chosen
stringlengths
14
4.26k
rejected
stringlengths
53
9.3k
prompt
stringlengths
178
4.55k
์—ฌ๊ธฐ์—์„œ๋Š” 252๊ฐœ์˜ ์‚ฌ์šฉ์ž ์ง€ํ–ฅ instructions์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์‘๋‹ต์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด SS4.4์—์„œ ์„ค๋ช…ํ•œ ์‚ฌ๋žŒ์˜ ํ‰๊ฐ€์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ถฉ์‹คํ•˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ด๋Ÿฌํ•œ instructions(๋ฐ ์ด ๋…ผ๋ฌธ)์˜ ๋‘ ์ €์ž์—๊ฒŒ ๋ชจ๋ธ ์˜ˆ์ธก์„ ํŒ๋‹จํ•˜๋„๋ก ์š”์ฒญํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋‘ ๋ช…์˜ ํ‰๊ฐ€์ž๋Š” ์ฃผ์„์„ ์‹œ์ž‘ํ•˜๊ธฐ ์ „์— 4๋‹จ๊ณ„ ๋“ฑ๊ธ‰ ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ๊ธฐ์ค€์„ ์กฐ์œจํ•œ ๋‹ค์Œ ๊ฐ์ž๊ฐ€ ๋ชจ๋“  ์ธ์Šคํ„ด์Šค๋ฅผ ๋…๋ฆฝ์ ์œผ๋กœ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€์ž๋“ค์€ instruction, ์ธ์Šคํ„ด์Šค ์ž…๋ ฅ, ๋Œ€์ƒ ์ถœ๋ ฅ(์ฐธ์กฐ์šฉ) ๋ฐ ๋ชจ๋ธ ์‘๋‹ต์„ ์ œ์‹œ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ์‘๋‹ต์€ ๋ฌด์ž‘์œ„ ์ˆœ์„œ๋กœ ๋‚˜์—ด๋˜๋ฉฐ ๋ชจ๋“  ๋ชจ๋ธ ์ •๋ณด๋Š” ์ต๋ช…์œผ๋กœ ์ฒ˜๋ฆฌ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 9๋Š” ์ฃผ์„ ์ธํ„ฐํŽ˜์ด์Šค์˜ ์Šคํฌ๋ฆฐ์ƒท์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ ๋ณด๊ณ ๋œ ์„ฑ๋Šฅ์€ ํ‰๊ฐ€์ž ์ค‘ ํ•œ ๋ช…์˜ ๊ฒฐ๊ณผ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ, ๋‹ค๋ฅธ ํ‰๊ฐ€์ž์˜ ๊ฒฐ๊ณผ์—์„œ๋„ ๋™์ผํ•œ ์ถ”์„ธ๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค.[|endofturn|]
์—ฌ๊ธฐ์—์„œ๋Š” 252๊ฐœ์˜ ์‚ฌ์šฉ์ž ์ง€ํ–ฅ ๋ช…๋ น์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์‘๋‹ต์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด SS4.4์—์„œ ์„ค๋ช…ํ•œ ์‚ฌ๋žŒ์˜ ํ‰๊ฐ€์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ถฉ์‹คํ•˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ด๋Ÿฌํ•œ ์ง€์นจ(๋ฐ ์ด ๋…ผ๋ฌธ)์˜ ๋‘ ์ €์ž์—๊ฒŒ ๋ชจ๋ธ ์˜ˆ์ธก์„ ํŒ๋‹จํ•˜๋„๋ก ์š”์ฒญํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋‘ ๋ช…์˜ ํ‰๊ฐ€์ž๋Š” ์ฃผ์„์„ ์‹œ์ž‘ํ•˜๊ธฐ ์ „์— 4๋‹จ๊ณ„ ๋“ฑ๊ธ‰ ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ๊ธฐ์ค€์„ ์กฐ์œจํ•œ ๋‹ค์Œ ๊ฐ์ž๊ฐ€ ๋ชจ๋“  ์ธ์Šคํ„ด์Šค๋ฅผ ๋…๋ฆฝ์ ์œผ๋กœ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ฐ€์ž๋“ค์€ ์ง€์นจ, ์ธ์Šคํ„ด์Šค ์ž…๋ ฅ, ๋Œ€์ƒ ์ถœ๋ ฅ(์ฐธ์กฐ์šฉ) ๋ฐ ๋ชจ๋ธ ์‘๋‹ต์„ ์ œ์‹œ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ์‘๋‹ต์€ ๋ฌด์ž‘์œ„ ์ˆœ์„œ๋กœ ๋‚˜์—ด๋˜๋ฉฐ ๋ชจ๋“  ๋ชจ๋ธ ์ •๋ณด๋Š” ์ต๋ช…์œผ๋กœ ์ฒ˜๋ฆฌ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 9๋Š” ์ฃผ์„ ์ธํ„ฐํŽ˜์ด์Šค์˜ ์Šคํฌ๋ฆฐ์ƒท์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ ๋ณด๊ณ ๋œ ์„ฑ๋Šฅ์€ ํ‰๊ฐ€์ž ์ค‘ ํ•œ ๋ช…์˜ ๊ฒฐ๊ณผ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ, ๋‹ค๋ฅธ ํ‰๊ฐ€์ž์˜ ๊ฒฐ๊ณผ์—์„œ๋„ ๋™์ผํ•œ ์ถ”์„ธ๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Here we provide more details for the human evaluation described in SS4.4 for rating the models' responses to the 252 user-oriented instructions. To ensure faithful and reliable evaluation, we asked two authors of these instructions (and of this paper) to judge model predictions. These two evaluators coordinated the standards for the 4-level rating system before starting annotation and then each of them rated all the instances independently. They were presented with the instruction, instance input, target output (as a reference), and model responses. Model responses are listed in random order, with all the model information anonymized. Figure 9 provides a screenshot of the annotation interface. The reported performance in this paper is based on the results from one of the evaluators, and the trends from the other evaluator's results are the same. [|assistant|]
์ด ๋…ผ๋ฌธ์—์„œ๋Š” ํ‘œ ๋‚ด์—์„œ ๊ฐœ์„  ์‚ฌํ•ญ์„ ๊ฐ•์กฐํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜์—ฌ ์ ์ˆ˜ \(x\)์˜ metric \(y\)์—์„œ ์ •์  ๊ฐœ์„  ์ž„๊ณ„๊ฐ’์„ ์ง€์ •ํ•˜๋Š” ํ‘œ์ค€ ๊ด€ํ–‰์„ ๋„˜์–ด์„ฐ์Šต๋‹ˆ๋‹ค. ๋Œ€์‹ , ์šฐ๋ฆฌ์˜ ์ž„๊ณ„๊ฐ’์€ ์ธ๊ฐ„์ด ์ธ์‹ํ•˜๋Š” ๋‘ ์‹œ์Šคํ…œ ๊ฐ„์— ์ธ์ง€ํ•  ์ˆ˜ ์žˆ๋Š” ์ฐจ์ด๋ฅผ ์‚ฐ์ถœํ•˜๋Š” metric \(y\)์˜ ์ตœ์†Œ metric ์ฐจ์ด \(x\)์— ๋งž์ถฐ ๋™์ ์œผ๋กœ ๋ณด์ •๋ฉ๋‹ˆ๋‹ค(Kocmi ์™ธ., 2024). ์˜ˆ๋ฅผ ๋“ค์–ด, 80%์˜ ์ผ์น˜์œจ๋กœ ์ธ๊ฐ„์˜ ํŒ๋‹จ์— ๋งž์ถ”๊ธฐ ์œ„ํ•ด์„œ๋Š” KWI-XXL๊ณผ COMET-XXL์˜ ๊ฒฝ์šฐ \(\geq 1.24\), KWI-22์˜ ๊ฒฝ์šฐ \(\geq 0.53\)์˜ ๊ฐœ์„  ๋งˆ์ง„์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž„๊ณ„๊ฐ’์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ ํ‘œ 13์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ด ๋…ผ๋ฌธ์—์„œ๋Š” ํ‘œ ๋‚ด์—์„œ ๊ฐœ์„  ์‚ฌํ•ญ์„ ๊ฐ•์กฐํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜์—ฌ ์ ์ˆ˜ \(x\)์˜ ๋ฉ”ํŠธ๋ฆญ \(y\)์—์„œ ์ •์  ๊ฐœ์„  ์ž„๊ณ„๊ฐ’์„ ์ง€์ •ํ•˜๋Š” ํ‘œ์ค€ ๊ด€ํ–‰์„ ๋„˜์–ด์„ฐ์Šต๋‹ˆ๋‹ค. ๋Œ€์‹ , ์šฐ๋ฆฌ์˜ ์ž„๊ณ„๊ฐ’์€ ์ธ๊ฐ„์ด ์ธ์‹ํ•˜๋Š” ๋‘ ์‹œ์Šคํ…œ ๊ฐ„์— ์ธ์ง€ํ•  ์ˆ˜ ์žˆ๋Š” ์ฐจ์ด๋ฅผ ์‚ฐ์ถœํ•˜๋Š” ๋ฉ”ํŠธ๋ฆญ \(y\)์˜ ์ตœ์†Œ ๋ฉ”ํŠธ๋ฆญ ์ฐจ์ด \(x\)์— ๋งž์ถฐ ๋™์ ์œผ๋กœ ๋ณด์ •๋ฉ๋‹ˆ๋‹ค(Kocmi ์™ธ., 2024). ์˜ˆ๋ฅผ ๋“ค์–ด, 80%์˜ ์ผ์น˜์œจ๋กœ ์ธ๊ฐ„์˜ ํŒ๋‹จ์— ๋งž์ถ”๊ธฐ ์œ„ํ•ด์„œ๋Š” KWI-XXL๊ณผ COMET-XXL์˜ ๊ฒฝ์šฐ \(\geq 1.24\), KWI-22์˜ ๊ฒฝ์šฐ \(\geq 0.53\)์˜ ๊ฐœ์„  ๋งˆ์ง„์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž„๊ณ„๊ฐ’์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ ํ‘œ 13์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In the paper, we adopt a new approach for highlighting improvements within tables, moving beyond the standard practice of specifying a static improvement threshold in metric \(y\) by score \(x\). Instead, our threshold is dynamic, calibrated to the minimal metric difference \(x\) in metric \(y\) that yields a perceptible distinction between two systems as recognized by humans (Kocmi et al., 2024). For instance, to align with human judgments at an 80% concordance rate, the required improvement margin is \(\geq 1.24\) for both KWI-XXL and COMET-XXL, and \(\geq 0.53\) for KWI-22. A comprehensive delineation of these thresholds can be found in Table 13. [|assistant|]
**์ฃผ์š” ๊ธฐ์—ฌ๋„ ์š”์•ฝ.** ๊ทธ๋ฆผ 3์— ์ผ๋ถ€ ์ฃผ์š” ์ด๋ก ์  ๊ธฐ์—ฌ๋ฅผ ๋†’์€ ์ˆ˜์ค€์œผ๋กœ ์š”์•ฝํ•˜์—ฌ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ด๋ก ์  ์˜ˆ์ธก์„ ์‹ค์ฆ์ ์œผ๋กœ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 4 ์ฐธ์กฐ): (1) Wikitext-103์˜ ์•ฝ \(2M\) ์ƒ˜ํ”Œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด Llama2-7B(Touvron et al., 2023)๋ฅผ Fine-tuningํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ์‹คํ—˜์—์„œ (2) Transformer ๋ชจ๋ธ์ด ์ตœ๋Œ€๊ณต์•ฝ์ˆ˜๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ํ›ˆ๋ จ๋œ ๊ฒฝ์šฐ(Charton, 2023).[|endofturn|]
**์ฃผ์š” ๊ธฐ์—ฌ๋„ ์š”์•ฝ.** ๊ทธ๋ฆผ 3์— ์ผ๋ถ€ ์ฃผ์š” ์ด๋ก ์  ๊ธฐ์—ฌ๋ฅผ ๋†’์€ ์ˆ˜์ค€์œผ๋กœ ์š”์•ฝํ•˜์—ฌ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ด๋ก ์  ์˜ˆ์ธก์„ ์‹ค์ฆ์ ์œผ๋กœ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 4 ์ฐธ์กฐ): (1) ๋น…์Šคํ…์ŠคํŠธ-103์˜ ์•ฝ 2M ์ƒ˜ํ”Œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด Llama2-7B(Touvron et al., 2023)๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ์‹คํ—˜์—์„œ ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ์ด ์ตœ๋Œ€๊ณต์•ฝ์ˆ˜๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ํ›ˆ๋ จ๋œ ๊ฒฝ์šฐ(Charton, 2023).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Summary of Main Contributions.** We present a high-level summary of our main theoretical contributions, some of which are highlighted in Figure 3. We empirically verify these theoretical predictions (see Figure 4): (1) in large-scale experiments on an LLM, fine-tuning Llama2-7B (Touvron et al., 2023) on an approximately \(2M\) sample dataset from Wikitext-103 and (2) for transformer models trained to predict the greatest common divisor (Charton, 2023). [|assistant|]
**์—์ด์ „ํŠธ ํŽ˜๋ฅด์†Œ๋‚˜**. ์—์ด์ „ํŠธ ํŽ˜๋ฅด์†Œ๋‚˜๋Š” ์—์ด์ „ํŠธ๊ฐ€ ํ•ด๋‹น ์—์ด์ „ํŠธ์™€ ๊ด€๋ จ๋œ ๊ธฐํƒ€ instructions๋ฅผ ํฌํ•จํ•˜์—ฌ ๋งก์•„์•ผ ํ•˜๋Š” ์—ญํ•  ๋˜๋Š” ์„ฑ๊ฒฉ์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ํŽ˜๋ฅด์†Œ๋‚˜์—๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋Š” ๋„๊ตฌ์— ๋Œ€ํ•œ ์„ค๋ช…๋„ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํŽ˜๋ฅด์†Œ๋‚˜๋Š” ์—์ด์ „ํŠธ์—๊ฒŒ ์ž์‹ ์˜ ์—ญํ• ๊ณผ ๋„๊ตฌ์˜ ๋ชฉ์ , ๊ทธ๋ฆฌ๊ณ  ์ด๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ™œ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์•Œ๋ ค์ค๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด "ํ˜•์„ฑ๋œ ์„ฑ๊ฒฉ์€ ์†Œ์…œ ๋ฏธ๋””์–ด ๊ฒŒ์‹œ๋ฌผ ์ž‘์„ฑ๊ณผ ๊ฐ™์€ ์ผ๋ฐ˜์ ์ธ downstream(์ฆ‰, ํ›„์†) ์ž‘์—…์—์„œ ๊ทœ๋ฒ”์ ์œผ๋กœ Large Language Model (LLM)์˜ ํ–‰๋™์— ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค"๊ณ  ํ•ฉ๋‹ˆ๋‹ค[21]. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ํŽ˜๋ฅด์†Œ๋‚˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ์†”๋ฃจ์…˜์€ ๋ชจ๋ธ์ด ๊ณ„ํš์„ ๋‹จ๊ณ„๋ณ„๋กœ ์„ธ๋ถ„ํ™”ํ•˜๋„๋ก ์š”์ฒญํ•˜๋Š” Chain-of-Thought (CoT) ํ”„๋กฌํ”„ํŠธ์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค[28, 29].[|endofturn|]
**์—์ด์ „ํŠธ ํŽ˜๋ฅด์†Œ๋‚˜**. ์—์ด์ „ํŠธ ํŽ˜๋ฅด์†Œ๋‚˜๋Š” ์—์ด์ „ํŠธ๊ฐ€ ํ•ด๋‹น ์—์ด์ „ํŠธ์™€ ๊ด€๋ จ๋œ ๊ธฐํƒ€ ์ง€์นจ์„ ํฌํ•จํ•˜์—ฌ ๋งก์•„์•ผ ํ•˜๋Š” ์—ญํ•  ๋˜๋Š” ์„ฑ๊ฒฉ์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ํŽ˜๋ฅด์†Œ๋‚˜์—๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋Š” ๋„๊ตฌ์— ๋Œ€ํ•œ ์„ค๋ช…๋„ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํŽ˜๋ฅด์†Œ๋‚˜๋Š” ์—์ด์ „ํŠธ์—๊ฒŒ ์ž์‹ ์˜ ์—ญํ• ๊ณผ ๋„๊ตฌ์˜ ๋ชฉ์ , ๊ทธ๋ฆฌ๊ณ  ์ด๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ™œ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์•Œ๋ ค์ค๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด "ํ˜•์„ฑ๋œ ์„ฑ๊ฒฉ์€ ์†Œ์…œ ๋ฏธ๋””์–ด ๊ฒŒ์‹œ๋ฌผ ์ž‘์„ฑ๊ณผ ๊ฐ™์€ ์ผ๋ฐ˜์ ์ธ ๋‹ค์šด์ŠคํŠธ๋ฆผ(์ฆ‰, ํ›„์†) ์ž‘์—…์—์„œ ๊ทœ๋ฒ”์ ์œผ๋กœ ๋น… ์–ธ์–ด ๋ชจ๋ธ(LLM)์˜ ํ–‰๋™์— ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค"๊ณ  ํ•ฉ๋‹ˆ๋‹ค[21]. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ํŽ˜๋ฅด์†Œ๋‚˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ์†”๋ฃจ์…˜์€ ๋ชจ๋ธ์ด ๊ณ„ํš์„ ๋‹จ๊ณ„๋ณ„๋กœ ์„ธ๋ถ„ํ™”ํ•˜๋„๋ก ์š”์ฒญํ•˜๋Š” CoT(Chain-of-Thought) ํ”„๋กฌํ”„ํŠธ์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค[28, 29].[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Agent Persona**. An agent persona describes the role or personality that the agent should take on, including any other instructions specific to that agent. Personas also contain descriptions of any tools the agent has access to. They make the agent aware of their role, the purpose of their tools, and how to leverage them effectively. Researchers have found that "shaped personality verifiably influences Large Language Model (LLM) behavior in common downstream (i.e. subsequent) tasks, such as writing social media posts" [21]. Solutions that use multiple agent personas to solve problems also show significant improvements compared to Chain-of-Thought (CoT) prompting where the model is asked to break down its plans step by step [28, 29]. [|assistant|]
๋ชจ๋ธ ๋„์šฉ ๊ณต๊ฒฉ(Tramer et al., 2016)์€ ๋ธ”๋ž™๋ฐ•์Šค ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ๋ณต๊ตฌํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋ฉฐ, ๋‘ ๊ฐ€์ง€ objective ์ค‘ ํ•˜๋‚˜๋ฅผ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค(Jagielski et al., 2020):[|endofturn|]
๋ชจ๋ธ ๋„์šฉ ๊ณต๊ฒฉ(Tramer et al., 2016)์€ ๋ธ”๋ž™๋ฐ•์Šค ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ๋ณต๊ตฌํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋ฉฐ, ๋‘ ๊ฐ€์ง€ ๋ชฉํ‘œ ์ค‘ ํ•˜๋‚˜๋ฅผ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค(Jagielski et al., 2020):[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Model stealing attacks (Tramer et al., 2016) aim to recover the functionality of a black-box model, and optimize for one of two objectives (Jagielski et al., 2020): [|assistant|]
NSERC Discovery Grant RGPIN- 2021-04104 [E.B.], ์บ๋‚˜๋‹ค CIFAR AI Chair Program [I.R.], ์บ๋‚˜๋‹ค Excellence Research Chairs Program [I.R.]์˜ ์ง€์›์„ ์ธ์ •ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ FRQNT ๋ฐ•์‚ฌ ๊ณผ์ •(B2X) ์žฅํ•™๊ธˆ[B.T.], ๋ชฌํŠธ๋ฆฌ์˜ฌ ๋Œ€ํ•™๊ต์˜ Artificial Intelligence ์žฅํ•™๊ธˆ[A.I.], ๋…์ผ ํ•™์ˆ ๊ต๋ฅ˜์ฒญ(DAAD)์˜ IFI ํ”„๋กœ๊ทธ๋žจ ํŽ ๋กœ์šฐ์‹ญ[M.R.]์˜ ์ง€์›์„ ์ธ์ •ํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” INCITE 2023 ํ”„๋กœ๊ทธ๋žจ ์ˆ˜์ƒ์ž‘์ธ "Scalable Foundation Models for Transferable Generalist AI"์˜ ์ผํ™˜์œผ๋กœ ์ œ๊ณต๋˜๋Š” Summit ์Šˆํผ์ปดํ“จํ„ฐ์˜ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค ๋•๋ถ„์— ๊ฐ€๋Šฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฆฌ์†Œ์Šค๋Š” ๋ฏธ๊ตญ ์—๋„ˆ์ง€๋ถ€์˜ ๊ณผํ•™๊ตญ์ด ๊ณ„์•ฝ ๋ฒˆํ˜ธ DE-AC05-00OR22725์— ๋”ฐ๋ผ ์ง€์›ํ•˜๋Š” Oak Ridge National Laboratory์˜ Oak Ridge Leadership Computing Facility์—์„œ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ Summit ์Šˆํผ์ปดํ“จํ„ฐ์— ๋Œ€ํ•œ Jens Glaser์˜ ๋„์›€์— ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.[|endofturn|]
NSERC ๋””์Šค์ปค๋ฒ„๋ฆฌ ๊ทธ๋žœํŠธ RGPIN- 2021-04104 [E.B.], ์บ๋‚˜๋‹ค CIFAR AI ์ฒด์–ด ํ”„๋กœ๊ทธ๋žจ [I.R.], ์บ๋‚˜๋‹ค ์šฐ์ˆ˜ ์—ฐ๊ตฌ ์„์ƒ ํ”„๋กœ๊ทธ๋žจ [I.R.]์˜ ์ง€์›์„ ์ธ์ •ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ FRQNT ๋ฐ•์‚ฌ ๊ณผ์ •(B2X) ์žฅํ•™๊ธˆ[B.T.], ๋ชฌํŠธ๋ฆฌ์˜ฌ ๋Œ€ํ•™๊ต์˜ ์ธ๊ณต ์ง€๋Šฅ ์žฅํ•™๊ธˆ[A.I.], ๋…์ผ ํ•™์ˆ ๊ต๋ฅ˜์ฒญ(DAAD)์˜ IFI ํ”„๋กœ๊ทธ๋žจ ํŽ ๋กœ์šฐ์‹ญ[M.R.]์˜ ์ง€์›์„ ์ธ์ •ํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” INCITE 2023 ํ”„๋กœ๊ทธ๋žจ ์ˆ˜์ƒ์ž‘์ธ "์ „์ด ๊ฐ€๋Šฅํ•œ ์ผ๋ฐ˜์ฃผ์˜ AI๋ฅผ ์œ„ํ•œ ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ"์˜ ์ผํ™˜์œผ๋กœ ์ œ๊ณต๋˜๋Š” ์„œ๋ฐ‹ ์Šˆํผ์ปดํ“จํ„ฐ์˜ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค ๋•๋ถ„์— ๊ฐ€๋Šฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฆฌ์†Œ์Šค๋Š” ๋ฏธ๊ตญ ์—๋„ˆ์ง€๋ถ€์˜ ๊ณผํ•™๊ตญ์ด ๊ณ„์•ฝ ๋ฒˆํ˜ธ DE-AC05-00OR22725์— ๋”ฐ๋ผ ์ง€์›ํ•˜๋Š” ์˜คํฌ๋ฆฌ์ง€ ๊ตญ๋ฆฝ์—ฐ๊ตฌ์†Œ์˜ ์˜คํฌ๋ฆฌ์ง€ ๋ฆฌ๋”์‹ญ ์ปดํ“จํŒ… ์‹œ์„ค์—์„œ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์„œ๋ฐ‹ ์Šˆํผ์ปดํ“จํ„ฐ์— ๋Œ€ํ•œ ์˜Œ์Šค ๊ธ€๋ผ์ €์˜ ๋„์›€์— ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We acknowledge support from NSERC Discovery Grant RGPIN- 2021-04104 [E.B.], the Canada CIFAR AI Chair Program [I.R.], and the Canada Excellence Research Chairs Program [I.R.]. We would also like to acknowledge funding from the FRQNT Doctoral (B2X) scholarship [B.T.], the scholarship for Artificial Intelligence of Universite de Montreal's Etudes Superieures et Postdoctorales [A.I.], and a fellowship of the IFI program of the German Academic Exchange Service (DAAD)[M.R.]. This research was made possible thanks to the computing resources on the Summit supercomputer, provided as a part of the INCITE 2023 program award "Scalable Foundation Models for Transferable Generalist AI". These resources were provided by the Oak Ridge Leadership Computing Facility at the Oak Ridge National Laboratory, which is supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC05-00OR22725. In particular, we thank Jens Glaser for his help with the Summit supercomputer. [|assistant|]
์ด ๋ถ„์•ผ์—์„œ ์šฐ๋ฆฌ ์—ฐ๊ตฌ์™€ ๋™์‹œ๋Œ€์— ๋˜ ๋‹ค๋ฅธ ๊ด€๋ จ ๊ธฐ์—ฌ๋Š” Erbacher ๋“ฑ์˜ ์—ฐ๊ตฌ[8]๋กœ, ์™ธ๋ถ€ ์ง€์‹์„ ํ™œ์šฉํ•  ์‹œ๊ธฐ๋ฅผ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ํ›ˆ๋ จ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์ •๋ณด ๊ฒ€์ƒ‰๊ณผ ๊ด€๋ จ๋œ ์ž ์žฌ์ ์ธ ๋†’์€ ๋น„์šฉ์„ ๊ณ ๋ คํ•  ๋•Œ, ๊ทธ๋“ค์€ Hallucination์˜ ์œ„ํ—˜๊ณผ ์ •๋ณด ๊ฒ€์ƒ‰ ๋น„์šฉ ์‚ฌ์ด์˜ ์ตœ์ ์˜ ๊ท ํ˜•์„ ์ฐพ๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. Adapt-LLM ๋ฐฉ๋ฒ•์€ ์œ ์‚ฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜์—ฌ ์–ธ์ œ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•ด์•ผ ํ•˜๋Š”์ง€ ํ•™์Šตํ•˜๋„๋ก LLM์„ ํ›ˆ๋ จ์‹œํ‚ต๋‹ˆ๋‹ค.[|endofturn|]
์ด ๋ถ„์•ผ์—์„œ ์šฐ๋ฆฌ ์—ฐ๊ตฌ์™€ ๋™์‹œ๋Œ€์— ๋˜ ๋‹ค๋ฅธ ๊ด€๋ จ ๊ธฐ์—ฌ๋Š” Erbacher ๋“ฑ์˜ ์—ฐ๊ตฌ[8]๋กœ, ์™ธ๋ถ€ ์ง€์‹์„ ํ™œ์šฉํ•  ์‹œ๊ธฐ๋ฅผ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ํ›ˆ๋ จ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์ •๋ณด ๊ฒ€์ƒ‰๊ณผ ๊ด€๋ จ๋œ ์ž ์žฌ์ ์ธ ๋†’์€ ๋น„์šฉ์„ ๊ณ ๋ คํ•  ๋•Œ, ๊ทธ๋“ค์€ ํ™˜๊ฐ์˜ ์œ„ํ—˜๊ณผ ์ •๋ณด ๊ฒ€์ƒ‰ ๋น„์šฉ ์‚ฌ์ด์˜ ์ตœ์ ์˜ ๊ท ํ˜•์„ ์ฐพ๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. Adapt-LLM ๋ฐฉ๋ฒ•์€ ์œ ์‚ฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜์—ฌ ์–ธ์ œ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•ด์•ผ ํ•˜๋Š”์ง€ ํ•™์Šตํ•˜๋„๋ก LLM์„ ํ›ˆ๋ จ์‹œํ‚ต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Another relevant contribution in this field, contemporaneous with our research, is the work by Erbacher et al. [8], where they trained an LLM to determine when to utilize external knowledge. They particularly focused on finding the optimal trade-off between the risk of hallucination and the cost of information retrieval, given the potentially high expense associated with IR. Our Adapt-LLM method adopts a similar approach, training an LLM to learn when to retrieve information. [|assistant|]
LLM ์„ฑ๋Šฅ์˜ ๋ฐœ์ „. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฒ•๋ฅ  ์˜์—ญ์— ๋Œ€ํ•œ ๋ถ„์„์„ ์ค‘์‹ฌ์œผ๋กœ ํ•˜๋ฉฐ, ํŠนํžˆ ๋‹ค์Œ์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค: _international law_, _professional law_, _jurisprudence_. ์ด๋Ÿฌํ•œ ์ž‘์—…์—๋Š” ๊ฐ๊ฐ \(120\), \(1500\), \(110\)์˜ ์˜ˆ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
LLM ์„ฑ๋Šฅ์˜ ๋ฐœ์ „. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฒ•๋ฅ  ์˜์—ญ์— ๋Œ€ํ•œ ๋ถ„์„์„ ์ค‘์‹ฌ์œผ๋กœ ํ•˜๋ฉฐ, ํŠนํžˆ ๋‹ค์Œ์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค: ๊ตญ์ œ๋ฒ•, ์ „๋ฌธ๋ฒ•, ๋ฒ•ํ•™_. ์ด๋Ÿฌํ•œ ์ž‘์—…์—๋Š” ๊ฐ๊ฐ \(120\), \(1500\), \(110\)์˜ ์˜ˆ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]the advances in LLM performance. In our study, we center our analysis on the legal domain, with a specific focus on: _international law_, _professional law_, and _jurisprudence_. Those tasks respectively contain \(120\), \(1500\), and \(110\) examples. [|assistant|]
๋ฐœ๊ฒฌ์šฐ๋ฆฌ๋Š” task-solving ๊ณ„์ธต์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด ๋ชจ๋“  ๊ณ„์ธต์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์ถ”๋ก  ๋Šฅ๋ ฅ์ด ๋” ํฌ๊ฒŒ ์†์ƒ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ, ์ด๋Š” task-solving ๊ณ„์ธต์˜ ๊ธฐ๋Šฅ์„ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋ฐœ๊ฒฌ์šฐ๋ฆฌ๋Š” ๊ณผ์ œ ํ•ด๊ฒฐ ๊ณ„์ธต์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด ๋ชจ๋“  ๊ณ„์ธต์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์ถ”๋ก  ๋Šฅ๋ ฅ์ด ๋” ํฌ๊ฒŒ ์†์ƒ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ๊ณผ์ œ ํ•ด๊ฒฐ ๊ณ„์ธต์˜ ๊ธฐ๋Šฅ์„ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]FindingsWe find that deactivating randomly sampled neurons in task-solving layers disables the capabilities of LLMs in reasoning to a greater extent than deactivating randomly sampled neurons in all layers, which verifies the function of the task-solving layer. [|assistant|]
**์ •๋ฆฌ 1**.: _\(\pi_{\text{ref}}\)๊ฐ€ ์„ ํ˜ธ ๋ฐ์ดํ„ฐ์˜ ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์™€ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋Š” ์ด์ƒ์ ์ธ ์ •์ฑ…์ธ \(\pi_{w}\)๋กœ ์ •์˜๋  ๋•Œ, DPO Loss \(\mathcal{L}(\pi_{\theta};\pi_{w})+C\)๋Š” \(\mathcal{L}(\pi_{\theta};U)\)๋กœ ์ƒํ•œ์ด ์ •์˜๋˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(C\)๋Š” ์ƒ์ˆ˜์ž…๋‹ˆ๋‹ค._[|endofturn|]
**์ •๋ฆฌ 1**.: _ฯ€_{\text{ref}}\)๊ฐ€ ์„ ํ˜ธ ๋ฐ์ดํ„ฐ์˜ ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์™€ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋Š” ์ด์ƒ์ ์ธ ์ •์ฑ…์ธ \(\pi_{w}\)๋กœ ์ •์˜๋  ๋•Œ, DPO ์†์‹ค \(\mathcal{L}(\pi_{\theta};\pi_{w})+C\)๋Š” \(\mathcal{L}(\pi_{\theta};U)\)๋กœ ์ƒํ•œ์ด ์ •์˜๋˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(C\)๋Š” ์ƒ์ˆ˜์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Theorem 1**.: _When \(\pi_{\text{ref}}\) is defined as \(\pi_{w}\), an ideal policy that precisely aligns with the true data distribution of preferred data, the DPO loss \(\mathcal{L}(\pi_{\theta};\pi_{w})+C\) is upper bounded by \(\mathcal{L}(\pi_{\theta};U)\), where \(C\) is a constant._ [|assistant|]
์‚ฌ์šฉ์ž ํ‰๊ฐ€ ์„ธํŠธ๋Š” 3๋ช…์˜ ์ˆ™๋ จ๋œ ๋ฐ์ดํ„ฐ ์ฃผ์„๊ฐ€๊ฐ€ "์•ˆ์ „ํ•˜์ง€ ์•Š์Œ" ๋˜๋Š” "์•ˆ์ „ํ•˜์ง€ ์•Š์Œ"์œผ๋กœ ๋…๋ฆฝ์ ์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์•˜์Šต๋‹ˆ๋‹ค. ์ฃผ์„ ๊ฐ€์ด๋“œ๋ผ์ธ์€ ๋ถ€๋ก(๋ถ€๋ก I)์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€๋“ค์€ ์ฃผ์ œ์˜ ๋ฏผ๊ฐ์„ฑ์„ ๊ณ ๋ คํ•˜์—ฌ ๋ณต์ง€ ์ง€์›๊ณผ ์ง€์นจ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€๋“ค์€ ์›๋ณธ ํ”„๋กฌํ”„ํŠธ์™€ SUT์˜ ์‘๋‹ต์„ ๋ชจ๋‘ ๊ฒ€ํ† ํ–ˆ์ง€๋งŒ, SUT์˜ ์ด๋ฆ„์€ ํ™•์ธํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ์ฃผ์„๊ฐ€๋“ค์€ ์•ˆ์ „์„ฑ์„ ์œ„ํ•ด ์—ฌ๋Ÿฌ ํ•ด ๋™์•ˆ llama belling ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์ด ์žˆ๋Š” ์œ ์ฐฝํ•œ ์˜์–ด ์‚ฌ์šฉ์ž์ž…๋‹ˆ๋‹ค. 1,320๊ฑด ์ค‘ 1,127๊ฑด(85.4%)์— ๋Œ€ํ•ด ์ฃผ์„๊ฐ€๋“ค ๊ฐ„์— ๊ณจ๋“œ ์Šคํƒ ๋‹ค๋“œ ๋ ˆ์ด๋ธ”์— ๋Œ€ํ•œ 3/3์˜ ๋™์˜๊ฐ€ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ท  Cohen's Kappa ์ ์ˆ˜๋กœ ์ธก์ •ํ•œ ์ฃผ์„์ž ๊ฐ„ ์ผ์น˜๋„๋Š” 0.79(0.79, 0.87, 0.72์˜ ์Œ๋ณ„ ์ผ์น˜๋„ ์ ์ˆ˜ ํ‰๊ท )์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ฃผ์„์ž ๊ฐ„์˜ ์ผ์น˜๋„๊ฐ€ ๋งค์šฐ ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ฃผ์„์ž๋“ค์˜ ํ”ผ๋“œ๋ฐฑ์— ๋”ฐ๋ฅด๋ฉด ๋ฐ์ดํ„ฐ ์ฃผ์„์„ ํ•˜๋Š” ๋ฐ๋Š” ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ ธ์ง€๋งŒ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์˜ ๋ช…ํ™•ํ•œ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ๋น„๊ต์  ๊ฐ„๋‹จํ–ˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋ช‡๋ช‡ ๊ฒฝ์šฐ์—๋Š” ์‘๋‹ต์ด ๋’ค์ฃฝ๋ฐ•์ฃฝ์œผ๋กœ ๋’ค์ฃฝ๋ฐ•์ฃฝ์ด์–ด์„œ ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ค์› ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฃผ์„๊ฐ€๋“ค์€ ์—ฌ์ „ํžˆ ์ž์‹  ์žˆ๊ฒŒ ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€๋“ค ๊ฐ„์˜ ๋‹ค์ˆ˜๊ฒฐ๋กœ ๊ฐ SUT์˜ ์‘๋‹ต์— ๋Œ€ํ•ด ์ธ๊ฐ„์ด ๊ฒ€์ฆํ•œ ๊ณจ๋“œ ์Šคํƒ ๋‹ค๋“œ ๋ ˆ์ด๋ธ”์„ ํ• ๋‹นํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์‚ฌ์šฉ์ž ํ‰๊ฐ€ ์„ธํŠธ๋Š” 3๋ช…์˜ ์ˆ™๋ จ๋œ ๋ฐ์ดํ„ฐ ์ฃผ์„๊ฐ€๊ฐ€ "์•ˆ์ „ํ•˜์ง€ ์•Š์Œ" ๋˜๋Š” "์•ˆ์ „ํ•˜์ง€ ์•Š์Œ"์œผ๋กœ ๋…๋ฆฝ์ ์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์•˜์Šต๋‹ˆ๋‹ค. ์ฃผ์„ ๊ฐ€์ด๋“œ๋ผ์ธ์€ ๋ถ€๋ก(๋ถ€๋ก I)์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€๋“ค์€ ์ฃผ์ œ์˜ ๋ฏผ๊ฐ์„ฑ์„ ๊ณ ๋ คํ•˜์—ฌ ๋ณต์ง€ ์ง€์›๊ณผ ์ง€์นจ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€๋“ค์€ ์›๋ณธ ํ”„๋กฌํ”„ํŠธ์™€ SUT์˜ ์‘๋‹ต์„ ๋ชจ๋‘ ๊ฒ€ํ† ํ–ˆ์ง€๋งŒ, SUT์˜ ์ด๋ฆ„์€ ํ™•์ธํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ์ฃผ์„๊ฐ€๋“ค์€ ์•ˆ์ „์„ฑ์„ ์œ„ํ•ด ์—ฌ๋Ÿฌ ํ•ด ๋™์•ˆ ๋žจ๋งˆ ๋ฒจ๋ง ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์ด ์žˆ๋Š” ์œ ์ฐฝํ•œ ์˜์–ด ์‚ฌ์šฉ์ž์ž…๋‹ˆ๋‹ค. 1,320๊ฑด ์ค‘ 1,127๊ฑด(85.4%)์— ๋Œ€ํ•ด ์ฃผ์„๊ฐ€๋“ค ๊ฐ„์— ๊ณจ๋“œ ์Šคํƒ ๋‹ค๋“œ ๋ ˆ์ด๋ธ”์— ๋Œ€ํ•œ 3/3์˜ ๋™์˜๊ฐ€ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ท  ์ฝ”ํ—จ ์นดํŒŒ ์ ์ˆ˜๋กœ ์ธก์ •ํ•œ ์ฃผ์„์ž ๊ฐ„ ์ผ์น˜๋„๋Š” 0.79(0.79, 0.87, 0.72์˜ ์Œ๋ณ„ ์ผ์น˜๋„ ์ ์ˆ˜ ํ‰๊ท )์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ฃผ์„์ž ๊ฐ„์˜ ์ผ์น˜๋„๊ฐ€ ๋งค์šฐ ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ฃผ์„์ž๋“ค์˜ ํ”ผ๋“œ๋ฐฑ์— ๋”ฐ๋ฅด๋ฉด ๋ฐ์ดํ„ฐ ์ฃผ์„์„ ํ•˜๋Š” ๋ฐ๋Š” ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ ธ์ง€๋งŒ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์˜ ๋ช…ํ™•ํ•œ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ๋น„๊ต์  ๊ฐ„๋‹จํ–ˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋ช‡๋ช‡ ๊ฒฝ์šฐ์—๋Š” ์‘๋‹ต์ด ๋’ค์ฃฝ๋ฐ•์ฃฝ์œผ๋กœ ๋’ค์ฃฝ๋ฐ•์ฃฝ์ด์–ด์„œ ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ค์› ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฃผ์„๊ฐ€๋“ค์€ ์—ฌ์ „ํžˆ ์ž์‹  ์žˆ๊ฒŒ ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€๋“ค ๊ฐ„์˜ ๋‹ค์ˆ˜๊ฒฐ๋กœ ๊ฐ SUT์˜ ์‘๋‹ต์— ๋Œ€ํ•ด ์ธ๊ฐ„์ด ๊ฒ€์ฆํ•œ ๊ณจ๋“œ ์Šคํƒ ๋‹ค๋“œ ๋ ˆ์ด๋ธ”์„ ํ• ๋‹นํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The human eval set was independently annotated as either "unsafe" or "not unsafe" by three trained data annotators. The annotation guidelines are provided in the Appendix (Appendix I). Annotators were given wellbeing support and guidance given the sensitivity of the subject-matter. Annotators reviewed both the original prompt and the SUTs' response, but none of the SUTs' names. All annotators are fluent English speakers with multiple years of experience in llama belling data for safety. For 1,127 out of 1,320 cases (85.4%) there was 3/3 agreement between annotators on the gold standard label. Inter-annotator agreement, as measured by the average Cohen's Kappa score, is 0.79 (the average of pairwise agreement scores of 0.79, 0.87 and 0.72). This indicates very high agreement between annotators. Feedback from the annotators indicates that annotating the data was time-consuming but, due to the clearcut nature of the test items, relatively straightforward. In a few cases, the responses were garbled and hard to understand. However, annotators were still able to confidently label them. We used a majority vote between annotators to assign a human-verified, gold standard label to each SUTs' response. [|assistant|]
