question
stringlengths
101
2.69k
chosen
stringlengths
43
2.27k
rejected
stringlengths
42
2.18k
Figure 3: Data distribution of the 43 programming languages used for pre-training. The number within each bar indicates the percent of the overall code distribution that an individual language comprises.
Figure 3: ์‚ฌ์ „ ํ›ˆ๋ จ์— ์‚ฌ์šฉ๋œ 43๊ฐœ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์˜ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ. ๊ฐ ๋ง‰๋Œ€ ์•ˆ์˜ ์ˆซ์ž๋Š” ๊ฐœ๋ณ„ ์–ธ์–ด๊ฐ€ ์ „์ฒด ์ฝ”๋“œ ๋ถ„ํฌ์—์„œ ์ฐจ์ง€ํ•˜๋Š” ๋น„์œจ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 3: ์‚ฌ์ „ ํ•™์Šต์— ์‚ฌ์šฉ๋œ 43๊ฐœ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์˜ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ. ๊ฐ ๋ง‰๋Œ€ ์•ˆ์˜ ์ˆซ์ž๋Š” ๊ฐœ๋ณ„ ์–ธ์–ด๊ฐ€ ์ „์ฒด ์ฝ”๋“œ ๋ถ„ํฌ์—์„œ ์ฐจ์ง€ํ•˜๋Š” ๋น„์œจ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
where, \(f(\theta;\cdot)\) is a LLM with parameters \(\theta\), \(y\) is the task output, \(x\) is an input in target task data \(\mathcal{T}\), and \(\mathcal{L}_{t}\) is the target task loss or metric.
์—ฌ๊ธฐ์„œ \(f(\theta;\cdot)\)๋Š” ๋งค๊ฐœ๋ณ€์ˆ˜ \(\theta\)๋ฅผ ๊ฐ€์ง„ LLM์ด๋ฉฐ, \(y\)๋Š” ์ž‘์—…์˜ ์ถœ๋ ฅ, \(x\)๋Š” ๋Œ€์ƒ ์ž‘์—… ๋ฐ์ดํ„ฐ \(\mathcal{T}\)์˜ ์ž…๋ ฅ, \(\mathcal{L}_{t}\)๋Š” ๋Œ€์ƒ ์ž‘์—…์˜ ์†์‹ค ๋˜๋Š” ์ง€ํ‘œ์ž…๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ, \(f(\theta;\cdot)\)๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ \(\theta\)๋ฅผ ๊ฐ–๋Š” LLM์ด๊ณ , \(y\)๋Š” ์ž‘์—… ์ถœ๋ ฅ์ด๋ฉฐ, \(x\)๋Š” ๋Œ€์ƒ ์ž‘์—… ๋ฐ์ดํ„ฐ \(\mathcal{T}\)์˜ ์ž…๋ ฅ์ด๊ณ , \(\mathcal{L}_{t}\)๋Š” ๋Œ€์ƒ ์ž‘์—… ์†์‹ค ๋˜๋Š” ๋ฉ”ํŠธ๋ฆญ์ž…๋‹ˆ๋‹ค.
Footnote 3: For example, we found that using ## before the Question can lead to a noticeable improvement to **phi-3-mini**โ€™s results across many benchmarks, but we did not do such changes in the prompts.
Footnote 3: ์˜ˆ๋ฅผ ๋“ค์–ด, ์งˆ๋ฌธ ์•ž์— ##๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์—ฌ๋Ÿฌ ๋ฒค์น˜๋งˆํฌ์—์„œ **phi-3-mini**์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ˆˆ์— ๋„๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์ง€๋งŒ, ์šฐ๋ฆฌ๋Š” ํ”„๋กฌํ”„ํŠธ์—์„œ ๊ทธ๋Ÿฌํ•œ ๋ณ€๊ฒฝ์„ ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
๊ฐ์ฃผ 3: ์˜ˆ๋ฅผ ๋“ค์–ด, ์งˆ๋ฌธ ์•ž์— ##๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋งŽ์€ ๋ฒค์น˜๋งˆํฌ์—์„œ **phi-3-mini**์˜ ๊ฒฐ๊ณผ๊ฐ€ ๋ˆˆ์— ๋„๊ฒŒ ๊ฐœ์„ ๋˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์ง€๋งŒ ํ”„๋กฌํ”„ํŠธ์—์„œ๋Š” ๊ทธ๋Ÿฌํ•œ ๋ณ€๊ฒฝ์„ ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
Proof.: Let us consider two reward functions from the same class, such that \(r^{\prime}(x,y)=r(x,y)+f(x)\) and, let us denote as \(\pi_{r}\) and \(\pi_{r^{\prime}}\) the corresponding optimal policies.
์ฆ๋ช…: ๋™์ผํ•œ ํด๋ž˜์Šค์—์„œ ๋‘ ๊ฐœ์˜ ๋ณด์ƒ ํ•จ์ˆ˜ \(r^{\prime}(x,y)=r(x,y)+f(x)\)๋ฅผ ๊ณ ๋ คํ•ด ๋ณด์ž. ๊ทธ๋ฆฌ๊ณ  ํ•ด๋‹นํ•˜๋Š” ์ตœ์  ์ •์ฑ…์„ \(\pi_{r}\)์™€ \(\pi_{r^{\prime}}\)๋กœ ๋‚˜ํƒ€๋‚ด์ž.
์ฆ๋ช….: ๋™์ผํ•œ ํด๋ž˜์Šค์˜ ๋‘ ๋ณด์ƒ ํ•จ์ˆ˜๋ฅผ ๊ณ ๋ คํ•˜์—ฌ \(r^{\ํ”„๋ผ์ž„}(x,y)=r(x,y)+f(x)\)๊ฐ€ ๋˜๋„๋ก ํ•˜๊ณ , ํ•ด๋‹น ์ตœ์  ์ •์ฑ…์„ \(\pi_{r}\) ๋ฐ \(\pi_{r^{\ํ”„๋ผ์ž„}}\)๋กœ ํ‘œ์‹œํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.
The results in Section 4.1 indicate that, given a fixed amount of compute for training, selecting data from larger and larger source datasets is a promising method to improve language model performance.
์„น์…˜ 4.1์˜ ๊ฒฐ๊ณผ๋Š” ๊ณ ์ •๋œ ์–‘์˜ ์ปดํ“จํŒ… ์ž์›์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•  ๋•Œ, ๋” ํฌ๊ณ  ๋” ํฐ ์†Œ์Šค ๋ฐ์ดํ„ฐ์…‹์—์„œ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ์–ธ์–ด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์œ ๋งํ•œ ๋ฐฉ๋ฒ•์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์„น์…˜ 4.1์˜ ๊ฒฐ๊ณผ๋Š” ๊ณ ์ •๋œ ์–‘์˜ ์ปดํ“จํŒ… ํŠธ๋ ˆ์ด๋‹์ด ์ฃผ์–ด์กŒ์„ ๋•Œ, ์ ์  ๋” ํฐ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ์–ธ์–ด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์œ ๋งํ•œ ๋ฐฉ๋ฒ•์ž„์„ ๋‚˜ํƒ€๋‚ธ๋‹ค.
We justify our choice of the news API's. To begin with, we experiment with 5 eligible APIs to news corpus that accept retrieval date ranges, which, for our purpose, must be specified to prevent leakage.
์šฐ๋ฆฌ๋Š” ๋‰ด์Šค API ์„ ํƒ์˜ ์ •๋‹น์„ฑ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ์šฐ์„ , ์šฐ๋ฆฌ๋Š” ๊ฒ€์ƒ‰ ๋‚ ์งœ ๋ฒ”์œ„๋ฅผ ์ˆ˜๋ฝํ•˜๋Š” ๋‰ด์Šค ์ฝ”ํผ์Šค์— ์ ํ•ฉํ•œ 5๊ฐœ์˜ API๋ฅผ ์‹คํ—˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์šฐ๋ฆฌ์˜ ๋ชฉ์ ์„ ์œ„ํ•ด ๋ˆ„์ถœ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ˜๋“œ์‹œ ์ง€์ •๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๋‰ด์Šค API์˜ ์„ ํƒ์„ ์ •๋‹นํ™”ํ•ฉ๋‹ˆ๋‹ค. ์šฐ์„ , ๊ฒ€์ƒ‰ ๋‚ ์งœ ๋ฒ”์œ„๋ฅผ ํ—ˆ์šฉํ•˜๋Š” ๋‰ด์Šค ๋ง๋ญ‰์น˜์— ๋Œ€ํ•ด 5๊ฐœ์˜ ์ ๊ฒฉ API๋ฅผ ์‹คํ—˜ํ•ด ๋ณด์•˜๋Š”๋ฐ, ์ด๋Š” ๋ชฉ์ ์ƒ ์œ ์ถœ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ์ง€์ •ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Figure 3: (a) The skills distribution of _Evol-Instruct_ testset, and (b) The difficulty and complexity level distribution between the testset of Vicuna, Alpaca (Self-Instruct), and our _Evol-Instruct_.
Figure 3: (a) _Evol-Instruct_ ํ…Œ์ŠคํŠธ์…‹์˜ ๊ธฐ์ˆ  ๋ถ„ํฌ, (b) Vicuna, Alpaca (Self-Instruct), ๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ์˜ _Evol-Instruct_ ํ…Œ์ŠคํŠธ์…‹ ๊ฐ„์˜ ๋‚œ์ด๋„ ๋ฐ ๋ณต์žก๋„ ์ˆ˜์ค€ ๋ถ„ํฌ.
๊ทธ๋ฆผ 3: (a) _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ์Šคํ‚ฌ ๋ถ„ํฌ, (b) Vicuna, Alpaca(Self-Instruct)์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ ์šฐ๋ฆฌ์˜ _Evol-Instruct_์˜ ๋‚œ์ด๋„ ๋ฐ ๋ณต์žก๋„ ์ˆ˜์ค€ ๋ถ„ํฌ.
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.
Figure 5: **LegalBench-Instruct์—์„œ instruct ๋ชจ๋ธ๋“ค์˜ ๋น„๊ต**. SaulLM-7B-Instruct๋Š” ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ตœ๊ณ ์˜ Mistral-Instruct ๋ชจ๋ธ์„ 6์  ์ฐจ์ด๋กœ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
๊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.
GPT2 with Smaller MLP.Mistral has a larger MLP layer, and it is often believed that the MLP layer serves primarily for storing knowledge, in contrast to the Attention layer. But is this truly the case?
GPT2๋Š” ๋” ์ž‘์€ MLP๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Mistral์€ ๋” ํฐ MLP ๋ ˆ์ด์–ด๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ MLP ๋ ˆ์ด์–ด๋Š” Attention ๋ ˆ์ด์–ด์™€๋Š” ๋‹ฌ๋ฆฌ ์ฃผ๋กœ ์ง€์‹์„ ์ €์žฅํ•˜๋Š” ์—ญํ• ์„ ํ•œ๋‹ค๊ณ  ์—ฌ๊ฒจ์ง‘๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ด๊ฒƒ์ด ์ •๋ง ์‚ฌ์‹ค์ผ๊นŒ์š”?
๋” ์ž‘์€ MLP๋ฅผ ์‚ฌ์šฉํ•˜๋Š” GPT2.Mistral์€ MLP ๋ ˆ์ด์–ด๊ฐ€ ๋” ํฌ๋ฉฐ, MLP ๋ ˆ์ด์–ด๋Š” ์ฃผ๋กœ ์ฃผ์˜๋ ฅ ๋ ˆ์ด์–ด์™€๋Š” ๋‹ฌ๋ฆฌ ์ง€์‹์„ ์ €์žฅํ•˜๋Š” ์—ญํ• ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์—ฌ๊ฒจ์ง€๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์‹ค์ œ๋กœ ๊ทธ๋Ÿฐ๊ฐ€์š”?
**Dealing with Ambiguities:** Resolving or excluding ambiguous or contradictory data that might confuse the model during training. This can help the model to provide more definite and reliable answers.
**๋ชจํ˜ธ์„ฑ ์ฒ˜๋ฆฌ:** ํ›ˆ๋ จ ์ค‘ ๋ชจ๋ธ์„ ํ˜ผ๋ž€์Šค๋Ÿฝ๊ฒŒ ํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ์ƒ๋ฐ˜๋˜๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ํ•ด๊ฒฐํ•˜๊ฑฐ๋‚˜ ์ œ์™ธํ•˜๋Š” ๊ฒƒ. ์ด๋Š” ๋ชจ๋ธ์ด ๋” ๋ช…ํ™•ํ•˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**๋ชจํ˜ธ์„ฑ ์ฒ˜๋ฆฌ:** ํ•™์Šต ์ค‘์— ๋ชจ๋ธ์„ ํ˜ผ๋™์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ๋ชจ์ˆœ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ํ•ด๊ฒฐํ•˜๊ฑฐ๋‚˜ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ๋ชจ๋ธ์ด ๋ณด๋‹ค ๋ช…ํ™•ํ•˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Despite our best efforts to filter content that does not qualify as natural language, and to deduplicate documents, our pipeline may let through documents that may be considered as errors or redundant.
์ตœ๋Œ€ํ•œ ์ž์—ฐ์–ด๋กœ ๊ฐ„์ฃผ๋˜์ง€ ์•Š๋Š” ์ฝ˜ํ…์ธ ๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ณ  ๋ฌธ์„œ๋ฅผ ์ค‘๋ณต ์ œ๊ฑฐํ•˜๋ ค๋Š” ์ตœ์„ ์˜ ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์šฐ๋ฆฌ์˜ ํŒŒ์ดํ”„๋ผ์ธ์€ ์˜ค๋ฅ˜๋กœ ๊ฐ„์ฃผ๋˜๊ฑฐ๋‚˜ ์ค‘๋ณต๋œ ๋ฌธ์„œ๋ฅผ ํ†ต๊ณผ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ž์—ฐ์–ด์— ํ•ด๋‹นํ•˜์ง€ ์•Š๋Š” ์ฝ˜ํ…์ธ ๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ณ  ๋ฌธ์„œ ์ค‘๋ณต์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด ์ตœ์„ ์„ ๋‹คํ–ˆ์ง€๋งŒ, ํŒŒ์ดํ”„๋ผ์ธ์—์„œ ์˜ค๋ฅ˜ ๋˜๋Š” ์ค‘๋ณต์œผ๋กœ ๊ฐ„์ฃผ๋  ์ˆ˜ ์žˆ๋Š” ๋ฌธ์„œ๊ฐ€ ํ†ต๊ณผ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In this section, we provide the accuracy curves of our experiments across various datasets when utilizing different LLMs. From these curves, we demonstrate that our method has the following properties:
์ด ์„น์…˜์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ์…‹์„ ํ™œ์šฉํ•˜์—ฌ ์„œ๋กœ ๋‹ค๋ฅธ LLM์„ ์‚ฌ์šฉํ•  ๋•Œ์˜ ์‹คํ—˜ ์ •ํ™•๋„ ๊ณก์„ ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณก์„ ์„ ํ†ตํ•ด ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์ด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŠน์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋‹ค์–‘ํ•œ LLM์„ ํ™œ์šฉํ•  ๋•Œ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์‹คํ—˜์˜ ์ •ํ™•๋„ ๊ณก์„ ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณก์„ ์„ ํ†ตํ•ด ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŠน์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค:
In this paper, our primary focus is on \(\mathsf{bioS}(N)\) for \(N\) ranging between 10K and 20M. Notably, \(\mathsf{bioS}(20M)\) encompasses approximately 1B bits of knowledge (refer to Theorem 3.2).
์ด ๋…ผ๋ฌธ์—์„œ ์šฐ๋ฆฌ์˜ ์ฃผ์š” ์ดˆ์ ์€ \(N\)์ด 10K์—์„œ 20M ์‚ฌ์ด์ธ \(\mathsf{bioS}(N)\)์— ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, \(\mathsf{bioS}(20M)\)์€ ์•ฝ 10์–ต ๋น„ํŠธ์˜ ์ง€์‹์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค (์ •๋ฆฌ 3.2๋ฅผ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค).
