question
stringlengths
101
4.47k
chosen
stringlengths
3
4.25k
rejected
stringlengths
39
9.29k
Continue PretrainingWe continue pretrain the full-attention model using sequence parallelism [43] and distributed attention. This is to say, we do not use any sparse or linear attention, but use a brute force implementation of the full attention. We continue pretrain the Yi 6B/ 34B base model on the data mixture of (1). original pretraining data, as is introduced in section 2; (2). length-upsampled long-context data, where the long documents are mostly from books; (3). multi-document question-answering synthetic data, where we construct QA pairs where the answer contains a recitation of the related paragraph before the answer. Our data approach mostly follows the data engineering practice in Fu et al. [22] and Yu et al. [87]. We continue pretrain the model on 5B tokens with 4M batch size, which translate to 100 optimization steps. Aligning with the concurrent work from Fu et al. [22], we observe that such light-weight continue pretraining is already able to enable a strong performance on Needle-in-a-Haystack test, as we will show in Figure 6.
์‚ฌ์ „ ํ•™์Šต ๊ณ„์†ํ•˜๊ธฐ. ์šฐ๋ฆฌ๋Š” sequence parallelism [43]๊ณผ distributed attention์„ ์‚ฌ์šฉํ•˜์—ฌ full-attention ๋ชจ๋ธ์„ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, sparse ๋˜๋Š” linear attention์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ณ  full attention์— ๋Œ€ํ•œ brute force ๊ตฌํ˜„์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์„น์…˜ 2์—์„œ ์†Œ๊ฐœํ•œ (1) ์›๋ž˜์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ; (2) ๊ธธ์ด๊ฐ€ ์—…์ƒ˜ํ”Œ๋ง๋œ long-context ๋ฐ์ดํ„ฐ(๊ธด ๋ฌธ์„œ๋Š” ๋Œ€๋ถ€๋ถ„ ์ฑ…์—์„œ ๊ฐ€์ ธ์˜ด); (3) ๋‹ต๋ณ€์— ๋‹ต๋ณ€ ์•ž์˜ ๊ด€๋ จ ๋‹จ๋ฝ์„ ์ธ์šฉํ•˜๋Š” QA ์Œ์„ ๊ตฌ์„ฑํ•˜๋Š” multi-document question-answering synthetic ๋ฐ์ดํ„ฐ์˜ ํ˜ผํ•ฉ๋ฌผ์— ๋Œ€ํ•ด Yi 6B/34B base ๋ชจ๋ธ์„ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฐ์ดํ„ฐ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋Œ€๋ถ€๋ถ„ Fu et al. [22] ๋ฐ Yu et al. [87]์˜ ๋ฐ์ดํ„ฐ ์—”์ง€๋‹ˆ์–ด๋ง ๊ด€ํ–‰์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” 4M batch size๋กœ 5B tokens์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•˜์—ฌ 100 optimization ๋‹จ๊ณ„๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค. Fu et al. [22]์˜ ๋™์‹œ ์ž‘์—…๊ณผ ์ผ์น˜ํ•˜๋„๋ก, ์šฐ๋ฆฌ๋Š” ๊ทธ๋ฆผ 6์—์„œ ๋ณด์—ฌ์งˆ ๊ฒƒ์ฒ˜๋Ÿผ ์ด๋Ÿฌํ•œ ๊ฐ€๋ฒผ์šด ์›จ์ดํŠธ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šต์ด ์ด๋ฏธ Needle-in-a-Haystack ํ…Œ์ŠคํŠธ์—์„œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.
์‚ฌ์ „ ํ•™์Šต ๊ณ„์†ํ•˜๊ธฐ์‹œํ€€์Šค ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ[43]์™€ ๋ถ„์‚ฐ ์ฃผ์˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ „์ฒด ์ฃผ์˜ ๋ชจ๋ธ์„ ์‚ฌ์ „ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์ŠคํŒŒ์Šค ๋˜๋Š” ์„ ํ˜• ์ฃผ์˜๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ณ  ์ „์ฒด ์ฃผ์˜์— ๋Œ€ํ•œ ๋ฌด์ฐจ๋ณ„ ๋Œ€์ž… ๊ตฌํ˜„์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 2์—์„œ ์†Œ๊ฐœํ•œ (1)์˜ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ๋ฌผ์— ๋Œ€ํ•ด Yi 6B/34B ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ์‚ฌ์ „ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค; (2). ๊ธธ์ด๊ฐ€ ์—…์ƒ˜ํ”Œ๋ง๋œ ๊ธด ์ปจํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ(๊ธด ๋ฌธ์„œ๋Š” ๋Œ€๋ถ€๋ถ„ ์ฑ…์—์„œ ๊ฐ€์ ธ์˜ด); (3). ๋‹ต๋ณ€์— ๋‹ต๋ณ€ ์•ž์˜ ๊ด€๋ จ ๋‹จ๋ฝ์„ ์ธ์šฉํ•˜๋Š” QA ์Œ์„ ๊ตฌ์„ฑํ•˜๋Š” ๋‹ค์ค‘ ๋ฌธ์„œ ์งˆ์˜์‘๋‹ต ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ. ๋ฐ์ดํ„ฐ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋Œ€๋ถ€๋ถ„ Fu et al. [22] ๋ฐ Yu et al. [87]. 4M ๋ฐฐ์น˜ ํฌ๊ธฐ๋กœ 5B ํ† ํฐ์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ๊ณ„์† ํ•™์Šตํ•˜์—ฌ 100๊ฐœ์˜ ์ตœ์ ํ™” ๋‹จ๊ณ„๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค. Fu et al. [22]์˜ ๋™์‹œ ์ž‘์—…๊ณผ ์ผ์น˜ํ•˜๋„๋ก, ์šฐ๋ฆฌ๋Š” ๊ทธ๋ฆผ 6์—์„œ ๋ณด์—ฌ์งˆ ๊ฒƒ์ฒ˜๋Ÿผ ์ด๋Ÿฌํ•œ ๊ฐ€๋ฒผ์šด ์›จ์ดํŠธ ๊ณ„์† ํ•™์Šต์ด ์ด๋ฏธ ๊ฑด์ดˆ๋”๋ฏธ์—์„œ ๋ฐ”๋Š˜์„ ์ฐพ๋Š” ํ…Œ์ŠคํŠธ์—์„œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.
Overall we find OLMo to have a competitive fit, especially given its training data was explicitly decontaminated against Paloma. As seen through the comparison of final models (see shapes) as well intermediate checkpoints (see dashed lines), the OLMo results follow similar scaling trends of other models. Note that the performance of intermediate checkpoints is influenced by where that checkpoint occurs in the learning rate schedule. So models trained for fewer steps will tend to have steeper training curves without necessarily being more sample efficient if training duration were fixed across all models. MPT-7B, nevertheless, stands out as improving ahead of the other models in this subplot. This could be due to a number of factors, including pretraining data composition and its match to the domains in Paloma (e.g., MPT trains on 27% non-Common Crawl data rather than 18% for LLAMA, 12.2% for RedPajama, and 11.2% for OLMo) as well as various data preprocessing decisions (e.g., MPT's use of semantic deduplication by Abbas et al., 2023, on C4).
์ „๋ฐ˜์ ์œผ๋กœ OLMo๋Š” ํŠนํžˆ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๊ฐ€ Paloma์— ๋Œ€ํ•ด ๋ช…์‹œ์ ์œผ๋กœ ์†Œ๋…๋˜์—ˆ๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•  ๋•Œ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ํŒ๋‹จ๋ฉ๋‹ˆ๋‹ค. ์ตœ์ข… ๋ชจ๋ธ(๋ชจ์–‘ ์ฐธ์กฐ)๊ณผ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ(์ ์„  ์ฐธ์กฐ)์˜ ๋น„๊ต์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, OLMo ๊ฒฐ๊ณผ๋Š” ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ scaling ์ถ”์„ธ๋ฅผ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์˜ ์„ฑ๋Šฅ์€ Learning Rate ์ผ์ •์—์„œ ํ•ด๋‹น ์ฒดํฌํฌ์ธํŠธ๊ฐ€ ๋ฐœ์ƒํ•˜๋Š” ์‹œ์ ์— ์˜ํ–ฅ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ชจ๋“  ๋ชจ๋ธ์—์„œ ํ›ˆ๋ จ ๊ธฐ๊ฐ„์ด ๊ณ ์ •๋˜์–ด ์žˆ๋‹ค๋ฉด ํ›ˆ๋ จ ๋‹จ๊ณ„๊ฐ€ ์ ์€ ๋ชจ๋ธ์€ ๋ฐ˜๋“œ์‹œ ์ƒ˜ํ”Œ ํšจ์œจ์ด ๋†’์•„์ง€๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ํ›ˆ๋ จ ๊ณก์„ ์ด ๋” ๊ฐ€ํŒŒ๋ฅด๊ฒŒ ๋Š˜์–ด๋‚˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  MPT-7B๋Š” ์ด ํ•˜์œ„ ํ”Œ๋กฏ์—์„œ ๋‹ค๋ฅธ ๋ชจ๋ธ๋ณด๋‹ค ์•ž์„œ ๊ฐœ์„ ๋œ ๊ฒƒ์œผ๋กœ ๋ˆˆ์— ๋•๋‹ˆ๋‹ค. ์ด๋Š” ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ๊ณผ Paloma์˜ ๋„๋ฉ”์ธ๊ณผ์˜ ์ผ์น˜(์˜ˆ: MPT๋Š” LLAMA์˜ ๊ฒฝ์šฐ 18%, RedPajama์˜ ๊ฒฝ์šฐ 12.2%, OLMo์˜ ๊ฒฝ์šฐ 11.2%๊ฐ€ ์•„๋‹Œ 27%์˜ ๋น„์ปค๋จผ ํฌ๋กค๋ง ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต) ๋ฐ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ ๊ฒฐ์ •(์˜ˆ: MPT์˜ C4์— ๋Œ€ํ•œ Abbas ๋“ฑ(2023)์˜ ์˜๋ฏธ๋ก ์  ์ค‘๋ณต ์ œ๊ฑฐ ์‚ฌ์šฉ) ๋“ฑ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์š”์ธ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ „๋ฐ˜์ ์œผ๋กœ OLMo๋Š” ํŠนํžˆ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๊ฐ€ ํŒ”๋กœ๋งˆ์— ๋Œ€ํ•ด ๋ช…์‹œ์ ์œผ๋กœ ์†Œ๋…๋˜์—ˆ๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•  ๋•Œ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ํŒ๋‹จ๋ฉ๋‹ˆ๋‹ค. ์ตœ์ข… ๋ชจ๋ธ(๋ชจ์–‘ ์ฐธ์กฐ)๊ณผ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ(์ ์„  ์ฐธ์กฐ)์˜ ๋น„๊ต์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, OLMo ๊ฒฐ๊ณผ๋Š” ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ํ™•์žฅ ์ถ”์„ธ๋ฅผ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์˜ ์„ฑ๋Šฅ์€ ํ•™์Šต ์†๋„ ์ผ์ •์—์„œ ํ•ด๋‹น ์ฒดํฌํฌ์ธํŠธ๊ฐ€ ๋ฐœ์ƒํ•˜๋Š” ์‹œ์ ์— ์˜ํ–ฅ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ชจ๋“  ๋ชจ๋ธ์—์„œ ํ›ˆ๋ จ ๊ธฐ๊ฐ„์ด ๊ณ ์ •๋˜์–ด ์žˆ๋‹ค๋ฉด ํ›ˆ๋ จ ๋‹จ๊ณ„๊ฐ€ ์ ์€ ๋ชจ๋ธ์€ ๋ฐ˜๋“œ์‹œ ์ƒ˜ํ”Œ ํšจ์œจ์ด ๋†’์•„์ง€๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ํ›ˆ๋ จ ๊ณก์„ ์ด ๋” ๊ฐ€ํŒŒ๋ฅด๊ฒŒ ๋Š˜์–ด๋‚˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  MPT-7B๋Š” ์ด ํ•˜์œ„ ํ”Œ๋กฏ์—์„œ ๋‹ค๋ฅธ ๋ชจ๋ธ๋ณด๋‹ค ์•ž์„œ ๊ฐœ์„ ๋œ ๊ฒƒ์œผ๋กœ ๋ˆˆ์— ๋•๋‹ˆ๋‹ค. ์ด๋Š” ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ๊ณผ ํŒ”๋กœ๋งˆ์˜ ๋„๋ฉ”์ธ๊ณผ์˜ ์ผ์น˜(์˜ˆ: MPT๋Š” LLAMA์˜ ๊ฒฝ์šฐ 18%, RedPajama์˜ ๊ฒฝ์šฐ 12.2%, OLMo์˜ ๊ฒฝ์šฐ 11.2%๊ฐ€ ์•„๋‹Œ 27%์˜ ๋น„์ปค๋จผ ํฌ๋กค๋ง ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต) ๋ฐ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ ๊ฒฐ์ •(์˜ˆ: MPT์˜ C4์— ๋Œ€ํ•œ Abbas ๋“ฑ(2023)์˜ ์˜๋ฏธ๋ก ์  ์ค‘๋ณต ์ œ๊ฑฐ ์‚ฌ์šฉ) ๋“ฑ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์š”์ธ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Lu et al. (2023) introduces multilingual dictionaries in prompts to enhance machine translation. Li et al. (2023) proposes chain-of-knowledge (CoK-Li), which obtains structured knowledge from a knowledge base via a query generator to perform knowledge-guided reasoning. Wang et al. (2023) (CoK-Wang) also retrieves structured knowledge from KB. Moreover, it estimates the reasoning chains in terms of factuality and faithfulness and prompts models to rethink unreliable reasonings, which mitigates the knowledge retrieval errors in CoK-Li. KD-CoT Wang et al. (2023) addresses factual reasoning problems through a multi-turn QA approach. They design a feedback-augmented retriever for retrieving relevant external knowledge in each round of QA to calibrate the reasoning process. Other studies use the model's own memory as external knowledge. For example, Memory-of-Thought Li and Qiu (2023) first performs pre-thinking to save the high-confidence thoughts into external memory, and during inference, it lets the LLM recall relevant memory to aid reasoning.
Lu et al.(2023)์€ ๊ธฐ๊ณ„ ๋ฒˆ์—ญ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด prompts์— ๋‹ค๊ตญ์–ด ์‚ฌ์ „์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. Li et al. (2023)์€ ์ฟผ๋ฆฌ ์ƒ์„ฑ๊ธฐ๋ฅผ ํ†ตํ•ด ์ง€์‹ ๋ฒ ์ด์Šค์—์„œ ๊ตฌ์กฐํ™”๋œ ์ง€์‹์„ ํš๋“ํ•˜์—ฌ ์ง€์‹ ๊ธฐ๋ฐ˜ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๋Š” Chain-of-Knowledge(CoK-Li)๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Wang et al. (2023)(CoK-Wang)๋„ KB์—์„œ ๊ตฌ์กฐํ™”๋œ ์ง€์‹์„ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์‚ฌ์‹ค์„ฑ๊ณผ ์ถฉ์‹ค์„ฑ ์ธก๋ฉด์—์„œ ์ถ”๋ก  ์‚ฌ์Šฌ์„ ์ถ”์ •ํ•˜๊ณ  ๋ชจ๋ธ์ด ์‹ ๋ขฐํ•  ์ˆ˜ ์—†๋Š” ์ถ”๋ก ์„ ๋‹ค์‹œ ์ƒ๊ฐํ•˜๋„๋ก promptsํ•˜์—ฌ CoK-Li์˜ ์ง€์‹ ๊ฒ€์ƒ‰ ์˜ค๋ฅ˜๋ฅผ ์™„ํ™”ํ•ฉ๋‹ˆ๋‹ค. KD-CoT Wang et al. (2023)์€ ๋‹ค์ค‘ ํ„ด QA ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ์‚ฌ์‹ค ์ถ”๋ก  ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ QA์˜ ๊ฐ ๋ผ์šด๋“œ์—์„œ ๊ด€๋ จ ์™ธ๋ถ€ ์ง€์‹์„ ๊ฒ€์ƒ‰ํ•˜๊ธฐ ์œ„ํ•ด feedback-augmented retriever๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ ์ถ”๋ก  ๊ณผ์ •์„ ๋ณด์ •ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ชจ๋ธ ์ž์ฒด์˜ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์™ธ๋ถ€ ์ง€์‹์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Memory-of-Thought Li and Qiu (2023)๋Š” ๋จผ์ € ์‚ฌ์ „ ์ƒ๊ฐ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์‹ ๋ขฐ๋„๊ฐ€ ๋†’์€ ์ƒ๊ฐ์„ ์™ธ๋ถ€ ๋ฉ”๋ชจ๋ฆฌ์— ์ €์žฅํ•˜๊ณ , ์ถ”๋ก  ์ค‘์— LLM์ด ๊ด€๋ จ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๋ถˆ๋Ÿฌ์˜ค๋„๋ก ํ•˜์—ฌ ์ถ”๋ก ์„ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค.
Lu et al.(2023)์€ ๊ธฐ๊ณ„ ๋ฒˆ์—ญ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ํ”„๋กฌํ”„ํŠธ์— ๋‹ค๊ตญ์–ด ์‚ฌ์ „์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. Li et al. (2023)์€ ์ฟผ๋ฆฌ ์ƒ์„ฑ๊ธฐ๋ฅผ ํ†ตํ•ด ์ง€์‹ ๋ฒ ์ด์Šค์—์„œ ๊ตฌ์กฐํ™”๋œ ์ง€์‹์„ ํš๋“ํ•˜์—ฌ ์ง€์‹ ๊ธฐ๋ฐ˜ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๋Š” Chain-of-Knowledge(CoK-Li)๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Wang et al. (2023)(CoK-Wang)๋„ KB์—์„œ ๊ตฌ์กฐํ™”๋œ ์ง€์‹์„ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์‚ฌ์‹ค์„ฑ๊ณผ ์ถฉ์‹ค์„ฑ ์ธก๋ฉด์—์„œ ์ถ”๋ก  ์‚ฌ์Šฌ์„ ์ถ”์ •ํ•˜๊ณ  ๋ชจ๋ธ์ด ์‹ ๋ขฐํ•  ์ˆ˜ ์—†๋Š” ์ถ”๋ก ์„ ๋‹ค์‹œ ์ƒ๊ฐํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ CoK-Li์˜ ์ง€์‹ ๊ฒ€์ƒ‰ ์˜ค๋ฅ˜๋ฅผ ์™„ํ™”ํ•ฉ๋‹ˆ๋‹ค. KD-CoT Wang et al. (2023)์€ ๋‹ค์ค‘ ํ„ด QA ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ์‚ฌ์‹ค ์ถ”๋ก  ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ QA์˜ ๊ฐ ๋ผ์šด๋“œ์—์„œ ๊ด€๋ จ ์™ธ๋ถ€ ์ง€์‹์„ ๊ฒ€์ƒ‰ํ•˜๊ธฐ ์œ„ํ•ด ํ”ผ๋“œ๋ฐฑ ์ฆ๊ฐ• ๊ฒ€์ƒ‰๊ธฐ๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ ์ถ”๋ก  ๊ณผ์ •์„ ๋ณด์ •ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ชจ๋ธ ์ž์ฒด์˜ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์™ธ๋ถ€ ์ง€์‹์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Memory-of-Thought Li and Qiu (2023)๋Š” ๋จผ์ € ์‚ฌ์ „ ์ƒ๊ฐ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์‹ ๋ขฐ๋„๊ฐ€ ๋†’์€ ์ƒ๊ฐ์„ ์™ธ๋ถ€ ๋ฉ”๋ชจ๋ฆฌ์— ์ €์žฅํ•˜๊ณ , ์ถ”๋ก  ์ค‘์— LLM์ด ๊ด€๋ จ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๋ถˆ๋Ÿฌ์˜ค๋„๋ก ํ•˜์—ฌ ์ถ”๋ก ์„ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค.
To help inform the development of the necessary fairness, bias, and equity frameworks, it is important to employ a participatory approach to solicit representative views across a wide range of patient demographics, as well as clinical and health equity domain experts. Such evaluation frameworks should be complemented by extensive model red teaming and an adversarial approach to identifying any remaining gaps and failure modes. Recent advances in red teaming LLMs could be useful in this scenario [98, 99, 100, 101]. These practices should not only inform the evaluation of the final model, but also its development and iterative refinement. Model development should follow the established data and model reporting practices and provide transparency into the training data and the associated decision processes [102, 103, 104]. The dialogue research dataset contributing to AMIE training data in our study was de-identified, reducing the availability of socio-economic factors, patient demographics, and information about clinical settings and locations.
ํ•„์š”ํ•œ ๊ณต์ •์„ฑ, ํŽธํ–ฅ์„ฑ ๋ฐ ํ˜•ํ‰์„ฑ ํ”„๋ ˆ์ž„์›Œํฌ์˜ ๊ฐœ๋ฐœ์— ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ํ™˜์ž ์ธ๊ตฌ ํ†ต๊ณ„์™€ ์ž„์ƒ ๋ฐ ๊ฑด๊ฐ• ํ˜•ํ‰์„ฑ ์˜์—ญ ์ „๋ฌธ๊ฐ€๋ฅผ ๋Œ€์ƒ์œผ๋กœ ๋Œ€ํ‘œ์ ์ธ ๊ฒฌํ•ด๋ฅผ ์ˆ˜๋ ดํ•˜๋Š” ์ฐธ์—ฌ์  ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ๋ชจ๋ธ red teaming ๋ฐ adversarial ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ๋‚จ์•„ ์žˆ๋Š” ๊ฒฉ์ฐจ์™€ ์žฅ์•  ๋ชจ๋“œ๋ฅผ ์‹๋ณ„ํ•˜๋Š” ๋ฐ ๋ณด์™„๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์˜ red teaming LLM์˜ ๋ฐœ์ „์€ ์ด ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์œ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[98, 99, 100, 101]. ์ด๋Ÿฌํ•œ ๊ด€ํ–‰์€ ์ตœ์ข… ๋ชจ๋ธ์˜ ํ‰๊ฐ€๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๊ฐœ๋ฐœ ๋ฐ ๋ฐ˜๋ณต์ ์ธ ๊ฐœ์„ ์—๋„ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ ๊ฐœ๋ฐœ์€ ๊ธฐ์กด์˜ ๋ฐ์ดํ„ฐ ๋ฐ ๋ชจ๋ธ ๋ณด๊ณ  ๊ด€ํ–‰์„ ๋”ฐ๋ฅด๊ณ  ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์™€ ๊ด€๋ จ ์˜์‚ฌ ๊ฒฐ์ • ํ”„๋กœ์„ธ์Šค์— ๋Œ€ํ•œ ํˆฌ๋ช…์„ฑ์„ ์ œ๊ณตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค[102, 103, 104]. ์ด ์—ฐ๊ตฌ์—์„œ AMIE ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๊ธฐ์—ฌํ•œ ๋Œ€ํ™” ์—ฐ๊ตฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋น„์‹๋ณ„ ์ฒ˜๋ฆฌ๋˜์–ด ์‚ฌํšŒ๊ฒฝ์ œ์  ์š”์ธ, ํ™˜์ž ์ธ๊ตฌ ํ†ต๊ณ„, ์ž„์ƒ ํ™˜๊ฒฝ ๋ฐ ์œ„์น˜์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐํšŒ๊ฐ€ ์ค„์—ˆ์Šต๋‹ˆ๋‹ค.
ํ•„์š”ํ•œ ๊ณต์ •์„ฑ, ํŽธํ–ฅ์„ฑ ๋ฐ ํ˜•ํ‰์„ฑ ํ”„๋ ˆ์ž„์›Œํฌ์˜ ๊ฐœ๋ฐœ์— ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ํ™˜์ž ์ธ๊ตฌ ํ†ต๊ณ„์™€ ์ž„์ƒ ๋ฐ ๊ฑด๊ฐ• ํ˜•ํ‰์„ฑ ์˜์—ญ ์ „๋ฌธ๊ฐ€๋ฅผ ๋Œ€์ƒ์œผ๋กœ ๋Œ€ํ‘œ์ ์ธ ๊ฒฌํ•ด๋ฅผ ์ˆ˜๋ ดํ•˜๋Š” ์ฐธ์—ฌ์  ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ๋ชจ๋ธ ๋ ˆ๋“œ ํŒ€ ๋ฐ ์ ๋Œ€์  ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ๋‚จ์•„ ์žˆ๋Š” ๊ฒฉ์ฐจ์™€ ์žฅ์•  ๋ชจ๋“œ๋ฅผ ์‹๋ณ„ํ•˜๋Š” ๋ฐ ๋ณด์™„๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์˜ ๋ ˆ๋“œ ํŒ€ LLM์˜ ๋ฐœ์ „์€ ์ด ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์œ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[98, 99, 100, 101]. ์ด๋Ÿฌํ•œ ๊ด€ํ–‰์€ ์ตœ์ข… ๋ชจ๋ธ์˜ ํ‰๊ฐ€๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๊ฐœ๋ฐœ ๋ฐ ๋ฐ˜๋ณต์ ์ธ ๊ฐœ์„ ์—๋„ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ ๊ฐœ๋ฐœ์€ ๊ธฐ์กด์˜ ๋ฐ์ดํ„ฐ ๋ฐ ๋ชจ๋ธ ๋ณด๊ณ  ๊ด€ํ–‰์„ ๋”ฐ๋ฅด๊ณ  ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์™€ ๊ด€๋ จ ์˜์‚ฌ ๊ฒฐ์ • ํ”„๋กœ์„ธ์Šค์— ๋Œ€ํ•œ ํˆฌ๋ช…์„ฑ์„ ์ œ๊ณตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค[102, 103, 104]. ์ด ์—ฐ๊ตฌ์—์„œ AMIE ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๊ธฐ์—ฌํ•œ ๋Œ€ํ™” ์—ฐ๊ตฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋น„์‹๋ณ„ ์ฒ˜๋ฆฌ๋˜์–ด ์‚ฌํšŒ๊ฒฝ์ œ์  ์š”์ธ, ํ™˜์ž ์ธ๊ตฌ ํ†ต๊ณ„, ์ž„์ƒ ํ™˜๊ฒฝ ๋ฐ ์œ„์น˜์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐํšŒ๊ฐ€ ์ค„์—ˆ์Šต๋‹ˆ๋‹ค.
**Towards Reference-Free Metrics** Neural-based, reference-dependent metrics like COMET-22 demonstrate greater consistency with reference-free metrics and robustness compared to BLEU. For instance, with COMET-22, our models show significant improvements like other reference-free models over ALMA-13B-LoRA and comparable performance to GPT-4, e.g., 87.74 (Ours) vs. 87.68 (GPT-4) when en\(\rightarrow\)xx. However, it is important to note that, according to reference-free metrics, gold references are often inferior to system-generated translations, potentially indicating quality issues in the references that could impact COMET-22 evaluations. Consequently, inconsistencies still exist between COMET-22 and reference-free models like XCOMET. For example, XCOMET rates ALMA-R model on average higher than WMT winners (89.11 vs. 87.13), while COMET-22 favors WMT winners (85.21 vs. 85.60). In line with the recommendations in Freitag et al. (2023), we advocate for the use of reference-free models to circumvent the potential quality issues of references.
**์ฐธ์กฐ ์—†๋Š” ๋ฉ”ํŠธ๋ฆญ์„ ํ–ฅํ•˜์—ฌ** COMET-22์™€ ๊ฐ™์€ ์‹ ๊ฒฝ ๊ธฐ๋ฐ˜์˜ ์ฐธ์กฐ ์˜์กด ๋ฉ”ํŠธ๋ฆญ์€ BLEU์— ๋น„ํ•ด ์ฐธ์กฐ ์—†๋Š” ๋ฉ”ํŠธ๋ฆญ๊ณผ ๋” ํฐ ์ผ๊ด€์„ฑ๊ณผ robustness๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, COMET-22๋ฅผ ์‚ฌ์šฉํ•œ ๋ชจ๋ธ์€ ALMA-13B-LoRA์— ๋น„ํ•ด ๋‹ค๋ฅธ ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ƒ๋‹นํ•œ ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ๋ณด์˜€์œผ๋ฉฐ, en\(\rightarrow\)xx์ผ ๋•Œ 87.74(์šฐ๋ฆฌ) ๋Œ€ 87.68(GPT-4) ๋“ฑ GPT-4์™€ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฐธ์กฐ ์—†๋Š” ๋ฉ”ํŠธ๋ฆญ์— ๋”ฐ๋ฅด๋ฉด ๊ณจ๋“œ ์ฐธ์กฐ๊ฐ€ ์‹œ์Šคํ…œ ์ƒ์„ฑ ๋ฒˆ์—ญ๋ณด๋‹ค ์—ด๋“ฑํ•œ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ, ์ด๋Š” COMET-22 ํ‰๊ฐ€์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ์ฐธ์กฐ์˜ ํ’ˆ์งˆ ๋ฌธ์ œ๋ฅผ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ COMET-22์™€ XCOMET๊ณผ ๊ฐ™์€ ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ ๊ฐ„์—๋Š” ์—ฌ์ „ํžˆ ๋ถˆ์ผ์น˜๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, XCOMET์€ ํ‰๊ท ์ ์œผ๋กœ ALMA-R ๋ชจ๋ธ์„ WMT ์ˆ˜์ƒ์ž๋ณด๋‹ค ๋†’๊ฒŒ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค(89.11 ๋Œ€ 87.13), ๋ฐ˜๋ฉด COMET-22๋Š” WMT ์ˆ˜์ƒ์ž๋ฅผ ์„ ํ˜ธํ•ฉ๋‹ˆ๋‹ค(85.21 ๋Œ€ 85.60). Freitag ๋“ฑ(2023)์˜ ๊ถŒ์žฅ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ์ฐธ์กฐ์˜ ์ž ์žฌ์ ์ธ ํ’ˆ์งˆ ๋ฌธ์ œ๋ฅผ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•  ๊ฒƒ์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค.
**์ฐธ์กฐ ์—†๋Š” ๋ฉ”ํŠธ๋ฆญ์„ ํ–ฅํ•˜์—ฌ** COMET-22์™€ ๊ฐ™์€ ์‹ ๊ฒฝ ๊ธฐ๋ฐ˜์˜ ์ฐธ์กฐ ์˜์กด ๋ฉ”ํŠธ๋ฆญ์€ BLEU์— ๋น„ํ•ด ์ฐธ์กฐ ์—†๋Š” ๋ฉ”ํŠธ๋ฆญ๊ณผ ๋” ํฐ ์ผ๊ด€์„ฑ๊ณผ ๊ฐ•๊ฑด์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, COMET-22๋ฅผ ์‚ฌ์šฉํ•œ ๋ชจ๋ธ์€ ALMA-13B-LoRA์— ๋น„ํ•ด ๋‹ค๋ฅธ ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ƒ๋‹นํ•œ ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ๋ณด์˜€์œผ๋ฉฐ, en\(\rightarrow\)xx์ผ ๋•Œ 87.74(์šฐ๋ฆฌ) ๋Œ€ 87.68(GPT-4) ๋“ฑ GPT-4์™€ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฐธ์กฐ ์—†๋Š” ๋ฉ”ํŠธ๋ฆญ์— ๋”ฐ๋ฅด๋ฉด ๊ณจ๋“œ ์ฐธ์กฐ๊ฐ€ ์‹œ์Šคํ…œ ์ƒ์„ฑ ๋ฒˆ์—ญ๋ณด๋‹ค ์—ด๋“ฑํ•œ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ, ์ด๋Š” COMET-22 ํ‰๊ฐ€์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ์ฐธ์กฐ์˜ ํ’ˆ์งˆ ๋ฌธ์ œ๋ฅผ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ COMET-22์™€ XCOMET๊ณผ ๊ฐ™์€ ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ ๊ฐ„์—๋Š” ์—ฌ์ „ํžˆ ๋ถˆ์ผ์น˜๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, XCOMET์€ ํ‰๊ท ์ ์œผ๋กœ ALMA-R ๋ชจ๋ธ์„ WMT ์ˆ˜์ƒ์ž๋ณด๋‹ค ๋†’๊ฒŒ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค(89.11 ๋Œ€ 87.13), ๋ฐ˜๋ฉด COMET-22๋Š” WMT ์ˆ˜์ƒ์ž๋ฅผ ์„ ํ˜ธํ•ฉ๋‹ˆ๋‹ค(85.21 ๋Œ€ 85.60). Freitag ๋“ฑ(2023)์˜ ๊ถŒ์žฅ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ์ฐธ์กฐ์˜ ์ž ์žฌ์ ์ธ ํ’ˆ์งˆ ๋ฌธ์ œ๋ฅผ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•  ๊ฒƒ์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค.
Meanwhile, the use of synthetic data has gained prominence in the paradigm of aligning pre-trained LLMs via instruction fine-tuning, RLHF (Ouyang et al., 2022), and instruction backtranslation (Li et al., 2023b). Recently, in the context of pre-training, synthetic datawas used to generate datasets such as Tiny Stories (Eldan & Li, 2023) and Textbook quality synthetic data (Gunasekar et al., 2023; Li et al., 2023c). These were used to train smaller language models (like the Phi model family) that were as performant as larger language models on certain tasks. However, their data generation process stays largely opaque, and prohibitively expensive, requiring prompting a GPT-3.5 model for generating billions of tokens. Additionally, such data generation can create a large "knowledge bias" by specifically generating data pertaining to tasks that we want to perform well on. While synthetic data has shown promise, it is unclear if this is because of the higher quality nature of synthetic data, or because of strategic topic selection (Maini, 2023).
ํ•œํŽธ, Instruction Fine-tuning, RLHF(Ouyang et al., 2022), Instruction backtranslation(Li et al., 2023b)์„ ํ†ตํ•ด ์‚ฌ์ „ ํ•™์Šต๋œ LLM์„ ์กฐ์ •ํ•˜๋Š” ํŒจ๋Ÿฌ๋‹ค์ž„์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์‚ฌ์šฉ์ด ์ฃผ๋ชฉ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” ์‚ฌ์ „ ํ•™์Šต์˜ ๋งฅ๋ฝ์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Tiny Stories(Eldan & Li, 2023) ๋ฐ Textbook quality synthetic data(Gunasekar et al., 2023; Li et al., 2023c)์™€ ๊ฐ™์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋Š” ํŠน์ • ์ž‘์—…์—์„œ ๋” ํฐ ์–ธ์–ด ๋ชจ๋ธ๋งŒํผ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ ๋” ์ž‘์€ ์–ธ์–ด ๋ชจ๋ธ(Phi ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ๊ณผ ๊ฐ™์€)์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋Š” ์—ฌ์ „ํžˆ ๋ถˆํˆฌ๋ช…ํ•˜๊ณ  ๋ง‰๋Œ€ํ•œ ๋น„์šฉ์ด ์†Œ์š”๋˜์–ด ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ํ† ํฐ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด GPT-3.5 ๋ชจ๋ธ์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋‚ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์€ ํŠน์ • ์ž‘์—…๊ณผ ๊ด€๋ จ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•จ์œผ๋กœ์จ ํฐ '์ง€์‹ ํŽธํ–ฅ'์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ, ์ด๊ฒƒ์ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์ด ๋” ๋†’๊ธฐ ๋•Œ๋ฌธ์ธ์ง€ ์•„๋‹ˆ๋ฉด ์ „๋žต์ ์ธ ์ฃผ์ œ ์„ ํƒ ๋•Œ๋ฌธ์ธ์ง€๋Š” ํ™•์‹คํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค(Maini, 2023).
ํ•œํŽธ, ์ง€์นจ ๋ฏธ์„ธ ์กฐ์ •, RLHF(Ouyang et al., 2022), ์ง€์นจ ์—ญ๋ฒˆ์—ญ(Li et al., 2023b)์„ ํ†ตํ•ด ์‚ฌ์ „ ํ•™์Šต๋œ LLM์„ ์กฐ์ •ํ•˜๋Š” ํŒจ๋Ÿฌ๋‹ค์ž„์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์‚ฌ์šฉ์ด ์ฃผ๋ชฉ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ์—๋Š” ์‚ฌ์ „ ํ•™์Šต์˜ ๋งฅ๋ฝ์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Tiny Stories(Eldan & Li, 2023) ๋ฐ ํ…์ŠคํŠธ๋ถ ํ’ˆ์งˆ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ(Gunasekar et al., 2023; Li et al., 2023c)์™€ ๊ฐ™์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋Š” ํŠน์ • ์ž‘์—…์—์„œ ๋” ํฐ ์–ธ์–ด ๋ชจ๋ธ๋งŒํผ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ ๋” ์ž‘์€ ์–ธ์–ด ๋ชจ๋ธ(Phi ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ๊ณผ ๊ฐ™์€)์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋Š” ์—ฌ์ „ํžˆ ๋ถˆํˆฌ๋ช…ํ•˜๊ณ  ๋ง‰๋Œ€ํ•œ ๋น„์šฉ์ด ์†Œ์š”๋˜์–ด ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ํ† ํฐ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด GPT-3.5 ๋ชจ๋ธ์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋‚ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์€ ํŠน์ • ์ž‘์—…๊ณผ ๊ด€๋ จ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•จ์œผ๋กœ์จ ํฐ '์ง€์‹ ํŽธํ–ฅ'์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ, ์ด๊ฒƒ์ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์ด ๋” ๋†’๊ธฐ ๋•Œ๋ฌธ์ธ์ง€ ์•„๋‹ˆ๋ฉด ์ „๋žต์ ์ธ ์ฃผ์ œ ์„ ํƒ ๋•Œ๋ฌธ์ธ์ง€๋Š” ํ™•์‹คํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค(Maini, 2023).
**Adapter Tuning**. Adapter tuning incorporates small neural network modules (called _adapter_) into the Transformer models [398]. To implement the adapter module, a bottleneck architecture has been proposed in [398, 399], which first compresses the original feature vector into a smaller dimension (followed by a nonlinear transformation) and then recovers it to the original dimension. The adapter modules would be integrated into each Transformer layer, typically using a serial insertion after each of the two core parts (_i.e.,_ attention layer and feed-forward layer) of a Transformer layer. Alternatively, parallel adapters [400] can be also used in Transformer layers, where it places two adapter modules in parallel with the attention layer and feed-forward layer accordingly. During fine-tuning, the adapter modules would be optimized according to the specific task goals, while the parameters of the original language model are frozen in this process. In this way, we can effectively reduce the number of trainable parameters during fine-tuning.
**Adapter Tuning**. Adapter Tuning์€ Transformer ๋ชจ๋ธ์— _adapter_๋ผ๊ณ  ํ•˜๋Š” ์ž‘์€ ์‹ ๊ฒฝ๋ง ๋ชจ๋“ˆ์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค[398]. adapter ๋ชจ๋“ˆ์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด [398, 399]์—์„œ bottleneck architecture๊ฐ€ ์ œ์•ˆ๋˜์—ˆ๋Š”๋ฐ, ์ด๋Š” ๋จผ์ € ์›๋ณธ ํŠน์ง• ๋ฒกํ„ฐ๋ฅผ ๋” ์ž‘์€ ์ฐจ์›์œผ๋กœ ์••์ถ•ํ•œ ๋‹ค์Œ(๋น„์„ ํ˜• ๋ณ€ํ™˜ ํ›„) ์›๋ž˜ ์ฐจ์›์œผ๋กœ ๋ณต์›ํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. adapter ๋ชจ๋“ˆ์€ ์ผ๋ฐ˜์ ์œผ๋กœ Transformer layer์˜ ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ๋ถ€๋ถ„(_์ฆ‰, attention layer ๋ฐ feed-forward layer) ๊ฐ๊ฐ์— ์ง๋ ฌ๋กœ ์‚ฝ์ž…ํ•˜์—ฌ Transformer layer์— ํ†ตํ•ฉ๋ฉ๋‹ˆ๋‹ค. ๋˜๋Š” Transformer layer์— parallel adapters[400]๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜๋„ ์žˆ๋Š”๋ฐ, ์ด ๊ฒฝ์šฐ ๋‘ ๊ฐœ์˜ adapter ๋ชจ๋“ˆ์„ attention layer ๋ฐ feed-forward layer์™€ ๋ณ‘๋ ฌ๋กœ ๋ฐฐ์น˜ํ•ฉ๋‹ˆ๋‹ค. Fine-tuningํ•˜๋Š” ๋™์•ˆ adapter ๋ชจ๋“ˆ์€ ํŠน์ • ์ž‘์—… ๋ชฉํ‘œ์— ๋”ฐ๋ผ ์ตœ์ ํ™”๋˜๋ฉฐ, ์ด ๊ณผ์ •์—์„œ ์›๋ž˜ ์–ธ์–ด ๋ชจ๋ธ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ๋™๊ฒฐ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ Fine-tuning ์ค‘์— ํ›ˆ๋ จ ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ์ˆ˜๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์–ด๋Œ‘ํ„ฐ ํŠœ๋‹**. ์–ด๋Œ‘ํ„ฐ ํŠœ๋‹์€ ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ์— _์–ด๋Œ‘ํ„ฐ_๋ผ๊ณ  ํ•˜๋Š” ์ž‘์€ ์‹ ๊ฒฝ๋ง ๋ชจ๋“ˆ์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค[398]. ์–ด๋Œ‘ํ„ฐ ๋ชจ๋“ˆ์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด [398, 399]์—์„œ ๋ณ‘๋ชฉ ์•„ํ‚คํ…์ฒ˜๊ฐ€ ์ œ์•ˆ๋˜์—ˆ๋Š”๋ฐ, ์ด๋Š” ๋จผ์ € ์›๋ณธ ํŠน์ง• ๋ฒกํ„ฐ๋ฅผ ๋” ์ž‘์€ ์ฐจ์›์œผ๋กœ ์••์ถ•ํ•œ ๋‹ค์Œ(๋น„์„ ํ˜• ๋ณ€ํ™˜ ํ›„) ์›๋ž˜ ์ฐจ์›์œผ๋กœ ๋ณต์›ํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ์–ด๋Œ‘ํ„ฐ ๋ชจ๋“ˆ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํŠธ๋žœ์Šคํฌ๋จธ ๋ ˆ์ด์–ด์˜ ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ๋ถ€๋ถ„(_์ฆ‰, ์ฃผ์˜๋„ ๋ ˆ์ด์–ด ๋ฐ ํ”ผ๋“œํฌ์›Œ๋“œ ๋ ˆ์ด์–ด) ๊ฐ๊ฐ์— ์ง๋ ฌ๋กœ ์‚ฝ์ž…ํ•˜์—ฌ ํŠธ๋žœ์Šคํฌ๋จธ ๋ ˆ์ด์–ด์— ํ†ตํ•ฉ๋ฉ๋‹ˆ๋‹ค. ๋˜๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ๋ ˆ์ด์–ด์— ๋ณ‘๋ ฌ ์–ด๋Œ‘ํ„ฐ[400]๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜๋„ ์žˆ๋Š”๋ฐ, ์ด ๊ฒฝ์šฐ ๋‘ ๊ฐœ์˜ ์–ด๋Œ‘ํ„ฐ ๋ชจ๋“ˆ์„ ์ฃผ์˜๋„ ๋ ˆ์ด์–ด ๋ฐ ํ”ผ๋“œํฌ์›Œ๋“œ ๋ ˆ์ด์–ด์™€ ๋ณ‘๋ ฌ๋กœ ๋ฐฐ์น˜ํ•ฉ๋‹ˆ๋‹ค. ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋™์•ˆ ์–ด๋Œ‘ํ„ฐ ๋ชจ๋“ˆ์€ ํŠน์ • ์ž‘์—… ๋ชฉํ‘œ์— ๋”ฐ๋ผ ์ตœ์ ํ™”๋˜๋ฉฐ, ์ด ๊ณผ์ •์—์„œ ์›๋ž˜ ์–ธ์–ด ๋ชจ๋ธ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ๋™๊ฒฐ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ๋ฏธ์„ธ ์กฐ์ • ์ค‘์— ํ›ˆ๋ จ ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ์ˆ˜๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Code-based Approaches.** Although text-based approaches sound intuitive, they cannot guarantee faithful execution of the plan, which may lead to failure even when the plan is sound. To address this issue, code-based approaches have been proposed to generate more verifiable plans in the form of executable code in programming languages, _e.g.,_ Python or PDDL. In this way, LLMs are first prompted to generate the program and then utilize a deterministic solver to execute it. For example, FaithVolCt [442] and PAL [443] decompose a reasoning task into two stages: at the first stage, the LLM generates a plan conditioned on the query; at the second stage, a deterministic solver executes the plan to derive the final answer. Furthermore, code-based approaches can be applied to embodied agents in a similar way. For example, PROGPROMPT [530] and LLM+P [531] first utilize LLMs to generate plans in the form of python functions or PDDL files, and then leverage a virtual agent or classical planner to solve the problem according to the code-based plans.
**Code-based Approaches**. ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ง๊ด€์ ์œผ๋กœ ๋“ค๋ฆฌ์ง€๋งŒ ๊ณ„ํš์˜ ์ถฉ์‹คํ•œ ์‹คํ–‰์„ ๋ณด์žฅํ•  ์ˆ˜ ์—†์œผ๋ฏ€๋กœ ๊ณ„ํš์ด ๊ฑด์ „ํ•œ ๊ฒฝ์šฐ์—๋„ ์‹คํŒจ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด(์˜ˆ: Python ๋˜๋Š” PDDL)๋กœ ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ์ฝ”๋“œ ํ˜•ํƒœ์˜ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๋Š” ์ฝ”๋“œ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ LLMs์€ ๋จผ์ € ํ”„๋กœ๊ทธ๋žจ์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋‚ธ ๋‹ค์Œ ๊ฒฐ์ •๋ก ์  ์†”๋ฒ„๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, FaithVolCt [442] ๋ฐ PAL [443]์€ ์ถ”๋ก  ์ž‘์—…์„ ๋‘ ๋‹จ๊ณ„๋กœ ๋ถ„ํ•ดํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” LLM์ด ์ฟผ๋ฆฌ์— ๋”ฐ๋ผ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๊ณ , ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” ๊ฒฐ์ •๋ก ์  ์†”๋ฒ„๊ฐ€ ๊ณ„ํš์„ ์‹คํ–‰ํ•˜์—ฌ ์ตœ์ข… ๋‹ต์„ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ฝ”๋“œ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์œ ์‚ฌํ•œ ๋ฐฉ์‹์œผ๋กœ ๊ตฌํ˜„๋œ ์—์ด์ „ํŠธ์—๋„ ์ ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, PROGPROMPT [530] ๋ฐ LLM+P [531]์€ ๋จผ์ € LLMs์„ ํ™œ์šฉํ•˜์—ฌ Python ํ•จ์ˆ˜ ๋˜๋Š” PDDL ํŒŒ์ผ ํ˜•ํƒœ์˜ ๊ณ„ํš์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ๊ฐ€์ƒ ์—์ด์ „ํŠธ ๋˜๋Š” ๊ณ ์ „์  ํ”Œ๋ž˜๋„ˆ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ฝ”๋“œ ๊ธฐ๋ฐ˜ ๊ณ„ํš์— ๋”ฐ๋ผ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค.
**์ฝ”๋“œ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹**. ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ง๊ด€์ ์œผ๋กœ ๋“ค๋ฆฌ์ง€๋งŒ ๊ณ„ํš์˜ ์ถฉ์‹คํ•œ ์‹คํ–‰์„ ๋ณด์žฅํ•  ์ˆ˜ ์—†์œผ๋ฏ€๋กœ ๊ณ„ํš์ด ๊ฑด์ „ํ•œ ๊ฒฝ์šฐ์—๋„ ์‹คํŒจ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด(์˜ˆ: Python ๋˜๋Š” PDDL)๋กœ ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ์ฝ”๋“œ ํ˜•ํƒœ์˜ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๋Š” ์ฝ”๋“œ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ LLM์€ ๋จผ์ € ํ”„๋กœ๊ทธ๋žจ์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋‚ธ ๋‹ค์Œ ๊ฒฐ์ •๋ก ์  ์†”๋ฒ„๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, FaithVolCt [442] ๋ฐ PAL [443]์€ ์ถ”๋ก  ์ž‘์—…์„ ๋‘ ๋‹จ๊ณ„๋กœ ๋ถ„ํ•ดํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” LLM์ด ์ฟผ๋ฆฌ์— ๋”ฐ๋ผ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๊ณ , ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” ๊ฒฐ์ •๋ก ์  ์†”๋ฒ„๊ฐ€ ๊ณ„ํš์„ ์‹คํ–‰ํ•˜์—ฌ ์ตœ์ข… ๋‹ต์„ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ฝ”๋“œ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์œ ์‚ฌํ•œ ๋ฐฉ์‹์œผ๋กœ ๊ตฌํ˜„๋œ ์—์ด์ „ํŠธ์—๋„ ์ ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, PROGPROMPT [530] ๋ฐ LLM+P [531]์€ ๋จผ์ € LLM์„ ํ™œ์šฉํ•˜์—ฌ ํŒŒ์ด์ฌ ํ•จ์ˆ˜ ๋˜๋Š” PDDL ํŒŒ์ผ ํ˜•ํƒœ์˜ ๊ณ„ํš์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ๊ฐ€์ƒ ์—์ด์ „ํŠธ ๋˜๋Š” ๊ณ ์ „์  ํ”Œ๋ž˜๋„ˆ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ฝ”๋“œ ๊ธฐ๋ฐ˜ ๊ณ„ํš์— ๋”ฐ๋ผ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค.
The amount of knowledge elicited through the standard instruction-tuning is still limited, even though the perplexity of documents is minimized, a phenomenon we refer to as the "perplexity curse". Our next question is how can we improve the ability of LLMs to absorb knowledge from documents to mitigate the perplexity curse. The main challenge is the gap between the way knowledge is presented in raw documents and how it is accessed through question-answering. We found that QA pairs are generally straightforward, while documents tend to be more complex and cluttered, weaving many factual statements together in a more intricate manner. Using Fig. 3 as an example, the answer to the question "who handled the editing of Oppenheimer" is included in a sentence in the middle of the article "Editing was handled by Jennifer Lame...", which does not explicitly mention "Oppenheimer". During training, LLMs must understand the context and deduce that "editing" refers to "the editing of Oppenheimer" to effectively encode this knowledge in the parameters.
ํ‘œ์ค€ Instruction Tuning์„ ํ†ตํ•ด ์ด๋Œ์–ด๋‚ด๋Š” ์ง€์‹์˜ ์–‘์€ ์—ฌ์ „ํžˆ ์ œํ•œ์ ์ด๋ฉฐ, ๋ฌธ์„œ์˜ Perplexity๊ฐ€ ์ตœ์†Œํ™”๋˜๋”๋ผ๋„ 'Perplexity Curse'๋ผ๊ณ  ๋ถ€๋ฅด๋Š” ํ˜„์ƒ์ด ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ ์งˆ๋ฌธ์€ Perplexity Curse๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋ฌธ์„œ์—์„œ ์ง€์‹์„ ํก์ˆ˜ํ•˜๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ์–ด๋–ป๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š”๊ฐ€ ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ฐ€์žฅ ํฐ ๊ณผ์ œ๋Š” ์›์‹œ ๋ฌธ์„œ์— ์ง€์‹์ด ์ œ์‹œ๋˜๋Š” ๋ฐฉ์‹๊ณผ Question-Answering์„ ํ†ตํ•ด ์•ก์„ธ์Šค๋˜๋Š” ๋ฐฉ์‹ ์‚ฌ์ด์˜ ๊ฒฉ์ฐจ์ž…๋‹ˆ๋‹ค. Question-Answering ์Œ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐ„๋‹จํ•œ ๋ฐ˜๋ฉด, ๋ฌธ์„œ๋Š” ๋” ๋ณต์žกํ•˜๊ณ  ์–ด์ˆ˜์„ ํ•œ ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฉฐ, ๋งŽ์€ ์‚ฌ์‹ค ์ง„์ˆ ์„ ๋” ๋ณต์žกํ•œ ๋ฐฉ์‹์œผ๋กœ ์—ฎ์–ด๋‚ด๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3์„ ์˜ˆ๋กœ ๋“ค์–ด๋ณด๋ฉด, "์˜คํŽœํ•˜์ด๋จธ์˜ ํŽธ์ง‘์„ ๋ˆ„๊ฐ€ ์ฒ˜๋ฆฌํ–ˆ๋‚˜์š”?"๋ผ๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต์€ "ํŽธ์ง‘์€ ์ œ๋‹ˆํผ ๋ ˆ์ž„์ด ์ฒ˜๋ฆฌํ–ˆ์Šต๋‹ˆ๋‹ค..."๋ผ๋Š” ๊ธฐ์‚ฌ ์ค‘๊ฐ„์— ์žˆ๋Š” ๋ฌธ์žฅ์— ํฌํ•จ๋˜์–ด ์žˆ๋Š”๋ฐ, ์—ฌ๊ธฐ์—๋Š” "์˜คํŽœํ•˜์ด๋จธ"๊ฐ€ ๋ช…์‹œ์ ์œผ๋กœ ์–ธ๊ธ‰๋˜์–ด ์žˆ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ•™์Šตํ•˜๋Š” ๋™์•ˆ LLM์€ ๋ฌธ๋งฅ์„ ์ดํ•ดํ•˜๊ณ  'ํŽธ์ง‘'์ด '์˜คํŽœํ•˜์ด๋จธ์˜ ํŽธ์ง‘'์„ ์˜๋ฏธํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์ถ”๋ก ํ•˜์—ฌ ์ด ์ง€์‹์„ ํŒŒ๋ผ๋ฏธํ„ฐ์— ํšจ๊ณผ์ ์œผ๋กœ ์ธ์ฝ”๋”ฉํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
ํ‘œ์ค€ ์ง€์นจ ํŠœ๋‹์„ ํ†ตํ•ด ์ด๋Œ์–ด๋‚ด๋Š” ์ง€์‹์˜ ์–‘์€ ์—ฌ์ „ํžˆ ์ œํ•œ์ ์ด๋ฉฐ, ๋ฌธ์„œ์˜ ๋‹นํ˜น๊ฐ์ด ์ตœ์†Œํ™”๋˜๋”๋ผ๋„ '๋‹นํ˜น๊ฐ์˜ ์ €์ฃผ'๋ผ๊ณ  ๋ถ€๋ฅด๋Š” ํ˜„์ƒ์ด ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ ์งˆ๋ฌธ์€ ๋‹นํ˜น๊ฐ์˜ ์ €์ฃผ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋ฌธ์„œ์—์„œ ์ง€์‹์„ ํก์ˆ˜ํ•˜๋Š” LLM์˜ ๋Šฅ๋ ฅ์„ ์–ด๋–ป๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š”๊ฐ€ ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ฐ€์žฅ ํฐ ๊ณผ์ œ๋Š” ์›์‹œ ๋ฌธ์„œ์— ์ง€์‹์ด ์ œ์‹œ๋˜๋Š” ๋ฐฉ์‹๊ณผ ์งˆ๋ฌธ-๋‹ต๋ณ€์„ ํ†ตํ•ด ์•ก์„ธ์Šค๋˜๋Š” ๋ฐฉ์‹ ์‚ฌ์ด์˜ ๊ฒฉ์ฐจ์ž…๋‹ˆ๋‹ค. ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€ ์Œ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐ„๋‹จํ•œ ๋ฐ˜๋ฉด, ๋ฌธ์„œ๋Š” ๋” ๋ณต์žกํ•˜๊ณ  ์–ด์ˆ˜์„ ํ•œ ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฉฐ, ๋งŽ์€ ์‚ฌ์‹ค ์ง„์ˆ ์„ ๋” ๋ณต์žกํ•œ ๋ฐฉ์‹์œผ๋กœ ์—ฎ์–ด๋‚ด๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3์„ ์˜ˆ๋กœ ๋“ค์–ด๋ณด๋ฉด, "์˜คํŽœํ•˜์ด๋จธ์˜ ํŽธ์ง‘์„ ๋ˆ„๊ฐ€ ์ฒ˜๋ฆฌํ–ˆ๋‚˜์š”?"๋ผ๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต์€ "ํŽธ์ง‘์€ ์ œ๋‹ˆํผ ๋ ˆ์ž„์ด ์ฒ˜๋ฆฌํ–ˆ์Šต๋‹ˆ๋‹ค..."๋ผ๋Š” ๊ธฐ์‚ฌ ์ค‘๊ฐ„์— ์žˆ๋Š” ๋ฌธ์žฅ์— ํฌํ•จ๋˜์–ด ์žˆ๋Š”๋ฐ, ์—ฌ๊ธฐ์—๋Š” "์˜คํŽœํ•˜์ด๋จธ"๊ฐ€ ๋ช…์‹œ์ ์œผ๋กœ ์–ธ๊ธ‰๋˜์–ด ์žˆ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ•™์Šตํ•˜๋Š” ๋™์•ˆ LLM์€ ๋ฌธ๋งฅ์„ ์ดํ•ดํ•˜๊ณ  'ํŽธ์ง‘'์ด '์˜คํŽœํ•˜์ด๋จธ์˜ ํŽธ์ง‘'์„ ์˜๋ฏธํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์ถ”๋ก ํ•˜์—ฌ ์ด ์ง€์‹์„ ํŒŒ๋ผ๋ฏธํ„ฐ์— ํšจ๊ณผ์ ์œผ๋กœ ์ธ์ฝ”๋”ฉํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
**ZeRO.** ZeRO [333] technique, proposed by the DeepSpeed [74] library, focuses on the issue of memory redundancy in data parallelism. As mentioned before, data parallelism requires each GPU to store the same copy of a LLM, including model parameters, model gradients, and optimizer parameters. Whereas, not all of the above data is necessary to be retained on each GPU, which would cause a memory redundancy problem. To resolve it, the ZeRO technique aims to retain only a fraction of data on each GPU, while the rest data can be retrieved from other GPUs when required. Specifically, ZeRO provides three solutions, depending on how the three parts of the data are stored, namely optimizer state partitioning, gradient partitioning, and parameter partitioning. Empirical results indicate that the first two solutions do not increase the communication overhead, and the third solution increases about 50% communication overhead but saves memory proportional to the number of GPUs. PyTorch has implemented a similar technique as ZeRO, called FSDP [339].
**ZeRO.** ZeRO [333] ๊ธฐ์ˆ ์€ DeepSpeed [74] ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์—์„œ ์ œ์•ˆํ•œ ๊ฒƒ์œผ๋กœ, ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ์—์„œ ๋ฉ”๋ชจ๋ฆฌ ์ค‘๋ณต ๋ฌธ์ œ์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์•ž์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•ด์„œ๋Š” ๊ฐ GPU๊ฐ€ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ, ๋ชจ๋ธ ๊ทธ๋ผ๋””์–ธํŠธ, Optimizer ํŒŒ๋ผ๋ฏธํ„ฐ ๋“ฑ LLM์˜ ๋™์ผํ•œ ๋ณต์‚ฌ๋ณธ์„ ์ €์žฅํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์œ„์˜ ๋ชจ๋“  ๋ฐ์ดํ„ฐ๊ฐ€ ๊ฐ GPU์— ์œ ์ง€๋˜์–ด์•ผ ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ด๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ค‘๋ณต ๋ฌธ์ œ๋ฅผ ์ผ์œผํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ZeRO ๊ธฐ์ˆ ์€ ๊ฐ GPU์— ์ผ๋ถ€ ๋ฐ์ดํ„ฐ๋งŒ ์œ ์ง€ํ•˜๊ณ  ๋‚˜๋จธ์ง€ ๋ฐ์ดํ„ฐ๋Š” ํ•„์š”ํ•  ๋•Œ ๋‹ค๋ฅธ GPU์—์„œ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ZeRO๋Š” ๋ฐ์ดํ„ฐ์˜ ์„ธ ๋ถ€๋ถ„์„ ์ €์žฅํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋”ฐ๋ผ Optimizer ์ƒํƒœ ํŒŒํ‹ฐ์…”๋‹, Gradient ํŒŒํ‹ฐ์…”๋‹, ํŒŒ๋ผ๋ฏธํ„ฐ ํŒŒํ‹ฐ์…”๋‹ ๋“ฑ ์„ธ ๊ฐ€์ง€ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์ฒ˜์Œ ๋‘ ๊ฐ€์ง€ ์†”๋ฃจ์…˜์€ ํ†ต์‹  ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ์ฆ๊ฐ€์‹œํ‚ค์ง€ ์•Š์œผ๋ฉฐ, ์„ธ ๋ฒˆ์งธ ์†”๋ฃจ์…˜์€ ์•ฝ 50%์˜ ํ†ต์‹  ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ์ฆ๊ฐ€์‹œํ‚ค์ง€๋งŒ GPU ์ˆ˜์— ๋น„๋ก€ํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ ˆ์•ฝํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. PyTorch๋Š” ZeRO์™€ ์œ ์‚ฌํ•œ ๊ธฐ์ˆ ์ธ FSDP [339]๋ฅผ ๊ตฌํ˜„ํ–ˆ์Šต๋‹ˆ๋‹ค.
**ZeRO.** ZeRO [333] ๊ธฐ์ˆ ์€ DeepSpeed [74] ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์—์„œ ์ œ์•ˆํ•œ ๊ฒƒ์œผ๋กœ, ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ์—์„œ ๋ฉ”๋ชจ๋ฆฌ ์ค‘๋ณต ๋ฌธ์ œ์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์•ž์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•ด์„œ๋Š” ๊ฐ GPU๊ฐ€ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ, ๋ชจ๋ธ ๊ทธ๋ผ๋””์–ธํŠธ, ์ตœ์ ํ™” ํŒŒ๋ผ๋ฏธํ„ฐ ๋“ฑ LLM์˜ ๋™์ผํ•œ ๋ณต์‚ฌ๋ณธ์„ ์ €์žฅํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์œ„์˜ ๋ชจ๋“  ๋ฐ์ดํ„ฐ๊ฐ€ ๊ฐ GPU์— ์œ ์ง€๋˜์–ด์•ผ ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ด๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ค‘๋ณต ๋ฌธ์ œ๋ฅผ ์ผ์œผํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ZeRO ๊ธฐ์ˆ ์€ ๊ฐ GPU์— ์ผ๋ถ€ ๋ฐ์ดํ„ฐ๋งŒ ์œ ์ง€ํ•˜๊ณ  ๋‚˜๋จธ์ง€ ๋ฐ์ดํ„ฐ๋Š” ํ•„์š”ํ•  ๋•Œ ๋‹ค๋ฅธ GPU์—์„œ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ZeRO๋Š” ๋ฐ์ดํ„ฐ์˜ ์„ธ ๋ถ€๋ถ„์„ ์ €์žฅํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋”ฐ๋ผ ์ตœ์ ํ™” ์ƒํƒœ ํŒŒํ‹ฐ์…”๋‹, ๊ทธ๋ผ๋””์–ธํŠธ ํŒŒํ‹ฐ์…”๋‹, ํŒŒ๋ผ๋ฏธํ„ฐ ํŒŒํ‹ฐ์…”๋‹ ๋“ฑ ์„ธ ๊ฐ€์ง€ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์ฒ˜์Œ ๋‘ ๊ฐ€์ง€ ์†”๋ฃจ์…˜์€ ํ†ต์‹  ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ์ฆ๊ฐ€์‹œํ‚ค์ง€ ์•Š์œผ๋ฉฐ, ์„ธ ๋ฒˆ์งธ ์†”๋ฃจ์…˜์€ ์•ฝ 50%์˜ ํ†ต์‹  ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ์ฆ๊ฐ€์‹œํ‚ค์ง€๋งŒ GPU ์ˆ˜์— ๋น„๋ก€ํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ ˆ์•ฝํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. PyTorch๋Š” ZeRO์™€ ์œ ์‚ฌํ•œ ๊ธฐ์ˆ ์ธ FSDP [339]๋ฅผ ๊ตฌํ˜„ํ–ˆ์Šต๋‹ˆ๋‹ค.
Let us mention the study of Bertrand et al. (2023) in the context of image generation, which sheds light on the critical role of data composition in the stability and effectiveness in generative models. They explore scenarios involving a mix of clean data, representative of the true distribution, and synthesized data from previous iterations of the generator. Their analysis reveals that if the data mix consists exclusively of synthesized data, the generative process is likely to degenerate over time, leading to what they describe as a 'clueless generator'. Thus, the generator collapses: it progressively loses its ability to capture the essence of the data distribution it was intended to model. Conversely, they found that when the proportion of clean data in the mix is sufficiently high, the generator, under certain technical conditions, retains the capability to learn and accurately reflect the true data distribution. This work sheds light on the critical role of data composition in the stability and effectiveness of generative models.
์ด๋ฏธ์ง€ ์ƒ์„ฑ์˜ ๋งฅ๋ฝ์—์„œ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ์•ˆ์ •์„ฑ๊ณผ ํšจ๊ณผ์— ์žˆ์–ด ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ์˜ ์ค‘์š”ํ•œ ์—ญํ• ์„ ์กฐ๋ช…ํ•˜๋Š” Bertrand ๋“ฑ(2023)์˜ ์—ฐ๊ตฌ๋ฅผ ์–ธ๊ธ‰ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ์‹ค์ œ ๋ถ„ํฌ๋ฅผ ๋Œ€ํ‘œํ•˜๋Š” ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ์™€ ์ƒ์„ฑ๊ธฐ์˜ ์ด์ „ ๋ฐ˜๋ณต์—์„œ ํ•ฉ์„ฑ๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ํ˜ผํ•ฉ๋œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์˜ ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด ๋ฐ์ดํ„ฐ ๋ฏน์Šค๊ฐ€ ์ „์ ์œผ๋กœ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ ๊ฒฝ์šฐ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๊ฐ€ ๋ณ€์งˆ๋˜์–ด 'clueless generator'๋กœ ์ด์–ด์งˆ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ƒ์„ฑ๊ธฐ๋Š” ๋ถ•๊ดด๋˜์–ด ๋ชจ๋ธ๋งํ•˜๋ ค๊ณ  ํ–ˆ๋˜ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์˜ ๋ณธ์งˆ์„ ์ ์ง„์ ์œผ๋กœ ์ƒ์‹คํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ, ๋ฏน์Šค์—์„œ ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ์˜ ๋น„์œจ์ด ์ถฉ๋ถ„ํžˆ ๋†’์œผ๋ฉด ํŠน์ • ๊ธฐ์ˆ ์  ์กฐ๊ฑด์—์„œ ์ƒ์„ฑ๊ธฐ๊ฐ€ ํ•™์Šตํ•˜๊ณ  ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๋Š” ๊ธฐ๋Šฅ์„ ์œ ์ง€ํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์ƒ์„ฑ ๋ชจ๋ธ์˜ ์•ˆ์ •์„ฑ๊ณผ ํšจ๊ณผ์— ์žˆ์–ด ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ์˜ ์ค‘์š”ํ•œ ์—ญํ• ์„ ์กฐ๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์ด๋ฏธ์ง€ ์ƒ์„ฑ์˜ ๋งฅ๋ฝ์—์„œ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ์•ˆ์ •์„ฑ๊ณผ ํšจ๊ณผ์— ์žˆ์–ด ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ์˜ ์ค‘์š”ํ•œ ์—ญํ• ์„ ์กฐ๋ช…ํ•˜๋Š” Bertrand ๋“ฑ(2023)์˜ ์—ฐ๊ตฌ๋ฅผ ์–ธ๊ธ‰ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ์‹ค์ œ ๋ถ„ํฌ๋ฅผ ๋Œ€ํ‘œํ•˜๋Š” ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ์™€ ์ƒ์„ฑ๊ธฐ์˜ ์ด์ „ ๋ฐ˜๋ณต์—์„œ ํ•ฉ์„ฑ๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ํ˜ผํ•ฉ๋œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์˜ ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด ๋ฐ์ดํ„ฐ ๋ฏน์Šค๊ฐ€ ์ „์ ์œผ๋กœ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ ๊ฒฝ์šฐ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๊ฐ€ ๋ณ€์งˆ๋˜์–ด '๋ฌด์ง€ํ•œ ์ƒ์„ฑ๊ธฐ'๋กœ ์ด์–ด์งˆ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ƒ์„ฑ๊ธฐ๋Š” ๋ถ•๊ดด๋˜์–ด ๋ชจ๋ธ๋งํ•˜๋ ค๊ณ  ํ–ˆ๋˜ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์˜ ๋ณธ์งˆ์„ ํฌ์ฐฉํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์ ์ง„์ ์œผ๋กœ ์ƒ์‹คํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ, ๋ฏน์Šค์—์„œ ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ์˜ ๋น„์œจ์ด ์ถฉ๋ถ„ํžˆ ๋†’์œผ๋ฉด ํŠน์ • ๊ธฐ์ˆ ์  ์กฐ๊ฑด์—์„œ ์ƒ์„ฑ๊ธฐ๊ฐ€ ํ•™์Šตํ•˜๊ณ  ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๋Š” ๊ธฐ๋Šฅ์„ ์œ ์ง€ํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์ƒ์„ฑ ๋ชจ๋ธ์˜ ์•ˆ์ •์„ฑ๊ณผ ํšจ๊ณผ์— ์žˆ์–ด ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ์˜ ์ค‘์š”ํ•œ ์—ญํ• ์„ ์กฐ๋ช…ํ•ฉ๋‹ˆ๋‹ค.
In many of our experiments, we compare models trained with replay to models trained without it. When making such comparisons, we keep the amount of compute constant for training both models. That is, we correspondingly reduce the number of tokens seen from the new dataset to accommodate the additional tokens seen from the replay buffer. We refer to this use of replay as _compute-equivalent replay_. For instance, suppose datasets \(\mathcal{D}_{0}\) and \(\mathcal{D}_{1}\) each contain 100B tokens. We wish to compare model (a) trained sequentially on \(\mathcal{D}_{0}\) and \(\mathcal{D}_{1}\) to model (b) trained sequentially on \(\mathcal{D}_{0}\) and \(\mathcal{D}_{1}\) with 5% compute equivalent replay. Model (a) will see all tokens from both datasets for a total of 200B unique tokens. Model (b) will see 100B unique tokens of \(\mathcal{D}_{0}\) and 95B unique tokens of \(\mathcal{D}_{1}\) plus 5B replayed tokens from \(\mathcal{D}_{0}\) for a total of 200B tokens. In this way, both compared models expend the same amount of compute.
๋งŽ์€ ์‹คํ—˜์—์„œ replay๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ๊ณผ replay ์—†์ด ํ•™์Šต๋œ ๋ชจ๋ธ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋น„๊ต๋ฅผ ํ•  ๋•Œ ๋‘ ๋ชจ๋ธ์„ ๋ชจ๋‘ ํ•™์Šตํ•  ๋•Œ compute ์–‘์„ ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, replay ๋ฒ„ํผ์—์„œ ๋ณด์ด๋Š” ์ถ”๊ฐ€ ํ† ํฐ์„ ์ˆ˜์šฉํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ณด์ด๋Š” ํ† ํฐ ์ˆ˜๋ฅผ ๊ทธ์— ๋”ฐ๋ผ ์ค„์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ replay ์‚ฌ์šฉ์„ _compute-equivalent replay_๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ฐ์ดํ„ฐ ์„ธํŠธ \(\mathcal{D}_{0}\)๊ณผ \(\mathcal{D}_{1}\)์— ๊ฐ๊ฐ 100B ํ† ํฐ์ด ํฌํ•จ๋˜์–ด ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” \(\mathcal{D}_{0}\)์™€ \(\mathcal{D}_{1}\)์— ์ˆœ์ฐจ์ ์œผ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ (a)๋ฅผ 5% compute equivalent replay์™€ ํ•จ๊ป˜ ์ˆœ์ฐจ์ ์œผ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ (b)์™€ ๋น„๊ตํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ (a)๋Š” ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ชจ๋“  ํ† ํฐ์„ ์ด 200B์˜ ๊ณ ์œ  ํ† ํฐ์œผ๋กœ ๋ณด๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ (b)๋Š” \(\mathcal{D}_{0}\)์˜ ๊ณ ์œ  ํ† ํฐ 100B์™€ \(\mathcal{D}_{1}\)์˜ ๊ณ ์œ  ํ† ํฐ 95B์™€ \(\mathcal{D}_{0}\)์—์„œ replay๋œ ํ† ํฐ 5B๋ฅผ ํ•ฉ์ณ ์ด 200B ํ† ํฐ์„ ๋ณด๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ๋น„๊ต๋˜๋Š” ๋‘ ๋ชจ๋ธ์€ ๋™์ผํ•œ ์–‘์˜ compute๋ฅผ ์†Œ๋น„ํ•ฉ๋‹ˆ๋‹ค.
๋งŽ์€ ์‹คํ—˜์—์„œ ๋ฆฌํ”Œ๋ ˆ์ด๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ๊ณผ ๋ฆฌํ”Œ๋ ˆ์ด ์—†์ด ํ•™์Šต๋œ ๋ชจ๋ธ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋น„๊ต๋ฅผ ํ•  ๋•Œ ๋‘ ๋ชจ๋ธ์„ ๋ชจ๋‘ ํ•™์Šตํ•  ๋•Œ ๊ณ„์‚ฐ๋Ÿ‰์„ ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๋ฆฌํ”Œ๋ ˆ์ด ๋ฒ„ํผ์—์„œ ๋ณด์ด๋Š” ์ถ”๊ฐ€ ํ† ํฐ์„ ์ˆ˜์šฉํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ณด์ด๋Š” ํ† ํฐ ์ˆ˜๋ฅผ ๊ทธ์— ๋”ฐ๋ผ ์ค„์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฆฌํ”Œ๋ ˆ์ด ์‚ฌ์šฉ์„ _๊ณ„์‚ฐ ๋™๋“ฑ ๋ฆฌํ”Œ๋ ˆ์ด_๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ฐ์ดํ„ฐ ์„ธํŠธ \(\mathcal{D}_{0}\)๊ณผ \(\mathcal{D}_{1}\)์— ๊ฐ๊ฐ 100B ํ† ํฐ์ด ํฌํ•จ๋˜์–ด ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” \(\mathcal{D}_{0}\)์™€ \(\mathcal{D}_{1}\)์— ์ˆœ์ฐจ์ ์œผ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ (a)๋ฅผ 5% ๊ณ„์‚ฐ๋Ÿ‰๊ณผ ๋™์ผํ•œ ๋ชจ๋ธ (b)์™€ ๋น„๊ตํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ (a)๋Š” ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ชจ๋“  ํ† ํฐ์„ ์ด 200B์˜ ๊ณ ์œ  ํ† ํฐ์œผ๋กœ ๋ณด๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ (b)๋Š” \(\mathcal{D}_{0}\)์˜ ๊ณ ์œ  ํ† ํฐ 100B์™€ \(\mathcal{D}_{1}\)์˜ ๊ณ ์œ  ํ† ํฐ 95B์™€ \(\mathcal{D}_{0}\)์—์„œ ๋ฆฌํ”Œ๋ ˆ์ด๋œ ํ† ํฐ 5B๋ฅผ ํ•ฉ์ณ ์ด 200B ํ† ํฐ์„ ๋ณด๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ๋น„๊ต๋˜๋Š” ๋‘ ๋ชจ๋ธ์€ ๋™์ผํ•œ ์–‘์˜ ๊ณ„์‚ฐ์„ ์†Œ๋น„ํ•ฉ๋‹ˆ๋‹ค.
Although it is expedient to leverage an existing tokenizer (_e.g.,_ OPT [90] and GPT-3 [55] utilize the tokenizer of GPT-2 [26]), using a tokenizer specially designed for the pre-training corpus can be highly beneficial [78], especially for the corpus that consists of diverse domains, languages, and formats. Therefore, recent LLMs often train the customized tokenizers specially for the pre-training corpus with the SentencePiece library [226], which includes Byte-level BPE and Unigram tokenization. A note is that normalization techniques in BPE, such as NFKC [227], may degrade the tokenization performance [34, 64, 78]. When extending existing LLMs (_i.e.,_ continual pre-training or instruction tuning), we should be also aware of the potential side effect with customized tokenizers. For example, LLaMA trains the BPE tokenizer based on a pre-training corpus mainly consisting of English texts, and the derived vocabulary might be less capable in processing non-English data, _e.g.,_ taking longer inference latency to generate Chinese texts.
