question
stringlengths
101
2.69k
chosen
stringlengths
43
2.27k
rejected
stringlengths
42
2.18k
Scaling inverter.We train inverter models of varying size to see the effect of model scale on inversion performance. We note that the number of parameters in the inverter has a very large impact; with larger inverter models performing significantly better: under ablation settings, with the same number of training steps, the T5-large inverter achieves \(24\%\) higher BLEU score than T5-small. This finding indicates that we may more accurately invert language models simply by scaling our system.
์Šค์ผ€์ผ๋ง ์ธ๋ฒ„ํ„ฐ. ์šฐ๋ฆฌ๋Š” ์ธ๋ฒ„ํ„ฐ ๋ชจ๋ธ์˜ ํฌ๊ธฐ๋ฅผ ๋‹ค์–‘ํ•˜๊ฒŒ ์กฐ์ •ํ•˜์—ฌ ๋ชจ๋ธ ์Šค์ผ€์ผ์ด ์ธ๋ฒ„์ „ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ธ๋ฒ„ํ„ฐ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ๋งค์šฐ ํฐ ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค๋Š” ์ ์„ ์ฃผ๋ชฉํ•ฉ๋‹ˆ๋‹ค. ๋” ํฐ ์ธ๋ฒ„ํ„ฐ ๋ชจ๋ธ์ด ํ›จ์”ฌ ๋” ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋ฉฐ, ๋™์ผํ•œ ํ•™์Šต ๋‹จ๊ณ„ ์ˆ˜์—์„œ T5-large ์ธ๋ฒ„ํ„ฐ๋Š” T5-small๋ณด๋‹ค \(24\%\) ๋” ๋†’์€ BLEU ์ ์ˆ˜๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐœ๊ฒฌ์€ ๋‹จ์ˆœํžˆ ์‹œ์Šคํ…œ์„ ์Šค์ผ€์ผ๋งํ•จ์œผ๋กœ์จ ์–ธ์–ด ๋ชจ๋ธ์„ ๋” ์ •ํ™•ํ•˜๊ฒŒ ์ธ๋ฒ„ํŠธํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
์Šค์ผ€์ผ๋ง ์ธ๋ฒ„ํ„ฐ. ๋‹ค์–‘ํ•œ ํฌ๊ธฐ์˜ ์ธ๋ฒ„ํ„ฐ ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผœ ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ๋ฐ˜์ „ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ธ๋ฒ„ํ„ฐ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ๋งค์šฐ ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋ฉฐ, ๋” ํฐ ์ธ๋ฒ„ํ„ฐ ๋ชจ๋ธ์ด ํ›จ์”ฌ ๋” ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ์ ˆ์ œ ์„ค์ •์—์„œ ๋™์ผํ•œ ์ˆ˜์˜ ํ›ˆ๋ จ ๋‹จ๊ณ„๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ T5-large ์ธ๋ฒ„ํ„ฐ๋Š” T5-small๋ณด๋‹ค \(24\%\) ๋” ๋†’์€ BLEU ์ ์ˆ˜๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ๋‹จ์ˆœํžˆ ์‹œ์Šคํ…œ์„ ํ™•์žฅํ•จ์œผ๋กœ์จ ์–ธ์–ด ๋ชจ๋ธ์„ ๋” ์ •ํ™•ํ•˜๊ฒŒ ๋ฐ˜์ „์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
The AI community has witnessed the release of several open-source LLMs, where the model weights are made publicly available. EleutherAI has contributed GPT-NeoX-20B [35] and GPT-J-GB [36]. Google has released UL2-20B [37]. Tsinghua University has introduced GLM-130B [7]. Meta has released OPT [9] and LLaMA [8]. It is worth noting that while these open-source models have made valuable contributions, they generally do not exhibit the same level of performance as their closed-source counterparts.