์ด ๋…ผ๋ฌธ์—์„œ๋Š” 10K์—์„œ 20M ์‚ฌ์ด์˜ \(N\)์— ๋Œ€ํ•œ \(\mathsf{bioS}(N)\)์— ์ฃผ๋กœ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, \(\mathsf{bioS}(20M)\)๋Š” ์•ฝ 1B ๋น„ํŠธ์˜ ์ง€์‹์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค(์ •๋ฆฌ 3.2 ์ฐธ์กฐ).
loss are more learnable and better aligned with the desired distribution, naturally excluding tokens that are either irrelevant or of low quality. Below, we provide a detailed description of each step.
์†์‹ค์€ ๋” ํ•™์Šต ๊ฐ€๋Šฅํ•˜๊ณ  ์›ํ•˜๋Š” ๋ถ„ํฌ์™€ ๋” ์ž˜ ๋งž์œผ๋ฉฐ, ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ๊ด€๋ จ์ด ์—†๊ฑฐ๋‚˜ ํ’ˆ์งˆ์ด ๋‚ฎ์€ ํ† ํฐ์„ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. ์•„๋ž˜์—์„œ๋Š” ๊ฐ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์†์‹ค์€ ๋” ์‰ฝ๊ฒŒ ๋ฐฐ์šธ ์ˆ˜ ์žˆ๊ณ  ์›ํ•˜๋Š” ๋ถ„ํฌ์— ๋” ์ž˜ ๋ถ€ํ•ฉํ•˜๋ฉฐ, ๊ด€๋ จ์„ฑ์ด ์—†๊ฑฐ๋‚˜ ํ’ˆ์งˆ์ด ๋‚ฎ์€ ํ† ํฐ์„ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. ์•„๋ž˜์—์„œ๋Š” ๊ฐ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
_We train models on this mixture, ensuring each piece of useful data is seen for 100 exposures, thus making the total training 8 times longer compared to 100 exposures without junk (i.e., Figure 1(b)).
์ด ํ˜ผํ•ฉ๋ฌผ๋กœ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ์œ ์šฉํ•œ ๋ฐ์ดํ„ฐ์˜ ๊ฐ ๋ถ€๋ถ„์ด 100๋ฒˆ ๋…ธ์ถœ๋˜๋„๋ก ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์“ธ๋ชจ์—†๋Š” ๋ฐ์ดํ„ฐ ์—†์ด 100๋ฒˆ ๋…ธ์ถœ๋˜๋Š” ๊ฒฝ์šฐ์— ๋น„ํ•ด ์ด ํ›ˆ๋ จ ์‹œ๊ฐ„์ด 8๋ฐฐ ๋” ๊ธธ์–ด์ง‘๋‹ˆ๋‹ค (์ฆ‰, Figure 1(b)).
_์ด ํ˜ผํ•ฉ๋ฌผ์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผœ ์œ ์šฉํ•œ ๋ฐ์ดํ„ฐ์˜ ๊ฐ ์กฐ๊ฐ์ด 100ํšŒ ๋…ธ์ถœ๋˜๋„๋ก ํ•˜์—ฌ ์ •ํฌ๊ฐ€ ์—†๋Š” 100ํšŒ ๋…ธ์ถœ(์ฆ‰, ๊ทธ๋ฆผ 1(b))์— ๋น„ํ•ด ์ด ํ•™์Šต ์‹œ๊ฐ„์ด 8๋ฐฐ ๋” ๊ธธ์–ด์ง‘๋‹ˆ๋‹ค.
Please consider the following forecasting question and its background information. After that, I will give you a news article and ask you to rate its relevance with respect to the forecasting question.
๋‹ค์Œ ์˜ˆ์ธก ์งˆ๋ฌธ๊ณผ ๊ทธ ๋ฐฐ๊ฒฝ ์ •๋ณด๋ฅผ ๊ณ ๋ คํ•ด ์ฃผ์„ธ์š”. ๊ทธ ํ›„, ๋‰ด์Šค ๊ธฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์˜ˆ์ธก ์งˆ๋ฌธ๊ณผ์˜ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•ด ๋‹ฌ๋ผ๊ณ  ์š”์ฒญํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๋‹ค์Œ ์˜ˆ์ธก ์งˆ๋ฌธ๊ณผ ๊ทธ ๋ฐฐ๊ฒฝ ์ •๋ณด๋ฅผ ๊ณ ๋ คํ•ด ์ฃผ์„ธ์š”. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋‰ด์Šค ๊ธฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์˜ˆ์ธก ์งˆ๋ฌธ๊ณผ ๊ด€๋ จํ•˜์—ฌ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•ด ๋‹ฌ๋ผ๊ณ  ์š”์ฒญํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
When RALM was first proposed, many researchers used this method because it aligned with their original ideas, particularly those of Lewis et al. (2020), Guu et al. (2020), and Izacard and Grave (2020).
RALM์ด ์ฒ˜์Œ ์ œ์•ˆ๋˜์—ˆ์„ ๋•Œ, ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด ์ด ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ–ˆ๋Š”๋ฐ, ์ด๋Š” ํŠนํžˆ Lewis et al. (2020), Guu et al. (2020), Izacard and Grave (2020)์˜ ์›๋ž˜ ์•„์ด๋””์–ด์™€ ์ผ์น˜ํ–ˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
RALM์ด ์ฒ˜์Œ ์ œ์•ˆ๋˜์—ˆ์„ ๋•Œ, ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด ์ด ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ–ˆ๋Š”๋ฐ, ์ด๋Š” Lewis ๋“ฑ(2020), Guu ๋“ฑ(2020), Izacard and Grave(2020)์˜ ์›๋ž˜ ์•„์ด๋””์–ด, ํŠนํžˆ ์ด ๋ฐฉ๋ฒ•์ด ๊ทธ๋“ค์˜ ์›๋ž˜ ์•„์ด๋””์–ด์™€ ์ผ์น˜ํ–ˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค.
LMs for OpenAI. Overall, OpenAI has explored two major approaches to further improving the GPT-3 model, _i.e._, training on code data and alignment with human preference, which are detailed as follows.
OpenAI์˜ LMs. ์ „๋ฐ˜์ ์œผ๋กœ OpenAI๋Š” GPT-3 ๋ชจ๋ธ์„ ๋”์šฑ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ํƒ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฆ‰, ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ•™์Šต๊ณผ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž์ถ˜ ์กฐ์ •์ž…๋‹ˆ๋‹ค. ์ด ๋‘ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ž์„ธํžˆ ์„ค๋ช…๋ฉ๋‹ˆ๋‹ค.
OpenAI์šฉ LM. ์ „๋ฐ˜์ ์œผ๋กœ OpenAI๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ๊ณผ ์ธ๊ฐ„ ์„ ํ˜ธ๋„์— ๋Œ€ํ•œ ์ •๋ ฌ์ด๋ผ๋Š” GPT-3 ๋ชจ๋ธ์„ ๋”์šฑ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ชจ์ƒ‰ํ•ด ์™”์œผ๋ฉฐ, ์ด๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ž์„ธํžˆ ์„ค๋ช…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
Given the classification task definition and the class labels, generate an input that corresponds to each of the class labels. If the task doesn't require input, just generate the correct class label.
๋ถ„๋ฅ˜ ์ž‘์—… ์ •์˜์™€ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”์ด ์ฃผ์–ด์ง€๋ฉด, ๊ฐ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”์— ํ•ด๋‹นํ•˜๋Š” ์ž…๋ ฅ์„ ์ƒ์„ฑํ•˜์‹ญ์‹œ์˜ค. ์ž‘์—…์— ์ž…๋ ฅ์ด ํ•„์š”ํ•˜์ง€ ์•Š์€ ๊ฒฝ์šฐ, ์˜ฌ๋ฐ”๋ฅธ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”๋งŒ ์ƒ์„ฑํ•˜์‹ญ์‹œ์˜ค.
๋ถ„๋ฅ˜ ์ž‘์—… ์ •์˜์™€ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”์ด ์ฃผ์–ด์ง€๋ฉด ๊ฐ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”์— ํ•ด๋‹นํ•˜๋Š” ์ž…๋ ฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ž‘์—…์— ์ž…๋ ฅ์ด ํ•„์š”ํ•˜์ง€ ์•Š์€ ๊ฒฝ์šฐ ์˜ฌ๋ฐ”๋ฅธ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”๋งŒ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
Figure 1: **The performance-vs-loss curves of 1.5B, 6B, and 32B models. Each data point is the loss (\(x\)-axis) and performance (\(y\)-axis) of the intermediate checkpoint of one of the three models.
๊ทธ๋ฆผ 1: 1.5B, 6B, 32B ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ๋Œ€ ์†์‹ค ๊ณก์„ . ๊ฐ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋Š” ์„ธ ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์˜ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์˜ ์†์‹ค (\(x\)-์ถ•)๊ณผ ์„ฑ๋Šฅ (\(y\)-์ถ•)์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1: 1.5B, 6B ๋ฐ 32B ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ๋Œ€ ์†์‹ค ๊ณก์„ . ๊ฐ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋Š” ์„ธ ๊ฐ€์ง€ ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์˜ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์˜ ์†์‹ค(\(x\)-์ถ•)๊ณผ ์„ฑ๋Šฅ(\(y\)-์ถ•)์ž…๋‹ˆ๋‹ค.
Large language models exhibit promising accomplishments, but the main question that arises is how effectively they function and how their performance can be assessed in specific tasks or applications.
๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธ์€ ์œ ๋งํ•œ ์„ฑ๊ณผ๋ฅผ ๋ณด์ด์ง€๋งŒ, ์ฃผ์š”ํ•œ ์งˆ๋ฌธ์€ ์ด๋“ค์ด ํŠน์ • ์ž‘์—…์ด๋‚˜ ์‘์šฉ์—์„œ ์–ผ๋งˆ๋‚˜ ํšจ๊ณผ์ ์œผ๋กœ ์ž‘๋™ํ•˜๋ฉฐ ๊ทธ ์„ฑ๋Šฅ์„ ์–ด๋–ป๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€์ž…๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ์œ ๋งํ•œ ์„ฑ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์ง€๋งŒ, ๊ฐ€์žฅ ์ค‘์š”ํ•œ ์งˆ๋ฌธ์€ ํŠน์ • ์ž‘์—…์ด๋‚˜ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ ์–ผ๋งˆ๋‚˜ ํšจ๊ณผ์ ์œผ๋กœ ์ž‘๋™ํ•˜๊ณ  ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š”์ง€์— ๋Œ€ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Defining \(B_{crit}(L)\)Under the constraint of Equation 3.8, we can derive the critical batch size at \(L\) which minimizes the trade-off between time (\(S/S_{min}\)) and computation (\(E/E_{min}\)):
Equation 3.8์˜ ์ œ์•ฝ ์กฐ๊ฑด ํ•˜์—์„œ, ์‹œ๊ฐ„ (\(S/S_{min}\))๊ณผ ๊ณ„์‚ฐ (\(E/E_{min}\)) ๊ฐ„์˜ ๊ท ํ˜•์„ ์ตœ์†Œํ™”ํ•˜๋Š” \(L\)์—์„œ์˜ ์ž„๊ณ„ ๋ฐฐ์น˜ ํฌ๊ธฐ \(B_{crit}(L)\)๋ฅผ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
B_{crit}(L)\) ์ •์˜ํ•˜๊ธฐ3.8 ๋ฐฉ์ •์‹์˜ ์ œ์•ฝ ์กฐ๊ฑด ํ•˜์—์„œ ์‹œ๊ฐ„(\(S/S_{min}\))๊ณผ ๊ณ„์‚ฐ(\(E/E_{min}\)) ์‚ฌ์ด์˜ ์ ˆ์ถฉ์ ์„ ์ตœ์†Œํ™”ํ•˜๋Š” \(L\)์—์„œ ์ž„๊ณ„ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
In constrast, the storage manager is responsible for the long-term preservation of data, overseeing the storage of information that needs to be retained indefinitely, beyond the active lifespan of any
๋ฐ˜๋ฉด, ์Šคํ† ๋ฆฌ์ง€ ๊ด€๋ฆฌ์ž๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์žฅ๊ธฐ์ ์œผ๋กœ ๋ณด์กดํ•˜๋Š” ์—ญํ• ์„ ๋‹ด๋‹นํ•˜๋ฉฐ, ํ™œ์„ฑ ์ˆ˜๋ช…์„ ๋„˜์–ด์„œ ๋ฌด๊ธฐํ•œ์œผ๋กœ ๋ณด์กดํ•ด์•ผ ํ•˜๋Š” ์ •๋ณด๋ฅผ ๊ด€๋ฆฌํ•ฉ๋‹ˆ๋‹ค.
์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์Šคํ† ๋ฆฌ์ง€ ๊ด€๋ฆฌ์ž๋Š” ๋ฐ์ดํ„ฐ์˜ ์žฅ๊ธฐ ๋ณด์กด์„ ๋‹ด๋‹นํ•˜๋ฉฐ, ํ™œ์„ฑ ์ˆ˜๋ช…์„ ๋„˜์–ด ๋ฌด๊ธฐํ•œ ๋ณด์กดํ•ด์•ผ ํ•˜๋Š” ์ •๋ณด์˜ ์ €์žฅ์„ ๊ฐ๋…ํ•ฉ๋‹ˆ๋‹ค.
Specialized Knowledge TasksThe key message from the results in Table 2 is that synthetic data can not impart 'new knowledge'. It can only help pre-train faster, which was also the premise of our work.
Table 2์˜ ๊ฒฐ๊ณผ์—์„œ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ์ฃผ์š” ๋ฉ”์‹œ์ง€๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ '์ƒˆ๋กœ์šด ์ง€์‹'์„ ์ „๋‹ฌํ•  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๋‹จ์ง€ ๋” ๋น ๋ฅธ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ๋„์šธ ์ˆ˜ ์žˆ์„ ๋ฟ์ด๋ฉฐ, ์ด๋Š” ๋˜ํ•œ ์šฐ๋ฆฌ์˜ ์ž‘์—…์˜ ์ „์ œ์˜€์Šต๋‹ˆ๋‹ค.
์ „๋ฌธ ์ง€์‹ ์ž‘์—…ํ‘œ 2์˜ ๊ฒฐ๊ณผ์—์„œ ์ฃผ์š” ๋ฉ”์‹œ์ง€๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ '์ƒˆ๋กœ์šด ์ง€์‹'์„ ์ „๋‹ฌํ•  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋” ๋น ๋ฅด๊ฒŒ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ๋ฟ์ด๋ฉฐ, ์ด๋Š” ์šฐ๋ฆฌ ์ž‘์—…์˜ ์ „์ œ์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค.
_TL;DR: Our analysis indicates that our human-labeled data has a relatively minimal impact, probably due to a high proportion of tied translations and potential human bias in the evaluation process._
์š”์•ฝ: ์šฐ๋ฆฌ์˜ ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด, ์ธ๊ฐ„์ด ๋ผ๋ฒจ๋งํ•œ ๋ฐ์ดํ„ฐ๋Š” ๋น„๊ต์  ์ตœ์†Œํ•œ์˜ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋ฉฐ, ์ด๋Š” ์•„๋งˆ๋„ ๋™์  ๋ฒˆ์—ญ์˜ ๋†’์€ ๋น„์œจ๊ณผ ํ‰๊ฐ€ ๊ณผ์ •์—์„œ์˜ ์ž ์žฌ์ ์ธ ์ธ๊ฐ„ ํŽธํ–ฅ ๋•Œ๋ฌธ์ผ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
_TL;DR: ๋ถ„์„ ๊ฒฐ๊ณผ, ์ธ๊ฐ„์ด ๋ผ๋ฒจ์„ ๋ถ™์ธ ๋ฐ์ดํ„ฐ์˜ ์˜ํ–ฅ์ด ์ƒ๋Œ€์ ์œผ๋กœ ๋ฏธ๋ฏธํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ, ์ด๋Š” ๋ฌถ์Œ ๋ฒˆ์—ญ์˜ ๋น„์œจ์ด ๋†’๊ณ  ํ‰๊ฐ€ ๊ณผ์ •์—์„œ ์ธ๊ฐ„ ํŽธํ–ฅ์„ฑ์ด ์žˆ์—ˆ๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 5: SFT data scaling curve. Compared with UltraChat and its cleaned version UltraChat 200K, our SFT data demonstrates clear scaling advantages. We attribute its steep slope to the data quality.