๊ธฐ์กด tokenizer(์˜ˆ: OPT[90] ๋ฐ GPT-3[55]๋Š” GPT-2[26]์˜ tokenizer๋ฅผ ํ™œ์šฉํ•จ)๋ฅผ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ด ํŽธ๋ฆฌํ•˜์ง€๋งŒ, ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค๋ฅผ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ tokenizer๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ํŠนํžˆ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ, ์–ธ์–ด ๋ฐ ํ˜•์‹์œผ๋กœ ๊ตฌ์„ฑ๋œ ์ฝ”ํผ์Šค์— ๋งค์šฐ ์œ ๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[78]. ๋”ฐ๋ผ์„œ ์ตœ๊ทผ์˜ LLM์€ ๋ฐ”์ดํŠธ ์ˆ˜์ค€ BPE์™€ Unigram tokenization์„ ํฌํ•จํ•˜๋Š” SentencePiece ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ[226]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค๋ฅผ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ๋งž์ถคํ˜• tokenizer๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ์ฃผ์˜ํ•  ์ ์€ NFKC [227]์™€ ๊ฐ™์€ BPE์˜ ์ •๊ทœํ™” ๊ธฐ์ˆ ์ด tokenization ์„ฑ๋Šฅ [34, 64, 78]์„ ์ €ํ•˜์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ๊ธฐ์กด LLM์„ ํ™•์žฅํ•  ๋•Œ(์ฆ‰, ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ๋˜๋Š” Instruction Tuning) ์‚ฌ์šฉ์ž ์ง€์ • tokenizer์™€ ๊ด€๋ จ๋œ ์ž ์žฌ์ ์ธ ๋ถ€์ž‘์šฉ๋„ ์ธ์ง€ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, LLaMA๋Š” ์ฃผ๋กœ ์˜์–ด ํ…์ŠคํŠธ๋กœ ๊ตฌ์„ฑ๋œ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ BPE tokenizer๋ฅผ ํ›ˆ๋ จํ•˜๋ฏ€๋กœ ํŒŒ์ƒ๋œ ์–ดํœ˜๋Š” ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋” ์˜ค๋žœ ์ถ”๋ก  ์ง€์—ฐ ์‹œ๊ฐ„์ด ๊ฑธ๋ฆฌ๋Š” ๋“ฑ ๋น„์˜์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ๋Šฅ๋ ฅ์ด ๋–จ์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ธฐ์กด ํ† ํฐํ™” ๋„๊ตฌ(์˜ˆ: OPT[90] ๋ฐ GPT-3[55]๋Š” GPT-2[26]์˜ ํ† ํฐํ™” ๋„๊ตฌ๋ฅผ ํ™œ์šฉํ•จ)๋ฅผ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ด ํŽธ๋ฆฌํ•˜์ง€๋งŒ, ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค๋ฅผ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ํ† ํฐํ™” ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ํŠนํžˆ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ, ์–ธ์–ด ๋ฐ ํ˜•์‹์œผ๋กœ ๊ตฌ์„ฑ๋œ ์ฝ”ํผ์Šค์— ๋งค์šฐ ์œ ๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[78]. ๋”ฐ๋ผ์„œ ์ตœ๊ทผ์˜ LLM์€ ๋ฐ”์ดํŠธ ์ˆ˜์ค€ BPE์™€ ์–ธ๋‹ˆ๊ทธ๋žจ ํ† ํฐํ™”๋ฅผ ํฌํ•จํ•˜๋Š” SentencePiece ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ[226]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค๋ฅผ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ๋งž์ถคํ˜• ํ† ํฐํ™”๊ธฐ๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ์ฃผ์˜ํ•  ์ ์€ NFKC [227]์™€ ๊ฐ™์€ BPE์˜ ์ •๊ทœํ™” ๊ธฐ์ˆ ์ด ํ† ํฐํ™” ์„ฑ๋Šฅ [34, 64, 78]์„ ์ €ํ•˜์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ๊ธฐ์กด LLM์„ ํ™•์žฅํ•  ๋•Œ(์ฆ‰, ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ๋˜๋Š” ์ง€์นจ ํŠœ๋‹) ์‚ฌ์šฉ์ž ์ง€์ • ํ† ํฐํ™”๊ธฐ์™€ ๊ด€๋ จ๋œ ์ž ์žฌ์ ์ธ ๋ถ€์ž‘์šฉ๋„ ์ธ์ง€ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, LLaMA๋Š” ์ฃผ๋กœ ์˜์–ด ํ…์ŠคํŠธ๋กœ ๊ตฌ์„ฑ๋œ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ BPE ํ† ํฐํ™”๊ธฐ๋ฅผ ํ›ˆ๋ จํ•˜๋ฏ€๋กœ ํŒŒ์ƒ๋œ ์–ดํœ˜๋Š” ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋” ์˜ค๋žœ ์ถ”๋ก  ์ง€์—ฐ ์‹œ๊ฐ„์ด ๊ฑธ๋ฆฌ๋Š” ๋“ฑ ๋น„์˜์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ๋Šฅ๋ ฅ์ด ๋–จ์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Domain Specific Open-Book ExamIn this paper, we focused on a narrower but increasingly popular domain than the general open book exam, called the domain specific open book exam. In domain specific open book exams, we know apriori the domain in which the LLM will be tested - used for inference. The LLM can respond to the prompt using use any and all information from this specific domain, which it has been fine-tuned on. Examples of domain specific examples include enterprise documents, latest news, code repositories belonging to an organization, etc. In all these scenarios, the LLM will be used to respond to the questions, whose answers can be found within a collection of documents (a small practical domain). The retrieval technique itself has little to no impact on the mechanism (though it may impact the accuracy). This paper mainly studies this, domain specific open-book setting and how to adapt a pre-trained LLM to this specific domain, including how to make it more robust to a varying number of retrieved documents and distractors.
๋„๋ฉ”์ธ๋ณ„ ์˜คํ”ˆ๋ถ ์‹œํ—˜์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ผ๋ฐ˜ ์˜คํ”ˆ๋ถ ์‹œํ—˜๋ณด๋‹ค ๋ฒ”์œ„๊ฐ€ ์ข์ง€๋งŒ ์ ์  ์ธ๊ธฐ๋ฅผ ์–ป๊ณ  ์žˆ๋Š” ๋„๋ฉ”์ธ๋ณ„ ์˜คํ”ˆ๋ถ ์‹œํ—˜์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์Šต๋‹ˆ๋‹ค. ๋„๋ฉ”์ธ๋ณ„ ์˜คํ”ˆ๋ถ ์‹œํ—˜์—์„œ๋Š” ์ถ”๋ก ์— ์‚ฌ์šฉ๋˜๋Š” LLM์ด ํ…Œ์ŠคํŠธ๋  ๋„๋ฉ”์ธ์„ ์‚ฌ์ „์— ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLM์€ ์ด ํŠน์ • ๋„๋ฉ”์ธ์˜ ๋ชจ๋“  ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ์— ์‘๋‹ตํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด ๋„๋ฉ”์ธ์— ๋Œ€ํ•ด Fine-tuning์„ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค. ๋„๋ฉ”์ธ๋ณ„ ์˜ˆ์ œ์˜ ์˜ˆ๋กœ๋Š” ๊ธฐ์—… ๋ฌธ์„œ, ์ตœ์‹  ๋‰ด์Šค, ์กฐ์ง์— ์†ํ•œ ์ฝ”๋“œ ๋ฆฌํฌ์ง€ํ† ๋ฆฌ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋“  ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ LLM์€ ๋ฌธ์„œ ๋ชจ์Œ(์ž‘์€ ์‹ค์šฉ์  ๋„๋ฉ”์ธ) ๋‚ด์—์„œ ๋‹ต์„ ์ฐพ์„ ์ˆ˜ ์žˆ๋Š” ์งˆ๋ฌธ์— ์‘๋‹ตํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. Retrieval ๊ธฐ์ˆ  ์ž์ฒด๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๊ฑฐ์˜ ๋˜๋Š” ์ „ํ˜€ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค(์ •ํ™•๋„์—๋Š” ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Œ). ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ฃผ๋กœ ์ด๋Ÿฌํ•œ ๋„๋ฉ”์ธ๋ณ„ ์˜คํ”ˆ๋ถ ์„ค์ •๊ณผ ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ ๋ฐ ์‚ฐ๋งŒ ์š”์†Œ์˜ ๋‹ค์–‘ํ•œ ์ˆ˜์— ๋Œ€ํ•ด ๋”์šฑ ๊ฒฌ๊ณ ํ•˜๊ฒŒ ๋งŒ๋“ค๊ธฐ ์œ„ํ•œ ๋ฐฉ๋ฒ•์„ ํฌํ•จํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต๋œ LLM์„ ์ด ํŠน์ • ๋„๋ฉ”์ธ์— ์ ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
๋„๋ฉ”์ธ๋ณ„ ์˜คํ”ˆ๋ถ ์‹œํ—˜์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ผ๋ฐ˜ ์˜คํ”ˆ๋ถ ์‹œํ—˜๋ณด๋‹ค ๋ฒ”์œ„๊ฐ€ ์ข์ง€๋งŒ ์ ์  ์ธ๊ธฐ๋ฅผ ์–ป๊ณ  ์žˆ๋Š” ๋„๋ฉ”์ธ๋ณ„ ์˜คํ”ˆ๋ถ ์‹œํ—˜์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์Šต๋‹ˆ๋‹ค. ๋„๋ฉ”์ธ๋ณ„ ์˜คํ”ˆ๋ถ ์‹œํ—˜์—์„œ๋Š” ์ถ”๋ก ์— ์‚ฌ์šฉ๋˜๋Š” LLM์ด ํ…Œ์ŠคํŠธ๋  ๋„๋ฉ”์ธ์„ ์‚ฌ์ „์— ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLM์€ ์ด ํŠน์ • ๋„๋ฉ”์ธ์˜ ๋ชจ๋“  ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ์— ์‘๋‹ตํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด ๋„๋ฉ”์ธ์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •์„ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค. ๋„๋ฉ”์ธ๋ณ„ ์˜ˆ์ œ์˜ ์˜ˆ๋กœ๋Š” ๊ธฐ์—… ๋ฌธ์„œ, ์ตœ์‹  ๋‰ด์Šค, ์กฐ์ง์— ์†ํ•œ ์ฝ”๋“œ ๋ฆฌํฌ์ง€ํ† ๋ฆฌ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋“  ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ LLM์€ ๋ฌธ์„œ ๋ชจ์Œ(์ž‘์€ ์‹ค์šฉ์  ๋„๋ฉ”์ธ) ๋‚ด์—์„œ ๋‹ต์„ ์ฐพ์„ ์ˆ˜ ์žˆ๋Š” ์งˆ๋ฌธ์— ์‘๋‹ตํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰ ๊ธฐ์ˆ  ์ž์ฒด๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๊ฑฐ์˜ ๋˜๋Š” ์ „ํ˜€ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค(์ •ํ™•๋„์—๋Š” ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Œ). ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ฃผ๋กœ ์ด๋Ÿฌํ•œ ๋„๋ฉ”์ธ๋ณ„ ์˜คํ”ˆ๋ถ ์„ค์ •๊ณผ ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ ๋ฐ ์‚ฐ๋งŒ ์š”์†Œ์˜ ๋‹ค์–‘ํ•œ ์ˆ˜์— ๋Œ€ํ•ด ๋”์šฑ ๊ฒฌ๊ณ ํ•˜๊ฒŒ ๋งŒ๋“ค๊ธฐ ์œ„ํ•œ ๋ฐฉ๋ฒ•์„ ํฌํ•จํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต๋œ LLM์„ ์ด ํŠน์ • ๋„๋ฉ”์ธ์— ์ ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
Assisting human evaluation.Recent studies have shown that in many cases, synthetic judgements from large-scale LMs (LLMs) can serve as qualified, fast, and low-cost alternatives to actual human evaluation (Gilardi et al., 2023a). Using GPT-4 as the judge, Alpaca Eval (Li et al., 2023b) and MT Bench (Zheng et al., 2023) are two popular benchmarks that measure the comprehensive abilities of LM-based ChatBot. In coding tasks, synthetic environment is a common choice to aid human evaluation, as humans can make the assessment more efficiently via actual executions and analysis on running logs. Gu et al. (2024) propose CRUXEval, a code execution reasoning benchmark consisting of 800 Python functions generated by CodeLAMA-34B. Similarly, Liu et al. (2024a) introduce CodeMind, a framework to gauge the code reasoning abilities of LLMs on Independent Execution Reasoning (IER), Dependent Execution Reasoning (DER), and Specification Reasoning (SR). All these evaluations based on synthetic data show strong correlation with real human judgements.
์ธ๊ฐ„ ํ‰๊ฐ€ ์ง€์›. ์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๋Œ€๊ทœ๋ชจ LLM์˜ ํ•ฉ์„ฑ ํŒ๋‹จ์ด ์‹ค์ œ ์ธ๊ฐ„ ํ‰๊ฐ€์˜ ์ ๊ฒฉํ•˜๊ณ  ๋น ๋ฅด๋ฉฐ ์ €๋ ดํ•œ ๋Œ€์•ˆ์ด ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค(Gilardi et al., 2023a). GPT-4๋ฅผ ์‹ฌ์‚ฌ์ž๋กœ ์‚ฌ์šฉํ•˜๋Š” Alpaca Eval(Li et al., 2023b)๊ณผ MT Bench(Zheng et al., 2023)๋Š” LLM ๊ธฐ๋ฐ˜ ChatBot์˜ ์ข…ํ•ฉ์ ์ธ ๋Šฅ๋ ฅ์„ ์ธก์ •ํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ์ธ๊ธฐ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ์ž…๋‹ˆ๋‹ค. ์ฝ”๋”ฉ ์ž‘์—…์—์„œ๋Š” ์‹ค์ œ ์‹คํ–‰๊ณผ ์‹คํ–‰ ๋กœ๊ทธ ๋ถ„์„์„ ํ†ตํ•ด ๋ณด๋‹ค ํšจ์œจ์ ์œผ๋กœ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ํ•ฉ์„ฑ ํ™˜๊ฒฝ์ด ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ์ง€์›ํ•˜๋Š” ๋ฐ ์ผ๋ฐ˜์ ์œผ๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค. Gu ๋“ฑ(2024)์€ CodeLAMA-34B์— ์˜ํ•ด ์ƒ์„ฑ๋œ 800๊ฐœ์˜ Python ํ•จ์ˆ˜๋กœ ๊ตฌ์„ฑ๋œ ์ฝ”๋“œ ์‹คํ–‰ ์ถ”๋ก  ๋ฒค์น˜๋งˆํฌ์ธ CRUXEval์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Liu ๋“ฑ(2024a)์€ Independent Execution Reasoning (IER), Dependent Execution Reasoning (DER), Specification Reasoning (SR)์— ๋Œ€ํ•œ LLM์˜ ์ฝ”๋“œ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•œ ํ”„๋ ˆ์ž„์›Œํฌ์ธ CodeMind๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ด๋Ÿฌํ•œ ๋ชจ๋“  ํ‰๊ฐ€๋Š” ์‹ค์ œ ์‚ฌ๋žŒ์˜ ํŒ๋‹จ๊ณผ ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์ธ๊ฐ„ ํ‰๊ฐ€ ์ง€์›.์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๋Œ€๊ทœ๋ชจ LM(LLM)์˜ ํ•ฉ์„ฑ ํŒ๋‹จ์ด ์‹ค์ œ ์ธ๊ฐ„ ํ‰๊ฐ€์˜ ์ ๊ฒฉํ•˜๊ณ  ๋น ๋ฅด๋ฉฐ ์ €๋ ดํ•œ ๋Œ€์•ˆ์ด ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค(Gilardi et al., 2023a). GPT-4๋ฅผ ์‹ฌ์‚ฌ์ž๋กœ ์‚ฌ์šฉํ•˜๋Š” Alpaca Eval(Li et al., 2023b)๊ณผ MT Bench(Zheng et al., 2023)๋Š” LM ๊ธฐ๋ฐ˜ ์ฑ—๋ด‡์˜ ์ข…ํ•ฉ์ ์ธ ๋Šฅ๋ ฅ์„ ์ธก์ •ํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ์ธ๊ธฐ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ์ž…๋‹ˆ๋‹ค. ์ฝ”๋”ฉ ์ž‘์—…์—์„œ๋Š” ์‹ค์ œ ์‹คํ–‰๊ณผ ์‹คํ–‰ ๋กœ๊ทธ ๋ถ„์„์„ ํ†ตํ•ด ๋ณด๋‹ค ํšจ์œจ์ ์œผ๋กœ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ํ•ฉ์„ฑ ํ™˜๊ฒฝ์ด ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ์ง€์›ํ•˜๋Š” ๋ฐ ์ผ๋ฐ˜์ ์œผ๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค. Gu ๋“ฑ(2024)์€ CodeLAMA-34B์— ์˜ํ•ด ์ƒ์„ฑ๋œ 800๊ฐœ์˜ ํŒŒ์ด์ฌ ํ•จ์ˆ˜๋กœ ๊ตฌ์„ฑ๋œ ์ฝ”๋“œ ์‹คํ–‰ ์ถ”๋ก  ๋ฒค์น˜๋งˆํฌ์ธ CRUXEval์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Liu ๋“ฑ(2024a)์€ ๋…๋ฆฝ ์‹คํ–‰ ์ถ”๋ก (IER), ์ข…์† ์‹คํ–‰ ์ถ”๋ก (DER), ์‚ฌ์–‘ ์ถ”๋ก (SR)์— ๋Œ€ํ•œ LLM์˜ ์ฝ”๋“œ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•œ ํ”„๋ ˆ์ž„์›Œํฌ์ธ CodeMind๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ด๋Ÿฌํ•œ ๋ชจ๋“  ํ‰๊ฐ€๋Š” ์‹ค์ œ ์‚ฌ๋žŒ์˜ ํŒ๋‹จ๊ณผ ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
ChatGPT (OpenAI, 2022), evolved from InstructGPT (Ouyang et al., 2022), serves as an advanced conversational AI model capable of conducting context-aware, human-like interactions. Its success set the stage for the development of GPT-4 (OpenAI, 2023), a more sophisticated LLM, demonstrating even greater potential in natural language understanding, generation, and various NLP tasks, especially for its multi-modal and reasoning abilities. These models have catalyzed new research directions and applications, intensifying interest in exploring the potential of Artificial General Intelligence (AGI). Exhibiting impressive performance across multiple benchmarks, they have also demonstrated capabilities for few-shot learning and adaptability to new tasks, significantly driving the expansion of NLP research. Consequently, they have inspired both researchers and industry professionals to further harness their potential across a wide array of applications, including sentiment analysis, machine translation, question-answering systems, and more.
ChatGPT(OpenAI, 2022)๋Š” InstructGPT(Ouyang et al., 2022)์—์„œ ๋ฐœ์ „ํ•œ ๊ฒƒ์œผ๋กœ, ์ƒํ™ฉ์— ๋งž๋Š” ์ธ๊ฐ„๊ณผ ์œ ์‚ฌํ•œ ๋Œ€ํ™”๋ฅผ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ณ ๊ธ‰ ๋Œ€ํ™”ํ˜• AI ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์˜ ์„ฑ๊ณต์€ ๋”์šฑ ์ •๊ตํ•œ LLM์ธ GPT-4(OpenAI, 2023)์˜ ๊ฐœ๋ฐœ์„ ์œ„ํ•œ ๋ฐœํŒ์ด ๋˜์—ˆ์œผ๋ฉฐ, ์ž์—ฐ์–ด ์ดํ•ด, ์ƒ์„ฑ ๋ฐ ๋‹ค์–‘ํ•œ NLP ์ž‘์—…, ํŠนํžˆ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ฐ ์ถ”๋ก  ๋Šฅ๋ ฅ์—์„œ ๋” ํฐ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ƒˆ๋กœ์šด ์—ฐ๊ตฌ ๋ฐฉํ–ฅ๊ณผ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์„ ์ด‰์ง„ํ•˜์—ฌ Artificial General Intelligence (AGI)์˜ ์ž ์žฌ๋ ฅ์„ ํƒ๊ตฌํ•˜๋Š” ๋ฐ ๋Œ€ํ•œ ๊ด€์‹ฌ์„ ๊ฐ•ํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ธ์ƒ์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค€ ์ด ๋ชจ๋ธ์€ few-shot learning๊ณผ ์ƒˆ๋กœ์šด ์ž‘์—…์— ๋Œ€ํ•œ ์ ์‘๋ ฅ์„ ํ†ตํ•ด NLP ์—ฐ๊ตฌ์˜ ํ™•์žฅ์„ ํฌ๊ฒŒ ์ฃผ๋„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ์—ฐ๊ตฌ์ž์™€ ์—…๊ณ„ ์ „๋ฌธ๊ฐ€ ๋ชจ๋‘ ๊ฐ์ • ๋ถ„์„, ๊ธฐ๊ณ„ ๋ฒˆ์—ญ, question-answering ์‹œ์Šคํ…œ ๋“ฑ ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋ถ„์•ผ์—์„œ ์ด๋“ค์˜ ์ž ์žฌ๋ ฅ์„ ํ™œ์šฉํ•˜๋ ค๋Š” ๋…ธ๋ ฅ์„ ๊ณ„์†ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
ChatGPT(OpenAI, 2022)๋Š” InstructGPT(Ouyang et al., 2022)์—์„œ ๋ฐœ์ „ํ•œ ๊ฒƒ์œผ๋กœ, ์ƒํ™ฉ์— ๋งž๋Š” ์ธ๊ฐ„๊ณผ ์œ ์‚ฌํ•œ ๋Œ€ํ™”๋ฅผ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ณ ๊ธ‰ ๋Œ€ํ™”ํ˜• AI ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์˜ ์„ฑ๊ณต์€ ๋”์šฑ ์ •๊ตํ•œ LLM์ธ GPT-4(OpenAI, 2023)์˜ ๊ฐœ๋ฐœ์„ ์œ„ํ•œ ๋ฐœํŒ์ด ๋˜์—ˆ์œผ๋ฉฐ, ์ž์—ฐ์–ด ์ดํ•ด, ์ƒ์„ฑ ๋ฐ ๋‹ค์–‘ํ•œ NLP ์ž‘์—…, ํŠนํžˆ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ฐ ์ถ”๋ก  ๋Šฅ๋ ฅ์—์„œ ๋” ํฐ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ƒˆ๋กœ์šด ์—ฐ๊ตฌ ๋ฐฉํ–ฅ๊ณผ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์„ ์ด‰์ง„ํ•˜์—ฌ ์ธ๊ณต ์ผ๋ฐ˜ ์ง€๋Šฅ(AGI)์˜ ์ž ์žฌ๋ ฅ์„ ํƒ๊ตฌํ•˜๋Š” ๋ฐ ๋Œ€ํ•œ ๊ด€์‹ฌ์„ ๊ฐ•ํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ธ์ƒ์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค€ ์ด ๋ชจ๋ธ์€ ์†Œ์ˆ˜ ์ƒท ํ•™์Šต๊ณผ ์ƒˆ๋กœ์šด ์ž‘์—…์— ๋Œ€ํ•œ ์ ์‘๋ ฅ์„ ํ†ตํ•ด NLP ์—ฐ๊ตฌ์˜ ํ™•์žฅ์„ ํฌ๊ฒŒ ์ฃผ๋„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ์—ฐ๊ตฌ์ž์™€ ์—…๊ณ„ ์ „๋ฌธ๊ฐ€ ๋ชจ๋‘ ๊ฐ์ • ๋ถ„์„, ๊ธฐ๊ณ„ ๋ฒˆ์—ญ, ์งˆ์˜์‘๋‹ต ์‹œ์Šคํ…œ ๋“ฑ ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋ถ„์•ผ์—์„œ ์ด๋“ค์˜ ์ž ์žฌ๋ ฅ์„ ํ™œ์šฉํ•˜๋ ค๋Š” ๋…ธ๋ ฅ์„ ๊ณ„์†ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Chain StructurePAL (Gao et al., 2023) and PoT (Chen et al., 2022) introduce programming languages to describe the reasoning process, thereby converting the reasoning problem into the implementation of an executable program to obtain the final answer. Since the program execution is deterministic and performs arithmetic computations accurately, this approach shows excellent performance in mathematical reasoning. Besides, symbol sequence is another type of thought representation. Chain-of-Symbol (Hu et al., 2023) represents the complex environments with condensed symbolic chain representations during planning, which reduces the complexity of the simulation environment. Chain structure variants are shown in Figure 2(c,d) Algorithm of Thought (Sel et al., 2023) injects algorithmic capabilities into the model, making the model's reasoning more logical by adding examples based on algorithms. Its absence of the huge search space of tree search (Long, 2023; Yao et al., 2023) saves computational resources and achieves excellent performance.
์ฒด์ธ ๊ตฌ์กฐPAL(Gao et al., 2023)๊ณผ PoT(Chen et al., 2022)๋Š” ์ถ”๋ก  ๊ณผ์ •์„ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด๋ฅผ ๋„์ž…ํ•˜์—ฌ ์ถ”๋ก  ๋ฌธ์ œ๋ฅผ ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ํ”„๋กœ๊ทธ๋žจ์˜ ๊ตฌํ˜„์œผ๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ ์ตœ์ข… ๋‹ต์„ ์–ป์Šต๋‹ˆ๋‹ค. ํ”„๋กœ๊ทธ๋žจ ์‹คํ–‰์€ ๊ฒฐ์ •๋ก ์ ์ด๊ณ  ์‚ฐ์ˆ  ์—ฐ์‚ฐ์„ ์ •ํ™•ํ•˜๊ฒŒ ์ˆ˜ํ–‰ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ˆ˜ํ•™์  ์ถ”๋ก ์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐํ˜ธ ์‹œํ€€์Šค๋Š” ๋˜ ๋‹ค๋ฅธ ์œ ํ˜•์˜ ์‚ฌ๊ณ  ํ‘œํ˜„์ž…๋‹ˆ๋‹ค. Chain-of-Symbol (Hu et al., 2023)์€ ๊ณ„ํšํ•˜๋Š” ๋™์•ˆ ๋ณต์žกํ•œ ํ™˜๊ฒฝ์„ ์••์ถ•๋œ symbolic chain ํ‘œํ˜„์œผ๋กœ ํ‘œํ˜„ํ•˜์—ฌ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์˜ ๋ณต์žก์„ฑ์„ ์ค„์ž…๋‹ˆ๋‹ค. ์ฒด์ธ ๊ตฌ์กฐ ๋ณ€ํ˜•์€ ๊ทธ๋ฆผ 2(c,d) Algorithm of Thought (Sel et al., 2023)์— ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๊ธฐ๋Šฅ์„ ๋ชจ๋ธ์— ์‚ฝ์ž…ํ•˜์—ฌ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์˜ˆ์ œ๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ ๋ชจ๋ธ์˜ ์ถ”๋ก ์„ ๋ณด๋‹ค ๋…ผ๋ฆฌ์ ์œผ๋กœ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ํŠธ๋ฆฌ ๊ฒ€์ƒ‰์˜ ๊ฑฐ๋Œ€ํ•œ ๊ฒ€์ƒ‰ ๊ณต๊ฐ„(Long, 2023; Yao et al., 2023)์ด ์—†๊ธฐ ๋•Œ๋ฌธ์— ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค๋ฅผ ์ ˆ์•ฝํ•˜๊ณ  ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
์ฒด์ธ ๊ตฌ์กฐPAL(Gao et al., 2023)๊ณผ PoT(Chen et al., 2022)๋Š” ์ถ”๋ก  ๊ณผ์ •์„ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด๋ฅผ ๋„์ž…ํ•˜์—ฌ ์ถ”๋ก  ๋ฌธ์ œ๋ฅผ ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ํ”„๋กœ๊ทธ๋žจ์˜ ๊ตฌํ˜„์œผ๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ ์ตœ์ข… ๋‹ต์„ ์–ป์Šต๋‹ˆ๋‹ค. ํ”„๋กœ๊ทธ๋žจ ์‹คํ–‰์€ ๊ฒฐ์ •๋ก ์ ์ด๊ณ  ์‚ฐ์ˆ  ์—ฐ์‚ฐ์„ ์ •ํ™•ํ•˜๊ฒŒ ์ˆ˜ํ–‰ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ˆ˜ํ•™์  ์ถ”๋ก ์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐํ˜ธ ์‹œํ€€์Šค๋Š” ๋˜ ๋‹ค๋ฅธ ์œ ํ˜•์˜ ์‚ฌ๊ณ  ํ‘œํ˜„์ž…๋‹ˆ๋‹ค. ์ฒด์ธ ์˜ค๋ธŒ ์‹ฌ๋ณผ(Hu et al., 2023)์€ ๊ณ„ํšํ•˜๋Š” ๋™์•ˆ ๋ณต์žกํ•œ ํ™˜๊ฒฝ์„ ์••์ถ•๋œ ์‹ฌ๋ณผ ์ฒด์ธ ํ‘œํ˜„์œผ๋กœ ํ‘œํ˜„ํ•˜์—ฌ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์˜ ๋ณต์žก์„ฑ์„ ์ค„์ž…๋‹ˆ๋‹ค. ์ฒด์ธ ๊ตฌ์กฐ ๋ณ€ํ˜•์€ ๊ทธ๋ฆผ 2(c,d) ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์˜ค๋ธŒ ์‚ฌ๊ณ (Sel et al., 2023)์— ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๊ธฐ๋Šฅ์„ ๋ชจ๋ธ์— ์‚ฝ์ž…ํ•˜์—ฌ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์˜ˆ์ œ๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ ๋ชจ๋ธ์˜ ์ถ”๋ก ์„ ๋ณด๋‹ค ๋…ผ๋ฆฌ์ ์œผ๋กœ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ํŠธ๋ฆฌ ๊ฒ€์ƒ‰์˜ ๊ฑฐ๋Œ€ํ•œ ๊ฒ€์ƒ‰ ๊ณต๊ฐ„(Long, 2023; Yao et al., 2023)์ด ์—†๊ธฐ ๋•Œ๋ฌธ์— ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค๋ฅผ ์ ˆ์•ฝํ•˜๊ณ  ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
One of the many benefits of synthetic data is that it can be generated at scale, providing an abundant supply of training and testing data for AI models. This is particularly valuable in domains where real-world data is scarce or difficult to obtain (e.g., weather data covering all conditions (Lam et al., 2023; Li et al., 2023)). Second, synthetic data can be tailored to specific requirements, such as ensuring a balanced representation of different classes by introducing controlled variations (e.g., up-weighting low-resource languages in multilingual language learning (Przystupa and Abdul-Mageed, 2019)). This level of control over data characteristics can improve model performance and generalization. Third, synthetic data can help mitigate privacy concerns by creating anonymized or de-identified datasets that do not contain sensitive personal information (El Emam et al., 2020; Howe et al., 2017). This is crucial in domains such as healthcare, where patient privacy is of utmost importance (Dahmen and Cook, 2019; Wei et al., 2019).
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๋งŽ์€ ์ด์  ์ค‘ ํ•˜๋‚˜๋Š” ๋Œ€๊ทœ๋ชจ๋กœ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์–ด AI ๋ชจ๋ธ์„ ์œ„ํ•œ ํ’๋ถ€ํ•œ ํ•™์Šต ๋ฐ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๊ฑฐ๋‚˜ ํ™•๋ณดํ•˜๊ธฐ ์–ด๋ ค์šด ๋ถ„์•ผ(์˜ˆ: ๋ชจ๋“  ์กฐ๊ฑด์„ ํฌ๊ด„ํ•˜๋Š” ๋‚ ์”จ ๋ฐ์ดํ„ฐ(Lam et al., 2023; Li et al., 2023))์—์„œ ํŠนํžˆ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์ œ์–ด๋œ ๋ณ€ํ˜•์„ ๋„์ž…ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ํด๋ž˜์Šค์˜ ๊ท ํ˜• ์žกํžŒ ํ‘œํ˜„์„ ๋ณด์žฅํ•˜๋Š” ๋“ฑ ํŠน์ • ์š”๊ตฌ ์‚ฌํ•ญ์— ๋งž๊ฒŒ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ๋‹ค๊ตญ์–ด ์–ธ์–ด ํ•™์Šต์—์„œ ์ €์ž์› ์–ธ์–ด์˜ ๊ฐ€์ค‘์น˜ ์ƒํ–ฅ(Przystupa and Abdul-Mageed, 2019)). ์ด๋Ÿฌํ•œ ์ˆ˜์ค€์˜ ๋ฐ์ดํ„ฐ ํŠน์„ฑ ์ œ์–ด๋Š” ๋ชจ๋ธ ์„ฑ๋Šฅ๊ณผ generalization์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์…‹์งธ, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๋ฏผ๊ฐํ•œ ๊ฐœ์ธ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์ง€ ์•Š์€ ์ต๋ช…ํ™” ๋˜๋Š” ๋น„์‹๋ณ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(El Emam et al., 2020; Howe et al., 2017). ์ด๋Š” ํ™˜์ž์˜ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ๊ฐ€ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ์˜๋ฃŒ ๋ถ„์•ผ์—์„œ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค(Dahmen and Cook, 2019; Wei et al., 2019).
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๋งŽ์€ ์ด์  ์ค‘ ํ•˜๋‚˜๋Š” ๋Œ€๊ทœ๋ชจ๋กœ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์–ด AI ๋ชจ๋ธ์„ ์œ„ํ•œ ํ’๋ถ€ํ•œ ํ•™์Šต ๋ฐ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๊ฑฐ๋‚˜ ํ™•๋ณดํ•˜๊ธฐ ์–ด๋ ค์šด ๋ถ„์•ผ(์˜ˆ: ๋ชจ๋“  ์กฐ๊ฑด์„ ํฌ๊ด„ํ•˜๋Š” ๋‚ ์”จ ๋ฐ์ดํ„ฐ(Lam et al., 2023; Li et al., 2023))์—์„œ ํŠนํžˆ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์ œ์–ด๋œ ๋ณ€ํ˜•์„ ๋„์ž…ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ํด๋ž˜์Šค์˜ ๊ท ํ˜• ์žกํžŒ ํ‘œํ˜„์„ ๋ณด์žฅํ•˜๋Š” ๋“ฑ ํŠน์ • ์š”๊ตฌ ์‚ฌํ•ญ์— ๋งž๊ฒŒ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ๋‹ค๊ตญ์–ด ์–ธ์–ด ํ•™์Šต์—์„œ ์ €์ž์› ์–ธ์–ด์˜ ๊ฐ€์ค‘์น˜ ์ƒํ–ฅ(Przystupa and Abdul-Mageed, 2019)). ์ด๋Ÿฌํ•œ ์ˆ˜์ค€์˜ ๋ฐ์ดํ„ฐ ํŠน์„ฑ ์ œ์–ด๋Š” ๋ชจ๋ธ ์„ฑ๋Šฅ๊ณผ ์ผ๋ฐ˜ํ™”๋ฅผ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์…‹์งธ, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๋ฏผ๊ฐํ•œ ๊ฐœ์ธ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์ง€ ์•Š์€ ์ต๋ช…ํ™” ๋˜๋Š” ๋น„์‹๋ณ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(El Emam et al., 2020; Howe et al., 2017). ์ด๋Š” ํ™˜์ž์˜ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ๊ฐ€ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ์˜๋ฃŒ ๋ถ„์•ผ์—์„œ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค(Dahmen and Cook, 2019; Wei et al., 2019).
Even more recently Ethayarajh et al. proposed a new alignment approach called the Kahneman-Tversky Optimization (KTO) [136]. Unlike existing state-of-the-art approaches, KTO does not require paired preference data (\(x\), \(y_{w}\), \(y_{l}\)), and it only needs (x,y) and knowledge of whether \(y\) is desirable or undesirable. KTO-aligned models are shown to be good or better than DPO-aligned models at scales from 1B to 30B, despite not using paired preferences. KTO is also far easier to use in the real world than preference optimization methods, as the kind of data it needs is far more abundant. As an example, every retail company has a lot of customer interaction data and whether that interaction was successful (e.g., purchase made) or unsuccessful (e.g., no purchase made). However, They have little to no counterfactual data (i.e., what would have made an unsuccessful customer interaction \(y_{l}\) into a successful one \(y_{w}\)). Fig 31 shows a high-level comparison between KTO and other alignment approaches discussed above.
๋” ์ตœ๊ทผ์— Ethayarajh ๋“ฑ์€ Kahneman-Tversky Optimization (KTO)๋ผ๋Š” ์ƒˆ๋กœ์šด ์ •๋ ฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค[136]. ๊ธฐ์กด์˜ ์ตœ์ฒจ๋‹จ ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ ๋‹ฌ๋ฆฌ KTO๋Š” ์Œ์„ ์ด๋ฃฌ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ(\(x\), \(y_{w}\), \(y_{l}\))๊ฐ€ ํ•„์š”ํ•˜์ง€ ์•Š์œผ๋ฉฐ, (x,y)์™€ \(y\)๊ฐ€ ๋ฐ”๋žŒ์งํ•œ์ง€ ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์€์ง€์— ๋Œ€ํ•œ ์ง€์‹๋งŒ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. KTO-aligned ๋ชจ๋ธ์€ ์Œ์„ ์ด๋ฃฌ ์„ ํ˜ธ๋„๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  1B์—์„œ 30B ๊ทœ๋ชจ์—์„œ DPO-aligned ๋ชจ๋ธ๋ณด๋‹ค ๋” ์ข‹๊ฑฐ๋‚˜ ์ข‹์€ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ KTO๋Š” ํ•„์š”ํ•œ ๋ฐ์ดํ„ฐ์˜ ์ข…๋ฅ˜๊ฐ€ ํ›จ์”ฌ ํ’๋ถ€ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์‹ค์ œ ํ™˜๊ฒฝ์—์„œ preference optimization ๋ฐฉ๋ฒ•๋ณด๋‹ค ํ›จ์”ฌ ๋” ์‰ฝ๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ชจ๋“  ์†Œ๋งค ํšŒ์‚ฌ๋Š” ๋งŽ์€ ๊ณ ๊ฐ ์ƒํ˜ธ ์ž‘์šฉ ๋ฐ์ดํ„ฐ์™€ ๊ทธ ์ƒํ˜ธ ์ž‘์šฉ์ด ์„ฑ๊ณต์ (์˜ˆ: ๊ตฌ๋งค)์ด์—ˆ๋Š”์ง€ ์‹คํŒจ์ (์˜ˆ: ๊ตฌ๋งคํ•˜์ง€ ์•Š์Œ)์ด์—ˆ๋Š”์ง€์— ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ณด์œ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ทธ๋“ค์€ ์–ด๋–ค ๋ฐ์ดํ„ฐ๋„ ๊ฐ€์ง€๊ณ  ์žˆ์ง€ ์•Š์Šต๋‹ˆ๋‹ค(์ฆ‰, ์‹คํŒจํ•œ ๊ณ ๊ฐ ์ƒํ˜ธ ์ž‘์šฉ \(y_{l}\)์„ ์„ฑ๊ณตํ•œ ์ƒํ˜ธ ์ž‘์šฉ \(y_{w}\)์œผ๋กœ ๋งŒ๋“œ๋Š” ๋ฐ ์–ด๋–ค ๊ฒƒ์ด ๋„์›€์ด ๋˜์—ˆ์„๊นŒ์š”?). ๊ทธ๋ฆผ 31์€ ์œ„์—์„œ ๋…ผ์˜ํ•œ ๋‹ค๋ฅธ ์ •๋ ฌ ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ KTO๋ฅผ ๊ฐ„๋‹จํžˆ ๋น„๊ตํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๋” ์ตœ๊ทผ์— Ethayarajh ๋“ฑ์€ Kahneman-Tversky ์ตœ์ ํ™”(KTO)๋ผ๋Š” ์ƒˆ๋กœ์šด ์ •๋ ฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค[136]. ๊ธฐ์กด์˜ ์ตœ์ฒจ๋‹จ ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ ๋‹ฌ๋ฆฌ KTO๋Š” ์Œ์„ ์ด๋ฃฌ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ(\(x\), \(y_{w}\), \(y_{l}\))๊ฐ€ ํ•„์š”ํ•˜์ง€ ์•Š์œผ๋ฉฐ, (x,y)์™€ \(y\)๊ฐ€ ๋ฐ”๋žŒ์งํ•œ์ง€ ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์€์ง€์— ๋Œ€ํ•œ ์ง€์‹๋งŒ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. KTO ์ •๋ ฌ ๋ชจ๋ธ์€ ์Œ์„ ์ด๋ฃฌ ์„ ํ˜ธ๋„๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  1B์—์„œ 30B ๊ทœ๋ชจ์—์„œ DPO ์ •๋ ฌ ๋ชจ๋ธ๋ณด๋‹ค ๋” ์ข‹๊ฑฐ๋‚˜ ์ข‹์€ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ KTO๋Š” ํ•„์š”ํ•œ ๋ฐ์ดํ„ฐ์˜ ์ข…๋ฅ˜๊ฐ€ ํ›จ์”ฌ ํ’๋ถ€ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์‹ค์ œ ํ™˜๊ฒฝ์—์„œ ์„ ํ˜ธ๋„ ์ตœ์ ํ™” ๋ฐฉ๋ฒ•๋ณด๋‹ค ํ›จ์”ฌ ๋” ์‰ฝ๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ชจ๋“  ์†Œ๋งค ํšŒ์‚ฌ๋Š” ๋งŽ์€ ๊ณ ๊ฐ ์ƒํ˜ธ ์ž‘์šฉ ๋ฐ์ดํ„ฐ์™€ ๊ทธ ์ƒํ˜ธ ์ž‘์šฉ์ด ์„ฑ๊ณต์ (์˜ˆ: ๊ตฌ๋งค)์ด์—ˆ๋Š”์ง€ ์‹คํŒจ์ (์˜ˆ: ๊ตฌ๋งคํ•˜์ง€ ์•Š์Œ)์ด์—ˆ๋Š”์ง€์— ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ณด์œ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ทธ๋“ค์€ ์–ด๋–ค ๋ฐ์ดํ„ฐ๋„ ๊ฐ€์ง€๊ณ  ์žˆ์ง€ ์•Š์Šต๋‹ˆ๋‹ค(์ฆ‰, ์‹คํŒจํ•œ ๊ณ ๊ฐ ์ƒํ˜ธ ์ž‘์šฉ \(y_{l}\)์„ ์„ฑ๊ณตํ•œ ์ƒํ˜ธ ์ž‘์šฉ \(y_{w}\)์œผ๋กœ ๋งŒ๋“œ๋Š” ๋ฐ ์–ด๋–ค ๊ฒƒ์ด ๋„์›€์ด ๋˜์—ˆ์„๊นŒ์š”?). ๊ทธ๋ฆผ 31์€ ์œ„์—์„œ ๋…ผ์˜ํ•œ ๋‹ค๋ฅธ ์ •๋ ฌ ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ KTO๋ฅผ ๊ฐ„๋‹จํžˆ ๋น„๊ตํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Dolma is built using a pipeline of (1) language filtering, (2) quality filtering, (3) content filtering, (4) deduplication, (5) multi-source mixing, and (6) tokenization. We refer the reader to the Dolma report (Soldaini et al., 2024) for more details about its design principles, details about its construction, and a more detailed summary of its contents. The report provides additional analyses and experimental results from training language models on intermediate states of Dolma to share what we learned about important data curation practices, including the role of content or quality filters, deduplication, and mixing data from multiple sources. We keep documents from each source separate, both during curation as well as in the final release. We open-sourced our high-performance data curation tools; this toolkit can be used to further experiment on Dolma, reproduce our work, and enable fast and easy curation of pretraining corpora. Finally, we also open-sourced our WIMBD tool (Elazar et al., 2023) to help with dataset analysis.
Dolma๋Š” (1) ์–ธ์–ด ํ•„ํ„ฐ๋ง, (2) ํ’ˆ์งˆ ํ•„ํ„ฐ๋ง, (3) ์ฝ˜ํ…์ธ  ํ•„ํ„ฐ๋ง, (4) ์ค‘๋ณต ์ œ๊ฑฐ, (5) ๋‹ค์ค‘ ์†Œ์Šค ๋ฏน์‹ฑ, (6) tokenization ํŒŒ์ดํ”„๋ผ์ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ตฌ์ถ•๋ฉ๋‹ˆ๋‹ค. ์„ค๊ณ„ ์›์น™์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ, ๊ตฌ์ถ•์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ, ๋‚ด์šฉ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์š”์•ฝ์€ Dolma ๋ณด๊ณ ์„œ(Soldaini et al., 2024)๋ฅผ ์ฐธ์กฐํ•˜์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค. ์ด ๋ณด๊ณ ์„œ๋Š” ์ฝ˜ํ…์ธ  ํ•„ํ„ฐ ๋˜๋Š” ํ’ˆ์งˆ ํ•„ํ„ฐ์˜ ์—ญํ• , ์ค‘๋ณต ์ œ๊ฑฐ, ์—ฌ๋Ÿฌ ์†Œ์Šค์˜ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋“ฑ ์ค‘์š”ํ•œ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ๊ด€ํ–‰์— ๋Œ€ํ•ด ๋ฐฐ์šด ๋‚ด์šฉ์„ ๊ณต์œ ํ•˜๊ธฐ ์œ„ํ•ด Dolma์˜ ์ค‘๊ฐ„ ์ƒํƒœ์—์„œ language models๋ฅผ ํ›ˆ๋ จํ•œ ์ถ”๊ฐ€ ๋ถ„์„ ๋ฐ ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ํ๋ ˆ์ด์…˜ ์ค‘์—๋Š” ๋ฌผ๋ก  ์ตœ์ข… ๋ฆด๋ฆฌ์Šค์—์„œ๋„ ๊ฐ ์†Œ์Šค์˜ ๋ฌธ์„œ๋ฅผ ๋ณ„๋„๋กœ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ๊ณ ์„ฑ๋Šฅ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ๋„๊ตฌ๋ฅผ ์˜คํ”ˆ ์†Œ์Šคํ™”ํ•˜์—ฌ Dolma์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์‹คํ—˜, ์ €ํฌ์˜ ์ž‘์—… ์žฌํ˜„, ์‚ฌ์ „ ํ•™์Šต corpora์˜ ๋น ๋ฅด๊ณ  ์‰ฌ์šด ํ๋ ˆ์ด์…˜์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ์ด ํˆดํ‚ท์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ถ„์„์— ๋„์›€์ด ๋˜๋Š” WIMBD ํˆด(Elazar et al., 2023)๋„ ์˜คํ”ˆ ์†Œ์Šคํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค.
Dolma๋Š” (1) ์–ธ์–ด ํ•„ํ„ฐ๋ง, (2) ํ’ˆ์งˆ ํ•„ํ„ฐ๋ง, (3) ์ฝ˜ํ…์ธ  ํ•„ํ„ฐ๋ง, (4) ์ค‘๋ณต ์ œ๊ฑฐ, (5) ๋‹ค์ค‘ ์†Œ์Šค ๋ฏน์‹ฑ, (6) ํ† ํฐํ™” ํŒŒ์ดํ”„๋ผ์ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ตฌ์ถ•๋ฉ๋‹ˆ๋‹ค. ์„ค๊ณ„ ์›์น™์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ, ๊ตฌ์ถ•์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ, ๋‚ด์šฉ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์š”์•ฝ์€ Dolma ๋ณด๊ณ ์„œ(Soldaini et al., 2024)๋ฅผ ์ฐธ์กฐํ•˜์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค. ์ด ๋ณด๊ณ ์„œ๋Š” ์ฝ˜ํ…์ธ  ํ•„ํ„ฐ ๋˜๋Š” ํ’ˆ์งˆ ํ•„ํ„ฐ์˜ ์—ญํ• , ์ค‘๋ณต ์ œ๊ฑฐ, ์—ฌ๋Ÿฌ ์†Œ์Šค์˜ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋“ฑ ์ค‘์š”ํ•œ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ๊ด€ํ–‰์— ๋Œ€ํ•ด ๋ฐฐ์šด ๋‚ด์šฉ์„ ๊ณต์œ ํ•˜๊ธฐ ์œ„ํ•ด Dolma์˜ ์ค‘๊ฐ„ ์ƒํƒœ์—์„œ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•œ ์ถ”๊ฐ€ ๋ถ„์„ ๋ฐ ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ํ๋ ˆ์ด์…˜ ์ค‘์—๋Š” ๋ฌผ๋ก  ์ตœ์ข… ๋ฆด๋ฆฌ์Šค์—์„œ๋„ ๊ฐ ์†Œ์Šค์˜ ๋ฌธ์„œ๋ฅผ ๋ณ„๋„๋กœ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ๊ณ ์„ฑ๋Šฅ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ๋„๊ตฌ๋ฅผ ์˜คํ”ˆ ์†Œ์Šคํ™”ํ•˜์—ฌ Dolma์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์‹คํ—˜, ์ €ํฌ์˜ ์ž‘์—… ์žฌํ˜„, ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ๋ผ์˜ ๋น ๋ฅด๊ณ  ์‰ฌ์šด ํ๋ ˆ์ด์…˜์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ์ด ํˆดํ‚ท์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ถ„์„์— ๋„์›€์ด ๋˜๋Š” WIMBD ํˆด(Elazar et al., 2023)๋„ ์˜คํ”ˆ ์†Œ์Šคํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค.
To tackle the challenge of enhancing large language models' (LLMs) ability to sift through irrelevant text within the retrieval pipeline, our analysis revealed that training solely with oracle (highly relevant) documents can inadvertently diminish the model's ability to discern and disregard irrelevant information. To address this, our algorithm, RAFT, adopts a strategy that integrates oracle documents with a mix of irrelevant ones. This methodology prompts us to investigate the ideal fraction of negative (irrelevant) documents to incorporate throughout the training process and to assess how well this training approach adapts to different volumes of documents encountered by the Retrieval-Augmented Generation (RAG) during the test phase. Our aim is to refine the balance between relevant and irrelevant information to strengthen the model's efficiency in identifying and utilizing pertinent content. Notice that Sec 4.5 looked at what P% of training data should include distractors, while in this section, we study test-time scenarios.
๊ฒ€์ƒ‰ ํŒŒ์ดํ”„๋ผ์ธ ๋‚ด์—์„œ ๊ด€๋ จ์„ฑ์ด ์—†๋Š” ํ…์ŠคํŠธ๋ฅผ ์„ ๋ณ„ํ•˜๋Š” Large Language Models (LLMs)์˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด oracle (๊ด€๋ จ์„ฑ์ด ๋†’์€) ๋ฌธ์„œ๋กœ๋งŒ ํ•™์Šตํ•˜๋ฉด ๊ด€๋ จ์„ฑ์ด ์—†๋Š” ์ •๋ณด๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ๋ฌด์‹œํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์ด ์ €ํ•˜๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ๋ถ„์„ ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜์™”์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ €ํฌ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ RAFT๋Š” oracle ๋ฌธ์„œ์™€ ๊ด€๋ จ ์—†๋Š” ๋ฌธ์„œ๋ฅผ ํ˜ผํ•ฉํ•˜์—ฌ ํ†ตํ•ฉํ•˜๋Š” ์ „๋žต์„ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•๋ก ์„ ํ†ตํ•ด ํ›ˆ๋ จ ๊ณผ์ •์—์„œ ํ†ตํ•ฉํ•  ๋ถ€์ •(๊ด€๋ จ ์—†๋Š”) ๋ฌธ์„œ์˜ ์ด์ƒ์ ์ธ ๋น„์œจ์„ ์กฐ์‚ฌํ•˜๊ณ , ํ…Œ์ŠคํŠธ ๋‹จ๊ณ„์—์„œ Retrieval-Augmented Generation (RAG)์ด ์ ‘ํ•˜๋Š” ๋‹ค์–‘ํ•œ ์–‘์˜ ๋ฌธ์„œ์— ์ด ํ›ˆ๋ จ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์–ผ๋งˆ๋‚˜ ์ž˜ ์ ์‘ํ•˜๋Š”์ง€ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ชฉํ‘œ๋Š” ๊ด€๋ จ์„ฑ ์žˆ๋Š” ์ •๋ณด์™€ ๊ด€๋ จ ์—†๋Š” ์ •๋ณด ์‚ฌ์ด์˜ ๊ท ํ˜•์„ ๊ฐœ์„ ํ•˜์—ฌ ๊ด€๋ จ์„ฑ ์žˆ๋Š” ์ฝ˜ํ…์ธ ๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ํ™œ์šฉํ•˜๋Š” ๋ชจ๋ธ์˜ ํšจ์œจ์„ฑ์„ ๊ฐ•ํ™”ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์„น์…˜ 4.5์—์„œ๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ P%์— ์‚ฐ๋งŒ ์š”์†Œ๋ฅผ ํฌํ•จ์‹œ์ผœ์•ผ ํ•˜๋Š”์ง€ ์‚ดํŽด๋ดค์ง€๋งŒ, ์ด ์„น์…˜์—์„œ๋Š” ํ…Œ์ŠคํŠธ ์‹œ๊ฐ„ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
๊ฒ€์ƒ‰ ํŒŒ์ดํ”„๋ผ์ธ ๋‚ด์—์„œ ๊ด€๋ จ์„ฑ์ด ์—†๋Š” ํ…์ŠคํŠธ๋ฅผ ์„ ๋ณ„ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์˜ค๋ผํด(๊ด€๋ จ์„ฑ์ด ๋†’์€) ๋ฌธ์„œ๋กœ๋งŒ ํ•™์Šตํ•˜๋ฉด ๊ด€๋ จ์„ฑ์ด ์—†๋Š” ์ •๋ณด๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ๋ฌด์‹œํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์ด ์ €ํ•˜๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ๋ถ„์„ ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜์™”์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ €ํฌ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ RAFT๋Š” ์˜ค๋ผํด ๋ฌธ์„œ์™€ ๊ด€๋ จ ์—†๋Š” ๋ฌธ์„œ๋ฅผ ํ˜ผํ•ฉํ•˜์—ฌ ํ†ตํ•ฉํ•˜๋Š” ์ „๋žต์„ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•๋ก ์„ ํ†ตํ•ด ํ›ˆ๋ จ ๊ณผ์ •์—์„œ ํ†ตํ•ฉํ•  ๋ถ€์ •(๊ด€๋ จ ์—†๋Š”) ๋ฌธ์„œ์˜ ์ด์ƒ์ ์ธ ๋น„์œจ์„ ์กฐ์‚ฌํ•˜๊ณ , ํ…Œ์ŠคํŠธ ๋‹จ๊ณ„์—์„œ ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์ƒ์„ฑ(RAG)์ด ์ ‘ํ•˜๋Š” ๋‹ค์–‘ํ•œ ์–‘์˜ ๋ฌธ์„œ์— ์ด ํ›ˆ๋ จ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์–ผ๋งˆ๋‚˜ ์ž˜ ์ ์‘ํ•˜๋Š”์ง€ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ชฉํ‘œ๋Š” ๊ด€๋ จ์„ฑ ์žˆ๋Š” ์ •๋ณด์™€ ๊ด€๋ จ ์—†๋Š” ์ •๋ณด ์‚ฌ์ด์˜ ๊ท ํ˜•์„ ๊ฐœ์„ ํ•˜์—ฌ ๊ด€๋ จ์„ฑ ์žˆ๋Š” ์ฝ˜ํ…์ธ ๋ฅผ ์‹๋ณ„ํ•˜๊ณ  ํ™œ์šฉํ•˜๋Š” ๋ชจ๋ธ์˜ ํšจ์œจ์„ฑ์„ ๊ฐ•ํ™”ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์„น์…˜ 4.5์—์„œ๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ P%์— ์‚ฐ๋งŒ ์š”์†Œ๋ฅผ ํฌํ•จ์‹œ์ผœ์•ผ ํ•˜๋Š”์ง€ ์‚ดํŽด๋ดค์ง€๋งŒ, ์ด ์„น์…˜์—์„œ๋Š” ํ…Œ์ŠคํŠธ ์‹œ๊ฐ„ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
As AI model size grows, neural _scaling laws_ have become a crucial tool to predict the improvements of large models when increasing capacity and the size of original (human or natural) training data. Yet, the widespread use of popular models means that the ecosystem of online data and text will co-evolve to progressively contain increased amounts of synthesized data. In this paper we ask: _How will the scaling laws change in the inevitable regime where synthetic data makes its way into the training corpus?_ Will future models, still improve, or be doomed to degenerate up to total _(model) collapse_? We develop a theoretical framework of model collapse through the lens of scaling laws. We discover a wide range of decay phenomena, analyzing loss of scaling, shifted scaling with number of generations, the "un-learning" of skills, and grokking when mixing human and synthesized data. Our theory is validated by large-scale experiments with a transformer on an arithmetic task and text generation using the large language model Llama2.
AI ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ์ปค์ง์— ๋”ฐ๋ผ ์‹ ๊ฒฝ _scaling laws_์€ ์šฉ๋Ÿ‰์„ ๋Š˜๋ฆฌ๊ณ  ์›๋ณธ(์ธ๊ฐ„ ๋˜๋Š” ์ž์—ฐ) ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํฌ๊ธฐ๋ฅผ ๋Š˜๋ฆด ๋•Œ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ๊ฐœ์„ ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ๋„๊ตฌ๊ฐ€ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋Œ€์ค‘์ ์ธ ๋ชจ๋ธ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ์‚ฌ์šฉ์€ ์˜จ๋ผ์ธ ๋ฐ์ดํ„ฐ ๋ฐ ํ…์ŠคํŠธ ์ƒํƒœ๊ณ„๊ฐ€ ์ง„ํ™”ํ•˜์—ฌ ์ ์ง„์ ์œผ๋กœ ๋” ๋งŽ์€ ์–‘์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํฌํ•จํ•˜๊ฒŒ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์งˆ๋ฌธ์„ ๋˜์ง‘๋‹ˆ๋‹ค: _ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•™์Šต ์ฝ”ํผ์Šค์— ํฌํ•จ๋  ์ˆ˜๋ฐ–์— ์—†๋Š” ์ฒด์ œ์—์„œ scaling laws์€ ์–ด๋–ป๊ฒŒ ๋ณ€ํ•  ๊ฒƒ์ธ๊ฐ€?_ ๋ฏธ๋ž˜์˜ ๋ชจ๋ธ์€ ์—ฌ์ „ํžˆ ๊ฐœ์„ ๋  ๊ฒƒ์ธ๊ฐ€, ์•„๋‹ˆ๋ฉด ์™„์ „ํžˆ _(๋ชจ๋ธ) ๋ถ•๊ดด_์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ํ‡ด๋ณดํ•  ์ˆ˜๋ฐ–์— ์—†์„ ๊ฒƒ์ธ๊ฐ€? ์ €ํฌ๋Š” scaling laws์˜ ๋ Œ์ฆˆ๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ ๋ถ•๊ดด์˜ ์ด๋ก ์  ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์†์‹ค์˜ ํ™•์žฅ์„ฑ ์ƒ์‹ค, ์„ธ๋Œ€ ์ˆ˜์— ๋”ฐ๋ฅธ ํ™•์žฅ์„ฑ ์ด๋™, ์ธ๊ฐ„ ๋ฐ์ดํ„ฐ์™€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํ˜ผํ•ฉํ•  ๋•Œ์˜ ๊ธฐ์ˆ  '์–ธ๋Ÿฌ๋‹' ๋ฐ ๊ทธ๋กœํ‚น ๋ถ„์„ ๋“ฑ ๊ด‘๋ฒ”์œ„ํ•œ ๊ฐ์‡  ํ˜„์ƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ด๋ก ์€ ์‚ฐ์ˆ  ์ž‘์—…์— ๋Œ€ํ•œ transformer์™€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ Llama2๋ฅผ ์‚ฌ์šฉํ•œ ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ๋Œ€์ƒ์œผ๋กœ ํ•œ ๋Œ€๊ทœ๋ชจ ์‹คํ—˜์„ ํ†ตํ•ด ๊ฒ€์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
AI ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ์ปค์ง์— ๋”ฐ๋ผ ์‹ ๊ฒฝ _์Šค์ผ€์ผ๋ง ๋ฒ•์น™_์€ ์šฉ๋Ÿ‰์„ ๋Š˜๋ฆฌ๊ณ  ์›๋ณธ(์ธ๊ฐ„ ๋˜๋Š” ์ž์—ฐ) ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํฌ๊ธฐ๋ฅผ ๋Š˜๋ฆด ๋•Œ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ๊ฐœ์„ ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ๋„๊ตฌ๊ฐ€ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋Œ€์ค‘์ ์ธ ๋ชจ๋ธ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ์‚ฌ์šฉ์€ ์˜จ๋ผ์ธ ๋ฐ์ดํ„ฐ ๋ฐ ํ…์ŠคํŠธ ์ƒํƒœ๊ณ„๊ฐ€ ์ง„ํ™”ํ•˜์—ฌ ์ ์ง„์ ์œผ๋กœ ๋” ๋งŽ์€ ์–‘์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํฌํ•จํ•˜๊ฒŒ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์งˆ๋ฌธ์„ ๋˜์ง‘๋‹ˆ๋‹ค: _ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•™์Šต ์ฝ”ํผ์Šค์— ํฌํ•จ๋  ์ˆ˜๋ฐ–์— ์—†๋Š” ์ฒด์ œ์—์„œ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์€ ์–ด๋–ป๊ฒŒ ๋ณ€ํ•  ๊ฒƒ์ธ๊ฐ€?_ ๋ฏธ๋ž˜์˜ ๋ชจ๋ธ์€ ์—ฌ์ „ํžˆ ๊ฐœ์„ ๋  ๊ฒƒ์ธ๊ฐ€, ์•„๋‹ˆ๋ฉด ์™„์ „ํžˆ _(๋ชจ๋ธ) ๋ถ•๊ดด_์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ํ‡ด๋ณดํ•  ์ˆ˜๋ฐ–์— ์—†์„ ๊ฒƒ์ธ๊ฐ€? ์ €ํฌ๋Š” ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์˜ ๋ Œ์ฆˆ๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ ๋ถ•๊ดด์˜ ์ด๋ก ์  ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์†์‹ค์˜ ํ™•์žฅ์„ฑ ์ƒ์‹ค, ์„ธ๋Œ€ ์ˆ˜์— ๋”ฐ๋ฅธ ํ™•์žฅ์„ฑ ์ด๋™, ์ธ๊ฐ„ ๋ฐ์ดํ„ฐ์™€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํ˜ผํ•ฉํ•  ๋•Œ์˜ ๊ธฐ์ˆ  '์–ธ๋Ÿฌ๋‹' ๋ฐ ๊ทธ๋กœํ‚น ๋ถ„์„ ๋“ฑ ๊ด‘๋ฒ”์œ„ํ•œ ๊ฐ์‡  ํ˜„์ƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ด๋ก ์€ ์‚ฐ์ˆ  ์ž‘์—…์— ๋Œ€ํ•œ ํŠธ๋žœ์Šคํฌ๋จธ์™€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ Llama2๋ฅผ ์‚ฌ์šฉํ•œ ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ๋Œ€์ƒ์œผ๋กœ ํ•œ ๋Œ€๊ทœ๋ชจ ์‹คํ—˜์„ ํ†ตํ•ด ๊ฒ€์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
Hence, the ability of models to effectively translate is contingent upon the availability of high-quality translation pairs (Xu et al., 2023; Maillard et al., 2023). Furthermore, prevalent evaluation tools such as BLEU (Papineni et al., 2002) and COMET-22 (Rei et al., 2022) predominantly rely on reference-based metrics. However, the precision of these evaluations is sensitive to and compromised by substandard references (Kocmi et al., 2023; Freitag et al., 2023). Recent research (Xu et al., 2023; Kocmi et al., 2023; Freitag et al., 2023) has shifted attention towards assessing the quality of parallel datasets, indicating that target references may not consistently represent the highest quality. In Figure 2, we take a translation example from the FLORES-200 dataset, and compare the gold reference translation with outputs from the best ALMA model and GPT-4. This comparison reveals that the gold reference is a flawed translation, as it omits part of information, whereas the system-generated outputs demonstrate superior quality.
๋”ฐ๋ผ์„œ ๋ชจ๋ธ์ด ํšจ๊ณผ์ ์œผ๋กœ ๋ฒˆ์—ญํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์€ ๊ณ ํ’ˆ์งˆ ๋ฒˆ์—ญ ์Œ์˜ ๊ฐ€์šฉ์„ฑ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค(Xu et al., 2023; Maillard et al., 2023). ๋˜ํ•œ BLEU(Papineni et al., 2002) ๋ฐ COMET-22(Rei et al., 2022)์™€ ๊ฐ™์€ ์ผ๋ฐ˜์ ์ธ ํ‰๊ฐ€ ๋„๊ตฌ๋Š” ์ฃผ๋กœ reference-based metrics์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ํ‰๊ฐ€์˜ ์ •ํ™•๋„๋Š” ํ‘œ์ค€ ์ดํ•˜์˜ ์ฐธ์กฐ์— ๋ฏผ๊ฐํ•˜๊ณ  ์ด์— ์˜ํ•ด ์†์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Kocmi et al., 2023; Freitag et al., 2023). ์ตœ๊ทผ ์—ฐ๊ตฌ(Xu et al., 2023; Kocmi et al., 2023; Freitag et al., 2023)์—์„œ๋Š” parallel datasets์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ ๊ด€์‹ฌ์„ ์˜ฎ๊ฒจ๊ฐ€๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Š” target references๊ฐ€ ์ผ๊ด€๋˜๊ฒŒ ์ตœ๊ณ  ํ’ˆ์งˆ์„ ๋‚˜ํƒ€๋‚ด์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2์—์„œ๋Š” FLORES-200 ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ฒˆ์—ญ ์˜ˆ์‹œ๋ฅผ ๋“ค์–ด gold reference ๋ฒˆ์—ญ๊ณผ ์ตœ๊ณ  ALMA ๋ชจ๋ธ ๋ฐ GPT-4์˜ ์ถœ๋ ฅ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ด ๋น„๊ต๋ฅผ ํ†ตํ•ด gold reference๋Š” ์ •๋ณด์˜ ์ผ๋ถ€๋ฅผ ์ƒ๋žตํ•˜๋Š” ๊ฒฐํ•จ์ด ์žˆ๋Š” ๋ฒˆ์—ญ์ธ ๋ฐ˜๋ฉด, ์‹œ์Šคํ…œ ์ƒ์„ฑ ์ถœ๋ ฅ์€ ๋” ์šฐ์ˆ˜ํ•œ ํ’ˆ์งˆ์„ ๋ณด์—ฌ์ฃผ๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋”ฐ๋ผ์„œ ๋ชจ๋ธ์ด ํšจ๊ณผ์ ์œผ๋กœ ๋ฒˆ์—ญํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์€ ๊ณ ํ’ˆ์งˆ ๋ฒˆ์—ญ ์Œ์˜ ๊ฐ€์šฉ์„ฑ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค(Xu et al., 2023; Maillard et al., 2023). ๋˜ํ•œ BLEU(Papineni et al., 2002) ๋ฐ COMET-22(Rei et al., 2022)์™€ ๊ฐ™์€ ์ผ๋ฐ˜์ ์ธ ํ‰๊ฐ€ ๋„๊ตฌ๋Š” ์ฃผ๋กœ ์ฐธ์กฐ ๊ธฐ๋ฐ˜ ๋ฉ”ํŠธ๋ฆญ์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ํ‰๊ฐ€์˜ ์ •ํ™•๋„๋Š” ํ‘œ์ค€ ์ดํ•˜์˜ ์ฐธ์กฐ์— ๋ฏผ๊ฐํ•˜๊ณ  ์ด์— ์˜ํ•ด ์†์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Kocmi et al., 2023; Freitag et al., 2023). ์ตœ๊ทผ ์—ฐ๊ตฌ(Xu et al., 2023; Kocmi et al., 2023; Freitag et al., 2023)์—์„œ๋Š” ๋ณ‘๋ ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ ๊ด€์‹ฌ์„ ์˜ฎ๊ฒจ๊ฐ€๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋Œ€์ƒ ์ฐธ์กฐ๊ฐ€ ์ผ๊ด€๋˜๊ฒŒ ์ตœ๊ณ  ํ’ˆ์งˆ์„ ๋‚˜ํƒ€๋‚ด์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2์—์„œ๋Š” FLORES-200 ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ฒˆ์—ญ ์˜ˆ์‹œ๋ฅผ ๋“ค์–ด ๊ณจ๋“œ ์ฐธ์กฐ ๋ฒˆ์—ญ๊ณผ ์ตœ๊ณ  ALMA ๋ชจ๋ธ ๋ฐ GPT-4์˜ ์ถœ๋ ฅ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ด ๋น„๊ต๋ฅผ ํ†ตํ•ด ๊ณจ๋“œ ์ฐธ์กฐ๋Š” ์ •๋ณด์˜ ์ผ๋ถ€๋ฅผ ์ƒ๋žตํ•˜๋Š” ๊ฒฐํ•จ์ด ์žˆ๋Š” ๋ฒˆ์—ญ์ธ ๋ฐ˜๋ฉด, ์‹œ์Šคํ…œ ์ƒ์„ฑ ์ถœ๋ ฅ์€ ๋” ์šฐ์ˆ˜ํ•œ ํ’ˆ์งˆ์„ ๋ณด์—ฌ์ฃผ๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Bradley-Terry model [5], then fine-tune a language model to maximize the given reward using reinforcement learning algorithms, commonly REINFORCE [45], proximal policy optimization (PPO; [37]), or variants [32]. A closely-related line of work leverages LLMs fine-tuned for instruction following with human feedback to generate additional synthetic preference data for targeted attributes such as safety or harmlessness [2], using only weak supervision from humans in the form of a text rubric for the LLM's annotations. These methods represent a convergence of two bodies of work: one body of work on training language models with reinforcement learning for a variety of objectives [33; 27; 46] and another body of work on general methods for learning from human preferences [12; 19]. Despite the appeal of using relative human preferences, fine-tuning large language models with reinforcement learning remains a major practical challenge; this work provides a theoretically-justified approach to optimizing relative preferences without RL.
๋ธŒ๋ž˜๋“ค๋ฆฌ-ํ…Œ๋ฆฌ ๋ชจ๋ธ[5]์„ ์‚ฌ์šฉํ•œ ๋‹ค์Œ Reinforcement Learning ์•Œ๊ณ ๋ฆฌ์ฆ˜, ์ผ๋ฐ˜์ ์œผ๋กœ REINFORCE[45], Proximal Policy Optimization (PPO; [37]) ๋˜๋Š” ๋ณ€ํ˜•[32]์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฃผ์–ด์ง„ ๋ณด์ƒ์„ ์ตœ๋Œ€ํ™”ํ•˜๋„๋ก ์–ธ์–ด ๋ชจ๋ธ์„ Fine-tuningํ•ฉ๋‹ˆ๋‹ค. ์ด์™€ ๋ฐ€์ ‘ํ•œ ๊ด€๋ จ์ด ์žˆ๋Š” ์—ฐ๊ตฌ ๋ถ„์•ผ๋Š” ์•ˆ์ „์„ฑ์ด๋‚˜ ๋ฌดํ•ด์„ฑ๊ณผ ๊ฐ™์€ ํŠน์ • ์†์„ฑ์— ๋Œ€ํ•ด ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ด Instruction ์ค€์ˆ˜๋ฅผ ์œ„ํ•ด Fine-tuning๋œ LLM์„ ํ™œ์šฉํ•˜์—ฌ LLM์˜ ์ฃผ์„์— ๋Œ€ํ•œ ํ…์ŠคํŠธ ๋ฃจ๋ธŒ๋ฆญ ํ˜•ํƒœ์˜ ์ธ๊ฐ„์˜ ์•ฝํ•œ ๊ฐ๋…๋งŒ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๊ฐ€์ ์ธ ํ•ฉ์„ฑ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค[2]. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์€ ๋‹ค์–‘ํ•œ ๋ชฉํ‘œ๋ฅผ ์œ„ํ•ด Reinforcement Learning์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ํ•œ ๊ฐ€์ง€ ์ž‘์—…[33; 27; 46]๊ณผ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ๋˜ ๋‹ค๋ฅธ ์ž‘์—…[12; 19]์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ž‘์—…์˜ ์œตํ•ฉ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ธ๊ฐ„์˜ ์ƒ๋Œ€์  ์„ ํ˜ธ๋„๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์˜ ๋งค๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , Reinforcement Learning์„ ํ†ตํ•ด ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ Fine-tuningํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์ฃผ์š”ํ•œ ์‹ค์ œ ๊ณผ์ œ๋กœ ๋‚จ์•„ ์žˆ์œผ๋ฉฐ, ์ด ์—ฐ๊ตฌ๋Š” RL ์—†์ด ์ƒ๋Œ€์  ์„ ํ˜ธ๋„๋ฅผ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ด๋ก ์ ์œผ๋กœ ์ •๋‹นํ™”๋œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๋ธŒ๋ž˜๋“ค๋ฆฌ-ํ…Œ๋ฆฌ ๋ชจ๋ธ[5]์„ ์‚ฌ์šฉํ•œ ๋‹ค์Œ ๊ฐ•ํ™” ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜, ์ผ๋ฐ˜์ ์œผ๋กœ REINFORCE[45], ๊ทผ์ ‘ ์ •์ฑ… ์ตœ์ ํ™”(PPO; [37]) ๋˜๋Š” ๋ณ€ํ˜•[32]์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฃผ์–ด์ง„ ๋ณด์ƒ์„ ์ตœ๋Œ€ํ™”ํ•˜๋„๋ก ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด์™€ ๋ฐ€์ ‘ํ•œ ๊ด€๋ จ์ด ์žˆ๋Š” ์—ฐ๊ตฌ ๋ถ„์•ผ๋Š” ์•ˆ์ „์„ฑ์ด๋‚˜ ๋ฌดํ•ด์„ฑ๊ณผ ๊ฐ™์€ ํŠน์ • ์†์„ฑ์— ๋Œ€ํ•ด ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ด ์ง€์‹œ ์ค€์ˆ˜๋ฅผ ์œ„ํ•ด ๋ฏธ์„ธ ์กฐ์ •๋œ LLM์„ ํ™œ์šฉํ•˜์—ฌ LLM์˜ ์ฃผ์„์— ๋Œ€ํ•œ ํ…์ŠคํŠธ ๋ฃจ๋ธŒ๋ฆญ ํ˜•ํƒœ์˜ ์ธ๊ฐ„์˜ ์•ฝํ•œ ๊ฐ๋…๋งŒ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๊ฐ€์ ์ธ ํ•ฉ์„ฑ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค[2]. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์€ ๋‹ค์–‘ํ•œ ๋ชฉํ‘œ๋ฅผ ์œ„ํ•ด ๊ฐ•ํ™” ํ•™์Šต์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ํ•œ ๊ฐ€์ง€ ์ž‘์—…[33; 27; 46]๊ณผ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ๋˜ ๋‹ค๋ฅธ ์ž‘์—…[12; 19]์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ž‘์—…์˜ ์œตํ•ฉ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ธ๊ฐ„์˜ ์ƒ๋Œ€์  ์„ ํ˜ธ๋„๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์˜ ๋งค๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ๊ฐ•ํ™” ํ•™์Šต์„ ํ†ตํ•ด ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์ฃผ์š”ํ•œ ์‹ค์ œ ๊ณผ์ œ๋กœ ๋‚จ์•„ ์žˆ์œผ๋ฉฐ, ์ด ์—ฐ๊ตฌ๋Š” RL ์—†์ด ์ƒ๋Œ€์  ์„ ํ˜ธ๋„๋ฅผ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ด๋ก ์ ์œผ๋กœ ์ •๋‹นํ™”๋œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
* For the reversal tasks of determining the person's full name, only in the entity-preserving reversal case do accuracies become non-trivial. Both token/word reversals completely fail in such tasks. * When determining a person's full name given only the birth date, the reversal task accuracy remains near zero; this is because dates are treated as three entities in our adopted entity detection method, so their ordering is _not preserved_ in the reversal. * If the reversal tasks are simplified to determining the person's _last name only_, then word-level reversal suffices, and token-level reversal also yields non-trivial accuracies. * Some readers may find it surprising that an entity-preserving method can determine the person's full name but not the person's last name. This is a known phenomenon (Allen-Zhu & Li, 2023b): a language model may completely fail at retrieving _later_ tokens of a knowledge piece (such as the last name) without the help of spelling out earlier tokens -- they can be viewed as a Chain of Thought (CoT).