AI ์ปค๋ฎค๋‹ˆํ‹ฐ๋Š” ์—ฌ๋Ÿฌ ์˜คํ”ˆ ์†Œ์Šค LLM์˜ ์ถœ์‹œ๋ฅผ ๋ชฉ๊ฒฉํ–ˆ์œผ๋ฉฐ, ์ด ๋ชจ๋ธ์˜ ๊ฐ€์ค‘์น˜๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์ด์šฉ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. EleutherAI๋Š” GPT-NeoX-20B [35]์™€ GPT-J-GB [36]๋ฅผ ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. Google์€ UL2-20B [37]๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์นญํ™”๋Œ€ํ•™๊ต๋Š” GLM-130B [7]๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. Meta๋Š” OPT [9]์™€ LLaMA [8]๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ๋“ค์ด ๊ฐ€์น˜ ์žˆ๋Š” ๊ธฐ์—ฌ๋ฅผ ํ–ˆ๋‹ค๋Š” ์ ์€ ์ฃผ๋ชฉํ•  ๋งŒํ•˜์ง€๋งŒ, ์ผ๋ฐ˜์ ์œผ๋กœ ์ด๋“ค์€ ํ์‡„ํ˜• ๋ชจ๋ธ๋“ค๋งŒํผ์˜ ์„ฑ๋Šฅ์„ ๋ณด์ด์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค.
AI ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ๋Š” ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๊ฐ€ ๊ณต๊ฐœ์ ์œผ๋กœ ์ œ๊ณต๋˜๋Š” ์—ฌ๋Ÿฌ ์˜คํ”ˆ ์†Œ์Šค LLM์ด ์ถœ์‹œ๋˜๋Š” ๊ฒƒ์„ ๋ชฉ๊ฒฉํ–ˆ์Šต๋‹ˆ๋‹ค. EleutherAI๋Š” GPT-NeoX-20B[35]์™€ GPT-J-GB[36]๋ฅผ ๊ธฐ์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค. Google์€ UL2-20B[37]๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์นญํ™”๋Œ€ํ•™๊ต๋Š” GLM-130B[7]๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. Meta๋Š” OPT[9]์™€ LLaMA[8]๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์€ ๊ท€์ค‘ํ•œ ๊ธฐ์—ฌ๋ฅผ ํ–ˆ์ง€๋งŒ ์ผ๋ฐ˜์ ์œผ๋กœ ๋น„๊ณต๊ฐœ ๋ชจ๋ธ๊ณผ ๋™์ผํ•œ ์ˆ˜์ค€์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์ง€ ์•Š๋Š”๋‹ค๋Š” ์ ์— ์ฃผ๋ชฉํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
Continual learning (CL) approaches aim to learn from an evolving data distribution, adapting to novel data while retaining knowledge gathered through prior training (French, 1999; Rolnick et al., 2019; Caccia et al., 2020; Lesort et al., 2021). The key challenge of continual learning is to avoid forgetting past information, while also adapting to novel information. This trade-off is known as the rigidity-plasticity dilemma (Mermillod et al., 2013; Ostapenko et al., 2019; Riemer et al., 2019).
์ง€์† ํ•™์Šต(CL) ์ ‘๊ทผ๋ฒ•์€ ์ง„ํ™”ํ•˜๋Š” ๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜์—ฌ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์— ์ ์‘ํ•˜๋ฉด์„œ ์ด์ „ ํ›ˆ๋ จ์„ ํ†ตํ•ด ์ˆ˜์ง‘ํ•œ ์ง€์‹์„ ์œ ์ง€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค (French, 1999; Rolnick et al., 2019; Caccia et al., 2020; Lesort et al., 2021). ์ง€์† ํ•™์Šต์˜ ์ฃผ์š” ๊ณผ์ œ๋Š” ๊ณผ๊ฑฐ ์ •๋ณด๋ฅผ ์žŠ์ง€ ์•Š์œผ๋ฉด์„œ๋„ ์ƒˆ๋กœ์šด ์ •๋ณด์— ์ ์‘ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๊ท ํ˜•์„ ์œ ์ง€ํ•˜๋Š” ๋ฌธ์ œ๋Š” ๊ฒฝ์ง์„ฑ-๊ฐ€์†Œ์„ฑ ๋”œ๋ ˆ๋งˆ๋กœ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค (Mermillod et al., 2013; Ostapenko et al., 2019; Riemer et al., 2019).