Figure 5: SFT ๋ฐ์ดํ„ฐ ์Šค์ผ€์ผ๋ง ๊ณก์„ . UltraChat ๋ฐ ๊ทธ ์ •์ œ๋œ ๋ฒ„์ „์ธ UltraChat 200K์™€ ๋น„๊ตํ–ˆ์„ ๋•Œ, ์šฐ๋ฆฌ์˜ SFT ๋ฐ์ดํ„ฐ๋Š” ๋ช…ํ™•ํ•œ ์Šค์ผ€์ผ๋ง ์ด์ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๋Ÿฌํ•œ ๊ฐ€ํŒŒ๋ฅธ ๊ธฐ์šธ๊ธฐ๋ฅผ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์— ๊ธฐ์ธํ•œ๋‹ค๊ณ  ๋ด…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 5: SFT ๋ฐ์ดํ„ฐ ์Šค์ผ€์ผ๋ง ๊ณก์„ . UltraChat ๋ฐ ์ •๋ฆฌ๋œ ๋ฒ„์ „์ธ UltraChat 200K์™€ ๋น„๊ตํ–ˆ์„ ๋•Œ, ์šฐ๋ฆฌ์˜ SFT ๋ฐ์ดํ„ฐ๋Š” ๋ถ„๋ช…ํ•œ ํ™•์žฅ์„ฑ ์ด์ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ ๊ฐ€ํŒŒ๋ฅธ ๊ธฐ์šธ๊ธฐ๋Š” ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์— ๊ธฐ์ธํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
This section describes our pretraining setup, including our distributed training framework (Section 3.1), optimizer settings (Section 3.2), data preparation (Section 3.3), and hardware (Section 3.4).
์ด ์„น์…˜์—์„œ๋Š” ๋ถ„์‚ฐ ํ•™์Šต ํ”„๋ ˆ์ž„์›Œํฌ(์„น์…˜ 3.1), ์˜ตํ‹ฐ๋งˆ์ด์ € ์„ค์ •(์„น์…˜ 3.2), ๋ฐ์ดํ„ฐ ์ค€๋น„(์„น์…˜ 3.3), ํ•˜๋“œ์›จ์–ด(์„น์…˜ 3.4)๋ฅผ ํฌํ•จํ•œ ์šฐ๋ฆฌ์˜ ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋ถ„์‚ฐ ํ•™์Šต ํ”„๋ ˆ์ž„์›Œํฌ(3.1์ ˆ), ์ตœ์ ํ™” ๋„๊ตฌ ์„ค์ •(3.2์ ˆ), ๋ฐ์ดํ„ฐ ์ค€๋น„(3.3์ ˆ), ํ•˜๋“œ์›จ์–ด(3.4์ ˆ)๋ฅผ ํฌํ•จํ•œ ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
The original draft was polished by OpenAI GPT-4 for grammatical corrections and clarity improvements. We would like to thank our community members for their contributions to our open-source projects.
์›๋ณธ ์ดˆ์•ˆ์€ ๋ฌธ๋ฒ• ์ˆ˜์ • ๋ฐ ๋ช…ํ™•์„ฑ ํ–ฅ์ƒ์„ ์œ„ํ•ด OpenAI GPT-4์— ์˜ํ•ด ๋‹ค๋“ฌ์–ด์กŒ์Šต๋‹ˆ๋‹ค. ์˜คํ”ˆ ์†Œ์Šค ํ”„๋กœ์ ํŠธ์— ๊ธฐ์—ฌํ•ด ์ฃผ์‹  ์ปค๋ฎค๋‹ˆํ‹ฐ ๊ตฌ์„ฑ์›๋“ค๊ป˜ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
์›๋ณธ ์ดˆ์•ˆ์€ ๋ฌธ๋ฒ• ์ˆ˜์ • ๋ฐ ๋ช…ํ™•์„ฑ ๊ฐœ์„ ์„ ์œ„ํ•ด OpenAI GPT-4์—์„œ ๋‹ค๋“ฌ์—ˆ์Šต๋‹ˆ๋‹ค. ์˜คํ”ˆ ์†Œ์Šค ํ”„๋กœ์ ํŠธ์— ๊ธฐ์—ฌํ•ด ์ฃผ์‹  ์ปค๋ฎค๋‹ˆํ‹ฐ ํšŒ์› ์—ฌ๋Ÿฌ๋ถ„๊ป˜ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
- [Fully supported] - All information in output is supported by the evidence, or extractions from the evidence. This is only applicable when the output and part of the evidence are almost identical.
- [์™„์ „ํžˆ ์ง€์›๋จ] - ์ถœ๋ ฅ์˜ ๋ชจ๋“  ์ •๋ณด๊ฐ€ ์ฆ๊ฑฐ ๋˜๋Š” ์ฆ๊ฑฐ์—์„œ ์ถ”์ถœ๋œ ๋‚ด์šฉ์— ์˜ํ•ด ์ง€์›๋ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ถœ๋ ฅ๊ณผ ์ฆ๊ฑฐ์˜ ์ผ๋ถ€๊ฐ€ ๊ฑฐ์˜ ๋™์ผํ•  ๋•Œ๋งŒ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค.
- [์™„์ „ ์ง€์›]- ์ถœ๋ ฅ์˜ ๋ชจ๋“  ์ •๋ณด๋Š” ์ฆ๊ฑฐ ๋˜๋Š” ์ฆ๊ฑฐ์—์„œ ์ถ”์ถœํ•œ ์ •๋ณด์— ์˜ํ•ด ์ง€์›๋ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ถœ๋ ฅ๊ณผ ์ฆ๊ฑฐ์˜ ์ผ๋ถ€๊ฐ€ ๊ฑฐ์˜ ๋™์ผํ•œ ๊ฒฝ์šฐ์—๋งŒ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค.
Figure 2: Histograms depicting the proportion of questions where Adapt-LLM trained on NQ (left) and Adapt-LLM trained on SQuAD (right) ask for extra context for different popularity score intervals.
Figure 2: NQ์—์„œ ํ›ˆ๋ จ๋œ Adapt-LLM(์™ผ์ชฝ)๊ณผ SQuAD์—์„œ ํ›ˆ๋ จ๋œ Adapt-LLM(์˜ค๋ฅธ์ชฝ)์ด ๋‹ค๋ฅธ ์ธ๊ธฐ ์ ์ˆ˜ ๊ตฌ๊ฐ„์—์„œ ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๋ฅผ ์š”์ฒญํ•˜๋Š” ์งˆ๋ฌธ์˜ ๋น„์œจ์„ ๋‚˜ํƒ€๋‚ด๋Š” ํžˆ์Šคํ† ๊ทธ๋žจ.
๊ทธ๋ฆผ 2: NQ๋กœ ํ•™์Šต๋œ Adapt-LLM(์™ผ์ชฝ)๊ณผ SQuAD๋กœ ํ•™์Šต๋œ Adapt-LLM(์˜ค๋ฅธ์ชฝ)์ด ๋‹ค์–‘ํ•œ ์ธ๊ธฐ๋„ ์ ์ˆ˜ ๊ฐ„๊ฒฉ์— ๋Œ€ํ•ด ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๋ฅผ ์š”๊ตฌํ•˜๋Š” ์งˆ๋ฌธ์˜ ๋น„์œจ์„ ๋‚˜ํƒ€๋‚ด๋Š” ํžˆ์Šคํ† ๊ทธ๋žจ์ž…๋‹ˆ๋‹ค.
We can observe that in both Mistral and Vicuna, the intersection with English from other languages is relatively limited, suggesting that English possesses a predominant number of language-specific
Mistral๊ณผ Vicuna ๋ชจ๋‘์—์„œ ๋‹ค๋ฅธ ์–ธ์–ด์™€์˜ ์˜์–ด ๊ต์ง‘ํ•ฉ์ด ์ƒ๋Œ€์ ์œผ๋กœ ์ œํ•œ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์˜์–ด๊ฐ€ ์–ธ์–ด๋ณ„๋กœ ์ง€๋ฐฐ์ ์ธ ์ˆ˜๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
Mistral๊ณผ Vicuna ๋ชจ๋‘์—์„œ ๋‹ค๋ฅธ ์–ธ์–ด์˜ ์˜์–ด์™€์˜ ๊ต์ฐจ์ ์ด ์ƒ๋Œ€์ ์œผ๋กœ ์ œํ•œ๋˜์–ด ์žˆ๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์˜์–ด๊ฐ€ ์–ธ์–ด๋ณ„๋กœ ์šฐ์„ธํ•œ ์ˆ˜๋ฅผ ๋ณด์œ ํ•˜๊ณ  ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
Upon publication, we will make our code and final model checkpoints publicly available. A preliminary version of this work was made available as an ICML 2023 workshop paper in (Gupta et al., 2023).
์ถœํŒ ํ›„, ์šฐ๋ฆฌ๋Š” ์ฝ”๋“œ์™€ ์ตœ์ข… ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๊ณต๊ฐœํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ์ดˆ๊ธฐ ๋ฒ„์ „์€ ICML 2023 ์›Œํฌ์ˆ ๋…ผ๋ฌธ์œผ๋กœ (Gupta et al., 2023) ๊ณต๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
๊ฒŒ์‹œ ์‹œ ์ฝ”๋“œ์™€ ์ตœ์ข… ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ์˜ˆ๋น„ ๋ฒ„์ „์€ (Gupta et al., 2023)์˜ ICML 2023 ์›Œํฌ์ƒต ๋…ผ๋ฌธ์œผ๋กœ ๊ณต๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
The base version of the dataset is self-contained, but the multimodal version is interlaced with links to images-these are not distributed as part of the dataset, and constitute an external source.
๋ฐ์ดํ„ฐ์…‹์˜ ๊ธฐ๋ณธ ๋ฒ„์ „์€ ๋…๋ฆฝ์ ์ด์ง€๋งŒ, ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ฒ„์ „์€ ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ๋งํฌ์™€ ์–ฝํ˜€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ด๋ฏธ์ง€๋Š” ๋ฐ์ดํ„ฐ์…‹์˜ ์ผ๋ถ€๋กœ ๋ฐฐํฌ๋˜์ง€ ์•Š์œผ๋ฉฐ, ์™ธ๋ถ€ ์†Œ์Šค๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ธฐ๋ณธ ๋ฒ„์ „์€ ๋…๋ฆฝ์ ์ด์ง€๋งŒ, ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ฒ„์ „์€ ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ๋งํฌ์™€ ์ธํ„ฐ๋ ˆ์ด์Šค๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ผ๋ถ€๋กœ ๋ฐฐํฌ๋˜์ง€ ์•Š๊ณ  ์™ธ๋ถ€ ์†Œ์Šค๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
Prompting LLMs to generate book summaries:To summarize book-length documents, we adopt the hierarchical merging strategy from (Chang et al., 2023b); see Figure 1 for an illustration of the method.
LLMs์—๊ฒŒ ์ฑ… ์š”์•ฝ์„ ์ƒ์„ฑํ•˜๋„๋ก ์ง€์‹œํ•˜๊ธฐ: ์ฑ… ๊ธธ์ด์˜ ๋ฌธ์„œ๋ฅผ ์š”์•ฝํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” (Chang et al., 2023b)์—์„œ ์ œ์•ˆํ•œ ๊ณ„์ธต์  ๋ณ‘ํ•ฉ ์ „๋žต์„ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์„ค๋ช…์€ Figure 1์„ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค.
LLM์ด ์ฑ… ์š”์•ฝ์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ:์ฑ… ๊ธธ์ด์˜ ๋ฌธ์„œ๋ฅผ ์š”์•ฝํ•˜๊ธฐ ์œ„ํ•ด (Chang et al., 2023b)์˜ ๊ณ„์ธต์  ๋ณ‘ํ•ฉ ์ „๋žต์„ ์ฑ„ํƒํ•˜๋ฉฐ, ์ด ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์˜ˆ๋Š” ๊ทธ๋ฆผ 1์„ ์ฐธ์กฐํ•˜์„ธ์š”.
In addition to the above practical strategies and tricks, existing work has also used other tricks, _e.g.,_ concatenating multiple examples into a single sequence to approach the max length [355].
์œ„์˜ ์‹ค์šฉ์ ์ธ ์ „๋žต๊ณผ ์š”๋ น ์™ธ์—๋„, ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์—ฌ๋Ÿฌ ์˜ˆ์ œ๋ฅผ ํ•˜๋‚˜์˜ ์‹œํ€€์Šค๋กœ ์—ฐ๊ฒฐํ•˜์—ฌ ์ตœ๋Œ€ ๊ธธ์ด์— ์ ‘๊ทผํ•˜๋Š” ๋“ฑ์˜ ๋‹ค๋ฅธ ์š”๋ น๋„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค [355].
์œ„์˜ ์‹ค์ œ ์ „๋žต๊ณผ ํŠธ๋ฆญ ์™ธ์—๋„ ๊ธฐ์กด ์ž‘์—…์—์„œ๋Š” ์ตœ๋Œ€ ๊ธธ์ด์— ์ ‘๊ทผํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์˜ˆ์ œ๋ฅผ ๋‹จ์ผ ์‹œํ€€์Šค๋กœ ์—ฐ๊ฒฐํ•˜๋Š” ๋“ฑ ๋‹ค๋ฅธ ํŠธ๋ฆญ๋„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค[355].
2. Increased endurance: HIIT can improve endurance by increasing the body's ability to use oxygen during exercise. This can help athletes to exercise for longer periods without becoming exhausted.
2. ์ง€๊ตฌ๋ ฅ ํ–ฅ์ƒ: HIIT๋Š” ์šด๋™ ์ค‘ ์‹ ์ฒด๊ฐ€ ์‚ฐ์†Œ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์ฆ๊ฐ€์‹œ์ผœ ์ง€๊ตฌ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์šด๋™์„ ์ˆ˜๋“ค์ด ์ง€์น˜์ง€ ์•Š๊ณ  ๋” ์˜ค๋žœ ์‹œ๊ฐ„ ๋™์•ˆ ์šด๋™ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ค๋‹ˆ๋‹ค.
2. ์ง€๊ตฌ๋ ฅ ํ–ฅ์ƒ: HIIT๋Š” ์šด๋™ ์ค‘ ์‚ฐ์†Œ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์‹ ์ฒด์˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œ์ผœ ์ง€๊ตฌ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์šด๋™์„ ์ˆ˜๊ฐ€ ์ง€์น  ๋•Œ๊นŒ์ง€ ์˜ค๋žซ๋™์•ˆ ์šด๋™ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
where \(c:=1-1/\beta\in(0,1)\), and we have used the elementary fact that \(\sum_{i\geq k+1}i^{-\beta}\asymp k^{-(\beta-1)}\) for large \(k\). For the second sum, we will need the following lemma.
์—ฌ๊ธฐ์„œ \(c:=1-1/\beta\in(0,1)\)์ด๋ฉฐ, ์šฐ๋ฆฌ๋Š” \(k\)๊ฐ€ ํด ๋•Œ \(\sum_{i\geq k+1}i^{-\beta}\asymp k^{-(\beta-1)}\)๋ผ๋Š” ๊ธฐ๋ณธ์ ์ธ ์‚ฌ์‹ค์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ํ•ฉ์„ ์œ„ํ•ด์„œ๋Š” ๋‹ค์Œ์˜ ๋ณด์กฐ์ •๋ฆฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(c:=1-1/\beta\in(0,1)\), ๊ทธ๋ฆฌ๊ณ  ํฐ \(k\)์— ๋Œ€ํ•ด \(\sum_{i\geq k+1}i^{-\beta}\asymp k^{-(\beta-1)}\)๋ผ๋Š” ๊ธฐ๋ณธ์ ์ธ ์‚ฌ์‹ค์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ํ•ฉ์˜ ๊ฒฝ์šฐ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ ˆ๋งˆ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
For both downstream and perplexity evaluation, we use our fixed evaluation pipeline to compare results against publicly available models. We also report a separate evaluation of our adapted model.
๋‹ค์šด์ŠคํŠธ๋ฆผ ํ‰๊ฐ€์™€ ํผํ”Œ๋ ‰์‹œํ‹ฐ ํ‰๊ฐ€ ๋ชจ๋‘์—์„œ, ์šฐ๋ฆฌ๋Š” ๊ณ ์ •๋œ ํ‰๊ฐ€ ํŒŒ์ดํ”„๋ผ์ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐœ๋œ ๋ชจ๋ธ๋“ค๊ณผ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ ์‘๋œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๋ณ„๋„์˜ ํ‰๊ฐ€๋„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.