* ์‚ฌ๋žŒ์˜ ์ „์ฒด ์ด๋ฆ„์„ ๊ฒฐ์ •ํ•˜๋Š” reversal ์ž‘์—…์˜ ๊ฒฝ์šฐ, entity-preserving reversal์˜ ๊ฒฝ์šฐ์—๋งŒ ์ •ํ™•๋„๊ฐ€ ์‚ฌ์†Œํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ† ํฐ/๋‹จ์–ด reversal์€ ์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ ์™„์ „ํžˆ ์‹คํŒจํ•ฉ๋‹ˆ๋‹ค. * ์ƒ๋…„์›”์ผ๋งŒ ์ฃผ์–ด์กŒ์„ ๋•Œ ์‚ฌ๋žŒ์˜ ์ „์ฒด ์ด๋ฆ„์„ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒฝ์šฐ, reversal ์ž‘์—… ์ •ํ™•๋„๋Š” 0์— ๊ฐ€๊นŒ์šด ์ˆ˜์ค€์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋‚ ์งœ๊ฐ€ ์ฑ„ํƒ๋œ entity detection ๋ฐฉ๋ฒ•์—์„œ ์„ธ ๊ฐœ์˜ entity๋กœ ์ฒ˜๋ฆฌ๋˜๊ธฐ ๋•Œ๋ฌธ์— reversal์—์„œ ๋‚ ์งœ์˜ ์ˆœ์„œ๊ฐ€ _๋ณด์กด๋˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. * reversal ์ž‘์—…์„ ์‚ฌ๋žŒ์˜ _์„ฑ๋งŒ_ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‹จ์ˆœํ™”ํ•˜๋ฉด ๋‹จ์–ด ์ˆ˜์ค€์˜ reversal๋กœ ์ถฉ๋ถ„ํ•˜๋ฉฐ ํ† ํฐ ์ˆ˜์ค€์˜ reversal๋„ ๋น„ ์‚ฌ์†Œํ•œ ์ •ํ™•๋„๋ฅผ ์‚ฐ์ถœํ•ฉ๋‹ˆ๋‹ค. * ์ผ๋ถ€ ๋…์ž๋“ค์€ entity-preserving ๋ฐฉ๋ฒ•์œผ๋กœ ์‚ฌ๋žŒ์˜ ์ „์ฒด ์ด๋ฆ„์€ ๊ฒฐ์ •ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ์‚ฌ๋žŒ์˜ ์„ฑ์€ ๊ฒฐ์ •ํ•  ์ˆ˜ ์—†๋Š” ๊ฒƒ์ด ๋†€๋ž๋‹ค๊ณ  ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์•Œ๋ ค์ง„ ํ˜„์ƒ์ž…๋‹ˆ๋‹ค(Allen-Zhu & Li, 2023b): ์–ธ์–ด ๋ชจ๋ธ์€ ์•ž์˜ ํ† ํฐ์„ ์ฒ ์žํ•˜๋Š” ๋ฐ ๋„์›€ ์—†์ด๋Š” ์ง€์‹ ์กฐ๊ฐ์˜ _ํ›„๋ฐ˜_ ํ† ํฐ(์˜ˆ: ์„ฑ)์„ ๊ฒ€์ƒ‰ํ•˜๋Š” ๋ฐ ์™„์ „ํžˆ ์‹คํŒจํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” Chain of Thought (CoT)๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
* ์‚ฌ๋žŒ์˜ ์ „์ฒด ์ด๋ฆ„์„ ๊ฒฐ์ •ํ•˜๋Š” ๋ฐ˜์ „ ์ž‘์—…์˜ ๊ฒฝ์šฐ, ์‹ค์ฒด ๋ณด์กด ๋ฐ˜์ „์˜ ๊ฒฝ์šฐ์—๋งŒ ์ •ํ™•๋„๊ฐ€ ์‚ฌ์†Œํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ† ํฐ/๋‹จ์–ด ๋ฐ˜์ „์€ ์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ ์™„์ „ํžˆ ์‹คํŒจํ•ฉ๋‹ˆ๋‹ค. * ์ƒ๋…„์›”์ผ๋งŒ ์ฃผ์–ด์กŒ์„ ๋•Œ ์‚ฌ๋žŒ์˜ ์ „์ฒด ์ด๋ฆ„์„ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒฝ์šฐ, ๋ฐ˜์ „ ์ž‘์—… ์ •ํ™•๋„๋Š” 0์— ๊ฐ€๊นŒ์šด ์ˆ˜์ค€์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋‚ ์งœ๊ฐ€ ์ฑ„ํƒ๋œ ์—”ํ‹ฐํ‹ฐ ๊ฐ์ง€ ๋ฐฉ๋ฒ•์—์„œ ์„ธ ๊ฐœ์˜ ์—”ํ‹ฐํ‹ฐ๋กœ ์ฒ˜๋ฆฌ๋˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฐ˜์ „์—์„œ ๋‚ ์งœ์˜ ์ˆœ์„œ๊ฐ€ _๋ณด์กด๋˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. * ๋ฐ˜์ „ ์ž‘์—…์„ ์‚ฌ๋žŒ์˜ _์„ฑ๋งŒ_ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‹จ์ˆœํ™”ํ•˜๋ฉด ๋‹จ์–ด ์ˆ˜์ค€์˜ ๋ฐ˜์ „์œผ๋กœ ์ถฉ๋ถ„ํ•˜๋ฉฐ ํ† ํฐ ์ˆ˜์ค€์˜ ๋ฐ˜์ „๋„ ๋น„ ์‚ฌ์†Œํ•œ ์ •ํ™•๋„๋ฅผ ์‚ฐ์ถœํ•ฉ๋‹ˆ๋‹ค. * ์ผ๋ถ€ ๋…์ž๋“ค์€ ์—”ํ‹ฐํ‹ฐ ๋ณด์กด ๋ฐฉ๋ฒ•์œผ๋กœ ์‚ฌ๋žŒ์˜ ์ „์ฒด ์ด๋ฆ„์€ ๊ฒฐ์ •ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ์‚ฌ๋žŒ์˜ ์„ฑ์€ ๊ฒฐ์ •ํ•  ์ˆ˜ ์—†๋Š” ๊ฒƒ์ด ๋†€๋ž๋‹ค๊ณ  ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์•Œ๋ ค์ง„ ํ˜„์ƒ์ž…๋‹ˆ๋‹ค(Allen-Zhu & Li, 2023b): ์–ธ์–ด ๋ชจ๋ธ์€ ์•ž์˜ ํ† ํฐ์„ ์ฒ ์žํ•˜๋Š” ๋ฐ ๋„์›€ ์—†์ด๋Š” ์ง€์‹ ์กฐ๊ฐ์˜ _ํ›„๋ฐ˜_ ํ† ํฐ(์˜ˆ: ์„ฑ)์„ ๊ฒ€์ƒ‰ํ•˜๋Š” ๋ฐ ์™„์ „ํžˆ ์‹คํŒจํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ(CoT)๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Large Language Models (LLMs) have drawn a lot of attention due to their strong performance on a wide range of natural language tasks, since the release of ChatGPT in November 2022. LLMs' ability of general-purpose language understanding and generation is acquired by training billions of model's parameters on massive amounts of text data, as predicted by scaling laws [1, 2]. The research area of LLMs, while very recent, is evolving rapidly in many different ways. In this paper, we review some of the most prominent LLMs, including three popular LLM families (GPT, LLaMA, PaLM), and discuss their characteristics, contributions and limitations. We also give an overview of techniques developed to build, and augment LLMs. We then survey popular datasets prepared for LLM training, fine-tuning, and evaluation, review widely used LLM evaluation metrics, and compare the performance of several popular LLMs on a set of representative benchmarks. Finally, we conclude the paper by discussing open challenges and future research directions.
2022๋…„ 11์›” ChatGPT๊ฐ€ ์ถœ์‹œ๋œ ์ดํ›„, Large Language Models (LLMs)์€ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ž‘์—…์—์„œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์—ฌ ๋งŽ์€ ๊ด€์‹ฌ์„ ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLMs์˜ ๋ฒ”์šฉ ์–ธ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์€ [1, 2]์˜ scaling laws์— ๋”ฐ๋ผ ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ•™์Šตํ•จ์œผ๋กœ์จ ์Šต๋“๋ฉ๋‹ˆ๋‹ค. LLMs์˜ ์—ฐ๊ตฌ ๋ถ„์•ผ๋Š” ๋งค์šฐ ์ตœ๊ทผ์˜ ๋ถ„์•ผ์ด์ง€๋งŒ ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ ๋น ๋ฅด๊ฒŒ ์ง„ํ™”ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์„ธ ๊ฐ€์ง€ ์ธ๊ธฐ ์žˆ๋Š” LLM ์ œํ’ˆ๊ตฐ(GPT, LLaMA, PaLM)์„ ํฌํ•จํ•˜์—ฌ ๊ฐ€์žฅ ๋Œ€ํ‘œ์ ์ธ LLM์— ๋Œ€ํ•ด ๊ฒ€ํ† ํ•˜๊ณ , ๊ทธ ํŠน์ง•, ๊ธฐ์—ฌ๋„ ๋ฐ ํ•œ๊ณ„์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์„ ๊ตฌ์ถ•ํ•˜๊ณ  ์ฆ๊ฐ•ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐœ๋ฐœ๋œ ๊ธฐ์ˆ ์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ LLM ํ›ˆ๋ จ, Fine-tuning ๋ฐ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ค€๋น„๋œ ์ธ๊ธฐ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์กฐ์‚ฌํ•˜๊ณ , ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” LLM ํ‰๊ฐ€ ๋ฉ”ํŠธ๋ฆญ์„ ๊ฒ€ํ† ํ•˜๋ฉฐ, ๋Œ€ํ‘œ์ ์ธ ๋ฒค์น˜๋งˆํฌ ์„ธํŠธ์—์„œ ๋ช‡ ๊ฐ€์ง€ ์ธ๊ธฐ ์žˆ๋Š” LLM์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๊ฐœ๋ฐฉํ˜• ๊ณผ์ œ์™€ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•˜์—ฌ ๋…ผ๋ฌธ์„ ๋งˆ๋ฌด๋ฆฌํ•ฉ๋‹ˆ๋‹ค.
2022๋…„ 11์›” ChatGPT๊ฐ€ ์ถœ์‹œ๋œ ์ดํ›„, ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ž‘์—…์—์„œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์—ฌ ๋งŽ์€ ๊ด€์‹ฌ์„ ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLM์˜ ๋ฒ”์šฉ ์–ธ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์€ [1, 2]์˜ ํ™•์žฅ ๋ฒ•์น™์— ๋”ฐ๋ผ ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ•™์Šตํ•จ์œผ๋กœ์จ ์Šต๋“๋ฉ๋‹ˆ๋‹ค. LLM์˜ ์—ฐ๊ตฌ ๋ถ„์•ผ๋Š” ๋งค์šฐ ์ตœ๊ทผ์˜ ๋ถ„์•ผ์ด์ง€๋งŒ ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ ๋น ๋ฅด๊ฒŒ ์ง„ํ™”ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์„ธ ๊ฐ€์ง€ ์ธ๊ธฐ ์žˆ๋Š” LLM ์ œํ’ˆ๊ตฐ(GPT, LLaMA, PaLM)์„ ํฌํ•จํ•˜์—ฌ ๊ฐ€์žฅ ๋Œ€ํ‘œ์ ์ธ LLM์— ๋Œ€ํ•ด ๊ฒ€ํ† ํ•˜๊ณ , ๊ทธ ํŠน์ง•, ๊ธฐ์—ฌ๋„ ๋ฐ ํ•œ๊ณ„์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์„ ๊ตฌ์ถ•ํ•˜๊ณ  ์ฆ๊ฐ•ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐœ๋ฐœ๋œ ๊ธฐ์ˆ ์— ๋Œ€ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ LLM ํ›ˆ๋ จ, ๋ฏธ์„ธ ์กฐ์ • ๋ฐ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ค€๋น„๋œ ์ธ๊ธฐ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์กฐ์‚ฌํ•˜๊ณ , ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” LLM ํ‰๊ฐ€ ๋ฉ”ํŠธ๋ฆญ์„ ๊ฒ€ํ† ํ•˜๋ฉฐ, ๋Œ€ํ‘œ์ ์ธ ๋ฒค์น˜๋งˆํฌ ์„ธํŠธ์—์„œ ๋ช‡ ๊ฐ€์ง€ ์ธ๊ธฐ ์žˆ๋Š” LLM์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๊ฐœ๋ฐฉํ˜• ๊ณผ์ œ์™€ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•˜์—ฌ ๋…ผ๋ฌธ์„ ๋งˆ๋ฌด๋ฆฌํ•ฉ๋‹ˆ๋‹ค.
LLMs acquire language generation and instruction-following capabilities through pretraining and instruction tuning. However, English holds a dominant position in the field of natural language processing, possessing the most abundant collection of text data from various domains. LLMs trained on English-dominant corpora exhibit inferior performance on other non-English languages. Extrapolating LLMs to non-English languages poses a highly valuable research challenge. Common extrapolation approaches consist of the following three steps: (1) extending the vocabulary to add tokens of the target language, and thus enhancing encoding expressiveness to that language. (2) further pretraining to transfer language generation capabilities of LLMs to the target language. The required training scale for this step is generally on the order of billions of tokens, significantly less than the trillions of tokens needed for training from scratch. (3) conducting SFT in the target language to transfer instruction-following capabilities of LLMs.
LLM์€ ์‚ฌ์ „ ํ•™์Šต๊ณผ Instruction Tuning์„ ํ†ตํ•ด ์–ธ์–ด ์ƒ์„ฑ ๋ฐ ์ง€์นจ ์ถ”์ข… ๊ธฐ๋Šฅ์„ ์Šต๋“ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์˜์–ด๋Š” ๋‹ค์–‘ํ•œ ์˜์—ญ์˜ ๊ฐ€์žฅ ํ’๋ถ€ํ•œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ๋ณด์œ ํ•˜๊ณ  ์žˆ๋Š” ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ ์ง€๋ฐฐ์ ์ธ ์œ„์น˜๋ฅผ ์ฐจ์ง€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜์–ด๊ฐ€ ์ง€๋ฐฐ์ ์ธ ์ฝ”ํผ๋ผ๋กœ ํ•™์Šต๋œ LLM์€ ๋‹ค๋ฅธ ๋น„์˜์–ด ์–ธ์–ด์— ๋Œ€ํ•ด ์—ด๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. LLM์„ ๋น„์˜์–ด ์–ธ์–ด๋กœ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ๊ฐ€์น˜ ์žˆ๋Š” ์—ฐ๊ตฌ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ์™ธ์‚ฝ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋‹ค์Œ ์„ธ ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: (1) ๋Œ€์ƒ ์–ธ์–ด์˜ ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜์—ฌ ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜๊ณ , ์ด๋ฅผ ํ†ตํ•ด ํ•ด๋‹น ์–ธ์–ด์— ๋Œ€ํ•œ ์ธ์ฝ”๋”ฉ ํ‘œํ˜„๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. (2) LLM์˜ ์–ธ์–ด ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ๋Œ€์ƒ ์–ธ์–ด๋กœ ์ „์†กํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ์‚ฌ์ „ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์— ํ•„์š”ํ•œ ํ›ˆ๋ จ ๊ทœ๋ชจ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ํ† ํฐ ์ˆ˜์ค€์œผ๋กœ, ์ฒ˜์Œ๋ถ€ํ„ฐ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ์ˆ˜์กฐ ๊ฐœ์˜ ํ† ํฐ๋ณด๋‹ค ํ›จ์”ฌ ์ ์Šต๋‹ˆ๋‹ค. (3) ๋Œ€์ƒ ์–ธ์–ด๋กœ SFT๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ LLM์˜ ์ง€์นจ ์ถ”์ข… ๊ธฐ๋Šฅ์„ ์ „์†กํ•ฉ๋‹ˆ๋‹ค.
LLM์€ ์‚ฌ์ „ ํ•™์Šต๊ณผ ์ง€์นจ ํŠœ๋‹์„ ํ†ตํ•ด ์–ธ์–ด ์ƒ์„ฑ ๋ฐ ์ง€์นจ ์ถ”์ข… ๊ธฐ๋Šฅ์„ ์Šต๋“ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์˜์–ด๋Š” ๋‹ค์–‘ํ•œ ์˜์—ญ์˜ ๊ฐ€์žฅ ํ’๋ถ€ํ•œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ๋ณด์œ ํ•˜๊ณ  ์žˆ๋Š” ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ ์ง€๋ฐฐ์ ์ธ ์œ„์น˜๋ฅผ ์ฐจ์ง€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜์–ด๊ฐ€ ์ง€๋ฐฐ์ ์ธ ์ฝ”ํผ๋ผ๋กœ ํ•™์Šต๋œ LLM์€ ๋‹ค๋ฅธ ๋น„์˜์–ด ์–ธ์–ด์— ๋Œ€ํ•ด ์—ด๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. LLM์„ ๋น„์˜์–ด ์–ธ์–ด๋กœ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ๊ฐ€์น˜ ์žˆ๋Š” ์—ฐ๊ตฌ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ์™ธ์‚ฝ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋‹ค์Œ ์„ธ ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: (1) ๋Œ€์ƒ ์–ธ์–ด์˜ ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜์—ฌ ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜๊ณ , ์ด๋ฅผ ํ†ตํ•ด ํ•ด๋‹น ์–ธ์–ด์— ๋Œ€ํ•œ ์ธ์ฝ”๋”ฉ ํ‘œํ˜„๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. (2) LLM์˜ ์–ธ์–ด ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ๋Œ€์ƒ ์–ธ์–ด๋กœ ์ „์†กํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ์‚ฌ์ „ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์— ํ•„์š”ํ•œ ํ›ˆ๋ จ ๊ทœ๋ชจ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ํ† ํฐ ์ˆ˜์ค€์œผ๋กœ, ์ฒ˜์Œ๋ถ€ํ„ฐ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ์ˆ˜์กฐ ๊ฐœ์˜ ํ† ํฐ๋ณด๋‹ค ํ›จ์”ฌ ์ ์Šต๋‹ˆ๋‹ค. (3) ๋Œ€์ƒ ์–ธ์–ด๋กœ SFT๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ LLM์˜ ์ง€์นจ ์ถ”์ข… ๊ธฐ๋Šฅ์„ ์ „์†กํ•ฉ๋‹ˆ๋‹ค.
**Annotation Details.** Three researchers manually inspected 100 instances randomly sampled from training data. To ensure the quality and consistency of the annotation process, one researcher designed the annotation instruction describing the definition of each aspect (see Table 1) as well as 4 demonstration examples for each aspect category. Figure 3 displays demonstration examples for logicality. These examples are chosen randomly from the remaining training set, which was not included in the initial set of 100 annotation samples, and are manually annotated based on aspect descriptions. Upon the completion of annotation, we gauged the inter-rater agreement rates to validate the reliability of submitted annotation results by computing the average Fleiss' Kappa coefficient (Fleiss, 1971). The mean Kappa score among the three annotators across all aspect categories was 0.56, indicating a moderate agreement rate. Ultimately, we take the mode of annotated labels submitted by three annotators and consider it as a ground truth.
**Annotation Details.** ์„ธ ๋ช…์˜ ์—ฐ๊ตฌ์ž๊ฐ€ ํ•™์Šต ๋ฐ์ดํ„ฐ์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋งํ•œ 100๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋ฅผ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์„ ํ”„๋กœ์„ธ์Šค์˜ ํ’ˆ์งˆ๊ณผ ์ผ๊ด€์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ํ•œ ์—ฐ๊ตฌ์ž๊ฐ€ ๊ฐ ์ธก๋ฉด์˜ ์ •์˜๋ฅผ ์„ค๋ช…ํ•˜๋Š” annotation instruction(ํ‘œ 1 ์ฐธ์กฐ)๊ณผ ๊ฐ ์ธก๋ฉด ์นดํ…Œ๊ณ ๋ฆฌ์— ๋Œ€ํ•œ 4๊ฐœ์˜ ๋ฐ๋ชจ ์˜ˆ์ œ๋ฅผ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3์€ ๋…ผ๋ฆฌ์„ฑ์— ๋Œ€ํ•œ ๋ฐ๋ชจ ์˜ˆ์ œ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ์˜ˆ์ œ๋Š” ์ฒ˜์Œ 100๊ฐœ์˜ ์ฃผ์„ ์ƒ˜ํ”Œ ์„ธํŠธ์— ํฌํ•จ๋˜์ง€ ์•Š์€ ๋‚˜๋จธ์ง€ ํ•™์Šต ์„ธํŠธ์—์„œ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋˜๋ฉฐ, ์ธก๋ฉด ์„ค๋ช…์— ๋”ฐ๋ผ ์ˆ˜๋™์œผ๋กœ ์ฃผ์„์ด ์ง€์ •๋ฉ๋‹ˆ๋‹ค. ์ฃผ์„์ด ์™„๋ฃŒ๋˜๋ฉด ํ‰๊ท  Fleiss' Kappa ๊ณ„์ˆ˜(Fleiss, 1971)๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ ์ œ์ถœ๋œ ์ฃผ์„ ๊ฒฐ๊ณผ์˜ ์‹ ๋ขฐ์„ฑ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด 3๋ช…์˜ ์ฃผ์„์ž ๊ฐ„์˜ ์ผ์น˜์œจ์„ ์ธก์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ์ธก๋ฉด ๋ฒ”์ฃผ์— ๋Œ€ํ•œ ์„ธ ๋ช…์˜ ์ฃผ์„์ž ๊ฐ„์˜ ํ‰๊ท  Kappa ์ ์ˆ˜๋Š” 0.56์œผ๋กœ, ์ค‘๊ฐ„ ์ •๋„์˜ ์ผ์น˜์œจ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ถ๊ทน์ ์œผ๋กœ ์„ธ ๋ช…์˜ ์ฃผ์„์ž๊ฐ€ ์ œ์ถœํ•œ ์ฃผ์„ ๋ ˆ์ด๋ธ”์˜ ๋ชจ๋“œ๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์‹ค์ธก๊ฐ’์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค.
**์ฃผ์„ ์„ธ๋ถ€ ์‚ฌํ•ญ**. ์„ธ ๋ช…์˜ ์—ฐ๊ตฌ์ž๊ฐ€ ํ•™์Šต ๋ฐ์ดํ„ฐ์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋งํ•œ 100๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋ฅผ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์„ ํ”„๋กœ์„ธ์Šค์˜ ํ’ˆ์งˆ๊ณผ ์ผ๊ด€์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ํ•œ ์—ฐ๊ตฌ์ž๊ฐ€ ๊ฐ ์ธก๋ฉด์˜ ์ •์˜๋ฅผ ์„ค๋ช…ํ•˜๋Š” ์ฃผ์„ ์ง€์นจ(ํ‘œ 1 ์ฐธ์กฐ)๊ณผ ๊ฐ ์ธก๋ฉด ์นดํ…Œ๊ณ ๋ฆฌ์— ๋Œ€ํ•œ 4๊ฐœ์˜ ๋ฐ๋ชจ ์˜ˆ์ œ๋ฅผ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3์€ ๋…ผ๋ฆฌ์„ฑ์— ๋Œ€ํ•œ ๋ฐ๋ชจ ์˜ˆ์ œ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ์˜ˆ์ œ๋Š” ์ฒ˜์Œ 100๊ฐœ์˜ ์ฃผ์„ ์ƒ˜ํ”Œ ์„ธํŠธ์— ํฌํ•จ๋˜์ง€ ์•Š์€ ๋‚˜๋จธ์ง€ ํ•™์Šต ์„ธํŠธ์—์„œ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋˜๋ฉฐ, ์ธก๋ฉด ์„ค๋ช…์— ๋”ฐ๋ผ ์ˆ˜๋™์œผ๋กœ ์ฃผ์„์ด ์ง€์ •๋ฉ๋‹ˆ๋‹ค. ์ฃผ์„์ด ์™„๋ฃŒ๋˜๋ฉด ํ‰๊ท  ํ”Œ๋ผ์ด์Šค ์นดํŒŒ ๊ณ„์ˆ˜(Fleiss, 1971)๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ ์ œ์ถœ๋œ ์ฃผ์„ ๊ฒฐ๊ณผ์˜ ์‹ ๋ขฐ์„ฑ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด 3๋ช…์˜ ์ฃผ์„์ž ๊ฐ„์˜ ์ผ์น˜์œจ์„ ์ธก์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ์ธก๋ฉด ๋ฒ”์ฃผ์— ๋Œ€ํ•œ ์„ธ ๋ช…์˜ ์ฃผ์„์ž ๊ฐ„์˜ ํ‰๊ท  ์นดํŒŒ ์ ์ˆ˜๋Š” 0.56์œผ๋กœ, ์ค‘๊ฐ„ ์ •๋„์˜ ์ผ์น˜์œจ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ถ๊ทน์ ์œผ๋กœ ์„ธ ๋ช…์˜ ์ฃผ์„์ž๊ฐ€ ์ œ์ถœํ•œ ์ฃผ์„ ๋ ˆ์ด๋ธ”์˜ ๋ชจ๋“œ๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์‹ค์ธก๊ฐ’์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค.
To calculate carbon emissions, we multiply the total power consumption by a carbon intensity factor, measured in kg CO\({}_{2}\) emitted per KWh, based on the physical location of the data center where each model was trained. The model trained on A100-40GB GPUs was trained in Australia, so we assume a carbon intensity factor of 0.610, the national average for Australia in 2022.10 The model trained on MI250X GPUs was trained in the LUMI supercomputer, which runs on 100% renewable, carbon-neutral energy, so we assume a carbon intensity factor of 0. LUMI is powered entirely by hydroelectric power and some sources (Ubierna et al., 2022) measure the carbon intensity factor of hydroelectric power to be 0.024, which would imply total carbon emissions of 3.54 tCO\({}_{2}\)eq.11 However, we rely on the official LUMI data for our calculations, and thus we estimate total pretraining emissions of **69.78 tCO\({}_{2}\)eq.12** In Table 8 we compare our models with other previously released models based on publicly available information.
ํƒ„์†Œ ๋ฐฐ์ถœ๋Ÿ‰์„ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ๋ชจ๋ธ์ด ํ•™์Šต๋œ ๋ฐ์ดํ„ฐ ์„ผํ„ฐ์˜ ๋ฌผ๋ฆฌ์  ์œ„์น˜๋ฅผ ๊ธฐ์ค€์œผ๋กœ KWh๋‹น ๋ฐฐ์ถœ๋œ CO\({}_{2}\) ํ‚ฌ๋กœ๊ทธ๋žจ ๋‹จ์œ„์˜ ํƒ„์†Œ ๊ฐ•๋„ ๊ณ„์ˆ˜๋ฅผ ์ด ์ „๋ ฅ ์†Œ๋น„๋Ÿ‰์— ๊ณฑํ•ฉ๋‹ˆ๋‹ค. A100-40GB GPU๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์€ ํ˜ธ์ฃผ์—์„œ ํ•™์Šต๋˜์—ˆ์œผ๋ฏ€๋กœ 2022๋…„ ํ˜ธ์ฃผ์˜ ๊ตญ๊ฐ€ ํ‰๊ท ์ธ 0.610์˜ ํƒ„์†Œ ๊ฐ•๋„ ๊ณ„์ˆ˜๋ฅผ ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. MI250X GPU๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์€ 100% ์žฌ์ƒ ๊ฐ€๋Šฅํ•œ ํƒ„์†Œ ์ค‘๋ฆฝ ์—๋„ˆ์ง€๋กœ ์ž‘๋™ํ•˜๋Š” LUMI ์Šˆํผ์ปดํ“จํ„ฐ์—์„œ ํ•™์Šต๋˜์—ˆ์œผ๋ฏ€๋กœ ํƒ„์†Œ ๊ฐ•๋„ ๊ณ„์ˆ˜ 0. LUMI๋Š” ์ „์ ์œผ๋กœ ์ˆ˜๋ ฅ ๋ฐœ์ „์œผ๋กœ ๊ตฌ๋™๋˜๋ฉฐ ์ผ๋ถ€ ์†Œ์Šค(Ubierna et al., 2022)๋Š” ์ˆ˜๋ ฅ ๋ฐœ์ „์˜ ํƒ„์†Œ ๊ฐ•๋„ ๊ณ„์ˆ˜๋ฅผ 0.024๋กœ ์ธก์ •ํ•˜์—ฌ ์ด ํƒ„์†Œ ๋ฐฐ์ถœ๋Ÿ‰์ด 3.54 tCO\({}_{2}\)eq๋ผ๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ณ„์‚ฐ์—๋Š” ๊ณต์‹ LUMI ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฏ€๋กœ ์ด ์‚ฌ์ „ ํ•™์Šต ๋ฐฐ์ถœ๋Ÿ‰์€ **69.78 tCO\({}_{2}\)eq**์ž…๋‹ˆ๋‹ค. ํ‘œ 8์—์„œ๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ •๋ณด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ด์ „์— ์ถœ์‹œ๋œ ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•œ ๋ชจ๋ธ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
ํƒ„์†Œ ๋ฐฐ์ถœ๋Ÿ‰์„ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ๋ชจ๋ธ์ด ํ•™์Šต๋œ ๋ฐ์ดํ„ฐ ์„ผํ„ฐ์˜ ๋ฌผ๋ฆฌ์  ์œ„์น˜๋ฅผ ๊ธฐ์ค€์œผ๋กœ KWh๋‹น ๋ฐฐ์ถœ๋œ CO\({}_{2}\) ํ‚ฌ๋กœ๊ทธ๋žจ ๋‹จ์œ„์˜ ํƒ„์†Œ ๊ฐ•๋„ ๊ณ„์ˆ˜๋ฅผ ์ด ์ „๋ ฅ ์†Œ๋น„๋Ÿ‰์— ๊ณฑํ•ฉ๋‹ˆ๋‹ค. A100-40GB GPU๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์€ ํ˜ธ์ฃผ์—์„œ ํ•™์Šต๋˜์—ˆ์œผ๋ฏ€๋กœ 2022๋…„ ํ˜ธ์ฃผ์˜ ๊ตญ๊ฐ€ ํ‰๊ท ์ธ 0.610์˜ ํƒ„์†Œ ๊ฐ•๋„ ๊ณ„์ˆ˜๋ฅผ ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค.10 MI250X GPU๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์€ 100% ์žฌ์ƒ ๊ฐ€๋Šฅํ•œ ํƒ„์†Œ ์ค‘๋ฆฝ ์—๋„ˆ์ง€๋กœ ์ž‘๋™ํ•˜๋Š” LUMI ์Šˆํผ์ปดํ“จํ„ฐ์—์„œ ํ•™์Šต๋˜์—ˆ์œผ๋ฏ€๋กœ ํƒ„์†Œ ๊ฐ•๋„ ๊ณ„์ˆ˜ 0. LUMI๋Š” ์ „์ ์œผ๋กœ ์ˆ˜๋ ฅ ๋ฐœ์ „์œผ๋กœ ๊ตฌ๋™๋˜๋ฉฐ ์ผ๋ถ€ ์†Œ์Šค(Ubierna et al., 2022)๋Š” ์ˆ˜๋ ฅ ๋ฐœ์ „์˜ ํƒ„์†Œ ๊ฐ•๋„ ๊ณ„์ˆ˜๋ฅผ 0.024๋กœ ์ธก์ •ํ•˜์—ฌ ์ด ํƒ„์†Œ ๋ฐฐ์ถœ๋Ÿ‰์ด 3.54 tCO\({}_{2}\)eq.11์ด๋ผ๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ณ„์‚ฐ์—๋Š” ๊ณต์‹ LUMI ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฏ€๋กœ ์ด ์‚ฌ์ „ ํ•™์Šต ๋ฐฐ์ถœ๋Ÿ‰์€ **69.78 tCO\({}_{2}\)eq.12** ํ‘œ 8์—์„œ๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ •๋ณด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ด์ „์— ์ถœ์‹œ๋œ ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•œ ๋ชจ๋ธ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
\(\bullet\)_Prompt learning with sufficient data._ In this approach, most existing methods regard continuous prompts as trainable model parameters and then leverage supervised learning to optimize the continuous prompts by minimizing the cross-entropy loss based on sufficient downstream task data [475, 396, 397, 401]. As discussed in Section 5.3.1, prefix tuning [396] prepends a sequence of prefixes (_i.e.,_ a set of trainable continuous vectors) to each Transformer layer in language models, while prompt tuning [397] only incorporates trainable prompt vectors at the input layer. By fixing the large-scale parameters of LLMs and only tuning continuous prompt vector, this kind of approaches can be extremely parameter-efficient (Section 5.3). However, these approaches are typically independent of the inputs, lacking sufficient consideration of input semantics. Therefore, the authors in [475] propose context tuning, where the continuous prompts are derived based on the input text and learned through the downstream task losses.
\(\bullet\)_์ถฉ๋ถ„ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•œ Prompt learning._ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ ๋Œ€๋ถ€๋ถ„์˜ ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ์—ฐ์† prompts๋ฅผ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊ฐ„์ฃผํ•œ ๋‹ค์Œ, ์ถฉ๋ถ„ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—… ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ Loss๋ฅผ ์ตœ์†Œํ™”ํ•˜์—ฌ ์—ฐ์† prompts๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ์ง€๋„ ํ•™์Šต์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค [475, 396, 397, 401]. ์„น์…˜ 5.3.1์—์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, prefix tuning [396]์€ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฐ Transformer ๋ ˆ์ด์–ด์— ์ผ๋ จ์˜ ์ ‘๋‘์‚ฌ(์ฆ‰, ํ•™์Šต ๊ฐ€๋Šฅํ•œ ์—ฐ์† ๋ฒกํ„ฐ ์ง‘ํ•ฉ)๋ฅผ ์„ ํ–‰ํ•˜๋Š” ๋ฐ˜๋ฉด, prompt tuning [397]์€ ์ž…๋ ฅ ๋ ˆ์ด์–ด์—๋งŒ ํ•™์Šต ๊ฐ€๋Šฅํ•œ prompt ๋ฒกํ„ฐ๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. LLM์˜ ๋Œ€๊ทœ๋ชจ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ณ ์ •ํ•˜๊ณ  ์—ฐ์† prompt ๋ฒกํ„ฐ๋งŒ ํŠœ๋‹ํ•จ์œผ๋กœ์จ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋งค์šฐ ํŒŒ๋ผ๋ฏธํ„ฐ ํšจ์œจ์ ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์„น์…˜ 5.3). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ž…๋ ฅ๊ณผ ๋…๋ฆฝ์ ์ด์–ด์„œ ์ž…๋ ฅ ์˜๋ฏธ๋ก ์— ๋Œ€ํ•œ ์ถฉ๋ถ„ํ•œ ๊ณ ๋ ค๊ฐ€ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ [475]์˜ ์ €์ž๋“ค์€ ์—ฐ์† prompts๊ฐ€ ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋„์ถœ๋˜๊ณ  ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—… Loss๋ฅผ ํ†ตํ•ด ํ•™์Šต๋˜๋Š” context tuning์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_์ถฉ๋ถ„ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•œ ํ”„๋กฌํ”„ํŠธ ํ•™์Šต._ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ ๋Œ€๋ถ€๋ถ„์˜ ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ์—ฐ์† ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊ฐ„์ฃผํ•œ ๋‹ค์Œ, ์ถฉ๋ถ„ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—… ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ ์†์‹ค์„ ์ตœ์†Œํ™”ํ•˜์—ฌ ์—ฐ์† ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ์ง€๋„ ํ•™์Šต์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค [475, 396, 397, 401]. ์„น์…˜ 5.3.1์—์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ์ ‘๋‘์‚ฌ ํŠœ๋‹ [396]์€ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฐ ํŠธ๋žœ์Šคํฌ๋จธ ๋ ˆ์ด์–ด์— ์ผ๋ จ์˜ ์ ‘๋‘์‚ฌ(์ฆ‰, ํ•™์Šต ๊ฐ€๋Šฅํ•œ ์—ฐ์† ๋ฒกํ„ฐ ์ง‘ํ•ฉ)๋ฅผ ์„ ํ–‰ํ•˜๋Š” ๋ฐ˜๋ฉด, ํ”„๋กฌํ”„ํŠธ ํŠœ๋‹ [397]์€ ์ž…๋ ฅ ๋ ˆ์ด์–ด์—๋งŒ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํ”„๋กฌํ”„ํŠธ ๋ฒกํ„ฐ๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. LLM์˜ ๋Œ€๊ทœ๋ชจ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ณ ์ •ํ•˜๊ณ  ์—ฐ์† ํ”„๋กฌํ”„ํŠธ ๋ฒกํ„ฐ๋งŒ ํŠœ๋‹ํ•จ์œผ๋กœ์จ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋งค์šฐ ํŒŒ๋ผ๋ฏธํ„ฐ ํšจ์œจ์ ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์„น์…˜ 5.3). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ž…๋ ฅ๊ณผ ๋…๋ฆฝ์ ์ด์–ด์„œ ์ž…๋ ฅ ์˜๋ฏธ๋ก ์— ๋Œ€ํ•œ ์ถฉ๋ถ„ํ•œ ๊ณ ๋ ค๊ฐ€ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ [475]์˜ ์ €์ž๋“ค์€ ์—ฐ์† ํ”„๋กฌํ”„ํŠธ๊ฐ€ ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋„์ถœ๋˜๊ณ  ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—… ์†์‹ค์„ ํ†ตํ•ด ํ•™์Šต๋˜๋Š” ์ปจํ…์ŠคํŠธ ํŠœ๋‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
our framework can be flexibly applied to a wide range of LMs, we use FLAN-T5 (Longpre et al., 2023) models in this work because they are open-source and instruction-tuned on both QA and CoT data, enabling us to use it off-the-shelf without additional training or prompt engineering.1 Our experiment results show that LM-guided CoT prompting outperforms both the standard prompting and the original CoT prompting. More precisely, we find that (1) LM-guided CoT with KD and self-consistency (SC) decoding strategy (Wang et al., 2022b) maximizes the performance gain; (2) RL contributes to a slight increase in overall rationale quality and task performance; (3) choosing the highest-quality rationales for the large LM does not always guarantee improved task performance. This work presents a unique alternative to the direct optimization of the large LM through fine-tuning the comparatively smaller LM. Moreover, the clear separation of two fundamental sub-tasks within CoT reasoning grants practitioners greater control over each task.
์šฐ๋ฆฌ์˜ ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ๊ด‘๋ฒ”์œ„ํ•œ LMs์— ์œ ์—ฐํ•˜๊ฒŒ ์ ์šฉ๋  ์ˆ˜ ์žˆ์ง€๋งŒ, ์ด ์ž‘์—…์—์„œ๋Š” FLAN-T5(Longpre et al., 2023) ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๋Š”๋ฐ, ์ด ๋ชจ๋ธ์€ ์˜คํ”ˆ ์†Œ์Šค์ด๋ฉฐ QA ๋ฐ CoT ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด Instruction-tuned ๋˜์–ด ์ถ”๊ฐ€ ํ›ˆ๋ จ์ด๋‚˜ ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง ์—†์ด๋„ ๋ฐ”๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, LM ๊ฐ€์ด๋“œ CoT ํ”„๋กฌํ”„ํŠธ๊ฐ€ ํ‘œ์ค€ ํ”„๋กฌํ”„ํŠธ์™€ ๊ธฐ์กด CoT ํ”„๋กฌํ”„ํŠธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋” ์ •ํ™•ํ•˜๊ฒŒ๋Š” (1) KD ๋ฐ self-consistency (SC) ๋””์ฝ”๋”ฉ ์ „๋žต(Wang et al., 2022b)์„ ์‚ฌ์šฉํ•œ LM ๊ฐ€์ด๋“œ CoT๊ฐ€ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ทน๋Œ€ํ™”ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. (2) Reinforcement Learning (RL)์€ ์ „๋ฐ˜์ ์ธ ๊ทผ๊ฑฐ ํ’ˆ์งˆ๊ณผ ์ž‘์—… ์„ฑ๋Šฅ์„ ์•ฝ๊ฐ„ ์ฆ๊ฐ€์‹œํ‚ค๋Š” ๋ฐ ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค. (3) ๋Œ€๊ทœ๋ชจ LM์— ๊ฐ€์žฅ ๋†’์€ ํ’ˆ์งˆ์˜ ๊ทผ๊ฑฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ํ•ญ์ƒ ์ž‘์—… ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ์ƒ๋Œ€์ ์œผ๋กœ ์ž‘์€ LM์„ Fine-tuningํ•˜์—ฌ ๋Œ€๊ทœ๋ชจ LM์„ ์ง์ ‘ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ๋Œ€ํ•œ ๋…ํŠนํ•œ ๋Œ€์•ˆ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ CoT ์ถ”๋ก  ๋‚ด์—์„œ ๋‘ ๊ฐ€์ง€ ๊ธฐ๋ณธ ํ•˜์œ„ ์ž‘์—…์„ ๋ช…ํ™•ํ•˜๊ฒŒ ๋ถ„๋ฆฌํ•จ์œผ๋กœ์จ ์‹ค๋ฌด์ž๊ฐ€ ๊ฐ ์ž‘์—…์„ ๋” ์ž˜ ์ œ์–ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์šฐ๋ฆฌ์˜ ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ๊ด‘๋ฒ”์œ„ํ•œ LM์— ์œ ์—ฐํ•˜๊ฒŒ ์ ์šฉ๋  ์ˆ˜ ์žˆ์ง€๋งŒ, ์ด ์ž‘์—…์—์„œ๋Š” ํ”Œ๋žœ-T5(Longpre et al., 2023) ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๋Š”๋ฐ, ์ด ๋ชจ๋ธ์€ ์˜คํ”ˆ ์†Œ์Šค์ด๋ฉฐ QA ๋ฐ CoT ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์ง€์‹œ ์กฐ์ •๋˜์–ด ์ถ”๊ฐ€ ํ›ˆ๋ จ์ด๋‚˜ ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง ์—†์ด๋„ ๋ฐ”๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.1 ์‹คํ—˜ ๊ฒฐ๊ณผ, LM ๊ฐ€์ด๋“œ CoT ํ”„๋กฌํ”„ํŠธ๊ฐ€ ํ‘œ์ค€ ํ”„๋กฌํ”„ํŠธ์™€ ๊ธฐ์กด CoT ํ”„๋กฌํ”„ํŠธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋” ์ •ํ™•ํ•˜๊ฒŒ๋Š” (1) KD ๋ฐ ์ž์ฒด ์ผ๊ด€์„ฑ(SC) ๋””์ฝ”๋”ฉ ์ „๋žต(Wang et al., 2022b)์„ ์‚ฌ์šฉํ•œ LM ๊ฐ€์ด๋“œ CoT๊ฐ€ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ทน๋Œ€ํ™”ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. (2) RL์€ ์ „๋ฐ˜์ ์ธ ๊ทผ๊ฑฐ ํ’ˆ์งˆ๊ณผ ์ž‘์—… ์„ฑ๋Šฅ์„ ์•ฝ๊ฐ„ ์ฆ๊ฐ€์‹œํ‚ค๋Š” ๋ฐ ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค. (3) ๋Œ€๊ทœ๋ชจ LM์— ๊ฐ€์žฅ ๋†’์€ ํ’ˆ์งˆ์˜ ๊ทผ๊ฑฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ํ•ญ์ƒ ์ž‘์—… ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ์ƒ๋Œ€์ ์œผ๋กœ ์ž‘์€ LM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ๋Œ€๊ทœ๋ชจ LM์„ ์ง์ ‘ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ๋Œ€ํ•œ ๋…ํŠนํ•œ ๋Œ€์•ˆ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ CoT ์ถ”๋ก  ๋‚ด์—์„œ ๋‘ ๊ฐ€์ง€ ๊ธฐ๋ณธ ํ•˜์œ„ ์ž‘์—…์„ ๋ช…ํ™•ํ•˜๊ฒŒ ๋ถ„๋ฆฌํ•จ์œผ๋กœ์จ ์‹ค๋ฌด์ž๊ฐ€ ๊ฐ ์ž‘์—…์„ ๋” ์ž˜ ์ œ์–ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In Kaplan et al. (2020), OpenAI studies the relationship between the loss function and the number of tokens. In their experiments, they assume that consuming more steps is equivalent to consuming more time. Under this assumption, OpenAI defines a critical batch size that achieves a certain loss without consuming too many steps or tokens. This rationale is valid if the experiments are provided with unlimited GPUs (at least with in the scope of the experiments). Since GPUs are unlimited, enlarging batch size will not increase the single step duration but will decrease the total number of steps. However, in our experiment, since we have fixed resource (number of GPUs), we observe that doubling the batch size almost equals doubling the single-step time. Therefore, enlarging batch size to decrease total training steps has minimal effect on the total training time. In light of this observation, we drop the goal of "not consuming too many steps" and turn towards minimizing the token quantity to achieve the lowest loss, instead.
Kaplan ๋“ฑ(2020)์—์„œ OpenAI๋Š” Loss ํ•จ์ˆ˜์™€ ํ† ํฐ ์ˆ˜ ์‚ฌ์ด์˜ ๊ด€๊ณ„๋ฅผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ด ์‹คํ—˜์—์„œ OpenAI๋Š” ๋” ๋งŽ์€ ๋‹จ๊ณ„๋ฅผ ์†Œ๋น„ํ•˜๋Š” ๊ฒƒ์ด ๋” ๋งŽ์€ ์‹œ๊ฐ„์„ ์†Œ๋น„ํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฐ€์ • ํ•˜์— OpenAI๋Š” ๋„ˆ๋ฌด ๋งŽ์€ ๋‹จ๊ณ„๋‚˜ ํ† ํฐ์„ ์†Œ๋น„ํ•˜์ง€ ์•Š๊ณ  ํŠน์ • Loss๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ์ž„๊ณ„ Batch Size๋ฅผ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ทผ๊ฑฐ๋Š” ์‹คํ—˜์— ๋ฌด์ œํ•œ GPU(์ ์–ด๋„ ์‹คํ—˜ ๋ฒ”์œ„ ๋‚ด์—์„œ)๊ฐ€ ์ œ๊ณต๋˜๋Š” ๊ฒฝ์šฐ ์œ ํšจํ•ฉ๋‹ˆ๋‹ค. GPU๊ฐ€ ๋ฌด์ œํ•œ์ด๋ฏ€๋กœ Batch Size๋ฅผ ๋Š˜๋ฆฌ๋ฉด ๋‹จ์ผ ๋‹จ๊ณ„ ์‹œ๊ฐ„์ด ๋Š˜์–ด๋‚˜์ง€ ์•Š๊ณ  ์ด ๋‹จ๊ณ„ ์ˆ˜๊ฐ€ ์ค„์–ด๋“ญ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ์˜ ์‹คํ—˜์—์„œ๋Š” ๋ฆฌ์†Œ์Šค(GPU ์ˆ˜)๊ฐ€ ๊ณ ์ •๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— Batch Size๋ฅผ ๋‘ ๋ฐฐ๋กœ ๋Š˜๋ฆฌ๋ฉด ๋‹จ์ผ ๋‹จ๊ณ„ ์‹œ๊ฐ„์ด ๋‘ ๋ฐฐ๋กœ ๋Š˜์–ด๋‚˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ํ›ˆ๋ จ ๋‹จ๊ณ„๋ฅผ ์ค„์ด๊ธฐ ์œ„ํ•ด Batch Size๋ฅผ ๋Š˜๋ฆฌ๋Š” ๊ฒƒ์€ ์ด ํ›ˆ๋ จ ์‹œ๊ฐ„์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์ด ๋ฏธ๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์— ๋น„์ถ”์–ด ๋ณผ ๋•Œ, '๋„ˆ๋ฌด ๋งŽ์€ ๋‹จ๊ณ„๋ฅผ ์†Œ๋น„ํ•˜์ง€ ์•Š๊ธฐ'๋ผ๋Š” ๋ชฉํ‘œ๋ฅผ ํฌ๊ธฐํ•˜๊ณ  ๋Œ€์‹  ์ตœ์†Œ Loss๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ํ† ํฐ ์ˆ˜๋ฅผ ์ตœ์†Œํ™”ํ•˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ ์ „ํ™˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
Kaplan ๋“ฑ(2020)์—์„œ OpenAI๋Š” ์†์‹ค ํ•จ์ˆ˜์™€ ํ† ํฐ ์ˆ˜ ์‚ฌ์ด์˜ ๊ด€๊ณ„๋ฅผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ด ์‹คํ—˜์—์„œ OpenAI๋Š” ๋” ๋งŽ์€ ๋‹จ๊ณ„๋ฅผ ์†Œ๋น„ํ•˜๋Š” ๊ฒƒ์ด ๋” ๋งŽ์€ ์‹œ๊ฐ„์„ ์†Œ๋น„ํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฐ€์ • ํ•˜์— OpenAI๋Š” ๋„ˆ๋ฌด ๋งŽ์€ ๋‹จ๊ณ„๋‚˜ ํ† ํฐ์„ ์†Œ๋น„ํ•˜์ง€ ์•Š๊ณ  ํŠน์ • ์†์‹ค์„ ๋‹ฌ์„ฑํ•˜๋Š” ์ž„๊ณ„ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ทผ๊ฑฐ๋Š” ์‹คํ—˜์— ๋ฌด์ œํ•œ GPU(์ ์–ด๋„ ์‹คํ—˜ ๋ฒ”์œ„ ๋‚ด์—์„œ)๊ฐ€ ์ œ๊ณต๋˜๋Š” ๊ฒฝ์šฐ ์œ ํšจํ•ฉ๋‹ˆ๋‹ค. GPU๊ฐ€ ๋ฌด์ œํ•œ์ด๋ฏ€๋กœ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ๋Š˜๋ฆฌ๋ฉด ๋‹จ์ผ ๋‹จ๊ณ„ ์‹œ๊ฐ„์ด ๋Š˜์–ด๋‚˜์ง€ ์•Š๊ณ  ์ด ๋‹จ๊ณ„ ์ˆ˜๊ฐ€ ์ค„์–ด๋“ญ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ์˜ ์‹คํ—˜์—์„œ๋Š” ๋ฆฌ์†Œ์Šค(GPU ์ˆ˜)๊ฐ€ ๊ณ ์ •๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ๋‘ ๋ฐฐ๋กœ ๋Š˜๋ฆฌ๋ฉด ๋‹จ์ผ ๋‹จ๊ณ„ ์‹œ๊ฐ„์ด ๋‘ ๋ฐฐ๋กœ ๋Š˜์–ด๋‚˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ํ›ˆ๋ จ ๋‹จ๊ณ„๋ฅผ ์ค„์ด๊ธฐ ์œ„ํ•ด ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ๋Š˜๋ฆฌ๋Š” ๊ฒƒ์€ ์ด ํ›ˆ๋ จ ์‹œ๊ฐ„์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์ด ๋ฏธ๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์— ๋น„์ถ”์–ด ๋ณผ ๋•Œ, '๋„ˆ๋ฌด ๋งŽ์€ ๋‹จ๊ณ„๋ฅผ ์†Œ๋น„ํ•˜์ง€ ์•Š๊ธฐ'๋ผ๋Š” ๋ชฉํ‘œ๋ฅผ ํฌ๊ธฐํ•˜๊ณ  ๋Œ€์‹  ์ตœ์†Œ ์†์‹ค์„ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ํ† ํฐ ์ˆ˜๋ฅผ ์ตœ์†Œํ™”ํ•˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ ์ „ํ™˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
**Performance Improvement.** Despite being tuned on a moderate number of instances, instruction tuning has become an important way to improve or unlock the abilities of LLMs [69]. Recent studies have experimented with language models in multiple scales (ranging from 77M to 540B), showing that the models of different scales can all benefit from instruction tuning [69, 345], yielding improved performance as the parameter scale increases [94]. Further, smaller models with instruction tuning can even perform better than larger models without fine-tuning [28, 69]. Besides the model scale, instruction tuning demonstrates consistent improvements in various model architectures, pre-trainingobjectives, and model adaptation methods [69]. In practice, instruction tuning offers a general approach to enhancing the abilities of existing language models [69] (including small-sized PLMs). Also, it is much less costly than pre-training, since the amount of instruction data required by LLMs is significantly smaller than pre-training data.
**Performance Improvement.** ์ ๋‹นํ•œ ์ˆ˜์˜ ์ธ์Šคํ„ด์Šค์—์„œ ํŠœ๋‹๋˜์—ˆ์ง€๋งŒ, Instruction Tuning์€ LLM์˜ ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ฑฐ๋‚˜ ํ™œ์šฉํ•˜๋Š” ์ค‘์š”ํ•œ ๋ฐฉ๋ฒ•์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[69]. ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๊ทœ๋ชจ(77M~540B)์˜ ์–ธ์–ด ๋ชจ๋ธ์„ ์‹คํ—˜ํ•œ ๊ฒฐ๊ณผ, ๋‹ค์–‘ํ•œ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์ด ๋ชจ๋‘ Instruction Tuning์˜ ํ˜œํƒ์„ ๋ˆ„๋ฆด ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[69, 345]. ๋˜ํ•œ Instruction Tuning์ด ์ ์šฉ๋œ ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ์€ Fine-tuningํ•˜์ง€ ์•Š์€ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค[28, 69]. ๋ชจ๋ธ ๊ทœ๋ชจ ์™ธ์—๋„ Instruction Tuning์€ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜, ์‚ฌ์ „ ํ•™์Šต ๋ชฉํ‘œ ๋ฐ ๋ชจ๋ธ ์ ์‘ ๋ฐฉ๋ฒ•์—์„œ ์ผ๊ด€๋œ ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค[69]. ์‹ค์ œ๋กœ Instruction Tuning์€ ๊ธฐ์กด ์–ธ์–ด ๋ชจ๋ธ(์†Œ๊ทœ๋ชจ PLM ํฌํ•จ)์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์ผ๋ฐ˜์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค[69]. ๋˜ํ•œ Instruction Tuning์€ LLM์— ํ•„์š”ํ•œ Instruction ๋ฐ์ดํ„ฐ์˜ ์–‘์ด ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ณด๋‹ค ํ›จ์”ฌ ์ ๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์ „ ํ•™์Šต๋ณด๋‹ค ๋น„์šฉ์ด ํ›จ์”ฌ ์ ๊ฒŒ ๋“ญ๋‹ˆ๋‹ค.
**์„ฑ๋Šฅ ๊ฐœ์„ **. ์ ๋‹นํ•œ ์ˆ˜์˜ ์ธ์Šคํ„ด์Šค์—์„œ ํŠœ๋‹๋˜์—ˆ์ง€๋งŒ, ๋ช…๋ น์–ด ํŠœ๋‹์€ LLM์˜ ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ฑฐ๋‚˜ ํ™œ์šฉํ•˜๋Š” ์ค‘์š”ํ•œ ๋ฐฉ๋ฒ•์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[69]. ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๊ทœ๋ชจ(77M~540B)์˜ ์–ธ์–ด ๋ชจ๋ธ์„ ์‹คํ—˜ํ•œ ๊ฒฐ๊ณผ, ๋‹ค์–‘ํ•œ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์ด ๋ชจ๋‘ ๋ช…๋ น์–ด ํŠœ๋‹์˜ ํ˜œํƒ์„ ๋ˆ„๋ฆด ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[69, 345]. ๋˜ํ•œ ๋ช…๋ น์–ด ํŠœ๋‹์ด ์ ์šฉ๋œ ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ์€ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์ง€ ์•Š์€ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค[28, 69]. ๋ชจ๋ธ ๊ทœ๋ชจ ์™ธ์—๋„ ๋ช…๋ น์–ด ํŠœ๋‹์€ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜, ์‚ฌ์ „ ํ•™์Šต ๋ชฉํ‘œ ๋ฐ ๋ชจ๋ธ ์ ์‘ ๋ฐฉ๋ฒ•์—์„œ ์ผ๊ด€๋œ ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค[69]. ์‹ค์ œ๋กœ ๋ช…๋ น์–ด ํŠœ๋‹์€ ๊ธฐ์กด ์–ธ์–ด ๋ชจ๋ธ(์†Œ๊ทœ๋ชจ PLM ํฌํ•จ)์˜ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์ผ๋ฐ˜์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค[69]. ๋˜ํ•œ ๋ช…๋ น์–ด ํŠœ๋‹์€ LLM์— ํ•„์š”ํ•œ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ์˜ ์–‘์ด ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ณด๋‹ค ํ›จ์”ฌ ์ ๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์ „ ํ•™์Šต๋ณด๋‹ค ๋น„์šฉ์ด ํ›จ์”ฌ ์ ๊ฒŒ ๋“ญ๋‹ˆ๋‹ค.
The concurrent work by Allen-Zhu & Li (2023) investigates a related set of failures and potential solutions. Exploring the capability to answer questions based on synthetic biographies, they examine several data augmentation strategies, including incorporating instruction tuning data into pre-training, generating multiple unique biography entries, permuting biography sentences, and substituting pronouns or partial names with full names. They discover that augmentation during the pre-training phase is essential for enhancing downstream question answering performance across various tasks. However, in real pre-training data, some augmentations may not be feasible -- for instance, permuting sentences could degrade language model quality, and it remains uncertain how to best rewrite data during augmentation. Reverse training addresses this issue by presenting a distinct language task (the reversed language) to the language model, thereby avoiding interference with the primary task of left-to-right natural language modeling.
Allen-Zhu & Li(2023)์˜ ๋™์‹œ ์ž‘์—…์€ ๊ด€๋ จ๋œ ์ผ๋ จ์˜ ์‹คํŒจ์™€ ์ž ์žฌ์  ํ•ด๊ฒฐ์ฑ…์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ์ „๊ธฐ ๊ธฐ๋ฐ˜์œผ๋กœ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๊ธฐ๋Šฅ์„ ํƒ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์ „ ํ•™์Šต์— Instruction Tuning ๋ฐ์ดํ„ฐ ํ†ตํ•ฉ, ์—ฌ๋Ÿฌ ๊ฐœ์˜ ๊ณ ์œ ํ•œ ์ „๊ธฐ ํ•ญ๋ชฉ ์ƒ์„ฑ, ์ „๊ธฐ ๋ฌธ์žฅ ์ˆœ์—ด, ๋Œ€๋ช…์‚ฌ ๋˜๋Š” ๋ถ€๋ถ„ ์ด๋ฆ„์„ ์ „์ฒด ์ด๋ฆ„์œผ๋กœ ๋Œ€์ฒดํ•˜๋Š” ๋“ฑ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ์ „๋žต์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์—์„œ์˜ ์ฆ๊ฐ•์ด ํ•„์ˆ˜์ ์ด๋ผ๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์—์„œ๋Š” ์ผ๋ถ€ ์ฆ๊ฐ•์ด ๋ถˆ๊ฐ€๋Šฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ฌธ์žฅ์„ ์ˆœ์—ดํ•˜๋ฉด ์–ธ์–ด ๋ชจ๋ธ์˜ ํ’ˆ์งˆ์ด ์ €ํ•˜๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ฆ๊ฐ• ์ค‘์— ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐ€์žฅ ์ž˜ ๋‹ค์‹œ ์ž‘์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์•„์ง ๋ถˆํ™•์‹คํ•ฉ๋‹ˆ๋‹ค. ์—ญ ํ›ˆ๋ จ์€ ์–ธ์–ด ๋ชจ๋ธ์— ์™ผ์ชฝ์—์„œ ์˜ค๋ฅธ์ชฝ์œผ๋กœ์˜ ์ž์—ฐ์–ด ๋ชจ๋ธ๋ง์ด๋ผ๋Š” ์ฃผ์š” ์ž‘์—…๊ณผ ๊ฐ„์„ญํ•˜์ง€ ์•Š๋„๋ก ํ•˜๊ธฐ ์œ„ํ•ด ์–ธ์–ด ๋ชจ๋ธ์— ๋ณ„๊ฐœ์˜ ์–ธ์–ด ์ž‘์—…(์—ญ์–ด)์„ ์ œ์‹œํ•จ์œผ๋กœ์จ ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค.
Allen-Zhu & Li(2023)์˜ ๋™์‹œ ์ž‘์—…์€ ๊ด€๋ จ๋œ ์ผ๋ จ์˜ ์‹คํŒจ์™€ ์ž ์žฌ์  ํ•ด๊ฒฐ์ฑ…์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ์ „๊ธฐ ๊ธฐ๋ฐ˜์œผ๋กœ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” ๊ธฐ๋Šฅ์„ ํƒ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์ „ ํ•™์Šต์— ์ง€์นจ ํŠœ๋‹ ๋ฐ์ดํ„ฐ ํ†ตํ•ฉ, ์—ฌ๋Ÿฌ ๊ฐœ์˜ ๊ณ ์œ ํ•œ ์ „๊ธฐ ํ•ญ๋ชฉ ์ƒ์„ฑ, ์ „๊ธฐ ๋ฌธ์žฅ ์ˆœ์—ด, ๋Œ€๋ช…์‚ฌ ๋˜๋Š” ๋ถ€๋ถ„ ์ด๋ฆ„์„ ์ „์ฒด ์ด๋ฆ„์œผ๋กœ ๋Œ€์ฒดํ•˜๋Š” ๋“ฑ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ์ „๋žต์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์—์„œ์˜ ์ฆ๊ฐ•์ด ํ•„์ˆ˜์ ์ด๋ผ๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์—์„œ๋Š” ์ผ๋ถ€ ์ฆ๊ฐ•์ด ๋ถˆ๊ฐ€๋Šฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ฌธ์žฅ์„ ์ˆœ์—ดํ•˜๋ฉด ์–ธ์–ด ๋ชจ๋ธ์˜ ํ’ˆ์งˆ์ด ์ €ํ•˜๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ฆ๊ฐ• ์ค‘์— ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐ€์žฅ ์ž˜ ๋‹ค์‹œ ์ž‘์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์•„์ง ๋ถˆํ™•์‹คํ•ฉ๋‹ˆ๋‹ค. ์—ญ ํ›ˆ๋ จ์€ ์–ธ์–ด ๋ชจ๋ธ์— ์™ผ์ชฝ์—์„œ ์˜ค๋ฅธ์ชฝ์œผ๋กœ์˜ ์ž์—ฐ์–ด ๋ชจ๋ธ๋ง์ด๋ผ๋Š” ์ฃผ์š” ์ž‘์—…๊ณผ ๊ฐ„์„ญํ•˜์ง€ ์•Š๋„๋ก ํ•˜๊ธฐ ์œ„ํ•ด ์–ธ์–ด ๋ชจ๋ธ์— ๋ณ„๊ฐœ์˜ ์–ธ์–ด ์ž‘์—…(์—ญ์–ด)์„ ์ œ์‹œํ•จ์œผ๋กœ์จ ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค.
On top of this, we add an additional step of MinHash [8] de-duplication at the document-level. The parameters for MinHash are 20 hashes per signature, 20 buckets, and 1 row per bucket. These parameters are the default parameters in the spark implementation of MinHashLSH, and we did not do a hyperparameter sweep on these parameters due to compute limitations. Previous work has attempted running MinHash with much more aggressive parameters: Lee et al. [27] and Penedo et al. [39] use \(20\) buckets, \(450\) hashes per bucket, and \(9000\) signatures per hash. We conjecture that more aggressive MinHash would remove more templates, resulting in a higher-quality starting dataset, potentially making the SemDeDup step of D4 less necessary. Abbas et al. [1] did find that the performance of MinHash from Lee et al. [27] and SemDeDup are comparable at a fixed data selection ratio of 3.9% on C4, indicating that SemDeDup filters out similar data to aggressive MinHash does. We leave sweeping over these hyperparameters as future work.
์—ฌ๊ธฐ์— ๋ฌธ์„œ ์ˆ˜์ค€์—์„œ MinHash [8] ์ค‘๋ณต ์ œ๊ฑฐ ๋‹จ๊ณ„๋ฅผ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. MinHash์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” ์„œ๋ช…๋‹น 20๊ฐœ์˜ ํ•ด์‹œ, 20๊ฐœ์˜ ๋ฒ„ํ‚ท, ๋ฒ„ํ‚ท๋‹น 1๊ฐœ์˜ ํ–‰์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” MinHashLSH์˜ spark ๊ตฌํ˜„์—์„œ ๊ธฐ๋ณธ ๋งค๊ฐœ๋ณ€์ˆ˜์ด๋ฉฐ, ์ปดํ“จํŒ… ์ œํ•œ์œผ๋กœ ์ธํ•ด ์ด๋Ÿฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜์— ๋Œ€ํ•œ hyperparameter sweep์„ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” ํ›จ์”ฌ ๋” ๊ณต๊ฒฉ์ ์ธ ๋งค๊ฐœ๋ณ€์ˆ˜๋กœ MinHash๋ฅผ ์‹คํ–‰ํ•˜๋ ค๊ณ  ์‹œ๋„ํ–ˆ์Šต๋‹ˆ๋‹ค: Lee ๋“ฑ [27]๊ณผ Penedo ๋“ฑ [39]์€ ๋ฒ„ํ‚ท 20๊ฐœ, ๋ฒ„ํ‚ท๋‹น ํ•ด์‹œ 450๊ฐœ, ํ•ด์‹œ๋‹น ์„œ๋ช… 9000๊ฐœ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋” ๊ณต๊ฒฉ์ ์ธ MinHash๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋” ๋งŽ์€ ํ…œํ”Œ๋ฆฟ์„ ์ œ๊ฑฐํ•˜์—ฌ ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ์‹œ์ž‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ D4์˜ SemDeDup ๋‹จ๊ณ„๊ฐ€ ๋œ ํ•„์š”ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ถ”์ธกํ•ฉ๋‹ˆ๋‹ค. Abbas ๋“ฑ [1]์€ C4์—์„œ ๊ณ ์ •๋œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋น„์œจ 3.9%์—์„œ Lee ๋“ฑ [27]์˜ MinHash์™€ SemDeDup์˜ ์„ฑ๋Šฅ์ด ๋น„์Šทํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋Š”๋ฐ, ์ด๋Š” SemDeDup์ด ๊ณต๊ฒฉ์ ์ธ MinHash์™€ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฑธ๋Ÿฌ๋‚ธ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ hyperparameter๋ฅผ sweepํ•˜๋Š” ๊ฒƒ์€ ํ–ฅํ›„ ์ž‘์—…์œผ๋กœ ๋‚จ๊ฒจ๋‘์—ˆ์Šต๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์— ๋ฌธ์„œ ์ˆ˜์ค€์—์„œ MinHash [8] ์ค‘๋ณต ์ œ๊ฑฐ ๋‹จ๊ณ„๋ฅผ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. MinHash์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” ์„œ๋ช…๋‹น 20๊ฐœ์˜ ํ•ด์‹œ, 20๊ฐœ์˜ ๋ฒ„ํ‚ท, ๋ฒ„ํ‚ท๋‹น 1๊ฐœ์˜ ํ–‰์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” MinHashLSH์˜ ์ŠคํŒŒํฌ ๊ตฌํ˜„์—์„œ ๊ธฐ๋ณธ ๋งค๊ฐœ๋ณ€์ˆ˜์ด๋ฉฐ, ์ปดํ“จํŒ… ์ œํ•œ์œผ๋กœ ์ธํ•ด ์ด๋Ÿฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜์— ๋Œ€ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์Šค์œ•์„ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” ํ›จ์”ฌ ๋” ๊ณต๊ฒฉ์ ์ธ ๋งค๊ฐœ๋ณ€์ˆ˜๋กœ MinHash๋ฅผ ์‹คํ–‰ํ•˜๋ ค๊ณ  ์‹œ๋„ํ–ˆ์Šต๋‹ˆ๋‹ค: Lee ๋“ฑ [27]๊ณผ Penedo ๋“ฑ [39]์€ ๋ฒ„ํ‚ท 20๊ฐœ, ๋ฒ„ํ‚ท๋‹น ํ•ด์‹œ 450๊ฐœ, ํ•ด์‹œ๋‹น ์„œ๋ช… 9000๊ฐœ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋” ๊ณต๊ฒฉ์ ์ธ MinHash๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋” ๋งŽ์€ ํ…œํ”Œ๋ฆฟ์„ ์ œ๊ฑฐํ•˜์—ฌ ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ์‹œ์ž‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ D4์˜ SemDeDup ๋‹จ๊ณ„๊ฐ€ ๋œ ํ•„์š”ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ถ”์ธกํ•ฉ๋‹ˆ๋‹ค. Abbas ๋“ฑ [1]์€ C4์—์„œ ๊ณ ์ •๋œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋น„์œจ 3.9%์—์„œ Lee ๋“ฑ [27]์˜ MinHash์™€ SemDeDup์˜ ์„ฑ๋Šฅ์ด ๋น„์Šทํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋Š”๋ฐ, ์ด๋Š” SemDeDup์ด ๊ณต๊ฒฉ์ ์ธ MinHash์™€ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฑธ๋Ÿฌ๋‚ธ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์Šค์œ•ํ•˜๋Š” ๊ฒƒ์€ ํ–ฅํ›„ ์ž‘์—…์œผ๋กœ ๋‚จ๊ฒจ๋‘์—ˆ์Šต๋‹ˆ๋‹ค.
On the other hand, recent work has also shown promising results in mitigating hallucinations using synthetic data. For example, GPT-4 (OpenAI, 2023) was trained using a reward model that leveraged synthetic hallucination data in order to perform reinforcement learning (Zhang et al.,2023c). This method resulted in a significant improvement in performance on the TruthfulQA (Lin et al., 2022) dataset (Zhang et al., 2023c). Similarly, Jones et al. (2023) designed a synthetic task where hallucinations can be readily evaluated, utilizing this task to optimize LIM outputs by learning a continuous postfix via prefix-tuning. Tian et al. (2023) uses automated fact-checking and confidence scores to rank factuality scores of model response pairs, which are then used to finetune language models with DPO (Rafailov et al., 2023) to improve their factuality. Continued research in using synthetic data to mitigate hallucinations is still limited, however, by the lack of synthetic tasks for which hallucinations can be scalably evaluated.
๋ฐ˜๋ฉด, ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ hallucinations์„ ์™„ํ™”ํ•˜๋Š” ๋ฐ ์žˆ์–ด์„œ๋„ ์œ ๋งํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-4(OpenAI, 2023)๋Š” reinforcement learning์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ hallucination ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜๋Š” reward model์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Zhang et al.,2023c). ์ด ๋ฐฉ๋ฒ•์€ TruthfulQA(Lin et al., 2022) ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์™”์Šต๋‹ˆ๋‹ค(Zhang et al., 2023c). ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Jones et al. (2023)์€ hallucinations์„ ์‰ฝ๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š” ํ•ฉ์„ฑ ์ž‘์—…์„ ์„ค๊ณ„ํ•˜์—ฌ prefix-tuning์„ ํ†ตํ•ด ์—ฐ์† ์ ‘๋ฏธ์‚ฌ๋ฅผ ํ•™์Šตํ•˜์—ฌ LIM ์ถœ๋ ฅ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. Tian et al. (2023)์€ automated fact-checking ๋ฐ confidence scores๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ ์‘๋‹ต ์Œ์˜ factuality scores๋ฅผ ์ˆœ์œ„ํ™”ํ•œ ๋‹ค์Œ, ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ DPO(Rafailov et al., 2023)๋กœ language models๋ฅผ fine-tuneํ•˜์—ฌ ์‚ฌ์‹ค์„ฑ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. hallucinations์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์—ฐ๊ตฌ๋Š” ๊ณ„์† ์ œํ•œ์ ์ด์ง€๋งŒ, hallucinations์„ ํ™•์žฅ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š” ํ•ฉ์„ฑ ์ž‘์—…์ด ๋ถ€์กฑํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ œํ•œ์ ์ž…๋‹ˆ๋‹ค.
๋ฐ˜๋ฉด, ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ™˜๊ฐ์„ ์™„ํ™”ํ•˜๋Š” ๋ฐ ์žˆ์–ด์„œ๋„ ์œ ๋งํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-4(OpenAI, 2023)๋Š” ๊ฐ•ํ™” ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ํ™˜๊ฐ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜๋Š” ๋ณด์ƒ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Zhang et al.,2023c). ์ด ๋ฐฉ๋ฒ•์€ TruthfulQA(Lin et al., 2022) ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์™”์Šต๋‹ˆ๋‹ค(Zhang et al., 2023c). ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Jones et al. (2023)์€ ํ™˜๊ฐ์„ ์‰ฝ๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š” ํ•ฉ์„ฑ ์ž‘์—…์„ ์„ค๊ณ„ํ•˜์—ฌ ์ ‘๋‘์‚ฌ ํŠœ๋‹์„ ํ†ตํ•ด ์—ฐ์† ์ ‘๋ฏธ์‚ฌ๋ฅผ ํ•™์Šตํ•˜์—ฌ LIM ์ถœ๋ ฅ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. Tian et al. (2023)์€ ์ž๋™ํ™”๋œ ์‚ฌ์‹ค ํ™•์ธ ๋ฐ ์‹ ๋ขฐ๋„ ์ ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ ์‘๋‹ต ์Œ์˜ ์‚ฌ์‹ค์„ฑ ์ ์ˆ˜๋ฅผ ์ˆœ์œ„ํ™”ํ•œ ๋‹ค์Œ, ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ DPO(Rafailov et al., 2023)๋กœ ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์‚ฌ์‹ค์„ฑ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ํ™˜๊ฐ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์—ฐ๊ตฌ๋Š” ๊ณ„์† ์ œํ•œ์ ์ด์ง€๋งŒ, ํ™˜๊ฐ์„ ํ™•์žฅ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š” ํ•ฉ์„ฑ ์ž‘์—…์ด ๋ถ€์กฑํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ œํ•œ์ ์ž…๋‹ˆ๋‹ค.
\(\bullet\)_Human-based evaluation_. Unlike automatic evaluation for basic abilities, human evaluation typically considers more factors or abilities in real-world use, such as human alignment and tool manipulation. In this evaluation approach, test tasks are usually in the form of open-ended questions, and human evaluators are invited to make judgments on the quality of answers generated by LLMs. Typically, there are two main types of scoring methods for human evaluators: pairwise comparison and single-answer grading. In pairwise comparison, given the same question, humans are assigned two answers from different models to determine which one is better, while in single-answer grading, they only need to score a single answer at a time. For example, HELM [520] employs humans to perform single-answer grading on summarization and disinformation tasks, while Chatbot Arena [727] constructs a crowdsourcing platform that allows users to engage in conversations with two anonymous chat LLMs and report pairwise comparison results.
\(\bullet\)_์ธ๊ฐ„ ๊ธฐ๋ฐ˜ ํ‰๊ฐ€_. ๊ธฐ๋ณธ ๋Šฅ๋ ฅ์— ๋Œ€ํ•œ ์ž๋™ ํ‰๊ฐ€์™€ ๋‹ฌ๋ฆฌ, ์ธ๊ฐ„ ํ‰๊ฐ€๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„์˜ ์ •๋ ฌ ๋ฐ ๋„๊ตฌ ์กฐ์ž‘๊ณผ ๊ฐ™์€ ์‹ค์ œ ์‚ฌ์šฉ ์‹œ ๋” ๋งŽ์€ ์š”์†Œ ๋˜๋Š” ๋Šฅ๋ ฅ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด ํ‰๊ฐ€ ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ ํ…Œ์ŠคํŠธ ๊ณผ์ œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐœ๋ฐฉํ˜• ์งˆ๋ฌธ์˜ ํ˜•ํƒœ์ด๋ฉฐ, ์ธ๊ฐ„ ํ‰๊ฐ€์ž๋Š” LLM์ด ์ƒ์„ฑํ•œ ๋‹ต๋ณ€์˜ ํ’ˆ์งˆ์— ๋Œ€ํ•œ ํŒ๋‹จ์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„ ํ‰๊ฐ€์ž๋ฅผ ์œ„ํ•œ ์ฑ„์  ๋ฐฉ๋ฒ•์—๋Š” pairwise comparison์™€ single-answer grading์˜ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์œ ํ˜•์ด ์žˆ์Šต๋‹ˆ๋‹ค. pairwise comparison์—์„œ๋Š” ๋™์ผํ•œ ์งˆ๋ฌธ์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ์ธ๊ฐ„์€ ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ์—์„œ ๋‘ ๊ฐœ์˜ ๋‹ต๋ณ€์„ ํ• ๋‹น๋ฐ›์•„ ์–ด๋Š ์ชฝ์ด ๋” ๋‚˜์€์ง€ ๊ฒฐ์ •ํ•˜๋Š” ๋ฐ˜๋ฉด, single-answer grading์—์„œ๋Š” ํ•œ ๋ฒˆ์— ํ•˜๋‚˜์˜ ๋‹ต๋ณ€๋งŒ ์ฑ„์ ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, HELM [520]์€ ์š”์•ฝ ๋ฐ ํ—ˆ์œ„ ์ •๋ณด ์ž‘์—…์— ๋Œ€ํ•ด ์ธ๊ฐ„์ด single-answer grading์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ํ•˜๋ฉฐ, Chatbot Arena [727]์€ ์‚ฌ์šฉ์ž๊ฐ€ ๋‘ ๋ช…์˜ ์ต๋ช… ์ฑ„ํŒ… LLM๊ณผ ๋Œ€ํ™”์— ์ฐธ์—ฌํ•˜๊ณ  pairwise comparison ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•  ์ˆ˜ ์žˆ๋Š” ํฌ๋ผ์šฐ๋“œ์†Œ์‹ฑ ํ”Œ๋žซํผ์„ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_์ธ๊ฐ„ ๊ธฐ๋ฐ˜ ํ‰๊ฐ€_. ๊ธฐ๋ณธ ๋Šฅ๋ ฅ์— ๋Œ€ํ•œ ์ž๋™ ํ‰๊ฐ€์™€ ๋‹ฌ๋ฆฌ, ์ธ๊ฐ„ ํ‰๊ฐ€๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„์˜ ์ •๋ ฌ ๋ฐ ๋„๊ตฌ ์กฐ์ž‘๊ณผ ๊ฐ™์€ ์‹ค์ œ ์‚ฌ์šฉ ์‹œ ๋” ๋งŽ์€ ์š”์†Œ ๋˜๋Š” ๋Šฅ๋ ฅ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด ํ‰๊ฐ€ ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ ํ…Œ์ŠคํŠธ ๊ณผ์ œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐœ๋ฐฉํ˜• ์งˆ๋ฌธ์˜ ํ˜•ํƒœ์ด๋ฉฐ, ์ธ๊ฐ„ ํ‰๊ฐ€์ž๋Š” LLM์ด ์ƒ์„ฑํ•œ ๋‹ต๋ณ€์˜ ํ’ˆ์งˆ์— ๋Œ€ํ•œ ํŒ๋‹จ์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„ ํ‰๊ฐ€์ž๋ฅผ ์œ„ํ•œ ์ฑ„์  ๋ฐฉ๋ฒ•์—๋Š” ์Œ๋ณ„ ๋น„๊ต์™€ ๋‹จ์ผ ๋‹ต์•ˆ ์ฑ„์ ์˜ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์œ ํ˜•์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์Œ๋ณ„ ๋น„๊ต์—์„œ๋Š” ๋™์ผํ•œ ์งˆ๋ฌธ์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ์ธ๊ฐ„์€ ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ์—์„œ ๋‘ ๊ฐœ์˜ ๋‹ต๋ณ€์„ ํ• ๋‹น๋ฐ›์•„ ์–ด๋Š ์ชฝ์ด ๋” ๋‚˜์€์ง€ ๊ฒฐ์ •ํ•˜๋Š” ๋ฐ˜๋ฉด, ๋‹จ์ผ ๋‹ต๋ณ€ ์ฑ„์ ์—์„œ๋Š” ํ•œ ๋ฒˆ์— ํ•˜๋‚˜์˜ ๋‹ต๋ณ€๋งŒ ์ฑ„์ ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, HELM [520]์€ ์š”์•ฝ ๋ฐ ํ—ˆ์œ„ ์ •๋ณด ์ž‘์—…์— ๋Œ€ํ•ด ์ธ๊ฐ„์ด ๋‹จ์ผ ๋‹ต๋ณ€ ์ฑ„์ ์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ํ•˜๋ฉฐ, Chatbot Arena [727]์€ ์‚ฌ์šฉ์ž๊ฐ€ ๋‘ ๋ช…์˜ ์ต๋ช… ์ฑ„ํŒ… LLM๊ณผ ๋Œ€ํ™”์— ์ฐธ์—ฌํ•˜๊ณ  ์Œ๋ณ„ ๋น„๊ต ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•  ์ˆ˜ ์žˆ๋Š” ํฌ๋ผ์šฐ๋“œ์†Œ์‹ฑ ํ”Œ๋žซํผ์„ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค.