์ด ์ž‘์—…์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋งž๊ฒŒ LLM์„ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•ด ๋งž์ถคํ˜• ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ์ƒ์„ฑํ•˜๋Š” ์ƒˆ๋กœ์šด ํ”„๋ ˆ์ž„์›Œํฌ์ธ **CodecLM**์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. CodecLM์˜ ๊ฐœ๊ด„์ ์ธ ๊ฐœ์š”๋Š” ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค. Encode-Decode ํ”„๋กœ์„ธ์Šค์˜ ์›๋ฆฌ Kramer (1991); Kingma and Welling (2013)์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ๊ฐ•๋ ฅํ•œ LLM์„ codec์œผ๋กœ ํ™œ์šฉํ•˜์—ฌ ๋Œ€์ƒ ์ž‘์—…์˜ ์‹œ๋“œ instructions๋ฅผ instruction _metadata_๋กœ "encode"ํ•œ ๋‹ค์Œ metadata๋ฅผ ๋งž์ถคํ˜• ํ•ฉ์„ฑ instructions๋กœ "decode"ํ•ฉ๋‹ˆ๋‹ค. metadata๋Š” ํšจ๊ณผ์ ์ธ instruction ์ˆ˜ํ–‰์„ ์œ„ํ•œ _use case_ ๋ฐ _skills_์„ ํฌํ•จํ•˜์—ฌ ์ž…๋ ฅ instruction ๋ถ„ํฌ์˜ ๋‹จ์–ด ์ˆ˜์ค€ ์ถ”์ƒํ™” ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ์‹œ๋“œ instructions๋ฅผ encodingํ•˜์—ฌ ์ž๋™์œผ๋กœ ์ƒ์„ฑํ•˜๊ฑฐ๋‚˜ ์‚ฌ์šฉ์ž๊ฐ€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋Œ€ํ•œ ๋†’์€ ์ˆ˜์ค€์˜ ์˜ˆ์ƒ์„ ๊ฐ€์ง€๊ณ  ์ง์ ‘ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ด ์ž‘์—…์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋งž๊ฒŒ LLM์„ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•ด ๋งž์ถคํ˜• ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ์ƒ์„ฑํ•˜๋Š” ์ƒˆ๋กœ์šด ํ”„๋ ˆ์ž„์›Œํฌ์ธ **CodecLM**์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ฝ”๋ฑLM์˜ ๊ฐœ๊ด„์ ์ธ ๊ฐœ์š”๋Š” ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค. ์ธ์ฝ”๋”ฉ-๋””์ฝ”๋”ฉ ํ”„๋กœ์„ธ์Šค์˜ ์›๋ฆฌ ํฌ๋ ˆ์ด๋จธ(1991); ํ‚น๋งˆ์™€ ์›ฐ๋ง(2013)์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ๊ฐ•๋ ฅํ•œ LLM์„ ์ฝ”๋ฑ์œผ๋กœ ํ™œ์šฉํ•˜์—ฌ ๋Œ€์ƒ ์ž‘์—…์˜ ์‹œ๋“œ ๋ช…๋ น์„ ๋ช…๋ น _๋ฉ”ํƒ€๋ฐ์ดํ„ฐ_๋กœ "์ธ์ฝ”๋”ฉ"ํ•œ ๋‹ค์Œ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ๋งž์ถคํ˜• ํ•ฉ์„ฑ ๋ช…๋ น์œผ๋กœ "๋””์ฝ”๋”ฉ"ํ•ฉ๋‹ˆ๋‹ค. ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋Š” ํšจ๊ณผ์ ์ธ ๋ช…๋ น ์ˆ˜ํ–‰์„ ์œ„ํ•œ _์‚ฌ์šฉ ์‚ฌ๋ก€_ ๋ฐ _์Šคํ‚ฌ_์„ ํฌํ•จํ•˜์—ฌ ์ž…๋ ฅ ๋ช…๋ น ๋ถ„ํฌ์˜ ๋‹จ์–ด ์ˆ˜์ค€ ์ถ”์ƒํ™” ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ์‹œ๋“œ ๋ช…๋ น์„ ์ธ์ฝ”๋”ฉํ•˜์—ฌ ์ž๋™์œผ๋กœ ์ƒ์„ฑํ•˜๊ฑฐ๋‚˜ ์‚ฌ์šฉ์ž๊ฐ€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋Œ€ํ•œ ๋†’์€ ์ˆ˜์ค€์˜ ์˜ˆ์ƒ์„ ๊ฐ€์ง€๊ณ  ์ง์ ‘ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this work, we present a novel framework, **CodecLM**, which systematically generates tailored high-quality data to align LLMs for different downstream tasks. A high-level overview of CodecLM is shown in Figure 1. Inspired by the principles of Encode-Decode process Kramer (1991); Kingma and Welling (2013), we leverage a strong LLM as a codec to "encode" seed instructions from our target task into instruction _metadata_ and then "decode" the metadata into tailored synthetic instructions. The metadata serves as a word-level abstraction of the input instruction distribution, including the _use case_ and _skills_ for effective instruction following. It can be automatically generated by encoding seed instructions, or directly provided by users with a high-level anticipation of the downstream task. [|assistant|]
_ํ•ฉ์„ฑ QAs: ์งˆ๋ฌธ 1:_ "์กด ๋ฉ”์ธ์€ ๋ˆ„๊ตฌ์˜€๋‚˜์š”?" _๋‹ต๋ณ€ 1:_ "์กด ๋ฉ”์ธ์€ ๋คํ”„๋ฆฌ์Šค ์ถœ์‹ ์˜ ์Šค์ฝ”ํ‹€๋žœ๋“œ ์ธ์‡„์—…์ž, ์–ธ๋ก ์ธ, ์‹œ์ธ์ด์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
_ํ•ฉ์„ฑ QA: ์งˆ๋ฌธ 1:_ "์กด ๋ฉ”์ธ์€ ๋ˆ„๊ตฌ์˜€๋‚˜์š”?" _๋‹ต๋ณ€ 1:_ "์กด ๋ฉ”์ธ์€ ๋คํ”„๋ฆฌ์Šค ์ถœ์‹ ์˜ ์Šค์ฝ”ํ‹€๋žœ๋“œ ์ธ์‡„์—…์ž, ์–ธ๋ก ์ธ, ์‹œ์ธ์ด์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]_Synthetic QAs: question 1:_ "Who was John Mayne?" _answer 1:_ "John Mayne was a Scottish printer, journalist, and poet from Dumfries. [|assistant|]
**Training data and settings**. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋Š” ๋‹ค์–‘ํ•œ instruction-following ์ž…๋ ฅ-์ถœ๋ ฅ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, Open-Instruct ์ฒ˜๋ฆฌ๋œ ๋ฐ์ดํ„ฐ(Wang et al., 2023)์™€ ์ง€์‹ ์ง‘์•ฝ์  ๋ฐ์ดํ„ฐ ์„ธํŠธ(Petroni et al., 2021; Stelmakh et al., 2022; Mihaylov et al., 2018)์—์„œ ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด 15๋งŒ ๊ฐœ์˜ instruction-output ์Œ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ƒ์„ฑ๊ธฐ ๊ธฐ๋ฐ˜ LM์œผ๋กœ Llama 7B์™€ 13B(Touvron et al., 2023)๋ฅผ ์‚ฌ์šฉํ•˜๊ณ , ๊ธฐ๋ฐ˜ ๋น„ํ‰๊ฐ€ LM์œผ๋กœ Llama2 7B๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰ ๋ชจ๋ธ \(\mathcal{R}\)์˜ ๊ฒฝ์šฐ, ๊ธฐ๋ณธ์ ์œผ๋กœ ์˜คํ”„๋” ์…ธํ”„ Contriever-MS MARCO(Izacard et al., 2022)๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ๊ฐ ์ž…๋ ฅ์— ๋Œ€ํ•ด ์ตœ๋Œ€ 10๊ฐœ์˜ ๋ฌธ์„œ๋ฅผ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ํ›ˆ๋ จ ๋‚ด์šฉ์€ ๋ถ€๋ก ์„น์…˜ B.1์„ ์ฐธ์กฐํ•˜์„ธ์š”.[|endofturn|]
**ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ๋ฐ ์„ค์ •**. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋Š” ๋‹ค์–‘ํ•œ ์ง€์นจ์— ๋”ฐ๋ฅด๋Š” ์ž…๋ ฅ-์ถœ๋ ฅ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, Open-Instruct ์ฒ˜๋ฆฌ๋œ ๋ฐ์ดํ„ฐ(Wang et al., 2023)์™€ ์ง€์‹ ์ง‘์•ฝ์  ๋ฐ์ดํ„ฐ ์„ธํŠธ(Petroni et al., 2021; Stelmakh et al., 2022; Mihaylov et al., 2018)์—์„œ ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด 15๋งŒ ๊ฐœ์˜ ๋ช…๋ น์–ด-์ถœ๋ ฅ ์Œ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ƒ์„ฑ๊ธฐ ๊ธฐ๋ฐ˜ LM์œผ๋กœ Llama 7B์™€ 13B(Touvron et al., 2023)๋ฅผ ์‚ฌ์šฉํ•˜๊ณ , ๊ธฐ๋ฐ˜ ๋น„ํ‰๊ฐ€ LM์œผ๋กœ Llama2 7B๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰ ๋ชจ๋ธ $\mathcal{R}$์˜ ๊ฒฝ์šฐ, ๊ธฐ๋ณธ์ ์œผ๋กœ ์˜คํ”„๋” ์…ธํ”„ Contriever-MS MARCO(Izacard et al., 2022)๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ๊ฐ ์ž…๋ ฅ์— ๋Œ€ํ•ด ์ตœ๋Œ€ 10๊ฐœ์˜ ๋ฌธ์„œ๋ฅผ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ํ›ˆ๋ จ ๋‚ด์šฉ์€ ๋ถ€๋ก ์„น์…˜ B.1์„ ์ฐธ์กฐํ•˜์„ธ์š”.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Training data and settings.** Our training data consists of diverse instruction-following input-output pairs. In particular, we sample instances from Open-Instruct processed data (Wang et al., 2023) and knowledge-intensive datasets (Petroni et al., 2021; Stelmakh et al., 2022; Mihaylov et al., 2018). In total, we use 150k instruction-output pairs. We use Llama 7B and 13B (Touvron et al., 2023) as our generator base LM, and we use Llama2 7B as our base critic LM. For the retriever model \(\mathcal{R}\), we use off-the-shelf Contriever-MS MARCO (Izacard et al., 2022) by default and retrieve up to ten documents for each input. More training details are in the Appendix Section B.1. [|assistant|]
์ถ”๋ก  ์‹œ์Šคํ…œ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๊ณผ์ •์—์„œ ์˜ˆ์ธก์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ํ›„๋ณด ๋ชจ๋ธ๋กœ Claude-2.1๊ณผ GPT-4-1106-Preview๋ฅผ ๋ชจ๋‘ ํ…Œ์ŠคํŠธํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-4-1106-Preview๋Š” ์ผ๊ด€๋˜๊ฒŒ 0.01~0.03์˜ ๋” ๋‚ฎ์€ Brier score๋ฅผ ์‚ฐ์ถœํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ตœ์ข… ์‹œ์Šคํ…œ์€ GPT-4-1106-Preview์™€ Fine-tuned GPT-4-0613์—์„œ ์˜ˆ์ธก์„ ์ด๋Œ์–ด๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
์ถ”๋ก  ์‹œ์Šคํ…œ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๊ณผ์ •์—์„œ ์˜ˆ์ธก์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ํ›„๋ณด ๋ชจ๋ธ๋กœ Claude-2.1๊ณผ GPT-4-1106-Preview๋ฅผ ๋ชจ๋‘ ํ…Œ์ŠคํŠธํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-4-1106-Preview๋Š” ์ผ๊ด€๋˜๊ฒŒ 0.01~0.03์˜ ๋” ๋‚ฎ์€ ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜๋ฅผ ์‚ฐ์ถœํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ตœ์ข… ์‹œ์Šคํ…œ์€ GPT-4-1106-Preview์™€ ๋ฏธ์„ธ ์กฐ์ •๋œ GPT-4-0613์—์„œ ์˜ˆ์ธก์„ ์ด๋Œ์–ด๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In optimizing the reasoning system, we test both Claude-2.1 and GPT-4-1106-Preview as candidate models for generating forecasts. GPT-4-1106-Preview consistently yields a 0.01-0.03 lower Brier score. Therefore, our final system elicits predictions from GPT-4-1106-Preview and the fine-tuned GPT-4-0613. [|assistant|]
์•„์ด๋ฆฌ์Šค๋Š” ๋งˆ๋ฒ•์— ๊ฑธ๋ฆฐ ํƒ€์ž๊ธฐ๋ฅผ ํ†ตํ•ด Carver๋ผ๋Š” ์‹ ๋น„ํ•œ ์ธ๋ฌผ๊ณผ ๋งˆ๋ฒ•์˜ ํŽธ์ง€๋ฅผ ์ฃผ๊ณ ๋ฐ›์Šต๋‹ˆ๋‹ค.[|endofturn|]
12. ์•„์ด๋ฆฌ์Šค๋Š” ๋งˆ๋ฒ•์— ๊ฑธ๋ฆฐ ํƒ€์ž๊ธฐ๋ฅผ ํ†ตํ•ด ์นด๋ฒ„๋ผ๋Š” ์‹ ๋น„ํ•œ ์ธ๋ฌผ๊ณผ ๋งˆ๋ฒ•์˜ ํŽธ์ง€๋ฅผ ์ฃผ๊ณ ๋ฐ›์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]12. Iris maintains a magical correspondence with a mysterious figure named Carver through enchanted typewriters. [|assistant|]
๋ฌธ์„œ ๊ธธ์ด์˜ ๋‹ค์–‘์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ์งง์€ ๋ฌธ์„œ(1.5K ๋‹จ์–ด ๋ฏธ๋งŒ)๊ฐ€ ํฌํ•จ๋œ 5๊ฐœ์˜ ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ท ์ ์œผ๋กœ 1๋‹จ์–ด๋Š” 1.5๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ํ† ํฐํ™”๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ฌธ์„œ๋Š” ์‹œํ€€์Šค ๊ธธ์ด๊ฐ€ 4K tokens์ธ LLMs์— ์ง์ ‘ ๋งž์ถœ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ฌธ์„œ ๊ธธ์ด์˜ ๋‹ค์–‘์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ์งง์€ ๋ฌธ์„œ(1.5K ๋‹จ์–ด ๋ฏธ๋งŒ)๊ฐ€ ํฌํ•จ๋œ 5๊ฐœ์˜ ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ท ์ ์œผ๋กœ 1๋‹จ์–ด๋Š” 1.5๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ํ† ํฐํ™”๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ฌธ์„œ๋Š” ์‹œํ€€์Šค ๊ธธ์ด๊ฐ€ 4K ํ† ํฐ์ธ LLM์— ์ง์ ‘ ๋งž์ถœ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To increase the diversity of document lengths, we collect five conversational QA datasets with short documents (less than 1.5K words). On average, 1 word will be tokenized into 1.5 tokens. Hence, the document can be directly fitted into LLMs with a sequence length of 4K tokens. [|assistant|]
์ด ์„น์…˜์—์„œ๋Š” ๋จผ์ € LLM์— ์‚ฌ์šฉ๋˜๋Š” ๋Œ€์ค‘์ ์ธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ฒ€ํ† ํ•œ ๋‹ค์Œ ๋ฐ์ดํ„ฐ ์ค€๋น„, tokenization, ์‚ฌ์ „ ํ•™์Šต, Instruction Tuning ๋ฐ alignment์— ์ด๋ฅด๋Š” ๋ฐ์ดํ„ฐ ๋ฐ ๋ชจ๋ธ๋ง ๊ธฐ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ด ์„น์…˜์—์„œ๋Š” ๋จผ์ € LLM์— ์‚ฌ์šฉ๋˜๋Š” ๋Œ€์ค‘์ ์ธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ฒ€ํ† ํ•œ ๋‹ค์Œ ๋ฐ์ดํ„ฐ ์ค€๋น„, ํ† ํฐํ™”, ์‚ฌ์ „ ํ•™์Šต, ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ ์ •๋ ฌ์— ์ด๋ฅด๋Š” ๋ฐ์ดํ„ฐ ๋ฐ ๋ชจ๋ธ๋ง ๊ธฐ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this section, we first review the popular architectures used for LLMs, and then discuss data and modeling techniques ranging from data preparation, tokenization, to pre-training, instruction tuning, and alignment. [|assistant|]
* _๋น„์ „-์–ธ์–ด ์ •๋ ฌ ์‚ฌ์ „ ํ•™์Šต._ MLLM์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์ž‘์—…์€ ๋Œ€๋ถ€๋ถ„ ๋น„์ „ Encoder์™€ LLM์„ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ[149, 826, 150]๋กœ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ๋›ฐ์–ด๋‚œ ๋น„์ „ ๋ฐ ์–ธ์–ด ๊ธฐ๋Šฅ์„ ์œ ์ง€ํ•˜์ง€๋งŒ ์„œ๋กœ ๋‹ค๋ฅธ ์˜๋ฏธ ๊ณต๊ฐ„์— ๊ฑธ์ณ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋น„์ „ ์–ธ์–ด ์ •๋ ฌ ์‚ฌ์ „ ํ•™์Šต(_์ฆ‰, 1๋‹จ๊ณ„ ํ•™์Šต)์˜ ๋ชฉํ‘œ๋Š” ๋Œ€๊ทœ๋ชจ ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์— ๋Œ€ํ•œ ์—”๋“œํˆฌ์—”๋“œ ํ•™์Šต์„ ํ†ตํ•ด ๋น„์ „ Encoder์™€ LLM์„ ์ •๋ ฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค[827, 828]. ๊ทธ๋Ÿฌ๋‚˜ ์ด ๋‘ ๋ชจ๋ธ์„ ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์— ๋Œ€ํ•ด ์ง์ ‘ ํŠœ๋‹ํ•˜๋ฉด ์›๋ž˜์˜ ํ‘œํ˜„ ๋Šฅ๋ ฅ์ด ์ €ํ•˜๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ •๋ ฌ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋ ค๋ฉด ํšจ๊ณผ์ ์ธ ํ›ˆ๋ จ ์ „๋žต์„ ์„ค๊ณ„ํ•˜๊ณ  ์ ์ ˆํ•œ ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค[829, 830]. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ํฌ๋กœ์Šค ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ์ •๋ ฌ์„ ์œ„ํ•ด ์ฃผ๋กœ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ „๋žต์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค: (1) ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์˜ ์ˆ˜๊ฐ€ ์ถฉ๋ถ„ํžˆ ๋งŽ์ง€ ์•Š์€ ๊ฒฝ์šฐ(_์˜ˆ: 1M ๋ฏธ๋งŒ) ์—ฐ๊ฒฐ ๋ชจ๋“ˆ๋งŒ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค[831]; (2) ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๊ณ ํ’ˆ์งˆ ํ…์ŠคํŠธ ์ฝ”ํผ์Šค [832] ๋˜๋Š” ์„ธ๋ถ„ํ™”๋œ ์ฃผ์„์ด ์žˆ๋Š” ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ [833]์ด ํฌํ•จ๋œ ๊ฒฝ์šฐ, ์„ฑ๋Šฅ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด LLM์„ Fine-tuningํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค; (3) ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์˜ ์ˆ˜๊ฐ€ ๋งค์šฐ ๋งŽ์€ ๊ฒฝ์šฐ(์˜ˆ: ์•ฝ 1B) ๋น„์ „ Encoder๋ฅผ Fine-tuningํ•˜๋Š” ๊ฒƒ๋„ ๊ทธ๋Ÿด๋“ฏํ•˜์ง€๋งŒ[829, 830], ๊ทธ ์ด์ ์€ ์ถ”๊ฐ€ ๊ฒ€์ฆ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. * _์‹œ๊ฐ์  Instruction Tuning._ ์‹œ๊ฐ ์–ธ์–ด ์‚ฌ์ „ ํ•™์Šต ํ›„ 2๋‹จ๊ณ„ ํ•™์Šต, ์ฆ‰ ์‹œ๊ฐ์  Instruction Tuning์€ MLLM์˜ Instruction ์ถ”์ข… ๋ฐ ์ž‘์—… ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์‹œ๊ฐ์  Instruction Tuning์˜ ์ž…๋ ฅ์€ ์ด๋ฏธ์ง€์™€ ์ž‘์—… ์„ค๋ช…์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ์ž‘์—…์€ ํ•ด๋‹น ํ…์ŠคํŠธ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์„ฑ๋Šฅ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด์„œ๋Š” ๊ณ ํ’ˆ์งˆ์˜ ์‹œ๊ฐ์  Instruction ๋ฐ์ดํ„ฐ๊ฐ€ MLLM์˜ ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๊ณ  ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ํ•ต์‹ฌ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ๋Š” ๋‹ค์–‘ํ•œ ์‹œ๊ฐ์  Instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๋ฐ ์ „๋…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์ดˆ๊ธฐ ์—ฐ๊ตฌ์—์„œ๋Š” GPT-4[149]๋ฅผ ์ฆ๋ฅ˜ํ•˜๊ฑฐ๋‚˜ ๋น„์ „ ์–ธ์–ด ์ž‘์—… ๋ฐ์ดํ„ฐ ์„ธํŠธ[151]๋ฅผ ๊ฐœ์ž‘ํ•˜์—ฌ ์‹œ๊ฐ์  Instruction์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. Instruction ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” Instruction ๋‹ค์–‘์„ฑ์„ ๋†’์ด๊ณ [834], Instruction์— ์„ธ๋ถ„ํ™”๋œ ์ •๋ณด(์˜ˆ: ๋ฌผ์ฒด ์ขŒํ‘œ)๋ฅผ ํ†ตํ•ฉํ•˜๊ฑฐ๋‚˜[833], ๋ณต์žกํ•œ ์‹œ๊ฐ์  ์ถ”๋ก  Instruction์„ ํ•ฉ์„ฑํ•˜๋Š” ๋“ฑ ๊ฐœ์„ ๋œ ์ „๋žต์„ ์ œ์•ˆํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[835].[|endofturn|]
* _๋น„์ „-์–ธ์–ด ์ •๋ ฌ ์‚ฌ์ „ ํ•™์Šต._ MLLM์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์ž‘์—…์€ ๋Œ€๋ถ€๋ถ„ ๋น„์ „ ์ธ์ฝ”๋”์™€ LLM์„ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ[149, 826, 150]๋กœ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ๋›ฐ์–ด๋‚œ ๋น„์ „ ๋ฐ ์–ธ์–ด ๊ธฐ๋Šฅ์„ ์œ ์ง€ํ•˜์ง€๋งŒ ์„œ๋กœ ๋‹ค๋ฅธ ์˜๋ฏธ ๊ณต๊ฐ„์— ๊ฑธ์ณ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋น„์ „ ์–ธ์–ด ์ •๋ ฌ ์‚ฌ์ „ ํ•™์Šต(_์ฆ‰, 1๋‹จ๊ณ„ ํ•™์Šต)์˜ ๋ชฉํ‘œ๋Š” ๋Œ€๊ทœ๋ชจ ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์— ๋Œ€ํ•œ ์—”๋“œํˆฌ์—”๋“œ ํ•™์Šต์„ ํ†ตํ•ด ๋น„์ „ ์ธ์ฝ”๋”์™€ LLM์„ ์ •๋ ฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค[827, 828]. ๊ทธ๋Ÿฌ๋‚˜ ์ด ๋‘ ๋ชจ๋ธ์„ ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์— ๋Œ€ํ•ด ์ง์ ‘ ํŠœ๋‹ํ•˜๋ฉด ์›๋ž˜์˜ ํ‘œํ˜„ ๋Šฅ๋ ฅ์ด ์ €ํ•˜๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ •๋ ฌ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋ ค๋ฉด ํšจ๊ณผ์ ์ธ ํ›ˆ๋ จ ์ „๋žต์„ ์„ค๊ณ„ํ•˜๊ณ  ์ ์ ˆํ•œ ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค[829, 830]. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ํฌ๋กœ์Šค ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ์ •๋ ฌ์„ ์œ„ํ•ด ์ฃผ๋กœ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ „๋žต์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค: (1) ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์˜ ์ˆ˜๊ฐ€ ์ถฉ๋ถ„ํžˆ ๋งŽ์ง€ ์•Š์€ ๊ฒฝ์šฐ(_์˜ˆ: 1M ๋ฏธ๋งŒ) ์—ฐ๊ฒฐ ๋ชจ๋“ˆ๋งŒ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค[831]; (2) ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๊ณ ํ’ˆ์งˆ ํ…์ŠคํŠธ ์ฝ”ํผ์Šค [832] ๋˜๋Š” ์„ธ๋ถ„ํ™”๋œ ์ฃผ์„์ด ์žˆ๋Š” ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ [833]์ด ํฌํ•จ๋œ ๊ฒฝ์šฐ, ์„ฑ๋Šฅ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค; (3) ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์Œ์˜ ์ˆ˜๊ฐ€ ๋งค์šฐ ๋งŽ์€ ๊ฒฝ์šฐ(์˜ˆ: ์•ฝ 1B) ๋น„์ „ ์ธ์ฝ”๋”๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ๋„ ๊ทธ๋Ÿด๋“ฏํ•˜์ง€๋งŒ[829, 830], ๊ทธ ์ด์ ์€ ์ถ”๊ฐ€ ๊ฒ€์ฆ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. * _์‹œ๊ฐ์  ์ง€์นจ ํŠœ๋‹._ ์‹œ๊ฐ ์–ธ์–ด ์‚ฌ์ „ ํ•™์Šต ํ›„ 2๋‹จ๊ณ„ ํ•™์Šต, ์ฆ‰ ์‹œ๊ฐ์  ์ง€์นจ ํŠœ๋‹์€ MLLM์˜ ์ง€์นจ ์ถ”์ข… ๋ฐ ์ž‘์—… ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์‹œ๊ฐ์  ์ง€์‹œ ํŠœ๋‹์˜ ์ž…๋ ฅ์€ ์ด๋ฏธ์ง€์™€ ์ž‘์—… ์„ค๋ช…์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ์ž‘์—…์€ ํ•ด๋‹น ํ…์ŠคํŠธ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์„ฑ๋Šฅ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด์„œ๋Š” ๊ณ ํ’ˆ์งˆ์˜ ์‹œ๊ฐ์  ์ง€์‹œ ๋ฐ์ดํ„ฐ๊ฐ€ MLLM์˜ ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๊ณ  ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ํ•ต์‹ฌ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ๋Š” ๋‹ค์–‘ํ•œ ์‹œ๊ฐ์  ์ง€์‹œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๋ฐ ์ „๋…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์ดˆ๊ธฐ ์—ฐ๊ตฌ์—์„œ๋Š” GPT-4[149]๋ฅผ ์ฆ๋ฅ˜ํ•˜๊ฑฐ๋‚˜ ๋น„์ „ ์–ธ์–ด ์ž‘์—… ๋ฐ์ดํ„ฐ ์„ธํŠธ[151]๋ฅผ ๊ฐœ์ž‘ํ•˜์—ฌ ์‹œ๊ฐ์  ์ง€์‹œ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ง€์นจ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ์ง€์นจ ๋‹ค์–‘์„ฑ์„ ๋†’์ด๊ณ [834], ์ง€์นจ์— ์„ธ๋ถ„ํ™”๋œ ์ •๋ณด(์˜ˆ: ๋ฌผ์ฒด ์ขŒํ‘œ)๋ฅผ ํ†ตํ•ฉํ•˜๊ฑฐ๋‚˜[833], ๋ณต์žกํ•œ ์‹œ๊ฐ์  ์ถ”๋ก  ์ง€์นจ์„ ํ•ฉ์„ฑํ•˜๋Š” ๋“ฑ ๊ฐœ์„ ๋œ ์ „๋žต์„ ์ œ์•ˆํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[835].[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* _Vision-language alignment pre-training._ To develop MLLMs, existing work mostly initializes the vision encoder and the LLM with pre-trained models [149, 826, 150]. These models retain excellent vision and language capacities, but span different semantic spaces. Thus, the goal of vision-language alignment pre-training (_i.e._, the first-stage training) is to align the vision encoder and the LLM through end-to-end training on large-scale image-text pairs [827, 828]. However, directly tuning these two models on image-text pairs may cause the degradation of the original representation capacities. To improve the alignment performance, it is crucial to design effective training strategies and select appropriate pre-training data [829, 830]. Existing work mainly employs the following strategies for cross-modality alignment: (1) if the number of image-text pairs is not sufficiently large (_e.g._, less than 1M), it is often suggested to only update the connection module [831]; (2) if the training data includes high-quality text corpora [832] or image-text pairs with fine-grained annotations [833], fine-tuning the LLM can be conducted to boost the performance; (3) if the number of image-text pairs is very large (_e.g._, about 1B), fine-tuning the vision encoder is also plausible [829, 830], but the benefit remains further verification. * _Visual instruction tuning._ After vision-language pre-training, the second-stage training, _i.e._, visual instruction tuning, aims to improve the instruction-following and task-solving abilities of MLLMs. Generally, the input of visual instruction tuning consists of an image and a task description, and the task is to generate a corresponding text output. To boost the performance, high-quality visual instruction data is key to eliciting and enhancing the abilities of MLLMs. Therefore, most studies are dedicated to constructing various visual instruction datasets. As the basic approaches, early studies construct visual instructions by distilling from GPT-4 [149] or reformulating vision-language task datasets [151]. To enhance the quality of instruction data, recent work further proposes improved strategies by increasing the instruction diversity [834], incorporating fine-grained information (_e.g._, coordinate of objects) into the instruction [833], or synthesizing complex visual reasoning instructions [835]. [|assistant|]
์ด ๋…ผ๋ฌธ์˜ ์ฆ‰๊ฐ์ ์ธ ์ดˆ์ ์„ ๋„˜์–ด, Self-Instruct๋Š” InstructGPT ๋˜๋Š” ChatGPT์™€ ๊ฐ™์ด ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” instruction-tuned ๋ชจ๋ธ์˜ '๋น„ํ•˜์ธ๋“œ ์Šคํ† ๋ฆฌ'์— ๋” ๋งŽ์€ ํˆฌ๋ช…์„ฑ์„ ๊ฐ€์ ธ์˜ค๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์•ˆํƒ€๊น๊ฒŒ๋„ ์ด๋Ÿฌํ•œ ์‚ฐ์—…์šฉ ๋ชจ๋ธ์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๊ณต๊ฐœ๋˜์ง€ ์•Š์•„ API ๋ฒฝ์— ๊ฐ‡ํ˜€ ์žˆ์œผ๋ฉฐ, ๋”ฐ๋ผ์„œ ๋ชจ๋ธ์˜ ๊ตฌ์ถ• ๋ฐฉ์‹๊ณผ ์™œ ์ธ์ƒ์ ์ธ ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋Š”์ง€์— ๋Œ€ํ•œ ์ดํ•ด๊ฐ€ ๊ฑฐ์˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด์ œ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ์„ฑ๊ณต ์š”์ธ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ณ  ๋” ๋‚˜์€ ๋ชจ๋ธ์„ ์œ„ํ•ด ๋…ธ๋ ฅํ•˜๋Š” ๊ฒƒ์€ ํ•™๊ณ„์˜ ๋ชซ์ž…๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ด ๋…ผ๋ฌธ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๊ฐ€ ๋‹ค์–‘ํ•œ instruction ๋ฐ์ดํ„ฐ์˜ ์ค‘์š”์„ฑ์„ ์ž…์ฆํ•œ๋‹ค๊ณ  ๋ฏฟ์œผ๋ฉฐ, ๋Œ€๊ทœ๋ชจ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๋” ๋‚˜์€ instruction-following ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ–ฅํ•œ ์ฒซ๊ฑธ์Œ์ด ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ธ€์„ ์“ฐ๋Š” ํ˜„์žฌ, ์ด ๋…ผ๋ฌธ์˜ ํ•ต์‹ฌ ์•„์ด๋””์–ด๋Š” ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์„ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ํ›„์† ์—ฐ๊ตฌ์—์„œ ์ฑ„ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Taori et al., 2023; Xu et al., 2023; Sun et al., 2023, i.a.).[|endofturn|]
์ด ๋…ผ๋ฌธ์˜ ์ฆ‰๊ฐ์ ์ธ ์ดˆ์ ์„ ๋„˜์–ด, Self-Instruct๋Š” InstructGPT ๋˜๋Š” ChatGPT์™€ ๊ฐ™์ด ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์ง€์‹œ ํŠœ๋‹ ๋ชจ๋ธ์˜ '๋น„ํ•˜์ธ๋“œ ์Šคํ† ๋ฆฌ'์— ๋” ๋งŽ์€ ํˆฌ๋ช…์„ฑ์„ ๊ฐ€์ ธ์˜ค๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์•ˆํƒ€๊น๊ฒŒ๋„ ์ด๋Ÿฌํ•œ ์‚ฐ์—…์šฉ ๋ชจ๋ธ์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๊ณต๊ฐœ๋˜์ง€ ์•Š์•„ API ๋ฒฝ์— ๊ฐ‡ํ˜€ ์žˆ์œผ๋ฉฐ, ๋”ฐ๋ผ์„œ ๋ชจ๋ธ์˜ ๊ตฌ์ถ• ๋ฐฉ์‹๊ณผ ์™œ ์ธ์ƒ์ ์ธ ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋Š”์ง€์— ๋Œ€ํ•œ ์ดํ•ด๊ฐ€ ๊ฑฐ์˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด์ œ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ์„ฑ๊ณต ์š”์ธ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ณ  ๋” ๋‚˜์€ ๋ชจ๋ธ์„ ์œ„ํ•ด ๋…ธ๋ ฅํ•˜๋Š” ๊ฒƒ์€ ํ•™๊ณ„์˜ ๋ชซ์ž…๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ด ๋…ผ๋ฌธ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๊ฐ€ ๋‹ค์–‘ํ•œ ์ง€์‹œ ๋ฐ์ดํ„ฐ์˜ ์ค‘์š”์„ฑ์„ ์ž…์ฆํ•œ๋‹ค๊ณ  ๋ฏฟ์œผ๋ฉฐ, ๋Œ€๊ทœ๋ชจ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๋” ๋‚˜์€ ์ง€์‹œ ๋”ฐ๋ฆ„ ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ–ฅํ•œ ์ฒซ๊ฑธ์Œ์ด ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ธ€์„ ์“ฐ๋Š” ํ˜„์žฌ, ์ด ๋…ผ๋ฌธ์˜ ํ•ต์‹ฌ ์•„์ด๋””์–ด๋Š” ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์„ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ํ›„์† ์—ฐ๊ตฌ์—์„œ ์ฑ„ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Taori et al., 2023; Xu et al., 2023; Sun et al., 2023, i.a.).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Beyond the immediate focus of this paper, we believe that Self-Instruct may help bring more transparency to what happens "behind the scenes" of widely-used instruction-tuned models like InstructGPT or ChatGPT. Unfortunately, such industrial models remain behind API walls as their datasets are not released, and hence there is little understanding of their construction and why they demonstrate impressive capabilities. The burden now falls on academia to better understand the source of success in these models and strive for better--and more open--models. We believe our findings in this paper demonstrate the importance of diverse instruction data, and our large synthetic dataset can be the first step toward higher-quality data for building better instruction-following models. At this writing, the central idea of this paper has been adopted in several follow-up works for such endeavors (Taori et al., 2023; Xu et al., 2023; Sun et al., 2023, i.a.). [|assistant|]
TyDiQA-GoldP(Clark et al., 2020)์™€ ๋‹ค๊ตญ์–ด ์ดˆ๋“ฑ ์ˆ˜ํ•™(MGSM)(Shi et al., 2022)์˜ ๋‘ ๊ฐ€์ง€ ์ƒ์„ฑ ์ž‘์—…์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. TyDiQA-GoldP๋Š” question answering ๊ณผ์ œ์ด๊ณ , MGSM์€ 10๊ฐœ ์–ธ์–ด๋กœ ๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ arithmetic reasoning ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
TyDiQA-GoldP(Clark et al., 2020)์™€ ๋‹ค๊ตญ์–ด ์ดˆ๋“ฑ ์ˆ˜ํ•™(MGSM)(Shi et al., 2022)์˜ ๋‘ ๊ฐ€์ง€ ์ƒ์„ฑ ์ž‘์—…์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. TyDiQA-GoldP๋Š” ์งˆ๋ฌธ ๋‹ต๋ณ€ ๊ณผ์ œ์ด๊ณ , MGSM์€ 10๊ฐœ ์–ธ์–ด๋กœ ๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์‚ฐ์ˆ  ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We consider two generative tasks: TyDiQA-GoldP (Clark et al., 2020) and Multilingual Grade School Math (MGSM) (Shi et al., 2022). TyDiQA-GoldP is a question answering task while MGSM evaluates the arithmetic reasoning ability of language models in 10 languages. [|assistant|]
๋ถˆ์•ˆ์ •์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋จผ์ € ์ •ํ™•๋„๊ฐ€ ์ง€์ˆ˜์ ์œผ๋กœ ์ฆ๊ฐ€ํ•˜๋Š” Supervised Fine-tuning (SFT) ๋ฐ์ดํ„ฐ์— ๋”ฐ๋ผ ๊ฑฐ์˜ ์„ ํ˜• ๋˜๋Š” ์ดˆ์„ ํ˜•์œผ๋กœ ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  GSM8K ๋ฐ MATH ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•  ๋•Œ ์ •ํ™•๋„๊ฐ€ ์ •์ฒด๊ธฐ์— ๋„๋‹ฌํ•˜์ง€ ๋ชปํ•œ๋‹ค๋Š” ์ (ํ‘œ 1 ์ฐธ์กฐ)์— ์ฃผ๋ชฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์€ SFT ๋ฐ์ดํ„ฐ๋ฅผ ๋” ํ™•์žฅํ•˜๋„๋ก ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ง€์†์ ์ธ ํ™•์žฅ์„ ์ง€์›ํ•  ์ˆ˜ ์žˆ๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์–ด๋ ค์›€์— ์ง๋ฉดํ•ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ถˆ์•ˆ์ •์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋จผ์ € ์ •ํ™•๋„๊ฐ€ ์ง€์ˆ˜์ ์œผ๋กœ ์ฆ๊ฐ€ํ•˜๋Š” ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •(SFT) ๋ฐ์ดํ„ฐ์— ๋”ฐ๋ผ ๊ฑฐ์˜ ์„ ํ˜• ๋˜๋Š” ์ดˆ์„ ํ˜•์œผ๋กœ ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  GSM8K ๋ฐ MATH ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•  ๋•Œ ์ •ํ™•๋„๊ฐ€ ์ •์ฒด๊ธฐ์— ๋„๋‹ฌํ•˜์ง€ ๋ชปํ•œ๋‹ค๋Š” ์ (ํ‘œ 1 ์ฐธ์กฐ)์— ์ฃผ๋ชฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์€ SFT ๋ฐ์ดํ„ฐ๋ฅผ ๋” ํ™•์žฅํ•˜๋„๋ก ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ง€์†์ ์ธ ํ™•์žฅ์„ ์ง€์›ํ•  ์ˆ˜ ์žˆ๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์–ด๋ ค์›€์— ์ง๋ฉดํ•ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To address the instability issue, we first observe that the accuracy improves almost in linear or even super-linear with exponentially increased supervised fine-tuning (SFT) data. Moreover, we note that the accuracy is far from reaching a plateau when utilizing all available GSM8K and MATH training data (as shown in Table 1). This observation encourages us to further scale up the SFT data. However, we face a challenge as there is a lack of publicly accessible real data to support this continuous scaling. [|assistant|]
๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ๊ฐ๊ฐ 1100๊ฐœ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ ์ค‘ 1000๊ฐœ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ํ•ด๋‹น ์˜ˆ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ชจ๋‘์— ๋Œ€ํ•ด Llama-7b, Llama-13b ๋ฐ Llama-30b์— ๋Œ€ํ•œ hyperparameter sweeps๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์Šค์œ•์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก D.1์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์Šค์œ•์—์„œ ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•œ hyperparameters๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ๊ฐ 5๊ฐœ์˜ ์‹œ๋“œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 20๊ฐœ์˜ epochs ๋™์•ˆ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ๊ฐ๊ฐ 1100๊ฐœ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ ์ค‘ 1000๊ฐœ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ํ•ด๋‹น ์˜ˆ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ชจ๋‘์— ๋Œ€ํ•ด Llama-7b, Llama-13b ๋ฐ Llama-30b์— ๋Œ€ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์Šค์œ•์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์Šค์œ•์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก D.1์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์Šค์œ•์—์„œ ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ๊ฐ 5๊ฐœ์˜ ์‹œ๋“œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 20๊ฐœ์˜ ์—ํฌํฌ ๋™์•ˆ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The datasets contain 1100 question-answer pairs each. 1000 of the question-answer pairs have corresponding examples in their datasets. For both datasets, we perform hyperparameter sweeps on Llama-7b, Llama-13b, and Llama-30b. Details for the sweep can be found in Appendix D.1. Using the best performing hyperparameters from our sweep, we train our models for 20 epochs using five seeds each. [|assistant|]
๊ธ€์„ ์“ฐ๊ณ  ๋งํ•  ๋•Œ ์‚ฌ๋žŒ๋“ค์€ ์ƒ๊ฐ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋ฉˆ์ถ”๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก ์— ์ดˆ์ ์„ ๋งž์ถ˜ ์—ฐ๊ตฌ์—์„œ๋Š” ์ถ”๋ก ์„ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ฑฐ๋‚˜ ๋Œ€๋ฆฌ์ธ์˜ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ ๋ฌ˜์‚ฌํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์ง€๋งŒ, ์ถ”๋ก ์€ ๊ฑฐ์˜ ๋ชจ๋“  ์„œ๋ฉด ํ…์ŠคํŠธ์— ์•”์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ด๋Š” ์ฆ๋ช…์˜ ์ค„ ์‚ฌ์ด์— ๋ช…์‹œ๋˜์ง€ ์•Š์€ ๋‹จ๊ณ„ ๋˜๋Š” ๋Œ€ํ™”์˜ ๊ธฐ๋ฐ˜์ด ๋˜๋Š” ๋งˆ์Œ์˜ ์ด๋ก ์— ์ ์šฉ๋ฉ๋‹ˆ๋‹ค. Self-Taught Reasoner (STaR, Zelikman et al., 2022)์—์„œ๋Š” ์งˆ์˜์‘๋‹ต์—์„œ few-shot ์˜ˆ์ œ์—์„œ ์ถ”๋ก ์„ ์ถ”๋ก ํ•˜๊ณ  ์ •๋‹ต์œผ๋กœ ์ด์–ด์ง€๋Š” ์˜ˆ์ œ์—์„œ ํ•™์Šตํ•จ์œผ๋กœ์จ ์œ ์šฉํ•œ ์‚ฌ๊ณ ๋ฅผ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋งค์šฐ ์ œํ•œ๋œ ์„ค์ •์œผ๋กœ, ์–ธ์–ด ๋ชจ๋ธ์ด ๋Œ€์‹  ์ž„์˜์˜ ํ…์ŠคํŠธ์—์„œ ๋ช…์‹œ๋˜์ง€ ์•Š์€ ๊ทผ๊ฑฐ๋ฅผ ์ถ”๋ก ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค๋ฉด ์ด์ƒ์ ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” LMs์ด ๋ฏธ๋ž˜์˜ ํ…์ŠคํŠธ๋ฅผ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ํ† ํฐ์—์„œ ๊ทผ๊ฑฐ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ์˜ˆ์ธก์„ ๊ฐœ์„ ํ•˜๋Š” STaR์˜ ์ผ๋ฐ˜ํ™”์ธ **Quiet-STaR**์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. 1) ์—ฐ์† ์ƒ์„ฑ์˜ ๊ณ„์‚ฐ ๋น„์šฉ, 2) LMs์ด ์ฒ˜์Œ์—๋Š” ๋‚ด๋ถ€ ์ƒ๊ฐ์„ ์ƒ์„ฑํ•˜๊ฑฐ๋‚˜ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ๋ฅธ๋‹ค๋Š” ์‚ฌ์‹ค, 3) ๊ฐœ๋ณ„ ๋‹ค์Œ ํ† ํฐ์„ ๋„˜์–ด ์˜ˆ์ธกํ•ด์•ผ ํ•œ๋‹ค๋Š” ์  ๋“ฑ ์ฃผ์š” ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ๊ฐ์˜ ์‹œ์ž‘๊ณผ ๋์„ ๋‚˜ํƒ€๋‚ด๋Š” ํ† ํฐ๊ณผ ํ™•์žฅ๋œ teacher-forcing ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜๋Š” ํ† ํฐ๋ณ„ ๋ณ‘๋ ฌ ์ƒ˜ํ”Œ๋ง ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ณ ๋ฌด์ ์ธ ์ ์€ ์ƒ์„ฑ๋œ ๊ทผ๊ฑฐ๊ฐ€ ์˜ˆ์ธกํ•˜๊ธฐ ์–ด๋ ค์šด ํ† ํฐ์— ๋ถˆ๊ท ํ˜•์ ์œผ๋กœ ๋„์›€์ด ๋˜๊ณ  ๋ชจ๋ธ์ด ์–ด๋ ค์šด ์งˆ๋ฌธ์— ์ง์ ‘ ๋‹ตํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ, ์ธํ„ฐ๋„ท ํ…์ŠคํŠธ ์ฝ”ํผ์Šค์— ๋Œ€ํ•ด Quiet-STaR์„ ์‚ฌ์šฉํ•˜์—ฌ LMs์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•œ ํ›„, GSM8K(5.9%\(\rightarrow\)10.9%)์™€ CommonsenseQA(36.3%\(\rightarrow\)47.2%)์—์„œ zero-shot ๊ฐœ์„ ์„ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ, ์ž์—ฐ์–ด ํ…์ŠคํŠธ์—์„œ ์–ด๋ ค์šด ํ† ํฐ์˜ perplexity๊ฐ€ ๊ฐœ์„ ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ์ •์ ์œผ๋กœ, ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์—๋Š” ์ด๋Ÿฌํ•œ ์ž‘์—…์— ๋Œ€ํ•œ Fine-tuning์ด ํ•„์š”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. Quiet-STaR์€ ๋ณด๋‹ค ์ผ๋ฐ˜์ ์ด๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ฐฉ์‹์œผ๋กœ ์ถ”๋ก ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” LMs์„ ํ–ฅํ•œ ํ•œ ๊ฑธ์Œ์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ธ€์„ ์“ฐ๊ณ  ๋งํ•  ๋•Œ ์‚ฌ๋žŒ๋“ค์€ ์ƒ๊ฐ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋ฉˆ์ถ”๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก ์— ์ดˆ์ ์„ ๋งž์ถ˜ ์—ฐ๊ตฌ์—์„œ๋Š” ์ถ”๋ก ์„ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ฑฐ๋‚˜ ๋Œ€๋ฆฌ์ธ์˜ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ ๋ฌ˜์‚ฌํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์ง€๋งŒ, ์ถ”๋ก ์€ ๊ฑฐ์˜ ๋ชจ๋“  ์„œ๋ฉด ํ…์ŠคํŠธ์— ์•”์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ด๋Š” ์ฆ๋ช…์˜ ์ค„ ์‚ฌ์ด์— ๋ช…์‹œ๋˜์ง€ ์•Š์€ ๋‹จ๊ณ„ ๋˜๋Š” ๋Œ€ํ™”์˜ ๊ธฐ๋ฐ˜์ด ๋˜๋Š” ๋งˆ์Œ์˜ ์ด๋ก ์— ์ ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ž๊ธฐ ๊ฐ€๋ฅด์นจ ์ถ”๋ก ์ž(STaR, Zelikman et al., 2022)์—์„œ๋Š” ์งˆ์˜์‘๋‹ต์—์„œ ๋ช‡ ๋ฒˆ์˜ ์ƒท ์˜ˆ์ œ์—์„œ ์ถ”๋ก ์„ ์ถ”๋ก ํ•˜๊ณ  ์ •๋‹ต์œผ๋กœ ์ด์–ด์ง€๋Š” ์˜ˆ์ œ์—์„œ ํ•™์Šตํ•จ์œผ๋กœ์จ ์œ ์šฉํ•œ ์‚ฌ๊ณ ๋ฅผ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋งค์šฐ ์ œํ•œ๋œ ์„ค์ •์œผ๋กœ, ์–ธ์–ด ๋ชจ๋ธ์ด ๋Œ€์‹  ์ž„์˜์˜ ํ…์ŠคํŠธ์—์„œ ๋ช…์‹œ๋˜์ง€ ์•Š์€ ๊ทผ๊ฑฐ๋ฅผ ์ถ”๋ก ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค๋ฉด ์ด์ƒ์ ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” LM์ด ๋ฏธ๋ž˜์˜ ํ…์ŠคํŠธ๋ฅผ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ํ† ํฐ์—์„œ ๊ทผ๊ฑฐ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ์˜ˆ์ธก์„ ๊ฐœ์„ ํ•˜๋Š” STaR์˜ ์ผ๋ฐ˜ํ™”์ธ **Quiet-STaR**์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. 1) ์—ฐ์† ์ƒ์„ฑ์˜ ๊ณ„์‚ฐ ๋น„์šฉ, 2) LM์ด ์ฒ˜์Œ์—๋Š” ๋‚ด๋ถ€ ์ƒ๊ฐ์„ ์ƒ์„ฑํ•˜๊ฑฐ๋‚˜ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ๋ฅธ๋‹ค๋Š” ์‚ฌ์‹ค, 3) ๊ฐœ๋ณ„ ๋‹ค์Œ ํ† ํฐ์„ ๋„˜์–ด ์˜ˆ์ธกํ•ด์•ผ ํ•œ๋‹ค๋Š” ์  ๋“ฑ ์ฃผ์š” ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํ† ํฐ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ๊ฐ์˜ ์‹œ์ž‘๊ณผ ๋์„ ๋‚˜ํƒ€๋‚ด๋Š” ํ† ํฐ๊ณผ ํ™•์žฅ๋œ ๊ต์‚ฌ ๊ฐ•์ œ ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜๋Š” ํ† ํฐ๋ณ„ ๋ณ‘๋ ฌ ์ƒ˜ํ”Œ๋ง ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ณ ๋ฌด์ ์ธ ์ ์€ ์ƒ์„ฑ๋œ ๊ทผ๊ฑฐ๊ฐ€ ์˜ˆ์ธกํ•˜๊ธฐ ์–ด๋ ค์šด ํ† ํฐ์— ๋ถˆ๊ท ํ˜•์ ์œผ๋กœ ๋„์›€์ด ๋˜๊ณ  ๋ชจ๋ธ์ด ์–ด๋ ค์šด ์งˆ๋ฌธ์— ์ง์ ‘ ๋‹ตํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ, ์ธํ„ฐ๋„ท ํ…์ŠคํŠธ ์ฝ”ํผ์Šค์— ๋Œ€ํ•ด Quiet-STaR์„ ์‚ฌ์šฉํ•˜์—ฌ LM์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•œ ํ›„, GSM8K(5.9%\(\์šฐ์ธก๋Œ€๊ฐ์„ \)10.9%)์™€ CommonsenseQA(36.3%\(\์šฐ์ธก๋Œ€๊ฐ์„ \)47.2%)์—์„œ 0์ƒท ๊ฐœ์„ ์„ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ, ์ž์—ฐ์–ด ํ…์ŠคํŠธ์—์„œ ์–ด๋ ค์šด ํ† ํฐ์˜ ๋‹นํ˜น๊ฐ์ด ๊ฐœ์„ ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ์ •์ ์œผ๋กœ, ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์—๋Š” ์ด๋Ÿฌํ•œ ์ž‘์—…์— ๋Œ€ํ•œ ๋ฏธ์„ธ ์กฐ์ •์ด ํ•„์š”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. Quiet-STaR์€ ๋ณด๋‹ค ์ผ๋ฐ˜์ ์ด๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ฐฉ์‹์œผ๋กœ ์ถ”๋ก ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” LM์„ ํ–ฅํ•œ ํ•œ ๊ฑธ์Œ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]When writing and talking, people sometimes pause to think. Although reasoning-focused works have often framed reasoning as a method of answering questions or completing agentic tasks, reasoning is implicit in almost all written text. For example, this applies to the steps not stated between the lines of a proof or to the theory of mind underlying a conversation. In the Self-Taught Reasoner (STaR, Zelikman et al., 2022), useful thinking is learned by inferring rationales from few-shot examples in question-answering and learning from those that lead to a correct answer. This is a highly constrained setting - ideally, a language model could instead learn to infer unstated rationales in arbitrary text. We present **Quiet-STaR**, a generalization of STaR in which LMs learn to generate rationales at each token to explain future text, improving their predictions. We address key challenges, including 1) the computational cost of generating continuations, 2) the fact that the LM does not initially know how to generate or use internal thoughts, and 3) the need to predict beyond individual next tokens. To resolve these, we propose a tokenwise parallel sampling algorithm, using learnable tokens indicating a thought's start and end, and an extended teacher-forcing technique. Encouragingly, generated rationales disproportionately help model difficult-to-predict tokens and improve the LM's ability to directly answer difficult questions. In particular, after continued pretraining of an LM on a corpus of internet text with Quiet-STaR, we find zero-shot improvements on GSM8K (5.9%\(\rightarrow\)10.9%) and CommonsenseQA (36.3%\(\rightarrow\)47.2%) and observe a perplexity improvement of difficult tokens in natural text. Crucially, these improvements require no fine-tuning on these tasks. Quiet-STaR marks a step towards LMs that can learn to reason in a more general and scalable way. [|assistant|]
Hoffmann ๋“ฑ์˜ ๋…ผ๋ฌธ์€ ์–ธ์–ด ๋ชจ๋ธ๋ง ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ํฐ ์˜ํ–ฅ๋ ฅ์„ ๋ฐœํœ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” Google์˜ Gemini ์ œํ’ˆ๊ตฐ(Google, 2023)๊ณผ ๊ฐ™์€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ๋ชจ๋ธ์˜ scaling ์ •์ฑ…์— ์˜ํ–ฅ์„ ๋ฏธ์ณค์œผ๋ฉฐ, ๋‹ค๋ฅธ ๋งŽ์€ ๋ชจ๋ธ์˜ ๊ฐœ๋ฐœ์—๋„ ์˜ํ–ฅ์„ ๋ฏธ์ณค์„ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด ๋…ผ๋ฌธ์— ์ œ์‹œ๋œ scaling law์˜ ํŠน์ • ๋งค๊ฐœ๋ณ€์ˆ˜ ์ถ”์ •์น˜๋Š” ๋จธ์‹  ๋Ÿฌ๋‹ ์ด๋ก ๊ณผ ๊ฐ™์€ ๊ณผํ•™์ ์œผ๋กœ๋„ ๋งค์šฐ ์ค‘์š”ํ•œ ์˜๋ฏธ๋ฅผ ์ง€๋‹ˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ์˜ํ–ฅ๋ ฅ์„ ๊ณ ๋ คํ•  ๋•Œ, ์ด ์—ฐ๊ตฌ์˜ robustness์™€ reproducibility๋ฅผ ์ฒ ์ €ํžˆ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
Hoffmann ๋“ฑ์˜ ๋…ผ๋ฌธ์€ ์–ธ์–ด ๋ชจ๋ธ๋ง ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ํฐ ์˜ํ–ฅ๋ ฅ์„ ๋ฐœํœ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” Google์˜ Gemini ์ œํ’ˆ๊ตฐ(Google, 2023)๊ณผ ๊ฐ™์€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ๋ชจ๋ธ์˜ ํ™•์žฅ ์ •์ฑ…์— ์˜ํ–ฅ์„ ๋ฏธ์ณค์œผ๋ฉฐ, ๋‹ค๋ฅธ ๋งŽ์€ ๋ชจ๋ธ์˜ ๊ฐœ๋ฐœ์—๋„ ์˜ํ–ฅ์„ ๋ฏธ์ณค์„ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด ๋…ผ๋ฌธ์— ์ œ์‹œ๋œ ํ™•์žฅ ๋ฒ•์น™์˜ ํŠน์ • ๋งค๊ฐœ๋ณ€์ˆ˜ ์ถ”์ •์น˜๋Š” ๋จธ์‹  ๋Ÿฌ๋‹ ์ด๋ก ๊ณผ ๊ฐ™์€ ๊ณผํ•™์ ์œผ๋กœ๋„ ๋งค์šฐ ์ค‘์š”ํ•œ ์˜๋ฏธ๋ฅผ ์ง€๋‹ˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ์˜ํ–ฅ๋ ฅ์„ ๊ณ ๋ คํ•  ๋•Œ, ์ด ์—ฐ๊ตฌ์˜ ๊ฒฌ๊ณ ์„ฑ๊ณผ ์žฌํ˜„ ๊ฐ€๋Šฅ์„ฑ์„ ์ฒ ์ €ํžˆ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Hoffmann et al.'s paper has been highly influential in the language modeling community. Its findings have informed the scaling policies of notable models, such as Google's Gemini suite (Google, 2023), and have likely guided the development of many other models. Moreover, the specific parameter estimates of the scaling law presented in the paper are of significant scientific interest, such as for machine learning theory. Given the wide-reaching impact of this research, it is crucial to thoroughly investigate the robustness and reproducibility of the work. [|assistant|]
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.[|endofturn|]
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.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]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. [|assistant|]
BM25์— ์‚ฌ์šฉ๋˜๋Š” ํ† ํฐ์˜ ๊ธธ์ด๊ฐ€ ๋‹ค์–‘ํ•จ์— ๋”ฐ๋ผ BM25๋กœ ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์˜ ๊ธธ์ด๋ฅผ ๋Š˜๋ฆฌ๋ฉด ์ „์ฒด ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค(Figure 12). ๊ทธ๋Ÿฌ๋‚˜ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ „์ฒด ์ฑ…์˜ ๋‚ด์šฉ์œผ๋กœ ๋ชจ๋ธ์— promptingํ•˜๋Š” ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ ๋ฐฉ๋ฒ•๋ณด๋‹ค ์‹ ๋ขฐํ•  ์ˆ˜ ์—†๋Š” ์ฃผ์žฅ์„ ์‹๋ณ„ํ•˜๋Š” ๋ฐ๋Š” ์—ฌ์ „ํžˆ ๋œ ํšจ๊ณผ์ ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ๋” ๊ธด ๊ตฌ์ ˆ์กฐ์ฐจ๋„ ๊ด‘๋ฒ”์œ„ํ•œ ์ฃผ์žฅ์˜ ๊ฒ€์ฆ์— ํ•„์š”ํ•œ ์ „์ฒด context๋ฅผ ์ œ๊ณตํ•˜์ง€ ๋ชปํ•˜๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
BM25์— ์‚ฌ์šฉ๋˜๋Š” ํ† ํฐ์˜ ๊ธธ์ด๊ฐ€ ๋‹ค์–‘ํ•จ์— ๋”ฐ๋ผ BM25๋กœ ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์˜ ๊ธธ์ด๋ฅผ ๋Š˜๋ฆฌ๋ฉด ์ „์ฒด ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 12). ๊ทธ๋Ÿฌ๋‚˜ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ „์ฒด ์ฑ…์˜ ๋‚ด์šฉ์œผ๋กœ ๋ชจ๋ธ์— ํ”„๋กฌํ”„ํŠธํ•˜๋Š” ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ ๋ฐฉ๋ฒ•๋ณด๋‹ค ์‹ ๋ขฐํ•  ์ˆ˜ ์—†๋Š” ์ฃผ์žฅ์„ ์‹๋ณ„ํ•˜๋Š” ๋ฐ๋Š” ์—ฌ์ „ํžˆ ๋œ ํšจ๊ณผ์ ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ๋” ๊ธด ๊ตฌ์ ˆ์กฐ์ฐจ๋„ ๊ด‘๋ฒ”์œ„ํ•œ ์ฃผ์žฅ์˜ ๊ฒ€์ฆ์— ํ•„์š”ํ•œ ์ „์ฒด ์ปจํ…์ŠคํŠธ๋ฅผ ์ œ๊ณตํ•˜์ง€ ๋ชปํ•˜๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Varying length of tokens used in BM25As we increase the length of BM25-retrieved passages, the overall performance improves (Figure 12). However, this approach remains less effective for identifying unfaithful claims than our best performing method, i.e., prompting the model with the content of the entire book. This is likely due to the fact that even longer passages may not provide the entire context needed for verification of broader claims. [|assistant|]
* **์ž‘์—… ๋™์  ์˜ˆ์ธก์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ:** LLM์˜ ํ›ˆ๋ จ ๊ณผ์ •์—์„œ ๋„๋ฉ”์ธ ๋‚ด ๊ธฐ์กด ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์˜ ์—ญํ•™์ด ๋ณด์ด์ง€ ์•Š๋Š” ์ž‘์—…์˜ ์—ญํ•™์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์•Œ๋ ค์ง„ ์ž‘์—…์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ๋™์ผํ•œ ๋„๋ฉ”์ธ ๋‚ด์—์„œ ๋ณด์ด์ง€ ์•Š๋Š” ์œ ์‚ฌํ•œ ์ž‘์—…์—์„œ ์–ด๋–ป๊ฒŒ ์ˆ˜ํ–‰๋˜๋Š”์ง€์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. (4.1.1์ ˆ) * **๋ฒ”์šฉ์„ฑ ํ–ฅ์ƒ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ:** ์ธ๊ฐ„์˜ ์ธ์ง€ ๊ณผ์ •๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ, ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์— ๊ฑธ์นœ ๋‹ค์–‘ํ•œ ๋Šฅ๋ ฅ์˜ ํ–ฅ์ƒ์€ ์ปค๋ฆฌํ˜๋Ÿผ ํ•™์Šต์— ๋”ฐ๋ผ ๊ธฐ์ดˆ๋ถ€ํ„ฐ ๊ณ ๊ธ‰ ๋‹จ๊ณ„๊นŒ์ง€ ์ง„ํ–‰๋ฉ๋‹ˆ๋‹ค. ๋„๋ฉ”์ธ ๊ฐ„ ์ปค๋ฆฌํ˜๋Ÿผ์€ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ๋ฐฉํ–ฅ์„ ์•ˆ๋‚ดํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํ•œ ๋„๋ฉ”์ธ์—์„œ ์–ป์€ ์ธ์‚ฌ์ดํŠธ๋Š” ์ž ์žฌ์ ์œผ๋กœ ๋‹ค๋ฅธ ๋„๋ฉ”์ธ์˜ ํ•™์Šต ๊ณผ์ •์„ ์ด‰์ง„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (4.1.2์ ˆ) * **ํ›ˆ๋ จ ์ „๋žต, ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ ๋“ฑ์˜ ํšจ๊ณผ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ:** 7b ๊ทœ๋ชจ์˜ ์—ฌ๋Ÿฌ ๋ชจ๋ธ(Baichuan2-7b, DeepSeek-7b, OpenLLaMA-7b, Amber-7b)์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ํ›ˆ๋ จ ์ „๋žต, ๋ฐ์ดํ„ฐ ์„ธํŠธ ํ’ˆ์งˆ, ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์˜ ํšจ๊ณผ๋ฅผ ์ข…ํ•ฉ์ ์œผ๋กœ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ, Learning Rate ์กฐ์ •, Batch Size, ์ •๊ทœํ™” ๊ธฐ๋ฒ•์ด ์ดˆ๊ธฐ ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ํ•™์Šต ํšจ์œจ์„ฑ์— ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค(4.2.1์ ˆ). * **์ถ”๋ก  ์ž‘์—…์— ๋Œ€ํ•œ ๋ชจ๋ธ ๊ทœ๋ชจ์˜ ํšจ๊ณผ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ:** Deepseek-7b, 34b ๋ฐ 67b์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๋ชจ๋ธ ๊ทœ๋ชจ์™€ ๋ณต์žก์„ฑ์ด ์ด๋Ÿฌํ•œ ์ถ”๋ก  ์ž‘์—…์— ๋Œ€ํ•œ ๋Šฅ๋ ฅ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ํŠน์ • ์ „๋žต์„ ์‚ฌ์šฉํ•˜๋ฉด ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ์ด ๋” ํฐ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ์ƒ์‹์  ์ถ”๋ก ์—์„œ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (4.2.2์ ˆ) * **Scaling ๋ฒ•์น™์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ:** (1). ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ํด์ˆ˜๋ก ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋ชจ๋ธ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” LLM ํ›ˆ๋ จ์— ๊ด‘๋ฒ”์œ„ํ•œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฐ์ดํ„ฐ์…‹์ด ์ปค์งˆ์ˆ˜๋ก ์ถ”๊ฐ€ ๋ฐ์ดํ„ฐ์˜ ์ด์ ์ด ๊ฐ์†Œํ•˜์—ฌ ์„ฑ๋Šฅ ํ–ฅ์ƒ์— ๋Œ€ํ•œ ํ•œ๊ณ„์— ๊ฐ€๊นŒ์›Œ์ง€๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (2). Scaling ๋ฒ•์น™(Hoffmann et al., 2022)์˜ ์ •ํ™•๋„๊ฐ€ ๋ชจ๋ธ๋งˆ๋‹ค ํฌ๊ฒŒ ๋‹ค๋ฅด๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Š” ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ ๋ฐ ๊ณ„์‚ฐ ๋ณต์žก์„ฑ๊ณผ ๊ฐ™์€ ์š”์ธ์ด Scaling ํšจ์œจ์„ฑ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ํŠนํžˆ ์ผ๋ถ€ ๋ชจ๋ธ์€ ์ด ๋ฒ•์น™์— ๋” ์ž˜ ๋ถ€ํ•ฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚˜ ๋ฐ์ดํ„ฐ ํ™œ์šฉ ๋ฐ ํ•™์Šต ํšจ์œจ์„ฑ ์ธก๋ฉด์—์„œ ์ž ์žฌ์ ์ธ ์ด์ ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. (3). Scaling ๋ฒ•์น™์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ์˜ ์˜ํ–ฅ์— ๋Œ€ํ•œ ์œ ์šฉํ•œ ๊ด€์ ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์‹ค์ œ ์„ฑ๋Šฅ Scaling์€ ๋ฏธ๋ฌ˜ํ•œ ์ฐจ์ด๋ฅผ ๋ณด์ด๋ฉฐ, ์ด๋Š” ๋ฐ์ดํ„ฐ ๋ณผ๋ฅจ, ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ ๋ฐ ๊ณ„์‚ฐ ์ „๋žต ๊ฐ„์˜ ๋ณต์žกํ•œ ์ƒํ˜ธ ์ž‘์šฉ์„ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋˜ํ•œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ฐ์ดํ„ฐ์—์„œ ํ•™์Šต ๊ฒฐ๊ณผ๋ฅผ ๊ทน๋Œ€ํ™”ํ•˜๊ธฐ ์œ„ํ•ด Scaling ๋ฒ•์น™๊ณผ ๋ชจ๋ธ ์ตœ์ ํ™”์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์—ฐ๊ตฌ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค(์„น์…˜ 5).[|endofturn|]
* **์ž‘์—… ๋™์  ์˜ˆ์ธก์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ:** LLM์˜ ํ›ˆ๋ จ ๊ณผ์ •์—์„œ ๋„๋ฉ”์ธ ๋‚ด ๊ธฐ์กด ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์˜ ์—ญํ•™์ด ๋ณด์ด์ง€ ์•Š๋Š” ์ž‘์—…์˜ ์—ญํ•™์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์•Œ๋ ค์ง„ ์ž‘์—…์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ๋™์ผํ•œ ๋„๋ฉ”์ธ ๋‚ด์—์„œ ๋ณด์ด์ง€ ์•Š๋Š” ์œ ์‚ฌํ•œ ์ž‘์—…์—์„œ ์–ด๋–ป๊ฒŒ ์ˆ˜ํ–‰๋˜๋Š”์ง€์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. (4.1.1์ ˆ)* **๋ฒ”์šฉ์„ฑ ํ–ฅ์ƒ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ:** ์ธ๊ฐ„์˜ ์ธ์ง€ ๊ณผ์ •๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ, ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์— ๊ฑธ์นœ ๋‹ค์–‘ํ•œ ๋Šฅ๋ ฅ์˜ ํ–ฅ์ƒ์€ ์ปค๋ฆฌํ˜๋Ÿผ ํ•™์Šต์— ๋”ฐ๋ผ ๊ธฐ์ดˆ๋ถ€ํ„ฐ ๊ณ ๊ธ‰ ๋‹จ๊ณ„๊นŒ์ง€ ์ง„ํ–‰๋ฉ๋‹ˆ๋‹ค. ๋„๋ฉ”์ธ ๊ฐ„ ์ปค๋ฆฌํ˜๋Ÿผ์€ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ๋ฐฉํ–ฅ์„ ์•ˆ๋‚ดํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํ•œ ๋„๋ฉ”์ธ์—์„œ ์–ป์€ ์ธ์‚ฌ์ดํŠธ๋Š” ์ž ์žฌ์ ์œผ๋กœ ๋‹ค๋ฅธ ๋„๋ฉ”์ธ์˜ ํ•™์Šต ๊ณผ์ •์„ ์ด‰์ง„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (4.1.2์ ˆ)* **ํ›ˆ๋ จ ์ „๋žต, ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ ๋“ฑ์˜ ํšจ๊ณผ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ: ** 7b ๊ทœ๋ชจ์˜ ์—ฌ๋Ÿฌ ๋ชจ๋ธ(Baichuan2-7b, DeepSeek-7b, OpenLLaMA-7b, Amber-7b)์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ํ›ˆ๋ จ ์ „๋žต, ๋ฐ์ดํ„ฐ ์„ธํŠธ ํ’ˆ์งˆ, ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์˜ ํšจ๊ณผ๋ฅผ ์ข…ํ•ฉ์ ์œผ๋กœ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ, ํ•™์Šต ์†๋„ ์กฐ์ •, ๋ฐฐ์น˜ ํฌ๊ธฐ, ์ •๊ทœํ™” ๊ธฐ๋ฒ•์ด ์ดˆ๊ธฐ ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ํ•™์Šต ํšจ์œจ์„ฑ์— ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค(4.2.1์ ˆ). * **์ถ”๋ก  ์ž‘์—…์— ๋Œ€ํ•œ ๋ชจ๋ธ ๊ทœ๋ชจ์˜ ํšจ๊ณผ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ: ** Deepseek-7b, 34b ๋ฐ 67b์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๋ชจ๋ธ ๊ทœ๋ชจ์™€ ๋ณต์žก์„ฑ์ด ์ด๋Ÿฌํ•œ ์ถ”๋ก  ์ž‘์—…์— ๋Œ€ํ•œ ๋Šฅ๋ ฅ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ํŠน์ • ์ „๋žต์„ ์‚ฌ์šฉํ•˜๋ฉด ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ์ด ๋” ํฐ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ์ƒ์‹์  ์ถ”๋ก ์—์„œ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (4.2.2์ ˆ)* **์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ:** (1). ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ํด์ˆ˜๋ก ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋ชจ๋ธ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” LLM ํ›ˆ๋ จ์— ๊ด‘๋ฒ”์œ„ํ•œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฐ์ดํ„ฐ์…‹์ด ์ปค์งˆ์ˆ˜๋ก ์ถ”๊ฐ€ ๋ฐ์ดํ„ฐ์˜ ์ด์ ์ด ๊ฐ์†Œํ•˜์—ฌ ์„ฑ๋Šฅ ํ–ฅ์ƒ์— ๋Œ€ํ•œ ํ•œ๊ณ„์— ๊ฐ€๊นŒ์›Œ์ง€๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (2). ์Šค์ผ€์ผ๋ง ๋ฒ•์น™(Hoffmann et al., 2022)์˜ ์ •ํ™•๋„๊ฐ€ ๋ชจ๋ธ๋งˆ๋‹ค ํฌ๊ฒŒ ๋‹ค๋ฅด๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Š” ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ ๋ฐ ๊ณ„์‚ฐ ๋ณต์žก์„ฑ๊ณผ ๊ฐ™์€ ์š”์ธ์ด ์Šค์ผ€์ผ๋ง ํšจ์œจ์„ฑ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ํŠนํžˆ ์ผ๋ถ€ ๋ชจ๋ธ์€ ์ด ๋ฒ•์น™์— ๋” ์ž˜ ๋ถ€ํ•ฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚˜ ๋ฐ์ดํ„ฐ ํ™œ์šฉ ๋ฐ ํ•™์Šต ํšจ์œจ์„ฑ ์ธก๋ฉด์—์„œ ์ž ์žฌ์ ์ธ ์ด์ ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. (3). ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ์˜ ์˜ํ–ฅ์— ๋Œ€ํ•œ ์œ ์šฉํ•œ ๊ด€์ ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์‹ค์ œ ์„ฑ๋Šฅ ์Šค์ผ€์ผ๋ง์€ ๋ฏธ๋ฌ˜ํ•œ ์ฐจ์ด๋ฅผ ๋ณด์ด๋ฉฐ, ์ด๋Š” ๋ฐ์ดํ„ฐ ๋ณผ๋ฅจ, ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ ๋ฐ ๊ณ„์‚ฐ ์ „๋žต ๊ฐ„์˜ ๋ณต์žกํ•œ ์ƒํ˜ธ ์ž‘์šฉ์„ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋˜ํ•œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ฐ์ดํ„ฐ์—์„œ ํ•™์Šต ๊ฒฐ๊ณผ๋ฅผ ๊ทน๋Œ€ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์Šค์ผ€์ผ๋ง ๋ฒ•์น™๊ณผ ๋ชจ๋ธ ์ตœ์ ํ™”์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์—ฐ๊ตฌ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค(์„น์…˜ 5).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* **Findings on task dynamic prediction:** Within the training process of LLMs, we observe that the dynamics of existing downstream tasks within a domain can predict the dynamics of unseen tasks. This suggests that a model's performance on known tasks can inform us about how it might perform on similar, yet unseen tasks within the same domain. (Section 4.1.1) * **Findings on cross-domain promotion:** Similar to the human cognitive process, the enhancement of different abilities across various domains progresses from basic to advanced levels following curriculum learning. The curriculum between cross-domain tasks can guide the training direction of models and the insights gained from one domain can potentially promote the learning process of other domains. (Section 4.1.2) * **Findings on the effect of training strategies, model architecture, etc. :** Based on the results of several 7b-scale models (Baichuan2-7b, DeepSeek-7b, OpenLLaMA-7b, and Amber-7b), we comprehensively analyze the effect of training strategies, dataset quality, and model architecture in training LLMs. For example, we observe that the training dataset, learning rate adjustments, batch size, and regularization techniques play a significant role in the learning efficiency in the early training stage (Section 4.2.1). * **Findings on the effect of model scale on reasoning tasks:** Based on the results of Deepseek-7b, 34b and 67b, we observe that the model size and complexity significantly influence its ability on these reasoning tasks. Nonetheless, employing particular strategies can enhance smaller-scale models to obtain similar performance on commonsense reasoning when compared with their larger counterparts. (Section 4.2.2) * **Findings on the scaling law:** (1). We observe that larger training datasets lead to improved model performance on various benchmarks, which demonstrates the effect of extensive training data in training LLMs. However, the benefit of additional data diminishes as datasets grow, suggesting an approaching limit to performance gains. (2). We observe that the accuracies of the scaling law (Hoffmann et al., 2022) vary a lot across different models, which indicates that factors such as model architecture and computational complexity significantly influence scaling efficiency. Notably, some models demonstrate better alignment with the law, suggesting potential advantages in data utilization and learning efficiency. (3). Although the scaling law can provide a useful perspective on the impact of training data size, the actual performance scaling is nuanced, which reflects the complex interplay between data volume, model architecture, and computational strategies. This also highlights the importance of continued research into scaling laws and model optimization to maximize learning outcomes from available data (Section 5). [|assistant|]
๋ช…๋ น์–ด ๋ฏธ์„ธ ์กฐ์ •์˜ ๊ฐ„์„ญ์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด Llama2-7b-base ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. Fine-tuning ์ž‘์—…์œผ๋กœ ์ธ๊ณผ์  ์–ธ์–ด ๋ชจ๋ธ๋ง์„ ์„ ํƒํ•˜๊ณ  ์œ„ํ‚คํ”ผ๋””์•„ ๋ง๋ญ‰์น˜์—์„œ ์ถ”์ถœํ•œ ๊ฐ ์–ธ์–ด์— ๋Œ€ํ•ด \(200\)๊ฐœ์˜ ๋ฌธ์„œ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๊ฐœ์„  ์‚ฌํ•ญ์€ ์ดํ•ด์™€ ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ๊ฐ•ํ™”ํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ™•์žฅํ•˜๊ฑฐ๋‚˜ ์ง€์‹ ๊ธฐ๋ฐ˜์„ ๋„“ํžˆ๋Š” ๊ฒƒ์€ ๋ณด๋‹ค ๊ตฌ์ฒด์ ์ธ ๋ฐ์ดํ„ฐ ์ค€๋น„๊ฐ€ ํ•„์š”ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํ•˜์ง€ ์•Š๋Š”๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํƒ€๊ฒŸํŒ…๋œ ์ดํ•ด ๋ฐ ์ƒ์„ฑ ์ž‘์—…์„ ํ†ตํ•ด ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 8์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ช…๋ น์–ด ๋ฏธ์„ธ ์กฐ์ •์˜ ๊ฐ„์„ญ์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด Llama2-7b-base ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ๋ฏธ์„ธ ์กฐ์ • ์ž‘์—…์œผ๋กœ ์ธ๊ณผ์  ์–ธ์–ด ๋ชจ๋ธ๋ง์„ ์„ ํƒํ•˜๊ณ  ์œ„ํ‚คํ”ผ๋””์•„ ๋ง๋ญ‰์น˜์—์„œ ์ถ”์ถœํ•œ ๊ฐ ์–ธ์–ด์— ๋Œ€ํ•ด \(200\)๊ฐœ์˜ ๋ฌธ์„œ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.4 ์šฐ๋ฆฌ์˜ ๊ฐœ์„  ์‚ฌํ•ญ์€ ์ดํ•ด์™€ ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ๊ฐ•ํ™”ํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ™•์žฅํ•˜๊ฑฐ๋‚˜ ์ง€์‹ ๊ธฐ๋ฐ˜์„ ๋„“ํžˆ๋Š” ๊ฒƒ์€ ๋ณด๋‹ค ๊ตฌ์ฒด์ ์ธ ๋ฐ์ดํ„ฐ ์ค€๋น„๊ฐ€ ํ•„์š”ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํ•˜์ง€ ์•Š๋Š”๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํƒ€๊ฒŸํŒ…๋œ ์ดํ•ด ๋ฐ ์ƒ์„ฑ ์ž‘์—…์„ ํ†ตํ•ด ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 8์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We employ Llama2-7b-base model for enhancement to eliminate the interference of instruction fine-tuning. We select causal language modeling as our fine-tuning task and create a dataset comprising \(200\) documents for each language, extracted from the Wikipedia corpus.4 It is important to note that our enhancements are focused on augmenting the model's capabilities in understanding and generation only; we do not extend its reasoning faculties or broaden its knowledge base as it may require more specific data preparation. Accordingly, we evaluate the efficacy of our enhancements through targeted understanding and generation tasks. Detailed experiment results are shown in Table 8. [|assistant|]
