question
stringlengths
101
4.47k
chosen
stringlengths
3
4.25k
rejected
stringlengths
39
9.29k
**LLM-Enhanced IR Models.** As another major research direction, LLMs can be employed to improve existing IR models (_e.g._, small models). A common challenge faced by existing IR models is the lack of relevant judgment annotation [786, 787]. To tackle this problem, LLMs can be instructed to annotate positive or negative documents for a given query [788], or to generate corresponding queries based on a set of documents in the corpus by referring to a few demonstrations [789, 790]. In addition to training data augmentation, LLM has the potential to improve existing IR models by refining the search-oriented informativeness of both queries and documents. In IR systems, the input queries may be constrained by a user's cognitive and cultural competency, making it challenging to accurately express the real intent, and irrelevant content present in documents can also impact the relevance evaluation with the query. As a solution, LLM can be utilized to rewrite the query for enhancing the understanding of the query intent and incorporating additional knowledge into the query through well-designed instructions. The rewritten query can take the form of an improved version of the original query [791], a document in the corpus that related to the query [792], or an expansion of the query that concatenated with a pseudo generated document [793]. In addition, documents can also be expanded with queries that are generated based on the original documents using LLMs for context extension [794].
**LLM ๊ฐ•ํ™” IR ๋ชจ๋ธ**. ๋˜ ๋‹ค๋ฅธ ์ฃผ์š” ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์œผ๋กœ, LLM์€ ๊ธฐ์กด IR ๋ชจ๋ธ(_์˜ˆ: ์†Œํ˜• ๋ชจ๋ธ)์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด IR ๋ชจ๋ธ์ด ์ง๋ฉดํ•œ ๊ณตํ†ต์ ์ธ ๋ฌธ์ œ๋Š” ๊ด€๋ จ ํŒ๋‹จ ์ฃผ์„์ด ๋ถ€์กฑํ•˜๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค[786, 787]. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด LLM์€ ์ฃผ์–ด์ง„ ์ฟผ๋ฆฌ์— ๋Œ€ํ•ด ๊ธ์ • ๋˜๋Š” ๋ถ€์ • ๋ฌธ์„œ๋ฅผ ์ฃผ์„์œผ๋กœ ๋‹ฌ๊ฑฐ๋‚˜, ๋ช‡ ๊ฐ€์ง€ ๋ฐ๋ชจ๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ์ฝ”ํผ์Šค์˜ ๋ฌธ์„œ ์ง‘ํ•ฉ์— ๋”ฐ๋ผ ํ•ด๋‹น ์ฟผ๋ฆฌ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ์ง€์‹œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[788, 789, 790]. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ์™ธ์—๋„ LLM์€ ์ฟผ๋ฆฌ์™€ ๋ฌธ์„œ ๋ชจ๋‘์˜ ๊ฒ€์ƒ‰ ์ง€ํ–ฅ์  ์ •๋ณด์„ฑ์„ ๊ฐœ์„ ํ•˜์—ฌ ๊ธฐ์กด IR ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. IR ์‹œ์Šคํ…œ์—์„œ ์ž…๋ ฅ ์ฟผ๋ฆฌ๋Š” ์‚ฌ์šฉ์ž์˜ ์ธ์ง€ ๋ฐ ๋ฌธํ™”์  ์—ญ๋Ÿ‰์— ์˜ํ•ด ์ œ์•ฝ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์‹ค์ œ ์˜๋„๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ํ‘œํ˜„ํ•˜๊ธฐ ์–ด๋ ต๊ณ , ๋ฌธ์„œ์— ์กด์žฌํ•˜๋Š” ๊ด€๋ จ ์—†๋Š” ์ฝ˜ํ…์ธ ๋„ ์ฟผ๋ฆฌ์™€์˜ ๊ด€๋ จ์„ฑ ํ‰๊ฐ€์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์— ๋Œ€ํ•œ ํ•ด๊ฒฐ์ฑ…์œผ๋กœ LLM์„ ํ™œ์šฉํ•˜์—ฌ ์ฟผ๋ฆฌ ์˜๋„์˜ ์ดํ•ด๋ฅผ ๋†’์ด๊ณ  ์ž˜ ์„ค๊ณ„๋œ instructions์„ ํ†ตํ•ด ์ฟผ๋ฆฌ์— ์ถ”๊ฐ€ ์ง€์‹์„ ํ†ตํ•ฉํ•˜๊ธฐ ์œ„ํ•ด ์ฟผ๋ฆฌ๋ฅผ ๋‹ค์‹œ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์žฌ์ž‘์„ฑ๋œ ์ฟผ๋ฆฌ๋Š” ์›๋ž˜ ์ฟผ๋ฆฌ์˜ ๊ฐœ์„ ๋œ ๋ฒ„์ „[791], ์ฟผ๋ฆฌ์™€ ๊ด€๋ จ๋œ ์ฝ”ํผ์Šค์˜ ๋ฌธ์„œ[792], ๋˜๋Š” ์˜์‚ฌ ์ƒ์„ฑ ๋ฌธ์„œ์™€ ์—ฐ๊ฒฐ๋œ ์ฟผ๋ฆฌ์˜ ํ™•์žฅ[793]์˜ ํ˜•ํƒœ๋ฅผ ์ทจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฌธ๋งฅ ํ™•์žฅ์„ ์œ„ํ•ด LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์›๋ณธ ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ์„ฑ๋œ ์ฟผ๋ฆฌ๋กœ ๋ฌธ์„œ๋ฅผ ํ™•์žฅํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค[794].
**LLM ๊ฐ•ํ™” IR ๋ชจ๋ธ**. ๋˜ ๋‹ค๋ฅธ ์ฃผ์š” ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์œผ๋กœ, LLM์€ ๊ธฐ์กด IR ๋ชจ๋ธ(_์˜ˆ: ์†Œํ˜• ๋ชจ๋ธ)์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด IR ๋ชจ๋ธ์ด ์ง๋ฉดํ•œ ๊ณตํ†ต์ ์ธ ๋ฌธ์ œ๋Š” ๊ด€๋ จ ํŒ๋‹จ ์ฃผ์„์ด ๋ถ€์กฑํ•˜๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค[786, 787]. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด LLM์€ ์ฃผ์–ด์ง„ ์ฟผ๋ฆฌ์— ๋Œ€ํ•ด ๊ธ์ • ๋˜๋Š” ๋ถ€์ • ๋ฌธ์„œ๋ฅผ ์ฃผ์„์œผ๋กœ ๋‹ฌ๊ฑฐ๋‚˜, ๋ช‡ ๊ฐ€์ง€ ๋ฐ๋ชจ๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ์ฝ”ํผ์Šค์˜ ๋ฌธ์„œ ์ง‘ํ•ฉ์— ๋”ฐ๋ผ ํ•ด๋‹น ์ฟผ๋ฆฌ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ์ง€์‹œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[788, 789, 790]. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ์™ธ์—๋„ LLM์€ ์ฟผ๋ฆฌ์™€ ๋ฌธ์„œ ๋ชจ๋‘์˜ ๊ฒ€์ƒ‰ ์ง€ํ–ฅ์  ์ •๋ณด์„ฑ์„ ๊ฐœ์„ ํ•˜์—ฌ ๊ธฐ์กด IR ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. IR ์‹œ์Šคํ…œ์—์„œ ์ž…๋ ฅ ์ฟผ๋ฆฌ๋Š” ์‚ฌ์šฉ์ž์˜ ์ธ์ง€ ๋ฐ ๋ฌธํ™”์  ์—ญ๋Ÿ‰์— ์˜ํ•ด ์ œ์•ฝ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์‹ค์ œ ์˜๋„๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ํ‘œํ˜„ํ•˜๊ธฐ ์–ด๋ ต๊ณ , ๋ฌธ์„œ์— ์กด์žฌํ•˜๋Š” ๊ด€๋ จ ์—†๋Š” ์ฝ˜ํ…์ธ ๋„ ์ฟผ๋ฆฌ์™€์˜ ๊ด€๋ จ์„ฑ ํ‰๊ฐ€์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์— ๋Œ€ํ•œ ํ•ด๊ฒฐ์ฑ…์œผ๋กœ LLM์„ ํ™œ์šฉํ•˜์—ฌ ์ฟผ๋ฆฌ ์˜๋„์˜ ์ดํ•ด๋ฅผ ๋†’์ด๊ณ  ์ž˜ ์„ค๊ณ„๋œ ์ง€์นจ์„ ํ†ตํ•ด ์ฟผ๋ฆฌ์— ์ถ”๊ฐ€ ์ง€์‹์„ ํ†ตํ•ฉํ•˜๊ธฐ ์œ„ํ•ด ์ฟผ๋ฆฌ๋ฅผ ๋‹ค์‹œ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์žฌ์ž‘์„ฑ๋œ ์ฟผ๋ฆฌ๋Š” ์›๋ž˜ ์ฟผ๋ฆฌ์˜ ๊ฐœ์„ ๋œ ๋ฒ„์ „[791], ์ฟผ๋ฆฌ์™€ ๊ด€๋ จ๋œ ์ฝ”ํผ์Šค์˜ ๋ฌธ์„œ[792], ๋˜๋Š” ์˜์‚ฌ ์ƒ์„ฑ ๋ฌธ์„œ์™€ ์—ฐ๊ฒฐ๋œ ์ฟผ๋ฆฌ์˜ ํ™•์žฅ[793]์˜ ํ˜•ํƒœ๋ฅผ ์ทจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฌธ๋งฅ ํ™•์žฅ์„ ์œ„ํ•ด LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์›๋ณธ ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ์„ฑ๋œ ์ฟผ๋ฆฌ๋กœ ๋ฌธ์„œ๋ฅผ ํ™•์žฅํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค[794].
What does infinite training data mean?The above steps require training models with a fixed set of hyperparameters on infinite data until convergence. In practical scenarios, however, data is always finite in practice. Nevertheless, it is feasible to have _relatively_ "infinite training data" for a given model. Say we have an extremely small model with only \(N=10\) parameters. For this micro model, due to its very limited capacity, no matter how much training data we use to train it (_e.g._, 1B, 2B, or 1T tokens5), the model's performance cannot be improved, and the training dynamic (_i.e._, the test loss at any given training step) is almost the same. This is to say, beyond a _critical size_, this micro model is unaware of the growth of the data size. In this case, we can say that the training data is _relatively_ infinite for this model. In fact, for a given parameter with \(N\) parameters, this critical size could be computed (see Eq. 4.4 in [16]). Here, we give a simple empirical method to deduce whether a data size is (relatively) infinite for a \(N\)-parameter model. Note that overfitting should not happen when the training data is (relatively) infinite. As a result, we can train the \(N\)-parameter model and observe whether the training and test losses diverge. If the dynamics of the training and test losses are almost the same everywhere (_i.e._, no overfitting), we can safely infer that the training data is (relatively) infinite for the given \(N\)-parameter model.
๋ฌดํ•œํ•œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ž€ ๋ฌด์—‡์„ ์˜๋ฏธํ•˜๋‚˜์š”?์œ„์˜ ๋‹จ๊ณ„๋Š” ์ˆ˜๋ ดํ•  ๋•Œ๊นŒ์ง€ ๋ฌดํ•œํ•œ ๋ฐ์ดํ„ฐ๋กœ ๊ณ ์ •๋œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์„ธํŠธ๋กœ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ํ•ญ์ƒ ์œ ํ•œํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ฃผ์–ด์ง„ ๋ชจ๋ธ์— ๋Œ€ํ•ด _์ƒ๋Œ€์ ์œผ๋กœ_ "๋ฌดํ•œํ•œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ"๋ฅผ ๊ฐ–๋Š” ๊ฒƒ์€ ์‹คํ˜„ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ \(N=10\) ๊ฐœ๋ฟ์ธ ๋งค์šฐ ์ž‘์€ ๋ชจ๋ธ์ด ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด ๋งˆ์ดํฌ๋กœ ๋ชจ๋ธ์€ ์šฉ๋Ÿ‰์ด ๋งค์šฐ ์ œํ•œ๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— 1B, 2B ๋˜๋Š” 1T ํ† ํฐ5์™€ ๊ฐ™์ด ์•„๋ฌด๋ฆฌ ๋งŽ์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•ด๋„ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์—†์œผ๋ฉฐ ํ›ˆ๋ จ ๋™์ (_์ฆ‰, ์ฃผ์–ด์ง„ ํ›ˆ๋ จ ๋‹จ๊ณ„์—์„œ์˜ ํ…Œ์ŠคํŠธ Loss)๋„ ๊ฑฐ์˜ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์ž„๊ณ„ ํฌ๊ธฐ_๋ฅผ ๋„˜์–ด์„œ๋ฉด ์ด ๋งˆ์ดํฌ๋กœ ๋ชจ๋ธ์€ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ์˜ ์ฆ๊ฐ€๋ฅผ ์ธ์‹ํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๊ฐ€ ์ด ๋ชจ๋ธ์— ๋Œ€ํ•ด _์ƒ๋Œ€์ ์œผ๋กœ_ ๋ฌดํ•œ๋Œ€๋ผ๊ณ  ๋งํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ, N๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ์žˆ๋Š” ํŠน์ • ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๊ฒฝ์šฐ ์ด ์ž„๊ณ„ ํฌ๊ธฐ๋Š” ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(16]์˜ Eq. 4.4 ์ฐธ์กฐ). ์—ฌ๊ธฐ์„œ๋Š” ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๊ฐ€ (์ƒ๋Œ€์ ์œผ๋กœ) ๋ฌดํ•œํ•œ์ง€ ์—ฌ๋ถ€๋ฅผ \(N\) ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ์ถ”๋ก ํ•˜๋Š” ๊ฐ„๋‹จํ•œ ๊ฒฝํ—˜์  ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๊ฐ€ (์ƒ๋Œ€์ ์œผ๋กœ) ๋ฌดํ•œ๋Œ€์ผ ๋•Œ๋Š” ๊ณผ์ ํ•ฉ์ด ๋ฐœ์ƒํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ, \(N\) ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ํ›ˆ๋ จ ๋ฐ ํ…Œ์ŠคํŠธ Loss๊ฐ€ ๊ฐˆ๋ผ์ง€๋Š”์ง€ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ Loss์™€ ํ…Œ์ŠคํŠธ Loss์˜ ๋™์—ญํ•™์ด ๋ชจ๋“  ๊ณณ์—์„œ ๊ฑฐ์˜ ๋™์ผํ•˜๋‹ค๋ฉด(_์ฆ‰, ๊ณผ์ ํ•ฉ์ด ๋ฐœ์ƒํ•˜์ง€ ์•Š์Œ_), ์ฃผ์–ด์ง„ \(N\) ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๊ฐ€ (์ƒ๋Œ€์ ์œผ๋กœ) ๋ฌดํ•œ๋Œ€๋ผ๊ณ  ์•ˆ์ „ํ•˜๊ฒŒ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ฌดํ•œํ•œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ž€ ๋ฌด์—‡์„ ์˜๋ฏธํ•˜๋‚˜์š”?์œ„์˜ ๋‹จ๊ณ„๋Š” ์ˆ˜๋ ดํ•  ๋•Œ๊นŒ์ง€ ๋ฌดํ•œํ•œ ๋ฐ์ดํ„ฐ๋กœ ๊ณ ์ •๋œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์„ธํŠธ๋กœ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ํ•ญ์ƒ ์œ ํ•œํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ฃผ์–ด์ง„ ๋ชจ๋ธ์— ๋Œ€ํ•ด _์ƒ๋Œ€์ ์œผ๋กœ_ "๋ฌดํ•œํ•œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ"๋ฅผ ๊ฐ–๋Š” ๊ฒƒ์€ ์‹คํ˜„ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ \(N=10\) ๊ฐœ๋ฟ์ธ ๋งค์šฐ ์ž‘์€ ๋ชจ๋ธ์ด ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด ๋งˆ์ดํฌ๋กœ ๋ชจ๋ธ์€ ์šฉ๋Ÿ‰์ด ๋งค์šฐ ์ œํ•œ๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— 1B, 2B ๋˜๋Š” 1T ํ† ํฐ5์™€ ๊ฐ™์ด ์•„๋ฌด๋ฆฌ ๋งŽ์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•ด๋„ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์—†์œผ๋ฉฐ ํ›ˆ๋ จ ๋™์ (_์ฆ‰, ์ฃผ์–ด์ง„ ํ›ˆ๋ จ ๋‹จ๊ณ„์—์„œ์˜ ํ…Œ์ŠคํŠธ ์†์‹ค)๋„ ๊ฑฐ์˜ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์ž„๊ณ„ ํฌ๊ธฐ_๋ฅผ ๋„˜์–ด์„œ๋ฉด ์ด ๋งˆ์ดํฌ๋กœ ๋ชจ๋ธ์€ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ์˜ ์ฆ๊ฐ€๋ฅผ ์ธ์‹ํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๊ฐ€ ์ด ๋ชจ๋ธ์— ๋Œ€ํ•ด _์ƒ๋Œ€์ ์œผ๋กœ_ ๋ฌดํ•œ๋Œ€๋ผ๊ณ  ๋งํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ, N๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ์žˆ๋Š” ํŠน์ • ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๊ฒฝ์šฐ ์ด ์ž„๊ณ„ ํฌ๊ธฐ๋Š” ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(16]์˜ Eq. 4.4 ์ฐธ์กฐ). ์—ฌ๊ธฐ์„œ๋Š” ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๊ฐ€ (์ƒ๋Œ€์ ์œผ๋กœ) ๋ฌดํ•œํ•œ์ง€ ์—ฌ๋ถ€๋ฅผ \(N\) ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ์ถ”๋ก ํ•˜๋Š” ๊ฐ„๋‹จํ•œ ๊ฒฝํ—˜์  ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๊ฐ€ (์ƒ๋Œ€์ ์œผ๋กœ) ๋ฌดํ•œ๋Œ€์ผ ๋•Œ๋Š” ๊ณผ์ ํ•ฉ์ด ๋ฐœ์ƒํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ, \(N\) ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ํ›ˆ๋ จ ๋ฐ ํ…Œ์ŠคํŠธ ์†์‹ค์ด ๊ฐˆ๋ผ์ง€๋Š”์ง€ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์†์‹ค๊ณผ ํ…Œ์ŠคํŠธ ์†์‹ค์˜ ๋™์—ญํ•™์ด ๋ชจ๋“  ๊ณณ์—์„œ ๊ฑฐ์˜ ๋™์ผํ•˜๋‹ค๋ฉด(_์ฆ‰, ๊ณผ์ ํ•ฉ์ด ๋ฐœ์ƒํ•˜์ง€ ์•Š์Œ_), ์ฃผ์–ด์ง„ \(N\) ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๊ฐ€ (์ƒ๋Œ€์ ์œผ๋กœ) ๋ฌดํ•œ๋Œ€๋ผ๊ณ  ์•ˆ์ „ํ•˜๊ฒŒ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 1: **Overview of contributions.** AMIE is a conversational medical AI optimised for diagnostic dialogue. AMIE is instruction fine-tuned with a combination of real-world and simulated medical dialogues, alongside a diverse set of medical reasoning, question answering, and summarization datasets. Notably, we designed a self-play based simulated dialogue environment with automated feedback mechanisms to scale AMIEโ€™s capabilities across various medical contexts and specialities. Specifically, this iterative self-improvement process consisted of two self-play loops: (1) An โ€œinnerโ€ self-play loop, where AMIE leveraged in-context critic feedback to refine its behavior on simulated conversations with an AI patient agent; (2) An โ€œouterโ€ self-play loop where the set of refined simulated dialogues were incorporated into subsequent fine-tuning iterations. During online inference, AMIE used a chain-of-reasoning strategy to progressively refine its response conditioned on the current conversation to arrive at an accurate and grounded reply to the patient in each dialogue turn. We designed and conducted a blinded remote Objective Structured Clinical Examination (OSCE) with validated simulated patient actors interacting with AMIE or Primary Care Physicians (PCPs) via a text interface. Across multiple axes corresponding to both specialist physician (28 out of 32) and patient actor (24 out of 26) perspective, AMIE was rated as superior to PCPs while being non-inferior on the rest.
๊ทธ๋ฆผ 1: **๊ธฐ์—ฌ๋„ ๊ฐœ์š”.** AMIE๋Š” ์ง„๋‹จ ๋Œ€ํ™”์— ์ตœ์ ํ™”๋œ ๋Œ€ํ™”ํ˜• ์˜๋ฃŒ AI์ž…๋‹ˆ๋‹ค. AMIE๋Š” ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ์ถ”๋ก , question answering ๋ฐ ์š”์•ฝ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ํ•จ๊ป˜ ์‹ค์ œ ์˜๋ฃŒ ๋Œ€ํ™” ๋ฐ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ์˜๋ฃŒ ๋Œ€ํ™”์˜ ์กฐํ•ฉ์œผ๋กœ Instruction Fine-tuning๋ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ์ƒํ™ฉ๊ณผ ์ „๋ฌธ ๋ถ„์•ผ์— ๊ฑธ์ณ AMIE์˜ ์—ญ๋Ÿ‰์„ ํ™•์žฅํ•˜๊ธฐ ์œ„ํ•ด ์ž๋™ํ™”๋œ ํ”ผ๋“œ๋ฐฑ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ฐ–์ถ˜ self-play ๊ธฐ๋ฐ˜์˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™” ํ™˜๊ฒฝ์„ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์ด๋Ÿฌํ•œ ๋ฐ˜๋ณต์ ์ธ ์ž๊ธฐ ๊ฐœ์„  ํ”„๋กœ์„ธ์Šค๋Š” ๋‘ ๊ฐ€์ง€ self-play ๋ฃจํ”„๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค: (1) AMIE๊ฐ€ ์ธ๊ณต์ง€๋Šฅ ํ™˜์ž ์—์ด์ „ํŠธ์™€์˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™”์—์„œ ์ž์‹ ์˜ ํ–‰๋™์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด in-context critic feedback์„ ํ™œ์šฉํ•˜๋Š” "๋‚ด๋ถ€" self-play ๋ฃจํ”„, (2) ๊ฐœ์„ ๋œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™” ์„ธํŠธ๊ฐ€ ํ›„์† Fine-tuning ๋ฐ˜๋ณต์— ํ†ตํ•ฉ๋˜๋Š” "์™ธ๋ถ€" self-play ๋ฃจํ”„. ์˜จ๋ผ์ธ ์ถ”๋ก  ์ค‘์— AMIE๋Š” chain-of-reasoning ์ „๋žต์„ ์‚ฌ์šฉํ•˜์—ฌ ํ˜„์žฌ ๋Œ€ํ™”์— ๋”ฐ๋ผ ์‘๋‹ต์„ ์ ์ง„์ ์œผ๋กœ ๊ฐœ์„ ํ•˜์—ฌ ๊ฐ ๋Œ€ํ™” ํ„ด์—์„œ ํ™˜์ž์—๊ฒŒ ์ •ํ™•ํ•˜๊ณ  ๊ทผ๊ฑฐ ์žˆ๋Š” ๋‹ต๋ณ€์„ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๊ฒ€์ฆ๋œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜์ž ๋ฐฐ์šฐ๊ฐ€ ํ…์ŠคํŠธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ํ†ตํ•ด AMIE ๋˜๋Š” ์ฃผ์น˜์˜(PCP)์™€ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ๋งน๊ฒ€ ์›๊ฒฉ Objective Structured Clinical Examination (OSCE)์„ ์„ค๊ณ„ํ•˜๊ณ  ์‹ค์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฌธ์˜(32๋ช… ์ค‘ 28๋ช…)์™€ ํ™˜์ž ๋ฐฐ์šฐ(26๋ช… ์ค‘ 24๋ช…)์˜ ๊ด€์ ์— ํ•ด๋‹นํ•˜๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ถ•์—์„œ AMIE๋Š” PCP๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ๊ฒƒ์œผ๋กœ ํ‰๊ฐ€๋˜์—ˆ์œผ๋ฉฐ, ๋‚˜๋จธ์ง€ ์ถ•์—์„œ๋Š” ์—ด๋“ฑํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1: **๊ธฐ์—ฌ๋„ ๊ฐœ์š”.** AMIE๋Š” ์ง„๋‹จ ๋Œ€ํ™”์— ์ตœ์ ํ™”๋œ ๋Œ€ํ™”ํ˜• ์˜๋ฃŒ AI์ž…๋‹ˆ๋‹ค. AMIE๋Š” ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ์ถ”๋ก , ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ ์š”์•ฝ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ํ•จ๊ป˜ ์‹ค์ œ ์˜๋ฃŒ ๋Œ€ํ™” ๋ฐ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ์˜๋ฃŒ ๋Œ€ํ™”์˜ ์กฐํ•ฉ์œผ๋กœ ์ง€์‹œ ๋ฏธ์„ธ ์กฐ์ •๋ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ์ƒํ™ฉ๊ณผ ์ „๋ฌธ ๋ถ„์•ผ์— ๊ฑธ์ณ AMIE์˜ ์—ญ๋Ÿ‰์„ ํ™•์žฅํ•˜๊ธฐ ์œ„ํ•ด ์ž๋™ํ™”๋œ ํ”ผ๋“œ๋ฐฑ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ฐ–์ถ˜ ์…€ํ”„ ํ”Œ๋ ˆ์ด ๊ธฐ๋ฐ˜์˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™” ํ™˜๊ฒฝ์„ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์ด๋Ÿฌํ•œ ๋ฐ˜๋ณต์ ์ธ ์ž๊ธฐ ๊ฐœ์„  ํ”„๋กœ์„ธ์Šค๋Š” ๋‘ ๊ฐ€์ง€ ์ž๊ธฐ ํ”Œ๋ ˆ์ด ๋ฃจํ”„๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค: (1) AMIE๊ฐ€ ์ธ๊ณต์ง€๋Šฅ ํ™˜์ž ์—์ด์ „ํŠธ์™€์˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™”์—์„œ ์ž์‹ ์˜ ํ–‰๋™์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ปจํ…์ŠคํŠธ ๋‚ด ๋น„ํ‰๊ฐ€ ํ”ผ๋“œ๋ฐฑ์„ ํ™œ์šฉํ•˜๋Š” "๋‚ด๋ถ€" ์ž๊ธฐ ํ”Œ๋ ˆ์ด ๋ฃจํ”„, (2) ๊ฐœ์„ ๋œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™” ์„ธํŠธ๊ฐ€ ํ›„์† ๋ฏธ์„ธ ์กฐ์ • ๋ฐ˜๋ณต์— ํ†ตํ•ฉ๋˜๋Š” "์™ธ๋ถ€" ์ž๊ธฐ ํ”Œ๋ ˆ์ด ๋ฃจํ”„. ์˜จ๋ผ์ธ ์ถ”๋ก  ์ค‘์— AMIE๋Š” ์—ฐ์‡„ ์ถ”๋ก  ์ „๋žต์„ ์‚ฌ์šฉํ•˜์—ฌ ํ˜„์žฌ ๋Œ€ํ™”์— ๋”ฐ๋ผ ์‘๋‹ต์„ ์ ์ง„์ ์œผ๋กœ ๊ฐœ์„ ํ•˜์—ฌ ๊ฐ ๋Œ€ํ™” ํ„ด์—์„œ ํ™˜์ž์—๊ฒŒ ์ •ํ™•ํ•˜๊ณ  ๊ทผ๊ฑฐ ์žˆ๋Š” ๋‹ต๋ณ€์„ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๊ฒ€์ฆ๋œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜์ž ๋ฐฐ์šฐ๊ฐ€ ํ…์ŠคํŠธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ํ†ตํ•ด AMIE ๋˜๋Š” ์ฃผ์น˜์˜(PCP)์™€ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ๋งน๊ฒ€ ์›๊ฒฉ ๊ฐ๊ด€์  ๊ตฌ์กฐํ™” ์ž„์ƒ ์‹œํ—˜(OSCE)์„ ์„ค๊ณ„ํ•˜๊ณ  ์‹ค์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฌธ์˜(32๋ช… ์ค‘ 28๋ช…)์™€ ํ™˜์ž ๋ฐฐ์šฐ(26๋ช… ์ค‘ 24๋ช…)์˜ ๊ด€์ ์— ํ•ด๋‹นํ•˜๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ถ•์—์„œ AMIE๋Š” PCP๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ๊ฒƒ์œผ๋กœ ํ‰๊ฐ€๋˜์—ˆ์œผ๋ฉฐ, ๋‚˜๋จธ์ง€ ์ถ•์—์„œ๋Š” ์—ด๋“ฑํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
Despite the considerable strides made by LLaMA and Alpaca in NLP, they exhibit inherent limitations concerning native support for Chinese language tasks. Their vocabularies contain only a few hundred Chinese tokens, substantially hindering their efficiency in encoding and decoding Chinese text. Building on our previous work with the Chinese BERT series (Cui et al., 2021) and Chinese minority-oriented multilingual pre-trained models (Yang et al., 2022), in this technical report, we propose the development of Chinese LLaMA and Alpaca models with enhanced capabilities for understanding and generating Chinese content. We extend the original LLaMA's vocabulary with an additional 20,000 Chinese tokens, significantly improving its proficiency in processing and generating Chinese text. To ensure efficient training and deployment of these models, we employ the Low-Rank Adaptation (LoRA) approach (Hu et al., 2021), enabling us to train and fine-tune the models without excessive computational costs. We anticipate our preliminary study to enhance the Chinese understanding and generation capabilities of LLaMA and Alpaca serves as a foundation for researchers aiming to adapt these models to other languages. By showcasing the feasibility and effectiveness of our approach, we offer insights and methodologies that can be employed to extend vocabularies and improve the performance of LLaMA and Alpaca models in various languages. An overview of the proposed models is depicted in Figure 1.
LLaMA์™€ Alpaca๋Š” ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ ์ƒ๋‹นํ•œ ๋ฐœ์ „์„ ์ด๋ฃจ์—ˆ์ง€๋งŒ, ์ค‘๊ตญ์–ด ์ž‘์—…์„ ๊ธฐ๋ณธ์ ์œผ๋กœ ์ง€์›ํ•˜๋Š” ๋ฐ๋Š” ์—ฌ์ „ํžˆ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์˜ ์–ดํœ˜์—๋Š” ์ˆ˜๋ฐฑ ๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ๋งŒ ํฌํ•จ๋˜์–ด ์žˆ์–ด ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ๋ฅผ ์ธ์ฝ”๋”ฉํ•˜๊ณ  ๋””์ฝ”๋”ฉํ•˜๋Š” ๋ฐ ํšจ์œจ์„ฑ์ด ํ˜„์ €ํžˆ ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค. ์ด์ „์— ๊ฐœ๋ฐœํ•œ ์ค‘๊ตญ์–ด BERT ์‹œ๋ฆฌ์ฆˆ(Cui et al., 2021)์™€ ์ค‘๊ตญ ์†Œ์ˆ˜ ๋ฏผ์กฑ ์ค‘์‹ฌ์˜ ๋‹ค๊ตญ์–ด ์‚ฌ์ „ ํ•™์Šต ๋ชจ๋ธ(Yang et al., 2022)์„ ๊ธฐ๋ฐ˜์œผ๋กœ, ์ด๋ฒˆ ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ์—์„œ๋Š” ์ค‘๊ตญ์–ด ์ฝ˜ํ…์ธ ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๊ธฐ๋Šฅ์„ ๊ฐ•ํ™”ํ•œ ์ค‘๊ตญ์–ด LLaMA์™€ Alpaca ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ LLaMA์˜ ์–ดํœ˜๋ฅผ 20,000๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ์œผ๋กœ ํ™•์žฅํ•˜์—ฌ ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ ์ฒ˜๋ฆฌ ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ํšจ์œจ์ ์œผ๋กœ ํ›ˆ๋ จํ•˜๊ณ  ๋ฐฐํฌํ•˜๊ธฐ ์œ„ํ•ด Low-Rank Adaptation (LoRA) ์ ‘๊ทผ ๋ฐฉ์‹(Hu et al., 2021)์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณผ๋„ํ•œ ๊ณ„์‚ฐ ๋น„์šฉ ์—†์ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  Fine-tuningํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ๊ฐ•ํ™”ํ•œ LLaMA์™€ Alpaca์˜ ์˜ˆ๋น„ ์—ฐ๊ตฌ๊ฐ€ ๋‹ค๋ฅธ ์–ธ์–ด์— ์ด ๋ชจ๋ธ์„ ์ ์šฉํ•˜๋ ค๋Š” ์—ฐ๊ตฌ์ž๋“ค์—๊ฒŒ ๋ฐ‘๊ฑฐ๋ฆ„์ด ๋  ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํƒ€๋‹น์„ฑ๊ณผ ํšจ๊ณผ๋ฅผ ์ž…์ฆํ•จ์œผ๋กœ์จ ๋‹ค์–‘ํ•œ ์–ธ์–ด์—์„œ LLaMA์™€ Alpaca ๋ชจ๋ธ์˜ ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜๊ณ  ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ธ์‚ฌ์ดํŠธ์™€ ๋ฐฉ๋ฒ•๋ก ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ œ์•ˆ๋œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ฐœ์š”๋Š” ๊ทธ๋ฆผ 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
LLaMA์™€ ์•ŒํŒŒ์นด๋Š” ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ ์ƒ๋‹นํ•œ ๋ฐœ์ „์„ ์ด๋ฃจ์—ˆ์ง€๋งŒ, ์ค‘๊ตญ์–ด ์ž‘์—…์„ ๊ธฐ๋ณธ์ ์œผ๋กœ ์ง€์›ํ•˜๋Š” ๋ฐ๋Š” ์—ฌ์ „ํžˆ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์˜ ์–ดํœ˜์—๋Š” ์ˆ˜๋ฐฑ ๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ๋งŒ ํฌํ•จ๋˜์–ด ์žˆ์–ด ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ๋ฅผ ์ธ์ฝ”๋”ฉํ•˜๊ณ  ๋””์ฝ”๋”ฉํ•˜๋Š” ๋ฐ ํšจ์œจ์„ฑ์ด ํ˜„์ €ํžˆ ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค. ์ด์ „์— ๊ฐœ๋ฐœํ•œ ์ค‘๊ตญ์–ด BERT ์‹œ๋ฆฌ์ฆˆ(Cui et al., 2021)์™€ ์ค‘๊ตญ ์†Œ์ˆ˜ ๋ฏผ์กฑ ์ค‘์‹ฌ์˜ ๋‹ค๊ตญ์–ด ์‚ฌ์ „ ํ•™์Šต ๋ชจ๋ธ(Yang et al., 2022)์„ ๊ธฐ๋ฐ˜์œผ๋กœ, ์ด๋ฒˆ ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ์—์„œ๋Š” ์ค‘๊ตญ์–ด ์ฝ˜ํ…์ธ ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๊ธฐ๋Šฅ์„ ๊ฐ•ํ™”ํ•œ ์ค‘๊ตญ์–ด LLaMA์™€ ์•ŒํŒŒ์นด ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ LLaMA์˜ ์–ดํœ˜๋ฅผ 20,000๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ์œผ๋กœ ํ™•์žฅํ•˜์—ฌ ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ ์ฒ˜๋ฆฌ ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ํšจ์œจ์ ์œผ๋กœ ํ›ˆ๋ จํ•˜๊ณ  ๋ฐฐํฌํ•˜๊ธฐ ์œ„ํ•ด ์ €๋žญ๋„ ์ ์‘(LoRA) ์ ‘๊ทผ ๋ฐฉ์‹(Hu et al., 2021)์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณผ๋„ํ•œ ๊ณ„์‚ฐ ๋น„์šฉ ์—†์ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ๋ฏธ์„ธ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ๊ฐ•ํ™”ํ•œ LLaMA์™€ ์•ŒํŒŒ์นด์˜ ์˜ˆ๋น„ ์—ฐ๊ตฌ๊ฐ€ ๋‹ค๋ฅธ ์–ธ์–ด์— ์ด ๋ชจ๋ธ์„ ์ ์šฉํ•˜๋ ค๋Š” ์—ฐ๊ตฌ์ž๋“ค์—๊ฒŒ ๋ฐ‘๊ฑฐ๋ฆ„์ด ๋  ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํƒ€๋‹น์„ฑ๊ณผ ํšจ๊ณผ๋ฅผ ์ž…์ฆํ•จ์œผ๋กœ์จ ๋‹ค์–‘ํ•œ ์–ธ์–ด์—์„œ LLaMA์™€ ์•ŒํŒŒ์นด ๋ชจ๋ธ์˜ ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜๊ณ  ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ธ์‚ฌ์ดํŠธ์™€ ๋ฐฉ๋ฒ•๋ก ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ œ์•ˆ๋œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ฐœ์š”๋Š” ๊ทธ๋ฆผ 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
Next, we evaluate fine-tuning performance of DPO on summarization and single-turn dialogue. For summarization, automatic evaluation metrics such as ROUGE can be poorly correlated with human preferences [38], and prior work has found that fine-tuning LMs using PPO on human preferences to provide more effective summaries. We evaluate different methods by sampling completions on the test split of TL;DR summarization dataset, and computing the average win rate against reference completions in the test set. The completions for all methods are sampled at temperatures varying from 0.0 to 1.0, and the win rates are shown in Figure 2 (right). DPO, PPO and Preferred-FT all fine-tune the same GPT-J SFT model4. We find that DPO has a win rate of approximately 61% at a temperature of 0.0, exceeding the performance of PPO at 57% at its optimal sampling temperature of 0.0. DPO also achieves a higher maximum win rate compared to the best of \(N\) baseline. We note that we did not meaningfully tune DPO's \(\beta\) hyperparameter, so these results may underestimate DPO's potential. Moreover, we find DPO to be much more robust to the sampling temperature than PPO, the performance of which can degrade to that of the base GPT-J model at high temperatures. Preferred-FT does not improve significantly over the SFT model. We also compare DPO and PPO head-to-head in human evaluations in Section 6.4, where DPO samples at temperature 0.25 were preferred 58% times over PPO samples at temperature 0.
๋‹ค์Œ์œผ๋กœ ์š”์•ฝ ๋ฐ ๋‹จ์ผ ํ„ด ๋Œ€ํ™”์— ๋Œ€ํ•œ DPO์˜ Fine-tuning ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์š”์•ฝ์˜ ๊ฒฝ์šฐ ROUGE์™€ ๊ฐ™์€ ์ž๋™ ํ‰๊ฐ€ ๋ฉ”ํŠธ๋ฆญ์€ ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„์™€ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ๋‚ฎ์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„์— ๋Œ€ํ•ด PPO๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LM์„ Fine-tuningํ•˜์—ฌ ๋” ํšจ๊ณผ์ ์ธ ์š”์•ฝ์„ ์ œ๊ณตํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค. TL;DR ์š”์•ฝ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํ…Œ์ŠคํŠธ ๋ถ„ํ• ์—์„œ ์™„์„ฑ์„ ์ƒ˜ํ”Œ๋งํ•˜๊ณ  ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ์ฐธ์กฐ ์™„์„ฑ์— ๋Œ€ํ•œ ํ‰๊ท  ์Šน๋ฅ ์„ ๊ณ„์‚ฐํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ฐฉ๋ฒ•์˜ ์™„์„ฑ์€ 0.0์—์„œ 1.0๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์˜จ๋„์—์„œ ์ƒ˜ํ”Œ๋ง๋˜๋ฉฐ, ์Šน๋ฅ ์€ ๊ทธ๋ฆผ 2(์˜ค๋ฅธ์ชฝ)์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. DPO, PPO ๋ฐ Preferred-FT๋Š” ๋ชจ๋‘ ๋™์ผํ•œ GPT-J SFT ๋ชจ๋ธ์„ Fine-tuningํ•ฉ๋‹ˆ๋‹ค. ์˜จ๋„๊ฐ€ 0.0์ผ ๋•Œ DPO์˜ ์Šน๋ฅ ์€ ์•ฝ 61%๋กœ ์ตœ์ ์˜ ์ƒ˜ํ”Œ๋ง ์˜จ๋„์ธ 0.0์—์„œ 57%์ธ PPO์˜ ์„ฑ๋Šฅ์„ ์ดˆ๊ณผํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ DPO๋Š” ์ตœ๊ณ  \(N\) ๊ธฐ์ค€์„ ๋ณด๋‹ค ๋” ๋†’์€ ์ตœ๋Œ€ ์Šน๋ฅ ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. DPO์˜ \(\beta\) ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์˜๋ฏธ ์žˆ๊ฒŒ ์กฐ์ •ํ•˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” DPO์˜ ์ž ์žฌ๋ ฅ์„ ๊ณผ์†Œํ‰๊ฐ€ํ•œ ๊ฒƒ์ผ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ DPO๋Š” ์ƒ˜ํ”Œ๋ง ์˜จ๋„์— ๋Œ€ํ•ด PPO๋ณด๋‹ค ํ›จ์”ฌ ๋” Robustํ•˜๋ฉฐ, ์˜จ๋„๊ฐ€ ๋†’์œผ๋ฉด ์„ฑ๋Šฅ์ด ๊ธฐ๋ณธ GPT-J ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์œผ๋กœ ์ €ํ•˜๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Preferred-FT๋Š” SFT ๋ชจ๋ธ์— ๋น„ํ•ด ํฌ๊ฒŒ ๊ฐœ์„ ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์„น์…˜ 6.4์˜ ์ธ๊ฐ„ ํ‰๊ฐ€์—์„œ DPO์™€ PPO๋ฅผ ์ง์ ‘ ๋น„๊ตํ–ˆ๋Š”๋ฐ, ์˜จ๋„ 0.25์—์„œ ์ƒ˜ํ”Œ๋งํ•œ DPO๊ฐ€ ์˜จ๋„ 0์—์„œ ์ƒ˜ํ”Œ๋งํ•œ PPO๋ณด๋‹ค 58% ๋” ์„ ํ˜ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
๋‹ค์Œ์œผ๋กœ ์š”์•ฝ ๋ฐ ๋‹จ์ผ ํ„ด ๋Œ€ํ™”์— ๋Œ€ํ•œ DPO์˜ ๋ฏธ์„ธ ์กฐ์ • ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์š”์•ฝ์˜ ๊ฒฝ์šฐ ROUGE์™€ ๊ฐ™์€ ์ž๋™ ํ‰๊ฐ€ ๋ฉ”ํŠธ๋ฆญ์€ ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„์™€ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ๋‚ฎ์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„์— ๋Œ€ํ•ด PPO๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ๋” ํšจ๊ณผ์ ์ธ ์š”์•ฝ์„ ์ œ๊ณตํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค. TL;DR ์š”์•ฝ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํ…Œ์ŠคํŠธ ๋ถ„ํ• ์—์„œ ์™„์„ฑ์„ ์ƒ˜ํ”Œ๋งํ•˜๊ณ  ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ์ฐธ์กฐ ์™„์„ฑ์— ๋Œ€ํ•œ ํ‰๊ท  ์Šน๋ฅ ์„ ๊ณ„์‚ฐํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ฐฉ๋ฒ•์˜ ์™„์„ฑ์€ 0.0์—์„œ 1.0๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์˜จ๋„์—์„œ ์ƒ˜ํ”Œ๋ง๋˜๋ฉฐ, ์Šน๋ฅ ์€ ๊ทธ๋ฆผ 2(์˜ค๋ฅธ์ชฝ)์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. DPO, PPO ๋ฐ Preferred-FT๋Š” ๋ชจ๋‘ ๋™์ผํ•œ GPT-J SFT ๋ชจ๋ธ4๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์˜จ๋„๊ฐ€ 0.0์ผ ๋•Œ DPO์˜ ์Šน๋ฅ ์€ ์•ฝ 61%๋กœ ์ตœ์ ์˜ ์ƒ˜ํ”Œ๋ง ์˜จ๋„์ธ 0.0์—์„œ 57%์ธ PPO์˜ ์„ฑ๋Šฅ์„ ์ดˆ๊ณผํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ DPO๋Š” ์ตœ๊ณ  \(N\) ๊ธฐ์ค€์„ ๋ณด๋‹ค ๋” ๋†’์€ ์ตœ๋Œ€ ์Šน๋ฅ ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. DPO์˜ \(\๋ฒ ํƒ€\) ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์˜๋ฏธ ์žˆ๊ฒŒ ์กฐ์ •ํ•˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” DPO์˜ ์ž ์žฌ๋ ฅ์„ ๊ณผ์†Œํ‰๊ฐ€ํ•œ ๊ฒƒ์ผ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ DPO๋Š” ์ƒ˜ํ”Œ๋ง ์˜จ๋„์— ๋Œ€ํ•ด PPO๋ณด๋‹ค ํ›จ์”ฌ ๋” ๊ฐ•๋ ฅํ•˜๋ฉฐ, ์˜จ๋„๊ฐ€ ๋†’์œผ๋ฉด ์„ฑ๋Šฅ์ด ๊ธฐ๋ณธ GPT-J ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์œผ๋กœ ์ €ํ•˜๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Preferred-FT๋Š” SFT ๋ชจ๋ธ์— ๋น„ํ•ด ํฌ๊ฒŒ ๊ฐœ์„ ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์„น์…˜ 6.4์˜ ์ธ๊ฐ„ ํ‰๊ฐ€์—์„œ DPO์™€ PPO๋ฅผ ์ง์ ‘ ๋น„๊ตํ–ˆ๋Š”๋ฐ, ์˜จ๋„ 0.25์—์„œ ์ƒ˜ํ”Œ๋งํ•œ DPO๊ฐ€ ์˜จ๋„ 0์—์„œ ์ƒ˜ํ”Œ๋งํ•œ PPO๋ณด๋‹ค 58% ๋” ์„ ํ˜ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
* **ARC Easy (ARC-E)**: The Easy set of the AI2 Reasoning Challenge (Clark et al., 2018) features questions from the same source as ARC-C but are considered less challenging and do not require as advanced reasoning skills. * **BoolQ**: A dataset consisting of boolean (yes/no) questions, focusing on reading comprehension and general understanding of natural language text (Clark et al., 2019). * **Winogrande (Wino.)**: This dataset challenges models on common sense reasoning in a language context, focusing on pronoun disambiguation tasks (ai2, 2019). * **PIQA**: Physical Interaction Question Answering tests the understanding of everyday physical processes, an aspect of practical common sense (Bisk et al., 2020). * **HellaSwag**: This dataset evaluates a model's ability to complete scenarios in a contextually and logically coherent manner, requiring both language understanding and common sense reasoning (Zellers et al., 2019). * **TruthfulQA**: Focused on the generation of truthful, accurate answers, this dataset challenges models on their ability to discern and reproduce factually correct information (Lin et al., 2021). * **OpenBookQA (OBQA)**: OpenBookQA requires understanding a wide array of facts and concepts, thereby evaluating the model's broader knowledge and reasoning skills (Mihaylov et al., 2018). * **LogiQA-2**: This dataset involves logical reasoning, testing the model's capability to understand and apply logical constructs and principles (Liu et al., 2023b).
* **ARC Easy(ARC-E)**: AI2 ์ถ”๋ก  ์ฑŒ๋ฆฐ์ง€(Clark et al., 2018)์˜ Easy ์„ธํŠธ๋Š” ARC-C์™€ ๋™์ผํ•œ ์ถœ์ฒ˜์˜ ๋ฌธ์ œ๋ฅผ ์ถœ์ œํ•˜์ง€๋งŒ ๋œ ์–ด๋ ค์šด ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋˜๋ฉฐ ๊ณ ๊ธ‰ ์ถ”๋ก  ๊ธฐ์ˆ ์ด ํ•„์š”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. * **BoolQ**: ๋…ํ•ด๋ ฅ๊ณผ ์ž์—ฐ์–ด ํ…์ŠคํŠธ์˜ ์ผ๋ฐ˜์ ์ธ ์ดํ•ด์— ์ค‘์ ์„ ๋‘” ๋ถ€์šธ(์˜ˆ/์•„๋‹ˆ์˜ค) ์งˆ๋ฌธ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ(Clark et al., 2019). * **Winogrande(Wino.)**: ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋Œ€๋ช…์‚ฌ ํ•ด๋… ์ž‘์—…์— ์ค‘์ ์„ ๋‘” ์–ธ์–ด ๋งฅ๋ฝ์—์„œ ์ƒ์‹ ์ถ”๋ก ์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ๋„์ „์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค(ai2, 2019). * **PIQA**: ์ผ์ƒ์ ์ธ ๋ฌผ๋ฆฌ์  ๊ณผ์ •์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฌผ๋ฆฌ์  ์ƒํ˜ธ์ž‘์šฉ ์งˆ๋ฌธ ๋‹ต๋ณ€์€ ์‹ค์šฉ์ ์ธ ์ƒ์‹์˜ ํ•œ ์ธก๋ฉด์ธ ์ผ์ƒ์ ์ธ ๋ฌผ๋ฆฌ์  ๊ณผ์ •์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค(Bisk et al., 2020). * **HellaSwag**: ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋ฌธ๋งฅ์ ์ด๊ณ  ๋…ผ๋ฆฌ์ ์œผ๋กœ ์ผ๊ด€๋œ ๋ฐฉ์‹์œผ๋กœ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์™„์„ฑํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๋ฉฐ, ์–ธ์–ด ์ดํ•ด์™€ ์ƒ์‹ ์ถ”๋ก ์ด ๋ชจ๋‘ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค(Zellers et al., 2019). * **TruthfulQA**: ์ง„์‹คํ•˜๊ณ  ์ •ํ™•ํ•œ ๋‹ต๋ณ€ ์ƒ์„ฑ์— ์ค‘์ ์„ ๋‘” ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์‚ฌ์‹ค์ ์œผ๋กœ ์ •ํ™•ํ•œ ์ •๋ณด๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ์žฌํ˜„ํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์— ๋„์ „ํ•ฉ๋‹ˆ๋‹ค(Lin et al., 2021). * **OpenBookQA(OBQA)**: OpenBookQA๋Š” ๋‹ค์–‘ํ•œ ์‚ฌ์‹ค๊ณผ ๊ฐœ๋…์„ ์ดํ•ดํ•ด์•ผ ํ•˜๋ฏ€๋กœ ๋ชจ๋ธ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ์ง€์‹๊ณผ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค(Mihaylov et al., 2018). * **LogiQA-2**: ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋…ผ๋ฆฌ์  ์ถ”๋ก ์„ ํฌํ•จํ•˜๋ฉฐ, ๋ชจ๋ธ์ด ๋…ผ๋ฆฌ์  ๊ตฌ์„ฑ๊ณผ ์›๋ฆฌ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ ์šฉํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค(Liu et al., 2023b).
* **ARC Easy(ARC-E)**: AI2 ์ถ”๋ก  ์ฑŒ๋ฆฐ์ง€(Clark et al., 2018)์˜ Easy ์„ธํŠธ๋Š” ARC-C์™€ ๋™์ผํ•œ ์ถœ์ฒ˜์˜ ๋ฌธ์ œ๋ฅผ ์ถœ์ œํ•˜์ง€๋งŒ ๋œ ์–ด๋ ค์šด ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋˜๋ฉฐ ๊ณ ๊ธ‰ ์ถ”๋ก  ๊ธฐ์ˆ ์ด ํ•„์š”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. * **BoolQ**: ๋…ํ•ด๋ ฅ๊ณผ ์ž์—ฐ์–ด ํ…์ŠคํŠธ์˜ ์ผ๋ฐ˜์ ์ธ ์ดํ•ด์— ์ค‘์ ์„ ๋‘” ๋ถ€์šธ(์˜ˆ/์•„๋‹ˆ์˜ค) ์งˆ๋ฌธ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ(Clark et al., 2019). * **Winogrande(Wino.)**: ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋Œ€๋ช…์‚ฌ ํ•ด๋… ์ž‘์—…์— ์ค‘์ ์„ ๋‘” ์–ธ์–ด ๋งฅ๋ฝ์—์„œ ์ƒ์‹ ์ถ”๋ก ์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ๋„์ „์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค(ai2, 2019). * **PIQA**: ์ผ์ƒ์ ์ธ ๋ฌผ๋ฆฌ์  ๊ณผ์ •์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฌผ๋ฆฌ์  ์ƒํ˜ธ์ž‘์šฉ ์งˆ๋ฌธ ๋‹ต๋ณ€์€ ์‹ค์šฉ์ ์ธ ์ƒ์‹์˜ ํ•œ ์ธก๋ฉด์ธ ์ผ์ƒ์ ์ธ ๋ฌผ๋ฆฌ์  ๊ณผ์ •์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค(Bisk et al., 2020). * **ํ—ฌ๋ผ์Šค์›จ๊ทธ**: ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋ฌธ๋งฅ์ ์ด๊ณ  ๋…ผ๋ฆฌ์ ์œผ๋กœ ์ผ๊ด€๋œ ๋ฐฉ์‹์œผ๋กœ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์™„์„ฑํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๋ฉฐ, ์–ธ์–ด ์ดํ•ด์™€ ์ƒ์‹ ์ถ”๋ก ์ด ๋ชจ๋‘ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค(Zellers et al., 2019). * **์ง„์‹คํ•œQA**: ์ง„์‹คํ•˜๊ณ  ์ •ํ™•ํ•œ ๋‹ต๋ณ€ ์ƒ์„ฑ์— ์ค‘์ ์„ ๋‘” ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์‚ฌ์‹ค์ ์œผ๋กœ ์ •ํ™•ํ•œ ์ •๋ณด๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ์žฌํ˜„ํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์— ๋„์ „ํ•ฉ๋‹ˆ๋‹ค(Lin et al., 2021). * **OpenBookQA(OBQA)**: OpenBookQA๋Š” ๋‹ค์–‘ํ•œ ์‚ฌ์‹ค๊ณผ ๊ฐœ๋…์„ ์ดํ•ดํ•ด์•ผ ํ•˜๋ฏ€๋กœ ๋ชจ๋ธ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ์ง€์‹๊ณผ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค(Mihaylov et al., 2018). * **LogiQA-2**: ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋…ผ๋ฆฌ์  ์ถ”๋ก ์„ ํฌํ•จํ•˜๋ฉฐ, ๋ชจ๋ธ์ด ๋…ผ๋ฆฌ์  ๊ตฌ์„ฑ๊ณผ ์›๋ฆฌ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ ์šฉํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค(Liu et al., 2023b).
To evaluate this ability, existing studies consider multiple criteria for human alignment, such as helpfulness, honesty, and safety [368, 46, 170]. For helpfulness and honesty, adversarial question answering tasks (_e.g._, TruthfulQA [556]) can be utilized to examine LLM's ability in detecting possible falsehood in the text [36, 81]. Furthermore, harmlessness can be also evaluated by several existing benchmarks, _e.g._, CrowS-Pairs [603] and Winogender [604]. Despite the automatic evaluation with the above datasets, human evaluation is still a more direct way to effectively test the human alignment ability of LLMs. OpenAI invites many experts in domains related to AI risks to evaluate and improve the behaviors of GPT-4 when encountering risky contents [46]. In addition, for other aspects of human alignment (_e.g._, truthfulness), several studies propose to use specific instructions and devise annotation rules to guide the annotation process [81]. Empirical studies have revealed that these strategies can greatly improve the human alignment ability of LLMs [170]. For instance, after alignment tuning on data collected through interactions with experts, the incorrect behavior rate of GPT-4 can be largely reduced when it deals with sensitive or disallowed prompts. In addition, high-quality pre-training data can reduce the effort required for alignment [46]. For instance, Galactic is potentially more harmless due to the less biased contents in the scientific corpus [35].
์ด ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์œ ์šฉ์„ฑ, ์ •์ง์„ฑ, ์•ˆ์ „์„ฑ ๋“ฑ ์ธ๊ฐ„์˜ ์ •๋ ฌ์„ ์œ„ํ•œ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ธฐ์ค€์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค[368, 46, 170]. ์œ ์šฉ์„ฑ๊ณผ ์ •์ง์„ฑ์˜ ๊ฒฝ์šฐ, ์ ๋Œ€์  ์งˆ๋ฌธ ๋‹ต๋ณ€ ๊ณผ์ œ(_์˜ˆ: TruthfulQA [556])๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ…์ŠคํŠธ์—์„œ ๊ฐ€๋Šฅํ•œ ํ—ˆ๊ตฌ๋ฅผ ๊ฐ์ง€ํ•˜๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ๊ฒ€์‚ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[36, 81]. ๋˜ํ•œ, ๋ฌดํ•ด์„ฑ์€ ๊ธฐ์กด ๋ฒค์น˜๋งˆํฌ(์˜ˆ: CrowS-Pairs [603] ๋ฐ Winogender [604])๋ฅผ ํ†ตํ•ด์„œ๋„ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ„์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ†ตํ•œ ์ž๋™ ํ‰๊ฐ€์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ธ๊ฐ„ ํ‰๊ฐ€๋Š” ์—ฌ์ „ํžˆ LLM์˜ ์ธ๊ฐ„ ์ •๋ ฌ ๋Šฅ๋ ฅ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ…Œ์ŠคํŠธํ•  ์ˆ˜ ์žˆ๋Š” ๋ณด๋‹ค ์ง์ ‘์ ์ธ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. OpenAI๋Š” AI ์œ„ํ—˜๊ณผ ๊ด€๋ จ๋œ ์˜์—ญ์˜ ๋งŽ์€ ์ „๋ฌธ๊ฐ€๋“ค์„ ์ดˆ๋Œ€ํ•˜์—ฌ ์œ„ํ—˜ํ•œ ์ฝ˜ํ…์ธ ๊ฐ€ ๋ฐœ์ƒํ–ˆ์„ ๋•Œ GPT-4์˜ ํ–‰๋™์„ ํ‰๊ฐ€ํ•˜๊ณ  ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค[46]. ๋˜ํ•œ, ์ง„์‹ค์„ฑ ๋“ฑ ์ธ๊ฐ„ ์ •๋ ฌ์˜ ๋‹ค๋ฅธ ์ธก๋ฉด์— ๋Œ€ํ•ด์„œ๋Š” ํŠน์ • Instruction์„ ์‚ฌ์šฉํ•˜๊ณ  ์ฃผ์„ ๊ทœ์น™์„ ๊ณ ์•ˆํ•˜์—ฌ ์ฃผ์„ ํ”„๋กœ์„ธ์Šค๋ฅผ ์•ˆ๋‚ดํ•  ๊ฒƒ์„ ์ œ์•ˆํ•˜๋Š” ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค[81]. ์‹ค์ฆ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์ด๋Ÿฌํ•œ ์ „๋žต์€ LLM์˜ ์ธ๊ฐ„ ์ •๋ ฌ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[170]. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ „๋ฌธ๊ฐ€์™€์˜ ์ƒํ˜ธ ์ž‘์šฉ์„ ํ†ตํ•ด ์ˆ˜์ง‘ํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ alignment tuning์„ ๊ฑฐ์นœ ํ›„, GPT-4๊ฐ€ ๋ฏผ๊ฐํ•˜๊ฑฐ๋‚˜ ํ—ˆ์šฉ๋˜์ง€ ์•Š๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ ์ž˜๋ชป๋œ ๋™์ž‘ ๋น„์œจ์„ ํฌ๊ฒŒ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๊ณ ํ’ˆ์งˆ์˜ pre-training ๋ฐ์ดํ„ฐ๋Š” ์ •๋ ฌ์— ํ•„์š”ํ•œ ๋…ธ๋ ฅ์„ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[46]. ์˜ˆ๋ฅผ ๋“ค์–ด, Galactic์€ ๊ณผํ•™ ์ฝ”ํผ์Šค์˜ ํŽธํ–ฅ์„ฑ์ด ๋œํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ž ์žฌ์ ์œผ๋กœ ๋œ ํ•ด๋กœ์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[35].
์ด ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์œ ์šฉ์„ฑ, ์ •์ง์„ฑ, ์•ˆ์ „์„ฑ ๋“ฑ ์ธ๊ฐ„์˜ ์ •๋ ฌ์„ ์œ„ํ•œ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ธฐ์ค€์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค[368, 46, 170]. ์œ ์šฉ์„ฑ๊ณผ ์ •์ง์„ฑ์˜ ๊ฒฝ์šฐ, ์ ๋Œ€์  ์งˆ๋ฌธ ๋‹ต๋ณ€ ๊ณผ์ œ(_์˜ˆ: TruthfulQA [556])๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ…์ŠคํŠธ์—์„œ ๊ฐ€๋Šฅํ•œ ํ—ˆ๊ตฌ๋ฅผ ๊ฐ์ง€ํ•˜๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ๊ฒ€์‚ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[36, 81]. ๋˜ํ•œ, ๋ฌดํ•ด์„ฑ์€ ๊ธฐ์กด ๋ฒค์น˜๋งˆํฌ(์˜ˆ: CrowS-Pairs [603] ๋ฐ Winogender [604])๋ฅผ ํ†ตํ•ด์„œ๋„ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ„์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ†ตํ•œ ์ž๋™ ํ‰๊ฐ€์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ธ๊ฐ„ ํ‰๊ฐ€๋Š” ์—ฌ์ „ํžˆ LLM์˜ ์ธ๊ฐ„ ์ •๋ ฌ ๋Šฅ๋ ฅ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ…Œ์ŠคํŠธํ•  ์ˆ˜ ์žˆ๋Š” ๋ณด๋‹ค ์ง์ ‘์ ์ธ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. OpenAI๋Š” AI ์œ„ํ—˜๊ณผ ๊ด€๋ จ๋œ ์˜์—ญ์˜ ๋งŽ์€ ์ „๋ฌธ๊ฐ€๋“ค์„ ์ดˆ๋Œ€ํ•˜์—ฌ ์œ„ํ—˜ํ•œ ์ฝ˜ํ…์ธ ๊ฐ€ ๋ฐœ์ƒํ–ˆ์„ ๋•Œ GPT-4์˜ ํ–‰๋™์„ ํ‰๊ฐ€ํ•˜๊ณ  ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค[46]. ๋˜ํ•œ, ์ง„์‹ค์„ฑ ๋“ฑ ์ธ๊ฐ„ ์ •๋ ฌ์˜ ๋‹ค๋ฅธ ์ธก๋ฉด์— ๋Œ€ํ•ด์„œ๋Š” ํŠน์ • ์ง€์นจ์„ ์‚ฌ์šฉํ•˜๊ณ  ์ฃผ์„ ๊ทœ์น™์„ ๊ณ ์•ˆํ•˜์—ฌ ์ฃผ์„ ํ”„๋กœ์„ธ์Šค๋ฅผ ์•ˆ๋‚ดํ•  ๊ฒƒ์„ ์ œ์•ˆํ•˜๋Š” ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค[81]. ์‹ค์ฆ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์ด๋Ÿฌํ•œ ์ „๋žต์€ LLM์˜ ์ธ๊ฐ„ ์ •๋ ฌ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[170]. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ „๋ฌธ๊ฐ€์™€์˜ ์ƒํ˜ธ ์ž‘์šฉ์„ ํ†ตํ•ด ์ˆ˜์ง‘ํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ •๋ ฌ ํŠœ๋‹์„ ๊ฑฐ์นœ ํ›„, GPT-4๊ฐ€ ๋ฏผ๊ฐํ•˜๊ฑฐ๋‚˜ ํ—ˆ์šฉ๋˜์ง€ ์•Š๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ ์ž˜๋ชป๋œ ๋™์ž‘ ๋น„์œจ์„ ํฌ๊ฒŒ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๊ณ ํ’ˆ์งˆ์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋Š” ์ •๋ ฌ์— ํ•„์š”ํ•œ ๋…ธ๋ ฅ์„ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[46]. ์˜ˆ๋ฅผ ๋“ค์–ด, Galactic์€ ๊ณผํ•™ ์ฝ”ํผ์Šค์˜ ํŽธํ–ฅ์„ฑ์ด ๋œํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ž ์žฌ์ ์œผ๋กœ ๋œ ํ•ด๋กœ์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[35].
With respect to the large deviation from Chinchilla Optimal \(\frac{N_{opt}}{D_{opt}}\), we notice that their scaling experiment was conducted in a not very recent configuration. To compare with more recent configuration such as Llama2 (Touvron et al., 2023), we extract the training loss data from Llama2 paper (left part) in Appendix Figure 18 and estimate the compute optimal \(\frac{D_{opt}}{N_{opt}}\) in their paper using the right part of Figure 18. Since they use Cosine LRS, the loss is not optimal in the middle of the training, depicted by the concave curve during training in the right figure of Figure 18. We fill the concave part with a straight line to estimate the optimal loss envelope if they had used the WSD LRS. After that, the compute model size should roughly be the regime that a model's loss curve is about to intersect with a larger model's loss curve. With this intuition, the 13B model is about to intersect with 34B model at \(10^{5}\) EFlops (\(10^{1}\)s Flops), and the 34B model is about to intersect with 70B model at \(5\times 10^{5}\) EFlops. Therefore, we estimate the \(\frac{D_{opt}}{N_{opt}}\) to be roughly \(\frac{5\times 10^{5}}{6\times 34^{2}}\sim\frac{10^{5}}{6\times 13^{2}}\), which is \(70\sim 100\). Therefore, under this approximate comparison, their data-model ratio is closer to ours. And our configuration can absorb more data into a smaller model compared to previous ones. However, we note that the above estimate are only a rough one.
์นœ์น ๋ผ ์ตœ์ ๊ฐ’์—์„œ ํฐ ํŽธ์ฐจ๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ๊ณผ ๊ด€๋ จํ•˜์—ฌ, ๊ทธ๋“ค์˜ scaling ์‹คํ—˜์€ ๊ทธ๋ฆฌ ์ตœ๊ทผ์˜ ๊ตฌ์„ฑ์—์„œ ์ˆ˜ํ–‰๋œ ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Llama2(Touvron et al., 2023)์™€ ๊ฐ™์€ ์ตœ์‹  ๊ตฌ์„ฑ๊ณผ ๋น„๊ตํ•˜๊ธฐ ์œ„ํ•ด ๋ถ€๋ก ๊ทธ๋ฆผ 18์˜ Llama2 ๋…ผ๋ฌธ์—์„œ ํ›ˆ๋ จ loss ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•˜๊ณ  ๊ทธ๋ฆผ 18์˜ ์˜ค๋ฅธ์ชฝ ๋ถ€๋ถ„์„ ์‚ฌ์šฉํ•˜์—ฌ ๋…ผ๋ฌธ์—์„œ compute optimal \(\frac{D_{opt}}{N_{opt}}\)๋ฅผ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค. Cosine LRS๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— loss๋Š” ํ›ˆ๋ จ ์ค‘์— ์˜ค๋ชฉํ•œ ๊ณก์„ ์œผ๋กœ ํ‘œ์‹œ๋˜๋Š” ํ›ˆ๋ จ ์ค‘๋ฐ˜์— ์ตœ์ ์ด ์•„๋‹™๋‹ˆ๋‹ค. 18 ๊ทธ๋ฆผ์˜ ์˜ค๋ฅธ์ชฝ ๊ทธ๋ฆผ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ์˜ค๋ชฉํ•œ ๋ถ€๋ถ„์„ ์ง์„ ์œผ๋กœ ์ฑ„์›Œ WSD LRS๋ฅผ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ์˜ ์ตœ์  loss ์™ธํ”ผ๋ฅผ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ํ›„ compute ๋ชจ๋ธ ํฌ๊ธฐ๋Š” ๋ชจ๋ธ์˜ loss ๊ณก์„ ์ด ๋” ํฐ ๋ชจ๋ธ์˜ loss ๊ณก์„ ๊ณผ ๊ต์ฐจํ•˜๊ธฐ ์ง์ „์˜ ๋ ˆ์ง์ด ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง๊ด€์„ ํ†ตํ•ด 13B ๋ชจ๋ธ์€ \(10^{5}\) EFlops (\(10^{1}\)s Flops)์—์„œ 34B ๋ชจ๋ธ๊ณผ ๊ต์ฐจํ•˜๊ณ , 34B ๋ชจ๋ธ์€ \(5\times 10^{5}\) EFlops์—์„œ 70B ๋ชจ๋ธ๊ณผ ๊ต์ฐจํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ \(\frac{D_{opt}}{N_{opt}}\)๋Š” ๋Œ€๋žต \(\frac{5\times 10^{5}}{6\times 34^{2}}\sim\frac{10^{5}}{6\times 13^{2}}\), ์ฆ‰, \(70\sim 100\)๋กœ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ๋Œ€๋žต์ ์ธ ๋น„๊ต์— ๋”ฐ๋ฅด๋ฉด, ์ด๋“ค์˜ ๋ฐ์ดํ„ฐ-๋ชจ๋ธ ๋น„์œจ์€ ์šฐ๋ฆฌ์™€ ๋” ๊ฐ€๊น์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ์˜ ๊ตฌ์„ฑ์€ ์ด์ „ ๊ตฌ์„ฑ์— ๋น„ํ•ด ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋ฅผ ๋” ์ž‘์€ ๋ชจ๋ธ์— ํก์ˆ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์œ„์˜ ์ถ”์ •์น˜๋Š” ๋Œ€๋žต์ ์ธ ์ถ”์ •์น˜์ผ ๋ฟ์ด๋ผ๋Š” ์ ์— ์œ ์˜ํ•ฉ๋‹ˆ๋‹ค.
์นœ์น ๋ผ ์ตœ์ ๊ฐ’์—์„œ ํฐ ํŽธ์ฐจ๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ๊ณผ ๊ด€๋ จํ•˜์—ฌ, ๊ทธ๋“ค์˜ ์Šค์ผ€์ผ๋ง ์‹คํ—˜์€ ๊ทธ๋ฆฌ ์ตœ๊ทผ์˜ ๊ตฌ์„ฑ์—์„œ ์ˆ˜ํ–‰๋œ ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Llama2(Touvron et al., 2023)์™€ ๊ฐ™์€ ์ตœ์‹  ๊ตฌ์„ฑ๊ณผ ๋น„๊ตํ•˜๊ธฐ ์œ„ํ•ด ๋ถ€๋ก ๊ทธ๋ฆผ 18์˜ Llama2 ๋…ผ๋ฌธ์—์„œ ํ›ˆ๋ จ ์†์‹ค ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•˜๊ณ  ๊ทธ๋ฆผ 18์˜ ์˜ค๋ฅธ์ชฝ ๋ถ€๋ถ„์„ ์‚ฌ์šฉํ•˜์—ฌ ๋…ผ๋ฌธ์—์„œ ๊ณ„์‚ฐ ์ตœ์  \(\frac{D_{opt}}{N_{opt}}\)๋ฅผ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค. ์ฝ”์‚ฌ์ธ LRS๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์†์‹ค์€ ํ›ˆ๋ จ ์ค‘์— ์˜ค๋ชฉํ•œ ๊ณก์„ ์œผ๋กœ ํ‘œ์‹œ๋˜๋Š” ํ›ˆ๋ จ ์ค‘๋ฐ˜์— ์ตœ์ ์ด ์•„๋‹™๋‹ˆ๋‹ค. 18 ๊ทธ๋ฆผ์˜ ์˜ค๋ฅธ์ชฝ ๊ทธ๋ฆผ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ์˜ค๋ชฉํ•œ ๋ถ€๋ถ„์„ ์ง์„ ์œผ๋กœ ์ฑ„์›Œ WSD LRS๋ฅผ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ์˜ ์ตœ์  ์†์‹ค ์™ธํ”ผ๋ฅผ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ํ›„ ๊ณ„์‚ฐ ๋ชจ๋ธ ํฌ๊ธฐ๋Š” ๋ชจ๋ธ์˜ ์†์‹ค ๊ณก์„ ์ด ๋” ํฐ ๋ชจ๋ธ์˜ ์†์‹ค ๊ณก์„ ๊ณผ ๊ต์ฐจํ•˜๊ธฐ ์ง์ „์˜ ๋ ˆ์ง์ด ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง๊ด€์„ ํ†ตํ•ด 13B ๋ชจ๋ธ์€ \(10^{5}\) EFlops (\(10^{1}\)s Flops)์—์„œ 34B ๋ชจ๋ธ๊ณผ ๊ต์ฐจํ•˜๊ณ , 34B ๋ชจ๋ธ์€ \(5\times 10^{5}\) EFlops์—์„œ 70B ๋ชจ๋ธ๊ณผ ๊ต์ฐจํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ \(\frac{D_{opt}}{N_{opt}}\)๋Š” ๋Œ€๋žต \(\frac{5\times 10^{5}}{6\times 34^{2}}\sim\frac{10^{5}}{6\times 13^{2}}\), ์ฆ‰, \(70\sim 100\)๋กœ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ๋Œ€๋žต์ ์ธ ๋น„๊ต์— ๋”ฐ๋ฅด๋ฉด, ์ด๋“ค์˜ ๋ฐ์ดํ„ฐ-๋ชจ๋ธ ๋น„์œจ์€ ์šฐ๋ฆฌ์™€ ๋” ๊ฐ€๊น์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ์˜ ๊ตฌ์„ฑ์€ ์ด์ „ ๊ตฌ์„ฑ์— ๋น„ํ•ด ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋ฅผ ๋” ์ž‘์€ ๋ชจ๋ธ์— ํก์ˆ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์œ„์˜ ์ถ”์ •์น˜๋Š” ๋Œ€๋žต์ ์ธ ์ถ”์ •์น˜์ผ ๋ฟ์ด๋ผ๋Š” ์ ์— ์œ ์˜ํ•ฉ๋‹ˆ๋‹ค.
* **Domain knowledge deficit**: A language model may lack comprehensive expertise in a specific domain to which it has not been exposed. For example, a model trained exclusively on texts written by William Shakespeare would perform poorly when asked about the works of Mark Twain. * **Outdated Information**: LLMs invariably have a cutoff date determined by their training dataset. Consequently, any events, discoveries, or changes occurring after the last training update will not be within the model's knowledge without access to external sources. * **Immemorization**: Sometimes, a model is exposed to knowledge during its training process but does not retain it. This is especially true for rare facts that appear in the training dataset only scarcely (Kandpal et al., 2023). * **Forgetting**: Language models often undergo additional training after the pre-training phase (fine-tuning). In some cases, this might lead to a phenomenon called _catastrophic forgetting_(Kirkpatrick et al., 2017; Goodfellow et al., 2013; Chen et al., 2020; Luo et al., 2023), where models lose some of the knowledge they had prior to the fine-tuning process. * **Reasoning Failure**: In certain instances, a language model might possess relevant knowledge about a fact but fail to utilize it properly. This is particularly evident in complex multi-step reasoning tasks (Tan et al., 2023) or when posed with different questions about the same fact, resulting in disparate outcomes (Berglund et al., 2023).
* **๋„๋ฉ”์ธ ์ง€์‹ ๊ฒฐํ•**: ์–ธ์–ด ๋ชจ๋ธ์€ ๋…ธ์ถœ๋˜์ง€ ์•Š์€ ํŠน์ • ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์ „๋ฌธ ์ง€์‹์ด ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์œŒ๋ฆฌ์—„ ์…ฐ์ต์Šคํ”ผ์–ด๊ฐ€ ์“ด ํ…์ŠคํŠธ๋กœ๋งŒ ํ•™์Šตํ•œ ๋ชจ๋ธ์€ ๋งˆํฌ ํŠธ์›จ์ธ์˜ ์ž‘ํ’ˆ์— ๋Œ€ํ•ด ์งˆ๋ฌธ์„ ๋ฐ›์œผ๋ฉด ์„ฑ๋Šฅ์ด ์ €ํ•˜๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * **Outdated Information**: LLM์€ ํ•ญ์ƒ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์˜ํ•ด ๊ฒฐ์ •๋œ ๋งˆ๊ฐ์ผ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋งˆ์ง€๋ง‰ ํ•™์Šต ์—…๋ฐ์ดํŠธ ์ดํ›„ ๋ฐœ์ƒํ•œ ๋ชจ๋“  ์ด๋ฒคํŠธ, ๋ฐœ๊ฒฌ ๋˜๋Š” ๋ณ€๊ฒฝ ์‚ฌํ•ญ์€ ์™ธ๋ถ€ ์†Œ์Šค์— ์•ก์„ธ์Šคํ•˜์ง€ ์•Š๊ณ ๋Š” ๋ชจ๋ธ์˜ ์ง€์‹์— ํฌํ•จ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. * **Immemorization**: ๋•Œ๋•Œ๋กœ ๋ชจ๋ธ์€ ํ•™์Šต ๊ณผ์ •์—์„œ ์ง€์‹์— ๋…ธ์ถœ๋˜์ง€๋งŒ ์ด๋ฅผ ๊ธฐ์–ตํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๊ฑฐ์˜ ๋‚˜ํƒ€๋‚˜์ง€ ์•Š๋Š” ํฌ๊ท€ํ•œ ์‚ฌ์‹ค์˜ ๊ฒฝ์šฐ ํŠนํžˆ ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค(Kandpal et al., 2023). * **Forgetting**: ์–ธ์–ด ๋ชจ๋ธ์€ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„ ์ดํ›„์— ์ถ”๊ฐ€ ํ•™์Šต(Fine-tuning)์„ ๊ฑฐ์น˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ฒฝ์šฐ์— ๋”ฐ๋ผ ์ด๋กœ ์ธํ•ด ๋ชจ๋ธ์ด Fine-tuning ํ”„๋กœ์„ธ์Šค ์ด์ „์˜ ์ง€์‹ ์ค‘ ์ผ๋ถ€๋ฅผ ์žƒ์–ด๋ฒ„๋ฆฌ๋Š” 'catastrophic forgetting'(Kirkpatrick et al., 2017; Goodfellow et al., 2013; Chen et al., 2020; Luo et al., 2023)์ด๋ผ๋Š” ํ˜„์ƒ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * **Reasoning Failure**: ํŠน์ • ๊ฒฝ์šฐ ์–ธ์–ด ๋ชจ๋ธ์ด ์‚ฌ์‹ค์— ๋Œ€ํ•œ ๊ด€๋ จ ์ง€์‹์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์ง€๋งŒ ์ด๋ฅผ ์ œ๋Œ€๋กœ ํ™œ์šฉํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ณต์žกํ•œ ๋‹ค๋‹จ๊ณ„ ์ถ”๋ก  ๊ณผ์ œ(Tan et al., 2023)์—์„œ ๋˜๋Š” ๋™์ผํ•œ ์‚ฌ์‹ค์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ์งˆ๋ฌธ์„ ๋ฐ›์•˜์„ ๋•Œ ๊ฒฐ๊ณผ๊ฐ€ ์ƒ์ดํ•˜๊ฒŒ ๋‚˜ํƒ€๋‚  ๋•Œ ํŠนํžˆ ๋‘๋“œ๋Ÿฌ์ง‘๋‹ˆ๋‹ค(Berglund et al., 2023).
* **๋„๋ฉ”์ธ ์ง€์‹ ๊ฒฐํ•**: ์–ธ์–ด ๋ชจ๋ธ์€ ๋…ธ์ถœ๋˜์ง€ ์•Š์€ ํŠน์ • ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์ „๋ฌธ ์ง€์‹์ด ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์œŒ๋ฆฌ์—„ ์…ฐ์ต์Šคํ”ผ์–ด๊ฐ€ ์“ด ํ…์ŠคํŠธ๋กœ๋งŒ ํ•™์Šตํ•œ ๋ชจ๋ธ์€ ๋งˆํฌ ํŠธ์›จ์ธ์˜ ์ž‘ํ’ˆ์— ๋Œ€ํ•ด ์งˆ๋ฌธ์„ ๋ฐ›์œผ๋ฉด ์„ฑ๋Šฅ์ด ์ €ํ•˜๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * **๊ตฌํ˜• ์ •๋ณด**: LLM์€ ํ•ญ์ƒ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์˜ํ•ด ๊ฒฐ์ •๋œ ๋งˆ๊ฐ์ผ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋งˆ์ง€๋ง‰ ํ•™์Šต ์—…๋ฐ์ดํŠธ ์ดํ›„ ๋ฐœ์ƒํ•œ ๋ชจ๋“  ์ด๋ฒคํŠธ, ๋ฐœ๊ฒฌ ๋˜๋Š” ๋ณ€๊ฒฝ ์‚ฌํ•ญ์€ ์™ธ๋ถ€ ์†Œ์Šค์— ์•ก์„ธ์Šคํ•˜์ง€ ์•Š๊ณ ๋Š” ๋ชจ๋ธ์˜ ์ง€์‹์— ํฌํ•จ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. * **๊ธฐ์–ต ์ƒ์‹ค**: ๋•Œ๋•Œ๋กœ ๋ชจ๋ธ์€ ํ•™์Šต ๊ณผ์ •์—์„œ ์ง€์‹์— ๋…ธ์ถœ๋˜์ง€๋งŒ ์ด๋ฅผ ๊ธฐ์–ตํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๊ฑฐ์˜ ๋‚˜ํƒ€๋‚˜์ง€ ์•Š๋Š” ํฌ๊ท€ํ•œ ์‚ฌ์‹ค์˜ ๊ฒฝ์šฐ ํŠนํžˆ ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค(Kandpal et al., 2023). * **์žŠ์–ด๋ฒ„๋ฆผ**: ์–ธ์–ด ๋ชจ๋ธ์€ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„ ์ดํ›„์— ์ถ”๊ฐ€ ํ•™์Šต(๋ฏธ์„ธ ์กฐ์ •)์„ ๊ฑฐ์น˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ฒฝ์šฐ์— ๋”ฐ๋ผ ์ด๋กœ ์ธํ•ด ๋ชจ๋ธ์ด ๋ฏธ์„ธ ์กฐ์ • ํ”„๋กœ์„ธ์Šค ์ด์ „์˜ ์ง€์‹ ์ค‘ ์ผ๋ถ€๋ฅผ ์žƒ์–ด๋ฒ„๋ฆฌ๋Š” '์žฌ์•™์  ๋ง๊ฐ'(Kirkpatrick et al., 2017; Goodfellow et al., 2013; Chen et al., 2020; Luo et al., 2023)์ด๋ผ๋Š” ํ˜„์ƒ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * **์ถ”๋ก  ์‹คํŒจ**: ํŠน์ • ๊ฒฝ์šฐ ์–ธ์–ด ๋ชจ๋ธ์ด ์‚ฌ์‹ค์— ๋Œ€ํ•œ ๊ด€๋ จ ์ง€์‹์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์ง€๋งŒ ์ด๋ฅผ ์ œ๋Œ€๋กœ ํ™œ์šฉํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ณต์žกํ•œ ๋‹ค๋‹จ๊ณ„ ์ถ”๋ก  ๊ณผ์ œ(Tan et al., 2023)์—์„œ ๋˜๋Š” ๋™์ผํ•œ ์‚ฌ์‹ค์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ์งˆ๋ฌธ์„ ๋ฐ›์•˜์„ ๋•Œ ๊ฒฐ๊ณผ๊ฐ€ ์ƒ์ดํ•˜๊ฒŒ ๋‚˜ํƒ€๋‚  ๋•Œ ํŠนํžˆ ๋‘๋“œ๋Ÿฌ์ง‘๋‹ˆ๋‹ค(Berglund et al., 2023).
In the classification task, S4nd [30] proposes a multi-dimensional and multi-polar graphics component to expand the modeling capability of multi-dimensional data continuous signals, which can model large-scale visual data into dynamic multi-dimensional linear signals. VMamba [60] uses linear complexity to capture the full range of sensory fields, introduces traversal of spatial information across scan blocks, and converts non-causal visual images into ordered patch sequences. Vim [61] uses a bidirectional state-space model to compress visual representation information and understand the global context through location embedding and visual information. Li et al. present Mamba-ND [68], an extension of Mamba designed to handle arbitrary multi-dimensional data by processing input data across dimensions in a row-major order. The authors of [57] design S5 based on S4 to establish the relationship between S5 and S4, utilize multi-input multi-output SSM, and use the state space layer of parallel scanning for long-distance sequence modeling. Baron et al. [76] design a new 2-dimensional State Space Layer for Spatial Inductive Bias. The core goals of this layer are to achieve perception of 2-D position, dynamic spatial localization, and translation and alignment invariance. Chen et al. [79] are the first to integrate residuals into the original VMamba, and maintain the inherent global and local state characteristics of the original VMamba for food classification. Yang et al.
๋ถ„๋ฅ˜ ์ž‘์—…์—์„œ S4nd [30]๋Š” ๋‹ค์ฐจ์› ๋ฐ ๋‹ค๊ทน ๊ทธ๋ž˜ํ”ฝ ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ์ œ์•ˆํ•˜์—ฌ ๋‹ค์ฐจ์› ๋ฐ์ดํ„ฐ ์—ฐ์† ์‹ ํ˜ธ์˜ ๋ชจ๋ธ๋ง ๊ธฐ๋Šฅ์„ ํ™•์žฅํ•˜์—ฌ ๋Œ€๊ทœ๋ชจ ์‹œ๊ฐ ๋ฐ์ดํ„ฐ๋ฅผ ๋™์  ๋‹ค์ฐจ์› ์„ ํ˜• ์‹ ํ˜ธ๋กœ ๋ชจ๋ธ๋งํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. VMamba [60]๋Š” linear complexity๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ „์ฒด ๊ฐ๊ฐ ์˜์—ญ์„ ํฌ์ฐฉํ•˜๊ณ , ์Šค์บ” ๋ธ”๋ก ๊ฐ„์˜ ๊ณต๊ฐ„ ์ •๋ณด ์ˆœํšŒ๋ฅผ ๋„์ž…ํ•˜๋ฉฐ, ๋น„์ธ๊ณผ์  ์‹œ๊ฐ ์ด๋ฏธ์ง€๋ฅผ ์ •๋ ฌ๋œ patch sequences๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค. Vim [61]์€ ์–‘๋ฐฉํ–ฅ state-space model์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹œ๊ฐ์  ํ‘œํ˜„ ์ •๋ณด๋ฅผ ์••์ถ•ํ•˜๊ณ  ์œ„์น˜ embedding ๋ฐ ์‹œ๊ฐ ์ •๋ณด๋ฅผ ํ†ตํ•ด ์ „์—ญ ์ปจํ…์ŠคํŠธ๋ฅผ ์ดํ•ดํ•ฉ๋‹ˆ๋‹ค. Li ๋“ฑ์€ Mamba-ND [68]๋ฅผ ์ œ์‹œํ•˜๋Š”๋ฐ, ์ด๋Š” ํ–‰๋ ฌ ์ˆœ์„œ๋กœ ์ฐจ์›์„ ๊ฐ€๋กœ์ง€๋ฅด๋Š” ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌํ•˜์—ฌ ์ž„์˜์˜ ๋‹ค์ฐจ์› ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋„๋ก ์„ค๊ณ„๋œ Mamba์˜ ํ™•์žฅ์ž…๋‹ˆ๋‹ค. [57]์˜ ์ €์ž๋Š” S4๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ S5๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ S5์™€ S4์˜ ๊ด€๊ณ„๋ฅผ ์„ค์ •ํ•˜๊ณ , ๋‹ค์ค‘ ์ž…๋ ฅ ๋‹ค์ค‘ ์ถœ๋ ฅ SSM์„ ํ™œ์šฉํ•˜๋ฉฐ, ์žฅ๊ฑฐ๋ฆฌ ์‹œํ€€์Šค ๋ชจ๋ธ๋ง์„ ์œ„ํ•ด ๋ณ‘๋ ฌ ์Šค์บ”์˜ state-space layer๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Baron ๋“ฑ [76]์€ ๊ณต๊ฐ„์  ๊ท€๋‚ฉ์  ๋ฐ”์ด์–ด์Šค๋ฅผ ์œ„ํ•œ ์ƒˆ๋กœ์šด 2์ฐจ์› state-space layer๋ฅผ ์„ค๊ณ„ํ•ฉ๋‹ˆ๋‹ค. ์ด layer์˜ ํ•ต์‹ฌ ๋ชฉํ‘œ๋Š” 2D ์œ„์น˜, ๋™์  ๊ณต๊ฐ„ ๋กœ์ปฌ๋ผ์ด์ œ์ด์…˜, ๋ณ€ํ™˜ ๋ฐ ์ •๋ ฌ ๋ถˆ๋ณ€์„ฑ์„ ์ธ์‹ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. Chen ๋“ฑ [79]์€ ์ตœ์ดˆ๋กœ ์ž”์ฐจ๋ฅผ ์›๋ž˜์˜ VMamba์— ํ†ตํ•ฉํ–ˆ์œผ๋ฉฐ, ์Œ์‹ ๋ถ„๋ฅ˜๋ฅผ ์œ„ํ•ด ์›๋ž˜ VMamba ๊ณ ์œ ์˜ ์ „์—ญ ๋ฐ ๋กœ์ปฌ ์ƒํƒœ ํŠน์„ฑ์„ ์œ ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค. Yang et al.
๋ถ„๋ฅ˜ ์ž‘์—…์—์„œ S4nd [30]๋Š” ๋‹ค์ฐจ์› ๋ฐ ๋‹ค๊ทน ๊ทธ๋ž˜ํ”ฝ ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ์ œ์•ˆํ•˜์—ฌ ๋‹ค์ฐจ์› ๋ฐ์ดํ„ฐ ์—ฐ์† ์‹ ํ˜ธ์˜ ๋ชจ๋ธ๋ง ๊ธฐ๋Šฅ์„ ํ™•์žฅํ•˜์—ฌ ๋Œ€๊ทœ๋ชจ ์‹œ๊ฐ ๋ฐ์ดํ„ฐ๋ฅผ ๋™์  ๋‹ค์ฐจ์› ์„ ํ˜• ์‹ ํ˜ธ๋กœ ๋ชจ๋ธ๋งํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. VMamba [60]๋Š” ์„ ํ˜• ๋ณต์žก์„ฑ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ „์ฒด ๊ฐ๊ฐ ์˜์—ญ์„ ํฌ์ฐฉํ•˜๊ณ , ์Šค์บ” ๋ธ”๋ก ๊ฐ„์˜ ๊ณต๊ฐ„ ์ •๋ณด ์ˆœํšŒ๋ฅผ ๋„์ž…ํ•˜๋ฉฐ, ๋น„์ธ๊ณผ์  ์‹œ๊ฐ ์ด๋ฏธ์ง€๋ฅผ ์ •๋ ฌ๋œ ํŒจ์น˜ ์‹œํ€€์Šค๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค. Vim [61]์€ ์–‘๋ฐฉํ–ฅ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹œ๊ฐ์  ํ‘œํ˜„ ์ •๋ณด๋ฅผ ์••์ถ•ํ•˜๊ณ  ์œ„์น˜ ์ž„๋ฒ ๋”ฉ ๋ฐ ์‹œ๊ฐ ์ •๋ณด๋ฅผ ํ†ตํ•ด ์ „์—ญ ์ปจํ…์ŠคํŠธ๋ฅผ ์ดํ•ดํ•ฉ๋‹ˆ๋‹ค. Li ๋“ฑ์€ Mamba-ND [68]๋ฅผ ์ œ์‹œํ•˜๋Š”๋ฐ, ์ด๋Š” ํ–‰๋ ฌ ์ˆœ์„œ๋กœ ์ฐจ์›์„ ๊ฐ€๋กœ์ง€๋ฅด๋Š” ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌํ•˜์—ฌ ์ž„์˜์˜ ๋‹ค์ฐจ์› ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋„๋ก ์„ค๊ณ„๋œ Mamba์˜ ํ™•์žฅ์ž…๋‹ˆ๋‹ค. [57]์˜ ์ €์ž๋Š” S4๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ S5๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ S5์™€ S4์˜ ๊ด€๊ณ„๋ฅผ ์„ค์ •ํ•˜๊ณ , ๋‹ค์ค‘ ์ž…๋ ฅ ๋‹ค์ค‘ ์ถœ๋ ฅ SSM์„ ํ™œ์šฉํ•˜๋ฉฐ, ์žฅ๊ฑฐ๋ฆฌ ์‹œํ€€์Šค ๋ชจ๋ธ๋ง์„ ์œ„ํ•ด ๋ณ‘๋ ฌ ์Šค์บ”์˜ ์ƒํƒœ ๊ณต๊ฐ„ ๋ ˆ์ด์–ด๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Baron ๋“ฑ [76]์€ ๊ณต๊ฐ„์  ๊ท€๋‚ฉ์  ๋ฐ”์ด์–ด์Šค๋ฅผ ์œ„ํ•œ ์ƒˆ๋กœ์šด 2์ฐจ์› ์ƒํƒœ ๊ณต๊ฐ„ ๋ ˆ์ด์–ด๋ฅผ ์„ค๊ณ„ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ ˆ์ด์–ด์˜ ํ•ต์‹ฌ ๋ชฉํ‘œ๋Š” 2D ์œ„์น˜, ๋™์  ๊ณต๊ฐ„ ๋กœ์ปฌ๋ผ์ด์ œ์ด์…˜, ๋ณ€ํ™˜ ๋ฐ ์ •๋ ฌ ๋ถˆ๋ณ€์„ฑ์„ ์ธ์‹ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. Chen ๋“ฑ [79]์€ ์ตœ์ดˆ๋กœ ์ž”์ฐจ๋ฅผ ์›๋ž˜์˜ VMamba์— ํ†ตํ•ฉํ–ˆ์œผ๋ฉฐ, ์Œ์‹ ๋ถ„๋ฅ˜๋ฅผ ์œ„ํ•ด ์›๋ž˜ VMamba ๊ณ ์œ ์˜ ์ „์—ญ ๋ฐ ๋กœ์ปฌ ์ƒํƒœ ํŠน์„ฑ์„ ์œ ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค. Yang et al.
Table 3 reports the average accuracy of each model for our English evaluation tasks and the normalized accuracy for the German Hella Swag evaluation task. We do not report the average German evaluation score as it is not informative due to evaluations having near-random chance accuracy (see Table 11). We observe that English models consistently outperform German models on the English evaluations. However, the strong replay used with the 25% replay German model helps to reduce this gap. English models' English evaluation performance is very similar with a range of 1.19 between the highest and lowest values. We suspect that there is significant noise in the evaluation process for base models of this size and believe that the differences are likely not significant. That being said, the continually pre-trained model with LR re-warming, LR re-decaying, and replay does improve on the \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) model. When evaluating German-trained models on English evaluation tasks, we see consistent improvements for models using more replay. We note that once again the model trained with LR re-warming, LR re-decaying, and replay does improve on the \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) model. Turning to the German Hella Swag results we observe that German models consistently outperform their English counterparts. Among German-trained models, the continually trained models outperform the union-trained model and the model trained exclusively on German.
ํ‘œ 3์€ ์˜์–ด ํ‰๊ฐ€ ์ž‘์—…์— ๋Œ€ํ•œ ๊ฐ ๋ชจ๋ธ์˜ ํ‰๊ท  ์ •ํ™•๋„์™€ ๋…์ผ์–ด Hella Swag ํ‰๊ฐ€ ์ž‘์—…์— ๋Œ€ํ•œ ์ •๊ทœํ™”๋œ ์ •ํ™•๋„๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€๊ฐ€ ๊ฑฐ์˜ ๋ฌด์ž‘์œ„๋กœ ์ •ํ™•๋„๊ฐ€ ๋‚ฎ๊ธฐ ๋•Œ๋ฌธ์— ์˜์–ด ํ‰๊ฐ€ ์ ์ˆ˜์˜ ํ‰๊ท ์€ ์œ ์˜๋ฏธํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ ๋ณด๊ณ ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค(ํ‘œ 11 ์ฐธ์กฐ). ์˜์–ด ๋ชจ๋ธ์ด ์˜์–ด ํ‰๊ฐ€์—์„œ ๋…์ผ์–ด ๋ชจ๋ธ๋ณด๋‹ค ์ผ๊ด€๋˜๊ฒŒ ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ 25% replay ๋…์ผ์–ด ๋ชจ๋ธ๊ณผ ํ•จ๊ป˜ ์‚ฌ์šฉ๋œ ๊ฐ•๋ ฅํ•œ replay๋Š” ์ด๋Ÿฌํ•œ ๊ฒฉ์ฐจ๋ฅผ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ์˜์–ด ๋ชจ๋ธ์˜ ์˜์–ด ํ‰๊ฐ€ ์„ฑ๋Šฅ์€ ์ตœ๊ณ ๊ฐ’๊ณผ ์ตœ์ €๊ฐ’ ์‚ฌ์ด์˜ 1.19 ๋ฒ”์œ„๋กœ ๋งค์šฐ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ํฌ๊ธฐ์˜ ๊ธฐ๋ณธ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ‰๊ฐ€ ๊ณผ์ •์— ์ƒ๋‹นํ•œ ๋…ธ์ด์ฆˆ๊ฐ€ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋ฉฐ, ์ด๋Ÿฌํ•œ ์ฐจ์ด๋Š” ์œ ์˜๋ฏธํ•˜์ง€ ์•Š์„ ๊ฒƒ์œผ๋กœ ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Learning Rate ์žฌ๊ฐ€์—ด, Learning Rate ์žฌ๊ฐ์‡  ๋ฐ replay๋กœ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์€ \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ์˜์–ด ํ‰๊ฐ€ ์ž‘์—…์—์„œ ๋…์ผ์–ด๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•  ๋•Œ, ๋” ๋งŽ์€ replay๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ์— ๋Œ€ํ•ด ์ผ๊ด€๋œ ๊ฐœ์„  ์‚ฌํ•ญ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์‹œ ํ•œ ๋ฒˆ Learning Rate ์žฌ๊ฐ€์—ด, Learning Rate ์žฌ๊ฐ์‡  ๋ฐ replay๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์ด \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•œ๋‹ค๋Š” ์ ์— ์ฃผ๋ชฉํ•ฉ๋‹ˆ๋‹ค. ๋…์ผ์–ด Hella Swag ๊ฒฐ๊ณผ๋ฅผ ์‚ดํŽด๋ณด๋ฉด ๋…์ผ์–ด ๋ชจ๋ธ์ด ์˜์–ด ๋ชจ๋ธ๋ณด๋‹ค ์ผ๊ด€๋˜๊ฒŒ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋…์ผ์–ด๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ ์ค‘์—์„œ๋„ ์ง€์†์ ์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์ด ํ•ฉ์ง‘ํ•ฉ ํ›ˆ๋ จ ๋ชจ๋ธ๊ณผ ๋…์ผ์–ด๋กœ๋งŒ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค.
ํ‘œ 3์€ ์˜์–ด ํ‰๊ฐ€ ์ž‘์—…์— ๋Œ€ํ•œ ๊ฐ ๋ชจ๋ธ์˜ ํ‰๊ท  ์ •ํ™•๋„์™€ ๋…์ผ์–ด ํ—ฌ๋ผ ์Šค์›ฉ ํ‰๊ฐ€ ์ž‘์—…์— ๋Œ€ํ•œ ์ •๊ทœํ™”๋œ ์ •ํ™•๋„๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€๊ฐ€ ๊ฑฐ์˜ ๋ฌด์ž‘์œ„๋กœ ์ •ํ™•๋„๊ฐ€ ๋‚ฎ๊ธฐ ๋•Œ๋ฌธ์— ์˜์–ด ํ‰๊ฐ€ ์ ์ˆ˜์˜ ํ‰๊ท ์€ ์œ ์˜๋ฏธํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ ๋ณด๊ณ ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค(ํ‘œ 11 ์ฐธ์กฐ). ์˜์–ด ๋ชจ๋ธ์ด ์˜์–ด ํ‰๊ฐ€์—์„œ ๋…์ผ์–ด ๋ชจ๋ธ๋ณด๋‹ค ์ผ๊ด€๋˜๊ฒŒ ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ 25% ๋ฆฌํ”Œ๋ ˆ์ด ๋…์ผ์–ด ๋ชจ๋ธ๊ณผ ํ•จ๊ป˜ ์‚ฌ์šฉ๋œ ๊ฐ•๋ ฅํ•œ ๋ฆฌํ”Œ๋ ˆ์ด๋Š” ์ด๋Ÿฌํ•œ ๊ฒฉ์ฐจ๋ฅผ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ์˜์–ด ๋ชจ๋ธ์˜ ์˜์–ด ํ‰๊ฐ€ ์„ฑ๋Šฅ์€ ์ตœ๊ณ ๊ฐ’๊ณผ ์ตœ์ €๊ฐ’ ์‚ฌ์ด์˜ 1.19 ๋ฒ”์œ„๋กœ ๋งค์šฐ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ํฌ๊ธฐ์˜ ๊ธฐ๋ณธ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ‰๊ฐ€ ๊ณผ์ •์— ์ƒ๋‹นํ•œ ๋…ธ์ด์ฆˆ๊ฐ€ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋ฉฐ, ์ด๋Ÿฌํ•œ ์ฐจ์ด๋Š” ์œ ์˜๋ฏธํ•˜์ง€ ์•Š์„ ๊ฒƒ์œผ๋กœ ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ LR ์žฌ๊ฐ€์—ด, LR ์žฌ๊ฐ์‡  ๋ฐ ๋ฆฌํ”Œ๋ ˆ์ด๋กœ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์€ \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ์˜์–ด ํ‰๊ฐ€ ์ž‘์—…์—์„œ ๋…์ผ์–ด๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•  ๋•Œ, ๋” ๋งŽ์€ ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ์— ๋Œ€ํ•ด ์ผ๊ด€๋œ ๊ฐœ์„  ์‚ฌํ•ญ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์‹œ ํ•œ ๋ฒˆ LR ์žฌ๊ฐ€์—ด, LR ์žฌ๊ฐ์‡  ๋ฐ ๋ฆฌํ”Œ๋ ˆ์ด๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์ด \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•œ๋‹ค๋Š” ์ ์— ์ฃผ๋ชฉํ•ฉ๋‹ˆ๋‹ค. ๋…์ผ์–ด ํ—ฌ๋ผ ์Šค์›ฉ ๊ฒฐ๊ณผ๋ฅผ ์‚ดํŽด๋ณด๋ฉด ๋…์ผ์–ด ๋ชจ๋ธ์ด ์˜์–ด ๋ชจ๋ธ๋ณด๋‹ค ์ผ๊ด€๋˜๊ฒŒ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋…์ผ์–ด๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ ์ค‘์—์„œ๋„ ์ง€์†์ ์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์ด ํ•ฉ์ง‘ํ•ฉ ํ›ˆ๋ จ ๋ชจ๋ธ๊ณผ ๋…์ผ์–ด๋กœ๋งŒ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค.
**Stage 1:**: we train the parameters of the ViT and the projection module using an image resolution of \(224^{2}\). The training leverages a substantial dataset comprising \(100\) million image-text pairs from LAION-400M [66]. The primary objective is to enhance the ViT's knowledge acquisition within our specified architecture and to achieve better alignment between the ViT and the LLM. **Stage 2:**: we scale up the image resolution of ViT to \(448^{2}\), aiming to further boost the model's capability for discerning intricate visual details. The dataset used in this stage includes \(20\) million image-text pairs derived from LAION-400M. Additionally, we incorporate around \(4.8\) million image-text pairs from diverse sources, _e.g._, CLLaVA [45], LLaVAR [91], Flickr [85], VQAv2 [25], RefCOCO [37], Visual7w [95] and so on. **Stage 3:**: the parameters of the entire model are trained. The primary goal is to enhance the model's proficiency in multimodal chat interactions, thereby endowing it with the ability to seamlessly integrate and interpret visual and linguistic inputs. To this end, the training dataset encompasses a diverse range of sources, totalling approximately \(1\) million image-text pairs, including GQA [32], VizWiz VQA [26], TextCaps [71], OCR-VQA [51], Visual Genome [39], ShareGPT4V [6] and so on. To ensure data balancing, we impose a cap on the maximum data contribution from any single source, restricting it to no more than \(50,000\) pairs.
**1๋‹จ๊ณ„**: ์ด๋ฏธ์ง€ ํ•ด์ƒ๋„ \(224^{2}\)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ViT์™€ projection ๋ชจ๋“ˆ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ์€ LAION-400M[66]์˜ 1์–ต ๊ฐœ image-text ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐฉ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์š” ๋ชฉํ‘œ๋Š” ์ง€์ •๋œ ์•„ํ‚คํ…์ฒ˜ ๋‚ด์—์„œ ViT์˜ ์ง€์‹ ์Šต๋“์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ViT์™€ LLM ๊ฐ„์˜ ์ •๋ ฌ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. **2๋‹จ๊ณ„**: ViT์˜ ์ด๋ฏธ์ง€ ํ•ด์ƒ๋„๋ฅผ \(448^{2}\)๋กœ ํ™•์žฅํ•˜์—ฌ ๋ณต์žกํ•œ ์‹œ๊ฐ์  ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ์‹๋ณ„ํ•˜๋Š” ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—์„œ ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” LAION-400M์—์„œ ํŒŒ์ƒ๋œ 2์ฒœ๋งŒ ๊ฐœ์˜ image-text ์Œ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ ์†Œ์Šค(์˜ˆ: CLLaVA [45], LLaVAR [91], Flickr [85], VQAv2 [25], RefCOCO [37], Visual7w [95] ๋“ฑ)์—์„œ ์•ฝ 480๋งŒ ๊ฐœ์˜ image-text ์Œ์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. **3๋‹จ๊ณ„**: ์ „์ฒด ๋ชจ๋ธ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์š” ๋ชฉํ‘œ๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์ฑ„ํŒ… ์ƒํ˜ธ ์ž‘์šฉ์—์„œ ๋ชจ๋ธ์˜ ์ˆ™๋ จ๋„๋ฅผ ํ–ฅ์ƒ์‹œ์ผœ ์‹œ๊ฐ์  ๋ฐ ์–ธ์–ด์  ์ž…๋ ฅ์„ ์›ํ™œํ•˜๊ฒŒ ํ†ตํ•ฉํ•˜๊ณ  ํ•ด์„ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๋ถ€์—ฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” GQA [32], VizWiz VQA [26], TextCaps [71], OCR-VQA [51], Visual Genome [39], ShareGPT4V [6] ๋“ฑ์„ ํฌํ•จํ•˜์—ฌ ์•ฝ 100๋งŒ ๊ฐœ์˜ image-text ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋‹ค์–‘ํ•œ ์†Œ์Šค๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ๊ท ํ˜•์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ๋‹จ์ผ ์†Œ์Šค์˜ ์ตœ๋Œ€ ๋ฐ์ดํ„ฐ ๊ธฐ์—ฌ๋„๋ฅผ ์ œํ•œํ•˜์—ฌ 50,000์Œ ์ดํ•˜๋กœ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค.