๋‹ค์šด์ŠคํŠธ๋ฆผ ๋ฐ ๋‹นํ™ฉ์Šค๋Ÿฌ์›€ ํ‰๊ฐ€ ๋ชจ๋‘์— ๋Œ€ํ•ด ๊ณ ์ •๋œ ํ‰๊ฐ€ ํŒŒ์ดํ”„๋ผ์ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ๊ณผ์˜ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•œ๋‹ค. ๋˜ํ•œ ์šฐ๋ฆฌ๋Š” ์šฐ๋ฆฌ์˜ ์ ์‘๋œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๋ณ„๋„์˜ ํ‰๊ฐ€๋ฅผ ๋ณด๊ณ ํ•œ๋‹ค.
##There are certain important aspects of college preparation. The student preparing for colleges should define their goals and interests. The student should also plan financially and academically.
## ๋Œ€ํ•™ ์ค€๋น„์—๋Š” ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์ธก๋ฉด์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ•™์„ ์ค€๋น„ํ•˜๋Š” ํ•™์ƒ์€ ์ž์‹ ์˜ ๋ชฉํ‘œ์™€ ๊ด€์‹ฌ์‚ฌ๋ฅผ ์ •์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•™์ƒ์€ ์žฌ์ •์  ๋ฐ ํ•™๋ฌธ์ ์œผ๋กœ ๊ณ„ํš์„ ์„ธ์›Œ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
## ๋Œ€ํ•™ ์ค€๋น„์—๋Š” ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์ธก๋ฉด์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ•™์„ ์ค€๋น„ํ•˜๋Š” ํ•™์ƒ์€ ์ž์‹ ์˜ ๋ชฉํ‘œ์™€ ๊ด€์‹ฌ์‚ฌ๋ฅผ ์ •์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•™์ƒ์€ ์žฌ์ •์ ์œผ๋กœ๋‚˜ ํ•™์—…์ ์œผ๋กœ ๊ณ„ํš์„ ์„ธ์›Œ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
However, there are also some risks associated with HIIT, including the risk of injury, especially if the athlete is not properly warmed up or does not have adequate recovery time between workouts.
๊ทธ๋Ÿฌ๋‚˜ HIIT์—๋Š” ๋ถ€์ƒ์˜ ์œ„ํ—˜์„ ํฌํ•จํ•œ ๋ช‡ ๊ฐ€์ง€ ์œ„ํ—˜์ด ์žˆ์œผ๋ฉฐ, ํŠนํžˆ ์šด๋™ ์„ ์ˆ˜๊ฐ€ ์ œ๋Œ€๋กœ ์ค€๋น„ ์šด๋™์„ ํ•˜์ง€ ์•Š์•˜๊ฑฐ๋‚˜ ์šด๋™ ์‚ฌ์ด์— ์ถฉ๋ถ„ํ•œ ํšŒ๋ณต ์‹œ๊ฐ„์„ ๊ฐ€์ง€์ง€ ์•Š์€ ๊ฒฝ์šฐ์— ๊ทธ๋Ÿฌํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ HIIT์—๋Š” ๋ถ€์ƒ ์œ„ํ—˜, ํŠนํžˆ ์šด๋™ ์„ ์ˆ˜๊ฐ€ ์ œ๋Œ€๋กœ ์›Œ๋ฐ์—…์„ ํ•˜์ง€ ์•Š์•˜๊ฑฐ๋‚˜ ์šด๋™ ์‚ฌ์ด์— ์ถฉ๋ถ„ํ•œ ํšŒ๋ณต ์‹œ๊ฐ„์ด ์—†๋Š” ๊ฒฝ์šฐ ๋ถ€์ƒ ์œ„ํ—˜ ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ์œ„ํ—˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
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.
ALFWorld์˜ _์„ฑ๊ณต๋ฅ _๊ณผ WebShop์˜ _ํ‰๊ท  ์ ์ˆ˜/์„ฑ๊ณต๋ฅ _์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์˜ ๊ธธ์ด๋ฅผ ์ค„์ด๊ณ  EOS ํ† ํฐ์œผ๋กœ ์ค„ ๋ฐ”๊ฟˆ์„ ์‚ฌ์šฉํ•˜๋Š” ReAct [449]๋ฅผ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค.
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.
To address these problems and improve encoding efficiency, we propose to extend LLaMA vocabulary with additional Chinese tokens and adapt the model for the extended vocabulary (Yang et al., 2022).
์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ  ์ธ์ฝ”๋”ฉ ํšจ์œจ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” LLaMA ์–ดํœ˜์— ์ถ”๊ฐ€์ ์ธ ์ค‘๊ตญ์–ด ํ† ํฐ์„ ํ™•์žฅํ•˜๊ณ  ํ™•์žฅ๋œ ์–ดํœ˜์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์กฐ์ •ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค (Yang et al., 2022).
์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ  ์ธ์ฝ”๋”ฉ ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ์ค‘๊ตญ์–ด ํ† ํฐ์œผ๋กœ LLaMA ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜๊ณ  ํ™•์žฅ๋œ ์–ดํœ˜์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์กฐ์ •ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•œ๋‹ค(Yang et al., 2022).
Clinical history-taking and diagnostic dialogue through which clinicians derive diagnosis and management plans represent a complex skill [19] whose optimal conduct is highly dependent on context.
์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™”๋Š” ์ž„์ƒ์˜๊ฐ€ ์ง„๋‹จ ๋ฐ ๊ด€๋ฆฌ ๊ณ„ํš์„ ๋„์ถœํ•˜๋Š” ๋ณต์žกํ•œ ๊ธฐ์ˆ ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ๊ทธ ์ตœ์ ์˜ ์ˆ˜ํ–‰์€ ์ƒํ™ฉ์— ํฌ๊ฒŒ ์˜์กดํ•ฉ๋‹ˆ๋‹ค [19].
์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ž„์ƒ์˜๊ฐ€ ์ง„๋‹จ ๋ฐ ๊ด€๋ฆฌ ๊ณ„ํš์„ ๋„์ถœํ•˜๋Š” ์ง„๋‹จ ๋Œ€ํ™”๋Š” ์ตœ์ ์˜ ์ˆ˜ํ–‰์ด ์ƒํ™ฉ์— ๋”ฐ๋ผ ํฌ๊ฒŒ ๋‹ฌ๋ผ์ง€๋Š” ๋ณต์žกํ•œ ๊ธฐ์ˆ  [19]์„ ๋‚˜ํƒ€๋‚ธ๋‹ค.
FIRST provide a one-sentence comparison of the two responses and explain \ which you feel is more helpful. SECOND, on a new line, state only "A" or "B" to indicate which response is more helpful.
๋จผ์ € ๋‘ ์‘๋‹ต์„ ํ•œ ๋ฌธ์žฅ์œผ๋กœ ๋น„๊ตํ•˜๊ณ , ์–ด๋–ค ์‘๋‹ต์ด ๋” ๋„์›€์ด ๋˜๋Š”์ง€ ์„ค๋ช…ํ•˜์„ธ์š”. ๋‘ ๋ฒˆ์งธ๋กœ, ์ƒˆ๋กœ์šด ์ค„์— "A" ๋˜๋Š” "B"๋งŒ ์ ์–ด ์–ด๋–ค ์‘๋‹ต์ด ๋” ๋„์›€์ด ๋˜๋Š”์ง€ ํ‘œ์‹œํ•˜์„ธ์š”.
๋จผ์ € ๋‘ ๋‹ต๋ณ€์„ ํ•œ ๋ฌธ์žฅ์œผ๋กœ ๋น„๊ตํ•˜๊ณ  ์–ด๋–ค ๋‹ต๋ณ€์ด ๋” ๋„์›€์ด ๋œ๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋Š”์ง€ ์„ค๋ช…ํ•˜์„ธ์š”. ๋‘ ๋ฒˆ์งธ๋กœ, ์ƒˆ ์ค„์— "A" ๋˜๋Š” "B"๋ฅผ ํ‘œ์‹œํ•˜์—ฌ ์–ด๋–ค ๋‹ต๋ณ€์ด ๋” ๋„์›€์ด ๋˜๋Š”์ง€ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.
Figure 3: The impact of mask ratio when training on Llama2-7B with GSM8K. We compare MFT of two settings: Fixed mask ratio without warm-up, and a linear warm-up of the mask ratio starting from 0.
Figure 3: GSM8K๋กœ Llama2-7B๋ฅผ ํ›ˆ๋ จํ•  ๋•Œ ๋งˆ์Šคํฌ ๋น„์œจ์˜ ์˜ํ–ฅ. ์šฐ๋ฆฌ๋Š” ๋‘ ๊ฐ€์ง€ ์„ค์ •์˜ MFT๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค: ์›Œ๋ฐ์—… ์—†์ด ๊ณ ์ •๋œ ๋งˆ์Šคํฌ ๋น„์œจ๊ณผ 0์—์„œ ์‹œ์ž‘ํ•˜๋Š” ๋งˆ์Šคํฌ ๋น„์œจ์˜ ์„ ํ˜• ์›Œ๋ฐ์—….
๊ทธ๋ฆผ 3: GSM8K๋กœ Llama2-7B๋ฅผ ํ›ˆ๋ จํ•  ๋•Œ ๋งˆ์Šคํฌ ๋น„์œจ์˜ ์˜ํ–ฅ. ๋‘ ์„ค์ •์˜ MFT๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค: ์›Œ๋ฐ์—…์ด ์—†๋Š” ๊ณ ์ • ๋งˆ์Šคํฌ ๋น„์œจ๊ณผ 0์—์„œ ์‹œ์ž‘ํ•˜๋Š” ๋งˆ์Šคํฌ ๋น„์œจ์˜ ์„ ํ˜• ์›Œ๋ฐ์—…์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค.
EvaluationWe consider two baselines in our experiments: LLaVA-1.6-Mistral-7B [31], one of our source models, and Japanese Stable VLM [43] a Japanese VLM trained from scratch on Japanese datasets.
ํ‰๊ฐ€ ์‹คํ—˜์—์„œ๋Š” ๋‘ ๊ฐ€์ง€ ๊ธฐ์ค€ ๋ชจ๋ธ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค: ์†Œ์Šค ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์ธ LLaVA-1.6-Mistral-7B [31]์™€ ์ผ๋ณธ์–ด ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ•™์Šต๋œ ์ผ๋ณธ์–ด VLM์ธ Japanese Stable VLM [43]์ž…๋‹ˆ๋‹ค.
ํ‰๊ฐ€์‹คํ—˜์—์„œ๋Š” ์†Œ์Šค ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์ธ LLaVA-1.6-Mistral-7B [31]์™€ ์ผ๋ณธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ•™์Šต๋œ ์ผ๋ณธ์–ด VLM์ธ ์ผ๋ณธ์–ด Stable VLM [43]์˜ ๋‘ ๊ฐ€์ง€ ๋ฒ ์ด์Šค๋ผ์ธ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค.
Figure 7: Percentage of the most frequent entities for each fact composition type of TwoHopFact. The expanded forms of the abbreviations used for the fact composition types are listed in Table 4.
Figure 7: TwoHopFact์˜ ๊ฐ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์— ๋Œ€ํ•œ ๊ฐ€์žฅ ๋นˆ๋ฒˆํ•œ ์—”ํ‹ฐํ‹ฐ์˜ ๋น„์œจ. ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์— ์‚ฌ์šฉ๋œ ์•ฝ์–ด์˜ ํ™•์žฅํ˜•์€ Table 4์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 7: TwoHopFact์˜ ๊ฐ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์— ๋Œ€ํ•œ ๊ฐ€์žฅ ๋นˆ๋ฒˆํ•œ ์—”ํ‹ฐํ‹ฐ์˜ ๋ฐฑ๋ถ„์œจ. ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์— ์‚ฌ์šฉ๋œ ์•ฝ์–ด์˜ ํ™•์žฅ๋œ ํ˜•ํƒœ๋Š” ํ‘œ 4์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
In this experiment, our objective is to once again evaluate the effectiveness of the Adapt-LLM model, this time focusing on its ability to accurately determine when additional context is needed.
์ด๋ฒˆ ์‹คํ—˜์—์„œ ์šฐ๋ฆฌ์˜ ๋ชฉํ‘œ๋Š” ์ถ”๊ฐ€์ ์ธ ๋ฌธ๋งฅ์ด ํ•„์š”ํ•  ๋•Œ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ํŒ๋‹จํ•˜๋Š” Adapt-LLM ๋ชจ๋ธ์˜ ํšจ๊ณผ๋ฅผ ๋‹ค์‹œ ํ•œ ๋ฒˆ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
์ด๋ฒˆ ์‹คํ—˜์—์„œ๋Š” ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๊ฐ€ ํ•„์š”ํ•œ ์‹œ๊ธฐ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ํŒ๋‹จํ•˜๋Š” ๋Šฅ๋ ฅ์— ์ดˆ์ ์„ ๋งž์ถฐ Adapt-LLM ๋ชจ๋ธ์˜ ํšจ๊ณผ๋ฅผ ๋‹ค์‹œ ํ•œ ๋ฒˆ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
_Remark 3.4_ (why "sum of 3").: It is essential to obtain a lower bound that is the _sum_ of the three components; neglecting any may result in a suboptimal bound (see examples in Appendix A.4).
_์ฃผ์„ 3.4_ (์™œ "3์˜ ํ•ฉ"์ธ๊ฐ€).: ์„ธ ๊ฐ€์ง€ ๊ตฌ์„ฑ ์š”์†Œ์˜ _ํ•ฉ_์ธ ํ•˜ํ•œ์„ ์–ป๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์–ด๋Š ํ•˜๋‚˜๋ผ๋„ ๋ฌด์‹œํ•˜๋ฉด ์ตœ์  ์ดํ•˜์˜ ๊ฒฝ๊ณ„๊ฐ€ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค (๋ถ€๋ก A.4์˜ ์˜ˆ์‹œ๋ฅผ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค).
_Remark 3.4_ ("3์˜ ํ•ฉ"์ด์œ ): ์„ธ ๊ฐ€์ง€ ๊ตฌ์„ฑ ์š”์†Œ์˜ _ํ•ฉ๊ณ„_์ธ ํ•˜ํ•œ์„ ์–ป๋Š” ๊ฒƒ์ด ํ•„์ˆ˜์ ์ด๋ฉฐ, ์–ด๋Š ํ•˜๋‚˜๋ผ๋„ ๋ฌด์‹œํ•˜๋ฉด ์ตœ์ ์ด ์•„๋‹Œ ๋ฐ”์šด๋“œ๊ฐ€ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๋ถ€๋ก A.4์˜ ์˜ˆ ์ฐธ์กฐ).
We hope future work can leverage CodecLM as a flexible data synthesis framework for LLM alignment, so that advances in the field can be integrated into CodecLM to reduce its current limitations.
์šฐ๋ฆฌ๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ๊ฐ€ CodecLM์„ LLM ์ •๋ ฌ์„ ์œ„ํ•œ ์œ ์—ฐํ•œ ๋ฐ์ดํ„ฐ ํ•ฉ์„ฑ ํ”„๋ ˆ์ž„์›Œํฌ๋กœ ํ™œ์šฉํ•˜์—ฌ, ์ด ๋ถ„์•ผ์˜ ๋ฐœ์ „์ด CodecLM์— ํ†ตํ•ฉ๋˜์–ด ํ˜„์žฌ์˜ ํ•œ๊ณ„๋ฅผ ์ค„์ผ ์ˆ˜ ์žˆ๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ CodecLM์„ LLM ์ •๋ ฌ์„ ์œ„ํ•œ ์œ ์—ฐํ•œ ๋ฐ์ดํ„ฐ ํ•ฉ์„ฑ ํ”„๋ ˆ์ž„์›Œํฌ๋กœ ํ™œ์šฉํ•˜์—ฌ ์ด ๋ถ„์•ผ์˜ ๋ฐœ์ „์„ CodecLM์— ํ†ตํ•ฉํ•˜์—ฌ ํ˜„์žฌ์˜ ํ•œ๊ณ„๋ฅผ ์ค„์ผ ์ˆ˜ ์žˆ๊ธฐ๋ฅผ ๊ธฐ๋Œ€ํ•œ๋‹ค.