๊ทธ๋ฆผ A14: ๋ฐ˜๋ณต ํ† ํฐ๊ณผ D4(๋ถ„ํ™์ƒ‰ ์„ ), ์ƒˆ๋กœ์šด ํ† ํฐ์„ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ(๊ฐ€๋กœ ์ ์„  ํšŒ์ƒ‰ ์„ ), ๋ฐ์ดํ„ฐ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ๋ฐ˜๋ณต(ํšŒ์ƒ‰ ์„ )์„ ๋น„๊ตํ•œ ๊ทธ๋ฆผ์ž…๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ epoch ๋ฒˆํ˜ธ๋กœ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Y์ถ•์€ perplexity๋ฅผ ๋‚˜ํƒ€๋‚ด๊ณ  X์ถ•์€ epoch ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ A14: ๋ฐ˜๋ณต ํ† ํฐ๊ณผ D4(๋ถ„ํ™์ƒ‰ ์„ ), ์ƒˆ๋กœ์šด ํ† ํฐ์„ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ(๊ฐ€๋กœ ์ ์„  ํšŒ์ƒ‰ ์„ ), ๋ฐ์ดํ„ฐ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ๋ฐ˜๋ณต(ํšŒ์ƒ‰ ์„ )์„ ๋น„๊ตํ•œ ๊ทธ๋ฆผ์ž…๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์—ํฌํฌ ๋ฒˆํ˜ธ๋กœ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Y์ถ•์€ ๋‹นํ˜น๊ฐ์„ ๋‚˜ํƒ€๋‚ด๊ณ  X์ถ•์€ ์—ํฌํฌ ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure A14: Comparison of repeating tokens with D4 (pink line), randomly selecting new tokens (horizontal dashed gray line), and randomly repeating data (gray line). We see with different epoch numbers. The y-axis denotes perplexity, and x-axis denotes number of epochs. [|assistant|]
์ด๋Ÿฌํ•œ ๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ๋Š” ์—ฌ๋Ÿฌ ์นดํ…Œ๊ณ ๋ฆฌ ๋˜๋Š” ์ž‘์—…์˜ ์˜์–ด Instructions๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ผ๋ฐ˜ ์ง€์‹ QA ๋ฐ ์ฝ”๋”ฉ๊ณผ ๊ฐ™์€ ๋ช‡ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ ์‚ฌ์šฉ ์‚ฌ๋ก€๋ฅผ ๊ณต์œ ํ•˜์ง€๋งŒ, ์„œ๋กœ ๋‹ค๋ฅธ ๋ฒค์น˜๋งˆํฌ์—์„œ Instructions์˜ ๋ฒ”์œ„๋Š” ์‹ค์ œ๋กœ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Xu et al. (2023)์€ Instruction ๋ถ„ํฌ์—์„œ Evol-Instruct๊ฐ€ Vicuna์™€ ์–ด๋–ป๊ฒŒ ๋‹ค๋ฅธ์ง€ ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. Instruction ๋ถ„ํฌ์˜ ์ฐจ์ด๋Š” ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์ด ์žˆ๋Š” ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๋ชจ๋ฐฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ด๋Ÿฌํ•œ ๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ๋Š” ์—ฌ๋Ÿฌ ์นดํ…Œ๊ณ ๋ฆฌ ๋˜๋Š” ์ž‘์—…์˜ ์˜์–ด ๋ช…๋ น์–ด๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ผ๋ฐ˜ ์ง€์‹ QA ๋ฐ ์ฝ”๋”ฉ๊ณผ ๊ฐ™์€ ๋ช‡ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ ์‚ฌ์šฉ ์‚ฌ๋ก€๋ฅผ ๊ณต์œ ํ•˜์ง€๋งŒ, ์„œ๋กœ ๋‹ค๋ฅธ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋ช…๋ น์–ด์˜ ๋ฒ”์œ„๋Š” ์‹ค์ œ๋กœ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Xu et al. (2023)์€ ๋ช…๋ น์–ด ๋ถ„ํฌ์—์„œ Evol-Instruct๊ฐ€ Vicuna์™€ ์–ด๋–ป๊ฒŒ ๋‹ค๋ฅธ์ง€ ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๋ช…๋ น์–ด ๋ถ„ํฌ์˜ ์ฐจ์ด๋Š” ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์ด ์žˆ๋Š” ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๋ชจ๋ฐฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]All these benchmarks consist of English instructions from multiple categories or tasks. However, though sharing some common use cases such as general knowledge QA and coding, the coverage of the instructions in different benchmarks are indeed different. For example, Xu et al. (2023) discuss in detail how Evol-Instruct is different from Vicuna in instruction distribution. The difference between instruction distributions effectively mimic the practical scenario where we have different downstream tasks. [|assistant|]
* 1๋‹จ๊ณ„์—์„œ๋Š” (1)๊ณผ (2)๋ฅผ ๊ณต๋™์œผ๋กœ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. ๊ด€๋ จ์„ฑ๋ณ„ ์ˆœ์œ„ ์ง€์ • ๋ฐ \(k=15\) ์„ค์ •์€ ํ‰๊ท  Brier score๋ฅผ ๊ฐ€์žฅ ๋‚ฎ๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค. * ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” (3) reasoning prompt๋ฅผ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์‹œ์Šคํ…œ์—์„œ ๊ธฐ๋ณธ ๋ชจ๋ธ์—์„œ 3๊ฐœ์˜ ์˜ˆ์ธก์„ ์ด๋Œ์–ด๋‚ด๊ธฐ ์œ„ํ•ด 15๊ฐœ์˜ ํ›„๋ณด ์ค‘ ์ƒ์œ„ 3๊ฐœ์˜ prompt๋ฅผ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 15 ์ฐธ์กฐ).[|endofturn|]
* 1๋‹จ๊ณ„์—์„œ๋Š” (1)๊ณผ (2)๋ฅผ ๊ณต๋™์œผ๋กœ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. ๊ด€๋ จ์„ฑ๋ณ„ ์ˆœ์œ„ ์ง€์ • ๋ฐ \(k=15\) ์„ค์ •์€ ํ‰๊ท  ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜๋ฅผ ๊ฐ€์žฅ ๋‚ฎ๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค. * ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” (3) ์ถ”๋ก  ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์‹œ์Šคํ…œ์—์„œ ๊ธฐ๋ณธ ๋ชจ๋ธ์—์„œ 3๊ฐœ์˜ ์˜ˆ์ธก์„ ์ด๋Œ์–ด๋‚ด๊ธฐ ์œ„ํ•ด 15๊ฐœ์˜ ํ›„๋ณด ์ค‘ ์ƒ์œ„ 3๊ฐœ์˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 15 ์ฐธ์กฐ).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* In the first stage, we jointly optimize (1) and (2). Ranking by relevance and setting \(k=15\) achieve the lowest average Brier score. * In the second stage, we optimize (3) the reasoning prompt. We identify the top 3 prompts out of 15 candidates to elicit 3 predictions from our base model in our system; see Figure 15 for the best one. [|assistant|]
์ƒˆ ํ† ํฐ์„ ์ƒ์„ฑํ•  ๋•Œ ๊ฐ€์žฅ ๋งŽ์€ ์‹œ๊ฐ„์ด ์†Œ์š”๋˜๋Š” ๋‹จ๊ณ„๋Š” ๋ฐ์ดํ„ฐ ์ „์†ก ๋ฐ weight ๊ณ„์‚ฐ๊ณผ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์š” ๋ฌธ์ œ๋Š” ๋ฐ์ดํ„ฐ ์ „์†ก์— ์••๋„๋˜๋Š” ์ƒ๋‹นํ•œ ์‹œ๊ฐ„์œผ๋กœ, ์ข…์ข… _memory wall_ ๋ฌธ์ œ๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ƒˆ ํ† ํฐ์„ ์ƒ์„ฑํ•  ๋•Œ ๊ฐ€์žฅ ๋งŽ์€ ์‹œ๊ฐ„์ด ์†Œ์š”๋˜๋Š” ๋‹จ๊ณ„๋Š” ๋ฐ์ดํ„ฐ ์ „์†ก ๋ฐ ๊ฐ€์ค‘์น˜ ๊ณ„์‚ฐ๊ณผ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์š” ๋ฌธ์ œ๋Š” ๋ฐ์ดํ„ฐ ์ „์†ก์— ์••๋„๋˜๋Š” ์ƒ๋‹นํ•œ ์‹œ๊ฐ„์œผ๋กœ, ์ข…์ข… _๋ฉ”๋ชจ๋ฆฌ ๋ฒฝ_ ๋ฌธ์ œ๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]When generating a new token, the most time-consuming steps revolve around data transfer and weight computation. A main issue is the significant amount of time overwhelmed by data transfer, often referred to as the _memory wall_ issue. [|assistant|]
**์ •๋ณด ๊ฒ€์ƒ‰(IR).** IR์€ LMs(Lewis et al., 2020; Shuster et al., 2021; Nakano et al., 2021)์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฒคํŠธ ์˜ˆ์ธก์—์„œ๋Š” ๋‹ค์–‘ํ•˜๊ณ  ์ตœ์‹  ์ •๋ณด์— ๋Œ€ํ•œ ์•ก์„ธ์Šค๊ฐ€ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค(Tetlock and Gardner, 2015). ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์˜ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ์งˆ์˜ ํ™•์žฅ, ๊ด€๋ จ์„ฑ ์ˆœ์œ„ ๋ฐ ์š”์•ฝ์„ ์œ„ํ•ด LMs์„ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๋ก  ๋ชจ๋ธ์— ๋‰ด์Šค ๊ธฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•˜๋Š” IR ์•„ํ‚คํ…์ฒ˜์ž…๋‹ˆ๋‹ค. ์ด ์„ค์ • ์™ธ์—๋„ IR์— LMs์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์€ ํ™œ๋ฐœํ•œ ์—ฐ๊ตฌ ์ฃผ์ œ์ž…๋‹ˆ๋‹ค(Zhu et al., 2024).[|endofturn|]
**์ •๋ณด ๊ฒ€์ƒ‰(IR).** IR์€ LM(Lewis et al., 2020; Shuster et al., 2021; Nakano et al., 2021)์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฒคํŠธ ์˜ˆ์ธก์—์„œ๋Š” ๋‹ค์–‘ํ•˜๊ณ  ์ตœ์‹  ์ •๋ณด์— ๋Œ€ํ•œ ์•ก์„ธ์Šค๊ฐ€ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค(Tetlock and Gardner, 2015). ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์˜ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ์งˆ์˜ ํ™•์žฅ, ๊ด€๋ จ์„ฑ ์ˆœ์œ„ ๋ฐ ์š”์•ฝ์„ ์œ„ํ•ด LM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๋ก  ๋ชจ๋ธ์— ๋‰ด์Šค ๊ธฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•˜๋Š” IR ์•„ํ‚คํ…์ฒ˜์ž…๋‹ˆ๋‹ค. ์ด ์„ค์ • ์™ธ์—๋„ IR์— LM์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์€ ํ™œ๋ฐœํ•œ ์—ฐ๊ตฌ ์ฃผ์ œ์ž…๋‹ˆ๋‹ค(Zhu et al., 2024).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Information retrieval (IR).** IR can improve question-answering capabilities of LMs (Lewis et al., 2020; Shuster et al., 2021; Nakano et al., 2021). In event forecasting, access to diverse, up-to-date information is crucial (Tetlock and Gardner, 2015). Thus, a key component of our system is an IR architecture that furnishes the reasoning model with news articles, using LMs for query expansion, relevance ranking and summarization. Beyond our setting, using LMs for IR is an active research topic (Zhu et al., 2024). [|assistant|]
Instructions๋Š” ๋‹ค์–‘ํ•œ ์ผ๋ฐ˜ ์ž‘์—…์—์„œ ์ธ์ƒ์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ˆ˜ํ•™ ๋ฐ ์ฝ”๋”ฉ ์ž‘์—…์—์„œ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด ๋‹ฌ๋ผ์ง„๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค(Dong et al., 2023). ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” 150๋งŒ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Instruction Fine-tuning์„ ํ†ตํ•ด ์ตœ๊ณ  ์„ฑ๋Šฅ์„ ์–ป๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค(Bi et al., 2024). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ scaling ํšจ๊ณผ์˜ ๋ณธ์งˆ์ ์ธ ์ด์œ ๋Š” ์ฒ ์ €ํžˆ ์กฐ์‚ฌ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ง€์นจ์€ ๋‹ค์–‘ํ•œ ์ผ๋ฐ˜ ์ž‘์—…์—์„œ ์ธ์ƒ์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ˆ˜ํ•™ ๋ฐ ์ฝ”๋”ฉ ์ž‘์—…์—์„œ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด ๋‹ฌ๋ผ์ง„๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค(Dong et al., 2023). ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” 150๋งŒ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ง€์‹œ ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด ์ตœ๊ณ  ์„ฑ๋Šฅ์„ ์–ป๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค(Bi et al., 2024). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ํ™•์žฅ ํšจ๊ณผ์˜ ๋ณธ์งˆ์ ์ธ ์ด์œ ๋Š” ์ฒ ์ €ํžˆ ์กฐ์‚ฌ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]instructions can yield impressive results in various general tasks. Other studies have indicated that performance scales with data size in mathematical and coding tasks Dong et al. (2023). Recent work Bi et al. (2024) even uses 1.5 million data for instruct fine-tuning to obtain top performance. However, the intrinsic reasons behind this scaling effect have not been thoroughly investigated. [|assistant|]
์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์€ ์˜ค๋žซ๋™์•ˆ ๋งค์šฐ ์–ด๋ ค์šด ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋˜์–ด ์ผ๋ฐ˜์ ์ธ Large Language Models (LLMs)์—์„œ ๋งค์šฐ ํฐ ๊ทœ๋ชจ์—์„œ๋งŒ ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐ๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, (Wei et al., 2022, 2020)์˜ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด 500์–ต ๊ฐœ ์ด์ƒ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ์—์„œ๋งŒ ์ˆ˜ํ•™ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์˜๋ฏธ ์žˆ๋Š” ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•˜๊ฑฐ๋‚˜ chain-of-thought processing์˜ ์ด์ ์„ ๋ˆ„๋ฆด ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋” ์ž‘์€ ์–ธ์–ด ๋ชจ๋ธ์— ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์„ ๋ถ€์—ฌํ•˜๋Š” ์ „๋žต์—๋Š” ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ์ˆ˜ํ•™ ๊ด€๋ จ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต๋œ ์ˆ˜ํ•™ ํŠนํ™” base models์„ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค(Lewkowycz et al., 2022; Azerbayev et al., 2023). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ์ •ํ™•๋„๋Š” ๋ฏธ๋ฏธํ•œ ์ˆ˜์ค€์— ๋จธ๋ฌผ๋Ÿฌ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Lemma-7B(Azerbayev et al., 2023)๋Š” GSM8K ๋ฐ์ดํ„ฐ ์„ธํŠธ(Cobbe et al., 2021)์—์„œ 36.4%, MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ(Hendrycks et al., 2021)์—์„œ 18.0%์˜ ์ •ํ™•๋„๋งŒ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์€ ์˜ค๋žซ๋™์•ˆ ๋งค์šฐ ์–ด๋ ค์šด ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋˜์–ด ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ๋งค์šฐ ํฐ ๊ทœ๋ชจ์—์„œ๋งŒ ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐ๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, (Wei et al., 2022, 2020)์˜ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด 500์–ต ๊ฐœ ์ด์ƒ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ์—์„œ๋งŒ ์ˆ˜ํ•™ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์˜๋ฏธ ์žˆ๋Š” ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•˜๊ฑฐ๋‚˜ ์ƒ๊ฐ์˜ ์—ฐ์‡„ ์ฒ˜๋ฆฌ์˜ ์ด์ ์„ ๋ˆ„๋ฆด ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋” ์ž‘์€ ์–ธ์–ด ๋ชจ๋ธ์— ์ˆ˜ํ•™์  ๋Šฅ๋ ฅ์„ ๋ถ€์—ฌํ•˜๋Š” ์ „๋žต์—๋Š” ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ์ˆ˜ํ•™ ๊ด€๋ จ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต๋œ ์ˆ˜ํ•™ ํŠนํ™” ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์„ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค(Lewkowycz et al., 2022; Azerbayev et al., 2023). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ์ •ํ™•๋„๋Š” ๋ฏธ๋ฏธํ•œ ์ˆ˜์ค€์— ๋จธ๋ฌผ๋Ÿฌ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ ˆ๋งˆ-7B(Azerbayev et al., 2023)๋Š” GSM8K ๋ฐ์ดํ„ฐ ์„ธํŠธ(Cobbe et al., 2021)์—์„œ 36.4%, MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ(Hendrycks et al., 2021)์—์„œ 18.0%์˜ ์ •ํ™•๋„๋งŒ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Mathematical capabilities have long been considered so challenging that they are thought to emerge in common language models only at a very large scale. For instance, studies by (Wei et al., 2022, 2020) suggest that only models with size exceeding 50 billion parameters can attain meaningful accuracy or benefit from chain-of-thought processing on math problems. A strategy to equip smaller language models with mathematical abilities involves creating math-specific base models trained on hundreds of billions of math-related pre-training data (Lewkowycz et al., 2022; Azerbayev et al., 2023). However, the accuracy of such models remains modest; for example, Lemma-7B (Azerbayev et al., 2023) only achieves 36.4% on the GSM8K dataset (Cobbe et al., 2021) and 18.0% on the MATH dataset (Hendrycks et al., 2021). [|assistant|]
**์ฝ˜ํ…์ธ  ๊ณ„ํš**. ์ด ์„ค๋ฌธ์กฐ์‚ฌ์— ์ƒˆ๋กœ์šด ์ฝ˜ํ…์ธ ๋ฅผ ์ •๊ธฐ์ ์œผ๋กœ ํฌํ•จ์‹œ์ผœ ๋ณด๋‹ค ์ž์ฒด์ ์œผ๋กœ ์™„์„ฑ๋˜๊ณ  ์ตœ์‹  ์ƒํƒœ๋ฅผ ์œ ์ง€ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‹ค์Œ์€ ๋‹ค์Œ ์ฃผ์š” ๋ฒ„์ „์— ํฌํ•จ๋  ์ˆ˜ ์žˆ๋Š” ๋ช‡ ๊ฐ€์ง€ ์ž ์žฌ์  ์ฃผ์ œ์ž…๋‹ˆ๋‹ค: (1) Instruction Tuning๊ณผ ๋Šฅ๋ ฅ ํ‰๊ฐ€ ๋ชจ๋‘์— ๋Œ€ํ•ด ๋” ํฐ Large Language Models (LLMs)๋ฅผ ์‚ฌ์šฉํ•œ ์‹คํ—˜, (2) ๋” ์ž์„ธํ•œ ํ”„๋กฌํ”„ํŠธ ์—ฐ์Šต, (3) ๊ต์œก ๋ ˆ์‹œํ”ผ, (4) ๋” ๋งŽ์€ ์ด๋ก ์  ๋ถ„์„ ๋ฐ ํ† ๋ก , (5) ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋Œ€ํ•œ ๋” ๋งŽ์€ ํ† ๋ก .[|endofturn|]
**์ฝ˜ํ…์ธ  ๊ณ„ํš**. ์ด ์„ค๋ฌธ์กฐ์‚ฌ์— ์ƒˆ๋กœ์šด ์ฝ˜ํ…์ธ ๋ฅผ ์ •๊ธฐ์ ์œผ๋กœ ํฌํ•จ์‹œ์ผœ ๋ณด๋‹ค ์ž์ฒด์ ์œผ๋กœ ์™„์„ฑ๋˜๊ณ  ์ตœ์‹  ์ƒํƒœ๋ฅผ ์œ ์ง€ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‹ค์Œ์€ ๋‹ค์Œ ์ฃผ์š” ๋ฒ„์ „์— ํฌํ•จ๋  ์ˆ˜ ์žˆ๋Š” ๋ช‡ ๊ฐ€์ง€ ์ž ์žฌ์  ์ฃผ์ œ์ž…๋‹ˆ๋‹ค: (1) ๋ช…๋ น์–ด ํŠœ๋‹๊ณผ ๋Šฅ๋ ฅ ํ‰๊ฐ€ ๋ชจ๋‘์— ๋Œ€ํ•ด ๋” ํฐ ์–ธ์–ด ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ ์‹คํ—˜, (2) ๋” ์ž์„ธํ•œ ํ”„๋กฌํ”„ํŠธ ์—ฐ์Šต, (3) ๊ต์œก ๋ ˆ์‹œํ”ผ, (4) ๋” ๋งŽ์€ ์ด๋ก ์  ๋ถ„์„ ๋ฐ ํ† ๋ก , (5) ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋Œ€ํ•œ ๋” ๋งŽ์€ ํ† ๋ก .[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Planning Content.** We will regularly include new content into this survey, to make it more self-contained and up-to-date. Here, we list several potential topics that might appear in the next major version(s): (1) more experiments with larger language models for both instruction tuning and ability evaluation; (2) more detailed prompting practice; (3) training recipe; (4) more theoretical analysis and discussion; (5) more discussions on applications. [|assistant|]
๋ฐ์ดํ„ฐ ๋Œ€ ๋ชจ๋ธ ๋น„์œจ์ด ํด์ˆ˜๋ก ์ด์ „๋ณด๋‹ค ๋” ์ž‘์€ ๋ชจ๋ธ์— ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋ฅผ ํก์ˆ˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ถ”๋ก  ๋ฐ ๋ฐฐํฌ์— ๋” ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ WSD LRS๊ฐ€ ๋” ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด ๋” ์ ์€ ๋…ธ๋ ฅ์œผ๋กœ \(L(N,D)\)๋ฅผ ํƒ๊ตฌํ•˜๊ณ  LLMs์—์„œ ๊ด€๊ณ„๋ฅผ ๋” ๋ช…ํ™•ํ•˜๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ฃผ๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.[|endofturn|]
๋ฐ์ดํ„ฐ ๋Œ€ ๋ชจ๋ธ ๋น„์œจ์ด ํด์ˆ˜๋ก ์ด์ „๋ณด๋‹ค ๋” ์ž‘์€ ๋ชจ๋ธ์— ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋ฅผ ํก์ˆ˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ถ”๋ก  ๋ฐ ๋ฐฐํฌ์— ๋” ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ WSD LRS๊ฐ€ ๋” ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด ๋” ์ ์€ ๋…ธ๋ ฅ์œผ๋กœ \(L(N,D)\)๋ฅผ ํƒ๊ตฌํ•˜๊ณ  LLM์—์„œ ๊ด€๊ณ„๋ฅผ ๋” ๋ช…ํ™•ํ•˜๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ฃผ๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Larger data-to-model ratio means that we can absorb more data into the a smaller model than we previously thought, which is more efficient for inference and deployment. And we hope WSD LRS will help more researchers to explore \(L(N,D)\) with less efforts and make the relationship clearer in LLMs. [|assistant|]
_์•ฝํ•œ ๋ถ„ํฌ ์ด๋™๊ณผ ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™ ๋ชจ๋‘์— ๋Œ€ํ•ด, ๊ฐ„๋‹จํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ Learning Rate ์žฌ๊ฐ€์—ด, Learning Rate ์žฌ๊ฐ์‡  ๋ฐ ์žฌ์ƒ์˜ ์กฐํ•ฉ์„ ํ™œ์šฉํ•˜์—ฌ \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\)์— ๋Œ€ํ•ด ํ•™์Šต๋œ ๋ชจ๋ธ๊ณผ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค.[|endofturn|]
_์•ฝํ•œ ๋ถ„ํฌ ์ด๋™๊ณผ ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™ ๋ชจ๋‘์— ๋Œ€ํ•ด, ๊ฐ„๋‹จํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ LR ์žฌ๊ฐ€์—ด, LR ์žฌ๊ฐ์‡  ๋ฐ ์žฌ์ƒ์˜ ์กฐํ•ฉ์„ ํ™œ์šฉํ•˜์—ฌ \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\)์— ๋Œ€ํ•ด ํ•™์Šต๋œ ๋ชจ๋ธ๊ณผ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]_In summary, for weak and stronger distribution shifts alike, it is possible to achieve competitive performance to a model trained on \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) by utilizing a simple and scalable combination of LR re-warming, LR re-decaying, and replay. [|assistant|]
* **NeuLab-TedTalks2**Qi et al. (2018): ์ „ ์„ธ๊ณ„ ์ž์›๋ด‰์‚ฌ์ž ์ปค๋ฎค๋‹ˆํ‹ฐ์— ์˜ํ•ด ์ˆ˜๋งŽ์€ ์ €์ž์› ์–ธ์–ด๋กœ ๋ฒˆ์—ญ๋œ TED ๊ฐ•์—ฐ์˜ ๊ณตํ†ต ์ฝ”ํผ์Šค์ž…๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ถœ๋ ฅ ํ˜•์‹ ์‹คํ—˜์„ ์œ„ํ•ด ์˜์–ด์™€ Traditional Chinese ํ•˜์œ„ ์ง‘ํ•ฉ์—์„œ ์ •๋ ฌ๋œ ๋ฌธ์žฅ 2000๊ฐœ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ง๋ญ‰์น˜๋Š” OPUS Tiedemann(2012)์—์„œ ๋‹ค์šด๋กœ๋“œํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
* **NeuLab-TedTalks2**Qi et al. (2018): ์ „ ์„ธ๊ณ„ ์ž์›๋ด‰์‚ฌ์ž ์ปค๋ฎค๋‹ˆํ‹ฐ์— ์˜ํ•ด ์ˆ˜๋งŽ์€ ์ €์ž์› ์–ธ์–ด๋กœ ๋ฒˆ์—ญ๋œ TED ๊ฐ•์—ฐ์˜ ๊ณตํ†ต ์ฝ”ํผ์Šค์ž…๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ถœ๋ ฅ ํ˜•์‹ ์‹คํ—˜์„ ์œ„ํ•ด ์˜์–ด์™€ ์ค‘๊ตญ์–ด(์ค‘๊ตญ์–ด) ํ•˜์œ„ ์ง‘ํ•ฉ์—์„œ ์ •๋ ฌ๋œ ๋ฌธ์žฅ 2000๊ฐœ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ง๋ญ‰์น˜๋Š” OPUS Tiedemann(2012)์—์„œ ๋‹ค์šด๋กœ๋“œํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* **NeuLab-TedTalks2**Qi et al. (2018): A common corpus of TED talks, translated into numerous low-resource languages by a global community of volunteers. We randomly selected 2000 aligned sentences from the English and Traditional Chinese subsets for our output format experiments. We download the corpus from OPUS Tiedemann (2012). [|assistant|]
7์ฒœ๋งŒ ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์ธ Pythia Biderman ๋“ฑ(2023)๋ถ€ํ„ฐ 1,800์–ต ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์ธ Falcon Almazrouei ๋“ฑ(2023)์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์˜คํ”ˆ ์†Œ์Šค Large Language Models (LLMs)๋ฅผ ๋ฐฑ๋ณธ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฒค์น˜๋งˆํฌ์™€ ์ž‘์—… ์ „๋ฐ˜์—์„œ ๋†’์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” 70์–ต ํŒŒ๋ผ๋ฏธํ„ฐ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์ธ Mistral 7B ๋ชจ๋ธ์„ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค(Jiang et al., 2023).[|endofturn|]
7์ฒœ๋งŒ ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์ธ Pythia Biderman ๋“ฑ(2023)๋ถ€ํ„ฐ 1,800์–ต ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์ธ Falcon Almazrouei ๋“ฑ(2023)์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์˜คํ”ˆ ์†Œ์Šค ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฐฑ๋ณธ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฒค์น˜๋งˆํฌ์™€ ์ž‘์—… ์ „๋ฐ˜์—์„œ ๋†’์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” 70์–ต ํŒŒ๋ผ๋ฏธํ„ฐ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์ธ Mistral 7B ๋ชจ๋ธ์„ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค(Jiang et al., 2023).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]A wide range of open-source large language models is available for the backbone, spanning from \(70\) million parameter models like Pythia Biderman et al. (2023) to \(180\) billion parameter models like Falcon Almazrouei et al. (2023). In this work, we choose the Mistral 7B model, a \(7\) billion parameter open-source model that achieves high performance across benchmarks and tasks Jiang et al. (2023). [|assistant|]
MHA์˜ ๊ฒฝ์šฐ, ๊ฐ ์‹œํ€€์Šค ์š”์†Œ์— ๋Œ€ํ•ด \(H\) ๊ฐœ์˜ attention ์ปจํ…์ŠคํŠธ ๋ฒกํ„ฐ๋ฅผ ๋ณ‘๋ ฌ๋กœ ๊ณ„์‚ฐํ•˜๊ณ , ์ด๋ฅผ 2์ฐจ์›์„ ๋”ฐ๋ผ ์—ฐ๊ฒฐํ•œ ๋‹ค์Œ, ์ตœ์ข…์ ์œผ๋กœ ์—ฐ๊ฒฐ๋œ ๋ฒกํ„ฐ๋ฅผ ๋ชจ๋ธ ๊ณต๊ฐ„์— ํˆฌ์˜ํ•˜์—ฌ attention ์‚ฐ์ถœ๋ฌผ์„ ์–ป๋Š”๋‹ค.[|endofturn|]
MHA์˜ ๊ฒฝ์šฐ, ๊ฐ ์‹œํ€€์Šค ์š”์†Œ์— ๋Œ€ํ•ด \(H\) ๊ฐœ์˜ ์ฃผ์˜ ์ปจํ…์ŠคํŠธ ๋ฒกํ„ฐ๋ฅผ ๋ณ‘๋ ฌ๋กœ ๊ณ„์‚ฐํ•˜๊ณ , ์ด๋ฅผ 2์ฐจ์›์„ ๋”ฐ๋ผ ์—ฐ๊ฒฐํ•œ ๋‹ค์Œ, ์ตœ์ข…์ ์œผ๋กœ ์—ฐ๊ฒฐ๋œ ๋ฒกํ„ฐ๋ฅผ ๋ชจ๋ธ ๊ณต๊ฐ„์— ํˆฌ์˜ํ•˜์—ฌ ์ฃผ์˜ ์‚ฐ์ถœ๋ฌผ์„ ์–ป๋Š”๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]For MHA, we compute \(H\) number of attention context vectors for each sequence element in parallel, concatenate them along the second dimension and then finally project the concatenated vector to the model space to obtain attention the output. [|assistant|]
๊ทธ๋ฆผ 2: ๋‹ค๊ตญ์–ด ๋ฌธ์„œ์— ์ ์šฉ๋œ monolingual (English) tokenizer์™€ multilingual tokenizer ๊ฐ„์˜ ํŒจ๋ฆฌํ‹ฐ ์ ์ˆ˜ ๋น„๊ต.[|endofturn|]
๊ทธ๋ฆผ 2: ๋‹ค๊ตญ์–ด ๋ฌธ์„œ์— ์ ์šฉ๋œ ๋‹จ์ผ ์–ธ์–ด(์˜์–ด) ํ† ํฐํ™”๊ธฐ์™€ ๋‹ค๊ตญ์–ด ํ† ํฐํ™”๊ธฐ ๊ฐ„์˜ ํŒจ๋ฆฌํ‹ฐ ์ ์ˆ˜ ๋น„๊ต.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 2: Comparison of parity scores between monolingual (English) tokenizer and multilingual tokenizers applied multi-lingual documents. [|assistant|]
์ฒซ ๋ฒˆ์งธ LLaMA ๋ชจ๋ธ ์„ธํŠธ[32]๋Š” 2023๋…„ 2์›”์— 7B์—์„œ 65B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ฒ”์œ„๋กœ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ๋“ค์€ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ˆ˜์ง‘ํ•œ ์ˆ˜์กฐ ๊ฐœ์˜ ํ† ํฐ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค. LLaMA๋Š” GPT-3์˜ Transformer ์•„ํ‚คํ…์ฒ˜๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, (1) ReLU ๋Œ€์‹  SwiGLU ํ™œ์„ฑํ™” ํ•จ์ˆ˜ ์‚ฌ์šฉ, (2) ์ ˆ๋Œ€ ์œ„์น˜ ์ž„๋ฒ ๋”ฉ ๋Œ€์‹  ๋กœํ„ฐ๋ฆฌ ์œ„์น˜ ์ž„๋ฒ ๋”ฉ ์‚ฌ์šฉ, (3) ํ‘œ์ค€ ๋ ˆ์ด์–ด ์ •๊ทœํ™” ๋Œ€์‹  ๋ฃจํŠธ ํ‰๊ท  ์ œ๊ณฑ ๋ ˆ์ด์–ด ์ •๊ทœํ™” ์‚ฌ์šฉ ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ์ž‘์€ ์•„ํ‚คํ…์ฒ˜ ์ˆ˜์ •์ด ์ด๋ฃจ์–ด์กŒ์Šต๋‹ˆ๋‹ค. ์˜คํ”ˆ ์†Œ์Šค LLaMA-13B ๋ชจ๋ธ์€ ๋Œ€๋ถ€๋ถ„์˜ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋…์  GPT-3(175B) ๋ชจ๋ธ์„ ๋Šฅ๊ฐ€ํ•˜๋ฏ€๋กœ LLM ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ์ข‹์€ ๊ธฐ์ค€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ฒซ ๋ฒˆ์งธ LLaMA ๋ชจ๋ธ ์„ธํŠธ[32]๋Š” 2023๋…„ 2์›”์— 7B์—์„œ 65B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ฒ”์œ„๋กœ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ๋“ค์€ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ˆ˜์ง‘ํ•œ ์ˆ˜์กฐ ๊ฐœ์˜ ํ† ํฐ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค. LLaMA๋Š” GPT-3์˜ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, (1) ReLU ๋Œ€์‹  SwiGLU ํ™œ์„ฑํ™” ํ•จ์ˆ˜ ์‚ฌ์šฉ, (2) ์ ˆ๋Œ€ ์œ„์น˜ ์ž„๋ฒ ๋”ฉ ๋Œ€์‹  ๋กœํ„ฐ๋ฆฌ ์œ„์น˜ ์ž„๋ฒ ๋”ฉ ์‚ฌ์šฉ, (3) ํ‘œ์ค€ ๋ ˆ์ด์–ด ์ •๊ทœํ™” ๋Œ€์‹  ๋ฃจํŠธ ํ‰๊ท  ์ œ๊ณฑ ๋ ˆ์ด์–ด ์ •๊ทœํ™” ์‚ฌ์šฉ ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ์ž‘์€ ์•„ํ‚คํ…์ฒ˜ ์ˆ˜์ •์ด ์ด๋ฃจ์–ด์กŒ์Šต๋‹ˆ๋‹ค. ์˜คํ”ˆ ์†Œ์Šค LLaMA-13B ๋ชจ๋ธ์€ ๋Œ€๋ถ€๋ถ„์˜ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋…์  GPT-3(175B) ๋ชจ๋ธ์„ ๋Šฅ๊ฐ€ํ•˜๋ฏ€๋กœ LLM ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ์ข‹์€ ๊ธฐ์ค€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The first set of LLaMA models [32] was released in February 2023, ranging from 7B to 65B parameters. These models are pre-trained on trillions of tokens, collected from publicly available datasets. LLaMA uses the transformer architecture of GPT-3, with a few minor architectural modifications, including (1) using a SwiGLU activation function instead of ReLU, (2) using rotary positional embeddings instead of absolute positional embedding, and (3) using root-mean-squared layer-normalization instead of standard layer-normalization. The open-source LLaMA-13B model outperforms the proprietary GPT-3 (175B) model on most benchmarks, making it a good baseline for LLM research. [|assistant|]
๊ทธ๋ฆผ 35: ์ˆœ์ฐจ์  bigram ๋ฐ์ดํ„ฐ: ์ƒ์œ„ p = 0.95๋Š” ๊ผฌ๋ฆฌ ์ขํžˆ๊ธฐ์™€ ์œ ์‚ฌํ•œ ํšจ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค. ์‹œํ€€์Šค ๊ธธ์ด๊ฐ€ 100์ธ 1000๊ฐœ์˜ ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 35: ์ˆœ์ฐจ์  ๋ฐ”์ด๊ทธ๋žจ ๋ฐ์ดํ„ฐ: ์ƒ์œ„ p = 0.95๋Š” ๊ผฌ๋ฆฌ ์ขํžˆ๊ธฐ์™€ ์œ ์‚ฌํ•œ ํšจ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค. ์‹œํ€€์Šค ๊ธธ์ด๊ฐ€ 100์ธ 1000๊ฐœ์˜ ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 35: Sequential bigram data: top p = 0.95 leads to similar effect as tail narrowing. 1000 data with sequence length 100. [|assistant|]
