text
stringlengths
74
4.35k
arxiv-translator
stringlengths
51
2.38k
deepl
stringlengths
33
2.26k
**LoRA**(Hu et al., 2021) We only freeze the encoder for the encoder-decoder LM and the entire model for the decoder-only LM. We use the optimal rank \(r\) of 4 and adapt both \(W_{q}\) and \(W_{v}\) in the self-attention module, which corresponds to the best performing hyperparameter setting in the paper.
**LoRA**(Hu ๋“ฑ, 2021) ์ธ์ฝ”๋”-๋””์ฝ”๋” LM์— ๋Œ€ํ•œ ์ธ์ฝ”๋”์™€ ๋””์ฝ”๋” ์ „์šฉ LM์— ๋Œ€ํ•œ ์ „์ฒด ๋ชจ๋ธ๋งŒ ๋™๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ ์˜ ์ˆœ์œ„ \(r\) 4๋ฅผ ์‚ฌ์šฉํ•˜๊ณ , ์ž์ฒด ์ฃผ์˜ ๋ชจ๋“ˆ์—์„œ \(W_{q}\)์™€ \(W_{v}\)๋ฅผ ๋ชจ๋‘ ์ ์šฉํ•˜๋ฉฐ, ์ด๋Š” ๋…ผ๋ฌธ์—์„œ ๊ฐ€์žฅ ์ž˜ ์ˆ˜ํ–‰๋˜๋Š” ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์„ค์ •์— ํ•ด๋‹นํ•œ๋‹ค.
**LoRA**(Hu et al., 2021) ์ธ์ฝ”๋”-๋””์ฝ”๋” LM์˜ ๊ฒฝ์šฐ ์ธ์ฝ”๋”๋งŒ ๊ณ ์ •ํ•˜๊ณ  ๋””์ฝ”๋” ์ „์šฉ LM์˜ ๊ฒฝ์šฐ ์ „์ฒด ๋ชจ๋ธ์„ ๊ณ ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ ์˜ ๋žญํฌ \(r\) 4๋ฅผ ์‚ฌ์šฉํ•˜๊ณ , ๋…ผ๋ฌธ์—์„œ ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์„ค์ •์— ํ•ด๋‹นํ•˜๋Š” ์ž๊ธฐ ์ฃผ์˜ ๋ชจ๋“ˆ์—์„œ \(W_{q}\) ๋ฐ \(W_{v}\) ๋ชจ๋‘๋ฅผ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค.
1. **Parallel rationale generation (think, Subsection 4.2)**: In parallel across \(n\) tokens \(x_{i}\) in an input sequence \(x_{0:n}\), we generate \(r\) rationales of length \(t\): \(c_{i}=(c_{i1},\dots,c_{it})\), resulting in \(n\times r\) rationale candidates. We insert learned \(<\)\(\mid\)start\(\mid\)\(>\) and \(<\)\(\mid\)end\(\mid\)\(>\) tokens to mark each rationale's start and end. 2. **Mixing post-rationale and base predictions (talk, Subsection 4.3)**: From the hidden state output after each rationale, we train a "mixing head" - a shallow MLP producing a weight determining how much the post-rationale next-token predicted logits should be incorporated compared to the base language model predicted logits. This approach eases distribution shift early in finetuning, due to introducing rationales. 3. **Optimizing rationale generation (learn, Subsection 4.4)**: We optimize the rationale generation parameters (start/end tokens and LM weights) to increase the likelihood of rationales that make future text more probable. We use REINFORCE to provide a learning signal to rationales based on their impact on future-token prediction. To reduce variance, we apply a teacher-forcing trick to include in the loss the likelihood of predicting not only the token after the thought but also later tokens.
1. **๋ณ‘๋ ฌ ๊ทผ๊ฑฐ ์ƒ์„ฑ(์‚ฌ๊ณ , ํ•˜์œ„ ์„น์…˜ 4.2)**: ์ž…๋ ฅ ์‹œํ€€์Šค \(x_{0:n}\)์—์„œ \(n\) ํ† ํฐ \(x_{i}\)์— ๊ฑธ์ณ ๋ณ‘๋ ฌ๋กœ \(r\) ๊ธธ์ด์˜ ๊ทผ๊ฑฐ \(t\): \(c_{i}=(c_{i1},\dots,c_{it})\๋ฅผ ์ƒ์„ฑํ•˜์—ฌ \(n\times r\)์˜ ๊ทผ๊ฑฐ ํ›„๋ณด๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต๋œ \(<\)\(\mid\)start\(\mid\)\(>\)์™€ \(<\)\(\mid\)end\(\mid\)\(>\) ํ† ํฐ์„ ์‚ฝ์ž…ํ•˜์—ฌ ๊ฐ ๊ทผ๊ฑฐ์˜ ์‹œ์ž‘๊ณผ ๋์„ ํ‘œ์‹œํ•œ๋‹ค. 2. **ํ•ฉ๋ฆฌ ํ›„ ๋ฐ ๊ธฐ๋ณธ ์˜ˆ์ธก ํ˜ผํ•ฉ(ํ† ํฌ, ํ•˜์œ„ ์„น์…˜ 4.3)**: ๊ฐ ๊ทผ๊ฑฐ ์ดํ›„์˜ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ ์ถœ๋ ฅ์—์„œ "ํ˜ผํ•ฉ ํ—ค๋“œ" - ๊ธฐ๋ณธ ์–ธ์–ด ๋ชจ๋ธ ์˜ˆ์ธก ๋กœ์ง“๊ณผ ๋น„๊ตํ•˜์—ฌ ํ•ฉ๋ฆฌ ํ›„ ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก ๋กœ์ง“์ด ์–ผ๋งˆ๋‚˜ ํ†ตํ•ฉ๋˜์–ด์•ผ ํ•˜๋Š”์ง€๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฐ€์ค‘์น˜๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์–•์€ MLP๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ๋ฒ•์€ ๊ทผ๊ฑฐ๋ฅผ ๋„์ž…ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฏธ์„ธ ์กฐ์ • ์ดˆ๊ธฐ์— ๋ถ„ํฌ ์ด๋™์„ ์™„ํ™”ํ•œ๋‹ค. 3. **๊ทผ๊ฑฐ ์ƒ์„ฑ ์ตœ์ ํ™”(ํ•™์Šต, ํ•˜์œ„ ์„น์…˜ 4.4)**: ๊ทผ๊ฑฐ ์ƒ์„ฑ ๋งค๊ฐœ ๋ณ€์ˆ˜(์‹œ์ž‘/๋ ํ† ํฐ ๋ฐ LM ๊ฐ€์ค‘์น˜)๋ฅผ ์ตœ์ ํ™”ํ•˜์—ฌ ๋ฏธ๋ž˜ ํ…์ŠคํŠธ๋ฅผ ๋” ํ™•๋ฅ ์ ์œผ๋กœ ๋งŒ๋“œ๋Š” ๊ทผ๊ฑฐ์˜ ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” REINFORCE๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏธ๋ž˜ ํ† ํฐ ์˜ˆ์ธก์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ถ”๋ก ์— ๋Œ€ํ•œ ํ•™์Šต ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•œ๋‹ค. ๋ถ„์‚ฐ ๊ฐ์†Œ๋ฅผ ์œ„ํ•ด ๊ต์‚ฌ ๊ฐ•์ œ ํŠธ๋ฆญ์„ ์ ์šฉํ•˜์—ฌ ์‚ฌ๊ณ  ํ›„ ํ† ํฐ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋‚˜์ค‘์— ํ† ํฐ์„ ์˜ˆ์ธกํ•  ๊ฐ€๋Šฅ์„ฑ์„ ์†์‹ค์— ํฌํ•จ์‹œํ‚จ๋‹ค.
1. **๋ณ‘๋ ฌ ๊ทผ๊ฑฐ ์ƒ์„ฑ(4.2์ ˆ ์ฐธ์กฐ)**: ์ž…๋ ฅ ์‹œํ€€์Šค \(x_{0:n}\)์—์„œ \(n\) ํ† ํฐ \(x_{i}\)์— ๊ฑธ์ณ ๋ณ‘๋ ฌ๋กœ ๊ธธ์ด \(t\)์˜ \(r\) ๊ทผ๊ฑฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค: \(c_{i}=(c_{i1},\dots,c_{it})\), ๊ทธ ๊ฒฐ๊ณผ \(n\๋ฐฐ r\) ๊ฐœ์˜ ๊ทผ๊ฑฐ ํ›„๋ณด๊ฐ€ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํ•™์Šต๋œ \(<\)\(\mid\)start\(\mid\)\(>\) ๋ฐ \(<\)\(\mid\)end\(\mid\)\(>\) ํ† ํฐ์„ ์‚ฝ์ž…ํ•˜์—ฌ ๊ฐ ๊ทผ๊ฑฐ์˜ ์‹œ์ž‘๊ณผ ๋์„ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. 2. **์‚ฌํ›„ ๊ทผ๊ฑฐ์™€ ๊ธฐ๋ณธ ์˜ˆ์ธก ํ˜ผํ•ฉ(ํ† ํฌ, ํ•˜์œ„ ์„น์…˜ 4.3)**: ๊ฐ ๊ทผ๊ฑฐ ์ดํ›„์˜ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ ์ถœ๋ ฅ์—์„œ, ๊ธฐ๋ณธ ์–ธ์–ด ๋ชจ๋ธ ์˜ˆ์ธก ๋กœ์ง“๊ณผ ๋น„๊ตํ•ด ์‚ฌํ›„ ๊ทผ๊ฑฐ ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก ๋กœ์ง“์ด ์–ผ๋งˆ๋‚˜ ํ†ตํ•ฉ๋˜์–ด์•ผ ํ•˜๋Š”์ง€๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฐ€์ค‘์น˜๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์–•์€ MLP์ธ 'ํ˜ผํ•ฉ ํ—ค๋“œ'๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ทผ๊ฑฐ๋ฅผ ๋„์ž…ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฏธ์„ธ ์กฐ์ • ์ดˆ๊ธฐ์— ๋ถ„ํฌ ์ด๋™์„ ์™„ํ™”ํ•ฉ๋‹ˆ๋‹ค. 3. **๊ทผ๊ฑฐ ์ƒ์„ฑ ์ตœ์ ํ™”(ํ•™์Šต, ํ•˜์œ„ ์„น์…˜ 4.4)**: ๊ทผ๊ฑฐ ์ƒ์„ฑ ๋งค๊ฐœ๋ณ€์ˆ˜(์‹œ์ž‘/์ข…๋ฃŒ ํ† ํฐ ๋ฐ LM ๊ฐ€์ค‘์น˜)๋ฅผ ์ตœ์ ํ™”ํ•˜์—ฌ ํ–ฅํ›„ ํ…์ŠคํŠธ์˜ ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์ด๋Š” ๊ทผ๊ฑฐ์˜ ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค. ๋ฏธ๋ž˜ ํ† ํฐ ์˜ˆ์ธก์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์— ๋”ฐ๋ผ ๊ทผ๊ฑฐ์— ํ•™์Šต ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด REINFORCE๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํŽธ์ฐจ๋ฅผ ์ค„์ด๊ธฐ ์œ„ํ•ด ๊ต์‚ฌ ๊ฐ•์ œ ํŠธ๋ฆญ์„ ์ ์šฉํ•˜์—ฌ ์ƒ๊ฐ ์ดํ›„์˜ ํ† ํฐ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ดํ›„ ํ† ํฐ์„ ์˜ˆ์ธกํ•  ๊ฐ€๋Šฅ์„ฑ๋„ ์†์‹ค์— ํฌํ•จ์‹œํ‚ต๋‹ˆ๋‹ค.
Table 4 summarizes by listing the time (# of samples) when the mixture models finally learn a GCD that a purely AI-trained model cannot learn, and the delay (in millions samples) since the previous GCD was learned (see also Figure 30 to illustrate the comparison between the clean and the AI-trained model):
ํ‘œ 4๋Š” ํ˜ผํ•ฉ๋ฌผ ๋ชจ๋ธ๋“ค์ด ์ˆœ์ˆ˜ํ•˜๊ฒŒ AI-ํŠธ๋ ˆ์ด๋‹๋œ ๋ชจ๋ธ์ด ํ•™์Šตํ•  ์ˆ˜ ์—†๋Š” GCD๋ฅผ ์ตœ์ข…์ ์œผ๋กœ ํ•™์Šตํ•˜๋Š” ์‹œ๊ฐ„(#์˜ ์ƒ˜ํ”Œ)์„ ๋‚˜์—ดํ•˜์—ฌ ์š”์•ฝํ•˜๊ณ , ์ด์ „ GCD๊ฐ€ ํ•™์Šต๋˜์—ˆ๊ธฐ ์ดํ›„์˜ ์ง€์—ฐ(๋ฐฑ๋งŒ ์ƒ˜ํ”Œ ๋‹จ์œ„)์„ ์š”์•ฝํ•œ๋‹ค(์ฒญ์ • ๋ชจ๋ธ๊ณผ AI-ํŠธ๋ ˆ์ด๋‹๋œ ๋ชจ๋ธ ๊ฐ„์˜ ๋น„๊ต๋ฅผ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ๋„ 30 ์ฐธ์กฐ):
ํ‘œ 4๋Š” ์ˆœ์ˆ˜ AI ํ•™์Šต ๋ชจ๋ธ๋กœ๋Š” ํ•™์Šตํ•  ์ˆ˜ ์—†๋Š” GCD๋ฅผ ํ˜ผํ•ฉ ๋ชจ๋ธ์ด ์ตœ์ข…์ ์œผ๋กœ ํ•™์Šตํ•˜๋Š” ์‹œ๊ฐ„(์ƒ˜ํ”Œ ์ˆ˜)๊ณผ ์ด์ „ GCD ํ•™์Šต ์ดํ›„ ์ง€์—ฐ ์‹œ๊ฐ„(์ˆ˜๋ฐฑ๋งŒ ์ƒ˜ํ”Œ ๋‹จ์œ„)์„ ๋‚˜์—ดํ•˜์—ฌ ์š”์•ฝํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค(ํด๋ฆฐ ๋ชจ๋ธ๊ณผ AI ํ•™์Šต ๋ชจ๋ธ ๊ฐ„์˜ ๋น„๊ต๋Š” ๊ทธ๋ฆผ 30์„ ์ฐธ์กฐํ•˜์„ธ์š”):
How to improve upon SLM?There are many natural extensions of SLM, _e.g.,_ reweighting tokens instead of selecting may improve robustness; using a reference model as a reward model to guide pretraining with reinforcement learning; adopting multiple reference models to reduce overfitting; designing token-level curriculum learning and iterative strategies for continuous improvements, _etc._
SLM์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ• SLM์˜ ๋งŽ์€ ์ž์—ฐ์ ์ธ ํ™•์žฅ, ์˜ˆ๋ฅผ ๋“ค์–ด ํ† ํฐ์„ ์„ ํƒํ•˜๋Š” ๋Œ€์‹  ์žฌ๊ฐ€์ค‘์น˜๋Š” ๊ฒฌ๊ณ ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค; ๊ฐ•ํ™” ํ•™์Šต์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์•ˆ๋‚ดํ•˜๊ธฐ ์œ„ํ•ด ๋ณด์ƒ ๋ชจ๋ธ๋กœ ์ฐธ์กฐ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ๋‹ค; ๊ณผ์ ํ•ฉ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์ฐธ์กฐ ๋ชจ๋ธ์„ ์ฑ„ํƒํ•œ๋‹ค; ์ง€์†์ ์ธ ๊ฐœ์„ ์„ ์œ„ํ•ด ํ† ํฐ ์ˆ˜์ค€ ์ปค๋ฆฌํ˜๋Ÿผ ํ•™์Šต ๋ฐ ๋ฐ˜๋ณต ์ „๋žต์„ ์„ค๊ณ„ํ•œ๋‹ค. _etc_
์„ ํƒ ๋Œ€์‹  ํ† ํฐ์— ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜๋ฉด ๊ฒฌ๊ณ ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๊ณ , ๊ฐ•ํ™” ํ•™์Šต์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์•ˆ๋‚ดํ•˜๊ธฐ ์œ„ํ•ด ์ฐธ์กฐ ๋ชจ๋ธ์„ ๋ณด์ƒ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•˜๊ณ , ๊ณผ์ ํ•ฉ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์ฐธ์กฐ ๋ชจ๋ธ์„ ์ฑ„ํƒํ•˜๊ณ , ์ง€์†์ ์ธ ๊ฐœ์„ ์„ ์œ„ํ•ด ํ† ํฐ ์ˆ˜์ค€์˜ ์ปค๋ฆฌํ˜๋Ÿผ ํ•™์Šต๊ณผ ๋ฐ˜๋ณต ์ „๋žต์„ ์„ค๊ณ„ํ•˜๋Š” ๋“ฑ SLM์„ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์ด ๋งŽ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
language models, which have led to the emergence of a variety of popular models, including LLaVA [149], MiniGPT-4 [150], InstructBLIP [151], and PandaGPT [152]. The release of LLMA has greatly advanced the research progress of LLMs. To summarize the research work conducted on LLaMA, we present a brief evolutionary graph in Figure 5.
์–ธ์–ด ๋ชจ๋ธ์€ LLaVA[149], MiniGPT-4[150], InstructBLIP[151], PandaGPT[152] ๋“ฑ ๋‹ค์–‘ํ•œ ์ธ๊ธฐ ๋ชจ๋ธ์ด ๋“ฑ์žฅํ•˜๊ฒŒ ๋˜์—ˆ๋‹ค. LLMA์˜ ์ถœ์‹œ๋Š” LLM์˜ ์—ฐ๊ตฌ ์ง„์ „์„ ํฌ๊ฒŒ ๋ฐœ์ „์‹œ์ผฐ๋‹ค. LLaMA์— ๋Œ€ํ•ด ์ˆ˜ํ–‰๋œ ์—ฐ๊ตฌ ์ž‘์—…์„ ์š”์•ฝํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 5์— ๊ฐ„๋žตํ•œ ์ง„ํ™” ๊ทธ๋ž˜ํ”„๋ฅผ ์ œ์‹œํ•œ๋‹ค.
์–ธ์–ด ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ๋†’์•„์ง€๋ฉด์„œ LLaVA [149], MiniGPT-4 [150], InstructBLIP [151], PandaGPT [152] ๋“ฑ ๋‹ค์–‘ํ•œ ์ธ๊ธฐ ๋ชจ๋ธ์ด ๋“ฑ์žฅํ•˜๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. LLMA์˜ ์ถœ์‹œ๋Š” LLM์˜ ์—ฐ๊ตฌ ๋ฐœ์ „์„ ํฌ๊ฒŒ ์ง„์ „์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. LLaMA์— ๋Œ€ํ•ด ์ˆ˜ํ–‰๋œ ์—ฐ๊ตฌ ์ž‘์—…์„ ์š”์•ฝํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 5์— ๊ฐ„๋žตํ•œ ์ง„ํ™” ๊ทธ๋ž˜ํ”„๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
1. Burpees: Burpees are a full-body workout that combines squats, a push-up, and a jump to produce a high number of calories burned in a short amount of time.
1. ๋ฒ„ํ”ผ: ๋ฒ„ํ”ผ๋Š” ์Šค์ฟผํŠธ, ํŒ”๊ตฝํ˜€ํŽด๊ธฐ, ์ ํ”„๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ ์งง์€ ์‹œ๊ฐ„์— ๋งŽ์€ ์นผ๋กœ๋ฆฌ๋ฅผ ์†Œ๋ชจํ•˜๋Š” ์ „์‹  ์šด๋™์ด๋‹ค.
1. ๋ฒ„ํ”ผ: ๋ฒ„ํ”ผ๋Š” ์Šค์ฟผํŠธ, ํŒ”๊ตฝํ˜€ํŽด๊ธฐ, ์ ํ”„๋ฅผ ๊ฒฐํ•ฉํ•œ ์ „์‹  ์šด๋™์œผ๋กœ ์งง์€ ์‹œ๊ฐ„์— ๋งŽ์€ ์นผ๋กœ๋ฆฌ๋ฅผ ์†Œ๋ชจํ•  ์ˆ˜ ์žˆ๋Š” ์šด๋™์ž…๋‹ˆ๋‹ค.
We begin by validating that the attacks developed in the prior sections can effectively recover the full logit vector given a limited query interface. In Table 3 we report the average number of bits of agreement between the true logit vector and the recovered logit vector, as well as the (amortized) number of queries required to recover one full logit vector.
์ด์ „ ์„น์…˜์—์„œ ๊ฐœ๋ฐœ๋œ ๊ณต๊ฒฉ๋“ค์ด ์ œํ•œ๋œ ์งˆ์˜ ์ธํ„ฐํŽ˜์ด์Šค์—์„œ ์ „์ฒด ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ฒ€์ฆํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•œ๋‹ค. ํ‘œ 3์—์„œ ์šฐ๋ฆฌ๋Š” ์ง„์ •ํ•œ ๋กœ์ง“ ๋ฒกํ„ฐ์™€ ๋ณต๊ตฌ๋œ ๋กœ์ง“ ๋ฒกํ„ฐ ์‚ฌ์ด์˜ ํ‰๊ท  ์ผ์น˜ ๋น„ํŠธ ์ˆ˜์™€ ํ•˜๋‚˜์˜ ์™„์ „ํ•œ ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ๋ณต๊ตฌํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ (์ƒ๊ฐ๋œ) ์ฟผ๋ฆฌ ์ˆ˜๋ฅผ ๋ณด๊ณ ํ•œ๋‹ค.
๋จผ์ € ์ด์ „ ์„น์…˜์—์„œ ๊ฐœ๋ฐœํ•œ ๊ณต๊ฒฉ์ด ์ œํ•œ๋œ ์ฟผ๋ฆฌ ์ธํ„ฐํŽ˜์ด์Šค์—์„œ ์ „์ฒด ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 3์—์„œ๋Š” ์‹ค์ œ ๋กœ์ง“ ๋ฒกํ„ฐ์™€ ๋ณต๊ตฌ๋œ ๋กœ์ง“ ๋ฒกํ„ฐ ๊ฐ„์˜ ํ‰๊ท  ์ผ์น˜ ๋น„ํŠธ ์ˆ˜์™€ ํ•˜๋‚˜์˜ ์ „์ฒด ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ๋ณต๊ตฌํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ (์ƒ๊ฐ๋œ) ์ฟผ๋ฆฌ ์ˆ˜๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.
Footnote 5: We do not append a <eos> token at the end of documents because we only use the first section, which does not signify the conclusion of the entire article.
๊ฐ์ฃผ 5: ์ „์ฒด ๊ธฐ์‚ฌ์˜ ๊ฒฐ๋ก ์„ ์˜๋ฏธํ•˜์ง€ ์•Š๋Š” ์ฒซ ๋ฒˆ์งธ ์„น์…˜๋งŒ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฌธ์„œ ๋์— <eos> ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜์ง€ ์•Š๋Š”๋‹ค.
๊ฐ์ฃผ 5: ์ „์ฒด ๋ฌธ์„œ์˜ ๊ฒฐ๋ก ์„ ์˜๋ฏธํ•˜์ง€ ์•Š๋Š” ์ฒซ ๋ฒˆ์งธ ์„น์…˜๋งŒ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฌธ์„œ ๋์— <eos> ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
PersonA: Urinalysis is a test that evaluates a sample of your urine. It is used to detect and assess a wide range of disorders, such as urinary tract infection, kidney disease, and diabetes. It examines the appearance, concentration, and content of urine.
PersonA: ์†Œ๋ณ€๊ฒ€์‚ฌ๋Š” ๋‹น์‹ ์˜ ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒ€์‚ฌ์ด๋‹ค. ์š”๋กœ ๊ฐ์—ผ, ์‹ ์žฅ ์งˆํ™˜, ๋‹น๋‡จ๋ณ‘๊ณผ ๊ฐ™์€ ๊ด‘๋ฒ”์œ„ํ•œ ์žฅ์• ๋ฅผ ๊ฐ์ง€ํ•˜๊ณ  ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋œ๋‹ค. ์†Œ๋ณ€์˜ ๋ชจ์–‘, ๋†๋„, ํ•จ์œ ๋Ÿ‰ ๋“ฑ์„ ๊ฒ€์‚ฌํ•œ๋‹ค.
PersonA: ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒ€์‚ฌ์ž…๋‹ˆ๋‹ค. ์š”๋กœ ๊ฐ์—ผ, ์‹ ์žฅ ์งˆํ™˜, ๋‹น๋‡จ๋ณ‘ ๋“ฑ ๋‹ค์–‘ํ•œ ์งˆํ™˜์„ ๊ฐ์ง€ํ•˜๊ณ  ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์†Œ๋ณ€์˜ ๋ชจ์–‘, ๋†๋„ ๋ฐ ํ•จ๋Ÿ‰์„ ๊ฒ€์‚ฌํ•ฉ๋‹ˆ๋‹ค.
Manually constructed X-of-thought methods expand on in-context learning by adding different types of step-by-step intermediate reasoning processes to demonstrations. They allow LLMs to mimic and generate reasoning paths. Although manual XoT methods provide greater interpretability as well as trustworthiness for human understanding and outperform on complex tasks, i.e., mathematical reasoning, commonsense reasoning, symbolic reasoning, etc., manual annotating of rationales entails significant costs and suffers from drawbacks such as difficulty in demonstration selection and task generalization. Specifically, different tasks require different ways of demonstrations. Therefore, other works attempt to construct the reasoning path automatically, as discussed in SS4.1.2.
์ˆ˜๋™์œผ๋กœ ๊ตฌ์„ฑ๋œ X-of-thought ๋ฐฉ๋ฒ•์€ ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ๋‹จ๊ณ„๋ณ„ ์ค‘๊ฐ„ ์ถ”๋ก  ๊ณผ์ •์„ ๋ฐ๋ชจ์— ์ถ”๊ฐ€ํ•˜์—ฌ ์ธ-์ปจํ…์ŠคํŠธ ํ•™์Šต์„ ํ™•์žฅํ•œ๋‹ค. ๊ทธ๋“ค์€ LLM๋“ค์ด ์ถ”๋ก  ๊ฒฝ๋กœ๋“ค์„ ๋ชจ๋ฐฉํ•˜๊ณ  ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•œ๋‹ค. ์ˆ˜๋™ XoT ๋ฐฉ๋ฒ•์€ ๋ณต์žกํ•œ ์ž‘์—…, ์ฆ‰ ์ˆ˜ํ•™์  ์ถ”๋ก , ์ƒ์‹ ์ถ”๋ก , ์ƒ์ง• ์ถ”๋ก  ๋“ฑ์— ๋Œ€ํ•œ ์ธ๊ฐ„์˜ ์ดํ•ด์™€ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ์‹ ๋ขฐ์„ฑ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋” ํฐ ํ•ด์„ ๊ฐ€๋Šฅ์„ฑ์„ ์ œ๊ณตํ•˜์ง€๋งŒ, ๊ทผ๊ฑฐ์˜ ์ˆ˜๋™ ์ฃผ์„์€ ์ƒ๋‹นํ•œ ๋น„์šฉ์„ ์ˆ˜๋ฐ˜ํ•˜๊ณ  ์‹œ์—ฐ ์„ ํƒ ๋ฐ ์ž‘์—… ์ผ๋ฐ˜ํ™”์˜ ์–ด๋ ค์›€๊ณผ ๊ฐ™์€ ๋‹จ์ ์„ ๊ฒช๋Š”๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ๋‹ค๋ฅธ ์ž‘์—…๋“ค์€ ๋‹ค๋ฅธ ๋ฐฉ์‹์˜ ๋ฐ๋ชจ๋ฅผ ํ•„์š”๋กœ ํ•œ๋‹ค. ๋”ฐ๋ผ์„œ, ๋‹ค๋ฅธ ์ž‘์—…๋“ค์€ SS4.1.2์—์„œ ๋…ผ์˜๋œ ๋ฐ”์™€ ๊ฐ™์ด ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์ž๋™์œผ๋กœ ๊ตฌ์„ฑํ•˜๋ ค๊ณ  ์‹œ๋„ํ•œ๋‹ค.
์ˆ˜๋™์œผ๋กœ ๊ตฌ์ถ•๋œ X-of-thought ๋ฐฉ๋ฒ•์€ ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ๋‹จ๊ณ„๋ณ„ ์ค‘๊ฐ„ ์ถ”๋ก  ๊ณผ์ •์„ ๋ฐ๋ชจ์— ์ถ”๊ฐ€ํ•˜์—ฌ ์ƒํ™ฉ ๋‚ด ํ•™์Šต์„ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด LLM์€ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ๋ชจ๋ฐฉํ•˜๊ณ  ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ˆ˜๋™ XoT ๋ฐฉ๋ฒ•์€ ์ˆ˜ํ•™์  ์ถ”๋ก , ์ƒ์‹์  ์ถ”๋ก , ์ƒ์ง•์  ์ถ”๋ก  ๋“ฑ ๋ณต์žกํ•œ ์ž‘์—…์—์„œ ์ธ๊ฐ„์˜ ์ดํ•ด์— ๋Œ€ํ•œ ์‹ ๋ขฐ์„ฑ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ํ•ด์„ ๊ฐ€๋Šฅ์„ฑ๋„ ๋›ฐ์–ด๋‚˜์ง€๋งŒ, ์ถ”๋ก ์˜ ์ˆ˜๋™ ์ฃผ์„ ๋‹ฌ๊ธฐ๋Š” ์ƒ๋‹นํ•œ ๋น„์šฉ์„ ์ˆ˜๋ฐ˜ํ•˜๊ณ  ๋ฐ๋ชจ ์„ ํƒ ๋ฐ ์ž‘์—… ์ผ๋ฐ˜ํ™”์˜ ์–ด๋ ค์›€๊ณผ ๊ฐ™์€ ๋‹จ์ ์„ ์•ˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ๊ณผ์ œ๋งˆ๋‹ค ์„œ๋กœ ๋‹ค๋ฅธ ์‹œ์—ฐ ๋ฐฉ์‹์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ SS4.1.2์—์„œ ์„ค๋ช…ํ•œ ๊ฒƒ์ฒ˜๋Ÿผ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ์ž๋™์œผ๋กœ ๊ตฌ์„ฑํ•˜๋ ค๋Š” ์ž‘์—…๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Human alignment_. For human alignment, we select TruthfulQA [556] to measure whether a LLM is truthful in generating answers to questions, CrowS-Pairs [603] and WinoSender [604] to assess the stereotypes in LLMs, RealIoxicityPromps [605] to evaluate the extent to which LLMs generate toxic language, and HaluEval [602] to test the ability of LLMs to recognize hallucination. As the test set of Real-Toxicity-Prompts is too large, we randomly sample 10000 examples from it for evaluation. We follow LLaMA [57] to report the zero-shot performance, and compute the _accuracy_ of identifying a claim as true for TruthfulQA, _accuracy_ of recognizing biased sentences (high perplexity) for CrowS-Pairs, _corference resolution accuracy (the/she/they)_ for WinoGender, _toxicity score_ for RealToxicityPrompts, and _average accuracy_ of recognizing hallucinations for HaluEval. For TruthfulQA, we follow existing work [57] that utilizes text-davinci-003 to replace humans for scoring. For CrowS-Pairs and WinoGender, we follow the experimental settings of LLaMA [57] to compute the perplexity and coreference resolution score. For RealToxicityPrompts, we utilize the Perspective-API47 for toxicity evaluation.
\(\bullet\)_์ธ๊ฐ„ ์ •๋ ฌ_. ์ธ๊ฐ„ ์ •๋ ฌ์„ ์œ„ํ•ด TruthfulQA [556]์„ ์„ ํƒํ•˜์—ฌ LLM์ด ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ง„์‹ค์„ฑ ์—ฌ๋ถ€๋ฅผ ์ธก์ •ํ•˜๊ณ , CrowS-Pairs [603] ๋ฐ WinoSender [604]๋ฅผ ์„ ํƒํ•˜์—ฌ LLM์—์„œ ๊ณ ์ •๊ด€๋…์„ ํ‰๊ฐ€ํ•˜๊ณ , RealIoxicityPromps [605]๋ฅผ ์„ ํƒํ•˜์—ฌ LLM์ด ๋…์„ฑ ์–ธ์–ด๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ •๋„๋ฅผ ํ‰๊ฐ€ํ•˜๊ณ , HaluEval [602]๋ฅผ ์„ ํƒํ•˜์—ฌ LLM์ด ํ™˜๊ฐ์„ ์ธ์‹ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•œ๋‹ค. ์‹ค์ œ ๋…์„ฑ ํ”„๋กฌํ”„ํŠธ์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๊ฐ€ ๋„ˆ๋ฌด ํฌ๊ธฐ ๋•Œ๋ฌธ์— ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด 10000๊ฐœ์˜ ์˜ˆ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋งํ•œ๋‹ค. ์šฐ๋ฆฌ๋Š” LLaMA [57]์„ ๋”ฐ๋ผ ์ œ๋กœ์ƒท ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•˜๊ณ , TruthfulQA์— ๋Œ€ํ•ด ์ฐธ์œผ๋กœ ํด๋ ˆ์ž„์„ ์‹๋ณ„ํ•˜๋Š” _์ •ํ™•๋„_, CrowS-Pairs์— ๋Œ€ํ•ด ํŽธํ–ฅ๋œ ๋ฌธ์žฅ(๋†’์€ ๋ณต์žก๋„)์„ ์ธ์‹ํ•˜๋Š” _์ •ํ™•๋„_, WinoGender์— ๋Œ€ํ•ด _์ƒ๊ด€ ๋ถ„ํ•ด ์ •ํ™•๋„(the/she/they)_, RealToxicityPrompt์— ๋Œ€ํ•ด _๋…์„ฑ ์ ์ˆ˜_ ๋ฐ HaluEval์— ๋Œ€ํ•ด ํ™˜๊ฐ์„ ์ธ์‹ํ•˜๋Š” _ํ‰๊ท  ์ •ํ™•๋„_๋ฅผ ๊ณ„์‚ฐํ•œ๋‹ค. TruthfulQA์˜ ๊ฒฝ์šฐ ํ…์ŠคํŠธ ๋‹ค๋นˆ์น˜-003์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฑ„์ ์„ ์œ„ํ•ด ์‚ฌ๋žŒ์„ ๋Œ€์ฒดํ•˜๋Š” ๊ธฐ์กด ์ž‘์—… [57]์„ ๋”ฐ๋ฅธ๋‹ค. CrowS-Pairs์™€ WinoGender์˜ ๊ฒฝ์šฐ, ์šฐ๋ฆฌ๋Š” ๋ณต์žก๋„์™€ ์ƒํ˜ธ์ฐธ์กฐ ํ•ด์ƒ๋„ ์ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด LLaMA [57]์˜ ์‹คํ—˜ ์„ค์ •์„ ๋”ฐ๋ฅธ๋‹ค. RealToxicityPrompts์˜ ๊ฒฝ์šฐ ๋…์„ฑ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด Perspective-API47์„ ํ™œ์šฉํ•œ๋‹ค.
(\(\๋ถˆ๋ฆฟ\)_์ธ๊ฐ„ ์ •๋ ฌ_. ์ธ๊ฐ„ ์ •๋ ฌ์˜ ๊ฒฝ์šฐ, LLM์ด ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•  ๋•Œ ์ง„์‹คํ•œ์ง€ ์—ฌ๋ถ€๋ฅผ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด TruthfulQA [556], LLM์˜ ๊ณ ์ •๊ด€๋…์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด CrowS-Pairs [603] ๋ฐ WinoSender [604], LLM์ด ๋…์„ฑ ์–ธ์–ด๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ •๋„๋ฅผ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด RealIoxicityPromps [605], LLM์˜ ํ™˜๊ฐ์„ ์ธ์‹ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•˜๋Š” HaluEval [602]๋ฅผ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ ๋…์„ฑ ํ”„๋กฌํ”„ํŠธ์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๊ฐ€ ๋„ˆ๋ฌด ๋งŽ๊ธฐ ๋•Œ๋ฌธ์— ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ๋ฌด์ž‘์œ„๋กœ 10000๊ฐœ์˜ ์˜ˆ์‹œ๋ฅผ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค. LLaMA [57]๋ฅผ ๋”ฐ๋ผ ์ œ๋กœ ์ƒท ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•˜๊ณ , TruthfulQA์˜ ๊ฒฝ์šฐ ์ฃผ์žฅ์„ ์ฐธ์œผ๋กœ ์‹๋ณ„ํ•˜๋Š” _์ •ํ™•๋„_, CrowS-Pairs์˜ ๊ฒฝ์šฐ ํŽธํ–ฅ๋œ ๋ฌธ์žฅ(๋†’์€ ๋‚œํ•ด์„ฑ)์„ ์ธ์‹ํ•˜๋Š” _์ •ํ™•๋„_, WinoGender์˜ ๊ฒฝ์šฐ _์—ฐ๊ด€์„ฑ ํ•ด๊ฒฐ ์ •ํ™•๋„(the/she/they)_, RealToxicityPrompts์˜ ๊ฒฝ์šฐ _๋…์„ฑ ์ ์ˆ˜_, HaluEval์˜ ๊ฒฝ์šฐ _ํ‰๊ท  ํ™˜๊ฐ ์ธ์‹_์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. TruthfulQA์˜ ๊ฒฝ์šฐ, ์‚ฌ๋žŒ์„ ๋Œ€์ฒดํ•˜์—ฌ ์ฑ„์ ํ•˜๋Š” ๋ฐ ํ…์ŠคํŠธ-davinci-003์„ ํ™œ์šฉํ•˜๋Š” ๊ธฐ์กด ์ž‘์—…[57]์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. CrowS-Pairs์™€ WinoGender์˜ ๊ฒฝ์šฐ, ๋‹นํ˜น๊ฐ ๋ฐ ํ•ต์‹ฌ ์„ ํ˜ธ๋„ ์ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด LLaMA[57]์˜ ์‹คํ—˜ ์„ค์ •์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์‹ค์ œ ๋…์„ฑ ํ”„๋กฌํ”„ํŠธ์˜ ๊ฒฝ์šฐ, ๋…์„ฑ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด Perspective-API47์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค.
BaselinesWe also train several baselines. Two baselines are trained on \(\mathcal{D}_{0}\) and \(\mathcal{D}_{1}\) respectively while the third is trained on \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\). We consider the model trained on \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) to be an upper bound on performance as it represents an expensive full re-training. The baselines trained on individual datasets can be seen as compute-equivalent alternatives to continual pre-training (e.g., one could opt to train a model from random initialization on \(\mathcal{D}_{1}\) instead of continually pre-training it).
๊ธฐ์ค€์„  ์šฐ๋ฆฌ๋Š” ๋˜ํ•œ ์—ฌ๋Ÿฌ ๊ธฐ์ค€์„ ์„ ํ›ˆ๋ จํ•œ๋‹ค. ๋‘ ๊ฐœ์˜ ๊ธฐ์ค€์„ ์€ ๊ฐ๊ฐ \(\mathcal{D}_{0}\)์™€ \(\mathcal{D}_{1}\)์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋˜๊ณ  ์„ธ ๋ฒˆ์งธ ๊ธฐ์ค€์„ ์€ \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\)์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋œ๋‹ค. ์šฐ๋ฆฌ๋Š” \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\)๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ๊ณ ๊ฐ€์˜ ์™„์ „ ์žฌํ•™์Šต์„ ๋‚˜ํƒ€๋‚ด๊ธฐ ๋•Œ๋ฌธ์— ์„ฑ๋Šฅ์˜ ์ƒํ•œ์œผ๋กœ ๊ฐ„์ฃผํ•œ๋‹ค. ๊ฐœ๋ณ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋œ ๊ธฐ์ค€์„ ์€ ์—ฐ์† ์‚ฌ์ „ ํ›ˆ๋ จ์— ๋Œ€ํ•œ ๊ณ„์‚ฐ๊ณผ ๋™๋“ฑํ•œ ๋Œ€์•ˆ์œผ๋กœ ๋ณผ ์ˆ˜ ์žˆ๋‹ค(์˜ˆ๋ฅผ ๋“ค์–ด, ์—ฐ์† ์‚ฌ์ „ ํ›ˆ๋ จ ๋Œ€์‹  \(\mathcal{D}_{1}\)์—์„œ ๋ฌด์ž‘์œ„ ์ดˆ๊ธฐํ™”๋กœ๋ถ€ํ„ฐ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋„๋ก ์„ ํƒํ•  ์ˆ˜ ์žˆ๋‹ค).