์ง€์† ํ•™์Šต(CL) ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ง„ํ™”ํ•˜๋Š” ๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๊ณ , ์ด์ „ ํ•™์Šต์„ ํ†ตํ•ด ์ˆ˜์ง‘ํ•œ ์ง€์‹์„ ์œ ์ง€ํ•˜๋ฉด์„œ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์— ์ ์‘ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค(French, 1999; Rolnick et al., 2019; Caccia et al., 2020; Lesort et al., 2021). ์ง€์† ํ•™์Šต์˜ ํ•ต์‹ฌ ๊ณผ์ œ๋Š” ๊ณผ๊ฑฐ ์ •๋ณด๋ฅผ ์žŠ์–ด๋ฒ„๋ฆฌ๋Š” ๊ฒƒ์„ ๋ฐฉ์ง€ํ•˜๋Š” ๋™์‹œ์— ์ƒˆ๋กœ์šด ์ •๋ณด์— ์ ์‘ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํŠธ๋ ˆ์ด๋“œ์˜คํ”„๋ฅผ ๊ฒฝ์ง์„ฑ-๊ฐ€์†Œ์„ฑ ๋”œ๋ ˆ๋งˆ(Mermillod et al., 2013; Ostapenko et al., 2019; Riemer et al., 2019)๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค.
**Results.** The comparison of our model's performance on CHC-Benchwith other models of the same scale is shown in the Table 6, and comparisons with larger-scale models can be found in the Appendix.E.3. In CHC-Benchone can assess the expertise of models in specific domains. For instance, Deepseek-coder-1.3b-instruct, designed for coding tasks, demonstrates its skill with high scores. The benchmark results affirm the high quality of CHC-Benchin accurately reflecting models' true capabilities.
**๊ฒฐ๊ณผ.** CHC-Bench์—์„œ ์šฐ๋ฆฌ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋™์ผํ•œ ๊ทœ๋ชจ์˜ ๋‹ค๋ฅธ ๋ชจ๋ธ๋“ค๊ณผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ๋Š” Table 6์— ๋‚˜์™€ ์žˆ์œผ๋ฉฐ, ๋” ํฐ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ๋“ค๊ณผ์˜ ๋น„๊ต๋Š” Appendix.E.3์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. CHC-Bench์—์„œ๋Š” ํŠน์ • ๋„๋ฉ”์ธ์—์„œ ๋ชจ๋ธ์˜ ์ „๋ฌธ์„ฑ์„ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ฝ”๋”ฉ ์ž‘์—…์„ ์œ„ํ•ด ์„ค๊ณ„๋œ Deepseek-coder-1.3b-instruct๋Š” ๋†’์€ ์ ์ˆ˜๋กœ ๊ทธ ๋Šฅ๋ ฅ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ๋ฒค์น˜๋งˆํฌ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์˜ ์‹ค์ œ ๋Šฅ๋ ฅ์„ ์ •ํ™•ํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๋Š” CHC-Bench์˜ ๋†’์€ ํ’ˆ์งˆ์„ ํ™•์ธ์‹œ์ผœ์ค๋‹ˆ๋‹ค.
**๊ฒฐ๊ณผ**. CHC-Bench์—์„œ ๋™์ผํ•œ ๊ทœ๋ชจ์˜ ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•œ ๋ชจ๋ธ ์„ฑ๋Šฅ์€ ํ‘œ 6์— ๋‚˜์™€ ์žˆ์œผ๋ฉฐ, ๋” ํฐ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ๋Š” ๋ถ€๋ก E.3์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. CHC-Bench์—์„œ๋Š” ํŠน์ • ๋„๋ฉ”์ธ์—์„œ ๋ชจ๋ธ์˜ ์ „๋ฌธ์„ฑ์„ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ์ฝ”๋”ฉ ์ž‘์—…์„ ์œ„ํ•ด ์„ค๊ณ„๋œ Deepseek-coder-1.3b-instruct๋Š” ๋†’์€ ์ ์ˆ˜๋กœ ๊ทธ ๊ธฐ์ˆ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ฒค์น˜๋งˆํฌ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์˜ ์‹ค์ œ ์—ญ๋Ÿ‰์„ ์ •ํ™•ํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๋Š” CHC-Benchin์˜ ๋†’์€ ํ’ˆ์งˆ์„ ํ™•์ธ์‹œ์ผœ ์ค๋‹ˆ๋‹ค.