Here, \(W_{K}\in\mathbb{R}^{d_{model}\times d_{bay}}\), \(W_{V}\in\mathbb{R}^{d_{model}\times d_{value}}\) and \(W_{Q}\in\mathbb{R}^{d_{model}\times d_{bay}}\) are trainable projection matrices.
์—ฌ๊ธฐ์„œ \(W_{K}\in\mathbb{R}^{d_{model}\times d_{bay}}\), \(W_{V}\in\mathbb{R}^{d_{model}\times d_{value}}\) ๋ฐ \(W_{Q}\in\mathbb{R}^{d_{model}\times d_{bay}}\)๋Š” ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํˆฌ์˜ ํ–‰๋ ฌ์ž…๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(W_{K}\in\mathbb{R}^{d_{model}\times d_{bay}}\), \(W_{V}\in\mathbb{R}^{d_{model}\times d_{value}}\), \(W_{Q}\in\mathbb{R}^{d_{model}\times d_{bay}}\)๋Š” ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํˆฌ์˜ ํ–‰๋ ฌ์ž…๋‹ˆ๋‹ค.
Figure 3: Average compute (GFLOPS) required to process a single word within (a) multilingual, (b) English, and (c) German documents within a full **training** pass (including the backward pass).
Figure 3: ์ „์ฒด **ํ›ˆ๋ จ** ๊ณผ์ •(์—ญ์ „ํŒŒ ํฌํ•จ)์—์„œ (a) ๋‹ค๊ตญ์–ด, (b) ์˜์–ด, (c) ๋…์ผ์–ด ๋ฌธ์„œ ๋‚ด ๋‹จ์–ด ํ•˜๋‚˜๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ํ‰๊ท  ์—ฐ์‚ฐ๋Ÿ‰(GFLOPS).
๊ทธ๋ฆผ 3: ์ „์ฒด **ํ›ˆ๋ จ** ํŒจ์Šค(์—ญ๋ฐฉํ–ฅ ํŒจ์Šค ํฌํ•จ) ๋‚ด์—์„œ (a) ๋‹ค๊ตญ์–ด, (b) ์˜์–ด ๋ฐ (c) ๋…์ผ์–ด ๋ฌธ์„œ ๋‚ด์—์„œ ๋‹จ์ผ ๋‹จ์–ด๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ํ‰๊ท  ๊ณ„์‚ฐ(GFLOPS)์ž…๋‹ˆ๋‹ค.
Release requirementsTo ensure the integrity of test results, model publishers (i.e., organizations who make SUTs available) commit to adhering to the following rules, which may change over time:
์ถœ์‹œ ์š”๊ตฌ ์‚ฌํ•ญ: ํ…Œ์ŠคํŠธ ๊ฒฐ๊ณผ์˜ ๋ฌด๊ฒฐ์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ ๊ฒŒ์‹œ์ž(์ฆ‰, SUT๋ฅผ ์ œ๊ณตํ•˜๋Š” ์กฐ์ง)๋Š” ๋‹ค์Œ ๊ทœ์น™์„ ์ค€์ˆ˜ํ•  ๊ฒƒ์„ ์•ฝ์†ํ•˜๋ฉฐ, ์ด ๊ทœ์น™์€ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ๋ณ€๊ฒฝ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
๋ฆด๋ฆฌ์Šค ์š”๊ตฌ ์‚ฌํ•ญํ…Œ์ŠคํŠธ ๊ฒฐ๊ณผ์˜ ๋ฌด๊ฒฐ์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ ๊ฒŒ์‹œ์ž(์˜ˆ: SUT์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ์กฐ์ง)๋Š” ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ๋ณ€๊ฒฝ๋  ์ˆ˜ ์žˆ๋Š” ๋‹ค์Œ ๊ทœ์น™์„ ์ค€์ˆ˜ํ•˜๊ธฐ๋กœ ์•ฝ์ •ํ•ฉ๋‹ˆ๋‹ค:
Footnote 20: The modelโ€™s ability to learn from junk data is negligible; each person in bioS(\(N^{\prime}\)) appears only 0.2 times during training when \(N=200k\), or 0.05 times when \(N=50k\).
Footnote 20: ๋ชจ๋ธ์ด ๋ถˆํ•„์š”ํ•œ ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๋Š” ๋Šฅ๋ ฅ์€ ๋ฌด์‹œํ•  ๋งŒํ•ฉ๋‹ˆ๋‹ค. bioS(\(N^{\prime}\))์˜ ๊ฐ ์‚ฌ๋žŒ์€ \(N=200k\)์ผ ๋•Œ ํ›ˆ๋ จ ์ค‘์— ๋‹จ 0.2๋ฒˆ, \(N=50k\)์ผ ๋•Œ๋Š” 0.05๋ฒˆ๋งŒ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.
๊ฐ์ฃผ 20: ๋ชจ๋ธ์ด ์ •ํฌ ๋ฐ์ดํ„ฐ์—์„œ ํ•™์Šตํ•˜๋Š” ๋Šฅ๋ ฅ์€ ๋ฏธ๋ฏธํ•œ ์ˆ˜์ค€์ด๋ฉฐ, bioS(\(N^{\ํ”„๋ผ์ž„}\))์˜ ๊ฐ ์‚ฌ๋žŒ์€ \(N=200k\)์ผ ๋•Œ ํ›ˆ๋ จ ์ค‘ 0.2ํšŒ, \(N=50k\)์ผ ๋•Œ 0.05ํšŒ๋งŒ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.
Generation TaskDeactivation methods: (i) randomly sampled neurons in the generating layers. (ii) randomly sampled neurons in all layers. (iv) language-specific neurons in the generating layers.
์ƒ์„ฑ ์ž‘์—… ๋น„ํ™œ์„ฑํ™” ๋ฐฉ๋ฒ•: (i) ์ƒ์„ฑ ๋ ˆ์ด์–ด์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋‰ด๋Ÿฐ. (ii) ๋ชจ๋“  ๋ ˆ์ด์–ด์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋‰ด๋Ÿฐ. (iv) ์ƒ์„ฑ ๋ ˆ์ด์–ด์—์„œ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ.
์„ธ๋Œ€ ์ž‘์—…๋น„ํ™œ์„ฑํ™” ๋ฐฉ๋ฒ•: (i) ์ƒ์„ฑ ๋ ˆ์ด์–ด์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋‰ด๋Ÿฐ. (ii) ๋ชจ๋“  ๋ ˆ์ด์–ด์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋‰ด๋Ÿฐ. (iv) ์ƒ์„ฑ ๋ ˆ์ด์–ด์˜ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ.
Figure 9: \(|\) R-GSM example where the original problem can be correctly solved by all LLMs, but GPT-3.5-Turbo fails on the reordered version while all the other LLMs still solve it correctly.
Figure 9: ์›๋ž˜ ๋ฌธ์ œ๋Š” ๋ชจ๋“  LLM์ด ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์žฌ์ •๋ ฌ๋œ ๋ฒ„์ „์—์„œ๋Š” GPT-3.5-Turbo๊ฐ€ ์‹คํŒจํ•˜๊ณ  ๋‹ค๋ฅธ ๋ชจ๋“  LLM์€ ์—ฌ์ „ํžˆ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ•ด๊ฒฐํ•˜๋Š” R-GSM ์˜ˆ์ œ์ž…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 9: ์›๋ž˜ ๋ฌธ์ œ๋Š” ๋ชจ๋“  LLM์ด ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์ˆœ์„œ๊ฐ€ ๋ฐ”๋€ ๋ฒ„์ „์—์„œ๋Š” GPT-3.5-Turbo๊ฐ€ ์‹คํŒจํ•˜๋Š” ๋ฐ˜๋ฉด ๋‹ค๋ฅธ ๋ชจ๋“  LLM์€ ์—ฌ์ „ํžˆ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ•ด๊ฒฐํ•˜๋Š” \(|\) R-GSM ์˜ˆ์ œ.
**Reading comprehension:** Reading a given article or paragraph and answering questions about it. We use RACE [23] for English and C3 [42] for Chinese. Both are based on multi-choice questions.
**๋…ํ•ด:** ์ฃผ์–ด์ง„ ๊ธฐ์‚ฌ๋‚˜ ๋‹จ๋ฝ์„ ์ฝ๊ณ  ์ด์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์˜์–ด๋ฅผ ์œ„ํ•ด RACE [23]๋ฅผ, ์ค‘๊ตญ์–ด๋ฅผ ์œ„ํ•ด C3 [42]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋‘˜ ๋‹ค ๋‹ค์ง€์„ ๋‹คํ˜• ์งˆ๋ฌธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
**๋…ํ•ด๋ ฅ: ์ฃผ์–ด์ง„ ๊ธฐ์‚ฌ๋‚˜ ๋‹จ๋ฝ์„ ์ฝ๊ณ  ๊ทธ์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. ์˜์–ด๋Š” RACE [23], ์ค‘๊ตญ์–ด๋Š” C3 [42]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋‘˜ ๋‹ค ๊ฐ๊ด€์‹ ๋ฌธ์ œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
**Evaluation of MLLM.** After introducing the approaches to developing MLLMs, we further discuss how to effectively assess the multimodal capabilities of MLLMs from the following three aspects.
MLLM์˜ ํ‰๊ฐ€. MLLM์„ ๊ฐœ๋ฐœํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ์†Œ๊ฐœํ•œ ํ›„, ๋‹ค์Œ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์—์„œ MLLM์˜ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๊ธฐ๋Šฅ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค.
**MLLM์˜ ํ‰๊ฐ€**. MLLM ๊ฐœ๋ฐœ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์†Œ๊ฐœํ•œ ํ›„, ๋‹ค์Œ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์—์„œ MLLM์˜ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๊ธฐ๋Šฅ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์ž์„ธํžˆ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค.
Table 2 provides insights into the characteristics of the three datasets involved in our experimental procedure, including the total number of questions and the average number of words per ques
Table 2๋Š” ์‹คํ—˜ ์ ˆ์ฐจ์— ํฌํ•จ๋œ ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ์…‹์˜ ํŠน์„ฑ์— ๋Œ€ํ•œ ํ†ต์ฐฐ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ด ์งˆ๋ฌธ ์ˆ˜์™€ ์งˆ๋ฌธ๋‹น ํ‰๊ท  ๋‹จ์–ด ์ˆ˜๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
ํ‘œ 2๋Š” ์ด ์งˆ๋ฌธ ์ˆ˜์™€ ์งˆ๋ฌธ๋‹น ํ‰๊ท  ๋‹จ์–ด ์ˆ˜๋ฅผ ํฌํ•จํ•˜์—ฌ ์‹คํ—˜ ์ ˆ์ฐจ์— ํฌํ•จ๋œ ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํŠน์„ฑ์— ๋Œ€ํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
2020; Wu et al., 2021; Xiao et al., 2021; Li et al., 2022; Gupta and Gupta, 2022) is more challenging as it introduces additional temporal information compared to visual multi-modal reasoning.
2020; Wu et al., 2021; Xiao et al., 2021; Li et al., 2022; Gupta and Gupta, 2022)๋Š” ์‹œ๊ฐ์  ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์ถ”๋ก ์— ๋น„ํ•ด ์ถ”๊ฐ€์ ์ธ ์‹œ๊ฐ„ ์ •๋ณด๋ฅผ ๋„์ž…ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋” ๋„์ „์ ์ž…๋‹ˆ๋‹ค.
์‹œ๊ฐ์  ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์— ๋น„ํ•ด ์ถ”๊ฐ€์ ์ธ ์‹œ๊ฐ„์  ์ •๋ณด๋ฅผ ๋„์ž…ํ•˜๊ธฐ ๋•Œ๋ฌธ์— 2020; Wu et al., 2021; Xiao et al., 2021; Li et al., 2022; Gupta and Gupta, 2022)๋Š” ๋” ์–ด๋ ต๋‹ค.
In this part, we present discussions regarding two fundamental questions related to CoT prompting, _i.e._, _"when does CoT prompting work for LLMs"_ and _"why can LLMs perform CoT reasoning"_.
์ด ๋ถ€๋ถ„์—์„œ๋Š” CoT ํ”„๋กฌํ”„ํŠธ์™€ ๊ด€๋ จ๋œ ๋‘ ๊ฐ€์ง€ ๊ทผ๋ณธ์ ์ธ ์งˆ๋ฌธ, ์ฆ‰ "LLMs์—์„œ CoT ํ”„๋กฌํ”„ํŠธ๊ฐ€ ์–ธ์ œ ์ž‘๋™ํ•˜๋Š”๊ฐ€"์™€ "์™œ LLMs๊ฐ€ CoT ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€"์— ๋Œ€ํ•œ ๋…ผ์˜๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ถ€๋ถ„์—์„œ๋Š” CoT ํ”„๋กฌํ”„ํŠธ์™€ ๊ด€๋ จ๋œ ๋‘ ๊ฐ€์ง€ ๊ทผ๋ณธ์ ์ธ ์งˆ๋ฌธ, ์ฆ‰ _"์–ธ์ œ CoT ํ”„๋กฌํ”„ํŠธ๊ฐ€ LLM์— ์ž‘๋™ํ•˜๋Š”๊ฐ€"_์™€ _"์™œ LLM์ด CoT ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€"_์— ๋Œ€ํ•œ ๋…ผ์˜๋ฅผ ์ œ์‹œํ•œ๋‹ค.
**S&P**Plausible but only partially supported. The Reference [1] does not explicitly mention the person is born in Moscow and there is no description about the Russian Guild of Film Producers.
S&P: ๊ทธ๋Ÿด๋“ฏํ•˜์ง€๋งŒ ๋ถ€๋ถ„์ ์œผ๋กœ๋งŒ ์ง€์›๋ฉ๋‹ˆ๋‹ค. ์ฐธ๊ณ  ๋ฌธํ—Œ [1]์—๋Š” ๊ทธ ์‚ฌ๋žŒ์ด ๋ชจ์Šคํฌ๋ฐ”์—์„œ ํƒœ์–ด๋‚ฌ๋‹ค๋Š” ๋ช…์‹œ์ ์ธ ์–ธ๊ธ‰์ด ์—†์œผ๋ฉฐ, ๋Ÿฌ์‹œ์•„ ์˜ํ™” ์ œ์ž‘์ž ๊ธธ๋“œ์— ๋Œ€ํ•œ ์„ค๋ช…๋„ ์—†์Šต๋‹ˆ๋‹ค.
**S&P**๊ทธ๋Ÿด๋“ฏํ•˜์ง€๋งŒ ๋ถ€๋ถ„์ ์œผ๋กœ๋งŒ ๋’ท๋ฐ›์นจ๋ฉ๋‹ˆ๋‹ค. ์ฐธ์กฐ [1]์—๋Š” ํ•ด๋‹น ์ธ๋ฌผ์ด ๋ชจ์Šคํฌ๋ฐ”์—์„œ ํƒœ์–ด๋‚ฌ๋‹ค๋Š” ์–ธ๊ธ‰์ด ๋ช…์‹œ์ ์œผ๋กœ ์—†์œผ๋ฉฐ ๋Ÿฌ์‹œ์•„ ์˜ํ™” ์ œ์ž‘์ž ๊ธธ๋“œ์— ๋Œ€ํ•œ ์„ค๋ช…์ด ์—†์Šต๋‹ˆ๋‹ค.
Our web documents in the corpora consist of Oscars1Abadji et al. (2021), that were generated by the ungoliant pipeline2 based on three Common Crawl WET Archives (2022-27, 2022-49 and 2023-14).
์šฐ๋ฆฌ ๋ง๋ญ‰์น˜์˜ ์›น ๋ฌธ์„œ๋Š” Oscars1Abadji et al. (2021)๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์„ธ ๊ฐœ์˜ Common Crawl WET Archives (2022-27, 2022-49, 2023-14)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ungoliant ํŒŒ์ดํ”„๋ผ์ธ2์— ์˜ํ•ด ์ƒ์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
๋ง๋ญ‰์น˜์˜ ์›น ๋ฌธ์„œ๋Š” ์„ธ ๊ฐœ์˜ Common Crawl WET ์•„์นด์ด๋ธŒ(2022-27, 2022-49 ๋ฐ 2023-14)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ungoliant ํŒŒ์ดํ”„๋ผ์ธ2์— ์˜ํ•ด ์ƒ์„ฑ๋œ Oscars1Abadji ๋“ฑ(2021)์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.