๊ธฐ์ค€์„ ์šฐ๋ฆฌ๋Š” ๋˜ํ•œ ์—ฌ๋Ÿฌ ๊ธฐ์ค€์„ ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๊ฐœ์˜ ๊ธฐ์ค€์„ ์€ ๊ฐ๊ฐ \(\mathcal{D}_{0}\) ๋ฐ \(\mathcal{D}_{1}\)์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋˜๊ณ , ์„ธ ๋ฒˆ์งธ ๊ธฐ์ค€์„ ์€ \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\)์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋ฉ๋‹ˆ๋‹ค. (\(\mathcal{D}_{0}\cup\mathcal{D}_{1}\)์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์€ ๋น„์šฉ์ด ๋งŽ์ด ๋“œ๋Š” ์ „์ฒด ์žฌํ›ˆ๋ จ์„ ๋‚˜ํƒ€๋‚ด๋ฏ€๋กœ ์„ฑ๋Šฅ์˜ ์ƒํ•œ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ๊ฐœ๋ณ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋œ ๊ธฐ์ค€์„ ์€ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์— ๋Œ€ํ•œ ์ปดํ“จํŒ… ๋™๋“ฑํ•œ ๋Œ€์•ˆ์œผ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จํ•˜๋Š” ๋Œ€์‹  \(\mathcal{D}_{1}\)์—์„œ ๋ฌด์ž‘์œ„ ์ดˆ๊ธฐํ™”๋ฅผ ํ†ตํ•ด ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ์Œ).
In the following experiments (SS 4 and SS 5), we use Llama-2 (7B and 70B) [11] since it is one of the best-performing LLMs. We use Wikipedia articles classified under the "2023" Category including topics from diverse domains such as films, arts, economics, politics, events, etc.3 The likelihood that this factual information is not
๋‹ค์Œ ์‹คํ—˜(SS 4 ๋ฐ SS 5)์—์„œ Llama-2(7B ๋ฐ 70B) [11]์€ ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ LLM ์ค‘ ํ•˜๋‚˜์ด๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์šฉํ•œ๋‹ค. ์šฐ๋ฆฌ๋Š” ์˜ํ™”, ์˜ˆ์ˆ , ๊ฒฝ์ œ, ์ •์น˜, ์‚ฌ๊ฑด ๋“ฑ๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์˜์—ญ์˜ ์ฃผ์ œ๋ฅผ ํฌํ•จํ•˜๋Š” "2023" ๋ฒ”์ฃผ์— ๋”ฐ๋ผ ๋ถ„๋ฅ˜๋œ ์œ„ํ‚คํ”ผ๋””์•„ ๊ธฐ์‚ฌ๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. 3 ์ด๋Ÿฌํ•œ ์‚ฌ์‹ค ์ •๋ณด๊ฐ€ ์•„๋‹ ๊ฐ€๋Šฅ์„ฑ์€
๋‹ค์Œ ์‹คํ—˜(SS 4์™€ SS 5)์—์„œ๋Š” ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ LLM ์ค‘ ํ•˜๋‚˜์ธ Llama-2(7B์™€ 70B)[11]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ํ™”, ์˜ˆ์ˆ , ๊ฒฝ์ œ, ์ •์น˜, ์‚ฌ๊ฑด ๋“ฑ ๋‹ค์–‘ํ•œ ์˜์—ญ์˜ ์ฃผ์ œ๋ฅผ ํฌํ•จํ•˜์—ฌ "2023" ์นดํ…Œ๊ณ ๋ฆฌ๋กœ ๋ถ„๋ฅ˜๋œ ์œ„ํ‚ค๋ฐฑ๊ณผ ๋ฌธ์„œ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.3 ์ด ์‚ฌ์‹ค์  ์ •๋ณด๊ฐ€
**Rationale Distillation.** For rationale generation from the teacher model, we randomly sampled 15K examples from each of the two datasets, resulting in a total of 30K examples. After filtering invalid rationales, the dataset was reduced to a total of 23K examples. 90% of data was used for training, and the remaining 10% was used for validation. For training, we used 8 NVIDIA Tesla V100 GPUs with 16GB configurations. Hyperparameters for training are as follows: 3e-3 for learning rates, 5 epochs, 64 batch size.
**Rationale Distillation.** ๊ต์‚ฌ ๋ชจ๋ธ์—์„œ ๊ทผ๊ฑฐ ์ƒ์„ฑ์„ ์œ„ํ•ด ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐ๊ฐ์—์„œ 15K ๊ฐœ์˜ ์˜ˆ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ์ด 30K ๊ฐœ์˜ ์˜ˆ๋ฅผ ์ƒ์„ฑํ–ˆ๋‹ค. ์ž˜๋ชป๋œ ๊ทผ๊ฑฐ๋ฅผ ํ•„ํ„ฐ๋งํ•œ ํ›„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ด 23K๊ฐœ์˜ ์˜ˆ์ œ๋กœ ์ถ•์†Œ๋˜์—ˆ๋‹ค. 90%์˜ ๋ฐ์ดํ„ฐ๋Š” ํ›ˆ๋ จ์— ์‚ฌ์šฉ๋˜์—ˆ๊ณ  ๋‚˜๋จธ์ง€ 10%๋Š” ๊ฒ€์ฆ์— ์‚ฌ์šฉ๋˜์—ˆ๋‹ค. ๊ต์œก์„ ์œ„ํ•ด 16GB ๊ตฌ์„ฑ์˜ NVIDIA ํ…Œ์Šฌ๋ผ V100 GPU 8๊ฐœ๋ฅผ ์‚ฌ์šฉํ–ˆ๋‹ค. ํ•™์Šต์— ๋Œ€ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ํ•™์Šต ์†๋„์— ๋Œ€ํ•œ 3e-3, 5 ์—ํญ, 64 ๋ฐฐ์น˜ ํฌ๊ธฐ์ž…๋‹ˆ๋‹ค.
**๊ต์‚ฌ ๋ชจ๋ธ์—์„œ ๊ทผ๊ฑฐ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฐ๊ฐ 15,000๊ฐœ์˜ ์‚ฌ๋ก€๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ์ด 30,000๊ฐœ์˜ ์‚ฌ๋ก€๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์œ ํšจํ•˜์ง€ ์•Š์€ ๊ทผ๊ฑฐ๋ฅผ ํ•„ํ„ฐ๋งํ•œ ํ›„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ด 23,000๊ฐœ์˜ ์‚ฌ๋ก€๋กœ ์ถ•์†Œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ์˜ 90%๋Š” ํ›ˆ๋ จ์— ์‚ฌ์šฉ๋˜์—ˆ๊ณ  ๋‚˜๋จธ์ง€ 10%๋Š” ๊ฒ€์ฆ์— ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ์—๋Š” 16GB ๊ตฌ์„ฑ์˜ NVIDIA Tesla V100 GPU 8๊ฐœ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•™์Šต์„ ์œ„ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค: ํ•™์Šต ์†๋„ 3e-3, ์—ํฌํฌ 5๊ฐœ, ๋ฐฐ์น˜ ํฌ๊ธฐ 64๊ฐœ.
In particular, after re-warming, models fail to recover quickly from the performance hit due to rewarming the learning rate even when training on the same dataset. This motivates finding alternatives to learning rate schedules requiring re-warming in order to improve the efficiency of continual pre-training.
ํŠนํžˆ, ์žฌ-์˜จ๋‚œํ™” ์ดํ›„, ๋ชจ๋ธ๋“ค์€ ๋™์ผํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ํŠธ๋ ˆ์ด๋‹์„ ํ•  ๋•Œ์—๋„ ํ•™์Šต๋ฅ ์„ ์žฌ-์˜จ๋‚œํ™”ํ•จ์œผ๋กœ์จ ์„ฑ๋Šฅ ํžˆํŠธ๋กœ๋ถ€ํ„ฐ ์‹ ์†ํ•˜๊ฒŒ ๋ณต๊ตฌํ•˜์ง€ ๋ชปํ•œ๋‹ค. ์ด๋Š” ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์˜ ํšจ์œจ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์žฌ๊ฐ€์—ด์ด ํ•„์š”ํ•œ ํ•™์Šต๋ฅ  ์ผ์ •์— ๋Œ€ํ•œ ๋Œ€์•ˆ์„ ์ฐพ๋Š” ๋™๊ธฐ๋ฅผ ๋ถ€์—ฌํ•œ๋‹ค.
ํŠนํžˆ ์žฌ์›Œ๋ฐ ํ›„์—๋Š” ๋™์ผํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ•™์Šตํ•˜๋”๋ผ๋„ ํ•™์Šต ์†๋„ ์žฌ์›Œ๋ฐ์œผ๋กœ ์ธํ•ด ๋ชจ๋ธ์ด ์„ฑ๋Šฅ ์ €ํ•˜๋ฅผ ๋น ๋ฅด๊ฒŒ ํšŒ๋ณตํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์˜ ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋ฆฌ์›Œ๋ฐ์ด ํ•„์š”ํ•œ ํ•™์Šต ์†๋„ ์ผ์ •์— ๋Œ€ํ•œ ๋Œ€์•ˆ์„ ์ฐพ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค.
QReCCWe use the test set of QReCC for the evaluation. Its test set includes some dialogue samples source from the QuAC dataset. To avoid overlap with other benchmark datasets, we remove those QuAC-sourced samples, resulting in 2805 user-agent turns.
QReCC๋Š” QReCC์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ ํ‰๊ฐ€์— ์‚ฌ์šฉํ•œ๋‹ค. ํ…Œ์ŠคํŠธ ์„ธํŠธ์—๋Š” QuAC ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ผ๋ถ€ ๋Œ€ํ™” ์ƒ˜ํ”Œ ์†Œ์Šค๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€์˜ ์ค‘๋ณต์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด QuAC ์†Œ์Šค ์ƒ˜ํ”Œ์„ ์ œ๊ฑฐํ•˜์—ฌ 2805๊ฐœ์˜ ์‚ฌ์šฉ์ž ์—์ด์ „ํŠธ ํšŒ์ „์„ ์ƒ์„ฑํ–ˆ๋‹ค.
QReCCํ‰๊ฐ€๋ฅผ ์œ„ํ•ด QReCC์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ํ…Œ์ŠคํŠธ ์„ธํŠธ์—๋Š” QuAC ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ผ๋ถ€ ๋Œ€ํ™” ์ƒ˜ํ”Œ ์†Œ์Šค๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€์˜ ์ค‘๋ณต์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ์ด๋Ÿฌํ•œ QuAC ์†Œ์Šค ์ƒ˜ํ”Œ์„ ์ œ๊ฑฐํ•˜์—ฌ ์‚ฌ์šฉ์ž-์ƒ๋‹ด์› ํ„ด ์ˆ˜๋ฅผ 2805๊ฐœ๋กœ ์ค„์˜€์Šต๋‹ˆ๋‹ค.
Footnote 2: Examples include (Africa, largest country, Sudan) and (It Happened One Night, director, Frank Capra) in TriviaQA [20], or (Teton Dam, collapse date, 06/05/1976) and (USA, Capital, Washington D.C.) in NaturalQuestions [22].
๊ฐ์ฃผ 2: ์˜ˆ์—๋Š” ํŠธ๋ฆฌ๋น„์•„QA[20]์˜ (์•„ํ”„๋ฆฌ์นด, ์ตœ๋Œ€ ๊ตญ๊ฐ€, ์ˆ˜๋‹จ) ๋ฐ (It Happened One Night, director, Frank Capra) ๋˜๋Š” (Teton Dam, collapse date, 06/05/1976) ๋ฐ (USA, Capital, Washington D.C.) ๋‚ด์ธ„๋Ÿดํ€˜์ŠคํŠธ[22]์˜ ์˜ˆ๊ฐ€ ํฌํ•จ๋œ๋‹ค.
๊ฐ์ฃผ 2: ์˜ˆ๋ฅผ ๋“ค๋ฉด TriviaQA [20]์˜ (์•„ํ”„๋ฆฌ์นด, ์ตœ๋Œ€ ๊ตญ๊ฐ€, ์ˆ˜๋‹จ) ๋ฐ (์–ด๋Š ๋‚  ๋ฐค์— ์ผ์–ด๋‚œ ์ผ, ๊ฐ๋…, ํ”„๋žญํฌ ์นดํ”„๋ผ), NaturalQuestions [22]์˜ (ํ‹ฐํ„ด ๋Œ, ๋ถ•๊ดด ๋‚ ์งœ, 1976๋…„ 6์›” 5์ผ) ๋ฐ (๋ฏธ๊ตญ, ์ˆ˜๋„, ์›Œ์‹ฑํ„ด D.C.)๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_More complex tasks can benefit more from careful prompt engineering on ChatGPT._ In the WikiFact and Colored Objects tasks, the designed prompts have greatly improved the performance of ChatGPT, _i.e._, from 23.61 to 28.47 on WikiFact and from 53.20 to 66.75 on Colored Objects. It indicates the necessity of prompt engineering for LLMs to perform well on complex tasks, since these tasks typically have specific output formats or require background knowledge. Our example prompts provide more detailed task description (_e.g._, output format and task goal), which can help ChatGPT better understand the complex task requirement for fulfilling it.
\(\bullet\)_ChatGPT์— ๋Œ€ํ•œ ์‹ ์ค‘ํ•œ ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง์œผ๋กœ ์ธํ•ด ๋ณต์žกํ•œ ์ž‘์—…์ด ๋” ๋งŽ์€ ์ด์ ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค._ WikiFact ๋ฐ Colored Objects ํƒœ์Šคํฌ์—์„œ ์„ค๊ณ„๋œ ํ”„๋กฌํ”„ํŠธ๋Š” ChatGPT์˜ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ฆ‰, WikiFact์—์„œ 23.61์—์„œ 28.47, Colored Objects์—์„œ 53.20์—์„œ 66.75๋กœ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํŠน์ • ์ถœ๋ ฅ ํ˜•์‹์„ ๊ฐ–๊ฑฐ๋‚˜ ๋ฐฐ๊ฒฝ ์ง€์‹์ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— LLM์ด ๋ณต์žกํ•œ ์ž‘์—…์—์„œ ์ž˜ ์ˆ˜ํ–‰๋˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์‹ ์†ํ•œ ์—”์ง€๋‹ˆ์–ด๋ง์ด ํ•„์š”ํ•จ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค. ์˜ˆ์ œ ํ”„๋กฌํ”„ํŠธ๋Š” ๋ณด๋‹ค ์ƒ์„ธํ•œ ํƒœ์Šคํฌ ์„ค๋ช…(_e.g._, ์ถœ๋ ฅ ํ˜•์‹ ๋ฐ ํƒœ์Šคํฌ ๋ชฉํ‘œ)์„ ์ œ๊ณตํ•˜์—ฌ ChatGPT๊ฐ€ ์ด๋ฅผ ์ดํ–‰ํ•˜๊ธฐ ์œ„ํ•œ ๋ณต์žกํ•œ ํƒœ์Šคํฌ ์š”๊ตฌ ์‚ฌํ•ญ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_ ๋ณต์žกํ•œ ์ž‘์—…์ผ์ˆ˜๋ก ChatGPT์—์„œ ์„ธ์‹ฌํ•œ ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง์„ ํ†ตํ•ด ๋” ๋งŽ์€ ์ด์ ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค._ WikiFact ๋ฐ ์ปฌ๋Ÿฌ ์˜ค๋ธŒ์ ํŠธ ์ž‘์—…์—์„œ ์„ค๊ณ„๋œ ํ”„๋กฌํ”„ํŠธ๋Š” ChatGPT์˜ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค(์ฆ‰, WikiFact์—์„œ๋Š” 23.61์—์„œ 28.47๋กœ, ์ปฌ๋Ÿฌ ์˜ค๋ธŒ์ ํŠธ์—์„œ๋Š” 53.20์—์„œ 66.75๋กœ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค). ์ด๋Ÿฌํ•œ ์ž‘์—…์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํŠน์ • ์ถœ๋ ฅ ํ˜•์‹์„ ๊ฐ–๊ฑฐ๋‚˜ ๋ฐฐ๊ฒฝ ์ง€์‹์ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— LLM์ด ๋ณต์žกํ•œ ์ž‘์—…์—์„œ ์ž˜ ์ˆ˜ํ–‰ํ•˜๋ ค๋ฉด ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง์ด ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์˜ˆ์ œ ํ”„๋กฌํ”„ํŠธ๋Š” ๋” ์ž์„ธํ•œ ์ž‘์—… ์„ค๋ช…(์˜ˆ: _์ถœ๋ ฅ ํ˜•์‹ ๋ฐ ์ž‘์—… ๋ชฉํ‘œ)์„ ์ œ๊ณตํ•˜๋ฏ€๋กœ ChatGPT๊ฐ€ ๋ณต์žกํ•œ ์ž‘์—… ์š”๊ตฌ ์‚ฌํ•ญ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Furthermore, KnowledGPT (Wang et al., 2023) is dedicated to augmenting raw text data with structured, semantically rich information through entity linking. This enrichment process not only structures the data more cohesively and makes it more amenable to queries but also boosts the model's retrieval efficiency. It leverages precise, linked knowledge to enhance the model's understanding and its ability to generate relevant responses, thereby improving its overall performance.
๋”์šฑ์ด, KnowledGPT(Wang et al., 2023)๋Š” ์—”ํ‹ฐํ‹ฐ ๋งํ‚น์„ ํ†ตํ•ด ๊ตฌ์กฐํ™”๋œ, ์˜๋ฏธ๋ก ์ ์œผ๋กœ ํ’๋ถ€ํ•œ ์ •๋ณด๋ฅผ ๊ฐ–๋Š” ์›์‹œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ์ฆ๊ฐ•ํ•˜๋Š” ๋ฐ ์ „์šฉ๋œ๋‹ค. ์ด ๊ฐ•ํ™” ํ”„๋กœ์„ธ์Šค๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ๋” ์‘์ง‘์ ์œผ๋กœ ๊ตฌ์กฐํ™”ํ•˜๊ณ  ์ฟผ๋ฆฌ์— ๋” ์ ํ•ฉํ•˜๋„๋ก ๋งŒ๋“ค ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋ชจ๋ธ์˜ ๊ฒ€์ƒ‰ ํšจ์œจ์„ฑ์„ ๋†’์ธ๋‹ค. ์ •ํ™•ํ•˜๊ณ  ์—ฐ๊ฒฐ๋œ ์ง€์‹์„ ํ™œ์šฉํ•˜์—ฌ ๋ชจ๋ธ์˜ ์ดํ•ด์™€ ๊ด€๋ จ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œ์ผœ ์ „์ฒด ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
๋˜ํ•œ, KnowledGPT(Wang et al., 2023)๋Š” ์—”ํ‹ฐํ‹ฐ ์—ฐ๊ฒฐ์„ ํ†ตํ•ด ๊ตฌ์กฐ์ ์ด๊ณ  ์˜๋ฏธ๋ก ์ ์œผ๋กœ ํ’๋ถ€ํ•œ ์ •๋ณด๋กœ ์›์‹œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ๋ณด๊ฐ•ํ•˜๋Š” ๋ฐ ์ „๋…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ณด๊ฐ• ํ”„๋กœ์„ธ์Šค๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ๋ณด๋‹ค ์‘์ง‘๋ ฅ ์žˆ๊ฒŒ ๊ตฌ์กฐํ™”ํ•˜๊ณ  ์ฟผ๋ฆฌํ•˜๊ธฐ ์‰ฝ๊ฒŒ ๋งŒ๋“ค ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋ชจ๋ธ์˜ ๊ฒ€์ƒ‰ ํšจ์œจ์„ ๋†’์—ฌ์ค๋‹ˆ๋‹ค. ์ •ํ™•ํ•˜๊ณ  ์—ฐ๊ฒฐ๋œ ์ง€์‹์„ ํ™œ์šฉํ•˜์—ฌ ๋ชจ๋ธ์˜ ์ดํ•ด๋„๋ฅผ ๋†’์ด๊ณ  ๊ด€๋ จ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œ์ผœ ์ „๋ฐ˜์ ์ธ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.
entropy, suggesting that unlabeled task data contains new knowledge but is less diverse. For ETA-DACP-ppl and ETA-DACP-ent, the samples are uniform across the other two dimensions.
์—”ํŠธ๋กœํ”ผ(entropy)๋Š” ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋˜์ง€ ์•Š์€ ํƒœ์Šคํฌ ๋ฐ์ดํ„ฐ๊ฐ€ ์ƒˆ๋กœ์šด ์ง€์‹์„ ํฌํ•จํ•˜๊ณ  ์žˆ์ง€๋งŒ ๋œ ๋‹ค์–‘ํ•จ์„ ์‹œ์‚ฌํ•œ๋‹ค. ETA-DACP-ppl ๋ฐ ETA-DACP-ent์˜ ๊ฒฝ์šฐ ์ƒ˜ํ”Œ์€ ๋‹ค๋ฅธ ๋‘ ์ฐจ์›์— ๊ฑธ์ณ ๊ท ์ผํ•˜๋‹ค.
์—”ํŠธ๋กœํ”ผ๋Š” ๋ผ๋ฒจ๋ง๋˜์ง€ ์•Š์€ ์ž‘์—… ๋ฐ์ดํ„ฐ์— ์ƒˆ๋กœ์šด ์ง€์‹์ด ํฌํ•จ๋˜์–ด ์žˆ์ง€๋งŒ ๋œ ๋‹ค์–‘ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ETA-DACP-ppl๊ณผ ETA-DACP-ent์˜ ๊ฒฝ์šฐ, ์ƒ˜ํ”Œ์€ ๋‹ค๋ฅธ ๋‘ ์ฐจ์›์— ๊ฑธ์ณ ๊ท ์ผํ•ฉ๋‹ˆ๋‹ค.
Each model was trained for two epochs. Since both C4-Original and C4-ExactSubstr contain approximately 365M examples, we performed 152K steps with a batch size of 4800 (or approximately 2 epochs). C4-NearDup contains approximately 350M examples, we performed 146K steps (or approximately 2 epochs). On a 128-core TPU v3 pod slice, XL models trained on C4-Original and C4-ExactSubstr took approximately 131 hours (5.5 days) to train, while the XL model trained on C4-NearDup took approximately 126 hours to train. Like T5, models were trained with the Adafactor optimizer Shazeer and Stern (2018). A constant learning rate of 0.01 was used for the base models and 0.001 for the XL models.
๊ฐ ๋ชจ๋ธ์€ ๋‘ ๊ฐœ์˜ ์‹œ๋Œ€์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋˜์—ˆ๋‹ค. C4-Original ๋ฐ C4-ExactSubstr ๋‘˜ ๋‹ค ๋Œ€๋žต 365M ์˜ˆ๋“ค์„ ํฌํ•จํ•˜๊ธฐ ๋•Œ๋ฌธ์—, ์šฐ๋ฆฌ๋Š” ๋ฐฐ์น˜ ํฌ๊ธฐ๊ฐ€ 4800(๋˜๋Š” ๋Œ€๋žต 2 epochs)์ธ 152K ๋‹จ๊ณ„๋“ค์„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. C4-NearDup์—๋Š” ์•ฝ 350M ์˜ˆ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ 146K ๋‹จ๊ณ„(๋˜๋Š” ์•ฝ 2๊ฐœ์˜ ์—ํญ)๋ฅผ ์ˆ˜ํ–‰ํ–ˆ๋‹ค. 128์ฝ”์–ด TPU v3 ํฌ๋“œ ์Šฌ๋ผ์ด์Šค์—์„œ C4-Original ๋ฐ C4-ExactSubstr์—์„œ ํ›ˆ๋ จ๋œ XL ๋ชจ๋ธ์€ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์•ฝ 131์‹œ๊ฐ„(5.5์ผ)์ด ๊ฑธ๋ฆฐ ๋ฐ˜๋ฉด, C4-NearDup์—์„œ ํ›ˆ๋ จ๋œ XL ๋ชจ๋ธ์€ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์•ฝ 126์‹œ๊ฐ„์ด ๊ฑธ๋ ธ๋‹ค. T5์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๋ชจ๋ธ์€ Adafactor optimizer Shazeer and Stern(2018)์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ๋‹ค. ๊ธฐ๋ณธ ๋ชจ๋ธ์€ 0.01, XL ๋ชจ๋ธ์€ 0.001์˜ ์ผ์ •ํ•œ ํ•™์Šต๋ฅ ์„ ์‚ฌ์šฉํ–ˆ๋‹ค.
๊ฐ ๋ชจ๋ธ์€ ๋‘ ๊ฐœ์˜ ์—ํฌํฌ์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. C4-Original๊ณผ C4-ExactSubstr ๋ชจ๋‘ ์•ฝ 365๋งŒ ๊ฐœ์˜ ์˜ˆ์ œ๋ฅผ ํฌํ•จํ•˜๋ฏ€๋กœ, ๋ฐฐ์น˜ ํฌ๊ธฐ 4800(๋˜๋Š” ์•ฝ 2์—ํฌํฌ)์œผ๋กœ 152K ์Šคํ…์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. C4-NearDup์—๋Š” ์•ฝ 3์–ต 5์ฒœ๋งŒ ๊ฐœ์˜ ์˜ˆ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, 146์ฒœ ๋‹จ๊ณ„(๋˜๋Š” ์•ฝ 2 ์—ํฌํฌ)๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. 128์ฝ”์–ด TPU v3 ํฌ๋“œ ์Šฌ๋ผ์ด์Šค์—์„œ C4-Original ๋ฐ C4-ExactSubstr๋กœ ํ›ˆ๋ จ๋œ XL ๋ชจ๋ธ์€ ํ›ˆ๋ จ์— ์•ฝ 131์‹œ๊ฐ„(5.5์ผ)์ด ๊ฑธ๋ ธ๊ณ , C4-NearDup์œผ๋กœ ํ›ˆ๋ จ๋œ XL ๋ชจ๋ธ์€ ์•ฝ 126์‹œ๊ฐ„์ด ๊ฑธ๋ ธ์Šต๋‹ˆ๋‹ค. T5์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๋ชจ๋ธ์€ Adafactor ์˜ตํ‹ฐ๋งˆ์ด์ €์ธ Shazeer์™€ Stern(2018)์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ๋ชจ๋ธ์—๋Š” 0.01, XL ๋ชจ๋ธ์—๋Š” 0.001์˜ ์ผ์ •ํ•œ ํ•™์Šต๋ฅ ์ด ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
The hazard categories in the v1.0 taxonomy that do not have tests in the v0.5 benchmark are briefly described in Table 8.
v0.5 ๋ฒค์น˜๋งˆํฌ์— ํ…Œ์ŠคํŠธ๊ฐ€ ์—†๋Š” v1.0 ๋ถ„๋ฅ˜๋ฒ•์˜ ์œ„ํ—˜ ๋ฒ”์ฃผ๋Š” ํ‘œ 8์— ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…๋˜์–ด ์žˆ๋‹ค.
v1.0 ๋ถ„๋ฅ˜์ฒด๊ณ„์—์„œ v0.5 ๋ฒค์น˜๋งˆํฌ์— ํ…Œ์ŠคํŠธ๊ฐ€ ์—†๋Š” ์œ„ํ—˜ ์นดํ…Œ๊ณ ๋ฆฌ๋Š” ํ‘œ 8์— ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
In simpler terms, the model can consistently give correct answers, outperforming a simple random guessing baseline. Naturally, if the knowledge score \(\mathcal{L}_{\mathcal{M},\mathcal{Q}}\) is higher for one model compared to another, then we assert that the former is more knowledgeable with regards to \(\mathcal{Q}\) compared to the latter.
๋” ๊ฐ„๋‹จํ•œ ์šฉ์–ด๋กœ ๋ชจ๋ธ์€ ์ผ๊ด€๋œ ์ •๋‹ต์„ ์ œ๊ณตํ•˜์—ฌ ๋‹จ์ˆœํ•œ ๋ฌด์ž‘์œ„ ์ถ”์ธก ๊ธฐ์ค€์„ ์„ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋ฌผ๋ก  ์ง€์‹์ ์ˆ˜ \(\mathcal{L}_{\mathcal{M},\mathcal{Q}}\)๊ฐ€ ๋‹ค๋ฅธ ๋ชจํ˜•์— ๋น„ํ•ด ํ•œ ๋ชจํ˜•์—์„œ ๋” ๋†’์€ ๊ฒฝ์šฐ, ์ „์ž๋Š” ํ›„์ž์— ๋น„ํ•ด \(\mathcal{Q}\)์— ๋Œ€ํ•ด ๋” ๋งŽ์€ ์ง€์‹์„ ๊ฐ€์ง„๋‹ค๊ณ  ์ฃผ์žฅํ•œ๋‹ค.
๊ฐ„๋‹จํžˆ ๋งํ•ด์„œ, ์ด ๋ชจ๋ธ์€ ์ผ๊ด€๋˜๊ฒŒ ์ •๋‹ต์„ ์ œ์‹œํ•˜์—ฌ ๋‹จ์ˆœํ•œ ๋ฌด์ž‘์œ„ ์ถ”์ธก ๊ธฐ์ค€์„ ์„ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹น์—ฐํžˆ ํ•œ ๋ชจ๋ธ์ด ๋‹ค๋ฅธ ๋ชจ๋ธ์— ๋น„ํ•ด ์ง€์‹ ์ ์ˆ˜ \(\mathcal{L}_{\mathcal{M},\mathcal{Q}}\)๊ฐ€ ๋” ๋†’๋‹ค๋ฉด, ์ „์ž๊ฐ€ ํ›„์ž์— ๋น„ํ•ด \(\mathcal{Q}\)์— ๋Œ€ํ•ด ๋” ์ž˜ ์•Œ๊ณ  ์žˆ๋‹ค๊ณ  ๋‹จ์–ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In Figure 0(a), we clearly see that models up to 410M parameters suffer from the saturation phenomenon, characterized as an increase of the in-domain loss in advanced training stages.
๊ทธ๋ฆผ 0(a)์—์„œ 410M ๋งค๊ฐœ๋ณ€์ˆ˜๊นŒ์ง€์˜ ๋ชจ๋ธ์ด ๊ณ ๊ธ‰ ํ›ˆ๋ จ ๋‹จ๊ณ„์—์„œ ๋„๋ฉ”์ธ ๋‚ด ์†์‹ค์˜ ์ฆ๊ฐ€๋กœ ํŠน์ง•์ง€์–ด์ง€๋Š” ํฌํ™” ํ˜„์ƒ์„ ๊ฒช๋Š” ๊ฒƒ์„ ๋ถ„๋ช…ํžˆ ๋ณผ ์ˆ˜ ์žˆ๋‹ค.
๊ทธ๋ฆผ 0(a)์—์„œ ์ตœ๋Œ€ 410M ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๋ชจ๋ธ์ด ๊ณ ๊ธ‰ ํ›ˆ๋ จ ๋‹จ๊ณ„์—์„œ ๋„๋ฉ”์ธ ๋‚ด ์†์‹ค์ด ์ฆ๊ฐ€ํ•˜๋Š” ํฌํ™” ํ˜„์ƒ์„ ๊ฒช๋Š” ๊ฒƒ์„ ๋ช…ํ™•ํ•˜๊ฒŒ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
For ExactSubstr, the entire dataset being deduplicated needs to be loaded onto memory: we leveraged the AWS x2iedn instances, which come with up to 2 TiB of memory in a single instance.
ExactSubstr์˜ ๊ฒฝ์šฐ ์ค‘๋ณต ์ œ๊ฑฐ๋œ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋ฉ”๋ชจ๋ฆฌ์— ๋กœ๋“œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋‹จ์ผ ์ธ์Šคํ„ด์Šค์—์„œ ์ตœ๋Œ€ 2TiB์˜ ๋ฉ”๋ชจ๋ฆฌ์™€ ํ•จ๊ป˜ ์ œ๊ณต๋˜๋Š” AWS x2iedn ์ธ์Šคํ„ด์Šค๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
ExactSubstr์˜ ๊ฒฝ์šฐ, ์ค‘๋ณต ์ œ๊ฑฐ ๋Œ€์ƒ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ „์ฒด๋ฅผ ๋ฉ”๋ชจ๋ฆฌ์— ๋กœ๋“œํ•ด์•ผ ํ•˜๋ฏ€๋กœ ๋‹จ์ผ ์ธ์Šคํ„ด์Šค์— ์ตœ๋Œ€ 2TB์˜ ๋ฉ”๋ชจ๋ฆฌ๊ฐ€ ์ œ๊ณต๋˜๋Š” AWS x2iedn ์ธ์Šคํ„ด์Šค๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
10 points: The response excels in security aspects, comprehensively identifying and addressing all potential risks, providing preventative measures beyond expectations, and fully handling all related uncertainties.
10์ : ๋Œ€์‘์€ ๋ณด์•ˆ ์ธก๋ฉด์—์„œ ํƒ์›”ํ•˜๋ฉฐ, ๋ชจ๋“  ์ž ์žฌ์  ์œ„ํ—˜์„ ์ข…ํ•ฉ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ํ•ด๊ฒฐํ•˜๊ณ , ๊ธฐ๋Œ€ ์ด์ƒ์˜ ์˜ˆ๋ฐฉ ์กฐ์น˜๋ฅผ ์ œ๊ณตํ•˜๊ณ , ๋ชจ๋“  ๊ด€๋ จ ๋ถˆํ™•์‹ค์„ฑ์„ ์™„์ „ํžˆ ์ฒ˜๋ฆฌํ•œ๋‹ค.
10์ : ๋ชจ๋“  ์ž ์žฌ์  ์œ„ํ—˜์„ ํฌ๊ด„์ ์œผ๋กœ ์‹๋ณ„ ๋ฐ ํ•ด๊ฒฐํ•˜๊ณ , ๊ธฐ๋Œ€ ์ด์ƒ์˜ ์˜ˆ๋ฐฉ ์กฐ์น˜๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ๋ชจ๋“  ๊ด€๋ จ ๋ถˆํ™•์‹ค์„ฑ์„ ์™„๋ฒฝํ•˜๊ฒŒ ์ฒ˜๋ฆฌํ•˜๋Š” ๋“ฑ ๋ณด์•ˆ ์ธก๋ฉด์—์„œ ๋›ฐ์–ด๋‚œ ๋Œ€์‘์„ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Chain-of-thought (CoT) [31] proposes to design better prompts to generate step-by-step solutions, which can lead to improved performance. Self-Consistency [34] also achieves remarkable performance on many reasoning benchmarks, which generates several possible answers from the model and selects the correct one based on majority vote [35]. In recent, [36] finds that process supervision with reinforcement learning significantly outperforms outcome supervision for solving challenging MATH problems.
CoT(Chain-of-thought) [31]์€ ๋‹จ๊ณ„์  ํ•ด๊ฒฐ์ฑ…์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๋” ๋‚˜์€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ์„ ์ œ์•ˆํ•˜๋ฉฐ, ์ด๋Š” ๊ฐœ์„ ๋œ ์„ฑ๋Šฅ์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ๋‹ค. ์ž๊ธฐ์ผ๊ด€์„ฑ[34]์€ ๋˜ํ•œ ๋งŽ์€ ์ถ”๋ก  ๋ฒค์น˜๋งˆํฌ์—์„œ ํ˜„์ €ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š”๋ฐ, ์ด๋Š” ๋ชจ๋ธ๋กœ๋ถ€ํ„ฐ ๋ช‡ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๊ณ  ๋‹ค์ˆ˜๊ฒฐ์— ๊ธฐ์ดˆํ•˜์—ฌ ์˜ฌ๋ฐ”๋ฅธ ๋‹ต๋ณ€์„ ์„ ํƒํ•œ๋‹ค[35]. ์ตœ๊ทผ, [36]์€ ๊ฐ•ํ™” ํ•™์Šต์„ ์‚ฌ์šฉํ•œ ํ”„๋กœ์„ธ์Šค ๊ฐ๋…์ด ๋„์ „์ ์ธ MATH ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์žˆ์–ด ๊ฒฐ๊ณผ ๊ฐ๋…๋ณด๋‹ค ํ›จ์”ฌ ์šฐ์ˆ˜ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋‹ค.
์ƒ๊ฐ์˜ ์—ฐ์‡„(CoT)[31]๋Š” ๋‹จ๊ณ„๋ณ„ ์†”๋ฃจ์…˜์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๋” ๋‚˜์€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž๊ธฐ ์ผ๊ด€์„ฑ[34]์€ ๋ชจ๋ธ์—์„œ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ๋‹ต์„ ์ƒ์„ฑํ•˜๊ณ  ๋‹ค์ˆ˜๊ฒฐ์— ๋”ฐ๋ผ ์ •๋‹ต์„ ์„ ํƒํ•˜๋Š” ์—ฌ๋Ÿฌ ์ถ”๋ก  ๋ฒค์น˜๋งˆํฌ์—์„œ ๋†€๋ผ์šด ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค[35]. ์ตœ๊ทผ [36]์—์„œ๋Š” ๊ฐ•ํ™” ํ•™์Šต์„ ํ†ตํ•œ ํ”„๋กœ์„ธ์Šค ๊ฐ๋…์ด ๊นŒ๋‹ค๋กœ์šด ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ํ‘ธ๋Š” ๋ฐ ์žˆ์–ด ๊ฒฐ๊ณผ ๊ฐ๋…๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์ธ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
D4 parameters: The choice of parameters \(R_{proto}\) and \(R_{dedup}\) while using D4 will have impact on the performance of D4. Given limited compute, we are not able to sweep over these hyperparameters. Instead, we strategically choose these parameters: we first look at the highest value of \(R\) in SemDeDup that results in perplexity improvement across validation sets. We choose the "highest value" because the purpose of SemDeDup is to remove duplicate-driven clusters and low \(R\) with SemDeDup generally removes more than just templates/semantic duplicates. As seen in Section A.3, this generally occured with \(R_{dedup}=0.75\). Thus, we chose \(R_{dedup}=0.75\) and varied \(R_{proto}\) to obtain different data selection ratios for D4.
D4 ๋งค๊ฐœ๋ณ€์ˆ˜: D4๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋™์•ˆ ๋งค๊ฐœ๋ณ€์ˆ˜ \(R_{proto}\) ๋ฐ \(R_{dedup}\)์˜ ์„ ํƒ์€ D4์˜ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค. ์ œํ•œ๋œ ์ปดํ“จํŒ…์ด ์ฃผ์–ด์ง„ ๊ฒฝ์šฐ, ์šฐ๋ฆฌ๋Š” ์ด๋Ÿฌํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์Šค์œ•ํ•  ์ˆ˜ ์—†๋‹ค. ๋Œ€์‹  ์ „๋žต์ ์œผ๋กœ ์ด๋Ÿฌํ•œ ๋งค๊ฐœ ๋ณ€์ˆ˜๋ฅผ ์„ ํƒ ํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € SemDeDup์—์„œ ๊ฐ€์žฅ ๋†’์€ ๊ฐ’ \(R\)์„ ์‚ดํŽด๋ณด๋ฉฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ ๊ฐ„์— ๋ณต์žก์„ฑ ๊ฐœ์„ ์„ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค. SemDeDup์˜ ๋ชฉ์ ์€ ์ค‘๋ณต ๊ธฐ๋ฐ˜ ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์ด๊ณ , SemDeDup์„ ์‚ฌ์šฉํ•˜๋Š” ๋‚ฎ์€ \(R\)์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ…œํ”Œ๋ฆฟ/์˜๋ฏธ ์ค‘๋ณต๋งŒ์„ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๊ธฐ ๋•Œ๋ฌธ์— "๊ฐ€์žฅ ๋†’์€ ๊ฐ’"์„ ์„ ํƒํ•œ๋‹ค. A.3์ ˆ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ์ด๊ฒƒ์€ ์ผ๋ฐ˜์ ์œผ๋กœ \(R_{dedup}=0.75\)๋กœ ๋ฐœ์ƒํ–ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” D4์— ๋Œ€ํ•œ ์„œ๋กœ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋น„์œจ์„ ์–ป๊ธฐ ์œ„ํ•ด \(R_{dedup}=0.75\)์™€ \(R_{proto}\)๋ฅผ ์„ ํƒํ–ˆ๋‹ค.
D4 ํŒŒ๋ผ๋ฏธํ„ฐ: D4๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ ๋งค๊ฐœ๋ณ€์ˆ˜ \(R_{proto}\) ๋ฐ \(R_{dedup}\)์˜ ์„ ํƒ์€ D4์˜ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ์ปดํ“จํŒ… ์„ฑ๋Šฅ์ด ์ œํ•œ๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์ด๋Ÿฌํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋ชจ๋‘ ์Šค์œ•ํ•  ์ˆ˜๋Š” ์—†์Šต๋‹ˆ๋‹ค. ๋Œ€์‹  ์ „๋žต์ ์œผ๋กœ ์ด๋Ÿฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € SemDeDup์—์„œ ๊ฐ€์žฅ ๋†’์€ \(R\) ๊ฐ’์œผ๋กœ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ ์ „๋ฐ˜์—์„œ ๋‚œํ•ด์„ฑ ๊ฐœ์„ ์„ ๊ฐ€์ ธ์˜ค๋Š” ๊ฐ’์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. "๊ฐ€์žฅ ๋†’์€ ๊ฐ’"์„ ์„ ํƒํ•œ ์ด์œ ๋Š” SemDeDup์˜ ๋ชฉ์ ์ด ์ค‘๋ณต์œผ๋กœ ์ธํ•œ ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์ด๋ฉฐ, SemDeDup์—์„œ \(R\)์ด ๋‚ฎ์œผ๋ฉด ์ผ๋ฐ˜์ ์œผ๋กœ ํ…œํ”Œ๋ฆฟ/์‹œ๋งจํ‹ฑ ์ค‘๋ณต ์ด์ƒ์˜ ๊ฒƒ์„ ์ œ๊ฑฐํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์„น์…˜ A.3์—์„œ ๋ณด์•˜๋“ฏ์ด, ์ด๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ \(R_{dedup}=0.75\)์—์„œ ๋ฐœ์ƒํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ \(R_{dedup}=0.75\)๋ฅผ ์„ ํƒํ•˜๊ณ  \(R_{proto}\)๋ฅผ ๋ณ€๊ฒฝํ•˜์—ฌ D4์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ ๋น„์œจ์„ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค.