Both deduplication methods identify overlap between the train set and the validation set (Table 2). For example, 4.6% of the C4 validation set and 14.4% of the RealNews validation set examples had an approximate duplicate in their respective training sets. Such duplication is problematic since it could cause evaluation metrics to be unfairly inflated for models that are better at memorizing their train sets. We evaluate the effect of this leakage on publicly released models in Section 6.3.
๋‘ ์ค‘๋ณต ์ œ๊ฑฐ ๋ฐฉ๋ฒ• ๋ชจ๋‘ ํ›ˆ๋ จ ์„ธํŠธ์™€ ๊ฒ€์ฆ ์„ธํŠธ ๊ฐ„์˜ ์ค‘๋ณต์„ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค (ํ‘œ 2). ์˜ˆ๋ฅผ ๋“ค์–ด, C4 ๊ฒ€์ฆ ์„ธํŠธ์˜ 4.6%์™€ RealNews ๊ฒ€์ฆ ์„ธํŠธ์˜ 14.4% ์˜ˆ์ œ๊ฐ€ ๊ฐ๊ฐ์˜ ํ›ˆ๋ จ ์„ธํŠธ์— ๋Œ€๋žต์ ์ธ ์ค‘๋ณต์„ ๊ฐ€์ง€๊ณ  ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ค‘๋ณต์€ ํ›ˆ๋ จ ์„ธํŠธ๋ฅผ ๋” ์ž˜ ๊ธฐ์–ตํ•˜๋Š” ๋ชจ๋ธ์— ๋Œ€ํ•ด ํ‰๊ฐ€ ์ง€ํ‘œ๊ฐ€ ๋ถ€๋‹นํ•˜๊ฒŒ ์ƒ์Šนํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฌธ์ œ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” 6.3์ ˆ์—์„œ ๊ณต๊ฐœ๋œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ด ๋ˆ„์ถœ์˜ ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
๋‘ ์ค‘๋ณต ์ œ๊ฑฐ ๋ฐฉ๋ฒ• ๋ชจ๋‘ ํ›ˆ๋ จ ์„ธํŠธ์™€ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ ๊ฐ„์˜ ์ค‘๋ณต์„ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค(ํ‘œ 2). ์˜ˆ๋ฅผ ๋“ค์–ด, C4 ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์˜ 4.6%, RealNews ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์˜ 14.4% ์˜ˆ์ œ๊ฐ€ ๊ฐ ํ›ˆ๋ จ ์„ธํŠธ์—์„œ ๋Œ€๋žต์ ์ธ ์ค‘๋ณต์„ ๊ฐ€์กŒ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ค‘๋ณต์€ ๋ชจ๋ธ์ด ํ›ˆ๋ จ ์„ธํŠธ๋ฅผ ๋” ์ž˜ ์™ธ์šฐ๋Š” ๊ฒฝ์šฐ ํ‰๊ฐ€ ๋ฉ”ํŠธ๋ฆญ์ด ๋ถ€๋‹นํ•˜๊ฒŒ ๋ถ€ํ’€๋ ค์งˆ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋ฌธ์ œ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ์ถœ์ด ๊ณต๊ฐœ๋œ ๋ชจ๋ธ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์€ ์„น์…˜ 6.3์—์„œ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
As for the concrete data-to-model ratio \(\frac{D_{opt}}{N_{opt}}\), we notice that there is a huge gap in compute optimal regime of between ours and Hoffmann et al. (2022) despite that the trend of \(\frac{D_{opt}}{N_{opt}}\), with compute \(C\) is aligned between ours and theirs. Specifically, the data size should be 192 times larger than the model size on average, as opposed to 20 times in Hoffmann et al. (2022). We note that this aligns with the observation in Section 4.3 and Figure 6.