* Step 1. Generate a new math question. We request the GPT-4 Turbo API to generate a brand-new question using a reference math question as a starting point. To improve the validity of the new questions, we incorporate three rules into the prompt: Firstly, the new question must obey common knowledge; secondly, it should be solvable independently of the original question; and thirdly, it must not include any answer responses. Besides, we have set specific formatting requirements for questions and answers tailored to various target datasets. * Step 2. Verify the question. We further enhance the quality of the generated questions by validating and refining them through attempted solutions. By integrating solving and verification steps into a single prompt, we have found that this approach consistently elevates the validity of questions across different benchmarks. * Step 3. Generate chain-of-thought (CoT) answers. We request GPT-4 Turbo to produce a chain-of-thought (CoT) answer response for each newly generated question.
* 1๋‹จ๊ณ„. ์ƒˆ๋กœ์šด ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ฐธ์กฐ ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ์ถœ๋ฐœ์ ์œผ๋กœ ์‚ผ์•„ ์ƒˆ๋กœ์šด ๋ฌธ์ œ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก GPT-4 Turbo API๋ฅผ ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด ๋ฌธ์ œ์˜ ํƒ€๋‹น์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ์„ธ ๊ฐ€์ง€ ๊ทœ์น™์„ ํ”„๋กฌํ”„ํŠธ์— ํ†ตํ•ฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ƒˆ๋กœ์šด ๋ฌธ์ œ๋Š” ์ƒ์‹์„ ์ค€์ˆ˜ํ•ด์•ผ ํ•˜๋ฉฐ, ๋‘˜์งธ, ์›๋ž˜ ๋ฌธ์ œ์™€ ๋…๋ฆฝ์ ์œผ๋กœ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•˜๋ฉฐ, ์…‹์งธ, ์–ด๋–ค ๋‹ต๋ณ€๋„ ํฌํ•จํ•˜์ง€ ์•Š์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ ๋Œ€์ƒ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋งž๊ฒŒ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์— ๋Œ€ํ•œ ํŠน์ • ํ˜•์‹ ์š”๊ฑด์„ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. * 2๋‹จ๊ณ„. ๋ฌธ์ œ๋ฅผ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ํ’€์ด๋ฅผ ์‹œ๋„ํ•˜์—ฌ ๋ฌธ์ œ์˜ ํ’ˆ์งˆ์„ ํ•œ์ธต ๋” ๋†’์—ฌ์ค๋‹ˆ๋‹ค. ํ’€์ด ๋‹จ๊ณ„์™€ ๊ฒ€์ฆ ๋‹จ๊ณ„๋ฅผ ํ•˜๋‚˜์˜ ํ”„๋กฌํ”„ํŠธ์— ํ†ตํ•ฉํ•จ์œผ๋กœ์จ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ผ๊ด€๋˜๊ฒŒ ์งˆ๋ฌธ์˜ ํƒ€๋‹น์„ฑ์„ ๋†’์ผ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. * 3๋‹จ๊ณ„. Chain-of-Thought (CoT) ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ ์ƒ์„ฑ๋œ ๊ฐ ์งˆ๋ฌธ์— ๋Œ€ํ•ด GPT-4 Turbo์— Chain-of-Thought (CoT) ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค.
* 1๋‹จ๊ณ„. ์ƒˆ๋กœ์šด ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ฐธ์กฐ ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ์ถœ๋ฐœ์ ์œผ๋กœ ์‚ผ์•„ ์ƒˆ๋กœ์šด ๋ฌธ์ œ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก GPT-4 ํ„ฐ๋ณด API๋ฅผ ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด ๋ฌธ์ œ์˜ ํƒ€๋‹น์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ์„ธ ๊ฐ€์ง€ ๊ทœ์น™์„ ํ”„๋กฌํ”„ํŠธ์— ํ†ตํ•ฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ƒˆ๋กœ์šด ๋ฌธ์ œ๋Š” ์ƒ์‹์„ ์ค€์ˆ˜ํ•ด์•ผ ํ•˜๋ฉฐ, ๋‘˜์งธ, ์›๋ž˜ ๋ฌธ์ œ์™€ ๋…๋ฆฝ์ ์œผ๋กœ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•˜๋ฉฐ, ์…‹์งธ, ์–ด๋–ค ๋‹ต๋ณ€๋„ ํฌํ•จํ•˜์ง€ ์•Š์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ ๋Œ€์ƒ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋งž๊ฒŒ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์— ๋Œ€ํ•œ ํŠน์ • ํ˜•์‹ ์š”๊ฑด์„ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. * 2๋‹จ๊ณ„. ๋ฌธ์ œ๋ฅผ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ํ’€์ด๋ฅผ ์‹œ๋„ํ•˜์—ฌ ๋ฌธ์ œ์˜ ํ’ˆ์งˆ์„ ํ•œ์ธต ๋” ๋†’์—ฌ์ค๋‹ˆ๋‹ค. ํ’€์ด ๋‹จ๊ณ„์™€ ๊ฒ€์ฆ ๋‹จ๊ณ„๋ฅผ ํ•˜๋‚˜์˜ ํ”„๋กฌํ”„ํŠธ์— ํ†ตํ•ฉํ•จ์œผ๋กœ์จ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ผ๊ด€๋˜๊ฒŒ ์งˆ๋ฌธ์˜ ํƒ€๋‹น์„ฑ์„ ๋†’์ผ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. * 3๋‹จ๊ณ„. ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ(CoT) ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ ์ƒ์„ฑ๋œ ๊ฐ ์งˆ๋ฌธ์— ๋Œ€ํ•ด GPT-4 ํ„ฐ๋ณด์— ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ(CoT) ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค.
To study latent two-hop reasoning with diverse types of fact composition, we introduce TwoHopFact dataset, which is based on Wikidata (Vrandecic and Krotzsch, 2014) and consists of 45,595 two-hop prompts of 52 types of fact composition. We experiment with LLaMA-2 (Touvron et al., 2023) 7B, 13B, and 70B. Our findings can be summarized as follows. Across a wide range of fact composition types for the two-hop prompts, we find substantial evidence for the first hop of the multi-hop reasoning. In about 70% of the times where we change the prompt to indirectly mention the bridge entity, the later layers of the transformer show increased bridge entity recall. For the second hop and overall traversal, the evidence appears weaker: in 60% of the cases where we increase entity recall score, consistency goes up. Likewise, in about 40% of the time, both hops work together (compared to a random 25% baseline); changing the descriptive mention increases the entity recall score, and increasing this recall score increases consistency.
๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ์‚ฌ์‹ค ๊ตฌ์„ฑ์œผ๋กœ ์ž ์žฌ์ ์ธ 2-hop ์ถ”๋ก ์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” Wikidata(Vrandecic and Krotzsch, 2014)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ 52๊ฐ€์ง€ ์œ ํ˜•์˜ ์‚ฌ์‹ค ๊ตฌ์„ฑ์œผ๋กœ ๊ตฌ์„ฑ๋œ 45,595๊ฐœ์˜ 2-hop ํ”„๋กฌํ”„ํŠธ๋กœ ๊ตฌ์„ฑ๋œ TwoHopFact ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. LLaMA-2(Touvron et al., 2023) 7B, 13B, 70B๋กœ ์‹คํ—˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์š”์•ฝํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 2-hop ํ”„๋กฌํ”„ํŠธ์˜ ๋‹ค์–‘ํ•œ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์— ๊ฑธ์ณ multi-hop ์ถ”๋ก ์˜ ์ฒซ ๋ฒˆ์งธ hop์— ๋Œ€ํ•œ ์ƒ๋‹นํ•œ ์ฆ๊ฑฐ๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋‹ค๋ฆฌ ๊ฐœ์ฒด๋ฅผ ๊ฐ„์ ‘์ ์œผ๋กœ ์–ธ๊ธ‰ํ•˜๋„๋ก ๋ณ€๊ฒฝํ•œ ๊ฒฝ์šฐ ์•ฝ 70%์—์„œ Transformer์˜ ํ›„๊ธฐ ๋ ˆ์ด์–ด์—์„œ ๋‹ค๋ฆฌ ๊ฐœ์ฒด ํšŒ์ƒ์ด ์ฆ๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ hop๊ณผ ์ „์ฒด ํƒ์ƒ‰์˜ ๊ฒฝ์šฐ ์ฆ๊ฑฐ๊ฐ€ ์•ฝํ•ด ๋ณด์ด๋Š”๋ฐ, ๊ฐœ์ฒด ํšŒ์ƒ ์ ์ˆ˜๋ฅผ ๋†’์ธ ๊ฒฝ์šฐ์˜ 60%์—์„œ ์ผ๊ด€์„ฑ์ด ๋†’์•„์ง‘๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์•ฝ 40%์˜ ๊ฒฝ์šฐ ๋‘ hop์ด ํ•จ๊ป˜ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค(๋ฌด์ž‘์œ„ 25% ๊ธฐ์ค€ ๋Œ€๋น„). ์„ค๋ช…์  ์–ธ๊ธ‰์„ ๋ณ€๊ฒฝํ•˜๋ฉด ๊ฐœ์ฒด ํšŒ์ƒ ์ ์ˆ˜๊ฐ€ ์ƒ์Šนํ•˜๊ณ , ์ด ํšŒ์ƒ ์ ์ˆ˜๊ฐ€ ์ƒ์Šนํ•˜๋ฉด ์ผ๊ด€์„ฑ์ด ๋†’์•„์ง‘๋‹ˆ๋‹ค.
๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ์‚ฌ์‹ค ๊ตฌ์„ฑ์œผ๋กœ ์ž ์žฌ์ ์ธ 2ํ™‰ ์ถ”๋ก ์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์œ„ํ‚ค๋ฐ์ดํ„ฐ(Vrandecic and Krotzsch, 2014)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ 52๊ฐ€์ง€ ์œ ํ˜•์˜ ์‚ฌ์‹ค ๊ตฌ์„ฑ์œผ๋กœ ๊ตฌ์„ฑ๋œ 45,595๊ฐœ์˜ 2ํ™‰ ํ”„๋กฌํ”„ํŠธ๋กœ ๊ตฌ์„ฑ๋œ TwoHopFact ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. LLaMA-2(Touvron et al., 2023) 7B, 13B, 70B๋กœ ์‹คํ—˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์š”์•ฝํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 2ํ™‰ ํ”„๋กฌํ”„ํŠธ์˜ ๋‹ค์–‘ํ•œ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์— ๊ฑธ์ณ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์˜ ์ฒซ ๋ฒˆ์งธ ํ™‰์— ๋Œ€ํ•œ ์ƒ๋‹นํ•œ ์ฆ๊ฑฐ๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋‹ค๋ฆฌ ๊ฐœ์ฒด๋ฅผ ๊ฐ„์ ‘์ ์œผ๋กœ ์–ธ๊ธ‰ํ•˜๋„๋ก ๋ณ€๊ฒฝํ•œ ๊ฒฝ์šฐ ์•ฝ 70%์—์„œ ํŠธ๋žœ์Šคํฌ๋จธ์˜ ํ›„๊ธฐ ๋ ˆ์ด์–ด์—์„œ ๋‹ค๋ฆฌ ๊ฐœ์ฒด ํšŒ์ƒ์ด ์ฆ๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ํ™‰๊ณผ ์ „์ฒด ํƒ์ƒ‰์˜ ๊ฒฝ์šฐ ์ฆ๊ฑฐ๊ฐ€ ์•ฝํ•ด ๋ณด์ด๋Š”๋ฐ, ๊ฐœ์ฒด ํšŒ์ƒ ์ ์ˆ˜๋ฅผ ๋†’์ธ ๊ฒฝ์šฐ์˜ 60%์—์„œ ์ผ๊ด€์„ฑ์ด ๋†’์•„์ง‘๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์•ฝ 40%์˜ ๊ฒฝ์šฐ ๋‘ ํ™‰์ด ํ•จ๊ป˜ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค(๋ฌด์ž‘์œ„ 25% ๊ธฐ์ค€ ๋Œ€๋น„). ์„ค๋ช…์  ์–ธ๊ธ‰์„ ๋ณ€๊ฒฝํ•˜๋ฉด ๊ฐœ์ฒด ํšŒ์ƒ ์ ์ˆ˜๊ฐ€ ์ƒ์Šนํ•˜๊ณ , ์ด ํšŒ์ƒ ์ ์ˆ˜๊ฐ€ ์ƒ์Šนํ•˜๋ฉด ์ผ๊ด€์„ฑ์ด ๋†’์•„์ง‘๋‹ˆ๋‹ค.
**Remaining Issues.** Despite the huge success, there are still several issues that limit the development and applications of LLM-based agents. First, with the explosive growth of the model scale, the efficiency of LLM-based agents, including both the time and memory overhead, becomes an important issue for large-scale deployment, especially for multi-agent systems with numerous instances of LLMs. Second, with the scaling of the number of LLM-based agents, more effective and efficient communication protocols and architectures are required to support the increased complexity of coordination among agents. Furthermore, building capable agents poses technical challenges for the capacities of LLMs like instruction following and long text modeling. Since existing LLMs are not specially optimized for instantiating agents, most public-sourced LLMs like LLaMA cannot effectively facilitate the development of agents. Therefore, it is crucial to develop capable, specialized models to serve as the core computation unit of agents.
**ํ˜„์•ˆ**. ๋ง‰๋Œ€ํ•œ ์„ฑ๊ณต์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ๊ฐœ๋ฐœ๊ณผ ์ ์šฉ์„ ์ œํ•œํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฌธ์ œ๊ฐ€ ์—ฌ์ „ํžˆ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ๊ธ‰์ฆํ•จ์— ๋”ฐ๋ผ ์‹œ๊ฐ„ ๋ฐ ๋ฉ”๋ชจ๋ฆฌ ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ํฌํ•จํ•œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ํšจ์œจ์„ฑ์€ ๋Œ€๊ทœ๋ชจ ๋ฐฐํฌ, ํŠนํžˆ ์ˆ˜๋งŽ์€ LLM ์ธ์Šคํ„ด์Šค๊ฐ€ ์žˆ๋Š” ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์˜ ๊ฒฝ์šฐ ์ค‘์š”ํ•œ ๋ฌธ์ œ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ์ˆ˜๊ฐ€ ํ™•์žฅ๋จ์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ ๊ฐ„์˜ ์กฐ์ • ๋ณต์žก์„ฑ ์ฆ๊ฐ€๋ฅผ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด ๋ณด๋‹ค ํšจ๊ณผ์ ์ด๊ณ  ํšจ์œจ์ ์ธ ํ†ต์‹  ํ”„๋กœํ† ์ฝœ๊ณผ ์•„ํ‚คํ…์ฒ˜๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์œ ๋Šฅํ•œ ์—์ด์ „ํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋ ค๋ฉด Instruction following ๋ฐ ๊ธด ํ…์ŠคํŠธ ๋ชจ๋ธ๋ง๊ณผ ๊ฐ™์€ LLM์˜ ์—ญ๋Ÿ‰์— ๋Œ€ํ•œ ๊ธฐ์ˆ ์  ๊ณผ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด LLM์€ ์—์ด์ „ํŠธ๋ฅผ ์ธ์Šคํ„ด์Šคํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ์ตœ์ ํ™”๋˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์—, ๋Œ€๋ถ€๋ถ„์˜ ๊ณต๊ฐœ ์†Œ์Šค LLM(์˜ˆ: LLaMA)์€ ์—์ด์ „ํŠธ ๊ฐœ๋ฐœ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ด‰์ง„ํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—์ด์ „ํŠธ์˜ ํ•ต์‹ฌ ๊ณ„์‚ฐ ๋‹จ์œ„ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์ด ์žˆ๋Š” ํŠน์ˆ˜ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
**ํ˜„์•ˆ**. ๋ง‰๋Œ€ํ•œ ์„ฑ๊ณต์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ๊ฐœ๋ฐœ๊ณผ ์ ์šฉ์„ ์ œํ•œํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฌธ์ œ๊ฐ€ ์—ฌ์ „ํžˆ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ๊ธ‰์ฆํ•จ์— ๋”ฐ๋ผ ์‹œ๊ฐ„ ๋ฐ ๋ฉ”๋ชจ๋ฆฌ ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ํฌํ•จํ•œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ํšจ์œจ์„ฑ์€ ๋Œ€๊ทœ๋ชจ ๋ฐฐํฌ, ํŠนํžˆ ์ˆ˜๋งŽ์€ LLM ์ธ์Šคํ„ด์Šค๊ฐ€ ์žˆ๋Š” ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์˜ ๊ฒฝ์šฐ ์ค‘์š”ํ•œ ๋ฌธ์ œ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ์ˆ˜๊ฐ€ ํ™•์žฅ๋จ์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ ๊ฐ„์˜ ์กฐ์ • ๋ณต์žก์„ฑ ์ฆ๊ฐ€๋ฅผ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด ๋ณด๋‹ค ํšจ๊ณผ์ ์ด๊ณ  ํšจ์œจ์ ์ธ ํ†ต์‹  ํ”„๋กœํ† ์ฝœ๊ณผ ์•„ํ‚คํ…์ฒ˜๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์œ ๋Šฅํ•œ ์—์ด์ „ํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋ ค๋ฉด ๋ช…๋ น์–ด ์ถ”์ข… ๋ฐ ๊ธด ํ…์ŠคํŠธ ๋ชจ๋ธ๋ง๊ณผ ๊ฐ™์€ LLM์˜ ์—ญ๋Ÿ‰์— ๋Œ€ํ•œ ๊ธฐ์ˆ ์  ๊ณผ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด LLM์€ ์—์ด์ „ํŠธ๋ฅผ ์ธ์Šคํ„ด์Šคํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ์ตœ์ ํ™”๋˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์—, ๋Œ€๋ถ€๋ถ„์˜ ๊ณต๊ฐœ ์†Œ์Šค LLM(์˜ˆ: LLaMA)์€ ์—์ด์ „ํŠธ ๊ฐœ๋ฐœ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ด‰์ง„ํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—์ด์ „ํŠธ์˜ ํ•ต์‹ฌ ๊ณ„์‚ฐ ๋‹จ์œ„ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์ด ์žˆ๋Š” ํŠน์ˆ˜ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
Similar to prior work Xu et al. (2023); Zhao et al. (2023), we compute the total ratio of wins and ties of a target LLM against the strong LLM, to indicate how much model capacity the target LLM recovers from the strong LLM (often treated as the upper bound performer). CRR simplifies the combinatorial pairwise comparisons between all target LLMs. We name the metric as _Capacity Recovery Ratio_ (CRR), where \(\texttt{CRR}=\frac{\texttt{wins}+\texttt{ties}}{\texttt{total comparisons}}\). In experiments, we observe that the number of ties often dominates the number of wins, since the strong LLM is much capable than the target model. So we do not put additional weights on wins in the calculation. To demonstrate CRR faithfully reflects model performance, we show the exact number of wins, ties and losses in Appendix A.5 on Evol-Instruct. We would like to emphasize our focus on the gap in CRR between different methods instead of the absolute value, since the absolute value may based on the specific LLM evaluator we choose.
์ด์ „ ์ž‘์—… Xu et al. (2023); Zhao et al. (2023)๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ, ๊ฐ•๋ ฅํ•œ LLM์— ๋Œ€ํ•œ ๋Œ€์ƒ LLM์˜ ์ด ์Šน๋ฆฌ ๋ฐ ๋™์  ๋น„์œจ์„ ๊ณ„์‚ฐํ•˜์—ฌ ๋Œ€์ƒ LLM์ด ๊ฐ•๋ ฅํ•œ LLM์—์„œ ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ํšŒ๋ณตํ–ˆ๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค(์ข…์ข… ์ƒํ•œ์„  ์ˆ˜ํ–‰์ž๋กœ ์ทจ๊ธ‰๋จ). CRR์€ ๋ชจ๋“  ๋Œ€์ƒ LLM ๊ฐ„์˜ ์กฐํ•ฉ์  ์Œ๋ณ„ ๋น„๊ต๋ฅผ ๋‹จ์ˆœํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฉ”ํŠธ๋ฆญ์„ _Capacity Recovery Ratio_ (CRR)์ด๋ผ๊ณ  ํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(\texttt{CRR}=\frac{\texttt{wins}+\texttt{ties}}{\texttt{total comparisons}}\)์ž…๋‹ˆ๋‹ค. ์‹คํ—˜์—์„œ ๊ฐ•๋ ฅํ•œ LLM์ด ๋Œ€์ƒ ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚˜๊ธฐ ๋•Œ๋ฌธ์— ๋™์  ์ˆ˜๊ฐ€ ์Šน๋ฆฌ ์ˆ˜๋ฅผ ์ง€๋ฐฐํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ณ„์‚ฐ์—์„œ ์Šน๋ฆฌ์— ์ถ”๊ฐ€ ๊ฐ€์ค‘์น˜๋ฅผ ๋‘์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. CRR์ด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ์ถฉ์‹คํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๋Š”์ง€ ์ž…์ฆํ•˜๊ธฐ ์œ„ํ•ด ๋ถ€๋ก A.5์—์„œ Evol-Instruct์— ๋Œ€ํ•œ ์ •ํ™•ํ•œ ์Šน๋ฆฌ, ๋™์  ๋ฐ ํŒจ๋ฐฐ ํšŸ์ˆ˜๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ ˆ๋Œ€๊ฐ’์€ ์šฐ๋ฆฌ๊ฐ€ ์„ ํƒํ•œ ํŠน์ • LLM ํ‰๊ฐ€์ž์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ ˆ๋Œ€๊ฐ’ ๋Œ€์‹  ์—ฌ๋Ÿฌ ๋ฐฉ๋ฒ• ๊ฐ„์˜ CRR ๊ฒฉ์ฐจ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.
์ด์ „ ์ž‘์—… Xu et al. (2023); Zhao et al. (2023)๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ, ๊ฐ•๋ ฅํ•œ LLM์— ๋Œ€ํ•œ ๋Œ€์ƒ LLM์˜ ์ด ์Šน๋ฆฌ ๋ฐ ๋™์  ๋น„์œจ์„ ๊ณ„์‚ฐํ•˜์—ฌ ๋Œ€์ƒ LLM์ด ๊ฐ•๋ ฅํ•œ LLM์—์„œ ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ํšŒ๋ณตํ–ˆ๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค(์ข…์ข… ์ƒํ•œ์„  ์ˆ˜ํ–‰์ž๋กœ ์ทจ๊ธ‰๋จ). CRR์€ ๋ชจ๋“  ๋Œ€์ƒ LLM ๊ฐ„์˜ ์กฐํ•ฉ์  ์Œ๋ณ„ ๋น„๊ต๋ฅผ ๋‹จ์ˆœํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฉ”ํŠธ๋ฆญ์„ _์šฉ๋Ÿ‰ ๋ณต๊ตฌ ๋น„์œจ_(CRR)์ด๋ผ๊ณ  ํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(\texttt{CRR}=\frac{\texttt{wins}+\texttt{ties}}{\texttt{total comparisons}}\)์ž…๋‹ˆ๋‹ค. ์‹คํ—˜์—์„œ ๊ฐ•๋ ฅํ•œ LLM์ด ๋Œ€์ƒ ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚˜๊ธฐ ๋•Œ๋ฌธ์— ๋™์  ์ˆ˜๊ฐ€ ์Šน๋ฆฌ ์ˆ˜๋ฅผ ์ง€๋ฐฐํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ณ„์‚ฐ์—์„œ ์Šน๋ฆฌ์— ์ถ”๊ฐ€ ๊ฐ€์ค‘์น˜๋ฅผ ๋‘์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. CRR์ด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ์ถฉ์‹คํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๋Š”์ง€ ์ž…์ฆํ•˜๊ธฐ ์œ„ํ•ด ๋ถ€๋ก A.5์—์„œ Evol-Instruct์— ๋Œ€ํ•œ ์ •ํ™•ํ•œ ์Šน๋ฆฌ, ๋™์  ๋ฐ ํŒจ๋ฐฐ ํšŸ์ˆ˜๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ ˆ๋Œ€๊ฐ’์€ ์šฐ๋ฆฌ๊ฐ€ ์„ ํƒํ•œ ํŠน์ • LLM ํ‰๊ฐ€์ž์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ ˆ๋Œ€๊ฐ’ ๋Œ€์‹  ์—ฌ๋Ÿฌ ๋ฐฉ๋ฒ• ๊ฐ„์˜ CRR ๊ฒฉ์ฐจ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_The effect of CoT prompting components._ The major distinction between CoT prompting and standard prompting is the incorporation of reasoning paths prior to the final answer. Thus, some researchers investigate the effects of different components in the reasoning paths. Specifically, a recent study identifies three key components in CoT prompting, namely _symbols_ (_e.g.,_ numerical quantities in arithmetic reasoning), _patterns_ (_e.g.,_ equations in arithmetic reasoning), and _text_ (_i.e.,_ the rest of tokens that are not symbols or patterns) [522]. It is shown that the latter two parts (_i.e.,_ patterns and text) are essential to the model performance, and removing either one would lead to a significant performance drop. However, the correctness of symbols and patterns does not seem critical. Further, there exists a symbiotic relationship between text and patterns: the text helps LLMs to generate useful patterns, and patterns aid LLMs to understand tasks and generate texts that help solve them [522].
\(\bullet\)_CoT prompting ๊ตฌ์„ฑ ์š”์†Œ์˜ ํšจ๊ณผ._ CoT prompting๊ณผ ํ‘œ์ค€ prompting์˜ ์ฃผ์š” ์ฐจ์ด์ ์€ ์ตœ์ข… ๋‹ต๋ณ€์— ์•ž์„œ reasoning ๊ฒฝ๋กœ๋ฅผ ํ†ตํ•ฉํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ผ๋ถ€ ์—ฐ๊ตฌ์ž๋“ค์€ reasoning ๊ฒฝ๋กœ์˜ ๋‹ค์–‘ํ•œ ๊ตฌ์„ฑ ์š”์†Œ์˜ ํšจ๊ณผ๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” CoT prompting์— ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ, ์ฆ‰ _๊ธฐํ˜ธ_(_์˜ˆ: arithmetic reasoning์˜ ์ˆ˜๋Ÿ‰), _ํŒจํ„ด_(_์˜ˆ: arithmetic reasoning์˜ ๋ฐฉ์ •์‹), _ํ…์ŠคํŠธ_(_์ฆ‰, ๊ธฐํ˜ธ๋‚˜ ํŒจํ„ด์ด ์•„๋‹Œ ๋‚˜๋จธ์ง€ ํ† ํฐ) [522]๋ฅผ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ํ›„์ž์˜ ๋‘ ๋ถ€๋ถ„(์ฆ‰, ํŒจํ„ด๊ณผ ํ…์ŠคํŠธ)์€ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ํ•„์ˆ˜์ ์ด๋ฉฐ, ์–ด๋Š ํ•˜๋‚˜๋ผ๋„ ์ œ๊ฑฐํ•˜๋ฉด ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ์ €ํ•˜๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐํ˜ธ์™€ ํŒจํ„ด์˜ ์ •ํ™•์„ฑ์€ ์ค‘์š”ํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ํ…์ŠคํŠธ์™€ ํŒจํ„ด ์‚ฌ์ด์—๋Š” ๊ณต์ƒ ๊ด€๊ณ„๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ํ…์ŠคํŠธ๋Š” LLM์ด ์œ ์šฉํ•œ ํŒจํ„ด์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๊ณ , ํŒจํ„ด์€ LLM์ด ์ž‘์—…์„ ์ดํ•ดํ•˜๊ณ  ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค[522].
\(\bullet\)_CoT ํ”„๋กฌํ”„ํŠธ ๊ตฌ์„ฑ ์š”์†Œ์˜ ํšจ๊ณผ._ CoT ํ”„๋กฌํ”„ํŠธ์™€ ํ‘œ์ค€ ํ”„๋กฌํ”„ํŠธ์˜ ์ฃผ์š” ์ฐจ์ด์ ์€ ์ตœ์ข… ๋‹ต๋ณ€์— ์•ž์„œ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ํ†ตํ•ฉํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ผ๋ถ€ ์—ฐ๊ตฌ์ž๋“ค์€ ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ๋‹ค์–‘ํ•œ ๊ตฌ์„ฑ ์š”์†Œ์˜ ํšจ๊ณผ๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” CoT ํ”„๋กฌํ”„ํŠธ์— ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ, ์ฆ‰ _๊ธฐํ˜ธ_(_์˜ˆ: ์‚ฐ์ˆ  ์ถ”๋ก ์˜ ์ˆ˜๋Ÿ‰), _ํŒจํ„ด_(_์˜ˆ: ์‚ฐ์ˆ  ์ถ”๋ก ์˜ ๋ฐฉ์ •์‹), _ํ…์ŠคํŠธ_(_์ฆ‰, ๊ธฐํ˜ธ๋‚˜ ํŒจํ„ด์ด ์•„๋‹Œ ๋‚˜๋จธ์ง€ ํ† ํฐ) [522]๋ฅผ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ํ›„์ž์˜ ๋‘ ๋ถ€๋ถ„(์ฆ‰, ํŒจํ„ด๊ณผ ํ…์ŠคํŠธ)์€ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ํ•„์ˆ˜์ ์ด๋ฉฐ, ์–ด๋Š ํ•˜๋‚˜๋ผ๋„ ์ œ๊ฑฐํ•˜๋ฉด ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ์ €ํ•˜๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐํ˜ธ์™€ ํŒจํ„ด์˜ ์ •ํ™•์„ฑ์€ ์ค‘์š”ํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ํ…์ŠคํŠธ์™€ ํŒจํ„ด ์‚ฌ์ด์—๋Š” ๊ณต์ƒ ๊ด€๊ณ„๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ํ…์ŠคํŠธ๋Š” LLM์ด ์œ ์šฉํ•œ ํŒจํ„ด์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๊ณ , ํŒจํ„ด์€ LLM์ด ์ž‘์—…์„ ์ดํ•ดํ•˜๊ณ  ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค[522].
Safety.Red teaming is a powerful technique for evaluating the safety and robustness of AI models (Casper et al., 2023b; Ganguli et al., 2022). By generating diverse and realistic scenarios designed to elicit unaligned or harmful outputs (Casper et al., 2023a), red teaming can expose vulnerabilities and weaknesses in AI systems (Perez et al., 2022). For example, Perez et al. (2023) use LMs to generate datasets for evaluating the behavior of other LMs. They end up producing 154 high-quality datasets which are verified by humans, and discover new cases of inverse scaling where LMs get worse with size. Hubinger et al. (2024) leverage synthetic data to trigger backdoor attacks to LMs at scale; they find LMs can exhibit deceptive behavior and create a false impression of safety under such attacks, and standard "safety training" could not remove such deception easily. These methods demonstrate the feasibility of using AI assistance to scale up human oversight (Bowman et al., 2022) over complex problems and unseen domains.
์•ˆ์ „์„ฑ. Red teaming์€ AI ๋ชจ๋ธ์˜ ์•ˆ์ „์„ฑ๊ณผ robustness๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค(Casper ์™ธ., 2023b; Ganguli ์™ธ., 2022). ์ •๋ ฌ๋˜์ง€ ์•Š๊ฑฐ๋‚˜ ์œ ํ•ดํ•œ ๊ฒฐ๊ณผ๋ฅผ ์ด๋Œ์–ด๋‚ด๋„๋ก ์„ค๊ณ„๋œ ๋‹ค์–‘ํ•˜๊ณ  ์‚ฌ์‹ค์ ์ธ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์ƒ์„ฑํ•จ์œผ๋กœ์จ Red teaming์€ AI ์‹œ์Šคํ…œ์˜ ์ทจ์•ฝ์ ๊ณผ ์•ฝ์ ์„ ๋…ธ์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Perez ์™ธ., 2022). ์˜ˆ๋ฅผ ๋“ค์–ด, Perez ๋“ฑ(2023)์€ LMs๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค๋ฅธ LMs์˜ ๋™์ž‘์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ๊ฒฐ๊ตญ ์ธ๊ฐ„์ด ๊ฒ€์ฆํ•œ 154๊ฐœ์˜ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์‚ฐํ•˜๊ณ , LMs๊ฐ€ ํฌ๊ธฐ๊ฐ€ ์ปค์งˆ์ˆ˜๋ก ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๋Š” inverse scaling์˜ ์ƒˆ๋กœ์šด ์‚ฌ๋ก€๋ฅผ ๋ฐœ๊ฒฌํ•ฉ๋‹ˆ๋‹ค. Hubinger ์™ธ(2024)๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋Œ€๊ทœ๋ชจ๋กœ LMs์— ๋ฐฑ๋„์–ด ๊ณต๊ฒฉ์„ ํŠธ๋ฆฌ๊ฑฐํ•˜๊ณ , ๊ทธ ๊ฒฐ๊ณผ LMs๊ฐ€ ์ด๋Ÿฌํ•œ ๊ณต๊ฒฉ ํ•˜์—์„œ ์†์ด๋Š” ํ–‰๋™์„ ๋ณด์ด๋ฉฐ ์•ˆ์ „์— ๋Œ€ํ•œ ์ž˜๋ชป๋œ ์ธ์ƒ์„ ์‹ฌ์–ด์ค„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํ‘œ์ค€ "safety training"์œผ๋กœ๋Š” ์ด๋Ÿฌํ•œ ์†์ž„์ˆ˜๋ฅผ ์‰ฝ๊ฒŒ ์ œ๊ฑฐํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์€ ๋ณต์žกํ•œ ๋ฌธ์ œ์™€ ๋ณด์ด์ง€ ์•Š๋Š” ์˜์—ญ์— ๋Œ€ํ•œ ์ธ๊ฐ„์˜ ๊ฐ๋…์„ ํ™•์žฅํ•˜๋Š” ๋ฐ AI ์ง€์›์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€๋Šฅํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(Bowman ์™ธ., 2022).
์•ˆ์ „์„ฑ.๋ ˆ๋“œ ํŒ€์€ AI ๋ชจ๋ธ์˜ ์•ˆ์ „์„ฑ๊ณผ ๊ฒฌ๊ณ ์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค(Casper ์™ธ., 2023b; Ganguli ์™ธ., 2022). ์ •๋ ฌ๋˜์ง€ ์•Š๊ฑฐ๋‚˜ ์œ ํ•ดํ•œ ๊ฒฐ๊ณผ๋ฅผ ์ด๋Œ์–ด๋‚ด๋„๋ก ์„ค๊ณ„๋œ ๋‹ค์–‘ํ•˜๊ณ  ์‚ฌ์‹ค์ ์ธ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์ƒ์„ฑํ•จ์œผ๋กœ์จ ๋ ˆ๋“œ ํŒ€์€ AI ์‹œ์Šคํ…œ์˜ ์ทจ์•ฝ์ ๊ณผ ์•ฝ์ ์„ ๋…ธ์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Perez ์™ธ., 2022). ์˜ˆ๋ฅผ ๋“ค์–ด, Perez ๋“ฑ(2023)์€ LM์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค๋ฅธ LM์˜ ๋™์ž‘์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ๊ฒฐ๊ตญ ์ธ๊ฐ„์ด ๊ฒ€์ฆํ•œ 154๊ฐœ์˜ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์‚ฐํ•˜๊ณ , LM์ด ํฌ๊ธฐ๊ฐ€ ์ปค์งˆ์ˆ˜๋ก ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๋Š” ์—ญ๊ทœ๋ชจ์˜ ์ƒˆ๋กœ์šด ์‚ฌ๋ก€๋ฅผ ๋ฐœ๊ฒฌํ•ฉ๋‹ˆ๋‹ค. Hubinger ์™ธ(2024)๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋Œ€๊ทœ๋ชจ๋กœ LM์— ๋ฐฑ๋„์–ด ๊ณต๊ฒฉ์„ ํŠธ๋ฆฌ๊ฑฐํ•˜๊ณ , ๊ทธ ๊ฒฐ๊ณผ LM์ด ์ด๋Ÿฌํ•œ ๊ณต๊ฒฉ ํ•˜์—์„œ ์†์ด๋Š” ํ–‰๋™์„ ๋ณด์ด๋ฉฐ ์•ˆ์ „์— ๋Œ€ํ•œ ์ž˜๋ชป๋œ ์ธ์ƒ์„ ์‹ฌ์–ด์ค„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํ‘œ์ค€ "์•ˆ์ „ ๊ต์œก"์œผ๋กœ๋Š” ์ด๋Ÿฌํ•œ ์†์ž„์ˆ˜๋ฅผ ์‰ฝ๊ฒŒ ์ œ๊ฑฐํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์€ ๋ณต์žกํ•œ ๋ฌธ์ œ์™€ ๋ณด์ด์ง€ ์•Š๋Š” ์˜์—ญ์— ๋Œ€ํ•œ ์ธ๊ฐ„์˜ ๊ฐ๋…์„ ํ™•์žฅํ•˜๋Š” ๋ฐ AI ์ง€์›์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€๋Šฅํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(Bowman ์™ธ., 2022).
\(\bullet\)_Conversation text._ Conversation data can enhance the conversational competence of LLMs [90] and potentially improve their performance on a range of question-answering tasks [56]. Researchers can utilize subsets of public conversation corpus (_e.g.,_ PushShift.io Reddit corpus) [202, 158] or collect conversation data from online social media. Since online conversational data often involves discussions among multiple participants, an effective processing way is to transform a conversation into a tree structure, where the utterance is linked to the one it responds to. In this way, the multi-party conversation tree can be divided into multiple sub-conversations, which can be collected in the pre-training corpus. Furthermore, a potential risk is that the excessive integration of dialogue data into LLMs may result in a side effect [90]: declarative instructions and direct interrogatives are erroneously perceived as the beginning of conversations, thus leading to a decline in the efficacy of the instructions.
\(\bullet\)_๋Œ€ํ™” ํ…์ŠคํŠธ._ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ๋Š” LLMs์˜ ๋Œ€ํ™” ์—ญ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  [90] ๋‹ค์–‘ํ•œ ์งˆ์˜์‘๋‹ต ์ž‘์—…์—์„œ ์ž ์žฌ์ ์œผ๋กœ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [56]. ์—ฐ๊ตฌ์ž๋“ค์€ ๊ณต๊ฐœ ๋Œ€ํ™” ์ฝ”ํผ์Šค์˜ ํ•˜์œ„ ์ง‘ํ•ฉ(_์˜ˆ:_ PushShift.io Reddit ์ฝ”ํผ์Šค) [202, 158]์„ ํ™œ์šฉํ•˜๊ฑฐ๋‚˜ ์˜จ๋ผ์ธ ์†Œ์…œ ๋ฏธ๋””์–ด์—์„œ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜จ๋ผ์ธ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ์—๋Š” ์ข…์ข… ์—ฌ๋Ÿฌ ์ฐธ๊ฐ€์ž ๊ฐ„์˜ ํ† ๋ก ์ด ํฌํ•จ๋˜๋ฏ€๋กœ ๋Œ€ํ™”๋ฅผ ํŠธ๋ฆฌ ๊ตฌ์กฐ๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ ๋ฐœํ™”๋ฅผ ์‘๋‹ตํ•˜๋Š” ๋ฐœํ™”์— ์—ฐ๊ฒฐํ•˜๋Š” ๊ฒƒ์ด ํšจ๊ณผ์ ์ธ ์ฒ˜๋ฆฌ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ๋‹ค์ž ๋Œ€ํ™” ํŠธ๋ฆฌ๋Š” ์—ฌ๋Ÿฌ ํ•˜์œ„ ๋Œ€ํ™”๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด ํ•˜์œ„ ๋Œ€ํ™”๋Š” ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์— ์ˆ˜์ง‘๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋Œ€ํ™” ๋ฐ์ดํ„ฐ๋ฅผ LLMs์— ๊ณผ๋„ํ•˜๊ฒŒ ํ†ตํ•ฉํ•˜๋ฉด ์„ ์–ธ์  instructions์™€ ์ง์ ‘์ ์ธ ์งˆ๋ฌธ์ด ๋Œ€ํ™”์˜ ์‹œ์ž‘์œผ๋กœ ์ž˜๋ชป ์ธ์‹๋˜์–ด instructions์˜ ํšจ์œจ์„ฑ์ด ์ €ํ•˜๋˜๋Š” ๋ถ€์ž‘์šฉ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [90].
\(\bullet\)_๋Œ€ํ™” ํ…์ŠคํŠธ._ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ๋Š” LLM์˜ ๋Œ€ํ™” ์—ญ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  [90] ๋‹ค์–‘ํ•œ ์งˆ์˜์‘๋‹ต ์ž‘์—…์—์„œ ์ž ์žฌ์ ์œผ๋กœ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [56]. ์—ฐ๊ตฌ์ž๋“ค์€ ๊ณต๊ฐœ ๋Œ€ํ™” ์ฝ”ํผ์Šค์˜ ํ•˜์œ„ ์ง‘ํ•ฉ(_์˜ˆ:_ PushShift.io Reddit ์ฝ”ํผ์Šค) [202, 158]์„ ํ™œ์šฉํ•˜๊ฑฐ๋‚˜ ์˜จ๋ผ์ธ ์†Œ์…œ ๋ฏธ๋””์–ด์—์„œ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜จ๋ผ์ธ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ์—๋Š” ์ข…์ข… ์—ฌ๋Ÿฌ ์ฐธ๊ฐ€์ž ๊ฐ„์˜ ํ† ๋ก ์ด ํฌํ•จ๋˜๋ฏ€๋กœ ๋Œ€ํ™”๋ฅผ ํŠธ๋ฆฌ ๊ตฌ์กฐ๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ ๋ฐœํ™”๋ฅผ ์‘๋‹ตํ•˜๋Š” ๋ฐœํ™”์— ์—ฐ๊ฒฐํ•˜๋Š” ๊ฒƒ์ด ํšจ๊ณผ์ ์ธ ์ฒ˜๋ฆฌ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ ๋‹ค์ž ๋Œ€ํ™” ํŠธ๋ฆฌ๋Š” ์—ฌ๋Ÿฌ ํ•˜์œ„ ๋Œ€ํ™”๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด ํ•˜์œ„ ๋Œ€ํ™”๋Š” ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์— ์ˆ˜์ง‘๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋Œ€ํ™” ๋ฐ์ดํ„ฐ๋ฅผ LLM์— ๊ณผ๋„ํ•˜๊ฒŒ ํ†ตํ•ฉํ•˜๋ฉด ์„ ์–ธ์  ์ง€์‹œ์™€ ์ง์ ‘์ ์ธ ์งˆ๋ฌธ์ด ๋Œ€ํ™”์˜ ์‹œ์ž‘์œผ๋กœ ์ž˜๋ชป ์ธ์‹๋˜์–ด ์ง€์‹œ์˜ ํšจ์œจ์„ฑ์ด ์ €ํ•˜๋˜๋Š” ๋ถ€์ž‘์šฉ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [90].
To apply Lemma F.1, we define \(\mathcal{Q}_{1}=[N_{0}]\), \(\mathcal{Q}_{2}=[N_{0}-1]\), and continue until \(\mathcal{Q}_{N}=[N_{0}-N+1]\). After uniformly randomly generating \(Q_{1},\ldots,Q_{N}\) from \(\mathcal{Q}_{1},\ldots,\mathcal{Q}_{N}\), we construct \((g_{1},\ldots,g_{N})\in[N]^{N}\) as follows: set \(g_{1}=Q_{1}\); for \(g_{2}\), set it to \(Q_{2}\) if \(Q_{2}<Q_{1}\), otherwise \(g_{2}=Q_{2}+1\); and in general, define \(g_{i}\) as the \(Q_{i}\)-th smallest element in \([N_{0}]\setminus\{g_{1},\ldots,g_{i-1}\}\). This method provides an alternative way to generate \(\mathcal{Z}=(g_{1},\ldots,g_{N})\), denoted as \(\mathcal{Z}(Q)\). For each \(i\in[N]\), we define \(P_{i}\) as follows: first, generate \(f=f(\mathcal{W}(\mathcal{Z}),R_{i})\) using fresh randomness \(R_{i}\). Set \(P_{i}\stackrel{{\text{\tiny def}}}{{=}}s\) if \(f\) is the \(s\)-th smallest element in \([N_{0}]\setminus\{g_{1},\ldots,g_{i-1}\}\), or a special symbol such as \(\varnothing\) if \(f\) is among \(\{g_{1},\ldots,g_{i-1}\}\).
๋ ˆ๋งˆ F.1์„ ์ ์šฉํ•˜๊ธฐ ์œ„ํ•ด \(\mathcal{Q}_{1}=[N_{0}]\), \(\mathcal{Q}_{2}=[N_{0}-1]\), ๊ทธ๋ฆฌ๊ณ  \(\mathcal{Q}_{N}=[N_{0}-N+1]\)๊นŒ์ง€ ๊ณ„์† ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ๊ท ์ผํ•˜๊ฒŒ ๋ฌด์ž‘์œ„๋กœ \(\mathcal{Q}_{1},\ldots,\mathcal{Q}_{N}\)์—์„œ \(Q_{1},\ldots,Q_{N}\)์„ ์ƒ์„ฑํ•œ ํ›„, ๋‹ค์Œ๊ณผ ๊ฐ™์ด \((g_{1},\ldots,g_{N})\in[N]^{N}\)์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค: \(g_{1}=Q_{1}\); \(g_{2}\)์˜ ๊ฒฝ์šฐ, \(Q_{2}<Q_{1}\)์ด๋ฉด \(Q_{2}\)๋กœ ์„ค์ •ํ•˜๊ณ , ๊ทธ๋ ‡์ง€ ์•Š์œผ๋ฉด \(g_{2}=Q_{2}+1\)์œผ๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค; ์ผ๋ฐ˜์ ์œผ๋กœ \(g_{i}\)๋Š” \([N_{0}]\setminus\{g_{1},\ldots,g_{i-1}\}\)์—์„œ \(Q_{i}\)-๋ฒˆ์งธ๋กœ ์ž‘์€ ์›์†Œ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ \(\mathcal{Z}=(g_{1},\ldots,g_{N})\)๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋˜ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•์„ ์ œ๊ณตํ•˜๋ฉฐ, ์ด๋ฅผ \(\mathcal{Z}(Q)\)๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๊ฐ \(i\in[N]\)์— ๋Œ€ํ•ด \(P_{i}\)๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ •์˜ํ•ฉ๋‹ˆ๋‹ค: ๋จผ์ €, ์‹ ์„ ํ•œ ๋‚œ์ˆ˜ \(R_{i}\)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ \(f=f(\mathcal{W}(\mathcal{Z}),R_{i})\)๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. \(f\)๊ฐ€ \([N_{0}]\setminus\{g_{1},\ldots,g_{i-1}\}\)์—์„œ \(s\) ๋ฒˆ์งธ๋กœ ์ž‘์€ ์›์†Œ์ธ ๊ฒฝ์šฐ \(P_{i}\stackrel{{\text{\tiny def}}}{{=}}s\)๋ฅผ ์„ค์ •ํ•˜๊ณ , \(f\)๊ฐ€ \(\{g_{1},\ldots,g_{i-1}\}\) ์ค‘ ํ•˜๋‚˜์ธ ๊ฒฝ์šฐ \(\varnothing\)์™€ ๊ฐ™์€ ํŠน์ˆ˜ ๊ธฐํ˜ธ๋ฅผ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.
๋ ˆ๋งˆ F.1์„ ์ ์šฉํ•˜๊ธฐ ์œ„ํ•ด \(\mathcal{Q}_{1}=[N_{0}]\), \(\mathcal{Q}_{2}=[N_{0}-1]\), ๊ทธ๋ฆฌ๊ณ  \(\mathcal{Q}_{N}=[N_{0}-N+1]\)๊นŒ์ง€ ๊ณ„์† ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ๊ท ์ผํ•˜๊ฒŒ ๋ฌด์ž‘์œ„๋กœ \(\mathcal{Q}_{1},\ldots,\mathcal{Q}_{N}\)์—์„œ \(Q_{1},\ldots,Q_{N}\)์„ ์ƒ์„ฑํ•œ ํ›„, ๋‹ค์Œ๊ณผ ๊ฐ™์ด \((g_{1},\ldots,g_{N})\in[N]^{N}\)์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค: g_{1}=Q_{1}\), \(g_{2}\)์˜ ๊ฒฝ์šฐ, \(Q_{2}<Q_{1}\), ๊ทธ๋ ‡์ง€ ์•Š์œผ๋ฉด \(g_{2}=Q_{2}+1\)์œผ๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค; ์ผ๋ฐ˜์ ์œผ๋กœ \(g_{i}\)๋Š” \([N_{0}]\setminus\{g_{1},\ldots,g_{i-1}\}\)์—์„œ \(Q_{i}\)-๋ฒˆ์งธ๋กœ ์ž‘์€ ์›์†Œ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ \(\mathcal{Z}=(g_{1},\ldots,g_{N})\)๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋˜ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•์„ ์ œ๊ณตํ•˜๋ฉฐ, ์ด๋ฅผ \(\mathcal{Z}(Q)\)๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๊ฐ \(i\in[N]\)์— ๋Œ€ํ•ด \(P_{i}\)๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ •์˜ํ•ฉ๋‹ˆ๋‹ค: ๋จผ์ €, ์‹ ์„ ํ•œ ๋‚œ์ˆ˜ \(R_{i}\)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ \(f=f(\mathcal{W}(\mathcal{Z}),R_{i})\)๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. f\)๊ฐ€ \([N_{0}]\setminus\{g_{1},\ldots,g_{i-1}\}\)์—์„œ \(s\) ๋ฒˆ์งธ๋กœ ์ž‘์€ ์›์†Œ์ธ ๊ฒฝ์šฐ \(P_{i}\stackrel{{\text{\tiny def}}}{{=}}s\)๋ฅผ ์„ค์ •ํ•˜๊ณ , \(f\)๊ฐ€ \(\{g_{1},\ldots,g_{i-1}\}\) ์ค‘ ํ•˜๋‚˜์ธ ๊ฒฝ์šฐ \(\varnothing\)์™€ ๊ฐ™์€ ํŠน์ˆ˜ ๊ธฐํ˜ธ๋ฅผ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.
Our studies rely on the two established tokenization algorithms, BPE and Unigram, and their implementation in the _Huggingface tokenizer_ library (Moi and Patry, 2023) and the _SentencePiece_ library (Kudo and Richardson, 2018). We considered both libraries in order to investigate the effect of differences in the pre-and post-processing steps and potential differences in the implementations. Due to missing pre-processing options for Huggingface's Unigram implementation, which causes a large discrepancy in the resulting vocabulary compared to SentencePiece's implementation of Unigram, we omitted the training of Unigram tokenizers based on Huggingface. Overall, we trained 24 different tokenizers, where one-half of the tokenizers were monolingual English tokenizers, and the other half of the tokenizers were multilingual tokenizers. Besides the tokenizer algorithm, language composition, and employed tokenizer library, we also varied the vocabulary size. Concrete tokenizer configurations are described in the Appendix B.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๋‘ ๊ฐ€์ง€ ํ™•๋ฆฝ๋œ ํ† ํฐํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ BPE์™€ Unigram๊ณผ ์ด๋ฅผ ๊ตฌํ˜„ํ•œ _Huggingface tokenizer_ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ(Moi and Patry, 2023)์™€ _SentencePiece_ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ(Kudo and Richardson, 2018)์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ๋ฐ ์‚ฌํ›„ ์ฒ˜๋ฆฌ ๋‹จ๊ณ„์˜ ์ฐจ์ด์™€ ์ž ์žฌ์ ์ธ ๊ตฌํ˜„ ์ฐจ์ด๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ๋ชจ๋‘ ๊ณ ๋ คํ–ˆ์Šต๋‹ˆ๋‹ค. Huggingface์˜ Unigram ๊ตฌํ˜„์— ๋Œ€ํ•œ ์‚ฌ์ „ ์ฒ˜๋ฆฌ ์˜ต์…˜์ด ๋ˆ„๋ฝ๋˜์–ด ๊ฒฐ๊ณผ ์–ดํœ˜์—์„œ SentencePiece์˜ Unigram ๊ตฌํ˜„๊ณผ ํฐ ์ฐจ์ด๊ฐ€ ๋ฐœ์ƒํ•˜๋ฏ€๋กœ, Huggingface ๊ธฐ๋ฐ˜์˜ Unigram tokenizer ํ›ˆ๋ จ์€ ์ƒ๋žตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „์ฒด์ ์œผ๋กœ 24๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ tokenizer๋ฅผ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, tokenizer์˜ ์ ˆ๋ฐ˜์€ ๋‹จ์ผ ์–ธ์–ด ์˜์–ด tokenizer, ๋‚˜๋จธ์ง€ ์ ˆ๋ฐ˜์€ ๋‹ค๊ตญ์–ด tokenizer์˜€์Šต๋‹ˆ๋‹ค. tokenizer ์•Œ๊ณ ๋ฆฌ์ฆ˜, ์–ธ์–ด ๊ตฌ์„ฑ, ์‚ฌ์šฉ๋œ tokenizer ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์™ธ์—๋„ ์–ดํœ˜ ํฌ๊ธฐ๋„ ๋‹ค์–‘ํ•˜๊ฒŒ ์กฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์ธ tokenizer ๊ตฌ์„ฑ์€ ๋ถ€๋ก B์— ์„ค๋ช…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๋‘ ๊ฐ€์ง€ ํ™•๋ฆฝ๋œ ํ† ํฐํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ BPE์™€ ์œ ๋‹ˆ๊ทธ๋žจ๊ณผ ์ด๋ฅผ ๊ตฌํ˜„ํ•œ _ํ—ˆ๊น…ํŽ˜์ด์Šค ํ† ํฐํ™”๊ธฐ_ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ(Moi and Patry, 2023)์™€ _SentencePiece_ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ(Kudo and Richardson, 2018)์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ๋ฐ ์‚ฌํ›„ ์ฒ˜๋ฆฌ ๋‹จ๊ณ„์˜ ์ฐจ์ด์™€ ์ž ์žฌ์ ์ธ ๊ตฌํ˜„ ์ฐจ์ด๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ๋ชจ๋‘ ๊ณ ๋ คํ–ˆ์Šต๋‹ˆ๋‹ค. ํ—ˆ๊น…ํŽ˜์ด์Šค์˜ ์œ ๋‹ˆ๊ทธ๋žจ ๊ตฌํ˜„์— ๋Œ€ํ•œ ์‚ฌ์ „ ์ฒ˜๋ฆฌ ์˜ต์…˜์ด ๋ˆ„๋ฝ๋˜์–ด ๊ฒฐ๊ณผ ์–ดํœ˜์—์„œ SentencePiece์˜ ์œ ๋‹ˆ๊ทธ๋žจ ๊ตฌํ˜„๊ณผ ํฐ ์ฐจ์ด๊ฐ€ ๋ฐœ์ƒํ•˜๋ฏ€๋กœ, ํ—ˆ๊น…ํŽ˜์ด์Šค ๊ธฐ๋ฐ˜์˜ ์œ ๋‹ˆ๊ทธ๋žจ ํ† ํฐํ™”๊ธฐ ํ›ˆ๋ จ์€ ์ƒ๋žตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „์ฒด์ ์œผ๋กœ 24๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ํ† ํฐํ™”๊ธฐ๋ฅผ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, ํ† ํฐํ™”๊ธฐ์˜ ์ ˆ๋ฐ˜์€ ๋‹จ์ผ ์–ธ์–ด ์˜์–ด ํ† ํฐํ™”๊ธฐ, ๋‚˜๋จธ์ง€ ์ ˆ๋ฐ˜์€ ๋‹ค๊ตญ์–ด ํ† ํฐํ™”๊ธฐ์˜€์Šต๋‹ˆ๋‹ค. ํ† ํฐํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜, ์–ธ์–ด ๊ตฌ์„ฑ, ์‚ฌ์šฉ๋œ ํ† ํฐํ™” ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์™ธ์—๋„ ์–ดํœ˜ ํฌ๊ธฐ๋„ ๋‹ค์–‘ํ•˜๊ฒŒ ์กฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์ธ ํ† ํฐํ™” ๊ตฌ์„ฑ์€ ๋ถ€๋ก B์— ์„ค๋ช…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Prompt transferring with scarce data._ Supervised learning approaches demand in sufficient training data to learn optimal continuous prompts, which may not work well in data-scarce domains and tasks. To address this problem, SPoT [476] proposes a prompt-based transfer learning approach, which first learns a single continuous prompt for several representative source tasks and then uses this prompt to initialize the prompt for a target task. However, this approach leverages the same prompt for solving all instances of the target task. For a single task, even a well-learned prompt may not be suitable for all the data instances from a large population. To address this issue, an improved method [477] designs an adaptive attention mechanism during the prompt transfer process to derive the target prompts, considering both task- and instance-level information. The prompt transfer paradigm can leverage the knowledge of data-sufficient source tasks encoded in source prompts for solving data-scarce target tasks.
\(\bullet\)_๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•œ ํ”„๋กฌํ”„ํŠธ ์ „์†ก._ Supervised learning ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ตœ์ ์˜ ์—ฐ์† ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด ์ถฉ๋ถ„ํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ํ•„์š”๋กœ ํ•˜๋ฉฐ, ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•œ ๋„๋ฉ”์ธ ๋ฐ ์ž‘์—…์—์„œ๋Š” ์ž˜ ์ž‘๋™ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด SPoT [476]์€ ๋จผ์ € ์—ฌ๋Ÿฌ ๋Œ€ํ‘œ์ ์ธ ์†Œ์Šค ์ž‘์—…์— ๋Œ€ํ•ด ๋‹จ์ผ ์—ฐ์† ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ•™์Šตํ•œ ๋‹ค์Œ ์ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋Œ€์ƒ ์ž‘์—…์˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ดˆ๊ธฐํ™”ํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ ๊ธฐ๋ฐ˜ transfer learning ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋Œ€์ƒ ์ž‘์—…์˜ ๋ชจ๋“  ์ธ์Šคํ„ด์Šค๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋™์ผํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋‹จ์ผ ์ž‘์—…์˜ ๊ฒฝ์šฐ ์ž˜ ํ•™์Šต๋œ ํ”„๋กฌํ”„ํŠธ๋„ ๋Œ€๊ทœ๋ชจ ๋ชจ์ง‘๋‹จ์˜ ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์ธ์Šคํ„ด์Šค์— ์ ํ•ฉํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๊ฐœ์„ ๋œ ๋ฐฉ๋ฒ• [477]์€ ํ”„๋กฌํ”„ํŠธ ์ „์†ก ๊ณผ์ •์—์„œ ์ž‘์—… ๋ฐ ์ธ์Šคํ„ด์Šค ์ˆ˜์ค€ ์ •๋ณด๋ฅผ ๋ชจ๋‘ ๊ณ ๋ คํ•˜์—ฌ ๋Œ€์ƒ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋„์ถœํ•˜๊ธฐ ์œ„ํ•ด adaptive attention ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ ์ „์†ก ํŒจ๋Ÿฌ๋‹ค์ž„์€ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•œ ๋Œ€์ƒ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์†Œ์Šค ํ”„๋กฌํ”„ํŠธ์— ์ธ์ฝ”๋”ฉ๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ์ถฉ๋ถ„ํ•œ ์†Œ์Šค ์ž‘์—…์˜ ์ง€์‹์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•œ ํ”„๋กฌํ”„ํŠธ ์ „์†ก._ ์ง€๋„ ํ•™์Šต ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ตœ์ ์˜ ์—ฐ์† ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด ์ถฉ๋ถ„ํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ํ•„์š”๋กœ ํ•˜๋ฉฐ, ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•œ ๋„๋ฉ”์ธ ๋ฐ ์ž‘์—…์—์„œ๋Š” ์ž˜ ์ž‘๋™ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด SPoT [476]์€ ๋จผ์ € ์—ฌ๋Ÿฌ ๋Œ€ํ‘œ์ ์ธ ์†Œ์Šค ์ž‘์—…์— ๋Œ€ํ•ด ๋‹จ์ผ ์—ฐ์† ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ•™์Šตํ•œ ๋‹ค์Œ ์ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋Œ€์ƒ ์ž‘์—…์˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ดˆ๊ธฐํ™”ํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ ๊ธฐ๋ฐ˜ ์ „์ด ํ•™์Šต ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋Œ€์ƒ ์ž‘์—…์˜ ๋ชจ๋“  ์ธ์Šคํ„ด์Šค๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋™์ผํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋‹จ์ผ ์ž‘์—…์˜ ๊ฒฝ์šฐ ์ž˜ ํ•™์Šต๋œ ํ”„๋กฌํ”„ํŠธ๋„ ๋Œ€๊ทœ๋ชจ ๋ชจ์ง‘๋‹จ์˜ ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์ธ์Šคํ„ด์Šค์— ์ ํ•ฉํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๊ฐœ์„ ๋œ ๋ฐฉ๋ฒ• [477]์€ ํ”„๋กฌํ”„ํŠธ ์ „์†ก ๊ณผ์ •์—์„œ ์ž‘์—… ๋ฐ ์ธ์Šคํ„ด์Šค ์ˆ˜์ค€ ์ •๋ณด๋ฅผ ๋ชจ๋‘ ๊ณ ๋ คํ•˜์—ฌ ๋Œ€์ƒ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋„์ถœํ•˜๊ธฐ ์œ„ํ•ด ์ ์‘ํ˜• ์ฃผ์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ ์ „์†ก ํŒจ๋Ÿฌ๋‹ค์ž„์€ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•œ ๋Œ€์ƒ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์†Œ์Šค ํ”„๋กฌํ”„ํŠธ์— ์ธ์ฝ”๋”ฉ๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ์ถฉ๋ถ„ํ•œ ์†Œ์Šค ์ž‘์—…์˜ ์ง€์‹์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Table 11 shows the CKL benchmark result of GPT-2 models continually pretrained on CC-RecentNews-Small for 8 epochs with a learning rate of 1e-3. By comparing the results in this table with those in Table 10, which is for models continually pretrained with a learning rate of 1e-4, the results in Table 11 shows worse performance on both IL and NQE. Unlike in Appendix E, increasing the learning rate does not result in better learning of new knowledge. Instead, NQE performance is even worse than GPT2-Initial for GPT2-Vanilla, GPT2-Recadam, and GPT2-MixReview. FUAR is _no gain_ for these cases by the definition of the metric because the denominator has the value of zero. This shows that a large learning rate for continual pretraining may lead to failure: neither retaining old knowledge nor acquiring new knowledge effectively. For parameter-expansion methods, because many parameters including the decoder are frozen during the continual training process, they seem to be less prone to the effect of a large learning rate.
ํ‘œ 11์€ ํ•™์Šต๋ฅ  1e-3์œผ๋กœ 8ํšŒ ๋™์•ˆ CC-RecentNews-Small์— ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ GPT-2 ๋ชจ๋ธ์˜ CKL ๋ฒค์น˜๋งˆํฌ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ์ด ํ‘œ์˜ ๊ฒฐ๊ณผ๋ฅผ ํ•™์Šต๋ฅ  1e-4๋กœ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ‘œ 10์˜ ๊ฒฐ๊ณผ์™€ ๋น„๊ตํ•˜๋ฉด ํ‘œ 11์˜ ๊ฒฐ๊ณผ๋Š” IL๊ณผ NQE ๋ชจ๋‘์—์„œ ๋” ๋‚˜์œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ถ€๋ก E์™€ ๋‹ฌ๋ฆฌ ํ•™์Šต๋ฅ (Learning Rate)์„ ๋†’์ด๋ฉด ์ƒˆ๋กœ์šด ์ง€์‹์„ ๋” ์ž˜ ํ•™์Šตํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ๋Œ€์‹ , NQE ์„ฑ๋Šฅ์€ GPT2-Vanilla, GPT2-Recadam ๋ฐ GPT2-MixReview์˜ ๊ฒฝ์šฐ GPT2-Initial๋ณด๋‹ค ๋” ๋‚˜๋น ์ง‘๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ ๋ถ„๋ชจ์— 0์ด๋ผ๋Š” ๊ฐ’์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฉ”ํŠธ๋ฆญ์˜ ์ •์˜์— ๋”ฐ๋ผ FUAR์€ '์ด๋“ ์—†์Œ'์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ํฐ ํ•™์Šต๋ฅ ์ด ์‹คํŒจ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด์ „ ์ง€์‹์„ ์œ ์ง€ํ•˜๊ฑฐ๋‚˜ ์ƒˆ๋กœ์šด ์ง€์‹์„ ํšจ๊ณผ์ ์œผ๋กœ ์Šต๋“ํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ํŒŒ๋ผ๋ฏธํ„ฐ ํ™•์žฅ ๋ฐฉ๋ฒ•์˜ ๊ฒฝ์šฐ, ์ง€์†์  ํ•™์Šต ๊ณผ์ •์—์„œ Decoder๋ฅผ ํฌํ•จํ•œ ๋งŽ์€ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ๋™๊ฒฐ๋˜๊ธฐ ๋•Œ๋ฌธ์— ํฐ ํ•™์Šต๋ฅ ์˜ ์˜ํ–ฅ์„ ๋œ ๋ฐ›๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค.
ํ‘œ 11์€ ํ•™์Šต๋ฅ  1e-3์œผ๋กœ 8ํšŒ ๋™์•ˆ CC-RecentNews-Small์— ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ GPT-2 ๋ชจ๋ธ์˜ CKL ๋ฒค์น˜๋งˆํฌ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ์ด ํ‘œ์˜ ๊ฒฐ๊ณผ๋ฅผ ํ•™์Šต๋ฅ  1e-4๋กœ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ‘œ 10์˜ ๊ฒฐ๊ณผ์™€ ๋น„๊ตํ•˜๋ฉด ํ‘œ 11์˜ ๊ฒฐ๊ณผ๋Š” IL๊ณผ NQE ๋ชจ๋‘์—์„œ ๋” ๋‚˜์œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ถ€๋ก E์™€ ๋‹ฌ๋ฆฌ ํ•™์Šต๋ฅ ์„ ๋†’์ด๋ฉด ์ƒˆ๋กœ์šด ์ง€์‹์„ ๋” ์ž˜ ํ•™์Šตํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ๋Œ€์‹ , NQE ์„ฑ๋Šฅ์€ GPT2-Vanilla, GPT2-Recadam ๋ฐ GPT2-MixReview์˜ ๊ฒฝ์šฐ GPT2-Initial๋ณด๋‹ค ๋” ๋‚˜๋น ์ง‘๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ ๋ถ„๋ชจ์— 0์ด๋ผ๋Š” ๊ฐ’์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฉ”ํŠธ๋ฆญ์˜ ์ •์˜์— ๋”ฐ๋ผ FUAR์€ '์ด๋“ ์—†์Œ'์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ํฐ ํ•™์Šต๋ฅ ์ด ์‹คํŒจ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด์ „ ์ง€์‹์„ ์œ ์ง€ํ•˜๊ฑฐ๋‚˜ ์ƒˆ๋กœ์šด ์ง€์‹์„ ํšจ๊ณผ์ ์œผ๋กœ ์Šต๋“ํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ํŒŒ๋ผ๋ฏธํ„ฐ ํ™•์žฅ ๋ฐฉ๋ฒ•์˜ ๊ฒฝ์šฐ, ์ง€์†์  ํ•™์Šต ๊ณผ์ •์—์„œ ๋””์ฝ”๋”๋ฅผ ํฌํ•จํ•œ ๋งŽ์€ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ๋™๊ฒฐ๋˜๊ธฐ ๋•Œ๋ฌธ์— ํฐ ํ•™์Šต๋ฅ ์˜ ์˜ํ–ฅ์„ ๋œ ๋ฐ›๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค.
The majority of explorations of AI as tools for conducting medical consultations have focused on "symptom checker" applications rather than a full natural dialogue, or on topics such as transcription of medical audio or the generation of plausible dialogue given clinical notes or summaries [60, 61, 62, 63]. Language models have been trained using clinical dialogue datasets but not comprehensively evaluated [64]. Studies have been grounded in messages between doctors and patients in commercial chat platforms (which may have altered doctor-patient engagement compared to 1:1 medical consultations) [65, 28, 66]. Many focused largely on predicting next turns in the recorded exchanges rather than clinically meaningful metrics. And to date, there have been no reported studies that have examined the quality of AI models for diagnostic dialogue using the same criteria that are used to examine and train human physicians in dialogue and communication skills; nor evaluating AI systems in common frameworks such as the OSCE.
์˜๋ฃŒ ์ƒ๋‹ด์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•œ ๋„๊ตฌ๋กœ์„œ์˜ AI๋ฅผ ํƒ๊ตฌํ•˜๋Š” ๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ๋Š” ์™„์ „ํ•œ ์ž์—ฐ์–ด ๋Œ€ํ™”๋ณด๋‹ค๋Š” '์ฆ์ƒ ํ™•์ธ๊ธฐ' ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์ดˆ์ ์„ ๋งž์ถ”๊ฑฐ๋‚˜ ์˜๋ฃŒ์šฉ ์˜ค๋””์˜ค ์ „์‚ฌ ๋˜๋Š” ์ž„์ƒ ๋…ธํŠธ ๋˜๋Š” ์š”์•ฝ์„ ๊ฐ์•ˆํ•˜์—ฌ ๊ทธ๋Ÿด๋“ฏํ•œ ๋Œ€ํ™”๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋“ฑ์˜ ์ฃผ์ œ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค [60, 61, 62, 63]. ์ž„์ƒ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Language Models๋ฅผ ํ•™์Šต์‹œ์ผฐ์ง€๋งŒ ์ข…ํ•ฉ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜์ง€๋Š” ์•Š์•˜์Šต๋‹ˆ๋‹ค [64]. ์—ฐ๊ตฌ๋Š” ์ƒ์—…์šฉ ์ฑ„ํŒ… ํ”Œ๋žซํผ์—์„œ ์˜์‚ฌ์™€ ํ™˜์ž ๊ฐ„์˜ ๋ฉ”์‹œ์ง€์— ๊ธฐ๋ฐ˜์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(1:1 ์˜๋ฃŒ ์ƒ๋‹ด๊ณผ ๋น„๊ตํ•˜์—ฌ ์˜์‚ฌ-ํ™˜์ž ์ฐธ์—ฌ๊ฐ€ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์Œ) [65, 28, 66]. ๋Œ€๋ถ€๋ถ„ ์ž„์ƒ์ ์œผ๋กœ ์˜๋ฏธ ์žˆ๋Š” ์ง€ํ‘œ๋ณด๋‹ค๋Š” ๊ธฐ๋ก๋œ ๋Œ€ํ™”์—์„œ ๋‹ค์Œ ํ„ด์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ๊นŒ์ง€ ์ธ๊ฐ„ ์˜์‚ฌ์˜ ๋Œ€ํ™” ๋ฐ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ๊ธฐ์ˆ ์„ ๊ฒ€ํ† ํ•˜๊ณ  ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๊ฒƒ๊ณผ ๋™์ผํ•œ ๊ธฐ์ค€์„ ์‚ฌ์šฉํ•˜์—ฌ ์ง„๋‹จ ๋Œ€ํ™”๋ฅผ ์œ„ํ•œ AI ๋ชจ๋ธ์˜ ํ’ˆ์งˆ์„ ์กฐ์‚ฌํ•œ ์—ฐ๊ตฌ๋Š” ๋ณด๊ณ ๋˜์ง€ ์•Š์•˜์œผ๋ฉฐ, OSCE์™€ ๊ฐ™์€ ์ผ๋ฐ˜์ ์ธ ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ AI ์‹œ์Šคํ…œ์„ ํ‰๊ฐ€ํ•œ ์—ฐ๊ตฌ๋„ ์—†์Šต๋‹ˆ๋‹ค.
์˜๋ฃŒ ์ƒ๋‹ด์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•œ ๋„๊ตฌ๋กœ์„œ์˜ AI๋ฅผ ํƒ๊ตฌํ•˜๋Š” ๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ๋Š” ์™„์ „ํ•œ ์ž์—ฐ์–ด ๋Œ€ํ™”๋ณด๋‹ค๋Š” '์ฆ์ƒ ํ™•์ธ๊ธฐ' ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์ดˆ์ ์„ ๋งž์ถ”๊ฑฐ๋‚˜ ์˜๋ฃŒ์šฉ ์˜ค๋””์˜ค ์ „์‚ฌ ๋˜๋Š” ์ž„์ƒ ๋…ธํŠธ ๋˜๋Š” ์š”์•ฝ์„ ๊ฐ์•ˆํ•˜์—ฌ ๊ทธ๋Ÿด๋“ฏํ•œ ๋Œ€ํ™”๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋“ฑ์˜ ์ฃผ์ œ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค [60, 61, 62, 63]. ์ž„์ƒ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผฐ์ง€๋งŒ ์ข…ํ•ฉ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜์ง€๋Š” ์•Š์•˜์Šต๋‹ˆ๋‹ค [64]. ์—ฐ๊ตฌ๋Š” ์ƒ์—…์šฉ ์ฑ„ํŒ… ํ”Œ๋žซํผ์—์„œ ์˜์‚ฌ์™€ ํ™˜์ž ๊ฐ„์˜ ๋ฉ”์‹œ์ง€์— ๊ธฐ๋ฐ˜์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(1:1 ์˜๋ฃŒ ์ƒ๋‹ด๊ณผ ๋น„๊ตํ•˜์—ฌ ์˜์‚ฌ-ํ™˜์ž ์ฐธ์—ฌ๊ฐ€ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์Œ) [65, 28, 66]. ๋Œ€๋ถ€๋ถ„ ์ž„์ƒ์ ์œผ๋กœ ์˜๋ฏธ ์žˆ๋Š” ์ง€ํ‘œ๋ณด๋‹ค๋Š” ๊ธฐ๋ก๋œ ๋Œ€ํ™”์—์„œ ๋‹ค์Œ ํ„ด์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ๊นŒ์ง€ ์ธ๊ฐ„ ์˜์‚ฌ์˜ ๋Œ€ํ™” ๋ฐ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ๊ธฐ์ˆ ์„ ๊ฒ€ํ† ํ•˜๊ณ  ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๊ฒƒ๊ณผ ๋™์ผํ•œ ๊ธฐ์ค€์„ ์‚ฌ์šฉํ•˜์—ฌ ์ง„๋‹จ ๋Œ€ํ™”๋ฅผ ์œ„ํ•œ AI ๋ชจ๋ธ์˜ ํ’ˆ์งˆ์„ ์กฐ์‚ฌํ•œ ์—ฐ๊ตฌ๋Š” ๋ณด๊ณ ๋˜์ง€ ์•Š์•˜์œผ๋ฉฐ, OSCE์™€ ๊ฐ™์€ ์ผ๋ฐ˜์ ์ธ ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ AI ์‹œ์Šคํ…œ์„ ํ‰๊ฐ€ํ•œ ์—ฐ๊ตฌ๋„ ์—†์Šต๋‹ˆ๋‹ค.
**Memorization.** In order to handle long-horizon tasks, it has become a key approach to aid plan refinement with _long-term memory_ in addition to utilizing the _short-term memory_ ofLMs through ICL. For example, Reflexion [450] stores the feedback from self-reflection into the memory, so previous feedback can be retrieved for plan refinement. Generative Agents [533] designs the memory stream mechanism as the core component of agents for action planning and reflection. Further, the skill library mechanism [445, 528] is proposed to store successful plans in the library, which can be reused and synthesized as complex plans for novel tasks. To implement the long-term memory mechanism, tools like vector databases (_e.g._, milvus [536]) can be used to encode plans or feedbacks into high-dimensional vectors for efficient storage and retrieval at a large scale. MemoryBank [537] further proposes the memory updating mechanism to allow memory forgetting and strengthening following the Ebbinghaus Forgetting Curve theory.