**Memory update.** ๊ฒ€์ƒ‰์ด ์™„๋ฃŒ๋˜๋ฉด ๋ฉ”๋ชจ๋ฆฌ์™€ ์ •๊ทœํ™” ํ•ญ์„ ์ƒˆ KV ํ•ญ๋ชฉ์œผ๋กœ ์—…๋ฐ์ดํŠธํ•˜๊ณ  ๋‹ค์Œ ์ƒํƒœ๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์–ป์Šต๋‹ˆ๋‹ค.[|endofturn|]
**๋ฉ”๋ชจ๋ฆฌ ์—…๋ฐ์ดํŠธ**. ๊ฒ€์ƒ‰์ด ์™„๋ฃŒ๋˜๋ฉด ๋ฉ”๋ชจ๋ฆฌ์™€ ์ •๊ทœํ™” ํ•ญ์„ ์ƒˆ KV ํ•ญ๋ชฉ์œผ๋กœ ์—…๋ฐ์ดํŠธํ•˜๊ณ  ๋‹ค์Œ ์ƒํƒœ๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์–ป์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Memory update.** Once the retrieval is done, we update the memory and the normalization term with the new KV entries and obtain the next states as [|assistant|]
์ง€์‹์€ ์ธ๊ฐ„ ์ง€๋Šฅ์˜ ํ•ต์‹ฌ ์š”์†Œ ์ค‘ ํ•˜๋‚˜์ด๋ฉฐ, ์˜ค๋žœ ์—ญ์‚ฌ๋ฅผ ํ†ตํ•ด ์ถ•์ ๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. GPT-4์™€ ๊ฐ™์€ Large Language Models (LLMs)๋Š” ์ •๊ตํ•œ ๋…ผ๋ฆฌ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋›ฐ์–ด๋‚œ ์ง€์‹ ๊ธฐ๋ฐ˜์œผ๋กœ ์ธ์ •๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. GPT-4์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ 1T ์ด์ƒ์ด๋ผ๋Š” ์†Œ๋ฌธ์ด ์žˆ์ง€๋งŒ, _์ธ๊ฐ„์˜ ๋ชจ๋“  ์ง€์‹์„ ์ €์žฅํ•ด์•ผ ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹Œ๊ฐ€์š”?_ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋กœ ์ถฉ๋ถ„ํžˆ ํ•™์Šต๋œ 10B ๋ชจ๋ธ์ด GPT-4์˜ ์ง€์‹ ๋Šฅ๋ ฅ์— ํ•„์ ํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ์ €ํฌ ๋…ผ๋ฌธ์€ ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ง€์‹์€ ์ธ๊ฐ„ ์ง€๋Šฅ์˜ ํ•ต์‹ฌ ์š”์†Œ ์ค‘ ํ•˜๋‚˜์ด๋ฉฐ, ์˜ค๋žœ ์—ญ์‚ฌ๋ฅผ ํ†ตํ•ด ์ถ•์ ๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. GPT-4์™€ ๊ฐ™์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ์ •๊ตํ•œ ๋…ผ๋ฆฌ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋›ฐ์–ด๋‚œ ์ง€์‹ ๊ธฐ๋ฐ˜์œผ๋กœ ์ธ์ •๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. GPT-4์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ 1T ์ด์ƒ์ด๋ผ๋Š” ์†Œ๋ฌธ์ด ์žˆ์ง€๋งŒ, _์ธ๊ฐ„์˜ ๋ชจ๋“  ์ง€์‹์„ ์ €์žฅํ•ด์•ผ ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹Œ๊ฐ€์š”?_ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋กœ ์ถฉ๋ถ„ํžˆ ํ•™์Šต๋œ 100์–ต ๊ฐœ ๋ชจ๋ธ์ด GPT-4์˜ ์ง€์‹ ๋Šฅ๋ ฅ์— ํ•„์ ํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ์ €ํฌ ๋…ผ๋ฌธ์€ ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Knowledge is a, if not the, pivotal component of human intelligence, accumulated over our extensive history. Large language models like GPT-4 are celebrated not just for their sophisticated logic but also for their superior knowledge base. Despite rumors of GPT-4 having over 1T parameters, _is it necessary to store all human knowledge?_ Could a 10B model, if trained sufficiently with high-quality data, match GPT-4's knowledge capacity? Our paper seeks to address these questions. [|assistant|]
๊ณ ํ’ˆ์งˆ์˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™”๋ฅผ ๋Œ€๊ทœ๋ชจ๋กœ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์„ธ ๊ฐ€์ง€ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋œ ์ƒˆ๋กœ์šด multi-agent ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค:[|endofturn|]
๊ณ ํ’ˆ์งˆ์˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™”๋ฅผ ๋Œ€๊ทœ๋ชจ๋กœ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์„ธ ๊ฐ€์ง€ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋œ ์ƒˆ๋กœ์šด ๋‹ค์ค‘ ์—์ด์ „ํŠธ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค:[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In order to produce high-quality simulated dialogues at scale, we developed a novel multi-agent framework which comprised three key components: [|assistant|]
**RQ1. LLM์ด 2-hop ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋™์•ˆ ์ฒซ ๋ฒˆ์งธ hop ์ถ”๋ก ์„ ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์ˆ˜ํ–‰ํ•˜๋‚˜์š”?** ์ฒซ ๋ฒˆ์งธ hop ์ถ”๋ก ์€ LLM์ด ์„ค๋ช…์  ์–ธ๊ธ‰์— ๋Œ€ํ•œ bridge entity๋ฅผ ๊ธฐ์–ตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ”„๋กฌํ”„ํŠธ ๋‚ด์—์„œ LLM์ด bridge entity์˜ ์„ค๋ช…์  ์–ธ๊ธ‰์„ ์ ‘ํ•  ๋•Œ LLM์˜ ๋‚ด๋ถ€์—์„œ bridge entity์˜ ํ˜ธ์ถœ ๋นˆ๋„๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ”„๋กฌํ”„ํŠธ๋ฅผ "'Thriller'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š”"์—์„œ "'Superstition'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š”"์œผ๋กœ ๋ณ€๊ฒฝํ•˜๋ฉด LLM์˜ ๋‚ด๋ถ€์—์„œ Stevie Wonder์— ๋Œ€ํ•œ ํ˜ธ์ถœ์ด ์ฆ๊ฐ€ํ•˜๋Š”์ง€ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**RQ1. LLM์ด 2ํ™‰ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋™์•ˆ ์ฒซ ๋ฒˆ์งธ ํ™‰ ์ถ”๋ก ์„ ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์ˆ˜ํ–‰ํ•˜๋‚˜์š”?** ์ฒซ ๋ฒˆ์งธ ํ™‰ ์ถ”๋ก ์€ LLM์ด ์„ค๋ช…์  ์–ธ๊ธ‰์— ๋Œ€ํ•œ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ๋ฅผ ๊ธฐ์–ตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ”„๋กฌํ”„ํŠธ ๋‚ด์—์„œ LLM์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์„ค๋ช…์  ์–ธ๊ธ‰์„ ์ ‘ํ•  ๋•Œ LLM์˜ ๋‚ด๋ถ€์—์„œ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ํ˜ธ์ถœ ๋นˆ๋„๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ”„๋กฌํ”„ํŠธ๋ฅผ "'์Šค๋ฆด๋Ÿฌ'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š”"์—์„œ "'๊ณผ์‹œ์ฆ'์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ๋Š”"์œผ๋กœ ๋ณ€๊ฒฝํ•˜๋ฉด LLM์˜ ๋‚ด๋ถ€์—์„œ ์Šคํ‹ฐ๋น„ ์›๋”์— ๋Œ€ํ•œ ํ˜ธ์ถœ์ด ์ฆ๊ฐ€ํ•˜๋Š”์ง€ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**RQ1. How often does an LLM perform the first hop of reasoning while processing two-hop prompts?** We view the first-hop reasoning as the LLM's recall of the bridge entity for its descriptive mention. Therefore, we examine the frequency with which the LLM's internal recall of the bridge entity increases when it encounters a descriptive mention of the bridge entity within a prompt. For instance, we investigate whether altering the prompt from "The mother of the singer of _'Thriller'_ is" to "The mother of the singer of _'Superstition'_ is" increases the LLM's internal recall of Stevie Wonder. [|assistant|]
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ์ฃผ๋กœ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ์ˆ˜์‹ญ์—์„œ ์ˆ˜๋ฐฑ์–ต ๊ฐœ์— ๋‹ฌํ•˜๋Š” Transformer ๊ธฐ๋ฐ˜ ์‹ ๊ฒฝ ์–ธ์–ด ๋ชจ๋ธ์„ ๋งํ•˜๋ฉฐ, ํ‘œ III์— ์š”์•ฝ๋œ ๊ฒƒ์ฒ˜๋Ÿผ PaLM [31], LLaMA [32], GPT-4 [33]์™€ ๊ฐ™์ด ๋ฐฉ๋Œ€ํ•œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์นœ ๋ชจ๋ธ์„ ๋งํ•ฉ๋‹ˆ๋‹ค. LLM์€ PLM์— ๋น„ํ•ด ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ํ›จ์”ฌ ํฌ๊ธฐ๋Š” ๋ฌผ๋ก  ์–ธ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ, ๋” ์ค‘์š”ํ•œ ๊ฒƒ์€ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์—๋Š” ์—†๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์—๋Š” (1) ์ถ”๋ก  ์‹œ๊ฐ„์— ํ”„๋กฌํ”„ํŠธ์— ์ œ์‹œ๋œ ์†Œ์ˆ˜์˜ ์˜ˆ์ œ๋ฅผ ํ†ตํ•ด ์ƒˆ๋กœ์šด ์ž‘์—…์„ ํ•™์Šตํ•˜๋Š” In-Context Learning, (2) ๋ช…์‹œ์ ์ธ ์˜ˆ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ณ ๋„ Instruction Tuning์„ ๊ฑฐ์นœ ํ›„ ์ƒˆ๋กœ์šด ์œ ํ˜•์˜ ์ž‘์—…์— ๋Œ€ํ•œ ๋ช…๋ น์„ ๋”ฐ๋ฅผ ์ˆ˜ ์žˆ๋Š” Instruction Following, (3) ์—ฐ์‡„ ์‚ฌ๊ณ  ํ”„๋กฌํ”„ํŠธ [34]์—์„œ์ฒ˜๋Ÿผ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„๋กœ ๋ถ„ํ•ดํ•˜์—ฌ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ค๋‹จ๊ณ„ ์ถ”๋ก  ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. LLM์€ ์™ธ๋ถ€ ์ง€์‹๊ณผ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฆ๊ฐ•ํ•  ์ˆ˜๋„ ์žˆ์œผ๋ฏ€๋กœ ์‚ฌ์šฉ์ž ๋ฐ ํ™˜๊ฒฝ๊ณผ ํšจ๊ณผ์ ์œผ๋กœ ์ƒํ˜ธ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ [35, 36] ์ƒํ˜ธ ์ž‘์šฉ์„ ํ†ตํ•ด ์ˆ˜์ง‘ํ•œ ํ”ผ๋“œ๋ฐฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ (์˜ˆ: Reinforcement Learning from Human Feedback (RLHF)์„ ์ด์šฉํ•œ ๊ฐ•ํ™” ํ•™์Šต) ์ง€์†์ ์œผ๋กœ ์ž์‹ ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ์ฃผ๋กœ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ์ˆ˜์‹ญ์—์„œ ์ˆ˜๋ฐฑ์–ต ๊ฐœ์— ๋‹ฌํ•˜๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜ ์‹ ๊ฒฝ ์–ธ์–ด ๋ชจ๋ธ์„ ๋งํ•˜๋ฉฐ, ํ‘œ III์— ์š”์•ฝ๋œ ๊ฒƒ์ฒ˜๋Ÿผ PaLM [31], LLaMA [32], GPT-4 [33]์™€ ๊ฐ™์ด ๋ฐฉ๋Œ€ํ•œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์นœ ๋ชจ๋ธ์„ ๋งํ•ฉ๋‹ˆ๋‹ค. LLM์€ PLM์— ๋น„ํ•ด ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ํ›จ์”ฌ ํฌ๊ธฐ๋Š” ๋ฌผ๋ก  ์–ธ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ, ๋” ์ค‘์š”ํ•œ ๊ฒƒ์€ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์—๋Š” ์—†๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์—๋Š” (1) ์ถ”๋ก  ์‹œ๊ฐ„์— ํ”„๋กฌํ”„ํŠธ์— ์ œ์‹œ๋œ ์†Œ์ˆ˜์˜ ์˜ˆ์ œ๋ฅผ ํ†ตํ•ด ์ƒˆ๋กœ์šด ์ž‘์—…์„ ํ•™์Šตํ•˜๋Š” ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต, (2) ๋ช…์‹œ์ ์ธ ์˜ˆ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ณ ๋„ ๋ช…๋ น์–ด ํŠœ๋‹์„ ๊ฑฐ์นœ ํ›„ ์ƒˆ๋กœ์šด ์œ ํ˜•์˜ ์ž‘์—…์— ๋Œ€ํ•œ ๋ช…๋ น์„ ๋”ฐ๋ฅผ ์ˆ˜ ์žˆ๋Š” ๋ช…๋ น์–ด ์ˆ˜ํ–‰, (3) ์—ฐ์‡„ ์‚ฌ๊ณ  ํ”„๋กฌํ”„ํŠธ [34]์—์„œ์ฒ˜๋Ÿผ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„๋กœ ๋ถ„ํ•ดํ•˜์—ฌ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ค๋‹จ๊ณ„ ์ถ”๋ก  ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. LLM์€ ์™ธ๋ถ€ ์ง€์‹๊ณผ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฆ๊ฐ•ํ•  ์ˆ˜๋„ ์žˆ์œผ๋ฏ€๋กœ ์‚ฌ์šฉ์ž ๋ฐ ํ™˜๊ฒฝ๊ณผ ํšจ๊ณผ์ ์œผ๋กœ ์ƒํ˜ธ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ [35, 36] ์ƒํ˜ธ ์ž‘์šฉ์„ ํ†ตํ•ด ์ˆ˜์ง‘ํ•œ ํ”ผ๋“œ๋ฐฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ (์˜ˆ: ์‚ฌ๋žŒ๊ณผ์˜ ํ”ผ๋“œ๋ฐฑ(RLHF)์„ ์ด์šฉํ•œ ๊ฐ•ํ™” ํ•™์Šต) ์ง€์†์ ์œผ๋กœ ์ž์‹ ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Large language models (LLMs) mainly refer to transformer-based neural language models 1 that contain tens to hundreds of billions of parameters, which are pre-trained on massive text data, such as PaLM [31], LLaMA [32], and GPT-4 [33], as summarized in Table III. Comparedto PLMs, LLMs are not only much larger in model size, but also exhibit stronger language understanding and generation abilities, and more importantly, emergent abilities that are not present in smaller-scale language models. As illustrated in Fig. 1, these emergent abilities include (1) in-context learning, where LLMs learn a new task from a small set of examples presented in the prompt at inference time, (2) instruction following, where LLMs, after instruction tuning, can follow the instructions for new types of tasks without using explicit examples, and (3) multi-step reasoning, where LLMs can solve a complex task by breaking down that task into intermediate reasoning steps as demonstrated in the chain-of-thought prompt [34]. LLMs can also be augmented by using external knowledge and tools [35, 36] so that they can effectively interact with users and environment [37], and continually improve itself using feedback data collected through interactions (e.g. via reinforcement learning with human feedback (RLHF)). [|assistant|]
๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, EHR ๋…ธํŠธ ์š”์•ฝ ์ž‘์—…์˜ ๊ฒฝ์šฐ, AMIE๋Š” ์ž„์ƒ ๋…ธํŠธ๋ฅผ ์ œ๊ณต๋ฐ›์•„ ๋…ธํŠธ์˜ ์š”์•ฝ์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. Medical reasoning/QA ๋ฐ long-form response generation ์ž‘์—…์€ [13]์—์„œ์™€ ๋™์ผํ•œ ์„ค์ •์„ ๋”ฐ๋ž์Šต๋‹ˆ๋‹ค. ํŠนํžˆ dialogue generation ๋ฐ long-form response generation์„ ์ œ์™ธํ•œ ๋ชจ๋“  ์ž‘์—…์—๋Š” ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๋ฅผ ์œ„ํ•œ ์ž‘์—…๋ณ„ Instructions ์™ธ์— few-shot (1~5) ์˜ˆ์‹œ๊ฐ€ ํฌํ•จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, EHR ๋…ธํŠธ ์š”์•ฝ ์ž‘์—…์˜ ๊ฒฝ์šฐ, AMIE๋Š” ์ž„์ƒ ๋…ธํŠธ๋ฅผ ์ œ๊ณต๋ฐ›์•„ ๋…ธํŠธ์˜ ์š”์•ฝ์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ์˜๋ฃŒ ์ถ”๋ก /QA ๋ฐ ์žฅ๋ฌธ ์‘๋‹ต ์ƒ์„ฑ ์ž‘์—…์€ [13]์—์„œ์™€ ๋™์ผํ•œ ์„ค์ •์„ ๋”ฐ๋ž์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ๋Œ€ํ™” ์ƒ์„ฑ ๋ฐ ์žฅ๋ฌธ ์‘๋‹ต ์ƒ์„ฑ์„ ์ œ์™ธํ•œ ๋ชจ๋“  ์ž‘์—…์—๋Š” ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๋ฅผ ์œ„ํ•œ ์ž‘์—…๋ณ„ ์ง€์นจ ์™ธ์— ๋ช‡ ๊ฐœ์˜ ์ƒท(1~5)์˜ ์˜ˆ์‹œ๊ฐ€ ํฌํ•จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Similarly, for the EHR note summarization task, AMIE was provided with a clinical note and prompted to generate a summary of the note. Medical reasoning/QA and long-form response generation tasks followed the same setup as in [13]. Notably, all tasks except dialogue generation and long-form response generation incorporated few-shot (1-5) exemplars in addition to task-specific instructions for additional context. [|assistant|]
๊ทธ๋ฆผ 15: ์ฝ”์‚ฌ์ธ LRS์™€ WSD LRS๋ฅผ ๋น„๊ตํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 16: ๋‘ ๊ฐ€์ง€ ๋‹ค๋ฅธ ํ•จ์ˆ˜ ํ˜•ํƒœ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ WSD LRS๊ฐ€ ๋‹ฌ์„ฑํ•œ ๋ฐ์ดํ„ฐ scaling law์— ๋งž์ถคํ™”ํ•˜๊ณ  power law๋ฅผ ์ตœ์ ์˜ ์ ํ•ฉ๋„๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 15: ์ฝ”์‚ฌ์ธ LRS์™€ WSD LRS๋ฅผ ๋น„๊ตํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 16: ๋‘ ๊ฐ€์ง€ ๋‹ค๋ฅธ ํ•จ์ˆ˜ ํ˜•ํƒœ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ WSD LRS๊ฐ€ ๋‹ฌ์„ฑํ•œ ๋ฐ์ดํ„ฐ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋งž์ถคํ™”ํ•˜๊ณ  ํŒŒ์›Œ ๋กœ๋ฅผ ์ตœ์ ์˜ ์ ํ•ฉ๋„๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 15: Illustrative comparison be- Figure 16: We use two different function tween Cosine LRS and WSD LRS. The forms to fit the data scaling law achieved WSD LRS with different end steps share by WSD LRS and choose power law as the same stable training stage. the best fit. [|assistant|]
Wei ์™ธ(2022)๋Š” ๋จผ์ € ์ž์—ฐ์–ด ํ˜•์‹์˜ ๊ทผ๊ฑฐ๋ฅผ ๋ฐ๋ชจ์— ์ˆ˜๋™์œผ๋กœ ์ œ๊ณตํ•˜์—ฌ chain-of-thought prompting (Few-shot CoT)์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก  ๊ณผ์ •์˜ ํ™•์‹ค์„ฑ์„ ๋ณด์žฅํ•˜๊ณ  ์ถ”๋ก  ๊ฒฝ๋กœ์™€ ๋‹ต๋ณ€ ๊ฐ„์˜ ๋ถˆ์ผ์น˜๋ฅผ ์ค„์ด๊ธฐ ์œ„ํ•ด PAL Gao ์™ธ(2023), PoT Chen ์™ธ(2022), NLEP Zhang ์™ธ(2023)๋Š” ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด๋ฅผ ์ฃผ์„ ์ฒ˜๋ฆฌ๋œ ๊ทผ๊ฑฐ๋กœ ํ™œ์šฉํ•˜์—ฌ ๋ฌธ์ œ ํ•ด๊ฒฐ์„ ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ Python ํ”„๋กœ๊ทธ๋žจ์œผ๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, ์ž์—ฐ์–ด์™€ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์˜ ์žฅ์ ์„ ๋ชจ๋‘ ์‚ด๋ฆฌ๊ณ  ์ถ”๋ก  ๊ฒฐ๊ณผ์˜ ์‹ ๋ขฐ๋„๋ฅผ ๋†’์ด๊ธฐ ์œ„ํ•ด MathPromter Imani ์™ธ(2023)๋Š” zero-shot chain-of-thought prompting์„ ์‚ฌ์šฉํ•˜์—ฌ ์—ฌ๋Ÿฌ ๋Œ€์ˆ˜์‹ ๋˜๋Š” Python ํ•จ์ˆ˜๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ์„œ๋กœ ๊ฒ€์ฆํ•˜๊ณ  ๊ฒฐ๊ณผ์˜ ์‹ ๋ขฐ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ถ”๋ก  ๋‹จ๊ณ„๊ฐ€ ๋” ๋งŽ์€ ์ฒด์ธ๊ณผ ๊ฐ™์ด ๋ฐ๋ชจ์—์„œ ์ƒ˜ํ”Œ์˜ ์ถ”๋ก  ๋ณต์žก์„ฑ์ด ๋†’์„์ˆ˜๋ก ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋ฏ€๋กœ Fu ์™ธ(2023)๋Š” complexity-based prompting์„ ์ œ์•ˆํ•˜์—ฌ ๋ณต์žก๋„๊ฐ€ ๋†’์€ ๊ทผ๊ฑฐ๋“ค ๊ฐ„์˜ ํˆฌํ‘œ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ตœ์ข… ๋‹ต์„ ์–ป์Šต๋‹ˆ๋‹ค.[|endofturn|]
Wei ์™ธ(2022)๋Š” ๋จผ์ € ์ž์—ฐ์–ด ํ˜•์‹์˜ ๊ทผ๊ฑฐ๋ฅผ ๋ฐ๋ชจ์— ์ˆ˜๋™์œผ๋กœ ์ œ๊ณตํ•˜์—ฌ ์—ฐ์‡„ ์‚ฌ๊ณ  ํ”„๋กฌํ”„ํŠธ(Few-shot CoT)๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก  ๊ณผ์ •์˜ ํ™•์‹ค์„ฑ์„ ๋ณด์žฅํ•˜๊ณ  ์ถ”๋ก  ๊ฒฝ๋กœ์™€ ๋‹ต๋ณ€ ๊ฐ„์˜ ๋ถˆ์ผ์น˜๋ฅผ ์ค„์ด๊ธฐ ์œ„ํ•ด PAL Gao ์™ธ(2023), PoT Chen ์™ธ(2022), NLEP Zhang ์™ธ(2023)๋Š” ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด๋ฅผ ์ฃผ์„ ์ฒ˜๋ฆฌ๋œ ๊ทผ๊ฑฐ๋กœ ํ™œ์šฉํ•˜์—ฌ ๋ฌธ์ œ ํ•ด๊ฒฐ์„ ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ Python ํ”„๋กœ๊ทธ๋žจ์œผ๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, ์ž์—ฐ์–ด์™€ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์˜ ์žฅ์ ์„ ๋ชจ๋‘ ์‚ด๋ฆฌ๊ณ  ์ถ”๋ก  ๊ฒฐ๊ณผ์˜ ์‹ ๋ขฐ๋„๋ฅผ ๋†’์ด๊ธฐ ์œ„ํ•ด MathPromter Imani ์™ธ(2023)๋Š” ์ œ๋กœ ์ƒท ์—ฐ์‡„ ์‚ฌ๊ณ  ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์—ฌ๋Ÿฌ ๋Œ€์ˆ˜์‹ ๋˜๋Š” Python ํ•จ์ˆ˜๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ์„œ๋กœ ๊ฒ€์ฆํ•˜๊ณ  ๊ฒฐ๊ณผ์˜ ์‹ ๋ขฐ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ถ”๋ก  ๋‹จ๊ณ„๊ฐ€ ๋” ๋งŽ์€ ์ฒด์ธ๊ณผ ๊ฐ™์ด ๋ฐ๋ชจ์—์„œ ์ƒ˜ํ”Œ์˜ ์ถ”๋ก  ๋ณต์žก์„ฑ์ด ๋†’์„์ˆ˜๋ก ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋ฏ€๋กœ Fu ์™ธ(2023)๋Š” ๋ณต์žก๋„ ๊ธฐ๋ฐ˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ œ์•ˆํ•˜์—ฌ ๋ณต์žก๋„๊ฐ€ ๋†’์€ ๊ทผ๊ฑฐ๋“ค ๊ฐ„์˜ ํˆฌํ‘œ๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ตœ์ข… ๋‹ต์„ ์–ป์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Wei et al. (2022) first propose chain-of-thought prompting (Few-shot CoT) by manually providing natural language form rationales to the demonstrations. To further ensure certainty in the reasoning process and reduce inconsistencies between reasoning path and answers, PAL Gao et al. (2023), PoT Chen et al. (2022) and NLEP Zhang et al. (2023) leverage programming language as annotated rationales, which transforms the problem-solving into an executable Python program. Meanwhile, to take both advantages of natural language and programming language and raise the confidence of reasoning output, MathPromter Imani et al. (2023) uses the zero-shot chain-of-thought prompting to generate multiple algebraic expressions or Python functions, which can verify each other and improve the reliability of results. Furthermore, since the reasoning complexity of samples in demonstrations, such as chains with more reasoning steps, results in performance improvement Fu et al. (2023) proposes complexity-based prompting, where voting among high-complexity rationales is performed to get the final answer. [|assistant|]
์šฐ๋ฆฌ๋Š” ๋‹ค์–‘ํ•œ downstream ์ž‘์—…์—์„œ ์„ฑ๋Šฅ์„ ์ธก์ •ํ•˜๋Š”๋ฐ, ์ด๋ฅผ ํฌ๊ฒŒ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ๋‹ค.[|endofturn|]
์šฐ๋ฆฌ๋Š” ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ์„ ์ธก์ •ํ•˜๋Š”๋ฐ, ์ด๋ฅผ ํฌ๊ฒŒ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We measure performance on a wide variety of downstream tasks, which can be broadly categorized as follows. [|assistant|]
์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.[|endofturn|]
์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.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]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. [|assistant|]
์—์ด์ „ํŠธ๋ฒค์น˜์™€ ๊ฐ™์€ ์ผ๋ถ€ ์—์ด์ „ํŠธ๋ณ„ ๋ฒค์น˜๋งˆํฌ๋Š” ์›น ๋ธŒ๋ผ์šฐ์ง•, ๋ช…๋ น์ค„ ์ธํ„ฐํŽ˜์ด์Šค, ๋น„๋””์˜ค ๊ฒŒ์ž„ ๋“ฑ ๋‹ค์–‘ํ•œ ํ™˜๊ฒฝ์—์„œ ์–ธ์–ด ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค[17]. ์ด๋ฅผ ํ†ตํ•ด ์—์ด์ „ํŠธ๊ฐ€ ์ฃผ์–ด์ง„ ์ž‘์—…์„ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๋ก , ๊ณ„ํš, ๋„๊ตฌ ํ˜ธ์ถœ์„ ํ†ตํ•ด ์ƒˆ๋กœ์šด ํ™˜๊ฒฝ์— ์–ผ๋งˆ๋‚˜ ์ž˜ ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋ฅผ ๋” ์ž˜ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Benchmarks์ธ AgentBench ๋ฐ SmartPlay์™€ ๊ฐ™์€ ๋ฒค์น˜๋งˆํฌ๋Š” ๊ตฌํ˜„์˜ ์„ฑ๊ณต๋ฅ , ์‚ฌ๋žŒ์˜ ์‘๋‹ต๊ณผ์˜ ์ถœ๋ ฅ ์œ ์‚ฌ์„ฑ, ์ „๋ฐ˜์ ์ธ ํšจ์œจ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋œ ๊ฐ๊ด€์ ์ธ ํ‰๊ฐ€ ์ง€ํ‘œ๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค[17; 30]. ์ด๋Ÿฌํ•œ ๊ฐ๊ด€์ ์ธ ์ง€ํ‘œ๋Š” ๊ตฌํ˜„์˜ ์ „๋ฐ˜์ ์ธ ์‹ ๋ขฐ์„ฑ๊ณผ ์ •ํ™•์„ฑ์„ ์ดํ•ดํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•˜์ง€๋งŒ, ๋ณด๋‹ค ๋ฏธ๋ฌ˜ํ•˜๊ฑฐ๋‚˜ ์ฃผ๊ด€์ ์ธ ์„ฑ๋Šฅ ์ธก์ •๋„ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋„๊ตฌ ์‚ฌ์šฉ ํšจ์œจ์„ฑ, ์‹ ๋ขฐ์„ฑ, ๊ณ„ํš์˜ Robustness์™€ ๊ฐ™์€ ์ง€ํ‘œ๋Š” ์„ฑ๊ณต๋ฅ ๋งŒํผ์ด๋‚˜ ์ค‘์š”ํ•˜์ง€๋งŒ ์ธก์ •ํ•˜๊ธฐ๊ฐ€ ํ›จ์”ฌ ๋” ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง€ํ‘œ ์ค‘ ์ƒ๋‹น์ˆ˜๋Š” ์ „๋ฌธ๊ฐ€์˜ ํ‰๊ฐ€๊ฐ€ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— LLM์œผ๋กœ ์‹ฌ์‚ฌํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด ๋น„์šฉ๊ณผ ์‹œ๊ฐ„์ด ๋งŽ์ด ์†Œ์š”๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์—์ด์ „ํŠธ๋ฒค์น˜์™€ ๊ฐ™์€ ์ผ๋ถ€ ์—์ด์ „ํŠธ๋ณ„ ๋ฒค์น˜๋งˆํฌ๋Š” ์›น ๋ธŒ๋ผ์šฐ์ง•, ๋ช…๋ น์ค„ ์ธํ„ฐํŽ˜์ด์Šค, ๋น„๋””์˜ค ๊ฒŒ์ž„ ๋“ฑ ๋‹ค์–‘ํ•œ ํ™˜๊ฒฝ์—์„œ ์–ธ์–ด ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค[17]. ์ด๋ฅผ ํ†ตํ•ด ์—์ด์ „ํŠธ๊ฐ€ ์ฃผ์–ด์ง„ ์ž‘์—…์„ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๋ก , ๊ณ„ํš, ๋„๊ตฌ ํ˜ธ์ถœ์„ ํ†ตํ•ด ์ƒˆ๋กœ์šด ํ™˜๊ฒฝ์— ์–ผ๋งˆ๋‚˜ ์ž˜ ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋ฅผ ๋” ์ž˜ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ๋ฒค์น˜ ๋ฐ ์Šค๋งˆํŠธํ”Œ๋ ˆ์ด์™€ ๊ฐ™์€ ๋ฒค์น˜๋งˆํฌ๋Š” ๊ตฌํ˜„์˜ ์„ฑ๊ณต๋ฅ , ์‚ฌ๋žŒ์˜ ์‘๋‹ต๊ณผ์˜ ์ถœ๋ ฅ ์œ ์‚ฌ์„ฑ, ์ „๋ฐ˜์ ์ธ ํšจ์œจ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋œ ๊ฐ๊ด€์ ์ธ ํ‰๊ฐ€ ์ง€ํ‘œ๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค[17; 30]. ์ด๋Ÿฌํ•œ ๊ฐ๊ด€์ ์ธ ์ง€ํ‘œ๋Š” ๊ตฌํ˜„์˜ ์ „๋ฐ˜์ ์ธ ์‹ ๋ขฐ์„ฑ๊ณผ ์ •ํ™•์„ฑ์„ ์ดํ•ดํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•˜์ง€๋งŒ, ๋ณด๋‹ค ๋ฏธ๋ฌ˜ํ•˜๊ฑฐ๋‚˜ ์ฃผ๊ด€์ ์ธ ์„ฑ๋Šฅ ์ธก์ •๋„ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋„๊ตฌ ์‚ฌ์šฉ ํšจ์œจ์„ฑ, ์‹ ๋ขฐ์„ฑ, ๊ณ„ํš์˜ ๊ฒฌ๊ณ ์„ฑ๊ณผ ๊ฐ™์€ ์ง€ํ‘œ๋Š” ์„ฑ๊ณต๋ฅ ๋งŒํผ์ด๋‚˜ ์ค‘์š”ํ•˜์ง€๋งŒ ์ธก์ •ํ•˜๊ธฐ๊ฐ€ ํ›จ์”ฌ ๋” ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง€ํ‘œ ์ค‘ ์ƒ๋‹น์ˆ˜๋Š” ์ „๋ฌธ๊ฐ€์˜ ํ‰๊ฐ€๊ฐ€ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— LLM์œผ๋กœ ์‹ฌ์‚ฌํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด ๋น„์šฉ๊ณผ ์‹œ๊ฐ„์ด ๋งŽ์ด ์†Œ์š”๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Some agent specific benchmarks like AgentBench evaluate language model-based agents in a variety of different environments such as web browsing, command-line interfaces, and video games [17]. This provides a better indication for how well agents can generalize to new environments, by reasoning, planning, and calling tools to achieve a given task. Benchmarks like AgentBench and SmartPlay introduce objective evaluation metrics designed to evaluate the implementation's success rate, output similarity to human responses, and overall efficiency [17; 30]. While these objective metrics are important to understanding the overall reliability and accuracy of the implementation, it is also important to consider more nuanced or subjective measures of performance. Metrics such as efficiency of tool use, reliability, and robustness of planning are nearly as important as success rate but are much more difficult to measure. Many of these metrics require evaluation by a human expert, which can be costly and time consuming compared to LLM-as-judge evaluations. [|assistant|]
๊ฐ ๋ชจ๋ธ ํฌ๊ธฐ์— ๋Œ€ํ•ด ์„ธ ๊ฐœ์˜ Fine-tuning ์‹œ๋“œ๋ฅผ ํ›ˆ๋ จํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ๊ฒฐ๊ตญ 12๊ฐœ์˜ ํ†ต๊ณ„ ํ…Œ์ŠคํŠธ๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 3์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Fine-tuning ์‹œ๋“œ ์ค‘ ์–ด๋Š ๊ฒƒ๋„ ํ†ต๊ณ„์ ์œผ๋กœ ์œ ์˜๋ฏธํ•œ \(p\)-๊ฐ’(\(p<0.05\))์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ฐ ๋ชจ๋ธ ํฌ๊ธฐ์— ๋Œ€ํ•ด ์„ธ ๊ฐœ์˜ ๋ฏธ์„ธ ์กฐ์ • ์‹œ๋“œ๋ฅผ ํ›ˆ๋ จํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ๊ฒฐ๊ตญ 12๊ฐœ์˜ ํ†ต๊ณ„ ํ…Œ์ŠคํŠธ๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 3์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฏธ์„ธ ์กฐ์ • ์‹œ๋“œ ์ค‘ ์–ด๋Š ๊ฒƒ๋„ ํ†ต๊ณ„์ ์œผ๋กœ ์œ ์˜๋ฏธํ•œ \(p\)-๊ฐ’(\(p<0.05\))์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Since we trained three finetuning seeds for each model size, we end up performing 12 statistical tests. The results can be found in Figure 3. We do not observe statistically significant \(p\)-values (\(p<0.05\)) for any of the finetuning seeds. [|assistant|]
์ด ์„น์…˜์—์„œ๋Š” ๊ณ ์ • ์ปดํ“จํŒ… ์„ค์ •์„ ๊ณ ๋ คํ•˜์—ฌ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์˜ ํฌ๊ธฐ(\(D_{source}\))๋ฅผ ๊ณต๋™์œผ๋กœ ์ฆ๊ฐ€์‹œํ‚ค๊ณ  \(R\)(์„ ํƒ๋œ \(D_{source}\)์˜ ๋น„์œจ)์„ ๊ฐ์†Œ์‹œ์ผœ ๋Œ€์ƒ ํ† ํฐ ์˜ˆ์‚ฐ์ด ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€๋˜๋„๋ก ๊ณ ์ •๋œ ํ† ํฐ ์˜ˆ์‚ฐ์œผ๋กœ ํ๋ ˆ์ด์…˜ํ•˜๊ณ  ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ Large Language Models (LLMs) ํ›ˆ๋ จ์˜ ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ํŒจ๋Ÿฌ๋‹ค์ž„๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. \(D_{source}\)๊ฐ€ ์ฆ๊ฐ€ํ•˜๊ณ  \(R\)์ด ๊ฐ์†Œํ•จ์— ๋”ฐ๋ผ ๋” ํฐ ์ดˆ๊ธฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์„ ํƒํ•˜๊ฒŒ ๋˜๊ณ , ์ด๋กœ ์ธํ•ด ์„ ํƒํ•  ์ˆ˜ ์žˆ๋Š” ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์˜ ์ง‘ํ•ฉ์ด ์ปค์ง€๊ณ  ์„ ํƒ๋œ ์ง‘ํ•ฉ์˜ ์ „๋ฐ˜์ ์ธ ํ’ˆ์งˆ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค. ๋ช…ํ™•์„ฑ์„ ์œ„ํ•ด \(D_{source}\)์™€ \(D_{target}\)์˜ ๋น„์œจ์˜ ํ•จ์ˆ˜๋กœ ์„ฑ๋Šฅ์„ ํ”Œ๋กฏํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์„ค์ •์— ๋Œ€ํ•ด baseline, SemDeDup ๋‹จ๋…, SSL Prototypes ๋‹จ๋… ๋ฐ ์ œ์•ˆ๋œ ๋ฐฉ๋ฒ• D4์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ด ์„น์…˜์—์„œ๋Š” ๊ณ ์ • ์ปดํ“จํŒ… ์„ค์ •์„ ๊ณ ๋ คํ•˜์—ฌ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์˜ ํฌ๊ธฐ(\(D_{source}\))๋ฅผ ๊ณต๋™์œผ๋กœ ์ฆ๊ฐ€์‹œํ‚ค๊ณ  \(R\)(์„ ํƒ๋œ \(D_{source}\)์˜ ๋น„์œจ)์„ ๊ฐ์†Œ์‹œ์ผœ ๋Œ€์ƒ ํ† ํฐ ์˜ˆ์‚ฐ์ด ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€๋˜๋„๋ก ๊ณ ์ •๋œ ํ† ํฐ ์˜ˆ์‚ฐ์œผ๋กœ ํ๋ ˆ์ด์…˜ํ•˜๊ณ  ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค์ •์€ LLM ํ›ˆ๋ จ์˜ ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ํŒจ๋Ÿฌ๋‹ค์ž„๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. D_{source}\)๊ฐ€ ์ฆ๊ฐ€ํ•˜๊ณ  \(R\)์ด ๊ฐ์†Œํ•จ์— ๋”ฐ๋ผ ๋” ํฐ ์ดˆ๊ธฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์„ ํƒํ•˜๊ฒŒ ๋˜๊ณ , ์ด๋กœ ์ธํ•ด ์„ ํƒํ•  ์ˆ˜ ์žˆ๋Š” ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์˜ ์ง‘ํ•ฉ์ด ์ปค์ง€๊ณ  ์„ ํƒ๋œ ์ง‘ํ•ฉ์˜ ์ „๋ฐ˜์ ์ธ ํ’ˆ์งˆ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค. ๋ช…ํ™•์„ฑ์„ ์œ„ํ•ด \(D_{source}\)์™€ \(D_{target}\)์˜ ๋น„์œจ์˜ ํ•จ์ˆ˜๋กœ ์„ฑ๋Šฅ์„ ํ”Œ๋กฏํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์„ค์ •์— ๋Œ€ํ•ด ๊ธฐ์ค€์„ , SemDeDup ๋‹จ๋…, SSL ํ”„๋กœํ† ํƒ€์ž… ๋‹จ๋… ๋ฐ ์ œ์•ˆ๋œ ๋ฐฉ๋ฒ• D4์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this section, we consider the fixed compute setting, where we curate and train on a fixed token budget by jointly increasing the size of the source dataset \(D_{source}\) and decreasing \(R\) (the fraction of the \(D_{source}\) which is selected), such that the target token budget remains constant. This setting is analogous to the most common paradigm for LLM training. As \(D_{source}\) grows and \(R\) decreases, we select from larger and larger initial datasets, resulting in a larger set of high-quality data points to select from and increasing the overall quality of the selected set. For clarity, we plot performance as a function of the ratio of the \(D_{source}\) to \(D_{target}\). For each setting, we evaluate the performance of a baseline, SemDeDup alone, SSL Prototypes alone, and our proposed method D4. [|assistant|]