๊ตฌ์ฒด์ ์ธ ๋ฐ์ดํ„ฐ ๋Œ€ ๋ชจ๋ธ ๋น„์œจ \(\frac{D_{opt}}{N_{opt}}\)์— ๋Œ€ํ•ด, ๊ณ„์‚ฐ ์ตœ์ ํ™” ์ฒด์ œ์—์„œ ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ์™€ Hoffmann et al. (2022)์˜ ๊ฒฐ๊ณผ ์‚ฌ์ด์— ํฐ ์ฐจ์ด๊ฐ€ ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋น„๋ก ๊ณ„์‚ฐ \(C\)์™€ ํ•จ๊ป˜ \(\frac{D_{opt}}{N_{opt}}\)์˜ ๊ฒฝํ–ฅ์€ ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ์™€ ๊ทธ๋“ค์˜ ๊ฒฐ๊ณผ๊ฐ€ ์ผ์น˜ํ•˜์ง€๋งŒ, ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๋Š” ํ‰๊ท ์ ์œผ๋กœ ๋ชจ๋ธ ํฌ๊ธฐ๋ณด๋‹ค 192๋ฐฐ ๋” ์ปค์•ผ ํ•˜๋ฉฐ, ์ด๋Š” Hoffmann et al. (2022)์—์„œ ์ œ์‹œ๋œ 20๋ฐฐ์™€๋Š” ๋Œ€์กฐ์ ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์„น์…˜ 4.3๊ณผ ๊ทธ๋ฆผ 6์—์„œ์˜ ๊ด€์ฐฐ๊ณผ ์ผ์น˜ํ•จ์„ ์ฃผ๋ชฉํ•ฉ๋‹ˆ๋‹ค.
๊ตฌ์ฒด์ ์ธ ๋ฐ์ดํ„ฐ ๋Œ€ ๋ชจ๋ธ ๋น„์œจ \frac{D_{opt}}{N_{opt}}\)์˜ ๊ฒฝ์šฐ, ๊ณ„์‚ฐ ์ตœ์  ๋ ˆ์ง์—์„œ ์šฐ๋ฆฌ์™€ Hoffmann et al. (2022) ์‚ฌ์ด์— ํฐ ๊ฒฉ์ฐจ๊ฐ€ ์žˆ๋Š”๋ฐ, ๊ณ„์‚ฐ \(C\)๋ฅผ ์‚ฌ์šฉํ•œ \frac{D_{opt}}{N_{opt}}\)์˜ ์ถ”์„ธ๋Š” ์šฐ๋ฆฌ์™€ ๊ทธ๋“ค์˜ ์ถ”์„ธ๊ฐ€ ์ผ์น˜ํ•œ๋‹ค๋Š” ์ ์— ์ฃผ๋ชฉํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๋Š” ํ‰๊ท ์ ์œผ๋กœ ๋ชจ๋ธ ํฌ๊ธฐ๋ณด๋‹ค 192๋ฐฐ ๋” ์ปค์•ผ ํ•˜๋Š”๋ฐ, Hoffmann et al. (2022)์—์„œ๋Š” 20๋ฐฐ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์„น์…˜ 4.3 ๋ฐ ๊ทธ๋ฆผ 6์˜ ๊ด€์ฐฐ๊ณผ ์ผ์น˜ํ•œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ฉ๋‹ˆ๋‹ค.
We show that one particular source of bias, duplicated training examples, is pervasive: all four common NLP datasets we studied contained duplicates. Additionally, all four corresponding validation sets contained text duplicated in the training set. While naive deduplication is straightforward (and the datasets we consider already perform some naive form of deduplication), performing thorough deduplication at scale is both computationally challenging and requires sophisticated techniques.
์šฐ๋ฆฌ๋Š” ์ค‘๋ณต๋œ ํ›ˆ๋ จ ์˜ˆ์ œ๋ผ๋Š” ํŠน์ •ํ•œ ํŽธํ–ฅ์˜ ์›์ฒœ์ด ๋„๋ฆฌ ํผ์ ธ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค: ์šฐ๋ฆฌ๊ฐ€ ์—ฐ๊ตฌํ•œ ๋„ค ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ NLP ๋ฐ์ดํ„ฐ์…‹ ๋ชจ๋‘ ์ค‘๋ณต์„ ํฌํ•จํ•˜๊ณ  ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ถ”๊ฐ€๋กœ, ๋„ค ๊ฐ€์ง€ ํ•ด๋‹น ๊ฒ€์ฆ ์„ธํŠธ ๋ชจ๋‘ ํ›ˆ๋ จ ์„ธํŠธ์— ์ค‘๋ณต๋œ ํ…์ŠคํŠธ๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋‹จ์ˆœํ•œ ์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๊ฐ„๋‹จํ•˜์ง€๋งŒ (์šฐ๋ฆฌ๊ฐ€ ๊ณ ๋ คํ•œ ๋ฐ์ดํ„ฐ์…‹์€ ์ด๋ฏธ ์ผ๋ถ€ ๋‹จ์ˆœํ•œ ํ˜•ํƒœ์˜ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค), ๋Œ€๊ทœ๋ชจ๋กœ ์ฒ ์ €ํ•œ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์€ ๊ณ„์‚ฐ์ ์œผ๋กœ ๋„์ „์ ์ด๋ฉฐ ์ •๊ตํ•œ ๊ธฐ์ˆ ์„ ํ•„์š”๋กœ ํ•ฉ๋‹ˆ๋‹ค.
๋ชจ๋ฐฉ ํ•™์Šต ์˜ˆ์ œ์˜ ์ค‘๋ณต๊ณผ ๊ฐ™์€ ํŽธํ–ฅ์˜ ํ•œ ๊ฐ€์ง€ ํŠน์ • ์›์ธ์ด ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์กด์žฌํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค: ์šฐ๋ฆฌ๊ฐ€ ์—ฐ๊ตฌํ•œ ๋„ค ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ชจ๋‘ ์ค‘๋ณต์ด ํฌํ•จ๋˜์–ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋„ค ๊ฐ€์ง€ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ ๋ชจ๋‘ ํ•™์Šต ์„ธํŠธ์— ์ค‘๋ณต๋œ ํ…์ŠคํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ˆœ์ง„ํ•œ ์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๊ฐ„๋‹จํ•˜์ง€๋งŒ(์šฐ๋ฆฌ๊ฐ€ ๊ณ ๋ คํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ด๋ฏธ ์ผ๋ถ€ ์ˆœ์ง„ํ•œ ํ˜•ํƒœ์˜ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค), ๋Œ€๊ทœ๋ชจ๋กœ ์ฒ ์ €ํ•œ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์€ ๊ณ„์‚ฐ์ ์œผ๋กœ ์–ด๋ ต๊ณ  ์ •๊ตํ•œ ๊ธฐ์ˆ ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
For both instruction tuning and evaluation, we adopt the following prompt: _"The following is a conversation between a human and an AI assistant. The AI assistant gives helpful, detailed, and polite answers to the user's questions.\(\backslash n\) [\(|\)Humann\(|\):\(\{input\}\backslash n|AI|\)]."_. To reproduce our results, we release the code and data at the link: [https://github.com/RUCAIBox/LLMSurvey/tree/main/Experiments](https://github.com/RUCAIBox/LLMSurvey/tree/main/Experiments).
Instruction ํŠœ๋‹๊ณผ ํ‰๊ฐ€ ๋ชจ๋‘์— ๋Œ€ํ•ด, ์šฐ๋ฆฌ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค: _"๋‹ค์Œ์€ ์ธ๊ฐ„๊ณผ AI ์–ด์‹œ์Šคํ„ดํŠธ ๊ฐ„์˜ ๋Œ€ํ™”์ž…๋‹ˆ๋‹ค. AI ์–ด์‹œ์Šคํ„ดํŠธ๋Š” ์‚ฌ์šฉ์ž์˜ ์งˆ๋ฌธ์— ๋Œ€ํ•ด ๋„์›€์ด ๋˜๊ณ , ์ƒ์„ธํ•˜๋ฉฐ, ๊ณต์†ํ•œ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.\(\backslash n\) [\(|\)Humann\(|\):\(\{input\}\backslash n|AI|\)]."_. ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋ฅผ ์žฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์ฝ”๋“œ์™€ ๋ฐ์ดํ„ฐ๋ฅผ ๋‹ค์Œ ๋งํฌ์—์„œ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค: [https://github.com/RUCAIBox/LLMSurvey/tree/main/Experiments](https://github.com/RUCAIBox/LLMSurvey/tree/main/Experiments).