Equal contribution. See author contributions for details. Work done during their internships at Microsoft Research Asia. EUR [email protected]; [email protected] authors.
๋™์ผํ•œ ๊ธฐ์—ฌ. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ์ €์ž ๊ธฐ์—ฌ๋„๋ฅผ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค. Microsoft Research Asia์—์„œ ์ธํ„ด์‹ญ ๋™์•ˆ ์ˆ˜ํ–‰๋œ ์ž‘์—…์ž…๋‹ˆ๋‹ค. EUR [email protected]; [email protected] ์—ฐ๋ฝ ์ €์ž.
๋™๋“ฑํ•œ ๊ธฐ์—ฌ. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ์ €์ž ๊ธฐ์—ฌ๋„๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. Microsoft Research Asia์—์„œ ์ธํ„ด์œผ๋กœ ๊ทผ๋ฌดํ•œ ๊ธฐ๊ฐ„ ๋™์•ˆ ์ˆ˜ํ–‰ํ•œ ์ž‘์—…. EUR [email protected]; [email protected]๊ต์‹  ์ €์ž.
After the death of Jurning mother, lost and Mrs. Price grow slower and decides to nearly adults inside early experiences about lost and Mrs. Price wedding but later becomes very at this point.
Jurning ์–ด๋จธ๋‹ˆ์˜ ์ฃฝ์Œ ์ดํ›„, ์žƒ์–ด๋ฒ„๋ฆฐ ๊ฒƒ๊ณผ ํ”„๋ผ์ด์Šค ๋ถ€์ธ์€ ์ ์  ๋Š๋ ค์ง€๊ณ  ๊ฑฐ์˜ ์„ฑ์ธ์ด ๋˜๊ธฐ ์ „์˜ ์ดˆ๊ธฐ ๊ฒฝํ—˜์— ๋Œ€ํ•ด ๊ฒฐํ˜ผ์„ ์ƒ๊ฐํ•˜์ง€๋งŒ ๋‚˜์ค‘์—๋Š” ๋งค์šฐ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค.
์ฃผ๋‹ ์–ด๋จธ๋‹ˆ๊ฐ€ ์ฃฝ์€ ํ›„, ์žƒ์–ด๋ฒ„๋ฆฐ ๊ฐ€๊ฒฉ ๋ถ€์ธ์€ ๋”๋””๊ฒŒ ์ž๋ผ๊ธฐ ์‹œ์ž‘ํ•˜๊ณ  ์žƒ์–ด๋ฒ„๋ฆฐ ๊ฐ€๊ฒฉ ๋ถ€์ธ์˜ ๊ฒฐํ˜ผ์‹์— ๋Œ€ํ•œ ์ดˆ๊ธฐ ๊ฒฝํ—˜์„ ํ†ตํ•ด ๊ฑฐ์˜ ์„ฑ์ธ์ด ๋˜์ง€๋งŒ ๋‚˜์ค‘์— ์ด ์‹œ์ ์—์„œ ๋งค์šฐ ๋ฉ๋‹ˆ๋‹ค.
The WMT competition winners for each direction as reported in WMT'21 and WMT'22 correspond to those used by Hendy et al. (2023). For more detailed information, we direct readers to this paper.
WMT'21 ๋ฐ WMT'22์—์„œ ๋ณด๊ณ ๋œ ๊ฐ ๋ฐฉํ–ฅ์˜ WMT ๋Œ€ํšŒ ์šฐ์Šน์ž๋Š” Hendy et al. (2023)์—์„œ ์‚ฌ์šฉ๋œ ๊ฒƒ๊ณผ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ์ •๋ณด๋Š” ์ด ๋…ผ๋ฌธ์„ ์ฐธ์กฐํ•˜์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค.
WMT'21 ๋ฐ WMT'22์— ๋ณด๊ณ ๋œ ๊ฐ ๋ฐฉํ–ฅ์˜ WMT ๊ฒฝ์Ÿ ์šฐ์Šน์ž๋Š” Hendy ๋“ฑ(2023)์—์„œ ์‚ฌ์šฉํ•œ ๊ฒƒ๊ณผ ์ผ์น˜ํ•œ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ์ด ๋…ผ๋ฌธ์„ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค.
_(B) **Later-Stage Dynamics.** As soon as_ \(T\gtrsim\overline{T}\)_, it holds that_ \(E_{test}\asymp(\pi T)^{-c}\)_. Similarly, we recover the unpolluted sample-size law scaling_ \(T^{-c}\)_.
_(B) **ํ›„๊ธฐ ๋‹จ๊ณ„ ๋™์—ญํ•™.** \(T\gtrsim\overline{T}\)๊ฐ€ ๋˜๋ฉด, \(E_{test}\asymp(\pi T)^{-c}\)๊ฐ€ ์„ฑ๋ฆฝํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ์˜ค์—ผ๋˜์ง€ ์•Š์€ ์ƒ˜ํ”Œ ํฌ๊ธฐ ๋ฒ•์น™ \(T^{-c}\)๋ฅผ ํšŒ๋ณตํ•ฉ๋‹ˆ๋‹ค.
_(B) **ํ›„๊ธฐ ๋™์—ญํ•™**._ \(T\gtrsim\overline{T}\)_, it holds that_ \(E_{test}\asymp(\pi T)^{-c}\)_. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ์˜ค์—ผ๋˜์ง€ ์•Š์€ ์ƒ˜ํ”Œ ํฌ๊ธฐ ๋ฒ•์น™ ์Šค์ผ€์ผ๋ง_ \(T^{-c}\)_.
Commonsense reasoning is one of the important capabilities each model can obtain. This capability denotes the ability of the model to use prior knowledge in combination with reasoning skills.
์ƒ์‹์  ์ถ”๋ก ์€ ๊ฐ ๋ชจ๋ธ์ด ์–ป์„ ์ˆ˜ ์žˆ๋Š” ์ค‘์š”ํ•œ ๋Šฅ๋ ฅ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์ด ๋Šฅ๋ ฅ์€ ๋ชจ๋ธ์ด ์‚ฌ์ „ ์ง€์‹์„ ์ถ”๋ก  ๊ธฐ์ˆ ๊ณผ ๊ฒฐํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์ƒ์‹ ์ถ”๋ก ์€ ๊ฐ ๋ชจ๋ธ์ด ํš๋“ํ•  ์ˆ˜ ์žˆ๋Š” ์ค‘์š”ํ•œ ๊ธฐ๋Šฅ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์ด ๊ธฐ๋Šฅ์€ ๋ชจ๋ธ์ด ์ถ”๋ก  ๊ธฐ์ˆ ๊ณผ ๊ฒฐํ•ฉํ•˜์—ฌ ์‚ฌ์ „ ์ง€์‹์„ ์‚ฌ์šฉํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
By following these steps, it is possible to train an LLM using RLHF methods based on InstructGPT to create a human assistant that exceeds open assistant in performance, accuracy, and utility.
์ด๋Ÿฌํ•œ ๋‹จ๊ณ„๋ฅผ ๋”ฐ๋ฅด๋ฉด InstructGPT๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ RLHF ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์„ฑ๋Šฅ, ์ •ํ™•์„ฑ ๋ฐ ์œ ์šฉ์„ฑ์—์„œ open assistant๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ธ๊ฐ„ ๋ณด์กฐ์ž๋ฅผ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด LLM์„ ํ›ˆ๋ จ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด ๋‹จ๊ณ„๋ฅผ ์ˆ˜ํ–‰ํ•˜๋ฉด InstructGPT๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ RLHF ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ ํ›ˆ๋ จํ•˜์—ฌ ์„ฑ๋Šฅ, ์ •ํ™•์„ฑ ๋ฐ ์œ ์šฉ์„ฑ ์ธก๋ฉด์—์„œ ์˜คํ”ˆ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ธ๊ฐ„ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Will the dataset be distributed under a copyright or other intellectual property (IP) license and/or under applicable terms of use (ToU)?** The dataset is distributed under a CC-BY license.
**๋ฐ์ดํ„ฐ์…‹์ด ์ €์ž‘๊ถŒ ๋˜๋Š” ๊ธฐํƒ€ ์ง€์  ์žฌ์‚ฐ๊ถŒ (IP) ๋ผ์ด์„ ์Šค ๋ฐ/๋˜๋Š” ์ ์šฉ ๊ฐ€๋Šฅํ•œ ์ด์šฉ ์•ฝ๊ด€ (ToU) ํ•˜์— ๋ฐฐํฌ๋ฉ๋‹ˆ๊นŒ?** ๋ฐ์ดํ„ฐ์…‹์€ CC-BY ๋ผ์ด์„ ์Šค ํ•˜์— ๋ฐฐํฌ๋ฉ๋‹ˆ๋‹ค.
**๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ €์ž‘๊ถŒ ๋˜๋Š” ๊ธฐํƒ€ ์ง€์  ์žฌ์‚ฐ๊ถŒ(IP) ๋ผ์ด์„ ์Šค ๋ฐ/๋˜๋Š” ํ•ด๋‹น ์‚ฌ์šฉ ์•ฝ๊ด€(ToU)์— ๋”ฐ๋ผ ๋ฐฐํฌ๋˜๋‚˜์š”?** ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” CC-BY ๋ผ์ด์„ ์Šค์— ๋”ฐ๋ผ ๋ฐฐํฌ๋ฉ๋‹ˆ๋‹ค.
In sDPO, we propose to use the available preference datasets in a stepwise manner instead of using them all at once. The comparison of the overall flow of DPO and sDPO is presented in Fig. 1.
sDPO์—์„œ๋Š” ๋ชจ๋“  ์„ ํ˜ธ ๋ฐ์ดํ„ฐ์…‹์„ ํ•œ ๋ฒˆ์— ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  ๋‹จ๊ณ„์ ์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. DPO์™€ sDPO์˜ ์ „์ฒด ํ๋ฆ„ ๋น„๊ต๋Š” Fig. 1์— ์ œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
sDPO์—์„œ๋Š” ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ•œ ๋ฒˆ์— ๋ชจ๋‘ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  ๋‹จ๊ณ„๋ณ„ ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์šฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•œ๋‹ค. DPO์™€ sDPO์˜ ์ „์ฒด ํ๋ฆ„ ๋น„๊ต๋Š” Fig. 1์— ๋‚˜์™€ ์žˆ๋‹ค.
A single head in the vanilla MHA computes its attention context \(A_{dot}\in\mathbb{R}^{N\times d_{value}}\) from sequence of input segments \(X\in\mathbb{R}^{N\times d_{model}}\) as follows.
๋ฐ”๋‹๋ผ MHA์˜ ๋‹จ์ผ ํ—ค๋“œ๋Š” ์ž…๋ ฅ ์‹œํ€€์Šค \(X\in\mathbb{R}^{N\times d_{model}}\)๋กœ๋ถ€ํ„ฐ ์ฃผ์˜ ์ปจํ…์ŠคํŠธ \(A_{dot}\in\mathbb{R}^{N\times d_{value}}\)๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
๋ฐ”๋‹๋ผ MHA์˜ ๋‹จ์ผ ํ—ค๋“œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ž…๋ ฅ ์„ธ๊ทธ๋จผํŠธ ์‹œํ€€์Šค \(X\in\mathbb{R}^{N\times d_{model}}\)๋กœ๋ถ€ํ„ฐ ์ฃผ์˜ ์ปจํ…์ŠคํŠธ \(A_{dot}\in\mathbb{R}^{N\times d_{value}}\)๋ฅผ ๊ณ„์‚ฐํ•œ๋‹ค.
We also train a multimodal version of MiniCPM, namely MiniCPM-V. The details of MiniCPM-V will be introduced in a separate paper. Here, we only list a snapshot of the performance in Table 11.
์šฐ๋ฆฌ๋Š” ๋˜ํ•œ MiniCPM์˜ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ฒ„์ „์ธ MiniCPM-V๋ฅผ ํ›ˆ๋ จ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. MiniCPM-V์˜ ์„ธ๋ถ€ ์‚ฌํ•ญ์€ ๋ณ„๋„์˜ ๋…ผ๋ฌธ์—์„œ ์†Œ๊ฐœ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” Table 11์— ์„ฑ๋Šฅ์˜ ์Šค๋ƒ…์ƒท๋งŒ ๋‚˜์—ดํ•ฉ๋‹ˆ๋‹ค.
๋˜ํ•œ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ฒ„์ „์˜ MiniCPM, ์ฆ‰ MiniCPM-V๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. MiniCPM-V์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ณ„๋„์˜ ๋…ผ๋ฌธ์—์„œ ์†Œ๊ฐœํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” ํ‘œ 11์— ์„ฑ๋Šฅ์˜ ์Šค๋ƒ…์ƒท๋งŒ ๋‚˜์—ดํ•ฉ๋‹ˆ๋‹ค.
Preprocess the data: The data needs to be preprocessed to remove noise and irrelevant information. This step is crucial to ensure that the model learns the correct patterns and relationships.
๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ: ๋ฐ์ดํ„ฐ์—์„œ ๋…ธ์ด์ฆˆ์™€ ๊ด€๋ จ ์—†๋Š” ์ •๋ณด๋ฅผ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด ์ „์ฒ˜๋ฆฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๋Š” ๋ชจ๋ธ์ด ์˜ฌ๋ฐ”๋ฅธ ํŒจํ„ด๊ณผ ๊ด€๊ณ„๋ฅผ ํ•™์Šตํ•˜๋„๋ก ๋ณด์žฅํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ: ๋ฐ์ดํ„ฐ๋ฅผ ์ „์ฒ˜๋ฆฌํ•˜์—ฌ ๋…ธ์ด์ฆˆ์™€ ๊ด€๋ จ ์—†๋Š” ์ •๋ณด๋ฅผ ์ œ๊ฑฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๋Š” ๋ชจ๋ธ์ด ์˜ฌ๋ฐ”๋ฅธ ํŒจํ„ด๊ณผ ๊ด€๊ณ„๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
Since SFT and RLHF are two major adaptation tuning methods for LLMs, it is important to understand the connections and difference between them. Next, we make some discussions on this issue30.
SFT์™€ RLHF๋Š” LLMs์˜ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์ ์‘ ํŠœ๋‹ ๋ฐฉ๋ฒ•์ด๊ธฐ ๋•Œ๋ฌธ์—, ์ด๋“ค ๊ฐ„์˜ ์—ฐ๊ฒฐ์„ฑ๊ณผ ์ฐจ์ด์ ์„ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, ์ด ๋ฌธ์ œ์— ๋Œ€ํ•ด ๋ช‡ ๊ฐ€์ง€ ๋…ผ์˜๋ฅผ ์ง„ํ–‰ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.
SFT์™€ RLHF๋Š” LLM์˜ ์ฃผ์š” ์ ์‘ ํŠœ๋‹ ๋ฐฉ๋ฒ•์ด๋ฏ€๋กœ, ์ด ๋‘˜ ์‚ฌ์ด์˜ ์—ฐ๊ด€์„ฑ๊ณผ ์ฐจ์ด์ ์„ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, ์ด ๋ฌธ์ œ์— ๋Œ€ํ•ด ๋ช‡ ๊ฐ€์ง€ ๋…ผ์˜๋ฅผ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
Since their salaries increased simultaneously, Lylah's salary also increased by 40% of what she was making four years ago. So, we need to add 40% of $28000 to her salary from four years ago.
๊ทธ๋“ค์˜ ๊ธ‰์—ฌ๊ฐ€ ๋™์‹œ์— ์ธ์ƒ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์—, Lylah์˜ ๊ธ‰์—ฌ๋„ 4๋…„ ์ „ ๊ทธ๋…€๊ฐ€ ๋ฐ›๋˜ ๊ธˆ์•ก์˜ 40%๋งŒํผ ์ฆ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ, 4๋…„ ์ „ ๊ทธ๋…€์˜ ๊ธ‰์—ฌ์— $28000์˜ 40%๋ฅผ ๋”ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๋‘ ์‚ฌ๋žŒ์˜ ๊ธ‰์—ฌ๊ฐ€ ๋™์‹œ์— ์ธ์ƒ๋˜์—ˆ์œผ๋ฏ€๋กœ ๋ฆฌ๋ผ์˜ ๊ธ‰์—ฌ๋„ 4๋…„ ์ „ ๊ธ‰์—ฌ์˜ 40%๊ฐ€ ์ธ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ 4๋…„ ์ „ ๊ธ‰์—ฌ์— 28000๋‹ฌ๋Ÿฌ์˜ 40%๋ฅผ ๋”ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Figure. 2 (Right) shows the results of models using a mixture of GSM8K and MATH in a 1:1 ratio. These models exhibit balanced scaling behaviors in both in-domain and out-of-domain benchmarks.