**Memorization**. ์žฅ๊ธฐ ์˜ˆ์ธก ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด In-Context Learning์„ ํ†ตํ•ด LMs์˜ _๋‹จ๊ธฐ ๊ธฐ์–ต_์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ _์žฅ๊ธฐ ๊ธฐ์–ต_์œผ๋กœ ๊ณ„ํš ๊ตฌ์ฒดํ™”๋ฅผ ์ง€์›ํ•˜๋Š” ๊ฒƒ์ด ํ•ต์‹ฌ ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Reflexion [450]์€ ์ž๊ธฐ ์„ฑ์ฐฐ์—์„œ ์–ป์€ ํ”ผ๋“œ๋ฐฑ์„ ๋ฉ”๋ชจ๋ฆฌ์— ์ €์žฅํ•˜์—ฌ ์ด์ „ ํ”ผ๋“œ๋ฐฑ์„ ๊ฒ€์ƒ‰ํ•˜์—ฌ ๊ณ„ํš์„ ๊ตฌ์ฒดํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Generative Agents [533]๋Š” ํ–‰๋™ ๊ณ„ํš ๋ฐ ๋ฐ˜์„ฑ์„ ์œ„ํ•œ ์—์ด์ „ํŠธ์˜ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๋ฉ”๋ชจ๋ฆฌ ์ŠคํŠธ๋ฆผ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์„ฑ๊ณต์ ์ธ ๊ณ„ํš์„ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์— ์ €์žฅํ•˜์—ฌ ์ƒˆ๋กœ์šด ์ž‘์—…์— ๋Œ€ํ•ด ๋ณต์žกํ•œ ๊ณ„ํš์œผ๋กœ ์žฌ์‚ฌ์šฉ ๋ฐ ํ•ฉ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์Šคํ‚ฌ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ๋ฉ”์ปค๋‹ˆ์ฆ˜ [445, 528]์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์žฅ๊ธฐ ๋ฉ”๋ชจ๋ฆฌ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค(_์˜ˆ: milvus [536])์™€ ๊ฐ™์€ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ณ„ํš์ด๋‚˜ ํ”ผ๋“œ๋ฐฑ์„ ๊ณ ์ฐจ์› ๋ฒกํ„ฐ๋กœ ์ธ์ฝ”๋”ฉํ•˜์—ฌ ๋Œ€๊ทœ๋ชจ๋กœ ํšจ์œจ์ ์œผ๋กœ ์ €์žฅํ•˜๊ณ  ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. MemoryBank [537]์€ Ebbinghaus Forgetting Curve ์ด๋ก ์— ๋”ฐ๋ผ ๋ฉ”๋ชจ๋ฆฌ ์žŠ์–ด๋ฒ„๋ฆผ ๋ฐ ๊ฐ•ํ™”๋ฅผ ํ—ˆ์šฉํ•˜๋Š” ๋ฉ”๋ชจ๋ฆฌ ์—…๋ฐ์ดํŠธ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ถ”๊ฐ€๋กœ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
**์•”๊ธฐ**. ์žฅ๊ธฐ ์˜ˆ์ธก ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ICL์„ ํ†ตํ•ด LM์˜ _๋‹จ๊ธฐ ๊ธฐ์–ต_์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ _์žฅ๊ธฐ ๊ธฐ์–ต_์œผ๋กœ ๊ณ„ํš ๊ตฌ์ฒดํ™”๋ฅผ ์ง€์›ํ•˜๋Š” ๊ฒƒ์ด ํ•ต์‹ฌ ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Reflexion [450]์€ ์ž๊ธฐ ์„ฑ์ฐฐ์—์„œ ์–ป์€ ํ”ผ๋“œ๋ฐฑ์„ ๋ฉ”๋ชจ๋ฆฌ์— ์ €์žฅํ•˜์—ฌ ์ด์ „ ํ”ผ๋“œ๋ฐฑ์„ ๊ฒ€์ƒ‰ํ•˜์—ฌ ๊ณ„ํš์„ ๊ตฌ์ฒดํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ƒ์„ฑ ์—์ด์ „ํŠธ [533]๋Š” ํ–‰๋™ ๊ณ„ํš ๋ฐ ๋ฐ˜์„ฑ์„ ์œ„ํ•œ ์—์ด์ „ํŠธ์˜ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๋ฉ”๋ชจ๋ฆฌ ์ŠคํŠธ๋ฆผ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์„ฑ๊ณต์ ์ธ ๊ณ„ํš์„ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์— ์ €์žฅํ•˜์—ฌ ์ƒˆ๋กœ์šด ์ž‘์—…์— ๋Œ€ํ•ด ๋ณต์žกํ•œ ๊ณ„ํš์œผ๋กœ ์žฌ์‚ฌ์šฉ ๋ฐ ํ•ฉ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์Šคํ‚ฌ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ๋ฉ”์ปค๋‹ˆ์ฆ˜ [445, 528]์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์žฅ๊ธฐ ๋ฉ”๋ชจ๋ฆฌ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค(_์˜ˆ: milvus [536])์™€ ๊ฐ™์€ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ณ„ํš์ด๋‚˜ ํ”ผ๋“œ๋ฐฑ์„ ๊ณ ์ฐจ์› ๋ฒกํ„ฐ๋กœ ์ธ์ฝ”๋”ฉํ•˜์—ฌ ๋Œ€๊ทœ๋ชจ๋กœ ํšจ์œจ์ ์œผ๋กœ ์ €์žฅํ•˜๊ณ  ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. MemoryBank [537]์€ Ebbinghaus Forgetting Curve ์ด๋ก ์— ๋”ฐ๋ผ ๋ฉ”๋ชจ๋ฆฌ ์žŠ์–ด๋ฒ„๋ฆผ ๋ฐ ๊ฐ•ํ™”๋ฅผ ํ—ˆ์šฉํ•˜๋Š” ๋ฉ”๋ชจ๋ฆฌ ์—…๋ฐ์ดํŠธ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ถ”๊ฐ€๋กœ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
where \(q\) is the query, for example, the task that the agent should address or the context in which the agent is situated. \(M\) is the set of all memories. \(s^{rec}(\cdot)\), \(s^{rel}(\cdot)\) and \(s^{imp}(\cdot)\) are the scoring functions for measuring the recency, relevance, and importance of the memory \(m\). These scoring functions can be implemented using various methods, for example, \(s^{rel}(q,m)\) can be realized based on LSH, ANNOY, HNSW, FAISS and so on. It should be noted that \(s^{imp}\) only reflects the characters of the memory itself, thus it is unrelated to the query \(q\). \(\alpha\), \(\beta\) and \(\gamma\) are balancing parameters. By assigning them with different values, one can obtain various memory reading strategies. For example, by setting \(\alpha=\gamma=0\), many studies [16, 30, 38, 42] only consider the relevance score \(s^{rel}\) for memory reading. By assigning \(\alpha=\beta=\gamma=1.0\), [20] equally weights all the above three metrics to extract information from memory.
์—ฌ๊ธฐ์„œ \(q\)๋Š” ์—์ด์ „ํŠธ๊ฐ€ ํ•ด๊ฒฐํ•ด์•ผ ํ•˜๋Š” ์ž‘์—… ๋˜๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ฒ˜ํ•œ ์ƒํ™ฉ๊ณผ ๊ฐ™์€ query์ž…๋‹ˆ๋‹ค. \(M\)์€ ๋ชจ๋“  ๋ฉ”๋ชจ๋ฆฌ์˜ ์ง‘ํ•ฉ์ž…๋‹ˆ๋‹ค. \(s^{rec}(\cdot)\), \(s^{rel}(\cdot)\) ๋ฐ \(s^{imp}(\cdot)\)๋Š” ๋ฉ”๋ชจ๋ฆฌ \(m\)์˜ ์ตœ๊ทผ์„ฑ, ๊ด€๋ จ์„ฑ ๋ฐ ์ค‘์š”๋„๋ฅผ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•œ scoring functions์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ scoring functions๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์˜ˆ๋ฅผ ๋“ค์–ด \(s^{rel}(q,m)\)์€ LSH, ANNOY, HNSW, FAISS ๋“ฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. \(s^{imp}\)๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ž์ฒด์˜ ํŠน์„ฑ๋งŒ ๋ฐ˜์˜ํ•˜๋ฏ€๋กœ query \(q\)์™€๋Š” ๊ด€๋ จ์ด ์—†๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. \(\alpha\), \(\beta\) ๋ฐ \(\gamma\)๋Š” balancing parameters์ž…๋‹ˆ๋‹ค. ์ด ๊ฐ’์„ ๋‹ค๋ฅด๊ฒŒ ์ง€์ •ํ•˜๋ฉด ๋‹ค์–‘ํ•œ ๋ฉ”๋ชจ๋ฆฌ ์ฝ๊ธฐ ์ „๋žต์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(\alpha=\gamma=0\)์œผ๋กœ ์„ค์ •ํ•˜๋ฉด ๋งŽ์€ ์—ฐ๊ตฌ[16, 30, 38, 42]์—์„œ ๋ฉ”๋ชจ๋ฆฌ ์ฝ๊ธฐ์— ๊ด€๋ จ ์ ์ˆ˜ \(s^{rel}\)๋งŒ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. \(\alpha=\beta=\gamma=1.0\), [20]์€ ์œ„์˜ ์„ธ ๊ฐ€์ง€ metrics์— ๋ชจ๋‘ ๋™์ผํ•œ ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(q\)๋Š” ์—์ด์ „ํŠธ๊ฐ€ ํ•ด๊ฒฐํ•ด์•ผ ํ•˜๋Š” ์ž‘์—… ๋˜๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ฒ˜ํ•œ ์ƒํ™ฉ๊ณผ ๊ฐ™์€ ์ฟผ๋ฆฌ์ž…๋‹ˆ๋‹ค. M\)์€ ๋ชจ๋“  ๋ฉ”๋ชจ๋ฆฌ์˜ ์ง‘ํ•ฉ์ž…๋‹ˆ๋‹ค. s^{rec}(\cdot)\), \(s^{rel}(\cdot)\) ๋ฐ \(s^{imp}(\cdot)\)๋Š” ๋ฉ”๋ชจ๋ฆฌ \(m\)์˜ ์ตœ๊ทผ์„ฑ, ๊ด€๋ จ์„ฑ ๋ฐ ์ค‘์š”๋„๋ฅผ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•œ ์ ์ˆ˜ ํ•จ์ˆ˜์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ์ˆ˜ ํ•จ์ˆ˜๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์˜ˆ๋ฅผ ๋“ค์–ด \(s^{rel}(q,m)\)์€ LSH, ANNOY, HNSW, FAISS ๋“ฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. s^{imp}\)๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ž์ฒด์˜ ํŠน์„ฑ๋งŒ ๋ฐ˜์˜ํ•˜๋ฏ€๋กœ ์ฟผ๋ฆฌ \(q\)์™€๋Š” ๊ด€๋ จ์ด ์—†๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. \(\์•ŒํŒŒ\), \(\๋ฒ ํƒ€\), \(\๊ฐ๋งˆ\)๋Š” ๋ฐธ๋Ÿฐ์‹ฑ ํŒŒ๋ผ๋ฏธํ„ฐ์ž…๋‹ˆ๋‹ค. ์ด ๊ฐ’์„ ๋‹ค๋ฅด๊ฒŒ ์ง€์ •ํ•˜๋ฉด ๋‹ค์–‘ํ•œ ๋ฉ”๋ชจ๋ฆฌ ์ฝ๊ธฐ ์ „๋žต์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(\์•ŒํŒŒ=\๊ฐ๋งˆ=0\)์œผ๋กœ ์„ค์ •ํ•˜๋ฉด ๋งŽ์€ ์—ฐ๊ตฌ[16, 30, 38, 42]์—์„œ ๋ฉ”๋ชจ๋ฆฌ ์ฝ๊ธฐ์— ๊ด€๋ จ ์ ์ˆ˜ \(s^{rel}\)๋งŒ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์•ŒํŒŒ=\๋ฒ ํƒ€=\๊ฐ๋งˆ=1.0\), [20]์€ ์œ„์˜ ์„ธ ๊ฐ€์ง€ ๋ฉ”ํŠธ๋ฆญ์— ๋ชจ๋‘ ๋™์ผํ•œ ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค.
This report introduces EEVE-Korean-v1.0, a Korean adaptation of large language models that exhibit remarkable capabilities across English and Korean text understanding. Building on recent highly capable but English-centric LLMs, such as SOLAR-10.7B and Phi-2, where non-English texts are inefficiently processed with English-centric tokenizers, we present an efficient and effective vocabulary expansion (EEVE) method, which encompasses parameter freezing and subword initialization. In contrast to previous efforts that believe new embeddings require trillions of training tokens, we show that our method can significantly boost non-English proficiency within just 2 billion tokens. Surpassing most instruction-tuned LLMs on the Open Ko-LLM Leaderboard, as of January 2024, our model EEVE-Korean-10.88-v1.0 ranks as the leading Korean pre-trained model in the open-source community, according to Hugging Face's leaderboard. We open-source our models on Huggingface to empower the open research community in various languages.
์ด ๋ณด๊ณ ์„œ๋Š” ์˜์–ด์™€ ํ•œ๊ตญ์–ด ํ…์ŠคํŠธ ์ดํ•ด์— ๊ฑธ์ณ ๋†€๋ผ์šด ์—ญ๋Ÿ‰์„ ๋ฐœํœ˜ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ํ•œ๊ตญ์–ด ์ ์‘ํŒ์ธ EEVE-Korean-v1.0์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ ์˜์–ด ์ค‘์‹ฌ์˜ ๊ณ ์„ฑ๋Šฅ LLM(์˜ˆ: SOLAR-10.7B ๋ฐ Phi-2)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋˜, ์˜์–ด ์ค‘์‹ฌ์˜ ํ† ํฐํ™”๊ธฐ๋กœ ๋น„์˜์–ด ํ…์ŠคํŠธ๋ฅผ ๋น„ํšจ์œจ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๋Š” ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํŒŒ๋ผ๋ฏธํ„ฐ ๋™๊ฒฐ ๋ฐ ํ•˜์œ„ ๋‹จ์–ด ์ดˆ๊ธฐํ™”๋ฅผ ํฌํ•จํ•˜๋Š” ํšจ์œจ์ ์ด๊ณ  ํšจ๊ณผ์ ์ธ ์–ดํœ˜ ํ™•์žฅ(EEVE) ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด ์ž„๋ฒ ๋”ฉ์— ์ˆ˜์กฐ ๊ฐœ์˜ ํ•™์Šต ํ† ํฐ์ด ํ•„์š”ํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋Š” ๊ธฐ์กด์˜ ๋…ธ๋ ฅ๊ณผ ๋‹ฌ๋ฆฌ, ์ €ํฌ๋Š” ๋‹จ 20์–ต ๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ๋น„์˜์–ด ์‹ค๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. 2024๋…„ 1์›” ํ˜„์žฌ Open Ko-LLM ๋ฆฌ๋”๋ณด๋“œ์—์„œ ๋Œ€๋ถ€๋ถ„์˜ Instruction-tuned LLM์„ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ €ํฌ ๋ชจ๋ธ EEVE-Korean-10.88-v1.0์€ Hugging Face ๋ฆฌ๋”๋ณด๋“œ์— ๋”ฐ๋ฅด๋ฉด ์˜คํ”ˆ ์†Œ์Šค ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ์„ ๋„์ ์ธ ํ•œ๊ตญ์–ด ์‚ฌ์ „ ํ•™์Šต ๋ชจ๋ธ๋กœ ์„ ์ •๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์–ธ์–ด์˜ ์˜คํ”ˆ ๋ฆฌ์„œ์น˜ ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ์—ญ๋Ÿ‰์„ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•ด Hugging Face์— ๋ชจ๋ธ์„ ์˜คํ”ˆ ์†Œ์Šคํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค.
์ด ๋ณด๊ณ ์„œ๋Š” ์˜์–ด์™€ ํ•œ๊ตญ์–ด ํ…์ŠคํŠธ ์ดํ•ด์— ๊ฑธ์ณ ๋†€๋ผ์šด ์—ญ๋Ÿ‰์„ ๋ฐœํœ˜ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ํ•œ๊ตญ์–ด ์ ์‘ํŒ์ธ EEVE-Korean-v1.0์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ ์˜์–ด ์ค‘์‹ฌ์˜ ๊ณ ์„ฑ๋Šฅ LLM(์˜ˆ: SOLAR-10.7B ๋ฐ Phi-2)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋˜, ์˜์–ด ์ค‘์‹ฌ์˜ ํ† ํฐํ™”๊ธฐ๋กœ ๋น„์˜์–ด ํ…์ŠคํŠธ๋ฅผ ๋น„ํšจ์œจ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๋Š” ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํŒŒ๋ผ๋ฏธํ„ฐ ๋™๊ฒฐ ๋ฐ ํ•˜์œ„ ๋‹จ์–ด ์ดˆ๊ธฐํ™”๋ฅผ ํฌํ•จํ•˜๋Š” ํšจ์œจ์ ์ด๊ณ  ํšจ๊ณผ์ ์ธ ์–ดํœ˜ ํ™•์žฅ(EEVE) ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด ์ž„๋ฒ ๋”ฉ์— ์ˆ˜์กฐ ๊ฐœ์˜ ํ•™์Šต ํ† ํฐ์ด ํ•„์š”ํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋Š” ๊ธฐ์กด์˜ ๋…ธ๋ ฅ๊ณผ ๋‹ฌ๋ฆฌ, ์ €ํฌ๋Š” ๋‹จ 20์–ต ๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ๋น„์˜์–ด ์‹ค๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. 2024๋…„ 1์›” ํ˜„์žฌ Open Ko-LLM ๋ฆฌ๋”๋ณด๋“œ์—์„œ ๋Œ€๋ถ€๋ถ„์˜ ์ง€์‹œ ํŠœ๋‹ LLM์„ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ €ํฌ ๋ชจ๋ธ EEVE-Korean-10.88-v1.0์€ ํ—ˆ๊น…ํŽ˜์ด์Šค ๋ฆฌ๋”๋ณด๋“œ์— ๋”ฐ๋ฅด๋ฉด ์˜คํ”ˆ ์†Œ์Šค ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ์„ ๋„์ ์ธ ํ•œ๊ตญ์–ด ์‚ฌ์ „ ํ•™์Šต ๋ชจ๋ธ๋กœ ์„ ์ •๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์–ธ์–ด์˜ ์˜คํ”ˆ ๋ฆฌ์„œ์น˜ ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ์—ญ๋Ÿ‰์„ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ—ˆ๊น…ํŽ˜์ด์Šค์— ๋ชจ๋ธ์„ ์˜คํ”ˆ ์†Œ์Šคํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค.
In the context of continual pre-training of LLMs, we have seen that learning rate re-warming and re-decaying is important for adaptation and found that the forgetting is easily mitigated with replay in this setting--at seemingly little cost to adaptation. Given their powerful ability to enhance adaptation and mitigate forgetting simultaneously, we proposed the simple and scalable combination of LR re-warming, LR re-decaying, and replay for continually pre-training LLMs at scale. We showed that these strategies enable continual pre-training to achieve average performance on par with expensively re-training from scratch on all data, across two distribution shifts (weak & strong) and two model scales (405M & 10B). Upon further analysis, we identified a pathology of LR re-warming and, inspired by previous work, proposed infinite learning rate schedules for continually pre-training LLMs. In initial experiments, our schedules achieve performance on par with cosine decay while circumventing the need for LR re-warming.
LLM์˜ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์˜ ๊ฒฝ์šฐ, ์ ์‘์„ ์œ„ํ•ด Learning Rate ์žฌ๊ฐ€์—ด๊ณผ ์žฌ๊ฐ์†Œ๊ฐ€ ์ค‘์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์œผ๋ฉฐ, ์ด ์„ค์ •์—์„œ ์žฌ์ƒ์„ ํ†ตํ•ด ์žŠ์–ด๋ฒ„๋ฆผ์„ ์‰ฝ๊ฒŒ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ ์‘์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ์žŠ์–ด๋ฒ„๋ฆผ์„ ๋™์‹œ์— ์™„ํ™”ํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ๋Šฅ๋ ฅ์„ ๋ฐ”ํƒ•์œผ๋กœ, ์ €ํฌ๋Š” ๋Œ€๊ทœ๋ชจ LLM์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ„๋‹จํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ Learning Rate ์žฌ๊ฐ€์—ด, Learning Rate ์žฌ๊ฐ์†Œ, ์žฌ์ƒ์˜ ์กฐํ•ฉ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ „๋žต์„ ํ†ตํ•ด ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์ด ๋‘ ๊ฐ€์ง€ ๋ถ„ํฌ ๋ณ€ํ™”(์•ฝํ•œ ๋ณ€ํ™”์™€ ๊ฐ•ํ•œ ๋ณ€ํ™”)์™€ ๋‘ ๊ฐ€์ง€ ๋ชจ๋ธ ๊ทœ๋ชจ(405M ๋ฐ 10B)์— ๊ฑธ์ณ ๋ชจ๋“  ๋ฐ์ดํ„ฐ์—์„œ ์ฒ˜์Œ๋ถ€ํ„ฐ ๋‹ค์‹œ ํ•™์Šตํ•˜๋Š” ๊ฒƒ๊ณผ ๋น„์Šทํ•œ ํ‰๊ท  ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ถ”๊ฐ€ ๋ถ„์„ ๊ฒฐ๊ณผ, Learning Rate ์žฌ๊ฐ€์—ด์˜ ๋ณ‘๋ฆฌ๋ฅผ ํ™•์ธํ•˜๊ณ  ์ด์ „ ์—ฐ๊ตฌ์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ LLM์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•œ ๋ฌดํ•œ Learning Rate ์ผ์ •์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ์‹คํ—˜์—์„œ ์šฐ๋ฆฌ์˜ ์ผ์ •์€ ์ฝ”์‚ฌ์ธ ๊ฐ์†Œ์™€ ๋™๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋ฉด์„œ Learning Rate ์žฌ๊ฐ€์—ด์˜ ํ•„์š”์„ฑ์„ ํ”ผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
LLM์˜ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์˜ ๊ฒฝ์šฐ, ์ ์‘์„ ์œ„ํ•ด ํ•™์Šต ์†๋„ ์žฌ๊ฐ€์—ด๊ณผ ์žฌ๊ฐ์†Œ๊ฐ€ ์ค‘์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์œผ๋ฉฐ, ์ด ์„ค์ •์—์„œ ์žฌ์ƒ์„ ํ†ตํ•ด ์žŠ์–ด๋ฒ„๋ฆผ์„ ์‰ฝ๊ฒŒ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ ์‘์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ์žŠ์–ด๋ฒ„๋ฆผ์„ ๋™์‹œ์— ์™„ํ™”ํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ๋Šฅ๋ ฅ์„ ๋ฐ”ํƒ•์œผ๋กœ, ์ €ํฌ๋Š” ๋Œ€๊ทœ๋ชจ LLM์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ„๋‹จํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ LR ์žฌ๊ฐ€์—ด, LR ์žฌ๊ฐ์†Œ, ์žฌ์ƒ์˜ ์กฐํ•ฉ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ „๋žต์„ ํ†ตํ•ด ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์ด ๋‘ ๊ฐ€์ง€ ๋ถ„ํฌ ๋ณ€ํ™”(์•ฝํ•œ ๋ณ€ํ™”์™€ ๊ฐ•ํ•œ ๋ณ€ํ™”)์™€ ๋‘ ๊ฐ€์ง€ ๋ชจ๋ธ ๊ทœ๋ชจ(405M ๋ฐ 10B)์— ๊ฑธ์ณ ๋ชจ๋“  ๋ฐ์ดํ„ฐ์—์„œ ์ฒ˜์Œ๋ถ€ํ„ฐ ๋‹ค์‹œ ํ•™์Šตํ•˜๋Š” ๊ฒƒ๊ณผ ๋น„์Šทํ•œ ํ‰๊ท  ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ถ”๊ฐ€ ๋ถ„์„ ๊ฒฐ๊ณผ, LR ์žฌ๊ฐ€์—ด์˜ ๋ณ‘๋ฆฌ๋ฅผ ํ™•์ธํ•˜๊ณ  ์ด์ „ ์—ฐ๊ตฌ์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ LLM์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•œ ๋ฌดํ•œ ํ•™์Šต ์†๋„ ์ผ์ •์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ์‹คํ—˜์—์„œ ์šฐ๋ฆฌ์˜ ์ผ์ •์€ ์ฝ”์‚ฌ์ธ ๊ฐ์†Œ์™€ ๋™๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋ฉด์„œ LR ์žฌ๊ฐ€์—ด์˜ ํ•„์š”์„ฑ์„ ํ”ผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 2 reports the validation losses for \(\mathcal{D}_{0}\) and \(\mathcal{D}_{1}\) for all models throughout the first 50B tokens of continued pre-training on \(\mathcal{D}_{1}\). The top row reports results for the weak distribution shift, while the bottom row reports results for the strong distribution shift. Across both distribution shifts, we observe initially that models using shorter linear warmup initially forget and adapt faster than their longer warmup counterparts. This happens because they increase the LR faster which leads to faster forgetting and adaptation. In all scenarios, however, these initial differences diminish throughout training, leaving all models with relatively similar forgetting and adaptation after 50B tokens. Thus, in the settings explored, **the duration of linear warm-up phase does not appear to be impactful on forgetting or adaptation when continuing to pre-train.** With this in mind, we set a linear warmup duration of 1% of training iterations for all subsequent experiments.
๊ทธ๋ฆผ 2๋Š” \(\mathcal{D}_{1}\)์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์˜ ์ฒซ 50B ํ† ํฐ ์ „์ฒด์— ๊ฑธ์ณ ๋ชจ๋“  ๋ชจ๋ธ์— ๋Œ€ํ•œ \(\mathcal{D}_{0}\) ๋ฐ \(\mathcal{D}_{1}\)์— ๋Œ€ํ•œ validation losses๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์œ„์ชฝ ํ–‰์€ ์•ฝํ•œ ๋ถ„ํฌ ๋ณ€ํ™”์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•˜๊ณ , ์•„๋ž˜์ชฝ ํ–‰์€ ๊ฐ•ํ•œ ๋ถ„ํฌ ๋ณ€ํ™”์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๋ถ„ํฌ ๋ณ€ํ™” ๋ชจ๋‘์—์„œ ์ฒ˜์Œ์—๋Š” ์ดˆ๊ธฐ warmup์ด ์งง์€ ๋ชจ๋ธ์ด ๊ธด warmup ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋นจ๋ฆฌ ์žŠ์–ด๋ฒ„๋ฆฌ๊ณ  ์ ์‘ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” Learning Rate๋ฅผ ๋” ๋นจ๋ฆฌ ์ฆ๊ฐ€์‹œ์ผœ ๋” ๋นจ๋ฆฌ ์žŠ์–ด๋ฒ„๋ฆฌ๊ณ  ์ ์‘ํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋“  ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์ด๋Ÿฌํ•œ ์ดˆ๊ธฐ ์ฐจ์ด๋Š” ํ›ˆ๋ จ์ด ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ ์ค„์–ด๋“ค์–ด 50B ํ† ํฐ ์ดํ›„์—๋Š” ๋ชจ๋“  ๋ชจ๋ธ์ด ๋น„๊ต์  ๋น„์Šทํ•œ ์žŠ์–ด๋ฒ„๋ฆผ๊ณผ ์ ์‘์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํƒ์ƒ‰ํ•œ ์„ค์ •์—์„œ linear warm-up ๋‹จ๊ณ„์˜ ์ง€์† ์‹œ๊ฐ„์€ ์‚ฌ์ „ ํ•™์Šต์„ ๊ณ„์†ํ•  ๋•Œ ์žŠ์–ด๋ฒ„๋ฆผ์ด๋‚˜ ์ ์‘์— ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ์—ผ๋‘์— ๋‘๊ณ  ๋ชจ๋“  ํ›„์† ์‹คํ—˜์— ๋Œ€ํ•ด linear warmup ๊ธฐ๊ฐ„์„ ํ›ˆ๋ จ ๋ฐ˜๋ณต์˜ 1%๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2๋Š” \(\mathcal{D}_{1}\)์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์˜ ์ฒซ 50B ํ† ํฐ ์ „์ฒด์— ๊ฑธ์ณ ๋ชจ๋“  ๋ชจ๋ธ์— ๋Œ€ํ•œ \(\mathcal{D}_{0}\) ๋ฐ \(\mathcal{D}_{1}\)์— ๋Œ€ํ•œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์†์‹ค์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์œ„์ชฝ ํ–‰์€ ์•ฝํ•œ ๋ถ„ํฌ ๋ณ€ํ™”์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•˜๊ณ , ์•„๋ž˜์ชฝ ํ–‰์€ ๊ฐ•ํ•œ ๋ถ„ํฌ ๋ณ€ํ™”์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๋ถ„ํฌ ๋ณ€ํ™” ๋ชจ๋‘์—์„œ ์ฒ˜์Œ์—๋Š” ์ดˆ๊ธฐ ์›Œ๋ฐ์—…์ด ์งง์€ ๋ชจ๋ธ์ด ๊ธด ์›Œ๋ฐ์—… ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋นจ๋ฆฌ ์žŠ์–ด๋ฒ„๋ฆฌ๊ณ  ์ ์‘ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” LR์„ ๋” ๋นจ๋ฆฌ ์ฆ๊ฐ€์‹œ์ผœ ๋” ๋นจ๋ฆฌ ์žŠ์–ด๋ฒ„๋ฆฌ๊ณ  ์ ์‘ํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋“  ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์ด๋Ÿฌํ•œ ์ดˆ๊ธฐ ์ฐจ์ด๋Š” ํ›ˆ๋ จ์ด ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ ์ค„์–ด๋“ค์–ด 500์–ต ํ† ํฐ ์ดํ›„์—๋Š” ๋ชจ๋“  ๋ชจ๋ธ์ด ๋น„๊ต์  ๋น„์Šทํ•œ ์žŠ์–ด๋ฒ„๋ฆผ๊ณผ ์ ์‘์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํƒ์ƒ‰ํ•œ ์„ค์ •์—์„œ ์„ ํ˜• ์›Œ๋ฐ์—… ๋‹จ๊ณ„์˜ ์ง€์† ์‹œ๊ฐ„์€ ์‚ฌ์ „ ํ•™์Šต์„ ๊ณ„์†ํ•  ๋•Œ ์žŠ์–ด๋ฒ„๋ฆผ์ด๋‚˜ ์ ์‘์— ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ์—ผ๋‘์— ๋‘๊ณ  ๋ชจ๋“  ํ›„์† ์‹คํ—˜์— ๋Œ€ํ•ด ์„ ํ˜• ์›Œ๋ฐ์—… ๊ธฐ๊ฐ„์„ ํ›ˆ๋ จ ๋ฐ˜๋ณต์˜ 1%๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
Doubanis a social network and database that allows users to create content related to literature and artistic works such as films, books, TV series, music, etc. We sample data from books, movies, and TV series, extracting metadata that includes ratings, detailed information on actors/crew, and long reviews. Then, we design three tasks in total: synopsis generation, review generation, and recommendations. For each task, we manually design various prompt templates and used these templates in combination with metadata to construct instructions. For synopsis generation and review generation, we construct instructions using prompt templates combined with movie or TV series names, with responses generated by Douban users. Then we remove responses with lengths shorter than a threshold and delete personal information and irrelevant content(e.g., "Subscribe our Official Accounts"). Additionally, we manually adjusted some instructions to add more complex implicit intents, aligning better with the details of the response.
๋‘๋ฐ˜์€ ์‚ฌ์šฉ์ž๊ฐ€ ์˜ํ™”, ์ฑ…, TV ์‹œ๋ฆฌ์ฆˆ, ์Œ์•… ๋“ฑ๊ณผ ๊ฐ™์€ ๋ฌธํ•™ ๋ฐ ์˜ˆ์ˆ  ์ž‘ํ’ˆ๊ณผ ๊ด€๋ จ๋œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์†Œ์…œ ๋„คํŠธ์›Œํฌ์ด์ž ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์ž…๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ฑ…, ์˜ํ™”, TV ์‹œ๋ฆฌ์ฆˆ์—์„œ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ํ‰์ , ๋ฐฐ์šฐ/์ œ์ž‘์ง„์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด, ์žฅ๋ฌธ ๋ฆฌ๋ทฐ ๋“ฑ์˜ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ์„ธ ๊ฐ€์ง€ ์ž‘์—…, ์ฆ‰ ์‹œ๋†‰์‹œ์Šค ์ƒ์„ฑ, ๋ฆฌ๋ทฐ ์ƒ์„ฑ, ์ถ”์ฒœ์„ ์„ค๊ณ„ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์ž‘์—…์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•œ prompt ํ…œํ”Œ๋ฆฟ์„ ์ˆ˜๋™์œผ๋กœ ์„ค๊ณ„ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ํ…œํ”Œ๋ฆฟ์„ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ์™€ ์กฐํ•ฉํ•˜์—ฌ instructions๋ฅผ ๊ตฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹œ๋†‰์‹œ์Šค ์ƒ์„ฑ ๋ฐ ๋ฆฌ๋ทฐ ์ƒ์„ฑ์˜ ๊ฒฝ์šฐ, ์˜ํ™” ๋˜๋Š” TV ์‹œ๋ฆฌ์ฆˆ ์ด๋ฆ„๊ณผ ๊ฒฐํ•ฉ๋œ prompt ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ instructions๋ฅผ ๊ตฌ์„ฑํ•˜๊ณ , ๋‘๋ฐ˜ ์‚ฌ์šฉ์ž๊ฐ€ ์ƒ์„ฑํ•œ ์‘๋‹ต์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ธธ์ด๊ฐ€ ์ž„๊ณ„๊ฐ’๋ณด๋‹ค ์งง์€ ์‘๋‹ต์„ ์ œ๊ฑฐํ•˜๊ณ  ๊ฐœ์ธ ์ •๋ณด ๋ฐ ๊ด€๋ จ ์—†๋Š” ์ฝ˜ํ…์ธ (์˜ˆ: "๊ณต์‹ ๊ณ„์ • ๊ตฌ๋…")๋ฅผ ์‚ญ์ œํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์‘๋‹ต์˜ ์„ธ๋ถ€ ์‚ฌํ•ญ์— ๋” ์ž˜ ๋ถ€ํ•ฉํ•˜๋„๋ก ์ผ๋ถ€ instructions๋ฅผ ์ˆ˜๋™์œผ๋กœ ์กฐ์ •ํ•˜์—ฌ ๋” ๋ณต์žกํ•œ ์•”๋ฌต์  ์˜๋„๋ฅผ ์ถ”๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.
๋‘๋ฐ˜์€ ์‚ฌ์šฉ์ž๊ฐ€ ์˜ํ™”, ์ฑ…, TV ์‹œ๋ฆฌ์ฆˆ, ์Œ์•… ๋“ฑ๊ณผ ๊ฐ™์€ ๋ฌธํ•™ ๋ฐ ์˜ˆ์ˆ  ์ž‘ํ’ˆ๊ณผ ๊ด€๋ จ๋œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์†Œ์…œ ๋„คํŠธ์›Œํฌ์ด์ž ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์ž…๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ฑ…, ์˜ํ™”, TV ์‹œ๋ฆฌ์ฆˆ์—์„œ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ํ‰์ , ๋ฐฐ์šฐ/์ œ์ž‘์ง„์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด, ์žฅ๋ฌธ ๋ฆฌ๋ทฐ ๋“ฑ์˜ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ์„ธ ๊ฐ€์ง€ ์ž‘์—…, ์ฆ‰ ์‹œ๋†‰์‹œ์Šค ์ƒ์„ฑ, ๋ฆฌ๋ทฐ ์ƒ์„ฑ, ์ถ”์ฒœ์„ ์„ค๊ณ„ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์ž‘์—…์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•œ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์„ ์ˆ˜๋™์œผ๋กœ ์„ค๊ณ„ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ํ…œํ”Œ๋ฆฟ์„ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ์™€ ์กฐํ•ฉํ•˜์—ฌ ์ง€์นจ์„ ๊ตฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹œ๋†‰์‹œ์Šค ์ƒ์„ฑ ๋ฐ ๋ฆฌ๋ทฐ ์ƒ์„ฑ์˜ ๊ฒฝ์šฐ, ์˜ํ™” ๋˜๋Š” TV ์‹œ๋ฆฌ์ฆˆ ์ด๋ฆ„๊ณผ ๊ฒฐํ•ฉ๋œ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ง€์นจ์„ ๊ตฌ์„ฑํ•˜๊ณ , ๋‘๋ฐ˜ ์‚ฌ์šฉ์ž๊ฐ€ ์ƒ์„ฑํ•œ ์‘๋‹ต์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ธธ์ด๊ฐ€ ์ž„๊ณ„๊ฐ’๋ณด๋‹ค ์งง์€ ์‘๋‹ต์„ ์ œ๊ฑฐํ•˜๊ณ  ๊ฐœ์ธ ์ •๋ณด ๋ฐ ๊ด€๋ จ ์—†๋Š” ์ฝ˜ํ…์ธ (์˜ˆ: "๊ณต์‹ ๊ณ„์ • ๊ตฌ๋…")๋ฅผ ์‚ญ์ œํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์‘๋‹ต์˜ ์„ธ๋ถ€ ์‚ฌํ•ญ์— ๋” ์ž˜ ๋ถ€ํ•ฉํ•˜๋„๋ก ์ผ๋ถ€ ์ง€์นจ์„ ์ˆ˜๋™์œผ๋กœ ์กฐ์ •ํ•˜์—ฌ ๋” ๋ณต์žกํ•œ ์•”๋ฌต์  ์˜๋„๋ฅผ ์ถ”๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.
As mentioned in Section 3.4, we use the same hyperparameters and configurations as the original OPT model architecture from Zhang et al. [59]. We describe these hyperparameters briefly in Table A1. We chose these configurations because they are openly available and have been used as the standard in many previous works [1, 13, 29, 48, 59]. All models use GELU activation [18], Adam optimizer [26] with \(\beta_{1}=0.9\), \(\beta_{2}=0.95\), \(\epsilon=10^{-8}\), weight decay set to \(0.1\), and we clip gradient norms at 1.0. We use a polynomial learning rate schedule, where learning rate warms up from 0.0 to peak learning rate over the first 375 million tokens, and is then annealed to (\(0.1\) * Peak LR) over the remaining (\(T_{target}-375\)) M tokens. We train all our models in fully sharded data parallel mode [2] using Megatron-LM Tensor Parallelism [45] with fp16 precision. For reproducibility (and perhaps the only difference from the original configuration in Zhang et al. [59]) is that we do not use dropout.
3.4์ ˆ์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด Zhang ๋“ฑ์˜ ์›๋ž˜ OPT ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์™€ ๋™์ผํ•œ hyperparameters์™€ configurations์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค[59]. ์ด๋Ÿฌํ•œ hyperparameters์— ๋Œ€ํ•ด์„œ๋Š” ํ‘œ A1์— ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ configurations์„ ์„ ํƒํ•œ ์ด์œ ๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ณ  ์ด์ „์˜ ๋งŽ์€ ์ž‘์—… [1, 13, 29, 48, 59]์—์„œ ํ‘œ์ค€์œผ๋กœ ์‚ฌ์šฉ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์€ GELU activation [18], Adam optimizer [26]๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, \(\beta_{1}=0.9\), \(\beta_{2}=0.95\), \(\epsilon=10^{-8}\), weight decay๋Š” \(0.1\)๋กœ ์„ค์ •ํ•˜๊ณ  gradient norms๋ฅผ 1.0์œผ๋กœ clipํ•ฉ๋‹ˆ๋‹ค. ๋‹คํ•ญ์‹ learning rate ์ผ์ •์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฒ˜์Œ 3์–ต 7,500๋งŒ ํ† ํฐ์— ๊ฑธ์ณ learning rate๊ฐ€ 0.0์—์„œ ์ตœ๋Œ€ learning rate๋กœ ์›Œ๋ฐ์—…๋œ ๋‹ค์Œ ๋‚˜๋จธ์ง€ (\(T_{target}-375)) M ํ† ํฐ์— ๊ฑธ์ณ (\(0.1\) * Peak LR)๋กœ ์–ด๋‹๋ง๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์€ fp16 ์ •๋ฐ€๋„๋กœ Megatron-LM ํ…์„œ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ [45]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์™„์ „ํžˆ ์ƒค๋“œํ™”๋œ ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ ๋ชจ๋“œ [2]๋กœ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์žฌํ˜„์„ฑ์„ ์œ„ํ•ด(๊ทธ๋ฆฌ๊ณ  ์•„๋งˆ๋„ Zhang ๋“ฑ์˜ ์›๋ž˜ ๊ตฌ์„ฑ๊ณผ๋Š” ์œ ์ผํ•œ ์ฐจ์ด์ [59]) dropout์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
3.4์ ˆ์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด Zhang ๋“ฑ์˜ ์›๋ž˜ OPT ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์™€ ๋™์ผํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ์™€ ๊ตฌ์„ฑ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค[59]. ์ด๋Ÿฌํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ์— ๋Œ€ํ•ด์„œ๋Š” ํ‘œ A1์— ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ตฌ์„ฑ์„ ์„ ํƒํ•œ ์ด์œ ๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ณ  ์ด์ „์˜ ๋งŽ์€ ์ž‘์—… [1, 13, 29, 48, 59]์—์„œ ํ‘œ์ค€์œผ๋กœ ์‚ฌ์šฉ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์€ GELU ํ™œ์„ฑํ™” [18], Adam ์˜ตํ‹ฐ๋งˆ์ด์ € [26]๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, \(\beta_{1}=0.9\), \(\beta_{2}=0.95\), \(\epsilon=10^{-8}\), ๊ฐ€์ค‘์น˜ ๊ฐ์‡ ๋Š” \(0.1\)๋กœ ์„ค์ •ํ•˜๊ณ  ๊ทธ๋ผ๋””์–ธํŠธ ๊ทœ๋ฒ”์„ 1.0์œผ๋กœ ํด๋ฆฝํ•ฉ๋‹ˆ๋‹ค. ๋‹คํ•ญ์‹ ํ•™์Šต ์†๋„ ์ผ์ •์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฒ˜์Œ 3์–ต 7,500๋งŒ ํ† ํฐ์— ๊ฑธ์ณ ํ•™์Šต ์†๋„๊ฐ€ 0.0์—์„œ ์ตœ๋Œ€ ํ•™์Šต ์†๋„๋กœ ์›Œ๋ฐ์—…๋œ ๋‹ค์Œ ๋‚˜๋จธ์ง€ (\(T_{target}-375)) M ํ† ํฐ์— ๊ฑธ์ณ (\(0.1\) * Peak LR)๋กœ ์–ด๋‹๋ง๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์€ fp16 ์ •๋ฐ€๋„๋กœ Megatron-LM ํ…์„œ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ [45]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์™„์ „ํžˆ ์ƒค๋“œํ™”๋œ ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ ๋ชจ๋“œ [2]๋กœ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์žฌํ˜„์„ฑ์„ ์œ„ํ•ด(๊ทธ๋ฆฌ๊ณ  ์•„๋งˆ๋„ Zhang ๋“ฑ์˜ ์›๋ž˜ ๊ตฌ์„ฑ๊ณผ๋Š” ์œ ์ผํ•œ ์ฐจ์ด์ [59]) ๋“œ๋กญ์•„์›ƒ์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
**Several works investigate teaching language model agents to reason in planning (Chen et al., 2023; Gandhi et al., 2023; Qiao et al., 2024), or to use specific tools or memory (Yao et al., 2022; Lanchantin et al., 2024; Schick et al., 2024), while others investigate how one may distill the reasoning from a large language model into a smaller language model Ho et al. (2022); Li et al. (2022); Hsieh et al. (2023). Notably however, Pan et al. (2024) demonstratesthat these feedback loops may result in reward hacking. Zelikman et al. (2023) shows how a bootstrapping loop can be implemented where a model repeatedly improves a code-improver using the same code-improver and Haluptzok et al. (2023) shows how language models can bootstrap their programming ability with programming puzzles Schuster et al. (2021). Other works have employed a similar strategy for using language models to solve inductive reasoning tasks or to model real-world systems (Wang et al., 2023; Qiu et al., 2023; Zhu et al., 2023; Li et al., 2024).
**์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ ๊ณ„ํš(Chen et al., 2023; Gandhi et al., 2023; Qiao et al., 2024)์—์„œ ์ถ”๋ก ํ•˜๋„๋ก ์–ธ์–ด ๋ชจ๋ธ ์—์ด์ „ํŠธ๋ฅผ ๊ฐ€๋ฅด์น˜๊ฑฐ๋‚˜ ํŠน์ • ๋„๊ตฌ ๋˜๋Š” ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜๋„๋ก(Yao et al., 2022; Lanchantin et al., 2024; Schick et al., 2024) ์—ฐ๊ตฌํ•˜๋Š” ํ•œํŽธ, ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ์ถ”๋ก ์„ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ๋กœ ์ฆ๋ฅ˜ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Ho et al., 2022; Li et al., 2022; Hsieh et al., 2023). ๊ทธ๋Ÿฌ๋‚˜ ํŠนํžˆ Pan et al. (2024)์€ ์ด๋Ÿฌํ•œ ํ”ผ๋“œ๋ฐฑ ๋ฃจํ”„๊ฐ€ ๋ณด์ƒ ํ•ดํ‚น์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Zelikman et al. (2023)์€ ๋ชจ๋ธ์ด ๋™์ผํ•œ ์ฝ”๋“œ ๊ฐœ์„ ์ž๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฝ”๋“œ ๊ฐœ์„ ์ž๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฐœ์„ ํ•˜๋Š” bootstrapping loop๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ฃผ๊ณ , Haluptzok et al. (2023)์€ ์–ธ์–ด ๋ชจ๋ธ์ด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ํผ์ฆ์„ ํ†ตํ•ด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋Šฅ๋ ฅ์„ bootstrappingํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(Schuster et al., 2021). ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” inductive reasoning ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ฑฐ๋‚˜ ์‹ค์ œ ์‹œ์Šคํ…œ์„ ๋ชจ๋ธ๋งํ•˜๊ธฐ ์œ„ํ•ด ์–ธ์–ด ๋ชจ๋ธ์— ์œ ์‚ฌํ•œ ์ „๋žต์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค(Wang et al., 2023; Qiu et al., 2023; Zhu et al., 2023; Li et al., 2024).
**์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ ๊ณ„ํš(Chen et al., 2023; Gandhi et al., 2023; Qiao et al., 2024)์—์„œ ์ถ”๋ก ํ•˜๋„๋ก ์–ธ์–ด ๋ชจ๋ธ ์—์ด์ „ํŠธ๋ฅผ ๊ฐ€๋ฅด์น˜๊ฑฐ๋‚˜ ํŠน์ • ๋„๊ตฌ ๋˜๋Š” ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜๋„๋ก(Yao et al., 2022; Lanchantin et al., 2024; Schick et al., 2024) ์—ฐ๊ตฌํ•˜๋Š” ํ•œํŽธ, ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ์ถ”๋ก ์„ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ๋กœ ์ฆ๋ฅ˜ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Ho et al., 2022); Li et al. (2022); Hsieh et al. (2023). ๊ทธ๋Ÿฌ๋‚˜ ํŠนํžˆ Pan et al. (2024)์€ ์ด๋Ÿฌํ•œ ํ”ผ๋“œ๋ฐฑ ๋ฃจํ”„๊ฐ€ ๋ณด์ƒ ํ•ดํ‚น์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Zelikman et al. (2023)์€ ๋ชจ๋ธ์ด ๋™์ผํ•œ ์ฝ”๋“œ ๊ฐœ์„ ์ž๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฝ”๋“œ ๊ฐœ์„ ์ž๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฐœ์„ ํ•˜๋Š” ๋ถ€ํŠธ์ŠคํŠธ๋žฉ ๋ฃจํ”„๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ฃผ๊ณ , Haluptzok et al. (2023)์€ ์–ธ์–ด ๋ชจ๋ธ์ด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ํผ์ฆ์„ ํ†ตํ•ด ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋Šฅ๋ ฅ์„ ๋ถ€ํŠธ์ŠคํŠธ๋žฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(Schuster et al., 2021). ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ์œ ๋„ ์ถ”๋ก  ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ฑฐ๋‚˜ ์‹ค์ œ ์‹œ์Šคํ…œ์„ ๋ชจ๋ธ๋งํ•˜๊ธฐ ์œ„ํ•ด ์–ธ์–ด ๋ชจ๋ธ์— ์œ ์‚ฌํ•œ ์ „๋žต์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค(Wang et al., 2023; Qiu et al., 2023; Zhu et al., 2023; Li et al., 2024).
human labelers to annotate the preference for these pairs. The annotation process can be conducted in multiple forms, and a common approach is to annotate by ranking the generated candidate texts, which can reduce the inconsistency among annotators. Then, the RM is trained to predict the human-preferred output. In InstructGPT, labelers rank model-generated outputs from best to worst, and the RM (_i.e._, 6B GPT-3) is trained to predict the ranking. Note that, in recent work [371], the annotation of preference on response pairs has been conducted by an AI agent (usually an aligned LLM) instead of humans, which is called _"reinforcement learning from AI feedback (RLAIF)"_. LLMs trained with typical RLHF algorithms tend to generate harmless responses with less helpfulness, which is called _evasion problem_[371]. To guarantee both the harmlessness and helpfulness, RLAIF generates the AI feedback based on pre-set alignment principles in instructions [371, 372], which can also reduce the efforts of human annotation.
์ด๋Ÿฌํ•œ ์Œ์— ๋Œ€ํ•œ ์„ ํ˜ธ๋„๋ฅผ ์ฃผ์„์„ ๋‹ฌ๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ๋ ˆ์ด๋ธ”๋Ÿฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์„ ํ”„๋กœ์„ธ์Šค๋Š” ์—ฌ๋Ÿฌ ํ˜•ํƒœ๋กœ ์ˆ˜ํ–‰๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ƒ์„ฑ๋œ ํ›„๋ณด ํ…์ŠคํŠธ์˜ ์ˆœ์œ„๋ฅผ ๋งค๊ธฐ๋Š” ๋ฐฉ์‹์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์•„ ์ฃผ์„์ž ๊ฐ„์˜ ๋ถˆ์ผ์น˜๋ฅผ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ RM์€ ์‚ฌ๋žŒ์ด ์„ ํ˜ธํ•˜๋Š” ์ถœ๋ ฅ์„ ์˜ˆ์ธกํ•˜๋„๋ก ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. InstructGPT์—์„œ ๋ ˆ์ด๋ธ”๋Ÿฌ๋Š” ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ ์ถœ๋ ฅ์„ ๊ฐ€์žฅ ์ข‹์€ ๊ฒƒ๋ถ€ํ„ฐ ์ตœ์•…์˜ ๊ฒƒ๊นŒ์ง€ ์ˆœ์œ„๋ฅผ ๋งค๊ธฐ๊ณ , RM(_์ฆ‰, 6B GPT-3)์€ ์ˆœ์œ„๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ[371]์—์„œ๋Š” ์‘๋‹ต ์Œ์— ๋Œ€ํ•œ ์„ ํ˜ธ๋„ ์ฃผ์„์ด ์ธ๊ฐ„ ๋Œ€์‹  AI ์—์ด์ „ํŠธ(์ผ๋ฐ˜์ ์œผ๋กœ ์ •๋ ฌ๋œ LLM)์— ์˜ํ•ด ์ˆ˜ํ–‰๋˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋ฅผ _"Reinforcement Learning from AI Feedback (RLAIF)"_์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ RLHF ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ํ›ˆ๋ จ๋œ LLM์€ ์œ ์šฉ์„ฑ์ด ๋–จ์–ด์ง€๋Š” ๋ฌดํ•ดํ•œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š”๋ฐ, ์ด๋ฅผ _evasion problem_[371]์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋ฌดํ•ด์„ฑ๊ณผ ์œ ์šฉ์„ฑ์„ ๋ชจ๋‘ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด RLAIF๋Š” instructions์— ๋ฏธ๋ฆฌ ์„ค์ •๋œ ์ •๋ ฌ ์›์น™์„ ๊ธฐ๋ฐ˜์œผ๋กœ AI ํ”ผ๋“œ๋ฐฑ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค[371, 372], ์ด๋ฅผ ํ†ตํ•ด ์ธ๊ฐ„์˜ ์ฃผ์„ ์ž‘์—…๋„ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ์Œ์— ๋Œ€ํ•œ ์„ ํ˜ธ๋„๋ฅผ ์ฃผ์„์„ ๋‹ฌ๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ๋ ˆ์ด๋ธ”๋Ÿฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์„ ํ”„๋กœ์„ธ์Šค๋Š” ์—ฌ๋Ÿฌ ํ˜•ํƒœ๋กœ ์ˆ˜ํ–‰๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ƒ์„ฑ๋œ ํ›„๋ณด ํ…์ŠคํŠธ์˜ ์ˆœ์œ„๋ฅผ ๋งค๊ธฐ๋Š” ๋ฐฉ์‹์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์•„ ์ฃผ์„์ž ๊ฐ„์˜ ๋ถˆ์ผ์น˜๋ฅผ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ RM์€ ์‚ฌ๋žŒ์ด ์„ ํ˜ธํ•˜๋Š” ์ถœ๋ ฅ์„ ์˜ˆ์ธกํ•˜๋„๋ก ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. InstructGPT์—์„œ ๋ ˆ์ด๋ธ”๋Ÿฌ๋Š” ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ ์ถœ๋ ฅ์„ ๊ฐ€์žฅ ์ข‹์€ ๊ฒƒ๋ถ€ํ„ฐ ์ตœ์•…์˜ ๊ฒƒ๊นŒ์ง€ ์ˆœ์œ„๋ฅผ ๋งค๊ธฐ๊ณ , RM(_์ฆ‰, 6B GPT-3)์€ ์ˆœ์œ„๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ[371]์—์„œ๋Š” ์‘๋‹ต ์Œ์— ๋Œ€ํ•œ ์„ ํ˜ธ๋„ ์ฃผ์„์ด ์ธ๊ฐ„ ๋Œ€์‹  AI ์—์ด์ „ํŠธ(์ผ๋ฐ˜์ ์œผ๋กœ ์ •๋ ฌ๋œ LLM)์— ์˜ํ•ด ์ˆ˜ํ–‰๋˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋ฅผ _"AI ํ”ผ๋“œ๋ฐฑ์œผ๋กœ๋ถ€ํ„ฐ์˜ ๊ฐ•ํ™” ํ•™์Šต(RLAIF)"_์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ RLHF ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ํ›ˆ๋ จ๋œ LLM์€ ์œ ์šฉ์„ฑ์ด ๋–จ์–ด์ง€๋Š” ๋ฌดํ•ดํ•œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š”๋ฐ, ์ด๋ฅผ _ํšŒํ”ผ ๋ฌธ์ œ_[371]๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋ฌดํ•ด์„ฑ๊ณผ ์œ ์šฉ์„ฑ์„ ๋ชจ๋‘ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด RLAIF๋Š” ์ง€์นจ์— ๋ฏธ๋ฆฌ ์„ค์ •๋œ ์ •๋ ฌ ์›์น™์„ ๊ธฐ๋ฐ˜์œผ๋กœ AI ํ”ผ๋“œ๋ฐฑ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค[371, 372], ์ด๋ฅผ ํ†ตํ•ด ์ธ๊ฐ„์˜ ์ฃผ์„ ์ž‘์—…๋„ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
For unanswerable case, we consider the model indicating that the question cannot be answered as correct. For answerable cases, we consider the model not indicating the question is unanswerable as correct (i.e., the model giving an answer). We find that ChatQA always generates "Sorry. I cannot find the answer based on the context" when it finds the question is unanswerable, since we use this sentence as a response for unanswerable questions in the stage-2 tuning. As for Llama2-Chat, GPT-3.5-turbo, and GPT-4, their outputs for unanswerable questions usually follow some specific patterns. Below, we list the heuristic matching patterns derived from all generated samples in DoQA and QuAC datasets to determine if the model suggests that the question is unanswerable. When we find the generated output contains the pattern listed below, we interpret it as an indication that the question is unanswerable, and we will replace its answer with "Sorry. I cannot find the answer based on the context." for the F1 evaluation.
๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ, ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Œ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๋ชจ๋ธ์„ ์ •๋‹ต์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ๋‹ต๋ณ€ ๊ฐ€๋Šฅํ•œ ๊ฒฝ์šฐ, ๋ชจ๋ธ์ด ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Œ์„ ๋‚˜ํƒ€๋‚ด์ง€ ์•Š๋Š” ๋ชจ๋ธ์„ ์ •๋‹ต์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค(์ฆ‰, ๋ชจ๋ธ์ด ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๊ฒฝ์šฐ). ChatQA๋Š” 2๋‹จ๊ณ„ tuning ๋‹จ๊ณ„์—์„œ ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์‘๋‹ต์œผ๋กœ ์ด ๋ฌธ์žฅ์„ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ ํ•ญ์ƒ "์ฃ„์†กํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ๋งฅ์— ๋”ฐ๋ผ ๋‹ต๋ณ€์„ ์ฐพ์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค."๋ฅผ ์ƒ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Llama2-Chat, GPT-3.5-turbo ๋ฐ GPT-4์˜ ๊ฒฝ์šฐ, ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์ถœ๋ ฅ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ช‡ ๊ฐ€์ง€ ํŠน์ • ํŒจํ„ด์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์•„๋ž˜์—์„œ๋Š” DoQA ๋ฐ QuAC ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ƒ์„ฑ๋œ ๋ชจ๋“  ์ƒ˜ํ”Œ์—์„œ ๋„์ถœ๋œ heuristic matching ํŒจํ„ด์„ ๋‚˜์—ดํ•˜์—ฌ ๋ชจ๋ธ์ด ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Œ์„ ์ œ์•ˆํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ƒ์„ฑ๋œ ์ถœ๋ ฅ์— ์•„๋ž˜์— ๋‚˜์—ด๋œ ํŒจํ„ด์ด ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉด ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Œ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฒƒ์œผ๋กœ ํ•ด์„ํ•˜๊ณ , F1 ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ๋‹ต๋ณ€์„ "์ฃ„์†กํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ๋งฅ์— ๋”ฐ๋ผ ๋‹ต๋ณ€์„ ์ฐพ์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค."๋กœ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค.
๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ, ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Œ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๋ชจ๋ธ์„ ์ •๋‹ต์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ๋‹ต๋ณ€ ๊ฐ€๋Šฅํ•œ ๊ฒฝ์šฐ, ๋ชจ๋ธ์ด ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Œ์„ ๋‚˜ํƒ€๋‚ด์ง€ ์•Š๋Š” ๋ชจ๋ธ์„ ์ •๋‹ต์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค(์ฆ‰, ๋ชจ๋ธ์ด ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๊ฒฝ์šฐ). ChatQA๋Š” 2๋‹จ๊ณ„ ํŠœ๋‹ ๋‹จ๊ณ„์—์„œ ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์‘๋‹ต์œผ๋กœ ์ด ๋ฌธ์žฅ์„ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ ํ•ญ์ƒ "์ฃ„์†กํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ๋งฅ์— ๋”ฐ๋ผ ๋‹ต๋ณ€์„ ์ฐพ์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค."๋ฅผ ์ƒ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Llama2-Chat, GPT-3.5-turbo ๋ฐ GPT-4์˜ ๊ฒฝ์šฐ, ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์ถœ๋ ฅ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ช‡ ๊ฐ€์ง€ ํŠน์ • ํŒจํ„ด์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์•„๋ž˜์—์„œ๋Š” DoQA ๋ฐ QuAC ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ƒ์„ฑ๋œ ๋ชจ๋“  ์ƒ˜ํ”Œ์—์„œ ๋„์ถœ๋œ ํœด๋ฆฌ์Šคํ‹ฑ ๋งค์นญ ํŒจํ„ด์„ ๋‚˜์—ดํ•˜์—ฌ ๋ชจ๋ธ์ด ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Œ์„ ์ œ์•ˆํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ƒ์„ฑ๋œ ์ถœ๋ ฅ์— ์•„๋ž˜์— ๋‚˜์—ด๋œ ํŒจํ„ด์ด ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉด ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†์Œ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฒƒ์œผ๋กœ ํ•ด์„ํ•˜๊ณ , F1 ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ๋‹ต๋ณ€์„ "์ฃ„์†กํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ๋งฅ์— ๋”ฐ๋ผ ๋‹ต๋ณ€์„ ์ฐพ์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค."๋กœ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค.
Generative AI systems have the potential to cause harm in myriad ways, affecting different people, groups, societies and environments across the globe [71]. This includes physical, emotional, financial, allocative, reputational, representational, and psychological harms [72, 16, 73]. Such harms can be caused by using generative AI systems [74], being excluded from them [75], being represented or described by them [76, 77], or being subjected to decisions made by them [78]. Key considerations when assessing harm include whether the harm is tangible or intangible, short- or long-term in duration, highly severe or less severe in nature, inflicted on oneself or on others, or internalized or externalized in its expression [79, 71, 80, 81]. Experiences of harm are often shaped by the context in which the harm is inflicted and can be affected by a range of risk factors. Aspects like the users' background, life experiences, personality, and past behavior can all impact whether they experience harm [82, 83, 84, 85].
์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI ์‹œ์Šคํ…œ์€ ์ „ ์„ธ๊ณ„์˜ ๋‹ค์–‘ํ•œ ์‚ฌ๋žŒ, ๊ทธ๋ฃน, ์‚ฌํšŒ ๋ฐ ํ™˜๊ฒฝ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ ํ”ผํ•ด๋ฅผ ์ž…ํž ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[71]. ์—ฌ๊ธฐ์—๋Š” ์‹ ์ฒด์ , ์ •์„œ์ , ์žฌ์ •์ , ํ• ๋‹น์ , ํ‰ํŒ์ , ๋Œ€ํ‘œ์ , ์‹ฌ๋ฆฌ์  ํ”ผํ•ด๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค[72, 16, 73]. ์ด๋Ÿฌํ•œ ํ”ผํ•ด๋Š” generative AI ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•  ๋•Œ[74], generative AI ์‹œ์Šคํ…œ์—์„œ ์ œ์™ธ๋  ๋•Œ[75], generative AI ์‹œ์Šคํ…œ์— ์˜ํ•ด ํ‘œํ˜„๋˜๊ฑฐ๋‚˜ ์„ค๋ช…๋  ๋•Œ[76, 77], ๋˜๋Š” generative AI ์‹œ์Šคํ…œ์— ์˜ํ•ด ๋‚ด๋ ค์ง„ ๊ฒฐ์ •์— ๋…ธ์ถœ๋  ๋•Œ[78] ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ”ผํ•ด๋ฅผ ํ‰๊ฐ€ํ•  ๋•Œ ๊ณ ๋ คํ•ด์•ผ ํ•  ์ฃผ์š” ์‚ฌํ•ญ์œผ๋กœ๋Š” ํ”ผํ•ด๊ฐ€ ์œ ํ˜•์ ์ด๊ฑฐ๋‚˜ ๋ฌดํ˜•์ ์ธ์ง€ ์—ฌ๋ถ€, ํ”ผํ•ด์˜ ์ง€์† ์‹œ๊ฐ„์ด ๋‹จ๊ธฐ์ธ์ง€ ์žฅ๊ธฐ์ธ์ง€, ํ”ผํ•ด์˜ ์ •๋„๊ฐ€ ๋งค์šฐ ์‹ฌ๊ฐํ•œ์ง€ ๋˜๋Š” ๋œ ์‹ฌ๊ฐํ•œ์ง€, ์ž์‹ ์—๊ฒŒ ๊ฐ€ํ•ด์ง€๋Š”์ง€ ๋˜๋Š” ํƒ€์ธ์—๊ฒŒ ๊ฐ€ํ•ด์ง€๋Š”์ง€, ํ”ผํ•ด๊ฐ€ ๋‚ด๋ฉดํ™”๋˜๋Š”์ง€ ๋˜๋Š” ์™ธ๋ถ€ํ™”๋˜๋Š”์ง€ ์—ฌ๋ถ€ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค[79, 71, 80, 81]. ํ”ผํ•ด ๊ฒฝํ—˜์€ ํ”ผํ•ด๊ฐ€ ๊ฐ€ํ•ด์ง€๋Š” ๋งฅ๋ฝ์— ๋”ฐ๋ผ ํ˜•์„ฑ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ ๋‹ค์–‘ํ•œ ์œ„ํ—˜ ์š”์ธ์— ์˜ํ•ด ์˜ํ–ฅ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž์˜ ๋ฐฐ๊ฒฝ, ์ƒํ™œ ๊ฒฝํ—˜, ์„ฑ๊ฒฉ, ๊ณผ๊ฑฐ ํ–‰๋™๊ณผ ๊ฐ™์€ ์ธก๋ฉด์€ ๋ชจ๋‘ ํ”ผํ•ด ๊ฒฝํ—˜ ์—ฌ๋ถ€์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[82, 83, 84, 85].
์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI ์‹œ์Šคํ…œ์€ ์ „ ์„ธ๊ณ„์˜ ๋‹ค์–‘ํ•œ ์‚ฌ๋žŒ, ๊ทธ๋ฃน, ์‚ฌํšŒ ๋ฐ ํ™˜๊ฒฝ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ ํ”ผํ•ด๋ฅผ ์ž…ํž ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[71]. ์—ฌ๊ธฐ์—๋Š” ์‹ ์ฒด์ , ์ •์„œ์ , ์žฌ์ •์ , ํ• ๋‹น์ , ํ‰ํŒ์ , ๋Œ€ํ‘œ์ , ์‹ฌ๋ฆฌ์  ํ”ผํ•ด๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค[72, 16, 73]. ์ด๋Ÿฌํ•œ ํ”ผํ•ด๋Š” ์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•  ๋•Œ[74], ์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI ์‹œ์Šคํ…œ์—์„œ ์ œ์™ธ๋  ๋•Œ[75], ์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI ์‹œ์Šคํ…œ์— ์˜ํ•ด ํ‘œํ˜„๋˜๊ฑฐ๋‚˜ ์„ค๋ช…๋  ๋•Œ[76, 77], ๋˜๋Š” ์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI ์‹œ์Šคํ…œ์— ์˜ํ•ด ๋‚ด๋ ค์ง„ ๊ฒฐ์ •์— ๋…ธ์ถœ๋  ๋•Œ[78] ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ”ผํ•ด๋ฅผ ํ‰๊ฐ€ํ•  ๋•Œ ๊ณ ๋ คํ•ด์•ผ ํ•  ์ฃผ์š” ์‚ฌํ•ญ์œผ๋กœ๋Š” ํ”ผํ•ด๊ฐ€ ์œ ํ˜•์ ์ด๊ฑฐ๋‚˜ ๋ฌดํ˜•์ ์ธ์ง€ ์—ฌ๋ถ€, ํ”ผํ•ด์˜ ์ง€์† ์‹œ๊ฐ„์ด ๋‹จ๊ธฐ์ธ์ง€ ์žฅ๊ธฐ์ธ์ง€, ํ”ผํ•ด์˜ ์ •๋„๊ฐ€ ๋งค์šฐ ์‹ฌ๊ฐํ•œ์ง€ ๋˜๋Š” ๋œ ์‹ฌ๊ฐํ•œ์ง€, ์ž์‹ ์—๊ฒŒ ๊ฐ€ํ•ด์ง€๋Š”์ง€ ๋˜๋Š” ํƒ€์ธ์—๊ฒŒ ๊ฐ€ํ•ด์ง€๋Š”์ง€, ํ”ผํ•ด๊ฐ€ ๋‚ด๋ฉดํ™”๋˜๋Š”์ง€ ๋˜๋Š” ์™ธ๋ถ€ํ™”๋˜๋Š”์ง€ ์—ฌ๋ถ€ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค[79, 71, 80, 81]. ํ”ผํ•ด ๊ฒฝํ—˜์€ ํ”ผํ•ด๊ฐ€ ๊ฐ€ํ•ด์ง€๋Š” ๋งฅ๋ฝ์— ๋”ฐ๋ผ ํ˜•์„ฑ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ ๋‹ค์–‘ํ•œ ์œ„ํ—˜ ์š”์ธ์— ์˜ํ•ด ์˜ํ–ฅ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž์˜ ๋ฐฐ๊ฒฝ, ์ƒํ™œ ๊ฒฝํ—˜, ์„ฑ๊ฒฉ, ๊ณผ๊ฑฐ ํ–‰๋™๊ณผ ๊ฐ™์€ ์ธก๋ฉด์€ ๋ชจ๋‘ ํ”ผํ•ด ๊ฒฝํ—˜ ์—ฌ๋ถ€์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[82, 83, 84, 85].
**Organization of this review.** In this paper, we first give a preliminary preview of the working principle of the State Space Model in Section 2. Then, in Section 3, we focus on reviewing the related works of SSMs from multiple aspects, including origin and variation of SSMs, natural language processing, computer vision, graph, multi-modal and multi-media, point cloud/event stream, time series data, and other domains. An overview of the structure and key State Space Model related papers reviewed in this survey is illustrated in Fig. 3. More importantly, we conduct extensive experiments on multiple downstream tasks to validate the effectiveness of SSMs in Section 4. The downstream tasks involve single-/ multi-label classification, visual object tracking, pixel-level segmentation, image-to-text generation, and person/vehicle re-identification. We also propose several possible research directions to the theory and applications of SSMs in Section 5. Finally, we give a conclusion about this paper in Section 6.
**์ด ๋ฆฌ๋ทฐ์˜ ๊ตฌ์„ฑ**. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋จผ์ € ์„น์…˜ 2์—์„œ State Space Model์˜ ์ž‘๋™ ์›๋ฆฌ์— ๋Œ€ํ•œ ์˜ˆ๋น„์ ์ธ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์„น์…˜ 3์—์„œ๋Š” SSM์˜ ๊ธฐ์›๊ณผ ๋ณ€ํ˜•, ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ, ์ปดํ“จํ„ฐ ๋น„์ „, ๊ทธ๋ž˜ํ”„, ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ฐ ๋ฉ€ํ‹ฐ๋ฏธ๋””์–ด, ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ/์ด๋ฒคํŠธ ์ŠคํŠธ๋ฆผ, ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ ๋ฐ ๊ธฐํƒ€ ์˜์—ญ์„ ํฌํ•จํ•œ ์—ฌ๋Ÿฌ ์ธก๋ฉด์—์„œ SSM์˜ ๊ด€๋ จ ์—ฐ๊ตฌ๋ฅผ ๊ฒ€ํ† ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ์—์„œ ๊ฒ€ํ† ๋œ ๊ตฌ์กฐ์™€ ์ฃผ์š” State Space Model ๊ด€๋ จ ๋…ผ๋ฌธ์— ๋Œ€ํ•œ ๊ฐœ์š”๋Š” ๊ทธ๋ฆผ 3์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”์šฑ ์ค‘์š”ํ•œ ๊ฒƒ์€, ์„น์…˜ 4์—์„œ ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์—ฌ SSM์˜ ํšจ๊ณผ๋ฅผ ๊ฒ€์ฆํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—๋Š” ๋‹จ์ผ/๋‹ค์ค‘ ๋ ˆ์ด๋ธ” ๋ถ„๋ฅ˜, ์‹œ๊ฐ์  ๊ฐ์ฒด ์ถ”์ , ํ”ฝ์…€ ์ˆ˜์ค€ ๋ถ„ํ• , ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์ƒ์„ฑ, ์‚ฌ๋žŒ/์ฐจ๋Ÿ‰ ์žฌ์‹๋ณ„ ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์„น์…˜ 5์—์„œ๋Š” SSM์˜ ์ด๋ก ๊ณผ ์‘์šฉ์— ๋Œ€ํ•ด ๋ช‡ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ์„น์…˜ 6์—์„œ๋Š” ์ด ๋…ผ๋ฌธ์— ๋Œ€ํ•œ ๊ฒฐ๋ก ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
**์ด ๋ฆฌ๋ทฐ์˜ ๊ตฌ์„ฑ**. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋จผ์ € ์„น์…˜ 2์—์„œ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์˜ ์ž‘๋™ ์›๋ฆฌ์— ๋Œ€ํ•œ ์˜ˆ๋น„์ ์ธ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์„น์…˜ 3์—์„œ๋Š” SSM์˜ ๊ธฐ์›๊ณผ ๋ณ€ํ˜•, ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ, ์ปดํ“จํ„ฐ ๋น„์ „, ๊ทธ๋ž˜ํ”„, ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ฐ ๋ฉ€ํ‹ฐ๋ฏธ๋””์–ด, ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ/์ด๋ฒคํŠธ ์ŠคํŠธ๋ฆผ, ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ ๋ฐ ๊ธฐํƒ€ ์˜์—ญ์„ ํฌํ•จํ•œ ์—ฌ๋Ÿฌ ์ธก๋ฉด์—์„œ SSM์˜ ๊ด€๋ จ ์—ฐ๊ตฌ๋ฅผ ๊ฒ€ํ† ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ์—์„œ ๊ฒ€ํ† ๋œ ๊ตฌ์กฐ์™€ ์ฃผ์š” ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ ๊ด€๋ จ ๋…ผ๋ฌธ์— ๋Œ€ํ•œ ๊ฐœ์š”๋Š” ๊ทธ๋ฆผ 3์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”์šฑ ์ค‘์š”ํ•œ ๊ฒƒ์€, ์„น์…˜ 4์—์„œ ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์—ฌ SSM์˜ ํšจ๊ณผ๋ฅผ ๊ฒ€์ฆํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—๋Š” ๋‹จ์ผ/๋‹ค์ค‘ ๋ ˆ์ด๋ธ” ๋ถ„๋ฅ˜, ์‹œ๊ฐ์  ๊ฐ์ฒด ์ถ”์ , ํ”ฝ์…€ ์ˆ˜์ค€ ๋ถ„ํ• , ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์ƒ์„ฑ, ์‚ฌ๋žŒ/์ฐจ๋Ÿ‰ ์žฌ์‹๋ณ„ ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์„น์…˜ 5์—์„œ๋Š” SSM์˜ ์ด๋ก ๊ณผ ์‘์šฉ์— ๋Œ€ํ•ด ๋ช‡ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ์„น์…˜ 6์—์„œ๋Š” ์ด ๋…ผ๋ฌธ์— ๋Œ€ํ•œ ๊ฒฐ๋ก ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
**Simulated Dialogue.** The use of simulated data allowed us to quickly scale training to a broad set of conditions and patient contexts, while the injection of knowledge from search encouraged these dialogues to remain grounded and realistic. Though the simulated patients encompassed a wide range of conditions, they failed to capture the full range of potential patient backgrounds, personalities, and motivations. Through the inner self-play procedure, we were able to iteratively improve the simulated dialogue we generated and used in fine-tuning. However, these improvements were limited by our ability to articulate what makes a good dialogue in the critic instructions, the critic's ability to produce effective feedback, and AMIE's ability to adapt to such feedback. For example, in the simulated environment we impose that AMIE reaches a proposed differential and testing/treatment plan for the patient, but such an endpoint may be unrealistic for some conditions, especially in the virtual chat-based setting.
**Simulated Dialogue**. ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ด‘๋ฒ”์œ„ํ•œ ์กฐ๊ฑด๊ณผ ํ™˜์ž ์ƒํ™ฉ์— ๋Œ€ํ•œ ๊ต์œก์„ ๋น ๋ฅด๊ฒŒ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์—ˆ๊ณ , ๊ฒ€์ƒ‰์—์„œ ์ง€์‹์„ ์‚ฝ์ž…ํ•˜๋ฉด ์ด๋Ÿฌํ•œ ๋Œ€ํ™”๊ฐ€ ๊ทผ๊ฑฐ ์žˆ๊ณ  ํ˜„์‹ค๊ฐ์„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜์ž๋Š” ๋‹ค์–‘ํ•œ ์กฐ๊ฑด์„ ํฌ๊ด„ํ•˜์ง€๋งŒ ์ž ์žฌ์ ์ธ ํ™˜์ž์˜ ๋ฐฐ๊ฒฝ, ์„ฑ๊ฒฉ ๋ฐ ๋™๊ธฐ๋ฅผ ๋ชจ๋‘ ํฌ์ฐฉํ•˜์ง€๋Š” ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‚ด๋ฉด์˜ ์ž๊ธฐ ํ”Œ๋ ˆ์ด ์ ˆ์ฐจ๋ฅผ ํ†ตํ•ด Fine-tuning์— ์‚ฌ์šฉํ•˜๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™”๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์€ ๋น„ํ‰๊ฐ€ ์ง€์นจ์—์„œ ์ข‹์€ ๋Œ€ํ™”๋ฅผ ๋งŒ๋“œ๋Š” ์š”์†Œ๋ฅผ ๋ช…ํ™•ํžˆ ํ‘œํ˜„ํ•˜๋Š” ๋Šฅ๋ ฅ, ๋น„ํ‰๊ฐ€๊ฐ€ ํšจ๊ณผ์ ์ธ ํ”ผ๋“œ๋ฐฑ์„ ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ, AMIE๊ฐ€ ์ด๋Ÿฌํ•œ ํ”ผ๋“œ๋ฐฑ์— ์ ์‘ํ•˜๋Š” ๋Šฅ๋ ฅ์— ์˜ํ•ด ์ œํ•œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์—์„œ๋Š” AMIE๊ฐ€ ํ™˜์ž์— ๋Œ€ํ•œ ์ฐจ๋“ฑ ์ง„๋‹จ ๋ฐ ๊ฒ€์‚ฌ/์น˜๋ฃŒ ๊ณ„ํš์— ๋„๋‹ฌํ•˜๋„๋ก ๊ฐ•์š”ํ•˜์ง€๋งŒ, ์ผ๋ถ€ ์งˆํ™˜์˜ ๊ฒฝ์šฐ, ํŠนํžˆ ๊ฐ€์ƒ ์ฑ„ํŒ… ๊ธฐ๋ฐ˜ ํ™˜๊ฒฝ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์ข…์ฐฉ์ ์ด ๋น„ํ˜„์‹ค์ ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**๋ชจ์˜ ๋Œ€ํ™”**. ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ด‘๋ฒ”์œ„ํ•œ ์กฐ๊ฑด๊ณผ ํ™˜์ž ์ƒํ™ฉ์— ๋Œ€ํ•œ ๊ต์œก์„ ๋น ๋ฅด๊ฒŒ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์—ˆ๊ณ , ๊ฒ€์ƒ‰์—์„œ ์ง€์‹์„ ์‚ฝ์ž…ํ•˜๋ฉด ์ด๋Ÿฌํ•œ ๋Œ€ํ™”๊ฐ€ ๊ทผ๊ฑฐ ์žˆ๊ณ  ํ˜„์‹ค๊ฐ์„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜์ž๋Š” ๋‹ค์–‘ํ•œ ์กฐ๊ฑด์„ ํฌ๊ด„ํ•˜์ง€๋งŒ ์ž ์žฌ์ ์ธ ํ™˜์ž์˜ ๋ฐฐ๊ฒฝ, ์„ฑ๊ฒฉ ๋ฐ ๋™๊ธฐ๋ฅผ ๋ชจ๋‘ ํฌ์ฐฉํ•˜์ง€๋Š” ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‚ด๋ฉด์˜ ์ž๊ธฐ ํ”Œ๋ ˆ์ด ์ ˆ์ฐจ๋ฅผ ํ†ตํ•ด ๋ฏธ์„ธ ์กฐ์ •์— ์‚ฌ์šฉํ•˜๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™”๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์€ ๋น„ํ‰๊ฐ€ ์ง€์นจ์—์„œ ์ข‹์€ ๋Œ€ํ™”๋ฅผ ๋งŒ๋“œ๋Š” ์š”์†Œ๋ฅผ ๋ช…ํ™•ํžˆ ํ‘œํ˜„ํ•˜๋Š” ๋Šฅ๋ ฅ, ๋น„ํ‰๊ฐ€๊ฐ€ ํšจ๊ณผ์ ์ธ ํ”ผ๋“œ๋ฐฑ์„ ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ, AMIE๊ฐ€ ์ด๋Ÿฌํ•œ ํ”ผ๋“œ๋ฐฑ์— ์ ์‘ํ•˜๋Š” ๋Šฅ๋ ฅ์— ์˜ํ•ด ์ œํ•œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์—์„œ๋Š” AMIE๊ฐ€ ํ™˜์ž์— ๋Œ€ํ•œ ์ฐจ๋“ฑ ์ง„๋‹จ ๋ฐ ๊ฒ€์‚ฌ/์น˜๋ฃŒ ๊ณ„ํš์— ๋„๋‹ฌํ•˜๋„๋ก ๊ฐ•์š”ํ•˜์ง€๋งŒ, ์ผ๋ถ€ ์งˆํ™˜์˜ ๊ฒฝ์šฐ, ํŠนํžˆ ๊ฐ€์ƒ ์ฑ„ํŒ… ๊ธฐ๋ฐ˜ ํ™˜๊ฒฝ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์ข…์ฐฉ์ ์ด ๋น„ํ˜„์‹ค์ ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Dialog-Enabled Resolving Agents (DERA) [177] are specialized AI agents that can engage in dialogue, resolve queries, and make decisions based on interactive exchanges. DERA is developed based on the idea of utilizing multiple agents within a dialog context, each with specific roles and functions. These agents can include Researchers, who gather and analyze information, and Deciders, who make final judgments based on the information provided. This division of roles allows for a well-organized and efficient approach to problem-solving and decision-making. DERA is particularly advantageous in scenarios requiring complex decision-making and problem-solving, such as those in medical diagnostics or customer service. The collaborative and interactive nature of DERA agents allows them to handle intricate queries with a level of depth and nuance that single-agent systems might struggle with. Moreover, this approach aligns well with human decision-making processes, making AI reasoning more relatable and trustworthy.
