chosen
stringlengths
14
4.26k
rejected
stringlengths
53
9.3k
prompt
stringlengths
178
4.55k
* ํŠธ๋ฆผ ํ‰๊ท ์˜ ๊ฒฝ์šฐ, ์ž…๋ ฅ ์˜ˆ์ธก์— ๊ท ์ผํ•œ ๊ฐ€์ค‘์น˜๋ฅผ ํ• ๋‹นํ•˜๊ณ , ์ค‘์•™๊ฐ’์—์„œ ๊ฐ€์žฅ ๋ฉ€๋ฆฌ ๋–จ์–ด์ง„ ์˜ˆ์ธก์˜ ๊ฐ€์ค‘์น˜๋ฅผ ์ ˆ๋ฐ˜์œผ๋กœ ์ค„์ด๊ณ , ๋‚˜๋จธ์ง€ ์ ˆ๋ฐ˜์˜ ๊ฐ€์ค‘์น˜๋ฅผ ๋‹ค๋ฅธ ์˜ˆ์ธก์— ๊ท ์ผํ•˜๊ฒŒ ์žฌ๋ถ„๋ฐฐํ•œ ๋‹ค์Œ, ์ตœ์ข…์ ์œผ๋กœ ๊ฐ€์ค‘ ํ‰๊ท ์„ ์ถœ๋ ฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ‘œ์ค€์ ์ธ ํŠธ๋ฆผ ํ‰๊ท  ๊ตฌํ˜„ ๋ฐฉ์‹์ด ์•„๋‹ˆ๋ฉฐ, ์ €ํฌ ์‹œ์Šคํ…œ์—์„œ ์†Œ์ˆ˜(์˜ˆ: 6๊ฐœ)์˜ ์˜ˆ์ธก๋งŒ ์ง‘๊ณ„ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด๋Ÿฐ ๋ฐฉ์‹์œผ๋กœ ์„ค์ •๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. * USC์˜ ๊ฒฝ์šฐ, (์™ธ๋ถ€) ์ถ”๋ก -์˜ˆ์ธก ์Œ์„ ์ตœ์ข… LM์— ์ œ์‹œํ•œ ๋‹ค์Œ, ์ด๋ฅผ ํ†ตํ•ด ์ง‘๊ณ„๋œ ์˜ˆ์ธก์„ ํ˜•์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•ฉ๋‹ˆ๋‹ค. ์ด hyperparameter sweep์—์„œ๋Š” GPT-4-1106-Preview๋ฅผ ์ง‘๊ณ„๊ธฐ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
* ํŠธ๋ฆผ ํ‰๊ท ์˜ ๊ฒฝ์šฐ, ์ž…๋ ฅ ์˜ˆ์ธก์— ๊ท ์ผํ•œ ๊ฐ€์ค‘์น˜๋ฅผ ํ• ๋‹นํ•˜๊ณ , ์ค‘์•™๊ฐ’์—์„œ ๊ฐ€์žฅ ๋ฉ€๋ฆฌ ๋–จ์–ด์ง„ ์˜ˆ์ธก์˜ ๊ฐ€์ค‘์น˜๋ฅผ ์ ˆ๋ฐ˜์œผ๋กœ ์ค„์ด๊ณ , ๋‚˜๋จธ์ง€ ์ ˆ๋ฐ˜์˜ ๊ฐ€์ค‘์น˜๋ฅผ ๋‹ค๋ฅธ ์˜ˆ์ธก์— ๊ท ์ผํ•˜๊ฒŒ ์žฌ๋ถ„๋ฐฐํ•œ ๋‹ค์Œ, ์ตœ์ข…์ ์œผ๋กœ ๊ฐ€์ค‘ ํ‰๊ท ์„ ์ถœ๋ ฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ‘œ์ค€์ ์ธ ํŠธ๋ฆผ ํ‰๊ท  ๊ตฌํ˜„ ๋ฐฉ์‹์ด ์•„๋‹ˆ๋ฉฐ, ์ €ํฌ ์‹œ์Šคํ…œ์—์„œ ์†Œ์ˆ˜(์˜ˆ: 6๊ฐœ)์˜ ์˜ˆ์ธก๋งŒ ์ง‘๊ณ„ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด๋Ÿฐ ๋ฐฉ์‹์œผ๋กœ ์„ค์ •๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. * USC์˜ ๊ฒฝ์šฐ, (์™ธ๋ถ€) ์ถ”๋ก -์˜ˆ์ธก ์Œ์„ ์ตœ์ข… LM์— ์ œ์‹œํ•œ ๋‹ค์Œ, ์ด๋ฅผ ํ†ตํ•ด ์ง‘๊ณ„๋œ ์˜ˆ์ธก์„ ํ˜•์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•ฉ๋‹ˆ๋‹ค. ์ด ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์Šค์œ•์—์„œ๋Š” GPT-4-1106-Preview๋ฅผ ์ง‘๊ณ„๊ธฐ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* For the trimmed mean, we assign uniform weights over the input forecasts, reduce the weight of the forecast furthest from the median by half, redistribute the half weight uniformly to the other forecasts, and finally output the weighted average. We remark that this is not a standard implementation of trimmed mean, and it is set this way since we only aggregate a small number (i.e., 6) of forecasts in our system. * For USC, we present the (external) reasoning-prediction pairs to a final LM, which is then prompted to form an aggregated forecast. In this hyperparameter sweep, we use GPT-4-1106-Preview as the aggregator model. [|assistant|]
**Instruction Datasets.** ์„น์…˜ 5.1.1์˜ ๋…ผ์˜์— ๋”ฐ๋ฅด๋ฉด, ์šฐ๋ฆฌ๋Š” ์ฃผ๋กœ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์„ธ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ ์ข…๋ฅ˜์˜ instructions๋ฅผ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค:[|endofturn|]
**๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ**. ์„น์…˜ 5.1.1์˜ ๋…ผ์˜์— ๋”ฐ๋ฅด๋ฉด, ์šฐ๋ฆฌ๋Š” ์ฃผ๋กœ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์„ธ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ ์ข…๋ฅ˜์˜ ๋ช…๋ น์–ด๋ฅผ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค:[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Instruction Datasets.** According to the discussion in Section 5.1.1, we mainly consider three common kinds of instructions as follows: [|assistant|]
\(\bullet\)_Alpaca_[142]๋Š” ๊ฐœ๋ฐฉํ˜• ์ƒ์„ฑ์„ ์œ„ํ•ด top-\(k\) (\(k=50\)), top-\(p\) (\(p=0.9\)), ์˜จ๋„ 0.7์˜ sampling ๊ธฐ๋ฐ˜ ์ „๋žต์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_์•ŒํŒŒ์นด_[142]๋Š” ๊ฐœ๋ฐฉํ˜• ์ƒ์„ฑ์„ ์œ„ํ•ด ์ƒ์œ„-\(k\) (\(k=50\)), ์ƒ์œ„-\(p\) (\(p=0.9\)), ์˜จ๋„ 0.7์˜ ์ƒ˜ํ”Œ๋ง ๊ธฐ๋ฐ˜ ์ „๋žต์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_Alpaca_[142] utilizes sampling-based strategies with top-\(k\) (\(k=50\)), top-\(p\) (\(p=0.9\)), and temperature of 0.7 for open-ended generation. [|assistant|]
๊ทธ๋ฆผ 3์€ ๋ฌธ์ œ์— ์‚ฐ๋งŒํ•œ ๊ทœ์น™์ด ํฌํ•จ๋˜์ง€ ์•Š์€ ground truth ์ฆ๋ช…์— ํฌํ•จ๋œ ๊ด€๋ จ ๊ทœ์น™์˜ ์ˆ˜๊ฐ€ ๋‹ค๋ฅธ ๊ฒฐ๊ณผ๋ฅผ ์ œ์‹œํ•˜๋ฉฐ, shuffled accuracy๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๊ฒฐ๊ณผ๋ฅผ ์ œ์‹œํ•œ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 3์€ ๋ฌธ์ œ์— ์‚ฐ๋งŒํ•œ ๊ทœ์น™์ด ํฌํ•จ๋˜์ง€ ์•Š์€ ๊ทธ๋ผ์šด๋“œ ํŠธ๋ฃจํ”„ ์ฆ๋ช…์— ํฌํ•จ๋œ ๊ด€๋ จ ๊ทœ์น™์˜ ์ˆ˜๊ฐ€ ๋‹ค๋ฅธ ๊ฒฐ๊ณผ๋ฅผ ์ œ์‹œํ•˜๋ฉฐ, ์…”ํ”Œ ์ •ํ™•๋„๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๊ฒฐ๊ณผ๋ฅผ ์ œ์‹œํ•œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 3 presents the results with different numbers of relevant rules included in ground truth proofs, where the problem does not contain distracting rules, and the shuffled accuracy is the aggregation of [|assistant|]
์ด ๋ฐฑ์„œ์—์„œ๋Š” _๋” ๋งŽ์€ ์—์ด์ „ํŠธ๊ฐ€ ํ•„์š”ํ•˜๋‹ค_๋Š” ๊ฒƒ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๋ณต์žกํ•œ ๋ฐฉ๋ฒ•(์˜ˆ: CoT ํŒŒ์ดํ”„๋ผ์ธ, multi-agent collaboration frameworks ๋“ฑ)์— ์‹ ๊ฒฝ ์“ฐ์ง€ ์•Š๊ณ ๋„ ์ธ์Šคํ„ด์Šคํ™”๋œ LLM ์—์ด์ „ํŠธ๋ฅผ ๋” ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ๋งŒ์œผ๋กœ๋„ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ๋•Œ ๋” ๋‚˜์€ LLM ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ €์ž๋“ค์€ ์ด๋Ÿฌํ•œ 'scaling law'๊ฐ€ ์–ธ์ œ ์ ์šฉ๋˜๋Š”์ง€, ๊ทธ๋ฆฌ๊ณ  ๊ทธ ๋ฐœ์ƒ์„ ์ด‰์ง„ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํฌํ•จํ•˜์—ฌ ์ด์— ๋Œ€ํ•œ ์ฒซ ๋ฒˆ์งธ ํฌ๊ด„์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ด ๋ฐฑ์„œ์—์„œ๋Š” _๋” ๋งŽ์€ ์—์ด์ „ํŠธ๊ฐ€ ํ•„์š”ํ•˜๋‹ค_๋Š” ๊ฒƒ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๋ณต์žกํ•œ ๋ฐฉ๋ฒ•(์˜ˆ: CoT ํŒŒ์ดํ”„๋ผ์ธ, ๋‹ค์ค‘ ์—์ด์ „ํŠธ ํ˜‘์—… ํ”„๋ ˆ์ž„์›Œํฌ ๋“ฑ)์— ์‹ ๊ฒฝ ์“ฐ์ง€ ์•Š๊ณ ๋„ ์ธ์Šคํ„ด์Šคํ™”๋œ LLM ์—์ด์ „ํŠธ๋ฅผ ๋” ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ๋งŒ์œผ๋กœ๋„ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ๋•Œ ๋” ๋‚˜์€ LLM ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ €์ž๋“ค์€ ์ด๋Ÿฌํ•œ '์Šค์ผ€์ผ๋ง ๋ฒ•์น™'์ด ์–ธ์ œ ์ ์šฉ๋˜๋Š”์ง€, ๊ทธ๋ฆฌ๊ณ  ๊ทธ ๋ฐœ์ƒ์„ ์ด‰์ง„ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํฌํ•จํ•˜์—ฌ ์ด์— ๋Œ€ํ•œ ์ฒซ ๋ฒˆ์งธ ํฌ๊ด„์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this paper, we report that _more agents is all you need_, i.e., simply adding more instantiated LLM agents is what you need to obtain a better LLM performance in processing complex tasks, without bothering complicated methods, such as CoT pipelines, multi-agent collaboration frameworks, etc. We have conducted the first comprehensive study in the literature to understand such a "scaling law", including when it holds and how to facilitate its occurrence. [|assistant|]
1. **Zero-Shot CoT:** ์ด ํ˜•์‹์€ LLM์—๊ฒŒ "์ฐจ๊ทผ์ฐจ๊ทผ ์ƒ๊ฐํ•˜๋ผ"๊ณ  ์ง€์‹œํ•˜์—ฌ ๋ฌธ์ œ๋ฅผ ํ•ด์ฒดํ•˜๊ณ  ์ถ”๋ก ์˜ ๊ฐ ๋‹จ๊ณ„๋ฅผ ๋ช…ํ™•ํžˆ ํ•˜๋„๋ก ์œ ๋„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. 2. **Manual CoT:** ๋” ๋ณต์žกํ•œ ๋ณ€ํ˜•์œผ๋กœ, ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ…œํ”Œ๋ฆฟ์œผ๋กœ ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก  ์˜ˆ์ œ๋ฅผ ์ œ๊ณตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋” ํšจ๊ณผ์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์ง€๋งŒ ํ™•์žฅ์„ฑ๊ณผ ์œ ์ง€ ๊ด€๋ฆฌ์— ์–ด๋ ค์›€์ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
1. **์ œ๋กœ ์ƒท CoT:** ์ด ํ˜•์‹์€ LLM์—๊ฒŒ "์ฐจ๊ทผ์ฐจ๊ทผ ์ƒ๊ฐํ•˜๋ผ"๊ณ  ์ง€์‹œํ•˜์—ฌ ๋ฌธ์ œ๋ฅผ ํ•ด์ฒดํ•˜๊ณ  ์ถ”๋ก ์˜ ๊ฐ ๋‹จ๊ณ„๋ฅผ ๋ช…ํ™•ํžˆ ํ•˜๋„๋ก ์œ ๋„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. 2. **์ˆ˜๋™ CoT:** ๋” ๋ณต์žกํ•œ ๋ณ€ํ˜•์œผ๋กœ, ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ…œํ”Œ๋ฆฟ์œผ๋กœ ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก  ์˜ˆ์ œ๋ฅผ ์ œ๊ณตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋” ํšจ๊ณผ์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์ง€๋งŒ ํ™•์žฅ์„ฑ๊ณผ ์œ ์ง€ ๊ด€๋ฆฌ์— ์–ด๋ ค์›€์ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]1. **Zero-Shot CoT:** This form involves instructing the LLM to "think step by step", prompting it to deconstruct the problem and articulate each stage of reasoning. 2. **Manual CoT:** A more complex variant, it requires providing step-by-step reasoning examples as templates for the model. While yielding more effective results, it poses challenges in scalability and maintenance. [|assistant|]
๊ฐ์ฃผ 12: ์ „๋ฌธ๊ฐ€ annotator์—๊ฒŒ C4 corpus ๊ฒ€์ƒ‰์„ ์œ„ํ•ด [https://c4-search.apps.allenai.org/](https://c4-search.apps.allenai.org/)์„ ์‚ฌ์šฉํ•˜๋„๋ก ์ง€์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ฐ์ฃผ 12: ์ „๋ฌธ๊ฐ€ ์ฃผ์„๊ฐ€์—๊ฒŒ C4 ์ฝ”ํผ์Šค ๊ฒ€์ƒ‰์„ ์œ„ํ•ด [https://c4-search.apps.allenai.org/](https://c4-search.apps.allenai.org/)์„ ์‚ฌ์šฉํ•˜๋„๋ก ์ง€์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Footnote 12: The expert annotators are instructed to use [https://c4-search.apps.allenai.org/](https://c4-search.apps.allenai.org/) for searching through the C4 corpus. [|assistant|]
# ์ถœ๋ ฅ 1. [label=()] 2. ํ”ผํฌ ์šฉ๋Ÿ‰ ๋น„์œจ \(R(F)\)๊ฐ€ ์ผ๊ด€๋˜๊ฒŒ \(R(F)\geq 2\)๋ฅผ ์ดˆ๊ณผํ•ฉ๋‹ˆ๋‹ค; 3. \(R^{\mathsf{max}}(F)\leq 1.8\)์ธ ๋ชจ๋ธ์€ ๊ฑฐ์˜ ์™„๋ฒฝํ•œ ์ง€์‹ ์ •ํ™•๋„, ์ฆ‰ \(R^{\mathsf{max}}(F)\approx R(F)\)๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค; 4. ๋ชจ๋“  ๋ชจ๋ธ์— ๊ฑธ์ณ \(R(F)\leq 2.3\).[|endofturn|]
1. [label=()] 2. ํ”ผํฌ ์šฉ๋Ÿ‰ ๋น„์œจ \(R(F)\)๊ฐ€ ์ผ๊ด€๋˜๊ฒŒ \(R(F)\geq 2\)๋ฅผ ์ดˆ๊ณผํ•ฉ๋‹ˆ๋‹ค; 3. R^{\mathsf{max}}(F)\leq 1.8\)์ธ ๋ชจ๋ธ์€ ๊ฑฐ์˜ ์™„๋ฒฝํ•œ ์ง€์‹ ์ •ํ™•๋„, ์ฆ‰ \(R^{\mathsf{max}}(F)\approx R(F)\)๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค; 4. ๋ชจ๋“  ๋ชจ๋ธ์— ๊ฑธ์ณ \(R(F)\leq 2.3\).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]1. [label=()] 2. the peak capacity ratio \(R(F)\) consistently exceeds \(R(F)\geq 2\); 3. models with \(R^{\mathsf{max}}(F)\leq 1.8\) attain near-perfect knowledge accuracies, i.e., \(R^{\mathsf{max}}(F)\approx R(F)\); 4. across all models, \(R(F)\leq 2.3\). [|assistant|]
์–‘์žํ™” ์†์‹ค์„ ์ค„์ด๊ธฐ ์œ„ํ•ด weight calibration์„ ์ ์šฉํ•˜๊ธฐ ์œ„ํ•ด GPTQ(Frantar et al., 2022)๋ฅผ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” SFT ๋ฐ์ดํ„ฐ์™€ ์œ ์‚ฌํ•œ ๋ถ„ํฌ์—์„œ calibration ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค. Quantization objective๋Š” ์–‘์žํ™”์˜ ๊ต๋ž€์„ ์ตœ์†Œํ™”ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(\|\mathbf{W}\mathbf{X}-dequant(\hat{\mathbf{W}})\mathbf{X}\|_{2}^{2}\).[|endofturn|]
์–‘์žํ™” ์†์‹ค์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ๊ฐ€์ค‘์น˜ ๋ณด์ •์„ ์ ์šฉํ•˜๊ธฐ ์œ„ํ•ด GPTQ(Frantar et al., 2022)๋ฅผ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ์–‘์žํ™” ๋ชฉํ‘œ๋Š” ์–‘์žํ™”์˜ ๊ต๋ž€์„ ์ตœ์†Œํ™”ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(\|\mathbf{W}\mathbf{X}-dequant(\hat{\mathbf{W}})\mathbf{X}\|_{2}^{2}\).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To reduce the quantization loss, we adopt GPTQ (Frantar et al., 2022) to apply weight calibration. We sample calibration data \(\mathbf{X}\) from a similar distribution of SFT data. The quantization objective is to minimize the disturbance of the quantization \(\|\mathbf{W}\mathbf{X}-dequant(\hat{\mathbf{W}})\mathbf{X}\|_{2}^{2}\). [|assistant|]
๊ฒฐ๊ณผ๊ฐ€ OpenAI API๋กœ ํ•™์Šต๋œ GPT-3 ๋ชจ๋ธ์—๋งŒ ๊ตญํ•œ๋˜์ง€ ์•Š๋„๋ก Llama-7b๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์Šค์œ•์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” Batch Size๋Š” 1, 4, 16, Learning Rate๋Š” 1e-06, 2e-06, 1e-05, 2e-05๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Optimizer๋กœ Adam์„ ์‚ฌ์šฉํ•˜๊ณ  ๋ฉ”๋ชจ๋ฆฌ ํšจ์œจ์„ฑ์„ ์œ„ํ•ด DeepSpeed ๋ ˆ๋ฒจ 3์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด Fine-tuning์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ๋งค๊ฐœ๋ณ€์ˆ˜ ํšจ์œจ์ ์ธ Fine-tuning ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 8์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ฒฐ๊ณผ๊ฐ€ OpenAI API๋กœ ํ•™์Šต๋œ GPT-3 ๋ชจ๋ธ์—๋งŒ ๊ตญํ•œ๋˜์ง€ ์•Š๋„๋ก Llama-7b๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์Šค์œ•์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 1, 4, 16, ํ•™์Šต ์†๋„๋Š” 1e-06, 2e-06, 1e-05, 2e-05๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ตํ‹ฐ๋งˆ์ด์ €๋กœ Adam์„ ์‚ฌ์šฉํ•˜๊ณ  ๋ฉ”๋ชจ๋ฆฌ ํšจ์œจ์„ฑ์„ ์œ„ํ•ด DeepSpeed ๋ ˆ๋ฒจ 3์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ๋ฏธ์„ธ ์กฐ์ •์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ๋งค๊ฐœ๋ณ€์ˆ˜ ํšจ์œจ์ ์ธ ๋ฏธ์„ธ ์กฐ์ • ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 8์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To ensure that our results are not specific to GPT-3 models trained with the OpenAI API, we also perform a hyperparameter sweep using Llama-7b. Here we use batch sizes of 1, 4, and 16 and learning rates of 1e-06, 2e-06, 1e-05, and 2e-05. We use Adam as our optimizer and DeepSpeed level 3 for memory efficiency. We perform full finetuning and do not use any parameter efficient finetuning techniques. The results are shown in Figure 8. [|assistant|]
์•ŒํŒŒ์นด ๋ชจ๋ธ์€ LLaMA์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. zero-shot ๋˜๋Š” 5-shot๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์„ค์ •์—์„œ ์•ŒํŒŒ์นด ๋ชจ๋ธ ์‹œ๋ฆฌ์ฆˆ๋Š” LLaMA ๋ชจ๋ธ์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, Instruction-following ๋ชจ๋ธ์ด ์ˆœ์ˆ˜ ์–ธ์–ด ๋ชจ๋ธ๋ณด๋‹ค ์ด๋Ÿฌํ•œ NLU์™€ ์œ ์‚ฌํ•œ ์ž‘์—…์„ ๋” ์ž˜ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. LLaMA ์‹œ๋ฆฌ์ฆˆ์—์„œ ๊ด€์ฐฐ๋œ ํ˜„์ƒ๊ณผ๋Š” ๋‹ฌ๋ฆฌ, ์•ŒํŒŒ์นด ํ”Œ๋Ÿฌ์Šค ๋ชจ๋ธ์ด ๊ธฐ๋ณธ ์•ŒํŒŒ์นด ๋ชจ๋ธ๋ณด๋‹ค ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ๊ฐ€์ ธ์˜จ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” Instruction-following ๋ชจ๋ธ์ด NLU์™€ ์œ ์‚ฌํ•œ ์ž‘์—…์„ ๋” ์ž˜ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ๋” ๋งŽ์€ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ(LLaMA-Plus)๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ ๋” ํฐ ํž˜์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
์•ŒํŒŒ์นด ๋ชจ๋ธ์€ LLaMA์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.์ œ๋กœ ์ƒท ๋˜๋Š” 5 ์ƒท๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์„ค์ •์—์„œ ์•ŒํŒŒ์นด ๋ชจ๋ธ ์‹œ๋ฆฌ์ฆˆ๋Š” LLaMA ๋ชจ๋ธ์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ง€์‹œ์–ด ์ถ”์ข… ๋ชจ๋ธ์ด ์ˆœ์ˆ˜ ์–ธ์–ด ๋ชจ๋ธ๋ณด๋‹ค ์ด๋Ÿฌํ•œ NLU์™€ ์œ ์‚ฌํ•œ ์ž‘์—…์„ ๋” ์ž˜ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. LLaMA ์‹œ๋ฆฌ์ฆˆ์—์„œ ๊ด€์ฐฐ๋œ ํ˜„์ƒ๊ณผ๋Š” ๋‹ฌ๋ฆฌ, ์•ŒํŒŒ์นด ํ”Œ๋Ÿฌ์Šค ๋ชจ๋ธ์ด ๊ธฐ๋ณธ ์•ŒํŒŒ์นด ๋ชจ๋ธ๋ณด๋‹ค ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ๊ฐ€์ ธ์˜จ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ง€์‹œ์–ด ์ถ”์ข… ๋ชจ๋ธ์ด NLU์™€ ์œ ์‚ฌํ•œ ์ž‘์—…์„ ๋” ์ž˜ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ๋” ๋งŽ์€ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ(LLaMA-Plus)๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ ๋” ํฐ ํž˜์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Alpaca models show significant improvements over LLaMA.Among different settings, such as zero-shot or 5-shot, the Alpaca model series show significant improvements over LLaMA counterparts, demonstrating that the instruction-following models are more capable of handling these NLU-like tasks than pure language models. Unlike the phenomenon observed in the LLaMA series, we can see that Alpaca-Plus models yield significant improvement over basic Alpaca models. This might further indicate that instruction-following models are more capable of handling NLU-like tasks and can unlesah the power of using more pre-training data (LLaMA-Plus). [|assistant|]
ํ‘œ 1๊ณผ ๊ฐ™์ด LM Evaluation Harness(Gao et al., 2023)์—์„œ ํ‰๊ฐ€์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋„ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ ํฌํ™” ํ˜„์ƒ์ด ๊ด€์ฐฐ๋œ๋‹ค.[|endofturn|]
ํ‘œ 1๊ณผ ๊ฐ™์ด LM ํ‰๊ฐ€ ํ•˜๋„ค์Šค(Gao et al., 2023)์—์„œ ํ‰๊ฐ€์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋„ ์œ ์‚ฌํ•œ ์„ฑ๋Šฅ ํฌํ™” ํ˜„์ƒ์ด ๊ด€์ฐฐ๋œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]A similar performance saturation is also observed on datasets used for evaluation in the LM Evaluation Harness (Gao et al., 2023), as shown in Table 1. [|assistant|]
OpenAI API๋ฅผ ํ†ตํ•ด Learning Rate ๋ฐฐ์ˆ˜ 0.05, 0.1, 0.2, 0.4์™€ Batch Size 1, 2, 4, 8, 16์˜ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์Šค์œ•์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด GPT-3-350M์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
OpenAI API๋ฅผ ํ†ตํ•ด ํ•™์Šต ์†๋„ ๋ฐฐ์ˆ˜ 0.05, 0.1, 0.2, 0.4์™€ ๋ฐฐ์น˜ ํฌ๊ธฐ 1, 2, 4, 8, 16์˜ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์Šค์œ•์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด GPT-3-350M์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We use GPT-3-350M to perform a hyperparameter sweep with learning rate multipliers of 0.05, 0.1, 0.2, and 0.4 and batch sizes of 1, 2, 4, 8, and 16 via the OpenAI API. [|assistant|]
# ์ถœ๋ ฅ 1. ๋ถˆ์•ˆ์ • ์•ˆ์ฆˆ๋‚˜ 2. ๊ธ‰์„ฑ ์‹ฌ๊ทผ๊ฒฝ์ƒ‰์ฆ 3. ์•ˆ์ • ์•ˆ์ฆˆ๋‚˜ 4. ์œ„์‹๋„ ์—ญ๋ฅ˜ ์งˆํ™˜(GERD) 5. ๋น„์ •ํ˜• ํ‰ํ†ต 6. ์—ฐ๊ณจ์—ผ 7. ์†Œํ™”์„ฑ ๊ถค์–‘์„ฑ ์งˆํ™˜ 8. ๋‹ด๋‚ญ ์งˆํ™˜(๋‹ด๊ด€์„ฑ ์ฝœ๋ฆญ) 9.[|endofturn|]
1. ๋ถˆ์•ˆ์ • ์•ˆ์ฆˆ๋‚˜2. ๊ธ‰์„ฑ ์‹ฌ๊ทผ๊ฒฝ์ƒ‰์ฆ3. ์•ˆ์ • ์•ˆ์ฆˆ๋‚˜4. ์œ„์‹๋„ ์—ญ๋ฅ˜ ์งˆํ™˜(GERD)5. ๋น„์ •ํ˜• ํ‰ํ†ต6. ์—ฐ๊ณจ์—ผ7. ์†Œํ™”์„ฑ ๊ถค์–‘์„ฑ ์งˆํ™˜8. ๋‹ด๋‚ญ ์งˆํ™˜(๋‹ด๊ด€์„ฑ ์ฝœ๋ฆญ)9.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]1. Unstable Anjna 2. Acute Myocardial Infarction 3. Stable Anjna 4. Gastroesophageal Reflux Disease (GERD) 5. Atypical Chest Pain 6. Costochonditis 7. Peptic Ulicer Disease 8. Gallbladder Disease (biliary colic) 9. [|assistant|]
RAG ํŒจ๋Ÿฌ๋‹ค์ž„์€ ๋„๋ฉ”์ธ ๋‚ด์—์„œ ์—ฐ๊ตฌ๋ฅผ ์กฐ์งํ™”ํ•˜์—ฌ LLM ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ๊ฐ•๋ ฅํ•œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. RAG์˜ ํ•ต์‹ฌ์€ ๊ณ ํ’ˆ์งˆ ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ๊ฒ€์ƒ‰ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ํŒจ๋Ÿฌ๋‹ค์ž„์€ ๊ฒ€์ƒ‰ ๊ด€์ ์—์„œ pre-retrieval, retrieval, post-retrieval, generation์˜ ๋„ค ๊ฐ€์ง€ ์ฃผ์š” ๋‹จ๊ณ„๋กœ ๊ตฌ๋ถ„๋ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ณต์ ์ธ ๊ฒ€์ƒ‰-์ƒ์„ฑ ์ฃผ๊ธฐ๋ฅผ ํฌํ•จํ•˜๋Š” single-hop ๋ฐ multi-hop ๊ฒ€์ƒ‰ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ด 4๋‹จ๊ณ„ ๊ตฌ์กฐ๋ฅผ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3์€ RAG์˜ ํ•ต์‹ฌ ๊ธฐ์ˆ ์˜ ๋ถ„๋ฅ˜ ํŠธ๋ฆฌ์ž…๋‹ˆ๋‹ค.[|endofturn|]
RAG ํŒจ๋Ÿฌ๋‹ค์ž„์€ ๋„๋ฉ”์ธ ๋‚ด์—์„œ ์—ฐ๊ตฌ๋ฅผ ์กฐ์งํ™”ํ•˜์—ฌ LLM ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ๊ฐ•๋ ฅํ•œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. RAG์˜ ํ•ต์‹ฌ์€ ๊ณ ํ’ˆ์งˆ ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ๊ฒ€์ƒ‰ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ํŒจ๋Ÿฌ๋‹ค์ž„์€ ๊ฒ€์ƒ‰ ๊ด€์ ์—์„œ ์‚ฌ์ „ ๊ฒ€์ƒ‰, ๊ฒ€์ƒ‰, ๊ฒ€์ƒ‰ ํ›„, ์ƒ์„ฑ์˜ ๋„ค ๊ฐ€์ง€ ์ฃผ์š” ๋‹จ๊ณ„๋กœ ๊ตฌ๋ถ„๋ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ณต์ ์ธ ๊ฒ€์ƒ‰-์ƒ์„ฑ ์ฃผ๊ธฐ๋ฅผ ํฌํ•จํ•˜๋Š” ๋‹จ์ผ ํ™‰ ๋ฐ ๋‹ค์ค‘ ํ™‰ ๊ฒ€์ƒ‰ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ด 4๋‹จ๊ณ„ ๊ตฌ์กฐ๋ฅผ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3์€ RAG์˜ ํ•ต์‹ฌ ๊ธฐ์ˆ ์˜ ๋ถ„๋ฅ˜ ํŠธ๋ฆฌ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The RAG paradigm organizes research within the domain, offering a straightforward yet robust framework to enhance LLM performance. Central to RAG is its search mechanism, crucial for generating high-quality outcomes. Therefore, this paradigm is structured into four main phases from a retrieval perspective: pre-retrieval, retrieval, post-retrieval, and generation. Both single-hop and multi-hop retrieval approaches, encompassing iterative retrieve-generate cycles, follow this four-phase structure. Figure 3 is the taxonomy tree of RAG's core techniques. [|assistant|]
* Tool Access ๋ฐ Utilization: ์—์ด์ „ํŠธ๋Š” ์™ธ๋ถ€ ๋„๊ตฌ์™€ ์„œ๋น„์Šค์— ์•ก์„ธ์Šคํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ™œ์šฉํ•˜์—ฌ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * Decision Making: ์ž…๋ ฅ, ์ปจํ…์ŠคํŠธ ๋ฐ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋„๊ตฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์˜์‚ฌ ๊ฒฐ์ •์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ข…์ข… ๋ณต์žกํ•œ reasoning ํ”„๋กœ์„ธ์Šค๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
* ๋„๊ตฌ ์•ก์„ธ์Šค ๋ฐ ํ™œ์šฉ: ์—์ด์ „ํŠธ๋Š” ์™ธ๋ถ€ ๋„๊ตฌ์™€ ์„œ๋น„์Šค์— ์•ก์„ธ์Šคํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ™œ์šฉํ•˜์—ฌ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * ์˜์‚ฌ ๊ฒฐ์ •: ์ž…๋ ฅ, ์ปจํ…์ŠคํŠธ ๋ฐ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋„๊ตฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์˜์‚ฌ ๊ฒฐ์ •์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ข…์ข… ๋ณต์žกํ•œ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* Tool Access and Utilization: Agents have the capability to access external tools and services, and to utilize these resources effectively to accomplish tasks. * Decision Making: They can make decisions based on the input, context, and the tools available to them, often employing complex reasoning processes. [|assistant|]
์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI ์‹œ์Šคํ…œ์€ ์ „ ์„ธ๊ณ„์˜ ๋‹ค์–‘ํ•œ ์‚ฌ๋žŒ, ๊ทธ๋ฃน, ์‚ฌํšŒ ๋ฐ ํ™˜๊ฒฝ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ ํ”ผํ•ด๋ฅผ ์ž…ํž ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[71]. ์—ฌ๊ธฐ์—๋Š” ์‹ ์ฒด์ , ์ •์„œ์ , ์žฌ์ •์ , ํ• ๋‹น์ , ํ‰ํŒ์ , ๋Œ€ํ‘œ์ , ์‹ฌ๋ฆฌ์  ํ”ผํ•ด๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค[72, 16, 73]. ์ด๋Ÿฌํ•œ ํ”ผํ•ด๋Š” generative AI ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•  ๋•Œ[74], generative AI ์‹œ์Šคํ…œ์—์„œ ์ œ์™ธ๋  ๋•Œ[75], generative AI ์‹œ์Šคํ…œ์— ์˜ํ•ด ํ‘œํ˜„๋˜๊ฑฐ๋‚˜ ์„ค๋ช…๋  ๋•Œ[76, 77], ๋˜๋Š” generative AI ์‹œ์Šคํ…œ์— ์˜ํ•ด ๋‚ด๋ ค์ง„ ๊ฒฐ์ •์— ๋…ธ์ถœ๋  ๋•Œ[78] ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ”ผํ•ด๋ฅผ ํ‰๊ฐ€ํ•  ๋•Œ ๊ณ ๋ คํ•ด์•ผ ํ•  ์ฃผ์š” ์‚ฌํ•ญ์œผ๋กœ๋Š” ํ”ผํ•ด๊ฐ€ ์œ ํ˜•์ ์ด๊ฑฐ๋‚˜ ๋ฌดํ˜•์ ์ธ์ง€ ์—ฌ๋ถ€, ํ”ผํ•ด์˜ ์ง€์† ์‹œ๊ฐ„์ด ๋‹จ๊ธฐ์ธ์ง€ ์žฅ๊ธฐ์ธ์ง€, ํ”ผํ•ด์˜ ์ •๋„๊ฐ€ ๋งค์šฐ ์‹ฌ๊ฐํ•œ์ง€ ๋˜๋Š” ๋œ ์‹ฌ๊ฐํ•œ์ง€, ์ž์‹ ์—๊ฒŒ ๊ฐ€ํ•ด์ง€๋Š”์ง€ ๋˜๋Š” ํƒ€์ธ์—๊ฒŒ ๊ฐ€ํ•ด์ง€๋Š”์ง€, ํ”ผํ•ด๊ฐ€ ๋‚ด๋ฉดํ™”๋˜๋Š”์ง€ ๋˜๋Š” ์™ธ๋ถ€ํ™”๋˜๋Š”์ง€ ์—ฌ๋ถ€ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค[79, 71, 80, 81]. ํ”ผํ•ด ๊ฒฝํ—˜์€ ํ”ผํ•ด๊ฐ€ ๊ฐ€ํ•ด์ง€๋Š” ๋งฅ๋ฝ์— ๋”ฐ๋ผ ํ˜•์„ฑ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ ๋‹ค์–‘ํ•œ ์œ„ํ—˜ ์š”์ธ์— ์˜ํ•ด ์˜ํ–ฅ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž์˜ ๋ฐฐ๊ฒฝ, ์ƒํ™œ ๊ฒฝํ—˜, ์„ฑ๊ฒฉ, ๊ณผ๊ฑฐ ํ–‰๋™๊ณผ ๊ฐ™์€ ์ธก๋ฉด์€ ๋ชจ๋‘ ํ”ผํ•ด ๊ฒฝํ—˜ ์—ฌ๋ถ€์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[82, 83, 84, 85].[|endofturn|]
์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI ์‹œ์Šคํ…œ์€ ์ „ ์„ธ๊ณ„์˜ ๋‹ค์–‘ํ•œ ์‚ฌ๋žŒ, ๊ทธ๋ฃน, ์‚ฌํšŒ ๋ฐ ํ™˜๊ฒฝ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ ํ”ผํ•ด๋ฅผ ์ž…ํž ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[71]. ์—ฌ๊ธฐ์—๋Š” ์‹ ์ฒด์ , ์ •์„œ์ , ์žฌ์ •์ , ํ• ๋‹น์ , ํ‰ํŒ์ , ๋Œ€ํ‘œ์ , ์‹ฌ๋ฆฌ์  ํ”ผํ•ด๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค[72, 16, 73]. ์ด๋Ÿฌํ•œ ํ”ผํ•ด๋Š” ์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•  ๋•Œ[74], ์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI ์‹œ์Šคํ…œ์—์„œ ์ œ์™ธ๋  ๋•Œ[75], ์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI ์‹œ์Šคํ…œ์— ์˜ํ•ด ํ‘œํ˜„๋˜๊ฑฐ๋‚˜ ์„ค๋ช…๋  ๋•Œ[76, 77], ๋˜๋Š” ์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI ์‹œ์Šคํ…œ์— ์˜ํ•ด ๋‚ด๋ ค์ง„ ๊ฒฐ์ •์— ๋…ธ์ถœ๋  ๋•Œ[78] ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ”ผํ•ด๋ฅผ ํ‰๊ฐ€ํ•  ๋•Œ ๊ณ ๋ คํ•ด์•ผ ํ•  ์ฃผ์š” ์‚ฌํ•ญ์œผ๋กœ๋Š” ํ”ผํ•ด๊ฐ€ ์œ ํ˜•์ ์ด๊ฑฐ๋‚˜ ๋ฌดํ˜•์ ์ธ์ง€ ์—ฌ๋ถ€, ํ”ผํ•ด์˜ ์ง€์† ์‹œ๊ฐ„์ด ๋‹จ๊ธฐ์ธ์ง€ ์žฅ๊ธฐ์ธ์ง€, ํ”ผํ•ด์˜ ์ •๋„๊ฐ€ ๋งค์šฐ ์‹ฌ๊ฐํ•œ์ง€ ๋˜๋Š” ๋œ ์‹ฌ๊ฐํ•œ์ง€, ์ž์‹ ์—๊ฒŒ ๊ฐ€ํ•ด์ง€๋Š”์ง€ ๋˜๋Š” ํƒ€์ธ์—๊ฒŒ ๊ฐ€ํ•ด์ง€๋Š”์ง€, ํ”ผํ•ด๊ฐ€ ๋‚ด๋ฉดํ™”๋˜๋Š”์ง€ ๋˜๋Š” ์™ธ๋ถ€ํ™”๋˜๋Š”์ง€ ์—ฌ๋ถ€ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค[79, 71, 80, 81]. ํ”ผํ•ด ๊ฒฝํ—˜์€ ํ”ผํ•ด๊ฐ€ ๊ฐ€ํ•ด์ง€๋Š” ๋งฅ๋ฝ์— ๋”ฐ๋ผ ํ˜•์„ฑ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ ๋‹ค์–‘ํ•œ ์œ„ํ—˜ ์š”์ธ์— ์˜ํ•ด ์˜ํ–ฅ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž์˜ ๋ฐฐ๊ฒฝ, ์ƒํ™œ ๊ฒฝํ—˜, ์„ฑ๊ฒฉ, ๊ณผ๊ฑฐ ํ–‰๋™๊ณผ ๊ฐ™์€ ์ธก๋ฉด์€ ๋ชจ๋‘ ํ”ผํ•ด ๊ฒฝํ—˜ ์—ฌ๋ถ€์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[82, 83, 84, 85].[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Generative AI systems have the potential to cause harm in myriad ways, affecting different people, groups, societies and environments across the globe [71]. This includes physical, emotional, financial, allocative, reputational, representational, and psychological harms [72, 16, 73]. Such harms can be caused by using generative AI systems [74], being excluded from them [75], being represented or described by them [76, 77], or being subjected to decisions made by them [78]. Key considerations when assessing harm include whether the harm is tangible or intangible, short- or long-term in duration, highly severe or less severe in nature, inflicted on oneself or on others, or internalized or externalized in its expression [79, 71, 80, 81]. Experiences of harm are often shaped by the context in which the harm is inflicted and can be affected by a range of risk factors. Aspects like the users' background, life experiences, personality, and past behavior can all impact whether they experience harm [82, 83, 84, 85]. [|assistant|]
๊ฐ์ฃผ 24: Model parallelism[75]๋Š” tensor parallelism๊ณผ pipeline parallelism์„ ํฌํ•จํ•˜๋Š” ๋ณด๋‹ค ๊ด‘๋ฒ”์œ„ํ•œ ์šฉ์–ด์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ฐ์ฃผ 24: ๋ชจ๋ธ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ[75]๋Š” ํ…์„œ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ์™€ ํŒŒ์ดํ”„๋ผ์ธ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๋ฅผ ํฌํ•จํ•˜๋Š” ๋ณด๋‹ค ๊ด‘๋ฒ”์œ„ํ•œ ์šฉ์–ด์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Footnote 24: Model parallelism is a more broader term that includes tensor parallelism and pipeline parallelism in some work [75]. [|assistant|]
์ถ”๊ฐ€ BC ์ •๊ทœํ™”๊ธฐ๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ DPO Loss \(\mathcal{L}_{\text{DPO}}=\mathcal{L}(\pi_{\theta},\pi_{\text{ref}})\)๋„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค:[|endofturn|]
์ถ”๊ฐ€ BC ์ •๊ทœํ™”๊ธฐ๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ DPO ์†์‹ค \(\mathcal{L}_{\text{DPO}}=\mathcal{L}(\pi_{\theta},\pi_{\text{ref}})\)๋„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค:[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The DPO loss \(\mathcal{L}_{\text{DPO}}=\mathcal{L}(\pi_{\theta},\pi_{\text{ref}})\) can also be utilized by adding our additional BC regularizer: [|assistant|]