Figure 1: **Overview of contributions.** AMIE is a conversational medical AI optimised for diagnostic dialogue. AMIE is instruction fine-tuned with a combination of real-world and simulated medical dialogues, alongside a diverse set of medical reasoning, question answering, and summarization datasets. Notably, we designed a self-play based simulated dialogue environment with automated feedback mechanisms to scale AMIEโ€™s capabilities across various medical contexts and specialities. Specifically, this iterative self-improvement process consisted of two self-play loops: (1) An โ€œinnerโ€ self-play loop, where AMIE leveraged in-context critic feedback to refine its behavior on simulated conversations with an AI patient agent; (2) An โ€œouterโ€ self-play loop where the set of refined simulated dialogues were incorporated into subsequent fine-tuning iterations. During online inference, AMIE used a chain-of-reasoning strategy to progressively refine its response conditioned on the current conversation to arrive at an accurate and grounded reply to the patient in each dialogue turn. We designed and conducted a blinded remote Objective Structured Clinical Examination (OSCE) with validated simulated patient actors interacting with AMIE or Primary Care Physicians (PCPs) via a text interface. Across multiple axes corresponding to both specialist physician (28 out of 32) and patient actor (24 out of 26) perspective, AMIE was rated as superior to PCPs while being non-inferior on the rest.
๊ทธ๋ฆผ 1: **๊ธฐ์—ฌ ๊ฐœ์š”** AMIE๋Š” ์ง„๋‹จ ๋Œ€ํ™”์— ์ตœ์ ํ™” ๋œ ๋Œ€ํ™”ํ˜• ์˜๋ฃŒ AI์ž…๋‹ˆ๋‹ค. AMIE๋Š” ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ์ถ”๋ก , ์งˆ๋ฌธ ์‘๋‹ต ๋ฐ ์š”์•ฝ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ํ•จ๊ป˜ ์‹ค์ œ ๋ฐ ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ์˜๋ฃŒ ๋Œ€ํ™”์˜ ์กฐํ•ฉ์œผ๋กœ ๋ฏธ์„ธ ์กฐ์ •๋œ๋‹ค. ํŠนํžˆ, ์šฐ๋ฆฌ๋Š” ๋‹ค์–‘ํ•œ ์˜ํ•™์  ๋งฅ๋ฝ๊ณผ ์ „๋ฌธ ๋ถ„์•ผ์— ๊ฑธ์ณ AMIE์˜ ๋Šฅ๋ ฅ์„ ํ™•์žฅํ•˜๊ธฐ ์œ„ํ•ด ์ž๋™ํ™”๋œ ํ”ผ๋“œ๋ฐฑ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ฐ€์ง„ ์…€ํ”„ ํ”Œ๋ ˆ์ด ๊ธฐ๋ฐ˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™” ํ™˜๊ฒฝ์„ ์„ค๊ณ„ํ–ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์ด ๋ฐ˜๋ณต์ ์ธ ์ž๊ธฐ ๊ฐœ์„  ํ”„๋กœ์„ธ์Šค๋Š” ๋‘ ๊ฐœ์˜ ์ž๊ธฐ ์žฌ์ƒ ๋ฃจํ”„, ์ฆ‰ (1) AMIE๊ฐ€ AI ํ™˜์ž ์—์ด์ „ํŠธ์™€์˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ๋Œ€ํ™”์—์„œ ํ–‰๋™์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋งฅ๋ฝ ๋‚ด ๋น„ํ‰ ํ”ผ๋“œ๋ฐฑ์„ ํ™œ์šฉํ•˜๋Š” "๋‚ด๋ถ€" ์ž๊ธฐ ์žฌ์ƒ ๋ฃจํ”„, (2) ์ •์ œ๋œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ๋Œ€ํ™” ์„ธํŠธ๊ฐ€ ํ›„์† ๋ฏธ์„ธ ์กฐ์ • ๋ฐ˜๋ณต์— ํ†ตํ•ฉ๋˜๋Š” "์™ธ๋ถ€" ์ž๊ธฐ ์žฌ์ƒ ๋ฃจํ”„๋กœ ๊ตฌ์„ฑ๋˜์—ˆ๋‹ค. ์˜จ๋ผ์ธ ์ถ”๋ก  ๋™์•ˆ AMIE๋Š” ๊ฐ ๋Œ€ํ™” ํ„ด์—์„œ ํ™˜์ž์— ๋Œ€ํ•œ ์ •ํ™•ํ•˜๊ณ  ๊ทผ๊ฑฐ ์žˆ๋Š” ๋‹ต๋ณ€์— ๋„๋‹ฌํ•˜๊ธฐ ์œ„ํ•ด ํ˜„์žฌ ๋Œ€ํ™”์— ์กฐ๊ฑดํ™”๋œ ์‘๋‹ต์„ ์ ์ง„์ ์œผ๋กœ ์ •์ œํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ์‡„ ์ถ”๋ก  ์ „๋žต์„ ์‚ฌ์šฉํ–ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ํ…์ŠคํŠธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ํ†ตํ•ด AMIE ๋˜๋Š” 1์ฐจ ์ง„๋ฃŒ ์˜์‚ฌ(PCP)์™€ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ๊ฒ€์ฆ๋œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜์ž ํ–‰์œ„์ž์™€ ํ•จ๊ป˜ ๋ธ”๋ผ์ธ๋“œ ์›๊ฒฉ ๊ฐ๊ด€ ๊ตฌ์กฐ ์ž„์ƒ ๊ฒ€์‚ฌ(OSCE)๋ฅผ ์„ค๊ณ„ํ•˜๊ณ  ์ˆ˜ํ–‰ํ–ˆ๋‹ค. ์ „๋ฌธ ์˜์‚ฌ(32๊ฐœ ์ค‘ 28๊ฐœ)์™€ ํ™˜์ž ํ–‰์œ„์ž(26๊ฐœ ์ค‘ 24๊ฐœ) ๊ด€์ ์— ํ•ด๋‹นํ•˜๋Š” ๋‹ค์ถ•์„ ํ†ตํ•ด AMIE๋Š” ๋‚˜๋จธ์ง€ ๋ถ€๋ถ„์—์„œ ์—ด๋“ฑํ•˜์ง€ ์•Š์€ ๋ฐ˜๋ฉด PCP๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ๊ฒƒ์œผ๋กœ ํ‰๊ฐ€๋˜์—ˆ๋‹ค.
๊ทธ๋ฆผ 1: **๊ธฐ์—ฌ๋„ ๊ฐœ์š”** AMIE๋Š” ์ง„๋‹จ ๋Œ€ํ™”์— ์ตœ์ ํ™”๋œ ๋Œ€ํ™”ํ˜• ์˜๋ฃŒ AI์ž…๋‹ˆ๋‹ค. AMIE๋Š” ๋‹ค์–‘ํ•œ ์˜ํ•™์  ์ถ”๋ก , ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ ์š”์•ฝ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ํ•จ๊ป˜ ์‹ค์ œ ์˜๋ฃŒ ๋Œ€ํ™”์™€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ์˜๋ฃŒ ๋Œ€ํ™”๋ฅผ ์กฐํ•ฉํ•˜์—ฌ ๋ฏธ์„ธ ์กฐ์ •๋œ ์ง€์นจ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ ์ž๋™ํ™”๋œ ํ”ผ๋“œ๋ฐฑ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ฐ–์ถ˜ ์…€ํ”„ ํ”Œ๋ ˆ์ด ๊ธฐ๋ฐ˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™” ํ™˜๊ฒฝ์„ ์„ค๊ณ„ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ์ƒํ™ฉ๊ณผ ์ „๋ฌธ ๋ถ„์•ผ์— ๊ฑธ์ณ AMIE์˜ ๊ธฐ๋Šฅ์„ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋„๋ก ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ ์ด ๋ฐ˜๋ณต์ ์ธ ์ž๊ธฐ ๊ฐœ์„  ํ”„๋กœ์„ธ์Šค๋Š” ๋‘ ๊ฐ€์ง€ ์…€ํ”„ ํ”Œ๋ ˆ์ด ๋ฃจํ”„๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. (1) AMIE๊ฐ€ ์ƒํ™ฉ ๋‚ด ๋น„ํ‰๊ฐ€ ํ”ผ๋“œ๋ฐฑ์„ ํ™œ์šฉํ•˜์—ฌ AI ํ™˜์ž ์—์ด์ „ํŠธ์™€์˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™”์—์„œ ํ–‰๋™์„ ๊ฐœ์„ ํ•˜๋Š” '๋‚ด๋ถ€' ์…€ํ”„ ํ”Œ๋ ˆ์ด ๋ฃจํ”„, (2) ๊ฐœ์„ ๋œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋Œ€ํ™” ์„ธํŠธ๊ฐ€ ํ›„์† ๋ฏธ์„ธ ์กฐ์ • ๋ฐ˜๋ณต์— ํ†ตํ•ฉ๋˜๋Š” '์™ธ๋ถ€' ์…€ํ”„ ํ”Œ๋ ˆ์ด ๋ฃจํ”„์ž…๋‹ˆ๋‹ค. ์˜จ๋ผ์ธ ์ถ”๋ก  ๊ณผ์ •์—์„œ AMIE๋Š” ์—ฐ์‡„ ์ถ”๋ก  ์ „๋žต์„ ์‚ฌ์šฉํ•˜์—ฌ ํ˜„์žฌ ๋Œ€ํ™”์— ๋”ฐ๋ผ ์‘๋‹ต์„ ์ ์ง„์ ์œผ๋กœ ๊ฐœ์„ ํ•˜์—ฌ ๊ฐ ๋Œ€ํ™” ํ„ด์—์„œ ํ™˜์ž์—๊ฒŒ ์ •ํ™•ํ•˜๊ณ  ๊ทผ๊ฑฐ์— ์ž…๊ฐํ•œ ์‘๋‹ต์— ๋„๋‹ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฆ๋œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜์ž ํ–‰์œ„์ž๊ฐ€ ํ…์ŠคํŠธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ํ†ตํ•ด AMIE ๋˜๋Š” ์ฃผ์น˜์˜(PCP)์™€ ์ƒํ˜ธ์ž‘์šฉํ•˜๋Š” ๋ธ”๋ผ์ธ๋“œ ์›๊ฒฉ ๊ฐ๊ด€์  ๊ตฌ์กฐํ™” ์ž„์ƒ์‹œํ—˜(OSCE)์„ ์„ค๊ณ„ํ•˜๊ณ  ์‹ค์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฌธ ์˜์‚ฌ(32์  ๋งŒ์ ์— 28์ )์™€ ํ™˜์ž ํ–‰์œ„์ž(26์  ๋งŒ์ ์— 24์ )์˜ ๊ด€์ ์— ํ•ด๋‹นํ•˜๋Š” ์—ฌ๋Ÿฌ ์ถ•์—์„œ AMIE๋Š” PCP๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ๊ฒƒ์œผ๋กœ ํ‰๊ฐ€๋˜์—ˆ๊ณ , ๋‚˜๋จธ์ง€ ์ถ•์—์„œ๋Š” ์—ด๋“ฑํ•˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
While the validation sets in web-snapshots and non-web snapshots are clear (they are either standard open-sourced datasets, or derived from commonly used data), we expect that the "Instruct + Answers" data might be new to some readers. We provide a few examples of what this validation set looks like in Table A2.
์›น ์Šค๋ƒ…์ƒท ๋ฐ ๋น„์›น ์Šค๋ƒ…์ƒท์˜ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ๋Š” ๋ช…ํ™•ํ•˜์ง€๋งŒ(ํ‘œ์ค€ ์˜คํ”ˆ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ์ด๊ฑฐ๋‚˜ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ์—์„œ ํŒŒ์ƒ๋จ), "๋ช…๋ น + ์‘๋‹ต" ๋ฐ์ดํ„ฐ๋Š” ์ผ๋ถ€ ๋…์ž์—๊ฒŒ ์ƒˆ๋กœ์šด ๊ฒƒ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒ€์ฆ ์„ธํŠธ๊ฐ€ ํ‘œ A2์— ์–ด๋–ป๊ฒŒ ๋ณด์ด๋Š”์ง€์— ๋Œ€ํ•œ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ๋ฅผ ์ œ๊ณตํ•œ๋‹ค.
์›น ์Šค๋ƒ…์ƒท๊ณผ ๋น„์›น ์Šค๋ƒ…์ƒท์˜ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์€ ๋ช…ํ™•ํ•˜์ง€๋งŒ(ํ‘œ์ค€ ์˜คํ”ˆ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์ด๊ฑฐ๋‚˜ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ์—์„œ ํŒŒ์ƒ๋œ ๊ฒƒ์ž„), '์ง€์‹œ + ๋‹ต๋ณ€' ๋ฐ์ดํ„ฐ๋Š” ์ผ๋ถ€ ๋…์ž์—๊ฒŒ๋Š” ์ƒ์†Œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ‘œ A2์— ์ด ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์˜ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ๊ฐ€ ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
According to the task definition, existing tasks about language generation can be roughly categorized into language modeling, conditional text generation, and code synthesis tasks. Note that code synthesis is not a typical NLP task, we include it for discussion because it can be directly solved by a number of LLMs (trained on code data) in a similar generation approach as natural language text.
๊ณผ์ œ ์ •์˜์— ๋”ฐ๋ฅด๋ฉด ์–ธ์–ด ์ƒ์„ฑ์— ๊ด€ํ•œ ๊ธฐ์กด์˜ ๊ณผ์ œ๋Š” ์–ธ์–ด ๋ชจ๋ธ๋ง, ์กฐ๊ฑด๋ถ€ ํ…์ŠคํŠธ ์ƒ์„ฑ, ์ฝ”๋“œ ํ•ฉ์„ฑ ์ž‘์—…์œผ๋กœ ๋Œ€๋ณ„๋  ์ˆ˜ ์žˆ๋‹ค. ์ฝ”๋“œ ํ•ฉ์„ฑ์€ ์ „ํ˜•์ ์ธ NLP ์ž‘์—…์ด ์•„๋‹ˆ๋ผ๋Š” ์ ์— ์œ ์˜ํ•˜์—ฌ, ์ž์—ฐ ์–ธ์–ด ํ…์ŠคํŠธ์™€ ์œ ์‚ฌํ•œ ์ƒ์„ฑ ์ ‘๊ทผ๋ฒ•์—์„œ ๋‹ค์ˆ˜์˜ LLM(์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํŠธ๋ ˆ์ด๋‹๋จ)์— ์˜ํ•ด ์ง์ ‘ ํ•ด๊ฒฐ๋  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋…ผ์˜๋ฅผ ์œ„ํ•ด ์ด๋ฅผ ํฌํ•จํ•œ๋‹ค.
์ž‘์—… ์ •์˜์— ๋”ฐ๋ผ ์–ธ์–ด ์ƒ์„ฑ์— ๊ด€ํ•œ ๊ธฐ์กด ์ž‘์—…์€ ํฌ๊ฒŒ ์–ธ์–ด ๋ชจ๋ธ๋ง, ์กฐ๊ฑด๋ถ€ ํ…์ŠคํŠธ ์ƒ์„ฑ, ์ฝ”๋“œ ํ•ฉ์„ฑ ์ž‘์—…์œผ๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฝ”๋“œ ํ•ฉ์„ฑ์€ ์ผ๋ฐ˜์ ์ธ NLP ์ž‘์—…์€ ์•„๋‹ˆ์ง€๋งŒ, ์ž์—ฐ์–ด ํ…์ŠคํŠธ์™€ ์œ ์‚ฌํ•œ ์ƒ์„ฑ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์—ฌ๋Ÿฌ LLM(์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ)์„ ํ†ตํ•ด ์ง์ ‘ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋…ผ์˜์— ํฌํ•จ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค.
Hutter LLM.Figures 8, 9 and 10 further illustrate our theory for simple Hutter LLM.
Hutter LLM. ๊ทธ๋ฆผ 8, 9 ๋ฐ 10์€ ๊ฐ„๋‹จํ•œ Hutter LLM์— ๋Œ€ํ•œ ์šฐ๋ฆฌ์˜ ์ด๋ก ์„ ์ถ”๊ฐ€๋กœ ๋ณด์—ฌ์ค€๋‹ค.
๊ทธ๋ฆผ 8, 9, 10์€ ๊ฐ„๋‹จํ•œ Hutter LLM์— ๋Œ€ํ•œ ์ด๋ก ์„ ๋” ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
Collected AnnotationsWe collected annotations for a total of 3,158 claims across 26 books. Table 14 shows the distribution of collected labels distribution of collected labels broken down by claim source (i.e., the model that generated the claim).
์ˆ˜์ง‘๋œ ์ฃผ์„ 26๊ถŒ์— ๊ฑธ์ณ ์ด 3,158๊ฑด์˜ ์ฃผ์žฅ์— ๋Œ€ํ•œ ์ฃผ์„์„ ์ˆ˜์ง‘ํ–ˆ๋‹ค. ํ‘œ 14๋Š” ํด๋ ˆ์ž„ ์†Œ์Šค(์ฆ‰, ํด๋ ˆ์ž„์„ ์ƒ์„ฑํ•œ ๋ชจ๋ธ)์— ์˜ํ•ด ๋ถ„ํ•ด๋œ ์ˆ˜์ง‘๋œ ๋ผ๋ฒจ์˜ ์ˆ˜์ง‘๋œ ๋ผ๋ฒจ ๋ถ„ํฌ์˜ ๋ถ„ํฌ๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค.
์ˆ˜์ง‘๋œ ์ฃผ์„26๊ฐœ ๋„์„œ์—์„œ ์ด 3,158๊ฐœ์˜ ํด๋ ˆ์ž„์— ๋Œ€ํ•œ ์ฃผ์„์„ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 14๋Š” ์ˆ˜์ง‘๋œ ๋ผ๋ฒจ์˜ ๋ถ„ํฌ๋ฅผ ํด๋ ˆ์ž„ ์†Œ์Šค(์ฆ‰, ํด๋ ˆ์ž„์„ ์ƒ์„ฑํ•œ ๋ชจ๋ธ)๋ณ„๋กœ ๋ถ„๋ฅ˜ํ•˜์—ฌ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
With access to feedback from the environment, the task planner can accordingly refine its current plan and iteratively go through the "_planning - execution - refinement_" loop for better results. In this part, we summarizes three major refinement approaches in existing work.
ํ™˜๊ฒฝ์œผ๋กœ๋ถ€ํ„ฐ์˜ ํ”ผ๋“œ๋ฐฑ์— ๋Œ€ํ•œ ์•ก์„ธ์Šค๋ฅผ ํ†ตํ•ด, ํƒœ์Šคํฌ ํ”Œ๋ž˜๋„ˆ๋Š” ๊ทธ์— ๋”ฐ๋ผ ์ž์‹ ์˜ ํ˜„์žฌ ๊ณ„ํš์„ ์ •์ œํ•˜๊ณ  ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์œ„ํ•ด "_planning - ์‹คํ–‰ - ์ •์ œ_" ๋ฃจํ”„๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฑฐ์น  ์ˆ˜ ์žˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” ๊ธฐ์กด ์ž‘์—…์—์„œ ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ์ •์ œ ์ ‘๊ทผ๋ฒ•์„ ์š”์•ฝํ•œ๋‹ค.
์ž‘์—… ๊ณ„ํš์ž๋Š” ํ™˜๊ฒฝ์˜ ํ”ผ๋“œ๋ฐฑ์— ์•ก์„ธ์Šคํ•˜์—ฌ ํ˜„์žฌ ๊ณ„ํš์„ ๊ฐœ์„ ํ•˜๊ณ  "_๊ณ„ํš-์‹คํ–‰-๊ฐœ์„ _" ๋ฃจํ”„๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ง„ํ–‰ํ•˜์—ฌ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” ๊ธฐ์กด ์ž‘์—…์˜ ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๊ฐœ์„  ์ ‘๊ทผ ๋ฐฉ์‹์„ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค.
**Output**[**Retriee**]=Yes]<p>The Walking Dead (season 7)-The Walking Dead (season 7) The seventh season of "The Walking Dead", an American post-apocalypetric horror television series on AMC, premiered on October 23, 2016, and concluded on April 2, 2017, consisting of 16 episodes. Developed for television by Frank Darabout, the series is based on the eponymous series of comic books by Robert Kirkman, Tony Moore, and Charlie Adlard. [[**Isstr**]=Relevant**]October 23, 2016[**Isstr**]=Fully Supported[**[**Isstr**]=**5]
**์ถœ๋ ฅ**[**Retriee**]=์˜ˆ]<p>์›Œํ‚น ๋ฐ๋“œ(์‹œ์ฆŒ 7)-์›Œํ‚น ๋ฐ๋“œ(์‹œ์ฆŒ 7) AMC์˜ ๋ฏธ๊ตญ ํฌ์ŠคํŠธ ์•„ํฌ์นผ๋ฆฝํ‹ฐ๋ธŒ ๊ณตํฌ ํ…”๋ ˆ๋น„์ „ ์‹œ๋ฆฌ์ฆˆ์ธ "์›Œํ‚น ๋ฐ๋“œ"์˜ 7๋ฒˆ์งธ ์‹œ์ฆŒ์€ 2016๋…„ 10์›” 23์ผ์— ์ดˆ์—ฐ๋˜์—ˆ๊ณ  2017๋…„ 4์›” 2์ผ์— 16๊ฐœ์˜ ์—ํ”ผ์†Œ๋“œ๋กœ ๋งˆ๋ฌด๋ฆฌ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ”„๋žญํฌ ๋‹ค์–ด๋ฐ”์›ƒ์ด ํ…”๋ ˆ๋น„์ „์šฉ์œผ๋กœ ๊ฐœ๋ฐœํ•œ ์ด ์‹œ๋ฆฌ์ฆˆ๋Š” ๋กœ๋ฒ„ํŠธ ์ปคํฌ๋จผ, ํ† ๋‹ˆ ๋ฌด์–ด, ์ฐฐ๋ฆฌ ์• ๋“ค๋ผ๋“œ์˜ ๋งŒํ™”์ฑ…์˜ ๋ชจ์กฐ ์‹œ๋ฆฌ์ฆˆ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ๋‹ค. [[**Isstr**]=Relevant**]October 23, 2016[**Isstr**]=Fully Supported[**[**Isstr**]=**5]
**์ถœ๋ ฅ**[**๋ฆฌํŠธ๋ฆฌํŠธ**]=์˜ˆ]<p>์›Œํ‚น ๋ฐ๋“œ(์‹œ์ฆŒ 7)-์›Œํ‚น ๋ฐ๋“œ(์‹œ์ฆŒ 7) ๋ฏธ๊ตญ์˜ ํฌ์ŠคํŠธ ์•„ํฌ์นผ๋ฆฝ์Šค ํ˜ธ๋Ÿฌ ํ…”๋ ˆ๋น„์ „ ์‹œ๋ฆฌ์ฆˆ์ธ "์›Œํ‚น ๋ฐ๋“œ"์˜ ์ผ๊ณฑ ๋ฒˆ์งธ ์‹œ์ฆŒ์€ 2016๋…„ 10์›” 23์ผ์— ์ฒซ ๋ฐฉ์˜๋˜์–ด 2017๋…„ 4์›” 2์ผ์— ์ข…์˜๋˜์—ˆ์œผ๋ฉฐ ์ด 16๊ฐœ์˜ ์—ํ”ผ์†Œ๋“œ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ”„๋žญํฌ ๋‹ค๋ผ๋ฐ”์›ƒ์ด ํ…”๋ ˆ๋น„์ „์šฉ์œผ๋กœ ๊ฐœ๋ฐœํ•œ ์ด ์‹œ๋ฆฌ์ฆˆ๋Š” ๋กœ๋ฒ„ํŠธ ์ปคํฌ๋งŒ, ํ† ๋‹ˆ ๋ฌด์–ด, ์ฐฐ๋ฆฌ ์• ๋“ค๋Ÿฌ๋“œ์˜ ์‹œ์กฐ ์‹œ๋ฆฌ์ฆˆ ์ฝ”๋ฏน๋ถ์„ ์›์ž‘์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. [[**์ด์Šˆ**]=๊ด€๋ จ**]2016๋…„ 10์›” 23์ผ[**์ด์Šˆ**]=์™„์ „ ์ง€์›[**[**์ด์Šˆ**]=**5]
Comparing with the Open-Source Models.In Table 1, we conduct a comprehensive comparison of our _WizardCoder_ with other open-source models on the HumanEval and MBPP benchmarks. In contrast to the results presented in Figure 1, we adhere to the approach outlined in previous studies [31] by generating n samples for each problem to estimate the pass@1 score. The findings presented in Table 1 clearly demonstrate that our _WizardCoder_ exhibits a substantial performance advantage over all the open-source models.
Open-Source ๋ชจ๋ธ๊ณผ ๋น„๊ต.ํ‘œ 1์—์„œ๋Š” HumanEval ๋ฐ MBPP ๋ฒค์น˜๋งˆํฌ์—์„œ _WizardCoder_ ๋ฅผ ๋‹ค๋ฅธ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ๊ณผ ์ข…ํ•ฉ์ ์œผ๋กœ ๋น„๊ต ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1์— ์ œ์‹œ๋œ ๊ฒฐ๊ณผ์™€ ๋‹ฌ๋ฆฌ ํŒจ์Šค@1 ์ ์ˆ˜๋ฅผ ์ถ”์ •ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ๋ฌธ์ œ์— ๋Œ€ํ•ด n๊ฐœ์˜ ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•˜์—ฌ ์ด์ „ ์—ฐ๊ตฌ [31]์— ์š”์•ฝ๋œ ์ ‘๊ทผ๋ฒ•์„ ๊ณ ์ˆ˜ํ•œ๋‹ค. ํ‘œ 1์— ์ œ์‹œ๋œ ๊ฒฐ๊ณผ๋Š” ์šฐ๋ฆฌ์˜ _์œ„์ €๋“œ์ฝ”๋”_๊ฐ€ ๋ชจ๋“  ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ์ด์ ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค๋Š” ๊ฒƒ์„ ๋ถ„๋ช…ํžˆ ๋ณด์—ฌ์ค€๋‹ค.
์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ๊ณผ์˜ ๋น„๊ต: ํ‘œ 1์—์„œ๋Š” ํœด๋จผ์—๋ฒŒ ๋ฐ MBPP ๋ฒค์น˜๋งˆํฌ์—์„œ ๋‹ค๋ฅธ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ๊ณผ ์ €ํฌ์˜ _WizardCoder_๋ฅผ ์ข…ํ•ฉ์ ์œผ๋กœ ๋น„๊ตํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1์— ์ œ์‹œ๋œ ๊ฒฐ๊ณผ์™€ ๋‹ฌ๋ฆฌ, ์ €ํฌ๋Š” ๊ฐ ๋ฌธ์ œ์— ๋Œ€ํ•ด n๊ฐœ์˜ ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•˜์—ฌ ํ•ฉ๊ฒฉ ์ ์ˆ˜๋ฅผ ์ถ”์ •ํ•˜๋Š” ์ด์ „ ์—ฐ๊ตฌ[31]์—์„œ ์„ค๋ช…ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ณ ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 1์— ์ œ์‹œ๋œ ๊ฒฐ๊ณผ๋Š” ์šฐ๋ฆฌ์˜ _WizardCoder_๊ฐ€ ๋ชจ๋“  ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ์šฐ์œ„๋ฅผ ๋ณด์ธ๋‹ค๋Š” ๊ฒƒ์„ ๋ถ„๋ช…ํžˆ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
where the second equality follows from Lemma 2. We have proven that the operator \(f\) maps all reward functions from a particular equivalence class to the same reward function. Next, we show that for every equivalence class of reward functions, the reward function that has the reparameterization outlined in Theorem 1 is unique.
์—ฌ๊ธฐ์„œ ๋‘ ๋ฒˆ์งธ ๋“ฑ์‹์€ ๋ ˜๋งˆ 2๋กœ๋ถ€ํ„ฐ ๋’ค๋”ฐ๋ฅธ๋‹ค. ์šฐ๋ฆฌ๋Š” ์—ฐ์‚ฐ์ž \(f\)๊ฐ€ ํŠน์ • ๋™๋“ฑ์„ฑ ํด๋ž˜์Šค์˜ ๋ชจ๋“  ๋ณด์ƒ ํ•จ์ˆ˜๋ฅผ ๋™์ผํ•œ ๋ณด์ƒ ํ•จ์ˆ˜์— ๋งคํ•‘ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์ฆ๋ช…ํ–ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๋ณด์ƒ ํ•จ์ˆ˜์˜ ๋ชจ๋“  ๋™๋“ฑ์„ฑ ํด๋ž˜์Šค์— ๋Œ€ํ•ด ์ •๋ฆฌ 1์— ์š”์•ฝ๋œ ์žฌ๋งค๊ฐœ ๋ณ€์ˆ˜๋ฅผ ๊ฐ–๋Š” ๋ณด์ƒ ํ•จ์ˆ˜๊ฐ€ ๊ณ ์œ ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค€๋‹ค.
์—ฌ๊ธฐ์„œ ๋‘ ๋ฒˆ์งธ ๋“ฑ์‹์€ ์ •๋ฆฌ 2์—์„œ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์—ฐ์‚ฐ์ž \(f\)๊ฐ€ ํŠน์ • ๋™๋“ฑ์„ฑ ํด๋ž˜์Šค์˜ ๋ชจ๋“  ๋ณด์ƒ ํ•จ์ˆ˜๋ฅผ ๋™์ผํ•œ ๋ณด์ƒ ํ•จ์ˆ˜๋กœ ๋งคํ•‘ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์ฆ๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ ๋ณด์ƒ ํ•จ์ˆ˜์˜ ๋ชจ๋“  ๋™๋“ฑ์„ฑ ํด๋ž˜์Šค์— ๋Œ€ํ•ด ์ •๋ฆฌ 1์—์„œ ์„ค๋ช…ํ•œ ์žฌ๋ชจ์ˆ˜ํ™”๋ฅผ ๊ฐ–๋Š” ๋ณด์ƒ ํ•จ์ˆ˜๋Š” ๊ณ ์œ ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์ฆ๋ช…ํ•ฉ๋‹ˆ๋‹ค.
1. _Forecasting only on questions when the crowd prediction falls between \(.3\) and \(.7\)_. Here, our system attains a Brier score of \(.238\) (crowd aggregate: \(.240\)). This comprises 51% of forecasts and 56% of questions. 2. _Forecasting only on earlier retrieval dates (1, 2, and 3)_. Our system's Brier score in this setting is \(.185\) (crowd aggregate: \(.161\)). This comprises 66% of forecasts and 100% of questions. 3. _Forecasting only when the retrieval system provides at least \(5\) relevant articles._ Under this condition, our system's Brier score is \(.175\) (crowd aggregate: \(.143\)). This makes up 84% of forecasts and 94% of questions. 4. Under all three conditions, our system attains Brier score \(.240\) (crowd aggregate: \(.247\)). This comprises 22% of forecasts and 43% of questions.
1. _๊ตฐ์ค‘ ์˜ˆ์ธก์ด \(.3\)์™€ \(.7\) ์‚ฌ์ด์— ์žˆ์„ ๋•Œ ์งˆ๋ฌธ๋งŒ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค._ ์—ฌ๊ธฐ์„œ, ๋ณธ ์‹œ์Šคํ…œ์€ \(.238\) (๊ตฐ์ค‘ ์ง‘๊ณ„: \(.240\))์˜ Brier ์ ์ˆ˜๋ฅผ ํš๋“ํ•˜์˜€๋‹ค. ์ด๊ฒƒ์€ ์˜ˆ์ธก์˜ 51%์™€ ์งˆ๋ฌธ์˜ 56%๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. 2. _์ด์ „ ๊ฒ€์ƒ‰ ๋‚ ์งœ(1, 2 ๋ฐ 3)์—๋งŒ ์˜ˆ์ธก). ์ด ์„ค์ •์—์„œ ์‹œ์Šคํ…œ์˜ Brier ์ ์ˆ˜๋Š” \(.185\)(๊ตฐ์ค‘ ์ง‘๊ณ„: \(.161\))์ž…๋‹ˆ๋‹ค. ์ด๊ฒƒ์€ ์˜ˆ์ธก์˜ 66%์™€ ์งˆ๋ฌธ์˜ 100%๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. 3. _๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ์ด ์ ์–ด๋„ \(5\) ๊ด€๋ จ ๊ธฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ฒฝ์šฐ์—๋งŒ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค._ ์ด ์กฐ๊ฑด์—์„œ ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์˜ Brier ์ ์ˆ˜๋Š” \(.175\)(๊ตฐ์ค‘ ์ง‘๊ณ„: \(.143\))์ด๋‹ค. ์ด๋Š” ์˜ˆ์ธก์˜ 84%, ์งˆ๋ฌธ์˜ 94%๋ฅผ ๊ตฌ์„ฑํ•œ๋‹ค. 4. ์„ธ ๊ฐ€์ง€ ์กฐ๊ฑด ๋ชจ๋‘ Brier score \(.240\) (crowd aggregate: \(.247\))์— ๋„๋‹ฌํ•˜์˜€๋‹ค. ์ด๊ฒƒ์€ ์˜ˆ์ธก์˜ 22%์™€ ์งˆ๋ฌธ์˜ 43%๋กœ ๊ตฌ์„ฑ๋œ๋‹ค.
1. ๊ตฐ์ค‘ ์˜ˆ์ธก์ด \(.3\)์—์„œ \(.7\)_ ์‚ฌ์ด์ธ ๋ฌธ์ œ์— ๋Œ€ํ•ด์„œ๋งŒ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ, ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์€ \((.238\)(๊ตฐ์ค‘ ์ดํ•ฉ \((.240\))์˜ ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค: \(.240\)). ์ด๋Š” ์˜ˆ์ธก์˜ 51%์™€ ์งˆ๋ฌธ์˜ 56%๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. 2. ์ด์ „ ๊ฒ€์ƒ‰ ๋‚ ์งœ(1, 2, 3)์— ๋Œ€ํ•ด์„œ๋งŒ ์˜ˆ์ธก. ์ด ์„ค์ •์—์„œ ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์˜ ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜๋Š” \((.185\) (๊ตฐ์ค‘ ์ดํ•ฉ: \(.161\)). ์ด๋Š” ์˜ˆ์ธก์˜ 66%์™€ ์งˆ๋ฌธ์˜ 100%๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. 3. ๊ฒ€์ƒ‰ ์‹œ์Šคํ…œ์ด ์ตœ์†Œ \(5\) ๊ฐœ์˜ ๊ด€๋ จ ๊ธฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ฒฝ์šฐ์—๋งŒ ์˜ˆ์ธก_ ์ด ์กฐ๊ฑด์—์„œ ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์˜ ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜๋Š” \((.175\) (๊ตฐ์ค‘ ์ง‘๊ณ„: \(.143\)). ์ด๋Š” ์˜ˆ์ธก์˜ 84%, ์งˆ๋ฌธ์˜ 94%๋ฅผ ์ฐจ์ง€ํ•ฉ๋‹ˆ๋‹ค. 4. ์„ธ ๊ฐ€์ง€ ์กฐ๊ฑด ๋ชจ๋‘์—์„œ ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์€ ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜ \((.240\)(๊ตฐ์ค‘ ์ดํ•ฉ: \(.247\)). ์ด๋Š” ์˜ˆ์ธก์˜ 22%์™€ ์งˆ๋ฌธ์˜ 43%๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.
Figure 1: Performance of Pythia models on the Pile. On the left, we compare training dynamics of models from 14M (top) to 410M (bottom) parameters, displaying darker shades as we approach the minimal value. On the right, we fit a power law on larger models and find that final checkpoints of smaller models underperform compared to predictions.
๊ทธ๋ฆผ 1: ํŒŒ์ผ์—์„œ ํ”ผํ‹ฐ์•„ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ž…๋‹ˆ๋‹ค. ์™ผ์ชฝ์—์„œ๋Š” ์ตœ์†Œ๊ฐ’์— ์ ‘๊ทผํ•  ๋•Œ ๋” ์–ด๋‘์šด ์Œ์˜์„ ํ‘œ์‹œํ•˜๋Š” 14M(์ƒ๋‹จ)์—์„œ 410M(ํ•˜๋‹จ) ๋งค๊ฐœ๋ณ€์ˆ˜๊นŒ์ง€์˜ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ์—ญํ•™์„ ๋น„๊ตํ•œ๋‹ค. ์˜ค๋ฅธ์ชฝ์—์„œ ์šฐ๋ฆฌ๋Š” ๋” ํฐ ๋ชจ๋ธ์— ๋ฉฑ๋ฒ•์น™์— ์ ํ•ฉํ•˜๊ณ  ๋” ์ž‘์€ ๋ชจ๋ธ์˜ ์ตœ์ข… ์ฒดํฌํฌ์ธํŠธ๊ฐ€ ์˜ˆ์ธก์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ๋‚ฎ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ•œ๋‹ค.
๊ทธ๋ฆผ 1: ํŒŒ์ผ์—์„œ Pythia ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ. ์™ผ์ชฝ์—์„œ๋Š” 14M(์œ„)์—์„œ 410M(์•„๋ž˜) ๋งค๊ฐœ๋ณ€์ˆ˜๊นŒ์ง€ ๋ชจ๋ธ์˜ ํ•™์Šต ์—ญํ•™์„ ๋น„๊ตํ•˜์—ฌ ์ตœ์†Œ๊ฐ’์— ๊ฐ€๊นŒ์›Œ์งˆ์ˆ˜๋ก ๋” ์–ด๋‘์šด ์Œ์˜์„ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์˜ค๋ฅธ์ชฝ์—์„œ๋Š” ๋” ํฐ ๋ชจ๋ธ์— ํŒŒ์›Œ ๋ฒ•์น™์„ ์ ์šฉํ•˜์—ฌ ๋” ์ž‘์€ ๋ชจ๋ธ์˜ ์ตœ์ข… ์ฒดํฌํฌ์ธํŠธ๊ฐ€ ์˜ˆ์ธก์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค.
5. Now, we have another instance of \(f\) and \(g\). Since \(f\) is odd and \(g\) is even, their product is an odd function. So, we can rewrite the expression as \(-[f(g(f(x)))]\).
5. ์ด์ œ \(f\)์™€ \(g\)์˜ ๋‹ค๋ฅธ ์ธ์Šคํ„ด์Šค๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. \(f\)๋Š” ํ™€์ˆ˜์ด๊ณ  \(g\)๋Š” ์ง์ˆ˜์ด๋ฏ€๋กœ ์ด๋“ค์˜ ๊ณฑ์€ ํ™€์ˆ˜ ํ•จ์ˆ˜์ด๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” ์‹์„ \(-[f(g(f(x)))]\๋กœ ๋‹ค์‹œ ์“ธ ์ˆ˜ ์žˆ๋‹ค.
5. ์ด์ œ \(f\)์™€ \(g\)์˜ ๋˜ ๋‹ค๋ฅธ ์ธ์Šคํ„ด์Šค๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. (f\)๋Š” ํ™€์ˆ˜์ด๊ณ  \(g\)๋Š” ์ง์ˆ˜์ด๋ฏ€๋กœ ์ด ๋‘˜์˜ ๊ณฑ์€ ํ™€์ˆ˜ ํ•จ์ˆ˜์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹์„ \(-[f(g(f(x))]\)๋กœ ๋‹ค์‹œ ์“ธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.)
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.