**1๋‹จ๊ณ„**: ์ด๋ฏธ์ง€ ํ•ด์ƒ๋„ \(224^{2}\)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ViT์™€ ํˆฌ์˜ ๋ชจ๋“ˆ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ์€ LAION-400M[66]์˜ 1์–ต ๊ฐœ ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐฉ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์š” ๋ชฉํ‘œ๋Š” ์ง€์ •๋œ ์•„ํ‚คํ…์ฒ˜ ๋‚ด์—์„œ ViT์˜ ์ง€์‹ ์Šต๋“์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ViT์™€ LLM ๊ฐ„์˜ ์ •๋ ฌ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. **2๋‹จ๊ณ„**: ViT์˜ ์ด๋ฏธ์ง€ ํ•ด์ƒ๋„๋ฅผ \(448^{2}\)๋กœ ํ™•์žฅํ•˜์—ฌ ๋ณต์žกํ•œ ์‹œ๊ฐ์  ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ์‹๋ณ„ํ•˜๋Š” ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—์„œ ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” LAION-400M์—์„œ ํŒŒ์ƒ๋œ 2์ฒœ๋งŒ ๊ฐœ์˜ ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ ์†Œ์Šค(์˜ˆ: CLLaVA [45], LLaVAR [91], Flickr [85], VQAv2 [25], RefCOCO [37], Visual7w [95] ๋“ฑ)์—์„œ ์•ฝ 480๋งŒ ๊ฐœ์˜ ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. **3๋‹จ๊ณ„**: ์ „์ฒด ๋ชจ๋ธ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์š” ๋ชฉํ‘œ๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์ฑ„ํŒ… ์ƒํ˜ธ ์ž‘์šฉ์—์„œ ๋ชจ๋ธ์˜ ์ˆ™๋ จ๋„๋ฅผ ํ–ฅ์ƒ์‹œ์ผœ ์‹œ๊ฐ์  ๋ฐ ์–ธ์–ด์  ์ž…๋ ฅ์„ ์›ํ™œํ•˜๊ฒŒ ํ†ตํ•ฉํ•˜๊ณ  ํ•ด์„ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๋ถ€์—ฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” GQA [32], VizWiz VQA [26], TextCaps [71], OCR-VQA [51], Visual Genome [39], ShareGPT4V [6] ๋“ฑ์„ ํฌํ•จํ•˜์—ฌ ์•ฝ 100๋งŒ ๊ฐœ์˜ ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋‹ค์–‘ํ•œ ์†Œ์Šค๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ๊ท ํ˜•์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ๋‹จ์ผ ์†Œ์Šค์˜ ์ตœ๋Œ€ ๋ฐ์ดํ„ฐ ๊ธฐ์—ฌ๋„๋ฅผ ์ œํ•œํ•˜์—ฌ 50,000์Œ ์ดํ•˜๋กœ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค.
For the medical image based analysis, Guo et al. [97] introduce a medical MR-CT deformable registration method based on the Mamba framework, named MambaMorph. The key to this method lies in achieving voxel-level spatial correspondence capture across different imaging modalities, which is crucial for medical image analysis. Xie et al. introduce a new polyp segmentation model ProMamba [99] based on the Vision Mamba architecture and prompts technology. It is the first time to introduce the Vision Mamba and prompts into polyp segmentation. Wu et al. [100] introduce a novel neural network for medical image segmentation based on SSM and SS2D, called High-order Vision Mamba UNet (H-vmunet), which gradually reduces the introduction of redundant information through advanced interaction and enhances the ability of SS2D to learn local features at each interaction stage. Vivim is proposed by Yang et al. [98], which targets effectively compressing long-term spatio-temporal representations into sequences of different scales through the designed time Mamba blocks for medical video object segmentation. Due to the scarcity of large labeled medical datasets, Wang et al. [72] propose the Weak-Mamba-Unet architecture, which attempts to address this challenge by training a Mamba-based UNet in a weakly-supervised manner. It leverages convolutional neural networks (CNNs), Vision Transformers (ViTs), and Vmamba to predict the data labels, then, generate dense pseudo labels.
์˜๋ฃŒ ์ด๋ฏธ์ง€ ๊ธฐ๋ฐ˜ ๋ถ„์„์„ ์œ„ํ•ด Guo et al. [97]์€ Mamba ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์˜๋ฃŒ์šฉ MR-CT ๋ณ€ํ˜• ๋“ฑ๋ก ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•˜๋Š”๋ฐ, ์ด๋ฅผ MambaMorph๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์˜ ํ•ต์‹ฌ์€ ์˜๋ฃŒ ์ด๋ฏธ์ง€ ๋ถ„์„์— ์ค‘์š”ํ•œ ๋‹ค์–‘ํ•œ ์ด๋ฏธ์ง• ๋ฐฉ์‹ ๊ฐ„์˜ voxel ์ˆ˜์ค€ ๊ณต๊ฐ„ ๋Œ€์‘ ์บก์ฒ˜๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ์žˆ์Šต๋‹ˆ๋‹ค. Xie et al.์€ Vision Mamba ์•„ํ‚คํ…์ฒ˜์™€ prompts ๊ธฐ์ˆ ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒˆ๋กœ์šด ์šฉ์ข… ๋ถ„ํ•  ๋ชจ๋ธ ProMamba [99]๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์šฉ์ข… ๋ถ„ํ• ์— Vision Mamba์™€ prompts๋ฅผ ๋„์ž…ํ•œ ๊ฒƒ์€ ์ด๋ฒˆ์ด ์ฒ˜์Œ์ž…๋‹ˆ๋‹ค. Wu et al. [100]์€ ์˜๋ฃŒ ์ด๋ฏธ์ง€ ๋ถ„ํ• ์„ ์œ„ํ•ด SSM๊ณผ SS2D๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ƒˆ๋กœ์šด ์‹ ๊ฒฝ๋ง์ธ ๊ณ ์ฐจ Vision Mamba UNet (H-vmunet)์„ ์†Œ๊ฐœํ•˜๋Š”๋ฐ, ์ด๋Š” ๊ณ ๊ธ‰ ์ƒํ˜ธ ์ž‘์šฉ์„ ํ†ตํ•ด ์ค‘๋ณต ์ •๋ณด์˜ ๋„์ž…์„ ์ ์ง„์ ์œผ๋กœ ์ค„์ด๊ณ  ๊ฐ ์ƒํ˜ธ ์ž‘์šฉ ๋‹จ๊ณ„์—์„œ SS2D์˜ ๋กœ์ปฌ ํŠน์ง• ํ•™์Šต ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. Vivim์€ ์˜๋ฃŒ์šฉ ๋น„๋””์˜ค ๊ฐ์ฒด ๋ถ„ํ• ์„ ์œ„ํ•ด ์„ค๊ณ„๋œ ์‹œ๊ฐ„ Mamba ๋ธ”๋ก์„ ํ†ตํ•ด ์žฅ๊ธฐ์ ์ธ ์‹œ๊ณต๊ฐ„ ํ‘œํ˜„์„ ๋‹ค์–‘ํ•œ ์Šค์ผ€์ผ์˜ ์‹œํ€€์Šค๋กœ ํšจ๊ณผ์ ์œผ๋กœ ์••์ถ•ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” Yang et al. [98]์— ์˜ํ•ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ์˜๋ฃŒ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ถ€์กฑ์œผ๋กœ ์ธํ•ด Wang et al. [72]๋Š” ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์•ฝํ•œ ๊ฐ๋… ๋ฐฉ์‹์œผ๋กœ Mamba ๊ธฐ๋ฐ˜ UNet์„ ํ›ˆ๋ จํ•˜๋Š” Weak-Mamba-Unet ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ์•„ํ‚คํ…์ฒ˜๋Š” convolutional neural networks (CNNs), Vision Transformers (ViTs) ๋ฐ Vmamba๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ๋ ˆ์ด๋ธ”์„ ์˜ˆ์ธกํ•œ ๋‹ค์Œ ๋ฐ€๋„ ์žˆ๋Š” ์˜์‚ฌ ๋ ˆ์ด๋ธ”์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
์˜๋ฃŒ ์ด๋ฏธ์ง€ ๊ธฐ๋ฐ˜ ๋ถ„์„์„ ์œ„ํ•ด Guo et al. [97]์€ Mamba ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์˜๋ฃŒ์šฉ MR-CT ๋ณ€ํ˜• ๋“ฑ๋ก ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•˜๋Š”๋ฐ, ์ด๋ฅผ MambaMorph๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์˜ ํ•ต์‹ฌ์€ ์˜๋ฃŒ ์ด๋ฏธ์ง€ ๋ถ„์„์— ์ค‘์š”ํ•œ ๋‹ค์–‘ํ•œ ์ด๋ฏธ์ง• ๋ฐฉ์‹ ๊ฐ„์˜ ๋ณต์…€ ์ˆ˜์ค€ ๊ณต๊ฐ„ ๋Œ€์‘ ์บก์ฒ˜๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ์žˆ์Šต๋‹ˆ๋‹ค. Xie et al.์€ ๋น„์ „ ๋ง˜๋ฐ” ์•„ํ‚คํ…์ฒ˜์™€ ํ”„๋กฌํ”„ํŠธ ๊ธฐ์ˆ ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒˆ๋กœ์šด ์šฉ์ข… ๋ถ„ํ•  ๋ชจ๋ธ ProMamba [99]๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์šฉ์ข… ๋ถ„ํ• ์— ๋น„์ „ ๋ง˜๋ฐ”์™€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋„์ž…ํ•œ ๊ฒƒ์€ ์ด๋ฒˆ์ด ์ฒ˜์Œ์ž…๋‹ˆ๋‹ค. Wu et al. [100]์€ ์˜๋ฃŒ ์ด๋ฏธ์ง€ ๋ถ„ํ• ์„ ์œ„ํ•ด SSM๊ณผ SS2D๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ƒˆ๋กœ์šด ์‹ ๊ฒฝ๋ง์ธ ๊ณ ์ฐจ ๋น„์ „ ๋ง˜๋ฐ” ์œ ๋„ท(H-vmunet)์„ ์†Œ๊ฐœํ•˜๋Š”๋ฐ, ์ด๋Š” ๊ณ ๊ธ‰ ์ƒํ˜ธ ์ž‘์šฉ์„ ํ†ตํ•ด ์ค‘๋ณต ์ •๋ณด์˜ ๋„์ž…์„ ์ ์ง„์ ์œผ๋กœ ์ค„์ด๊ณ  ๊ฐ ์ƒํ˜ธ ์ž‘์šฉ ๋‹จ๊ณ„์—์„œ SS2D์˜ ๋กœ์ปฌ ํŠน์ง• ํ•™์Šต ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. Vivim์€ ์˜๋ฃŒ์šฉ ๋น„๋””์˜ค ๊ฐ์ฒด ๋ถ„ํ• ์„ ์œ„ํ•ด ์„ค๊ณ„๋œ ์‹œ๊ฐ„ ๋ง˜๋ฐ” ๋ธ”๋ก์„ ํ†ตํ•ด ์žฅ๊ธฐ์ ์ธ ์‹œ๊ณต๊ฐ„ ํ‘œํ˜„์„ ๋‹ค์–‘ํ•œ ์Šค์ผ€์ผ์˜ ์‹œํ€€์Šค๋กœ ํšจ๊ณผ์ ์œผ๋กœ ์••์ถ•ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” Yang et al. [98]์— ์˜ํ•ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ์˜๋ฃŒ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ถ€์กฑ์œผ๋กœ ์ธํ•ด Wang et al. [72]๋Š” ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์•ฝํ•œ ๊ฐ๋… ๋ฐฉ์‹์œผ๋กœ Mamba ๊ธฐ๋ฐ˜ UNet์„ ํ›ˆ๋ จํ•˜๋Š” Weak-Mamba-Unet ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ์•„ํ‚คํ…์ฒ˜๋Š” ์ปจ๋ณผ๋ฃจ์…˜ ์‹ ๊ฒฝ๋ง(CNN), ๋น„์ „ ํŠธ๋žœ์Šคํฌ๋จธ(ViT) ๋ฐ Vmamba๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ๋ ˆ์ด๋ธ”์„ ์˜ˆ์ธกํ•œ ๋‹ค์Œ ๋ฐ€๋„ ์žˆ๋Š” ์˜์‚ฌ ๋ ˆ์ด๋ธ”์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
(4) _Self-driven Evolution._ In LMA3 [98], the agent can autonomously set goals for itself, and gradually improve its capability by exploring the environment and receiving feedback from a reward function. Following this mechanism, the agent can acquire knowledge and develop capabilities according to its own preferences. In SALLM-MS [99], by integrating advanced large language models like GPT-4 into a multi-agent system, agents can adapt and perform complex tasks, showcasing advanced communication capabilities, thereby realizing self-driven evolution in their interactions with the environment. In CLMTWA [100], by using a large language model as a teacher and a weaker language model as a student, the teacher can generate and communicate natural language explanations to improve the student's reasoning skills via theory of mind. The teacher can also personalize its explanations for the student and intervene only when necessary, based on the expected utility of intervention. In NLSOM [101], different agents communicate and collaborate through natural language to solve tasks that a single agent cannot solve. This can be seen as a form of self-driven learning, utilizing the exchange of information and knowledge between multiple agents. However, unlike other models such as LMA3, SALLM-MS, and CLMTWA, NLSOM allows for dynamic adjustment of agent roles, tasks, and relationships based on the task requirements and feedback from other agents or the environment.
(4) _Self-driven Evolution._ LMA3 [98]์—์„œ ์—์ด์ „ํŠธ๋Š” ์Šค์Šค๋กœ ๋ชฉํ‘œ๋ฅผ ์„ค์ •ํ•˜๊ณ  ํ™˜๊ฒฝ์„ ํƒ์ƒ‰ํ•˜๊ณ  ๋ณด์ƒ ํ•จ์ˆ˜์˜ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์•„ ์ ์ฐจ์ ์œผ๋กœ ์—ญ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ๋Š” ์ž์‹ ์˜ ์„ ํ˜ธ๋„์— ๋”ฐ๋ผ ์ง€์‹์„ ์Šต๋“ํ•˜๊ณ  ์—ญ๋Ÿ‰์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. SALLM-MS [99]์—์„œ๋Š” GPT-4์™€ ๊ฐ™์€ ๊ณ ๊ธ‰ Large Language Models (LLMs)๋ฅผ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์— ํ†ตํ•ฉํ•˜์—ฌ ์—์ด์ „ํŠธ๊ฐ€ ๋ณต์žกํ•œ ์ž‘์—…์„ ์กฐ์ •ํ•˜๊ณ  ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์—ฌ ๊ณ ๊ธ‰ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๊ณ  ํ™˜๊ฒฝ๊ณผ์˜ ์ƒํ˜ธ์ž‘์šฉ์—์„œ ์ž์œจ์  ์ง„ํ™”๋ฅผ ์‹คํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. CLMTWA [100]์—์„œ๋Š” Large Language Models (LLMs)๋ฅผ ๊ต์‚ฌ๋กœ, ์•ฝํ•œ ์–ธ์–ด ๋ชจ๋ธ์„ ํ•™์ƒ์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๊ต์‚ฌ๊ฐ€ ์ž์—ฐ์–ด ์„ค๋ช…์„ ์ƒ์„ฑํ•˜๊ณ  ์ „๋‹ฌํ•˜์—ฌ ๋งˆ์Œ ์ด๋ก ์„ ํ†ตํ•ด ํ•™์ƒ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ต์‚ฌ๋Š” ๊ฐœ์ž…์˜ ๊ธฐ๋Œ€ ํšจ์šฉ์— ๋”ฐ๋ผ ํ•™์ƒ์„ ์œ„ํ•œ ์„ค๋ช…์„ ๊ฐœ์ธํ™”ํ•˜๊ณ  ํ•„์š”ํ•œ ๊ฒฝ์šฐ์—๋งŒ ๊ฐœ์ž…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. NLSOM [101]์—์„œ๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๊ฐ€ ์ž์—ฐ์–ด๋ฅผ ํ†ตํ•ด ์†Œํ†ตํ•˜๊ณ  ํ˜‘๋ ฅํ•˜์—ฌ ๋‹จ์ผ ์—์ด์ „ํŠธ๊ฐ€ ํ•ด๊ฒฐํ•  ์ˆ˜ ์—†๋Š” ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ์ •๋ณด ๋ฐ ์ง€์‹ ๊ตํ™˜์„ ํ™œ์šฉํ•˜๋Š” ์ž๊ธฐ ์ฃผ๋„์  ํ•™์Šต์˜ ํ•œ ํ˜•ํƒœ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LMA3, SALLM-MS, CLMTWA์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋‹ฌ๋ฆฌ NLSOM์€ ์ž‘์—… ์š”๊ตฌ ์‚ฌํ•ญ๊ณผ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ ๋˜๋Š” ํ™˜๊ฒฝ์˜ ํ”ผ๋“œ๋ฐฑ์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ ์—ญํ• , ์ž‘์—… ๋ฐ ๊ด€๊ณ„๋ฅผ ๋™์ ์œผ๋กœ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
(4) _์ž์œจ์  ์ง„ํ™”._ LMA3 [98]์—์„œ ์—์ด์ „ํŠธ๋Š” ์Šค์Šค๋กœ ๋ชฉํ‘œ๋ฅผ ์„ค์ •ํ•˜๊ณ  ํ™˜๊ฒฝ์„ ํƒ์ƒ‰ํ•˜๊ณ  ๋ณด์ƒ ํ•จ์ˆ˜์˜ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์•„ ์ ์ฐจ์ ์œผ๋กœ ์—ญ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ๋Š” ์ž์‹ ์˜ ์„ ํ˜ธ๋„์— ๋”ฐ๋ผ ์ง€์‹์„ ์Šต๋“ํ•˜๊ณ  ์—ญ๋Ÿ‰์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. SALLM-MS [99]์—์„œ๋Š” GPT-4์™€ ๊ฐ™์€ ๊ณ ๊ธ‰ ์–ธ์–ด ๋ชจ๋ธ์„ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์— ํ†ตํ•ฉํ•˜์—ฌ ์—์ด์ „ํŠธ๊ฐ€ ๋ณต์žกํ•œ ์ž‘์—…์„ ์กฐ์ •ํ•˜๊ณ  ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์—ฌ ๊ณ ๊ธ‰ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๊ณ  ํ™˜๊ฒฝ๊ณผ์˜ ์ƒํ˜ธ์ž‘์šฉ์—์„œ ์ž์œจ์  ์ง„ํ™”๋ฅผ ์‹คํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. CLMTWA [100]์—์„œ๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ต์‚ฌ๋กœ, ์•ฝํ•œ ์–ธ์–ด ๋ชจ๋ธ์„ ํ•™์ƒ์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๊ต์‚ฌ๊ฐ€ ์ž์—ฐ์–ด ์„ค๋ช…์„ ์ƒ์„ฑํ•˜๊ณ  ์ „๋‹ฌํ•˜์—ฌ ๋งˆ์Œ ์ด๋ก ์„ ํ†ตํ•ด ํ•™์ƒ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ต์‚ฌ๋Š” ๊ฐœ์ž…์˜ ๊ธฐ๋Œ€ ํšจ์šฉ์— ๋”ฐ๋ผ ํ•™์ƒ์„ ์œ„ํ•œ ์„ค๋ช…์„ ๊ฐœ์ธํ™”ํ•˜๊ณ  ํ•„์š”ํ•œ ๊ฒฝ์šฐ์—๋งŒ ๊ฐœ์ž…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. NLSOM [101]์—์„œ๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๊ฐ€ ์ž์—ฐ์–ด๋ฅผ ํ†ตํ•ด ์†Œํ†ตํ•˜๊ณ  ํ˜‘๋ ฅํ•˜์—ฌ ๋‹จ์ผ ์—์ด์ „ํŠธ๊ฐ€ ํ•ด๊ฒฐํ•  ์ˆ˜ ์—†๋Š” ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ์ •๋ณด ๋ฐ ์ง€์‹ ๊ตํ™˜์„ ํ™œ์šฉํ•˜๋Š” ์ž๊ธฐ ์ฃผ๋„์  ํ•™์Šต์˜ ํ•œ ํ˜•ํƒœ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LMA3, SALLM-MS, CLMTWA์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋‹ฌ๋ฆฌ NLSOM์€ ์ž‘์—… ์š”๊ตฌ ์‚ฌํ•ญ๊ณผ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ ๋˜๋Š” ํ™˜๊ฒฝ์˜ ํ”ผ๋“œ๋ฐฑ์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ ์—ญํ• , ์ž‘์—… ๋ฐ ๊ด€๊ณ„๋ฅผ ๋™์ ์œผ๋กœ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Nowadays, LLMs are posing a significant impact on the AI community, and the advent of ChatGPT and GPT-4 leads to the rethinking of the possibilities of artificial general intelligence (AGI). OpenAI has published a technical article entitled _"Planning for AGI and beyond"_, which discusses the short-term and long-term plans to approach AGI [40], and a more recent paper has argued that GPT-4 might be considered as an early version of an AGI system [41]. The research areas of AI are being revolutionized by the rapid progress of LLMs. In the field of NLP, LLMs can serve as a general-purpose language task solver (to some extent), and the research paradigm has been shifting towards the use of LLMs. In the field of IR, traditional search engines are challenged by the new information seeking way through AI chatbots (_i.e.,_ ChatGPT), and _New Bing3_ presents an initial attempt that enhances the search results based on LLMs. In the field of CV, the researchers try to develop ChatGPT-like vision-language models that can better serve multimodal dialogues [42, 43, 44, 45], and GPT-4 [46] has supported multimodal input by integrating the visual information. This new wave of technology would potentially lead to a prosperous ecosystem of real-world applications based on LLMs. For instance, Microsoft 365 is being empowered by LLMs (_i.e.,_ Copilot) to automate the office work, and OpenAI supports the use of plugins in ChatGPT for implementing special functions.
์š”์ฆ˜ Large Language Models (LLMs)์€ AI ์ปค๋ฎค๋‹ˆํ‹ฐ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜๊ณ  ์žˆ์œผ๋ฉฐ, ChatGPT์™€ GPT-4์˜ ๋“ฑ์žฅ์œผ๋กœ ์ธ๊ณต ์ผ๋ฐ˜ ์ง€๋Šฅ(AGI)์˜ ๊ฐ€๋Šฅ์„ฑ์— ๋Œ€ํ•œ ์žฌ๊ณ ๊ฐ€ ์ด์–ด์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. OpenAI๋Š” AGI์— ์ ‘๊ทผํ•˜๊ธฐ ์œ„ํ•œ ๋‹จ๊ธฐ ๋ฐ ์žฅ๊ธฐ ๊ณ„ํš์„ ๋…ผ์˜ํ•˜๋Š” _"Planning for AGI and beyond"๋ผ๋Š” ์ œ๋ชฉ์˜ ๊ธฐ์ˆ  ๋ฌธ์„œ๋ฅผ ๋ฐœํ‘œํ–ˆ์œผ๋ฉฐ[40], ์ตœ๊ทผ ๋…ผ๋ฌธ์—์„œ๋Š” GPT-4๋ฅผ AGI ์‹œ์Šคํ…œ์˜ ์ดˆ๊ธฐ ๋ฒ„์ „์œผ๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ–ˆ์Šต๋‹ˆ๋‹ค[41]. AI์˜ ์—ฐ๊ตฌ ๋ถ„์•ผ๋Š” LLM์˜ ๊ธ‰์†ํ•œ ๋ฐœ์ „์œผ๋กœ ํ˜์‹ ์„ ๊ฑฐ๋“ญํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ LLM์€ ๋ฒ”์šฉ ์–ธ์–ด ์ž‘์—… ์†”๋ฒ„ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฐ๊ตฌ ํŒจ๋Ÿฌ๋‹ค์ž„์€ LLM ์‚ฌ์šฉ์œผ๋กœ ์ „ํ™˜๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ •๋ณด ๊ฒ€์ƒ‰(IR) ๋ถ„์•ผ์—์„œ๋Š” AI ์ฑ—๋ด‡(_์˜ˆ:_ ChatGPT)๋ฅผ ํ†ตํ•œ ์ƒˆ๋กœ์šด ์ •๋ณด ๊ฒ€์ƒ‰ ๋ฐฉ์‹์— ๊ธฐ์กด ๊ฒ€์ƒ‰ ์—”์ง„์ด ๋„์ „๋ฐ›๊ณ  ์žˆ์œผ๋ฉฐ, _New Bing3_๋Š” LLM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๋ฅผ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์ดˆ๊ธฐ ์‹œ๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ปดํ“จํ„ฐ ๋น„์ „(CV) ๋ถ„์•ผ์—์„œ๋Š” ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋Œ€ํ™”์— ๋” ์ž˜ ๋ถ€ํ•ฉํ•  ์ˆ˜ ์žˆ๋Š” ChatGPT์™€ ์œ ์‚ฌํ•œ ๋น„์ „ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•ด ๋…ธ๋ ฅํ•˜๊ณ  ์žˆ์œผ๋ฉฐ[42, 43, 44, 45], GPT-4[46]๋Š” ์‹œ๊ฐ ์ •๋ณด๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ž…๋ ฅ์„ ์ง€์›ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ƒˆ๋กœ์šด ๊ธฐ์ˆ ์€ ์ž ์žฌ์ ์œผ๋กœ LLM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์‹ค์ œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ๋ฒˆ์˜ํ•˜๋Š” ์ƒํƒœ๊ณ„๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Microsoft 365๋Š” ์‚ฌ๋ฌด ์ž‘์—…์„ ์ž๋™ํ™”ํ•˜๊ธฐ ์œ„ํ•ด LLM(_์˜ˆ:_ Copilot)์œผ๋กœ ๊ฐ•ํ™”๋˜๊ณ  ์žˆ์œผ๋ฉฐ, OpenAI๋Š” ํŠน์ˆ˜ ๊ธฐ๋Šฅ์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด ChatGPT์— ํ”Œ๋Ÿฌ๊ทธ์ธ ์‚ฌ์šฉ์„ ์ง€์›ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
์š”์ฆ˜ LLM์€ AI ์ปค๋ฎค๋‹ˆํ‹ฐ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜๊ณ  ์žˆ์œผ๋ฉฐ, ChatGPT์™€ GPT-4์˜ ๋“ฑ์žฅ์œผ๋กœ ์ธ๊ณต ์ผ๋ฐ˜ ์ง€๋Šฅ(AGI)์˜ ๊ฐ€๋Šฅ์„ฑ์— ๋Œ€ํ•œ ์žฌ๊ณ ๊ฐ€ ์ด์–ด์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. OpenAI๋Š” AGI์— ์ ‘๊ทผํ•˜๊ธฐ ์œ„ํ•œ ๋‹จ๊ธฐ ๋ฐ ์žฅ๊ธฐ ๊ณ„ํš์„ ๋…ผ์˜ํ•˜๋Š” _"AGI์™€ ๊ทธ ์ดํ›„๋ฅผ ์œ„ํ•œ ๊ณ„ํš"์ด๋ผ๋Š” ์ œ๋ชฉ์˜ ๊ธฐ์ˆ  ๋ฌธ์„œ๋ฅผ ๋ฐœํ‘œํ–ˆ์œผ๋ฉฐ[40], ์ตœ๊ทผ ๋…ผ๋ฌธ์—์„œ๋Š” GPT-4๋ฅผ AGI ์‹œ์Šคํ…œ์˜ ์ดˆ๊ธฐ ๋ฒ„์ „์œผ๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ–ˆ์Šต๋‹ˆ๋‹ค[41]. AI์˜ ์—ฐ๊ตฌ ๋ถ„์•ผ๋Š” LLM์˜ ๊ธ‰์†ํ•œ ๋ฐœ์ „์œผ๋กœ ํ˜์‹ ์„ ๊ฑฐ๋“ญํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ LLM์€ ๋ฒ”์šฉ ์–ธ์–ด ์ž‘์—… ์†”๋ฒ„ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฐ๊ตฌ ํŒจ๋Ÿฌ๋‹ค์ž„์€ LLM ์‚ฌ์šฉ์œผ๋กœ ์ „ํ™˜๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. IR ๋ถ„์•ผ์—์„œ๋Š” AI ์ฑ—๋ด‡(_์˜ˆ:_ ChatGPT๋ฅผ ํ†ตํ•œ ์ƒˆ๋กœ์šด ์ •๋ณด ๊ฒ€์ƒ‰ ๋ฐฉ์‹์— ๊ธฐ์กด ๊ฒ€์ƒ‰ ์—”์ง„์ด ๋„์ „๋ฐ›๊ณ  ์žˆ์œผ๋ฉฐ, _New Bing3_๋Š” LLM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๋ฅผ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์ดˆ๊ธฐ ์‹œ๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. CV ๋ถ„์•ผ์—์„œ๋Š” ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋Œ€ํ™”์— ๋” ์ž˜ ๋ถ€ํ•ฉํ•  ์ˆ˜ ์žˆ๋Š” ChatGPT์™€ ์œ ์‚ฌํ•œ ๋น„์ „ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•ด ๋…ธ๋ ฅํ•˜๊ณ  ์žˆ์œผ๋ฉฐ[42, 43, 44, 45], GPT-4[46]๋Š” ์‹œ๊ฐ ์ •๋ณด๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ž…๋ ฅ์„ ์ง€์›ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ƒˆ๋กœ์šด ๊ธฐ์ˆ ์€ ์ž ์žฌ์ ์œผ๋กœ LLM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์‹ค์ œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ๋ฒˆ์˜ํ•˜๋Š” ์ƒํƒœ๊ณ„๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Microsoft 365๋Š” ์‚ฌ๋ฌด ์ž‘์—…์„ ์ž๋™ํ™”ํ•˜๊ธฐ ์œ„ํ•ด LLM(_์˜ˆ:_ Copilot)์œผ๋กœ ๊ฐ•ํ™”๋˜๊ณ  ์žˆ์œผ๋ฉฐ, OpenAI๋Š” ํŠน์ˆ˜ ๊ธฐ๋Šฅ์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด ChatGPT์— ํ”Œ๋Ÿฌ๊ทธ์ธ ์‚ฌ์šฉ์„ ์ง€์›ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
**Overall Training Suggestion.** In practice, the above training techniques, especially 3D parallelism, are often jointly used to improve the training throughput and large model loading. For instance, researchers have incorporated 8-way data parallelism, 4-way tensor parallelism, and 12-way pipeline parallelism, enabling the training of BLOOM [78] on 384 A100 GPUs. Currently, open-source libraries like DeepSpeed [74], Colossal-AI [189], and Alpa [340] can well support the three parallel training methods. To reduce the memory redundancy, ZeRO, FSDP, and activation recomputation techniques [77, 341] can be also employed for training LLMs, which have already been integrated into DeepSpeed, PyTorch, and Megatron-LM. In addition, the mixed precision training technique such as BF16 can be also leveraged to improve the training efficiency and reduce GPU memory usage, while it requires necessary support on hardware (_e.g._, A100 GPU). Because training large models is a time-intensive process, it would be useful to forecast the model performance and detect abnormal issues at an early stage. For this purpose, GPT-4 [46] has recently introduced a new mechanism called _predictable scaling_ built on a deep learning stack, enabling the performance prediction of large models with a much smaller model, which might be quite useful for developing LLMs. In practice, one can further leverage the supporting training techniques of mainstream deep learning frameworks.
**์ „๋ฐ˜์ ์ธ ํ›ˆ๋ จ ์ œ์•ˆ**. ์‹ค์ œ๋กœ, ์œ„์˜ ํ›ˆ๋ จ ๊ธฐ๋ฒ•, ํŠนํžˆ 3D parallelism์ด ์ข…์ข… ๊ณต๋™์œผ๋กœ ์‚ฌ์šฉ๋˜์–ด ํ›ˆ๋ จ ์ฒ˜๋ฆฌ๋Ÿ‰๊ณผ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ ๋กœ๋”ฉ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์—ฐ๊ตฌ์ž๋“ค์€ 8-way data parallelism, 4-way tensor parallelism, 12-way pipeline parallelism์„ ํ†ตํ•ฉํ•˜์—ฌ 384๊ฐœ์˜ A100 GPU์—์„œ BLOOM [78]์„ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ DeepSpeed [74], Colossal-AI [189], Alpa [340]์™€ ๊ฐ™์€ ์˜คํ”ˆ ์†Œ์Šค ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋Š” ์„ธ ๊ฐ€์ง€ parallel training ๋ฐฉ๋ฒ•์„ ์ž˜ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ ์ค‘๋ณต์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ZeRO, FSDP ๋ฐ activation recomputation ๊ธฐ์ˆ  [77, 341]์„ LLM ํ›ˆ๋ จ์—๋„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ด๋ฏธ DeepSpeed, PyTorch ๋ฐ Megatron-LM์— ํ†ตํ•ฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ BF16๊ณผ ๊ฐ™์€ mixed precision training ๊ธฐ์ˆ ์„ ํ™œ์šฉํ•˜์—ฌ ํ›ˆ๋ จ ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ณ  GPU ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰์„ ์ค„์ผ ์ˆ˜ ์žˆ์ง€๋งŒ, ํ•˜๋“œ์›จ์–ด(_์˜ˆ: A100 GPU)์—์„œ ํ•„์š”ํ•œ ์ง€์›์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ฆฌ๋Š” ๊ณผ์ •์ด๋ฏ€๋กœ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•˜๊ณ  ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ๋น„์ •์ƒ์ ์ธ ๋ฌธ์ œ๋ฅผ ๊ฐ์ง€ํ•˜๋Š” ๊ฒƒ์ด ์œ ์šฉํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ์ตœ๊ทผ GPT-4 [46]๋Š” ๋”ฅ ๋Ÿฌ๋‹ ์Šคํƒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ 'predictable scaling'์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๋„์ž…ํ•˜์—ฌ ํ›จ์”ฌ ์ž‘์€ ๋ชจ๋ธ๋กœ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” LLM ๊ฐœ๋ฐœ์— ๋งค์šฐ ์œ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ์ฃผ๋ฅ˜ ๋”ฅ๋Ÿฌ๋‹ ํ”„๋ ˆ์ž„์›Œํฌ์˜ ์ง€์› ํ›ˆ๋ จ ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์ „๋ฐ˜์ ์ธ ํ›ˆ๋ จ ์ œ์•ˆ**. ์‹ค์ œ๋กœ, ์œ„์˜ ํ›ˆ๋ จ ๊ธฐ๋ฒ•, ํŠนํžˆ 3D ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๊ฐ€ ์ข…์ข… ๊ณต๋™์œผ๋กœ ์‚ฌ์šฉ๋˜์–ด ํ›ˆ๋ จ ์ฒ˜๋ฆฌ๋Ÿ‰๊ณผ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ ๋กœ๋”ฉ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์—ฐ๊ตฌ์ž๋“ค์€ 8๋ฐฉํ–ฅ ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ, 4๋ฐฉํ–ฅ ํ…์„œ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ, 12๋ฐฉํ–ฅ ํŒŒ์ดํ”„๋ผ์ธ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ 384๊ฐœ์˜ A100 GPU์—์„œ BLOOM [78]์„ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ DeepSpeed [74], Colossal-AI [189], Alpa [340]์™€ ๊ฐ™์€ ์˜คํ”ˆ ์†Œ์Šค ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋Š” ์„ธ ๊ฐ€์ง€ ๋ณ‘๋ ฌ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์„ ์ž˜ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ ์ค‘๋ณต์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ZeRO, FSDP ๋ฐ ํ™œ์„ฑํ™” ์žฌ๊ณ„์‚ฐ ๊ธฐ์ˆ  [77, 341]์„ LLM ํ›ˆ๋ จ์—๋„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ด๋ฏธ DeepSpeed, PyTorch ๋ฐ Megatron-LM์— ํ†ตํ•ฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ BF16๊ณผ ๊ฐ™์€ ํ˜ผํ•ฉ ์ •๋ฐ€๋„ ํ›ˆ๋ จ ๊ธฐ์ˆ ์„ ํ™œ์šฉํ•˜์—ฌ ํ›ˆ๋ จ ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ณ  GPU ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰์„ ์ค„์ผ ์ˆ˜ ์žˆ์ง€๋งŒ, ํ•˜๋“œ์›จ์–ด(_์˜ˆ: A100 GPU)์—์„œ ํ•„์š”ํ•œ ์ง€์›์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ฆฌ๋Š” ๊ณผ์ •์ด๋ฏ€๋กœ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•˜๊ณ  ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ๋น„์ •์ƒ์ ์ธ ๋ฌธ์ œ๋ฅผ ๊ฐ์ง€ํ•˜๋Š” ๊ฒƒ์ด ์œ ์šฉํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ์ตœ๊ทผ GPT-4 [46]๋Š” ๋”ฅ ๋Ÿฌ๋‹ ์Šคํƒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ '์˜ˆ์ธก ๊ฐ€๋Šฅํ•œ ์Šค์ผ€์ผ๋ง'์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๋„์ž…ํ•˜์—ฌ ํ›จ์”ฌ ์ž‘์€ ๋ชจ๋ธ๋กœ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” LLM ๊ฐœ๋ฐœ์— ๋งค์šฐ ์œ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ์ฃผ๋ฅ˜ ๋”ฅ๋Ÿฌ๋‹ ํ”„๋ ˆ์ž„์›Œํฌ์˜ ์ง€์› ํ›ˆ๋ จ ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Effect of pre-training data on LM performance:** Gao et al. [16] trains variants of GPT-2 [40] models from scratch to compare the "Pile" dataset to CommonCrawl-derived corpora. Radford et al. [40] demonstrates the positive impact of the quality filters and data de-duplication methods used to curate MassiveWeb by training 1.4B parameter models from scratch. Hernandez et al. [19] quantifies the effect of various amounts of artificially created data duplication and provides analysis on interpreting the changes in the behaviour of the models trained on duplicated data. Concurrently to our work, Xie et al. [56] propose using importance resampling to align the distribution of web data to high-quality reference corpora such as Wikipedia. Similarly, Gururangan et al. [17] explores data selection strategies for adapting LMs to a task-specific corpus. Another line of recent work explores how data mixture affects pre-training, with Xie et al. [55] demonstrating impressive improvements in downstream accuracy and perplexity across all datasets for 8B parameter models trained on the Pile. Similarly, Longpre et al. [30] explores the role of text quality, toxicity, age, and domain distribution of training data on LLM performance. Outside of data curation, there has been a recent surge of work exploring the impact of repeating data [5, 37, 57], generally concluding that repeating tokens is worse than training on new tokens (which we question in Section 4.2).
**์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๊ฐ€ LM ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ:** Gao ๋“ฑ[16]์€ GPT-2[40] ๋ชจ๋ธ์˜ ๋ณ€ํ˜•์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ›ˆ๋ จํ•˜์—ฌ "Pile" ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ CommonCrawl์—์„œ ํŒŒ์ƒ๋œ corpora์™€ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. Radford ๋“ฑ[40]์€ 1.4B parameter ๋ชจ๋ธ์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ›ˆ๋ จํ•˜์—ฌ MassiveWeb์„ ํ๋ ˆ์ด์…˜ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ ํ’ˆ์งˆ ํ•„ํ„ฐ์™€ ๋ฐ์ดํ„ฐ ์ค‘๋ณต ์ œ๊ฑฐ ๋ฐฉ๋ฒ•์˜ ๊ธ์ •์ ์ธ ์˜ํ–ฅ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. Hernandez ๋“ฑ[19]์€ ์ธ์œ„์ ์œผ๋กœ ์ƒ์„ฑ๋œ ๋‹ค์–‘ํ•œ ์–‘์˜ ๋ฐ์ดํ„ฐ ์ค‘๋ณต์˜ ์˜ํ–ฅ์„ ์ •๋Ÿ‰ํ™”ํ•˜๊ณ  ์ค‘๋ณต๋œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ๋™์ž‘์„ ํ•ด์„ํ•˜๋Š” ๋ถ„์„์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ ์—ฐ๊ตฌ์™€ ๋™์‹œ์— Xie ๋“ฑ[56]์€ ์ค‘์š”๋„ resampling์„ ์‚ฌ์šฉํ•˜์—ฌ ์›น ๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ๋ฅผ Wikipedia์™€ ๊ฐ™์€ ๊ณ ํ’ˆ์งˆ ์ฐธ์กฐ corpora์— ๋งž์ถ”๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, Gururangan ๋“ฑ[17]์€ LM์„ ์ž‘์—…๋ณ„ corpora์— ๋งž๊ฒŒ ์กฐ์ •ํ•˜๊ธฐ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์„ ํƒ์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์˜ ๋˜ ๋‹ค๋ฅธ ์—ฐ๊ตฌ๋“ค์€ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์ด ์‚ฌ์ „ ํ•™์Šต์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ํƒ๊ตฌํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, Xie ๋“ฑ[55]์€ Pile์—์„œ ํ•™์Šตํ•œ 8B parameter ๋ชจ๋ธ์— ๋Œ€ํ•ด ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ downstream ์ •ํ™•๋„์™€ perplexity๊ฐ€ ํฌ๊ฒŒ ํ–ฅ์ƒ๋จ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Longpre ๋“ฑ[30]์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ํ…์ŠคํŠธ ํ’ˆ์งˆ, ๋…์„ฑ, ์—ฐ๋ น ๋ฐ ๋„๋ฉ”์ธ ๋ถ„ํฌ๊ฐ€ LLM ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ์™ธ์—๋„ ์ตœ๊ทผ์—๋Š” ๋ฐ˜๋ณต๋˜๋Š” ๋ฐ์ดํ„ฐ์˜ ์˜ํ–ฅ์„ ํƒ๊ตฌํ•˜๋Š” ์—ฐ๊ตฌ๊ฐ€ ๊ธ‰์ฆํ•˜๊ณ  ์žˆ์œผ๋ฉฐ[5, 37, 57], ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒˆ๋กœ์šด ํ† ํฐ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋ฐ˜๋ณต๋˜๋Š” ํ† ํฐ์ด ๋” ๋‚˜์˜๋‹ค๋Š” ๊ฒฐ๋ก ์„ ๋‚ด๋ฆฌ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(4.2์ ˆ์—์„œ ์ด์— ๋Œ€ํ•ด ์˜๋ฌธ์„ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค).
**์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๊ฐ€ LM ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ:** Gao ๋“ฑ[16]์€ GPT-2[40] ๋ชจ๋ธ์˜ ๋ณ€ํ˜•์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ›ˆ๋ จํ•˜์—ฌ "Pile" ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ CommonCrawl์—์„œ ํŒŒ์ƒ๋œ ์ฝ”ํผ๋ผ์™€ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. Radford ๋“ฑ[40]์€ 1.4B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ›ˆ๋ จํ•˜์—ฌ MassiveWeb์„ ํ๋ ˆ์ด์…˜ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ ํ’ˆ์งˆ ํ•„ํ„ฐ์™€ ๋ฐ์ดํ„ฐ ์ค‘๋ณต ์ œ๊ฑฐ ๋ฐฉ๋ฒ•์˜ ๊ธ์ •์ ์ธ ์˜ํ–ฅ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. Hernandez ๋“ฑ[19]์€ ์ธ์œ„์ ์œผ๋กœ ์ƒ์„ฑ๋œ ๋‹ค์–‘ํ•œ ์–‘์˜ ๋ฐ์ดํ„ฐ ์ค‘๋ณต์˜ ์˜ํ–ฅ์„ ์ •๋Ÿ‰ํ™”ํ•˜๊ณ  ์ค‘๋ณต๋œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ๋™์ž‘์„ ํ•ด์„ํ•˜๋Š” ๋ถ„์„์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ ์—ฐ๊ตฌ์™€ ๋™์‹œ์— Xie ๋“ฑ[56]์€ ์ค‘์š”๋„ ๋ฆฌ์ƒ˜ํ”Œ๋ง์„ ์‚ฌ์šฉํ•˜์—ฌ ์›น ๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ๋ฅผ Wikipedia์™€ ๊ฐ™์€ ๊ณ ํ’ˆ์งˆ ์ฐธ์กฐ ์ฝ”ํผ์Šค์— ๋งž์ถ”๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ๊ตฌ๋ฃจ๋ผ๊ฐ„ ๋“ฑ[17]์€ LM์„ ์ž‘์—…๋ณ„ ์ฝ”ํผ์Šค์— ๋งž๊ฒŒ ์กฐ์ •ํ•˜๊ธฐ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์„ ํƒ์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์˜ ๋˜ ๋‹ค๋ฅธ ์—ฐ๊ตฌ๋“ค์€ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์ด ์‚ฌ์ „ ํ•™์Šต์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ํƒ๊ตฌํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, Xie ๋“ฑ[55]์€ Pile์—์„œ ํ•™์Šตํ•œ 8B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ •ํ™•๋„์™€ ๋‹นํ˜น๊ฐ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋จ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Longpre ๋“ฑ[30]์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ํ…์ŠคํŠธ ํ’ˆ์งˆ, ๋…์„ฑ, ์—ฐ๋ น ๋ฐ ๋„๋ฉ”์ธ ๋ถ„ํฌ๊ฐ€ LLM ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ์™ธ์—๋„ ์ตœ๊ทผ์—๋Š” ๋ฐ˜๋ณต๋˜๋Š” ๋ฐ์ดํ„ฐ์˜ ์˜ํ–ฅ์„ ํƒ๊ตฌํ•˜๋Š” ์—ฐ๊ตฌ๊ฐ€ ๊ธ‰์ฆํ•˜๊ณ  ์žˆ์œผ๋ฉฐ[5, 37, 57], ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒˆ๋กœ์šด ํ† ํฐ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋ฐ˜๋ณต๋˜๋Š” ํ† ํฐ์ด ๋” ๋‚˜์˜๋‹ค๋Š” ๊ฒฐ๋ก ์„ ๋‚ด๋ฆฌ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(4.2์ ˆ์—์„œ ์ด์— ๋Œ€ํ•ด ์˜๋ฌธ์„ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค).
Instruction-following LMs.A series of works have found evidence that vanilla LMs can be effective at following general language instructions if tuned with annotated "instructional" data--datasets containing language instructional commands and their desired outcomes based on human annotation (Weller et al., 2020, Mishra et al., 2022, 2022, 2022, 2022, 2022, 2022, 2022). Additionally, they show a direct correlation between the size and diversity of the "instructional" data and the generalizability of resulting models to unseen tasks (Wang et al., 2022, 2022). However, since these developments largely focus on existing NLP tasks and depend on human-annotated instructions, this poses a bottleneck for progress toward more generalizable models (e.g., see Fig. 5a in Wang et al., 2022). Our work aims to move beyond classical NLP tasks and tackle the challenges of creating diverse instruction data by employing pretrained LMs. InstructGPT (Ouyang et al., 2022) shares a similar goal as ours in building more general-purpose LMs, and has demonstrated remarkable performance in following diverse user instructions. However, as a commercial system, their construction process still remains quite opaque. In particular, the role of _data_ has remained understudied due to limited transparency and the private user data they used in their study. Addressing such challenges necessitates the creation of a large-scale, public dataset covering a broad range of tasks.
๋ช…๋ น์–ด๋ฅผ ๋”ฐ๋ฅด๋Š” LMs. ์ผ๋ จ์˜ ์—ฐ๊ตฌ์—์„œ, annotation๋œ "instructional" ๋ฐ์ดํ„ฐ(์ธ๊ฐ„์˜ ์ฃผ์„์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์–ธ์–ด ๋ช…๋ น์–ด์™€ ์›ํ•˜๋Š” ๊ฒฐ๊ณผ๊ฐ€ ํฌํ•จ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ)๋กœ ์กฐ์ •ํ•˜๋ฉด ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ๋ช…๋ น์–ด๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๋”ฐ๋ฅผ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ฆ๊ฑฐ๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(Weller et al., 2020, Mishra et al., 2022, 2022, 2022, 2022, 2022, 2022, 2022, 2022, 2022). ๋˜ํ•œ, "instructional" ๋ฐ์ดํ„ฐ์˜ ๊ทœ๋ชจ์™€ ๋‹ค์–‘์„ฑ๊ณผ ๊ฒฐ๊ณผ ๋ชจ๋ธ์ด ๋ณด์ด์ง€ ์•Š๋Š” ์ž‘์—…์— ๋Œ€ํ•œ ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅ์„ฑ ์‚ฌ์ด์—๋Š” ์ง์ ‘์ ์ธ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Wang et al., 2022, 2022). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๊ฐœ๋ฐœ์€ ๋Œ€๋ถ€๋ถ„ ๊ธฐ์กด NLP ์ž‘์—…์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹ฌ์•„์•ผ ํ•˜๋Š” instruction์— ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ณด๋‹ค ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ์„ ํ–ฅํ•œ ์ง„์ „์— ๋ณ‘๋ชฉ ํ˜„์ƒ์„ ์ดˆ๋ž˜ํ•ฉ๋‹ˆ๋‹ค(์˜ˆ: Wang et al., 2022์˜ ๊ทธ๋ฆผ 5a ์ฐธ์กฐ). ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๊ธฐ์กด NLP ์ž‘์—…์„ ๋„˜์–ด์„œ ์‚ฌ์ „ ํ•™์Šต๋œ LMs์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ instruction ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. InstructGPT(Ouyang et al., 2022)๋Š” ๋ณด๋‹ค ๋ฒ”์šฉ์ ์ธ LMs์„ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์— ์žˆ์–ด ์šฐ๋ฆฌ์™€ ๋น„์Šทํ•œ ๋ชฉํ‘œ๋ฅผ ๊ณต์œ ํ•˜๋ฉฐ, ๋‹ค์–‘ํ•œ ์‚ฌ์šฉ์ž instruction์„ ๋”ฐ๋ฅด๋Š” ๋†€๋ผ์šด ์„ฑ๋Šฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ƒ์šฉ ์‹œ์Šคํ…œ์œผ๋กœ์„œ ๊ทธ ๊ตฌ์ถ• ๊ณผ์ •์€ ์—ฌ์ „ํžˆ ์ƒ๋‹นํžˆ ๋ถˆํˆฌ๋ช…ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ œํ•œ๋œ ํˆฌ๋ช…์„ฑ๊ณผ ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋œ ๋น„๊ณต๊ฐœ ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ๋กœ ์ธํ•ด _data_์˜ ์—ญํ• ์€ ์•„์ง ์—ฐ๊ตฌ๋˜์ง€ ์•Š์€ ์ฑ„๋กœ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋ ค๋ฉด ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—…์„ ํฌ๊ด„ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ๊ณต๊ฐœ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“ค์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๋ช…๋ น์–ด๋ฅผ ๋”ฐ๋ฅด๋Š” LM.์ผ๋ จ์˜ ์—ฐ๊ตฌ์—์„œ, annotation๋œ '๋ช…๋ น์–ด' ๋ฐ์ดํ„ฐ(์ธ๊ฐ„์˜ ์ฃผ์„์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์–ธ์–ด ๋ช…๋ น์–ด์™€ ์›ํ•˜๋Š” ๊ฒฐ๊ณผ๊ฐ€ ํฌํ•จ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ)๋กœ ์กฐ์ •ํ•˜๋ฉด ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ๋ช…๋ น์–ด๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๋”ฐ๋ฅผ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ฆ๊ฑฐ๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(Weller et al., 2020, Mishra et al., 2022, 2022, 2022, 2022, 2022, 2022, 2022, 2022, 2022). ๋˜ํ•œ, '๋ช…๋ น์–ด' ๋ฐ์ดํ„ฐ์˜ ๊ทœ๋ชจ์™€ ๋‹ค์–‘์„ฑ๊ณผ ๊ฒฐ๊ณผ ๋ชจ๋ธ์ด ๋ณด์ด์ง€ ์•Š๋Š” ์ž‘์—…์— ๋Œ€ํ•œ ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅ์„ฑ ์‚ฌ์ด์—๋Š” ์ง์ ‘์ ์ธ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Wang et al., 2022, 2022). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๊ฐœ๋ฐœ์€ ๋Œ€๋ถ€๋ถ„ ๊ธฐ์กด NLP ์ž‘์—…์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹ฌ์•„์•ผ ํ•˜๋Š” ๋ช…๋ น์–ด์— ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ณด๋‹ค ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ์„ ํ–ฅํ•œ ์ง„์ „์— ๋ณ‘๋ชฉ ํ˜„์ƒ์„ ์ดˆ๋ž˜ํ•ฉ๋‹ˆ๋‹ค(์˜ˆ: Wang et al., 2022์˜ ๊ทธ๋ฆผ 5a ์ฐธ์กฐ). ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๊ธฐ์กด NLP ์ž‘์—…์„ ๋„˜์–ด์„œ ์‚ฌ์ „ ํ•™์Šต๋œ LM์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. InstructGPT(Ouyang et al., 2022)๋Š” ๋ณด๋‹ค ๋ฒ”์šฉ์ ์ธ LM์„ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์— ์žˆ์–ด ์šฐ๋ฆฌ์™€ ๋น„์Šทํ•œ ๋ชฉํ‘œ๋ฅผ ๊ณต์œ ํ•˜๋ฉฐ, ๋‹ค์–‘ํ•œ ์‚ฌ์šฉ์ž ๋ช…๋ น์„ ๋”ฐ๋ฅด๋Š” ๋†€๋ผ์šด ์„ฑ๋Šฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ƒ์šฉ ์‹œ์Šคํ…œ์œผ๋กœ์„œ ๊ทธ ๊ตฌ์ถ• ๊ณผ์ •์€ ์—ฌ์ „ํžˆ ์ƒ๋‹นํžˆ ๋ถˆํˆฌ๋ช…ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ œํ•œ๋œ ํˆฌ๋ช…์„ฑ๊ณผ ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋œ ๋น„๊ณต๊ฐœ ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ๋กœ ์ธํ•ด _๋ฐ์ดํ„ฐ_์˜ ์—ญํ• ์€ ์•„์ง ์—ฐ๊ตฌ๋˜์ง€ ์•Š์€ ์ฑ„๋กœ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋ ค๋ฉด ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—…์„ ํฌ๊ด„ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ๊ณต๊ฐœ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“ค์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Instruction Following.Synthetic data can serve as a promising approach for training instruction-following models, particularly in scenarios where real-world data is scarce, expensive, or challenging to obtain. Self-instruct (Wang et al., 2022a) and Stanford Alpaca (Taori et al., 2023) are both using LLMs to generate instruction following data which covers a wide range of scenarios. They first pick a small set of "seed instruction following samples" and then ask the LLMs to imitate the format to generate more demonstrations. One concern of this type of method is how to keep the generated data high quality, which involves the complexity of queries (Liu et al., 2023d), the diversity of semantics (Ding et al., 2023), and the scale of the synthetic dataset (Yuan et al., 2023). To this end, Xu et al. (2023) propose Evol-Instruct which adds complexity to simple instructions via prompting. Mukherjee et al. (2023) leverage LLMs to revise the instructions and responses iteratively to include high-quality explanation traces in the FLAN dataset (Wei et al., 2022), and they find the trained model has improved performance in many NLP tasks. UltraChat (Ding et al., 2023) is large-scale and multi-round synthetic dialogue dataset, which is generated by two separate ChatGPT Turbo API models--one serves as the user role while the other serves as the assistant. They instruct the user model with carefully designed prompts to mimic real human user behaviors.
๋ช…๋ น์–ด ๋”ฐ๋ผํ•˜๊ธฐ. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ํŠนํžˆ ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๊ฑฐ๋‚˜ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ฑฐ๋‚˜ ํ™•๋ณดํ•˜๊ธฐ ์–ด๋ ค์šด ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ Instruction-following ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์œ ๋งํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Self-instruct(Wang et al., 2022a)์™€ Stanford Alpaca(Taori et al., 2023)๋Š” ๋ชจ๋‘ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํฌ๊ด„ํ•˜๋Š” Instruction-following ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € 'seed Instruction-following samples'์˜ ์ž‘์€ ์ง‘ํ•ฉ์„ ์„ ํƒํ•œ ๋‹ค์Œ, LLM์ด ์ด ํ˜•์‹์„ ๋ชจ๋ฐฉํ•˜์—ฌ ๋” ๋งŽ์€ ๋ฐ๋ชจ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ๋ฐฉ๋ฒ•์—์„œ ํ•œ ๊ฐ€์ง€ ์šฐ๋ ค๋˜๋Š” ์ ์€ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์„ ๋†’์ด ์œ ์ง€ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด๋ฉฐ, ์—ฌ๊ธฐ์—๋Š” ์ฟผ๋ฆฌ์˜ ๋ณต์žก์„ฑ(Liu et al., 2023d), ์˜๋ฏธ์˜ ๋‹ค์–‘์„ฑ(Ding et al., 2023), ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ทœ๋ชจ(Yuan et al., 2023) ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด Xu ๋“ฑ(2023)์€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด ๊ฐ„๋‹จํ•œ ๋ช…๋ น์–ด์— ๋ณต์žก์„ฑ์„ ์ถ”๊ฐ€ํ•˜๋Š” Evol-Instruct๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Mukherjee ๋“ฑ(2023)์€ LLM์„ ํ™œ์šฉํ•˜์—ฌ ๋ช…๋ น์–ด์™€ ์‘๋‹ต์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ˆ˜์ •ํ•˜์—ฌ FLAN ๋ฐ์ดํ„ฐ ์„ธํŠธ(Wei ๋“ฑ, 2022)์— ๊ณ ํ’ˆ์งˆ ์„ค๋ช… ์ถ”์ ์„ ํฌํ•จ์‹œํ‚ค๊ณ , ํ•™์Šต๋œ ๋ชจ๋ธ์ด ๋งŽ์€ NLP ์ž‘์—…์—์„œ ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. UltraChat(Ding et al., 2023)์€ ๋Œ€๊ทœ๋ชจ ๋‹ค์ค‘ ๋ผ์šด๋“œ ํ•ฉ์„ฑ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ, ํ•œ ๋ชจ๋ธ์€ ์‚ฌ์šฉ์ž ์—ญํ• ์„, ๋‹ค๋ฅธ ๋ชจ๋ธ์€ ์–ด์‹œ์Šคํ„ดํŠธ ์—ญํ• ์„ ํ•˜๋Š” ๋‘ ๊ฐœ์˜ ๋ณ„๋„์˜ ChatGPT Turbo API ๋ชจ๋ธ์— ์˜ํ•ด ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž ๋ชจ๋ธ์— ์‹ ์ค‘ํ•˜๊ฒŒ ์„ค๊ณ„๋œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด ์‹ค์ œ ์‚ฌ์šฉ์ž ํ–‰๋™์„ ๋ชจ๋ฐฉํ•˜๋„๋ก ์ง€์‹œํ•ฉ๋‹ˆ๋‹ค.
๋ช…๋ น์–ด ๋”ฐ๋ผํ•˜๊ธฐ.ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ํŠนํžˆ ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๊ฑฐ๋‚˜ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ฑฐ๋‚˜ ํ™•๋ณดํ•˜๊ธฐ ์–ด๋ ค์šด ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋ช…๋ น์–ด ์ถ”์ข… ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์œ ๋งํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์…€ํ”„์ธ์ŠคํŠธ๋ŸญํŠธ(Wang et al., 2022a)์™€ ์Šคํƒ ํฌ๋“œ ์•ŒํŒŒ์นด(Taori et al., 2023)๋Š” ๋ชจ๋‘ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํฌ๊ด„ํ•˜๋Š” ๋ช…๋ น์–ด ์ถ”์ข… ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € '์‹œ๋“œ ๋ช…๋ น์–ด ์ถ”์ข… ์ƒ˜ํ”Œ'์˜ ์ž‘์€ ์ง‘ํ•ฉ์„ ์„ ํƒํ•œ ๋‹ค์Œ, LLM์ด ์ด ํ˜•์‹์„ ๋ชจ๋ฐฉํ•˜์—ฌ ๋” ๋งŽ์€ ๋ฐ๋ชจ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ๋ฐฉ๋ฒ•์—์„œ ํ•œ ๊ฐ€์ง€ ์šฐ๋ ค๋˜๋Š” ์ ์€ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์„ ๋†’์ด ์œ ์ง€ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด๋ฉฐ, ์—ฌ๊ธฐ์—๋Š” ์ฟผ๋ฆฌ์˜ ๋ณต์žก์„ฑ(Liu et al., 2023d), ์˜๋ฏธ์˜ ๋‹ค์–‘์„ฑ(Ding et al., 2023), ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ทœ๋ชจ(Yuan et al., 2023) ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด Xu ๋“ฑ(2023)์€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด ๊ฐ„๋‹จํ•œ ๋ช…๋ น์–ด์— ๋ณต์žก์„ฑ์„ ์ถ”๊ฐ€ํ•˜๋Š” Evol-Instruct๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Mukherjee ๋“ฑ(2023)์€ LLM์„ ํ™œ์šฉํ•˜์—ฌ ๋ช…๋ น์–ด์™€ ์‘๋‹ต์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ˆ˜์ •ํ•˜์—ฌ FLAN ๋ฐ์ดํ„ฐ ์„ธํŠธ(Wei ๋“ฑ, 2022)์— ๊ณ ํ’ˆ์งˆ ์„ค๋ช… ์ถ”์ ์„ ํฌํ•จ์‹œํ‚ค๊ณ , ํ•™์Šต๋œ ๋ชจ๋ธ์ด ๋งŽ์€ NLP ์ž‘์—…์—์„œ ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. UltraChat(Ding et al., 2023)์€ ๋Œ€๊ทœ๋ชจ ๋‹ค์ค‘ ๋ผ์šด๋“œ ํ•ฉ์„ฑ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ, ํ•œ ๋ชจ๋ธ์€ ์‚ฌ์šฉ์ž ์—ญํ• ์„, ๋‹ค๋ฅธ ๋ชจ๋ธ์€ ์–ด์‹œ์Šคํ„ดํŠธ ์—ญํ• ์„ ํ•˜๋Š” ๋‘ ๊ฐœ์˜ ๋ณ„๋„์˜ ChatGPT ํ„ฐ๋ณด API ๋ชจ๋ธ์— ์˜ํ•ด ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž ๋ชจ๋ธ์— ์‹ ์ค‘ํ•˜๊ฒŒ ์„ค๊ณ„๋œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด ์‹ค์ œ ์‚ฌ์šฉ์ž ํ–‰๋™์„ ๋ชจ๋ฐฉํ•˜๋„๋ก ์ง€์‹œํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_Task-level predictability_. Existing research of scaling laws are mostly conducted in terms of language modeling loss (_e.g._, per-token cross-entropy loss in nats [30]), while in practice we are more concerned about the performance of LLMs on actual tasks. Thus, a basic problem is that how the decrease of language modeling loss translates into the improvement of task performance [58]. Intuitively, a model with a smaller language modeling loss tends to yield a better performance on downstream tasks, since language modeling loss can be considered as a general measure of the overall model capacity. GPT-4 [46] has reported that some capabilities (_e.g._, coding ability) can be accurately predicted via scaling law. Despite that, readers should be aware that a direct decrease in language modeling loss does not always indicate an improvement of model performance on downstream tasks. Specially, the phenomenon of _inverse scaling_ would occur for some tasks, where task performance surprisingly becomes worse as the language modeling loss decreases [62]. Overall, it is more difficult to explore and characterize task-level scaling laws, since it might be also dependent on task-related information (task metric, task difficulty, etc.). Furthermore, some capacities (_e.g._, in-context learning [55]) are unpredictable according to the scaling law, which can be observed only when the model size exceeds a certain level (as discussed below).
\(\bullet\)_์ž‘์—… ์ˆ˜์ค€์˜ ์˜ˆ์ธก ๊ฐ€๋Šฅ์„ฑ_. Scaling laws์— ๋Œ€ํ•œ ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ๋Œ€๋ถ€๋ถ„ ์–ธ์–ด ๋ชจ๋ธ๋ง Loss(_์˜ˆ: nats[30]์˜ ํ† ํฐ๋‹น ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ Loss)์˜ ๊ด€์ ์—์„œ ์ˆ˜ํ–‰๋˜๋Š” ๋ฐ˜๋ฉด, ์‹ค์ œ๋กœ๋Š” ์‹ค์ œ ์ž‘์—…์—์„œ LLM์˜ ์„ฑ๋Šฅ์— ๋” ๊ด€์‹ฌ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ธฐ๋ณธ์ ์ธ ๋ฌธ์ œ๋Š” ์–ธ์–ด ๋ชจ๋ธ๋ง Loss์˜ ๊ฐ์†Œ๊ฐ€ ์ž‘์—… ์„ฑ๋Šฅ์˜ ๊ฐœ์„ ์œผ๋กœ ์–ด๋–ป๊ฒŒ ์ด์–ด์ง€๋Š”์ง€์— ๋Œ€ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค[58]. ์ง๊ด€์ ์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ๋ง Loss๋Š” ์ „์ฒด ๋ชจ๋ธ ์šฉ๋Ÿ‰์˜ ์ผ๋ฐ˜์ ์ธ ์ฒ™๋„๋กœ ๊ฐ„์ฃผ๋  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์–ธ์–ด ๋ชจ๋ธ๋ง Loss๊ฐ€ ์ž‘์€ ๋ชจ๋ธ์ผ์ˆ˜๋ก ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. GPT-4 [46]๋Š” ์ผ๋ถ€ ๊ธฐ๋Šฅ(์˜ˆ: ์ฝ”๋”ฉ ๋Šฅ๋ ฅ)์„ Scaling law๋ฅผ ํ†ตํ•ด ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ณด๊ณ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋…์ž๋“ค์€ ์–ธ์–ด ๋ชจ๋ธ๋ง Loss์˜ ์ง์ ‘์ ์ธ ๊ฐ์†Œ๊ฐ€ ํ•ญ์ƒ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ๋ชจ๋ธ ์„ฑ๋Šฅ์˜ ํ–ฅ์ƒ์„ ์˜๋ฏธํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ผ๋Š” ์ ์„ ์ธ์ง€ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์–ธ์–ด ๋ชจ๋ธ๋ง Loss๊ฐ€ ๊ฐ์†Œํ•จ์— ๋”ฐ๋ผ ์ž‘์—… ์„ฑ๋Šฅ์ด ๋†€๋ž๊ฒŒ๋„ ์•…ํ™”๋˜๋Š” _์—ญ์  Scaling_ ํ˜„์ƒ์ด ์ผ๋ถ€ ์ž‘์—…์—์„œ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[62]. ์ „๋ฐ˜์ ์œผ๋กœ ์ž‘์—… ์ˆ˜์ค€์˜ Scaling laws๋ฅผ ํƒ์ƒ‰ํ•˜๊ณ  ํŠน์„ฑํ™”ํ•˜๋Š” ๊ฒƒ์€ ์ž‘์—… ๊ด€๋ จ ์ •๋ณด(์ž‘์—… metric, ์ž‘์—… ๋‚œ์ด๋„ ๋“ฑ)์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋” ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ผ๋ถ€ ๊ธฐ๋Šฅ(์˜ˆ: In-Context Learning [55])์€ Scaling law์— ๋”ฐ๋ผ ์˜ˆ์ธกํ•  ์ˆ˜ ์—†์œผ๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ํŠน์ • ์ˆ˜์ค€์„ ์ดˆ๊ณผํ•  ๋•Œ๋งŒ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์•„๋ž˜์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด).
\(\bullet\)_์ž‘์—… ์ˆ˜์ค€์˜ ์˜ˆ์ธก ๊ฐ€๋Šฅ์„ฑ_. ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋Œ€ํ•œ ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ๋Œ€๋ถ€๋ถ„ ์–ธ์–ด ๋ชจ๋ธ๋ง ์†์‹ค(_์˜ˆ: nats[30]์˜ ํ† ํฐ๋‹น ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ ์†์‹ค)์˜ ๊ด€์ ์—์„œ ์ˆ˜ํ–‰๋˜๋Š” ๋ฐ˜๋ฉด, ์‹ค์ œ๋กœ๋Š” ์‹ค์ œ ์ž‘์—…์—์„œ LLM์˜ ์„ฑ๋Šฅ์— ๋” ๊ด€์‹ฌ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ธฐ๋ณธ์ ์ธ ๋ฌธ์ œ๋Š” ์–ธ์–ด ๋ชจ๋ธ๋ง ์†์‹ค์˜ ๊ฐ์†Œ๊ฐ€ ์ž‘์—… ์„ฑ๋Šฅ์˜ ๊ฐœ์„ ์œผ๋กœ ์–ด๋–ป๊ฒŒ ์ด์–ด์ง€๋Š”์ง€์— ๋Œ€ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค[58]. ์ง๊ด€์ ์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ๋ง ์†์‹ค์€ ์ „์ฒด ๋ชจ๋ธ ์šฉ๋Ÿ‰์˜ ์ผ๋ฐ˜์ ์ธ ์ฒ™๋„๋กœ ๊ฐ„์ฃผ๋  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์–ธ์–ด ๋ชจ๋ธ๋ง ์†์‹ค์ด ์ž‘์€ ๋ชจ๋ธ์ผ์ˆ˜๋ก ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. GPT-4 [46]๋Š” ์ผ๋ถ€ ๊ธฐ๋Šฅ(์˜ˆ: ์ฝ”๋”ฉ ๋Šฅ๋ ฅ)์„ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ํ†ตํ•ด ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ณด๊ณ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋…์ž๋“ค์€ ์–ธ์–ด ๋ชจ๋ธ๋ง ์†์‹ค์˜ ์ง์ ‘์ ์ธ ๊ฐ์†Œ๊ฐ€ ํ•ญ์ƒ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ๋ชจ๋ธ ์„ฑ๋Šฅ์˜ ํ–ฅ์ƒ์„ ์˜๋ฏธํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ผ๋Š” ์ ์„ ์ธ์ง€ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์–ธ์–ด ๋ชจ๋ธ๋ง ์†์‹ค์ด ๊ฐ์†Œํ•จ์— ๋”ฐ๋ผ ์ž‘์—… ์„ฑ๋Šฅ์ด ๋†€๋ž๊ฒŒ๋„ ์•…ํ™”๋˜๋Š” _์—ญ์  ์Šค์ผ€์ผ๋ง_ ํ˜„์ƒ์ด ์ผ๋ถ€ ์ž‘์—…์—์„œ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[62]. ์ „๋ฐ˜์ ์œผ๋กœ ์ž‘์—… ์ˆ˜์ค€์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ํƒ์ƒ‰ํ•˜๊ณ  ํŠน์„ฑํ™”ํ•˜๋Š” ๊ฒƒ์€ ์ž‘์—… ๊ด€๋ จ ์ •๋ณด(์ž‘์—… ๋ฉ”ํŠธ๋ฆญ, ์ž‘์—… ๋‚œ์ด๋„ ๋“ฑ)์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋” ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ผ๋ถ€ ๊ธฐ๋Šฅ(์˜ˆ: ์ปจํ…์ŠคํŠธ ๋‚ด ํ•™์Šต [55])์€ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋”ฐ๋ผ ์˜ˆ์ธกํ•  ์ˆ˜ ์—†์œผ๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ํŠน์ • ์ˆ˜์ค€์„ ์ดˆ๊ณผํ•  ๋•Œ๋งŒ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์•„๋ž˜์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด).
Performance Analysis.To answer the efficiency question, we conduct a comparative analysis between AIOS employing FIFO scheduling and a non-scheduled approach, wherein the aforementioned three agents run concurrently. In the non-scheduled setting, the three agents are executed following a predefined sequential order: Math Agent, Narrative Agent, and Rec Agent. We employ two metrics for assessing temporal efficiency: waiting time (the interval from agent request submission to commencement) and turnaround time (the duration from agent request submission to completion). As each agent will send multiple requests to the LLM, the waiting time and turnaround time of each agent are calculated as the average of waiting time and turnaround time of all its sent requests, respectively. To mitigate randomness, we execute these three agents, both with and without scheduling, in five separate trials to report the results. As shown in Table 5, the non-scheduled approach exhibits good performance for agents earlier in the sequence, but at the expense of extended waiting time and turnaround time for agents later in the sequence. Conversely, AIOS's scheduling mechanism efficiently regulates both waiting time and turnaround time, a benefit that becomes particularly evident for agent requests submitted later by agents, especially when LLM is large. This suggests the importance of our scheduling to accommodate parallel operations of multiple agents.
์„ฑ๋Šฅ ๋ถ„์„. ํšจ์œจ์„ฑ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด FIFO ์Šค์ผ€์ค„๋ง์„ ์‚ฌ์šฉํ•˜๋Š” AIOS์™€ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์„ธ ๊ฐ€์ง€ ์—์ด์ „ํŠธ๊ฐ€ ๋™์‹œ์— ์‹คํ–‰๋˜๋Š” ๋น„์Šค์ผ€์ค„ ๋ฐฉ์‹ ๊ฐ„์˜ ๋น„๊ต ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋น„์Šค์ผ€์ค„ ๋ฐฉ์‹์—์„œ๋Š” Math Agent, Narrative Agent, Rec Agent์˜ ์‚ฌ์ „ ์ •์˜๋œ ์ˆœ์„œ์— ๋”ฐ๋ผ ์„ธ ๊ฐ€์ง€ ์—์ด์ „ํŠธ๊ฐ€ ์‹คํ–‰๋ฉ๋‹ˆ๋‹ค. ์‹œ๊ฐ„์  ํšจ์œจ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐ€์ง€ ๋ฉ”ํŠธ๋ฆญ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค: ๋Œ€๊ธฐ ์‹œ๊ฐ„(์—์ด์ „ํŠธ ์š”์ฒญ ์ œ์ถœ๋ถ€ํ„ฐ ์‹œ์ž‘๊นŒ์ง€์˜ ๊ฐ„๊ฒฉ)๊ณผ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„(์—์ด์ „ํŠธ ์š”์ฒญ ์ œ์ถœ๋ถ€ํ„ฐ ์™„๋ฃŒ๊นŒ์ง€์˜ ์‹œ๊ฐ„)์ž…๋‹ˆ๋‹ค. ๊ฐ ์—์ด์ „ํŠธ๋Š” LLM์— ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์š”์ฒญ์„ ๋ณด๋‚ด๋ฏ€๋กœ ๊ฐ ์—์ด์ „ํŠธ์˜ ๋Œ€๊ธฐ ์‹œ๊ฐ„๊ณผ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์€ ๊ฐ๊ฐ ์ „์†ก๋œ ๋ชจ๋“  ์š”์ฒญ์˜ ๋Œ€๊ธฐ ์‹œ๊ฐ„๊ณผ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์˜ ํ‰๊ท ์œผ๋กœ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. ๋ฌด์ž‘์œ„์„ฑ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ด ์„ธ ๊ฐ€์ง€ ์—์ด์ „ํŠธ๋ฅผ ์Šค์ผ€์ค„๋ง์„ ํ•˜๊ฑฐ๋‚˜ ํ•˜์ง€ ์•Š์€ ์ƒํƒœ๋กœ 5๋ฒˆ์˜ ๋ณ„๋„ ์‹คํ—˜์„ ํ†ตํ•ด ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 5์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ๋น„์Šค์ผ€์ค„ ๋ฐฉ์‹์€ ์ˆœ์„œ๊ฐ€ ๋น ๋ฅธ ์—์ด์ „ํŠธ์˜ ๊ฒฝ์šฐ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ด์ง€๋งŒ, ์ˆœ์„œ๊ฐ€ ๋Šฆ์€ ์—์ด์ „ํŠธ์˜ ๊ฒฝ์šฐ ๋Œ€๊ธฐ ์‹œ๊ฐ„๊ณผ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์ด ๊ธธ์–ด์ง€๋Š” ๋‹จ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ, AIOS์˜ ์Šค์ผ€์ค„๋ง ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ๋Œ€๊ธฐ ์‹œ๊ฐ„๊ณผ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์„ ํšจ์œจ์ ์œผ๋กœ ์กฐ์ ˆํ•˜๋ฉฐ, ์ด๋Š” ํŠนํžˆ LLM์ด ํฐ ๊ฒฝ์šฐ ์—์ด์ „ํŠธ๊ฐ€ ๋‚˜์ค‘์— ์ œ์ถœํ•˜๋Š” ์—์ด์ „ํŠธ ์š”์ฒญ์— ๋Œ€ํ•ด ๋”์šฑ ๋ถ„๋ช…ํ•˜๊ฒŒ ๋“œ๋Ÿฌ๋‚ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ๋ณ‘๋ ฌ ์ž‘์—…์„ ์ˆ˜์šฉํ•˜๊ธฐ ์œ„ํ•œ ์Šค์ผ€์ค„๋ง์˜ ์ค‘์š”์„ฑ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
์„ฑ๋Šฅ ๋ถ„์„.ํšจ์œจ์„ฑ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด FIFO ์Šค์ผ€์ค„๋ง์„ ์‚ฌ์šฉํ•˜๋Š” AIOS์™€ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์„ธ ๊ฐ€์ง€ ์—์ด์ „ํŠธ๊ฐ€ ๋™์‹œ์— ์‹คํ–‰๋˜๋Š” ๋น„์Šค์ผ€์ค„ ๋ฐฉ์‹ ๊ฐ„์˜ ๋น„๊ต ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋น„์Šค์ผ€์ค„ ๋ฐฉ์‹์—์„œ๋Š” ์ˆ˜ํ•™ ์—์ด์ „ํŠธ, ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ ์—์ด์ „ํŠธ, ๋ฆฌํฌ ์—์ด์ „ํŠธ์˜ ์‚ฌ์ „ ์ •์˜๋œ ์ˆœ์„œ์— ๋”ฐ๋ผ ์„ธ ๊ฐ€์ง€ ์—์ด์ „ํŠธ๊ฐ€ ์‹คํ–‰๋ฉ๋‹ˆ๋‹ค. ์‹œ๊ฐ„์  ํšจ์œจ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐ€์ง€ ๋ฉ”ํŠธ๋ฆญ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค: ๋Œ€๊ธฐ ์‹œ๊ฐ„(์—์ด์ „ํŠธ ์š”์ฒญ ์ œ์ถœ๋ถ€ํ„ฐ ์‹œ์ž‘๊นŒ์ง€์˜ ๊ฐ„๊ฒฉ)๊ณผ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„(์—์ด์ „ํŠธ ์š”์ฒญ ์ œ์ถœ๋ถ€ํ„ฐ ์™„๋ฃŒ๊นŒ์ง€์˜ ์‹œ๊ฐ„)์ž…๋‹ˆ๋‹ค. ๊ฐ ์—์ด์ „ํŠธ๋Š” LLM์— ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์š”์ฒญ์„ ๋ณด๋‚ด๋ฏ€๋กœ ๊ฐ ์—์ด์ „ํŠธ์˜ ๋Œ€๊ธฐ ์‹œ๊ฐ„๊ณผ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์€ ๊ฐ๊ฐ ์ „์†ก๋œ ๋ชจ๋“  ์š”์ฒญ์˜ ๋Œ€๊ธฐ ์‹œ๊ฐ„๊ณผ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์˜ ํ‰๊ท ์œผ๋กœ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. ๋ฌด์ž‘์œ„์„ฑ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ด ์„ธ ๊ฐ€์ง€ ์—์ด์ „ํŠธ๋ฅผ ์Šค์ผ€์ค„๋ง์„ ํ•˜๊ฑฐ๋‚˜ ํ•˜์ง€ ์•Š์€ ์ƒํƒœ๋กœ 5๋ฒˆ์˜ ๋ณ„๋„ ์‹คํ—˜์„ ํ†ตํ•ด ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 5์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ๋น„์Šค์ผ€์ค„ ๋ฐฉ์‹์€ ์ˆœ์„œ๊ฐ€ ๋น ๋ฅธ ์—์ด์ „ํŠธ์˜ ๊ฒฝ์šฐ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ด์ง€๋งŒ, ์ˆœ์„œ๊ฐ€ ๋Šฆ์€ ์—์ด์ „ํŠธ์˜ ๊ฒฝ์šฐ ๋Œ€๊ธฐ ์‹œ๊ฐ„๊ณผ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์ด ๊ธธ์–ด์ง€๋Š” ๋‹จ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ, AIOS์˜ ์Šค์ผ€์ค„๋ง ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ๋Œ€๊ธฐ ์‹œ๊ฐ„๊ณผ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์„ ํšจ์œจ์ ์œผ๋กœ ์กฐ์ ˆํ•˜๋ฉฐ, ์ด๋Š” ํŠนํžˆ LLM์ด ํฐ ๊ฒฝ์šฐ ์—์ด์ „ํŠธ๊ฐ€ ๋‚˜์ค‘์— ์ œ์ถœํ•˜๋Š” ์—์ด์ „ํŠธ ์š”์ฒญ์— ๋Œ€ํ•ด ๋”์šฑ ๋ถ„๋ช…ํ•˜๊ฒŒ ๋“œ๋Ÿฌ๋‚ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ๋ณ‘๋ ฌ ์ž‘์—…์„ ์ˆ˜์šฉํ•˜๊ธฐ ์œ„ํ•œ ์Šค์ผ€์ค„๋ง์˜ ์ค‘์š”์„ฑ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
**Code Synthesis.** In addition to generating high-quality natural language text, existing LLMs also show strong abilities to generate formal language, especially computer programs (_i.e._, code) that satisfy specific conditions, called _code synthesis_[635]. Unlike natural language generation, as the generated code can be directly checked by execution with corresponding compilers or interpreters, existing work mostly evaluates the quality of the generated code from LLMs by calculating the pass rate against the test cases, _i.e._, pass@_k42_. Recently, several code benchmarks focusing on functional correctness are proposed to assess the code synthesis abilities of LLMs, such as APPS [378], HumanEval [105], and MBPP [208]. Typically, they consist of diverse programming problems, with text specification and test cases for correctness checking. To improve such an ability, it is key to fine-tuning (or pre-training) LLMs on code data, which can effectively adapt LLMs to code synthesis tasks [86]. In addition, existing work has proposed new strategies to generate code, _e.g._, sampling multiple candidate solutions [208] and planning-guided decoding [636], which can be considered as the imitation of bug-fixing and code-planning processes by programmers. Impressively, LLMs have recently shown competitive performance with humans by achieving a ranking of the top 28% among users on the programming contest platform Codeforces [114].
**Code Synthesis**. ๊ธฐ์กด LLM์€ ๊ณ ํ’ˆ์งˆ์˜ ์ž์—ฐ์–ด ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ํŠน์ • ์กฐ๊ฑด์„ ๋งŒ์กฑํ•˜๋Š” ๊ณต์‹ ์–ธ์–ด, ํŠนํžˆ ์ปดํ“จํ„ฐ ํ”„๋กœ๊ทธ๋žจ(_์ฆ‰, ์ฝ”๋“œ)์„ ์ƒ์„ฑํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋Š”๋ฐ, ์ด๋ฅผ _code synthesis_๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค[635]. ์ƒ์„ฑ๋œ ์ฝ”๋“œ๋Š” ํ•ด๋‹น ์ปดํŒŒ์ผ๋Ÿฌ๋‚˜ ์ธํ„ฐํ”„๋ฆฌํ„ฐ๋ฅผ ํ†ตํ•ด ์ง์ ‘ ์‹คํ–‰ํ•˜์—ฌ ๊ฒ€์‚ฌํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์ž์—ฐ์–ด ์ƒ์„ฑ๊ณผ ๋‹ฌ๋ฆฌ ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์ฃผ๋กœ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค์— ๋Œ€ํ•œ ํ•ฉ๊ฒฉ๋ฅ , ์ฆ‰ pass@_k42_๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ LLM์—์„œ ์ƒ์„ฑ๋œ ์ฝ”๋“œ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” ๊ธฐ๋Šฅ์  ์ •ํ™•์„ฑ์— ์ดˆ์ ์„ ๋งž์ถ˜ ์—ฌ๋Ÿฌ ์ฝ”๋“œ ๋ฒค์น˜๋งˆํฌ(์˜ˆ: APPS [378], HumanEval [105], MBPP [208])๊ฐ€ ์ œ์•ˆ๋˜์–ด LLM์˜ code synthesis ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ •ํ™•์„ฑ ๊ฒ€์‚ฌ๋ฅผ ์œ„ํ•œ ํ…์ŠคํŠธ ์‚ฌ์–‘๊ณผ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๊ฐ€ ํฌํ•จ๋œ ๋‹ค์–‘ํ•œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋ฌธ์ œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด LLM์„ Fine-tuning(๋˜๋Š” ์‚ฌ์ „ ํ•™์Šต)ํ•˜๋Š” ๊ฒƒ์ด ํ•ต์‹ฌ์ด๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด LLM์„ code synthesis ์ž‘์—…์— ํšจ๊ณผ์ ์œผ๋กœ ์ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[86]. ๋˜ํ•œ ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์ฝ”๋“œ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ์ƒˆ๋กœ์šด ์ „๋žต, _์˜ˆ๋ฅผ ๋“ค์–ด ์—ฌ๋Ÿฌ ํ›„๋ณด ์†”๋ฃจ์…˜์„ ์ƒ˜ํ”Œ๋ง[208]ํ•˜๊ณ  ๊ณ„ํš ์•ˆ๋‚ด ๋””์ฝ”๋”ฉ[636]ํ•˜๋Š” ๋“ฑ ํ”„๋กœ๊ทธ๋ž˜๋จธ์˜ ๋ฒ„๊ทธ ์ˆ˜์ • ๋ฐ ์ฝ”๋“œ ๊ณ„ํš ํ”„๋กœ์„ธ์Šค๋ฅผ ๋ชจ๋ฐฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ๋Š” ์ƒˆ๋กœ์šด ์ „๋žต์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋†€๋ž๊ฒŒ๋„ LLM์€ ์ตœ๊ทผ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ฒฝ์ง„๋Œ€ํšŒ ํ”Œ๋žซํผ์ธ Codeforces์—์„œ ์‚ฌ์šฉ์ž ์ค‘ ์ƒ์œ„ 28%์˜ ์ˆœ์œ„๋ฅผ ๊ธฐ๋กํ•˜์—ฌ ์ธ๊ฐ„๊ณผ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค[114].
**์ฝ”๋“œ ํ•ฉ์„ฑ**. ๊ธฐ์กด LLM์€ ๊ณ ํ’ˆ์งˆ์˜ ์ž์—ฐ์–ด ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ํŠน์ • ์กฐ๊ฑด์„ ๋งŒ์กฑํ•˜๋Š” ๊ณต์‹ ์–ธ์–ด, ํŠนํžˆ ์ปดํ“จํ„ฐ ํ”„๋กœ๊ทธ๋žจ(_์ฆ‰, ์ฝ”๋“œ)์„ ์ƒ์„ฑํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋Š”๋ฐ, ์ด๋ฅผ _์ฝ”๋“œ ํ•ฉ์„ฑ_์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค[635]. ์ƒ์„ฑ๋œ ์ฝ”๋“œ๋Š” ํ•ด๋‹น ์ปดํŒŒ์ผ๋Ÿฌ๋‚˜ ์ธํ„ฐํ”„๋ฆฌํ„ฐ๋ฅผ ํ†ตํ•ด ์ง์ ‘ ์‹คํ–‰ํ•˜์—ฌ ๊ฒ€์‚ฌํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์ž์—ฐ์–ด ์ƒ์„ฑ๊ณผ ๋‹ฌ๋ฆฌ ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์ฃผ๋กœ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค์— ๋Œ€ํ•œ ํ•ฉ๊ฒฉ๋ฅ , ์ฆ‰ pass@_k42_๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ LLM์—์„œ ์ƒ์„ฑ๋œ ์ฝ”๋“œ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” ๊ธฐ๋Šฅ์  ์ •ํ™•์„ฑ์— ์ดˆ์ ์„ ๋งž์ถ˜ ์—ฌ๋Ÿฌ ์ฝ”๋“œ ๋ฒค์น˜๋งˆํฌ(์˜ˆ: APPS [378], HumanEval [105], MBPP [208])๊ฐ€ ์ œ์•ˆ๋˜์–ด LLM์˜ ์ฝ”๋“œ ํ•ฉ์„ฑ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ •ํ™•์„ฑ ๊ฒ€์‚ฌ๋ฅผ ์œ„ํ•œ ํ…์ŠคํŠธ ์‚ฌ์–‘๊ณผ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๊ฐ€ ํฌํ•จ๋œ ๋‹ค์–‘ํ•œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋ฌธ์ œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด LLM์„ ๋ฏธ์„ธ ์กฐ์ •(๋˜๋Š” ์‚ฌ์ „ ํ•™์Šต)ํ•˜๋Š” ๊ฒƒ์ด ํ•ต์‹ฌ์ด๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด LLM์„ ์ฝ”๋“œ ํ•ฉ์„ฑ ์ž‘์—…์— ํšจ๊ณผ์ ์œผ๋กœ ์ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[86]. ๋˜ํ•œ ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์ฝ”๋“œ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ์ƒˆ๋กœ์šด ์ „๋žต, _์˜ˆ๋ฅผ ๋“ค์–ด ์—ฌ๋Ÿฌ ํ›„๋ณด ์†”๋ฃจ์…˜์„ ์ƒ˜ํ”Œ๋ง[208]ํ•˜๊ณ  ๊ณ„ํš ์•ˆ๋‚ด ๋””์ฝ”๋”ฉ[636]ํ•˜๋Š” ๋“ฑ ํ”„๋กœ๊ทธ๋ž˜๋จธ์˜ ๋ฒ„๊ทธ ์ˆ˜์ • ๋ฐ ์ฝ”๋“œ ๊ณ„ํš ํ”„๋กœ์„ธ์Šค๋ฅผ ๋ชจ๋ฐฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ๋Š” ์ƒˆ๋กœ์šด ์ „๋žต์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋†€๋ž๊ฒŒ๋„ LLM์€ ์ตœ๊ทผ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ฒฝ์ง„๋Œ€ํšŒ ํ”Œ๋žซํผ์ธ Codeforces์—์„œ ์‚ฌ์šฉ์ž ์ค‘ ์ƒ์œ„ 28%์˜ ์ˆœ์œ„๋ฅผ ๊ธฐ๋กํ•˜์—ฌ ์ธ๊ฐ„๊ณผ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค[114].
Instruction tuning has emerged as the key in aligning large language models (LLMs) with specific task instructions, thereby mitigating the discrepancy between the next-token prediction objective and users' actual goals. To reduce the labor and time cost to collect or annotate data by humans, researchers start to explore the use of LLMs to generate instructionaligned synthetic data. Recent works focus on generating diverse instructions and applying LLM to increase instruction complexity, often neglecting downstream use cases. It remains unclear how to _tailor_ high-quality data to elicit better instruction-following abilities in different target instruction distributions and LLMs. To this end, we introduce **CodecLM**, a general framework for adaptively generating high-quality synthetic data for LLM alignment with different downstream instruction distributions and LLMs. Drawing on the Encode-Decode principles, we use LLMs as codecs to guide the data generation process. We first _encode_ seed instructions into metadata, which are concise keywords generated on-the-fly to capture the target instruction distribution, and then _decode_ metadata to create tailored instructions. We also introduce Self-Rubrics and Contrastive Filtering during decoding to tailor data-efficient samples. Extensive experiments on four open-domain instruction following benchmarks validate the effectiveness of CodecLM over the current state-of-the-arts.
Instruction Tuning์€ Large Language Models (LLMs)์„ ํŠน์ • ์ž‘์—… ์ง€์นจ์— ๋งž์ถ”๋Š” ๋ฐ ์žˆ์–ด ํ•ต์‹ฌ์ ์ธ ์—ญํ• ์„ ํ•˜๋ฉฐ, ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก ๋ชฉํ‘œ์™€ ์‚ฌ์šฉ์ž์˜ ์‹ค์ œ ๋ชฉํ‘œ ๊ฐ„์˜ ๋ถˆ์ผ์น˜๋ฅผ ์™„ํ™”ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ์ธ๊ฐ„์ด ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ฑฐ๋‚˜ ์ฃผ์„์„ ๋‹ฌ๊ธฐ ์œ„ํ•œ ๋…ธ๋™๋ ฅ๊ณผ ์‹œ๊ฐ„ ๋น„์šฉ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ง€์นจ์— ๋ถ€ํ•ฉํ•˜๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํƒ์ƒ‰ํ•˜๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์–‘ํ•œ Instructions์„ ์ƒ์„ฑํ•˜๊ณ  Instruction์˜ ๋ณต์žก์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด LLM์„ ์ ์šฉํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์œผ๋ฉฐ, ๋‹ค์šด์ŠคํŠธ๋ฆผ ์‚ฌ์šฉ ์‚ฌ๋ก€๋Š” ์ข…์ข… ์†Œํ™€ํžˆ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋Œ€์ƒ Instruction ๋ถ„ํฌ์™€ LLM์—์„œ ๋” ๋‚˜์€ Instruction ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๊ธฐ ์œ„ํ•ด ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ _๋งž์ถคํ™”_ํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์•„์ง ๋ช…ํ™•ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ์ €ํฌ๋Š” ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ Instruction ๋ถ„ํฌ์™€ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ LLM ์ •๋ ฌ์„ ์œ„ํ•œ ๊ณ ํ’ˆ์งˆ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ ์‘์ ์œผ๋กœ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ์ผ๋ฐ˜ ํ”„๋ ˆ์ž„์›Œํฌ์ธ **CodecLM**์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. Encode-Decode ์›์น™์„ ๋ฐ”ํƒ•์œผ๋กœ LLM์„ ์ฝ”๋ฑ์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์‹œ๋“œ ๋ช…๋ น์„ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋กœ _์ธ์ฝ”๋”ฉ_ํ•˜๊ณ , ์ด ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋Š” ๋Œ€์ƒ ๋ช…๋ น ๋ถ„ํฌ๋ฅผ ํฌ์ฐฉํ•˜๊ธฐ ์œ„ํ•ด ์ฆ‰์„์—์„œ ์ƒ์„ฑ๋œ ๊ฐ„๊ฒฐํ•œ ํ‚ค์›Œ๋“œ์ด๋ฉฐ, ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ _๋””์ฝ”๋”ฉ_ํ•˜์—ฌ ๋งž์ถคํ˜• ๋ช…๋ น์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋””์ฝ”๋”ฉ ์ค‘์— Self-Rubrics์™€ Contrastive Filtering์„ ๋„์ž…ํ•˜์—ฌ ๋ฐ์ดํ„ฐ ํšจ์œจ์ ์ธ ์ƒ˜ํ”Œ์„ ๋งž์ถคํ™”ํ•ฉ๋‹ˆ๋‹ค. 4๊ฐœ์˜ ์˜คํ”ˆ ๋„๋ฉ”์ธ Instruction ์ˆ˜ํ–‰ ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด ํ˜„์žฌ ์ตœ์‹  ๊ธฐ์ˆ ๋ณด๋‹ค CodecLM์˜ ํšจ๊ณผ๋ฅผ ๊ฒ€์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.
๊ต์œก ํŠœ๋‹์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์„ ํŠน์ • ์ž‘์—… ์ง€์นจ์— ๋งž์ถ”๋Š” ๋ฐ ์žˆ์–ด ํ•ต์‹ฌ์ ์ธ ์—ญํ• ์„ ํ•˜๋ฉฐ, ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก ๋ชฉํ‘œ์™€ ์‚ฌ์šฉ์ž์˜ ์‹ค์ œ ๋ชฉํ‘œ ๊ฐ„์˜ ๋ถˆ์ผ์น˜๋ฅผ ์™„ํ™”ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ์ธ๊ฐ„์ด ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ฑฐ๋‚˜ ์ฃผ์„์„ ๋‹ฌ๊ธฐ ์œ„ํ•œ ๋…ธ๋™๋ ฅ๊ณผ ์‹œ๊ฐ„ ๋น„์šฉ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ง€์นจ์— ๋ถ€ํ•ฉํ•˜๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํƒ์ƒ‰ํ•˜๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ์ง€์นจ์„ ์ƒ์„ฑํ•˜๊ณ  ์ง€์นจ์˜ ๋ณต์žก์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด LLM์„ ์ ์šฉํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์œผ๋ฉฐ, ๋‹ค์šด์ŠคํŠธ๋ฆผ ์‚ฌ์šฉ ์‚ฌ๋ก€๋Š” ์ข…์ข… ์†Œํ™€ํžˆ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋Œ€์ƒ ์ง€์นจ ๋ถ„ํฌ์™€ LLM์—์„œ ๋” ๋‚˜์€ ์ง€์นจ ์ˆ˜ํ–‰ ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๊ธฐ ์œ„ํ•ด ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ _๋งž์ถคํ™”_ํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์•„์ง ๋ช…ํ™•ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ์ €ํฌ๋Š” ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ง€์นจ ๋ถ„ํฌ์™€ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ LLM ์ •๋ ฌ์„ ์œ„ํ•œ ๊ณ ํ’ˆ์งˆ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ ์‘์ ์œผ๋กœ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ์ผ๋ฐ˜ ํ”„๋ ˆ์ž„์›Œํฌ์ธ **CodecLM**์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ธ์ฝ”๋”ฉ-๋””์ฝ”๋”ฉ ์›์น™์„ ๋ฐ”ํƒ•์œผ๋กœ LLM์„ ์ฝ”๋ฑ์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์‹œ๋“œ ๋ช…๋ น์„ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋กœ _์ธ์ฝ”๋”ฉ_ํ•˜๊ณ , ์ด ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋Š” ๋Œ€์ƒ ๋ช…๋ น ๋ถ„ํฌ๋ฅผ ํฌ์ฐฉํ•˜๊ธฐ ์œ„ํ•ด ์ฆ‰์„์—์„œ ์ƒ์„ฑ๋œ ๊ฐ„๊ฒฐํ•œ ํ‚ค์›Œ๋“œ์ด๋ฉฐ, ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ _๋””์ฝ”๋”ฉ_ํ•˜์—ฌ ๋งž์ถคํ˜• ๋ช…๋ น์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋””์ฝ”๋”ฉ ์ค‘์— ์…€ํ”„ ๋ฃจ๋ธŒ๋ฆญ๊ณผ ๋Œ€์กฐ ํ•„ํ„ฐ๋ง์„ ๋„์ž…ํ•˜์—ฌ ๋ฐ์ดํ„ฐ ํšจ์œจ์ ์ธ ์ƒ˜ํ”Œ์„ ๋งž์ถคํ™”ํ•ฉ๋‹ˆ๋‹ค. 4๊ฐœ์˜ ์˜คํ”ˆ ๋„๋ฉ”์ธ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด ํ˜„์žฌ ์ตœ์‹  ๊ธฐ์ˆ ๋ณด๋‹ค CodecLM์˜ ํšจ๊ณผ๋ฅผ ๊ฒ€์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Model training._ Different from LLMs, MLLMs are not trained from scratch, but instead developed based on pre-trained language and vision models. Existing work employs a typical two-stage approach for training MLLMs, _i.e.,_ vision-language alignment pre-training and visual instruction tuning. In essence, existing MLLMs aim to (1) preserve the inherent capabilities and parametric knowledge of LLMs as possible, and meanwhile (2) effectively adapt to multimodal tasks by leveraging the pre-trained LLMs and visual encoders. To achieve the above two goals, two typical training strategies are often employed for visual instruction tuning, either only optimizing the connection module [151] or fine-tuning both the connector module and LLM component [851]. As we can see, the former can reserve the original capacities of LLMs but likely have a weak an adaptation performance, while the latter can fully adapt to multimodal tasks but suffer from the loss of original capacities of LLMs. More efforts should be made to investigate how to effectively balance the two aspects, so as to achieving improved multimodal capacities. In addition, existing MLLMs are still overly dependent on the capacities of LLMs, which pose the limits on many multimodal tasks (_e.g.,_ space positioning). It will be meaningful to explore improved training approaches of language models, so that multimodal information can be also utilized in this process.