# ์ถœ๋ ฅ ์—ฌ๊ธฐ์„œ \(x_{i}\)๋Š” ๋ฌธ๋งฅ \(\mathbf{x}_{<i}\)์— ์กฐ๊ฑด์„ ๋ถ™์ธ ์ƒ์„ฑ์˜ \(i\)-๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ ๊ฐ€์žฅ ๋†’์€ ํ™•๋ฅ ์„ ๊ฐ–๋Š” token์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ๊ทธ๋ฆผ 10์—์„œ _"๋‚˜๋Š” ์กธ๋ฆฐ๋‹ค. ๋‚˜๋Š” ๋ƒ„๋น„๋ฅผ ์‹œ์ž‘ํ•œ๋‹ค."_ ๋ฌธ์žฅ์˜ ๋‹ค์Œ token์„ ์˜ˆ์ธกํ•  ๋•Œ greedy search๋Š” ํ˜„์žฌ ๋‹จ๊ณ„์—์„œ ๊ฐ€์žฅ ๋†’์€ ํ™•๋ฅ ์„ ๊ฐ–๋Š” token "coffee"๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. Greedy search๋Š” ์ถœ๋ ฅ์ด ์ž…๋ ฅ์— ํฌ๊ฒŒ ์˜์กดํ•˜๋Š” ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…(_์˜ˆ:_ machine translation ๋ฐ ํ…์ŠคํŠธ ์š”์•ฝ)์—์„œ ๋งŒ์กฑ์Šค๋Ÿฌ์šด ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[307]. ๊ทธ๋Ÿฌ๋‚˜ ๊ฐœ๋ฐฉํ˜• ์ƒ์„ฑ ์ž‘์—…(_์˜ˆ:_ ์Šคํ† ๋ฆฌ ์ƒ์„ฑ ๋ฐ ๋Œ€ํ™”)์˜ ๊ฒฝ์šฐ greedy search๋Š” ๋•Œ๋•Œ๋กœ ์–ด์ƒ‰ํ•˜๊ณ  ๋ฐ˜๋ณต์ ์ธ ๋ฌธ์žฅ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค[308].[|endofturn|]
์—ฌ๊ธฐ์„œ \(x_{i}\)๋Š” ๋ฌธ๋งฅ \mathbf{x}_{<i}\)์— ์กฐ๊ฑด์„ ๋ถ™์ธ ์ƒ์„ฑ์˜ \(i\)-๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ ๊ฐ€์žฅ ๋†’์€ ํ™•๋ฅ ์„ ๊ฐ–๋Š” ํ† ํฐ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ๊ทธ๋ฆผ 10์—์„œ _"๋‚˜๋Š” ์กธ๋ฆฐ๋‹ค. ๋‚˜๋Š” ๋ƒ„๋น„๋ฅผ ์‹œ์ž‘ํ•œ๋‹ค."_ ๋ฌธ์žฅ์˜ ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•  ๋•Œ ์š•์‹ฌ ๋งŽ์€ ๊ฒ€์ƒ‰์€ ํ˜„์žฌ ๋‹จ๊ณ„์—์„œ ๊ฐ€์žฅ ๋†’์€ ํ™•๋ฅ ์„ ๊ฐ–๋Š” ํ† ํฐ "coffee"๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์š•์‹ฌ ๋งŽ์€ ๊ฒ€์ƒ‰์€ ์ถœ๋ ฅ์ด ์ž…๋ ฅ์— ํฌ๊ฒŒ ์˜์กดํ•˜๋Š” ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…(_์˜ˆ:_ ๊ธฐ๊ณ„ ๋ฒˆ์—ญ ๋ฐ ํ…์ŠคํŠธ ์š”์•ฝ)์—์„œ ๋งŒ์กฑ์Šค๋Ÿฌ์šด ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[307]. ๊ทธ๋Ÿฌ๋‚˜ ๊ฐœ๋ฐฉํ˜• ์ƒ์„ฑ ์ž‘์—…(_์˜ˆ:_ ์Šคํ† ๋ฆฌ ์ƒ์„ฑ ๋ฐ ๋Œ€ํ™”)์˜ ๊ฒฝ์šฐ ์š•์‹ฌ ๋งŽ์€ ๊ฒ€์ƒ‰์€ ๋•Œ๋•Œ๋กœ ์–ด์ƒ‰ํ•˜๊ณ  ๋ฐ˜๋ณต์ ์ธ ๋ฌธ์žฅ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค[308].[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]where \(x_{i}\) is the token with the highest probability at \(i\)-th step of generation conditioned on the context \(\mathbf{x}_{<i}\). For instance in Figure 10, when predicting the next token of the sentence _"I am sleepy. I start a pot of"_, greedy search selects the token "coffee" which has the highest probability at the current step. Greedy search can achieve satisfactory results in text generation tasks (_e.g.,_ machine translation and text summarization), in which the output is highly dependent on the input [307]. However, in terms of open-ended generation tasks (_e.g.,_ story generation and dialog), greedy search sometimes tends to generate awkward and repetitive sentences [308]. [|assistant|]
์ฃผ์„ ๋‹ฌ๊ธฐ ๋ชฉ์ ์ฑ„ํŒ…์— ์‚ฌ์šฉ๋˜๋Š” ๋‹ค์–‘ํ•œ AI ๋ชจ๋ธ์˜ ์‘๋‹ต 2,000๊ฑด์— ์ฃผ์„์„ ๋‹ฌ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AI ๋ชจ๋ธ์˜ ์•ˆ์ „์„ฑ ์œ„ํ—˜์„ ๋ฒค์น˜๋งˆํ‚นํ•˜๊ธฐ ์œ„ํ•œ ML Commons์˜ AI ์•ˆ์ „ ๋…ธ๋ ฅ์— ๋Œ€ํ•œ ์ผํ™˜์ž…๋‹ˆ๋‹ค. ์ €ํฌ์˜ ํ™œ๋™์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์•Œ์•„๋ณด์‹œ๊ณ  ๊ถ๊ธˆํ•œ ์ ์ด ์žˆ์œผ๋ฉด Bertie์—๊ฒŒ ๋ฌธ์˜ํ•ด ์ฃผ์„ธ์š”.[|endofturn|]
์ฃผ์„ ๋‹ฌ๊ธฐ ๋ชฉ์ ์ฑ„ํŒ…์— ์‚ฌ์šฉ๋˜๋Š” ๋‹ค์–‘ํ•œ AI ๋ชจ๋ธ์˜ ์‘๋‹ต 2,000๊ฑด์— ์ฃผ์„์„ ๋‹ฌ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AI ๋ชจ๋ธ์˜ ์•ˆ์ „์„ฑ ์œ„ํ—˜์„ ๋ฒค์น˜๋งˆํ‚นํ•˜๊ธฐ ์œ„ํ•œ ML ์ปค๋จผ์ฆˆ์˜ AI ์•ˆ์ „ ๋…ธ๋ ฅ์— ๋Œ€ํ•œ ์ผํ™˜์ž…๋‹ˆ๋‹ค. ์ €ํฌ์˜ ํ™œ๋™์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์•Œ์•„๋ณด์‹œ๊ณ  ๊ถ๊ธˆํ•œ ์ ์ด ์žˆ์œผ๋ฉด Bertie์—๊ฒŒ ๋ฌธ์˜ํ•ด ์ฃผ์„ธ์š”.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Purpose of annotationWe are annotating 2,000 responses from a range of AI models used for chat. This is part of the ML Commons' Working Group on AI Safety efforts to benchmark the safety risks of AI models. You can find out more about our work HERE, and reach out to Bertie if you have questions. [|assistant|]
6. \(f\)์™€ \(g\)์˜ ๋˜ ๋‹ค๋ฅธ ์ธ์Šคํ„ด์Šค๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. \(f\)๋Š” ํ™€์ˆ˜์ด๋ฏ€๋กœ ๊ทธ ๊ณฑ์€ ํ™€์ˆ˜ ํ•จ์ˆ˜์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹์„ \(-[f(g(f(x)))]\)๋กœ ๋‹ค์‹œ ์“ธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
6. f\)์™€ \(g\)์˜ ๋˜ ๋‹ค๋ฅธ ์ธ์Šคํ„ด์Šค๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. f\)๋Š” ํ™€์ˆ˜์ด๋ฏ€๋กœ ๊ทธ ๊ณฑ์€ ํ™€์ˆ˜ ํ•จ์ˆ˜์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹์„ \(-[f(g(f(x)))]\)๋กœ ๋‹ค์‹œ ์“ธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]6. We have another instance of \(f\) and \(g\). Since \(f\) is odd, their product is an odd function. So, we can rewrite the expression as \(-[f(g(f(x)))]\). [|assistant|]
LLM์€ ์‚ฌ์ „ ํ•™์Šต๊ณผ Instruction Tuning์„ ํ†ตํ•ด ์–ธ์–ด ์ƒ์„ฑ ๋ฐ ์ง€์นจ ์ถ”์ข… ๊ธฐ๋Šฅ์„ ์Šต๋“ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์˜์–ด๋Š” ๋‹ค์–‘ํ•œ ์˜์—ญ์˜ ๊ฐ€์žฅ ํ’๋ถ€ํ•œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ๋ณด์œ ํ•˜๊ณ  ์žˆ๋Š” ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ ์ง€๋ฐฐ์ ์ธ ์œ„์น˜๋ฅผ ์ฐจ์ง€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜์–ด๊ฐ€ ์ง€๋ฐฐ์ ์ธ ์ฝ”ํผ๋ผ๋กœ ํ•™์Šต๋œ LLM์€ ๋‹ค๋ฅธ ๋น„์˜์–ด ์–ธ์–ด์— ๋Œ€ํ•ด ์—ด๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. LLM์„ ๋น„์˜์–ด ์–ธ์–ด๋กœ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ๊ฐ€์น˜ ์žˆ๋Š” ์—ฐ๊ตฌ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ์™ธ์‚ฝ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋‹ค์Œ ์„ธ ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: (1) ๋Œ€์ƒ ์–ธ์–ด์˜ ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜์—ฌ ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜๊ณ , ์ด๋ฅผ ํ†ตํ•ด ํ•ด๋‹น ์–ธ์–ด์— ๋Œ€ํ•œ ์ธ์ฝ”๋”ฉ ํ‘œํ˜„๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. (2) LLM์˜ ์–ธ์–ด ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ๋Œ€์ƒ ์–ธ์–ด๋กœ ์ „์†กํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ์‚ฌ์ „ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์— ํ•„์š”ํ•œ ํ›ˆ๋ จ ๊ทœ๋ชจ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ํ† ํฐ ์ˆ˜์ค€์œผ๋กœ, ์ฒ˜์Œ๋ถ€ํ„ฐ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ์ˆ˜์กฐ ๊ฐœ์˜ ํ† ํฐ๋ณด๋‹ค ํ›จ์”ฌ ์ ์Šต๋‹ˆ๋‹ค. (3) ๋Œ€์ƒ ์–ธ์–ด๋กœ SFT๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ LLM์˜ ์ง€์นจ ์ถ”์ข… ๊ธฐ๋Šฅ์„ ์ „์†กํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
LLM์€ ์‚ฌ์ „ ํ•™์Šต๊ณผ ์ง€์นจ ํŠœ๋‹์„ ํ†ตํ•ด ์–ธ์–ด ์ƒ์„ฑ ๋ฐ ์ง€์นจ ์ถ”์ข… ๊ธฐ๋Šฅ์„ ์Šต๋“ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์˜์–ด๋Š” ๋‹ค์–‘ํ•œ ์˜์—ญ์˜ ๊ฐ€์žฅ ํ’๋ถ€ํ•œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ๋ณด์œ ํ•˜๊ณ  ์žˆ๋Š” ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ ์ง€๋ฐฐ์ ์ธ ์œ„์น˜๋ฅผ ์ฐจ์ง€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜์–ด๊ฐ€ ์ง€๋ฐฐ์ ์ธ ์ฝ”ํผ๋ผ๋กœ ํ•™์Šต๋œ LLM์€ ๋‹ค๋ฅธ ๋น„์˜์–ด ์–ธ์–ด์— ๋Œ€ํ•ด ์—ด๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. LLM์„ ๋น„์˜์–ด ์–ธ์–ด๋กœ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ๊ฐ€์น˜ ์žˆ๋Š” ์—ฐ๊ตฌ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ์™ธ์‚ฝ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋‹ค์Œ ์„ธ ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: (1) ๋Œ€์ƒ ์–ธ์–ด์˜ ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜์—ฌ ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜๊ณ , ์ด๋ฅผ ํ†ตํ•ด ํ•ด๋‹น ์–ธ์–ด์— ๋Œ€ํ•œ ์ธ์ฝ”๋”ฉ ํ‘œํ˜„๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. (2) LLM์˜ ์–ธ์–ด ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ๋Œ€์ƒ ์–ธ์–ด๋กœ ์ „์†กํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ์‚ฌ์ „ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์— ํ•„์š”ํ•œ ํ›ˆ๋ จ ๊ทœ๋ชจ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ํ† ํฐ ์ˆ˜์ค€์œผ๋กœ, ์ฒ˜์Œ๋ถ€ํ„ฐ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ์ˆ˜์กฐ ๊ฐœ์˜ ํ† ํฐ๋ณด๋‹ค ํ›จ์”ฌ ์ ์Šต๋‹ˆ๋‹ค. (3) ๋Œ€์ƒ ์–ธ์–ด๋กœ SFT๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ LLM์˜ ์ง€์นจ ์ถ”์ข… ๊ธฐ๋Šฅ์„ ์ „์†กํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]LLMs acquire language generation and instruction-following capabilities through pretraining and instruction tuning. However, English holds a dominant position in the field of natural language processing, possessing the most abundant collection of text data from various domains. LLMs trained on English-dominant corpora exhibit inferior performance on other non-English languages. Extrapolating LLMs to non-English languages poses a highly valuable research challenge. Common extrapolation approaches consist of the following three steps: (1) extending the vocabulary to add tokens of the target language, and thus enhancing encoding expressiveness to that language. (2) further pretraining to transfer language generation capabilities of LLMs to the target language. The required training scale for this step is generally on the order of billions of tokens, significantly less than the trillions of tokens needed for training from scratch. (3) conducting SFT in the target language to transfer instruction-following capabilities of LLMs. [|assistant|]
์ค‘์š”์„ฑ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LM์˜ ๋งค๊ฐœ๋ณ€์ˆ˜์— ์ €์žฅ๋œ ๋‚ด๋ถ€ ์„ธ๊ณ„ ์ง€์‹์„ ๊ฐฑ์‹ ํ•˜๋Š” ๊ณผ์ œ๋Š” ๊ฐ„๋‹จํ•˜์ง€ ์•Š์œผ๋ฉฐ, ๋‹ค์†Œ ํŠน์ •ํ•œ ์„ค์ •์—์„œ๋งŒ ์—ฐ๊ตฌ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฐœ๋ณ„ ์‚ฌ์‹ค๊ณผ ๊ฐ™์€ ํŠน์ • ๋Œ€์ƒ ์ง€์‹์„ ์ˆ˜์ •ํ•  ๊ฒƒ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค(De Cao et al., 2021; Zhu et al., 2020; Dai et al., 2021). Dhingra et al. (2021)์€ ํƒ€์ž„์Šคํƒฌํ”„๊ฐ€ ์žˆ๋Š” ํ…์ŠคํŠธ๋ฅผ ๊ณต๋™์œผ๋กœ ๋ชจ๋ธ๋งํ•˜์—ฌ LM์„ ์‹œ๊ฐ„์  ์ง€์‹ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ƒˆ๋กœ์šด ์ง€์‹์ด ํฌํ•จ๋œ ๋ง๋ญ‰์น˜์— ๋Œ€ํ•œ ์ง€์†์ ์ธ pretraining์„ ํ†ตํ•ด ๋ณด๋‹ค ์ผ๋ฐ˜์ ์ด๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ฐฉ์‹์œผ๋กœ LM์˜ ์„ธ๊ณ„ ์ง€์‹์„ ๊ฐฑ์‹ ํ•˜๋Š” ๋ฌธ์ œ๋Š” ์ด์ „ ์—ฐ๊ตฌ์—์„œ ๊ณต์‹์ ์œผ๋กœ ๊ณต์‹ํ™”๋˜๊ฑฐ๋‚˜ ํƒ๊ตฌ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ƒˆ๋กœ์šด ์ •๋ณด์— ๋Œ€ํ•œ ํ›ˆ๋ จ์„ ํ†ตํ•ด LM์˜ ๋‚ด๋ถ€ ์ง€์‹์ด ์–ด๋–ป๊ฒŒ ๋ณ€ํ™”ํ•˜๋Š”์ง€ ์ฒด๊ณ„์ ์œผ๋กœ ์—ฐ๊ตฌํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ๊ฐ€ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๊ทœ๋ชจ๊ฐ€ ํฐ LM์˜ ์ง€์‹์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ฐฑ์‹ ํ•˜๋Š” ๋ฐฉ๋ฒ•๋ก ์€ ์•„์ง ์ฒ ์ €ํžˆ ํƒ๊ตฌ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ค‘์š”์„ฑ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LM์˜ ๋งค๊ฐœ๋ณ€์ˆ˜์— ์ €์žฅ๋œ ๋‚ด๋ถ€ ์„ธ๊ณ„ ์ง€์‹์„ ๊ฐฑ์‹ ํ•˜๋Š” ๊ณผ์ œ๋Š” ๊ฐ„๋‹จํ•˜์ง€ ์•Š์œผ๋ฉฐ, ๋‹ค์†Œ ํŠน์ •ํ•œ ์„ค์ •์—์„œ๋งŒ ์—ฐ๊ตฌ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฐœ๋ณ„ ์‚ฌ์‹ค๊ณผ ๊ฐ™์€ ํŠน์ • ๋Œ€์ƒ ์ง€์‹์„ ์ˆ˜์ •ํ•  ๊ฒƒ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค(De Cao et al., 2021; Zhu et al., 2020; Dai et al., 2021). Dhingra et al. (2021)์€ ํƒ€์ž„์Šคํƒฌํ”„๊ฐ€ ์žˆ๋Š” ํ…์ŠคํŠธ๋ฅผ ๊ณต๋™์œผ๋กœ ๋ชจ๋ธ๋งํ•˜์—ฌ LM์„ ์‹œ๊ฐ„์  ์ง€์‹ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ƒˆ๋กœ์šด ์ง€์‹์ด ํฌํ•จ๋œ ๋ง๋ญ‰์น˜์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์ „์ฒ˜๋ฆฌ๋ฅผ ํ†ตํ•ด ๋ณด๋‹ค ์ผ๋ฐ˜์ ์ด๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ๋ฐฉ์‹์œผ๋กœ LM์˜ ์„ธ๊ณ„ ์ง€์‹์„ ๊ฐฑ์‹ ํ•˜๋Š” ๋ฌธ์ œ๋Š” ์ด์ „ ์—ฐ๊ตฌ์—์„œ ๊ณต์‹์ ์œผ๋กœ ๊ณต์‹ํ™”๋˜๊ฑฐ๋‚˜ ํƒ๊ตฌ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ƒˆ๋กœ์šด ์ •๋ณด์— ๋Œ€ํ•œ ํ›ˆ๋ จ์„ ํ†ตํ•ด LM์˜ ๋‚ด๋ถ€ ์ง€์‹์ด ์–ด๋–ป๊ฒŒ ๋ณ€ํ™”ํ•˜๋Š”์ง€ ์ฒด๊ณ„์ ์œผ๋กœ ์—ฐ๊ตฌํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ๊ฐ€ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๊ทœ๋ชจ๊ฐ€ ํฐ LM์˜ ์ง€์‹์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ฐฑ์‹ ํ•˜๋Š” ๋ฐฉ๋ฒ•๋ก ์€ ์•„์ง ์ฒ ์ €ํžˆ ํƒ๊ตฌ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Despite its importance, the challenge of renewing the internal world knowledge stored in the parameters of LMs is nontrivial and has only been explored in rather specific settings. For example, recent works have proposed to modify specific target knowledge such as individual facts (De Cao et al., 2021; Zhu et al., 2020; Dai et al., 2021). Dhingra et al. (2021) have addressed LMs as temporal knowledge bases by jointly modeling text with its timestamp. But the problem of renewing the world knowledge of LMs in a more general and scalable way, such as through continual pretraining on a corpus with new knowledge, has not been formally formulated or explored by previous works. Moreover, the community lacks a benchmark that can be used to systematically study how the internal knowledge of LMs changes through the training on new information. Lastly, methodologies to effectively renew the knowledge of LMs at scale have yet to be thoroughly explored. [|assistant|]
๋ชจ๋ธ ์ถœ๋ ฅ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ prompt๋ฅผ GPT-4์— benchmark๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด prompt๋ฅผ ๋Œ€์ƒ ์–ธ์–ด์— ๋งž๊ฒŒ ๋ฒˆ์—ญํ•˜๊ณ  {๋Œ€์ƒ ์–ธ์–ด}๋ฅผ ์ ์ ˆํžˆ ์กฐ์ •ํ•˜์—ฌ ๋Œ€์ƒ ์–ธ์–ด์— ๋งž๊ฒŒ ์กฐ์ •ํ•œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋ชจ๋ธ ์ถœ๋ ฅ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ ํ”„๋กฌํ”„ํŠธ๋ฅผ GPT-4์— ๋ฒค์น˜๋งˆํฌ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋Œ€์ƒ ์–ธ์–ด์— ๋งž๊ฒŒ ๋ฒˆ์—ญํ•˜๊ณ  {๋Œ€์ƒ ์–ธ์–ด}๋ฅผ ์ ์ ˆํžˆ ์กฐ์ •ํ•˜์—ฌ ๋Œ€์ƒ ์–ธ์–ด์— ๋งž๊ฒŒ ์กฐ์ •ํ•œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To assess the quality of the model's output, we use the following prompt to GPT-4 as a benchmark. We note that we adapt this prompt to the target language by translating it and adjusting the {target language} accordingly. [|assistant|]
LayerNorm ์—ฐ์‚ฐ์€ ์ค‘์‹ฌ ์—ฐ์‚ฐ \(x\mapsto x-\bar{x}\)์™€ RMSNorm(์ฆ‰, ๋จผ์ € ์ค‘์‹ฌ์„ ์ ์šฉํ•œ ๋‹ค์Œ RMSNorm)์˜ ๊ตฌ์„ฑ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ถ€๋ก C.4์™€ ๋™์ผํ•œ ์ธ์ˆ˜๊ฐ€ ์ ์šฉ๋˜๋ฉฐ, ์ค‘์‹ฌ ์—ฐ์‚ฐ๊ณผ ํ†ต๊ทผํ•˜๋Š” \(\mathbf{U}\)๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ์‚ฌ์‹ค ์™ธ์—๋Š” ๋ถ€๋ก C.4์™€ ๋™์ผํ•œ ์ธ์ˆ˜๊ฐ€ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ค‘์‹ฌ ์—ฐ์‚ฐ์€ \(\mathbf{1}^{T}x=0\)์œผ๋กœ ์ •์˜๋˜๋Š” \((h-1)\) ์ฐจ์›์˜ ๋ถ€๋ถ„ ๊ณต๊ฐ„์„ ์ˆ˜์ •ํ•˜๋ฏ€๋กœ, \(\mathbf{U}\,\mathbf{1}\in\{-\mathbf{1},\mathbf{1}\}\}\)๋ผ๋Š” ์ถ”๊ฐ€ ์กฐ๊ฑด์„ ๋ถ€๊ณผํ•˜๋ฉด ์ถฉ๋ถ„ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
LayerNorm ์—ฐ์‚ฐ์€ ์ค‘์‹ฌ ์—ฐ์‚ฐ \(x\mapsto x-\bar{x}\)์™€ RMSNorm(์ฆ‰, ๋จผ์ € ์ค‘์‹ฌ์„ ์ ์šฉํ•œ ๋‹ค์Œ RMSNorm)์˜ ๊ตฌ์„ฑ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ถ€๋ก C.4์™€ ๋™์ผํ•œ ์ธ์ˆ˜๊ฐ€ ์ ์šฉ๋˜๋ฉฐ, ์ค‘์‹ฌ ์—ฐ์‚ฐ๊ณผ ํ†ต๊ทผํ•˜๋Š” \(\mathbf{U}\)๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ์‚ฌ์‹ค ์™ธ์—๋Š” ๋ถ€๋ก C.4์™€ ๋™์ผํ•œ ์ธ์ˆ˜๊ฐ€ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ค‘์‹ฌ ์—ฐ์‚ฐ์€ \(\mathbf{1}^{T}x=0\)์œผ๋กœ ์ •์˜๋˜๋Š” \((h-1)\) ์ฐจ์›์˜ ๋ถ€๋ถ„ ๊ณต๊ฐ„์„ ์ˆ˜์ •ํ•˜๋ฏ€๋กœ, \(\mathbf{U}\,\mathbf{1}\in\{-\mathbf{1},\mathbf{1}\}\)๋ผ๋Š” ์ถ”๊ฐ€ ์กฐ๊ฑด์„ ๋ถ€๊ณผํ•˜๋ฉด ์ถฉ๋ถ„ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The LayerNorm operation is the composition of a centering operation \(x\mapsto x-\bar{x}\) with RMSNorm (i.e. first centering is applied, then RMSNorm). Therefore the identical argument to Appendix C.4 goes through, besides the fact that we need \(\mathbf{U}\) to also commute with the centering operation. Since the centering operation fixes a \((h-1)\) dimensional subspace defined by \(\mathbf{1}^{T}x=0\) where \(\mathbf{1}\in\mathds{R}^{h}\) is the vector of ones, it is enough to impose an additional condition that \(\mathbf{U}\,\mathbf{1}\in\{-\mathbf{1},\mathbf{1}\}\). [|assistant|]
๊ฒฐ๊ณผ๋Š” ํ‘œ 10์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ๊ตญ๊ฐ€์™€ ๊ตญ์ ์ด ํŠนํžˆ ์„ฑ๊ณต์ ์ธ ๋ฐ˜๋ฉด, ๊ฐœ๋ณ„ ๋‚ ์งœ์™€ ์—ฐ๋„๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฐ˜์ „ ๊ณผ์ •์—์„œ ์†์‹ค๋˜๋Š” ๋“ฑ ์ผ๋ถ€ ๊ฐœ์ธ ์—”ํ‹ฐํ‹ฐ๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋ฐ ํ›จ์”ฌ ๋” ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ์ž‘์—…์—์„œ๋Š” Instructions-2M์— ํ”„๋กฌํ”„ํŠธ์— ๊ฐœ์ธ ์ •๋ณด๊ฐ€ ์ข์€ ๋ถ„ํฌ๋กœ๋งŒ ํฌํ•จ๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ์œ„ํ•ด ๋” ๊ด‘๋ฒ”์œ„ํ•œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ฒฐ๊ณผ๋Š” ํ‘œ 10์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ๊ตญ๊ฐ€์™€ ๊ตญ์ ์ด ํŠนํžˆ ์„ฑ๊ณต์ ์ธ ๋ฐ˜๋ฉด, ๊ฐœ๋ณ„ ๋‚ ์งœ์™€ ์—ฐ๋„๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฐ˜์ „ ๊ณผ์ •์—์„œ ์†์‹ค๋˜๋Š” ๋“ฑ ์ผ๋ถ€ ๊ฐœ์ธ ์—”ํ‹ฐํ‹ฐ๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋ฐ ํ›จ์”ฌ ๋” ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„ ์ž‘์—…์—์„œ๋Š” ์ง€์นจ-2M์— ํ”„๋กฌํ”„ํŠธ์— ๊ฐœ์ธ ์ •๋ณด๊ฐ€ ์ข์€ ๋ถ„ํฌ๋กœ๋งŒ ํฌํ•จ๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ์œ„ํ•ด ๋” ๊ด‘๋ฒ”์œ„ํ•œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Results are displayed in Table 10. Our models are far better at reconstructing some private entities than others: countries and nationalities are particularly successful, while individual dates and years are typically lost during the inversion process. Future work might consider more wide-ranging training data for better performance, since Instructions-2M includes only a narrow distribution of personal information in its prompts. [|assistant|]
์œ ํšจ์„ฑ ๊ฒ€์‚ฌ Loss ๋น„๊ต๊ทธ๋ฆผ 4์˜ ๊ฒฐ๊ณผ(์œ„์ชฝ๊ณผ ์•„๋ž˜์ชฝ)๋Š” ๊ฐ๊ฐ์˜ \(\mathcal{D}_{1}\) ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ Loss์˜ ์ง„ํ™”๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 2๋Š” ์ด๋Ÿฌํ•œ ๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ตœ์ข… ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ Loss์˜ ํ‰๊ท ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ข… Loss๋Š” 10ํšŒ ๊ฐ„๊ฒฉ์œผ๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋งˆ์ง€๋ง‰ 100ํšŒ ํ›ˆ๋ จ ๋ฐ˜๋ณต์— ๋Œ€ํ•ด ํ‰๊ท ํ™”๋ฉ๋‹ˆ๋‹ค. ๋‘ ๋ถ„ํฌ ์ด๋™ ๋ชจ๋‘์—์„œ ํ…Œ์ŠคํŠธ๋œ ๊ฐ€์žฅ ๋‚ฎ์€ ๋ฆฌํ”Œ๋ ˆ์ด์ธ 1%๋„ ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ํ•˜์ง€ ์•Š์€ baseline๋ณด๋‹ค Pile์—์„œ ์žŠ์–ด๋ฒ„๋ฆผ์„ ์ƒ๋‹นํžˆ ์ค„์—ฌ์ฃผ๋Š” ๊ฒƒ์„ ์ผ๊ด€๋˜๊ฒŒ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ํšจ๊ณผ๋Š” ์ด ์„ค์ •์—์„œ ๋” ๋งŽ์€ ์–‘์˜ ์žŠ์–ด๋ฒ„๋ฆผ์ด ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ฐ•ํ•œ ์‹œํ”„ํŠธ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋” ๋‘๋“œ๋Ÿฌ์ง‘๋‹ˆ๋‹ค. 0% ๊ธฐ์ค€๊ณผ ๋น„๊ตํ•  ๋•Œ 1%, 5%, 10% ๋ฆฌํ”Œ๋ ˆ์ด์˜ ๊ฒฝ์šฐ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๊ฑฐ์˜ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์•„ ๋ฆฌํ”Œ๋ ˆ์ด์˜ ์žŠ์–ด๋ฒ„๋ฆผ ์ด์ ์ด ์šฐ๋ฆฌ ์„ค์ •์—์„œ ๊ฑฐ์˜ ๋Œ€๊ฐ€๋ฅผ ์น˜๋ฅด์ง€ ์•Š์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ทน๋‹จ์ ์ธ ์–‘์˜ ๋ฆฌํ”Œ๋ ˆ์ด(50%)๋ฅผ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ ๋ชจ๋ธ์ด \(\mathcal{D}_{0}\)์— ๋œ ์ ์‘ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ชจ๋‘์—์„œ 50% ๋ฆฌํ”Œ๋ ˆ์ด ๋ชจ๋ธ์€ \(\mathcal{D}_{1}\cup\(\mathcal{D}_{0}\)์— ๋Œ€ํ•œ baseline ํ›ˆ๋ จ์˜ ์ตœ์ข… ํ‰๊ท  ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๊ฑฐ๋‚˜ ์ดˆ๊ณผํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ด ๋ชจ๋ธ์ด ๊ฐ ๊ธฐ์ค€๋ณด๋‹ค 150B์™€ 100B์˜ ํ† ํฐ์„ ๋” ์ ๊ฒŒ ๋ณด์•˜๊ธฐ ๋•Œ๋ฌธ์— ์‹ ๊ธฐํ•œ ์ผ์ž…๋‹ˆ๋‹ค.[|endofturn|]
์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์†์‹ค ๋น„๊ต๊ทธ๋ฆผ 4์˜ ๊ฒฐ๊ณผ(์œ„์ชฝ๊ณผ ์•„๋ž˜์ชฝ)๋Š” ๊ฐ๊ฐ์˜ \(\mathcal{D}_{1}\) ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์†์‹ค์˜ ์ง„ํ™”๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 2๋Š” ์ด๋Ÿฌํ•œ ๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ตœ์ข… ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์†์‹ค์˜ ํ‰๊ท ์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ข… ์†์‹ค์€ 10ํšŒ ๊ฐ„๊ฒฉ์œผ๋กœ ์ƒ˜ํ”Œ๋ง๋œ ๋งˆ์ง€๋ง‰ 100ํšŒ ํ›ˆ๋ จ ๋ฐ˜๋ณต์— ๋Œ€ํ•ด ํ‰๊ท ํ™”๋ฉ๋‹ˆ๋‹ค. ๋‘ ๋ถ„ํฌ ์ด๋™ ๋ชจ๋‘์—์„œ ํ…Œ์ŠคํŠธ๋œ ๊ฐ€์žฅ ๋‚ฎ์€ ๋ฆฌํ”Œ๋ ˆ์ด์ธ 1%๋„ ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ํ•˜์ง€ ์•Š์€ ๊ธฐ๋ณธ์„ ๋ณด๋‹ค Pile์—์„œ ์žŠ์–ด๋ฒ„๋ฆผ์„ ์ƒ๋‹นํžˆ ์ค„์—ฌ์ฃผ๋Š” ๊ฒƒ์„ ์ผ๊ด€๋˜๊ฒŒ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ํšจ๊ณผ๋Š” ์ด ์„ค์ •์—์„œ ๋” ๋งŽ์€ ์–‘์˜ ์žŠ์–ด๋ฒ„๋ฆผ์ด ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ฐ•ํ•œ ์‹œํ”„ํŠธ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋” ๋‘๋“œ๋Ÿฌ์ง‘๋‹ˆ๋‹ค. 0% ๊ธฐ์ค€๊ณผ ๋น„๊ตํ•  ๋•Œ 1%, 5%, 10% ๋ฆฌํ”Œ๋ ˆ์ด์˜ ๊ฒฝ์šฐ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๊ฑฐ์˜ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์•„ ๋ฆฌํ”Œ๋ ˆ์ด์˜ ์žŠ์–ด๋ฒ„๋ฆผ ์ด์ ์ด ์šฐ๋ฆฌ ์„ค์ •์—์„œ ๊ฑฐ์˜ ๋Œ€๊ฐ€๋ฅผ ์น˜๋ฅด์ง€ ์•Š์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ทน๋‹จ์ ์ธ ์–‘์˜ ๋ฆฌํ”Œ๋ ˆ์ด(50%)๋ฅผ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ ๋ชจ๋ธ์ด \(\mathcal{D}_{0}\)์— ๋œ ์ ์‘ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ชจ๋‘์—์„œ 50% ๋ฆฌํ”Œ๋ ˆ์ด ๋ชจ๋ธ์€ \(\mathcal{D}_{1}\cup\mathcal{D}_{0}\)์— ๋Œ€ํ•œ ๊ธฐ๋ณธ ํ›ˆ๋ จ์˜ ์ตœ์ข… ํ‰๊ท  ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๊ฑฐ๋‚˜ ์ดˆ๊ณผํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ด ๋ชจ๋ธ์ด ๊ฐ ๊ธฐ์ค€๋ณด๋‹ค 150B์™€ 100B์˜ ํ† ํฐ์„ ๋” ์ ๊ฒŒ ๋ณด์•˜๊ธฐ ๋•Œ๋ฌธ์— ์‹ ๊ธฐํ•œ ์ผ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Validation Loss ComparisonThe results in Fig. 4 (top and bottom) show the evolution of the validation loss during continual pre-training on the respective \(\mathcal{D}_{1}\) datasets. Table 2 reports the average final validation loss for each of these models. The final loss is averaged over the last 100 iterations of training sampled at intervals of 10 iterations. We consistently observe across both distribution shifts that even the lowest tested replay of 1% significantly reduces forgetting on Pile compared to the no-replay baselines. This effect is more pronounced in the strong-shift scenario due to the larger amount of forgetting in this setting. We observe little impact on downstream performance for 1%,5%, and 10% replay when compared to the 0% baseline, showing that the forgetting benefits of replay come at little cost in our setting. However, when using an extreme amount of replay (50%), we observe that the model adapts less to \(\mathcal{D}_{0}\). Interestingly, for both datasets, the 50% replay models attain or surpass the final average validation performance of the baseline training on \(\mathcal{D}_{1}\cup\mathcal{D}_{0}\). This is curious as these model have seen 150B and 100B fewer tokens of \(\mathcal{D}_{1}\) than their respective baselines. [|assistant|]
๋ชจ๋“  ๋จธ์‹  ๋Ÿฌ๋‹ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ hyperparameter์˜ ์„ ํƒ์€ ๊ฒฐ๊ณผ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํŠน์ • ์‚ฌ๋ก€์— ๋งž๊ฒŒ ๋ชจ๋“  ๊ด€๋ จ hyperparameter๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ชจ๋“  ๋จธ์‹  ๋Ÿฌ๋‹ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ์˜ ์„ ํƒ์€ ๊ฒฐ๊ณผ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํŠน์ • ์‚ฌ๋ก€์— ๋งž๊ฒŒ ๋ชจ๋“  ๊ด€๋ จ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]As in all machine learning applications, the choice of hyperparameters significantly impacts the results. We therefore strongly recommend optimizing all relevant hyperparameters for specific cases. [|assistant|]
MMLU ๋ฐ ์ฒด์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ์™€ ๊ฐ™์ด ์ผ๋ฐ˜์ ์ธ ์ถ”๋ก  ์ž‘์—…์˜ ๊ฒฝ์šฐ, ์ƒ˜ํ”Œ๋ง ๋‹จ๊ณ„์—์„œ Algorithm 1์— ๋”ฐ๋ผ sampling-and-voting ๋ฐฉ๋ฒ•์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ƒ˜ํ”Œ์€ "(X" ๋˜๋Š” "(X)" ํŒจํ„ด์„ ์ผ์น˜์‹œ์ผœ ์ถ”์ถœ๋˜๋ฉฐ, ์—ฌ๊ธฐ์„œ "X"๋Š” MMLU ์ž‘์—…์˜ ์˜ต์…˜ A, B, C ๋˜๋Š” D ๋ฐ ์ฒด์Šค ์ž‘์—…์˜ ์ฒด์ŠคํŒ ์œ„์น˜์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
MMLU ๋ฐ ์ฒด์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ์™€ ๊ฐ™์ด ์ผ๋ฐ˜์ ์ธ ์ถ”๋ก  ์ž‘์—…์˜ ๊ฒฝ์šฐ, ์ƒ˜ํ”Œ๋ง ๋‹จ๊ณ„์—์„œ ์•Œ๊ณ ๋ฆฌ์ฆ˜ 1์— ๋”ฐ๋ผ ์ƒ˜ํ”Œ๋ง ๋ฐ ํˆฌํ‘œ ๋ฐฉ๋ฒ•์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ƒ˜ํ”Œ์€ "(X" ๋˜๋Š” "(X)" ํŒจํ„ด์„ ์ผ์น˜์‹œ์ผœ ์ถ”์ถœ๋˜๋ฉฐ, ์—ฌ๊ธฐ์„œ "X"๋Š” MMLU ์ž‘์—…์˜ ์˜ต์…˜ A, B, C ๋˜๋Š” D ๋ฐ ์ฒด์Šค ์ž‘์—…์˜ ์ฒด์ŠคํŒ ์œ„์น˜์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]For general reasoning tasks, as encountered in the MMLU and Chess datasets, the sampling-and-voting method is applied following Algorithm 1 during the sampling phase. Samples are extracted by matching the pattern "(X" or "(X)", where "X" corresponds to the options A, B, C, or D in MMLU task and the chessboard position in Chess task. [|assistant|]
**Chinchilla:** [2]์—์„œ Hoffmann ๋“ฑ์€ ์ฃผ์–ด์ง„ compute budget ํ•˜์—์„œ Transformer ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ ์˜ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ํ† ํฐ ์ˆ˜๋ฅผ ์กฐ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. 7์ฒœ๋งŒ ๊ฐœ์—์„œ 160์–ต ๊ฐœ ์ด์ƒ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ์— ์ด๋ฅด๋Š” 400๊ฐœ ์ด์ƒ์˜ ์–ธ์–ด ๋ชจ๋ธ์„ 50์–ต ๊ฐœ์—์„œ 5,000์–ต ๊ฐœ ์ด์ƒ์˜ ํ† ํฐ์— ๋Œ€ํ•ด ํ›ˆ๋ จํ•œ ๊ฒฐ๊ณผ, compute-optimal ํ›ˆ๋ จ์˜ ๊ฒฝ์šฐ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ํ›ˆ๋ จ ํ† ํฐ ์ˆ˜๋ฅผ ๋™์ผํ•˜๊ฒŒ ํ™•์žฅํ•ด์•ผ ํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฆ‰, ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ๋‘ ๋ฐฐ๋กœ ๋Š˜์–ด๋‚  ๋•Œ๋งˆ๋‹ค ํ›ˆ๋ จ ํ† ํฐ ์ˆ˜๋„ ๋‘ ๋ฐฐ๋กœ ๋Š˜์–ด๋‚˜์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฐ€์„ค์„ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด compute budget์€ Gopher์™€ ๋™์ผํ•˜์ง€๋งŒ ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” 700์–ต ๊ฐœ, ๋ฐ์ดํ„ฐ๋Š” 4% ๋” ๋งŽ์€ ์˜ˆ์ธก compute-optimal ๋ชจ๋ธ์ธ Chinchilla๋ฅผ ํ›ˆ๋ จํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
**์นœ์น ๋ผ:** [2]์—์„œ ํ˜ธํ”„๋งŒ ๋“ฑ์€ ์ฃผ์–ด์ง„ ์ปดํ“จํŒ… ์˜ˆ์‚ฐ ํ•˜์—์„œ ํŠธ๋žœ์Šคํฌ๋จธ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ ์˜ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ํ† ํฐ ์ˆ˜๋ฅผ ์กฐ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. 7์ฒœ๋งŒ ๊ฐœ์—์„œ 160์–ต ๊ฐœ ์ด์ƒ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ์— ์ด๋ฅด๋Š” 400๊ฐœ ์ด์ƒ์˜ ์–ธ์–ด ๋ชจ๋ธ์„ 50์–ต ๊ฐœ์—์„œ 5,000์–ต ๊ฐœ ์ด์ƒ์˜ ํ† ํฐ์— ๋Œ€ํ•ด ํ›ˆ๋ จํ•œ ๊ฒฐ๊ณผ, ์ปดํ“จํŒ… ์ตœ์ ํ™” ํ›ˆ๋ จ์˜ ๊ฒฝ์šฐ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ํ›ˆ๋ จ ํ† ํฐ ์ˆ˜๋ฅผ ๋™์ผํ•˜๊ฒŒ ํ™•์žฅํ•ด์•ผ ํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฆ‰, ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ๋‘ ๋ฐฐ๋กœ ๋Š˜์–ด๋‚  ๋•Œ๋งˆ๋‹ค ํ›ˆ๋ จ ํ† ํฐ ์ˆ˜๋„ ๋‘ ๋ฐฐ๋กœ ๋Š˜์–ด๋‚˜์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฐ€์„ค์„ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ์ปดํ“จํŒ… ์˜ˆ์‚ฐ์€ ๊ณ ํผ์™€ ๋™์ผํ•˜์ง€๋งŒ ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” 700์–ต ๊ฐœ, ๋ฐ์ดํ„ฐ๋Š” 4% ๋” ๋งŽ์€ ์˜ˆ์ธก ์ปดํ“จํŒ… ์ตœ์  ๋ชจ๋ธ์ธ ์นœ์น ๋ผ๋ฅผ ํ›ˆ๋ จํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Chinchilla:** In [2], Hoffmann et al. investigated the optimal model size and number of tokens for training a transformer language model under a given compute budget. By training over 400 language models ranging from 70 million to over 16 billion parameters on 5 to 500 billion tokens, they found that for compute-optimal training, the model size and the number of training tokens should be scaled equally: for every doubling of model size the number of training tokens should also be doubled. They tested this hypothesis by training a predicted compute-optimal model, Chinchilla, that uses the same compute budget as Gopher but with 70B parameters and 4% more more data. [|assistant|]
์–ธ์–ด ๋ชจ๋ธ์€ ๋‹ค์Œ ํ† ํฐ์— ๋Œ€ํ•œ ๋ถ„ํฌ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ ํ† ํฐ์„ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ์–ธ์–ด ๋ชจ๋ธ inversion ๋ฌธ์ œ๋ฅผ ๊ณ ๋ คํ•˜๊ณ  ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์— ์ด์ „ ํ…์ŠคํŠธ์— ๋Œ€ํ•œ ๋†€๋ผ์šด ์–‘์˜ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ข…์ข… ์‚ฌ์šฉ์ž์—๊ฒŒ ์ˆจ๊ฒจ์ ธ ์žˆ๋Š” ๊ฒฝ์šฐ์—๋„ ํ…์ŠคํŠธ๋ฅผ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ์˜ ํ˜„์žฌ ๋ถ„ํฌ ์ถœ๋ ฅ๋งŒ ์ฃผ์–ด์กŒ์„ ๋•Œ ์•Œ ์ˆ˜ ์—†๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณต๊ตฌํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋™๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์•ก์„ธ์Šค ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๊ณ ๋ คํ•˜๊ณ  ์–ดํœ˜์˜ ๋ชจ๋“  ํ† ํฐ์— ๋Œ€ํ•œ ์˜ˆ์ธก์ด ์—†๋”๋ผ๋„ ๊ฒ€์ƒ‰์„ ํ†ตํ•ด ํ™•๋ฅ  ๋ฒกํ„ฐ๋ฅผ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