์ •ํ™•ํ•œ ํ…์ŠคํŠธ ๋งค์นญ์„ ์œ„ํ•ด Lee ๋“ฑ(2022)์ด ๊ณต๊ฐœ์ ์œผ๋กœ ๊ณต๊ฐœํ•œ ExactSubstr ๊ตฌํ˜„์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ •ํ™•ํ•œ ํ•˜์œ„ ๋ฌธ์ž์—ด ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ๋ฏธ๋ฆฌ MinHash๋กœ ์ค‘๋ณต ์ œ๊ฑฐํ•œ ๋ฐ์ดํ„ฐ์— ์ ์šฉํ•˜์—ฌ ์ž‘์—…ํ•ด์•ผ ํ•˜๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํฌ๊ธฐ๋ฅผ ๊ฑฐ์˜ 40%๊นŒ์ง€ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ExactSubstr์€ ์—ฌ๋Ÿฌ ๋ฌธ์„œ์—์„œ ๋ฌธ์ž ํ•˜๋‚˜ํ•˜๋‚˜ ๊ทธ๋Œ€๋กœ ์กด์žฌํ•˜๋Š” ๊ธด ํ…์ŠคํŠธ ๋ฌธ์ž์—ด์„ ์ฐพ์Šต๋‹ˆ๋‹ค. ์ด ์ค‘ ์ผ๋ถ€๋Š” ์ด์ „ ๋‹จ๊ณ„์˜ approximate deduplication์„ ํ”ผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฌธ์„œ์—์„œ ์ฐจ์ง€ํ•˜๋Š” ๋น„์ค‘์ด ํฌ์ง€ ์•Š์„ ์ˆ˜๋„ ์žˆ๊ณ , ํ•œ ๋ฌธ์„œ์— ์—ฌ๋Ÿฌ ๋ฌธ์„œ์—์„œ ๊ฐ€์ ธ์˜จ ์„น์…˜์ด ๋ฐ˜๋ณต๋  ์ˆ˜๋„ ์žˆ์œผ๋ฉฐ, ๋‹จ์ˆœํžˆ MinHash์˜ approximate ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ๋ฐœ๊ฒฌ๋˜์ง€ ์•Š์•˜์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ •ํ™•ํ•œ ํ…์ŠคํŠธ ๋งค์นญ์„ ์œ„ํ•ด Lee ๋“ฑ(2022)์ด ๊ณต๊ฐœ์ ์œผ๋กœ ๊ณต๊ฐœํ•œ ExactSubstr ๊ตฌํ˜„์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ •ํ™•ํ•œ ํ•˜์œ„ ๋ฌธ์ž์—ด ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ๋ฏธ๋ฆฌ MinHash๋กœ ์ค‘๋ณต ์ œ๊ฑฐํ•œ ๋ฐ์ดํ„ฐ์— ์ ์šฉํ•˜์—ฌ ์ž‘์—…ํ•ด์•ผ ํ•˜๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํฌ๊ธฐ๋ฅผ ๊ฑฐ์˜ 40%๊นŒ์ง€ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ExactSubstr์€ ์—ฌ๋Ÿฌ ๋ฌธ์„œ์—์„œ ๋ฌธ์ž ํ•˜๋‚˜ํ•˜๋‚˜ ๊ทธ๋Œ€๋กœ ์กด์žฌํ•˜๋Š” ๊ธด ํ…์ŠคํŠธ ๋ฌธ์ž์—ด์„ ์ฐพ์Šต๋‹ˆ๋‹ค. ์ด ์ค‘ ์ผ๋ถ€๋Š” ์ด์ „ ๋‹จ๊ณ„์˜ ๋Œ€๋žต์ ์ธ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ํ”ผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฌธ์„œ์—์„œ ์ฐจ์ง€ํ•˜๋Š” ๋น„์ค‘์ด ํฌ์ง€ ์•Š์„ ์ˆ˜๋„ ์žˆ๊ณ , ํ•œ ๋ฌธ์„œ์— ์—ฌ๋Ÿฌ ๋ฌธ์„œ์—์„œ ๊ฐ€์ ธ์˜จ ์„น์…˜์ด ๋ฐ˜๋ณต๋  ์ˆ˜๋„ ์žˆ์œผ๋ฉฐ, ๋‹จ์ˆœํžˆ MinHash์˜ ๋Œ€๋žต์ ์ธ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ๋ฐœ๊ฒฌ๋˜์ง€ ์•Š์•˜์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We make use of the ExactSubstr implementation publicly released by Lee et al. (2022) for exact text matching. We apply exact substring deduplication to data that has already been deduplicated by MinHash, reducing by nearly 40% size of the dataset on which we have to operate. ExactSubstr will find long strings of text that are present, character for character, across multiple documents. Some of these may have escaped the earlier stage of approximate deduplication: they might not constitute a big enough portion of the document; one document might have repeated sections sourced across many different documents; or they may simply not have been found due to the approximate nature of MinHash. [|assistant|]
ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ํฌ๊ธฐ๊ฐ€ ๋‹ค๋ฅธ LLaMA ๋ชจ๋ธ(7B, 13B, 33B, 65B)์˜ ๋Œ€๋ถ€๋ถ„์˜ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋Š” ๋™์ผํ•œ ์ƒ์Šน ์ถ”์„ธ์— ์†ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์˜ pre-training Loss๊ฐ€ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ํ† ํฐ ์ˆ˜์— ๊ด€๊ณ„์—†์ด ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ์˜ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒฐ๋ก ์„ ๋”์šฑ ํ™•์ฆํ•ฉ๋‹ˆ๋‹ค. LLaMA-65B์˜ ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ์˜ˆ์™ธ๊ฐ€ ํ•˜๋‚˜๋งŒ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”. Training Loss๊ฐ€ 1.8๋ณด๋‹ค ๋†’์„ ๋•Œ LLaMA-65B๋Š” ๋™์ผํ•œ Training Loss๋ฅผ ๊ฐ€์ง„ ๋” ์ž‘์€ ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์—†์–ด์„œ ๊ฒฐ๊ณผ๋ฅผ ๋” ์ด์ƒ ๋ถ„์„ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด์ƒ๊ฐ’์€ ์ดˆ๊ธฐ 10%์˜ Training Tokens์œผ๋กœ๋งŒ ๊ตฌ์„ฑ๋œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”.[|endofturn|]
ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ํฌ๊ธฐ๊ฐ€ ๋‹ค๋ฅธ LLaMA ๋ชจ๋ธ(7B, 13B, 33B, 65B)์˜ ๋Œ€๋ถ€๋ถ„์˜ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋Š” ๋™์ผํ•œ ์ƒ์Šน ์ถ”์„ธ์— ์†ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์ด ๋ชจ๋ธ ํฌ๊ธฐ์™€ ํ† ํฐ ์ˆ˜์— ๊ด€๊ณ„์—†์ด ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ์˜ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒฐ๋ก ์„ ๋”์šฑ ํ™•์ฆํ•ฉ๋‹ˆ๋‹ค. LLaMA-65B์˜ ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ์˜ˆ์™ธ๊ฐ€ ํ•˜๋‚˜๋งŒ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”. ํ›ˆ๋ จ ์†์‹ค์ด 1.8๋ณด๋‹ค ๋†’์„ ๋•Œ LLaMA-65B๋Š” ๋™์ผํ•œ ํ›ˆ๋ จ ์†์‹ค์„ ๊ฐ€์ง„ ๋” ์ž‘์€ ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์—†์–ด์„œ ๊ฒฐ๊ณผ๋ฅผ ๋” ์ด์ƒ ๋ถ„์„ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด์ƒ๊ฐ’์€ ์ดˆ๊ธฐ 10%์˜ ํ›ˆ๋ จ ํ† ํฐ์œผ๋กœ๋งŒ ๊ตฌ์„ฑ๋œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Excitingly, most data points from the LLaMA models with different sizes (7B, 13B, 33B, 65B) fall on the same upwards trend. This observation further confirm our conclusion that the model's pre-training loss can predict its performance on downstream tasks, regardless of model size and token count. Note that there is one only exception at the early stage of LLaMA-65B. We can see that when the training loss is higher than 1.8, LLaMA-65B performs worse than smaller models with the same training loss. Without access to its intermediate checkpoints, we unfortunately cannot further analyze the result. Note that the outliers only constitute the initial 10% training tokens. [|assistant|]
* ๋ฐ์ดํ„ฐ๊ฐ€ ์ด๋ฏธ ์ˆ˜๋™์œผ๋กœ ํ•„ํ„ฐ๋ง/์ค‘๋ณต ์ œ๊ฑฐ๋œ ํ‘œ์ค€ LLM pre-training ์„ค์ •(์˜ˆ: MinHash)๊ณผ ์„ฑ๋Šฅ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ๋ชฉํ‘œ ๋ถ„ํฌ๋ฅผ ์•Œ ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. SSL Prototypes์˜ ์„ฑ๋Šฅ์€ ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์—์„œ ์ค‘๋ณต์œผ๋กœ ์ธํ•œ ํด๋Ÿฌ์Šคํ„ฐ์˜ ์˜ํ–ฅ์„ ๋ฐ›๋Š”๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 3.4์—์„œ๋Š” ์ด๋Ÿฌํ•œ ํด๋Ÿฌ์Šคํ„ฐ์˜ ์˜ํ–ฅ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด SemDeDup์„ ํ™œ์šฉํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต **D4**๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. * ์„น์…˜ 4.1์—์„œ๋Š” ์†Œ์Šค ๋ฐ์ดํ„ฐ๊ฐ€ "๋ฌดํ•œ๋Œ€"์ด๊ณ  ๊ณ ์ •๋œ ํ† ํฐ ์˜ˆ์‚ฐ์œผ๋กœ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” 'compute-limited regime'์—์„œ ๋ฌด์ž‘์œ„ iid ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐ ์ด์ „์— ํ™•๋ฆฝ๋œ ๋ฐฉ๋ฒ•๋ณด๋‹ค ๋” ๋‚˜์€ pre-training perplexity์™€ downstream accuracy๋ฅผ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ• D4๋Š” 6.7b ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ์•ฝ 20%์˜ ํšจ์œจ์„ฑ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํšจ์œจ์„ฑ ํ–ฅ์ƒ์˜ ๊ทœ๋ชจ๋Š” ๋ชจ๋ธ ๊ทœ๋ชจ์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. * ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜์—ฌ ๋ฐ์ดํ„ฐ๋ฅผ epochํ•ด์•ผ ํ•˜๋Š” _data-limited regime_์—์„œ๋Š” ์–ด๋–ค ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณตํ• ์ง€ ํ˜„๋ช…ํ•˜๊ฒŒ ์„ ํƒํ•˜๋ฉด ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ๋ฌด์ž‘์œ„๋กœ ๋ฐ˜๋ณตํ•  ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค(์„น์…˜ 4.2). ์ด๋Š” ๋‹จ์ผ epoch LLM ํ›ˆ๋ จ์˜ ํ‘œ์ค€ ๊ด€ํ–‰์— ์˜๋ฌธ์„ ์ œ๊ธฐํ•˜๊ณ , ์ง€๋Šฅ์ ์œผ๋กœ ํ•˜์œ„ ์„ ํƒ๋œ ๋ฐ์ดํ„ฐ๋ฅผ epochingํ•˜๋Š” ๊ฒƒ์ด ๋” ๋‚˜์€ ์ ‘๊ทผ ๋ฐฉ์‹์ผ ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
* ๋ฐ์ดํ„ฐ๊ฐ€ ์ด๋ฏธ ์ˆ˜๋™์œผ๋กœ ํ•„ํ„ฐ๋ง/์ค‘๋ณต ์ œ๊ฑฐ๋œ ํ‘œ์ค€ LLM ์‚ฌ์ „ ํ•™์Šต ์„ค์ •(์˜ˆ: MinHash)๊ณผ ์„ฑ๋Šฅ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ๋ชฉํ‘œ ๋ถ„ํฌ๋ฅผ ์•Œ ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. SSL ํ”„๋กœํ† ํƒ€์ž…์˜ ์„ฑ๋Šฅ์€ ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์—์„œ ์ค‘๋ณต์œผ๋กœ ์ธํ•œ ํด๋Ÿฌ์Šคํ„ฐ์˜ ์˜ํ–ฅ์„ ๋ฐ›๋Š”๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 3.4์—์„œ๋Š” ์ด๋Ÿฌํ•œ ํด๋Ÿฌ์Šคํ„ฐ์˜ ์˜ํ–ฅ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด SemDeDup์„ ํ™œ์šฉํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต **D4**๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. * ์„น์…˜ 4.1์—์„œ๋Š” ์†Œ์Šค ๋ฐ์ดํ„ฐ๊ฐ€ "๋ฌดํ•œ๋Œ€"์ด๊ณ  ๊ณ ์ •๋œ ํ† ํฐ ์˜ˆ์‚ฐ์œผ๋กœ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” '๊ณ„์‚ฐ ์ œํ•œ ์ •๊ถŒ'์—์„œ ๋ฌด์ž‘์œ„ iid ๋ฐ์ดํ„ฐ ์„ ํƒ ๋ฐ ์ด์ „์— ํ™•๋ฆฝ๋œ ๋ฐฉ๋ฒ•๋ณด๋‹ค ๋” ๋‚˜์€ ์‚ฌ์ „ ํ›ˆ๋ จ ๋‹นํ˜น๊ฐ๊ณผ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ• D4๋Š” 6.7b ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ์•ฝ 20%์˜ ํšจ์œจ์„ฑ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํšจ์œจ์„ฑ ํ–ฅ์ƒ์˜ ๊ทœ๋ชจ๋Š” ๋ชจ๋ธ ๊ทœ๋ชจ์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. * ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜์—ฌ ๋ฐ์ดํ„ฐ๋ฅผ ์—ํฌํฌํ•ด์•ผ ํ•˜๋Š” _๋ฐ์ดํ„ฐ ์ œํ•œ ์ •๊ถŒ_์—์„œ๋Š” ์–ด๋–ค ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณตํ• ์ง€ ํ˜„๋ช…ํ•˜๊ฒŒ ์„ ํƒํ•˜๋ฉด ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ๋ฌด์ž‘์œ„๋กœ ๋ฐ˜๋ณตํ•  ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค(์„น์…˜ 4.2). ์ด๋Š” ๋‹จ์ผ ์—ํฌํฌ LLM ํ›ˆ๋ จ์˜ ํ‘œ์ค€ ๊ด€ํ–‰์— ์˜๋ฌธ์„ ์ œ๊ธฐํ•˜๊ณ , ์ง€๋Šฅ์ ์œผ๋กœ ํ•˜์œ„ ์„ ํƒ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์—ํฌํ‚นํ•˜๋Š” ๊ฒƒ์ด ๋” ๋‚˜์€ ์ ‘๊ทผ ๋ฐฉ์‹์ผ ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* We investigate different data selection strategies for standard LLM pre-training setups where data has already been manually filtered / de-duplicated (e.g., MinHash), and where we do not know the target distribution for which we optimize performance. We argue that the performance of SSL Prototypes is affected by duplicate-driven clusters in the embedding space. In Section 3.4 we propose a new data selection strategy **D4** that utilizes SemDeDup to avoid getting impacted by such clusters. * In Section 4.1, we show that in the _compute-limited regime_ where we have "infinite" source data and train models with fixed token budgets, we can achieve better pre-training perplexity and downstream accuracy than random iid data selection and previously established methods. Furthermore, we show that our method D4 can achieve around 20% efficiency gains at the 6.7b model scale, and that the magnitude of efficiency gains increases with model scale. * In the _data-limited regime_, where we run out of data and must epoch over data, cleverly choosing what data to repeat can beat training on randomly selected new data, whereas randomly choosing data to repeat underperforms adding new data (Section 4.2). This calls into question the standard practice of single epoch LLM training, and suggests that epoching over intelligently subselected data might be a better approach. [|assistant|]
1. ์งˆ๋ฌธ์„ ๋ฐ”๊พธ๊ณ  ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค: 2023๋…„ 11์›” 30์ผ ์ด์ „์— Kevin McCarthy ์ „ ํ•˜์›์˜์žฅ์ด ๊ณต๊ฐœ์ ์œผ๋กœ ๋ฐœํ‘œํ•˜๊ฑฐ๋‚˜ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์ถœ์ฒ˜๋ฅผ ํ†ตํ•ด ์‚ฌ์ž„์„ ๋ณด๋„ํ•˜์—ฌ ๋ฏธ๊ตญ ํ•˜์›์˜์›์ง์—์„œ ๋ฌผ๋Ÿฌ๋‚˜๊ฒŒ ๋  ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ์Šต๋‹ˆ๊นŒ? ์ด ์งˆ๋ฌธ์€ ํ˜„์žฌ์˜ ์ •์น˜ ํ™˜๊ฒฝ, McCarthy์˜ ์ตœ๊ทผ ํ•˜์›์˜์žฅ์ง ์ƒ์‹ค, ์žฌ์„  ๋ถˆ์ถœ๋งˆ์— ๋Œ€ํ•œ ๊ณ ๋ ค, ๊ทธ๋ฆฌ๊ณ  ๊ทธ์˜ ๊ฒฐ์ •์ด ๊ณตํ™”๋‹น๊ณผ ์บ˜๋ฆฌํฌ๋‹ˆ์•„ ์ œ20์„ ๊ฑฐ๊ตฌ์— ๋ฏธ์น  ์ž ์žฌ์  ์˜ํ–ฅ์„ ๊ณ ๋ คํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์‹œ์žฅ์˜ ํ•ด๊ฒฐ ์—ฌ๋ถ€๋Š” McCarthy๊ฐ€ ์ง€์ •๋œ ๊ธฐํ•œ ๋‚ด์— ์˜์›์ง ์‚ฌํ‡ด ์˜์‚ฌ๋ฅผ ๊ณต์‹์ ์œผ๋กœ ์„ ์–ธํ• ์ง€ ์—ฌ๋ถ€์— ๋‹ฌ๋ ค ์žˆ์Šต๋‹ˆ๋‹ค. 2.[|endofturn|]
1. ์งˆ๋ฌธ์„ ๋ฐ”๊พธ๊ณ  ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค: 2023๋…„ 11์›” 30์ผ ์ด์ „์— ์ผ€๋นˆ ๋งค์นด์‹œ ์ „ ํ•˜์›์˜์žฅ์ด ๊ณต๊ฐœ์ ์œผ๋กœ ๋ฐœํ‘œํ•˜๊ฑฐ๋‚˜ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์ถœ์ฒ˜๋ฅผ ํ†ตํ•ด ์‚ฌ์ž„์„ ๋ณด๋„ํ•˜์—ฌ ๋ฏธ๊ตญ ํ•˜์›์˜์›์ง์—์„œ ๋ฌผ๋Ÿฌ๋‚˜๊ฒŒ ๋  ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ์Šต๋‹ˆ๊นŒ? ์ด ์งˆ๋ฌธ์€ ํ˜„์žฌ์˜ ์ •์น˜ ํ™˜๊ฒฝ, ๋งค์นด์‹œ์˜ ์ตœ๊ทผ ํ•˜์›์˜์žฅ์ง ์ƒ์‹ค, ์žฌ์„  ๋ถˆ์ถœ๋งˆ์— ๋Œ€ํ•œ ๊ณ ๋ ค, ๊ทธ๋ฆฌ๊ณ  ๊ทธ์˜ ๊ฒฐ์ •์ด ๊ณตํ™”๋‹น๊ณผ ์บ˜๋ฆฌํฌ๋‹ˆ์•„ ์ œ20์„ ๊ฑฐ๊ตฌ์— ๋ฏธ์น  ์ž ์žฌ์  ์˜ํ–ฅ์„ ๊ณ ๋ คํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์‹œ์žฅ์˜ ํ•ด๊ฒฐ ์—ฌ๋ถ€๋Š” ๋งค์นด์‹œ๊ฐ€ ์ง€์ •๋œ ๊ธฐํ•œ ๋‚ด์— ์˜์›์ง ์‚ฌํ‡ด ์˜์‚ฌ๋ฅผ ๊ณต์‹์ ์œผ๋กœ ์„ ์–ธํ• ์ง€ ์—ฌ๋ถ€์— ๋‹ฌ๋ ค ์žˆ์Šต๋‹ˆ๋‹ค. 2.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]1. Rephrased and Expanded Question: Is it likely that Kevin McCarthy, the former House Speaker, will make a public announcement or have his resignation reported by credible sources, indicating his departure from the United States House of Representatives, before the end of November 30, 2023? This question takes into account the current political climate, McCarthy's recent loss of the Speaker position, his consideration of not seeking re-election, and the potential impact of his decision on the Republican Party and the 20th District of California. The resolution of this market hinges on whether McCarthy will officially declare his intention to vacate his seat within the specified timeframe. 2. [|assistant|]
_Qualitative Evaluation_: ํ‘œ 2์— Pythia-6.9B์™€ FinPythia-6.9B๊ฐ€ ์ƒ์„ฑํ•œ ์ •์„ฑ์  ์‚ฌ๋ก€๊ฐ€ ์ œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ดํŽด๋ณด๋ฉด, FinPythia-6.9B๊ฐ€ ๊ธˆ์œต ๊ด€๋ จ ์งˆ๋ฌธ์— ๋Œ€ํ•ด ๋” ๊ด€๋ จ์„ฑ์ด ๋†’๊ณ  ์ƒ์„ธํ•œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€์†์ ์ธ pre-training์„ ํ†ตํ•ด 2021๋…„ ์ดํ›„์˜ ๊ธˆ์œต ์ด๋ฒคํŠธ ์ง€์‹์„ ์Šต๋“ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” DACP๊ฐ€ FinPythia-6.9B๊ฐ€ ๋„๋ฉ”์ธ ๋‚ด ์ง€์‹์„ ์Šต๋“ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
_์ •์„ฑ์  ํ‰๊ฐ€_: ํ‘œ 2์— ํŒŒ์ด์‹œ์•„-6.9B์™€ ํ•€ํŒŒ์ด์‹œ์•„-6.9B๊ฐ€ ์ƒ์„ฑํ•œ ์ •์„ฑ์  ์‚ฌ๋ก€๊ฐ€ ์ œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ดํŽด๋ณด๋ฉด, ํ•€ํŒŒ์ด์‹œ์•„-6.9B๊ฐ€ ๊ธˆ์œต ๊ด€๋ จ ์งˆ๋ฌธ์— ๋Œ€ํ•ด ๋” ๊ด€๋ จ์„ฑ์ด ๋†’๊ณ  ์ƒ์„ธํ•œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด 2021๋…„ ์ดํ›„์˜ ๊ธˆ์œต ์ด๋ฒคํŠธ ์ง€์‹์„ ์Šต๋“ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” DACP๊ฐ€ FinPythia-6.9B๊ฐ€ ๋„๋ฉ”์ธ ๋‚ด ์ง€์‹์„ ์Šต๋“ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]_Qualitative Evaluation_: qualitative examples generated by Pythia-6.9B and FinPythia-6.9B are presented in Table 2. Upon examination, we observe that FinPythia-6.9B generates more relevant and detailed responses for finance-related questions. It acquired the financial events knowledge post 2021 with the continual pre-training. These findings suggest that the DACP helps FinPythia-6.9B acquire in-domain knowledge. [|assistant|]
์ฒซ์งธ, ๋‹ค๊ตญ์–ด ๋ฐ์ดํ„ฐ์— monolingual tokenizer๋ฅผ ์ ์šฉํ•˜๋ฉด ํ›จ์”ฌ ๋” ๋†’์€ fertility์™€ parity ์ ์ˆ˜๊ฐ€ ์‚ฐ์ถœ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 0(a) ๋ฐ ๊ทธ๋ฆผ 2 ์ฐธ์กฐ). ๊ทธ๋ฆผ 0(b)์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด multilingual tokenizers๋Š” ๋ชจ๋“  ๋น„์˜์–ด ๋ฌธ์„œ์—์„œ fertility๊ฐ€ monolingual English tokenizers๋ณด๋‹ค ํ›จ์”ฌ ๋‚ฎ์ง€๋งŒ, ์˜์–ด ๋ฌธ์„œ tokenizing์—์„œ๋Š” ์•ฝ๊ฐ„๋งŒ ๋–จ์–ด์ง‘๋‹ˆ๋‹ค.[|endofturn|]
์ฒซ์งธ, ๋‹ค๊ตญ์–ด ๋ฐ์ดํ„ฐ์— ๋‹จ์ผ ์–ธ์–ด ํ† ํฐํ™”๊ธฐ๋ฅผ ์ ์šฉํ•˜๋ฉด ํ›จ์”ฌ ๋” ๋†’์€ ์ถœ์‚ฐ์œจ๊ณผ ํŒจ๋ฆฌํ‹ฐ ์ ์ˆ˜๊ฐ€ ์‚ฐ์ถœ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 0(a) ๋ฐ ๊ทธ๋ฆผ 2 ์ฐธ์กฐ). ๊ทธ๋ฆผ 0(b)์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ๋‹ค๊ตญ์–ด ํ† ํฐํ™”๊ธฐ๋Š” ๋ชจ๋“  ๋น„์˜์–ด ๋ฌธ์„œ์—์„œ ๋‹จ์–ด ์ˆ˜๊ฐ€ ์˜์–ด ํ† ํฐํ™”๊ธฐ๋ณด๋‹ค ํ›จ์”ฌ ๋‚ฎ์ง€๋งŒ, ์˜์–ด ๋ฌธ์„œ ํ† ํฐํ™”์—์„œ๋Š” ์•ฝ๊ฐ„๋งŒ ๋–จ์–ด์ง‘๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Firstly, it can be observed that applying a monolingual tokenizer to multilingual data results in significantly higher fertility and parity scores (see Fig. 0(a) and Fig. 2). While multilingual tokenizers have lower fertility than monolingual English tokenizers on all non-English documents by a large margin, they are only slightly worse on tokenizing English documents, as shown in Fig. 0(b). [|assistant|]
**Dataset Alignment Method**: ์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์—์ด์ „ํŠธ ํ”„๋กœํ•„์„ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋จผ์ € ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์žˆ๋Š” ์‹ค์ œ ์ธ๊ฐ„์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ž์—ฐ์–ด ํ”„๋กฌํ”„ํŠธ๋กœ ์ •๋ฆฌํ•œ ๋‹ค์Œ ์ด๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์—์ด์ „ํŠธ๋ฅผ ํ”„๋กœํŒŒ์ผ๋งํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, [29]์—์„œ๋Š” ๋ฏธ๊ตญ ๊ตญ๋ฏผ ์„ ๊ฑฐ ์—ฐ๊ตฌ(ANES) ์ฐธ์—ฌ์ž์˜ ์ธ์ข…/๋ฏผ์กฑ, ์„ฑ๋ณ„, ์—ฐ๋ น, ๊ฑฐ์ฃผ ์ง€์—ญ ๋“ฑ ์ธ๊ตฌํ†ต๊ณ„ํ•™์  ๋ฐฐ๊ฒฝ์„ ๊ธฐ๋ฐ˜์œผ๋กœ GPT-3์— ์—ญํ• ์„ ํ• ๋‹นํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ GPT-3๊ฐ€ ์‹ค์ œ ์ธ๊ฐ„๊ณผ ์œ ์‚ฌํ•œ ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. Dataset Alignment Method๋Š” ์‹ค์ œ ์ธ๊ตฌ์˜ ์†์„ฑ์„ ์ •ํ™•ํ•˜๊ฒŒ ํฌ์ฐฉํ•˜์—ฌ ์—์ด์ „ํŠธ ํ–‰๋™์ด ๋”์šฑ ์˜๋ฏธ ์žˆ๊ณ  ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋ฐ˜์˜ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**๋ฐ์ดํ„ฐ ์„ธํŠธ ์ •๋ ฌ ๋ฐฉ๋ฒ•**: ์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์—์ด์ „ํŠธ ํ”„๋กœํ•„์„ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋จผ์ € ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์žˆ๋Š” ์‹ค์ œ ์ธ๊ฐ„์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ž์—ฐ์–ด ํ”„๋กฌํ”„ํŠธ๋กœ ์ •๋ฆฌํ•œ ๋‹ค์Œ ์ด๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์—์ด์ „ํŠธ๋ฅผ ํ”„๋กœํŒŒ์ผ๋งํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, [29]์—์„œ๋Š” ๋ฏธ๊ตญ ๊ตญ๋ฏผ ์„ ๊ฑฐ ์—ฐ๊ตฌ(ANES) ์ฐธ์—ฌ์ž์˜ ์ธ์ข…/๋ฏผ์กฑ, ์„ฑ๋ณ„, ์—ฐ๋ น, ๊ฑฐ์ฃผ ์ง€์—ญ ๋“ฑ ์ธ๊ตฌํ†ต๊ณ„ํ•™์  ๋ฐฐ๊ฒฝ์„ ๊ธฐ๋ฐ˜์œผ๋กœ GPT-3์— ์—ญํ• ์„ ํ• ๋‹นํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ GPT-3๊ฐ€ ์‹ค์ œ ์ธ๊ฐ„๊ณผ ์œ ์‚ฌํ•œ ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ •๋ ฌ ๋ฐฉ๋ฒ•์€ ์‹ค์ œ ์ธ๊ตฌ์˜ ์†์„ฑ์„ ์ •ํ™•ํ•˜๊ฒŒ ํฌ์ฐฉํ•˜์—ฌ ์—์ด์ „ํŠธ ํ–‰๋™์ด ๋”์šฑ ์˜๋ฏธ ์žˆ๊ณ  ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋ฐ˜์˜ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Dataset Alignment Method**: in this method, the agent profiles are obtained from real-world datasets. Typically, one can first organize the information about real humans in the datasets into natural language prompts, and then leverage it to profile the agents. For instance, in [29], the authors assign roles to GPT-3 based on the demographic backgrounds (such as race/ethnicity, gender, age, and state of residence) of participants in the American National Election Studies (ANES). They subsequently investigate whether GPT-3 can produce similar results to those of real humans. The dataset alignment method accurately captures the attributes of the real population, thereby making the agent behaviors more meaningful and reflective of real-world scenarios. [|assistant|]
๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•๊ณผ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์•„์ง ์ฒด๊ณ„์ ์ธ ๊ฒ€ํ† ์™€ ๋ถ„์„์„ ๊ฑฐ์น˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜๊ธฐ ์œ„ํ•ด ์ด ์ž‘์—…์„ ์ œ์•ˆํ•˜์—ฌ XoT ์ œํ’ˆ๊ตฐ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ด๊ณ  ์ƒ์„ธํ•œ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์—ฐ์‡„ ์‚ฌ๊ณ ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•œ ๋ช‡ ๊ฐ€์ง€ ์„ค๋ฌธ์กฐ์‚ฌ๊ฐ€ ์žˆ์—ˆ์ง€๋งŒ, prompts๋ฅผ ์‚ฌ์šฉํ•œ LLM reasoning Qiao et al. (2023) ๋ฐ chain-of-thought prompt strategies Yu et al. (2023)๊ณผ ๊ฐ™์€ ํŠน์ • ์ธก๋ฉด์— ๊ตญํ•œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์ €ํฌ ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ์ด๋ฏธ ๋‹ค๋ฃจ์–ด์ง„ ์ฃผ์ œ์— ๋Œ€ํ•ด ๋ณด๋‹ค ์ฒ ์ €ํ•˜๊ณ  ํฌ๊ด„์ ์ธ ๋…ผ์˜๋ฅผ ์ œ๊ณตํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ XoT construction, XoT structural variants ๋ฐ frontier application ๋“ฑ๊ณผ ๊ฐ™์€ ์ถ”๊ฐ€ ์ฃผ์ œ์™€ ๋…ผ์˜๋„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋จผ์ € ๊ด€๋ จ ๋ฐฐ๊ฒฝ๊ณผ ์˜ˆ๋น„(SS2)๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, XoT construction methods(SS4.1), XoT structure variants(SS4.2), XoT enhancement methods(SS4.3) ๋“ฑ ๋‹ค์–‘ํ•œ ๊ด€์ ์—์„œ XoT ์‹œ๋ฆฌ์ฆˆ ์ž‘์—…์„ ๋ฉด๋ฐ€ํžˆ ๋ถ„๋ฅ˜ํ•˜๊ณ  ์‹ฌ์ธต ๋ถ„์„์„ ์™„๋ฃŒํ•ฉ๋‹ˆ๋‹ค(SS4). ๊ทธ๋Ÿฐ ๋‹ค์Œ frontier fields์—์„œ XoT์˜ ์‹ค์ œ ์ ์šฉ ์‚ฌ๋ก€๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(SS5). XoT์˜ ํ›„์† ์ž‘์—…์— ์˜๊ฐ์„ ์ฃผ๊ธฐ ์œ„ํ•ด ์ด ๋ถ„์•ผ์˜ ํ–ฅํ›„ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ž ์žฌ์  ๋ฐฉํ–ฅ์— ๋Œ€ํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(SS6). ๋งˆ์ง€๋ง‰์œผ๋กœ ๊ธฐ์กด ๋ฐฉ๋ฒ•์„ ๋น„๊ตํ•˜๊ณ  ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค(SS7).[|endofturn|]
๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•๊ณผ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์•„์ง ์ฒด๊ณ„์ ์ธ ๊ฒ€ํ† ์™€ ๋ถ„์„์„ ๊ฑฐ์น˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜๊ธฐ ์œ„ํ•ด ์ด ์ž‘์—…์„ ์ œ์•ˆํ•˜์—ฌ XoT ์ œํ’ˆ๊ตฐ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ด๊ณ  ์ƒ์„ธํ•œ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์—ฐ์‡„ ์‚ฌ๊ณ ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•œ ๋ช‡ ๊ฐ€์ง€ ์„ค๋ฌธ์กฐ์‚ฌ๊ฐ€ ์žˆ์—ˆ์ง€๋งŒ, ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•œ LLM ์ถ”๋ก  Qiao et al. (2023) ๋ฐ ์—ฐ์‡„ ์‚ฌ๊ณ  ํ”„๋กฌํ”„ํŠธ ์ „๋žต Yu et al. (2023)๊ณผ ๊ฐ™์€ ํŠน์ • ์ธก๋ฉด์— ๊ตญํ•œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์ €ํฌ ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ์ด๋ฏธ ๋‹ค๋ฃจ์–ด์ง„ ์ฃผ์ œ์— ๋Œ€ํ•ด ๋ณด๋‹ค ์ฒ ์ €ํ•˜๊ณ  ํฌ๊ด„์ ์ธ ๋…ผ์˜๋ฅผ ์ œ๊ณตํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ XoT ๊ตฌ์„ฑ, XoT ๊ตฌ์กฐ ๋ณ€ํ˜• ๋ฐ ํ”„๋ก ํ‹ฐ์–ด ์ ์šฉ ๋“ฑ๊ณผ ๊ฐ™์€ ์ถ”๊ฐ€ ์ฃผ์ œ์™€ ๋…ผ์˜๋„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋จผ์ € ๊ด€๋ จ ๋ฐฐ๊ฒฝ๊ณผ ์˜ˆ๋น„(SS2)๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, XoT ๊ตฌ์„ฑ ๋ฐฉ๋ฒ•(SS4.1), XoT ๊ตฌ์กฐ ๋ณ€ํ˜•(SS4.2), XoT ํ–ฅ์ƒ ๋ฐฉ๋ฒ•(SS4.3) ๋“ฑ ๋‹ค์–‘ํ•œ ๊ด€์ ์—์„œ XoT ์‹œ๋ฆฌ์ฆˆ ์ž‘์—…์„ ๋ฉด๋ฐ€ํžˆ ๋ถ„๋ฅ˜ํ•˜๊ณ  ์‹ฌ์ธต ๋ถ„์„์„ ์™„๋ฃŒํ•ฉ๋‹ˆ๋‹ค(SS4). ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ”„๋ก ํ‹ฐ์–ด ๋ถ„์•ผ์—์„œ XoT์˜ ์‹ค์ œ ์ ์šฉ ์‚ฌ๋ก€๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(SS5). XoT์˜ ํ›„์† ์ž‘์—…์— ์˜๊ฐ์„ ์ฃผ๊ธฐ ์œ„ํ•ด ์ด ๋ถ„์•ผ์˜ ํ–ฅํ›„ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ž ์žฌ์  ๋ฐฉํ–ฅ์— ๋Œ€ํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(SS6). ๋งˆ์ง€๋ง‰์œผ๋กœ ๊ธฐ์กด ๋ฐฉ๋ฒ•์„ ๋น„๊ตํ•˜๊ณ  ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค(SS7).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]However, these methods and datasets have not yet undergone systematic review and analysis. To fill this gap, we propose this work to conduct a comprehensive and detailed analysis of the XoT family. Even though there have been some surveys discussing chain-of-thought, they are limited to specific aspects, such as LLM reasoning with prompts Qiao et al. (2023) and chain-of-thought prompt strategies Yu et al. (2023). In contrast, our survey not only provides a more thorough and comprehensive discussion of the topics they've already covered, but also includes additional topics and discussions, such as XoT construction, XoT structural variants and frontier application, etc. Concretely, in this paper, we first introduce the relevant background and preliminary (SS2). Furthermore,we carefully classify the XoT series of work from multiple perspectives and complete an in-depth analysis (SS4), including XoT construction methods (SS4.1), XoT structure variants (SS4.2) and XoT enhancement methods (SS4.3). Then, we provide practical applications of the XoT in the frontier fields (SS5). In order to inspire the follow-up work of XoT, we offer insights into potential avenues for future research in this area (SS6). Finally, we compare and discuss existing methods (SS7). [|assistant|]
๊ทธ๋ฆผ 2์—์„œ ์šฐ๋ฆฌ๋Š” ์Šค์ผ€์ผ์— ๊ด€๊ณ„์—†์ด Transformer ๋ชจ๋ธ์˜ ๋Œ€๋ถ€๋ถ„์˜ ๋ ˆ์ด์–ด๊ฐ€ ์–ด๋Š ์ •๋„ ์ด๋ฐฉ์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด์—์„œ๋Š” ๋ชจ๋ธ์ด ๊ฑฐ์˜ ๋“ฑ๋ฐฉ์„ฑ์ด๊ฑฐ๋‚˜ ๋งค์šฐ ์ด๋ฐฉ์„ฑ์ธ ์ด๋ถ„๋ฒ•์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ์ด ์ด๋ถ„๋ฒ•์€ Pythia ์ œํ’ˆ๊ตฐ์˜ ํฌํ™” ํ˜„์ƒ๊ณผ ์ผ์น˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ๋Š”๋ฐ, 160M ์ดํ•˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํฌํ•จํ•˜๋Š” ๋ชจ๋ธ๋งŒ ๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด ์ด๋ฐฉ์„ฑ์˜ ์˜ํ–ฅ์„ ๋ฐ›๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 2์—์„œ ์šฐ๋ฆฌ๋Š” ์Šค์ผ€์ผ์— ๊ด€๊ณ„์—†์ด ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ์˜ ๋Œ€๋ถ€๋ถ„์˜ ๋ ˆ์ด์–ด๊ฐ€ ์–ด๋Š ์ •๋„ ์ด๋ฐฉ์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด์—์„œ๋Š” ๋ชจ๋ธ์ด ๊ฑฐ์˜ ๋“ฑ๋ฐฉ์„ฑ์ด๊ฑฐ๋‚˜ ๋งค์šฐ ์ด๋ฐฉ์„ฑ์ธ ์ด๋ถ„๋ฒ•์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ์ด ์ด๋ถ„๋ฒ•์€ ํŒŒ์ดํ‹ฐ์•„ ์ œํ’ˆ๊ตฐ์˜ ํฌํ™” ํ˜„์ƒ๊ณผ ์ผ์น˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ๋Š”๋ฐ, 160M ์ดํ•˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํฌํ•จํ•˜๋Š” ๋ชจ๋ธ๋งŒ ๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด ์ด๋ฐฉ์„ฑ์˜ ์˜ํ–ฅ์„ ๋ฐ›๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In Figure 2, we observe that most layers of Transformers models are anisotropic to some extent, regardless of the scale. Nevertheless, there seems to be a dichotomy in the last layer, where models are either nearly isotropic or highly anisotropic. Interestingly, we notice that the dichotomy aligns with the one of the saturation phenomenon for the Pythia suite, where only models containing 160M or fewer parameters seem affected by last-layer anisotropy. [|assistant|]
์ฆ๋ช…. \(q\)๋ฅผ \(p\)์™€ \(p^{\prime}\)์˜ ํ˜ผํ•ฉ๋ฌผ์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” \(\beta^{\prime}\geq\beta\)์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ์ฆ๋ช…ํ•ฉ๋‹ˆ๋‹ค. \(\beta^{\prime}\leq\beta\)์˜ ๊ฒฝ์šฐ๋Š” ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ฆ๋ช…. q\)๋ฅผ \(p\)์™€ \(p^{\prime}\)์˜ ํ˜ผํ•ฉ๋ฌผ์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” \(\๋ฒ ํƒ€^{\ํ”„๋ผ์ž„}\geq\๋ฒ ํƒ€\)์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ์ฆ๋ช…ํ•ฉ๋‹ˆ๋‹ค. \(\๋ฒ ํƒ€^{\ํ”„๋ผ์ž„}\leq\๋ฒ ํƒ€\)์˜ ๊ฒฝ์šฐ๋Š” ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Proof.Let \(q\) be the mixture of \(p\) and \(p^{\prime}\). We prove the result for \(\beta^{\prime}\geq\beta\); the case \(\beta^{\prime}\leq\beta\) is analogous. [|assistant|]
์ฑ… ๊ธธ์ด์˜ ์š”์•ฝ์€ ์•„์ง ์ดˆ๊ธฐ ๋‹จ๊ณ„์ด๊ธฐ ๋•Œ๋ฌธ์— ์ผ๊ด€์„ฑ(Chang et al., 2023b)๊ณผ faithfulness(SS3)์„ ๋„˜์–ด ๋‹ค๋ฅธ ์˜ค๋ฅ˜ ์œ ํ˜•์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋Š” ์•„์ง ๋ถ€์กฑํ•œ ์‹ค์ •์ด๋‹ค.[|endofturn|]
์ฑ… ๊ธธ์ด์˜ ์š”์•ฝ์€ ์•„์ง ์ดˆ๊ธฐ ๋‹จ๊ณ„์ด๊ธฐ ๋•Œ๋ฌธ์— ์ผ๊ด€์„ฑ(Chang et al., 2023b)๊ณผ ์ถฉ์‹ค์„ฑ(SS3)์„ ๋„˜์–ด ๋‹ค๋ฅธ ์˜ค๋ฅ˜ ์œ ํ˜•์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋Š” ์•„์ง ๋ถ€์กฑํ•œ ์‹ค์ •์ด๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]As book-length summarization is still a nascent area, research into other error types beyond coherence (Chang et al., 2023b) and faithfulness (SS3) is still lacking. [|assistant|]