Figure. 2 (์˜ค๋ฅธ์ชฝ)์€ GSM8K์™€ MATH๋ฅผ 1:1 ๋น„์œจ๋กœ ํ˜ผํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ•œ ๋ชจ๋ธ์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ๋“ค์€ ๋„๋ฉ”์ธ ๋‚ด ๋ฐ ๋„๋ฉ”์ธ ์™ธ ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ท ํ˜• ์žกํžŒ ํ™•์žฅ ํ–‰๋™์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2 (์˜ค๋ฅธ์ชฝ)๋Š” GSM8K์™€ MATH๋ฅผ 1:1 ๋น„์œจ๋กœ ํ˜ผํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ•œ ๋ชจ๋ธ์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ๋„๋ฉ”์ธ ๋‚ด ๋ฐ ๋„๋ฉ”์ธ ์™ธ ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ท ํ˜• ์žกํžŒ ์Šค์ผ€์ผ๋ง ๋™์ž‘์„ ๋ณด์—ฌ์ค€๋‹ค.
**QUESTION:** Jim spends 2 hours watching TV and then decides to go to bed and reads for half as long. He does this 3 times a week. How many hours does he spend on TV and reading in 4 weeks?
**์งˆ๋ฌธ:** Jim์€ TV๋ฅผ 2์‹œ๊ฐ„ ์‹œ์ฒญํ•œ ํ›„ ์ž ์ž๋ฆฌ์— ๋“ค๊ธฐ๋กœ ํ•˜๊ณ  ๊ทธ ์ ˆ๋ฐ˜์˜ ์‹œ๊ฐ„ ๋™์•ˆ ๋…์„œ๋ฅผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Š” ์ด๊ฒƒ์„ ์ผ์ฃผ์ผ์— 3๋ฒˆ ํ•ฉ๋‹ˆ๋‹ค. 4์ฃผ ๋™์•ˆ TV ์‹œ์ฒญ๊ณผ ๋…์„œ์— ๋ช‡ ์‹œ๊ฐ„์„ ์†Œ๋น„ํ•ฉ๋‹ˆ๊นŒ?
**์งˆ๋ฌธ:** ์ง์€ TV๋ฅผ ์‹œ์ฒญํ•˜๋Š” ๋ฐ 2์‹œ๊ฐ„์„ ๋ณด๋‚ธ ํ›„ ์ž ์ž๋ฆฌ์— ๋“ค๊ธฐ๋กœ ๊ฒฐ์ •ํ•˜๊ณ  ์ ˆ๋ฐ˜ ์ •๋„๋งŒ ์ฝ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Š” ์ผ์ฃผ์ผ์— ์„ธ ๋ฒˆ ์ด๋ ‡๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. 4์ฃผ ๋™์•ˆ ์ง์€ TV์™€ ๋…์„œ์— ๋ช‡ ์‹œ๊ฐ„์„ ํ• ์• ํ•˜๋‚˜์š”?
For base models, we consider both common language models, i.e., LLaMA-2 7B/13B/70B/Mistral-7B, and math-specific models, such as Llemma-7B, to assess the generality of the proposed approach.
๊ธฐ๋ณธ ๋ชจ๋ธ๋กœ๋Š” ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ๋ชจ๋ธ์ธ LLaMA-2 7B/13B/70B/Mistral-7B์™€ ์ˆ˜ํ•™ ์ „์šฉ ๋ชจ๋ธ์ธ Llemma-7B๋ฅผ ๋ชจ๋‘ ๊ณ ๋ คํ•˜์—ฌ ์ œ์•ˆ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์ผ๋ฐ˜์„ฑ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, ๊ณตํ†ต ์–ธ์–ด ๋ชจ๋ธ(์˜ˆ: LLaMA-2 7B/13B/70B/Mistral-7B)๊ณผ ์ˆ˜ํ•™ ์ „๋ฌธ ๋ชจ๋ธ(์˜ˆ: Llemma-7B)์„ ๋ชจ๋‘ ๊ณ ๋ คํ•˜์—ฌ ์ œ์•ˆ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์ผ๋ฐ˜์„ฑ์„ ํ‰๊ฐ€ํ•œ๋‹ค.
ically, the Yi-6B model's superior performance against models with significantly more parameters challenges the notion that parameter count alone is a sufficient predictor of model efficacy.
ํŠนํžˆ, Yi-6B ๋ชจ๋ธ์˜ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๋งŒ์œผ๋กœ ๋ชจ๋ธ์˜ ํšจ์œจ์„ฑ์„ ์ถฉ๋ถ„ํžˆ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฐœ๋…์— ๋„์ „ํ•ฉ๋‹ˆ๋‹ค.
๊ธฐ์ˆ ์ ์œผ๋กœ, ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ ํ›จ์”ฌ ๋” ๋งŽ์€ ๋ชจ๋ธ์— ๋Œ€ํ•œ Yi-6B ๋ชจ๋ธ์˜ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๋งŒ์œผ๋กœ ๋ชจ๋ธ ํšจ๋Šฅ์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฐœ๋…์— ๋„์ „ํ•œ๋‹ค.
In this section, we apply the PLND method to selected languages and models in order to confirm the existence of language-specific neurons and investigate the relationships between languages.
์ด ์„น์…˜์—์„œ๋Š” PLND ๋ฐฉ๋ฒ•์„ ์„ ํƒ๋œ ์–ธ์–ด์™€ ๋ชจ๋ธ์— ์ ์šฉํ•˜์—ฌ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์˜ ์กด์žฌ๋ฅผ ํ™•์ธํ•˜๊ณ  ์–ธ์–ด ๊ฐ„์˜ ๊ด€๊ณ„๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์˜ ์กด์žฌ๋ฅผ ํ™•์ธํ•˜๊ณ  ์–ธ์–ด ๊ฐ„์˜ ๊ด€๊ณ„๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด PLND ๋ฐฉ๋ฒ•์„ ์„ ํƒ๋œ ์–ธ์–ด์™€ ๋ชจ๋ธ์— ์ ์šฉํ•œ๋‹ค.
- Given his history of strong financial support and political influence, McCarthy may feel compelled to stay and influence the political process, especially with the upcoming 2024 elections.
- ๊ฐ•๋ ฅํ•œ ์žฌ์ • ์ง€์›๊ณผ ์ •์น˜์  ์˜ํ–ฅ๋ ฅ์„ ๊ฐ€์ง„ ๊ทธ์˜ ์ด๋ ฅ์„ ๊ณ ๋ คํ•  ๋•Œ, McCarthy๋Š” ํŠนํžˆ ๋‹ค๊ฐ€์˜ค๋Š” 2024๋…„ ์„ ๊ฑฐ๋ฅผ ์•ž๋‘๊ณ  ์ •์น˜ ๊ณผ์ •์— ๋‚จ์•„ ์˜ํ–ฅ์„ ๋ฏธ์ณ์•ผ ํ•œ๋‹ค๋Š” ์••๋ฐ•์„ ๋Š๋‚„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
- ๊ฐ•๋ ฅํ•œ ์žฌ์ •์  ์ง€์›๊ณผ ์ •์น˜์  ์˜ํ–ฅ๋ ฅ์„ ๋ฐœํœ˜ํ•ด ์˜จ ๊ทธ์˜ ์—ญ์‚ฌ๋ฅผ ๊ณ ๋ คํ•  ๋•Œ, ๋งค์นด์‹œ๋Š” ํŠนํžˆ 2024๋…„ ๋Œ€์„ ์„ ์•ž๋‘๊ณ  ์ •์น˜ ๊ณผ์ •์— ๊ณ„์† ์ฐธ์—ฌํ•˜๊ณ  ์˜ํ–ฅ๋ ฅ์„ ํ–‰์‚ฌํ•ด์•ผ ํ•œ๋‹ค๊ณ  ๋Š๋‚„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Xiao Wang,, Shiao Wang, Yuhe Ding, Yuehang Li, Wentao Wu, Yao Rong, Weizhe Kong, Ju Huang, Shihao Li, Haoxiang Yang, Ziwen Wang, Bo Jiang, Chenglong Li, Yaowei Wang, Yonghong Tian,, Jin Tang
Xiao Wang, Shiao Wang, Yuhe Ding, Yuehang Li, Wentao Wu, Yao Rong, Weizhe Kong, Ju Huang, Shihao Li, Haoxiang Yang, Ziwen Wang, Bo Jiang, Chenglong Li, Yaowei Wang, Yonghong Tian, Jin Tang
์ƒค์˜ค ์™•, ์‹œ์•„์˜ค ์™•, ์œ ํ—ˆ ๋”ฉ, ์œ„ํ•œ ๋ฆฌ, ์›ํƒ€์˜ค ์šฐ, ์•ผ์˜ค ๋ฃฝ, ์œ„์ œ ์ฝฉ, ์ฃผ ํ™ฉ, ์‹œํ•˜์˜ค ๋ฆฌ, ํ•˜์˜ค์ƒน ์–‘, ์ง€์›ฌ ์™•, ๋ณด ์žฅ, ์ฒธ๋ฃฝ ๋ฆฌ, ์•ผ์˜ค์›จ์ด ์™•, ์šฉํ™ ํ‹ฐ์•ˆ, ์ง„ํƒ•
This section will introduce some of the data sources commonly used in RALM and categorise them into structured and unstructured data. Figure 5 illustrates the categorization of data sources.
์ด ์„น์…˜์—์„œ๋Š” RALM์—์„œ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ ์†Œ์Šค ๋ช‡ ๊ฐ€์ง€๋ฅผ ์†Œ๊ฐœํ•˜๊ณ  ์ด๋ฅผ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ์™€ ๋น„๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. Figure 5๋Š” ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ๋ถ„๋ฅ˜๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” RALM์—์„œ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ์†Œ๊ฐœํ•˜๊ณ  ์ด๋ฅผ ์ •ํ˜• ๋ฐ์ดํ„ฐ์™€ ๋น„์ •ํ˜• ๋ฐ์ดํ„ฐ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 5๋Š” ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ๋ถ„๋ฅ˜๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
We use "Answer the following question with a short span, or a full and complete answer." for SQA and TopiOCQA, since these datasets have both short and long answers based on the questions.
SQA์™€ TopiOCQA์— ๋Œ€ํ•ด์„œ๋Š” "๋‹ค์Œ ์งˆ๋ฌธ์— ์งง์€ ๋‹ต๋ณ€ ๋˜๋Š” ์™„์ „ํ•˜๊ณ  ์™„๋ฒฝํ•œ ๋‹ต๋ณ€์œผ๋กœ ๋‹ตํ•˜์‹œ์˜ค."๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ์…‹๋“ค์€ ์งˆ๋ฌธ์— ๋”ฐ๋ผ ์งง์€ ๋‹ต๋ณ€๊ณผ ๊ธด ๋‹ต๋ณ€์„ ๋ชจ๋‘ ํฌํ•จํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ์งˆ๋ฌธ์— ๋”ฐ๋ผ ์งง์€ ๋‹ต๋ณ€๊ณผ ๊ธด ๋‹ต๋ณ€์ด ๋ชจ๋‘ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— SQA์™€ TopiOCQA์—๋Š” "์งง์€ ๋ฒ”์œ„ ๋˜๋Š” ์™„์ „ํ•˜๊ณ  ์™„์ „ํ•œ ๋‹ต๋ณ€์œผ๋กœ ๋‹ค์Œ ์งˆ๋ฌธ์— ๋‹ตํ•˜์„ธ์š”."๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
Substituting \(S_{min}\) with \(S\)Finally, after establishing the analytical relationship between \(S_{min}\) and \(S\), we can substitute \(S_{min}\) with \(S\) to link \(S\) with \(L\).
๋งˆ์ง€๋ง‰์œผ๋กœ, \(S_{min}\)๊ณผ \(S\) ์‚ฌ์ด์˜ ๋ถ„์„์  ๊ด€๊ณ„๋ฅผ ํ™•๋ฆฝํ•œ ํ›„, \(S_{min}\)์„ \(S\)๋กœ ๋Œ€์ฒดํ•˜์—ฌ \(S\)์™€ \(L\)์„ ์—ฐ๊ฒฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
S_{min}\)์„ \(S\)๋กœ ๋Œ€์ž…๋งˆ์ง€๋ง‰์œผ๋กœ, \(S_{min}\)๊ณผ \(S\) ์‚ฌ์ด์˜ ๋ถ„์„์  ๊ด€๊ณ„๋ฅผ ์„ค์ •ํ•œ ํ›„, \(S_{min}\)์„ \(S\)๋กœ ๋Œ€์ฒดํ•˜์—ฌ \(S\)์™€ \(L\)์„ ์—ฐ๊ฒฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In our study, since the preferred data is selected by reference-free models and the same models are used for evaluation, we investigate the potential for 'cheating' in the scoring process.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ์—์„œ, ์„ ํ˜ธ๋˜๋Š” ๋ฐ์ดํ„ฐ๋Š” ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ๋“ค์— ์˜ํ•ด ์„ ํƒ๋˜๊ณ  ๋™์ผํ•œ ๋ชจ๋ธ๋“ค์ด ํ‰๊ฐ€์— ์‚ฌ์šฉ๋˜๊ธฐ ๋•Œ๋ฌธ์—, ์šฐ๋ฆฌ๋Š” ์ ์ˆ˜ ๋งค๊ธฐ๊ธฐ ๊ณผ์ •์—์„œ '๋ถ€์ •ํ–‰์œ„'์˜ ๊ฐ€๋Šฅ์„ฑ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ์— ์˜ํ•ด ์„ ํ˜ธ ๋ฐ์ดํ„ฐ๊ฐ€ ์„ ํƒ๋˜๊ณ  ๋™์ผํ•œ ๋ชจ๋ธ์ด ํ‰๊ฐ€์— ์‚ฌ์šฉ๋˜๊ธฐ ๋•Œ๋ฌธ์— ์ฑ„์  ๊ณผ์ •์—์„œ '๋ถ€์ •ํ–‰์œ„'์˜ ๊ฐ€๋Šฅ์„ฑ์„ ์กฐ์‚ฌํ•œ๋‹ค.
These enhancement techniques lead to the improved GPT-3 models with stronger capacities, which are called GPT-3.5 models by OpenAI (see the discussion about the OpenAI API in Section 3.1).
์ด๋Ÿฌํ•œ ํ–ฅ์ƒ ๊ธฐ์ˆ ๋“ค์€ ๋” ๊ฐ•๋ ฅํ•œ ๋Šฅ๋ ฅ์„ ๊ฐ€์ง„ ๊ฐœ์„ ๋œ GPT-3 ๋ชจ๋ธ๋กœ ์ด์–ด์ง€๋ฉฐ, ์ด๋Š” OpenAI์— ์˜ํ•ด GPT-3.5 ๋ชจ๋ธ๋กœ ๋ถˆ๋ฆฝ๋‹ˆ๋‹ค (OpenAI API์— ๋Œ€ํ•œ ๋…ผ์˜๋Š” 3.1์ ˆ์„ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค).
์ด๋Ÿฌํ•œ ํ–ฅ์ƒ ๊ธฐ๋ฒ•์€ ๋” ๊ฐ•๋ ฅํ•œ ์šฉ๋Ÿ‰์„ ๊ฐ–์ถ˜ ๊ฐœ์„ ๋œ GPT-3 ๋ชจ๋ธ๋กœ ์ด์–ด์ง€๋ฉฐ, ์ด๋ฅผ OpenAI์—์„œ๋Š” GPT-3.5 ๋ชจ๋ธ์ด๋ผ๊ณ  ๋ถ€๋ฆ…๋‹ˆ๋‹ค(์„น์…˜ 3.1์˜ OpenAI API์— ๋Œ€ํ•œ ๋…ผ์˜ ์ฐธ์กฐ).