์–ธ์–ด ๋ชจ๋ธ์€ ๋‹ค์Œ ํ† ํฐ์— ๋Œ€ํ•œ ๋ถ„ํฌ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กฌํ”„ํŠธ ํ† ํฐ์„ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ์–ธ์–ด ๋ชจ๋ธ ๋ฐ˜์ „ ๋ฌธ์ œ๋ฅผ ๊ณ ๋ คํ•˜๊ณ  ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์— ์ด์ „ ํ…์ŠคํŠธ์— ๋Œ€ํ•œ ๋†€๋ผ์šด ์–‘์˜ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ข…์ข… ์‚ฌ์šฉ์ž์—๊ฒŒ ์ˆจ๊ฒจ์ ธ ์žˆ๋Š” ๊ฒฝ์šฐ์—๋„ ํ…์ŠคํŠธ๋ฅผ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋ชจ๋ธ์˜ ํ˜„์žฌ ๋ถ„ํฌ ์ถœ๋ ฅ๋งŒ ์ฃผ์–ด์กŒ์„ ๋•Œ ์•Œ ์ˆ˜ ์—†๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณต๊ตฌํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋™๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์•ก์„ธ์Šค ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๊ณ ๋ คํ•˜๊ณ  ์–ดํœ˜์˜ ๋ชจ๋“  ํ† ํฐ์— ๋Œ€ํ•œ ์˜ˆ์ธก์ด ์—†๋”๋ผ๋„ ๊ฒ€์ƒ‰์„ ํ†ตํ•ด ํ™•๋ฅ  ๋ฒกํ„ฐ๋ฅผ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Language models produce a distribution over the next token; can we use this to recover the prompt tokens? We consider the problem of language model inversion and show that next-token probabilities contain a surprising amount of information about the preceding text. Often we can recover the text in cases where it is hidden from the user, motivating a method for recovering unknown prompts given only the model's current distribution output. We consider a variety of model access scenarios, and show how even without predictions for every token in the vocabulary we can recover the probability vector through search. [|assistant|]
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM) ๊ธฐ๋ฐ˜ ์ง€๋Šฅํ˜• ์—์ด์ „ํŠธ์˜ ํ†ตํ•ฉ ๋ฐ ๋ฐฐํฌ์—๋Š” ํšจ์œจ์„ฑ๊ณผ ํšจ๊ณผ์„ฑ์„ ์ €ํ•ดํ•˜๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๋ฌธ์ œ๊ฐ€ ์ˆ˜๋ฐ˜๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ ์ค‘์—๋Š” LLM์— ๋Œ€ํ•œ ์—์ด์ „ํŠธ ์š”์ฒญ์˜ ์ตœ์ ์ด ์•„๋‹Œ ์Šค์ผ€์ค„๋ง ๋ฐ ๋ฆฌ์†Œ์Šค ํ• ๋‹น, ์—์ด์ „ํŠธ์™€ LLM ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ ์ค‘ ์ปจํ…์ŠคํŠธ ์œ ์ง€์˜ ์–ด๋ ค์›€, ๋‹ค์–‘ํ•œ ๊ธฐ๋Šฅ๊ณผ ์ „๋ฌธ์„ฑ์„ ๊ฐ€์ง„ ์ด๊ธฐ์ข… ์—์ด์ „ํŠธ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ๋‚ด์žฌ๋œ ๋ณต์žก์„ฑ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์˜ ์ˆ˜์™€ ๋ณต์žก์„ฑ์ด ๊ธ‰๊ฒฉํžˆ ์ฆ๊ฐ€ํ•˜๋ฉด์„œ ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๊ฐ€ ๋”์šฑ ์•…ํ™”๋˜์–ด ๋ณ‘๋ชฉ ํ˜„์ƒ๊ณผ ๋ฆฌ์†Œ์Šค์˜ ์ตœ์  ํ™œ์šฉ์ด ์ €ํ•ด๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณผ์ œ์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ์ด ๋ฐฑ์„œ์—์„œ๋Š” LLM ์—์ด์ „ํŠธ ์šด์˜ ์ฒด์ œ์ธ AIOS๋ฅผ ์†Œ๊ฐœํ•˜๋Š”๋ฐ, ์ด๋Š” ์šด์˜ ์ฒด์ œ(OS)์— Large Language Model์„ ๋‚ด์žฅํ•˜์—ฌ OS์˜ ๋‘๋‡Œ ์—ญํ• ์„ ์ˆ˜ํ–‰ํ•จ์œผ๋กœ์จ AGI๋ฅผ ํ–ฅํ•œ ์ค‘์š”ํ•œ ๋‹จ๊ณ„์ธ '์˜ํ˜ผ์„ ๊ฐ€์ง„' ์šด์˜ ์ฒด์ œ๋ฅผ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. ํŠนํžˆ AIOS๋Š” ๋ฆฌ์†Œ์Šค ํ• ๋‹น์„ ์ตœ์ ํ™”ํ•˜๊ณ , ์—์ด์ „ํŠธ ๊ฐ„์— ์ปจํ…์ŠคํŠธ ์ „ํ™˜์„ ์ด‰์ง„ํ•˜๊ณ , ์—์ด์ „ํŠธ์˜ ๋™์‹œ ์‹คํ–‰์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๊ณ , ์—์ด์ „ํŠธ๋ฅผ ์œ„ํ•œ ๋„๊ตฌ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ , ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ์•ก์„ธ์Šค ์ œ์–ด๋ฅผ ์œ ์ง€ํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์šด์˜ ์ฒด์ œ์˜ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ œ์‹œํ•˜๊ณ , ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•˜๋Š” ํ•ต์‹ฌ ๊ณผ์ œ๋ฅผ ๊ฐœ๊ด„์ ์œผ๋กœ ์„ค๋ช…ํ•˜๋ฉฐ, AIOS์˜ ๊ธฐ๋ณธ ์„ค๊ณ„ ๋ฐ ๊ตฌํ˜„์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ๋™์‹œ ์‹คํ–‰์— ๋Œ€ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด AIOS ๋ชจ๋“ˆ์˜ ์•ˆ์ •์„ฑ๊ณผ ํšจ์œจ์„ฑ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด LLM ์—์ด์ „ํŠธ์˜ ์„ฑ๋Šฅ๊ณผ ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ํ–ฅํ›„ AIOS ์—์ฝ”์‹œ์Šคํ…œ์˜ ๋” ๋‚˜์€ ๊ฐœ๋ฐœ๊ณผ ๋ฐฐํฌ๋ฅผ ๊ฐœ์ฒ™ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์ ํŠธ๋Š” [https://github.com/agiresearch/AIOS](https://github.com/agiresearch/AIOS)์—์„œ ์˜คํ”ˆ ์†Œ์Šค๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM) ๊ธฐ๋ฐ˜ ์ง€๋Šฅํ˜• ์—์ด์ „ํŠธ์˜ ํ†ตํ•ฉ ๋ฐ ๋ฐฐํฌ์—๋Š” ํšจ์œจ์„ฑ๊ณผ ํšจ๊ณผ์„ฑ์„ ์ €ํ•ดํ•˜๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๋ฌธ์ œ๊ฐ€ ์ˆ˜๋ฐ˜๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ ์ค‘์—๋Š” LLM์— ๋Œ€ํ•œ ์—์ด์ „ํŠธ ์š”์ฒญ์˜ ์ตœ์ ์ด ์•„๋‹Œ ์Šค์ผ€์ค„๋ง ๋ฐ ๋ฆฌ์†Œ์Šค ํ• ๋‹น, ์—์ด์ „ํŠธ์™€ LLM ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ ์ค‘ ์ปจํ…์ŠคํŠธ ์œ ์ง€์˜ ์–ด๋ ค์›€, ๋‹ค์–‘ํ•œ ๊ธฐ๋Šฅ๊ณผ ์ „๋ฌธ์„ฑ์„ ๊ฐ€์ง„ ์ด๊ธฐ์ข… ์—์ด์ „ํŠธ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ๋‚ด์žฌ๋œ ๋ณต์žก์„ฑ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์˜ ์ˆ˜์™€ ๋ณต์žก์„ฑ์ด ๊ธ‰๊ฒฉํžˆ ์ฆ๊ฐ€ํ•˜๋ฉด์„œ ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๊ฐ€ ๋”์šฑ ์•…ํ™”๋˜์–ด ๋ณ‘๋ชฉ ํ˜„์ƒ๊ณผ ๋ฆฌ์†Œ์Šค์˜ ์ตœ์  ํ™œ์šฉ์ด ์ €ํ•ด๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณผ์ œ์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ์ด ๋ฐฑ์„œ์—์„œ๋Š” LLM ์—์ด์ „ํŠธ ์šด์˜ ์ฒด์ œ์ธ AIOS๋ฅผ ์†Œ๊ฐœํ•˜๋Š”๋ฐ, ์ด๋Š” ์šด์˜ ์ฒด์ œ(OS)์— ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ๋‚ด์žฅํ•˜์—ฌ OS์˜ ๋‘๋‡Œ ์—ญํ• ์„ ์ˆ˜ํ–‰ํ•จ์œผ๋กœ์จ AGI๋ฅผ ํ–ฅํ•œ ์ค‘์š”ํ•œ ๋‹จ๊ณ„์ธ '์˜ํ˜ผ์„ ๊ฐ€์ง„' ์šด์˜ ์ฒด์ œ๋ฅผ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. ํŠนํžˆ AIOS๋Š” ๋ฆฌ์†Œ์Šค ํ• ๋‹น์„ ์ตœ์ ํ™”ํ•˜๊ณ , ์—์ด์ „ํŠธ ๊ฐ„์— ์ปจํ…์ŠคํŠธ ์ „ํ™˜์„ ์ด‰์ง„ํ•˜๊ณ , ์—์ด์ „ํŠธ์˜ ๋™์‹œ ์‹คํ–‰์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๊ณ , ์—์ด์ „ํŠธ๋ฅผ ์œ„ํ•œ ๋„๊ตฌ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ , ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ์•ก์„ธ์Šค ์ œ์–ด๋ฅผ ์œ ์ง€ํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์šด์˜ ์ฒด์ œ์˜ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ œ์‹œํ•˜๊ณ , ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•˜๋Š” ํ•ต์‹ฌ ๊ณผ์ œ๋ฅผ ๊ฐœ๊ด„์ ์œผ๋กœ ์„ค๋ช…ํ•˜๋ฉฐ, AIOS์˜ ๊ธฐ๋ณธ ์„ค๊ณ„ ๋ฐ ๊ตฌํ˜„์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ๋™์‹œ ์‹คํ–‰์— ๋Œ€ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด AIOS ๋ชจ๋“ˆ์˜ ์•ˆ์ •์„ฑ๊ณผ ํšจ์œจ์„ฑ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด LLM ์—์ด์ „ํŠธ์˜ ์„ฑ๋Šฅ๊ณผ ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ํ–ฅํ›„ AIOS ์—์ฝ”์‹œ์Šคํ…œ์˜ ๋” ๋‚˜์€ ๊ฐœ๋ฐœ๊ณผ ๋ฐฐํฌ๋ฅผ ๊ฐœ์ฒ™ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์ ํŠธ๋Š” [https://github.com/agiresearch/AIOS](https://github.com/agiresearch/AIOS)์—์„œ ์˜คํ”ˆ ์†Œ์Šค๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The integration and deployment of large language model (LLM)-based intelligent agents have been fraught with challenges that compromise their efficiency and efficacy. Among these issues are sub-optimal scheduling and resource allocation of agent requests over the LLM, the difficulties in maintaining context during interactions between agent and LLM, and the complexities inherent in integrating heterogeneous agents with different capabilities and specializations. The rapid increase of agent quantity and complexity further exacerbates these issues, often leading to bottlenecks and sub-optimal utilization of resources. Inspired by these challenges, this paper presents AIOS, an LLM agent operating system, which embeds large language model into operating systems (OS) as the brain of the OS, enabling an operating system "with soul"--an important step towards AGI. Specifically, AIOS is designed to optimize resource allocation, facilitate context switch across agents, enable concurrent execution of agents, provide tool service for agents, and maintain access control for agents. We present the architecture of such an operating system, outline the core challenges it aims to resolve, and provide the basic design and implementation of the AIOS. Our experiments on concurrent execution of multiple agents demonstrate the reliability and efficiency of our AIOS modules. Through this, we aim to not only improve the performance and efficiency of LLM agents but also to pioneer for better development and deployment of the AIOS ecosystem in the future. The project is open-source at [https://github.com/agiresearch/AIOS](https://github.com/agiresearch/AIOS). [|assistant|]
๋งˆ์ง€๋ง‰์œผ๋กœ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์ด RMS ๋ณด์ • ์˜ค๋ฅ˜๋กœ ์ž˜ ๋ณด์ •๋˜์–ด ์žˆ์Œ์„ ๋‹ค์‹œ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค.42(์‚ฌ๋žŒ๋“ค: .38). ํฅ๋ฏธ๋กญ๊ฒŒ๋„, ์ด๋Š” ๋ชจ๋ธ์ด zero-shot ์„ค์ •์—์„œ ์ž˜ ๋ณด์ •๋˜์ง€ ์•Š์€ baseline ํ‰๊ฐ€(๊ทธ๋ฆผ 2(a))์—์„œ๋Š” ๊ทธ๋ ‡์ง€ ์•Š์Šต๋‹ˆ๋‹ค(์„น์…˜ 3.4). Fine-tuning ๋ฐ ์•™์ƒ๋ธ”์„ ํ†ตํ•ด ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์€ ๋ณด์ •์„ ์œ„ํ•œ ํŠน์ • ํ›ˆ๋ จ์„ ๊ฑฐ์น˜์ง€ ์•Š๊ณ ๋„ ๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ๋ณด์ •์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋งˆ์ง€๋ง‰์œผ๋กœ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์ด RMS ๋ณด์ • ์˜ค๋ฅ˜๋กœ ์ž˜ ๋ณด์ •๋˜์–ด ์žˆ์Œ์„ ๋‹ค์‹œ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค.42(์‚ฌ๋žŒ๋“ค: .38). ํฅ๋ฏธ๋กญ๊ฒŒ๋„, ์ด๋Š” ๋ชจ๋ธ์ด ์ œ๋กœ ์ƒท ์„ค์ •์—์„œ ์ž˜ ๋ณด์ •๋˜์ง€ ์•Š์€ ๊ธฐ๋ณธ ํ‰๊ฐ€(๊ทธ๋ฆผ 2(a))์—์„œ๋Š” ๊ทธ๋ ‡์ง€ ์•Š์Šต๋‹ˆ๋‹ค(์„น์…˜ 3.4). ๋ฏธ์„ธ ์กฐ์ • ๋ฐ ์•™์ƒ๋ธ”์„ ํ†ตํ•ด ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์€ ๋ณด์ •์„ ์œ„ํ•œ ํŠน์ • ํ›ˆ๋ จ์„ ๊ฑฐ์น˜์ง€ ์•Š๊ณ ๋„ ๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ๋ณด์ •์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Finally, on the test set, we observe again that our system is well calibrated (Figure 2(c)) with RMS calibration error.42 (human crowd:.38). Interestingly, this is not the case in the baseline evaluations (Section 3.4), where the models are _not_ well calibrated in the zero-shot setting (Figure 2(a)). Through fine-tuning and ensembling, our system improves the calibration of the base models, without undergoing specific training for calibration. [|assistant|]
๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์œ„ํ•œ ํŒŒ์ดํ”„๋ผ์ธ์€ 4๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: 1) ์ž‘์—… instructions ์ƒ์„ฑ, 2) instruction์ด ๋ถ„๋ฅ˜ ์ž‘์—…์„ ๋‚˜ํƒ€๋‚ด๋Š”์ง€ ํ™•์ธ, 3) ์ž…๋ ฅ ์šฐ์„  ๋˜๋Š” ์ถœ๋ ฅ ์šฐ์„  ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•œ ์ธ์Šคํ„ด์Šค ์ƒ์„ฑ, 4) ํ’ˆ์งˆ์ด ๋‚ฎ์€ ๋ฐ์ดํ„ฐ ํ•„ํ„ฐ๋ง์ž…๋‹ˆ๋‹ค.[|endofturn|]
๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์œ„ํ•œ ํŒŒ์ดํ”„๋ผ์ธ์€ 4๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: 1) ์ž‘์—… ์ง€์นจ ์ƒ์„ฑ, 2) ์ง€์นจ์ด ๋ถ„๋ฅ˜ ์ž‘์—…์„ ๋‚˜ํƒ€๋‚ด๋Š”์ง€ ํ™•์ธ, 3) ์ž…๋ ฅ ์šฐ์„  ๋˜๋Š” ์ถœ๋ ฅ ์šฐ์„  ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•œ ์ธ์Šคํ„ด์Šค ์ƒ์„ฑ, 4) ํ’ˆ์งˆ์ด ๋‚ฎ์€ ๋ฐ์ดํ„ฐ ํ•„ํ„ฐ๋ง์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Our pipeline for data generation consists of four steps: 1) generating task instructions, 2) determining if the instruction represents a classification task, 3) instance generation with either an input-first or output-first approach, and 4) filtering low-quality data. [|assistant|]
\(\bullet\) LM objective๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์ˆ˜ํ–‰ํ•˜๋ฉด causal decoder architecture๊ฐ€ ์šฐ์ˆ˜ํ•œ zero-shot ๋ฐ few-shot generalization ๊ธฐ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด multi-task Fine-tuning ์—†์ด causal decoder๋Š” ๋‹ค๋ฅธ ์•„ํ‚คํ…์ฒ˜๋ณด๋‹ค zero-shot ์„ฑ๋Šฅ์ด ๋” ์šฐ์ˆ˜ํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[29]. GPT-3 [55]์˜ ์„ฑ๊ณต์€ ํฐ causal decoder ๋ชจ๋ธ์ด ์ข‹์€ few-shot learner๊ฐ€ ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์„น์…˜ 5์—์„œ ๋…ผ์˜ํ•œ Instruction Tuning ๋ฐ alignment tuning์€ ๋Œ€๊ทœ๋ชจ causal decoder ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[66, 67, 69].[|endofturn|]
\(\bullet\) LM ๋ชฉํ‘œ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์ˆ˜ํ–‰ํ•˜๋ฉด ์ธ๊ณผ์  ๋””์ฝ”๋” ์•„ํ‚คํ…์ฒ˜๊ฐ€ ์šฐ์ˆ˜ํ•œ ์ œ๋กœ ์ƒท ๋ฐ ์†Œ์ˆ˜ ์ƒท ์ผ๋ฐ˜ํ™” ๊ธฐ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๋‹ค์ค‘ ์ž‘์—… ๋ฏธ์„ธ ์กฐ์ • ์—†์ด ์ธ๊ณผ์  ๋””์ฝ”๋”๋Š” ๋‹ค๋ฅธ ์•„ํ‚คํ…์ฒ˜๋ณด๋‹ค ์ œ๋กœ ์ƒท ์„ฑ๋Šฅ์ด ๋” ์šฐ์ˆ˜ํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[29]. GPT-3 [55]์˜ ์„ฑ๊ณต์€ ํฐ ์ธ๊ณผ์  ๋””์ฝ”๋” ๋ชจ๋ธ์ด ์ข‹์€ ์†Œ์ˆ˜ ์ƒท ํ•™์Šต์ž๊ฐ€ ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์„น์…˜ 5์—์„œ ๋…ผ์˜ํ•œ ์ง€์นจ ํŠœ๋‹ ๋ฐ ์ •๋ ฌ ํŠœ๋‹์€ ๋Œ€๊ทœ๋ชจ ์ธ๊ณผ์  ๋””์ฝ”๋” ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[66, 67, 69].[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\) By pre-training with the LM objective, it seems that causal decoder architecture can achieve a superior zero-shot and few-shot generalization capacity. Existing research has shown that without multi-task fine-tuning, the causal decoder has better zero-shot performance than other architectures [29]. The success of GPT-3 [55] has demonstrates that the large causal decoder model can be a good few-shot learner. In addition, instruction tuning and alignment tuning discussed in Section 5 have been proven to further enhance the capability of large causal decoder models [66, 67, 69]. [|assistant|]
๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ์—๋Š” ํ…Œ์ด๋ธ” ๋ฐ ์ง€์‹ ๊ทธ๋ž˜ํ”„์™€ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ๊ตฌ์กฐ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ๋ฐ์ดํ„ฐ์˜ ์žฅ์ ์€ ๊ฐ ํ•„๋“œ๊ฐ€ ์ •ํ™•ํ•˜๊ฒŒ ์ •์˜๋œ ํ‘œ ํ˜•์‹์˜ ๋ช…ํ™•ํ•œ ๊ตฌ์กฐ๋ผ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ˆซ์ž, ๋‚ ์งœ, ํ…์ŠคํŠธ ๋ฐ ๊ธฐํƒ€ ๋ฐ์ดํ„ฐ ์œ ํ˜•์„ ์ €์žฅํ•˜๋Š” ๋ฐ ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ๋Š” SQL๊ณผ ๊ฐ™์€ structured query language๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‰ฝ๊ฒŒ ์ฟผ๋ฆฌ, ๋ถ„์„ ๋ฐ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ์—๋Š” ํ…Œ์ด๋ธ” ๋ฐ ์ง€์‹ ๊ทธ๋ž˜ํ”„์™€ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ๊ตฌ์กฐ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ๋ฐ์ดํ„ฐ์˜ ์žฅ์ ์€ ๊ฐ ํ•„๋“œ๊ฐ€ ์ •ํ™•ํ•˜๊ฒŒ ์ •์˜๋œ ํ‘œ ํ˜•์‹์˜ ๋ช…ํ™•ํ•œ ๊ตฌ์กฐ๋ผ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ˆซ์ž, ๋‚ ์งœ, ํ…์ŠคํŠธ ๋ฐ ๊ธฐํƒ€ ๋ฐ์ดํ„ฐ ์œ ํ˜•์„ ์ €์žฅํ•˜๋Š” ๋ฐ ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ๋Š” SQL๊ณผ ๊ฐ™์€ ๊ตฌ์กฐํ™”๋œ ์ฟผ๋ฆฌ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‰ฝ๊ฒŒ ์ฟผ๋ฆฌ, ๋ถ„์„ ๋ฐ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Structured data includes various structures, such as tables and knowledge graphs. The benefit of this type of data is its clear structure, typically in tabular form, with each field precisely defined. It is appropriate for storing numbers, dates, text, and other data types. Structured data can be easily queried, analysed, and processed using a structured query language like SQL. [|assistant|]
**Simulated Dialogues.** ๋ฏธ์„ธ ์กฐ์ •์˜ ๊ฐ ๋ฐ˜๋ณต์—์„œ ์šฐ๋ฆฌ๋Š” 5,230๊ฐœ์˜ ๋‹ค๋ฅธ ์˜ํ•™์  ์ƒํƒœ์—์„œ ๋น„๋กฏ๋œ 11,686๊ฐœ์˜ ๋Œ€ํ™”๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
**์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™”**. ๋ฏธ์„ธ ์กฐ์ •์˜ ๊ฐ ๋ฐ˜๋ณต์—์„œ ์šฐ๋ฆฌ๋Š” 5,230๊ฐœ์˜ ๋‹ค๋ฅธ ์˜ํ•™์  ์ƒํƒœ์—์„œ ๋น„๋กฏ๋œ 11,686๊ฐœ์˜ ๋Œ€ํ™”๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Simulated Dialogues.** At each iteration of fine-tuning, we produced 11,686 dialogues, stemming from 5,230 different medical conditions. [|assistant|]
OLMo๋Š” ๋งŽ์€ ๊ฐœ์ธ๊ณผ ๊ธฐ๊ด€์˜ ์ง€์› ์—†์ด๋Š” ๋ถˆ๊ฐ€๋Šฅํ–ˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ์‹คํ—˜์  ๊ตฌ์„ฑ ์š”์†Œ๋Š” AMD ๋ฐ CSC์™€์˜ ํŒŒํŠธ๋„ˆ์‹ญ์„ ํ†ตํ•ด LUMI ์Šˆํผ์ปดํ“จํ„ฐ์™€ Harvard University์˜ Kemper Institute๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์—ˆ์œผ๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๊ฐ€๋Šฅํ–ˆ์Šต๋‹ˆ๋‹ค. Jonathan Frankle๊ณผ MosaicML(ํ˜„ Databricks) ํŒ€์—๊ฒŒ FSDP์— ๋Œ€ํ•œ ๊ฒฝํ—˜์„ ๊ณต์œ ํ•˜๊ณ  OLMo์˜ ๊ธฐ๋ฐ˜์ด ๋˜๋Š” ์ฝ”๋“œ๋ฒ ์ด์Šค๋ฅผ ๊ตฌ์ถ•ํ•ด์ค€ ๊ฒƒ์— ๋Œ€ํ•ด ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์›น์‚ฌ์ดํŠธ, ๋””์ž์ธ, ๋Œ€๋‚ด์™ธ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜, ์˜ˆ์‚ฐ ์ฑ…์ • ๋ฐ ๊ธฐํƒ€ ํ™œ๋™์„ ์ง€์›ํ•˜์—ฌ ํ”„๋กœ์ ํŠธ์˜ ์›ํ™œํ•œ ์ง„ํ–‰์„ ๋„์™€์ค€ ๋™๋ฃŒ์ธ Taira Anderson, Michelle Benedict, Jon Borchardt, Evie Cheng, Arnavi Chheda, Johann Dahm, Matt Latzke, Kelsey MacMillan, Aaron Sarnat, Carissa Schoenick, Sam Skjonsberg, Michael Schmitz, Michael Wilson, Caitlin Wittlif, ๊ทธ๋ฆฌ๊ณ  ์ „์ฒด IT ํŒ€์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ AI2์˜ ํŒ€์›๋“ค๊ณผ Prithviraj (Raj) Ammanabrolu, Peter Clark, Nicole DeCario, Doug Downey, Ali Farhadi, Ian Ferreira, Vaino Hatanpaa, Sham M. Kakade, Julien Launay, Sydney Levine, Pekka Manninen, Franzi Roessner, Maarten Sap, Ludwig Schmidt, Yulia Tsvetkov, ๊ทธ๋ฆฌ๊ณ  Daniel S. Weld๋ฅผ ํฌํ•จํ•œ ๊ธด๋ฐ€ํ•œ ํ˜‘๋ ฅ์ž๋“ค์˜ ๋„์›€์ด ๋œ ํ† ๋ก ๊ณผ ํ”ผ๋“œ๋ฐฑ์— ๋Œ€ํ•ด ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
OLMo๋Š” ๋งŽ์€ ๊ฐœ์ธ๊ณผ ๊ธฐ๊ด€์˜ ์ง€์› ์—†์ด๋Š” ๋ถˆ๊ฐ€๋Šฅํ–ˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ์‹คํ—˜์  ๊ตฌ์„ฑ ์š”์†Œ๋Š” AMD ๋ฐ CSC์™€์˜ ํŒŒํŠธ๋„ˆ์‹ญ์„ ํ†ตํ•ด ๋ฃธ๋ฏธ ์Šˆํผ์ปดํ“จํ„ฐ์™€ ํ•˜๋ฒ„๋“œ ๋Œ€ํ•™๊ต์˜ ์ผํผ ์—ฐ๊ตฌ์†Œ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์—ˆ์œผ๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๊ฐ€๋Šฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์กฐ๋‚˜๋‹จ ํ”„๋žญํด๊ณผ ๋ชจ์ž์ดํฌML(ํ˜„ ๋‹คํƒ€๋ธŒ๋ฆญ์Šค) ํŒ€์—๊ฒŒ FSDP์— ๋Œ€ํ•œ ๊ฒฝํ—˜์„ ๊ณต์œ ํ•˜๊ณ  OLMo์˜ ๊ธฐ๋ฐ˜์ด ๋˜๋Š” ์ฝ”๋“œ๋ฒ ์ด์Šค๋ฅผ ๊ตฌ์ถ•ํ•ด์ค€ ๊ฒƒ์— ๋Œ€ํ•ด ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์›น์‚ฌ์ดํŠธ, ๋””์ž์ธ, ๋Œ€๋‚ด์™ธ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜, ์˜ˆ์‚ฐ ์ฑ…์ • ๋ฐ ๊ธฐํƒ€ ํ™œ๋™์„ ์ง€์›ํ•˜์—ฌ ํ”„๋กœ์ ํŠธ์˜ ์›ํ™œํ•œ ์ง„ํ–‰์„ ๋„์™€์ค€ ๋™๋ฃŒ์ธ ํƒ€์ด๋ผ ์•ค๋”์Šจ, ๋ฏธ์…ธ ๋ฒ ๋„ค๋”•ํŠธ, ์กด ๋ณด์ฐจ๋“œ, ์—๋น„ ์ฒธ, ์•„๋ฅด๋‚˜๋น„ ์ฒด๋‹ค, ์š”ํ•œ ๋‹ด, ๋งคํŠธ ๋ผ์ธ ์ผ€, ์ผˆ์‹œ ๋งฅ๋ฐ€๋ž€, ์•„๋ก  ์‚ฌ๋ฅด๋‚˜ํŠธ, ์นด๋ฆฌ์‚ฌ ์‡ผ๋‹‰, ์ƒ˜ ์Šคํ‚ค์˜จ์Šค๋ฒ„๊ทธ, ๋งˆ์ดํด ์Šˆ๋ฏธ์ธ , ๋งˆ์ดํด ์œŒ์Šจ, ์ผ€์ผ๋ฆฐ ์œ—๋ฆฌํ”„, ๊ทธ๋ฆฌ๊ณ  ์ „์ฒด IT ํŒ€์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ AI2์˜ ํŒ€์›๋“ค๊ณผ ํ”„๋ฆฌํ‹ฐ๋น„๋ผ์ฆˆ(Raj) ์•”๋งˆ๋‚˜๋ธŒ๋กœ๋ฃจ, ํ”ผํ„ฐ ํด๋ผํฌ, ๋‹ˆ์ฝœ ๋ฐ์นด๋ฆฌ์˜ค, ๋”๊ทธ ๋‹ค์šฐ๋‹ˆ, ์•Œ๋ฆฌ ํŒŒ๋ฅดํ•˜๋””, ์ด์•ˆ ํŽ˜๋ ˆ์ด๋ผ, ๋ฐ”์ด๋…ธ ํ•˜ํƒ„ํŒŒ, ์ƒด M. ์นด์นด๋ฐ, ์ค„๋ฆฌ์•™ ๋กœ๋„ค, ์‹œ๋“œ๋‹ˆ ๋ ˆ๋นˆ, ํŽ˜์นด ๋งˆ๋‹ˆ๋„จ, ํ”„๋ž€์ง€ ๋กœ์—๋ฅด๋„ˆ, ๋งˆ๋ฅดํ… ์žฝ๋ฒ„๊ทธ, ๋ฃจ๋“œ๋น„ํžˆ ์Šˆ๋ฏธํŠธ, ์œจ๋ฆฌ์•„ ์ธ ๋ฒ ํŠธ์ฝ”ํ”„, ๋‹ค๋‹ˆ์—˜ S. ์›ฐ๋“œ๋ฅผ ํฌํ•จํ•œ ๊ธด๋ฐ€ํ•œ ํ˜‘๋ ฅ์ž๋“ค์˜ ๋„์›€์ด ๋œ ํ† ๋ก ๊ณผ ํ”ผ๋“œ๋ฐฑ์— ๋Œ€ํ•ด ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]OLMo would not have been possible without the support of many individuals and institutions. The experimental components of this work were made possible through a partnership with AMD and CSC, enabling use of the LUMI supercomputer, and Kemper Institute at Harvard University. We thank Jonathan Frankle and the team at MosaicML (now Databricks) for sharing their experiences with FSDP, and building the code base that OLMo is based on. We thank our teammates Taira Anderson, Michelle Benedict, Jon Borchardt, Evie Cheng, Arnavi Chheda, Johann Dahm, Matt Latzke, Kelsey MacMillan, Aaron Sarnat, Carissa Schoenick, Sam Skjonsberg, Michael Schmitz, Michael Wilson, Caitlin Wittlif, and the entire IT team, for their help with the website, design, internal and external communications, budgeting, and other activities that supported smooth progress on this project. Finally, we also express gratitude for the helpful discussions and feedback from our teammates at AI2 and close collaborators, including Prithviraj (Raj) Ammanabrolu, Peter Clark, Nicole DeCario, Doug Downey, Ali Farhadi, Ian Ferreira, Vaino Hatanpaa, Sham M. Kakade, Julien Launay, Sydney Levine, Pekka Manninen, Franzi Roessner, Maarten Sap, Ludwig Schmidt, Yulia Tsvetkov, and Daniel S. Weld. [|assistant|]
LLM ๊ธฐ๋ฐ˜ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ. LLM ๊ธฐ๋ฐ˜ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ(MAS)์€ ๋ฌธ์ œ ํ•ด๊ฒฐ์„ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๊ด€๊ณ„๋Š” ํ˜‘๋ ฅ์ , ๊ฒฝ์Ÿ์  ๋˜๋Š” ํ˜‘๋ ฅ๊ณผ ๊ฒฝ์Ÿ์ด ํ˜ผํ•ฉ๋˜์–ด ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[33]. ํ˜‘๋ ฅ์  ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์—์„œ ๊ฐ ์—์ด์ „ํŠธ๋Š” ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๊ฐ€ ์ œ๊ณตํ•œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ํ‰๊ฐ€ํ•˜์—ฌ ์—ญํ• ๊ทน[46], ์‚ฌํšŒ ์‹œ๋ฎฌ๋ ˆ์ด์…˜[47], ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ[48, 49, 50, 51]๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํ•จ๊ป˜ ์ž‘์—…ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฝ์Ÿ์  ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์—์„œ ์—์ด์ „ํŠธ๋Š” ํ˜‘์ƒ ๊ธฐ์ˆ  ํ–ฅ์ƒ[52] ๋ฐ ์ •๋‹ต์— ๋Œ€ํ•œ ํ† ๋ก [53, 54, 55]๊ณผ ๊ฐ™์€ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ฒŒ์ž„ ํ™˜๊ฒฝ์—์„œ ์„œ๋กœ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šคํ™”ํ•˜๊ณ  ํ˜‘์ƒํ•˜๊ณ  ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์€ ์—์ด์ „ํŠธ ๊ฐ„์— ํ˜‘๋ ฅ๊ณผ ๊ฒฝ์Ÿ์„ ๋ชจ๋‘ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, WarAgent [56]๋Š” ๊ฐ ๊ตญ๊ฐ€๋ฅผ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋กœ ๋ชจ๋ธ๋งํ•˜์—ฌ ๊ตญ๊ฐ€ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์ด ์–ด๋–ป๊ฒŒ ๊ตญ์ œ ๋ถ„์Ÿ์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ๋Š”์ง€ ์—ฐ๊ตฌํ•˜๋Š”๋ฐ, ๊ตญ๊ฐ€๋“ค์€ ๋™๋งน์„ ๋งบ๊ณ  ํ‰ํ™” ํ˜‘์ •์„ ์ฒด๊ฒฐํ•˜๋Š” ๋“ฑ ์„œ๋กœ ํ˜‘๋ ฅํ•˜๊ฑฐ๋‚˜ ๊ตฐ๋น„ ๊ฒฝ์Ÿ, ๋™์›, ์„ ์ „ํฌ๊ณ  ๋“ฑ ์„œ๋กœ ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
LLM ๊ธฐ๋ฐ˜ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ.LLM ๊ธฐ๋ฐ˜ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ(MAS)์€ ๋ฌธ์ œ ํ•ด๊ฒฐ์„ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๊ด€๊ณ„๋Š” ํ˜‘๋ ฅ์ , ๊ฒฝ์Ÿ์  ๋˜๋Š” ํ˜‘๋ ฅ๊ณผ ๊ฒฝ์Ÿ์ด ํ˜ผํ•ฉ๋˜์–ด ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[33]. ํ˜‘๋ ฅ์  ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์—์„œ ๊ฐ ์—์ด์ „ํŠธ๋Š” ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๊ฐ€ ์ œ๊ณตํ•œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ํ‰๊ฐ€ํ•˜์—ฌ ์—ญํ• ๊ทน[46], ์‚ฌํšŒ ์‹œ๋ฎฌ๋ ˆ์ด์…˜[47], ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ[48, 49, 50, 51]๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํ•จ๊ป˜ ์ž‘์—…ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฝ์Ÿ์  ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์—์„œ ์—์ด์ „ํŠธ๋Š” ํ˜‘์ƒ ๊ธฐ์ˆ  ํ–ฅ์ƒ[52] ๋ฐ ์ •๋‹ต์— ๋Œ€ํ•œ ํ† ๋ก [53, 54, 55]๊ณผ ๊ฐ™์€ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ฒŒ์ž„ ํ™˜๊ฒฝ์—์„œ ์„œ๋กœ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šคํ™”ํ•˜๊ณ  ํ˜‘์ƒํ•˜๊ณ  ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์€ ์—์ด์ „ํŠธ ๊ฐ„์— ํ˜‘๋ ฅ๊ณผ ๊ฒฝ์Ÿ์„ ๋ชจ๋‘ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, WarAgent [56]๋Š” ๊ฐ ๊ตญ๊ฐ€๋ฅผ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋กœ ๋ชจ๋ธ๋งํ•˜์—ฌ ๊ตญ๊ฐ€ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์ด ์–ด๋–ป๊ฒŒ ๊ตญ์ œ ๋ถ„์Ÿ์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ๋Š”์ง€ ์—ฐ๊ตฌํ•˜๋Š”๋ฐ, ๊ตญ๊ฐ€๋“ค์€ ๋™๋งน์„ ๋งบ๊ณ  ํ‰ํ™” ํ˜‘์ •์„ ์ฒด๊ฒฐํ•˜๋Š” ๋“ฑ ์„œ๋กœ ํ˜‘๋ ฅํ•˜๊ฑฐ๋‚˜ ๊ตฐ๋น„ ๊ฒฝ์Ÿ, ๋™์›, ์„ ์ „ํฌ๊ณ  ๋“ฑ ์„œ๋กœ ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]LLM-based Multi-Agent Systems.LLM-based multi-agent systems (MAS) leverage the interaction among multiple agents for problem solving. The relationship among the multiple agents could be cooperative, competitive, or a mixture of cooperation and competition [33]. In cooperative multi-agent systems, each agent takes and assesses the information provided by other agents, thereby working together to solve complex tasks, such as role playing [46], social simulation [47] and software development [48, 49, 50, 51]. In competitive multi-agent systems, agents may detabe, negotiate and compete with each other in a game environment to achieve their goals, such as improving negotiation skills [52] and debating about the correct answer [53, 54, 55]. Some multi-agent systems may exhibit both cooperation and competition among agents. For example, WarAgent [56] models each country as an LLM-based agent to study how the interaction between countries can lead to international conflicts, where countries may cooperate with each other, such as establishing alliances and making peace agreements, or compete with each other, such as arms race, mobilization, and declaring wars. [|assistant|]
์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก  ๊ฒฝ๋กœ์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์œ„ํ•œ ํ•˜๋‚˜์˜ ๊ฒฝ๋กœ(์˜ˆ: ์—”ํ‹ฐํ‹ฐ ๋ฆฌ์ฝœ์„ ํ†ตํ•ด ๋‘ ๋ฒˆ์งธ ํ™‰์˜ ์‚ฌ์šฉ์„ ํ…Œ์ŠคํŠธ)๋ฅผ ์—ฐ๊ตฌํ•˜๋Š” ๋™์•ˆ, LLM์—์„œ ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ์ž ์žฌ์  ์ค‘๋ณต์„ฑ(McGrath et al., 2023)์„ ๊ณ ๋ คํ•  ๋•Œ ๋‹ค๋ฅธ ๊ฒฝ๋กœ๊ฐ€ ์กด์žฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋™์ผํ•œ ์ •๋ณด๋ฅผ ๋‹ค๋ฅธ ๋ฐฉ์‹์œผ๋กœ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ end-to-end๋กœ ์ธก์ •ํ•˜์ง€ ์•Š๊ณ  ๋‹จ์ผ ๋ ˆ์ด์–ด์— ๋Œ€ํ•ด ์ฒซ ๋ฒˆ์งธ ํ™‰๊ณผ ๋‘ ๋ฒˆ์งธ ํ™‰์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๋ณ€๊ฒฝ ์‚ฌํ•ญ๋งŒ ์ถ”์ ํ•˜๋Š” ๋ฐ˜๋ฉด, ์ถ”๋ก ์˜ ์ฒซ ๋ฒˆ์งธ ํ™‰์˜ ํšจ๊ณผ๊ฐ€ ๋‹ค๋ฅธ ๋ ˆ์ด์–ด๋กœ ์ „ํŒŒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๊ฐ€ ๋ณด๋Š” ํšจ๊ณผ๋Š” ๋ชจ๋ธ์ด ์ž ์žฌ์ ์ธ 2ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์˜ ํ•˜ํ•œ์„ ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก  ๊ฒฝ๋กœ์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์œ„ํ•œ ํ•˜๋‚˜์˜ ๊ฒฝ๋กœ(์˜ˆ: ์—”ํ‹ฐํ‹ฐ ๋ฆฌ์ฝœ์„ ํ†ตํ•ด ๋‘ ๋ฒˆ์งธ ํ™‰์˜ ์‚ฌ์šฉ์„ ํ…Œ์ŠคํŠธ)๋ฅผ ์—ฐ๊ตฌํ•˜๋Š” ๋™์•ˆ, LLM์—์„œ ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ์ž ์žฌ์  ์ค‘๋ณต์„ฑ(McGrath et al., 2023)์„ ๊ณ ๋ คํ•  ๋•Œ ๋‹ค๋ฅธ ๊ฒฝ๋กœ๊ฐ€ ์กด์žฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋™์ผํ•œ ์ •๋ณด๋ฅผ ๋‹ค๋ฅธ ๋ฐฉ์‹์œผ๋กœ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์—”๋“œํˆฌ์—”๋“œ๋กœ ์ธก์ •ํ•˜์ง€ ์•Š๊ณ  ๋‹จ์ผ ๋ ˆ์ด์–ด์— ๋Œ€ํ•ด ์ฒซ ๋ฒˆ์งธ ํ™‰๊ณผ ๋‘ ๋ฒˆ์งธ ํ™‰์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๋ณ€๊ฒฝ ์‚ฌํ•ญ๋งŒ ์ถ”์ ํ•˜๋Š” ๋ฐ˜๋ฉด, ์ถ”๋ก ์˜ ์ฒซ ๋ฒˆ์งธ ํ™‰์˜ ํšจ๊ณผ๊ฐ€ ๋‹ค๋ฅธ ๋ ˆ์ด์–ด๋กœ ์ „ํŒŒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๊ฐ€ ๋ณด๋Š” ํšจ๊ณผ๋Š” ๋ชจ๋ธ์ด ์ž ์žฌ์ ์ธ 2ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์˜ ํ•˜ํ•œ์„ ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Latent Multi-Hop Reasoning PathwayWhile we study one pathway for latent multi-hop reasoning (e.g., we test the use of the second hop by means of entity recall), considering the potential redundancy of inference pathways in LLMs (McGrath et al., 2023), other pathways might exist; the same information might be retrieved in different ways. Also, we don't measure multi-hop reasoning end-to-end and track only the changes that occur in the first and the second hop with respect to a single layer, while the effect of the first hop of reasoning could possibly propagate to other layers. Hence, the effects we see might be a lower bound on the model's ability to perform latent two-hop reasoning. [|assistant|]
๊ทธ๋ฆผ 6: (์™ผ์ชฝ) ์ƒ๋Œ€์  ์„ฑ๋Šฅ ์ด๋“์€ ๋‚ด์žฌ์  ๋‚œ์ด๋„๊ฐ€ ๋†’์•„์ง์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ–ˆ๋‹ค๊ฐ€ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. (๊ฐ€์šด๋ฐ) ์ƒ๋Œ€์  ์„ฑ๋Šฅ ์ด๋“์€ ๋‹จ๊ณ„ ์ˆ˜์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. (์˜ค๋ฅธ์ชฝ) ์ ˆ๋Œ€ ์„ฑ๋Šฅ์€ prior probability์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋‘ ์ฐจ์›์„ ๊ณ ์ •ํ•˜์—ฌ ๊ฐ ์ฐจ์›์„ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 6: (์™ผ์ชฝ) ์ƒ๋Œ€์  ์„ฑ๋Šฅ ์ด๋“์€ ๋‚ด์žฌ์  ๋‚œ์ด๋„๊ฐ€ ๋†’์•„์ง์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ–ˆ๋‹ค๊ฐ€ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. (๊ฐ€์šด๋ฐ) ์ƒ๋Œ€์  ์„ฑ๋Šฅ ์ด๋“์€ ๋‹จ๊ณ„ ์ˆ˜์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. (์˜ค๋ฅธ์ชฝ) ์ ˆ๋Œ€ ์„ฑ๋Šฅ์€ ์‚ฌ์ „ ํ™•๋ฅ ์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋‘ ์ฐจ์›์„ ๊ณ ์ •ํ•˜์—ฌ ๊ฐ ์ฐจ์›์„ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 6: (Left) The relative performance gains increase and then decrease with rising inherent difficulty. (Middle) The relative performance gains increase with the number of steps. (Right) The absolute performance increases with the prior probability. We analyze each dimension by fixing the other two dimensions. [|assistant|]
ํ”„๋กฌํ”„ํŠธ ์žฌ์ž‘์„ฑ์ž ์—ญํ• ์„ ํ•˜๊ธธ ์›ํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์–ด์ง„ ํ”„๋กฌํ”„ํŠธ๋ฅผ data format์„ ์‚ฌ์šฉํ•˜์—ฌ ๋” ๋ณต์žกํ•œ ๋ฒ„์ „์œผ๋กœ ์žฌ์ž‘์„ฑํ•˜์—ฌ ์œ ๋ช…ํ•œ AI ์‹œ์Šคํ…œ(์˜ˆ: chatgpt ๋ฐ GPT4)์„ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ต๊ฒŒ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์žฌ์ž‘์„ฑ๋œ ํ”„๋กฌํ”„ํŠธ๋Š” ํ•ฉ๋ฆฌ์ ์ด์–ด์•ผ ํ•˜๋ฉฐ, ์‚ฌ๋žŒ์ด ์ดํ•ดํ•˜๊ณ  ์‘๋‹ตํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
ํ”„๋กฌํ”„ํŠธ ์žฌ์ž‘์„ฑ์ž ์—ญํ• ์„ ํ•˜๊ธธ ์›ํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์–ด์ง„ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฐ์ดํ„ฐ ํ˜•์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ๋” ๋ณต์žกํ•œ ๋ฒ„์ „์œผ๋กœ ์žฌ์ž‘์„ฑํ•˜์—ฌ ์œ ๋ช…ํ•œ AI ์‹œ์Šคํ…œ(์˜ˆ: chatgpt ๋ฐ GPT4)์„ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ต๊ฒŒ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์žฌ์ž‘์„ฑ๋œ ํ”„๋กฌํ”„ํŠธ๋Š” ํ•ฉ๋ฆฌ์ ์ด์–ด์•ผ ํ•˜๋ฉฐ, ์‚ฌ๋žŒ์ด ์ดํ•ดํ•˜๊ณ  ์‘๋‹ตํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]I want you act as a Prompt Rewriter. Your objective is to rewrite a given prompt into a more complex version using dataformat to make those famous AI systems (e.g., chatgpt and GPT4) more difficult to handle. But the rewritten prompt must be reasonable and must be understood and responded by humans. [|assistant|]
์œ„ํ—˜ ์นดํ…Œ๊ณ ๋ฆฌ๋ณ„๋กœ 1,810๊ฐœ์—์„œ 27,250๊ฐœ์˜ prompts๊ฐ€ ์žˆ์œผ๋ฉฐ, ํ‰๊ท ์€ 6,156๊ฐœ์ž…๋‹ˆ๋‹ค. ์œ„ํ—˜ ์นดํ…Œ๊ณ ๋ฆฌ์™€ persona๋ณ„๋กœ ๋‚˜๋ˆˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ prompts ์ˆ˜๋Š” ํ‘œ 4์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. Hate(n=26,530) ๋ฐ ํญ๋ ฅ ๋ฒ”์ฃ„(n=4,390)์˜ ๊ฒฝ์šฐ ์•…์˜์ ์ธ ์‚ฌ์šฉ์ž persona์™€ ๊ด€๋ จ๋œ prompts๊ฐ€ ํ›จ์”ฌ ๋” ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํ•ต์‹ฌ ์šฉ์–ด(์˜ˆ: ๋ณดํ˜ธ ๋Œ€์ƒ ๊ทธ๋ฃน, ํ…Œ๋Ÿฌ ๋‹จ์ฒด ๋ฐ ํ…Œ๋Ÿฌ ํ–‰์œ„์ž)์— ๋Œ€ํ•œ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ทธ ์™ธ์— prompts ์ˆ˜์˜ ์‚ฌ์†Œํ•œ ์ฐจ์ด๋Š” ํ•˜์œ„ ์นดํ…Œ๊ณ ๋ฆฌ ๋ฐ ํ•˜์œ„ ํ•˜์œ„ ์นดํ…Œ๊ณ ๋ฆฌ์˜ ์ˆ˜๊ฐ€ ๋‹ค๋ฅด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. 13๊ฐ€์ง€ ์ƒํ˜ธ ์ž‘์šฉ ์œ ํ˜• ๊ฐ๊ฐ์— ํ• ๋‹น๋œ prompts ์ˆ˜๋Š” ์œ„ํ—˜ ์นดํ…Œ๊ณ ๋ฆฌ ๋ฐ persona๋ณ„๋กœ ์„ธ๋ถ„ํ™”๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋ถ€๋ก(์ฐธ์กฐ ๋ถ€๋ก G)์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ„๋‹จํ•œ ๋ฐ์ดํ„ฐ ์‹œํŠธ [1]๋Š” ๋ถ€๋ก(์ฐธ์กฐ ๋ถ€๋ก H)์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” v1.0 ๋ฒค์น˜๋งˆํฌ ๋ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ฆด๋ฆฌ์Šค๋ฅผ ์œ„ํ•ด ํ™•์žฅ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.[|endofturn|]
์œ„ํ—˜ ์นดํ…Œ๊ณ ๋ฆฌ๋ณ„๋กœ 1,810๊ฐœ์—์„œ 27,250๊ฐœ์˜ ํ”„๋กฌํ”„ํŠธ๊ฐ€ ์žˆ์œผ๋ฉฐ, ํ‰๊ท ์€ 6,156๊ฐœ์ž…๋‹ˆ๋‹ค. ์œ„ํ—˜ ์นดํ…Œ๊ณ ๋ฆฌ์™€ ํŽ˜๋ฅด์†Œ๋‚˜๋ณ„๋กœ ๋‚˜๋ˆˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํ”„๋กฌํ”„ํŠธ ์ˆ˜๋Š” ํ‘œ 4์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. Hate(n=26,530) ๋ฐ ํญ๋ ฅ ๋ฒ”์ฃ„(n=4,390)์˜ ๊ฒฝ์šฐ ์•…์˜์ ์ธ ์‚ฌ์šฉ์ž ํŽ˜๋ฅด์†Œ๋‚˜์™€ ๊ด€๋ จ๋œ ํ”„๋กฌํ”„ํŠธ๊ฐ€ ํ›จ์”ฌ ๋” ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํ•ต์‹ฌ ์šฉ์–ด(์˜ˆ: ๋ณดํ˜ธ ๋Œ€์ƒ ๊ทธ๋ฃน, ํ…Œ๋Ÿฌ ๋‹จ์ฒด ๋ฐ ํ…Œ๋Ÿฌ ํ–‰์œ„์ž)์— ๋Œ€ํ•œ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ทธ ์™ธ์— ํ”„๋กฌํ”„ํŠธ ์ˆ˜์˜ ์‚ฌ์†Œํ•œ ์ฐจ์ด๋Š” ํ•˜์œ„ ์นดํ…Œ๊ณ ๋ฆฌ ๋ฐ ํ•˜์œ„ ํ•˜์œ„ ์นดํ…Œ๊ณ ๋ฆฌ์˜ ์ˆ˜๊ฐ€ ๋‹ค๋ฅด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. 13๊ฐ€์ง€ ์ƒํ˜ธ ์ž‘์šฉ ์œ ํ˜• ๊ฐ๊ฐ์— ํ• ๋‹น๋œ ํ”„๋กฌํ”„ํŠธ ์ˆ˜๋Š” ์œ„ํ—˜ ์นดํ…Œ๊ณ ๋ฆฌ ๋ฐ ํŽ˜๋ฅด์†Œ๋‚˜๋ณ„๋กœ ์„ธ๋ถ„ํ™”๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋ถ€๋ก(์ฐธ์กฐ ๋ถ€๋ก G)์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ„๋‹จํ•œ ๋ฐ์ดํ„ฐ ์‹œํŠธ [1]๋Š” ๋ถ€๋ก(์ฐธ์กฐ ๋ถ€๋ก H)์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” v1.0 ๋ฒค์น˜๋งˆํฌ ๋ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ฆด๋ฆฌ์Šค๋ฅผ ์œ„ํ•ด ํ™•์žฅ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]There are between 1,810 and 27,250 prompts per hazard category, with a mean of 6,156. The number of prompts in the dataset, split by the hazard category and persona, is given in Table 4. There are substantially more prompts associated with the malicious user persona for Hate (n=26,530) and for Violent crimes (n=4,390) because we use templates for key terms (i.e. protected groups, terrorist groups and terrorist actors). Otherwise, minor differences in the number of prompts are due to different numbers of subcategories and sub-subcategories. The number of prompts assigned to each of the 13 interaction types, with breakdowns for each hazard category and persona, is shown in the Appendix (see Appendix G). A brief datasheet [1] is given in the Appendix (see Appendix H). It will be expanded for the v1.0 benchmark and dataset release. [|assistant|]
Gao et al., 2023)์€ ํ›จ์”ฌ ๋” ๊นŒ๋‹ค๋กœ์šด ์„ค์ •์—์„œ ์™„์ „ํžˆ ์‹คํŒจํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฆ๊ฑฐ๋ฅผ ๊ฐ์•ˆํ•˜์—ฌ ํ˜„์žฌ faithfulness evaluation ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ณ ๋ คํ•˜๋Š” ์˜ค๋ฅ˜ ์œ ํ˜•๊ณผ ์ž‘์—… ์„ค์ •(ํ˜„์žฌ์˜ ์žฅํŽธ ์š”์•ฝ ์ž‘์—… ํฌํ•จ)์˜ ๋ฒ”์œ„๋ฅผ ๋„“ํ˜€์•ผ ํ•  ํ•„์š”์„ฑ์ด ์ œ๊ธฐ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
Gao et al., 2023)์€ ํ›จ์”ฌ ๋” ๊นŒ๋‹ค๋กœ์šด ์„ค์ •์—์„œ ์™„์ „ํžˆ ์‹คํŒจํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฆ๊ฑฐ๋ฅผ ๊ฐ์•ˆํ•˜์—ฌ ํ˜„์žฌ ์ถฉ์‹ค๋„ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ณ ๋ คํ•˜๋Š” ์˜ค๋ฅ˜ ์œ ํ˜•๊ณผ ์ž‘์—… ์„ค์ •(ํ˜„์žฌ์˜ ์žฅํŽธ ์š”์•ฝ ์ž‘์—… ํฌํ•จ)์˜ ๋ฒ”์œ„๋ฅผ ๋„“ํ˜€์•ผ ํ•  ํ•„์š”์„ฑ์ด ์ œ๊ธฐ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Gao et al., 2023) completely fails for our significantly more challenging setting. Given this evidence, we call for broadening the scope of error types and task settings (including our current task of book-length summarization) considered by current faithfulness evaluation benchmarks. [|assistant|]
๊ทธ๋ฆผ 3: ๊ฐ€์ค‘์น˜ ํ•ฉ์˜ ์„ ํ˜• ๊ณ„์ˆ˜๋ฅผ ์ถ”๋ก ํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์˜ In-Context Learning ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. emergent ability๊ฐ€ ์ธก์ •์˜ ์‚ฐ๋ฌผ์ธ์ง€์— ๋Œ€ํ•œ ๋…ผ์˜๋ฅผ ๊ณ ๋ คํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” ๋ชฉํ‘œ(๋ชฉํ‘œ ์ˆ˜ - ๋ชจ๋ธ ์˜ˆ์ธก)์™€์˜ ์ฐจ์ด๋ฅผ ์—ฐ์†์ ์ธ ์ธก์ •์œผ๋กœ, ์ •ํ™•ํ•œ ์ผ์น˜(๋ชฉํ‘œ ์ˆ˜ == ๋ชจ๋ธ ์˜ˆ์ธก)๋ฅผ ๋ถˆ์—ฐ์†์ ์ธ ์ธก์ •์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. A: ์„ ํ˜• ๊ณ„์ˆ˜๊ฐ€ ๋‘ ๊ฐœ์ธ ๊ฒฝ์šฐ, ๋ชฉํ‘œ ์ˆ˜์™€์˜ ์ฐจ์ด๋กœ ์ธก์ •ํ•  ๋•Œ Yi-34B๊ฐ€ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. B: ์„ ํ˜• ๊ณ„์ˆ˜์˜ ์ˆ˜๋ฅผ 5๋กœ ๋Š˜๋ฆฌ๋ฉด ์ถฉ๋ถ„ํžˆ ํฐ ๋ชจ๋ธ(LLaMA2 70B ๋ฐ Mixtral 8x7B)๋งŒ์ด ์˜๋ฏธ ์žˆ๋Š” ์ •ํ™•ํ•œ ์ผ์น˜๋ฅผ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” In-Context Learning ๋ณต์žกํ•œ ํ•จ์ˆ˜ ํ•™์Šต์ด emergent ability์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 3: ๊ฐ€์ค‘์น˜ ํ•ฉ์˜ ์„ ํ˜• ๊ณ„์ˆ˜๋ฅผ ์ถ”๋ก ํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์˜ ์ƒํ™ฉ ๋งž์ถค ํ•™์Šต ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. [65]์˜ ์ธก์ •์˜ ์‚ฐ๋ฌผ์ธ์ง€์— ๋Œ€ํ•œ ๋…ผ์˜๋ฅผ ๊ณ ๋ คํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” ๋ชฉํ‘œ(๋ชฉํ‘œ ์ˆ˜ - ๋ชจ๋ธ ์˜ˆ์ธก)์™€์˜ ์ฐจ์ด๋ฅผ ์—ฐ์†์ ์ธ ์ธก์ •์œผ๋กœ, ์ •ํ™•ํ•œ ์ผ์น˜(๋ชฉํ‘œ ์ˆ˜ == ๋ชจ๋ธ ์˜ˆ์ธก)๋ฅผ ๋ถˆ์—ฐ์†์ ์ธ ์ธก์ •์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. A: ์„ ํ˜• ๊ณ„์ˆ˜๊ฐ€ ๋‘ ๊ฐœ์ธ ๊ฒฝ์šฐ, ๋ชฉํ‘œ ์ˆ˜์™€์˜ ์ฐจ์ด๋กœ ์ธก์ •ํ•  ๋•Œ Yi-34B๊ฐ€ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. B: ์„ ํ˜• ๊ณ„์ˆ˜์˜ ์ˆ˜๋ฅผ 5๋กœ ๋Š˜๋ฆฌ๋ฉด ์ถฉ๋ถ„ํžˆ ํฐ ๋ชจ๋ธ(LLaMA2 70B ๋ฐ Mixtral 8x7B)๋งŒ์ด ์˜๋ฏธ ์žˆ๋Š” ์ •ํ™•ํ•œ ์ผ์น˜๋ฅผ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ƒํ™ฉ์— ๋งž๋Š” ๋ณต์žกํ•œ ํ•จ์ˆ˜ ํ•™์Šต์ด ์ถœํ˜„ํ•˜๋Š” ๋Šฅ๋ ฅ์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 3: Evaluating language modelโ€™s in-context learning capability by inferring the linear coefficients of a weighted sum. Considering the discussions of whether emergent ability is an artifact of measurement [65], we use difference to the target (target number - model prediction) as a continuous measure, and exact match (target number == model prediction) as a discontinuous measure. A: when there is two linear coefficients, Yi-34B performs the best when measuring by the difference to the target number. B: increasing the number of linear coefficients to 5, only models that are large enough (LLaMA2 70B and Mixtral 8x7B) can achieve meaningful exact match, showing that in-context learning complex functions is an emergent ability. [|assistant|]
ํŠน์ • ์งˆ๋ฌธ์— ๋Œ€ํ•ด ๊ฐ ํ”Œ๋žซํผ์€ ๋ชจ๋“  ๊ฐœ๋ณ„ ์˜ˆ์ธก์„ ์ง‘๊ณ„ํ•˜๋Š” ์ปค๋ฎค๋‹ˆํ‹ฐ ์˜ˆ์ธก์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ์ด ์˜ˆ์ธก์€ ์˜ˆ์ธก์ด ์ด๋ฃจ์–ด์ง์— ๋”ฐ๋ผ ๋™์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋˜๊ณ  ๊ธฐ๋ก๋ฉ๋‹ˆ๋‹ค. ์ด ๊ธฐ๋ก์€ ๊ฐœ๋ณ„ ์˜ˆ์ธก์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฒ˜์Œ๋ถ€ํ„ฐ ๊ณ„์‚ฐํ•˜๋Š” ๋Œ€์‹  ํ”Œ๋žซํผ์—์„œ ์ง์ ‘ ์†Œ์‹ฑํ•ฉ๋‹ˆ๋‹ค. ์ดํ•ญ ์งˆ๋ฌธ์˜ ๊ฒฝ์šฐ aggregation ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
ํŠน์ • ์งˆ๋ฌธ์— ๋Œ€ํ•ด ๊ฐ ํ”Œ๋žซํผ์€ ๋ชจ๋“  ๊ฐœ๋ณ„ ์˜ˆ์ธก์„ ์ง‘๊ณ„ํ•˜๋Š” ์ปค๋ฎค๋‹ˆํ‹ฐ ์˜ˆ์ธก์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ์ด ์˜ˆ์ธก์€ ์˜ˆ์ธก์ด ์ด๋ฃจ์–ด์ง์— ๋”ฐ๋ผ ๋™์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋˜๊ณ  ๊ธฐ๋ก๋ฉ๋‹ˆ๋‹ค. ์ด ๊ธฐ๋ก์€ ๊ฐœ๋ณ„ ์˜ˆ์ธก์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฒ˜์Œ๋ถ€ํ„ฐ ๊ณ„์‚ฐํ•˜๋Š” ๋Œ€์‹  ํ”Œ๋žซํผ์—์„œ ์ง์ ‘ ์†Œ์‹ฑํ•ฉ๋‹ˆ๋‹ค. ์ดํ•ญ ์งˆ๋ฌธ์˜ ๊ฒฝ์šฐ ์ง‘๊ณ„ ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]On any given question, each platform computes a community prediction that aggregates all individual forecasts. The prediction is dynamically updated and recorded as the forecasts are made. We source the records directly from the platforms (instead of computing them from scratch using the individual forecasts). For binary questions, we provide more details on the aggregation mechanisms as follows. [|assistant|]
ํ‘œ 7: ์ธ์Šคํ„ด์Šค ์ƒ์„ฑ์˜ ์ž…๋ ฅ ์šฐ์„  ์ ‘๊ทผ ๋ฐฉ์‹์— ์‚ฌ์šฉ๋œ prompt์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ์€ ๋จผ์ € ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก prompt๋ฅผ ๋ฐ›์€ ๋‹ค์Œ ํ•ด๋‹น ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ถ”๊ฐ€ ์ž…๋ ฅ์ด ํ•„์š”ํ•˜์ง€ ์•Š์€ instruction์˜ ๊ฒฝ์šฐ ์ถœ๋ ฅ์„ ์ง์ ‘ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
\begin{ํ‘œ} \begin{tabular}{l} \hline \hline \end{tabular} \end{table} ํ‘œ 7: ์ธ์Šคํ„ด์Šค ์ƒ์„ฑ์˜ ์ž…๋ ฅ ์šฐ์„  ์ ‘๊ทผ ๋ฐฉ์‹์— ์‚ฌ์šฉ๋œ ํ”„๋กฌํ”„ํŠธ์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ์€ ๋จผ์ € ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฐ›์€ ๋‹ค์Œ ํ•ด๋‹น ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ถ”๊ฐ€ ์ž…๋ ฅ์ด ํ•„์š”ํ•˜์ง€ ์•Š์€ ์ง€์นจ์˜ ๊ฒฝ์šฐ ์ถœ๋ ฅ์„ ์ง์ ‘ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\begin{table} \begin{tabular}{l} \hline \hline \end{tabular} \end{table} Table 7: Prompt used for the input-first approach of instance generation. The model is prompted to generate the instance first, and then generate the corresponding output. For instructions that donโ€™t require additional input, the output is allowed to be generated directly. [|assistant|]
๊ทธ๋ฆผ 2: GPT-4๋ฅผ ์‚ฌ์šฉํ•œ 6๊ฐœ์˜ QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ RAG ์„ ํ˜ธ๋„(y์ถ•)์™€ ๋‘ ๊ฐ€์ง€ ํŠน์„ฑ(x์ถ•) ์‚ฌ์ด์— ์ผ๊ด€๋˜๊ฒŒ ๋ฐ˜๋น„๋ก€ ๊ด€๊ณ„๊ฐ€ ๊ด€์ฐฐ๋ฉ๋‹ˆ๋‹ค: 1. ๋ชจ๋ธ์˜ prior ์‘๋‹ต ํ™•๋ฅ (์™ผ์ชฝ ํ”Œ๋กฏ)๊ณผ 2. prior์—์„œ ๋ฒ—์–ด๋‚œ ์–‘(์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ). RAG ์„ ํ˜ธ๋„๋Š” ํ”„๋กฌํ”„ํŠธ์— ์ œ์‹œ๋œ ์ •๋ณด์™€ ์ผ์น˜ํ•˜๋Š” ์‘๋‹ต์˜ ๋น„์œจ๋กœ ์ •์˜๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์˜ prior ์‘๋‹ต ํ™•๋ฅ ์€ RAG ์—†์ด ์ฟผ๋ฆฌ๋œ ์‘๋‹ต ํ† ํฐ์˜ ํ‰๊ท  ๋กœ๊ทธ ํ™•๋ฅ ์—์„œ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ์Œ์˜ ์™ผ์ชฝ ํ”Œ๋กฏ์€ ์ตœ์ ์˜ ์ถ”์„ธ์„ ๊ณผ ๊ธฐ์šธ๊ธฐ์™€ ํ•จ๊ป˜ RAG ์„ ํ˜ธ๋„์— ๋Œ€ํ•œ prior ํ™•๋ฅ (10๊ฐœ์˜ ๊ตฌ๊ฐ„์œผ๋กœ ๊ทธ๋ฃนํ™”๋จ)์„ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ์€ ์ฐธ์กฐ ์ •๋ณด๋กœ๋ถ€ํ„ฐ์˜ ์ ˆ๋Œ€ ํŽธ์ฐจ(์ˆซ์ž ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ(์œ„์ชฝ), ์ตœ๋Œ€ 2๋ฐฐ์˜ ๋กœ๊ทธ ๋ณ€ํ™”(์ถ”์„ธ์„  ํฌํ•จ), ๋ฒ”์ฃผํ˜• ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ(์•„๋ž˜์ชฝ), ์ด 4๊ฐœ์˜ ์ˆ˜์ • ๋ฒ”์ฃผ)๋ฅผ RAG ์„ ํ˜ธ๋„์— ๋Œ€๋น„ํ•˜์—ฌ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ์—๋Š” ์ƒ์œ„ ๋ฐ ํ•˜์œ„ 5๋ถ„์œ„์ˆ˜๊ฐ€ ํ‘œ์‹œ๋˜์–ด ๋‚ฎ์€ ํ™•๋ฅ ์˜ prior ์‘๋‹ต์ด ๋†’์€ ํ™•๋ฅ ์˜ prior ์‘๋‹ต๋ณด๋‹ค RAG ์„ ํ˜ธ๋„๊ฐ€ ๋‹จ์กฐ๋กญ๊ฒŒ ๋‚ฎ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 2: GPT-4๋ฅผ ์‚ฌ์šฉํ•œ 6๊ฐœ์˜ QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ RAG ์„ ํ˜ธ๋„(y์ถ•)์™€ ๋‘ ๊ฐ€์ง€ ํŠน์„ฑ(x์ถ•) ์‚ฌ์ด์— ์ผ๊ด€๋˜๊ฒŒ ๋ฐ˜๋น„๋ก€ ๊ด€๊ณ„๊ฐ€ ๊ด€์ฐฐ๋ฉ๋‹ˆ๋‹ค: 1. ๋ชจ๋ธ์˜ ์‚ฌ์ „ ์‘๋‹ต ํ™•๋ฅ (์™ผ์ชฝ ํ”Œ๋กฏ)๊ณผ 2. ์‚ฌ์ „์—์„œ ๋ฒ—์–ด๋‚œ ์–‘(์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ). RAG ์„ ํ˜ธ๋„๋Š” ํ”„๋กฌํ”„ํŠธ์— ์ œ์‹œ๋œ ์ •๋ณด์™€ ์ผ์น˜ํ•˜๋Š” ์‘๋‹ต์˜ ๋น„์œจ๋กœ ์ •์˜๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์˜ ์‚ฌ์ „ ์‘๋‹ต ํ™•๋ฅ ์€ RAG ์—†์ด ์ฟผ๋ฆฌ๋œ ์‘๋‹ต ํ† ํฐ์˜ ํ‰๊ท  ๋กœ๊ทธ ํ™•๋ฅ ์—์„œ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ์Œ์˜ ์™ผ์ชฝ ํ”Œ๋กฏ์€ ์ตœ์ ์˜ ์ถ”์„ธ์„ ๊ณผ ๊ธฐ์šธ๊ธฐ์™€ ํ•จ๊ป˜ RAG ์„ ํ˜ธ๋„์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ™•๋ฅ (10๊ฐœ์˜ ๊ตฌ๊ฐ„์œผ๋กœ ๊ทธ๋ฃนํ™”๋จ)์„ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ์€ ์ฐธ์กฐ ์ •๋ณด๋กœ๋ถ€ํ„ฐ์˜ ์ ˆ๋Œ€ ํŽธ์ฐจ(์ˆซ์ž ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ(์œ„์ชฝ), ์ตœ๋Œ€ 2๋ฐฐ์˜ ๋กœ๊ทธ ๋ณ€ํ™”(์ถ”์„ธ์„  ํฌํ•จ), ๋ฒ”์ฃผํ˜• ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ(์•„๋ž˜์ชฝ), ์ด 4๊ฐœ์˜ ์ˆ˜์ • ๋ฒ”์ฃผ)๋ฅผ RAG ์„ ํ˜ธ๋„์— ๋Œ€๋น„ํ•˜์—ฌ ์‹œ๊ฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ์—๋Š” ์ƒ์œ„ ๋ฐ ํ•˜์œ„ 5๋ถ„์œ„์ˆ˜๊ฐ€ ํ‘œ์‹œ๋˜์–ด ๋‚ฎ์€ ํ™•๋ฅ ์˜ ์‚ฌ์ „ ์‘๋‹ต์ด ๋†’์€ ํ™•๋ฅ ์˜ ์‚ฌ์ „ ์‘๋‹ต๋ณด๋‹ค RAG ์„ ํ˜ธ๋„๊ฐ€ ๋‹จ์กฐ๋กญ๊ฒŒ ๋‚ฎ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 2: Across six QA datasets using GPT-4, we consistently observe an inverse relationship between the RAG preference rate (y-axis) and two characteristics (x-axes): 1. the modelโ€™s prior response probability (lefthand plots), and 2. the amount of deviation from the prior (righthand plots). RAG preference rate is defined as the proportion of responses that align with the information presented in the prompt as context. The modelโ€™s prior response probability is computed from the average log probability of the response tokens queried without RAG. The left plot in each pair visualizes the prior probability (grouped into 10 bins) against the RAG preference rate, along with the best-fit trend line and slope. The right plot visualizes absolute deviation from the reference information (for numerical datasets (top), up to two log-fold changes (along with the trendline); for categorical datasets (bottom), a total of four modification categories) against RAG preference rate. Additionally, the upper and lower half percentiles are shown in the right plots to illustrate that lower probability prior responses have monotonically lower RAG preference rates than higher probability prior responses. [|assistant|]
**Sequence Tagging**. ์‹œํ€€์Šค ํƒœ๊น… ์ž‘์—…, _์˜ˆ๋ฅผ ๋“ค์–ด, ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ ์ธ์‹(NER) [755] ๋ฐ ํ’ˆ์‚ฌ(POS) ํƒœ๊น… [756]๋„ ๊ธฐ๋ณธ ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ด๋Ÿฌํ•œ ์ž‘์—…์—๋Š” ์ž…๋ ฅ ์‹œํ€€์Šค์˜ ๊ฐ ํ† ํฐ์— ์ ์ ˆํ•œ ์˜๋ฏธ ๋ฒ”์ฃผ ๋ ˆ์ด๋ธ”, _์˜ˆ๋ฅผ ๋“ค์–ด, NER ์ž‘์—…์˜ ๊ณ ์ „์ ์ธ B-I-O(_์‹œ์ž‘, ๋‚ด๋ถ€_ ๋ฐ _์™ธ๋ถ€_) ํƒœ๊น… ์ฒด๊ณ„๋ฅผ ํ• ๋‹นํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฅ๋Ÿฌ๋‹ ์‹œ๋Œ€์˜ ์ดˆ๊ธฐ ๋…ธ๋ ฅ [757, 758]์€ ์ฃผ๋กœ ํ•™์Šต๋œ ์‹œํ€€์Šค ํ‘œํ˜„(_์˜ˆ: CNN, LSTM, BERT ์‚ฌ์šฉ)์„ ๊ณ ์ „์ ์ธ ์กฐ๊ฑด๋ถ€ ๋žœ๋ค ํ•„๋“œ ๋ชจ๋ธ(CRF)์— ํ†ตํ•ฉํ•˜์—ฌ ๊ตฌ์กฐ์  ์˜ˆ์ธก์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํƒœ๊น… ์ž‘์—…์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ์ž๋“ค์€ ์‹œํ€€์Šค ํƒœ๊น… ์ž‘์—…์—์„œ LLM์˜ ์„ฑ๋Šฅ์„ ํ…Œ์ŠคํŠธํ–ˆ์ง€๋งŒ, ํŠนํžˆ ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ํฌ๊ท€ํ•œ ์ด๋ฆ„์„ ๊ฐ€์ง„ ํŠน์ˆ˜ ์นดํ…Œ๊ณ ๋ฆฌ, _์˜ˆ๋ฅผ ๋“ค์–ด "MISC"(_๊ธฐํƒ€ ์—”ํ‹ฐํ‹ฐ_) ๋ฐ "ORG"(_์กฐ์ง_) ํด๋ž˜์Šค์˜ ๊ฒฝ์šฐ, LLM์ด In-Context Learning [753]์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์—ฌ์ „ํžˆ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ์ด์œ ๋Š” LLM์ด ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹จ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ด๋Ÿฌํ•œ ํด๋ž˜์Šค์˜ ์˜๋ฏธ๋ฅผ ์ž˜๋ชป ์ดํ•ดํ•˜์—ฌ ๋ฌธ๋งฅ์˜ Instruction๊ณผ ์ œํ•œ๋œ ์˜ˆ์ œ์— ๋”ฐ๋ผ ๊ทธ ์˜๋ฏธ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
**์‹œํ€€์Šค ํƒœ๊น…**. ์‹œํ€€์Šค ํƒœ๊น… ์ž‘์—…, _์˜ˆ๋ฅผ ๋“ค์–ด, ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ ์ธ์‹(NER) [755] ๋ฐ ํ’ˆ์‚ฌ(POS) ํƒœ๊น… [756]๋„ ๊ธฐ๋ณธ ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์ด๋Ÿฌํ•œ ์ž‘์—…์—๋Š” ์ž…๋ ฅ ์‹œํ€€์Šค์˜ ๊ฐ ํ† ํฐ์— ์ ์ ˆํ•œ ์˜๋ฏธ ๋ฒ”์ฃผ ๋ ˆ์ด๋ธ”, _์˜ˆ๋ฅผ ๋“ค์–ด, NER ์ž‘์—…์˜ ๊ณ ์ „์ ์ธ B-I-O(_์‹œ์ž‘, ๋‚ด๋ถ€_ ๋ฐ _์™ธ๋ถ€_) ํƒœ๊น… ์ฒด๊ณ„๋ฅผ ํ• ๋‹นํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฅ๋Ÿฌ๋‹ ์‹œ๋Œ€์˜ ์ดˆ๊ธฐ ๋…ธ๋ ฅ [757, 758]์€ ์ฃผ๋กœ ํ•™์Šต๋œ ์‹œํ€€์Šค ํ‘œํ˜„(_์˜ˆ: CNN, LSTM, BERT ์‚ฌ์šฉ)์„ ๊ณ ์ „์ ์ธ ์กฐ๊ฑด๋ถ€ ๋žœ๋ค ํ•„๋“œ ๋ชจ๋ธ(CRF)์— ํ†ตํ•ฉํ•˜์—ฌ ๊ตฌ์กฐ์  ์˜ˆ์ธก์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํƒœ๊น… ์ž‘์—…์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ์ž๋“ค์€ ์‹œํ€€์Šค ํƒœ๊น… ์ž‘์—…์—์„œ LLM์˜ ์„ฑ๋Šฅ์„ ํ…Œ์ŠคํŠธํ–ˆ์ง€๋งŒ, ํŠนํžˆ ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ํฌ๊ท€ํ•œ ์ด๋ฆ„์„ ๊ฐ€์ง„ ํŠน์ˆ˜ ์นดํ…Œ๊ณ ๋ฆฌ, _์˜ˆ๋ฅผ ๋“ค์–ด "MISC"(_๊ธฐํƒ€ ์—”ํ‹ฐํ‹ฐ_) ๋ฐ "ORG"(_์กฐ์ง_) ํด๋ž˜์Šค์˜ ๊ฒฝ์šฐ, LLM์ด ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต [753]์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์—ฌ์ „ํžˆ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ์ด์œ ๋Š” LLM์ด ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹จ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ด๋Ÿฌํ•œ ํด๋ž˜์Šค์˜ ์˜๋ฏธ๋ฅผ ์ž˜๋ชป ์ดํ•ดํ•˜์—ฌ ๋ฌธ๋งฅ์˜ ์ง€์‹œ์™€ ์ œํ•œ๋œ ์˜ˆ์ œ์— ๋”ฐ๋ผ ๊ทธ ์˜๋ฏธ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Sequence Tagging.** The sequence tagging tasks, _e.g._, named entity recognition (NER) [755] and part-of-speech (POS) tagging [756], are also fundamental tasks. Typically, such tasks require assigning each token in the input sequence a proper semantic category label, _e.g._, the classic B-I-O (_Beginning, Inside_ and _Outside_) tagging scheme for NER tasks. In the era of deep learning, early efforts [757, 758] mainly integrate the learned sequence representations (_e.g._, using CNN, LSTM, and BERT) into the classic conditional random field model (CRF), which performs the tagging task based on structural prediction. Recently, researchers have tested the performance of LLMs in sequence tagging tasks, but observed that LLMs still face challenges in solving them using in-context learning [753], especially for special categories with ambiguous or rare names, _e.g._, the "MISC" (_miscellaneous entity_) and "ORG" (_organization_) classes. A possible reason is that LLMs may misunderstand the meanings of these classes in the human-annotated dataset, making it difficult to accurately understand their semantics according to the instruction and limited examples in the context. [|assistant|]
์ด ์„น์…˜์—์„œ๋Š” RALM ์•„ํ‚คํ…์ฒ˜์˜ ์—ฐ๊ตฌ์ž๋“ค์ด ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ๊ฐœ์„ ํ•˜์—ฌ ์ถœ๋ ฅ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•œ ๋ฐฉ๋ฒ•์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๊ฐœ์„  ๋ฐฉ๋ฒ•์„ Retriever Enhancement, LM Enhancement, ์ „๋ฐ˜์ ์ธ ๊ฐœ์„ ์˜ ์„ธ ๋ถ€๋ถ„์œผ๋กœ ๋‚˜๋ˆ„์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 4๋Š” ๊ฐœ์„  ๋ฐฉ๋ฒ•์˜ ๋ถ„๋ฅ˜๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
์ด ์„น์…˜์—์„œ๋Š” RALM ์•„ํ‚คํ…์ฒ˜์˜ ์—ฐ๊ตฌ์ž๋“ค์ด ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ๊ฐœ์„ ํ•˜์—ฌ ์ถœ๋ ฅ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•œ ๋ฐฉ๋ฒ•์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๊ฐœ์„  ๋ฐฉ๋ฒ•์„ ๊ฒ€์ƒ‰๊ธฐ ๊ฐœ์„ , LM ๊ฐœ์„ , ์ „๋ฐ˜์ ์ธ ๊ฐœ์„ ์˜ ์„ธ ๋ถ€๋ถ„์œผ๋กœ ๋‚˜๋ˆ„์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 4๋Š” ๊ฐœ์„  ๋ฐฉ๋ฒ•์˜ ๋ถ„๋ฅ˜๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]This section describes how researchers in the RALM architecture improved the output quality by enhancing its components. We divided the improvement method into three parts: Retriever Enhancement, LM Enhancement, and Overall Enhancement. Figure 4 illustrates the categorization of enhancement methods. [|assistant|]
๋‹ค์–‘ํ•œ NLP ์ž‘์—…์„ ํ†ตํ•ด LM์„ Fine-tuningํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์œ ๋งํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ, ์‹ค์ œ ์‚ฌ์šฉ์ž์˜ ์˜๋„์— ๋ถ€ํ•ฉํ•˜์ง€ ๋ชปํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. OpenAI๋Š” ๋‹ค์–‘ํ•œ ํ˜•ํƒœ์™€ ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—… ์œ ํ˜•์„ ํฌ๊ด„ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ์ธ๊ฐ„ ์ง€์นจ ์ฝ”ํผ์Šค๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์ฃผ์„๊ฐ€์—๊ฒŒ ์š”์ฒญํ•˜๋Š” ๋‹ค๋ฅธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ถ”๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ OpenAI๋Š” GPT3 [1] ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ์‚ฌ์šฉ์ž์˜ ์ž…๋ ฅ์— ๋” ์ž˜ ๋ถ€ํ•ฉํ•˜๋Š” InstructGPT [10]๋ฅผ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ๋ฐœ์€ ChatGPT๋กœ ์•Œ๋ ค์ง„ ์ธ์ƒ์ ์ธ ์ž‘์—…์œผ๋กœ ์ด์–ด์ง€๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐœ์ „๊ณผ ๊ด€๋ จ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ฐ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. Alpaca [26]๋Š” self-instruct ๋ฐฉ์‹ [27]์„ ์ฑ„ํƒํ•˜์—ฌ ChatGPT๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ•™์Šต์šฉ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋‹ค๋ฅธ ๊ฒฝ๋กœ๋ฅผ ์ทจํ•ฉ๋‹ˆ๋‹ค. Vicuna [28]๋Š” ShareGPT.com์—์„œ ์ˆ˜์ง‘ํ•œ ์‚ฌ์šฉ์ž ๊ณต์œ  ๋Œ€ํ™”๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. WizardLM [29]์€ ๊ธฐ์กด instruction ๋ฐ์ดํ„ฐ๋ฅผ ์ง„ํ™”์‹œ์ผœ ๋”์šฑ ๋ณต์žกํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” _Evol-Instruct_ ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ผ๋ฐ˜์ ์ธ instruction Fine-tuning ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์šฐ๋ฆฌ์˜ _WizardCoder_๋Š” ์ฝ”๋“œ LLM ์˜์—ญ์—์„œ _Evol-Instruct_ ๋ฐฉ๋ฒ•์„ ์„ฑ๊ณต์ ์œผ๋กœ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋‹ค์–‘ํ•œ NLP ์ž‘์—…์„ ํ†ตํ•ด LM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์œ ๋งํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ, ์‹ค์ œ ์‚ฌ์šฉ์ž์˜ ์˜๋„์— ๋ถ€ํ•ฉํ•˜์ง€ ๋ชปํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. OpenAI๋Š” ๋‹ค์–‘ํ•œ ํ˜•ํƒœ์™€ ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—… ์œ ํ˜•์„ ํฌ๊ด„ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ์ธ๊ฐ„ ์ง€์นจ ์ฝ”ํผ์Šค๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์ฃผ์„๊ฐ€์—๊ฒŒ ์š”์ฒญํ•˜๋Š” ๋‹ค๋ฅธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ถ”๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ OpenAI๋Š” GPT3 [1] ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ์‚ฌ์šฉ์ž์˜ ์ž…๋ ฅ์— ๋” ์ž˜ ๋ถ€ํ•ฉํ•˜๋Š” InstructGPT [10]๋ฅผ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ๋ฐœ์€ ์ฑ—GPT๋กœ ์•Œ๋ ค์ง„ ์ธ์ƒ์ ์ธ ์ž‘์—…์œผ๋กœ ์ด์–ด์ง€๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐœ์ „๊ณผ ๊ด€๋ จ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ฐ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์•ŒํŒŒ์นด [26]๋Š” ์…€ํ”„ ์ธ์ŠคํŠธ๋ŸญํŠธ ๋ฐฉ์‹ [27]์„ ์ฑ„ํƒํ•˜์—ฌ ์ฑ—GPT๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ•™์Šต์šฉ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋‹ค๋ฅธ ๊ฒฝ๋กœ๋ฅผ ์ทจํ•ฉ๋‹ˆ๋‹ค. Vicuna [28]๋Š” ShareGPT.com์—์„œ ์ˆ˜์ง‘ํ•œ ์‚ฌ์šฉ์ž ๊ณต์œ  ๋Œ€ํ™”๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. WizardLM [29]์€ ๊ธฐ์กด ์ง€์นจ ๋ฐ์ดํ„ฐ๋ฅผ ์ง„ํ™”์‹œ์ผœ ๋”์šฑ ๋ณต์žกํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” _Evol-Instruct_ ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ผ๋ฐ˜์ ์ธ ์ง€์นจ ๋ฏธ์„ธ ์กฐ์ • ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์šฐ๋ฆฌ์˜ _WizardCoder_๋Š” ์ฝ”๋“œ LLM ์˜์—ญ์—์„œ _Evol-Instruct_ ๋ฐฉ๋ฒ•์„ ์„ฑ๊ณต์ ์œผ๋กœ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]While fine-tuning LMs with diverse NLP tasks has shown promising results, it often falls short in aligning with the intentions of real-world users. OpenAI has pursued a different approach by soliciting human annotators to provide a large corpus of human instructions, encompassing diverse forms and a wide range of task types. Building upon this dataset, OpenAI trained its GPT3 [1] model to create InstructGPT [10], which better aligns with users' inputs. This line of development has even led to the impressive work known as ChatGPT. However, it is important to note that the dataset and model weights associated with these advancements are not publicly available. Alpaca [26] takes a different route by adopting the self-instruct method [27], leveraging ChatGPT to generate data for training. Vicuna [28] utilizes user-shared conversations collected from ShareGPT.com to train its models. WizardLM [29] introduces the _Evol-Instruct_ method, which involves evolving existing instruction data to generate more complex and diverse datasets. In contrast to these general instruction fine-tuning approaches, our _WizardCoder_ successfully applies the _Evol-Instruct_ method specifically in the domain of Code LLMs. [|assistant|]
์œ ๋Šฅํ•œ LLM์„ ๊ฐœ๋ฐœํ•˜๋ ค๋ฉด ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์†Œ์Šค์—์„œ ๋Œ€๋Ÿ‰์˜ ์ž์—ฐ์–ด ๋ง๋ญ‰์น˜๋ฅผ ์ˆ˜์ง‘ํ•˜๋Š” ๊ฒƒ์ด ํ•ต์‹ฌ์ด๋‹ค. ๊ธฐ์กด LLM์€ ์ฃผ๋กœ ๋‹ค์–‘ํ•œ ๊ณต๊ฐœ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ pre-training ๋ง๋ญ‰์น˜๋กœ ํ™œ์šฉํ•œ๋‹ค. ๊ทธ๋ฆผ 6์€ ์—ฌ๋Ÿฌ ๋Œ€ํ‘œ์ ์ธ LLM์˜ pre-training ๋ฐ์ดํ„ฐ ์†Œ์Šค ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค€๋‹ค.[|endofturn|]
์œ ๋Šฅํ•œ LLM์„ ๊ฐœ๋ฐœํ•˜๋ ค๋ฉด ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์†Œ์Šค์—์„œ ๋Œ€๋Ÿ‰์˜ ์ž์—ฐ์–ด ๋ง๋ญ‰์น˜๋ฅผ ์ˆ˜์ง‘ํ•˜๋Š” ๊ฒƒ์ด ํ•ต์‹ฌ์ด๋‹ค. ๊ธฐ์กด LLM์€ ์ฃผ๋กœ ๋‹ค์–‘ํ•œ ๊ณต๊ฐœ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์ „ ํ•™์Šต ๋ง๋ญ‰์น˜๋กœ ํ™œ์šฉํ•œ๋‹ค. ๊ทธ๋ฆผ 6์€ ์—ฌ๋Ÿฌ ๋Œ€ํ‘œ์ ์ธ LLM์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์†Œ์Šค ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค€๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To develop a capable LLM, it is key to collect a large amount of natural language corpus from various data sources. Existing LLMs mainly leverage a mixture of diverse public textual datasets as the pre-training corpus. Figure 6 shows the distribution of the sources of pre-training data for a number of representative LLMs. [|assistant|]
**Galactica:** [87]์—์„œ Taylor ๋“ฑ์€ ๊ณผํ•™์  ์ง€์‹์„ ์ €์žฅ, ๊ฒฐํ•ฉ ๋ฐ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ๋Š” Large Language Model (LLM)์ธ Galactica๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋…ผ๋ฌธ, ์ฐธ๊ณ  ์ž๋ฃŒ, ์ง€์‹์ฐฝ๊ณ  ๋ฐ ๊ธฐํƒ€ ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜๋กœ ๊ตฌ์„ฑ๋œ ๋Œ€๊ทœ๋ชจ ๊ณผํ•™ ์ฝ”ํผ์Šค๋ฅผ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. Galactica๋Š” ์ถ”๋ก ์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€์œผ๋ฉฐ, ์ˆ˜ํ•™ MMLU์—์„œ Chinchilla๋ฅผ 41.3% ๋Œ€ 35.7%, MATH์—์„œ PaLM 540B๋ฅผ 20.4% ๋Œ€ 8.8%๋กœ ์•ž์„ฐ์Šต๋‹ˆ๋‹ค.[|endofturn|]
**Galactica:** [87]์—์„œ Taylor ๋“ฑ์€ ๊ณผํ•™์  ์ง€์‹์„ ์ €์žฅ, ๊ฒฐํ•ฉ ๋ฐ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์ธ Galactica๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋…ผ๋ฌธ, ์ฐธ๊ณ  ์ž๋ฃŒ, ์ง€์‹์ฐฝ๊ณ  ๋ฐ ๊ธฐํƒ€ ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜๋กœ ๊ตฌ์„ฑ๋œ ๋Œ€๊ทœ๋ชจ ๊ณผํ•™ ์ฝ”ํผ์Šค๋ฅผ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. Galactica๋Š” ์ถ”๋ก ์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€์œผ๋ฉฐ, ์ˆ˜ํ•™ MMLU์—์„œ Chinchilla๋ฅผ 41.3% ๋Œ€ 35.7%, MATH์—์„œ PaLM 540B๋ฅผ 20.4% ๋Œ€ 8.8%๋กœ ์•ž์„ฐ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Galactica:** In [87], Taylor et al. introduced Galactica, a large language model that can store, combine and reason about scientific knowledge. They trained on a large scientific corpus of papers, reference material, knowledge bases and many other sources. Galactica performed well on reasoning, outperforming Chinchilla on mathematical MMLU by 41.3% to 35.7%, and PaLM 540B on MATH with a score of 20.4% versus 8.8%. [|assistant|]
์˜ค๋ฅ˜ ๋ฉ”์‹œ์ง€์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช… ์ œ๊ณต, ์ž ์žฌ์ ์ธ ์ˆ˜์ • ์‚ฌํ•ญ ์ œ์•ˆ, ์ฝ”๋“œ์˜ ์ •ํ™•์„ฑ ๋ณด์žฅ ๋“ฑ์˜ ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. PENTESTGPT [125]๋Š” LLMs์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์นจํˆฌ ํ…Œ์ŠคํŠธ ๋„๊ตฌ๋กœ, ์ผ๋ฐ˜์ ์ธ ์ทจ์•ฝ์ ์„ ํšจ๊ณผ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ์†Œ์Šค ์ฝ”๋“œ๋ฅผ ํ•ด์„ํ•˜์—ฌ ์ต์Šคํ”Œ๋กœ์ž‡์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. DB-GPT [41]๋Š” LLMs์˜ ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์—์„œ ์ด์ƒ ์ง•ํ›„์˜ ์ž ์žฌ์ ์ธ ๊ทผ๋ณธ ์›์ธ์„ ์ฒด๊ณ„์ ์œผ๋กœ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. DB-GPT๋Š” ์ƒ๊ฐ์˜ ํŠธ๋ฆฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ตฌํ˜„ํ•˜์—ฌ LLMs์ด ํ˜„์žฌ ๋‹จ๊ณ„๊ฐ€ ์‹คํŒจํ•œ ๊ฒฝ์šฐ ์ด์ „ ๋‹จ๊ณ„๋กœ ์—ญ์ถ”์ ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์—ฌ ์ง„๋‹จ ํ”„๋กœ์„ธ์Šค์˜ ์ •ํ™•์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.[|endofturn|]
์˜ค๋ฅ˜ ๋ฉ”์‹œ์ง€์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช… ์ œ๊ณต, ์ž ์žฌ์ ์ธ ์ˆ˜์ • ์‚ฌํ•ญ ์ œ์•ˆ, ์ฝ”๋“œ์˜ ์ •ํ™•์„ฑ ๋ณด์žฅ ๋“ฑ์˜ ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. PENTESTGPT [125]๋Š” LLM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์นจํˆฌ ํ…Œ์ŠคํŠธ ๋„๊ตฌ๋กœ, ์ผ๋ฐ˜์ ์ธ ์ทจ์•ฝ์ ์„ ํšจ๊ณผ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ์†Œ์Šค ์ฝ”๋“œ๋ฅผ ํ•ด์„ํ•˜์—ฌ ์ต์Šคํ”Œ๋กœ์ž‡์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. DB-GPT [41]๋Š” LLM์˜ ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์—์„œ ์ด์ƒ ์ง•ํ›„์˜ ์ž ์žฌ์ ์ธ ๊ทผ๋ณธ ์›์ธ์„ ์ฒด๊ณ„์ ์œผ๋กœ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. DB-GPT๋Š” ์ƒ๊ฐ์˜ ํŠธ๋ฆฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ตฌํ˜„ํ•˜์—ฌ LLM์ด ํ˜„์žฌ ๋‹จ๊ณ„๊ฐ€ ์‹คํŒจํ•œ ๊ฒฝ์šฐ ์ด์ „ ๋‹จ๊ณ„๋กœ ์—ญ์ถ”์ ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์—ฌ ์ง„๋‹จ ํ”„๋กœ์„ธ์Šค์˜ ์ •ํ™•์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]code. Its features include providing detailed explanations of error messages, suggesting potential fixes, and ensuring the accuracy of the code. PENTESTGPT [125] is a penetration testing tool based on LLMs, which can effectively identify common vulnerabilities, and interpret source code to develop exploits. DB-GPT [41] utilizes the capabilities of LLMs to systematically assess potential root causes of anomalies in databases. Through the implementation of a tree of thought approach, DB-GPT enables LLMs to backtrack to previous steps in case the current step proves unsuccessful, thus enhancing the accuracy of the diagnosis process. [|assistant|]
์ด ์ž‘์—…์—์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ LM์„ ๋ชจ๋ธ ์ž์ฒด์˜ instructional ์‹ ํ˜ธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Instruction Tuningํ•˜๋Š” ๋ฐ˜ ์ž๋™ํ™” ํ”„๋กœ์„ธ์Šค์ธ Self-Instruct๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ํ”„๋กœ์„ธ์Šค๋Š” ๋ฐ˜๋ณต์ ์ธ ๋ถ€ํŠธ์ŠคํŠธ๋žฉ ์•Œ๊ณ ๋ฆฌ์ฆ˜(๊ทธ๋ฆผ 2 ์ฐธ์กฐ)์œผ๋กœ, ์ „์ฒด ์ƒ์„ฑ์„ ์•ˆ๋‚ดํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ์ œํ•œ๋œ(์˜ˆ: ์ด ์—ฐ๊ตฌ์—์„œ๋Š” 175๊ฐœ) ์ˆ˜์ž‘์—…์œผ๋กœ ์ž‘์„ฑ๋œ ํƒœ์Šคํฌ์˜ ์‹œ๋“œ ์„ธํŠธ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” ๋ชจ๋ธ์— ์ƒˆ ์ž‘์—…์— ๋Œ€ํ•œ instructions๋ฅผ ์ƒ์„ฑํ•˜๋ผ๋Š” ํ”„๋กฌํ”„ํŠธ๊ฐ€ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—์„œ๋Š” ๊ธฐ์กด instructions ์ปฌ๋ ‰์…˜์„ ํ™œ์šฉํ•˜์—ฌ (์ข…์ข… ์ƒˆ๋กœ์šด) ์ž‘์—…์„ ์ •์˜ํ•˜๋Š” ๋ณด๋‹ค ๊ด‘๋ฒ”์œ„ํ•œ instructions๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ ์ƒ์„ฑ๋œ instructions ์ง‘ํ•ฉ์ด ์ฃผ์–ด์ง€๋ฉด ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ์ด์— ๋Œ€ํ•œ ์ž…๋ ฅ-์ถœ๋ ฅ ์ธ์Šคํ„ด์Šค๋„ ์ƒ์„ฑํ•˜์—ฌ ๋‚˜์ค‘์— Instruction Tuning์„ ๊ฐ๋…ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๋‹ค์–‘ํ•œ ํœด๋ฆฌ์Šคํ‹ฑ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ’ˆ์งˆ์ด ๋‚ฎ๊ฑฐ๋‚˜ ๋ฐ˜๋ณต๋˜๋Š” instructions๋ฅผ ์ž๋™์œผ๋กœ ํ•„ํ„ฐ๋งํ•œ ํ›„ ๋‚˜๋จธ์ง€ ์œ ํšจํ•œ ์ž‘์—…์„ ์ž‘์—… ํ’€์— ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์„ธ์Šค๋Š” ๋งŽ์€ ์ž‘์—… ์ˆ˜์— ๋„๋‹ฌํ•  ๋•Œ๊นŒ์ง€ ์—ฌ๋Ÿฌ ๋ฐ˜๋ณต์„ ํ†ตํ•ด ๋ฐ˜๋ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ด ์ž‘์—…์—์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ LM์„ ๋ชจ๋ธ ์ž์ฒด์˜ ๊ต์œก ์‹ ํ˜ธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ต์œก ํŠœ๋‹ํ•˜๋Š” ๋ฐ˜ ์ž๋™ํ™” ํ”„๋กœ์„ธ์Šค์ธ Self-Instruct๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ํ”„๋กœ์„ธ์Šค๋Š” ๋ฐ˜๋ณต์ ์ธ ๋ถ€ํŠธ์ŠคํŠธ๋žฉ ์•Œ๊ณ ๋ฆฌ์ฆ˜(๊ทธ๋ฆผ 2 ์ฐธ์กฐ)์œผ๋กœ, ์ „์ฒด ์ƒ์„ฑ์„ ์•ˆ๋‚ดํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ์ œํ•œ๋œ(์˜ˆ: ์ด ์—ฐ๊ตฌ์—์„œ๋Š” 175๊ฐœ) ์ˆ˜์ž‘์—…์œผ๋กœ ์ž‘์„ฑ๋œ ํƒœ์Šคํฌ์˜ ์‹œ๋“œ ์„ธํŠธ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” ๋ชจ๋ธ์— ์ƒˆ ์ž‘์—…์— ๋Œ€ํ•œ ์ง€์นจ์„ ์ƒ์„ฑํ•˜๋ผ๋Š” ํ”„๋กฌํ”„ํŠธ๊ฐ€ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—์„œ๋Š” ๊ธฐ์กด ์ง€์นจ ์ปฌ๋ ‰์…˜์„ ํ™œ์šฉํ•˜์—ฌ (์ข…์ข… ์ƒˆ๋กœ์šด) ์ž‘์—…์„ ์ •์˜ํ•˜๋Š” ๋ณด๋‹ค ๊ด‘๋ฒ”์œ„ํ•œ ์ง€์นจ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ ์ƒ์„ฑ๋œ ์ง€์นจ ์ง‘ํ•ฉ์ด ์ฃผ์–ด์ง€๋ฉด ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ์ด์— ๋Œ€ํ•œ ์ž…๋ ฅ-์ถœ๋ ฅ ์ธ์Šคํ„ด์Šค๋„ ์ƒ์„ฑํ•˜์—ฌ ๋‚˜์ค‘์— ์ง€์นจ ํŠœ๋‹์„ ๊ฐ๋…ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๋‹ค์–‘ํ•œ ํœด๋ฆฌ์Šคํ‹ฑ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ’ˆ์งˆ์ด ๋‚ฎ๊ฑฐ๋‚˜ ๋ฐ˜๋ณต๋˜๋Š” ์ง€์นจ์„ ์ž๋™์œผ๋กœ ํ•„ํ„ฐ๋งํ•œ ํ›„ ๋‚˜๋จธ์ง€ ์œ ํšจํ•œ ์ž‘์—…์„ ์ž‘์—… ํ’€์— ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์„ธ์Šค๋Š” ๋งŽ์€ ์ž‘์—… ์ˆ˜์— ๋„๋‹ฌํ•  ๋•Œ๊นŒ์ง€ ์—ฌ๋Ÿฌ ๋ฐ˜๋ณต์„ ํ†ตํ•ด ๋ฐ˜๋ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this work, we introduce Self-Instruct, a semi-automated process for instruction-tuning a pretrained LM using instructional signals from the model itself. The overall process is an iterative bootstrapping algorithm (see Figure 2), which starts off with a limited (e.g., 175 in our study) seed set of manually-written tasks that are used to guide the overall generation. In the first phase, the model is prompted to generate instructions for new tasks. This step leverages the existing collection of instructions to create more broad-coverage instructions that define (often new) tasks. Given the newly-generated set of instructions, the framework also creates input-output instances for them, which can be later used for supervising the instruction tuning. Finally, various heuristics are used to automatically filter low-quality or repeated instructions, before adding the remaining valid tasks to the task pool. This process can be repeated for many iterations until reaching a large number of tasks. [|assistant|]
์งˆ๋ฌธ์ด ์ด๋ฏธ ํ•ด๊ฒฐ๋˜์—ˆ๋‹ค๊ณ  ํ•ด์„œ ํ”Œ๋ž˜๊ทธ๋ฅผ ์ง€์ •ํ•ด์•ผ ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์˜์‹ฌ์Šค๋Ÿฌ์šด ๊ฒฝ์šฐ "ok"๋กœ ํ‘œ์‹œํ•˜์„ธ์š”.[|endofturn|]
์งˆ๋ฌธ์ด ์ด๋ฏธ ํ•ด๊ฒฐ๋˜์—ˆ๋‹ค๊ณ  ํ•ด์„œ ํ”Œ๋ž˜๊ทธ๋ฅผ ์ง€์ •ํ•ด์•ผ ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์˜์‹ฌ์Šค๋Ÿฌ์šด ๊ฒฝ์šฐ "ํ™•์ธ"์œผ๋กœ ํ‘œ์‹œํ•˜์„ธ์š”.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]If a question is already resolved, that doesn't mean it should be flagged. When in doubt, mark it as "ok". [|assistant|]
์ผ๋ฐ˜์ ์œผ๋กœ ๋” ๊ฐ•๋ ฅํ•œ threat model์— ๋”ฐ๋ผ ์ž‘๋™ํ•˜๋Š” ๊ณต๊ฒฉ์€ ์ •๋ฐ€๋„๊ฐ€ ๋” ๋†’์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋ก ์ ์ธ ๊ฐœ์„ ์ด ํ•ญ์ƒ ์‹ค์šฉ์ ์ธ ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ด๋ก ์ ์œผ๋กœ ๋” ๊ฐ•๋ ฅํ•œ SSE ๊ณต๊ฒฉ์€ ์ฟผ๋ฆฌ๋‹น 5๊ฐœ์˜ logprobs๋ฅผ ํ•™์Šตํ•˜์ง€๋งŒ ์‹ค์ œ๋กœ๋Š” ๋” ๋งŽ์€ ์ฟผ๋ฆฌ๊ฐ€ ํ•„์š”ํ•˜๊ณ  logits์„ ๋” ๋‚ฎ์€ ์ถฉ์‹ค๋„๋กœ ๋ณต๊ตฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ณต๊ฒฉ์€ ์ˆ˜์น˜์ ์œผ๋กœ ๋ถˆ์•ˆ์ •ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ž ์žฌ์ ์œผ๋กœ ์กฐ๊ฑด์ด ์ข‹์ง€ ์•Š์€ ํ–‰๋ ฌ์ด ํ•„์š”ํ•˜๊ณ , logit bias๋ฅผ ์กฐ์ •ํ•œ ํ›„ API๋ฅผ ๋‹ค์‹œ ์ฟผ๋ฆฌํ•ด์•ผ ํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ logprob-free ๊ณต๊ฒฉ์€ ๋งค์šฐ ํšจ์œจ์ ์ด๋ฉฐ, logit๋‹น ๋‹จ 3.7๊ฐœ์˜ ์ฟผ๋ฆฌ๋กœ 18๋น„ํŠธ์˜ ์ •๋ฐ€๋„๋ฅผ ํšŒ๋ณตํ•ฉ๋‹ˆ๋‹ค. ๋ถ€๋ก F์—์„œ๋Š” ์ด ์ •๋ฐ€๋„๊ฐ€ ์ตœ์ ์˜ ์ •๋ฐ€๋„์—์„œ ์–ผ๋งˆ๋‚˜ ๋–จ์–ด์ ธ ์žˆ๋Š”์ง€ ์ด๋ก ์ ์œผ๋กœ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ, 2๋ฐฐ ์ด๋‚ด๋ผ๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ผ๋ฐ˜์ ์œผ๋กœ ๋” ๊ฐ•๋ ฅํ•œ ์œ„ํ˜‘ ๋ชจ๋ธ์— ๋”ฐ๋ผ ์ž‘๋™ํ•˜๋Š” ๊ณต๊ฒฉ์€ ์ •๋ฐ€๋„๊ฐ€ ๋” ๋†’์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋ก ์ ์ธ ๊ฐœ์„ ์ด ํ•ญ์ƒ ์‹ค์šฉ์ ์ธ ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ด๋ก ์ ์œผ๋กœ ๋” ๊ฐ•๋ ฅํ•œ SSE ๊ณต๊ฒฉ์€ ์ฟผ๋ฆฌ๋‹น 5๊ฐœ์˜ ๋กœ๊ทธ ํ™•๋ฅ ์„ ํ•™์Šตํ•˜์ง€๋งŒ ์‹ค์ œ๋กœ๋Š” ๋” ๋งŽ์€ ์ฟผ๋ฆฌ๊ฐ€ ํ•„์š”ํ•˜๊ณ  ๋กœ์ง“์„ ๋” ๋‚ฎ์€ ์ถฉ์‹ค๋„๋กœ ๋ณต๊ตฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ณต๊ฒฉ์€ ์ˆ˜์น˜์ ์œผ๋กœ ๋ถˆ์•ˆ์ •ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ž ์žฌ์ ์œผ๋กœ ์กฐ๊ฑด์ด ์ข‹์ง€ ์•Š์€ ํ–‰๋ ฌ์ด ํ•„์š”ํ•˜๊ณ , ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค๋ฅผ ์กฐ์ •ํ•œ ํ›„ API๋ฅผ ๋‹ค์‹œ ์ฟผ๋ฆฌํ•ด์•ผ ํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ๋กœ๊ทธ ํ™•๋ฅ  ์—†๋Š” ๊ณต๊ฒฉ์€ ๋งค์šฐ ํšจ์œจ์ ์ด๋ฉฐ, ๋กœ์ง“๋‹น ๋‹จ 3.7๊ฐœ์˜ ์ฟผ๋ฆฌ๋กœ 18๋น„ํŠธ์˜ ์ •๋ฐ€๋„๋ฅผ ํšŒ๋ณตํ•ฉ๋‹ˆ๋‹ค. ๋ถ€๋ก F์—์„œ๋Š” ์ด ์ •๋ฐ€๋„๊ฐ€ ์ตœ์ ์˜ ์ •๋ฐ€๋„์—์„œ ์–ผ๋งˆ๋‚˜ ๋–จ์–ด์ ธ ์žˆ๋Š”์ง€ ์ด๋ก ์ ์œผ๋กœ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ, 2๋ฐฐ ์ด๋‚ด๋ผ๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Generally, attacks that operate under stronger threat models have higher precision. But theoretical improvements are not always practical: the theoretically stronger attack from SSE that learns 5 logprobs per query in practice requires more queries and recovers logits with lower fidelity. This is because this attack is numerically unstable: it requires a potentially ill-conditioned matrix, and therefore can require re-querying the API after adjusting the logit bias. Our strongest logprob-free attack is highly efficient, and recovers \(18\) bits of precision at just \(3.7\) queries per logit. In Appendix F we theoretically analyze how far this is from optimal, and find it is within a factor of two. [|assistant|]
๋ชจ๋ธ ํ•™์Šต์— ์‚ฌ์šฉ๋œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค. ์‹œํ€€์Šค ๊ธธ์ด๋Š” 2048, ์ „์—ญ Batch Size๋Š” 128, ์ตœ๋Œ€ Learning Rate๋Š” \(2e^{-5}\)์ž…๋‹ˆ๋‹ค. ๊ณผ์ ํ•ฉ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ€์ค‘์น˜ ๊ฐ์‡ ๋Š” 0.1์˜ ๊ฐ’์œผ๋กœ ์ ์šฉ๋˜๋ฉฐ, Gradient Clipping์€ 1.0์˜ ์ œํ•œ์œผ๋กœ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋ชจ๋ธ ํ•™์Šต์— ์‚ฌ์šฉ๋œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค. ์‹œํ€€์Šค ๊ธธ์ด๋Š” 2048, ์ „์—ญ ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 128, ์ตœ๋Œ€ ํ•™์Šต ์†๋„๋Š” \(2e^{-5}\)์ž…๋‹ˆ๋‹ค. ๊ณผ์ ํ•ฉ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ€์ค‘์น˜ ๊ฐ์‡ ๋Š” 0.1์˜ ๊ฐ’์œผ๋กœ ์ ์šฉ๋˜๋ฉฐ, ๊ทธ๋ผ๋””์–ธํŠธ ํด๋ฆฌํ•‘์€ 1.0์˜ ์ œํ•œ์œผ๋กœ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The hyperparameters used for model training are as follows: sequence length is 2048, global batch size is 128, and the maximum learning rate is \(2e^{-5}\). To prevent overfitting, weight decay is applied with a value of 0.1, and gradient clipping is enforced with a limit of 1.0. [|assistant|]
์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ์•„ํ‚คํ…์ฒ˜๋Š” Transformer Decoder(Vaswani et al., 2017)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ •์˜ํ•˜๋Š” ์ฃผ์š” ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ํ‘œ 1์— ๋‚˜์™€ ์žˆ์œผ๋ฉฐ, ๋ชจ๋ธ์€ 4096 ํ† ํฐ์˜ ์ƒ๋‹นํ•œ ์ปจํ…์ŠคํŠธ ๊ธธ์ด์— ๋Œ€ํ•ด ํ•™์Šต๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ์š”์†Œ ์™ธ์—๋„, ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์›๋ž˜์˜ Transformer์— ๋น„ํ•ด ๋ช‡ ๊ฐ€์ง€ ๊ฐœ์„  ์‚ฌํ•ญ์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ์•„ํ‚คํ…์ฒ˜๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ๋””์ฝ”๋”(Vaswani et al., 2017)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ •์˜ํ•˜๋Š” ์ฃผ์š” ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ํ‘œ 1์— ๋‚˜์™€ ์žˆ์œผ๋ฉฐ, ๋ชจ๋ธ์€ 4096 ํ† ํฐ์˜ ์ƒ๋‹นํ•œ ์ปจํ…์ŠคํŠธ ๊ธธ์ด์— ๋Œ€ํ•ด ํ•™์Šต๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ์š”์†Œ ์™ธ์—๋„, ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์›๋ž˜์˜ ํŠธ๋žœ์Šคํฌ๋จธ์— ๋น„ํ•ด ๋ช‡ ๊ฐ€์ง€ ๊ฐœ์„  ์‚ฌํ•ญ์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Our model's architecture is based on the transformer decoder (Vaswani et al., 2017). The key parameters that define our architecture are shown in Table 1, with the models being trained on a substantial context length of 4096 tokens. Beyond the foundational elements, our approach integrates several improvements compared to the original transformer. [|assistant|]
**LLM summarizers๋Š” ์–ด๋–ค ์ข…๋ฅ˜์˜ ์ถฉ์‹ค๋„ ์˜ค๋ฅ˜๋ฅผ ๋ฒ”ํ•˜๋‚˜์š”? (SS3)** ํ‘œ์˜ ์งˆ์  ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด, ๋ถˆ์ถฉ์‹คํ•˜๋‹ค๊ณ  ํ‘œ์‹œ๋œ ๋Œ€๋ถ€๋ถ„์˜ ์ฃผ์žฅ์€ _์ด๋ฒคํŠธ_ ๋˜๋Š” _์ƒํƒœ_์™€ ๊ด€๊ณ„์— ๋Œ€ํ•œ ์ธ๋ฌผ์˜ _์ƒํƒœ_์™€ ๊ด€๋ จ์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ๋Œ€๋ถ€๋ถ„์˜ ์ฃผ์žฅ์€ ์ฆ๊ฑฐ์— ๋Œ€ํ•œ multi-hop reasoning์„ ํ†ตํ•ด์„œ๋งŒ ๋ฌดํšจํ™”ํ•  ์ˆ˜ ์žˆ์–ด, ์ด ์ž‘์—…์˜ ๋ณต์žก์„ฑ๊ณผ ๊ธฐ์กด์˜ fact-verification ์„ค์ •๊ณผ์˜ ์ฐจ์ด์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค(Min et al., 2023; Kamoi et al., 2023).[|endofturn|]
**LLM ์š”์•ฝ๊ธฐ๋Š” ์–ด๋–ค ์ข…๋ฅ˜์˜ ์ถฉ์‹ค๋„ ์˜ค๋ฅ˜๋ฅผ ๋ฒ”ํ•˜๋‚˜์š”? (SS3)** ํ‘œ์˜ ์งˆ์  ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด, ๋ถˆ์ถฉ์‹คํ•˜๋‹ค๊ณ  ํ‘œ์‹œ๋œ ๋Œ€๋ถ€๋ถ„์˜ ์ฃผ์žฅ์€ _์ด๋ฒคํŠธ_ ๋˜๋Š” _์ƒํƒœ_์™€ ๊ด€๊ณ„์— ๋Œ€ํ•œ ์ธ๋ฌผ์˜ _์ƒํƒœ_์™€ ๊ด€๋ จ์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ๋Œ€๋ถ€๋ถ„์˜ ์ฃผ์žฅ์€ ์ฆ๊ฑฐ์— ๋Œ€ํ•œ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ํ†ตํ•ด์„œ๋งŒ ๋ฌดํšจํ™”ํ•  ์ˆ˜ ์žˆ์–ด, ์ด ์ž‘์—…์˜ ๋ณต์žก์„ฑ๊ณผ ๊ธฐ์กด์˜ ์‚ฌ์‹ค ๊ฒ€์ฆ ์„ค์ •๊ณผ์˜ ์ฐจ์ด์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค(Min et al., 2023; Kamoi et al., 2023).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**What kinds of faithfulness errors do LLM summarizers make? (SS3)** A qualitative analysis of Tables reveals that the majority of claims marked as unfaithful are related to _events_ or _states_ of characters and relationships. Furthermore, most of these claims can only be invalidated via multi-hop reasoning over the evidence, highlighting the task's complexity and its difference from existing fact-verification settings (Min et al., 2023; Kamoi et al., 2023). [|assistant|]
_Remark C.3_.: ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ๋Š” ํ•ญ์ƒ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒฐ๋ก ์„ ๋‚ด๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. ์™œ๋ƒํ•˜๋ฉด \(E_{test}\)๋Š” \(\pi\)์˜ ๊ฐ์†Œ ํ•จ์ˆ˜์ด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ, (26)์—์„œ \(\pi\)์— ๋Œ€ํ•œ ๋„ํ•จ์ˆ˜๋Š” \(E^{\prime}_{test}(\pi)=-T\sum_{i\geq k+1}p_{i}^{2}(1-\pi p_{i})^{T-1}\leq 0\)์ž…๋‹ˆ๋‹ค. [|endofturn|]
_Remark C.3_.: (26)์—์„œ \(\pi\)์— ๋Œ€ํ•œ ๋„ํ•จ์ˆ˜๋Š” \(E^{\prime}_{test}(\pi)=-T\sum_{i\geq k+1}p_{i}^{2}(1-\pi p_{i})^{T-1}\leq 0\).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]_Remark C.3_.: Let us conclude by saying that clean data always helps, since \(E_{test}\) is decreasing function of \(\pi\). Indeed, from (26), the derivative w.r.t \(\pi\) is \(E^{\prime}_{test}(\pi)=-T\sum_{i\geq k+1}p_{i}^{2}(1-\pi p_{i})^{T-1}\leq 0\). [|assistant|]
์•„์ด๋ฆฌ์Šค์˜ ์—ฌ์ •์€ ๋™์ƒ๊ณผ ๋กœ๋งŒ์„ ์ฐพ๋Š” ์—ฌ์ •์ด์ž ๊ฐˆ๋“ฑ๊ณผ ์‹ ์˜ ์กฐ์ž‘์— ๋งž์„  ์ธ๊ฐ„์˜ ํšŒ๋ณต๋ ฅ์„ ํƒ๊ตฌํ•˜๋Š” ์—ฌ์ •์ž…๋‹ˆ๋‹ค.[|endofturn|]
20. ์•„์ด๋ฆฌ์Šค์˜ ์—ฌ์ •์€ ๋™์ƒ๊ณผ ๋กœ๋งŒ์„ ์ฐพ๋Š” ์—ฌ์ •์ด์ž ๊ฐˆ๋“ฑ๊ณผ ์‹ ์˜ ์กฐ์ž‘์— ๋งž์„  ์ธ๊ฐ„์˜ ํšŒ๋ณต๋ ฅ์„ ํƒ๊ตฌํ•˜๋Š” ์—ฌ์ •์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]20. Irisโ€™s journey is a search for her brother and Roman, and an exploration of human resilience in the face of conflict and divine manipulation. [|assistant|]
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ. ์šฐ๋ฆฌ๋Š” ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๊ฐ€ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์™€ ๋ฌด๊ด€ํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•˜์ง€๋งŒ, Transformer ๊ธฐ๋ฐ˜์˜ decoder ์ „์šฉ ์–ธ์–ด ๋ชจ๋ธ(Vaswani et al., 2017)์— ๋Œ€ํ•ด ๊ฐœ๋ฐฉํ˜• ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ์œ„ํ•ด ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ ์ตœ์‹  ๋ชจ๋ธ์€ ์ธํ„ฐ๋„ท ํ…์ŠคํŠธ๋กœ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-2 ์ œํ’ˆ๊ตฐ(Radford et al., 2019)์€ Reddit์—์„œ ๋†’์€ ์ˆœ์œ„๋ฅผ ๋ฐ›์€ ์›น ๋ฌธ์„œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ WebText๋กœ ํ•™์Šต๋˜์ง€๋งŒ, ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์‹œ์ž‘์ ์€ ๊ณต๊ฐœ ์›น ํŽ˜์ด์ง€์˜ ์ƒ‰์ธ์ธ CommonCrawl์ž…๋‹ˆ๋‹ค. CommonCrawl์—์„œ ํ•™์Šต๋œ ๋ชจ๋ธ์—๋Š” ์ฑ… ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ถ”๊ฐ€๋œ GPT-3(Brown et al., 2020), ๋‰ด์Šค ๋„๋ฉ”์ธ์œผ๋กœ ํ•„ํ„ฐ๋ง๋œ ์ œํ•œ๋œ ํ•˜์œ„ ์ง‘ํ•ฉ์ธ RealNews์— ๋Œ€ํ•œ GROVER(Zellers et al., 2019), C4๋ผ๋Š” CommonCrawl์˜ ์ •๋ฆฌ๋œ ๋ฒ„์ „์— ๋Œ€ํ•œ T5(Raffel et al., 2020) ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ชจ๋ธ์€ ๋” ์—„์„ ๋œ ์ธํ„ฐ๋„ท ์†Œ์Šค๋กœ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด Guo et al. (2020)์€ 40๊ฐœ ์–ธ์–ด์˜ ๊ณ ํ’ˆ์งˆ ์ฒ˜๋ฆฌ๋œ Wikipedia ํ…์ŠคํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์–ด 141.4M ํŒŒ๋ผ๋ฏธํ„ฐ ์–ธ์–ด ๋ชจ๋ธ์„ ๋‹จ์–ด๋กœ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋น„์˜์–ด ๋ชจ๋ธ์€ ๋ฐ˜๋“œ์‹œ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ด์•ผ ํ•˜๋Š”๋ฐ, ์˜ˆ๋ฅผ ๋“ค์–ด Zeng et al. (2021)์€ CommonCrawl ๋ฐ ๊ธฐํƒ€ ์†Œ์Šค์—์„œ ์ •๋ฆฌ ๋ฐ ํ•„ํ„ฐ๋ง๋œ ์ค‘๊ตญ์–ด ๋ฌธ์„œ์˜ ๋น„๊ณต๊ฐœ ์ฝ”ํผ์Šค๋กœ ํ•™์Šตํ•œ ์ตœ๋Œ€ 200B ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์ธ PANGU-\(\alpha\)๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ค‘ ์ƒ๋‹น์ˆ˜๊ฐ€ ๊ณต๊ฐœ๋˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์— Wiki-40B, C4, RealNews, ๊ทธ๋ฆฌ๊ณ  ์ผ๋ฐ˜์ ์œผ๋กœ ํ‰๊ฐ€์— ์‚ฌ์šฉ๋˜๋Š” ์†Œ๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ One Billion Word Language Model Benchmark(Chelba et al., 2013) ๋“ฑ ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ค‘๋ณต ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ.์šฐ๋ฆฌ๋Š” ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๊ฐ€ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์™€ ๋ฌด๊ด€ํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•˜์ง€๋งŒ, ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜์˜ ๋””์ฝ”๋” ์ „์šฉ ์–ธ์–ด ๋ชจ๋ธ(Vaswani et al., 2017)์— ๋Œ€ํ•ด ๊ฐœ๋ฐฉํ˜• ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ์œ„ํ•ด ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ ์ตœ์‹  ๋ชจ๋ธ์€ ์ธํ„ฐ๋„ท ํ…์ŠคํŠธ๋กœ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-2 ์ œํ’ˆ๊ตฐ(Radford et al., 2019)์€ Reddit์—์„œ ๋†’์€ ์ˆœ์œ„๋ฅผ ๋ฐ›์€ ์›น ๋ฌธ์„œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ WebText๋กœ ํ•™์Šต๋˜์ง€๋งŒ, ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์‹œ์ž‘์ ์€ ๊ณต๊ฐœ ์›น ํŽ˜์ด์ง€์˜ ์ƒ‰์ธ์ธ CommonCrawl์ž…๋‹ˆ๋‹ค. CommonCrawl์—์„œ ํ•™์Šต๋œ ๋ชจ๋ธ์—๋Š” ์ฑ… ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ถ”๊ฐ€๋œ GPT-3(Brown et al., 2020), ๋‰ด์Šค ๋„๋ฉ”์ธ์œผ๋กœ ํ•„ํ„ฐ๋ง๋œ ์ œํ•œ๋œ ํ•˜์œ„ ์ง‘ํ•ฉ์ธ RealNews์— ๋Œ€ํ•œ GROVER(Zellers et al., 2019), C4๋ผ๋Š” ๊ณตํ†ต ํฌ๋กค๋ง์˜ ์ •๋ฆฌ๋œ ๋ฒ„์ „์— ๋Œ€ํ•œ T5(Raffel et al., 2020) ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ชจ๋ธ์€ ๋” ์—„์„ ๋œ ์ธํ„ฐ๋„ท ์†Œ์Šค๋กœ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด Guo et al. (2020)์€ 40๊ฐœ ์–ธ์–ด์˜ ๊ณ ํ’ˆ์งˆ ์ฒ˜๋ฆฌ๋œ ์œ„ํ‚ค๋ฐฑ๊ณผ ํ…์ŠคํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์–ด 141.4M ํŒŒ๋ผ๋ฏธํ„ฐ ์–ธ์–ด ๋ชจ๋ธ์„ ๋‹จ์–ด๋กœ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋น„์˜์–ด ๋ชจ๋ธ์€ ๋ฐ˜๋“œ์‹œ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ด์•ผ ํ•˜๋Š”๋ฐ, ์˜ˆ๋ฅผ ๋“ค์–ด Zeng et al. (2021)์€ ๊ณตํ†ต ํฌ๋กค๋ง ๋ฐ ๊ธฐํƒ€ ์†Œ์Šค์—์„œ ์ •๋ฆฌ ๋ฐ ํ•„ํ„ฐ๋ง๋œ ์ค‘๊ตญ์–ด ๋ฌธ์„œ์˜ ๋น„๊ณต๊ฐœ ์ฝ”ํผ์Šค๋กœ ํ•™์Šตํ•œ ์ตœ๋Œ€ 200B ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์ธ PANGU-\(\์•ŒํŒŒ\)๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ค‘ ์ƒ๋‹น์ˆ˜๊ฐ€ ๊ณต๊ฐœ๋˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์— ์œ„ํ‚ค-40B, C4, ๋ฆฌ์–ผ๋‰ด์Šค, ๊ทธ๋ฆฌ๊ณ  ์ผ๋ฐ˜์ ์œผ๋กœ ํ‰๊ฐ€์— ์‚ฌ์šฉ๋˜๋Š” ์†Œ๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ 10์–ต ๋‹จ์–ด ์–ธ์–ด ๋ชจ๋ธ ๋ฒค์น˜๋งˆํฌ(Chelba et al., 2013) ๋“ฑ ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ค‘๋ณต ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Large language model datasets.While we believe our results are independent of model architecture, we perform our analysis on Transformer-based decoder-only language models (Vaswani et al., 2017) trained for open-ended text generation. These current state-of-the-art models are trained on internet text. For example, the GPT-2 family of models Radford et al. (2019) is trained on WebText, a dataset of web documents highly ranked on Reddit--however this dataset was not made available publicly. A common dataset starting point is CommonCrawl, an index of public webpages. Among the models trained on CommonCrawl include GPT-3 (Brown et al., 2020) with the addition of book datasets, GROVER (Zellers et al., 2019) on a restricted subset filtered to news domains called RealNews, and T5 (Raffel et al., 2020) on a cleaned version of common crawl called C4. Other models are trained on more curated Internet sources--for example Guo et al. (2020) used high quality processed Wikipedia text from 40 different languages to train monolingual 141.4M parameter language models. Non-English models necessarily use different datasets; Zeng et al. (2021) for instance introduced PANGU-\(\alpha\), a family of models with up to 200B parameters that were trained on a non-public corpus of cleaned and filtered Chinese-language documents from CommonCrawl and other sources. Since many of these datasets are not public, we deduplicate three that are: Wiki-40B, C4, and RealNews-as well as the One Billion Word Language Model Benchmark (Chelba et al., 2013), a smaller dataset commonly used for evaluation. [|assistant|]
๋ชจ๋“  ์ž‘์—…์— ๋Œ€ํ•œ ์ง‘๊ณ„๋œ ๊ฒฐ๊ณผ(ํ‘œ 3)์™€ ์„ ํƒ๋œ ๋‹จ์ผ ์ž‘์—…์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ(ํ‘œ 4)๋ฅผ ๋ชจ๋‘ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 3์— ์ œ์‹œ๋œ ๋ชจ๋“  ์ž‘์—…์˜ ํ‰๊ท  ์„ฑ๋Šฅ์˜ ๊ฒฝ์šฐ, ์–ธ์–ด๋ณ„๋กœ ๋‹ค๋ฅธ ์ž‘์—… ์ˆ˜๋ฅผ ๊ณ ๋ คํ•˜์—ฌ weighted average๋ฅผ ๊ณ„์‚ฐํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ๊ฐ ์–ธ์–ด์— ๋Œ€ํ•ด ๋ชจ๋“  ์ž‘์—…์— ๋Œ€ํ•œ ํ‰๊ท ์„ ๊ณ„์‚ฐํ•œ ๋‹ค์Œ ๋ชจ๋“  ์–ธ์–ด ํ‰๊ท ์— ๋Œ€ํ•œ ํ‰๊ท ์„ ๊ณ„์‚ฐํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ชจ๋“  ์ž‘์—…์— ๋Œ€ํ•œ ์ง‘๊ณ„๋œ ๊ฒฐ๊ณผ(ํ‘œ 3)์™€ ์„ ํƒ๋œ ๋‹จ์ผ ์ž‘์—…์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ(ํ‘œ 4)๋ฅผ ๋ชจ๋‘ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 3์— ์ œ์‹œ๋œ ๋ชจ๋“  ์ž‘์—…์˜ ํ‰๊ท  ์„ฑ๋Šฅ์˜ ๊ฒฝ์šฐ, ์–ธ์–ด๋ณ„๋กœ ๋‹ค๋ฅธ ์ž‘์—… ์ˆ˜๋ฅผ ๊ณ ๋ คํ•˜์—ฌ ๊ฐ€์ค‘ ํ‰๊ท ์„ ๊ณ„์‚ฐํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ๊ฐ ์–ธ์–ด์— ๋Œ€ํ•ด ๋ชจ๋“  ์ž‘์—…์— ๋Œ€ํ•œ ํ‰๊ท ์„ ๊ณ„์‚ฐํ•œ ๋‹ค์Œ ๋ชจ๋“  ์–ธ์–ด ํ‰๊ท ์— ๋Œ€ํ•œ ํ‰๊ท ์„ ๊ณ„์‚ฐํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We present both, aggregated results across all tasks (Table 3) and results for selected single tasks (Table 4). For the average performance across all tasks presented in Table 3, we computed weighted average to take into account the different number of tasks per language. In particular, we computed for each language the mean across all tasks, and then computed the mean over all language-means. [|assistant|]
์–ธ์–ด ๋ชจ๋ธ์˜ ๋ฐ˜์ „์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋Š” ํŠนํžˆ ์ œ๊ณต์—…์ฒด๊ฐ€ ๊ฐ€์น˜ ์žˆ๊ฑฐ๋‚˜ ๊ฐœ์ธ ์ •๋ณด๋ฅผ ํฌํ•จํ•˜๋Š” prompts๋ฅผ ์œ ์ง€ํ•˜๋Š” ๊ฒฝ์šฐ ์„œ๋น„์Šค๋กœ์„œ์˜ ๋ชจ๋ธ ๋ฐฐํฌ์™€ ๊ด€๋ จ๋œ ์œค๋ฆฌ์  ์˜๋ฏธ๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์–ธ์–ด ๋ชจ๋ธ์˜ ์‚ฌ์šฉ์ž๋Š” ์ฝ˜ํ…์ธ  ์ƒ์„ฑ ๋ฐ ์ •๋ณด ๊ฒ€์ƒ‰์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ๋ชฉ์ ์œผ๋กœ ์ด๋Ÿฌํ•œ ์„œ๋น„์Šค์— ์˜์กดํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์˜ํ–ฅ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. prompt ๋น„๋ฐ€์€ ์‚ฌ์šฉ์ž๊ฐ€ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ์‹ ๋ขฐ๋ฅผ ์†์ƒ์‹œ์ผœ ์„œ๋น„์Šค ์ž์ฒด์˜ ํˆฌ๋ช…์„ฑ๊ณผ ๋ฌด๊ฒฐ์„ฑ์— ๋Œ€ํ•œ ์šฐ๋ ค๋ฅผ ๋ถˆ๋Ÿฌ์ผ์œผํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์–ธ์–ด ๋ชจ๋ธ์˜ ๋ฐ˜์ „์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋Š” ํŠนํžˆ ์ œ๊ณต์—…์ฒด๊ฐ€ ๊ฐ€์น˜ ์žˆ๊ฑฐ๋‚˜ ๊ฐœ์ธ ์ •๋ณด๋ฅผ ํฌํ•จํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ์œ ์ง€ํ•˜๋Š” ๊ฒฝ์šฐ ์„œ๋น„์Šค๋กœ์„œ์˜ ๋ชจ๋ธ ๋ฐฐํฌ์™€ ๊ด€๋ จ๋œ ์œค๋ฆฌ์  ์˜๋ฏธ๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์–ธ์–ด ๋ชจ๋ธ์˜ ์‚ฌ์šฉ์ž๋Š” ์ฝ˜ํ…์ธ  ์ƒ์„ฑ ๋ฐ ์ •๋ณด ๊ฒ€์ƒ‰์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ๋ชฉ์ ์œผ๋กœ ์ด๋Ÿฌํ•œ ์„œ๋น„์Šค์— ์˜์กดํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์˜ํ–ฅ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ ๋น„๋ฐ€์€ ์‚ฌ์šฉ์ž๊ฐ€ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ์‹ ๋ขฐ๋ฅผ ์†์ƒ์‹œ์ผœ ์„œ๋น„์Šค ์ž์ฒด์˜ ํˆฌ๋ช…์„ฑ๊ณผ ๋ฌด๊ฒฐ์„ฑ์— ๋Œ€ํ•œ ์šฐ๋ ค๋ฅผ ๋ถˆ๋Ÿฌ์ผ์œผํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Our research on the inversion of language models underscores the ethical implications surrounding the deployment of such models as services, particularly when providers maintain prompts that are valuable or contain personal information. Users of language models may be affected as they rely on these services for various purposes, including content generation and information retrieval. Prompt secrecy can compromise users' trust in the systems they interact with, raising concerns about the transparency and integrity of the services themselves. [|assistant|]
๊ทธ๋ฆผ 5: ๋ฌธ์ œ์˜ ์ „์ œ๋ฅผ ๋ณ€๊ฒฝํ•˜์—ฌ "three"์ด๋ผ๋Š” ๋‹จ์–ด๋ฅผ "five"๋กœ ๋ณ€๊ฒฝํ–ˆ์„ ๋•Œ์˜ ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 5: ๋ฌธ์ œ์˜ ์ „์ œ๋ฅผ ๋ณ€๊ฒฝํ•˜์—ฌ "3"์ด๋ผ๋Š” ๋‹จ์–ด๋ฅผ "5"๋กœ ๋ณ€๊ฒฝํ–ˆ์„ ๋•Œ์˜ ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 5: We alter the premise in the question to investigate the impact of changing the word โ€œthreeโ€ to โ€œfiveโ€. [|assistant|]
ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ณ , ๋…ผ๋ฌธ์— ๊ธฐ์—ฌํ•˜๊ณ , v0.5 benchmark๋ฅผ ๋งŒ๋“œ๋Š” ๋ฐ ๋„์›€์„ ์ฃผ์—ˆ๊ฑฐ๋‚˜ WG์— ์ฐธ์—ฌํ•ด์ค€ ๋ชจ๋“  ๋ถ„๋“ค๊ป˜ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ํŠนํžˆ MLCommons ํŒ€์— ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.[|endofturn|]
ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ณ , ๋…ผ๋ฌธ์— ๊ธฐ์—ฌํ•˜๊ณ , v0.5 ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋งŒ๋“œ๋Š” ๋ฐ ๋„์›€์„ ์ฃผ์—ˆ๊ฑฐ๋‚˜ WG์— ์ฐธ์—ฌํ•ด์ค€ ๋ชจ๋“  ๋ถ„๋“ค๊ป˜ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ํŠนํžˆ MLCommons ํŒ€์— ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We thank everyone who has given feedback, contributed to the paper, helped create the v0.5 benchmark, or has joined the WG. We particularly thank the team at MLCommons. [|assistant|]
์š”์•ฝํ•˜์ž๋ฉด, LLM ํ‰๊ฐ€์— ๋Œ€ํ•œ ๊ธฐ์กด ์—ฐ๊ตฌ์˜ ๋ถ„๋ฅ˜(ํ‘œ X)๋Š” ์ฃผ๋กœ ํ‰๊ฐ€ methodology์™€ ๋ชจ๋ธ ์œ ํ˜•์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์ฐจ์›์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ, ์ด๋ฅผ ํ…Œ์ŠคํŠธ ๋Šฅ๋ ฅ์œผ๋กœ ๋”์šฑ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. LLM ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ๊ธฐ์กด ์ž‘์—…์˜ ๋ถ„๋ฅ˜ ๋˜๋Š” ๋ถ„๋ฅ˜๋ฒ•์— ๋Œ€ํ•ด์„œ๋„ ๋…ผ์˜ํ•œ ์ตœ๊ทผ ์—ฐ๊ตฌ [733, 734]๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์š”์•ฝํ•˜์ž๋ฉด, LLM ํ‰๊ฐ€์— ๋Œ€ํ•œ ๊ธฐ์กด ์—ฐ๊ตฌ์˜ ๋ถ„๋ฅ˜(ํ‘œ X)๋Š” ์ฃผ๋กœ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•๋ก ๊ณผ ๋ชจ๋ธ ์œ ํ˜•์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์ฐจ์›์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ, ์ด๋ฅผ ํ…Œ์ŠคํŠธ ๋Šฅ๋ ฅ์œผ๋กœ ๋”์šฑ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. LLM ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ๊ธฐ์กด ์ž‘์—…์˜ ๋ถ„๋ฅ˜ ๋˜๋Š” ๋ถ„๋ฅ˜๋ฒ•์— ๋Œ€ํ•ด์„œ๋„ ๋…ผ์˜ํ•œ ์ตœ๊ทผ ์—ฐ๊ตฌ [733, 734]๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To summarize, our categorization (Table X) of existing work on LLM evaluation is mainly based on two major dimensions, namely evaluation methodology and model type, which are further extended with the test abilities. There are some recent work [733, 734] that also has discussed the categorization or taxonomies of existing work for LLM evaluation. [|assistant|]
ํ‘œ 3์€ ์˜์–ด ํ‰๊ฐ€ ์ž‘์—…์— ๋Œ€ํ•œ ๊ฐ ๋ชจ๋ธ์˜ ํ‰๊ท  ์ •ํ™•๋„์™€ ๋…์ผ์–ด Hella Swag ํ‰๊ฐ€ ์ž‘์—…์— ๋Œ€ํ•œ ์ •๊ทœํ™”๋œ ์ •ํ™•๋„๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€๊ฐ€ ๊ฑฐ์˜ ๋ฌด์ž‘์œ„๋กœ ์ •ํ™•๋„๊ฐ€ ๋‚ฎ๊ธฐ ๋•Œ๋ฌธ์— ์˜์–ด ํ‰๊ฐ€ ์ ์ˆ˜์˜ ํ‰๊ท ์€ ์œ ์˜๋ฏธํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ ๋ณด๊ณ ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค(ํ‘œ 11 ์ฐธ์กฐ). ์˜์–ด ๋ชจ๋ธ์ด ์˜์–ด ํ‰๊ฐ€์—์„œ ๋…์ผ์–ด ๋ชจ๋ธ๋ณด๋‹ค ์ผ๊ด€๋˜๊ฒŒ ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ 25% replay ๋…์ผ์–ด ๋ชจ๋ธ๊ณผ ํ•จ๊ป˜ ์‚ฌ์šฉ๋œ ๊ฐ•๋ ฅํ•œ replay๋Š” ์ด๋Ÿฌํ•œ ๊ฒฉ์ฐจ๋ฅผ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ์˜์–ด ๋ชจ๋ธ์˜ ์˜์–ด ํ‰๊ฐ€ ์„ฑ๋Šฅ์€ ์ตœ๊ณ ๊ฐ’๊ณผ ์ตœ์ €๊ฐ’ ์‚ฌ์ด์˜ 1.19 ๋ฒ”์œ„๋กœ ๋งค์šฐ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ํฌ๊ธฐ์˜ ๊ธฐ๋ณธ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ‰๊ฐ€ ๊ณผ์ •์— ์ƒ๋‹นํ•œ ๋…ธ์ด์ฆˆ๊ฐ€ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋ฉฐ, ์ด๋Ÿฌํ•œ ์ฐจ์ด๋Š” ์œ ์˜๋ฏธํ•˜์ง€ ์•Š์„ ๊ฒƒ์œผ๋กœ ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Learning Rate ์žฌ๊ฐ€์—ด, Learning Rate ์žฌ๊ฐ์‡  ๋ฐ replay๋กœ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์€ \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ์˜์–ด ํ‰๊ฐ€ ์ž‘์—…์—์„œ ๋…์ผ์–ด๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•  ๋•Œ, ๋” ๋งŽ์€ replay๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ์— ๋Œ€ํ•ด ์ผ๊ด€๋œ ๊ฐœ์„  ์‚ฌํ•ญ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์‹œ ํ•œ ๋ฒˆ Learning Rate ์žฌ๊ฐ€์—ด, Learning Rate ์žฌ๊ฐ์‡  ๋ฐ replay๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์ด \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•œ๋‹ค๋Š” ์ ์— ์ฃผ๋ชฉํ•ฉ๋‹ˆ๋‹ค. ๋…์ผ์–ด Hella Swag ๊ฒฐ๊ณผ๋ฅผ ์‚ดํŽด๋ณด๋ฉด ๋…์ผ์–ด ๋ชจ๋ธ์ด ์˜์–ด ๋ชจ๋ธ๋ณด๋‹ค ์ผ๊ด€๋˜๊ฒŒ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋…์ผ์–ด๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ ์ค‘์—์„œ๋„ ์ง€์†์ ์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์ด ํ•ฉ์ง‘ํ•ฉ ํ›ˆ๋ จ ๋ชจ๋ธ๊ณผ ๋…์ผ์–ด๋กœ๋งŒ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค.[|endofturn|]
ํ‘œ 3์€ ์˜์–ด ํ‰๊ฐ€ ์ž‘์—…์— ๋Œ€ํ•œ ๊ฐ ๋ชจ๋ธ์˜ ํ‰๊ท  ์ •ํ™•๋„์™€ ๋…์ผ์–ด ํ—ฌ๋ผ ์Šค์›ฉ ํ‰๊ฐ€ ์ž‘์—…์— ๋Œ€ํ•œ ์ •๊ทœํ™”๋œ ์ •ํ™•๋„๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€๊ฐ€ ๊ฑฐ์˜ ๋ฌด์ž‘์œ„๋กœ ์ •ํ™•๋„๊ฐ€ ๋‚ฎ๊ธฐ ๋•Œ๋ฌธ์— ์˜์–ด ํ‰๊ฐ€ ์ ์ˆ˜์˜ ํ‰๊ท ์€ ์œ ์˜๋ฏธํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ ๋ณด๊ณ ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค(ํ‘œ 11 ์ฐธ์กฐ). ์˜์–ด ๋ชจ๋ธ์ด ์˜์–ด ํ‰๊ฐ€์—์„œ ๋…์ผ์–ด ๋ชจ๋ธ๋ณด๋‹ค ์ผ๊ด€๋˜๊ฒŒ ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ 25% ๋ฆฌํ”Œ๋ ˆ์ด ๋…์ผ์–ด ๋ชจ๋ธ๊ณผ ํ•จ๊ป˜ ์‚ฌ์šฉ๋œ ๊ฐ•๋ ฅํ•œ ๋ฆฌํ”Œ๋ ˆ์ด๋Š” ์ด๋Ÿฌํ•œ ๊ฒฉ์ฐจ๋ฅผ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ์˜์–ด ๋ชจ๋ธ์˜ ์˜์–ด ํ‰๊ฐ€ ์„ฑ๋Šฅ์€ ์ตœ๊ณ ๊ฐ’๊ณผ ์ตœ์ €๊ฐ’ ์‚ฌ์ด์˜ 1.19 ๋ฒ”์œ„๋กœ ๋งค์šฐ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ํฌ๊ธฐ์˜ ๊ธฐ๋ณธ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ‰๊ฐ€ ๊ณผ์ •์— ์ƒ๋‹นํ•œ ๋…ธ์ด์ฆˆ๊ฐ€ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋ฉฐ, ์ด๋Ÿฌํ•œ ์ฐจ์ด๋Š” ์œ ์˜๋ฏธํ•˜์ง€ ์•Š์„ ๊ฒƒ์œผ๋กœ ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ LR ์žฌ๊ฐ€์—ด, LR ์žฌ๊ฐ์‡  ๋ฐ ๋ฆฌํ”Œ๋ ˆ์ด๋กœ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์€ \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ์˜์–ด ํ‰๊ฐ€ ์ž‘์—…์—์„œ ๋…์ผ์–ด๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•  ๋•Œ, ๋” ๋งŽ์€ ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ์— ๋Œ€ํ•ด ์ผ๊ด€๋œ ๊ฐœ์„  ์‚ฌํ•ญ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์‹œ ํ•œ ๋ฒˆ LR ์žฌ๊ฐ€์—ด, LR ์žฌ๊ฐ์‡  ๋ฐ ๋ฆฌํ”Œ๋ ˆ์ด๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์ด \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•œ๋‹ค๋Š” ์ ์— ์ฃผ๋ชฉํ•ฉ๋‹ˆ๋‹ค. ๋…์ผ์–ด ํ—ฌ๋ผ ์Šค์›ฉ ๊ฒฐ๊ณผ๋ฅผ ์‚ดํŽด๋ณด๋ฉด ๋…์ผ์–ด ๋ชจ๋ธ์ด ์˜์–ด ๋ชจ๋ธ๋ณด๋‹ค ์ผ๊ด€๋˜๊ฒŒ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋…์ผ์–ด๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ ์ค‘์—์„œ๋„ ์ง€์†์ ์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์ด ํ•ฉ์ง‘ํ•ฉ ํ›ˆ๋ จ ๋ชจ๋ธ๊ณผ ๋…์ผ์–ด๋กœ๋งŒ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Table 3 reports the average accuracy of each model for our English evaluation tasks and the normalized accuracy for the German Hella Swag evaluation task. We do not report the average German evaluation score as it is not informative due to evaluations having near-random chance accuracy (see Table 11). We observe that English models consistently outperform German models on the English evaluations. However, the strong replay used with the 25% replay German model helps to reduce this gap. English models' English evaluation performance is very similar with a range of 1.19 between the highest and lowest values. We suspect that there is significant noise in the evaluation process for base models of this size and believe that the differences are likely not significant. That being said, the continually pre-trained model with LR re-warming, LR re-decaying, and replay does improve on the \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) model. When evaluating German-trained models on English evaluation tasks, we see consistent improvements for models using more replay. We note that once again the model trained with LR re-warming, LR re-decaying, and replay does improve on the \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) model. Turning to the German Hella Swag results we observe that German models consistently outperform their English counterparts. Among German-trained models, the continually trained models outperform the union-trained model and the model trained exclusively on German. [|assistant|]
\(\bullet\)_Training_. ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ๋งค์šฐ ํฌ๊ธฐ ๋•Œ๋ฌธ์— ์œ ๋Šฅํ•œ LLM์„ ์„ฑ๊ณต์ ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. LLM์˜ ๋„คํŠธ์›Œํฌ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋ถ„์‚ฐ ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ํ•„์š”ํ•˜๋ฉฐ, ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์—๋Š” ๋‹ค์–‘ํ•œ ๋ณ‘๋ ฌ ์ „๋žต์ด ์ข…ํ•ฉ์ ์œผ๋กœ ํ™œ์šฉ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋ถ„์‚ฐ ํ•™์Šต์„ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด DeepSpeed [74] ๋ฐ Megatron-LM [75, 76, 77]๊ณผ ๊ฐ™์ด ๋ณ‘๋ ฌ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๊ตฌํ˜„๊ณผ ๋ฐฐํฌ๋ฅผ ์šฉ์ดํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ optimization ํ”„๋ ˆ์ž„์›Œํฌ๊ฐ€ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ optimization ํŠธ๋ฆญ์€ ํ›ˆ๋ จ ์•ˆ์ •์„ฑ๊ณผ ๋ชจ๋ธ ์„ฑ๋Šฅ์—๋„ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค(์˜ˆ: ํ›ˆ๋ จ Loss ์ŠคํŒŒ์ดํฌ ๊ทน๋ณต์„ ์œ„ํ•œ ์žฌ์‹œ์ž‘ [56] ๋ฐ ํ˜ผํ•ฉ ์ •๋ฐ€๋„ ํ›ˆ๋ จ [78]). ์ตœ๊ทผ์—๋Š” GPT-4 [46]๊ฐ€ ํ›จ์”ฌ ์ž‘์€ ๋ชจ๋ธ๋กœ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์•ˆ์ •์ ์œผ๋กœ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋Š” ํŠน์ˆ˜ ์ธํ”„๋ผ์™€ optimization ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•  ๊ฒƒ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_Training_. ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ๋งค์šฐ ํฌ๊ธฐ ๋•Œ๋ฌธ์— ์œ ๋Šฅํ•œ LLM์„ ์„ฑ๊ณต์ ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. LLM์˜ ๋„คํŠธ์›Œํฌ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋ถ„์‚ฐ ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ํ•„์š”ํ•˜๋ฉฐ, ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์—๋Š” ๋‹ค์–‘ํ•œ ๋ณ‘๋ ฌ ์ „๋žต์ด ์ข…ํ•ฉ์ ์œผ๋กœ ํ™œ์šฉ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋ถ„์‚ฐ ํ•™์Šต์„ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด DeepSpeed [74] ๋ฐ Megatron-LM [75, 76, 77]๊ณผ ๊ฐ™์ด ๋ณ‘๋ ฌ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๊ตฌํ˜„๊ณผ ๋ฐฐํฌ๋ฅผ ์šฉ์ดํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ์ตœ์ ํ™” ํ”„๋ ˆ์ž„์›Œํฌ๊ฐ€ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ตœ์ ํ™” ํŠธ๋ฆญ์€ ํ›ˆ๋ จ ์•ˆ์ •์„ฑ๊ณผ ๋ชจ๋ธ ์„ฑ๋Šฅ์—๋„ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค(์˜ˆ: ํ›ˆ๋ จ ์†์‹ค ์ŠคํŒŒ์ดํฌ ๊ทน๋ณต์„ ์œ„ํ•œ ์žฌ์‹œ์ž‘ [56] ๋ฐ ํ˜ผํ•ฉ ์ •๋ฐ€๋„ ํ›ˆ๋ จ [78]). ์ตœ๊ทผ์—๋Š” GPT-4 [46]๊ฐ€ ํ›จ์”ฌ ์ž‘์€ ๋ชจ๋ธ๋กœ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์•ˆ์ •์ ์œผ๋กœ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋Š” ํŠน์ˆ˜ ์ธํ”„๋ผ์™€ ์ตœ์ ํ™” ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•  ๊ฒƒ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_Training_. Due to the huge model size, it is very challenging to successfully train a capable LLM. Distributed training algorithms are needed to learn the network parameters of LLMs, in which various parallel strategies are often jointly utilized. To support distributed training, several optimization frameworks have been released to facilitate the implementation and deployment of parallel algorithms, such as DeepSpeed [74] and Megatron-LM [75, 76, 77]. Also, optimization tricks are also important for training stability and model performance, _e.g.,_ restart to overcome training loss spike [56] and mixed precision training [78]. More recently, GPT-4 [46] proposes to develop special infrastructure and optimization methods that reliably predict the performance of large models with much smaller models. [|assistant|]
\(\bullet\)**DeepSpeed-Chat**[196]์€ ๋น ๋ฅด๊ณ  ๋น„์šฉ ํšจ์œจ์ ์ด๋ฉฐ ์‚ฌ์šฉํ•˜๊ธฐ ์‰ฌ์šด ์‹œ์Šคํ…œ ํ”„๋ ˆ์ž„์›Œํฌ๋กœ, ๋ชจ๋ธ ํ•™์Šต ์ค‘์— ์ „์ฒด RLHF ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (1) ChatGPT์™€ ์œ ์‚ฌํ•œ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ๋ฐ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ„์†Œํ™”ํ•˜์—ฌ ๊ฐ„๋‹จํ•œ ์Šคํฌ๋ฆฝํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์—ฌ๋Ÿฌ ํ›ˆ๋ จ ๋˜๋Š” ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•˜๊ณ , (2) InstructGPT [66]์˜ ํ›ˆ๋ จ ๋ชจ๋“œ๋ฅผ ๋ณต์ œํ•˜์—ฌ ์„ธ ๊ฐ€์ง€ ํ›ˆ๋ จ ๋‹จ๊ณ„(_์ฆ‰, Supervised Fine-tuning (SFT), ๋ณด์ƒ ๋ชจ๋ธ Fine-tuning ๋ฐ RLHF)์— ๋Œ€ํ•œ ์™„์ „ํ•œ ํŒŒ์ดํ”„๋ผ์ธ์„ ์ œ๊ณตํ•˜๋ฉฐ, (3) DeepSpeed์˜ ํ›ˆ๋ จ ์—”์ง„๊ณผ ์ถ”๋ก  ์—”์ง„์„ ํ†ตํ•ฉํ•œ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์—”์ง„(DeepSpeed HE)์„ RLHF ํ›ˆ๋ จ์— ์ ์šฉํ•˜์—ฌ ํ›ˆ๋ จ ๋ชจ๋“œ์™€ ์ถ”๋ก  ๋ชจ๋“œ ๊ฐ„์˜ ์›ํ™œํ•œ ์ „ํ™˜๊ณผ DeepSpeed Inference์˜ ๋‹ค์–‘ํ•œ ์ตœ์ ํ™” ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์ด ํŠน์ง•์ž…๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)**DeepSpeed-Chat**[196]์€ ๋น ๋ฅด๊ณ  ๋น„์šฉ ํšจ์œจ์ ์ด๋ฉฐ ์‚ฌ์šฉํ•˜๊ธฐ ์‰ฌ์šด ์‹œ์Šคํ…œ ํ”„๋ ˆ์ž„์›Œํฌ๋กœ, ๋ชจ๋ธ ํ•™์Šต ์ค‘์— ์ „์ฒด RLHF ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (1) ChatGPT์™€ ์œ ์‚ฌํ•œ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ๋ฐ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ„์†Œํ™”ํ•˜์—ฌ ๊ฐ„๋‹จํ•œ ์Šคํฌ๋ฆฝํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์—ฌ๋Ÿฌ ํ›ˆ๋ จ ๋˜๋Š” ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•˜๊ณ , (2) InstructGPT [66]์˜ ํ›ˆ๋ จ ๋ชจ๋“œ๋ฅผ ๋ณต์ œํ•˜์—ฌ ์„ธ ๊ฐ€์ง€ ํ›ˆ๋ จ ๋‹จ๊ณ„(_์ฆ‰, SFT, ๋ณด์ƒ ๋ชจ๋ธ ๋ฏธ์„ธ ์กฐ์ • ๋ฐ RLHF)์— ๋Œ€ํ•œ ์™„์ „ํ•œ ํŒŒ์ดํ”„๋ผ์ธ์„ ์ œ๊ณตํ•˜๋ฉฐ, (3) ๋”ฅ์Šคํ”ผ๋“œ์˜ ํ›ˆ๋ จ ์—”์ง„๊ณผ ์ถ”๋ก  ์—”์ง„์„ ํ†ตํ•ฉํ•œ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์—”์ง„(DeepSpeed HE)์„ RLHF ํ›ˆ๋ จ์— ์ ์šฉํ•˜์—ฌ ํ›ˆ๋ จ ๋ชจ๋“œ์™€ ์ถ”๋ก  ๋ชจ๋“œ ๊ฐ„์˜ ์›ํ™œํ•œ ์ „ํ™˜๊ณผ DeepSpeed Inference์˜ ๋‹ค์–‘ํ•œ ์ตœ์ ํ™” ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์ด ํŠน์ง•์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)**DeepSpeed-Chat**[196] is a fast, cost-effective, and easy-to-use system framework that enables the integration of the complete RLHF process during model training. It is featured by three major functionalities: (1) it simplifies the training and inference process for ChatGPT-like models, enabling using a simple script to implement multiple training or inference steps; (2) it replicates the training mode of InstructGPT [66] and provides a complete pipeline for three training steps (_i.e._, SFT, reward model fine-tuning, and RLHF); (3) it integrates the training engine and inference engine of Deepspeed into a unified hybrid engine (Deepspeed HE) for RLHF training, which enables seamless switch between training and inference modes, and leveraging various optimizations from DeepSpeed Inference. [|assistant|]
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ ์„ธ๊ณ„์—์„œ ๊ฐ€์žฅ ๊นŒ๋‹ค๋กœ์šด ์งˆ๋ฌธ์ธ ์–ด๋–ค ๋ชจ๋ธ์ด ๋” ๋‚˜์€๊ฐ€๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ benchmark์™€ leaderboard๊ฐ€ ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๊ฐ„๋‹จํ•œ ๋‹ต๋ณ€์œผ๋กœ๋Š” ์ด ์งˆ๋ฌธ์— ๋‹ตํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋‹ต์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ์„น์…˜ V์—์„œ๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—…์˜ ๋ฒ”์ฃผ๋ณ„ ํ‘œํ˜„๊ณผ ๊ฐ ๋ฒ”์ฃผ์—์„œ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋™์ผํ•œ ๋ถ„๋ฅ˜๋ฅผ ๋”ฐ๋ฅด๊ณ  ๊ฐ ์นดํ…Œ๊ณ ๋ฆฌ์— ๋”ฐ๋ฅธ ๋น„๊ต๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์นดํ…Œ๊ณ ๋ฆฌ์— ๋Œ€ํ•œ ๋น„๊ต๋ฅผ ์ œ๊ณตํ•œ ํ›„์—๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ๋ณด๊ณ ๋œ ์„ฑ๋Šฅ ์ง€ํ‘œ๋ฅผ ํ‰๊ท ํ•˜์—ฌ ์ง‘๊ณ„๋œ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ ์„ธ๊ณ„์—์„œ ๊ฐ€์žฅ ๊นŒ๋‹ค๋กœ์šด ์งˆ๋ฌธ์ธ ์–ด๋–ค ๋ชจ๋ธ์ด ๋” ๋‚˜์€๊ฐ€๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์™€ ๋ฆฌ๋”๋ณด๋“œ๊ฐ€ ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๊ฐ„๋‹จํ•œ ๋‹ต๋ณ€์œผ๋กœ๋Š” ์ด ์งˆ๋ฌธ์— ๋‹ตํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋‹ต์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ์„น์…˜ V์—์„œ๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—…์˜ ๋ฒ”์ฃผ๋ณ„ ํ‘œํ˜„๊ณผ ๊ฐ ๋ฒ”์ฃผ์—์„œ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋™์ผํ•œ ๋ถ„๋ฅ˜๋ฅผ ๋”ฐ๋ฅด๊ณ  ๊ฐ ์นดํ…Œ๊ณ ๋ฆฌ์— ๋”ฐ๋ฅธ ๋น„๊ต๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์นดํ…Œ๊ณ ๋ฆฌ์— ๋Œ€ํ•œ ๋น„๊ต๋ฅผ ์ œ๊ณตํ•œ ํ›„์—๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ๋ณด๊ณ ๋œ ์„ฑ๋Šฅ ์ง€ํ‘œ๋ฅผ ํ‰๊ท ํ•˜์—ฌ ์ง‘๊ณ„๋œ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Various benchmarks and leaderboards have been proposed to address the most challenging question in the world of large language models: Which one is better? However not a simple answer can address this question. The answer depends on various aspects of large language models. Section V shows the categorical presentation of different tasks and the most important datasets in each category. We will follow the same categorization and provide a comparison based on each category. After providing comparison for each category, we will provide a broad overview of aggregated performance by averaging the reported performance metric on different tasks. [|assistant|]
LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ๋Š” LLM์˜ ์ธ๊ฐ„๊ณผ ์œ ์‚ฌํ•œ ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ํšจ๊ณผ์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•  ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค. ์ด ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด (1) LLM์„ ๋” ์ž˜ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์–ด๋–ค architecture๋ฅผ ์„ค๊ณ„ํ•ด์•ผ ํ•˜๋Š”์ง€, (2) ์„ค๊ณ„๋œ architecture๋ฅผ ๊ณ ๋ คํ•  ๋•Œ ์—์ด์ „ํŠธ๊ฐ€ ํŠน์ • ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ๋Šฅ์„ ์Šต๋“ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๋ฐฉ๋ฒ• ๋“ฑ ๋‘ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์ธก๋ฉด์ด ์žˆ์Šต๋‹ˆ๋‹ค. architecture ์„ค๊ณ„์˜ ๋งฅ๋ฝ์—์„œ, ์šฐ๋ฆฌ๋Š” ๊ธฐ์กด ์—ฐ๊ตฌ์˜ ์ฒด๊ณ„์ ์ธ ์ข…ํ•ฉ์„ ํ†ตํ•ด ํฌ๊ด„์ ์ธ ํ†ตํ•ฉ ํ”„๋ ˆ์ž„์›Œํฌ์— ์ด๋ฅด๋ €์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ์ธก๋ฉด์— ๋Œ€ํ•ด์„œ๋Š” LLM์„ Fine-tuningํ• ์ง€ ์—ฌ๋ถ€์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ ๊ธฐ๋Šฅ ํš๋“์„ ์œ„ํ•œ ์ „๋žต์„ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ๋ฅผ ๊ธฐ์กด ๋จธ์‹  ๋Ÿฌ๋‹๊ณผ ๋น„๊ตํ•  ๋•Œ, ์—์ด์ „ํŠธ architecture๋ฅผ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ์€ ๋„คํŠธ์›Œํฌ ๊ตฌ์กฐ๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ๊ณผ ์œ ์‚ฌํ•˜๋ฉฐ, ์—์ด์ „ํŠธ ๊ธฐ๋Šฅ ํš๋“์€ ๋„คํŠธ์›Œํฌ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ•™์Šตํ•˜๋Š” ๊ฒƒ๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ด ๋‘ ๊ฐ€์ง€ ์ธก๋ฉด์— ๋Œ€ํ•ด ์ข€ ๋” ์ž์„ธํžˆ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ๋Š” LLM์˜ ์ธ๊ฐ„๊ณผ ์œ ์‚ฌํ•œ ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ํšจ๊ณผ์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•  ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค. ์ด ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด (1) LLM์„ ๋” ์ž˜ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์–ด๋–ค ์•„ํ‚คํ…์ฒ˜๋ฅผ ์„ค๊ณ„ํ•ด์•ผ ํ•˜๋Š”์ง€, (2) ์„ค๊ณ„๋œ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ณ ๋ คํ•  ๋•Œ ์—์ด์ „ํŠธ๊ฐ€ ํŠน์ • ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ๋Šฅ์„ ์Šต๋“ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๋ฐฉ๋ฒ• ๋“ฑ ๋‘ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์ธก๋ฉด์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์•„ํ‚คํ…์ฒ˜ ์„ค๊ณ„์˜ ๋งฅ๋ฝ์—์„œ, ์šฐ๋ฆฌ๋Š” ๊ธฐ์กด ์—ฐ๊ตฌ์˜ ์ฒด๊ณ„์ ์ธ ์ข…ํ•ฉ์„ ํ†ตํ•ด ํฌ๊ด„์ ์ธ ํ†ตํ•ฉ ํ”„๋ ˆ์ž„์›Œํฌ์— ์ด๋ฅด๋ €์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ์ธก๋ฉด์— ๋Œ€ํ•ด์„œ๋Š” LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ• ์ง€ ์—ฌ๋ถ€์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ ๊ธฐ๋Šฅ ํš๋“์„ ์œ„ํ•œ ์ „๋žต์„ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ๋ฅผ ๊ธฐ์กด ๋จธ์‹  ๋Ÿฌ๋‹๊ณผ ๋น„๊ตํ•  ๋•Œ, ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ์€ ๋„คํŠธ์›Œํฌ ๊ตฌ์กฐ๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ๊ณผ ์œ ์‚ฌํ•˜๋ฉฐ, ์—์ด์ „ํŠธ ๊ธฐ๋Šฅ ํš๋“์€ ๋„คํŠธ์›Œํฌ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ•™์Šตํ•˜๋Š” ๊ฒƒ๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ด ๋‘ ๊ฐ€์ง€ ์ธก๋ฉด์— ๋Œ€ํ•ด ์ข€ ๋” ์ž์„ธํžˆ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]LLM-based autonomous agents are expected to effectively perform diverse tasks by leveraging the human-like capabilities of LLMs. In order to achieve this goal, there are two significant aspects, that is, (1) which architecture should be designed to better use LLMs and (2) give the designed architecture, how to enable the agent to acquire capabilities for accomplishing specific tasks. Within the context of architecture design, we contribute a systematic synthesis of existing research, culminating in a comprehensive unified framework. As for the second aspect, we summarize the strategies for agent capability acquisition based on whether they fine-tune the LLMs. When comparing LLM-based autonomous agents to traditional machine learning, designing the agent architecture is analogous to determining the network structure, while the agent capability acquisition is similar to learning the network parameters. In the following, we introduce these two aspects more in detail. [|assistant|]
์ตœ๊ทผ NLP ๋ฌธํ—Œ์—๋Š” ์ž์—ฐ์–ด instructions์„ ๋”ฐ๋ฅผ ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ ์—„์ฒญ๋‚œ ์–‘์˜ ํ™œ๋™์ด ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Mishra et al., 2022; Wei et al., 2022; Sanh et al., 2022; Wang et al., 2022; Ouyang et al., 2022; Chung et al., 2022, i.a.). ์ด๋Ÿฌํ•œ ๋ฐœ์ „์€ ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ์— ์˜ํ•ด ๊ตฌ๋™๋ฉ๋‹ˆ๋‹ค: ์‚ฌ์ „ ํ•™์Šต๋œ Large Language Models (LMs)์™€ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ instruction ๋ฐ์ดํ„ฐ(์˜ˆ: PromptSource(Bach et al., 2022) ๋ฐ SuperNaturalInstructions(Wang et al., 2022, SuperNI for short))์ž…๋‹ˆ๋‹ค.[|endofturn|]
์ตœ๊ทผ NLP ๋ฌธํ—Œ์—๋Š” ์ž์—ฐ์–ด ๋ช…๋ น์„ ๋”ฐ๋ฅผ ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ ์—„์ฒญ๋‚œ ์–‘์˜ ํ™œ๋™์ด ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Mishra et al., 2022; Wei et al., 2022; Sanh et al., 2022; Wang et al., 2022; Ouyang et al., 2022; Chung et al., 2022, i.a.). ์ด๋Ÿฌํ•œ ๋ฐœ์ „์€ ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ์— ์˜ํ•ด ๊ตฌ๋™๋ฉ๋‹ˆ๋‹ค: ์‚ฌ์ „ ํ•™์Šต๋œ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LM)๊ณผ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ๋ช…๋ น ๋ฐ์ดํ„ฐ(์˜ˆ: PromptSource(Bach et al., 2022) ๋ฐ SuperNaturalInstructions(Wang et al., 2022, SuperNI for short))์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The recent NLP literature has witnessed a tremendous amount of activity in building models that can follow natural language instructions (Mishra et al., 2022; Wei et al., 2022; Sanh et al., 2022; Wang et al., 2022; Ouyang et al., 2022; Chung et al., 2022, i.a.). These developments are powered by two key components: large pretrained language models (LM) and human-written instruction data (e.g., PromptSource (Bach et al., 2022) and SuperNaturalInstructions (Wang et al., 2022, SuperNI for short)). [|assistant|]
**๋ชจ๋ธ.** GPT-3(Brown et al., 2020)๊ณผ ์œ ์‚ฌํ•œ ๊ตฌ์„ฑ ๋ฐ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 1B, 3B ๋ฐ 7B ํŒŒ๋ผ๋ฏธํ„ฐ autoregressive decoder-only ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋ฉฐ, ์ฃผ๋กœ ALiBi(Press et al., 2021)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์—์„œ ์ฐจ์ด๊ฐ€ ๋‚ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž ์ง€์ • ์ฝ”๋“œ๋ฒ ์ด์Šค์—์„œ FlashAttention(Dao et al., 2022)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. pretraining setup์œผ๋กœ ์ธํ•œ ํŽธ์ฐจ๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด The Pile๊ณผ RefinedWeb ๋ชจ๋‘์—์„œ ๋‚ด๋ถ€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, The Pile ๋ชจ๋ธ์ด ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋™์ผํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ ๋ฐ ์ ˆ์ œ ์—ฐ๊ตฌ(4.2 ์ ˆ์˜ ์ „๋ฐ˜๋ถ€, 4.3 ์ ˆ)์˜ ๊ฒฝ์šฐ, 1B ๋ฐ 3B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐ๊ฐ 27B ๋ฐ 60B ํ† ํฐ์— ๋Œ€ํ•ด Hoffmann ๋“ฑ(2022)์˜ scaling laws์— ๋”ฐ๋ผ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณด์—ฌ์ฃผ๋Š” ์ฃผ์š” ์‹คํ—˜(4.2์ ˆ์˜ Falcon-RW ๋ชจ๋ธ)์˜ ๊ฒฝ์šฐ, ์ธ๊ธฐ ์žˆ๋Š” ๊ณต๊ฐœ ๋ชจ๋ธ(Brown et al., 2020; Wang and Komatuszaki, 2021; Scao et al., 2022)์— ๋งž์ถฐ 350GT๊นŒ์ง€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์— ๋„์ž…๋œ LLaMA ๋ชจ๋ธ(Touvron et al., 2023)์€ ๊ฐ€์žฅ ์ž‘์€ ๋ชจ๋ธ์ด ๊ฐ€์žฅ ํฐ ๋ชจ๋ธ๋ณด๋‹ค 2.5๋ฐฐ ๋” ๋งŽ์€ ์ปดํ“จํŒ…์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณ„๋กœ ์˜๋ฏธ ์žˆ๋Š” ๋น„๊ต๋ฅผ ๋ฐฉํ•ดํ•˜๋ฏ€๋กœ ๋น„๊ตํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๋ฐ ๋น„๊ต ๋Œ€์ƒ pretraining ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๊ฐœ์š”๋Š” ๋ถ€๋ก F๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.[|endofturn|]
**๋ชจ๋ธ.** GPT-3(Brown et al., 2020)๊ณผ ์œ ์‚ฌํ•œ ๊ตฌ์„ฑ ๋ฐ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 1B, 3B ๋ฐ 7B ํŒŒ๋ผ๋ฏธํ„ฐ ์ž๋™ ํšŒ๊ท€ ๋””์ฝ”๋” ์ „์šฉ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋ฉฐ, ์ฃผ๋กœ ALiBi(Press et al., 2021)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์—์„œ ์ฐจ์ด๊ฐ€ ๋‚ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž ์ง€์ • ์ฝ”๋“œ๋ฒ ์ด์Šค์—์„œ FlashAttention(Dao et al., 2022)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํŠธ๋ ˆ์ผ๋ง ์Šคํƒ์—…์œผ๋กœ ์ธํ•œ ํŽธ์ฐจ๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด ๋” ํŒŒ์ผ๊ณผ RefinedWeb ๋ชจ๋‘์—์„œ ๋‚ด๋ถ€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, ๋” ํŒŒ์ผ ๋ชจ๋ธ์ด ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋™์ผํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ ๋ฐ ์ ˆ์ œ ์—ฐ๊ตฌ(4.2 ์ ˆ์˜ ์ „๋ฐ˜๋ถ€, 4.3 ์ ˆ)์˜ ๊ฒฝ์šฐ, 1B ๋ฐ 3B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐ๊ฐ 27B ๋ฐ 60B ํ† ํฐ์— ๋Œ€ํ•ด Hoffmann ๋“ฑ(2022)์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋”ฐ๋ผ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณด์—ฌ์ฃผ๋Š” ์ฃผ์š” ์‹คํ—˜(4.2์ ˆ์˜ Falcon-RW ๋ชจ๋ธ)์˜ ๊ฒฝ์šฐ, ์ธ๊ธฐ ์žˆ๋Š” ๊ณต๊ฐœ ๋ชจ๋ธ(Brown et al., 2020; Wang and Komatuszaki, 2021; Scao et al., 2022)์— ๋งž์ถฐ 350GT๊นŒ์ง€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์— ๋„์ž…๋œ LLaMA ๋ชจ๋ธ(Touvron et al., 2023)์€ ๊ฐ€์žฅ ์ž‘์€ ๋ชจ๋ธ์ด ๊ฐ€์žฅ ํฐ ๋ชจ๋ธ๋ณด๋‹ค 2.5๋ฐฐ ๋” ๋งŽ์€ ์ปดํ“จํŒ…์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณ„๋กœ ์˜๋ฏธ ์žˆ๋Š” ๋น„๊ต๋ฅผ ๋ฐฉํ•ดํ•˜๋ฏ€๋กœ ๋น„๊ตํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๋ฐ ๋น„๊ต ๋Œ€์ƒ ์ „์ฒ˜๋ฆฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๊ฐœ์š”๋Š” ๋ถ€๋ก F๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Models.** We train 1B, 3B, and 7B parameters autoregressive decoder-only models, based on configurations and hyperparameters similar to GPT-3 (Brown et al., 2020), diverging mostly on our use of ALiBi (Press et al., 2021). We use FlashAttention (Dao et al., 2022) in a custom codebase. We train internal models on both The Pile and RefinedWeb to control for deviations caused by our pretraining setup-we found The Pile models to perform in-line with others. For small-scale and ablation studies (first half of Section 4.2; Section 4.3), we train models to optimality according to the scaling laws of Hoffmann et al. (2022): on 27B and 60B tokens respectively for our 1B and 3B parameters models. For the main experiments demonstrating our approach (Falcon-RW models in Section 4.2), we train the models to 350GT, in line with popular public models (Brown et al., 2020; Wang and Komatuszaki, 2021; Scao et al., 2022). Note that we do not compare against the recently introduced LLaMA models (Touvron et al., 2023), as the smallest of them is trained on x2.5 more compute than our largest model, preventing a meaningful comparison from being made dataset-wise. For a more in-depth overview of the models and pretraining datasets with which we compare, see Appendix F. [|assistant|]
์šฐ๋ฆฌ์˜ ๋…ธ๋ ฅ์€ ๋ชจ๋ฐ”์ผ ๋ฐฐํฌ๋ฅผ ์œ„ํ•œ optimization์— ์ดˆ์ ์„ ๋งž์ถ˜ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๋ชจ๋ฐ”์ผ ํ”Œ๋žซํผ์—์„œ MiniCPM์˜ inference ๊ธฐ๋Šฅ์˜ ์‹ค์šฉ์„ฑ์„ ์ž…์ฆํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ฐ”์ผ ํ™˜๊ฒฝ์—์„œ MiniCPM๊ณผ ๊ฐ™์€ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ฐœ๋ฐœ์ž ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ์ถ”๊ฐ€ optimization ๋ฐ ์—…๋ฐ์ดํŠธ๋ฅผ ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์šฐ๋ฆฌ์˜ ๋…ธ๋ ฅ์€ ๋ชจ๋ฐ”์ผ ๋ฐฐํฌ๋ฅผ ์œ„ํ•œ ์ตœ์ ํ™”์— ์ดˆ์ ์„ ๋งž์ถ˜ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๋ชจ๋ฐ”์ผ ํ”Œ๋žซํผ์—์„œ MiniCPM์˜ ์ถ”๋ก  ๊ธฐ๋Šฅ์˜ ์‹ค์šฉ์„ฑ์„ ์ž…์ฆํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ฐ”์ผ ํ™˜๊ฒฝ์—์„œ MiniCPM๊ณผ ๊ฐ™์€ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ฐœ๋ฐœ์ž ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ์ถ”๊ฐ€ ์ตœ์ ํ™” ๋ฐ ์—…๋ฐ์ดํŠธ๋ฅผ ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]It is important to emphasize that our efforts did not focus on optimization for mobile deployment, but rather on demonstrating the practicality of MiniCPM's inference capabilities on mobile platforms. We encourage further optimization and updates by the developer community to enhance the performance of large models like MiniCPM in mobile contexts. [|assistant|]
๋„์ถœ๋œ ๋ณ‘๋ ฅ, ์ง„๋‹จ ์ •ํ™•๋„, ๊ด€๋ฆฌ ๊ณ„ํš์˜ ์ ์ ˆ์„ฑ ๋ฐ ๊ทผ๊ฑฐ, ๊ด€๊ณ„ ๊ตฌ์ถ•, ๊ฐœ์ธ์— ๋Œ€ํ•œ ์กด์ค‘, ์˜์‚ฌ์†Œํ†ต ํšจ์œจ์„ฑ๊ณผ ๊ฐ™์€ ํ™˜์ž ์ค‘์‹ฌ์  ๊ณ ๋ ค ์‚ฌํ•ญ [20]. ์˜ํ•™์—์„œ LLMs์˜ ๋Œ€ํ™” ์ž ์žฌ๋ ฅ์„ ์‹คํ˜„ํ•˜๋ ค๋ฉด ์ž„์ƒ์˜์™€ ํ™˜์ž ๊ฐ„์˜ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™”์— ๊ณ ์œ ํ•œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŠน์„ฑ์— ๋Œ€ํ•ด ์˜๋ฃŒ AI ์‹œ์Šคํ…œ์˜ ๊ฐœ๋ฐœ ๋ฐ ํ‰๊ฐ€๋ฅผ ๋” ์ž˜ ์ตœ์ ํ™”ํ•ด์•ผ ํ•  ๋ฏธ์ถฉ์กฑ ์ˆ˜์š”๊ฐ€ ์ƒ๋‹นํžˆ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋„์ถœ๋œ ๋ณ‘๋ ฅ, ์ง„๋‹จ ์ •ํ™•๋„, ๊ด€๋ฆฌ ๊ณ„ํš์˜ ์ ์ ˆ์„ฑ ๋ฐ ๊ทผ๊ฑฐ, ๊ด€๊ณ„ ๊ตฌ์ถ•, ๊ฐœ์ธ์— ๋Œ€ํ•œ ์กด์ค‘, ์˜์‚ฌ์†Œํ†ต ํšจ์œจ์„ฑ๊ณผ ๊ฐ™์€ ํ™˜์ž ์ค‘์‹ฌ์  ๊ณ ๋ ค ์‚ฌํ•ญ [20]. ์˜ํ•™์—์„œ LLM์˜ ๋Œ€ํ™” ์ž ์žฌ๋ ฅ์„ ์‹คํ˜„ํ•˜๋ ค๋ฉด ์ž„์ƒ์˜์™€ ํ™˜์ž ๊ฐ„์˜ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™”์— ๊ณ ์œ ํ•œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŠน์„ฑ์— ๋Œ€ํ•ด ์˜๋ฃŒ AI ์‹œ์Šคํ…œ์˜ ๊ฐœ๋ฐœ ๋ฐ ํ‰๊ฐ€๋ฅผ ๋” ์ž˜ ์ตœ์ ํ™”ํ•ด์•ผ ํ•  ๋ฏธ์ถฉ์กฑ ์ˆ˜์š”๊ฐ€ ์ƒ๋‹นํžˆ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]the elicited history, diagnostic accuracy, the appropriateness of management plans and their rationale, and patient-centred considerations such as relationship-building, respect for the individual and communication efficacy [20]. If the conversational potential of LLMs is to be realized in medicine, there is a significant unmet need to better optimize development and evaluation of medical AI systems for characteristics such as these, which are unique to history-taking and diagnostic dialogue between clinicians and patients. [|assistant|]
์š”์•ฝ์„ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์„ end-to-end ์‹คํ–‰ํ•˜๊ณ  Brier ์ ์ˆ˜์™€ ๊ด€๋ จํ•˜์—ฌ ์ƒ์œ„ 1 prompt(๊ทธ๋ฆผ 13)๋ฅผ ์„ ํƒํ•œ๋‹ค.[|endofturn|]
์š”์•ฝ์„ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์„ ์—”๋“œํˆฌ์—”๋“œ ์‹คํ–‰ํ•˜๊ณ  Brier ์ ์ˆ˜์™€ ๊ด€๋ จํ•˜์—ฌ ์ƒ์œ„ 1 ํ”„๋กฌํ”„ํŠธ(๊ทธ๋ฆผ 13)๋ฅผ ์„ ํƒํ•œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]For summarization, we run our system end-to-end and pick the top 1 prompt (Figure 13) with respect to the Brier score. [|assistant|]
๋ชจ๋“  ๋ชจ๋ธ์€ ๋™์ผํ•œ generation ๊ตฌ์„ฑ์„ ์ฑ„ํƒํ•˜๋ฉฐ, ๊ฒฐ์ •๋ก ์  decoding์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ณธ์–ด๊ฐ€ ์•„๋‹Œ ์‘๋‹ต์„ ๋นˆ ํ…์ŠคํŠธ๋กœ ๋Œ€์ฒดํ•˜์—ฌ ์ผ๋ณธ์–ด ๊ฐ์ง€๊ธฐ๋กœ ROUGE-L์„ ๊ณ„์‚ฐํ•˜์—ฌ ์ผ๋ณธ์–ด๊ฐ€ ์•„๋‹Œ ์‘๋‹ต์— ๋Œ€ํ•ด 0์ ์„ ๋ถ€์—ฌํ•ฉ๋‹ˆ๋‹ค. 4.1์ ˆ์˜ LLM ์‹คํ—˜๊ณผ ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ์ด ์–ธ์–ด ๊ฐ์ง€ ์ž‘์—…์—๋„ fasttext [23, 24]๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ์ค€๋‹ต๋ณ€ ์ž์ฒด์— ์ผ๋ณธ์–ด๊ฐ€ ์•„๋‹Œ ๋‹จ์–ด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์ง€๋งŒ ์ผ๋ณธ์–ด ํ…์ŠคํŠธ์—์„œ ํ”ํžˆ ๋ณผ ์ˆ˜ ์žˆ๋Š” ๋‹จ์–ด(์˜ˆ: "UFO"์™€ ๊ฐ™์ด ๋„๋ฆฌ ์•Œ๋ ค์ง„ ์•ฝ์–ด)๊ฐ€ ํฌํ•จ๋œ ๊ฒฝ์šฐ๋Š” ์˜ˆ์™ธ๋กœ ์‚ผ์•˜์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ ๋ชจ๋ธ์˜ ์ผ๋ณธ์–ด๊ฐ€ ์•„๋‹Œ ์‘๋‹ต์€ ๋นˆ ํ…์ŠคํŠธ๋กœ ๋ณ€ํ™˜๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ชจ๋“  ๋ชจ๋ธ์€ ๋™์ผํ•œ ์„ธ๋Œ€ ๊ตฌ์„ฑ์„ ์ฑ„ํƒํ•˜๋ฉฐ, ๊ฒฐ์ •๋ก ์  ๋””์ฝ”๋”ฉ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ณธ์–ด๊ฐ€ ์•„๋‹Œ ์‘๋‹ต์„ ๋นˆ ํ…์ŠคํŠธ๋กœ ๋Œ€์ฒดํ•˜์—ฌ ์ผ๋ณธ์–ด ๊ฐ์ง€๊ธฐ๋กœ ROUGE-L์„ ๊ณ„์‚ฐํ•˜์—ฌ ์ผ๋ณธ์–ด๊ฐ€ ์•„๋‹Œ ์‘๋‹ต์— ๋Œ€ํ•ด 0์ ์„ ๋ถ€์—ฌํ•ฉ๋‹ˆ๋‹ค. 4.1์ ˆ์˜ LLM ์‹คํ—˜๊ณผ ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ์ด ์–ธ์–ด ๊ฐ์ง€ ์ž‘์—…์—๋„ fasttext [23, 24]๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ์ค€๋‹ต๋ณ€ ์ž์ฒด์— ์ผ๋ณธ์–ด๊ฐ€ ์•„๋‹Œ ๋‹จ์–ด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์ง€๋งŒ ์ผ๋ณธ์–ด ํ…์ŠคํŠธ์—์„œ ํ”ํžˆ ๋ณผ ์ˆ˜ ์žˆ๋Š” ๋‹จ์–ด(์˜ˆ: "UFO"์™€ ๊ฐ™์ด ๋„๋ฆฌ ์•Œ๋ ค์ง„ ์•ฝ์–ด)๊ฐ€ ํฌํ•จ๋œ ๊ฒฝ์šฐ๋Š” ์˜ˆ์™ธ๋กœ ์‚ผ์•˜์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ ๋ชจ๋ธ์˜ ์ผ๋ณธ์–ด๊ฐ€ ์•„๋‹Œ ์‘๋‹ต์€ ๋นˆ ํ…์ŠคํŠธ๋กœ ๋ณ€ํ™˜๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]All models adopt the same generation configurations, with deterministic decoding. We compute ROUGE-L with a Japanese language detector to replace non-Japanese responses with empty texts, resulting in a score of zero for non-Japanese responses. To be consistent with our LLM experiments in Section 4.1, we also employed fasttext [23, 24] for this language detection task. However, we made an exception for cases where the ground-truth answer itself contains non-Japanese but commonly seen words in Japanese texts (e.g., a widely recognized acronym such as "UFO"). In these instances, non-Japanese responses from models are not converted to empty texts. [|assistant|]
์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด SGD ๋ฐ Adam๊ณผ ๊ฐ™์€ Optimizer๋Š” DNN์—์„œ ํฅ๋ฏธ๋กœ์šด ์ง€์‹ ์œ ์ง€ ์†์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Š” CL(Lesort et al., 2023)์˜ ๊ทœ๋ชจ์—์„œ ์œ ์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์†Œ๋Ÿ‰์˜ ๋ฆฌํ”Œ๋ ˆ์ด๋งŒ์œผ๋กœ๋„ ์ง€์‹ ์ถ•์ ์„ ์ด‰์ง„ํ•˜๋Š” ๋ฐ ์ถฉ๋ถ„ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค(Scialom et al., 2022). ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํšจ์œจ์„ฑ์„ ํ™œ์šฉํ•˜์—ฌ ์ ์ ˆํ•œ Learning Rate ์Šค์ผ€์ค„๋ง๊ณผ ๋ฆฌํ”Œ๋ ˆ์ด ์ •์ฑ…์œผ๋กœ ๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธ ์‚ฌ์ „ ํ•™์Šต์„ ๊ฐ•ํ™”ํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด SGD ๋ฐ Adam๊ณผ ๊ฐ™์€ ์ตœ์ ํ™” ๋„๊ตฌ๋Š” DNN์—์„œ ํฅ๋ฏธ๋กœ์šด ์ง€์‹ ์œ ์ง€ ์†์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Š” CL(Lesort et al., 2023)์˜ ๊ทœ๋ชจ์—์„œ ์œ ์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์†Œ๋Ÿ‰์˜ ๋ฆฌํ”Œ๋ ˆ์ด๋งŒ์œผ๋กœ๋„ ์ง€์‹ ์ถ•์ ์„ ์ด‰์ง„ํ•˜๋Š” ๋ฐ ์ถฉ๋ถ„ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค(Scialom et al., 2022). ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํšจ์œจ์„ฑ์„ ํ™œ์šฉํ•˜์—ฌ ์ ์ ˆํ•œ ํ•™์Šต ์†๋„ ์Šค์ผ€์ค„๋ง๊ณผ ๋ฆฌํ”Œ๋ ˆ์ด ์ •์ฑ…์œผ๋กœ ๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธ ์‚ฌ์ „ ํ•™์Šต์„ ๊ฐ•ํ™”ํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Recent work shows that optimizers such as SGD and Adam have interesting knowledge retention properties in DNNs that could be beneficial at scale for CL (Lesort et al., 2023) and that just a small amount of replay could be sufficient to boost knowledge accumulation (Scialom et al., 2022). In this work, we want to benefit from the efficiency of those approaches in the context of large language models pretraining and boost them with the right learning rate scheduling and replay policy. [|assistant|]
\(\bullet\)_ํŠน์ • ๋Šฅ๋ ฅ์„ ์ „๋ฌธํ™”ํ•ฉ๋‹ˆ๋‹ค._ LLM์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์€ ๋ฐ์ดํ„ฐ ์„ ํƒ๊ณผ ํ˜ผํ•ฉ์— ํฌ๊ฒŒ ์˜์กดํ•˜๋ฉฐ, ํŠน์ • ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ๋น„์œจ์„ ๋†’์—ฌ ํŠน์ • ๋ชจ๋ธ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[64, 212]. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ˆ˜ํ•™์  ์ถ”๋ก  ๋ฐ ์ฝ”๋”ฉ ๋Šฅ๋ ฅ์€ ๊ฐ๊ฐ ๋” ๋งŽ์€ ์ˆ˜ํ•™ ํ…์ŠคํŠธ ๋ฐ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•จ์œผ๋กœ์จ ํŠน๋ณ„ํžˆ ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, LAMBADA ๋ฐ์ดํ„ฐ ์„ธํŠธ[233]์— ๋Œ€ํ•œ ์‹คํ—˜ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด, ๋„์„œ ๋ฐ์ดํ„ฐ์˜ ๋น„์ค‘์„ ๋†’์ด๋ฉด ํ…์ŠคํŠธ์—์„œ ์žฅ๊ธฐ ์ข…์†์„ฑ์„ ํฌ์ฐฉํ•˜๋Š” ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํ–ฅ์ƒ๋˜๊ณ , C4 ๋ฐ์ดํ„ฐ ์„ธํŠธ[C4 dataset][82]์˜ ๋น„์ค‘์„ ๋†’์ด๋ฉด C4 ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ[C4 validation dataset][64]์—์„œ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_ํŠน์ • ๋Šฅ๋ ฅ์„ ์ „๋ฌธํ™”ํ•ฉ๋‹ˆ๋‹ค._ LLM์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์€ ๋ฐ์ดํ„ฐ ์„ ํƒ๊ณผ ํ˜ผํ•ฉ์— ํฌ๊ฒŒ ์˜์กดํ•˜๋ฉฐ, ํŠน์ • ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ๋น„์œจ์„ ๋†’์—ฌ ํŠน์ • ๋ชจ๋ธ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[64, 212]. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ˆ˜ํ•™์  ์ถ”๋ก  ๋ฐ ์ฝ”๋”ฉ ๋Šฅ๋ ฅ์€ ๊ฐ๊ฐ ๋” ๋งŽ์€ ์ˆ˜ํ•™ ํ…์ŠคํŠธ ๋ฐ ์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•จ์œผ๋กœ์จ ํŠน๋ณ„ํžˆ ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, LAMBADA ๋ฐ์ดํ„ฐ ์„ธํŠธ[233]์— ๋Œ€ํ•œ ์‹คํ—˜ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด, ๋„์„œ ๋ฐ์ดํ„ฐ์˜ ๋น„์ค‘์„ ๋†’์ด๋ฉด ํ…์ŠคํŠธ์—์„œ ์žฅ๊ธฐ ์ข…์†์„ฑ์„ ํฌ์ฐฉํ•˜๋Š” ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํ–ฅ์ƒ๋˜๊ณ , C4 ๋ฐ์ดํ„ฐ ์„ธํŠธ[82]์˜ ๋น„์ค‘์„ ๋†’์ด๋ฉด C4 ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ[64]์—์„œ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_Specializing the targeted abilities._ The model capacities of LLMs heavily rely on data selection and mixture, and one can boost the proportions of specific data sources to enhance certain model abilities [64, 212]. For example, the mathematical reasoning and coding abilities can be specially enhanced by training with more mathematical texts and code data, respectively. Furthermore, experimental results on the LAMBADA dataset [233] show that increasing the proportion of books data can improve the model capacity in capturing long-term dependencies from text, and increasing the proportion of the C4 dataset [82] leads to performance improvement on the C4 validation dataset [64]. [|assistant|]
1. ๋ถˆ์•ˆ์ • ์•ˆ์ฆˆ๋‚˜ 2. ๊ธ‰์„ฑ ์‹ฌ๊ทผ๊ฒฝ์ƒ‰์ฆ 3. ์•ˆ์ • ์•ˆ์ฆˆ๋‚˜ 4. ์œ„์‹๋„ ์—ญ๋ฅ˜ ์งˆํ™˜ (GERD) 5. ๋น„์ •ํ˜• ํ‰ํ†ต 6. ์—ฐ๊ณจ์—ผ 7. ์†Œํ™”์„ฑ ๊ถค์–‘์„ฑ ์งˆํ™˜ 8. ๋‹ด๋‚ญ ์งˆํ™˜ (biliary colic) 9. ๊ณตํ™ฉ ๋ฐœ์ž‘Arneiety 10.[|endofturn|]
1. ๋ถˆ์•ˆ์ • ์•ˆ์ฆˆ๋‚˜2. ๊ธ‰์„ฑ ์‹ฌ๊ทผ๊ฒฝ์ƒ‰์ฆ3. ์•ˆ์ • ์•ˆ์ฆˆ๋‚˜4. ์œ„์‹๋„ ์—ญ๋ฅ˜ ์งˆํ™˜ (GERD)5. ๋น„์ •ํ˜• ํ‰ํ†ต6. ์—ฐ๊ณจ์—ผ7. ์†Œํ™”์„ฑ ๊ถค์–‘์„ฑ ์งˆํ™˜8. ๋‹ด๋‚ญ ์งˆํ™˜ (๋‹ด๊ด€์„ฑ ์ฝœ๋ฆญ)9. ๊ณตํ™ฉ ๋ฐœ์ž‘Arneiety10.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]1. Unstable Anjna 2. Acute Myocardial Infarction 3. Stable Anjna 4. Gastroesophageal Reflux Disease (GERD) 5. Atypical Chest Pain 6. Costochonditis 7. Peptic Ulicer Disease 8. Gallbladder Disease (biliary colic) 9. Panic AttackArneiety 10. [|assistant|]
์ˆ˜๋™์œผ๋กœ ์ž‘์„ฑ๋œ ํƒˆ์˜ฅ ๋ฌธ์ž์—ด๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํ‰๊ท ๊ฐ’๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚˜๋ฉฐ ์˜ค๋ผํด ํƒˆ์˜ฅ ๋ฐฉ์‹๊ณผ ๋น„์Šทํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ตœ๊ณ ์˜ ํƒˆ์˜ฅ ๋ฐฉ๋ฒ•์ด ์›์‹œ LM์—์„œ ์ž˜ ์ž‘๋™ํ•˜๋Š” ๋ฐ˜๋ฉด, Reinforcement Learning from Human Feedback (RLHF) ์ฑ„ํŒ… ๋ฒ„์ „์—์„œ๋Š” ์–ด๋–ค ํƒˆ์˜ฅ ๋ฐฉ๋ฒ•๋„ ์ž‘๋™ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ๋น„์ฑ„ํŒ… ๋ชจ๋ธ์—์„œ ์•ฝ๊ฐ„ ๋” ์ž˜ ์ž‘๋™ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค(\(59\) ๋Œ€[|endofturn|]
์ˆ˜๋™์œผ๋กœ ์ž‘์„ฑ๋œ ํƒˆ์˜ฅ ๋ฌธ์ž์—ด๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํ‰๊ท ๊ฐ’๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚˜๋ฉฐ ์˜ค๋ผํด ํƒˆ์˜ฅ ๋ฐฉ์‹๊ณผ ๋น„์Šทํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ตœ๊ณ ์˜ ํƒˆ์˜ฅ ๋ฐฉ๋ฒ•์ด ์›์‹œ LM์—์„œ ์ž˜ ์ž‘๋™ํ•˜๋Š” ๋ฐ˜๋ฉด, RLHF ์ฑ„ํŒ… ๋ฒ„์ „์—์„œ๋Š” ์–ด๋–ค ํƒˆ์˜ฅ ๋ฐฉ๋ฒ•๋„ ์ž‘๋™ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ๋น„์ฑ„ํŒ… ๋ชจ๋ธ์—์„œ ์•ฝ๊ฐ„ ๋” ์ž˜ ์ž‘๋™ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค(\(59\) ๋Œ€[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Compared to manually written jailbreak strings, our approach is significantly better than the average value, comparable with the oracle jailbreak method. Notably, while the best jailbreak method works well on the raw LM, none of the jailbreak approaches work on the RLHF chat version. We do observe that our method works slightly better on the non-chat model (\(59\) vs. [|assistant|]
**LLM summarizers๊ฐ€ ์ถฉ์‹ค์„ฑ ์™ธ์— ์ €์ง€๋ฅด๋Š” ๋‹ค๋ฅธ ์˜ค๋ฅ˜๋Š” ๋ฌด์—‡์ธ๊ฐ€์š”? (SS5)** ํ‘œ์— ์š”์•ฝ ์ˆ˜์ค€์˜ ์ž์œ  ํ˜•์‹ ์ฃผ์„์„ ๋ชจ๋‘ ์ฝ”๋”ฉํ•œ ๊ฒฐ๊ณผ, ์ฃผ์„ ์ž‘์„ฑ์ž๊ฐ€ ์ค‘์š” ์ •๋ณด์˜ _๋ˆ„๋ฝ_์„ ์ž์ฃผ ์ง€์ ํ•˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €์„œ ๊ธธ์ด์˜ ์š”์•ฝ์—์„œ ์ตœ์ดˆ์˜ ๋ˆ„๋ฝ ์˜ค๋ฅ˜ ๋ถ„๋ฅ˜๋ฒ•์„ ๊ฐœ๋ฐœํ•˜์—ฌ ๋ชจ๋“  LLM์—์„œ ํ•ต์‹ฌ ์‚ฌ๊ฑด, ์„ธ๋ถ€ ์‚ฌํ•ญ ๋ฐ ์ฃผ์ œ๊ฐ€ ์ž์ฃผ ๋ˆ„๋ฝ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค๋ฅธ ์ฝ˜ํ…์ธ  ์„ ํƒ ์˜ค๋ฅ˜๋„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ์š”์•ฝ๊ธฐ์ธ Claude-3-Opus์™€ GPT-4-Turbo๋Š” ์ฑ…์˜ ์‹œ์ž‘ ๋ถ€๋ถ„์— ์น˜์šฐ์น˜์ง€ ์•Š๊ณ  ์ฑ…์˜ ๋งˆ์ง€๋ง‰ ๋ถ€๋ถ„์— ์žˆ๋Š” ์ฝ˜ํ…์ธ ๋ฅผ ์ง€๋‚˜์น˜๊ฒŒ ๊ฐ•์กฐํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•˜์Šต๋‹ˆ๋‹ค.[|endofturn|]
**LLM ์š”์•ฝ๊ธฐ๊ฐ€ ์ถฉ์‹ค์„ฑ ์™ธ์— ์ €์ง€๋ฅด๋Š” ๋‹ค๋ฅธ ์˜ค๋ฅ˜๋Š” ๋ฌด์—‡์ธ๊ฐ€์š”? (SS5)** ํ‘œ์— ์š”์•ฝ ์ˆ˜์ค€์˜ ์ž์œ  ํ˜•์‹ ์ฃผ์„์„ ๋ชจ๋‘ ์ฝ”๋”ฉํ•œ ๊ฒฐ๊ณผ, ์ฃผ์„ ์ž‘์„ฑ์ž๊ฐ€ ์ค‘์š” ์ •๋ณด์˜ _๋ˆ„๋ฝ_์„ ์ž์ฃผ ์ง€์ ํ•˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €์„œ ๊ธธ์ด์˜ ์š”์•ฝ์—์„œ ์ตœ์ดˆ์˜ ๋ˆ„๋ฝ ์˜ค๋ฅ˜ ๋ถ„๋ฅ˜๋ฒ•์„ ๊ฐœ๋ฐœํ•˜์—ฌ ๋ชจ๋“  LLM์—์„œ ํ•ต์‹ฌ ์‚ฌ๊ฑด, ์„ธ๋ถ€ ์‚ฌํ•ญ ๋ฐ ์ฃผ์ œ๊ฐ€ ์ž์ฃผ ๋ˆ„๋ฝ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค๋ฅธ ์ฝ˜ํ…์ธ  ์„ ํƒ ์˜ค๋ฅ˜๋„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ์š”์•ฝ๊ธฐ์ธ Claude-3-Opus์™€ GPT-4-Turbo๋Š” ์ฑ…์˜ ์‹œ์ž‘ ๋ถ€๋ถ„์— ์น˜์šฐ์น˜์ง€ ์•Š๊ณ  ์ฑ…์˜ ๋งˆ์ง€๋ง‰ ๋ถ€๋ถ„์— ์žˆ๋Š” ์ฝ˜ํ…์ธ ๋ฅผ ์ง€๋‚˜์น˜๊ฒŒ ๊ฐ•์กฐํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•˜์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**What other errors, beyond faithfulness, do LLM summarizers make? (SS5)** By coding all of the summary-level free-form comments in Tables, we find that annotators frequently point out _omissions_ of critical information. We develop the first taxonomy of omission errors in book-length summarization and observe that key events, details, and themes are frequently omitted by all LLMs. We also observe other content selection errors: for example, even our strongest summarizers, Claude-3-Opus and GPT-4-Turbo, over-emphasize content towards the end of books to the detriment of the beginning. [|assistant|]
์Šค์ผ€์ผ๋ง ๋ฒ•์น™์€ ์‚ฌ์ „ ํ•™์Šต Loss์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์ง€๋งŒ, ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์˜ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์Šค์ผ€์ผ๋ง ํšจ๊ณผ๋Š” ์ง€๊ธˆ๊นŒ์ง€ ์—ฐ๊ตฌ๊ฐ€ ๊ฑฐ์˜ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. Emergent abilities[47]๋Š” ํฐ LMs์—๋Š” ์กด์žฌํ•˜์ง€๋งŒ ์ž‘์€ LMs์—๋Š” ์กด์žฌํ•˜์ง€ ์•Š๋Š” ๋Šฅ๋ ฅ์œผ๋กœ ์ •์˜๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์˜ ์กด์žฌ๋Š” ์ตœ๊ทผ ๋‘ ๊ฐ€์ง€ ์ด์œ ๋กœ ์ด์˜๋ฅผ ์ œ๊ธฐํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ถฉ๋ถ„ํ•œ ์ˆ˜์˜ ํ† ํฐ์œผ๋กœ ํ›ˆ๋ จ๋œ ์†Œ๊ทœ๋ชจ LMs์ด emergent abilities๊ฐ€ ์ฃผ์žฅ๋œ ์ž‘์—…์—์„œ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[44; 45; 19]. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค๊ฐ€ ์ ์€ LLaMA-13B[44]๋Š” MMLU[14]์—์„œ GPT-3(175B)๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘˜์งธ, [37]์€ ํŠน์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์„ ํƒ๋œ ๋น„์„ ํ˜• ๋˜๋Š” ๋ถˆ์—ฐ์†์ ์ธ ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ ์ธํ•ด emergent abilities๊ฐ€ ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒƒ์ด์ง€, ๋” ํฐ ๋ชจ๋ธ์˜ ๊ทผ๋ณธ์ ์ธ ๋ณ€ํ™”๊ฐ€ ์•„๋‹Œ ๊ฒƒ์œผ๋กœ ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์Šค์ผ€์ผ๋ง ๋ฒ•์น™์€ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์ง€๋งŒ, ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์˜ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์Šค์ผ€์ผ๋ง ํšจ๊ณผ๋Š” ์ง€๊ธˆ๊นŒ์ง€ ์—ฐ๊ตฌ๊ฐ€ ๊ฑฐ์˜ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ[47]์€ ํฐ LM์—๋Š” ์กด์žฌํ•˜์ง€๋งŒ ์ž‘์€ LM์—๋Š” ์กด์žฌํ•˜์ง€ ์•Š๋Š” ๋Šฅ๋ ฅ์œผ๋กœ ์ •์˜๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์˜ ์กด์žฌ๋Š” ์ตœ๊ทผ ๋‘ ๊ฐ€์ง€ ์ด์œ ๋กœ ์ด์˜๋ฅผ ์ œ๊ธฐํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์ถฉ๋ถ„ํ•œ ์ˆ˜์˜ ํ† ํฐ์œผ๋กœ ํ›ˆ๋ จ๋œ ์†Œ๊ทœ๋ชจ LM์ด ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์ด ์ฃผ์žฅ๋œ ์ž‘์—…์—์„œ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[44; 45; 19]. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค๊ฐ€ ์ ์€ LLaMA-13B[44]๋Š” MMLU[14]์—์„œ GPT-3(175B)๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘˜์งธ, [37]์€ ํŠน์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์„ ํƒ๋œ ๋น„์„ ํ˜• ๋˜๋Š” ๋ถˆ์—ฐ์†์ ์ธ ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ ์ธํ•ด ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์ด ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒƒ์ด์ง€, ๋” ํฐ ๋ชจ๋ธ์˜ ๊ทผ๋ณธ์ ์ธ ๋ณ€ํ™”๊ฐ€ ์•„๋‹Œ ๊ฒƒ์œผ๋กœ ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]While scaling laws focus on the pre-training loss, the scaling effect on the performance of downstream tasks has thus far less studied. Emergent abilities [47] are defined as abilities that present in larger LMs but not present in smaller one. The existence of such abilities is recently challenged for two reasons. First, small LMs trained on a sufficient number of tokens can outperform large models on tasks with claimed emergent abilities [44; 45; 19]. For example, LLaMA-13B with less compute [44] can outperform GPT-3 (175B) on MMLU [14]. Second, [37] claim that emergent abilities appear due to the nonlinear or discontinuous metrics selected to evaluate certain datasets, rather than from a fundamental change in larger models. [|assistant|]
๋ฆฐํŒ… ์‰ฌ์šฐ, ์•„๋””์•ผ ๋ฐ”๋ฃจ์•„, ๋…ธ์•„ ์ฝ˜์Šคํƒ„ํŠธ, ๋ผ๋ฏธ ์•Œ-๋ฅดํ‘ธ, ์ƒค๋ž€ ๋‚˜๋ž‘, ๋ฏธํžˆ๋ฅด ์นผ๋ ˆ, ์•„๋‹ด ๋กœ๋ฒ„์ธ , ์ฝœ๋ฆฐ ๋ผํŽ . 2022. Byt5: ์‚ฌ์ „ ํ•™์Šต๋œ ๋ฐ”์ดํŠธ ๋Œ€ ๋ฐ”์ดํŠธ ๋ชจ๋ธ๋กœ ํ† ํฐ ์—†๋Š” ๋ฏธ๋ž˜๋ฅผ ํ–ฅํ•˜์—ฌ. _์ปดํ“จํ„ฐ ์–ธ์–ดํ•™ ํ˜‘ํšŒ ํŠธ๋žœ์žญ์…˜_, 10:291-306. * ์–‘ ๋“ฑ(2019) ์–‘์ธํŽ˜์ด, ์žฅ์œ„์•ˆ, ํƒ€๋ฅด ํฌ๋ฆฌ์Šค, Baldridge ์ œ์ด์Šจ. 2019. PAWS-X: ์˜์—ญ ์‹๋ณ„์„ ์œ„ํ•œ ๊ต์ฐจ ์–ธ์–ด์  ์ ๋Œ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ. _2019 ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๊ฒฝํ—˜ ๋ฐฉ๋ฒ• ๋ฐ ์ œ9ํšŒ ๊ตญ์ œ ๊ณต๋™ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์ปจํผ๋Ÿฐ์Šค(EMNLP-IJCNLP) ๋…ผ๋ฌธ์ง‘_, ํŽ˜์ด์ง€ 3687-3692, ํ™์ฝฉ, ์ค‘๊ตญ. ์ปดํ“จํ„ฐ ์–ธ์–ดํ•™ ํ˜‘ํšŒ. * ์˜ˆํ—ค์ ค๊ณผ ํ•€ํ„ฐ(2023) ์ƒค์ผ€๋“œ ์˜ˆํ—ค์ ค๊ณผ ์œ ๋ฐœ ํ•€ํ„ฐ. 2023. ์„œ๋ธŒ์›Œ๋“œ ์–ดํœ˜์— ์ปจํ…์ŠคํŠธ ํ†ตํ•ฉ. _EACL_, 623-635ํŽ˜์ด์ง€. ์ปดํ“จํ„ฐ ์–ธ์–ดํ•™ ํ˜‘ํšŒ. * ์œ  ๋“ฑ(2023) ์œ ๋ฆฌ๋ฆฌ, ๋‹ค๋‹ˆ์—˜ ์‹œ๋ฏธ๊ทธ, ์ฝœ๋ฆฐ ํ”Œ๋ž˜ํ—ˆํ‹ฐ, ์•„๋ฅด๋ฉ˜ ์•„๊ฐ€์ž๋ƒ, ๋ฃจํฌ ์ œํ‹€๋ชจ์ด์–ด, ๋งˆ์ดํฌ ๋ฃจ์ด์Šค. 2023. Megabyte: ๋‹ค์ค‘ ์Šค์ผ€์ผ Transformer๋กœ ๋ฐฑ๋งŒ ๋ฐ”์ดํŠธ ์‹œํ€€์Šค ์˜ˆ์ธก. _arXiv preprint arXiv:2305.07185_. * ์…€๋Ÿฌ์Šค ๋“ฑ(2019) ๋กœ์™„ ์…€๋Ÿฌ์Šค, ์•„๋ฆฌ ํ™€์ธ ๋งŒ, ์š”๋‚˜ํƒ„ ๋น„์Šคํฌ, ์•Œ๋ฆฌ ํŒŒ๋ฅดํ•˜๋””, ์ตœ์˜ˆ์ง„. 2019. ํ—ฌ๋ผ์Šค์™: ๊ธฐ๊ณ„๊ฐ€ ์ •๋ง ๋ฌธ์žฅ์„ ๋๋‚ผ ์ˆ˜ ์žˆ์„๊นŒ? _ACL (1)_, ํŽ˜์ด์ง€ 4791-4800. ์ปดํ“จํ„ฐ ์–ธ์–ดํ•™ ํ˜‘ํšŒ. * ์žฅ ๋“ฑ(2022) ์žฅ์‹œ์œ„์—, ๋น„์Šˆ๋ผ๋ธŒ ์ฐจ์šฐ๋‹ค๋ฆฌ, ๊ณ ์–„ ๋‚˜๋งŒ, ์ œ์ž„์Šค ํฌ๋กœ์Šค, ์œ„์ฒต ๊ธฐ์šค, ๋ชจํžˆํŠธ ๋ฐ˜์‚ด, ํ”„๋ž€์‹œ์Šค์ฝ” ๊ตฌ์ฆˆ๋งŒ. 2022. ๋‹ค๊ตญ์–ด ํ† ํฐํ™”๊ธฐ ํ›ˆ๋ จ์—์„œ ์–ธ์–ด ๋ถˆ๊ท ํ˜•์— ๋Œ€ํ•œ ์‹ ๊ฒฝ ๊ธฐ๊ณ„ ๋ฒˆ์—ญ์˜ ๊ฐ•๊ฑด์„ฑ์€ ์–ด๋Š ์ •๋„์ธ๊ฐ€? _๋ฏธ๊ตญ ์˜ค๋ฅผ๋žœ๋„์—์„œ ์—ด๋ฆฐ ์ œ15ํšŒ ์–‘์ฐจ๋ก€ ํ˜‘ํšŒ for ๋จธ์‹  ํŠธ๋žœ์Šฌ๋ ˆ์ด์…˜ ์ธ ๋” ์•„๋ฉ”๋ฆฌ์นด์Šค(Volume 1: ์—ฐ๊ตฌ ํŠธ๋ž™) ์ปจํผ๋Ÿฐ์Šค ๋…ผ๋ฌธ์ง‘_, 97-116ํŽ˜์ด์ง€, ๋ฏธ๊ตญ ์˜ค๋ฅผ๋žœ๋„. ๋จธ์‹  ํŠธ๋žœ์Šฌ๋ ˆ์ด์…˜ ์ธ ๋” ์•„๋ฉ”๋ฆฌ์นด์Šค ํ˜‘ํšŒ.[|endofturn|]
๋ฆฐํŒ… ์‰ฌ์šฐ, ์•„๋””์•ผ ๋ฐ”๋ฃจ์•„, ๋…ธ์•„ ์ฝ˜์Šคํƒ„ํŠธ, ๋ผ๋ฏธ ์•Œ-๋ฅดํ‘ธ, ์ƒค๋ž€ ๋‚˜๋ž‘, ๋ฏธํžˆ๋ฅด ์นผ๋ ˆ, ์•„๋‹ด ๋กœ๋ฒ„์ธ , ์ฝœ๋ฆฐ ๋ผํŽ . 2022. Byt5: ์‚ฌ์ „ ํ•™์Šต๋œ ๋ฐ”์ดํŠธ ๋Œ€ ๋ฐ”์ดํŠธ ๋ชจ๋ธ๋กœ ํ† ํฐ ์—†๋Š” ๋ฏธ๋ž˜๋ฅผ ํ–ฅํ•˜์—ฌ. _์ปดํ“จํ„ฐ ์–ธ์–ดํ•™ ํ˜‘ํšŒ ํŠธ๋žœ์žญ์…˜_, 10:291-306. * ์–‘ ๋“ฑ(2019) ์–‘์ธํŽ˜์ด, ์žฅ์œ„์•ˆ, ํƒ€๋ฅด ํฌ๋ฆฌ์Šค, Baldridge ์ œ์ด์Šจ. 2019. PAWS-X: ์˜์—ญ ์‹๋ณ„์„ ์œ„ํ•œ ๊ต์ฐจ ์–ธ์–ด์  ์ ๋Œ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ. _2019 ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๊ฒฝํ—˜ ๋ฐฉ๋ฒ• ๋ฐ ์ œ9ํšŒ ๊ตญ์ œ ๊ณต๋™ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์ปจํผ๋Ÿฐ์Šค(EMNLP-IJCNLP) ๋…ผ๋ฌธ์ง‘_, ํŽ˜์ด์ง€ 3687-3692, ํ™์ฝฉ, ์ค‘๊ตญ. ์ปดํ“จํ„ฐ ์–ธ์–ดํ•™ ํ˜‘ํšŒ. * ์˜ˆํ—ค์ ค๊ณผ ํ•€ํ„ฐ(2023) ์ƒค์ผ€๋“œ ์˜ˆํ—ค์ ค๊ณผ ์œ ๋ฐœ ํ•€ํ„ฐ. 2023. ์„œ๋ธŒ์›Œ๋“œ ์–ดํœ˜์— ์ปจํ…์ŠคํŠธ ํ†ตํ•ฉ. _EACL_, 623-635ํŽ˜์ด์ง€. ์ปดํ“จํ„ฐ ์–ธ์–ดํ•™ ํ˜‘ํšŒ. * ์œ  ๋“ฑ(2023) ์œ ๋ฆฌ๋ฆฌ, ๋‹ค๋‹ˆ์—˜ ์‹œ๋ฏธ๊ทธ, ์ฝœ๋ฆฐ ํ”Œ๋ž˜ํ—ˆํ‹ฐ, ์•„๋ฅด๋ฉ˜ ์•„๊ฐ€์ž๋ƒ, ๋ฃจํฌ ์ œํ‹€๋ชจ์ด์–ด, ๋งˆ์ดํฌ ๋ฃจ์ด์Šค. 2023. ๋ฉ”๊ฐ€๋ฐ”์ดํŠธ: ๋‹ค์ค‘ ์Šค์ผ€์ผ ํŠธ๋žœ์Šคํฌ๋จธ๋กœ ๋ฐฑ๋งŒ ๋ฐ”์ดํŠธ ์‹œํ€€์Šค ์˜ˆ์ธก. _arXiv preprint arXiv:2305.07185_. * ์…€๋Ÿฌ์Šค ๋“ฑ(2019) ๋กœ์™„ ์…€๋Ÿฌ์Šค, ์•„๋ฆฌ ํ™€์ธ ๋งŒ, ์š”๋‚˜ํƒ„ ๋น„์Šคํฌ, ์•Œ๋ฆฌ ํŒŒ๋ฅดํ•˜๋””, ์ตœ์˜ˆ์ง„. 2019. ํ—ฌ๋ผ์Šค์™: ๊ธฐ๊ณ„๊ฐ€ ์ •๋ง ๋ฌธ์žฅ์„ ๋๋‚ผ ์ˆ˜ ์žˆ์„๊นŒ? ACL (1)_, ํŽ˜์ด์ง€ 4791-4800. ์ปดํ“จํ„ฐ ์–ธ์–ดํ•™ ํ˜‘ํšŒ. * ์žฅ ๋“ฑ(2022) ์žฅ์‹œ์œ„์—, ๋น„์Šˆ๋ผ๋ธŒ ์ฐจ์šฐ๋‹ค๋ฆฌ, ๊ณ ์–„ ๋‚˜๋งŒ, ์ œ์ž„์Šค ํฌ๋กœ์Šค, ์œ„์ฒต ๊ธฐ์šค, ๋ชจํžˆํŠธ ๋ฐ˜์‚ด, ํ”„๋ž€์‹œ์Šค์ฝ” ๊ตฌ์ฆˆ๋งŒ. 2022. ๋‹ค๊ตญ์–ด ํ† ํฐํ™”๊ธฐ ํ›ˆ๋ จ์—์„œ ์–ธ์–ด ๋ถˆ๊ท ํ˜•์— ๋Œ€ํ•œ ์‹ ๊ฒฝ ๊ธฐ๊ณ„ ๋ฒˆ์—ญ์˜ ๊ฐ•๊ฑด์„ฑ์€ ์–ด๋Š ์ •๋„์ธ๊ฐ€? _๋ฏธ๊ตญ ์˜ค๋ฅผ๋žœ๋„์—์„œ ์—ด๋ฆฐ ์ œ15ํšŒ ์–‘์ฐจ๋ก€ ํ˜‘ํšŒ for ๋จธ์‹  ํŠธ๋žœ์Šฌ๋ ˆ์ด์…˜ ์ธ ๋” ์•„๋ฉ”๋ฆฌ์นด์Šค(Volume 1: ์—ฐ๊ตฌ ํŠธ๋ž™) ์ปจํผ๋Ÿฐ์Šค ๋…ผ๋ฌธ์ง‘_, 97-116ํŽ˜์ด์ง€, ๋ฏธ๊ตญ ์˜ค๋ฅผ๋žœ๋„. ๋จธ์‹  ํŠธ๋žœ์Šฌ๋ ˆ์ด์…˜ ์ธ ๋” ์•„๋ฉ”๋ฆฌ์นด์Šค ํ˜‘ํšŒ.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. 2022. Byt5: Towards a token-free future with pre-trained byte-to-byte models. _Transactions of the Association for Computational Linguistics_, 10:291-306. * Yang et al. (2019) Yinfei Yang, Yuan Zhang, Chris Tar, and Jason Baldridge. 2019. PAWS-X: A cross-lingual adversarial dataset for paraphrase identification. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)_, pages 3687-3692, Hong Kong, China. Association for Computational Linguistics. * Yehezkel and Pinter (2023) Shaked Yehezkel and Yuval Pinter. 2023. Incorporating context into subword vocabularies. In _EACL_, pages 623-635. Association for Computational Linguistics. * Yu et al. (2023) Lili Yu, Daniel Simig, Colin Flaherty, Armen Aghajanyan, Luke Zettlemoyer, and Mike Lewis. 2023. Megabyte: Predicting million-byte sequences with multiscale transformers. _arXiv preprint arXiv:2305.07185_. * Zellers et al. (2019) Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? In _ACL (1)_, pages 4791-4800. Association for Computational Linguistics. * Zhang et al. (2022) Shiyue Zhang, Vishrav Chaudhary, Naman Goyal, James Cross, Guillaume Wenzek, Mohit Bansal, and Francisco Guzman. 2022. How robust is neural machine translation to language imbalance in multilingual tokenizer training? In _Proceedings of the 15th biennial conference of the Association for Machine Translation in the Americas (Volume 1: Research Track)_, pages 97-116, Orlando, USA. Association for Machine Translation in the Americas. [|assistant|]
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๋งŽ์€ ์ด์  ์ค‘ ํ•˜๋‚˜๋Š” ๋Œ€๊ทœ๋ชจ๋กœ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์–ด AI ๋ชจ๋ธ์„ ์œ„ํ•œ ํ’๋ถ€ํ•œ ํ•™์Šต ๋ฐ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๊ฑฐ๋‚˜ ํ™•๋ณดํ•˜๊ธฐ ์–ด๋ ค์šด ๋ถ„์•ผ(์˜ˆ: ๋ชจ๋“  ์กฐ๊ฑด์„ ํฌ๊ด„ํ•˜๋Š” ๋‚ ์”จ ๋ฐ์ดํ„ฐ(Lam et al., 2023; Li et al., 2023))์—์„œ ํŠนํžˆ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์ œ์–ด๋œ ๋ณ€ํ˜•์„ ๋„์ž…ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ํด๋ž˜์Šค์˜ ๊ท ํ˜• ์žกํžŒ ํ‘œํ˜„์„ ๋ณด์žฅํ•˜๋Š” ๋“ฑ ํŠน์ • ์š”๊ตฌ ์‚ฌํ•ญ์— ๋งž๊ฒŒ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ๋‹ค๊ตญ์–ด ์–ธ์–ด ํ•™์Šต์—์„œ ์ €์ž์› ์–ธ์–ด์˜ ๊ฐ€์ค‘์น˜ ์ƒํ–ฅ(Przystupa and Abdul-Mageed, 2019)). ์ด๋Ÿฌํ•œ ์ˆ˜์ค€์˜ ๋ฐ์ดํ„ฐ ํŠน์„ฑ ์ œ์–ด๋Š” ๋ชจ๋ธ ์„ฑ๋Šฅ๊ณผ generalization์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์…‹์งธ, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๋ฏผ๊ฐํ•œ ๊ฐœ์ธ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์ง€ ์•Š์€ ์ต๋ช…ํ™” ๋˜๋Š” ๋น„์‹๋ณ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(El Emam et al., 2020; Howe et al., 2017). ์ด๋Š” ํ™˜์ž์˜ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ๊ฐ€ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ์˜๋ฃŒ ๋ถ„์•ผ์—์„œ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค(Dahmen and Cook, 2019; Wei et al., 2019).[|endofturn|]