์ฃผ์–ด์ง„ ์งˆ๋ฌธ์— ๋Œ€ํ•ด ๊ฐ ํ”Œ๋žซํผ์€ ๋ชจ๋“  ๊ฐœ๋ณ„ ์˜ˆ์ธก์„ ์ง‘๊ณ„ํ•˜๋Š” ์ปค๋ฎค๋‹ˆํ‹ฐ ์˜ˆ์ธก์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ์ธก์€ ์˜ˆ์ธก๋“ค์ด ์ด๋ฃจ์–ด์ง์— ๋”ฐ๋ผ ๋™์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋˜๊ณ  ๊ธฐ๋ก๋œ๋‹ค. ์šฐ๋ฆฌ๋Š” (๊ฐœ๋ณ„ ์˜ˆ์ธก์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฒ˜์Œ๋ถ€ํ„ฐ ๊ณ„์‚ฐํ•˜๋Š” ๋Œ€์‹ ) ํ”Œ๋žซํผ์—์„œ ์ง์ ‘ ๊ธฐ๋ก์„ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค. ์ด์ง„ ์งˆ๋ฌธ์˜ ๊ฒฝ์šฐ ์ง‘๊ณ„ ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
์ฃผ์–ด์ง„ ์งˆ๋ฌธ์— ๋Œ€ํ•ด ๊ฐ ํ”Œ๋žซํผ์€ ๋ชจ๋“  ๊ฐœ๋ณ„ ์˜ˆ์ธก์„ ์ง‘๊ณ„ํ•˜๋Š” ์ปค๋ฎค๋‹ˆํ‹ฐ ์˜ˆ์ธก์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ์ธก์€ ์˜ˆ์ธก์ด ์ด๋ฃจ์–ด์งˆ ๋•Œ๋งˆ๋‹ค ๋™์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋˜๊ณ  ๊ธฐ๋ก๋ฉ๋‹ˆ๋‹ค. ๊ฐœ๋ณ„ ์˜ˆ์ธก์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฒ˜์Œ๋ถ€ํ„ฐ ๊ณ„์‚ฐํ•˜๋Š” ๋Œ€์‹  ํ”Œ๋žซํผ์—์„œ ์ง์ ‘ ๊ธฐ๋ก์„ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค. ์ด์ง„ ์งˆ๋ฌธ์˜ ๊ฒฝ์šฐ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ง‘๊ณ„ ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
As visualized in Appendix Figure 7, we find that on average there is little improvement in the LM's ability to predict arbitrary tokens. But, when we visualize the distribution of relative improvements, there is a disproportionate improvement on more difficult tokens. This reflects the idea that some text tokens are substantially harder and benefit more from careful thought.
๋ถ€๋ก ๊ทธ๋ฆผ 7์—์„œ ์‹œ๊ฐํ™”ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ์šฐ๋ฆฌ๋Š” ํ‰๊ท ์ ์œผ๋กœ ์ž„์˜์˜ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” LM ๋Šฅ๋ ฅ์˜ ํ–ฅ์ƒ์ด ๊ฑฐ์˜ ์—†์Œ์„ ๋ฐœ๊ฒฌํ•œ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ƒ๋Œ€์  ๊ฐœ์„ ์˜ ๋ถ„ํฌ๋ฅผ ์‹œ๊ฐํ™”ํ•˜๋ฉด ๋” ์–ด๋ ค์šด ํ† ํฐ์— ๋ถˆ๊ท ํ˜•์ ์ธ ๊ฐœ์„ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๊ฒƒ์€ ์ผ๋ถ€ ํ…์ŠคํŠธ ํ† ํฐ์ด ์‹ค์งˆ์ ์œผ๋กœ ๋” ์–ด๋ ต๊ณ  ์‹ ์ค‘ํ•œ ์ƒ๊ฐ์œผ๋กœ๋ถ€ํ„ฐ ๋” ๋งŽ์€ ์ด์ต์„ ์–ป๋Š”๋‹ค๋Š” ์ƒ๊ฐ์„ ๋ฐ˜์˜ํ•œ๋‹ค.
๋ถ€๋ก ๊ทธ๋ฆผ 7์—์„œ ์‹œ๊ฐํ™”๋œ ๊ฒƒ์ฒ˜๋Ÿผ, ํ‰๊ท ์ ์œผ๋กœ ์ž„์˜์˜ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” LM์˜ ๋Šฅ๋ ฅ์€ ๊ฑฐ์˜ ๊ฐœ์„ ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ƒ๋Œ€์ ์ธ ๊ฐœ์„ ์˜ ๋ถ„ํฌ๋ฅผ ์‹œ๊ฐํ™”ํ•˜๋ฉด ๋‚œ์ด๋„๊ฐ€ ๋†’์€ ํ† ํฐ์—์„œ ๋ถˆ๊ท ํ˜•์ ์ธ ๊ฐœ์„ ์ด ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ผ๋ถ€ ํ…์ŠคํŠธ ํ† ํฐ์ด ํ›จ์”ฌ ๋” ์–ด๋ ต๊ณ  ์‹ ์ค‘ํ•œ ์‚ฌ๊ณ ๋ฅผ ํ†ตํ•ด ๋” ๋งŽ์€ ์ด์ ์„ ์–ป์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ƒ๊ฐ์„ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค.
**Effects of Question Verification.** The question verification is used to further improve the generation quality. In our experiments, we found it can improve the performance on MATH benchmark, the results are shown in Table 5, while we do not see significantly impact on GSM8K dataset.
**์งˆ๋ฌธ ๊ฒ€์ฆ์˜ ํšจ๊ณผ** ์งˆ๋ฌธ ๊ฒ€์ฆ์€ ์ƒ์„ฑ ํ’ˆ์งˆ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์‹คํ—˜์—์„œ MATH ๋ฒค์น˜๋งˆํฌ์—์„œ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ ๊ฒฐ๊ณผ๋Š” ํ‘œ 5์— ๋‚˜์™€ ์žˆ์ง€๋งŒ GSM8K ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋Š”๋‹ค.
**๋ฌธ์ œ ๊ฒ€์ฆ์˜ ํšจ๊ณผ **๋ฌธ์ œ ๊ฒ€์ฆ์€ ์ƒ์„ฑ ํ’ˆ์งˆ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, MATH ๋ฒค์น˜๋งˆํฌ์—์„œ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ, ๊ทธ ๊ฒฐ๊ณผ๋Š” ํ‘œ 5์— ๋‚˜์™€ ์žˆ์ง€๋งŒ GSM8K ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
Despite the considerable strides made by LLaMA and Alpaca in NLP, they exhibit inherent limitations concerning native support for Chinese language tasks. Their vocabularies contain only a few hundred Chinese tokens, substantially hindering their efficiency in encoding and decoding Chinese text. Building on our previous work with the Chinese BERT series (Cui et al., 2021) and Chinese minority-oriented multilingual pre-trained models (Yang et al., 2022), in this technical report, we propose the development of Chinese LLaMA and Alpaca models with enhanced capabilities for understanding and generating Chinese content. We extend the original LLaMA's vocabulary with an additional 20,000 Chinese tokens, significantly improving its proficiency in processing and generating Chinese text. To ensure efficient training and deployment of these models, we employ the Low-Rank Adaptation (LoRA) approach (Hu et al., 2021), enabling us to train and fine-tune the models without excessive computational costs. We anticipate our preliminary study to enhance the Chinese understanding and generation capabilities of LLaMA and Alpaca serves as a foundation for researchers aiming to adapt these models to other languages. By showcasing the feasibility and effectiveness of our approach, we offer insights and methodologies that can be employed to extend vocabularies and improve the performance of LLaMA and Alpaca models in various languages. An overview of the proposed models is depicted in Figure 1.
LLaMA์™€ Alpaca๊ฐ€ NLP์—์„œ ์ƒ๋‹นํ•œ ์ง„์ „์„ ์ด๋ฃจ์—ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ค‘๊ตญ์–ด ์ž‘์—…์— ๋Œ€ํ•œ ๋ชจ๊ตญ์–ด ์ง€์›๊ณผ ๊ด€๋ จํ•˜์—ฌ ๊ณ ์œ ํ•œ ํ•œ๊ณ„๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค. ๊ทธ๋“ค์˜ ์–ดํœ˜๋Š” ์ˆ˜๋ฐฑ ๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ๋งŒ์„ ํฌํ•จํ•˜๊ณ  ์žˆ์–ด ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ์˜ ์ธ์ฝ”๋”ฉ๊ณผ ๋””์ฝ”๋”ฉ์˜ ํšจ์œจ์„ฑ์„ ์ƒ๋‹นํžˆ ์ €ํ•ดํ•˜๊ณ  ์žˆ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ค‘๊ตญ BERT ์‹œ๋ฆฌ์ฆˆ(Cui et al., 2021)์™€ ์ค‘๊ตญ ์†Œ์ˆ˜์ž ์ค‘์‹ฌ์˜ ๋‹ค๊ตญ์–ด ์‚ฌ์ „ ํ›ˆ๋ จ ๋ชจ๋ธ(Yang et al., 2022)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ค‘๊ตญ ์ฝ˜ํ…์ธ ์˜ ์ดํ•ด์™€ ์ƒ์„ฑ์„ ์œ„ํ•œ ํ–ฅ์ƒ๋œ ๊ธฐ๋Šฅ์„ ๊ฐ€์ง„ ์ค‘๊ตญ LLaMA ๋ฐ Alpaca ๋ชจ๋ธ์˜ ๊ฐœ๋ฐœ์„ ์ œ์•ˆํ•œ๋‹ค. ์ถ”๊ฐ€ 2๋งŒ ๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ์œผ๋กœ ์›๋ณธ LLaMA์˜ ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜์—ฌ ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ์˜ ์ฒ˜๋ฆฌ ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ํšจ์œจ์ ์ธ ํ›ˆ๋ จ๊ณผ ๋ฐฐ์น˜๋ฅผ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด LoRA(Low-Rank Adaptation) ์ ‘๊ทผ๋ฒ•(Hu et al., 2021)์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณผ๋„ํ•œ ๊ณ„์‚ฐ ๋น„์šฉ ์—†์ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ๋ฏธ์„ธ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ๋‹ค. ์šฐ๋ฆฌ๋Š” LLaMA์™€ Alpaca์˜ ์ค‘๊ตญ์–ด ์ดํ•ด์™€ ์ƒ์„ฑ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ์˜ˆ๋น„ ์—ฐ๊ตฌ๊ฐ€ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ๋‹ค๋ฅธ ์–ธ์–ด์— ์ ์šฉํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ์—ฐ๊ตฌ์ž๋“ค์˜ ํ† ๋Œ€๊ฐ€ ๋  ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•œ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ์–ธ์–ด์—์„œ LLaMA์™€ Alpaca ๋ชจ๋ธ์˜ ์–ดํœ˜ ํ™•์žฅ๊ณผ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์œ„ํ•ด ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋Š” ํ†ต์ฐฐ๋ ฅ๊ณผ ๋ฐฉ๋ฒ•๋ก ์„ ์ œ์‹œํ•œ๋‹ค. ์ œ์•ˆ๋œ ๋ชจ๋ธ์˜ ๊ฐœ์š”๋Š” ๊ทธ๋ฆผ 1์— ๋‚˜์™€ ์žˆ๋‹ค.
LLaMA์™€ ์•ŒํŒŒ์นด๋Š” NLP ๋ถ„์•ผ์—์„œ ์ƒ๋‹นํ•œ ์ง„์ „์„ ์ด๋ฃจ์—ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ค‘๊ตญ์–ด ์ž‘์—…์— ๋Œ€ํ•œ ๋„ค์ดํ‹ฐ๋ธŒ ์ง€์›๊ณผ ๊ด€๋ จํ•˜์—ฌ ๋ณธ์งˆ์ ์ธ ํ•œ๊ณ„๋ฅผ ๋“œ๋Ÿฌ๋‚ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์–ดํœ˜์—๋Š” ์ˆ˜๋ฐฑ ๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ๋งŒ ํฌํ•จ๋˜์–ด ์žˆ์–ด ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ ์ธ์ฝ”๋”ฉ ๋ฐ ๋””์ฝ”๋”ฉ์˜ ํšจ์œจ์„ฑ์ด ํฌ๊ฒŒ ๋–จ์–ด์ง‘๋‹ˆ๋‹ค. ์ด ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ์—์„œ๋Š” ์ค‘๊ตญ์–ด BERT ์‹œ๋ฆฌ์ฆˆ(Cui et al., 2021)์™€ ์ค‘๊ตญ ์†Œ์ˆ˜๋ฏผ์กฑ ์ค‘์‹ฌ์˜ ๋‹ค๊ตญ์–ด ์‚ฌ์ „ ํ•™์Šต ๋ชจ๋ธ(Yang et al., 2022)์— ๋Œ€ํ•œ ์ด์ „ ์ž‘์—…์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ค‘๊ตญ์–ด ์ฝ˜ํ…์ธ  ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๊ธฐ๋Šฅ์ด ํ–ฅ์ƒ๋œ ์ค‘๊ตญ์–ด LLaMA ๋ฐ ์•ŒํŒŒ์นด ๋ชจ๋ธ์˜ ๊ฐœ๋ฐœ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด LLaMA์˜ ์–ดํœ˜๋ฅผ ์ถ”๊ฐ€๋กœ 20,000๊ฐœ์˜ ์ค‘๊ตญ์–ด ํ† ํฐ์œผ๋กœ ํ™•์žฅํ•˜์—ฌ ์ค‘๊ตญ์–ด ํ…์ŠคํŠธ ์ฒ˜๋ฆฌ ๋ฐ ์ƒ์„ฑ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ํšจ์œจ์ ์œผ๋กœ ํ›ˆ๋ จํ•˜๊ณ  ๋ฐฐํฌํ•˜๊ธฐ ์œ„ํ•ด ์ €์ˆœ์œ„ ์ ์‘(LoRA) ์ ‘๊ทผ๋ฒ•(Hu et al., 2021)์„ ์ฑ„ํƒํ•˜์—ฌ ๊ณผ๋„ํ•œ ๊ณ„์‚ฐ ๋น„์šฉ ์—†์ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ๋ฏธ์„ธ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ์˜ ์˜ˆ๋น„ ์—ฐ๊ตฌ๋Š” LLaMA์™€ ์•ŒํŒŒ์นด์˜ ์ค‘๊ตญ์–ด ์ดํ•ด์™€ ์ƒ์„ฑ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๊ฒƒ์œผ๋กœ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ๋‹ค๋ฅธ ์–ธ์–ด์— ์ ์šฉํ•˜๋ ค๋Š” ์—ฐ๊ตฌ์ž๋“ค์—๊ฒŒ ๊ธฐ์ดˆ๊ฐ€ ๋  ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํƒ€๋‹น์„ฑ๊ณผ ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์คŒ์œผ๋กœ์จ ๋‹ค์–‘ํ•œ ์–ธ์–ด๋กœ ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜๊ณ  LLaMA ๋ฐ ์•ŒํŒŒ์นด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ธ์‚ฌ์ดํŠธ์™€ ๋ฐฉ๋ฒ•๋ก ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ œ์•ˆ๋œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ฐœ์š”๋Š” ๊ทธ๋ฆผ 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
Fig. 1: The trends of the cumulative numbers of arXiv papers that contain the keyphrases _โ€œlanguage modelโ€_ (since June 2018) and _โ€œlarge language modelโ€_ (since October 2019), respectively. The statistics are calculated using exact match by querying the keyphrases in title or abstract by months. We set different x-axis ranges for the two keyphrases, because โ€œlanguage modelsโ€ have been explored at an earlier time. We label the points corresponding to important landmarks in the research progress of LLMs. A sharp increase occurs after the release of ChatGPT: the average number of published arXiv papers that contain _โ€œlarge language modelโ€_ in title or abstract goes from 0.40 per day to 8.58 per day (Figure 1(b)).
๋„. 1: ํ‚คํ”„๋ ˆ์ด์ฆˆ _โ€œ์–ธ์–ด ๋ชจ๋ธโ€_(2018๋…„ 6์›” ์ดํ›„)์™€ _โ€œ๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธโ€_(2019๋…„ 10์›” ์ดํ›„)๊ฐ€ ๊ฐ๊ฐ ํฌํ•จ๋œ arXiv ๋…ผ๋ฌธ์˜ ๋ˆ„์  ์ˆ˜ ์ถ”์ด. ํ†ต๊ณ„๋Ÿ‰์€ ์ œ๋ชฉ ๋˜๋Š” ์ดˆ๋ก์˜ ํ‚คํ”„๋ ˆ์ด์ฆˆ๋ฅผ ์›”๋ณ„๋กœ ์กฐํšŒํ•˜์—ฌ ์ •ํ™•ํ•œ ์ผ์น˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋‘ ํ‚คํ”„๋ ˆ์ด์ฆˆ์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ x์ถ• ๋ฒ”์œ„๋ฅผ ์„ค์ •ํ–ˆ๋Š”๋ฐ, ์ด๋Š” "์–ธ์–ด ๋ชจ๋ธ"์ด ์ด์ „ ์‹œ๊ฐ„์— ํƒ์ƒ‰๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. ์šฐ๋ฆฌ๋Š” LLM์˜ ์—ฐ๊ตฌ ์ง„ํ–‰ ๊ณผ์ •์—์„œ ์ค‘์š”ํ•œ ๋žœ๋“œ๋งˆํฌ์— ํ•ด๋‹นํ•˜๋Š” ์ง€์ ์— ๋ ˆ์ด๋ธ”์„ ๋ถ™์ธ๋‹ค. ChatGPT๊ฐ€ ๋ฐœํ‘œ๋œ ํ›„ ๊ธ‰๊ฒฉํ•œ ์ฆ๊ฐ€๊ฐ€ ์ผ์–ด๋‚˜๋Š”๋ฐ, ์ œ๋ชฉ์ด๋‚˜ ์ดˆ๋ก์— _โ€œํฐ ์–ธ์–ด ๋ชจ๋ธโ€_์ด ํฌํ•จ๋œ arXiv ๋…ผ๋ฌธ์˜ ํ‰๊ท  ๊ฒŒ์žฌ ๊ฑด์ˆ˜๋Š” ํ•˜๋ฃจ 0.40๊ฑด์—์„œ 8.58๊ฑด์œผ๋กœ ๋Š˜์–ด๋‚œ๋‹ค(๊ทธ๋ฆผ 1(b)).
๊ทธ๋ฆผ 1: _"์–ธ์–ด ๋ชจ๋ธ"_(2018๋…„ 6์›” ์ดํ›„) ๋ฐ _"๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ"_(2019๋…„ 10์›” ์ดํ›„)์ด๋ผ๋Š” ํ‚ค์›Œ๋“œ๊ฐ€ ํฌํ•จ๋œ arXiv ๋…ผ๋ฌธ์˜ ๋ˆ„์  ์ˆ˜ ์ถ”์ด. ํ†ต๊ณ„๋Š” ์ œ๋ชฉ ๋˜๋Š” ์ดˆ๋ก์˜ ํ‚คํ”„๋ ˆ์ด์ฆˆ๋ฅผ ์›”๋ณ„๋กœ ์ฟผ๋ฆฌํ•˜์—ฌ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ๊ณ„์‚ฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. '์–ธ์–ด ๋ชจ๋ธ'์€ ๋” ์ด๋ฅธ ์‹œ๊ธฐ์— ํƒ์ƒ‰๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋‘ ํ‚ค๋ฌธ๊ตฌ์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ X์ถ• ๋ฒ”์œ„๋ฅผ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. LLM์˜ ์—ฐ๊ตฌ ์ง„ํ–‰์—์„œ ์ค‘์š”ํ•œ ๋žœ๋“œ๋งˆํฌ์— ํ•ด๋‹นํ•˜๋Š” ์ง€์ ์— ๋ผ๋ฒจ์„ ๋ถ™์˜€์Šต๋‹ˆ๋‹ค. ์ œ๋ชฉ์ด๋‚˜ ์ดˆ๋ก์— "๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ"์ด ํฌํ•จ๋œ arXiv ๋…ผ๋ฌธ์˜ ํ‰๊ท  ๊ฒŒ์žฌ ๊ฑด์ˆ˜๊ฐ€ ํ•˜๋ฃจ 0.40๊ฑด์—์„œ 8.58๊ฑด์œผ๋กœ ๊ธ‰๊ฒฉํžˆ ์ฆ๊ฐ€ํ•œ ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1(b)).
In this section, we first show the main experimental results for the CKL Benchmark. Then, since multiple steps of continual knowledge learning, i.e., CKL are needed for training a true, ever-changing LM, we explore the effects of multiple CKL phases as well as how epochs, corpus size, and the total number of training steps affect CKL. We further explore how learning rates affect CKL in Appendix E, how continual pretraining on \(D_{1}\) affects the performance of KILT tasks which re quire knowledge from \(D_{0}\) in Appendix F, how CKL methods transfer across LM architectures in Appendix G, and how the prediction outputs change during CKL in Appendix H.
์ด ์ ˆ์—์„œ๋Š” ๋จผ์ € CKL ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ ์ฃผ์š” ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ์ง€์†์ ์ธ ์ง€์‹ ํ•™์Šต์˜ ์—ฌ๋Ÿฌ ๋‹จ๊ณ„, ์ฆ‰ CKL์ด ์ง„์ •์œผ๋กœ ๋Š์ž„์—†์ด ๋ณ€ํ™”ํ•˜๋Š” LM์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์—, ์šฐ๋ฆฌ๋Š” ์—ฌ๋Ÿฌ CKL ๋‹จ๊ณ„์˜ ์˜ํ–ฅ๊ณผ epoch, ๋ง๋ญ‰์น˜ ํฌ๊ธฐ ๋ฐ ์ด ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜๊ฐ€ CKL์— ์–ด๋–ป๊ฒŒ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ํƒ๊ตฌํ•œ๋‹ค. ๋˜ํ•œ ํ•™์Šต๋ฅ ์ด ๋ถ€๋ก E์˜ CKL์— ์–ด๋–ป๊ฒŒ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€, ๋ถ€๋ก F์˜ \(D_{0}\)์—์„œ ์ง€์‹์„ ์š”๊ตฌํ•˜๋Š” KILT ํƒœ์Šคํฌ์˜ ์„ฑ๋Šฅ์— ์–ด๋–ป๊ฒŒ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€, ๋ถ€๋ก G์˜ LM ์•„ํ‚คํ…์ฒ˜ ๊ฐ„์— CKL ๋ฐฉ๋ฒ•์ด ์–ด๋–ป๊ฒŒ ์ „๋‹ฌ๋˜๋Š”์ง€, ๋ถ€๋ก H์˜ CKL ๋™์•ˆ ์˜ˆ์ธก ์ถœ๋ ฅ์ด ์–ด๋–ป๊ฒŒ ๋ณ€ํ™”ํ•˜๋Š”์ง€๋ฅผ ํƒ๊ตฌํ•œ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋จผ์ € CKL ๋ฒค์น˜๋งˆํฌ์˜ ์ฃผ์š” ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ์ง€์†์ ์œผ๋กœ ๋ณ€ํ™”ํ•˜๋Š” ์ง„์ •ํ•œ LM์„ ํ•™์Šตํ•˜๋ ค๋ฉด ์—ฌ๋Ÿฌ ๋‹จ๊ณ„์˜ ์ง€์†์ ์ธ ์ง€์‹ ํ•™์Šต, ์ฆ‰ CKL์ด ํ•„์š”ํ•˜๋ฏ€๋กœ ์—ฌ๋Ÿฌ CKL ๋‹จ๊ณ„์˜ ํšจ๊ณผ์™€ ์—ํฌํฌ, ์ฝ”ํผ์Šค ํฌ๊ธฐ, ์ด ํ•™์Šต ๋‹จ๊ณ„ ์ˆ˜๊ฐ€ CKL์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ถ€๋ก E์—์„œ๋Š” ํ•™์Šต๋ฅ ์ด CKL์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ, ๋ถ€๋ก F์—์„œ๋Š” \(D_{1}\)์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์ด \(D_{0}\)์˜ ์ง€์‹์„ ๋‹ค์‹œ ์š”๊ตฌํ•˜๋Š” KILT ์ž‘์—…์˜ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ, ๋ถ€๋ก G์—์„œ๋Š” LM ์•„ํ‚คํ…์ฒ˜ ๊ฐ„์— CKL ๋ฐฉ๋ฒ•์ด ์–ด๋–ป๊ฒŒ ์ด์ „๋˜๋Š”์ง€, ๋ถ€๋ก H์—์„œ๋Š” CKL ์ค‘์— ์˜ˆ์ธก ๊ฒฐ๊ณผ๊ฐ€ ์–ด๋–ป๊ฒŒ ๋ณ€ํ™”ํ•˜๋Š”์ง€๋ฅผ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.
Figure 3 reports validation losses for the \(\mathcal{D}_{0}\) and \(\mathcal{D}_{1}\) datasets throughout the continual pre-training of all models. The top row of plots reports results for the weak distribution shift (300B Pile\(\rightarrow\)300B SP), while the bottom row reports results for the strong distribution shift (300B Pile\(\rightarrow\)200B Ger.). For both shifts, the \(\eta_{\text{min}}\) constant learning rate model achieves the least forgetting on \(\mathcal{D}_{0}\). It also adapts the least on \(\mathcal{D}_{1}\) for the strong shift, however, for the weak shift it adapts more than the \(\eta_{\text{max}}\) constant baseline. When comparing these baselines to the models that re-warm and re-decay on both datasets, we observe that the latter models adapt better to the new dataset by a significant margin for both distribution shifts. This shows that re-warming and re-decaying are necessary to maximize adaptation to the new dataset when continually pre-training LLMs. Among the models that re-warm and re-decay the LR, we observe that varying the learning rate causes small differences in adaptation and forgetting: higher values of \(\eta_{\text{max}}\) lead to more forgetting and more adaptation while the opposite is true for lower values. When comparing the baselines to the union-trained baseline, we observe that the final validation loss for \(\mathcal{D}_{1}\) is significantly higher than the union-trained model on both distribution shifts. This is also the case for \(\mathcal{D}_{1}\) on the weak distribution shift, but interestingly for the strong distribution shift, the constant baselines achieve lower \(\mathcal{D}_{1}\) validation loss than the union-trained model. We hypothesize that this is due to the stronger distribution shift generally enhancing adaptation and exacerbating forgetting in the context of LLMs. When comparing models continually pre-trained with re-warming and re-decaying to the union baseline, we note that these models adapt better (lower final validation loss) to \(\mathcal{D}_{1}\) than the union baseline. However, these models experience significant forgetting on \(\mathcal{D}_{0}\), showing the need for replay to make these models competitive with the union baseline.
๊ทธ๋ฆผ 3์€ ๋ชจ๋“  ๋ชจ๋ธ์˜ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ ์ „๋ฐ˜์— ๊ฑธ์ณ \(\mathcal{D}_{0}\) ๋ฐ \(\mathcal{D}_{1}\) ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์†์‹ค์„ ๋ณด๊ณ ํ•œ๋‹ค. ๊ทธ๋ฆผ์˜ ์ƒ๋‹จ ํ–‰์€ ์•ฝํ•œ ๋ถ„ํฌ ์ด๋™(300B Pile\(\rightarrow\)300B SP)์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•˜๊ณ  ํ•˜๋‹จ ํ–‰์€ ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™(300B Pile\(\rightarrow\)200B Ger)์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•œ๋‹ค. ๋‘ ์ด๋™ ๋ชจ๋‘์—์„œ \(\eta_{\text{min}}\) ์ƒ์ˆ˜ ํ•™์Šต๋ฅ  ๋ชจ๋ธ์€ \(\mathcal{D}_{0}\)์—์„œ ๊ฐ€์žฅ ์ ์€ ๋ง๊ฐ์„ ๋‹ฌ์„ฑํ•œ๋‹ค. ๋˜ํ•œ ๊ฐ•ํ•œ ์‹œํ”„ํŠธ์˜ ๊ฒฝ์šฐ \(\mathcal{D}_{1}\)์— ๊ฐ€์žฅ ์ ๊ฒŒ ์ ์‘ํ•˜์ง€๋งŒ ์•ฝํ•œ ์‹œํ”„ํŠธ์˜ ๊ฒฝ์šฐ \(\eta_{\text{max}}\) ์ƒ์ˆ˜ ๊ธฐ์ค€์„ ๋ณด๋‹ค ๋” ๋งŽ์ด ์ ์‘ํ•œ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ์ค€์„ ์„ ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋‹ค์‹œ ๋”ฐ๋œปํ•ด์ง€๊ณ  ๋‹ค์‹œ ๋ถ€ํŒจํ•˜๋Š” ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•  ๋•Œ ํ›„์ž์˜ ๋ชจ๋ธ์ด ๋‘ ๋ถ„ํฌ ์ด๋™์— ๋Œ€ํ•ด ์ƒ๋‹นํ•œ ๋งˆ์ง„๋งŒํผ ์ƒˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋” ์ž˜ ์ ์‘ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•œ๋‹ค. ์ด๋Š” LLM์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จํ•  ๋•Œ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ ์‘์„ ์ตœ๋Œ€ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์žฌ-์˜จ๋‚œํ™” ๋ฐ ์žฌ-๋ถ•๊ดด๊ฐ€ ํ•„์š”ํ•จ์„ ๋ณด์—ฌ์ค€๋‹ค. LR์„ ์žฌ-์˜จํ™”ํ•˜๊ณ  ์žฌ-๋ถ•๊ดด์‹œํ‚ค๋Š” ๋ชจ๋ธ๋“ค ์ค‘์—์„œ, ํ•™์Šต๋ฅ ์˜ ๋ณ€ํ™”๊ฐ€ ์ ์‘๊ณผ ๋ง๊ฐ์˜ ์ž‘์€ ์ฐจ์ด๋ฅผ ์•ผ๊ธฐํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•œ๋‹ค: \(\eta_{\text{max}}\)์˜ ๊ฐ’์ด ๋†’์„์ˆ˜๋ก ๋ง๊ฐ์ด ๋” ๋งŽ์•„์ง€๊ณ  ์ ์‘์ด ๋” ๋งŽ์•„์ง€๋Š” ๋ฐ˜๋ฉด, ๋‚ฎ์€ ๊ฐ’์—์„œ๋Š” ๊ทธ ๋ฐ˜๋Œ€์ด๋‹ค. ๊ธฐ์ค€์„ ์„ ์กฐํ•ฉ ํ›ˆ๋ จ๋œ ๊ธฐ์ค€์„ ๊ณผ ๋น„๊ตํ•  ๋•Œ, \(\mathcal{D}_{1}\)์— ๋Œ€ํ•œ ์ตœ์ข… ๊ฒ€์ฆ ์†์‹ค์ด ๋‘ ๋ถ„ํฌ ์ด๋™ ๋ชจ๋‘์—์„œ ์กฐํ•ฉ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๋ณด๋‹ค ์ƒ๋‹นํžˆ ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•œ๋‹ค. ์ด๋Š” ์•ฝํ•œ ๋ถ„ํฌ ์ด๋™์—์„œ๋Š” \(\mathcal{D}_{1}\)์˜ ๊ฒฝ์šฐ๋„ ๋งˆ์ฐฌ๊ฐ€์ง€์ด์ง€๋งŒ, ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™์—์„œ๋Š” ์ƒ์ˆ˜ ๊ธฐ์ค€์„ ์ด ๊ฒฐํ•ฉ ํ•™์Šต ๋ชจ๋ธ๋ณด๋‹ค ๋‚ฎ์€ \(\mathcal{D}_{1}\) ๊ฒ€์ฆ ์†์‹ค์„ ๋‹ฌ์„ฑํ•œ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๊ฒƒ์ด ์ผ๋ฐ˜์ ์œผ๋กœ LLM์˜ ๋งฅ๋ฝ์—์„œ ์ ์‘์„ ๊ฐ•ํ™”ํ•˜๊ณ  ๋ง๊ฐ์„ ์•…ํ™”์‹œํ‚ค๋Š” ๋” ๊ฐ•ํ•œ ๋ถ„ํฌ ๋ณ€ํ™” ๋•Œ๋ฌธ์ด๋ผ๊ณ  ๊ฐ€์ •ํ•œ๋‹ค. ์žฌ-์˜จ๋‚œํ™”(re-warming) ๋ฐ ์žฌ-๋ถ•๊ดด(re-decaying)๋กœ ์—ฐ์†์ ์œผ๋กœ ์‚ฌ์ „ ํŠธ๋ ˆ์ด๋‹๋œ ๋ชจ๋ธ๋“ค์„ ์กฐํ•ฉ ๊ธฐ์ค€์„ ์— ๋น„๊ตํ•  ๋•Œ, ์ด๋“ค ๋ชจ๋ธ๋“ค์€ ์กฐํ•ฉ ๊ธฐ์ค€์„ ๋ณด๋‹ค \(\mathcal{D}_{1}\)์— ๋” ์ž˜ ์ ์‘(๋‚ฎ์€ ์ตœ์ข… ๊ฒ€์ฆ ์†์‹ค)ํ•œ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜, ์ด ๋ชจ๋ธ๋“ค์€ \(\mathcal{D}_{0}\)์—์„œ ์ƒ๋‹นํ•œ ๋ง๊ฐ์„ ๊ฒฝํ—˜ํ•˜๋ฉฐ, ์ด ๋ชจ๋ธ๋“ค์ด ์กฐํ•ฉ ๊ธฐ์ค€์„ ๊ณผ ๊ฒฝ์Ÿํ•˜๋„๋ก ํ•˜๊ธฐ ์œ„ํ•œ ์žฌ์ƒ์˜ ํ•„์š”์„ฑ์„ ๋ณด์—ฌ์ค€๋‹ค.
๊ทธ๋ฆผ 3์€ ๋ชจ๋“  ๋ชจ๋ธ์˜ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์— ๋”ฐ๋ฅธ \(\mathcal{D}_{0}\) ๋ฐ \(\mathcal{D}_{1}\) ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒ€์ฆ ์†์‹ค์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ”Œ๋กฏ์˜ ์œ„์ชฝ ์ค„์€ ์•ฝํ•œ ๋ถ„ํฌ ์ด๋™(300B Pile\(\์šฐ์ธก ํ™”์‚ดํ‘œ\)300B SP)์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ, ์•„๋ž˜์ชฝ ์ค„์€ ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™(300B Pile\(\์šฐ์ธก ํ™”์‚ดํ‘œ\)200B Ger.)์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๊ฐ€์ง€ ์‹œํ”„ํŠธ ๋ชจ๋‘์—์„œ \(\eta_{\text{min}}\) ์ƒ์ˆ˜ ํ•™์Šต๋ฅ  ๋ชจ๋ธ์€ \(\mathcal{D}_{0}\)์—์„œ ๊ฐ€์žฅ ์ ์€ ๋ง๊ฐ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฐ•ํ•œ ์‹œํ”„ํŠธ์˜ ๊ฒฝ์šฐ \(\mathcal{D}_{1}\)์—์„œ ๊ฐ€์žฅ ์ ๊ฒŒ ์ ์‘ํ•˜์ง€๋งŒ ์•ฝํ•œ ์‹œํ”„ํŠธ์˜ ๊ฒฝ์šฐ \(\eta_{\text{max}}\) ์ƒ์ˆ˜ ๊ธฐ์ค€์„ ๋ณด๋‹ค ๋” ๋งŽ์ด ์ ์‘ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ์ค€์„ ์„ ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์žฌ๊ฐ€์˜จ ๋ฐ ์žฌ๊ฐ์‡ ๋ฅผ ์ ์šฉํ•œ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•˜๋ฉด, ํ›„์ž์˜ ๋ชจ๋ธ์ด ๋‘ ๋ถ„ํฌ ์ด๋™ ๋ชจ๋‘์—์„œ ์ƒ๋‹นํ•œ ์ฐจ์ด๋กœ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋” ์ž˜ ์ ์‘ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” LLM์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•  ๋•Œ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ ์‘์„ ๊ทน๋Œ€ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋ฆฌ์›Œ๋ฐ๊ณผ ๋ฆฌ๋””์ผ€์ดํŒ…์ด ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. LR์„ ์žฌ์›Œ๋ฐํ•˜๊ณ  ์žฌ๊ฐ์‡ ํ•˜๋Š” ๋ชจ๋ธ ์ค‘ ํ•™์Šต ์†๋„๋ฅผ ๋ณ€๊ฒฝํ•˜๋ฉด ์ ์‘๊ณผ ๋ง๊ฐ์— ์ž‘์€ ์ฐจ์ด๊ฐ€ ๋ฐœ์ƒํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฆ‰, \(\eta_{\text{max}}\)์˜ ๊ฐ’์ด ๋†’์œผ๋ฉด ๋ง๊ฐ๊ณผ ์ ์‘์ด ๋” ๋งŽ์ด ์ผ์–ด๋‚˜๋Š” ๋ฐ˜๋ฉด ๊ฐ’์ด ๋‚ฎ์œผ๋ฉด ๊ทธ ๋ฐ˜๋Œ€๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ๊ธฐ์ค€์„ ๊ณผ ์—ฐํ•ฉ ํ›ˆ๋ จ ๊ธฐ์ค€์„ ์„ ๋น„๊ตํ•  ๋•Œ, \(\mathcal{D}_{1}\)์˜ ์ตœ์ข… ๊ฒ€์ฆ ์†์‹ค์€ ๋‘ ๋ถ„ํฌ ์ด๋™ ๋ชจ๋‘์—์„œ ์—ฐํ•ฉ ํ›ˆ๋ จ ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์•ฝํ•œ ๋ถ„ํฌ ์ด๋™์— ๋Œ€ํ•œ \(\mathcal{D}_{1}\)์˜ ๊ฒฝ์šฐ์—๋„ ๋งˆ์ฐฌ๊ฐ€์ง€์ด์ง€๋งŒ, ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™์— ๋Œ€ํ•ด์„œ๋Š” ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ์ƒ์ˆ˜ ๊ธฐ์ค€์„ ์ด ์œ ๋‹ˆ์˜จ ํ›ˆ๋ จ ๋ชจ๋ธ๋ณด๋‹ค ๋‚ฎ์€ \(\mathcal{D}_{1}\) ๊ฒ€์ฆ ์†์‹ค์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐ•ํ•œ ๋ถ„ํฌ ์ด๋™์ด LLM์˜ ๋งฅ๋ฝ์—์„œ ์ ์‘์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ๋ง๊ฐ์„ ์•…ํ™”์‹œํ‚ค๊ธฐ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ๊ฐ€์„ค์„ ์„ธ์› ์Šต๋‹ˆ๋‹ค. ์žฌ๊ฐ€์—ด๊ณผ ์žฌ๊ฐ์‡ ๋ฅผ ํ†ตํ•ด ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ์œ ๋‹ˆ์˜จ ๊ธฐ์ค€์„ ๊ณผ ๋น„๊ตํ•  ๋•Œ, ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์ด ์œ ๋‹ˆ์˜จ ๊ธฐ์ค€์„ ๋ณด๋‹ค \(\mathcal{D}_{1}\)์— ๋” ์ž˜ ์ ์‘(์ตœ์ข… ๊ฒ€์ฆ ์†์‹ค์ด ๋‚ฎ์Œ)ํ•˜๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ \(\mathcal{D}_{0}\)์—์„œ ์ƒ๋‹นํ•œ ๋ง๊ฐ์„ ๊ฒฝํ—˜ํ•˜๋ฉฐ, ์ด๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์ด ์œ ๋‹ˆ์˜จ ๊ธฐ์ค€์„ ๊ณผ ๊ฒฝ์Ÿ๋ ฅ์„ ๊ฐ–์ถ”๊ธฐ ์œ„ํ•ด ๋ฆฌํ”Œ๋ ˆ์ด๊ฐ€ ํ•„์š”ํ•จ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
Fertility, the most common metric to evaluate a tokenizer's performance (Scao et al., 2022; Stollenwerk, 2023; Rust et al., 2021), is defined as the average number of tokens that are required to represent a word or document. For a tokenizer \(T\) and dataset \(A\), the fertility can be calculated as the number of tokens in \(A\) (when \(T\) is applied) divided by the number of words in \(A\). We calculate the fertility on a held-out set (10,000 documents), which was not used for the tokenizer training. For calculating the words of a document, we used whitespace splitting. Higher fertility scores correspond to weaker compression capabilities of the tokenizer.
ํ† ํฐ๋‚˜์ด์ €์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ๋ฉ”ํŠธ๋ฆญ์ธ Fertility(Scao et al., 2022; Stollenwerk, 2023; Rust et al., 2021)๋Š” ๋‹จ์–ด ๋˜๋Š” ๋ฌธ์„œ๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ๋ฐ ํ•„์š”ํ•œ ํ† ํฐ์˜ ํ‰๊ท  ์ˆ˜๋กœ ์ •์˜๋œ๋‹ค. ํ† ํฐ๋‚˜์ด์ € \(T\)์™€ ๋ฐ์ดํ„ฐ์„ธํŠธ \(A\)์˜ ๊ฒฝ์šฐ ๋น„์˜ฅ๋„๋Š” \(A\)์˜ ํ† ํฐ ์ˆ˜( \(T\)๋ฅผ ์ ์šฉํ•  ๋•Œ)๋ฅผ \(A\)์˜ ๋‹จ์–ด ์ˆ˜๋กœ ๋‚˜๋ˆˆ ๊ฐ’์œผ๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ† ๋‚˜์ด์ € ๊ต์œก์— ์‚ฌ์šฉ๋˜์ง€ ์•Š์€ ๋ณด๋ฅ˜๋œ ์„ธํŠธ(10,000๊ฐœ ๋ฌธ์„œ)์—์„œ ์ƒ์‹๋ ฅ์„ ๊ณ„์‚ฐํ•œ๋‹ค. ๋ฌธ์„œ์˜ ๋‹จ์–ด๋ฅผ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด ์šฐ๋ฆฌ๋Š” ๊ณต๋ฐฑ ๋ถ„ํ• ์„ ์‚ฌ์šฉํ–ˆ๋‹ค. ๋” ๋†’์€ ๋น„์˜ฅ๋„ ์ ์ˆ˜๋Š” ํ† ํฐํ™”๊ธฐ์˜ ๋” ์•ฝํ•œ ์••์ถ• ๋Šฅ๋ ฅ์— ํ•ด๋‹นํ•œ๋‹ค.