\(\bullet\)_๋ชจ๋ธ ํ•™์Šต._ MLLM์€ LLM๊ณผ ๋‹ฌ๋ฆฌ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ•™์Šตํ•˜์ง€ ์•Š๊ณ  ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ฐ ๋น„์ „ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” MLLM ํ•™์Šต์„ ์œ„ํ•ด ์ผ๋ฐ˜์ ์ธ 2๋‹จ๊ณ„ ์ ‘๊ทผ ๋ฐฉ์‹, ์ฆ‰ ๋น„์ „-์–ธ์–ด ์ •๋ ฌ ์‚ฌ์ „ ํ•™์Šต๊ณผ ์‹œ๊ฐ์  Instruction Tuning์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด MLLM์€ ๋ณธ์งˆ์ ์œผ๋กœ (1) LLM์˜ ๊ณ ์œ ํ•œ ๊ธฐ๋Šฅ๊ณผ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ์ง€์‹์„ ์ตœ๋Œ€ํ•œ ๋ณด์กดํ•˜๋Š” ๋™์‹œ์— (2) ์‚ฌ์ „ ํ•™์Šต๋œ LLM๊ณผ ์‹œ๊ฐ Encoder๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ž‘์—…์— ํšจ๊ณผ์ ์œผ๋กœ ์ ์‘ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์œ„์˜ ๋‘ ๊ฐ€์ง€ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์‹œ๊ฐ์  Instruction Tuning์—๋Š” ์—ฐ๊ฒฐ ๋ชจ๋“ˆ๋งŒ ์ตœ์ ํ™”ํ•˜๊ฑฐ๋‚˜ ์ปค๋„ฅํ„ฐ ๋ชจ๋“ˆ๊ณผ LLM ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ๋ชจ๋‘ Fine-tuningํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ ํ›ˆ๋ จ ์ „๋žต์ด ์ข…์ข… ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค[851]. ์•ž์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ์ „์ž๋Š” LLM์˜ ์›๋ž˜ ๊ธฐ๋Šฅ์„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ์ ์‘ ์„ฑ๋Šฅ์ด ์•ฝํ•  ์ˆ˜ ์žˆ๊ณ , ํ›„์ž๋Š” ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ž‘์—…์— ์™„์ „ํžˆ ์ ์‘ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ LLM์˜ ์›๋ž˜ ๊ธฐ๋Šฅ์ด ์†์‹ค๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋‘ ๊ฐ€์ง€ ์ธก๋ฉด์˜ ๊ท ํ˜•์„ ํšจ๊ณผ์ ์œผ๋กœ ๋งž์ถ”๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•˜๋Š” ๋…ธ๋ ฅ์ด ๋”์šฑ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด์˜ MLLM์€ ์—ฌ์ „ํžˆ LLM์˜ ์—ญ๋Ÿ‰์— ์ง€๋‚˜์น˜๊ฒŒ ์˜์กดํ•˜๊ณ  ์žˆ์–ด ๋งŽ์€ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ž‘์—…(_์˜ˆ:_ ๊ณต๊ฐ„ ์œ„์น˜ ์ง€์ •)์— ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์—์„œ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ •๋ณด๋„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฐœ์„ ๋œ ํ›ˆ๋ จ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์€ ์˜๋ฏธ๊ฐ€ ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
\(\bullet\)_๋ชจ๋ธ ํ•™์Šต._ MLLM์€ LLM๊ณผ ๋‹ฌ๋ฆฌ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ•™์Šตํ•˜์ง€ ์•Š๊ณ  ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ฐ ๋น„์ „ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” MLLM ํ•™์Šต์„ ์œ„ํ•ด ์ผ๋ฐ˜์ ์ธ 2๋‹จ๊ณ„ ์ ‘๊ทผ ๋ฐฉ์‹, ์ฆ‰ ๋น„์ „-์–ธ์–ด ์ •๋ ฌ ์‚ฌ์ „ ํ•™์Šต๊ณผ ์‹œ๊ฐ์  ์ง€์นจ ํŠœ๋‹์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด MLLM์€ ๋ณธ์งˆ์ ์œผ๋กœ (1) LLM์˜ ๊ณ ์œ ํ•œ ๊ธฐ๋Šฅ๊ณผ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ์ง€์‹์„ ์ตœ๋Œ€ํ•œ ๋ณด์กดํ•˜๋Š” ๋™์‹œ์— (2) ์‚ฌ์ „ ํ•™์Šต๋œ LLM๊ณผ ์‹œ๊ฐ ์ธ์ฝ”๋”๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ž‘์—…์— ํšจ๊ณผ์ ์œผ๋กœ ์ ์‘ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์œ„์˜ ๋‘ ๊ฐ€์ง€ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์‹œ๊ฐ์  ์ง€์‹œ ํŠœ๋‹์—๋Š” ์—ฐ๊ฒฐ ๋ชจ๋“ˆ๋งŒ ์ตœ์ ํ™”ํ•˜๊ฑฐ๋‚˜ ์ปค๋„ฅํ„ฐ ๋ชจ๋“ˆ๊ณผ LLM ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ๋ชจ๋‘ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ ํ›ˆ๋ จ ์ „๋žต์ด ์ข…์ข… ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค[851]. ์•ž์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ์ „์ž๋Š” LLM์˜ ์›๋ž˜ ๊ธฐ๋Šฅ์„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ์ ์‘ ์„ฑ๋Šฅ์ด ์•ฝํ•  ์ˆ˜ ์žˆ๊ณ , ํ›„์ž๋Š” ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ž‘์—…์— ์™„์ „ํžˆ ์ ์‘ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ LLM์˜ ์›๋ž˜ ๊ธฐ๋Šฅ์ด ์†์‹ค๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋‘ ๊ฐ€์ง€ ์ธก๋ฉด์˜ ๊ท ํ˜•์„ ํšจ๊ณผ์ ์œผ๋กœ ๋งž์ถ”๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•˜๋Š” ๋…ธ๋ ฅ์ด ๋”์šฑ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด์˜ MLLM์€ ์—ฌ์ „ํžˆ LLM์˜ ์—ญ๋Ÿ‰์— ์ง€๋‚˜์น˜๊ฒŒ ์˜์กดํ•˜๊ณ  ์žˆ์–ด ๋งŽ์€ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ž‘์—…(_์˜ˆ:_ ๊ณต๊ฐ„ ์œ„์น˜ ์ง€์ •)์— ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์—์„œ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ •๋ณด๋„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฐœ์„ ๋œ ํ›ˆ๋ จ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์€ ์˜๋ฏธ๊ฐ€ ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Large Language Models (LMs) are known to encode world knowledge in their parameters as they pretrain on a vast amount of web corpus, which is often utilized for performing knowledge-dependent downstream tasks such as question answering, fact-checking, and open dialogue. In real-world scenarios, the world knowledge stored in the LMs can quickly become outdated as the world changes, but it is non-trivial to avoid catastrophic forgetting and reliably acquire new knowledge while preserving invariant knowledge. To push the community towards better maintenance of ever-changing LMs, we formulate a new continual learning (CL) problem called Continual Knowledge Learning (CKL). We construct a new benchmark and metric to quantify the retention of time-invariant world knowledge, the update of outdated knowledge, and the acquisition of new knowledge. We adopt applicable recent methods from literature to create several strong baselines. Through extensive experiments, we find that CKL exhibits unique challenges that are not addressed in previous CL setups, where parameter expansion is necessary to reliably retain and learn knowledge simultaneously. By highlighting the critical causes of knowledge forgetting, we show that CKL is a challenging and important problem that helps us better understand and train ever-changing LMs. The benchmark datasets, model checkpoints, and code to reproduce our results are available at this https URL.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LMs)์€ ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์›น ์ฝ”ํผ์Šค๋ฅผ ์‚ฌ์ „ ํ•™์Šตํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋งค๊ฐœ๋ณ€์ˆ˜์— ์„ธ๊ณ„ ์ง€์‹์„ ์ธ์ฝ”๋”ฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์•Œ๋ ค์ ธ ์žˆ์œผ๋ฉฐ, ์ด๋Š” question answering, fact-checking, open dialogue ๋“ฑ ์ง€์‹์— ์˜์กดํ•˜๋Š” ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐ ์ž์ฃผ ํ™œ์šฉ๋ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” ์„ธ๊ณ„๊ฐ€ ๋ณ€ํ™”ํ•จ์— ๋”ฐ๋ผ LMs์— ์ €์žฅ๋œ ์„ธ๊ณ„ ์ง€์‹์ด ๋น ๋ฅด๊ฒŒ ๋‚™ํ›„๋  ์ˆ˜ ์žˆ์ง€๋งŒ, catastrophic forgetting์„ ํ”ผํ•˜๊ณ  ๋ถˆ๋ณ€ ์ง€์‹์„ ๋ณด์กดํ•˜๋ฉด์„œ ์ƒˆ๋กœ์šด ์ง€์‹์„ ์•ˆ์ •์ ์œผ๋กœ ์Šต๋“ํ•˜๋Š” ๊ฒƒ์€ ๊ฐ„๋‹จํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด ์ปค๋ฎค๋‹ˆํ‹ฐ๊ฐ€ ๋Š์ž„์—†์ด ๋ณ€ํ™”ํ•˜๋Š” LMs์„ ๋” ์ž˜ ๊ด€๋ฆฌํ•  ์ˆ˜ ์žˆ๋„๋ก, ์šฐ๋ฆฌ๋Š” Continual Knowledge Learning (CKL)์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ์ง€์†์  ํ•™์Šต(CL) ๋ฌธ์ œ๋ฅผ ๊ณต์‹ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์‹œ๊ฐ„์— ๋ณ€ํ•˜์ง€ ์•Š๋Š” ์„ธ๊ณ„ ์ง€์‹์˜ ์œ ์ง€, ์˜ค๋ž˜๋œ ์ง€์‹์˜ ์—…๋ฐ์ดํŠธ, ์ƒˆ๋กœ์šด ์ง€์‹์˜ ์Šต๋“์„ ์ •๋Ÿ‰ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด ๋ฒค์น˜๋งˆํฌ์™€ ๋ฉ”ํŠธ๋ฆญ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ฌธํ—Œ์—์„œ ์ ์šฉ ๊ฐ€๋Šฅํ•œ ์ตœ๊ทผ์˜ ๋ฐฉ๋ฒ•์„ ์ฑ„ํƒํ•˜์—ฌ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ฐ•๋ ฅํ•œ ๊ธฐ์ค€์„ ์„ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด CKL์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ํ™•์žฅ์ด ํ•„์š”ํ•œ ์ด์ „ CL ์„ค์ •์—์„œ๋Š” ๋‹ค๋ฃจ์ง€ ๋ชปํ–ˆ๋˜ ๊ณ ์œ ํ•œ ๊ณผ์ œ๋ฅผ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ง€์‹์„ ๋™์‹œ์— ์•ˆ์ •์ ์œผ๋กœ ์œ ์ง€ํ•˜๊ณ  ํ•™์Šตํ•˜๋Š” ๋ฐ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ง€์‹ ๋ง๊ฐ์˜ ์ค‘์š”ํ•œ ์›์ธ์„ ๊ฐ•์กฐํ•จ์œผ๋กœ์จ CKL์€ ๋Š์ž„์—†์ด ๋ณ€ํ™”ํ•˜๋Š” LMs์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ณ  ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ๋„์ „์ ์ด๊ณ  ์ค‘์š”ํ•œ ๋ฌธ์ œ์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ, ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ, ๊ฒฐ๊ณผ ์žฌํ˜„์„ ์œ„ํ•œ ์ฝ”๋“œ๋Š” ์ด https URL์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LM)์€ ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์›น ์ฝ”ํผ์Šค๋ฅผ ์‚ฌ์ „ ํ•™์Šตํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋งค๊ฐœ๋ณ€์ˆ˜์— ์„ธ๊ณ„ ์ง€์‹์„ ์ธ์ฝ”๋”ฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์•Œ๋ ค์ ธ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์งˆ๋ฌธ ๋‹ต๋ณ€, ์‚ฌ์‹ค ํ™•์ธ, ์˜คํ”ˆ ๋Œ€ํ™” ๋“ฑ ์ง€์‹์— ์˜์กดํ•˜๋Š” ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐ ์ž์ฃผ ํ™œ์šฉ๋ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” ์„ธ๊ณ„๊ฐ€ ๋ณ€ํ™”ํ•จ์— ๋”ฐ๋ผ LM์— ์ €์žฅ๋œ ์„ธ๊ณ„ ์ง€์‹์ด ๋น ๋ฅด๊ฒŒ ๋‚™ํ›„๋  ์ˆ˜ ์žˆ์ง€๋งŒ, ํŒŒ๊ตญ์  ๋ง๊ฐ์„ ํ”ผํ•˜๊ณ  ๋ถˆ๋ณ€ ์ง€์‹์„ ๋ณด์กดํ•˜๋ฉด์„œ ์ƒˆ๋กœ์šด ์ง€์‹์„ ์•ˆ์ •์ ์œผ๋กœ ์Šต๋“ํ•˜๋Š” ๊ฒƒ์€ ๊ฐ„๋‹จํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด ์ปค๋ฎค๋‹ˆํ‹ฐ๊ฐ€ ๋Š์ž„์—†์ด ๋ณ€ํ™”ํ•˜๋Š” LM์„ ๋” ์ž˜ ๊ด€๋ฆฌํ•  ์ˆ˜ ์žˆ๋„๋ก, ์šฐ๋ฆฌ๋Š” ์ง€์†์  ์ง€์‹ ํ•™์Šต(CKL)์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ์ง€์†์  ํ•™์Šต(CL) ๋ฌธ์ œ๋ฅผ ๊ณต์‹ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์‹œ๊ฐ„์— ๋ณ€ํ•˜์ง€ ์•Š๋Š” ์„ธ๊ณ„ ์ง€์‹์˜ ์œ ์ง€, ์˜ค๋ž˜๋œ ์ง€์‹์˜ ์—…๋ฐ์ดํŠธ, ์ƒˆ๋กœ์šด ์ง€์‹์˜ ์Šต๋“์„ ์ •๋Ÿ‰ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด ๋ฒค์น˜๋งˆํฌ์™€ ๋ฉ”ํŠธ๋ฆญ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ฌธํ—Œ์—์„œ ์ ์šฉ ๊ฐ€๋Šฅํ•œ ์ตœ๊ทผ์˜ ๋ฐฉ๋ฒ•์„ ์ฑ„ํƒํ•˜์—ฌ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ฐ•๋ ฅํ•œ ๊ธฐ์ค€์„ ์„ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด CKL์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ํ™•์žฅ์ด ํ•„์š”ํ•œ ์ด์ „ CL ์„ค์ •์—์„œ๋Š” ๋‹ค๋ฃจ์ง€ ๋ชปํ–ˆ๋˜ ๊ณ ์œ ํ•œ ๊ณผ์ œ๋ฅผ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ง€์‹์„ ๋™์‹œ์— ์•ˆ์ •์ ์œผ๋กœ ์œ ์ง€ํ•˜๊ณ  ํ•™์Šตํ•˜๋Š” ๋ฐ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ง€์‹ ๋ง๊ฐ์˜ ์ค‘์š”ํ•œ ์›์ธ์„ ๊ฐ•์กฐํ•จ์œผ๋กœ์จ CKL์€ ๋Š์ž„์—†์ด ๋ณ€ํ™”ํ•˜๋Š” LM์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ณ  ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ๋„์ „์ ์ด๊ณ  ์ค‘์š”ํ•œ ๋ฌธ์ œ์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ, ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ, ๊ฒฐ๊ณผ ์žฌํ˜„์„ ์œ„ํ•œ ์ฝ”๋“œ๋Š” ์ด https URL์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Primary training objective._ Since the alignment data typically consists of an input instruction and an output response, the primary training loss is still the traditional cross-entropy loss for sequence-to-sequence learning. Based on this loss, many studies propose a number of improvement variants for enhancing the supervised alignment tuning. For example, CoH [388] constructs the training data by prepending _"A helpful answer:"_ and _"An unhelpful answer:"_ to the annotated good and bad responses, respectively, and only compute losses for those response tokens with special masking. Quark [384] sorts model responses into different quantiles with varying alignment quality, it prepends a special reward token to each model response to represent the reward level of the response. Further, to enable the preference modeling via the maximum likelihood objective, DPO [389] first reparameterizes the response rewards using the policy model (_i.e._, the language model being optimized), and then the original reward modelling objective can be reformulated only based on the policy model. In this way, DPO removes the explicit reward modeling step, and optimizing the new learning objective only involving the policy model is equivalent to optimizing the rewards. Furthermore, FIGA [386] designs a fine-grained contrastive loss that aims to encourage desirable tokens, penalize undesirable ones, and disregard trivial tokens.
\(\bullet\)_Primary training objective._ ์ •๋ ฌ ๋ฐ์ดํ„ฐ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ž…๋ ฅ instruction๊ณผ ์ถœ๋ ฅ ์‘๋‹ต์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฏ€๋กœ, Primary training loss๋Š” ์—ฌ์ „ํžˆ ์‹œํ€€์Šค ๋Œ€ ์‹œํ€€์Šค ํ•™์Šต์„ ์œ„ํ•œ ๊ธฐ์กด์˜ cross-entropy loss์ž…๋‹ˆ๋‹ค. ์ด loss๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋งŽ์€ ์—ฐ๊ตฌ์—์„œ supervised alignment tuning์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ๊ฐœ์„  ๋ณ€ํ˜•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, CoH [388]๋Š” ์ฃผ์„ ์ฒ˜๋ฆฌ๋œ ์ข‹์€ ์‘๋‹ต๊ณผ ๋‚˜์œ ์‘๋‹ต์— ๊ฐ๊ฐ _"๋„์›€์ด ๋˜๋Š” ๋‹ต๋ณ€:"_ ๋ฐ _"๋„์›€์ด ๋˜์ง€ ์•Š๋Š” ๋‹ต๋ณ€:"_๋ฅผ ์•ž์— ๋ถ™์—ฌ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•˜๊ณ , ํŠน์ˆ˜ ๋งˆ์Šคํ‚น์ด ์žˆ๋Š” ์‘๋‹ต ํ† ํฐ์— ๋Œ€ํ•ด์„œ๋งŒ ์†์‹ค์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. Quark [384]๋Š” ๋ชจ๋ธ ์‘๋‹ต์„ ์ •๋ ฌ ํ’ˆ์งˆ์ด ๋‹ค๋ฅธ ๋‹ค์–‘ํ•œ ๋ถ„์œ„์ˆ˜๋กœ ์ •๋ ฌํ•˜๊ณ , ๊ฐ ๋ชจ๋ธ ์‘๋‹ต์— ํŠน์ˆ˜ ๋ณด์ƒ ํ† ํฐ์„ ์•ž์— ๋ถ™์—ฌ ์‘๋‹ต์˜ ๋ณด์ƒ ์ˆ˜์ค€์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ตœ๋Œ€ ๊ฐ€๋Šฅ์„ฑ objective๋ฅผ ํ†ตํ•ด ์„ ํ˜ธ๋„ ๋ชจ๋ธ๋ง์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด DPO [389]๋Š” ๋จผ์ € policy model(_์ฆ‰, ์ตœ์ ํ™” ์ค‘์ธ language model)์„ ์‚ฌ์šฉํ•˜์—ฌ ์‘๋‹ต ๋ณด์ƒ์„ ์žฌ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”ํ•œ ๋‹ค์Œ, policy model๋งŒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์›๋ž˜์˜ reward modeling objective๋ฅผ ์žฌ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ DPO๋Š” ๋ช…์‹œ์ ์ธ reward modeling ๋‹จ๊ณ„๋ฅผ ์ œ๊ฑฐํ•˜๊ณ  policy model๋งŒ ํฌํ•จํ•˜๋Š” ์ƒˆ๋กœ์šด ํ•™์Šต objective๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๊ฒƒ์€ ๋ณด์ƒ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๊ฒƒ๊ณผ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, FIGA [386]๋Š” ๋ฐ”๋žŒ์งํ•œ ํ† ํฐ์„ ์žฅ๋ คํ•˜๊ณ  ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์€ ํ† ํฐ์€ ๋ถˆ์ด์ต์„ ์ฃผ๋ฉฐ ์‚ฌ์†Œํ•œ ํ† ํฐ์€ ๋ฌด์‹œํ•˜๋Š” ์„ธ๋ถ„ํ™”๋œ contrastive loss๋ฅผ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_1์ฐจ ํ›ˆ๋ จ ๋ชฉํ‘œ._ ์ •๋ ฌ ๋ฐ์ดํ„ฐ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ž…๋ ฅ ๋ช…๋ น์–ด์™€ ์ถœ๋ ฅ ์‘๋‹ต์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฏ€๋กœ, 1์ฐจ ํ›ˆ๋ จ ์†์‹ค์€ ์—ฌ์ „ํžˆ ์‹œํ€€์Šค ๋Œ€ ์‹œํ€€์Šค ํ•™์Šต์„ ์œ„ํ•œ ๊ธฐ์กด์˜ ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ ์†์‹ค์ž…๋‹ˆ๋‹ค. ์ด ์†์‹ค์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋งŽ์€ ์—ฐ๊ตฌ์—์„œ ์ง€๋„ ์ •๋ ฌ ํŠœ๋‹์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ๊ฐœ์„  ๋ณ€ํ˜•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, CoH [388]๋Š” ์ฃผ์„ ์ฒ˜๋ฆฌ๋œ ์ข‹์€ ์‘๋‹ต๊ณผ ๋‚˜์œ ์‘๋‹ต์— ๊ฐ๊ฐ _"๋„์›€์ด ๋˜๋Š” ๋‹ต๋ณ€:"_ ๋ฐ _"๋„์›€์ด ๋˜์ง€ ์•Š๋Š” ๋‹ต๋ณ€:"_๋ฅผ ์•ž์— ๋ถ™์—ฌ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•˜๊ณ , ํŠน์ˆ˜ ๋งˆ์Šคํ‚น์ด ์žˆ๋Š” ์‘๋‹ต ํ† ํฐ์— ๋Œ€ํ•ด์„œ๋งŒ ์†์‹ค์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. Quark [384]๋Š” ๋ชจ๋ธ ์‘๋‹ต์„ ์ •๋ ฌ ํ’ˆ์งˆ์ด ๋‹ค๋ฅธ ๋‹ค์–‘ํ•œ ๋ถ„์œ„์ˆ˜๋กœ ์ •๋ ฌํ•˜๊ณ , ๊ฐ ๋ชจ๋ธ ์‘๋‹ต์— ํŠน์ˆ˜ ๋ณด์ƒ ํ† ํฐ์„ ์•ž์— ๋ถ™์—ฌ ์‘๋‹ต์˜ ๋ณด์ƒ ์ˆ˜์ค€์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ตœ๋Œ€ ๊ฐ€๋Šฅ์„ฑ ๋ชฉํ‘œ๋ฅผ ํ†ตํ•ด ์„ ํ˜ธ๋„ ๋ชจ๋ธ๋ง์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด DPO [389]๋Š” ๋จผ์ € ์ •์ฑ… ๋ชจ๋ธ(_์ฆ‰, ์ตœ์ ํ™” ์ค‘์ธ ์–ธ์–ด ๋ชจ๋ธ)์„ ์‚ฌ์šฉํ•˜์—ฌ ์‘๋‹ต ๋ณด์ƒ์„ ์žฌ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”ํ•œ ๋‹ค์Œ, ์ •์ฑ… ๋ชจ๋ธ๋งŒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์›๋ž˜์˜ ๋ณด์ƒ ๋ชจ๋ธ๋ง ๋ชฉํ‘œ๋ฅผ ์žฌ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ DPO๋Š” ๋ช…์‹œ์ ์ธ ๋ณด์ƒ ๋ชจ๋ธ๋ง ๋‹จ๊ณ„๋ฅผ ์ œ๊ฑฐํ•˜๊ณ  ์ •์ฑ… ๋ชจ๋ธ๋งŒ ํฌํ•จํ•˜๋Š” ์ƒˆ๋กœ์šด ํ•™์Šต ๋ชฉํ‘œ๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๊ฒƒ์€ ๋ณด์ƒ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๊ฒƒ๊ณผ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, FIGA [386]๋Š” ๋ฐ”๋žŒ์งํ•œ ํ† ํฐ์„ ์žฅ๋ คํ•˜๊ณ  ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์€ ํ† ํฐ์€ ๋ถˆ์ด์ต์„ ์ฃผ๋ฉฐ ์‚ฌ์†Œํ•œ ํ† ํฐ์€ ๋ฌด์‹œํ•˜๋Š” ์„ธ๋ถ„ํ™”๋œ ๋Œ€์กฐ ์†์‹ค์„ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค.
In reasoning tasks, even a minor error can cascade into inaccurate results, leading to suboptimal performance of large language models in such domains. Earlier fine-tuning approaches sought to mitigate this by leveraging more precise supervisory signals from human labeling, larger models, or self-sampling, although at a high cost. Conversely, we develop a method that avoids external resources, relying instead on introducing perturbations to the input. Our training approach randomly masks certain tokens within the chain of thought, a technique we found to be particularly effective for reasoning tasks. When applied to fine-tuning with GSM8K, this method achieved a 5% improvement in accuracy over standard supervised fine-tuning with a few codes modified and no additional labeling effort. Furthermore, it is complementary to existing methods. When integrated with related data augmentation methods, it leads to an average improvement of 3% improvement in GSM8K accuracy and 1% improvement in MATH accuracy across five datasets of various quality and size, as well as two base models. We further investigate the mechanisms behind this improvement through case studies and quantitative analysis, suggesting that our approach may provide superior support for the model in capturing long-distance dependencies, especially those related to questions. This enhancement could deepen understanding of premises in questions and prior steps.
์ถ”๋ก  ๊ณผ์ œ์—์„œ๋Š” ์‚ฌ์†Œํ•œ ์˜ค๋ฅ˜๋„ ๋ถ€์ •ํ™•ํ•œ ๊ฒฐ๊ณผ๋กœ ์ด์–ด์ ธ ํ•ด๋‹น ์˜์—ญ์—์„œ Large Language Models์˜ ์„ฑ๋Šฅ์ด ์ตœ์ ์ด ๋˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์ „์˜ Fine-tuning ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ธ๊ฐ„ ๋ผ๋ฒจ๋ง์˜ ๋ณด๋‹ค ์ •ํ™•ํ•œ ๊ฐ๋… ์‹ ํ˜ธ, ๋” ํฐ ๋ชจ๋ธ ๋˜๋Š” ์ž์ฒด ์ƒ˜ํ”Œ๋ง์„ ํ™œ์šฉํ•˜์—ฌ ์ด๋ฅผ ์™„ํ™”ํ•˜๊ณ ์ž ํ–ˆ์ง€๋งŒ, ๊ทธ ๋Œ€๊ฐ€๊ฐ€ ์ปธ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ ์ €ํฌ๋Š” ์™ธ๋ถ€ ๋ฆฌ์†Œ์Šค๋ฅผ ํ”ผํ•˜๊ณ  ๋Œ€์‹  ์ž…๋ ฅ์— ๊ต๋ž€์„ ๋„์ž…ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์ถ”๋ก  ์ž‘์—…์— ํŠนํžˆ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ๋ฐํ˜€์ง„ ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ๋‚ด์—์„œ ํŠน์ • ํ† ํฐ์„ ๋ฌด์ž‘์œ„๋กœ ๋งˆ์Šคํ‚นํ•˜๋Š” ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ GSM8K๋กœ Fine-tuningํ•˜๋Š” ๋ฐ ์ ์šฉํ•˜๋ฉด ๋ช‡ ๊ฐ€์ง€ ์ฝ”๋“œ๋ฅผ ์ˆ˜์ •ํ•˜๊ณ  ์ถ”๊ฐ€ ๋ผ๋ฒจ๋ง ์ž‘์—… ์—†์ด๋„ ํ‘œ์ค€ Supervised Fine-tuning๋ณด๋‹ค ์ •ํ™•๋„๊ฐ€ 5% ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด ๋ฐฉ๋ฒ•๊ณผ ๋ณด์™„์ ์ž…๋‹ˆ๋‹ค. ๊ด€๋ จ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐฉ๋ฒ•๊ณผ ํ†ตํ•ฉํ•˜๋ฉด ๋‹ค์–‘ํ•œ ํ’ˆ์งˆ๊ณผ ํฌ๊ธฐ์˜ 5๊ฐœ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ 2๊ฐœ์˜ ๊ธฐ๋ณธ ๋ชจ๋ธ์—์„œ ํ‰๊ท ์ ์œผ๋กœ GSM8K ์ •ํ™•๋„๋Š” 3% ํ–ฅ์ƒ, MATH ์ •ํ™•๋„๋Š” 1% ํ–ฅ์ƒ๋˜๋Š” ํšจ๊ณผ๋ฅผ ๊ฐ€์ ธ์™”์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์‚ฌ๋ก€ ์—ฐ๊ตฌ์™€ ์ •๋Ÿ‰์  ๋ถ„์„์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ถ”๊ฐ€๋กœ ์กฐ์‚ฌํ•˜์—ฌ, ์ด ์ ‘๊ทผ ๋ฐฉ์‹์ด ํŠนํžˆ ์งˆ๋ฌธ๊ณผ ๊ด€๋ จ๋œ ์žฅ๊ฑฐ๋ฆฌ ์ข…์†์„ฑ, ํŠนํžˆ ์งˆ๋ฌธ๊ณผ ๊ด€๋ จ๋œ ์žฅ๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์„ ํฌ์ฐฉํ•˜๋Š” ๋ฐ ์žˆ์–ด ๋ชจ๋ธ์„ ๋” ์šฐ์ˆ˜ํ•˜๊ฒŒ ์ง€์›ํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์€ ์งˆ๋ฌธ์˜ ์ „์ œ์™€ ์ด์ „ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ๊นŠ๊ฒŒ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ถ”๋ก  ๊ณผ์ œ์—์„œ๋Š” ์‚ฌ์†Œํ•œ ์˜ค๋ฅ˜๋„ ๋ถ€์ •ํ™•ํ•œ ๊ฒฐ๊ณผ๋กœ ์ด์–ด์ ธ ํ•ด๋‹น ์˜์—ญ์—์„œ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ์ตœ์ ์ด ๋˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์ „์˜ ๋ฏธ์„ธ ์กฐ์ • ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ธ๊ฐ„ ๋ผ๋ฒจ๋ง์˜ ๋ณด๋‹ค ์ •ํ™•ํ•œ ๊ฐ๋… ์‹ ํ˜ธ, ๋” ํฐ ๋ชจ๋ธ ๋˜๋Š” ์ž์ฒด ์ƒ˜ํ”Œ๋ง์„ ํ™œ์šฉํ•˜์—ฌ ์ด๋ฅผ ์™„ํ™”ํ•˜๊ณ ์ž ํ–ˆ์ง€๋งŒ, ๊ทธ ๋Œ€๊ฐ€๊ฐ€ ์ปธ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ ์ €ํฌ๋Š” ์™ธ๋ถ€ ๋ฆฌ์†Œ์Šค๋ฅผ ํ”ผํ•˜๊ณ  ๋Œ€์‹  ์ž…๋ ฅ์— ๊ต๋ž€์„ ๋„์ž…ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์ถ”๋ก  ์ž‘์—…์— ํŠนํžˆ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ๋ฐํ˜€์ง„ ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ๋‚ด์—์„œ ํŠน์ • ํ† ํฐ์„ ๋ฌด์ž‘์œ„๋กœ ๋งˆ์Šคํ‚นํ•˜๋Š” ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ GSM8K๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋ฐ ์ ์šฉํ•˜๋ฉด ๋ช‡ ๊ฐ€์ง€ ์ฝ”๋“œ๋ฅผ ์ˆ˜์ •ํ•˜๊ณ  ์ถ”๊ฐ€ ๋ผ๋ฒจ๋ง ์ž‘์—… ์—†์ด๋„ ํ‘œ์ค€ ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •๋ณด๋‹ค ์ •ํ™•๋„๊ฐ€ 5% ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด ๋ฐฉ๋ฒ•๊ณผ ๋ณด์™„์ ์ž…๋‹ˆ๋‹ค. ๊ด€๋ จ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐฉ๋ฒ•๊ณผ ํ†ตํ•ฉํ•˜๋ฉด ๋‹ค์–‘ํ•œ ํ’ˆ์งˆ๊ณผ ํฌ๊ธฐ์˜ 5๊ฐœ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ 2๊ฐœ์˜ ๊ธฐ๋ณธ ๋ชจ๋ธ์—์„œ ํ‰๊ท ์ ์œผ๋กœ GSM8K ์ •ํ™•๋„๋Š” 3% ํ–ฅ์ƒ, MATH ์ •ํ™•๋„๋Š” 1% ํ–ฅ์ƒ๋˜๋Š” ํšจ๊ณผ๋ฅผ ๊ฐ€์ ธ์™”์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์‚ฌ๋ก€ ์—ฐ๊ตฌ์™€ ์ •๋Ÿ‰์  ๋ถ„์„์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ถ”๊ฐ€๋กœ ์กฐ์‚ฌํ•˜์—ฌ, ์ด ์ ‘๊ทผ ๋ฐฉ์‹์ด ํŠนํžˆ ์งˆ๋ฌธ๊ณผ ๊ด€๋ จ๋œ ์žฅ๊ฑฐ๋ฆฌ ์ข…์†์„ฑ, ํŠนํžˆ ์งˆ๋ฌธ๊ณผ ๊ด€๋ จ๋œ ์žฅ๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์„ ํฌ์ฐฉํ•˜๋Š” ๋ฐ ์žˆ์–ด ๋ชจ๋ธ์„ ๋” ์šฐ์ˆ˜ํ•˜๊ฒŒ ์ง€์›ํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์€ ์งˆ๋ฌธ์˜ ์ „์ œ์™€ ์ด์ „ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ๊นŠ๊ฒŒ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Task Generalization.** Instruction tuning encourages the model to understand natural language instructions for task completion. It endows LLMs with the ability (often considered as an emergent ability) to follow human instructions [31] to perform specific tasks without demonstrations, even on unseen tasks [69]. A large number of studies have confirmed the effectiveness of instruction tuning to achieve superior performance on both seen and unseen tasks [95, 345]. Also, instruction tuning has been shown to be useful in alleviating several weaknesses of LLMs (_e.g._, repetitive generation or complementing the input without accomplishing a certain task) [66, 69], leading to a superior capacity to solve real-world tasks for LLMs. Furthermore, LLMs trained with instruction tuning can generalize to related tasks across languages. For example, BLOOMZ-P3 [94] is fine-tuned based on BLOOM [78] using English-only task collection P3 [167]. Interestingly, BLOOMZ-P3 can achieve a more than 50% improvement in multilingual sentence completion tasks compared to BLOOM, which shows that instruction tuning can help LLMs acquire general task skills from English-only datasets and transfer such skills into other languages [94]. In addition, it has been found that using English-only instructions can produce satisfactory results on multilingual tasks [94], which helps reduce the effort of instruction engineering for a specific language.
**์ž‘์—… ์ผ๋ฐ˜ํ™”**. Instruction Tuning์€ ๋ชจ๋ธ์ด ์ž‘์—… ์™„๋ฃŒ๋ฅผ ์œ„ํ•œ ์ž์—ฐ์–ด ๋ช…๋ น์–ด๋ฅผ ์ดํ•ดํ•˜๋„๋ก ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด LLM์€ ์‹œ์—ฐ ์—†์ด๋„ ๋ณด์ด์ง€ ์•Š๋Š” ์ž‘์—…์—์„œ๋„ ํŠน์ • ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด [31] ์ธ๊ฐ„์˜ ์ง€์‹œ๋ฅผ ๋”ฐ๋ฅด๋Š” ๋Šฅ๋ ฅ(์ข…์ข… ์ถœํ˜„ํ•˜๋Š” ๋Šฅ๋ ฅ์œผ๋กœ ๊ฐ„์ฃผ๋จ)์„ ๊ฐ–์ถ”๊ฒŒ ๋ฉ๋‹ˆ๋‹ค[69]. ๋งŽ์€ ์—ฐ๊ตฌ์—์„œ Instruction Tuning์ด ๋ณด์ด๋Š” ์ž‘์—…๊ณผ ๋ณด์ด์ง€ ์•Š๋Š” ์ž‘์—… ๋ชจ๋‘์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์ด ํ™•์ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[95, 345]. ๋˜ํ•œ Instruction Tuning์€ LLM์˜ ๋ช‡ ๊ฐ€์ง€ ์•ฝ์ (_์˜ˆ: ๋ฐ˜๋ณต์ ์ธ ์ƒ์„ฑ ๋˜๋Š” ํŠน์ • ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š๊ณ  ์ž…๋ ฅ์„ ๋ณด์™„ํ•˜๋Š” ๊ฒƒ)์„ ์™„ํ™”ํ•˜๋Š” ๋ฐ ์œ ์šฉํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[66, 69], ์ด๋ฅผ ํ†ตํ•ด LLM์˜ ์‹ค์ œ ์ž‘์—… ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ Instruction Tuning์œผ๋กœ ํ›ˆ๋ จ๋œ LLM์€ ์–ธ์–ด ๊ฐ„ ๊ด€๋ จ ์ž‘์—…์œผ๋กœ ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, BLOOMZ-P3 [94]๋Š” ์˜์–ด ์ „์šฉ ์ž‘์—… ์ˆ˜์ง‘ P3 [167]์„ ์‚ฌ์šฉํ•˜์—ฌ BLOOM [78]์„ ๊ธฐ๋ฐ˜์œผ๋กœ Fine-tuningํ•ฉ๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ BLOOMZ-P3๋Š” BLOOM์— ๋น„ํ•ด ๋‹ค๊ตญ์–ด ๋ฌธ์žฅ ์™„์„ฑ ์ž‘์—…์—์„œ 50% ์ด์ƒ์˜ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” Instruction Tuning์ด LLM์ด ์˜์–ด ์ „์šฉ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ผ๋ฐ˜์ ์ธ ์ž‘์—… ๊ธฐ์ˆ ์„ ์Šต๋“ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ ์„ ๋‹ค๋ฅธ ์–ธ์–ด๋กœ ์ „์†กํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค[94]. ๋˜ํ•œ ์˜์–ด ์ „์šฉ ๋ช…๋ น์–ด๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋‹ค๊ตญ์–ด ์ž‘์—…์—์„œ๋„ ๋งŒ์กฑ์Šค๋Ÿฌ์šด ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ[94], ์ด๋Š” ํŠน์ • ์–ธ์–ด์— ๋Œ€ํ•œ Instruction ์—”์ง€๋‹ˆ์–ด๋ง์˜ ๋…ธ๋ ฅ์„ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
**์ž‘์—… ์ผ๋ฐ˜ํ™”**. ๋ช…๋ น์–ด ํŠœ๋‹์€ ๋ชจ๋ธ์ด ์ž‘์—… ์™„๋ฃŒ๋ฅผ ์œ„ํ•œ ์ž์—ฐ์–ด ๋ช…๋ น์–ด๋ฅผ ์ดํ•ดํ•˜๋„๋ก ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด LLM์€ ์‹œ์—ฐ ์—†์ด๋„ ๋ณด์ด์ง€ ์•Š๋Š” ์ž‘์—…์—์„œ๋„ ํŠน์ • ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด [31] ์ธ๊ฐ„์˜ ์ง€์‹œ๋ฅผ ๋”ฐ๋ฅด๋Š” ๋Šฅ๋ ฅ(์ข…์ข… ์ถœํ˜„ํ•˜๋Š” ๋Šฅ๋ ฅ์œผ๋กœ ๊ฐ„์ฃผ๋จ)์„ ๊ฐ–์ถ”๊ฒŒ ๋ฉ๋‹ˆ๋‹ค[69]. ๋งŽ์€ ์—ฐ๊ตฌ์—์„œ ๋ช…๋ น์–ด ํŠœ๋‹์ด ๋ณด์ด๋Š” ์ž‘์—…๊ณผ ๋ณด์ด์ง€ ์•Š๋Š” ์ž‘์—… ๋ชจ๋‘์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์ด ํ™•์ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[95, 345]. ๋˜ํ•œ ๋ช…๋ น์–ด ํŠœ๋‹์€ LLM์˜ ๋ช‡ ๊ฐ€์ง€ ์•ฝ์ (_์˜ˆ: ๋ฐ˜๋ณต์ ์ธ ์ƒ์„ฑ ๋˜๋Š” ํŠน์ • ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š๊ณ  ์ž…๋ ฅ์„ ๋ณด์™„ํ•˜๋Š” ๊ฒƒ)์„ ์™„ํ™”ํ•˜๋Š” ๋ฐ ์œ ์šฉํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[66, 69], ์ด๋ฅผ ํ†ตํ•ด LLM์˜ ์‹ค์ œ ์ž‘์—… ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ช…๋ น์–ด ํŠœ๋‹์œผ๋กœ ํ›ˆ๋ จ๋œ LLM์€ ์–ธ์–ด ๊ฐ„ ๊ด€๋ จ ์ž‘์—…์œผ๋กœ ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, BLOOMZ-P3 [94]๋Š” ์˜์–ด ์ „์šฉ ์ž‘์—… ์ˆ˜์ง‘ P3 [167]์„ ์‚ฌ์šฉํ•˜์—ฌ BLOOM [78]์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ BLOOMZ-P3๋Š” BLOOM์— ๋น„ํ•ด ๋‹ค๊ตญ์–ด ๋ฌธ์žฅ ์™„์„ฑ ์ž‘์—…์—์„œ 50% ์ด์ƒ์˜ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋ช…๋ น์–ด ํŠœ๋‹์ด LLM์ด ์˜์–ด ์ „์šฉ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ผ๋ฐ˜์ ์ธ ์ž‘์—… ๊ธฐ์ˆ ์„ ์Šต๋“ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ ์„ ๋‹ค๋ฅธ ์–ธ์–ด๋กœ ์ „์†กํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค[94]. ๋˜ํ•œ ์˜์–ด ์ „์šฉ ๋ช…๋ น์–ด๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋‹ค๊ตญ์–ด ์ž‘์—…์—์„œ๋„ ๋งŒ์กฑ์Šค๋Ÿฌ์šด ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ[94], ์ด๋Š” ํŠน์ • ์–ธ์–ด์— ๋Œ€ํ•œ ๋ช…๋ น์–ด ์—”์ง€๋‹ˆ์–ด๋ง์˜ ๋…ธ๋ ฅ์„ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
Towards high-fidelity and more efficient scalable oversight.As AI models become increasingly complex and autonomous, it becomes challenging to monitor and assess their behavior using traditional oversight methods that rely on human supervision or real-world data (Amodei et al., 2016). Future research should explore the use of synthetic data for high-fidelity scalable oversight of these advanced systems. Existing methods typically simulate a certain scenario in social iterations, such as debate (Leike et al., 2018), reflection (Zhang et al., 2023), or revisions (Liu et al., 2023) to obtain synthetic data, while new approaches could cover more comprehensive scenarios and more modalities (Sun et al., 2023), as recent studies have found many issues of simulation that only coversa narrowed down (Cheng et al., 2023) or over-simplified (Zhou et al., 2024) scenes. Looking forward, another growing direction could be how to achieve scalable oversight more efficiently--given that we have the full control over the synthetic data generation, we can probably provide more targeted oversights with less synthetic data. As the need for effective AI governance and regulation grows, synthetic data will play an increasingly vital role in enabling more trustworthy scalable oversight mechanisms that promote robust, accountable, and safe deployment of AI technologies for the benefit of society (Askell et al., 2021; Bowman et al., 2022).
๊ณ ์ถฉ์‹ค๋„์˜ ๋ณด๋‹ค ํšจ์œจ์ ์ธ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ oversight๋ฅผ ํ–ฅํ•˜์—ฌ. AI ๋ชจ๋ธ์ด ์ ์  ๋” ๋ณต์žกํ•˜๊ณ  ์ž์œจ์ ์œผ๋กœ ๋ณ€ํ™”ํ•จ์— ๋”ฐ๋ผ ์ธ๊ฐ„์˜ ๊ฐ๋…์ด๋‚˜ ์‹ค์ œ ๋ฐ์ดํ„ฐ์— ์˜์กดํ•˜๋Š” ๊ธฐ์กด์˜ oversight ๋ฐฉ๋ฒ•์œผ๋กœ๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ํ–‰๋™์„ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๊ณ  ํ‰๊ฐ€ํ•˜๊ธฐ๊ฐ€ ์ ์  ๋” ์–ด๋ ค์›Œ์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Amodei et al., 2016). ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๊ณ ๊ธ‰ ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ๊ณ ์ถฉ์‹ค๋„ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ oversight๋ฅผ ์œ„ํ•ด synthetic data์˜ ์‚ฌ์šฉ์„ ํƒ๊ตฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ† ๋ก (Leike et al., 2018), ์„ฑ์ฐฐ(Zhang et al., 2023) ๋˜๋Š” ์ˆ˜์ •(Liu et al., 2023)๊ณผ ๊ฐ™์€ ์‚ฌํšŒ์  ๋ฐ˜๋ณต์—์„œ ํŠน์ • ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜์—ฌ synthetic data๋ฅผ ์–ป์ง€๋งŒ, ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฒ”์œ„๊ฐ€ ์ข์•„์ง„(Cheng et al., 2023) ๋˜๋Š” ์ง€๋‚˜์น˜๊ฒŒ ๋‹จ์ˆœํ™”๋œ(Zhou et al., 2024) ์žฅ๋ฉด๋งŒ์„ ๋‹ค๋ฃจ๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜์˜ ๋งŽ์€ ๋ฌธ์ œ๋ฅผ ๋ฐœ๊ฒฌํ•จ์— ๋”ฐ๋ผ ์ƒˆ๋กœ์šด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ์‹œ๋‚˜๋ฆฌ์˜ค์™€ ๋” ๋งŽ์€ modality(Sun et al., 2023)๋ฅผ ๋‹ค๋ฃฐ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์•ž์œผ๋กœ ๋‚˜์•„๊ฐˆ ๋ฐฉํ–ฅ์€ synthetic data ์ƒ์„ฑ์„ ์™„์ „ํžˆ ์ œ์–ดํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— synthetic data๋ฅผ ๋” ์ ๊ฒŒ ์‚ฌ์šฉํ•˜์—ฌ ๋” ํƒ€๊ฒŸํŒ…๋œ oversight๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์—์„œ ๋ณด๋‹ค ํšจ์œจ์ ์œผ๋กœ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ oversight๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํšจ๊ณผ์ ์ธ AI ๊ฑฐ๋ฒ„๋„Œ์Šค ๋ฐ ๊ทœ์ œ์— ๋Œ€ํ•œ ํ•„์š”์„ฑ์ด ์ปค์ง์— ๋”ฐ๋ผ synthetic data๋Š” ์‚ฌํšŒ๋ฅผ ์œ„ํ•ด ๊ฐ•๋ ฅํ•˜๊ณ  ์ฑ…์ž„๊ฐ ์žˆ๊ณ  ์•ˆ์ „ํ•œ AI ๊ธฐ์ˆ  ๋ฐฐํฌ๋ฅผ ์ด‰์ง„ํ•˜๋Š” ๋ณด๋‹ค ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ํ™•์žฅ ๊ฐ€๋Šฅํ•œ oversight ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐ ์ ์  ๋” ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค(Askell et al., 2021; Bowman et al., 2022).
๊ณ ์ถฉ์‹ค๋„์˜ ๋ณด๋‹ค ํšจ์œจ์ ์ธ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๊ฐ๋…์„ ํ–ฅํ•˜์—ฌ.AI ๋ชจ๋ธ์ด ์ ์  ๋” ๋ณต์žกํ•˜๊ณ  ์ž์œจ์ ์œผ๋กœ ๋ณ€ํ™”ํ•จ์— ๋”ฐ๋ผ ์ธ๊ฐ„์˜ ๊ฐ๋…์ด๋‚˜ ์‹ค์ œ ๋ฐ์ดํ„ฐ์— ์˜์กดํ•˜๋Š” ๊ธฐ์กด์˜ ๊ฐ๋… ๋ฐฉ๋ฒ•์œผ๋กœ๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ํ–‰๋™์„ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๊ณ  ํ‰๊ฐ€ํ•˜๊ธฐ๊ฐ€ ์ ์  ๋” ์–ด๋ ค์›Œ์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Amodei et al., 2016). ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๊ณ ๊ธ‰ ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ๊ณ ์ถฉ์‹ค๋„ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๊ฐ๋…์„ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์‚ฌ์šฉ์„ ํƒ๊ตฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ† ๋ก (Leike et al., 2018), ์„ฑ์ฐฐ(Zhang et al., 2023) ๋˜๋Š” ์ˆ˜์ •(Liu et al., 2023)๊ณผ ๊ฐ™์€ ์‚ฌํšŒ์  ๋ฐ˜๋ณต์—์„œ ํŠน์ • ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜์—ฌ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์–ป์ง€๋งŒ, ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฒ”์œ„๊ฐ€ ์ข์•„์ง„(Cheng et al., 2023) ๋˜๋Š” ์ง€๋‚˜์น˜๊ฒŒ ๋‹จ์ˆœํ™”๋œ(Zhou et al., 2024) ์žฅ๋ฉด๋งŒ์„ ๋‹ค๋ฃจ๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜์˜ ๋งŽ์€ ๋ฌธ์ œ๋ฅผ ๋ฐœ๊ฒฌํ•จ์— ๋”ฐ๋ผ ์ƒˆ๋กœ์šด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ์‹œ๋‚˜๋ฆฌ์˜ค์™€ ๋” ๋งŽ์€ ๋ฐฉ์‹(Sun et al., 2023)์„ ๋‹ค๋ฃฐ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์•ž์œผ๋กœ ๋‚˜์•„๊ฐˆ ๋ฐฉํ–ฅ์€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์™„์ „ํžˆ ์ œ์–ดํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ๋” ์ ๊ฒŒ ์‚ฌ์šฉํ•˜์—ฌ ๋” ํƒ€๊ฒŸํŒ…๋œ ๊ฐ๋…์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์—์„œ ๋ณด๋‹ค ํšจ์œจ์ ์œผ๋กœ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๊ฐ๋…์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํšจ๊ณผ์ ์ธ AI ๊ฑฐ๋ฒ„๋„Œ์Šค ๋ฐ ๊ทœ์ œ์— ๋Œ€ํ•œ ํ•„์š”์„ฑ์ด ์ปค์ง์— ๋”ฐ๋ผ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์‚ฌํšŒ๋ฅผ ์œ„ํ•ด ๊ฐ•๋ ฅํ•˜๊ณ  ์ฑ…์ž„๊ฐ ์žˆ๊ณ  ์•ˆ์ „ํ•œ AI ๊ธฐ์ˆ  ๋ฐฐํฌ๋ฅผ ์ด‰์ง„ํ•˜๋Š” ๋ณด๋‹ค ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๊ฐ๋… ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐ ์ ์  ๋” ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค(Askell et al., 2021; Bowman et al., 2022).
**Formatting Synthetic Data.** To reduce the burden of human annotation or manual collection, several semi-automated approaches [143] have been proposed for constructing instances by feeding existing instances into LLMs to synthesize diverse task descriptions and instances. As illustrated in Figure 11(c), the Self-Instruct method only needs 175 instances as the initial task pool. Then, they randomly select a few instances from the pool as demonstrations and prompt a LLM to generate new instructions and corresponding input-output pairs. After the quality and diversity filtering, newly generated instances would be added into the task pool. Hence, the synthetic method is an effective and economical way to generate large-scale instruction data for LLMs. However, the instances generated by the Self-Instruct method might be simplistic or lack the diversity. To improve the quality of synthetic int ructions, WizardLM [346] introduces Evol-Instruct by proposing in-depth and in-breadth evolving to enrich the complexity and diversity of the instances. Furthermore, Self-Align [347] establishes multiple human-aligned principles to filter the synthesized instances. It then employs these instances to train a LLM in order to yield more aligned instances. To enhance the quality of the instance output, researchers directly adopt human-written texts as the output and synthesize corresponding instructions using ICL examples [348].
**Formatting Synthetic Data.** ์ธ์  ์ฃผ์„์ด๋‚˜ ์ˆ˜๋™ ์ˆ˜์ง‘์˜ ๋ถ€๋‹ด์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์ธ์Šคํ„ด์Šค๋ฅผ LLM์— ์ž…๋ ฅํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—… ์„ค๋ช…๊ณผ ์ธ์Šคํ„ด์Šค๋ฅผ ํ•ฉ์„ฑํ•จ์œผ๋กœ์จ ์ธ์Šคํ„ด์Šค๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฐ˜ ์ž๋™ํ™”๋œ ์ ‘๊ทผ ๋ฐฉ์‹ [143]์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 11(c)์— ์˜ˆ์‹œ๋œ ๋ฐ”์™€ ๊ฐ™์ด, Self-Instruct ๋ฐฉ๋ฒ•์€ ์ดˆ๊ธฐ ์ž‘์—… ํ’€๋กœ 175๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋งŒ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ’€์—์„œ ๋ช‡ ๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ•˜์—ฌ ๋ฐ๋ชจ๋กœ ์‚ฌ์šฉํ•˜๊ณ  LLM์— ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ์ƒˆ๋กœ์šด instructions์™€ ํ•ด๋‹น ์ž…๋ ฅ-์ถœ๋ ฅ ์Œ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํ’ˆ์งˆ ๋ฐ ๋‹ค์–‘์„ฑ ํ•„ํ„ฐ๋ง ํ›„ ์ƒˆ๋กœ ์ƒ์„ฑ๋œ ์ธ์Šคํ„ด์Šค๋Š” ์ž‘์—… ํ’€์— ์ถ”๊ฐ€๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ•ฉ์„ฑ ๋ฐฉ๋ฒ•์€ LLM์„ ์œ„ํ•œ ๋Œ€๊ทœ๋ชจ instruction ๋ฐ์ดํ„ฐ๋ฅผ ํšจ๊ณผ์ ์ด๊ณ  ๊ฒฝ์ œ์ ์œผ๋กœ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Self-Instruct ๋ฐฉ๋ฒ•์œผ๋กœ ์ƒ์„ฑ๋œ ์ธ์Šคํ„ด์Šค๋Š” ๋‹จ์ˆœํ•˜๊ฑฐ๋‚˜ ๋‹ค์–‘์„ฑ์ด ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ๋œ instructions์˜ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด WizardLM [346]์€ ์ธ์Šคํ„ด์Šค์˜ ๋ณต์žก์„ฑ๊ณผ ๋‹ค์–‘์„ฑ์„ ํ’๋ถ€ํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด ์‹ฌ์ธต์ ์ด๊ณ  ํญ๋„“๊ฒŒ ์ง„ํ™”ํ•˜๋Š” Evol-Instruct๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Self-Align [347]์€ ํ•ฉ์„ฑ๋œ ์ธ์Šคํ„ด์Šค๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์ธ๊ฐ„ ์ •๋ ฌ ์›์น™์„ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ์ธ์Šคํ„ด์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ ํ›ˆ๋ จํ•˜์—ฌ ๋” ์ž˜ ์ •๋ ฌ๋œ ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ธ์Šคํ„ด์Šค ์ถœ๋ ฅ์˜ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ ํ…์ŠคํŠธ๋ฅผ ์ถœ๋ ฅ์œผ๋กœ ์ง์ ‘ ์ฑ„ํƒํ•˜๊ณ  In-Context Learning (ICL) ์˜ˆ์ œ [348]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•ด๋‹น instructions๋ฅผ ํ•ฉ์„ฑํ•ฉ๋‹ˆ๋‹ค.
**ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ํฌ๋งทํŒ…**. ์ธ์  ์ฃผ์„์ด๋‚˜ ์ˆ˜๋™ ์ˆ˜์ง‘์˜ ๋ถ€๋‹ด์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์ธ์Šคํ„ด์Šค๋ฅผ LLM์— ์ž…๋ ฅํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—… ์„ค๋ช…๊ณผ ์ธ์Šคํ„ด์Šค๋ฅผ ํ•ฉ์„ฑํ•จ์œผ๋กœ์จ ์ธ์Šคํ„ด์Šค๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฐ˜ ์ž๋™ํ™”๋œ ์ ‘๊ทผ ๋ฐฉ์‹ [143]์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 11(c)์— ์˜ˆ์‹œ๋œ ๋ฐ”์™€ ๊ฐ™์ด, Self-Instruct ๋ฐฉ๋ฒ•์€ ์ดˆ๊ธฐ ์ž‘์—… ํ’€๋กœ 175๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋งŒ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ’€์—์„œ ๋ช‡ ๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ•˜์—ฌ ๋ฐ๋ชจ๋กœ ์‚ฌ์šฉํ•˜๊ณ  LLM์— ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ์ƒˆ๋กœ์šด ์ง€์นจ๊ณผ ํ•ด๋‹น ์ž…๋ ฅ-์ถœ๋ ฅ ์Œ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํ’ˆ์งˆ ๋ฐ ๋‹ค์–‘์„ฑ ํ•„ํ„ฐ๋ง ํ›„ ์ƒˆ๋กœ ์ƒ์„ฑ๋œ ์ธ์Šคํ„ด์Šค๋Š” ์ž‘์—… ํ’€์— ์ถ”๊ฐ€๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ•ฉ์„ฑ ๋ฐฉ๋ฒ•์€ LLM์„ ์œ„ํ•œ ๋Œ€๊ทœ๋ชจ ๋ช…๋ น ๋ฐ์ดํ„ฐ๋ฅผ ํšจ๊ณผ์ ์ด๊ณ  ๊ฒฝ์ œ์ ์œผ๋กœ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Self-Instruct ๋ฐฉ๋ฒ•์œผ๋กœ ์ƒ์„ฑ๋œ ์ธ์Šคํ„ด์Šค๋Š” ๋‹จ์ˆœํ•˜๊ฑฐ๋‚˜ ๋‹ค์–‘์„ฑ์ด ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ๋œ ๋ช…๋ น์–ด์˜ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด WizardLM [346]์€ ์ธ์Šคํ„ด์Šค์˜ ๋ณต์žก์„ฑ๊ณผ ๋‹ค์–‘์„ฑ์„ ํ’๋ถ€ํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด ์‹ฌ์ธต์ ์ด๊ณ  ํญ๋„“๊ฒŒ ์ง„ํ™”ํ•˜๋Š” Evol-Instruct๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Self-Align [347]์€ ํ•ฉ์„ฑ๋œ ์ธ์Šคํ„ด์Šค๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์ธ๊ฐ„ ์ •๋ ฌ ์›์น™์„ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ์ธ์Šคํ„ด์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ ํ›ˆ๋ จํ•˜์—ฌ ๋” ์ž˜ ์ •๋ ฌ๋œ ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ธ์Šคํ„ด์Šค ์ถœ๋ ฅ์˜ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ ํ…์ŠคํŠธ๋ฅผ ์ถœ๋ ฅ์œผ๋กœ ์ง์ ‘ ์ฑ„ํƒํ•˜๊ณ  ICL ์˜ˆ์ œ [348]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•ด๋‹น ์ง€์นจ์„ ํ•ฉ์„ฑํ•ฉ๋‹ˆ๋‹ค.
is also essential for specific tasks. For example, retrieved documents are highly useful for open-domain question answering as supporting evidence. Both the quality of the retrieved documents and their relevance to the question have an impact on the generated answers [461]. Thus, it needs to include such information in a proper prompt pattern or expression format. Furthermore, in-context task exemplers are also helpful for eliciting LLMs to accomplish a complex task, which can better depict the task goal, the special output formats, and the mapping relation between input and output. * _Prompt style._ For different LLMs, it is important to design a suitable prompt style for eliciting their abilities to solve specific tasks. Overall, one should express the prompt as a clear question or detailed instruction that can be well understood and answered. In some cases, it is also useful to add the prefix or suffix to better guide LLMs. For example, using the prefix _"Let us think step by step"_ can help elicit LLMs perform step-by-step reasoning, and using the prefix _"You are an expert on this task (or in this domain)"_ can boost the performance of LLMs in some specific tasks. Further, for chat-based LLMs (_e.g._, ChatGPT), instead of directly feeding a long or complex task prompt, it is suggested to decompose it into multiple prompts for the sub-tasks and then feed them into LLMs via a multi-turn conversation [448].
๋˜ํ•œ ํŠน์ • ์ž‘์—…์—๋„ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ๋Š” ๋‹ต๋ณ€์˜ ๊ทผ๊ฑฐ ์ž๋ฃŒ๋กœ์„œ ์˜คํ”ˆ ๋„๋ฉ”์ธ question answering์— ๋งค์šฐ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ์˜ ํ’ˆ์งˆ๊ณผ ์งˆ๋ฌธ๊ณผ์˜ ๊ด€๋ จ์„ฑ ๋ชจ๋‘ ์ƒ์„ฑ๋œ ๋‹ต๋ณ€์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค[461]. ๋”ฐ๋ผ์„œ ์ ์ ˆํ•œ prompt ํŒจํ„ด์ด๋‚˜ ํ‘œํ˜„ ํ˜•์‹์— ์ด๋Ÿฌํ•œ ์ •๋ณด๋ฅผ ํฌํ•จ์‹œ์ผœ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, in-context task ์˜ˆ์‹œ๋Š” ๋ณต์žกํ•œ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก LLM์„ ์ด๋Œ์–ด๋‚ด๋Š” ๋ฐ์—๋„ ๋„์›€์ด ๋˜๋ฉฐ, ์ž‘์—… ๋ชฉํ‘œ, ํŠน์ˆ˜ ์ถœ๋ ฅ ํ˜•์‹, ์ž…๋ ฅ๊ณผ ์ถœ๋ ฅ ๊ฐ„์˜ ๋งคํ•‘ ๊ด€๊ณ„๋ฅผ ๋” ์ž˜ ๋ฌ˜์‚ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * _Prompt style._ LLM๋งˆ๋‹ค ํŠน์ • ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๊ธฐ ์œ„ํ•ด ์ ํ•ฉํ•œ prompt ์Šคํƒ€์ผ์„ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ prompt๋Š” ๋ช…ํ™•ํ•œ ์งˆ๋ฌธ์ด๋‚˜ ์ž˜ ์ดํ•ด๋˜๊ณ  ๋‹ต๋ณ€ํ•  ์ˆ˜ ์žˆ๋Š” ์ƒ์„ธํ•œ instruction์œผ๋กœ ํ‘œํ˜„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฝ์šฐ์— ๋”ฐ๋ผ์„œ๋Š” ์ ‘๋‘์‚ฌ๋‚˜ ์ ‘๋ฏธ์‚ฌ๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ LLM์„ ๋” ์ž˜ ์•ˆ๋‚ดํ•˜๋Š” ๊ฒƒ๋„ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ ‘๋‘์‚ฌ _"์ฐจ๊ทผ์ฐจ๊ทผ ์ƒ๊ฐํ•ด ๋ด…์‹œ๋‹ค"_๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด LLM์ด ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์œ ๋„ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ ‘๋‘์‚ฌ _"๋‹น์‹ ์€ ์ด ์ž‘์—…(๋˜๋Š” ์ด ๋„๋ฉ”์ธ)์˜ ์ „๋ฌธ๊ฐ€์ž…๋‹ˆ๋‹ค."_๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ผ๋ถ€ ํŠน์ • ์ž‘์—…์—์„œ LLM์˜ ์„ฑ๋Šฅ์„ ๋†’์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ฑ„ํŒ… ๊ธฐ๋ฐ˜ LLM(_์˜ˆ: ChatGPT)์˜ ๊ฒฝ์šฐ, ๊ธธ๊ฑฐ๋‚˜ ๋ณต์žกํ•œ ์ž‘์—… prompt๋ฅผ ์ง์ ‘ ์ž…๋ ฅํ•˜๋Š” ๋Œ€์‹  ํ•˜์œ„ ์ž‘์—…์— ๋Œ€ํ•œ ์—ฌ๋Ÿฌ prompt๋กœ ๋ถ„ํ•ดํ•œ ๋‹ค์Œ ๋ฉ€ํ‹ฐ ํ„ด ๋Œ€ํ™”๋ฅผ ํ†ตํ•ด LLM์— ์ž…๋ ฅํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค[448].
๋˜ํ•œ ํŠน์ • ์ž‘์—…์—๋„ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ๋Š” ๋‹ต๋ณ€์˜ ๊ทผ๊ฑฐ ์ž๋ฃŒ๋กœ์„œ ์˜คํ”ˆ ๋„๋ฉ”์ธ ์งˆ๋ฌธ์— ๋งค์šฐ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ์˜ ํ’ˆ์งˆ๊ณผ ์งˆ๋ฌธ๊ณผ์˜ ๊ด€๋ จ์„ฑ ๋ชจ๋‘ ์ƒ์„ฑ๋œ ๋‹ต๋ณ€์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค[461]. ๋”ฐ๋ผ์„œ ์ ์ ˆํ•œ ํ”„๋กฌํ”„ํŠธ ํŒจํ„ด์ด๋‚˜ ํ‘œํ˜„ ํ˜•์‹์— ์ด๋Ÿฌํ•œ ์ •๋ณด๋ฅผ ํฌํ•จ์‹œ์ผœ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ƒํ™ฉ์— ๋งž๋Š” ์ž‘์—… ์˜ˆ์‹œ๋Š” ๋ณต์žกํ•œ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก LLM์„ ์ด๋Œ์–ด๋‚ด๋Š” ๋ฐ์—๋„ ๋„์›€์ด ๋˜๋ฉฐ, ์ž‘์—… ๋ชฉํ‘œ, ํŠน์ˆ˜ ์ถœ๋ ฅ ํ˜•์‹, ์ž…๋ ฅ๊ณผ ์ถœ๋ ฅ ๊ฐ„์˜ ๋งคํ•‘ ๊ด€๊ณ„๋ฅผ ๋” ์ž˜ ๋ฌ˜์‚ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * ํ”„๋กฌํ”„ํŠธ ์Šคํƒ€์ผ._ LLM๋งˆ๋‹ค ํŠน์ • ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๊ธฐ ์œ„ํ•ด ์ ํ•ฉํ•œ ํ”„๋กฌํ”„ํŠธ ์Šคํƒ€์ผ์„ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ ํ”„๋กฌํ”„ํŠธ๋Š” ๋ช…ํ™•ํ•œ ์งˆ๋ฌธ์ด๋‚˜ ์ž˜ ์ดํ•ด๋˜๊ณ  ๋‹ต๋ณ€ํ•  ์ˆ˜ ์žˆ๋Š” ์ƒ์„ธํ•œ ์ง€์นจ์œผ๋กœ ํ‘œํ˜„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฝ์šฐ์— ๋”ฐ๋ผ์„œ๋Š” ์ ‘๋‘์‚ฌ๋‚˜ ์ ‘๋ฏธ์‚ฌ๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ LLM์„ ๋” ์ž˜ ์•ˆ๋‚ดํ•˜๋Š” ๊ฒƒ๋„ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ ‘๋‘์‚ฌ _"์ฐจ๊ทผ์ฐจ๊ทผ ์ƒ๊ฐํ•ด ๋ด…์‹œ๋‹ค"_๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด LLM์ด ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์œ ๋„ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ ‘๋‘์‚ฌ _"๋‹น์‹ ์€ ์ด ์ž‘์—…(๋˜๋Š” ์ด ๋„๋ฉ”์ธ)์˜ ์ „๋ฌธ๊ฐ€์ž…๋‹ˆ๋‹ค."_๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ผ๋ถ€ ํŠน์ • ์ž‘์—…์—์„œ LLM์˜ ์„ฑ๋Šฅ์„ ๋†’์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ฑ„ํŒ… ๊ธฐ๋ฐ˜ LLM(_์˜ˆ: ChatGPT)์˜ ๊ฒฝ์šฐ, ๊ธธ๊ฑฐ๋‚˜ ๋ณต์žกํ•œ ์ž‘์—… ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ง์ ‘ ์ž…๋ ฅํ•˜๋Š” ๋Œ€์‹  ํ•˜์œ„ ์ž‘์—…์— ๋Œ€ํ•œ ์—ฌ๋Ÿฌ ํ”„๋กฌํ”„ํŠธ๋กœ ๋ถ„ํ•ดํ•œ ๋‹ค์Œ ๋ฉ€ํ‹ฐ ํ„ด ๋Œ€ํ™”๋ฅผ ํ†ตํ•ด LLM์— ์ž…๋ ฅํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค[448].
**Balancing the Data Distribution.** Since instruction tuning involves a mixture of different tasks, it is important to balance the proportion of different tasks during fine-tuning. A widely used method is the _examples-proportional mixing_ strategy [82], _i.e.,_ combining all the datasets and sampling each instance equally from the mixed datasets. Furthermore, increasing the sampling ratio of high-quality collections (_e.g.,_ FLAN [67] and P3 [167]) can generally lead to performance improvement according to recent findings [95, 69]. Further, it is common to set a _maximum cap_ to control the maximum number of examples that a dataset can contain during instruction tuning [82], which is set to prevent larger datasets from overwhelming the entire distribution [82, 95]. In practice, the maximum cap is typically set to several thousands or tens of thousands according to different datasets [67, 69]. Recently, it has been empirically found that existing instruction datasets (Table III) mainly focus on enhancing LLMs' capabilities in certain aspects, and a single dataset alone cannot lead to a comprehensive enhancement in model capacity [353]. Therefore, it is often suggested to use a mixture of existing instruction datasets to achieve a balanced improvement in different capacities, including NLP task data (_e.g.,_ FLAN v2 [292]), chat data (_e.g.,_ ShareGPT [148]), and synthetic data (_e.g.,_ GPT4-Alpaca [354]).
**๋ฐ์ดํ„ฐ ๋ถ„ํฌ์˜ ๊ท ํ˜• ์œ ์ง€**. Instruction Tuning์—๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—…์ด ํ˜ผํ•ฉ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ Fine-tuningํ•˜๋Š” ๋™์•ˆ ๋‹ค์–‘ํ•œ ์ž‘์—…์˜ ๋น„์œจ์„ ๊ท ํ˜• ์žˆ๊ฒŒ ์œ ์ง€ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋ฐฉ๋ฒ•์€ _examples-proportional mixing_ ์ „๋žต[82]์œผ๋กœ, ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ฒฐํ•ฉํ•˜๊ณ  ํ˜ผํ•ฉ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฐ ์ธ์Šคํ„ด์Šค๋ฅผ ๋™์ผํ•˜๊ฒŒ ์ƒ˜ํ”Œ๋งํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ตœ๊ทผ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด ๊ณ ํ’ˆ์งˆ ์ปฌ๋ ‰์…˜(์˜ˆ: FLAN [67] ๋ฐ P3 [167])์˜ ์ƒ˜ํ”Œ๋ง ๋น„์œจ์„ ๋†’์ด๋ฉด ์ผ๋ฐ˜์ ์œผ๋กœ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[95, 69]. ๋˜ํ•œ Instruction Tuning ์ค‘์— ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ํฌํ•จ๋  ์ˆ˜ ์žˆ๋Š” ์ตœ๋Œ€ ์˜ˆ์ œ ์ˆ˜๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด _maximum cap_์„ ์„ค์ •ํ•˜๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ด๋ฉฐ, ์ด๋Š” ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ „์ฒด ๋ถ„ํฌ๋ฅผ ์••๋„ํ•˜๋Š” ๊ฒƒ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ์„ค์ •๋ฉ๋‹ˆ๋‹ค[82, 95]. ์‹ค์ œ๋กœ maximum cap์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋”ฐ๋ผ ์ˆ˜์ฒœ ๋˜๋Š” ์ˆ˜๋งŒ ๊ฐœ๋กœ ์„ค์ •๋ฉ๋‹ˆ๋‹ค[67, 69]. ์ตœ๊ทผ์—๋Š” ๊ธฐ์กด Instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ(ํ‘œ III)๊ฐ€ ์ฃผ๋กœ ํŠน์ • ์ธก๋ฉด์—์„œ LLMs์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ๋‹จ์ผ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋งŒ์œผ๋กœ๋Š” ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ์ข…ํ•ฉ์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์—†๋‹ค๋Š” ์‚ฌ์‹ค์ด ๊ฒฝํ—˜์ ์œผ๋กœ ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค[353]. ๋”ฐ๋ผ์„œ ๋‹ค์–‘ํ•œ ๋Šฅ๋ ฅ(์˜ˆ: NLP ์ž‘์—… ๋ฐ์ดํ„ฐ[292], ์ฑ„ํŒ… ๋ฐ์ดํ„ฐ[148], ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ[354])์˜ ๊ท ํ˜• ์žกํžŒ ํ–ฅ์ƒ์„ ์œ„ํ•ด ๊ธฐ์กด Instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ˜ผํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.
**๋ฐ์ดํ„ฐ ๋ถ„ํฌ์˜ ๊ท ํ˜• ์œ ์ง€**. ๋ช…๋ น์–ด ํŠœ๋‹์—๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—…์ด ํ˜ผํ•ฉ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋™์•ˆ ๋‹ค์–‘ํ•œ ์ž‘์—…์˜ ๋น„์œจ์„ ๊ท ํ˜• ์žˆ๊ฒŒ ์œ ์ง€ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋ฐฉ๋ฒ•์€ _์˜ˆ์ œ ๋น„๋ก€ ํ˜ผํ•ฉ_ ์ „๋žต[82]์œผ๋กœ, ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ฒฐํ•ฉํ•˜๊ณ  ํ˜ผํ•ฉ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฐ ์ธ์Šคํ„ด์Šค๋ฅผ ๋™์ผํ•˜๊ฒŒ ์ƒ˜ํ”Œ๋งํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ตœ๊ทผ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด ๊ณ ํ’ˆ์งˆ ์ปฌ๋ ‰์…˜(์˜ˆ: FLAN [67] ๋ฐ P3 [167])์˜ ์ƒ˜ํ”Œ๋ง ๋น„์œจ์„ ๋†’์ด๋ฉด ์ผ๋ฐ˜์ ์œผ๋กœ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[95, 69]. ๋˜ํ•œ ๋ช…๋ น์–ด ํŠœ๋‹ ์ค‘์— ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ํฌํ•จ๋  ์ˆ˜ ์žˆ๋Š” ์ตœ๋Œ€ ์˜ˆ์ œ ์ˆ˜๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด _์ตœ๋Œ€ ํ•œ๋„_๋ฅผ ์„ค์ •ํ•˜๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ด๋ฉฐ, ์ด๋Š” ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ „์ฒด ๋ถ„ํฌ๋ฅผ ์••๋„ํ•˜๋Š” ๊ฒƒ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ์„ค์ •๋ฉ๋‹ˆ๋‹ค[82, 95]. ์‹ค์ œ๋กœ ์ตœ๋Œ€ ํ•œ๋„๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋”ฐ๋ผ ์ˆ˜์ฒœ ๋˜๋Š” ์ˆ˜๋งŒ ๊ฐœ๋กœ ์„ค์ •๋ฉ๋‹ˆ๋‹ค[67, 69]. ์ตœ๊ทผ์—๋Š” ๊ธฐ์กด ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ(ํ‘œ III)๊ฐ€ ์ฃผ๋กœ ํŠน์ • ์ธก๋ฉด์—์„œ LLM์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ๋‹จ์ผ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋งŒ์œผ๋กœ๋Š” ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ์ข…ํ•ฉ์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์—†๋‹ค๋Š” ์‚ฌ์‹ค์ด ๊ฒฝํ—˜์ ์œผ๋กœ ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค[353]. ๋”ฐ๋ผ์„œ ๋‹ค์–‘ํ•œ ๋Šฅ๋ ฅ(์˜ˆ: NLP ์ž‘์—… ๋ฐ์ดํ„ฐ[292], ์ฑ„ํŒ… ๋ฐ์ดํ„ฐ[148], ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ[354])์˜ ๊ท ํ˜• ์žกํžŒ ํ–ฅ์ƒ์„ ์œ„ํ•ด ๊ธฐ์กด ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ˜ผํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.
cluding GPT-4 (Achiam et al., 2023), were trained on predominantly human-generated text; similarly, diffusion models like DALL-E (Ramesh et al., 2021), Stable Diffusion (Rombach et al., 2022), Midjourney (Midjourney, 2023) are trained on web-scale image datasets. These training corpora already potentially exhaust all the available clean data on the internet. A growing number of synthetic data generated with these increasingly popular models starts to populate the web, often indistinguishable from "real" data. We have thus already raced into the future where our training corpora are irreversibly mixed with synthetic data and this situation stands to get worse. Recent works call attention to the potential dramatic deterioration in the resulting models, an effect referred to as _"model collapse"_(Shumailov et al., 2023). Facets of this phenomenon have been demonstrated _empirically_ in various settings (Hataya et al., 2023; Martinez et al., 2023; Bohacek and Farid, 2023; Briesch et al., 2023; Guo et al., 2023). Theoretically, a few works are emerging to analyze the effect of iterative training on self-generated (or mixed) data (see Related Work): (Shumailov et al., 2023) coin the term _"model collapse"_ to characterize complete reversion to the mean, Alemohammad et al. (2023) analyze _"self-consuming loops"_ and Bertrand et al. (2023) show that iterative synthetic training leads to a _"clueless generator"_.
GPT-4(Achiam et al., 2023)๋ฅผ ๋น„๋กฏํ•œ ๋Œ€๋ถ€๋ถ„์˜ ์ธ๊ณต์ง€๋Šฅ์€ ์ฃผ๋กœ ์‚ฌ๋žŒ์ด ์ƒ์„ฑํ•œ ํ…์ŠคํŠธ๋กœ ํ•™์Šต๋˜์—ˆ์œผ๋ฉฐ, DALL-E(Ramesh et al., 2021), Stable Diffusion(Rombach et al., 2022), Midjourney(Midjourney, 2023)์™€ ๊ฐ™์€ diffusion ๋ชจ๋ธ๋„ ์›น ์Šค์ผ€์ผ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•™์Šต ์ฝ”ํผ์Šค๋Š” ์ด๋ฏธ ์ธํ„ฐ๋„ท์—์„œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์†Œ์ง„ํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ ์  ๋” ๋Œ€์ค‘ํ™”๋˜๊ณ  ์žˆ๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋ธ๋กœ ์ƒ์„ฑ๋œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ์›น์„ ์ฑ„์šฐ๊ธฐ ์‹œ์ž‘ํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ์ข…์ข… "์‹ค์ œ" ๋ฐ์ดํ„ฐ์™€ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” ์ด๋ฏธ ํ•™์Šต ์ฝ”ํผ์Šค๊ฐ€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ๋น„๊ฐ€์—ญ์ ์œผ๋กœ ํ˜ผํ•ฉ๋˜๋Š” ๋ฏธ๋ž˜์— ์ง„์ž…ํ–ˆ์œผ๋ฉฐ ์ด๋Ÿฌํ•œ ์ƒํ™ฉ์€ ๋”์šฑ ์•…ํ™”๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฒฐ๊ณผ ๋ชจ๋ธ์˜ ์ž ์žฌ์ ์ธ ๊ธ‰๊ฒฉํ•œ ์•…ํ™”์— ์ฃผ๋ชฉํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋ฅผ _"model collapse"_(Shumailov et al., 2023)๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ˜„์ƒ์˜ ์ธก๋ฉด์€ ๋‹ค์–‘ํ•œ ํ™˜๊ฒฝ์—์„œ _"๊ฒฝํ—˜์ ์œผ๋กœ"_ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Hataya et al., 2023; Martinez et al., 2023; Bohacek and Farid, 2023; Briesch et al., 2023; Guo et al., 2023). ์ด๋ก ์ ์œผ๋กœ, ์ž์ฒด ์ƒ์„ฑ(๋˜๋Š” ํ˜ผํ•ฉ) ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋ฐ˜๋ณต ํ•™์Šต์˜ ํšจ๊ณผ๋ฅผ ๋ถ„์„ํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ์ž‘์—…์ด ๋“ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(๊ด€๋ จ ์—ฐ๊ตฌ ์ฐธ์กฐ): (Shumailov et al., 2023)์€ _"model collapse"_๋ผ๋Š” ์šฉ์–ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ท ์œผ๋กœ ์™„์ „ํžˆ ๋˜๋Œ์•„๊ฐ€๋Š” ๊ฒƒ์„ ํŠน์ง•์ง“๊ณ , Alemohammad et al. (2023)์€ _"self-consuming loops"_๋ฅผ ๋ถ„์„ํ•˜๊ณ , Bertrand et al. (2023)์€ ๋ฐ˜๋ณต์ ์ธ ํ•ฉ์„ฑ ํ•™์Šต์ด _"clueless generator"_๋กœ ์ด์–ด์ง„๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
GPT-4(Achiam et al., 2023)๋ฅผ ๋น„๋กฏํ•œ ๋Œ€๋ถ€๋ถ„์˜ ์ธ๊ณต์ง€๋Šฅ์€ ์ฃผ๋กœ ์‚ฌ๋žŒ์ด ์ƒ์„ฑํ•œ ํ…์ŠคํŠธ๋กœ ํ•™์Šต๋˜์—ˆ์œผ๋ฉฐ, DALL-E(Ramesh et al., 2021), Stable Diffusion(Rombach et al., 2022), Midjourney(Midjourney, 2023)์™€ ๊ฐ™์€ ํ™•์‚ฐ ๋ชจ๋ธ๋„ ์›น ์Šค์ผ€์ผ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•™์Šต ์ฝ”ํผ์Šค๋Š” ์ด๋ฏธ ์ธํ„ฐ๋„ท์—์„œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์†Œ์ง„ํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ ์  ๋” ๋Œ€์ค‘ํ™”๋˜๊ณ  ์žˆ๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋ธ๋กœ ์ƒ์„ฑ๋œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ์›น์„ ์ฑ„์šฐ๊ธฐ ์‹œ์ž‘ํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ์ข…์ข… "์‹ค์ œ" ๋ฐ์ดํ„ฐ์™€ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” ์ด๋ฏธ ํ•™์Šต ์ฝ”ํผ์Šค๊ฐ€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ๋น„๊ฐ€์—ญ์ ์œผ๋กœ ํ˜ผํ•ฉ๋˜๋Š” ๋ฏธ๋ž˜์— ์ง„์ž…ํ–ˆ์œผ๋ฉฐ ์ด๋Ÿฌํ•œ ์ƒํ™ฉ์€ ๋”์šฑ ์•…ํ™”๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฒฐ๊ณผ ๋ชจ๋ธ์˜ ์ž ์žฌ์ ์ธ ๊ธ‰๊ฒฉํ•œ ์•…ํ™”์— ์ฃผ๋ชฉํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋ฅผ _"๋ชจ๋ธ ๋ถ•๊ดด"_(Shumailov et al., 2023)๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ˜„์ƒ์˜ ์ธก๋ฉด์€ ๋‹ค์–‘ํ•œ ํ™˜๊ฒฝ์—์„œ _"๊ฒฝํ—˜์ ์œผ๋กœ"_ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Hataya et al., 2023; Martinez et al., 2023; Bohacek and Farid, 2023; Briesch et al., 2023; Guo et al., 2023). ์ด๋ก ์ ์œผ๋กœ, ์ž์ฒด ์ƒ์„ฑ(๋˜๋Š” ํ˜ผํ•ฉ) ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋ฐ˜๋ณต ํ•™์Šต์˜ ํšจ๊ณผ๋ฅผ ๋ถ„์„ํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ์ž‘์—…์ด ๋“ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(๊ด€๋ จ ์—ฐ๊ตฌ ์ฐธ์กฐ): (Shumailov et al., 2023)์€ _"๋ชจ๋ธ ๋ถ•๊ดด"_๋ผ๋Š” ์šฉ์–ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ท ์œผ๋กœ ์™„์ „ํžˆ ๋˜๋Œ์•„๊ฐ€๋Š” ๊ฒƒ์„ ํŠน์ง•์ง“๊ณ , Alemohammad et al. (2023)์€ _"์ž๊ธฐ ์†Œ๋น„ ๋ฃจํ”„"_๋ฅผ ๋ถ„์„ํ•˜๊ณ , Bertrand et al. (2023)์€ ๋ฐ˜๋ณต์ ์ธ ํ•ฉ์„ฑ ํ•™์Šต์ด _"๋ฌด์ง€ํ•œ ์ƒ์„ฑ์ž"๋กœ ์ด์–ด์ง„๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
The quantization level is strictly bound to the memory usage and inference speed, and thus a trade-off must be made when choosing a proper quantization level. As we can see, the 8-bit quantization method has almost the same or even lower perplexities compared to the original FP16 model, demonstrating that it is a good choice for deploying LLMs on personal computers, with only half size of the FP16 one. The 6-bit models also achieve decent PPLs comparable to the 8-bit one, making it a better balance of speed and performance. When we use a more aggressive quantization level, the performance drastically decreases (i.e., higher PPL), especially for 3-bit and 2-bit. We also discover that larger models are less sensitive to quantization methods than smaller ones. For example, the performance of 33B models changes much more mildly than the others. A similar result is also observed when comparing Plus-7B and Plus-13B models. This might indicate that though 2-bit and 3-bit quantization are less effective for smaller models, it might be a promising way to deploy larger models without significant performance loss. This is extremely helpful when the users only have limited computing resources and still want to try large language models. This might also imply that the quantized training method may become a main-stream approach for training large language models, especially for those with limited training resources.
์–‘์žํ™” ์ˆ˜์ค€์€ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰๊ณผ ์ถ”๋ก  ์†๋„์— ์—„๊ฒฉํ•˜๊ฒŒ ์ œํ•œ๋˜๋ฏ€๋กœ ์ ์ ˆํ•œ quantization ์ˆ˜์ค€์„ ์„ ํƒํ•  ๋•Œ ์ ˆ์ถฉ์•ˆ์„ ์ฐพ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ณด์‹œ๋‹ค์‹œํ”ผ 8-bit quantization ๋ฐฉ์‹์€ ์›๋ž˜ FP16 ๋ชจ๋ธ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ๊ฑฐ์˜ ๋™์ผํ•˜๊ฑฐ๋‚˜ ์‹ฌ์ง€์–ด ๋” ๋‚ฎ์€ perplexity๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์–ด FP16 ๋ชจ๋ธ์˜ ์ ˆ๋ฐ˜ ํฌ๊ธฐ๋งŒ์œผ๋กœ ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ์— LLM์„ ๋ฐฐํฌํ•˜๋Š” ๋ฐ ์ข‹์€ ์„ ํƒ์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. 6-bit ๋ชจ๋ธ๋„ 8-bit ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์ˆ˜์ค€์˜ ๋›ฐ์–ด๋‚œ PPL์„ ๋‹ฌ์„ฑํ•˜์—ฌ ์†๋„์™€ ์„ฑ๋Šฅ์˜ ๊ท ํ˜•์ด ๋” ์ž˜ ์žกํ˜€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋” ๊ณต๊ฒฉ์ ์ธ quantization ์ˆ˜์ค€์„ ์‚ฌ์šฉํ•˜๋ฉด ํŠนํžˆ 3-bit์™€ 2-bit์˜ ๊ฒฝ์šฐ ์„ฑ๋Šฅ์ด ๊ธ‰๊ฒฉํžˆ ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค(์ฆ‰, PPL์ด ๋†’์•„์ง). ๋˜ํ•œ ๋” ํฐ ๋ชจ๋ธ์ด ๋” ์ž‘์€ ๋ชจ๋ธ๋ณด๋‹ค quantization ๋ฐฉ์‹์— ๋œ ๋ฏผ๊ฐํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 33B ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ ๋‹ค๋ฅธ ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ์˜จํ™”ํ•˜๊ฒŒ ๋ณ€ํ™”ํ•ฉ๋‹ˆ๋‹ค. Plus-7B์™€ Plus-13B ๋ชจ๋ธ์„ ๋น„๊ตํ•  ๋•Œ๋„ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๊ฐ€ ๊ด€์ฐฐ๋ฉ๋‹ˆ๋‹ค. ์ด๋Š” 2-bit ๋ฐ 3-bit quantization์ด ์†Œํ˜• ๋ชจ๋ธ์—๋Š” ๋œ ํšจ๊ณผ์ ์ด์ง€๋งŒ, ์„ฑ๋Šฅ ์†์‹ค ์—†์ด ๋Œ€ํ˜• ๋ชจ๋ธ์„ ๋ฐฐํฌํ•˜๋Š” ๋ฐ ์œ ๋งํ•œ ๋ฐฉ๋ฒ•์ด ๋  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋Š” ์‚ฌ์šฉ์ž๊ฐ€ ์ œํ•œ๋œ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค๋งŒ ๊ฐ€์ง€๊ณ  ์žˆ์ง€๋งŒ ์—ฌ์ „ํžˆ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๊ณ ์ž ํ•  ๋•Œ ๋งค์šฐ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ quantized training ๋ฐฉ๋ฒ•์ด ํŠนํžˆ ์ œํ•œ๋œ ํ›ˆ๋ จ ๋ฆฌ์†Œ์Šค๋ฅผ ๊ฐ€์ง„ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์ฃผ๋ฅ˜ ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋  ์ˆ˜ ์žˆ์Œ์„ ์˜๋ฏธํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
์–‘์žํ™” ์ˆ˜์ค€์€ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰๊ณผ ์ถ”๋ก  ์†๋„์— ์—„๊ฒฉํ•˜๊ฒŒ ์ œํ•œ๋˜๋ฏ€๋กœ ์ ์ ˆํ•œ ์–‘์žํ™” ์ˆ˜์ค€์„ ์„ ํƒํ•  ๋•Œ ์ ˆ์ถฉ์•ˆ์„ ์ฐพ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ณด์‹œ๋‹ค์‹œํ”ผ 8๋น„ํŠธ ์–‘์žํ™” ๋ฐฉ์‹์€ ์›๋ž˜ FP16 ๋ชจ๋ธ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ๊ฑฐ์˜ ๋™์ผํ•˜๊ฑฐ๋‚˜ ์‹ฌ์ง€์–ด ๋” ๋‚ฎ์€ ๋‚œํ•ด์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์–ด FP16 ๋ชจ๋ธ์˜ ์ ˆ๋ฐ˜ ํฌ๊ธฐ๋งŒ์œผ๋กœ ๊ฐœ์ธ์šฉ ์ปดํ“จํ„ฐ์— LLM์„ ๋ฐฐํฌํ•˜๋Š” ๋ฐ ์ข‹์€ ์„ ํƒ์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. 6๋น„ํŠธ ๋ชจ๋ธ๋„ 8๋น„ํŠธ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์ˆ˜์ค€์˜ ๋›ฐ์–ด๋‚œ PPL์„ ๋‹ฌ์„ฑํ•˜์—ฌ ์†๋„์™€ ์„ฑ๋Šฅ์˜ ๊ท ํ˜•์ด ๋” ์ž˜ ์žกํ˜€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋” ๊ณต๊ฒฉ์ ์ธ ์–‘์žํ™” ์ˆ˜์ค€์„ ์‚ฌ์šฉํ•˜๋ฉด ํŠนํžˆ 3๋น„ํŠธ์™€ 2๋น„ํŠธ์˜ ๊ฒฝ์šฐ ์„ฑ๋Šฅ์ด ๊ธ‰๊ฒฉํžˆ ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค(์ฆ‰, PPL์ด ๋†’์•„์ง). ๋˜ํ•œ ๋” ํฐ ๋ชจ๋ธ์ด ๋” ์ž‘์€ ๋ชจ๋ธ๋ณด๋‹ค ์–‘์žํ™” ๋ฐฉ์‹์— ๋œ ๋ฏผ๊ฐํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 33B ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ ๋‹ค๋ฅธ ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ์˜จํ™”ํ•˜๊ฒŒ ๋ณ€ํ™”ํ•ฉ๋‹ˆ๋‹ค. Plus-7B์™€ Plus-13B ๋ชจ๋ธ์„ ๋น„๊ตํ•  ๋•Œ๋„ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๊ฐ€ ๊ด€์ฐฐ๋ฉ๋‹ˆ๋‹ค. ์ด๋Š” 2๋น„ํŠธ ๋ฐ 3๋น„ํŠธ ์–‘์žํ™”๊ฐ€ ์†Œํ˜• ๋ชจ๋ธ์—๋Š” ๋œ ํšจ๊ณผ์ ์ด์ง€๋งŒ, ์„ฑ๋Šฅ ์†์‹ค ์—†์ด ๋Œ€ํ˜• ๋ชจ๋ธ์„ ๋ฐฐํฌํ•˜๋Š” ๋ฐ ์œ ๋งํ•œ ๋ฐฉ๋ฒ•์ด ๋  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋Š” ์‚ฌ์šฉ์ž๊ฐ€ ์ œํ•œ๋œ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค๋งŒ ๊ฐ€์ง€๊ณ  ์žˆ์ง€๋งŒ ์—ฌ์ „ํžˆ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๊ณ ์ž ํ•  ๋•Œ ๋งค์šฐ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์–‘์žํ™” ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์ด ํŠนํžˆ ์ œํ•œ๋œ ํ›ˆ๋ จ ๋ฆฌ์†Œ์Šค๋ฅผ ๊ฐ€์ง„ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์ฃผ๋ฅ˜ ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋  ์ˆ˜ ์žˆ์Œ์„ ์˜๋ฏธํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Scaling the instructions._ It has been widely shown that scaling the number of tasks can largely enhance the generalization ability of LLMs [28, 67, 88]. With the increasing of the task number, the model performance initially shows a continuous growth pattern, while the gain becomes negligible when it reaches a certain level [69, 88]. A plausible speculation is that a certain number of representative tasks can provide relatively sufficient knowledge and adding more tasks may not bring additional gains [69]. Also, it is beneficial to enhance the diversity of the task descriptions in several aspects, such as length, structure, and creativity [28]. As for the number of instances per task, it has been found that a small number of instances can usually saturate the generalization performance of the model to perform a specific task [67, 69]. Specially, several recent work [349, 350] has explored the effect of fine-tuning with a small amount of high-quality instruction data (_e.g.,_ one or a few thousand instances), showing very promising results on the evaluation tasks. In contrast, another line of studies continue to explore the scaling effect of instruction data [351, 352]. For example, Orca [351] scales up the synthesized instances to 5 million with step-by-step explanations, and it achieves superior performance across a wide range of tasks compared to the methods tuned with instruction data.
\(\bullet\)_Instruction ์Šค์ผ€์ผ๋ง._ ์ž‘์—… ์ˆ˜๋ฅผ ํ™•์žฅํ•˜๋ฉด LLM์˜ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋„๋ฆฌ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค[28, 67, 88]. ์ž‘์—… ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๋ชจ๋ธ ์„ฑ๋Šฅ์€ ์ฒ˜์Œ์—๋Š” ์ง€์†์ ์ธ ์„ฑ์žฅ ํŒจํ„ด์„ ๋ณด์ด๋Š” ๋ฐ˜๋ฉด, ์ผ์ • ์ˆ˜์ค€์— ๋„๋‹ฌํ•˜๋ฉด ๊ทธ ์ด๋“์€ ๋ฌด์‹œํ•  ์ˆ˜ ์žˆ์„ ์ •๋„๋กœ ์ค„์–ด๋“ญ๋‹ˆ๋‹ค[69, 88]. ๊ทธ๋Ÿด๋“ฏํ•œ ์ถ”์ธก์€ ์ผ์ • ์ˆ˜์˜ ๋Œ€ํ‘œ์ ์ธ ์ž‘์—…์ด ์ƒ๋Œ€์ ์œผ๋กœ ์ถฉ๋ถ„ํ•œ ์ง€์‹์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ์ž‘์—…์„ ๋” ์ถ”๊ฐ€ํ•ด๋„ ์ถ”๊ฐ€์ ์ธ ์ด๋“์„ ๊ฐ€์ ธ์˜ค์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค[69]. ๋˜ํ•œ ๊ธธ์ด, ๊ตฌ์กฐ, ์ฐฝ์˜์„ฑ ๋“ฑ ์—ฌ๋Ÿฌ ์ธก๋ฉด์—์„œ ์ž‘์—… ์„ค๋ช…์˜ ๋‹ค์–‘์„ฑ์„ ๋†’์ด๋Š” ๊ฒƒ์ด ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค[28]. ์ž‘์—…๋‹น ์ธ์Šคํ„ด์Šค ์ˆ˜์˜ ๊ฒฝ์šฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ์†Œ์ˆ˜์˜ ์ธ์Šคํ„ด์Šค๊ฐ€ ํŠน์ • ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ์„ฑ๋Šฅ์„ ํฌํ™” ์ƒํƒœ๋กœ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[67, 69]. ํŠนํžˆ, ์ตœ๊ทผ ๋ช‡๋ช‡ ์—ฐ๊ตฌ[349, 350]์—์„œ๋Š” ์†Œ๋Ÿ‰์˜ ๊ณ ํ’ˆ์งˆ Instruction ๋ฐ์ดํ„ฐ(์˜ˆ: ์ˆ˜์ฒœ ๊ฐœ์˜ ์ธ์Šคํ„ด์Šค)๋กœ Fine-tuningํ•˜๋Š” ํšจ๊ณผ๋ฅผ ํƒ๊ตฌํ•˜์—ฌ ํ‰๊ฐ€ ์ž‘์—…์—์„œ ๋งค์šฐ ์œ ๋งํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” Instruction ๋ฐ์ดํ„ฐ์˜ ์Šค์ผ€์ผ๋ง ํšจ๊ณผ๋ฅผ ๊ณ„์† ํƒ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[351, 352]. ์˜ˆ๋ฅผ ๋“ค์–ด, Orca[351]๋Š” ํ•ฉ์„ฑ๋œ ์ธ์Šคํ„ด์Šค๋ฅผ ๋‹จ๊ณ„๋ณ„ ์„ค๋ช…๊ณผ ํ•จ๊ป˜ 5๋ฐฑ๋งŒ ๊ฐœ๋กœ ํ™•์žฅํ•˜์—ฌ Instruction ๋ฐ์ดํ„ฐ๋กœ ์กฐ์ •๋œ ๋ฐฉ๋ฒ•์— ๋น„ํ•ด ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—…์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_์ธ์ŠคํŠธ๋Ÿญ์…˜ ์Šค์ผ€์ผ๋ง._ ์ž‘์—… ์ˆ˜๋ฅผ ํ™•์žฅํ•˜๋ฉด LLM์˜ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋„๋ฆฌ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค[28, 67, 88]. ์ž‘์—… ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๋ชจ๋ธ ์„ฑ๋Šฅ์€ ์ฒ˜์Œ์—๋Š” ์ง€์†์ ์ธ ์„ฑ์žฅ ํŒจํ„ด์„ ๋ณด์ด๋Š” ๋ฐ˜๋ฉด, ์ผ์ • ์ˆ˜์ค€์— ๋„๋‹ฌํ•˜๋ฉด ๊ทธ ์ด๋“์€ ๋ฌด์‹œํ•  ์ˆ˜ ์žˆ์„ ์ •๋„๋กœ ์ค„์–ด๋“ญ๋‹ˆ๋‹ค[69, 88]. ๊ทธ๋Ÿด๋“ฏํ•œ ์ถ”์ธก์€ ์ผ์ • ์ˆ˜์˜ ๋Œ€ํ‘œ์ ์ธ ์ž‘์—…์ด ์ƒ๋Œ€์ ์œผ๋กœ ์ถฉ๋ถ„ํ•œ ์ง€์‹์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ์ž‘์—…์„ ๋” ์ถ”๊ฐ€ํ•ด๋„ ์ถ”๊ฐ€์ ์ธ ์ด๋“์„ ๊ฐ€์ ธ์˜ค์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค[69]. ๋˜ํ•œ ๊ธธ์ด, ๊ตฌ์กฐ, ์ฐฝ์˜์„ฑ ๋“ฑ ์—ฌ๋Ÿฌ ์ธก๋ฉด์—์„œ ์ž‘์—… ์„ค๋ช…์˜ ๋‹ค์–‘์„ฑ์„ ๋†’์ด๋Š” ๊ฒƒ์ด ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค[28]. ์ž‘์—…๋‹น ์ธ์Šคํ„ด์Šค ์ˆ˜์˜ ๊ฒฝ์šฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ์†Œ์ˆ˜์˜ ์ธ์Šคํ„ด์Šค๊ฐ€ ํŠน์ • ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ์„ฑ๋Šฅ์„ ํฌํ™” ์ƒํƒœ๋กœ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[67, 69]. ํŠนํžˆ, ์ตœ๊ทผ ๋ช‡๋ช‡ ์—ฐ๊ตฌ[349, 350]์—์„œ๋Š” ์†Œ๋Ÿ‰์˜ ๊ณ ํ’ˆ์งˆ ๋ช…๋ น ๋ฐ์ดํ„ฐ(์˜ˆ: ์ˆ˜์ฒœ ๊ฐœ์˜ ์ธ์Šคํ„ด์Šค)๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ํšจ๊ณผ๋ฅผ ํƒ๊ตฌํ•˜์—ฌ ํ‰๊ฐ€ ์ž‘์—…์—์„œ ๋งค์šฐ ์œ ๋งํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ง€์‹œ ๋ฐ์ดํ„ฐ์˜ ์Šค์ผ€์ผ๋ง ํšจ๊ณผ๋ฅผ ๊ณ„์† ํƒ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[351, 352]. ์˜ˆ๋ฅผ ๋“ค์–ด, Orca[351]๋Š” ํ•ฉ์„ฑ๋œ ์ธ์Šคํ„ด์Šค๋ฅผ ๋‹จ๊ณ„๋ณ„ ์„ค๋ช…๊ณผ ํ•จ๊ป˜ 5๋ฐฑ๋งŒ ๊ฐœ๋กœ ํ™•์žฅํ•˜์—ฌ ์ง€์‹œ ๋ฐ์ดํ„ฐ๋กœ ์กฐ์ •๋œ ๋ฐฉ๋ฒ•์— ๋น„ํ•ด ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—…์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.
Inspired by InstructGPT[2] and PRMs[36], we train two reward models to predict the quality of the instructions and the correctness of each step in the answer respectively:1. Instruction Reward Model (IRM): This model aims to judge the quality of the evolved instructions on three aspects: i) Definition, ii) Precision, and iii) Integrity. To produce the ranking list training data of IRM, for each instruction, we firstly use ChatGPT and Wizard-E 4 to generate 2-4 evolved instructions respectively. Then we leverage Wizard-E to rank the quality of those 4-8 instructions. Footnote 4: Wizard-E named Wizard-E-Generator, which is an Alpha version fine-tuned Llama model specifically used to execute _Evol-Instruct_ without APIs. 2. Process-supervised Reward Model (PRM): As there is no powerful open-source math reasoning LLMs before this work, there is no simple way to support highly precise process supervision without professional human-labelers and close-source ChatGPT. Therefore, we depend on ChatGPT to provide process supervision, and ask it to assess the correctness of each step in the solutions generated by our model. 3. PPO training. We evolve the original math (GSM8k + MATH) instructions by 8 turns, increasing the data size from 15k to 96k. We use IRM and PRM to generate the instruction reward (\(r^{I}\)) and the answer reward (\(r^{A}\)). Then apply a product as the final reward \(r=r^{I}\cdot r^{A}\).
์ธ์ŠคํŠธ๋ŸญํŠธGPT[2]์™€ PRM[36]์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ๋‹ต๋ณ€์˜ ๊ฐ ๋‹จ๊ณ„์˜ ์ง€์นจ ํ’ˆ์งˆ๊ณผ ์ •ํ™•์„ฑ์„ ๊ฐ๊ฐ ์˜ˆ์ธกํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค:1. Instruction Reward Model (IRM): ์ด ๋ชจ๋ธ์€ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์—์„œ ์ง„ํ™”๋œ instructions์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค: i) ์ •์˜, ii) ์ •๋ฐ€๋„, iii) ๋ฌด๊ฒฐ์„ฑ. IRM์˜ ์ˆœ์œ„ ๋ชฉ๋ก ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ instruction์— ๋Œ€ํ•ด ๋จผ์ € ChatGPT์™€ Wizard-E 4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ๊ฐ 2-4๊ฐœ์˜ ์ง„ํ™”๋œ instructions๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ Wizard-E๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ด๋Ÿฌํ•œ 4~8๊ฐœ์˜ instructions์˜ ํ’ˆ์งˆ์„ ์ˆœ์œ„ํ™”ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ์ฃผ 4: Wizard-E๋Š” Wizard-E-Generator๋ผ๋Š” ์ด๋ฆ„์œผ๋กœ, API ์—†์ด _Evol-Instruct_๋ฅผ ์‹คํ–‰ํ•˜๋Š” ๋ฐ ํŠน๋ณ„ํžˆ ์‚ฌ์šฉ๋˜๋Š” ์•ŒํŒŒ ๋ฒ„์ „์˜ Fine-tuned Llama ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.2. Process-supervised Reward Model (PRM): ์ด ์ž‘์—… ์ด์ „์—๋Š” ๊ฐ•๋ ฅํ•œ ์˜คํ”ˆ ์†Œ์Šค ์ˆ˜ํ•™ ์ถ”๋ก  LLM์ด ์—†์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ „๋ฌธ ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ์™€ ํ์‡„ํ˜• ChatGPT ์—†์ด ๋งค์šฐ ์ •๋ฐ€ํ•œ ํ”„๋กœ์„ธ์Šค ๊ฐ๋…์„ ์ง€์›ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ„๋‹จํ•œ ๋ฐฉ๋ฒ•์ด ์—†์—ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” ChatGPT์— ์˜์กดํ•˜์—ฌ ๋ชจ๋ธ์ด ์ƒ์„ฑํ•œ ์†”๋ฃจ์…˜์˜ ๊ฐ ๋‹จ๊ณ„์˜ ์ •ํ™•์„ฑ์„ ํ‰๊ฐ€ํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. 3. PPO training. ์›๋ž˜์˜ ์ˆ˜ํ•™(GSM8k + MATH) instructions๋ฅผ 8๋ฒˆ ์ง„ํ™”์‹œ์ผœ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๋ฅผ 15k์—์„œ 96k๋กœ ๋Š˜๋ฆฝ๋‹ˆ๋‹ค. IRM๊ณผ PRM์„ ์‚ฌ์šฉํ•˜์—ฌ instruction reward (\(r^{I}\))๊ณผ answer reward (\(r^{A}\))๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ œํ’ˆ์„ ์ตœ์ข… ๋ณด์ƒ \(r=r^{I}\cdot r^{A}\)๋กœ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค.
์ธ์ŠคํŠธ๋ŸญํŠธGPT[2]์™€ PRM[36]์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ๋‹ต๋ณ€์˜ ๊ฐ ๋‹จ๊ณ„์˜ ์ง€์นจ ํ’ˆ์งˆ๊ณผ ์ •ํ™•์„ฑ์„ ๊ฐ๊ฐ ์˜ˆ์ธกํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค:1. ์ง€์นจ ๋ณด์ƒ ๋ชจ๋ธ(IRM): ์ด ๋ชจ๋ธ์€ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์—์„œ ์ง„ํ™”๋œ ์ง€์นจ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค: i) ์ •์˜, ii) ์ •๋ฐ€๋„, iii) ๋ฌด๊ฒฐ์„ฑ. IRM์˜ ์ˆœ์œ„ ๋ชฉ๋ก ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ์ง€์นจ์— ๋Œ€ํ•ด ๋จผ์ € ChatGPT์™€ Wizard-E 4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ๊ฐ 2-4๊ฐœ์˜ ์ง„ํ™”๋œ ์ง€์นจ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ Wizard-E๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ด๋Ÿฌํ•œ 4~8๊ฐœ์˜ ์ง€์นจ์˜ ํ’ˆ์งˆ์„ ์ˆœ์œ„ํ™”ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ์ฃผ 4: Wizard-E๋Š” Wizard-E-Generator๋ผ๋Š” ์ด๋ฆ„์œผ๋กœ, API ์—†์ด _Evol-Instruct_๋ฅผ ์‹คํ–‰ํ•˜๋Š” ๋ฐ ํŠน๋ณ„ํžˆ ์‚ฌ์šฉ๋˜๋Š” ์•ŒํŒŒ ๋ฒ„์ „์˜ ๋ฏธ์„ธ ์กฐ์ •๋œ Llama ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.2. ํ”„๋กœ์„ธ์Šค ๊ฐ๋… ๋ณด์ƒ ๋ชจ๋ธ(PRM): ์ด ์ž‘์—… ์ด์ „์—๋Š” ๊ฐ•๋ ฅํ•œ ์˜คํ”ˆ ์†Œ์Šค ์ˆ˜ํ•™ ์ถ”๋ก  LLM์ด ์—†์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ „๋ฌธ ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ์™€ ํ์‡„ํ˜• ChatGPT ์—†์ด ๋งค์šฐ ์ •๋ฐ€ํ•œ ํ”„๋กœ์„ธ์Šค ๊ฐ๋…์„ ์ง€์›ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ„๋‹จํ•œ ๋ฐฉ๋ฒ•์ด ์—†์—ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” ChatGPT์— ์˜์กดํ•˜์—ฌ ๋ชจ๋ธ์ด ์ƒ์„ฑํ•œ ์†”๋ฃจ์…˜์˜ ๊ฐ ๋‹จ๊ณ„์˜ ์ •ํ™•์„ฑ์„ ํ‰๊ฐ€ํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. 3. PPO ํ›ˆ๋ จ. ์›๋ž˜์˜ ์ˆ˜ํ•™(GSM8k + MATH) ์ง€์นจ์„ 8๋ฒˆ ์ง„ํ™”์‹œ์ผœ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๋ฅผ 15k์—์„œ 96k๋กœ ๋Š˜๋ฆฝ๋‹ˆ๋‹ค. IRM๊ณผ PRM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ง€์นจ ๋ณด์ƒ(\(r^{I}\))๊ณผ ๋‹ต๋ณ€ ๋ณด์ƒ(\(r^{A}\))์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ œํ’ˆ์„ ์ตœ์ข… ๋ณด์ƒ \(r=r^{I}\cdot r^{A}\)๋กœ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค.
**Natural Science Education**: LLM-based agents can communicate with humans fluently, often being utilized to develop agent-based educational tools [115, 117, 118, 119]. For example, [115] develops agent-based education systems to facilitate students learning of experimental design, methodologies, and analysis. The objective of these systems is to enhance students' critical thinking and problem-solving skills, while also fostering a deeper comprehension of scientific principles. Math Agents [117] can assist researchers in exploring, discovering, solving and proving mathematical problems. Additionally, it can communicate with humans and aids them in understanding and using mathematics. [118] utilize the capabilities of CodeX [119] to automatically solve and explain university-level mathematical problems, which can be used as education tools to teach students and researchers. CodeHelp [120] is an education agent for programming. It offers many useful features, such as setting course-specific keywords, monitoring student queries, and providing feedback to the system. EduChat [86] is an LLM-based agent designed specifically for the education domain. It provides personalized, equitable, and empathetic educational support to teachers, students, and parents through dialogue. FreeText [121] is an agent that utilizes LLMs to automatically assess students' responses to open-ended questions and offer feedback.