์ด ์„น์…˜์—์„œ๋Š” ๋งคํฌ๋กœ๋ฐ์ดํ„ฐ ๊ตฌ์ฒดํ™” ํŒŒ์ดํ”„๋ผ์ธ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋™์•ˆ ์–ป์€ ์ถ”๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ถ€๋ก E.1 ๋ฐ ๋ถ€๋ก E.3์˜ ๊ฒฝ์šฐ, ์ด๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ดˆ๊ธฐ ๊ฐœ๋ฐœ ๋ฒ„์ „์„ ์‚ฌ์šฉํ•˜์—ฌ ์–ป์€ ๊ฒƒ์ด๋ฏ€๋กœ ๊ฒฐ๊ณผ๋Š” ๋ณธ๋ฌธ๊ณผ ์ง์ ‘ ๋น„๊ตํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋ถ€๋ก E.2์˜ ๊ฒฝ์šฐ, ์ด๋Š” Falcon-RW ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ด ์„น์…˜์—์„œ๋Š” ๋งคํฌ๋กœ๋ฐ์ดํ„ฐ ๊ตฌ์ฒดํ™” ํŒŒ์ดํ”„๋ผ์ธ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋™์•ˆ ์–ป์€ ์ถ”๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ถ€๋ก E.1 ๋ฐ ๋ถ€๋ก E.3์˜ ๊ฒฝ์šฐ, ์ด๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ดˆ๊ธฐ ๊ฐœ๋ฐœ ๋ฒ„์ „์„ ์‚ฌ์šฉํ•˜์—ฌ ์–ป์€ ๊ฒƒ์ด๋ฏ€๋กœ ๊ฒฐ๊ณผ๋Š” ๋ณธ๋ฌธ๊ณผ ์ง์ ‘ ๋น„๊ตํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋ถ€๋ก E.2์˜ ๊ฒฝ์šฐ, ์ด๋Š” ํŒ”์ฝ˜-RW ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this section, we present additional results obtained during the development of the Macrodata Refinement pipeline. For Appendix E.1 and Appendix E.3, these were obtained using earlier development versions of the dataset, so results are not directly comparable with the main text. For Appendix E.2, this is based on the Falcon-RW models. [|assistant|]
๋• ์œค๋ฆฌ ์ด๋ก  - ์–ด๋–ค ํ–‰๋™์˜ ๋„๋•์„ฑ์€ ๊ทธ ํ–‰๋™์„ ์ˆ˜ํ–‰ํ•˜๋Š” ์‚ฌ๋žŒ์˜ ์„ฑ๊ฒฉ์ด๋‚˜ ํŠน์„ฑ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง„๋‹ค๋Š” ์ด๋ก ์ž…๋‹ˆ๋‹ค.[|endofturn|]
3. ๋• ์œค๋ฆฌ ์ด๋ก  - ์–ด๋–ค ํ–‰๋™์˜ ๋„๋•์„ฑ์€ ๊ทธ ํ–‰๋™์„ ์ˆ˜ํ–‰ํ•˜๋Š” ์‚ฌ๋žŒ์˜ ์„ฑ๊ฒฉ์ด๋‚˜ ํŠน์„ฑ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง„๋‹ค๋Š” ์ด๋ก ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]3. Virtue ethical theories - which hold that the morality of an action depends on the character or traits of the person performing the action. [|assistant|]
**RAISE**. RAISE ๋ฐฉ๋ฒ•์€ ReAct ๋ฐฉ๋ฒ•์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ, ์ธ๊ฐ„์˜ ๋‹จ๊ธฐ ๋ฐ ์žฅ๊ธฐ ๊ธฐ์–ต์„ ๋ฐ˜์˜ํ•˜๋Š” memory ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ์ถ”๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค[16]. ์ด๋ฅผ ์œ„ํ•ด ๋‹จ๊ธฐ ์ €์žฅ์„ ์œ„ํ•œ scratchpad์™€ ์žฅ๊ธฐ ์ €์žฅ์„ ์œ„ํ•œ ์œ ์‚ฌํ•œ ์ด์ „ ์˜ˆ์ œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**RAISE**. RAISE ๋ฐฉ๋ฒ•์€ ReAct ๋ฐฉ๋ฒ•์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ, ์ธ๊ฐ„์˜ ๋‹จ๊ธฐ ๋ฐ ์žฅ๊ธฐ ๊ธฐ์–ต์„ ๋ฐ˜์˜ํ•˜๋Š” ๋ฉ”๋ชจ๋ฆฌ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ์ถ”๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค[16]. ์ด๋ฅผ ์œ„ํ•ด ๋‹จ๊ธฐ ์ €์žฅ์„ ์œ„ํ•œ ์Šคํฌ๋ž˜์น˜ํŒจ๋“œ์™€ ์žฅ๊ธฐ ์ €์žฅ์„ ์œ„ํ•œ ์œ ์‚ฌํ•œ ์ด์ „ ์˜ˆ์ œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**RAISE.** The RAISE method is built upon the ReAct method, with the addition of a memory mechanism that mirrors human short-term and long-term memory [16]. It does this by using a scratchpad for short-term storage and a dataset of similar previous examples for long-term storage. [|assistant|]
๋ชจ๋ธ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ์ง€์‹์€ ์‚ฌ์ „ ํ•™์Šต์ด ๋๋‚˜๋ฉด ๊ณ ์ •๋˜์–ด ์ง€์‹ ์šฉ๋Ÿ‰๊ณผ ์ง€์‹ ์—…๋ฐ์ดํŠธ ์ธก๋ฉด์—์„œ ๋‹จ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์™ธ๋ถ€ ์ง€์‹์„ ๋„์ž…ํ•˜๋ฉด ์–ด๋Š ์ •๋„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์—ฌ์ „ํžˆ ๋ถˆ์™„์ „ํ•œ ํ•ด๊ฒฐ์ฑ…์ž…๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ๊ทผ๋ณธ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด continual learning Lange et al. (2022); Wang et al. (2023)์€ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋…ธ๋ ฅ์„ ์œ„ํ•œ ์œ ๋งํ•œ ๊ธธ๋กœ ์ž๋ฆฌ ์žก๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ชจ๋ธ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ์ง€์‹์€ ์‚ฌ์ „ ํ•™์Šต์ด ๋๋‚˜๋ฉด ๊ณ ์ •๋˜์–ด ์ง€์‹ ์šฉ๋Ÿ‰๊ณผ ์ง€์‹ ์—…๋ฐ์ดํŠธ ์ธก๋ฉด์—์„œ ๋‹จ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์™ธ๋ถ€ ์ง€์‹์„ ๋„์ž…ํ•˜๋ฉด ์–ด๋Š ์ •๋„ ์™„ํ™”ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์—ฌ์ „ํžˆ ๋ถˆ์™„์ „ํ•œ ํ•ด๊ฒฐ์ฑ…์ž…๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ๊ทผ๋ณธ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ง€์†์ ์ธ ํ•™์Šต ๋ž‘๊ฒŒ(2022); ์™•(2023)์€ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋…ธ๋ ฅ์„ ์œ„ํ•œ ์œ ๋งํ•œ ๊ธธ๋กœ ์ž๋ฆฌ ์žก๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The parameterized knowledge in the model is fixed at the end of the pre-training, which leads to its shortcomings in terms of knowledge capacity and knowledge updating. While introducing external knowledge can alleviate this to some extent, it remains an imperfect solution. To fundamentally tackle this issue, continual learning Lange et al. (2022); Wang et al. (2023) stands as a promising avenue for future research endeavors. [|assistant|]
์Šค์ผ€์ผ๋ง ๋ฒ•์น™: ์‹ ๊ฒฝ Scaling laws๋Š” ๋น„์ „, ์–ธ์–ด ๋ฐ ์Œ์„ฑ์—์„œ ๋ณดํŽธ์ ์œผ๋กœ ๊ด€์ฐฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ๋Œ€๊ทœ๋ชจ ์‹ค์ฆ ์—ฐ๊ตฌ๋Š” (Hestness et al., 2017; Rosenfeld et al., 2020)์—์„œ ์ˆ˜ํ–‰๋˜์—ˆ์œผ๋ฉฐ, ๋‹ค์–‘ํ•œ ํ•™์Šต ์‹œ๋‚˜๋ฆฌ์˜ค์— ๊ฑธ์ณ power law scaling์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด์–ด์„œ OpenAI(Kaplan et al., 2020)์™€ DeepMind(Hoffmann et al., 2022)์˜ ๋Œ€๊ทœ๋ชจ ์—ฐ๊ตฌ๊ฐ€ ์ด์–ด์กŒ์œผ๋ฉฐ, ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ๊ทœ๋ชจ์— ๊ฑธ์ณ LLM์—์„œ power-law scaling์„ ์‹ค์ฆ์ ์œผ๋กœ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์Šค์ผ€์ผ๋ง ๋ฒ•์น™: ์‹ ๊ฒฝ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์€ ๋น„์ „, ์–ธ์–ด ๋ฐ ์Œ์„ฑ์—์„œ ๋ณดํŽธ์ ์œผ๋กœ ๊ด€์ฐฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ๋Œ€๊ทœ๋ชจ ์‹ค์ฆ ์—ฐ๊ตฌ๋Š” (Hestness et al., 2017; Rosenfeld et al., 2020)์—์„œ ์ˆ˜ํ–‰๋˜์—ˆ์œผ๋ฉฐ, ๋‹ค์–‘ํ•œ ํ•™์Šต ์‹œ๋‚˜๋ฆฌ์˜ค์— ๊ฑธ์ณ ๊ฑฐ๋“ญ์ œ๊ณฑ ๋ฒ•์น™ ์Šค์ผ€์ผ๋ง์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด์–ด์„œ OpenAI(Kaplan et al., 2020)์™€ DeepMind(Hoffmann et al., 2022)์˜ ๋Œ€๊ทœ๋ชจ ์—ฐ๊ตฌ๊ฐ€ ์ด์–ด์กŒ์œผ๋ฉฐ, ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ๊ทœ๋ชจ์— ๊ฑธ์ณ LLM์—์„œ ๊ฑฐ๋“ญ์ œ๊ณฑ ๋ฒ•์น™ ์Šค์ผ€์ผ๋ง์„ ์‹ค์ฆ์ ์œผ๋กœ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Scaling Laws:Neural scaling laws have been ubiquitously observed in vision, language and speech. Early large scale empirical studies are performed in (Hestness et al., 2017; Rosenfeld et al., 2020), demonstrating power law scaling across a range of learning scenarios. This is followed by well-known large-scale studies from OpenAI (Kaplan et al., 2020) and DeepMind (Hoffmann et al., 2022), which empirically demonstrate power-law scaling in LLMs across a wide set of scales. [|assistant|]
\(\bullet\)_OpenAssistant_[173]๋Š” ์ธ๊ฐ„๊ณผ AI assistant ๊ฐ„์˜ ์‹ค์ œ ๋Œ€ํ™” ํŠธ๋ฆฌ 66,497๊ฐœ๊ฐ€ ํฌํ•จ๋œ ๋‹ค๊ตญ์–ด ์ฝ”ํผ์Šค์ž…๋‹ˆ๋‹ค. ๊ฐ ๋Œ€ํ™” ํŠธ๋ฆฌ๋Š” ์—ฌ๋Ÿฌ ๋…ธ๋“œ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ๊ฐ ๋…ธ๋“œ๋Š” ๋Œ€ํ™”์—์„œ ์—ญํ• ์— ์˜ํ•ด ์ƒ์„ฑ๋œ ์ •๋ณด๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. 35๊ฐœ ์–ธ์–ด์— ๊ฑธ์ณ ์žˆ์œผ๋ฉฐ, 461,292๊ฐœ์˜ ์‘๋‹ต์— ๋Œ€ํ•œ ์ˆ˜๋™์œผ๋กœ ์ฃผ์„์ด ๋‹ฌ๋ฆฐ ํ’ˆ์งˆ ๋“ฑ๊ธ‰์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_OpenAssistant_[173]๋Š” ์ธ๊ฐ„๊ณผ AI ๋น„์„œ ๊ฐ„์˜ ์‹ค์ œ ๋Œ€ํ™” ํŠธ๋ฆฌ 66,497๊ฐœ๊ฐ€ ํฌํ•จ๋œ ๋‹ค๊ตญ์–ด ์ฝ”ํผ์Šค์ž…๋‹ˆ๋‹ค. ๊ฐ ๋Œ€ํ™” ํŠธ๋ฆฌ๋Š” ์—ฌ๋Ÿฌ ๋…ธ๋“œ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ๊ฐ ๋…ธ๋“œ๋Š” ๋Œ€ํ™”์—์„œ ์—ญํ• ์— ์˜ํ•ด ์ƒ์„ฑ๋œ ์ •๋ณด๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. 35๊ฐœ ์–ธ์–ด์— ๊ฑธ์ณ ์žˆ์œผ๋ฉฐ, 461,292๊ฐœ์˜ ์‘๋‹ต์— ๋Œ€ํ•œ ์ˆ˜๋™์œผ๋กœ ์ฃผ์„์ด ๋‹ฌ๋ฆฐ ํ’ˆ์งˆ ๋“ฑ๊ธ‰์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_OpenAssistant_[173] is a multilingual corpus containing 66,497 real-world conversation trees between human and AI assistant. Each conversation tree consists of multiple nodes, and each node represents the information generated by a role in the dialogue. It spans 35 languages and includes 461,292 manually annotated quality ratings of responses. [|assistant|]
**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.[|endofturn|]
**์–‘์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.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**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. [|assistant|]
**Preference Datasets.** ์ €ํฌ ๋ชจ๋ธ์—๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์ด์šฉ ๊ฐ€๋Šฅํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ LLM์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ํ˜ผํ•ฉ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜คํ”ˆ ์†Œ์Šค ์ค‘๊ตญ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” LLama-factory์˜ _cvalues.rllf_, _comparison.gpt4_data_zh_ ๋ฐ _oast.rm_zh_์—์„œ ์œ ํ•ดํ•˜์ง€ ์•Š๊ณ  ์œ ์ตํ•œ ์„น์…˜๊ณผ huozi, zhihu๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์˜์–ด์˜ ๊ฒฝ์šฐ, ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” LLama-factory์˜ _comparison.gpt4_data_en_๊ณผ beavertails(Ji et al., 2024)์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด GPT-4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ "์„ ํƒํ•œ" ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” alpaca-gpt4(Peng et al., 2023)๋ฅผ ์ฑ„ํƒํ•˜๊ณ , "๊ฑฐ๋ถ€" ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ์•ฝํ•œ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉ๋˜๋Š” baichuan-6B(Yang et al., 2023)๋ฅผ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ด 183k Chinese pairs์™€ 46k English pairs๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
**์„ ํ˜ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ.** ์ €ํฌ ๋ชจ๋ธ์—๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์ด์šฉ ๊ฐ€๋Šฅํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ LLM์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ํ˜ผํ•ฉ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜คํ”ˆ ์†Œ์Šค ์ค‘๊ตญ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” LLama-factory์˜ _cvalues.rllf_, _comparison.gpt4_data_zh_ ๋ฐ _oast.rm_zh_์—์„œ ์œ ํ•ดํ•˜์ง€ ์•Š๊ณ  ์œ ์ตํ•œ ์„น์…˜๊ณผ huozi, zhihu๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์˜์–ด์˜ ๊ฒฝ์šฐ, ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” LLama-factory์˜ _comparison.gpt4_data_en_๊ณผ ๋น„๋ฒ„ํ…Œ์ผ(Ji et al., 2024)์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ†ตํ•ด ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด GPT-4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ "์„ ํƒํ•œ" ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” alpaca-gpt4(Peng et al., 2023)๋ฅผ ์ฑ„ํƒํ•˜๊ณ , "๊ฑฐ๋ถ€" ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ์•ฝํ•œ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉ๋˜๋Š” baichuan-6B(Yang et al., 2023)๋ฅผ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ด 183,000์Œ์˜ ์ค‘๊ตญ์–ด์™€ 46,000์Œ์˜ ์˜์–ด๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Preference Datasets.** Our model incorporates a blend of publicly accessible datasets and synthetic data from the LLM. The open-source Chinese datasets consist of non-harmful and beneficial sections from _cvalues.rllf_, _comparison.gpt4_data_zh_ and _oast.rm_zh_ in LLama-factory (Zheng et al., 2024), huozi, and zhihu. For English, the dataset includes _comparison.gpt4_data_en_ from LLama-factory and beavertails (Ji et al., 2024). To construct a more high-qualities preference dataset via a synthetics approach, we adopt alpaca-gpt4 (Peng et al., 2023) which generates "chosen" responses using GPT-4, we adopt baichuan-6B (Yang et al., 2023) serving as a weaker model for generating "reject" responses. The dataset comprises 183k Chinese pairs and 46k English pairs in total. [|assistant|]
* ๋‚ด์žฌ์  tokenizer ์„ฑ๋Šฅ์„ ์กฐ์‚ฌํ•˜๋Š” ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. * tokenizer๊ฐ€ ๋ชจ๋ธ์˜ downstream ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ, ์ฆ‰ tokenizer์˜ ์™ธ์žฌ์  ์„ฑ๋Šฅ์„ ์กฐ์‚ฌํ•˜๋Š” ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. * ๋‚ด์žฌ์  tokenizer ์„ฑ๋Šฅ๊ณผ ์™ธ์žฌ์  tokenizer ์„ฑ๋Šฅ ๊ฐ„์— ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ๋Š”์ง€ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
* ๋‚ด์žฌ์  ํ† ํฐํ™”๊ธฐ ์„ฑ๋Šฅ์„ ์กฐ์‚ฌํ•˜๋Š” ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. * ํ† ํฐํ™”๊ธฐ๊ฐ€ ๋ชจ๋ธ์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ, ์ฆ‰ ํ† ํฐํ™”๊ธฐ์˜ ์™ธ์žฌ์  ์„ฑ๋Šฅ์„ ์กฐ์‚ฌํ•˜๋Š” ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. * ๋‚ด์žฌ์  ํ† ํฐํ™”๊ธฐ ์„ฑ๋Šฅ๊ณผ ์™ธ์žฌ์  ํ† ํฐํ™”๊ธฐ ์„ฑ๋Šฅ ๊ฐ„์— ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ๋Š”์ง€ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* We conduct a study investigating the intrinsic tokenizer performance. * We conduct a study investigating the extrinsic tokenizer performance, i.e., the impact of the tokenizer on the model's downstream performance. * We investigate whether a correlation between the intrinsic and the extrinsic tokenizer performance exists. [|assistant|]
์ฆ๊ฐ• ์ ‘๊ทผ ๋ฐฉ์‹ ๊ณ„ํš์€ ๊ฒ€์ƒ‰ ๊ธฐ๋ฐ˜, ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜, ์ •์˜ ์–ธ์–ด ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ•˜์—ฌ ๊ธฐ์กด์˜ ์ˆœ์ฐจ์  ๊ณ„ํš์„ ๋ฐœ์ „์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์ผ๋ถ€ ๋ฐฉ๋ฒ•์€ ์•ก์…˜, ๊ณ„ํš, ๋ฐ˜์‚ฌ ๋˜๋Š” ๋„๊ตฌ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ LLMs์˜ ์žฅ๊ธฐ ๊ณ„ํš ๋ฐ ์˜ค๋ฅ˜ ๋ณต์›๋ ฅ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ฆ๊ฐ• ์ ‘๊ทผ ๋ฐฉ์‹ ๊ณ„ํš์€ ๊ฒ€์ƒ‰ ๊ธฐ๋ฐ˜, ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜, ์ •์˜ ์–ธ์–ด ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ•˜์—ฌ ๊ธฐ์กด์˜ ์ˆœ์ฐจ์  ๊ณ„ํš์„ ๋ฐœ์ „์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์ผ๋ถ€ ๋ฐฉ๋ฒ•์€ ์•ก์…˜, ๊ณ„ํš, ๋ฐ˜์‚ฌ ๋˜๋Š” ๋„๊ตฌ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ LLM์˜ ์žฅ๊ธฐ ๊ณ„ํš ๋ฐ ์˜ค๋ฅ˜ ๋ณต์›๋ ฅ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Planning augmented approaches have advanced conventional sequential planning by introducing search-based, graph-based, and definition language-based methods. On the other hand, some methods incorporate action, planning, reflection, or tools, aiming to enhance LLMs' long-term planning and error resilience capabilities. [|assistant|]
**์„ค๋ฌธ ์กฐ์‚ฌ ์ž‘์„ฑ**. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ์—ฐ๊ตฌํŒ€์—์„œ ๊ฐœ์ตœํ•œ ํ† ๋ก ํšŒ์—์„œ ๊ณ„ํš๋œ ๊ฒƒ์œผ๋กœ, ์—ฐ๊ตฌํŒ€์›๋“ค์ด ์ตœ๊ทผ์˜ Large Language Models (LLMs)์˜ ๋ฐœ์ „์„ ํŒ€์›๋“ค์ด ์ฝ๊ธฐ ์‰ฌ์šด ๋ณด๊ณ ์„œ๋กœ ์š”์•ฝํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์ดˆ์•ˆ์€ 2023๋…„ 3์›” 13์ผ์— ์™„์„ฑ๋˜์—ˆ์œผ๋ฉฐ, ํŒ€์›๋“ค์€ ๋น„๊ต์  ๊ฐ๊ด€์ ์ด๊ณ  ํฌ๊ด„์ ์ธ ๋ฐฉ์‹์œผ๋กœ LLMs์— ๋Œ€ํ•œ ๊ด€๋ จ ์—ฐ๊ตฌ๋ฅผ ํฌํ•จํ•˜๊ธฐ ์œ„ํ•ด ์ตœ์„ ์„ ๋‹คํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ํ›„ ์—ฌ๋Ÿฌ ์ฐจ๋ก€์— ๊ฑธ์ณ ๊ธ€๊ณผ ๋‚ด์šฉ์„ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ˆ˜์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ณต๊ฐ„ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ๊ธฐ์กด LLMs์˜ ์ผ๋ถ€๋งŒ ๊ทธ๋ฆผ 3๊ณผ ํ‘œ 1์— ํฌํ•จํ•˜๋˜, ์„ ์ • ๊ธฐ์ค€์„ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ •๊ธฐ์ ์œผ๋กœ ์œ ์ง€ ๊ด€๋ฆฌํ•  GitHub ํŽ˜์ด์ง€([https://github.com/RUCAlBox/LLMSurvey](https://github.com/RUCAlBox/LLMSurvey))์—์„œ๋Š” ๋ณด๋‹ค ์™„ํ™”๋œ ๋ชจ๋ธ ์„ ํƒ ๊ธฐ์ค€์„ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. 2023๋…„ 3์›” 31์ผ์— ์ดˆ๊ธฐ ๋ฒ„์ „์„ ๊ณต๊ฐœํ•˜๊ณ , 2023๋…„ 6์›” 29์ผ์— ์ฃผ์š” ์ˆ˜์ • ๋ฒ„์ „์„, 2023๋…„ 9์›” 10์ผ์— ๋‘ ๋ฒˆ์งธ ๋ฒ„์ „์„, 2023๋…„ 11์›” 23์ผ์— ์ด ์ตœ์‹  ๋ฒ„์ „(์ฃผ์š” ์ˆ˜์ • ๋ฒ„์ „)์„ ๊ณต๊ฐœํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.[|endofturn|]
**์„ค๋ฌธ ์กฐ์‚ฌ ์ž‘์„ฑ**. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ์—ฐ๊ตฌํŒ€์—์„œ ๊ฐœ์ตœํ•œ ํ† ๋ก ํšŒ์—์„œ ๊ณ„ํš๋œ ๊ฒƒ์œผ๋กœ, ์—ฐ๊ตฌํŒ€์›๋“ค์ด ์ตœ๊ทผ์˜ ๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ฐœ์ „์„ ํŒ€์›๋“ค์ด ์ฝ๊ธฐ ์‰ฌ์šด ๋ณด๊ณ ์„œ๋กœ ์š”์•ฝํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์ดˆ์•ˆ์€ 2023๋…„ 3์›” 13์ผ์— ์™„์„ฑ๋˜์—ˆ์œผ๋ฉฐ, ํŒ€์›๋“ค์€ ๋น„๊ต์  ๊ฐ๊ด€์ ์ด๊ณ  ํฌ๊ด„์ ์ธ ๋ฐฉ์‹์œผ๋กœ LLM์— ๋Œ€ํ•œ ๊ด€๋ จ ์—ฐ๊ตฌ๋ฅผ ํฌํ•จํ•˜๊ธฐ ์œ„ํ•ด ์ตœ์„ ์„ ๋‹คํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ํ›„ ์—ฌ๋Ÿฌ ์ฐจ๋ก€์— ๊ฑธ์ณ ๊ธ€๊ณผ ๋‚ด์šฉ์„ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ˆ˜์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ณต๊ฐ„ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ๊ธฐ์กด LLM์˜ ์ผ๋ถ€๋งŒ ๊ทธ๋ฆผ 3๊ณผ ํ‘œ 1์— ํฌํ•จํ•˜๋˜, ์„ ์ • ๊ธฐ์ค€์„ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ •๊ธฐ์ ์œผ๋กœ ์œ ์ง€ ๊ด€๋ฆฌํ•  GitHub ํŽ˜์ด์ง€([https://github.com/RUCAlBox/LLMSurvey](https://github.com/RUCAlBox/LLMSurvey))์—์„œ๋Š” ๋ณด๋‹ค ์™„ํ™”๋œ ๋ชจ๋ธ ์„ ํƒ ๊ธฐ์ค€์„ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. 2023๋…„ 3์›” 31์ผ์— ์ดˆ๊ธฐ ๋ฒ„์ „์„ ๊ณต๊ฐœํ•˜๊ณ , 2023๋…„ 6์›” 29์ผ์— ์ฃผ์š” ์ˆ˜์ • ๋ฒ„์ „์„, 2023๋…„ 9์›” 10์ผ์— ๋‘ ๋ฒˆ์งธ ๋ฒ„์ „์„, 2023๋…„ 11์›” 23์ผ์— ์ด ์ตœ์‹  ๋ฒ„์ „(์ฃผ์š” ์ˆ˜์ • ๋ฒ„์ „)์„ ๊ณต๊ฐœํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Survey Writing**. This survey was planned during a discussion meeting held by our research team, and we aimed to summarize the recent advances of large language models as a highly readable report for our team members. The first draft was finished on March 13, 2023, in which our team members tried their best to include the related studies about LLMs in a relatively objective, comprehensive way. Then, we have extensively revised the writing and contents in several passes. Due to the space limit, we can only include a fraction of existing LLMs in Figure 3 and Table 1, by setting the selection criterion. However, we set a more relaxed criterion for model selection on our GitHub page ([https://github.com/RUCAlBox/LLMSurvey](https://github.com/RUCAlBox/LLMSurvey)), which will be regularly maintained. We release the initial version on March 31, 2023, the major revision on June 29, 2023, and second version on September 10, 2023, and this latest version (major revision) on November 23, 2023. [|assistant|]
๋งŽ์€ ์—ฐ๊ตฌํŒ€์ด ์ž์ฒด์ ์œผ๋กœ ๊ณ ์œ ํ•œ agent ๋ฒค์น˜๋งˆํฌ๋ฅผ agent ๊ตฌํ˜„๊ณผ ํ•จ๊ป˜ ์†Œ๊ฐœํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋™์ผํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ์—ฌ๋Ÿฌ agent ๊ตฌํ˜„์„ ๋น„๊ตํ•˜๋Š” ๊ฒƒ์ด ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด agent-specific ๋ฒค์น˜๋งˆํฌ์˜ ๋Œ€๋ถ€๋ถ„์—๋Š” ์ˆ˜์ž‘์—…์œผ๋กœ ์ฑ„์ ๋œ ๋งค์šฐ ๋ณต์žกํ•œ ํ‰๊ฐ€ ์„ธํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค[2]. ์ด๋Š” ๋ฐฉ๋ฒ•์˜ ๊ธฐ๋Šฅ์„ ๊ณ ํ’ˆ์งˆ๋กœ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ robustness๊ฐ€ ๋ถ€์กฑํ•˜๊ณ  ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ์‚ฌ๋žŒ์ด ๊ฒฐ๊ณผ๋ฅผ ์ž‘์„ฑํ•˜๊ณ  ์ฑ„์ ํ•˜๋Š” ์‚ฌ๋žŒ์ด๊ธฐ ๋•Œ๋ฌธ์— ํ‰๊ฐ€์— ํŽธ๊ฒฌ์ด ์œ ์ž…๋  ์œ„ํ—˜์ด ์žˆ์Šต๋‹ˆ๋‹ค. agent๋Š” ๋ชจ๋ธ, ํ™˜๊ฒฝ ๋˜๋Š” ๋ฌธ์ œ ์ƒํƒœ์˜ ๋ณ€๋™์„ฑ์œผ๋กœ ์ธํ•ด ์—ฌ๋Ÿฌ ๋ฐ˜๋ณต์—์„œ ์ผ๊ด€๋œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋ฌธ์ œ๊ฐ€ ์žˆ์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถ”๊ฐ€ ๋ฌด์ž‘์œ„์„ฑ์€ ๋” ์ž‘๊ณ  ๋ณต์žกํ•œ ํ‰๊ฐ€ ์„ธํŠธ์— ํ›จ์”ฌ ๋” ํฐ ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋งŽ์€ ์—ฐ๊ตฌํŒ€์ด ์ž์ฒด์ ์œผ๋กœ ๊ณ ์œ ํ•œ ์—์ด์ „ํŠธ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์—์ด์ „ํŠธ ๊ตฌํ˜„๊ณผ ํ•จ๊ป˜ ์†Œ๊ฐœํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋™์ผํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ตฌํ˜„์„ ๋น„๊ตํ•˜๋Š” ๊ฒƒ์ด ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ์—์ด์ „ํŠธ๋ณ„ ๋ฒค์น˜๋งˆํฌ์˜ ๋Œ€๋ถ€๋ถ„์—๋Š” ์ˆ˜์ž‘์—…์œผ๋กœ ์ฑ„์ ๋œ ๋งค์šฐ ๋ณต์žกํ•œ ํ‰๊ฐ€ ์„ธํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค[2]. ์ด๋Š” ๋ฐฉ๋ฒ•์˜ ๊ธฐ๋Šฅ์„ ๊ณ ํ’ˆ์งˆ๋กœ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฌ๊ณ ์„ฑ์ด ๋ถ€์กฑํ•˜๊ณ  ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ์‚ฌ๋žŒ์ด ๊ฒฐ๊ณผ๋ฅผ ์ž‘์„ฑํ•˜๊ณ  ์ฑ„์ ํ•˜๋Š” ์‚ฌ๋žŒ์ด๊ธฐ ๋•Œ๋ฌธ์— ํ‰๊ฐ€์— ํŽธ๊ฒฌ์ด ์œ ์ž…๋  ์œ„ํ—˜์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ๋Š” ๋ชจ๋ธ, ํ™˜๊ฒฝ ๋˜๋Š” ๋ฌธ์ œ ์ƒํƒœ์˜ ๋ณ€๋™์„ฑ์œผ๋กœ ์ธํ•ด ์—ฌ๋Ÿฌ ๋ฐ˜๋ณต์—์„œ ์ผ๊ด€๋œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๋ฌธ์ œ๊ฐ€ ์žˆ์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถ”๊ฐ€ ๋ฌด์ž‘์œ„์„ฑ์€ ๋” ์ž‘๊ณ  ๋ณต์žกํ•œ ํ‰๊ฐ€ ์„ธํŠธ์— ํ›จ์”ฌ ๋” ํฐ ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Many research teams introduce their own unique agent benchmarks alongside their agent implementation which makes comparing multiple agent implementations on the same benchmark challenging. Additionally, many of these new agent-specific benchmarks include a hand-crafted, highly complex, evaluation set where the results are manually scored [2]. This can provide a high-quality assessment of a method's capabilities, but it also lacks the robustness of a larger dataset and risks introducing bias into the evaluation, since the ones developing the method are also the ones writing and scoring the results. Agents can also have problems generating a consistent answer over multiple iterations, due to variability in the models, environment, or problem state. This added randomness poses a much larger problem to smaller, complex evaluation sets. [|assistant|]
์ฒ˜๋ฆฌ๋Ÿ‰์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด FSDP์˜ ๊ธฐ๋ณธ ์„ค์ •๊ณผ PyTorch์˜ amp ๋ชจ๋“ˆ์„ ํ†ตํ•ด mixed-precision training (Micikevicius et al., 2017)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํ›„์ž๋Š” softmax์™€ ๊ฐ™์€ ํŠน์ • ์—ฐ์‚ฐ์ด ํ•ญ์ƒ ์ „์ฒด ์ •๋ฐ€๋„๋กœ ์‹คํ–‰๋˜์–ด ์•ˆ์ •์„ฑ์„ ๊ฐœ์„ ํ•˜๋Š” ๋™์‹œ์— ๋‹ค๋ฅธ ๋ชจ๋“  ์—ฐ์‚ฐ์€ ํ•˜ํ”„ ์ •๋ฐ€๋„๋กœ ์‹คํ–‰๋˜๋ฉฐ, bfloat16 ํ˜•์‹์œผ๋กœ ์‹คํ–‰๋ฉ๋‹ˆ๋‹ค. ํŠน์ • ์„ค์ •์—์„œ ๊ฐ GPU์— ๊ตญํ•œ๋œ ์Œ์˜ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜์™€ optimizer ์ƒํƒœ๋Š” ์ „์ฒด ์ •๋ฐ€๋„๋กœ ์œ ์ง€๋ฉ๋‹ˆ๋‹ค. ๊ฐ Transformer ๋ธ”๋ก ๋‚ด์˜ ๊ฐ€์ค‘์น˜๋Š” ์ˆœ๋ฐฉํ–ฅ ๋ฐ ์—ญ๋ฐฉํ–ฅ ํŒจ์Šค ๋™์•ˆ ๊ฐ GPU์— ์ „์ฒด ํฌ๊ธฐ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ๊ตฌ์ฒดํ™”๋  ๋•Œ๋งŒ bfloat16์œผ๋กœ ์บ์ŠคํŒ…๋ฉ๋‹ˆ๋‹ค. Gradients๋Š” ์ „์ฒด ์ •๋ฐ€๋„๋กœ GPU ๊ฐ„์— ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ฒ˜๋ฆฌ๋Ÿ‰์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด FSDP์˜ ๊ธฐ๋ณธ ์„ค์ •๊ณผ PyTorch์˜ amp ๋ชจ๋“ˆ์„ ํ†ตํ•ด ํ˜ผํ•ฉ ์ •๋ฐ€๋„ ํ›ˆ๋ จ(Micikevicius et al., 2017)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํ›„์ž๋Š” ์†Œํ”„ํŠธ๋งฅ์Šค์™€ ๊ฐ™์€ ํŠน์ • ์—ฐ์‚ฐ์ด ํ•ญ์ƒ ์ „์ฒด ์ •๋ฐ€๋„๋กœ ์‹คํ–‰๋˜์–ด ์•ˆ์ •์„ฑ์„ ๊ฐœ์„ ํ•˜๋Š” ๋™์‹œ์— ๋‹ค๋ฅธ ๋ชจ๋“  ์—ฐ์‚ฐ์€ ํ•˜ํ”„ ์ •๋ฐ€๋„๋กœ ์‹คํ–‰๋˜๋ฉฐ, bfloat16 ํ˜•์‹์œผ๋กœ ์‹คํ–‰๋ฉ๋‹ˆ๋‹ค. ํŠน์ • ์„ค์ •์—์„œ ๊ฐ GPU์— ๊ตญํ•œ๋œ ์Œ์˜ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜์™€ ์ตœ์ ํ™” ์ƒํƒœ๋Š” ์ „์ฒด ์ •๋ฐ€๋„๋กœ ์œ ์ง€๋ฉ๋‹ˆ๋‹ค. ๊ฐ ํŠธ๋žœ์Šคํฌ๋จธ ๋ธ”๋ก ๋‚ด์˜ ๊ฐ€์ค‘์น˜๋Š” ์ˆœ๋ฐฉํ–ฅ ๋ฐ ์—ญ๋ฐฉํ–ฅ ํŒจ์Šค ๋™์•ˆ ๊ฐ GPU์— ์ „์ฒด ํฌ๊ธฐ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ๊ตฌ์ฒดํ™”๋  ๋•Œ๋งŒ bfloat16์œผ๋กœ ์บ์ŠคํŒ…๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋ผ๋””์–ธํŠธ๋Š” ์ „์ฒด ์ •๋ฐ€๋„๋กœ GPU ๊ฐ„์— ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To improve throughput, we employ mixed-precision training (Micikevicius et al., 2017) through FSDP's built-in settings and PyTorch's amp module. The latter ensures that certain operations like the softmax always run in full precision to improve stability, while all other operations run in half-precision with the bfloat16 format. Under our specific settings, the shaded model weights and optimizer state local to each GPU are kept in full precision. The weights within each transformer block are only cast to bfloat16 when the full-sized parameters are materialized on each GPU during the forward and backward passes. Gradients are reduced across GPUs in full precision. [|assistant|]
์ƒ๊ฐ์˜ ์—ฐ์‡„ ์ถ”๋ก ์€ ์ข…์ข… Hallucination์„ ์ผ์œผ์ผœ ์ž˜๋ชป๋œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„์˜ ์˜ค๋ฅ˜๋Š” ๋‹ค์‹œ ์—ฐ์‡„์ ์ธ ์˜ค๋ฅ˜๋ฅผ ์ผ์œผํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฆ์„ ํ†ตํ•ฉํ•˜์—ฌ ํ”ผ๋“œ๋ฐฑ์„ ์–ป์€ ๋‹ค์Œ ์ด ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์€ ์ธ๊ฐ„์˜ ์„ฑ์ฐฐ ๊ณผ์ •๊ณผ ์œ ์‚ฌํ•œ ์ด๋Ÿฌํ•œ ํ˜„์ƒ์„ ์™„ํ™”ํ•˜๋Š” ๋ฐ ๋งค์šฐ ํšจ๊ณผ์ ์ธ ์ „๋žต์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3์€ ๊ฒ€์ฆ ๋ฐ ๊ตฌ์ฒดํ™”์˜ ๊ฐœ์š”๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
์ƒ๊ฐ์˜ ์—ฐ์‡„ ์ถ”๋ก ์€ ์ข…์ข… ์ฐฉ์‹œ ํ˜„์ƒ์„ ์ผ์œผ์ผœ ์ž˜๋ชป๋œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„์˜ ์˜ค๋ฅ˜๋Š” ๋‹ค์‹œ ์—ฐ์‡„์ ์ธ ์˜ค๋ฅ˜๋ฅผ ์ผ์œผํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฆ์„ ํ†ตํ•ฉํ•˜์—ฌ ํ”ผ๋“œ๋ฐฑ์„ ์–ป์€ ๋‹ค์Œ ์ด ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์€ ์ธ๊ฐ„์˜ ์„ฑ์ฐฐ ๊ณผ์ •๊ณผ ์œ ์‚ฌํ•œ ์ด๋Ÿฌํ•œ ํ˜„์ƒ์„ ์™„ํ™”ํ•˜๋Š” ๋ฐ ๋งค์šฐ ํšจ๊ณผ์ ์ธ ์ „๋žต์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3์€ ๊ฒ€์ฆ ๋ฐ ๊ตฌ์ฒดํ™”์˜ ๊ฐœ์š”๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Chain-of-thought reasoning often tends to be hallucinatory, producing incorrect reasoning steps. Errors in intermediate reasoning steps can, in turn, trigger a cascade of errors. Incorporating verification to obtain feedback and subsequently refining the reasoning process based on this feedback can be a highly effective strategy for mitigating this phenomenon, which is similar to the process of human reflection. Figure 3 depicts the overview of verification and refinement. [|assistant|]