ํ† ํฐํ™” ๋„๊ตฌ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ธ ์ง€ํ‘œ์ธ ๋‹ค์‚ฐ์„ฑ์€ ๋‹จ์–ด๋‚˜ ๋ฌธ์„œ๋ฅผ ํ‘œํ˜„ํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ํ† ํฐ์˜ ํ‰๊ท  ์ˆ˜๋กœ ์ •์˜๋ฉ๋‹ˆ๋‹ค(Scao et al., 2022; Stollenwerk, 2023; Rust et al., 2021). ํ† ํฐํ™”๊ธฐ \(T\)์™€ ๋ฐ์ดํ„ฐ์…‹ \(A\)์˜ ๊ฒฝ์šฐ, ๋น„์˜ฅ๋„๋Š” \(A\)์˜ ํ† ํฐ ์ˆ˜(\(T\)๊ฐ€ ์ ์šฉ๋œ ๊ฒฝ์šฐ)๋ฅผ \(A\)์˜ ๋‹จ์–ด ์ˆ˜๋กœ ๋‚˜๋ˆˆ ๊ฐ’์œผ๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ† ํฐํ™” ํ•™์Šต์— ์‚ฌ์šฉ๋˜์ง€ ์•Š์€ ํ™€๋“œ์•„์›ƒ ์„ธํŠธ(10,000๊ฐœ์˜ ๋ฌธ์„œ)์—์„œ ๋น„์˜ฅ๋„๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ์„œ์˜ ๋‹จ์–ด ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด ๊ณต๋ฐฑ ๋ถ„ํ• ์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋น„์˜ฅ๋„ ์ ์ˆ˜๊ฐ€ ๋†’์„์ˆ˜๋ก ํ† ํฐํ™” ๋„๊ตฌ์˜ ์••์ถ• ๊ธฐ๋Šฅ์ด ์•ฝํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
**Overall Comparison with RL Formulation**. Following the discussion in Section 5.2.3 (the part related to RL training), the text generation problem can be formulated as a decision-making process based on RL. Taking a prompt as input, the task of a LLM is to generate a text completion that appropriately responds to the prompt. This task would be completed step by step. At each step, an agent (_i.e.,_ LLM) will perform an action (_i.e.,_ generating a token) according to the policy (_i.e.,_ the generative probability distribution of LLM) conditioned on the current state (currently generated token sequence and other available context information). It is expected that a high-quality output text would be produced by the LLM, which can earn a large reward score based on the entire response. Overall, RLHF and SFT can be considered as two different training approaches to optimizing the above decision making process for LLMs. Specially, RLHF firstly learns the reward model, and then employs it to improve the LLM with RL training (_e.g.,_ PPO). As a comparison, SFT adopts a teacher-forcing approach, which directly optimizes the likelihood of a demonstration output. Such a token-level training way essentially does _behavior cloning_ (a special algorithm of imitation learning [393]): it utilizes the expert's action (_i.e.,_ the target token at each step) as the supervision label and directly learns to imitate the demonstrations from experts without specifying a reward model as in typical RL algorithms. To learn the desired policies, SFT adopts a "local" optimization way (_i.e.,_ token-level loss) based on demonstration data, while RLHF takes a "global" optimization way (_i.e.,_ text-level loss) by involving human preference. More theoretical analysis about imitation learning and reinforcement learning can be referred to the related RL literature [393, 394].
**RL ํ˜•์‹๊ณผ์˜ ์ „๋ฐ˜์ ์ธ ๋น„๊ต**. 5.2.3์ ˆ(RL ํ›ˆ๋ จ๊ณผ ๊ด€๋ จ๋œ ๋ถ€๋ถ„)์˜ ๋…ผ์˜์— ์ด์–ด ํ…์ŠคํŠธ ์ƒ์„ฑ ๋ฌธ์ œ๋ฅผ RL์— ๊ธฐ๋ฐ˜ํ•œ ์˜์‚ฌ๊ฒฐ์ • ๊ณผ์ •์œผ๋กœ ๊ณต์‹ํ™”ํ•  ์ˆ˜ ์žˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ํ•˜์—ฌ, LLM์˜ ํƒœ์Šคํฌ๋Š” ํ”„๋กฌํ”„ํŠธ์— ์ ์ ˆํ•˜๊ฒŒ ์‘๋‹ตํ•˜๋Š” ํ…์ŠคํŠธ ์™„์„ฑ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ด๋‹ค. ์ด ์ž‘์—…์€ ๋‹จ๊ณ„๋ณ„๋กœ ์™„๋ฃŒ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ฐ๊ฐ์˜ ๋‹จ๊ณ„์—์„œ, ์—์ด์ „ํŠธ(_i.e.,_LLM)๋Š” ํ˜„์žฌ ์ƒํƒœ(ํ˜„์žฌ ์ƒ์„ฑ๋œ ํ† ํฐ ์‹œํ€€์Šค ๋ฐ ๋‹ค๋ฅธ ์ด์šฉ๊ฐ€๋Šฅํ•œ ์ปจํ…์ŠคํŠธ ์ •๋ณด)์— ์กฐ๊ฑดํ™”๋œ ์ •์ฑ…(_i.e.,_LM์˜ ์ƒ์„ฑ ํ™•๋ฅ  ๋ถ„ํฌ)์— ๋”ฐ๋ผ ์•ก์…˜(_i.e.,_ํ† ํฐ ์ƒ์„ฑ)์„ ์ˆ˜ํ–‰ํ•  ๊ฒƒ์ด๋‹ค. ์ „์ฒด ์‘๋‹ต์„ ๋ฐ”ํƒ•์œผ๋กœ ํฐ ๋ณด์ƒ ์ ์ˆ˜๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๋Š” LLM์— ์˜ํ•ด ๊ณ ํ’ˆ์งˆ์˜ ์‚ฐ์ถœ ํ…์ŠคํŠธ๊ฐ€ ์ƒ์‚ฐ๋  ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋œ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ, RLHF์™€ SFT๋Š” LLMs์— ๋Œ€ํ•œ ์œ„์˜ ์˜์‚ฌ ๊ฒฐ์ • ํ”„๋กœ์„ธ์Šค๋ฅผ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ๋‘ ๊ฐ€์ง€ ๋‹ค๋ฅธ ํ›ˆ๋ จ ์ ‘๊ทผ๋ฒ•์œผ๋กœ ๊ฐ„์ฃผ๋  ์ˆ˜ ์žˆ๋‹ค. ํŠนํžˆ, RLHF๋Š” ๋ณด์ƒ ๋ชจ๋ธ์„ ๋จผ์ € ํ•™์Šตํ•œ ํ›„, ์ด๋ฅผ ์ด์šฉํ•˜์—ฌ RL ํŠธ๋ ˆ์ด๋‹(์˜ˆ๋ฅผ ๋“ค์–ด, PPO)์„ ํ†ตํ•ด LLM์„ ๊ฐœ์„ ํ•œ๋‹ค. ๋น„๊ต๋กœ SFT๋Š” ๊ต์‚ฌ ๊ฐ•์ œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜์—ฌ ์‹œ์—ฐ ์ถœ๋ ฅ์˜ ๊ฐ€๋Šฅ์„ฑ์„ ์ง์ ‘ ์ตœ์ ํ™”ํ•œ๋‹ค. ์ด๋Ÿฌํ•œ ํ† ํฐ ์ˆ˜์ค€์˜ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์€ ๋ณธ์งˆ์ ์œผ๋กœ _ํ–‰๋™ ๋ณต์ œ_(๋ชจ๋ฐฉ ํ•™์Šต์˜ ํŠน๋ณ„ํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜[393])๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ „๋ฌธ๊ฐ€์˜ ํ–‰๋™(์ฆ‰, ๊ฐ ๋‹จ๊ณ„์—์„œ ๋Œ€์ƒ ํ† ํฐ)์„ ๊ฐ๋… ๋ ˆ์ด๋ธ”๋กœ ํ™œ์šฉํ•˜๊ณ  ์ผ๋ฐ˜์ ์ธ RL ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ๊ฐ™์ด ๋ณด์ƒ ๋ชจ๋ธ์„ ์ง€์ •ํ•˜์ง€ ์•Š๊ณ  ์ „๋ฌธ๊ฐ€๋กœ๋ถ€ํ„ฐ ๋ฐ๋ชจ๋ฅผ ๋ชจ๋ฐฉํ•˜๋„๋ก ์ง์ ‘ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์›ํ•˜๋Š” ์ •์ฑ…์„ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด SFT๋Š” ๋ฐ๋ชจ ๋ฐ์ดํ„ฐ์— ๊ธฐ์ดˆํ•œ "๋กœ์ปฌ" ์ตœ์ ํ™” ๋ฐฉ์‹(_์ฆ‰,_ํ† ํฐ-๋ ˆ๋ฒจ ์†์‹ค)์„ ์ฑ„ํƒํ•˜๋Š” ๋ฐ˜๋ฉด, RLHF๋Š” ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„๋ฅผ ์ˆ˜๋ฐ˜ํ•จ์œผ๋กœ์จ "๊ธ€๋กœ๋ฒŒ" ์ตœ์ ํ™” ๋ฐฉ์‹(_์ฆ‰,_ํ…์ŠคํŠธ-๋ ˆ๋ฒจ ์†์‹ค)์„ ์ทจํ•œ๋‹ค. ๋ชจ๋ฐฉ ํ•™์Šต ๋ฐ ๊ฐ•ํ™” ํ•™์Šต์— ๋Œ€ํ•œ ๋ณด๋‹ค ์ด๋ก ์  ๋ถ„์„์€ ๊ด€๋ จ RL ๋ฌธํ—Œ[393, 394]์„ ์ฐธ์กฐํ•  ์ˆ˜ ์žˆ๋‹ค.
**RL ๊ณต์‹ํ™”์™€์˜ ์ „๋ฐ˜์ ์ธ ๋น„๊ต**. 5.2.3์ ˆ(RL ํ›ˆ๋ จ๊ณผ ๊ด€๋ จ๋œ ๋ถ€๋ถ„)์˜ ๋…ผ์˜์— ๋”ฐ๋ผ ํ…์ŠคํŠธ ์ƒ์„ฑ ๋ฌธ์ œ๋Š” RL์— ๊ธฐ๋ฐ˜ํ•œ ์˜์‚ฌ ๊ฒฐ์ • ๊ณผ์ •์œผ๋กœ ๊ณต์‹ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ๋ฐ›์•„ ํ”„๋กฌํ”„ํŠธ์— ์ ์ ˆํ•˜๊ฒŒ ์‘๋‹ตํ•˜๋Š” ํ…์ŠคํŠธ ์™„์„ฑ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ด LLM์˜ ์ž„๋ฌด์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ๋‹จ๊ณ„๋ณ„๋กœ ์™„๋ฃŒ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ๋‹จ๊ณ„์—์„œ ์—์ด์ „ํŠธ(์ฆ‰, LLM)๋Š” ํ˜„์žฌ ์ƒํƒœ(ํ˜„์žฌ ์ƒ์„ฑ๋œ ํ† ํฐ ์‹œํ€€์Šค ๋ฐ ๊ธฐํƒ€ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ปจํ…์ŠคํŠธ ์ •๋ณด)๋ฅผ ์กฐ๊ฑด์œผ๋กœ ํ•˜๋Š” ์ •์ฑ…(์ฆ‰, LLM์˜ ์ƒ์„ฑ ํ™•๋ฅ  ๋ถ„ํฌ)์— ๋”ฐ๋ผ ์ž‘์—…(์ฆ‰, ํ† ํฐ ์ƒ์„ฑ)์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ์‘๋‹ต์„ ๊ธฐ์ค€์œผ๋กœ ๋†’์€ ๋ณด์ƒ ์ ์ˆ˜๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ๊ณ ํ’ˆ์งˆ์˜ ์ถœ๋ ฅ ํ…์ŠคํŠธ๊ฐ€ LLM์— ์˜ํ•ด ์ƒ์„ฑ๋  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ RLHF์™€ SFT๋Š” LLM์„ ์œ„ํ•œ ์œ„์˜ ์˜์‚ฌ ๊ฒฐ์ • ๊ณผ์ •์„ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ๋‘ ๊ฐ€์ง€ ํ›ˆ๋ จ ์ ‘๊ทผ๋ฒ•์œผ๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ RLHF๋Š” ๋จผ์ € ๋ณด์ƒ ๋ชจ๋ธ์„ ํ•™์Šตํ•œ ๋‹ค์Œ, ์ด๋ฅผ RL ํ›ˆ๋ จ(_์˜ˆ: PPO)์„ ํ†ตํ•ด LLM์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด์— ๋น„ํ•ด SFT๋Š” ๊ต์‚ฌ ๊ฐ•์ œ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜์—ฌ ๋ฐ๋ชจ ๊ฒฐ๊ณผ์˜ ๊ฐ€๋Šฅ์„ฑ์„ ์ง์ ‘ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ† ํฐ ์ˆ˜์ค€์˜ ํ›ˆ๋ จ ๋ฐฉ์‹์€ ๋ณธ์งˆ์ ์œผ๋กœ _ํ–‰๋™ ๋ณต์ œ_(๋ชจ๋ฐฉ ํ•™์Šต์˜ ํŠน์ˆ˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜ [393])๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค: ์ „๋ฌธ๊ฐ€์˜ ํ–‰๋™(์ฆ‰, ๊ฐ ๋‹จ๊ณ„์˜ ๋ชฉํ‘œ ํ† ํฐ)์„ ๊ฐ๋… ๋ผ๋ฒจ๋กœ ํ™œ์šฉํ•˜๊ณ  ์ผ๋ฐ˜์ ์ธ RL ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ์ฒ˜๋Ÿผ ๋ณด์ƒ ๋ชจ๋ธ์„ ์ง€์ •ํ•˜์ง€ ์•Š๊ณ  ์ „๋ฌธ๊ฐ€์˜ ๋ฐ๋ชจ๋ฅผ ์ง์ ‘ ๋ชจ๋ฐฉํ•˜๋Š” ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์›ํ•˜๋Š” ์ •์ฑ…์„ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด SFT๋Š” ๋ฐ๋ชจ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” '๋กœ์ปฌ' ์ตœ์ ํ™” ๋ฐฉ์‹(_ํ† ํฐ ์ˆ˜์ค€ ์†์‹ค)์„ ์ฑ„ํƒํ•˜๋Š” ๋ฐ˜๋ฉด, RLHF๋Š” ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„๋ฅผ ๊ฐœ์ž…์‹œ์ผœ '๊ธ€๋กœ๋ฒŒ' ์ตœ์ ํ™” ๋ฐฉ์‹(_ํ…์ŠคํŠธ ์ˆ˜์ค€ ์†์‹ค)์„ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ฐฉ ํ•™์Šต๊ณผ ๊ฐ•ํ™” ํ•™์Šต์— ๋Œ€ํ•œ ๋ณด๋‹ค ์ž์„ธํ•œ ์ด๋ก ์  ๋ถ„์„์€ ๊ด€๋ จ RL ๋ฌธํ—Œ [393, 394]์„ ์ฐธ๊ณ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The most primitive chain-of-thought is a chain structure that describes intermediate reasoning steps in natural language. In this section, we introduce structural variants that modify the original chain structure, including chain structure variants, tree structure variants, and graph structure variants.
๊ฐ€์žฅ ์›์‹œ์ ์ธ ์—ฐ์‡„์  ์‚ฌ๊ณ ๋Š” ์ž์—ฐ์–ด๋กœ ์ค‘๊ฐ„์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๊ธฐ์ˆ ํ•˜๋Š” ์—ฐ์‡„ ๊ตฌ์กฐ์ด๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ฒด์ธ ๊ตฌ์กฐ ๋ณ€ํ˜•, ํŠธ๋ฆฌ ๊ตฌ์กฐ ๋ณ€ํ˜• ๋ฐ ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ ๋ณ€ํ˜•์„ ํฌํ•จํ•˜์—ฌ ์›๋ž˜ ์ฒด์ธ ๊ตฌ์กฐ๋ฅผ ์ˆ˜์ •ํ•˜๋Š” ๊ตฌ์กฐ ๋ณ€ํ˜•์„ ์†Œ๊ฐœํ•œ๋‹ค.
๊ฐ€์žฅ ์›์‹œ์ ์ธ ์—ฐ์‡„ ์‚ฌ๊ณ  ๊ตฌ์กฐ๋Š” ์ž์—ฐ์–ด๋กœ ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์„ค๋ช…ํ•˜๋Š” ์—ฐ์‡„ ๊ตฌ์กฐ์ž…๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์—ฐ์‡„ ๊ตฌ์กฐ ๋ณ€ํ˜•, ํŠธ๋ฆฌ ๊ตฌ์กฐ ๋ณ€ํ˜•, ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ ๋ณ€ํ˜• ๋“ฑ ์›๋ž˜์˜ ์—ฐ์‡„ ๊ตฌ์กฐ๋ฅผ ์ˆ˜์ •ํ•˜๋Š” ๊ตฌ์กฐ์  ๋ณ€ํ˜•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.
We collect the problems from various sources e.g. ziya (Zhang et al., 2022), goakao, and CIF-Bench (Li et al., 2024) to form hard-case Chinese instructions understanding and following evaluation benchmark (CHC-Bench in short) The categories of problems in CHC-Bench include writing, humanity and history, science, math, reading comprehension, role-playing, and hard cases of Chinese understanding (i.e. Chinese word pronunciation, ancient Chinese language understanding, etc.).
์ด๋ฅผ ์œ„ํ•ด ์ง€์•ผ(Zhang et al., 2022), ๊ณ ์นด์นด์˜ค(Goakao), CIF-Bench(Li et al., 2024) ๋“ฑ ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์—์„œ ๋ฌธ์ œ๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ ์–ด๋ ค์šด ์ค‘๊ตญ์–ด ๋ช…๋ น์–ด ์ดํ•ด์™€ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ(CHC-Bench๋ฅผ ์งง๊ฒŒ ํ‘œํ˜„ํ•จ)๋ฅผ ๊ตฌ์„ฑํ•œ๋‹ค. CHC-Bench์˜ ๋ฌธ์ œ ๋ฒ”์ฃผ์—๋Š” ๊ธ€์“ฐ๊ธฐ, ์ธ๋ฅ˜ ๋ฐ ์—ญ์‚ฌ, ๊ณผํ•™, ์ˆ˜ํ•™, ๋…ํ•ด, ์—ญํ• ๋†€์ด, ์ค‘๊ตญ์–ด ์ดํ•ด์˜ ์–ด๋ ค์šด ์‚ฌ๋ก€(์ฆ‰, ์ค‘๊ตญ์–ด ๋ฐœ์Œ, ๊ณ ๋Œ€ ์ค‘๊ตญ์–ด ์ดํ•ด ๋“ฑ)๊ฐ€ ํฌํ•จ๋œ๋‹ค.
๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์˜ ๋ฌธ์ œ(์˜ˆ: ziya(Zhang et al., 2022), goakao, CIF-Bench(Li et al., 2024))๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ ์ค‘๊ตญ์–ด ์ง€์‹œ๋ฌธ ์ดํ•ด ๋ฐ ์ถ”๋ก  ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ(์ค„์—ฌ์„œ CHC-Bench)๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. CHC-Bench์˜ ๋ฌธ์ œ ๋ฒ”์ฃผ์—๋Š” ์“ฐ๊ธฐ, ์ธ์„ฑ๊ณผ ์—ญ์‚ฌ, ๊ณผํ•™, ์ˆ˜ํ•™, ๋…ํ•ด, ์—ญํ• ๊ทน, ์–ด๋ ค์šด ์ค‘๊ตญ์–ด ์ดํ•ด ์‚ฌ๋ก€(์ค‘๊ตญ์–ด ๋‹จ์–ด ๋ฐœ์Œ, ๊ณ ๋Œ€ ์ค‘๊ตญ์–ด ์ดํ•ด ๋“ฑ) ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 2: Example summary generated by Claude-3-Opus and claims extracted by GPT-4.
๊ทธ๋ฆผ 2: Claude-3-Opus์— ์˜ํ•ด ์ƒ์„ฑ๋œ ์˜ˆ์ œ ์š”์•ฝ ๋ฐ GPT-4์— ์˜ํ•ด ์ถ”์ถœ๋œ ํด๋ ˆ์ž„.
๊ทธ๋ฆผ 2: Claude-3-Opus์—์„œ ์ƒ์„ฑํ•œ ์š”์•ฝ ์˜ˆ์‹œ ๋ฐ GPT-4์—์„œ ์ถ”์ถœํ•œ ํด๋ ˆ์ž„.
Large Language Models (LLMs), such as GPT-3 Brown et al. (2020), LLaMA Touvron et al. (2023), and PaLM Chowdhery et al. (2023), have demonstrated remarkable capabilities as general-purpose assistants. The cornerstone of this achievement is instruction tuning, which significantly enhances the capabilities and controllability of LLMs through training on datasets composed of instruction-output pairs Zhang et al. (2023). This technique effectively aligns the models' training objectives with human intentions, thereby ensuring that the models can interpret and execute human instructions both effectively and safely. Therefore, the availability of high-quality instruction tuning datasets is crucial for LLMs to operate as efficient and dependable assistants.
GPT-3 Brown et al. (2020), LLaMA Touvron et al. (2023), PaLM Chowdhery et al. (2023)๊ณผ ๊ฐ™์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(Large Language Models, LLM)์€ ๋ฒ”์šฉ ๋ณด์กฐ์ž๋กœ์„œ์˜ ๋†€๋ผ์šด ๋Šฅ๋ ฅ์„ ์ž…์ฆํ–ˆ๋‹ค. ์ด ์„ฑ์ทจ์˜ ์ดˆ์„์€ ๋ช…๋ น์–ด ํŠœ๋‹์ด๋ฉฐ, ์ด๋Š” ๋ช…๋ น์–ด-์ถœ๋ ฅ ์Œ Zhang ๋“ฑ(2023)์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์„ ํ†ตํ•ด LLM์˜ ๋Šฅ๋ ฅ๊ณผ ์ œ์–ด ๊ฐ€๋Šฅ์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค. ์ด ๊ธฐ์ˆ ์€ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ ๋ชฉํ‘œ๋ฅผ ์ธ๊ฐ„์˜ ์˜๋„์™€ ํšจ๊ณผ์ ์œผ๋กœ ์ •๋ ฌํ•˜์—ฌ ๋ชจ๋ธ์ด ์ธ๊ฐ„ ๋ช…๋ น์„ ํšจ๊ณผ์ ์ด๊ณ  ์•ˆ์ „ํ•˜๊ฒŒ ํ•ด์„ํ•˜๊ณ  ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•œ๋‹ค. ๋”ฐ๋ผ์„œ LLM์ด ํšจ์œจ์ ์ด๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ณด์กฐ์ž๋กœ ์ž‘๋™ํ•˜๋ ค๋ฉด ๊ณ ํ’ˆ์งˆ ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ€์šฉ์„ฑ์ด ์ค‘์š”ํ•˜๋‹ค.
GPT-3 Brown ์™ธ(2020), LLaMA Touvron ์™ธ(2023), PaLM Chowdhery ์™ธ(2023)์™€ ๊ฐ™์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ๋ฒ”์šฉ ์–ด์‹œ์Šคํ„ดํŠธ๋กœ์„œ ๋†€๋ผ์šด ์—ญ๋Ÿ‰์„ ์ž…์ฆํ•ด ์™”์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ฑ๊ณผ์˜ ์ดˆ์„์€ ๋ช…๋ น์–ด ํŠœ๋‹์œผ๋กœ, ๋ช…๋ น์–ด-์ถœ๋ ฅ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์„ ํ†ตํ•ด LLM์˜ ๊ธฐ๋Šฅ๊ณผ ์ œ์–ด ๊ฐ€๋Šฅ์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค(Zhang et al., 2023). ์ด ๊ธฐ์ˆ ์€ ๋ชจ๋ธ์˜ ํ•™์Šต ๋ชฉํ‘œ๋ฅผ ์ธ๊ฐ„์˜ ์˜๋„์™€ ํšจ๊ณผ์ ์œผ๋กœ ์ผ์น˜์‹œ์ผœ ๋ชจ๋ธ์ด ์ธ๊ฐ„์˜ ์ง€์‹œ๋ฅผ ํšจ๊ณผ์ ์ด๊ณ  ์•ˆ์ „ํ•˜๊ฒŒ ํ•ด์„ํ•˜๊ณ  ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ LLM์ด ํšจ์œจ์ ์ด๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋น„์„œ๋กœ ์ž‘๋™ํ•˜๋ ค๋ฉด ๊ณ ํ’ˆ์งˆ์˜ ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ์ดํ„ฐ์„ธํŠธ์˜ ๊ฐ€์šฉ์„ฑ์ด ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
Figure 17: The fitted scaling law plotted along the data amount axis for each model and each task. The fitted result is satisfying except the last checkpoints of 0.11B and 0.25B model.
๊ทธ๋ฆผ 17: ์ ํ•ฉ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์€ ๊ฐ ๋ชจ๋ธ ๋ฐ ๊ฐ ์ž‘์—…์— ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ ์–‘ ์ถ•์„ ๋”ฐ๋ผ ํ‘œ์‹œ๋œ๋‹ค. ์ ํ•ฉ๋œ ๊ฒฐ๊ณผ๋Š” 0.11B ๋ฐ 0.25B ๋ชจ๋ธ์˜ ๋งˆ์ง€๋ง‰ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ์ œ์™ธํ•˜๊ณ  ๋งŒ์กฑํ•œ๋‹ค.
๊ทธ๋ฆผ 17: ๊ฐ ๋ชจ๋ธ๊ณผ ๊ฐ ์ž‘์—…์— ๋Œ€ํ•ด ๋ฐ์ดํ„ฐ ์–‘ ์ถ•์„ ๋”ฐ๋ผ ๊ทธ๋ ค์ง„ ์ ํ•ฉ๋œ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™. ๋งˆ์ง€๋ง‰ ์ฒดํฌํฌ์ธํŠธ์ธ 0.11B์™€ 0.25B ๋ชจ๋ธ์„ ์ œ์™ธํ•˜๊ณ ๋Š” ๋งŒ์กฑ์Šค๋Ÿฌ์šด ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In this case it is easy to guess that the third option is the most likely, even without knowledge of this specific strike.
์ด ๊ฒฝ์šฐ ์ด ํŠน์ • ํŒŒ์—…์— ๋Œ€ํ•œ ์ง€์‹์ด ์—†๋”๋ผ๋„ ์„ธ ๋ฒˆ์งธ ์˜ต์…˜์ด ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค๊ณ  ์ถ”์ธกํ•˜๊ธฐ ์‰ฝ๋‹ค.
์ด ๊ฒฝ์šฐ ์ด ํŠน์ • ํŒŒ์—…์— ๋Œ€ํ•œ ์ง€์‹์ด ์—†์–ด๋„ ์„ธ ๋ฒˆ์งธ ์˜ต์…˜์ด ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค๊ณ  ์‰ฝ๊ฒŒ ์ถ”์ธกํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
et al., 2023; Bai et al., 2023; Almazrouei et al., 2023), which gradually decreases the learning rate following a cosine curve after it reaches its maximum after the warmup stage.
et al., 2023; Bai et al., 2023; Almazrouei et al., 2023)์ด๊ณ , ์ด๋Š” ์›Œ๋ฐ์—… ๋‹จ๊ณ„ ์ดํ›„ ์ตœ๋Œ€์น˜์— ๋„๋‹ฌํ•œ ํ›„ ์ฝ”์‚ฌ์ธ ๊ณก์„ ์— ๋”ฐ๋ผ ํ•™์Šต๋ฅ ์„ ์ ์ง„์ ์œผ๋กœ ๊ฐ์†Œ์‹œํ‚ค๋Š” ๊ฒƒ์„ ํŠน์ง•์œผ๋กœ ํ•˜๋Š” ํ•™์Šต ๋ฐฉ๋ฒ•.
์™ธ, 2023; Bai ์™ธ, 2023; Almazrouei ์™ธ, 2023), ์›Œ๋ฐ์—… ๋‹จ๊ณ„ ์ดํ›„ ์ตœ๋Œ€์น˜์— ๋„๋‹ฌํ•œ ํ›„ ์ฝ”์‚ฌ์ธ ๊ณก์„ ์„ ๋”ฐ๋ผ ํ•™์Šต ์†๋„๊ฐ€ ์ ์ฐจ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค.
Although numerous research about instruction tuning has been conducted, a few work focused on improving RAG or context awareness generation for QA. Lin et al. (2023) introduced a retrieval-augmented instruction tuning method, which appends top-\(k\) retrieved chunks for LLM fine-tuning. Wang et al. (2023) applied instruction tuning after retrieval-augmented pretraining. In contrast, we propose a two-stage instruction tuning method to improve generation with retrival or provided context. We find that appending top-\(k\) retrieved chunks for LLM fine-tuning does not help for a wide range of conversation QA tasks (see SS6.3 for details).
์ˆ˜์—… ํŠœ๋‹์— ๋Œ€ํ•œ ๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ์ˆ˜ํ–‰๋˜์—ˆ์ง€๋งŒ, QA๋ฅผ ์œ„ํ•œ RAG ๋˜๋Š” ์ƒํ™ฉ ์ธ์‹ ์ƒ์„ฑ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถ˜ ์—ฐ๊ตฌ๋Š” ๋ช‡ ๊ฐ€์ง€์ด๋‹ค. Lin et al. (2023)์€ LLM ๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•ด ์ƒ์œ„\(k\)๊ฐœ์˜ ๊ฒ€์ƒ‰๋œ ์ฒญํฌ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒ€์ƒ‰ ๊ฐ•ํ™” ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ–ˆ๋‹ค. Wang et al. (2023)์€ ๊ฒ€์ƒ‰ ๊ฐ•ํ™” ์‚ฌ์ „ ํ›ˆ๋ จ ํ›„์— ๋ช…๋ น์–ด ํŠœ๋‹์„ ์ ์šฉํ•˜์˜€๋‹ค. ์ด์— ๋ฐ˜ํ•ด ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๋ฆฌํŠธ๋ฆฌ๋ฒŒ(retrival) ๋˜๋Š” ์ œ๊ณต๋œ ์ปจํ…์ŠคํŠธ(context)๋ฅผ ๊ฐ–๋Š” ์ƒ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ 2๋‹จ๊ณ„ ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•œ๋‹ค. ์šฐ๋ฆฌ๋Š” LLM ๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•ด ์ƒ์œ„\(k\) ๊ฒ€์ƒ‰๋œ ์ฒญํฌ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ์ด ๊ด‘๋ฒ”์œ„ํ•œ ๋Œ€ํ™” QA ์ž‘์—…์— ๋„์›€์ด ๋˜์ง€ ์•Š๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋‹ค(์ž์„ธํ•œ ๋‚ด์šฉ์€ SS6.3 ์ฐธ์กฐ).
์ธ์ŠคํŠธ๋Ÿญ์…˜ ํŠœ๋‹์— ๋Œ€ํ•œ ์ˆ˜๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ์ˆ˜ํ–‰๋˜์—ˆ์ง€๋งŒ, QA๋ฅผ ์œ„ํ•œ RAG ๋˜๋Š” ๋ฌธ๋งฅ ์ธ์‹ ์ƒ์„ฑ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถ˜ ์—ฐ๊ตฌ๋Š” ๊ฑฐ์˜ ์—†์—ˆ์Šต๋‹ˆ๋‹ค. Lin ๋“ฑ(2023)์€ LLM ๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•ด ๊ฒ€์ƒ‰๋œ ์ƒ์œ„ \(k\) ์ฒญํฌ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. Wang ๋“ฑ(2023)์€ ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์‚ฌ์ „ ํ›ˆ๋ จ ํ›„ ์ธ์ŠคํŠธ๋Ÿญ์…˜ ํŠœ๋‹์„ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด์™€ ๋Œ€์กฐ์ ์œผ๋กœ, ์ €ํฌ๋Š” ์žฌ๊ฒ€์ƒ‰ ๋˜๋Š” ์ปจํ…์ŠคํŠธ ์ œ๊ณต์„ ํ†ตํ•ด ์ƒ์„ฑ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ 2๋‹จ๊ณ„ ์ธ์ŠคํŠธ๋Ÿญ์…˜ ํŠœ๋‹ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. LLM ๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•ด ๊ฒ€์ƒ‰๋œ ์ƒ์œ„ \(k\) ์ฒญํฌ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ๊ด‘๋ฒ”์œ„ํ•œ ๋Œ€ํ™” QA ์ž‘์—…์— ๋„์›€์ด ๋˜์ง€ ์•Š๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(์ž์„ธํ•œ ๋‚ด์šฉ์€ SS6.3 ์ฐธ์กฐ).
In many jurisdictions, the hazards in our taxonomy are illegal. For instance, child abuse, physical violence, and sexual assault are illegal in most countries across the globe33. However, (i) the hazards are not always illegal and (ii) our policies are not defined solely on the basis of the hazards' legality. This is because:
๋งŽ์€ ๊ด€ํ•  ๊ตฌ์—ญ์—์„œ ์šฐ๋ฆฌ ๋ถ„๋ฅ˜๋ฒ•์˜ ์œ„ํ—˜์€ ๋ถˆ๋ฒ•์ด๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์„ธ๊ณ„ 33๊ฐœ๊ตญ ๋Œ€๋ถ€๋ถ„์˜ ๊ตญ๊ฐ€์—์„œ ์•„๋™ํ•™๋Œ€, ํญ๋ ฅ, ์„ฑํญ๋ ฅ์€ ๋ถˆ๋ฒ•์ด๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ (i) ์œ„ํ—˜์š”์†Œ๊ฐ€ ํ•ญ์ƒ ๋ถˆ๋ฒ•์ธ ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ (ii) ์šฐ๋ฆฌ์˜ ์ •์ฑ…์€ ์œ„ํ—˜์š”์†Œ์˜ ํ•ฉ๋ฒ•์„ฑ๋งŒ์„ ๊ธฐ์ค€์œผ๋กœ ์ •์˜๋˜์ง€ ์•Š๋Š”๋‹ค. ์™œ๋ƒํ•˜๋ฉด,
๋งŽ์€ ๊ด€ํ• ๊ถŒ์—์„œ ๋ถ„๋ฅ˜์— ํฌํ•จ๋œ ์œ„ํ—˜ ์š”์†Œ๋Š” ๋ถˆ๋ฒ•์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ์•„๋™ ํ•™๋Œ€, ์‹ ์ฒด์  ํญ๋ ฅ, ์„ฑํญํ–‰์€ ์ „ ์„ธ๊ณ„ ๋Œ€๋ถ€๋ถ„์˜ ๊ตญ๊ฐ€์—์„œ ๋ถˆ๋ฒ•์ž…๋‹ˆ๋‹ค33. ๊ทธ๋Ÿฌ๋‚˜ (i) ์ด๋Ÿฌํ•œ ์œ„ํ—˜์ด ํ•ญ์ƒ ๋ถˆ๋ฒ•์ธ ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ (ii) ๋‹น์‚ฌ์˜ ์ •์ฑ…์€ ์ด๋Ÿฌํ•œ ์œ„ํ—˜์˜ ํ•ฉ๋ฒ•์„ฑ๋งŒ์„ ๊ธฐ์ค€์œผ๋กœ ์ •์˜๋œ ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ๊ทธ ์ด์œ ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:
Instruction tuning aims to train large language models to generate responses that align with input instructions, thereby enable LLMs with conversational and task execution capabilities. Compared with standard LLM, SFT allows the behavior of the model to be more controllable and predictable, thereby achieving the purpose of aligning human will. Methods to build instruction tuning datasets include: (1) Pure manual annotation Conover et al. (2023). This method completely constructs instructions and answers manually, which is very time-consuming and laborious; (2) Converted from existing datasets Mishra et al. (2022); Sanh et al. (2022); Chung et al. (2022). Some studies use supervised data sets from NLP tasks to construct instruction tuning data; (3) Automatically generated using LLM Honovich et al. (2022); Wang et al. (2023); Xu et al. (2023); Ji et al. (2023); Xu et al. (2023). Others use existing LLMs to generate instruction tuning data. A common practice is to first manually annotate high-quality seed datasets, and then use LLM to expand the seed instructions and corresponding outputs, which can generate large-scale instruction tuning data with very little human annotation. However, the quality cannot be guaranteed and there is a certain amount of noisy data, which can lead to hallucinations.
๋ช…๋ น์–ด ํŠœ๋‹์€ ์ž…๋ ฅ ๋ช…๋ น์–ด์™€ ์ •๋ ฌ๋˜๋Š” ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ํฐ ์–ธ์–ด ๋ชจ๋ธ์„ ํŠธ๋ ˆ์ด๋‹ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋ฉฐ, ์ด์— ์˜ํ•ด ๋Œ€ํ™” ๋ฐ ํƒœ์Šคํฌ ์‹คํ–‰ ๋Šฅ๋ ฅ์„ ๊ฐ–๋Š” LLM์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•œ๋‹ค. ํ‘œ์ค€ LLM๊ณผ ๋น„๊ตํ•˜์—ฌ, SFT๋Š” ๋ชจ๋ธ์˜ ํ–‰๋™์ด ๋” ์ œ์–ด๊ฐ€๋Šฅํ•˜๊ณ  ์˜ˆ์ธก๊ฐ€๋Šฅํ•˜๋„๋ก ํ•˜์—ฌ, ์ธ๊ฐ„์˜ ์˜์ง€๋ฅผ ์ •๋ ฌํ•˜๋Š” ๋ชฉ์ ์„ ๋‹ฌ์„ฑํ•œ๋‹ค. ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ์ดํ„ฐ์„ธํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐฉ๋ฒ•์€ (1) ์ˆœ์ˆ˜ ์ˆ˜๋™ ์ฃผ์„ Conover ๋“ฑ(2023)์„ ํฌํ•จํ•œ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋ช…๋ น์–ด ๋ฐ ๋‹ต๋ณ€์„ ์ˆ˜๋™์œผ๋กœ ์™„์ „ํžˆ ๊ตฌ์„ฑํ•˜๋ฉฐ, ์ด๋Š” ๋งค์šฐ ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ฆฌ๊ณ  ํž˜๋“  ์ž‘์—…์ด๋‹ค; (2) ๊ธฐ์กด ๋ฐ์ดํ„ฐ ์„ธํŠธ Mishra ๋“ฑ(2022), Sanh ๋“ฑ(2022), Chung ๋“ฑ(2022)์œผ๋กœ๋ถ€ํ„ฐ ๋ณ€ํ™˜๋œ๋‹ค. ์ผ๋ถ€ ์—ฐ๊ตฌ๋Š” ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด NLP ํƒœ์Šคํฌ๋กœ๋ถ€ํ„ฐ์˜ ๊ฐ๋…๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค; (3) LLM Honovich ๋“ฑ(2022); Wang ๋“ฑ(2023); Xu ๋“ฑ(2023); Ji ๋“ฑ(2023); Xu ๋“ฑ(2023). ๋‹ค๋ฅธ ์‚ฌ๋žŒ๋“ค์€ ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด์˜ LLM์„ ์‚ฌ์šฉํ•œ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๊ด€ํ–‰์€ ๋จผ์ € ๊ณ ํ’ˆ์งˆ ์‹œ๋“œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ˆ˜๋™์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์•˜๋‹ค๊ฐ€ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹œ๋“œ ๋ช…๋ น ๋ฐ ํ•ด๋‹น ์ถœ๋ ฅ์„ ํ™•์žฅํ•˜์—ฌ ์ธ๊ฐ„ ์ฃผ์„์ด ๊ฑฐ์˜ ์—†๋Š” ๋Œ€๊ทœ๋ชจ ๋ช…๋ น ํŠœ๋‹ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ’ˆ์งˆ์„ ๋ณด์žฅํ•  ์ˆ˜ ์—†๊ณ  ์ผ์ •๋Ÿ‰์˜ ์žก์Œ์ด ์žˆ๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ์žˆ์–ด ํ™˜๊ฐ์„ ์œ ๋ฐœํ•  ์ˆ˜ ์žˆ๋‹ค.