**์ž์—ฐ๊ณผํ•™ ๊ต์œก**: LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ์ธ๊ฐ„๊ณผ ์œ ์ฐฝํ•˜๊ฒŒ ์†Œํ†ตํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜ ๊ต์œก ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ์ž์ฃผ ํ™œ์šฉ๋ฉ๋‹ˆ๋‹ค[115, 117, 118, 119]. ์˜ˆ๋ฅผ ๋“ค์–ด, [115]๋Š” ์‹คํ—˜ ์„ค๊ณ„, ๋ฐฉ๋ฒ•๋ก  ๋ฐ ๋ถ„์„์— ๋Œ€ํ•œ ํ•™์ƒ์˜ ํ•™์Šต์„ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜ ๊ต์œก ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹œ์Šคํ…œ์˜ ๋ชฉํ‘œ๋Š” ํ•™์ƒ๋“ค์˜ ๋น„ํŒ์  ์‚ฌ๊ณ ์™€ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋™์‹œ์— ๊ณผํ•™์  ์›๋ฆฌ์— ๋Œ€ํ•œ ๊นŠ์€ ์ดํ•ด๋ฅผ ์ด‰์ง„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. Math Agents [117]๋Š” ์—ฐ๊ตฌ์ž๋“ค์ด ์ˆ˜ํ•™์  ๋ฌธ์ œ๋ฅผ ํƒ์ƒ‰, ๋ฐœ๊ฒฌ, ํ•ด๊ฒฐ ๋ฐ ์ฆ๋ช…ํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ธ๊ฐ„๊ณผ ์†Œํ†ตํ•˜์—ฌ ์ˆ˜ํ•™์„ ์ดํ•ดํ•˜๊ณ  ์‚ฌ์šฉํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. [118]์€ CodeX [119]์˜ ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜์—ฌ ๋Œ€ํ•™ ์ˆ˜์ค€์˜ ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ์ž๋™์œผ๋กœ ํ’€๊ณ  ์„ค๋ช…ํ•˜๋ฉฐ, ์ด๋ฅผ ๊ต์œก ๋„๊ตฌ๋กœ ํ™œ์šฉํ•˜์—ฌ ํ•™์ƒ๊ณผ ์—ฐ๊ตฌ์ž๋ฅผ ๊ฐ€๋ฅด์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. CodeHelp [120]๋Š” ํ”„๋กœ๊ทธ๋ž˜๋ฐ์„ ์œ„ํ•œ ๊ต์œก์šฉ ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค. ์ด ์—์ด์ „ํŠธ๋Š” ์ฝ”์Šค๋ณ„ ํ‚ค์›Œ๋“œ ์„ค์ •, ํ•™์ƒ ์ฟผ๋ฆฌ ๋ชจ๋‹ˆํ„ฐ๋ง, ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ํ”ผ๋“œ๋ฐฑ ์ œ๊ณต ๋“ฑ ๋‹ค์–‘ํ•œ ์œ ์šฉํ•œ ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. EduChat [86]์€ ๊ต์œก ์˜์—ญ์„ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค. ์ด ์—์ด์ „ํŠธ๋Š” ๋Œ€ํ™”๋ฅผ ํ†ตํ•ด ๊ต์‚ฌ, ํ•™์ƒ, ํ•™๋ถ€๋ชจ์—๊ฒŒ ๊ฐœ์ธํ™”๋˜๊ณ  ๊ณตํ‰ํ•˜๋ฉฐ ๊ณต๊ฐ์ ์ธ ๊ต์œก ์ง€์›์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. FreeText [121]๋Š” LLM์„ ํ™œ์šฉํ•˜์—ฌ ์ฃผ๊ด€์‹ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ํ•™์ƒ์˜ ๋‹ต๋ณ€์„ ์ž๋™์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ณ  ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๋Š” ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค.
**์ž์—ฐ๊ณผํ•™ ๊ต์œก**: LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ์ธ๊ฐ„๊ณผ ์œ ์ฐฝํ•˜๊ฒŒ ์†Œํ†ตํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜ ๊ต์œก ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ์ž์ฃผ ํ™œ์šฉ๋ฉ๋‹ˆ๋‹ค[115, 117, 118, 119]. ์˜ˆ๋ฅผ ๋“ค์–ด, [115]๋Š” ์‹คํ—˜ ์„ค๊ณ„, ๋ฐฉ๋ฒ•๋ก  ๋ฐ ๋ถ„์„์— ๋Œ€ํ•œ ํ•™์ƒ์˜ ํ•™์Šต์„ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜ ๊ต์œก ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹œ์Šคํ…œ์˜ ๋ชฉํ‘œ๋Š” ํ•™์ƒ๋“ค์˜ ๋น„ํŒ์  ์‚ฌ๊ณ ์™€ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋™์‹œ์— ๊ณผํ•™์  ์›๋ฆฌ์— ๋Œ€ํ•œ ๊นŠ์€ ์ดํ•ด๋ฅผ ์ด‰์ง„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ˆ˜ํ•™ ์—์ด์ „ํŠธ [117]๋Š” ์—ฐ๊ตฌ์ž๋“ค์ด ์ˆ˜ํ•™์  ๋ฌธ์ œ๋ฅผ ํƒ์ƒ‰, ๋ฐœ๊ฒฌ, ํ•ด๊ฒฐ ๋ฐ ์ฆ๋ช…ํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ธ๊ฐ„๊ณผ ์†Œํ†ตํ•˜์—ฌ ์ˆ˜ํ•™์„ ์ดํ•ดํ•˜๊ณ  ์‚ฌ์šฉํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. [118]์€ CodeX [119]์˜ ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜์—ฌ ๋Œ€ํ•™ ์ˆ˜์ค€์˜ ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ์ž๋™์œผ๋กœ ํ’€๊ณ  ์„ค๋ช…ํ•˜๋ฉฐ, ์ด๋ฅผ ๊ต์œก ๋„๊ตฌ๋กœ ํ™œ์šฉํ•˜์—ฌ ํ•™์ƒ๊ณผ ์—ฐ๊ตฌ์ž๋ฅผ ๊ฐ€๋ฅด์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. CodeHelp [120]๋Š” ํ”„๋กœ๊ทธ๋ž˜๋ฐ์„ ์œ„ํ•œ ๊ต์œก์šฉ ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค. ์ด ์—์ด์ „ํŠธ๋Š” ์ฝ”์Šค๋ณ„ ํ‚ค์›Œ๋“œ ์„ค์ •, ํ•™์ƒ ์ฟผ๋ฆฌ ๋ชจ๋‹ˆํ„ฐ๋ง, ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ํ”ผ๋“œ๋ฐฑ ์ œ๊ณต ๋“ฑ ๋‹ค์–‘ํ•œ ์œ ์šฉํ•œ ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์—๋“€์ฑ— [86]์€ ๊ต์œก ์˜์—ญ์„ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค. ์ด ์—์ด์ „ํŠธ๋Š” ๋Œ€ํ™”๋ฅผ ํ†ตํ•ด ๊ต์‚ฌ, ํ•™์ƒ, ํ•™๋ถ€๋ชจ์—๊ฒŒ ๊ฐœ์ธํ™”๋˜๊ณ  ๊ณตํ‰ํ•˜๋ฉฐ ๊ณต๊ฐ์ ์ธ ๊ต์œก ์ง€์›์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. FreeText [121]๋Š” LLM์„ ํ™œ์šฉํ•˜์—ฌ ์ฃผ๊ด€์‹ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ํ•™์ƒ์˜ ๋‹ต๋ณ€์„ ์ž๋™์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ณ  ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๋Š” ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค.
Socio-technical entanglementIt can be difficult to pinpoint, and causally explain the origins of, the harm that is inflicted through the use of generative AI systems. For instance, experts often disagree on whether a given AI output is hazardous [130], the time horizon over which harms from AI systems manifest can be months if not years, and the impact of AI can be multifaceted and subtle rather than deterministic and direct [131]. This is because AI systems are socio-technically entangled, which means that "the interaction of technical and social components determines whether risk manifests" rather than either component singularly [16]. Further, this entanglement makes it challenging to predict what harms may be caused when a generative AI system meets existing socio-technical contexts, and it is difficult to precisely pinpoint their causal impact. Indeed, assessing the causal impact of AI models on the people who interact with them is a well-established (and largely unresolved) research question in social media studies [132, 133, 134, 135, 136]. One approach is to consider counterfactuals. For instance, Mazeika et al. [114] argue that safety assessments of models should consider what is enabled by using an AI model "above and beyond what a human could accomplish with a search engine." Examples exist in the algorithmic audit literature, but this is methodologically difficult to implement [137].
์‚ฌํšŒ ๊ธฐ์ˆ ์  ์–ฝํž˜Generative AI ์‹œ์Šคํ…œ์˜ ์‚ฌ์šฉ์œผ๋กœ ์ธํ•ด ๊ฐ€ํ•ด์ง€๋Š” ํ”ผํ•ด์˜ ๊ธฐ์›์„ ์ •ํ™•ํžˆ ํŒŒ์•…ํ•˜๊ณ  ์ธ๊ณผ์ ์œผ๋กœ ์„ค๋ช…ํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ „๋ฌธ๊ฐ€๋“ค์€ ํŠน์ • AI ์ถœ๋ ฅ์ด ์œ„ํ—˜ํ•œ์ง€ ์—ฌ๋ถ€์— ๋Œ€ํ•ด ์ข…์ข… ์˜๊ฒฌ์ด ์ผ์น˜ํ•˜์ง€ ์•Š์œผ๋ฉฐ, AI ์‹œ์Šคํ…œ์œผ๋กœ ์ธํ•œ ํ”ผํ•ด๊ฐ€ ๋‚˜ํƒ€๋‚˜๋Š” ์‹œ๊ธฐ๋Š” ์ˆ˜๋…„์ด ์•„๋‹Œ ์ˆ˜๊ฐœ์›”์ด ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, AI์˜ ์˜ํ–ฅ์€ ๊ฒฐ์ •๋ก ์ ์ด๊ณ  ์ง์ ‘์ ์ธ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๋‹ค๋ฉด์ ์ด๊ณ  ๋ฏธ๋ฌ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[131]. ์ด๋Š” AI ์‹œ์Šคํ…œ์ด socio-technically entangled, ์ฆ‰ "์œ„ํ—˜์ด ๋‚˜ํƒ€๋‚˜๋Š”์ง€ ์—ฌ๋ถ€๋Š” ๋‘ ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ๋‹จ๋…์œผ๋กœ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๊ธฐ์ˆ ์  ์š”์†Œ์™€ ์‚ฌํšŒ์  ์š”์†Œ์˜ ์ƒํ˜ธ ์ž‘์šฉ์— ์˜ํ•ด ๊ฒฐ์ •๋œ๋‹ค"๋Š” ์˜๋ฏธ์ž…๋‹ˆ๋‹ค[16]. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ์–ฝํž˜์œผ๋กœ ์ธํ•ด generative AI ์‹œ์Šคํ…œ์ด ๊ธฐ์กด์˜ socio-technical contexts์™€ ๋งŒ๋‚ฌ์„ ๋•Œ ์–ด๋–ค ํ”ผํ•ด๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์˜ˆ์ธกํ•˜๊ธฐ ์–ด๋ ต๊ณ , ๊ทธ ์ธ๊ณผ์  ์˜ํ–ฅ์„ ์ •ํ™•ํžˆ ํŒŒ์•…ํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ AI ๋ชจ๋ธ์ด ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ๋ฏธ์น˜๋Š” ์ธ๊ณผ์  ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ์†Œ์…œ ๋ฏธ๋””์–ด ์—ฐ๊ตฌ์—์„œ ์ž˜ ์ •๋ฆฝ๋œ(๊ทธ๋ฆฌ๊ณ  ๋Œ€๋ถ€๋ถ„ ํ•ด๊ฒฐ๋˜์ง€ ์•Š์€) ์—ฐ๊ตฌ ์งˆ๋ฌธ์ž…๋‹ˆ๋‹ค[132, 133, 134, 135, 136]. ํ•œ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์€ counterfactuals๋ฅผ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Mazeika ๋“ฑ[114]์€ ๋ชจ๋ธ์˜ ์•ˆ์ „์„ฑ ํ‰๊ฐ€์—์„œ AI ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•จ์œผ๋กœ์จ "์ธ๊ฐ„์ด ๊ฒ€์ƒ‰ ์—”์ง„์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ ์ด์ƒ์˜" ๊ธฐ๋Šฅ์ด ํ™œ์„ฑํ™”๋˜๋Š”์ง€ ๊ณ ๋ คํ•ด์•ผ ํ•œ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๊ฐ์‚ฌ ๋ฌธํ—Œ์—๋Š” ์˜ˆ๊ฐ€ ์กด์žฌํ•˜์ง€๋งŒ, ์ด๋Š” ๋ฐฉ๋ฒ•๋ก ์ ์œผ๋กœ ๊ตฌํ˜„ํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค[137].
์‚ฌํšŒ ๊ธฐ์ˆ ์  ์–ฝํž˜Generative AI ์‹œ์Šคํ…œ์˜ ์‚ฌ์šฉ์œผ๋กœ ์ธํ•ด ๊ฐ€ํ•ด์ง€๋Š” ํ”ผํ•ด์˜ ๊ธฐ์›์„ ์ •ํ™•ํžˆ ํŒŒ์•…ํ•˜๊ณ  ์ธ๊ณผ์ ์œผ๋กœ ์„ค๋ช…ํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ „๋ฌธ๊ฐ€๋“ค์€ ํŠน์ • AI ์ถœ๋ ฅ์ด ์œ„ํ—˜ํ•œ์ง€ ์—ฌ๋ถ€์— ๋Œ€ํ•ด ์ข…์ข… ์˜๊ฒฌ์ด ์ผ์น˜ํ•˜์ง€ ์•Š์œผ๋ฉฐ, AI ์‹œ์Šคํ…œ์œผ๋กœ ์ธํ•œ ํ”ผํ•ด๊ฐ€ ๋‚˜ํƒ€๋‚˜๋Š” ์‹œ๊ธฐ๋Š” ์ˆ˜๋…„์ด ์•„๋‹Œ ์ˆ˜๊ฐœ์›”์ด ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, AI์˜ ์˜ํ–ฅ์€ ๊ฒฐ์ •๋ก ์ ์ด๊ณ  ์ง์ ‘์ ์ธ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๋‹ค๋ฉด์ ์ด๊ณ  ๋ฏธ๋ฌ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[131]. ์ด๋Š” AI ์‹œ์Šคํ…œ์ด ์‚ฌํšŒ ๊ธฐ์ˆ ์ ์œผ๋กœ ์–ฝํ˜€ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ธ๋ฐ, ์ด๋Š” "์œ„ํ—˜์ด ๋‚˜ํƒ€๋‚˜๋Š”์ง€ ์—ฌ๋ถ€๋Š” ๋‘ ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ๋‹จ๋…์œผ๋กœ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๊ธฐ์ˆ ์  ์š”์†Œ์™€ ์‚ฌํšŒ์  ์š”์†Œ์˜ ์ƒํ˜ธ ์ž‘์šฉ์— ์˜ํ•ด ๊ฒฐ์ •๋œ๋‹ค"๋Š” ์˜๋ฏธ์ž…๋‹ˆ๋‹ค[16]. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ์–ฝํž˜์œผ๋กœ ์ธํ•ด ์ƒ์„ฑํ˜• AI ์‹œ์Šคํ…œ์ด ๊ธฐ์กด์˜ ์‚ฌํšŒ ๊ธฐ์ˆ ์  ๋งฅ๋ฝ๊ณผ ๋งŒ๋‚ฌ์„ ๋•Œ ์–ด๋–ค ํ”ผํ•ด๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์˜ˆ์ธกํ•˜๊ธฐ ์–ด๋ ต๊ณ , ๊ทธ ์ธ๊ณผ์  ์˜ํ–ฅ์„ ์ •ํ™•ํžˆ ํŒŒ์•…ํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ AI ๋ชจ๋ธ์ด ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ๋ฏธ์น˜๋Š” ์ธ๊ณผ์  ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ์†Œ์…œ ๋ฏธ๋””์–ด ์—ฐ๊ตฌ์—์„œ ์ž˜ ์ •๋ฆฝ๋œ(๊ทธ๋ฆฌ๊ณ  ๋Œ€๋ถ€๋ถ„ ํ•ด๊ฒฐ๋˜์ง€ ์•Š์€) ์—ฐ๊ตฌ ์งˆ๋ฌธ์ž…๋‹ˆ๋‹ค[132, 133, 134, 135, 136]. ํ•œ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋ฐ˜๋Œ€ ์‚ฌ๋ก€๋ฅผ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Mazeika ๋“ฑ[114]์€ ๋ชจ๋ธ์˜ ์•ˆ์ „์„ฑ ํ‰๊ฐ€์—์„œ AI ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•จ์œผ๋กœ์จ "์ธ๊ฐ„์ด ๊ฒ€์ƒ‰ ์—”์ง„์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ ์ด์ƒ์˜" ๊ธฐ๋Šฅ์ด ํ™œ์„ฑํ™”๋˜๋Š”์ง€ ๊ณ ๋ คํ•ด์•ผ ํ•œ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๊ฐ์‚ฌ ๋ฌธํ—Œ์—๋Š” ์˜ˆ๊ฐ€ ์กด์žฌํ•˜์ง€๋งŒ, ์ด๋Š” ๋ฐฉ๋ฒ•๋ก ์ ์œผ๋กœ ๊ตฌํ˜„ํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค[137].
**Demonstration Format.** After selecting task examples, the next step is to integrate and format them into a natural language prompt for LLMs. A straightforward method is to instantiate a pre-defined template with the corresponding input-output pairs [36]. To construct more informative templates, recent studies consider adding task descriptions [69] or enhancing the reasoning capability of LLMs with chain-of-thought prompts [33]. For instance, in [166], the authors collect a large-scale dataset with task descriptions written by humans. After tuning with this dataset, the performance on seen tasks can be boosted, and LLMs can also generalize to unseen tasks to some extent. To reduce the annotation costs, a semi-automated approach has been proposed in [143] by employing a seed set consisting of human-written task descriptions to guide LLMs to generate task descriptions for new tasks. Since it is costly to manually annotate demonstration formats for different tasks, some work also studies how to automatically generate high-quality ones. As two representative methods, Auto-CoT [434] leverages LLMs with the zero-shot prompt _"Let's think step by step"_ for generating intermediate reasoning steps, while least-tomost prompting [439] first queries LLMs to perform problem decomposition and then utilizes LLMs to sequentially solve sub-problems based on the intermediate answers to previously solved ones.
**Demonstration Format.** ์ž‘์—… ์˜ˆ์ œ๋ฅผ ์„ ํƒํ•œ ํ›„ ๋‹ค์Œ ๋‹จ๊ณ„๋Š” ์ด๋ฅผ LLM์šฉ ์ž์—ฐ์–ด prompt์— ํ†ตํ•ฉํ•˜๊ณ  ํ˜•์‹์„ ์ง€์ •ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ฐ„๋‹จํ•œ ๋ฐฉ๋ฒ•์€ ์‚ฌ์ „ ์ •์˜๋œ template์„ ํ•ด๋‹น input-output ์Œ [36]์œผ๋กœ ์ธ์Šคํ„ด์Šคํ™”ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ณด๋‹ค ์œ ์ตํ•œ template์„ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ์ž‘์—… ์„ค๋ช… [69]์„ ์ถ”๊ฐ€ํ•˜๊ฑฐ๋‚˜ chain-of-thought prompts [33]๋ฅผ ํ†ตํ•ด LLM์˜ ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, [166]์—์„œ ์ €์ž๋“ค์€ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ์ž‘์—… ์„ค๋ช…์ด ํฌํ•จ๋œ ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ tuningํ•œ ํ›„, ๋ณผ ์ˆ˜ ์žˆ๋Š” ์ž‘์—…์˜ ์„ฑ๋Šฅ์„ ๋†’์ผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, LLM์€ ๋˜ํ•œ ๋ณด์ด์ง€ ์•Š๋Š” ์ž‘์—…์—๋„ ์–ด๋Š ์ •๋„ ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์„ ๋น„์šฉ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด [143]์—์„œ๋Š” ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ์ž‘์—… ์„ค๋ช…์œผ๋กœ ๊ตฌ์„ฑ๋œ ์‹œ๋“œ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์ด ์ƒˆ๋กœ์šด ์ž‘์—…์— ๋Œ€ํ•œ ์ž‘์—… ์„ค๋ช…์„ ์ƒ์„ฑํ•˜๋„๋ก ์•ˆ๋‚ดํ•˜๋Š” ๋ฐ˜์ž๋™ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๋Œ€ํ•œ demonstration formats์„ ์ˆ˜๋™์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ๋Š” ๋ฐ๋Š” ๋งŽ์€ ๋น„์šฉ์ด ๋“ค๊ธฐ ๋•Œ๋ฌธ์— ๊ณ ํ’ˆ์งˆ์˜ demonstration formats์„ ์ž๋™์œผ๋กœ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•๋„ ์—ฐ๊ตฌ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ๋‘ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์œผ๋กœ, Auto-CoT [434]๋Š” ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด LLM์— zero-shot prompt _"Let's think step by step"_๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, least-to-most prompting [439]๋Š” ๋จผ์ € LLM์— ๋ฌธ์ œ ๋ถ„ํ•ด๋ฅผ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์ฟผ๋ฆฌํ•œ ๋‹ค์Œ, ์ด์ „์— ํ’€์—ˆ๋˜ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์ค‘๊ฐ„ ๋‹ต๋ณ€์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜์—ฌ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•˜์œ„ ๋ฌธ์ œ๋ฅผ ์ˆœ์ฐจ์ ์œผ๋กœ ํ‘ธ๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค.
**๋ฐ๋ชจ ํ˜•์‹**. ์ž‘์—… ์˜ˆ์ œ๋ฅผ ์„ ํƒํ•œ ํ›„ ๋‹ค์Œ ๋‹จ๊ณ„๋Š” ์ด๋ฅผ LLM์šฉ ์ž์—ฐ์–ด ํ”„๋กฌํ”„ํŠธ์— ํ†ตํ•ฉํ•˜๊ณ  ํ˜•์‹์„ ์ง€์ •ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ฐ„๋‹จํ•œ ๋ฐฉ๋ฒ•์€ ์‚ฌ์ „ ์ •์˜๋œ ํ…œํ”Œ๋ฆฟ์„ ํ•ด๋‹น ์ž…๋ ฅ-์ถœ๋ ฅ ์Œ [36]์œผ๋กœ ์ธ์Šคํ„ด์Šคํ™”ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ณด๋‹ค ์œ ์ตํ•œ ํ…œํ”Œ๋ฆฟ์„ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ์ž‘์—… ์„ค๋ช… [69]์„ ์ถ”๊ฐ€ํ•˜๊ฑฐ๋‚˜ ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ํ”„๋กฌํ”„ํŠธ [33]๋ฅผ ํ†ตํ•ด LLM์˜ ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, [166]์—์„œ ์ €์ž๋“ค์€ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ์ž‘์—… ์„ค๋ช…์ด ํฌํ•จ๋œ ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ํŠœ๋‹ํ•œ ํ›„, ๋ณผ ์ˆ˜ ์žˆ๋Š” ์ž‘์—…์˜ ์„ฑ๋Šฅ์„ ๋†’์ผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, LLM์€ ๋˜ํ•œ ๋ณด์ด์ง€ ์•Š๋Š” ์ž‘์—…์—๋„ ์–ด๋Š ์ •๋„ ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์„ ๋น„์šฉ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด [143]์—์„œ๋Š” ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ์ž‘์—… ์„ค๋ช…์œผ๋กœ ๊ตฌ์„ฑ๋œ ์‹œ๋“œ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์ด ์ƒˆ๋กœ์šด ์ž‘์—…์— ๋Œ€ํ•œ ์ž‘์—… ์„ค๋ช…์„ ์ƒ์„ฑํ•˜๋„๋ก ์•ˆ๋‚ดํ•˜๋Š” ๋ฐ˜์ž๋™ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๋Œ€ํ•œ ๋ฐ๋ชจ ํ˜•์‹์„ ์ˆ˜๋™์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ๋Š” ๋ฐ๋Š” ๋งŽ์€ ๋น„์šฉ์ด ๋“ค๊ธฐ ๋•Œ๋ฌธ์— ๊ณ ํ’ˆ์งˆ์˜ ๋ฐ๋ชจ ํ˜•์‹์„ ์ž๋™์œผ๋กœ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•๋„ ์—ฐ๊ตฌ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ๋‘ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์œผ๋กœ, Auto-CoT [434]๋Š” ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด LLM์— ์ œ๋กœ ์ƒท ํ”„๋กฌํ”„ํŠธ _"์ฐจ๊ทผ์ฐจ๊ทผ ์ƒ๊ฐํ•ด ๋ด…์‹œ๋‹ค"_๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ์ตœ์†Œ-์ตœ๋Œ€ ํ”„๋กฌํ”„ํŠธ [439]๋Š” ๋จผ์ € LLM์— ๋ฌธ์ œ ๋ถ„ํ•ด๋ฅผ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์ฟผ๋ฆฌํ•œ ๋‹ค์Œ, ์ด์ „์— ํ’€์—ˆ๋˜ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์ค‘๊ฐ„ ๋‹ต๋ณ€์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜์—ฌ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•˜์œ„ ๋ฌธ์ œ๋ฅผ ์ˆœ์ฐจ์ ์œผ๋กœ ํ‘ธ๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค.
In this study, we introduce CT-LLM, a 2B large language model (LLM) that illustrates a pivotal shift towards prioritizing the Chinese language in developing LLMs. Uniquely initiated from scratch, CT-LLM diverges from the conventional methodology by primarily incorporating Chinese textual data, utilizing an extensive corpus of 1,200 billion tokens, including 800 billion Chinese tokens, 300 billion English tokens, and 100 billion code tokens. This strategic composition facilitates the model's exceptional proficiency in understanding and processing Chinese, a capability further enhanced through alignment techniques. Demonstrating remarkable performance on the CHC-Bench, CT-LLM excels in Chinese language tasks, and showcases its adeptness in English through SFT. This research challenges the prevailing paradigm of training LLMs predominantly on English corpora and then adapting them to other languages, broadening the horizons for LLM training methodologies. By open-sourcing the full process of training a Chinese LLM, including a detailed data processing procedure with the obtained Massive Appropriate Pretraining Chinese Corpus (MAP-CC), a well-chosen multidisciplinary Chinese Hard Case Benchmark (CHC-Bench), and the 2B-size Chinese Tiny LLM (CT-LLM), we aim to foster further exploration and innovation in both academia and industry, paving the way for more inclusive and versatile language models.
์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ค‘๊ตญ์–ด LLM ๊ฐœ๋ฐœ์—์„œ ์ค‘๊ตญ์–ด ์šฐ์„ ์ˆœ์œ„๋ฅผ ๊ฐ•์กฐํ•˜๋Š” ์ค‘์š”ํ•œ ์ „ํ™˜์„ ๋ณด์—ฌ์ฃผ๋Š” 2B Large Language Model (LLM)์ธ CT-LLM์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ๋ฐฉ๋ฒ•๋ก ์—์„œ ๋ฒ—์–ด๋‚˜ ์ฒ˜์Œ๋ถ€ํ„ฐ ์ฒ˜์Œ์œผ๋กœ ์‹œ์ž‘ํ•œ CT-LLM์€ 8000์–ต ๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ, 3000์–ต ๊ฐœ์˜ ์˜์–ด ํ† ํฐ, 1000์–ต ๊ฐœ์˜ ์ฝ”๋“œ ํ† ํฐ์„ ํฌํ•จํ•œ 1์กฐ 2000์–ต ๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๊ด‘๋ฒ”์œ„ํ•œ ๋ง๋ญ‰์น˜๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ์ฃผ๋กœ ํ†ตํ•ฉํ•˜๋Š” ๋…ํŠนํ•œ ๋ฐฉ๋ฒ•์„ ์ฑ„ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ „๋žต์  ๊ตฌ์„ฑ ๋•๋ถ„์— ์ค‘๊ตญ์–ด๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ฒ˜๋ฆฌํ•˜๋Š” ๋ชจ๋ธ์˜ ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๊ฒŒ ๋˜์—ˆ์œผ๋ฉฐ, alignment ๊ธฐ์ˆ ์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. CHC-Bench์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค€ CT-LLM์€ ์ค‘๊ตญ์–ด ์ž‘์—…์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋ฉฐ, Supervised Fine-tuning (SFT)๋ฅผ ํ†ตํ•ด ์˜์–ด์— ๋Œ€ํ•œ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ์˜์–ด ๋ง๋ญ‰์น˜๋กœ LLM์„ ํ›ˆ๋ จํ•œ ๋‹ค์Œ ๋‹ค๋ฅธ ์–ธ์–ด์— ์ ์šฉํ•˜๋Š” ๊ธฐ์กด์˜ ํŒจ๋Ÿฌ๋‹ค์ž„์— ๋„์ „ํ•˜์—ฌ LLM ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก ์˜ ์ง€ํ‰์„ ๋„“ํž™๋‹ˆ๋‹ค. ํš๋“ํ•œ Massive Appropriate Pretraining Chinese Corpus (MAP-CC)๋ฅผ ์‚ฌ์šฉํ•œ ์ƒ์„ธํ•œ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ ์ ˆ์ฐจ, ์ž˜ ์„ ํƒ๋œ ๋‹คํ•™์ œ Chinese Hard Case Benchmark (CHC-Bench), 2B ํฌ๊ธฐ์˜ Chinese Tiny LLM (CT-LLM) ๋“ฑ ์ค‘๊ตญ์–ด LLM ํ›ˆ๋ จ์˜ ์ „ ๊ณผ์ •์„ ์˜คํ”ˆ ์†Œ์‹ฑํ•จ์œผ๋กœ์จ ํ•™๊ณ„์™€ ์—…๊ณ„์—์„œ ๋”์šฑ ํฌ๊ด„์ ์ด๊ณ  ๋‹ค์žฌ๋‹ค๋Šฅํ•œ ์–ธ์–ด ๋ชจ๋ธ์„ ์œ„ํ•œ ๊ธธ์„ ์—ด์–ด๊ฐ€๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ค‘๊ตญ์–ด LLM ๊ฐœ๋ฐœ์—์„œ ์ค‘๊ตญ์–ด ์šฐ์„ ์ˆœ์œ„๋ฅผ ๊ฐ•์กฐํ•˜๋Š” ์ค‘์š”ํ•œ ์ „ํ™˜์„ ๋ณด์—ฌ์ฃผ๋Š” 2B ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์ธ CT-LLM์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ๋ฐฉ๋ฒ•๋ก ์—์„œ ๋ฒ—์–ด๋‚˜ ์ฒ˜์Œ๋ถ€ํ„ฐ ์ฒ˜์Œ์œผ๋กœ ์‹œ์ž‘ํ•œ CT-LLM์€ 8,000์–ต ๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ, 3,000์–ต ๊ฐœ์˜ ์˜์–ด ํ† ํฐ, 1,000์–ต ๊ฐœ์˜ ์ฝ”๋“œ ํ† ํฐ์„ ํฌํ•จํ•œ 1,2,000์–ต ๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๊ด‘๋ฒ”์œ„ํ•œ ๋ง๋ญ‰์น˜๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ์ฃผ๋กœ ํ†ตํ•ฉํ•˜๋Š” ๋…ํŠนํ•œ ๋ฐฉ๋ฒ•์„ ์ฑ„ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ „๋žต์  ๊ตฌ์„ฑ ๋•๋ถ„์— ์ค‘๊ตญ์–ด๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ฒ˜๋ฆฌํ•˜๋Š” ๋ชจ๋ธ์˜ ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๊ฒŒ ๋˜์—ˆ์œผ๋ฉฐ, ์ •๋ ฌ ๊ธฐ์ˆ ์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. CHC-Bench์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค€ CT-LLM์€ ์ค‘๊ตญ์–ด ์ž‘์—…์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋ฉฐ, SFT๋ฅผ ํ†ตํ•ด ์˜์–ด์— ๋Œ€ํ•œ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ์˜์–ด ๋ง๋ญ‰์น˜๋กœ LLM์„ ํ›ˆ๋ จํ•œ ๋‹ค์Œ ๋‹ค๋ฅธ ์–ธ์–ด์— ์ ์šฉํ•˜๋Š” ๊ธฐ์กด์˜ ํŒจ๋Ÿฌ๋‹ค์ž„์— ๋„์ „ํ•˜์—ฌ LLM ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก ์˜ ์ง€ํ‰์„ ๋„“ํž™๋‹ˆ๋‹ค. ํš๋“ํ•œ ๋Œ€๊ทœ๋ชจ ์ ์ ˆํ•œ ์‚ฌ์ „ ํ•™์Šต ์ค‘๊ตญ์–ด ๋ง๋ญ‰์น˜(MAP-CC)๋ฅผ ์‚ฌ์šฉํ•œ ์ƒ์„ธํ•œ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ ์ ˆ์ฐจ, ์ž˜ ์„ ํƒ๋œ ๋‹คํ•™์ œ ์ค‘๊ตญ์–ด ํ•˜๋“œ ์ผ€์ด์Šค ๋ฒค์น˜๋งˆํฌ(CHC-Bench), 2B ํฌ๊ธฐ์˜ ์ค‘๊ตญ์–ด ํƒ€์ด๋‹ˆ LLM(CT-LLM) ๋“ฑ ์ค‘๊ตญ์–ด LLM ํ›ˆ๋ จ์˜ ์ „ ๊ณผ์ •์„ ์˜คํ”ˆ ์†Œ์‹ฑํ•จ์œผ๋กœ์จ ํ•™๊ณ„์™€ ์—…๊ณ„์—์„œ ๋”์šฑ ํฌ๊ด„์ ์ด๊ณ  ๋‹ค์žฌ๋‹ค๋Šฅํ•œ ์–ธ์–ด ๋ชจ๋ธ์„ ์œ„ํ•œ ๊ธธ์„ ์—ด์–ด๊ฐ€๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
Zheng et al. [69] describe a novel network architecture called FD-Vision Mamba (FDVM-Net) designed to correct exposure abnormalities in endoscopic images. This is crucial for maintaining image quality to assist healthcare professionals in their decision-making process. FDVM-Net operates in the frequency domain and reconstructs the endoscopic image's frequency representation to improve exposure. Xing et al. [62] propose the SegmamBA module to enhance 3D feature modeling, which uses gated spatial convolution internally to enhance feature representation in spatial dimension to deal with long-distance dependency. Yue et al. [81] propose the MedMamba based on the State Space Model and convolutional layers for medical image classification. It can effectively capture long-range dependencies while maintaining the ability to extract local features, which is suitable for medical images of different modalities. Huang et al. [78] inherit the advantages of the original Mamba linear complexity and global receptive field, and propose an arbitrary mask mechanism to adapt the Mamba to the image reconstruction task, termed MambaMIR-GAN. Schiff et al. [87] extend the Mamba block into a component supporting bi-directional BiMamba and a MambaDNA supporting RC equivariant. Then, they use the MambaDNA as a basis for Caducucus and incorporate pre-training and fine-tuning strategies for DNA Sequence Modeling.
Zheng et al. [69]์€ ๋‚ด์‹œ๊ฒฝ ์ด๋ฏธ์ง€์˜ ๋…ธ์ถœ ์ด์ƒ์„ ๋ณด์ •ํ•˜๋„๋ก ์„ค๊ณ„๋œ FD-Vision Mamba(FDVM-Net)๋ผ๋Š” ์ƒˆ๋กœ์šด ๋„คํŠธ์›Œํฌ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์˜๋ฃŒ ์ „๋ฌธ๊ฐ€์˜ ์˜์‚ฌ ๊ฒฐ์ • ๊ณผ์ •์„ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด ์ด๋ฏธ์ง€ ํ’ˆ์งˆ์„ ์œ ์ง€ํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. FDVM-Net์€ ์ฃผํŒŒ์ˆ˜ ์˜์—ญ์—์„œ ์ž‘๋™ํ•˜๋ฉฐ ๋‚ด์‹œ๊ฒฝ ์ด๋ฏธ์ง€์˜ ์ฃผํŒŒ์ˆ˜ ํ‘œํ˜„์„ ์žฌ๊ตฌ์„ฑํ•˜์—ฌ ๋…ธ์ถœ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. Xing et al. [62]๋Š” 3D ํŠน์ง• ๋ชจ๋ธ๋ง์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด SegmamBA ๋ชจ๋“ˆ์„ ์ œ์•ˆํ•˜๋Š”๋ฐ, ์ด ๋ชจ๋“ˆ์€ ๋‚ด๋ถ€์ ์œผ๋กœ ๊ฒŒ์ดํŠธ ๊ณต๊ฐ„ convolution์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐ„ ์ฐจ์›์—์„œ ํŠน์ง• ํ‘œํ˜„์„ ํ–ฅ์ƒ์‹œ์ผœ ์›๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์„ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. Yue et al. [81]์€ ์˜๋ฃŒ ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜๋ฅผ ์œ„ํ•ด State Space Model๊ณผ convolutional layers๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ MedMamba๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ๊ตญ๋ถ€์ ์ธ ํŠน์ง•์„ ์ถ”์ถœํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์œ ์ง€ํ•˜๋ฉด์„œ ์žฅ๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์„ ํšจ๊ณผ์ ์œผ๋กœ ํฌ์ฐฉํ•  ์ˆ˜ ์žˆ์–ด ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์˜ ์˜๋ฃŒ ์ด๋ฏธ์ง€์— ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. Huang et al. [78]์€ ์›๋ž˜ Mamba ์„ ํ˜• ๋ณต์žก์„ฑ๊ณผ ์ „์—ญ ์ˆ˜์šฉ ํ•„๋“œ์˜ ์žฅ์ ์„ ๊ณ„์Šนํ•˜์—ฌ ์ž„์˜ ๋งˆ์Šคํฌ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ œ์•ˆํ•˜์—ฌ Mamba๋ฅผ ์ด๋ฏธ์ง€ ์žฌ๊ตฌ์„ฑ ์ž‘์—…์— ์ ์šฉํ•˜๋Š” MambaMIR-GAN์ด๋ผ๋Š” ์šฉ์–ด๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Schiff et al. [87]์€ Mamba ๋ธ”๋ก์„ ์–‘๋ฐฉํ–ฅ BiMamba๋ฅผ ์ง€์›ํ•˜๋Š” ๊ตฌ์„ฑ ์š”์†Œ์™€ RC ๋“ฑ๋ณ€์„ฑ์„ ์ง€์›ํ•˜๋Š” MambaDNA๋กœ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ๊ทธ๋“ค์€ MambaDNA๋ฅผ Caducucus์˜ ๊ธฐ์ดˆ๋กœ ์‚ฌ์šฉํ•˜๊ณ  DNA Sequence Modeling์„ ์œ„ํ•œ pre-training ๋ฐ Fine-tuning ์ „๋žต์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค.
Zheng et al. [69]์€ ๋‚ด์‹œ๊ฒฝ ์ด๋ฏธ์ง€์˜ ๋…ธ์ถœ ์ด์ƒ์„ ๋ณด์ •ํ•˜๋„๋ก ์„ค๊ณ„๋œ FD-Vision Mamba(FDVM-Net)๋ผ๋Š” ์ƒˆ๋กœ์šด ๋„คํŠธ์›Œํฌ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์˜๋ฃŒ ์ „๋ฌธ๊ฐ€์˜ ์˜์‚ฌ ๊ฒฐ์ • ๊ณผ์ •์„ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด ์ด๋ฏธ์ง€ ํ’ˆ์งˆ์„ ์œ ์ง€ํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. FDVM-Net์€ ์ฃผํŒŒ์ˆ˜ ์˜์—ญ์—์„œ ์ž‘๋™ํ•˜๋ฉฐ ๋‚ด์‹œ๊ฒฝ ์ด๋ฏธ์ง€์˜ ์ฃผํŒŒ์ˆ˜ ํ‘œํ˜„์„ ์žฌ๊ตฌ์„ฑํ•˜์—ฌ ๋…ธ์ถœ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. Xing et al. [62]๋Š” 3D ํŠน์ง• ๋ชจ๋ธ๋ง์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด SegmamBA ๋ชจ๋“ˆ์„ ์ œ์•ˆํ•˜๋Š”๋ฐ, ์ด ๋ชจ๋“ˆ์€ ๋‚ด๋ถ€์ ์œผ๋กœ ๊ฒŒ์ดํŠธ ๊ณต๊ฐ„ ์ปจ๋ณผ๋ฃจ์…˜์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐ„ ์ฐจ์›์—์„œ ํŠน์ง• ํ‘œํ˜„์„ ํ–ฅ์ƒ์‹œ์ผœ ์›๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์„ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. Yue et al. [81]์€ ์˜๋ฃŒ ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜๋ฅผ ์œ„ํ•ด ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ๊ณผ ์ปจ๋ณผ๋ฃจ์…˜ ๋ ˆ์ด์–ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ MedMamba๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ๊ตญ๋ถ€์ ์ธ ํŠน์ง•์„ ์ถ”์ถœํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์œ ์ง€ํ•˜๋ฉด์„œ ์žฅ๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์„ ํšจ๊ณผ์ ์œผ๋กœ ํฌ์ฐฉํ•  ์ˆ˜ ์žˆ์–ด ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์˜ ์˜๋ฃŒ ์ด๋ฏธ์ง€์— ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. Huang et al. [78]์€ ์›๋ž˜ ๋ง˜๋ฐ” ์„ ํ˜• ๋ณต์žก์„ฑ๊ณผ ์ „์—ญ ์ˆ˜์šฉ ํ•„๋“œ์˜ ์žฅ์ ์„ ๊ณ„์Šนํ•˜์—ฌ ์ž„์˜ ๋งˆ์Šคํฌ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ œ์•ˆํ•˜์—ฌ ๋ง˜๋ฐ”๋ฅผ ์ด๋ฏธ์ง€ ์žฌ๊ตฌ์„ฑ ์ž‘์—…์— ์ ์šฉํ•˜๋Š” MambaMIR-GAN์ด๋ผ๋Š” ์šฉ์–ด๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Schiff et al. [87]์€ ๋ง˜๋ฐ” ๋ธ”๋ก์„ ์–‘๋ฐฉํ–ฅ BiMamba๋ฅผ ์ง€์›ํ•˜๋Š” ๊ตฌ์„ฑ ์š”์†Œ์™€ RC ๋“ฑ๋ณ€์„ฑ์„ ์ง€์›ํ•˜๋Š” MambaDNA๋กœ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ๊ทธ๋“ค์€ MambaDNA๋ฅผ Caducucus์˜ ๊ธฐ์ดˆ๋กœ ์‚ฌ์šฉํ•˜๊ณ  DNA ์‹œํ€€์Šค ๋ชจ๋ธ๋ง์„ ์œ„ํ•œ ์‚ฌ์ „ ํ•™์Šต ๋ฐ ๋ฏธ์„ธ ์กฐ์ • ์ „๋žต์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค.
We now investigate the benefits of re-warming and re-decaying the learning rate (e.g., following a cosine schedule) for different values of \(\eta_{\text{max}}\). Specifically, we compare these models to two natural baselines: a model that does not re-warm, staying constant at \(\eta_{\text{min}}\) (\(3\cdot 10^{-5}\)), and a model that re-warms to the pre-training \(\eta_{\text{max}}\) (\(3\cdot 10^{-4}\)) but does not re-decay. We use the same two-dataset settings: we first pre-train on the Pile (\(\mathcal{D}_{0}\)) for 300B tokens and continually pre-train our model on SlimPjama (weak shift) or German Common Crawl (strong shift) as our \(\mathcal{D}_{1}\) datasets. The continual pre-training is conducted for the full size (300B and 200B tokens, respectively) of the datasets. The models that re-warm and re-decay the LR consider three strategies: re-warming to half the pre-training's \(\eta_{\text{max}}\) (\(1.5\cdot 10^{-4}\)), re-warming to the same \(\eta_{\text{max}}\) as pre-training (\(3\cdot 10^{-4}\)), and re-warming to twice the \(\eta_{\text{max}}\) of pre-training (\(6\cdot 10^{-4}\)). In all cases, the learning rate is cosine-decayed after linear warmup to reach \(\eta_{\text{min}}=0.1\cdot\eta_{\text{max}}\) by the end of training. Finally, we consider models trained on \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) as a third baseline to provide an upper bound on performance.
์ด์ œ ๋‹ค์–‘ํ•œ \(\eta_{\text{max}}\) ๊ฐ’์— ๋Œ€ํ•ด Learning Rate๋ฅผ ์žฌ๊ฐ€์—ดํ•˜๊ณ  ์žฌ๊ฐ์‡ ํ•˜๋Š” ๊ฒƒ(์˜ˆ: ์ฝ”์‚ฌ์ธ ์Šค์ผ€์ค„์— ๋”ฐ๋ฆ„)์˜ ์ด์ ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ์žฌ๊ฐ€์—ดํ•˜์ง€ ์•Š๊ณ  \(\eta_{\text{min}}\)(3\cdot 10^{-5}\))์—์„œ ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€ํ•˜๋Š” ๋ชจ๋ธ๊ณผ ์žฌ๊ฐ€์—ดํ•˜์ง€ ์•Š๊ณ  ์‚ฌ์ „ ํ•™์Šต \(\eta_{\text{max}}\)(3\cdot 10^{-4}\))๋กœ ์žฌ๊ฐ€์—ดํ•˜์ง€๋งŒ ์žฌ๊ฐ์‡ ํ•˜์ง€ ์•Š๋Š” ๋ชจ๋ธ์˜ ๋‘ ๊ฐ€์ง€ ์ž์—ฐ ๊ธฐ์ค€๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋™์ผํ•œ ๋‘ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์„ค์ •์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € Pile(\(\mathcal{D}_{0}\))์—์„œ 300B ํ† ํฐ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•˜๊ณ , SlimPjama(์•ฝํ•œ ์‹œํ”„ํŠธ) ๋˜๋Š” German Common Crawl(๊ฐ•ํ•œ ์‹œํ”„ํŠธ)์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ์—ฐ์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜์—ฌ \(\mathcal{D}_{1}\) ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์—ฐ์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ „์ฒด ํฌ๊ธฐ(๊ฐ๊ฐ 300B ๋ฐ 200B ํ† ํฐ)์— ๋Œ€ํ•ด ์ˆ˜ํ–‰๋ฉ๋‹ˆ๋‹ค. Learning Rate๋ฅผ ์žฌ๊ฐ€์—ดํ•˜๊ณ  ์žฌ๊ฐ์‡ ํ•˜๋Š” ๋ชจ๋ธ์€ ์‚ฌ์ „ ํ•™์Šต์˜ \(\eta_{\text{max}}\)(1.5\cdot 10^{-4}\), ์‚ฌ์ „ ํ•™์Šต๊ณผ ๋™์ผํ•œ \(\eta_{\text{max}}\)(3\cdot 10^{-4}\), ์‚ฌ์ „ ํ•™์Šต์˜ \(\eta_{\text{max}}\)(6\cdot 10^{-4}\), ์‚ฌ์ „ ํ•™์Šต์˜ ๋‘ ๋ฐฐ๋กœ ์žฌ๊ฐ€์—ดํ•˜๋Š” ์„ธ ๊ฐ€์ง€ ์ „๋žต์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๊ฒฝ์šฐ์— ์žˆ์–ด์„œ, Learning Rate๋Š” ์„ ํ˜• ์›Œ๋ฐ์—… ํ›„ ์ฝ”์‚ฌ์ธ ๊ฐ์†Œํ•˜์—ฌ ํ›ˆ๋ จ์ด ๋๋‚  ๋•Œ๊นŒ์ง€ \(\eta_{\text{min}}=0.1\cdot\eta_{\text{max}}\)์— ๋„๋‹ฌํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์„ฑ๋Šฅ์˜ ์ƒํ•œ์„ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\)์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ์„ธ ๋ฒˆ์งธ ๋ฒ ์ด์Šค๋ผ์ธ์œผ๋กœ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค.
์ด์ œ ๋‹ค์–‘ํ•œ \(\eta_{\text{max}}\) ๊ฐ’์— ๋Œ€ํ•ด ํ•™์Šต ์†๋„๋ฅผ ์žฌ๊ฐ€์—ดํ•˜๊ณ  ์žฌ๊ฐ์‡ ํ•˜๋Š” ๊ฒƒ(์˜ˆ: ์ฝ”์‚ฌ์ธ ์Šค์ผ€์ค„์— ๋”ฐ๋ฆ„)์˜ ์ด์ ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ์žฌ๊ฐ€์—ดํ•˜์ง€ ์•Š๊ณ  \(\eta_{\text{min}}\)(3\cdot 10^{-5}\))์—์„œ ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€ํ•˜๋Š” ๋ชจ๋ธ๊ณผ ์žฌ๊ฐ€์—ดํ•˜์ง€ ์•Š๊ณ  ์‚ฌ์ „ ํ•™์Šต \(\eta_{\text{max}}\)(3\cdot 10^{-4}\))๋กœ ์žฌ๊ฐ€์—ดํ•˜์ง€๋งŒ ์žฌ๊ฐ์‡ ํ•˜์ง€ ์•Š๋Š” ๋ชจ๋ธ์˜ ๋‘ ๊ฐ€์ง€ ์ž์—ฐ ๊ธฐ์ค€๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋™์ผํ•œ ๋‘ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์„ค์ •์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € Pile(\mathcal{D}_{0}\))์—์„œ 300B ํ† ํฐ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•˜๊ณ , SlimPjama(์•ฝํ•œ ์‹œํ”„ํŠธ) ๋˜๋Š” German Common Crawl(๊ฐ•ํ•œ ์‹œํ”„ํŠธ)์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ์—ฐ์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜์—ฌ \(\mathcal{D}_{1}\) ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์—ฐ์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ „์ฒด ํฌ๊ธฐ(๊ฐ๊ฐ 300B ๋ฐ 200B ํ† ํฐ)์— ๋Œ€ํ•ด ์ˆ˜ํ–‰๋ฉ๋‹ˆ๋‹ค. LR์„ ์žฌ๊ฐ€์—ดํ•˜๊ณ  ์žฌ๊ฐ์‡ ํ•˜๋Š” ๋ชจ๋ธ์€ ์‚ฌ์ „ ํ•™์Šต์˜ \(\eta_{\text{max}}\)(1.5\cdot 10^{-4}\), ์‚ฌ์ „ ํ•™์Šต๊ณผ ๋™์ผํ•œ \(\eta_{\text{max}}\)(3\cdot 10^{-4}\), ์‚ฌ์ „ ํ•™์Šต์˜ \(\eta_{\text{max}}\)(6\cdot 10^{-4}\), ์‚ฌ์ „ ํ•™์Šต์˜ ๋‘ ๋ฐฐ๋กœ ์žฌ๊ฐ€์—ดํ•˜๋Š” ์„ธ ๊ฐ€์ง€ ์ „๋žต์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๊ฒฝ์šฐ์— ์žˆ์–ด์„œ, ํ•™์Šต๋ฅ ์€ ์„ ํ˜• ์›Œ๋ฐ์—… ํ›„ ์ฝ”์‚ฌ์ธ ๊ฐ์†Œํ•˜์—ฌ ํ›ˆ๋ จ์ด ๋๋‚  ๋•Œ๊นŒ์ง€ \(\eta_{\text{min}}=0.1\cdot\eta_{\text{max}}\)์— ๋„๋‹ฌํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์„ฑ๋Šฅ์˜ ์ƒํ•œ์„ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\)์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ์„ธ ๋ฒˆ์งธ ๋ฒ ์ด์Šค๋ผ์ธ์œผ๋กœ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค.
1. Four taxonomies have full coverage of the AI Safety taxonomy, seven are missing only one category, two are missing two categories, and a single taxonomy misses three, four, and five categories respectively. 2. Nearly all of the taxonomies have additional categories (e.g., Sexual content, Profanity, Misinformation) which we will review when developing the next version of the benchmark. 3. Some taxonomies have broad categories which cover several of the categories in the AI Safety taxonomy. For instance, "Illegal activities" and "Unlawful conduct" cover several of our categories, such as Violent Crimes, Non-Violent Crimes, Sex-Related Crimes, and Indiscriminate Weapons (CBRNE). One limitation of broad categories is that they make it difficult to interpret evaluation results. 4. Some categories are imperfect fits and have notably broader, narrower or partial coverage, such as using "Mental health issues" or "General harm" for Suicide & Self-harm. Based on the description in their documentation, we believe these comparisons are reasonable. 5. Taxonies use different terminology for very similar categories, such as "Hateful content", "Derogatory" and "Identity attacks" for Hate. 6. Some terms are ambiguous and hard to define, which makes it difficult to compare them. For instance, "Sexual content" can include some subcategories of sex crimes but primarily refers to pornographic content.
1. 4๊ฐœ์˜ ๋ถ„๋ฅ˜๋ฒ•์€ AI Safety taxonomy๋ฅผ ์™„์ „ํžˆ ํฌ๊ด„ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, 7๊ฐœ์˜ ๋ถ„๋ฅ˜๋ฒ•์€ ํ•œ ๊ฐ€์ง€ ๋ฒ”์ฃผ๊ฐ€ ๋ˆ„๋ฝ๋˜์–ด ์žˆ๊ณ , 2๊ฐœ์˜ ๋ถ„๋ฅ˜๋ฒ•์€ ๋‘ ๊ฐ€์ง€ ๋ฒ”์ฃผ๊ฐ€ ๋ˆ„๋ฝ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋‹จ์ผ ๋ถ„๋ฅ˜๋ฒ•์€ ๊ฐ๊ฐ ์„ธ ๊ฐ€์ง€, ๋„ค ๊ฐ€์ง€, ๋‹ค์„ฏ ๊ฐ€์ง€ ๋ฒ”์ฃผ๊ฐ€ ๋ˆ„๋ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. 2. ๊ฑฐ์˜ ๋ชจ๋“  ๋ถ„๋ฅ˜๋ฒ•์—๋Š” ์ถ”๊ฐ€ ๋ฒ”์ฃผ(์˜ˆ: ์„ฑ์ ์ธ ์ฝ˜ํ…์ธ , ์š•์„ค, ์ž˜๋ชป๋œ ์ •๋ณด)๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋‹ค์Œ ๋ฒ„์ „์˜ benchmark๋ฅผ ๊ฐœ๋ฐœํ•  ๋•Œ ๊ฒ€ํ† ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. 3. ์ผ๋ถ€ ๋ถ„๋ฅ˜๋ฒ•์—๋Š” AI Safety taxonomy์˜ ์—ฌ๋Ÿฌ ๋ฒ”์ฃผ๋ฅผ ํฌ๊ด„ํ•˜๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ๋ฒ”์ฃผ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, '๋ถˆ๋ฒ• ํ™œ๋™'๊ณผ '๋ถˆ๋ฒ• ํ–‰์œ„'๋Š” ํญ๋ ฅ ๋ฒ”์ฃ„, ๋น„ํญ๋ ฅ ๋ฒ”์ฃ„, ์„ฑ ๊ด€๋ จ ๋ฒ”์ฃ„, ๋ฌด์ฐจ๋ณ„ ๋ฌด๊ธฐ(CBRNE) ๋“ฑ ์—ฌ๋Ÿฌ ๋ฒ”์ฃผ๋ฅผ ํฌ๊ด„ํ•ฉ๋‹ˆ๋‹ค. ๊ด‘๋ฒ”์œ„ํ•œ ๋ฒ”์ฃผ์˜ ํ•œ ๊ฐ€์ง€ ํ•œ๊ณ„๋Š” ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ํ•ด์„ํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. 4. ์ž์‚ด ๋ฐ ์žํ•ด์™€ ๊ฐ™์ด '์ •์‹  ๊ฑด๊ฐ• ๋ฌธ์ œ' ๋˜๋Š” '์ผ๋ฐ˜์ ์ธ ํ•ด์•…'์„ ์‚ฌ์šฉํ•˜๋Š” ๋“ฑ ์ผ๋ถ€ ๋ฒ”์ฃผ๋Š” ๋ถˆ์™„์ „ํ•œ ์ ํ•ฉ์„ฑ์„ ๊ฐ€์ง€๋ฉฐ ์ƒ๋‹นํžˆ ๋” ๋„“๊ฑฐ๋‚˜ ์ข๊ฑฐ๋‚˜ ๋ถ€๋ถ„์ ์œผ๋กœ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค. ๋ฌธ์„œ์— ์„ค๋ช…๋œ ๋‚ด์šฉ์— ๋”ฐ๋ฅด๋ฉด ์ด๋Ÿฌํ•œ ๋น„๊ต๊ฐ€ ํ•ฉ๋ฆฌ์ ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. 5. 'ํ˜์˜ค ์ฝ˜ํ…์ธ ', '๊ฒฝ๋ฉธ์ ', '์‹ ์› ๊ณต๊ฒฉ' ๋“ฑ ๋งค์šฐ ์œ ์‚ฌํ•œ ๋ฒ”์ฃผ์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•œ ์šฉ์–ด๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. 6. ์ผ๋ถ€ ์šฉ์–ด๋Š” ๋ชจํ˜ธํ•˜๊ณ  ์ •์˜ํ•˜๊ธฐ ์–ด๋ ค์›Œ ๋น„๊ตํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, '์„ฑ์  ์ฝ˜ํ…์ธ '์—๋Š” ์„ฑ๋ฒ”์ฃ„์˜ ์ผ๋ถ€ ํ•˜์œ„ ๋ฒ”์ฃผ๊ฐ€ ํฌํ•จ๋  ์ˆ˜ ์žˆ์ง€๋งŒ ์ฃผ๋กœ ์Œ๋ž€๋ฌผ ์ฝ˜ํ…์ธ ๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
1. 4๊ฐœ์˜ ๋ถ„๋ฅ˜๋ฒ•์€ AI ์•ˆ์ „ ๋ถ„๋ฅ˜๋ฒ•์„ ์™„์ „ํžˆ ํฌ๊ด„ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, 7๊ฐœ์˜ ๋ถ„๋ฅ˜๋ฒ•์€ ํ•œ ๊ฐ€์ง€ ๋ฒ”์ฃผ๊ฐ€ ๋ˆ„๋ฝ๋˜์–ด ์žˆ๊ณ , 2๊ฐœ์˜ ๋ถ„๋ฅ˜๋ฒ•์€ ๋‘ ๊ฐ€์ง€ ๋ฒ”์ฃผ๊ฐ€ ๋ˆ„๋ฝ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋‹จ์ผ ๋ถ„๋ฅ˜๋ฒ•์€ ๊ฐ๊ฐ ์„ธ ๊ฐ€์ง€, ๋„ค ๊ฐ€์ง€, ๋‹ค์„ฏ ๊ฐ€์ง€ ๋ฒ”์ฃผ๊ฐ€ ๋ˆ„๋ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. 2. ๊ฑฐ์˜ ๋ชจ๋“  ๋ถ„๋ฅ˜๋ฒ•์—๋Š” ์ถ”๊ฐ€ ๋ฒ”์ฃผ(์˜ˆ: ์„ฑ์ ์ธ ์ฝ˜ํ…์ธ , ์š•์„ค, ์ž˜๋ชป๋œ ์ •๋ณด)๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋‹ค์Œ ๋ฒ„์ „์˜ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ฐœ๋ฐœํ•  ๋•Œ ๊ฒ€ํ† ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. 3. ์ผ๋ถ€ ๋ถ„๋ฅ˜๋ฒ•์—๋Š” AI ์•ˆ์ „ ๋ถ„๋ฅ˜๋ฒ•์˜ ์—ฌ๋Ÿฌ ๋ฒ”์ฃผ๋ฅผ ํฌ๊ด„ํ•˜๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ๋ฒ”์ฃผ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, '๋ถˆ๋ฒ• ํ™œ๋™'๊ณผ '๋ถˆ๋ฒ• ํ–‰์œ„'๋Š” ํญ๋ ฅ ๋ฒ”์ฃ„, ๋น„ํญ๋ ฅ ๋ฒ”์ฃ„, ์„ฑ ๊ด€๋ จ ๋ฒ”์ฃ„, ๋ฌด์ฐจ๋ณ„ ๋ฌด๊ธฐ(CBRNE) ๋“ฑ ์—ฌ๋Ÿฌ ๋ฒ”์ฃผ๋ฅผ ํฌ๊ด„ํ•ฉ๋‹ˆ๋‹ค. ๊ด‘๋ฒ”์œ„ํ•œ ๋ฒ”์ฃผ์˜ ํ•œ ๊ฐ€์ง€ ํ•œ๊ณ„๋Š” ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ํ•ด์„ํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. 4. ์ž์‚ด ๋ฐ ์žํ•ด์™€ ๊ฐ™์ด '์ •์‹  ๊ฑด๊ฐ• ๋ฌธ์ œ' ๋˜๋Š” '์ผ๋ฐ˜์ ์ธ ํ•ด์•…'์„ ์‚ฌ์šฉํ•˜๋Š” ๋“ฑ ์ผ๋ถ€ ๋ฒ”์ฃผ๋Š” ๋ถˆ์™„์ „ํ•œ ์ ํ•ฉ์„ฑ์„ ๊ฐ€์ง€๋ฉฐ ์ƒ๋‹นํžˆ ๋” ๋„“๊ฑฐ๋‚˜ ์ข๊ฑฐ๋‚˜ ๋ถ€๋ถ„์ ์œผ๋กœ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค. ๋ฌธ์„œ์— ์„ค๋ช…๋œ ๋‚ด์šฉ์— ๋”ฐ๋ฅด๋ฉด ์ด๋Ÿฌํ•œ ๋น„๊ต๊ฐ€ ํ•ฉ๋ฆฌ์ ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. 5. ํ˜์˜ค ์ฝ˜ํ…์ธ ', '๊ฒฝ๋ฉธ์ ', '์‹ ์› ๊ณต๊ฒฉ' ๋“ฑ ๋งค์šฐ ์œ ์‚ฌํ•œ ๋ฒ”์ฃผ์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•œ ์šฉ์–ด๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. 6. ์ผ๋ถ€ ์šฉ์–ด๋Š” ๋ชจํ˜ธํ•˜๊ณ  ์ •์˜ํ•˜๊ธฐ ์–ด๋ ค์›Œ ๋น„๊ตํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, '์„ฑ์  ์ฝ˜ํ…์ธ '์—๋Š” ์„ฑ๋ฒ”์ฃ„์˜ ์ผ๋ถ€ ํ•˜์œ„ ๋ฒ”์ฃผ๊ฐ€ ํฌํ•จ๋  ์ˆ˜ ์žˆ์ง€๋งŒ ์ฃผ๋กœ ์Œ๋ž€๋ฌผ ์ฝ˜ํ…์ธ ๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
**Order effect for human logical reasoning.** Although the premise order does not matter in deductive reasoning, several studies show that the premise order can impact the human reasoning performance (Dekeyser et al., 2000; Girotto et al., 1997). Dekeyser et al. (2000) described _co-reference_ as a human preference of premise order; i.e., humans prefer the premises to be presented in an order where they can draw immediate conclusions after seeing each one. In this work, we show that LLMs also have such a preference, and they achieve the best performance when the ordering of rules follows the ground truth proof. Girotto et al. (1997) studied how the premise order affects logical reasoning for humans, and found that the premise order has a significant effect in solving _modus tollens_ problems (i.e., if P, then Q; not Q; therefore, not P), but not _modus ponens_ problems (i.e., if P, then Q; P; therefore, Q). However, differing from our work, they studied the influence of different ordering between rules and facts, e.g., their experiments on _modus tollens_ problems show that presenting negation statements (not Q) before rules (if P, then Q) improves the performance over the reverse order. On the other hand, our work focuses on _modus ponens_ problems that are easier for both humans and LLMs, and we show that the LLM performance is still quite sensitive to the ordering of the premises.
**Order effect for human logical reasoning.** ์—ฐ์—ญ์  ์ถ”๋ก ์—์„œ๋Š” ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ค‘์š”ํ•˜์ง€ ์•Š์ง€๋งŒ, ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ธ๊ฐ„์˜ ์ถ”๋ก  ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค(Dekeyser et al., 2000; Girotto et al., 1997). Dekeyser ๋“ฑ(2000)์€ _๊ณต๋™ ์ฐธ์กฐ_๋ฅผ ์ „์ œ ์ˆœ์„œ์— ๋Œ€ํ•œ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„, ์ฆ‰ ์ธ๊ฐ„์€ ๊ฐ ์ „์ œ๋ฅผ ๋ณธ ํ›„ ์ฆ‰๊ฐ์ ์ธ ๊ฒฐ๋ก ์„ ๋„์ถœํ•  ์ˆ˜ ์žˆ๋Š” ์ˆœ์„œ๋กœ ์ „์ œ๋ฅผ ์ œ์‹œํ•˜๋Š” ๊ฒƒ์„ ์„ ํ˜ธํ•œ๋‹ค๊ณ  ์„ค๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” LLMs๋„ ์ด๋Ÿฌํ•œ ์„ ํ˜ธ๋„๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ๊ทœ์น™์˜ ์ˆœ์„œ๊ฐ€ ์‹ค์ธก ์ฆ๋ช…์„ ๋”ฐ๋ฅผ ๋•Œ ๊ฐ€์žฅ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. Girotto ๋“ฑ(1997)์€ ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ธ๊ฐ„์˜ ๋…ผ๋ฆฌ์  ์ถ”๋ก ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ์—ฐ๊ตฌํ•œ ๊ฒฐ๊ณผ, ์ „์ œ ์ˆœ์„œ๊ฐ€ _๋ชจ๋‘์Šค ํ†จ๋ Œ์Šค_ ๋ฌธ์ œ(์ฆ‰, P์ด๋ฉด Q, Q๊ฐ€ ์•„๋‹ˆ๋ฏ€๋กœ P๊ฐ€ ์•„๋‹ˆ๋‹ค)๋ฅผ ํ‘ธ๋Š” ๋ฐ๋Š” ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€๋งŒ _๋ชจ๋‘์Šค ํฌ๋„จ์Šค_ ๋ฌธ์ œ(์ฆ‰, P์ด๋ฉด Q, P์ด๋ฏ€๋กœ Q)์—๋Š” ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ ๊ทธ๋“ค์€ ๊ทœ์น™๊ณผ ์‚ฌ์‹ค ์‚ฌ์ด์˜ ๋‹ค๋ฅธ ์ˆœ์„œ์˜ ์˜ํ–ฅ, ์˜ˆ๋ฅผ ๋“ค์–ด _๋ชจ๋‘์Šค ํ†จ๋ Œ์Šค_ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด ๋ถ€์ • ์ง„์ˆ (Q๊ฐ€ ์•„๋‹˜)์„ ๊ทœ์น™(P์ด๋ฉด Q)๋ณด๋‹ค ๋จผ์ € ์ œ์‹œํ•˜๋ฉด ๊ทธ ๋ฐ˜๋Œ€ ์ˆœ์„œ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์šฐ๋ฆฌ ์—ฐ๊ตฌ๋Š” ์ธ๊ฐ„๊ณผ LLMs ๋ชจ๋‘์—๊ฒŒ ๋” ์‰ฌ์šด _๋ชจ๋‘์Šค ํฌ๋„จ์Šค_ ๋ฌธ์ œ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, LLMs ์„ฑ๋Šฅ์ด ์ „์ œ์˜ ์ˆœ์„œ์— ์—ฌ์ „ํžˆ ๋งค์šฐ ๋ฏผ๊ฐํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
**์ธ๊ฐ„์˜ ๋…ผ๋ฆฌ์  ์ถ”๋ก ์— ๋Œ€ํ•œ ์ˆœ์„œ ํšจ๊ณผ.** ์—ฐ์—ญ์  ์ถ”๋ก ์—์„œ๋Š” ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ค‘์š”ํ•˜์ง€ ์•Š์ง€๋งŒ, ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ธ๊ฐ„์˜ ์ถ”๋ก  ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค(Dekeyser et al., 2000; Girotto et al., 1997). Dekeyser ๋“ฑ(2000)์€ _๊ณต๋™ ์ฐธ์กฐ_๋ฅผ ์ „์ œ ์ˆœ์„œ์— ๋Œ€ํ•œ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„, ์ฆ‰ ์ธ๊ฐ„์€ ๊ฐ ์ „์ œ๋ฅผ ๋ณธ ํ›„ ์ฆ‰๊ฐ์ ์ธ ๊ฒฐ๋ก ์„ ๋„์ถœํ•  ์ˆ˜ ์žˆ๋Š” ์ˆœ์„œ๋กœ ์ „์ œ๋ฅผ ์ œ์‹œํ•˜๋Š” ๊ฒƒ์„ ์„ ํ˜ธํ•œ๋‹ค๊ณ  ์„ค๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” LLM๋„ ์ด๋Ÿฌํ•œ ์„ ํ˜ธ๋„๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ๊ทœ์น™์˜ ์ˆœ์„œ๊ฐ€ ์‹ค์ธก ์ฆ๋ช…์„ ๋”ฐ๋ฅผ ๋•Œ ๊ฐ€์žฅ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. Girotto ๋“ฑ(1997)์€ ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ธ๊ฐ„์˜ ๋…ผ๋ฆฌ์  ์ถ”๋ก ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ์—ฐ๊ตฌํ•œ ๊ฒฐ๊ณผ, ์ „์ œ ์ˆœ์„œ๊ฐ€ _๋ชจ๋‘์Šค ํ†จ๋ Œ์Šค_ ๋ฌธ์ œ(์ฆ‰, P์ด๋ฉด Q, Q๊ฐ€ ์•„๋‹ˆ๋ฏ€๋กœ P๊ฐ€ ์•„๋‹ˆ๋‹ค)๋ฅผ ํ‘ธ๋Š” ๋ฐ๋Š” ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€๋งŒ _๋ชจ๋‘์Šค ํฌ๋„จ์Šค_ ๋ฌธ์ œ(์ฆ‰, P์ด๋ฉด Q, P์ด๋ฏ€๋กœ Q)์—๋Š” ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ ๊ทธ๋“ค์€ ๊ทœ์น™๊ณผ ์‚ฌ์‹ค ์‚ฌ์ด์˜ ๋‹ค๋ฅธ ์ˆœ์„œ์˜ ์˜ํ–ฅ, ์˜ˆ๋ฅผ ๋“ค์–ด _๋ชจ๋‘์Šค ํ†จ๋ Œ์Šค_ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด ๋ถ€์ • ์ง„์ˆ (Q๊ฐ€ ์•„๋‹˜)์„ ๊ทœ์น™(P์ด๋ฉด Q)๋ณด๋‹ค ๋จผ์ € ์ œ์‹œํ•˜๋ฉด ๊ทธ ๋ฐ˜๋Œ€ ์ˆœ์„œ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์šฐ๋ฆฌ ์—ฐ๊ตฌ๋Š” ์ธ๊ฐ„๊ณผ LLM ๋ชจ๋‘์—๊ฒŒ ๋” ์‰ฌ์šด _๋ชจ๋‘์Šค ํฌ๋„จ์Šค_ ๋ฌธ์ œ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, LLM ์„ฑ๋Šฅ์ด ์ „์ œ์˜ ์ˆœ์„œ์— ์—ฌ์ „ํžˆ ๋งค์šฐ ๋ฏผ๊ฐํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
Further improving quality and diversity of synthetic data.While existing methods for generating synthetic data have shown promise, there is still room for improvement in terms of creating high-quality, attributed synthetic samples that closely mimic real-world data. Future research should focus on developing new advanced techniques (or based on existing ones such as Generative Adversarial Networks (GANs) (Goodfellow et al., 2020) or Diffusion Models (Ho et al., 2020), _inter alia_) that can control and manipulate specific attributes of the generated data, enabling the creation of diverse and customizable synthetic datasets. Additionally, researchers should explore methods that can incorporate domain-specific knowledge to ensure the generated data adheres to the underlying constraints and patterns present in the target domain (e.g., via Retrieval Augmented Generation (RAG) (Borgeaud et al., 2022; Lewis et al., 2020)) while maintaining the data quality. By advancing the state-of-the-art in attributed synthetic data generation, we can unlock new opportunities for privacy-preserving analysis (Assefa et al., 2020), and model training across various fields, from healthcare (e.g., synthetic medical images (Frid-Adar et al., 2018; Wei et al., 2019)) and finance (e.g., simulated trading trajectories (Zheng et al., 2022)) to social sciences (Argyle et al., 2023; Park et al., 2023) and beyond.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋”์šฑ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ, ์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฉด๋ฐ€ํžˆ ๋ชจ๋ฐฉํ•˜๋Š” ๊ณ ํ’ˆ์งˆ์˜ ์†์„ฑ์„ ๊ฐ€์ง„ ํ•ฉ์„ฑ ์ƒ˜ํ”Œ์„ ๋งŒ๋“œ๋Š” ์ธก๋ฉด์—์„œ๋Š” ์—ฌ์ „ํžˆ ๊ฐœ์„ ์˜ ์—ฌ์ง€๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ๋Š” ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ํŠน์ • ์†์„ฑ์„ ์ œ์–ดํ•˜๊ณ  ์กฐ์ž‘ํ•  ์ˆ˜ ์žˆ๋Š” ์ƒˆ๋กœ์šด ๊ณ ๊ธ‰ ๊ธฐ์ˆ (๋˜๋Š” Generative Adversarial Networks (GANs)(Goodfellow et al., 2020) ๋˜๋Š” Diffusion Models (Ho et al., 2020) ๋“ฑ ๊ธฐ์กด ๊ธฐ์ˆ ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๊ธฐ์ˆ )์„ ๊ฐœ๋ฐœํ•˜์—ฌ ๋‹ค์–‘ํ•˜๊ณ  ์‚ฌ์šฉ์ž ์ง€์ • ๊ฐ€๋Šฅํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์—ฐ๊ตฌ์ž๋“ค์€ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์„ ์œ ์ง€ํ•˜๋ฉด์„œ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋Œ€์ƒ ๋„๋ฉ”์ธ์— ์กด์žฌํ•˜๋Š” ๊ธฐ๋ณธ์ ์ธ ์ œ์•ฝ ์กฐ๊ฑด๊ณผ ํŒจํ„ด์„ ์ค€์ˆ˜ํ•˜๋„๋ก ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ํ†ตํ•ฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค(์˜ˆ: Retrieval-augmented generation (RAG)(Borgeaud et al., 2022; Lewis et al., 2020)). ์†์„ฑ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์˜ ์ตœ์ฒจ๋‹จ ๊ธฐ์ˆ ์„ ๋ฐœ์ „์‹œํ‚ด์œผ๋กœ์จ ํ”„๋ผ์ด๋ฒ„์‹œ ๋ณดํ˜ธ ๋ถ„์„(Assefa et al., 2020)์„ ์œ„ํ•œ ์ƒˆ๋กœ์šด ๊ธฐํšŒ๋ฅผ ์—ด์–ด์ค„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์˜๋ฃŒ(์˜ˆ: ํ•ฉ์„ฑ ์˜๋ฃŒ ์ด๋ฏธ์ง€(Frid-Adar et al., 2018; Wei et al., 2019)), ๊ธˆ์œต(์˜ˆ: ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ๊ฑฐ๋ž˜ ๊ถค์ (Zheng et al., 2022)) ๋ถ„์•ผ์—์„œ๋ถ€ํ„ฐ ์‚ฌํšŒ ๊ณผํ•™(Argyle et al., 2023; Park et al., 2023)์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์—์„œ ๋ชจ๋ธ ํ•™์Šต์„ ์œ„ํ•œ ์ƒˆ๋กœ์šด ๊ธฐํšŒ๋ฅผ ์—ด์–ด์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋”์šฑ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ, ์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฉด๋ฐ€ํžˆ ๋ชจ๋ฐฉํ•˜๋Š” ๊ณ ํ’ˆ์งˆ์˜ ์†์„ฑ์„ ๊ฐ€์ง„ ํ•ฉ์„ฑ ์ƒ˜ํ”Œ์„ ๋งŒ๋“œ๋Š” ์ธก๋ฉด์—์„œ๋Š” ์—ฌ์ „ํžˆ ๊ฐœ์„ ์˜ ์—ฌ์ง€๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ๋Š” ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ํŠน์ • ์†์„ฑ์„ ์ œ์–ดํ•˜๊ณ  ์กฐ์ž‘ํ•  ์ˆ˜ ์žˆ๋Š” ์ƒˆ๋กœ์šด ๊ณ ๊ธ‰ ๊ธฐ์ˆ (๋˜๋Š” ์ƒ์„ฑ์  ์ ๋Œ€ ์‹ ๊ฒฝ๋ง(GAN)(Goodfellow et al., 2020) ๋˜๋Š” ํ™•์‚ฐ ๋ชจ๋ธ(Ho et al., 2020) ๋“ฑ ๊ธฐ์กด ๊ธฐ์ˆ ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๊ธฐ์ˆ )์„ ๊ฐœ๋ฐœํ•˜์—ฌ ๋‹ค์–‘ํ•˜๊ณ  ์‚ฌ์šฉ์ž ์ง€์ • ๊ฐ€๋Šฅํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์—ฐ๊ตฌ์ž๋“ค์€ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์„ ์œ ์ง€ํ•˜๋ฉด์„œ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋Œ€์ƒ ๋„๋ฉ”์ธ์— ์กด์žฌํ•˜๋Š” ๊ธฐ๋ณธ์ ์ธ ์ œ์•ฝ ์กฐ๊ฑด๊ณผ ํŒจํ„ด์„ ์ค€์ˆ˜ํ•˜๋„๋ก ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ํ†ตํ•ฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค(์˜ˆ: ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์ƒ์„ฑ(RAG)(Borgeaud et al., 2022; Lewis et al., 2020)). ์†์„ฑ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์˜ ์ตœ์ฒจ๋‹จ ๊ธฐ์ˆ ์„ ๋ฐœ์ „์‹œํ‚ด์œผ๋กœ์จ ํ”„๋ผ์ด๋ฒ„์‹œ ๋ณดํ˜ธ ๋ถ„์„(Assefa et al., 2020)์„ ์œ„ํ•œ ์ƒˆ๋กœ์šด ๊ธฐํšŒ๋ฅผ ์—ด์–ด์ค„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์˜๋ฃŒ(์˜ˆ: ํ•ฉ์„ฑ ์˜๋ฃŒ ์ด๋ฏธ์ง€(Frid-Adar et al., 2018; Wei et al., 2019)), ๊ธˆ์œต(์˜ˆ: ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ๊ฑฐ๋ž˜ ๊ถค์ (Zheng et al., 2022)) ๋ถ„์•ผ์—์„œ๋ถ€ํ„ฐ ์‚ฌํšŒ ๊ณผํ•™(Argyle et al., 2023; Park et al., 2023)์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์—์„œ ๋ชจ๋ธ ํ•™์Šต์„ ์œ„ํ•œ ์ƒˆ๋กœ์šด ๊ธฐํšŒ๋ฅผ ์—ด์–ด์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Back-translation augmentation.Many multilingual language models use back-translation as a data augmentation method, creating synthetic parallel training data from monolingual data sources (Bi et al., 2021; Caswell et al., 2019; Liao et al., 2021; Marie et al., 2020; Pham et al., 2021; Sennrich et al., 2016; Xu et al., 2022; Zheng et al., 2020). For example, Sennrich et al. (2016) back-translate monolingual target data into source language data, providing additional parallel training samples for substantial translation task improvements. Researchers have also explored different sampling methods for back-translation (e.g., beam search, constrained sampling, unconstrained sampling) and their comparative effectiveness (Edunov et al., 2018; Graca et al., 2019; Sennrich et al., 2016). Xu et al. (2022) emphasize the importance of the weight and quality of synthetic data for optimal NMT performance using back-translation. They propose a method to optimize the ratio between search methods and a gamma score to balance estimated importance weight and quality. However, some limitations exist with back-translation-based synthetic data generation. For example, the quality and diversity of synthetic data depends on the performance of the back-translation method. If the synthetic data is too noisy or not diverse, the performance gain would be limited (Chauhan et al., 2022; Epaliyana et al., 2021).
๋ฐฑ-translation ์ฆ๊ฐ•. ๋งŽ์€ ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ์€ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐฉ๋ฒ•์œผ๋กœ ๋ฐฑ-translation์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์ผ ์–ธ์–ด ๋ฐ์ดํ„ฐ ์†Œ์Šค์—์„œ ํ•ฉ์„ฑ ํ‰ํ–‰ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค(Bi et al., 2021; Caswell et al., 2019; Liao et al., 2021; Marie et al., 2020; Pham et al., 2021; Sennrich et al., 2016; Xu et al., 2022; Zheng et al., 2020). ์˜ˆ๋ฅผ ๋“ค์–ด, Sennrich et al. (2016)์€ ๋‹จ์ผ ์–ธ์–ด ๋Œ€์ƒ ๋ฐ์ดํ„ฐ๋ฅผ ์†Œ์Šค ์–ธ์–ด ๋ฐ์ดํ„ฐ๋กœ ๋ฐฑ-translateํ•˜์—ฌ ์ƒ๋‹นํ•œ ๋ฒˆ์—ญ ์ž‘์—… ๊ฐœ์„ ์„ ์œ„ํ•œ ์ถ”๊ฐ€ ๋ณ‘๋ ฌ ํ•™์Šต ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ž๋“ค์€ ๋ฐฑ-translation์„ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ์ƒ˜ํ”Œ๋ง ๋ฐฉ๋ฒ•(์˜ˆ: beam search, constrained sampling, unconstrained sampling)๊ณผ ๊ทธ ๋น„๊ต ํšจ๊ณผ(Edunov et al., 2018; Graca et al., 2019; Sennrich et al., 2016)๋„ ํƒ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. Xu et al. (2022)์€ ๋ฐฑ-translation์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ ์˜ NMT ์„ฑ๋Šฅ์„ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€์ค‘์น˜์™€ ํ’ˆ์งˆ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ์ถ”์ • ์ค‘์š”๋„ ๊ฐ€์ค‘์น˜์™€ ํ’ˆ์งˆ์˜ ๊ท ํ˜•์„ ๋งž์ถ”๊ธฐ ์œ„ํ•ด ๊ฒ€์ƒ‰ ๋ฐฉ๋ฒ• ๊ฐ„์˜ ๋น„์œจ๊ณผ ๊ฐ๋งˆ ์ ์ˆ˜๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฐฑ-translation ๊ธฐ๋ฐ˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์—๋Š” ๋ช‡ ๊ฐ€์ง€ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์€ ๋ฐฑ-translation ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ๋„ˆ๋ฌด ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ๊ฑฐ๋‚˜ ๋‹ค์–‘ํ•˜์ง€ ์•Š์œผ๋ฉด ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ์ œํ•œ๋ฉ๋‹ˆ๋‹ค(Chauhan et al., 2022; Epaliyana et al., 2021).
๋ฐฑ ํŠธ๋žœ์Šฌ๋ ˆ์ด์…˜ ์ฆ๊ฐ•.๋งŽ์€ ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ์€ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐฉ๋ฒ•์œผ๋กœ ๋ฐฑ ํŠธ๋žœ์Šฌ๋ ˆ์ด์…˜์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์ผ ์–ธ์–ด ๋ฐ์ดํ„ฐ ์†Œ์Šค์—์„œ ํ•ฉ์„ฑ ํ‰ํ–‰ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค(Bi et al., 2021; Caswell et al., 2019; Liao et al., 2021; Marie et al., 2020; Pham et al., 2021; Sennrich et al., 2016; Xu et al., 2022; Zheng et al., 2020). ์˜ˆ๋ฅผ ๋“ค์–ด, Sennrich et al. (2016)์€ ๋‹จ์ผ ์–ธ์–ด ๋Œ€์ƒ ๋ฐ์ดํ„ฐ๋ฅผ ์†Œ์Šค ์–ธ์–ด ๋ฐ์ดํ„ฐ๋กœ ์—ญ๋ฒˆ์—ญํ•˜์—ฌ ์ƒ๋‹นํ•œ ๋ฒˆ์—ญ ์ž‘์—… ๊ฐœ์„ ์„ ์œ„ํ•œ ์ถ”๊ฐ€ ๋ณ‘๋ ฌ ํ•™์Šต ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ž๋“ค์€ ๋ฐฑ ํŠธ๋žœ์Šฌ๋ ˆ์ด์…˜์„ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ์ƒ˜ํ”Œ๋ง ๋ฐฉ๋ฒ•(์˜ˆ: ๋น” ๊ฒ€์ƒ‰, ์ œํ•œ ์ƒ˜ํ”Œ๋ง, ์ œํ•œ๋˜์ง€ ์•Š์€ ์ƒ˜ํ”Œ๋ง)๊ณผ ๊ทธ ๋น„๊ต ํšจ๊ณผ(Edunov et al., 2018; Graca et al., 2019; Sennrich et al., 2016)๋„ ํƒ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. Xu et al. (2022)์€ ๋ฐฑ ํŠธ๋žœ์Šฌ๋ ˆ์ด์…˜์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ ์˜ NMT ์„ฑ๋Šฅ์„ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€์ค‘์น˜์™€ ํ’ˆ์งˆ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ์ถ”์ • ์ค‘์š”๋„ ๊ฐ€์ค‘์น˜์™€ ํ’ˆ์งˆ์˜ ๊ท ํ˜•์„ ๋งž์ถ”๊ธฐ ์œ„ํ•ด ๊ฒ€์ƒ‰ ๋ฐฉ๋ฒ• ๊ฐ„์˜ ๋น„์œจ๊ณผ ๊ฐ๋งˆ ์ ์ˆ˜๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฐฑ ํŠธ๋žœ์Šฌ๋ ˆ์ด์…˜ ๊ธฐ๋ฐ˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์—๋Š” ๋ช‡ ๊ฐ€์ง€ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์€ ๋ฐฑ ํŠธ๋žœ์Šฌ๋ ˆ์ด์…˜ ๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ๋„ˆ๋ฌด ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ๊ฑฐ๋‚˜ ๋‹ค์–‘ํ•˜์ง€ ์•Š์œผ๋ฉด ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ์ œํ•œ๋ฉ๋‹ˆ๋‹ค(Chauhan et al., 2022; Epaliyana et al., 2021).
When we aim to enhance the capabilities of a LLM in a specific language, vocabulary extension is an intuitively reasonable approach. In this section, we evaluate the impact of vocabulary extension through the LLM-Eval benchmark, and the experimental results are presented in table 1. Initially, we collected one million Chinese sentences from the internet (approximately 0.5 billion tokens) and further pretrain the original LLaMA without vocabulary extension. Surprisingly, we find that this model significantly outperform the vocabulary-extended Chinese LLaMA, across settings of 1K, 5K, and 950K instruction tuning. This discovery is thought-privoking, given that the Chinese LLaMA underwent further Chinese pretraining on 30 billion tokens, a much larger volume than our 0.5 billion tokens. Moreover, within the 950K setting, we include results from extending the vocabulary on original LLaMA and training it with the same 0.5 billion tokens, to mitigate the influence of training data discrepancy. The outcomes remain consistent. This indicates that vocabulary extension is not a favorable choice within training scales of tens of billions of tokens. While we don't negate the efficacy of vocabulary extension in settings involving larger-scale pretraining (such as trillions of tokens), as reported in other literatures [23], this already leans more towards retraining than mere language transfer.
ํŠน์ • ์–ธ์–ด์—์„œ LLM์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•  ๋•Œ ์–ดํœ˜ ํ™•์žฅ์€ ์ง๊ด€์ ์œผ๋กœ ํ•ฉ๋ฆฌ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” LLM-Eval ๋ฒค์น˜๋งˆํฌ๋ฅผ ํ†ตํ•ด ์–ดํœ˜ ํ™•์žฅ์˜ ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 1์— ์ œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ € ์ธํ„ฐ๋„ท์—์„œ ํ•œ ๋ฐฑ๋งŒ ๊ฐœ์˜ ์ค‘๊ตญ์–ด ๋ฌธ์žฅ(์•ฝ 0.50์–ต ํ† ํฐ)์„ ์ˆ˜์ง‘ํ•˜์—ฌ ์–ดํœ˜ ํ™•์žฅ ์—†์ด ์›๋ž˜ LLaMA๋ฅผ ์ถ”๊ฐ€๋กœ ์‚ฌ์ „ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋†€๋ž๊ฒŒ๋„ ์ด ๋ชจ๋ธ์€ 1K, 5K, 950K Instruction Tuning ์„ค์ •์—์„œ ์–ดํœ˜๊ฐ€ ํ™•์žฅ๋œ ์ค‘๊ตญ์–ด LLaMA๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐœ๊ฒฌ์€ ์ค‘๊ตญ์–ด LLaMA๊ฐ€ 0.5์–ต ํ† ํฐ๋ณด๋‹ค ํ›จ์”ฌ ๋งŽ์€ ์–‘์ธ 300์–ต ํ† ํฐ์— ๋Œ€ํ•ด ์ถ”๊ฐ€ ์ค‘๊ตญ์–ด ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์ณค๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•  ๋•Œ ์ƒ๊ฐ์„ ์ž๊ทนํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ 950K ์„ค์ •์—์„œ๋Š” ์›๋ž˜ LLaMA์˜ ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜๊ณ  ๋™์ผํ•œ 0.5์–ต ํ† ํฐ์œผ๋กœ ํ•™์Šต์‹œ์ผœ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ๋ถˆ์ผ์น˜์˜ ์˜ํ–ฅ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๊ฒฐ๊ณผ๋ฅผ ํฌํ•จํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ๋Š” ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ˆ˜๋ฐฑ์–ต ํ† ํฐ ๊ทœ๋ชจ์˜ ํ›ˆ๋ จ์—์„œ๋Š” ์–ดํœ˜ ํ™•์žฅ์ด ๋ฐ”๋žŒ์งํ•œ ์„ ํƒ์ด ์•„๋‹˜์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ฌธํ—Œ[23]์—์„œ ๋ณด๊ณ ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ๋” ํฐ ๊ทœ๋ชจ์˜ ์‚ฌ์ „ ํ•™์Šต(์˜ˆ: ์ˆ˜์กฐ ํ† ํฐ)์ด ํฌํ•จ๋œ ์„ค์ •์—์„œ๋Š” ์–ดํœ˜ ํ™•์žฅ์˜ ํšจ๊ณผ๋ฅผ ๋ถ€์ •ํ•˜์ง€ ์•Š์ง€๋งŒ, ์ด๋Š” ์ด๋ฏธ ๋‹จ์ˆœํ•œ ์–ธ์–ด ์ „๋‹ฌ๋ณด๋‹ค๋Š” ์žฌ๊ต์œก์— ๋” ๊ฐ€๊น์Šต๋‹ˆ๋‹ค.
ํŠน์ • ์–ธ์–ด์—์„œ LLM์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•  ๋•Œ ์–ดํœ˜ ํ™•์žฅ์€ ์ง๊ด€์ ์œผ๋กœ ํ•ฉ๋ฆฌ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” LLM-Eval ๋ฒค์น˜๋งˆํฌ๋ฅผ ํ†ตํ•ด ์–ดํœ˜ ํ™•์žฅ์˜ ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 1์— ์ œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ € ์ธํ„ฐ๋„ท์—์„œ ํ•œ ๋ฐฑ๋งŒ ๊ฐœ์˜ ์ค‘๊ตญ์–ด ๋ฌธ์žฅ(์•ฝ 0.50์–ต ํ† ํฐ)์„ ์ˆ˜์ง‘ํ•˜์—ฌ ์–ดํœ˜ ํ™•์žฅ ์—†์ด ์›๋ž˜ LLaMA๋ฅผ ์ถ”๊ฐ€๋กœ ์‚ฌ์ „ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋†€๋ž๊ฒŒ๋„ ์ด ๋ชจ๋ธ์€ 1K, 5K, 950K ๋ช…๋ น์–ด ํŠœ๋‹ ์„ค์ •์—์„œ ์–ดํœ˜๊ฐ€ ํ™•์žฅ๋œ ์ค‘๊ตญ์–ด LLaMA๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐœ๊ฒฌ์€ ์ค‘๊ตญ์–ด LLaMA๊ฐ€ 0.5์–ต ํ† ํฐ๋ณด๋‹ค ํ›จ์”ฌ ๋งŽ์€ ์–‘์ธ 300์–ต ํ† ํฐ์— ๋Œ€ํ•ด ์ถ”๊ฐ€ ์ค‘๊ตญ์–ด ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์ณค๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•  ๋•Œ ์ƒ๊ฐ์„ ์ž๊ทนํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ 950K ์„ค์ •์—์„œ๋Š” ์›๋ž˜ LLaMA์˜ ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜๊ณ  ๋™์ผํ•œ 0.5์–ต ํ† ํฐ์œผ๋กœ ํ•™์Šต์‹œ์ผœ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ๋ถˆ์ผ์น˜์˜ ์˜ํ–ฅ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๊ฒฐ๊ณผ๋ฅผ ํฌํ•จํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ๋Š” ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ˆ˜๋ฐฑ์–ต ํ† ํฐ ๊ทœ๋ชจ์˜ ํ›ˆ๋ จ์—์„œ๋Š” ์–ดํœ˜ ํ™•์žฅ์ด ๋ฐ”๋žŒ์งํ•œ ์„ ํƒ์ด ์•„๋‹˜์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ฌธํ—Œ[23]์—์„œ ๋ณด๊ณ ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ๋” ํฐ ๊ทœ๋ชจ์˜ ์‚ฌ์ „ ํ•™์Šต(์˜ˆ: ์ˆ˜์กฐ ํ† ํฐ)์ด ํฌํ•จ๋œ ์„ค์ •์—์„œ๋Š” ์–ดํœ˜ ํ™•์žฅ์˜ ํšจ๊ณผ๋ฅผ ๋ถ€์ •ํ•˜์ง€ ์•Š์ง€๋งŒ, ์ด๋Š” ์ด๋ฏธ ๋‹จ์ˆœํ•œ ์–ธ์–ด ์ „๋‹ฌ๋ณด๋‹ค๋Š” ์žฌ๊ต์œก์— ๋” ๊ฐ€๊น์Šต๋‹ˆ๋‹ค.
In terms of fine-tuning, whether through Supervised Fine Tuning (SFT) or Reinforcement Learning (RL), the primary aim is to gather high-quality supervisory signals for model improvement. For SFT, the approach generally involves leveraging larger models or manual efforts to collect high-quality data (An et al., 2023; Yu et al., 2023; Yue et al., 2023; Luo et al., 2023). MetaMath(Yu et al., 2023; Gou et al., 2023) augments its dataset by rewriting mathematical queries and provide answers with the help of an LLM. It also leverages backward reasoning to enhance the model's capabilities. MAmmoTH(Yue et al., 2023) compiles diverse mathematical datasets for training, enriching them with Chains of Thought (CoT) and Programs of Thought (PoT) rationales for reasoning. RL-based methods strive to gather superior training signals through self-sampling and reward supervision. Lightman et al. (2023) proposes to use process supervision to give more precise signals than outcome supervision. Some strategies adopt the process supervision into the online RL training (Wang et al., 2023b; Xi et al., 2024; Luo et al., 2023). RFT(Yuan et al., 2023) is a simplified offline method of standard online RL. It enhances its training dataset with accurate reasoning paths by generating and collecting them via a SFT model. This approach is further extended by ReST (Singh et al., 2023) through an iterative process.
๋ฏธ์„ธ ์กฐ์ •์˜ ๊ฒฝ์šฐ, Supervised Fine-tuning (SFT) ๋˜๋Š” Reinforcement Learning (RL)์„ ํ†ตํ•ด ๋ชจ๋ธ ๊ฐœ์„ ์„ ์œ„ํ•œ ๊ณ ํ’ˆ์งˆ ๊ฐ๋… ์‹ ํ˜ธ๋ฅผ ์ˆ˜์ง‘ํ•˜๋Š” ๊ฒƒ์ด ์ฃผ์š” ๋ชฉํ‘œ์ž…๋‹ˆ๋‹ค. SFT์˜ ๊ฒฝ์šฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ๋” ํฐ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜๊ฑฐ๋‚˜ ์ˆ˜๋™์œผ๋กœ ๋…ธ๋ ฅํ•˜์—ฌ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐฉ์‹์„ ์ทจํ•ฉ๋‹ˆ๋‹ค(An et al., 2023; Yu et al., 2023; Yue et al., 2023; Luo et al., 2023). MetaMath(Yu et al., 2023; Gou et al., 2023)๋Š” ์ˆ˜ํ•™์  ์ฟผ๋ฆฌ๋ฅผ ๋‹ค์‹œ ์ž‘์„ฑํ•˜์—ฌ LLM์˜ ๋„์›€์„ ๋ฐ›์•„ ๋‹ต์„ ์ œ๊ณตํ•จ์œผ๋กœ์จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ญ ์ถ”๋ก ์„ ํ™œ์šฉํ•˜์—ฌ ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. MAmmoTH(Yue et al., 2023)๋Š” ํ•™์Šต์„ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ์ˆ˜ํ•™ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ปดํŒŒ์ผํ•˜๊ณ  ์ถ”๋ก ์„ ์œ„ํ•œ Chains of Thought (CoT)๊ณผ Programs of Thought (PoT) ๊ทผ๊ฑฐ๋กœ ์ด๋ฅผ ๊ฐ•ํ™”ํ•ฉ๋‹ˆ๋‹ค. RL ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์€ ์ž์ฒด ์ƒ˜ํ”Œ๋ง๊ณผ ๋ณด์ƒ ๊ฐ๋…์„ ํ†ตํ•ด ๋” ์šฐ์ˆ˜ํ•œ ํ›ˆ๋ จ ์‹ ํ˜ธ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ธฐ ์œ„ํ•ด ๋…ธ๋ ฅํ•ฉ๋‹ˆ๋‹ค. Lightman ๋“ฑ(2023)์€ ๊ฒฐ๊ณผ ๊ฐ๋…๋ณด๋‹ค ๋” ์ •ํ™•ํ•œ ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋กœ์„ธ์Šค ๊ฐ๋…์„ ์‚ฌ์šฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ถ€ ์ „๋žต์—์„œ๋Š” ์˜จ๋ผ์ธ RL ํ›ˆ๋ จ์— ํ”„๋กœ์„ธ์Šค ๊ฐ๋…์„ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค(Wang et al., 2023b; Xi et al., 2024; Luo et al., 2023). RFT(Yuan et al., 2023)๋Š” ํ‘œ์ค€ ์˜จ๋ผ์ธ RL์˜ ๊ฐ„์†Œํ™”๋œ ์˜คํ”„๋ผ์ธ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ SFT ๋ชจ๋ธ์„ ํ†ตํ•ด ์ƒ์„ฑํ•˜๊ณ  ์ˆ˜์ง‘ํ•œ ์ •ํ™•ํ•œ ์ถ”๋ก  ๊ฒฝ๋กœ๋กœ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋ฐ˜๋ณต ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ด ReST(Singh et al., 2023)์— ์˜ํ•ด ๋”์šฑ ํ™•์žฅ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
๋ฏธ์„ธ ์กฐ์ •์˜ ๊ฒฝ์šฐ, ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •(SFT) ๋˜๋Š” ๊ฐ•ํ™” ํ•™์Šต(RL)์„ ํ†ตํ•ด ๋ชจ๋ธ ๊ฐœ์„ ์„ ์œ„ํ•œ ๊ณ ํ’ˆ์งˆ ๊ฐ๋… ์‹ ํ˜ธ๋ฅผ ์ˆ˜์ง‘ํ•˜๋Š” ๊ฒƒ์ด ์ฃผ์š” ๋ชฉํ‘œ์ž…๋‹ˆ๋‹ค. SFT์˜ ๊ฒฝ์šฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ๋” ํฐ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜๊ฑฐ๋‚˜ ์ˆ˜๋™์œผ๋กœ ๋…ธ๋ ฅํ•˜์—ฌ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐฉ์‹์„ ์ทจํ•ฉ๋‹ˆ๋‹ค(An et al., 2023; Yu et al., 2023; Yue et al., 2023; Luo et al., 2023). MetaMath(Yu et al., 2023; Gou et al., 2023)๋Š” ์ˆ˜ํ•™์  ์ฟผ๋ฆฌ๋ฅผ ๋‹ค์‹œ ์ž‘์„ฑํ•˜์—ฌ LLM์˜ ๋„์›€์„ ๋ฐ›์•„ ๋‹ต์„ ์ œ๊ณตํ•จ์œผ๋กœ์จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ญ ์ถ”๋ก ์„ ํ™œ์šฉํ•˜์—ฌ ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. MAmmoTH(Yue et al., 2023)๋Š” ํ•™์Šต์„ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ์ˆ˜ํ•™ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ปดํŒŒ์ผํ•˜๊ณ  ์ถ”๋ก ์„ ์œ„ํ•œ ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ(CoT)๊ณผ ์ƒ๊ฐ์˜ ํ”„๋กœ๊ทธ๋žจ(PoT) ๊ทผ๊ฑฐ๋กœ ์ด๋ฅผ ๊ฐ•ํ™”ํ•ฉ๋‹ˆ๋‹ค. RL ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์€ ์ž์ฒด ์ƒ˜ํ”Œ๋ง๊ณผ ๋ณด์ƒ ๊ฐ๋…์„ ํ†ตํ•ด ๋” ์šฐ์ˆ˜ํ•œ ํ›ˆ๋ จ ์‹ ํ˜ธ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ธฐ ์œ„ํ•ด ๋…ธ๋ ฅํ•ฉ๋‹ˆ๋‹ค. Lightman ๋“ฑ(2023)์€ ๊ฒฐ๊ณผ ๊ฐ๋…๋ณด๋‹ค ๋” ์ •ํ™•ํ•œ ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋กœ์„ธ์Šค ๊ฐ๋…์„ ์‚ฌ์šฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ถ€ ์ „๋žต์—์„œ๋Š” ์˜จ๋ผ์ธ RL ํ›ˆ๋ จ์— ํ”„๋กœ์„ธ์Šค ๊ฐ๋…์„ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค(Wang et al., 2023b; Xi et al., 2024; Luo et al., 2023). RFT(Yuan et al., 2023)๋Š” ํ‘œ์ค€ ์˜จ๋ผ์ธ RL์˜ ๊ฐ„์†Œํ™”๋œ ์˜คํ”„๋ผ์ธ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ SFT ๋ชจ๋ธ์„ ํ†ตํ•ด ์ƒ์„ฑํ•˜๊ณ  ์ˆ˜์ง‘ํ•œ ์ •ํ™•ํ•œ ์ถ”๋ก  ๊ฒฝ๋กœ๋กœ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋ฐ˜๋ณต ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ด ReST(Singh et al., 2023)์— ์˜ํ•ด ๋”์šฑ ํ™•์žฅ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
Several existing works evaluate the impact of pre-training and model scale on continual learning. Cossu et al. (2022) investigate pre-training scenarios for language and vision. They find that unsupervised and self-supervised pre-training plays a fundamental role in mitigating forgetting, while supervision hurts performance. Similarly, Mehta et al. (2023) find that pre-trained models forget less than randomly initialized models, due to their weights lying in flatter regions of the loss landscape. They also find that larger models forget less which is connected to the findings of Ramasesh et al. (2022); Mirzadeh et al. (2022). The former finds that pre-trained models forget less as they are scaled up, suggesting that it may be due to the hidden representations growing more orthogonal with scale. The latter finds that wider neural networks forget less compared to their parameter-equivalent deeper counterparts. Hernandez et al. (2021) establish scaling laws for transfer: equations that can predict the performance of a neural network on a new task as a function of its parameter count and pre-training dataset size. The authors find that this positive transfer consistently improves as the parameter count increases. Finally, Scialom et al. (2022) show that autoregressive LLMs have a strong ability to learn continually which they hypothesize is related to their pre-training objective.
์‚ฌ์ „ ํ•™์Šต๊ณผ ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ์ง€์† ํ•™์Šต์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ธฐ์กด ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๊ฐ€ ์—ฌ๋Ÿฌ ๊ฐœ ์žˆ์Šต๋‹ˆ๋‹ค. Cossu ๋“ฑ(2022)์€ ์–ธ์–ด์™€ ๋น„์ „์— ๋Œ€ํ•œ pre-training ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋น„์ง€๋„ ๋ฐ self-supervised pre-training์ด ์žŠ์–ด๋ฒ„๋ฆผ์„ ์™„ํ™”ํ•˜๋Š” ๋ฐ ๊ทผ๋ณธ์ ์ธ ์—ญํ• ์„ ํ•˜๋Š” ๋ฐ˜๋ฉด, ์ง€๋„๋Š” ์„ฑ๋Šฅ์— ํ•ด๋ฅผ ๋ผ์นœ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Mehta ๋“ฑ(2023)์€ pre-trained ๋ชจ๋ธ์€ ๋ฌด์ž‘์œ„๋กœ ์ดˆ๊ธฐํ™”๋œ ๋ชจ๋ธ๋ณด๋‹ค ๊ฐ€์ค‘์น˜๊ฐ€ loss landscape์˜ ๋” ํ‰ํ‰ํ•œ ์˜์—ญ์— ์œ„์น˜ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋œ ์žŠ์–ด๋ฒ„๋ฆฐ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋” ํฐ ๋ชจ๋ธ์ด ๋œ ์žŠ์–ด๋ฒ„๋ฆฐ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ๋Š”๋ฐ, ์ด๋Š” Ramasesh ๋“ฑ(2022)๊ณผ Mirzadeh ๋“ฑ(2022)์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์™€๋„ ์—ฐ๊ฒฐ๋ฉ๋‹ˆ๋‹ค. ์ „์ž๋Š” pre-trained ๋ชจ๋ธ์ด ๊ทœ๋ชจ๊ฐ€ ์ปค์ง์— ๋”ฐ๋ผ ๋œ ์žŠ์–ด๋ฒ„๋ฆฌ๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋Š”๋ฐ, ์ด๋Š” ์ˆจ๊ฒจ์ง„ ํ‘œํ˜„์ด ๊ทœ๋ชจ์— ๋”ฐ๋ผ ๋” ์ง๊ตํ•˜๊ฒŒ ์ปค์ง€๋Š” ๊ฒƒ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํ›„์ž๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ์™€ ๋™์ผํ•œ ๋” ๊นŠ์€ ๋ชจ๋ธ์— ๋น„ํ•ด ๋” ๋„“์€ ์‹ ๊ฒฝ๋ง์ด ๋œ ์žŠ์–ด๋ฒ„๋ฆฌ๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. Hernandez ๋“ฑ(2021)์€ scaling laws for transfer, ์ฆ‰ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜์™€ pre-training ๋ฐ์ดํ„ฐ ์„ธํŠธ ํฌ๊ธฐ์˜ ํ•จ์ˆ˜๋กœ์„œ ์ƒˆ๋กœ์šด ์ž‘์—…์—์„œ ์‹ ๊ฒฝ๋ง์˜ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ์ •์‹์„ ํ™•๋ฆฝํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €์ž๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์ด๋Ÿฌํ•œ ์–‘์˜ ์ „์ด๊ฐ€ ์ง€์†์ ์œผ๋กœ ๊ฐœ์„ ๋˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, Scialom ๋“ฑ(2022)์€ autoregressive LLM์ด ์ง€์†์ ์œผ๋กœ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ•๋ ฅํ•œ ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Š” pre-training objective์™€ ๊ด€๋ จ์ด ์žˆ๋‹ค๊ณ  ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค.
์‚ฌ์ „ ํ•™์Šต๊ณผ ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ์ง€์† ํ•™์Šต์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ธฐ์กด ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๊ฐ€ ์—ฌ๋Ÿฌ ๊ฐœ ์žˆ์Šต๋‹ˆ๋‹ค. Cossu ๋“ฑ(2022)์€ ์–ธ์–ด์™€ ๋น„์ „์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋น„์ง€๋„ ๋ฐ ์ž๊ธฐ ์ง€๋„ ์‚ฌ์ „ ํ•™์Šต์ด ์žŠ์–ด๋ฒ„๋ฆผ์„ ์™„ํ™”ํ•˜๋Š” ๋ฐ ๊ทผ๋ณธ์ ์ธ ์—ญํ• ์„ ํ•˜๋Š” ๋ฐ˜๋ฉด, ์ง€๋„๋Š” ์„ฑ๋Šฅ์— ํ•ด๋ฅผ ๋ผ์นœ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Mehta ๋“ฑ(2023)์€ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์€ ๋ฌด์ž‘์œ„๋กœ ์ดˆ๊ธฐํ™”๋œ ๋ชจ๋ธ๋ณด๋‹ค ๊ฐ€์ค‘์น˜๊ฐ€ ์†์‹ค ์กฐ๋ง์˜ ๋” ํ‰ํ‰ํ•œ ์˜์—ญ์— ์œ„์น˜ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋œ ์žŠ์–ด๋ฒ„๋ฆฐ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋” ํฐ ๋ชจ๋ธ์ด ๋œ ์žŠ์–ด๋ฒ„๋ฆฐ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ๋Š”๋ฐ, ์ด๋Š” Ramasesh ๋“ฑ(2022)๊ณผ Mirzadeh ๋“ฑ(2022)์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์™€๋„ ์—ฐ๊ฒฐ๋ฉ๋‹ˆ๋‹ค. ์ „์ž๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ๊ทœ๋ชจ๊ฐ€ ์ปค์ง์— ๋”ฐ๋ผ ๋œ ์žŠ์–ด๋ฒ„๋ฆฌ๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋Š”๋ฐ, ์ด๋Š” ์ˆจ๊ฒจ์ง„ ํ‘œํ˜„์ด ๊ทœ๋ชจ์— ๋”ฐ๋ผ ๋” ์ง๊ตํ•˜๊ฒŒ ์ปค์ง€๋Š” ๊ฒƒ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํ›„์ž๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ์™€ ๋™์ผํ•œ ๋” ๊นŠ์€ ๋ชจ๋ธ์— ๋น„ํ•ด ๋” ๋„“์€ ์‹ ๊ฒฝ๋ง์ด ๋œ ์žŠ์–ด๋ฒ„๋ฆฌ๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. Hernandez ๋“ฑ(2021)์€ ์ „์ด์— ๋Œ€ํ•œ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™, ์ฆ‰ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜์™€ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ ํฌ๊ธฐ์˜ ํ•จ์ˆ˜๋กœ์„œ ์ƒˆ๋กœ์šด ์ž‘์—…์—์„œ ์‹ ๊ฒฝ๋ง์˜ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ์ •์‹์„ ํ™•๋ฆฝํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €์ž๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์ด๋Ÿฌํ•œ ์–‘์˜ ์ „์ด๊ฐ€ ์ง€์†์ ์œผ๋กœ ๊ฐœ์„ ๋˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, Scialom ๋“ฑ(2022)์€ ์ž๋™ ํšŒ๊ท€ LLM์ด ์ง€์†์ ์œผ๋กœ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ•๋ ฅํ•œ ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์‚ฌ์ „ ํ•™์Šต ๋ชฉํ‘œ์™€ ๊ด€๋ จ์ด ์žˆ๋‹ค๊ณ  ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค.
**Discussion on Effect of Data Quality.** For pre-training, the quality of pre-training data is vital to the model capacities of LLMs. Existing work has shown that pre-training on the low-quality corpus, such as noisy, toxic, and duplicate data, would largely hurt the performance of models [64, 214, 216, 219]. Recent studies, such as T5 [82], GLaM [112], and Gopher [64], have investigated the influence of data quality on the LLMs' capacities. By comparing the performance of models trained on the filtered and unfiltered corpus, they have reached the similar conclusion that pre-training LLMs on cleaned data can improve the model performance. More specifically, the duplication of data may result in "_double descent_" (referring to the phenomenon of performance initially deteriorating and subsequently improving) [214, 228], or even overwhelm the training process [214]. In addition, it has been shown that duplicate data degrades the ability of LLMs to copy from the context, which might further affect the generalization capacity of LLMs using in-context learning [214]. Therefore, as suggested in [56, 64, 78, 212], it is essential to utilize preprocessing methods like quality filtering, toxic filtering and deduplication to carefully clean the pre-training corpus (as illustrated in Section 4.1.2), to improve stability of the training process and avoid affecting the model performance.
**๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์˜ ์˜ํ–ฅ์— ๋Œ€ํ•œ ๋…ผ์˜**. ์‚ฌ์ „ ํ•™์Šต์˜ ๊ฒฝ์šฐ, ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์€ LLM์˜ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๋…ธ์ด์ฆˆ, ๋…์„ฑ, ์ค‘๋ณต ๋ฐ์ดํ„ฐ ๋“ฑ ํ’ˆ์งˆ์ด ๋‚ฎ์€ ์ฝ”ํผ์Šค์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šตํ•˜๋ฉด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ์ €ํ•˜๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค[64, 214, 216, 219]. T5[82], GLaM[112], Gopher[64] ๋“ฑ ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์ด LLM์˜ ์—ญ๋Ÿ‰์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์กฐ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•„ํ„ฐ๋ง๋œ ์ฝ”ํผ์Šค์™€ ํ•„ํ„ฐ๋ง๋˜์ง€ ์•Š์€ ์ฝ”ํผ์Šค์—์„œ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ, ์ •๋ฆฌ๋œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด LLM์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋ฉด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ์œ ์‚ฌํ•œ ๊ฒฐ๋ก ์— ๋„๋‹ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ณด๋‹ค ๊ตฌ์ฒด์ ์œผ๋กœ, ๋ฐ์ดํ„ฐ์˜ ์ค‘๋ณต์œผ๋กœ ์ธํ•ด ์„ฑ๋Šฅ์ด ์ฒ˜์Œ์—๋Š” ์•…ํ™”๋˜์—ˆ๋‹ค๊ฐ€ ๊ฐœ์„ ๋˜๋Š” ํ˜„์ƒ('_double descent_'์ด๋ผ๊ณ  ํ•จ)์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ[214, 228], ์‹ฌ์ง€์–ด ํ•™์Šต ๊ณผ์ •์„ ์••๋„ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค[214]. ๋˜ํ•œ ์ค‘๋ณต ๋ฐ์ดํ„ฐ๋Š” LLM์ด ๋ฌธ๋งฅ์—์„œ ๋ณต์‚ฌํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์ €ํ•˜์‹œ์ผœ In-Context Learning์— ์‚ฌ์šฉ๋˜๋Š” LLM์˜ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[214]. ๋”ฐ๋ผ์„œ [56, 64, 78, 212]์—์„œ ์ œ์•ˆํ•œ ๋ฐ”์™€ ๊ฐ™์ด ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ์ •๋ฆฌํ•˜์—ฌ ํ•™์Šต ๊ณผ์ •์˜ ์•ˆ์ •์„ฑ์„ ๊ฐœ์„ ํ•˜๊ณ  ๋ชจ๋ธ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ์ฃผ์ง€ ์•Š๋„๋ก ํ’ˆ์งˆ ํ•„ํ„ฐ๋ง, ๋…์„ฑ ํ•„ํ„ฐ๋ง ๋ฐ ์ค‘๋ณต ์ œ๊ฑฐ์™€ ๊ฐ™์€ ์ „์ฒ˜๋ฆฌ ๋ฐฉ๋ฒ•์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ด ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค(4.1.2์ ˆ์—์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด).
**๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์˜ ์˜ํ–ฅ์— ๋Œ€ํ•œ ๋…ผ์˜**. ์‚ฌ์ „ ํ•™์Šต์˜ ๊ฒฝ์šฐ, ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์€ LLM์˜ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๋…ธ์ด์ฆˆ, ๋…์„ฑ, ์ค‘๋ณต ๋ฐ์ดํ„ฐ ๋“ฑ ํ’ˆ์งˆ์ด ๋‚ฎ์€ ์ฝ”ํผ์Šค์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šตํ•˜๋ฉด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ์ €ํ•˜๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค[64, 214, 216, 219]. T5[82], GLaM[112], Gopher[64] ๋“ฑ ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์ด LLM์˜ ์—ญ๋Ÿ‰์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์กฐ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•„ํ„ฐ๋ง๋œ ์ฝ”ํผ์Šค์™€ ํ•„ํ„ฐ๋ง๋˜์ง€ ์•Š์€ ์ฝ”ํผ์Šค์—์„œ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ, ์ •๋ฆฌ๋œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด LLM์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋ฉด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ์œ ์‚ฌํ•œ ๊ฒฐ๋ก ์— ๋„๋‹ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ณด๋‹ค ๊ตฌ์ฒด์ ์œผ๋กœ, ๋ฐ์ดํ„ฐ์˜ ์ค‘๋ณต์œผ๋กœ ์ธํ•ด ์„ฑ๋Šฅ์ด ์ฒ˜์Œ์—๋Š” ์•…ํ™”๋˜์—ˆ๋‹ค๊ฐ€ ๊ฐœ์„ ๋˜๋Š” ํ˜„์ƒ('_์ด์ค‘ ํ•˜๊ฐ•'์ด๋ผ๊ณ  ํ•จ)์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ[214, 228], ์‹ฌ์ง€์–ด ํ•™์Šต ๊ณผ์ •์„ ์••๋„ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค[214]. ๋˜ํ•œ ์ค‘๋ณต ๋ฐ์ดํ„ฐ๋Š” LLM์ด ๋ฌธ๋งฅ์—์„œ ๋ณต์‚ฌํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์ €ํ•˜์‹œ์ผœ ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต์— ์‚ฌ์šฉ๋˜๋Š” LLM์˜ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[214]. ๋”ฐ๋ผ์„œ [56, 64, 78, 212]์—์„œ ์ œ์•ˆํ•œ ๋ฐ”์™€ ๊ฐ™์ด ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ์ •๋ฆฌํ•˜์—ฌ ํ•™์Šต ๊ณผ์ •์˜ ์•ˆ์ •์„ฑ์„ ๊ฐœ์„ ํ•˜๊ณ  ๋ชจ๋ธ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ์ฃผ์ง€ ์•Š๋„๋ก ํ’ˆ์งˆ ํ•„ํ„ฐ๋ง, ๋…์„ฑ ํ•„ํ„ฐ๋ง ๋ฐ ์ค‘๋ณต ์ œ๊ฑฐ์™€ ๊ฐ™์€ ์ „์ฒ˜๋ฆฌ ๋ฐฉ๋ฒ•์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ด ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค(4.1.2์ ˆ์—์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด).
While we observe consistent _average_ perplexity improvements, Section A.3 demonstrates that this perplexity improvement varies greatly across validation sets. More importantly, data selection always impairs performance on web snapshot validation sets such as CC-dedup, CommonCrawl, and C4. To investigate why this occurs, we embed each validation set into the same embedding space as the training set and search for the nearest neighbors to validation points in the training set for our 1.3B baseline model. In the left plot of Figure 5, we show that validation sets drawn from the same distribution as web-snapshots are closer to training set compared to other validation sets, while the right plot of Figure 5 shows that data selection disproportionately affects these web-snapshot validation sets: on the top-right plot, we see that web validation sets reside in regions of the embedding space which are sparsified as a result of data selection (e.g. regions of space close to cluster centroids in the training set), and in the bottom-right plot we see that these points are also the most affected by data selection, since their perplexity after data selection significantly increases. Moreover, the middle-right plot shows that these validation points have the lowest perplexity before pruning indicating that these points are "easy" points, perhaps due to their proximity to the training set.
์ผ๊ด€๋œ _ํ‰๊ท _ perplexity ๊ฐœ์„ ์„ ๊ด€์ฐฐํ•˜๋Š” ๋ฐ˜๋ฉด, ์„น์…˜ A.3์—์„œ๋Š” ์ด๋Ÿฌํ•œ perplexity ๊ฐœ์„ ์ด ๊ฒ€์ฆ ์„ธํŠธ๋งˆ๋‹ค ํฌ๊ฒŒ ๋‹ค๋ฅด๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋” ์ค‘์š”ํ•œ ๊ฒƒ์€ ๋ฐ์ดํ„ฐ ์„ ํƒ์ด ํ•ญ์ƒ CC-dedup, CommonCrawl ๋ฐ C4์™€ ๊ฐ™์€ ์›น ์Šค๋ƒ…์ƒท ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์˜ ์„ฑ๋Šฅ์„ ์†์ƒ์‹œํ‚จ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ˜„์ƒ์ด ๋ฐœ์ƒํ•˜๋Š” ์ด์œ ๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋ฅผ ํ›ˆ๋ จ ์„ธํŠธ์™€ ๋™์ผํ•œ embedding ๊ณต๊ฐ„์— ํฌํ•จ์‹œํ‚ค๊ณ  1.3B baseline ๋ชจ๋ธ์— ๋Œ€ํ•ด ํ›ˆ๋ จ ์„ธํŠธ์—์„œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง€์ ์— ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ด์›ƒ์„ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 5์˜ ์™ผ์ชฝ plot์—์„œ๋Š” ์›น ์Šค๋ƒ…์ƒท๊ณผ ๋™์ผํ•œ ๋ถ„ํฌ์—์„œ ์ถ”์ถœํ•œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๊ฐ€ ๋‹ค๋ฅธ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์— ๋น„ํ•ด ํ›ˆ๋ จ ์„ธํŠธ์— ๋” ๊ฐ€๊น๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ๊ณ , ๊ทธ๋ฆผ 5์˜ ์˜ค๋ฅธ์ชฝ plot์—์„œ๋Š” ๋ฐ์ดํ„ฐ ์„ ํƒ์ด ์ด๋Ÿฌํ•œ ์›น ์Šค๋ƒ…์ƒท ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์— ๋ถˆ๊ท ํ˜•ํ•˜๊ฒŒ ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์˜ค๋ฅธ์ชฝ ์ƒ๋‹จ plot์—์„œ๋Š” ์›น ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๊ฐ€ embedding ๊ณต๊ฐ„์˜ ํฌ์†Œํ™”๋œ ์˜์—ญ(์˜ˆ: ํ›ˆ๋ จ ์„ธํŠธ์˜ ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์— ๊ฐ€๊นŒ์šด ๊ณต๊ฐ„ ์˜์—ญ)์— ์œ„์น˜ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์˜ค๋ฅธ์ชฝ ํ•˜๋‹จ plot์—์„œ๋Š” ๋ฐ์ดํ„ฐ ์„ ํƒ ํ›„ perplexity๊ฐ€ ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด๋Ÿฌํ•œ ์ ๋“ค์ด ๋ฐ์ดํ„ฐ ์„ ํƒ์˜ ์˜ํ–ฅ์„ ๊ฐ€์žฅ ๋งŽ์ด ๋ฐ›๋Š”๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฐ€์šด๋ฐ ์˜ค๋ฅธ์ชฝ plot์€ ์ด๋Ÿฌํ•œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ ์ด ๊ฐ€์žฅ ๋‚ฎ์€ perplexity๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ด๋Š” ์ด๋Ÿฌํ•œ ์ ๋“ค์ด ํ›ˆ๋ จ ์„ธํŠธ์— ๊ฐ€๊นŒ์›Œ์„œ "์‰ฌ์šด" ์ ์ผ ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์ผ๊ด€๋œ _ํ‰๊ท _ ๋‚œํ•ด์„ฑ ๊ฐœ์„ ์„ ๊ด€์ฐฐํ•˜๋Š” ๋ฐ˜๋ฉด, ์„น์…˜ A.3์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๋‚œํ•ด์„ฑ ๊ฐœ์„ ์ด ๊ฒ€์ฆ ์„ธํŠธ๋งˆ๋‹ค ํฌ๊ฒŒ ๋‹ค๋ฅด๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋” ์ค‘์š”ํ•œ ๊ฒƒ์€ ๋ฐ์ดํ„ฐ ์„ ํƒ์ด ํ•ญ์ƒ CC-dedup, CommonCrawl ๋ฐ C4์™€ ๊ฐ™์€ ์›น ์Šค๋ƒ…์ƒท ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์˜ ์„ฑ๋Šฅ์„ ์†์ƒ์‹œํ‚จ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ˜„์ƒ์ด ๋ฐœ์ƒํ•˜๋Š” ์ด์œ ๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋ฅผ ํ›ˆ๋ จ ์„ธํŠธ์™€ ๋™์ผํ•œ ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์— ํฌํ•จ์‹œํ‚ค๊ณ  1.3B ๊ธฐ์ค€ ๋ชจ๋ธ์— ๋Œ€ํ•ด ํ›ˆ๋ จ ์„ธํŠธ์—์„œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง€์ ์— ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ด์›ƒ์„ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 5์˜ ์™ผ์ชฝ ํ”Œ๋กฏ์—์„œ๋Š” ์›น ์Šค๋ƒ…์ƒท๊ณผ ๋™์ผํ•œ ๋ถ„ํฌ์—์„œ ์ถ”์ถœํ•œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๊ฐ€ ๋‹ค๋ฅธ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์— ๋น„ํ•ด ํ›ˆ๋ จ ์„ธํŠธ์— ๋” ๊ฐ€๊น๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ๊ณ , ๊ทธ๋ฆผ 5์˜ ์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ์—์„œ๋Š” ๋ฐ์ดํ„ฐ ์„ ํƒ์ด ์ด๋Ÿฌํ•œ ์›น ์Šค๋ƒ…์ƒท ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์— ๋ถˆ๊ท ํ˜•ํ•˜๊ฒŒ ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์˜ค๋ฅธ์ชฝ ์ƒ๋‹จ ํ”Œ๋กฏ์—์„œ๋Š” ์›น ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๊ฐ€ ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์˜ ํฌ์†Œํ™”๋œ ์˜์—ญ(์˜ˆ: ํ›ˆ๋ จ ์„ธํŠธ์˜ ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์— ๊ฐ€๊นŒ์šด ๊ณต๊ฐ„ ์˜์—ญ)์— ์œ„์น˜ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์˜ค๋ฅธ์ชฝ ํ•˜๋‹จ ํ”Œ๋กฏ์—์„œ๋Š” ๋ฐ์ดํ„ฐ ์„ ํƒ ํ›„ ๋‚œํ•ด์„ฑ์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด๋Ÿฌํ•œ ์ ๋“ค์ด ๋ฐ์ดํ„ฐ ์„ ํƒ์˜ ์˜ํ–ฅ์„ ๊ฐ€์žฅ ๋งŽ์ด ๋ฐ›๋Š”๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฐ€์šด๋ฐ ์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ์€ ์ด๋Ÿฌํ•œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ ์ด ๊ฐ€์žฅ ๋‚ฎ์€ ๋‚œํ•ด์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ด๋Š” ์ด๋Ÿฌํ•œ ์ ๋“ค์ด ํ›ˆ๋ จ ์„ธํŠธ์— ๊ฐ€๊นŒ์›Œ์„œ "์‰ฌ์šด" ์ ์ผ ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
\(\bullet\)_GPT-3_. GPT-3 [55] was released in 2020, which scaled the model parameters to an ever larger size of 175B. In the GPT-3's paper, it formally introduced the concept of in-context learning (ICL)17, which utilizes LLMs in a few-shot or zero-shot way. ICL can teach (or instruct) LLMs to understand the tasks in the form of natural language text. With ICL, the pre-training and utilization of LLMs converge to the same language modeling paradigm: pre-training predicts the following text sequence conditioned on the context, while ICL predicts the correct task solution, which can be also formatted as a text sequence, given the task description and demonstrations. GPT-3 not only demonstrates very excellent performance in a variety of NLP tasks, but also on a number of specially designed tasks that require the abilities of reasoning or domain adaptation. Although the GPT-3's paper does not explicitly discuss the emergent abilities of LLMs, we can observe large performance leap that might transcend the basic scaling law [30], _e.g.,_ larger models have significantly stronger ICL ability (illustrated in the original Figure 1.2 of the GPT-3's paper [55]). Overall, GPT-3 can be viewed as a remarkable landmark in the journey evolving from PLMs to LLMs. It has empirically proved that scaling the neural networks to a significant size can lead to a huge increase in model capacity.
\(\bullet\)_GPT-3_. GPT-3 [55]๋Š” 2020๋…„์— ์ถœ์‹œ๋˜์—ˆ์œผ๋ฉฐ, ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ 175B์˜ ๋” ํฐ ๊ทœ๋ชจ๋กœ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” ๋ช‡ ๋ฒˆ์˜ ์ƒท ๋˜๋Š” ์ œ๋กœ ์ƒท ๋ฐฉ์‹์œผ๋กœ LLM์„ ํ™œ์šฉํ•˜๋Š” In-Context Learning (ICL)17์˜ ๊ฐœ๋…์„ ๊ณต์‹์ ์œผ๋กœ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ICL์€ ์ž์—ฐ์–ด ํ…์ŠคํŠธ ํ˜•ํƒœ๋กœ LLM์ด ์ž‘์—…์„ ์ดํ•ดํ•˜๋„๋ก ๊ฐ€๋ฅด์น˜๊ฑฐ๋‚˜ ์ง€์‹œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ICL์„ ํ†ตํ•ด ์‚ฌ์ „ ํ•™์Šต๊ณผ LLM์˜ ํ™œ์šฉ์€ ๋™์ผํ•œ ์–ธ์–ด ๋ชจ๋ธ๋ง ํŒจ๋Ÿฌ๋‹ค์ž„์œผ๋กœ ์ˆ˜๋ ด๋ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์‚ฌ์ „ ํ•™์Šต์€ ๋ฌธ๋งฅ์— ๋”ฐ๋ผ ๋‹ค์Œ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ˜๋ฉด, ICL์€ ์ž‘์—… ์„ค๋ช…๊ณผ ๋ฐ๋ชจ๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ํ…์ŠคํŠธ ์‹œํ€€์Šค๋กœ ํฌ๋งทํ•  ์ˆ˜ ์žˆ๋Š” ์˜ฌ๋ฐ”๋ฅธ ์ž‘์—… ์†”๋ฃจ์…˜์„ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. GPT-3๋Š” ๋‹ค์–‘ํ•œ NLP ์ž‘์—…๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ถ”๋ก ์ด๋‚˜ ๋„๋ฉ”์ธ ์ ์‘ ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•œ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ํŠน์ˆ˜ ์„ค๊ณ„๋œ ์ž‘์—…์—์„œ๋„ ๋งค์šฐ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” LLM์˜ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด ๋ช…์‹œ์ ์œผ๋กœ ๋…ผ์˜ํ•˜์ง€ ์•Š์•˜์ง€๋งŒ, ๋” ํฐ ๋ชจ๋ธ์ด ํ›จ์”ฌ ๋” ๊ฐ•๋ ฅํ•œ ICL ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค๋Š” ์ (GPT-3 ๋…ผ๋ฌธ [55]์˜ ์›๋ž˜ ๊ทธ๋ฆผ 1.2์— ์„ค๋ช…๋จ)์—์„œ ๊ธฐ๋ณธ์ ์ธ Scaling laws [30]์„ ๋„˜์–ด์„œ๋Š” ํฐ ์„ฑ๋Šฅ์˜ ๋„์•ฝ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ GPT-3๋Š” PLM์—์„œ LLM์œผ๋กœ ์ง„ํ™”ํ•˜๋Š” ์—ฌ์ •์—์„œ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์ด์ •ํ‘œ๋ผ๊ณ  ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ ๊ฒฝ๋ง์„ ์ƒ๋‹นํ•œ ๊ทœ๋ชจ๋กœ ํ™•์žฅํ•˜๋ฉด ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ์ฆ๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_GPT-3_. GPT-3 [55]๋Š” 2020๋…„์— ์ถœ์‹œ๋˜์—ˆ์œผ๋ฉฐ, ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ 175B์˜ ๋” ํฐ ๊ทœ๋ชจ๋กœ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” ๋ช‡ ๋ฒˆ์˜ ์ƒท ๋˜๋Š” ์ œ๋กœ ์ƒท ๋ฐฉ์‹์œผ๋กœ LLM์„ ํ™œ์šฉํ•˜๋Š” ์ปจํ…์ŠคํŠธ ๋‚ด ํ•™์Šต(ICL)17์˜ ๊ฐœ๋…์„ ๊ณต์‹์ ์œผ๋กœ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ICL์€ ์ž์—ฐ์–ด ํ…์ŠคํŠธ ํ˜•ํƒœ๋กœ LLM์ด ์ž‘์—…์„ ์ดํ•ดํ•˜๋„๋ก ๊ฐ€๋ฅด์น˜๊ฑฐ๋‚˜ ์ง€์‹œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ICL์„ ํ†ตํ•ด ์‚ฌ์ „ ํ•™์Šต๊ณผ LLM์˜ ํ™œ์šฉ์€ ๋™์ผํ•œ ์–ธ์–ด ๋ชจ๋ธ๋ง ํŒจ๋Ÿฌ๋‹ค์ž„์œผ๋กœ ์ˆ˜๋ ด๋ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์‚ฌ์ „ ํ•™์Šต์€ ๋ฌธ๋งฅ์— ๋”ฐ๋ผ ๋‹ค์Œ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ˜๋ฉด, ICL์€ ์ž‘์—… ์„ค๋ช…๊ณผ ๋ฐ๋ชจ๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ํ…์ŠคํŠธ ์‹œํ€€์Šค๋กœ ํฌ๋งทํ•  ์ˆ˜ ์žˆ๋Š” ์˜ฌ๋ฐ”๋ฅธ ์ž‘์—… ์†”๋ฃจ์…˜์„ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. GPT-3๋Š” ๋‹ค์–‘ํ•œ NLP ์ž‘์—…๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ถ”๋ก ์ด๋‚˜ ๋„๋ฉ”์ธ ์ ์‘ ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•œ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ํŠน์ˆ˜ ์„ค๊ณ„๋œ ์ž‘์—…์—์„œ๋„ ๋งค์šฐ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” LLM์˜ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด ๋ช…์‹œ์ ์œผ๋กœ ๋…ผ์˜ํ•˜์ง€ ์•Š์•˜์ง€๋งŒ, ๋” ํฐ ๋ชจ๋ธ์ด ํ›จ์”ฌ ๋” ๊ฐ•๋ ฅํ•œ ICL ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค๋Š” ์ (GPT-3 ๋…ผ๋ฌธ [55]์˜ ์›๋ž˜ ๊ทธ๋ฆผ 1.2์— ์„ค๋ช…๋จ)์—์„œ ๊ธฐ๋ณธ์ ์ธ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™ [30]์„ ๋„˜์–ด์„œ๋Š” ํฐ ์„ฑ๋Šฅ์˜ ๋„์•ฝ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ GPT-3๋Š” PLM์—์„œ LLM์œผ๋กœ ์ง„ํ™”ํ•˜๋Š” ์—ฌ์ •์—์„œ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์ด์ •ํ‘œ๋ผ๊ณ  ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ ๊ฒฝ๋ง์„ ์ƒ๋‹นํ•œ ๊ทœ๋ชจ๋กœ ํ™•์žฅํ•˜๋ฉด ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ์ฆ๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Visual instruction data._ Extensive work [856, 831] has empirically found that both quantity and quality of visual instructions have an important impact on model performance of MLLMs. One basic way to construct visual instructions is to leverage the exceptional capability of LLMs to synthesize instructions based on text descriptions of images [851]. To further enhance the quality of instructions, one can construct fine-grained visual instructions with the help of human annotation [857, 833] or synthesize more complex data through carefully-designed prompts [835]. Despite the effectiveness of the above LLM-based approaches, one primary question emerges as to whether a LLM (_i.e._, text generation model without training on any images) possesses the ability to generate sufficiently good visual instructions solely based on verbalized visual information (_e.g._, captions and coordinates). Specially, existing work has also revealed that visual instructions generated by LLMs sometimes contain misinterpretations about the visual information, _e.g._, object hallucination [844]. Therefore, it is crucial to design effective verification methods to control the quality of instruction data generated by LLMs [835]. Furthermore, it still needs more investigation about what makes good visual instructions and how visual instructions elicit specific multimodal abilities in MLLMs.
\(\bullet\)_Visual instruction data._ ๊ด‘๋ฒ”์œ„ํ•œ ์—ฐ๊ตฌ[856, 831]์— ๋”ฐ๋ฅด๋ฉด ์‹œ๊ฐ์  ์ง€์นจ์˜ ์–‘๊ณผ ์งˆ ๋ชจ๋‘ MLLM์˜ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ์ค‘์š”ํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์‹œ๊ฐ์  ์ง€์นจ์„ ๊ตฌ์„ฑํ•˜๋Š” ๊ธฐ๋ณธ์ ์ธ ๋ฐฉ๋ฒ• ์ค‘ ํ•˜๋‚˜๋Š” ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ํ…์ŠคํŠธ ์„ค๋ช…์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ง€์นจ์„ ํ•ฉ์„ฑํ•˜๋Š” LLM์˜ ๋›ฐ์–ด๋‚œ ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค[851]. ์ง€์นจ์˜ ํ’ˆ์งˆ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์‚ฌ๋žŒ์˜ ์ฃผ์„[857, 833]์„ ํ†ตํ•ด ์„ธ๋ถ„ํ™”๋œ ์‹œ๊ฐ์  ์ง€์นจ์„ ๊ตฌ์„ฑํ•˜๊ฑฐ๋‚˜ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ค๊ณ„๋œ prompts[835]๋ฅผ ํ†ตํ•ด ๋” ๋ณต์žกํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ํ•ฉ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ„์˜ LLM ๊ธฐ๋ฐ˜ ์ ‘๊ทผ๋ฒ•์˜ ํšจ๊ณผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , LLM(_์ฆ‰, ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ํ•™์Šต์ด ์—†๋Š” ํ…์ŠคํŠธ ์ƒ์„ฑ ๋ชจ๋ธ_)์ด ์–ธ์–ดํ™”๋œ ์‹œ๊ฐ ์ •๋ณด(_e.g._, captions ๋ฐ coordinates)๋งŒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ถฉ๋ถ„ํžˆ ์ข‹์€ ์‹œ๊ฐ์  ์ง€์นจ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š”์ง€์— ๋Œ€ํ•œ ์ฃผ์š” ์งˆ๋ฌธ์ด ์ œ๊ธฐ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ๊ธฐ์กด ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด LLM์ด ์ƒ์„ฑํ•œ ์‹œ๊ฐ์  ์ง€์นจ์—๋Š” ๋•Œ๋•Œ๋กœ ์‹œ๊ฐ์  ์ •๋ณด์— ๋Œ€ํ•œ ์˜คํ•ด, _e.g._, object hallucination[844]์ด ํฌํ•จ๋˜์–ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ LLM์ด ์ƒ์„ฑํ•œ ์ง€์นจ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์„ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด ํšจ๊ณผ์ ์ธ verification ๋ฐฉ๋ฒ•์„ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค[835]. ๋˜ํ•œ ์ข‹์€ ์‹œ๊ฐ์  ์ง€์นจ์ด ๋˜๋Š” ์š”์†Œ์™€ ์‹œ๊ฐ์  ์ง€์นจ์ด MLLM์—์„œ ํŠน์ • ๋‹ค์ค‘ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๋Š” ๋ฐฉ์‹์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๊ฐ€ ๋” ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_์‹œ๊ฐ์  ์ง€์นจ ๋ฐ์ดํ„ฐ._ ๊ด‘๋ฒ”์œ„ํ•œ ์—ฐ๊ตฌ[856, 831]์— ๋”ฐ๋ฅด๋ฉด ์‹œ๊ฐ์  ์ง€์นจ์˜ ์–‘๊ณผ ์งˆ ๋ชจ๋‘ MLLM์˜ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ์ค‘์š”ํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์‹œ๊ฐ์  ์ง€์นจ์„ ๊ตฌ์„ฑํ•˜๋Š” ๊ธฐ๋ณธ์ ์ธ ๋ฐฉ๋ฒ• ์ค‘ ํ•˜๋‚˜๋Š” ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ํ…์ŠคํŠธ ์„ค๋ช…์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ง€์นจ์„ ํ•ฉ์„ฑํ•˜๋Š” LLM์˜ ๋›ฐ์–ด๋‚œ ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค[851]. ์ง€์นจ์˜ ํ’ˆ์งˆ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์‚ฌ๋žŒ์˜ ์ฃผ์„[857, 833]์„ ํ†ตํ•ด ์„ธ๋ถ„ํ™”๋œ ์‹œ๊ฐ์  ์ง€์นจ์„ ๊ตฌ์„ฑํ•˜๊ฑฐ๋‚˜ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ค๊ณ„๋œ ํ”„๋กฌํ”„ํŠธ[835]๋ฅผ ํ†ตํ•ด ๋” ๋ณต์žกํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ํ•ฉ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ„์˜ LLM ๊ธฐ๋ฐ˜ ์ ‘๊ทผ๋ฒ•์˜ ํšจ๊ณผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , LLM(_์ฆ‰, ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ํ•™์Šต์ด ์—†๋Š” ํ…์ŠคํŠธ ์ƒ์„ฑ ๋ชจ๋ธ)์ด ์–ธ์–ดํ™”๋œ ์‹œ๊ฐ ์ •๋ณด(์˜ˆ: ์บก์…˜ ๋ฐ ์ขŒํ‘œ)๋งŒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ถฉ๋ถ„ํžˆ ์ข‹์€ ์‹œ๊ฐ์  ์ง€์นจ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š”์ง€์— ๋Œ€ํ•œ ์ฃผ์š” ์งˆ๋ฌธ์ด ์ œ๊ธฐ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ๊ธฐ์กด ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด LLM์ด ์ƒ์„ฑํ•œ ์‹œ๊ฐ์  ์ง€์นจ์—๋Š” ๋•Œ๋•Œ๋กœ ์‹œ๊ฐ์  ์ •๋ณด์— ๋Œ€ํ•œ ์˜คํ•ด, ์˜ˆ๋ฅผ ๋“ค์–ด ์‚ฌ๋ฌผ ํ™˜๊ฐ[844]์ด ํฌํ•จ๋˜์–ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ LLM์ด ์ƒ์„ฑํ•œ ์ง€์นจ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์„ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด ํšจ๊ณผ์ ์ธ ๊ฒ€์ฆ ๋ฐฉ๋ฒ•์„ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค[835]. ๋˜ํ•œ ์ข‹์€ ์‹œ๊ฐ์  ์ง€์นจ์ด ๋˜๋Š” ์š”์†Œ์™€ ์‹œ๊ฐ์  ์ง€์นจ์ด MLLM์—์„œ ํŠน์ • ๋‹ค์ค‘ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๋Š” ๋ฐฉ์‹์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๊ฐ€ ๋” ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
Different from traditional LLMs, autonomous agent usually has to play as specific roles (_e.g._, program coder, researcher and chemist) for accomplishing different tasks. Thus, the capability of the agent for role-playing is very important. Although LLMs can effectively simulate many common roles such as movie reviewers, there are still various roles and aspects that they struggle to capture accurately. To begin with, LLMs are usually trained based on web-corpus, thus for the roles which are seldom discussed on the web or the newly emerging roles, LLMs may not simulate them well. In addition, previous research [30] has shown that existing LLMs may not well model the human cognitive psychology characters, leading to the lack of self-awareness in conversation scenarios. Potential solution to these problems may include fine-tuning LLMs or carefully designing the agent prompts/architectures [183]. For example, one can firstly collect real-human data for uncommon roles or psychology characters, and then leverage it to fine-tune LLMs. However, how to ensure that fine-tuned model still perform well for the common roles may pose further challenges. Beyond fine-tuning, one can also design tailored agent prompts/architectures to enhance the capability of LLM on role-playing. However, finding the optimal prompts/architectures is not easy, since their designing spaces are too large.
๊ธฐ์กด LLM๊ณผ ๋‹ฌ๋ฆฌ ์ž์œจ ์—์ด์ „ํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ํŠน์ • ์—ญํ• (_์˜ˆ: ํ”„๋กœ๊ทธ๋žจ ์ฝ”๋”, ์—ฐ๊ตฌ์› ๋ฐ ํ™”ํ•™์ž)์„ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—ญํ• ๊ทน์— ๋Œ€ํ•œ ์—์ด์ „ํŠธ์˜ ์—ญ๋Ÿ‰์€ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. LLM์€ ์˜ํ™” ๋ฆฌ๋ทฐ์–ด์™€ ๊ฐ™์€ ๋งŽ์€ ์ผ๋ฐ˜์ ์ธ ์—ญํ• ์„ ํšจ๊ณผ์ ์œผ๋กœ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์—ฌ์ „ํžˆ ์ •ํ™•ํ•˜๊ฒŒ ํฌ์ฐฉํ•˜์ง€ ๋ชปํ•˜๋Š” ๋‹ค์–‘ํ•œ ์—ญํ• ๊ณผ ์ธก๋ฉด์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ์„ , LLM์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์›น ์ฝ”ํผ์Šค๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•™์Šต๋˜๋ฏ€๋กœ ์›น์—์„œ ๊ฑฐ์˜ ๋…ผ์˜๋˜์ง€ ์•Š๊ฑฐ๋‚˜ ์ƒˆ๋กœ ๋“ฑ์žฅํ•˜๋Š” ์—ญํ• ์˜ ๊ฒฝ์šฐ LLM์ด ์ด๋ฅผ ์ž˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด์ „ ์—ฐ๊ตฌ[30]์— ๋”ฐ๋ฅด๋ฉด ๊ธฐ์กด LLM์€ ์ธ๊ฐ„์˜ ์ธ์ง€ ์‹ฌ๋ฆฌ์  ํŠน์„ฑ์„ ์ž˜ ๋ชจ๋ธ๋งํ•˜์ง€ ๋ชปํ•˜์—ฌ ๋Œ€ํ™” ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์ž๊ธฐ ์ธ์‹์ด ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์ž ์žฌ์ ์ธ ํ•ด๊ฒฐ์ฑ…์—๋Š” LLM์„ Fine-tuningํ•˜๊ฑฐ๋‚˜ ์—์ด์ „ํŠธ prompts/architectures๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ[183]์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋จผ์ € ํ”ํ•˜์ง€ ์•Š์€ ์—ญํ• ์ด๋‚˜ ์‹ฌ๋ฆฌ์  ํŠน์„ฑ์— ๋Œ€ํ•œ ์‹ค์ œ ์ธ๊ฐ„ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•œ ๋‹ค์Œ ์ด๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์„ Fine-tuningํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Fine-tuning๋œ ๋ชจ๋ธ์ด ์ผ๋ฐ˜์ ์ธ ์—ญํ• ์— ๋Œ€ํ•ด์„œ๋„ ์—ฌ์ „ํžˆ ์ž˜ ์ž‘๋™ํ•˜๋„๋ก ๋ณด์žฅํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์ถ”๊ฐ€์ ์ธ ๊ณผ์ œ๊ฐ€ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Fine-tuning ์™ธ์—๋„ ๋งž์ถคํ˜• ์—์ด์ „ํŠธ prompts/architectures๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ ์—ญํ• ๊ทน์— ๋Œ€ํ•œ LLM์˜ ์—ญ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ prompts/architectures์˜ ์„ค๊ณ„ ๊ณต๊ฐ„์ด ๋„ˆ๋ฌด ์ปค์„œ ์ตœ์ ์˜ prompts/architectures๋ฅผ ์ฐพ๋Š” ๊ฒƒ์€ ์‰ฝ์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
๊ธฐ์กด LLM๊ณผ ๋‹ฌ๋ฆฌ ์ž์œจ ์—์ด์ „ํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ํŠน์ • ์—ญํ• (_์˜ˆ: ํ”„๋กœ๊ทธ๋žจ ์ฝ”๋”, ์—ฐ๊ตฌ์› ๋ฐ ํ™”ํ•™์ž)์„ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—ญํ• ๊ทน์— ๋Œ€ํ•œ ์—์ด์ „ํŠธ์˜ ์—ญ๋Ÿ‰์€ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. LLM์€ ์˜ํ™” ๋ฆฌ๋ทฐ์–ด์™€ ๊ฐ™์€ ๋งŽ์€ ์ผ๋ฐ˜์ ์ธ ์—ญํ• ์„ ํšจ๊ณผ์ ์œผ๋กœ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์—ฌ์ „ํžˆ ์ •ํ™•ํ•˜๊ฒŒ ํฌ์ฐฉํ•˜์ง€ ๋ชปํ•˜๋Š” ๋‹ค์–‘ํ•œ ์—ญํ• ๊ณผ ์ธก๋ฉด์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ์„ , LLM์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์›น ์ฝ”ํผ์Šค๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•™์Šต๋˜๋ฏ€๋กœ ์›น์—์„œ ๊ฑฐ์˜ ๋…ผ์˜๋˜์ง€ ์•Š๊ฑฐ๋‚˜ ์ƒˆ๋กœ ๋“ฑ์žฅํ•˜๋Š” ์—ญํ• ์˜ ๊ฒฝ์šฐ LLM์ด ์ด๋ฅผ ์ž˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด์ „ ์—ฐ๊ตฌ[30]์— ๋”ฐ๋ฅด๋ฉด ๊ธฐ์กด LLM์€ ์ธ๊ฐ„์˜ ์ธ์ง€ ์‹ฌ๋ฆฌ์  ํŠน์„ฑ์„ ์ž˜ ๋ชจ๋ธ๋งํ•˜์ง€ ๋ชปํ•˜์—ฌ ๋Œ€ํ™” ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์ž๊ธฐ ์ธ์‹์ด ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์ž ์žฌ์ ์ธ ํ•ด๊ฒฐ์ฑ…์—๋Š” LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๊ฑฐ๋‚˜ ์—์ด์ „ํŠธ ํ”„๋กฌํ”„ํŠธ/์•„ํ‚คํ…์ฒ˜๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ[183]์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋จผ์ € ํ”ํ•˜์ง€ ์•Š์€ ์—ญํ• ์ด๋‚˜ ์‹ฌ๋ฆฌ์  ํŠน์„ฑ์— ๋Œ€ํ•œ ์‹ค์ œ ์ธ๊ฐ„ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•œ ๋‹ค์Œ ์ด๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ์ด ์ผ๋ฐ˜์ ์ธ ์—ญํ• ์— ๋Œ€ํ•ด์„œ๋„ ์—ฌ์ „ํžˆ ์ž˜ ์ž‘๋™ํ•˜๋„๋ก ๋ณด์žฅํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์ถ”๊ฐ€์ ์ธ ๊ณผ์ œ๊ฐ€ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฏธ์„ธ ์กฐ์ • ์™ธ์—๋„ ๋งž์ถคํ˜• ์—์ด์ „ํŠธ ํ”„๋กฌํ”„ํŠธ/์•„ํ‚คํ…์ฒ˜๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ ์—ญํ• ๊ทน์— ๋Œ€ํ•œ LLM์˜ ์—ญ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ”„๋กฌํ”„ํŠธ/์•„ํ‚คํ…์ฒ˜์˜ ์„ค๊ณ„ ๊ณต๊ฐ„์ด ๋„ˆ๋ฌด ์ปค์„œ ์ตœ์ ์˜ ํ”„๋กฌํ”„ํŠธ/์•„ํ‚คํ…์ฒ˜๋ฅผ ์ฐพ๋Š” ๊ฒƒ์€ ์‰ฝ์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
**Metrics**: In order to objectively evaluate the effectiveness of the agents, designing proper metrics is significant, which may influence the evaluation accuracy and comprehensiveness. Ideal evaluation metrics should precisely reflect the quality of the agents, and align with the human feelings when using them in real-world scenarios. In existing work, we can conclude the following representative evaluation metrics. (1) _Task success metrics:_ These metrics measure how well an agent can complete tasks and achieve goals. Common metrics include success rate [12, 57, 59, 22], reward/score [22, 59, 138], coverage [16], and accuracy [18, 40, 102]. Higher values indicate greater task completion ability. (2) _Human similarity metrics:_ These metrics quantify the degree to which the agent behaviors closely resembles that of humans. Typical examples include trajectory/location accuracy [164, 38], dialogue similarities [79, 102], and mimicry of human responses [29, 102]. Higher similarity suggests better human simulation performance. (3) _Efficiency metrics:_ In contrast to the aforementioned metrics used to evaluate the agent effectiveness, these metrics aim to assess the efficiency of agent. Commonly considered metrics encompass the length of planning [57], the cost associated with development [18], the speed of inference [16, 38], and number of clarification dialogues [138].
**Metrics**: ์—์ด์ „ํŠธ์˜ ํšจ๊ณผ๋ฅผ ๊ฐ๊ด€์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ ์ ˆํ•œ metrics๋ฅผ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋ฉฐ, ์ด๋Š” ํ‰๊ฐ€์˜ ์ •ํ™•์„ฑ๊ณผ ํฌ๊ด„์„ฑ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์ƒ์ ์ธ ํ‰๊ฐ€ metrics๋Š” ์—์ด์ „ํŠธ์˜ ํ’ˆ์งˆ์„ ์ •ํ™•ํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๊ณ  ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์‚ฌ์šฉํ•  ๋•Œ ์‚ฌ๋žŒ์˜ ๊ฐ์ •๊ณผ ์ผ์น˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋Œ€ํ‘œ์ ์ธ ํ‰๊ฐ€ metrics๋ฅผ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (1) _Task success metrics:_ ์ด๋Ÿฌํ•œ metrics๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๊ณ  ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ์–ผ๋งˆ๋‚˜ ์ž˜ํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ metrics๋กœ๋Š” ์„ฑ๊ณต๋ฅ [12, 57, 59, 22], ๋ณด์ƒ/์ ์ˆ˜[22, 59, 138], ์ปค๋ฒ„๋ฆฌ์ง€[16], ์ •ํ™•๋„[18, 40, 102] ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ’์ด ๋†’์„์ˆ˜๋ก ์ž‘์—… ์™„๋ฃŒ ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. (2) _Human similarity metrics:_ ์ด๋Ÿฌํ•œ metrics๋Š” ์—์ด์ „ํŠธ์˜ ํ–‰๋™์ด ์ธ๊ฐ„์˜ ํ–‰๋™๊ณผ ์–ผ๋งˆ๋‚˜ ์œ ์‚ฌํ•œ์ง€๋ฅผ ์ •๋Ÿ‰ํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ๋Š” ๊ถค์ /์œ„์น˜ ์ •ํ™•๋„[164, 38], ๋Œ€ํ™” ์œ ์‚ฌ์„ฑ[79, 102], ์ธ๊ฐ„ ๋ฐ˜์‘ ๋ชจ๋ฐฉ[29, 102] ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์œ ์‚ฌ๋„๊ฐ€ ๋†’์„์ˆ˜๋ก ์ธ๊ฐ„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. (3) _Efficiency metrics:_ ์•ž์„œ ์„ค๋ช…ํ•œ ์—์ด์ „ํŠธ์˜ ํšจ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” metrics์™€๋Š” ๋‹ฌ๋ฆฌ, ์ด metrics๋Š” ์—์ด์ „ํŠธ์˜ ํšจ์œจ์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๊ณ ๋ ค๋˜๋Š” metrics์—๋Š” ๊ณ„ํš์˜ ๊ธธ์ด[57], ๊ฐœ๋ฐœ๊ณผ ๊ด€๋ จ๋œ ๋น„์šฉ[18], ์ถ”๋ก  ์†๋„[16, 38], ๋ช…ํ™•ํ™” ๋Œ€ํ™” ํšŸ์ˆ˜[138] ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
**๋ฉ”ํŠธ๋ฆญ**: ์—์ด์ „ํŠธ์˜ ํšจ๊ณผ๋ฅผ ๊ฐ๊ด€์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ ์ ˆํ•œ ๋ฉ”ํŠธ๋ฆญ์„ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋ฉฐ, ์ด๋Š” ํ‰๊ฐ€์˜ ์ •ํ™•์„ฑ๊ณผ ํฌ๊ด„์„ฑ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์ƒ์ ์ธ ํ‰๊ฐ€ ์ง€ํ‘œ๋Š” ์—์ด์ „ํŠธ์˜ ํ’ˆ์งˆ์„ ์ •ํ™•ํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๊ณ  ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์‚ฌ์šฉํ•  ๋•Œ ์‚ฌ๋žŒ์˜ ๊ฐ์ •๊ณผ ์ผ์น˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋Œ€ํ‘œ์ ์ธ ํ‰๊ฐ€ ์ง€ํ‘œ๋ฅผ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (1) _์ž‘์—… ์„ฑ๊ณต ๋ฉ”ํŠธ๋ฆญ:_ ์ด๋Ÿฌํ•œ ๋ฉ”ํŠธ๋ฆญ์€ ์—์ด์ „ํŠธ๊ฐ€ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๊ณ  ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ์–ผ๋งˆ๋‚˜ ์ž˜ํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ๋Š” ์„ฑ๊ณต๋ฅ [12, 57, 59, 22], ๋ณด์ƒ/์ ์ˆ˜[22, 59, 138], ์ปค๋ฒ„๋ฆฌ์ง€[16], ์ •ํ™•๋„[18, 40, 102] ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ’์ด ๋†’์„์ˆ˜๋ก ์ž‘์—… ์™„๋ฃŒ ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. (2) _์ธ๊ฐ„ ์œ ์‚ฌ์„ฑ ๋ฉ”ํŠธ๋ฆญ:_ ์ด๋Ÿฌํ•œ ๋ฉ”ํŠธ๋ฆญ์€ ์—์ด์ „ํŠธ์˜ ํ–‰๋™์ด ์ธ๊ฐ„์˜ ํ–‰๋™๊ณผ ์–ผ๋งˆ๋‚˜ ์œ ์‚ฌํ•œ์ง€๋ฅผ ์ •๋Ÿ‰ํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ๋Š” ๊ถค์ /์œ„์น˜ ์ •ํ™•๋„[164, 38], ๋Œ€ํ™” ์œ ์‚ฌ์„ฑ[79, 102], ์ธ๊ฐ„ ๋ฐ˜์‘ ๋ชจ๋ฐฉ[29, 102] ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์œ ์‚ฌ๋„๊ฐ€ ๋†’์„์ˆ˜๋ก ์ธ๊ฐ„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. (3) _ํšจ์œจ์„ฑ ๋ฉ”ํŠธ๋ฆญ:_ ์•ž์„œ ์„ค๋ช…ํ•œ ์—์ด์ „ํŠธ์˜ ํšจ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋ฉ”ํŠธ๋ฆญ๊ณผ๋Š” ๋‹ฌ๋ฆฌ, ์ด ๋ฉ”ํŠธ๋ฆญ์€ ์—์ด์ „ํŠธ์˜ ํšจ์œจ์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๊ณ ๋ ค๋˜๋Š” ๋ฉ”ํŠธ๋ฆญ์—๋Š” ๊ณ„ํš์˜ ๊ธธ์ด[57], ๊ฐœ๋ฐœ๊ณผ ๊ด€๋ จ๋œ ๋น„์šฉ[18], ์ถ”๋ก  ์†๋„[16, 38], ๋ช…ํ™•ํ™” ๋Œ€ํ™” ํšŸ์ˆ˜[138] ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
\begin{table} \begin{tabular}{l l l} \hline \hline **Category** & **Description** & **Number of links** \\ \hline adult & adult websites: from eroticism to hard pornography & 4516478 \\ phishing & phishing websites, malwares, etc. & 42445 \\ dating & dating websites & 3829 \\ gambling & online casino & 1365 \\ filehosting & websites hosting files, videos, pictures, music & 909 \\ ddos & websites related to ddos attacks & 421 \\ agressif & hate, racism, etc & 390 \\ chat & online chat websites & 244 \\ mixed adult & websites with some adult content & 153 \\ ariel & French regulated gambling websites & 69 \\ \hline \hline \end{tabular} \end{table} Table 13: **We select categories likely to contain adult or malicious content, as well as spam or unstructured text.**We curated the lists based on manual inspection of the data, informed by top hits reported by ToxicBERT. For the strict subword matching, we included words that were unequivocally related to adult content (e.g., groupsex). We avoided partial unclear matches (e.g., ass), that may be part of neutral words (e.g., massachusetts). In the soft word list, we included words that do not constitute a sufficient reason to discard the document on their own, but which are suspicious when multiple words from the list result in a match. This helped with keeping medical or legal content unaffected (e.g., a single match of dick).
\begin{table} \begin{tabular}{l l l} \hline \hline **์นดํ…Œ๊ณ ๋ฆฌ** & **์„ค๋ช…** & **๋งํฌ ์ˆ˜** \\ \hline ์„ฑ์ธ & ์—๋กœํ‹ฐ์‹œ์ฆ˜์—์„œ ํ•˜๋“œ ํฌ๋ฅด๋…ธ์— ์ด๋ฅด๋Š” ์„ฑ์ธ ์›น์‚ฌ์ดํŠธ & 4516478 \\ ํ”ผ์‹ฑ & ํ”ผ์‹ฑ ์›น์‚ฌ์ดํŠธ, ๋ฉ€์›จ์–ด ๋“ฑ & 42445 \\ ๋ฐ์ดํŠธ & ๋ฐ์ดํŠธ ์›น์‚ฌ์ดํŠธ & 3829 \\ ๊ฐฌ๋ธ”๋ง & ์˜จ๋ผ์ธ ์นด์ง€๋…ธ & 1365 \\ ํŒŒ์ผ ํ˜ธ์ŠคํŒ… & ํŒŒ์ผ, ๋™์˜์ƒ, ์‚ฌ์ง„, ์Œ์•…์„ ํ˜ธ์ŠคํŒ…ํ•˜๋Š” ์›น์‚ฌ์ดํŠธ & 909 \\ ddos & ddos ๊ณต๊ฒฉ๊ณผ ๊ด€๋ จ๋œ ์›น์‚ฌ์ดํŠธ & 421 \\ ๊ณต๊ฒฉ์  & ์ฆ์˜ค, ์ธ์ข…์ฐจ๋ณ„ ๋“ฑ & 390 \\ ์ฑ„ํŒ… & ์˜จ๋ผ์ธ ์ฑ„ํŒ… ์›น์‚ฌ์ดํŠธ & 244 \\ ํ˜ผํ•ฉ ์„ฑ์ธ & ์ผ๋ถ€ ์„ฑ์ธ ์ฝ˜ํ…์ธ ๊ฐ€ ํฌํ•จ๋œ ์›น์‚ฌ์ดํŠธ & 153 \\ ariel & ํ”„๋ž‘์Šค ๊ทœ์ œ ๋„๋ฐ• ์›น์‚ฌ์ดํŠธ & 69 \\ \hline \hline \end{tabular} \end{table} ํ‘œ 13: **์„ฑ์ธ ๋˜๋Š” ์•…์„ฑ ์ฝ˜ํ…์ธ , ์ŠคํŒธ ๋˜๋Š” ๋น„์ •ํ˜• ํ…์ŠคํŠธ๊ฐ€ ํฌํ•จ๋  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ์นดํ…Œ๊ณ ๋ฆฌ๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.**ToxicBERT๊ฐ€ ๋ณด๊ณ ํ•œ ์ƒ์œ„ ๊ฒ€์ƒ‰์–ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์‚ฌํ•˜์—ฌ ๋ชฉ๋ก์„ ํ๋ ˆ์ด์…˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—„๊ฒฉํ•œ ํ•˜์œ„ ๋‹จ์–ด ๋งค์นญ์˜ ๊ฒฝ์šฐ ์„ฑ์ธ ์ฝ˜ํ…์ธ ์™€ ๋ช…๋ฐฑํžˆ ๊ด€๋ จ๋œ ๋‹จ์–ด(์˜ˆ: groupsex)๋ฅผ ํฌํ•จ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ค‘๋ฆฝ์ ์ธ ๋‹จ์–ด(์˜ˆ: massachusetts)์˜ ์ผ๋ถ€์ผ ์ˆ˜ ์žˆ๋Š” ๋ถ€๋ถ„์ ์œผ๋กœ ๋ถˆ๋ถ„๋ช…ํ•œ ์ผ์น˜(์˜ˆ: ass)๋Š” ํ”ผํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ถ€๋“œ๋Ÿฌ์šด ๋‹จ์–ด ๋ชฉ๋ก์—๋Š” ๋ฌธ์„œ๋ฅผ ํ๊ธฐํ•  ์ถฉ๋ถ„ํ•œ ์ด์œ ๊ฐ€ ๋˜์ง€๋Š” ์•Š์ง€๋งŒ ๋ชฉ๋ก์˜ ์—ฌ๋Ÿฌ ๋‹จ์–ด๊ฐ€ ์ผ์น˜ํ•  ๊ฒฝ์šฐ ์˜์‹ฌ์Šค๋Ÿฌ์šด ๋‹จ์–ด๋ฅผ ํฌํ•จ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์˜๋ฃŒ ๋˜๋Š” ๋ฒ•๋ฅ  ์ฝ˜ํ…์ธ ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๊ณ  ์œ ์ง€ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค(์˜ˆ: dick์˜ ๋‹จ์ผ ์ผ์น˜).
\begin{table} \begin{tabular}{l l l} \hline \hline **์นดํ…Œ๊ณ ๋ฆฌ** & **์„ค๋ช…** & **๋งํฌ ์ˆ˜** \\ \hline ์„ฑ์ธ & ์—๋กœํ‹ฐ์‹œ์ฆ˜์—์„œ ํ•˜๋“œ ํฌ๋ฅด๋…ธ์— ์ด๋ฅด๋Š” ์„ฑ์ธ ์›น์‚ฌ์ดํŠธ & 4516478 \\ ํ”ผ์‹ฑ & ํ”ผ์‹ฑ ์›น์‚ฌ์ดํŠธ, ๋ฉ€์›จ์–ด ๋“ฑ & 42445 \\ ๋ฐ์ดํŠธ & ๋ฐ์ดํŠธ ์›น์‚ฌ์ดํŠธ & 3829 \\ ๊ฐฌ๋ธ”๋ง & ์˜จ๋ผ์ธ ์นด์ง€๋…ธ & 1365 \\ ํŒŒ์ผ ํ˜ธ์ŠคํŒ… & ํŒŒ์ผ, ๋™์˜์ƒ, ์‚ฌ์ง„, ์Œ์•…์„ ํ˜ธ์ŠคํŒ…ํ•˜๋Š” ์›น์‚ฌ์ดํŠธ & 909 \\ ddos & ddos ๊ณต๊ฒฉ๊ณผ ๊ด€๋ จ๋œ ์›น์‚ฌ์ดํŠธ & 421 \\ ๊ณต๊ฒฉ์  & ์ฆ์˜ค, ์ธ์ข…์ฐจ๋ณ„ ๋“ฑ & 390 \\ ์ฑ„ํŒ… & ์˜จ๋ผ์ธ ์ฑ„ํŒ… ์›น์‚ฌ์ดํŠธ & 244 \\ ํ˜ผํ•ฉ ์„ฑ์ธ & ์ผ๋ถ€ ์„ฑ์ธ ์ฝ˜ํ…์ธ ๊ฐ€ ํฌํ•จ๋œ ์›น์‚ฌ์ดํŠธ & 153 \\ ariel & ํ”„๋ž‘์Šค ๊ทœ์ œ ๋„๋ฐ• ์›น์‚ฌ์ดํŠธ & 69 \\ \hline \hline \end{tabular} \end{table} ํ‘œ 13: **์„ฑ์ธ ๋˜๋Š” ์•…์„ฑ ์ฝ˜ํ…์ธ , ์ŠคํŒธ ๋˜๋Š” ๋น„์ •ํ˜• ํ…์ŠคํŠธ๊ฐ€ ํฌํ•จ๋  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ์นดํ…Œ๊ณ ๋ฆฌ๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.**ToxicBERT๊ฐ€ ๋ณด๊ณ ํ•œ ์ƒ์œ„ ๊ฒ€์ƒ‰์–ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์‚ฌํ•˜์—ฌ ๋ชฉ๋ก์„ ํ๋ ˆ์ด์…˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—„๊ฒฉํ•œ ํ•˜์œ„ ๋‹จ์–ด ๋งค์นญ์˜ ๊ฒฝ์šฐ ์„ฑ์ธ ์ฝ˜ํ…์ธ ์™€ ๋ช…๋ฐฑํžˆ ๊ด€๋ จ๋œ ๋‹จ์–ด(์˜ˆ: ๊ทธ๋ฃน์„น์Šค)๋ฅผ ํฌํ•จ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ค‘๋ฆฝ์ ์ธ ๋‹จ์–ด(์˜ˆ: ๋งค์‚ฌ์ถ”์„ธ์ธ )์˜ ์ผ๋ถ€์ผ ์ˆ˜ ์žˆ๋Š” ๋ถ€๋ถ„์ ์œผ๋กœ ๋ถˆ๋ถ„๋ช…ํ•œ ์ผ์น˜(์˜ˆ: ์—‰๋ฉ์ด)๋Š” ํ”ผํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ถ€๋“œ๋Ÿฌ์šด ๋‹จ์–ด ๋ชฉ๋ก์—๋Š” ๋ฌธ์„œ๋ฅผ ํ๊ธฐํ•  ์ถฉ๋ถ„ํ•œ ์ด์œ ๊ฐ€ ๋˜์ง€๋Š” ์•Š์ง€๋งŒ ๋ชฉ๋ก์˜ ์—ฌ๋Ÿฌ ๋‹จ์–ด๊ฐ€ ์ผ์น˜ํ•  ๊ฒฝ์šฐ ์˜์‹ฌ์Šค๋Ÿฌ์šด ๋‹จ์–ด๋ฅผ ํฌํ•จ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์˜๋ฃŒ ๋˜๋Š” ๋ฒ•๋ฅ  ์ฝ˜ํ…์ธ ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๊ณ  ์œ ์ง€ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค(์˜ˆ: ๋”•์˜ ๋‹จ์ผ ์ผ์น˜).
We conduct a human study to verify the reliability of GPT-4's judgments, using the results of the TL:DR summarization experiment and two different GPT-4 prompts. The **GPT-4 (S)** (simple) prompt simply asks for which summary better-summarizes the important information in the post. The **GPT-4 (C)** (concise) prompt also asks for which summary is more concise; we evaluate this prompt because we find that GPT-4 prefers longer, more repetitive summaries than humans do with the **GPT-4 (S)** prompt. See Appendix C.2 for the complete prompts. We perform three comparisons, using the highest (DPO, temp. middle-performing (SFT, temp. 0.25) method with the aim of covering a diversity of sample qualities; all three methods are compared against greedily-sampled PPO (its best-performing temperature). We find that with both prompts, GPT-4 tends to agree with humans about as often as humans agree with each other, suggesting that GPT-4 is a reasonable proxy for human evaluations (due to limited human raters, we only collect multiple human judgments for the DPO and PPO-1 comparisons). Overall, the **GPT-4 (C)** prompt generally provides win rates more representative of humans; we therefore use this prompt for the main results in Section 6.2. For additional details about the human study, including the web interface presented to raters and the list of human volunteers, see Appendix D.3.
TL:DR ์š”์•ฝ ์‹คํ—˜์˜ ๊ฒฐ๊ณผ์™€ ๋‘ ๊ฐ€์ง€ ๋‹ค๋ฅธ GPT-4 prompts๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ GPT-4์˜ ํŒ๋‹จ์˜ ์‹ ๋ขฐ์„ฑ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. **GPT-4 (S)**(๋‹จ์ˆœ) prompt๋Š” ์–ด๋–ค ์š”์•ฝ์ด ๊ฒŒ์‹œ๋ฌผ์˜ ์ค‘์š”ํ•œ ์ •๋ณด๋ฅผ ๋” ์ž˜ ์š”์•ฝํ•˜๋Š”์ง€ ๋ฌป๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. **GPT-4 (C)**(๊ฐ„๊ฒฐ) prompt๋Š” ์–ด๋–ค ์š”์•ฝ์ด ๋” ๊ฐ„๊ฒฐํ•œ์ง€ ๋ฌป๋Š” ๊ฒƒ์œผ๋กœ, GPT-4๊ฐ€ **GPT-4 (S)** prompt์—์„œ ์ธ๊ฐ„๋ณด๋‹ค ๋” ๊ธธ๊ณ  ๋ฐ˜๋ณต์ ์ธ ์š”์•ฝ์„ ์„ ํ˜ธํ•˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ•˜์—ฌ ์ด prompt๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด prompts๋Š” ๋ถ€๋ก C.2๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ๋‹ค์–‘ํ•œ ์ƒ˜ํ”Œ ํ’ˆ์งˆ์„ ๋‹ค๋ฃจ๊ธฐ ์œ„ํ•ด ๊ฐ€์žฅ ๋†’์€(DPO, ์˜จ๋„ ์ค‘๊ฐ„ ์„ฑ๋Šฅ(SFT, ์˜จ๋„ 0.25) ๋ฐฉ๋ฒ•๊ณผ ์š•์‹ฌ ์ƒ˜ํ”Œ๋ง๋œ PPO(์ตœ๊ณ  ์„ฑ๋Šฅ ์˜จ๋„)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์„ธ ๊ฐ€์ง€ ๋น„๊ต๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋‘ prompts ๋ชจ๋‘์—์„œ GPT-4๋Š” ์ธ๊ฐ„์ด ์„œ๋กœ ๋™์˜ํ•˜๋Š” ๋นˆ๋„๋งŒํผ ์ธ๊ฐ„๊ณผ ๋™์˜ํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์–ด GPT-4๊ฐ€ ์ธ๊ฐ„ ํ‰๊ฐ€์˜ ํ•ฉ๋ฆฌ์ ์ธ ๋Œ€๋ฆฌ์ž์ž„์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค(์ธ๊ฐ„ ํ‰๊ฐ€์ž๊ฐ€ ์ œํ•œ๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— DPO ๋ฐ PPO-1 ๋น„๊ต์— ๋Œ€ํ•ด์„œ๋งŒ ์—ฌ๋Ÿฌ ์ธ๊ฐ„์˜ ํŒ๋‹จ์„ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค). ์ „๋ฐ˜์ ์œผ๋กœ **GPT-4 (C)** prompt๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„์„ ๋” ๋Œ€ํ‘œํ•˜๋Š” ์Šน๋ฅ ์„ ์ œ๊ณตํ•˜๋ฏ€๋กœ ์„น์…˜ 6.2์˜ ์ฃผ์š” ๊ฒฐ๊ณผ์—๋Š” ์ด prompt๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€์ž์—๊ฒŒ ์ œ์‹œ๋œ ์›น ์ธํ„ฐํŽ˜์ด์Šค์™€ ์ธ๊ฐ„ ์ง€์›์ž ๋ชฉ๋ก์„ ํฌํ•จํ•œ ์ธ๊ฐ„ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก D.3์„ ์ฐธ์กฐํ•˜์„ธ์š”.
TL:DR ์š”์•ฝ ์‹คํ—˜์˜ ๊ฒฐ๊ณผ์™€ ๋‘ ๊ฐ€์ง€ ๋‹ค๋ฅธ GPT-4 ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ GPT-4์˜ ํŒ๋‹จ์˜ ์‹ ๋ขฐ์„ฑ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. GPT-4 (S)**(๋‹จ์ˆœ) ํ”„๋กฌํ”„ํŠธ๋Š” ์–ด๋–ค ์š”์•ฝ์ด ๊ฒŒ์‹œ๋ฌผ์˜ ์ค‘์š”ํ•œ ์ •๋ณด๋ฅผ ๋” ์ž˜ ์š”์•ฝํ•˜๋Š”์ง€ ๋ฌป๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. GPT-4 (C)**(๊ฐ„๊ฒฐ) ํ”„๋กฌํ”„ํŠธ๋Š” ์–ด๋–ค ์š”์•ฝ์ด ๋” ๊ฐ„๊ฒฐํ•œ์ง€ ๋ฌป๋Š” ๊ฒƒ์œผ๋กœ, GPT-4๊ฐ€ **GPT-4 (S)** ํ”„๋กฌํ”„ํŠธ์—์„œ ์ธ๊ฐ„๋ณด๋‹ค ๋” ๊ธธ๊ณ  ๋ฐ˜๋ณต์ ์ธ ์š”์•ฝ์„ ์„ ํ˜ธํ•˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ•˜์—ฌ ์ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ํ”„๋กฌํ”„ํŠธ๋Š” ๋ถ€๋ก C.2๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ๋‹ค์–‘ํ•œ ์ƒ˜ํ”Œ ํ’ˆ์งˆ์„ ๋‹ค๋ฃจ๊ธฐ ์œ„ํ•ด ๊ฐ€์žฅ ๋†’์€(DPO, ์˜จ๋„ ์ค‘๊ฐ„ ์„ฑ๋Šฅ(SFT, ์˜จ๋„ 0.25) ๋ฐฉ๋ฒ•๊ณผ ์š•์‹ฌ ์ƒ˜ํ”Œ๋ง๋œ PPO(์ตœ๊ณ  ์„ฑ๋Šฅ ์˜จ๋„)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์„ธ ๊ฐ€์ง€ ๋น„๊ต๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋‘ ํ”„๋กฌํ”„ํŠธ ๋ชจ๋‘์—์„œ GPT-4๋Š” ์ธ๊ฐ„์ด ์„œ๋กœ ๋™์˜ํ•˜๋Š” ๋นˆ๋„๋งŒํผ ์ธ๊ฐ„๊ณผ ๋™์˜ํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์–ด GPT-4๊ฐ€ ์ธ๊ฐ„ ํ‰๊ฐ€์˜ ํ•ฉ๋ฆฌ์ ์ธ ๋Œ€๋ฆฌ์ž์ž„์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค(์ธ๊ฐ„ ํ‰๊ฐ€์ž๊ฐ€ ์ œํ•œ๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— DPO ๋ฐ PPO-1 ๋น„๊ต์— ๋Œ€ํ•ด์„œ๋งŒ ์—ฌ๋Ÿฌ ์ธ๊ฐ„์˜ ํŒ๋‹จ์„ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค). ์ „๋ฐ˜์ ์œผ๋กœ **GPT-4 (C)** ํ”„๋กฌํ”„ํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„์„ ๋” ๋Œ€ํ‘œํ•˜๋Š” ์Šน๋ฅ ์„ ์ œ๊ณตํ•˜๋ฏ€๋กœ ์„น์…˜ 6.2์˜ ์ฃผ์š” ๊ฒฐ๊ณผ์—๋Š” ์ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€์ž์—๊ฒŒ ์ œ์‹œ๋œ ์›น ์ธํ„ฐํŽ˜์ด์Šค์™€ ์ธ๊ฐ„ ์ง€์›์ž ๋ชฉ๋ก์„ ํฌํ•จํ•œ ์ธ๊ฐ„ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก D.3์„ ์ฐธ์กฐํ•˜์„ธ์š”.
Mathematical capabilities were previously believed to emerge in common language models only at a very large scale or require extensive math-related pre-training. This paper shows that the LLaMA-2 7B model with common pre-training already exhibits strong mathematical abilities, as evidenced by its impressive accuracy of 97.7% and 72.0% on the GSM8K and MATH benchmarks, respectively, when selecting the best response from 256 random generations. The primary issue with the current base model is the difficulty in consistently eliciting its inherent mathematical capabilities. Notably, the accuracy for the first answer drops to 49.5% and 7.9% on the GSM8K and MATH benchmarks, respectively. We find that simply scaling up the SFT data can significantly enhance the reliability of generating correct answers. However, the potential for extensive scaling is constrained by the scarcity of publicly available math questions. To overcome this limitation, we employ synthetic data, which proves to be nearly as effective as real data and shows no clear saturation when scaled up to approximately one million samples. This straightforward approach achieves an accuracy of 82.6% on GSM8K and 40.6% on MATH using LLaMA-2 7B models, surpassing previous models by 14.2% and 20.8%, respectively. We also provide insights into scaling behaviors across different reasoning complexities and error types.
์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์€ ์ด์ „์—๋Š” ๋งค์šฐ ํฐ ๊ทœ๋ชจ์—์„œ๋งŒ ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ๋‚˜ํƒ€๋‚˜๊ฑฐ๋‚˜ ๊ด‘๋ฒ”์œ„ํ•œ ์ˆ˜ํ•™ ๊ด€๋ จ ์‚ฌ์ „ ํ•™์Šต์ด ํ•„์š”ํ•œ ๊ฒƒ์œผ๋กœ ์—ฌ๊ฒจ์กŒ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ตœ๊ณ ์˜ ์‘๋‹ต์„ 256๊ฐœ์˜ ๋ฌด์ž‘์œ„ ์ƒ์„ฑ์—์„œ ์„ ํƒํ•  ๋•Œ GSM8K ๋ฐ MATH ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ฐ๊ฐ 97.7%์™€ 72.0%์˜ ์ธ์ƒ์ ์ธ ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ฃผ๋Š” ๋ฐ”์™€ ๊ฐ™์ด ์ผ๋ฐ˜์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์นœ LLaMA-2 7B ๋ชจ๋ธ์ด ์ด๋ฏธ ๊ฐ•๋ ฅํ•œ ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ˜„์žฌ ๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ์ฃผ์š” ๋ฌธ์ œ๋Š” ๋‚ด์žฌ๋œ ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์„ ์ผ๊ด€๋˜๊ฒŒ ์ด๋Œ์–ด๋‚ด๊ธฐ ์–ด๋ ต๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ ์ฒซ ๋ฒˆ์งธ ๋‹ต๋ณ€์˜ ์ •ํ™•๋„๋Š” GSM8K ๋ฐ MATH ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ฐ๊ฐ 49.5%์™€ 7.9%๋กœ ๋–จ์–ด์ง‘๋‹ˆ๋‹ค. ๋‹จ์ˆœํžˆ Supervised Fine-tuning (SFT) ๋ฐ์ดํ„ฐ๋ฅผ ํ™•์žฅํ•˜๋ฉด ์ •๋‹ต ์ƒ์„ฑ์˜ ์‹ ๋ขฐ์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ˆ˜ํ•™ ๋ฌธ์ œ์˜ ๋ถ€์กฑ์œผ๋กœ ์ธํ•ด ๊ด‘๋ฒ”์œ„ํ•œ ํ™•์žฅ์˜ ์ž ์žฌ๋ ฅ์ด ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š”๋ฐ, ์ด๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ์™€ ๊ฑฐ์˜ ๋™๋“ฑํ•œ ํšจ๊ณผ๋ฅผ ๋ฐœํœ˜ํ•˜๋ฉฐ ์•ฝ 100๋งŒ ๊ฐœ์˜ ์ƒ˜ํ”Œ๋กœ ํ™•์žฅํ•  ๋•Œ ํฌํ™” ์ƒํƒœ๊ฐ€ ๋‚˜ํƒ€๋‚˜์ง€ ์•Š๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ด ๊ฐ„๋‹จํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ LLaMA-2 7B ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ GSM8K์—์„œ 82.6%, MATH์—์„œ 40.6%์˜ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ ์ด์ „ ๋ชจ๋ธ๋ณด๋‹ค ๊ฐ๊ฐ 14.2%์™€ 20.8% ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ ์ถ”๋ก  ๋ณต์žก์„ฑ๊ณผ ์˜ค๋ฅ˜ ์œ ํ˜•์— ๋”ฐ๋ฅธ ํ™•์žฅ ๋™์ž‘์— ๋Œ€ํ•œ ์ธ์‚ฌ์ดํŠธ๋„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์€ ์ด์ „์—๋Š” ๋งค์šฐ ํฐ ๊ทœ๋ชจ์—์„œ๋งŒ ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ๋‚˜ํƒ€๋‚˜๊ฑฐ๋‚˜ ๊ด‘๋ฒ”์œ„ํ•œ ์ˆ˜ํ•™ ๊ด€๋ จ ์‚ฌ์ „ ํ•™์Šต์ด ํ•„์š”ํ•œ ๊ฒƒ์œผ๋กœ ์—ฌ๊ฒจ์กŒ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ตœ๊ณ ์˜ ์‘๋‹ต์„ 256๊ฐœ์˜ ๋ฌด์ž‘์œ„ ์ƒ์„ฑ์—์„œ ์„ ํƒํ•  ๋•Œ GSM8K ๋ฐ MATH ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ฐ๊ฐ 97.7%์™€ 72.0%์˜ ์ธ์ƒ์ ์ธ ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ฃผ๋Š” ๋ฐ”์™€ ๊ฐ™์ด ์ผ๋ฐ˜์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์นœ LLaMA-2 7B ๋ชจ๋ธ์ด ์ด๋ฏธ ๊ฐ•๋ ฅํ•œ ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ˜„์žฌ ๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ์ฃผ์š” ๋ฌธ์ œ๋Š” ๋‚ด์žฌ๋œ ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์„ ์ผ๊ด€๋˜๊ฒŒ ์ด๋Œ์–ด๋‚ด๊ธฐ ์–ด๋ ต๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ ์ฒซ ๋ฒˆ์งธ ๋‹ต๋ณ€์˜ ์ •ํ™•๋„๋Š” GSM8K ๋ฐ MATH ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ฐ๊ฐ 49.5%์™€ 7.9%๋กœ ๋–จ์–ด์ง‘๋‹ˆ๋‹ค. ๋‹จ์ˆœํžˆ SFT ๋ฐ์ดํ„ฐ๋ฅผ ํ™•์žฅํ•˜๋ฉด ์ •๋‹ต ์ƒ์„ฑ์˜ ์‹ ๋ขฐ์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ˆ˜ํ•™ ๋ฌธ์ œ์˜ ๋ถ€์กฑ์œผ๋กœ ์ธํ•ด ๊ด‘๋ฒ”์œ„ํ•œ ํ™•์žฅ์˜ ์ž ์žฌ๋ ฅ์ด ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š”๋ฐ, ์ด๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ์™€ ๊ฑฐ์˜ ๋™๋“ฑํ•œ ํšจ๊ณผ๋ฅผ ๋ฐœํœ˜ํ•˜๋ฉฐ ์•ฝ 100๋งŒ ๊ฐœ์˜ ์ƒ˜ํ”Œ๋กœ ํ™•์žฅํ•  ๋•Œ ํฌํ™” ์ƒํƒœ๊ฐ€ ๋‚˜ํƒ€๋‚˜์ง€ ์•Š๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ด ๊ฐ„๋‹จํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ LLaMA-2 7B ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ GSM8K์—์„œ 82.6%, MATH์—์„œ 40.6%์˜ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ ์ด์ „ ๋ชจ๋ธ๋ณด๋‹ค ๊ฐ๊ฐ 14.2%์™€ 20.8% ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ ์ถ”๋ก  ๋ณต์žก์„ฑ๊ณผ ์˜ค๋ฅ˜ ์œ ํ˜•์— ๋”ฐ๋ฅธ ํ™•์žฅ ๋™์ž‘์— ๋Œ€ํ•œ ์ธ์‚ฌ์ดํŠธ๋„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
With the emergence of the Mamba [12], the SSM model is increasingly attracting attention and favor from current researchers. The following works are currently explored in language modeling task [41][45][46][47][48], deep noise suppression task [42], and clinical note understanding task [49]. To be specific, for the language modeling task, [41] mainly studies the application of gated state spaces in the direction of long-range language modeling and introduces a novel method called GSS (Gated State Space). It can be used on long sequence modeling and effectively reduce the number of participants. Their experiments demonstrate that it achieves 2-3 times faster than the DSS [38]. Grazzi et al. [45] exploit the Mamba on simple function estimation and natural language processing in context learning tasks and validate that the overall performance is indeed better than the S4 version and comparable to other Transformer networks. S4++ [46] finds two issues of S4 architecture, i.e., the non-stationary state (NSS) and dependency bias, and proposes the State Memory Reply (SMR) mechanism to integrate multi-state information into the current state. They also integrate complex dependency bias via an interactive cross-attention mechanism and extensive experimental results show that S4++ outperforms S4 on multiple sequence modeling tasks, demonstrating significant performance gains.
๋ง˜๋ฐ”[12]์˜ ๋“ฑ์žฅ์œผ๋กœ SSM ๋ชจ๋ธ์€ ํ˜„์žฌ ์—ฐ๊ตฌ์ž๋“ค๋กœ๋ถ€ํ„ฐ ์ ์  ๋” ๋งŽ์€ ๊ด€์‹ฌ๊ณผ ํ˜ธํ‰์„ ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…[41][45][46][47][48], ์‹ฌ์ธต ์žก์Œ ์–ต์ œ ์ž‘์—…[42], ์ž„์ƒ ๋…ธํŠธ ์ดํ•ด ์ž‘์—…[49]์—์„œ ๋‹ค์Œ ์ž‘์—…์„ ์—ฐ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…์˜ ๊ฒฝ์šฐ [41]์€ ์ฃผ๋กœ ์žฅ๊ฑฐ๋ฆฌ ์–ธ์–ด ๋ชจ๋ธ๋ง ๋ฐฉํ–ฅ์—์„œ ๊ฒŒ์ดํŠธ ์ƒํƒœ ๊ณต๊ฐ„์˜ ์ ์šฉ์„ ์—ฐ๊ตฌํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, GSS(Gated State Space)๋ผ๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๊ธด ์‹œํ€€์Šค ๋ชจ๋ธ๋ง์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ์ฐธ๊ฐ€์ž ์ˆ˜๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์ด ๋ฐฉ๋ฒ•์€ DSS [38]๋ณด๋‹ค 2~3๋ฐฐ ๋” ๋น ๋ฅธ ์†๋„๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. Grazzi ๋“ฑ[45]์€ ๊ฐ„๋‹จํ•œ ํ•จ์ˆ˜ ์ถ”์ •๊ณผ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ In-Context Learning ์ž‘์—…์—์„œ ๋ง˜๋ฐ”๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ „์ฒด ์„ฑ๋Šฅ์ด ์‹ค์ œ๋กœ S4 ๋ฒ„์ „๋ณด๋‹ค ์šฐ์ˆ˜ํ•˜๊ณ  ๋‹ค๋ฅธ Transformer ๋„คํŠธ์›Œํฌ์™€ ๋น„์Šทํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๊ฒ€์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. S4++ [46]๋Š” S4 ์•„ํ‚คํ…์ฒ˜์˜ ๋‘ ๊ฐ€์ง€ ๋ฌธ์ œ, ์ฆ‰ ๋น„๊ณ ์ • ์ƒํƒœ(NSS)์™€ ์ข…์†์„ฑ ํŽธํ–ฅ์„ ๋ฐœ๊ฒฌํ•˜๊ณ , ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์ค‘ ์ƒํƒœ ์ •๋ณด๋ฅผ ํ˜„์žฌ ์ƒํƒœ์— ํ†ตํ•ฉํ•˜๋Š” State Memory Reply (SMR) ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ƒํ˜ธ ์ž‘์šฉํ˜• ๊ต์ฐจ Attention ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ†ตํ•ด ๋ณต์žกํ•œ ์ข…์†์„ฑ ํŽธํ–ฅ์„ ํ†ตํ•ฉํ–ˆ์œผ๋ฉฐ, ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ํ†ตํ•ด S4++๊ฐ€ ์—ฌ๋Ÿฌ ์‹œํ€€์Šค ๋ชจ๋ธ๋ง ์ž‘์—…์—์„œ S4๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์—ฌ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.
๋ง˜๋ฐ”[12]์˜ ๋“ฑ์žฅ์œผ๋กœ SSM ๋ชจ๋ธ์€ ํ˜„์žฌ ์—ฐ๊ตฌ์ž๋“ค๋กœ๋ถ€ํ„ฐ ์ ์  ๋” ๋งŽ์€ ๊ด€์‹ฌ๊ณผ ํ˜ธํ‰์„ ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…[41][45][46][47][48], ์‹ฌ์ธต ์žก์Œ ์–ต์ œ ์ž‘์—…[42], ์ž„์ƒ ๋…ธํŠธ ์ดํ•ด ์ž‘์—…[49]์—์„œ ๋‹ค์Œ ์ž‘์—…์„ ์—ฐ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…์˜ ๊ฒฝ์šฐ [41]์€ ์ฃผ๋กœ ์žฅ๊ฑฐ๋ฆฌ ์–ธ์–ด ๋ชจ๋ธ๋ง ๋ฐฉํ–ฅ์—์„œ ๊ฒŒ์ดํŠธ ์ƒํƒœ ๊ณต๊ฐ„์˜ ์ ์šฉ์„ ์—ฐ๊ตฌํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, GSS(๊ฒŒ์ดํŠธ ์ƒํƒœ ๊ณต๊ฐ„)๋ผ๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๊ธด ์‹œํ€€์Šค ๋ชจ๋ธ๋ง์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ์ฐธ๊ฐ€์ž ์ˆ˜๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์ด ๋ฐฉ๋ฒ•์€ DSS [38]๋ณด๋‹ค 2~3๋ฐฐ ๋” ๋น ๋ฅธ ์†๋„๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. Grazzi ๋“ฑ[45]์€ ๊ฐ„๋‹จํ•œ ํ•จ์ˆ˜ ์ถ”์ •๊ณผ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ปจํ…์ŠคํŠธ ํ•™์Šต ์ž‘์—…์—์„œ ๋ง˜๋ฐ”๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ „์ฒด ์„ฑ๋Šฅ์ด ์‹ค์ œ๋กœ S4 ๋ฒ„์ „๋ณด๋‹ค ์šฐ์ˆ˜ํ•˜๊ณ  ๋‹ค๋ฅธ ํŠธ๋žœ์Šคํฌ๋จธ ๋„คํŠธ์›Œํฌ์™€ ๋น„์Šทํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๊ฒ€์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. S4++ [46]๋Š” S4 ์•„ํ‚คํ…์ฒ˜์˜ ๋‘ ๊ฐ€์ง€ ๋ฌธ์ œ, ์ฆ‰ ๋น„๊ณ ์ • ์ƒํƒœ(NSS)์™€ ์ข…์†์„ฑ ํŽธํ–ฅ์„ ๋ฐœ๊ฒฌํ•˜๊ณ , ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์ค‘ ์ƒํƒœ ์ •๋ณด๋ฅผ ํ˜„์žฌ ์ƒํƒœ์— ํ†ตํ•ฉํ•˜๋Š” ์ƒํƒœ ๋ฉ”๋ชจ๋ฆฌ ์‘๋‹ต(SMR) ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ƒํ˜ธ ์ž‘์šฉํ˜• ๊ต์ฐจ ์ฃผ์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ†ตํ•ด ๋ณต์žกํ•œ ์ข…์†์„ฑ ํŽธํ–ฅ์„ ํ†ตํ•ฉํ–ˆ์œผ๋ฉฐ, ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ํ†ตํ•ด S4++๊ฐ€ ์—ฌ๋Ÿฌ ์‹œํ€€์Šค ๋ชจ๋ธ๋ง ์ž‘์—…์—์„œ S4๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์—ฌ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.
**Open-Book QA.** Unlike closed-book QA, in open-book QA tasks, LLMs can extract useful evidence from the external knowledge base or document collections, and then answer the question based on the extracted evidence [653, 654, 655]. Typical open-book QA datasets (_e.g.,_ Natural Questions [554], OpenBookQA [566], and SQuAD [569]) have overlap with closed-book QA datasets, but they incorporate external data sources, _e.g.,_ Wikipedia. The metrics of accuracy and F1 score are widely used in open-book QA tasks for evaluation. To select relevant knowledge from external resources, LLMs are often paired with a text retriever (or even a search engine), which is trained independently or jointly with LLMs [653, 657, 81]. Also, previous work [658, 659, 660] has indicated that retrievers can assist LLMs in verifying and rectifying the reasoning path. In evaluation, existing studies mainly focus on testing how LLMs utilize the extracted knowledge to answer the question and show that the retrieved evidence can largely improve the accuracy of the generated answers, even enabling a smaller LLM to outperform \(10\times\) larger ones [653, 657]. Further, open-book QA tasks can be also employed to evaluate the recency of knowledge information. Pre-training or retrieving from outdated knowledge resources may cause LLMs to generate incorrect answers for time-sensitive questions [653].
**Open-Book QA**. ํ์‡„ํ˜• ์‹œํ—˜์ง€ QA์™€ ๋‹ฌ๋ฆฌ, Open-Book QA ๊ณผ์ œ์—์„œ๋Š” LLMs์ด ์™ธ๋ถ€ ์ง€์‹์ฐฝ๊ณ ๋‚˜ ๋ฌธ์„œ ์ปฌ๋ ‰์…˜์—์„œ ์œ ์šฉํ•œ ์ฆ๊ฑฐ๋ฅผ ์ถ”์ถœํ•œ ๋‹ค์Œ, ์ถ”์ถœํ•œ ์ฆ๊ฑฐ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์งˆ๋ฌธ์— ๋‹ตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[653, 654, 655]. ์ผ๋ฐ˜์ ์ธ Open-Book QA ๋ฐ์ดํ„ฐ ์„ธํŠธ(_์˜ˆ: Natural Questions[554], OpenBookQA[566], SQuAD[569])๋Š” ํ์‡„ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๊ฒน์น˜์ง€๋งŒ, ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ ์†Œ์Šค(์˜ˆ: Wikipedia)๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ •ํ™•๋„์™€ F1 ์ ์ˆ˜ ๋ฉ”ํŠธ๋ฆญ์€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด Open-Book QA ์ž‘์—…์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์™ธ๋ถ€ ๋ฆฌ์†Œ์Šค์—์„œ ๊ด€๋ จ ์ง€์‹์„ ์„ ํƒํ•˜๊ธฐ ์œ„ํ•ด LLMs์€ ์ข…์ข… ํ…์ŠคํŠธ ๊ฒ€์ƒ‰๊ธฐ(๋˜๋Š” ๊ฒ€์ƒ‰ ์—”์ง„)์™€ ์ง์„ ์ด๋ฃจ๋ฉฐ, ์ด ๊ฒ€์ƒ‰๊ธฐ๋Š” ๋…๋ฆฝ์ ์œผ๋กœ ๋˜๋Š” LLMs๊ณผ ๊ณต๋™์œผ๋กœ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค[653, 657, 81]. ๋˜ํ•œ ์ด์ „ ์—ฐ๊ตฌ[658, 659, 660]์— ๋”ฐ๋ฅด๋ฉด ๊ฒ€์ƒ‰๊ธฐ๋Š” LLMs์ด ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ๊ฒ€์ฆํ•˜๊ณ  ์ˆ˜์ •ํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€์—์„œ ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ LLMs์ด ์ถ”์ถœํ•œ ์ง€์‹์„ ํ™œ์šฉํ•˜์—ฌ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์œผ๋ฉฐ, ๊ฒ€์ƒ‰๋œ ์ฆ๊ฑฐ๊ฐ€ ์ƒ์„ฑ๋œ ๋‹ต๋ณ€์˜ ์ •ํ™•๋„๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผœ ๋” ์ž‘์€ LLMs์ด 10๋ฐฐ ๋” ํฐ LLMs๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค[653, 657]. ๋˜ํ•œ Open-Book QA ๊ณผ์ œ๋Š” ์ง€์‹ ์ •๋ณด์˜ ์ตœ์‹ ์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ์—๋„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ค๋ž˜๋œ ์ง€์‹ ๋ฆฌ์†Œ์Šค๋ฅผ ์‚ฌ์ „ ํ•™์Šตํ•˜๊ฑฐ๋‚˜ ๊ฒ€์ƒ‰ํ•˜๋ฉด LLMs์ด ์‹œ๊ฐ„ ๋ฏผ๊ฐํ•œ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์ž˜๋ชป๋œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[653].
**์˜คํ”ˆ๋ถ QA**. ํ์‡„ํ˜• ์‹œํ—˜์ง€ QA์™€ ๋‹ฌ๋ฆฌ, ์˜คํ”ˆ๋ถ QA ๊ณผ์ œ์—์„œ๋Š” LLM์ด ์™ธ๋ถ€ ์ง€์‹์ฐฝ๊ณ ๋‚˜ ๋ฌธ์„œ ์ปฌ๋ ‰์…˜์—์„œ ์œ ์šฉํ•œ ์ฆ๊ฑฐ๋ฅผ ์ถ”์ถœํ•œ ๋‹ค์Œ, ์ถ”์ถœํ•œ ์ฆ๊ฑฐ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์งˆ๋ฌธ์— ๋‹ตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[653, 654, 655]. ์ผ๋ฐ˜์ ์ธ ์˜คํ”ˆ๋ถ QA ๋ฐ์ดํ„ฐ ์„ธํŠธ(_์˜ˆ: ์ž์—ฐ์–ด ์งˆ๋ฌธ[554], ์˜คํ”ˆ๋ถQA[566], SQuAD[569])๋Š” ํ์‡„ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๊ฒน์น˜์ง€๋งŒ, ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ ์†Œ์Šค(์˜ˆ: ์œ„ํ‚คํ”ผ๋””์•„)๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ •ํ™•๋„์™€ F1 ์ ์ˆ˜ ๋ฉ”ํŠธ๋ฆญ์€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์˜คํ”ˆ๋ถ QA ์ž‘์—…์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์™ธ๋ถ€ ๋ฆฌ์†Œ์Šค์—์„œ ๊ด€๋ จ ์ง€์‹์„ ์„ ํƒํ•˜๊ธฐ ์œ„ํ•ด LLM์€ ์ข…์ข… ํ…์ŠคํŠธ ๊ฒ€์ƒ‰๊ธฐ(๋˜๋Š” ๊ฒ€์ƒ‰ ์—”์ง„)์™€ ์ง์„ ์ด๋ฃจ๋ฉฐ, ์ด ๊ฒ€์ƒ‰๊ธฐ๋Š” ๋…๋ฆฝ์ ์œผ๋กœ ๋˜๋Š” LLM๊ณผ ๊ณต๋™์œผ๋กœ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค[653, 657, 81]. ๋˜ํ•œ ์ด์ „ ์—ฐ๊ตฌ[658, 659, 660]์— ๋”ฐ๋ฅด๋ฉด ๊ฒ€์ƒ‰๊ธฐ๋Š” LLM์ด ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ๊ฒ€์ฆํ•˜๊ณ  ์ˆ˜์ •ํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€์—์„œ ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ LLM์ด ์ถ”์ถœํ•œ ์ง€์‹์„ ํ™œ์šฉํ•˜์—ฌ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์œผ๋ฉฐ, ๊ฒ€์ƒ‰๋œ ์ฆ๊ฑฐ๊ฐ€ ์ƒ์„ฑ๋œ ๋‹ต๋ณ€์˜ ์ •ํ™•๋„๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผœ ๋” ์ž‘์€ LLM์ด 10๋ฐฐ ๋” ํฐ LLM๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค[653, 657]. ๋˜ํ•œ ์˜คํ”ˆ๋ถ QA ๊ณผ์ œ๋Š” ์ง€์‹ ์ •๋ณด์˜ ์ตœ์‹ ์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ์—๋„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ค๋ž˜๋œ ์ง€์‹ ๋ฆฌ์†Œ์Šค๋ฅผ ์‚ฌ์ „ ํ•™์Šตํ•˜๊ฑฐ๋‚˜ ๊ฒ€์ƒ‰ํ•˜๋ฉด LLM์ด ์‹œ๊ฐ„ ๋ฏผ๊ฐํ•œ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์ž˜๋ชป๋œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[653].
**Application and Ecosystem.** As LLMs have shown strong capacities in solving various tasks, they can be applied in a broad range of real-world applications (_i.e._, following task-specific natural language instructions). As a remarkable progress, ChatGPT has potentially changed the way how humans access information, which has been additionally integrated in the release of _New Bing_. Generally, in the near future, it can be foreseen that LLMs would have a significant impact on information-seeking techniques, including both search engines and recommender systems. Furthermore, LLMs make it possible to develop more intelligent systems (_e.g._, autonomous AI agents) to tackle various complex tasks in real-world scenarios. Specially, Assistants API has been launched by OpenAI (featured by instructions, knowledge and tool use), enabling rapid development of agent-like assistants within the applications. This wave of technical innovation would lead to an ecosystem of LLM-empowered applications (_e.g._, OpenAI's GPT Store), which has a close connection with human life. Lastly, the rise of LLMs sheds light on the exploration of artificial general intelligence (AGI). It is promising to develop more smart AI systems than ever. However, in this development process, AI safety should be one of the primary concerns, _i.e._, making AI lead to good for humanity but not bad [40].
**Application and Ecosystem**. LLMs์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ์—ญ๋Ÿ‰์„ ๋ณด์—ฌ ์™”๊ธฐ ๋•Œ๋ฌธ์— ๊ด‘๋ฒ”์œ„ํ•œ ์‹ค์ œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜(_์˜ˆ: ์ž‘์—…๋ณ„ ์ž์—ฐ์–ด instructions ์ค€์ˆ˜)์— ์ ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์ง„์ „์œผ๋กœ, ChatGPT๋Š” ์ธ๊ฐ„์ด ์ •๋ณด์— ์ ‘๊ทผํ•˜๋Š” ๋ฐฉ์‹์„ ์ž ์žฌ์ ์œผ๋กœ ๋ณ€ํ™”์‹œ์ผฐ์œผ๋ฉฐ, ์ด๋Š” _New Bing_ ์ถœ์‹œ์—์„œ ์ถ”๊ฐ€๋กœ ํ†ตํ•ฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐ€๊นŒ์šด ๋ฏธ๋ž˜์—๋Š” ๊ฒ€์ƒ‰ ์—”์ง„๊ณผ ์ถ”์ฒœ ์‹œ์Šคํ…œ์„ ํฌํ•จํ•œ ์ •๋ณด ๊ฒ€์ƒ‰ ๊ธฐ์ˆ ์— LLMs์ด ์ƒ๋‹นํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น  ๊ฒƒ์œผ๋กœ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ LLMs์€ ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋‹ค์–‘ํ•œ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๋ณด๋‹ค ์ง€๋Šฅ์ ์ธ ์‹œ์Šคํ…œ(_์˜ˆ: ์ž์œจ AI ์—์ด์ „ํŠธ)์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. ํŠนํžˆ, OpenAI์—์„œ Assistant API(instructions, ์ง€์‹ ๋ฐ ๋„๊ตฌ ์‚ฌ์šฉ ๊ธฐ๋Šฅ์ด ํŠน์ง•)๋ฅผ ์ถœ์‹œํ•˜์—ฌ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋‚ด์—์„œ ์—์ด์ „ํŠธ์™€ ๊ฐ™์€ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ๋น ๋ฅด๊ฒŒ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ  ํ˜์‹ ์˜ ๋ฌผ๊ฒฐ์€ ์ธ๊ฐ„์˜ ์‚ถ๊ณผ ๋ฐ€์ ‘ํ•œ ๊ด€๋ จ์ด ์žˆ๋Š” LLMs ๊ธฐ๋ฐ˜ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ์—์ฝ”์‹œ์Šคํ…œ(_์˜ˆ: OpenAI์˜ GPT Store)์œผ๋กœ ์ด์–ด์งˆ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, LLMs์˜ ๋ถ€์ƒ์€ ์ธ๊ณต ์ผ๋ฐ˜ ์ง€๋Šฅ(AGI)์˜ ํƒ๊ตฌ์— ๋น›์„ ๋น„์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์–ด๋Š ๋•Œ๋ณด๋‹ค ๋” ๋˜‘๋˜‘ํ•œ AI ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐœ์ „ ๊ณผ์ •์—์„œ AI ์•ˆ์ „, ์ฆ‰ AI๊ฐ€ ์ธ๋ฅ˜์—๊ฒŒ ํ•ด๊ฐ€ ๋˜์ง€ ์•Š๊ณ  ์„ ํ•œ ๋ฐฉํ–ฅ์œผ๋กœ ์ด๋Œ ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๊ด€์‹ฌ์‚ฌ ์ค‘ ํ•˜๋‚˜๊ฐ€ ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค[40].
**์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋ฐ ์—์ฝ”์‹œ์Šคํ…œ**. LLM์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ์—ญ๋Ÿ‰์„ ๋ณด์—ฌ ์™”๊ธฐ ๋•Œ๋ฌธ์— ๊ด‘๋ฒ”์œ„ํ•œ ์‹ค์ œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜(_์˜ˆ: ์ž‘์—…๋ณ„ ์ž์—ฐ์–ด ์ง€์นจ ์ค€์ˆ˜)์— ์ ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์ง„์ „์œผ๋กœ, ChatGPT๋Š” ์ธ๊ฐ„์ด ์ •๋ณด์— ์ ‘๊ทผํ•˜๋Š” ๋ฐฉ์‹์„ ์ž ์žฌ์ ์œผ๋กœ ๋ณ€ํ™”์‹œ์ผฐ์œผ๋ฉฐ, ์ด๋Š” _New Bing_ ์ถœ์‹œ์—์„œ ์ถ”๊ฐ€๋กœ ํ†ตํ•ฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐ€๊นŒ์šด ๋ฏธ๋ž˜์—๋Š” ๊ฒ€์ƒ‰ ์—”์ง„๊ณผ ์ถ”์ฒœ ์‹œ์Šคํ…œ์„ ํฌํ•จํ•œ ์ •๋ณด ๊ฒ€์ƒ‰ ๊ธฐ์ˆ ์— LLM์ด ์ƒ๋‹นํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น  ๊ฒƒ์œผ๋กœ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์€ ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋‹ค์–‘ํ•œ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๋ณด๋‹ค ์ง€๋Šฅ์ ์ธ ์‹œ์Šคํ…œ(_์˜ˆ: ์ž์œจ AI ์—์ด์ „ํŠธ)์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. ํŠนํžˆ, OpenAI์—์„œ Assistant API(๋ช…๋ น์–ด, ์ง€์‹ ๋ฐ ๋„๊ตฌ ์‚ฌ์šฉ ๊ธฐ๋Šฅ์ด ํŠน์ง•)๋ฅผ ์ถœ์‹œํ•˜์—ฌ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋‚ด์—์„œ ์—์ด์ „ํŠธ์™€ ๊ฐ™์€ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ๋น ๋ฅด๊ฒŒ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ  ํ˜์‹ ์˜ ๋ฌผ๊ฒฐ์€ ์ธ๊ฐ„์˜ ์‚ถ๊ณผ ๋ฐ€์ ‘ํ•œ ๊ด€๋ จ์ด ์žˆ๋Š” LLM ๊ธฐ๋ฐ˜ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ์—์ฝ”์‹œ์Šคํ…œ(_์˜ˆ: OpenAI์˜ GPT Store)์œผ๋กœ ์ด์–ด์งˆ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, LLM์˜ ๋ถ€์ƒ์€ ์ธ๊ณต ์ผ๋ฐ˜ ์ง€๋Šฅ(AGI)์˜ ํƒ๊ตฌ์— ๋น›์„ ๋น„์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์–ด๋Š ๋•Œ๋ณด๋‹ค ๋” ๋˜‘๋˜‘ํ•œ AI ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐœ์ „ ๊ณผ์ •์—์„œ AI ์•ˆ์ „, ์ฆ‰ AI๊ฐ€ ์ธ๋ฅ˜์—๊ฒŒ ํ•ด๊ฐ€ ๋˜์ง€ ์•Š๊ณ  ์„ ํ•œ ๋ฐฉํ–ฅ์œผ๋กœ ์ด๋Œ ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๊ด€์‹ฌ์‚ฌ ์ค‘ ํ•˜๋‚˜๊ฐ€ ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค[40].
Retrieval setup details.By default, we use Contriever-MS MARCO to retrieve the top five documents from Wikipedia, and use official Wikipedia embeddings based on 2018 English Wikipedia. On PopQA, where question and answer pairs are created based on WikiData in 2022, we found that the 2018 Wikipedia sometimes lacks articles about some entities that have been more recently added to Wikipedia. Therefore, for PopQA, we used the December 2020 preprocessed Wikipedia corpus provided by Izacard et al. (2022b) and generated document embeddings.8 The issues of performance variance from different Wikipedia dumps have been reported by prior work (Asai et al., 2020; Izacard et al., 2022b). Yet, we observe limited effectiveness of such off-the-shelf retrieval models trained primarily on knowledge-intensive tasks for open-ended generation (e.g., instruction following). Recent or concurrent work studies instruction-tuning of retrieval systems (Asai et al., 2023b) or joint training of retrieval and LM components (Lin et al., 2023), while we leave exploring the effectiveness of such approaches for future work. For bio generation and open-domain QA tasks, we additionally retrieve five documents using Google Programmable Search9 and search documents from English Wikipedia. As this API only provides snippets, we retrieve Wikipedia introductory paragraphs for the corresponding entities.
๊ฒ€์ƒ‰ ์„ค์ • ์„ธ๋ถ€ ์ •๋ณด. ๊ธฐ๋ณธ์ ์œผ๋กœ ์œ„ํ‚คํ”ผ๋””์•„์—์„œ ์ƒ์œ„ 5๊ฐœ ๋ฌธ์„œ๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ธฐ ์œ„ํ•ด Contriever-MS MARCO๋ฅผ ์‚ฌ์šฉํ•˜๊ณ , 2018๋…„ ์˜์–ด ์œ„ํ‚คํ”ผ๋””์•„๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๊ณต์‹ ์œ„ํ‚คํ”ผ๋””์•„ ์ž„๋ฒ ๋”ฉ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. 2022๋…„์— ์œ„ํ‚ค๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€ ์Œ์ด ์ƒ์„ฑ๋˜๋Š” PopQA์—์„œ๋Š” 2018๋…„ ์œ„ํ‚คํ”ผ๋””์•„์— ์ตœ๊ทผ์— ์œ„ํ‚คํ”ผ๋””์•„์— ์ถ”๊ฐ€๋œ ์ผ๋ถ€ ๊ฐœ์ฒด์— ๋Œ€ํ•œ ๋ฌธ์„œ๊ฐ€ ๋ถ€์กฑํ•œ ๊ฒฝ์šฐ๊ฐ€ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ PopQA์˜ ๊ฒฝ์šฐ, Izacard ๋“ฑ(2022b)์ด ์ œ๊ณตํ•œ 2020๋…„ 12์›” ์ „์ฒ˜๋ฆฌ๋œ ์œ„ํ‚คํ”ผ๋””์•„ ๋ง๋ญ‰์น˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ์„œ ์ž„๋ฒ ๋”ฉ์„ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์œ„ํ‚คํ”ผ๋””์•„ ๋คํ”„์—์„œ ์„ฑ๋Šฅ ์ฐจ์ด ๋ฌธ์ œ๋Š” ์ด์ „ ์ž‘์—…(Asai et al., 2020; Izacard et al., 2022b)์—์„œ ๋ณด๊ณ ๋œ ๋ฐ” ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฃผ๋กœ ์ง€์‹ ์ง‘์•ฝ์ ์ธ ์ž‘์—…(์˜ˆ: Instruction ์ˆ˜ํ–‰)์— ๋Œ€ํ•ด ํ•™์Šต๋œ ์ด๋Ÿฌํ•œ ๊ธฐ์„ฑ ๊ฒ€์ƒ‰ ๋ชจ๋ธ์˜ ํšจ๊ณผ๋Š” ์ œํ•œ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋˜๋Š” ๋™์‹œ์— ์ง„ํ–‰๋˜๋Š” ์ž‘์—…์—์„œ๋Š” ๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ์˜ Instruction Tuning(Asai et al., 2023b) ๋˜๋Š” ๊ฒ€์ƒ‰ ๋ฐ LM ๊ตฌ์„ฑ ์š”์†Œ์˜ ๊ณต๋™ ํ•™์Šต(Lin et al., 2023)์„ ์—ฐ๊ตฌํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ํ–ฅํ›„ ์ž‘์—…์— ๋Œ€ํ•œ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํšจ๊ณผ๋ฅผ ํƒ์ƒ‰ํ•˜๋Š” ๊ฒƒ์„ ๋‚จ๊ฒจ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ”์ด์˜ค ์ƒ์„ฑ ๋ฐ ์˜คํ”ˆ ๋„๋ฉ”์ธ QA ์ž‘์—…์˜ ๊ฒฝ์šฐ, Google Programmable Search๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 5๊ฐœ์˜ ๋ฌธ์„œ๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ณ  ์˜์–ด ์œ„ํ‚คํ”ผ๋””์•„์˜ ๊ฒ€์ƒ‰ ๋ฌธ์„œ๋ฅผ ์ถ”๊ฐ€๋กœ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ์ด API๋Š” ์Šค๋‹ˆํŽซ๋งŒ ์ œ๊ณตํ•˜๋ฏ€๋กœ ํ•ด๋‹น ์—”ํ‹ฐํ‹ฐ์— ๋Œ€ํ•œ ์œ„ํ‚คํ”ผ๋””์•„ ์†Œ๊ฐœ ๋‹จ๋ฝ์„ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค.
๊ฒ€์ƒ‰ ์„ค์ • ์„ธ๋ถ€ ์ •๋ณด.๊ธฐ๋ณธ์ ์œผ๋กœ ์œ„ํ‚คํ”ผ๋””์•„์—์„œ ์ƒ์œ„ 5๊ฐœ ๋ฌธ์„œ๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ธฐ ์œ„ํ•ด Contriever-MS MARCO๋ฅผ ์‚ฌ์šฉํ•˜๊ณ , 2018๋…„ ์˜์–ด ์œ„ํ‚คํ”ผ๋””์•„๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๊ณต์‹ ์œ„ํ‚คํ”ผ๋””์•„ ์ž„๋ฒ ๋”ฉ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. 2022๋…„์— ์œ„ํ‚ค๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€ ์Œ์ด ์ƒ์„ฑ๋˜๋Š” PopQA์—์„œ๋Š” 2018๋…„ ์œ„ํ‚คํ”ผ๋””์•„์— ์ตœ๊ทผ์— ์œ„ํ‚คํ”ผ๋””์•„์— ์ถ”๊ฐ€๋œ ์ผ๋ถ€ ๊ฐœ์ฒด์— ๋Œ€ํ•œ ๋ฌธ์„œ๊ฐ€ ๋ถ€์กฑํ•œ ๊ฒฝ์šฐ๊ฐ€ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ PopQA์˜ ๊ฒฝ์šฐ, Izacard ๋“ฑ(2022b)์ด ์ œ๊ณตํ•œ 2020๋…„ 12์›” ์ „์ฒ˜๋ฆฌ๋œ ์œ„ํ‚คํ”ผ๋””์•„ ๋ง๋ญ‰์น˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ์„œ ์ž„๋ฒ ๋”ฉ์„ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.8 ๋‹ค์–‘ํ•œ ์œ„ํ‚คํ”ผ๋””์•„ ๋คํ”„์—์„œ ์„ฑ๋Šฅ ์ฐจ์ด ๋ฌธ์ œ๋Š” ์ด์ „ ์ž‘์—…(Asai et al., 2020; Izacard et al., 2022b)์—์„œ ๋ณด๊ณ ๋œ ๋ฐ” ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฃผ๋กœ ์ง€์‹ ์ง‘์•ฝ์ ์ธ ์ž‘์—…(์˜ˆ: ์ง€์นจ ์ˆ˜ํ–‰)์— ๋Œ€ํ•ด ํ•™์Šต๋œ ์ด๋Ÿฌํ•œ ๊ธฐ์„ฑ ๊ฒ€์ƒ‰ ๋ชจ๋ธ์˜ ํšจ๊ณผ๋Š” ์ œํ•œ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋˜๋Š” ๋™์‹œ์— ์ง„ํ–‰๋˜๋Š” ์ž‘์—…์—์„œ๋Š” ๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ์˜ ์ง€์นจ ์กฐ์ •(Asai et al., 2023b) ๋˜๋Š” ๊ฒ€์ƒ‰ ๋ฐ LM ๊ตฌ์„ฑ ์š”์†Œ์˜ ๊ณต๋™ ํ•™์Šต(Lin et al., 2023)์„ ์—ฐ๊ตฌํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ํ–ฅํ›„ ์ž‘์—…์— ๋Œ€ํ•œ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํšจ๊ณผ๋ฅผ ํƒ์ƒ‰ํ•˜๋Š” ๊ฒƒ์„ ๋‚จ๊ฒจ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ”์ด์˜ค ์ƒ์„ฑ ๋ฐ ์˜คํ”ˆ ๋„๋ฉ”์ธ QA ์ž‘์—…์˜ ๊ฒฝ์šฐ, Google ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ฐ€๋Šฅ ๊ฒ€์ƒ‰9๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 5๊ฐœ์˜ ๋ฌธ์„œ๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ณ  ์˜์–ด ์œ„ํ‚คํ”ผ๋””์•„์˜ ๊ฒ€์ƒ‰ ๋ฌธ์„œ๋ฅผ ์ถ”๊ฐ€๋กœ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ์ด API๋Š” ์Šค๋‹ˆํŽซ๋งŒ ์ œ๊ณตํ•˜๋ฏ€๋กœ ํ•ด๋‹น ์—”ํ‹ฐํ‹ฐ์— ๋Œ€ํ•œ ์œ„ํ‚คํ”ผ๋””์•„ ์†Œ๊ฐœ ๋‹จ๋ฝ์„ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค.
**Retrieval-Augmented Language Models** RAG enhances language models by integrating a retrieval module that sources relevant information from external knowledge bases, significantly improving performance across various NLP tasks, including language modeling (Guu et al., 2020; Borgeaud et al., 2022; Khandelwal et al., 2019; Shi et al., 2023; Lin et al., 2023; Shi et al., 2023; Asai et al., 2023; Xu et al., 2023; Wang et al., 2023) and open-domain question answering (Izacard et al., 2023; Lewis et al., 2020). This integration follows a "retrieve-and-read" paradigm where the retrieval module provides additional context from external sources, which the LM then uses to generate the final output. The retrieval process involves using the input as a query to fetch documents, which the LM incorporates for final predictions. For instance, Atlas (Izacard et al., 2023) fine-tunes T5 models with the retriever, treating documents as latent variables, while RETRO (Borgeaud et al., 2022) modifies the decoder-only architecture to include retrieved texts and conducts pre-training from scratch. kNN-LM (Khandelwal et al., 2019) interpolates between the LM's next token distribution and distributions computed from retrieved tokens at inference. (Shi et al., 2023; Ram et al., 2023) assume black-box access to an LM and combine it with either off-the-shelf or fine-tuned retriever.
**Retrieval-Augmented Language Models** RAG๋Š” ์™ธ๋ถ€ ์ง€์‹ ๋ฒ ์ด์Šค์—์„œ ๊ด€๋ จ ์ •๋ณด๋ฅผ ์†Œ์‹ฑํ•˜๋Š” retrieval ๋ชจ๋“ˆ์„ ํ†ตํ•ฉํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์„ ํ–ฅ์ƒ์‹œ์ผœ ์–ธ์–ด ๋ชจ๋ธ๋ง(Guu et al., 2020; Borgeaud et al., 2022; Khandelwal et al., 2019; Shi et al., 2023; Lin et al., 2023; Shi et al., 2023; Asai et al., 2023; Xu et al., 2023; Wang et al., 2023) ๋ฐ ์˜คํ”ˆ ๋„๋ฉ”์ธ ์งˆ๋ฌธ ๋‹ต๋ณ€(Izacard et al., 2023; Lewis et al., 2020)์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ†ตํ•ฉ์€ retrieval ๋ชจ๋“ˆ์ด ์™ธ๋ถ€ ์†Œ์Šค์—์„œ ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๋ฅผ ์ œ๊ณตํ•˜๊ณ , LM์ด ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข… ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๋Š” 'retrieve-and-read' ํŒจ๋Ÿฌ๋‹ค์ž„์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. retrieval ํ”„๋กœ์„ธ์Šค์—๋Š” ์ž…๋ ฅ์„ ์ฟผ๋ฆฌ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ์„œ๋ฅผ ๊ฐ€์ ธ์˜ค๋Š” ๊ฒƒ์œผ๋กœ, LM์ด ์ตœ์ข… ์˜ˆ์ธก์„ ์œ„ํ•ด ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Atlas(Izacard et al., 2023)๋Š” retriever๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ T5 ๋ชจ๋ธ์„ Fine-tuningํ•˜๊ณ  ๋ฌธ์„œ๋ฅผ ์ž ์žฌ ๋ณ€์ˆ˜๋กœ ์ทจ๊ธ‰ํ•˜๋Š” ๋ฐ˜๋ฉด, RETRO(Borgeaud et al., 2022)๋Š” retrieval๋œ ํ…์ŠคํŠธ๋ฅผ ํฌํ•จํ•˜๋„๋ก decoder ์ „์šฉ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ˆ˜์ •ํ•˜๊ณ  ์ฒ˜์Œ๋ถ€ํ„ฐ ์‚ฌ์ „ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. kNN-LM(Khandelwal et al., 2019)์€ ์ถ”๋ก  ์‹œ LM์˜ ๋‹ค์Œ ํ† ํฐ ๋ถ„ํฌ์™€ retrieval๋œ ํ† ํฐ์—์„œ ๊ณ„์‚ฐ๋œ ๋ถ„ํฌ ์‚ฌ์ด๋ฅผ ๋ณด๊ฐ„ํ•ฉ๋‹ˆ๋‹ค. (Shi et al., 2023; Ram et al., 2023)์€ LM์— ๋Œ€ํ•œ ๋ธ”๋ž™๋ฐ•์Šค ์•ก์„ธ์Šค๋ฅผ ๊ฐ€์ •ํ•˜๊ณ  ์ด๋ฅผ ๊ธฐ์„ฑ ๋˜๋Š” Fine-tuning๋œ retriever์™€ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค.
**๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์–ธ์–ด ๋ชจ๋ธ** RAG๋Š” ์™ธ๋ถ€ ์ง€์‹ ๋ฒ ์ด์Šค์—์„œ ๊ด€๋ จ ์ •๋ณด๋ฅผ ์†Œ์‹ฑํ•˜๋Š” ๊ฒ€์ƒ‰ ๋ชจ๋“ˆ์„ ํ†ตํ•ฉํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์„ ํ–ฅ์ƒ์‹œ์ผœ ์–ธ์–ด ๋ชจ๋ธ๋ง(Guu et al., 2020; Borgeaud et al., 2022; Khandelwal et al., 2019; Shi et al., 2023; Lin et al., 2023; Shi et al., 2023; Asai et al., 2023; Xu et al., 2023; Wang et al., 2023) ๋ฐ ์˜คํ”ˆ ๋„๋ฉ”์ธ ์งˆ๋ฌธ ๋‹ต๋ณ€(Izacard et al., 2023; Lewis et al., 2020)์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ†ตํ•ฉ์€ ๊ฒ€์ƒ‰ ๋ชจ๋“ˆ์ด ์™ธ๋ถ€ ์†Œ์Šค์—์„œ ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๋ฅผ ์ œ๊ณตํ•˜๊ณ , LM์ด ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข… ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๋Š” '๊ฒ€์ƒ‰ ๋ฐ ์ฝ๊ธฐ' ํŒจ๋Ÿฌ๋‹ค์ž„์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰ ํ”„๋กœ์„ธ์Šค์—๋Š” ์ž…๋ ฅ์„ ์ฟผ๋ฆฌ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ์„œ๋ฅผ ๊ฐ€์ ธ์˜ค๋Š” ๊ฒƒ์œผ๋กœ, LM์ด ์ตœ์ข… ์˜ˆ์ธก์„ ์œ„ํ•ด ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Atlas(Izacard et al., 2023)๋Š” ๊ฒ€์ƒ‰๊ธฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ T5 ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๊ณ  ๋ฌธ์„œ๋ฅผ ์ž ์žฌ ๋ณ€์ˆ˜๋กœ ์ทจ๊ธ‰ํ•˜๋Š” ๋ฐ˜๋ฉด, RETRO(Borgeaud et al., 2022)๋Š” ๊ฒ€์ƒ‰๋œ ํ…์ŠคํŠธ๋ฅผ ํฌํ•จํ•˜๋„๋ก ๋””์ฝ”๋” ์ „์šฉ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ˆ˜์ •ํ•˜๊ณ  ์ฒ˜์Œ๋ถ€ํ„ฐ ์‚ฌ์ „ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. kNN-LM(Khandelwal et al., 2019)์€ ์ถ”๋ก  ์‹œ LM์˜ ๋‹ค์Œ ํ† ํฐ ๋ถ„ํฌ์™€ ๊ฒ€์ƒ‰๋œ ํ† ํฐ์—์„œ ๊ณ„์‚ฐ๋œ ๋ถ„ํฌ ์‚ฌ์ด๋ฅผ ๋ณด๊ฐ„ํ•ฉ๋‹ˆ๋‹ค. (Shi et al., 2023; Ram et al., 2023)์€ LM์— ๋Œ€ํ•œ ๋ธ”๋ž™๋ฐ•์Šค ์•ก์„ธ์Šค๋ฅผ ๊ฐ€์ •ํ•˜๊ณ  ์ด๋ฅผ ๊ธฐ์„ฑ ๋˜๋Š” ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ฆฌํŠธ๋ฆฌ๋ฒ„์™€ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค.
As AutoEncoder language models excel at Natural Language Understanding (NLU) tasks, many RALM architectures(Thulke et al., 2021) (Cheng et al., 2024) (Sachan et al., 2021)utilise them for specific tasks, such as judgement. One of the most commonly used models is BERT and its improved versions.Devlin et al. (2018) proposed the BERT model, which was inspired by closed tasks(Taylor, 1953). RoBERTa(Liu et al., 2019) is trained using dynamic masking, full sentences without NSP loss, large mini-batches, and a larger byte-level BPE to address the lack of training of Bert's model. According to Jiang et al. (2020), BERT heavily relies on global self-attention blocks, resulting in a large memory footprint and computational cost. Although all attention heads query the entire input sequence, some only need to learn local dependencies, leading to computational redundancy. To address this issue, They proposed a new span-based dynamic convolution to replace these self-attention heads and directly model local dependencies. The new convolutional head, along with other self-attentive heads, forms a hybrid attention block. Furthermore, Sanh et al. (2019) was able to decrease the size of the BERT model by 40% while maintaining 97% of its language comprehension abilities and achieving a 60% increase in speed by implementing knowledge distillation during the pre-training phase.
์ž๋™ Encoder ์–ธ์–ด ๋ชจ๋ธ์€ ์ž์—ฐ์–ด ์ดํ•ด(NLU) ์ž‘์—…์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๊ธฐ ๋•Œ๋ฌธ์—, ๋งŽ์€ RALM ์•„ํ‚คํ…์ฒ˜(Thulke et al., 2021)(Cheng et al., 2024)(Sachan et al., 2021)๊ฐ€ ํŒ๋‹จ๊ณผ ๊ฐ™์€ ํŠน์ • ์ž‘์—…์— ์ด๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜๋Š” BERT์™€ ๊ทธ ๊ฐœ์„ ๋œ ๋ฒ„์ „์ž…๋‹ˆ๋‹ค.Devlin et al. (2018)์€ ํ์‡„ํ˜• ๊ณผ์ œ(Taylor, 1953)์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ BERT ๋ชจ๋ธ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. RoBERTa(Liu et al., 2019)๋Š” ๋™์  ๋งˆ์Šคํ‚น, NSP Loss๊ฐ€ ์—†๋Š” ์ „์ฒด ๋ฌธ์žฅ, ๋Œ€๊ทœ๋ชจ ๋ฏธ๋‹ˆ ๋ฐฐ์น˜, ๋” ํฐ ๋ฐ”์ดํŠธ ์ˆ˜์ค€ BPE๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•˜์—ฌ BERT ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ๋ถ€์กฑ์„ ํ•ด๊ฒฐํ–ˆ์Šต๋‹ˆ๋‹ค. Jiang et al. (2020)์— ๋”ฐ๋ฅด๋ฉด BERT๋Š” ๊ธ€๋กœ๋ฒŒ self-attention ๋ธ”๋ก์— ํฌ๊ฒŒ ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰๊ณผ ๊ณ„์‚ฐ ๋น„์šฉ์ด ํฝ๋‹ˆ๋‹ค. ๋ชจ๋“  attention ํ—ค๋“œ๊ฐ€ ์ „์ฒด ์ž…๋ ฅ ์‹œํ€€์Šค๋ฅผ ์ฟผ๋ฆฌํ•˜์ง€๋งŒ ์ผ๋ถ€๋Š” ๋กœ์ปฌ ์ข…์†์„ฑ๋งŒ ํ•™์Šตํ•˜๋ฉด ๋˜๊ธฐ ๋•Œ๋ฌธ์— ๊ณ„์‚ฐ ์ค‘๋ณต์ด ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ด๋“ค์€ ์ด๋Ÿฌํ•œ self-attention ํ—ค๋“œ๋ฅผ ๋Œ€์ฒดํ•˜๊ณ  ๋กœ์ปฌ ์ข…์†์„ฑ์„ ์ง์ ‘ ๋ชจ๋ธ๋งํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด ์ŠคํŒฌ ๊ธฐ๋ฐ˜ ๋™์  ์ปจ๋ณผ๋ฃจ์…˜์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ƒˆ๋กœ์šด ์ปจ๋ณผ๋ฃจ์…˜ ํ—ค๋“œ๋Š” ๋‹ค๋ฅธ self-attentive ํ—ค๋“œ์™€ ํ•จ๊ป˜ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ attention ๋ธ”๋ก์„ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Sanh et al. (2019)์€ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์—์„œ knowledge distillation์„ ๊ตฌํ˜„ํ•˜์—ฌ ์–ธ์–ด ์ดํ•ด ๋Šฅ๋ ฅ์˜ 97%๋ฅผ ์œ ์ง€ํ•˜๋ฉด์„œ BERT ๋ชจ๋ธ์˜ ํฌ๊ธฐ๋ฅผ 40% ์ค„์ด๊ณ  ์†๋„๋ฅผ 60% ๋†’์ผ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
์ž๋™ ์ธ์ฝ”๋” ์–ธ์–ด ๋ชจ๋ธ์€ ์ž์—ฐ์–ด ์ดํ•ด(NLU) ์ž‘์—…์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๊ธฐ ๋•Œ๋ฌธ์—, ๋งŽ์€ RALM ์•„ํ‚คํ…์ฒ˜(Thulke et al., 2021)(Cheng et al., 2024)(Sachan et al., 2021)๊ฐ€ ํŒ๋‹จ๊ณผ ๊ฐ™์€ ํŠน์ • ์ž‘์—…์— ์ด๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜๋Š” BERT์™€ ๊ทธ ๊ฐœ์„ ๋œ ๋ฒ„์ „์ž…๋‹ˆ๋‹ค.Devlin et al. (2018)์€ ํ์‡„ํ˜• ๊ณผ์ œ(Taylor, 1953)์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ BERT ๋ชจ๋ธ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. RoBERTa(Liu et al., 2019)๋Š” ๋™์  ๋งˆ์Šคํ‚น, NSP ์†์‹ค์ด ์—†๋Š” ์ „์ฒด ๋ฌธ์žฅ, ๋Œ€๊ทœ๋ชจ ๋ฏธ๋‹ˆ ๋ฐฐ์น˜, ๋” ํฐ ๋ฐ”์ดํŠธ ์ˆ˜์ค€ BPE๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•˜์—ฌ Bert ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ๋ถ€์กฑ์„ ํ•ด๊ฒฐํ–ˆ์Šต๋‹ˆ๋‹ค. Jiang et al. (2020)์— ๋”ฐ๋ฅด๋ฉด BERT๋Š” ๊ธ€๋กœ๋ฒŒ ์ž๊ธฐ ์ฃผ์˜ ๋ธ”๋ก์— ํฌ๊ฒŒ ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰๊ณผ ๊ณ„์‚ฐ ๋น„์šฉ์ด ํฝ๋‹ˆ๋‹ค. ๋ชจ๋“  ์ฃผ์˜ ํ—ค๋“œ๊ฐ€ ์ „์ฒด ์ž…๋ ฅ ์‹œํ€€์Šค๋ฅผ ์ฟผ๋ฆฌํ•˜์ง€๋งŒ ์ผ๋ถ€๋Š” ๋กœ์ปฌ ์ข…์†์„ฑ๋งŒ ํ•™์Šตํ•˜๋ฉด ๋˜๊ธฐ ๋•Œ๋ฌธ์— ๊ณ„์‚ฐ ์ค‘๋ณต์ด ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ด๋“ค์€ ์ด๋Ÿฌํ•œ ์ž๊ธฐ ์ฃผ์˜ ํ—ค๋“œ๋ฅผ ๋Œ€์ฒดํ•˜๊ณ  ๋กœ์ปฌ ์ข…์†์„ฑ์„ ์ง์ ‘ ๋ชจ๋ธ๋งํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด ์ŠคํŒฌ ๊ธฐ๋ฐ˜ ๋™์  ์ปจ๋ณผ๋ฃจ์…˜์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ƒˆ๋กœ์šด ์ปจ๋ณผ๋ฃจ์…˜ ํ—ค๋“œ๋Š” ๋‹ค๋ฅธ ์ž๊ธฐ ์ฃผ์˜ ํ—ค๋“œ์™€ ํ•จ๊ป˜ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์ฃผ์˜ ๋ธ”๋ก์„ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Sanh et al. (2019)์€ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์—์„œ ์ง€์‹ ์ฆ๋ฅ˜๋ฅผ ๊ตฌํ˜„ํ•˜์—ฌ ์–ธ์–ด ์ดํ•ด ๋Šฅ๋ ฅ์˜ 97%๋ฅผ ์œ ์ง€ํ•˜๋ฉด์„œ BERT ๋ชจ๋ธ์˜ ํฌ๊ธฐ๋ฅผ 40% ์ค„์ด๊ณ  ์†๋„๋ฅผ 60% ๋†’์ผ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
**Data Mixture.** Since each kind of data source is closely related to the development of certain capacities for LLMs (referring to the discussions in Section 4.1), it is important to set a suitable distribution to mix these data. The data mixture is generally set in a global level (_i.e.,_ the distribution of the entire pre-training data), and can be also locally set to varied proportions at different training stages. During pre-training, data samples from different sources would be selected according to the mixture proportions: more data will be sampled from a data source with a larger weight. Typically, existing LLMs such as LLaMA [57] may employ upsampling or downsampling on the full data of each source to create specific data mixtures as pre-training data. As Figure 6 illustrates, existing LLMs use different data mixtures to construct the pre-training data. As a representative model, the pre-training data of LLaMA [57] mainly consists of webpages (over 80%), alongside 6.5% of code-heavy data from GitHub and StackExchange, 4.5% from books, and 2.5% of scientific data sourced from arXiv, which has become an important reference for training general-purpose LLMs. Furthermore, special data mixtures can be used to facilitate different purposes. For example, Falcon [141] is trained on pure webpages, and CodeGen [86] largely increases the amount of code data.
**Data Mixture**. ๊ฐ ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ์ข…๋ฅ˜๋Š” LLM์˜ ํŠน์ • ์šฉ๋Ÿ‰ ๊ฐœ๋ฐœ๊ณผ ๋ฐ€์ ‘ํ•œ ๊ด€๋ จ์ด ์žˆ์œผ๋ฏ€๋กœ(4.1์ ˆ์˜ ๋…ผ์˜ ์ฐธ์กฐ), ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ํ˜ผํ•ฉํ•˜๊ธฐ ์œ„ํ•œ ์ ์ ˆํ•œ ๋ถ„ํฌ๋ฅผ ์„ค์ •ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ mixture๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ „์—ญ ์ˆ˜์ค€(_์ฆ‰, ์ „์ฒด pre-training ๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ)์œผ๋กœ ์„ค์ •๋˜๋ฉฐ, ๋‹ค์–‘ํ•œ ํ›ˆ๋ จ ๋‹จ๊ณ„์—์„œ ๋‹ค์–‘ํ•œ ๋น„์œจ๋กœ ๋กœ์ปฌ๋กœ ์„ค์ •ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. pre-training ์ค‘์— ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ์€ mixture ๋น„์œจ์— ๋”ฐ๋ผ ๋‹ค๋ฅธ ์†Œ์Šค์—์„œ ์„ ํƒ๋˜๋ฉฐ, ๊ฐ€์ค‘์น˜๊ฐ€ ํฐ ๋ฐ์ดํ„ฐ ์†Œ์Šค์—์„œ ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๊ฐ€ ์ƒ˜ํ”Œ๋ง๋ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ LLaMA [57]์™€ ๊ฐ™์€ ๊ธฐ์กด LLM์€ ๊ฐ ์†Œ์Šค์˜ ์ „์ฒด ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด upsampling ๋˜๋Š” downsampling์„ ์‚ฌ์šฉํ•˜์—ฌ pre-training ๋ฐ์ดํ„ฐ๋กœ ํŠน์ • ๋ฐ์ดํ„ฐ mixture๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 6์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ๊ธฐ์กด LLM์€ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ mixture๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ pre-training ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ๋ชจ๋ธ๋กœ์„œ, LLaMA [57]์˜ pre-training ๋ฐ์ดํ„ฐ๋Š” ์ฃผ๋กœ ์›น ํŽ˜์ด์ง€(80% ์ด์ƒ)๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, GitHub์™€ StackExchange์˜ ์ฝ”๋“œ๊ฐ€ ๋งŽ์€ ๋ฐ์ดํ„ฐ 6.5%, ๋„์„œ 4.5%, arXiv์—์„œ ๊ฐ€์ ธ์˜จ ๊ณผํ•™ ๋ฐ์ดํ„ฐ 2.5% ๋“ฑ์ด ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋ฒ”์šฉ LLM์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์ฐธ๊ณ  ์ž๋ฃŒ๊ฐ€ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ํŠน์ˆ˜ํ•œ ๋ฐ์ดํ„ฐ mixture๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ชฉ์ ์„ ์šฉ์ดํ•˜๊ฒŒ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Falcon [141]์€ ์ˆœ์ˆ˜ํ•œ ์›น ํŽ˜์ด์ง€์— ๋Œ€ํ•ด ํ•™์Šต๋˜๋ฉฐ, CodeGen [86]์€ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์˜ ์–‘์„ ํฌ๊ฒŒ ๋Š˜๋ฆฝ๋‹ˆ๋‹ค.
**๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ**. ๊ฐ ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ์ข…๋ฅ˜๋Š” LLM์˜ ํŠน์ • ์šฉ๋Ÿ‰ ๊ฐœ๋ฐœ๊ณผ ๋ฐ€์ ‘ํ•œ ๊ด€๋ จ์ด ์žˆ์œผ๋ฏ€๋กœ(4.1์ ˆ์˜ ๋…ผ์˜ ์ฐธ์กฐ), ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ํ˜ผํ•ฉํ•˜๊ธฐ ์œ„ํ•œ ์ ์ ˆํ•œ ๋ถ„ํฌ๋ฅผ ์„ค์ •ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ „์—ญ ์ˆ˜์ค€(_์ฆ‰, ์ „์ฒด ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ)์œผ๋กœ ์„ค์ •๋˜๋ฉฐ, ๋‹ค์–‘ํ•œ ํ›ˆ๋ จ ๋‹จ๊ณ„์—์„œ ๋‹ค์–‘ํ•œ ๋น„์œจ๋กœ ๋กœ์ปฌ๋กœ ์„ค์ •ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ์€ ํ˜ผํ•ฉ ๋น„์œจ์— ๋”ฐ๋ผ ๋‹ค๋ฅธ ์†Œ์Šค์—์„œ ์„ ํƒ๋˜๋ฉฐ, ๊ฐ€์ค‘์น˜๊ฐ€ ํฐ ๋ฐ์ดํ„ฐ ์†Œ์Šค์—์„œ ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๊ฐ€ ์ƒ˜ํ”Œ๋ง๋ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ LLaMA [57]์™€ ๊ฐ™์€ ๊ธฐ์กด LLM์€ ๊ฐ ์†Œ์Šค์˜ ์ „์ฒด ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์—…์ƒ˜ํ”Œ๋ง ๋˜๋Š” ๋‹ค์šด์ƒ˜ํ”Œ๋ง์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋กœ ํŠน์ • ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 6์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ๊ธฐ์กด LLM์€ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ๋ชจ๋ธ๋กœ์„œ, LLaMA [57]์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋Š” ์ฃผ๋กœ ์›น ํŽ˜์ด์ง€(80% ์ด์ƒ)๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, GitHub์™€ StackExchange์˜ ์ฝ”๋“œ๊ฐ€ ๋งŽ์€ ๋ฐ์ดํ„ฐ 6.5%, ๋„์„œ 4.5%, arXiv์—์„œ ๊ฐ€์ ธ์˜จ ๊ณผํ•™ ๋ฐ์ดํ„ฐ 2.5% ๋“ฑ์ด ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋ฒ”์šฉ LLM์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์ฐธ๊ณ  ์ž๋ฃŒ๊ฐ€ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ํŠน์ˆ˜ํ•œ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ชฉ์ ์„ ์šฉ์ดํ•˜๊ฒŒ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Falcon [141]์€ ์ˆœ์ˆ˜ํ•œ ์›น ํŽ˜์ด์ง€์— ๋Œ€ํ•ด ํ•™์Šต๋˜๋ฉฐ, CodeGen [86]์€ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์˜ ์–‘์„ ํฌ๊ฒŒ ๋Š˜๋ฆฝ๋‹ˆ๋‹ค.
Some methods adopt ranking, such as Cobbe et al. (2021), which trains a verifier to select high-confidence reasoning chains through ranking. Meanwhile, other methods select reasoning chains through a voting mechanism. Self-consistency Wang et al. (2023) selects the most consistent answer by majority voting among sampled reasoning chains based on final answers. Furthermore, Fu et al. (2023) proposes Complex CoT, which utilizes a complexity-based voting strategy that leans towards selecting answers generated by more complex reasoning chains. However, answer-based voting mechanisms do not take into account the correctness of reasoning chains. Miao et al. (2023) takes the reasoning steps into account when voting, which can obtain both consistent answers and trustworthy reasoning processes simultaneously. Moreover, to consider the relations between intermediate steps across chains, Yoran et al. (2023) mixes information between reasoning chains and selects the most relevant facts to perform meta-reason over multiple reasoning chains. GRACE Khalifa et al. (2023) trains a discriminator through contrastive learning and uses this discriminator to rank each intermediate reasoning step. Previous methods sample based on the probability distribution, while Diversity-of-Thought Naik et al. (2023) obtains multiple reasoning paths by prompting with different instructions.
์ผ๋ถ€ ๋ฐฉ๋ฒ•์€ ranking์„ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Cobbe ๋“ฑ(2021)์€ ranking์„ ํ†ตํ•ด ์‹ ๋ขฐ๋„๊ฐ€ ๋†’์€ ์ถ”๋ก  ์ฒด์ธ์„ ์„ ํƒํ•˜๋„๋ก ๊ฒ€์ฆ์ž๋ฅผ ํ›ˆ๋ จ์‹œํ‚ต๋‹ˆ๋‹ค. ํ•œํŽธ, ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•์€ ํˆฌํ‘œ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ†ตํ•ด ์ถ”๋ก  ์ฒด์ธ์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. Self-consistency Wang ๋“ฑ(2023)์€ ์ตœ์ข… ๋‹ต๋ณ€์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ˜ํ”Œ๋ง๋œ ์ถ”๋ก  ์ฒด์ธ ์ค‘ ๋‹ค์ˆ˜๊ฒฐ ํˆฌํ‘œ๋ฅผ ํ†ตํ•ด ๊ฐ€์žฅ ์ผ๊ด€์„ฑ์ด ๋†’์€ ๋‹ต๋ณ€์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, Fu ๋“ฑ(2023)์€ ๋” ๋ณต์žกํ•œ ์ถ”๋ก  ์ฒด์ธ์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๋‹ต๋ณ€์„ ์„ ํƒํ•˜๋Š” complexity-based voting ์ „๋žต์„ ํ™œ์šฉํ•˜๋Š” Complex CoT๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹ต๋ณ€ ๊ธฐ๋ฐ˜ voting ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ์ถ”๋ก  ์ฒด์ธ์˜ ์ •ํ™•์„ฑ์„ ๊ณ ๋ คํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. Miao ๋“ฑ(2023)์€ voting ์‹œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๊ณ ๋ คํ•˜์—ฌ ์ผ๊ด€๋œ ๋‹ต๋ณ€๊ณผ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค๋ฅผ ๋™์‹œ์— ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์—ฌ๋Ÿฌ ์ฒด์ธ ๊ฐ„์˜ ์ค‘๊ฐ„ ๋‹จ๊ณ„ ๊ฐ„์˜ ๊ด€๊ณ„๋ฅผ ๊ณ ๋ คํ•˜๊ธฐ ์œ„ํ•ด Yoran ๋“ฑ(2023)์€ ์ถ”๋ก  ์ฒด์ธ ๊ฐ„์˜ ์ •๋ณด๋ฅผ ํ˜ผํ•ฉํ•˜๊ณ  ๊ฐ€์žฅ ๊ด€๋ จ์„ฑ์ด ๋†’์€ ์‚ฌ์‹ค์„ ์„ ํƒํ•˜์—ฌ ์—ฌ๋Ÿฌ ์ถ”๋ก  ์ฒด์ธ์— ๋Œ€ํ•œ meta-reasoning์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. GRACE Khalifa ๋“ฑ(2023)์€ contrastive learning์„ ํ†ตํ•ด ํŒ๋ณ„์ž๋ฅผ ํ›ˆ๋ จํ•˜๊ณ  ์ด ํŒ๋ณ„์ž๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„์˜ ranking์„ ๋งค๊น๋‹ˆ๋‹ค. ์ด์ „ ๋ฐฉ๋ฒ•์€ ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ˜ํ”Œ๋งํ•˜๋Š” ๋ฐ˜๋ฉด, Diversity-of-Thought Naik ๋“ฑ(2023)์€ ์„œ๋กœ ๋‹ค๋ฅธ instructions์œผ๋กœ ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค.
์ผ๋ถ€ ๋ฐฉ๋ฒ•์€ ์ˆœ์œ„๋ฅผ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Cobbe ๋“ฑ(2021)์€ ์ˆœ์œ„๋ฅผ ํ†ตํ•ด ์‹ ๋ขฐ๋„๊ฐ€ ๋†’์€ ์ถ”๋ก  ์ฒด์ธ์„ ์„ ํƒํ•˜๋„๋ก ๊ฒ€์ฆ์ž๋ฅผ ํ›ˆ๋ จ์‹œํ‚ต๋‹ˆ๋‹ค. ํ•œํŽธ, ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•์€ ํˆฌํ‘œ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ†ตํ•ด ์ถ”๋ก  ์ฒด์ธ์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์ž์ฒด ์ผ๊ด€์„ฑ Wang ๋“ฑ(2023)์€ ์ตœ์ข… ๋‹ต๋ณ€์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ˜ํ”Œ๋ง๋œ ์ถ”๋ก  ์ฒด์ธ ์ค‘ ๋‹ค์ˆ˜๊ฒฐ ํˆฌํ‘œ๋ฅผ ํ†ตํ•ด ๊ฐ€์žฅ ์ผ๊ด€์„ฑ์ด ๋†’์€ ๋‹ต๋ณ€์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, Fu ๋“ฑ(2023)์€ ๋” ๋ณต์žกํ•œ ์ถ”๋ก  ์ฒด์ธ์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๋‹ต๋ณ€์„ ์„ ํƒํ•˜๋Š” ๋ณต์žก์„ฑ ๊ธฐ๋ฐ˜ ํˆฌํ‘œ ์ „๋žต์„ ํ™œ์šฉํ•˜๋Š” Complex CoT๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹ต๋ณ€ ๊ธฐ๋ฐ˜ ํˆฌํ‘œ ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ์ถ”๋ก  ์ฒด์ธ์˜ ์ •ํ™•์„ฑ์„ ๊ณ ๋ คํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. Miao ๋“ฑ(2023)์€ ํˆฌํ‘œ ์‹œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๊ณ ๋ คํ•˜์—ฌ ์ผ๊ด€๋œ ๋‹ต๋ณ€๊ณผ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค๋ฅผ ๋™์‹œ์— ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์—ฌ๋Ÿฌ ์ฒด์ธ ๊ฐ„์˜ ์ค‘๊ฐ„ ๋‹จ๊ณ„ ๊ฐ„์˜ ๊ด€๊ณ„๋ฅผ ๊ณ ๋ คํ•˜๊ธฐ ์œ„ํ•ด Yoran ๋“ฑ(2023)์€ ์ถ”๋ก  ์ฒด์ธ ๊ฐ„์˜ ์ •๋ณด๋ฅผ ํ˜ผํ•ฉํ•˜๊ณ  ๊ฐ€์žฅ ๊ด€๋ จ์„ฑ์ด ๋†’์€ ์‚ฌ์‹ค์„ ์„ ํƒํ•˜์—ฌ ์—ฌ๋Ÿฌ ์ถ”๋ก  ์ฒด์ธ์— ๋Œ€ํ•œ ๋ฉ”ํƒ€ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. GRACE Khalifa ๋“ฑ(2023)์€ ๋Œ€์กฐ ํ•™์Šต์„ ํ†ตํ•ด ํŒ๋ณ„์ž๋ฅผ ํ›ˆ๋ จํ•˜๊ณ  ์ด ํŒ๋ณ„์ž๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„์˜ ์ˆœ์œ„๋ฅผ ๋งค๊น๋‹ˆ๋‹ค. ์ด์ „ ๋ฐฉ๋ฒ•์€ ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ˜ํ”Œ๋งํ•˜๋Š” ๋ฐ˜๋ฉด, Diversity-of-Thought Naik ๋“ฑ(2023)์€ ์„œ๋กœ ๋‹ค๋ฅธ ์ง€์นจ์œผ๋กœ ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค.
Recently, a growing body of work has demonstrated the usefulness of custom tokens optimized to perform specific functions in the context of a neural network - for this reason, they have also been referred to as "function vectors." (Todd et al., 2023). One of the original instantiations of this was prompt-tuning (Lester et al., 2021) (and relatedly prefix-tuning (Li and Liang, 2021)), where the embeddings corresponding to the tokens of a prompt could be optimized to better accomplish a task. Others have applied meta-tokens to compress long prompts (Li et al., 2023; Jung and Kim, 2023) for efficiency. Most relevant to this work, Mu et al. (2024) optimized a token such that, when the tokens after it could not attend to the tokens before it (i.e., a context compression token), it would provide sufficient information to future tokens. Although we do not focus on compression, we share the problem of learning a token that affects attention and controls complex downstream behavior. In one related work, Goyal et al. (2023) show that learning a single "pause" token (essentially representing each token as two tokens) improves LM performance. However, unlike the thought tokens in our work, this pause token does not initialize a thought - instead, it can be seen as acting as the entirety of the thought. We find that reasoning in language is significantly more helpful.
์ตœ๊ทผ์—๋Š” ์‹ ๊ฒฝ๋ง์˜ ๋งฅ๋ฝ์—์„œ ํŠน์ • ๊ธฐ๋Šฅ์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์ตœ์ ํ™”๋œ ์‚ฌ์šฉ์ž ์ง€์ • ํ† ํฐ์˜ ์œ ์šฉ์„ฑ์„ ์ž…์ฆํ•˜๋Š” ์—ฐ๊ตฌ๊ฐ€ ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Ÿฌํ•œ ์ด์œ ๋กœ "function vectors"๋ผ๊ณ ๋„ ๋ถˆ๋ฆฝ๋‹ˆ๋‹ค. (Todd et al., 2023). ์ด๋ฅผ ๊ตฌํ˜„ํ•œ ์ตœ์ดˆ์˜ ์‚ฌ๋ก€ ์ค‘ ํ•˜๋‚˜๋Š” prompt-tuning (Lester et al., 2021)์ด๋ฉฐ, ํ”„๋กฌํ”„ํŠธ์˜ ํ† ํฐ์— ํ•ด๋‹นํ•˜๋Š” ์ž„๋ฒ ๋”ฉ์„ ์ตœ์ ํ™”ํ•˜์—ฌ ์ž‘์—…์„ ๋” ์ž˜ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆฌ๊ณ  ์ด์™€ ๊ด€๋ จํ•˜์—ฌ prefix-tuning (Li์™€ Liang, 2021)). ๋‹ค๋ฅธ ์‚ฌ๋žŒ๋“ค์€ ๋ฉ”ํƒ€ ํ† ํฐ์„ ์ ์šฉํ•˜์—ฌ ํšจ์œจ์„ฑ์„ ์œ„ํ•ด ๊ธด ํ”„๋กฌํ”„ํŠธ๋ฅผ ์••์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค(Li et al., 2023; Jung๊ณผ Kim, 2023). ์ด ์ž‘์—…๊ณผ ๊ฐ€์žฅ ๊ด€๋ จ์ด ์žˆ๋Š” ๊ฒƒ์€ Mu ๋“ฑ(2024)์ด ํ† ํฐ ๋’ค์— ์žˆ๋Š” ํ† ํฐ์ด ์•ž์— ์žˆ๋Š” ํ† ํฐ์— ์ฃผ์˜๋ฅผ ๊ธฐ์šธ์ผ ์ˆ˜ ์—†์„ ๋•Œ(์ฆ‰, context compression token) ๋ฏธ๋ž˜์˜ ํ† ํฐ์— ์ถฉ๋ถ„ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋„๋ก ํ† ํฐ์„ ์ตœ์ ํ™”ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์••์ถ•์— ์ดˆ์ ์„ ๋งž์ถ”์ง€๋Š” ์•Š์ง€๋งŒ, attention์„ ๋Œ๊ณ  ๋ณต์žกํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๋™์ž‘์„ ์ œ์–ดํ•˜๋Š” ํ† ํฐ์„ ํ•™์Šตํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ๊ณต์œ ํ•ฉ๋‹ˆ๋‹ค. ๊ด€๋ จ ์—ฐ๊ตฌ ์ค‘ ํ•˜๋‚˜์ธ Goyal ์™ธ(2023)๋Š” ๋‹จ์ผ '์ผ์‹œ ์ •์ง€' ํ† ํฐ์„ ํ•™์Šตํ•˜๋ฉด(๋ณธ์งˆ์ ์œผ๋กœ ๊ฐ ํ† ํฐ์„ ๋‘ ๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ํ‘œํ˜„) LM ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์ผ์‹œ ์ •์ง€ ํ† ํฐ์€ ์ €ํฌ ์ž‘์—…์˜ ์ƒ๊ฐ ํ† ํฐ๊ณผ ๋‹ฌ๋ฆฌ ์ƒ๊ฐ์„ ์ดˆ๊ธฐํ™”ํ•˜์ง€ ์•Š๊ณ , ๋Œ€์‹  ์ƒ๊ฐ์˜ ์ „์ฒด๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์—ญํ• ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์–ธ์–ด๋กœ ์ถ”๋ก ํ•˜๋Š” ๊ฒƒ์ด ํ›จ์”ฌ ๋” ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
์ตœ๊ทผ์—๋Š” ์‹ ๊ฒฝ๋ง์˜ ๋งฅ๋ฝ์—์„œ ํŠน์ • ๊ธฐ๋Šฅ์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์ตœ์ ํ™”๋œ ์‚ฌ์šฉ์ž ์ง€์ • ํ† ํฐ์˜ ์œ ์šฉ์„ฑ์„ ์ž…์ฆํ•˜๋Š” ์—ฐ๊ตฌ๊ฐ€ ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Ÿฌํ•œ ์ด์œ ๋กœ "ํ•จ์ˆ˜ ๋ฒกํ„ฐ"๋ผ๊ณ ๋„ ๋ถˆ๋ฆฝ๋‹ˆ๋‹ค. (Todd ์™ธ., 2023). ์ด๋ฅผ ๊ตฌํ˜„ํ•œ ์ตœ์ดˆ์˜ ์‚ฌ๋ก€ ์ค‘ ํ•˜๋‚˜๋Š” ํ”„๋กฌํ”„ํŠธ ํŠœ๋‹(Lester ์™ธ., 2021)์ด๋ฉฐ, ํ”„๋กฌํ”„ํŠธ์˜ ํ† ํฐ์— ํ•ด๋‹นํ•˜๋Š” ์ž„๋ฒ ๋”ฉ์„ ์ตœ์ ํ™”ํ•˜์—ฌ ์ž‘์—…์„ ๋” ์ž˜ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆฌ๊ณ  ์ด์™€ ๊ด€๋ จํ•˜์—ฌ ์ ‘๋‘์‚ฌ ํŠœ๋‹(Li์™€ Liang, 2021)). ๋‹ค๋ฅธ ์‚ฌ๋žŒ๋“ค์€ ๋ฉ”ํƒ€ ํ† ํฐ์„ ์ ์šฉํ•˜์—ฌ ํšจ์œจ์„ฑ์„ ์œ„ํ•ด ๊ธด ํ”„๋กฌํ”„ํŠธ๋ฅผ ์••์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค(Li et al., 2023; Jung๊ณผ Kim, 2023). ์ด ์ž‘์—…๊ณผ ๊ฐ€์žฅ ๊ด€๋ จ์ด ์žˆ๋Š” ๊ฒƒ์€ Mu ๋“ฑ(2024)์ด ํ† ํฐ ๋’ค์— ์žˆ๋Š” ํ† ํฐ์ด ์•ž์— ์žˆ๋Š” ํ† ํฐ์— ์ฃผ์˜๋ฅผ ๊ธฐ์šธ์ผ ์ˆ˜ ์—†์„ ๋•Œ(์ฆ‰, ์ปจํ…์ŠคํŠธ ์••์ถ• ํ† ํฐ) ๋ฏธ๋ž˜์˜ ํ† ํฐ์— ์ถฉ๋ถ„ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋„๋ก ํ† ํฐ์„ ์ตœ์ ํ™”ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์••์ถ•์— ์ดˆ์ ์„ ๋งž์ถ”์ง€๋Š” ์•Š์ง€๋งŒ, ์ฃผ์˜๋ฅผ ๋Œ๊ณ  ๋ณต์žกํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๋™์ž‘์„ ์ œ์–ดํ•˜๋Š” ํ† ํฐ์„ ํ•™์Šตํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ๊ณต์œ ํ•ฉ๋‹ˆ๋‹ค. ๊ด€๋ จ ์—ฐ๊ตฌ ์ค‘ ํ•˜๋‚˜์ธ Goyal ์™ธ(2023)๋Š” ๋‹จ์ผ '์ผ์‹œ ์ •์ง€' ํ† ํฐ์„ ํ•™์Šตํ•˜๋ฉด(๋ณธ์งˆ์ ์œผ๋กœ ๊ฐ ํ† ํฐ์„ ๋‘ ๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ํ‘œํ˜„) LM ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์ผ์‹œ ์ •์ง€ ํ† ํฐ์€ ์ €ํฌ ์ž‘์—…์˜ ์ƒ๊ฐ ํ† ํฐ๊ณผ ๋‹ฌ๋ฆฌ ์ƒ๊ฐ์„ ์ดˆ๊ธฐํ™”ํ•˜์ง€ ์•Š๊ณ , ๋Œ€์‹  ์ƒ๊ฐ์˜ ์ „์ฒด๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์—ญํ• ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์–ธ์–ด๋กœ ์ถ”๋ก ํ•˜๋Š” ๊ฒƒ์ด ํ›จ์”ฌ ๋” ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
* We investigate different data selection strategies for standard LLM pre-training setups where data has already been manually filtered / de-duplicated (e.g., MinHash), and where we do not know the target distribution for which we optimize performance. We argue that the performance of SSL Prototypes is affected by duplicate-driven clusters in the embedding space. In Section 3.4 we propose a new data selection strategy **D4** that utilizes SemDeDup to avoid getting impacted by such clusters. * In Section 4.1, we show that in the _compute-limited regime_ where we have "infinite" source data and train models with fixed token budgets, we can achieve better pre-training perplexity and downstream accuracy than random iid data selection and previously established methods. Furthermore, we show that our method D4 can achieve around 20% efficiency gains at the 6.7b model scale, and that the magnitude of efficiency gains increases with model scale. * In the _data-limited regime_, where we run out of data and must epoch over data, cleverly choosing what data to repeat can beat training on randomly selected new data, whereas randomly choosing data to repeat underperforms adding new data (Section 4.2). This calls into question the standard practice of single epoch LLM training, and suggests that epoching over intelligently subselected data might be a better approach.
* ๋ฐ์ดํ„ฐ๊ฐ€ ์ด๋ฏธ ์ˆ˜๋™์œผ๋กœ ํ•„ํ„ฐ๋ง/์ค‘๋ณต ์ œ๊ฑฐ๋œ ํ‘œ์ค€ LLM pre-training ์„ค์ •(์˜ˆ: MinHash)๊ณผ ์„ฑ๋Šฅ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ๋ชฉํ‘œ ๋ถ„ํฌ๋ฅผ ์•Œ ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. SSL Prototypes์˜ ์„ฑ๋Šฅ์€ ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์—์„œ ์ค‘๋ณต์œผ๋กœ ์ธํ•œ ํด๋Ÿฌ์Šคํ„ฐ์˜ ์˜ํ–ฅ์„ ๋ฐ›๋Š”๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 3.4์—์„œ๋Š” ์ด๋Ÿฌํ•œ ํด๋Ÿฌ์Šคํ„ฐ์˜ ์˜ํ–ฅ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด SemDeDup์„ ํ™œ์šฉํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต **D4**๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. * ์„น์…˜ 4.1์—์„œ๋Š” ์†Œ์Šค ๋ฐ์ดํ„ฐ๊ฐ€ "๋ฌดํ•œ๋Œ€"์ด๊ณ  ๊ณ ์ •๋œ ํ† ํฐ ์˜ˆ์‚ฐ์œผ๋กœ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” 'compute-limited regime'์—์„œ ๋ฌด์ž‘์œ„ iid ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐ ์ด์ „์— ํ™•๋ฆฝ๋œ ๋ฐฉ๋ฒ•๋ณด๋‹ค ๋” ๋‚˜์€ pre-training perplexity์™€ downstream accuracy๋ฅผ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ• D4๋Š” 6.7b ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ์•ฝ 20%์˜ ํšจ์œจ์„ฑ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํšจ์œจ์„ฑ ํ–ฅ์ƒ์˜ ๊ทœ๋ชจ๋Š” ๋ชจ๋ธ ๊ทœ๋ชจ์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. * ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜์—ฌ ๋ฐ์ดํ„ฐ๋ฅผ epochํ•ด์•ผ ํ•˜๋Š” _data-limited regime_์—์„œ๋Š” ์–ด๋–ค ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณตํ• ์ง€ ํ˜„๋ช…ํ•˜๊ฒŒ ์„ ํƒํ•˜๋ฉด ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ๋ฌด์ž‘์œ„๋กœ ๋ฐ˜๋ณตํ•  ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค(์„น์…˜ 4.2). ์ด๋Š” ๋‹จ์ผ epoch LLM ํ›ˆ๋ จ์˜ ํ‘œ์ค€ ๊ด€ํ–‰์— ์˜๋ฌธ์„ ์ œ๊ธฐํ•˜๊ณ , ์ง€๋Šฅ์ ์œผ๋กœ ํ•˜์œ„ ์„ ํƒ๋œ ๋ฐ์ดํ„ฐ๋ฅผ epochingํ•˜๋Š” ๊ฒƒ์ด ๋” ๋‚˜์€ ์ ‘๊ทผ ๋ฐฉ์‹์ผ ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
* ๋ฐ์ดํ„ฐ๊ฐ€ ์ด๋ฏธ ์ˆ˜๋™์œผ๋กœ ํ•„ํ„ฐ๋ง/์ค‘๋ณต ์ œ๊ฑฐ๋œ ํ‘œ์ค€ LLM ์‚ฌ์ „ ํ•™์Šต ์„ค์ •(์˜ˆ: MinHash)๊ณผ ์„ฑ๋Šฅ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ๋ชฉํ‘œ ๋ถ„ํฌ๋ฅผ ์•Œ ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. SSL ํ”„๋กœํ† ํƒ€์ž…์˜ ์„ฑ๋Šฅ์€ ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์—์„œ ์ค‘๋ณต์œผ๋กœ ์ธํ•œ ํด๋Ÿฌ์Šคํ„ฐ์˜ ์˜ํ–ฅ์„ ๋ฐ›๋Š”๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 3.4์—์„œ๋Š” ์ด๋Ÿฌํ•œ ํด๋Ÿฌ์Šคํ„ฐ์˜ ์˜ํ–ฅ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด SemDeDup์„ ํ™œ์šฉํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต **D4**๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. * ์„น์…˜ 4.1์—์„œ๋Š” ์†Œ์Šค ๋ฐ์ดํ„ฐ๊ฐ€ "๋ฌดํ•œ๋Œ€"์ด๊ณ  ๊ณ ์ •๋œ ํ† ํฐ ์˜ˆ์‚ฐ์œผ๋กœ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” '๊ณ„์‚ฐ ์ œํ•œ ์ •๊ถŒ'์—์„œ ๋ฌด์ž‘์œ„ iid ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐ ์ด์ „์— ํ™•๋ฆฝ๋œ ๋ฐฉ๋ฒ•๋ณด๋‹ค ๋” ๋‚˜์€ ์‚ฌ์ „ ํ›ˆ๋ จ ๋‹นํ˜น๊ฐ๊ณผ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ• D4๋Š” 6.7b ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ์•ฝ 20%์˜ ํšจ์œจ์„ฑ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํšจ์œจ์„ฑ ํ–ฅ์ƒ์˜ ๊ทœ๋ชจ๋Š” ๋ชจ๋ธ ๊ทœ๋ชจ์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. * ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜์—ฌ ๋ฐ์ดํ„ฐ๋ฅผ ์—ํฌํฌํ•ด์•ผ ํ•˜๋Š” _๋ฐ์ดํ„ฐ ์ œํ•œ ์ •๊ถŒ_์—์„œ๋Š” ์–ด๋–ค ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณตํ• ์ง€ ํ˜„๋ช…ํ•˜๊ฒŒ ์„ ํƒํ•˜๋ฉด ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ๋ฌด์ž‘์œ„๋กœ ๋ฐ˜๋ณตํ•  ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค(์„น์…˜ 4.2). ์ด๋Š” ๋‹จ์ผ ์—ํฌํฌ LLM ํ›ˆ๋ จ์˜ ํ‘œ์ค€ ๊ด€ํ–‰์— ์˜๋ฌธ์„ ์ œ๊ธฐํ•˜๊ณ , ์ง€๋Šฅ์ ์œผ๋กœ ํ•˜์œ„ ์„ ํƒ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์—ํฌํ‚นํ•˜๋Š” ๊ฒƒ์ด ๋” ๋‚˜์€ ์ ‘๊ทผ ๋ฐฉ์‹์ผ ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
**Remaining Issues.** Despite these efforts, there are still several challenges to address when applying LLMs in recommender systems. First, existing studies have shown that LLM-based recommendation models in zero/few-shot settings tend to perform worse than traditional ID-based recommenders [806, 807]. This indicates that LLMs might lack an understanding of personalized user behaviors and domain-specific collaborative semantics. Although instruction tuning alleviates this issue to some extent [809, 357], it can't fully reduce the semantic gap between LLMs and recommender systems, and also suffers from high tuning costs. Furthermore, recommender systems prioritize minimizing inference latency to enhance users' experience in low-resourced environments (_e.g._, phones), which poses a challenge to LLMs' inference speed as well as memory overhead. Therefore, it is important to explore improvement techniques, such as efficient tuning and quantization methods, to deploy LLMs efficiently and effectively in real-world recommender systems. In addition, existing LLMs have limited capacities in long context modeling, make it difficult to process the huge amount of user-item interaction data. Improved context length extension and context information utilization approaches should be developed to improve the modeling capacities of LLMs in long interaction sequences.
**ํ˜„์•ˆ**. ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ถ”์ฒœ ์‹œ์Šคํ…œ์— LLM์„ ์ ์šฉํ•  ๋•Œ ํ•ด๊ฒฐํ•ด์•ผ ํ•  ๋ช‡ ๊ฐ€์ง€ ๊ณผ์ œ๊ฐ€ ์—ฌ์ „ํžˆ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๊ธฐ์กด ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด zero/few-shot ์„ค์ •์—์„œ LLM ๊ธฐ๋ฐ˜ ์ถ”์ฒœ ๋ชจ๋ธ์€ ๊ธฐ์กด ID ๊ธฐ๋ฐ˜ ์ถ”์ฒœ ์‹œ์Šคํ…œ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค[806, 807]. ์ด๋Š” LLM์ด ๊ฐœ์ธํ™”๋œ ์‚ฌ์šฉ์ž ํ–‰๋™๊ณผ ๋„๋ฉ”์ธ๋ณ„ ํ˜‘์—… ์‹œ๋งจํ‹ฑ์— ๋Œ€ํ•œ ์ดํ•ด๊ฐ€ ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. Instruction Tuning์€ ์ด ๋ฌธ์ œ๋ฅผ ์–ด๋Š ์ •๋„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ[809, 357], LLM๊ณผ ์ถ”์ฒœ ์‹œ์Šคํ…œ ๊ฐ„์˜ ์˜๋ฏธ๋ก ์  ๊ฒฉ์ฐจ๋ฅผ ์™„์ „ํžˆ ์ค„์ผ ์ˆ˜๋Š” ์—†์œผ๋ฉฐ ๋†’์€ ํŠœ๋‹ ๋น„์šฉ์˜ ๋ฌธ์ œ๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ถ”์ฒœ ์‹œ์Šคํ…œ์€ ๋ฆฌ์†Œ์Šค๊ฐ€ ๋ถ€์กฑํ•œ ํ™˜๊ฒฝ(_์˜ˆ: ํœด๋Œ€ํฐ)์—์„œ ์‚ฌ์šฉ์ž์˜ ๊ฒฝํ—˜์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ถ”๋ก  ์ง€์—ฐ ์‹œ๊ฐ„์„ ์ตœ์†Œํ™”ํ•˜๋Š” ๊ฒƒ์„ ์šฐ์„ ์‹œํ•˜๊ธฐ ๋•Œ๋ฌธ์— LLM์˜ ์ถ”๋ก  ์†๋„์™€ ๋ฉ”๋ชจ๋ฆฌ ์˜ค๋ฒ„ํ—ค๋“œ์— ๋Œ€ํ•œ ๊ณผ์ œ๊ฐ€ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹ค์ œ ์ถ”์ฒœ ์‹œ์Šคํ…œ์—์„œ LLM์„ ํšจ์œจ์ ์ด๊ณ  ํšจ๊ณผ์ ์œผ๋กœ ๋ฐฐํฌํ•˜๊ธฐ ์œ„ํ•ด ํšจ์œจ์ ์ธ ํŠœ๋‹ ๋ฐ Quantization ๋ฐฉ๋ฒ•๊ณผ ๊ฐ™์€ ๊ฐœ์„  ๊ธฐ๋ฒ•์„ ํƒ์ƒ‰ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด LLM์€ ๊ธด ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ๋ง์— ์ œํ•œ๋œ ์šฉ๋Ÿ‰์„ ๊ฐ€์ง€๊ณ  ์žˆ์–ด ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์‚ฌ์šฉ์ž-์•„์ดํ…œ ์ƒํ˜ธ ์ž‘์šฉ ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๊ธด ์ƒํ˜ธ ์ž‘์šฉ ์‹œํ€€์Šค์—์„œ LLM์˜ ๋ชจ๋ธ๋ง ์šฉ๋Ÿ‰์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐœ์„ ๋œ ์ปจํ…์ŠคํŠธ ๊ธธ์ด ํ™•์žฅ ๋ฐ ์ปจํ…์ŠคํŠธ ์ •๋ณด ํ™œ์šฉ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
**ํ˜„์•ˆ**. ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ถ”์ฒœ ์‹œ์Šคํ…œ์— LLM์„ ์ ์šฉํ•  ๋•Œ ํ•ด๊ฒฐํ•ด์•ผ ํ•  ๋ช‡ ๊ฐ€์ง€ ๊ณผ์ œ๊ฐ€ ์—ฌ์ „ํžˆ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๊ธฐ์กด ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์ œ๋กœ/์†Œ์ˆ˜ ์ƒท ์„ค์ •์—์„œ LLM ๊ธฐ๋ฐ˜ ์ถ”์ฒœ ๋ชจ๋ธ์€ ๊ธฐ์กด ID ๊ธฐ๋ฐ˜ ์ถ”์ฒœ ์‹œ์Šคํ…œ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค[806, 807]. ์ด๋Š” LLM์ด ๊ฐœ์ธํ™”๋œ ์‚ฌ์šฉ์ž ํ–‰๋™๊ณผ ๋„๋ฉ”์ธ๋ณ„ ํ˜‘์—… ์‹œ๋งจํ‹ฑ์— ๋Œ€ํ•œ ์ดํ•ด๊ฐ€ ๋ถ€์กฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ง€์นจ ํŠœ๋‹์€ ์ด ๋ฌธ์ œ๋ฅผ ์–ด๋Š ์ •๋„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ[809, 357], LLM๊ณผ ์ถ”์ฒœ ์‹œ์Šคํ…œ ๊ฐ„์˜ ์˜๋ฏธ๋ก ์  ๊ฒฉ์ฐจ๋ฅผ ์™„์ „ํžˆ ์ค„์ผ ์ˆ˜๋Š” ์—†์œผ๋ฉฐ ๋†’์€ ํŠœ๋‹ ๋น„์šฉ์˜ ๋ฌธ์ œ๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ถ”์ฒœ ์‹œ์Šคํ…œ์€ ๋ฆฌ์†Œ์Šค๊ฐ€ ๋ถ€์กฑํ•œ ํ™˜๊ฒฝ(_์˜ˆ: ํœด๋Œ€ํฐ)์—์„œ ์‚ฌ์šฉ์ž์˜ ๊ฒฝํ—˜์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ถ”๋ก  ์ง€์—ฐ ์‹œ๊ฐ„์„ ์ตœ์†Œํ™”ํ•˜๋Š” ๊ฒƒ์„ ์šฐ์„ ์‹œํ•˜๊ธฐ ๋•Œ๋ฌธ์— LLM์˜ ์ถ”๋ก  ์†๋„์™€ ๋ฉ”๋ชจ๋ฆฌ ์˜ค๋ฒ„ํ—ค๋“œ์— ๋Œ€ํ•œ ๊ณผ์ œ๊ฐ€ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹ค์ œ ์ถ”์ฒœ ์‹œ์Šคํ…œ์—์„œ LLM์„ ํšจ์œจ์ ์ด๊ณ  ํšจ๊ณผ์ ์œผ๋กœ ๋ฐฐํฌํ•˜๊ธฐ ์œ„ํ•ด ํšจ์œจ์ ์ธ ํŠœ๋‹ ๋ฐ ์–‘์žํ™” ๋ฐฉ๋ฒ•๊ณผ ๊ฐ™์€ ๊ฐœ์„  ๊ธฐ๋ฒ•์„ ํƒ์ƒ‰ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด LLM์€ ๊ธด ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ๋ง์— ์ œํ•œ๋œ ์šฉ๋Ÿ‰์„ ๊ฐ€์ง€๊ณ  ์žˆ์–ด ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ์‚ฌ์šฉ์ž-์•„์ดํ…œ ์ƒํ˜ธ ์ž‘์šฉ ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๊ธด ์ƒํ˜ธ ์ž‘์šฉ ์‹œํ€€์Šค์—์„œ LLM์˜ ๋ชจ๋ธ๋ง ์šฉ๋Ÿ‰์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐœ์„ ๋œ ์ปจํ…์ŠคํŠธ ๊ธธ์ด ํ™•์žฅ ๋ฐ ์ปจํ…์ŠคํŠธ ์ •๋ณด ํ™œ์šฉ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_GPT-4_. As another remarkable progress, GPT-4 [46] was released in March 2023, which extended the text input to multimodal signals. Overall, GPT-4 has stronger capacities in solving complex tasks than GPT-3.5, showing a large performance improvement on many evaluation tasks. A recent study [41] investigated the capacities of GPT-4 by conducting qualitative tests with human-generated problems, spanning a diverse range of difficult tasks, and showed that GPT-4 can achieve more superior performance than prior GPT models such as ChatGPT. Furthermore, GPT-4 responds more safely to malicious or provocative queries, due to a six-month iterative alignment (with an additional safety reward signal in the RLHF training). In the technical report, OpenAI has emphasized how to safely develop GPT-4 and applied a number of intervention strategies to mitigate the possible issues of LLMs, such as hallucinations, privacy and overreliance. For example, they introduced the mechanism called _red teaming_[132] to reduce the harm or toxic content generation. As another important aspect, GPT-4 has been developed on a well-established deep learning infrastructure with improved optimization methods. They introduced a new mechanism called _predictable scaling_ that can accurately predict the final performance with a small proportion of compute during model training.
\(\bullet\)_GPT-4_. ๋˜ ๋‹ค๋ฅธ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์ง„์ „์œผ๋กœ, 2023๋…„ 3์›”์— ํ…์ŠคํŠธ ์ž…๋ ฅ์„ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์‹ ํ˜ธ๋กœ ํ™•์žฅํ•œ GPT-4 [46]๊ฐ€ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ GPT-4๋Š” GPT-3.5๋ณด๋‹ค ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๋Šฅ๋ ฅ์ด ๋” ๊ฐ•ํ•˜๋ฉฐ, ๋งŽ์€ ํ‰๊ฐ€ ์ž‘์—…์—์„œ ํฐ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ [41]์—์„œ๋Š” ์ธ๊ฐ„์ด ์ƒ์„ฑํ•œ ๋ฌธ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ •์„ฑ์  ํ…Œ์ŠคํŠธ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ฒ”์œ„์˜ ์–ด๋ ค์šด ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•จ์œผ๋กœ์จ GPT-4์˜ ์—ญ๋Ÿ‰์„ ์กฐ์‚ฌํ–ˆ์œผ๋ฉฐ, GPT-4๊ฐ€ ChatGPT์™€ ๊ฐ™์€ ์ด์ „ GPT ๋ชจ๋ธ๋ณด๋‹ค ๋” ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ GPT-4๋Š” ์•…์˜์ ์ด๊ฑฐ๋‚˜ ๋„๋ฐœ์ ์ธ ์ฟผ๋ฆฌ์— ๋Œ€ํ•ด ๋” ์•ˆ์ „ํ•˜๊ฒŒ ์‘๋‹ตํ•˜๋ฉฐ, ์ด๋Š” RLHF ํ›ˆ๋ จ์—์„œ ์ถ”๊ฐ€ ์•ˆ์ „ ๋ณด์ƒ ์‹ ํ˜ธ๋ฅผ ํ†ตํ•ด 6๊ฐœ์›”์— ๊ฑธ์ณ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ •๋ ฌ๋œ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ์—์„œ OpenAI๋Š” GPT-4๋ฅผ ์•ˆ์ „ํ•˜๊ฒŒ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๊ฐ•์กฐํ•˜๊ณ  hallucination, ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ, ๊ณผ์˜์กด๊ณผ ๊ฐ™์€ LLM์˜ ๊ฐ€๋Šฅํ•œ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ๊ฐœ์ž… ์ „๋žต์„ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ•ด๋ฅผ ์ž…ํžˆ๊ฑฐ๋‚˜ ์œ ๋…ํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด 'red teaming'[132]์ด๋ผ๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ค‘์š”ํ•œ ์ธก๋ฉด์œผ๋กœ์„œ GPT-4๋Š” ๊ฐœ์„ ๋œ optimization ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ์ž˜ ๊ตฌ์ถ•๋œ ๋”ฅ๋Ÿฌ๋‹ ์ธํ”„๋ผ์—์„œ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ํ•™์Šต ์ค‘์— ์ ์€ ๋น„์œจ์˜ ์ปดํ“จํŒ…์œผ๋กœ ์ตœ์ข… ์„ฑ๋Šฅ์„ ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋Š” 'predictable scaling'์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_GPT-4_. ๋˜ ๋‹ค๋ฅธ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์ง„์ „์œผ๋กœ, 2023๋…„ 3์›”์— ํ…์ŠคํŠธ ์ž…๋ ฅ์„ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์‹ ํ˜ธ๋กœ ํ™•์žฅํ•œ GPT-4 [46]๊ฐ€ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ GPT-4๋Š” GPT-3.5๋ณด๋‹ค ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๋Šฅ๋ ฅ์ด ๋” ๊ฐ•ํ•˜๋ฉฐ, ๋งŽ์€ ํ‰๊ฐ€ ์ž‘์—…์—์„œ ํฐ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ [41]์—์„œ๋Š” ์ธ๊ฐ„์ด ์ƒ์„ฑํ•œ ๋ฌธ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ •์„ฑ์  ํ…Œ์ŠคํŠธ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ฒ”์œ„์˜ ์–ด๋ ค์šด ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•จ์œผ๋กœ์จ GPT-4์˜ ์—ญ๋Ÿ‰์„ ์กฐ์‚ฌํ–ˆ์œผ๋ฉฐ, GPT-4๊ฐ€ ChatGPT์™€ ๊ฐ™์€ ์ด์ „ GPT ๋ชจ๋ธ๋ณด๋‹ค ๋” ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ GPT-4๋Š” ์•…์˜์ ์ด๊ฑฐ๋‚˜ ๋„๋ฐœ์ ์ธ ์ฟผ๋ฆฌ์— ๋Œ€ํ•ด ๋” ์•ˆ์ „ํ•˜๊ฒŒ ์‘๋‹ตํ•˜๋ฉฐ, ์ด๋Š” RLHF ํ›ˆ๋ จ์—์„œ ์ถ”๊ฐ€ ์•ˆ์ „ ๋ณด์ƒ ์‹ ํ˜ธ๋ฅผ ํ†ตํ•ด 6๊ฐœ์›”์— ๊ฑธ์ณ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ •๋ ฌ๋œ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ์—์„œ OpenAI๋Š” GPT-4๋ฅผ ์•ˆ์ „ํ•˜๊ฒŒ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๊ฐ•์กฐํ•˜๊ณ  ํ™˜๊ฐ, ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ, ๊ณผ์˜์กด๊ณผ ๊ฐ™์€ LLM์˜ ๊ฐ€๋Šฅํ•œ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ๊ฐœ์ž… ์ „๋žต์„ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ•ด๋ฅผ ์ž…ํžˆ๊ฑฐ๋‚˜ ์œ ๋…ํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด '๋ ˆ๋“œ ํŒ€'[132]์ด๋ผ๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ค‘์š”ํ•œ ์ธก๋ฉด์œผ๋กœ์„œ GPT-4๋Š” ๊ฐœ์„ ๋œ ์ตœ์ ํ™” ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ์ž˜ ๊ตฌ์ถ•๋œ ๋”ฅ๋Ÿฌ๋‹ ์ธํ”„๋ผ์—์„œ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ํ•™์Šต ์ค‘์— ์ ์€ ๋น„์œจ์˜ ์ปดํ“จํŒ…์œผ๋กœ ์ตœ์ข… ์„ฑ๋Šฅ์„ ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋Š” '์˜ˆ์ธก ๊ฐ€๋Šฅํ•œ ์Šค์ผ€์ผ๋ง'์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค.
Large Language Models (LLMs), such as ChatGPT and GPT-4, have dramatically transformed natural language processing research and shown promising strides towards Artificial General Intelligence (AGI). Nonetheless, the high costs associated with training and deploying LLMs present substantial obstacles to transparent, accessible academic research. While several large language models, such as LLaMA, have been open-sourced by the community, these predominantly focus on English corpora, limiting their usefulness for other languages. In this paper, we propose a method to augment LLaMA with capabilities for understanding and generating Chinese text and its ability to follow instructions. We achieve this by extending LLaMA's existing vocabulary with an additional 20,000 Chinese tokens, thereby improving its encoding efficiency and semantic understanding of Chinese. We further incorporate secondary pre-training using Chinese data and fine-tune the model with Chinese instruction datasets, significantly enhancing the model's ability to comprehend and execute instructions. Our experimental results indicate that the newly proposed model markedly enhances the original LLaMA's proficiency in understanding and generating Chinese content. Additionally, the results on the C-Eval dataset yield competitive performance among the models with several times the size of ours.
ChatGPT ๋ฐ GPT-4์™€ ๊ฐ™์€ Large Language Models (LLMs)์€ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์—ฐ๊ตฌ์— ํš๊ธฐ์ ์ธ ๋ณ€ํ™”๋ฅผ ๊ฐ€์ ธ์™”์œผ๋ฉฐ Artificial General Intelligence (AGI)์„ ํ–ฅํ•œ ์œ ๋งํ•œ ์ง„์ „์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์„ ํ›ˆ๋ จํ•˜๊ณ  ๋ฐฐํฌํ•˜๋Š” ๋ฐ ๋“œ๋Š” ๋†’์€ ๋น„์šฉ์€ ํˆฌ๋ช…ํ•˜๊ณ  ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ํ•™์ˆ  ์—ฐ๊ตฌ์— ์ƒ๋‹นํ•œ ์žฅ์• ๋ฌผ์ด ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLaMA์™€ ๊ฐ™์€ ๋ช‡๋ช‡ Large Language Models์€ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ์˜ํ•ด ์˜คํ”ˆ ์†Œ์Šคํ™”๋˜์—ˆ์ง€๋งŒ, ๋Œ€๋ถ€๋ถ„ ์˜์–ด ๋ง๋ญ‰์น˜์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์–ด ๋‹ค๋ฅธ ์–ธ์–ด์— ๋Œ€ํ•œ ์œ ์šฉ์„ฑ์ด ์ œํ•œ์ ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๊ธฐ๋Šฅ๊ณผ Instruction์„ ๋”ฐ๋ฅด๋Š” ๊ธฐ๋Šฅ์„ LLaMA์— ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด LLaMA์˜ ๊ธฐ์กด ์–ดํœ˜์— 20,000๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜์—ฌ ์ค‘๊ตญ์–ด์˜ ์ธ์ฝ”๋”ฉ ํšจ์œจ์„ฑ๊ณผ ์˜๋ฏธ๋ก ์  ์ดํ•ด๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 2์ฐจ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ฉํ•˜๊ณ  ์ค‘๊ตญ์–ด Instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ๋ชจ๋ธ์„ Fine-tuningํ•˜์—ฌ ๋ชจ๋ธ์˜ Instruction ์ดํ•ด ๋ฐ ์‹คํ–‰ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์ƒˆ๋กœ ์ œ์•ˆ๋œ ๋ชจ๋ธ์€ ์ค‘๊ตญ์–ด ์ฝ˜ํ…์ธ ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์žˆ์–ด ๊ธฐ์กด LLaMA์˜ ๋Šฅ๋ ฅ์„ ๋ˆˆ์— ๋„๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ C-Eval ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฐ๊ณผ๋Š” ์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ๋ช‡ ๋ฐฐ ํฌ๊ธฐ์˜ ๋ชจ๋ธ๋“ค ์‚ฌ์ด์—์„œ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
ChatGPT ๋ฐ GPT-4์™€ ๊ฐ™์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์—ฐ๊ตฌ์— ํš๊ธฐ์ ์ธ ๋ณ€ํ™”๋ฅผ ๊ฐ€์ ธ์™”์œผ๋ฉฐ ์ธ๊ณต ์ผ๋ฐ˜ ์ง€๋Šฅ(AGI)์„ ํ–ฅํ•œ ์œ ๋งํ•œ ์ง„์ „์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์„ ํ›ˆ๋ จํ•˜๊ณ  ๋ฐฐํฌํ•˜๋Š” ๋ฐ ๋“œ๋Š” ๋†’์€ ๋น„์šฉ์€ ํˆฌ๋ช…ํ•˜๊ณ  ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ํ•™์ˆ  ์—ฐ๊ตฌ์— ์ƒ๋‹นํ•œ ์žฅ์• ๋ฌผ์ด ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLaMA์™€ ๊ฐ™์€ ๋ช‡๋ช‡ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ์˜ํ•ด ์˜คํ”ˆ ์†Œ์Šคํ™”๋˜์—ˆ์ง€๋งŒ, ๋Œ€๋ถ€๋ถ„ ์˜์–ด ๋ง๋ญ‰์น˜์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์–ด ๋‹ค๋ฅธ ์–ธ์–ด์— ๋Œ€ํ•œ ์œ ์šฉ์„ฑ์ด ์ œํ•œ์ ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๊ธฐ๋Šฅ๊ณผ ๋ช…๋ น์–ด๋ฅผ ๋”ฐ๋ฅด๋Š” ๊ธฐ๋Šฅ์„ LLaMA์— ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด LLaMA์˜ ๊ธฐ์กด ์–ดํœ˜์— 20,000๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜์—ฌ ์ค‘๊ตญ์–ด์˜ ์ธ์ฝ”๋”ฉ ํšจ์œจ์„ฑ๊ณผ ์˜๋ฏธ๋ก ์  ์ดํ•ด๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 2์ฐจ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ฉํ•˜๊ณ  ์ค‘๊ตญ์–ด ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ๋ชจ๋ธ์˜ ๋ช…๋ น์–ด ์ดํ•ด ๋ฐ ์‹คํ–‰ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์ƒˆ๋กœ ์ œ์•ˆ๋œ ๋ชจ๋ธ์€ ์ค‘๊ตญ์–ด ์ฝ˜ํ…์ธ ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์žˆ์–ด ๊ธฐ์กด LLaMA์˜ ๋Šฅ๋ ฅ์„ ๋ˆˆ์— ๋„๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ C-Eval ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฐ๊ณผ๋Š” ์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ๋ช‡ ๋ฐฐ ํฌ๊ธฐ์˜ ๋ชจ๋ธ๋“ค ์‚ฌ์ด์—์„œ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
Raffel et al. (2020) introduces a unified framework for converting all text-based language problems into text-to-text format. The aim is to explore the potential of transfer learning techniques for natural language processing. In contrast to the original transformer, a simplified version of layer normalization is used, where activations are rescaled without additional biases. After applying layer normalization, a residual skip connection (He et al., 2016) adds the input of each subcomponent to its output. Srivastava et al. (2014) is applied to the feed-forward network, the skip connections, the attentional weights, and the inputs and outputs of the entire stack. The T5 model has been widely used as a language model by many researchers, such as Hofstatter et al. (2023), Sachan et al. (2021), and Singh et al. (2021). Additionally, Chung et al. (2022) proposed instruction tuning as an approach to improve model performance. The study focused on three aspects: the number of scaling tasks, the size of the scaled model, and the fine-tuning of chain of thought data. The results showed that larger model sizes and more fine-tuning tasks significantly improved model performance. Additionally, the study found that chain of thought(CoT) significantly improves inference level. Wang et al. (2023) used this approach to tune T5 and apply it to the RALM architecture.
Raffel et al. (2020)์€ ๋ชจ๋“  ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ฌธ์ œ๋ฅผ ํ…์ŠคํŠธ-ํ…์ŠคํŠธ ํ˜•์‹์œผ๋กœ ๋ณ€ํ™˜ํ•˜๊ธฐ ์œ„ํ•œ ํ†ตํ•ฉ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ „์ด ํ•™์Šต ๊ธฐ๋ฒ•์˜ ์ž ์žฌ๋ ฅ์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ž…๋‹ˆ๋‹ค. ์›๋ž˜์˜ Transformer์™€ ๋‹ฌ๋ฆฌ ํ™œ์„ฑํ™”๊ฐ€ ์ถ”๊ฐ€ ๋ฐ”์ด์–ด์Šค ์—†์ด ์žฌ์Šค์ผ€์ผ๋ง๋˜๋Š” ๋‹จ์ˆœํ™”๋œ ๋ฒ„์ „์˜ layer normalization์ด ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. layer normalization์„ ์ ์šฉํ•œ ํ›„, residual skip connection (He et al., 2016)์€ ๊ฐ ํ•˜์œ„ ๊ตฌ์„ฑ ์š”์†Œ์˜ ์ž…๋ ฅ์„ ์ถœ๋ ฅ์— ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Srivastava et al. (2014)์€ feed-forward network, skip connections, attentional weights, ์ „์ฒด ์Šคํƒ์˜ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ์— ์ ์šฉ๋ฉ๋‹ˆ๋‹ค. T5 ๋ชจ๋ธ์€ Hofstatter et al. (2023), Sachan et al. (2021), Singh et al. (2021)๊ณผ ๊ฐ™์€ ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์— ์˜ํ•ด ์–ธ์–ด ๋ชจ๋ธ๋กœ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ Chung et al. (2022)์€ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ Instruction Tuning์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” scaling ์ž‘์—…์˜ ์ˆ˜, scaling๋œ ๋ชจ๋ธ์˜ ํฌ๊ธฐ, chain of thought ๋ฐ์ดํ„ฐ์˜ Fine-tuning์˜ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ๋” ํฐ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ๋” ๋งŽ์€ Fine-tuning ์ž‘์—…์ด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด ์—ฐ๊ตฌ๋Š” chain of thought (CoT)์ด ์ถ”๋ก  ์ˆ˜์ค€์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. Wang et al. (2023)์€ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์—ฌ T5๋ฅผ ์กฐ์ •ํ•˜๊ณ  RALM ์•„ํ‚คํ…์ฒ˜์— ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
Raffel et al.(2020)์€ ๋ชจ๋“  ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ฌธ์ œ๋ฅผ ํ…์ŠคํŠธ-ํ…์ŠคํŠธ ํ˜•์‹์œผ๋กœ ๋ณ€ํ™˜ํ•˜๊ธฐ ์œ„ํ•œ ํ†ตํ•ฉ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ „์ด ํ•™์Šต ๊ธฐ๋ฒ•์˜ ์ž ์žฌ๋ ฅ์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ž…๋‹ˆ๋‹ค. ์›๋ž˜์˜ ํŠธ๋žœ์Šคํฌ๋จธ์™€ ๋‹ฌ๋ฆฌ ํ™œ์„ฑํ™”๊ฐ€ ์ถ”๊ฐ€ ๋ฐ”์ด์–ด์Šค ์—†์ด ์žฌ์Šค์ผ€์ผ๋ง๋˜๋Š” ๋‹จ์ˆœํ™”๋œ ๋ฒ„์ „์˜ ๋ ˆ์ด์–ด ์ •๊ทœํ™”๊ฐ€ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๋ ˆ์ด์–ด ์ •๊ทœํ™”๋ฅผ ์ ์šฉํ•œ ํ›„, ์ž”์ฐจ ์Šคํ‚ต ์—ฐ๊ฒฐ(He et al., 2016)์€ ๊ฐ ํ•˜์œ„ ๊ตฌ์„ฑ ์š”์†Œ์˜ ์ž…๋ ฅ์„ ์ถœ๋ ฅ์— ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Srivastava et al.(2014)์€ ํ”ผ๋“œํฌ์›Œ๋“œ ๋„คํŠธ์›Œํฌ, ์Šคํ‚ต ์—ฐ๊ฒฐ, ์ฃผ์˜ ๊ฐ€์ค‘์น˜, ์ „์ฒด ์Šคํƒ์˜ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ์— ์ ์šฉ๋ฉ๋‹ˆ๋‹ค. T5 ๋ชจ๋ธ์€ Hofstatter et al.(2023), Sachan et al.(2021), Singh et al.(2021)๊ณผ ๊ฐ™์€ ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์— ์˜ํ•ด ์–ธ์–ด ๋ชจ๋ธ๋กœ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ Chung et al.(2022)์€ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ๋ช…๋ น์–ด ํŠœ๋‹์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์Šค์ผ€์ผ๋ง ์ž‘์—…์˜ ์ˆ˜, ์Šค์ผ€์ผ๋ง๋œ ๋ชจ๋ธ์˜ ํฌ๊ธฐ, ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ๋ฐ์ดํ„ฐ์˜ ๋ฏธ์„ธ ์กฐ์ •์˜ ์„ธ ๊ฐ€์ง€ ์ธก๋ฉด์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ๋” ํฐ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ๋” ๋งŽ์€ ๋ฏธ์„ธ ์กฐ์ • ์ž‘์—…์ด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด ์—ฐ๊ตฌ๋Š” ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ(CoT)์ด ์ถ”๋ก  ์ˆ˜์ค€์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. Wang et al.(2023)์€ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์—ฌ T5๋ฅผ ์กฐ์ •ํ•˜๊ณ  RALM ์•„ํ‚คํ…์ฒ˜์— ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
In this paper, we show how to directly optimize a language model to adhere to human preferences, without explicit reward modeling or reinforcement learning. We propose _Direct Preference Optimization (DPO)_, an algorithm that implicitly optimizes the same objective as existing RLHF algorithms (reward maximization with a KL-divergence constraint) but is simple to implement and straightforward to train. Intuitively, the DPO update increases the relative log probability of preferred to dispreferred responses, but it incorporates a dynamic, per-example importance weight that prevents the model degeneration that we find occurs with a naive probability ratio objective. Like existing algorithms, DPO relies on a theoretical preference model (such as the Bradley-Terry model; [5]) that measures how well a given reward function aligns with empirical preference data. However, while existing methods use the preference model to define a preference loss to train a reward model and then train a policy that optimizes the learned reward model, DPO uses a change of variables to define the preference loss as a function of the policy directly. Given a dataset of human preferences over model responses, DPO can therefore optimize a policy using a simple binary cross entropy objective, producing the optimal policy to an implicit reward function fit to the preference data.
์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋ช…์‹œ์ ์ธ ๋ณด์ƒ ๋ชจ๋ธ๋ง์ด๋‚˜ Reinforcement Learning ์—†์ด ์–ธ์–ด ๋ชจ๋ธ์ด ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„๋ฅผ ์ค€์ˆ˜ํ•˜๋„๋ก ์ง์ ‘ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ธฐ์กด RLHF ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ๋™์ผํ•œ ๋ชฉ์ (KL-๋ถ„์‚ฐ ์ œ์•ฝ ์กฐ๊ฑด์„ ๊ฐ€์ง„ ๋ณด์ƒ ์ตœ๋Œ€ํ™”)์„ ์•”์‹œ์ ์œผ๋กœ ์ตœ์ ํ™”ํ•˜์ง€๋งŒ ๊ตฌํ˜„์ด ๊ฐ„๋‹จํ•˜๊ณ  ํ›ˆ๋ จ์ด ๊ฐ„๋‹จํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ _Direct Preference Optimization(DPO)_์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ง๊ด€์ ์œผ๋กœ DPO ์—…๋ฐ์ดํŠธ๋Š” ์„ ํ˜ธ๋˜๋Š” ์‘๋‹ต๊ณผ ๋ถˆ๋ฆฌํ•œ ์‘๋‹ต์˜ ์ƒ๋Œ€์  ๋กœ๊ทธ ํ™•๋ฅ ์„ ๋†’์ด์ง€๋งŒ, ์ˆœ์ง„ํ•œ ํ™•๋ฅ  ๋น„์œจ ๋ชฉํ‘œ์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๋ชจ๋ธ ๋ณ€์งˆ์„ ๋ฐฉ์ง€ํ•˜๋Š” ๋™์  ์˜ˆ์‹œ๋ณ„ ์ค‘์š”๋„ ๊ฐ€์ค‘์น˜๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ DPO๋Š” ์ฃผ์–ด์ง„ ๋ณด์ƒ ํ•จ์ˆ˜๊ฐ€ ๊ฒฝํ—˜์  ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ์™€ ์–ผ๋งˆ๋‚˜ ์ž˜ ์ผ์น˜ํ•˜๋Š”์ง€๋ฅผ ์ธก์ •ํ•˜๋Š” ์ด๋ก ์  ์„ ํ˜ธ๋„ ๋ชจ๋ธ(์˜ˆ: ๋ธŒ๋ž˜๋“ค๋ฆฌ-ํ…Œ๋ฆฌ ๋ชจ๋ธ[5])์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ์„ ํ˜ธ๋„ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์„ ํ˜ธ๋„ ์†์‹ค์„ ์ •์˜ํ•˜์—ฌ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚จ ๋‹ค์Œ ํ•™์Šต๋œ ๋ณด์ƒ ๋ชจ๋ธ์„ ์ตœ์ ํ™”ํ•˜๋Š” ์ •์ฑ…์„ ํ•™์Šต์‹œํ‚ค๋Š” ๋ฐ˜๋ฉด, DPO๋Š” ๋ณ€์ˆ˜ ๋ณ€๊ฒฝ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ •์ฑ…์˜ ํ•จ์ˆ˜๋กœ ์„ ํ˜ธ๋„ ์†์‹ค์„ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ DPO๋Š” ๋ชจ๋ธ ์‘๋‹ต์— ๋Œ€ํ•œ ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ๊ฐ„๋‹จํ•œ ์ด์ง„ ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ ๋ชฉํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ •์ฑ…์„ ์ตœ์ ํ™”ํ•˜์—ฌ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ์— ๋งž๋Š” ์•”์‹œ์  ๋ณด์ƒ ํ•จ์ˆ˜์— ๋Œ€ํ•œ ์ตœ์ ์˜ ์ •์ฑ…์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋ช…์‹œ์ ์ธ ๋ณด์ƒ ๋ชจ๋ธ๋ง์ด๋‚˜ ๊ฐ•ํ™” ํ•™์Šต ์—†์ด ์–ธ์–ด ๋ชจ๋ธ์ด ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„๋ฅผ ์ค€์ˆ˜ํ•˜๋„๋ก ์ง์ ‘ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ธฐ์กด RLHF ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ๋™์ผํ•œ ๋ชฉ์ (KL-๋ถ„์‚ฐ ์ œ์•ฝ ์กฐ๊ฑด์„ ๊ฐ€์ง„ ๋ณด์ƒ ์ตœ๋Œ€ํ™”)์„ ์•”์‹œ์ ์œผ๋กœ ์ตœ์ ํ™”ํ•˜์ง€๋งŒ ๊ตฌํ˜„์ด ๊ฐ„๋‹จํ•˜๊ณ  ํ›ˆ๋ จ์ด ๊ฐ„๋‹จํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ _Direct Preference Optimization(DPO)_์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ง๊ด€์ ์œผ๋กœ DPO ์—…๋ฐ์ดํŠธ๋Š” ์„ ํ˜ธ๋˜๋Š” ์‘๋‹ต๊ณผ ๋ถˆ๋ฆฌํ•œ ์‘๋‹ต์˜ ์ƒ๋Œ€์  ๋กœ๊ทธ ํ™•๋ฅ ์„ ๋†’์ด์ง€๋งŒ, ์ˆœ์ง„ํ•œ ํ™•๋ฅ  ๋น„์œจ ๋ชฉํ‘œ์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๋ชจ๋ธ ๋ณ€์งˆ์„ ๋ฐฉ์ง€ํ•˜๋Š” ๋™์  ์˜ˆ์‹œ๋ณ„ ์ค‘์š”๋„ ๊ฐ€์ค‘์น˜๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ DPO๋Š” ์ฃผ์–ด์ง„ ๋ณด์ƒ ํ•จ์ˆ˜๊ฐ€ ๊ฒฝํ—˜์  ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ์™€ ์–ผ๋งˆ๋‚˜ ์ž˜ ์ผ์น˜ํ•˜๋Š”์ง€๋ฅผ ์ธก์ •ํ•˜๋Š” ์ด๋ก ์  ์„ ํ˜ธ๋„ ๋ชจ๋ธ(์˜ˆ: ๋ธŒ๋ž˜๋“ค๋ฆฌ-ํ…Œ๋ฆฌ ๋ชจ๋ธ[5])์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ์„ ํ˜ธ๋„ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์„ ํ˜ธ๋„ ์†์‹ค์„ ์ •์˜ํ•˜์—ฌ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚จ ๋‹ค์Œ ํ•™์Šต๋œ ๋ณด์ƒ ๋ชจ๋ธ์„ ์ตœ์ ํ™”ํ•˜๋Š” ์ •์ฑ…์„ ํ•™์Šต์‹œํ‚ค๋Š” ๋ฐ˜๋ฉด, DPO๋Š” ๋ณ€์ˆ˜ ๋ณ€๊ฒฝ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ •์ฑ…์˜ ํ•จ์ˆ˜๋กœ ์„ ํ˜ธ๋„ ์†์‹ค์„ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ DPO๋Š” ๋ชจ๋ธ ์‘๋‹ต์— ๋Œ€ํ•œ ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ๊ฐ„๋‹จํ•œ ์ด์ง„ ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ ๋ชฉํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ •์ฑ…์„ ์ตœ์ ํ™”ํ•˜์—ฌ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ์— ๋งž๋Š” ์•”์‹œ์  ๋ณด์ƒ ํ•จ์ˆ˜์— ๋Œ€ํ•œ ์ตœ์ ์˜ ์ •์ฑ…์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
frequency in each corpus of the entities that serve as the ground truths, e.g., those entities appear more in corpus \(D_{0}\) than in \(D_{1}\). In order to get rid of the influence of frequency of entities when analyzing the source of the EM gap, we find overlapping _Person_ type answers from UL and NL, and analyze only the 67 probing sentences for both datasets each paired to one of these entities. As shown in Figure (b)b, the EM on UL is still much higher than that of NL. Manually analyzing these instances, we find that the probing sentences for NL ask for relatively more _fine-grained_ knowledge compared to UL, since the instances of UL by definition are overlapped close sentences with different answers in the corpus \(D_{0}\) and \(D_{1}\), that naturally make them be _coarse-grained_. For instance, the probing sentences for entity "Tim Walz" in UL and NL are " \(\xrightarrow{\ }\) is the governor of Minnesota this year." and " \(\xrightarrow{\ }\) is the governor of Minnesota calling for the evacuation of St. Paul.", respectively. We thus conjecture that the main causation of the EM gap to be UL consisting of instances requiring _coarse-grained_ knowledge, which is likely to have appeared more during in \(D_{1}\), while NL consisting of instances requiring _fine-grained_ knowledge, which is expected to likely have appeared less in \(D_{1}\).
๊ฐ ๋ง๋ญ‰์น˜์—์„œ ๊ธฐ์ค€ ์ง„์‹ค ์—ญํ• ์„ ํ•˜๋Š” ๊ฐœ์ฒด์˜ ๋นˆ๋„, ์˜ˆ๋ฅผ ๋“ค์–ด, ํ•ด๋‹น ๊ฐœ์ฒด๊ฐ€ \(D_{1}\)๋ณด๋‹ค ๋ง๋ญ‰์น˜ \(D_{0}\)์— ๋” ๋งŽ์ด ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒฝ์šฐ. EM ๊ฒฉ์ฐจ์˜ ์›์ธ์„ ๋ถ„์„ํ•  ๋•Œ ๊ฐœ์ฒด์˜ ๋นˆ๋„์— ์˜ํ•œ ์˜ํ–ฅ์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด UL๊ณผ NL์—์„œ ๊ฒน์น˜๋Š” _Person_ ์œ ํ˜•์˜ ๋‹ต์„ ์ฐพ์•„์„œ ์ด๋Ÿฌํ•œ ๊ฐœ์ฒด ์ค‘ ํ•˜๋‚˜์— ๊ฐ๊ฐ ์ง์„ ์ด๋ฃฌ ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ 67๊ฐœ์˜ ํƒ์ƒ‰ ๋ฌธ์žฅ๋งŒ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ (b)b์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด UL์˜ EM์€ ์—ฌ์ „ํžˆ NL๋ณด๋‹ค ํ›จ์”ฌ ๋†’์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ธ์Šคํ„ด์Šค๋ฅผ ์ˆ˜๋™์œผ๋กœ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ, NL์— ๋Œ€ํ•œ ํƒ์ƒ‰ ๋ฌธ์žฅ์€ ์ •์˜์ƒ ์ฝ”ํผ์Šค \(D_{0}\)์™€ \(D_{1}\)์—์„œ ๋‹ต์ด ๋‹ค๋ฅธ ๋ฌธ์žฅ์ด ๊ฒน์น˜๋Š” UL์— ๋น„ํ•ด ์ƒ๋Œ€์ ์œผ๋กœ ๋” _fine-grained_ ์ง€์‹์„ ์š”๊ตฌํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, UL๊ณผ NL์—์„œ "Tim Walz"๋ผ๋Š” ๊ฐœ์ฒด์— ๋Œ€ํ•œ ํƒ์ƒ‰ ๋ฌธ์žฅ์€ ๊ฐ๊ฐ " \(\xrightarrow{\ }\)๋Š” ์˜ฌํ•ด ๋ฏธ๋„ค์†Œํƒ€์˜ ์ฃผ์ง€์‚ฌ์ž…๋‹ˆ๋‹ค." ๋ฐ " \(\xrightarrow{\ }\)๋Š” ์„ธ์ธํŠธ ํด์˜ ์ฒ ์ˆ˜๋ฅผ ์š”๊ตฌํ•˜๋Š” ๋ฏธ๋„ค์†Œํƒ€ ์ฃผ์ง€์‚ฌ์ž…๋‹ˆ๋‹ค."์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ EM ๊ฒฉ์ฐจ์˜ ์ฃผ์š” ์›์ธ์€ _coarse-grained_ ์ง€์‹์ด ํ•„์š”ํ•œ ์ธ์Šคํ„ด์Šค๋กœ ๊ตฌ์„ฑ๋œ UL์€ \(D_{1}\)์—์„œ ๋” ๋งŽ์ด ๋‚˜ํƒ€๋‚ฌ์„ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ๋ฐ˜๋ฉด, _fine-grained_ ์ง€์‹์ด ํ•„์š”ํ•œ ์ธ์Šคํ„ด์Šค๋กœ ๊ตฌ์„ฑ๋œ NL์€ \(D_{1}\)์—์„œ ๋” ์ ๊ฒŒ ๋‚˜ํƒ€๋‚ฌ์„ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.
๊ฐ ๋ง๋ญ‰์น˜์—์„œ ๊ธฐ์ค€ ์ง„์‹ค ์—ญํ• ์„ ํ•˜๋Š” ๊ฐœ์ฒด์˜ ๋นˆ๋„, ์˜ˆ๋ฅผ ๋“ค์–ด, ํ•ด๋‹น ๊ฐœ์ฒด๊ฐ€ \(D_{1}\)๋ณด๋‹ค ๋ง๋ญ‰์น˜ \(D_{0}\)์— ๋” ๋งŽ์ด ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒฝ์šฐ. EM ๊ฒฉ์ฐจ์˜ ์›์ธ์„ ๋ถ„์„ํ•  ๋•Œ ๊ฐœ์ฒด์˜ ๋นˆ๋„์— ์˜ํ•œ ์˜ํ–ฅ์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด UL๊ณผ NL์—์„œ ๊ฒน์น˜๋Š” _Person_ ์œ ํ˜•์˜ ๋‹ต์„ ์ฐพ์•„์„œ ์ด๋Ÿฌํ•œ ๊ฐœ์ฒด ์ค‘ ํ•˜๋‚˜์— ๊ฐ๊ฐ ์ง์„ ์ด๋ฃฌ ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ 67๊ฐœ์˜ ํƒ์ƒ‰ ๋ฌธ์žฅ๋งŒ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ (b)b์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด UL์˜ EM์€ ์—ฌ์ „ํžˆ NL๋ณด๋‹ค ํ›จ์”ฌ ๋†’์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ธ์Šคํ„ด์Šค๋ฅผ ์ˆ˜๋™์œผ๋กœ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ, NL์— ๋Œ€ํ•œ ํƒ์ƒ‰ ๋ฌธ์žฅ์€ ์ •์˜์ƒ ์ฝ”ํผ์Šค \(D_{0}\)์™€ \(D_{1}\)์—์„œ ๋‹ต์ด ๋‹ค๋ฅธ ๋ฌธ์žฅ์ด ๊ฒน์น˜๋Š” UL์— ๋น„ํ•ด ์ƒ๋Œ€์ ์œผ๋กœ ๋” _์„ธ๋ถ„ํ™”๋œ_ ์ง€์‹์„ ์š”๊ตฌํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, UL๊ณผ NL์—์„œ "Tim Walz"๋ผ๋Š” ๊ฐœ์ฒด์— ๋Œ€ํ•œ ํƒ์ƒ‰ ๋ฌธ์žฅ์€ ๊ฐ๊ฐ " \(\xrightarrow{\ }\)๋Š” ์˜ฌํ•ด ๋ฏธ๋„ค์†Œํƒ€์˜ ์ฃผ์ง€์‚ฌ์ž…๋‹ˆ๋‹ค." ๋ฐ " \(\xrightarrow{\ }\)๋Š” ์„ธ์ธํŠธ ํด์˜ ์ฒ ์ˆ˜๋ฅผ ์š”๊ตฌํ•˜๋Š” ๋ฏธ๋„ค์†Œํƒ€ ์ฃผ์ง€์‚ฌ์ž…๋‹ˆ๋‹ค."์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ EM ๊ฒฉ์ฐจ์˜ ์ฃผ์š” ์›์ธ์€ _๋Œ€๋žต์ ์ธ_ ์ง€์‹์ด ํ•„์š”ํ•œ ์ธ์Šคํ„ด์Šค๋กœ ๊ตฌ์„ฑ๋œ UL์€ \(D_{1}\)์—์„œ ๋” ๋งŽ์ด ๋‚˜ํƒ€๋‚ฌ์„ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ๋ฐ˜๋ฉด, _๋ฏธ์„ธํ•œ_ ์ง€์‹์ด ํ•„์š”ํ•œ ์ธ์Šคํ„ด์Šค๋กœ ๊ตฌ์„ฑ๋œ NL์€ \(D_{1}\)์—์„œ ๋” ์ ๊ฒŒ ๋‚˜ํƒ€๋‚ฌ์„ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.
**Emergent Architectures.** The conventional Transformer architectures typically suffer from quadratic computational complexity. Because of this, efficiency has become an important issue when training and making inference with long inputs. To improve efficiency, some studies aim to devise new architectures for language modeling, including parameterized state space models (_e.g.,_ S4 [247], GSS [248], and H3 [249]), long convolutions like Hyea [250], and Transformer-like architectures that incorporate recursive update mechanisms (_e.g.,_ RWKV [251] and RetNet [252]). The key merits of these new architectures are twofold. First, these models can generate outputs recursively like RNNs, meaning that they only need to refer to the single previous state during decoding. It makes the decoding process more efficient as it eliminates the need to revisit all previous states as in conventional Transformers. Second, these models have the capacity to encode an entire sentence in parallel like Transformers. This contrasts with conventional RNNs which has to encode sentences on a token-by-token basis. Thus, they can benefit from the parallelism of GPUs with techniques such as Parallel Scan [253, 254], FFT [250, 251], and Chunkwise Recurrent [252]. These techniques enable models with these new architectures to be trained in a highly parallel and efficient manner.
**Emergent Architectures.** ๊ธฐ์กด์˜ Transformer ์•„ํ‚คํ…์ฒ˜๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ด์ฐจ์ ์ธ ๊ณ„์‚ฐ ๋ณต์žก์„ฑ์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ์ด ๋•Œ๋ฌธ์— ๊ธด ์ž…๋ ฅ์œผ๋กœ ํ›ˆ๋ จํ•˜๊ณ  ์ถ”๋ก ํ•  ๋•Œ ํšจ์œจ์„ฑ์ด ์ค‘์š”ํ•œ ๋ฌธ์ œ๊ฐ€ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ํŒŒ๋ผ๋ฏธํ„ฐํ™”๋œ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ(_์˜ˆ:_, S4 [247], GSS [248], H3 [249]), Hyea [250]์™€ ๊ฐ™์€ ๊ธด ์ปจ๋ณผ๋ฃจ์…˜, ์žฌ๊ท€์  ์—…๋ฐ์ดํŠธ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ†ตํ•ฉํ•œ Transformer ์œ ์‚ฌ ์•„ํ‚คํ…์ฒ˜(_์˜ˆ:_, RWKV [251], RetNet [252]) ๋“ฑ ์–ธ์–ด ๋ชจ๋ธ๋ง์„ ์œ„ํ•œ ์ƒˆ๋กœ์šด ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ณ ์•ˆํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ์•„ํ‚คํ…์ฒ˜์˜ ์ฃผ์š” ์žฅ์ ์€ ๋‘ ๊ฐ€์ง€์ž…๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ RNN์ฒ˜๋Ÿผ ์žฌ๊ท€์ ์œผ๋กœ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋””์ฝ”๋”ฉ ์ค‘์— ์ด์ „ ์ƒํƒœ ํ•˜๋‚˜๋งŒ ์ฐธ์กฐํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ธฐ์กด Transformer์ฒ˜๋Ÿผ ์ด์ „ ์ƒํƒœ๋ฅผ ๋ชจ๋‘ ๋‹ค์‹œ ๋ฐฉ๋ฌธํ•  ํ•„์š”๊ฐ€ ์—†์œผ๋ฏ€๋กœ ๋””์ฝ”๋”ฉ ๊ณผ์ •์ด ๋” ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ๋‘˜์งธ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ Transformer์ฒ˜๋Ÿผ ๋ฌธ์žฅ ์ „์ฒด๋ฅผ ๋ณ‘๋ ฌ๋กœ ์ธ์ฝ”๋”ฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํ† ํฐ ๋‹จ์œ„๋กœ ๋ฌธ์žฅ์„ ์ธ์ฝ”๋”ฉํ•ด์•ผ ํ•˜๋Š” ๊ธฐ์กด RNN๊ณผ ๋Œ€์กฐ์ ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Parallel Scan [253, 254], FFT [250, 251], Chunkwise Recurrent [252]์™€ ๊ฐ™์€ ๊ธฐ์ˆ ์„ ํ†ตํ•ด GPU์˜ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๋ฅผ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ ์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ์„ ๋งค์šฐ ๋ณ‘๋ ฌ์ ์ด๊ณ  ํšจ์œจ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์ด๋จธ์ „ํŠธ ์•„ํ‚คํ…์ฒ˜.** ๊ธฐ์กด์˜ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ด์ฐจ์ ์ธ ๊ณ„์‚ฐ ๋ณต์žก์„ฑ์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ์ด ๋•Œ๋ฌธ์— ๊ธด ์ž…๋ ฅ์œผ๋กœ ํ›ˆ๋ จํ•˜๊ณ  ์ถ”๋ก ํ•  ๋•Œ ํšจ์œจ์„ฑ์ด ์ค‘์š”ํ•œ ๋ฌธ์ œ๊ฐ€ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ํŒŒ๋ผ๋ฏธํ„ฐํ™”๋œ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ(_์˜ˆ:_, S4 [247], GSS [248], H3 [249]), Hyea [250]์™€ ๊ฐ™์€ ๊ธด ์ปจ๋ณผ๋ฃจ์…˜, ์žฌ๊ท€์  ์—…๋ฐ์ดํŠธ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ†ตํ•ฉํ•œ ํŠธ๋žœ์Šคํฌ๋จธ ์œ ์‚ฌ ์•„ํ‚คํ…์ฒ˜(_์˜ˆ:_, RWKV [251], RetNet [252]) ๋“ฑ ์–ธ์–ด ๋ชจ๋ธ๋ง์„ ์œ„ํ•œ ์ƒˆ๋กœ์šด ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ณ ์•ˆํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ์•„ํ‚คํ…์ฒ˜์˜ ์ฃผ์š” ์žฅ์ ์€ ๋‘ ๊ฐ€์ง€์ž…๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ RNN์ฒ˜๋Ÿผ ์žฌ๊ท€์ ์œผ๋กœ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋””์ฝ”๋”ฉ ์ค‘์— ์ด์ „ ์ƒํƒœ ํ•˜๋‚˜๋งŒ ์ฐธ์กฐํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ธฐ์กด ํŠธ๋žœ์Šคํฌ๋จธ์ฒ˜๋Ÿผ ์ด์ „ ์ƒํƒœ๋ฅผ ๋ชจ๋‘ ๋‹ค์‹œ ๋ฐฉ๋ฌธํ•  ํ•„์š”๊ฐ€ ์—†์œผ๋ฏ€๋กœ ๋””์ฝ”๋”ฉ ๊ณผ์ •์ด ๋” ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ๋‘˜์งธ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ํŠธ๋žœ์Šคํฌ๋จธ์ฒ˜๋Ÿผ ๋ฌธ์žฅ ์ „์ฒด๋ฅผ ๋ณ‘๋ ฌ๋กœ ์ธ์ฝ”๋”ฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํ† ํฐ ๋‹จ์œ„๋กœ ๋ฌธ์žฅ์„ ์ธ์ฝ”๋”ฉํ•ด์•ผ ํ•˜๋Š” ๊ธฐ์กด RNN๊ณผ ๋Œ€์กฐ์ ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ณ‘๋ ฌ ์Šค์บ” [253, 254], FFT [250, 251], ์ฒญํฌ ์™€์ด์ฆˆ ๋ฆฌ์ปค๋ŸฐํŠธ [252]์™€ ๊ฐ™์€ ๊ธฐ์ˆ ์„ ํ†ตํ•ด GPU์˜ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๋ฅผ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ ์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ์„ ๋งค์šฐ ๋ณ‘๋ ฌ์ ์ด๊ณ  ํšจ์œจ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 5 reports the validation loss during continual pre-training of 405M parameter models for weak (top) and strong (bottom) shifts. Table 3 reports the average (over the last 100 iterations) final loss value for these models. Since the transition from English to German represents a starker distribution shift than Pile to SlimPajama, training on German leads to significantly more forgetting on Pile (\(\mathcal{D}_{0}\)) for the continually pre-trained model without replay (0.27 vs 1.39 for weak and strong shifts respectively). However, choosing 25% replay to handle the starker shift significantly reduces the amount of forgetting on Pile, a reduction of 1.23 in terms of final loss. When comparing continually pre-trained models to baselines trained exclusively on \(\mathcal{D}_{1}\), we observe that the continually pre-trained models always have lower validation loss across both distribution shifts. When comparing the continually pre-trained models with the \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) baselines we find that both models achieve nearly identical (weak shift) or identical (strong shift) average final validation losses. This shows that for strong and weak distribution shifts, a simple and scalable combination of LR re-warming, LR re-decaying, and replay can achieve similar performance to the \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) baseline.
๊ทธ๋ฆผ 5๋Š” ์•ฝํ•œ shift(์œ„)์™€ ๊ฐ•ํ•œ shift(์•„๋ž˜)์— ๋Œ€ํ•ด 405M ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์˜ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘ validation loss๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 3์€ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ํ‰๊ท (๋งˆ์ง€๋ง‰ 100ํšŒ ๋ฐ˜๋ณต) ์ตœ์ข… loss ๊ฐ’์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์˜์–ด์—์„œ ๋…์ผ์–ด๋กœ ์ „ํ™˜ํ•˜๋Š” ๊ฒƒ์€ Pile์—์„œ SlimPajama๋กœ ์ „ํ™˜ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ๋šœ๋ ทํ•œ ๋ถ„ํฌ ์ด๋™์„ ๋‚˜ํƒ€๋‚ด๋ฏ€๋กœ, ๋…์ผ์–ด๋กœ ํ›ˆ๋ จํ•˜๋ฉด ๊ณ„์†์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์—์„œ ์žฌ์ƒ์ด ์—†๋Š” Pile(\(\mathcal{D}_{0}\))์— ๋Œ€ํ•ด ํ›จ์”ฌ ๋” ๋งŽ์€ ์žŠ์–ด๋ฒ„๋ฆผ์ด ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค(์•ฝํ•œ shift์™€ ๊ฐ•ํ•œ shift์˜ ๊ฒฝ์šฐ ๊ฐ๊ฐ 0.27 ๋Œ€ 1.39). ๊ทธ๋Ÿฌ๋‚˜ ๋” ๋šœ๋ ทํ•œ shift๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด 25% replay๋ฅผ ์„ ํƒํ•˜๋ฉด ์ตœ์ข… loss ์ธก๋ฉด์—์„œ 1.23์ด ๊ฐ์†Œํ•˜์—ฌ Pile์— ๋Œ€ํ•œ ์žŠ์–ด๋ฒ„๋ฆผ์˜ ์–‘์ด ํฌ๊ฒŒ ์ค„์–ด๋“ญ๋‹ˆ๋‹ค. ๊ณ„์† ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ \(\mathcal{D}_{1}\)์—์„œ๋งŒ ํ•™์Šต๋œ baseline๊ณผ ๋น„๊ตํ•˜๋ฉด, ๊ณ„์† ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ๋‘ ๋ถ„ํฌ ์ด๋™ ๋ชจ๋‘์—์„œ ํ•ญ์ƒ ๋” ๋‚ฎ์€ validation loss๋ฅผ ๊ฐ–๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณ„์† ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) baseline๊ณผ ๋น„๊ตํ•˜๋ฉด ๋‘ ๋ชจ๋ธ ๋ชจ๋‘ ๊ฑฐ์˜ ๋™์ผํ•œ(์•ฝํ•œ shift) ๋˜๋Š” ๋™์ผํ•œ(๊ฐ•ํ•œ shift) ํ‰๊ท  ์ตœ์ข… validation loss๋ฅผ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™๊ณผ ์•ฝํ•œ ๋ถ„ํฌ ์ด๋™์˜ ๊ฒฝ์šฐ, ๊ฐ„๋‹จํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ Learning Rate ์žฌ๊ฐ€์—ด, Learning Rate ์žฌ๊ฐ์†Œ ๋ฐ replay์˜ ์กฐํ•ฉ์œผ๋กœ \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) baseline๊ณผ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 5๋Š” ์•ฝํ•œ ์‹œํ”„ํŠธ(์œ„)์™€ ๊ฐ•ํ•œ ์‹œํ”„ํŠธ(์•„๋ž˜)์— ๋Œ€ํ•ด 405M ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์˜ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘ ๊ฒ€์ฆ ์†์‹ค์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 3์€ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ํ‰๊ท (๋งˆ์ง€๋ง‰ 100ํšŒ ๋ฐ˜๋ณต) ์ตœ์ข… ์†์‹ค ๊ฐ’์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์˜์–ด์—์„œ ๋…์ผ์–ด๋กœ ์ „ํ™˜ํ•˜๋Š” ๊ฒƒ์€ Pile์—์„œ SlimPajama๋กœ ์ „ํ™˜ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ๋šœ๋ ทํ•œ ๋ถ„ํฌ ์ด๋™์„ ๋‚˜ํƒ€๋‚ด๋ฏ€๋กœ, ๋…์ผ์–ด๋กœ ํ›ˆ๋ จํ•˜๋ฉด ๊ณ„์†์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์—์„œ ์žฌ์ƒ์ด ์—†๋Š” Pile(\(\mathcal{D}_{0}\))์— ๋Œ€ํ•ด ํ›จ์”ฌ ๋” ๋งŽ์€ ์žŠ์–ด๋ฒ„๋ฆผ์ด ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค(์•ฝํ•œ ์‹œํ”„ํŠธ์™€ ๊ฐ•ํ•œ ์‹œํ”„ํŠธ์˜ ๊ฒฝ์šฐ ๊ฐ๊ฐ 0.27 ๋Œ€ 1.39). ๊ทธ๋Ÿฌ๋‚˜ ๋” ๋šœ๋ ทํ•œ ์‹œํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด 25% ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ์„ ํƒํ•˜๋ฉด ์ตœ์ข… ์†์‹ค ์ธก๋ฉด์—์„œ 1.23์ด ๊ฐ์†Œํ•˜์—ฌ Pile์— ๋Œ€ํ•œ ์žŠ์–ด๋ฒ„๋ฆผ์˜ ์–‘์ด ํฌ๊ฒŒ ์ค„์–ด๋“ญ๋‹ˆ๋‹ค. ๊ณ„์† ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ $\mathcal{D}_{1}\)์—์„œ๋งŒ ํ•™์Šต๋œ ๋ฒ ์ด์Šค๋ผ์ธ๊ณผ ๋น„๊ตํ•˜๋ฉด, ๊ณ„์† ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ๋‘ ๋ถ„ํฌ ์ด๋™ ๋ชจ๋‘์—์„œ ํ•ญ์ƒ ๋” ๋‚ฎ์€ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์†์‹ค์„ ๊ฐ–๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณ„์† ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ $\mathcal{D}_{0}\cup\mathcal{D}_{1}\) ๋ฒ ์ด์Šค๋ผ์ธ๊ณผ ๋น„๊ตํ•˜๋ฉด ๋‘ ๋ชจ๋ธ ๋ชจ๋‘ ๊ฑฐ์˜ ๋™์ผํ•œ(์•ฝํ•œ ์‹œํ”„ํŠธ) ๋˜๋Š” ๋™์ผํ•œ(๊ฐ•ํ•œ ์‹œํ”„ํŠธ) ํ‰๊ท  ์ตœ์ข… ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์†์‹ค์„ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™๊ณผ ์•ฝํ•œ ๋ถ„ํฌ ์ด๋™์˜ ๊ฒฝ์šฐ, ๊ฐ„๋‹จํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ LR ์žฌ๊ฐ€์—ด, LR ์žฌ๊ฐ์†Œ ๋ฐ ๋ฆฌํ”Œ๋ ˆ์ด์˜ ์กฐํ•ฉ์œผ๋กœ $\mathcal{D}_{0}\cup\mathcal{D}_{1}\) ๋ฒ ์ด์Šค๋ผ์ธ๊ณผ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
**Exploration on Distribution Matching.** As shown by previous results, generating metadata extracted from the downstream instruction distribution indeed helps. However, in practice, the extracted or human-written metadata may not be able to precisely characterize the instruction distribution. Therefore, it is necessary to explore the performance of CodecLM when the distribution represented by instruction metadata does not fully match the test distribution. As the true test distribution is complicated and not known as a prior, we approximate various extent of distribution matching by random subsampling from the set of metadata \(\mathcal{M}\). To control the effect of data quantity, we keep the total number of instruction-response pairs the same for each case. For example, when subsampling \(20\%\) of \(\mathcal{M}\), we prompt the strong LLM to generate 5 times more instructions for each metadata accordingly. The result is shown in the upper part of Figure 4, and we did observe the trend that the better instruction metadata captures the underlying distribution, the better performance the target LLM can achieve. Moreover, when the metadata matching proportion is equal or greater than \(60\%\), we obtain close performance as the fully-matched result. This observation highlights CodecLM's robustness under potential instruction metadata mismatch.
**Exploration on Distribution Matching.** ์ด์ „ ๊ฒฐ๊ณผ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ๋‹ค์šด์ŠคํŠธ๋ฆผ instruction ๋ถ„ํฌ์—์„œ ์ถ”์ถœํ•œ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ด ์‹ค์ œ๋กœ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ๋กœ ์ถ”์ถœ๋œ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋‚˜ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋Š” instruction ๋ถ„ํฌ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ํŠน์„ฑํ™”ํ•˜์ง€ ๋ชปํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ instruction ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๊ฐ€ ๋‚˜ํƒ€๋‚ด๋Š” ๋ถ„ํฌ๊ฐ€ ํ…Œ์ŠคํŠธ ๋ถ„ํฌ์™€ ์™„์ „ํžˆ ์ผ์น˜ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ CodecLM์˜ ์„ฑ๋Šฅ์„ ํƒ์ƒ‰ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ ํ…Œ์ŠคํŠธ ๋ถ„ํฌ๋Š” ๋ณต์žกํ•˜๊ณ  ์‚ฌ์ „์œผ๋กœ ์•Œ ์ˆ˜ ์—†๊ธฐ ๋•Œ๋ฌธ์— ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ \(\mathcal{M}\)์—์„œ ๋ฌด์ž‘์œ„๋กœ ํ•˜์œ„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ •๋„์˜ ๋ถ„ํฌ ์ผ์น˜๋ฅผ ๊ทผ์‚ฌํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์ˆ˜๋Ÿ‰์˜ ํšจ๊ณผ๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ๊ฒฝ์šฐ์˜ ์ด instruction-response ์Œ ์ˆ˜๋ฅผ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(\mathcal{M}\)์˜ \(20\%\)๋ฅผ ํ•˜์œ„ ์ƒ˜ํ”Œ๋งํ•  ๋•Œ, ๊ฐ•๋ ฅํ•œ LLM์ด ๊ฐ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด 5๋ฐฐ ๋” ๋งŽ์€ instruction์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 4์˜ ์œ„์ชฝ์— ๋‚˜์™€ ์žˆ์œผ๋ฉฐ, instruction ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๊ฐ€ ๊ธฐ๋ณธ ๋ถ„ํฌ๋ฅผ ๋” ์ž˜ ํฌ์ฐฉํ• ์ˆ˜๋ก ๋Œ€์ƒ LLM์ด ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒฝํ–ฅ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ๋งค์นญ ๋น„์œจ์ด \(60\%\) ์ด์ƒ์ผ ๋•Œ ์™„์ „ํžˆ ์ผ์น˜ํ•˜๋Š” ๊ฒฐ๊ณผ์™€ ๊ฑฐ์˜ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ์„ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ์ž ์žฌ์ ์ธ instruction ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ๋ถˆ์ผ์น˜ ์‹œ CodecLM์˜ robustness๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
**๋ถ„ํฌ ๋งค์นญ์— ๋Œ€ํ•œ ํƒ์ƒ‰**. ์ด์ „ ๊ฒฐ๊ณผ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ๋‹ค์šด์ŠคํŠธ๋ฆผ ๋ช…๋ น์–ด ๋ถ„ํฌ์—์„œ ์ถ”์ถœํ•œ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ด ์‹ค์ œ๋กœ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ๋กœ ์ถ”์ถœ๋œ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋‚˜ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋Š” ๋ช…๋ น์–ด ๋ถ„ํฌ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ํŠน์„ฑํ™”ํ•˜์ง€ ๋ชปํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ช…๋ น์–ด ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๊ฐ€ ๋‚˜ํƒ€๋‚ด๋Š” ๋ถ„ํฌ๊ฐ€ ํ…Œ์ŠคํŠธ ๋ถ„ํฌ์™€ ์™„์ „ํžˆ ์ผ์น˜ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ CodecLM์˜ ์„ฑ๋Šฅ์„ ํƒ์ƒ‰ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ ํ…Œ์ŠคํŠธ ๋ถ„ํฌ๋Š” ๋ณต์žกํ•˜๊ณ  ์‚ฌ์ „์œผ๋กœ ์•Œ ์ˆ˜ ์—†๊ธฐ ๋•Œ๋ฌธ์— ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ \mathcal{M}\)์—์„œ ๋ฌด์ž‘์œ„๋กœ ํ•˜์œ„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ •๋„์˜ ๋ถ„ํฌ ์ผ์น˜๋ฅผ ๊ทผ์‚ฌํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์ˆ˜๋Ÿ‰์˜ ํšจ๊ณผ๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ๊ฒฝ์šฐ์˜ ์ด ๋ช…๋ น์–ด-์‘๋‹ต ์Œ ์ˆ˜๋ฅผ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \mathcal{M}\)์˜ \(20\%\)๋ฅผ ํ•˜์œ„ ์ƒ˜ํ”Œ๋งํ•  ๋•Œ, ๊ฐ•๋ ฅํ•œ LLM์ด ๊ฐ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด 5๋ฐฐ ๋” ๋งŽ์€ ๋ช…๋ น์–ด๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 4์˜ ์œ„์ชฝ์— ๋‚˜์™€ ์žˆ์œผ๋ฉฐ, ๋ช…๋ น ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๊ฐ€ ๊ธฐ๋ณธ ๋ถ„ํฌ๋ฅผ ๋” ์ž˜ ํฌ์ฐฉํ• ์ˆ˜๋ก ๋Œ€์ƒ LLM์ด ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒฝํ–ฅ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ๋งค์นญ ๋น„์œจ์ด 60\% ์ด์ƒ์ผ ๋•Œ ์™„์ „ํžˆ ์ผ์น˜ํ•˜๋Š” ๊ฒฐ๊ณผ์™€ ๊ฑฐ์˜ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ์„ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ์ž ์žฌ์ ์ธ ๋ช…๋ น ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ๋ถˆ์ผ์น˜ ์‹œ CodecLM์˜ ๊ฒฌ๊ณ ์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
Of the 660 items in the human eval set categorised as unsafe (based on the classifications provided by LlamaGuard), 358 are unsafe (based on the annotators' gold standard labels) and 302 are not unsafe. Accuracy is 54.2%. Of the 660 items in the human eval set categorised as not unsafe (based on the classifications provided by LlamaGuard), 88 are unsafe (based on the annotators' gold standard labels) and 572 are not unsafe. Accuracy is 86.7%. Overall, the accuracy of LlamaGuard is 70.4% on the human eval set. Due to the relatively small size of the human eval set, we do not analyse more finegrained categories, such as the performance of LlamaGuard on specific hazard categories or SUTs. The heavily imbalanced nature of the full dataset of 560,170 responses (of which only 533,125 or 95.2% are labelled Not Unsafe), and the better performance of the evaluator model on responses categorised as not unsafe, means that accuracy of LlamaGuard on the full dataset is higher than the eval set. If we weight the full dataset by the distribution of responses, accuracy is 85.6% (based on 86.7% x 95.2% plus 54.2% x 4.8%). Overall, although performance could be improved, LlamaGuard is sufficiently accurate at identifying unsafe and not unsafe responses to be used for the v0.5 Benchmark. We aim to improve performance of the evaluator model in future releases.
LlamaGuard์—์„œ ์ œ๊ณตํ•œ ๋ถ„๋ฅ˜๋ฅผ ๊ธฐ์ค€์œผ๋กœ unsafeํ•œ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋œ ์ธ๊ฐ„ ํ‰๊ฐ€ ์„ธํŠธ์˜ 660๊ฐœ ํ•ญ๋ชฉ ์ค‘ 358๊ฐœ ํ•ญ๋ชฉ์€ unsafeํ•œ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋˜์—ˆ์œผ๋ฉฐ(์ฃผ๊ด€์ž์˜ gold standard ๋ ˆ์ด๋ธ” ๊ธฐ์ค€), 302๊ฐœ ํ•ญ๋ชฉ์€ unsafeํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ •ํ™•๋„๋Š” 54.2%์ž…๋‹ˆ๋‹ค. ์ธ๊ฐ„ ํ‰๊ฐ€ ์„ธํŠธ์˜ 660๊ฐœ ํ•ญ๋ชฉ ์ค‘ unsafeํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋œ ํ•ญ๋ชฉ(LlamaGuard์—์„œ ์ œ๊ณตํ•œ ๋ถ„๋ฅ˜๋ฅผ ๊ธฐ์ค€) ์ค‘ 88๊ฐœ ํ•ญ๋ชฉ์€ unsafeํ•œ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋˜์—ˆ์œผ๋ฉฐ(์ฃผ๊ด€์ž์˜ gold standard ๋ ˆ์ด๋ธ” ๊ธฐ์ค€), 572๊ฐœ ํ•ญ๋ชฉ์€ unsafeํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ •ํ™•๋„๋Š” 86.7%์ž…๋‹ˆ๋‹ค. ์ „์ฒด์ ์œผ๋กœ ์ธ๊ฐ„ ํ‰๊ฐ€ ์„ธํŠธ์—์„œ LlamaGuard์˜ ์ •ํ™•๋„๋Š” 70.4%์ž…๋‹ˆ๋‹ค. ์ธ๊ฐ„ ํ‰๊ฐ€ ์„ธํŠธ์˜ ๊ทœ๋ชจ๊ฐ€ ์ƒ๋Œ€์ ์œผ๋กœ ์ž‘๊ธฐ ๋•Œ๋ฌธ์— ํŠน์ • hazard ๋ฒ”์ฃผ ๋˜๋Š” SUT์— ๋Œ€ํ•œ LlamaGuard์˜ ์„ฑ๋Šฅ๊ณผ ๊ฐ™์ด ๋” ์„ธ๋ถ„ํ™”๋œ ๋ฒ”์ฃผ๋ฅผ ๋ถ„์„ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. 560,170๊ฐœ์˜ ์‘๋‹ต์œผ๋กœ ๊ตฌ์„ฑ๋œ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ถˆ๊ท ํ˜•ํ•œ ํŠน์„ฑ(์ด ์ค‘ 533,125๊ฐœ ๋˜๋Š” 95.2%๋งŒ unsafeํ•˜์ง€ ์•Š์Œ์œผ๋กœ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋จ)๊ณผ unsafeํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋œ ์‘๋‹ต์— ๋Œ€ํ•œ evaluator ๋ชจ๋ธ์˜ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์œผ๋กœ ์ธํ•ด ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ LlamaGuard์˜ ์ •ํ™•๋„๊ฐ€ ํ‰๊ฐ€ ์„ธํŠธ๋ณด๋‹ค ๋†’์Šต๋‹ˆ๋‹ค. ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์‘๋‹ต ๋ถ„ํฌ์— ๋”ฐ๋ผ ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜๋ฉด ์ •ํ™•๋„๋Š” 85.6%(86.7% x 95.2% + 54.2% x 4.8% ๊ธฐ์ค€)์ž…๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, LlamaGuard๋Š” v0.5 Benchmark์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์„ ๋งŒํผ unsafeํ•œ ์‘๋‹ต์„ ์‹๋ณ„ํ•˜๋Š” ๋ฐ ์ถฉ๋ถ„ํžˆ ์ •ํ™•ํ•ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ๋ฆด๋ฆฌ์Šค์—์„œ๋Š” evaluator ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•  ๊ณ„ํš์ž…๋‹ˆ๋‹ค.
LlamaGuard์—์„œ ์ œ๊ณตํ•œ ๋ถ„๋ฅ˜๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋œ ์ธ๊ฐ„ ํ‰๊ฐ€ ์„ธํŠธ์˜ 660๊ฐœ ํ•ญ๋ชฉ ์ค‘ 358๊ฐœ ํ•ญ๋ชฉ์€ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋˜์—ˆ์œผ๋ฉฐ(์ฃผ๊ด€์ž์˜ ๊ณจ๋“œ ํ‘œ์ค€ ๋ ˆ์ด๋ธ” ๊ธฐ์ค€), 302๊ฐœ ํ•ญ๋ชฉ์€ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ •ํ™•๋„๋Š” 54.2%์ž…๋‹ˆ๋‹ค. ์ธ๊ฐ„ ํ‰๊ฐ€ ์„ธํŠธ์˜ 660๊ฐœ ํ•ญ๋ชฉ ์ค‘ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋œ ํ•ญ๋ชฉ(LlamaGuard์—์„œ ์ œ๊ณตํ•œ ๋ถ„๋ฅ˜๋ฅผ ๊ธฐ์ค€) ์ค‘ 88๊ฐœ ํ•ญ๋ชฉ์€ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋˜์—ˆ์œผ๋ฉฐ(์ฃผ๊ด€์ž์˜ ๊ณจ๋“œ ํ‘œ์ค€ ๋ ˆ์ด๋ธ” ๊ธฐ์ค€), 572๊ฐœ ํ•ญ๋ชฉ์€ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ •ํ™•๋„๋Š” 86.7%์ž…๋‹ˆ๋‹ค. ์ „์ฒด์ ์œผ๋กœ ์ธ๊ฐ„ ํ‰๊ฐ€ ์„ธํŠธ์—์„œ LlamaGuard์˜ ์ •ํ™•๋„๋Š” 70.4%์ž…๋‹ˆ๋‹ค. ์ธ๊ฐ„ ํ‰๊ฐ€ ์„ธํŠธ์˜ ๊ทœ๋ชจ๊ฐ€ ์ƒ๋Œ€์ ์œผ๋กœ ์ž‘๊ธฐ ๋•Œ๋ฌธ์— ํŠน์ • ์œ„ํ—˜ ๋ฒ”์ฃผ ๋˜๋Š” SUT์— ๋Œ€ํ•œ LlamaGuard์˜ ์„ฑ๋Šฅ๊ณผ ๊ฐ™์ด ๋” ์„ธ๋ถ„ํ™”๋œ ๋ฒ”์ฃผ๋ฅผ ๋ถ„์„ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. 560,170๊ฐœ์˜ ์‘๋‹ต์œผ๋กœ ๊ตฌ์„ฑ๋œ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ถˆ๊ท ํ˜•ํ•œ ํŠน์„ฑ(์ด ์ค‘ 533,125๊ฐœ ๋˜๋Š” 95.2%๋งŒ ์•ˆ์ „ํ•˜์ง€ ์•Š์Œ์œผ๋กœ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋จ)๊ณผ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋œ ์‘๋‹ต์— ๋Œ€ํ•œ ํ‰๊ฐ€์ž ๋ชจ๋ธ์˜ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์œผ๋กœ ์ธํ•ด ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ LlamaGuard์˜ ์ •ํ™•๋„๊ฐ€ ํ‰๊ฐ€ ์„ธํŠธ๋ณด๋‹ค ๋†’์Šต๋‹ˆ๋‹ค. ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์‘๋‹ต ๋ถ„ํฌ์— ๋”ฐ๋ผ ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜๋ฉด ์ •ํ™•๋„๋Š” 85.6%(86.7% x 95.2% + 54.2% x 4.8% ๊ธฐ์ค€)์ž…๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, LlamaGuard๋Š” v0.5 ๋ฒค์น˜๋งˆํฌ์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์„ ๋งŒํผ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ์‘๋‹ต์„ ์‹๋ณ„ํ•˜๋Š” ๋ฐ ์ถฉ๋ถ„ํžˆ ์ •ํ™•ํ•ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ๋ฆด๋ฆฌ์Šค์—์„œ๋Š” ํ‰๊ฐ€์ž ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•  ๊ณ„ํš์ž…๋‹ˆ๋‹ค.
**LLM-enhanced Recommendation Models.** In addition to instructing LLMs to directly provide recommendations, researchers also propose leveraging the universal knowledge encoded in LLMs to improve traditional recommender systems. Existing approaches in this line can be divided into three main categories. The first category employs LLMs to infer users' potential intention from their historical interaction data. Furthermore, traditional recommendation/search models employ the inferred intentions to improve the retrieval of relevant items [812, 813]. Additionally, several studies explore the use of LLMs as feature encoders. They employ LLMs to encode the side information of items and users (_e.g._, item's descriptions and user's reviews), thus deriving more informative representations of users and items. These representations are then fed into traditional recommender systems as augmented input [814, 815]. As another alternative approach, several studies [816, 817] adopt a distillation-like way to transfer LLM's capacities (_e.g._, semantic encoding) to improve traditional recommenders (_i.e._, small models). Specially, they align the hidden states of LLMs and traditional recommendation models via joint training. After training, since only the enhanced small model will be deployed online, it can avoid the huge overhead of LLMs in online service.
**LLM ๊ฐ•ํ™” ์ถ”์ฒœ ๋ชจ๋ธ.** ์—ฐ๊ตฌ์ž๋“ค์€ LLM์— ์ง์ ‘ ์ถ”์ฒœ์„ ์ œ๊ณตํ•˜๋„๋ก ์ง€์‹œํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ๊ธฐ์กด ์ถ”์ฒœ ์‹œ์Šคํ…œ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด LLM์— ์ธ์ฝ”๋”ฉ๋œ ๋ณดํŽธ์  ์ง€์‹์„ ํ™œ์šฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํฌ๊ฒŒ ์„ธ ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋ฒ”์ฃผ๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์šฉ์ž์˜ ๊ณผ๊ฑฐ ์ƒํ˜ธ ์ž‘์šฉ ๋ฐ์ดํ„ฐ์—์„œ ์‚ฌ์šฉ์ž์˜ ์ž ์žฌ์  ์˜๋„๋ฅผ ์ถ”๋ก ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด์˜ ์ถ”์ฒœ/๊ฒ€์ƒ‰ ๋ชจ๋ธ์€ ์ถ”๋ก ๋œ ์˜๋„๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ด€๋ จ ํ•ญ๋ชฉ์˜ ๊ฒ€์ƒ‰์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค[812, 813]. ๋˜ํ•œ, ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ LLM์„ ๊ธฐ๋Šฅ ์ธ์ฝ”๋”๋กœ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํƒ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์•„์ดํ…œ๊ณผ ์‚ฌ์šฉ์ž์˜ ๋ถ€๊ฐ€ ์ •๋ณด(์˜ˆ: ์•„์ดํ…œ์˜ ์„ค๋ช… ๋ฐ ์‚ฌ์šฉ์ž์˜ ๋ฆฌ๋ทฐ)๋ฅผ ์ธ์ฝ”๋”ฉํ•˜์—ฌ ์‚ฌ์šฉ์ž์™€ ์•„์ดํ…œ์— ๋Œ€ํ•œ ๋ณด๋‹ค ์œ ์ตํ•œ ํ‘œํ˜„์„ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ํ‘œํ˜„์„ ๊ธฐ์กด ์ถ”์ฒœ ์‹œ์Šคํ…œ์— ์ฆ๊ฐ•๋œ ์ž…๋ ฅ์œผ๋กœ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค[814, 815]. ๋˜ ๋‹ค๋ฅธ ๋Œ€์•ˆ์œผ๋กœ, ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ[816, 817]์—์„œ๋Š” ์ „ํ†ต์ ์ธ ์ถ”์ฒœ ์‹œ์Šคํ…œ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด LLM์˜ ์—ญ๋Ÿ‰(์˜ˆ: semantic encoding)์„ ์ „์ˆ˜ํ•˜๋Š” ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์ด๋“ค์€ ๊ณต๋™ ํ•™์Šต์„ ํ†ตํ•ด LLM๊ณผ ๊ธฐ์กด ์ถ”์ฒœ ๋ชจ๋ธ์˜ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ๋ฅผ ์ •๋ ฌํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต์ด ๋๋‚˜๋ฉด ๊ฐ•ํ™”๋œ ์†Œํ˜• ๋ชจ๋ธ๋งŒ ์˜จ๋ผ์ธ์— ๋ฐฐํฌ๋˜๋ฏ€๋กœ ์˜จ๋ผ์ธ ์„œ๋น„์Šค์—์„œ LLM์˜ ๋ง‰๋Œ€ํ•œ ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ํ”ผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**LLM ๊ฐ•ํ™” ์ถ”์ฒœ ๋ชจ๋ธ.** ์—ฐ๊ตฌ์ž๋“ค์€ LLM์— ์ง์ ‘ ์ถ”์ฒœ์„ ์ œ๊ณตํ•˜๋„๋ก ์ง€์‹œํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ๊ธฐ์กด ์ถ”์ฒœ ์‹œ์Šคํ…œ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด LLM์— ์ธ์ฝ”๋”ฉ๋œ ๋ณดํŽธ์  ์ง€์‹์„ ํ™œ์šฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํฌ๊ฒŒ ์„ธ ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋ฒ”์ฃผ๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์šฉ์ž์˜ ๊ณผ๊ฑฐ ์ƒํ˜ธ ์ž‘์šฉ ๋ฐ์ดํ„ฐ์—์„œ ์‚ฌ์šฉ์ž์˜ ์ž ์žฌ์  ์˜๋„๋ฅผ ์ถ”๋ก ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด์˜ ์ถ”์ฒœ/๊ฒ€์ƒ‰ ๋ชจ๋ธ์€ ์ถ”๋ก ๋œ ์˜๋„๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ด€๋ จ ํ•ญ๋ชฉ์˜ ๊ฒ€์ƒ‰์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค[812, 813]. ๋˜ํ•œ, ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ LLM์„ ๊ธฐ๋Šฅ ์ธ์ฝ”๋”๋กœ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํƒ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์•„์ดํ…œ๊ณผ ์‚ฌ์šฉ์ž์˜ ๋ถ€๊ฐ€ ์ •๋ณด(์˜ˆ: ์•„์ดํ…œ์˜ ์„ค๋ช… ๋ฐ ์‚ฌ์šฉ์ž์˜ ๋ฆฌ๋ทฐ)๋ฅผ ์ธ์ฝ”๋”ฉํ•˜์—ฌ ์‚ฌ์šฉ์ž์™€ ์•„์ดํ…œ์— ๋Œ€ํ•œ ๋ณด๋‹ค ์œ ์ตํ•œ ํ‘œํ˜„์„ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ํ‘œํ˜„์„ ๊ธฐ์กด ์ถ”์ฒœ ์‹œ์Šคํ…œ์— ์ฆ๊ฐ•๋œ ์ž…๋ ฅ์œผ๋กœ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค[814, 815]. ๋˜ ๋‹ค๋ฅธ ๋Œ€์•ˆ์œผ๋กœ, ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ[816, 817]์—์„œ๋Š” ์ „ํ†ต์ ์ธ ์ถ”์ฒœ ์‹œ์Šคํ…œ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด LLM์˜ ์—ญ๋Ÿ‰(์˜ˆ: ์˜๋ฏธ ์ธ์ฝ”๋”ฉ)์„ ์ „์ˆ˜ํ•˜๋Š” ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์ด๋“ค์€ ๊ณต๋™ ํ•™์Šต์„ ํ†ตํ•ด LLM๊ณผ ๊ธฐ์กด ์ถ”์ฒœ ๋ชจ๋ธ์˜ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ๋ฅผ ์ •๋ ฌํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต์ด ๋๋‚˜๋ฉด ๊ฐ•ํ™”๋œ ์†Œํ˜• ๋ชจ๋ธ๋งŒ ์˜จ๋ผ์ธ์— ๋ฐฐํฌ๋˜๋ฏ€๋กœ ์˜จ๋ผ์ธ ์„œ๋น„์Šค์—์„œ LLM์˜ ๋ง‰๋Œ€ํ•œ ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ํ”ผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Retrieval Aware Fine-Tuning (RAFT), presents a novel recipe to prepare fine-tuning data to tailor the models for domain specific open-book settings, equivalent to in-domain RAG In RAFT, we prepare the training data such that each data point contains a question (\(Q\)), a set of documents (\(D_{k}\)), and a corresponding Chain-of-though style answer (\(A^{*}\)) generated from one of the document (\(D^{*}\)). We differentiate between two types of documents: 'oracle' documents (\(D*\)) i.e. the documents from which the answer to the question can be deduced, and 'distractor' documents (\(D_{i}\)) that do not contain answer-relevant information. As an implementation detail, the 'oracle' document doesn't need to be a single document, but can be more than one document, as is the case in HotpotQA (Yang et al., 2018). Then, for \(P\) fraction of the questions (\(q_{i}\)) in the dataset, we retain the oracle document (\(d_{i}^{*}\)) along with distractor documents (\(d_{k-1}\)). For \((1-P)\) fraction of the questions (\(q_{i}\)) in the dataset, we include no oracle document and only include distractor documents (\(d_{k}\)). We then fine-tune the language model using the standard supervised training (SFT) technique, training it to generate answers from the provided documents and questions. Fig. 2 illustrates the high-level design principal for RAFT.
๋ฆฌํŠธ๋ฆฌ๋ฒŒ ์ธ์‹ Fine-tuning (RAFT)์€ ๋„๋ฉ”์ธ ํŠน์ • ์˜คํ”ˆ ๋ถ ์„ค์ •์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์กฐ์ •ํ•˜๊ธฐ ์œ„ํ•œ Fine-tuning ๋ฐ์ดํ„ฐ๋ฅผ ์ค€๋น„ํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ ˆ์‹œํ”ผ๋ฅผ ์ œ์‹œํ•˜๋ฉฐ, ์ด๋Š” ๋„๋ฉ”์ธ ๋‚ด RAG์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. RAFT์—์„œ๋Š” ๊ฐ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์— ์งˆ๋ฌธ(\(Q\)), ๋ฌธ์„œ ์ง‘ํ•ฉ(\(D_{k}\)), ๊ทธ๋ฆฌ๊ณ  ๋ฌธ์„œ ์ค‘ ํ•˜๋‚˜์—์„œ ์ƒ์„ฑ๋œ ํ•ด๋‹น ์‚ฌ๊ณ ์˜ ์‚ฌ์Šฌ ์Šคํƒ€์ผ ๋‹ต๋ณ€(\(A^{*}\))์„ ํฌํ•จํ•˜๋„๋ก ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ์ค€๋น„ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต์„ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฌธ์„œ(์ฆ‰, '์˜ค๋ผํด' ๋ฌธ์„œ(\(D*\))์™€ ๋‹ต๊ณผ ๊ด€๋ จ๋œ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์ง€ ์•Š์€ '๋ฐฉํ•ด ์š”์†Œ' ๋ฌธ์„œ(\(D_{i}\))์˜ ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์˜ ๋ฌธ์„œ๋ฅผ ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌํ˜„ ์„ธ๋ถ€ ์‚ฌํ•ญ์œผ๋กœ์„œ, '์˜ค๋ผํด' ๋ฌธ์„œ๊ฐ€ ํ•˜๋‚˜์˜ ๋ฌธ์„œ์ผ ํ•„์š”๋Š” ์—†์œผ๋ฉฐ, HotpotQA(Yang et al., 2018)์˜ ๊ฒฝ์šฐ์ฒ˜๋Ÿผ ํ•˜๋‚˜ ์ด์ƒ์˜ ๋ฌธ์„œ๊ฐ€ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์งˆ๋ฌธ(\(q_{i}\)) ์ค‘ \(P\) ๋ถ„์ˆ˜์— ๋Œ€ํ•ด ์˜ค๋ผํด ๋ฌธ์„œ(\(d_{i}^{*}\))๋ฅผ ๋ฐฉํ•ด ์š”์†Œ ๋ฌธ์„œ(\(d_{k-1}\))์™€ ํ•จ๊ป˜ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ \((1-P)\) ๋ถ„์ˆ˜์˜ ์งˆ๋ฌธ(\(q_{i}\))์— ๋Œ€ํ•ด์„œ๋Š” ์˜ค๋ผํด ๋ฌธ์„œ๋ฅผ ํฌํ•จํ•˜์ง€ ์•Š๊ณ  ๋ฐฉํ•ด ์š”์†Œ ๋ฌธ์„œ(\(d_{k}\))๋งŒ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ‘œ์ค€ ์ง€๋„ ํ›ˆ๋ จ(SFT) ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์„ Fine-tuningํ•˜์—ฌ ์ œ๊ณต๋œ ๋ฌธ์„œ์™€ ์งˆ๋ฌธ์—์„œ ๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2๋Š” RAFT์˜ ๋†’์€ ์ˆ˜์ค€์˜ ์„ค๊ณ„ ์›๋ฆฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๋ฆฌํŠธ๋ฆฌ๋ฒŒ ์ธ์‹ ๋ฏธ์„ธ ์กฐ์ •(RAFT)์€ ๋„๋ฉ”์ธ ํŠน์ • ์˜คํ”ˆ ๋ถ ์„ค์ •์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์กฐ์ •ํ•˜๊ธฐ ์œ„ํ•œ ๋ฏธ์„ธ ์กฐ์ • ๋ฐ์ดํ„ฐ๋ฅผ ์ค€๋น„ํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ ˆ์‹œํ”ผ๋ฅผ ์ œ์‹œํ•˜๋ฉฐ, ์ด๋Š” ๋„๋ฉ”์ธ ๋‚ด RAG์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. RAFT์—์„œ๋Š” ๊ฐ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์— ์งˆ๋ฌธ(\(Q\)), ๋ฌธ์„œ ์ง‘ํ•ฉ(\(D_{k}\)), ๊ทธ๋ฆฌ๊ณ  ๋ฌธ์„œ ์ค‘ ํ•˜๋‚˜์—์„œ ์ƒ์„ฑ๋œ ํ•ด๋‹น ์‚ฌ๊ณ ์˜ ์‚ฌ์Šฌ ์Šคํƒ€์ผ ๋‹ต๋ณ€(\(A^{*}\))์„ ํฌํ•จํ•˜๋„๋ก ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ์ค€๋น„ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต์„ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฌธ์„œ(์ฆ‰, '์˜ค๋ผํด' ๋ฌธ์„œ(\(D*\))์™€ ๋‹ต๊ณผ ๊ด€๋ จ๋œ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์ง€ ์•Š์€ '๋ฐฉํ•ด ์š”์†Œ' ๋ฌธ์„œ(\(D_{i}\))์˜ ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์˜ ๋ฌธ์„œ๋ฅผ ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌํ˜„ ์„ธ๋ถ€ ์‚ฌํ•ญ์œผ๋กœ์„œ, '์˜ค๋ผํด' ๋ฌธ์„œ๊ฐ€ ํ•˜๋‚˜์˜ ๋ฌธ์„œ์ผ ํ•„์š”๋Š” ์—†์œผ๋ฉฐ, HotpotQA(Yang et al., 2018)์˜ ๊ฒฝ์šฐ์ฒ˜๋Ÿผ ํ•˜๋‚˜ ์ด์ƒ์˜ ๋ฌธ์„œ๊ฐ€ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์งˆ๋ฌธ(\(q_{i}\)) ์ค‘ \(P\) ๋ถ„์ˆ˜์— ๋Œ€ํ•ด ์˜ค๋ผํด ๋ฌธ์„œ(\(d_{i}^{*}\))๋ฅผ ๋ฐฉํ•ด ์š”์†Œ ๋ฌธ์„œ(\(d_{k-1}\))์™€ ํ•จ๊ป˜ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ \((1-P)\) ๋ถ„์ˆ˜์˜ ์งˆ๋ฌธ(\(q_{i}\))์— ๋Œ€ํ•ด์„œ๋Š” ์˜ค๋ผํด ๋ฌธ์„œ๋ฅผ ํฌํ•จํ•˜์ง€ ์•Š๊ณ  ๋ฐฉํ•ด ์š”์†Œ ๋ฌธ์„œ(\(d_{k}\))๋งŒ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ‘œ์ค€ ์ง€๋„ ํ›ˆ๋ จ(SFT) ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์ œ๊ณต๋œ ๋ฌธ์„œ์™€ ์งˆ๋ฌธ์—์„œ ๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2๋Š” RAFT์˜ ๋†’์€ ์ˆ˜์ค€์˜ ์„ค๊ณ„ ์›๋ฆฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
Training and generating with critics.Training LLMs with reinforcement learning (e.g., Proximal Policy Optimization or PPO; Schulman et al. 2017) from human feedback (RLHF) has proven effective in aligning LLMs with human preferences (Ouyang et al., 2022). Wu et al. (2023) introduce fine-grained RLHF with multiple reward models. Though our work also studies fine-grained critique on retrieval and generation, we train our target LM on task examples augmented with reflection tokens from a critic model offline, with a far lower training cost compared to RLHF. In addition, reflection tokens in Self-Rag enable controllable generation at inference, while RLHF focuses on human preference alignment during training. Other works use general control tokens to guide LM generation (Lu et al., 2022; Korbak et al., 2023), while Self-Rag uses reflection tokens to decide the need for retrieval and to self-evaluate generation quality. Xie et al. (2023) propose a self-evaluation-guided decoding framework, but they focus only on reasoning tasks with one evaluation dimension (reasoning path consistency) and without retrieval. Recent work on LLM refinement (Dhuliawala et al., 2023; Madaan et al., 2023; Paul et al., 2023) prompts a model to generate task output, natural language feedback and refined task output iteratively, but at the cost of inference efficiency.
๋น„ํ‰๊ฐ€์™€ ํ•จ๊ป˜ ํ›ˆ๋ จ ๋ฐ ์ƒ์„ฑ. ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ(RLHF)์œผ๋กœ๋ถ€ํ„ฐ์˜ Reinforcement Learning(์˜ˆ: Proximal Policy Optimization ๋˜๋Š” PPO; Schulman et al. 2017)์„ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋”ฐ๋ผ LLM์„ ์กฐ์ •ํ•˜๋Š” ๋ฐ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Ouyang et al., 2022). Wu et al. (2023)์€ ์—ฌ๋Ÿฌ ๋ณด์ƒ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์„ธ๋ถ„ํ™”๋œ RLHF๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ์—ฐ๊ตฌ์—์„œ๋„ ๊ฒ€์ƒ‰ ๋ฐ ์ƒ์„ฑ์— ๋Œ€ํ•œ ์„ธ๋ถ„ํ™”๋œ ๋น„ํ‰์„ ์—ฐ๊ตฌํ•˜์ง€๋งŒ, RLHF์— ๋น„ํ•ด ํ›จ์”ฌ ๋‚ฎ์€ ํ›ˆ๋ จ ๋น„์šฉ์œผ๋กœ ์˜คํ”„๋ผ์ธ์—์„œ ๋น„ํ‰ ๋ชจ๋ธ์˜ ๋ฐ˜์˜ ํ† ํฐ์œผ๋กœ ์ฆ๊ฐ•๋œ ์ž‘์—… ์˜ˆ์ œ์— ๋Œ€ํ•ด ๋Œ€์ƒ LM์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Self-Rag์˜ ๋ฐ˜์‚ฌ ํ† ํฐ์€ ์ถ”๋ก  ์ค‘์— ์ œ์–ด ๊ฐ€๋Šฅํ•œ ์ƒ์„ฑ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๋ฐ˜๋ฉด, RLHF๋Š” ํ›ˆ๋ จ ์ค‘์— ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„ ์กฐ์ •์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์ž‘์—…์—์„œ๋Š” ์ผ๋ฐ˜์ ์ธ ์ œ์–ด ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ LM ์ƒ์„ฑ์„ ์•ˆ๋‚ดํ•˜๋Š” ๋ฐ˜๋ฉด(Lu et al., 2022; Korbak et al., 2023), Self-Rag์€ ๋ฐ˜์‚ฌ ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰์˜ ํ•„์š”์„ฑ์„ ๊ฒฐ์ •ํ•˜๊ณ  ์ƒ์„ฑ ํ’ˆ์งˆ์„ ์ž์ฒด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Xie et al. (2023)์€ ์ž์ฒด ํ‰๊ฐ€ ๊ธฐ๋ฐ˜ ๋””์ฝ”๋”ฉ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•˜์ง€๋งŒ, ์ด๋Š” ๊ฒ€์ƒ‰ ์—†์ด ํ•˜๋‚˜์˜ ํ‰๊ฐ€ ์ฐจ์›(์ถ”๋ก  ๊ฒฝ๋กœ ์ผ๊ด€์„ฑ)๋งŒ ์žˆ๋Š” ์ถ”๋ก  ์ž‘์—…์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ LLM refinement(Dhuliawala et al., 2023; Madaan et al., 2023; Paul et al., 2023) ์ž‘์—…์—์„œ๋Š” ๋ชจ๋ธ์ด ์ž‘์—… ์ถœ๋ ฅ, ์ž์—ฐ์–ด ํ”ผ๋“œ๋ฐฑ ๋ฐ ์ˆ˜์ •๋œ ์ž‘์—… ์ถœ๋ ฅ์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•˜์ง€๋งŒ ์ถ”๋ก  ํšจ์œจ์„ฑ์˜ ๋Œ€๊ฐ€๋ฅผ ์น˜๋ฅด๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.
๋น„ํ‰๊ฐ€์™€ ํ•จ๊ป˜ ํ›ˆ๋ จ ๋ฐ ์ƒ์„ฑ.์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ(RLHF)์œผ๋กœ๋ถ€ํ„ฐ์˜ ๊ฐ•ํ™” ํ•™์Šต(์˜ˆ: ๊ทผ์ ‘ ์ •์ฑ… ์ตœ์ ํ™” ๋˜๋Š” PPO; Schulman et al. 2017)์„ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋”ฐ๋ผ LLM์„ ์กฐ์ •ํ•˜๋Š” ๋ฐ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Ouyang et al., 2022). Wu et al. (2023)์€ ์—ฌ๋Ÿฌ ๋ณด์ƒ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์„ธ๋ถ„ํ™”๋œ RLHF๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ์—ฐ๊ตฌ์—์„œ๋„ ๊ฒ€์ƒ‰ ๋ฐ ์ƒ์„ฑ์— ๋Œ€ํ•œ ์„ธ๋ถ„ํ™”๋œ ๋น„ํ‰์„ ์—ฐ๊ตฌํ•˜์ง€๋งŒ, RLHF์— ๋น„ํ•ด ํ›จ์”ฌ ๋‚ฎ์€ ํ›ˆ๋ จ ๋น„์šฉ์œผ๋กœ ์˜คํ”„๋ผ์ธ์—์„œ ๋น„ํ‰ ๋ชจ๋ธ์˜ ๋ฐ˜์˜ ํ† ํฐ์œผ๋กœ ์ฆ๊ฐ•๋œ ์ž‘์—… ์˜ˆ์ œ์— ๋Œ€ํ•ด ๋Œ€์ƒ LM์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Self-Rag์˜ ๋ฐ˜์‚ฌ ํ† ํฐ์€ ์ถ”๋ก  ์ค‘์— ์ œ์–ด ๊ฐ€๋Šฅํ•œ ์ƒ์„ฑ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๋ฐ˜๋ฉด, RLHF๋Š” ํ›ˆ๋ จ ์ค‘์— ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„ ์กฐ์ •์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์ž‘์—…์—์„œ๋Š” ์ผ๋ฐ˜์ ์ธ ์ œ์–ด ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ LM ์ƒ์„ฑ์„ ์•ˆ๋‚ดํ•˜๋Š” ๋ฐ˜๋ฉด(Lu et al., 2022; Korbak et al., 2023), Self-Rag์€ ๋ฐ˜์‚ฌ ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰์˜ ํ•„์š”์„ฑ์„ ๊ฒฐ์ •ํ•˜๊ณ  ์ƒ์„ฑ ํ’ˆ์งˆ์„ ์ž์ฒด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Xie et al. (2023)์€ ์ž์ฒด ํ‰๊ฐ€ ๊ธฐ๋ฐ˜ ๋””์ฝ”๋”ฉ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•˜์ง€๋งŒ, ์ด๋Š” ๊ฒ€์ƒ‰ ์—†์ด ํ•˜๋‚˜์˜ ํ‰๊ฐ€ ์ฐจ์›(์ถ”๋ก  ๊ฒฝ๋กœ ์ผ๊ด€์„ฑ)๋งŒ ์žˆ๋Š” ์ถ”๋ก  ์ž‘์—…์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ LLM ๊ตฌ์ฒดํ™”(Dhuliawala et al., 2023; Madaan et al., 2023; Paul et al., 2023) ์ž‘์—…์—์„œ๋Š” ๋ชจ๋ธ์ด ์ž‘์—… ์ถœ๋ ฅ, ์ž์—ฐ์–ด ํ”ผ๋“œ๋ฐฑ ๋ฐ ์ˆ˜์ •๋œ ์ž‘์—… ์ถœ๋ ฅ์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•˜์ง€๋งŒ ์ถ”๋ก  ํšจ์œจ์„ฑ์˜ ๋Œ€๊ฐ€๋ฅผ ์น˜๋ฅด๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.
The recent success of Large Language Models (LLMs) has been predominantly driven by curating the training dataset composition, scaling of model architectures and dataset sizes and advancements in pretraining objectives, leaving tokenizer influence as a blind spot. Shedding light on this underexplored area, we conduct a comprehensive study on the influence of tokenizer choice on LLM downstream performance by training 24 mono- and multilingual LLMs at a 2.6 B parameter scale, ablating different tokenizer algorithms and parameterizations. Our studies highlight that the tokenizer choice can significantly impact the model's downstream performance and training costs. In particular, we find that the common tokenizer evaluation metrics _fertility_ and _parity_ are not always predictive of model downstream performance, rendering these metrics a questionable proxy for the model's downstream performance. Furthermore, we show that multilingual tokenizers trained on the five most frequent European languages require vocabulary size increases of factor three in comparison to English. While English-centric tokenizers have been applied to the training of multi-lingual LLMs in the past, we find that this approach results in a severe downstream performance degradation and additional training costs of up to 68%, due to an inefficient tokenization vocabulary.
์ตœ๊ทผ Large Language Models (LLMs)์˜ ์„ฑ๊ณต์€ ์ฃผ๋กœ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ตฌ์„ฑ, ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ ๋ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ ํฌ๊ธฐ์˜ ํ™•์žฅ, ์‚ฌ์ „ ํ›ˆ๋ จ ๋ชฉํ‘œ์˜ ๋ฐœ์ „์— ์˜ํ•ด ์ฃผ๋„๋˜์–ด ์™”์œผ๋ฉฐ, tokenizer์˜ ์˜ํ–ฅ์€ ์‚ฌ๊ฐ์ง€๋Œ€์— ๋†“์—ฌ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์ œ ๋ง‰ ํƒ๊ตฌ๋˜๊ธฐ ์‹œ์ž‘ํ•œ ์ด ๋ถ„์•ผ์— ์กฐ๋ช…์„ ๋น„์ถ”๊ธฐ ์œ„ํ•ด 24๊ฐœ์˜ ๋‹จ์ผ ์–ธ์–ด ๋ฐ ๋‹ค๊ตญ์–ด LLM์„ 2.6B ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ๋กœ ํ›ˆ๋ จํ•˜๊ณ  ๋‹ค์–‘ํ•œ tokenizer ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์ œ๊ฑฐํ•จ์œผ๋กœ์จ tokenizer ์„ ํƒ์ด LLM downstream ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, tokenizer ์„ ํƒ์ด ๋ชจ๋ธ์˜ downstream ์„ฑ๋Šฅ๊ณผ ํ›ˆ๋ จ ๋น„์šฉ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Œ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ผ๋ฐ˜์ ์ธ tokenizer ํ‰๊ฐ€ ์ง€ํ‘œ์ธ _fertility_์™€ _parity_๊ฐ€ ํ•ญ์ƒ ๋ชจ๋ธ์˜ downstream ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ด๋Ÿฌํ•œ ์ง€ํ‘œ๊ฐ€ ๋ชจ๋ธ์˜ downstream ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ์˜์‹ฌ์Šค๋Ÿฌ์šด ๋Œ€๋ฆฌ ์ง€ํ‘œ์ž„์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๊ฐ€์žฅ ๋นˆ๋ฒˆํ•œ 5๊ฐœ์˜ ์œ ๋Ÿฝ ์–ธ์–ด๋กœ ํ›ˆ๋ จ๋œ ๋‹ค๊ตญ์–ด tokenizer๋Š” ์˜์–ด์— ๋น„ํ•ด ์–ดํœ˜ ํฌ๊ธฐ๊ฐ€ 3๋ฐฐ ์ด์ƒ ์ฆ๊ฐ€ํ•ด์•ผ ํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ณผ๊ฑฐ์—๋Š” ์˜์–ด ์ค‘์‹ฌ์˜ tokenizer๊ฐ€ ๋‹ค๊ตญ์–ด LLM ํ›ˆ๋ จ์— ์ ์šฉ๋˜์—ˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ tokenizer ์–ดํœ˜๊ฐ€ ๋น„ํšจ์œจ์ ์ด๊ธฐ ๋•Œ๋ฌธ์— downstream ์„ฑ๋Šฅ์ด ์‹ฌ๊ฐํ•˜๊ฒŒ ์ €ํ•˜๋˜๊ณ  ์ตœ๋Œ€ 68%์˜ ์ถ”๊ฐ€ ํ›ˆ๋ จ ๋น„์šฉ์ด ๋ฐœ์ƒํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
์ตœ๊ทผ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์˜ ์„ฑ๊ณต์€ ์ฃผ๋กœ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ตฌ์„ฑ, ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ ๋ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ ํฌ๊ธฐ์˜ ํ™•์žฅ, ์‚ฌ์ „ ํ›ˆ๋ จ ๋ชฉํ‘œ์˜ ๋ฐœ์ „์— ์˜ํ•ด ์ฃผ๋„๋˜์–ด ์™”์œผ๋ฉฐ, ํ† ํฐํ™” ๋„๊ตฌ์˜ ์˜ํ–ฅ์€ ์‚ฌ๊ฐ์ง€๋Œ€์— ๋†“์—ฌ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์ œ ๋ง‰ ํƒ๊ตฌ๋˜๊ธฐ ์‹œ์ž‘ํ•œ ์ด ๋ถ„์•ผ์— ์กฐ๋ช…์„ ๋น„์ถ”๊ธฐ ์œ„ํ•ด 24๊ฐœ์˜ ๋‹จ์ผ ์–ธ์–ด ๋ฐ ๋‹ค๊ตญ์–ด LLM์„ 2.6B ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ๋กœ ํ›ˆ๋ จํ•˜๊ณ  ๋‹ค์–‘ํ•œ ํ† ํฐํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์ œ๊ฑฐํ•จ์œผ๋กœ์จ ํ† ํฐํ™” ๋„๊ตฌ ์„ ํƒ์ด LLM ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, ํ† ํฐํ™” ๋„๊ตฌ ์„ ํƒ์ด ๋ชจ๋ธ์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ๊ณผ ํ›ˆ๋ จ ๋น„์šฉ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Œ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ผ๋ฐ˜์ ์ธ ํ† ํฐํ™” ํ‰๊ฐ€ ์ง€ํ‘œ์ธ _๋‹ค์‚ฐ์„ฑ_๊ณผ _ํŒจ๋ฆฌํ‹ฐ_๊ฐ€ ํ•ญ์ƒ ๋ชจ๋ธ์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ด๋Ÿฌํ•œ ์ง€ํ‘œ๊ฐ€ ๋ชจ๋ธ์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ์˜์‹ฌ์Šค๋Ÿฌ์šด ๋Œ€๋ฆฌ ์ง€ํ‘œ์ž„์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๊ฐ€์žฅ ๋นˆ๋ฒˆํ•œ 5๊ฐœ์˜ ์œ ๋Ÿฝ ์–ธ์–ด๋กœ ํ›ˆ๋ จ๋œ ๋‹ค๊ตญ์–ด ํ† ํฐํ™” ๋„๊ตฌ๋Š” ์˜์–ด์— ๋น„ํ•ด ์–ดํœ˜ ํฌ๊ธฐ๊ฐ€ 3๋ฐฐ ์ด์ƒ ์ฆ๊ฐ€ํ•ด์•ผ ํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ณผ๊ฑฐ์—๋Š” ์˜์–ด ์ค‘์‹ฌ์˜ ํ† ํฐํ™” ๋„๊ตฌ๊ฐ€ ๋‹ค๊ตญ์–ด LLM ํ›ˆ๋ จ์— ์ ์šฉ๋˜์—ˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํ† ํฐํ™” ์–ดํœ˜๊ฐ€ ๋น„ํšจ์œจ์ ์ด๊ธฐ ๋•Œ๋ฌธ์— ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์ด ์‹ฌ๊ฐํ•˜๊ฒŒ ์ €ํ•˜๋˜๊ณ  ์ตœ๋Œ€ 68%์˜ ์ถ”๊ฐ€ ํ›ˆ๋ จ ๋น„์šฉ์ด ๋ฐœ์ƒํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
**Instruction Tuning Perplexity**. The evaluation mentioned above metrics presents an inherent trade-off. Though accuracy on downstream tasks is typically viewed as a more concrete representation of a language model's real-world value, its variance tends to be higher due to the limited number of examples in these tasks and the step-wise behavior of accuracy as a metric. In contrast, perplexity, as a metric, is smoother while still exhibiting a strong correlation with performance [43]. Therefore as a middle ground between the two evaluation metrics, we propose evaluating the perplexity on a sample drawn from the instruction-tuning dataset used for fine-tuning OPT-IML [21]. This dataset spans over 1500 unique NLP tasks and comprises a wide array of prompt-answer pairs and therefore is representative of the _average_ NLP task. It has been carefully crafted by merging extensive task collections such as Super-NaturalInstructions [53] and PromptSource [3]. We refer the reader to Table 2.1 in [21] for a comprehensive breakdown. This approach allows us to balance practical performance measures and statistical consistency in evaluation. We note that this metric can simply be considered as perplexity on another validation set, where the validation set is filled with examples used for instruction-tuning (we are **not** fine-tuning on this dataset).
**Instruction Tuning Perplexity**. ์œ„์—์„œ ์–ธ๊ธ‰ํ•œ ํ‰๊ฐ€๋Š” ๋ณธ์งˆ์ ์ธ ์žฅ๋‹จ์ ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์˜ ์ •ํ™•๋„๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์˜ ์‹ค์ œ ๊ฐ€์น˜๋ฅผ ๋ณด๋‹ค ๊ตฌ์ฒด์ ์œผ๋กœ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋˜์ง€๋งŒ, ์ด๋Ÿฌํ•œ ์ž‘์—…์˜ ์ œํ•œ๋œ ์˜ˆ์ œ ์ˆ˜์™€ ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ์„œ ์ •ํ™•๋„์˜ ๋‹จ๊ณ„์  ๋™์ž‘์œผ๋กœ ์ธํ•ด ๊ทธ ํŽธ์ฐจ๊ฐ€ ๋” ๋†’์€ ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, perplexity๋Š” ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ์„œ ๋” ๋ถ€๋“œ๋Ÿฌ์šด ๋™์‹œ์— ์„ฑ๋Šฅ๊ณผ ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค[43]. ๋”ฐ๋ผ์„œ ๋‘ ๊ฐ€์ง€ ํ‰๊ฐ€ ์ง€ํ‘œ์˜ ์ค‘๊ฐ„ ์ง€์ ์œผ๋กœ์„œ, ์šฐ๋ฆฌ๋Š” OPT-IML [21]์„ Fine-tuningํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ Instruction Tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ถ”์ถœํ•œ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ perplexity๋ฅผ ํ‰๊ฐ€ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 1500๊ฐœ ์ด์ƒ์˜ ๊ณ ์œ ํ•œ NLP ์ž‘์—…์— ๊ฑธ์ณ ์žˆ์œผ๋ฉฐ ๋‹ค์–‘ํ•œ ํ”„๋กฌํ”„ํŠธ-๋‹ต๋ณ€ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ _ํ‰๊ท _ NLP ์ž‘์—…์„ ๋Œ€ํ‘œํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” Super-NaturalInstructions [53] ๋ฐ PromptSource [3]๊ณผ ๊ฐ™์€ ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—… ์ปฌ๋ ‰์…˜์„ ๋ณ‘ํ•ฉํ•˜์—ฌ ์‹ ์ค‘ํ•˜๊ฒŒ ์ œ์ž‘๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ [21]์˜ ํ‘œ 2.1์„ ์ฐธ์กฐํ•˜์„ธ์š”. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ์‹ค์šฉ์ ์ธ ์„ฑ๋Šฅ ์ธก์ •๊ณผ ํ‰๊ฐ€์˜ ํ†ต๊ณ„์  ์ผ๊ด€์„ฑ์„ ๊ท ํ˜• ์žˆ๊ฒŒ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฉ”ํŠธ๋ฆญ์€ ๋‹จ์ˆœํžˆ ๋‹ค๋ฅธ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์— ๋Œ€ํ•œ perplexity๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์€ Instruction Tuning์— ์‚ฌ์šฉ๋œ ์˜ˆ์ œ๋กœ ์ฑ„์›Œ์ง‘๋‹ˆ๋‹ค(์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด **Fine-tuning**ํ•˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค).
**์ง€์นจ ํŠœ๋‹ ๋‹นํ˜น๊ฐ**. ์œ„์—์„œ ์–ธ๊ธ‰ํ•œ ํ‰๊ฐ€๋Š” ๋ณธ์งˆ์ ์ธ ์žฅ๋‹จ์ ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์˜ ์ •ํ™•๋„๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์˜ ์‹ค์ œ ๊ฐ€์น˜๋ฅผ ๋ณด๋‹ค ๊ตฌ์ฒด์ ์œผ๋กœ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋˜์ง€๋งŒ, ์ด๋Ÿฌํ•œ ์ž‘์—…์˜ ์ œํ•œ๋œ ์˜ˆ์ œ ์ˆ˜์™€ ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ์„œ ์ •ํ™•๋„์˜ ๋‹จ๊ณ„์  ๋™์ž‘์œผ๋กœ ์ธํ•ด ๊ทธ ํŽธ์ฐจ๊ฐ€ ๋” ๋†’์€ ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ๋ณต์žก๋„๋Š” ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ์„œ ๋” ๋ถ€๋“œ๋Ÿฌ์šด ๋™์‹œ์— ์„ฑ๋Šฅ๊ณผ ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค[43]. ๋”ฐ๋ผ์„œ ๋‘ ๊ฐ€์ง€ ํ‰๊ฐ€ ์ง€ํ‘œ์˜ ์ค‘๊ฐ„ ์ง€์ ์œผ๋กœ์„œ, ์šฐ๋ฆฌ๋Š” OPT-IML [21]์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ ๋ช…๋ น์–ด ์กฐ์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ถ”์ถœํ•œ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ๋‹นํ˜น๊ฐ์„ ํ‰๊ฐ€ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 1500๊ฐœ ์ด์ƒ์˜ ๊ณ ์œ ํ•œ NLP ์ž‘์—…์— ๊ฑธ์ณ ์žˆ์œผ๋ฉฐ ๋‹ค์–‘ํ•œ ํ”„๋กฌํ”„ํŠธ-๋‹ต๋ณ€ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ _ํ‰๊ท _ NLP ์ž‘์—…์„ ๋Œ€ํ‘œํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” Super-NaturalInstructions [53] ๋ฐ PromptSource [3]๊ณผ ๊ฐ™์€ ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—… ์ปฌ๋ ‰์…˜์„ ๋ณ‘ํ•ฉํ•˜์—ฌ ์‹ ์ค‘ํ•˜๊ฒŒ ์ œ์ž‘๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ [21]์˜ ํ‘œ 2.1์„ ์ฐธ์กฐํ•˜์„ธ์š”. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ์‹ค์šฉ์ ์ธ ์„ฑ๋Šฅ ์ธก์ •๊ณผ ํ‰๊ฐ€์˜ ํ†ต๊ณ„์  ์ผ๊ด€์„ฑ์„ ๊ท ํ˜• ์žˆ๊ฒŒ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฉ”ํŠธ๋ฆญ์€ ๋‹จ์ˆœํžˆ ๋‹ค๋ฅธ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์— ๋Œ€ํ•œ ๋‹นํ˜น๊ฐ์œผ๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์€ ๋ช…๋ น์–ด ํŠœ๋‹์— ์‚ฌ์šฉ๋œ ์˜ˆ์ œ๋กœ ์ฑ„์›Œ์ง‘๋‹ˆ๋‹ค(์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด **๋ฏธ์„ธ ์กฐ์ •**ํ•˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค).
Large language model (LLM) pre-training has been largely democratized and open-sourced, allowing various academic labs, and industries to pre-train custom LLMs. Yet, a key differentiator between these models is the composition and size of the data used to train them. Data curation strategies are required to filter out scrapes of the web that are unstructured and/or poorly phrased (Eisenstein, 2013). While some of these strategies have been made public (Brown et al., 2020; Wenzek et al., 2020; Penedo et al., 2023), most state-of-the-art data curation techniques are unknown to the research community, and only anecdotal evidence remains. Research on data curation requires multiple rounds of re-training, making it an expensive endeavour to document techniques that lead to practical improvements. On the other hand, scaling laws for language models (such as Chinchilla scaling laws (Hoffmann et al., 2022)) show that with increasing model sizes, we should also increase both the training compute and data size linearly. This is infeasible because (a) high-quality data is limited (Villalobos et al., 2022), and repeating for even a small number of epochs (4 or more) results in diminishing returns or overfitting (Muennighoff et al., 2023; Touvron et al., 2023; Xue et al., 2023); and (b) pre-training for such long durations is prohibitively expensive.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM) ์‚ฌ์ „ ํ•™์Šต์€ ํฌ๊ฒŒ ๋ฏผ์ฃผํ™”๋˜๊ณ  ์˜คํ”ˆ ์†Œ์Šคํ™”๋˜์–ด ๋‹ค์–‘ํ•œ ํ•™์ˆ  ์—ฐ๊ตฌ์†Œ์™€ ์‚ฐ์—…๊ณ„์—์„œ ๋งž์ถคํ˜• LLM์„ ์‚ฌ์ „ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ ๊ฐ„์˜ ์ฃผ์š” ์ฐจ๋ณ„ํ™” ์š”์†Œ๋Š” ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ์˜ ๊ตฌ์„ฑ๊ณผ ํฌ๊ธฐ์ž…๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ์ „๋žต์€ ๋น„์ •ํ˜•์ ์ด๊ฑฐ๋‚˜ ์ž˜๋ชป ํ‘œํ˜„๋œ ์›น ์Šคํฌ๋žฉ์„ ๊ฑธ๋Ÿฌ๋‚ด๊ธฐ ์œ„ํ•ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค(Eisenstein, 2013). ์ด๋Ÿฌํ•œ ์ „๋žต ์ค‘ ์ผ๋ถ€๋Š” ๊ณต๊ฐœ๋˜์—ˆ์ง€๋งŒ(Brown et al., 2020; Wenzek et al., 2020; Penedo et al., 2023), ๋Œ€๋ถ€๋ถ„์˜ ์ตœ์‹  ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ๊ธฐ์ˆ ์€ ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ์•Œ๋ ค์ง€์ง€ ์•Š์•˜์œผ๋ฉฐ ์ผํ™”์ ์ธ ์ฆ๊ฑฐ๋งŒ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜์— ๋Œ€ํ•œ ์—ฐ๊ตฌ์—๋Š” ์—ฌ๋Ÿฌ ๋ฒˆ์˜ ์žฌํ•™์Šต์ด ํ•„์š”ํ•˜๋ฏ€๋กœ ์‹ค์งˆ์ ์ธ ๊ฐœ์„ ์„ ๊ฐ€์ ธ์˜ค๋Š” ๊ธฐ์ˆ ์„ ๋ฌธ์„œํ™”ํ•˜๋Š” ๋ฐ ๋งŽ์€ ๋น„์šฉ์ด ์†Œ์š”๋ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์–ธ์–ด ๋ชจ๋ธ์˜ Scaling laws(์˜ˆ: Chinchilla Scaling laws(Hoffmann et al., 2022))์— ๋”ฐ๋ฅด๋ฉด ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ์ปค์ง€๋ฉด ํ•™์Šต ์ปดํ“จํŒ…๊ณผ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๋„ ์„ ํ˜•์ ์œผ๋กœ ๋Š˜๋ ค์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ (a) ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋Š” ์ œํ•œ์ ์ด๋ฉฐ(Villalobos et al., 2022), ์ ์€ ์ˆ˜์˜ ์—ํฌํฌ(4 ์ด์ƒ)๋ฅผ ๋ฐ˜๋ณตํ•˜๋ฉด ์ˆ˜์ต ๊ฐ์†Œ ๋˜๋Š” ๊ณผ์ ํ•ฉ์ด ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ์—(Muennighoff et al., 2023; Touvron et al., 2023; Xue et al., 2023) ์ด๋Š” ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค; ๊ทธ๋ฆฌ๊ณ  (b) ๊ทธ๋ ‡๊ฒŒ ์˜ค๋žœ ๊ธฐ๊ฐ„ ๋™์•ˆ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์€ ๋น„์šฉ์ด ๋„ˆ๋ฌด ๋งŽ์ด ๋“ญ๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM) ์‚ฌ์ „ ํ•™์Šต์€ ํฌ๊ฒŒ ๋ฏผ์ฃผํ™”๋˜๊ณ  ์˜คํ”ˆ ์†Œ์Šคํ™”๋˜์–ด ๋‹ค์–‘ํ•œ ํ•™์ˆ  ์—ฐ๊ตฌ์†Œ์™€ ์‚ฐ์—…๊ณ„์—์„œ ๋งž์ถคํ˜• LLM์„ ์‚ฌ์ „ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ ๊ฐ„์˜ ์ฃผ์š” ์ฐจ๋ณ„ํ™” ์š”์†Œ๋Š” ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ์˜ ๊ตฌ์„ฑ๊ณผ ํฌ๊ธฐ์ž…๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ์ „๋žต์€ ๋น„์ •ํ˜•์ ์ด๊ฑฐ๋‚˜ ์ž˜๋ชป ํ‘œํ˜„๋œ ์›น ์Šคํฌ๋žฉ์„ ๊ฑธ๋Ÿฌ๋‚ด๊ธฐ ์œ„ํ•ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค(Eisenstein, 2013). ์ด๋Ÿฌํ•œ ์ „๋žต ์ค‘ ์ผ๋ถ€๋Š” ๊ณต๊ฐœ๋˜์—ˆ์ง€๋งŒ(Brown et al., 2020; Wenzek et al., 2020; Penedo et al., 2023), ๋Œ€๋ถ€๋ถ„์˜ ์ตœ์‹  ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ๊ธฐ์ˆ ์€ ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ์•Œ๋ ค์ง€์ง€ ์•Š์•˜์œผ๋ฉฐ ์ผํ™”์ ์ธ ์ฆ๊ฑฐ๋งŒ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜์— ๋Œ€ํ•œ ์—ฐ๊ตฌ์—๋Š” ์—ฌ๋Ÿฌ ๋ฒˆ์˜ ์žฌํ•™์Šต์ด ํ•„์š”ํ•˜๋ฏ€๋กœ ์‹ค์งˆ์ ์ธ ๊ฐœ์„ ์„ ๊ฐ€์ ธ์˜ค๋Š” ๊ธฐ์ˆ ์„ ๋ฌธ์„œํ™”ํ•˜๋Š” ๋ฐ ๋งŽ์€ ๋น„์šฉ์ด ์†Œ์š”๋ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์–ธ์–ด ๋ชจ๋ธ์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™(์˜ˆ: ์นœ์น ๋ผ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™(Hoffmann et al., 2022))์— ๋”ฐ๋ฅด๋ฉด ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ์ปค์ง€๋ฉด ํ•™์Šต ์ปดํ“จํŒ…๊ณผ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๋„ ์„ ํ˜•์ ์œผ๋กœ ๋Š˜๋ ค์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ (a) ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋Š” ์ œํ•œ์ ์ด๋ฉฐ(Villalobos et al., 2022), ์ ์€ ์ˆ˜์˜ ์—ํฌํฌ(4 ์ด์ƒ)๋ฅผ ๋ฐ˜๋ณตํ•˜๋ฉด ์ˆ˜์ต ๊ฐ์†Œ ๋˜๋Š” ๊ณผ์ ํ•ฉ์ด ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ์—(Muennighoff et al., 2023; Touvron et al., 2023; Xue et al., 2023) ์ด๋Š” ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค; ๊ทธ๋ฆฌ๊ณ  (b) ๊ทธ๋ ‡๊ฒŒ ์˜ค๋žœ ๊ธฐ๊ฐ„ ๋™์•ˆ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์€ ๋น„์šฉ์ด ๋„ˆ๋ฌด ๋งŽ์ด ๋“ญ๋‹ˆ๋‹ค.
Learned FiltersWe use learned filters to address nuanced cases that exceed the capabilities of standard heuristic rules. Notably, the Chinese content extracted from Common Crawl present unique challenges, particularly with a higher ratio of inappropriate content like pornography and gambling. Traditional heuristic-rule-based filters struggle to effectively identify and eliminate all harmful content. To enhance our filtering process, we have integrated a suite of learned scorers for filtering, namely the perplexity scorer, quality scorer, safety scorer, and document coherence scorer: (1). the _Perplexity Scorer_, utilizing the KenLM library as per CCNet [80], evaluates a vast array of web documents, discarding those with perplexity scores largely above average; (2). the _Quality Scorer_ is a classifier trained to recognize and favor pages similar to Wikipedia in quality and assign scores accordingly. Documents that fail to meet the quality standard are subsequently removed; (3). the _Document Coherence Scorer_ identifies low-quality web documents that consist of disparate sentences or paragraphs, thus being incoherence. Such documents are either segmented for further analysis or removed entirely. (4). the _Safety Scorer_ identifies and removes web documents containing toxic content, such as violence, pornography, and political propaganda.
ํ•™์Šต๋œ ํ•„ํ„ฐํ•™์Šต๋œ ํ•„ํ„ฐ๋Š” ํ‘œ์ค€ heuristic ๊ทœ์น™์˜ ๊ธฐ๋Šฅ์„ ์ดˆ๊ณผํ•˜๋Š” ๋ฏธ๋ฌ˜ํ•œ ์‚ฌ๋ก€๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, Common Crawl์—์„œ ์ถ”์ถœํ•œ ์ค‘๊ตญ์–ด ์ฝ˜ํ…์ธ ๋Š” ์Œ๋ž€๋ฌผ์ด๋‚˜ ๋„๋ฐ•๊ณผ ๊ฐ™์€ ๋ถ€์ ์ ˆํ•œ ์ฝ˜ํ…์ธ ์˜ ๋น„์œจ์ด ๋†’์•„ ๋…ํŠนํ•œ ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด์˜ heuristic ๊ทœ์น™ ๊ธฐ๋ฐ˜ ํ•„ํ„ฐ๋Š” ๋ชจ๋“  ์œ ํ•ดํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ์ œ๊ฑฐํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ํ•„ํ„ฐ๋ง ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ํ•„ํ„ฐ๋ง์„ ์œ„ํ•œ ํ•™์Šต ์Šค์ฝ”์–ด๋ง ์ œํ’ˆ๊ตฐ, ์ฆ‰ Perplexity Scorer, Quality Scorer, Safety Scorer, Document Coherence Scorer๋ฅผ ํ†ตํ•ฉํ–ˆ์Šต๋‹ˆ๋‹ค: (1). KenLM ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ CCNet [80]์— ๋”ฐ๋ผ ํ™œ์šฉํ•˜๋Š” _Perplexity Scorer_๋Š” ๋ฐฉ๋Œ€ํ•œ ์›น ๋ฌธ์„œ๋ฅผ ํ‰๊ฐ€ํ•˜์—ฌ Perplexity ์ ์ˆ˜๊ฐ€ ํฌ๊ฒŒ ํ‰๊ท  ์ด์ƒ์ธ ๋ฌธ์„œ๋Š” ๋ฒ„๋ฆฝ๋‹ˆ๋‹ค; (2). _Quality Scorer_๋Š” ํ’ˆ์งˆ ์ธก๋ฉด์—์„œ ์œ„ํ‚คํ”ผ๋””์•„์™€ ์œ ์‚ฌํ•œ ํŽ˜์ด์ง€๋ฅผ ์ธ์‹ํ•˜๊ณ  ์ด์— ๋”ฐ๋ผ ์ ์ˆ˜๋ฅผ ๋ถ€์—ฌํ•˜๋„๋ก ํ›ˆ๋ จ๋œ ๋ถ„๋ฅ˜๊ธฐ์ž…๋‹ˆ๋‹ค. ํ’ˆ์งˆ ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฌธ์„œ๋Š” ์ดํ›„ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค; (3). _Document Coherence Scorer_๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ๋ฌธ์žฅ์ด๋‚˜ ๋‹จ๋ฝ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์ผ๊ด€์„ฑ์ด ์—†๋Š” ํ’ˆ์งˆ์ด ๋‚ฎ์€ ์›น ๋ฌธ์„œ๋ฅผ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์„œ๋Š” ์ถ”๊ฐ€ ๋ถ„์„์„ ์œ„ํ•ด ์„ธ๋ถ„ํ™”๋˜๊ฑฐ๋‚˜ ์™„์ „ํžˆ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. (4). _Safety Scorer_๋Š” ํญ๋ ฅ, ์Œ๋ž€๋ฌผ, ์ •์น˜์  ์„ ์ „ ๋“ฑ ์œ ๋…ํ•œ ์ฝ˜ํ…์ธ ๊ฐ€ ํฌํ•จ๋œ ์›น ๋ฌธ์„œ๋ฅผ ์‹๋ณ„ํ•˜์—ฌ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
ํ•™์Šต๋œ ํ•„ํ„ฐํ•™์Šต๋œ ํ•„ํ„ฐ๋Š” ํ‘œ์ค€ ํœด๋ฆฌ์Šคํ‹ฑ ๊ทœ์น™์˜ ๊ธฐ๋Šฅ์„ ์ดˆ๊ณผํ•˜๋Š” ๋ฏธ๋ฌ˜ํ•œ ์‚ฌ๋ก€๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ปค๋จผ ํฌ๋กค๋ง์—์„œ ์ถ”์ถœํ•œ ์ค‘๊ตญ์–ด ์ฝ˜ํ…์ธ ๋Š” ์Œ๋ž€๋ฌผ์ด๋‚˜ ๋„๋ฐ•๊ณผ ๊ฐ™์€ ๋ถ€์ ์ ˆํ•œ ์ฝ˜ํ…์ธ ์˜ ๋น„์œจ์ด ๋†’์•„ ๋…ํŠนํ•œ ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด์˜ ํœด๋ฆฌ์Šคํ‹ฑ ๊ทœ์น™ ๊ธฐ๋ฐ˜ ํ•„ํ„ฐ๋Š” ๋ชจ๋“  ์œ ํ•ดํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ์ œ๊ฑฐํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ํ•„ํ„ฐ๋ง ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ํ•„ํ„ฐ๋ง์„ ์œ„ํ•œ ํ•™์Šต ์Šค์ฝ”์–ด๋ง ์ œํ’ˆ๊ตฐ, ์ฆ‰ ๋‹นํ˜น๊ฐ ์Šค์ฝ”์–ด๋ง, ํ’ˆ์งˆ ์Šค์ฝ”์–ด๋ง, ์•ˆ์ „์„ฑ ์Šค์ฝ”์–ด๋ง, ๋ฌธ์„œ ์ผ๊ด€์„ฑ ์Šค์ฝ”์–ด๋ง์„ ํ†ตํ•ฉํ–ˆ์Šต๋‹ˆ๋‹ค: (1). KenLM ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ CCNet [80]์— ๋”ฐ๋ผ ํ™œ์šฉํ•˜๋Š” _Perplexity Scorer_๋Š” ๋ฐฉ๋Œ€ํ•œ ์›น ๋ฌธ์„œ๋ฅผ ํ‰๊ฐ€ํ•˜์—ฌ ๋‹นํ˜น๊ฐ ์ ์ˆ˜๊ฐ€ ํฌ๊ฒŒ ํ‰๊ท  ์ด์ƒ์ธ ๋ฌธ์„œ๋Š” ๋ฒ„๋ฆฝ๋‹ˆ๋‹ค; (2). ํ’ˆ์งˆ ์Šค์ฝ”์–ด๋ง๊ธฐ๋Š” ํ’ˆ์งˆ ์ธก๋ฉด์—์„œ ์œ„ํ‚คํ”ผ๋””์•„์™€ ์œ ์‚ฌํ•œ ํŽ˜์ด์ง€๋ฅผ ์ธ์‹ํ•˜๊ณ  ์ด์— ๋”ฐ๋ผ ์ ์ˆ˜๋ฅผ ๋ถ€์—ฌํ•˜๋„๋ก ํ›ˆ๋ จ๋œ ๋ถ„๋ฅ˜๊ธฐ์ž…๋‹ˆ๋‹ค. ํ’ˆ์งˆ ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฌธ์„œ๋Š” ์ดํ›„ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค; (3). ๋ฌธ์„œ ์ผ๊ด€์„ฑ ์ ์ˆ˜์ž_๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ๋ฌธ์žฅ์ด๋‚˜ ๋‹จ๋ฝ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์ผ๊ด€์„ฑ์ด ์—†๋Š” ํ’ˆ์งˆ์ด ๋‚ฎ์€ ์›น ๋ฌธ์„œ๋ฅผ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์„œ๋Š” ์ถ”๊ฐ€ ๋ถ„์„์„ ์œ„ํ•ด ์„ธ๋ถ„ํ™”๋˜๊ฑฐ๋‚˜ ์™„์ „ํžˆ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. (4). ์•ˆ์ „์„ฑ ์ ์ˆ˜์ž_๋Š” ํญ๋ ฅ, ์Œ๋ž€๋ฌผ, ์ •์น˜์  ์„ ์ „ ๋“ฑ ์œ ๋…ํ•œ ์ฝ˜ํ…์ธ ๊ฐ€ ํฌํ•จ๋œ ์›น ๋ฌธ์„œ๋ฅผ ์‹๋ณ„ํ•˜์—ฌ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
**Handcrafting Method**: in this method, agent profiles are manually specified. For instance, if one would like to design agents with different personalities, he can use "you are an outgoing person" or "you are an introverted person" to profile the agent. The handcrafting method has been leveraged in a lot of previous work to indicate the agent profiles. For example, Generative Agent [22] describes the agent by the information like name, objectives, and relationships with other agents. MetaGPT [23], ChatDev [18], and Self-collaboration [24] predefine various roles and their corresponding responsibilities in software development, manually assigning distinct profiles to each agent to facilitate collaboration. PTLLM [25] aims to explore and quantify personality traits displayed in texts generated by LLMs. This method guides LLMs in generating diverse responses by manfully defining various agent characters through the use of personality assessment tools such as IPIP-NEO [26] and BFI [27]. [28] studies the toxicity of the LLM output by manually prompting LLMs with different roles, such as politicians, journalists and businesspersons. In general, the handcrafting method is very flexible, since one can assign any profile information to the agents. However, it can be also labor-intensive, particularly when dealing with a large number of agents.
**Handcrafting ๋ฐฉ๋ฒ•**: ์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” agent ํ”„๋กœํ•„์„ ์ˆ˜๋™์œผ๋กœ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋‹ค์–‘ํ•œ ์„ฑ๊ฒฉ์„ ๊ฐ€์ง„ agent๋ฅผ ์„ค๊ณ„ํ•˜๊ณ ์ž ํ•˜๋Š” ๊ฒฝ์šฐ '๋‹น์‹ ์€ ์™ธํ–ฅ์ ์ธ ์‚ฌ๋žŒ์ž…๋‹ˆ๋‹ค' ๋˜๋Š” '๋‹น์‹ ์€ ๋‚ด์„ฑ์ ์ธ ์‚ฌ๋žŒ์ž…๋‹ˆ๋‹ค'๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ agent์˜ ํ”„๋กœํ•„์„ ์ง€์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Handcrafting ๋ฐฉ๋ฒ•์€ ์ด์ „์˜ ๋งŽ์€ ์ž‘์—…์—์„œ agent ํ”„๋กœํ•„์„ ๋‚˜ํƒ€๋‚ด๋Š” ๋ฐ ํ™œ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Generative Agent[22]๋Š” ์ด๋ฆ„, ๋ชฉํ‘œ, ๋‹ค๋ฅธ agent์™€์˜ ๊ด€๊ณ„ ๋“ฑ์˜ ์ •๋ณด๋กœ agent๋ฅผ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. MetaGPT[23], ChatDev[18], Self-collaboration[24]์€ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ์—์„œ ๋‹ค์–‘ํ•œ ์—ญํ• ๊ณผ ๊ทธ์— ์ƒ์‘ํ•˜๋Š” ์ฑ…์ž„์„ ์‚ฌ์ „ ์ •์˜ํ•˜๊ณ  ๊ฐ agent์— ๊ณ ์œ ํ•œ ํ”„๋กœํ•„์„ ์ˆ˜๋™์œผ๋กœ ํ• ๋‹นํ•˜์—ฌ ํ˜‘์—…์„ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค. PTLLM[25]์€ LLM์ด ์ƒ์„ฑํ•œ ํ…์ŠคํŠธ์— ํ‘œ์‹œ๋œ ์„ฑ๊ฒฉ ํŠน์„ฑ์„ ํƒ์ƒ‰ํ•˜๊ณ  ์ •๋Ÿ‰ํ™”ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ IPIP-NEO[26] ๋ฐ BFI[27]์™€ ๊ฐ™์€ ์„ฑ๊ฒฉ ํ‰๊ฐ€ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ agent ์บ๋ฆญํ„ฐ๋ฅผ ์ˆ˜๋™์œผ๋กœ ์ •์˜ํ•จ์œผ๋กœ์จ LLM์ด ๋‹ค์–‘ํ•œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. [28]์€ ์ •์น˜์ธ, ์–ธ๋ก ์ธ, ์‚ฌ์—…๊ฐ€ ๋“ฑ ๋‹ค์–‘ํ•œ ์—ญํ• ์„ ๊ฐ€์ง„ LLM์— ์ˆ˜๋™์œผ๋กœ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ œ๊ณตํ•˜์—ฌ LLM ์ถœ๋ ฅ์˜ ๋…์„ฑ์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ Handcrafting ๋ฐฉ์‹์€ agent์— ๋ชจ๋“  ํ”„๋กœํ•„ ์ •๋ณด๋ฅผ ํ• ๋‹นํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋งค์šฐ ์œ ์—ฐํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํŠนํžˆ ๋งŽ์€ ์ˆ˜์˜ agent๋ฅผ ๋‹ค๋ฃฐ ๋•Œ๋Š” ๋…ธ๋™ ์ง‘์•ฝ์ ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**ํ•ธ๋“œํฌ๋ž˜ํ”„ํŒ… ๋ฐฉ๋ฒ•**: ์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” ์—์ด์ „ํŠธ ํ”„๋กœํ•„์„ ์ˆ˜๋™์œผ๋กœ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋‹ค์–‘ํ•œ ์„ฑ๊ฒฉ์„ ๊ฐ€์ง„ ์—์ด์ „ํŠธ๋ฅผ ์„ค๊ณ„ํ•˜๊ณ ์ž ํ•˜๋Š” ๊ฒฝ์šฐ '๋‹น์‹ ์€ ์™ธํ–ฅ์ ์ธ ์‚ฌ๋žŒ์ž…๋‹ˆ๋‹ค' ๋˜๋Š” '๋‹น์‹ ์€ ๋‚ด์„ฑ์ ์ธ ์‚ฌ๋žŒ์ž…๋‹ˆ๋‹ค'๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์—์ด์ „ํŠธ์˜ ํ”„๋กœํ•„์„ ์ง€์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•ธ๋“œํฌ๋ž˜ํ”„ํŒ… ๋ฐฉ๋ฒ•์€ ์ด์ „์˜ ๋งŽ์€ ์ž‘์—…์—์„œ ์—์ด์ „ํŠธ ํ”„๋กœํ•„์„ ๋‚˜ํƒ€๋‚ด๋Š” ๋ฐ ํ™œ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ƒ์„ฑ ์—์ด์ „ํŠธ[22]๋Š” ์ด๋ฆ„, ๋ชฉํ‘œ, ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์™€์˜ ๊ด€๊ณ„ ๋“ฑ์˜ ์ •๋ณด๋กœ ์—์ด์ „ํŠธ๋ฅผ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๋ฉ”ํƒ€GPT [23], ChatDev [18], ์…€ํ”„ ํ˜‘์—… [24]์€ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ์—์„œ ๋‹ค์–‘ํ•œ ์—ญํ• ๊ณผ ๊ทธ์— ์ƒ์‘ํ•˜๋Š” ์ฑ…์ž„์„ ์‚ฌ์ „ ์ •์˜ํ•˜๊ณ  ๊ฐ ์—์ด์ „ํŠธ์— ๊ณ ์œ ํ•œ ํ”„๋กœํ•„์„ ์ˆ˜๋™์œผ๋กœ ํ• ๋‹นํ•˜์—ฌ ํ˜‘์—…์„ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค. PTLLM [25]์€ LLM์ด ์ƒ์„ฑํ•œ ํ…์ŠคํŠธ์— ํ‘œ์‹œ๋œ ์„ฑ๊ฒฉ ํŠน์„ฑ์„ ํƒ์ƒ‰ํ•˜๊ณ  ์ •๋Ÿ‰ํ™”ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ IPIP-NEO [26] ๋ฐ BFI [27]์™€ ๊ฐ™์€ ์„ฑ๊ฒฉ ํ‰๊ฐ€ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์—์ด์ „ํŠธ ์บ๋ฆญํ„ฐ๋ฅผ ์ˆ˜๋™์œผ๋กœ ์ •์˜ํ•จ์œผ๋กœ์จ LLM์ด ๋‹ค์–‘ํ•œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. [28]์€ ์ •์น˜์ธ, ์–ธ๋ก ์ธ, ์‚ฌ์—…๊ฐ€ ๋“ฑ ๋‹ค์–‘ํ•œ ์—ญํ• ์„ ๊ฐ€์ง„ LLM์— ์ˆ˜๋™์œผ๋กœ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ œ๊ณตํ•˜์—ฌ LLM ์ถœ๋ ฅ์˜ ๋…์„ฑ์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ํ•ธ๋“œํฌ๋ž˜ํ”„ํŒ… ๋ฐฉ์‹์€ ์—์ด์ „ํŠธ์— ๋ชจ๋“  ํ”„๋กœํ•„ ์ •๋ณด๋ฅผ ํ• ๋‹นํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋งค์šฐ ์œ ์—ฐํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํŠนํžˆ ๋งŽ์€ ์ˆ˜์˜ ์—์ด์ „ํŠธ๋ฅผ ๋‹ค๋ฃฐ ๋•Œ๋Š” ๋…ธ๋™ ์ง‘์•ฝ์ ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
All authors contribute substantially to the MiniCPM project. Shengding Hu lead and participated in all aspects of the projects. This included the scaling experiments (conducted alongside Yuge Tu), babysitting the training of MiniCPM base models, and contributing to various other parts of the research. Shengding Hu wrote the paper. Chaoqun He was responsible for evaluating MiniCPM, while Ganqu Cui handled the RLHF training. Xiang Long, Zhi Zheng, Xinrong Zhang and Shengding Hu extended the context window to 128K. The MoE research was conducted by Yewei Fang and Zhi Zheng. Weilin Zhao and Kaihuo Zhang contributed to the training and inference infrastructure. The open-sourcing of MiniCPM was prepared by Yuxiang Huang and Shengding Hu. Shengding Hu, along with Chenyang Zhao, also provided analysis on the WSD scheduler's training dynamics. Zheng Leng Thai developed the tokenizer. The development of MiniCPM-V was carried out by Chongyi Wang and Yuan Yao. The training corpus of MiniCPM was prepared by Jie Zhou, Jie Cai, Shengding Hu, Zhi Zheng, and Zhongwu Zhai. The paper was proofread by Xingrong Zhang and Chaoqun He. Insightful instructions on training MiniCPM were provided by Xu Han, Ning Ding, and Zhiyuan Liu. Finally, Zhiyuan Liu, Maosong Sun, Guoyang Zeng, Chao Jia, and Dahai Li offered essential resources for the training of MiniCPM.
๋ชจ๋“  ์ €์ž๋Š” MiniCPM ํ”„๋กœ์ ํŠธ์— ํฌ๊ฒŒ ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. Shengding Hu๋Š” ํ”„๋กœ์ ํŠธ์˜ ๋ชจ๋“  ์ธก๋ฉด์„ ์ฃผ๋„ํ•˜๊ณ  ์ฐธ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” scaling ์‹คํ—˜(Yuge Tu์™€ ํ•จ๊ป˜ ์ˆ˜ํ–‰), MiniCPM ๊ธฐ๋ณธ ๋ชจ๋ธ ํ›ˆ๋ จ ๋ณด์กฐ, ์—ฐ๊ตฌ์˜ ๋‹ค๋ฅธ ์—ฌ๋Ÿฌ ๋ถ€๋ถ„์— ๊ธฐ์—ฌํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. Shengding Hu๊ฐ€ ๋…ผ๋ฌธ์„ ์ž‘์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. Chaoqun He๋Š” MiniCPM์„ ํ‰๊ฐ€ํ•˜๋Š” ์—…๋ฌด๋ฅผ ๋‹ด๋‹นํ–ˆ๊ณ , Ganqu Cui๋Š” RLHF ํ›ˆ๋ จ์„ ๋‹ด๋‹นํ–ˆ์Šต๋‹ˆ๋‹ค. Xiang Long, Zhi Zheng, Xinrong Zhang, Shengding Hu๋Š” ์ปจํ…์ŠคํŠธ ์ฐฝ์„ 128K๋กœ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. MoE ์—ฐ๊ตฌ๋Š” Yewei Fang๊ณผ Zhi Zheng์ด ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. Weilin Zhao์™€ Kaihuo Zhang์€ ํ›ˆ๋ จ ๋ฐ ์ถ”๋ก  ์ธํ”„๋ผ์— ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. MiniCPM์˜ ์˜คํ”ˆ ์†Œ์‹ฑ์€ Yuxiang Huang๊ณผ Shengding Hu๊ฐ€ ์ค€๋น„ํ–ˆ์Šต๋‹ˆ๋‹ค. Shengding Hu์™€ Chenyang Zhao๋Š” WSD ์Šค์ผ€์ค„๋Ÿฌ์˜ ํ›ˆ๋ จ ๋™์—ญํ•™์— ๋Œ€ํ•œ ๋ถ„์„๋„ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. Zheng Leng Thai๊ฐ€ ํ† ํฐํ™” ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. MiniCPM-V์˜ ๊ฐœ๋ฐœ์€ Chongyi Wang๊ณผ Yuan Yao๊ฐ€ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. MiniCPM์˜ ํ›ˆ๋ จ ๋ง๋ญ‰์น˜๋Š” Jie Zhou, Jie Cai, Shengding Hu, Zhi Zheng, Zhongwu Zhai๊ฐ€ ์ค€๋น„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์€ Xingrong Zhang๊ณผ Chaoqun He๊ฐ€ ๊ต์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. MiniCPM ํ›ˆ๋ จ์— ๋Œ€ํ•œ ํ†ต์ฐฐ๋ ฅ ์žˆ๋Š” ์ง€์นจ์€ Xu Han, Ning Ding, Zhiyuan Liu๊ฐ€ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, Zhiyuan Liu, Maosong Sun, Guoyang Zeng, Chao Jia, Dahai Li๋Š” MiniCPM ํ›ˆ๋ จ์„ ์œ„ํ•œ ํ•„์ˆ˜ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค.
๋ชจ๋“  ์ €์ž๋Š” ๋ฏธ๋‹ˆCPM ํ”„๋กœ์ ํŠธ์— ํฌ๊ฒŒ ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์Š๋”ฉ ํ›„๋Š” ํ”„๋กœ์ ํŠธ์˜ ๋ชจ๋“  ์ธก๋ฉด์„ ์ฃผ๋„ํ•˜๊ณ  ์ฐธ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์Šค์ผ€์ผ๋ง ์‹คํ—˜(ํˆฌ๊ฒŒ ํˆฌ์™€ ํ•จ๊ป˜ ์ˆ˜ํ–‰), ๋ฏธ๋‹ˆCPM ๊ธฐ๋ณธ ๋ชจ๋ธ ๊ต์œก ๋ณด์กฐ, ์—ฐ๊ตฌ์˜ ๋‹ค๋ฅธ ์—ฌ๋Ÿฌ ๋ถ€๋ถ„์— ๊ธฐ์—ฌํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์Š๋”ฉ ํ›„๊ฐ€ ๋…ผ๋ฌธ์„ ์ž‘์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฐจ์˜ค์ฟค ํ—ˆ๋Š” ๋ฏธ๋‹ˆCPM์„ ํ‰๊ฐ€ํ•˜๋Š” ์—…๋ฌด๋ฅผ ๋‹ด๋‹นํ–ˆ๊ณ , ๊ฐ„์ฟค cui๋Š” RLHF ๊ต์œก์„ ๋‹ด๋‹นํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹œ์•™ ๋กฑ, ์ง€ ์ •, ์‹ ๋กฑ ์žฅ, ์Š๋”ฉ ํ›„๋Š” ์ปจํ…์ŠคํŠธ ์ฐฝ์„ 128K๋กœ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. MoE ์—ฐ๊ตฌ๋Š” ์˜ˆ์›จ์ด ํŒก๊ณผ ์ง€ ์ •์ด ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์›จ์ผ๋ฆฐ ์ž์˜ค์™€ ์นด์ดํ›„์˜ค ์žฅ์€ ํ›ˆ๋ จ ๋ฐ ์ถ”๋ก  ์ธํ”„๋ผ์— ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฏธ๋‹ˆCPM์˜ ์˜คํ”ˆ ์†Œ์‹ฑ์€ ์œ ์‹œ์•™ ํ™ฉ๊ณผ ์Š๋”ฉ ํ›„๊ฐ€ ์ค€๋น„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์Š๋”ฉ ํ›„์™€ ์ฒธ์–‘ ์ž์˜ค๋Š” WSD ์Šค์ผ€์ค„๋Ÿฌ์˜ ํ›ˆ๋ จ ๋™์—ญํ•™์— ๋Œ€ํ•œ ๋ถ„์„๋„ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ • ๋ ˆ์ž‰ ํƒ€์ด๊ฐ€ ํ† ํฐํ™” ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฏธ๋‹ˆCPM-V์˜ ๊ฐœ๋ฐœ์€ ์™•์ถฉ์ด์™€ ์•ผ์˜ค์œ„์•ˆ์ด ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฏธ๋‹ˆCPM์˜ ํ›ˆ๋ จ ๋ง๋ญ‰์น˜๋Š” ์ง€์— ์ฃผ, ์ง€์— ์ฐจ์ด, ์Š๋”ฉ ํ›„, ์ง€ ์ •, ์ค‘์šฐ ์ž์ด๊ฐ€ ์ค€๋น„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์€ ์žฅ์‹ฑ๋กฑ๊ณผ ์ฐจ์˜ค์ฟค ํ—ˆ๊ฐ€ ๊ต์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฏธ๋‹ˆCPM ํ›ˆ๋ จ์— ๋Œ€ํ•œ ํ†ต์ฐฐ๋ ฅ ์žˆ๋Š” ์ง€์นจ์€ ์„œํ•œ, ๋”ฉ๋‹, ๋ฆฌ์šฐ์ง€์œ„์•ˆ์ด ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๋ฆฌ์šฐ์ง€์œ„์•ˆ, ๋งˆ์˜ค์†ก ์„ , ์ €์šฐ๊ฐ€์˜ค์–‘, ์ฐจ์˜ค์ง€์•„, ๋ฆฌ๋‹คํ•˜์ด๋Š” ๋ฏธ๋‹ˆCPM ํ›ˆ๋ จ์„ ์œ„ํ•œ ํ•„์ˆ˜ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค.
The experimental results are presented in Table 1. We observed the following: (1) **chat vector enables the model to follow instruction:** as shown in table 1, \(\text{llama2}\rightarrow\text{CP}+\text{chat}\) vector have comparable result with \(\text{llama2}\rightarrow\text{CP}\rightarrow\text{FT}\) in both \(\text{Trandtional}\) Chinese LLaMA and Chinese LLaMA. This states the chat vector contains information about the instruction following, which the model can use to guide its output. (2) **FT and chat vector have a complementary effect:** FT and adding a chat vector together gives better performance than doing either one alone. (3) \(\text{llama2-chat}\rightarrow\text{CP}+\text{FT}\)** makes the chat ability lost:** although it outperform than the \(\text{llama2}\rightarrow\text{CP}\rightarrow\text{FT}\), it is worse than the \(\text{llama2}\) CP \(+\text{chat}\) vector or \(\text{llama2}\rightarrow\text{CP}\rightarrow\text{FT}+\text{chat}\) vector. This indicates that \(\text{llama2-chat}\rightarrow\text{CP}+\text{FT}\) will cause information loss, and also shows the importance of using the chat vector. To conclude, the chat vector augments model capabilities in following instructions, independent of the CP model's native language. Furthermore, introducing the chat vector subsequent to fine-tuning elevates generation prowess.