2. Recovery: HIIT involves short recovery periods, which can help to increase muscle strength and endurance. Other forms of exercise, such as jogging, may not provide enough recovery time.
2. ํšŒ๋ณต: HIIT๋Š” ์งง์€ ํšŒ๋ณต ๊ธฐ๊ฐ„์„ ํฌํ•จํ•˜๋ฉฐ, ์ด๋Š” ๊ทผ๋ ฅ๊ณผ ์ง€๊ตฌ๋ ฅ์„ ์ฆ๊ฐ€์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์กฐ๊น…๊ณผ ๊ฐ™์€ ๋‹ค๋ฅธ ํ˜•ํƒœ์˜ ์šด๋™์€ ์ถฉ๋ถ„ํ•œ ํšŒ๋ณต ์‹œ๊ฐ„์„ ์ œ๊ณตํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
2. ํšŒ๋ณต: HIIT๋Š” ์งง์€ ํšŒ๋ณต ๊ธฐ๊ฐ„์„ ํฌํ•จํ•˜๋ฏ€๋กœ ๊ทผ๋ ฅ๊ณผ ์ง€๊ตฌ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์กฐ๊น…๊ณผ ๊ฐ™์€ ๋‹ค๋ฅธ ํ˜•ํƒœ์˜ ์šด๋™์€ ์ถฉ๋ถ„ํ•œ ํšŒ๋ณต ์‹œ๊ฐ„์„ ์ œ๊ณตํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 4: _Naive_ and _overall_ efficiency gain of data selection via D4 relative to the total cost of training as a function of model size on Instruct + Answers perplexity at \(R=0.25\).
Figure 4: Instruct + Answers ๋‹นํ˜น๋„์—์„œ \(R=0.25\)์ผ ๋•Œ ๋ชจ๋ธ ํฌ๊ธฐ์˜ ํ•จ์ˆ˜๋กœ์„œ D4๋ฅผ ํ†ตํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ์˜ _Naive_ ๋ฐ _overall_ ํšจ์œจ์„ฑ ์ฆ๊ฐ€๋ฅผ ์ „์ฒด ํ›ˆ๋ จ ๋น„์šฉ๊ณผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ.
๊ทธ๋ฆผ 4: Instruct + Answers ๋‚œํ•ด๋„์—์„œ ๋ชจ๋ธ ํฌ๊ธฐ์— ๋Œ€ํ•œ ํ•จ์ˆ˜๋กœ์„œ D4๋ฅผ ํ†ตํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ์˜ _์ˆœ์ง„ํ•œ_ ํšจ์œจ์„ฑ๊ณผ ์ด ํ›ˆ๋ จ ๋น„์šฉ์— ๋Œ€ํ•œ _์ „์ฒด์ ์ธ_ ํšจ์œจ์„ฑ ์ด๋“(\(R=0.25\)).
Figure 12: Same as Figure 2, but including models with accuracies below 50% (which may overlap with higher-accuracy models). The _peak_ capacity ratios consistently exceed \(R(F)\geq 2\).
Figure 12: Figure 2์™€ ๋™์ผํ•˜์ง€๋งŒ, ์ •ํ™•๋„๊ฐ€ 50% ์ดํ•˜์ธ ๋ชจ๋ธ๋„ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค (์ด๋Š” ๋” ๋†’์€ ์ •ํ™•๋„์˜ ๋ชจ๋ธ๊ณผ ๊ฒน์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค). _์ตœ๋Œ€_ ์šฉ๋Ÿ‰ ๋น„์œจ์€ ์ผ๊ด€๋˜๊ฒŒ \(R(F)\geq 2\)๋ฅผ ์ดˆ๊ณผํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 12: ๊ทธ๋ฆผ 2์™€ ๋™์ผํ•˜์ง€๋งŒ ์ •ํ™•๋„๊ฐ€ 50% ๋ฏธ๋งŒ์ธ ๋ชจ๋ธ(์ •ํ™•๋„๊ฐ€ ๋†’์€ ๋ชจ๋ธ๊ณผ ๊ฒน์น  ์ˆ˜ ์žˆ์Œ)์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ํ”ผํฌ_ ์šฉ๋Ÿ‰ ๋น„์œจ์€ ์ผ๊ด€๋˜๊ฒŒ \(R(F)\geq 2\)๋ฅผ ์ดˆ๊ณผํ•ฉ๋‹ˆ๋‹ค.
Footnote 2: Most duplicates we saw were automatically generated pages, such as the outcomes of sports games. This shows the strength of manual curation for creating high-quality datasets.
Footnote 2: ์šฐ๋ฆฌ๊ฐ€ ๋ณธ ๋Œ€๋ถ€๋ถ„์˜ ์ค‘๋ณต ํŽ˜์ด์ง€๋Š” ์Šคํฌ์ธ  ๊ฒฝ๊ธฐ ๊ฒฐ๊ณผ์™€ ๊ฐ™์€ ์ž๋™ ์ƒ์„ฑ๋œ ํŽ˜์ด์ง€์˜€์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ์…‹์„ ๋งŒ๋“ค๊ธฐ ์œ„ํ•œ ์ˆ˜์ž‘์—… ํ๋ ˆ์ด์…˜์˜ ๊ฐ•์ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๊ฐ์ฃผ 2: ์šฐ๋ฆฌ๊ฐ€ ๋ณธ ๋Œ€๋ถ€๋ถ„์˜ ์ค‘๋ณต์€ ์Šคํฌ์ธ  ๊ฒฝ๊ธฐ ๊ฒฐ๊ณผ์™€ ๊ฐ™์ด ์ž๋™์œผ๋กœ ์ƒ์„ฑ๋œ ํŽ˜์ด์ง€์˜€์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“œ๋Š” ๋ฐ ์žˆ์–ด ์ˆ˜๋™ ํ๋ ˆ์ด์…˜์˜ ๊ฐ•์ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
Some of the real-world datasets used in the development of AMIE are open-source (MedQA). The scenario packs from UK used in the OSCE study are also available for download on the internet.
AMIE ๊ฐœ๋ฐœ์— ์‚ฌ์šฉ๋œ ์ผ๋ถ€ ์‹ค์ œ ๋ฐ์ดํ„ฐ์…‹์€ ์˜คํ”ˆ ์†Œ์Šค์ž…๋‹ˆ๋‹ค (MedQA). OSCE ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋œ ์˜๊ตญ์˜ ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ๋„ ์ธํ„ฐ๋„ท์—์„œ ๋‹ค์šด๋กœ๋“œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
AMIE ๊ฐœ๋ฐœ์— ์‚ฌ์šฉ๋œ ์‹ค์ œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ค‘ ์ผ๋ถ€๋Š” ์˜คํ”ˆ ์†Œ์Šค(MedQA)์ž…๋‹ˆ๋‹ค. OSCE ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋œ ์˜๊ตญ์˜ ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ๋„ ์ธํ„ฐ๋„ท์—์„œ ๋‹ค์šด๋กœ๋“œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
For the stage-1 SFT data used in stage-2 instruction tuning, we keep the same format template as in stage-1. In other words, the {Context for Latest Question} and {Instruction} are empty.
2๋‹จ๊ณ„ instruction tuning์— ์‚ฌ์šฉ๋˜๋Š” 1๋‹จ๊ณ„ SFT ๋ฐ์ดํ„ฐ์˜ ๊ฒฝ์šฐ, 1๋‹จ๊ณ„์™€ ๋™์ผํ•œ ํ˜•์‹ ํ…œํ”Œ๋ฆฟ์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, {Context for Latest Question}์™€ {Instruction}์€ ๋น„์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
2๋‹จ๊ณ„ ์ง€์นจ ํŠœ๋‹์— ์‚ฌ์šฉ๋˜๋Š” 1๋‹จ๊ณ„ SFT ๋ฐ์ดํ„ฐ์˜ ๊ฒฝ์šฐ, 1๋‹จ๊ณ„์™€ ๋™์ผํ•œ ํ˜•์‹ ํ…œํ”Œ๋ฆฟ์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, '{์ตœ์‹  ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์ปจํ…์ŠคํŠธ}'์™€ '{์ง€์นจ}'์€ ๋น„์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
Finally, to demonstrate why it is reasonable to assume that models cannot generally answer questions about new information, with better than random success, look at the following example:
๋งˆ์ง€๋ง‰์œผ๋กœ, ๋ชจ๋ธ์ด ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒˆ๋กœ์šด ์ •๋ณด์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ๋ฌด์ž‘์œ„ ์„ฑ๊ณต๋ฅ ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๊ณผ๋กœ ๋‹ตํ•  ์ˆ˜ ์—†๋‹ค๊ณ  ๊ฐ€์ •ํ•˜๋Š” ๊ฒƒ์ด ์™œ ํ•ฉ๋ฆฌ์ ์ธ์ง€ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด, ๋‹ค์Œ ์˜ˆ๋ฅผ ์‚ดํŽด๋ณด์‹ญ์‹œ์˜ค:
๋งˆ์ง€๋ง‰์œผ๋กœ, ๋ชจ๋ธ์ด ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒˆ๋กœ์šด ์ •๋ณด์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ๋ฌด์ž‘์œ„๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๊ณต๋ฅ ๋กœ ๋Œ€๋‹ตํ•  ์ˆ˜ ์—†๋‹ค๊ณ  ๊ฐ€์ •ํ•˜๋Š” ๊ฒƒ์ด ํ•ฉ๋ฆฌ์ ์ธ ์ด์œ ๋ฅผ ๋ณด์—ฌ์ฃผ๊ธฐ ์œ„ํ•ด ๋‹ค์Œ ์˜ˆ๋ฅผ ์‚ดํŽด๋ณด์„ธ์š”:
detailed information about access requests, agent activities, and any modifications to the access control parameters, which help to safeguard against potential privilege attacks [64; 65].
์ ‘๊ทผ ์š”์ฒญ, ์—์ด์ „ํŠธ ํ™œ๋™ ๋ฐ ์ ‘๊ทผ ์ œ์–ด ๋งค๊ฐœ๋ณ€์ˆ˜์˜ ๋ชจ๋“  ์ˆ˜์ • ์‚ฌํ•ญ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด๋Š” ์ž ์žฌ์ ์ธ ๊ถŒํ•œ ๊ณต๊ฒฉ์„ ๋ฐฉ์ง€ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค [64; 65].
์•ก์„ธ์Šค ์š”์ฒญ, ์—์ด์ „ํŠธ ํ™œ๋™ ๋ฐ ์•ก์„ธ์Šค ์ œ์–ด ๋งค๊ฐœ๋ณ€์ˆ˜์˜ ๋ชจ๋“  ์ˆ˜์ • ์‚ฌํ•ญ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด๋Š” ์ž ์žฌ์ ์ธ ๊ถŒํ•œ ๊ณต๊ฒฉ์œผ๋กœ๋ถ€ํ„ฐ ๋ณดํ˜ธํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค[64; 65].
**Evaluation of Fine-tuned LLMs.** Fine-tuned LLMs in this part refer to the model checkpoints obtained after instruction tuning or alignment tuning based on pre-trained model weights45.
**๋ฏธ์„ธ ์กฐ์ •๋œ LLM์˜ ํ‰๊ฐ€.** ์ด ๋ถ€๋ถ„์—์„œ ๋ฏธ์„ธ ์กฐ์ •๋œ LLM์€ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ instruction tuning ๋˜๋Š” alignment tuning ํ›„์— ์–ป์€ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
**๋ฏธ์„ธ ์กฐ์ •๋œ LLM์˜ ํ‰๊ฐ€**. ์ด ๋ถ€๋ถ„์—์„œ ๋ฏธ์„ธ ์กฐ์ •๋œ LLM์€ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜45๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ช…๋ น์–ด ํŠœ๋‹ ๋˜๋Š” ์ •๋ ฌ ํŠœ๋‹ ํ›„ ์–ป์€ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
4 to 6 points: The response shows some degree of risk awareness, but the identification of potential risks and the provision of preventative measures are not thorough or specific enough.
4์ ์—์„œ 6์ : ์‘๋‹ต์€ ์–ด๋Š ์ •๋„์˜ ์œ„ํ—˜ ์ธ์‹์„ ๋ณด์—ฌ์ฃผ์ง€๋งŒ, ์ž ์žฌ์  ์œ„ํ—˜์˜ ์‹๋ณ„๊ณผ ์˜ˆ๋ฐฉ ์กฐ์น˜์˜ ์ œ๊ณต์ด ์ถฉ๋ถ„ํžˆ ์ฒ ์ €ํ•˜๊ฑฐ๋‚˜ ๊ตฌ์ฒด์ ์ด์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
4~6์ : ์‘๋‹ต์€ ์–ด๋Š ์ •๋„์˜ ์œ„ํ—˜ ์ธ์‹์„ ๋ณด์—ฌ ์ฃผ์ง€๋งŒ ์ž ์žฌ์  ์œ„ํ—˜์„ ์‹๋ณ„ํ•˜๊ณ  ์˜ˆ๋ฐฉ ์กฐ์น˜๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์ด ์ถฉ๋ถ„ํžˆ ์ฒ ์ €ํ•˜๊ฑฐ๋‚˜ ๊ตฌ์ฒด์ ์ด์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
This adds only a single scalar value as training parameter per head while allowing a learnable trade-off between the long-term and local information flows in the model (Wu et al., 2022).
์ด๋Š” ๊ฐ ํ—ค๋“œ๋‹น ํ•˜๋‚˜์˜ ์Šค์นผ๋ผ ๊ฐ’์„ ํ›ˆ๋ จ ๋งค๊ฐœ๋ณ€์ˆ˜๋กœ ์ถ”๊ฐ€ํ•˜๋ฉด์„œ ๋ชจ๋ธ ๋‚ด์—์„œ ์žฅ๊ธฐ ๋ฐ ๋กœ์ปฌ ์ •๋ณด ํ๋ฆ„ ๊ฐ„์˜ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ๊ท ํ˜•์„ ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค (Wu et al., 2022).
์ด๋Š” ๋ชจ๋ธ์˜ ์žฅ๊ธฐ ๋ฐ ๋กœ์ปฌ ์ •๋ณด ํ๋ฆ„ ๊ฐ„์— ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํŠธ๋ ˆ์ด๋“œ์˜คํ”„๋ฅผ ํ—ˆ์šฉํ•˜๋ฉด์„œ ํ—ค๋“œ๋‹น ๋‹จ์ผ ์Šค์นผ๋ผ ๊ฐ’๋งŒ์„ ํ›ˆ๋ จ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ์ถ”๊ฐ€ํ•œ๋‹ค(Wu et al., 2022).
Footnote 2: We omit the bias term as it often models the frequency of the token (Kobayashi et al., 2023), which we do not want to consider for measuring the internal recall of an entity.
Footnote 2: ์šฐ๋ฆฌ๋Š” ํŽธํ–ฅ ํ•ญ์„ ์ƒ๋žตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ข…์ข… ํ† ํฐ์˜ ๋นˆ๋„๋ฅผ ๋ชจ๋ธ๋งํ•˜๋Š”๋ฐ ์‚ฌ์šฉ๋˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค(Kobayashi et al., 2023). ์ด๋Š” ์—”ํ‹ฐํ‹ฐ์˜ ๋‚ด๋ถ€ ํšŒ์ƒ์„ ์ธก์ •ํ•˜๋Š” ๋ฐ ๊ณ ๋ คํ•˜๊ณ  ์‹ถ์ง€ ์•Š์€ ์š”์†Œ์ž…๋‹ˆ๋‹ค.
๊ฐ์ฃผ 2: ๋ฐ”์ด์–ด์Šค ํ•ญ์€ ์ข…์ข… ํ† ํฐ์˜ ๋นˆ๋„๋ฅผ ๋ชจ๋ธ๋งํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ƒ๋žตํ•ฉ๋‹ˆ๋‹ค(Kobayashi et al., 2023), ์ด๋Š” ์—”ํ‹ฐํ‹ฐ์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์„ ์ธก์ •ํ•  ๋•Œ ๊ณ ๋ คํ•˜๊ณ  ์‹ถ์ง€ ์•Š์Šต๋‹ˆ๋‹ค.