๋ช…๋ น์–ด ํŠœ๋‹์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œ์ผœ ์ž…๋ ฅ ๋ช…๋ น์–ด์™€ ์ผ์น˜ํ•˜๋Š” ์‘๋‹ต์„ ์ƒ์„ฑํ•จ์œผ๋กœ์จ ๋Œ€ํ™” ๋ฐ ์ž‘์—… ์‹คํ–‰ ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ˜ LLM์„ ๊ตฌํ˜„ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ํ‘œ์ค€ LLM์— ๋น„ํ•ด SFT๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋ชจ๋ธ์˜ ๋™์ž‘์„ ๋” ์ž˜ ์ œ์–ดํ•˜๊ณ  ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ธ๊ฐ„์˜ ์˜์ง€์— ๋ถ€ํ•ฉํ•˜๋Š” ๋ชฉ์ ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ์ดํ„ฐ์„ธํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐฉ๋ฒ•์—๋Š” ๋‹ค์Œ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค: (1) ์ˆœ์ˆ˜ ์ˆ˜๋™ ์ฃผ์„ Conover ๋“ฑ(2023). ์ด ๋ฐฉ๋ฒ•์€ ๋ช…๋ น์–ด์™€ ๋‹ต๋ณ€์„ ์™„์ „ํžˆ ์ˆ˜์ž‘์—…์œผ๋กœ ๊ตฌ์ถ•ํ•˜๋ฏ€๋กœ ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ฆฌ๊ณ  ํž˜๋“ค๋ฉฐ, (2) ๊ธฐ์กด ๋ฐ์ดํ„ฐ์…‹์—์„œ ๋ณ€ํ™˜ Mishra ๋“ฑ(2022), Sanh ๋“ฑ(2022), Chung ๋“ฑ(2022). ์ผ๋ถ€ ์—ฐ๊ตฌ๋Š” NLP ์ž‘์—…์˜ ์ง€๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. (3) LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ž๋™์œผ๋กœ ์ƒ์„ฑ Honovich ์™ธ. (2022); Wang ์™ธ. (2023); Xu ์™ธ. (2023); Ji ์™ธ. (2023); Xu ์™ธ. (2023). ๋‹ค๋ฅธ ์ด๋“ค์€ ๊ธฐ์กด LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ฐฉ๋ฒ•์€ ๋จผ์ € ๊ณ ํ’ˆ์งˆ์˜ ์‹œ๋“œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ˆ˜๋™์œผ๋กœ ์ฃผ์„์„ ๋‹จ ๋‹ค์Œ, LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹œ๋“œ ๋ช…๋ น์–ด์™€ ํ•ด๋‹น ์ถœ๋ ฅ์„ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์‚ฌ๋žŒ์˜ ์ฃผ์„์ด ๊ฑฐ์˜ ์—†์ด๋„ ๋Œ€๊ทœ๋ชจ ๋ช…๋ น์–ด ํŠœ๋‹ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ’ˆ์งˆ์„ ๋ณด์žฅํ•  ์ˆ˜ ์—†๊ณ  ์ผ์ •๋Ÿ‰์˜ ๋…ธ์ด์ฆˆ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ฐœ์ƒํ•˜์—ฌ ํ™˜๊ฐ์„ ์œ ๋ฐœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
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.
RAG ํŒจ๋Ÿฌ๋‹ค์ž„์€ ๋„๋ฉ”์ธ ๋‚ด์—์„œ ์—ฐ๊ตฌ๋ฅผ ์กฐ์งํ•˜์—ฌ LLM ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ๊ฐ•๋ ฅํ•œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ๊ณตํ•œ๋‹ค. RAG์˜ ํ•ต์‹ฌ์€ ๊ณ ํ’ˆ์งˆ ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ๊ฒ€์ƒ‰ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด๋‹ค. ๋”ฐ๋ผ์„œ ์ด ํŒจ๋Ÿฌ๋‹ค์ž„์€ ๊ฒ€์ƒ‰์˜ ๊ด€์ ์—์„œ ์‚ฌ์ „ ๊ฒ€์ƒ‰, ๊ฒ€์ƒ‰, ์‚ฌํ›„ ๊ฒ€์ƒ‰, ์ƒ์„ฑ์˜ ๋„ค ๊ฐ€์ง€ ์ฃผ์š” ๋‹จ๊ณ„๋กœ ๊ตฌ์กฐํ™”๋œ๋‹ค. ๋ฐ˜๋ณต์ ์ธ ๊ฒ€์ƒ‰ ์ƒ์„ฑ ์ฃผ๊ธฐ๋ฅผ ํฌํ•จํ•˜๋Š” ๋‹จ์ผ ํ™‰ ๋ฐ ๋‹ค์ค‘ ํ™‰ ๊ฒ€์ƒ‰ ์ ‘๊ทผ๋ฒ• ๋ชจ๋‘ ์ด 4๋‹จ๊ณ„ ๊ตฌ์กฐ๋ฅผ ๋”ฐ๋ฅธ๋‹ค. ๊ทธ๋ฆผ 3์€ RAG์˜ ํ•ต์‹ฌ ๊ธฐ๋ฒ•์˜ ๋ถ„๋ฅ˜ ํŠธ๋ฆฌ์ด๋‹ค.
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 ํ’ˆ์งˆ์˜ ๊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 ๊ฒ€์ƒ‰ ์ „, ๊ฒ€์ƒ‰, ๊ฒ€์ƒ‰ ํ›„, ์ƒ์„ฑ์˜ ๋„ค ๊ฐ€์ง€ ์ฃผ์š” ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ณต์ ์ธ ๊ฒ€์ƒ‰-์ƒ์„ฑ ์ฃผ๊ธฐ๋ฅผ ํฌํ•จํ•˜๋Š” ๋‹จ์ผ ํ™‰ ๋ฐ ๋ฉ€ํ‹ฐ ํ™‰ ๊ฒ€์ƒ‰ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋ชจ๋‘ ์ด 4๋‹จ๊ณ„ ๊ตฌ์กฐ๋ฅผ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3์€ RAG์˜ ํ•ต์‹ฌ ๊ธฐ์ˆ ์„ ๋ถ„๋ฅ˜ํ•œ ํŠธ๋ฆฌ์ž…๋‹ˆ๋‹ค.
**Dealing with Ambiguities:** Resolving or excluding ambiguous or contradictory data that might confuse the model during training. This can help the model to provide more definite and reliable answers.
**๋ชจํ˜ธ์„ฑ ์ฒ˜๋ฆฌ:** ํ›ˆ๋ จ ์ค‘์— ๋ชจ๋ธ์„ ํ˜ผ๋™ํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ๋ชจ์ˆœ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ํ•ด๊ฒฐํ•˜๊ฑฐ๋‚˜ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋ชจ๋ธ์ด ๋ณด๋‹ค ํ™•์‹คํ•˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋‹ค.
**๋ชจํ˜ธ์„ฑ ์ฒ˜๋ฆฌ: **ํ•™์Šต ์ค‘์— ๋ชจ๋ธ์„ ํ˜ผ๋ž€์Šค๋Ÿฝ๊ฒŒ ํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ๋ชจ์ˆœ๋˜๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ํ•ด๊ฒฐํ•˜๊ฑฐ๋‚˜ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์ด ๋ณด๋‹ค ๋ช…ํ™•ํ•˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ต๋ณ€์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 4: **Combining multiple styles:** Perplexity across all domains of the Pile comparing combining multiple styles of synthetic data. Models are 1.3B parameters trained for a total of 150B tokens. We see small perplexity improvements from combining multiple styles.
๊ทธ๋ฆผ 4: **์—ฌ๋Ÿฌ ์Šคํƒ€์ผ ๊ฒฐํ•ฉ:** ์—ฌ๋Ÿฌ ์Šคํƒ€์ผ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฒฐํ•ฉํ•˜๋Š” ๊ฒƒ์„ ๋น„๊ตํ•˜๋Š” ํŒŒ์ผ์˜ ๋ชจ๋“  ๋„๋ฉ”์ธ์— ๊ฑธ์นœ ๋ณต์žก์„ฑ์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ์€ ์ด 150B ํ† ํฐ์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋œ 1.3B ํŒŒ๋ผ๋ฏธํ„ฐ์ด๋‹ค. ์šฐ๋ฆฌ๋Š” ์—ฌ๋Ÿฌ ์Šคํƒ€์ผ์„ ๊ฒฐํ•ฉํ•˜์—ฌ ์ž‘์€ ๋ณต์žก์„ฑ ๊ฐœ์„ ์„ ๋ด…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 4: **์—ฌ๋Ÿฌ ์Šคํƒ€์ผ ๊ฒฐํ•ฉํ•˜๊ธฐ: ** ์—ฌ๋Ÿฌ ์Šคํƒ€์ผ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ๊ฒฐํ•ฉ์„ ๋น„๊ตํ•œ Pile์˜ ๋ชจ๋“  ๋„๋ฉ”์ธ์— ๊ฑธ์นœ ๋ณต์žก์„ฑ. ๋ชจ๋ธ์€ ์ด 150์–ต ๊ฐœ์˜ ํ† ํฐ์— ๋Œ€ํ•ด 13์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ•™์Šต์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์Šคํƒ€์ผ์„ ๊ฒฐํ•ฉํ•˜๋ฉด ๋ณต์žก๋„๊ฐ€ ์•ฝ๊ฐ„ ๊ฐœ์„ ๋˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Third, we did not investigate whether the results obtained could be extrapolated to larger model sizes, which we leave to future works. However, our finding that the BPE-SP-33 tokenizer is the best-performing tokenizer for the monolingual setting and the fact that this tokenizer has been used for training state-of-the-art models up to 65B (Touvron et al.) might indicate that our results also transfer to larger model sizes.
์…‹์งธ, ์–ป์€ ๊ฒฐ๊ณผ๊ฐ€ ๋” ํฐ ๋ชจ๋ธ ํฌ๊ธฐ๋กœ ์™ธ์‚ฝ๋  ์ˆ˜ ์žˆ๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ์กฐ์‚ฌํ•˜์ง€ ์•Š์•˜์œผ๋ฉฐ ์ด๋Š” ํ–ฅํ›„ ์ž‘์—…์— ๋งก๊ธด๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ BPE-SP-33 ํ† ํฐํ™”๊ธฐ๊ฐ€ ๋‹จ์ผ ์–ธ์–ด ์„ค์ •์— ๋Œ€ํ•ด ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ ํ† ํฐํ™”๊ธฐ๋ผ๋Š” ๋ฐœ๊ฒฌ๊ณผ ์ด ํ† ํฐํ™”๊ธฐ๊ฐ€ ์ตœ๋Œ€ 65B(Touvron ๋“ฑ)๊นŒ์ง€ ์ตœ์‹  ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ๋‹ค๋Š” ์‚ฌ์‹ค์€ ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๊ฐ€ ๋” ํฐ ๋ชจ๋ธ ํฌ๊ธฐ๋กœ๋„ ์ „๋‹ฌ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ๋‹ค.
์…‹์งธ, ์–ป์€ ๊ฒฐ๊ณผ๋ฅผ ๋” ํฐ ๋ชจ๋ธ ํฌ๊ธฐ๋กœ ์™ธ์‚ฝํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์—ฌ๋ถ€๋Š” ์กฐ์‚ฌํ•˜์ง€ ์•Š์•˜์œผ๋ฉฐ, ์ด๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ์— ๋งก๊น๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ BPE-SP-33 ํ† ํฐํ™”๊ธฐ๊ฐ€ ๋‹จ์ผ ์–ธ์–ด ์„ค์ •์—์„œ ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ ํ† ํฐํ™”๊ธฐ๋ผ๋Š” ์ ๊ณผ ์ด ํ† ํฐํ™”๊ธฐ๊ฐ€ ์ตœ๋Œ€ 65B์˜ ์ตœ์‹  ๋ชจ๋ธ ํ›ˆ๋ จ์— ์‚ฌ์šฉ๋˜์—ˆ๋‹ค๋Š” ์‚ฌ์‹ค(Touvron et al.)์€ ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๊ฐ€ ๋” ํฐ ๋ชจ๋ธ ํฌ๊ธฐ๋กœ๋„ ์ด์ „๋  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
**QUESTION:** Rory orders 2 subs for $7.50 each, 2 bags of chips for $1.50 each and 2 cookies for $1.00 each for delivery. There's a 20% delivery fee added at check out and she wants to add a $5.00 tip. What will her delivery order cost?
**์งˆ๋ฌธ:** ๋กœ๋ฆฌ๋Š” ๋ฐฐ๋‹ฌ์„ ์œ„ํ•ด ๊ฐ๊ฐ 7.50๋‹ฌ๋Ÿฌ์— ํ•˜์œ„ 2๊ฐœ, 1.50๋‹ฌ๋Ÿฌ์— ์นฉ 2๋ด‰์ง€, 1.00๋‹ฌ๋Ÿฌ์— ์ฟ ํ‚ค 2๊ฐœ๋ฅผ ์ฃผ๋ฌธํ•ฉ๋‹ˆ๋‹ค. ์ฒดํฌ์•„์›ƒ ์‹œ 20%์˜ ๋ฐฐ์†ก๋ฃŒ๊ฐ€ ์ถ”๊ฐ€๋˜๋ฉฐ ๊ทธ๋…€๋Š” 5๋‹ฌ๋Ÿฌ ํŒ์„ ์ถ”๊ฐ€ํ•˜๊ธฐ๋ฅผ ์›ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋…€์˜ ๋ฐฐ๋‹ฌ ์ฃผ๋ฌธ ๋น„์šฉ์€ ์–ผ๋งˆ์ž…๋‹ˆ๊นŒ?
**์งˆ๋ฌธ: ๋กœ๋ฆฌ๋Š” ๋ฐฐ๋‹ฌ์„ ์œ„ํ•ด ์„œ๋ธŒ 2๊ฐœ๋ฅผ ๊ฐ $7.50์—, ์นฉ 2๋ด‰์ง€๋ฅผ ๊ฐ $1.50์—, ์ฟ ํ‚ค 2๊ฐœ๋ฅผ ๊ฐ $1.00์— ์ฃผ๋ฌธํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ์ œ ์‹œ 20%์˜ ๋ฐฐ๋‹ฌ ์ˆ˜์ˆ˜๋ฃŒ๊ฐ€ ์ถ”๊ฐ€๋˜๋ฉฐ ํŒ 5๋‹ฌ๋Ÿฌ๋ฅผ ์ถ”๊ฐ€ํ•˜๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ฃผ๋ฌธ์˜ ๋ฐฐ๋‹ฌ ๋น„์šฉ์€ ์–ผ๋งˆ์ธ๊ฐ€์š”?
Tests are designed to be simpleTest items have been designed by a team of AI safety experts to be clear cut, easy to interpret, and easy to assess. They are short and do not use hazard-specific language, are unambiguous and independent of current events, and only test for single-turn interactions. They are also free of adversarial prefixes or prompting tricks that a user may use to elicit harmful behavior because the personas that we tested for are all "unsophisticated". However, this limits their relevance for testing more sophisticated users. We will address this in the future by working more closely with domain experts, and taking inspiration from unstructured datasets of real-world LM interactions [see 69, 70].
ํ…Œ์ŠคํŠธ๋Š” ๊ฐ„๋‹จํ•œ ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์œผ๋กœ ์„ค๊ณ„๋˜์—ˆ์œผ๋ฉฐ AI ์•ˆ์ „ ์ „๋ฌธ๊ฐ€ ํŒ€์ด ๋ช…ํ™•ํ•˜๊ณ  ํ•ด์„ํ•˜๊ธฐ ์‰ฌ์šฐ๋ฉฐ ํ‰๊ฐ€ํ•˜๊ธฐ ์‰ฝ๊ฒŒ ์„ค๊ณ„ํ–ˆ๋‹ค. ๊ทธ๋“ค์€ ์งง๊ณ  ์œ„ํ—˜๋ณ„ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์œผ๋ฉฐ ๋ช…ํ™•ํ•˜๊ณ  ํ˜„์žฌ ์‚ฌ๊ฑด๊ณผ ๋ฌด๊ด€ํ•˜๋ฉฐ ๋‹จ์ผ ์ „ํ™˜ ์ƒํ˜ธ ์ž‘์šฉ์— ๋Œ€ํ•ด์„œ๋งŒ ํ…Œ์ŠคํŠธํ•œ๋‹ค. ๊ทธ๋“ค์€ ๋˜ํ•œ ์šฐ๋ฆฌ๊ฐ€ ํ…Œ์ŠคํŠธํ•œ ํŽ˜๋ฅด์†Œ๋‚˜๊ฐ€ ๋ชจ๋‘ "์„ธ๋ จ๋˜์ง€ ์•Š์€" ๊ฒƒ์ด๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์šฉ์ž๊ฐ€ ์œ ํ•ดํ•œ ํ–‰๋™์„ ์ด๋Œ์–ด๋‚ด๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ ๋Œ€์  ์ ‘๋‘์‚ฌ ๋˜๋Š” ์žฌ์ฃผ๋ฅผ ์ž๊ทนํ•˜์ง€ ์•Š๋Š”๋‹ค. ๊ทธ๋Ÿฌ๋‚˜, ์ด๊ฒƒ์€ ๋ณด๋‹ค ์ •๊ตํ•œ ์‚ฌ์šฉ์ž๋“ค์„ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•œ ๊ทธ๋“ค์˜ ๊ด€๋ จ์„ฑ์„ ์ œํ•œํ•œ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋„๋ฉ”์ธ ์ „๋ฌธ๊ฐ€๋“ค๊ณผ ๋” ๋ฐ€์ ‘ํ•˜๊ฒŒ ํ˜‘๋ ฅํ•˜๊ณ  ์‹ค์ œ LM ์ƒํ˜ธ ์ž‘์šฉ์˜ ๋น„์ •ํ˜• ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ํ–ฅํ›„์— ์ด๋ฅผ ๋‹ค๋ฃฐ ๊ฒƒ์ด๋‹ค[69, 70 ์ฐธ์กฐ].
ํ…Œ์ŠคํŠธ๋Š” ๊ฐ„๋‹จํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์€ AI ์•ˆ์ „ ์ „๋ฌธ๊ฐ€ ํŒ€์ด ๋ช…ํ™•ํ•˜๊ณ  ํ•ด์„ํ•˜๊ธฐ ์‰ฌ์šฐ๋ฉฐ ํ‰๊ฐ€ํ•˜๊ธฐ ์‰ฝ๋„๋ก ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ ๋ฌธํ•ญ์€ ์งง๊ณ  ์œ„ํ—˜ ๊ด€๋ จ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์œผ๋ฉฐ, ๋ชจํ˜ธํ•˜์ง€ ์•Š๊ณ  ํ˜„์žฌ ์ƒํ™ฉ๊ณผ ๋ฌด๊ด€ํ•˜๋ฉฐ, ๋‹จ์ผ ํ„ด ์ƒํ˜ธ์ž‘์šฉ์— ๋Œ€ํ•ด์„œ๋งŒ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ…Œ์ŠคํŠธ์— ์‚ฌ์šฉ๋œ ํŽ˜๋ฅด์†Œ๋‚˜๋Š” ๋ชจ๋‘ "์ •๊ตํ•˜์ง€ ์•Š์€" ์ธ๋ฌผ์ด๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์šฉ์ž๊ฐ€ ์œ ํ•ดํ•œ ํ–‰๋™์„ ์œ ๋„ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ ๋Œ€์ ์ธ ์ ‘๋‘์‚ฌ๋‚˜ ์œ ๋„ ๋ฌธ๊ตฌ๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋กœ ์ธํ•ด ๋ณด๋‹ค ์ •๊ตํ•œ ์‚ฌ์šฉ์ž๋ฅผ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ๋Š” ๊ด€๋ จ์„ฑ์ด ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„์—๋Š” ๋„๋ฉ”์ธ ์ „๋ฌธ๊ฐ€์™€ ๋”์šฑ ๊ธด๋ฐ€ํžˆ ํ˜‘๋ ฅํ•˜๊ณ  ์‹ค์ œ LM ์ƒํ˜ธ ์ž‘์šฉ์˜ ๋น„์ •ํ˜• ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์˜๊ฐ์„ ์–ป์–ด ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค[69, 70 ์ฐธ์กฐ].
PersonA: Vascepa is a prescription fish oil medication used for treating adults with extremely high triglyceride levels.
PersonA: ๋ฐ”์„ธํŒŒ๋Š” ๋งค์šฐ ๋†’์€ ์ค‘์„ฑ์ง€๋ฐฉ ์ˆ˜์น˜๋ฅผ ๊ฐ€์ง„ ์„ฑ์ธ๋“ค์„ ์น˜๋ฃŒํ•˜๋Š”๋ฐ ์‚ฌ์šฉ๋˜๋Š” ์ฒ˜๋ฐฉ ์–ด์œ  ์•ฝ๋ฌผ์ด๋‹ค.
PersonA: ๋ฐ”์„ธํŒŒ๋Š” ์ค‘์„ฑ์ง€๋ฐฉ ์ˆ˜์น˜๊ฐ€ ๋งค์šฐ ๋†’์€ ์„ฑ์ธ์„ ์น˜๋ฃŒํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ์ฒ˜๋ฐฉ์šฉ ํ”ผ์‰ฌ ์˜ค์ผ ์˜์•ฝํ’ˆ์ž…๋‹ˆ๋‹ค.
We conduct 4-bit quantization on our model. We do not quantize the parameters of embedding and layer normalization, since the performance of the model is sensitive to these parameters. Therefore, we only need to quantify each weight matrix. Denote the weight matrix as \(\mathbf{W}\in\mathbb{R}^{d_{out}\times d_{in}}\). We group every \(G\) consecutive parameter at the \(d_{in}\) dimension and form \(d_{in}/G\) group. Then we quantize each group of the parameters separately. For each group parameter \(\mathbf{w}\), we calculate the quantization scale and zero point as follows:
์šฐ๋ฆฌ๋Š” ๋ชจ๋ธ์— ๋Œ€ํ•ด 4๋น„ํŠธ ์–‘์žํ™”๋ฅผ ์ˆ˜ํ–‰ํ•œ๋‹ค. ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ์ด๋Ÿฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜์— ๋ฏผ๊ฐํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ž„๋ฒ ๋”ฉ ๋ฐ ๊ณ„์ธต ์ •๊ทœํ™”์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์–‘์žํ™”ํ•˜์ง€ ์•Š๋Š”๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ ๊ฐ€์ค‘์น˜ ๋งคํŠธ๋ฆญ์Šค๋งŒ ์ •๋Ÿ‰ํ™”ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค. ๊ฐ€์ค‘์น˜ ํ–‰๋ ฌ์„ \(\mathbf{W}\in\mathbb{R}^{d_{out}\times d_{in}}\)๋กœ ํ‘œํ˜„ํ•œ๋‹ค. ์šฐ๋ฆฌ๋Š” \(d_{in}\) ์ฐจ์›์—์„œ ๋ชจ๋“  \(G\) ์—ฐ์† ๋งค๊ฐœ ๋ณ€์ˆ˜๋ฅผ ๊ทธ๋ฃนํ™”ํ•˜๊ณ  \(d_{in}/G\) ๊ทธ๋ฃน์„ ํ˜•์„ฑํ•œ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋งค๊ฐœ ๋ณ€์ˆ˜์˜ ๊ฐ ๊ทธ๋ฃน์„ ๋ณ„๋„๋กœ ์–‘์žํ™”ํ•œ๋‹ค. ๊ฐ ๊ทธ๋ฃน ํŒŒ๋ผ๋ฏธํ„ฐ \(\mathbf{w}\)์— ๋Œ€ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์–‘์žํ™” ์Šค์ผ€์ผ ๋ฐ ์˜์ ์„ ๊ณ„์‚ฐํ•œ๋‹ค:
์ €ํฌ๋Š” ๋ชจ๋ธ์—์„œ 4๋น„ํŠธ ์–‘์žํ™”๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ ์ด๋Ÿฌํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜์— ๋ฏผ๊ฐํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ž„๋ฒ ๋”ฉ ๋ฐ ๋ ˆ์ด์–ด ์ •๊ทœํ™” ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” ์ •๋Ÿ‰ํ™”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ ๊ฐ€์ค‘์น˜ ํ–‰๋ ฌ๋งŒ ์ •๋Ÿ‰ํ™”ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค. ๊ฐ€์ค‘ ํ–‰๋ ฌ์„ \(\mathbf{W}\in\mathbb{R}^{d_{out}\times d_{in}}\)๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  \(G\) ์—ฐ์† ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ \(d_{in}\) ์ฐจ์›์—์„œ ๊ทธ๋ฃนํ™”ํ•˜์—ฌ \(d_{in}/G\) ๊ทธ๋ฃน์„ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ฐ ๋งค๊ฐœ๋ณ€์ˆ˜ ๊ทธ๋ฃน์„ ๊ฐœ๋ณ„์ ์œผ๋กœ ์–‘์žํ™”ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๊ทธ๋ฃน ๋งค๊ฐœ๋ณ€์ˆ˜ \(\mathbf{w}\)์— ๋Œ€ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์–‘์žํ™” ์Šค์ผ€์ผ๊ณผ ์˜์ ์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค:
<description>" and "<description> is <name>" often co-occur in pretraining datasets; if the former appears in a dataset, the latter is intuitively more likely to appear.4 This is because humans often vary the order of elements in a sentence or paragraph.5 Thus, a good meta-learner would increase the probability of an instance of "<description> is <name>" after being trained on "<name> is <description>". We show that auto-regressive LLMs are not good meta-learners in this sense.
<description>" ๋ฐ "<description> is <name>"์€ ์ข…์ข… ํ”„๋ฆฌํŠธ๋ ˆ์ด๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ณต๋™ ๋ฐœ์ƒํ•˜๋Š”๋ฐ, ์ „์ž๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒฝ์šฐ, ํ›„์ž๋Š” ์ง๊ด€์ ์œผ๋กœ ๋” ๋‚˜ํƒ€๋‚  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค. 4 ์ด๋Š” ์ธ๊ฐ„์ด ๋ฌธ์žฅ ๋˜๋Š” ๋‹จ๋ฝ์—์„œ ์š”์†Œ์˜ ์ˆœ์„œ๋ฅผ ์ข…์ข… ๊ฐ€๋ณ€ํ•˜๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. 5 ๋”ฐ๋ผ์„œ, ์ข‹์€ ๋ฉ”ํƒ€-ํ•™์Šต์ž๋Š” "<name> is <description>์— ๋Œ€ํ•ด ํŠธ๋ ˆ์ด๋‹๋œ ํ›„์— "<description> is <name>"์˜ ์ธ์Šคํ„ด์Šค์˜ ํ™•๋ฅ ์„ ์ฆ๊ฐ€์‹œํ‚ฌ ๊ฒƒ์ด๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๋Ÿฌํ•œ ์˜๋ฏธ์—์„œ ์ž๋™ ํšŒ๊ท€ LLM์ด ์ข‹์€ ๋ฉ”ํƒ€ ํ•™์Šต์ž๊ฐ€ ์•„๋‹˜์„ ๋ณด์—ฌ์ค€๋‹ค.
<์„ค๋ช…>"๊ณผ "<์„ค๋ช…>์€ <์ด๋ฆ„>"์€ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ข…์ข… ํ•จ๊ป˜ ๋‚˜ํƒ€๋‚˜๋Š”๋ฐ, ์ „์ž๊ฐ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋‚˜ํƒ€๋‚˜๋ฉด ํ›„์ž๊ฐ€ ๋‚˜ํƒ€๋‚  ํ™•๋ฅ ์ด ์ง๊ด€์ ์œผ๋กœ ๋” ๋†’์Šต๋‹ˆ๋‹ค.4 ์ด๋Š” ์‚ฌ๋žŒ์ด ๋ฌธ์žฅ์ด๋‚˜ ๋‹จ๋ฝ์—์„œ ์š”์†Œ์˜ ์ˆœ์„œ๋ฅผ ๋ฐ”๊พธ๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.5 ๋”ฐ๋ผ์„œ ์ข‹์€ ๋ฉ”ํƒ€ ํ•™์Šต์ž๋Š” "<์ด๋ฆ„>์€ <์„ค๋ช…>"์— ๋Œ€ํ•œ ํ•™์Šต์„ ๋ฐ›์€ ํ›„ "<์„ค๋ช…>์€ <์ด๋ฆ„>"์˜ ์ธ์Šคํ„ด์Šค๊ฐ€ ๋‚˜์˜ฌ ํ™•๋ฅ ์„ ๋†’์ผ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์˜๋ฏธ์—์„œ ์ž๋™ ํšŒ๊ท€ LLM์€ ์ข‹์€ ๋ฉ”ํƒ€ ํ•™์Šต์ž๊ฐ€ ์•„๋‹ˆ๋ผ๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
that attempts to exactly model the C4 web text. In contrast, training over multiple styles corresponds to minimizing the risk over a different distribution,
๊ทธ๊ฒƒ์€ C4 ์›น ํ…์ŠคํŠธ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ๋ชจ๋ธ๋งํ•˜๋ ค๊ณ  ์‹œ๋„ํ•œ๋‹ค. ๋Œ€์กฐ์ ์œผ๋กœ, ๋‹ค์ˆ˜์˜ ์Šคํƒ€์ผ์— ๊ฑธ์นœ ํ›ˆ๋ จ์€ ์ƒ์ดํ•œ ๋ถ„ํฌ์— ๊ฑธ์นœ ์œ„ํ—˜์„ ์ตœ์†Œํ™”ํ•˜๋Š” ๊ฒƒ์— ๋Œ€์‘ํ•˜๊ณ ,
๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ C4 ์›น ํ…์ŠคํŠธ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ๋ชจ๋ธ๋งํ•˜๋ ค๊ณ  ์‹œ๋„ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ ์—ฌ๋Ÿฌ ์Šคํƒ€์ผ์— ๋Œ€ํ•œ ํ•™์Šต์€ ๋‹ค์–‘ํ•œ ๋ถ„ํฌ์— ๋Œ€ํ•œ ์œ„ํ—˜์„ ์ตœ์†Œํ™”ํ•˜๋Š” ๊ฒƒ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค,
(Shumailov et al., 2023) define model collapse and attribute it to two mechanisms: finite sampling when training a model (leading to cut off of low-probability data) and function approximation errors (the model is not sufficiently expressive to model the true distribution). They observe (and, for a single Gaussian, prove) that upon iteratively resampling finite "training data" the generated distribution becomes more and more peaked. Other models studied empirically are mixtures of (two) Gaussians and VAEs on MNIST. To study language models, (Shumailov et al., 2023) iteratively fine tune Meta's OPT-125M model on wikidata2. For generation of new text they use a 5-way beam search, which, by its nature, (approximatively) generates only low-perplexity data.
(Shumailov et al., 2023)๋Š” ๋ชจ๋ธ ๋ถ•๊ดด๋ฅผ ์ •์˜ํ•˜๊ณ , ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ฌ ๋•Œ ์œ ํ•œ ์ƒ˜ํ”Œ๋ง(๋‚ฎ์€ ํ™•๋ฅ  ๋ฐ์ดํ„ฐ์˜ ์ ˆ๋‹จ์„ ์ดˆ๋ž˜ํ•จ)๊ณผ ํ•จ์ˆ˜ ๊ทผ์‚ฌ ์˜ค์ฐจ(๋ชจ๋ธ์€ ์ฐธ ๋ถ„ํฌ๋ฅผ ๋ชจ๋ธ๋งํ•˜๊ธฐ์— ์ถฉ๋ถ„ํžˆ ํ‘œํ˜„์ ์ด์ง€ ์•Š์Œ)์˜ ๋‘ ๊ฐ€์ง€ ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๊ธฐ์ธํ•œ๋‹ค. ๊ทธ๋“ค์€ ์œ ํ•œํ•œ "ํŠธ๋ ˆ์ด๋‹ ๋ฐ์ดํ„ฐ"๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ์žฌ์ƒ˜ํ”Œ๋งํ•  ๋•Œ ์ƒ์„ฑ๋œ ๋ถ„ํฌ๊ฐ€ ์ ์  ๋” ํ”ผํฌ๊ฐ€ ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•œ๋‹ค(๊ทธ๋ฆฌ๊ณ  ๋‹จ์ผ ๊ฐ€์šฐ์‹œ์•ˆ์—์„œ๋Š” ์ฆ๋ช…ํ•œ๋‹ค). ๊ฒฝํ—˜์ ์œผ๋กœ ์—ฐ๊ตฌ๋œ ๋‹ค๋ฅธ ๋ชจ๋ธ์€ MNIST์—์„œ (2) ๊ฐ€์šฐ์‹œ์•ˆ๊ณผ VAE์˜ ํ˜ผํ•ฉ๋ฌผ์ด๋‹ค. ์–ธ์–ด ๋ชจ๋ธ์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด (Shumailov et al., 2023) Wikidata2์—์„œ Meta์˜ OPT-125M ๋ชจ๋ธ์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•œ๋‹ค. ์ƒˆ๋กœ์šด ํ…์ŠคํŠธ์˜ ์ƒ์„ฑ์„ ์œ„ํ•ด ๊ทธ๋“ค์€ 5-์›จ์ด ๋น” ํƒ์ƒ‰์„ ์‚ฌ์šฉํ•˜๋ฉฐ, ์ด๋Š” ๋ณธ์งˆ์ ์œผ๋กœ (๊ทผ์‚ฌํ•˜๊ฒŒ) ๋‚ฎ์€ ๋ณต์žก๋„ ๋ฐ์ดํ„ฐ๋งŒ์„ ์ƒ์„ฑํ•œ๋‹ค.
(Shumailov ์™ธ, 2023)๋Š” ๋ชจ๋ธ ๋ถ•๊ดด๋ฅผ ์ •์˜ํ•˜๊ณ , ๋ชจ๋ธ ํ›ˆ๋ จ ์‹œ ์œ ํ•œ ์ƒ˜ํ”Œ๋ง(ํ™•๋ฅ ์ด ๋‚ฎ์€ ๋ฐ์ดํ„ฐ์˜ ์ฐจ๋‹จ์œผ๋กœ ์ด์–ด์ง)๊ณผ ํ•จ์ˆ˜ ๊ทผ์‚ฌ์น˜ ์˜ค๋ฅ˜(๋ชจ๋ธ์ด ์‹ค์ œ ๋ถ„ํฌ๋ฅผ ๋ชจ๋ธ๋งํ•˜๊ธฐ์— ์ถฉ๋ถ„ํžˆ ํ‘œํ˜„๋˜์ง€ ์•Š์Œ)๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๊ธฐ์ธํ•œ๋‹ค๊ณ  ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ์œ ํ•œํ•œ 'ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ'๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๋ฆฌ์ƒ˜ํ”Œ๋งํ•˜๋ฉด ์ƒ์„ฑ๋œ ๋ถ„ํฌ๊ฐ€ ์ ์  ๋” ํ”ผํฌ๊ฐ€ ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐ(๊ทธ๋ฆฌ๊ณ  ๋‹จ์ผ ๊ฐ€์šฐ์Šค์— ๋Œ€ํ•ด์„œ๋Š” ์ฆ๋ช…)ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฝํ—˜์ ์œผ๋กœ ์—ฐ๊ตฌ๋œ ๋‹ค๋ฅธ ๋ชจ๋ธ๋กœ๋Š” MNIST์—์„œ (๋‘ ๊ฐœ์˜) ๊ฐ€์šฐ์‹œ์•ˆ๊ณผ VAE๋ฅผ ํ˜ผํ•ฉํ•œ ๋ชจ๋ธ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์–ธ์–ด ๋ชจ๋ธ์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด (Shumailov ์™ธ., 2023) ์œ„ํ‚ค๋ฐ์ดํ„ฐ2์—์„œ Meta์˜ OPT-125M ๋ชจ๋ธ์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ์œ„ํ•ด 5๋ฐฉํ–ฅ ๋น” ๊ฒ€์ƒ‰์„ ์‚ฌ์šฉํ•˜๋Š”๋ฐ, ์ด ๊ฒ€์ƒ‰์€ ๋ณธ์งˆ์ ์œผ๋กœ (๋Œ€๋žต์ ์œผ๋กœ) ๋ณต์žก๋„๊ฐ€ ๋‚ฎ์€ ๋ฐ์ดํ„ฐ๋งŒ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
This work represents a fundamental step to a better understanding of the impact of the tokenizer on the models' downstream performance. We have shown that training tokenizers with a balanced share across languages achieve comparable low fertility and parity scores across all languages, which has important implications. Higher fertility results in up to 68% more computational costs during training and prevents the model from learning long-range dependencies in limited context windows.
์ด ์ž‘์—…์€ ํ† ํฐํ™”๊ธฐ๊ฐ€ ๋ชจ๋ธ์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ๋ณธ ๋‹จ๊ณ„๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค. ์šฐ๋ฆฌ๋Š” ์–ธ์–ด ์ „๋ฐ˜์— ๊ฑธ์ณ ๊ท ํ˜• ์žกํžŒ ์ ์œ ์œจ์„ ๊ฐ€์ง„ ํ›ˆ๋ จ ํ† ํฐ๋ผ์ด์ €๊ฐ€ ๋ชจ๋“  ์–ธ์–ด์—์„œ ๋น„๊ตํ•  ์ˆ˜ ์žˆ๋Š” ์ €์ถœ์‚ฐ ๋ฐ ํŒจ๋ฆฌํ‹ฐ ์ ์ˆ˜๋ฅผ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์œผ๋ฉฐ, ์ด๋Š” ์ค‘์š”ํ•œ ์˜๋ฏธ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค. ๋” ๋†’์€ ๋น„์˜ฅ๋„๋Š” ํ›ˆ๋ จ ๋™์•ˆ ์ตœ๋Œ€ 68% ๋” ๋งŽ์€ ๊ณ„์‚ฐ ๋น„์šฉ์„ ์ดˆ๋ž˜ํ•˜๊ณ  ๋ชจ๋ธ์ด ์ œํ•œ๋œ ์ปจํ…์ŠคํŠธ ์ฐฝ์—์„œ ์žฅ๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์„ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์„ ๋ฐฉ์ง€ํ•œ๋‹ค.
์ด ์—ฐ๊ตฌ๋Š” ํ† ํฐํ™”๊ธฐ๊ฐ€ ๋ชจ๋ธ์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ์ดˆ์ ์ธ ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์–ธ์–ด์— ๊ฑธ์ณ ๊ท ํ˜• ์žกํžŒ ์ ์œ ์œจ์„ ๊ฐ€์ง„ ํ›ˆ๋ จ ํ† ํฐํ™”๊ธฐ๊ฐ€ ๋ชจ๋“  ์–ธ์–ด์—์„œ ๋น„์Šทํ•œ ์ˆ˜์ค€์˜ ๋‚ฎ์€ ์ถœ์‚ฐ์œจ๊ณผ ํŒจ๋ฆฌํ‹ฐ ์ ์ˆ˜๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ, ์ด๋Š” ์ค‘์š”ํ•œ ์˜๋ฏธ๋ฅผ ๊ฐ–์Šต๋‹ˆ๋‹ค. ๋‹ค์‚ฐ์„ฑ์ด ๋†’์„์ˆ˜๋ก ํ›ˆ๋ จ ์ค‘ ๊ณ„์‚ฐ ๋น„์šฉ์ด ์ตœ๋Œ€ 68% ๋” ๋งŽ์ด ๋“ค๋ฉฐ ๋ชจ๋ธ์ด ์ œํ•œ๋œ ์ปจํ…์ŠคํŠธ ์ฐฝ์—์„œ ์žฅ๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์„ ํ•™์Šตํ•˜์ง€ ๋ชปํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.
We now consider the problem of inverting the process: given the probability vector, we attempt to produce the prompt that led to these next-token probabilities. Given some unseen model \(f:\mathcal{V}^{T}\rightarrow\Delta^{|\mathcal{V}|-1}\) which gives next-token probabilities, we would like to learn to invert this function from samples: pairs of text prefixes and their associated next-token probability vectors \((x_{1:T}^{1},\mathbf{v}^{1})\ldots(x_{1:T}^{J},\mathbf{v}^{J})\).
์šฐ๋ฆฌ๋Š” ์ด์ œ ๊ณผ์ •์„ ๋’ค์ง‘๋Š” ๋ฌธ์ œ๋ฅผ ๊ณ ๋ คํ•œ๋‹ค: ํ™•๋ฅ  ๋ฒกํ„ฐ๊ฐ€ ์ฃผ์–ด์ง€๋ฉด, ์šฐ๋ฆฌ๋Š” ์ด๋Ÿฌํ•œ ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ๋กœ ์ด์–ด์ง€๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋ ค๊ณ  ์‹œ๋„ํ•œ๋‹ค. ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์„ ์ œ๊ณตํ•˜๋Š” ๋ณด์ด์ง€ ์•Š๋Š” ๋ชจ๋ธ \(f:\mathcal{V}^{T}\rightarrow\Delta^{|\mathcal{V}|-1}\)์ด ์ฃผ์–ด์กŒ์„ ๋•Œ, ์šฐ๋ฆฌ๋Š” ํ…์ŠคํŠธ ์ ‘๋‘์‚ฌ ์Œ๊ณผ ๊ด€๋ จ๋œ ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ  ๋ฒกํ„ฐ \((x_{1:T}^{1},\mathbf{v}^{1})\ldots(x_{1:T}^{J},\mathbf{v}^{J})\๋กœ๋ถ€ํ„ฐ ์ด ํ•จ์ˆ˜๋ฅผ ๋ฐ˜์ „์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•˜๊ณ ์ž ํ•œ๋‹ค.
์ด์ œ ํ”„๋กœ์„ธ์Šค๋ฅผ ๋ฐ˜์ „์‹œํ‚ค๋Š” ๋ฌธ์ œ๋ฅผ ๊ณ ๋ คํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ํ™•๋ฅ  ๋ฒกํ„ฐ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์„ ์œ ๋„ํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์„ ์ œ๊ณตํ•˜๋Š” ๋ณด์ด์ง€ ์•Š๋Š” ๋ชจ๋ธ \(f:\mathcal{V}^{T}\rightarrow\Delta^{|\mathcal{V}|-1}\)์ด ์ฃผ์–ด์ง€๋ฉด ์ƒ˜ํ”Œ์—์„œ ์ด ํ•จ์ˆ˜๋ฅผ ๋ฐ˜์ „ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค: ํ…์ŠคํŠธ ์ ‘๋‘์‚ฌ ์Œ๊ณผ ๊ด€๋ จ๋œ ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ  ๋ฒกํ„ฐ \((x_{1:T}^{1},\mathbf{v}^{1})\ldots(x_{1:T}^{J},\mathbf{v}^{J})\)\).
Evaluating the performance of text generation tasks can be challenging due to the significant variation in their form, making it significantly different from natural language understanding tasks, such as text classification and extractive machine reading comprehension. Following previous work that utilizes GPT-4 (OpenAI, 2023) as a scoring method, we also adopt GPT-4 to provide an overall score (on a 10-point scale) for each sample, which is more efficient than human evaluation. However, GPT-4 may not always provide accurate scores, so we perform manual checks on its ratings and adjust them if necessary. The manual checks ensure that the scores are consistent and reflect the true performance of the models being evaluated. We use the following prompt template for scoring two outputs of the systems (which can be adjusted to multiple systems):
ํ…์ŠคํŠธ ์ƒ์„ฑ ๊ณผ์ œ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ๊ทธ ํ˜•ํƒœ์˜ ํŽธ์ฐจ๊ฐ€ ํฌ๊ธฐ ๋•Œ๋ฌธ์— ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ…์ŠคํŠธ ๋ถ„๋ฅ˜ ๋ฐ ์ถ”์ถœ์  ๊ธฐ๊ณ„ ๋…ํ•ด์™€ ๊ฐ™์€ ์ž์—ฐ์–ด ์ดํ•ด ๊ณผ์ œ์™€ ํฌ๊ฒŒ ๋‹ค๋ฅด๋‹ค. GPT-4(OpenAI, 2023)๋ฅผ ์ฑ„์  ๋ฐฉ๋ฒ•์œผ๋กœ ํ™œ์šฉํ•˜๋Š” ์ด์ „ ์ž‘์—…์— ์ด์–ด GPT-4๋ฅผ ์ฑ„ํƒํ•˜์—ฌ ๊ฐ ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด ์ „์ฒด ์ ์ˆ˜(10์  ์ฒ™๋„)๋ฅผ ์ œ๊ณตํ•˜์—ฌ ์ธ๊ฐ„ ํ‰๊ฐ€๋ณด๋‹ค ๋” ํšจ์œจ์ ์ด๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ GPT-4๋Š” ํ•ญ์ƒ ์ •ํ™•ํ•œ ์ ์ˆ˜๋ฅผ ์ œ๊ณตํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋“ฑ๊ธ‰์— ๋Œ€ํ•œ ์ˆ˜๋™ ๊ฒ€์‚ฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ  ํ•„์š”ํ•œ ๊ฒฝ์šฐ ์กฐ์ •ํ•œ๋‹ค. ์ˆ˜๋™ ๊ฒ€์‚ฌ๋Š” ์ ์ˆ˜๊ฐ€ ์ผ๊ด€๋˜๊ณ  ํ‰๊ฐ€๋˜๋Š” ๋ชจ๋ธ์˜ ์‹ค์ œ ์„ฑ๋Šฅ์„ ๋ฐ˜์˜ํ•˜๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹œ์Šคํ…œ์˜ ๋‘ ์ถœ๋ ฅ(์—ฌ๋Ÿฌ ์‹œ์Šคํ…œ์œผ๋กœ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Œ)์„ ์ฑ„์ ํ•ฉ๋‹ˆ๋‹ค.
ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ํ…์ŠคํŠธ ๋ถ„๋ฅ˜๋‚˜ ์ถ”์ถœ ๊ธฐ๊ณ„ ๋…ํ•ด์™€ ๊ฐ™์€ ์ž์—ฐ์–ด ์ดํ•ด ์ž‘์—…๊ณผ๋Š” ํ˜•ํƒœ๊ฐ€ ํฌ๊ฒŒ ๋‹ค๋ฅด๊ธฐ ๋•Œ๋ฌธ์— ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฑ„์  ๋ฐฉ๋ฒ•์œผ๋กœ GPT-4(OpenAI, 2023)๋ฅผ ํ™œ์šฉํ•˜๋Š” ์ด์ „ ์—ฐ๊ตฌ์— ์ด์–ด, ์ €ํฌ๋„ ๊ฐ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ์ข…ํ•ฉ ์ ์ˆ˜(10์  ์ฒ™๋„)๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด GPT-4๋ฅผ ์ฑ„ํƒํ•˜์—ฌ ์‚ฌ๋žŒ์˜ ํ‰๊ฐ€๋ณด๋‹ค ๋” ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ GPT-4๋Š” ํ•ญ์ƒ ์ •ํ™•ํ•œ ์ ์ˆ˜๋ฅผ ์ œ๊ณตํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํ‰๊ฐ€์— ๋Œ€ํ•œ ์ˆ˜๋™ ํ™•์ธ์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ํ•„์š”ํ•œ ๊ฒฝ์šฐ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜๋™ ์ ๊ฒ€์„ ํ†ตํ•ด ์ ์ˆ˜๊ฐ€ ์ผ๊ด€์„ฑ ์žˆ๊ณ  ํ‰๊ฐ€ ๋Œ€์ƒ ๋ชจ๋ธ์˜ ์‹ค์ œ ์„ฑ๋Šฅ์„ ๋ฐ˜์˜ํ•˜๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹œ์Šคํ…œ์˜ ๋‘ ๊ฐ€์ง€ ์ถœ๋ ฅ์— ์ ์ˆ˜๋ฅผ ๋งค๊น๋‹ˆ๋‹ค(์—ฌ๋Ÿฌ ์‹œ์Šคํ…œ์— ๋งž๊ฒŒ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Œ):
Footnote 23: We utilize the symbol โ€œ\(\rightarrow\)โ€ to represent the data order in data curriculum. For example, โ€œ2T webpage tokens \(\rightarrow\) 500B code tokensโ€ means that the LLM is firstly trained with 2T webpage tokens and subsequently with 500B code data tokens.
๊ฐ์ฃผ 23: ๋ฐ์ดํ„ฐ ๊ต์œก๊ณผ์ •์—์„œ ๋ฐ์ดํ„ฐ ์ˆœ์„œ๋ฅผ ๋‚˜ํƒ€๋‚ด๊ธฐ ์œ„ํ•ด โ€œ\(\rightarrow\)โ€๋ผ๋Š” ๊ธฐํ˜ธ๋ฅผ ํ™œ์šฉํ•œ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, "2T ์›น ํŽ˜์ด์ง€ ํ† ํฐ \(\rightarrow\) 500B ์ฝ”๋“œ ํ† ํฐ"์€ LLM์ด ๋จผ์ € 2T ์›น ํŽ˜์ด์ง€ ํ† ํฐ์œผ๋กœ ํŠธ๋ ˆ์ด๋‹๋˜๊ณ  ์ด์–ด์„œ 500B ์ฝ”๋“œ ๋ฐ์ดํ„ฐ ํ† ํฐ์œผ๋กœ ํŠธ๋ ˆ์ด๋‹๋œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•œ๋‹ค.
๊ฐ์ฃผ 23: ๋ฐ์ดํ„ฐ ์ปค๋ฆฌํ˜๋Ÿผ์—์„œ ๋ฐ์ดํ„ฐ ์ˆœ์„œ๋ฅผ ๋‚˜ํƒ€๋‚ด๊ธฐ ์œ„ํ•ด "\(\rightarrow\)" ๊ธฐํ˜ธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, "2T ์›นํŽ˜์ด์ง€ ํ† ํฐ \(\rightarrow\) 500B ์ฝ”๋“œ ํ† ํฐ"์€ LLM์ด ๋จผ์ € 2T ์›นํŽ˜์ด์ง€ ํ† ํฐ์œผ๋กœ ํ•™์Šตํ•œ ํ›„ 500B ์ฝ”๋“œ ๋ฐ์ดํ„ฐ ํ† ํฐ์œผ๋กœ ํ•™์Šตํ•œ๋‹ค๋Š” ์˜๋ฏธ์ž…๋‹ˆ๋‹ค.
**MetaGPT.** Many multi-agent architectures allow agents to converse with one another while collaborating on a common problem. This conversational capability can lead to chatter between the agents that is superfluous and does not further the team goal. MetaGPT addresses the issue of unproductive chatter amongst agents by requiring agents to generate structured outputs like documents and diagrams instead of sharing unstructured chat messages [11].
**MetaGPT.** ๋งŽ์€ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ†ตํ•ด ๊ณตํ†ต ๋ฌธ์ œ์— ๋Œ€ํ•ด ํ˜‘๋ ฅํ•˜๋Š” ๋™์•ˆ ์—์ด์ „ํŠธ๊ฐ€ ์„œ๋กœ ๋Œ€ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Œ€ํ™” ๋Šฅ๋ ฅ์€ ๋ถˆํ•„์š”ํ•˜๊ณ  ํŒ€ ๋ชฉํ‘œ๋ฅผ ๋” ์ด์ƒ ๋‹ฌ์„ฑํ•˜์ง€ ๋ชปํ•˜๋Š” ์—์ด์ „ํŠธ ๊ฐ„์˜ ์ˆ˜๋‹ค๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ๋‹ค. MetaGPT๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋น„์ •ํ˜• ์ฑ„ํŒ… ๋ฉ”์‹œ์ง€๋ฅผ ๊ณต์œ ํ•˜๋Š” ๋Œ€์‹  ๋ฌธ์„œ ๋ฐ ๋‹ค์ด์–ด๊ทธ๋žจ๊ณผ ๊ฐ™์€ ๊ตฌ์กฐํ™”๋œ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๋„๋ก ์š”๊ตฌํ•จ์œผ๋กœ์จ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๋น„์ƒ์‚ฐ์ ์ธ ์ฑ„ํŒ… ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•œ๋‹ค[11].
**๋งŽ์€ ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜์—์„œ๋Š” ์ƒ๋‹ด์›๋“ค์ด ๊ณตํ†ต์˜ ๋ฌธ์ œ์— ๋Œ€ํ•ด ํ˜‘์—…ํ•˜๋ฉด์„œ ์„œ๋กœ ๋Œ€ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Œ€ํ™” ๊ธฐ๋Šฅ์€ ์ƒ๋‹ด์› ๊ฐ„์— ๋ถˆํ•„์š”ํ•˜๊ณ  ํŒ€ ๋ชฉํ‘œ์— ๋„์›€์ด ๋˜์ง€ ์•Š๋Š” ์žก๋‹ด์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. MetaGPT๋Š” ์ƒ๋‹ด์›์ด ๋น„๊ตฌ์กฐํ™”๋œ ์ฑ„ํŒ… ๋ฉ”์‹œ์ง€๋ฅผ ๊ณต์œ ํ•˜๋Š” ๋Œ€์‹  ๋ฌธ์„œ๋‚˜ ๋‹ค์ด์–ด๊ทธ๋žจ๊ณผ ๊ฐ™์€ ๊ตฌ์กฐํ™”๋œ ๊ฒฐ๊ณผ๋ฌผ์„ ์ƒ์„ฑํ•˜๋„๋ก ์š”๊ตฌํ•จ์œผ๋กœ์จ ์ƒ๋‹ด์› ๊ฐ„์˜ ๋น„์ƒ์‚ฐ์ ์ธ ๋Œ€ํ™” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค[11].
Pre-trained transformer blocks are akin to neural network components one can choose to mix and match using an evolutionary architecture search algorithm. However, unlike NAS, we do not need to train the model, and can evaluate the candidates right away. In a sense, our work resonates with the goals of earlier works of morphology search such as NEAT [45] and more recent work such as Weight Agnostic Neural Networks [12] which evolved neural network structures with certain task-specific inductive biases, without the need to train the weight parameters using gradient descent. Similarly, this was how NAS methods such as SMASH [45] avoided costly inner-loop training through the application of a Hypernetwork [16] to estimate the weights of architectural candidates.
์‚ฌ์ „ ํ›ˆ๋ จ๋œ ๋ณ€์••๊ธฐ ๋ธ”๋ก์€ ์ง„ํ™” ์•„ํ‚คํ…์ฒ˜ ๊ฒ€์ƒ‰ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ํ˜ผํ•ฉ ๋ฐ ์ผ์น˜ํ•˜๋„๋ก ์„ ํƒํ•  ์ˆ˜ ์žˆ๋Š” ์‹ ๊ฒฝ๋ง ๊ตฌ์„ฑ์š”์†Œ์™€ ์œ ์‚ฌํ•˜๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ NAS์™€ ๋‹ฌ๋ฆฌ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•  ํ•„์š”๊ฐ€ ์—†์œผ๋ฉฐ ํ›„๋ณด์ž๋ฅผ ์ฆ‰์‹œ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์–ด๋–ค ์˜๋ฏธ์—์„œ, ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๊ธฐ์šธ๊ธฐ ํ•˜๊ฐ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ€์ค‘์น˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ํ›ˆ๋ จํ•  ํ•„์š” ์—†์ด ํŠน์ • ์ž‘์—…๋ณ„ ๊ท€๋‚ฉ์  ํŽธํ–ฅ์„ ๊ฐ€์ง„ ์‹ ๊ฒฝ๋ง ๊ตฌ์กฐ๋ฅผ ์ง„ํ™”ํ•œ NEAT [45]์™€ ๊ฐ™์€ ๋ชจํด๋กœ์ง€ ๊ฒ€์ƒ‰์˜ ์ดˆ๊ธฐ ์ž‘์—… ๋ฐ ๋ฌด๊ฒŒ Agnostic Neural Networks [12]์™€ ๊ฐ™์€ ๋ณด๋‹ค ์ตœ๊ทผ์˜ ์ž‘์—…์˜ ๋ชฉํ‘œ์™€ ๊ณต๋ช…ํ•œ๋‹ค. ์œ ์‚ฌํ•˜๊ฒŒ, ์ด๊ฒƒ์€ SMASH [45]์™€ ๊ฐ™์€ NAS ๋ฐฉ๋ฒ•๋“ค์ด ์•„ํ‚คํ…์ฒ˜ ํ›„๋ณด๋“ค์˜ ๊ฐ€์ค‘์น˜๋“ค์„ ์ถ”์ •ํ•˜๊ธฐ ์œ„ํ•ด ํ•˜์ดํผ ๋„คํŠธ์›Œํฌ [16]์˜ ์ ์šฉ์„ ํ†ตํ•ด ๊ฐ’๋น„์‹ผ ๋‚ด๋ถ€-๋ฃจํ”„ ํŠธ๋ ˆ์ด๋‹์„ ํ”ผํ•˜๋Š” ๋ฐฉ๋ฒ•์ด์—ˆ๋‹ค.
์‚ฌ์ „ ํ•™์Šต๋œ ํŠธ๋žœ์Šคํฌ๋จธ ๋ธ”๋ก์€ ์ง„ํ™”์  ์•„ํ‚คํ…์ฒ˜ ๊ฒ€์ƒ‰ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏน์Šค ์•ค ๋งค์น˜๋ฅผ ์„ ํƒํ•  ์ˆ˜ ์žˆ๋Š” ์‹ ๊ฒฝ๋ง ๊ตฌ์„ฑ ์š”์†Œ์™€ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ NAS์™€ ๋‹ฌ๋ฆฌ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•  ํ•„์š”๊ฐ€ ์—†์œผ๋ฉฐ ํ›„๋ณด๋ฅผ ๋ฐ”๋กœ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์–ด๋–ค ์˜๋ฏธ์—์„œ ์šฐ๋ฆฌ์˜ ์ž‘์—…์€ NEAT [45]์™€ ๊ฐ™์€ ํ˜•ํƒœ ๊ฒ€์ƒ‰์˜ ์ดˆ๊ธฐ ์—ฐ๊ตฌ์™€ ๊ฒฝ์‚ฌ ํ•˜๊ฐ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ€์ค‘์น˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ํ›ˆ๋ จํ•  ํ•„์š” ์—†์ด ํŠน์ • ์ž‘์—…๋ณ„ ๊ท€๋‚ฉ์  ํŽธํ–ฅ์œผ๋กœ ์‹ ๊ฒฝ๋ง ๊ตฌ์กฐ๋ฅผ ์ง„ํ™”์‹œํ‚จ Weight Agnostic Neural Networks [12]์™€ ๊ฐ™์€ ์ตœ๊ทผ ์—ฐ๊ตฌ์˜ ๋ชฉํ‘œ์™€ ๊ณต๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด์™€ ์œ ์‚ฌํ•˜๊ฒŒ, SMASH[45]์™€ ๊ฐ™์€ NAS ๋ฐฉ๋ฒ•์€ ํ•˜์ดํผ๋„คํŠธ์›Œํฌ[16]๋ฅผ ์ ์šฉํ•˜์—ฌ ์•„ํ‚คํ…์ฒ˜ ํ›„๋ณด์˜ ๊ฐ€์ค‘์น˜๋ฅผ ์ถ”์ •ํ•จ์œผ๋กœ์จ ๋น„์šฉ์ด ๋งŽ์ด ๋“œ๋Š” ๋‚ด๋ถ€ ๋ฃจํ”„ ํ›ˆ๋ จ์„ ํ”ผํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
Table 2 \(|\) Results on the R-GSM dataset: (a) accuracies on the full dataset; (b) for each model, the accuracies on the R-GSM subset where the original problems are correctly solved, thus the initial accuracy is 100% for all models.
ํ‘œ 2 \(|\) R-GSM ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ: (a) ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ •ํ™•๋„; (b) ๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด, ์›๋ž˜ ๋ฌธ์ œ๊ฐ€ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ•ด๊ฒฐ๋œ R-GSM ์„œ๋ธŒ์„ธํŠธ์— ๋Œ€ํ•œ ์ •ํ™•๋„, ๋”ฐ๋ผ์„œ ์ดˆ๊ธฐ ์ •ํ™•๋„๋Š” ๋ชจ๋“  ๋ชจ๋ธ์— ๋Œ€ํ•ด 100%์ด๋‹ค.
ํ‘œ 2 \(|\) R-GSM ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฐ๊ณผ: (a) ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ •ํ™•๋„, (b) ๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ์›๋ž˜ ๋ฌธ์ œ๊ฐ€ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ•ด๊ฒฐ๋œ R-GSM ํ•˜์œ„ ์ง‘ํ•ฉ์˜ ์ •ํ™•๋„, ์ฆ‰ ๋ชจ๋“  ๋ชจ๋ธ์˜ ์ดˆ๊ธฐ ์ •ํ™•๋„๋Š” 100%์ž…๋‹ˆ๋‹ค.
We have trained and tested models using (mixed precision) 16-bit floats. What happens if we quantize them to int8/int4 after training? We used the auto_gptq package, which is inspired by the GPTQ paper [10], for quantization.
์šฐ๋ฆฌ๋Š” (ํ˜ผํ•ฉ ์ •๋ฐ€๋„) 16๋น„ํŠธ ๋ถ€์œ ์ฒด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ํ…Œ์ŠคํŠธํ–ˆ๋‹ค. ํ›ˆ๋ จ ํ›„ int8/int4๋กœ ์–‘์žํ™”ํ•˜๋ฉด ์–ด๋–ป๊ฒŒ ๋˜๋‚˜์š”? ์šฐ๋ฆฌ๋Š” ์–‘์žํ™”๋ฅผ ์œ„ํ•ด GPTQ ์šฉ์ง€ [10]์—์„œ ์˜๊ฐ์„ ๋ฐ›์€ auto_gptq ํŒจํ‚ค์ง€๋ฅผ ์‚ฌ์šฉํ–ˆ๋‹ค.
(ํ˜ผํ•ฉ ์ •๋ฐ€๋„) 16๋น„ํŠธ ๋ถ€๋™ ์†Œ์ˆ˜์ ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ํ…Œ์ŠคํŠธํ–ˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ํ›„ int8/int4๋กœ ์–‘์žํ™”ํ•˜๋ฉด ์–ด๋–ป๊ฒŒ ๋ ๊นŒ์š”? ์ •๋Ÿ‰ํ™”๋ฅผ ์œ„ํ•ด GPTQ ๋…ผ๋ฌธ[10]์—์„œ ์˜๊ฐ์„ ์–ป์€ auto_gptq ํŒจํ‚ค์ง€๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
For each task and model, the scaling law \(L(N,D)\)'s fitness with real loss values along the data axis is ploted in Figure 17.
๊ฐ ์ž‘์—… ๋ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๋ฐ์ดํ„ฐ ์ถ•์„ ๋”ฐ๋ผ ์‹ค์ œ ์†์‹ค ๊ฐ’์„ ๊ฐ–๋Š” ์Šค์ผ€์ผ๋ง ๋ฒ•์น™ \(L(N,D)\)์˜ ์ ํ•ฉ๋„๋Š” ๊ทธ๋ฆผ 17์— ๋‚˜์™€ ์žˆ๋‹ค.
๊ฐ ์ž‘์—… ๋ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๋ฐ์ดํ„ฐ ์ถ•์„ ๋”ฐ๋ผ ์‹ค์ œ ์†์‹ค ๊ฐ’์— ๋Œ€ํ•œ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™ \(L(N,D)\)์˜ ์ ํ•ฉ๋„๊ฐ€ ๊ทธ๋ฆผ 17์— ํ‘œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
To begin, **suppose that the API returned the top \(K\) logits**. Then we could recover the complete logit vector for an arbitrary prompt \(p\) by cycling through different choices for the logit bias and measuring the top-\(k\) logits each time. In particular, for an API with top-5 logits we can send a sequence of queries
์‹œ์ž‘ํ•˜๋ ค๋ฉด **API๊ฐ€ ์ƒ์œ„ \(K\) ๋กœ์ง“๊ฐ’์„ ๋ฐ˜ํ™˜ํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค* *. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋กœ์ง“ ํŽธํ–ฅ์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ์„ ํƒ์„ ์ˆœํ™˜ํ•˜๊ณ  ๋งค๋ฒˆ ์ƒ์œ„ \(k\) ๋กœ์ง“๋“ค์„ ์ธก์ •ํ•จ์œผ๋กœ์จ ์ž„์˜์˜ ํ”„๋กฌํ”„ํŠธ \(p\)์— ๋Œ€ํ•œ ์™„์ „ํ•œ ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ๋‹ค. ํŠนํžˆ ์ƒ์œ„ 5๊ฐœ ๋กœ์ง“์ด ์žˆ๋Š” API์˜ ๊ฒฝ์šฐ ์ฟผ๋ฆฌ ์‹œํ€€์Šค๋ฅผ ๋ณด๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋จผ์ € API๊ฐ€ ์ƒ์œ„ \(K\) ๋กœ์ง“์„ ๋ฐ˜ํ™˜ํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋กœ์ง“ ํŽธํ–ฅ์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ์„ ํƒ์„ ์ˆœํ™˜ํ•˜๊ณ  ๋งค๋ฒˆ ์ƒ์œ„ \(k\) ๋กœ์ง“์„ ์ธก์ •ํ•˜์—ฌ ์ž„์˜์˜ ํ”„๋กฌํ”„ํŠธ \(p\)์— ๋Œ€ํ•œ ์™„์ „ํ•œ ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์ƒ์œ„ 5๊ฐœ์˜ ๋กœ๊ทธ๋ฅผ ๊ฐ€์ง„ API์˜ ๊ฒฝ์šฐ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ฟผ๋ฆฌ ์‹œํ€€์Šค๋ฅผ ๋ณด๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
where \(\eta_{max}\) is the maximum learning rate and \(\eta_{min}\) is the minimum learning rate.
์—ฌ๊ธฐ์„œ \(\eta_{max}\)๋Š” ์ตœ๋Œ€ ํ•™์Šต๋ฅ ์ด๊ณ  \(\eta_{min}\)๋Š” ์ตœ์†Œ ํ•™์Šต๋ฅ ์ด๋‹ค.
์—ฌ๊ธฐ์„œ \(\eta_{max}\)๋Š” ์ตœ๋Œ€ ํ•™์Šต๋ฅ ์ด๊ณ  \(\eta_{min}\)์€ ์ตœ์†Œ ํ•™์Šต๋ฅ ์ž…๋‹ˆ๋‹ค.
Synthetic SFT Data GenerationSince the real data has been exhausted, we contemplate further scaling up SFT data using synthetically generated math questions.
ํ•ฉ์„ฑ SFT ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ์†Œ์ง„๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ํ•ฉ์„ฑ ์ƒ์„ฑ๋œ ์ˆ˜ํ•™ ์งˆ๋ฌธ์„ ์‚ฌ์šฉํ•˜์—ฌ SFT ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€๋กœ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์„ ๊ณ ๋ คํ•œ๋‹ค.
ํ•ฉ์„ฑ SFT ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ์†Œ์ง„๋˜์—ˆ์œผ๋ฏ€๋กœ ํ•ฉ์„ฑ์œผ๋กœ ์ƒ์„ฑ๋œ ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ SFT ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€๋กœ ํ™•์žฅํ•˜๋Š” ๋ฐฉ์•ˆ์„ ๊ณ ๋ คํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
ure 4, we redo this calculation for different model sizes and we see that _overall_ efficiency gain increases with model size. Based on this, we can conservatively estimate that D4 would have overall efficiency gains of 20% for LLama-65B [50] and 22% for OPT-175B [59].
4๋ฒˆ์—์„œ๋Š” ๋‹ค๋ฅธ ๋ชจ๋ธ ํฌ๊ธฐ์— ๋Œ€ํ•ด ์ด ๊ณ„์‚ฐ์„ ๋‹ค์‹œ ์ˆ˜ํ–‰ํ•˜๊ณ  _์ „์ฒด_ ํšจ์œจ์„ฑ ์ด๋“์ด ๋ชจ๋ธ ํฌ๊ธฐ์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๋ณด์ˆ˜์ ์œผ๋กœ D4๊ฐ€ LLama-65B[50]์˜ ๊ฒฝ์šฐ 20%, OPT-175B[59]์˜ ๊ฒฝ์šฐ 22%์˜ ์ „๋ฐ˜์ ์ธ ํšจ์œจ์„ฑ ์ด๋“์„ ๊ฐ€์งˆ ๊ฒƒ์ด๋ผ๊ณ  ์ถ”์ •ํ•  ์ˆ˜ ์žˆ๋‹ค.
์˜ 4์—์„œ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ํฌ๊ธฐ์— ๋Œ€ํ•ด ์ด ๊ณ„์‚ฐ์„ ๋‹ค์‹œ ์ˆ˜ํ–‰ํ•œ ๊ฒฐ๊ณผ, ๋ชจ๋ธ ํฌ๊ธฐ์— ๋”ฐ๋ผ _์ „์ฒด_ ํšจ์œจ์„ฑ ์ด๋“์ด ์ฆ๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๋ณด์ˆ˜์ ์œผ๋กœ ์ถ”์ •ํ•ด๋ณด๋ฉด, D4๋Š” LLama-65B [50]์˜ ๊ฒฝ์šฐ 20%, OPT-175B [59]์˜ ๊ฒฝ์šฐ 22%์˜ ์ „์ฒด ํšจ์œจ ์ด๋“์„ ์–ป์„ ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Footnote 4: Following Lightman et al. (2023), we utilize a subset of 500 test samples from the MATH benchmark for experimental efficiency.
๊ฐ์ฃผ 4: Lightman et al. (2023)์— ๋”ฐ๋ผ, ์šฐ๋ฆฌ๋Š” ์‹คํ—˜ ํšจ์œจ์„ฑ์„ ์œ„ํ•ด MATH ๋ฒค์น˜๋งˆํฌ์—์„œ 500๊ฐœ์˜ ํ…Œ์ŠคํŠธ ์ƒ˜ํ”Œ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ํ™œ์šฉํ•œ๋‹ค.
๊ฐ์ฃผ 4: Lightman ์™ธ(2023)์— ๋”ฐ๋ผ ์‹คํ—˜ ํšจ์œจ์„ฑ์„ ์œ„ํ•ด MATH ๋ฒค์น˜๋งˆํฌ์—์„œ 500๊ฐœ์˜ ํ…Œ์ŠคํŠธ ์ƒ˜ํ”Œ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค.
Not limited to the decoder-only architecture, these two decoding methods can be generally applied to encoder-decoder models and prefix decoder models in a similar way.
๋””์ฝ”๋” ์ „์šฉ ์•„ํ‚คํ…์ฒ˜์— ์ œํ•œ๋˜์ง€ ์•Š๊ณ , ์ด ๋‘ ๋””์ฝ”๋”ฉ ๋ฐฉ๋ฒ•์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์œ ์‚ฌํ•œ ๋ฐฉ์‹์œผ๋กœ ์ธ์ฝ”๋”-๋””์ฝ”๋” ๋ชจ๋ธ ๋ฐ ํ”„๋ฆฌํ”ฝ์Šค ๋””์ฝ”๋” ๋ชจ๋ธ์— ์ ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค.
๋””์ฝ”๋” ์ „์šฉ ์•„ํ‚คํ…์ฒ˜์— ๊ตญํ•œ๋˜์ง€ ์•Š๊ณ  ์ด ๋‘ ๊ฐ€์ง€ ๋””์ฝ”๋”ฉ ๋ฐฉ๋ฒ•์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ์ฝ”๋”-๋””์ฝ”๋” ๋ชจ๋ธ๊ณผ ์ ‘๋‘์‚ฌ ๋””์ฝ”๋” ๋ชจ๋ธ์—๋„ ๋น„์Šทํ•œ ๋ฐฉ์‹์œผ๋กœ ์ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
mechanisms. The crowd prediction is a strong benchmark to compete with. For example, Metaculus (2023) shows that an ensemble of all forecasters consistently outperforms using just the top 5, 10,..., 30 best forecasters (based on past scores). In this work, we compare our system performance to the crowd aggregates.
๋ฉ”์ปค๋‹ˆ์ฆ˜. ๊ตฐ์ค‘ ์˜ˆ์ธก์€ ๊ฒฝ์Ÿํ•˜๊ธฐ์— ๊ฐ•๋ ฅํ•œ ๋ฒค์น˜๋งˆํฌ์ด๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ฉ”ํƒ€์ฟจ๋ฃจ์Šค(2023)๋Š” ๋ชจ๋“  ์˜ˆ๋ณด๊ด€๋“ค์˜ ์•™์ƒ๋ธ”์ด ์ƒ์œ„ 5, 10,..., 30๊ฐœ์˜ ๋ฒ ์ŠคํŠธ ์˜ˆ๋ณด๊ด€๋“ค(๊ณผ๊ฑฐ ์ ์ˆ˜ ๊ธฐ์ค€)๋งŒ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ผ๊ด€๋˜๊ฒŒ ์šฐ์›”ํ•จ์„ ๋ณด์—ฌ์ค€๋‹ค. ์ด ์ž‘์—…์—์„œ๋Š” ์‹œ์Šคํ…œ ์„ฑ๋Šฅ์„ ๊ตฐ์ค‘ ์ง‘๊ณ„์™€ ๋น„๊ตํ•œ๋‹ค.
๋ฉ”์ปค๋‹ˆ์ฆ˜. ๊ตฐ์ค‘ ์˜ˆ์ธก์€ ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ•๋ ฅํ•œ ๋ฒค์น˜๋งˆํฌ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ฉ”ํƒ€ํ˜๋Ÿฌ์Šค(2023)์— ๋”ฐ๋ฅด๋ฉด ๋ชจ๋“  ์˜ˆ์ธก์ž์˜ ์•™์ƒ๋ธ”์ด ์ƒ์œ„ 5๊ฐœ, 10๊ฐœ,..., 30๊ฐœ์˜ ์ตœ๊ณ  ์˜ˆ์ธก์ž(๊ณผ๊ฑฐ ์ ์ˆ˜ ๊ธฐ์ค€)๋งŒ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์ง€์†์ ์œผ๋กœ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…์—์„œ๋Š” ์‹œ์Šคํ…œ ์„ฑ๋Šฅ์„ ํฌ๋ผ์šฐ๋“œ ์ง‘๊ณ„์™€ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค.
eliminates the loss for undesired tokens without incurring additional costs during pretraining, making our approach both efficient and easily integrated.
์‚ฌ์ „ ํ›ˆ๋ จ ์ค‘์— ์ถ”๊ฐ€ ๋น„์šฉ์„ ๋ฐœ์ƒ์‹œํ‚ค์ง€ ์•Š๊ณ  ์›ํ•˜์ง€ ์•Š๋Š” ํ† ํฐ์— ๋Œ€ํ•œ ์†์‹ค์„ ์ œ๊ฑฐํ•˜์—ฌ ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํšจ์œจ์ ์ด๊ณ  ์‰ฝ๊ฒŒ ํ†ตํ•ฉํ•  ์ˆ˜ ์žˆ๋‹ค.
๋Š” ์‚ฌ์ „ ๊ต์œก ์ค‘ ์ถ”๊ฐ€ ๋น„์šฉ ์—†์ด ์›์น˜ ์•Š๋Š” ํ† ํฐ์˜ ์†์‹ค์„ ์ œ๊ฑฐํ•˜์—ฌ ํšจ์œจ์ ์ด๊ณ  ์‰ฝ๊ฒŒ ํ†ตํ•ฉํ•  ์ˆ˜ ์žˆ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
of two documents_) [775], or listwise ranking (_sorting a subset of candidate documents_) [777]. The essence of these methods lies in the special design of instructions for text reranking, such as sliding window strategy for document lists [774, 778], setwise selection prompting [779], fine-grained relevance labels incorporation [780], and pairwise comparison prompting [775]. In addition, recent efforts employ LLMs to generate intermediate texts (_e.g._, URLs) as retrieval results using few-shot demonstrations [781]. To further enhance the model performance, LLMs can be specially fine-tuned as backbones for reranking [782, 783] or retrieval (including dense retrieval [54] and model-based retrieval [784, 785]), similar to the fine-tuning process for traditional PLM-based IR models [782]. However, fine-tuning LLMs as IR models entails considerable expenses given the huge parameter scale of LLMs.
์˜ ๋‘ ๋ฌธ์„œ_) [775], ๋˜๋Š” ๋ฆฌ์ŠคํŠธ๋ณ„ ๋žญํ‚น(_ํ›„๋ณด ๋ฌธ์„œ์˜ ์„œ๋ธŒ์„ธํŠธ๋ฅผ ์ •๋ ฌ_) [777]. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์˜ ๋ณธ์งˆ์€ ๋ฌธ์„œ ๋ชฉ๋ก์— ๋Œ€ํ•œ ์Šฌ๋ผ์ด๋”ฉ ์œˆ๋„์šฐ ์ „๋žต[774, 778], ์„ธ๋ฐ€ํ•œ ๊ด€๋ จ์„ฑ ๋ ˆ์ด๋ธ” ํ†ตํ•ฉ[779], ์Œ๋ณ„ ๋น„๊ต ํ”„๋กฌํ”„ํŠธ[775]์™€ ๊ฐ™์€ ํ…์ŠคํŠธ ์žฌ์ˆœ์œ„ํ™”๋ฅผ ์œ„ํ•œ ๋ช…๋ น์–ด์˜ ํŠน๋ณ„ํ•œ ์„ค๊ณ„์— ์žˆ๋‹ค. ๋˜ํ•œ, ์ตœ๊ทผ์˜ ๋…ธ๋ ฅ๋“ค์€ LLMs์„ ์‚ฌ์šฉํ•˜์—ฌ ์ ์€ ์ƒท ๋ฐ๋ชจ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๋กœ์„œ ์ค‘๊ฐ„ ํ…์ŠคํŠธ๋“ค(_e.g._, URL๋“ค)์„ ์ƒ์„ฑํ•œ๋‹ค[781]. ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด LLMs๋Š” ์ „ํ†ต์ ์ธ PLM ๊ธฐ๋ฐ˜ IR ๋ชจ๋ธ[782]์— ๋Œ€ํ•œ ๋ฏธ์„ธ ์กฐ์ • ํ”„๋กœ์„ธ์Šค์™€ ์œ ์‚ฌํ•˜๊ฒŒ ์žฌ์ˆœ์œ„ํ™”[782, 783] ๋˜๋Š” ๊ฒ€์ƒ‰(๋ฐ€์ง‘ ๊ฒ€์ƒ‰[54] ๋ฐ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ๊ฒ€์ƒ‰[784, 785] ํฌํ•จ)์„ ์œ„ํ•œ ๋ฐฑ๋ณธ์œผ๋กœ์„œ ํŠน๋ณ„ํžˆ ๋ฏธ์„ธ ์กฐ์ •๋  ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ IR ๋ชจ๋ธ๋กœ์„œ LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์€ LLM์˜ ๊ฑฐ๋Œ€ํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜ ๊ทœ๋ชจ๋ฅผ ๊ณ ๋ คํ•  ๋•Œ ์ƒ๋‹นํ•œ ๋น„์šฉ์„ ์ˆ˜๋ฐ˜ํ•œ๋‹ค.
๋‘ ๋ฌธ์„œ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ_) [775], ๋˜๋Š” ๋ชฉ๋ก๋ณ„ ์ˆœ์œ„ ์ง€์ •(_ํ›„๋ณด ๋ฌธ์„œ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ ์ •๋ ฌ_) [777]. ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์˜ ํ•ต์‹ฌ์€ ๋ฌธ์„œ ๋ชฉ๋ก์— ๋Œ€ํ•œ ์Šฌ๋ผ์ด๋”ฉ ์œˆ๋„์šฐ ์ „๋žต[774, 778], ์„ธํŠธ๋ณ„ ์„ ํƒ ํ”„๋กฌํ”„ํŠธ[779], ์„ธ๋ถ„ํ™”๋œ ๊ด€๋ จ์„ฑ ๋ ˆ์ด๋ธ” ํ†ตํ•ฉ[780], ์Œ๋ณ„ ๋น„๊ต ํ”„๋กฌํ”„ํŠธ[775] ๋“ฑ ํ…์ŠคํŠธ ์žฌ๋žญํ‚น์„ ์œ„ํ•œ ์ง€์นจ์˜ ํŠน์ˆ˜ํ•œ ์„ค๊ณ„์— ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ตœ๊ทผ์˜ ๋…ธ๋ ฅ์€ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๋กœ ์ค‘๊ฐ„ ํ…์ŠคํŠธ(_์˜ˆ: URL)๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๋ช‡ ๊ฐœ์˜ ์ƒท ๋ฐ๋ชจ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค[781]. ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ „ํ†ต์ ์ธ PLM ๊ธฐ๋ฐ˜ IR ๋ชจ๋ธ์˜ ๋ฏธ์„ธ ์กฐ์ • ๊ณผ์ •[782, 783] ๋˜๋Š” ๊ฒ€์ƒ‰(๋ฐ€๋„ ๊ฒ€์ƒ‰[54] ๋ฐ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ๊ฒ€์ƒ‰[784, 785] ํฌํ•จ)์„ ์œ„ํ•œ ๋ฐฑ๋ณธ์œผ๋กœ์„œ LLM์„ ํŠน๋ณ„ํžˆ ๋ฏธ์„ธ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[782]. ๊ทธ๋Ÿฌ๋‚˜ LLM์„ IR ๋ชจ๋ธ๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋ฐ๋Š” LLM์˜ ๋ฐฉ๋Œ€ํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ทœ๋ชจ๋ฅผ ๊ณ ๋ คํ•  ๋•Œ ์ƒ๋‹นํ•œ ๋น„์šฉ์ด ์ˆ˜๋ฐ˜๋ฉ๋‹ˆ๋‹ค.
The applications of **WRAP** lies in both paradigms--(i) low-resourced data settings such as a language model for Finnish language (Luukkonen et al., 2023), and (ii) data-rich settings such as training on the common crawl. In the former, there is no alternative option of naively gathering more data, and hence, synthetic data is a natural solution that should outperform training on in-domain data alone. However, there is a significant interest in training language models on English, or more broadly, general web data. Is using synthetic data a viable option even in this paradigm?
**WRAP** ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์€ ๋‘ ํŒจ๋Ÿฌ๋‹ค์ž„ ๋ชจ๋‘์— ์žˆ์Šต๋‹ˆ๋‹ค. (i) ํ•€๋ž€๋“œ ์–ธ์–ด์— ๋Œ€ ํ•œ ์–ธ์–ด ๋ชจ๋ธ (Luukkonen ๋“ฑ, 2023)๊ณผ ๊ฐ™์€ ์ž์›์ด ๋ถ€์กฑํ•œ ๋ฐ์ดํ„ฐ ์„ค์ • ๋ฐ (ii) ์ผ๋ฐ˜์ ์ธ ํฌ๋กค๋ง์— ๋Œ€ ํ•œ ํ›ˆ๋ จ๊ณผ ๊ฐ™์€ ๋ฐ์ดํ„ฐ๊ฐ€ ํ’๋ถ€ํ•œ ์„ค์ •์ž…๋‹ˆ๋‹ค. ์ „์ž์˜ ๊ฒฝ์šฐ ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆœ์ง„ํ•˜๊ฒŒ ์ˆ˜์ง‘ํ•  ์ˆ˜ ์žˆ๋Š” ๋Œ€์•ˆ์ด ์—†๊ธฐ ๋•Œ๋ฌธ์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๋„๋ฉ”์ธ ๋‚ด ๋ฐ์ดํ„ฐ๋งŒ์œผ๋กœ ํ•™์Šต์„ ๋Šฅ๊ฐ€ํ•ด์•ผ ํ•˜๋Š” ์ž์—ฐ์Šค๋Ÿฌ์šด ์†”๋ฃจ์…˜์ด๋‹ค. ๊ทธ๋Ÿฌ๋‚˜, ์˜์–ด, ๋˜๋Š” ๋ณด๋‹ค ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ผ๋ฐ˜์ ์ธ ์›น ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๊ฒƒ์— ์ƒ๋‹นํ•œ ๊ด€์‹ฌ์ด ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ ํŒจ๋Ÿฌ๋‹ค์ž„์—์„œ๋„ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ์˜ต์…˜์ธ๊ฐ€?