๋ช…๋ น์–ด ํŠœ๋‹๊ณผ ํ‰๊ฐ€ ๋ชจ๋‘์— ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค: _"๋‹ค์Œ์€ ์ธ๊ฐ„๊ณผ AI ์–ด์‹œ์Šคํ„ดํŠธ ๊ฐ„์˜ ๋Œ€ํ™”์ž…๋‹ˆ๋‹ค. AI ์–ด์‹œ์Šคํ„ดํŠธ๋Š” ์‚ฌ์šฉ์ž์˜ ์งˆ๋ฌธ์— ์œ ์šฉํ•˜๊ณ  ์ƒ์„ธํ•˜๋ฉฐ ์ •์ค‘ํ•œ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(\backslash n\) [\(|\)์ธ๊ฐ„n\(|\):\(\{์ž…๋ ฅ\}\backslash n|AI|\)]."_. ๊ฒฐ๊ณผ๋ฅผ ์žฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด ์ฝ”๋“œ์™€ ๋ฐ์ดํ„ฐ๋ฅผ ๋งํฌ์— ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค: [https://github.com/RUCAIBox/LLMSurvey/tree/main/Experiments](https://github.com/RUCAIBox/LLMSurvey/tree/main/Experiments).
We introduced D4, a method for data curation on LLMs that improves training efficiency by 20% across multiple model scales, with larger gains at increased model scale. We also demonstrated that, in contrast to common practice, repeating data via epoching can be beneficial for LLM training, but only if the data subset is intelligently selected. While we have shown encouraging efficiency gains and performance improvements via D4, our work has several limitations and many future directions.
์šฐ๋ฆฌ๋Š” ์—ฌ๋Ÿฌ ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ํ›ˆ๋ จ ํšจ์œจ์„ฑ์„ 20% ํ–ฅ์ƒ์‹œํ‚ค๋Š” LLM ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ๋ฐฉ๋ฒ•์ธ D4๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ์ปค์งˆ์ˆ˜๋ก ๋” ํฐ ์ด์ ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ผ๋ฐ˜์ ์ธ ๊ด€ํ–‰๊ณผ ๋‹ฌ๋ฆฌ, ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ์ด LLM ํ›ˆ๋ จ์— ์œ ์ตํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ, ์ด๋Š” ๋ฐ์ดํ„ฐ ํ•˜์œ„ ์ง‘ํ•ฉ์ด ์ง€๋Šฅ์ ์œผ๋กœ ์„ ํƒ๋œ ๊ฒฝ์šฐ์—๋งŒ ํ•ด๋‹น๋ฉ๋‹ˆ๋‹ค. D4๋ฅผ ํ†ตํ•ด ๊ณ ๋ฌด์ ์ธ ํšจ์œจ์„ฑ ํ–ฅ์ƒ๊ณผ ์„ฑ๋Šฅ ๊ฐœ์„ ์„ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ, ์šฐ๋ฆฌ์˜ ์ž‘์—…์—๋Š” ๋ช‡ ๊ฐ€์ง€ ํ•œ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ ๋งŽ์€ ๋ฏธ๋ž˜ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์ด ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค.
์—ฌ๋Ÿฌ ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ํ›ˆ๋ จ ํšจ์œจ์„ฑ์„ 20% ๊ฐœ์„ ํ•˜๋Š” LLM์— ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ๋ฐฉ๋ฒ•์ธ D4๋ฅผ ์†Œ๊ฐœํ–ˆ์œผ๋ฉฐ, ๋ชจ๋ธ ๊ทœ๋ชจ๊ฐ€ ์ปค์งˆ์ˆ˜๋ก ๋” ํฐ ์ด๋“์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ผ๋ฐ˜์ ์ธ ๊ด€ํ–‰๊ณผ๋Š” ๋‹ฌ๋ฆฌ ์—ํฌ์นญ์„ ํ†ตํ•ด ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ์ด LLM ํ›ˆ๋ จ์— ์œ ๋ฆฌํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋ฐ์ดํ„ฐ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์ง€๋Šฅ์ ์œผ๋กœ ์„ ํƒํ•ด์•ผ๋งŒ ๊ทธ ํšจ๊ณผ๋ฅผ ๋ณผ ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. D4๋ฅผ ํ†ตํ•ด ๊ณ ๋ฌด์ ์ธ ํšจ์œจ์„ฑ ํ–ฅ์ƒ๊ณผ ์„ฑ๋Šฅ ๊ฐœ์„ ์„ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ, ์ด ์ž‘์—…์—๋Š” ๋ช‡ ๊ฐ€์ง€ ํ•œ๊ณ„์™€ ํ–ฅํ›„ ