question
stringlengths
101
4.47k
chosen
stringlengths
3
4.25k
rejected
stringlengths
39
9.29k
Ensure your student can answer the new question without the given question. If you want to use some numbers, conditions or background in the given question, please restate them to ensure no information is omitted in your new question.
ํ•™์ƒ์ด ์ฃผ์–ด์ง„ ์งˆ๋ฌธ ์—†์ด ์ƒˆ๋กœ์šด ์งˆ๋ฌธ์— ๋‹ตํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์‹ญ์‹œ์˜ค. ์ฃผ์–ด์ง„ ์งˆ๋ฌธ์—์„œ ์ผ๋ถ€ ์ˆซ์ž, ์กฐ๊ฑด ๋˜๋Š” ๋ฐฐ๊ฒฝ์„ ์‚ฌ์šฉํ•˜๊ณ  ์‹ถ๋‹ค๋ฉด, ์ƒˆ๋กœ์šด ์งˆ๋ฌธ์—์„œ ์ •๋ณด๊ฐ€ ๋ˆ„๋ฝ๋˜์ง€ ์•Š๋„๋ก ์ด๋ฅผ ๋‹ค์‹œ ์„œ์ˆ ํ•ด ์ฃผ์‹ญ์‹œ์˜ค.
ํ•™์ƒ์ด ์ฃผ์–ด์ง„ ๋ฌธ์ œ ์—†์ด๋„ ์ƒˆ๋กœ์šด ๋ฌธ์ œ์— ๋‹ตํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์–ด์ง„ ๋ฌธ์ œ์—์„œ ์ˆซ์ž, ์กฐ๊ฑด ๋˜๋Š” ๋ฐฐ๊ฒฝ์„ ์‚ฌ์šฉํ•˜๋ ค๋Š” ๊ฒฝ์šฐ, ์ƒˆ๋กœ์šด ๋ฌธ์ œ์—์„œ ์ •๋ณด๊ฐ€ ๋ˆ„๋ฝ๋˜์ง€ ์•Š๋„๋ก ๋‹ค์‹œ ์ž‘์„ฑํ•˜์„ธ์š”.
Zhewei Wei received his Ph.D. of Computer Science and Engineering from Hong Kong University of Science and Technology. He did postdoctoral research in Aarhus University from 2012 to 2014, and joined Renmin University of China in 2014.
Zhewei Wei๋Š” ํ™์ฝฉ๊ณผํ•™๊ธฐ์ˆ ๋Œ€ํ•™๊ต์—์„œ ์ปดํ“จํ„ฐ ๊ณผํ•™ ๋ฐ ๊ณตํ•™ ๋ฐ•์‚ฌ ํ•™์œ„๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋Š” 2012๋…„๋ถ€ํ„ฐ 2014๋…„๊นŒ์ง€ ์˜ค๋ฅดํ›„์Šค ๋Œ€ํ•™๊ต์—์„œ ๋ฐ•์‚ฌ ํ›„ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์œผ๋ฉฐ, 2014๋…„์— ์ค‘๊ตญ ์ธ๋ฏผ๋Œ€ํ•™๊ต์— ํ•ฉ๋ฅ˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
์ œ์›จ์ด ์œ„๋Š” ํ™์ฝฉ๊ณผํ•™๊ธฐ์ˆ ๋Œ€ํ•™๊ต์—์„œ ์ปดํ“จํ„ฐ๊ณผํ•™ ๋ฐ ๊ณตํ•™ ๋ฐ•์‚ฌ ํ•™์œ„๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. 2012๋…„๋ถ€ํ„ฐ 2014๋…„๊นŒ์ง€ ์˜ค๋ฅดํ›„์Šค ๋Œ€ํ•™๊ต์—์„œ ๋ฐ•์‚ฌํ›„ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์œผ๋ฉฐ, 2014๋…„ ์ค‘๊ตญ ์ธ๋ฏผ๋Œ€ํ•™๊ต์— ํ•ฉ๋ฅ˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
Our work paves the way for a better understanding of the consequences of the softmax bottleneck on language modeling, and for the conception of language models that better embrace the complexity of the target probability distribution.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ์–ธ์–ด ๋ชจ๋ธ๋ง์—์„œ ์†Œํ”„ํŠธ๋งฅ์Šค ๋ณ‘๋ชฉ ํ˜„์ƒ์˜ ๊ฒฐ๊ณผ์— ๋Œ€ํ•œ ๋” ๋‚˜์€ ์ดํ•ด์™€ ๋ชฉํ‘œ ํ™•๋ฅ  ๋ถ„ํฌ์˜ ๋ณต์žก์„ฑ์„ ๋” ์ž˜ ์ˆ˜์šฉํ•˜๋Š” ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ตฌ์ƒ์„ ์œ„ํ•œ ๊ธธ์„ ์—ด์–ด์ค๋‹ˆ๋‹ค.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ์†Œํ”„ํŠธ๋งฅ์Šค ๋ณ‘๋ชฉ ํ˜„์ƒ์ด ์–ธ์–ด ๋ชจ๋ธ๋ง์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ณ , ๋ชฉํ‘œ ํ™•๋ฅ  ๋ถ„ํฌ์˜ ๋ณต์žก์„ฑ์„ ๋” ์ž˜ ํฌ์šฉํ•˜๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ์„ค๊ณ„ํ•˜๋Š” ๋ฐ ๊ธธ์„ ๋‹ฆ๋Š”๋‹ค.
Footnote 16: _...claims are very focused on the idea of themes of digital age and the story doesnโ€™t cover that at all. Its not even based on a modern world.โ€ โ€“ authorโ€™s social media accounts are mentioned at the very end of the book.
Footnote 16: _...์ฃผ์žฅ์€ ๋””์ง€ํ„ธ ์‹œ๋Œ€์˜ ์ฃผ์ œ์— ๋งค์šฐ ์ง‘์ค‘๋˜์–ด ์žˆ์œผ๋ฉฐ ์ด์•ผ๊ธฐ๋Š” ๊ทธ๊ฒƒ์„ ์ „ํ˜€ ๋‹ค๋ฃจ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ˜„๋Œ€ ์„ธ๊ณ„๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜์ง€๋„ ์•Š์Šต๋‹ˆ๋‹ค.โ€ โ€“ ์ €์ž์˜ ์†Œ์…œ ๋ฏธ๋””์–ด ๊ณ„์ •์€ ์ฑ…์˜ ๋งจ ๋์— ์–ธ๊ธ‰๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
16์ชฝ ์ฃผ: _...๋””์ง€ํ„ธ ์‹œ๋Œ€์˜ ์ฃผ์ œ์— ๋Œ€ํ•œ ์ฃผ์žฅ์€ ๋งค์šฐ ์ง‘์ค‘๋˜์–ด ์žˆ์œผ๋ฉฐ ์ด์•ผ๊ธฐ๋Š” ์ด๋ฅผ ์ „ํ˜€ ๋‹ค๋ฃจ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ˜„๋Œ€ ์„ธ๊ณ„๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜์ง€๋„ ์•Š์Šต๋‹ˆ๋‹ค." - ์ €์ž์˜ ์†Œ์…œ ๋ฏธ๋””์–ด ๊ณ„์ •์€ ์ฑ… ๋งˆ์ง€๋ง‰์— ์–ธ๊ธ‰๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
In our main experiments in Section 4 we demonstrate that LLMs trained with synthetic rephrases are a better backbone for zero-shot question-answering tasks as the model learns the question-answer format and style during pre-training.
์šฐ๋ฆฌ์˜ ์ฃผ์š” ์‹คํ—˜์ธ 4์žฅ์—์„œ, ํ•ฉ์„ฑ๋œ ์žฌ๊ตฌ์„ฑ์„ ํ†ตํ•ด ํ›ˆ๋ จ๋œ LLMs๊ฐ€ ์‚ฌ์ „ ํ›ˆ๋ จ ๋™์•ˆ ์งˆ๋ฌธ-๋‹ต๋ณ€ ํ˜•์‹๊ณผ ์Šคํƒ€์ผ์„ ํ•™์Šตํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ œ๋กœ์ƒท ์งˆ๋ฌธ-๋‹ต๋ณ€ ์ž‘์—…์— ๋” ๋‚˜์€ ๋ฐฑ๋ณธ์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค.
์„น์…˜ 4์˜ ์ฃผ์š” ์‹คํ—˜์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฌธ๊ตฌ๋กœ ํ›ˆ๋ จ๋œ LLM์ด ์‚ฌ์ „ ํ›ˆ๋ จ ์ค‘์— ์งˆ๋ฌธ-๋‹ต๋ณ€ ํ˜•์‹๊ณผ ์Šคํƒ€์ผ์„ ํ•™์Šตํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ œ๋กœ ์ƒท ์งˆ๋ฌธ-๋‹ต๋ณ€ ์ž‘์—…์— ๋” ๋‚˜์€ ๋ฐฑ๋ณธ์ด๋ผ๋Š” ๊ฒƒ์„ ์ž…์ฆํ•œ๋‹ค.
Recently, several comprehensive benchmarks [70, 364, 520] have been released for the evaluation of LLMs. In this part, we introduce several widely used benchmarks, _i.e._, MMLU, BIG-bench, HELM, and a series of human exam benchmarks.
์ตœ๊ทผ LLMs์˜ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์ข…ํ•ฉ์ ์ธ ๋ฒค์น˜๋งˆํฌ [70, 364, 520]๊ฐ€ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” MMLU, BIG-bench, HELM ๋ฐ ์ผ๋ จ์˜ ์ธ๊ฐ„ ์‹œํ—˜ ๋ฒค์น˜๋งˆํฌ ๋“ฑ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์—ฌ๋Ÿฌ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.
์ตœ๊ทผ์—๋Š” LLM์˜ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ๋ช‡ ๊ฐ€์ง€ ์ข…ํ•ฉ์ ์ธ ๋ฒค์น˜๋งˆํฌ[70, 364, 520]๊ฐ€ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋ฒค์น˜๋งˆํฌ, ์ฆ‰ MMLU, BIG-bench, HELM ๋ฐ ์ผ๋ จ์˜ ์ธ๊ฐ„ ์‹œํ—˜ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.
An ethical dilemma is a situation where there are competing moral claims, and it is not clear which course of action is morally right. For example, a doctor may have to decide between saving the life of a pregnant woman or the fetus.
์œค๋ฆฌ์  ๋”œ๋ ˆ๋งˆ๋Š” ์ƒ์ถฉํ•˜๋Š” ๋„๋•์  ์ฃผ์žฅ๋“ค์ด ์กด์žฌํ•˜๋ฉฐ, ์–ด๋–ค ํ–‰๋™์ด ๋„๋•์ ์œผ๋กœ ์˜ณ์€์ง€ ๋ช…ํ™•ํ•˜์ง€ ์•Š์€ ์ƒํ™ฉ์„ ๋งํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์˜์‚ฌ๋Š” ์ž„์‹ ํ•œ ์—ฌ์„ฑ์˜ ์ƒ๋ช…๊ณผ ํƒœ์•„์˜ ์ƒ๋ช… ์ค‘ ํ•˜๋‚˜๋ฅผ ๊ตฌํ•ด์•ผ ํ•˜๋Š” ๊ฒฐ์ •์„ ๋‚ด๋ ค์•ผ ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
์œค๋ฆฌ์  ๋”œ๋ ˆ๋งˆ๋Š” ์ƒ๋ฐ˜๋˜๋Š” ๋„๋•์  ์ฃผ์žฅ์ด ์žˆ๋Š” ์ƒํ™ฉ์œผ๋กœ, ์–ด๋–ค ํ–‰๋™ ๋ฐฉ์นจ์ด ๋„๋•์ ์œผ๋กœ ์˜ณ์€์ง€ ๋ช…ํ™•ํ•˜์ง€ ์•Š์€ ์ƒํ™ฉ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ์˜์‚ฌ๋Š” ์ž„์‚ฐ๋ถ€์™€ ํƒœ์•„์˜ ์ƒ๋ช…์„ ๊ตฌํ•˜๋Š” ๊ฒƒ ์ค‘ ํ•˜๋‚˜๋ฅผ ์„ ํƒํ•ด์•ผ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Footnote 21: The total number of human verified responses is not equal to the number of hazard categories multiplied by the number of models, multiplied by 10 (i.e., 910 == 7 x 13 x 10) as some combinations did not have 10 responses.
Footnote 21: ์ธ๊ฐ„์— ์˜ํ•ด ๊ฒ€์ฆ๋œ ์‘๋‹ต์˜ ์ด ์ˆ˜๋Š” ์œ„ํ—˜ ๋ฒ”์ฃผ ์ˆ˜์— ๋ชจ๋ธ ์ˆ˜๋ฅผ ๊ณฑํ•˜๊ณ  ๋‹ค์‹œ 10์„ ๊ณฑํ•œ ๊ฐ’๊ณผ ๊ฐ™์ง€ ์•Š์Šต๋‹ˆ๋‹ค (์ฆ‰, 910 == 7 x 13 x 10), ์ด๋Š” ์ผ๋ถ€ ์กฐํ•ฉ์ด 10๊ฐœ์˜ ์‘๋‹ต์„ ๊ฐ€์ง€์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
๊ฐ์ฃผ 21: ์ผ๋ถ€ ์กฐํ•ฉ์— 10๊ฐœ์˜ ์‘๋‹ต์ด ์—†์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ธ๊ฐ„์ด ํ™•์ธํ•œ ์ด ์‘๋‹ต ์ˆ˜๋Š” ์œ„ํ—˜ ๋ฒ”์ฃผ ์ˆ˜์— ๋ชจ๋ธ ์ˆ˜๋ฅผ ๊ณฑํ•œ ์ˆ˜(์˜ˆ: 910 == 7 x 13 x 10)์™€ ๊ฐ™์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
Jurning maintains winning what further, Nash becomes natural when he initially discussed her concerns about his motion. When he was in the middle of a school, she was able to learn a more complex, and the story is not a good choice.
Jurning์€ ๊ทธ์˜ ๋™์ž‘์— ๋Œ€ํ•œ ๊ทธ๋…€์˜ ์šฐ๋ ค๋ฅผ ์ฒ˜์Œ ๋…ผ์˜ํ–ˆ์„ ๋•Œ ์Šน๋ฆฌ๋ฅผ ์œ ์ง€ํ•˜๋ฉฐ, Nash๋Š” ์ž์—ฐ์Šค๋Ÿฌ์›Œ์ง‘๋‹ˆ๋‹ค. ๊ทธ๊ฐ€ ํ•™๊ต ์ค‘๊ฐ„์— ์žˆ์„ ๋•Œ, ๊ทธ๋…€๋Š” ๋” ๋ณต์žกํ•œ ๊ฒƒ์„ ๋ฐฐ์šธ ์ˆ˜ ์žˆ์—ˆ๊ณ , ์ด์•ผ๊ธฐ๋Š” ์ข‹์€ ์„ ํƒ์ด ์•„๋‹™๋‹ˆ๋‹ค.
์ €๋‹์€ ์ดํ›„์—๋„ ์Šน๋ฆฌ๋ฅผ ์œ ์ง€ํ•˜๋ฉฐ, ์ฒ˜์Œ์— ์ž์‹ ์˜ ๋™์ž‘์— ๋Œ€ํ•œ ์šฐ๋ ค๋ฅผ ๋…ผ์˜ํ–ˆ์„ ๋•Œ ๋‚ด์‰ฌ๋Š” ์ž์—ฐ์Šค๋Ÿฌ์›Œ์กŒ๋‹ค. ๊ทธ๊ฐ€ ํ•™๊ต ์ค‘๋ฐ˜์— ์žˆ์„ ๋•Œ ๊ทธ๋…€๋Š” ๋” ๋ณต์žกํ•œ ๊ฒƒ์„ ๋ฐฐ์šธ ์ˆ˜ ์žˆ์—ˆ๊ณ , ๊ทธ ์ด์•ผ๊ธฐ๋Š” ์ข‹์€ ์„ ํƒ์ด ์•„๋‹ˆ์—ˆ๋‹ค.
In this section, we explore X-of-thought reasoning through three different categorizations: the construction of X-of-thought (SS4.1), the structural variants of X-of-thought (SS4.2), and the enhanced methods of X-of-thought (SS4.3).
์ด ์„น์…˜์—์„œ๋Š” ์„ธ ๊ฐ€์ง€ ๋‹ค๋ฅธ ๋ถ„๋ฅ˜๋ฅผ ํ†ตํ•ด X-of-thought ์ถ”๋ก ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค: X-of-thought์˜ ๊ตฌ์„ฑ (SS4.1), X-of-thought์˜ ๊ตฌ์กฐ์  ๋ณ€ํ˜• (SS4.2), ๊ทธ๋ฆฌ๊ณ  X-of-thought์˜ ํ–ฅ์ƒ๋œ ๋ฐฉ๋ฒ• (SS4.3).
์ด ์„น์…˜์—์„œ๋Š” ์„ธ ๊ฐ€์ง€ ๋ฒ”์ฃผ๋ฅผ ํ†ตํ•ด X-of-thought ์ถ”๋ก ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค: X-of-thought์˜ ๊ตฌ์„ฑ(SS4.1), X-of-thought์˜ ๊ตฌ์กฐ์  ๋ณ€ํ˜•(SS4.2), X-of-thought์˜ ํ–ฅ์ƒ๋œ ๋ฐฉ๋ฒ•(SS4.3)์ž…๋‹ˆ๋‹ค.
**Lemma F.1**.: _Let \(\mathcal{Q}_{1},\ldots,\mathcal{Q}_{k}\) be fixed sets (we call domains), and assume that for each \(i\in[k]\), \(Q_{i}\) is independently and randomly chosen from its corresponding domain \(\mathcal{Q}_{i}\).
Lemma F.1: \(\mathcal{Q}_{1},\ldots,\mathcal{Q}_{k}\)๋ฅผ ๊ณ ์ •๋œ ์ง‘ํ•ฉ(๋„๋ฉ”์ธ์ด๋ผ๊ณ  ๋ถ€๋ฆ„)์ด๋ผ๊ณ  ํ•˜์ž. ๊ฐ \(i\in[k]\)์— ๋Œ€ํ•ด \(Q_{i}\)๊ฐ€ ํ•ด๋‹น ๋„๋ฉ”์ธ \(\mathcal{Q}_{i}\)์—์„œ ๋…๋ฆฝ์ ์œผ๋กœ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ๋‹ค๊ณ  ๊ฐ€์ •ํ•˜์ž.
**์ •๋ฆฌ F.1**.: _Let \(\mathcal{Q}_{1},\ldots,\mathcal{Q}_{k}\)๋ฅผ ๊ณ ์ •๋œ ์ง‘ํ•ฉ(๋„๋ฉ”์ธ์ด๋ผ๊ณ  ํ•จ)์œผ๋กœ ํ•˜๊ณ , ๊ฐ \(i\in[k]\)์— ๋Œ€ํ•ด \(Q_{i}\)๊ฐ€ ํ•ด๋‹น ๋„๋ฉ”์ธ \(\mathcal{Q}_{i}\)์—์„œ ๋…๋ฆฝ์ ์œผ๋กœ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค.
Figure 10: **Location mentions in all binary questions in our full dataset**. We visualize all location mentions in our full dataset on a world map. This shows that the dataset provides a diverse coverage of topics across the globe.
Figure 10: **์ „์ฒด ๋ฐ์ดํ„ฐ์…‹์˜ ๋ชจ๋“  ์ด์ง„ ์งˆ๋ฌธ์—์„œ ์–ธ๊ธ‰๋œ ์œ„์น˜**. ์šฐ๋ฆฌ๋Š” ์ „์ฒด ๋ฐ์ดํ„ฐ์…‹์—์„œ ์–ธ๊ธ‰๋œ ๋ชจ๋“  ์œ„์น˜๋ฅผ ์„ธ๊ณ„ ์ง€๋„์— ์‹œ๊ฐํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ฐ์ดํ„ฐ์…‹์ด ์ „ ์„ธ๊ณ„์ ์œผ๋กœ ๋‹ค์–‘ํ•œ ์ฃผ์ œ๋ฅผ ๋‹ค๋ฃจ๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 10: **์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ชจ๋“  ์ด์ง„ ์งˆ๋ฌธ์— ์–ธ๊ธ‰๋œ ์œ„์น˜**. ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ชจ๋“  ์œ„์น˜๋ฅผ ์›”๋“œ ๋งต์— ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ „ ์„ธ๊ณ„์˜ ๋‹ค์–‘ํ•œ ์ฃผ์ œ์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•˜๊ฒŒ ๋‹ค๋ฃจ๊ณ  ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The raw dataset contains questions that are ill-defined, overly personal, or of niche interests. Furthermore, recent questions are highly unbalanced, with over 80% of questions since June 1, 2023 coming from Manifold and Polymarket.
์›์‹œ ๋ฐ์ดํ„ฐ์…‹์—๋Š” ์ •์˜๊ฐ€ ๋ถˆ๋ช…ํ™•ํ•˜๊ฑฐ๋‚˜, ์ง€๋‚˜์น˜๊ฒŒ ๊ฐœ์ธ์ ์ด๊ฑฐ๋‚˜, ํŠน์ • ๊ด€์‹ฌ์‚ฌ์— ๊ตญํ•œ๋œ ์งˆ๋ฌธ๋“ค์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒŒ๋‹ค๊ฐ€, ์ตœ๊ทผ ์งˆ๋ฌธ๋“ค์€ ๋งค์šฐ ๋ถˆ๊ท ํ˜•ํ•œ๋ฐ, 2023๋…„ 6์›” 1์ผ ์ดํ›„์˜ ์งˆ๋ฌธ ์ค‘ 80% ์ด์ƒ์ด Manifold์™€ Polymarket์—์„œ ๋‚˜์˜จ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
์›์‹œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์ž˜๋ชป ์ •์˜๋˜๊ฑฐ๋‚˜ ์ง€๋‚˜์น˜๊ฒŒ ๊ฐœ์ธ์ ์ด๊ฑฐ๋‚˜ ํ‹ˆ์ƒˆ ๊ด€์‹ฌ์‚ฌ์— ๋Œ€ํ•œ ์งˆ๋ฌธ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ตœ๊ทผ ์งˆ๋ฌธ์€ 6์›” 1์ผ ์ดํ›„ ์งˆ๋ฌธ์˜ 80% ์ด์ƒ์ด Manifold์™€ Polymarket์—์„œ ์ถœ์ œ๋˜๋Š” ๋“ฑ ๋ถˆ๊ท ํ˜•ํ•œ ํŽธ์ž…๋‹ˆ๋‹ค.
**Who was involved in the data collection process and how were they compensated?** Apart from the AI-safety working group members, 12 civil society experts were interviewed for creating the taxonomy for AI safety evaluation prompts.
**๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๊ณผ์ •์— ๋ˆ„๊ฐ€ ์ฐธ์—ฌํ–ˆ์œผ๋ฉฐ, ๊ทธ๋“ค์€ ์–ด๋–ป๊ฒŒ ๋ณด์ƒ๋ฐ›์•˜๋‚˜์š”?** AI ์•ˆ์ „ ์ž‘์—… ๊ทธ๋ฃน ๊ตฌ์„ฑ์› ์™ธ์—๋„, AI ์•ˆ์ „ ํ‰๊ฐ€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์œ„ํ•œ ๋ถ„๋ฅ˜ ์ฒด๊ณ„๋ฅผ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด 12๋ช…์˜ ์‹œ๋ฏผ ์‚ฌํšŒ ์ „๋ฌธ๊ฐ€๋“ค์ด ์ธํ„ฐ๋ทฐ๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค.
**๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๊ณผ์ •์— ์–ด๋–ค ์‚ฌ๋žŒ๋“ค์ด ์ฐธ์—ฌํ–ˆ์œผ๋ฉฐ ์–ด๋–ป๊ฒŒ ๋ณด์ƒ์„ ๋ฐ›์•˜๋‚˜์š”?** AI ์•ˆ์ „ ํ‰๊ฐ€ ํ”„๋กฌํ”„ํŠธ์˜ ๋ถ„๋ฅ˜๋ฅผ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด AI ์•ˆ์ „ ์—ฐ๊ตฌ ๊ทธ๋ฃน ๊ตฌ์„ฑ์› ์™ธ์—๋„ 12๋ช…์˜ ์‹œ๋ฏผ ์‚ฌํšŒ ์ „๋ฌธ๊ฐ€๋ฅผ ์ธํ„ฐ๋ทฐํ–ˆ์Šต๋‹ˆ๋‹ค.
In summary, CoT prompting provides a general and flexible approach to eliciting the reasoning ability of LLMs. There are also some preliminary attempts to extend this technique to solve multimodal [523] and multilingual tasks [524].
์š”์•ฝํ•˜์ž๋ฉด, CoT ํ”„๋กฌํ”„ํŠธ๋Š” LLM์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๋Š” ์ผ๋ฐ˜์ ์ด๊ณ  ์œ ์—ฐํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด ๊ธฐ์ˆ ์„ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ [523] ๋ฐ ๋‹ค๊ตญ์–ด ์ž‘์—… [524]์— ์ ์šฉํ•˜๋ ค๋Š” ์ดˆ๊ธฐ ์‹œ๋„๋“ค๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
์š”์•ฝํ•˜๋ฉด, CoT ํ”„๋กฌํ”„ํŠธ๋Š” LLM์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๋Š” ์ผ๋ฐ˜์ ์ด๊ณ  ์œ ์—ฐํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ธฐ์ˆ ์„ ํ™•์žฅํ•˜์—ฌ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ[523] ๋ฐ ๋‹ค๊ตญ์–ด ์ž‘์—…[524]์„ ํ•ด๊ฒฐํ•˜๋ ค๋Š” ๋ช‡ ๊ฐ€์ง€ ์˜ˆ๋น„์ ์ธ ์‹œ๋„๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
Arithmetic reasoning is another challenging reasoning capability to achieve. GSM8K for example contains grade school mathematical questions with respect to their answers. Table IX provides an insight for different model comparisons.
์‚ฐ์ˆ ์  ์ถ”๋ก ์€ ๋˜ ๋‹ค๋ฅธ ๋„์ „์ ์ธ ์ถ”๋ก  ๋Šฅ๋ ฅ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GSM8K๋Š” ์ดˆ๋“ฑํ•™๊ต ์ˆ˜์ค€์˜ ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ๋‹ต๊ณผ ํ•จ๊ป˜ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Table IX๋Š” ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ๋น„๊ต์— ๋Œ€ํ•œ ํ†ต์ฐฐ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์‚ฐ์ˆ  ์ถ”๋ก ์€ ๋‹ฌ์„ฑํ•˜๊ธฐ ์–ด๋ ค์šด ๋˜ ๋‹ค๋ฅธ ์ถ”๋ก  ๋Šฅ๋ ฅ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด GSM8K์—๋Š” ๋‹ต๊ณผ ๊ด€๋ จ๋œ ์ดˆ๋“ฑํ•™๊ต ์ˆ˜ํ•™ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ‘œ IX๋Š” ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ๋น„๊ต์— ๋Œ€ํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
2. Heart Health: HIIT workouts can place a significant amount of stress on the heart, which can be dangerous for individuals with heart conditions or histories. Athletes should consult a doctor before starting any new workt regimen.
2. ์‹ฌ์žฅ ๊ฑด๊ฐ•: HIIT ์šด๋™์€ ์‹ฌ์žฅ์— ์ƒ๋‹นํ•œ ์ŠคํŠธ๋ ˆ์Šค๋ฅผ ์ค„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์‹ฌ์žฅ ์งˆํ™˜์ด๋‚˜ ๊ด€๋ จ ๋ณ‘๋ ฅ์ด ์žˆ๋Š” ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ์œ„ํ—˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šด๋™ ์„ ์ˆ˜๋“ค์€ ์ƒˆ๋กœ์šด ์šด๋™ ํ”„๋กœ๊ทธ๋žจ์„ ์‹œ์ž‘ํ•˜๊ธฐ ์ „์— ์˜์‚ฌ์™€ ์ƒ๋‹ดํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
2. ์‹ฌ์žฅ ๊ฑด๊ฐ•: HIIT ์šด๋™์€ ์‹ฌ์žฅ์— ์ƒ๋‹นํ•œ ์ŠคํŠธ๋ ˆ์Šค๋ฅผ ๊ฐ€ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์‹ฌ์žฅ ์งˆํ™˜์ด ์žˆ๊ฑฐ๋‚˜ ๋ณ‘๋ ฅ์ด ์žˆ๋Š” ์‚ฌ๋žŒ์—๊ฒŒ๋Š” ์œ„ํ—˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šด๋™์„ ์ˆ˜๋Š” ์ƒˆ๋กœ์šด ์šด๋™ ์š”๋ฒ•์„ ์‹œ์ž‘ํ•˜๊ธฐ ์ „์— ์˜์‚ฌ์™€ ์ƒ๋‹ดํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
The scaling law's curve fits seem to be closely aligned with the actual data points in all three graphs. This indicates that Eq.(1) effectively models the relationship between training data size and model performance in these cases.
์Šค์ผ€์ผ๋ง ๋ฒ•์น™์˜ ๊ณก์„  ํ”ผํŒ…์€ ์„ธ ๊ฐœ์˜ ๊ทธ๋ž˜ํ”„ ๋ชจ๋‘์—์„œ ์‹ค์ œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์™€ ๋ฐ€์ ‘ํ•˜๊ฒŒ ์ผ์น˜ํ•˜๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๋ณด์ž…๋‹ˆ๋‹ค. ์ด๋Š” Eq.(1)์ด ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ์—์„œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ์™€ ๋ชจ๋ธ ์„ฑ๋Šฅ ๊ฐ„์˜ ๊ด€๊ณ„๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๋ชจ๋ธ๋งํ•˜๊ณ  ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์Šค์ผ€์ผ๋ง ๋ฒ•์น™์˜ ๊ณก์„ ์€ ์„ธ ๊ทธ๋ž˜ํ”„ ๋ชจ๋‘์—์„œ ์‹ค์ œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์™€ ๋ฐ€์ ‘ํ•˜๊ฒŒ ์ผ์น˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ธ๋‹ค. ์ด๋Š” Eq.(1)์ด ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ์™€ ๋ชจ๋ธ ์„ฑ๋Šฅ ์‚ฌ์ด์˜ ๊ด€๊ณ„๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๋ชจ๋ธ๋งํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค.
feedback (RLHF) based on the base model. The weights of these models are denoted as \(\theta_{PLM}\) and \(\theta_{chat}\), respectively, where \(\theta_{PLM},\theta_{chat}\in\mathbb{R}^{d}\), and \(d\) is the number of parameters.
๊ธฐ๋ณธ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ํ”ผ๋“œ๋ฐฑ (RLHF). ์ด ๋ชจ๋ธ๋“ค์˜ ๊ฐ€์ค‘์น˜๋Š” ๊ฐ๊ฐ \(\theta_{PLM}\) ๋ฐ \(\theta_{chat}\)๋กœ ํ‘œ์‹œ๋˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(\theta_{PLM}, \theta_{chat}\in\mathbb{R}^{d}\)์ด๊ณ , \(d\)๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ์ˆ˜์ž…๋‹ˆ๋‹ค.
๋ฒ ์ด์Šค ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ”ผ๋“œ๋ฐฑ(RLHF)์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ๊ฐ€์ค‘์น˜๋Š” ๊ฐ๊ฐ \(\theta_{PLM}\) ๋ฐ \(\theta_{chat}\)๋กœ ํ‘œ์‹œ๋˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(\theta_{PLM},\theta_{chat}\in\mathbb{R}^{d}\), \(d\)๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ์ˆ˜์ž…๋‹ˆ๋‹ค.
**Input** Given a chat history separated by new lines, generates an informative, knowledgeable and engaging response: Hello, I love German shepherd dogs and you? - My favorite dog, they are officially known as German Shepherd Dogs?
**์ž…๋ ฅ** ์—ฌ๋Ÿฌ ์ค„๋กœ ๊ตฌ๋ถ„๋œ ์ฑ„ํŒ… ๊ธฐ๋ก์„ ๋ฐ”ํƒ•์œผ๋กœ, ์ •๋ณด๊ฐ€ ํ’๋ถ€ํ•˜๊ณ  ์ง€์‹์ด ์žˆ์œผ๋ฉฐ ํฅ๋ฏธ๋กœ์šด ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค: ์•ˆ๋…•ํ•˜์„ธ์š”, ์ €๋Š” ์ €๋จผ ์…ฐํผ๋“œ ๊ฐœ๋ฅผ ์ข‹์•„ํ•ฉ๋‹ˆ๋‹ค. ๋‹น์‹ ์€์š”? - ์ œ๊ฐ€ ๊ฐ€์žฅ ์ข‹์•„ํ•˜๋Š” ๊ฐœ์ž…๋‹ˆ๋‹ค. ์ด๋“ค์€ ๊ณต์‹์ ์œผ๋กœ ์ €๋จผ ์…ฐํผ๋“œ ๊ฐœ๋กœ ์•Œ๋ ค์ ธ ์žˆ๋‚˜์š”?
**์ž…๋ ฅ** ์ƒˆ ์ค„๋กœ ๊ตฌ๋ถ„๋œ ์ฑ„ํŒ… ๊ธฐ๋ก์ด ์ฃผ์–ด์ง€๋ฉด ์œ ์ตํ•˜๊ณ , ์ง€์‹์ด ํ’๋ถ€ํ•˜๋ฉฐ, ๋งค๋ ฅ์ ์ธ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค: ์•ˆ๋…•ํ•˜์„ธ์š”, ์ €๋Š” ์ €๋จผ ์…ฐํผ๋“œ ๊ฐœ๋ฅผ ์ข‹์•„ํ•ด์š”? - ์ œ๊ฐ€ ๊ฐ€์žฅ ์ข‹์•„ํ•˜๋Š” ๊ฐœ, ๊ณต์‹์ ์œผ๋กœ ์ €๋จผ ์…ฐํผ๋“œ ๊ฐœ๋กœ ์•Œ๋ ค์ ธ ์žˆ๋‚˜์š”?
**The Milestones of Language Models**. Based on all the exploration efforts, two major milestones have been achieved by OpenAI, namely ChatGPT [131] and GPT-4 [46], which have largely raised the capacity bar of existing AI systems.
**์–ธ์–ด ๋ชจ๋ธ์˜ ์ด์ •ํ‘œ**. ๋ชจ๋“  ํƒ๊ตฌ ๋…ธ๋ ฅ์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ, OpenAI๋Š” ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์ด์ •ํ‘œ์ธ ChatGPT [131]์™€ GPT-4 [46]๋ฅผ ๋‹ฌ์„ฑํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ๊ธฐ์กด AI ์‹œ์Šคํ…œ์˜ ์—ญ๋Ÿ‰์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค.
**์–ธ์–ด ๋ชจ๋ธ์˜ ์ด์ •ํ‘œ**. ๋ชจ๋“  ํƒ์ƒ‰ ๋…ธ๋ ฅ์„ ๋ฐ”ํƒ•์œผ๋กœ OpenAI๋Š” ๊ธฐ์กด AI ์‹œ์Šคํ…œ์˜ ์—ญ๋Ÿ‰ ๊ธฐ์ค€์„ ํฌ๊ฒŒ ๋†’์ธ ChatGPT [131]์™€ GPT-4 [46]๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์ด์ •ํ‘œ๋ฅผ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.
**Question:** A treasure hunter found a buried treasure chest filled with gems. There were 175 diamonds, 35 fewer rubies than diamonds, and twice the number of emeralds than the rubies. How many of the gems were there in the chest?
**์งˆ๋ฌธ:** ํ•œ ๋ณด๋ฌผ ์‚ฌ๋ƒฅ๊พผ์ด ๋ณด์„์œผ๋กœ ๊ฐ€๋“ ์ฐฌ ๋ฌปํžŒ ๋ณด๋ฌผ ์ƒ์ž๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์ด์•„๋ชฌ๋“œ๋Š” 175๊ฐœ, ๋ฃจ๋น„๋Š” ๋‹ค์ด์•„๋ชฌ๋“œ๋ณด๋‹ค 35๊ฐœ ์ ์—ˆ๊ณ , ์—๋ฉ”๋ž„๋“œ๋Š” ๋ฃจ๋น„์˜ ๋‘ ๋ฐฐ์˜€์Šต๋‹ˆ๋‹ค. ์ƒ์ž ์•ˆ์— ๋ณด์„์€ ์ด ๋ช‡ ๊ฐœ์˜€์„๊นŒ์š”?
**์งˆ๋ฌธ:** ํ•œ ๋ณด๋ฌผ ์‚ฌ๋ƒฅ๊พผ์ด ๋ณด๋ฌผ ์ƒ์ž๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ณด๋ฌผ ์ƒ์ž์—๋Š” 175๊ฐœ์˜ ๋‹ค์ด์•„๋ชฌ๋“œ, ๋‹ค์ด์•„๋ชฌ๋“œ๋ณด๋‹ค 35๊ฐœ ์ ์€ ๋ฃจ๋น„, ๋ฃจ๋น„๋ณด๋‹ค ๋‘ ๋ฐฐ ๋งŽ์€ ์—๋ฉ”๋ž„๋“œ๊ฐ€ ๋“ค์–ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ƒ์ž ์•ˆ์—๋Š” ๋ช‡ ๊ฐœ์˜ ๋ณด์„์ด ๋“ค์–ด ์žˆ์—ˆ์„๊นŒ์š”?
The data we retrieve at this stage, called RW-Raw, corresponds to what we can extract with the minimal amount of filtering. At this stage, only 48% of the original documents are left, mostly filtered out by language identification.
์ด ๋‹จ๊ณ„์—์„œ ์šฐ๋ฆฌ๊ฐ€ ์ˆ˜์ง‘ํ•œ ๋ฐ์ดํ„ฐ์ธ RW-Raw๋Š” ์ตœ์†Œํ•œ์˜ ํ•„ํ„ฐ๋ง์œผ๋กœ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—์„œ๋Š” ์›๋ณธ ๋ฌธ์„œ์˜ 48%๋งŒ ๋‚จ์•„ ์žˆ์œผ๋ฉฐ, ๋Œ€๋ถ€๋ถ„์€ ์–ธ์–ด ์‹๋ณ„์— ์˜ํ•ด ํ•„ํ„ฐ๋ง๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
์ด ๋‹จ๊ณ„์—์„œ ๊ฒ€์ƒ‰ํ•˜๋Š” ๋ฐ์ดํ„ฐ(RW-Raw๋ผ๊ณ  ํ•จ)๋Š” ์ตœ์†Œํ•œ์˜ ํ•„ํ„ฐ๋ง์œผ๋กœ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—์„œ๋Š” ์›๋ฌธ์˜ 48%๋งŒ ๋‚จ๊ฒŒ ๋˜๋ฉฐ, ๋Œ€๋ถ€๋ถ„ ์–ธ์–ด ์‹๋ณ„์— ์˜ํ•ด ๊ฑธ๋Ÿฌ์ง‘๋‹ˆ๋‹ค.