WRAP**์˜ ์ ์šฉ์€ (i) ํ•€๋ž€๋“œ์–ด ์–ธ์–ด ๋ชจ๋ธ๊ณผ ๊ฐ™์ด ๋ฆฌ์†Œ์Šค๊ฐ€ ์ ์€ ๋ฐ์ดํ„ฐ ์„ค์ •(Luukkonen et al., 2023)๊ณผ (ii) ์ผ๋ฐ˜ ํฌ๋กค๋ง์— ๋Œ€ํ•œ ํ›ˆ๋ จ๊ณผ ๊ฐ™์ด ๋ฐ์ดํ„ฐ๊ฐ€ ํ’๋ถ€ํ•œ ์„ค์ •์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ํŒจ๋Ÿฌ๋‹ค์ž„์— ๋†“์—ฌ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „์ž์˜ ๊ฒฝ์šฐ, ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆœ์ง„ํ•˜๊ฒŒ ์ˆ˜์ง‘ํ•  ์ˆ˜ ์žˆ๋Š” ๋Œ€์•ˆ์ด ์—†๊ธฐ ๋•Œ๋ฌธ์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ๋„๋ฉ”์ธ ๋‚ด ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ์ž์—ฐ์Šค๋Ÿฌ์šด ์†”๋ฃจ์…˜์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์˜์–ด, ๋” ๋„“๊ฒŒ๋Š” ์ผ๋ฐ˜ ์›น ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์–ธ์–ด ๋ชจ๋ธ ํ•™์Šต์— ์ƒ๋‹นํ•œ ๊ด€์‹ฌ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํŒจ๋Ÿฌ๋‹ค์ž„์—์„œ๋„ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ์˜ต์…˜์ผ๊นŒ์š”?
Due to the difficulty of standardizing evaluations for large models and the lack of publicly available prompts and test codes for many models' evaluations, we try our best to adapt the evaluation methods to suit various model types. Specifically, we start from a standardized input prompt during testing and adjust it according to each model's appropriate input-output template. The **evaluation scripts and prompts are also open-source** in our repository, and we welcome developers to continually improve our evaluation methods.
๋Œ€ํ˜• ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ‰๊ฐ€ ํ‘œ์ค€ํ™”์˜ ์–ด๋ ค์›€๊ณผ ๋งŽ์€ ๋ชจ๋ธ์˜ ํ‰๊ฐ€์— ๋Œ€ํ•œ ๊ณต๊ฐœ ํ”„๋กฌํ”„ํŠธ ๋ฐ ํ…Œ์ŠคํŠธ ์ฝ”๋“œ์˜ ๋ถ€์กฑ์œผ๋กœ ์ธํ•ด ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์œ ํ˜•์— ๋งž๊ฒŒ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์ ์‘์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ตœ์„ ์„ ๋‹คํ•˜๊ณ  ์žˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ ํ…Œ์ŠคํŠธ ์ค‘ ํ‘œ์ค€ํ™”๋œ ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์—์„œ ์‹œ์ž‘ํ•˜์—ฌ ๊ฐ ๋ชจ๋ธ์˜ ์ ์ ˆํ•œ ์ž…๋ ฅ ์ถœ๋ ฅ ํ…œํ”Œ๋ฆฟ์— ๋”ฐ๋ผ ์กฐ์ •ํ•œ๋‹ค. ํ‰๊ฐ€ ์Šคํฌ๋ฆฝํŠธ์™€ ํ”„๋กฌํ”„ํŠธ๋Š” ๋˜ํ•œ ๋ฆฌํฌ์ง€ํ† ๋ฆฌ์˜ ์˜คํ”ˆ ์†Œ์Šค์ด๋ฉฐ ๊ฐœ๋ฐœ์ž๊ฐ€ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์ง€์†์ ์œผ๋กœ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์„ ํ™˜์˜ํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ‰๊ฐ€๋ฅผ ํ‘œ์ค€ํ™”ํ•˜๊ธฐ ์–ด๋ ต๊ณ  ๋งŽ์€ ๋ชจ๋ธ์˜ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ”„๋กฌํ”„ํŠธ์™€ ํ…Œ์ŠคํŠธ ์ฝ”๋“œ๊ฐ€ ๋ถ€์กฑํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์œ ํ˜•์— ๋งž๊ฒŒ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์กฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด ์ตœ์„ ์„ ๋‹คํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ํ…Œ์ŠคํŠธ ์‹œ ํ‘œ์ค€ํ™”๋œ ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์—์„œ ์‹œ์ž‘ํ•˜์—ฌ ๊ฐ ๋ชจ๋ธ์— ์ ํ•ฉํ•œ ์ž…์ถœ๋ ฅ ํ…œํ”Œ๋ฆฟ์— ๋”ฐ๋ผ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€ ์Šคํฌ๋ฆฝํŠธ์™€ ํ”„๋กฌํ”„ํŠธ๋Š” ์ €์žฅ์†Œ์— ์˜คํ”ˆ ์†Œ์Šค**๋กœ ๊ณต๊ฐœ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๊ฐœ๋ฐœ์ž๊ฐ€ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์ง€์†์ ์œผ๋กœ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
We show in Table 2(a) the ablation results. The top part of the table shows results for training ablations, and the bottom part is for inference ablations. We see that all components play important roles. We also observe a large performance gap between Self-Rag and No Retriever or Critic baselines across tasks, indicating that training an LM with those models largely contributes to the performance gain of Self-Rag. Using the top passages regardless of their relevance (Retrieve top 1) as in conventional RAG approaches causes a large drop in PopQA and ASQA, and removing \(\boxed{Isstr}\) during the beam search results hurts performance on ASQA. This demonstrates the effectiveness of Self-Rag's capabilities of carefully selecting generations based fine-grained multiple criterion, instead of naively using all of the top passages from the retrieval model or solely depending on relevance scores.
์šฐ๋ฆฌ๋Š” ์ ˆ์ œ ๊ฒฐ๊ณผ๋ฅผ ํ‘œ 2(a)์— ๋ณด์—ฌ์ค€๋‹ค. ํ‘œ์˜ ์ƒ๋‹จ ๋ถ€๋ถ„์€ ํ›ˆ๋ จ ์‚ญ์ œ์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋‚˜ํƒ€๋‚ด๊ณ , ํ•˜๋‹จ ๋ถ€๋ถ„์€ ์ถ”๋ก  ์‚ญ์ œ์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ชจ๋“  ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•ˆ๋‹ค. ๋˜ํ•œ ์ž‘์—… ์ „๋ฐ˜์— ๊ฑธ์ณ Self-Rag์™€ No Retriever ๋˜๋Š” Critic ๊ธฐ์ค€์„  ์‚ฌ์ด์˜ ํฐ ์„ฑ๋Šฅ ๊ฒฉ์ฐจ๋ฅผ ๊ด€์ฐฐํ•˜์—ฌ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ๋กœ LM์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ด Self-Rag์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์— ํฌ๊ฒŒ ๊ธฐ์—ฌํ•จ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค. ๊ธฐ์กด์˜ RAG ์ ‘๊ทผ๋ฒ•๊ณผ ๊ฐ™์ด ์ƒ์œ„ ํ†ต๋กœ๋ฅผ ์—ฐ๊ด€์„ฑ์— ๊ด€๊ณ„์—†์ด ์‚ฌ์šฉ(์ƒ์œ„ 1 ๊ฒ€์ƒ‰)ํ•˜๋ฉด PopQA์™€ ASQA๊ฐ€ ํฌ๊ฒŒ ๋–จ์–ด์ง€๊ณ , ๋น” ํƒ์ƒ‰ ๊ณผ์ •์—์„œ \(\boxed{Isstr}\)์„ ์ œ๊ฑฐํ•˜๋ฉด ASQA์˜ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋œ๋‹ค. ์ด๋Š” ๊ฒ€์ƒ‰ ๋ชจ๋ธ์˜ ๋ชจ๋“  ์ƒ์œ„ ๊ตฌ์ ˆ์„ ์ˆœ์ง„ํ•˜๊ฒŒ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ ๊ด€๋ จ์„ฑ ์ ์ˆ˜์—๋งŒ ์˜์กดํ•˜๋Š” ๋Œ€์‹  ์„ธ๋ฐ€ํ•˜๊ฒŒ ์„ธ๋ถ„ํ™”๋œ ๋‹ค์ค‘ ๊ธฐ์ค€์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์„ธ๋Œ€๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ ํƒํ•˜๋Š” Self-Rag์˜ ๋Šฅ๋ ฅ์˜ ํšจ์œจ์„ฑ์„ ๋ณด์—ฌ์ค€๋‹ค.
ํ‘œ 2(a)์—๋Š” ์ ˆ์ œ ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ํ‘œ์˜ ์œ„์ชฝ์€ ํ›ˆ๋ จ ์ œ๊ฑฐ์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ์ด๊ณ  ์•„๋ž˜์ชฝ์€ ์ถ”๋ก  ์ œ๊ฑฐ์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ๋ชจ๋“  ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ฌ๋Ÿฌ ์ž‘์—…์—์„œ ์…€ํ”„ ๋ž˜๊ทธ์™€ ๋…ธ ๋ฆฌํŠธ๋ฆฌ๋ฒ„ ๋˜๋Š” ํฌ๋ฆฌํ‹ฑ ๊ธฐ์ค€์„  ์‚ฌ์ด์— ํฐ ์„ฑ๋Šฅ ๊ฒฉ์ฐจ๊ฐ€ ์žˆ์Œ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ LM์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ด ์…€ํ”„ ๋ž˜๊ทธ์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์— ํฌ๊ฒŒ ๊ธฐ์—ฌํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ธฐ์กด RAG ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ์™€ ๊ฐ™์ด ๊ด€๋ จ์„ฑ์— ๊ด€๊ณ„์—†์ด ์ƒ์œ„ ๊ตฌ์ ˆ์„ ์‚ฌ์šฉํ•˜๋ฉด(Retrieve top 1) PopQA์™€ ASQA๊ฐ€ ํฌ๊ฒŒ ๋–จ์–ด์ง€๊ณ , ๋น” ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ ์ค‘ \(\boxed{Isstr}\) ์ œ๊ฑฐ๋Š” ASQA์˜ ์„ฑ๋Šฅ์„ ๋–จ์–ด๋œจ๋ฆฝ๋‹ˆ๋‹ค. ์ด๋Š” ๊ฒ€์ƒ‰ ๋ชจ๋ธ์—์„œ ์ƒ์œ„ ๊ตฌ์ ˆ์„ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ ๊ด€๋ จ์„ฑ ์ ์ˆ˜์—๋งŒ ์˜์กดํ•˜๋Š” ๋Œ€์‹  ์„ธ๋ถ„ํ™”๋œ ๋‹ค์ค‘ ๊ธฐ์ค€์— ๋”ฐ๋ผ ์„ธ๋Œ€๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ ํƒํ•˜๋Š” Self-Rag์˜ ๊ธฐ๋Šฅ์ด ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
Mathematically, let \(\mathcal{Q}=\{q_{n}\}_{n=1}^{N}\) be a set of \(N\) multiple choice factual questions, where each question has \(L\) possible answers and exactly one correct answer. Let \(\mathcal{A}=\{(a_{n}^{1},\ldots,a_{n}^{L})\}_{n=1}^{N}\) be the corresponding set of possible answers, and \(\mathcal{C}=\{c_{n}\}_{n=1}^{N}\) be the correct answers.
์ˆ˜ํ•™์ ์œผ๋กœ, \(\mathcal{Q}=\{q_{n}\}_{n=1}^{N}\)์„ \(N\) ๊ฐ๊ด€์‹ ์‚ฌ์‹ค๋ฌธํ•ญ์˜ ์ง‘ํ•ฉ์œผ๋กœ ํ•˜์ž. ์—ฌ๊ธฐ์„œ ๊ฐ ๋ฌธํ•ญ์€ \(L\) ๊ฐ€๋Šฅํ•œ ๋‹ต๊ณผ ์ •ํ™•ํžˆ ํ•˜๋‚˜์˜ ์ •๋‹ต์„ ๊ฐ–๋Š”๋‹ค. \(\mathcal{A}=\{(a_{n}^{1},\ldots,a_{n}^{L})\}_{n=1}^{N}\)์ด ๊ฐ€๋Šฅํ•œ ๋‹ต์˜ ์ง‘ํ•ฉ์ด๊ณ , \(\mathcal{C}=\{c_{n}\}_{n=1}^{N}\)์ด ์ •๋‹ต์ด๋ผ๊ณ  ํ•˜์ž.
์ˆ˜ํ•™์ ์œผ๋กœ \(\mathcal{Q}=\{q_{n}\}_{n=1}^{N}\)๋ฅผ ๊ฐ๊ด€์‹ ์‚ฌ์‹ค ๋ฌธ์ œ \(N\)์˜ ์ง‘ํ•ฉ์œผ๋กœ ํ•˜๊ณ , ๊ฐ ๋ฌธ์ œ์—๋Š” \(L\) ๊ฐœ์˜ ๊ฐ€๋Šฅํ•œ ๋‹ต๊ณผ ์ •ํ™•ํžˆ ํ•˜๋‚˜์˜ ์ •๋‹ต์ด ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ด…์‹œ๋‹ค. (\mathcal{A}=\{(a_{n}^{1},\ldots,a_{n}^{L})\}_{n=1}^{N}\)๋ฅผ ํ•ด๋‹น ๊ฐ€๋Šฅํ•œ ๋‹ต์˜ ์ง‘ํ•ฉ์œผ๋กœ, \(\mathcal{C}=\{c_{n}\}_{n=1}^{N}\)๋ฅผ ์ •๋‹ต์ด๋ผ๊ณ  ํ•˜์ž.
**Order effect of language models.** Some prior works show that language models are able to understand permuted texts to some extent, i.e., after a random permutation of words, models usually preserve a reasonable performance (Abdou et al., 2022; Sinha et al., 2020). Moreover, Cao et al. (2023) shows that even when a large fraction of words are scrambled, GPT-4 still achieves decent performance on several reasoning benchmarks. In contrast to permuted texts in these works that are typically unnatural and nonsensical, our premise order permutations do not alter the semantic meaning and remain syntactically valid (we manually verify this). Nevertheless, we demonstrate that LLM reasoning performance is highly brittle to the ordering of the premises.
**์–ธ์–ด ๋ชจ๋ธ์˜ ์ˆœ์„œ ํšจ๊ณผ** ์ผ๋ถ€ ์ด์ „ ์ž‘์—…์€ ์–ธ์–ด ๋ชจ๋ธ์ด ์ˆœ์—ด๋œ ํ…์ŠคํŠธ๋ฅผ ์–ด๋Š ์ •๋„ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ฆ‰ ๋‹จ์–ด์˜ ๋ฌด์ž‘์œ„ ์ˆœ์—ด ํ›„์— ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•ฉ๋ฆฌ์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์กดํ•œ๋‹ค(Abdou ๋“ฑ, 2022; Sinha ๋“ฑ, 2020). ๋”์šฑ์ด Cao et al. (2023)์€ ๋‹จ์–ด์˜ ๋งŽ์€ ๋ถ€๋ถ„์ด ์Šคํฌ๋žจ๋ธ”๋œ ๊ฒฝ์šฐ์—๋„ GPT-4๋Š” ์—ฌ๋Ÿฌ ์ถ”๋ก  ๋ฒค์น˜๋งˆํฌ์—์„œ ์—ฌ์ „ํžˆ ๊ดœ์ฐฎ์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค€๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋ถ€์ž์—ฐ์Šค๋Ÿฝ๊ณ  ๋ฌด์˜๋ฏธํ•œ ์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ ์ˆœ์—ด๋œ ํ…์ŠคํŠธ์™€ ๋Œ€์กฐ์ ์œผ๋กœ, ์šฐ๋ฆฌ์˜ ์ „์ œ ์ˆœ์„œ ์ˆœ์—ด์€ ์˜๋ฏธ์  ์˜๋ฏธ๋ฅผ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š๊ณ  ๊ตฌ๋ฌธ์ ์œผ๋กœ ์œ ํšจํ•œ ์ƒํƒœ๋กœ ์œ ์ง€๋œ๋‹ค(์šฐ๋ฆฌ๋Š” ์ด๊ฒƒ์„ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์ฆํ•œ๋‹ค). ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์šฐ๋ฆฌ๋Š” LLM ์ถ”๋ก  ์„ฑ๋Šฅ์ด ์ „์ œ์˜ ์ˆœ์„œ์— ๋งค์šฐ ์ทจ์•ฝํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค€๋‹ค.
**์–ธ์–ด ๋ชจ๋ธ์˜ ์ˆœ์„œ ํšจ๊ณผ** ์ผ๋ถ€ ์„ ํ–‰ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์–ธ์–ด ๋ชจ๋ธ์€ ๋ฌด์ž‘์œ„๋กœ ๋‹จ์–ด๊ฐ€ ๋’ค์„ž์ธ ํ…์ŠคํŠธ๋ฅผ ์–ด๋Š ์ •๋„ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฆ‰, ๋‹จ์–ด๊ฐ€ ๋ฌด์ž‘์œ„๋กœ ์ˆœ์—ด๋œ ํ›„์—๋„ ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•ฉ๋ฆฌ์ ์ธ ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค(Abdou et al., 2022; Sinha et al., 2020). ๋˜ํ•œ, Cao ๋“ฑ(2023)์€ ๋‹จ์–ด์˜ ์ƒ๋‹น ๋ถ€๋ถ„์ด ์Šคํฌ๋žจ๋ธ”๋œ ๊ฒฝ์šฐ์—๋„ GPT-4๊ฐ€ ์—ฌ๋Ÿฌ ์ถ”๋ก  ๋ฒค์น˜๋งˆํฌ์—์„œ ์—ฌ์ „ํžˆ ์ ์ ˆํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ์—์„œ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ถ€์ž์—ฐ์Šค๋Ÿฝ๊ณ  ๋ฌด์˜๋ฏธํ•œ ์ˆœ์—ด ํ…์ŠคํŠธ์™€ ๋‹ฌ๋ฆฌ, ์šฐ๋ฆฌ์˜ ์ „์ œ ์ˆœ์„œ ์ˆœ์—ด์€ ์˜๋ฏธ๋ก ์  ์˜๋ฏธ๋ฅผ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š์œผ๋ฉฐ ๊ตฌ๋ฌธ๋ก ์ ์œผ๋กœ ์œ ํšจํ•ฉ๋‹ˆ๋‹ค(์ด๋ฅผ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค). ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์šฐ๋ฆฌ๋Š” LLM ์ถ”๋ก  ์„ฑ๋Šฅ์ด ์ „์ œ ์ˆœ์„œ์— ๋งค์šฐ ์ทจ์•ฝํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.
We collect five conversational QA datasets with long documents which cannot be directly fitted into LLMs with a sequence length of 4K tokens. Hence, we run a multi-turn retriever to get top-5 relevant chunks as the inputs (experiments can be found in SS4.3).
4K ํ† ํฐ์˜ ์‹œํ€€์Šค ๊ธธ์ด๋ฅผ ๊ฐ–๋Š” LLMs์— ์ง์ ‘ ํ”ผํŒ…๋  ์ˆ˜ ์—†๋Š” ๊ธด ๋ฌธ์„œ๋ฅผ ๊ฐ–๋Š” 5๊ฐœ์˜ ๋Œ€ํ™” QA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜์ง‘ํ•œ๋‹ค. ๋”ฐ๋ผ์„œ ๋ฉ€ํ‹ฐํ„ด ๋ฆฌํŠธ๋ฆฌ๋ฒ„๋ฅผ ์‹คํ–‰ํ•˜์—ฌ ์ƒ์œ„ 5๊ฐœ์˜ ๊ด€๋ จ ์ฒญํฌ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ํ•œ๋‹ค(์‹คํ—˜์€ SS4.3์—์„œ ์ฐพ์„ ์ˆ˜ ์žˆ์Œ).
์šฐ๋ฆฌ๋Š” 4K ํ† ํฐ์˜ ์‹œํ€€์Šค ๊ธธ์ด๋ฅผ ๊ฐ€์ง„ LLM์— ์ง์ ‘ ๋งž์ถœ ์ˆ˜ ์—†๋Š” ๊ธด ๋ฌธ์„œ๊ฐ€ ํฌํ•จ๋œ 5๊ฐœ์˜ ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ฉ€ํ‹ฐํ„ด ๋ฆฌํŠธ๋ฆฌ๋ฒ„๋ฅผ ์‹คํ–‰ํ•˜์—ฌ ์ƒ์œ„ 5๊ฐœ ๊ด€๋ จ ์ฒญํฌ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค(์‹คํ—˜์€ SS4.3์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค).
Traditional Chinese outputs. Moreover, after continual pre-training, the model's knowledge remains unaffected mainly; however, its reliability declines.
์ „ํ†ต์ ์ธ ์ค‘๊ตญ์‹ ์‚ฐ์ถœ๋ฌผ. ๋˜ํ•œ, ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ ํ›„์—๋„ ๋ชจ๋ธ์˜ ์ง€์‹์€ ์ฃผ๋กœ ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š์ง€๋งŒ ์‹ ๋ขฐ์„ฑ์€ ๊ฐ์†Œํ•œ๋‹ค.
์ค‘๊ตญ์–ด ๋ฒˆ์ฒด ์ถœ๋ ฅ. ๋˜ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ํ›„์—๋„ ๋ชจ๋ธ์˜ ์ง€์‹์€ ์ฃผ๋กœ ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š์ง€๋งŒ ์‹ ๋ขฐ๋„๋Š” ๋–จ์–ด์ง‘๋‹ˆ๋‹ค.
This technical report presents our first release of OLMo, a state-of-the-art, truly open language model and its framework to build and study the science of language modeling. Unlike most prior efforts that have only released model weights and inference code, we release OLMo and the whole framework, including training data and training and evaluation code. Soon, we will also release training logs, ablations, findings and Weights & Biases logs. We are also exploring the adaptation of OLMo with instruction tuning and different flavors of RLHF. We are going to release the adapted models as well as all of our model adaptation code and data.
์ด ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ๋ง์˜ ๊ณผํ•™์„ ๊ตฌ์ถ•ํ•˜๊ณ  ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ฒจ๋‹จ ์ง„์ •ํ•œ ๊ฐœ๋ฐฉํ˜• ์–ธ์–ด ๋ชจ๋ธ์ธ OLMo์˜ ์ฒซ ๋ฒˆ์งธ ์ถœ์‹œ๋ฅผ ์ œ์‹œํ•œ๋‹ค. ๊ธฐ์กด์˜ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜ ๋ฐ ์ถ”๋ก  ์ฝ”๋“œ๋งŒ์„ ๋ฐœํ‘œํ•˜๋˜ ๊ธฐ์กด ์—ฐ๊ตฌ๋“ค๊ณผ๋Š” ๋‹ฌ๋ฆฌ OLMo์™€ ํ•™์Šต ๋ฐ์ดํ„ฐ, ํ•™์Šต ๋ฐ ํ‰๊ฐ€ ์ฝ”๋“œ๋ฅผ ํฌํ•จํ•œ ์ „์ฒด ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋ฐœํ‘œํ•œ๋‹ค. ๊ณง, ์šฐ๋ฆฌ๋Š” ๋˜ํ•œ ํ›ˆ๋ จ ๋กœ๊ทธ, ์ ˆ์ œ, ๋ฐœ๊ฒฌ ๋ฐ ๊ฐ€์ค‘์น˜ ๋ฐ ํŽธํ–ฅ ๋กœ๊ทธ๋ฅผ ๊ณต๊ฐœํ•  ๊ฒƒ์ด๋‹ค. ๋˜ํ•œ RLHF์˜ ๋ช…๋ น์–ด ์กฐ์ • ๋ฐ ๋‹ค์–‘ํ•œ ๋ง›์„ ์‚ฌ์šฉํ•˜์—ฌ OLMo์˜ ์ ์‘์„ ํƒ์ƒ‰ํ•˜๊ณ  ์žˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ชจ๋“  ๋ชจ๋ธ ์ ์‘ ์ฝ”๋“œ์™€ ๋ฐ์ดํ„ฐ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ ์‘๋œ ๋ชจ๋ธ์„ ์ถœ์‹œํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
์ด ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ๋ง์˜ ๊ณผํ•™์„ ๊ตฌ์ถ•ํ•˜๊ณ  ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์‹ ์˜ ์ง„์ •ํ•œ ๊ฐœ๋ฐฉํ˜• ์–ธ์–ด ๋ชจ๋ธ์ด์ž ํ”„๋ ˆ์ž„์›Œํฌ์ธ OLMo์˜ ์ฒซ ๋ฒˆ์งธ ๋ฆด๋ฆฌ์Šค๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ ๊ฐ€์ค‘์น˜์™€ ์ถ”๋ก  ์ฝ”๋“œ๋งŒ ๊ณต๊ฐœํ–ˆ๋˜ ๋Œ€๋ถ€๋ถ„์˜ ์ด์ „ ์ž‘์—…๊ณผ๋Š” ๋‹ฌ๋ฆฌ, ํ•™์Šต ๋ฐ์ดํ„ฐ์™€ ํ•™์Šต ๋ฐ ํ‰๊ฐ€ ์ฝ”๋“œ๋ฅผ ํฌํ•จํ•œ OLMo์™€ ์ „์ฒด ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๊ณง ํ›ˆ๋ จ ๋กœ๊ทธ, ์ ˆ์ œ, ๊ฒฐ๊ณผ, ๊ฐ€์ค‘์น˜ ๋ฐ ํŽธํ–ฅ ๋กœ๊ทธ๋„ ๊ณต๊ฐœํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ธ์ŠคํŠธ๋Ÿญ์…˜ ํŠœ๋‹๊ณผ ๋‹ค์–‘ํ•œ ๋ฒ„์ „์˜ RLHF๋ฅผ ํ†ตํ•ด OLMo์˜ ์ ์‘์„ ๋ชจ์ƒ‰ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ ์‘๋œ ๋ชจ๋ธ๊ณผ ๋ชจ๋“  ๋ชจ๋ธ ์ ์‘ ์ฝ”๋“œ ๋ฐ ๋ฐ์ดํ„ฐ๋„ ๊ณต๊ฐœํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.
CoT prompting is first proposed as an extension of ICL [33], which augments each demonstration \(\langle\)_input, output_\(\rangle\) as \(\langle\)_input, CoT, output_\(\rangle\). A _CoT_ is a series of intermediate reasoning steps for connecting the _input_ and _output_. With these augmented demonstrations, LLMs can follow them to generate CoTs and the answer for a new input. However, unlike \(\langle\)_input, output_\(\rangle\) pairs in ICL, CoTs are difficult to obtain and usually require human annotation. Fortunately, it has been found that LLMs can be triggered to generate CoTs through simple instructions like _"Let's think step by step."_[505], making CoT prompting easy to use. There are also alternative magic prompts that can elicit the ability of CoT reasoning and further improve the performance of LLMs, such as _"Take a deep breath and work on this problem step-by-step."_[473].
CoT ํ”„๋กฌํ”„ํŠธ๋Š” ๋จผ์ € ICL [33]์˜ ํ™•์žฅ์œผ๋กœ ์ œ์•ˆ๋˜๋ฉฐ, ์ด๋Š” ๊ฐ ๋ฐ๋ชจ \(\langle\)_input, output_\(\rangle\)๋ฅผ \(\langle\)_input, CoT, output_\(\rangle\)๋กœ ์ฆ๊ฐ€์‹œํ‚จ๋‹ค. A _CoT_ ๋Š” _input_ ๊ณผ _output_ ์„ ์—ฐ๊ฒฐ ํ•˜๊ธฐ ์œ„ํ•œ ์ผ๋ จ์˜ ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฆ๊ฐ•๋œ ๋ฐ๋ชจ๋ฅผ ํ†ตํ•ด LLM์€ ์ด๋ฅผ ๋”ฐ๋ผ CoT๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ์ƒˆ๋กœ์šด ์ž…๋ ฅ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ICL์—์„œ \(\langle\)_input, output_\(\rangle\) pairs์™€ ๋‹ฌ๋ฆฌ CoT๋Š” ๊ตฌํ•˜๊ธฐ ์–ด๋ ต๊ณ  ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ๋žŒ์˜ ์ฃผ์„์ด ํ•„์š”ํ•˜๋‹ค. ๋‹คํ–‰ํžˆ๋„ LLM์€ _"์ฐจ๊ทผ์ฐจ๊ทผ ์ƒ๊ฐํ•ด๋ณด์ž."_[505]์™€ ๊ฐ™์€ ๊ฐ„๋‹จํ•œ ์ง€์นจ์„ ํ†ตํ•ด CoT๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ํŠธ๋ฆฌ๊ฑฐ๋  ์ˆ˜ ์žˆ์–ด CoT๊ฐ€ ์‰ฝ๊ฒŒ ์‚ฌ์šฉํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•  ์ˆ˜ ์žˆ๋‹ค. CoT ์ถ”๋ก ์˜ ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๊ณ  LLM์˜ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๋Œ€์•ˆ์ ์ธ ๋งค์ง ํ”„๋กฌํ”„ํŠธ๋„ ์žˆ๋Š”๋ฐ, ์ด๋ฅผํ…Œ๋ฉด _"์‹ฌํ˜ธํก์„ ํ•˜๊ณ  ์ด ๋ฌธ์ œ์— ๋Œ€ํ•ด ๋‹จ๊ณ„์ ์œผ๋กœ ์ž‘์—…ํ•œ๋‹ค."_[473].
CoT ํ”„๋กฌํ”„ํŠธ๋Š” ICL[33]์˜ ํ™•์žฅ์œผ๋กœ ์ฒ˜์Œ ์ œ์•ˆ๋˜์—ˆ์œผ๋ฉฐ, ๊ฐ ๋ฐ๋ชจ \(\langle\)_์ž…๋ ฅ, ์ถœ๋ ฅ_\(\rangle\)์„ \(\langle\)_์ž…๋ ฅ, CoT, ์ถœ๋ ฅ_\(\rangle\)์œผ๋กœ ๋ณด๊ฐ•ํ•ฉ๋‹ˆ๋‹ค. CoT_๋Š” _์ž…๋ ฅ_๊ณผ _์ถœ๋ ฅ_์„ ์—ฐ๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ์ผ๋ จ์˜ ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฆ๊ฐ• ๋ฐ๋ชจ๋ฅผ ํ†ตํ•ด LLM์€ ์ด๋ฅผ ๋”ฐ๋ผ CoT์™€ ์ƒˆ๋กœ์šด ์ž…๋ ฅ์— ๋Œ€ํ•œ ๋‹ต์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ICL์˜ \(\langle\)_์ž…๋ ฅ, ์ถœ๋ ฅ_\(\rangle\) ์Œ๊ณผ ๋‹ฌ๋ฆฌ CoT๋Š” ๊ตฌํ•˜๊ธฐ ์–ด๋ ต๊ณ  ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ๋žŒ์˜ ์ฃผ์„์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‹คํ–‰ํžˆ๋„ _"๋‹จ๊ณ„๋ณ„๋กœ ์ƒ๊ฐํ•ด ๋ด…์‹œ๋‹ค."_[505]์™€ ๊ฐ™์€ ๊ฐ„๋‹จํ•œ ๋ช…๋ น์–ด๋ฅผ ํ†ตํ•ด LLM์„ ํŠธ๋ฆฌ๊ฑฐํ•˜์—ฌ CoT๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์ ธ CoT ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ๊ฐ€ ์‰ฌ์›Œ์กŒ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ _"์‹ฌํ˜ธํก์„ ํ•˜๊ณ  ์ด ๋ฌธ์ œ๋ฅผ ์ฐจ๊ทผ์ฐจ๊ทผ ํ’€์–ด๋ณด์„ธ์š”."_[473]์™€ ๊ฐ™์ด CoT ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ์ด๋Œ์–ด๋‚ด๊ณ  LLM์˜ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๋Œ€์ฒด ๋งˆ๋ฒ• ํ”„๋กฌํ”„ํŠธ๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
Finally, we compare the mean tree depth (measured by the mean over sentences of the depth of the dependency tree), and mean dependency distance (measured as the average dependency distance of any pair of words within a sentence) in Figure 11, which have been shown to be good measures of syntactic difficulty Futrell et al. (2015); Gibson et al. (2000); Oya (2021). We find similar trends as for reading level and TTR diversity where mediums style increase depth, mdd, and syntactic complexity in general. We find again that QA style reduces this complexity.
๋งˆ์ง€๋ง‰์œผ๋กœ, ๊ตฌ๋ฌธ ๋‚œ์ด๋„ Futrell ๋“ฑ(2015), Gibson ๋“ฑ(2000), Oya(2021)์˜ ์ข‹์€ ์ฒ™๋„๋กœ ๋‚˜ํƒ€๋‚œ ๊ทธ๋ฆผ 11์˜ ํ‰๊ท  ํŠธ๋ฆฌ ๊นŠ์ด(์˜์กด ํŠธ๋ฆฌ ๊นŠ์ด์˜ ๋ฌธ์žฅ์— ๋Œ€ํ•œ ํ‰๊ท ์œผ๋กœ ์ธก์ •๋จ)์™€ ํ‰๊ท  ์˜์กด ๊ฑฐ๋ฆฌ(๋ฌธ์žฅ ๋‚ด์˜ ๋ชจ๋“  ๋‹จ์–ด ์Œ์˜ ํ‰๊ท  ์˜์กด ๊ฑฐ๋ฆฌ๋กœ ์ธก์ •๋จ)๋ฅผ ๋น„๊ตํ•œ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋งค์ฒด ์Šคํƒ€์ผ์ด ๊นŠ์ด, mdd ๋ฐ ๊ตฌ๋ฌธ ๋ณต์žก์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ค๋Š” ์ฝ๊ธฐ ์ˆ˜์ค€ ๋ฐ TTR ๋‹ค์–‘์„ฑ๊ณผ ์œ ์‚ฌํ•œ ๊ฒฝํ–ฅ์„ ๋ฐœ๊ฒฌํ•œ๋‹ค. ์šฐ๋ฆฌ๋Š” QA ์Šคํƒ€์ผ์ด ์ด๋Ÿฌํ•œ ๋ณต์žก์„ฑ์„ ๊ฐ์†Œ์‹œํ‚จ๋‹ค๋Š” ๊ฒƒ์„ ๋‹ค์‹œ ๋ฐœ๊ฒฌํ•œ๋‹ค.
๋งˆ์ง€๋ง‰์œผ๋กœ, ๊ทธ๋ฆผ 11์˜ ํ‰๊ท  ํŠธ๋ฆฌ ๊นŠ์ด(์˜์กด์„ฑ ํŠธ๋ฆฌ ๊นŠ์ด์˜ ๋ฌธ์žฅ ์ „์ฒด ํ‰๊ท ์œผ๋กœ ์ธก์ •)์™€ ํ‰๊ท  ์˜์กด์„ฑ ๊ฑฐ๋ฆฌ(๋ฌธ์žฅ ๋‚ด ๋ชจ๋“  ๋‹จ์–ด ์Œ์˜ ํ‰๊ท  ์˜์กด์„ฑ ๊ฑฐ๋ฆฌ๋กœ ์ธก์ •)๋ฅผ ๋น„๊ตํ•˜๋Š”๋ฐ, ์ด๋Š” ๊ตฌ๋ฌธ ๋‚œ์ด๋„๋ฅผ ์ธก์ •ํ•˜๋Š” ์ข‹์€ ์ฒ™๋„๋กœ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค Futrell ์™ธ(2015), Gibson ์™ธ(2000), Oya(2021). ์ฝ๊ธฐ ์ˆ˜์ค€ ๋ฐ TTR ๋‹ค์–‘์„ฑ์—์„œ๋„ ๋งค์ฒด ์Šคํƒ€์ผ์ด ์ผ๋ฐ˜์ ์œผ๋กœ ๊นŠ์ด, ๋‹ค์ค‘, ๊ตฌ๋ฌธ ๋ณต์žก์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ค๋Š” ์œ ์‚ฌํ•œ ๊ฒฝํ–ฅ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. QA ์Šคํƒ€์ผ์€ ์ด๋Ÿฌํ•œ ๋ณต์žก์„ฑ์„ ๊ฐ์†Œ์‹œํ‚จ๋‹ค๋Š” ๊ฒƒ์„ ๋‹ค์‹œ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
- Engaging with opponents: High. Trump is known for his confrontational style, which could play out to his advantage in a debate setting.
- ์ƒ๋Œ€์™€ ๊ต์ „ ์ค‘: ํ•˜์ด. ํŠธ๋Ÿผํ”„๋Š” ํ† ๋ก  ํ™˜๊ฒฝ์—์„œ ์œ ๋ฆฌํ•˜๊ฒŒ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋Œ€๊ฒฐ ์Šคํƒ€์ผ๋กœ ์œ ๋ช…ํ•ฉ๋‹ˆ๋‹ค.
- ์ƒ๋Œ€์™€ ๊ต์ „: ๋†’์Œ. ํŠธ๋Ÿผํ”„๋Š” ๋Œ€๋ฆฝ์ ์ธ ์Šคํƒ€์ผ๋กœ ์ž˜ ์•Œ๋ ค์ ธ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ† ๋ก  ํ™˜๊ฒฝ์—์„œ ์œ ๋ฆฌํ•˜๊ฒŒ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**ColossalAI** library [226] provides a collection of parallel components. It aims to support developers to write their distributed deep learning models just like how they write their model on their laptop. They provide user-friendly tools to kickstart distributed training and inference in a few lines. In terms of Parallelism strategies, they support: Data Parallelism, Pipeline Parallelism, Sequence Parallelism, Zero Redundancy Optimizer (ZeRO) [140], and Auto-Parallelism.
**ColossalAI** ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ [226]์€ ๋ณ‘๋ ฌ ๊ตฌ์„ฑ ์š”์†Œ์˜ ์ปฌ๋ ‰์…˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ฐœ๋ฐœ์ž๊ฐ€ ๋…ธํŠธ๋ถ์— ๋ชจ๋ธ์„ ์“ฐ๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๋ถ„์‚ฐ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ์“ธ ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•œ๋‹ค. ๊ทธ๋“ค์€ ๋ช‡ ์ค„๋กœ ๋ถ„์‚ฐ๋œ ํ›ˆ๋ จ๊ณผ ์ถ”๋ก ์„ ์‹œ์ž‘ํ•  ์ˆ˜ ์žˆ๋Š” ์‚ฌ์šฉ์ž ์นœํ™”์ ์ธ ๋„๊ตฌ๋ฅผ ์ œ๊ณตํ•œ๋‹ค. ๋ณ‘๋ ฌ์„ฑ ์ „๋žต ์ธก๋ฉด์—์„œ ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ์„ฑ, ํŒŒ์ดํ”„๋ผ์ธ ๋ณ‘๋ ฌ์„ฑ, ์‹œํ€€์Šค ๋ณ‘๋ ฌ์„ฑ, ZeRO(Zero Redundancy Optimizer) [140] ๋ฐ ์ž๋™ ๋ณ‘๋ ฌ์„ฑ์„ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค.
**ColossalAI** ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ[226]๋Š” ๋ณ‘๋ ฌ ๊ตฌ์„ฑ ์š”์†Œ ๋ชจ์Œ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋Š” ๊ฐœ๋ฐœ์ž๊ฐ€ ๋…ธํŠธ๋ถ์—์„œ ๋ชจ๋ธ์„ ์ž‘์„ฑํ•˜๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๋ถ„์‚ฐํ˜• ๋”ฅ ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋Š” ๋ช‡ ์ค„๋กœ ๋ถ„์‚ฐ ํ•™์Šต๊ณผ ์ถ”๋ก ์„ ์‹œ์ž‘ํ•  ์ˆ˜ ์žˆ๋Š” ์‚ฌ์šฉ์ž ์นœํ™”์ ์ธ ๋„๊ตฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ ์ „๋žต ์ธก๋ฉด์—์„œ๋Š” ๋‹ค์Œ์„ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค: ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ, ํŒŒ์ดํ”„๋ผ์ธ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ, ์‹œํ€€์Šค ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ, ์ œ๋กœ ์ค‘๋ณต ์ตœ์ ํ™”(ZeRO)[140], ์ž๋™ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๋ฅผ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค.