชจ๋‘์Šค ํ†จ๋ Œ์Šค_ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด ๋ถ€์ • ์ง„์ˆ (Q๊ฐ€ ์•„๋‹˜)์„ ๊ทœ์น™(P์ด๋ฉด Q)๋ณด๋‹ค ๋จผ์ € ์ œ์‹œํ•˜๋ฉด ๊ทธ ๋ฐ˜๋Œ€ ์ˆœ์„œ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์šฐ๋ฆฌ ์—ฐ๊ตฌ๋Š” ์ธ๊ฐ„๊ณผ LLMs ๋ชจ๋‘์—๊ฒŒ ๋” ์‰ฌ์šด _๋ชจ๋‘์Šค ํฌ๋„จ์Šค_ ๋ฌธ์ œ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, LLMs ์„ฑ๋Šฅ์ด ์ „์ œ์˜ ์ˆœ์„œ์— ์—ฌ์ „ํžˆ ๋งค์šฐ ๋ฏผ๊ฐํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
**์ธ๊ฐ„์˜ ๋…ผ๋ฆฌ์  ์ถ”๋ก ์— ๋Œ€ํ•œ ์ˆœ์„œ ํšจ๊ณผ.** ์—ฐ์—ญ์  ์ถ”๋ก ์—์„œ๋Š” ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ค‘์š”ํ•˜์ง€ ์•Š์ง€๋งŒ, ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ธ๊ฐ„์˜ ์ถ”๋ก  ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค(Dekeyser et al., 2000; Girotto et al., 1997). Dekeyser ๋“ฑ(2000)์€ _๊ณต๋™ ์ฐธ์กฐ_๋ฅผ ์ „์ œ ์ˆœ์„œ์— ๋Œ€ํ•œ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„, ์ฆ‰ ์ธ๊ฐ„์€ ๊ฐ ์ „์ œ๋ฅผ ๋ณธ ํ›„ ์ฆ‰๊ฐ์ ์ธ ๊ฒฐ๋ก ์„ ๋„์ถœํ•  ์ˆ˜ ์žˆ๋Š” ์ˆœ์„œ๋กœ ์ „์ œ๋ฅผ ์ œ์‹œํ•˜๋Š” ๊ฒƒ์„ ์„ ํ˜ธํ•œ๋‹ค๊ณ  ์„ค๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” LLM๋„ ์ด๋Ÿฌํ•œ ์„ ํ˜ธ๋„๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ๊ทœ์น™์˜ ์ˆœ์„œ๊ฐ€ ์‹ค์ธก ์ฆ๋ช…์„ ๋”ฐ๋ฅผ ๋•Œ ๊ฐ€์žฅ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. Girotto ๋“ฑ(1997)์€ ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ธ๊ฐ„์˜ ๋…ผ๋ฆฌ์  ์ถ”๋ก ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ์—ฐ๊ตฌํ•œ ๊ฒฐ๊ณผ, ์ „์ œ ์ˆœ์„œ๊ฐ€ _๋ชจ๋‘์Šค ํ†จ๋ Œ์Šค_ ๋ฌธ์ œ(์ฆ‰, P์ด๋ฉด Q, Q๊ฐ€ ์•„๋‹ˆ๋ฏ€๋กœ P๊ฐ€ ์•„๋‹ˆ๋‹ค)๋ฅผ ํ‘ธ๋Š” ๋ฐ๋Š” ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€๋งŒ _๋ชจ๋‘์Šค ํฌ๋„จ์Šค_ ๋ฌธ์ œ(์ฆ‰, P์ด๋ฉด Q, P์ด๋ฏ€๋กœ Q)์—๋Š” ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ ๊ทธ๋“ค์€ ๊ทœ์น™๊ณผ ์‚ฌ์‹ค ์‚ฌ์ด์˜ ๋‹ค๋ฅธ ์ˆœ์„œ์˜ ์˜ํ–ฅ, ์˜ˆ๋ฅผ ๋“ค์–ด _๋ชจ๋‘์Šค ํ†จ๋ Œ์Šค_ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด ๋ถ€์ • ์ง„์ˆ (Q๊ฐ€ ์•„๋‹˜)์„ ๊ทœ์น™(P์ด๋ฉด Q)๋ณด๋‹ค ๋จผ์ € ์ œ์‹œํ•˜๋ฉด ๊ทธ ๋ฐ˜๋Œ€ ์ˆœ์„œ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์šฐ๋ฆฌ ์—ฐ๊ตฌ๋Š” ์ธ๊ฐ„๊ณผ LLM ๋ชจ๋‘์—๊ฒŒ ๋” ์‰ฌ์šด _๋ชจ๋‘์Šค ํฌ๋„จ์Šค_ ๋ฌธ์ œ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, LLM ์„ฑ๋Šฅ์ด ์ „์ œ์˜ ์ˆœ์„œ์— ์—ฌ์ „ํžˆ ๋งค์šฐ ๋ฏผ๊ฐํ•˜๋‹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 for human logical reasoning.** Although the premise order does not matter in deductive reasoning, several studies show that the premise order can impact the human reasoning performance (Dekeyser et al., 2000; Girotto et al., 1997). Dekeyser et al. (2000) described _co-reference_ as a human preference of premise order; i.e., humans prefer the premises to be presented in an order where they can draw immediate conclusions after seeing each one. In this work, we show that LLMs also have such a preference, and they achieve the best performance when the ordering of rules follows the ground truth proof. Girotto et al. (1997) studied how the premise order affects logical reasoning for humans, and found that the premise order has a significant effect in solving _modus tollens_ problems (i.e., if P, then Q; not Q; therefore, not P), but not _modus ponens_ problems (i.e., if P, then Q; P; therefore, Q). However, differing from our work, they studied the influence of different ordering between rules and facts, e.g., their experiments on _modus tollens_ problems show that presenting negation statements (not Q) before rules (if P, then Q) improves the performance over the reverse order. On the other hand, our work focuses on _modus ponens_ problems that are easier for both humans and LLMs, and we show that the LLM performance is still quite sensitive to the ordering of the premises. [|assistant|]
๋งˆ์ง€๋ง‰์œผ๋กœ, ์„น์…˜ 7์—์„œ๋Š” Learning Rate ์žฌ๊ฐ€์—ด์ด ์›์น˜ ์•Š๋Š” ๋ง๊ฐ์„ ์œ ๋ฐœํ•  ์ˆ˜ ์žˆ์Œ์„ ์„ค๋ช…ํ•˜๊ณ , ์ด๋ฅผ ์šฐํšŒํ•  ์ˆ˜ ์žˆ๋Š” ์œ ๋งํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ ๋ฌดํ•œ Learning Rate ์ผ์ •์„ ์†Œ๊ฐœํ•˜๋ฉฐ, ์ด๋Ÿฌํ•œ ์ผ์ •์„ baseline๊ณผ ๋น„๊ตํ•œ๋‹ค.[|endofturn|]
๋งˆ์ง€๋ง‰์œผ๋กœ, ์„น์…˜ 7์—์„œ๋Š” LR ์žฌ๊ฐ€์—ด์ด ์›์น˜ ์•Š๋Š” ๋ง๊ฐ์„ ์œ ๋ฐœํ•  ์ˆ˜ ์žˆ์Œ์„ ์„ค๋ช…ํ•˜๊ณ , ์ด๋ฅผ ์šฐํšŒํ•  ์ˆ˜ ์žˆ๋Š” ์œ ๋งํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ ๋ฌดํ•œ ํ•™์Šต ์†๋„ ์ผ์ •์„ ์†Œ๊ฐœํ•˜๋ฉฐ, ์ด๋Ÿฌํ•œ ์ผ์ •์„ ๊ธฐ์ค€๊ณผ ๋น„๊ตํ•œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Finally, in section 7, we illustrate LR re-warming can cause unwanted forgetting, introduce infinite learning rate schedules as a promising way to circumvent it, and compare these schedules to baselines. [|assistant|]
128-core TPU v3์—์„œ ์•ฝ 131์‹œ๊ฐ„ ๋˜๋Š” 5.5์ผ ๋™์•ˆ ํ›ˆ๋ จํ–ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€๋žต ์ค‘๋ณต๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์›๋ณธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณด๋‹ค 3.9% ์ž‘๊ณ  63์‹œ๊ฐ„/epoch์— ํ›ˆ๋ จ๋˜๋ฏ€๋กœ ๋‘ epoch์— ๋Œ€ํ•ด ์•ฝ 5์‹œ๊ฐ„์˜ ์ปดํ“จํŒ… ์‹œ๊ฐ„์„ ์ ˆ์•ฝํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. XL-Originalmodel์€ ๋ถ๋ฏธ์—์„œ ํ›ˆ๋ จ๋˜์—ˆ๊ณ , XL-ExactSubstr๊ณผ XL-NearDup์€ ๋Œ€๋งŒ์—์„œ ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ ์—๋„ˆ์ง€ ์–‘์„ ์ถ”์ •ํ•˜๊ธฐ ์œ„ํ•ด Patterson ๋“ฑ(2021)์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‹œ๊ฐ„๋‹น \(MWh\)/core์˜ ์–‘์„ ๊ณ„์‚ฐํ•˜๊ณ  ์‚ฌ์šฉ๋Ÿ‰์„ ๊ณฑํ•˜์—ฌ ๊ณ„์‚ฐํ–ˆ์Šต๋‹ˆ๋‹ค(์ด๋Ÿฌํ•œ ๊ฐ’์„ ๊ณ„์‚ฐํ•œ ๋ฐฉ๋ฒ•์€ ํ‘œ 6์„ ์ฐธ์กฐํ•˜์„ธ์š”). ๊ฐ„๋‹จํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด ๋Œ€๋งŒ ๋ฐ์ดํ„ฐ์„ผํ„ฐ์˜ ์ถ”์ •์น˜๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. XL-Original ๋ฐ XL-ExactSubstr์˜ 2๊ฐœ epoch๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ๋Š” \(5.86MWh\)๊ฐ€ ์‚ฌ์šฉ๋œ ๊ฒƒ์œผ๋กœ ์ถ”์ •๋ฉ๋‹ˆ๋‹ค. XL-NearDup์€ ๋” ์ ์€ ๋‹จ๊ณ„๋กœ ํ›ˆ๋ จ๋˜๋ฉฐ \(5.63MWh\)๊ฐ€ ์‚ฌ์šฉ๋œ ๊ฒƒ์œผ๋กœ ์ถ”์ •๋ฉ๋‹ˆ๋‹ค. ๊ฐ ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ๋Š” 64-core TPU v3 pod slice์—์„œ ์•ฝ 3์ผ์ด ์†Œ์š”๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” ์•ฝ 1.61MWh๊ฐ€ ์†Œ์š”๋˜๋Š” ๊ฒƒ์œผ๋กœ ์ถ”์ •๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
128์ฝ”์–ด TPU v3์—์„œ ์•ฝ 131์‹œ๊ฐ„ ๋˜๋Š” 5.5์ผ ๋™์•ˆ ํ›ˆ๋ จํ–ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€๋žต ์ค‘๋ณต๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์›๋ณธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณด๋‹ค 3.9% ์ž‘๊ณ  63์‹œ๊ฐ„/์—ํฌํฌ์— ํ›ˆ๋ จ๋˜๋ฏ€๋กœ ๋‘ ์—ํฌํฌ์— ๋Œ€ํ•ด ์•ฝ 5์‹œ๊ฐ„์˜ ์ปดํ“จํŒ… ์‹œ๊ฐ„์„ ์ ˆ์•ฝํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. XL-Originalmodel์€ ๋ถ๋ฏธ์—์„œ ํ›ˆ๋ จ๋˜์—ˆ๊ณ , XL-ExactSubstr๊ณผ XL-NearDup์€ ๋Œ€๋งŒ์—์„œ ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ ์—๋„ˆ์ง€ ์–‘์„ ์ถ”์ •ํ•˜๊ธฐ ์œ„ํ•ด Patterson ๋“ฑ(2021)์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‹œ๊ฐ„๋‹น \(MWh\)/์ฝ”์–ด์˜ ์–‘์„ ๊ณ„์‚ฐํ•˜๊ณ  ์‚ฌ์šฉ๋Ÿ‰์„ ๊ณฑํ•˜์—ฌ ๊ณ„์‚ฐํ–ˆ์Šต๋‹ˆ๋‹ค(์ด๋Ÿฌํ•œ ๊ฐ’์„ ๊ณ„์‚ฐํ•œ ๋ฐฉ๋ฒ•์€ ํ‘œ 6์„ ์ฐธ์กฐํ•˜์„ธ์š”). ๊ฐ„๋‹จํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด ๋Œ€๋งŒ ๋ฐ์ดํ„ฐ์„ผํ„ฐ์˜ ์ถ”์ •์น˜๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. XL-Original ๋ฐ XL-ExactSubstr์˜ 2๊ฐœ ์—ํฌํฌ๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ๋Š” \(5.86MWh\)๊ฐ€ ์‚ฌ์šฉ๋œ ๊ฒƒ์œผ๋กœ ์ถ”์ •๋ฉ๋‹ˆ๋‹ค. XL-NearDup์€ ๋” ์ ์€ ๋‹จ๊ณ„๋กœ ํ›ˆ๋ จ๋˜๋ฉฐ \(5.63MWh\)๊ฐ€ ์‚ฌ์šฉ๋œ ๊ฒƒ์œผ๋กœ ์ถ”์ •๋ฉ๋‹ˆ๋‹ค. ๊ฐ ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ๋Š” 64์ฝ”์–ด TPU v3 ํฌ๋“œ ์Šฌ๋ผ์ด์Šค์—์„œ ์•ฝ 3์ผ์ด ์†Œ์š”๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” ์•ฝ 1.61MWh๊ฐ€ ์†Œ์š”๋˜๋Š” ๊ฒƒ์œผ๋กœ ์ถ”์ •๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We trained for approximately 131 hours or 5.5 days on a 128-core TPU v3. The approximate deduplicated dataset is 3.9% smaller than the original dataset and trains in 63 hours/epoch, saving us around 5 hours of compute time for the two epochs. The XL-Originalmodel was trained in North America where the XL-ExactSubstr and XL-NearDup were trained in Taiwan. We used data from Patterson et al. (2021) to estimate amount of energy used in training these models by computing the amount of \(MWh\)/hour/core and multiplying by our usage (see Table 6 for how we computed these values). For simplicity, we use estimates from Taiwaninese datacenters as an estimate. We estimate training 2 epochs of XL-Original and XL-ExactSubstr uses \(5.86MWh\). XL-NearDup is trained for fewer steps and we estimate uses \(5.63MWh\). Training each base model was approximately 3 days on a 64-core TPU v3 pod slice which uses an estimated \(1.61MWh\). [|assistant|]
์ค‘๊ฐ„ ๋ฐ qa ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด SimCSE objective(Gao et al., 2021)๋กœ ํ•™์Šต๋œ pre-trained BERT ๋ชจ๋ธ์˜ ๋ฌธ์žฅ ์ž„๋ฒ ๋”ฉ์˜ cosine similarity๋ฅผ ๊ทธ๋ฆผ 8(a) ๋ฐ (b)์— ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์œ ์‚ฌ์„ฑ์„ ๊ณ„์‚ฐํ•  ๋•Œ outlier๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ค‘๊ฐ„ ๋ฐ qa ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด SimCSE ๋ชฉํ‘œ(Gao et al., 2021)๋กœ ํ•™์Šต๋œ ์‚ฌ์ „ ํ•™์Šต๋œ BERT ๋ชจ๋ธ์˜ ๋ฌธ์žฅ ์ž„๋ฒ ๋”ฉ์˜ ์ฝ”์‚ฌ์ธ ์œ ์‚ฌ์„ฑ์„ ๊ทธ๋ฆผ 8(a) ๋ฐ (b)์— ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์œ ์‚ฌ์„ฑ์„ ๊ณ„์‚ฐํ•  ๋•Œ ์ด์ƒ๊ฐ’์„ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We show the cosine similarity of the sentence embeddings from a pre-trained BERT model trained with SimCSE objective (Gao et al., 2021) for medium and qa prompts in Figure 8(a) and (b). When computing similarity, we remove outliers. [|assistant|]
์—ฌ๊ธฐ์— ๋ฌธ์„œ ์ˆ˜์ค€์—์„œ MinHash [8] ์ค‘๋ณต ์ œ๊ฑฐ ๋‹จ๊ณ„๋ฅผ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. MinHash์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” ์„œ๋ช…๋‹น 20๊ฐœ์˜ ํ•ด์‹œ, 20๊ฐœ์˜ ๋ฒ„ํ‚ท, ๋ฒ„ํ‚ท๋‹น 1๊ฐœ์˜ ํ–‰์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” MinHashLSH์˜ spark ๊ตฌํ˜„์—์„œ ๊ธฐ๋ณธ ๋งค๊ฐœ๋ณ€์ˆ˜์ด๋ฉฐ, ์ปดํ“จํŒ… ์ œํ•œ์œผ๋กœ ์ธํ•ด ์ด๋Ÿฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜์— ๋Œ€ํ•œ hyperparameter sweep์„ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” ํ›จ์”ฌ ๋” ๊ณต๊ฒฉ์ ์ธ ๋งค๊ฐœ๋ณ€์ˆ˜๋กœ MinHash๋ฅผ ์‹คํ–‰ํ•˜๋ ค๊ณ  ์‹œ๋„ํ–ˆ์Šต๋‹ˆ๋‹ค: Lee ๋“ฑ [27]๊ณผ Penedo ๋“ฑ [39]์€ ๋ฒ„ํ‚ท 20๊ฐœ, ๋ฒ„ํ‚ท๋‹น ํ•ด์‹œ 450๊ฐœ, ํ•ด์‹œ๋‹น ์„œ๋ช… 9000๊ฐœ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋” ๊ณต๊ฒฉ์ ์ธ MinHash๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋” ๋งŽ์€ ํ…œํ”Œ๋ฆฟ์„ ์ œ๊ฑฐํ•˜์—ฌ ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ์‹œ์ž‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ D4์˜ SemDeDup ๋‹จ๊ณ„๊ฐ€ ๋œ ํ•„์š”ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ถ”์ธกํ•ฉ๋‹ˆ๋‹ค. Abbas ๋“ฑ [1]์€ C4์—์„œ ๊ณ ์ •๋œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋น„์œจ 3.9%์—์„œ Lee ๋“ฑ [27]์˜ MinHash์™€ SemDeDup์˜ ์„ฑ๋Šฅ์ด ๋น„์Šทํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋Š”๋ฐ, ์ด๋Š” SemDeDup์ด ๊ณต๊ฒฉ์ ์ธ MinHash์™€ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฑธ๋Ÿฌ๋‚ธ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ hyperparameter๋ฅผ sweepํ•˜๋Š” ๊ฒƒ์€ ํ–ฅํ›„ ์ž‘์—…์œผ๋กœ ๋‚จ๊ฒจ๋‘์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์—ฌ๊ธฐ์— ๋ฌธ์„œ ์ˆ˜์ค€์—์„œ MinHash [8] ์ค‘๋ณต ์ œ๊ฑฐ ๋‹จ๊ณ„๋ฅผ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. MinHash์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” ์„œ๋ช…๋‹น 20๊ฐœ์˜ ํ•ด์‹œ, 20๊ฐœ์˜ ๋ฒ„ํ‚ท, ๋ฒ„ํ‚ท๋‹น 1๊ฐœ์˜ ํ–‰์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” MinHashLSH์˜ ์ŠคํŒŒํฌ ๊ตฌํ˜„์—์„œ ๊ธฐ๋ณธ ๋งค๊ฐœ๋ณ€์ˆ˜์ด๋ฉฐ, ์ปดํ“จํŒ… ์ œํ•œ์œผ๋กœ ์ธํ•ด ์ด๋Ÿฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜์— ๋Œ€ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์Šค์œ•์„ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” ํ›จ์”ฌ ๋” ๊ณต๊ฒฉ์ ์ธ ๋งค๊ฐœ๋ณ€์ˆ˜๋กœ MinHash๋ฅผ ์‹คํ–‰ํ•˜๋ ค๊ณ  ์‹œ๋„ํ–ˆ์Šต๋‹ˆ๋‹ค: Lee ๋“ฑ [27]๊ณผ Penedo ๋“ฑ [39]์€ ๋ฒ„ํ‚ท 20๊ฐœ, ๋ฒ„ํ‚ท๋‹น ํ•ด์‹œ 450๊ฐœ, ํ•ด์‹œ๋‹น ์„œ๋ช… 9000๊ฐœ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋” ๊ณต๊ฒฉ์ ์ธ MinHash๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋” ๋งŽ์€ ํ…œํ”Œ๋ฆฟ์„ ์ œ๊ฑฐํ•˜์—ฌ ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ์‹œ์ž‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ D4์˜ SemDeDup ๋‹จ๊ณ„๊ฐ€ ๋œ ํ•„์š”ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ถ”์ธกํ•ฉ๋‹ˆ๋‹ค. Abbas ๋“ฑ [1]์€ C4์—์„œ ๊ณ ์ •๋œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋น„์œจ 3.9%์—์„œ Lee ๋“ฑ [27]์˜ MinHash์™€ SemDeDup์˜ ์„ฑ๋Šฅ์ด ๋น„์Šทํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋Š”๋ฐ, ์ด๋Š” SemDeDup์ด ๊ณต๊ฒฉ์ ์ธ MinHash์™€ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฑธ๋Ÿฌ๋‚ธ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์Šค์œ•ํ•˜๋Š” ๊ฒƒ์€ ํ–ฅํ›„ ์ž‘์—…์œผ๋กœ ๋‚จ๊ฒจ๋‘์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]On top of this, we add an additional step of MinHash [8] de-duplication at the document-level. The parameters for MinHash are 20 hashes per signature, 20 buckets, and 1 row per bucket. These parameters are the default parameters in the spark implementation of MinHashLSH, and we did not do a hyperparameter sweep on these parameters due to compute limitations. Previous work has attempted running MinHash with much more aggressive parameters: Lee et al. [27] and Penedo et al. [39] use \(20\) buckets, \(450\) hashes per bucket, and \(9000\) signatures per hash. We conjecture that more aggressive MinHash would remove more templates, resulting in a higher-quality starting dataset, potentially making the SemDeDup step of D4 less necessary. Abbas et al. [1] did find that the performance of MinHash from Lee et al. [27] and SemDeDup are comparable at a fixed data selection ratio of 3.9% on C4, indicating that SemDeDup filters out similar data to aggressive MinHash does. We leave sweeping over these hyperparameters as future work. [|assistant|]
\(\bullet\)_Embeddings_. ์ด ํ˜•์‹์—์„œ๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ •๋ณด๊ฐ€ embedding ๋ฒกํ„ฐ๋กœ ์ธ์ฝ”๋”ฉ๋˜์–ด ๋ฉ”๋ชจ๋ฆฌ ๊ฒ€์ƒ‰ ๋ฐ ์ฝ๊ธฐ ํšจ์œจ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, MemoryBank [39]๋Š” ๊ฐ ๋ฉ”๋ชจ๋ฆฌ ์„ธ๊ทธ๋จผํŠธ๋ฅผ embedding ๋ฒกํ„ฐ๋กœ ์ธ์ฝ”๋”ฉํ•˜์—ฌ ๊ฒ€์ƒ‰์„ ์œ„ํ•œ ์ธ๋ฑ์‹ฑ๋œ ์ฝ”ํผ์Šค๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. [16]์€ ์ฐธ์กฐ ๊ณ„ํš์„ embeddings๋กœ ํ‘œํ˜„ํ•˜์—ฌ ์ผ์น˜ ๋ฐ ์žฌ์‚ฌ์šฉ์„ ์šฉ์ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ChatDev [18]์€ ๊ฒ€์ƒ‰์„ ์œ„ํ•ด ๋Œ€ํ™” ๊ธฐ๋ก์„ ๋ฒกํ„ฐ๋กœ ์ธ์ฝ”๋”ฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_์ž„๋ฒ ๋”ฉ_. ์ด ํ˜•์‹์—์„œ๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ •๋ณด๊ฐ€ ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ๋กœ ์ธ์ฝ”๋”ฉ๋˜์–ด ๋ฉ”๋ชจ๋ฆฌ ๊ฒ€์ƒ‰ ๋ฐ ์ฝ๊ธฐ ํšจ์œจ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, MemoryBank [39]๋Š” ๊ฐ ๋ฉ”๋ชจ๋ฆฌ ์„ธ๊ทธ๋จผํŠธ๋ฅผ ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ๋กœ ์ธ์ฝ”๋”ฉํ•˜์—ฌ ๊ฒ€์ƒ‰์„ ์œ„ํ•œ ์ธ๋ฑ์‹ฑ๋œ ์ฝ”ํผ์Šค๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. [16]์€ ์ฐธ์กฐ ๊ณ„ํš์„ ์ž„๋ฒ ๋”ฉ์œผ๋กœ ํ‘œํ˜„ํ•˜์—ฌ ์ผ์น˜ ๋ฐ ์žฌ์‚ฌ์šฉ์„ ์šฉ์ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ChatDev [18]์€ ๊ฒ€์ƒ‰์„ ์œ„ํ•ด ๋Œ€ํ™” ๊ธฐ๋ก์„ ๋ฒกํ„ฐ๋กœ ์ธ์ฝ”๋”ฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_Embeddings_. In this format, memory information is encoded into embedding vectors, which can enhance the memory retrieval and reading efficiency. For instance, MemoryBank [39] encodes each memory segment into an embedding vector,which creates an indexed corpus for retrieval. [16] represents reference plans as embeddings to facilitate matching and reuse. Furthermore, ChatDev [18] encodes dialogue history into vectors for retrieval. [|assistant|]
# ์ถœ๋ ฅ K=2์ผ ๋•Œ Equation 18์€ Bradley-Terry ๋ชจ๋ธ๋กœ ์ถ•์†Œ๋œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”. ๊ทธ๋Ÿฌ๋‚˜ ์ผ๋ฐ˜์ ์ธ Plackett-Luce ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ์—๋„ Equation 5์˜ ๊ฒฐ๊ณผ๋ฅผ ํ™œ์šฉํ•˜๊ณ  ์ตœ์ ์˜ ์ •์ฑ…์œผ๋กœ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ๋ณด์ƒ ํ•จ์ˆ˜๋ฅผ ๋Œ€์ž…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
K=2\)์ผ ๋•Œ ๋ฐฉ์ •์‹ 18์€ ๋ธŒ๋ž˜๋“ค๋ฆฌ-ํ…Œ๋ฆฌ ๋ชจ๋ธ๋กœ ์ถ•์†Œ๋œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”. ๊ทธ๋Ÿฌ๋‚˜ ์ผ๋ฐ˜์ ์ธ ํ”Œ๋ผ์ผ“-๋ฃจ์Šค ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ์—๋„ Eq. 5์˜ ๊ฒฐ๊ณผ๋ฅผ ํ™œ์šฉํ•˜๊ณ  ์ตœ์ ์˜ ์ •์ฑ…์œผ๋กœ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ๋ณด์ƒ ํ•จ์ˆ˜๋ฅผ ๋Œ€์ž…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Notice that when \(K=2\), Equation 18 reduces to the Bradley-Terry model. However, for the general Plackett-Luce model, we can still utilize the results of Eq. 5 and substitute the reward function parameterized by its optimal policy. [|assistant|]
์ผ๋ถ€ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๊ฐ€ ์Šค์ผ€์ค„๋Ÿฌ์— ์˜ํ•ด ์ค‘๊ฐ„ ๋‹จ๊ณ„์—์„œ ์ค‘์ง€๋œ ๊ฒฝ์šฐ, context ๊ด€๋ฆฌ์ž๋Š” snapshot ๊ธฐ๋Šฅ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‘๋‹ต ์ƒ์„ฑ์„ ์œ„ํ•ด ํƒ์ƒ‰ ์ค‘์ธ ๋ชจ๋“  ์ค‘๊ฐ„ ํ™•๋ฅ  ๋ฐ ๊ฒฝ๋กœ๋ฅผ ํฌํ•จํ•˜์—ฌ LLM์˜ beam search ํŠธ๋ฆฌ์˜ ํ˜„์žฌ ์ƒํƒœ๋ฅผ ์บก์ฒ˜ํ•˜๊ณ  ์ €์žฅํ•ฉ๋‹ˆ๋‹ค. ์žฌ๊ฐœ ์‹œ restoration ๊ธฐ๋Šฅ์„ ์‚ฌ์šฉํ•˜์—ฌ snapshot์—์„œ ์ €์žฅ๋œ ์ƒํƒœ๋ฅผ ๋‹ค์‹œ ๋กœ๋“œํ•˜๋ฉด LLM์ด ์ค‘๋‹จ๋œ ์ง€์ ๋ถ€ํ„ฐ ์ •ํ™•ํžˆ ์ตœ์ข… ๋‹ต์— ๋„๋‹ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: _ํŒŒ๋ฆฌ์˜ ๋‚ ์”จ ๊ฒ€์ƒ‰_. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ context ๊ด€๋ฆฌ์ž๋Š” ํ•œ ์—์ด์ „ํŠธ์˜ ์š”์ฒญ์ด ์ผ์‹œ์ ์œผ๋กœ ์ค‘๋‹จ๋˜์–ด๋„ ์ง„ํ–‰ ์ƒํ™ฉ์ด ์†์‹ค๋˜์ง€ ์•Š๋„๋ก ํ•˜์—ฌ ์‘๋‹ต ์ƒ์„ฑ์˜ ํ’ˆ์งˆ๊ณผ ํšจ์œจ์„ฑ์„ ์ €ํ•˜์‹œํ‚ค์ง€ ์•Š์œผ๋ฉด์„œ ๋ฆฌ์†Œ์Šค ์‚ฌ์šฉ์„ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ผ๋ถ€ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๊ฐ€ ์Šค์ผ€์ค„๋Ÿฌ์— ์˜ํ•ด ์ค‘๊ฐ„ ๋‹จ๊ณ„์—์„œ ์ค‘์ง€๋œ ๊ฒฝ์šฐ, ์ปจํ…์ŠคํŠธ ๊ด€๋ฆฌ์ž๋Š” ์Šค๋ƒ…์ƒท ๊ธฐ๋Šฅ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‘๋‹ต ์ƒ์„ฑ์„ ์œ„ํ•ด ํƒ์ƒ‰ ์ค‘์ธ ๋ชจ๋“  ์ค‘๊ฐ„ ํ™•๋ฅ  ๋ฐ ๊ฒฝ๋กœ๋ฅผ ํฌํ•จํ•˜์—ฌ LLM์˜ ๋น” ๊ฒ€์ƒ‰ ํŠธ๋ฆฌ์˜ ํ˜„์žฌ ์ƒํƒœ๋ฅผ ์บก์ฒ˜ํ•˜๊ณ  ์ €์žฅํ•ฉ๋‹ˆ๋‹ค. ์žฌ๊ฐœ ์‹œ ๋ณต์› ๊ธฐ๋Šฅ์„ ์‚ฌ์šฉํ•˜์—ฌ ์Šค๋ƒ…์ƒท์—์„œ ์ €์žฅ๋œ ์ƒํƒœ๋ฅผ ๋‹ค์‹œ ๋กœ๋“œํ•˜๋ฉด LLM์ด ์ค‘๋‹จ๋œ ์ง€์ ๋ถ€ํ„ฐ ์ •ํ™•ํžˆ ์ตœ์ข… ๋‹ต์— ๋„๋‹ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: _ํŒŒ๋ฆฌ์˜ ๋‚ ์”จ ๊ฒ€์ƒ‰_. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ์ปจํ…์ŠคํŠธ ๊ด€๋ฆฌ์ž๋Š” ํ•œ ์—์ด์ „ํŠธ์˜ ์š”์ฒญ์ด ์ผ์‹œ์ ์œผ๋กœ ์ค‘๋‹จ๋˜์–ด๋„ ์ง„ํ–‰ ์ƒํ™ฉ์ด ์†์‹ค๋˜์ง€ ์•Š๋„๋ก ํ•˜์—ฌ ์‘๋‹ต ์ƒ์„ฑ์˜ ํ’ˆ์งˆ๊ณผ ํšจ์œจ์„ฑ์„ ์ €ํ•˜์‹œํ‚ค์ง€ ์•Š์œผ๋ฉด์„œ ๋ฆฌ์†Œ์Šค ์‚ฌ์šฉ์„ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]When such generation process has been suspended by the scheduler at an intermediate step, the context manager uses the snapshot function to capture and store the current state of the LLM's beam search tree, including all intermediate probabilities and paths being explored for generating the response. Upon resumption, the restoration function is employed to reload the saved state from the snapshot, allowing the LLM to continue its generation process exactly from the point of suspension to reach the final answer: _Search weather in Paris_. In this way, the context manager ensures that the temporary suspension of one agent's request does not lead to a loss of progress, thereby optimizing resource use without compromising the quality and efficiency of response generation. [|assistant|]
# ์ถœ๋ ฅ ์‹ ๋ขฐ ๊ตฌ๊ฐ„ 0.001์„ ์–ป์œผ๋ ค๋ฉด ๋ช‡ ๋ฒˆ์˜ training ์‹คํ–‰์ด ํ•„์š”ํ• ๊นŒ์š”? ํ‘œ์ค€ ์˜ค์ฐจ๋Š” \(\sqrt{N}\)์—์„œ ์ถ•์†Œ๋˜๋ฏ€๋กœ ์‹คํ—˜ ํšŸ์ˆ˜๋ฅผ \(50^{2}\) ๋˜๋Š” 2500๋ฐฐ๋กœ ๋Š˜๋ ค์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, Hoffmann ๋“ฑ์ด ๋ณด๊ณ ํ•œ ๊ฒƒ๋งŒํผ ์ด˜์ด˜ํ•œ ์‹ ๋ขฐ ๊ตฌ๊ฐ„์„ ์–ป์œผ๋ ค๋ฉด ์•ฝ \(240\๋ฐฐ 2116=600,000\)์˜ training ๊ฒฐ๊ณผ์— ์•ก์„ธ์Šคํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์‹ ๋ขฐ ๊ตฌ๊ฐ„ 0.001์„ ์–ป์œผ๋ ค๋ฉด ๋ช‡ ๋ฒˆ์˜ ํ›ˆ๋ จ ์‹คํ–‰์ด ํ•„์š”ํ• ๊นŒ์š”? ํ‘œ์ค€ ์˜ค์ฐจ๋Š” \(\sqrt{N}\)์—์„œ ์ถ•์†Œ๋˜๋ฏ€๋กœ ์‹คํ—˜ ํšŸ์ˆ˜๋ฅผ \(50^{2}\) ๋˜๋Š” 2500๋ฐฐ๋กœ ๋Š˜๋ ค์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, Hoffmann ๋“ฑ์ด ๋ณด๊ณ ํ•œ ๊ฒƒ๋งŒํผ ์ด˜์ด˜ํ•œ ์‹ ๋ขฐ ๊ตฌ๊ฐ„์„ ์–ป์œผ๋ ค๋ฉด ์•ฝ \(240\๋ฐฐ 2116=600,000\)์˜ ํ›ˆ๋ จ ๊ฒฐ๊ณผ์— ์•ก์„ธ์Šคํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]How many training runs would we need to observe to get a confidence interval of 0.001? Since the standard errors shrink in \(\sqrt{N}\), we would need to increase the number of experiments by a factor of \(50^{2}\) or 2500. That means that we would need to have access to the results from nearly \(240\times 2116=600,000\) training runs to obtain a confidence interval as tight as that reported by Hoffmann et al. [|assistant|]
์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์œ„์—์„œ ์„ค๋ช…ํ•œ ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. Gorilla APIBench ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์ด๋ฏธ ๋‹ต๋ณ€์— ์ถ”๋ก ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. Fig. 3์— ์ƒ์„ฑ ๋‹จ๊ณ„์˜ ์˜ˆ๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์ƒ์„ธํ•œ ์ถ”๋ก  ๋‹ต๋ณ€์—๋Š” #begin_quote## ๋ฐ ##end_quote## ๋‚ด๋ถ€์˜ ์›๋ณธ ์ปจํ…์ŠคํŠธ์—์„œ ์ธ์šฉ๋ฌธ๊ณผ ์ธ์šฉ๋ฌธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฒฐ๋ก ์— ๋„๋‹ฌํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ์„น์…˜์—์„œ ์ƒ์„ธํ•œ ์ถ”๋ก  ๋‹จ๋ฝ์„ ์ถ”๊ฐ€ํ•˜๋ฉด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค.[|endofturn|]
์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์œ„์—์„œ ์„ค๋ช…ํ•œ ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ณ ๋ฆด๋ผ APIBench ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์ด๋ฏธ ๋‹ต๋ณ€์— ์ถ”๋ก ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. Fig. 3์— ์ƒ์„ฑ ๋‹จ๊ณ„์˜ ์˜ˆ๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์ƒ์„ธํ•œ ์ถ”๋ก  ๋‹ต๋ณ€์—๋Š” #begin_quote## ๋ฐ ##end_quote## ๋‚ด๋ถ€์˜ ์›๋ณธ ์ปจํ…์ŠคํŠธ์—์„œ ์ธ์šฉ๋ฌธ๊ณผ ์ธ์šฉ๋ฌธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฒฐ๋ก ์— ๋„๋‹ฌํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ์„น์…˜์—์„œ ์ƒ์„ธํ•œ ์ถ”๋ก  ๋‹จ๋ฝ์„ ์ถ”๊ฐ€ํ•˜๋ฉด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]For all the datasets in our experiments, we generate the answers using the technique described above. Note that the Gorilla APIBench dataset, already includes reasoning in the answers. We provide an example of the generation step in Fig. 3, the detailed reasoning answer includes a citation from the original context inside #begin_quote## and ##end_quote## as well as the detailed explanation on how to reach the conclusion based on the citations. We demonstrate that adding detailed reasoning paragraphs helps boost the model's performance in our experiment section. [|assistant|]
๋จธ์‹  ๋ฒˆ์—ญ์„ ์œ„ํ•ด ์ฒ˜์Œ ์ œ์•ˆ๋œ Transformer ์–ธ์–ด ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜๋Š” Encoder์™€ Decoder๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. Encoder๋Š” N = 6์˜ ๋™์ผํ•œ Transformer ๋ ˆ์ด์–ด ์Šคํƒ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ ˆ์ด์–ด์—๋Š” ๋‘ ๊ฐœ์˜ ํ•˜์œ„ ๋ ˆ์ด์–ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ๋Š” multi-head self-attention ๋ ˆ์ด์–ด์ด๊ณ , ๋‹ค๋ฅธ ํ•˜๋‚˜๋Š” ๋‹จ์ˆœํ•œ ์œ„์น˜๋ณ„ ์™„์ „ ์—ฐ๊ฒฐ ํ”ผ๋“œ ํฌ์›Œ๋“œ ๋„คํŠธ์›Œํฌ์ž…๋‹ˆ๋‹ค. Decoder๋Š” 6๊ฐœ์˜ ๋™์ผํ•œ ๋ ˆ์ด์–ด ์Šคํƒ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๊ฐ Encoder ๋ ˆ์ด์–ด์˜ ๋‘ ๊ฐœ์˜ ํ•˜์œ„ ๋ ˆ์ด์–ด ์™ธ์—๋„ Decoder์—๋Š” Encoder ์Šคํƒ์˜ ์ถœ๋ ฅ์— ๋Œ€ํ•ด multi-head attention์„ ์ˆ˜ํ–‰ํ•˜๋Š” ์„ธ ๋ฒˆ์งธ ํ•˜์œ„ ๋ ˆ์ด์–ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. Attention ๊ธฐ๋Šฅ์€ ์ฟผ๋ฆฌ์™€ ํ‚ค-๊ฐ’ ์Œ ์„ธํŠธ๋ฅผ ์ถœ๋ ฅ์— ๋งคํ•‘ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ ์ฟผ๋ฆฌ, ํ‚ค, ๊ฐ’ ๋ฐ ์ถœ๋ ฅ์€ ๋ชจ๋‘ ๋ฒกํ„ฐ์ž…๋‹ˆ๋‹ค. ์ถœ๋ ฅ์€ ๊ฐ’์˜ ๊ฐ€์ค‘์น˜ ํ•ฉ์œผ๋กœ ๊ณ„์‚ฐ๋˜๋ฉฐ, ๊ฐ ๊ฐ’์— ํ• ๋‹น๋œ ๊ฐ€์ค‘์น˜๋Š” ํ•ด๋‹น ํ‚ค์™€ ์ฟผ๋ฆฌ์˜ ํ˜ธํ™˜์„ฑ ํ•จ์ˆ˜์— ์˜ํ•ด ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. \(d_{model}\) ์ฐจ์›์˜ ํ‚ค, ๊ฐ’, ์ฟผ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์ผ attention ๊ธฐ๋Šฅ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋Œ€์‹ , ํ•™์Šต๋œ ์„œ๋กœ ๋‹ค๋ฅธ ์„ ํ˜• ํˆฌ์˜์„ ํ†ตํ•ด ์ฟผ๋ฆฌ, ํ‚ค, ๊ฐ’ \(h\)๋ฅผ ๊ฐ๊ฐ \(d_{k}\), \(d_{k}\), \(d_{v}\) ์ฐจ์›์œผ๋กœ ์„ ํ˜•์ ์œผ๋กœ ํˆฌ์˜ํ•˜๋Š” ๊ฒƒ์ด ์œ ๋ฆฌํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์œ„์น˜ ์ธ์ฝ”๋”ฉ์€ ์‹œํ€€์Šค์—์„œ ํ† ํฐ์˜ ์ƒ๋Œ€์  ๋˜๋Š” ์ ˆ๋Œ€์  ์œ„์น˜์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์œตํ•ฉํ•˜๊ธฐ ์œ„ํ•ด ํ†ตํ•ฉ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋จธ์‹  ๋ฒˆ์—ญ์„ ์œ„ํ•ด ์ฒ˜์Œ ์ œ์•ˆ๋œ ํŠธ๋žœ์Šคํฌ๋จธ ์–ธ์–ด ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜๋Š” ์ธ์ฝ”๋”์™€ ๋””์ฝ”๋”๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ธ์ฝ”๋”๋Š” N = 6์˜ ๋™์ผํ•œ ํŠธ๋žœ์Šคํฌ๋จธ ๋ ˆ์ด์–ด ์Šคํƒ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ ˆ์ด์–ด์—๋Š” ๋‘ ๊ฐœ์˜ ํ•˜์œ„ ๋ ˆ์ด์–ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ๋Š” ๋‹ค์ค‘ ํ—ค๋“œ ์ž๊ธฐ ์ฃผ์˜ ๋ ˆ์ด์–ด์ด๊ณ , ๋‹ค๋ฅธ ํ•˜๋‚˜๋Š” ๋‹จ์ˆœํ•œ ์œ„์น˜๋ณ„ ์™„์ „ ์—ฐ๊ฒฐ ํ”ผ๋“œ ํฌ์›Œ๋“œ ๋„คํŠธ์›Œํฌ์ž…๋‹ˆ๋‹ค. ๋””์ฝ”๋”๋Š” 6๊ฐœ์˜ ๋™์ผํ•œ ๋ ˆ์ด์–ด ์Šคํƒ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ์ธ์ฝ”๋” ๋ ˆ์ด์–ด์˜ ๋‘ ๊ฐœ์˜ ํ•˜์œ„ ๋ ˆ์ด์–ด ์™ธ์—๋„ ๋””์ฝ”๋”์—๋Š” ์ธ์ฝ”๋” ์Šคํƒ์˜ ์ถœ๋ ฅ์— ๋Œ€ํ•ด ๋‹ค์ค‘ ํ—ค๋“œ ์ฃผ์˜๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ์„ธ ๋ฒˆ์งธ ํ•˜์œ„ ๋ ˆ์ด์–ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์˜ ๊ธฐ๋Šฅ์€ ์ฟผ๋ฆฌ์™€ ํ‚ค-๊ฐ’ ์Œ ์„ธํŠธ๋ฅผ ์ถœ๋ ฅ์— ๋งคํ•‘ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ ์ฟผ๋ฆฌ, ํ‚ค, ๊ฐ’ ๋ฐ ์ถœ๋ ฅ์€ ๋ชจ๋‘ ๋ฒกํ„ฐ์ž…๋‹ˆ๋‹ค. ์ถœ๋ ฅ์€ ๊ฐ’์˜ ๊ฐ€์ค‘์น˜ ํ•ฉ์œผ๋กœ ๊ณ„์‚ฐ๋˜๋ฉฐ, ๊ฐ ๊ฐ’์— ํ• ๋‹น๋œ ๊ฐ€์ค‘์น˜๋Š” ํ•ด๋‹น ํ‚ค์™€ ์ฟผ๋ฆฌ์˜ ํ˜ธํ™˜์„ฑ ํ•จ์ˆ˜์— ์˜ํ•ด ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. d_{model}$ ์ฐจ์›์˜ ํ‚ค, ๊ฐ’, ์ฟผ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์ผ ์ฃผ์˜ ๊ธฐ๋Šฅ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋Œ€์‹ , ํ•™์Šต๋œ ์„œ๋กœ ๋‹ค๋ฅธ ์„ ํ˜• ํˆฌ์˜์„ ํ†ตํ•ด ์ฟผ๋ฆฌ, ํ‚ค, ๊ฐ’ $h$๋ฅผ ๊ฐ๊ฐ $d_{k}$, $d_{k}$, $d_{v}$ ์ฐจ์›์œผ๋กœ ์„ ํ˜•์ ์œผ๋กœ ํˆฌ์˜ํ•˜๋Š” ๊ฒƒ์ด ์œ ๋ฆฌํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์œ„์น˜ ์ธ์ฝ”๋”ฉ์€ ์‹œํ€€์Šค์—์„œ ํ† ํฐ์˜ ์ƒ๋Œ€์  ๋˜๋Š” ์ ˆ๋Œ€์  ์œ„์น˜์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์œตํ•ฉํ•˜๊ธฐ ์œ„ํ•ด ํ†ตํ•ฉ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The Transformer language model architecture, originally proposed for machine translation, consists of an encoder and a decoder. The encoder is composed of a stack of N = 6 identical Transformer layers. Each layer has two sub-layers. The first one is a multi-head self-attention layer, and the other one is a simple position-wise fully connected feed-forward network. The decoder is composed of a stack of 6 identical layers. In addition to the two sub-layers in each encoder layer, the decoder has a third sub-layer, which performs multi-head attention over the output of the encoder stack. The attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is computed as a weighted sum of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key. Instead of performing a single attention function with \(d_{model}\) dimensional keys, values and queries, it is found to be beneficial to linearly project the queries, keys and values \(h\) with different, learned linear projections to \(d_{k}\), \(d_{k}\) and \(d_{v}\) dimensions, respectively. Positional encoding is incorporated to fuse information about the relative or absolute position of the tokens in the sequence. [|assistant|]
์‹œ์Šคํ…œ์—์„œ Fine-tuned GPT-4๋ฅผ Fine-tuned GPT-3.5๋กœ ๋Œ€์ฒดํ•˜๊ณ  ์„น์…˜ 6.1์—์„œ์™€ ๋™์ผํ•œ ๋ฐฉ๋ฒ•๋ก ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์šฐ๋ฆฌ์˜ Brier score๋Š” ์ด์ „ ์ ์ˆ˜์ธ .179์— ๋น„ํ•ด .182๋กœ ์•ฝ๊ฐ„๋งŒ ๋‚˜๋น ์ง€๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์‹œ์Šคํ…œ์—์„œ ๋ฏธ์„ธ ์กฐ์ •๋œ GPT-4๋ฅผ ๋ฏธ์„ธ ์กฐ์ •๋œ GPT-3.5๋กœ ๋Œ€์ฒดํ•˜๊ณ  ์„น์…˜ 6.1์—์„œ์™€ ๋™์ผํ•œ ๋ฐฉ๋ฒ•๋ก ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์šฐ๋ฆฌ์˜ ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜๋Š” ์ด์ „ ์ ์ˆ˜์ธ .179์— ๋น„ํ•ด .182๋กœ ์•ฝ๊ฐ„๋งŒ ๋‚˜๋น ์ง€๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We replace fine-tuned GPT-4 in our system with fine-tuned GPT-3.5, and evaluate using the same methodology as in Section 6.1. We find here that our Brier score is only slightly worse:.182 compared to the previous score of.179. [|assistant|]
๋‹ค์Œ์œผ๋กœ, ๋ชจ๋ธ์ด ์ž‘์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ด์ „์— ๋ณด์ง€ ๋ชปํ–ˆ๋˜ ์ƒˆ๋กœ์šด ์ง€์‹์„ ํ•™์Šตํ•  ๋•Œ Fine-tuning ๋‹จ๊ณ„์— ์—ญ ํ›ˆ๋ จ์„ ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ์„น์…˜ 3.3์—์„œ ์„ค๋ช…ํ•œ ๊ฒƒ๊ณผ ๋™์ผํ•œ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ๊ณผ ์ถ”๊ฐ€๋กœ LLama-2 7B ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ€์ƒ์˜ ์‚ฌ์‹ค๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์ถ”๊ฐ€๋กœ Fine-tuningํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ๋Š” "[์ด๋ฆ„]์€ [์„ค๋ช…]" ๋˜๋Š” ๊ทธ ๋ฐ˜๋Œ€ ํ˜•์‹์˜ ๋ฌธ์žฅ์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ์ด๋ฆ„๊ณผ ์„ค๋ช…์€ ๋ฌด์ž‘์œ„๋กœ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ€์ƒ์  ํŠน์„ฑ์€ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ํ•ด๋‹น ์‚ฌ์‹ค์„ ๋ณผ ์ˆ˜ ์—†์œผ๋ฉฐ, Fine-tuning ์ค‘์— ์ง€์ •๋œ ๋ฐฉํ–ฅ์œผ๋กœ๋งŒ ๋ณผ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ชจ๋ธ์ด ํ•™์Šต ์ค‘์— ๋ณธ ๊ฒƒ๊ณผ ๊ฐ™์€ ๋ฐฉํ–ฅ ๋˜๋Š” ๋ฐ˜๋Œ€ ๋ฐฉํ–ฅ์œผ๋กœ ์ด๋Ÿฌํ•œ ์‚ฌ์‹ค์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋‹ค์Œ์œผ๋กœ, ๋ชจ๋ธ์ด ์ž‘์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ด์ „์— ๋ณด์ง€ ๋ชปํ–ˆ๋˜ ์ƒˆ๋กœ์šด ์ง€์‹์„ ํ•™์Šตํ•  ๋•Œ ๋ฏธ์„ธ ์กฐ์ • ๋‹จ๊ณ„์— ์—ญ ํ›ˆ๋ จ์„ ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ์„น์…˜ 3.3์—์„œ ์„ค๋ช…ํ•œ ๊ฒƒ๊ณผ ๋™์ผํ•œ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ๊ณผ ์ถ”๊ฐ€๋กœ LLama-2 7B ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ€์ƒ์˜ ์‚ฌ์‹ค๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์ถ”๊ฐ€๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ๋Š” "[์ด๋ฆ„]์€ [์„ค๋ช…]" ๋˜๋Š” ๊ทธ ๋ฐ˜๋Œ€ ํ˜•์‹์˜ ๋ฌธ์žฅ์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ์ด๋ฆ„๊ณผ ์„ค๋ช…์€ ๋ฌด์ž‘์œ„๋กœ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ€์ƒ์  ํŠน์„ฑ์€ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ํ•ด๋‹น ์‚ฌ์‹ค์„ ๋ณผ ์ˆ˜ ์—†์œผ๋ฉฐ, ๋ฏธ์„ธ ์กฐ์ • ์ค‘์— ์ง€์ •๋œ ๋ฐฉํ–ฅ์œผ๋กœ๋งŒ ๋ณผ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ชจ๋ธ์ด ํ•™์Šต ์ค‘์— ๋ณธ ๊ฒƒ๊ณผ ๊ฐ™์€ ๋ฐฉํ–ฅ ๋˜๋Š” ๋ฐ˜๋Œ€ ๋ฐฉํ–ฅ์œผ๋กœ ์ด๋Ÿฌํ•œ ์‚ฌ์‹ค์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We next explore if our reverse training can be applied to the finetuning stage when the model is learning new, previously unseen knowledge from a small training dataset. We use the same pre-trained models described in Section 3.3 and an additional LLama-2 7B model, and further finetune them on a dataset made up of fictitious facts. These data are made up of statements of the form "[name] is [description]" (or the reverse) where the names and descriptions are randomly generated. The fictitious nature of this dataset guarantees that those facts are not seen during pre-training, and are only seen in the specified direction during finetuning. The model is then tested to see if it is able to learn these facts in the same or reverse direction that it has seen during training. [|assistant|]
์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๋ชจ๋ธ์˜ ์ง€์‹๊ณผ ์‚ฌ์‹ค ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ์–ตํ•˜๊ณ  ์ดํ•ดํ•˜๋ฉฐ ๊ฒ€์ƒ‰ํ•˜๋Š” ๋Šฅ๋ ฅ์— ๋Œ€ํ•œ ํ‰๊ฐ€์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” _knowledge injection_(Wang et al., 2020; Chen et al., 2022; Liu et al., 2020; Lauscher et al., 2020)์˜ ๊ฐœ๋…์„ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ํ…์ŠคํŠธ ์ฝ”ํผ์Šค ํ˜•ํƒœ์˜ ์ง€์‹ ๋ฒ ์ด์Šค๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์— ์ด ์ง€์‹์„ ๊ฐ€๋ฅด์น˜๋Š” ๊ฐ€์žฅ ์ข‹์€ ๋ฐฉ๋ฒ•์€ ๋ฌด์—‡์ผ๊นŒ์š”?[|endofturn|]
์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๋ชจ๋ธ์˜ ์ง€์‹๊ณผ ์‚ฌ์‹ค ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ์–ตํ•˜๊ณ  ์ดํ•ดํ•˜๋ฉฐ ๊ฒ€์ƒ‰ํ•˜๋Š” ๋Šฅ๋ ฅ์— ๋Œ€ํ•œ ํ‰๊ฐ€์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” _์ง€์‹ ์ฃผ์ž…_(Wang et al., 2020; Chen et al., 2022; Liu et al., 2020; Lauscher et al., 2020)์˜ ๊ฐœ๋…์„ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ํ…์ŠคํŠธ ์ฝ”ํผ์Šค ํ˜•ํƒœ์˜ ์ง€์‹ ๋ฒ ์ด์Šค๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์— ์ด ์ง€์‹์„ ๊ฐ€๋ฅด์น˜๋Š” ๊ฐ€์žฅ ์ข‹์€ ๋ฐฉ๋ฒ•์€ ๋ฌด์—‡์ผ๊นŒ์š”?[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this work, we focus on the evaluation of a model's knowledge and its ability to memorize, understand, and retrieve factual data. We aim to understand the concept of _knowledge injection_(Wang et al., 2020; Chen et al., 2022; Liu et al., 2020; Lauscher et al., 2020). Given some knowledge base in the form of a text corpus, what is the best way to teach a pre-trained model this knowledge? [|assistant|]
1. ์‚ฌ์šฉ ์‚ฌ๋ก€, ํ…Œ์ŠคํŠธ ๋Œ€์ƒ ์‹œ์Šคํ…œ(SUT) ์œ ํ˜•, ์–ธ์–ด ๋ฐ ์ปจํ…์ŠคํŠธ, ํŽ˜๋ฅด์†Œ๋‚˜, ํ…Œ์ŠคํŠธ ๋ฐ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ(2์žฅ ์ฐธ์กฐ)์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ง€์ •ํ•˜๊ณ  ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ์›์น™์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹. 2. ์ •์˜ ๋ฐ ํ•˜์œ„ ๋ฒ”์ฃผ๊ฐ€ ์žˆ๋Š” 13๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ์— ๋Œ€ํ•œ ๋ถ„๋ฅ˜(์„น์…˜ 3 ์ฐธ์กฐ). 3. 7๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ์— ๋Œ€ํ•œ ํ…Œ์ŠคํŠธ, ๊ฐ๊ฐ ๊ณ ์œ ํ•œ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ(์ฆ‰, prompts)์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค(์„น์…˜ 4 ์ฐธ์กฐ). ์ด 43,090๊ฐœ์˜ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ…œํ”Œ๋ฆฟ์œผ๋กœ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. 4. ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ AI ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ์ฑ„์  ์‹œ์Šคํ…œ์€ ๊ฐœ๋ฐฉ์ ์ด๊ณ  ์„ค๋ช… ๊ฐ€๋Šฅํ•˜๋ฉฐ ๋‹ค์–‘ํ•œ ์‚ฌ์šฉ ์‚ฌ๋ก€์— ๋งž๊ฒŒ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์„น์…˜ 5 ์ฐธ์กฐ). 5. ๋ฒค์น˜๋งˆํฌ์—์„œ AI ์‹œ์Šคํ…œ์˜ ์•ˆ์ „์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ”Œ๋žซํผ ๋ฐ ๋‹ค์šด๋กœ๋“œ ๊ฐ€๋Šฅํ•œ ๋„๊ตฌ์ธ **ModelBench**๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.2 ๊ฐ์ฃผ 2: [https://github.com/mlcommons/modelbench](https://github.com/mlcommons/modelbench)6. 6. ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ 12๊ฐœ ์ด์ƒ์˜ ์ฑ„ํŒ… ํŠœ๋‹ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋ฒค์น˜๋งˆํ‚นํ•œ ํ‰๊ฐ€ ๋ณด๊ณ ์„œ ์˜ˆ์‹œ์ž…๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์€ ์ต๋ช…ํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค(์„น์…˜ 6 ์ฐธ์กฐ).[|endofturn|]
1. ์‚ฌ์šฉ ์‚ฌ๋ก€, ํ…Œ์ŠคํŠธ ๋Œ€์ƒ ์‹œ์Šคํ…œ(SUT) ์œ ํ˜•, ์–ธ์–ด ๋ฐ ์ปจํ…์ŠคํŠธ, ํŽ˜๋ฅด์†Œ๋‚˜, ํ…Œ์ŠคํŠธ ๋ฐ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ(2์žฅ ์ฐธ์กฐ)์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ง€์ •ํ•˜๊ณ  ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ์›์น™์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹. 2. ์ •์˜ ๋ฐ ํ•˜์œ„ ๋ฒ”์ฃผ๊ฐ€ ์žˆ๋Š” 13๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ์— ๋Œ€ํ•œ ๋ถ„๋ฅ˜(์„น์…˜ 3 ์ฐธ์กฐ). 3. 7๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ์— ๋Œ€ํ•œ ํ…Œ์ŠคํŠธ, ๊ฐ๊ฐ ๊ณ ์œ ํ•œ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ(์ฆ‰, ํ”„๋กฌํ”„ํŠธ)์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค(์„น์…˜ 4 ์ฐธ์กฐ). ์ด 43,090๊ฐœ์˜ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ…œํ”Œ๋ฆฟ์œผ๋กœ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. 4. ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ AI ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ์ฑ„์  ์‹œ์Šคํ…œ์€ ๊ฐœ๋ฐฉ์ ์ด๊ณ  ์„ค๋ช… ๊ฐ€๋Šฅํ•˜๋ฉฐ ๋‹ค์–‘ํ•œ ์‚ฌ์šฉ ์‚ฌ๋ก€์— ๋งž๊ฒŒ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์„น์…˜ 5 ์ฐธ์กฐ). 5. ๋ฒค์น˜๋งˆํฌ์—์„œ AI ์‹œ์Šคํ…œ์˜ ์•ˆ์ „์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ”Œ๋žซํผ ๋ฐ ๋‹ค์šด๋กœ๋“œ ๊ฐ€๋Šฅํ•œ ๋„๊ตฌ์ธ **ModelBench**๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.2 ๊ฐ์ฃผ 2: [https://github.com/mlcommons/modelbench](https://github.com/mlcommons/modelbench)6. 6. ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ 12๊ฐœ ์ด์ƒ์˜ ์ฑ„ํŒ… ํŠœ๋‹ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋ฒค์น˜๋งˆํ‚นํ•œ ํ‰๊ฐ€ ๋ณด๊ณ ์„œ ์˜ˆ์‹œ์ž…๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์€ ์ต๋ช…ํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค(์„น์…˜ 6 ์ฐธ์กฐ).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]1. A principled approach to specifying and constructing the benchmark, which comprises use cases, types of systems under test (SUTs), language and context, personas, tests, and test items (see Section 2). 2. A taxonomy of 13 hazard categories with definitions and subcategories (see Section 3). 3. Tests for seven of the hazard categories, each comprising a unique set of test items, i.e., prompts (see Section 4). There are 43,090 test items in total, which we created with templates. 4. A grading system for AI systems against the benchmark that is open, explainable, and can be adjusted for a range of use cases (see Section 5). 5. An openly available platform, and downloadable tool, called **ModelBench** that can be used to evaluate the safety of AI systems on the benchmark.2 Footnote 2: [https://github.com/mlcommons/modelbench](https://github.com/mlcommons/modelbench) 6. An example evaluation report which benchmarks the performance of over a dozen openly available chat-tuned language models. All models have been anonymized (see Section 6). [|assistant|]
**๋ฉ”๋ชจ๋ฆฌ ๊ตฌ์กฐ**: LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„์˜ ๊ธฐ์–ต ๊ณผ์ •์— ๋Œ€ํ•œ ์ธ์ง€ ๊ณผํ•™ ์—ฐ๊ตฌ์—์„œ ํŒŒ์ƒ๋œ ์›๋ฆฌ์™€ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ธ๊ฐ„์˜ ๊ธฐ์–ต์€ ์ง€๊ฐ ์ž…๋ ฅ์„ ๋“ฑ๋กํ•˜๋Š” ๊ฐ๊ฐ ๊ธฐ์–ต์—์„œ ์ผ์‹œ์ ์œผ๋กœ ์ •๋ณด๋ฅผ ์œ ์ง€ํ•˜๋Š” ๋‹จ๊ธฐ๊ธฐ์–ต, ์žฅ๊ธฐ๊ฐ„์— ๊ฑธ์ณ ์ •๋ณด๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ์žฅ๊ธฐ๊ธฐ์–ต์œผ๋กœ ์ผ๋ฐ˜์ ์œผ๋กœ ์ง„ํ–‰๋ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ž๋“ค์€ ์—์ด์ „ํŠธ ๋ฉ”๋ชจ๋ฆฌ ๊ตฌ์กฐ๋ฅผ ์„ค๊ณ„ํ•  ๋•Œ ์ด๋Ÿฌํ•œ ์ธ๊ฐ„ ๊ธฐ์–ต์˜ ์ธก๋ฉด์—์„œ ์˜๊ฐ์„ ์–ป์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ๋‹จ๊ธฐ๊ธฐ์–ต์€ Transformer ์•„ํ‚คํ…์ฒ˜์— ์˜ํ•ด ์ œํ•œ๋œ ์ปจํ…์ŠคํŠธ ์ฐฝ ๋‚ด์˜ ์ž…๋ ฅ ์ •๋ณด์™€ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์žฅ๊ธฐ๊ธฐ์–ต์€ ์—์ด์ „ํŠธ๊ฐ€ ํ•„์š”์— ๋”ฐ๋ผ ๋น ๋ฅด๊ฒŒ ์ฟผ๋ฆฌํ•˜๊ณ  ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋Š” ์™ธ๋ถ€ ๋ฒกํ„ฐ ์ €์žฅ์†Œ์™€ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ๋‹จ๊ธฐ๊ธฐ์–ต๊ณผ ์žฅ๊ธฐ๊ธฐ์–ต์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๋‘ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ฉ”๋ชจ๋ฆฌ ๊ตฌ์กฐ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**๋ฉ”๋ชจ๋ฆฌ ๊ตฌ์กฐ**: LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„์˜ ๊ธฐ์–ต ๊ณผ์ •์— ๋Œ€ํ•œ ์ธ์ง€ ๊ณผํ•™ ์—ฐ๊ตฌ์—์„œ ํŒŒ์ƒ๋œ ์›๋ฆฌ์™€ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ธ๊ฐ„์˜ ๊ธฐ์–ต์€ ์ง€๊ฐ ์ž…๋ ฅ์„ ๋“ฑ๋กํ•˜๋Š” ๊ฐ๊ฐ ๊ธฐ์–ต์—์„œ ์ผ์‹œ์ ์œผ๋กœ ์ •๋ณด๋ฅผ ์œ ์ง€ํ•˜๋Š” ๋‹จ๊ธฐ๊ธฐ์–ต, ์žฅ๊ธฐ๊ฐ„์— ๊ฑธ์ณ ์ •๋ณด๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ์žฅ๊ธฐ๊ธฐ์–ต์œผ๋กœ ์ผ๋ฐ˜์ ์œผ๋กœ ์ง„ํ–‰๋ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ž๋“ค์€ ์—์ด์ „ํŠธ ๋ฉ”๋ชจ๋ฆฌ ๊ตฌ์กฐ๋ฅผ ์„ค๊ณ„ํ•  ๋•Œ ์ด๋Ÿฌํ•œ ์ธ๊ฐ„ ๊ธฐ์–ต์˜ ์ธก๋ฉด์—์„œ ์˜๊ฐ์„ ์–ป์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ๋‹จ๊ธฐ๊ธฐ์–ต์€ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜์— ์˜ํ•ด ์ œํ•œ๋œ ์ปจํ…์ŠคํŠธ ์ฐฝ ๋‚ด์˜ ์ž…๋ ฅ ์ •๋ณด์™€ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์žฅ๊ธฐ๊ธฐ์–ต์€ ์—์ด์ „ํŠธ๊ฐ€ ํ•„์š”์— ๋”ฐ๋ผ ๋น ๋ฅด๊ฒŒ ์ฟผ๋ฆฌํ•˜๊ณ  ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋Š” ์™ธ๋ถ€ ๋ฒกํ„ฐ ์ €์žฅ์†Œ์™€ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ๋‹จ๊ธฐ๊ธฐ์–ต๊ณผ ์žฅ๊ธฐ๊ธฐ์–ต์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๋‘ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ฉ”๋ชจ๋ฆฌ ๊ตฌ์กฐ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Memory Structures**: LLM-based autonomous agents usually incorporate principles and mechanisms derived from cognitive science research on human memory processes. Human memory follows a general progression from sensory memory that registers perceptual inputs, to short-term memory that maintains information transiently, to long-term memory that consolidates information over extended periods. When designing the agent memory structures, researchers take inspiration from these aspects of human memory. In specific, short-term memory is analogous to the input information within the context window constrained by the transformer architecture. Long-term memory resembles the external vector storage that agents can rapidly query and retrieve from as needed. In the following, we introduce two commonly used memory structures based on the short- and long-term memories. [|assistant|]
๋‘ ๊ฐ€์ง€ ๊ธฐ๋ฒ•์„ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜์—ฌ ๋„ค ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐ๊ฐ์„ ์ค‘๋ณต ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ๋ฐ์ดํ„ฐ ๋ถ„ํ• ์—์„œ ํ…์ŠคํŠธ๊ฐ€ ์ค‘๋ณต๋˜๋Š” ๊ฒฝ์šฐ ํ…Œ์ŠคํŠธ ๋˜๋Š” ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์— ์‚ฌ๋ณธ์„ ์œ ์ง€ํ•˜๊ณ  train ์ง‘ํ•ฉ์—์„œ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์„ ์šฐ์„ ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋‘ ๊ฐ€์ง€ ๊ธฐ๋ฒ•์„ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜์—ฌ ๋„ค ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐ๊ฐ์„ ์ค‘๋ณต ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ๋ฐ์ดํ„ฐ ๋ถ„ํ• ์—์„œ ํ…์ŠคํŠธ๊ฐ€ ์ค‘๋ณต๋˜๋Š” ๊ฒฝ์šฐ ํ…Œ์ŠคํŠธ ๋˜๋Š” ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์— ์‚ฌ๋ณธ์„ ์œ ์ง€ํ•˜๊ณ  ํ›ˆ๋ จ ์ง‘ํ•ฉ์—์„œ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์„ ์šฐ์„ ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We deduplicate each of the four datasets with both of our two techniques. When text was duplicated across multiple data splits, we prioritized keeping a copy in the test or validation set and removing it from the train set. [|assistant|]
ํ…์ŠคํŠธ ์ƒ์„ฑ์€ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ๊ฐ€์žฅ ๊ธฐ๋ณธ์ ์ธ ๊ธฐ๋Šฅ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. Alpaca-Plus-7B์™€ Alpaca-Plus-13B์— ๋น„ํ•ด Alpaca-33B๋Š” ์ด ์นดํ…Œ๊ณ ๋ฆฌ์—์„œ ์—ด๋“ฑํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 5๋Š” ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…์˜ ์˜ˆ์ž…๋‹ˆ๋‹ค. Alpaca-Plus-7B์™€ Alpaca-Plus-13B ๋ชจ๋‘ ์‚ฌ์šฉ์ž prompt์˜ ์š”๊ตฌ ์‚ฌํ•ญ์„ ์ถฉ์กฑํ•˜๋Š” ์˜ฌ๋ฐ”๋ฅธ ๊ธ€์ž ์Šคํƒ€์ผ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Alpaca-Plus-13B๋Š” ์‹ ์ฒญ์ž๊ฐ€ ๋น„์ž ์‹ ์ฒญ์— ํ•„์š”ํ•œ ๋ชจ๋“  ์ž๋ฃŒ๋ฅผ ์ฒ ์ €ํžˆ ์ค€๋น„ํ–ˆ์Œ์„ ๋‚˜ํƒ€๋‚ด์–ด ๊ฐ€์žฅ ํฌ๊ด„์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์„ธ ์‹œ์Šคํ…œ ์ค‘ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์ƒ์„ฑ ํ’ˆ์งˆ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Alpaca-33B๋Š” ํŽธ์ง€ ์Šคํƒ€์ผ์„ ๋”ฐ๋ฅด์ง€ ์•Š๊ณ  ๋‚ด์šฉ์ด ๋‹ค์†Œ ๋‹จ์ˆœํ™”๋˜์–ด ์žˆ์–ด ๋‹ค๋ฅธ ์‹œ์Šคํ…œ์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์ด ๋ถ„๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋” ์ž‘์€ ๋ชจ๋ธ๋กœ ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ด ๋” ์ ์€ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ํฐ ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
ํ…์ŠคํŠธ ์ƒ์„ฑ์€ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ๊ฐ€์žฅ ๊ธฐ๋ณธ์ ์ธ ๊ธฐ๋Šฅ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์•ŒํŒŒ์นด ํ”Œ๋Ÿฌ์Šค-7B์™€ ์•ŒํŒŒ์นด ํ”Œ๋Ÿฌ์Šค-13B์— ๋น„ํ•ด ์•ŒํŒŒ์นด-33B๋Š” ์ด ์นดํ…Œ๊ณ ๋ฆฌ์—์„œ ์—ด๋“ฑํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 5๋Š” ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…์˜ ์˜ˆ์ž…๋‹ˆ๋‹ค. ์•ŒํŒŒ์นด ํ”Œ๋Ÿฌ์Šค-7B์™€ ์•ŒํŒŒ์นด ํ”Œ๋Ÿฌ์Šค-13B ๋ชจ๋‘ ์‚ฌ์šฉ์ž ํ”„๋กฌํ”„ํŠธ์˜ ์š”๊ตฌ ์‚ฌํ•ญ์„ ์ถฉ์กฑํ•˜๋Š” ์˜ฌ๋ฐ”๋ฅธ ๊ธ€์ž ์Šคํƒ€์ผ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์•ŒํŒŒ์นด ํ”Œ๋Ÿฌ์Šค-13B๋Š” ์‹ ์ฒญ์ž๊ฐ€ ๋น„์ž ์‹ ์ฒญ์— ํ•„์š”ํ•œ ๋ชจ๋“  ์ž๋ฃŒ๋ฅผ ์ฒ ์ €ํžˆ ์ค€๋น„ํ–ˆ์Œ์„ ๋‚˜ํƒ€๋‚ด์–ด ๊ฐ€์žฅ ํฌ๊ด„์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ์„ธ ์‹œ์Šคํ…œ ์ค‘ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์ƒ์„ฑ ํ’ˆ์งˆ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์•ŒํŒŒ์นด-33B๋Š” ํŽธ์ง€ ์Šคํƒ€์ผ์„ ๋”ฐ๋ฅด์ง€ ์•Š๊ณ  ๋‚ด์šฉ์ด ๋‹ค์†Œ ๋‹จ์ˆœํ™”๋˜์–ด ์žˆ์–ด ๋‹ค๋ฅธ ์‹œ์Šคํ…œ์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์ด ๋ถ„๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋” ์ž‘์€ ๋ชจ๋ธ๋กœ ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ด ๋” ์ ์€ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ํฐ ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Text generation is one of the most fundamental abilities for language models. Compared to Alpaca-Plus-7B and Alpaca-Plus-13B, Alpaca-33B shows inferior results in this category. Table 5 shows an example of a text generation task. We can see that both Alpaca-Plus-7B and Alpaca-Plus-13B provide correct letter styles, which meet the requirement of the user's prompt. Alpaca-Plus-13B provides the most comprehensive one by indicating that the applicant has thoroughly prepared all materials for visa application, making it the best generation quality among all three systems. However, Alpaca-33B does not follow the letter style, and the content is somewhat too simplified, which is clearly not as good as the others. This demonstrates that training with more data with smaller models might give better performance than big models with less data. [|assistant|]
LLaMA ์ œํ’ˆ๊ตฐ์ด ๋น ๋ฅด๊ฒŒ ์„ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋” ๋งŽ์€ instruction-following ๋ชจ๋ธ์ด LLaMA ๋˜๋Š” LLaMA-2๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์ถ•๋˜์—ˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์—๋Š” Code LLaMA [66], Gorilla [67], Giraffe [68], Vigogne [69], Tulu 65B [70], Long LLaMA [71], Stable Beluga2 [72] ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ฝ”๋“œ LLaMA [66], ๊ณ ๋ฆด๋ผ [67], ๊ธฐ๋ฆฐ [68], ๋น„๊ณ ๋„ค [69], ํˆด๋ฃจ 65B [70], ๋กฑ LLaMA [71], ์Šคํ…Œ์ด๋ธ” ๋ฒจ๋ฃจ๊ฐ€2 [72] ๋“ฑ LLaMA ๋˜๋Š” LLaMA-2๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋” ๋งŽ์€ ๋ช…๋ น์–ด ์ถ”์ข… ๋ชจ๋ธ์ด ๊ตฌ์ถ•๋˜๋ฉด์„œ LLaMA ์ œํ’ˆ๊ตฐ์ด ๋น ๋ฅด๊ฒŒ ์„ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The LLaMA family is growing rapidly, as more instruction-following models have been built on LLaMA or LLaMA-2, including Code LLaMA [66], Gorilla [67], Giraffe [68], Vigogne [69], Tulu 65B [70], Long LLaMA [71], and Stable Beluga2 [72], just to name a few. [|assistant|]
**์ถœ๋ ฅ**[**Retriee**]=Yes]<p>์›Œํ‚น ๋ฐ๋“œ (์‹œ์ฆŒ 7)-์›Œํ‚น ๋ฐ๋“œ (์‹œ์ฆŒ 7) AMC์˜ ๋ฏธ๊ตญ ํฌ์ŠคํŠธ ์•„ํฌ์นผ๋ฆฝ์Šค ๊ณตํฌ ํ…”๋ ˆ๋น„์ „ ์‹œ๋ฆฌ์ฆˆ์ธ "์›Œํ‚น ๋ฐ๋“œ"์˜ ์ผ๊ณฑ ๋ฒˆ์งธ ์‹œ์ฆŒ์€ 2016๋…„ 10์›” 23์ผ์— ์ฒซ ๋ฐฉ์†ก๋˜์–ด 2017๋…„ 4์›” 2์ผ์— 16๊ฐœ์˜ ์—ํ”ผ์†Œ๋“œ๋กœ ๊ตฌ์„ฑ๋œ ์—ํ”ผ์†Œ๋“œ๊ฐ€ ๋ฐฉ์˜๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ”„๋žญํฌ ๋‹ค๋ผ๋ฐ”์›ƒ์ด ํ…”๋ ˆ๋น„์ „์šฉ์œผ๋กœ ๊ฐœ๋ฐœํ•œ ์ด ์‹œ๋ฆฌ์ฆˆ๋Š” ๋กœ๋ฒ„ํŠธ ์ปคํฌ๋จผ, ํ† ๋‹ˆ ๋ฌด์–ด, ์ฐฐ๋ฆฌ ์• ๋“ค๋ผ๋“œ์˜ ์‹œ์กฐ ๋งŒํ™”๋ฅผ ์›์ž‘์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**์ถœ๋ ฅ**[**Retriee**]=Yes]<p>์›Œํ‚น ๋ฐ๋“œ (์‹œ์ฆŒ 7)-์›Œํ‚น ๋ฐ๋“œ (์‹œ์ฆŒ 7) AMC์˜ ๋ฏธ๊ตญ ํฌ์ŠคํŠธ ์•„ํฌ์นผ๋ฆฝํŠธ ๊ณตํฌ ํ…”๋ ˆ๋น„์ „ ์‹œ๋ฆฌ์ฆˆ์ธ "์›Œํ‚น ๋ฐ๋“œ"์˜ ์ผ๊ณฑ ๋ฒˆ์งธ ์‹œ์ฆŒ์€ 2016๋…„ 10์›” 23์ผ์— ์ฒซ ๋ฐฉ์†ก๋˜์–ด 2017๋…„ 4์›” 2์ผ์— 16๊ฐœ์˜ ์—ํ”ผ์†Œ๋“œ๋กœ ๊ตฌ์„ฑ๋œ ์—ํ”ผ์†Œ๋“œ๊ฐ€ ๋ฐฉ์˜๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ”„๋žญํฌ ๋‹ค๋ผ๋ฐ”์›ƒ์ด ํ…”๋ ˆ๋น„์ „์šฉ์œผ๋กœ ๊ฐœ๋ฐœํ•œ ์ด ์‹œ๋ฆฌ์ฆˆ๋Š” ๋กœ๋ฒ„ํŠธ ์ปคํฌ๋จผ, ํ† ๋‹ˆ ๋ฌด์–ด, ์ฐฐ๋ฆฌ ์• ๋“ค๋ผ๋“œ์˜ ์‹œ์กฐ ๋งŒํ™”๋ฅผ ์›์ž‘์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Output**[**Retriee**]=Yes]<p>The Walking Dead (season 7)-The Walking Dead (season 7) The seventh season of "The Walking Dead", an American post-apocalypetric horror television series on AMC, premiered on October 23, 2016, and concluded on April 2, 2017, consisting of 16 episodes. Developed for television by Frank Darabout, the series is based on the eponymous series of comic books by Robert Kirkman, Tony Moore, and Charlie Adlard. [|assistant|]
์—ฌ๊ธฐ์„œ \(i\)๋Š” ํ˜„์žฌ ์‚ฌ์ „ ํ•™์Šต ์ค‘์ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ธ๋ฑ์Šค์ด๊ณ , \(\gamma_{j}=0\)์€ \(j=0\)์ธ ๊ฒฝ์šฐ, ๊ทธ๋ ‡์ง€ ์•Š์€ ๊ฒฝ์šฐ \(1\)์ž…๋‹ˆ๋‹ค. ํ›„์ž๋Š” replay๊ฐ€ ์‚ฌ์šฉ๋˜์ง€ ์•Š๋Š” ์ฒซ ๋ฒˆ์งธ ๋ฐ์ดํ„ฐ ์„ธํŠธ \(\mathcal{D}_{0}\)๋ฅผ ์ œ์™ธํ•˜๊ณ ๋Š” compute-equivalent replay๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— \(\mathcal{D}_{i}\)์˜ \(s_{i}\cdot(1-\alpha)\) ํ† ํฐ๋งŒ ๋ณผ ์ˆ˜ ์žˆ๊ณ , ๋”ฐ๋ผ์„œ \(\mathcal{D}_{0}\)์˜ ๋ชจ๋“  \(s_{0}\) ํ† ํฐ์„ ๋ณผ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.[|endofturn|]
์—ฌ๊ธฐ์„œ \(i\)๋Š” ํ˜„์žฌ ์‚ฌ์ „ ํ•™์Šต ์ค‘์ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ธ๋ฑ์Šค์ด๊ณ , \(\gamma_{j}=0\)์€ \(j=0\)์ธ ๊ฒฝ์šฐ, ๊ทธ๋ ‡์ง€ ์•Š์€ ๊ฒฝ์šฐ \(1\)์ž…๋‹ˆ๋‹ค. ํ›„์ž๋Š” ๋ฆฌํ”Œ๋ ˆ์ด๊ฐ€ ์‚ฌ์šฉ๋˜์ง€ ์•Š๋Š” ์ฒซ ๋ฒˆ์งธ ๋ฐ์ดํ„ฐ ์„ธํŠธ \(\mathcal{D}_{0}\)๋ฅผ ์ œ์™ธํ•˜๊ณ ๋Š” ๊ณ„์‚ฐ ๋“ฑ๊ฐ€ ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— \(\mathcal{D}_{i}\)์˜ \(s_{i}\cdot(1-\alpha)\) ํ† ํฐ๋งŒ ๋ณผ ์ˆ˜ ์žˆ๊ณ , ๋”ฐ๋ผ์„œ \(\mathcal{D}_{0}\)์˜ ๋ชจ๋“  \(s_{0}\) ํ† ํฐ์„ ๋ณผ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]where \(i\) is the index of the dataset on which we are currently pretraining, and \(\gamma_{j}=0\) if \(j=0\), and \(1\) otherwise. The latter is because when pretraining on \(\mathcal{D}_{i}\), we only see \(s_{i}\cdot(1-\alpha)\) tokens of \(\mathcal{D}_{i}\) because we use compute-equivalent replay, except for the first dataset \(\mathcal{D}_{0}\) where replay is not used, and where we hence see all \(s_{0}\) tokens of \(\mathcal{D}_{0}\). [|assistant|]
์—ฌ์ „ํžˆ SSM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•  ์—ฌ์ง€๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ImageNet [2]๊ณผ ๊ฐ™์€ ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด pre-trained๋œ SSM์€ ๋งŽ์€ downstream ์ž‘์—…์—์„œ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜์ง€๋งŒ, SSM์˜ base, large ๋ฐ huge ๋ฒ„์ „์€ ๊ฑฐ์˜ ๊ณต๊ฐœ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Š” CV ์ž‘์—…์˜ ๋†’์€ ์„ฑ๋Šฅ์— ์žฅ์• ๊ฐ€ ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์—ฌ์ „ํžˆ SSM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•  ์—ฌ์ง€๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ImageNet [2]๊ณผ ๊ฐ™์€ ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต๋œ SSM์€ ๋งŽ์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜์ง€๋งŒ, SSM์˜ ๊ธฐ๋ณธ, ๋Œ€ํ˜• ๋ฐ ๊ฑฐ๋Œ€ ๋ฒ„์ „์€ ๊ฑฐ์˜ ๊ณต๊ฐœ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Š” CV ์ž‘์—…์˜ ๋†’์€ ์„ฑ๋Šฅ์— ์žฅ์• ๊ฐ€ ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]still room for performance improvement based on SSM. The SSMs pre-trained on the large-scale dataset, such as ImageNet [2], play a critical role in many downstream tasks, however, the base, large, and huge versions of SSMs are rarely released. We believe this may be an obstacle to the high performance on the CV tasks. [|assistant|]
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ์—๋Š” ๊ฐ€์žฅ ์ค‘์š”ํ•œ ํ•œ๊ณ„, ์ฆ‰ ํ…์ŠคํŠธ ์ฑ„ํŒ… ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์‚ฌ์šฉํ–ˆ๋‹ค๋Š” ์ ์ด ์žˆ๋‹ค. ์ด๋Š” ์ž ์žฌ์ ์œผ๋กœ ํ™˜์ž์™€ ์ง„๋‹จ ๋Œ€ํ™” ์ „๋ฌธ Large Language Models (LLMs) ๊ฐ„์˜ ๋Œ€๊ทœ๋ชจ ์ƒํ˜ธ ์ž‘์šฉ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜์ง€๋งŒ, ์›๊ฒฉ ์ƒ๋‹ด์„ ์œ„ํ•ด PCP์—๊ฒŒ๋Š” ์ต์ˆ™ํ•˜์ง€ ์•Š์€ ๊ฒƒ์ด๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” (์›๊ฒฉ) ์˜๋ฃŒ์˜ ์ผ๋ฐ˜์ ์ธ ๊ด€ํ–‰์„ ๋Œ€ํ‘œํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋˜์–ด์„œ๋Š” ์•ˆ ๋œ๋‹ค.[|endofturn|]
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ์—๋Š” ๊ฐ€์žฅ ์ค‘์š”ํ•œ ํ•œ๊ณ„, ์ฆ‰ ํ…์ŠคํŠธ ์ฑ„ํŒ… ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์‚ฌ์šฉํ–ˆ๋‹ค๋Š” ์ ์ด ์žˆ๋‹ค. ์ด๋Š” ์ž ์žฌ์ ์œผ๋กœ ํ™˜์ž์™€ ์ง„๋‹จ ๋Œ€ํ™” ์ „๋ฌธ LLM ๊ฐ„์˜ ๋Œ€๊ทœ๋ชจ ์ƒํ˜ธ ์ž‘์šฉ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜์ง€๋งŒ, ์›๊ฒฉ ์ƒ๋‹ด์„ ์œ„ํ•ด PCP์—๊ฒŒ๋Š” ์ต์ˆ™ํ•˜์ง€ ์•Š์€ ๊ฒƒ์ด๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” (์›๊ฒฉ) ์˜๋ฃŒ์˜ ์ผ๋ฐ˜์ ์ธ ๊ด€ํ–‰์„ ๋Œ€ํ‘œํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋˜์–ด์„œ๋Š” ์•ˆ ๋œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Our research has important limitations, most notably that we utilized a text-chat interface, which although enabling potentially large-scale interaction between patients and LLMs specialized for diagnostic dialogue, was unfamiliar to PCPs for remote consultation. Thus our study should not be regarded as representative of usual practice in (tele)medicine. [|assistant|]
๊ทธ๋ฆผ 10: WSD Scheduler๋ฅผ ์‚ฌ์šฉํ•œ scaling ์‹คํ—˜์˜ ์ ํ•ฉ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ์ˆ˜ํ‰์„ ์˜ ๊ฒ€์€ ์ ์€ ๋™์ผํ•œ ๋ชจ๋ธ ํฌ๊ธฐ ๋‚ด์—์„œ ์„œ๋กœ ๋‹ค๋ฅธ ์ปดํ“จํŒ…์—์„œ ๊ฐ์‡ ๋œ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 10: WSD ์Šค์ผ€์ค„๋Ÿฌ๋ฅผ ์‚ฌ์šฉํ•œ ํ™•์žฅ ์‹คํ—˜์˜ ์ ํ•ฉ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ์ˆ˜ํ‰์„ ์˜ ๊ฒ€์€ ์ ์€ ๋™์ผํ•œ ๋ชจ๋ธ ํฌ๊ธฐ ๋‚ด์—์„œ ์„œ๋กœ ๋‹ค๋ฅธ ์ปดํ“จํŒ…์—์„œ ๊ฐ์‡ ๋œ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 10: The fit result of the scaling experiment with WSD Scheduler. The black dots in a horizontal line denotes the decayed checkpoints in different compute within a same model size. [|assistant|]
์˜๋ฃŒ ๋Œ€ํ™”์—์„œ ํ™˜์ž ๋˜๋Š” ์˜์‚ฌ ์—ญํ• ์„ ์ˆ˜ํ–‰ํ•˜๊ณ , ์˜๋ฃŒ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ณ  ์ถ”๋ก ํ•˜๊ณ , EHR ๋…ธํŠธ๋ฅผ ์š”์•ฝํ•˜๋Š” ๋“ฑ์˜ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ์ž‘์—…๋ณ„ instructions๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ AMIE๋ฅผ Fine-tuningํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ๋ณธ LLM์˜ ์ฒซ ๋ฒˆ์งธ Fine-tuning์€ ์ •์ ์ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋งŒ ์‚ฌ์šฉํ•œ ๋ฐ˜๋ฉด, ์ดํ›„์˜ Fine-tuning ๋‹จ๊ณ„์—์„œ๋Š” 2.2.1์ ˆ์—์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด ์…€ํ”„ ํ”Œ๋ ˆ์ด ๋‚ด๋ถ€ ๋ฃจํ”„๋ฅผ ํ†ตํ•ด ์ƒ์„ฑ๋œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™”๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์˜๋ฃŒ ๋Œ€ํ™”์—์„œ ํ™˜์ž ๋˜๋Š” ์˜์‚ฌ ์—ญํ• ์„ ์ˆ˜ํ–‰ํ•˜๊ณ , ์˜๋ฃŒ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ณ  ์ถ”๋ก ํ•˜๊ณ , EHR ๋…ธํŠธ๋ฅผ ์š”์•ฝํ•˜๋Š” ๋“ฑ์˜ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ์ž‘์—…๋ณ„ ์ง€์นจ์„ ์‚ฌ์šฉํ•˜์—ฌ AMIE๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ๋ณธ LLM์˜ ์ฒซ ๋ฒˆ์งธ ๋ฏธ์„ธ ์กฐ์ •์€ ์ •์ ์ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋งŒ ์‚ฌ์šฉํ•œ ๋ฐ˜๋ฉด, ์ดํ›„์˜ ๋ฏธ์„ธ ์กฐ์ • ๋‹จ๊ณ„์—์„œ๋Š” 2.2.1์ ˆ์—์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด ์…€ํ”„ ํ”Œ๋ ˆ์ด ๋‚ด๋ถ€ ๋ฃจํ”„๋ฅผ ํ†ตํ•ด ์ƒ์„ฑ๋œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™”๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We employed task-specific instructions to fine-tune AMIE in playing either the patient or doctor role within medical dialogues, performing medical question answering and reasoning, and summarizing EHR notes. While the first round of fine-tuning from the base LLM only used the static datasets, subsequent rounds of fine-tuning leveraged the simulated dialogues generated through the self-play inner loop as described in Section 2.2.1. [|assistant|]
**Instruction Tuning for LLM Alignment.** ๋ช…๋ น์„ ์ถฉ์‹คํžˆ ๋”ฐ๋ฅด๊ณ  ๋‹ค์–‘ํ•œ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž์ถฐ LLM์„ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์ค‘์š”ํ•œ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค(Efrat and Levy, 2020). ์ดˆ๊ธฐ ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ๋‹ค์–‘ํ•œ ๊ณต๊ฐœ NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ชจ๋ธ์„ Fine-tuningํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ต์ฐจ ์ž‘์—… ์ผ๋ฐ˜ํ™”์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์Šต๋‹ˆ๋‹ค(Raffel et al., 2020; Wei et al., 2021; Aribandi et al., 2021; Victor et al., 2022; Chung et al., 2022). ์ตœ๊ทผ์—๋Š” ๋” ๊ด‘๋ฒ”์œ„ํ•œ ํ˜•์‹๊ณผ ์ž‘์—… ์œ ํ˜•์„ ํŠน์ง•์œผ๋กœ ํ•˜๋Š” ์˜คํ”ˆ ๋„๋ฉ”์ธ์— Instruction Tuning์„ ํ™•์žฅํ•œ ์—ฐ๊ตฌ์ž๋“ค์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ณ€ํ™”๋Š” ํฌ๋ผ์šฐ๋“œ์†Œ์‹ฑ์„ ํ†ตํ•ด ์ธ๊ฐ„์ด ์ƒ์„ฑํ•œ ๋ช…๋ น์–ด-์‘๋‹ต ์Œ(Ouyang et al., 2022; Kopf et al., 2023; Zhou et al., 2023)๊ณผ LLM์ด ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ(Taori et al., 2023; Chiang et al., 2023)๋ฅผ ํ™œ์šฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ CodecLM์€ Instruction metadata ๊ฐœ๋…์„ ํ™œ์šฉํ•˜์—ฌ ์ธ๊ฐ„์˜ ์ฃผ์„ ์—†์ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํŠน์ • ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋งž๊ฒŒ ์กฐ์ •ํ•˜๋Š” ๊ณ ์œ ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**LLM ์ •๋ ฌ์„ ์œ„ํ•œ ๋ช…๋ น์–ด ํŠœ๋‹**. ๋ช…๋ น์„ ์ถฉ์‹คํžˆ ๋”ฐ๋ฅด๊ณ  ๋‹ค์–‘ํ•œ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž์ถฐ LLM์„ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์ค‘์š”ํ•œ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค(Efrat and Levy, 2020). ์ดˆ๊ธฐ ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ๋‹ค์–‘ํ•œ ๊ณต๊ฐœ NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ต์ฐจ ์ž‘์—… ์ผ๋ฐ˜ํ™”์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์Šต๋‹ˆ๋‹ค(Raffel et al., 2020; Wei et al., 2021; Aribandi et al., 2021; Victor et al., 2022; Chung et al., 2022). ์ตœ๊ทผ์—๋Š” ๋” ๊ด‘๋ฒ”์œ„ํ•œ ํ˜•์‹๊ณผ ์ž‘์—… ์œ ํ˜•์„ ํŠน์ง•์œผ๋กœ ํ•˜๋Š” ์˜คํ”ˆ ๋„๋ฉ”์ธ์— ๋ช…๋ น์–ด ํŠœ๋‹์„ ํ™•์žฅํ•œ ์—ฐ๊ตฌ์ž๋“ค์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ณ€ํ™”๋Š” ํฌ๋ผ์šฐ๋“œ์†Œ์‹ฑ์„ ํ†ตํ•ด ์ธ๊ฐ„์ด ์ƒ์„ฑํ•œ ๋ช…๋ น์–ด-์‘๋‹ต ์Œ(Ouyang et al., 2022; Kopf et al., 2023; Zhou et al., 2023)๊ณผ LLM์ด ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ(Taori et al., 2023; Chiang et al., 2023)๋ฅผ ํ™œ์šฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ CodecLM์€ ๋ช…๋ น ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ๊ฐœ๋…์„ ํ™œ์šฉํ•˜์—ฌ ์ธ๊ฐ„์˜ ์ฃผ์„ ์—†์ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํŠน์ • ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋งž๊ฒŒ ์กฐ์ •ํ•˜๋Š” ๊ณ ์œ ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Instruction Tuning for LLM Alignment.** Tuning LLM to faithfully follow instructions and align with diverse human preferences remains a significant challenge Efrat and Levy (2020). Early research primarily focused on cross-task generalization, where models were fine-tuned on various public NLP datasets to improve performance on diverse tasks Raffel et al. (2020); Wei et al. (2021); Aribandi et al. (2021); Victor et al. (2022); Chung et al. (2022). More recently, researchers have extended instruction tuning to open-domains, characterized by a wider range of formats and task types. This shift has been driven by crowdsourcing human-generated instruction-response pairs Ouyang et al. (2022); Kopf et al. (2023); Zhou et al. (2023) and LLM-generated data Taori et al. (2023); Chiang et al. (2023). Unlike prior work, CodecLM presents a unique approach for tailoring synthetic data to specific downstream tasks without human annotation, utilizing the concept of instruction metadata. [|assistant|]
๊ทธ๋ฆผ 12์—์„œ๋Š” SLM ๋ฐฉ์‹์œผ๋กœ ์„ ํƒ๋œ ํ† ํฐ์˜ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ๋ฅผ ์ œ์‹œํ•˜์˜€์œผ๋ฉฐ, ํŒŒ๋ž€์ƒ‰์œผ๋กœ ํ‘œ์‹œ๋œ ๋‚ด์šฉ์€ pretraining ๊ณผ์ •์—์„œ ์‹ค์ œ๋กœ ์„ ํƒ๋œ ํ† ํฐ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 12์—์„œ๋Š” SLM ๋ฐฉ์‹์œผ๋กœ ์„ ํƒ๋œ ํ† ํฐ์˜ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ๋ฅผ ์ œ์‹œํ•˜์˜€์œผ๋ฉฐ, ํŒŒ๋ž€์ƒ‰์œผ๋กœ ํ‘œ์‹œ๋œ ๋‚ด์šฉ์€ ์‚ฌ์ „ ํ•™์Šต ๊ณผ์ •์—์„œ ์‹ค์ œ๋กœ ์„ ํƒ๋œ ํ† ํฐ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In Figure 12, we present several examples of tokens selected by the SLM method, with content marked in blue indicating the tokens actually chosen during the pretraining process. [|assistant|]
kens๋ฅผ ๋Œ€ํ‘œํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์ƒˆ๋กœ์šด tokenizer๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด 1.6B tokens๋งŒ ์‚ฌ์šฉํ•˜์—ฌ ๊ฑฐ์˜ ์ ˆ๋ฐ˜์œผ๋กœ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ฐจ์ด๋Š” Phi-2 ๋ฐ EEVE-Korean-2.8 ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๊ฐ๊ฐ 5.6B tokens๊ณผ 1.6B tokens์ด ํ•„์š”ํ•œ ๊ฒฝ์šฐ ๋”์šฑ ๋‘๋“œ๋Ÿฌ์ง‘๋‹ˆ๋‹ค. tokens ๊ธธ์ด์— ๋Œ€ํ•œ ์ปดํ“จํŒ… ๋ณต์žก์„ฑ์ด ์ด์ฐจ์ ์œผ๋กœ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•  ๋•Œ, ์ด๋Š” ๋‘ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์˜๋ฏธ๋กœ ํ•ด์„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋™์ผํ•œ GPU์—์„œ 4๋ฐฐ ์ด์ƒ ๊ธด ์‹œํ€€์Šค๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜๋Š” ๋‘˜์งธ, ๋™์ผํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฑฐ์˜ 4๋ฐฐ ๋” ๊ณ„์‚ฐ ํšจ์œจ์ ์œผ๋กœ ๋ชจ๋ธ์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์˜๋ฏธ์ž…๋‹ˆ๋‹ค. ์ด ์ฐจ์ด๋Š” Phi-2 ๋ฐ EEVE-Korean-2.8B tokenizer์˜ ๊ฒฝ์šฐ ๋”์šฑ ๋‘๋“œ๋Ÿฌ์ง‘๋‹ˆ๋‹ค.[|endofturn|]
kens๋ฅผ ๋Œ€ํ‘œํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์ƒˆ๋กœ์šด ํ† ํฐํ™” ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด 1.6B ํ† ํฐ๋งŒ ์‚ฌ์šฉํ•˜์—ฌ ๊ฑฐ์˜ ์ ˆ๋ฐ˜์œผ๋กœ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ฐจ์ด๋Š” Phi-2 ๋ฐ EEVE-Korean-2.8 ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๊ฐ๊ฐ 5.6B ํ† ํฐ๊ณผ 1.6B ํ† ํฐ์ด ํ•„์š”ํ•œ ๊ฒฝ์šฐ ๋”์šฑ ๋‘๋“œ๋Ÿฌ์ง‘๋‹ˆ๋‹ค. ํ† ํฐ ๊ธธ์ด์— ๋Œ€ํ•œ ์ปดํ“จํŒ… ๋ณต์žก์„ฑ์ด ์ด์ฐจ์ ์œผ๋กœ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•  ๋•Œ, ์ด๋Š” ๋‘ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์˜๋ฏธ๋กœ ํ•ด์„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋™์ผํ•œ GPU์—์„œ 4๋ฐฐ ์ด์ƒ ๊ธด ์‹œํ€€์Šค๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜๋Š” ๋‘˜์งธ, ๋™์ผํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฑฐ์˜ 4๋ฐฐ ๋” ๊ณ„์‚ฐ ํšจ์œจ์ ์œผ๋กœ ๋ชจ๋ธ์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์˜๋ฏธ์ž…๋‹ˆ๋‹ค. ์ด ์ฐจ์ด๋Š” Phi-2 ๋ฐ EEVE-Korean-2.8B ํ† ํฐํ™”๊ธฐ์˜ ๊ฒฝ์šฐ ๋”์šฑ ๋‘๋“œ๋Ÿฌ์ง‘๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]kens to represent them, but our new tokenizer can do so with almost half, using only 1.6B tokens. This difference becomes even more pronounced in the case of Phi-2 and EEVE-Korean-2.88 models, where they require 5.6B tokens and 1.6B tokens respectively. Considering that transformers have a quadratic-increasing computation complexity with respect to token length, this can be interpreted in two significant ways. First, it allows for processing sequences more than 4 times longer on the same GPU. Or second, it means our model can be trained nearly 4 times more computationally efficiently on the same dataset. This difference becomes even more pronounced in the case of the Phi-2 and EEVE-Korean-2.8B tokenizers. [|assistant|]
๋‹ค์Œ ๊ด€์ฐฐ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. Hoffmann ๋“ฑ์˜ ์ถ”์ •์น˜๊ฐ€ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ตœ์ ์ด์—ˆ๊ณ  ์šฐ๋ฆฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๊ทธ๋“ค์ด ์‚ฌ์šฉํ•œ ์›๋ณธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ผ๋ฐ˜์  ๋˜๋Š” ๋ฌด์ž‘์œ„ ํ•˜์œ„ ์ง‘ํ•ฉ์ด๋ผ๋ฉด, ์šฐ๋ฆฌ์˜ ์ถ”์ •์น˜๋Š” ์–ด๋–ค ๊ณต๋ถ„์‚ฐ ํ–‰๋ ฌ \(\Sigma\)์„ ๊ฐ€์ง„ ๊ทธ๋“ค์˜ ์ถ”์ •์น˜๋ฅผ ์ค‘์‹ฌ์œผ๋กœ ์ •๊ทœ ๋ถ„ํฌ๋ฅผ ๋Œ€๋žต์ ์œผ๋กœ ๋”ฐ๋ผ์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋‹ค์Œ ๊ด€์ฐฐ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ํ˜ธํ”„๋งŒ ๋“ฑ์˜ ์ถ”์ •์น˜๊ฐ€ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ตœ์ ์ด์—ˆ๊ณ  ์šฐ๋ฆฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๊ทธ๋“ค์ด ์‚ฌ์šฉํ•œ ์›๋ณธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ผ๋ฐ˜์  ๋˜๋Š” ๋ฌด์ž‘์œ„ ํ•˜์œ„ ์ง‘ํ•ฉ์ด๋ผ๋ฉด, ์šฐ๋ฆฌ์˜ ์ถ”์ •์น˜๋Š” ์–ด๋–ค ๊ณต๋ถ„์‚ฐ ํ–‰๋ ฌ \(\Sigma\)์„ ๊ฐ€์ง„ ๊ทธ๋“ค์˜ ์ถ”์ •์น˜๋ฅผ ์ค‘์‹ฌ์œผ๋กœ ์ •๊ทœ ๋ถ„ํฌ๋ฅผ ๋Œ€๋žต์ ์œผ๋กœ ๋”ฐ๋ผ์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]based on the following observation: if Hoffmann et al.'s estimates were optimal on the full dataset and our dataset was a generic or random subset of the original dataset they used, then our estimates should approximately follow a normal distribution centered around their estimates with some covariance matrix \(\Sigma\). [|assistant|]
**Validation Set Perplexity**. ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋Š” ์ฃผ๋กœ [59]์—์„œ ๊ฐ€์ ธ์˜ค๋Š”๋ฐ, ์—ฌ๊ธฐ์—๋Š” CommonCrawl, DM ์ˆ˜ํ•™, HackerNews, OpenSubtitles, OpenWebText2, ํ”„๋กœ์ ํŠธ ๊ตฌํ…๋ฒ ๋ฅดํฌ, USPTO, ์œ„ํ‚คํ”ผ๋””์•„ ๋“ฑ Pile [16]์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์—์„œ ํŒŒ์ƒ๋œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ PushShift.io Reddit ๋ฐ์ดํ„ฐ ์„ธํŠธ [4]์—์„œ ์–ป์€ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ(์—ฌ๊ธฐ์„œ๋Š” _redditflattened_์ด๋ผ๊ณ  ํ•จ)๋„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ _CC-dedup_์˜ ํ›ˆ๋ จ-๊ฒ€์ฆ ๋ถ„ํ• ์—์„œ ์–ป์€ ๊ฒ€์ฆ ์„ธํŠธ์™€ C4 [41]์˜ ๊ฒ€์ฆ ์„ธํŠธ์—์„œ Perplexity๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ ๋‹นํ™ฉ์Šค๋Ÿฌ์›€**. ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋Š” ์ฃผ๋กœ [59]์—์„œ ๊ฐ€์ ธ์˜ค๋Š”๋ฐ, ์—ฌ๊ธฐ์—๋Š” CommonCrawl, DM ์ˆ˜ํ•™, HackerNews, OpenSubtitles, OpenWebText2, ํ”„๋กœ์ ํŠธ ๊ตฌํ…๋ฒ ๋ฅดํฌ, USPTO, ์œ„ํ‚คํ”ผ๋””์•„ ๋“ฑ Pile [16]์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์—์„œ ํŒŒ์ƒ๋œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ PushShift.io Reddit ๋ฐ์ดํ„ฐ ์„ธํŠธ [4]์—์„œ ์–ป์€ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ(์—ฌ๊ธฐ์„œ๋Š” _redditflattened_์ด๋ผ๊ณ  ํ•จ)๋„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ _CC-dedup_์˜ ํ›ˆ๋ จ-๊ฒ€์ฆ ๋ถ„ํ• ์—์„œ ์–ป์€ ๊ฒ€์ฆ ์„ธํŠธ์™€ C4 [41]์˜ ๊ฒ€์ฆ ์„ธํŠธ์—์„œ ๋‹นํ™ฉ์Šค๋Ÿฌ์›€์„ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Validation Set Perplexity**. Our validation sets mainly come from [59], which includes validation sets derived from subsets of the Pile [16] such as CommonCrawl, DM Mathematics, HackerNews, OpenSubtitles, OpenWebText2, Project Gutenberg, USPTO, Wikipedia. We also include a validation set obtained from the PushShift.io Reddit dataset [4] (which we refer to as _redditflattened_). In addition, we measure perplexity on a validation set obtained from a train-validation split of our source dataset _CC-dedup_, and a validation set from C4 [41]. [|assistant|]
์ผ๊ด€์„ฑ์ด ์—†๋Š” ๊ฒƒ์€ ์ถ”๋ก ์˜ ๋˜ ๋‹ค๋ฅธ ์ฃผ์š” ๊ณผ์ œ์ด๋ฉฐ, Dua ๋“ฑ(2022)์€ ๋ชจ๋ธ์ด ์ผ๊ด€๋œ ๋‹ต์„ ์ œ๊ณตํ•  ๋•Œ๊นŒ์ง€ ์ด์ „ ์ถ”๋ก  ๊ฒฐ๊ณผ๋ฅผ ํ”„๋กฌํ”„ํŠธ๋กœ ๋ฐ˜๋ณต์ ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Paul ๋“ฑ(2023)์€ ์ถ”๋ก  ๊ณผ์ •์— ๋Œ€ํ•œ ๊ตฌ์กฐํ™”๋œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ๋น„ํ‰ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. Self-Refine(Madaan et al., 2023)์€ ๋ฐ˜๋ณต์ ์ธ ์ž์ฒด ํ”ผ๋“œ๋ฐฑ ๋ฐ ๊ฐœ์„ ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ถ”๋ก ์˜ ์˜ค๋ฅ˜๋ฅผ ์™„ํ™”ํ•ฉ๋‹ˆ๋‹ค. Self-Refine๊ณผ ๋น„๊ตํ•˜์—ฌ Reflexion(Shinn et al., 2023)์€ ๋ฐ˜์„ฑ์„ ์œ„ํ•œ Reinforcement Learning์„ ๋„์ž…ํ•˜์—ฌ ์˜์‚ฌ ๊ฒฐ์ • ๊ธฐ๋Šฅ์„ ์ถ”๊ฐ€๋กœ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฒ€์ฆ์„ ์œ„ํ•ด ์—ญ ์ถ”๋ก (Yu et al., 2023)์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. RCoT(Xue et al., 2023)๋Š” ์ถ”๋ก  ์ฒด์ธ์— ๋”ฐ๋ผ ์งˆ๋ฌธ์„ ์žฌ๊ตฌ์„ฑํ•˜๋ฉฐ, ์›๋ž˜ ์งˆ๋ฌธ๊ณผ์˜ ๋ถˆ์ผ์น˜๋Š” ์ถ”๋ก  ๊ณผ์ •์˜ ์˜ค๋ฅ˜๋ฅผ ๋…ธ์ถœ์‹œํ‚ต๋‹ˆ๋‹ค. FOBAR(Jiang et al., 2023)์™€ Self Verification(Weng et al., 2022)์€ ์งˆ๋ฌธ์˜ ์กฐ๊ฑด์„ ๋‹ต๋ณ€์—์„œ ์ถ”๋ก ํ•˜์—ฌ ๊ฒ€์ฆ์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. FOBAR๋Š” ์งˆ๋ฌธ์˜ ๋ณ€์ˆ˜๋ฅผ ์ถ”๋ก ํ•˜๊ณ , Self Verification์€ ์งˆ๋ฌธ์˜ ์กฐ๊ฑด์„ ์ถ”๋ก ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Huang ๋“ฑ(2023)์€ LLM์ด ์™ธ๋ถ€ ํ”ผ๋“œ๋ฐฑ ์—†์ด ์ž์ฒด ์ˆ˜์ •ํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์œผ๋ฉฐ, ์„ฑ๋Šฅ์ด ์ €ํ•˜๋  ์ˆ˜๋„ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ผ๊ด€์„ฑ์ด ์—†๋Š” ๊ฒƒ์€ ์ถ”๋ก ์˜ ๋˜ ๋‹ค๋ฅธ ์ฃผ์š” ๊ณผ์ œ์ด๋ฉฐ, Dua ๋“ฑ(2022)์€ ๋ชจ๋ธ์ด ์ผ๊ด€๋œ ๋‹ต์„ ์ œ๊ณตํ•  ๋•Œ๊นŒ์ง€ ์ด์ „ ์ถ”๋ก  ๊ฒฐ๊ณผ๋ฅผ ํ”„๋กฌํ”„ํŠธ๋กœ ๋ฐ˜๋ณต์ ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Paul ๋“ฑ(2023)์€ ์ถ”๋ก  ๊ณผ์ •์— ๋Œ€ํ•œ ๊ตฌ์กฐํ™”๋œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ๋น„ํ‰ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์…€ํ”„ ๋ฆฌํŒŒ์ธ(Madaan et al., 2023)์€ ๋ฐ˜๋ณต์ ์ธ ์ž์ฒด ํ”ผ๋“œ๋ฐฑ ๋ฐ ๊ฐœ์„ ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ถ”๋ก ์˜ ์˜ค๋ฅ˜๋ฅผ ์™„ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์…€ํ”„ ๋ฆฌํŒŒ์ธ๊ณผ ๋น„๊ตํ•˜์—ฌ ๋ฆฌํ”Œ๋ ‰์…˜(์‹  ๋“ฑ, 2023)์€ ๋ฐ˜์„ฑ์„ ์œ„ํ•œ ๊ฐ•ํ™” ํ•™์Šต์„ ๋„์ž…ํ•˜์—ฌ ์˜์‚ฌ ๊ฒฐ์ • ๊ธฐ๋Šฅ์„ ์ถ”๊ฐ€๋กœ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฒ€์ฆ์„ ์œ„ํ•ด ์—ญ ์ถ”๋ก (Yu et al., 2023)์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. RCoT(Xue et al., 2023)๋Š” ์ถ”๋ก  ์ฒด์ธ์— ๋”ฐ๋ผ ์งˆ๋ฌธ์„ ์žฌ๊ตฌ์„ฑํ•˜๋ฉฐ, ์›๋ž˜ ์งˆ๋ฌธ๊ณผ์˜ ๋ถˆ์ผ์น˜๋Š” ์ถ”๋ก  ๊ณผ์ •์˜ ์˜ค๋ฅ˜๋ฅผ ๋…ธ์ถœ์‹œํ‚ต๋‹ˆ๋‹ค. FOBAR(Jiang et al., 2023)์™€ ์ž์ฒด ๊ฒ€์ฆ(Weng et al., 2022)์€ ์งˆ๋ฌธ์˜ ์กฐ๊ฑด์„ ๋‹ต๋ณ€์—์„œ ์ถ”๋ก ํ•˜์—ฌ ๊ฒ€์ฆ์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. FOBAR๋Š” ์งˆ๋ฌธ์˜ ๋ณ€์ˆ˜๋ฅผ ์ถ”๋ก ํ•˜๊ณ , ์ž์ฒด ๊ฒ€์ฆ์€ ์งˆ๋ฌธ์˜ ์กฐ๊ฑด์„ ์ถ”๋ก ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Huang ๋“ฑ(2023)์€ LLM์ด ์™ธ๋ถ€ ํ”ผ๋“œ๋ฐฑ ์—†์ด ์ž์ฒด ์ˆ˜์ •ํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์œผ๋ฉฐ, ์„ฑ๋Šฅ์ด ์ €ํ•˜๋  ์ˆ˜๋„ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Inconsistency is another major challenge in reasoning, Dua et al. (2022) iteratively uses previous reasoning results as prompts until the model gives a consistent answer. Paul et al. (2023) trains a critic model to provide structured feedback on the reasoning process. Self-Refine (Madaan et al., 2023) performs iterative self-feedback and refinement to alleviate errors in reasoning. Compared with Self-Refine, Reflexion (Shinn et al., 2023) introduces reinforcement learning for reflection, which additionally brings decision-making capability. Meanwhile, some work introduces backward reasoning (Yu et al., 2023) for verification. RCoT (Xue et al., 2023) reconstructs the question according to the reasoning chains, and its inconsistency with the original question exposes errors in the reasoning process. FOBAR (Jiang et al., 2023) and Self Verification (Weng et al., 2022) perform verification by deducing the conditions in the question from the answer. FOBAR infers the variables in the question, and Self Verification infers the conditions in the question. However, Huang et al. (2023) finds that LLMs struggle to self-correct without external feedback, and it could even lead to a performance decline. [|assistant|]
26๊ฐœ์˜ 2.6B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ ๊ฐ๊ฐ์„ NVIDIA A100 GPU์—์„œ ํ•™์Šตํ–ˆ์œผ๋ฉฐ, ๊ฐ ๋ชจ๋ธ์˜ ํ•™์Šต์—๋Š” ์ตœ๋Œ€ 2304์‹œ๊ฐ„์˜ GPU ์‹œ๊ฐ„์ด ์†Œ์š”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ํ•™์Šต ๋น„์šฉ์€ ์•ฝ 59,000์‹œ๊ฐ„์˜ GPU ์‹œ๊ฐ„์ด ์†Œ์š”๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
26๊ฐœ์˜ 2.6B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ ๊ฐ๊ฐ์„ NVIDIA A100 GPU์—์„œ ํ•™์Šตํ–ˆ์œผ๋ฉฐ, ๊ฐ ๋ชจ๋ธ์˜ ํ•™์Šต์—๋Š” ์ตœ๋Œ€ 2304์‹œ๊ฐ„์˜ GPU ์‹œ๊ฐ„์ด ์†Œ์š”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ํ•™์Šต ๋น„์šฉ์€ ์•ฝ 59.000์‹œ๊ฐ„์˜ GPU ์‹œ๊ฐ„์ด ์†Œ์š”๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We trained each of our 26 2.6B parameter models on NVIDIA A100 GPUs, and the training of each model took up to 2304 GPU hours. Therefore, the total training costs amounted to \(\approx 59.000\) GPU hours. [|assistant|]
\(\bullet\)_PaLM_(540B) [56]์€ ์†Œ์…œ ๋ฏธ๋””์–ด ๋Œ€ํ™”, ํ•„ํ„ฐ๋ง๋œ ์›น ํŽ˜์ด์ง€, ์ฑ…, Github, ๋‹ค๊ตญ์–ด Wikipedia, ๋‰ด์Šค์—์„œ ์†Œ์‹ฑํ•œ 780B ํ† ํฐ์˜ pre-training ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_PaLM_(540B) [56]์€ ์†Œ์…œ ๋ฏธ๋””์–ด ๋Œ€ํ™”, ํ•„ํ„ฐ๋ง๋œ ์›น ํŽ˜์ด์ง€, ์ฑ…, Github, ๋‹ค๊ตญ์–ด ์œ„ํ‚คํ”ผ๋””์•„, ๋‰ด์Šค์—์„œ ์†Œ์‹ฑํ•œ 780B ํ† ํฐ์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_PaLM_ (540B) [56] uses a pre-training dataset of 780B tokens, which is sourced from social media conversations, filtered webpages, books, Github, multilingual Wikipedia, and news. [|assistant|]
์šฐ๋ฆฌ์˜ ๋„๋ฉ”์ธ ๋ง๋ญ‰์น˜๋ฅผ ํ๋ ˆ์ด์…˜ํ•˜๊ธฐ ์œ„ํ•œ ๋‘ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์†Œ์Šค, ์ฆ‰ ๊ธˆ์œต ๋‰ด์Šค CommonCrawl๊ณผ SEC filings์„ ์„ค๋ช…ํ•œ๋‹ค.[|endofturn|]
์šฐ๋ฆฌ์˜ ๋„๋ฉ”์ธ ๋ง๋ญ‰์น˜๋ฅผ ํ๋ ˆ์ด์…˜ํ•˜๊ธฐ ์œ„ํ•œ ๋‘ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์†Œ์Šค, ์ฆ‰ ๊ธˆ์œต ๋‰ด์Šค CommonCrawl๊ณผ SEC ์„œ๋ฅ˜๋ฅผ ์„ค๋ช…ํ•œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We describe the two data sources for curating our domain corpus: Financial News CommonCrawl and SEC filings. [|assistant|]
์•”๊ธฐ ๋ถ„ํฌ. ๊ทธ๋ฆผ 6์€ 4๊ฐ€์ง€ ์œ ํ˜•์˜ prompting์„ ์‚ฌ์šฉํ•  ๋•Œ ์ƒ์„ฑ๋œ ๋ชจ๋“  ์‹œํ€€์Šค์— ๋Œ€ํ•œ ์•”๊ธฐ๋Ÿ‰ ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค: ํ›ˆ๋ จ์—์„œ ์ค‘๋ณต์ด ์žˆ๋Š” ํ›ˆ๋ จ ์˜ˆ์ œ,[|endofturn|]
์•”๊ธฐ ๋ถ„ํฌ.๊ทธ๋ฆผ 6์€ 4๊ฐ€์ง€ ์œ ํ˜•์˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ ์ƒ์„ฑ๋œ ๋ชจ๋“  ์‹œํ€€์Šค์— ๋Œ€ํ•œ ์•”๊ธฐ๋Ÿ‰ ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค: ํ›ˆ๋ จ์—์„œ ์ค‘๋ณต์ด ์žˆ๋Š” ํ›ˆ๋ จ ์˜ˆ์ œ,[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Distribution of memorization.Figure 6 shows the distribution in memorization amount over all generated sequences when using four types of prompting: train example with duplicates in train, [|assistant|]
์ด๋ฅผ ํ†ตํ•ด ํ•™์Šต์šฉ 3,762๊ฐœ, ๊ฒ€์ฆ์šฉ 840๊ฐœ, ํ…Œ์ŠคํŠธ์šฉ 914๊ฐœ ๋“ฑ ์ด 5,516๊ฐœ์˜ ์ด์ง„ ์งˆ๋ฌธ ์„ธํŠธ๊ฐ€ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค(Table 1(a)). Table 12๋Š” ์ƒ˜ํ”Œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์— ๋Œ€ํ•œ ๋‚ด์šฉ์ด๋ฉฐ, ํ๋ ˆ์ด์…˜ ํ”„๋กœ์„ธ์Šค์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ Appendix C๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.[|endofturn|]
์ด๋ฅผ ํ†ตํ•ด ํ•™์Šต์šฉ 3,762๊ฐœ, ๊ฒ€์ฆ์šฉ 840๊ฐœ, ํ…Œ์ŠคํŠธ์šฉ 914๊ฐœ ๋“ฑ ์ด 5,516๊ฐœ์˜ ์ด์ง„ ์งˆ๋ฌธ ์„ธํŠธ๊ฐ€ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค(ํ‘œ 1(a)). ํ‘œ 12๋Š” ์ƒ˜ํ”Œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์— ๋Œ€ํ•œ ๋‚ด์šฉ์ด๋ฉฐ, ํ๋ ˆ์ด์…˜ ํ”„๋กœ์„ธ์Šค์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก C๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]This yields a set of 5,516 binary questions, including 3,762 for training, 840 for validation, and 914 for testing (Table 1(a)). See Table 12 for a sample data point and Appendix C for details about the curation process. [|assistant|]
๋ฐ์ดํ„ฐ ์„ ํƒ์˜ ํšจ๊ณผ๋Š” ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์ด ์›น ๋ฐ์ดํ„ฐ ์ฝ”ํผ์Šค์—์„œ ํŒŒ์ƒ๋œ ๊ฒƒ์ธ์ง€ ์—ฌ๋ถ€์— ๋”ฐ๋ผ ๊ฐœ๋ณ„ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์— ์ƒ๋‹นํ•œ ์ฐจ์ด๊ฐ€ ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์ž์„ธํ•œ ๋‚ด์šฉ๊ณผ ๋ถ„์„์€ ์„น์…˜ 4.4.1 ์ฐธ์กฐ). ์ด์— ์ฐฉ์•ˆํ•˜์—ฌ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์„ Web-snapshots (C4, CommonCrawl ๋ฐ CC-dedup)๊ณผ Non-web snapshots๋กœ ๋‚˜๋ˆ„๊ณ  ์ด๋Ÿฌํ•œ ์ง‘ํ•ฉ ๋‚ด์—์„œ ํ‰๊ท  Perplexity๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋ฐ์ดํ„ฐ ์„ ํƒ์˜ ํšจ๊ณผ๋Š” ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์ด ์›น ๋ฐ์ดํ„ฐ ์ฝ”ํผ์Šค์—์„œ ํŒŒ์ƒ๋œ ๊ฒƒ์ธ์ง€ ์—ฌ๋ถ€์— ๋”ฐ๋ผ ๊ฐœ๋ณ„ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์— ์ƒ๋‹นํ•œ ์ฐจ์ด๊ฐ€ ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์ž์„ธํ•œ ๋‚ด์šฉ๊ณผ ๋ถ„์„์€ ์„น์…˜ 4.4.1 ์ฐธ์กฐ). ์ด์— ์ฐฉ์•ˆํ•˜์—ฌ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์„ ์›น ์Šค๋ƒ…์ƒท(C4, CommonCrawl ๋ฐ CC-dedup)๊ณผ ๋น„์›น ์Šค๋ƒ…์ƒท์œผ๋กœ ๋‚˜๋ˆ„๊ณ  ์ด๋Ÿฌํ•œ ์ง‘ํ•ฉ ๋‚ด์—์„œ ํ‰๊ท  ๋‚œํ•ดํ•จ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We notice that the effects of data selection vary significantly on individual validation sets depending on whether the validation set was derived from a web data corpus or not (see more details and analysis in Section 4.4.1). Motivated by this, we split validation sets into Web-snapshots (C4, CommonCrawl, and CC-dedup) and Non-web snapshots, and report average perplexity within these sets. [|assistant|]
MiniCPM์€ 2024๋…„ 2์›” 1์ผ์— ๊ธฐ์ˆ  ๋ธ”๋กœ๊ทธ๋กœ ์ฒ˜์Œ ๊ฒŒ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์ดํ›„๋กœ ์ปค๋ฎค๋‹ˆํ‹ฐ๋กœ๋ถ€ํ„ฐ ์ˆ˜๋งŽ์€ ํ†ต์ฐฐ๋ ฅ ์žˆ๋Š” ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์•„ ์ด ๋…ผ๋ฌธ์˜ ๊ฐœ๋ฐœ์— ํฌ๊ฒŒ ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ท€์ค‘ํ•œ ๋…ผ์˜๋ฅผ ์ œ๊ณตํ•œ ์ฒญํŒ… ์ €์šฐ์™€ ์•„๋ฅด๋ฉ˜ ์•„๊ฐ€์ž์–€์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ ๋ธ”๋กœ๊ทธ์˜ ๋ชจํ˜ธํ•œ ๋ถ€๋ถ„์„ ๋ช…ํ™•ํžˆ ํ•˜๋Š” ๋ฐ ์„ธ์‹ฌํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•œ ํ”ผ๊ธด ์„ ๊ณผ ์˜Œ ์™•์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, MiniCPM์„ llama.cpp ๋“ฑ๊ณผ ๊ฐ™์€ inference ํ”„๋ ˆ์ž„์›Œํฌ์— ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ๋…ธ๋ ฅํ•œ ์˜คํ”ˆ ์†Œ์Šค ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ๋…ธ๋ ฅ์— ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.[|endofturn|]
MiniCPM์€ 2024๋…„ 2์›” 1์ผ์— ๊ธฐ์ˆ  ๋ธ”๋กœ๊ทธ๋กœ ์ฒ˜์Œ ๊ฒŒ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์ดํ›„๋กœ ์ปค๋ฎค๋‹ˆํ‹ฐ๋กœ๋ถ€ํ„ฐ ์ˆ˜๋งŽ์€ ํ†ต์ฐฐ๋ ฅ ์žˆ๋Š” ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์•„ ์ด ๋…ผ๋ฌธ์˜ ๊ฐœ๋ฐœ์— ํฌ๊ฒŒ ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ท€์ค‘ํ•œ ๋…ผ์˜๋ฅผ ์ œ๊ณตํ•œ ์ฒญํŒ… ์ €์šฐ์™€ ์•„๋ฅด๋ฉ˜ ์•„๊ฐ€์ž์–€์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ ๋ธ”๋กœ๊ทธ์˜ ๋ชจํ˜ธํ•œ ๋ถ€๋ถ„์„ ๋ช…ํ™•ํžˆ ํ•˜๋Š” ๋ฐ ์„ธ์‹ฌํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•œ ํ”ผ๊ธด ์„ ๊ณผ ์˜Œ ์™•์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, MiniCPM์„ llama.cpp ๋“ฑ๊ณผ ๊ฐ™์€ ์ถ”๋ก  ํ”„๋ ˆ์ž„์›Œํฌ์— ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ๋…ธ๋ ฅํ•œ ์˜คํ”ˆ ์†Œ์Šค ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ๋…ธ๋ ฅ์— ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]MiniCPM was initially published as a technical blog on February 1st, 2024. Since then, we have received numerous insightful feedback from the community, significantly contributing to the development of this paper. We extend our gratitude to Chunting Zhou and Armen Aghajanyan for their valuable discussions. Special thanks go to Feigin Sun and Yan Wang for their meticulous feedback on clarifying ambiguities in the blog. Additionally, we appreciate the open-source community's efforts in integrating MiniCPM into inference frameworks like llama.cpp, etc. [|assistant|]
๋งŽ์€ ์—ฐ๊ตฌ์—์„œ ๋‹ค์–‘ํ•œ ๊ด€์ ์—์„œ LMs์˜ ํ•™์Šต ์—ญํ•™์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. Carlini et al. (2022)์€ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ๋ฐ์ดํ„ฐ ์ค‘๋ณต ๋นˆ๋„์— ๋”ฐ๋ฅธ ์•”๊ธฐ ๋Šฅ๋ ฅ์„ ์ •๋Ÿ‰ํ™”ํ•ฉ๋‹ˆ๋‹ค. Tirumala et al. (2022)์€ ๋” ํฐ LMs์ด ๋œ ํŠ€์–ด๋‚˜๋“ค๋ฉด์„œ ๋” ๋น ๋ฅด๊ฒŒ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์•”๊ธฐํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. Xia et al. (2023)์€ perplexity๊ฐ€ ๋‹ค๋ฅธ ์š”์ธ๋ณด๋‹ค ๋ชจ๋ธ ๋™์ž‘์„ ๋” ์ž˜ ์˜ˆ์ธกํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Dery et al. (2022)์€ ๋ถ„๋ฅ˜ ์ž‘์—…๊ณผ RoBERTa ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข… ์ž‘์—… ์ธ์‹ ์‚ฌ์ „ ํ•™์Šต์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ์—ฐ๊ตฌ๋Š” ๋ณธ์ธ์ด ๋ณธ ๋ฌธ์„œ์—์„œ ์ •๋ณด๋ฅผ ๊ธฐ์–ตํ•˜๊ณ  ์ผ๋ฐ˜ํ™”ํ•˜์—ฌ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋Šฅ๋ ฅ์— ํŠน๋ณ„ํžˆ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ๋‹ค๋Š” ์ ์—์„œ ์ฐจ๋ณ„ํ™”๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋งŽ์€ ์—ฐ๊ตฌ์—์„œ ๋‹ค์–‘ํ•œ ๊ด€์ ์—์„œ LM์˜ ํ•™์Šต ์—ญํ•™์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. Carlini et al. (2022)์€ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ๋ฐ์ดํ„ฐ ์ค‘๋ณต ๋นˆ๋„์— ๋”ฐ๋ฅธ ์•”๊ธฐ ๋Šฅ๋ ฅ์„ ์ •๋Ÿ‰ํ™”ํ•ฉ๋‹ˆ๋‹ค. Tirumala et al. (2022)์€ ๋” ํฐ LM์ด ๋œ ํŠ€์–ด๋‚˜๋“ค๋ฉด์„œ ๋” ๋น ๋ฅด๊ฒŒ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์•”๊ธฐํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. Xia et al. (2023)์€ ๋‹นํ˜น๊ฐ์ด ๋‹ค๋ฅธ ์š”์ธ๋ณด๋‹ค ๋ชจ๋ธ ๋™์ž‘์„ ๋” ์ž˜ ์˜ˆ์ธกํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Dery et al. (2022)์€ ๋ถ„๋ฅ˜ ์ž‘์—…๊ณผ RoBERTa ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข… ์ž‘์—… ์ธ์‹ ์‚ฌ์ „ ํ•™์Šต์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ์—ฐ๊ตฌ๋Š” ๋ณธ์ธ์ด ๋ณธ ๋ฌธ์„œ์—์„œ ์ •๋ณด๋ฅผ ๊ธฐ์–ตํ•˜๊ณ  ์ผ๋ฐ˜ํ™”ํ•˜์—ฌ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋Šฅ๋ ฅ์— ํŠน๋ณ„ํžˆ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ๋‹ค๋Š” ์ ์—์„œ ์ฐจ๋ณ„ํ™”๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Many works study the training dynamics of LMs from different perspectives. Carlini et al. (2022) quantifies memorization across model sizes and the frequency of data duplication. Tirumala et al. (2022) finds that larger LMs memorize training data faster with less overfitting. Xia et al. (2023) show that perplexity is more predictive of model behaviors than other factors. Dery et al. (2022) studies end-task aware pre-training using classification tasks and RoBERTa models. Our work differs in that we specifically focus on the capacity of recalling and generalizing information from a seen document to answer questions. [|assistant|]
\(\bullet\)_Scaling_. ์ด์ „ ๋ถ€๋ถ„์—์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, Transformer ์–ธ์–ด ๋ชจ๋ธ์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋” ํฐ ๋ชจ๋ธ/๋ฐ์ดํ„ฐ ํฌ๊ธฐ์™€ ๋” ๋งŽ์€ ํ›ˆ๋ จ ์ปดํ“จํŒ…์ด ๋ชจ๋ธ ์šฉ๋Ÿ‰ ํ–ฅ์ƒ์œผ๋กœ ์ด์–ด์ง€๋Š” ๋šœ๋ ทํ•œ scaling ํšจ๊ณผ๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค[30, 34]. ๋Œ€ํ‘œ์ ์ธ ๋‘ ๊ฐ€์ง€ ๋ชจ๋ธ๋กœ์„œ GPT-3์™€ PaLM์€ ๊ฐ๊ฐ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ 175B์™€ 540B๋กœ ๋Š˜๋ ค scaling ํ•œ๊ณ„๋ฅผ ํƒ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ compute ์˜ˆ์‚ฐ์ด ์ œํ•œ๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— scaling laws๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ compute ๋ฆฌ์†Œ์Šค๋ฅผ ๋ณด๋‹ค ํšจ์œจ์ ์œผ๋กœ ํ• ๋‹นํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Chinchilla๋Š” ๋™์ผํ•œ compute ์˜ˆ์‚ฐ์œผ๋กœ ๋ฐ์ดํ„ฐ ๊ทœ๋ชจ๋ฅผ ๋Š˜๋ ค์„œ ๋” ๋งŽ์€ ํ›ˆ๋ จ ํ† ํฐ์„ ๊ฐ€์ง„ ๋ชจ๋ธ์ธ Gopher๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค[34]. ๋˜ํ•œ, ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์ด ๋ชจ๋ธ ์šฉ๋Ÿ‰์— ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋ฏ€๋กœ ๋ฐ์ดํ„ฐ scaling์€ ์‹ ์ค‘ํ•œ ์ •๋ฆฌ ๊ณผ์ •์„ ๊ฑฐ์ณ์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_์Šค์ผ€์ผ๋ง_. ์ด์ „ ๋ถ€๋ถ„์—์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ํŠธ๋žœ์Šคํฌ๋จธ ์–ธ์–ด ๋ชจ๋ธ์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋” ํฐ ๋ชจ๋ธ/๋ฐ์ดํ„ฐ ํฌ๊ธฐ์™€ ๋” ๋งŽ์€ ํ›ˆ๋ จ ์ปดํ“จํŒ…์ด ๋ชจ๋ธ ์šฉ๋Ÿ‰ ํ–ฅ์ƒ์œผ๋กœ ์ด์–ด์ง€๋Š” ๋šœ๋ ทํ•œ ์Šค์ผ€์ผ๋ง ํšจ๊ณผ๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค[30, 34]. ๋Œ€ํ‘œ์ ์ธ ๋‘ ๊ฐ€์ง€ ๋ชจ๋ธ๋กœ์„œ GPT-3์™€ PaLM์€ ๊ฐ๊ฐ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ 175B์™€ 540B๋กœ ๋Š˜๋ ค ์Šค์ผ€์ผ๋ง ํ•œ๊ณ„๋ฅผ ํƒ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ปดํ“จํŒ… ์˜ˆ์‚ฐ์ด ์ œํ•œ๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ์‚ฌ์šฉํ•˜์—ฌ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค๋ฅผ ๋ณด๋‹ค ํšจ์œจ์ ์œผ๋กœ ํ• ๋‹นํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์นจ์นœ์น ๋ผ๋Š” ๋™์ผํ•œ ์ปดํ“จํŒ… ์˜ˆ์‚ฐ์œผ๋กœ ๋ฐ์ดํ„ฐ ๊ทœ๋ชจ๋ฅผ ๋Š˜๋ ค์„œ ๋” ๋งŽ์€ ํ›ˆ๋ จ ํ† ํฐ์„ ๊ฐ€์ง„ ๋ชจ๋ธ์ธ ๊ณ ํผ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค[34]. ๋˜ํ•œ, ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์ด ๋ชจ๋ธ ์šฉ๋Ÿ‰์— ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋ฏ€๋กœ ๋ฐ์ดํ„ฐ ์Šค์ผ€์ผ๋ง์€ ์‹ ์ค‘ํ•œ ์ •๋ฆฌ ๊ณผ์ •์„ ๊ฑฐ์ณ์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_Scaling_. As discussed in previous parts, there exists an evident scaling effect in Transformer language models: larger model/data sizes and more training compute typically lead to an improved model capacity [30, 34]. As two representative models, GPT-3 and PaLM explored the scaling limits by increasing the model size to 175B and 540B, respectively. Since compute budget is usually limited, scaling laws can be further employed to conduct a more compute-efficient allocation of the compute resources. For example, Chinchilla (with more training tokens) outperforms its counterpart model Gopher (with a larger model size) by increasing the data scale with the same compute budget [34]. In addition, data scaling should be with careful cleaning process, since the quality of pre-training data plays a key role in the model capacity. [|assistant|]
๋ฐœ๊ฒฌ๊ฒฐ๊ณผ ๊ทธ๋ฆผ 2์˜ ๊ฒฐ๊ณผ๋Š” ๋„๋ฉ”์ธ ๋น„์œจ์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ๋„๋ฉ”์ธ Loss์˜ ์ •๋Ÿ‰์  ์˜ˆ์ธก ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋™์ผํ•œ ํฌ๊ธฐ๋กœ ๋™์ผํ•œ ์ˆ˜์˜ ๋‹จ๊ณ„๋กœ ํ›ˆ๋ จ๋œ ์ฒดํฌํฌ์ธํŠธ์˜ ๊ฒฝ์šฐ, ๊ณต์œ  ์ƒ์ˆ˜3์„ ๋บ€ ํ›„ ๋กœ๊ทธ ์Šค์ผ€์ผ์˜ ๋„๋ฉ”์ธ Loss๊ฐ€ ๋„๋ฉ”์ธ ๋น„์œจ์— ๋”ฐ๋ผ ์„ ํ˜•์ ์ธ ๊ด€๊ณ„๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋‘ ๋„๋ฉ”์ธ ๋ชจ๋‘์— ํ•ด๋‹น๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋ฐœ๊ฒฌ๊ฒฐ๊ณผ ๊ทธ๋ฆผ 2์˜ ๊ฒฐ๊ณผ๋Š” ๋„๋ฉ”์ธ ๋น„์œจ์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ๋„๋ฉ”์ธ ์†์‹ค์˜ ์ •๋Ÿ‰์  ์˜ˆ์ธก ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋™์ผํ•œ ํฌ๊ธฐ๋กœ ๋™์ผํ•œ ์ˆ˜์˜ ๋‹จ๊ณ„๋กœ ํ›ˆ๋ จ๋œ ์ฒดํฌํฌ์ธํŠธ์˜ ๊ฒฝ์šฐ, ๊ณต์œ  ์ƒ์ˆ˜3์„ ๋บ€ ํ›„ ๋กœ๊ทธ ์Šค์ผ€์ผ์˜ ๋„๋ฉ”์ธ ์†์‹ค์ด ๋„๋ฉ”์ธ ๋น„์œจ์— ๋”ฐ๋ผ ์„ ํ˜•์ ์ธ ๊ด€๊ณ„๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋‘ ๋„๋ฉ”์ธ ๋ชจ๋‘์— ํ•ด๋‹น๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]FindingsResults in Fig. 2 reveal the quantitative predictability of domain losses given the domain proportions. We encouragingly find that, for checkpoints with the same size and trained with the same number of steps, after subtracting a shared constant3, their domain losses in the log scale demonstrate a linear relationship to the domain proportion. This holds for both domains in our experiments. [|assistant|]
์ด์ „ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์ฃผ์š” ๋ฌธ์ œ๋Š” logit ๋ฒกํ„ฐ์— ๋Œ€ํ•œ ์„ ํ–‰์ด \([-B,0]\)์— ๊ฑธ์ณ ๊ท ์ผํ•œ ๊ฒฝ์šฐ์—๋„ ์ผ๋ถ€ ํ† ํฐ์ด ๋‹ค๋ฅธ ํ† ํฐ๋ณด๋‹ค ๋” ์ž์ฃผ ์ƒ˜ํ”Œ๋ง๋œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” "ํ•˜์ดํผ๋ ‰ํŠธ์•ต๊ธ€์˜ ์ค‘์‹ฌ" logit bias๊ฐ€ ํ•˜์ดํผ๋ ‰ํŠธ์•ต๊ธ€์„ argmax ์ขŒํ‘œ๋กœ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ๋™์ผํ•œ ํฌ๊ธฐ์˜ ๋ถ€๋ถ„์œผ๋กœ ๋ถ„ํ• ํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(\beta_{i}-\alpha_{i}\ll\beta_{j}-\alpha_{j}\)์ธ ๊ฒฝ์šฐ, \([\alpha_{i},\beta_{i}]\times[\alpha_{j},\beta_{j}]\)์— ๋Œ€ํ•œ ๊ท ์ผํ•œ ์„ ํ–‰ ํ•˜์—์„œ \(j\)๊ฐ€ \(i\)๋ณด๋‹ค ์ถœ๋ ฅ ํ† ํฐ์ด ๋  ๊ฐ€๋Šฅ์„ฑ์ด ํ›จ์”ฌ ๋” ๋†’์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์•Œ๊ณ ๋ฆฌ์ฆ˜ 3์—์„œ๋Š” ๋‹ค๋ฅธ logit์˜ ๊ด€์ ์—์„œ \(\mathsf{logit}_{i}\)๋ฅผ ํ•˜ํ•œ์œผ๋กœ ์ œํ•œํ•˜๋Š” ์ œ์•ฝ ์กฐ๊ฑด์„ ๊ฑฐ์˜ ์–ป์ง€ ๋ชปํ•˜๋ฏ€๋กœ \(\mathcal{C}\)์˜ ์™„ํ™”๊ฐ€ ๋” ์•ฝํ•ด์ง‘๋‹ˆ๋‹ค.[|endofturn|]
์ด์ „ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์ฃผ์š” ๋ฌธ์ œ๋Š” ๋กœ๊ทธ ๋ฒกํ„ฐ์— ๋Œ€ํ•œ ์„ ํ–‰์ด \([-B,0]\)์— ๊ฑธ์ณ ๊ท ์ผํ•œ ๊ฒฝ์šฐ์—๋„ ์ผ๋ถ€ ํ† ํฐ์ด ๋‹ค๋ฅธ ํ† ํฐ๋ณด๋‹ค ๋” ์ž์ฃผ ์ƒ˜ํ”Œ๋ง๋œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” "ํ•˜์ดํผ ๋ ‰ํŠธ ์•ต๊ธ€์˜ ์ค‘์‹ฌ" ๋กœ๊ทธํŠธ ๋น„์ฆˆ๊ฐ€ ํ•˜์ดํผ ๋ ‰ํŠธ ์•ต๊ธ€์„ ์ตœ๋Œ€ ์ขŒํ‘œ๋กœ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ๋™์ผํ•œ ํฌ๊ธฐ์˜ ๋ถ€๋ถ„์œผ๋กœ ๋ถ„ํ• ํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(\beta_{i}-\alpha_{i}\ll\beta_{j}-\alpha_{j}\)์ธ ๊ฒฝ์šฐ, \([\alpha_{i},\beta_{i}]\times[\alpha_{j},\beta_{j}]\)์— ๋Œ€ํ•œ ๊ท ์ผํ•œ ์„ ํ–‰ ํ•˜์—์„œ \(j\)๊ฐ€ \(i\)๋ณด๋‹ค ์ถœ๋ ฅ ํ† ํฐ์ด ๋  ๊ฐ€๋Šฅ์„ฑ์ด ํ›จ์”ฌ ๋” ๋†’์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์•Œ๊ณ ๋ฆฌ์ฆ˜ 3์—์„œ๋Š” ๋‹ค๋ฅธ ๋กœ๊ทธํŠธ์˜ ๊ด€์ ์—์„œ \(\mathsf{logit}_{i}\)๋ฅผ ํ•˜ํ•œ์œผ๋กœ ์ œํ•œํ•˜๋Š” ์ œ์•ฝ ์กฐ๊ฑด์„ ๊ฑฐ์˜ ์–ป์ง€ ๋ชปํ•˜๋ฏ€๋กœ \(\mathcal{C}\)์˜ ์™„ํ™”๊ฐ€ ๋” ์•ฝํ•ด์ง‘๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The main issue of the previous approach is that some tokens are sampled more often than others, even in the case our prior for the logit vector is uniform over \([-B,0]\). This is because the "centering of the hyperrectangle" logit bias does not partition the hyperrectangle into equally-sized parts labeled by the argmax coordinate. For example, if \(\beta_{i}-\alpha_{i}\ll\beta_{j}-\alpha_{j}\), under an uniform prior over \([\alpha_{i},\beta_{i}]\times[\alpha_{j},\beta_{j}]\), \(j\) will be much more likely to be the output token than \(i\). Hence, in Algorithm 3 we rarely get constraints lower-bounding \(\mathsf{logit}_{i}\) in terms of other logits, which makes for weaker relaxations of \(\mathcal{C}\). [|assistant|]
Vaswani et al. (2017)์„ ๊ธฐ๋ฐ˜์œผ๋กœ decoder-only Transformer ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ฑ„ํƒํ•˜๊ณ , ํ‘œ 1์— ์„ค๋ช…๋œ ๋Œ€๋กœ 1B ๋ฐ 7B ๋ณ€ํ˜•์„ ์ œ๊ณตํ•˜๋ฉฐ, 65B ๋ฒ„์ „์ด ๊ณง ์ถœ์‹œ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ํŠน์ • ์•„ํ‚คํ…์ฒ˜์—๋Š” PaLM(Chowdhery et al., 2022), LLaMA ์ œํ’ˆ๊ตฐ(Touvron et al., 2023a,b), OpenLM(Gururangan et al., 2023), Falcon(Almazrouei et al., 2023)๊ณผ ๊ฐ™์€ ์ตœ๊ทผ์˜ ๋‹ค๋ฅธ Large Language Models (LLMs)์— ๋”ฐ๋ผ Vaswani et al.(2017)์˜ ์ผ๋ฐ˜ Transformer์— ๋น„ํ•ด ๋ช‡ ๊ฐ€์ง€ ๊ฐœ์„  ์‚ฌํ•ญ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 2๋Š” ์ด๋Ÿฌํ•œ ๋‹ค๋ฅธ ์ œํ’ˆ๊ตฐ์˜ ๋น„์Šทํ•œ ํฌ๊ธฐ์˜ ๋ชจ๋ธ๊ณผ ์šฐ๋ฆฌ์˜ 7B ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ข…ํ•ฉ์ ์œผ๋กœ ๋น„๊ตํ•œ ํ‘œ์ž…๋‹ˆ๋‹ค.[|endofturn|]
Vaswani et al. (2017)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋””์ฝ”๋” ์ „์šฉ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ฑ„ํƒํ•˜๊ณ , ํ‘œ 1์— ์„ค๋ช…๋œ ๋Œ€๋กœ 1B ๋ฐ 7B ๋ณ€ํ˜•์„ ์ œ๊ณตํ•˜๋ฉฐ, 65B ๋ฒ„์ „์ด ๊ณง ์ถœ์‹œ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ํŠน์ • ์•„ํ‚คํ…์ฒ˜์—๋Š” PaLM(Chowdhery et al., 2022), LLaMA ์ œํ’ˆ๊ตฐ(Touvron et al., 2023a,b), OpenLM(Gururangan et al., 2023), Falcon(Almazrouei et al., 2023)๊ณผ ๊ฐ™์€ ์ตœ๊ทผ์˜ ๋‹ค๋ฅธ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์— ๋”ฐ๋ผ Vaswani et al.(2017)์˜ ์ผ๋ฐ˜ ํŠธ๋žœ์Šคํฌ๋จธ์— ๋น„ํ•ด ๋ช‡ ๊ฐ€์ง€ ๊ฐœ์„  ์‚ฌํ•ญ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 2๋Š” ์ด๋Ÿฌํ•œ ๋‹ค๋ฅธ ์ œํ’ˆ๊ตฐ์˜ ๋น„์Šทํ•œ ํฌ๊ธฐ์˜ ๋ชจ๋ธ๊ณผ ์šฐ๋ฆฌ์˜ 7B ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ข…ํ•ฉ์ ์œผ๋กœ ๋น„๊ตํ•œ ํ‘œ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We adopt a decoder-only transformer architecture based on Vaswani et al. (2017), and deliver 1B and 7B variants as described in Table 1, with a 65B version coming soon. Our specific architecture includes several improvements over the vanilla transformer from Vaswani et al. (2017) following other recent large language models like PaLM (Chowdhery et al., 2022), the LLaMA family (Touvron et al., 2023a,b), OpenLM (Gururangan et al., 2023), and Falcon (Almazrouei et al., 2023). Table 2 gives a comprehensive comparison of our 7B architecture to the similarly-sized models from these other families. [|assistant|]
๊ทธ๋ฆผ 18: "Reddit๊ฐ€ 2023๋…„ 7์›” 1์ผ ์ด์ „์— ์ œ์•ˆ๋œ API ์š”๊ธˆ ๊ฐ€๊ฒฉ ์ฑ…์ •์— ๋Œ€ํ•œ ๋ณ€๊ฒฝ ๋˜๋Š” ์—ฐ๊ธฐ๋ฅผ ๋ฐœํ‘œํ•  ๊ฒƒ์ธ๊ฐ€?"๋ผ๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•œ Fine-tuned ๋ชจ๋ธ์˜ ์‘๋‹ต์ž…๋‹ˆ๋‹ค(ํ•ด๊ฒฐ๋จ). ์ด ์˜ˆ์ธก์€ ์งˆ๋ฌธ์ด ๊ฐœ์„ค๋œ ์ง€ ํ•˜๋ฃจ ํ›„์ธ 2023๋…„ 6์›” 14์ผ์— ๊ฒ€์ƒ‰๋œ ๋‚ ์งœ์— ์ด๋ฃจ์–ด์กŒ์Šต๋‹ˆ๋‹ค. ๋‹น์‹œ ์ปค๋ฎค๋‹ˆํ‹ฐ๋Š” 40%์˜ ์˜ˆ์ธก์„ ๋‚ด๋ ธ๋Š”๋ฐ, ์ด ์‘๋‹ต์€ 20%๋ฅผ ์ œ์‹œํ•˜์—ฌ ๊ฒฐ๋ก ์— ๋” ๊ฐ€๊นŒ์šด ๊ฒƒ์œผ๋กœ ํŒ๋ช…๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 18: "Reddit๊ฐ€ 2023๋…„ 7์›” 1์ผ ์ด์ „์— ์ œ์•ˆ๋œ API ์š”๊ธˆ ๊ฐ€๊ฒฉ ์ฑ…์ •์— ๋Œ€ํ•œ ๋ณ€๊ฒฝ ๋˜๋Š” ์—ฐ๊ธฐ๋ฅผ ๋ฐœํ‘œํ•  ๊ฒƒ์ธ๊ฐ€?"๋ผ๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ์˜ ์‘๋‹ต์ž…๋‹ˆ๋‹ค(ํ•ด๊ฒฐ๋จ). ์ด ์˜ˆ์ธก์€ ์งˆ๋ฌธ์ด ๊ฐœ์„ค๋œ ์ง€ ํ•˜๋ฃจ ํ›„์ธ 2023๋…„ 6์›” 14์ผ์— ๊ฒ€์ƒ‰๋œ ๋‚ ์งœ์— ์ด๋ฃจ์–ด์กŒ์Šต๋‹ˆ๋‹ค. ๋‹น์‹œ ์ปค๋ฎค๋‹ˆํ‹ฐ๋Š” 40%์˜ ์˜ˆ์ธก์„ ๋‚ด๋ ธ๋Š”๋ฐ, ์ด ์‘๋‹ต์€ 20%๋ฅผ ์ œ์‹œํ•˜์—ฌ ๊ฒฐ๋ก ์— ๋” ๊ฐ€๊นŒ์šด ๊ฒƒ์œผ๋กœ ํŒ๋ช…๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 18: A response from our fine-tuned model on the question โ€œWill Reddit announce changes or a delay to its proposed API fee pricing before July 1, 2023?โ€ (resolved NO). The forecast was made at the retrieval date June, 14, 2023, one day after the question was opened. At the time, the community gives a forecast of 40%, where this response suggests 20% which turned out to be closer to the resolution. [|assistant|]
ํ•˜์ดํผ ๋ฉ€ํ‹ฐ ์–ธ์–ด LLM๊ทธ๋ฆผ 0(b)๋ฅผ ํ†ตํ•ด BLOOMZ๋„ "์ƒ๊ฐ" ์–ธ์–ด๋กœ ์˜์–ด๋ฅผ ๊ธฐ๋ณธ์œผ๋กœ ํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ BLOOMZ๋Š” ์ฃผ๋กœ ์˜์–ด๋ฅผ ์‚ฌ๊ณ ์˜ ์–ธ์–ด๋กœ ์‚ฌ์šฉํ•˜๋ฉด์„œ ์ด๋Ÿฌํ•œ ์–ธ์–ด๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์—ฌ๋Ÿฌ ์–ธ์–ด๋ฅผ ์ง€์›ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ทธ๋ฆผ 4๋Š” self-attention ๋ฐ feed-forward ๊ตฌ์กฐ ๋ชจ๋‘์—์„œ ์–ธ์–ด ๊ฐ„ ๋‰ด๋Ÿฐ ์ค‘๋ณต ์ •๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3์— ๋‚˜ํƒ€๋‚œ ๊ฒฐ๊ณผ์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ ์ค‘๋ณต์ด ํ˜„์ €ํžˆ ๊ฐ์†Œํ•˜์—ฌ ๊ฐœ๋ณ„ ์–ธ์–ด๊ฐ€ ๋” ๋†’์€ ์ˆ˜์ค€์˜ ๋…๋ฆฝ์„ฑ์„ ์œ ์ง€ํ•˜๊ณ  ์„œ๋กœ ๋‰ด๋Ÿฐ์„ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ๊ณต์œ ํ•˜์ง€ ์•Š์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
ํ•˜์ดํผ ๋ฉ€ํ‹ฐ ์–ธ์–ด LLM๊ทธ๋ฆผ 0(b)๋ฅผ ํ†ตํ•ด BLOOMZ๋„ "์ƒ๊ฐ" ์–ธ์–ด๋กœ ์˜์–ด๋ฅผ ๊ธฐ๋ณธ์œผ๋กœ ํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ BLOOMZ๋Š” ์ฃผ๋กœ ์˜์–ด๋ฅผ ์‚ฌ๊ณ ์˜ ์–ธ์–ด๋กœ ์‚ฌ์šฉํ•˜๋ฉด์„œ ์ด๋Ÿฌํ•œ ์–ธ์–ด๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์—ฌ๋Ÿฌ ์–ธ์–ด๋ฅผ ์ง€์›ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ทธ๋ฆผ 4๋Š” ์ž๊ธฐ ์ฃผ์˜ ๋ฐ ํ”ผ๋“œ ํฌ์›Œ๋“œ ๊ตฌ์กฐ ๋ชจ๋‘์—์„œ ์–ธ์–ด ๊ฐ„ ๋‰ด๋Ÿฐ ์ค‘๋ณต ์ •๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3์— ๋‚˜ํƒ€๋‚œ ๊ฒฐ๊ณผ์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ ์ค‘๋ณต์ด ํ˜„์ €ํžˆ ๊ฐ์†Œํ•˜์—ฌ ๊ฐœ๋ณ„ ์–ธ์–ด๊ฐ€ ๋” ๋†’์€ ์ˆ˜์ค€์˜ ๋…๋ฆฝ์„ฑ์„ ์œ ์ง€ํ•˜๊ณ  ์„œ๋กœ ๋‰ด๋Ÿฐ์„ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ๊ณต์œ ํ•˜์ง€ ์•Š์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Hyper-Multilingual LLMsFrom Figure 0(b) we observe that BLOOMZ also tends to default to English as its "thinking" language. Therefore, it can support multiple languages mainly because it has the ability to understand and generate these languages while using English as the language of thinking. Additionally, Figure 4 illustrates the degree of neuron overlap among languages within both the self-attention and feed-forward structures of BLOOMZ. In contrast to the findings shown in Figure 3, there is a marked reduction in overlap, indicating that individual languages maintain a higher degree of independence and do not extensively share neurons with one another. [|assistant|]
๊ทธ๋ฆผ 20: "ํŠธ๋Ÿผํ”„๊ฐ€ ์ฒซ RNC ํ† ๋ก ์— ์ฐธ์„ํ•  ๊ฒƒ์ธ๊ฐ€?"๋ผ๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•œ Fine-tuned ๋ชจ๋ธ์˜ ์‘๋‹ต์ž…๋‹ˆ๋‹ค(2023๋…„ 8์›” 23์ผ์— NO๋กœ ํ•ด๊ฒฐ๋จ). ์ด ์˜ˆ์ธก์€ ์งˆ๋ฌธ์ด ๊ฐœ์„ค๋œ ์ง€ 2์ผ ํ›„์ธ 2023๋…„ 6์›” 15์ผ์— Retrieval index 1์—์„œ ์ด๋ฃจ์–ด์กŒ์Šต๋‹ˆ๋‹ค. ๋‹น์‹œ ์ปค๋ฎค๋‹ˆํ‹ฐ๋Š” 60%๋ฅผ ์˜ˆ์ธกํ–ˆ์œผ๋ฉฐ, ์ด ์‘๋‹ต์—์„œ ๋ชจ๋ธ์ด ์˜ˆ์ธกํ•œ 55%์— ๊ฐ€๊นŒ์šด ์ˆ˜์น˜์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 20: "ํŠธ๋Ÿผํ”„๊ฐ€ ์ฒซ RNC ํ† ๋ก ์— ์ฐธ์„ํ•  ๊ฒƒ์ธ๊ฐ€?"๋ผ๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ์˜ ์‘๋‹ต์ž…๋‹ˆ๋‹ค(2023๋…„ 8์›” 23์ผ์— NO๋กœ ํ•ด๊ฒฐ๋จ). ์ด ์˜ˆ์ธก์€ ์งˆ๋ฌธ์ด ๊ฐœ์„ค๋œ ์ง€ 2์ผ ํ›„์ธ 2023๋…„ 6์›” 15์ผ์— ๊ฒ€์ƒ‰ ์ธ๋ฑ์Šค 1์—์„œ ์ด๋ฃจ์–ด์กŒ์Šต๋‹ˆ๋‹ค. ๋‹น์‹œ ์ปค๋ฎค๋‹ˆํ‹ฐ๋Š” 60%๋ฅผ ์˜ˆ์ธกํ–ˆ์œผ๋ฉฐ, ์ด ์‘๋‹ต์—์„œ ๋ชจ๋ธ์ด ์˜ˆ์ธกํ•œ 55%์— ๊ฐ€๊นŒ์šด ์ˆ˜์น˜์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 20: A response from our fine-tuned model on the question โ€œWill Trump attend the first RNC debate?โ€ (resolved NO on August, 23, 2023). The forecast was made at the retrieval index 1 on June 15, 2023, 2 days after the question was opened. At the time, the community predicted 60%, close to our modelโ€™s forecast of 55% from this response. [|assistant|]
๋ชจ๋ธ์˜ ๋ฒ•๋ฅ  ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด (i) 5๊ฐ€์ง€ ์œ ํ˜•์˜ ๋ฒ•๋ฅ  ๋ฌธ์„œ์— ๋Œ€ํ•œ ๋ฐฑ๋ณธ์˜ perplexity๋ฅผ ๋น„๊ตํ•˜๊ณ , (ii) ๋” ๊นŠ์€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด LegalBench๋ฅผ LegalBench-Instruct๋กœ ๊ฐœ์„ ํ•˜๊ณ , (iii) ์ถ”๊ฐ€ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์œ„ํ•ด MMLU์˜ ๋ฒ•๋ฅ  ์„น์…˜์— ์˜์กดํ•˜๋Š” 3๊ฐ€์ง€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋ชจ๋ธ์˜ ๋ฒ•๋ฅ  ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด (i) 5๊ฐ€์ง€ ์œ ํ˜•์˜ ๋ฒ•๋ฅ  ๋ฌธ์„œ์— ๋Œ€ํ•œ ๋ฐฑ๋ณธ์˜ ๋‹นํ˜น๊ฐ์„ ๋น„๊ตํ•˜๊ณ , (ii) ๋” ๊นŠ์€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด LegalBench-Instruct๋กœ LegalBench๋ฅผ ๊ฐœ์„ ํ•˜๊ณ , (iii) ์ถ”๊ฐ€ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์œ„ํ•ด MMLU์˜ ๋ฒ•๋ฅ  ์„น์…˜์— ์˜์กดํ•˜๋Š” 3๊ฐ€์ง€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To evaluate the model's legal abilities, we use \(3\) benchmarks (i) we compare the perplexity of the backbones on \(5\) types of legal documents, (ii) we enhance LegalBench with LegalBench-Instruct for deeper evaluation, (iii) we rely on the legal section of MMLU for additional insights. [|assistant|]
์ตœ์ข… ์ •ํ™•๋„๋Š” ์‚ฌ์ „ ํ•™์Šต ๋ชฉํ‘œ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ์ž˜ ์ธก์ •ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, LLM์˜ ๋Šฅ๋ ฅ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ‰๊ฐ€ ์ž‘์—…์˜ ์„ฑ๋Šฅ์— ๋”ฐ๋ผ ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ๋‹ค๋Š” ์ ์„ ์ „์ œ๋กœ ํ•˜๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ์ด ์–ด๋–ค ์‹์œผ๋กœ๋“  Instruction Tuning, Fine-tuning ๋˜๋Š” ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ์กฐ์ •๋˜์ง€ ์•Š์•˜๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ธ๊ธฐ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ ํ‰๊ฐ€๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์„ ํƒํ•œ ํ‰๊ฐ€ ์ž‘์—…์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ ๋ณธ๋ฌธ์˜ ์„น์…˜ 5.4์™€ ๋ถ€๋ก์˜ ์„น์…˜ A.5๋ฅผ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค.[|endofturn|]
์ตœ์ข… ์ •ํ™•๋„๋Š” ์‚ฌ์ „ ํ•™์Šต ๋ชฉํ‘œ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ์ž˜ ์ธก์ •ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, LLM์˜ ๋Šฅ๋ ฅ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ‰๊ฐ€ ์ž‘์—…์˜ ์„ฑ๋Šฅ์— ๋”ฐ๋ผ ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ๋‹ค๋Š” ์ ์„ ์ „์ œ๋กœ ํ•˜๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ์ด ์–ด๋–ค ์‹์œผ๋กœ๋“  ์ง€์‹œ ํŠœ๋‹, ๋ฏธ์„ธ ์กฐ์ • ๋˜๋Š” ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ์กฐ์ •๋˜์ง€ ์•Š์•˜๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ธ๊ธฐ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ ํ‰๊ฐ€๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์„ ํƒํ•œ ํ‰๊ฐ€ ์ž‘์—…์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ ๋ณธ๋ฌธ์˜ ์„น์…˜ 5.4์™€ ๋ถ€๋ก์˜ ์„น์…˜ A.5๋ฅผ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]While final accuracy provides a good measure of performance on the pre-training objective, LLMs abilities are typically judged by their performance on evaluation tasks. With the caveat that we use base models, that is our models have not been instruction-tuned, fine-tuned, or adapted to human preferences in any way, we present their evaluation on popular benchmarks in this section. We refer the reader to Sec. 5.4 of the main manuscript and Sec. A.5 of the appendix for a more detailed description of the chosen evaluation tasks. [|assistant|]
TwoHopFact์˜ ๊ฐ 2-hop ํ”„๋กฌํ”„ํŠธ \(\tau_{\text{2H}}\)์— ๋Œ€ํ•ด, ๋™์ผํ•œ ํŒฉํŠธ ๊ตฌ์„ฑ ์œ ํ˜•์—์„œ ํ•˜๋‚˜์˜ \(e_{1}^{\prime}\)๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ•˜๊ณ  ๋ฏธ๋ฆฌ ์ •์˜๋œ ํ›„๋ณด ๊ด€๊ณ„ ์ง‘ํ•ฉ(๋ถ€๋ก ํ‘œ 5์— ์ œ๊ณต๋จ)์—์„œ ํ•˜๋‚˜์˜ \(r_{1}^{\prime}\)๋ฅผ ์„ ํƒํ•˜์—ฌ \(\tau_{\text{2H}}^{\prime}\)๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์—”ํ‹ฐํ‹ฐ ๋˜๋Š” ๊ด€๊ณ„ ์น˜ํ™˜์„ ํ†ตํ•ด \(\tau_{\text{2H}}^{\prime}\)๋ฅผ \(\tau_{\text{2H}}\)๋กœ ๋Œ€์ฒดํ•  ๋•Œ \(e_{2}\)์˜ ๋ฆฌ์ฝœ์ด ์ฆ๊ฐ€ํ•˜๋Š” ๊ฒฝ์šฐ์˜ ์ƒ๋Œ€์  ๋นˆ๋„๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. ์ƒ๋Œ€ ๋นˆ๋„๊ฐ€ 0.5 ์ด์ƒ์ด๋ฉด LLM์ด ์ฒซ ๋ฒˆ์งธ hop ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ํ™•๋ฅ ์ด ์ด๋Ÿฌํ•œ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ๋ฌด์ž‘์œ„ ํ™•๋ฅ ์„ ์ดˆ๊ณผํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
TwoHopFact์˜ ๊ฐ 2ํ™‰ ํ”„๋กฌํ”„ํŠธ \(\tau_{\text{2H}}\)์— ๋Œ€ํ•ด, ๋™์ผํ•œ ํŒฉํŠธ ๊ตฌ์„ฑ ์œ ํ˜•์—์„œ ํ•˜๋‚˜์˜ \(e_{1}^{\prime}\)๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ•˜๊ณ  ๋ฏธ๋ฆฌ ์ •์˜๋œ ํ›„๋ณด ๊ด€๊ณ„ ์ง‘ํ•ฉ(๋ถ€๋ก ํ‘œ 5์— ์ œ๊ณต๋จ)์—์„œ ํ•˜๋‚˜์˜ \(r_{1}^{\prime}\)๋ฅผ ์„ ํƒํ•˜์—ฌ \(\tau_{\text{2H}}^{\prime}\)๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์—”ํ‹ฐํ‹ฐ ๋˜๋Š” ๊ด€๊ณ„ ์น˜ํ™˜์„ ํ†ตํ•ด \(\tau_{\text{2H}}^{\prime}\)๋ฅผ \(\tau_{\text{2H}}\)๋กœ ๋Œ€์ฒดํ•  ๋•Œ \(e_{2}\)์˜ ๋ฆฌ์ฝœ์ด ์ฆ๊ฐ€ํ•˜๋Š” ๊ฒฝ์šฐ์˜ ์ƒ๋Œ€์  ๋นˆ๋„๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. ์ƒ๋Œ€ ๋นˆ๋„๊ฐ€ 0.5 ์ด์ƒ์ด๋ฉด LLM์ด ์ฒซ ๋ฒˆ์งธ ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ํ™•๋ฅ ์ด ์ด๋Ÿฌํ•œ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ๋ฌด์ž‘์œ„ ํ™•๋ฅ ์„ ์ดˆ๊ณผํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]For each two-hop prompt \(\tau_{\text{2H}}\) in TwoHopFact, we randomly select one \(e_{1}^{\prime}\) from the same fact composition type and one \(r_{1}^{\prime}\) from a set of predefined candidate relations (provided in Appendix Table 5) to create \(\tau_{\text{2H}}^{\prime}\). We then measure the relative frequency of cases where replacing \(\tau_{\text{2H}}^{\prime}\) with \(\tau_{\text{2H}}\) via entity or relation substitution increases the recall of \(e_{2}\). A relative frequency above 0.5 suggests the LLM's chance to perform first-hop reasoning exceeds the random chance for these prompts. [|assistant|]
ํ‘œ 1์€ ์ผ๋ณธ์–ด ์ˆ˜ํ•™ ๋ฐ ์ „์ฒด ์ผ๋ณธ์–ด ๋ฒค์น˜๋งˆํฌ ๊ณผ์ œ์—์„œ LLM์˜ ์„ฑ๋Šฅ์„ ์š”์•ฝํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. MGSM-JA ์—ด์—๋Š” ์•ž์„œ ์„ค๋ช…ํ•œ metrics๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ MGSM ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ๊ฒฐ๊ณผ๊ฐ€ ๋ณด๊ณ ๋ฉ๋‹ˆ๋‹ค. ์ผ๋ณธ์–ด ๋ชจ๋ธ(Model 1)์€ ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์ด ์ œํ•œ์ ์ธ ๋ฐ˜๋ฉด, ์ˆ˜ํ•™ ๋ชจ๋ธ(Models 2 ๋ฐ 3)์€ ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์€ ๋›ฐ์–ด๋‚˜์ง€๋งŒ ์ผ๋ณธ์–ด์— ๋Œ€ํ•œ ์ง€๋ฐฐ๋ ฅ์ด ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์„ธ ๋ชจ๋ธ ๋ชจ๋‘ MGSM-JA์—์„œ ๋‚ฎ์€ ์ ์ˆ˜๋ฅผ ๋ฐ›์•˜์œผ๋ฉฐ ์ •ํ™•๋„๋Š” 30.0 ์ดํ•˜์ž…๋‹ˆ๋‹ค.[|endofturn|]
ํ‘œ 1์€ ์ผ๋ณธ์–ด ์ˆ˜ํ•™ ๋ฐ ์ „์ฒด ์ผ๋ณธ์–ด ๋ฒค์น˜๋งˆํฌ ๊ณผ์ œ์—์„œ LLM์˜ ์„ฑ๋Šฅ์„ ์š”์•ฝํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. MGSM-JA ์—ด์—๋Š” ์•ž์„œ ์„ค๋ช…ํ•œ ๋ฉ”ํŠธ๋ฆญ์„ ์‚ฌ์šฉํ•˜์—ฌ MGSM ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ๊ฒฐ๊ณผ๊ฐ€ ๋ณด๊ณ ๋ฉ๋‹ˆ๋‹ค. ์ผ๋ณธ์–ด ๋ชจ๋ธ(๋ชจ๋ธ 1)์€ ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์ด ์ œํ•œ์ ์ธ ๋ฐ˜๋ฉด, ์ˆ˜ํ•™ ๋ชจ๋ธ(๋ชจ๋ธ 2 ๋ฐ 3)์€ ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์€ ๋›ฐ์–ด๋‚˜์ง€๋งŒ ์ผ๋ณธ์–ด์— ๋Œ€ํ•œ ์ง€๋ฐฐ๋ ฅ์ด ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์„ธ ๋ชจ๋ธ ๋ชจ๋‘ MGSM-JA์—์„œ ๋‚ฎ์€ ์ ์ˆ˜๋ฅผ ๋ฐ›์•˜์œผ๋ฉฐ ์ •ํ™•๋„๋Š” 30.0 ์ดํ•˜์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Table 1 summarizes the performance of the LLMs on Japanese Math and overall Japanese benchmark tasks. The MGSM-JA column reports the results from the MGSM test set, employing the previously described metrics. The Japanese language model (Model 1) demonstrates limited mathematical proficiency, while the Math models (Models 2 and 3), though mathematically adept, show insufficient command of the Japanese language. Consequently, all three models score low scores on the MGSM-JA, with accuracy at or below 30.0. [|assistant|]
์ด์ œ ๊ณ ์ •๋œ ์œ ํ•œ Batch Size \(B\)์—์„œ Loss ๊ฐ’ \(L\)๊ณผ ํ›ˆ๋ จ ๋‹จ๊ณ„ \(S\) ์‚ฌ์ด์˜ ๊ด€๊ณ„์ธ Eq. 3.13์„ ๋ถ„์„ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ชจ๋“  ์–‘์€ ์ด์ „์— ๊ณ„์‚ฐ๋œ ์ƒ์ˆ˜ ์Šค์นผ๋ผ์ž…๋‹ˆ๋‹ค(\(N_{c},\alpha_{N},S_{c},\alpha_{S},B_{*}\) ๋ฐ \(\alpha_{B}\)). Loss ๊ฐ’ \(L(N,S,B)\)๋ฅผ ๊ณ„์‚ฐํ•˜๊ธฐ ์–ด๋ ค์šด ์ ์€ ๋ฐฉ์ •์‹์˜ ์–‘์ชฝ์— ๋‚˜ํƒ€๋‚˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ \(L(N,S,B)\)๋ฅผ ๋ถ„๋ฆฌํ•˜๋Š” ํ•ด์„์  ํ•ด๋Š” ์—†์ง€๋งŒ, ๊ทผ ์ฐพ๊ธฐ ๋ฐฉ๋ฒ•(์˜ˆ: ์ด๋“ฑ๋ถ„๋ฒ•)์„ ์‚ฌ์šฉํ•˜์—ฌ ์ˆ˜์น˜์ ์œผ๋กœ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 3.13์—์„œ \(L(N,S,B)\)๋Š” ์œ ์ผํ•œ ๋ฏธ์ง€ ์ˆ˜๋Ÿ‰์ด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.[|endofturn|]
์ด์ œ ๊ณ ์ •๋œ ์œ ํ•œ ๋ฐฐ์น˜ ํฌ๊ธฐ \(B\)์—์„œ ์†์‹ค ๊ฐ’ \(L\)๊ณผ ํ›ˆ๋ จ ๋‹จ๊ณ„ \(S\) ์‚ฌ์ด์˜ ๊ด€๊ณ„์ธ Eq. 3.13์„ ๋ถ„์„ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ชจ๋“  ์–‘์€ ์ด์ „์— ๊ณ„์‚ฐ๋œ ์ƒ์ˆ˜ ์Šค์นผ๋ผ์ž…๋‹ˆ๋‹ค(\(N_{c},\์•ŒํŒŒ_{N},S_{c},\์•ŒํŒŒ_{S},B_{*}\) ๋ฐ \(\์•ŒํŒŒ_{B}\)). ์†์‹ค ๊ฐ’ \(L(N,S,B)\)๋ฅผ ๊ณ„์‚ฐํ•˜๊ธฐ ์–ด๋ ค์šด ์ ์€ ๋ฐฉ์ •์‹์˜ ์–‘์ชฝ์— ๋‚˜ํƒ€๋‚˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ \(L(N,S,B)\)๋ฅผ ๋ถ„๋ฆฌํ•˜๋Š” ํ•ด์„์  ํ•ด๋Š” ์—†์ง€๋งŒ, ๊ทผ ์ฐพ๊ธฐ ๋ฐฉ๋ฒ•(์˜ˆ: ์ด๋“ฑ๋ถ„๋ฒ•)์„ ์‚ฌ์šฉํ•˜์—ฌ ์ˆ˜์น˜์ ์œผ๋กœ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 3.13์—์„œ \(L(N,S,B)\)๋Š” ์œ ์ผํ•œ ๋ฏธ์ง€ ์ˆ˜๋Ÿ‰์ด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Now, let us analyze Eq. 3.13, which is the relation between the loss value \(L\) and the training step \(S\) under a fixed finite batch size \(B\). All other quantities are constant scalars that have been worked out before (\(N_{c},\alpha_{N},S_{c},\alpha_{S},B_{*}\) and \(\alpha_{B}\)). The difficulty of computing the loss value \(L(N,S,B)\) is that it appears on both sides of the equation. Although there is no analytical solution to isolate \(L(N,S,B)\) here, we can numerically estimate \(L(N,S,B)\) by any root-finding method (_e.g._, the bisection method) because it is the only unknown quantity in Eq. 3.13. [|assistant|]
์•…์˜์ ์ธ ํ™œ๋™์„ ์˜๋„ํ•˜๋Š” ์„ฑ์ธ ์‚ฌ์šฉ์ž, ์ •๊ตํ•˜์ง€ ์•Š์€ ๋ฐฉ์‹์œผ๋กœ ํ–‰๋™ํ•˜๋Š” ์‚ฌ์šฉ์ž์ด ํŽ˜๋ฅด์†Œ๋‚˜๋Š” ํ•ด๋ฅผ ๋ผ์น˜๋ ค๋Š” ์˜๋„๋‚˜ ์˜์ง€๊ฐ€ ์žˆ๋‹ค๋Š” ์ ์—์„œ ์•…์˜์ ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ, ํ•ด๋ฅผ ์ž…ํžˆ๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ๋„๋ฉ”์ธ ์ง€์‹์ด ์ œํ•œ๋˜์–ด ์žˆ์–ด ์œ ํ•ดํ•œ ํ–‰๋™์— ๋Œ€ํ•œ ๊ธฐ๋ณธ์ ์ธ ์งˆ๋ฌธ์„ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ AI ๋ชจ๋ธ๊ณผ ์•ˆ์ „ ์กฐ์น˜๊ฐ€ ์–ด๋–ป๊ฒŒ ์œ„ํƒœ๋กœ์šธ ์ˆ˜ ์žˆ๋Š”์ง€์— ๋Œ€ํ•œ ์ดํ•ด๋„ ๋ฏธ๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋“ค์€ jailbreaking์ด๋‚˜ prompt injection๊ณผ ๊ฐ™์€ ๊ณ ๊ธ‰ ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.[|endofturn|]
์•…์˜์ ์ธ ํ™œ๋™์„ ์˜๋„ํ•˜๋Š” ์„ฑ์ธ ์‚ฌ์šฉ์ž, ์ •๊ตํ•˜์ง€ ์•Š์€ ๋ฐฉ์‹์œผ๋กœ ํ–‰๋™ํ•˜๋Š” ์‚ฌ์šฉ์ž์ด ํŽ˜๋ฅด์†Œ๋‚˜๋Š” ํ•ด๋ฅผ ๋ผ์น˜๋ ค๋Š” ์˜๋„๋‚˜ ์˜์ง€๊ฐ€ ์žˆ๋‹ค๋Š” ์ ์—์„œ ์•…์˜์ ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ, ํ•ด๋ฅผ ์ž…ํžˆ๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ๋„๋ฉ”์ธ ์ง€์‹์ด ์ œํ•œ๋˜์–ด ์žˆ์–ด ์œ ํ•ดํ•œ ํ–‰๋™์— ๋Œ€ํ•œ ๊ธฐ๋ณธ์ ์ธ ์งˆ๋ฌธ์„ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ AI ๋ชจ๋ธ๊ณผ ์•ˆ์ „ ์กฐ์น˜๊ฐ€ ์–ด๋–ป๊ฒŒ ์œ„ํƒœ๋กœ์šธ ์ˆ˜ ์žˆ๋Š”์ง€์— ๋Œ€ํ•œ ์ดํ•ด๋„ ๋ฏธ๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋“ค์€ ํƒˆ์˜ฅ์ด๋‚˜ ํ”„๋กฌํ”„ํŠธ ์‚ฝ์ž…๊ณผ ๊ฐ™์€ ๊ณ ๊ธ‰ ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Adult user intent on malicious activities, behaving in a non-sophisticated wayThis persona is malicious, in the sense that they have intent or willingness to inflict harm. Specifically, they possess limited domain knowledge on how to cause harm, which results in them asking basic questions about harmful actions. They also have minimal understanding of AImodels and how their safety measures could be compromised. As a result, they do not employ advanced techniques such as jailbreaking or prompt injection. [|assistant|]
๊ฒฐ๋ก ์ ์œผ๋กœ, fertility ๋ฐ parity์™€ ๊ฐ™์€ ๋‚ด์žฌ์  tokenizer ๋ฉ”ํŠธ๋ฆญ์€ ์กฐ์‹ฌํ•ด์„œ ์‚ฌ์šฉํ•ด์•ผ ํ•˜๋ฉฐ, ํŠน์ • ๋ฒ”์œ„ ๋‚ด์—์„œ๋งŒ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‚ฎ์€ fertility ์ ์ˆ˜๋Š” ํ•„์š”ํ•œ ๊ธฐ์ค€์œผ๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ์ถฉ๋ถ„ํ•œ ๊ธฐ์ค€์œผ๋กœ ๊ฐ„์ฃผํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ฒฐ๋ก ์ ์œผ๋กœ, ๋น„์˜ฅ๋„ ๋ฐ ํŒจ๋ฆฌํ‹ฐ์™€ ๊ฐ™์€ ๋‚ด์žฌ์  ํ† ํฐํ™” ๋ฉ”ํŠธ๋ฆญ์€ ์กฐ์‹ฌํ•ด์„œ ์‚ฌ์šฉํ•ด์•ผ ํ•˜๋ฉฐ, ํŠน์ • ๋ฒ”์œ„ ๋‚ด์—์„œ๋งŒ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‚ฎ์€ ์ถœ์‚ฐ์œจ ์ ์ˆ˜๋Š” ํ•„์š”ํ•œ ๊ธฐ์ค€์œผ๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ์ถฉ๋ถ„ํ•œ ๊ธฐ์ค€์œผ๋กœ ๊ฐ„์ฃผํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In conclusion, intrinsic tokenizer metrics such as fertility and parity need to be taken with a grain of salt and supposedly are only predictive of downstream model performance in certain bounds. Low fertility scores might be regarded as a necessary criterion but not as a sufficient one. [|assistant|]
์ค‘๋ณต ์ œ๊ฑฐํ•„ํ„ฐ๋ง ํ›„, ์šฐ๋ฆฌ๋Š” Penedo ๋“ฑ(2023) [56]์˜ ์ ˆ์ฐจ์— ๋”ฐ๋ผ ํฌ๊ด„์ ์ธ deduplication ํŒŒ์ดํ”„๋ผ์ธ์„ ๊ตฌํ˜„ํ•ฉ๋‹ˆ๋‹ค. ์ด ํŒŒ์ดํ”„๋ผ์ธ์€ ๋ฌธ์„œ ์ˆ˜์ค€์˜ MinHash deduplication๊ณผ ํ•˜์œ„ ๋ฌธ์„œ exact-match deduplication์„ ํ†ตํ•ฉํ•˜์—ฌ ๋ฌธ์„œ ๋‚ด ๋ฐ ๋ฌธ์„œ ๊ฐ„์˜ ์ค‘๋ณต ์ฝ˜ํ…์ธ ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‰ด์Šค, ๊ด‘๊ณ , ์ง€์‹ ๊ธฐ๋ฐ˜ ์ฝ˜ํ…์ธ ์™€ ๊ฐ™์€ ๋ ˆ์ด๋ธ”์„ ์˜ˆ์ธกํ•˜๋Š” topic model์„ ์‚ฌ์šฉํ•˜์—ฌ ์›น ๋ฌธ์„œ๋ฅผ ํŠน์ • ํ…Œ๋งˆ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ข… pretraining ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋Š” ์ •๋ณด ๋ฐ€๋„๋ฅผ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ๋Œ€๋ถ€๋ถ„ ๊ด‘๊ณ ์ธ ๋œ ์œ ์šฉํ•œ ์ฝ˜ํ…์ธ ๋ฅผ down-samplingํ•ฉ๋‹ˆ๋‹ค. Yi์˜ pretraining ๋ฐ์ดํ„ฐ์˜ ์ตœ์ข… ๊ตฌ์„ฑ์€ ๊ทธ๋ฆผ 2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ค‘๋ณต ์ œ๊ฑฐํ•„ํ„ฐ๋ง ํ›„, ์šฐ๋ฆฌ๋Š” Penedo ๋“ฑ(2023) [56]์˜ ์ ˆ์ฐจ์— ๋”ฐ๋ผ ํฌ๊ด„์ ์ธ ์ค‘๋ณต ์ œ๊ฑฐ ํŒŒ์ดํ”„๋ผ์ธ์„ ๊ตฌํ˜„ํ•ฉ๋‹ˆ๋‹ค. ์ด ํŒŒ์ดํ”„๋ผ์ธ์€ ๋ฌธ์„œ ์ˆ˜์ค€์˜ MinHash ์ค‘๋ณต ์ œ๊ฑฐ์™€ ํ•˜์œ„ ๋ฌธ์„œ ์ •ํ™• ์ผ์น˜ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ๋ฌธ์„œ ๋‚ด ๋ฐ ๋ฌธ์„œ ๊ฐ„์˜ ์ค‘๋ณต ์ฝ˜ํ…์ธ ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‰ด์Šค, ๊ด‘๊ณ , ์ง€์‹ ๊ธฐ๋ฐ˜ ์ฝ˜ํ…์ธ ์™€ ๊ฐ™์€ ๋ ˆ์ด๋ธ”์„ ์˜ˆ์ธกํ•˜๋Š” ํ† ํ”ฝ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์›น ๋ฌธ์„œ๋ฅผ ํŠน์ • ํ…Œ๋งˆ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ข… ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋Š” ์ •๋ณด ๋ฐ€๋„๋ฅผ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ๋Œ€๋ถ€๋ถ„ ๊ด‘๊ณ ์ธ ๋œ ์œ ์šฉํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ๋‹ค์šด ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค. Yi์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ์ตœ์ข… ๊ตฌ์„ฑ์€ ๊ทธ๋ฆผ 2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]DeduplicationAfter filtering, we implement a comprehensive deduplication pipeline following the procedure in Penedo et al. (2023) [56]. This pipeline integrates document-level MinHash deduplication and sub-document exact-match deduplication, effectively identifying and removing duplicate content within and across documents. We further categorize web documents into specific themes using a topic model predicting labels like as news, ads, and knowledge-based content. In the final pretraining dataset, we down-sample less helpful content, mostly advertisements, to ensure information density. The final composition of Yi's pretraining data is shown in Fig. 2. [|assistant|]
๊ณ„์ธต์  Encoder์™€ ๊ณ„๋‹จ์‹ ์„ ํƒ๊ธฐ์˜ ์•„ํ‚คํ…์ฒ˜, ์ž…๋ ฅ, ์ถœ๋ ฅ, ํŒŒ๋ผ๋ฏธํ„ฐ ๋“ฑ ์–ด๋–ป๊ฒŒ ๋” ์ž์„ธํ•œ ๋‚ด์šฉ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”?[|endofturn|]
๊ณ„์ธต์  ์ธ์ฝ”๋”์™€ ๊ณ„๋‹จ์‹ ์„ ํƒ๊ธฐ์˜ ์•„ํ‚คํ…์ฒ˜, ์ž…๋ ฅ, ์ถœ๋ ฅ, ํŒŒ๋ผ๋ฏธํ„ฐ ๋“ฑ ์–ด๋–ป๊ฒŒ ๋” ์ž์„ธํ•œ ๋‚ด์šฉ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”?[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]in detail. How can you provide more details of the hierarchical encoder and the cascaded selectors, such as their architectures, inputs, outputs, and parameters? [|assistant|]
๋ฐ์ดํ„ฐ ๋งค์นญ baseline์„ ๋ณด๋ฉด ๋ฐ์ดํ„ฐ ๋ฐ”์ธ๋”ฉ๋œ ๊ฒฝ์šฐ reverse training์ด ํ‘œ์ค€ ์ž‘์—…์—๋„ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. baseline์ด ํšจ๊ณผ์ ์œผ๋กœ ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋Š” compute-matched ์‚ฌ๋ก€์—์„œ๋Š” 5, 10๊ฐœ์˜ ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด reverse training์ด ์•ฝ๊ฐ„ ๋’ค์ฒ˜์ง‘๋‹ˆ๋‹ค. ์ค‘์š”ํ•œ ๊ฒƒ์€ ๋ฐ์ดํ„ฐ ๋งค์นญ๊ณผ compute-matched ๋ชจ๋‘์—์„œ baseline์— ๋น„ํ•ด reverse training์˜ ์—ญ ๋ฐฉํ–ฅ ์งˆ๋ฌธ์ด ํฌ๊ฒŒ ๊ฐœ์„ ๋˜์—ˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” pre-training ๋‹จ๊ณ„์—์„œ reverse training์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ reversal curse์— ๋Œ€ํ•ด robustํ•˜๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
๋ฐ์ดํ„ฐ ๋งค์นญ ๊ธฐ์ค€์„ ์„ ๋ณด๋ฉด ๋ฐ์ดํ„ฐ ๋ฐ”์ธ๋”ฉ๋œ ๊ฒฝ์šฐ ์—ญ ํ›ˆ๋ จ์ด ํ‘œ์ค€ ์ž‘์—…์—๋„ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์ค€์„ ์ด ํšจ๊ณผ์ ์œผ๋กœ ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋Š” ๊ณ„์‚ฐ ๋งค์นญ ์‚ฌ๋ก€์—์„œ๋Š” 5, 10๊ฐœ์˜ ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด ์—ญ ํ›ˆ๋ จ์ด ์•ฝ๊ฐ„ ๋’ค์ฒ˜์ง‘๋‹ˆ๋‹ค. ์ค‘์š”ํ•œ ๊ฒƒ์€ ๋ฐ์ดํ„ฐ ๋งค์นญ๊ณผ ๊ณ„์‚ฐ ๋งค์นญ ๋ชจ๋‘์—์„œ ๊ธฐ์ค€์„ ์— ๋น„ํ•ด ์—ญ ํ›ˆ๋ จ์˜ ์—ญ ๋ฐฉํ–ฅ ์งˆ๋ฌธ์ด ํฌ๊ฒŒ ๊ฐœ์„ ๋˜์—ˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์—์„œ ์—ญ ํ›ˆ๋ จ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ์—ญ์ „์˜ ์ €์ฃผ์— ๋Œ€ํ•ด ๊ฒฌ๊ณ ํ•˜๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]the data-matched baseline, showing that in the data-bound case, reverse training even helps on standard tasks. In the compute-matched case, where the baseline has effectively access to more data, reversal training lags slightly behind for 5, 10 samples. Importantly, in both the data-matched and compute-matched case we see significant improvement in the reverse direction questions for reverse training compared to either baseline. This demonstrates that reverse training can be employed during the pre-training stage to make the model robust against the reversal curse. [|assistant|]
์‹ค์ œ๋กœ ์‚ฌ์ „ ํ•™์Šต LLM์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹จ์ผ ๋ง๋ญ‰์น˜๊ฐ€ ์•„๋‹Œ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ํ˜ผํ•ฉ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 6 ์ฐธ์กฐ). ๋”ฐ๋ผ์„œ ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ธฐ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ(์˜ˆ: C4, OpenWebText, Pile)๋ฅผ ํ˜ผํ•ฉํ•œ ๋‹ค์Œ ์ถ”๊ฐ€ ์ฒ˜๋ฆฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ pre-training ๋ง๋ญ‰์น˜๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํŠน์ • ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์ ํ•ฉํ•œ LLM์„ ํ›ˆ๋ จํ•˜๋ ค๋ฉด pre-training ๋ฐ์ดํ„ฐ์˜ ํ•ด๋‹น ์ •๋ณด๋ฅผ ๋ณด๊ฐ•ํ•˜๊ธฐ ์œ„ํ•ด ๊ด€๋ จ ์†Œ์Šค(์˜ˆ: Wikipedia ๋ฐ BigQuery)์—์„œ ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๊ฒƒ๋„ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์‹ค์ œ๋กœ ์‚ฌ์ „ ํ•™์Šต LLM์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹จ์ผ ๋ง๋ญ‰์น˜๊ฐ€ ์•„๋‹Œ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ํ˜ผํ•ฉ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 6 ์ฐธ์กฐ). ๋”ฐ๋ผ์„œ ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ธฐ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ(์˜ˆ: C4, OpenWebText, Pile)๋ฅผ ํ˜ผํ•ฉํ•œ ๋‹ค์Œ ์ถ”๊ฐ€ ์ฒ˜๋ฆฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต ๋ง๋ญ‰์น˜๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํŠน์ • ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์ ํ•ฉํ•œ LLM์„ ํ›ˆ๋ จํ•˜๋ ค๋ฉด ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํ•ด๋‹น ์ •๋ณด๋ฅผ ๋ณด๊ฐ•ํ•˜๊ธฐ ์œ„ํ•ด ๊ด€๋ จ ์†Œ์Šค(์˜ˆ: Wikipedia ๋ฐ BigQuery)์—์„œ ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๊ฒƒ๋„ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In practice, it commonly requires a mixture of different data sources for pre-training LLMs (see Figure 6), instead of a single corpus. Therefore, existing studies commonly mix several ready-made datasets (_e.g._, C4, OpenWebText, and the Pile), and then perform further processing to obtain the pre-training corpus. Furthermore, to train the LLMs that are adaptive to specific applications, it is also important to extract data from relevant sources (_e.g._, Wikipedia and BigQuery) for enriching the corresponding information in pre-training data. [|assistant|]
์‹คํ—˜ 3์—์„œ๋Š” "Respond with <answer> when you see <question>" ํ˜•์‹์˜ question-answering instructions์— ๋Œ€ํ•ด LLM์„ Fine-tuningํ•˜๊ณ  "Q: <question> A: <answer>"๋กœ์˜ ์ผ๋ฐ˜ํ™”๋ฅผ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ 1์—์„œ์™€ ์œ ์‚ฌํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์‹คํ—˜ 3์—์„œ๋Š” "Q: <์งˆ๋ฌธ> A: <๋‹ต>" ํ˜•์‹์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์ง€์‹œ์— ๋Œ€ํ•ด LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๊ณ  ์ผ๋ฐ˜ํ™”๋ฅผ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ 1์—์„œ์™€ ์œ ์‚ฌํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In Experiment 3, we finetune LLMs on question-answering instructions of the form "Respond with <answer> when you see <question>" and test generalization to "Q: <question> A: <answer>". We find results similar to those in Experiment 1. [|assistant|]
**MODEL ANSWER:** ๋น„์˜ฌ๋ ˆํƒ€๋Š” 5 * 2 = <<5*2=10>>10 ํฌ๋ ˆ์šฉ์„ ์‚ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” 20 - 10 = $*20-10=10>>10์˜ ์ž”๋ˆ์„ ๋ฐ›๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
**๋ชจ๋ธ ๋‹ต๋ณ€:** ๋น„์˜ฌ๋ ˆํƒ€๋Š” 5 * 2 = <<5*2=10>>10 ํฌ๋ ˆ์šฉ์„ ์‚ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” 20 - 10 = $*20-10=10>>10์˜ ์ž”๋ˆ์„ ๋ฐ›๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**MODEL ANSWER:** Violetta needs to buy 5 * 2 = <<5*2=10>>10 crayons. She will get change of 20 - 10 = $*20-10=10>>10. [|assistant|]
๋„ค๋ชจํŠธ๋ก -4 15B์˜ ์ธ์ƒ์ ์ธ ๋‹ค๊ตญ์–ด ๋Šฅ๋ ฅ์„ ๋”์šฑ ์ž…์ฆํ•˜๋“ฏ, ํ‘œ 9๋Š” MGSM์—์„œ์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ชจ๋“  chain-of-thought ์„ค๋ช…์ด ๊ณผ์ œ์˜ ์–ธ์–ด๊ฐ€ ์•„๋‹Œ ์˜์–ด๋กœ ๋ชจ๋ธ์— ์ œ์‹œ๋˜๋Š” (Shi et al., 2022)์—์„œ ์†Œ๊ฐœ๋œ ์˜์–ด chain-of-thought ์„ค์ •์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ๊ณผ ๋‹ค๊ตญ์–ด ๋Šฅ๋ ฅ์˜ ๊ต์ฐจ์ ์„ ํ‰๊ฐ€ํ•˜๋Š” ์ด ์–ด๋ ค์šด ๊ณผ์ œ์—์„œ Nemotron-4 15B๋Š” ๋น„๊ต ๋Œ€์ƒ ๋ชจ๋ธ ์ค‘ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์œผ๋ฉฐ, ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ ์ˆ˜๋ณด๋‹ค ์•ฝ 30% ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋„ค๋ชจํŠธ๋ก -4 15B์˜ ์ธ์ƒ์ ์ธ ๋‹ค๊ตญ์–ด ๋Šฅ๋ ฅ์„ ๋”์šฑ ์ž…์ฆํ•˜๋“ฏ, ํ‘œ 9๋Š” MGSM์—์„œ์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ชจ๋“  ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ์„ค๋ช…์ด ๊ณผ์ œ์˜ ์–ธ์–ด๊ฐ€ ์•„๋‹Œ ์˜์–ด๋กœ ๋ชจ๋ธ์— ์ œ์‹œ๋˜๋Š” (Shi et al., 2022)์—์„œ ์†Œ๊ฐœ๋œ ์˜์–ด ์‚ฌ์Šฌ-of-thought ์„ค์ •์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ๊ณผ ๋‹ค๊ตญ์–ด ๋Šฅ๋ ฅ์˜ ๊ต์ฐจ์ ์„ ํ‰๊ฐ€ํ•˜๋Š” ์ด ์–ด๋ ค์šด ๊ณผ์ œ์—์„œ Nemotron-4 15B๋Š” ๋น„๊ต ๋Œ€์ƒ ๋ชจ๋ธ ์ค‘ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์œผ๋ฉฐ, ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ ์ˆ˜๋ณด๋‹ค ์•ฝ 30% ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Further demonstrating the impressive multilingual ability of Nemotron-4 15B, Table 9 shows the performance on MGSM. We report using the English chain-of-thought setting introduced in (Shi et al., 2022) where all chain of thought explanations are presented to the model in English rather than in the language of the task. On this challenging task which assesses the intersection of mathematical and multilingual ability, Nemotron-4 15B achieves the best performance amongst compared models and improves upon the closest score by nearly 30%. [|assistant|]
# ์ถœ๋ ฅ ์ถ”๋ก  ์‹œ๊ฐ„์—๋Š” 3. ๊ทธ๋ฆผ์˜ ํ˜•์‹์— ๋”ฐ๋ผ ์งˆ๋ฌธ์ด ์ฃผ์–ด์ง€๋ฉด greedy decoding์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜์˜ Llama-2๋ฅผ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด QA ์Œ 5๊ฐœ๋ฅผ In-Context ์˜ˆ์ œ๋กœ ์ถ”๊ฐ€ํ•˜์—ฌ QA ํ˜•์‹์„ ๋”ฐ๋ฅด๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€๋ถ€๋ถ„์˜ ์งˆ๋ฌธ์€ ๋‹จ์ˆœํ•œ ํŒฉํ† ์ด๋“œ ์งˆ๋ฌธ์ด๊ณ  ๋Œ€๋ถ€๋ถ„์˜ ๋‹ต์€ ๋น„๊ต์  ์งง๊ธฐ ๋•Œ๋ฌธ์— ์ •๊ทœํ™”(์˜ˆ: ๊ด€์‚ฌ ๋ฐ ๊ตฌ๋‘์  ์ œ๊ฑฐ) ํ›„ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ์ด ๊ธˆ ๋‹ต๊ณผ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ์ธก์ •ํ•˜๋Š” exact match (EM)๋ฅผ ๊ธฐ๋ณธ ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค Kwiatkowski et al. (2019). ๋” ๊ธด ์‘๋‹ต์„ ํ‰๊ฐ€ํ•˜๊ณ  ์‚ฌ์†Œํ•œ ์–ดํœ˜ ์ฐจ์ด๋ฅผ ์ˆ˜์šฉํ•˜๊ธฐ ์œ„ํ•ด ๊ธˆ ๋‹ต์ด ๋ชจ๋ธ์˜ ์ถœ๋ ฅ์— ๋‚˜ํƒ€๋‚˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋‹ต๋ณ€ ํšŒ์ˆ˜์™€ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ๊ณผ ๊ธˆ ๋‹ต ์‚ฌ์ด์˜ ๊ฐ€์žฅ ๊ธด ๊ณตํ†ต ์ˆ˜์—ด์„ ์ธก์ •ํ•˜๋Š” ROUGE-L๋„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ถ”๋ก  ์‹œ๊ฐ„์—๋Š” 3. ๊ทธ๋ฆผ์˜ ํ˜•์‹์— ๋”ฐ๋ผ ์งˆ๋ฌธ์ด ์ฃผ์–ด์ง€๋ฉด ํƒ์š•์Šค๋Ÿฌ์šด ๋””์ฝ”๋”ฉ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜์˜ Llama-2๋ฅผ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด QA ์Œ 5๊ฐœ๋ฅผ ์ธ์ปจํ…์ŠคํŠธ ์˜ˆ์ œ๋กœ ์ถ”๊ฐ€ํ•˜์—ฌ QA ํ˜•์‹์„ ๋”ฐ๋ฅด๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€๋ถ€๋ถ„์˜ ์งˆ๋ฌธ์€ ๋‹จ์ˆœํ•œ ํŒฉํ† ์ด๋“œ ์งˆ๋ฌธ์ด๊ณ  ๋Œ€๋ถ€๋ถ„์˜ ๋‹ต์€ ๋น„๊ต์  ์งง๊ธฐ ๋•Œ๋ฌธ์— ์ •๊ทœํ™”(์˜ˆ: ๊ด€์‚ฌ ๋ฐ ๊ตฌ๋‘์  ์ œ๊ฑฐ) ํ›„ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ์ด ๊ธˆ ๋‹ต๊ณผ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ์ธก์ •ํ•˜๋Š” ์ •ํ™•๋„ ์ผ์น˜(EM)๋ฅผ ๊ธฐ๋ณธ ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค Kwiatkowski et al. (2019). ๋” ๊ธด ์‘๋‹ต์„ ํ‰๊ฐ€ํ•˜๊ณ  ์‚ฌ์†Œํ•œ ์–ดํœ˜ ์ฐจ์ด๋ฅผ ์ˆ˜์šฉํ•˜๊ธฐ ์œ„ํ•ด ๊ธˆ ๋‹ต์ด ๋ชจ๋ธ์˜ ์ถœ๋ ฅ์— ๋‚˜ํƒ€๋‚˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋‹ต๋ณ€ ํšŒ์ˆ˜์™€ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ๊ณผ ๊ธˆ ๋‹ต ์‚ฌ์ด์˜ ๊ฐ€์žฅ ๊ธด ๊ณตํ†ต ์ˆ˜์—ด์„ ์ธก์ •ํ•˜๋Š” ROUGE-L๋„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]At inference time, we use greedy decoding to generate answers given questions as context, following the format in Fig. 3. To evaluate the original Llama-2, we add 5 QA pairs as in-context exemplars to make sure it follows the QA format. Since most questions are simple factoid questions and most answers are relatively short, we use exact match (EM) as our primary metric Kwiatkowski et al. (2019), which measures whether the model's output matches the gold answer exactly after normalization (e.g., remove articles and punctuations). To assess longer responses and accommodate minor lexical differences, we also report answer recall, which assesses if the gold answer appears in the model's output, and ROUGE-L, which measures the longest common subsequence between the model's output and the gold answer. [|assistant|]
๊ฐ์ฃผ 12: ์ด๋Ÿฌํ•œ ๋ฉ”ํŠธ๋ฆญ์€ ๋ถ€๋ถ„์ ์œผ๋กœ Carbonara์˜ AI ์—์ด์ „ํŠธ ๋ฐ ๋ชจ๋‹ˆํ„ฐ๋ง ํ”Œ๋žซํผ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ˆ˜์ง‘๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ฐ์ฃผ 12: ์ด๋Ÿฌํ•œ ๋ฉ”ํŠธ๋ฆญ์€ ๋ถ€๋ถ„์ ์œผ๋กœ ์นด๋ฅด๋ณด๋‚˜๋ผ์˜ AI ์—์ด์ „ํŠธ ๋ฐ ๋ชจ๋‹ˆํ„ฐ๋ง ํ”Œ๋žซํผ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ˆ˜์ง‘๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Footnote 12: These metrics were in part collected using Carbonaraโ€™s AI agent and monitoring platform. [|assistant|]
์ง€์‹์€ ๋‹ค์–‘ํ•œ ํ˜•ํƒœ์˜ ํ•™์Šต์„ ํ†ตํ•ด ์ „๋‹ฌ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: response distillation, feature distillation, API distillation์ž…๋‹ˆ๋‹ค. Response distillation์€ ๊ต์‚ฌ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ์—๋งŒ ๊ด€์‹ฌ์„ ๋‘๊ณ  ํ•™์ƒ ๋ชจ๋ธ์ด ๊ต์‚ฌ์ฒ˜๋Ÿผ ์ •ํ™•ํ•˜๊ฑฐ๋‚˜ ์ ์–ด๋„ ์œ ์‚ฌํ•˜๊ฒŒ (์˜ˆ์ธก์˜ ์˜๋ฏธ์—์„œ) ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๊ฐ€๋ฅด์น˜๋ ค๊ณ  ๋…ธ๋ ฅํ•ฉ๋‹ˆ๋‹ค. Feature distillation์€ ๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ค‘๊ฐ„ ๋ ˆ์ด์–ด๋„ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์ƒ ๋ชจ๋ธ์— ๋” ๋‚˜์€ ๋‚ด๋ถ€ ํ‘œํ˜„์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์ž‘์€ ๋ชจ๋ธ์ด ๊ต์‚ฌ ๋ชจ๋ธ๊ณผ ์œ ์‚ฌํ•œ ํ‘œํ˜„์„ ๊ฐ–๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ง€์‹์€ ๋‹ค์–‘ํ•œ ํ˜•ํƒœ์˜ ํ•™์Šต์„ ํ†ตํ•ด ์ „๋‹ฌ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: ์‘๋‹ต ์ฆ๋ฅ˜, ํŠน์ง• ์ฆ๋ฅ˜, API ์ฆ๋ฅ˜์ž…๋‹ˆ๋‹ค. ์‘๋‹ต ์ฆ๋ฅ˜๋Š” ๊ต์‚ฌ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ์—๋งŒ ๊ด€์‹ฌ์„ ๋‘๊ณ  ํ•™์ƒ ๋ชจ๋ธ์ด ๊ต์‚ฌ์ฒ˜๋Ÿผ ์ •ํ™•ํ•˜๊ฑฐ๋‚˜ ์ ์–ด๋„ ์œ ์‚ฌํ•˜๊ฒŒ (์˜ˆ์ธก์˜ ์˜๋ฏธ์—์„œ) ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๊ฐ€๋ฅด์น˜๋ ค๊ณ  ๋…ธ๋ ฅํ•ฉ๋‹ˆ๋‹ค. ํŠน์ง• ์ฆ๋ฅ˜๋Š” ๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ค‘๊ฐ„ ๋ ˆ์ด์–ด๋„ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์ƒ ๋ชจ๋ธ์— ๋” ๋‚˜์€ ๋‚ด๋ถ€ ํ‘œํ˜„์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์ž‘์€ ๋ชจ๋ธ์ด ๊ต์‚ฌ ๋ชจ๋ธ๊ณผ ์œ ์‚ฌํ•œ ํ‘œํ˜„์„ ๊ฐ–๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Knowledge can be transferred by different forms of learning: response distillation, feature distillation, and API distillation. Response distillation is concerned only with the outputs of the teacher model and tries to teach the student model how to exactly or at least similarly perform (in the sense of prediction) as the teacher. Feature distillation not only uses the last layer but also intermediate layers as well to create a better inner representation for the student model. This helps the smaller model to have a similar representation as the teacher model. [|assistant|]
์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋Š” ๋ชจ๋ธ์ด ๋ฌธ๋งฅ์—์„œ ์ˆซ์ž๋ฅผ ์ถ”์ถœํ•˜๊ฑฐ๋‚˜ ์‚ฐ์ˆ  ๊ณ„์‚ฐ์„ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•˜๋ฏ€๋กœ "๋ฌธ๋งฅ์—์„œ ์ˆซ์ž๋งŒ ๋‹ตํ•˜๊ฑฐ๋‚˜ +, -, *, /๋ฅผ ์‚ฌ์šฉํ•œ ์ˆ˜ํ•™ ์‚ฐ์ˆ ๋งŒ ๋‹ตํ•˜์„ธ์š”."๋ฅผ ConvFinQA์— ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ arithmetic ๊ณต์‹์„ ์ถ”์ถœํ•˜๊ณ  ๊ณ„์‚ฐ๊ธฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข… ๊ฒฐ๊ณผ๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋Š” ๋ชจ๋ธ์ด ๋ฌธ๋งฅ์—์„œ ์ˆซ์ž๋ฅผ ์ถ”์ถœํ•˜๊ฑฐ๋‚˜ ์‚ฐ์ˆ  ๊ณ„์‚ฐ์„ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•˜๋ฏ€๋กœ "๋ฌธ๋งฅ์—์„œ ์ˆซ์ž๋งŒ ๋‹ตํ•˜๊ฑฐ๋‚˜ +, -, *, /๋ฅผ ์‚ฌ์šฉํ•œ ์ˆ˜ํ•™ ์‚ฐ์ˆ ๋งŒ ๋‹ตํ•˜์„ธ์š”."๋ฅผ ConvFinQA์— ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ ์‚ฐ์ˆ  ๊ณต์‹์„ ์ถ”์ถœํ•˜๊ณ  ๊ณ„์‚ฐ๊ธฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข… ๊ฒฐ๊ณผ๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We use "Answer the following question with just a number from context or just the math arithmetic using +, -, *, or /." for ConvFinQA, since this dataset requires the model to either extract numbers from the context or do arithmetic calculation. We extract the arithmetic formula generated by the model and use a calculator to get its final result. [|assistant|]
๋˜ํ•œ ํ‘œ 5๋Š” ์–ธ์–ด์— ๋”ฐ๋ผ BPE ๋˜๋Š” Unigram ์ค‘ ์–ด๋Š ์ชฝ์ด ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š”์ง€ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ฒŒ๋ฅด๋งŒ์–ด์™€ ์˜์–ด๋Š” BPE ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์ด์ ์„ ๋ˆ„๋ ธ๋‹ค๋Š” ์ ์ด ์ฃผ๋ชฉํ•  ๋งŒํ•ฉ๋‹ˆ๋‹ค, ๋ฐ˜๋ฉด ํ”„๋ž‘์Šค์–ด์™€ ์ŠคํŽ˜์ธ์–ด๋Š” Unigram์˜ ์ด์ ์„ ๋ˆ„๋ ธ์Šต๋‹ˆ๋‹ค. ๋กœ๋งˆ์–ด์ธ ์ดํƒˆ๋ฆฌ์•„์–ด์— ๋Œ€ํ•œ ์‹คํ—˜์€ ๋‹ค๋ฅธ ๋‘ ๋กœ๋งˆ์–ด์ธ ์–ธ์–ด์™€ ๋‹ค๋ฅธ ํŒจํ„ด์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
๋˜ํ•œ ํ‘œ 5๋Š” ์–ธ์–ด์— ๋”ฐ๋ผ BPE ๋˜๋Š” ์œ ๋‹ˆ๊ทธ๋žจ ์ค‘ ์–ด๋Š ์ชฝ์ด ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š”์ง€ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ฒŒ๋ฅด๋งŒ์–ด์™€ ์˜์–ด๋Š” ๊ฒŒ๋ฅด๋งŒ์–ด์ธ ๋ฐ˜๋ฉด, ํ”„๋ž‘์Šค์–ด์™€ ์ŠคํŽ˜์ธ์–ด๋Š” ๋กœ๋งˆ์–ด์ธ ์–ธ์–ด๋กœ BPE ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์ด์ ์„ ๋ˆ„๋ ธ๋‹ค๋Š” ์ ์ด ์ฃผ๋ชฉํ•  ๋งŒํ•ฉ๋‹ˆ๋‹ค. ๋กœ๋งˆ์–ด์ธ ์ดํƒˆ๋ฆฌ์•„์–ด์— ๋Œ€ํ•œ ์‹คํ—˜์€ ๋‹ค๋ฅธ ๋‘ ๋กœ๋งˆ์–ด์ธ ์–ธ์–ด์™€ ๋‹ค๋ฅธ ํŒจํ„ด์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Furthermore, Table 5 shows that depending on the language, either the BPE or Unigram exhibits better performance. It is noteworthy that the Germanic languages German and English benefit from the BPE algorithm, whereas the Romanic languages French and Spanish benefited from Unigram. The experiments for Italian, a Romanic language as well, show a different pattern than the other two Romanic languages. [|assistant|]
๊ทธ๋ฆผ 5: **LegalBench-Instruct์—์„œ์˜ Instruct ๋ชจ๋ธ ๋น„๊ต**. SaulLM-7B-Instruct๋Š” ์ตœ๊ณ ์˜ Mistral-Instruct ๋ชจ๋ธ๋ณด๋‹ค 6์  ์ด์ƒ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์—ฌ ์ตœ์‹  ๊ธฐ์ˆ ์„ ํ™•๋ฆฝํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 5: **LegalBench-Instruct์—์„œ์˜ ๋ช…๋ น ๋ชจ๋ธ ๋น„๊ต**. SaulLM-7B-Instruct๋Š” ์ตœ๊ณ ์˜ Mistral-Instruct ๋ชจ๋ธ๋ณด๋‹ค 6์  ์ด์ƒ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์—ฌ ์ตœ์‹  ๊ธฐ์ˆ ์„ ํ™•๋ฆฝํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 5: **Comparison of instruct models on LegalBench-Instruct**. SaulLM-7B-Instruct establishes the state-of-the-art, outperforming the best Mistral-Instruct model by a significant 6 absolute points. [|assistant|]
* ์šฐ๋ฆฌ๋Š” ์—„๊ฒฉํ•œ ํ•„ํ„ฐ๋ง ์ ˆ์ฐจ๋ฅผ ํ†ตํ•ด ๋‹ฌ์„ฑํ•œ ์ธ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์— ๋งž๊ฒŒ ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ๊ณ ํ’ˆ์งˆ ์ค‘๊ตญ์–ด Instruction Fine-tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. * ์†Œ์…œ ๋ฏธ๋””์–ด, ๋ฐฑ๊ณผ์‚ฌ์ „, ์ „ํ†ต์ ์ธ NLP ์ž‘์—… ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์†Œ์Šค๊ฐ€ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ถ„์„์€ ์ค‘๊ตญ ์ธํ„ฐ๋„ท์—์„œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๋ฐ ํ•„์ˆ˜์ ์ธ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. * ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ์™€ ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด CQIA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ Fine-tuning๋œ ๋ชจ๋ธ์ด ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋Š” ๊ฒƒ์œผ๋กœ ํ™•์ธ๋˜์–ด ์ค‘๊ตญ NLP ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ๊ท€์ค‘ํ•œ ๋ฆฌ์†Œ์Šค๋กœ ์ž๋ฆฌ๋งค๊น€ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
* ์šฐ๋ฆฌ๋Š” ์—„๊ฒฉํ•œ ํ•„ํ„ฐ๋ง ์ ˆ์ฐจ๋ฅผ ํ†ตํ•ด ๋‹ฌ์„ฑํ•œ ์ธ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์— ๋งž๊ฒŒ ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ๊ณ ํ’ˆ์งˆ ์ค‘๊ตญ์–ด ์ง€์นจ ๋ฏธ์„ธ ์กฐ์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. * ์†Œ์…œ ๋ฏธ๋””์–ด, ๋ฐฑ๊ณผ์‚ฌ์ „, ์ „ํ†ต์ ์ธ NLP ์ž‘์—… ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์†Œ์Šค๊ฐ€ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ถ„์„์€ ์ค‘๊ตญ ์ธํ„ฐ๋„ท์—์„œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๋ฐ ํ•„์ˆ˜์ ์ธ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. * ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ์™€ ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด CQIA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ์ด ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋Š” ๊ฒƒ์œผ๋กœ ํ™•์ธ๋˜์–ด ์ค‘๊ตญ NLP ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ๊ท€์ค‘ํ•œ ๋ฆฌ์†Œ์Šค๋กœ ์ž๋ฆฌ๋งค๊น€ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* We propose a high-quality Chinese instruction fine-tuning dataset, specifically designed to align with human interaction, achieved through rigorous filtering procedures. * We explore the influence of various data sources, including social media, encyclopedias, and traditional NLP tasks, on model performance. Our analysis offers essential insights for selecting training data from the Chinese internet. * Various benchmark tests and human evaluations confirm that models fine-tuned on our CQIA dataset exhibit superior performance, thus establishing CQIA as a valuable resource for the Chinese NLP community. [|assistant|]
1. _WizardMath 7B_๋Š” MPT, Falcon, Baichuan-chat, Vicuna v1.3, ChatGLM 2, Qwen, GSM8k ๋ฐ MATH benchmarks์—์„œ Llama 1 ๋ฐ Llama 2๋ฅผ ํฌํ•จํ•˜์—ฌ ์•ฝ 7B์—์„œ 40B ๋ฒ”์œ„์˜ parameter ์ˆ˜๋ฅผ ๊ฐ€์ง„ ๋Œ€๋ถ€๋ถ„์˜ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์„ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. parameter ์ˆ˜๋Š” ํ›จ์”ฌ ์ ์ง€๋งŒ์š”. 2. _WizardMath 13B_๋Š” GSM8k์—์„œ Llama 1 65B(63.9 ๋Œ€ 50.9) ๋ฐ Llama 2 70B(63.9 ๋Œ€ 56.8)๋ณด๋‹ค ํ›จ์”ฌ ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ MATH์—์„œ๋„ Llama 1 65B(14.0 ๋Œ€ 10.6)์™€ Llama 2 70B(14.0 ๋Œ€ 13.5)๋ฅผ ๋ชจ๋‘ ํฌ๊ฒŒ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. 3. ๊ฐ€์žฅ ๋ฐฉ๋Œ€ํ•œ ๋ชจ๋ธ์ธ _WizardMath 70B_๋Š” ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜์–ด GSM8k์—์„œ Llama 2 70B(81.6 ๋Œ€ 56.8)๋ฅผ 24.8%๋ผ๋Š” ์ƒ๋‹นํ•œ ์ฐจ์ด๋กœ ๋Šฅ๊ฐ€ํ•˜๋Š” ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋™์‹œ์— MATH์—์„œ๋„ Llama 2 70B(22.7 ๋Œ€ 13.5)๋ฅผ 9.2% ์ฐจ์ด๋กœ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
1. _WizardMath 7B_๋Š” MPT, Falcon, Baichuan-chat, Vicuna v1.3, ChatGLM 2, Qwen, GSM8k ๋ฐ MATH ๋ฒค์น˜๋งˆํฌ์—์„œ Llama 1 ๋ฐ Llama 2๋ฅผ ํฌํ•จํ•˜์—ฌ ์•ฝ 7B์—์„œ 40B ๋ฒ”์œ„์˜ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๋ฅผ ๊ฐ€์ง„ ๋Œ€๋ถ€๋ถ„์˜ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์„ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๋Š” ํ›จ์”ฌ ์ ์ง€๋งŒ์š”. 2. _WizardMath 13B_๋Š” GSM8k์—์„œ Llama 1 65B(63.9 ๋Œ€ 50.9) ๋ฐ Llama 2 70B(63.9 ๋Œ€ 56.8)๋ณด๋‹ค ํ›จ์”ฌ ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ MATH์—์„œ๋„ Llama 1 65B(14.0 ๋Œ€ 10.6)์™€ Llama 2 70B(14.0 ๋Œ€ 13.5)๋ฅผ ๋ชจ๋‘ ํฌ๊ฒŒ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. 3. ๊ฐ€์žฅ ๋ฐฉ๋Œ€ํ•œ ๋ชจ๋ธ์ธ _WizardMath 70B_๋Š” ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜์–ด GSM8k์—์„œ Llama 2 70B(81.6 ๋Œ€ 56.8)๋ฅผ 24.8%๋ผ๋Š” ์ƒ๋‹นํ•œ ์ฐจ์ด๋กœ ๋Šฅ๊ฐ€ํ•˜๋Š” ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋™์‹œ์— MATH์—์„œ๋„ Llama 2 70B(22.7 ๋Œ€ 13.5)๋ฅผ 9.2% ์ฐจ์ด๋กœ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]1. _WizardMath 7B_ surpasses most open-source models with parameter counts ranging approximately from 7B to 40B, including MPT, Falcon, Baichuan-chat, Vicuna v1.3, ChatGLM 2, Qwen, Llama 1 and Llama 2 on the GSM8k and MATH benchmarks. Even though its parameter counts are significantly lower. 2. _WizardMath 13B_ is significantly superior to Llama 1 65B (63.9 vs. 50.9) and Llama 2 70B (63.9 vs. 56.8) on GSM8k. Additionally, it substantially outperforms both Llama 1 65B (14.0 vs. 10.6) and Llama 2 70B (14.0 vs. 13.5) on MATH. 3. _WizardMath 70B_, our most extensive model, exemplifies a substantial advancement in performance, surpassing Llama 2 70B (81.6 vs. 56.8) by a significant margin of 24.8% on GSM8k. Concurrently, it also outperforms Llama 2 70B (22.7 vs. 13.5) by a margin of 9.2% on MATH. [|assistant|]
์ด๋Ÿฌํ•œ ๋งคํ•‘์˜ ์ˆœ์œ„๊ฐ€ ์ผ๋ฐ˜์ ์ธ ์ˆจ๊ฒจ์ง„ ์ฐจ์› ์„ ํƒ์— ๋น„ํ•ด ์ƒ๋Œ€์ ์œผ๋กœ ๋†’์„ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•  ์ˆ˜ ์žˆ์Œ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ LM head์˜ ์ˆœ์œ„ ์ œํ•œ์˜ ์˜ํ–ฅ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ 1000 ๋ฏธ๋งŒ์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์›์„ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ ์„ฑ๋Šฅ์ด ๋ˆˆ์— ๋„๊ฒŒ ์ €ํ•˜๋˜๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ŠคํŽ™ํŠธ๋Ÿผ ๋ถ„์„์˜ ๋ Œ์ฆˆ๋ฅผ ํ†ตํ•ด ํฌํ™” ํ˜„์ƒ์„ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ, ์ž‘์€ ๋ชจ๋ธ์—๋งŒ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด์˜ ์ด๋ฐฉ์„ฑ์ด ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒƒ์ด ํฌํ™” ํ˜„์ƒ๊ณผ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž‘์€ ๋ชจ๋ธ์˜ LM head๊ฐ€ _spectral_ ํฌํ™”, ์ฆ‰ ํ‡ดํ™” ์ƒํƒœ๋กœ ์ด์–ด์ง€๋Š” ํŠน์ด๊ฐ’์˜ ๊ท ์ผํ™”๋ฅผ ๋™์‹œ์— ๊ฒช๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
์ด๋Ÿฌํ•œ ๋งคํ•‘์˜ ์ˆœ์œ„๊ฐ€ ์ผ๋ฐ˜์ ์ธ ์ˆจ๊ฒจ์ง„ ์ฐจ์› ์„ ํƒ์— ๋น„ํ•ด ์ƒ๋Œ€์ ์œผ๋กœ ๋†’์„ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•  ์ˆ˜ ์žˆ์Œ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ LM ํ—ค๋“œ์˜ ์ˆœ์œ„ ์ œํ•œ์˜ ์˜ํ–ฅ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ 1000 ๋ฏธ๋งŒ์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์›์„ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ ์„ฑ๋Šฅ์ด ๋ˆˆ์— ๋„๊ฒŒ ์ €ํ•˜๋˜๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ŠคํŽ™ํŠธ๋Ÿผ ๋ถ„์„์˜ ๋ Œ์ฆˆ๋ฅผ ํ†ตํ•ด ํฌํ™” ํ˜„์ƒ์„ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ, ์ž‘์€ ๋ชจ๋ธ์—๋งŒ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด์˜ ์ด๋ฐฉ์„ฑ์ด ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒƒ์ด ํฌํ™” ํ˜„์ƒ๊ณผ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž‘์€ ๋ชจ๋ธ์˜ LM ํ—ค๋“œ๊ฐ€ _์ŠคํŽ™ํŠธ๋Ÿผ_ ํฌํ™”, ์ฆ‰ ํ‡ดํ™” ์ƒํƒœ๋กœ ์ด์–ด์ง€๋Š” ํŠน์ด๊ฐ’์˜ ๊ท ์ผํ™”๋ฅผ ๋™์‹œ์— ๊ฒช๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We empirically confirm that the rank of such a mapping can be expected to be relatively high compared to regular hidden dimension choices. Moreover, we conduct experiments to measure the impact of constraining the rank of the LM head on the performance of a large language model. Our results show that performance noticeably drops when using a hidden dimension below 1000. We further analyze the saturation phenomenon through the lens of spectral analysis and find that the emergence of last-layer anisotropy that only affects small models can be correlated with saturation. We also show that the LM heads of small models concurrently suffer from _spectral_ saturation, i.e. a uniformization of singular values that leads to a degenerated state. [|assistant|]
๋ชจ๋ธ๋‹ค๊ตญ์–ด ์ž‘์—…์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋‘ ๊ฐœ์˜ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ, ์ฆ‰ _Vicuna-7b-v1.5_Chiang ๋“ฑ(2023)๊ณผ _Mistral-7b-v1.0_Jiang ๋“ฑ(2023)์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. Vicuna-7b-v1.5๋Š” Instruction Fine-tuning์„ ํ†ตํ•ด LlaMa 2 Touvron ๋“ฑ(2023)์„ ๊ธฐ๋ฐ˜์œผ๋กœ Fine-tuningํ•œ ์˜คํ”ˆ ์†Œ์Šค ์ฑ„ํŒ… ๋ชจ๋ธ์ด๋ฉฐ, ๋‹ค๊ตญ์–ด Instruction์— ๋Œ€ํ•œ ์šฐ์ˆ˜ํ•œ ์ถ”์ข… ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. Mistral-7b-v1.0์€ Mistral-7b์— ๋Œ€ํ•ด Fine-tuning๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ„๋‹จํ•˜๊ฒŒ ๋‘ ๋ชจ๋ธ์„ ๊ฐ๊ฐ ๋‚˜ํƒ€๋‚ด๊ธฐ ์œ„ํ•ด ์ดํ›„ Vicuna์™€ Mistral๋กœ ์•ฝ์นญํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋ชจ๋ธ๋‹ค๊ตญ์–ด ์ž‘์—…์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋‘ ๊ฐœ์˜ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ, ์ฆ‰ _Vicuna-7b-v1.5_Chiang ๋“ฑ(2023)๊ณผ _Mistral-7b-v1.0_Jiang ๋“ฑ(2023)์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. Vicuna-7b-v1.5๋Š” ์ง€์‹œ ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด LlaMa 2 Touvron ๋“ฑ(2023)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•œ ์˜คํ”ˆ ์†Œ์Šค ์ฑ„ํŒ… ๋ชจ๋ธ์ด๋ฉฐ, ๋‹ค๊ตญ์–ด ์ง€์‹œ์–ด์— ๋Œ€ํ•œ ์šฐ์ˆ˜ํ•œ ์ถ”์ข… ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค.3 Mistral-7b-v1.0์€ Mistral-7b์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ„๋‹จํ•˜๊ฒŒ ๋‘ ๋ชจ๋ธ์„ ๊ฐ๊ฐ ๋‚˜ํƒ€๋‚ด๊ธฐ ์œ„ํ•ด ์ดํ›„ Vicuna์™€ Mistral๋กœ ์•ฝ์นญํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]ModelsWe test two open-source models that perform well on multilingual tasks, including _Vicuna-7b-v1.5_Chiang et al. (2023) and _Mistral-7b-v1.0_Jiang et al. (2023). Vicuna-7b-v1.5 is an open-source chat model fine-tuned on top of LlaMa 2 Touvron et al. (2023) via instruction fine-tuning and has shown good multilingual instruction following ability.3 Mistral-7b-v1.0 is fine-tuned on Mistral-7b. For simplicity, we abbreviate them as Vicuna and Mistral hereafter to represent the two models respectively. [|assistant|]
๊ทธ๋ฆผ 11์— ํ•™์Šต ๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ์—์„œ ์ค‘๊ตญ์–ด ์ฝ”ํผ์Šค์˜ CommonCrawl_Chn์€ CommonCrawl ์›์‹œ ์ฝ”ํผ์Šค์—์„œ ํŒŒ์ƒ๋˜๋ฉฐ ์ฒ ์ €ํ•œ ์ •๋ฆฌ๋ฅผ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค. Dolma(Soldaini et al., 2024), C4(Raffel et al., 2019), Pile(Gao et al., 2020; Biderman et al., 2022)์€ ์˜์–ด ์ฝ”ํผ์Šค์ž…๋‹ˆ๋‹ค. ์ด๋“ค์€ MinHash ์•Œ๊ณ ๋ฆฌ์ฆ˜(Broder, 1997)์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‚ด๋ถ€ ์ฝ”ํผ์Šค์™€ ์ฝ”ํผ์Šค ๊ฐ„ ์ค‘๋ณต์„ ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฝ”๋“œ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์—๋Š” ์Šคํƒ(Kocetkov et al., 2022)๊ณผ ์Šคํƒ€์ฝ”๋” ๋ฆฌ(Li et al., 2023)๊ฐ€ ํฌํ•จ๋˜๋ฉฐ, ๋‚ด๋ถ€ ์ค‘๋ณต๊ณผ ๊ต์ฐจ ์ค‘๋ณต์„ ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ๋ฌผ์—๋Š” ์‡ ํ‡ด ๋‹จ๊ณ„์—์„œ UltraChat(Ding et al., 2023), SlimOrca(Lian et al., 2023;3b), OssInstruction(Wei et al., 2023), EvolInstruction(Xu et al., 2023) ๋“ฑ ๋”์šฑ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ์™€ ๋…์  ๋ฐ์ดํ„ฐ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. SFT ์ ‘๋ฏธ์‚ฌ๊ฐ€ ๋ถ™์€ ๋ฐ์ดํ„ฐ๋Š” LectCode ์งˆ๋ฌธ, ์ดˆ๋“ฑํ•™๊ต๋ถ€ํ„ฐ 12ํ•™๋…„๊นŒ์ง€(K12) ๊ต๊ณผ์„œ ๋ฐ ์งˆ๋ฌธ ๋“ฑ์„ ํฌํ•จํ•˜๋Š” ์ž์‚ฌ ๋…์  ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 11์— ํ•™์Šต ๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ์—์„œ ์ค‘๊ตญ์–ด ์ฝ”ํผ์Šค์˜ CommonCrawl_Chn์€ CommonCrawl ์›์‹œ ์ฝ”ํผ์Šค์—์„œ ํŒŒ์ƒ๋˜๋ฉฐ ์ฒ ์ €ํ•œ ์ •๋ฆฌ๋ฅผ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค. Dolma(Soldaini et al., 2024), C4(Raffel et al., 2019), Pile(Gao et al., 2020; Biderman et al., 2022)์€ ์˜์–ด ์ฝ”ํผ์Šค์ž…๋‹ˆ๋‹ค. ์ด๋“ค์€ MinHash ์•Œ๊ณ ๋ฆฌ์ฆ˜(Broder, 1997)์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‚ด๋ถ€ ์ฝ”ํผ์Šค์™€ ์ฝ”ํผ์Šค ๊ฐ„ ์ค‘๋ณต์„ ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฝ”๋“œ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์—๋Š” ์Šคํƒ(Kocetkov et al., 2022)๊ณผ ์Šคํƒ€์ฝ”๋” ๋ฆฌ(Li et al., 2023)๊ฐ€ ํฌํ•จ๋˜๋ฉฐ, ๋‚ด๋ถ€ ์ค‘๋ณต๊ณผ ๊ต์ฐจ ์ค‘๋ณต์„ ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ๋ฌผ์—๋Š” ์‡ ํ‡ด ๋‹จ๊ณ„์—์„œ UltraChat(Ding et al., 2023), SlimOrca(Lian et al., 2023;3b), OssInstruct(Wei et al., 2023), EvolInstruct(Xu et al., 2023) ๋“ฑ ๋”์šฑ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ์™€ ๋…์  ๋ฐ์ดํ„ฐ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. SFT ์ ‘๋ฏธ์‚ฌ๊ฐ€ ๋ถ™์€ ๋ฐ์ดํ„ฐ๋Š” LectCode ์งˆ๋ฌธ, ์ดˆ๋“ฑํ•™๊ต๋ถ€ํ„ฐ 12ํ•™๋…„๊นŒ์ง€(K12) ๊ต๊ณผ์„œ ๋ฐ ์งˆ๋ฌธ ๋“ฑ์„ ํฌํ•จํ•˜๋Š” ์ž์‚ฌ ๋…์  ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We introduce our training data distribution in Figure 11. In the figure, CommonCrawl_Chn in a Chinese Corpus is derived from CommonCrawl raw corpus and goes through thorough cleaning. Dolma (Soldaini et al., 2024), C4 (Raffel et al., 2019), and Pile (Gao et al., 2020; Biderman et al., 2022) are English corpora. They are deduplicated inner corpus and across corpus using MinHash algorithms (Broder, 1997). The Code Pre-train data contains the stack (Kocetkov et al., 2022) and StarCoder Li et al. (2023), with inner deduplication and cross deduplication. In the decay stage, the data mixture contains more diverse data and proprietary data, including UltraChat (Ding et al., 2023), SlimOrca (Lian et al., 2023;3b), OssInstruct (Wei et al., 2023), EvolInstruct (Xu et al., 2023). The data with the suffix SFT is our proprietary data including LectCode questions, Kindergarten through 12th grade (K12) textbooks and questions, etc. [|assistant|]
ALFWorld์˜ _์„ฑ๊ณต๋ฅ _๊ณผ WebShop์˜ _ํ‰๊ท  ์ ์ˆ˜/์„ฑ๊ณต๋ฅ _์„ ์ถ”๊ฐ€๋กœ ๊ณ„์‚ฐํ•œ๋‹ค. ๋˜ํ•œ ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์˜ ๊ธธ์ด๋ฅผ ์ค„์ด๊ณ  ์ค„ ๋ฐ”๊ฟˆ์„ EOS token์œผ๋กœ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ReAct [449]์„ ๋”ฐ๋ฅด๊ธฐ๋„ ํ•œ๋‹ค.[|endofturn|]
ALFWorld์˜ _์„ฑ๊ณต๋ฅ _๊ณผ WebShop์˜ _ํ‰๊ท  ์ ์ˆ˜/์„ฑ๊ณต๋ฅ _์„ ์ถ”๊ฐ€๋กœ ๊ณ„์‚ฐํ•œ๋‹ค. ๋˜ํ•œ ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์˜ ๊ธธ์ด๋ฅผ ์ค„์ด๊ณ  ์ค„ ๋ฐ”๊ฟˆ์„ EOS ํ† ํฐ์œผ๋กœ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ReAct [449]์„ ๋”ฐ๋ฅด๊ธฐ๋„ ํ•œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]pute _success rate_ for ALFWorld and _average score/success rate_ for WebShop. Further, we also follow ReAct [449] to reduce the length of the input prompt and utilize line break as the EOS token. [|assistant|]
์ด๋Ÿฌํ•œ ๋น„์šฉ ์ ˆ์ถฉ์„ ๋ถ„์„ํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๋น„์šฉ๊ณผ ์ถ”๊ฐ€ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์–ธ์–ด ๋ชจ๋ธ ํ•™์Šต ๋น„์šฉ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ์‹คํ—˜์˜ ๊ฒฝ์šฐ, ๋น ๋ฅธ ์ƒ์„ฑ์„ ์œ„ํ•ด vLLM(Kwon et al., 2023) ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, Mistral-7B๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ ๋‹จ์ผ A100์—์„œ ์‹œ๊ฐ„๋‹น 3M tokens์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (์ด ์ž‘์—…์—์„œ์™€ ๊ฐ™์ด) 85B tokens์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ๋Š” ์•ฝ 25,000์‹œ๊ฐ„์˜ GPU๊ฐ€ ์†Œ์š”๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ด๋Ÿฌํ•œ ๋น„์šฉ ์ ˆ์ถฉ์„ ๋ถ„์„ํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๋น„์šฉ๊ณผ ์ถ”๊ฐ€ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์–ธ์–ด ๋ชจ๋ธ ํ•™์Šต ๋น„์šฉ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ์‹คํ—˜์˜ ๊ฒฝ์šฐ, ๋น ๋ฅธ ์ƒ์„ฑ์„ ์œ„ํ•ด vLLM(Kwon et al., 2023) ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, Mistral-7B๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ ๋‹จ์ผ A100์—์„œ ์‹œ๊ฐ„๋‹น 3M ํ† ํฐ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (์ด ์ž‘์—…์—์„œ์™€ ๊ฐ™์ด) 850์–ต ํ† ํฐ์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ๋Š” ์•ฝ 25,000์‹œ๊ฐ„์˜ GPU๊ฐ€ ์†Œ์š”๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To analyze this cost trade-off, we compare the cost of generating synthetic data, versus that of training a language model on extra data. For our synthetic data generation experiments, we use the vLLM (Kwon et al., 2023) library for fast generation. In particular, we are able to generate 3M tokens per hour on a single A100 when using the Mistral-7B. Generating 85B tokens (as in our work) accounts for about 25K GPU hours. [|assistant|]
FLARE(Jiang et al., 2023)๋Š” ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์—์„œ ์ƒ๋‹น์ˆ˜์˜ ์˜ˆ/์•„๋‹ˆ์š” ์งˆ๋ฌธ์ด ํฌํ•จ๋œ StrategyQA๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ €์ž๋Š” ๋ชจ๋ธ์ด ์˜ˆ/์•„๋‹ˆ์š” ๋‹ต๋ณ€์„ ๊ฒฐ์ •ํ•˜๋Š” ์ •ํ™•ํ•œ ์ถ”๋ก  ๊ณผ์ •๊ณผ ์ตœ์ข… ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋„๋ก ์š”์ฒญํ•˜์—ฌ ๋‹ต๋ณ€์ด ๊ณจ๋“œ ๋‹ต๋ณ€๊ณผ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ In-Context ์ƒ˜ํ”Œ์„ ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ์— ํ†ตํ•ฉํ•˜๊ณ  COPA, HellaSwag ๋ฐ PIQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•จ์œผ๋กœ์จ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋Š” LLM-R(Wang et al., 2023) ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ITER-RETGEN(Gao et al., 2022) ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ๊ฐœ๋…์€ ๋ฐ˜๋ณต์ ์ธ ๋ฐฉ๋ฒ•๋ก ์„ ์‚ฌ์šฉํ•˜์—ฌ Retrieval๊ณผ Language Models๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ StrategyQA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Commonsense Reasoning ์ž‘์—…์„ ์œ„ํ•ด ํ›ˆ๋ จ๋˜์—ˆ์œผ๋ฉฐ 7๋ฒˆ์˜ ๋ฐ˜๋ณต์—์„œ ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ KG-BART(Shao et al., 2023)๋Š” Commonsense Reasoning ์ž‘์—…์˜ ์šฐ์„  ์ˆœ์œ„๋ฅผ ์ •ํ•˜๊ณ  Knowledge Graphs๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด ๋ถ„์•ผ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํŠน์ • ํ‰๊ฐ€ ์ง€ํ‘œ์—์„œ ์ธ๊ฐ„ ์ˆ˜์ค€์— ๊ทผ์ ‘ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋ชจ๋ธ์˜ Commonsense Reasoning ์ž‘์—… ์™„๋ฃŒ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
FLARE(Jiang et al., 2023)๋Š” ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์—์„œ ์ƒ๋‹น์ˆ˜์˜ ์˜ˆ/์•„๋‹ˆ์š” ์งˆ๋ฌธ์ด ํฌํ•จ๋œ StrategyQA๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ €์ž๋Š” ๋ชจ๋ธ์ด ์˜ˆ/์•„๋‹ˆ์š” ๋‹ต๋ณ€์„ ๊ฒฐ์ •ํ•˜๋Š” ์ •ํ™•ํ•œ ์ถ”๋ก  ๊ณผ์ •๊ณผ ์ตœ์ข… ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋„๋ก ์š”์ฒญํ•˜์—ฌ ๋‹ต๋ณ€์ด ๊ณจ๋“œ ๋‹ต๋ณ€๊ณผ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ƒํ™ฉ ์ƒ˜ํ”Œ์„ ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ์— ํ†ตํ•ฉํ•˜๊ณ  COPA, HellaSwag ๋ฐ PIQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•จ์œผ๋กœ์จ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋Š” LLM-R(Wang et al., 2023) ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ITER-RETGEN(Gao et al., 2022) ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ๊ฐœ๋…์€ ๋ฐ˜๋ณต์ ์ธ ๋ฐฉ๋ฒ•๋ก ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰๊ธฐ์™€ ์–ธ์–ด ๋ชจ๋ธ์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ StrategyQA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์‹ ์ถ”๋ก  ์ž‘์—…์„ ์œ„ํ•ด ํ›ˆ๋ จ๋˜์—ˆ์œผ๋ฉฐ 7๋ฒˆ์˜ ๋ฐ˜๋ณต์—์„œ ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ KG-BART(Shao et al., 2023)๋Š” ์ƒ์‹ ์ถ”๋ก  ์ž‘์—…์˜ ์šฐ์„  ์ˆœ์œ„๋ฅผ ์ •ํ•˜๊ณ  ์ง€์‹ ๊ทธ๋ž˜ํ”„๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด ๋ถ„์•ผ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํŠน์ • ํ‰๊ฐ€ ์ง€ํ‘œ์—์„œ ์ธ๊ฐ„ ์ˆ˜์ค€์— ๊ทผ์ ‘ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋ชจ๋ธ์˜ ์ƒ์‹ ์ถ”๋ก  ์ž‘์—… ์™„๋ฃŒ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]FLARE (Jiang et al., 2023) uses StrategyQA, which contains a significant number of yes/no questions from a diverse range of sources. In addition, the authors request that the model provide the exact reasoning process and the final answer that determines the yes/no answer, ensuring that the answer matches the gold answer exactly. The incorporation of in-context samples into the retrieved content, along with training using data from the COPA, HellaSwag, and PIQA datasets, has resulted in the development of LLM-R (Wang et al., 2023) model that exhibits excellent performance. The fundamental concept of the ITER-RETGEN (Gao et al., 2022) model employs an iterative methodology to integrate retrievers and language models. The model was trained for the Commonsense Reasoning task using the StrategyQA dataset and achieved its optimal performance at seven iterations. In contrast, KG-BART (Shao et al., 2023) is designed to prioritize the Commonsense Reasoning task and employs knowledge graphs to enhance its performance in this area. This approach has proven effective in significantly improving the model's ability to complete the Commonsense Reasoning task, with performance approaching that of human beings under certain evaluation metrics. [|assistant|]
README.md exists but content is empty. Use the Edit dataset card button to edit it.
Downloads last month
0
Edit dataset card