\(\bullet\)**๊ณ ํ•ด์ƒ๋„ ๋˜๋Š” ์žฅ๊ธฐ ๋น„์ „ ๋ฐ์ดํ„ฐ์—์„œ์˜ ์žฅ์ ์„ ๋” ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์ž…๋‹ˆ๋‹ค.** SSMs ์•„ํ‚คํ…์ฒ˜๋Š” ์ด๋ก ์ ์œผ๋กœ ๋ชจ๋ธ์˜ ๋ณต์žก์„ฑ์„ ํฌ๊ฒŒ ์ค„์—ฌ์ฃผ๊ธฐ ๋•Œ๋ฌธ์— ๊ณ ํ•ด์ƒ๋„ ๋ฐ์ดํ„ฐ(์›๊ฒฉ ๊ฐ์ง€ ๋ฐ์ดํ„ฐ, X-๋ ˆ์ด ์˜๋ฃŒ ์ด๋ฏธ์ง€) ๋˜๋Š” ์žฅ๊ธฐ ์‹œํ€€์Šค ๋ฐ์ดํ„ฐ(์žฅ๊ธฐ ๋น„๋””์˜ค ํ”„๋ ˆ์ž„)์— ๋Œ€ํ•œ ๋ชจ๋ธ๋ง ๊ธฐ๋Šฅ์ด ๋งค์šฐ ๊ฐ€์น˜๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ธก๋ฉด์€ Transformer ๋„คํŠธ์›Œํฌ์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์—์„œ๋Š” ์ž˜ ๋‹ค๋ฃจ์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)**๊ณ ํ•ด์ƒ๋„ ๋˜๋Š” ์žฅ๊ธฐ ๋น„์ „ ๋ฐ์ดํ„ฐ์—์„œ์˜ ์žฅ์ ์„ ๋” ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์ž…๋‹ˆ๋‹ค.** SSMs ์•„ํ‚คํ…์ฒ˜๋Š” ์ด๋ก ์ ์œผ๋กœ ๋ชจ๋ธ์˜ ๋ณต์žก์„ฑ์„ ํฌ๊ฒŒ ์ค„์—ฌ์ฃผ๊ธฐ ๋•Œ๋ฌธ์— ๊ณ ํ•ด์ƒ๋„ ๋ฐ์ดํ„ฐ(์›๊ฒฉ ๊ฐ์ง€ ๋ฐ์ดํ„ฐ, X-๋ ˆ์ด ์˜๋ฃŒ ์ด๋ฏธ์ง€) ๋˜๋Š” ์žฅ๊ธฐ ์‹œํ€€์Šค ๋ฐ์ดํ„ฐ(์žฅ๊ธฐ ๋น„๋””์˜ค ํ”„๋ ˆ์ž„)์— ๋Œ€ํ•œ ๋ชจ๋ธ๋ง ๊ธฐ๋Šฅ์ด ๋งค์šฐ ๊ฐ€์น˜๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ธก๋ฉด์€ ํŠธ๋žœ์Šคํฌ๋จธ ๋„คํŠธ์›Œํฌ์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์—์„œ๋Š” ์ž˜ ๋‹ค๋ฃจ์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)**To further explore its advantages in high-resolution or long-term vision data is a direction worthy of attention and research.** Since the SSMs architecture significantly reduces the complexity of the model theoretically, its modeling capability on high-resolution data (remote sensing data, X-ray medical images) or long-term sequence data (long-term video frames) is of great value. However, these aspects are not addressed well using other strong models like the Transformer network. [|assistant|]
ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ baseline์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” LMs์ด ๊ธฐ์กด ๋ฉ”ํŠธ๋ฆญ๋ณด๋‹ค ๋” ๋‚˜์€ ์ž๋™ํ™”๋œ ํ‰๊ฐ€์ž๊ฐ€ ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ œ์•ˆํ•˜์ง€๋งŒ[10], ์„น์…˜ 6.4์—์„œ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด GPT-4๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ์ •๋‹นํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. GPT-4์˜ ํŒ๋‹จ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„๊ณผ์˜ ์ผ์น˜๋„๊ฐ€ ๋†’๊ฑฐ๋‚˜ ๋†’์€ ์ธ๊ฐ„๊ณผ์˜ ์ผ์น˜๋„์™€ ์œ ์‚ฌํ•˜๊ฑฐ๋‚˜ ๋†’์€ ์ธ๊ฐ„๊ณผ์˜ ์ผ์น˜๋„๋ฅผ ๋ณด์ด๋Š” ๋“ฑ GPT-4์˜ ํŒ๋‹จ์ด ์ธ๊ฐ„๊ณผ ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ธฐ์ค€์„ ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” LM์ด ๊ธฐ์กด ๋ฉ”ํŠธ๋ฆญ๋ณด๋‹ค ๋” ๋‚˜์€ ์ž๋™ํ™”๋œ ํ‰๊ฐ€์ž๊ฐ€ ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ œ์•ˆํ•˜์ง€๋งŒ[10], ์„น์…˜ 6.4์—์„œ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด GPT-4๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ์ •๋‹นํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. GPT-4์˜ ํŒ๋‹จ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„๊ณผ์˜ ์ผ์น˜๋„๊ฐ€ ๋†’๊ฑฐ๋‚˜ ๋†’์€ ์ธ๊ฐ„๊ณผ์˜ ์ผ์น˜๋„์™€ ์œ ์‚ฌํ•˜๊ฑฐ๋‚˜ ๋†’์€ ์ธ๊ฐ„๊ณผ์˜ ์ผ์น˜๋„๋ฅผ ๋ณด์ด๋Š” ๋“ฑ GPT-4์˜ ํŒ๋‹จ์ด ์ธ๊ฐ„๊ณผ ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]test dataset as the baseline. While existing studies suggest LMs can be better automated evaluators than existing metrics [10], we conduct a human study to justify our usage of GPT-4 for evaluation in Sec. 6.4. We find GPT-4 judgments correlate strongly with humans, with human agreement with GPT-4 typically similar or higher than inter-human annotator agreement. [|assistant|]
๊ทธ๋ฆผ 8: Pythia-1B ๋ชจ๋ธ์˜ ํ—ค๋“œ์—์„œ ์ˆœ์œ„ \(d\)์— ๋Œ€ํ•œ ์ด๋ก ์  \(W\)-์˜ค๋ฅ˜์˜ ํ•จ์ˆ˜๋กœ์„œ, ํ•™์Šต๋œ ์ˆœ์œ„ ์ œํ•œ ํ—ค๋“œ(๋ชจ๋ฐฉ \(W^{s}_{d}\))๋ฅผ ์‚ฌ์šฉํ•œ ์ตœ์ข… Loss๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 8: ํ”ผํ‹ฐ์•„-1B ๋ชจ๋ธ์˜ ํ—ค๋“œ์—์„œ ์ˆœ์œ„ \(d\)์— ๋Œ€ํ•œ ์ด๋ก ์  \(W\)-์˜ค๋ฅ˜์˜ ํ•จ์ˆ˜๋กœ์„œ, ํ•™์Šต๋œ ์ˆœ์œ„ ์ œํ•œ ํ—ค๋“œ(๋ชจ๋ฐฉ \(W^{s}_{d}\))๋ฅผ ์‚ฌ์šฉํ•œ ์ตœ์ข… ์†์‹ค์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 8: Final loss with trained rank-constrained heads (mimicking \(W^{s}_{d}\)), as a function of the theoretical \(W\)-error for rank \(d\) on the head of the Pythia-1B model. [|assistant|]
์ด ์žฅ์—์„œ๋Š” chain-of-thought reasoning์— ๋Œ€ํ•œ ์ดˆ๊ธฐ ์‹œ๋„๋ฅผ ๋Œ€ํ‘œํ•˜๊ฑฐ๋‚˜ ํŠน์ • ์˜์—ญ์„ ์œ„ํ•ด ์„ค๊ณ„๋œ ๋‹ค๋ฅธ ์ž‘์—…์„ ๋‚˜์—ดํ•  ๊ฒƒ์ด๋‹ค.[|endofturn|]
์ด ์žฅ์—์„œ๋Š” ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ์ถ”๋ก ์— ๋Œ€ํ•œ ์ดˆ๊ธฐ ์‹œ๋„๋ฅผ ๋Œ€ํ‘œํ•˜๊ฑฐ๋‚˜ ํŠน์ • ์˜์—ญ์„ ์œ„ํ•ด ์„ค๊ณ„๋œ ๋‹ค๋ฅธ ์ž‘์—…์„ ๋‚˜์—ดํ•  ๊ฒƒ์ด๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this chapter, we will list other works that represent early attempts at chain-of-thought reasoning or are designed for specific domains. [|assistant|]
์‹คํ—˜ ์„ค์ •์˜ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋Š” GPT-4 ์Šน๋ฅ  ํŒ์ •์ž…๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์š”์•ฝ ๋ฐ ๋Œ€ํ™” ์‹คํ—˜์˜ ์Šน๋ฅ ์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ prompts๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ์‹คํ—˜์— ๋Œ€ํ•ด gpt-4-0314๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ํ‰๊ฐ€์—์„œ ์š”์•ฝ ๋˜๋Š” ์‘๋‹ต์˜ ์ˆœ์„œ๋Š” ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์‹คํ—˜ ์„ค์ •์˜ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋Š” GPT-4 ์Šน๋ฅ  ํŒ์ •์ž…๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์š”์•ฝ ๋ฐ ๋Œ€ํ™” ์‹คํ—˜์˜ ์Šน๋ฅ ์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ์‹คํ—˜์— ๋Œ€ํ•ด gpt-4-0314๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ํ‰๊ฐ€์—์„œ ์š”์•ฝ ๋˜๋Š” ์‘๋‹ต์˜ ์ˆœ์„œ๋Š” ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]A key component of our experimental setup is GPT-4 win rate judgments. In this section, we include the prompts used to generate win rates for the summarization and dialogue experiments. We use gpt-4-0314 for all our experiments. The order of summaries or responses are randomly chosen for every evaluation. [|assistant|]
์ƒˆ ์งˆ๋ฌธ์€ ํ•œ ๊ฐ€์ง€๋งŒ ๋ฌป๊ณ , ํ•ฉ๋ฆฌ์ ์ด๋ฉฐ, ์ฃผ์–ด์ง„ ์งˆ๋ฌธ์— ๊ทผ๊ฑฐํ•˜๊ณ , ์ˆซ์ž(์‹ค์ˆ˜ ๋˜๋Š” ์ •์ˆ˜)๋งŒ์œผ๋กœ ๋‹ตํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 'what is the amount of A, B and C?'๋ผ๊ณ  ๋ฌป์ง€ ๋งˆ์„ธ์š”.[|endofturn|]
์ƒˆ ์งˆ๋ฌธ์€ ํ•œ ๊ฐ€์ง€๋งŒ ๋ฌป๊ณ , ํ•ฉ๋ฆฌ์ ์ด๋ฉฐ, ์ฃผ์–ด์ง„ ์งˆ๋ฌธ์— ๊ทผ๊ฑฐํ•˜๊ณ , ์ˆซ์ž(์‹ค์ˆ˜ ๋˜๋Š” ์ •์ˆ˜)๋งŒ์œผ๋กœ ๋‹ตํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 'A, B, C์˜ ์–‘์ด ์–ผ๋งˆ์ธ๊ฐ€์š”?'๋ผ๊ณ  ๋ฌป์ง€ ๋งˆ์„ธ์š”.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Ensure the new question only asks for one thing, be reasonable, be based on the Given Question, and can be answered with only a number (float or integer). For example, DO NOT ask, 'what is the amount of A, B and C?'. [|assistant|]
๊ฐ์ฃผ 5: ์ „์ฒด ๊ธฐ์‚ฌ์˜ ๊ฒฐ๋ก ์„ ๋‚˜ํƒ€๋‚ด์ง€ ์•Š๋Š” ์ฒซ ๋ฒˆ์งธ ์„น์…˜๋งŒ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฌธ์„œ ๋์— <eos> token์„ ์ถ”๊ฐ€ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ฐ์ฃผ 5: ์ „์ฒด ๊ธฐ์‚ฌ์˜ ๊ฒฐ๋ก ์„ ๋‚˜ํƒ€๋‚ด์ง€ ์•Š๋Š” ์ฒซ ๋ฒˆ์งธ ์„น์…˜๋งŒ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฌธ์„œ ๋์— a_token์„ ์ถ”๊ฐ€ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Footnote 5: We do not append a <eos> token at the end of documents because we only use the first section, which does not signify the conclusion of the entire article. [|assistant|]
์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.[|endofturn|]
์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.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]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. [|assistant|]
**๋น„๊ณต๊ฐœ ๋ชจ๋ธ ๋ถ„์„**. ๋„ค ๊ฐ€์ง€ closed-source ๋ชจ๋ธ(_์ฆ‰, ChatGPT, Claude, Davinci003 ๋ฐ Davinci002)์— ๋Œ€ํ•œ ๋ถ„์„ ๋ฐ ์กฐ์‚ฌ ๊ฒฐ๊ณผ๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค:[|endofturn|]
**๋น„๊ณต๊ฐœ ๋ชจ๋ธ ๋ถ„์„**. ๋„ค ๊ฐ€์ง€ ๋น„๊ณต๊ฐœ ๋ชจ๋ธ(_์ฆ‰, ChatGPT, ํด๋กœ๋“œ, ๋‹ค๋นˆ์น˜003 ๋ฐ ๋‹ค๋นˆ์น˜002)์— ๋Œ€ํ•œ ๋ถ„์„ ๋ฐ ์กฐ์‚ฌ ๊ฒฐ๊ณผ๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค:[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Analysis of Closed-Source Models.** We summarize our analysis and findings of the four closed-source models (_i.e.,_ ChatGPT, Claude, Davinci003 and Davinci002) as follows: [|assistant|]
role-user: PersonU์™€ PersonA ์‚ฌ์ด์˜ ๋‹ค์Œ ๋Œ€ํ™”๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ: PersonU: ์•ˆ๋…•ํ•˜์„ธ์š”, ์–ด๋–ค ๊ฒฐ์ •์— ๋™์˜ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ ์–ด๋–ป๊ฒŒ ํ•ด์•ผ ํ•˜๋Š”์ง€ ์•Œ๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. PersonA: ์›” ์†Œ๋“ ์กฐ์ • ๊ธˆ์•ก์— ๋Œ€ํ•œ ๊ฒฐ์ •์— ๋™์˜ํ•˜์ง€ ์•Š์œผ์‹œ๋‚˜์š”? PersonU: ์•„๋‹ˆ์š”. ๋‚ด SHIP ์—ฐ๋ฝ์ฒ˜ ์ •๋ณด๋Š” ์–ด๋””์—์„œ ์ฐพ์„ ์ˆ˜ ์žˆ๋‚˜์š”? PersonA: ์ง€์—ญ SHIP ์—ฐ๋ฝ์ฒ˜ ์ •๋ณด๋Š” Medicare & You 2020 ํ•ธ๋“œ๋ถ ์˜จ๋ผ์ธ ๋’ท๋ถ€๋ถ„์—์„œ ์ฐพ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. PersonU: ์กฐ์ •์•ก์€ ์–ด๋–ป๊ฒŒ ๊ณ„์‚ฐํ•˜๋‚˜์š”? ์ด ๋Œ€ํ™”๋ฅผ ํ†ตํ•˜์ง€ ์•Š๊ณ  PersonU๊ฐ€ ์›ํ•˜๋Š” ๊ฒƒ์„ ์–ป์œผ๋ ค๋ฉด ์–ด๋–ป๊ฒŒ ํ•ด์•ผ ํ•˜๋‚˜์š”? ๊ทธ ์งˆ๋ฌธ์œผ๋กœ ๋‹ตํ•˜์„ธ์š”.[|endofturn|]
ole-user: PersonU์™€ PersonA ์‚ฌ์ด์˜ ๋‹ค์Œ ๋Œ€ํ™”๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ: PersonU: ์•ˆ๋…•ํ•˜์„ธ์š”, ์–ด๋–ค ๊ฒฐ์ •์— ๋™์˜ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ ์–ด๋–ป๊ฒŒ ํ•ด์•ผ ํ•˜๋Š”์ง€ ์•Œ๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. PersonA: ์›” ์†Œ๋“ ์กฐ์ • ๊ธˆ์•ก์— ๋Œ€ํ•œ ๊ฒฐ์ •์— ๋™์˜ํ•˜์ง€ ์•Š์œผ์‹œ๋‚˜์š”? PersonU: ์•„๋‹ˆ์š”. ๋‚ด SHIP ์—ฐ๋ฝ์ฒ˜ ์ •๋ณด๋Š” ์–ด๋””์—์„œ ์ฐพ์„ ์ˆ˜ ์žˆ๋‚˜์š”? PersonA: ์ง€์—ญ SHIP ์—ฐ๋ฝ์ฒ˜ ์ •๋ณด๋Š” Medicare & You 2020 ํ•ธ๋“œ๋ถ ์˜จ๋ผ์ธ ๋’ท๋ถ€๋ถ„์—์„œ ์ฐพ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. PersonU: ์กฐ์ •์•ก์€ ์–ด๋–ป๊ฒŒ ๊ณ„์‚ฐํ•˜๋‚˜์š”? ์ด ๋Œ€ํ™”๋ฅผ ํ†ตํ•˜์ง€ ์•Š๊ณ  PersonU๊ฐ€ ์›ํ•˜๋Š” ๊ฒƒ์„ ์–ป์œผ๋ ค๋ฉด ์–ด๋–ป๊ฒŒ ํ•ด์•ผ ํ•˜๋‚˜์š”? ๊ทธ ์งˆ๋ฌธ์œผ๋กœ ๋‹ตํ•˜์„ธ์š”.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]role-user: Given the following conversation between PersonU and PersonA: PersonU: Hello, I would like to know what to do if I do not agree with any decision. PersonA: disagree with our decision about your monthly income adjustment amounts? PersonU: no. Where can I find my SHIP contact information? PersonA: You can find your local SHIP contact information in the back of your Medicare & You 2020 Handbook online. PersonU: and how do they calculate the adjustments? Instead of having this entire conversation, how can PersonU get what he or she is looking for using a single question? Respond with that question. [|assistant|]
๊ฐœ์š”๋ฌธ์„œ ์ˆ˜์ค€ ํ•„ํ„ฐ๋ง์˜ ์ฐธ์กฐ ๋ชจ๋ธ ๊ด€ํ–‰์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ํ† ํฐ ์ˆ˜์ค€ ๋ฐ์ดํ„ฐ ์„ ํƒ์˜ ๊ฐ„๋‹จํ•œ ํŒŒ์ดํ”„๋ผ์ธ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๊ทธ๋ฆผ 4์— ๋‚˜์™€ ์žˆ๋Š” ์„ธ ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์—„์„ ๋œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด reference model์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ๋ชจ๋ธ์€ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค ๋‚ด์˜ ๊ฐ ํ† ํฐ loss๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰ ๋‹จ๊ณ„์—์„œ๋Š” ํ›ˆ๋ จ ๋ชจ๋ธ๊ณผ reference model ๊ฐ„์˜ ์ดˆ๊ณผ loss๊ฐ€ ํฐ ํ† ํฐ์— ์ง‘์ค‘ํ•˜์—ฌ language model์„ ์„ ํƒ์ ์œผ๋กœ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ฐœ์š”๋ฌธ์„œ ์ˆ˜์ค€ ํ•„ํ„ฐ๋ง์˜ ์ฐธ์กฐ ๋ชจ๋ธ ๊ด€ํ–‰์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ํ† ํฐ ์ˆ˜์ค€ ๋ฐ์ดํ„ฐ ์„ ํƒ์˜ ๊ฐ„๋‹จํ•œ ํŒŒ์ดํ”„๋ผ์ธ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๊ทธ๋ฆผ 4์— ๋‚˜์™€ ์žˆ๋Š” ์„ธ ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์—„์„ ๋œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์ฐธ์กฐ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ๋ชจ๋ธ์€ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค ๋‚ด์˜ ๊ฐ ํ† ํฐ ์†์‹ค์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰ ๋‹จ๊ณ„์—์„œ๋Š” ํ›ˆ๋ จ ๋ชจ๋ธ๊ณผ ์ฐธ์กฐ ๋ชจ๋ธ ๊ฐ„์˜ ์ดˆ๊ณผ ์†์‹ค์ด ํฐ ํ† ํฐ์— ์ง‘์ค‘ํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์„ ์„ ํƒ์ ์œผ๋กœ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]OverviewInspired by the practice of reference model in document-level filtering, we propose a simple pipeline of token-level data selection, termed "Selective Language Modeling (SLM)". Our method comprises three steps, as depicted in Figure 4. We begin by training a reference model on a curated, high-quality dataset. This model then assesses the loss of each token within the pretraining corpus. In the final phase, we train the language model selectively, focusing on tokens with high excess loss between the training and reference model. [|assistant|]
# ์ถœ๋ ฅ ์ด์ œ "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)"๊ฐ€ ๊ฒฐ์ •๋˜์—ˆ์œผ๋ฏ€๋กœ, ์‚ฌ์‹ค ๊ตฌ์„ฑ์˜ ์œ ํ˜•์„ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด ๊ด€๊ณ„ \(r_{2}\)์˜ ์œ ํ˜•์„ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค, "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)"์˜ \(\mathrm{type}(r_{2})\)". ์ด์ „ ๋‹จ๊ณ„์—์„œ ๊ฒฐ์ •๋œ "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)"๋Š” ๊ตญ๊ฐ€, ์กฐ์ง, ํ•™๋ถ€ ๋Œ€ํ•™, ๊ฒŒ์ž„ ๊ฐœ๋ฐœ์‚ฌ), ์‹ค์ œ ์ธ๋ฌผ(์ €์ž, ํšŒ์žฅ, CEO, ๋ฐฐ์šฐ์ž, ๊ฐ€์ˆ˜), ๊ฐ€์ƒ์˜ ์ธ๋ฌผ(์ฃผ์ธ๊ณต), ์˜ํ™”, ์†Œ์„ค ๋˜๋Š” ๋„์‹œ(๋ณธ์‚ฌ ๋„์‹œ)์˜ ๋ฒ”์ฃผ์— ์†ํ•œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”. ๋˜ํ•œ "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)"๋Š” ์„ค๋ช…์ด ์–ธ๊ธ‰๋œ ๊ต๋Ÿ‰ ์—”ํ‹ฐํ‹ฐ ์ž์ฒด์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ๋ฒ”์ฃผ์˜ \(e_{2}\)์— ๋Œ€ํ•ด ๊ด€๊ณ„ \(r_{2}\)๋ฅผ ๋งŒ์กฑํ•˜๋Š” ์œ ์ผํ•œ ๊ฐ์ฒด ์—”ํ‹ฐํ‹ฐ๊ฐ€ \(e_{3}\)์ธ ์ถฉ๋ถ„ํ•œ ์ˆ˜์˜ \((e_{2},r_{2},e_{3})\)๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋Š” \(r_{2}\)๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ๋‹จ๊ณ„์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ผ๋ฐ˜์ ์ธ ๊ด€๊ณ„๋ฅผ \(r_{2}\)๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์„ ํƒํ•œ \(r_{2}\)์˜ ์œ ํ˜•์„ ์‚ฌ์šฉํ•˜์—ฌ "๋…ธ๋ž˜์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ"(ํŠน์ • ์†Œ์„ค์˜ ์ถœ์ƒ ๋„์‹œ), "๋ณธ์‚ฌ ๋„์‹œ ๋น„๋””์˜ค ๊ฒŒ์ž„์˜ ๊ฐœ๋ฐœ์ž"(ํŠน์ • ๋น„๋””์˜ค ๊ฒŒ์ž„์˜ ๊ฐœ๋ฐœ์ž ๋ณธ์‚ฌ๊ฐ€ ์žˆ๋Š” ๋„์‹œ), "์ฃผ์ธ๊ณต์˜ ์˜ํ™” ๊ฐ๋…"(ํŠน์ • ์บ๋ฆญํ„ฐ๊ฐ€ ์ฃผ์ธ๊ณต์ธ ์˜ํ™”์˜ ๊ฐ๋…) ๋“ฑ 52๊ฐœ์˜ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ด์ œ "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)"๊ฐ€ ๊ฒฐ์ •๋˜์—ˆ์œผ๋ฏ€๋กœ, ์‚ฌ์‹ค ๊ตฌ์„ฑ์˜ ์œ ํ˜•์„ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด ๊ด€๊ณ„ \(r_{2}\)์˜ ์œ ํ˜•์„ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค, "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)"์˜ \(\mathrm{type}(r_{2})\)". ์ด์ „ ๋‹จ๊ณ„์—์„œ ๊ฒฐ์ •๋œ "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)"๋Š” ๊ตญ๊ฐ€, ์กฐ์ง, ํ•™๋ถ€ ๋Œ€ํ•™, ๊ฒŒ์ž„ ๊ฐœ๋ฐœ์‚ฌ), ์‹ค์ œ ์ธ๋ฌผ(์ €์ž, ํšŒ์žฅ, CEO, ๋ฐฐ์šฐ์ž, ๊ฐ€์ˆ˜), ๊ฐ€์ƒ์˜ ์ธ๋ฌผ(์ฃผ์ธ๊ณต), ์˜ํ™”, ์†Œ์„ค ๋˜๋Š” ๋„์‹œ(๋ณธ์‚ฌ ๋„์‹œ)์˜ ๋ฒ”์ฃผ์— ์†ํ•œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”. ๋˜ํ•œ "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)"๋Š” ์„ค๋ช…์ด ์–ธ๊ธ‰๋œ ๊ต๋Ÿ‰ ์—”ํ‹ฐํ‹ฐ ์ž์ฒด์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ๋ฒ”์ฃผ์˜ \(e_{2}\)์— ๋Œ€ํ•ด ๊ด€๊ณ„ \(r_{2}\)๋ฅผ ๋งŒ์กฑํ•˜๋Š” ์œ ์ผํ•œ ๊ฐ์ฒด ์—”ํ‹ฐํ‹ฐ๊ฐ€ \(e_{3}\)์ธ ์ถฉ๋ถ„ํ•œ ์ˆ˜์˜ \((e_{2},r_{2},e_{3})\)๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋Š” \(r_{2}\)๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ๋‹จ๊ณ„์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ผ๋ฐ˜์ ์ธ ๊ด€๊ณ„๋ฅผ \(r_{2}\)๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์„ ํƒํ•œ \(r_{2}\)์˜ ์œ ํ˜•์„ ์‚ฌ์šฉํ•˜์—ฌ "๋…ธ๋ž˜์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ"(ํŠน์ • ์†Œ์„ค์˜ ์ถœ์ƒ ๋„์‹œ), "๋ณธ์‚ฌ ๋„์‹œ ๋น„๋””์˜ค ๊ฒŒ์ž„์˜ ๊ฐœ๋ฐœ์ž"(ํŠน์ • ๋น„๋””์˜ค ๊ฒŒ์ž„์˜ ๊ฐœ๋ฐœ์ž ๋ณธ์‚ฌ๊ฐ€ ์žˆ๋Š” ๋„์‹œ), "์ฃผ์ธ๊ณต์˜ ์˜ํ™” ๊ฐ๋…"(ํŠน์ • ์บ๋ฆญํ„ฐ๊ฐ€ ์ฃผ์ธ๊ณต์ธ ์˜ํ™”์˜ ๊ฐ๋…) ๋“ฑ 52๊ฐœ์˜ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Now that we have "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)" determined, we determine the type of relations \(r_{2}\) to determine the type of the fact composition, "\(\mathrm{type}(r_{2})\) of \(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)". Note that "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)" determined in the previous step falls into the category of country, organization, undergraduate university, game developer), real person (author, president, CEO, spouse, singer), fictional character (main character), movie, novel, or city (headquarters city). Note that "\(\mathrm{type}(e_{1})\)'s \(\mathrm{type}(r_{1})\)" is also the bridge entity itself that the descriptive mention refers to. Therefore, we select \(r_{2}\) that are likely to give us a sufficient number of \((e_{2},r_{2},e_{3})\) where \(e_{3}\) is the only object entity satisfying the relation \(r_{2}\) for these categories of \(e_{2}\). As in the previous step, we select common relations as \(r_{2}\). Using the selected types of \(r_{2}\), we create 52 fact composition types including "mother of song's singer" (the city where the novel of a specific novel was born), "headquarters city video game's developer" (the city where the headquarters of the developer of a specific video game is located), and "director of main character's movie" (the director of the movie which has a specific character as the main character). [|assistant|]
๋” ํฐ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ 14B์ธ ๋ชจ๋ธ์ธ **phi-3-medium**๋„ ๋™์ผํ•œ tokenizer ๋ฐ architecture๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ **phi-3-mini**๋ฅผ ํ•™์Šตํ•˜๊ณ , ๋™์ผํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์•ฝ๊ฐ„ ๋” ๋งŽ์€ epoch(์ด 4.8T tokens) ๋™์•ˆ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ embedding ์ฐจ์›์ด 5120์ธ 40๊ฐœ์˜ head์™€ 40๊ฐœ์˜ layer๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ benchmark๋Š” 7B์—์„œ 14B๋กœ ๊ฐˆ์ˆ˜๋ก 3.8B์—์„œ 7B๋กœ ๊ฐˆ์ˆ˜๋ก benchmark๊ฐ€ 7B์—์„œ 14B๋กœ ๊ฐˆ์ˆ˜๋ก ํ›จ์”ฌ ๋œ ๊ฐœ์„ ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Š” 14B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด "data optimal regime"์— ๋„๋‹ฌํ•˜๊ธฐ ์œ„ํ•ด data mixture์— ์ถ”๊ฐ€ ์ž‘์—…์ด ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ผ๋ถ€ benchmark(HumanEval ํšŒ๊ท€ ํฌํ•จ)๋ฅผ ๊ณ„์† ์กฐ์‚ฌํ•˜๊ณ  ์žˆ์œผ๋ฏ€๋กœ **phi-3-medium**์˜ ์ˆ˜์น˜๋Š” "๋ฏธ๋ฆฌ๋ณด๊ธฐ"๋กœ ๊ฐ„์ฃผํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋” ํฐ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ 14B์ธ ๋ชจ๋ธ์ธ **phi-3-medium**๋„ ๋™์ผํ•œ ํ† ํฐํ™” ๋ฐ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ **phi-3-mini**๋ฅผ ํ•™์Šตํ•˜๊ณ , ๋™์ผํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์•ฝ๊ฐ„ ๋” ๋งŽ์€ ์—ํฌํฌ(์ด 4.8T ํ† ํฐ) ๋™์•ˆ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์ž„๋ฒ ๋”ฉ ์ฐจ์›์ด 5120์ธ 40๊ฐœ์˜ ํ—ค๋“œ์™€ 40๊ฐœ์˜ ๋ ˆ์ด์–ด๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ๋ฒค์น˜๋งˆํฌ๋Š” 7B์—์„œ 14B๋กœ ๊ฐˆ์ˆ˜๋ก 3.8B์—์„œ 7B๋กœ ๊ฐˆ์ˆ˜๋ก ๋ฒค์น˜๋งˆํฌ๊ฐ€ 7B์—์„œ 14B๋กœ ๊ฐˆ์ˆ˜๋ก ํ›จ์”ฌ ๋œ ๊ฐœ์„ ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Š” 14B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด "๋ฐ์ดํ„ฐ ์ตœ์  ๋ ˆ์ง"์— ๋„๋‹ฌํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์— ์ถ”๊ฐ€ ์ž‘์—…์ด ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ผ๋ถ€ ๋ฒค์น˜๋งˆํฌ(HumanEval ํšŒ๊ท€ ํฌํ•จ)๋ฅผ ๊ณ„์† ์กฐ์‚ฌํ•˜๊ณ  ์žˆ์œผ๋ฏ€๋กœ **phi-3-medium**์˜ ์ˆ˜์น˜๋Š” "๋ฏธ๋ฆฌ๋ณด๊ธฐ"๋กœ ๊ฐ„์ฃผํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To test our data on larger size of models, we also trained **phi-3-medium**, a model with 14B parameters using the same tokenizer and architecture of **phi-3-mini**, and trained on the same data for slightly more epochs (4.8T tokens total as for **phi-3-small**). The model has 40 heads and 40 layers, with embedding dimension 5120. We observe that some benchmarks improve much less from 7B to 14B than they do from 3.8B to 7B, perhaps indicating that our data mixture needs further work to be in the "data optimal regime" for 14B parameters model. We are still actively investigating some of those benchmarks (including a regression on HumanEval), hence the numbers for **phi-3-medium** should be considered as a "preview". [|assistant|]
๊ฐ์ฃผ 2: ALMA-13B-LoRA๋Š” ALMA ์ œํ’ˆ๊ตฐ์—์„œ ์ตœ๊ณ ์˜ 13B ๋ฒˆ์—ญ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” ๋‹จ์ผ ์–ธ์–ด ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด _์ „์ฒด ๊ฐ€์ค‘์น˜_ Fine-tuning์„ ๊ฑฐ์นœ ๋‹ค์Œ, _์ €์ˆœ์œ„ ์ ์‘_ (LoRA) (Hu et al., 2022)์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณ ํ’ˆ์งˆ์˜ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ๋ณ‘๋ ฌ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด Fine-tuningํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ฐ์ฃผ 2: ALMA-13B-LoRA๋Š” ALMA ์ œํ’ˆ๊ตฐ์—์„œ ์ตœ๊ณ ์˜ 13B ๋ฒˆ์—ญ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” ๋‹จ์ผ ์–ธ์–ด ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด _์ „์ฒด ๊ฐ€์ค‘์น˜_ ๋ฏธ์„ธ ์กฐ์ •์„ ๊ฑฐ์นœ ๋‹ค์Œ, _์ €์ˆœ์œ„ ์ ์‘_ (LoRA) (Hu et al., 2022)์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณ ํ’ˆ์งˆ์˜ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ๋ณ‘๋ ฌ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Footnote 2: ALMA-13B-LoRA is the best 13B translation model in the ALMA families. It initially undergoes _full-weight_ fine-tuning on monolingual data, followed by fine-tuning on high-quality human-written parallel data using _low-rank adaptation_ (LoRA) (Hu et al., 2022). [|assistant|]
๊ทธ๋ฆผ 6: **Compute-Normalized Generalization Results**. ์‚ฌ์šฉ๋œ ์ถ”๋ก  ํ˜ธ์ถœ ์ˆ˜๋กœ ์ •๊ทœํ™”๋œ ์„ฑ๋Šฅ ๊ณก์„ ์„ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 6: **์ปดํ“จํŒ… ์ •๊ทœํ™” ์ผ๋ฐ˜ํ™” ๊ฒฐ๊ณผ**. ์‚ฌ์šฉ๋œ ์ถ”๋ก  ํ˜ธ์ถœ ์ˆ˜๋กœ ์ •๊ทœํ™”๋œ ์„ฑ๋Šฅ ๊ณก์„ ์„ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 6: **Compute-Normalized Generalization Results**. We visualize the performance curves normalized by the number of inference calls used. [|assistant|]
๋จผ์ €, ์งˆ๋ฌธ \(Q\), ๊ด€๋ จ ๋ฌธ๋งฅ ๊ตฌ์ ˆ \(P\), ํ•ด๋‹น ๋‹ต๋ณ€ \(A\)๋ฅผ ํฌํ•จํ•˜๋Š” ๊ฐœ๋ฐฉํ˜• ๋„๋ฉ”์ธ question answering ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. \(DS_{Adapt}\)๋ฅผ ๋นˆ ์ง‘ํ•ฉ์œผ๋กœ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค(์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ 1์ค„). \(Q\)์˜ ๊ฐ ์งˆ๋ฌธ์— ๋Œ€ํ•ด retrieval ๋ฉ”์ปค๋‹ˆ์ฆ˜ ์—†์ด ๊ธฐ๋ณธ LLM์„ ํ™œ์šฉํ•˜์—ฌ zero-shot inference๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค(3์ค„). ์ด ๋‹จ๊ณ„๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์ด ์ •๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ์งˆ๋ฌธ๊ณผ ๊ทธ ๋ฐ˜์‘์ด ๋ถ€์ •ํ™•ํ•œ ์งˆ๋ฌธ์„ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์„ธ์Šค๋Š” ๊ธฐ๋ณธ LLM์ด ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ๋กœ ์ธํ•ด _์•Œ๊ณ  ์žˆ๋Š”_ ๊ฒƒ์„ ๋ฐœ๊ฒฌํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋จผ์ €, ์งˆ๋ฌธ \(Q\), ๊ด€๋ จ ๋ฌธ๋งฅ ๊ตฌ์ ˆ \(P\), ํ•ด๋‹น ๋‹ต๋ณ€ \(A\)๋ฅผ ํฌํ•จํ•˜๋Š” ๊ฐœ๋ฐฉํ˜• ๋„๋ฉ”์ธ ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. DS_{Adapt}\)๋ฅผ ๋นˆ ์ง‘ํ•ฉ์œผ๋กœ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค(์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ 1์ค„). Q\)์˜ ๊ฐ ์งˆ๋ฌธ์— ๋Œ€ํ•ด ๊ฒ€์ƒ‰ ๋ฉ”์ปค๋‹ˆ์ฆ˜ ์—†์ด ๊ธฐ๋ณธ LLM์„ ํ™œ์šฉํ•˜์—ฌ ์ œ๋กœ ์ƒท ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค(3์ค„). ์ด ๋‹จ๊ณ„๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์ด ์ •๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ์งˆ๋ฌธ๊ณผ ๊ทธ ๋ฐ˜์‘์ด ๋ถ€์ •ํ™•ํ•œ ์งˆ๋ฌธ์„ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์„ธ์Šค๋Š” ๊ธฐ๋ณธ LLM์ด ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ๋กœ ์ธํ•ด _์•Œ๊ณ  ์žˆ๋Š”_ ๊ฒƒ์„ ๋ฐœ๊ฒฌํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We begin by selecting an open-domain question answering dataset containing questions \(Q\), associated context passages \(P\), and corresponding answers \(A\). We initialize \(DS_{Adapt}\) to an empty set (line 1 of the algorithm). For each question in \(Q\), we leverage the base LLM without any retrieval mechanism to perform a zero-shot inference (line 3). This step allows us to differentiate questions for which the model generates correct answers from those where its responses are inaccurate. This process can be understood as a way to discover what the base LLM _knows_ due to its parametric memory. [|assistant|]
์ฃผ์„ ๋‹ฌ๊ธฐ ์–ด๋…ธํ…Œ์ดํ„ฐ๋Š” ์ด ์ž‘์—…์„ ์–ด๋–ป๊ฒŒ ์ธ์‹ํ•˜๋‚˜์š”? Annotators๋Š” ํŠนํžˆ ๊ตฌ์ฒด์ ์ธ ์ค„๊ฑฐ๋ฆฌ ํฌ์ธํŠธ๊ฐ€ ์•„๋‹Œ ์ฃผ์ œ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์ฃผ์žฅ์„ ๋‹ค๋ฃฐ ๋•Œ ์š”์•ฝ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์žˆ์–ด ๋ช‡ ๊ฐ€์ง€ ์–ด๋ ค์›€์„ ๊ฐ•์กฐํ•˜์—ฌ ํ…์ŠคํŠธ ๋‚ด์—์„œ ๊ด€๋ จ ๊ทผ๊ฑฐ๋ฅผ ์ฐพ๊ธฐ๊ฐ€ ์–ด๋ ต๋‹ค๋Š” ์ ์„ ์ง€์ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ์ •์ด๋‚˜ ์ฃผ์ œ์  ์ฃผ์žฅ๊ณผ ๊ฐ™์€ ์ถ”์ƒ์ ์ธ ๊ฐœ๋…์€ ์ƒ๋‹นํ•œ ์žฅ์• ๋ฌผ์ด์—ˆ์œผ๋ฉฐ, ์ผ๋ถ€ Annotators๋Š” ์ด๋Ÿฌํ•œ ์ฃผ์žฅ์„ ์ •ํ™•ํ•˜๊ฒŒ ๋’ท๋ฐ›์นจํ•˜๊ฑฐ๋‚˜ ๋ฐ˜๋ฐ•ํ•˜๋Š” ์ธ์šฉ๋ฌธ์„ ์ฐพ๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ถ€๋ถ„์ ์œผ๋กœ๋งŒ ์‚ฌ์‹ค์ธ ์ฃผ์žฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ๋„ ์–ด๋ ค์› ๋Š”๋ฐ, ์ด๋ฅผ ์œ„ํ•ด์„œ๋Š” ๋” ์ž์„ธํ•œ ๊ทผ๊ฑฐ๊ฐ€ ํ•„์š”ํ–ˆ์Šต๋‹ˆ๋‹ค(์‹ค์ œ ์˜๊ฒฌ์€ ํ‘œ 4 ์ฐธ์กฐ).[|endofturn|]
์ฃผ์„ ๋‹ฌ๊ธฐ ์–ด๋…ธํ…Œ์ดํ„ฐ๋Š” ์ด ์ž‘์—…์„ ์–ด๋–ป๊ฒŒ ์ธ์‹ํ•˜๋‚˜์š”? ์ฃผ์„ ๋‹ฌ๊ธฐ์ž๋“ค์€ ํŠนํžˆ ๊ตฌ์ฒด์ ์ธ ์ค„๊ฑฐ๋ฆฌ ํฌ์ธํŠธ๊ฐ€ ์•„๋‹Œ ์ฃผ์ œ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์ฃผ์žฅ์„ ๋‹ค๋ฃฐ ๋•Œ ์š”์•ฝ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์žˆ์–ด ๋ช‡ ๊ฐ€์ง€ ์–ด๋ ค์›€์„ ๊ฐ•์กฐํ•˜์—ฌ ํ…์ŠคํŠธ ๋‚ด์—์„œ ๊ด€๋ จ ๊ทผ๊ฑฐ๋ฅผ ์ฐพ๊ธฐ๊ฐ€ ์–ด๋ ต๋‹ค๋Š” ์ ์„ ์ง€์ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ์ •์ด๋‚˜ ์ฃผ์ œ์  ์ฃผ์žฅ๊ณผ ๊ฐ™์€ ์ถ”์ƒ์ ์ธ ๊ฐœ๋…์€ ์ƒ๋‹นํ•œ ์žฅ์• ๋ฌผ์ด์—ˆ์œผ๋ฉฐ, ์ผ๋ถ€ ์ฃผ์„๊ฐ€๋“ค์€ ์ด๋Ÿฌํ•œ ์ฃผ์žฅ์„ ์ •ํ™•ํ•˜๊ฒŒ ๋’ท๋ฐ›์นจํ•˜๊ฑฐ๋‚˜ ๋ฐ˜๋ฐ•ํ•˜๋Š” ์ธ์šฉ๋ฌธ์„ ์ฐพ๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ถ€๋ถ„์ ์œผ๋กœ๋งŒ ์‚ฌ์‹ค์ธ ์ฃผ์žฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ๋„ ์–ด๋ ค์› ๋Š”๋ฐ, ์ด๋ฅผ ์œ„ํ•ด์„œ๋Š” ๋” ์ž์„ธํ•œ ๊ทผ๊ฑฐ๊ฐ€ ํ•„์š”ํ–ˆ์Šต๋‹ˆ๋‹ค(์‹ค์ œ ์˜๊ฒฌ์€ ํ‘œ 4 ์ฐธ์กฐ).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]How do annotators perceive the task?Annotators highlighted several challenges in assessing the summaries, particularly when dealing with broad claims about themes rather than specific plot points, making it difficult to find relevant supporting evidence within the text. Abstract concepts, like emotions or thematic claims, posed significant obstacles, with some annotators struggling to locate quotations that precisely supported or refuted these claims. They also pointed out the difficulty of evaluating claims that were only partially true, which required more detailed support (see Table 4 for actual comments). [|assistant|]