๋Œ€ํ™” ์ง€์› ํ•ด๊ฒฐ ์—์ด์ „ํŠธ(DERA) [177]๋Š” ๋Œ€ํ™”์— ์ฐธ์—ฌํ•˜๊ณ , ์ฟผ๋ฆฌ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ , ๋Œ€ํ™” ์ƒํ˜ธ์ž‘์šฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์˜์‚ฌ ๊ฒฐ์ •์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋Š” ํŠน์ˆ˜ AI ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค. DERA๋Š” ๋Œ€ํ™” ์ปจํ…์ŠคํŠธ ๋‚ด์—์„œ ๊ฐ๊ฐ ํŠน์ • ์—ญํ• ๊ณผ ๊ธฐ๋Šฅ์„ ๊ฐ€์ง„ ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๋ฅผ ํ™œ์šฉํ•œ๋‹ค๋Š” ์•„์ด๋””์–ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—์ด์ „ํŠธ์—๋Š” ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ๋ถ„์„ํ•˜๋Š” Researchers์™€ ์ œ๊ณต๋œ ์ •๋ณด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์ตœ์ข… ํŒ๋‹จ์„ ๋‚ด๋ฆฌ๋Š” Deciders๊ฐ€ ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—ญํ•  ๋ถ„๋‹ด์€ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋ฐ ์˜์‚ฌ ๊ฒฐ์ •์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ด๊ณ  ํšจ์œจ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. DERA๋Š” ์˜๋ฃŒ ์ง„๋‹จ์ด๋‚˜ ๊ณ ๊ฐ ์„œ๋น„์Šค์™€ ๊ฐ™์ด ๋ณต์žกํ•œ ์˜์‚ฌ ๊ฒฐ์ •๊ณผ ๋ฌธ์ œ ํ•ด๊ฒฐ์ด ํ•„์š”ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ํŠนํžˆ ์œ ๋ฆฌํ•ฉ๋‹ˆ๋‹ค. DERA ์—์ด์ „ํŠธ์˜ ํ˜‘์—…์ ์ด๊ณ  ์ธํ„ฐ๋ž™ํ‹ฐ๋ธŒํ•œ ํŠน์„ฑ ๋•๋ถ„์— ๋‹จ์ผ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์—์„œ๋Š” ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ค์šด ๋ณต์žกํ•œ ์ฟผ๋ฆฌ๋ฅผ ๊นŠ์ด ์žˆ๊ณ  ๋ฏธ๋ฌ˜ํ•˜๊ฒŒ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ธ๊ฐ„์˜ ์˜์‚ฌ ๊ฒฐ์ • ๊ณผ์ •๊ณผ ์ž˜ ๋ถ€ํ•ฉํ•˜์—ฌ AI ์ถ”๋ก ์„ ๋”์šฑ ์นœ๊ทผํ•˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค.
๋Œ€ํ™” ์ง€์› ํ•ด๊ฒฐ ์—์ด์ „ํŠธ(DERA) [177]๋Š” ๋Œ€ํ™”์— ์ฐธ์—ฌํ•˜๊ณ , ์ฟผ๋ฆฌ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ , ๋Œ€ํ™” ์ƒํ˜ธ์ž‘์šฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์˜์‚ฌ ๊ฒฐ์ •์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋Š” ํŠน์ˆ˜ AI ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค. DERA๋Š” ๋Œ€ํ™” ์ปจํ…์ŠคํŠธ ๋‚ด์—์„œ ๊ฐ๊ฐ ํŠน์ • ์—ญํ• ๊ณผ ๊ธฐ๋Šฅ์„ ๊ฐ€์ง„ ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๋ฅผ ํ™œ์šฉํ•œ๋‹ค๋Š” ์•„์ด๋””์–ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—์ด์ „ํŠธ์—๋Š” ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ๋ถ„์„ํ•˜๋Š” ๋ฆฌ์„œ์ฒ˜์™€ ์ œ๊ณต๋œ ์ •๋ณด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์ตœ์ข… ํŒ๋‹จ์„ ๋‚ด๋ฆฌ๋Š” ์˜์‚ฌ ๊ฒฐ์ •์ž๊ฐ€ ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—ญํ•  ๋ถ„๋‹ด์€ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋ฐ ์˜์‚ฌ ๊ฒฐ์ •์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ด๊ณ  ํšจ์œจ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. DERA๋Š” ์˜๋ฃŒ ์ง„๋‹จ์ด๋‚˜ ๊ณ ๊ฐ ์„œ๋น„์Šค์™€ ๊ฐ™์ด ๋ณต์žกํ•œ ์˜์‚ฌ ๊ฒฐ์ •๊ณผ ๋ฌธ์ œ ํ•ด๊ฒฐ์ด ํ•„์š”ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ํŠนํžˆ ์œ ๋ฆฌํ•ฉ๋‹ˆ๋‹ค. DERA ์—์ด์ „ํŠธ์˜ ํ˜‘์—…์ ์ด๊ณ  ์ธํ„ฐ๋ž™ํ‹ฐ๋ธŒํ•œ ํŠน์„ฑ ๋•๋ถ„์— ๋‹จ์ผ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์—์„œ๋Š” ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ค์šด ๋ณต์žกํ•œ ์ฟผ๋ฆฌ๋ฅผ ๊นŠ์ด ์žˆ๊ณ  ๋ฏธ๋ฌ˜ํ•˜๊ฒŒ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ธ๊ฐ„์˜ ์˜์‚ฌ ๊ฒฐ์ • ๊ณผ์ •๊ณผ ์ž˜ ๋ถ€ํ•ฉํ•˜์—ฌ AI ์ถ”๋ก ์„ ๋”์šฑ ์นœ๊ทผํ•˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค.
Quiet-STaR represents a step towards language models that can learn to reason in a general and scalable way. By training on the rich spectrum of reasoning tasks implicit in diverse web text, rather than narrowly specializing for particular datasets, Quiet-STaR points the way to more robust and adaptable language models. Our results demonstrate the promise of this approach, with Quiet-STaR improving downstream reasoning performance while generating qualitatively meaningful rationales. We believe this also opens many potential future directions - for example, one may aim to ensemble thoughts in order to further improve the predictions for future tokens. Moreover, if the language model can predict when thought will be useful, for example by putting the mixing head before the prediction, then the predicted mixing weight could be used to dynamically allocate compute during generation. Future work can build on these insights to further close the gap between language model and human-like reasoning capabilities.
Quiet-STaR์€ ์ผ๋ฐ˜์ ์ด๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ฐฉ์‹์œผ๋กœ ์ถ”๋ก ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ํ–ฅํ•œ ํ•œ ๊ฑธ์Œ์ž…๋‹ˆ๋‹ค. ํŠน์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ข๊ฒŒ ํŠนํ™”๋˜์ง€ ์•Š๊ณ  ๋‹ค์–‘ํ•œ ์›น ํ…์ŠคํŠธ์— ๋‚ดํฌ๋œ ํ’๋ถ€ํ•œ ์ถ”๋ก  ๊ณผ์ œ์— ๋Œ€ํ•ด ํ•™์Šตํ•จ์œผ๋กœ์จ Quiet-STaR์€ ๋ณด๋‹ค ๊ฐ•๋ ฅํ•˜๊ณ  ์ ์‘๋ ฅ ์žˆ๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ํ–ฅํ•œ ๊ธธ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ๊ฒฐ๊ณผ๋Š” ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ๊ฐ€๋Šฅ์„ฑ์„ ์ž…์ฆํ•˜๋ฉฐ, Quiet-STaR์€ ์–‘์งˆ์˜ ์˜๋ฏธ ์žˆ๋Š” ๊ทผ๊ฑฐ๋ฅผ ์ƒ์„ฑํ•˜๋ฉด์„œ downstream reasoning ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ–ฅํ›„ ํ† ํฐ์— ๋Œ€ํ•œ ์˜ˆ์ธก์„ ๋”์šฑ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ƒ๊ฐ์„ ์•™์ƒ๋ธ”ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์˜ˆ์ธก ์ „์— mixing head๋ฅผ ๋„ฃ์Œ์œผ๋กœ์จ ์–ธ์–ด ๋ชจ๋ธ์ด ์ƒ๊ฐ์ด ์–ธ์ œ ์œ ์šฉํ• ์ง€ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๋ฉด, ์˜ˆ์ธก๋œ mixing weight๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ ์ค‘์— ์ปดํ“จํŒ…์„ ๋™์ ์œผ๋กœ ํ• ๋‹นํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ๊ณผ ์ธ๊ฐ„๊ณผ ๊ฐ™์€ ์ถ”๋ก  ๊ธฐ๋Šฅ ๊ฐ„์˜ ๊ฒฉ์ฐจ๋ฅผ ๋”์šฑ ์ขํž ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Quiet-STaR์€ ์ผ๋ฐ˜์ ์ด๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ฐฉ์‹์œผ๋กœ ์ถ”๋ก ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ํ–ฅํ•œ ํ•œ ๊ฑธ์Œ์ž…๋‹ˆ๋‹ค. ํŠน์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ข๊ฒŒ ํŠนํ™”๋˜์ง€ ์•Š๊ณ  ๋‹ค์–‘ํ•œ ์›น ํ…์ŠคํŠธ์— ๋‚ดํฌ๋œ ํ’๋ถ€ํ•œ ์ถ”๋ก  ๊ณผ์ œ์— ๋Œ€ํ•ด ํ•™์Šตํ•จ์œผ๋กœ์จ Quiet-STaR์€ ๋ณด๋‹ค ๊ฐ•๋ ฅํ•˜๊ณ  ์ ์‘๋ ฅ ์žˆ๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ํ–ฅํ•œ ๊ธธ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ๊ฒฐ๊ณผ๋Š” ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ๊ฐ€๋Šฅ์„ฑ์„ ์ž…์ฆํ•˜๋ฉฐ, Quiet-STaR์€ ์–‘์งˆ์˜ ์˜๋ฏธ ์žˆ๋Š” ๊ทผ๊ฑฐ๋ฅผ ์ƒ์„ฑํ•˜๋ฉด์„œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ถ”๋ก  ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ–ฅํ›„ ํ† ํฐ์— ๋Œ€ํ•œ ์˜ˆ์ธก์„ ๋”์šฑ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ƒ๊ฐ์„ ์•™์ƒ๋ธ”ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์˜ˆ์ธก ์ „์— ํ˜ผํ•ฉ ํ—ค๋“œ๋ฅผ ๋„ฃ์Œ์œผ๋กœ์จ ์–ธ์–ด ๋ชจ๋ธ์ด ์ƒ๊ฐ์ด ์–ธ์ œ ์œ ์šฉํ• ์ง€ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๋ฉด, ์˜ˆ์ธก๋œ ํ˜ผํ•ฉ ๊ฐ€์ค‘์น˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ ์ค‘์— ์ปดํ“จํŒ…์„ ๋™์ ์œผ๋กœ ํ• ๋‹นํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ๊ณผ ์ธ๊ฐ„๊ณผ ๊ฐ™์€ ์ถ”๋ก  ๊ธฐ๋Šฅ ๊ฐ„์˜ ๊ฒฉ์ฐจ๋ฅผ ๋”์šฑ ์ขํž ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
* 1 Introduction * 1.1 Overview of the MLCommons AI Safety Working Group * 1.2 The AI Safety Benchmark * 1.3 Infrastructure of the v0.5 benchmark * 1.4 Release of the v0.5 benchmark * 2 Scope and specification of the AI Safety Benchmark * 2.1 Systems Under Test (SUTs) * 2.2 Use cases * 2.3 Personas * 2.4 Unsafe responses and false refusal * 2.5 Test specification * 3 Taxonomy of hazard categories * 3.1 Overview of the taxonomy * 4 Test items * 4.1 Conceptualisation and design of the test items * 4.2 Creating the test items with sentence fragments and interaction types * 4.3 Dataset overview * 5 Grading SUTs * 5.1 Requirements of the grading system for the AI Safety Benchmark * 5.2 From test items to a benchmark * 6 Results * 6.1 Selection and implementation of SUTs * 6.2 Performance on the benchmark * 7 Limitations * 8 Previous work on AI safety * 8.1 AI safety * 8.2 Challenges in AI safety evaluation * 8.3 Techniques for AI safety evaluation * 8.4 Benchmarks for AI safety evaluation
* 1 ์†Œ๊ฐœ * 1.1 MLCommons AI ์•ˆ์ „ ์ž‘์—… ๊ทธ๋ฃน ๊ฐœ์š” * 1.2 AI ์•ˆ์ „ Benchmark * 1.3 v0.5 Benchmark์˜ ์ธํ”„๋ผ * 1.4 v0.5 Benchmark ์ถœ์‹œ * 2 AI ์•ˆ์ „ Benchmark์˜ ๋ฒ”์œ„ ๋ฐ ์‚ฌ์–‘ * 2.1 ํ…Œ์ŠคํŠธ ๋Œ€์ƒ ์‹œ์Šคํ…œ(SUT) * 2.2 ์‚ฌ์šฉ ์‚ฌ๋ก€ * 2.3 ํŽ˜๋ฅด์†Œ๋‚˜ * 2.4 ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ์‘๋‹ต ๋ฐ ๊ฑฐ์ง“ ๊ฑฐ๋ถ€ * 2.5 ํ…Œ์ŠคํŠธ ์‚ฌ์–‘ * 3 ์œ„ํ—˜ ๋ฒ”์ฃผ ๋ถ„๋ฅ˜๋ฒ• * 3.1 ๋ถ„๋ฅ˜๋ฒ• ๊ฐœ์š” * 4 ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ * 4.1 ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์˜ ๊ฐœ๋…ํ™” ๋ฐ ์„ค๊ณ„ * 4.2 ๋ฌธ์žฅ ์กฐ๊ฐ๊ณผ ์ƒํ˜ธ ์ž‘์šฉ ์œ ํ˜•์œผ๋กœ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ ๋งŒ๋“ค๊ธฐ * 4.3 ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐœ์š” * 5 SUT ๋“ฑ๊ธ‰ ๋ถ€์—ฌ * 5.1 AI ์•ˆ์ „ Benchmark๋ฅผ ์œ„ํ•œ ์ฑ„์  ์‹œ์Šคํ…œ์˜ ์š”๊ตฌ ์‚ฌํ•ญ * 5.2 ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์—์„œ Benchmark๋กœ * 6 ๊ฒฐ๊ณผ * 6.1 SUT ์„ ํƒ ๋ฐ ๊ตฌํ˜„ * 6.2 Benchmark์—์„œ์˜ ์„ฑ๋Šฅ * 7 ํ•œ๊ณ„ * 8 AI ์•ˆ์ „์— ๋Œ€ํ•œ ์ด์ „ ์ž‘์—… * 8.1 AI ์•ˆ์ „ * 8.2 AI ์•ˆ์ „ ํ‰๊ฐ€์˜ ๊ณผ์ œ * 8.3 AI ์•ˆ์ „ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ๊ธฐ๋ฒ• * 8.4 AI ์•ˆ์ „ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ Benchmark
* 1 ์†Œ๊ฐœ * 1.1 MLCommons AI ์•ˆ์ „ ์ž‘์—… ๊ทธ๋ฃน ๊ฐœ์š” * 1.2 AI ์•ˆ์ „ ๋ฒค์น˜๋งˆํฌ * 1.3 v0.5 ๋ฒค์น˜๋งˆํฌ์˜ ์ธํ”„๋ผ * 1.4 v0.5 ๋ฒค์น˜๋งˆํฌ ์ถœ์‹œ * 2 AI ์•ˆ์ „ ๋ฒค์น˜๋งˆํฌ์˜ ๋ฒ”์œ„ ๋ฐ ์‚ฌ์–‘ * 2.1 ํ…Œ์ŠคํŠธ ๋Œ€์ƒ ์‹œ์Šคํ…œ(SUT) * 2.2 ์‚ฌ์šฉ ์‚ฌ๋ก€ * 2.3 ํŽ˜๋ฅด์†Œ๋‚˜ * 2.4 ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ์‘๋‹ต ๋ฐ ๊ฑฐ์ง“ ๊ฑฐ๋ถ€ * 2.5 ํ…Œ์ŠคํŠธ ์‚ฌ์–‘ * 3 ์œ„ํ—˜ ๋ฒ”์ฃผ ๋ถ„๋ฅ˜๋ฒ• * 3.1 ๋ถ„๋ฅ˜๋ฒ• ๊ฐœ์š” * 4 ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ * 4.1 ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์˜ ๊ฐœ๋…ํ™” ๋ฐ ์„ค๊ณ„ * 4.2 ๋ฌธ์žฅ ์กฐ๊ฐ๊ณผ ์ƒํ˜ธ ์ž‘์šฉ ์œ ํ˜•์œผ๋กœ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ ๋งŒ๋“ค๊ธฐ * 4.3 ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐœ์š” * 5 SUT ๋“ฑ๊ธ‰ ๋ถ€์—ฌ * 5.1 AI ์•ˆ์ „ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์œ„ํ•œ ์ฑ„์  ์‹œ์Šคํ…œ์˜ ์š”๊ตฌ ์‚ฌํ•ญ * 5.2 ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์—์„œ ๋ฒค์น˜๋งˆํฌ๋กœ * 6 ๊ฒฐ๊ณผ * 6.1 SUT ์„ ํƒ ๋ฐ ๊ตฌํ˜„ * 6.2 ๋ฒค์น˜๋งˆํฌ์—์„œ์˜ ์„ฑ๋Šฅ * 7 ํ•œ๊ณ„ * 8 AI ์•ˆ์ „์— ๋Œ€ํ•œ ์ด์ „ ์ž‘์—… * 8.1 AI ์•ˆ์ „ * 8.2 AI ์•ˆ์ „ ํ‰๊ฐ€์˜ ๊ณผ์ œ * 8.3 AI ์•ˆ์ „ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ๊ธฐ๋ฒ• * 8.4 AI ์•ˆ์ „ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ
Table 1 presents a detailed analysis of the RAG studies discussed in this paper. The analysis shows that the majority of these studies have utilized external data sources to enrich the content of LLMs. A preference for multiple-hop over single-hop retrieval was noted, indicating that iterative search rounds generally yield superior results. In other words, most methods employ dense retrieval to secure higher quality candidate documents. Compared to modifying datasets in the pre-retrieval stage, more studies focus on manipulating the query to improve retrieval performance. Additionally, there is a significant emphasis on optimizing the retrieval phase, highlighting its crucial role in the research. However, there seems to be a scarcity of studies concentrating on customization in the generation stage, pointing to this as a potential area for future exploration. Overall, while the goal of RAG is to enhance the response quality of LLMs, greater efforts have been directed towards improving retrieval aspects.
ํ‘œ 1์€ ๋ณธ ๋…ผ๋ฌธ์—์„œ ๋…ผ์˜ํ•œ Retrieval-augmented generation (RAG) ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋ถ„์„์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ถ„์„ ๊ฒฐ๊ณผ, ๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ์—์„œ ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ํ™œ์šฉํ•˜์—ฌ Large Language Models (LLMs)์˜ ์ฝ˜ํ…์ธ ๋ฅผ ๋ณด๊ฐ•ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋‹จ์ผ ํ™‰ ๊ฒ€์ƒ‰๋ณด๋‹ค ๋‹ค์ค‘ ํ™‰ ๊ฒ€์ƒ‰์„ ์„ ํ˜ธํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ๋Š”๋ฐ, ์ด๋Š” ๋ฐ˜๋ณต์ ์ธ ๊ฒ€์ƒ‰ ๋ผ์šด๋“œ๊ฐ€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋” ์šฐ์ˆ˜ํ•œ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ฆ‰, ๋Œ€๋ถ€๋ถ„์˜ ๋ฐฉ๋ฒ•์€ ๊ณ ํ’ˆ์งˆ ํ›„๋ณด ๋ฌธ์„œ๋ฅผ ํ™•๋ณดํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ๋ฐ€๋„ ๊ฒ€์ƒ‰์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰ ์ „ ๋‹จ๊ณ„์—์„œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜์ •ํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด ๊ฒ€์ƒ‰ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ฟผ๋ฆฌ๋ฅผ ์กฐ์ž‘ํ•˜๋Š” ๋ฐ ๋” ๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ์ง‘์ค‘๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฒ€์ƒ‰ ๋‹จ๊ณ„๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์–ด ์—ฐ๊ตฌ์—์„œ ๊ฒ€์ƒ‰ ๋‹จ๊ณ„์˜ ์ค‘์š”ํ•œ ์—ญํ• ์„ ๊ฐ•์กฐํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ƒ์„ฑ ๋‹จ๊ณ„์—์„œ ์‚ฌ์šฉ์ž ์ง€์ •์— ์ง‘์ค‘ํ•˜๋Š” ์—ฐ๊ตฌ๋Š” ๊ฑฐ์˜ ์—†๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ด๋ฉฐ, ์ด๋ฅผ ํ–ฅํ›„ ์—ฐ๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ์  ์˜์—ญ์œผ๋กœ ์ง€์ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ RAG์˜ ๋ชฉํ‘œ๋Š” LLM์˜ ์‘๋‹ต ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์ด์ง€๋งŒ, ๊ฒ€์ƒ‰ ์ธก๋ฉด์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ๋” ๋งŽ์€ ๋…ธ๋ ฅ์ด ๊ธฐ์šธ์—ฌ์ ธ ์™”์Šต๋‹ˆ๋‹ค.
ํ‘œ 1์€ ๋ณธ ๋…ผ๋ฌธ์—์„œ ๋…ผ์˜ํ•œ RAG ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋ถ„์„์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ถ„์„ ๊ฒฐ๊ณผ, ๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ์—์„œ ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์˜ ์ฝ˜ํ…์ธ ๋ฅผ ๋ณด๊ฐ•ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋‹จ์ผ ํ™‰ ๊ฒ€์ƒ‰๋ณด๋‹ค ๋‹ค์ค‘ ํ™‰ ๊ฒ€์ƒ‰์„ ์„ ํ˜ธํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ๋Š”๋ฐ, ์ด๋Š” ๋ฐ˜๋ณต์ ์ธ ๊ฒ€์ƒ‰ ๋ผ์šด๋“œ๊ฐ€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋” ์šฐ์ˆ˜ํ•œ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ฆ‰, ๋Œ€๋ถ€๋ถ„์˜ ๋ฐฉ๋ฒ•์€ ๊ณ ํ’ˆ์งˆ ํ›„๋ณด ๋ฌธ์„œ๋ฅผ ํ™•๋ณดํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ๋ฐ€๋„ ๊ฒ€์ƒ‰์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰ ์ „ ๋‹จ๊ณ„์—์„œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜์ •ํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด ๊ฒ€์ƒ‰ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ฟผ๋ฆฌ๋ฅผ ์กฐ์ž‘ํ•˜๋Š” ๋ฐ ๋” ๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ์ง‘์ค‘๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฒ€์ƒ‰ ๋‹จ๊ณ„๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์–ด ์—ฐ๊ตฌ์—์„œ ๊ฒ€์ƒ‰ ๋‹จ๊ณ„์˜ ์ค‘์š”ํ•œ ์—ญํ• ์„ ๊ฐ•์กฐํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ƒ์„ฑ ๋‹จ๊ณ„์—์„œ ์‚ฌ์šฉ์ž ์ง€์ •์— ์ง‘์ค‘ํ•˜๋Š” ์—ฐ๊ตฌ๋Š” ๊ฑฐ์˜ ์—†๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ด๋ฉฐ, ์ด๋ฅผ ํ–ฅํ›„ ์—ฐ๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ์  ์˜์—ญ์œผ๋กœ ์ง€์ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ RAG์˜ ๋ชฉํ‘œ๋Š” LLM์˜ ์‘๋‹ต ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์ด์ง€๋งŒ, ๊ฒ€์ƒ‰ ์ธก๋ฉด์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ๋” ๋งŽ์€ ๋…ธ๋ ฅ์ด ๊ธฐ์šธ์—ฌ์ ธ ์™”์Šต๋‹ˆ๋‹ค.
Task: Extract all the country names in the paragraph, list them separated by commas. Example 1 Paragraph: Dr. No is the sixth novel by the English author Ian Fleming to feature his British Secret Service agent James Bond. Written at Fleming's Goldeneye estate in Jamaica, it was first published in the United Kingdom by Jonathan Cape in 1958. In the novel Bond looks into the disappearance in Jamaica of two fellow MI6 operatives who had been investigating Doctor No. Bond travels to No's Caribbean island and meets Honeychile Rider, who is there to collect shells. They are captured and taken to a luxurious facility carved into a mountain. The character of Doctor No, the son of a German missionary and a Chinese woman, was influenced by Sax Rohner's Fu Manchu stories. Dr. No was the first of Fleming's novels to face widespread negative reviews in Britain, but it was received more favourably in the United States. Output: English, British, Jamaica, the United Kingdom, German, Chinese, Britain, the United States.
๊ณผ์ œ: ๋‹จ๋ฝ์— ์žˆ๋Š” ๋ชจ๋“  ๊ตญ๊ฐ€ ์ด๋ฆ„์„ ์ถ”์ถœํ•˜์—ฌ ์‰ผํ‘œ๋กœ ๊ตฌ๋ถ„ํ•˜์—ฌ ๋‚˜์—ดํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ์‹œ 1 ๋‹จ๋ฝ: ๋‹ฅํ„ฐ ๋…ธ๋Š” ์˜๊ตญ ์ž‘๊ฐ€ ์ด์•ˆ ํ”Œ๋ ˆ๋ฐ์˜ ์—ฌ์„ฏ ๋ฒˆ์งธ ์†Œ์„ค๋กœ, ์˜๊ตญ ๋น„๋ฐ€ ์š”์› ์ œ์ž„์Šค ๋ณธ๋“œ๊ฐ€ ๋“ฑ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ž๋ฉ”์ด์นด์˜ ๊ณจ๋“ ์•„์ด ์ €ํƒ์—์„œ ์“ฐ์—ฌ์ง„ ์ด ์†Œ์„ค์€ 1958๋…„ ์˜๊ตญ์—์„œ ์กฐ๋‚˜๋‹จ ์ผ€์ดํ”„์— ์˜ํ•ด ์ฒ˜์Œ ์ถœํŒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์†Œ์„ค์—์„œ ๋ณธ๋“œ๋Š” ์ž๋ฉ”์ด์นด์—์„œ '๋‹ฅํ„ฐ ๋…ธ'๋ฅผ ์กฐ์‚ฌํ•˜๋˜ ๋™๋ฃŒ MI6 ์š”์› ๋‘ ๋ช…์ด ์‚ฌ๋ผ์ง„ ์‚ฌ๊ฑด์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ณธ๋“œ๋Š” ๋…ธ์˜ ์นด๋ฆฌ๋ธŒํ•ด ์„ฌ์œผ๋กœ ๊ฐ€์„œ ์กฐ๊ฐœ๋ฅผ ์ฑ„์ทจํ•˜๋Ÿฌ ์˜จ ํ—ˆ๋‹ˆ์น ๋ผ์ด๋”๋ฅผ ๋งŒ๋‚ฉ๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ์ฒดํฌ๋˜์–ด ์‚ฐ์†์— ๊นŽ์•„ ๋งŒ๋“  ํ˜ธํ™”๋กœ์šด ์‹œ์„ค๋กœ ๋Œ๋ ค๊ฐ‘๋‹ˆ๋‹ค. ๋…์ผ ์„ ๊ต์‚ฌ์™€ ์ค‘๊ตญ ์—ฌ์„ฑ ์‚ฌ์ด์—์„œ ํƒœ์–ด๋‚œ ๋‹ฅํ„ฐ ๋…ธ์˜ ์บ๋ฆญํ„ฐ๋Š” ์ƒ‰์Šค ๋กœ๋„ˆ์˜ ํ‘ธ๋งŒ์ถ” ์ด์•ผ๊ธฐ์˜ ์˜ํ–ฅ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋‹ฅํ„ฐ ๋…ธ๋Š” ํ”Œ๋ ˆ๋ฐ์˜ ์†Œ์„ค ์ค‘ ์˜๊ตญ์—์„œ ์ฒ˜์Œ์œผ๋กœ ๊ด‘๋ฒ”์œ„ํ•œ ๋ถ€์ •์ ์ธ ํ‰๊ฐ€์— ์ง๋ฉดํ–ˆ์ง€๋งŒ ๋ฏธ๊ตญ์—์„œ๋Š” ๋” ํ˜ธ์˜์ ์ธ ํ‰๊ฐ€๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ: English, British, Jamaica, the United Kingdom, German, Chinese, Britain, the United States.
๊ณผ์ œ: ๋‹จ๋ฝ์— ์žˆ๋Š” ๋ชจ๋“  ๊ตญ๊ฐ€ ์ด๋ฆ„์„ ์ถ”์ถœํ•˜์—ฌ ์‰ผํ‘œ๋กœ ๊ตฌ๋ถ„ํ•˜์—ฌ ๋‚˜์—ดํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ์‹œ 1 ๋‹จ๋ฝ: ๋‹ฅํ„ฐ ๋…ธ๋Š” ์˜๊ตญ ์ž‘๊ฐ€ ์ด์•ˆ ํ”Œ๋ ˆ๋ฐ์˜ ์—ฌ์„ฏ ๋ฒˆ์งธ ์†Œ์„ค๋กœ, ์˜๊ตญ ๋น„๋ฐ€ ์š”์› ์ œ์ž„์Šค ๋ณธ๋“œ๊ฐ€ ๋“ฑ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ž๋ฉ”์ด์นด์˜ ๊ณจ๋“ ์•„์ด ์ €ํƒ์—์„œ ์“ฐ์—ฌ์ง„ ์ด ์†Œ์„ค์€ 1958๋…„ ์˜๊ตญ์—์„œ ์กฐ๋‚˜๋‹จ ์ผ€์ดํ”„์— ์˜ํ•ด ์ฒ˜์Œ ์ถœํŒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์†Œ์„ค์—์„œ ๋ณธ๋“œ๋Š” ์ž๋ฉ”์ด์นด์—์„œ '๋‹ฅํ„ฐ ๋…ธ'๋ฅผ ์กฐ์‚ฌํ•˜๋˜ ๋™๋ฃŒ MI6 ์š”์› ๋‘ ๋ช…์ด ์‚ฌ๋ผ์ง„ ์‚ฌ๊ฑด์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ณธ๋“œ๋Š” ๋…ธ์˜ ์นด๋ฆฌ๋ธŒํ•ด ์„ฌ์œผ๋กœ ๊ฐ€์„œ ์กฐ๊ฐœ๋ฅผ ์ฑ„์ทจํ•˜๋Ÿฌ ์˜จ ํ—ˆ๋‹ˆ์น ๋ผ์ด๋”๋ฅผ ๋งŒ๋‚ฉ๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ์ฒดํฌ๋˜์–ด ์‚ฐ์†์— ๊นŽ์•„ ๋งŒ๋“  ํ˜ธํ™”๋กœ์šด ์‹œ์„ค๋กœ ๋Œ๋ ค๊ฐ‘๋‹ˆ๋‹ค. ๋…์ผ ์„ ๊ต์‚ฌ์™€ ์ค‘๊ตญ ์—ฌ์„ฑ ์‚ฌ์ด์—์„œ ํƒœ์–ด๋‚œ ๋‹ฅํ„ฐ ๋…ธ์˜ ์บ๋ฆญํ„ฐ๋Š” ์ƒ‰์Šค ๋กœ๋„ˆ์˜ ํ‘ธ๋งŒ์ถ” ์ด์•ผ๊ธฐ์˜ ์˜ํ–ฅ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋‹ฅํ„ฐ ๋…ธ๋Š” ํ”Œ๋ ˆ๋ฐ์˜ ์†Œ์„ค ์ค‘ ์˜๊ตญ์—์„œ ์ฒ˜์Œ์œผ๋กœ ๊ด‘๋ฒ”์œ„ํ•œ ๋ถ€์ •์ ์ธ ํ‰๊ฐ€์— ์ง๋ฉดํ–ˆ์ง€๋งŒ ๋ฏธ๊ตญ์—์„œ๋Š” ๋” ํ˜ธ์˜์ ์ธ ํ‰๊ฐ€๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ: ์˜์–ด, ์˜๊ตญ, ์ž๋ฉ”์ด์นด, ์˜๊ตญ, ๋…์ผ, ์ค‘๊ตญ, ์˜๊ตญ, ๋ฏธ๊ตญ.
Research by Khandelwal et al. (Khandelwal et al., 2020) demonstrates that accessing relevant information from the training dataset itself can significantly improve LLM performance, highlighting the effectiveness of RAG. Over time, RAG has evolved from a means of providing supplementary information to enabling multiple interactions between the retrieval and generation components. This involves conducting several rounds of retrieval to refine the accuracy of the information retrieved and iteratively improve the quality of the generated output. Platforms such as LangChain1 and Llamalndex2 have modularized the RAG approach, enhancing its adaptability and expanding its range of applications. Despite these platforms employing diverse methodologies to tackle different aspects of RAG--from multiple search iterations to iterative generation--they maintain adherence to the fundamental RAG workflow. This consistency is crucial for understanding their operation and pinpointing opportunities for further development.
Khandelwal ๋“ฑ์˜ ์—ฐ๊ตฌ(Khandelwal et al., 2020)์— ๋”ฐ๋ฅด๋ฉด, ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ž์ฒด์—์„œ ๊ด€๋ จ ์ •๋ณด์— ์•ก์„ธ์Šคํ•˜๋ฉด LLM ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” Retrieval-augmented generation (RAG)์˜ ํšจ๊ณผ๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ RAG๋Š” ๋ณด์กฐ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋Š” ์ˆ˜๋‹จ์—์„œ ๊ฒ€์ƒ‰ ๋ฐ ์ƒ์„ฑ ๊ตฌ์„ฑ ์š”์†Œ ๊ฐ„์˜ ์—ฌ๋Ÿฌ ์ƒํ˜ธ ์ž‘์šฉ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ฐœ์ „ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๊ฒ€์ƒ‰๋œ ์ •๋ณด์˜ ์ •ํ™•์„ฑ์„ ๊ฐœ์„ ํ•˜๊ณ  ์ƒ์„ฑ๋œ ์ถœ๋ ฅ์˜ ํ’ˆ์งˆ์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์ฐจ๋ก€ ๊ฒ€์ƒ‰์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. LangChain1 ๋ฐ Llamalndex2์™€ ๊ฐ™์€ ํ”Œ๋žซํผ์€ RAG ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ชจ๋“ˆํ™”ํ•˜์—ฌ ์ ์‘๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ์ ์šฉ ๋ฒ”์œ„๋ฅผ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ”Œ๋žซํผ์€ ๋‹ค์–‘ํ•œ ๊ฒ€์ƒ‰ ๋ฐ˜๋ณต๋ถ€ํ„ฐ ๋ฐ˜๋ณต์ ์ธ ์ƒ์„ฑ์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ RAG์˜ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•๋ก ์„ ์‚ฌ์šฉํ•˜์ง€๋งŒ, ๊ธฐ๋ณธ์ ์ธ RAG ์›Œํฌํ”Œ๋กœ์šฐ๋ฅผ ๊ณ ์ˆ˜ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ผ๊ด€์„ฑ์€ ์šด์˜์„ ์ดํ•ดํ•˜๊ณ  ์ถ”๊ฐ€ ๊ฐœ๋ฐœ ๊ธฐํšŒ๋ฅผ ์ •ํ™•ํžˆ ํŒŒ์•…ํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
Khandelwal ๋“ฑ์˜ ์—ฐ๊ตฌ(Khandelwal et al., 2020)์— ๋”ฐ๋ฅด๋ฉด, ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ž์ฒด์—์„œ ๊ด€๋ จ ์ •๋ณด์— ์•ก์„ธ์Šคํ•˜๋ฉด LLM ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” RAG์˜ ํšจ๊ณผ๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ RAG๋Š” ๋ณด์กฐ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋Š” ์ˆ˜๋‹จ์—์„œ ๊ฒ€์ƒ‰ ๋ฐ ์ƒ์„ฑ ๊ตฌ์„ฑ ์š”์†Œ ๊ฐ„์˜ ์—ฌ๋Ÿฌ ์ƒํ˜ธ ์ž‘์šฉ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ฐœ์ „ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๊ฒ€์ƒ‰๋œ ์ •๋ณด์˜ ์ •ํ™•์„ฑ์„ ๊ฐœ์„ ํ•˜๊ณ  ์ƒ์„ฑ๋œ ์ถœ๋ ฅ์˜ ํ’ˆ์งˆ์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์ฐจ๋ก€ ๊ฒ€์ƒ‰์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. LangChain1 ๋ฐ Llamalndex2์™€ ๊ฐ™์€ ํ”Œ๋žซํผ์€ RAG ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ชจ๋“ˆํ™”ํ•˜์—ฌ ์ ์‘๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ์ ์šฉ ๋ฒ”์œ„๋ฅผ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ”Œ๋žซํผ์€ ๋‹ค์–‘ํ•œ ๊ฒ€์ƒ‰ ๋ฐ˜๋ณต๋ถ€ํ„ฐ ๋ฐ˜๋ณต์ ์ธ ์ƒ์„ฑ์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ RAG์˜ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•๋ก ์„ ์‚ฌ์šฉํ•˜์ง€๋งŒ, ๊ธฐ๋ณธ์ ์ธ RAG ์›Œํฌํ”Œ๋กœ์šฐ๋ฅผ ๊ณ ์ˆ˜ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ผ๊ด€์„ฑ์€ ์šด์˜์„ ์ดํ•ดํ•˜๊ณ  ์ถ”๊ฐ€ ๊ฐœ๋ฐœ ๊ธฐํšŒ๋ฅผ ์ •ํ™•ํžˆ ํŒŒ์•…ํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
To support the v0.5 benchmark, MLCommons has developed an open-source evaluation tool, which consists of the ModelBench benchmark runner (which can be used to implement the benchmark) and the ModelGauge test execution engine (which contains the actual test items). This tool enables standardized, reproducible benchmark runs using versioned tests and SUTs. The tool is designed with a modular plug-in architecture, allowing model providers to easily implement and add new SUTs to the platform for evaluation. As the AI Safety Benchmark evolves, new versions of tests will be added to the platform. Details on how to access and use the platform can be found in the ModelBench Git repository on GitHub.6 ModelBench and ModelGauge were developed in collaboration with the Holistic Evaluation of Language Models [12] team at the Stanford Center for Research on Foundation Models (CRFM), and build upon the HELM team's experience of creating a widely-adopted open-source model evaluation framework for living leaderboards.
MLCommons๋Š” v0.5 benchmark๋ฅผ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด benchmark ์‹คํ–‰์ž(benchmark๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š”)์ธ ModelBench benchmark runner์™€ ์‹ค์ œ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์ด ํฌํ•จ๋œ ํ…Œ์ŠคํŠธ ์‹คํ–‰ ์—”์ง„์ธ ModelGauge๋กœ ๊ตฌ์„ฑ๋œ ์˜คํ”ˆ ์†Œ์Šค ํ‰๊ฐ€ ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋ฒ„์ „์ด ์ง€์ •๋œ ํ…Œ์ŠคํŠธ์™€ SUT๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‘œ์ค€ํ™”๋˜๊ณ  ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ benchmark ์‹คํ–‰์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋„๊ตฌ๋Š” ๋ชจ๋“ˆ์‹ ํ”Œ๋Ÿฌ๊ทธ์ธ ์•„ํ‚คํ…์ฒ˜๋กœ ์„ค๊ณ„๋˜์–ด ๋ชจ๋ธ ์ œ๊ณต์—…์ฒด๊ฐ€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ƒˆ๋กœ์šด SUT๋ฅผ ํ”Œ๋žซํผ์— ์‰ฝ๊ฒŒ ๊ตฌํ˜„ํ•˜๊ณ  ์ถ”๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. AI Safety Benchmark๊ฐ€ ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ ํ”Œ๋žซํผ์— ์ƒˆ๋กœ์šด ๋ฒ„์ „์˜ ํ…Œ์ŠคํŠธ๊ฐ€ ์ถ”๊ฐ€๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ”Œ๋žซํผ์— ์•ก์„ธ์Šคํ•˜๊ณ  ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ GitHub์˜ ModelBench Git ๋ฆฌํฌ์ง€ํ† ๋ฆฌ์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ModelBench์™€ ModelGauge๋Š” Stanford Center for Research on Foundation Models (CRFM)์˜ Holistic Evaluation of Language Models [12] ํŒ€๊ณผ ๊ณต๋™์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์œผ๋ฉฐ, ๋„๋ฆฌ ์ฑ„ํƒ๋œ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋งŒ๋“ค์–ด์˜จ HELM ํŒ€์˜ ๋ฆฌ๋น™ ๋ฆฌ๋”๋ณด๋“œ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์„ ๋ฐ”ํƒ•์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
MLCommons๋Š” v0.5 ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด ๋ฒค์น˜๋งˆํฌ ์‹คํ–‰์ž(๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š”)์ธ ModelBench ๋ฒค์น˜๋งˆํฌ ๋Ÿฐ๋„ˆ์™€ ์‹ค์ œ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์ด ํฌํ•จ๋œ ํ…Œ์ŠคํŠธ ์‹คํ–‰ ์—”์ง„์ธ ModelGauge๋กœ ๊ตฌ์„ฑ๋œ ์˜คํ”ˆ ์†Œ์Šค ํ‰๊ฐ€ ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋ฒ„์ „์ด ์ง€์ •๋œ ํ…Œ์ŠคํŠธ์™€ SUT๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‘œ์ค€ํ™”๋˜๊ณ  ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ ๋ฒค์น˜๋งˆํฌ ์‹คํ–‰์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋„๊ตฌ๋Š” ๋ชจ๋“ˆ์‹ ํ”Œ๋Ÿฌ๊ทธ์ธ ์•„ํ‚คํ…์ฒ˜๋กœ ์„ค๊ณ„๋˜์–ด ๋ชจ๋ธ ์ œ๊ณต์—…์ฒด๊ฐ€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ƒˆ๋กœ์šด SUT๋ฅผ ํ”Œ๋žซํผ์— ์‰ฝ๊ฒŒ ๊ตฌํ˜„ํ•˜๊ณ  ์ถ”๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. AI ์•ˆ์ „ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ ํ”Œ๋žซํผ์— ์ƒˆ๋กœ์šด ๋ฒ„์ „์˜ ํ…Œ์ŠคํŠธ๊ฐ€ ์ถ”๊ฐ€๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ”Œ๋žซํผ์— ์•ก์„ธ์Šคํ•˜๊ณ  ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ GitHub์˜ ModelBench Git ๋ฆฌํฌ์ง€ํ† ๋ฆฌ์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.6 ModelBench์™€ ModelGauge๋Š” ์Šคํƒ ํฌ๋“œ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ ์—ฐ๊ตฌ ์„ผํ„ฐ(CRFM)์˜ HOLISTIC ์–ธ์–ด ๋ชจ๋ธ ํ‰๊ฐ€[12] ํŒ€๊ณผ ๊ณต๋™์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์œผ๋ฉฐ, ๋„๋ฆฌ ์ฑ„ํƒ๋œ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋งŒ๋“ค์–ด์˜จ HELM ํŒ€์˜ ๋ฆฌ๋น™ ๋ฆฌ๋”๋ณด๋“œ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์„ ๋ฐ”ํƒ•์œผ๋กœ ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
A multitude of previous tasks have employed structured data in the form of knowledge graphs, with knowledge graph completion representing a pervasive application. The conventional methodology (Chen et al., 2022) (Saxena et al., 2022) for completion entails defining the task as a sequence-to-sequence process, wherein incomplete triples and entities are transformed into text sequences. However, this approach is constrained by its reliance on implicit reasoning, which significantly constrains the utility of the knowledge graph itself. ReSKGC (Yu and Yang, 2023) proposes the integration of retrieval augmentation techniques with knowledge graph completion. This integration entails the selection of semantically relevant triples from the knowledge graph and their utilization as evidence to inform the generation of output through explicit reasoning. This model employs data from the Wikidata5M and WikiKG90Mv2 datasets, demonstrating superior performance compared to other existing work in a range of conditions.
์ง€์‹ ๊ทธ๋ž˜ํ”„์˜ ํ˜•ํƒœ๋กœ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์ด์ „ ์ž‘์—…์€ ๋งŽ์•˜์œผ๋ฉฐ, ์ง€์‹ ๊ทธ๋ž˜ํ”„ ์™„์„ฑ์ด ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์ž…๋‹ˆ๋‹ค. ์™„์„ฑ์„ ์œ„ํ•œ ๊ธฐ์กด ๋ฐฉ๋ฒ•๋ก (Chen et al., 2022) (Saxena et al., 2022)์€ ๋ถˆ์™„์ „ํ•œ ํŠธ๋ฆฌํ”Œ๊ณผ ์—”ํ‹ฐํ‹ฐ๋ฅผ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋กœ ๋ณ€ํ™˜ํ•˜๋Š” sequence-to-sequence ํ”„๋กœ์„ธ์Šค๋กœ ์ž‘์—…์„ ์ •์˜ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ง€์‹ ๊ทธ๋ž˜ํ”„ ์ž์ฒด์˜ ์œ ์šฉ์„ฑ์„ ํฌ๊ฒŒ ์ œ์•ฝํ•˜๋Š” ์•”๋ฌต์  ์ถ”๋ก ์— ์˜์กดํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ReSKGC(Yu and Yang, 2023)๋Š” ์ง€์‹ ๊ทธ๋ž˜ํ”„ ์™„์„ฑ์— Retrieval augmentation ๊ธฐ์ˆ ์„ ํ†ตํ•ฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ํ†ตํ•ฉ์—๋Š” ์ง€์‹ ๊ทธ๋ž˜ํ”„์—์„œ ์˜๋ฏธ์ ์œผ๋กœ ๊ด€๋ จ๋œ ํŠธ๋ฆฌํ”Œ์„ ์„ ํƒํ•˜๊ณ  ๋ช…์‹œ์  ์ถ”๋ก ์„ ํ†ตํ•ด ์ถœ๋ ฅ ์ƒ์„ฑ์— ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋Š” ์ฆ๊ฑฐ๋กœ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ Wikidata5M ๋ฐ WikiKG90Mv2 ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ๋‹ค์–‘ํ•œ ์กฐ๊ฑด์—์„œ ๊ธฐ์กด์˜ ๋‹ค๋ฅธ ์ž‘์—…์— ๋น„ํ•ด ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์ง€์‹ ๊ทธ๋ž˜ํ”„์˜ ํ˜•ํƒœ๋กœ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์ด์ „ ์ž‘์—…์€ ๋งŽ์•˜์œผ๋ฉฐ, ์ง€์‹ ๊ทธ๋ž˜ํ”„ ์™„์„ฑ์ด ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์ž…๋‹ˆ๋‹ค. ์™„์„ฑ์„ ์œ„ํ•œ ๊ธฐ์กด ๋ฐฉ๋ฒ•๋ก (Chen et al., 2022) (Saxena et al., 2022)์€ ๋ถˆ์™„์ „ํ•œ ํŠธ๋ฆฌํ”Œ๊ณผ ์—”ํ‹ฐํ‹ฐ๋ฅผ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์‹œํ€€์Šค-ํˆฌ-์‹œํ€€์Šค ํ”„๋กœ์„ธ์Šค๋กœ ์ž‘์—…์„ ์ •์˜ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ง€์‹ ๊ทธ๋ž˜ํ”„ ์ž์ฒด์˜ ์œ ์šฉ์„ฑ์„ ํฌ๊ฒŒ ์ œ์•ฝํ•˜๋Š” ์•”๋ฌต์  ์ถ”๋ก ์— ์˜์กดํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ReSKGC(Yu and Yang, 2023)๋Š” ์ง€์‹ ๊ทธ๋ž˜ํ”„ ์™„์„ฑ์— ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ๊ธฐ์ˆ ์„ ํ†ตํ•ฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ํ†ตํ•ฉ์—๋Š” ์ง€์‹ ๊ทธ๋ž˜ํ”„์—์„œ ์˜๋ฏธ์ ์œผ๋กœ ๊ด€๋ จ๋œ ํŠธ๋ฆฌํ”Œ์„ ์„ ํƒํ•˜๊ณ  ๋ช…์‹œ์  ์ถ”๋ก ์„ ํ†ตํ•ด ์ถœ๋ ฅ ์ƒ์„ฑ์— ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋Š” ์ฆ๊ฑฐ๋กœ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ Wikidata5M ๋ฐ WikiKG90Mv2 ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ๋‹ค์–‘ํ•œ ์กฐ๊ฑด์—์„œ ๊ธฐ์กด์˜ ๋‹ค๋ฅธ ์ž‘์—…์— ๋น„ํ•ด ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
The Yi-34B model seems to exhibit a superior fit to the Scaling Law. This enhanced alignment with the Scaling Law suggests that Yi-34B may have architectural or algorithmic advantages that enable it to leverage data more effectively. Despite Yi-34B having fewer parameters than Deepseek-67B, it demonstrates that with an increase in data volume, Yi-34B can achieve comparable results post-training. This observation suggests that the quantity of model parameters is not the sole determinant of performance enhancement when sufficient data is available. In expanding the training datasets, Yi-34B appears to utilize its parameters more efficiently, potentially indicating a better generalization capacity or optimization that allows for superior learning from additional data. This could encourage the development of more data-centric approaches in the design of algorithms, focusing on how to better utilize the information available to enhance learning outcomes, even with a comparatively smaller set of parameters.
Yi-34B ๋ชจ๋ธ์€ Scaling Law์— ๋” ์šฐ์ˆ˜ํ•œ ์ ํ•ฉ์„ฑ์„ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. Scaling Law์™€์˜ ์ •๋ ฌ์ด ๊ฐ•ํ™”๋œ ๊ฒƒ์€ ๋ฐ์ดํ„ฐ๋ฅผ ๋” ํšจ๊ณผ์ ์œผ๋กœ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์•„ํ‚คํ…์ฒ˜ ๋˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์  ์ด์ ์ด ์žˆ์„ ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. Yi-34B๋Š” Deepseek-67B๋ณด๋‹ค ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ์ ๋”๋ผ๋„ ๋ฐ์ดํ„ฐ ๋ณผ๋ฅจ์ด ์ฆ๊ฐ€ํ•˜๋ฉด ํ›ˆ๋ จ ํ›„ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ์ถฉ๋ถ„ํ•œ ๋ฐ์ดํ„ฐ๊ฐ€ ์žˆ์„ ๋•Œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์˜ ์œ ์ผํ•œ ๊ฒฐ์ • ์š”์†Œ๊ฐ€ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ์–‘์ด ์•„๋‹ˆ๋ผ๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ™•์žฅํ•  ๋•Œ Yi-34B๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋” ํšจ์œจ์ ์œผ๋กœ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ด๋ฉฐ, ์ด๋Š” ์ž ์žฌ์ ์œผ๋กœ ๋” ๋‚˜์€ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ ๋˜๋Š” ์ตœ์ ํ™”๋ฅผ ํ†ตํ•ด ์ถ”๊ฐ€ ๋ฐ์ดํ„ฐ์—์„œ ๋” ๋‚˜์€ ํ•™์Šต์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋Š” ๋น„๊ต์  ์ ์€ ์ˆ˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ๋„ ํ•™์Šต ๊ฒฐ๊ณผ๋ฅผ ๊ฐœ์„ ํ•˜๋Š” ๋ฐฉ๋ฒ•์— ์ดˆ์ ์„ ๋งž์ถ˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์„ค๊ณ„์—์„œ ๋ณด๋‹ค ๋ฐ์ดํ„ฐ ์ค‘์‹ฌ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•˜๋„๋ก ์žฅ๋ คํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Yi-34B ๋ชจ๋ธ์€ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋” ์šฐ์ˆ˜ํ•œ ์ ํ•ฉ์„ฑ์„ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ์Šค์ผ€์ผ๋ง ๋ฒ•์น™๊ณผ์˜ ์ •๋ ฌ์ด ๊ฐ•ํ™”๋œ ๊ฒƒ์€ ๋ฐ์ดํ„ฐ๋ฅผ ๋” ํšจ๊ณผ์ ์œผ๋กœ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์•„ํ‚คํ…์ฒ˜ ๋˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์  ์ด์ ์ด ์žˆ์„ ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. Yi-34B๋Š” Deepseek-67B๋ณด๋‹ค ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ์ ๋”๋ผ๋„ ๋ฐ์ดํ„ฐ ๋ณผ๋ฅจ์ด ์ฆ๊ฐ€ํ•˜๋ฉด ํ›ˆ๋ จ ํ›„ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ์ถฉ๋ถ„ํ•œ ๋ฐ์ดํ„ฐ๊ฐ€ ์žˆ์„ ๋•Œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์˜ ์œ ์ผํ•œ ๊ฒฐ์ • ์š”์†Œ๊ฐ€ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ์–‘์ด ์•„๋‹ˆ๋ผ๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ™•์žฅํ•  ๋•Œ Yi-34B๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋” ํšจ์œจ์ ์œผ๋กœ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ด๋ฉฐ, ์ด๋Š” ์ž ์žฌ์ ์œผ๋กœ ๋” ๋‚˜์€ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ ๋˜๋Š” ์ตœ์ ํ™”๋ฅผ ํ†ตํ•ด ์ถ”๊ฐ€ ๋ฐ์ดํ„ฐ์—์„œ ๋” ๋‚˜์€ ํ•™์Šต์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋Š” ๋น„๊ต์  ์ ์€ ์ˆ˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ๋„ ํ•™์Šต ๊ฒฐ๊ณผ๋ฅผ ๊ฐœ์„ ํ•˜๋Š” ๋ฐฉ๋ฒ•์— ์ดˆ์ ์„ ๋งž์ถ˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์„ค๊ณ„์—์„œ ๋ณด๋‹ค ๋ฐ์ดํ„ฐ ์ค‘์‹ฌ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•˜๋„๋ก ์žฅ๋ คํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Optimization and EvaluationFor optimization, we use the AdamW optimizer with a warmup of 20 steps, a learning rate of \(1e-6\), a weight decay of 0.001, and a batch size of 8 (along with any necessary gradient accumulation to keep this fixed across runs). Moreover, our <!startofthought|> and <!endofthought|> embedding gradient weight is \(1e2\) and our policy weight is \(1e6\). We sample with temperature \(T=1\) during training and use greedy decoding for the thoughts during evaluation. We treat our samples as importance samples by computing the REINFORCE loss at temperature \(T=3\). Because we do not prompt the model with any examples, we directly compute the probability of the correct answer, conditioned on generating an answer - for example, for multiple choice questions between \(A\cdots E\), we compute the accuracy over the logits for tokens corresponding to \(A\cdots E\). Lastly, for our training, we select a random span of 256 tokens from each sample (or pad if there are fewer than 256 tokens).
์ตœ์ ํ™” ๋ฐ ํ‰๊ฐ€. ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ์›Œ๋ฐ์—… 20๋‹จ๊ณ„, Learning Rate \(1e-6\), ๊ฐ€์ค‘์น˜ ๊ฐ์‡  0.001, Batch Size 8(ํ•„์š”์— ๋”ฐ๋ผ ์‹คํ–‰ ์ „๋ฐ˜์— ๊ฑธ์ณ ์ด๋ฅผ ๊ณ ์ •ํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ผ๋””์–ธํŠธ ๋ˆ„์ )์˜ AdamW Optimizer๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, <!startofthought|> ๋ฐ <!endofthought|> ์ž„๋ฒ ๋”ฉ ๊ทธ๋ผ๋””์–ธํŠธ ๊ฐ€์ค‘์น˜๋Š” \(1e2\)์ด๊ณ  ์ •์ฑ… ๊ฐ€์ค‘์น˜๋Š” \(1e6\)์ž…๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์ค‘์—๋Š” ์˜จ๋„ \(T=1\)๋กœ ์ƒ˜ํ”Œ๋งํ•˜๊ณ  ํ‰๊ฐ€ ์ค‘์—๋Š” ํƒ์š•์Šค๋Ÿฌ์šด ๋””์ฝ”๋”ฉ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜จ๋„ \(T=3\)์—์„œ REINFORCE Loss๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ ์ƒ˜ํ”Œ์„ ์ค‘์š” ์ƒ˜ํ”Œ๋กœ ์ทจ๊ธ‰ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์— ์˜ˆ์ œ๋ฅผ ํ”„๋กฌํ”„ํŠธํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ ์ •๋‹ต์˜ ํ™•๋ฅ ์„ ์ง์ ‘ ๊ณ„์‚ฐํ•˜์—ฌ ์˜ˆ๋ฅผ ๋“ค์–ด, \(A\cdots E\) ์‚ฌ์ด์˜ ๊ฐ๊ด€์‹ ๋ฌธ์ œ์˜ ๊ฒฝ์šฐ, \(A\cdots E\)์— ํ•ด๋‹นํ•˜๋Š” ํ† ํฐ์— ๋Œ€ํ•œ ๋กœ์ง“์— ๋Œ€ํ•œ ์ •ํ™•๋„๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ํ›ˆ๋ จ์˜ ๊ฒฝ์šฐ ๊ฐ ์ƒ˜ํ”Œ์—์„œ 256๊ฐœ์˜ ํ† ํฐ์˜ ์ž„์˜ ๋ฒ”์œ„๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค(๋˜๋Š” ํ† ํฐ์ด 256๊ฐœ ๋ฏธ๋งŒ์ธ ๊ฒฝ์šฐ ํŒจ๋“œ).
์ตœ์ ํ™” ๋ฐ ํ‰๊ฐ€์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ์›Œ๋ฐ์—… 20๋‹จ๊ณ„, ํ•™์Šต ์†๋„ \(1e-6\), ๊ฐ€์ค‘์น˜ ๊ฐ์‡  0.001, ๋ฐฐ์น˜ ํฌ๊ธฐ 8(ํ•„์š”์— ๋”ฐ๋ผ ์‹คํ–‰ ์ „๋ฐ˜์— ๊ฑธ์ณ ์ด๋ฅผ ๊ณ ์ •ํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ผ๋””์–ธํŠธ ๋ˆ„์ )์˜ AdamW ์˜ตํ‹ฐ๋งˆ์ด์ €๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, <!startofthought|> ๋ฐ <!endofthought|> ์ž„๋ฒ ๋”ฉ ๊ทธ๋ผ๋””์–ธํŠธ ๊ฐ€์ค‘์น˜๋Š” \(1e2\)์ด๊ณ  ์ •์ฑ… ๊ฐ€์ค‘์น˜๋Š” \(1e6\)์ž…๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์ค‘์—๋Š” ์˜จ๋„ \(T=1\)๋กœ ์ƒ˜ํ”Œ๋งํ•˜๊ณ  ํ‰๊ฐ€ ์ค‘์—๋Š” ํƒ์š•์Šค๋Ÿฌ์šด ๋””์ฝ”๋”ฉ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜จ๋„ \(T=3\)์—์„œ REINFORCE ์†์‹ค์„ ๊ณ„์‚ฐํ•˜์—ฌ ์ƒ˜ํ”Œ์„ ์ค‘์š” ์ƒ˜ํ”Œ๋กœ ์ทจ๊ธ‰ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์— ์˜ˆ์ œ๋ฅผ ํ”„๋กฌํ”„ํŠธํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ ์ •๋‹ต์˜ ํ™•๋ฅ ์„ ์ง์ ‘ ๊ณ„์‚ฐํ•˜์—ฌ ์˜ˆ๋ฅผ ๋“ค์–ด, \(A\cdots E\) ์‚ฌ์ด์˜ ๊ฐ๊ด€์‹ ๋ฌธ์ œ์˜ ๊ฒฝ์šฐ, \(A\cdots E\)์— ํ•ด๋‹นํ•˜๋Š” ํ† ํฐ์— ๋Œ€ํ•œ ๋กœ์ง“์— ๋Œ€ํ•œ ์ •ํ™•๋„๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ํ›ˆ๋ จ์˜ ๊ฒฝ์šฐ ๊ฐ ์ƒ˜ํ”Œ์—์„œ 256๊ฐœ์˜ ํ† ํฐ์˜ ์ž„์˜ ๋ฒ”์œ„๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค(๋˜๋Š” ํ† ํฐ์ด 256๊ฐœ ๋ฏธ๋งŒ์ธ ๊ฒฝ์šฐ ํŒจ๋“œ).
Linear weight averaging not only works well for image processing and classification models, but is also effective for image generation models such as latent diffusion models. The release of Stable Diffusion [39] led to a variety of specialist fine-tunes of the model for a variety of different styles, from photorealism, to anime, to realistic painting styles. Popular open-source WebUIs [4] enabled hobbyists to experiment with merging different models using linear or spherical linear interpolation (SLERP) [47] of weight parameters of distinct fine-tuned Stable Diffusion models. This has led to the discovery of merged models that combine strengths of specialized models into a single model. For some time, the most popular Stable Diffusion models were neither the original base models nor the fine-tuned versions, but rather the merged models created by enthusiasts. This trend persists until the release of a more advanced base model, at which point the community's cycle of fine-tuning and merging is renewed.
์„ ํ˜• ๊ฐ€์ค‘์น˜ ํ‰๊ท ์€ ์ด๋ฏธ์ง€ ์ฒ˜๋ฆฌ ๋ฐ ๋ถ„๋ฅ˜ ๋ชจ๋ธ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ž ์žฌ ํ™•์‚ฐ ๋ชจ๋ธ๊ณผ ๊ฐ™์€ ์ด๋ฏธ์ง€ ์ƒ์„ฑ ๋ชจ๋ธ์—๋„ ํšจ๊ณผ์ ์ž…๋‹ˆ๋‹ค. Stable Diffusion [39]์˜ ์ถœ์‹œ๋กœ ์‚ฌ์‹ค์ฃผ์˜๋ถ€ํ„ฐ ์• ๋‹ˆ๋ฉ”์ด์…˜, ์‚ฌ์‹ค์ ์ธ ๊ทธ๋ฆผ ์Šคํƒ€์ผ ๋“ฑ ๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ์„ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ์ „๋ฌธ Fine-tunes ๋ชจ๋ธ์ด ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€์ค‘์ ์ธ ์˜คํ”ˆ ์†Œ์Šค WebUIs [4]๋ฅผ ํ†ตํ•ด ์• ํ˜ธ๊ฐ€๋“ค์€ ์„œ๋กœ ๋‹ค๋ฅธ Fine-tuned Stable Diffusion ๋ชจ๋ธ์˜ ๊ฐ€์ค‘์น˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์„ ํ˜• ๋˜๋Š” ๊ตฌํ˜• ์„ ํ˜• ๋ณด๊ฐ„(SLERP) [47]ํ•˜์—ฌ ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ์„ ๋ณ‘ํ•ฉํ•˜๋Š” ์‹คํ—˜์„ ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ „๋ฌธ ๋ชจ๋ธ์˜ ๊ฐ•์ ์„ ํ•˜๋‚˜์˜ ๋ชจ๋ธ๋กœ ๊ฒฐํ•ฉํ•œ ๋ณ‘ํ•ฉ ๋ชจ๋ธ์„ ๋ฐœ๊ฒฌํ•˜๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ•œ๋™์•ˆ ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” Stable Diffusion ๋ชจ๋ธ์€ ์›๋ž˜์˜ ๊ธฐ๋ณธ ๋ชจ๋ธ์ด๋‚˜ Fine-tuned ๋ฒ„์ „์ด ์•„๋‹ˆ๋ผ ์• ํ˜ธ๊ฐ€๋“ค์ด ๋งŒ๋“  ๋ณ‘ํ•ฉ ๋ชจ๋ธ์ด์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถ”์„ธ๋Š” ๋” ๋ฐœ์ „๋œ ๊ธฐ๋ณธ ๋ชจ๋ธ์ด ์ถœ์‹œ๋  ๋•Œ๊นŒ์ง€ ์ง€์†๋˜๋ฉฐ, ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ๋Š” Fine-tuning ๋ฐ ๋ณ‘ํ•ฉ ์ฃผ๊ธฐ๊ฐ€ ์ƒˆ๋กญ๊ฒŒ ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค.
์„ ํ˜• ๊ฐ€์ค‘์น˜ ํ‰๊ท ์€ ์ด๋ฏธ์ง€ ์ฒ˜๋ฆฌ ๋ฐ ๋ถ„๋ฅ˜ ๋ชจ๋ธ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ž ์žฌ ํ™•์‚ฐ ๋ชจ๋ธ๊ณผ ๊ฐ™์€ ์ด๋ฏธ์ง€ ์ƒ์„ฑ ๋ชจ๋ธ์—๋„ ํšจ๊ณผ์ ์ž…๋‹ˆ๋‹ค. Stable Diffusion [39]์˜ ์ถœ์‹œ๋กœ ์‚ฌ์‹ค์ฃผ์˜๋ถ€ํ„ฐ ์• ๋‹ˆ๋ฉ”์ด์…˜, ์‚ฌ์‹ค์ ์ธ ๊ทธ๋ฆผ ์Šคํƒ€์ผ ๋“ฑ ๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ์„ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ์ „๋ฌธ ๋ฏธ์„ธ ์กฐ์ • ๋ชจ๋ธ์ด ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€์ค‘์ ์ธ ์˜คํ”ˆ ์†Œ์Šค WebUI [4]๋ฅผ ํ†ตํ•ด ์• ํ˜ธ๊ฐ€๋“ค์€ ์„œ๋กœ ๋‹ค๋ฅธ ๋ฏธ์„ธ ์กฐ์ •๋œ Stable Diffusion ๋ชจ๋ธ์˜ ๊ฐ€์ค‘์น˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์„ ํ˜• ๋˜๋Š” ๊ตฌํ˜• ์„ ํ˜• ๋ณด๊ฐ„(SLERP) [47]ํ•˜์—ฌ ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ์„ ๋ณ‘ํ•ฉํ•˜๋Š” ์‹คํ—˜์„ ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ „๋ฌธ ๋ชจ๋ธ์˜ ๊ฐ•์ ์„ ํ•˜๋‚˜์˜ ๋ชจ๋ธ๋กœ ๊ฒฐํ•ฉํ•œ ๋ณ‘ํ•ฉ ๋ชจ๋ธ์„ ๋ฐœ๊ฒฌํ•˜๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ•œ๋™์•ˆ ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” Stable Diffusion ๋ชจ๋ธ์€ ์›๋ž˜์˜ ๊ธฐ๋ณธ ๋ชจ๋ธ์ด๋‚˜ ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ฒ„์ „์ด ์•„๋‹ˆ๋ผ ์• ํ˜ธ๊ฐ€๋“ค์ด ๋งŒ๋“  ๋ณ‘ํ•ฉ ๋ชจ๋ธ์ด์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถ”์„ธ๋Š” ๋” ๋ฐœ์ „๋œ ๊ธฐ๋ณธ ๋ชจ๋ธ์ด ์ถœ์‹œ๋  ๋•Œ๊นŒ์ง€ ์ง€์†๋˜๋ฉฐ, ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ๋Š” ๋ฏธ์„ธ ์กฐ์ • ๋ฐ ๋ณ‘ํ•ฉ ์ฃผ๊ธฐ๊ฐ€ ์ƒˆ๋กญ๊ฒŒ ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค.
And more recently, **Mixture of Experts (MoE)**[130, 131] have become very popular in LLM space too. MoEs enable models to be pre-trained with much less compute, which means one can dramatically scale up the model or dataset size with the same compute budget as a dense model. MoE consists of two main elements: **Sparse MoE layers**, which are used instead of dense feed-forward network (FFN) layers, and have a certain number of "experts" (e.g. 8), in which each expert is a neural network. In practice, the experts are FFNs, but they can also be more complex networks. **A gate network or router**, that determines which tokens are sent to which expert. It is worth noting that, one can send a token to more than one expert. How to route a token to an expert is one of the big decisions when working with MoEs - the router is composed of learned parameters and is pretrained at the same time as the rest of the network. Fig 29 provides an illustration of a Switch Transformer encoder block, which are used in MoE.
๊ทธ๋ฆฌ๊ณ  ์ตœ๊ทผ์—๋Š” **Mixture of Experts (MoE)**[130, 131]๋„ LLM ์˜์—ญ์—์„œ ํฐ ์ธ๊ธฐ๋ฅผ ์–ป๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. MoE๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ํ›จ์”ฌ ์ ์€ ์—ฐ์‚ฐ์œผ๋กœ ๋ชจ๋ธ์„ pre-trainedํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ dense ๋ชจ๋ธ๊ณผ ๋™์ผํ•œ ์—ฐ์‚ฐ ์˜ˆ์‚ฐ์œผ๋กœ ๋ชจ๋ธ ๋˜๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ ํฌ๊ธฐ๋ฅผ ํš๊ธฐ์ ์œผ๋กœ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. MoE๋Š” ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: **Sparse MoE layers**๋Š” dense feed-forward network (FFN) ๋ ˆ์ด์–ด ๋Œ€์‹  ์‚ฌ์šฉ๋˜๋ฉฐ, ๊ฐ ์ „๋ฌธ๊ฐ€๊ฐ€ ์‹ ๊ฒฝ๋ง์ธ ํŠน์ • ์ˆ˜์˜ "experts"(์˜ˆ: 8)๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ experts๋Š” FFN์ด์ง€๋งŒ ๋” ๋ณต์žกํ•œ ๋„คํŠธ์›Œํฌ์ผ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. **A gate network or router**๋Š” ์–ด๋–ค tokens์ด ์–ด๋–ค expert์—๊ฒŒ ์ „์†ก๋˜๋Š”์ง€ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ํ•œ token์„ ๋‘ ๊ฐœ ์ด์ƒ์˜ expert์—๊ฒŒ ๋ณด๋‚ผ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์ฃผ๋ชฉํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. token์„ expert์—๊ฒŒ ๋ผ์šฐํŒ…ํ•˜๋Š” ๋ฐฉ๋ฒ•์€ MoE๋กœ ์ž‘์—…ํ•  ๋•Œ ์ค‘์š”ํ•œ ๊ฒฐ์ • ์ค‘ ํ•˜๋‚˜์ด๋ฉฐ, router๋Š” ํ•™์Šต๋œ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ ๋„คํŠธ์›Œํฌ์˜ ๋‚˜๋จธ์ง€ ๋ถ€๋ถ„๊ณผ ๋™์‹œ์— pre-trained๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 29๋Š” MoE์— ์‚ฌ์šฉ๋˜๋Š” Switch Transformer encoder ๋ธ”๋ก์˜ ์˜ˆ์‹œ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆฌ๊ณ  ์ตœ๊ทผ์—๋Š” **์ „๋ฌธ๊ฐ€ ํ˜ผํ•ฉ(MoE)**[130, 131]๋„ LLM ์˜์—ญ์—์„œ ํฐ ์ธ๊ธฐ๋ฅผ ์–ป๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. MoE๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ํ›จ์”ฌ ์ ์€ ์—ฐ์‚ฐ์œผ๋กœ ๋ชจ๋ธ์„ ์‚ฌ์ „ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋ฐ€๋„ ๋ชจ๋ธ๊ณผ ๋™์ผํ•œ ์—ฐ์‚ฐ ์˜ˆ์‚ฐ์œผ๋กœ ๋ชจ๋ธ ๋˜๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ ํฌ๊ธฐ๋ฅผ ํš๊ธฐ์ ์œผ๋กœ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. MoE๋Š” ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: **์ŠคํŒŒ์Šค MoE ๋ ˆ์ด์–ด**๋Š” ๋ฐ€๋„ ํ”ผ๋“œํฌ์›Œ๋“œ ๋„คํŠธ์›Œํฌ(FFN) ๋ ˆ์ด์–ด ๋Œ€์‹  ์‚ฌ์šฉ๋˜๋ฉฐ, ๊ฐ ์ „๋ฌธ๊ฐ€๊ฐ€ ์‹ ๊ฒฝ๋ง์ธ ํŠน์ • ์ˆ˜์˜ "์ „๋ฌธ๊ฐ€"(์˜ˆ: 8)๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ์ „๋ฌธ๊ฐ€๋Š” FFN์ด์ง€๋งŒ ๋” ๋ณต์žกํ•œ ๋„คํŠธ์›Œํฌ์ผ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. **๊ฒŒ์ดํŠธ ๋„คํŠธ์›Œํฌ ๋˜๋Š” ๋ผ์šฐํ„ฐ**๋Š” ์–ด๋–ค ํ† ํฐ์ด ์–ด๋–ค ์ „๋ฌธ๊ฐ€์—๊ฒŒ ์ „์†ก๋˜๋Š”์ง€ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ํ•œ ํ† ํฐ์„ ๋‘ ๊ฐœ ์ด์ƒ์˜ ์ „๋ฌธ๊ฐ€์—๊ฒŒ ๋ณด๋‚ผ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์ฃผ๋ชฉํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ํ† ํฐ์„ ์ „๋ฌธ๊ฐ€์—๊ฒŒ ๋ผ์šฐํŒ…ํ•˜๋Š” ๋ฐฉ๋ฒ•์€ MoE๋กœ ์ž‘์—…ํ•  ๋•Œ ์ค‘์š”ํ•œ ๊ฒฐ์ • ์ค‘ ํ•˜๋‚˜์ด๋ฉฐ, ๋ผ์šฐํ„ฐ๋Š” ํ•™์Šต๋œ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ ๋„คํŠธ์›Œํฌ์˜ ๋‚˜๋จธ์ง€ ๋ถ€๋ถ„๊ณผ ๋™์‹œ์— ์‚ฌ์ „ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 29๋Š” MoE์— ์‚ฌ์šฉ๋˜๋Š” ์Šค์œ„์น˜ ํŠธ๋žœ์Šคํฌ๋จธ ์ธ์ฝ”๋” ๋ธ”๋ก์˜ ์˜ˆ์‹œ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
**Across data selection methods**: We first take the same configuration as Section 4.2, where we have a starting source dataset of 40B tokens, use each of our data selection methods with \(R=0.25\) to select a subset of documents, and repeat over these documents until we reach the target token budget of 40B tokens. Note that this is at the 1.3B model scale. In Figure A13 we see that repeating data selected by both SemDeDup and SSL prototypes also outperforms randomly selecting new data. However, we quickly notice that for _fixed_ data selection strategy (e.g. _fixed_ column in Figure A13), repeating tokens either outperforms or matched selecting new tokens. In other words: cleverly repeating tokens can outperform randomly selecting new tokens, but if we fix the data selection strategy (random, SemDeDup, SSL prototypes, or D4) then it is usually preferable to select new tokens. We also note in Figure A16 that D4 outperforms other methods, although by a smaller margin than in the fixed-compute regime.
**๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์— ๋”ฐ๋ฅธ ์ฐจ์ด**: ๋จผ์ € 40B tokens์˜ ์‹œ์ž‘ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์žˆ๊ณ , ๊ฐ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ \(R=0.25\)์™€ ํ•จ๊ป˜ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ์„œ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์„ ํƒํ•œ ๋‹ค์Œ, ๋ชฉํ‘œ token ์˜ˆ์‚ฐ์ธ 40B tokens์— ๋„๋‹ฌํ•  ๋•Œ๊นŒ์ง€ ์ด๋Ÿฌํ•œ ๋ฌธ์„œ์— ๋Œ€ํ•ด ๋ฐ˜๋ณตํ•˜๋Š” ์„น์…˜ 4.2์™€ ๋™์ผํ•œ ๊ตฌ์„ฑ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” 1.3B ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ์ด๋ฃจ์–ด์ง„๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”. ๊ทธ๋ฆผ A13์—์„œ ๋ณด๋ฉด SemDeDup๊ณผ SSL ํ”„๋กœํ† ํƒ€์ž…์œผ๋กœ ์„ ํƒํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ๋„ ๋ฌด์ž‘์œ„๋กœ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ _fixed_ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต(์˜ˆ: ๊ทธ๋ฆผ A13์˜ _fixed_ ์—ด)์˜ ๊ฒฝ์šฐ tokens์„ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ์ด ์ƒˆ๋กœ์šด tokens์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๊ฑฐ๋‚˜ ์ผ์น˜ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊ธˆ๋ฐฉ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฆ‰, ์˜๋ฆฌํ•˜๊ฒŒ tokens์„ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ์ด ๋ฌด์ž‘์œ„๋กœ ์ƒˆ๋กœ์šด tokens์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต(random, SemDeDup, SSL prototypes, ๋˜๋Š” D4)์„ ๊ณ ์ •ํ•˜๋ฉด ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒˆ๋กœ์šด tokens์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ๋ฐ”๋žŒ์งํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ทธ๋ฆผ A16์—์„œ D4๊ฐ€ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜์ง€๋งŒ fixed-compute ์ฒด์ œ๋ณด๋‹ค ์ž‘์€ ํญ ์ฐจ์ด๋กœ ์•ž์„œ ๋‚˜๊ฐ€๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์— ๋”ฐ๋ฅธ ์ฐจ์ด**: ๋จผ์ € 40B ํ† ํฐ์˜ ์‹œ์ž‘ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์žˆ๊ณ , ๊ฐ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐฉ๋ฒ•์„ \(R=0.25\)์™€ ํ•จ๊ป˜ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ์„œ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์„ ํƒํ•œ ๋‹ค์Œ, ๋ชฉํ‘œ ํ† ํฐ ์˜ˆ์‚ฐ์ธ 40B ํ† ํฐ์— ๋„๋‹ฌํ•  ๋•Œ๊นŒ์ง€ ์ด๋Ÿฌํ•œ ๋ฌธ์„œ์— ๋Œ€ํ•ด ๋ฐ˜๋ณตํ•˜๋Š” ์„น์…˜ 4.2์™€ ๋™์ผํ•œ ๊ตฌ์„ฑ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” 1.3B ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ์ด๋ฃจ์–ด์ง„๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”. ๊ทธ๋ฆผ A13์—์„œ ๋ณด๋ฉด SemDeDup๊ณผ SSL ํ”„๋กœํ† ํƒ€์ž…์œผ๋กœ ์„ ํƒํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ๋„ ๋ฌด์ž‘์œ„๋กœ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ณ ์ •๋œ_ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต(์˜ˆ: ๊ทธ๋ฆผ A13์˜ _๊ณ ์ •_ ์—ด)์˜ ๊ฒฝ์šฐ ํ† ํฐ์„ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ์ด ์ƒˆ๋กœ์šด ํ† ํฐ์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๊ฑฐ๋‚˜ ์ผ์น˜ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊ธˆ๋ฐฉ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฆ‰, ์˜๋ฆฌํ•˜๊ฒŒ ํ† ํฐ์„ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ์ด ๋ฌด์ž‘์œ„๋กœ ์ƒˆ๋กœ์šด ํ† ํฐ์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต(๋ฌด์ž‘์œ„, SemDeDup, SSL ํ”„๋กœํ† ํƒ€์ž… ๋˜๋Š” D4)์„ ๊ณ ์ •ํ•˜๋ฉด ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒˆ๋กœ์šด ํ† ํฐ์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ๋ฐ”๋žŒ์งํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ทธ๋ฆผ A16์—์„œ D4๊ฐ€ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜์ง€๋งŒ ๊ณ ์ • ์ปดํ“จํŒ… ์ฒด์ œ๋ณด๋‹ค ์ž‘์€ ํญ ์ฐจ์ด๋กœ ์•ž์„œ ๋‚˜๊ฐ€๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Multi-stage Instruction Tuning.** For instruction tuning, there are two kinds of important instruction data, namely task-formatted instructions and daily chat instructions. Generally, the former has a significantly larger volume than the latter. It is important to balance the training with the two kinds of instruction data. In addition to carefully mixing different instruction data, we can also adopt a multi-stage instruction tuning strategy [352], where LLMs are first fine-tuned with large-scale task-formatted instructions and subsequently fine-tuned on daily chat ones. To avoid the capacity forgetting issue, it is also useful to add an amount of task-formatted instructions at the second stage. Actually, such a multi-stage tuning strategy can be also applied to other settings for instruction tuning. For example, we can schedule different fine-tuning stages with progressively increased levels on difficulty and complexity, and gradually improve the capacities of LLMs to follow complex instructions.
**Multi-stage Instruction Tuning**. Instruction Tuning์—๋Š” ์ž‘์—… ํ˜•์‹์˜ instruction๊ณผ ์ผ์ƒ์ ์ธ ์ฑ„ํŒ… instruction์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ค‘์š”ํ•œ instruction ๋ฐ์ดํ„ฐ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ „์ž์˜ ์–‘์ด ํ›„์ž๋ณด๋‹ค ํ›จ์”ฌ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋‘ ์ข…๋ฅ˜์˜ instruction ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ์˜ ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ instruction ๋ฐ์ดํ„ฐ๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ํ˜ผํ•ฉํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„, LLM์„ ๋จผ์ € ๋Œ€๊ทœ๋ชจ ์ž‘์—… ํ˜•์‹์˜ instruction์œผ๋กœ Fine-tuningํ•œ ๋‹ค์Œ ์ผ์ƒ์ ์ธ ์ฑ„ํŒ… instruction์œผ๋กœ Fine-tuningํ•˜๋Š” Multi-stage Instruction Tuning ์ „๋žต[352]์„ ์ฑ„ํƒํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฉ๋Ÿ‰ ์žŠ์–ด๋ฒ„๋ฆผ ๋ฌธ์ œ๋ฅผ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ ์ž‘์—… ํ˜•์‹์˜ instruction์„ ์ผ์ •๋Ÿ‰ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ๋„ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ์ด๋Ÿฌํ•œ Multi-stage Tuning ์ „๋žต์€ Instruction Tuning์„ ์œ„ํ•œ ๋‹ค๋ฅธ ์„ค์ •์—๋„ ์ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ ์ฐจ ๋‚œ์ด๋„์™€ ๋ณต์žก๋„๋ฅผ ๋†’์ธ ๋‹ค์–‘ํ•œ Fine-tuning ๋‹จ๊ณ„๋ฅผ ๊ณ„ํšํ•˜๊ณ  ๋ณต์žกํ•œ instruction์„ ๋”ฐ๋ฅด๋Š” LLM์˜ ์—ญ๋Ÿ‰์„ ์ ์ง„์ ์œผ๋กœ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**๋‹ค๋‹จ๊ณ„ ๋ช…๋ น์–ด ํŠœ๋‹**. ๋ช…๋ น์–ด ํŠœ๋‹์—๋Š” ์ž‘์—… ํ˜•์‹์˜ ๋ช…๋ น์–ด์™€ ์ผ์ƒ์ ์ธ ์ฑ„ํŒ… ๋ช…๋ น์–ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ „์ž์˜ ์–‘์ด ํ›„์ž๋ณด๋‹ค ํ›จ์”ฌ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋‘ ์ข…๋ฅ˜์˜ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ์˜ ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ํ˜ผํ•ฉํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„, LLM์„ ๋จผ์ € ๋Œ€๊ทœ๋ชจ ์ž‘์—… ํ˜•์‹์˜ ๋ช…๋ น์–ด๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•œ ๋‹ค์Œ ์ผ์ƒ์ ์ธ ์ฑ„ํŒ… ๋ช…๋ น์–ด๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋‹ค๋‹จ๊ณ„ ๋ช…๋ น์–ด ํŠœ๋‹ ์ „๋žต[352]์„ ์ฑ„ํƒํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฉ๋Ÿ‰ ์žŠ์–ด๋ฒ„๋ฆผ ๋ฌธ์ œ๋ฅผ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ ์ž‘์—… ํ˜•์‹์˜ ๋ช…๋ น์–ด๋ฅผ ์ผ์ •๋Ÿ‰ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ๋„ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ์ด๋Ÿฌํ•œ ๋‹ค๋‹จ๊ณ„ ํŠœ๋‹ ์ „๋žต์€ ๋ช…๋ น์–ด ํŠœ๋‹์„ ์œ„ํ•œ ๋‹ค๋ฅธ ์„ค์ •์—๋„ ์ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ ์ฐจ ๋‚œ์ด๋„์™€ ๋ณต์žก๋„๋ฅผ ๋†’์ธ ๋‹ค์–‘ํ•œ ๋ฏธ์„ธ ์กฐ์ • ๋‹จ๊ณ„๋ฅผ ๊ณ„ํšํ•˜๊ณ  ๋ณต์žกํ•œ ๋ช…๋ น์„ ๋”ฐ๋ฅด๋Š” LLM์˜ ์—ญ๋Ÿ‰์„ ์ ์ง„์ ์œผ๋กœ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
deep learning in its core, is a set of mathematical functions applied to matrices, with a specific precision for model weights. Reducing the precision of the weights can be used to reduce the size of the model and also make it faster. As an example, Float-32 operations compared to Int-8 operations are slower. This process, which is called quantization, can be applied in different phases. Main approaches for model quantization can be categorized as: post training quantization and quantization-aware training. Post-training quantization is concerned with quantized trained models in two well-known methods: dynamic and static. Dynamic post-training quantization computes the range of quantization on the runtime and is slower compared to static. Quantization-aware training adds quantization criteria into training, and a quantized model is trained and optimized during training process. This approach ensures that the end model will have good performance and also does not need to be quantized after training.
๋”ฅ ๋Ÿฌ๋‹์˜ ํ•ต์‹ฌ์€ ํ–‰๋ ฌ์— ์ ์šฉ๋˜๋Š” ์ผ๋ จ์˜ ์ˆ˜ํ•™์  ํ•จ์ˆ˜๋กœ, ๋ชจ๋ธ ๊ฐ€์ค‘์น˜์— ๋Œ€ํ•œ ํŠน์ • ์ •๋ฐ€๋„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ€์ค‘์น˜์˜ ์ •๋ฐ€๋„๋ฅผ ๋‚ฎ์ถ”๋ฉด ๋ชจ๋ธ์˜ ํฌ๊ธฐ๋ฅผ ์ค„์ด๊ณ  ๋” ๋น ๋ฅด๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Float-32 ์—ฐ์‚ฐ์€ Int-8 ์—ฐ์‚ฐ์— ๋น„ํ•ด ์†๋„๊ฐ€ ๋Š๋ฆฝ๋‹ˆ๋‹ค. Quantization์ด๋ผ๊ณ  ํ•˜๋Š” ์ด ๊ณผ์ •์€ ์—ฌ๋Ÿฌ ๋‹จ๊ณ„๋กœ ์ ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ Quantization์˜ ์ฃผ์š” ์ ‘๊ทผ ๋ฐฉ์‹์€ ์‚ฌํ›„ ํ›ˆ๋ จ Quantization๊ณผ Quantization-aware training์œผ๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌํ›„ ํ›ˆ๋ จ Quantization์€ ๋™์  ๋ฐ ์ •์ ์ด๋ผ๋Š” ์ž˜ ์•Œ๋ ค์ง„ ๋‘ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์œผ๋กœ ํ›ˆ๋ จ๋œ Quantized ๋ชจ๋ธ์— ๊ด€ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋™์  ์‚ฌํ›„ ํ›ˆ๋ จ Quantization์€ ๋Ÿฐํƒ€์ž„์— Quantization ๋ฒ”์œ„๋ฅผ ๊ณ„์‚ฐํ•˜๋ฉฐ ์ •์  Quantization์— ๋น„ํ•ด ์†๋„๊ฐ€ ๋Š๋ฆฝ๋‹ˆ๋‹ค. Quantization-aware training์€ ํ›ˆ๋ จ์— Quantization ๊ธฐ์ค€์„ ์ถ”๊ฐ€ํ•˜๋ฉฐ, Quantized ๋ชจ๋ธ์€ ํ›ˆ๋ จ ๊ณผ์ •์—์„œ ํ›ˆ๋ จ ๋ฐ ์ตœ์ ํ™”๋ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ตœ์ข… ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋ณด์žฅํ•˜๊ณ  ํ›ˆ๋ จ ํ›„ Quantizationํ•  ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค.
๋”ฅ ๋Ÿฌ๋‹์˜ ํ•ต์‹ฌ์€ ํ–‰๋ ฌ์— ์ ์šฉ๋˜๋Š” ์ผ๋ จ์˜ ์ˆ˜ํ•™์  ํ•จ์ˆ˜๋กœ, ๋ชจ๋ธ ๊ฐ€์ค‘์น˜์— ๋Œ€ํ•œ ํŠน์ • ์ •๋ฐ€๋„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ€์ค‘์น˜์˜ ์ •๋ฐ€๋„๋ฅผ ๋‚ฎ์ถ”๋ฉด ๋ชจ๋ธ์˜ ํฌ๊ธฐ๋ฅผ ์ค„์ด๊ณ  ๋” ๋น ๋ฅด๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Float-32 ์—ฐ์‚ฐ์€ Int-8 ์—ฐ์‚ฐ์— ๋น„ํ•ด ์†๋„๊ฐ€ ๋Š๋ฆฝ๋‹ˆ๋‹ค. ์–‘์žํ™”๋ผ๊ณ  ํ•˜๋Š” ์ด ๊ณผ์ •์€ ์—ฌ๋Ÿฌ ๋‹จ๊ณ„๋กœ ์ ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ์–‘์žํ™”์˜ ์ฃผ์š” ์ ‘๊ทผ ๋ฐฉ์‹์€ ์‚ฌํ›„ ํ›ˆ๋ จ ์–‘์žํ™”์™€ ์–‘์žํ™” ์ธ์‹ ํ›ˆ๋ จ์œผ๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌํ›„ ํ›ˆ๋ จ ์–‘์žํ™”๋Š” ๋™์  ๋ฐ ์ •์ ์ด๋ผ๋Š” ์ž˜ ์•Œ๋ ค์ง„ ๋‘ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์œผ๋กœ ํ›ˆ๋ จ๋œ ์–‘์žํ™” ๋ชจ๋ธ์— ๊ด€ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋™์  ์‚ฌํ›„ ํ›ˆ๋ จ ์–‘์žํ™”๋Š” ๋Ÿฐํƒ€์ž„์— ์–‘์žํ™” ๋ฒ”์œ„๋ฅผ ๊ณ„์‚ฐํ•˜๋ฉฐ ์ •์  ์–‘์žํ™”์— ๋น„ํ•ด ์†๋„๊ฐ€ ๋Š๋ฆฝ๋‹ˆ๋‹ค. ์–‘์žํ™” ์ธ์‹ ํ›ˆ๋ จ์€ ํ›ˆ๋ จ์— ์–‘์žํ™” ๊ธฐ์ค€์„ ์ถ”๊ฐ€ํ•˜๋ฉฐ, ์–‘์žํ™”๋œ ๋ชจ๋ธ์€ ํ›ˆ๋ จ ๊ณผ์ •์—์„œ ํ›ˆ๋ จ ๋ฐ ์ตœ์ ํ™”๋ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ตœ์ข… ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋ณด์žฅํ•˜๊ณ  ํ›ˆ๋ จ ํ›„ ์–‘์žํ™”ํ•  ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค.
Recent advances in language modeling consist in pretraining highly parameterized neural networks on extremely large web-mined text corpora. Training and inference with such models can be costly in practice, which incentivizes the use of smaller counterparts. However, it has been observed that smaller models can suffer from saturation, characterized as a drop in performance at some advanced point in training followed by a plateau. In this paper, we find that such saturation can be explained by a mismatch between the hidden dimension of smaller models and the high rank of the target contextual probability distribution. This mismatch affects the performance of the linear prediction head used in such models through the well-known softmax bottleneck phenomenon. We measure the effect of the softmax bottleneck in various settings and find that models based on less than 1000 hidden dimensions tend to adopt degenerate latent representations in late pretraining, which leads to reduced evaluation performance.
์ตœ๊ทผ ์–ธ์–ด ๋ชจ๋ธ๋ง์˜ ๋ฐœ์ „์€ ๋งค์šฐ ํฐ ์›น ๋งˆ์ด๋‹ ํ…์ŠคํŠธ ์ฝ”ํผ๋ผ์— ๋Œ€ํ•ด ๊ณ ๋„๋กœ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ์‹ ๊ฒฝ๋ง์„ pretrainingํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ๋กœ ํ›ˆ๋ จํ•˜๊ณ  ์ถ”๋ก ํ•˜๋Š” ๋ฐ๋Š” ์‹ค์ œ๋กœ ๋งŽ์€ ๋น„์šฉ์ด ๋“ค ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋” ์ž‘์€ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๋„๋ก ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ž‘์€ ๋ชจ๋ธ์€ ํ›ˆ๋ จ์˜ ์–ด๋Š ์ •๋„ ์ง„ํ–‰๋œ ์‹œ์ ์—์„œ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๊ณ  ์ •์ฒด๋˜๋Š” ํŠน์ง•์„ ๊ฐ€์ง„ ํฌํ™” ํ˜„์ƒ์„ ๊ฒช์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๊ด€์ฐฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ํฌํ™” ํ˜„์ƒ์ด ์†Œํ˜• ๋ชจ๋ธ์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์›๊ณผ ๋ชฉํ‘œ ์ปจํ…์ŠคํŠธ ํ™•๋ฅ  ๋ถ„ํฌ์˜ ๋†’์€ ์ˆœ์œ„ ์‚ฌ์ด์˜ ๋ถˆ์ผ์น˜๋กœ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ถˆ์ผ์น˜๋Š” ์ž˜ ์•Œ๋ ค์ง„ softmax bottleneck ํ˜„์ƒ์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์— ์‚ฌ์šฉ๋˜๋Š” ์„ ํ˜• ์˜ˆ์ธก ํ—ค๋“œ์˜ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์„ค์ •์—์„œ softmax bottleneck ํ˜„์ƒ์˜ ์˜ํ–ฅ์„ ์ธก์ •ํ•œ ๊ฒฐ๊ณผ, 1000๊ฐœ ๋ฏธ๋งŒ์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์›์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๋ชจ๋ธ์€ ํ›„๋ฐ˜๋ถ€ pretraining์—์„œ ํ‡ดํ™”๋œ ์ž ์žฌ ํ‘œํ˜„์„ ์ฑ„ํƒํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ‰๊ฐ€ ์„ฑ๋Šฅ ์ €ํ•˜๋กœ ์ด์–ด์ง„๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
์ตœ๊ทผ ์–ธ์–ด ๋ชจ๋ธ๋ง์˜ ๋ฐœ์ „์€ ๋งค์šฐ ํฐ ์›น ๋งˆ์ด๋‹ ํ…์ŠคํŠธ ์ฝ”ํผ๋ผ์— ๋Œ€ํ•ด ๊ณ ๋„๋กœ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ์‹ ๊ฒฝ๋ง์„ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ๋กœ ํ›ˆ๋ จํ•˜๊ณ  ์ถ”๋ก ํ•˜๋Š” ๋ฐ๋Š” ์‹ค์ œ๋กœ ๋งŽ์€ ๋น„์šฉ์ด ๋“ค ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋” ์ž‘์€ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๋„๋ก ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ž‘์€ ๋ชจ๋ธ์€ ํ›ˆ๋ จ์˜ ์–ด๋Š ์ •๋„ ์ง„ํ–‰๋œ ์‹œ์ ์—์„œ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๊ณ  ์ •์ฒด๋˜๋Š” ํŠน์ง•์„ ๊ฐ€์ง„ ํฌํ™” ํ˜„์ƒ์„ ๊ฒช์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๊ด€์ฐฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ํฌํ™” ํ˜„์ƒ์ด ์†Œํ˜• ๋ชจ๋ธ์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์›๊ณผ ๋ชฉํ‘œ ์ปจํ…์ŠคํŠธ ํ™•๋ฅ  ๋ถ„ํฌ์˜ ๋†’์€ ์ˆœ์œ„ ์‚ฌ์ด์˜ ๋ถˆ์ผ์น˜๋กœ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ถˆ์ผ์น˜๋Š” ์ž˜ ์•Œ๋ ค์ง„ ์†Œํ”„ํŠธ๋งฅ์Šค ๋ณ‘๋ชฉ ํ˜„์ƒ์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์— ์‚ฌ์šฉ๋˜๋Š” ์„ ํ˜• ์˜ˆ์ธก ํ—ค๋“œ์˜ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์„ค์ •์—์„œ ์†Œํ”„ํŠธ๋งฅ์Šค ๋ณ‘๋ชฉ ํ˜„์ƒ์˜ ์˜ํ–ฅ์„ ์ธก์ •ํ•œ ๊ฒฐ๊ณผ, 1000๊ฐœ ๋ฏธ๋งŒ์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์›์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๋ชจ๋ธ์€ ํ›„๋ฐ˜๋ถ€ ์‚ฌ์ „ ํ•™์Šต์—์„œ ํ‡ดํ™”๋œ ์ž ์žฌ ํ‘œํ˜„์„ ์ฑ„ํƒํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ‰๊ฐ€ ์„ฑ๋Šฅ ์ €ํ•˜๋กœ ์ด์–ด์ง„๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
**Kernel Layer.** The kernel layer is divided into two primary components: the OS Kernel and the LLM Kernel, each serving the unique requirements of non-LLM and LLM-specific operations, respectively. This distinction allows the LLM kernel to focus on LLM specific tasks such as context management and agent scheduling, which are essential for handling LLM-related activities and are not typically within the purview of standard OS kernel functions. Our work primarily concentrates on enhancing the LLM kernel without making significant alterations to the existing OS kernel structure. The LLM kernel is equipped with several key modules, including the LLM system call interface, agent scheduler, context manager, memory manager, storage manager, tool manager, and access manager. These components are designed to address the diverse execution needs of agent applications, ensuring efficient management and execution within the AIOS framework. The specifics of these modules will be further detailed in Section 4.
**Kernel Layer.** Kernel layer๋Š” OS Kernel๊ณผ LLM Kernel, ๊ฐ๊ฐ ๋น„LLM ๋ฐ LLM-specific ์ž‘์—…์˜ ๊ณ ์œ ํ•œ ์š”๊ตฌ ์‚ฌํ•ญ์„ ์ œ๊ณตํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๋‚˜๋‰ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ตฌ๋ถ„์„ ํ†ตํ•ด LLM Kernel์€ LLM ๊ด€๋ จ ํ™œ๋™์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ํ•„์ˆ˜์ ์ธ context management ๋ฐ agent scheduling๊ณผ ๊ฐ™์€ LLM-specific ์ž‘์—…์— ์ง‘์ค‘ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ํ‘œ์ค€ OS Kernel ๊ธฐ๋Šฅ์˜ ๋ฒ”์œ„ ๋‚ด์— ์žˆ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ๊ธฐ์กด OS Kernel ๊ตฌ์กฐ๋ฅผ ํฌ๊ฒŒ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š๊ณ  LLM Kernel์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. LLM Kernel์—๋Š” LLM system call interface, agent scheduler, context manager, memory manager, storage manager, tool manager, access manager ๋“ฑ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ฃผ์š” ๋ชจ๋“ˆ์ด ํƒ‘์žฌ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ตฌ์„ฑ ์š”์†Œ๋Š” agent ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ๋‹ค์–‘ํ•œ ์‹คํ–‰ ์š”๊ตฌ ์‚ฌํ•ญ์„ ํ•ด๊ฒฐํ•˜๋„๋ก ์„ค๊ณ„๋˜์–ด AIOS ํ”„๋ ˆ์ž„์›Œํฌ ๋‚ด์—์„œ ํšจ์œจ์ ์ธ ๊ด€๋ฆฌ ๋ฐ ์‹คํ–‰์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋“ˆ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ์„น์…˜ 4์—์„œ ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
**์ปค๋„ ๊ณ„์ธต.** ์ปค๋„ ๊ณ„์ธต์€ OS ์ปค๋„๊ณผ LLM ํŠน์ • ์ž‘์—…์„ ๊ฐ๊ฐ ์œ„ํ•œ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๋‚˜๋‰˜๋ฉฐ, ๊ฐ๊ฐ ๋น„LLM ๋ฐ LLM ํŠน์ • ์ž‘์—…์˜ ๊ณ ์œ ํ•œ ์š”๊ตฌ ์‚ฌํ•ญ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ตฌ๋ถ„์„ ํ†ตํ•ด LLM ์ปค๋„์€ LLM ๊ด€๋ จ ํ™œ๋™์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ํ•„์ˆ˜์ ์ธ ์ปจํ…์ŠคํŠธ ๊ด€๋ฆฌ ๋ฐ ์—์ด์ „ํŠธ ์Šค์ผ€์ค„๋ง๊ณผ ๊ฐ™์€ LLM ํŠน์ • ์ž‘์—…์— ์ง‘์ค‘ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ํ‘œ์ค€ OS ์ปค๋„ ๊ธฐ๋Šฅ์˜ ๋ฒ”์œ„ ๋‚ด์— ์žˆ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ๊ธฐ์กด OS ์ปค๋„ ๊ตฌ์กฐ๋ฅผ ํฌ๊ฒŒ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š๊ณ  LLM ์ปค๋„์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. LLM ์ปค๋„์—๋Š” LLM ์‹œ์Šคํ…œ ํ˜ธ์ถœ ์ธํ„ฐํŽ˜์ด์Šค, ์—์ด์ „ํŠธ ์Šค์ผ€์ค„๋Ÿฌ, ์ปจํ…์ŠคํŠธ ๊ด€๋ฆฌ์ž, ๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ์ž, ์Šคํ† ๋ฆฌ์ง€ ๊ด€๋ฆฌ์ž, ๋„๊ตฌ ๊ด€๋ฆฌ์ž, ์•ก์„ธ์Šค ๊ด€๋ฆฌ์ž ๋“ฑ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ฃผ์š” ๋ชจ๋“ˆ์ด ํƒ‘์žฌ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ์—์ด์ „ํŠธ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ๋‹ค์–‘ํ•œ ์‹คํ–‰ ์š”๊ตฌ ์‚ฌํ•ญ์„ ํ•ด๊ฒฐํ•˜๋„๋ก ์„ค๊ณ„๋˜์–ด AIOS ํ”„๋ ˆ์ž„์›Œํฌ ๋‚ด์—์„œ ํšจ์œจ์ ์ธ ๊ด€๋ฆฌ ๋ฐ ์‹คํ–‰์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋“ˆ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ์„น์…˜ 4์—์„œ ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
We train the model in a _many-to-many_ multilingual machine translation manner, starting with ALMA-13B-LoRA as the initial checkpoint. During the training phase, we focus exclusively on updating the weights of the added LoRA parameters. These weights have a rank of 16 and only add an additional 12M parameters to the original 13B size of the model. We adhere to the default \(\beta\) value of 0.1 as suggested by Rafailov et al. (2023). The fine-tuning process of ALMA-13B-LoRA involves a batch size of 128, a warm-up ratio of 0.01, spanning a single epoch, and accommodating sequences with a maximum length of 512 tokens. To optimize training efficiency, we integrate the deepspeed tool (Rasley et al., 2020). We utilize the same prompt as Xu et al. (2023) and do not compute the loss for the prompt. While our primary focus is on the performance of 13B models, CPO markedly benefits 7B models as well. Consequently, we also release ALMA-7B-R and provide a detailed discussion of its performance in Appendix A.
์ดˆ๊ธฐ ์ฒดํฌํฌ์ธํŠธ๋กœ ALMA-13B-LoRA๋ฅผ ์‹œ์ž‘์œผ๋กœ _๋‹ค๋Œ€๋‹ค_ ๋‹ค๊ตญ์–ด ๊ธฐ๊ณ„ ๋ฒˆ์—ญ ๋ฐฉ์‹์œผ๋กœ ๋ชจ๋ธ์„ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋‹จ๊ณ„์—์„œ๋Š” ์ถ”๊ฐ€๋œ LoRA ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๊ฐ€์ค‘์น˜๋งŒ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐ ์ง‘์ค‘ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฐ€์ค‘์น˜๋Š” ์ˆœ์œ„๊ฐ€ 16์ด๋ฉฐ ๋ชจ๋ธ์˜ ์›๋ž˜ 13B ํฌ๊ธฐ์— 12M์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋งŒ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Rafailov ๋“ฑ(2023)์ด ์ œ์•ˆํ•œ ๊ธฐ๋ณธ \(\beta\) ๊ฐ’ 0.1์„ ์ค€์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ALMA-13B-LoRA์˜ Fine-tuning ํ”„๋กœ์„ธ์Šค์—๋Š” Batch Size 128, ์›Œ๋ฐ์—… ๋น„์œจ 0.01, ํ•œ ๋ฒˆ์˜ ์—ํฌํฌ์— ๊ฑธ์ณ ์ตœ๋Œ€ ๊ธธ์ด 512 ํ† ํฐ์˜ ์‹œํ€€์Šค๋ฅผ ์ˆ˜์šฉํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ํšจ์œจ์„ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด deepspeed ๋„๊ตฌ(Rasley et al., 2020)๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. Xu ๋“ฑ(2023)๊ณผ ๋™์ผํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ Loss๋ฅผ ๊ณ„์‚ฐํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. 13B ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์— ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์ง€๋งŒ, CPO๋Š” 7B ๋ชจ๋ธ์—๋„ ํ˜„์ €ํ•œ ์ด์ ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ALMA-7B-R๋„ ์ถœ์‹œํ•˜๊ณ  ๊ทธ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋…ผ์˜๋Š” ๋ถ€๋ก A์—์„œ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์ดˆ๊ธฐ ์ฒดํฌํฌ์ธํŠธ๋กœ ALMA-13B-LoRA๋ฅผ ์‹œ์ž‘์œผ๋กœ _๋‹ค๋Œ€๋‹ค_ ๋‹ค๊ตญ์–ด ๊ธฐ๊ณ„ ๋ฒˆ์—ญ ๋ฐฉ์‹์œผ๋กœ ๋ชจ๋ธ์„ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋‹จ๊ณ„์—์„œ๋Š” ์ถ”๊ฐ€๋œ LoRA ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๊ฐ€์ค‘์น˜๋งŒ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐ ์ง‘์ค‘ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฐ€์ค‘์น˜๋Š” ์ˆœ์œ„๊ฐ€ 16์ด๋ฉฐ ๋ชจ๋ธ์˜ ์›๋ž˜ 13B ํฌ๊ธฐ์— 12M์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋งŒ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Rafailov ๋“ฑ(2023)์ด ์ œ์•ˆํ•œ ๊ธฐ๋ณธ \(\beta\) ๊ฐ’ 0.1์„ ์ค€์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ALMA-13B-LoRA์˜ ๋ฏธ์„ธ ์กฐ์ • ํ”„๋กœ์„ธ์Šค์—๋Š” ๋ฐฐ์น˜ ํฌ๊ธฐ 128, ์›Œ๋ฐ์—… ๋น„์œจ 0.01, ํ•œ ๋ฒˆ์˜ ์—ํฌํฌ์— ๊ฑธ์ณ ์ตœ๋Œ€ ๊ธธ์ด 512 ํ† ํฐ์˜ ์‹œํ€€์Šค๋ฅผ ์ˆ˜์šฉํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ํšจ์œจ์„ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋”ฅ์Šคํ”ผ๋“œ ๋„๊ตฌ(Rasley et al., 2020)๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. Xu ๋“ฑ(2023)๊ณผ ๋™์ผํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ์†์‹ค์„ ๊ณ„์‚ฐํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. 13B ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์— ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์ง€๋งŒ, CPO๋Š” 7B ๋ชจ๋ธ์—๋„ ํ˜„์ €ํ•œ ์ด์ ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ALMA-7B-R๋„ ์ถœ์‹œํ•˜๊ณ  ๊ทธ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋…ผ์˜๋Š” ๋ถ€๋ก A์—์„œ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
Inspired by _Evol-Instruct_ and Process-supervised Reinforcement Learning, this work aims to enhance the mathematical reasoning abilities of the SOTA open-source LLM, Llama-2 [20]. As shown in the Figure 1, we propose a new method named _Reinforcement Learning from Evol-Instruct Feedback_ (_**RLEIF**_), which could firstly generate diverse math instructions data by math-specific _Evol-Instruct_, then we train an instruction reward model (IRM) and a process-supervised reward model (PRM) [36, 37, 38, 39, 40, 41], the former indicates the quality of the evolved instruction and the later receives feedback for each step in the solution. The brand-new _Evol-Instruct_ method includes two downward evolution and upward evolution progress to produce the grade school math and challenging math respectively. Initially, we re-generate, filter and finetune the original math instruction data from GSM8k [42] and MATH [43]. Immediately, we train the Llama-2 models to obtain the reward models and our _WizardMath_.
Evol-Instruct_์™€ Process-supervised Reinforcement Learning์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„, ์ด ์ž‘์—…์€ SOTA ์˜คํ”ˆ ์†Œ์Šค LLM์ธ Llama-2 [20]์˜ ์ˆ˜ํ•™์  ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด, ๋จผ์ € ์ˆ˜ํ•™ ํŠนํ™” _Evol-Instruct_๋ฅผ ํ†ตํ•ด ๋‹ค์–‘ํ•œ ์ˆ˜ํ•™ instructions ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•œ ๋‹ค์Œ instruction reward model (IRM)๊ณผ process-supervised reward model (PRM) [36, 37, 38, 39, 40, 41]์„ ํ›ˆ๋ จํ•˜๋Š” _Reinforcement Learning from Evol-Instruct Feedback_(_**RLEIF**_)๋ผ๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค, ์ „์ž๋Š” ์ง„ํ™”๋œ instruction์˜ ํ’ˆ์งˆ์„ ๋‚˜ํƒ€๋‚ด๊ณ  ํ›„์ž๋Š” ์†”๋ฃจ์…˜์˜ ๊ฐ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด _Evol-Instruct_ ๋ฐฉ๋ฒ•์—๋Š” ๊ฐ๊ฐ ์ดˆ๋“ฑํ•™๊ต ์ˆ˜ํ•™๊ณผ ๋„์ „ ์ˆ˜ํ•™์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐ€์ง€ ํ•˜ํ–ฅ ์ง„ํ™” ๋ฐ ์ƒํ–ฅ ์ง„ํ™” ์ง„ํ–‰์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” GSM8k [42] ๋ฐ MATH [43]์—์„œ ์›๋ž˜์˜ ์ˆ˜ํ•™ instruction ๋ฐ์ดํ„ฐ๋ฅผ ๋‹ค์‹œ ์ƒ์„ฑ, ํ•„ํ„ฐ๋ง ๋ฐ Fine-tuningํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰์‹œ, ์šฐ๋ฆฌ๋Š” reward models๊ณผ ์šฐ๋ฆฌ์˜ _WizardMath_๋ฅผ ์–ป๊ธฐ ์œ„ํ•ด Llama-2 ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.
Evol-Instruct_์™€ ํ”„๋กœ์„ธ์Šค ์Šˆํผ๋ฐ”์ด์ € ๊ฐ•ํ™” ํ•™์Šต์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„, ์ด ์ž‘์—…์€ SOTA ์˜คํ”ˆ ์†Œ์Šค LLM์ธ Llama-2 [20]์˜ ์ˆ˜ํ•™์  ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด, ๋จผ์ € ์ˆ˜ํ•™ ํŠนํ™” _Evol-Instruct_๋ฅผ ํ†ตํ•ด ๋‹ค์–‘ํ•œ ์ˆ˜ํ•™ ์ง€์นจ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•œ ๋‹ค์Œ ์ง€์นจ ๋ณด์ƒ ๋ชจ๋ธ(IRM)๊ณผ ํ”„๋กœ์„ธ์Šค ์Šˆํผ๋ฐ”์ด์ € ๋ณด์ƒ ๋ชจ๋ธ(PRM) [36, 37, 38, 39, 40, 41]์„ ํ›ˆ๋ จํ•˜๋Š” _Reinforcement Learning from Evol-Instruct Feedback_(_**RLEIF**_)๋ผ๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค, ์ „์ž๋Š” ์ง„ํ™”๋œ ์ง€์นจ์˜ ํ’ˆ์งˆ์„ ๋‚˜ํƒ€๋‚ด๊ณ  ํ›„์ž๋Š” ์†”๋ฃจ์…˜์˜ ๊ฐ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด _Evol-Instruct_ ๋ฐฉ๋ฒ•์—๋Š” ๊ฐ๊ฐ ์ดˆ๋“ฑํ•™๊ต ์ˆ˜ํ•™๊ณผ ๋„์ „ ์ˆ˜ํ•™์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐ€์ง€ ํ•˜ํ–ฅ ์ง„ํ™” ๋ฐ ์ƒํ–ฅ ์ง„ํ™” ์ง„ํ–‰์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” GSM8k [42] ๋ฐ MATH [43]์—์„œ ์›๋ž˜์˜ ์ˆ˜ํ•™ ์ง€์นจ ๋ฐ์ดํ„ฐ๋ฅผ ๋‹ค์‹œ ์ƒ์„ฑ, ํ•„ํ„ฐ๋ง ๋ฐ ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰์‹œ, ์šฐ๋ฆฌ๋Š” ๋ฆฌ์›Œ๋“œ ๋ชจ๋ธ๊ณผ ์šฐ๋ฆฌ์˜ _WizardMath_๋ฅผ ์–ป๊ธฐ ์œ„ํ•ด Llama-2 ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.
In this section, we highlight some simple ways to further accelerate the parallel generation algorithm. For example, note that one can reduce the attention's memory cost by computing the diagonal attention simply as elementwise (rather than pairwise) dot-products. That is, given two input embedding sequences of shapes \((b,t,l,d)\) and \((b,1,l,d)\) where \(t\) is the number of timesteps ahead, \(b\) is batch size, \(l\) is sequence length, and \(d\) is embedding dimension, we do not need to compute their pairwise attention of shape \((b,t,l,l)\), we only need to compute the attention for the paired elements along the diagonal of shape \((b,t,l)\). Additionally, to avoid generating continuations for all of the tokens (for example, if one wanted to apply a value function to determine where thoughts would be most useful), one can index into this generated attention mask. Notably, however, this also requires manipulation of the other inputs during the forward pass such as the positional embeddings.
์ด ์„น์…˜์—์„œ๋Š” ๋ณ‘๋ ฌ ์ƒ์„ฑ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๋”์šฑ ๊ฐ€์†ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๋ช‡ ๊ฐ€์ง€ ๊ฐ„๋‹จํ•œ ๋ฐฉ๋ฒ•์„ ์ค‘์ ์ ์œผ๋กœ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋Œ€๊ฐ์„  attention์„ ์Œ๋ณ„(์Œ๋ณ„) dot-products์ด ์•„๋‹Œ ์š”์†Œ๋ณ„ dot-products์œผ๋กœ ๊ณ„์‚ฐํ•˜์—ฌ attention์˜ ๋ฉ”๋ชจ๋ฆฌ ๋น„์šฉ์„ ์ค„์ผ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”. ์ฆ‰, \((b,t,l,d)\) ๋ฐ \((b,1,l,d)\) ๋ชจ์–‘์˜ ๋‘ ์ž…๋ ฅ embedding ์‹œํ€€์Šค๊ฐ€ ์ฃผ์–ด์ง€๋ฉด, ์—ฌ๊ธฐ์„œ \(t\)๋Š” ์•ž์œผ๋กœ์˜ ์‹œ๊ฐ„ ๊ฐ„๊ฒฉ ์ˆ˜, \(b\)๋Š” batch size, \(l\)์€ ์‹œํ€€์Šค ๊ธธ์ด, \(d\)๋Š” embedding ์ฐจ์›์ž…๋‹ˆ๋‹ค, ์šฐ๋ฆฌ๋Š” \((b,t,l,l)\) ๋ชจ์–‘์˜ ์Œ๋ณ„ attention์„ ๊ณ„์‚ฐํ•  ํ•„์š” ์—†์ด \((b,t,l)\) ๋ชจ์–‘์˜ ๋Œ€๊ฐ์„ ์„ ๋”ฐ๋ผ ์ง์„ ์ด๋ฃฌ ์š”์†Œ์— ๋Œ€ํ•œ attention๋งŒ ๊ณ„์‚ฐํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋“  ํ† ํฐ์— ๋Œ€ํ•ด ์—ฐ์†์„ ์ƒ์„ฑํ•˜์ง€ ์•Š์œผ๋ ค๋ฉด(์˜ˆ๋ฅผ ๋“ค์–ด, ์–ด๋””์—์„œ ์ƒ๊ฐ์ด ๊ฐ€์žฅ ์œ ์šฉํ• ์ง€ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด value function์„ ์ ์šฉํ•˜๋ ค๋Š” ๊ฒฝ์šฐ) ์ด๋ ‡๊ฒŒ ์ƒ์„ฑ๋œ attention mask์— ์ธ๋ฑ์‹ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋ฅผ ์œ„ํ•ด์„œ๋Š” positional embeddings๊ณผ ๊ฐ™์€ forward pass ์ค‘ ๋‹ค๋ฅธ ์ž…๋ ฅ์„ ์กฐ์ž‘ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋ณ‘๋ ฌ ์ƒ์„ฑ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๋”์šฑ ๊ฐ€์†ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๋ช‡ ๊ฐ€์ง€ ๊ฐ„๋‹จํ•œ ๋ฐฉ๋ฒ•์„ ์ค‘์ ์ ์œผ๋กœ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋Œ€๊ฐ์„  ์ฃผ์˜๋ฅผ ์Œ๋ณ„(์Œ๋ณ„) ๋„ํŠธ ๊ณฑ์ด ์•„๋‹Œ ์š”์†Œ๋ณ„ ๋„ํŠธ ๊ณฑ์œผ๋กœ ๊ณ„์‚ฐํ•˜์—ฌ ์ฃผ์˜์˜ ๋ฉ”๋ชจ๋ฆฌ ๋น„์šฉ์„ ์ค„์ผ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”. ์ฆ‰, (b,t,l,d)\) ๋ฐ \((b,1,l,d)\) ๋ชจ์–‘์˜ ๋‘ ์ž…๋ ฅ ์ž„๋ฒ ๋”ฉ ์‹œํ€€์Šค๊ฐ€ ์ฃผ์–ด์ง€๋ฉด, ์—ฌ๊ธฐ์„œ \(t\)๋Š” ์•ž์œผ๋กœ์˜ ์‹œ๊ฐ„ ๊ฐ„๊ฒฉ ์ˆ˜, \(b\)๋Š” ๋ฐฐ์น˜ ํฌ๊ธฐ, \(l\)์€ ์‹œํ€€์Šค ๊ธธ์ด, \(d\)๋Š” ์ž„๋ฒ ๋”ฉ ์ฐจ์›์ž…๋‹ˆ๋‹ค, ์šฐ๋ฆฌ๋Š” (b,t,l,l)\) ๋ชจ์–‘์˜ ์Œ๋ณ„ ์ฃผ์˜๋„๋ฅผ ๊ณ„์‚ฐํ•  ํ•„์š” ์—†์ด (b,t,l)\) ๋ชจ์–‘์˜ ๋Œ€๊ฐ์„ ์„ ๋”ฐ๋ผ ์ง์„ ์ด๋ฃฌ ์š”์†Œ์— ๋Œ€ํ•œ ์ฃผ์˜๋„๋งŒ ๊ณ„์‚ฐํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋“  ํ† ํฐ์— ๋Œ€ํ•ด ์—ฐ์†์„ ์ƒ์„ฑํ•˜์ง€ ์•Š์œผ๋ ค๋ฉด(์˜ˆ๋ฅผ ๋“ค์–ด, ์–ด๋””์—์„œ ์ƒ๊ฐ์ด ๊ฐ€์žฅ ์œ ์šฉํ• ์ง€ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ’ ํ•จ์ˆ˜๋ฅผ ์ ์šฉํ•˜๋ ค๋Š” ๊ฒฝ์šฐ) ์ด๋ ‡๊ฒŒ ์ƒ์„ฑ๋œ ์ฃผ์˜ ๋งˆ์Šคํฌ์— ์ธ๋ฑ์‹ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋ฅผ ์œ„ํ•ด์„œ๋Š” ํฌ์ง€์…”๋„ ์ž„๋ฒ ๋”ฉ๊ณผ ๊ฐ™์€ ํฌ์›Œ๋“œ ํŒจ์Šค ์ค‘ ๋‹ค๋ฅธ ์ž…๋ ฅ์„ ์กฐ์ž‘ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
**Empirical Analysis on Quantization Experiments**. To further help readers understand the impact of quantization on LLMs, we also conduct a group of experiments to investigate the inference performance of quantized models here. Specifically, we focus on the fine-tuned LLMaM models (_i.e.,_ 7B and 13B) using popular SFT datasets, including FLAN-v2 [69], Alpaca-52K [137] and ShareGPT [148]. For evaluation, we utilize the same tasks in Table IX, and follow the quantization settings in the study [423] examining the performance of quantized language models at three precision levels: 4-bit, 8-bit and 16-bit. The results are summarized in Table X. As can be observed from Table X, the results obtained with 8-bit and 4-bit weight quantization are close to the performance of 16-bit models while significantly reducing memory consumption. In practice, it is recommended to first examine the performance of 4-bit weight quantization for LLMs if reducing memory usage is a critical consideration for deployment.
**Quantization ์‹คํ—˜์— ๋Œ€ํ•œ ๊ฒฝํ—˜์  ๋ถ„์„**. ๋…์ž๋“ค์ด Quantization์ด LLMs์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋” ์ž˜ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋„๋ก, ์—ฌ๊ธฐ์„œ๋Š” Quantized ๋ชจ๋ธ์˜ ์ถ”๋ก  ์„ฑ๋Šฅ์„ ์กฐ์‚ฌํ•˜๋Š” ์ผ๋ จ์˜ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, FLAN-v2 [69], Alpaca-52K [137] ๋ฐ ShareGPT [148]๋ฅผ ๋น„๋กฏํ•œ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” SFT ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Fine-tuning๋œ LLMaM ๋ชจ๋ธ(_์ฆ‰_, 7B ๋ฐ 13B)์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ํ‘œ IX์—์„œ ๋™์ผํ•œ ์ž‘์—…์„ ํ™œ์šฉํ•˜๊ณ , 3๊ฐœ์˜ ์ •๋ฐ€๋„ ์ˆ˜์ค€์—์„œ Quantized ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์กฐ์‚ฌํ•˜๋Š” ์—ฐ๊ตฌ [423]์˜ Quantization ์„ค์ •์„ ๋”ฐ๋ฅด๋ฉฐ, 4-bit, 8-bit, 16-bit์˜ ์„ธ ๊ฐ€์ง€ ์ •๋ฐ€๋„ ์ˆ˜์ค€์—์„œ Quantized ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ๋Š” ํ‘œ X์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ‘œ X์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, 8-bit ๋ฐ 4-bit Weight Quantization์„ ์‚ฌ์šฉํ•œ ๊ฒฐ๊ณผ๋Š” 16-bit ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์— ๊ทผ์ ‘ํ•œ ๋ฐ˜๋ฉด ๋ฉ”๋ชจ๋ฆฌ ์†Œ๋น„๋Š” ํฌ๊ฒŒ ์ค„์˜€์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ๋ฐฐํฌ ์‹œ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰์„ ์ค„์ด๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•œ ๊ณ ๋ ค ์‚ฌํ•ญ์ธ ๊ฒฝ์šฐ, LLMs์— ๋Œ€ํ•ด 4-bit Weight Quantization์˜ ์„ฑ๋Šฅ์„ ๋จผ์ € ๊ฒ€ํ† ํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.
**์–‘์žํ™” ์‹คํ—˜์— ๋Œ€ํ•œ ๊ฒฝํ—˜์  ๋ถ„์„**. ๋…์ž๋“ค์ด ์–‘์žํ™”๊ฐ€ LLM์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋” ์ž˜ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋„๋ก, ์—ฌ๊ธฐ์„œ๋Š” ์–‘์žํ™”๋œ ๋ชจ๋ธ์˜ ์ถ”๋ก  ์„ฑ๋Šฅ์„ ์กฐ์‚ฌํ•˜๋Š” ์ผ๋ จ์˜ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, FLAN-v2 [69], Alpaca-52K [137] ๋ฐ ShareGPT [148]๋ฅผ ๋น„๋กฏํ•œ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” SFT ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏธ์„ธ ์กฐ์ •๋œ LLMaM ๋ชจ๋ธ(_์ฆ‰_, 7B ๋ฐ 13B)์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ํ‘œ IX์—์„œ ๋™์ผํ•œ ์ž‘์—…์„ ํ™œ์šฉํ•˜๊ณ , 3๊ฐœ์˜ ์ •๋ฐ€๋„ ์ˆ˜์ค€์—์„œ ์–‘์žํ™”๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์กฐ์‚ฌํ•˜๋Š” ์—ฐ๊ตฌ [423]์˜ ์–‘์žํ™” ์„ค์ •์„ ๋”ฐ๋ฅด๋ฉฐ, 4๋น„ํŠธ, 8๋น„ํŠธ, 16๋น„ํŠธ์˜ ์„ธ ๊ฐ€์ง€ ์ •๋ฐ€๋„ ์ˆ˜์ค€์—์„œ ์–‘์žํ™”๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ๋Š” ํ‘œ X์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ‘œ X์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, 8๋น„ํŠธ ๋ฐ 4๋น„ํŠธ ๊ฐ€์ค‘์น˜ ์–‘์žํ™”๋ฅผ ์‚ฌ์šฉํ•œ ๊ฒฐ๊ณผ๋Š” 16๋น„ํŠธ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์— ๊ทผ์ ‘ํ•œ ๋ฐ˜๋ฉด ๋ฉ”๋ชจ๋ฆฌ ์†Œ๋น„๋Š” ํฌ๊ฒŒ ์ค„์˜€์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ๋ฐฐํฌ ์‹œ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰์„ ์ค„์ด๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•œ ๊ณ ๋ ค ์‚ฌํ•ญ์ธ ๊ฒฝ์šฐ, LLM์— ๋Œ€ํ•ด 4๋น„ํŠธ ๊ฐ€์ค‘์น˜ ์–‘์žํ™”์˜ ์„ฑ๋Šฅ์„ ๋จผ์ € ๊ฒ€ํ† ํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Sampling-based methods._ LLMs are known to suffer from instability during inference, which can lead to unfaithfulness in the generated reasoning steps. To address this issue, some work proposes to sample multiple reasoning paths instead of using greedy decoding. As a representative solution, self-consistency [436] first generates several reasoning paths and then takes an ensemble over the corresponding answers, selecting the most consistent one through majority voting. However, such a method can still lead to wrong answers when most of the reasoning paths are misled. Considering this, the authors in [433] only vote on the \(k\) most complex reasoning paths based on their observation that reasoning paths with higher complexity (_e.g._, more reasoning steps) usually have better performance. Furthermore, MCR [510] proposes referring to the steps from other reasoning paths when generating the next step, and performs reasoning across multiple reasoning paths to generate the final answer.
\(\bullet\)_์ƒ˜ํ”Œ๋ง ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•._ LLM์€ ์ถ”๋ก  ์ค‘ ๋ถˆ์•ˆ์ •์„ฑ์„ ๊ฒช๋Š” ๊ฒƒ์œผ๋กœ ์•Œ๋ ค์ ธ ์žˆ์œผ๋ฉฐ, ์ด๋กœ ์ธํ•ด ์ƒ์„ฑ๋œ ์ถ”๋ก  ๋‹จ๊ณ„์— ๋ถˆ์ถฉ์‹ค์„ฑ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” greedy decoding์„ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์ƒ˜ํ”Œ๋งํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ์†”๋ฃจ์…˜์œผ๋กœ, self-consistency[436]๋Š” ๋จผ์ € ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์ƒ์„ฑํ•œ ๋‹ค์Œ ํ•ด๋‹น ๋‹ต์•ˆ์— ๋Œ€ํ•ด ์•™์ƒ๋ธ”์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ๊ณผ๋ฐ˜์ˆ˜ ํˆฌํ‘œ๋ฅผ ํ†ตํ•ด ๊ฐ€์žฅ ์ผ๊ด€์„ฑ์ด ๋†’์€ ๋‹ต์•ˆ์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์€ ๋Œ€๋ถ€๋ถ„์˜ ์ถ”๋ก  ๊ฒฝ๋กœ๊ฐ€ ์ž˜๋ชป๋œ ๊ฒฝ์šฐ ์—ฌ์ „ํžˆ ์˜ค๋‹ต์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๊ณ ๋ คํ•˜์—ฌ [433]์˜ ์ €์ž๋“ค์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ณต์žก๋„๊ฐ€ ๋†’์€ ์ถ”๋ก  ๊ฒฝ๋กœ(_์˜ˆ: ์ถ”๋ก  ๋‹จ๊ณ„๊ฐ€ ๋” ๋งŽ์Œ)๊ฐ€ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ธ๋‹ค๋Š” ๊ด€์ฐฐ์„ ๋ฐ”ํƒ•์œผ๋กœ ๊ฐ€์žฅ ๋ณต์žกํ•œ ์ถ”๋ก  ๊ฒฝ๋กœ \(k\)์— ๋Œ€ํ•ด์„œ๋งŒ ํˆฌํ‘œํ•˜๋„๋ก ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, MCR[510]์€ ๋‹ค์Œ ๋‹จ๊ณ„๋ฅผ ์ƒ์„ฑํ•  ๋•Œ ๋‹ค๋ฅธ ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ๋‹จ๊ณ„๋ฅผ ์ฐธ์กฐํ•˜๊ณ  ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ์— ๊ฑธ์ณ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ตœ์ข… ๋‹ต์„ ์ƒ์„ฑํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_์ƒ˜ํ”Œ๋ง ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•._ LLM์€ ์ถ”๋ก  ์ค‘ ๋ถˆ์•ˆ์ •์„ฑ์„ ๊ฒช๋Š” ๊ฒƒ์œผ๋กœ ์•Œ๋ ค์ ธ ์žˆ์œผ๋ฉฐ, ์ด๋กœ ์ธํ•ด ์ƒ์„ฑ๋œ ์ถ”๋ก  ๋‹จ๊ณ„์— ๋ถˆ์ถฉ์‹ค์„ฑ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ์š•์‹ฌ์Šค๋Ÿฌ์šด ๋””์ฝ”๋”ฉ์„ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์ƒ˜ํ”Œ๋งํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ์†”๋ฃจ์…˜์œผ๋กœ, ์ž๊ธฐ ์ผ๊ด€์„ฑ[436]์€ ๋จผ์ € ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์ƒ์„ฑํ•œ ๋‹ค์Œ ํ•ด๋‹น ๋‹ต์•ˆ์— ๋Œ€ํ•ด ์•™์ƒ๋ธ”์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ๊ณผ๋ฐ˜์ˆ˜ ํˆฌํ‘œ๋ฅผ ํ†ตํ•ด ๊ฐ€์žฅ ์ผ๊ด€์„ฑ์ด ๋†’์€ ๋‹ต์•ˆ์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์€ ๋Œ€๋ถ€๋ถ„์˜ ์ถ”๋ก  ๊ฒฝ๋กœ๊ฐ€ ์ž˜๋ชป๋œ ๊ฒฝ์šฐ ์—ฌ์ „ํžˆ ์˜ค๋‹ต์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๊ณ ๋ คํ•˜์—ฌ [433]์˜ ์ €์ž๋“ค์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ณต์žก๋„๊ฐ€ ๋†’์€ ์ถ”๋ก  ๊ฒฝ๋กœ(_์˜ˆ: ์ถ”๋ก  ๋‹จ๊ณ„๊ฐ€ ๋” ๋งŽ์Œ)๊ฐ€ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ธ๋‹ค๋Š” ๊ด€์ฐฐ์„ ๋ฐ”ํƒ•์œผ๋กœ ๊ฐ€์žฅ ๋ณต์žกํ•œ ์ถ”๋ก  ๊ฒฝ๋กœ \(k\)์— ๋Œ€ํ•ด์„œ๋งŒ ํˆฌํ‘œํ•˜๋„๋ก ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, MCR [510]์€ ๋‹ค์Œ ๋‹จ๊ณ„๋ฅผ ์ƒ์„ฑํ•  ๋•Œ ๋‹ค๋ฅธ ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ๋‹จ๊ณ„๋ฅผ ์ฐธ์กฐํ•˜๊ณ  ์—ฌ๋Ÿฌ ์ถ”๋ก  ๊ฒฝ๋กœ์— ๊ฑธ์ณ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ตœ์ข… ๋‹ต์„ ์ƒ์„ฑํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_Decoding strategy optimization_ aims to improve the sequential nature of the auto-regressive generation manner in different ways. As a representative study, _speculative decoding_[323, 324] first leverages a compact but efficient model (_e.g._, a \(n\)-gram model or a small PLM) to generate short segments and then utilizes the LLM to verify and correct these drafts. It can lead to a notable \(2\times\) to \(3\times\) speedup without compromising the generation quality. Researchers further suggest several variants to improve the efficiency of this approach, such as a learning-based method to combine several small models [325] and a stage-wise acceleration which employs a more smaller LM to accelerate the small LM first [326]. In addition, token-level early-exit techniques have been proposed enabling the generation of a token at lower Transformer layers, rather than passing through all the layers [327]. It can attain greater speedup, but at the cost of sacrificing generation quality.
\(\bullet\)_Decoding strategy optimization_๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ auto-regressive generation ๋ฐฉ์‹์˜ ์ˆœ์ฐจ์  ํŠน์„ฑ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ์—ฐ๊ตฌ๋กœ, _speculative decoding_[323, 324]์€ ๋จผ์ € ์ปดํŒฉํŠธํ•˜์ง€๋งŒ ํšจ์œจ์ ์ธ ๋ชจ๋ธ(_์˜ˆ: n-gram ๋ชจ๋ธ ๋˜๋Š” ์†Œํ˜• PLM)์„ ํ™œ์šฉํ•˜์—ฌ ์งง์€ ์„ธ๊ทธ๋จผํŠธ๋ฅผ ์ƒ์„ฑํ•œ ๋‹ค์Œ LLM์„ ํ™œ์šฉํ•˜์—ฌ ์ด๋Ÿฌํ•œ ์ดˆ์•ˆ์„ ๊ฒ€์ฆํ•˜๊ณ  ์ˆ˜์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ƒ์„ฑ ํ’ˆ์งˆ ์ €ํ•˜ ์—†์ด 2๋ฐฐ์—์„œ 3๋ฐฐ๊นŒ์ง€ ์†๋„๋ฅผ ๋†’์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ž๋“ค์€ ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์†Œํ˜• ๋ชจ๋ธ์„ ๊ฒฐํ•ฉํ•˜๋Š” ํ•™์Šต ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•[325]๊ณผ ๋‹จ๊ณ„๋ณ„ ๊ฐ€์†ํ™” ๋ฐฉ๋ฒ•[326] ๋“ฑ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ๋ณ€ํ˜•์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ token-level early-exit ๊ธฐ์ˆ ์ด ์ œ์•ˆ๋˜์–ด ๋ชจ๋“  Transformer ๋ ˆ์ด์–ด๋ฅผ ํ†ต๊ณผํ•˜๋Š” ๋Œ€์‹  ๋” ๋‚ฎ์€ Transformer ๋ ˆ์ด์–ด์—์„œ ํ† ํฐ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[327]. ์ด ๊ธฐ์ˆ ์€ ๋” ํฐ ์†๋„ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ์ƒ์„ฑ ํ’ˆ์งˆ์„ ํฌ์ƒํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_๋””์ฝ”๋”ฉ ์ „๋žต ์ตœ์ ํ™”_๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ ์ž๋™ ํšŒ๊ท€ ์ƒ์„ฑ ๋ฐฉ์‹์˜ ์ˆœ์ฐจ์  ํŠน์„ฑ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ์—ฐ๊ตฌ๋กœ, _์‚ฌ์ธก ๋””์ฝ”๋”ฉ_[323, 324]์€ ๋จผ์ € ์ปดํŒฉํŠธํ•˜์ง€๋งŒ ํšจ์œจ์ ์ธ ๋ชจ๋ธ(_์˜ˆ: n\๊ทธ๋žจ ๋ชจ๋ธ ๋˜๋Š” ์†Œํ˜• PLM)์„ ํ™œ์šฉํ•˜์—ฌ ์งง์€ ์„ธ๊ทธ๋จผํŠธ๋ฅผ ์ƒ์„ฑํ•œ ๋‹ค์Œ LLM์„ ํ™œ์šฉํ•˜์—ฌ ์ด๋Ÿฌํ•œ ์ดˆ์•ˆ์„ ๊ฒ€์ฆํ•˜๊ณ  ์ˆ˜์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ƒ์„ฑ ํ’ˆ์งˆ ์ €ํ•˜ ์—†์ด 2๋ฐฐ์—์„œ 3๋ฐฐ๊นŒ์ง€ ์†๋„๋ฅผ ๋†’์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ž๋“ค์€ ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์†Œํ˜• ๋ชจ๋ธ์„ ๊ฒฐํ•ฉํ•˜๋Š” ํ•™์Šต ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•[325]๊ณผ ๋‹จ๊ณ„๋ณ„ ๊ฐ€์†ํ™” ๋ฐฉ๋ฒ•[326] ๋“ฑ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ๋ณ€ํ˜•์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํ† ํฐ ์ˆ˜์ค€ ์กฐ๊ธฐ ์ข…๋ฃŒ ๊ธฐ์ˆ ์ด ์ œ์•ˆ๋˜์–ด ๋ชจ๋“  ๋ ˆ์ด์–ด๋ฅผ ํ†ต๊ณผํ•˜๋Š” ๋Œ€์‹  ๋” ๋‚ฎ์€ ํŠธ๋žœ์Šคํฌ๋จธ ๋ ˆ์ด์–ด์—์„œ ํ† ํฐ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[327]. ์ด ๊ธฐ์ˆ ์€ ๋” ํฐ ์†๋„ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ์ƒ์„ฑ ํ’ˆ์งˆ์„ ํฌ์ƒํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Math tasks. This is expected as \(\mathtt{Exam}\) contains more math quizzes and exam types(e.g., reading comprehension), potentially boosting the model's performance in most tasks. Interestingly, \(\mathtt{Ruozhiba}\) ranks second on average across all subsets. We conjecture this is because it may enhance the model's logical reasoning ability, thereby benefiting most of the instruct-following tasks. \(\mathtt{COIG-PC}\) demonstrates proficiency in evaluations of the knowledge dimension, such as C-Eval, yet underperforms in Belle-Eval. We attribute this discrepancy to its origin in traditional NLP datasets and the short length of responses, which can impair reasoning tasks and are less favored by model-based evaluators. The substantial gap between C-Eval and Belle-Eval highlights the importance of developing assessments that can comprehensively and accurately evaluate Chinese LLMs. Moreover, WikiHow scores only 5.8, which we believe is due to the lack of diversity in its "how-to" instructions.
์ˆ˜ํ•™ ๊ณผ์ œ. ์ด๋Š” \(\mathtt{Exam}\)์— ์ˆ˜ํ•™ ํ€ด์ฆˆ์™€ ์‹œํ—˜ ์œ ํ˜•(์˜ˆ: ๋…ํ•ด๋ ฅ)์ด ๋” ๋งŽ์ด ํฌํ•จ๋˜์–ด ์žˆ์–ด ๋Œ€๋ถ€๋ถ„์˜ ๊ณผ์ œ์—์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์˜ˆ์ƒ๋˜๋Š” ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ \(\mathtt{Ruozhiba}\)๋Š” ๋ชจ๋“  ํ•˜์œ„ ์ง‘ํ•ฉ์—์„œ ํ‰๊ท ์ ์œผ๋กœ 2์œ„๋ฅผ ์ฐจ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ชจ๋ธ์˜ ๋…ผ๋ฆฌ์  ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œ์ผœ ๋Œ€๋ถ€๋ถ„์˜ Instruction ์ˆ˜ํ–‰ ๊ณผ์ œ์— ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ถ”์ธกํ•ฉ๋‹ˆ๋‹ค. \(\mathtt{COIG-PC}\)๋Š” C-Eval๊ณผ ๊ฐ™์€ ์ง€์‹ ์ฐจ์›์˜ ํ‰๊ฐ€์—๋Š” ๋Šฅ์ˆ™ํ•˜์ง€๋งŒ Belle-Eval์—์„œ๋Š” ์„ฑ๋Šฅ์ด ์ €์กฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฐจ์ด๋Š” ์ „ํ†ต์ ์ธ NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์œ ๋ž˜ํ–ˆ์œผ๋ฉฐ ์‘๋‹ต ๊ธธ์ด๊ฐ€ ์งง์•„ ์ถ”๋ก  ๊ณผ์ œ๋ฅผ ๋ฐฉํ•ดํ•˜๊ณ  ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ํ‰๊ฐ€์ž๊ฐ€ ์„ ํ˜ธํ•˜์ง€ ์•Š๋Š” ๊ฒƒ์— ๊ธฐ์ธํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. C-Eval๊ณผ Belle-Eval ์‚ฌ์ด์˜ ์ƒ๋‹นํ•œ ๊ฒฉ์ฐจ๋Š” ์ค‘๊ตญ Large Language Models (LLMs)์„ ํฌ๊ด„์ ์ด๊ณ  ์ •ํ™•ํ•˜๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š” ํ‰๊ฐ€ ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์–ผ๋งˆ๋‚˜ ์ค‘์š”ํ•œ์ง€๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ WikiHow์˜ ์ ์ˆ˜๋Š” 5.8์ ์— ๋ถˆ๊ณผํ•œ๋ฐ, ์ด๋Š” '๋ฐฉ๋ฒ•' ์ง€์นจ์˜ ๋‹ค์–‘์„ฑ์ด ๋ถ€์กฑํ•˜๊ธฐ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
์ˆ˜ํ•™ ๊ณผ์ œ. ์ด๋Š” \(\mathtt{Exam}\)์— ์ˆ˜ํ•™ ํ€ด์ฆˆ์™€ ์‹œํ—˜ ์œ ํ˜•(์˜ˆ: ๋…ํ•ด๋ ฅ)์ด ๋” ๋งŽ์ด ํฌํ•จ๋˜์–ด ์žˆ์–ด ๋Œ€๋ถ€๋ถ„์˜ ๊ณผ์ œ์—์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์˜ˆ์ƒ๋˜๋Š” ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ \(\mathtt{Ruozhiba}\)๋Š” ๋ชจ๋“  ํ•˜์œ„ ์ง‘ํ•ฉ์—์„œ ํ‰๊ท ์ ์œผ๋กœ 2์œ„๋ฅผ ์ฐจ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ชจ๋ธ์˜ ๋…ผ๋ฆฌ์  ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œ์ผœ ๋Œ€๋ถ€๋ถ„์˜ ์ง€์‹œ์–ด ์ˆ˜ํ–‰ ๊ณผ์ œ์— ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ถ”์ธกํ•ฉ๋‹ˆ๋‹ค. \(\mathtt{COIG-PC}\)๋Š” C-Eval๊ณผ ๊ฐ™์€ ์ง€์‹ ์ฐจ์›์˜ ํ‰๊ฐ€์—๋Š” ๋Šฅ์ˆ™ํ•˜์ง€๋งŒ Belle-Eval์—์„œ๋Š” ์„ฑ๋Šฅ์ด ์ €์กฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฐจ์ด๋Š” ์ „ํ†ต์ ์ธ NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์œ ๋ž˜ํ–ˆ์œผ๋ฉฐ ์‘๋‹ต ๊ธธ์ด๊ฐ€ ์งง์•„ ์ถ”๋ก  ๊ณผ์ œ๋ฅผ ๋ฐฉํ•ดํ•˜๊ณ  ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ํ‰๊ฐ€์ž๊ฐ€ ์„ ํ˜ธํ•˜์ง€ ์•Š๋Š” ๊ฒƒ์— ๊ธฐ์ธํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. C-Eval๊ณผ Belle-Eval ์‚ฌ์ด์˜ ์ƒ๋‹นํ•œ ๊ฒฉ์ฐจ๋Š” ์ค‘๊ตญ LLM์„ ํฌ๊ด„์ ์ด๊ณ  ์ •ํ™•ํ•˜๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š” ํ‰๊ฐ€ ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์–ผ๋งˆ๋‚˜ ์ค‘์š”ํ•œ์ง€๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์œ„ํ‚คํ•˜์šฐ์˜ ์ ์ˆ˜๋Š” 5.8์ ์— ๋ถˆ๊ณผํ•œ๋ฐ, ์ด๋Š” '๋ฐฉ๋ฒ•' ์ง€์นจ์˜ ๋‹ค์–‘์„ฑ์ด ๋ถ€์กฑํ•˜๊ธฐ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
Most commercial LLMs, like ChatGPT (OpenAI, 2023), Gemini (Gemini Team, 2023), and Perplexity.ai already employ some version of RAG in their Web interfaces. For example, ChatGPT employs a Bing search whereas Gemini accesses Google Search results. Though RAG has quickly become a default feature of user-facing LLM systems, most evaluations of LLM capabilities are still performed on the non-RAG counterparts (Zheng et al., 2023). This is problematic, as a model's default and RAG-enabled responses can drastically diverge depending on the quality and accuracy of the retrieved content. This problem is compounded when considering that web results constantly change, and can contain outdated, incorrect, or harmful information (Dash et al., 2023; Daws, 2020; Nastasi et al., 2023). Thus, objective evaluations of RAG-enabled LLM behavior are as important as benchmarking their non-RAG counterparts, especially as RAG systems are increasingly relied upon to provide factual information in a myriad of domains.
ChatGPT(OpenAI, 2023), Gemini(Gemini Team, 2023), Perplexity.ai ๋“ฑ ๋Œ€๋ถ€๋ถ„์˜ ์ƒ์šฉ LLM์€ ์›น ์ธํ„ฐํŽ˜์ด์Šค์—์„œ ์ด๋ฏธ ์ผ๋ถ€ ๋ฒ„์ „์˜ Retrieval-augmented generation (RAG)๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ChatGPT๋Š” Bing ๊ฒ€์ƒ‰์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐ˜๋ฉด Gemini๋Š” Google ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์•ก์„ธ์Šคํ•ฉ๋‹ˆ๋‹ค. RAG๋Š” ์‚ฌ์šฉ์ž ๋Œ€๋ฉด LLM ์‹œ์Šคํ…œ์˜ ๊ธฐ๋ณธ ๊ธฐ๋Šฅ์œผ๋กœ ๋น ๋ฅด๊ฒŒ ์ž๋ฆฌ ์žก์•˜์ง€๋งŒ, ๋Œ€๋ถ€๋ถ„์˜ LLM ๊ธฐ๋Šฅ ํ‰๊ฐ€๋Š” ์—ฌ์ „ํžˆ ๋น„-RAG์— ๋Œ€ํ•œ ํ‰๊ฐ€๊ฐ€ ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Zheng et al., 2023). ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ์˜ ํ’ˆ์งˆ๊ณผ ์ •ํ™•๋„์— ๋”ฐ๋ผ ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ๋ฐ RAG ์ง€์› ์‘๋‹ต์ด ํฌ๊ฒŒ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ด๋Š” ๋ฌธ์ œ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋Š” ์›น ๊ฒฐ๊ณผ๊ฐ€ ์ง€์†์ ์œผ๋กœ ๋ณ€๊ฒฝ๋˜๊ณ  ์˜ค๋ž˜๋œ, ์ž˜๋ชป๋œ ๋˜๋Š” ์œ ํ•ดํ•œ ์ •๋ณด๋ฅผ ํฌํ•จํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•  ๋•Œ ๋”์šฑ ๋ณต์žกํ•ด์ง‘๋‹ˆ๋‹ค(Dash et al., 2023; Daws, 2020; Nastasi et al., 2023). ๋”ฐ๋ผ์„œ ๋‹ค์–‘ํ•œ ์˜์—ญ์—์„œ ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด RAG ์ง€์› LLM ์‹œ์Šคํ…œ์— ์ ์  ๋” ๋งŽ์ด ์˜์กดํ•˜๊ณ  ์žˆ๋Š” ์ƒํ™ฉ์—์„œ ๋น„-RAG ์ง€์› LLM์„ ๋ฒค์น˜๋งˆํ‚นํ•˜๋Š” ๊ฒƒ๋งŒํผ์ด๋‚˜ RAG ์ง€์› LLM ๋™์ž‘์— ๋Œ€ํ•œ ๊ฐ๊ด€์ ์ธ ํ‰๊ฐ€๊ฐ€ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
ChatGPT(OpenAI, 2023), Gemini(Gemini Team, 2023), Perplexity.ai ๋“ฑ ๋Œ€๋ถ€๋ถ„์˜ ์ƒ์šฉ LLM์€ ์›น ์ธํ„ฐํŽ˜์ด์Šค์—์„œ ์ด๋ฏธ ์ผ๋ถ€ ๋ฒ„์ „์˜ RAG๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ChatGPT๋Š” Bing ๊ฒ€์ƒ‰์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐ˜๋ฉด Gemini๋Š” Google ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์•ก์„ธ์Šคํ•ฉ๋‹ˆ๋‹ค. RAG๋Š” ์‚ฌ์šฉ์ž ๋Œ€๋ฉด LLM ์‹œ์Šคํ…œ์˜ ๊ธฐ๋ณธ ๊ธฐ๋Šฅ์œผ๋กœ ๋น ๋ฅด๊ฒŒ ์ž๋ฆฌ ์žก์•˜์ง€๋งŒ, ๋Œ€๋ถ€๋ถ„์˜ LLM ๊ธฐ๋Šฅ ํ‰๊ฐ€๋Š” ์—ฌ์ „ํžˆ ๋น„-RAG์— ๋Œ€ํ•œ ํ‰๊ฐ€๊ฐ€ ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Zheng et al., 2023). ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ์˜ ํ’ˆ์งˆ๊ณผ ์ •ํ™•๋„์— ๋”ฐ๋ผ ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ๋ฐ RAG ์ง€์› ์‘๋‹ต์ด ํฌ๊ฒŒ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ด๋Š” ๋ฌธ์ œ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋Š” ์›น ๊ฒฐ๊ณผ๊ฐ€ ์ง€์†์ ์œผ๋กœ ๋ณ€๊ฒฝ๋˜๊ณ  ์˜ค๋ž˜๋œ, ์ž˜๋ชป๋œ ๋˜๋Š” ์œ ํ•ดํ•œ ์ •๋ณด๋ฅผ ํฌํ•จํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•  ๋•Œ ๋”์šฑ ๋ณต์žกํ•ด์ง‘๋‹ˆ๋‹ค(Dash et al., 2023; Daws, 2020; Nastasi et al., 2023). ๋”ฐ๋ผ์„œ ๋‹ค์–‘ํ•œ ์˜์—ญ์—์„œ ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด RAG ์ง€์› LLM ์‹œ์Šคํ…œ์— ์ ์  ๋” ๋งŽ์ด ์˜์กดํ•˜๊ณ  ์žˆ๋Š” ์ƒํ™ฉ์—์„œ ๋น„-RAG ์ง€์› LLM์„ ๋ฒค์น˜๋งˆํ‚นํ•˜๋Š” ๊ฒƒ๋งŒํผ์ด๋‚˜ RAG ์ง€์› LLM ๋™์ž‘์— ๋Œ€ํ•œ ๊ฐ๊ด€์ ์ธ ํ‰๊ฐ€๊ฐ€ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
We first study how the performance varies with different numbers of epochs. As shown in Tab. 2, training for 1 epoch is insufficient, and the performance of 3, 5, or 10 epochs is similar. We fix the number of epochs to 3 and arrange the order of QA pairs and corresponding documents as shown in Fig. 6. The interleaved arrangement cycles through all the data 3 times, ensuring that in each epoch, questions either precede or follow their associated documents. On the other hand, the grouped arrangement clusters each example's 3 appearances together, guaranteeing that the repeated questions are positioned either before or after their respective repeated documents. As shown in Tab. 2, positioning QA pairs before corresponding documents achieves better performance in both grouped and interleaved arrangements, indicating that during PIT, the learning mechanism prioritizes understanding how to access knowledge before learning to absorb information from the more complex and information-dense documents.
๋จผ์ € ๋‹ค์–‘ํ•œ epoch ์ˆ˜์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด ์–ด๋–ป๊ฒŒ ๋‹ฌ๋ผ์ง€๋Š”์ง€ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ํ‘œ 2์™€ ๊ฐ™์ด 1ํšŒ ํ•™์Šต์€ ๋ถˆ์ถฉ๋ถ„ํ•˜๋ฉฐ 3, 5 ๋˜๋Š” 10ํšŒ ํ•™์Šต์˜ ์„ฑ๋Šฅ์€ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. epoch ์ˆ˜๋ฅผ 3์œผ๋กœ ๊ณ ์ •ํ•˜๊ณ  ๊ทธ๋ฆผ 6๊ณผ ๊ฐ™์ด QA ์Œ๊ณผ ํ•ด๋‹น ๋ฌธ์„œ์˜ ์ˆœ์„œ๋ฅผ ์ •๋ ฌํ•ฉ๋‹ˆ๋‹ค. interleaved ๋ฐฐ์—ด์€ ๋ชจ๋“  ๋ฐ์ดํ„ฐ๋ฅผ 3๋ฒˆ ์ˆœํ™˜ํ•˜๋ฏ€๋กœ ๊ฐ epoch์—์„œ ์งˆ๋ฌธ์ด ๊ด€๋ จ ๋ฌธ์„œ๋ณด๋‹ค ์•ž์„œ๊ฑฐ๋‚˜ ๋’ค์— ์˜ค๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— grouped ๋ฐฐ์—ด์€ ๊ฐ ์˜ˆ์ œ์˜ 3๋ฒˆ์˜ ๋“ฑ์žฅ์„ ํ•จ๊ป˜ ํด๋Ÿฌ์Šคํ„ฐ๋งํ•˜์—ฌ ๋ฐ˜๋ณต๋˜๋Š” ์งˆ๋ฌธ์ด ๊ฐ๊ฐ ๋ฐ˜๋ณต๋˜๋Š” ๋ฌธ์„œ๋ณด๋‹ค ์•ž์ด๋‚˜ ๋’ค์— ์œ„์น˜ํ•˜๋„๋ก ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 2์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, grouped ๋ฐฐ์—ด๊ณผ interleaved ๋ฐฐ์—ด ๋ชจ๋‘์—์„œ ํ•ด๋‹น ๋ฌธ์„œ ์•ž์— QA ์Œ์„ ๋ฐฐ์น˜ํ•˜๋ฉด ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” PIT ๋™์•ˆ ํ•™์Šต ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ๋” ๋ณต์žกํ•˜๊ณ  ์ •๋ณด๊ฐ€ ๋ฐ€์ง‘๋œ ๋ฌธ์„œ์—์„œ ์ •๋ณด๋ฅผ ํก์ˆ˜ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ฐฐ์šฐ๊ธฐ ์ „์— ์ง€์‹์— ์•ก์„ธ์Šคํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ดํ•ดํ•˜๋Š” ๋ฐ ์šฐ์„ ์ˆœ์œ„๋ฅผ ๋‘๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๋จผ์ € ๋‹ค์–‘ํ•œ ์—ํฌํฌ ์ˆ˜์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด ์–ด๋–ป๊ฒŒ ๋‹ฌ๋ผ์ง€๋Š”์ง€ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ํ‘œ 2์™€ ๊ฐ™์ด 1ํšŒ ํ•™์Šต์€ ๋ถˆ์ถฉ๋ถ„ํ•˜๋ฉฐ 3, 5 ๋˜๋Š” 10ํšŒ ํ•™์Šต์˜ ์„ฑ๋Šฅ์€ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์—ํฌํฌ ์ˆ˜๋ฅผ 3์œผ๋กœ ๊ณ ์ •ํ•˜๊ณ  ๊ทธ๋ฆผ 6๊ณผ ๊ฐ™์ด QA ์Œ๊ณผ ํ•ด๋‹น ๋ฌธ์„œ์˜ ์ˆœ์„œ๋ฅผ ์ •๋ ฌํ•ฉ๋‹ˆ๋‹ค. ์ธํ„ฐ๋ฆฌ๋ธŒ ๋ฐฐ์—ด์€ ๋ชจ๋“  ๋ฐ์ดํ„ฐ๋ฅผ 3๋ฒˆ ์ˆœํ™˜ํ•˜๋ฏ€๋กœ ๊ฐ ์—ํฌํฌ์—์„œ ์งˆ๋ฌธ์ด ๊ด€๋ จ ๋ฌธ์„œ๋ณด๋‹ค ์•ž์„œ๊ฑฐ๋‚˜ ๋’ค์— ์˜ค๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ๊ทธ๋ฃนํ™”๋œ ๋ฐฐ์—ด์€ ๊ฐ ์˜ˆ์ œ์˜ 3๋ฒˆ์˜ ๋“ฑ์žฅ์„ ํ•จ๊ป˜ ํด๋Ÿฌ์Šคํ„ฐ๋งํ•˜์—ฌ ๋ฐ˜๋ณต๋˜๋Š” ์งˆ๋ฌธ์ด ๊ฐ๊ฐ ๋ฐ˜๋ณต๋˜๋Š” ๋ฌธ์„œ๋ณด๋‹ค ์•ž์ด๋‚˜ ๋’ค์— ์œ„์น˜ํ•˜๋„๋ก ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 2์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ๊ทธ๋ฃนํ™”๋œ ๋ฐฐ์—ด๊ณผ ์ธํ„ฐ๋ฆฌ๋ธŒ ๋ฐฐ์—ด ๋ชจ๋‘์—์„œ ํ•ด๋‹น ๋ฌธ์„œ ์•ž์— QA ์Œ์„ ๋ฐฐ์น˜ํ•˜๋ฉด ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” PIT ๋™์•ˆ ํ•™์Šต ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ๋” ๋ณต์žกํ•˜๊ณ  ์ •๋ณด๊ฐ€ ๋ฐ€์ง‘๋œ ๋ฌธ์„œ์—์„œ ์ •๋ณด๋ฅผ ํก์ˆ˜ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ฐฐ์šฐ๊ธฐ ์ „์— ์ง€์‹์— ์•ก์„ธ์Šคํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ดํ•ดํ•˜๋Š” ๋ฐ ์šฐ์„ ์ˆœ์œ„๋ฅผ ๋‘๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
**Text Generation.** Text generation tasks, _e.g._, machine translation [624] and automatic summarization [548], are long-standing NLP tasks that have been widely studied, and there have been a number of deployed products and systems based on fine-tuned small models [764, 311]. Since the pre-training of LLMs is established on text prediction, they exhibit strong language generation abilities as commercial products [627] and humans [628], with the help of proper prompts [765, 766]. Additionally, LLMs are flexible to effectively handle special requirement in real-world application scenarios, _e.g._, document-level translation [767], and also enable natural language interaction with users to further improve the generation quality [768]. Despite the above success, recent work also reveals that LLMs are hard to well address the generation tasks about low-resource languages and domains, _e.g._, Marathi-to-English translation [769], due to their unbalanced training data across different languages.
**ํ…์ŠคํŠธ ์ƒ์„ฑ**. ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…(_์˜ˆ: ๊ธฐ๊ณ„ ๋ฒˆ์—ญ[624] ๋ฐ ์ž๋™ ์š”์•ฝ[548])์€ ์˜ค๋žซ๋™์•ˆ ์—ฐ๊ตฌ๋˜์–ด ์˜จ NLP ์ž‘์—…์œผ๋กœ, Fine-tuning๋œ ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ[764, 311]์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์—ฌ๋Ÿฌ ์ œํ’ˆ๊ณผ ์‹œ์Šคํ…œ์ด ๋ฐฐํฌ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. LLM์˜ ์‚ฌ์ „ ํ•™์Šต์€ ํ…์ŠคํŠธ ์˜ˆ์ธก์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ ์ ˆํ•œ ํ”„๋กฌํ”„ํŠธ[765, 766]์˜ ๋„์›€์„ ๋ฐ›์•„ ์ƒ์šฉ ์ œํ’ˆ[627] ๋ฐ ์ธ๊ฐ„[628]๊ณผ ๊ฐ™์€ ๊ฐ•๋ ฅํ•œ ์–ธ์–ด ์ƒ์„ฑ ๋Šฅ๋ ฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์€ ๋ฌธ์„œ ์ˆ˜์ค€ ๋ฒˆ์—ญ[767]๊ณผ ๊ฐ™์€ ์‹ค์ œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ํŠน๋ณ„ํ•œ ์š”๊ตฌ ์‚ฌํ•ญ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์‚ฌ์šฉ์ž์™€ ์ž์—ฐ์Šค๋Ÿฌ์šด ์–ธ์–ด๋กœ ์ƒํ˜ธ ์ž‘์šฉํ•˜์—ฌ ์ƒ์„ฑ ํ’ˆ์งˆ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[768]. ์œ„์™€ ๊ฐ™์€ ์„ฑ๊ณต์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด LLM์€ ๋‹ค์–‘ํ•œ ์–ธ์–ด์— ๊ฑธ์ณ ๋ถˆ๊ท ํ˜•ํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋กœ ์ธํ•ด ๋งˆ๋ผํ‹ฐ์–ด-์˜์–ด ๋ฒˆ์—ญ[769]๊ณผ ๊ฐ™์ด ์ €์ž์› ์–ธ์–ด ๋ฐ ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ์ƒ์„ฑ ์ž‘์—…์„ ์ž˜ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค.
**ํ…์ŠคํŠธ ์ƒ์„ฑ**. ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…(_์˜ˆ: ๊ธฐ๊ณ„ ๋ฒˆ์—ญ[624] ๋ฐ ์ž๋™ ์š”์•ฝ[548])์€ ์˜ค๋žซ๋™์•ˆ ์—ฐ๊ตฌ๋˜์–ด ์˜จ NLP ์ž‘์—…์œผ๋กœ, ๋ฏธ์„ธ ์กฐ์ •๋œ ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ[764, 311]์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์—ฌ๋Ÿฌ ์ œํ’ˆ๊ณผ ์‹œ์Šคํ…œ์ด ๋ฐฐํฌ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. LLM์˜ ์‚ฌ์ „ ํ•™์Šต์€ ํ…์ŠคํŠธ ์˜ˆ์ธก์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ ์ ˆํ•œ ํ”„๋กฌํ”„ํŠธ[765, 766]์˜ ๋„์›€์„ ๋ฐ›์•„ ์ƒ์šฉ ์ œํ’ˆ[627] ๋ฐ ์ธ๊ฐ„[628]๊ณผ ๊ฐ™์€ ๊ฐ•๋ ฅํ•œ ์–ธ์–ด ์ƒ์„ฑ ๋Šฅ๋ ฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์€ ๋ฌธ์„œ ์ˆ˜์ค€ ๋ฒˆ์—ญ[767]๊ณผ ๊ฐ™์€ ์‹ค์ œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ํŠน๋ณ„ํ•œ ์š”๊ตฌ ์‚ฌํ•ญ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์‚ฌ์šฉ์ž์™€ ์ž์—ฐ์Šค๋Ÿฌ์šด ์–ธ์–ด๋กœ ์ƒํ˜ธ ์ž‘์šฉํ•˜์—ฌ ์ƒ์„ฑ ํ’ˆ์งˆ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[768]. ์œ„์™€ ๊ฐ™์€ ์„ฑ๊ณต์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด LLM์€ ๋‹ค์–‘ํ•œ ์–ธ์–ด์— ๊ฑธ์ณ ๋ถˆ๊ท ํ˜•ํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋กœ ์ธํ•ด ๋งˆ๋ผํ‹ฐ์–ด-์˜์–ด ๋ฒˆ์—ญ[769]๊ณผ ๊ฐ™์ด ์ €์ž์› ์–ธ์–ด ๋ฐ ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ์ƒ์„ฑ ์ž‘์—…์„ ์ž˜ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค.
In contrast to most previous Code LLMs that primarily emphasize the pre-training process, there has been limited exploration of fine-grained instruction tuning in the Code domain. The introduction of instruction tuning initially aimed to enhance the generalization capabilities of LMs across different tasks [19; 20; 21; 22; 23; 24; 25]. OpenAI's InstructGPT [10], for instance, involved soliciting human annotators to provide explicit instructions to ensure alignment with users' intentions. Similarly, recent works such as Alpaca [26] employed the self-instruct [27] method, where ChatGPT generated the instruction data. Vicuna [28] utilized user-shared conversations collected from ShareGPT.com. WizardLM [29] introduced the _Evol-Instruct_ method, which involved evolving existing instruction data to generate more complex and diverse datasets. However, it is worth noting that all these approaches primarily focused on the general domain and lacked specific design considerations for the code domain.
์ฃผ๋กœ ์‚ฌ์ „ ํ•™์Šต ๊ณผ์ •์„ ๊ฐ•์กฐํ•˜๋Š” ๋Œ€๋ถ€๋ถ„์˜ ๊ธฐ์กด ์ฝ”๋“œ LLM๊ณผ๋Š” ๋‹ฌ๋ฆฌ ์ฝ”๋“œ ์˜์—ญ์—์„œ ์„ธ๋ถ„ํ™”๋œ Instruction Tuning์— ๋Œ€ํ•œ ํƒ๊ตฌ๋Š” ์ œํ•œ์ ์ด์—ˆ์Šต๋‹ˆ๋‹ค. Instruction Tuning์€ ์ฒ˜์Œ์—๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๊ฑธ์ณ LM์˜ ์ผ๋ฐ˜ํ™” ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ–ˆ์Šต๋‹ˆ๋‹ค[19; 20; 21; 22; 23; 24; 25]. ์˜ˆ๋ฅผ ๋“ค์–ด, OpenAI์˜ InstructGPT [10]์€ ์‚ฌ์šฉ์ž์˜ ์˜๋„์— ๋งž๊ฒŒ ์กฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด ๋ช…์‹œ์ ์ธ ์ง€์นจ์„ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์ฃผ์„๊ฐ€๋ฅผ ์š”์ฒญํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Alpaca [26]์™€ ๊ฐ™์€ ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ChatGPT๊ฐ€ Instruction ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” self-instruct [27] ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. Vicuna [28]๋Š” ShareGPT.com์—์„œ ์ˆ˜์ง‘ํ•œ ์‚ฌ์šฉ์ž ๊ณต์œ  ๋Œ€ํ™”๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. WizardLM [29]์€ ๋” ๋ณต์žกํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด Instruction ๋ฐ์ดํ„ฐ๋ฅผ ์ง„ํ™”์‹œํ‚ค๋Š” _Evol-Instruct_ ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ชจ๋“  ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ฃผ๋กœ ์ผ๋ฐ˜ ๋„๋ฉ”์ธ์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์œผ๋ฉฐ ์ฝ”๋“œ ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ๊ตฌ์ฒด์ ์ธ ์„ค๊ณ„ ๊ณ ๋ ค ์‚ฌํ•ญ์ด ๋ถ€์กฑํ–ˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
์ฃผ๋กœ ์‚ฌ์ „ ํ•™์Šต ๊ณผ์ •์„ ๊ฐ•์กฐํ•˜๋Š” ๋Œ€๋ถ€๋ถ„์˜ ๊ธฐ์กด ์ฝ”๋“œ LLM๊ณผ๋Š” ๋‹ฌ๋ฆฌ ์ฝ”๋“œ ์˜์—ญ์—์„œ ์„ธ๋ถ„ํ™”๋œ ๋ช…๋ น์–ด ํŠœ๋‹์— ๋Œ€ํ•œ ํƒ๊ตฌ๋Š” ์ œํ•œ์ ์ด์—ˆ์Šต๋‹ˆ๋‹ค. ๋ช…๋ น์–ด ํŠœ๋‹์€ ์ฒ˜์Œ์—๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๊ฑธ์ณ LM์˜ ์ผ๋ฐ˜ํ™” ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ–ˆ์Šต๋‹ˆ๋‹ค[19; 20; 21; 22; 23; 24; 25]. ์˜ˆ๋ฅผ ๋“ค์–ด, OpenAI์˜ InstructGPT [10]์€ ์‚ฌ์šฉ์ž์˜ ์˜๋„์— ๋งž๊ฒŒ ์กฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด ๋ช…์‹œ์ ์ธ ์ง€์นจ์„ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์ฃผ์„๊ฐ€๋ฅผ ์š”์ฒญํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Alpaca [26]์™€ ๊ฐ™์€ ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ChatGPT๊ฐ€ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์…€ํ”„ ์ธ์ŠคํŠธ๋ŸญํŠธ [27] ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. Vicuna [28]๋Š” ShareGPT.com์—์„œ ์ˆ˜์ง‘ํ•œ ์‚ฌ์šฉ์ž ๊ณต์œ  ๋Œ€ํ™”๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. WizardLM [29]์€ ๋” ๋ณต์žกํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ๋ฅผ ์ง„ํ™”์‹œํ‚ค๋Š” _Evol-Instruct_ ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ชจ๋“  ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ฃผ๋กœ ์ผ๋ฐ˜ ๋„๋ฉ”์ธ์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์œผ๋ฉฐ ์ฝ”๋“œ ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ๊ตฌ์ฒด์ ์ธ ์„ค๊ณ„ ๊ณ ๋ ค ์‚ฌํ•ญ์ด ๋ถ€์กฑํ–ˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
In this work, we introduce Self-Instruct, a semi-automated process for instruction-tuning a pretrained LM using instructional signals from the model itself. The overall process is an iterative bootstrapping algorithm (see Figure 2), which starts off with a limited (e.g., 175 in our study) seed set of manually-written tasks that are used to guide the overall generation. In the first phase, the model is prompted to generate instructions for new tasks. This step leverages the existing collection of instructions to create more broad-coverage instructions that define (often new) tasks. Given the newly-generated set of instructions, the framework also creates input-output instances for them, which can be later used for supervising the instruction tuning. Finally, various heuristics are used to automatically filter low-quality or repeated instructions, before adding the remaining valid tasks to the task pool. This process can be repeated for many iterations until reaching a large number of tasks.
์ด ์ž‘์—…์—์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ LM์„ ๋ชจ๋ธ ์ž์ฒด์˜ instructional ์‹ ํ˜ธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Instruction Tuningํ•˜๋Š” ๋ฐ˜ ์ž๋™ํ™” ํ”„๋กœ์„ธ์Šค์ธ Self-Instruct๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ํ”„๋กœ์„ธ์Šค๋Š” ๋ฐ˜๋ณต์ ์ธ ๋ถ€ํŠธ์ŠคํŠธ๋žฉ ์•Œ๊ณ ๋ฆฌ์ฆ˜(๊ทธ๋ฆผ 2 ์ฐธ์กฐ)์œผ๋กœ, ์ „์ฒด ์ƒ์„ฑ์„ ์•ˆ๋‚ดํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ์ œํ•œ๋œ(์˜ˆ: ์ด ์—ฐ๊ตฌ์—์„œ๋Š” 175๊ฐœ) ์ˆ˜์ž‘์—…์œผ๋กœ ์ž‘์„ฑ๋œ ํƒœ์Šคํฌ์˜ ์‹œ๋“œ ์„ธํŠธ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” ๋ชจ๋ธ์— ์ƒˆ ์ž‘์—…์— ๋Œ€ํ•œ instructions๋ฅผ ์ƒ์„ฑํ•˜๋ผ๋Š” ํ”„๋กฌํ”„ํŠธ๊ฐ€ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—์„œ๋Š” ๊ธฐ์กด instructions ์ปฌ๋ ‰์…˜์„ ํ™œ์šฉํ•˜์—ฌ (์ข…์ข… ์ƒˆ๋กœ์šด) ์ž‘์—…์„ ์ •์˜ํ•˜๋Š” ๋ณด๋‹ค ๊ด‘๋ฒ”์œ„ํ•œ instructions๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ ์ƒ์„ฑ๋œ instructions ์ง‘ํ•ฉ์ด ์ฃผ์–ด์ง€๋ฉด ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ์ด์— ๋Œ€ํ•œ ์ž…๋ ฅ-์ถœ๋ ฅ ์ธ์Šคํ„ด์Šค๋„ ์ƒ์„ฑํ•˜์—ฌ ๋‚˜์ค‘์— Instruction Tuning์„ ๊ฐ๋…ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๋‹ค์–‘ํ•œ ํœด๋ฆฌ์Šคํ‹ฑ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ’ˆ์งˆ์ด ๋‚ฎ๊ฑฐ๋‚˜ ๋ฐ˜๋ณต๋˜๋Š” instructions๋ฅผ ์ž๋™์œผ๋กœ ํ•„ํ„ฐ๋งํ•œ ํ›„ ๋‚˜๋จธ์ง€ ์œ ํšจํ•œ ์ž‘์—…์„ ์ž‘์—… ํ’€์— ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์„ธ์Šค๋Š” ๋งŽ์€ ์ž‘์—… ์ˆ˜์— ๋„๋‹ฌํ•  ๋•Œ๊นŒ์ง€ ์—ฌ๋Ÿฌ ๋ฐ˜๋ณต์„ ํ†ตํ•ด ๋ฐ˜๋ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด ์ž‘์—…์—์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ LM์„ ๋ชจ๋ธ ์ž์ฒด์˜ ๊ต์œก ์‹ ํ˜ธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ต์œก ํŠœ๋‹ํ•˜๋Š” ๋ฐ˜ ์ž๋™ํ™” ํ”„๋กœ์„ธ์Šค์ธ Self-Instruct๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ํ”„๋กœ์„ธ์Šค๋Š” ๋ฐ˜๋ณต์ ์ธ ๋ถ€ํŠธ์ŠคํŠธ๋žฉ ์•Œ๊ณ ๋ฆฌ์ฆ˜(๊ทธ๋ฆผ 2 ์ฐธ์กฐ)์œผ๋กœ, ์ „์ฒด ์ƒ์„ฑ์„ ์•ˆ๋‚ดํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ์ œํ•œ๋œ(์˜ˆ: ์ด ์—ฐ๊ตฌ์—์„œ๋Š” 175๊ฐœ) ์ˆ˜์ž‘์—…์œผ๋กœ ์ž‘์„ฑ๋œ ํƒœ์Šคํฌ์˜ ์‹œ๋“œ ์„ธํŠธ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” ๋ชจ๋ธ์— ์ƒˆ ์ž‘์—…์— ๋Œ€ํ•œ ์ง€์นจ์„ ์ƒ์„ฑํ•˜๋ผ๋Š” ํ”„๋กฌํ”„ํŠธ๊ฐ€ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—์„œ๋Š” ๊ธฐ์กด ์ง€์นจ ์ปฌ๋ ‰์…˜์„ ํ™œ์šฉํ•˜์—ฌ (์ข…์ข… ์ƒˆ๋กœ์šด) ์ž‘์—…์„ ์ •์˜ํ•˜๋Š” ๋ณด๋‹ค ๊ด‘๋ฒ”์œ„ํ•œ ์ง€์นจ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ ์ƒ์„ฑ๋œ ์ง€์นจ ์ง‘ํ•ฉ์ด ์ฃผ์–ด์ง€๋ฉด ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ์ด์— ๋Œ€ํ•œ ์ž…๋ ฅ-์ถœ๋ ฅ ์ธ์Šคํ„ด์Šค๋„ ์ƒ์„ฑํ•˜์—ฌ ๋‚˜์ค‘์— ์ง€์นจ ํŠœ๋‹์„ ๊ฐ๋…ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๋‹ค์–‘ํ•œ ํœด๋ฆฌ์Šคํ‹ฑ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ’ˆ์งˆ์ด ๋‚ฎ๊ฑฐ๋‚˜ ๋ฐ˜๋ณต๋˜๋Š” ์ง€์นจ์„ ์ž๋™์œผ๋กœ ํ•„ํ„ฐ๋งํ•œ ํ›„ ๋‚˜๋จธ์ง€ ์œ ํšจํ•œ ์ž‘์—…์„ ์ž‘์—… ํ’€์— ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์„ธ์Šค๋Š” ๋งŽ์€ ์ž‘์—… ์ˆ˜์— ๋„๋‹ฌํ•  ๋•Œ๊นŒ์ง€ ์—ฌ๋Ÿฌ ๋ฐ˜๋ณต์„ ํ†ตํ•ด ๋ฐ˜๋ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_RL fine-tuning_. At this step, aligning (_i.e._, fine-tuning) the LM is formalized as an RL problem. In this setting, the pre-trained LM acts as the policy that takes as input a prompt and returns an output text, the action space of it is the vocabulary, the state is the currently generated token sequence, and the reward is provided by the RM. To avoid eviating significantly from the initial (before tuning) LM, a penalty term is commonly incorporated into the reward function. For example, InstructGPT optimizes the LM against the RM using the PPO algorithm. For each input prompt, InstructGPT calculates the KL divergence between the generated results from the current LM and the initial LM as the penalty. It is noted that the second and final steps can be iterated in multiple turns for better aligning LLMs. Due to the instability of the RL algorithm, recent work [373] replaces the RL tuning with another supervised fine-tuning by reusing the best ranked samples with higher rewards.
\(\bullet\)_RL Fine-tuning_. ์ด ๋‹จ๊ณ„์—์„œ LM์„ ์ •๋ ฌํ•˜๋Š” (_์ฆ‰, Fine-tuning)์€ RL ๋ฌธ์ œ๋กœ ๊ณต์‹ํ™”๋ฉ๋‹ˆ๋‹ค. ์ด ์„ค์ •์—์„œ ์‚ฌ์ „ ํ•™์Šต๋œ LM์€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ๋ฐ›์•„ ์ถœ๋ ฅ ํ…์ŠคํŠธ๋ฅผ ๋ฐ˜ํ™˜ํ•˜๋Š” ์ •์ฑ… ์—ญํ• ์„ ํ•˜๋ฉฐ, ๊ทธ action ๊ณต๊ฐ„์€ ์–ดํœ˜, ์ƒํƒœ๋Š” ํ˜„์žฌ ์ƒ์„ฑ๋œ ํ† ํฐ ์‹œํ€€์Šค, ๋ณด์ƒ์€ RM์— ์˜ํ•ด ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ํŠœ๋‹ ์ „์˜ ์ดˆ๊ธฐ LM์—์„œ ํฌ๊ฒŒ ๋ฒ—์–ด๋‚˜์ง€ ์•Š๋„๋ก ๋ณด์ƒ ํ•จ์ˆ˜์— ํŽ˜๋„ํ‹ฐ ํ•ญ์„ ํ†ต์ƒ์ ์œผ๋กœ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, InstructGPT๋Š” PPO ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ RM์— ๋Œ€ํ•ด LM์„ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด InstructGPT๋Š” ํ˜„์žฌ LM์—์„œ ์ƒ์„ฑ๋œ ๊ฒฐ๊ณผ์™€ ์ดˆ๊ธฐ LM ์‚ฌ์ด์˜ KL ๋ฐœ์‚ฐ์„ ํŽ˜๋„ํ‹ฐ๋กœ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์™€ ๋งˆ์ง€๋ง‰ ๋‹จ๊ณ„๋Š” ๋” ๋‚˜์€ ์ •๋ ฌ์„ ์œ„ํ•ด ์—ฌ๋Ÿฌ ํ„ด์œผ๋กœ ๋ฐ˜๋ณตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ฉ๋‹ˆ๋‹ค. RL ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๋ถˆ์•ˆ์ •์„ฑ์œผ๋กœ ์ธํ•ด ์ตœ๊ทผ ์—ฐ๊ตฌ [373]์—์„œ๋Š” ๋ณด์ƒ์ด ๋” ๋†’์€ ์ตœ๊ณ  ์ˆœ์œ„ ์ƒ˜ํ”Œ์„ ์žฌ์‚ฌ์šฉํ•˜์—ฌ RL ํŠœ๋‹์„ ๋‹ค๋ฅธ Supervised Fine-tuning์œผ๋กœ ๋Œ€์ฒดํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_RL ๋ฏธ์„ธ ์กฐ์ •_. ์ด ๋‹จ๊ณ„์—์„œ LM์„ ์ •๋ ฌํ•˜๋Š” (_์ฆ‰, ๋ฏธ์„ธ ์กฐ์ •)์€ RL ๋ฌธ์ œ๋กœ ๊ณต์‹ํ™”๋ฉ๋‹ˆ๋‹ค. ์ด ์„ค์ •์—์„œ ์‚ฌ์ „ ํ•™์Šต๋œ LM์€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ๋ฐ›์•„ ์ถœ๋ ฅ ํ…์ŠคํŠธ๋ฅผ ๋ฐ˜ํ™˜ํ•˜๋Š” ์ •์ฑ… ์—ญํ• ์„ ํ•˜๋ฉฐ, ๊ทธ ์•ก์…˜ ๊ณต๊ฐ„์€ ์–ดํœ˜, ์ƒํƒœ๋Š” ํ˜„์žฌ ์ƒ์„ฑ๋œ ํ† ํฐ ์‹œํ€€์Šค, ๋ณด์ƒ์€ RM์— ์˜ํ•ด ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ํŠœ๋‹ ์ „์˜ ์ดˆ๊ธฐ LM์—์„œ ํฌ๊ฒŒ ๋ฒ—์–ด๋‚˜์ง€ ์•Š๋„๋ก ๋ณด์ƒ ํ•จ์ˆ˜์— ํŽ˜๋„ํ‹ฐ ํ•ญ์„ ํ†ต์ƒ์ ์œผ๋กœ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, InstructGPT๋Š” PPO ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ RM์— ๋Œ€ํ•ด LM์„ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด InstructGPT๋Š” ํ˜„์žฌ LM์—์„œ ์ƒ์„ฑ๋œ ๊ฒฐ๊ณผ์™€ ์ดˆ๊ธฐ LM ์‚ฌ์ด์˜ KL ๋ฐœ์‚ฐ์„ ํŽ˜๋„ํ‹ฐ๋กœ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์™€ ๋งˆ์ง€๋ง‰ ๋‹จ๊ณ„๋Š” ๋” ๋‚˜์€ ์ •๋ ฌ์„ ์œ„ํ•ด ์—ฌ๋Ÿฌ ํ„ด์œผ๋กœ ๋ฐ˜๋ณตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ฉ๋‹ˆ๋‹ค. RL ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๋ถˆ์•ˆ์ •์„ฑ์œผ๋กœ ์ธํ•ด ์ตœ๊ทผ ์—ฐ๊ตฌ [373]์—์„œ๋Š” ๋ณด์ƒ์ด ๋” ๋†’์€ ์ตœ๊ณ  ์ˆœ์œ„ ์ƒ˜ํ”Œ์„ ์žฌ์‚ฌ์šฉํ•˜์—ฌ RL ํŠœ๋‹์„ ๋‹ค๋ฅธ ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •์œผ๋กœ ๋Œ€์ฒดํ–ˆ์Šต๋‹ˆ๋‹ค.
**Planning with Feedback**: In many real-world scenarios, the agents need to make long-horizon planning to solve complex tasks. When facing these tasks, the above planning modules without feedback can be less effective due to the following reasons: firstly, generating a flawless plan directly from the beginning is extremely difficult as it needs to consider various complex preconditions. As a result, simply following the initial plan often leads to failure. Moreover, the execution of the plan may be hindered by unpredictable transition dynamics, rendering the initial plan non-executable. Simultaneously, when examining how humans tackle complex tasks, we find that individuals may iteratively make and revise their plans based on external feedback. To simulate such human capability, researchers have designed many planning modules, where the agent can receive feedback after taking actions. The feedback can be obtained from environments, humans, and models, which are detailed in the following.
**Planning with Feedback**: ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์žฅ๊ธฐ์ ์ธ ๊ณ„ํš์„ ์„ธ์›Œ์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์— ์ง๋ฉดํ–ˆ์„ ๋•Œ ํ”ผ๋“œ๋ฐฑ์ด ์—†๋Š” ์œ„์˜ planning ๋ชจ๋“ˆ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ด์œ ๋กœ ํšจ๊ณผ๊ฐ€ ๋–จ์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋‹ค์–‘ํ•œ ๋ณต์žกํ•œ ์ „์ œ ์กฐ๊ฑด์„ ๊ณ ๋ คํ•ด์•ผ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ฒ˜์Œ๋ถ€ํ„ฐ ์™„๋ฒฝํ•œ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ์ดˆ๊ธฐ ๊ณ„ํš์„ ๋‹จ์ˆœํžˆ ๋”ฐ๋ฅด๋ฉด ์‹คํŒจ๋กœ ์ด์–ด์ง€๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์˜ˆ์ธกํ•  ์ˆ˜ ์—†๋Š” ์ „ํ™˜ ์—ญํ•™์œผ๋กœ ์ธํ•ด ๊ณ„ํš ์‹คํ–‰์ด ๋ฐฉํ•ด๋ฅผ ๋ฐ›์•„ ์ดˆ๊ธฐ ๊ณ„ํš์„ ์‹คํ–‰ํ•  ์ˆ˜ ์—†๊ฒŒ ๋  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋™์‹œ์— ์ธ๊ฐ„์ด ๋ณต์žกํ•œ ์ž‘์—…์„ ์–ด๋–ป๊ฒŒ ํ•ด๊ฒฐํ•˜๋Š”์ง€ ์‚ดํŽด๋ณด๋ฉด, ๊ฐœ์ธ์ด ์™ธ๋ถ€ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ”ํƒ•์œผ๋กœ ๊ณ„ํš์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ๋งŒ๋“ค๊ณ  ์ˆ˜์ •ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ธ๊ฐ„์˜ ๋Šฅ๋ ฅ์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ ์—์ด์ „ํŠธ๊ฐ€ ํ–‰๋™์„ ์ทจํ•œ ํ›„ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ๋Š” ๋งŽ์€ planning ๋ชจ๋“ˆ์„ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ”ผ๋“œ๋ฐฑ์€ ํ™˜๊ฒฝ, ์ธ๊ฐ„, ๋ชจ๋ธ์—์„œ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋‹ค์Œ์—์„œ ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
**ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•œ ๊ณ„ํš**: ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์žฅ๊ธฐ์ ์ธ ๊ณ„ํš์„ ์„ธ์›Œ์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์— ์ง๋ฉดํ–ˆ์„ ๋•Œ ํ”ผ๋“œ๋ฐฑ์ด ์—†๋Š” ์œ„์˜ ๊ณ„ํš ๋ชจ๋“ˆ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ด์œ ๋กœ ํšจ๊ณผ๊ฐ€ ๋–จ์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๋‹ค์–‘ํ•œ ๋ณต์žกํ•œ ์ „์ œ ์กฐ๊ฑด์„ ๊ณ ๋ คํ•ด์•ผ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ฒ˜์Œ๋ถ€ํ„ฐ ์™„๋ฒฝํ•œ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ์ดˆ๊ธฐ ๊ณ„ํš์„ ๋‹จ์ˆœํžˆ ๋”ฐ๋ฅด๋ฉด ์‹คํŒจ๋กœ ์ด์–ด์ง€๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์˜ˆ์ธกํ•  ์ˆ˜ ์—†๋Š” ์ „ํ™˜ ์—ญํ•™์œผ๋กœ ์ธํ•ด ๊ณ„ํš ์‹คํ–‰์ด ๋ฐฉํ•ด๋ฅผ ๋ฐ›์•„ ์ดˆ๊ธฐ ๊ณ„ํš์„ ์‹คํ–‰ํ•  ์ˆ˜ ์—†๊ฒŒ ๋  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋™์‹œ์— ์ธ๊ฐ„์ด ๋ณต์žกํ•œ ์ž‘์—…์„ ์–ด๋–ป๊ฒŒ ํ•ด๊ฒฐํ•˜๋Š”์ง€ ์‚ดํŽด๋ณด๋ฉด, ๊ฐœ์ธ์ด ์™ธ๋ถ€ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ”ํƒ•์œผ๋กœ ๊ณ„ํš์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ๋งŒ๋“ค๊ณ  ์ˆ˜์ •ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ธ๊ฐ„์˜ ๋Šฅ๋ ฅ์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ ์—์ด์ „ํŠธ๊ฐ€ ํ–‰๋™์„ ์ทจํ•œ ํ›„ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ๋Š” ๋งŽ์€ ๊ณ„ํš ๋ชจ๋“ˆ์„ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ”ผ๋“œ๋ฐฑ์€ ํ™˜๊ฒฝ, ์ธ๊ฐ„, ๋ชจ๋ธ์—์„œ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋‹ค์Œ์—์„œ ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
Pretrained models are not always used as-is, but rather further fine-tuned to improve their performance, safety, and usability. Often models are first trained to follow instructions (Mishra et al., 2022; Wei et al., 2022; Sanh et al., 2022), and then further trained on human preferences (Ouyang et al., 2022) to improve the quality of their generations. We showcase the efficacy of using OLMo as a base model for further fine-tuning by training OLMo to be a general chat assistant following our Open Instruct (Tulu) data and training setup (Ivison et al., 2023). Our approach involves first performing instruction fine-tuning with a mixture of distilled and human-written instruction data and then further aligning the model with distilled preference data using Direct Preference Optimization (DPO) (Rafailov et al., 2023). We experimented with mixing the Tulu instruction data at the end of pretraining, as done in recent models such as DeepSeek-AI et al. (2024), but did not have conclusive findings.
์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์€ ํ•ญ์ƒ ๊ทธ๋Œ€๋กœ ์‚ฌ์šฉ๋˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์„ฑ๋Šฅ, ์•ˆ์ „์„ฑ ๋ฐ ์‚ฌ์šฉ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€๋กœ Fine-tuning๋ฉ๋‹ˆ๋‹ค. ์ข…์ข… ๋ชจ๋ธ์€ ๋จผ์ € Instruction์„ ๋”ฐ๋ฅด๋„๋ก ํ•™์Šตํ•œ ๋‹ค์Œ(Mishra et al., 2022; Wei et al., 2022; Sanh et al., 2022) ์ƒ์„ฑ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋Œ€ํ•ด ์ถ”๊ฐ€ ํ•™์Šต(Ouyang et al., 2022)์„ ๊ฑฐ์นฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” OLMo๋ฅผ Open Instruct(Tulu) ๋ฐ์ดํ„ฐ ๋ฐ ํ›ˆ๋ จ ์„ค์ •(Ivison et al., 2023)์— ๋”ฐ๋ผ ์ผ๋ฐ˜ ์ฑ„ํŒ… ์–ด์‹œ์Šคํ„ดํŠธ๋กœ ํ›ˆ๋ จํ•˜์—ฌ ์ถ”๊ฐ€ Fine-tuning์„ ์œ„ํ•œ ๊ธฐ๋ณธ ๋ชจ๋ธ๋กœ์„œ OLMo๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋จผ์ € ์ฆ๋ฅ˜ ๋ฐ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ Instruction ๋ฐ์ดํ„ฐ๋ฅผ ํ˜ผํ•ฉํ•˜์—ฌ Instruction Fine-tuning์„ ์ˆ˜ํ–‰ํ•œ ๋‹ค์Œ, Direct Preference Optimization (DPO)(Rafailov et al., 2023)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฆ๋ฅ˜๋œ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ์™€ ๋ชจ๋ธ์„ ์ถ”๊ฐ€๋กœ ์ •๋ ฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋ชจ๋ธ(DeepSeek-AI ๋“ฑ, 2024)์—์„œ์™€ ๊ฐ™์ด ์‚ฌ์ „ ํ•™์Šต์ด ๋๋‚  ๋•Œ Tulu Instruction ๋ฐ์ดํ„ฐ๋ฅผ ํ˜ผํ•ฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‹คํ—˜ํ•ด ๋ณด์•˜์ง€๋งŒ ๊ฒฐ์ •์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์–ป์ง€๋Š” ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค.
์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์€ ํ•ญ์ƒ ๊ทธ๋Œ€๋กœ ์‚ฌ์šฉ๋˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์„ฑ๋Šฅ, ์•ˆ์ „์„ฑ ๋ฐ ์‚ฌ์šฉ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€๋กœ ๋ฏธ์„ธ ์กฐ์ •๋ฉ๋‹ˆ๋‹ค. ์ข…์ข… ๋ชจ๋ธ์€ ๋จผ์ € ์ง€์นจ์„ ๋”ฐ๋ฅด๋„๋ก ํ•™์Šตํ•œ ๋‹ค์Œ(Mishra et al., 2022; Wei et al., 2022; Sanh et al., 2022) ์ƒ์„ฑ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋Œ€ํ•ด ์ถ”๊ฐ€ ํ•™์Šต(Ouyang et al., 2022)์„ ๊ฑฐ์นฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” OLMo๋ฅผ Open Instruct(Tulu) ๋ฐ์ดํ„ฐ ๋ฐ ํ›ˆ๋ จ ์„ค์ •(Ivison et al., 2023)์— ๋”ฐ๋ผ ์ผ๋ฐ˜ ์ฑ„ํŒ… ์–ด์‹œ์Šคํ„ดํŠธ๋กœ ํ›ˆ๋ จํ•˜์—ฌ ์ถ”๊ฐ€ ํ›ˆ๋ จ์„ ์œ„ํ•œ ๊ธฐ๋ณธ ๋ชจ๋ธ๋กœ์„œ OLMo๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋จผ์ € ์ฆ๋ฅ˜ ๋ฐ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ์ง€์นจ ๋ฐ์ดํ„ฐ๋ฅผ ํ˜ผํ•ฉํ•˜์—ฌ ์ง€์นจ ๋ฏธ์„ธ ์กฐ์ •์„ ์ˆ˜ํ–‰ํ•œ ๋‹ค์Œ, ์ง์ ‘ ์„ ํ˜ธ๋„ ์ตœ์ ํ™”(DPO)(Rafailov et al., 2023)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฆ๋ฅ˜๋œ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ์™€ ๋ชจ๋ธ์„ ์ถ”๊ฐ€๋กœ ์ •๋ ฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋ชจ๋ธ(DeepSeek-AI ๋“ฑ, 2024)์—์„œ์™€ ๊ฐ™์ด ์‚ฌ์ „ ํ•™์Šต์ด ๋๋‚  ๋•Œ ํˆด๋ฃจ ์ง€์นจ ๋ฐ์ดํ„ฐ๋ฅผ ํ˜ผํ•ฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‹คํ—˜ํ•ด ๋ณด์•˜์ง€๋งŒ ๊ฒฐ์ •์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์–ป์ง€๋Š” ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค.
Self-Rag trains an arbitrary LM to generate text with reflection tokens by unifying them as the next token prediction from the expanded model vocabulary. We train our generator LM on a diverse collection of text interleaved with reflection tokens and retrieved passages. Reflection tokens, inspired by reward models used in reinforcement learning (Ziegler et al., 2019; Ouyang et al., 2022), are inserted offline into the original corpus by a trained _critic_ model. This eliminates the need to host a critic model during training, reducing overhead. The critic model, in part, is supervised on a dataset of input, output, and corresponding reflection tokens collected by prompting a propriety LM (i.e., GPT-4; OpenAI 2023). While we draw inspiration from studies that use control tokens to start and guide text generation (Lu et al., 2022; Keskar et al., 2019), our trained LM uses critique tokens to assess its own predictions after each generated segment as an integral part of the generation output.
Self-Rag๋Š” ํ™•์žฅ๋œ ๋ชจ๋ธ ์–ดํœ˜์—์„œ ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก์œผ๋กœ ํ†ตํ•ฉํ•˜์—ฌ reflection ํ† ํฐ์œผ๋กœ ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ์ž„์˜์˜ LM์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” reflection ํ† ํฐ๊ณผ ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์ด interleaved๋œ ๋‹ค์–‘ํ•œ ํ…์ŠคํŠธ ์ปฌ๋ ‰์…˜์— ๋Œ€ํ•ด generator LM์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. Reinforcement learning(Ziegler et al., 2019; Ouyang et al., 2022)์— ์‚ฌ์šฉ๋˜๋Š” reward ๋ชจ๋ธ์—์„œ ์˜๊ฐ์„ ์–ป์€ reflection ํ† ํฐ์€ ํ›ˆ๋ จ๋œ _critic_ ๋ชจ๋ธ์— ์˜ํ•ด ์›๋ณธ ์ฝ”ํผ์Šค์— ์˜คํ”„๋ผ์ธ์œผ๋กœ ์‚ฝ์ž…๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ›ˆ๋ จ ์ค‘์— critic ๋ชจ๋ธ์„ ํ˜ธ์ŠคํŒ…ํ•  ํ•„์š”๊ฐ€ ์—†์–ด overhead๊ฐ€ ์ค„์–ด๋“ญ๋‹ˆ๋‹ค. critic ๋ชจ๋ธ์€ ๋ถ€๋ถ„์ ์œผ๋กœ propriety LM(์˜ˆ: GPT-4; OpenAI 2023)์„ promptํ•˜์—ฌ ์ˆ˜์ง‘ํ•œ ์ž…๋ ฅ, ์ถœ๋ ฅ ๋ฐ ํ•ด๋‹น reflection ํ† ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด supervised๋ฉ๋‹ˆ๋‹ค. control ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ์‹œ์ž‘ํ•˜๊ณ  ์•ˆ๋‚ดํ•˜๋Š” ์—ฐ๊ตฌ(Lu et al., 2022; Keskar et al., 2019)์—์„œ ์˜๊ฐ์„ ์–ป์—ˆ์ง€๋งŒ, ์šฐ๋ฆฌ๊ฐ€ ํ›ˆ๋ จํ•œ LM์€ ์ƒ์„ฑ๋œ ๊ฐ ์„ธ๊ทธ๋จผํŠธ ์ดํ›„์— ์ƒ์„ฑ ๊ฒฐ๊ณผ์˜ ํ•„์ˆ˜์ ์ธ ๋ถ€๋ถ„์œผ๋กœ critique ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ž์ฒด ์˜ˆ์ธก์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
Self-Rag๋Š” ํ™•์žฅ๋œ ๋ชจ๋ธ ์–ดํœ˜์—์„œ ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก์œผ๋กœ ํ†ตํ•ฉํ•˜์—ฌ ๋ฐ˜์‚ฌ ํ† ํฐ์œผ๋กœ ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ์ž„์˜์˜ LM์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ฐ˜์‚ฌ ํ† ํฐ๊ณผ ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์ด ์ธํ„ฐ๋ฆฌ๋ธŒ๋œ ๋‹ค์–‘ํ•œ ํ…์ŠคํŠธ ์ปฌ๋ ‰์…˜์— ๋Œ€ํ•ด ์ œ๋„ˆ๋ ˆ์ดํ„ฐ LM์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๊ฐ•ํ™” ํ•™์Šต(Ziegler et al., 2019; Ouyang et al., 2022)์— ์‚ฌ์šฉ๋˜๋Š” ๋ณด์ƒ ๋ชจ๋ธ์—์„œ ์˜๊ฐ์„ ์–ป์€ ๋ฐ˜์‚ฌ ํ† ํฐ์€ ํ›ˆ๋ จ๋œ _๋น„ํ‰๊ฐ€_ ๋ชจ๋ธ์— ์˜ํ•ด ์›๋ณธ ์ฝ”ํผ์Šค์— ์˜คํ”„๋ผ์ธ์œผ๋กœ ์‚ฝ์ž…๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ›ˆ๋ จ ์ค‘์— ๋น„ํ‰๊ฐ€ ๋ชจ๋ธ์„ ํ˜ธ์ŠคํŒ…ํ•  ํ•„์š”๊ฐ€ ์—†์–ด ์˜ค๋ฒ„ํ—ค๋“œ๊ฐ€ ์ค„์–ด๋“ญ๋‹ˆ๋‹ค. ๋น„ํ‰๊ฐ€ ๋ชจ๋ธ์€ ๋ถ€๋ถ„์ ์œผ๋กœ ํŠนํ—ˆ LM(์˜ˆ: GPT-4; OpenAI 2023)์„ ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ์ˆ˜์ง‘ํ•œ ์ž…๋ ฅ, ์ถœ๋ ฅ ๋ฐ ํ•ด๋‹น ๋ฐ˜์‚ฌ ํ† ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ๊ฐ๋…๋ฉ๋‹ˆ๋‹ค. ์ปจํŠธ๋กค ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ์‹œ์ž‘ํ•˜๊ณ  ์•ˆ๋‚ดํ•˜๋Š” ์—ฐ๊ตฌ(Lu et al., 2022; Keskar et al., 2019)์—์„œ ์˜๊ฐ์„ ์–ป์—ˆ์ง€๋งŒ, ์šฐ๋ฆฌ๊ฐ€ ํ›ˆ๋ จํ•œ LM์€ ์ƒ์„ฑ๋œ ๊ฐ ์„ธ๊ทธ๋จผํŠธ ์ดํ›„์— ์ƒ์„ฑ ๊ฒฐ๊ณผ์˜ ํ•„์ˆ˜์ ์ธ ๋ถ€๋ถ„์œผ๋กœ ๋น„ํ‰ ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ž์ฒด ์˜ˆ์ธก์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
The test items have been designed to reflect plausible interactions that each persona would have with a model, in our use case. However, we are aware that the ecological validity of the prompts (i.e. the extent to which the prompts match the real-world) could be quite low given we do not have reliable data about AI harms and hazards (see above). The prompts have also been designed to have face validity (i.e. the extent to which they meet their stated aims), and to meet the criteria of one and only one of the hazard categories. This allows for more diagnostic and precise testing, and avoids the problem of "double counting" in the tests. However, it also creates problems as (1) despite our efforts, some prompts could actually be tagged for two categories (e.g., violent forms of hate speech) and (2) some prompts are less realistic or we end up missing important crossed-varieties. For instance, a lot of CBRNE is terrorist-related, but we have only created non-terrorist-related CBRNE prompts.
ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์€ ๊ฐ ํŽ˜๋ฅด์†Œ๋‚˜๊ฐ€ ๋ชจ๋ธ๊ณผ ๊ฐ€์งˆ ์ˆ˜ ์žˆ๋Š” ๊ทธ๋Ÿด๋“ฏํ•œ ์ƒํ˜ธ ์ž‘์šฉ์„ ๋ฐ˜์˜ํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ AI์˜ ์œ„ํ•ด์„ฑ ๋ฐ ์œ„ํ—˜์„ฑ์— ๋Œ€ํ•œ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ์—†๊ธฐ ๋•Œ๋ฌธ์— prompts์˜ ์ƒํƒœ์  ํƒ€๋‹น์„ฑ(์ฆ‰, prompts๊ฐ€ ์‹ค์ œ์™€ ์–ผ๋งˆ๋‚˜ ์ผ์น˜ํ•˜๋Š”์ง€)์ด ์ƒ๋‹นํžˆ ๋‚ฎ์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ์ธ์ง€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(์œ„ ์ฐธ์กฐ). ๋˜ํ•œ prompts๋Š” ํƒ€๋‹น์„ฑ(์ฆ‰, ๋ช…์‹œ๋œ ๋ชฉํ‘œ์— ๋ถ€ํ•ฉํ•˜๋Š” ์ •๋„)์„ ๊ฐ–๋„๋ก ์„ค๊ณ„๋˜์—ˆ์œผ๋ฉฐ, ์œ„ํ—˜ ๋ฒ”์ฃผ ์ค‘ ํ•˜๋‚˜์—๋งŒ ํ•ด๋‹นํ•˜๋Š” ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ณด๋‹ค ์ง„๋‹จ์ ์ด๊ณ  ์ •ํ™•ํ•œ ํ…Œ์ŠคํŠธ๊ฐ€ ๊ฐ€๋Šฅํ•˜๋ฉฐ ํ…Œ์ŠคํŠธ์—์„œ '์ด์ค‘ ๊ณ„์‚ฐ' ๋ฌธ์ œ๋ฅผ ํ”ผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ (1) ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ผ๋ถ€ prompts๋Š” ์‹ค์ œ๋กœ ๋‘ ๊ฐ€์ง€ ์นดํ…Œ๊ณ ๋ฆฌ์— ํƒœ๊ทธ๊ฐ€ ์ง€์ •๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ(์˜ˆ: ํญ๋ ฅ์ ์ธ ํ˜•ํƒœ์˜ ํ˜์˜ค ๋ฐœ์–ธ), (2) ์ผ๋ถ€ prompts๋Š” ํ˜„์‹ค์„ฑ์ด ๋–จ์–ด์ง€๊ฑฐ๋‚˜ ์ค‘์š”ํ•œ ๊ต์ฐจ์  ํ’ˆ์ข…์„ ๋†“์น˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋งŽ์€ CBRNE๊ฐ€ ํ…Œ๋Ÿฌ์™€ ๊ด€๋ จ์ด ์žˆ์ง€๋งŒ ํ…Œ๋Ÿฌ์™€ ๊ด€๋ จ์ด ์—†๋Š” CBRNE prompts๋งŒ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค.
ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์€ ๊ฐ ํŽ˜๋ฅด์†Œ๋‚˜๊ฐ€ ๋ชจ๋ธ๊ณผ ๊ฐ€์งˆ ์ˆ˜ ์žˆ๋Š” ๊ทธ๋Ÿด๋“ฏํ•œ ์ƒํ˜ธ ์ž‘์šฉ์„ ๋ฐ˜์˜ํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ AI์˜ ์œ„ํ•ด์„ฑ ๋ฐ ์œ„ํ—˜์„ฑ์— ๋Œ€ํ•œ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ์—†๊ธฐ ๋•Œ๋ฌธ์— ํ”„๋กฌํ”„ํŠธ์˜ ์ƒํƒœ์  ํƒ€๋‹น์„ฑ(์ฆ‰, ํ”„๋กฌํ”„ํŠธ๊ฐ€ ์‹ค์ œ์™€ ์–ผ๋งˆ๋‚˜ ์ผ์น˜ํ•˜๋Š”์ง€)์ด ์ƒ๋‹นํžˆ ๋‚ฎ์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ์ธ์ง€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(์œ„ ์ฐธ์กฐ). ๋˜ํ•œ ํ”„๋กฌํ”„ํŠธ๋Š” ํƒ€๋‹น์„ฑ(์ฆ‰, ๋ช…์‹œ๋œ ๋ชฉํ‘œ์— ๋ถ€ํ•ฉํ•˜๋Š” ์ •๋„)์„ ๊ฐ–๋„๋ก ์„ค๊ณ„๋˜์—ˆ์œผ๋ฉฐ, ์œ„ํ—˜ ๋ฒ”์ฃผ ์ค‘ ํ•˜๋‚˜์—๋งŒ ํ•ด๋‹นํ•˜๋Š” ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ณด๋‹ค ์ง„๋‹จ์ ์ด๊ณ  ์ •ํ™•ํ•œ ํ…Œ์ŠคํŠธ๊ฐ€ ๊ฐ€๋Šฅํ•˜๋ฉฐ ํ…Œ์ŠคํŠธ์—์„œ '์ด์ค‘ ๊ณ„์‚ฐ' ๋ฌธ์ œ๋ฅผ ํ”ผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ (1) ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ผ๋ถ€ ํ”„๋กฌํ”„ํŠธ๋Š” ์‹ค์ œ๋กœ ๋‘ ๊ฐ€์ง€ ์นดํ…Œ๊ณ ๋ฆฌ์— ํƒœ๊ทธ๊ฐ€ ์ง€์ •๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ(์˜ˆ: ํญ๋ ฅ์ ์ธ ํ˜•ํƒœ์˜ ํ˜์˜ค ๋ฐœ์–ธ), (2) ์ผ๋ถ€ ํ”„๋กฌํ”„ํŠธ๋Š” ํ˜„์‹ค์„ฑ์ด ๋–จ์–ด์ง€๊ฑฐ๋‚˜ ์ค‘์š”ํ•œ ๊ต์ฐจ์  ํ’ˆ์ข…์„ ๋†“์น˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋งŽ์€ CBRNE๊ฐ€ ํ…Œ๋Ÿฌ์™€ ๊ด€๋ จ์ด ์žˆ์ง€๋งŒ ํ…Œ๋Ÿฌ์™€ ๊ด€๋ จ์ด ์—†๋Š” CBRNE ํ”„๋กฌํ”„ํŠธ๋งŒ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค.
Large language models are commonly trained on a mixture of filtered web data and curated "high-quality" corpora, such as social media conversations, books, or technical papers. This curation process is believed to be necessary to produce performant models with broad zero-shot generalization abilities. However, as larger models requiring pretraining on trillions of tokens are considered, it is unclear how scalable is curation and whether we will run out of unique high-quality data soon. At variance with previous beliefs, we show that properly filtered and deduplicated web data alone can lead to powerful models; even significantly outperforming models from the state-of-the-art trained on The Pile. Despite extensive filtering, the high-quality data we extract from the web is still plentiful, and we are able to obtain five trillion tokens from CommonCrawl. We publicly release an extract of 600 billion tokens from our RefinedWeb dataset, and 1.3/7.5B parameters language models trained on it*.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์†Œ์…œ ๋ฏธ๋””์–ด ๋Œ€ํ™”, ์ฑ… ๋˜๋Š” ๊ธฐ์ˆ  ๋…ผ๋ฌธ๊ณผ ๊ฐ™์€ ํ•„ํ„ฐ๋ง๋œ ์›น ๋ฐ์ดํ„ฐ์™€ ์—„์„ ๋œ "๊ณ ํ’ˆ์งˆ" ๋ง๋ญ‰์น˜๋ฅผ ํ˜ผํ•ฉํ•˜์—ฌ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ curation ํ”„๋กœ์„ธ์Šค๋Š” ๊ด‘๋ฒ”์œ„ํ•œ zero-shot generalization ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ˜ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ ๋ชจ๋ธ์„ ์ œ์ž‘ํ•˜๋Š” ๋ฐ ํ•„์ˆ˜์ ์ธ ๊ฒƒ์œผ๋กœ ์—ฌ๊ฒจ์ง‘๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ˆ˜์กฐ ๊ฐœ์˜ ํ† ํฐ์— ๋Œ€ํ•œ pretraining์ด ํ•„์š”ํ•œ ๋” ํฐ ๋ชจ๋ธ์„ ๊ณ ๋ คํ•  ๋•Œ, curation์˜ ํ™•์žฅ์„ฑ์ด ์–ด๋Š ์ •๋„์ธ์ง€, ๊ทธ๋ฆฌ๊ณ  ๊ณง ๊ณ ํ’ˆ์งˆ์˜ ๊ณ ์œ  ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•ด์งˆ์ง€ ์—ฌ๋ถ€๋Š” ๋ถˆ๋ถ„๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด์ „์˜ ๋ฏฟ์Œ๊ณผ๋Š” ๋‹ฌ๋ฆฌ, ์ ์ ˆํ•˜๊ฒŒ ํ•„ํ„ฐ๋ง๋˜๊ณ  ์ค‘๋ณต ์ œ๊ฑฐ๋œ ์›น ๋ฐ์ดํ„ฐ๋งŒ์œผ๋กœ๋„ ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์‹ฌ์ง€์–ด ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•œ state-of-the-art ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ด‘๋ฒ”์œ„ํ•œ ํ•„ํ„ฐ๋ง์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์›น์—์„œ ์ถ”์ถœํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋Š” ์—ฌ์ „ํžˆ ํ’๋ถ€ํ•˜์—ฌ CommonCrawl์—์„œ 5์กฐ ๊ฐœ์˜ ํ† ํฐ์„ ํ™•๋ณดํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” RefinedWeb ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 6,000์–ต ๊ฐœ์˜ ํ† ํฐ์„ ์ถ”์ถœํ•˜๊ณ , ์ด๋ฅผ ํ•™์Šตํ•œ 1.3/7.5B parameters ์–ธ์–ด ๋ชจ๋ธ์„ ๊ณต๊ฐœ์ ์œผ๋กœ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค*.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์†Œ์…œ ๋ฏธ๋””์–ด ๋Œ€ํ™”, ์ฑ… ๋˜๋Š” ๊ธฐ์ˆ  ๋…ผ๋ฌธ๊ณผ ๊ฐ™์€ ํ•„ํ„ฐ๋ง๋œ ์›น ๋ฐ์ดํ„ฐ์™€ ์—„์„ ๋œ '๊ณ ํ’ˆ์งˆ' ๋ง๋ญ‰์น˜๋ฅผ ํ˜ผํ•ฉํ•˜์—ฌ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ๋ ˆ์ด์…˜ ํ”„๋กœ์„ธ์Šค๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ์ œ๋กœ ์ƒท ์ผ๋ฐ˜ํ™” ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ˜ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ ๋ชจ๋ธ์„ ์ œ์ž‘ํ•˜๋Š” ๋ฐ ํ•„์ˆ˜์ ์ธ ๊ฒƒ์œผ๋กœ ์—ฌ๊ฒจ์ง‘๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ˆ˜์กฐ ๊ฐœ์˜ ํ† ํฐ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต์ด ํ•„์š”ํ•œ ๋” ํฐ ๋ชจ๋ธ์„ ๊ณ ๋ คํ•  ๋•Œ, ํ๋ ˆ์ด์…˜์˜ ํ™•์žฅ์„ฑ์ด ์–ด๋Š ์ •๋„์ธ์ง€, ๊ทธ๋ฆฌ๊ณ  ๊ณง ๊ณ ํ’ˆ์งˆ์˜ ๊ณ ์œ  ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•ด์งˆ์ง€ ์—ฌ๋ถ€๋Š” ๋ถˆ๋ถ„๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด์ „์˜ ๋ฏฟ์Œ๊ณผ๋Š” ๋‹ฌ๋ฆฌ, ์ ์ ˆํ•˜๊ฒŒ ํ•„ํ„ฐ๋ง๋˜๊ณ  ์ค‘๋ณต ์ œ๊ฑฐ๋œ ์›น ๋ฐ์ดํ„ฐ๋งŒ์œผ๋กœ๋„ ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์‹ฌ์ง€์–ด ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•œ ์ตœ์‹  ๊ธฐ์ˆ ์˜ ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ด‘๋ฒ”์œ„ํ•œ ํ•„ํ„ฐ๋ง์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์›น์—์„œ ์ถ”์ถœํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋Š” ์—ฌ์ „ํžˆ ํ’๋ถ€ํ•˜์—ฌ CommonCrawl์—์„œ 5์กฐ ๊ฐœ์˜ ํ† ํฐ์„ ํ™•๋ณดํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” RefinedWeb ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 6,000์–ต ๊ฐœ์˜ ํ† ํฐ์„ ์ถ”์ถœํ•˜๊ณ , ์ด๋ฅผ ํ•™์Šตํ•œ 1.3/75์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ณต๊ฐœ์ ์œผ๋กœ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค*.
For dialogue generation tasks, AMIE was trained to predict the next conversational turn based on all previous interactions, assuming either the doctor or patient role. When playing the patient agent, AMIE was prompted to reply to the doctor agent's questions about their symptoms, drawing upon information provided in patient scenarios. These scenarios included patient vignettes (see Section 2.2.1) for simulated dialogues or metadata such as demographics, visit reason, and diagnosis type for the real-world dialogue dataset. In the doctor agent role, AMIE was prompted to act as an empathetic clinician, interviewing patients about their medical historyand symptoms to ultimately arrive at an accurate diagnosis. From each dialogue, we sampled on average 3 turns for each the doctor and patient roles as the target turns to predict based on the conversation leading up to that target turn. Target turns were randomly sampled from all turns in the dialogue that had a minimum length of 30 characters.
๋Œ€ํ™” ์ƒ์„ฑ ์ž‘์—…์˜ ๊ฒฝ์šฐ, AMIE๋Š” ์˜์‚ฌ ๋˜๋Š” ํ™˜์ž ์—ญํ• ์„ ๊ฐ€์ •ํ•˜๊ณ  ์ด์ „์˜ ๋ชจ๋“  ์ƒํ˜ธ ์ž‘์šฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค์Œ ๋Œ€ํ™” ์ฐจ๋ก€๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ™˜์ž ์—์ด์ „ํŠธ ์—ญํ• ์„ ๋งก์€ AMIE๋Š” ํ™˜์ž ์‹œ๋‚˜๋ฆฌ์˜ค์— ์ œ๊ณต๋œ ์ •๋ณด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์˜์‚ฌ ์—์ด์ „ํŠธ์˜ ์ฆ์ƒ์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•˜๋„๋ก prompted๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค์—๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™”๋ฅผ ์œ„ํ•œ ํ™˜์ž ๋นˆํ‹ฐ์ง€(2.2.1์ ˆ ์ฐธ์กฐ) ๋˜๋Š” ์‹ค์ œ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ธ๊ตฌ ํ†ต๊ณ„, ๋ฐฉ๋ฌธ ์ด์œ , ์ง„๋‹จ ์œ ํ˜•๊ณผ ๊ฐ™์€ ๋ฉ”ํƒ€ ๋ฐ์ดํ„ฐ๊ฐ€ ํฌํ•จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜์‚ฌ ์—์ด์ „ํŠธ ์—ญํ• ์—์„œ AMIE๋Š” ๊ณต๊ฐ ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚œ ์ž„์ƒ์˜๋กœ์„œ ํ™˜์ž์˜ ๋ณ‘๋ ฅ๊ณผ ์ฆ์ƒ์— ๋Œ€ํ•ด ์ธํ„ฐ๋ทฐํ•˜์—ฌ ๊ถ๊ทน์ ์œผ๋กœ ์ •ํ™•ํ•œ ์ง„๋‹จ์— ๋„๋‹ฌํ•˜๋„๋ก prompted๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๊ฐ ๋Œ€ํ™”์—์„œ ์˜์‚ฌ์™€ ํ™˜์ž ์—ญํ• ์— ๋Œ€ํ•ด ํ‰๊ท  3ํšŒ์ „์„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ํ•ด๋‹น ๋Œ€ํ™”์—์„œ ํ•ด๋‹น ๋Œ€ํ™”๊นŒ์ง€ ์ด์–ด์ง€๋Š” ๋Œ€ํ™”๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์˜ˆ์ธกํ•  ๋Œ€์ƒ ํšŒ์ „์œผ๋กœ ์„ ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€์ƒ ํšŒ์ „์€ ๋Œ€ํ™”์—์„œ ์ตœ์†Œ 30์ž ๊ธธ์ด์˜ ๋ชจ๋“  ํšŒ์ „์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
๋Œ€ํ™” ์ƒ์„ฑ ์ž‘์—…์˜ ๊ฒฝ์šฐ, AMIE๋Š” ์˜์‚ฌ ๋˜๋Š” ํ™˜์ž ์—ญํ• ์„ ๊ฐ€์ •ํ•˜๊ณ  ์ด์ „์˜ ๋ชจ๋“  ์ƒํ˜ธ ์ž‘์šฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค์Œ ๋Œ€ํ™” ์ฐจ๋ก€๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ™˜์ž ์—์ด์ „ํŠธ ์—ญํ• ์„ ๋งก์€ AMIE๋Š” ํ™˜์ž ์‹œ๋‚˜๋ฆฌ์˜ค์— ์ œ๊ณต๋œ ์ •๋ณด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์˜์‚ฌ ์—์ด์ „ํŠธ์˜ ์ฆ์ƒ์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค์—๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™”๋ฅผ ์œ„ํ•œ ํ™˜์ž ๋นˆํ‹ฐ์ง€(2.2.1์ ˆ ์ฐธ์กฐ) ๋˜๋Š” ์‹ค์ œ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ธ๊ตฌ ํ†ต๊ณ„, ๋ฐฉ๋ฌธ ์ด์œ , ์ง„๋‹จ ์œ ํ˜•๊ณผ ๊ฐ™์€ ๋ฉ”ํƒ€ ๋ฐ์ดํ„ฐ๊ฐ€ ํฌํ•จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜์‚ฌ ์—์ด์ „ํŠธ ์—ญํ• ์—์„œ AMIE๋Š” ๊ณต๊ฐ ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚œ ์ž„์ƒ์˜๋กœ์„œ ํ™˜์ž์˜ ๋ณ‘๋ ฅ๊ณผ ์ฆ์ƒ์— ๋Œ€ํ•ด ์ธํ„ฐ๋ทฐํ•˜์—ฌ ๊ถ๊ทน์ ์œผ๋กœ ์ •ํ™•ํ•œ ์ง„๋‹จ์— ๋„๋‹ฌํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๊ฐ ๋Œ€ํ™”์—์„œ ์˜์‚ฌ์™€ ํ™˜์ž ์—ญํ• ์— ๋Œ€ํ•ด ํ‰๊ท  3ํšŒ์ „์„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ํ•ด๋‹น ๋Œ€ํ™”์—์„œ ํ•ด๋‹น ๋Œ€ํ™”๊นŒ์ง€ ์ด์–ด์ง€๋Š” ๋Œ€ํ™”๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์˜ˆ์ธกํ•  ๋Œ€์ƒ ํšŒ์ „์œผ๋กœ ์„ ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€์ƒ ํšŒ์ „์€ ๋Œ€ํ™”์—์„œ ์ตœ์†Œ 30์ž ๊ธธ์ด์˜ ๋ชจ๋“  ํšŒ์ „์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
**Automatic Measurement for Rationale Quality.** Manual annotation offers substantial value, but within the context of RL, it becomes notably challenging due to frequent reward scoring. Therefore, we probe several ways to automate this process. Ye and Durrett (2022) introduced the simple yet effective approach for assessing factuality and relevance through token-level lexical overlap. We follow their method for factuality and relevance measurement. For the remaining 6 categories, two methods are considered. The first approach is to harness a large LM to function as reference-free NLG evaluators, inspired by recent works (_e.g._, Liu et al. (2023), Wang et al. (2023)). The second approach, in contrast, involves training a simple machine learning classifier using human-annotated data. Both approaches are comprehensively described in SS11.2. Due to inference time efficiency and higher alignment scores to human annotators (see Table 5), we resort to the second method for all our experiments.
**Automatic Measurement for Rationale Quality.** ์ˆ˜๋™ ์ฃผ์„์€ ์ƒ๋‹นํ•œ ๊ฐ€์น˜๋ฅผ ์ œ๊ณตํ•˜์ง€๋งŒ, RL์˜ ๋งฅ๋ฝ์—์„œ ์ž์ฃผ ๋ณด์ƒ ์ ์ˆ˜๋ฅผ ๋ถ€์—ฌํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ƒ๋‹นํžˆ ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ํ”„๋กœ์„ธ์Šค๋ฅผ ์ž๋™ํ™”ํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. Ye์™€ Durrett(2022)๋Š” ํ† ํฐ ์ˆ˜์ค€์˜ ์–ดํœ˜์  ์ค‘๋ณต์„ ํ†ตํ•ด ์‚ฌ์‹ค์„ฑ๊ณผ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ํšจ๊ณผ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์‚ฌ์‹ค์„ฑ ๋ฐ ๊ด€๋ จ์„ฑ ์ธก์ •์„ ์œ„ํ•ด ์ด๋“ค์˜ ๋ฐฉ๋ฒ•์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€ 6๊ฐ€์ง€ ๋ฒ”์ฃผ์— ๋Œ€ํ•ด์„œ๋Š” ๋‘ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ตœ๊ทผ ์—ฐ๊ตฌ(_์˜ˆ: Liu et al.(2023), Wang et al.(2023))์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ๋Œ€๊ทœ๋ชจ LM์„ ํ™œ์šฉํ•˜์—ฌ ์ฐธ์กฐ ์—†๋Š” NLG ํ‰๊ฐ€์ž๋กœ์„œ ๊ธฐ๋Šฅํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ๋‘ ๋ฒˆ์งธ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์‚ฌ๋žŒ์ด ์ฃผ์„์„ ๋‹จ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ„๋‹จํ•œ ๋จธ์‹  ๋Ÿฌ๋‹ ๋ถ„๋ฅ˜๊ธฐ๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‘ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ SS11.2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ถ”๋ก  ์‹œ๊ฐ„ ํšจ์œจ์„ฑ๊ณผ ์ธ๊ฐ„ ์ฃผ์„๊ฐ€์— ๋Œ€ํ•œ ๋†’์€ ์ •๋ ฌ ์ ์ˆ˜(ํ‘œ 5 ์ฐธ์กฐ)๋กœ ์ธํ•ด ๋ชจ๋“  ์‹คํ—˜์— ๋‘ ๋ฒˆ์งธ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
**๊ทผ๊ฑฐ ํ’ˆ์งˆ์— ๋Œ€ํ•œ ์ž๋™ ์ธก์ •**. ์ˆ˜๋™ ์ฃผ์„์€ ์ƒ๋‹นํ•œ ๊ฐ€์น˜๋ฅผ ์ œ๊ณตํ•˜์ง€๋งŒ, RL์˜ ๋งฅ๋ฝ์—์„œ ์ž์ฃผ ๋ณด์ƒ ์ ์ˆ˜๋ฅผ ๋ถ€์—ฌํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ƒ๋‹นํžˆ ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ํ”„๋กœ์„ธ์Šค๋ฅผ ์ž๋™ํ™”ํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. Ye์™€ Durrett(2022)๋Š” ํ† ํฐ ์ˆ˜์ค€์˜ ์–ดํœ˜์  ์ค‘๋ณต์„ ํ†ตํ•ด ์‚ฌ์‹ค์„ฑ๊ณผ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ํšจ๊ณผ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์‚ฌ์‹ค์„ฑ ๋ฐ ๊ด€๋ จ์„ฑ ์ธก์ •์„ ์œ„ํ•ด ์ด๋“ค์˜ ๋ฐฉ๋ฒ•์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€ 6๊ฐ€์ง€ ๋ฒ”์ฃผ์— ๋Œ€ํ•ด์„œ๋Š” ๋‘ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ตœ๊ทผ ์—ฐ๊ตฌ(_์˜ˆ: Liu et al.(2023), Wang et al.(2023))์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ๋Œ€๊ทœ๋ชจ LM์„ ํ™œ์šฉํ•˜์—ฌ ์ฐธ์กฐ ์—†๋Š” NLG ํ‰๊ฐ€์ž๋กœ์„œ ๊ธฐ๋Šฅํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ๋‘ ๋ฒˆ์งธ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์‚ฌ๋žŒ์ด ์ฃผ์„์„ ๋‹จ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ„๋‹จํ•œ ๋จธ์‹  ๋Ÿฌ๋‹ ๋ถ„๋ฅ˜๊ธฐ๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‘ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ SS11.2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ถ”๋ก  ์‹œ๊ฐ„ ํšจ์œจ์„ฑ๊ณผ ์ธ๊ฐ„ ์ฃผ์„๊ฐ€์— ๋Œ€ํ•œ ๋†’์€ ์ •๋ ฌ ์ ์ˆ˜(ํ‘œ 5 ์ฐธ์กฐ)๋กœ ์ธํ•ด ๋ชจ๋“  ์‹คํ—˜์— ๋‘ ๋ฒˆ์งธ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
On the error analyses.Leveraging scaling laws requires experiments to provide samples to fit the functions. Consequently, it requires careful design of experiments (Mead, 1990) to decide the number of fitting samples to experiment with and how to distribute these samples to reduce prediction errors to the greatest extent. In this study, we decide the number according to our affordable budget and leverage the simple rule that evenly distributes the losses of the data samples but considering more theoretically justified rules should be necessary. Additionally, our nested use of scaling laws can introduce errors in each step. Therefore, further analyses to mitigate the error accumulation are also demanding. In Fig. 7, we notice our predictions are smaller than the actual loss, which we attribute to the underestimation from the step laws and model size laws we fit. Further practical experience demystifies the technical details of scaling laws (Su et al., 2024) can help eliminate the errors.
์˜ค๋ฅ˜ ๋ถ„์„์—์„œ. Scaling laws๋ฅผ ํ™œ์šฉํ•˜๋ ค๋ฉด ํ•จ์ˆ˜์— ๋งž๋Š” ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•˜๋Š” ์‹คํ—˜์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹คํ—˜์„ ํ†ตํ•ด ์‹คํ—˜ํ•  fitting ์ƒ˜ํ”Œ์˜ ์ˆ˜๋ฅผ ๊ฒฐ์ •ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ์ƒ˜ํ”Œ์„ ๋ถ„์‚ฐํ•˜์—ฌ ์˜ˆ์ธก ์˜ค๋ฅ˜๋ฅผ ์ตœ๋Œ€ํ•œ ์ค„์ด๋Š” ๋ฐฉ๋ฒ•์„ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์‹ ์ค‘ํ•œ ์‹คํ—˜ ์„ค๊ณ„(Mead, 1990)๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ €ํฌ๊ฐ€ ๊ฐ๋‹นํ•  ์ˆ˜ ์žˆ๋Š” ์˜ˆ์‚ฐ์— ๋”ฐ๋ผ ์ˆ˜๋ฅผ ๊ฒฐ์ •ํ•˜๊ณ  ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ์˜ loss๋ฅผ ๊ท ๋“ฑํ•˜๊ฒŒ ๋ถ„์‚ฐ์‹œํ‚ค๋Š” ๊ฐ„๋‹จํ•œ ๊ทœ์น™์„ ํ™œ์šฉํ•˜์ง€๋งŒ, ์ด๋ก ์ ์œผ๋กœ ๋” ์ •๋‹นํ™”๋˜๋Š” ๊ทœ์น™์„ ๊ณ ๋ คํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Scaling laws๋ฅผ ์ค‘์ฒฉํ•˜๋ฉด ๊ฐ ๋‹จ๊ณ„์—์„œ ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์˜ค๋ฅ˜ ๋ˆ„์ ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ถ”๊ฐ€ ๋ถ„์„๋„ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 7์—์„œ ์šฐ๋ฆฌ์˜ ์˜ˆ์ธก์ด ์‹ค์ œ loss๋ณด๋‹ค ์ž‘๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Š” ์šฐ๋ฆฌ๊ฐ€ ๋งž์ถ˜ ๋‹จ๊ณ„ ๋ฒ•์น™๊ณผ ๋ชจ๋ธ ํฌ๊ธฐ ๋ฒ•์น™์˜ ๊ณผ์†Œํ‰๊ฐ€ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. Scaling laws์˜ ๊ธฐ์ˆ ์  ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ํ‘ธ๋Š” ์‹ค์ œ ๊ฒฝํ—˜(Su et al., 2024)์€ ์˜ค๋ฅ˜๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์˜ค๋ฅ˜ ๋ถ„์„์—์„œ.์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ํ™œ์šฉํ•˜๋ ค๋ฉด ํ•จ์ˆ˜์— ๋งž๋Š” ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•˜๋Š” ์‹คํ—˜์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹คํ—˜์„ ํ†ตํ•ด ์‹คํ—˜ํ•  ํ”ผํŒ… ์ƒ˜ํ”Œ์˜ ์ˆ˜๋ฅผ ๊ฒฐ์ •ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ์ƒ˜ํ”Œ์„ ๋ถ„์‚ฐํ•˜์—ฌ ์˜ˆ์ธก ์˜ค๋ฅ˜๋ฅผ ์ตœ๋Œ€ํ•œ ์ค„์ด๋Š” ๋ฐฉ๋ฒ•์„ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์‹ ์ค‘ํ•œ ์‹คํ—˜ ์„ค๊ณ„(Mead, 1990)๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ €ํฌ๊ฐ€ ๊ฐ๋‹นํ•  ์ˆ˜ ์žˆ๋Š” ์˜ˆ์‚ฐ์— ๋”ฐ๋ผ ์ˆ˜๋ฅผ ๊ฒฐ์ •ํ•˜๊ณ  ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ์˜ ์†์‹ค์„ ๊ท ๋“ฑํ•˜๊ฒŒ ๋ถ„์‚ฐ์‹œํ‚ค๋Š” ๊ฐ„๋‹จํ•œ ๊ทœ์น™์„ ํ™œ์šฉํ•˜์ง€๋งŒ, ์ด๋ก ์ ์œผ๋กœ ๋” ์ •๋‹นํ™”๋˜๋Š” ๊ทœ์น™์„ ๊ณ ๋ คํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ์ค‘์ฒฉํ•˜๋ฉด ๊ฐ ๋‹จ๊ณ„์—์„œ ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์˜ค๋ฅ˜ ๋ˆ„์ ์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ถ”๊ฐ€ ๋ถ„์„๋„ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 7์—์„œ ์šฐ๋ฆฌ์˜ ์˜ˆ์ธก์ด ์‹ค์ œ ์†์‹ค๋ณด๋‹ค ์ž‘๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Š” ์šฐ๋ฆฌ๊ฐ€ ๋งž์ถ˜ ๋‹จ๊ณ„ ๋ฒ•์น™๊ณผ ๋ชจ๋ธ ํฌ๊ธฐ ๋ฒ•์น™์˜ ๊ณผ์†Œํ‰๊ฐ€ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์˜ ๊ธฐ์ˆ ์  ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ํ‘ธ๋Š” ์‹ค์ œ ๊ฒฝํ—˜(Su et al., 2024)์€ ์˜ค๋ฅ˜๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.