์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์‚ฌ์‹ค์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค: (1) **chat ๋ฒกํ„ฐ๋Š” ๋ชจ๋ธ์ด Instruction์„ ๋”ฐ๋ฅผ ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค:** ํ‘œ 1์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, \(\text{llama2}\rightarrow\text{CP}+\text{chat}\) ๋ฒกํ„ฐ๋Š” \(\text{llama2}\rightarrow\text{CP}\rightarrow\text{FT}\)์™€ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ์ด๋Š” chat ๋ฒกํ„ฐ์— Instruction์„ ๋”ฐ๋ฅด๋Š” ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋ชจ๋ธ์ด ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ถœ๋ ฅ์„ ์•ˆ๋‚ดํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. (2) **FT์™€ chat ๋ฒกํ„ฐ๋Š” ์ƒํ˜ธ ๋ณด์™„์ ์ธ ํšจ๊ณผ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค:** FT์™€ chat ๋ฒกํ„ฐ๋ฅผ ํ•จ๊ป˜ ์‚ฌ์šฉํ•˜๋ฉด ๋‘˜ ์ค‘ ํ•˜๋‚˜๋งŒ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. (3) \(\text{llama2-chat}\rightarrow\text{CP}+\text{FT}\)**๋Š” chat ๊ธฐ๋Šฅ์„ ์ƒ์‹ค์‹œํ‚ต๋‹ˆ๋‹ค:** \(\text{llama2}\rightarrow\text{CP}\rightarrow\text{FT}\)๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜์ง€๋งŒ, \(\text{llama2}\) CP \(+\text{chat}\) ๋ฒกํ„ฐ ๋˜๋Š” \(\text{llama2}\rightarrow\text{CP}\rightarrow\text{FT}+\text{chat}\) ๋ฒกํ„ฐ๋ณด๋‹ค ๋‚˜์ฉ๋‹ˆ๋‹ค. ์ด๋Š” \(\text{llama2-chat}\rightarrow\text{CP}+\text{FT}\)๊ฐ€ ์ •๋ณด ์†์‹ค์„ ์œ ๋ฐœํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, chat ๋ฒกํ„ฐ ์‚ฌ์šฉ์˜ ์ค‘์š”์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ฒฐ๋ก ์ ์œผ๋กœ, chat ๋ฒกํ„ฐ๋Š” CP ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ์–ธ์–ด์™€ ๊ด€๊ณ„์—†์ด Instruction์„ ๋”ฐ๋ฅด๋Š” ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋˜ํ•œ Fine-tuning ํ›„ chat ๋ฒกํ„ฐ๋ฅผ ๋„์ž…ํ•˜๋ฉด ์ƒ์„ฑ ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค.
์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์‚ฌ์‹ค์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค: (1) **์ฑ„ํŒ… ๋ฒกํ„ฐ๋Š” ๋ชจ๋ธ์ด ๋ช…๋ น์„ ๋”ฐ๋ฅผ ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค:** ํ‘œ 1์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, (\text{llama2}\rightarrow\text{CP}+\text{chat}\) ๋ฒกํ„ฐ๋Š” (\text{llama2}\rightarrow\text{CP}\rightarrow\text{FT}\)์™€ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ฑ„ํŒ… ๋ฒกํ„ฐ์— ๋ช…๋ น์–ด๋ฅผ ๋”ฐ๋ฅด๋Š” ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋ชจ๋ธ์ด ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ถœ๋ ฅ์„ ์•ˆ๋‚ดํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. (2) **FT์™€ ์ฑ„ํŒ… ๋ฒกํ„ฐ๋Š” ์ƒํ˜ธ ๋ณด์™„์ ์ธ ํšจ๊ณผ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค:** FT์™€ ์ฑ„ํŒ… ๋ฒกํ„ฐ๋ฅผ ํ•จ๊ป˜ ์‚ฌ์šฉํ•˜๋ฉด ๋‘˜ ์ค‘ ํ•˜๋‚˜๋งŒ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. (3) \(\text{llama2-chat}\rightarrow\text{CP}+\text{FT}\)**๋Š” ์ฑ„ํŒ… ๊ธฐ๋Šฅ์„ ์ƒ์‹ค์‹œํ‚ต๋‹ˆ๋‹ค:** \(\text{llama2}\rightarrow\text{CP}\rightarrow\text{FT}\)๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜์ง€๋งŒ, \(\text{llama2}\) CP \(+\text{chat}\) ๋ฒกํ„ฐ ๋˜๋Š” \(\text{llama2}\rightarrow\text{CP}\rightarrow\text{FT}+\text{chat}\) ๋ฒกํ„ฐ๋ณด๋‹ค ๋‚˜์ฉ๋‹ˆ๋‹ค. ์ด๋Š” (\text{llama2-chat}\rightarrow\text{CP}+\text{FT}\)๊ฐ€ ์ •๋ณด ์†์‹ค์„ ์œ ๋ฐœํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ฑ„ํŒ… ๋ฒกํ„ฐ ์‚ฌ์šฉ์˜ ์ค‘์š”์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ฒฐ๋ก ์ ์œผ๋กœ, ์ฑ„ํŒ… ๋ฒกํ„ฐ๋Š” CP ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ์–ธ์–ด์™€ ๊ด€๊ณ„์—†์ด ๋ช…๋ น์„ ๋”ฐ๋ฅด๋Š” ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฏธ์„ธ ์กฐ์ • ํ›„ ์ฑ„ํŒ… ๋ฒกํ„ฐ๋ฅผ ๋„์ž…ํ•˜๋ฉด ์ƒ์„ฑ ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค.
The Transformer language model architecture, originally proposed for machine translation, consists of an encoder and a decoder. The encoder is composed of a stack of N = 6 identical Transformer layers. Each layer has two sub-layers. The first one is a multi-head self-attention layer, and the other one is a simple position-wise fully connected feed-forward network. The decoder is composed of a stack of 6 identical layers. In addition to the two sub-layers in each encoder layer, the decoder has a third sub-layer, which performs multi-head attention over the output of the encoder stack. The attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is computed as a weighted sum of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key. Instead of performing a single attention function with \(d_{model}\) dimensional keys, values and queries, it is found to be beneficial to linearly project the queries, keys and values \(h\) with different, learned linear projections to \(d_{k}\), \(d_{k}\) and \(d_{v}\) dimensions, respectively. Positional encoding is incorporated to fuse information about the relative or absolute position of the tokens in the sequence.
๋จธ์‹  ๋ฒˆ์—ญ์„ ์œ„ํ•ด ์ฒ˜์Œ ์ œ์•ˆ๋œ Transformer ์–ธ์–ด ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜๋Š” Encoder์™€ Decoder๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. Encoder๋Š” N = 6์˜ ๋™์ผํ•œ Transformer ๋ ˆ์ด์–ด ์Šคํƒ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ ˆ์ด์–ด์—๋Š” ๋‘ ๊ฐœ์˜ ํ•˜์œ„ ๋ ˆ์ด์–ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ๋Š” multi-head self-attention ๋ ˆ์ด์–ด์ด๊ณ , ๋‹ค๋ฅธ ํ•˜๋‚˜๋Š” ๋‹จ์ˆœํ•œ ์œ„์น˜๋ณ„ ์™„์ „ ์—ฐ๊ฒฐ ํ”ผ๋“œ ํฌ์›Œ๋“œ ๋„คํŠธ์›Œํฌ์ž…๋‹ˆ๋‹ค. Decoder๋Š” 6๊ฐœ์˜ ๋™์ผํ•œ ๋ ˆ์ด์–ด ์Šคํƒ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๊ฐ Encoder ๋ ˆ์ด์–ด์˜ ๋‘ ๊ฐœ์˜ ํ•˜์œ„ ๋ ˆ์ด์–ด ์™ธ์—๋„ Decoder์—๋Š” Encoder ์Šคํƒ์˜ ์ถœ๋ ฅ์— ๋Œ€ํ•ด multi-head attention์„ ์ˆ˜ํ–‰ํ•˜๋Š” ์„ธ ๋ฒˆ์งธ ํ•˜์œ„ ๋ ˆ์ด์–ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. Attention ๊ธฐ๋Šฅ์€ ์ฟผ๋ฆฌ์™€ ํ‚ค-๊ฐ’ ์Œ ์„ธํŠธ๋ฅผ ์ถœ๋ ฅ์— ๋งคํ•‘ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ ์ฟผ๋ฆฌ, ํ‚ค, ๊ฐ’ ๋ฐ ์ถœ๋ ฅ์€ ๋ชจ๋‘ ๋ฒกํ„ฐ์ž…๋‹ˆ๋‹ค. ์ถœ๋ ฅ์€ ๊ฐ’์˜ ๊ฐ€์ค‘์น˜ ํ•ฉ์œผ๋กœ ๊ณ„์‚ฐ๋˜๋ฉฐ, ๊ฐ ๊ฐ’์— ํ• ๋‹น๋œ ๊ฐ€์ค‘์น˜๋Š” ํ•ด๋‹น ํ‚ค์™€ ์ฟผ๋ฆฌ์˜ ํ˜ธํ™˜์„ฑ ํ•จ์ˆ˜์— ์˜ํ•ด ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. \(d_{model}\) ์ฐจ์›์˜ ํ‚ค, ๊ฐ’, ์ฟผ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์ผ attention ๊ธฐ๋Šฅ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋Œ€์‹ , ํ•™์Šต๋œ ์„œ๋กœ ๋‹ค๋ฅธ ์„ ํ˜• ํˆฌ์˜์„ ํ†ตํ•ด ์ฟผ๋ฆฌ, ํ‚ค, ๊ฐ’ \(h\)๋ฅผ ๊ฐ๊ฐ \(d_{k}\), \(d_{k}\), \(d_{v}\) ์ฐจ์›์œผ๋กœ ์„ ํ˜•์ ์œผ๋กœ ํˆฌ์˜ํ•˜๋Š” ๊ฒƒ์ด ์œ ๋ฆฌํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์œ„์น˜ ์ธ์ฝ”๋”ฉ์€ ์‹œํ€€์Šค์—์„œ ํ† ํฐ์˜ ์ƒ๋Œ€์  ๋˜๋Š” ์ ˆ๋Œ€์  ์œ„์น˜์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์œตํ•ฉํ•˜๊ธฐ ์œ„ํ•ด ํ†ตํ•ฉ๋ฉ๋‹ˆ๋‹ค.
๋จธ์‹  ๋ฒˆ์—ญ์„ ์œ„ํ•ด ์ฒ˜์Œ ์ œ์•ˆ๋œ ํŠธ๋žœ์Šคํฌ๋จธ ์–ธ์–ด ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜๋Š” ์ธ์ฝ”๋”์™€ ๋””์ฝ”๋”๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ธ์ฝ”๋”๋Š” N = 6์˜ ๋™์ผํ•œ ํŠธ๋žœ์Šคํฌ๋จธ ๋ ˆ์ด์–ด ์Šคํƒ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ ˆ์ด์–ด์—๋Š” ๋‘ ๊ฐœ์˜ ํ•˜์œ„ ๋ ˆ์ด์–ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ๋Š” ๋‹ค์ค‘ ํ—ค๋“œ ์ž๊ธฐ ์ฃผ์˜ ๋ ˆ์ด์–ด์ด๊ณ , ๋‹ค๋ฅธ ํ•˜๋‚˜๋Š” ๋‹จ์ˆœํ•œ ์œ„์น˜๋ณ„ ์™„์ „ ์—ฐ๊ฒฐ ํ”ผ๋“œ ํฌ์›Œ๋“œ ๋„คํŠธ์›Œํฌ์ž…๋‹ˆ๋‹ค. ๋””์ฝ”๋”๋Š” 6๊ฐœ์˜ ๋™์ผํ•œ ๋ ˆ์ด์–ด ์Šคํƒ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ์ธ์ฝ”๋” ๋ ˆ์ด์–ด์˜ ๋‘ ๊ฐœ์˜ ํ•˜์œ„ ๋ ˆ์ด์–ด ์™ธ์—๋„ ๋””์ฝ”๋”์—๋Š” ์ธ์ฝ”๋” ์Šคํƒ์˜ ์ถœ๋ ฅ์— ๋Œ€ํ•ด ๋‹ค์ค‘ ํ—ค๋“œ ์ฃผ์˜๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ์„ธ ๋ฒˆ์งธ ํ•˜์œ„ ๋ ˆ์ด์–ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์˜ ๊ธฐ๋Šฅ์€ ์ฟผ๋ฆฌ์™€ ํ‚ค-๊ฐ’ ์Œ ์„ธํŠธ๋ฅผ ์ถœ๋ ฅ์— ๋งคํ•‘ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ ์ฟผ๋ฆฌ, ํ‚ค, ๊ฐ’ ๋ฐ ์ถœ๋ ฅ์€ ๋ชจ๋‘ ๋ฒกํ„ฐ์ž…๋‹ˆ๋‹ค. ์ถœ๋ ฅ์€ ๊ฐ’์˜ ๊ฐ€์ค‘์น˜ ํ•ฉ์œผ๋กœ ๊ณ„์‚ฐ๋˜๋ฉฐ, ๊ฐ ๊ฐ’์— ํ• ๋‹น๋œ ๊ฐ€์ค‘์น˜๋Š” ํ•ด๋‹น ํ‚ค์™€ ์ฟผ๋ฆฌ์˜ ํ˜ธํ™˜์„ฑ ํ•จ์ˆ˜์— ์˜ํ•ด ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. d_{model}$ ์ฐจ์›์˜ ํ‚ค, ๊ฐ’, ์ฟผ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์ผ ์ฃผ์˜ ๊ธฐ๋Šฅ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋Œ€์‹ , ํ•™์Šต๋œ ์„œ๋กœ ๋‹ค๋ฅธ ์„ ํ˜• ํˆฌ์˜์„ ํ†ตํ•ด ์ฟผ๋ฆฌ, ํ‚ค, ๊ฐ’ $h$๋ฅผ ๊ฐ๊ฐ $d_{k}$, $d_{k}$, $d_{v}$ ์ฐจ์›์œผ๋กœ ์„ ํ˜•์ ์œผ๋กœ ํˆฌ์˜ํ•˜๋Š” ๊ฒƒ์ด ์œ ๋ฆฌํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์œ„์น˜ ์ธ์ฝ”๋”ฉ์€ ์‹œํ€€์Šค์—์„œ ํ† ํฐ์˜ ์ƒ๋Œ€์  ๋˜๋Š” ์ ˆ๋Œ€์  ์œ„์น˜์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์œตํ•ฉํ•˜๊ธฐ ์œ„ํ•ด ํ†ตํ•ฉ๋ฉ๋‹ˆ๋‹ค.
**CommonCrawl.** CommonCrawl [163] is one of the largest open-source web crawling databases, containing a petabyte-scale data volume, which has been widely used as training data for existing LLMs. As the whole dataset is very large, existing studies mainly extract subsets of web pages from it within a specific period. However, due to the widespread existence of noisy and low-quality information in web data, it is necessary to perform data preprocessing before usage. Based on CommonCrawl, there are four filtered datasets that are commonly used in existing work: C4 [82], CC-Stories [155], CC-News [27], and RealNews [156]. The Colossal Clean Crawled Corpus (C4) includes five variants21, namely en (806G), en.noclean (6T), realnewslike (36G), web-textlike (17G), and multilingual (38T). The _en_ version has been utilized for pre-training T5 [82], LaMDA [68], Go-pher [64], and UL2 [89]. The multilingual C4, also called mC4, has been used in mT5 [83]. CC-Stories (31G) is composed of a subset of CommonCrawl data, in which the contents are made in a story-like way. Because the original source of CC-Stories is not available now, we include a reproduction version, _CC-Stories-R_[164], in Table II. Moreover, two news corpora extracted from CommonCrawl, _i.e._, REALNEWS (120G) and CC-News (76G), are also commonly used as the pre-training data.
**CommonCrawl.** CommonCrawl [163]์€ ํŽ˜ํƒ€๋ฐ”์ดํŠธ ๊ทœ๋ชจ์˜ ๋ฐ์ดํ„ฐ ๋ณผ๋ฅจ์„ ํฌํ•จํ•˜๋Š” ๊ฐ€์žฅ ํฐ ์˜คํ”ˆ ์†Œ์Šค ์›น ํฌ๋กค๋ง ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค ์ค‘ ํ•˜๋‚˜์ด๋ฉฐ, ๊ธฐ์กด LLMs์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ๋กœ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๋งค์šฐ ํฌ๊ธฐ ๋•Œ๋ฌธ์— ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์ฃผ๋กœ ํŠน์ • ๊ธฐ๊ฐ„ ๋‚ด ์›น ํŽ˜์ด์ง€์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์›น ๋ฐ์ดํ„ฐ์—๋Š” ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ๊ณ  ํ’ˆ์งˆ์ด ๋‚ฎ์€ ์ •๋ณด๊ฐ€ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์กด์žฌํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์šฉํ•˜๊ธฐ ์ „์— ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ๋ฅผ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. CommonCrawl์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ธฐ์กด ์ž‘์—…์—์„œ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ํ•„ํ„ฐ๋ง๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” C4 [82], CC-Stories [155], CC-News [27] ๋ฐ RealNews [156]์˜ ๋„ค ๊ฐ€์ง€๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. Colossal Clean Crawled Corpus (C4)์—๋Š” 5๊ฐ€์ง€ ๋ณ€ํ˜•21, ์ฆ‰ en(806G), en.noclean(6T), realnewslike(36G), web-textlike(17G), ๋‹ค๊ตญ์–ด(38T)๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. _en_ ๋ฒ„์ „์€ T5 [82], LaMDA [68], Go-pher [64], UL2 [89]์˜ ์‚ฌ์ „ ํ•™์Šต์— ํ™œ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. mC4๋ผ๊ณ ๋„ ํ•˜๋Š” ๋‹ค๊ตญ์–ด C4๋Š” mT5 [83]์—์„œ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. CC-Stories(31G)๋Š” CommonCrawl ๋ฐ์ดํ„ฐ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ๊ทธ ๋‚ด์šฉ์€ ์Šคํ† ๋ฆฌ์™€ ๊ฐ™์€ ๋ฐฉ์‹์œผ๋กœ ์ž‘์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ CC-Stories์˜ ์›๋ณธ ์†Œ์Šค๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๊ธฐ ๋•Œ๋ฌธ์— ํ‘œ II์— ๋ณต์ œ ๋ฒ„์ „์ธ _CC-Stories-R_[164]์„ ํฌํ•จํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ CommonCrawl์—์„œ ์ถ”์ถœํ•œ ๋‘ ๊ฐœ์˜ ๋‰ด์Šค ์ฝ”ํผ์Šค, ์ฆ‰ REALNEWS(120G)์™€ CC-News(76G)๋„ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋กœ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
**CommonCrawl.** CommonCrawl [163]์€ ํŽ˜ํƒ€๋ฐ”์ดํŠธ ๊ทœ๋ชจ์˜ ๋ฐ์ดํ„ฐ ๋ณผ๋ฅจ์„ ํฌํ•จํ•˜๋Š” ๊ฐ€์žฅ ํฐ ์˜คํ”ˆ ์†Œ์Šค ์›น ํฌ๋กค๋ง ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค ์ค‘ ํ•˜๋‚˜์ด๋ฉฐ, ๊ธฐ์กด LLM์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ๋กœ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๋งค์šฐ ํฌ๊ธฐ ๋•Œ๋ฌธ์— ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์ฃผ๋กœ ํŠน์ • ๊ธฐ๊ฐ„ ๋‚ด ์›น ํŽ˜์ด์ง€์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์›น ๋ฐ์ดํ„ฐ์—๋Š” ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ๊ณ  ํ’ˆ์งˆ์ด ๋‚ฎ์€ ์ •๋ณด๊ฐ€ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์กด์žฌํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์šฉํ•˜๊ธฐ ์ „์— ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ๋ฅผ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. CommonCrawl์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ธฐ์กด ์ž‘์—…์—์„œ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ํ•„ํ„ฐ๋ง๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” C4 [82], CC-Stories [155], CC-News [27] ๋ฐ RealNews [156]์˜ ๋„ค ๊ฐ€์ง€๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฑฐ๋Œ€ํ•œ ํด๋ฆฐ ํฌ๋กค๋ง ์ฝ”ํผ์Šค(C4)์—๋Š” 5๊ฐ€์ง€ ๋ณ€ํ˜•21, ์ฆ‰ en(806G), en.noclean(6T), realnewslike(36G), web-textlike(17G), ๋‹ค๊ตญ์–ด(38T)๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. en_ ๋ฒ„์ „์€ T5 [82], LaMDA [68], Go-pher [64], UL2 [89]์˜ ์‚ฌ์ „ ํ•™์Šต์— ํ™œ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. mC4๋ผ๊ณ ๋„ ํ•˜๋Š” ๋‹ค๊ตญ์–ด C4๋Š” mT5 [83]์—์„œ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. CC-Stories(31G)๋Š” CommonCrawl ๋ฐ์ดํ„ฐ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ๊ทธ ๋‚ด์šฉ์€ ์Šคํ† ๋ฆฌ์™€ ๊ฐ™์€ ๋ฐฉ์‹์œผ๋กœ ์ž‘์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ CC-Stories์˜ ์›๋ณธ ์†Œ์Šค๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๊ธฐ ๋•Œ๋ฌธ์— ํ‘œ II์— ๋ณต์ œ ๋ฒ„์ „์ธ _CC-Stories-R_[164]์„ ํฌํ•จํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ CommonCrawl์—์„œ ์ถ”์ถœํ•œ ๋‘ ๊ฐœ์˜ ๋‰ด์Šค ์ฝ”ํผ์Šค, ์ฆ‰ REALNEWS(120G)์™€ CC-News(76G)๋„ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋กœ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
Previous language model pre-training methods have uniformly applied a next-token prediction loss to all training tokens. Challenging this norm, we posit that _"Not all tokens in a corpus are equally important for language model training"_. Our initial analysis delves into token-level training dynamics of language model, revealing distinct loss patterns for different tokens. Leveraging these insights, we introduce a new language model called Rho-1. Unlike traditional LMs that learn to predict every next token in a corpus, Rho-1 employs Selective Language Modeling (SLM), which selectively trains on useful tokens that aligned with the desired distribution. This approach involves scoring pretraining tokens using a reference model, and then training the language model with a focused loss on tokens with higher excess loss. When continual pretraining on 15B OpenWebMath corpus, Rho-1 yields an absolute improvement in few-shot accuracy of up to 30% in 9 math tasks. After fine-tuning, Rho-1-1B and 7B achieved state-of-the-art results of 40.6% and 51.8% on MATH dataset, respectively -- matching DeepSeeMath with only 3% of the pretraining tokens. Furthermore, when pretraining on 80B general tokens, Rho-1 achieves 6.8% average enhancement across 15 diverse tasks, increasing both efficiency and performance of the language model pre-training.
์ด์ „ ์–ธ์–ด ๋ชจ๋ธ ์‚ฌ์ „ ํ•™์Šต ๋ฐฉ๋ฒ•์€ ๋ชจ๋“  ํ•™์Šต ํ† ํฐ์— ๋Œ€ํ•ด ํš์ผ์ ์œผ๋กœ ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก Loss๋ฅผ ์ ์šฉํ•ด ์™”์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ์กด์˜ ๊ด€ํ–‰์— ๋„์ „ํ•˜์—ฌ ์ €ํฌ๋Š” _"์ฝ”ํผ์Šค์˜ ๋ชจ๋“  ํ† ํฐ์ด ์–ธ์–ด ๋ชจ๋ธ ํ•™์Šต์— ๋˜‘๊ฐ™์ด ์ค‘์š”ํ•œ ๊ฒƒ์€ ์•„๋‹ˆ๋‹ค"๋ผ๋Š” ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค. ์ €ํฌ์˜ ์ดˆ๊ธฐ ๋ถ„์„์€ ์–ธ์–ด ๋ชจ๋ธ์˜ ํ† ํฐ ์ˆ˜์ค€ ํ•™์Šต ๋™์—ญํ•™์„ ํƒ๊ตฌํ•˜์—ฌ ํ† ํฐ๋งˆ๋‹ค ๋‹ค๋ฅธ Loss ํŒจํ„ด์„ ๋“œ๋Ÿฌ๋ƒ…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ํ™œ์šฉํ•˜์—ฌ Rho-1์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ์–ธ์–ด ๋ชจ๋ธ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. Rho-1์€ ๋ง๋ญ‰์น˜์˜ ๋ชจ๋“  ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๊ธฐ์กด LMs๊ณผ ๋‹ฌ๋ฆฌ ์›ํ•˜๋Š” ๋ถ„ํฌ์— ๋งž์ถฐ ์œ ์šฉํ•œ ํ† ํฐ์„ ์„ ํƒ์ ์œผ๋กœ ํ•™์Šตํ•˜๋Š” Selective Language Modeling (SLM)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์—๋Š” ์ฐธ์กฐ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต ํ† ํฐ์— ์ ์ˆ˜๋ฅผ ๋ถ€์—ฌํ•œ ๋‹ค์Œ ์ดˆ๊ณผ Loss๊ฐ€ ๋” ๋†’์€ ํ† ํฐ์— ์ดˆ์ ์„ ๋งž์ถ˜ Loss๋กœ ์–ธ์–ด ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๋Š” ๊ณผ์ •์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. 150์–ต ๊ฐœ์˜ OpenWebMath ๋ง๋ญ‰์น˜์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์—์„œ Rho-1์€ 9๊ฐœ์˜ ์ˆ˜ํ•™ ๊ณผ์ œ์—์„œ ์ตœ๋Œ€ 30%๊นŒ์ง€ few-shot ์ •ํ™•๋„๋ฅผ ์ ˆ๋Œ€์ ์œผ๋กœ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. Fine-tuning ํ›„ Rho-1-1B์™€ 7B๋Š” ๊ฐ๊ฐ MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 40.6%์™€ 51.8%์˜ ์ตœ์‹  ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ์‚ฌ์ „ ํ•™์Šต ํ† ํฐ์˜ 3%๋งŒ ์‚ฌ์šฉํ•œ DeepSeeMath์™€ ์ผ์น˜ํ•˜๋Š” ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ 800์–ต ๊ฐœ์˜ ์ผ๋ฐ˜ ํ† ํฐ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•œ Rho-1์€ 15๊ฐœ์˜ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ํ‰๊ท  6.8%์˜ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ ์‚ฌ์ „ ํ•™์Šต์˜ ํšจ์œจ์„ฑ๊ณผ ์„ฑ๋Šฅ์„ ๋ชจ๋‘ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค.
์ด์ „ ์–ธ์–ด ๋ชจ๋ธ ์‚ฌ์ „ ํ•™์Šต ๋ฐฉ๋ฒ•์€ ๋ชจ๋“  ํ•™์Šต ํ† ํฐ์— ๋Œ€ํ•ด ํš์ผ์ ์œผ๋กœ ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก ์†์‹ค์„ ์ ์šฉํ•ด ์™”์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ์กด์˜ ๊ด€ํ–‰์— ๋„์ „ํ•˜์—ฌ ์ €ํฌ๋Š” _"์ฝ”ํผ์Šค์˜ ๋ชจ๋“  ํ† ํฐ์ด ์–ธ์–ด ๋ชจ๋ธ ํ•™์Šต์— ๋˜‘๊ฐ™์ด ์ค‘์š”ํ•œ ๊ฒƒ์€ ์•„๋‹ˆ๋‹ค"๋ผ๋Š” ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค. ์ €ํฌ์˜ ์ดˆ๊ธฐ ๋ถ„์„์€ ์–ธ์–ด ๋ชจ๋ธ์˜ ํ† ํฐ ์ˆ˜์ค€ ํ•™์Šต ๋™์—ญํ•™์„ ํƒ๊ตฌํ•˜์—ฌ ํ† ํฐ๋งˆ๋‹ค ๋‹ค๋ฅธ ์†์‹ค ํŒจํ„ด์„ ๋“œ๋Ÿฌ๋ƒ…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ํ™œ์šฉํ•˜์—ฌ Rho-1์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ์–ธ์–ด ๋ชจ๋ธ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. Rho-1์€ ๋ง๋ญ‰์น˜์˜ ๋ชจ๋“  ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๊ธฐ์กด LM๊ณผ ๋‹ฌ๋ฆฌ ์›ํ•˜๋Š” ๋ถ„ํฌ์— ๋งž์ถฐ ์œ ์šฉํ•œ ํ† ํฐ์„ ์„ ํƒ์ ์œผ๋กœ ํ•™์Šตํ•˜๋Š” ์„ ํƒ์  ์–ธ์–ด ๋ชจ๋ธ๋ง(SLM)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์—๋Š” ์ฐธ์กฐ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต ํ† ํฐ์— ์ ์ˆ˜๋ฅผ ๋ถ€์—ฌํ•œ ๋‹ค์Œ ์ดˆ๊ณผ ์†์‹ค์ด ๋” ๋†’์€ ํ† ํฐ์— ์ดˆ์ ์„ ๋งž์ถ˜ ์†์‹ค๋กœ ์–ธ์–ด ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๋Š” ๊ณผ์ •์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. 150์–ต ๊ฐœ์˜ OpenWebMath ๋ง๋ญ‰์น˜์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์—์„œ Rho-1์€ 9๊ฐœ์˜ ์ˆ˜ํ•™ ๊ณผ์ œ์—์„œ ์ตœ๋Œ€ 30%๊นŒ์ง€ ๋ช‡ ๋ฒˆ์˜ ์ƒท ์ •ํ™•๋„๋ฅผ ์ ˆ๋Œ€์ ์œผ๋กœ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ๋ฏธ์„ธ ์กฐ์ • ํ›„ Rho-1-1B์™€ 7B๋Š” ๊ฐ๊ฐ MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 40.6%์™€ 51.8%์˜ ์ตœ์‹  ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ์‚ฌ์ „ ํ•™์Šต ํ† ํฐ์˜ 3%๋งŒ ์‚ฌ์šฉํ•œ DeepSeeMath์™€ ์ผ์น˜ํ•˜๋Š” ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ 800์–ต ๊ฐœ์˜ ์ผ๋ฐ˜ ํ† ํฐ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•œ Rho-1์€ 15๊ฐœ์˜ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ํ‰๊ท  6.8%์˜ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ ์‚ฌ์ „ ํ•™์Šต์˜ ํšจ์œจ์„ฑ๊ณผ ์„ฑ๋Šฅ์„ ๋ชจ๋‘ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค.
As discussed in a recent study [493], LLMs exhibit the abilities of both task recognition and task learning in ICL, but the two abilities seem to be possessed with different model scales. As shown in the experiments [493], the ability of task recognition is easier to obtain, and even a small LM with only 350M parameters can exhibit this ability, while task learning can only emerge for LLMs with at least 66B parameters. Another study [499] also supports this finding with specially designed experiments. They set up the tasks with flipped and semantically unrelated labels in the experiment, which require task learning when performing ICL. The results suggest that small LMs tend to disregard the labels and mainly depend on their prior knowledge to accomplish the task, while LLMs have the ability to surpass their prior knowledge and acquire new knowledge from demonstrations, resulting in better outcomes. Furthermore, to improve the task learning ability, Meta-In-Context Learning [500] proposes to include multiple related tasks instead of just a single one in the prompt. In addition, Symbol Tuning [501] fine-tunes LLMs on demonstrations with semantically unrelated labels (_e.g.,_ foo/bar instead of positive/negative for sentiment analysis), forcing LLMs to learn the task from demonstrations instead of relying on prior knowledge.
์ตœ๊ทผ ์—ฐ๊ตฌ[493]์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, LLM์€ In-Context Learning์—์„œ ์ž‘์—… ์ธ์‹๊ณผ ์ž‘์—… ํ•™์Šต์˜ ๋‘ ๊ฐ€์ง€ ๋Šฅ๋ ฅ์„ ๋ชจ๋‘ ๋ณด์—ฌ์ฃผ์ง€๋งŒ, ์ด ๋‘ ๊ฐ€์ง€ ๋Šฅ๋ ฅ์€ ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ ๊ทœ๋ชจ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ์‹คํ—˜[493]์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ์ž‘์—… ์ธ์‹ ๋Šฅ๋ ฅ์€ ํš๋“ํ•˜๊ธฐ๊ฐ€ ๋” ์‰ฝ๊ณ , 3์–ต 5์ฒœ๋งŒ ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋งŒ ์žˆ๋Š” ์ž‘์€ LM๋„ ์ด ๋Šฅ๋ ฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ์ž‘์—… ํ•™์Šต์€ ์ตœ์†Œ 660์–ต ๊ฐœ ์ด์ƒ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ LLM์—์„œ๋งŒ ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์—ฐ๊ตฌ[499]์—์„œ๋„ ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ์‹คํ—˜์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋’ท๋ฐ›์นจํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” In-Context Learning์„ ์ˆ˜ํ–‰ํ•  ๋•Œ ์ž‘์—… ํ•™์Šต์ด ํ•„์š”ํ•œ ๋’ค์ง‘ํžŒ ๋ ˆ์ด๋ธ”๊ณผ ์˜๋ฏธ์ ์œผ๋กœ ๊ด€๋ จ์ด ์—†๋Š” ๋ ˆ์ด๋ธ”๋กœ ํƒœ์Šคํฌ๋ฅผ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ์†Œ๊ทœ๋ชจ LM์€ ๋ ˆ์ด๋ธ”์„ ๋ฌด์‹œํ•˜๊ณ  ์ฃผ๋กœ ์‚ฌ์ „ ์ง€์‹์— ์˜์กดํ•˜์—ฌ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š” ๋ฐ˜๋ฉด, LLM์€ ์‚ฌ์ „ ์ง€์‹์„ ๋›ฐ์–ด๋„˜์–ด ๋ฐ๋ชจ์—์„œ ์ƒˆ๋กœ์šด ์ง€์‹์„ ์Šต๋“ํ•˜์—ฌ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ž‘์—… ํ•™์Šต ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด Meta-In-Context Learning[500]์€ ํ”„๋กฌํ”„ํŠธ์— ๋‹จ์ผ ์ž‘์—…์ด ์•„๋‹Œ ์—ฌ๋Ÿฌ ๊ด€๋ จ ์ž‘์—…์„ ํฌํ•จํ•˜๋„๋ก ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Symbol Tuning[501]์€ ์˜๋ฏธ์ ์œผ๋กœ ๊ด€๋ จ์ด ์—†๋Š” ๋ ˆ์ด๋ธ”(์˜ˆ: ๊ฐ์„ฑ ๋ถ„์„์˜ ๊ฒฝ์šฐ ๊ธ์ •/๋ถ€์ • ๋Œ€์‹  foo/bar)๋กœ ๋ฐ๋ชจ์— ๋Œ€ํ•œ LLM์„ Fine-tuningํ•˜์—ฌ LLM์ด ์‚ฌ์ „ ์ง€์‹์— ์˜์กดํ•˜์ง€ ์•Š๊ณ  ๋ฐ๋ชจ์—์„œ ์ž‘์—…์„ ํ•™์Šตํ•˜๋„๋ก ๊ฐ•์ œํ•ฉ๋‹ˆ๋‹ค.
์ตœ๊ทผ ์—ฐ๊ตฌ[493]์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, LLM์€ ICL์—์„œ ์ž‘์—… ์ธ์‹๊ณผ ์ž‘์—… ํ•™์Šต์˜ ๋‘ ๊ฐ€์ง€ ๋Šฅ๋ ฅ์„ ๋ชจ๋‘ ๋ณด์—ฌ์ฃผ์ง€๋งŒ, ์ด ๋‘ ๊ฐ€์ง€ ๋Šฅ๋ ฅ์€ ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ ๊ทœ๋ชจ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ์‹คํ—˜[493]์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ์ž‘์—… ์ธ์‹ ๋Šฅ๋ ฅ์€ ํš๋“ํ•˜๊ธฐ๊ฐ€ ๋” ์‰ฝ๊ณ , 3์–ต 5์ฒœ๋งŒ ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋งŒ ์žˆ๋Š” ์ž‘์€ LM๋„ ์ด ๋Šฅ๋ ฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ์ž‘์—… ํ•™์Šต์€ ์ตœ์†Œ 660์–ต ๊ฐœ ์ด์ƒ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ LLM์—์„œ๋งŒ ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์—ฐ๊ตฌ[499]์—์„œ๋„ ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ์‹คํ—˜์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋’ท๋ฐ›์นจํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ICL์„ ์ˆ˜ํ–‰ํ•  ๋•Œ ์ž‘์—… ํ•™์Šต์ด ํ•„์š”ํ•œ ๋’ค์ง‘ํžŒ ๋ ˆ์ด๋ธ”๊ณผ ์˜๋ฏธ์ ์œผ๋กœ ๊ด€๋ จ์ด ์—†๋Š” ๋ ˆ์ด๋ธ”๋กœ ํƒœ์Šคํฌ๋ฅผ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ์†Œ๊ทœ๋ชจ LM์€ ๋ ˆ์ด๋ธ”์„ ๋ฌด์‹œํ•˜๊ณ  ์ฃผ๋กœ ์‚ฌ์ „ ์ง€์‹์— ์˜์กดํ•˜์—ฌ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š” ๋ฐ˜๋ฉด, LLM์€ ์‚ฌ์ „ ์ง€์‹์„ ๋›ฐ์–ด๋„˜์–ด ๋ฐ๋ชจ์—์„œ ์ƒˆ๋กœ์šด ์ง€์‹์„ ์Šต๋“ํ•˜์—ฌ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ž‘์—… ํ•™์Šต ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๋ฉ”ํƒ€ ์ธ ์ปจํ…์ŠคํŠธ ํ•™์Šต[500]์€ ํ”„๋กฌํ”„ํŠธ์— ๋‹จ์ผ ์ž‘์—…์ด ์•„๋‹Œ ์—ฌ๋Ÿฌ ๊ด€๋ จ ์ž‘์—…์„ ํฌํ•จํ•˜๋„๋ก ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์‹ฌ๋ณผ ํŠœ๋‹[501]์€ ์˜๋ฏธ์ ์œผ๋กœ ๊ด€๋ จ์ด ์—†๋Š” ๋ ˆ์ด๋ธ”(์˜ˆ: ๊ฐ์„ฑ ๋ถ„์„์˜ ๊ฒฝ์šฐ ๊ธ์ •/๋ถ€์ • ๋Œ€์‹  foo/bar)๋กœ ๋ฐ๋ชจ์— ๋Œ€ํ•œ LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ LLM์ด ์‚ฌ์ „ ์ง€์‹์— ์˜์กดํ•˜์ง€ ์•Š๊ณ  ๋ฐ๋ชจ์—์„œ ์ž‘์—…์„ ํ•™์Šตํ•˜๋„๋ก ๊ฐ•์ œํ•ฉ๋‹ˆ๋‹ค.
Concurrent RAG work.A few concurrent works2 on RAG propose new training or prompting strategies to improve widely-adopted RAG approaches. Lin et al. (2023) fine-tune both the retriever and LM on instruction-tuning datasets in two steps. While we also train our model on diverse instruction-following datasets, Self-Rag enables retrieval on demand and selection of the best possible model output via fine-grained self-reflection, making it widely applicable and more robust and controllable. Yoran et al. (2023) use a natural language inference model and Xu et al. (2023) use a summarization model to filter out or compress retrieved passages before using them to prompt the LM to generate the output. Self-Rag processes passages in parallel and filters out irrelevant ones through self-reflection, without relying on external models at inference. Moreover, our self-reflection mechanism also evaluates other aspects of the model output quality including factuality. LATS (Zhou et al., 2023) prompt off-the-shelf LMs to search for relevant information for question answering tasks and to generate with tree search, guided by LM-generated value scores. While their value function simply indicates an overall score of each generation, Self-Rag trains to an arbitrary LM to learn to generate fine-grained self-reflection and customizable inference.
๋™์‹œ Retrieval-augmented generation (RAG) ์ž‘์—…. ๋™์‹œ ์ž‘์—…2์—์„œ RAG๋Š” ๋„๋ฆฌ ์ฑ„ํƒ๋œ RAG ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด ํ›ˆ๋ จ ๋˜๋Š” ํ”„๋กฌํ”„ํŠธ ์ „๋žต์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Lin ๋“ฑ(2023)์€ ๋‘ ๋‹จ๊ณ„๋กœ ๊ฒ€์ƒ‰๊ธฐ์™€ LM์„ ๋ชจ๋‘ Instruction Tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด Fine-tuningํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋„ ๋‹ค์–‘ํ•œ Instruction-following ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์ง€๋งŒ, Self-Rag๋Š” ์„ธ๋ถ„ํ™”๋œ ์ž๊ธฐ ์„ฑ์ฐฐ์„ ํ†ตํ•ด ํ•„์š”์— ๋”ฐ๋ผ ๊ฒ€์ƒ‰์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ๊ฐ€๋Šฅํ•œ ์ตœ์ƒ์˜ ๋ชจ๋ธ ์ถœ๋ ฅ์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋„๋ฆฌ ์ ์šฉํ•  ์ˆ˜ ์žˆ๊ณ  ๋”์šฑ ๊ฐ•๋ ฅํ•˜๊ณ  ์ œ์–ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์š”๋ž€ ๋“ฑ(2023)์€ ์ž์—ฐ์–ด ์ถ”๋ก  ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๊ณ , ์„œ ๋“ฑ(2023)์€ ์š”์•ฝ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์„ ํ•„ํ„ฐ๋งํ•˜๊ฑฐ๋‚˜ ์••์ถ•ํ•œ ํ›„ ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LM์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋‚ด ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. Self-Rag๋Š” ์ถ”๋ก  ์‹œ ์™ธ๋ถ€ ๋ชจ๋ธ์— ์˜์กดํ•˜์ง€ ์•Š๊ณ  ๋ณ‘๋ ฌ๋กœ ๊ตฌ์ ˆ์„ ์ฒ˜๋ฆฌํ•˜๊ณ  ์ž๊ธฐ ์„ฑ์ฐฐ์„ ํ†ตํ•ด ๊ด€๋ จ ์—†๋Š” ๊ตฌ์ ˆ์„ ๊ฑธ๋Ÿฌ๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ž๊ธฐ ์„ฑ์ฐฐ ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ์‚ฌ์‹ค์„ฑ์„ ํฌํ•จํ•œ ๋ชจ๋ธ ์ถœ๋ ฅ ํ’ˆ์งˆ์˜ ๋‹ค๋ฅธ ์ธก๋ฉด๋„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. LATS(Zhou et al., 2023)๋Š” ๊ธฐ์„ฑ LM์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋‚ด ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€ ์ž‘์—…์— ๊ด€๋ จ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ณ  LM์ด ์ƒ์„ฑํ•œ ๊ฐ’ ์ ์ˆ˜์— ๋”ฐ๋ผ ํŠธ๋ฆฌ ๊ฒ€์ƒ‰์„ ํ†ตํ•ด ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์˜ ๊ฐ€์น˜ ํ•จ์ˆ˜๋Š” ๋‹จ์ˆœํžˆ ๊ฐ ์ƒ์„ฑ์˜ ์ „์ฒด ์ ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ๋ฐ˜๋ฉด, Self-Rag๋Š” ์„ธ๋ถ„ํ™”๋œ ์ž๊ธฐ ์„ฑ์ฐฐ๊ณผ ์‚ฌ์šฉ์ž ์ง€์ • ๊ฐ€๋Šฅํ•œ ์ถ”๋ก ์„ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด ์ž„์˜์˜ LM์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.
๋™์‹œ RAG ์ž‘์—….๋™์‹œ ์ž‘์—…2์—์„œ RAG๋Š” ๋„๋ฆฌ ์ฑ„ํƒ๋œ RAG ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด ํ›ˆ๋ จ ๋˜๋Š” ํ”„๋กฌํ”„ํŠธ ์ „๋žต์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Lin ๋“ฑ(2023)์€ ๋‘ ๋‹จ๊ณ„๋กœ ๊ฒ€์ƒ‰๊ธฐ์™€ LM์„ ๋ชจ๋‘ ์ง€์นจ ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋„ ๋‹ค์–‘ํ•œ ์ง€์นจ ์ˆ˜ํ–‰ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์ง€๋งŒ, Self-Rag๋Š” ์„ธ๋ถ„ํ™”๋œ ์ž๊ธฐ ์„ฑ์ฐฐ์„ ํ†ตํ•ด ํ•„์š”์— ๋”ฐ๋ผ ๊ฒ€์ƒ‰์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ๊ฐ€๋Šฅํ•œ ์ตœ์ƒ์˜ ๋ชจ๋ธ ์ถœ๋ ฅ์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋„๋ฆฌ ์ ์šฉํ•  ์ˆ˜ ์žˆ๊ณ  ๋”์šฑ ๊ฐ•๋ ฅํ•˜๊ณ  ์ œ์–ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์š”๋ž€ ๋“ฑ(2023)์€ ์ž์—ฐ์–ด ์ถ”๋ก  ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๊ณ , ์„œ ๋“ฑ(2023)์€ ์š”์•ฝ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์„ ํ•„ํ„ฐ๋งํ•˜๊ฑฐ๋‚˜ ์••์ถ•ํ•œ ํ›„ ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LM์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋‚ด ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. Self-Rag๋Š” ์ถ”๋ก  ์‹œ ์™ธ๋ถ€ ๋ชจ๋ธ์— ์˜์กดํ•˜์ง€ ์•Š๊ณ  ๋ณ‘๋ ฌ๋กœ ๊ตฌ์ ˆ์„ ์ฒ˜๋ฆฌํ•˜๊ณ  ์ž๊ธฐ ์„ฑ์ฐฐ์„ ํ†ตํ•ด ๊ด€๋ จ ์—†๋Š” ๊ตฌ์ ˆ์„ ๊ฑธ๋Ÿฌ๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ž๊ธฐ ์„ฑ์ฐฐ ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ์‚ฌ์‹ค์„ฑ์„ ํฌํ•จํ•œ ๋ชจ๋ธ ์ถœ๋ ฅ ํ’ˆ์งˆ์˜ ๋‹ค๋ฅธ ์ธก๋ฉด๋„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. LATS(Zhou et al., 2023)๋Š” ๊ธฐ์„ฑ LM์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋‚ด ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€ ์ž‘์—…์— ๊ด€๋ จ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ณ  LM์ด ์ƒ์„ฑํ•œ ๊ฐ’ ์ ์ˆ˜์— ๋”ฐ๋ผ ํŠธ๋ฆฌ ๊ฒ€์ƒ‰์„ ํ†ตํ•ด ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์˜ ๊ฐ€์น˜ ํ•จ์ˆ˜๋Š” ๋‹จ์ˆœํžˆ ๊ฐ ์ƒ์„ฑ์˜ ์ „์ฒด ์ ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ๋ฐ˜๋ฉด, Self-Rag๋Š” ์„ธ๋ถ„ํ™”๋œ ์ž๊ธฐ ์„ฑ์ฐฐ๊ณผ ์‚ฌ์šฉ์ž ์ง€์ • ๊ฐ€๋Šฅํ•œ ์ถ”๋ก ์„ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด ์ž„์˜์˜ LM์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.
In order for large language model (LLM)-based assistants to effectively adapt to evolving information needs, it must be possible to update their factual knowledge through continued training on new data. The standard recipe for doing so involves continued pre-training on new documents followed by instruction-tuning on question-answer (QA) pairs. However, we find that LLMs trained with this recipe struggle to answer questions, even though the perplexity of documents is minimized. We found that QA pairs are generally straightforward, while documents are more complex, weaving many factual statements together in an intricate manner. Therefore, we hypothesize that it is beneficial to expose LLMs to QA pairs _before_ continued pre-training on documents so that the process of encoding knowledge from complex documents takes into account how this knowledge is accessed through questions. Based on this, we propose **pre-instruction-tuning (PIT)**, a method that instruction-tunes on questions prior to training on documents. This contrasts with standard instruction-tuning, which learns how to extract knowledge after training on documents. Extensive experiments and ablation studies demonstrate that PIT significantly enhances the ability of LLMs to absorb knowledge from new documents, outperforming standard instruction-tuning by 17.8%.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM) ๊ธฐ๋ฐ˜ ์–ด์‹œ์Šคํ„ดํŠธ๊ฐ€ ์ง„ํ™”ํ•˜๋Š” ์ •๋ณด ์š”๊ตฌ์— ํšจ๊ณผ์ ์œผ๋กœ ์ ์‘ํ•˜๋ ค๋ฉด ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ํ•™์Šต์„ ํ†ตํ•ด ์‚ฌ์‹ค์  ์ง€์‹์„ ์—…๋ฐ์ดํŠธํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•œ ํ‘œ์ค€ ๋ ˆ์‹œํ”ผ์—๋Š” ์ƒˆ๋กœ์šด ๋ฌธ์„œ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต์„ ๊ณ„์†ํ•œ ํ›„ question-answer (QA) ์Œ์— ๋Œ€ํ•œ Instruction Tuning์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ๋ ˆ์‹œํ”ผ๋กœ ํ•™์Šต๋œ LLM์€ ๋ฌธ์„œ์˜ Perplexity๊ฐ€ ์ตœ์†Œํ™”๋˜์—ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๋Š”๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ QA ์Œ์€ ๊ฐ„๋‹จํ•˜์ง€๋งŒ ๋ฌธ์„œ๋Š” ๋” ๋ณต์žกํ•˜์—ฌ ๋งŽ์€ ์‚ฌ์‹ค์  ์ง„์ˆ ์„ ๋ณต์žกํ•œ ๋ฐฉ์‹์œผ๋กœ ์—ฎ์–ด ๋„ฃ๋Š”๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ณต์žกํ•œ ๋ฌธ์„œ์˜ ์ง€์‹์„ ์ธ์ฝ”๋”ฉํ•˜๋Š” ๊ณผ์ •์—์„œ ์ด ์ง€์‹์ด ์งˆ๋ฌธ์„ ํ†ตํ•ด ์•ก์„ธ์Šค๋˜๋Š” ๋ฐฉ์‹์„ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ๋„๋ก ๋ฌธ์„œ์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ „์— LLM์— QA ์Œ์„ ๋…ธ์ถœ์‹œํ‚ค๋Š” ๊ฒƒ์ด ์œ ๋ฆฌํ•  ๊ฒƒ์ด๋ผ๋Š” ๊ฐ€์„ค์„ ์„ธ์› ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๋ฌธ์„œ ํ•™์Šต ์ „์— ์งˆ๋ฌธ์— ๋Œ€ํ•œ Instruction Tuning์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐฉ๋ฒ•์ธ **pre-instruction-tuning (PIT)**์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋ฌธ์„œ ํ•™์Šต ํ›„ ์ง€์‹ ์ถ”์ถœ ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•˜๋Š” ํ‘œ์ค€ Instruction Tuning๊ณผ ๋Œ€์กฐ๋ฉ๋‹ˆ๋‹ค. ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜๊ณผ ์‚ญ๊ฐ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด PIT๋Š” ์ƒˆ๋กœ์šด ๋ฌธ์„œ์—์„œ ์ง€์‹์„ ํก์ˆ˜ํ•˜๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผœ ํ‘œ์ค€ Instruction Tuning๋ณด๋‹ค 17.8% ๋” ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM) ๊ธฐ๋ฐ˜ ์–ด์‹œ์Šคํ„ดํŠธ๊ฐ€ ์ง„ํ™”ํ•˜๋Š” ์ •๋ณด ์š”๊ตฌ์— ํšจ๊ณผ์ ์œผ๋กœ ์ ์‘ํ•˜๋ ค๋ฉด ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ํ•™์Šต์„ ํ†ตํ•ด ์‚ฌ์‹ค์  ์ง€์‹์„ ์—…๋ฐ์ดํŠธํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•œ ํ‘œ์ค€ ๋ ˆ์‹œํ”ผ์—๋Š” ์ƒˆ๋กœ์šด ๋ฌธ์„œ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต์„ ๊ณ„์†ํ•œ ํ›„ ์งˆ๋ฌธ-๋‹ต๋ณ€(QA) ์Œ์— ๋Œ€ํ•œ ์ง€์‹œ ํŠœ๋‹์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ๋ ˆ์‹œํ”ผ๋กœ ํ•™์Šต๋œ LLM์€ ๋ฌธ์„œ์˜ ๋‹นํ˜น๊ฐ์ด ์ตœ์†Œํ™”๋˜์—ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๋Š”๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ QA ์Œ์€ ๊ฐ„๋‹จํ•˜์ง€๋งŒ ๋ฌธ์„œ๋Š” ๋” ๋ณต์žกํ•˜์—ฌ ๋งŽ์€ ์‚ฌ์‹ค์  ์ง„์ˆ ์„ ๋ณต์žกํ•œ ๋ฐฉ์‹์œผ๋กœ ์—ฎ์–ด ๋„ฃ๋Š”๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ณต์žกํ•œ ๋ฌธ์„œ์˜ ์ง€์‹์„ ์ธ์ฝ”๋”ฉํ•˜๋Š” ๊ณผ์ •์—์„œ ์ด ์ง€์‹์ด ์งˆ๋ฌธ์„ ํ†ตํ•ด ์•ก์„ธ์Šค๋˜๋Š” ๋ฐฉ์‹์„ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ๋„๋ก ๋ฌธ์„œ์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ „์— LLM์— QA ์Œ์„ ๋…ธ์ถœ์‹œํ‚ค๋Š” ๊ฒƒ์ด ์œ ๋ฆฌํ•  ๊ฒƒ์ด๋ผ๋Š” ๊ฐ€์„ค์„ ์„ธ์› ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๋ฌธ์„œ ํ•™์Šต ์ „์— ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์ง€์‹œ ํŠœ๋‹์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐฉ๋ฒ•์ธ **์‚ฌ์ „ ์ง€์‹œ ํŠœ๋‹(PIT)**์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋ฌธ์„œ ํ•™์Šต ํ›„ ์ง€์‹ ์ถ”์ถœ ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•˜๋Š” ํ‘œ์ค€ ์ง€์‹œ ํŠœ๋‹๊ณผ ๋Œ€์กฐ๋ฉ๋‹ˆ๋‹ค. ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜๊ณผ ์‚ญ๊ฐ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด PIT๋Š” ์ƒˆ๋กœ์šด ๋ฌธ์„œ์—์„œ ์ง€์‹์„ ํก์ˆ˜ํ•˜๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผœ ํ‘œ์ค€ ์ง€์‹œ ํŠœ๋‹๋ณด๋‹ค 17.8% ๋” ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค.
Recent advancements in LLMs have demonstrated their great potential to accomplish a wide range of tasks in the form of question-answering (QA). However, building autonomous agents is far from QA, since they need to fulfill specific roles and autonomously perceive and learn from the environment to evolve themselves like humans. To bridge the gap between traditional LLMs and autonomous agents, a crucial aspect is to design rational agent architectures to assist LLMs in maximizing their capabilities. Along this direction, previous work hasdeveloped a number of modules to enhance LLMs. In this section, we propose a unified framework to summarize these modules. In specific, the overall structure of our framework is illustrated Figure 2, which is composed of a profiling module, a memory module, a planning module, and an action module. The purpose of the profiling module is to identify the role of the agent. The memory and planning modules place the agent into a dynamic environment, enabling it to recall past behaviors and plan future actions. The action module is responsible for translating the agent's decisions into specific outputs. Within these modules, the profiling module impacts the memory and planning modules, and collectively, these three modules influence the action module. In the following, we detail these modules.
์ตœ๊ทผ LLM์˜ ๋ฐœ์ „์€ question-answering (QA) ํ˜•ํƒœ๋กœ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋ง‰๋Œ€ํ•œ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ž์œจ ์—์ด์ „ํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์€ QA์™€๋Š” ๊ฑฐ๋ฆฌ๊ฐ€ ๋ฉ€๋ฉฐ, ํŠน์ • ์—ญํ• ์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ํ™˜๊ฒฝ์„ ์ž์œจ์ ์œผ๋กœ ์ธ์‹ํ•˜๊ณ  ํ•™์Šตํ•˜์—ฌ ์ธ๊ฐ„์ฒ˜๋Ÿผ ์Šค์Šค๋กœ ์ง„ํ™”ํ•ด์•ผ ํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ธฐ์กด LLM๊ณผ ์ž์œจ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ํ•ฉ๋ฆฌ์ ์ธ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ LLM์ด ๊ทธ ๋Šฅ๋ ฅ์„ ๊ทน๋Œ€ํ™”ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•œ ์ธก๋ฉด์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉํ–ฅ์—์„œ ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” LLM์„ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ๋ชจ๋“ˆ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋“ˆ์„ ์š”์•ฝํ•˜๊ธฐ ์œ„ํ•œ ํ†ตํ•ฉ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ํ”„๋กœํŒŒ์ผ๋ง ๋ชจ๋“ˆ, ๋ฉ”๋ชจ๋ฆฌ ๋ชจ๋“ˆ, ๊ณ„ํš ๋ชจ๋“ˆ, ์•ก์…˜ ๋ชจ๋“ˆ๋กœ ๊ตฌ์„ฑ๋œ ํ”„๋ ˆ์ž„์›Œํฌ์˜ ์ „์ฒด ๊ตฌ์กฐ๋Š” ๊ทธ๋ฆผ 2์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค. ํ”„๋กœํŒŒ์ผ๋ง ๋ชจ๋“ˆ์˜ ๋ชฉ์ ์€ ์—์ด์ „ํŠธ์˜ ์—ญํ• ์„ ์‹๋ณ„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ๊ณ„ํš ๋ชจ๋“ˆ์€ ์—์ด์ „ํŠธ๋ฅผ ๋™์  ํ™˜๊ฒฝ์— ๋ฐฐ์น˜ํ•˜์—ฌ ๊ณผ๊ฑฐ ํ–‰๋™์„ ๊ธฐ์–ตํ•˜๊ณ  ๋ฏธ๋ž˜์˜ ํ–‰๋™์„ ๊ณ„ํšํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์•ก์…˜ ๋ชจ๋“ˆ์€ ์—์ด์ „ํŠธ์˜ ๊ฒฐ์ •์„ ํŠน์ • ์ถœ๋ ฅ์œผ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์—ญํ• ์„ ๋‹ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋“ˆ ๋‚ด์—์„œ ํ”„๋กœํŒŒ์ผ๋ง ๋ชจ๋“ˆ์€ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ๊ณ„ํš ๋ชจ๋“ˆ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋ฉฐ, ์ด ์„ธ ๋ชจ๋“ˆ์€ ์ด์ฒด์ ์œผ๋กœ ์•ก์…˜ ๋ชจ๋“ˆ์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋“ˆ์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์ตœ๊ทผ LLM์˜ ๋ฐœ์ „์€ ์งˆ์˜์‘๋‹ต(QA) ํ˜•ํƒœ๋กœ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋ง‰๋Œ€ํ•œ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ž์œจ ์—์ด์ „ํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์€ QA์™€๋Š” ๊ฑฐ๋ฆฌ๊ฐ€ ๋ฉ€๋ฉฐ, ํŠน์ • ์—ญํ• ์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ํ™˜๊ฒฝ์„ ์ž์œจ์ ์œผ๋กœ ์ธ์‹ํ•˜๊ณ  ํ•™์Šตํ•˜์—ฌ ์ธ๊ฐ„์ฒ˜๋Ÿผ ์Šค์Šค๋กœ ์ง„ํ™”ํ•ด์•ผ ํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ธฐ์กด LLM๊ณผ ์ž์œจ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ํ•ฉ๋ฆฌ์ ์ธ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ LLM์ด ๊ทธ ๋Šฅ๋ ฅ์„ ๊ทน๋Œ€ํ™”ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•œ ์ธก๋ฉด์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉํ–ฅ์—์„œ ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” LLM์„ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ๋ชจ๋“ˆ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋“ˆ์„ ์š”์•ฝํ•˜๊ธฐ ์œ„ํ•œ ํ†ตํ•ฉ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ํ”„๋กœํŒŒ์ผ๋ง ๋ชจ๋“ˆ, ๋ฉ”๋ชจ๋ฆฌ ๋ชจ๋“ˆ, ๊ณ„ํš ๋ชจ๋“ˆ, ์•ก์…˜ ๋ชจ๋“ˆ๋กœ ๊ตฌ์„ฑ๋œ ํ”„๋ ˆ์ž„์›Œํฌ์˜ ์ „์ฒด ๊ตฌ์กฐ๋Š” ๊ทธ๋ฆผ 2์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค. ํ”„๋กœํŒŒ์ผ๋ง ๋ชจ๋“ˆ์˜ ๋ชฉ์ ์€ ์—์ด์ „ํŠธ์˜ ์—ญํ• ์„ ์‹๋ณ„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ๊ณ„ํš ๋ชจ๋“ˆ์€ ์—์ด์ „ํŠธ๋ฅผ ๋™์  ํ™˜๊ฒฝ์— ๋ฐฐ์น˜ํ•˜์—ฌ ๊ณผ๊ฑฐ ํ–‰๋™์„ ๊ธฐ์–ตํ•˜๊ณ  ๋ฏธ๋ž˜์˜ ํ–‰๋™์„ ๊ณ„ํšํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์•ก์…˜ ๋ชจ๋“ˆ์€ ์—์ด์ „ํŠธ์˜ ๊ฒฐ์ •์„ ํŠน์ • ์ถœ๋ ฅ์œผ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์—ญํ• ์„ ๋‹ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋“ˆ ๋‚ด์—์„œ ํ”„๋กœํŒŒ์ผ๋ง ๋ชจ๋“ˆ์€ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ๊ณ„ํš ๋ชจ๋“ˆ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋ฉฐ, ์ด ์„ธ ๋ชจ๋“ˆ์€ ์ด์ฒด์ ์œผ๋กœ ์•ก์…˜ ๋ชจ๋“ˆ์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋“ˆ์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
Wang et al. (2022) iteratively acquires knowledge from the model, making progress in multi-hop QA. Zhou et al. (2023) proposes Least-to-Most Prompting, which initially breaks down the question into sub-questions in a top-down fashion, and subsequently, it solves a sub-question once at a time and leverages their solutions to facilitate subsequent sub-questions. Successive Prompting (Dua et al., 2022) takes a similar approach to Least-to-Most Prompting, and the difference is that it takes a decomposition with interleaved sub-questions and answers rather than two-stage decomposition. The above methods do not formulate tailored solutions for various sub-problems. Decomposed Prompting (Khot et al., 2023) designs a modular shared library, each dedicated to a class of subproblems, which can tailor more effective solutions to different classes of sub-problems. Apart from general tasks, some works focus on question decomposition on tabular reasoning. BINDER(Cheng et al., 2023) maps reasoning to a program in a neural-symbolic manner and obtains the final answer through a program executor such as Python or SQL. Ye et al. (2023) introduces DATER, which breaks down large tables into smaller ones and complex questions into simpler ones. The former reduces irrelevant information, while the latter reduces the complexity of reasoning.
Wang et al. (2022)๋Š” ๋ชจ๋ธ์—์„œ ์ง€์‹์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ํš๋“ํ•˜์—ฌ multi-hop QA ๋ถ„์•ผ์—์„œ ์ง„์ „์„ ์ด๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค. Zhou et al. (2023)๋Š” ์ตœ์ดˆ์— ์งˆ๋ฌธ์„ ํ•˜ํ–ฅ์‹์œผ๋กœ ํ•˜์œ„ ์งˆ๋ฌธ์œผ๋กœ ๋ถ„ํ•ดํ•œ ํ›„, ํ•œ ๋ฒˆ์— ํ•˜๋‚˜์˜ ํ•˜์œ„ ์งˆ๋ฌธ์„ ํ•ด๊ฒฐํ•˜๊ณ  ๊ทธ ํ•ด๊ฒฐ์ฑ…์„ ํ™œ์šฉํ•˜์—ฌ ํ›„์† ํ•˜์œ„ ์งˆ๋ฌธ์„ ์šฉ์ดํ•˜๊ฒŒ ํ•˜๋Š” Least-to-Most Prompting์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Successive Prompting (Dua et al., 2022)๋Š” Least-to-Most Prompting๊ณผ ์œ ์‚ฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ทจํ•˜๋ฉฐ, 2๋‹จ๊ณ„ ๋ถ„ํ•ด๊ฐ€ ์•„๋‹Œ ํ•˜์œ„ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์ด ์ธํ„ฐ๋ฆฌ๋ธŒ๋œ ๋ถ„ํ•ด๋ฅผ ์ทจํ•œ๋‹ค๋Š” ์ ์ด ์ฐจ์ด์ ์ž…๋‹ˆ๋‹ค. ์œ„์˜ ๋ฐฉ๋ฒ•์€ ๋‹ค์–‘ํ•œ ํ•˜์œ„ ๋ฌธ์ œ์— ๋Œ€ํ•œ ๋งž์ถคํ˜• ์†”๋ฃจ์…˜์„ ๊ณต์‹ํ™”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. Decomposed Prompting (Khot et al., 2023)๋Š” ๊ฐ๊ฐ ํ•˜์œ„ ๋ฌธ์ œ์˜ ํ•œ ์ข…๋ฅ˜์— ์ „๋…ํ•˜๋Š” ๋ชจ๋“ˆ์‹ ๊ณต์œ  ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ํ•˜์œ„ ๋ฌธ์ œ ํด๋ž˜์Šค์— ๋Œ€ํ•ด ๋ณด๋‹ค ํšจ๊ณผ์ ์ธ ์†”๋ฃจ์…˜์„ ๋งž์ถคํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ์ž‘์—… ์™ธ์—๋„ ์ผ๋ถ€ ์ž‘์—…์€ ํ‘œ ํ˜•์‹์˜ ์ถ”๋ก ์— ๋Œ€ํ•œ ์งˆ๋ฌธ ๋ถ„ํ•ด์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. BINDER (Cheng et al., 2023)๋Š” ์ถ”๋ก ์„ ์‹ ๊ฒฝ ๊ธฐํ˜ธ ๋ฐฉ์‹์œผ๋กœ ํ”„๋กœ๊ทธ๋žจ์— ๋งคํ•‘ํ•˜๊ณ  Python ๋˜๋Š” SQL๊ณผ ๊ฐ™์€ ํ”„๋กœ๊ทธ๋žจ ์‹คํ–‰๊ธฐ๋ฅผ ํ†ตํ•ด ์ตœ์ข… ๋‹ต์„ ์–ป์Šต๋‹ˆ๋‹ค. Ye et al. (2023)๋Š” ํฐ ํ…Œ์ด๋ธ”์„ ์ž‘์€ ํ…Œ์ด๋ธ”๋กœ, ๋ณต์žกํ•œ ์งˆ๋ฌธ์„ ๋” ๊ฐ„๋‹จํ•œ ์งˆ๋ฌธ์œผ๋กœ ๋ถ„ํ•ดํ•˜๋Š” DATER๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ „์ž๋Š” ๊ด€๋ จ ์—†๋Š” ์ •๋ณด๋ฅผ ์ค„์ด๊ณ  ํ›„์ž๋Š” ์ถ”๋ก ์˜ ๋ณต์žก์„ฑ์„ ์ค„์ž…๋‹ˆ๋‹ค.
Wang ์™ธ(2022)๋Š” ๋ชจ๋ธ์—์„œ ์ง€์‹์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ํš๋“ํ•˜์—ฌ ๋‹ค์ค‘ ํ™‰ QA ๋ถ„์•ผ์—์„œ ์ง„์ „์„ ์ด๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค. Zhou ์™ธ(2023)๋Š” ์ตœ์ดˆ์— ์งˆ๋ฌธ์„ ํ•˜ํ–ฅ์‹์œผ๋กœ ํ•˜์œ„ ์งˆ๋ฌธ์œผ๋กœ ๋ถ„ํ•ดํ•œ ํ›„, ํ•œ ๋ฒˆ์— ํ•˜๋‚˜์˜ ํ•˜์œ„ ์งˆ๋ฌธ์„ ํ•ด๊ฒฐํ•˜๊ณ  ๊ทธ ํ•ด๊ฒฐ์ฑ…์„ ํ™œ์šฉํ•˜์—ฌ ํ›„์† ํ•˜์œ„ ์งˆ๋ฌธ์„ ์šฉ์ดํ•˜๊ฒŒ ํ•˜๋Š” ์ตœ์†Œ-์ตœ๋Œ€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์—ฐ์† ํ”„๋กฌํ”„ํŠธ(Dua ์™ธ, 2022)๋Š” ์ตœ์†Œ-์ตœ๋Œ€ ํ”„๋กฌํ”„ํŠธ์™€ ์œ ์‚ฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ทจํ•˜๋ฉฐ, 2๋‹จ๊ณ„ ๋ถ„ํ•ด๊ฐ€ ์•„๋‹Œ ํ•˜์œ„ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์ด ์ธํ„ฐ๋ฆฌ๋ธŒ๋œ ๋ถ„ํ•ด๋ฅผ ์ทจํ•œ๋‹ค๋Š” ์ ์ด ์ฐจ์ด์ ์ž…๋‹ˆ๋‹ค. ์œ„์˜ ๋ฐฉ๋ฒ•์€ ๋‹ค์–‘ํ•œ ํ•˜์œ„ ๋ฌธ์ œ์— ๋Œ€ํ•œ ๋งž์ถคํ˜• ์†”๋ฃจ์…˜์„ ๊ณต์‹ํ™”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋ถ„ํ•ด๋œ ํ”„๋กฌํ”„ํŠธ(Khot ์™ธ, 2023)๋Š” ๊ฐ๊ฐ ํ•˜์œ„ ๋ฌธ์ œ์˜ ํ•œ ์ข…๋ฅ˜์— ์ „๋…ํ•˜๋Š” ๋ชจ๋“ˆ์‹ ๊ณต์œ  ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ํ•˜์œ„ ๋ฌธ์ œ ํด๋ž˜์Šค์— ๋Œ€ํ•ด ๋ณด๋‹ค ํšจ๊ณผ์ ์ธ ์†”๋ฃจ์…˜์„ ๋งž์ถคํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ์ž‘์—… ์™ธ์—๋„ ์ผ๋ถ€ ์ž‘์—…์€ ํ‘œ ํ˜•์‹์˜ ์ถ”๋ก ์— ๋Œ€ํ•œ ์งˆ๋ฌธ ๋ถ„ํ•ด์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. BINDER(Cheng ์™ธ, 2023)๋Š” ์ถ”๋ก ์„ ์‹ ๊ฒฝ ๊ธฐํ˜ธ ๋ฐฉ์‹์œผ๋กœ ํ”„๋กœ๊ทธ๋žจ์— ๋งคํ•‘ํ•˜๊ณ  Python ๋˜๋Š” SQL๊ณผ ๊ฐ™์€ ํ”„๋กœ๊ทธ๋žจ ์‹คํ–‰๊ธฐ๋ฅผ ํ†ตํ•ด ์ตœ์ข… ๋‹ต์„ ์–ป์Šต๋‹ˆ๋‹ค. Ye ์™ธ(2023)๋Š” ํฐ ํ…Œ์ด๋ธ”์„ ์ž‘์€ ํ…Œ์ด๋ธ”๋กœ, ๋ณต์žกํ•œ ์งˆ๋ฌธ์„ ๋” ๊ฐ„๋‹จํ•œ ์งˆ๋ฌธ์œผ๋กœ ๋ถ„ํ•ดํ•˜๋Š” DATER๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ „์ž๋Š” ๊ด€๋ จ ์—†๋Š” ์ •๋ณด๋ฅผ ์ค„์ด๊ณ  ํ›„์ž๋Š” ์ถ”๋ก ์˜ ๋ณต์žก์„ฑ์„ ์ค„์ž…๋‹ˆ๋‹ค.
Tree StructureThe original chain structure inherently limits the scope of exploration. Through the incorporation of tree structures and tree search algorithms, models gain the capability to efficiently explore and backtrack during the reasoning process (Long, 2023; Yao et al., 2023), as shown in Figure 2(e). Combined with self-assessment of intermediate thoughts, models can achieve global optimum solutions. The reasoning process of ToT involves uncertainty, which can potentially lead to cascading errors. TouT (Mo and Xin, 2023) introduces Monte Carlo Dropout in reasoning, taking into account the uncertainty. Yu et al. (2023) delves into analogous problems, harnessing their solutions to elevate the intricate reasoning abilities of LLMs. These analogous problems exhibit a tree-like structure, ultimately converging to solve the main problem. However, the current tree-of-thought has considerable limitations on task selection and requires specific prompt designing for each task, which hinders its widespread application. SoT (Ning et al., 2023) is another variant of the tree structure, which decomposes a problem into subproblems that can be processed in parallel and solved simultaneously to speed up reasoning. However, its utility is restricted to parallel decomposable problems and is not suited for complex reasoning tasks.
ํŠธ๋ฆฌ ๊ตฌ์กฐ์›๋ž˜์˜ ์ฒด์ธ ๊ตฌ์กฐ๋Š” ๋ณธ์งˆ์ ์œผ๋กœ ํƒ์ƒ‰ ๋ฒ”์œ„๋ฅผ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค. ํŠธ๋ฆฌ ๊ตฌ์กฐ์™€ ํŠธ๋ฆฌ ๊ฒ€์ƒ‰ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ†ตํ•ฉํ•จ์œผ๋กœ์จ ๋ชจ๋ธ์€ Figure 2(e)์™€ ๊ฐ™์ด ์ถ”๋ก  ๊ณผ์ •์—์„œ ํšจ์œจ์ ์œผ๋กœ ํƒ์ƒ‰ํ•˜๊ณ  ์—ญ์ถ”์ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ”๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ์ƒ๊ฐ์— ๋Œ€ํ•œ ์ž์ฒด ํ‰๊ฐ€์™€ ๊ฒฐํ•ฉํ•˜๋ฉด ๋ชจ๋ธ์€ ์ „์—ญ ์ตœ์  ์†”๋ฃจ์…˜์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ToT์˜ ์ถ”๋ก  ๊ณผ์ •์—๋Š” ๋ถˆํ™•์‹ค์„ฑ์ด ํฌํ•จ๋˜์–ด ์žˆ์–ด ์ž ์žฌ์ ์œผ๋กœ ์—ฐ์‡„์ ์ธ ์˜ค๋ฅ˜๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. TouT(Mo and Xin, 2023)๋Š” ๋ถˆํ™•์‹ค์„ฑ์„ ๊ณ ๋ คํ•˜์—ฌ ์ถ”๋ก ์— Monte Carlo Dropout์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. Yu et al. (2023)์€ ์œ ์‚ฌํ•œ ๋ฌธ์ œ๋ฅผ ํƒ๊ตฌํ•˜์—ฌ ๊ทธ ํ•ด๊ฒฐ์ฑ…์„ ํ™œ์šฉํ•˜์—ฌ LLM์˜ ๋ณต์žกํ•œ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ์‚ฌ ๋ฌธ์ œ๋Š” ํŠธ๋ฆฌ์™€ ๊ฐ™์€ ๊ตฌ์กฐ๋ฅผ ๋ณด์ด๋ฉฐ ๊ถ๊ทน์ ์œผ๋กœ ์ฃผ์š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์ˆ˜๋ ดํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ˜„์žฌ ์ƒ๊ฐ์˜ ํŠธ๋ฆฌ๋Š” ์ž‘์—… ์„ ํƒ์— ์ƒ๋‹นํ•œ ์ œํ•œ์ด ์žˆ์œผ๋ฉฐ ๊ฐ ์ž‘์—…์— ๋Œ€ํ•ด ํŠน์ • prompt ์„ค๊ณ„๊ฐ€ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ด‘๋ฒ”์œ„ํ•œ ์ ์šฉ์— ๋ฐฉํ•ด๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. SoT(Ning et al., 2023)๋Š” ํŠธ๋ฆฌ ๊ตฌ์กฐ์˜ ๋˜ ๋‹ค๋ฅธ ๋ณ€ํ˜•์œผ๋กœ, ๋ฌธ์ œ๋ฅผ ๋ณ‘๋ ฌ๋กœ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ํ•˜์œ„ ๋ฌธ์ œ๋กœ ๋ถ„ํ•ดํ•˜์—ฌ ๋™์‹œ์— ํ•ด๊ฒฐํ•จ์œผ๋กœ์จ ์ถ”๋ก  ์†๋„๋ฅผ ๋†’์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ๋ฐฉ๋ฒ•์€ ๋ณ‘๋ ฌ ๋ถ„ํ•ด ๊ฐ€๋Šฅํ•œ ๋ฌธ์ œ์—๋งŒ ์œ ์šฉํ•˜๋ฉฐ ๋ณต์žกํ•œ ์ถ”๋ก  ์ž‘์—…์—๋Š” ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
ํŠธ๋ฆฌ ๊ตฌ์กฐ์›๋ž˜์˜ ์ฒด์ธ ๊ตฌ์กฐ๋Š” ๋ณธ์งˆ์ ์œผ๋กœ ํƒ์ƒ‰ ๋ฒ”์œ„๋ฅผ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค. ํŠธ๋ฆฌ ๊ตฌ์กฐ์™€ ํŠธ๋ฆฌ ๊ฒ€์ƒ‰ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ†ตํ•ฉํ•จ์œผ๋กœ์จ ๋ชจ๋ธ์€ Figure 2(e)์™€ ๊ฐ™์ด ์ถ”๋ก  ๊ณผ์ •์—์„œ ํšจ์œจ์ ์œผ๋กœ ํƒ์ƒ‰ํ•˜๊ณ  ์—ญ์ถ”์ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ”๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ์ƒ๊ฐ์— ๋Œ€ํ•œ ์ž์ฒด ํ‰๊ฐ€์™€ ๊ฒฐํ•ฉํ•˜๋ฉด ๋ชจ๋ธ์€ ์ „์—ญ ์ตœ์  ์†”๋ฃจ์…˜์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ToT์˜ ์ถ”๋ก  ๊ณผ์ •์—๋Š” ๋ถˆํ™•์‹ค์„ฑ์ด ํฌํ•จ๋˜์–ด ์žˆ์–ด ์ž ์žฌ์ ์œผ๋กœ ์—ฐ์‡„์ ์ธ ์˜ค๋ฅ˜๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. TouT(Mo and Xin, 2023)๋Š” ๋ถˆํ™•์‹ค์„ฑ์„ ๊ณ ๋ คํ•˜์—ฌ ์ถ”๋ก ์— ๋ชฌํ…Œ์นด๋ฅผ๋กœ ๋“œ๋กญ์•„์›ƒ์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. Yu et al. (2023)์€ ์œ ์‚ฌํ•œ ๋ฌธ์ œ๋ฅผ ํƒ๊ตฌํ•˜์—ฌ ๊ทธ ํ•ด๊ฒฐ์ฑ…์„ ํ™œ์šฉํ•˜์—ฌ LLM์˜ ๋ณต์žกํ•œ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ์‚ฌ ๋ฌธ์ œ๋Š” ํŠธ๋ฆฌ์™€ ๊ฐ™์€ ๊ตฌ์กฐ๋ฅผ ๋ณด์ด๋ฉฐ ๊ถ๊ทน์ ์œผ๋กœ ์ฃผ์š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์ˆ˜๋ ดํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ˜„์žฌ ์ƒ๊ฐ์˜ ํŠธ๋ฆฌ๋Š” ์ž‘์—… ์„ ํƒ์— ์ƒ๋‹นํ•œ ์ œํ•œ์ด ์žˆ์œผ๋ฉฐ ๊ฐ ์ž‘์—…์— ๋Œ€ํ•ด ํŠน์ • ํ”„๋กฌํ”„ํŠธ ์„ค๊ณ„๊ฐ€ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ด‘๋ฒ”์œ„ํ•œ ์ ์šฉ์— ๋ฐฉํ•ด๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. SoT(Ning et al., 2023)๋Š” ํŠธ๋ฆฌ ๊ตฌ์กฐ์˜ ๋˜ ๋‹ค๋ฅธ ๋ณ€ํ˜•์œผ๋กœ, ๋ฌธ์ œ๋ฅผ ๋ณ‘๋ ฌ๋กœ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ํ•˜์œ„ ๋ฌธ์ œ๋กœ ๋ถ„ํ•ดํ•˜์—ฌ ๋™์‹œ์— ํ•ด๊ฒฐํ•จ์œผ๋กœ์จ ์ถ”๋ก  ์†๋„๋ฅผ ๋†’์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ๋ฐฉ๋ฒ•์€ ๋ณ‘๋ ฌ ๋ถ„ํ•ด ๊ฐ€๋Šฅํ•œ ๋ฌธ์ œ์—๋งŒ ์œ ์šฉํ•˜๋ฉฐ ๋ณต์žกํ•œ ์ถ”๋ก  ์ž‘์—…์—๋Š” ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
ResultFigure 9 shows the distribution of \(\textsc{CnstScore}\) computed with different styles of prompts \(\tau^{\prime}\) as written in the y-axis. The red case is the consistency score of the two-hop prompt that we mainly study in our work, which requires full multi-hop reasoning. Because no information to reason from is given in the input, \(\textsc{CnstScore}\) is significantly lower than the cases of other CoT-style prompts. The blue case is where what the descriptive mention refers to is given as the input, but what the LLM knows about the bridge entity's attribute needs to be internally recalled and referred to. The green cases are where the bridge entity's attribute, i.e., the answer to the prompt, is explicitly given in the input, and thus, the LLM does not need to refer to its answer to the one-hop prompt. The result demonstrates that the mean of \(\textsc{CnstScore}\) is higher for the blue cases where the model is forced to refer to its answer to the one-hop prompt than in the green cases where the model does not need to refer to the answer. The difference between the red and the blue cases would have come from the existence of the information of the descriptive mention's identity in the input prompt, which would have helped the LLM to use the connection to refer to what it knows about the bridge entity.
๊ฒฐ๊ณผ๊ทธ๋ฆผ 9๋Š” y์ถ•์— ๊ธฐ์žฌ๋œ ๊ฒƒ์ฒ˜๋Ÿผ ๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ์˜ prompts \(\tau^{\prime}\)๋กœ ๊ณ„์‚ฐ๋œ \(\textsc{CnstScore}\)์˜ ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ๋นจ๊ฐ„์ƒ‰ ์‚ฌ๋ก€๋Š” ์šฐ๋ฆฌ ์—ฐ๊ตฌ์—์„œ ์ฃผ๋กœ ์—ฐ๊ตฌํ•˜๋Š” 2-hop prompt์˜ ์ผ๊ด€์„ฑ ์ ์ˆ˜๋กœ, ์™„์ „ํ•œ multi-hop reasoning์ด ํ•„์š”ํ•˜๋‹ค. ์ž…๋ ฅ์— reasoningํ•  ์ •๋ณด๊ฐ€ ์ œ๊ณต๋˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— \(\textsc{CnstScore}\)๋Š” ๋‹ค๋ฅธ CoT ์Šคํƒ€์ผ์˜ prompts์˜ ๊ฒฝ์šฐ๋ณด๋‹ค ํ›จ์”ฌ ๋‚ฎ๋‹ค. ํŒŒ๋ž€์ƒ‰ ์‚ฌ๋ก€๋Š” ์„ค๋ช…์ด ์–ธ๊ธ‰ํ•˜๋Š” ๋‚ด์šฉ์ด ์ž…๋ ฅ์œผ๋กœ ์ฃผ์–ด์ง€์ง€๋งŒ, LLM์ด bridge entity์˜ ์†์„ฑ์— ๋Œ€ํ•ด ์•Œ๊ณ  ์žˆ๋Š” ๋‚ด์šฉ์„ ๋‚ด๋ถ€์ ์œผ๋กœ ๊ธฐ์–ตํ•˜์—ฌ ์ฐธ์กฐํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ์ด๋‹ค. ๋…น์ƒ‰ ์‚ฌ๋ก€๋Š” bridge entity์˜ ์†์„ฑ, ์ฆ‰ prompt์— ๋Œ€ํ•œ ๋‹ต์ด ์ž…๋ ฅ์— ๋ช…์‹œ์ ์œผ๋กœ ์ฃผ์–ด์ง€๋ฏ€๋กœ LLM์ด ํ•œ ๋‹จ๊ณ„ prompt์— ๋Œ€ํ•œ ๋‹ต์„ ์ฐธ์กฐํ•  ํ•„์š”๊ฐ€ ์—†๋Š” ๊ฒฝ์šฐ์ด๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ๋‹ต๋ณ€์„ ์ฐธ์กฐํ•  ํ•„์š”๊ฐ€ ์—†๋Š” ๋…น์ƒ‰ ์‚ฌ๋ก€๋ณด๋‹ค ๋ชจ๋ธ์ด ํ•œ ๋‹จ๊ณ„ prompt์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์ฐธ์กฐํ•˜๋„๋ก ๊ฐ•์ œํ•˜๋Š” ํŒŒ๋ž€์ƒ‰ ์‚ฌ๋ก€์—์„œ \(\textsc{CnstScore}\)์˜ ํ‰๊ท ์ด ๋” ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋นจ๊ฐ„์ƒ‰๊ณผ ํŒŒ๋ž€์ƒ‰ ์‚ฌ๋ก€์˜ ์ฐจ์ด๋Š” ์ž…๋ ฅ prompt์— ์„ค๋ช…์  ์–ธ๊ธ‰์˜ ์‹ ์› ์ •๋ณด๊ฐ€ ์กด์žฌํ•˜์—ฌ LLM์ด bridge entity์— ๋Œ€ํ•ด ์•Œ๊ณ  ์žˆ๋Š” ๋‚ด์šฉ์„ ์ฐธ์กฐํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜์—ˆ์„ ๊ฒƒ์ด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
๊ฒฐ๊ณผ๊ทธ๋ฆผ 9๋Š” y์ถ•์— ๊ธฐ์žฌ๋œ ๊ฒƒ์ฒ˜๋Ÿผ ๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ์˜ ํ”„๋กฌํ”„ํŠธ \(\tau^{\prime}\)๋กœ ๊ณ„์‚ฐ๋œ \(\textsc{CnstScore}\)์˜ ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ๋นจ๊ฐ„์ƒ‰ ์‚ฌ๋ก€๋Š” ์šฐ๋ฆฌ ์—ฐ๊ตฌ์—์„œ ์ฃผ๋กœ ์—ฐ๊ตฌํ•˜๋Š” 2ํ™‰ ํ”„๋กฌํ”„ํŠธ์˜ ์ผ๊ด€์„ฑ ์ ์ˆ˜๋กœ, ์™„์ „ํ•œ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์ด ํ•„์š”ํ•˜๋‹ค. ์ž…๋ ฅ์— ์ถ”๋ก ํ•  ์ •๋ณด๊ฐ€ ์ œ๊ณต๋˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— \(\textsc{CnstScore}\)๋Š” ๋‹ค๋ฅธ CoT ์Šคํƒ€์ผ์˜ ํ”„๋กฌํ”„ํŠธ์˜ ๊ฒฝ์šฐ๋ณด๋‹ค ํ›จ์”ฌ ๋‚ฎ๋‹ค. ํŒŒ๋ž€์ƒ‰ ์‚ฌ๋ก€๋Š” ์„ค๋ช…์ด ์–ธ๊ธ‰ํ•˜๋Š” ๋‚ด์šฉ์ด ์ž…๋ ฅ์œผ๋กœ ์ฃผ์–ด์ง€์ง€๋งŒ, LLM์ด ๋‹ค๋ฆฌ ์—”ํ‹ฐํ‹ฐ์˜ ์†์„ฑ์— ๋Œ€ํ•ด ์•Œ๊ณ  ์žˆ๋Š” ๋‚ด์šฉ์„ ๋‚ด๋ถ€์ ์œผ๋กœ ๊ธฐ์–ตํ•˜์—ฌ ์ฐธ์กฐํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ์ด๋‹ค. ๋…น์ƒ‰ ์‚ฌ๋ก€๋Š” ๋‹ค๋ฆฌ ์—”ํ‹ฐํ‹ฐ์˜ ์†์„ฑ, ์ฆ‰ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ๋‹ต์ด ์ž…๋ ฅ์— ๋ช…์‹œ์ ์œผ๋กœ ์ฃผ์–ด์ง€๋ฏ€๋กœ LLM์ด ํ•œ ๋‹จ๊ณ„ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ๋‹ต์„ ์ฐธ์กฐํ•  ํ•„์š”๊ฐ€ ์—†๋Š” ๊ฒฝ์šฐ์ด๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ๋‹ต๋ณ€์„ ์ฐธ์กฐํ•  ํ•„์š”๊ฐ€ ์—†๋Š” ๋…น์ƒ‰ ์‚ฌ๋ก€๋ณด๋‹ค ๋ชจ๋ธ์ด ํ•œ ๋‹จ๊ณ„ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์ฐธ์กฐํ•˜๋„๋ก ๊ฐ•์ œํ•˜๋Š” ํŒŒ๋ž€์ƒ‰ ์‚ฌ๋ก€์—์„œ \(\textsc{CnstScore}\)์˜ ํ‰๊ท ์ด ๋” ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋นจ๊ฐ„์ƒ‰๊ณผ ํŒŒ๋ž€์ƒ‰ ์‚ฌ๋ก€์˜ ์ฐจ์ด๋Š” ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์— ์„ค๋ช…์  ์–ธ๊ธ‰์˜ ์‹ ์› ์ •๋ณด๊ฐ€ ์กด์žฌํ•˜์—ฌ LLM์ด ๋‹ค๋ฆฌ ์—”ํ‹ฐํ‹ฐ์— ๋Œ€ํ•ด ์•Œ๊ณ  ์žˆ๋Š” ๋‚ด์šฉ์„ ์ฐธ์กฐํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜์—ˆ์„ ๊ฒƒ์ด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
Retrieval augmented generation (RAG) is often used to fix hallucinations and provide up-to-date knowledge for large language models (LLMs). However, in cases when the LLM alone incorrectly answers a question, does providing the correct retrieved content always fix the error? Conversely, in cases where the retrieved content is incorrect, does the LLM know to ignore the wrong information, or does it recapitulate the error? To answer these questions, we systematically analyze the tug-of-war between a LLM's internal knowledge (i.e. its prior) and the retrieved information in settings when they disagree. We test GPT-4 and other LLMs on question-answering abilities across datasets with and without reference documents. As expected, providing the correct retrieved information fixes most model mistakes (94% accuracy). However, when the reference document is perturbed with increasing levels of wrong values, the LLM is more likely to recite the incorrect, modified information when its internal prior is weaker but is more resistant when its prior is stronger. Similarly, we also find that the more the modified information deviates from the model's prior, the less likely the model is to prefer it. These results highlight an underlying tension between a model's prior knowledge and the information presented in reference documents.
๋ฆฌํŠธ๋ฆฌ๋ฒŒ ์ฆ๊ฐ• ์ƒ์„ฑ(RAG)์€ ์ข…์ข… ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์˜ hallucination์„ ์ˆ˜์ •ํ•˜๊ณ  ์ตœ์‹  ์ง€์‹์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LLM๋งŒ์œผ๋กœ ์งˆ๋ฌธ์— ์ž˜๋ชป ๋‹ตํ•˜๋Š” ๊ฒฝ์šฐ, ์˜ฌ๋ฐ”๋ฅธ ๊ฒ€์ƒ‰ ์ฝ˜ํ…์ธ ๋ฅผ ์ œ๊ณตํ•˜๋ฉด ํ•ญ์ƒ ์˜ค๋ฅ˜๊ฐ€ ์ˆ˜์ •๋˜๋‚˜์š”? ๋ฐ˜๋Œ€๋กœ ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ๊ฐ€ ์ž˜๋ชป๋œ ๊ฒฝ์šฐ LLM์ด ์ž˜๋ชป๋œ ์ •๋ณด๋ฅผ ๋ฌด์‹œํ•ด์•ผ ํ•˜๋Š”์ง€, ์•„๋‹ˆ๋ฉด ์˜ค๋ฅ˜๋ฅผ ๋ฐ˜๋ณตํ•ด์•ผ ํ•˜๋Š”์ง€ ์•Œ๊ณ  ์žˆ๋‚˜์š”? ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” LLM์˜ ๋‚ด๋ถ€ ์ง€์‹(์ฆ‰, prior)๊ณผ ๊ฒ€์ƒ‰๋œ ์ •๋ณด๊ฐ€ ์„œ๋กœ ๋™์˜ํ•˜์ง€ ์•Š๋Š” ์„ค์ •์—์„œ LLM์˜ ๋‚ด๋ถ€ ์ง€์‹๊ณผ ๊ฒ€์ƒ‰๋œ ์ •๋ณด ์‚ฌ์ด์˜ ์ค„๋‹ค๋ฆฌ๊ธฐ๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. ์ฐธ์กฐ ๋ฌธ์„œ๊ฐ€ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ์—†๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€ ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด GPT-4 ๋ฐ ๊ธฐํƒ€ LLM์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ์ƒ๋Œ€๋กœ ์˜ฌ๋ฐ”๋ฅธ ๊ฒ€์ƒ‰ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋ฉด ๋Œ€๋ถ€๋ถ„์˜ ๋ชจ๋ธ ์‹ค์ˆ˜(์ •ํ™•๋„ 94%)๋ฅผ ์ˆ˜์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฐธ์กฐ ๋ฌธ์„œ๊ฐ€ ์ ์  ๋” ๋งŽ์€ ์ˆ˜์ค€์˜ ์ž˜๋ชป๋œ ๊ฐ’์œผ๋กœ ๊ต๋ž€๋  ๋•Œ, LLM์€ ๋‚ด๋ถ€ prior๊ฐ€ ์•ฝํ•  ๋•Œ๋Š” ์ž˜๋ชป๋œ ์ˆ˜์ •๋œ ์ •๋ณด๋ฅผ ์•”์†กํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์ง€๋งŒ prior๊ฐ€ ๊ฐ•ํ•  ๋•Œ๋Š” ๋” ๊ฐ•ํ•œ ์ €ํ•ญ๋ ฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ˆ˜์ •๋œ ์ •๋ณด๊ฐ€ ๋ชจ๋ธ์˜ prior์—์„œ ๋ฒ—์–ด๋‚ ์ˆ˜๋ก ๋ชจ๋ธ์ด ์„ ํ˜ธํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋‚ฎ๋‹ค๋Š” ์‚ฌ์‹ค๋„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์˜ ์‚ฌ์ „ ์ง€์‹๊ณผ ์ฐธ์กฐ ๋ฌธ์„œ์— ์ œ์‹œ๋œ ์ •๋ณด ์‚ฌ์ด์— ๊ทผ๋ณธ์ ์ธ ๊ธด์žฅ์ด ์žˆ์Œ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
๋ฆฌํŠธ๋ฆฌ๋ฒŒ ์ฆ๊ฐ• ์ƒ์„ฑ(RAG)์€ ์ข…์ข… ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์˜ ํ™˜๊ฐ์„ ์ˆ˜์ •ํ•˜๊ณ  ์ตœ์‹  ์ง€์‹์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LLM๋งŒ์œผ๋กœ ์งˆ๋ฌธ์— ์ž˜๋ชป ๋‹ตํ•˜๋Š” ๊ฒฝ์šฐ, ์˜ฌ๋ฐ”๋ฅธ ๊ฒ€์ƒ‰ ์ฝ˜ํ…์ธ ๋ฅผ ์ œ๊ณตํ•˜๋ฉด ํ•ญ์ƒ ์˜ค๋ฅ˜๊ฐ€ ์ˆ˜์ •๋˜๋‚˜์š”? ๋ฐ˜๋Œ€๋กœ ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ๊ฐ€ ์ž˜๋ชป๋œ ๊ฒฝ์šฐ LLM์ด ์ž˜๋ชป๋œ ์ •๋ณด๋ฅผ ๋ฌด์‹œํ•ด์•ผ ํ•˜๋Š”์ง€, ์•„๋‹ˆ๋ฉด ์˜ค๋ฅ˜๋ฅผ ๋ฐ˜๋ณตํ•ด์•ผ ํ•˜๋Š”์ง€ ์•Œ๊ณ  ์žˆ๋‚˜์š”? ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” LLM์˜ ๋‚ด๋ถ€ ์ง€์‹(์ฆ‰, ์„ ํ—˜)๊ณผ ๊ฒ€์ƒ‰๋œ ์ •๋ณด๊ฐ€ ์„œ๋กœ ๋™์˜ํ•˜์ง€ ์•Š๋Š” ์„ค์ •์—์„œ LLM์˜ ๋‚ด๋ถ€ ์ง€์‹๊ณผ ๊ฒ€์ƒ‰๋œ ์ •๋ณด ์‚ฌ์ด์˜ ์ค„๋‹ค๋ฆฌ๊ธฐ๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. ์ฐธ์กฐ ๋ฌธ์„œ๊ฐ€ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ์—†๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€ ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด GPT-4 ๋ฐ ๊ธฐํƒ€ LLM์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ์ƒ๋Œ€๋กœ ์˜ฌ๋ฐ”๋ฅธ ๊ฒ€์ƒ‰ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋ฉด ๋Œ€๋ถ€๋ถ„์˜ ๋ชจ๋ธ ์‹ค์ˆ˜(์ •ํ™•๋„ 94%)๋ฅผ ์ˆ˜์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฐธ์กฐ ๋ฌธ์„œ๊ฐ€ ์ ์  ๋” ๋งŽ์€ ์ˆ˜์ค€์˜ ์ž˜๋ชป๋œ ๊ฐ’์œผ๋กœ ๊ต๋ž€๋  ๋•Œ, LLM์€ ๋‚ด๋ถ€ ์„ ํ—˜์ด ์•ฝํ•  ๋•Œ๋Š” ์ž˜๋ชป๋œ ์ˆ˜์ •๋œ ์ •๋ณด๋ฅผ ์•”์†กํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์ง€๋งŒ ์„ ํ—˜์ด ๊ฐ•ํ•  ๋•Œ๋Š” ๋” ๊ฐ•ํ•œ ์ €ํ•ญ๋ ฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ˆ˜์ •๋œ ์ •๋ณด๊ฐ€ ๋ชจ๋ธ์˜ ์„ ํ—˜์—์„œ ๋ฒ—์–ด๋‚ ์ˆ˜๋ก ๋ชจ๋ธ์ด ์„ ํ˜ธํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋‚ฎ๋‹ค๋Š” ์‚ฌ์‹ค๋„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์˜ ์‚ฌ์ „ ์ง€์‹๊ณผ ์ฐธ์กฐ ๋ฌธ์„œ์— ์ œ์‹œ๋œ ์ •๋ณด ์‚ฌ์ด์— ๊ทผ๋ณธ์ ์ธ ๊ธด์žฅ์ด ์žˆ์Œ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
As briefly described in Section 4.4, we observe that many of our validation sets are close (in cosine distance) to our training sets, and the impact of data selection is varies across individual validation sets. Individual validation sets live in different regions of the embedding space, and as such they are affected differently by data selection. For example, one could imagine that web-snapshot validation sets such as C4 is close to CC-dedup in the embedding space, while esoteric validation sets (such as Gutenberg PG 19 or DM Mathematics) might be far. To quantify this, we first find the nearest neighbors in the training set to each validation point in all of our validation sets. We then qualitatively check (see Table A8 and Table A9 for examples) that nearest-neighbors in the training set truly convey information about validation points. we observe significant overlap between training points and validation points. We then quantitatively analyze how close each validation set is to the training set: in Figure A12, we show the breakdown of this distribution for each validation set. We see a general trend, that web-snapshots validation sets are closest to the training set as they are skewed to the right, while more esoteric validation sets (Gutenberg, or Wikipedia (en)) are more centered or even slightly left-skewed.
4.4์ ˆ์—์„œ ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ–ˆ๋“ฏ์ด, ์šฐ๋ฆฌ์˜ validation ์„ธํŠธ ์ค‘ ์ƒ๋‹น์ˆ˜๊ฐ€ ์šฐ๋ฆฌ์˜ training ์„ธํŠธ์— (cosine distance์—์„œ) ๊ฐ€๊น๊ณ , ๋ฐ์ดํ„ฐ ์„ ํƒ์˜ ์˜ํ–ฅ์€ ๊ฐœ๋ณ„ validation ์„ธํŠธ๋งˆ๋‹ค ๋‹ค๋ฅด๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค. ๊ฐœ๋ณ„ validation ์„ธํŠธ๋Š” embedding ๊ณต๊ฐ„์˜ ์„œ๋กœ ๋‹ค๋ฅธ ์˜์—ญ์— ์กด์žฌํ•˜๋ฏ€๋กœ ๋ฐ์ดํ„ฐ ์„ ํƒ์˜ ์˜ํ–ฅ์„ ๋‹ค๋ฅด๊ฒŒ ๋ฐ›์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์›น ์Šค๋ƒ…์ƒท validation ์„ธํŠธ(์˜ˆ: C4)๋Š” embedding ๊ณต๊ฐ„์—์„œ CC-dedup์™€ ๊ฐ€๊น์ง€๋งŒ, ๋‚œํ•ดํ•œ validation ์„ธํŠธ(์˜ˆ: Gutenberg PG 19 ๋˜๋Š” DM Mathematics)๋Š” ๋ฉ€๋ฆฌ ๋–จ์–ด์ ธ ์žˆ์„ ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ์ •๋Ÿ‰ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋จผ์ € ๋ชจ๋“  validation ์„ธํŠธ์—์„œ ๊ฐ validation ์ ์— ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด training ์„ธํŠธ๋ฅผ ์ฐพ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ training ์„ธํŠธ์˜ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ด์›ƒ์ด validation ์ ์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์‹ค์ œ๋กœ ์ „๋‹ฌํ•˜๋Š”์ง€ ์ •์„ฑ์ ์œผ๋กœ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค(์˜ˆ์‹œ๋Š” ํ‘œ A8 ๋ฐ ํ‘œ A9 ์ฐธ์กฐ). training ์ ๊ณผ validation ์  ์‚ฌ์ด์— ์ƒ๋‹นํ•œ ์ค‘๋ณต์ด ์žˆ์Œ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ฐ validation ์„ธํŠธ๊ฐ€ training ์„ธํŠธ์— ์–ผ๋งˆ๋‚˜ ๊ฐ€๊นŒ์šด์ง€ ์ •๋Ÿ‰์ ์œผ๋กœ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค: ๊ทธ๋ฆผ A12์—๋Š” ๊ฐ validation ์„ธํŠธ์— ๋Œ€ํ•œ ์ด ๋ถ„ํฌ์˜ ๋ถ„์„์ด ํ‘œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์›น ์Šค๋ƒ…์ƒท validation ์„ธํŠธ๋Š” ์˜ค๋ฅธ์ชฝ์œผ๋กœ ์น˜์šฐ์ณ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— training ์„ธํŠธ์— ๊ฐ€์žฅ ๊ฐ€๊น๊ณ , ๋‚œํ•ดํ•œ validation ์„ธํŠธ(Gutenberg ๋˜๋Š” Wikipedia (en))๋Š” ์ค‘์•™์— ์œ„์น˜ํ•˜๊ฑฐ๋‚˜ ์•ฝ๊ฐ„ ์™ผ์ชฝ์œผ๋กœ ์น˜์šฐ์ณ ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ ์ถ”์„ธ๋ฅผ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
4.4์ ˆ์—์„œ ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ–ˆ๋“ฏ์ด, ์šฐ๋ฆฌ์˜ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ ์ค‘ ์ƒ๋‹น์ˆ˜๊ฐ€ ์šฐ๋ฆฌ์˜ ํ›ˆ๋ จ ์„ธํŠธ์— (์ฝ”์‚ฌ์ธ ๊ฑฐ๋ฆฌ์—์„œ) ๊ฐ€๊น๊ณ , ๋ฐ์ดํ„ฐ ์„ ํƒ์˜ ์˜ํ–ฅ์€ ๊ฐœ๋ณ„ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋งˆ๋‹ค ๋‹ค๋ฅด๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค. ๊ฐœ๋ณ„ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋Š” ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์˜ ์„œ๋กœ ๋‹ค๋ฅธ ์˜์—ญ์— ์กด์žฌํ•˜๋ฏ€๋กœ ๋ฐ์ดํ„ฐ ์„ ํƒ์˜ ์˜ํ–ฅ์„ ๋‹ค๋ฅด๊ฒŒ ๋ฐ›์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์›น ์Šค๋ƒ…์ƒท ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ(์˜ˆ: C4)๋Š” ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์—์„œ CC-์ค‘๋ณต ์ œ๊ฑฐ์™€ ๊ฐ€๊น์ง€๋งŒ, ๋‚œํ•ดํ•œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ(์˜ˆ: ๊ตฌํ…๋ฒ ๋ฅดํฌ PG 19 ๋˜๋Š” DM ์ˆ˜ํ•™)๋Š” ๋ฉ€๋ฆฌ ๋–จ์–ด์ ธ ์žˆ์„ ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ์ •๋Ÿ‰ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋จผ์ € ๋ชจ๋“  ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์—์„œ ๊ฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ ์— ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ํ›ˆ๋ จ ์„ธํŠธ๋ฅผ ์ฐพ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ›ˆ๋ จ ์„ธํŠธ์˜ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ด์›ƒ์ด ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ ์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์‹ค์ œ๋กœ ์ „๋‹ฌํ•˜๋Š”์ง€ ์ •์„ฑ์ ์œผ๋กœ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค(์˜ˆ์‹œ๋Š” ํ‘œ A8 ๋ฐ ํ‘œ A9 ์ฐธ์กฐ). ํ›ˆ๋ จ ์ ๊ณผ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์  ์‚ฌ์ด์— ์ƒ๋‹นํ•œ ์ค‘๋ณต์ด ์žˆ์Œ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๊ฐ€ ํ›ˆ๋ จ ์„ธํŠธ์— ์–ผ๋งˆ๋‚˜ ๊ฐ€๊นŒ์šด์ง€ ์ •๋Ÿ‰์ ์œผ๋กœ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค: ๊ทธ๋ฆผ A12์—๋Š” ๊ฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์— ๋Œ€ํ•œ ์ด ๋ถ„ํฌ์˜ ๋ถ„์„์ด ํ‘œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์›น ์Šค๋ƒ…์ƒท ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋Š” ์˜ค๋ฅธ์ชฝ์œผ๋กœ ์น˜์šฐ์ณ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ํ›ˆ๋ จ ์„ธํŠธ์— ๊ฐ€์žฅ ๊ฐ€๊น๊ณ , ๋‚œํ•ดํ•œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ(๊ตฌํ…๋ฒ ๋ฅดํฌ ๋˜๋Š” ์œ„ํ‚ค๋ฐฑ๊ณผ(en))๋Š” ์ค‘์•™์— ์œ„์น˜ํ•˜๊ฑฐ๋‚˜ ์•ฝ๊ฐ„ ์™ผ์ชฝ์œผ๋กœ ์น˜์šฐ์ณ ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ ์ถ”์„ธ๋ฅผ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.