Large-scale language models have showcased immense capabilities, but they also come with substantial overhead. Balancing the trade-off between performance and overhead may require significant attention in future research endeavors.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ์—„์ฒญ๋‚œ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ, ์ƒ๋‹นํ•œ ์˜ค๋ฒ„ํ—ค๋“œ๋„ ๋™๋ฐ˜ํ•ฉ๋‹ˆ๋‹ค. ์„ฑ๋Šฅ๊ณผ ์˜ค๋ฒ„ํ—ค๋“œ ๊ฐ„์˜ ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ๊ฒƒ์€ ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ ์ค‘์š”ํ•œ ๊ณผ์ œ๊ฐ€ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ์—„์ฒญ๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์ง€๋งŒ, ๊ทธ๋งŒํผ์˜ ์˜ค๋ฒ„ํ—ค๋“œ๋„ ์ˆ˜๋ฐ˜ํ•ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์„ฑ๋Šฅ๊ณผ ์˜ค๋ฒ„ํ—ค๋“œ ์‚ฌ์ด์˜ ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•œ ๊ด€์‹ฌ์‚ฌ๊ฐ€ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
While RAG is becoming standard practice in commercially available LLMs, the reliability of such systems is still understudied. Our experiments uncover several mechanisms that modulate the degree to which LLMs adhere to RAG systems.
์ƒ์—…์ ์œผ๋กœ ์ด์šฉ ๊ฐ€๋Šฅํ•œ LLM์—์„œ RAG๊ฐ€ ํ‘œ์ค€ ๊ด€ํ–‰์ด ๋˜์–ด๊ฐ€๊ณ  ์žˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ์‹œ์Šคํ…œ์˜ ์‹ ๋ขฐ์„ฑ์€ ์—ฌ์ „ํžˆ ์ถฉ๋ถ„ํžˆ ์—ฐ๊ตฌ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์‹คํ—˜์€ LLM์ด RAG ์‹œ์Šคํ…œ์„ ๋”ฐ๋ฅด๋Š” ์ •๋„๋ฅผ ์กฐ์ ˆํ•˜๋Š” ์—ฌ๋Ÿฌ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๋ฐํ˜€๋ƒ…๋‹ˆ๋‹ค.
RAG๊ฐ€ ์ƒ์šฉํ™”๋œ LLM์˜ ํ‘œ์ค€ ๊ด€ํ–‰์ด ๋˜๊ณ  ์žˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ์‹œ์Šคํ…œ์˜ ์‹ ๋ขฐ์„ฑ์€ ์•„์ง ์—ฐ๊ตฌ๋˜์ง€ ์•Š์€ ์ƒํƒœ์ž…๋‹ˆ๋‹ค. ์ €ํฌ์˜ ์‹คํ—˜์€ LLM์ด RAG ์‹œ์Šคํ…œ์„ ์ค€์ˆ˜ํ•˜๋Š” ์ •๋„๋ฅผ ์กฐ์ ˆํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
Figure 4: Data distribution of the 53 natural languages, aside from English,we used for pre-training. The number within each bar indicates the percent of the overall multilingual distribution that an individual language comprises.
Figure 4: ์‚ฌ์ „ ํ›ˆ๋ จ์— ์‚ฌ์šฉ๋œ ์˜์–ด๋ฅผ ์ œ์™ธํ•œ 53๊ฐœ ์ž์—ฐ์–ด์˜ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ. ๊ฐ ๋ง‰๋Œ€ ์•ˆ์˜ ์ˆซ์ž๋Š” ๊ฐœ๋ณ„ ์–ธ์–ด๊ฐ€ ์ „์ฒด ๋‹ค๊ตญ์–ด ๋ถ„ํฌ์—์„œ ์ฐจ์ง€ํ•˜๋Š” ๋น„์œจ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 4: ์‚ฌ์ „ ํ•™์Šต์— ์‚ฌ์šฉํ•œ ์˜์–ด๋ฅผ ์ œ์™ธํ•œ 53๊ฐœ ์ž์—ฐ์–ด์˜ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ. ๊ฐ ๋ง‰๋Œ€ ์•ˆ์˜ ์ˆซ์ž๋Š” ๊ฐœ๋ณ„ ์–ธ์–ด๊ฐ€ ์ „์ฒด ๋‹ค๊ตญ์–ด ๋ถ„ํฌ์—์„œ ์ฐจ์ง€ํ•˜๋Š” ๋น„์œจ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
Conversation quality was assessed using patient actor ratings, specialist ratings, and outputs from auto-evaluation. Figure A.5 and A.6 show two example consultations for the same simulated patient from AMIE and PCP, respectively.
๋Œ€ํ™” ํ’ˆ์งˆ์€ ํ™˜์ž ๋ฐฐ์šฐ ํ‰๊ฐ€, ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€ ๋ฐ ์ž๋™ ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ A.5์™€ A.6์€ ๊ฐ๊ฐ AMIE์™€ PCP์—์„œ ๋™์ผํ•œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜์ž์— ๋Œ€ํ•œ ๋‘ ๊ฐ€์ง€ ์˜ˆ์‹œ ์ƒ๋‹ด์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
ํ™˜์ž ๋ฐฐ์šฐ์ž ํ‰๊ฐ€, ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€ ๋ฐ ์ž๋™ ํ‰๊ฐ€์˜ ๊ฒฐ๊ณผ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋Œ€ํ™” ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ–ˆ๋‹ค. ๊ทธ๋ฆผ A.5 ๋ฐ A.6์€ ๊ฐ๊ฐ AMIE ๋ฐ PCP์˜ ๋™์ผํ•œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜์ž์— ๋Œ€ํ•œ ๋‘ ๊ฐ€์ง€ ์˜ˆ์‹œ ํ˜‘์˜๋ฅผ ๋ณด์—ฌ์ค€๋‹ค.
As LLMs have revolutionized the way how we develop AI algorithms, it poses significant impact on the research community. In this part, we briefly review the advances that led by LLMs for several representative research directions.
LLMs๋Š” AI ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๊ฐœ๋ฐœ ๋ฐฉ์‹์„ ํ˜์‹ ์ ์œผ๋กœ ๋ณ€ํ™”์‹œ์ผฐ์œผ๋ฉฐ, ์ด๋Š” ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” LLMs๊ฐ€ ์ฃผ๋„ํ•œ ์—ฌ๋Ÿฌ ๋Œ€ํ‘œ์ ์ธ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์—์„œ์˜ ๋ฐœ์ „์„ ๊ฐ„๋žตํžˆ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค.
LLM์€ AI ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐฉ์‹์— ํ˜๋ช…์„ ์ผ์œผ์ผฐ๊ธฐ ๋•Œ๋ฌธ์— ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ์ƒ๋‹นํ•œ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” ๋ช‡ ๊ฐ€์ง€ ๋Œ€ํ‘œ์ ์ธ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์— ๋Œ€ํ•ด LLM์ด ์ฃผ๋„ํ•œ ๋ฐœ์ „์„ ๊ฐ„๋žตํ•˜๊ฒŒ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค.
1. Reflection Tokens. 2. Self-Rag Training 3. Self-Rag Inference 3. Experimental Details 1. More Details of Training 2. More Details of Evaluations 4. Results 5.1 Analysis 5.2 Human Evaluation Examples 5.3 Qualitative Examples 5.
1. Reflection Tokens. 2. Self-Rag Training 3. Self-Rag Inference 3. ์‹คํ—˜ ์„ธ๋ถ€์‚ฌํ•ญ 1. ํ›ˆ๋ จ์˜ ์„ธ๋ถ€์‚ฌํ•ญ 2. ํ‰๊ฐ€์˜ ์„ธ๋ถ€์‚ฌํ•ญ 4. ๊ฒฐ๊ณผ 5.1 ๋ถ„์„ 5.2 ์ธ๊ฐ„ ํ‰๊ฐ€ ์˜ˆ์‹œ 5.3 ์ •์„ฑ์  ์˜ˆ์‹œ 5.
1. ๋ฐ˜์‚ฌ ํ† ํฐ. 2. ์…€ํ”„๋ž˜๊ทธ ํŠธ๋ ˆ์ด๋‹ 3. ์…€ํ”„๋ž˜๊ทธ ์ถ”๋ก 3. ์‹คํ—˜ ์„ธ๋ถ€ ์‚ฌํ•ญ 1. ํ›ˆ๋ จ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ 2. ํ‰๊ฐ€์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ4. ๊ฒฐ๊ณผ 5.1 ๋ถ„์„ 5.2 ์ธ๊ฐ„ ํ‰๊ฐ€ ์˜ˆ์‹œ 5.3 ์ •์„ฑ์  ์˜ˆ์‹œ5.
This section provides a comprehensive overview of the baseline models in our experiments. Subsequently, we mainly elucidate the performance metrics of our models on two prevalent mathematical benchmarks: GSM8k [42] and MATH [43].
์ด ์„น์…˜์—์„œ๋Š” ์‹คํ—˜์—์„œ ์‚ฌ์šฉ๋œ ๊ธฐ์ค€ ๋ชจ๋ธ๋“ค์— ๋Œ€ํ•œ ์ข…ํ•ฉ์ ์ธ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด์–ด์„œ, ์ฃผ๋กœ ๋‘ ๊ฐ€์ง€ ์ฃผ์š”ํ•œ ์ˆ˜ํ•™์  ๋ฒค์น˜๋งˆํฌ์ธ GSM8k [42]์™€ MATH [43]์—์„œ ์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ์ง€ํ‘œ๋ฅผ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ์‹คํ—˜์˜ ๊ธฐ์ค€ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ดํ›„์—๋Š” ์ฃผ๋กœ ๋‘ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ ์ˆ˜ํ•™ ๋ฒค์น˜๋งˆํฌ์ธ GSM8k [42] ๋ฐ MATH [43]์—์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ๋ฉ”ํŠธ๋ฆญ์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
One Step Missing Error (OSME). Chains of thought which were correct except that they were missing a single step. These chains of thoughts could be rewritten to be correct by adding in an additional reasoning step that was missed.
One Step Missing Error (OSME). ๋‹จ ํ•˜๋‚˜์˜ ๋‹จ๊ณ„๊ฐ€ ๋ˆ„๋ฝ๋œ ๊ฒƒ์„ ์ œ์™ธํ•˜๊ณ ๋Š” ์˜ฌ๋ฐ”๋ฅธ ์‚ฌ๊ณ ์˜ ์—ฐ์‡„. ์ด๋Ÿฌํ•œ ์‚ฌ๊ณ ์˜ ์—ฐ์‡„๋Š” ๋ˆ„๋ฝ๋œ ์ถ”๊ฐ€์ ์ธ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๋‹ค์‹œ ์ž‘์„ฑ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ•œ ๋‹จ๊ณ„ ๋ˆ„๋ฝ ์˜ค๋ฅ˜(OSME). ํ•œ ๋‹จ๊ณ„๋งŒ ๋ˆ„๋ฝ๋˜์—ˆ์„ ๋ฟ์€ ์ •๋‹ต์ธ ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ์€ ๋ˆ„๋ฝ๋œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๋‹ค์‹œ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Setting a threshold of duplicates.An important question is how long must a substring match be before it is counted as a duplicate. In Figure 5, we plot the frequency of substring matches within the four datasets we will consider.
์ค‘๋ณต์˜ ์ž„๊ณ„๊ฐ’ ์„ค์ •. ์ค‘์š”ํ•œ ์งˆ๋ฌธ์€ ์„œ๋ธŒ์ŠคํŠธ๋ง์ด ์ค‘๋ณต์œผ๋กœ ๊ฐ„์ฃผ๋˜๊ธฐ ์ „์— ์–ผ๋งˆ๋‚˜ ๊ธธ์–ด์•ผ ํ•˜๋Š”๊ฐ€์ž…๋‹ˆ๋‹ค. Figure 5์—์„œ๋Š” ์šฐ๋ฆฌ๊ฐ€ ๊ณ ๋ คํ•  ๋„ค ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹ ๋‚ด์—์„œ ์„œ๋ธŒ์ŠคํŠธ๋ง ์ผ์น˜ ๋นˆ๋„๋ฅผ ๊ทธ๋ž˜ํ”„๋กœ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์ค‘๋ณต์˜ ์ž„๊ณ„๊ฐ’ ์„ค์ •.์ค‘์š”ํ•œ ์งˆ๋ฌธ์€ ์ค‘๋ณต์œผ๋กœ ๊ฐ„์ฃผ๋˜๊ธฐ ์ „์— ํ•˜์œ„ ๋ฌธ์ž์—ด ์ผ์น˜๊ฐ€ ์–ผ๋งˆ๋‚˜ ์˜ค๋ž˜ ์ง€์†๋˜์–ด์•ผ ํ•˜๋Š”๊ฐ€ ํ•˜๋Š” ๊ฒƒ์ด๋‹ค. ๊ทธ๋ฆผ 5์—์„œ๋Š” ์šฐ๋ฆฌ๊ฐ€ ๊ณ ๋ คํ•  ๋„ค ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ ๋‚ด์—์„œ ํ•˜์œ„ ๋ฌธ์ž์—ด ์ผ์น˜ ๋นˆ๋„๋ฅผ ๊ทธ๋ž˜ํ”„๋กœ ํ‘œ์‹œํ•œ๋‹ค.
Self-Consistency [159] utilizes an ensemble-based method, where the LLM is prompted to generate multiple responses to the same query. The consistency among these responses serves as an indicator of their accuracy and reliability.
Self-Consistency [159]๋Š” ์•™์ƒ๋ธ” ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ, ๋™์ผํ•œ ์ฟผ๋ฆฌ์— ๋Œ€ํ•ด LLM์ด ์—ฌ๋Ÿฌ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‘๋‹ต๋“ค ๊ฐ„์˜ ์ผ๊ด€์„ฑ์€ ๊ทธ๋“ค์˜ ์ •ํ™•์„ฑ๊ณผ ์‹ ๋ขฐ์„ฑ์„ ๋‚˜ํƒ€๋‚ด๋Š” ์ง€ํ‘œ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
์ž๊ธฐ ์ผ๊ด€์„ฑ[159]์€ ์•™์ƒ๋ธ” ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์„ ํ™œ์šฉํ•˜๋Š”๋ฐ, ์—ฌ๊ธฐ์„œ LLM์€ ๋™์ผํ•œ ์ฟผ๋ฆฌ์— ๋Œ€ํ•ด ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‘๋‹ต ๊ฐ„์˜ ์ผ๊ด€์„ฑ์€ ์ •ํ™•์„ฑ๊ณผ ์‹ ๋ขฐ์„ฑ์˜ ์ง€ํ‘œ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
To delve into this, we examine GPT2\({}_{1/4}\), which is GPT2 with its MLP layer reduced from \(d\to 4d\to d\) to \(d\to d\to d\) (thus, \(1/4\) of its original size), and GPT2\({}_{0}\), which is GPT2 but without any MLP layer.
์ด๋ฅผ ์ž์„ธํžˆ ์‚ดํŽด๋ณด๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” GPT2\({}_{1/4}\)์™€ GPT2\({}_{0}\)๋ฅผ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. GPT2\({}_{1/4}\)๋Š” MLP ๋ ˆ์ด์–ด๊ฐ€ \(d\to 4d\to d\)์—์„œ \(d\to d\to d\)๋กœ ์ถ•์†Œ๋œ GPT2 (์ฆ‰, ์›๋ž˜ ํฌ๊ธฐ์˜ \(1/4\))์ด๋ฉฐ, GPT2\({}_{0}\)๋Š” MLP ๋ ˆ์ด์–ด๊ฐ€ ์ „ํ˜€ ์—†๋Š” GPT2์ž…๋‹ˆ๋‹ค.
์ด๋ฅผ ํƒ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด MLP ๋ ˆ์ด์–ด๊ฐ€ \(d\to 4d\to d\)์—์„œ \(d\to d\to d\)๋กœ ์ถ•์†Œ๋œ GPT2\({}_{1/4}\)(๋”ฐ๋ผ์„œ ์›๋ž˜ ํฌ๊ธฐ์˜ \(1/4\))์™€ MLP ๋ ˆ์ด์–ด๊ฐ€ ์—†๋Š” GPT2\({}_{0}\), ์ฆ‰ GPT2๋ฅผ ์‚ดํŽด๋ณธ๋‹ค.
Footnote 7: It is possible to explicitly compute the integral in the denominator as \(\sqrt{2\pi}(2\Phi(\delta)-1)+2e^{-\delta^{2}/2}/\delta\) where \(\Phi\) is the standard normal distributionโ€™s cumulative distribution function.
Footnote 7: ๋ถ„๋ชจ์˜ ์ ๋ถ„์„ \(\sqrt{2\pi}(2\Phi(\delta)-1)+2e^{-\delta^{2}/2}/\delta\)๋กœ ๋ช…์‹œ์ ์œผ๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ \(\Phi\)๋Š” ํ‘œ์ค€ ์ •๊ทœ ๋ถ„ํฌ์˜ ๋ˆ„์  ๋ถ„ํฌ ํ•จ์ˆ˜์ž…๋‹ˆ๋‹ค.
๊ฐ์ฃผ 7: ๋ถ„๋ชจ์˜ ์ ๋ถ„์€ ํ‘œ์ค€ ์ •๊ทœ ๋ถ„ํฌ์˜ ๋ˆ„์  ๋ถ„ํฌ ํ•จ์ˆ˜์ธ \(\Phi\)๊ฐ€ \(\sqrt{2\pi}(2\Phi(\delta)-1)+2e^{-\delta^{2}/2}/\delta\)๋กœ ๋ช…์‹œ์ ์œผ๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
This update rule (\(Linear+Delta\)) leaves the associative matrix unmodified if the KV binding already exists in the memory while still tracking the same normalization term as the former one (\(Linear\)) for numerical stability.
์ด ์—…๋ฐ์ดํŠธ ๊ทœ์น™ (\(Linear+Delta\))์€ KV ๋ฐ”์ธ๋”ฉ์ด ๋ฉ”๋ชจ๋ฆฌ์— ์ด๋ฏธ ์กด์žฌํ•˜๋Š” ๊ฒฝ์šฐ ์—ฐ๊ด€ ํ–‰๋ ฌ์„ ์ˆ˜์ •ํ•˜์ง€ ์•Š์œผ๋ฉด์„œ๋„ ์ˆ˜์น˜์  ์•ˆ์ •์„ฑ์„ ์œ„ํ•ด ์ด์ „ ๊ทœ์น™ (\(Linear\))๊ณผ ๋™์ผํ•œ ์ •๊ทœํ™” ํ•ญ์„ ์ถ”์ ํ•ฉ๋‹ˆ๋‹ค.
์ด ์—…๋ฐ์ดํŠธ ๊ทœ์น™(\(Linear+Delta\))์€ ์—ฐ๊ด€ ํ–‰๋ ฌ์„ ์ˆ˜์ •ํ•˜์ง€ ์•Š๊ณ , ์ˆ˜์น˜ ์•ˆ์ •์„ฑ์„ ์œ„ํ•ด ์ด์ „ ๊ทœ์น™(\(Linear\))๊ณผ ๋™์ผํ•œ ์ •๊ทœํ™” ํ•ญ์„ ์ถ”์ ํ•œ๋‹ค.
Fuzzy deduplication.We remove similar documents by applying MinHash (Broder, 1997): for each document, we compute a sketch and measure its approximate similarity with other documents, eventually removing pairs with high overlap.
์œ ์‚ฌํ•œ ๋ฌธ์„œ ์ œ๊ฑฐ. ์šฐ๋ฆฌ๋Š” MinHash (Broder, 1997)๋ฅผ ์ ์šฉํ•˜์—ฌ ์œ ์‚ฌํ•œ ๋ฌธ์„œ๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค: ๊ฐ ๋ฌธ์„œ์— ๋Œ€ํ•ด ์Šค์ผ€์น˜๋ฅผ ๊ณ„์‚ฐํ•˜๊ณ  ๋‹ค๋ฅธ ๋ฌธ์„œ์™€์˜ ๋Œ€๋žต์ ์ธ ์œ ์‚ฌ์„ฑ์„ ์ธก์ •ํ•˜์—ฌ, ๊ฒฐ๊ตญ ๋†’์€ ์ค‘๋ณต์„ ๊ฐ€์ง„ ์Œ์„ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
ํผ์ง€ ์ค‘๋ณต ์ œ๊ฑฐ.MinHash(Broder, 1997)๋ฅผ ์ ์šฉํ•˜์—ฌ ์œ ์‚ฌํ•œ ๋ฌธ์„œ๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค: ๊ฐ ๋ฌธ์„œ์— ๋Œ€ํ•ด ์Šค์ผ€์น˜๋ฅผ ๊ณ„์‚ฐํ•˜๊ณ  ๋‹ค๋ฅธ ๋ฌธ์„œ์™€์˜ ๋Œ€๋žต์ ์ธ ์œ ์‚ฌ์„ฑ์„ ์ธก์ •ํ•˜์—ฌ ๊ฒฐ๊ตญ ๊ฒน์น˜๋Š” ๋ถ€๋ถ„์ด ๋งŽ์€ ์Œ์„ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
A: Let's think step by step. She eats 3 daily for breakfast. She bakes 4 daily into muffins for her friends. That leaves 9. She sells 9 for \(\(2\)), so \(\(9\)times 2=\(518\)\()\). So Janet makes $18 selling duck eggs each day.
A: ๋‹จ๊ณ„๋ณ„๋กœ ์ƒ๊ฐํ•ด ๋ด…์‹œ๋‹ค. ๊ทธ๋…€๋Š” ์•„์นจ ์‹์‚ฌ๋กœ ๋งค์ผ 3๊ฐœ๋ฅผ ๋จน์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ์นœ๊ตฌ๋“ค์„ ์œ„ํ•ด ๋งค์ผ 4๊ฐœ๋ฅผ ๋จธํ•€์œผ๋กœ ๊ตฌ์›Œ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋ฉด 9๊ฐœ๊ฐ€ ๋‚จ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” 9๊ฐœ๋ฅผ \(2\)์— ํŒ๋‹ˆ๋‹ค, ๊ทธ๋ž˜์„œ \(9 \times 2 = 18\)์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Janet์€ ๋งค์ผ ์˜ค๋ฆฌ ์•Œ์„ ํŒ”์•„ 18๋‹ฌ๋Ÿฌ๋ฅผ ๋ฒŒ์–ด๋“ค์ž…๋‹ˆ๋‹ค.
A: ์ฐจ๊ทผ์ฐจ๊ทผ ์ƒ๊ฐํ•ด ๋ด…์‹œ๋‹ค. ๊ทธ๋…€๋Š” ์•„์นจ ์‹์‚ฌ๋กœ ๋งค์ผ 3๊ฐœ๋ฅผ ๋จน์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ์นœ๊ตฌ๋“ค์„ ์œ„ํ•ด ๋งค์ผ 4๊ฐœ๋ฅผ ๋จธํ•€์œผ๋กœ ๊ตฝ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋ฉด 9๊ฐœ๊ฐ€ ๋‚จ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” 9๊ฐœ๋ฅผ \(\(2\))์— ํŒ”๊ธฐ ๋•Œ๋ฌธ์— \(\(9\) ๊ณฑํ•˜๊ธฐ 2=\(518\)\()\). ๋”ฐ๋ผ์„œ ์ž๋„ท์€ ๋งค์ผ ์˜ค๋ฆฌ์•Œ์„ ํŒ”์•„ 18๋‹ฌ๋Ÿฌ๋ฅผ ๋ฒŒ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Dynamic teams where agents are brought in and out of the system based on need have also been shown to be effective. This ensures that all agents participating in the planning or execution of tasks are fit for that round of work.
๋™์  ํŒ€์€ ํ•„์š”์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ๊ฐ€ ์‹œ์Šคํ…œ์— ํˆฌ์ž…๋˜๊ฑฐ๋‚˜ ์ œ์™ธ๋˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์šด์˜๋˜๋ฉฐ, ์ด๋Š” ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ณ„ํš์ด๋‚˜ ์ž‘์—… ์‹คํ–‰์— ์ฐธ์—ฌํ•˜๋Š” ๋ชจ๋“  ์—์ด์ „ํŠธ๊ฐ€ ํ•ด๋‹น ์ž‘์—… ๋ผ์šด๋“œ์— ์ ํ•ฉํ•˜๋„๋ก ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค.
ํ•„์š”์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ๋ฅผ ์‹œ์Šคํ…œ์— ๋“œ๋‚˜๋“ค๊ฒŒ ํ•˜๋Š” ๋™์  ํŒ€๋„ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์ž‘์—…์˜ ๊ณ„ํš ๋˜๋Š” ์‹คํ–‰์— ์ฐธ์—ฌํ•˜๋Š” ๋ชจ๋“  ์—์ด์ „ํŠธ๊ฐ€ ํ•ด๋‹น ์ž‘์—…์— ์ ํ•ฉํ•œ์ง€ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Recall from Section 10 that our dataset is a mixture, with 1/8 of the tokens coming from \(\mathsf{bioS}(N)\) for various \(N\) (referred to as "useful data"), and the remaining 7/8 from "junk data." We explored three scenarios:
Section 10์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, ์šฐ๋ฆฌ์˜ ๋ฐ์ดํ„ฐ์…‹์€ ํ˜ผํ•ฉ๋œ ํ˜•ํƒœ๋กœ, ๋‹ค์–‘ํ•œ \(N\)์— ๋Œ€ํ•œ \(\mathsf{bioS}(N)\)์—์„œ 1/8์˜ ํ† ํฐ์ด ๋‚˜์˜ค๊ณ (์ด๋ฅผ "์œ ์šฉํ•œ ๋ฐ์ดํ„ฐ"๋ผ๊ณ  ํ•จ), ๋‚˜๋จธ์ง€ 7/8์€ "์“ธ๋ชจ์—†๋Š” ๋ฐ์ดํ„ฐ"๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์„ธ ๊ฐ€์ง€ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํƒ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค.
์„น์…˜ 10์—์„œ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์ด ํ˜ผํ•ฉ๋˜์–ด ์žˆ์œผ๋ฉฐ, ํ† ํฐ์˜ 1/8์€ ๋‹ค์–‘ํ•œ \(N\)์— ๋Œ€ํ•ด \(\mathsf{bioS}(N)\)์—์„œ ๋‚˜์˜ค๊ณ (์ด๋ฅผ "์œ ์šฉํ•œ ๋ฐ์ดํ„ฐ"๋ผ๊ณ  ํ•จ), ๋‚˜๋จธ์ง€ 7/8์€ "์ •ํฌ ๋ฐ์ดํ„ฐ"์—์„œ ๋‚˜์˜จ๋‹ค๋Š” ๊ฒƒ์„ ๊ธฐ์–ตํ•˜์„ธ์š”. ์šฐ๋ฆฌ๋Š” ์„ธ ๊ฐ€์ง€ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค:
We evaluate our models on a total of 13 different zero-shot benchmarks to assess their abilities across various natural language tasks like common sense reasoning, language and knowledge understanding and mathematical reasoning.
์šฐ๋ฆฌ๋Š” ์ƒ์‹ ์ถ”๋ก , ์–ธ์–ด ๋ฐ ์ง€์‹ ์ดํ•ด, ์ˆ˜ํ•™์  ์ถ”๋ก ๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ž‘์—…์—์„œ ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์ด 13๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ์ œ๋กœ์ƒท ๋ฒค์น˜๋งˆํฌ์—์„œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
์šฐ๋ฆฌ๋Š” ์ƒ์‹ ์ถ”๋ก , ์–ธ์–ด ๋ฐ ์ง€์‹ ์ดํ•ด, ์ˆ˜ํ•™์  ์ถ”๋ก ๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ž‘์—…์— ๊ฑธ์ณ ๊ทธ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์ด 13๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ 0์ƒท ๋ฒค์น˜๋งˆํฌ์—์„œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
adding AI-generated data might lead to better models, especially when generated by a stronger model (e.g. He et al. (2023); Shipard et al. (2023); Bansal and Grover (2023); Lin et al. (2023)). See Appendix A for more references.
AI๊ฐ€ ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•˜๋ฉด ๋” ๋‚˜์€ ๋ชจ๋ธ์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํŠนํžˆ ๋” ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๊ฒฝ์šฐ (์˜ˆ: He et al. (2023); Shipard et al. (2023); Bansal and Grover (2023); Lin et al. (2023))์— ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค. ๋” ๋งŽ์€ ์ฐธ๊ณ  ๋ฌธํ—Œ์€ ๋ถ€๋ก A๋ฅผ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค.
ํŠนํžˆ ๋” ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๊ฒฝ์šฐ(์˜ˆ: He et al. (2023); Shipard et al. (2023); Bansal and Grover (2023); Lin et al. (2023)) ๋” ๋‚˜์€ ๋ชจ๋ธ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ž์„ธํ•œ ์ฐธ๊ณ  ๋ฌธํ—Œ์€ ๋ถ€๋ก A๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
But as their eyes adjusted to the bright sunlight, they saw that the object was glowing with a soft, blue light. It seemed to pulse and undulate, as if it were alive. Without thinking, my character reached out and picked it up.
ํ•˜์ง€๋งŒ ๊ทธ๋“ค์˜ ๋ˆˆ์ด ๋ฐ์€ ํ–‡๋น›์— ์ ์‘ํ•˜์ž, ๊ทธ ๋ฌผ์ฒด๊ฐ€ ๋ถ€๋“œ๋Ÿฌ์šด ํŒŒ๋ž€๋น›์œผ๋กœ ๋น›๋‚˜๊ณ  ์žˆ๋Š” ๊ฒƒ์„ ๋ณด์•˜๋‹ค. ๊ทธ๊ฒƒ์€ ๋งˆ์น˜ ์‚ด์•„์žˆ๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๋งฅ๋™ํ•˜๊ณ  ๋ฌผ๊ฒฐ์น˜๋Š” ๊ฒƒ ๊ฐ™์•˜๋‹ค. ์ƒ๊ฐํ•  ๊ฒจ๋ฅผ๋„ ์—†์ด, ๋‚ด ์บ๋ฆญํ„ฐ๋Š” ์†์„ ๋ป—์–ด ๊ทธ๊ฒƒ์„ ์ง‘์–ด ๋“ค์—ˆ๋‹ค.
ํ•˜์ง€๋งŒ ๋ˆˆ์ด ๋ฐ์€ ํ–‡๋น›์— ์ ์‘ํ•˜์ž ๊ทธ ๋ฌผ์ฒด๊ฐ€ ๋ถ€๋“œ๋Ÿฌ์šด ํŒŒ๋ž€ ๋น›์œผ๋กœ ๋น›๋‚˜๋Š” ๊ฒƒ์„ ๋ณด์•˜์Šต๋‹ˆ๋‹ค. ๋งˆ์น˜ ์‚ด์•„์žˆ๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๋งฅ๋ฐ•์„ ๋›ฐ๊ณ  ๋ฌผ๊ฒฐ์น˜๋Š” ๊ฒƒ ๊ฐ™์•˜์Šต๋‹ˆ๋‹ค. ์ œ ์บ๋ฆญํ„ฐ๋Š” ์•„๋ฌด ์ƒ๊ฐ ์—†์ด ์†์„ ๋ป—์–ด ๊ทธ๊ฒƒ์„ ์ง‘์–ด ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค.
All benchmarks are publicly available for non-commercial research purposes, and we strictly limit their usage in this research work. We also carefully check these datasets and make sure that no personal information is involved.
๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ๋Š” ๋น„์ƒ์—…์  ์—ฐ๊ตฌ ๋ชฉ์ ์œผ๋กœ ๊ณต๊ฐœ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ๋Š” ์ด ์—ฐ๊ตฌ ์ž‘์—…์—์„œ ๊ทธ ์‚ฌ์šฉ์„ ์—„๊ฒฉํžˆ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ์…‹์„ ์‹ ์ค‘ํ•˜๊ฒŒ ๊ฒ€ํ† ํ•˜์—ฌ ๊ฐœ์ธ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์ง€ ์•Š๋„๋ก ํ™•์ธํ•ฉ๋‹ˆ๋‹ค.
๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ๋Š” ๋น„์ƒ์—…์  ์—ฐ๊ตฌ ๋ชฉ์ ์œผ๋กœ ๊ณต๊ฐœ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋ณธ ์—ฐ๊ตฌ ์ž‘์—…์—์„œ ์ด๋“ค์˜ ์‚ฌ์šฉ์„ ์—„๊ฒฉํ•˜๊ฒŒ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ๊ฒ€ํ† ํ•˜์—ฌ ๊ฐœ์ธ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์ง€ ์•Š์€์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค.
Tools in this context are external functions or services that LLMs can utilize. These tools extend the range of tasks an LLM can perform, from basic information retrieval to complex interactions with external databases or APIs.
์ด ๋ฌธ๋งฅ์—์„œ ๋„๊ตฌ๋Š” LLM์ด ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์™ธ๋ถ€ ํ•จ์ˆ˜๋‚˜ ์„œ๋น„์Šค๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋„๊ตฌ๋Š” ๊ธฐ๋ณธ์ ์ธ ์ •๋ณด ๊ฒ€์ƒ‰๋ถ€ํ„ฐ ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค๋‚˜ API์™€์˜ ๋ณต์žกํ•œ ์ƒํ˜ธ์ž‘์šฉ์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ LLM์ด ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ์ž‘์—…์˜ ๋ฒ”์œ„๋ฅผ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ๋งฅ๋ฝ์—์„œ ๋„๊ตฌ๋Š” LLM์ด ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์™ธ๋ถ€ ํ•จ์ˆ˜ ๋˜๋Š” ์„œ๋น„์Šค์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋„๊ตฌ๋Š” ๊ธฐ๋ณธ์ ์ธ ์ •๋ณด ๊ฒ€์ƒ‰๋ถ€ํ„ฐ ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค ๋˜๋Š” API์™€์˜ ๋ณต์žกํ•œ ์ƒํ˜ธ ์ž‘์šฉ์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ LLM์ด ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ์ž‘์—…์˜ ๋ฒ”์œ„๋ฅผ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค.
Now we will break down all the eggs from our starting number, $n$ =, into the pieces that we set up previously. For the beginning number, we have: This brings us back to our starting number, and we can simplify this expression.
์ด์ œ ์šฐ๋ฆฌ๋Š” ์‹œ์ž‘ ์ˆซ์ž์ธ $n$์„ ์ด์ „์— ์„ค์ •ํ•œ ์กฐ๊ฐ๋“ค๋กœ ๋‚˜๋ˆ„๊ฒ ์Šต๋‹ˆ๋‹ค. ์‹œ์ž‘ ์ˆซ์ž์— ๋Œ€ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํ‘œํ˜„์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: ์ด๋Š” ์šฐ๋ฆฌ๋ฅผ ์‹œ์ž‘ ์ˆซ์ž๋กœ ๋˜๋Œ๋ ค์ฃผ๋ฉฐ, ์ด ํ‘œํ˜„์„ ๋‹จ์ˆœํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด์ œ ์‹œ์ž‘ ๋ฒˆํ˜ธ์ธ $n$ =๋ถ€ํ„ฐ ๋ชจ๋“  ๊ณ„๋ž€์„ ์ด์ „์— ์„ค์ •ํ•œ ์กฐ๊ฐ์œผ๋กœ ๋‚˜๋ˆ„์–ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์‹œ์ž‘ ๋ฒˆํ˜ธ์˜ ๊ฒฝ์šฐ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค: ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์‹œ์ž‘ ๋ฒˆํ˜ธ๋กœ ๋Œ์•„์˜ค๊ณ  ์ด ์‹์„ ๋‹จ์ˆœํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
1. Injury: High-impact HIIT workouts can increase the risk of injury due to the rapid, explosive movements required. Athletes should properly warm up and cool down before and after HIIT workouts to reduce their risk of injury.
1. ๋ถ€์ƒ: ๊ณ ๊ฐ•๋„ HIIT ์šด๋™์€ ๋น ๋ฅด๊ณ  ํญ๋ฐœ์ ์ธ ์›€์ง์ž„์ด ์š”๊ตฌ๋˜๊ธฐ ๋•Œ๋ฌธ์— ๋ถ€์ƒ์˜ ์œ„ํ—˜์„ ์ฆ๊ฐ€์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šด๋™ ์„ ์ˆ˜๋“ค์€ HIIT ์šด๋™ ์ „ํ›„์— ์ ์ ˆํžˆ ์›Œ๋ฐ์—…๊ณผ ์ฟจ๋‹ค์šด์„ ํ•˜์—ฌ ๋ถ€์ƒ ์œ„ํ—˜์„ ์ค„์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
1. ๋ถ€์ƒ: ์ถฉ๊ฒฉ์ด ํฐ HIIT ์šด๋™์€ ๋น ๋ฅด๊ณ  ํญ๋ฐœ์ ์ธ ์›€์ง์ž„์ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ถ€์ƒ ์œ„ํ—˜์ด ๋†’์•„์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šด๋™์„ ์ˆ˜๋Š” ๋ถ€์ƒ ์œ„ํ—˜์„ ์ค„์ด๊ธฐ ์œ„ํ•ด HIIT ์šด๋™ ์ „ํ›„์— ์ ์ ˆํ•˜๊ฒŒ ์›Œ๋ฐ์—…๊ณผ ์ฟจ๋‹ค์šด์„ ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Table 7 shows the MMMU test set leaderboard by Yi-VL's release. We note that this area is currently actively under research, aligning with the community's advances, we will continuously improve the update Yi-VL's performance.
Table 7์€ Yi-VL์˜ ๋ฆด๋ฆฌ์Šค์— ๋”ฐ๋ฅธ MMMU ํ…Œ์ŠคํŠธ ์„ธํŠธ ๋ฆฌ๋”๋ณด๋“œ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ๋ถ„์•ผ๋Š” ํ˜„์žฌ ํ™œ๋ฐœํžˆ ์—ฐ๊ตฌ ์ค‘์ด๋ฉฐ, ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ๋ฐœ์ „์— ๋งž์ถฐ Yi-VL์˜ ์„ฑ๋Šฅ์„ ์ง€์†์ ์œผ๋กœ ๊ฐœ์„ ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
ํ‘œ 7์€ Yi-VL ์ถœ์‹œ์— ๋”ฐ๋ฅธ MMMU ํ…Œ์ŠคํŠธ ์„ธํŠธ ๋ฆฌ๋”๋ณด๋“œ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ๋ถ„์•ผ๋Š” ํ˜„์žฌ ํ™œ๋ฐœํ•˜๊ฒŒ ์—ฐ๊ตฌ๋˜๊ณ  ์žˆ๋Š” ๋ถ„์•ผ์ด๋ฉฐ, ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ๋ฐœ์ „์— ๋ฐœ๋งž์ถฐ Yi-VL์˜ ์„ฑ๋Šฅ์„ ์ง€์†์ ์œผ๋กœ ๊ฐœ์„ ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Fluency: Assesses how well the answer adheres to human language conventions, with coherent phrasing and clear expression. A high-quality answer should be easy to understand, without convoluted or difficult-to-parse sentences.
์œ ์ฐฝ์„ฑ: ๋‹ต๋ณ€์ด ์ธ๊ฐ„ ์–ธ์–ด์˜ ๊ทœ์น™์„ ์–ผ๋งˆ๋‚˜ ์ž˜ ๋”ฐ๋ฅด๋Š”์ง€ ํ‰๊ฐ€ํ•˜๋ฉฐ, ์ผ๊ด€๋œ ๋ฌธ๊ตฌ์™€ ๋ช…ํ™•ํ•œ ํ‘œํ˜„์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ์˜ ๋‹ต๋ณ€์€ ์ดํ•ดํ•˜๊ธฐ ์‰ฌ์›Œ์•ผ ํ•˜๋ฉฐ, ๋ณต์žกํ•˜๊ฑฐ๋‚˜ ํ•ด์„ํ•˜๊ธฐ ์–ด๋ ค์šด ๋ฌธ์žฅ์ด ์—†์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
์œ ์ฐฝ์„ฑ: ์ผ๊ด€๋œ ๋ฌธ๊ตฌ์™€ ๋ช…ํ™•ํ•œ ํ‘œํ˜„์œผ๋กœ ๋‹ต๋ณ€์ด ์ธ๊ฐ„์˜ ์–ธ์–ด ๊ทœ์น™์„ ์–ผ๋งˆ๋‚˜ ์ž˜ ์ค€์ˆ˜ํ•˜๋Š”์ง€๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ ๋‹ต๋ณ€์€ ๋ณต์žกํ•˜๊ฑฐ๋‚˜ ๊ตฌ๋ฌธ ๋ถ„์„ํ•˜๊ธฐ ์–ด๋ ค์šด ๋ฌธ์žฅ ์—†์ด ์ดํ•ดํ•˜๊ธฐ ์‰ฌ์›Œ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
**Result 2** (Figure 11 in Appendix A.3).: In the same 1000-exposure setting, peak capacity ratios for GPT2 trained on \(\mathsf{bioS}^{\mathsf{simple}}\) and \(\mathsf{bioR}\) are also approximately 2, albeit slightly lower.
**Result 2** (Appendix A.3์˜ Figure 11): ๋™์ผํ•œ 1000ํšŒ ๋…ธ์ถœ ์„ค์ •์—์„œ, \(\mathsf{bioS}^{\mathsf{simple}}\)์™€ \(\mathsf{bioR}\)๋กœ ํ›ˆ๋ จ๋œ GPT2์˜ ํ”ผํฌ ์šฉ๋Ÿ‰ ๋น„์œจ๋„ ์•ฝ 2๋กœ ๋‚˜ํƒ€๋‚˜์ง€๋งŒ, ์•ฝ๊ฐ„ ๋‚ฎ์Šต๋‹ˆ๋‹ค.
**๊ฒฐ๊ณผ 2** (๋ถ€๋ก A.3์˜ ๊ทธ๋ฆผ 11).: ๋™์ผํ•œ 1000 ๋…ธ์ถœ ์„ค์ •์—์„œ, \(\mathsf{bioS}^{\mathsf{simple}}\) ๋ฐ \(\mathsf{bioR}\)์— ๋Œ€ํ•ด ํ•™์Šต๋œ GPT2์˜ ํ”ผํฌ ์šฉ๋Ÿ‰ ๋น„์œจ๋„ ์•ฝ๊ฐ„ ๋‚ฎ์ง€๋งŒ ์•ฝ 2์ž…๋‹ˆ๋‹ค.
Removing the logit bias assumption.All our attacks require the ability to pass a logit bias. Model providers including Google and OpenAI provided this capability when we began the writing of this paper, but this could change.
๋กœ๊ทธ์ž‡ ๋ฐ”์ด์–ด์Šค ๊ฐ€์ • ์ œ๊ฑฐ. ์šฐ๋ฆฌ์˜ ๋ชจ๋“  ๊ณต๊ฒฉ์€ ๋กœ๊ทธ์ž‡ ๋ฐ”์ด์–ด์Šค๋ฅผ ์ „๋‹ฌํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ํ•„์š”๋กœ ํ•ฉ๋‹ˆ๋‹ค. Google๊ณผ OpenAI๋ฅผ ํฌํ•จํ•œ ๋ชจ๋ธ ์ œ๊ณต์ž๋“ค์€ ์šฐ๋ฆฌ๊ฐ€ ์ด ๋…ผ๋ฌธ์„ ์ž‘์„ฑํ•˜๊ธฐ ์‹œ์ž‘ํ–ˆ์„ ๋•Œ ์ด ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ–ˆ์ง€๋งŒ, ์ด๋Š” ๋ณ€๊ฒฝ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋กœ์ง“ ๋ฐ”์ด์–ด์Šค ๊ฐ€์ •์„ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๊ณต๊ฒฉ์—๋Š” ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค๋ฅผ ํ†ต๊ณผํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์„ ์ž‘์„ฑํ•  ๋•Œ Google๊ณผ OpenAI๋ฅผ ํฌํ•จํ•œ ๋ชจ๋ธ ์ œ๊ณต์—…์ฒด๋Š” ์ด ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ–ˆ์ง€๋งŒ, ์ด๋Š” ๋ณ€๊ฒฝ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
that larger data volumes and bigger model sizes enhance performance. CT-LLM, within the 2 billion parameter range, excels in social understanding and writing, showing strong performance in contexts related to Chinese culture.
๋” ํฐ ๋ฐ์ดํ„ฐ ๋ณผ๋ฅจ๊ณผ ๋” ํฐ ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. 20์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ ๋ฒ”์œ„ ๋‚ด์— ์žˆ๋Š” CT-LLM์€ ์‚ฌํšŒ์  ์ดํ•ด์™€ ๊ธ€์“ฐ๊ธฐ์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋ฉฐ, ์ค‘๊ตญ ๋ฌธํ™”์™€ ๊ด€๋ จ๋œ ๋งฅ๋ฝ์—์„œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๋” ํฐ ๋ฐ์ดํ„ฐ ๋ณผ๋ฅจ๊ณผ ๋” ํฐ ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. CT-LLM์€ 20์–ต ํŒŒ๋ผ๋ฏธํ„ฐ ๋ฒ”์œ„ ๋‚ด์—์„œ ์‚ฌํšŒ์  ์ดํ•ด์™€ ๊ธ€์“ฐ๊ธฐ ๋ถ„์•ผ์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ ์ค‘๊ตญ ๋ฌธํ™”์™€ ๊ด€๋ จ๋œ ๋งฅ๋ฝ์—์„œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
In this section, we briefly review the recent progress on the applications of LLMs in two aspects, namely the impact to research community and representative domains. Figure 18 shows a content organization of this section49.
์ด ์„น์…˜์—์„œ๋Š” ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ๊ณผ ๋Œ€ํ‘œ์ ์ธ ๋„๋ฉ”์ธ์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ธก๋ฉด์—์„œ LLMs์˜ ์ตœ๊ทผ ์‘์šฉ ์ง„์ „์— ๋Œ€ํ•ด ๊ฐ„๋žตํžˆ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. Figure 18์€ ์ด ์„น์…˜์˜ ๋‚ด์šฉ ๊ตฌ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค49.
์ด ์„น์…˜์—์„œ๋Š” ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์™€ ๋Œ€ํ‘œ์ ์ธ ๋„๋ฉ”์ธ์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ธก๋ฉด์—์„œ LLM์˜ ์‘์šฉ์— ๋Œ€ํ•œ ์ตœ๊ทผ์˜ ์ง„์ „์„ ๊ฐ„๋žตํ•˜๊ฒŒ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 18์€ ์ด ์„น์…˜49์˜ ๋‚ด์šฉ ๊ตฌ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
Footnote 14: Within 1000 exposures, itโ€™s likely that the same individual will have 1000 different biography paragraphs detailing the same knowledge (see Section 2.2). Therefore, 1000 exposures can occur within a single pass.
Footnote 14: 1000๋ฒˆ์˜ ๋…ธ์ถœ ๋™์•ˆ, ๋™์ผํ•œ ๊ฐœ์ธ์ด ๋™์ผํ•œ ์ง€์‹์„ ์ƒ์„ธํžˆ ์„ค๋ช…ํ•˜๋Š” 1000๊ฐœ์˜ ๋‹ค๋ฅธ ์ „๊ธฐ ๋‹จ๋ฝ์„ ๊ฐ€์งˆ ๊ฐ€๋Šฅ์„ฑ์ด ํฝ๋‹ˆ๋‹ค(์„น์…˜ 2.2 ์ฐธ์กฐ). ๋”ฐ๋ผ์„œ, 1000๋ฒˆ์˜ ๋…ธ์ถœ์€ ํ•œ ๋ฒˆ์˜ ํ†ต๊ณผ ๋‚ด์—์„œ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฐ์ฃผ 14: 1000ํšŒ ๋…ธ์ถœ ์ด๋‚ด์— ๋™์ผํ•œ ๊ฐœ์ธ์ด ๋™์ผํ•œ ์ง€์‹์„ ์ž์„ธํžˆ ์„ค๋ช…ํ•˜๋Š” 1000๊ฐœ์˜ ๋‹ค๋ฅธ ์•ฝ๋ ฅ ๋‹จ๋ฝ์„ ๋ณด์œ ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค(2.2์ ˆ ์ฐธ์กฐ). ๋”ฐ๋ผ์„œ 1000ํšŒ ๋…ธ์ถœ์€ ํ•œ ๋ฒˆ์˜ ํŒจ์Šค ๋‚ด์—์„œ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The choice of architecture and pre-training tasks may incur different inductive biases for LLMs, which would lead to different model capacities. In this part, we discuss one open issue about the architecture choice for LLMs.
์•„ํ‚คํ…์ฒ˜์™€ ์‚ฌ์ „ ํ•™์Šต ์ž‘์—…์˜ ์„ ํƒ์€ LLM์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ๊ท€๋‚ฉ์  ํŽธํ–ฅ์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ ์šฉ๋Ÿ‰์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” LLM์˜ ์•„ํ‚คํ…์ฒ˜ ์„ ํƒ์— ๊ด€ํ•œ ํ•˜๋‚˜์˜ ๋ฏธํ•ด๊ฒฐ ๋ฌธ์ œ๋ฅผ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค.
์•„ํ‚คํ…์ฒ˜์™€ ์‚ฌ์ „ ํ•™์Šต ์ž‘์—…์˜ ์„ ํƒ์€ LLM์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ๊ท€๋‚ฉ์  ํŽธํ–ฅ์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ ์šฉ๋Ÿ‰์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” LLM์˜ ์•„ํ‚คํ…์ฒ˜ ์„ ํƒ์— ๋Œ€ํ•œ ํ•œ ๊ฐ€์ง€ ๊ณต๊ฐœ๋œ ๋ฌธ์ œ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค.
We set the input length to 32K for fine-tuning and increase to 500K for evaluating. We use a generation temperature of 0.5 and \(top_{p}=0.95\) and set the number of decoding steps to 1024 to generate a summary of each book.
์ž…๋ ฅ ๊ธธ์ด๋Š” ๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•ด 32K๋กœ ์„ค์ •ํ•˜๊ณ  ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด 500K๋กœ ์ฆ๊ฐ€์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ƒ์„ฑ ์˜จ๋„๋Š” 0.5๋กœ ์„ค์ •ํ•˜๊ณ  \(top_{p}=0.95\)๋กœ ์„ค์ •ํ–ˆ์œผ๋ฉฐ, ๊ฐ ์ฑ…์˜ ์š”์•ฝ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๋””์ฝ”๋”ฉ ๋‹จ๊ณ„๋ฅผ 1024๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•ด ์ž…๋ ฅ ๊ธธ์ด๋ฅผ 32K๋กœ ์„ค์ •ํ•˜๊ณ  ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด 500K๋กœ ๋Š˜๋ฆฝ๋‹ˆ๋‹ค. ์ƒ์„ฑ ์˜จ๋„ 0.5์™€ \(top_{p}=0.95\)๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ๋””์ฝ”๋”ฉ ๋‹จ๊ณ„ ์ˆ˜๋ฅผ 1024๋กœ ์„ค์ •ํ•˜์—ฌ ๊ฐ ์ฑ…์˜ ์š”์•ฝ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
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.
์šฐ๋ฆฌ๋Š” ๋‘ ๊ฐ€์ง€ ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๋„ค ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹ ๊ฐ๊ฐ์„ ์ค‘๋ณต ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค. ํ…์ŠคํŠธ๊ฐ€ ์—ฌ๋Ÿฌ ๋ฐ์ดํ„ฐ ๋ถ„ํ• ์— ๊ฑธ์ณ ์ค‘๋ณต๋œ ๊ฒฝ์šฐ, ํ…Œ์ŠคํŠธ ๋˜๋Š” ๊ฒ€์ฆ ์„ธํŠธ์— ๋ณต์‚ฌ๋ณธ์„ ์œ ์ง€ํ•˜๊ณ  ํ•™์Šต ์„ธํŠธ์—์„œ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์„ ์šฐ์„ ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค.
๋‘ ๊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.
The parameterized knowledge within models is limited and outdated. Thus, factual mistakes often occur when facing knowledge-intensive tasks. Introducing external knowledge can mitigate this phenomenon, as shown in Figure 5.
๋ชจ๋ธ ๋‚ด์˜ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ์ง€์‹์€ ์ œํ•œ์ ์ด๊ณ  ์˜ค๋ž˜๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ง€์‹ ์ง‘์•ฝ์ ์ธ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•  ๋•Œ ์‚ฌ์‹ค์  ์˜ค๋ฅ˜๊ฐ€ ์ž์ฃผ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. Figure 5์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ์™ธ๋ถ€ ์ง€์‹์„ ๋„์ž…ํ•˜๋ฉด ์ด๋Ÿฌํ•œ ํ˜„์ƒ์„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ชจ๋ธ ๋‚ด์˜ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ์ง€์‹์€ ์ œํ•œ์ ์ด๊ณ  ์˜ค๋ž˜๋œ ์ง€์‹์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ง€์‹ ์ง‘์•ฝ์ ์ธ ์ž‘์—…์— ์ง๋ฉดํ•  ๋•Œ ์‚ฌ์‹ค ์˜ค๋ฅ˜๊ฐ€ ์ž์ฃผ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 5์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ์™ธ๋ถ€ ์ง€์‹์„ ๋„์ž…ํ•˜๋ฉด ์ด๋Ÿฌํ•œ ํ˜„์ƒ์„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 7: **Mixing Llama Generated Data with Original Data** Based on Figure 4 left, we further mix the generated data with original data, with ratio 98 to 2. Adding original data significantly mitigates the model collapse.
Figure 7: **Llama๊ฐ€ ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ์™€ ์›๋ณธ ๋ฐ์ดํ„ฐ์˜ ํ˜ผํ•ฉ** Figure 4์˜ ์™ผ์ชฝ์„ ๊ธฐ๋ฐ˜์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์›๋ณธ ๋ฐ์ดํ„ฐ์™€ 98 ๋Œ€ 2์˜ ๋น„์œจ๋กœ ์ถ”๊ฐ€๋กœ ํ˜ผํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์›๋ณธ ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•˜๋ฉด ๋ชจ๋ธ ๋ถ•๊ดด๋ฅผ ์ƒ๋‹นํžˆ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 7: **๋ผ๋งˆ ์ƒ์„ฑ ๋ฐ์ดํ„ฐ์™€ ์›๋ณธ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ** ๊ทธ๋ฆผ 4 ์™ผ์ชฝ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์›๋ณธ ๋ฐ์ดํ„ฐ์™€ 98 ๋Œ€ 2์˜ ๋น„์œจ๋กœ ์ถ”๊ฐ€๋กœ ํ˜ผํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์›๋ณธ ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•˜๋ฉด ๋ชจ๋ธ ๋ถ•๊ดด๋ฅผ ์ƒ๋‹นํžˆ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
As mentioned in Section 2, we explore the methods from the literature that have addressed the limitations of CL methods and thus are applicable to CKL. We also categorize these methods into the three main categories of CL.
2์žฅ์—์„œ ์–ธ๊ธ‰ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ์šฐ๋ฆฌ๋Š” CL ๋ฐฉ๋ฒ•์˜ ํ•œ๊ณ„๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ  CKL์— ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฌธํ—Œ์˜ ๋ฐฉ๋ฒ•๋“ค์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•๋“ค์„ CL์˜ ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๋ฒ”์ฃผ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค.
2์ ˆ์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, CL ๋ฐฉ๋ฒ•์˜ ํ•œ๊ณ„๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ  ๋”ฐ๋ผ์„œ CKL์— ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฌธํ—Œ์˜ ๋ฐฉ๋ฒ•์„ ํƒ์ƒ‰ํ•œ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์„ CL์˜ ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๋ฒ”์ฃผ๋กœ ๋ถ„๋ฅ˜ํ•œ๋‹ค.
We use "Answer the following questions in JUST a few words or one sentence." for Topi-OCQA, since this dataset has both short and long answers based on the questions, and the long answers are generally within one sentence.
Topi-OCQA์— ๋Œ€ํ•ด์„œ๋Š” "๋‹ค์Œ ์งˆ๋ฌธ์— ๋‹จ ๋ช‡ ๋‹จ์–ด ๋˜๋Š” ํ•œ ๋ฌธ์žฅ์œผ๋กœ ๋‹ตํ•˜์‹ญ์‹œ์˜ค."๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ์…‹์€ ์งˆ๋ฌธ์— ๋”ฐ๋ผ ์งง์€ ๋‹ต๋ณ€๊ณผ ๊ธด ๋‹ต๋ณ€์„ ๋ชจ๋‘ ํฌํ•จํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ๊ธด ๋‹ต๋ณ€๋„ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•œ ๋ฌธ์žฅ ๋‚ด์— ์žˆ์Šต๋‹ˆ๋‹ค.
์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์งˆ๋ฌธ์— ๋”ฐ๋ฅธ ์งง์€ ๋‹ต๋ณ€๊ณผ ๊ธด ๋‹ต๋ณ€์ด ๋ชจ๋‘ ์žˆ์œผ๋ฉฐ, ๊ธด ๋‹ต๋ณ€์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•œ ๋ฌธ์žฅ ์ด๋‚ด์ด๋ฏ€๋กœ Topi-OCQA์—๋Š” "๋‹ค์Œ ์งˆ๋ฌธ์— ๋ช‡ ๋‹จ์–ด ๋˜๋Š” ํ•œ ๋ฌธ์žฅ์œผ๋กœ๋งŒ ๋‹ตํ•˜์„ธ์š”."๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
Broader Impact.Similar to the other LLMs, our _WizardMath_ could also generate unethical, harmful, or misleading information sometimes. Therefore, future research to address the ethical and societal implications is needed.
๋” ๋„“์€ ์˜ํ–ฅ. ๋‹ค๋ฅธ LLM๋“ค๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ์šฐ๋ฆฌ์˜ _WizardMath_๋„ ๋•Œ๋•Œ๋กœ ๋น„์œค๋ฆฌ์ ์ด๊ฑฐ๋‚˜ ํ•ด๋กญ๊ฑฐ๋‚˜ ์˜คํ•ด์˜ ์†Œ์ง€๊ฐ€ ์žˆ๋Š” ์ •๋ณด๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์œค๋ฆฌ์  ๋ฐ ์‚ฌํšŒ์  ํ•จ์˜๋ฅผ ๋‹ค๋ฃจ๊ธฐ ์œ„ํ•œ ํ–ฅํ›„ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
๋” ๋„“์€ ์˜ํ–ฅ.๋‹ค๋ฅธ LLM๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ์šฐ๋ฆฌ์˜ _WizardMath_๋„ ๋•Œ๋•Œ๋กœ ๋น„์œค๋ฆฌ์ , ์œ ํ•ดํ•˜๊ฑฐ๋‚˜ ์˜คํ•ด์˜ ์†Œ์ง€๊ฐ€ ์žˆ๋Š” ์ •๋ณด๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์œค๋ฆฌ์ , ์‚ฌํšŒ์  ์˜ํ–ฅ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ํ–ฅํ›„ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
Automatic raters of faithfulness:We implement our automatic raters by prompting an LLM in a zero-shot manner to verify a single claim given evidence from the book (Table 13), where the evidence can be one of the following:
์‹ ๋ขฐ์„ฑ์˜ ์ž๋™ ํ‰๊ฐ€์ž: ์šฐ๋ฆฌ๋Š” ์ฑ…์—์„œ ์ œ๊ณต๋œ ์ฆ๊ฑฐ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๋‹จ์ผ ์ฃผ์žฅ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ์ œ๋กœ์ƒท ๋ฐฉ์‹์œผ๋กœ ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ์ž๋™ ํ‰๊ฐ€์ž๋ฅผ ๊ตฌํ˜„ํ•ฉ๋‹ˆ๋‹ค (ํ‘œ 13). ์—ฌ๊ธฐ์„œ ์ฆ๊ฑฐ๋Š” ๋‹ค์Œ ์ค‘ ํ•˜๋‚˜์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
์ถฉ์‹ค๋„์˜ ์ž๋™ ํ‰๊ฐ€์ž: ์ฑ…์˜ ์ฆ๊ฑฐ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ ํ•˜๋‚˜์˜ ์ฃผ์žฅ์„ ๊ฒ€์ฆํ•˜๋„๋ก LLM์„ ์ œ๋กœ ์ƒท ๋ฐฉ์‹์œผ๋กœ ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ์ž๋™ ํ‰๊ฐ€์ž๋ฅผ ๊ตฌํ˜„ํ•ฉ๋‹ˆ๋‹ค(ํ‘œ 13). ์—ฌ๊ธฐ์„œ ์ฆ๊ฑฐ๋Š” ๋‹ค์Œ ์ค‘ ํ•˜๋‚˜์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
1. Violent crimes 2. Non-violent crimes 3. Sex-related crimes 4. Child sexual exploitation 5. Indiscriminate weapons, Chemical, Biological, Radiological, Nuclear, and high yield Explosives (CBRNE) 6. Suicide & self-harm 7.
1. ํญ๋ ฅ ๋ฒ”์ฃ„ 2. ๋น„ํญ๋ ฅ ๋ฒ”์ฃ„ 3. ์„ฑ ๊ด€๋ จ ๋ฒ”์ฃ„ 4. ์•„๋™ ์„ฑ ์ฐฉ์ทจ 5. ๋ฌด์ฐจ๋ณ„ ๋ฌด๊ธฐ, ํ™”ํ•™, ์ƒ๋ฌผํ•™, ๋ฐฉ์‚ฌ์„ , ํ•ต ๋ฐ ๊ณ ์„ฑ๋Šฅ ํญ๋ฐœ๋ฌผ (CBRNE) 6. ์ž์‚ด ๋ฐ ์žํ•ด 7.
1. ํญ๋ ฅ ๋ฒ”์ฃ„2. ๋น„ํญ๋ ฅ ๋ฒ”์ฃ„3. ์„ฑ ๊ด€๋ จ ๋ฒ”์ฃ„4. ์•„๋™ ์„ฑ์ฐฉ์ทจ5. ๋ฌด์ฐจ๋ณ„ ๋ฌด๊ธฐ, ํ™”ํ•™, ์ƒ๋ฌผํ•™, ๋ฐฉ์‚ฌ์„ , ํ•ต ๋ฐ ๊ณ ์ถœ๋ ฅ ํญ๋ฐœ๋ฌผ(CBRNE)6. ์ž์‚ด ๋ฐ ์žํ•ด7.
where \(d\) is the embedding dimension, and serves a s proxy for model capacity; the exponent \(c_{q}\) depends both on \(\beta\) and the particular algorithm \(q\) used to update the memories in the model during training.
์—ฌ๊ธฐ์„œ \(d\)๋Š” ์ž„๋ฒ ๋”ฉ ์ฐจ์›์œผ๋กœ, ๋ชจ๋ธ ์šฉ๋Ÿ‰์˜ ๋Œ€๋ฆฌ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค; ์ง€์ˆ˜ \(c_{q}\)๋Š” \(\beta\)์™€ ํ›ˆ๋ จ ์ค‘ ๋ชจ๋ธ์˜ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ํŠน์ • ์•Œ๊ณ ๋ฆฌ์ฆ˜ \(q\)์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(d\)๋Š” ์ž„๋ฒ ๋”ฉ ์ฐจ์›์ด๋ฉฐ ๋ชจ๋ธ ์šฉ๋Ÿ‰์˜ ๋Œ€๋ฆฌ ์—ญํ• ์„ ํ•˜๋ฉฐ, ์ง€์ˆ˜ \(c_{q}\)๋Š” \(\beta\)์™€ ํ›ˆ๋ จ ์ค‘์— ๋ชจ๋ธ์˜ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ํŠน์ • ์•Œ๊ณ ๋ฆฌ์ฆ˜ \(q\)์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค.
This section is devoted to an analysis of the limitations of existing RALM architectures and a description of potential future developments. Figure 7 summarizes the limitations of existing RALMs and our proposed solutions.
์ด ์„น์…˜์€ ๊ธฐ์กด RALM ์•„ํ‚คํ…์ฒ˜์˜ ํ•œ๊ณ„์— ๋Œ€ํ•œ ๋ถ„์„๊ณผ ์ž ์žฌ์ ์ธ ๋ฏธ๋ž˜ ๊ฐœ๋ฐœ์— ๋Œ€ํ•œ ์„ค๋ช…์— ์ „๋…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Figure 7์€ ๊ธฐ์กด RALM์˜ ํ•œ๊ณ„์™€ ์šฐ๋ฆฌ๊ฐ€ ์ œ์•ˆํ•œ ํ•ด๊ฒฐ์ฑ…์„ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๊ธฐ์กด RALM ์•„ํ‚คํ…์ฒ˜์˜ ํ•œ๊ณ„์— ๋Œ€ํ•œ ๋ถ„์„๊ณผ ํ–ฅํ›„ ์ž ์žฌ์  ๊ฐœ๋ฐœ์— ๋Œ€ํ•œ ์„ค๋ช…์— ์ „๋…ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 7์€ ๊ธฐ์กด RALM์˜ ํ•œ๊ณ„์™€ ์ œ์•ˆ๋œ ์†”๋ฃจ์…˜์„ ์š”์•ฝํ•œ ๊ฒƒ์ด๋‹ค.
Harmlessness: Assesses whether the answer avoids unethical or harmful content, adhering to ethical norms. A high-quality answer should uphold ethical principles, avoiding the propagation of harmful or immoral information.
๋ฌดํ•ด์„ฑ: ๋‹ต๋ณ€์ด ๋น„์œค๋ฆฌ์ ์ด๊ฑฐ๋‚˜ ํ•ด๋กœ์šด ๋‚ด์šฉ์„ ํ”ผํ•˜๊ณ  ์œค๋ฆฌ์  ๊ทœ๋ฒ”์„ ์ค€์ˆ˜ํ•˜๋Š”์ง€ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ์˜ ๋‹ต๋ณ€์€ ์œค๋ฆฌ์  ์›์น™์„ ์ค€์ˆ˜ํ•˜๋ฉฐ, ํ•ด๋กญ๊ฑฐ๋‚˜ ๋น„๋„๋•์ ์ธ ์ •๋ณด์˜ ์ „ํŒŒ๋ฅผ ํ”ผํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๋ฌดํ•ด์„ฑ: ๋‹ต๋ณ€์ด ๋น„์œค๋ฆฌ์ ์ด๊ฑฐ๋‚˜ ์œ ํ•ดํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ํ”ผํ•˜๊ณ  ์œค๋ฆฌ์  ๊ทœ๋ฒ”์„ ์ค€์ˆ˜ํ•˜๋Š”์ง€ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ ๋‹ต๋ณ€์€ ์œ ํ•ดํ•˜๊ฑฐ๋‚˜ ๋ถ€๋„๋•ํ•œ ์ •๋ณด์˜ ํ™•์‚ฐ์„ ํ”ผํ•˜๋ฉด์„œ ์œค๋ฆฌ์  ์›์น™์„ ์ง€์ผœ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
**When CoT Promping Works For LLMs?** Since CoT reasoning is an emergent ability [31], it only has a positive effect on sufficiently large models (typically containing 10B or more parameters [33]) but not on small models.
**LLM์—์„œ CoT ํ”„๋กฌํ”„ํŒ…์ด ์–ธ์ œ ํšจ๊ณผ๊ฐ€ ์žˆ์„๊นŒ์š”?** CoT ์ถ”๋ก ์€ ์ถœํ˜„ ๋Šฅ๋ ฅ [31]์ด๊ธฐ ๋•Œ๋ฌธ์—, ์ถฉ๋ถ„ํžˆ ํฐ ๋ชจ๋ธ(์ผ๋ฐ˜์ ์œผ๋กœ 100์–ต ๊ฐœ ์ด์ƒ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํฌํ•จํ•˜๋Š” ๋ชจ๋ธ [33])์—์„œ๋งŒ ๊ธ์ •์ ์ธ ํšจ๊ณผ๋ฅผ ๊ฐ€์ง€๋ฉฐ, ์ž‘์€ ๋ชจ๋ธ์—์„œ๋Š” ๊ทธ๋ ‡์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
**CoT ํ”„๋กฌํ”„ํŠธ๊ฐ€ LLM์— ํšจ๊ณผ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ๋Š” ์–ธ์ œ์ธ๊ฐ€์š”?** CoT ์ถ”๋ก ์€ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์ด๊ธฐ ๋•Œ๋ฌธ์— [31] ์ถฉ๋ถ„ํžˆ ํฐ ๋ชจ๋ธ(์ผ๋ฐ˜์ ์œผ๋กœ 10B ์ด์ƒ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํฌํ•จํ•˜๋Š” ๋ชจ๋ธ [33])์—๋งŒ ๊ธ์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€๋งŒ ์ž‘์€ ๋ชจ๋ธ์—๋Š” ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
Hungarian National High School ExamThis evaluation benchmark is first introduced by Grok-1 (xAI, 2023), which is designed for evaluating the out-of-domain capability of math models. It consists of 33 challenging problems.
ํ—๊ฐ€๋ฆฌ ๊ตญ๊ฐ€ ๊ณ ๋“ฑํ•™๊ต ์‹œํ—˜ ์ด ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋Š” ์ˆ˜ํ•™ ๋ชจ๋ธ์˜ ๋„๋ฉ”์ธ ์™ธ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด Grok-1 (xAI, 2023)์— ์˜ํ•ด ์ฒ˜์Œ ๋„์ž…๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” 33๊ฐœ์˜ ๋„์ „์ ์ธ ๋ฌธ์ œ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
ํ—๊ฐ€๋ฆฌ ๊ตญ๊ฐ€ ๊ณ ๋“ฑํ•™๊ต ์‹œํ—˜์ด ํ‰๊ฐ€ ๊ธฐ์ค€์€ ๋จผ์ € Grok-1(xAI, 2023)์— ์˜ํ•ด ๋„์ž…๋˜์—ˆ์œผ๋ฉฐ, ์ˆ˜ํ•™ ๋ชจ๋ธ์˜ ๋„๋ฉ”์ธ ์™ธ ๊ธฐ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ํ‰๊ฐ€๋Š” 33๊ฐœ์˜ ๋‚œ์ด๋„ ๋ฌธ์ œ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
**Pros and Cons of Different Evaluation Approaches**. In the above, we have discussed different evaluation approaches to assess the abilities of LLMs. Next, we simply analyze the pros and cons of each evaluation approach.
**๋‹ค์–‘ํ•œ ํ‰๊ฐ€ ์ ‘๊ทผ๋ฒ•์˜ ์žฅ๋‹จ์ **. ์œ„์—์„œ ์šฐ๋ฆฌ๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ํ‰๊ฐ€ ์ ‘๊ทผ๋ฒ•์— ๋Œ€ํ•ด ๋…ผ์˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, ๊ฐ ํ‰๊ฐ€ ์ ‘๊ทผ๋ฒ•์˜ ์žฅ๋‹จ์ ์„ ๊ฐ„๋‹จํžˆ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค.
**๋‹ค์–‘ํ•œ ํ‰๊ฐ€ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์žฅ๋‹จ์ **. ์œ„์—์„œ๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ํ‰๊ฐ€ ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•ด ๋…ผ์˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ๊ฐ ํ‰๊ฐ€ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์žฅ๋‹จ์ ์„ ๊ฐ„๋‹จํžˆ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค.
To facilitate further research, we release our dataset: the largest and most recent forecasting dataset compiled from 5 real-world forecasting competitions. We discuss a few opportunities to improve these systems further.
์ถ”๊ฐ€ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” 5๊ฐœ์˜ ์‹ค์ œ ์˜ˆ์ธก ๋Œ€ํšŒ์—์„œ ์ˆ˜์ง‘๋œ ๊ฐ€์žฅ ํฌ๊ณ  ์ตœ์‹ ์˜ ์˜ˆ์ธก ๋ฐ์ดํ„ฐ์…‹์„ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹œ์Šคํ…œ์„ ๋”์šฑ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๋ช‡ ๊ฐ€์ง€ ๊ธฐํšŒ๋ฅผ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค.
์ถ”๊ฐ€ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด ์‹ค์ œ ์˜ˆ์ธก ๊ฒฝ์ง„๋Œ€ํšŒ 5๊ฐœ์—์„œ ์ˆ˜์ง‘ํ•œ ๊ฐ€์žฅ ํฌ๊ณ  ์ตœ์‹ ์˜ ์˜ˆ์ธก ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹œ์Šคํ…œ์„ ๋”์šฑ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๋ช‡ ๊ฐ€์ง€ ๊ธฐํšŒ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค.
Footnote 2: Formally, the LLMโ€™s likelihood of name \(n\) when prompted with the description \(d\), \(P_{\text{LLM}}(n|d)\), is not higher than the likelihood of a random name \(n_{r}\), namely \(P_{\text{LLM}}(n_{r}|d)\).
Footnote 2: ๊ณต์‹์ ์œผ๋กœ, ์„ค๋ช… \(d\)๋กœ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฐ›์•˜์„ ๋•Œ ์ด๋ฆ„ \(n\)์— ๋Œ€ํ•œ LLM์˜ ํ™•๋ฅ  \(P_{\text{LLM}}(n|d)\)๋Š” ์ž„์˜์˜ ์ด๋ฆ„ \(n_{r}\)์— ๋Œ€ํ•œ ํ™•๋ฅ  \(P_{\text{LLM}}(n_{r}|d)\)๋ณด๋‹ค ๋†’์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
๊ฐ์ฃผ 2: ๊ณต์‹์ ์œผ๋กœ, ์„ค๋ช… \(d\)๊ฐ€ ํ”„๋กฌํ”„ํŠธ๋  ๋•Œ LLM์˜ ์ด๋ฆ„ \(n\)์— ๋Œ€ํ•œ ๊ฐ€๋Šฅ์„ฑ \(P_{\text{LLM}}(n|d)\)์€ ์ž„์˜์˜ ์ด๋ฆ„ \(n_{r}\), ์ฆ‰ \(P_{\text{LLM}}(n_{r}|d)\)๋ณด๋‹ค ๋†’์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
PersonA: While using Vascepa, you may need frequent blood tests, and it's important to follow your doctor's instructions closely. It's important to talk to your doctor about any allergies or existing conditions you have.
PersonA: Vascepa๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋™์•ˆ ์ž์ฃผ ํ˜ˆ์•ก ๊ฒ€์‚ฌ๋ฅผ ๋ฐ›์•„์•ผ ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์˜์‚ฌ์˜ ์ง€์‹œ๋ฅผ ์ฒ ์ €ํžˆ ๋”ฐ๋ฅด๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์•Œ๋ ˆ๋ฅด๊ธฐ๋‚˜ ๊ธฐ์กด ์งˆํ™˜์— ๋Œ€ํ•ด ์˜์‚ฌ์™€ ์ƒ์˜ํ•˜๋Š” ๊ฒƒ๋„ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ๋ฐ”์„ธํŒŒ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋™์•ˆ ๋นˆ๋ฒˆํ•œ ํ˜ˆ์•ก ๊ฒ€์‚ฌ๊ฐ€ ํ•„์š”ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์˜์‚ฌ์˜ ์ง€์‹œ๋ฅผ ์ž˜ ๋”ฐ๋ฅด๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์•Œ๋ ˆ๋ฅด๊ธฐ๋‚˜ ๊ธฐ์กด ์งˆํ™˜์ด ์žˆ๋Š” ๊ฒฝ์šฐ ์˜์‚ฌ์™€ ์ƒ์˜ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
**Correctness of** [IsRel] -- [IsNew] -- Both [IsRel] and [IsNew] are correct, as the Reference [1] provides detailed information of the book including the genre, and the model output is fully supported by the reference.
[IsRel]์˜ ์ •ํ™•์„ฑ -- [IsNew]์˜ ์ •ํ™•์„ฑ -- [IsRel]๊ณผ [IsNew] ๋ชจ๋‘ ์ •ํ™•ํ•ฉ๋‹ˆ๋‹ค. Reference [1]์€ ์žฅ๋ฅด๋ฅผ ํฌํ•จํ•œ ์ฑ…์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ๋ชจ๋ธ ์ถœ๋ ฅ์€ ์ฐธ์กฐ์— ์˜ํ•ด ์™„์ „ํžˆ ๋’ท๋ฐ›์นจ๋ฉ๋‹ˆ๋‹ค.
**์ •ํ™•์„ฑ** [IsRel] -- [IsNew] -- ์ฐธ์กฐ [1]์€ ์žฅ๋ฅด๋ฅผ ํฌํ•จํ•œ ์ฑ…์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๊ณ  ๋ชจ๋ธ ์ถœ๋ ฅ์€ ์ฐธ์กฐ์— ์˜ํ•ด ์™„์ „ํžˆ ์ง€์›๋˜๋ฏ€๋กœ [IsRel]๊ณผ [IsNew]๋Š” ๋ชจ๋‘ ์˜ฌ๋ฐ”๋ฆ…๋‹ˆ๋‹ค.
**Result 12** (Figure 8(g)+8(h)).: When 7/8 of training tokens are from junk (i.e., bioS(\(N^{\prime}\)) for \(N^{\prime}=100M\)), adding a special token at the start of every useful data greatly improves capacity ratio:
**Result 12** (Figure 8(g)+8(h)).: ํ›ˆ๋ จ ํ† ํฐ์˜ 7/8์ด ์“ธ๋ชจ์—†๋Š” ๋ฐ์ดํ„ฐ(์ฆ‰, \(N^{\prime}=100M\)์ธ bioS(\(N^{\prime}\)))์ผ ๋•Œ, ์œ ์šฉํ•œ ๋ฐ์ดํ„ฐ์˜ ์‹œ์ž‘ ๋ถ€๋ถ„์— ํŠน์ˆ˜ ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜๋ฉด ์šฉ๋Ÿ‰ ๋น„์œจ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค:
**๊ฒฐ๊ณผ 12**(๊ทธ๋ฆผ 8(g)+8(h)).: ํ›ˆ๋ จ ํ† ํฐ์˜ 7/8์ด ์ •ํฌ์—์„œ ์œ ๋ž˜ํ•œ ๊ฒฝ์šฐ(์ฆ‰, \(N^{\ํ”„๋ผ์ž„}=100M\)์— ๋Œ€ํ•œ bioS(\(N^{\ํ”„๋ผ์ž„}\))์˜ ๊ฒฝ์šฐ), ์œ ์šฉํ•œ ๋ฐ์ดํ„ฐ์˜ ์‹œ์ž‘ ๋ถ€๋ถ„์— ํŠน์ˆ˜ ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜๋ฉด ์šฉ๋Ÿ‰ ๋น„์œจ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค:
We use "Answer the following question with a short span, or one sentence." for Topi-OCQA, since this dataset has both short and long answers based on the questions, and the long answers are generally within one sentence.
Topi-OCQA์— ๋Œ€ํ•ด์„œ๋Š” "๋‹ค์Œ ์งˆ๋ฌธ์— ์งง์€ ๋ฒ”์œ„ ๋˜๋Š” ํ•œ ๋ฌธ์žฅ์œผ๋กœ ๋‹ตํ•˜์‹ญ์‹œ์˜ค."๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ์…‹์€ ์งˆ๋ฌธ์— ๋”ฐ๋ผ ์งง์€ ๋‹ต๋ณ€๊ณผ ๊ธด ๋‹ต๋ณ€์„ ๋ชจ๋‘ ํฌํ•จํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ๊ธด ๋‹ต๋ณ€๋„ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•œ ๋ฌธ์žฅ ๋‚ด์— ์žˆ์Šต๋‹ˆ๋‹ค.
์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์งˆ๋ฌธ์— ๋”ฐ๋ผ ์งง์€ ๋‹ต๋ณ€๊ณผ ๊ธด ๋‹ต๋ณ€์ด ๋ชจ๋‘ ์žˆ์œผ๋ฉฐ ๊ธด ๋‹ต๋ณ€์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•œ ๋ฌธ์žฅ ์ด๋‚ด์ด๋ฏ€๋กœ Topi-OCQA์—๋Š” "์งง์€ ๋ฒ”์œ„ ๋˜๋Š” ํ•œ ๋ฌธ์žฅ์œผ๋กœ ๋‹ค์Œ ์งˆ๋ฌธ์— ๋‹ตํ•˜์„ธ์š”."๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
In Figure 3 (middle) we have seen an illustration of the translated scaling curves under n-fold synthesized data in the Hutter LLM. Figure 12 illustrates this phenomenon for the slightly more complex tailed bigram model.
Figure 3 (์ค‘๊ฐ„)์—์„œ๋Š” Hutter LLM์—์„œ n๋ฐฐ๋กœ ํ•ฉ์„ฑ๋œ ๋ฐ์ดํ„ฐ์— ๋”ฐ๋ฅธ ๋ณ€ํ™˜๋œ ์Šค์ผ€์ผ๋ง ๊ณก์„ ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. Figure 12๋Š” ์•ฝ๊ฐ„ ๋” ๋ณต์žกํ•œ ๊ผฌ๋ฆฌ ์ด์ค‘์–ด ๋ชจ๋ธ์— ๋Œ€ํ•ด ์ด ํ˜„์ƒ์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 3(๊ฐ€์šด๋ฐ)์—์„œ Hutter LLM์—์„œ n๋ฐฐ ํ•ฉ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์•„๋ž˜์—์„œ ๋ณ€ํ™˜๋œ ์Šค์ผ€์ผ๋ง ๊ณก์„ ์˜ ๊ทธ๋ฆผ์„ ๋ณด์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 12๋Š” ์•ฝ๊ฐ„ ๋” ๋ณต์žกํ•œ ๊ผฌ๋ฆฌ๊ฐ€ ์žˆ๋Š” ์ด์› ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ด๋Ÿฌํ•œ ํ˜„์ƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
We generate 200 problems for each number of required rules. Considering different premise orders and numbers of distracting rules, each problem includes 15 variants, resulting in a total of 27K problems in our benchmark.
์šฐ๋ฆฌ๋Š” ํ•„์š”ํ•œ ๊ทœ์น™ ์ˆ˜๋งˆ๋‹ค 200๊ฐœ์˜ ๋ฌธ์ œ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ „์ œ ์ˆœ์„œ์™€ ๋ฐฉํ•ด ๊ทœ์น™์˜ ์ˆ˜๋ฅผ ๊ณ ๋ คํ•˜์—ฌ, ๊ฐ ๋ฌธ์ œ๋Š” 15๊ฐœ์˜ ๋ณ€ํ˜•์„ ํฌํ•จํ•˜๋ฉฐ, ์šฐ๋ฆฌ ๋ฒค์น˜๋งˆํฌ์—๋Š” ์ด 27K๊ฐœ์˜ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
ํ•„์š”ํ•œ ๊ทœ์น™ ์ˆ˜์— ๋”ฐ๋ผ 200๊ฐœ์˜ ๋ฌธ์ œ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ „์ œ ์ˆœ์„œ์™€ ์‚ฐ๋งŒํ•œ ๊ทœ์น™์˜ ์ˆ˜๋ฅผ ๋‹ค๋ฅด๊ฒŒ ๊ณ ๋ คํ•˜์—ฌ ๊ฐ ๋ฌธ์ œ์—๋Š” 15๊ฐ€์ง€ ๋ณ€ํ˜•์ด ํฌํ•จ๋˜์–ด ๋ฒค์น˜๋งˆํฌ์—๋Š” ์ด 27K๊ฐœ์˜ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
their settings, at test time, the domain or documents can be different than the training time; whereas our paper studies a slightly opposite scenario where we only care about testing the LLM on the same set of documents.
์ด๋“ค์˜ ์„ค์ •์—์„œ๋Š” ํ…Œ์ŠคํŠธ ์‹œ์ ์— ๋„๋ฉ”์ธ์ด๋‚˜ ๋ฌธ์„œ๊ฐ€ ํ›ˆ๋ จ ์‹œ์ ๊ณผ ๋‹ค๋ฅผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์šฐ๋ฆฌ์˜ ๋…ผ๋ฌธ์€ ๋™์ผํ•œ ๋ฌธ์„œ ์ง‘ํ•ฉ์—์„œ LLM์„ ํ…Œ์ŠคํŠธํ•˜๋Š” ๊ฒƒ์—๋งŒ ๊ด€์‹ฌ์„ ๋‘๋Š” ์•ฝ๊ฐ„ ๋ฐ˜๋Œ€๋˜๋Š” ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
์„ค์ •์—์„œ ํ…Œ์ŠคํŠธ ์‹œ๊ฐ„์—๋Š” ๋„๋ฉ”์ธ ๋˜๋Š” ๋ฌธ์„œ๊ฐ€ ํ•™์Šต ์‹œ๊ฐ„๊ณผ ๋‹ค๋ฅผ ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ์ €ํฌ ๋…ผ๋ฌธ์—์„œ๋Š” ๋™์ผํ•œ ๋ฌธ์„œ ์„ธํŠธ์— ๋Œ€ํ•ด LLM์„ ํ…Œ์ŠคํŠธํ•˜๋Š” ๊ฒƒ์—๋งŒ ๊ด€์‹ฌ์„ ๊ฐ–๋Š” ์•ฝ๊ฐ„ ๋ฐ˜๋Œ€๋˜๋Š” ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
The system rates the relevance of each retrieved article at the scale of 1-6. Any article that receives a rating of \(\leq 3\) is discarded. We do not make an attempt to optimize this threshold or the prompt choice here.
์‹œ์Šคํ…œ์€ ๊ฒ€์ƒ‰๋œ ๊ฐ ๊ธฐ์‚ฌ์— ๋Œ€ํ•ด 1-6์˜ ์ฒ™๋„๋กœ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํ‰์ ์ด 3 ์ดํ•˜์ธ ๊ธฐ์‚ฌ๋Š” ๋ชจ๋‘ ํ๊ธฐ๋ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” ์ด ์ž„๊ณ„๊ฐ’์ด๋‚˜ ํ”„๋กฌํ”„ํŠธ ์„ ํƒ์„ ์ตœ์ ํ™”ํ•˜๋ ค๋Š” ์‹œ๋„๋ฅผ ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
์‹œ์Šคํ…œ์€ ๊ฒ€์ƒ‰๋œ ๊ฐ ๋ฌธ์„œ์˜ ๊ด€๋ จ์„ฑ์„ 1~6์  ์ฒ™๋„๋กœ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. 3์  ์ดํ•˜์˜ ํ‰์ ์„ ๋ฐ›์€ ๋ฌธ์„œ๋Š” ๋ชจ๋‘ ํ๊ธฐ๋ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” ์ด ์ž„๊ณ„๊ฐ’์ด๋‚˜ ํ”„๋กฌํ”„ํŠธ ์„ ํƒ์„ ์ตœ์ ํ™”ํ•˜๋ ค๊ณ  ์‹œ๋„ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
**QUESTION:** Henry made two stops during his 60-mile bike trip. He first stopped after 20 miles. His second stop was 15 miles before the end of the trip. How many miles did he travel between his first and second stops?
**์งˆ๋ฌธ:** Henry๋Š” 60๋งˆ์ผ ์ž์ „๊ฑฐ ์—ฌํ–‰ ์ค‘ ๋‘ ๋ฒˆ์˜ ์ •์ฐจ๋ฅผ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Š” ์ฒ˜์Œ 20๋งˆ์ผ์„ ๋‹ฌ๋ฆฐ ํ›„ ์ฒซ ๋ฒˆ์งธ๋กœ ์ •์ฐจํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ์ •์ฐจ๋Š” ์—ฌํ–‰์ด ๋๋‚˜๊ธฐ 15๋งˆ์ผ ์ „์ด์—ˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์ •์ฐจ์™€ ๋‘ ๋ฒˆ์งธ ์ •์ฐจ ์‚ฌ์ด์— ๊ทธ๋Š” ๋ช‡ ๋งˆ์ผ์„ ์ด๋™ํ–ˆ์Šต๋‹ˆ๊นŒ?
**์งˆ๋ฌธ:** ํ—จ๋ฆฌ๋Š” 60๋งˆ์ผ ์ž์ „๊ฑฐ ์—ฌํ–‰ ์ค‘์— ๋‘ ๋ฒˆ ์ •์ฐจํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Š” ์ฒ˜์Œ 20๋งˆ์ผ ํ›„์— ๋ฉˆ์ท„์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ์ •๊ฑฐ์žฅ์€ ์—ฌํ–‰์ด ๋๋‚˜๊ธฐ 15๋งˆ์ผ ์ „์ด์—ˆ์Šต๋‹ˆ๋‹ค. ํ—จ๋ฆฌ๋Š” ์ฒซ ๋ฒˆ์งธ ์ •๊ฑฐ์žฅ๊ณผ ๋‘ ๋ฒˆ์งธ ์ •๊ฑฐ์žฅ ์‚ฌ์ด๋ฅผ ๋ช‡ ๋งˆ์ผ์ด๋‚˜ ์ด๋™ํ–ˆ์„๊นŒ์š”?
**QUESTION:** Cynthia eats one serving of ice cream every night. She buys cartons of ice cream with 15 servings of ice cream per carton at a cost of $4.00 per carton. After 60 days, how much will she spend on ice cream?
**์งˆ๋ฌธ:** Cynthia๋Š” ๋งค์ผ ๋ฐค ์•„์ด์Šคํฌ๋ฆผ ํ•œ ์„œ๋น™์„ ๋จน์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ํ•œ ์ƒ์ž์— 15 ์„œ๋น™์ด ๋“ค์–ด ์žˆ๋Š” ์•„์ด์Šคํฌ๋ฆผ ์ƒ์ž๋ฅผ ์ƒ์ž๋‹น $4.00์— ๊ตฌ์ž…ํ•ฉ๋‹ˆ๋‹ค. 60์ผ ํ›„, ๊ทธ๋…€๋Š” ์•„์ด์Šคํฌ๋ฆผ์— ์–ผ๋งˆ๋ฅผ ์“ธ๊นŒ์š”?
**์งˆ๋ฌธ:** ์‹ ์‹œ์•„๋Š” ๋งค์ผ ๋ฐค ์•„์ด์Šคํฌ๋ฆผ ํ•œ ์„œ๋น™์„ ๋จน์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ํ•œ ์ƒ์ž๋‹น 15์ธ๋ถ„์˜ ์•„์ด์Šคํฌ๋ฆผ์ด ๋“ค์–ด ์žˆ๋Š” ์•„์ด์Šคํฌ๋ฆผ ์ƒ์ž๋ฅผ ์ƒ์ž๋‹น 4.00๋‹ฌ๋Ÿฌ์— ๊ตฌ์ž…ํ•ฉ๋‹ˆ๋‹ค. 60์ผ ํ›„, ๊ทธ๋…€๋Š” ์•„์ด์Šคํฌ๋ฆผ์— ์–ผ๋งˆ๋ฅผ ์ง€์ถœํ• ๊นŒ์š”?
MM-REACT [76] integrates various external models, such as VideoBERT for video summarization, X-decoder for image generation, and SpeechBERT for audio processing, enhancing its capability in diverse multimodal scenarios.
MM-REACT [76]๋Š” ๋น„๋””์˜ค ์š”์•ฝ์„ ์œ„ํ•œ VideoBERT, ์ด๋ฏธ์ง€ ์ƒ์„ฑ์„ ์œ„ํ•œ X-decoder, ์˜ค๋””์˜ค ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ SpeechBERT์™€ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์™ธ๋ถ€ ๋ชจ๋ธ์„ ํ†ตํ•ฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ์˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
MM-REACT [76]๋Š” ๋™์˜์ƒ ์š”์•ฝ์„ ์œ„ํ•œ VideoBERT, ์ด๋ฏธ์ง€ ์ƒ์„ฑ์„ ์œ„ํ•œ X-decoder, ์˜ค๋””์˜ค ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ SpeechBERT ๋“ฑ ๋‹ค์–‘ํ•œ ์™ธ๋ถ€ ๋ชจ๋ธ์„ ํ†ตํ•ฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ์˜ ๊ธฐ๋Šฅ์„ ๊ฐ•ํ™”ํ•˜์˜€๋‹ค.
Figure 10: The computation graph of mixture law with implicit domain aggregation. We take an case of 3 training domains and 4 implicit validation domains as example. The parameters correspond to the notations in Eqn. 8.
Figure 10: ์•”๋ฌต์  ๋„๋ฉ”์ธ ์ง‘๊ณ„๋ฅผ ํฌํ•จํ•œ ํ˜ผํ•ฉ ๋ฒ•์น™์˜ ๊ณ„์‚ฐ ๊ทธ๋ž˜ํ”„. ์šฐ๋ฆฌ๋Š” 3๊ฐœ์˜ ํ›ˆ๋ จ ๋„๋ฉ”์ธ๊ณผ 4๊ฐœ์˜ ์•”๋ฌต์  ๊ฒ€์ฆ ๋„๋ฉ”์ธ์˜ ์˜ˆ๋ฅผ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” ์‹ 8์˜ ํ‘œ๊ธฐ๋ฒ•์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 10: ์•”์‹œ์  ๋„๋ฉ”์ธ ์ง‘๊ณ„๊ฐ€ ์žˆ๋Š” ํ˜ผํ•ฉ ๋ฒ•์น™์˜ ๊ณ„์‚ฐ ๊ทธ๋ž˜ํ”„. 3๊ฐœ์˜ ํ›ˆ๋ จ ๋„๋ฉ”์ธ๊ณผ 4๊ฐœ์˜ ์•”์‹œ์  ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋„๋ฉ”์ธ์˜ ๊ฒฝ์šฐ๋ฅผ ์˜ˆ๋กœ ๋“ค์–ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” Eqn. 8์˜ ํ‘œ๊ธฐ๋ฒ•์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
Figure 4: Prediction results on the domain losses and overall losses in the three-domain experiment. The domain losses are fitted with Eqn. 7 and we obtain the total losses through explicit domain aggregation of Eqn. 8.
Figure 4: ์„ธ ๊ฐ€์ง€ ๋„๋ฉ”์ธ ์‹คํ—˜์—์„œ ๋„๋ฉ”์ธ ์†์‹ค๊ณผ ์ „์ฒด ์†์‹ค์— ๋Œ€ํ•œ ์˜ˆ์ธก ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ๋„๋ฉ”์ธ ์†์‹ค์€ Eqn. 7๋กœ ๋งž์ถฐ์กŒ์œผ๋ฉฐ, Eqn. 8์˜ ๋ช…์‹œ์  ๋„๋ฉ”์ธ ์ง‘๊ณ„๋ฅผ ํ†ตํ•ด ์ „์ฒด ์†์‹ค์„ ์–ป์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 4: 3๊ฐœ์˜ ๋„๋ฉ”์ธ ์‹คํ—˜์—์„œ ๋„๋ฉ”์ธ ์†์‹ค๊ณผ ์ „์ฒด ์†์‹ค์— ๋Œ€ํ•œ ์˜ˆ์ธก ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ๋„๋ฉ”์ธ ์†์‹ค์€ Eqn. 7๋กœ ๋งž์ถคํ™”๋˜๋ฉฐ Eqn. 8์˜ ๋ช…์‹œ์  ๋„๋ฉ”์ธ ์ง‘๊ณ„๋ฅผ ํ†ตํ•ด ์ด ์†์‹ค์„ ์–ป์Šต๋‹ˆ๋‹ค.
To calculate the revenue from repairing car tires on Thursday and Friday, we need to add up the revenue from each day. Therefore, the total revenue from repairing car tires on Thursday and Friday was $160 + $480 = $640.
๋ชฉ์š”์ผ๊ณผ ๊ธˆ์š”์ผ์— ์ž๋™์ฐจ ํƒ€์ด์–ด ์ˆ˜๋ฆฌ๋กœ ์ธํ•œ ์ˆ˜์ต์„ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๊ฐ ๋‚ ์˜ ์ˆ˜์ต์„ ํ•ฉ์‚ฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ชฉ์š”์ผ๊ณผ ๊ธˆ์š”์ผ์— ์ž๋™์ฐจ ํƒ€์ด์–ด ์ˆ˜๋ฆฌ๋กœ ์ธํ•œ ์ด ์ˆ˜์ต์€ $160 + $480 = $640์ž…๋‹ˆ๋‹ค.
๋ชฉ์š”์ผ๊ณผ ๊ธˆ์š”์ผ์— ์ž๋™์ฐจ ํƒ€์ด์–ด๋ฅผ ์ˆ˜๋ฆฌํ•˜์—ฌ ์–ป์€ ์ˆ˜์ต์„ ๊ณ„์‚ฐํ•˜๋ ค๋ฉด ๊ฐ ๋‚ ์งœ์˜ ์ˆ˜์ต์„ ํ•ฉ์‚ฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ชฉ์š”์ผ๊ณผ ๊ธˆ์š”์ผ์— ์ž๋™์ฐจ ํƒ€์ด์–ด๋ฅผ ์ˆ˜๋ฆฌํ•˜์—ฌ ์–ป์€ ์ด ์ˆ˜์ต์€ $160 + $480 = $640์ž…๋‹ˆ๋‹ค.
In summary, MiniCPM propounds a new stage in the development of small language models, exemplifying the latent potential within SLMs and advocating for a more scientific and sustainable approach toward scaling up LLMs.
์š”์•ฝํ•˜์ž๋ฉด, MiniCPM์€ ์†Œํ˜• ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฐœ๋ฐœ์—์„œ ์ƒˆ๋กœ์šด ๋‹จ๊ณ„๋ฅผ ์ œ์‹œํ•˜๋ฉฐ, SLM ๋‚ด์— ์ž ์žฌ๋œ ๊ฐ€๋Šฅ์„ฑ์„ ์˜ˆ์‹œํ•˜๊ณ  LLM์˜ ํ™•์žฅ์„ ์œ„ํ•œ ๋ณด๋‹ค ๊ณผํ•™์ ์ด๊ณ  ์ง€์† ๊ฐ€๋Šฅํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์˜นํ˜ธํ•ฉ๋‹ˆ๋‹ค.
์š”์•ฝํ•˜์ž๋ฉด, MiniCPM์€ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ฐœ์ „์— ์žˆ์–ด ์ƒˆ๋กœ์šด ๋‹จ๊ณ„๋ฅผ ์ œ์•ˆํ•˜๋ฉฐ, SLM ๋‚ด์˜ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๊ณ  LLM์„ ํ™•์žฅํ•˜๊ธฐ ์œ„ํ•œ ๋ณด๋‹ค ๊ณผํ•™์ ์ด๊ณ  ์ง€์† ๊ฐ€๋Šฅํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์˜นํ˜ธํ•ฉ๋‹ˆ๋‹ค.
**Instruct + Answers**: perplexity on instruction-tuning data from OPT-IML [21], where each sample contains both the instruction-tuning prompt and the answer (in Figure A4 this is referred to as "prompts_with_answers."
**Instruct + Answers**: OPT-IML [21]์˜ instruction-tuning ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ perplexity๋กœ, ๊ฐ ์ƒ˜ํ”Œ์—๋Š” instruction-tuning ํ”„๋กฌํ”„ํŠธ์™€ ๋‹ต๋ณ€์ด ๋ชจ๋‘ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค (Figure A4์—์„œ๋Š” ์ด๋ฅผ "prompts_with_answers"๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค).
**๋ช…๋ น์–ด + ๋‹ต๋ณ€**: OPT-IML [21]์˜ ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋‹นํ˜น๊ฐ, ์—ฌ๊ธฐ์„œ ๊ฐ ์ƒ˜ํ”Œ์—๋Š” ๋ช…๋ น์–ด ํŠœ๋‹ ํ”„๋กฌํ”„ํŠธ์™€ ๋‹ต๋ณ€(๊ทธ๋ฆผ A4์—์„œ๋Š” "prompts_with_answers"๋ผ๊ณ  ํ•จ)์ด ๋ชจ๋‘ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
since \(b\), \(k\), and \(T\) are all \(\ll N\). The left term is the complexity of grouping by the signatures, and the right represents the pathological worst case of all documents falling into the same \(B\) buckets.
\(b\), \(k\), ๊ทธ๋ฆฌ๊ณ  \(T\)๊ฐ€ ๋ชจ๋‘ \(N\)๋ณด๋‹ค ํ›จ์”ฌ ์ž‘๊ธฐ ๋•Œ๋ฌธ์—, ์™ผ์ชฝ ํ•ญ์€ ์„œ๋ช…๋ณ„๋กœ ๊ทธ๋ฃนํ™”ํ•˜๋Š” ๋ณต์žก๋„๋ฅผ ๋‚˜ํƒ€๋‚ด๊ณ , ์˜ค๋ฅธ์ชฝ ํ•ญ์€ ๋ชจ๋“  ๋ฌธ์„œ๊ฐ€ ๋™์ผํ•œ \(B\) ๋ฒ„ํ‚ท์— ์†ํ•˜๋Š” ๋ณ‘๋ฆฌํ•™์  ์ตœ์•…์˜ ๊ฒฝ์šฐ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
b\), \(k\), \(T\)๋Š” ๋ชจ๋‘ \(\ll N\)์ž…๋‹ˆ๋‹ค. ์™ผ์ชฝ ํ•ญ์€ ์„œ๋ช…๋ณ„๋กœ ๊ทธ๋ฃนํ™”ํ•˜๋Š” ๋ณต์žก์„ฑ์ด๊ณ , ์˜ค๋ฅธ์ชฝ์€ ๋ชจ๋“  ๋ฌธ์„œ๊ฐ€ ๋™์ผํ•œ \(B\) ๋ฒ„ํ‚ท์— ์†ํ•˜๋Š” ๋ณ‘๋ฆฌํ•™์  ์ตœ์•…์˜ ๊ฒฝ์šฐ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
1. Increased Endurance: HIIT can help athletes build their endurance by improving their ability to perform at high levels for shorter bursts of time, which can transfer to longer, more intense workouts and competition.
1. ์ง€๊ตฌ๋ ฅ ํ–ฅ์ƒ: HIIT๋Š” ์šด๋™์„ ์ˆ˜๋“ค์ด ์งง์€ ์‹œ๊ฐ„ ๋™์•ˆ ๋†’์€ ์ˆ˜์ค€์—์„œ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œ์ผœ, ๋” ๊ธธ๊ณ  ๊ฐ•๋„ ๋†’์€ ์šด๋™๊ณผ ๊ฒฝ๊ธฐ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ค๋‹ˆ๋‹ค.
1. ์ง€๊ตฌ๋ ฅ ํ–ฅ์ƒ: HIIT๋Š” ์งง์€ ์‹œ๊ฐ„ ๋™์•ˆ ๋†’์€ ์ˆ˜์ค€์œผ๋กœ ์šด๋™ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œ์ผœ ์šด๋™์„ ์ˆ˜๊ฐ€ ์ง€๊ตฌ๋ ฅ์„ ํ‚ค์šฐ๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋” ๊ธธ๊ณ  ๊ฐ•๋„ ๋†’์€ ์šด๋™๊ณผ ๊ฒฝ๊ธฐ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
CC-RecentNews-Small into two different splits by the published date of each article to simulate a setting where multiple CKL phases are needed, denoted as Small-P1 (05.2020 - 11.2020)) and Small-P2 (11.2020 - 04.2021).
CC-RecentNews-Small์„ ๊ฐ ๊ธฐ์‚ฌ ๋ฐœํ–‰์ผ์— ๋”ฐ๋ผ ๋‘ ๊ฐœ์˜ ๋‹ค๋ฅธ ๋ถ„ํ• ๋กœ ๋‚˜๋ˆ„์–ด ์—ฌ๋Ÿฌ CKL ๋‹จ๊ณ„๊ฐ€ ํ•„์š”ํ•œ ์„ค์ •์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ Small-P1 (2020๋…„ 5์›” - 2020๋…„ 11์›”)๊ณผ Small-P2 (2020๋…„ 11์›” - 2021๋…„ 4์›”)๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.
CC-RecentNews-Small์„ ๊ฐ ๊ธฐ์‚ฌ์˜ ๊ฒŒ์‹œ ๋‚ ์งœ๋ณ„๋กœ ๋‘ ๊ฐœ์˜ ๋‹ค๋ฅธ ๋ถ„ํ• ๋กœ ๋‚˜๋ˆ„์–ด ์—ฌ๋Ÿฌ CKL ๋‹จ๊ณ„๊ฐ€ ํ•„์š”ํ•œ ์„ค์ •์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค(Small-P1(05.2020 - 11.2020) ๋ฐ Small-P2(11.2020 - 04.2021)๋กœ ํ‘œ์‹œ๋จ).
3. Fast Results: Due to its high-intensity nature, HIIT can provide athletes with a significant training stimulus in a short amount of time, allowing them to see results faster than with traditional forms of exercise.
3. ๋น ๋ฅธ ๊ฒฐ๊ณผ: HIIT๋Š” ๋†’์€ ๊ฐ•๋„์˜ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ์šด๋™ ์„ ์ˆ˜๋“ค์—๊ฒŒ ์งง์€ ์‹œ๊ฐ„ ๋‚ด์— ์ƒ๋‹นํ•œ ํ›ˆ๋ จ ์ž๊ทน์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ „ํ†ต์ ์ธ ์šด๋™ ํ˜•ํƒœ๋ณด๋‹ค ๋” ๋น ๋ฅด๊ฒŒ ๊ฒฐ๊ณผ๋ฅผ ๋ณผ ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.
3. ๋น ๋ฅธ ๊ฒฐ๊ณผ: ๊ณ ๊ฐ•๋„ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด HIIT๋Š” ์šด๋™์„ ์ˆ˜์—๊ฒŒ ๋‹จ์‹œ๊ฐ„์— ์ƒ๋‹นํ•œ ํ›ˆ๋ จ ์ž๊ทน์„ ์ œ๊ณตํ•˜์—ฌ ๊ธฐ์กด ํ˜•ํƒœ์˜ ์šด๋™๋ณด๋‹ค ๋น ๋ฅด๊ฒŒ ๊ฒฐ๊ณผ๋ฅผ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Remark**.: _It's worth noting that many common LLMs Tunstall et al. (2023) include an additional step of to align the model with human preference Rafailov et al. (2023); Munos et al. (2023); von Werra et al. (2020).
**Remark**: _์ฃผ๋ชฉํ•  ์ ์€, Tunstall et al. (2023)์— ๋”ฐ๋ฅด๋ฉด ๋งŽ์€ ์ผ๋ฐ˜์ ์ธ LLM๋“ค์€ ๋ชจ๋ธ์„ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž์ถ”๊ธฐ ์œ„ํ•œ ์ถ”๊ฐ€ ๋‹จ๊ณ„๋ฅผ ํฌํ•จํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค (Rafailov et al., 2023; Munos et al., 2023; von Werra et al., 2020)._
**์ฐธ๊ณ **.: _๋งŽ์€ ์ผ๋ฐ˜์ ์ธ LLM์— ์ถ”๊ฐ€ ๋‹จ๊ณ„๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋‹ค๋Š” ์ ์— ์ฃผ๋ชฉํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠœ instal et al. (2023)์—๋Š” ๋ชจ๋ธ์„ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž์ถ”๋Š” ๋‹จ๊ณ„๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค(Rafailov et al., 2023; Munos et al., 2023; von Werra et al., 2020).
Which of the following summaries does a better job of summarizing the most \ important points in the given forum post, without including unimportant or \ irrelevant details? A good summary is both precise and concise.
๋‹ค์Œ ์š”์•ฝ ์ค‘ ์–ด๋Š ๊ฒƒ์ด ์ฃผ์–ด์ง„ ํฌ๋Ÿผ ๊ฒŒ์‹œ๋ฌผ์—์„œ ์ค‘์š”ํ•˜์ง€ ์•Š๊ฑฐ๋‚˜ ๊ด€๋ จ ์—†๋Š” ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ํฌํ•จํ•˜์ง€ ์•Š๊ณ  ๊ฐ€์žฅ ์ค‘์š”ํ•œ ์ ๋“ค์„ ๋” ์ž˜ ์š”์•ฝํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๊นŒ? ์ข‹์€ ์š”์•ฝ์€ ์ •ํ™•ํ•˜๊ณ  ๊ฐ„๊ฒฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๋‹ค์Œ ์ค‘ ์ค‘์š”ํ•˜์ง€ ์•Š๊ฑฐ๋‚˜ ๊ด€๋ จ ์—†๋Š” ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ํฌํ•จํ•˜์ง€ ์•Š๊ณ  ์ฃผ์–ด์ง„ ํฌ๋Ÿผ ๊ฒŒ์‹œ๋ฌผ์˜ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ์š”์ ์„ ๋” ์ž˜ ์š”์•ฝํ•œ ์š”์•ฝ์€ ๋ฌด์—‡์ž…๋‹ˆ๊นŒ? ์ข‹์€ ์š”์•ฝ์€ ์ •ํ™•ํ•˜๊ณ  ๊ฐ„๊ฒฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
As illustrated in Figure 3, when analyzing the performance of HumanEval and MABP in the code generation domain using 7b models, it becomes apparent that the initial performance is comparable across the various models.
Figure 3์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, 7b ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฝ”๋“œ ์ƒ์„ฑ ๋„๋ฉ”์ธ์—์„œ HumanEval๊ณผ MABP์˜ ์„ฑ๋Šฅ์„ ๋ถ„์„ํ•  ๋•Œ, ์ดˆ๊ธฐ ์„ฑ๋Šฅ์ด ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ๊ฐ„์— ์œ ์‚ฌํ•˜๋‹ค๋Š” ๊ฒƒ์ด ๋ช…ํ™•ํ•ด์ง‘๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 3์— ์˜ˆ์‹œ๋œ ๋ฐ”์™€ ๊ฐ™์ด 7b ๋ชจ๋ธ์„ ์ด์šฉํ•˜์—ฌ ์ฝ”๋“œ ์ƒ์„ฑ ์˜์—ญ์—์„œ HumanEval๊ณผ MABP์˜ ์„ฑ๋Šฅ์„ ๋ถ„์„ํ•ด ๋ณด๋ฉด ์ดˆ๊ธฐ ์„ฑ๋Šฅ์€ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ์—์„œ ๋น„์Šทํ•œ ์ˆ˜์ค€์ž„์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค.
_Thus, during this stage, we recover the unpolluted sample-size law scaling \(T^{-(1-1/\beta)}\), up to within a multiplicative constant \(\pi^{-(1-1/\beta)}\) (which can be seen as an increase in the price of data).
๋”ฐ๋ผ์„œ, ์ด ๋‹จ๊ณ„์—์„œ๋Š” ๊ณฑ์…ˆ ์ƒ์ˆ˜ \(\pi^{-(1-1/\beta)}\) ๋‚ด์—์„œ ์˜ค์—ผ๋˜์ง€ ์•Š์€ ์ƒ˜ํ”Œ ํฌ๊ธฐ ๋ฒ•์น™ ์Šค์ผ€์ผ๋ง \(T^{-(1-1/\beta)}\)์„ ํšŒ๋ณตํ•ฉ๋‹ˆ๋‹ค (์ด๋Š” ๋ฐ์ดํ„ฐ ๊ฐ€๊ฒฉ์˜ ์ฆ๊ฐ€๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค).
๋”ฐ๋ผ์„œ ์ด ๋‹จ๊ณ„์—์„œ๋Š” ์ƒ˜ํ”Œ ํฌ๊ธฐ ๋ฒ•์น™ \(T^{-(1-1/\beta)}\)์„ ๊ณฑ์…ˆ ์ƒ์ˆ˜ \(\pi^{-(1-1/\beta)}\) ์ด๋‚ด(๋ฐ์ดํ„ฐ ๊ฐ€๊ฒฉ์˜ ์ƒ์Šน์œผ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Œ)๋กœ ๋ณต๊ตฌํ•ฉ๋‹ˆ๋‹ค.
My character reached out a hand to touch the object, and as soon as their fingertips brushed against it, they felt a surge of electricity coursing through their body. They gased and stepped back, momentarily stunned.
๋‚ด ์บ๋ฆญํ„ฐ๋Š” ์†์„ ๋ป—์–ด ๊ทธ ๋ฌผ์ฒด๋ฅผ ๋งŒ์ง€๋ ค ํ–ˆ๊ณ , ์†๋์ด ๊ทธ๊ฒƒ์— ๋‹ฟ์ž๋งˆ์ž ๋ชธ ์ „์ฒด์— ์ „๊ธฐ๊ฐ€ ํ๋ฅด๋Š” ๊ฒƒ์„ ๋Š๊ผˆ๋‹ค. ๊ทธ๋“ค์€ ์ˆจ์„ ํ—๋–ก์ด๋ฉฐ ์ž ์‹œ ๋†€๋ผ์„œ ๋’ท๊ฑธ์Œ์งˆ์ณค๋‹ค.
๋‚ด ์บ๋ฆญํ„ฐ๊ฐ€ ๋ฌผ์ฒด๋ฅผ ๋งŒ์ง€๊ธฐ ์œ„ํ•ด ์†์„ ๋ป—์—ˆ๊ณ  ์†๋์ด ๋ฌผ์ฒด์— ๋‹ฟ์ž๋งˆ์ž ๋ชธ ์ „์ฒด์— ์ „๊ธฐ๊ฐ€ ํ๋ฅด๋Š” ๊ฒƒ์„ ๋Š๊ผˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ์ˆจ์„ ํ—๋–ก์ด๋ฉฐ ํ•œ ๋ฐœ์ง ๋ฌผ๋Ÿฌ์„ฐ์Šต๋‹ˆ๋‹ค.
We also conduct experiments by freezing specific modules of the model during continual pre-training. We aim to explore whether these designated modules preserve the abilities acquired during the alignment operations.
์šฐ๋ฆฌ๋Š” ๋˜ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ ๋™์•ˆ ๋ชจ๋ธ์˜ ํŠน์ • ๋ชจ๋“ˆ์„ ๊ณ ์ •ํ•˜์—ฌ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง€์ •๋œ ๋ชจ๋“ˆ์ด ์ •๋ ฌ ์ž‘์—… ์ค‘์— ํš๋“ํ•œ ๋Šฅ๋ ฅ์„ ์œ ์ง€ํ•˜๋Š”์ง€ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
๋˜ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋ชจ๋ธ์˜ ํŠน์ • ๋ชจ๋“ˆ์„ ๋™๊ฒฐํ•˜์—ฌ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง€์ •๋œ ๋ชจ๋“ˆ์ด ์ •๋ ฌ ์ž‘์—… ์ค‘์— ์Šต๋“ํ•œ ๋Šฅ๋ ฅ์„ ์œ ์ง€ํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
Figure A.8 shows the DDx accuracy achieved by AMIE and PCPs for each specialty based on specialist ratings. Specifically, we observed that AMIE's performance matched or surpassed PCPs performance for all specialties.
Figure A.8์€ ๊ฐ ์ „๋ฌธ ๋ถ„์•ผ์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ AMIE์™€ PCPs๊ฐ€ ๋‹ฌ์„ฑํ•œ DDx ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” AMIE์˜ ์„ฑ๋Šฅ์ด ๋ชจ๋“  ์ „๋ฌธ ๋ถ„์•ผ์—์„œ PCPs์˜ ์„ฑ๋Šฅ๊ณผ ์ผ์น˜ํ•˜๊ฑฐ๋‚˜ ์ด๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ A.8์€ ์ „๋ฌธ์˜ ๋“ฑ๊ธ‰์„ ๊ธฐ์ค€์œผ๋กœ ํ•œ ๊ฐ ์ „๋ฌธ ๋ถ„์•ผ์— ๋Œ€ํ•œ AMIE์™€ PCP๊ฐ€ ๋‹ฌ์„ฑํ•œ DDx ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ํŠนํžˆ, ๋ชจ๋“  ์ „๋ฌธ ๋ถ„์•ผ์—์„œ AMIE์˜ ์„ฑ๋Šฅ์ด PCP์˜ ์„ฑ๋Šฅ๊ณผ ์ผ์น˜ํ•˜๊ฑฐ๋‚˜ ์ด๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ด€์ฐฐ๋˜์—ˆ๋‹ค.
Figure 11: Comparison between synthetic and real data from the C4 corpus showing that synthetic data have higher syntactic complexity indicated by higher average tree depth, and higher mean dependency distance (MDD).
Figure 11: C4 ์ฝ”ํผ์Šค์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ํ‰๊ท  ํŠธ๋ฆฌ ๊นŠ์ด๊ฐ€ ๋” ๋†’๊ณ , ํ‰๊ท  ์˜์กด ๊ฑฐ๋ฆฌ(MDD)๊ฐ€ ๋” ๋†’์•„ ๋” ๋†’์€ ๊ตฌ๋ฌธ ๋ณต์žก์„ฑ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 11: C4 ๋ง๋ญ‰์น˜์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ๋” ๋†’์€ ํ‰๊ท  ํŠธ๋ฆฌ ๊นŠ์ด์™€ ๋” ๋†’์€ ํ‰๊ท  ์ข…์†์„ฑ ๊ฑฐ๋ฆฌ(MDD)๋กœ ๋‚˜ํƒ€๋‚ด๋Š” ๊ตฌ๋ฌธ์  ๋ณต์žก์„ฑ์ด ๋” ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.