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๋งŽ์€ ์ด์  ์ค‘ ํ•˜๋‚˜๋Š” ๋Œ€๊ทœ๋ชจ๋กœ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์–ด AI ๋ชจ๋ธ์„ ์œ„ํ•œ ํ’๋ถ€ํ•œ ํ•™์Šต ๋ฐ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๊ฑฐ๋‚˜ ํ™•๋ณดํ•˜๊ธฐ ์–ด๋ ค์šด ๋ถ„์•ผ(์˜ˆ: ๋ชจ๋“  ์กฐ๊ฑด์„ ํฌ๊ด„ํ•˜๋Š” ๋‚ ์”จ ๋ฐ์ดํ„ฐ(Lam et al., 2023; Li et al., 2023))์—์„œ ํŠนํžˆ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์ œ์–ด๋œ ๋ณ€ํ˜•์„ ๋„์ž…ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ํด๋ž˜์Šค์˜ ๊ท ํ˜• ์žกํžŒ ํ‘œํ˜„์„ ๋ณด์žฅํ•˜๋Š” ๋“ฑ ํŠน์ • ์š”๊ตฌ ์‚ฌํ•ญ์— ๋งž๊ฒŒ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ๋‹ค๊ตญ์–ด ์–ธ์–ด ํ•™์Šต์—์„œ ์ €์ž์› ์–ธ์–ด์˜ ๊ฐ€์ค‘์น˜ ์ƒํ–ฅ(Przystupa and Abdul-Mageed, 2019)). ์ด๋Ÿฌํ•œ ์ˆ˜์ค€์˜ ๋ฐ์ดํ„ฐ ํŠน์„ฑ ์ œ์–ด๋Š” ๋ชจ๋ธ ์„ฑ๋Šฅ๊ณผ ์ผ๋ฐ˜ํ™”๋ฅผ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์…‹์งธ, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๋ฏผ๊ฐํ•œ ๊ฐœ์ธ ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์ง€ ์•Š์€ ์ต๋ช…ํ™” ๋˜๋Š” ๋น„์‹๋ณ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(El Emam et al., 2020; Howe et al., 2017). ์ด๋Š” ํ™˜์ž์˜ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ๊ฐ€ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ์˜๋ฃŒ ๋ถ„์•ผ์—์„œ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค(Dahmen and Cook, 2019; Wei et al., 2019).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]One of the many benefits of synthetic data is that it can be generated at scale, providing an abundant supply of training and testing data for AI models. This is particularly valuable in domains where real-world data is scarce or difficult to obtain (e.g., weather data covering all conditions (Lam et al., 2023; Li et al., 2023)). Second, synthetic data can be tailored to specific requirements, such as ensuring a balanced representation of different classes by introducing controlled variations (e.g., up-weighting low-resource languages in multilingual language learning (Przystupa and Abdul-Mageed, 2019)). This level of control over data characteristics can improve model performance and generalization. Third, synthetic data can help mitigate privacy concerns by creating anonymized or de-identified datasets that do not contain sensitive personal information (El Emam et al., 2020; Howe et al., 2017). This is crucial in domains such as healthcare, where patient privacy is of utmost importance (Dahmen and Cook, 2019; Wei et al., 2019). [|assistant|]
๊ทธ๋ฆผ 5: LLaMA์—์„œ ์ˆ˜ํ–‰๋œ ์—ฐ๊ตฌ ์ž‘์—…์˜ ์ง„ํ™” ๊ทธ๋ž˜ํ”„์ž…๋‹ˆ๋‹ค. ์—„์ฒญ๋‚œ ์ˆ˜๋กœ ์ธํ•ด ์ด ๊ทธ๋ฆผ์— ๋ชจ๋“  LLaMA ๋ณ€ํ˜•์„ ํฌํ•จํ•  ์ˆ˜ ์—†์œผ๋ฉฐ, ํ›จ์”ฌ ๋” ๋›ฐ์–ด๋‚œ ์ž‘์—…๋„ ํฌํ•จํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ ์ง„์  ์—…๋ฐ์ดํŠธ๋ฅผ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด ์ด ๊ทธ๋ฆผ์˜ ์†Œ์Šค ํŒŒ์ผ์„ ๊ณต์œ ํ•˜๋ฉฐ, ๋…์ž๋“ค์ด ์›ํ•˜๋Š” ๋ชจ๋ธ์„ ํฌํ•จํ•˜๋„๋ก GitHub ํŽ˜์ด์ง€์— pull request๋ฅผ ์ œ์ถœํ•ด ์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 5: LLaMA์—์„œ ์ˆ˜ํ–‰๋œ ์—ฐ๊ตฌ ์ž‘์—…์˜ ์ง„ํ™” ๊ทธ๋ž˜ํ”„์ž…๋‹ˆ๋‹ค. ์—„์ฒญ๋‚œ ์ˆ˜๋กœ ์ธํ•ด ์ด ๊ทธ๋ฆผ์— ๋ชจ๋“  LLaMA ๋ณ€ํ˜•์„ ํฌํ•จํ•  ์ˆ˜ ์—†์œผ๋ฉฐ, ํ›จ์”ฌ ๋” ๋›ฐ์–ด๋‚œ ์ž‘์—…๋„ ํฌํ•จํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ ์ง„์  ์—…๋ฐ์ดํŠธ๋ฅผ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด ์ด ๊ทธ๋ฆผ์˜ ์†Œ์Šค ํŒŒ์ผ์„ ๊ณต์œ ํ•˜๋ฉฐ, ๋…์ž๋“ค์ด ์›ํ•˜๋Š” ๋ชจ๋ธ์„ ํฌํ•จํ•˜๋„๋ก GitHub ํŽ˜์ด์ง€์— ํ’€ ๋ฆฌํ€˜์ŠคํŠธ๋ฅผ ์ œ์ถœํ•ด ์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Fig. 5: An evolutionary graph of the research work conducted on LLaMA. Due to the huge number, we cannot include all the LLaMA variants in this figure, even much excellent work. To support incremental update, we share the source file of this figure, and welcome the readers to include the desired models by submitting the pull requests on our GitHub page. [|assistant|]
tion ๋ฐ ๋‹ต๋ณ€. NQ๋Š” ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์˜ ๊ธธ์ด ์ธก๋ฉด์—์„œ PopQA์— ๋” ๊ฐ€๊นŒ์›Œ ๋ณด์ด์ง€๋งŒ, SQuAD์—์„œ Adapt-LLM์„ ํ•™์Šต์‹œํ‚ค๋Š” ๋” ๋‚˜์€ ๊ฒฐ๊ณผ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ํ•ต์‹ฌ ์š”์†Œ๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์งˆ๋ฌธ ์ˆ˜(\(\sim\)87K, SQuAD์—์„œ \(\sim\)58K, NQ์—์„œ \(\sim\)58K)์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠน์ • ๋Œ€์ƒ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋” ์ ํ•ฉํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“œ๋Š” ์š”์ธ์„ ๊ทœ๋ช…ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ถ”๊ฐ€ ๋ถ„์„์ด ํ•„์š”ํ•˜์ง€๋งŒ(์ด ์—ฐ๊ตฌ์˜ ๋ฒ”์œ„๋ฅผ ๋ฒ—์–ด๋‚œ ๋‚ด์šฉ), ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” scale์ด ๋‹ค์‹œ ํ•œ ๋ฒˆ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
tion ๋ฐ ๋‹ต๋ณ€. NQ๋Š” ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์˜ ๊ธธ์ด ์ธก๋ฉด์—์„œ PopQA์— ๋” ๊ฐ€๊นŒ์›Œ ๋ณด์ด์ง€๋งŒ, SQuAD์—์„œ Adapt-LLM์„ ํ•™์Šต์‹œํ‚ค๋Š” ๋” ๋‚˜์€ ๊ฒฐ๊ณผ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ํ•ต์‹ฌ ์š”์†Œ๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์งˆ๋ฌธ ์ˆ˜(\(sim\)87K, SQuAD์—์„œ \(\sim\)58K, NQ์—์„œ \(\sim\)58K)์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠน์ • ๋Œ€์ƒ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋” ์ ํ•ฉํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“œ๋Š” ์š”์ธ์„ ๊ทœ๋ช…ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ถ”๊ฐ€ ๋ถ„์„์ด ํ•„์š”ํ•˜์ง€๋งŒ(์ด ์—ฐ๊ตฌ์˜ ๋ฒ”์œ„๋ฅผ ๋ฒ—์–ด๋‚œ ๋‚ด์šฉ), ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๊ทœ๋ชจ๊ฐ€ ๋‹ค์‹œ ํ•œ ๋ฒˆ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•  ์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|]tion and answer. While NQ appears to be closer to PopQA in terms of question and answer lengths, the key factor influencing the better results of training Adapt-LLM on SQuAD may be the number of questions in the training dataset (\(\sim\)87K in SQuAD and \(\sim\)58K in NQ). Further analyses are required to elucidate the factors that render a training dataset more suitable for a given target dataset (which is beyond the scope of our study), but these results suggest that scale may play once again a crucial role. [|assistant|]
์•„ํ‚คํ…์ฒ˜ ์šฐ๋ฆฌ๋Š” ์„ธ ๊ฐ€์ง€ ๊ทœ๋ชจ(์†Œํ˜•, ์ค‘ํ˜•, XL)์—์„œ decoder-only Transformer ๋ชจ๋ธ(Vaswani et al., 2017)์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ(ํŒŒ๋ผ๋ฏธํ„ฐ 1์–ต 2,800๋งŒ ๊ฐœ) ๋ชจ๋ธ์€ 12๊ฐœ์˜ ๋ ˆ์ด์–ด, 12๊ฐœ์˜ attention heads, 768์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์› ํฌ๊ธฐ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ๊ทœ๋ชจ(ํŒŒ๋ผ๋ฏธํ„ฐ 3์–ต 5์ฒœ๋งŒ ๊ฐœ) ๋ชจ๋ธ์€ 24๊ฐœ์˜ ๋ ˆ์ด์–ด, 16๊ฐœ์˜ attention heads, 1024์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์› ํฌ๊ธฐ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. XL ๊ทœ๋ชจ(ํŒŒ๋ผ๋ฏธํ„ฐ 13์–ต ๊ฐœ) ๋ชจ๋ธ์€ 24๊ฐœ์˜ ๋ ˆ์ด์–ด, 16๊ฐœ์˜ attention heads, 2048์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์› ํฌ๊ธฐ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋‘ ๋ชจ๋ธ ๋ชจ๋‘ dropout์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š์œผ๋ฉฐ ์ตœ๋Œ€ ์‹œํ€€์Šค ๊ธธ์ด๋Š” 1024์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ์€ NVIDIA์˜ Megatron-LM ๋ฆฌํฌ์ง€ํ† ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์•„ํ‚คํ…์ฒ˜์šฐ๋ฆฌ๋Š” ์„ธ ๊ฐ€์ง€ ๊ทœ๋ชจ(์†Œํ˜•, ์ค‘ํ˜•, XL)์—์„œ ๋””์ฝ”๋” ์ „์šฉ ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ(Vaswani et al., 2017)์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ(ํŒŒ๋ผ๋ฏธํ„ฐ 1์–ต 2,800๋งŒ ๊ฐœ) ๋ชจ๋ธ์€ 12๊ฐœ์˜ ๋ ˆ์ด์–ด, 12๊ฐœ์˜ ์ฃผ์˜ ํ—ค๋“œ, 768์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์› ํฌ๊ธฐ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ๊ทœ๋ชจ(ํŒŒ๋ผ๋ฏธํ„ฐ 3์–ต 5์ฒœ๋งŒ ๊ฐœ) ๋ชจ๋ธ์€ 24๊ฐœ์˜ ๋ ˆ์ด์–ด, 16๊ฐœ์˜ ์ฃผ์˜ ํ—ค๋“œ, 1024์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์› ํฌ๊ธฐ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. XL ๊ทœ๋ชจ(ํŒŒ๋ผ๋ฏธํ„ฐ 13์–ต ๊ฐœ) ๋ชจ๋ธ์€ 24๊ฐœ์˜ ๋ ˆ์ด์–ด, 16๊ฐœ์˜ ์ฃผ์˜ ํ—ค๋“œ, 2048์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์› ํฌ๊ธฐ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋‘ ๋ชจ๋ธ ๋ชจ๋‘ ๋“œ๋กญ์•„์›ƒ์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š์œผ๋ฉฐ ์ตœ๋Œ€ ์‹œํ€€์Šค ๊ธธ์ด๋Š” 1024์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ์€ NVIDIA์˜ Megatron-LM ๋ฆฌํฌ์ง€ํ† ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]ArchitectureWe train decoder-only transformer models (Vaswani et al., 2017) at three different scales, small, medium and XL. The small-scale (128M parameter) model consists of 12 layers, 12 attention heads, and a hidden dimension size of 768. The medium-scale (350M parameter) model consists of 24 layers, 16 attention heads, and a hidden dimension sizeof 1024. The XL-scale (1.3B parameter) model consists of 24 layers, 16 attention heads, and a hidden dimension size of 2048. We do not use dropout in either model and a maximum sequence length of 1024. The models are trained using NVIDIA's Megatron-LM repository. [|assistant|]
์ด์™€ ๊ฐ™์€ ์ •๋ ฌ๋œ ๋ชจ๋ธ์ด ์—†๊ฑฐ๋‚˜ ์ฐธ์กฐ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ œ์–ด๊ถŒ์„ ํฌ๊ธฐํ•˜์—ฌ ์•ˆ์ „ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์œ„์˜ ๋ฐฉ๋ฒ•์€ ์‹คํ˜„ ๊ฐ€๋Šฅํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€์‹ , DPO ํ•™์Šต์„ ์ง„ํ–‰ํ•  ๋•Œ _๋‹จ๊ณ„๋ณ„ ๋ฐฉ์‹_์œผ๋กœ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ(๋˜๋Š” ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” '๋‹จ๊ณ„๋ณ„ DPO'๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ๋‹จ๊ณ„์˜ ์ •๋ ฌ๋œ ๋ชจ๋ธ์ด ํ˜„์žฌ ๋‹จ๊ณ„์˜ reference model๋กœ ์‚ฌ์šฉ๋˜๋ฉฐ, ์ด๋Š” ๋” ์ •๋ ฌ๋œ reference model(_์ฆ‰, ๋” ๋‚˜์€ ํ•˜ํ•œ์„ _)์„ ํ™œ์šฉํ•˜๋Š” ๊ฒฐ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค.[|endofturn|]
์ด์™€ ๊ฐ™์€ ์ •๋ ฌ๋œ ๋ชจ๋ธ์ด ์—†๊ฑฐ๋‚˜ ์ฐธ์กฐ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ œ์–ด๊ถŒ์„ ํฌ๊ธฐํ•˜์—ฌ ์•ˆ์ „ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์œ„์˜ ๋ฐฉ๋ฒ•์€ ์‹คํ˜„ ๊ฐ€๋Šฅํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€์‹ , DPO ํ•™์Šต์„ ์ง„ํ–‰ํ•  ๋•Œ _๋‹จ๊ณ„๋ณ„ ๋ฐฉ์‹_์œผ๋กœ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ(๋˜๋Š” ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” '๋‹จ๊ณ„๋ณ„ DPO'๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ๋‹จ๊ณ„์˜ ์ •๋ ฌ๋œ ๋ชจ๋ธ์ด ํ˜„์žฌ ๋‹จ๊ณ„์˜ ๊ธฐ์ค€ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉ๋˜๋ฉฐ, ์ด๋Š” ๋” ์ •๋ ฌ๋œ ๊ธฐ์ค€ ๋ชจ๋ธ(_์ฆ‰, ๋” ๋‚˜์€ ํ•˜ํ•œ์„ )์„ ํ™œ์šฉํ•˜๋Š” ๊ฒฐ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Note that the above may not be feasible due to the absence of such aligned models, or the fact that it renounces control over the reference model, leading to safety concerns. Instead, we propose'step-wise DPO', named sDPO, where we use the preference datasets (or subsets of a preference dataset) in _a step-by-step manner_ when undergoing DPO training. The aligned model in the previous step is used as the reference model for the current step, which results in utilizing a more aligned reference model (_i.e._, a better lower bound). [|assistant|]
์ด์ „ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ํ…์ŠคํŠธ ์ •๋ณด๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ Transformer ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์˜ ์—ฌ๋Ÿฌ ๊ณ„์ธต์— ์„œ๋กœ ๋‹ค๋ฅธ ๊ธฐ๋Šฅ์ด ์กด์žฌํ•œ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค(Ethayarajh, 2019; Van Aken et al., 2019). ๋”ฐ๋ผ์„œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋ชจ๋ธ์˜ ํŠน์ • ๋ ˆ์ด์–ด๋ฅผ ๋™๊ฒฐํ•˜๋Š” ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์ฒ˜์Œ 10๊ฐœ ๋ ˆ์ด์–ด๋ฅผ ๋™๊ฒฐํ•˜๊ณ  ๋งˆ์ง€๋ง‰ 10๊ฐœ ๋ ˆ์ด์–ด๋ฅผ ๋™๊ฒฐํ•˜๋Š” ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜๋ฉฐ, ๊ฐ๊ฐ Freeze First 10๊ณผ Freeze Last 10์œผ๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ด์ „ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ํ…์ŠคํŠธ ์ •๋ณด๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์˜ ์—ฌ๋Ÿฌ ๊ณ„์ธต์— ์„œ๋กœ ๋‹ค๋ฅธ ๊ธฐ๋Šฅ์ด ์กด์žฌํ•œ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค(Ethayarajh, 2019; Van Aken et al., 2019). ๋”ฐ๋ผ์„œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋ชจ๋ธ์˜ ํŠน์ • ๋ ˆ์ด์–ด๋ฅผ ๋™๊ฒฐํ•˜๋Š” ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์ฒ˜์Œ 10๊ฐœ ๋ ˆ์ด์–ด๋ฅผ ๋™๊ฒฐํ•˜๊ณ  ๋งˆ์ง€๋ง‰ 10๊ฐœ ๋ ˆ์ด์–ด๋ฅผ ๋™๊ฒฐํ•˜๋Š” ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜๋ฉฐ, ๊ฐ๊ฐ Freeze First 10๊ณผ Freeze Last 10์œผ๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Previous studies have shown that distinct functionality exists in different layers of Transformer-based models when processing textual information Ethayarajh (2019); Van Aken et al. (2019). Consequently, we experiment with freezing specific layers of the model during continual pre-training. Specifically, we explore freezing the first ten layers and freezing the last ten layers, denoted as Freeze First 10 and Freeze Last 10, respectively. [|assistant|]
**Challenge**; Clark et al.2018). ์šฐ๋ฆฌ๋Š” accuracy๋ฅผ ํ‰๊ฐ€ ์ง€ํ‘œ๋กœ ์‚ฌ์šฉํ•˜๊ณ  ํ…Œ์ŠคํŠธ ์„ธํŠธ์— ๋Œ€ํ•ด ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ๋Œ€์ƒ ํด๋ž˜์Šค์˜ ๋‹ต๋ฅ ์„ ์ง‘๊ณ„ํ•ฉ๋‹ˆ๋‹ค(๋ถ€๋ก ์„น์…˜ B.2).[|endofturn|]
**Challenge**; Clark et al.2018). ์ •ํ™•๋„๋ฅผ ํ‰๊ฐ€ ์ง€ํ‘œ๋กœ ์‚ฌ์šฉํ•˜๊ณ  ํ…Œ์ŠคํŠธ ์„ธํŠธ์— ๋Œ€ํ•ด ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ๋Œ€์ƒ ํด๋ž˜์Šค์˜ ๋‹ต๋ฅ ์„ ์ง‘๊ณ„ํ•ฉ๋‹ˆ๋‹ค(๋ถ€๋ก ์„น์…˜ B.2).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Challenge**; Clark et al.2018). We use accuracy as an evaluation metric and report on the test set. We aggregate the answer probabilities of target classes for both of these datasets (Appendix Section B.2). [|assistant|]
RealNews๋Š” ๋‰ด์Šค ๋„๋ฉ”์ธ์˜ ๊ธฐ์‚ฌ๋กœ ๊ตฌ์„ฑ๋œ Common Crawl์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์ž…๋‹ˆ๋‹ค Zellers et al. (2019). ํ‰๊ท  ๊ธธ์ด 793 BPE ํ† ํฐ์˜ 31M ๋ฌธ์„œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. RealNews๋Š” ๊ฐ ๋ฌธ์„œ์˜ ์ฒ˜์Œ 100์ž์˜ ํ•ด์‹œ๋ฅผ bloom filter Bloom (1970)์— ์‚ฝ์ž…ํ•œ ๋‹ค์Œ ํ•ด์‹œ ์ถฉ๋Œ์ด ๋ฐœ์ƒํ•œ ๋ฌธ์„œ๋ฅผ ์ œ์™ธํ•˜์—ฌ ์ค‘๋ณต ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค. C4์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ URL์ด ์ค‘๋ณต๋œ ์˜ˆ๋Š” ์ œ์™ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
RealNews๋Š” ๋‰ด์Šค ๋„๋ฉ”์ธ์˜ ๊ธฐ์‚ฌ๋กœ ๊ตฌ์„ฑ๋œ Common Crawl์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์ž…๋‹ˆ๋‹ค Zellers et al. (2019). ํ‰๊ท  ๊ธธ์ด 793 BPE ํ† ํฐ์˜ 31M ๋ฌธ์„œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. RealNews๋Š” ๊ฐ ๋ฌธ์„œ์˜ ์ฒ˜์Œ 100์ž์˜ ํ•ด์‹œ๋ฅผ ๋ธ”๋ฃธ ํ•„ํ„ฐ Bloom (1970)์— ์‚ฝ์ž…ํ•œ ๋‹ค์Œ ํ•ด์‹œ ์ถฉ๋Œ์ด ๋ฐœ์ƒํ•œ ๋ฌธ์„œ๋ฅผ ์ œ์™ธํ•˜์—ฌ ์ค‘๋ณต ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค. C4์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ URL์ด ์ค‘๋ณต๋œ ์˜ˆ๋Š” ์ œ์™ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]RealNewsis a subset of the Common Crawl consisting of articles from news domains Zellers et al. (2019). It contains 31M documents with average length 793 BPE tokens. RealNews was deduplicated by inserting a hash of the first 100 characters of each document into a bloom filter Bloom (1970) and then excluding any document which resulted in a hash collision. Like C4, examples with duplicate URLs were excluded. [|assistant|]
์šฐ๋ฆฌ๋Š” LLM์ด ํŒŒ๋ผ๋ฏธํ„ฐ ๊ณต๊ฐ„์œผ๋กœ ์ง€์‹์„ ํšจ์œจ์ ์œผ๋กœ ์••์ถ•ํ•˜์—ฌ 8๋น„ํŠธ ํŒŒ๋ผ๋ฏธํ„ฐ์—์„œ๋„ 2๋น„ํŠธ/ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์•˜์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ทธ๋Ÿฌํ•œ ์ง€์‹์ด ์–ด๋–ป๊ฒŒ, ์–ด๋””์— ์ €์žฅ๋˜๋Š”์ง€์— ๋Œ€ํ•œ ์˜๋ฌธ์„ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์˜ˆ๋น„์ ์ธ ๋‹ต๋ณ€์€ ์ง€์‹์ด ๋ชจ๋ธ ๋‚ด์— ๊ทธ๋‹ค์ง€ ์ค‘๋ณต๋˜์ง€ ์•Š๋Š” ๋ฐฉ์‹์œผ๋กœ ์••์ถ•๋˜์–ด ์ €์žฅ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋น„์Šทํ•œ ํฌ๊ธฐ์˜ Attention ๋ ˆ์ด์–ด๋„ ์ง€์‹ ์ €์žฅ์— ๊ธฐ์—ฌํ•˜๊ธฐ ๋•Œ๋ฌธ์— MLP ๋ ˆ์ด์–ด๋งŒ์œผ๋กœ๋Š” ์ง€์‹์„ ์ €์žฅํ•  ์ˆ˜ ์—†์„ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค(๊ฒฐ๊ณผ 5 ์ฐธ์กฐ). ๋˜ํ•œ, ํŠนํžˆ ์šฉ๋Ÿ‰ ํ•œ๊ณ„์— ๊ฐ€๊นŒ์šด ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, ๋งˆ์ง€๋ง‰ Transformer ๋ ˆ์ด์–ด๋ฅผ ์ œ๊ฑฐํ•˜์—ฌ ๋‚จ์€ ์ง€์‹์„ "ํƒ์ƒ‰"ํ•˜๋ฉด "๋‚จ์€ ์ง€์‹"์ด ์ „์ฒด์˜ \(1-\frac{1}{L}\)๋ณด๋‹ค ํ›จ์”ฌ ์ ์„ ์ˆ˜ ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.19 ์ด๋Š” ์ง€์‹์ด ๊ฐœ๋ณ„ ๋ ˆ์ด์–ด๊ฐ€ ์•„๋‹Œ ์กฐํ•ฉ ์ž๋ฌผ์‡ ๊ฐ€ ์žˆ๋Š” ๊ธˆ๊ณ ์™€ ์œ ์‚ฌํ•œ ๋ณต์žกํ•œ ๋ฐฉ์‹์œผ๋กœ ์ €์žฅ๋˜๋ฉฐ, ํ•œ ๋ ˆ์ด์–ด๋ฅผ ์ œ๊ฑฐํ•˜๋ฉด ์ „์ฒด ์ง€์‹์˜ \(\frac{1}{L}\)๋ณด๋‹ค ํ›จ์”ฌ ๋” ๋งŽ์€ ์–‘์ด ์ œ๊ฑฐ๋  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์šฐ๋ฆฌ๋Š” LLM์ด ํŒŒ๋ผ๋ฏธํ„ฐ ๊ณต๊ฐ„์œผ๋กœ ์ง€์‹์„ ํšจ์œจ์ ์œผ๋กœ ์••์ถ•ํ•˜์—ฌ 8๋น„ํŠธ ํŒŒ๋ผ๋ฏธํ„ฐ์—์„œ๋„ 2๋น„ํŠธ/ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์•˜์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ทธ๋Ÿฌํ•œ ์ง€์‹์ด ์–ด๋–ป๊ฒŒ, ์–ด๋””์— ์ €์žฅ๋˜๋Š”์ง€์— ๋Œ€ํ•œ ์˜๋ฌธ์„ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์˜ˆ๋น„์ ์ธ ๋‹ต๋ณ€์€ ์ง€์‹์ด ๋ชจ๋ธ ๋‚ด์— ๊ทธ๋‹ค์ง€ ์ค‘๋ณต๋˜์ง€ ์•Š๋Š” ๋ฐฉ์‹์œผ๋กœ ์••์ถ•๋˜์–ด ์ €์žฅ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋น„์Šทํ•œ ํฌ๊ธฐ์˜ ์–ดํ…์…˜ ๋ ˆ์ด์–ด๋„ ์ง€์‹ ์ €์žฅ์— ๊ธฐ์—ฌํ•˜๊ธฐ ๋•Œ๋ฌธ์— MLP ๋ ˆ์ด์–ด๋งŒ์œผ๋กœ๋Š” ์ง€์‹์„ ์ €์žฅํ•  ์ˆ˜ ์—†์„ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค(๊ฒฐ๊ณผ 5 ์ฐธ์กฐ). ๋˜ํ•œ, ํŠนํžˆ ์šฉ๋Ÿ‰ ํ•œ๊ณ„์— ๊ฐ€๊นŒ์šด ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, ๋งˆ์ง€๋ง‰ ํŠธ๋žœ์Šคํฌ๋จธ ๋ ˆ์ด์–ด๋ฅผ ์ œ๊ฑฐํ•˜์—ฌ ๋‚จ์€ ์ง€์‹์„ "ํƒ์ƒ‰"ํ•˜๋ฉด "๋‚จ์€ ์ง€์‹"์ด ์ „์ฒด์˜ \(1-\frac{1}{L}\)๋ณด๋‹ค ํ›จ์”ฌ ์ ์„ ์ˆ˜ ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.19 ์ด๋Š” ์ง€์‹์ด ๊ฐœ๋ณ„ ๋ ˆ์ด์–ด๊ฐ€ ์•„๋‹Œ ์กฐํ•ฉ ์ž๋ฌผ์‡ ๊ฐ€ ์žˆ๋Š” ๊ธˆ๊ณ ์™€ ์œ ์‚ฌํ•œ ๋ณต์žกํ•œ ๋ฐฉ์‹์œผ๋กœ ์ €์žฅ๋˜๋ฉฐ, ํ•œ ๋ ˆ์ด์–ด๋ฅผ ์ œ๊ฑฐํ•˜๋ฉด ์ „์ฒด ์ง€์‹์˜ \(\frac{1}{L}\)๋ณด๋‹ค ํ›จ์”ฌ ๋” ๋งŽ์€ ์–‘์ด ์ œ๊ฑฐ๋  ์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 have seen that LLMs can efficiently compress knowledge into their parameter space, achieving 2bit/param even with 8-bit parameters. This raises the question: how and where is such knowledge stored? Our preliminary answer is that knowledge can be compactly stored within the model in a not-so-redundant manner. It is unlikely that the MLP layers alone store knowledge, as Attention layers, being of comparable sizes, also contribute to knowledge storage (c.f. Result 5). Moreover, particularly in models near the capacity boundary, removing the last transformer layer of an \(L\)-layer model to "probe" for remaining knowledge reveals that the "leftover knowledge" can be significantly less than \(1-\frac{1}{L}\) of the total.19 This suggests knowledge is stored not in individual layers but in a complex manner, akin to a safe with combination locks, where removing one layer may eliminate much more than \(\frac{1}{L}\) of the total knowledge. [|assistant|]
[|endofturn|]
**ํ‰๊ฐ€ ํ˜•์‹**. ํ‰๊ฐ€ ๊ฒฐ๊ณผ์˜ ์œ ํ˜•์— ๋”ฐ๋ผ ํ‰๊ฐ€ ํ˜•์‹์€ _์ ์ˆ˜ ๊ธฐ๋ฐ˜ ํ‰๊ฐ€_์™€ _์–ธ์–ด ๊ธฐ๋ฐ˜ ํ‰๊ฐ€_๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ ์ˆ˜ ๊ธฐ๋ฐ˜ ํ‰๊ฐ€๋Š” ์ธก์ • ๊ฐ€๋Šฅํ•œ ๋ฉ”ํŠธ๋ฆญ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€๋œ ํ…์ŠคํŠธ์— ๋Œ€ํ•œ ํ’ˆ์งˆ ์ ์ˆ˜(_์˜ˆ: ๋“ฑ๊ธ‰ ๋˜๋Š” ์ˆœ์œ„)๋ฅผ ํ• ๋‹นํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋ฐฉ๋ฒ•์€ ์Œ๋ณ„ ๋น„๊ต๋ฅผ ์ˆ˜ํ–‰ํ•˜์—ฌ ํŠน์ • ์ง€์นจ์— ๋”ฐ๋ผ ํ›„๋ณด ํ…์ŠคํŠธ์˜ ๋ถ€๋ถ„ ์ˆœ์„œ ๊ด€๊ณ„๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ, ํ‰๊ฐ€ ์ž‘์—…์„ ํฌ๊ฒŒ ๊ฐ„์†Œํ™”ํ•ฉ๋‹ˆ๋‹ค[354, 647, 727]. ๊ทธ๋Ÿฌ๋‚˜ ํ›„๋ณด์˜ ์ˆ˜๋ฅผ ๋Š˜๋ฆด ๋•Œ ๋น„ํšจ์œจ์ ์ธ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[727]. ํ‰๊ฐ€ ์ค‘์— ๊ณ ํ’ˆ์งˆ์˜ ์ฐธ์กฐ ํ…์ŠคํŠธ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ, LLM์€ ์ฐธ์กฐ์—์„œ ์ œ๊ณต๋œ ์ง€์นจ์— ๋”ฐ๋ผ ํ…์ŠคํŠธ์— ์ ์ˆ˜๋ฅผ ๋ถ€์—ฌํ•˜๋„๋ก ์ง€์‹œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[727, 728, 727, 716]. ๋ฐ˜๋ฉด์— ์–ธ์–ด ๊ธฐ๋ฐ˜ ํ‰๊ฐ€๋Š” ๋‹จ์ˆœํ•œ ์ •๋Ÿ‰์  ์ ์ˆ˜๋ฅผ ๋„˜์–ด ์งˆ์  ์„ค๋ช…์„ ์ œ๊ณตํ•˜๋Š” ๋น„ํ‰๊ณผ ์ œ์•ˆ์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค[889, 890, 371, 891, 892, 893, 894, 895, 371, 896, 372, 897, 898, 373, 899, 381, 899, 390, 391, 892, 893, 894, 895, 896, 897, 898, 999, 900, 899, 899, 910, 899, 899, 890, 891, 892, 893, 894, 895, 896, 897, 898, 899, 999, 900, 891, 899, 899, 892, 893, 894, 895, 896, 897, 898, 899, 988, 999, 990, 991, 992, 993, 994, 995, 996, 997, 998, 999, 1000, 1001, 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009, 1008, 1009, 1010, 1011, 1012, 1013, 1014, 1015, 1016, 1017, 1018, 1019, 1020, 1021, 1022, 1023, 1024, 1025, 1026, 1027, 1028, 1029, 1030, 1031, 1032, 1033, 1034, 1035, 1036, 1037, 1038, 1039, 1040, 1041, 1042, 1043, 1044, 1044, 1045, 1046, 1047, 1048, 1049, 1050, 1051, 1052, 1053, 1054, 1055, 1056, 1057, 1058, 1059, 1060, 1061, 1062, 1063, 1064, 1065, 1066, 1067, 1068, 1069, 1070, 1071, 1072, 1073, 1074, 1075, 1076, 1077, 1078, 1079, 1080, 1081, 1082, 1082, 1083, 1084, 1085, 1086, 1087, 1088, 1089, 1090, 1091, 1092, 1093, 1094, 1095, 1096, 1097, 1098, 1099, 1100, 1101, 1102, 1103, 1104, 1105, 1106, 1107, 1108, 1109, 1110, 1111, 1112, 1113, 1114, 1115, 1116, 1117, 1118, 1119, 1120, 1121, 1122, 1123, 1124, 1125, 1126, 1127, 1128, 1129, 1130, 1131, 1132, 1133, 1134, 1135, 1136, 1137, 1138, 1139, 1140, 1141, 1142, 1143, 1144, 1145, 1146, 1147, 1148, 1149, 1150, 1151, 1152, 1153, 1154, 1155, 1156, 1157, 1158, 1159, 1160, 1161, 1162, 1163, 1164, 1165, 1166, 1167, 1168, 1169, 1170, 1171, 1172, 1173, 1174, 1175, 1176, 1177, 1178, 1179, 1180, 1181, 1182, 1183, 1184, 1185, 1186, 1187, 1188, 1189, 1190, 1191, 1192, 1193, 1194, 1195, 1196, 1197, 1198, 1199, 1200, 1201, 1202, 1203, 1204, 1205, 1206, 1207, 1208, 1209, 1210, 1211, 1212, 1213, 1214, 1215, 1216, 1217, 1218, 1219, 1220, 1221, 1222, 1223, 1224, 1225, 1226, 1227, 1228, 1229, 1230, 1231, 1232, 1233, 1234, 1235, 1236, 1237, 1238, 1239, 1240, 1241, 1242, 1243, 1244, 1245, 1246, 1247, 1248, 1249, 1250, 1251, 1252, 1253, 1254, 1255, 1256, 1257, 1258, 1259, 1260, 1261, 1262, 1263, 1264, 1265, 1266, 1267, 1268, 1269, 1270, 1271, 1272, 1273, 1274, 1275, 1276, 1277, 1278, 1279, 1280, 1281, 1282, 1283, 1284, 1285, 1286, 1287, 1288, 1289, 1290, 1291, 1292, 1293, 1294, 1295, 1296, 1297, 1298, 1299, 1300, 1301, 1302, 1303, 1304, 1305, 1306, 1307, 1308, 1309, 1310, 1311, 1312, 1313, 1314, 1315, 1316, 1317, 1318, 1319, 1320, 1321, 1322, 1323, 1324, 1325, 1326, 1327, 1328, 1329, 1330, 1331, 1332, 1333, 1334, 1335, 1336, 1337, 1338, 1339, 1340, 1341, 1342, 1343, 1344, 1345, 1346, 1347, 1348, 1349, 1350, 1351, 1352, 1353, 1354, 1355, 1356, 1357, 1358, 1359, 1360, 1361, 1362, 1363, 1364, 1365, 1366, 1367, 1368, 1369, 1370, 1371, 1372, [|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Evaluation Formats.** Depending on the type of evaluation outcome, the evaluation format can be categorized into _score-based evaluation_ and _language-based evaluation_. Score-based evaluation employs measurable metrics to assign quality scores (_e.g._, ratings or rankings) for evaluated texts. A prevalent way is to conduct pairwise comparison, where LLMs are used to determine the partial order relation of candidate texts following specific guidelines [354, 647, 727], which greatly simplifies the evaluation task. However, it may face the inefficiency issue when scaling up the number of candidates [727]. When high-quality reference texts are available during evaluation, LLMs can be instructed to score texts under the guidance provided by references [727, 728, 716]. On the other hand, language-based evaluation focuses on generating critiques and suggestions, offering qualitative explanation beyond simple quantitative scoring [889, 890, 371, 891, 892, 893, 894, 895, 371, 896, 372, 897, 898, 373, 899, 381, 899, 390, 391, 892, 893, 894, 895, 896, 897, 898, 999, 900, 899, 899, 910, 899, 899, 890, 891, 892, 893, 894, 895, 896, 897, 898, 899, 999, 900, 891, 899, 899, 892, 893, 894, 895, 896, 897, 898, 899, 988, 999, 990, 991, 992, 993, 994, 995, 996, 997, 998, 999, 1000, 1001, 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009, 1008, 1009, 1010, 1011, 1012, 1013, 1014, 1015, 1016, 1017, 1018, 1019, 1020, 1021, 1022, 1023, 1024, 1025, 1026, 1027, 1028, 1029, 1030, 1031, 1032, 1033, 1034, 1035, 1036, 1037, 1038, 1039, 1040, 1041, 1042, 1043, 1044, 1044, 1045, 1046, 1047, 1048, 1049, 1050, 1051, 1052, 1053, 1054, 1055, 1056, 1057, 1058, 1059, 1060, 1061, 1062, 1063, 1064, 1065, 1066, 1067, 1068, 1069, 1070, 1071, 1072, 1073, 1074, 1075, 1076, 1077, 1078, 1079, 1080, 1081, 1082, 1preferences and the fairness of the evaluations made by LLMs [647, 727, 893, 897, 898]. As a representative benchmark, MT-Bench [727] evaluates the agreement between LLMs and human judgments, demonstrating that GPT-4 aligns closely with human preferences in no-tie comparisons on 80 multi-turn questions. In addition, to address potential biases arising from subjective human evaluations, LLMBar [897] manually designs outputs that are objectively worse but superficially appealing, which could mislead evaluators. The evaluation results reveal that even the most advanced LLMs still fall short of human-level evaluation in the challenging setting. [|assistant|]
๋žœ๋””๋Š” ๋†์žฅ์— ๋ง๊ณ ๋‚˜๋ฌด 60๊ทธ๋ฃจ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ง๊ณ ๋‚˜๋ฌด์˜ ์ ˆ๋ฐ˜๋ณด๋‹ค 5๊ทธ๋ฃจ ์ ์€ ์ฝ”์ฝ”๋„› ๋‚˜๋ฌด๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋žœ๋””์˜ ๋†์žฅ์—๋Š” ์ด ๋ช‡ ๊ทธ๋ฃจ์˜ ๋‚˜๋ฌด๊ฐ€ ์žˆ์„๊นŒ์š”?[|endofturn|]
๋žœ๋””๋Š” ๋†์žฅ์— ๋ง๊ณ ๋‚˜๋ฌด 60๊ทธ๋ฃจ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ง๊ณ ๋‚˜๋ฌด๋ณด๋‹ค ์ฝ”์ฝ”๋„› ๋‚˜๋ฌด๊ฐ€ ์ ˆ๋ฐ˜๋„ ์•ˆ ๋˜๋Š” 5๊ทธ๋ฃจ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋žœ๋””์˜ ๋†์žฅ์—๋Š” ์ด ๋ช‡ ๊ทธ๋ฃจ์˜ ๋‚˜๋ฌด๊ฐ€ ์žˆ์„๊นŒ์š”?[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Randy has 60 mango trees on his farm. He also has 5 less than half as many coconut trees as mango trees. How many trees does Randy have in all on his farm? [|assistant|]
**๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๊ณผ์ •์— ์–ด๋–ค ์‚ฌ๋žŒ๋“ค์ด ์ฐธ์—ฌํ–ˆ์œผ๋ฉฐ ์–ด๋–ป๊ฒŒ ๋ณด์ƒ์„ ๋ฐ›์•˜๋‚˜์š”?** AI ์•ˆ์ „ ํ‰๊ฐ€ ํ”„๋กฌํ”„ํŠธ์˜ taxonomy๋ฅผ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด AI ์•ˆ์ „ ์—ฐ๊ตฌ ๊ทธ๋ฃน ๊ตฌ์„ฑ์› ์™ธ์—๋„ 12๋ช…์˜ ์‹œ๋ฏผ ์‚ฌํšŒ ์ „๋ฌธ๊ฐ€๋ฅผ ์ธํ„ฐ๋ทฐํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
**๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๊ณผ์ •์— ์–ด๋–ค ์‚ฌ๋žŒ๋“ค์ด ์ฐธ์—ฌํ–ˆ์œผ๋ฉฐ ์–ด๋–ป๊ฒŒ ๋ณด์ƒ์„ ๋ฐ›์•˜๋‚˜์š”?** AI ์•ˆ์ „ ํ‰๊ฐ€ ํ”„๋กฌํ”„ํŠธ์˜ ๋ถ„๋ฅ˜๋ฅผ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด AI ์•ˆ์ „ ์—ฐ๊ตฌ ๊ทธ๋ฃน ๊ตฌ์„ฑ์› ์™ธ์—๋„ 12๋ช…์˜ ์‹œ๋ฏผ ์‚ฌํšŒ ์ „๋ฌธ๊ฐ€๋ฅผ ์ธํ„ฐ๋ทฐํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Who was involved in the data collection process and how were they compensated?** Apart from the AI-safety working group members, 12 civil society experts were interviewed for creating the taxonomy for AI safety evaluation prompts. [|assistant|]
**Long-form Medical Question Answering.** ์—ฌ๊ธฐ์„œ ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” MultiMedBench [12]์˜ HealthSearchQA, LiveQA ๋ฐ Medication QA์—์„œ 64๊ฐœ์˜ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€๊ฐ€ ์ž‘์„ฑํ•œ long-form ํ˜•์‹์˜ ๋‹ต๋ณ€์œผ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
**์žฅ๋ฌธ ์˜๋ฃŒ ์งˆ๋ฌธ ๋‹ต๋ณ€**. ์—ฌ๊ธฐ์„œ ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” MultiMedBench [12]์˜ HealthSearchQA, LiveQA ๋ฐ Medication QA์—์„œ 64๊ฐœ์˜ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€๊ฐ€ ์ž‘์„ฑํ•œ ์žฅ๋ฌธ ํ˜•์‹์˜ ๋‹ต๋ณ€์œผ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Long-form Medical Question Answering.** The dataset used here consisted of expert-crafted long-form responses to 64 questions from HealthSearchQA, LiveQA, and Medication QA in MultiMedBench [12]. [|assistant|]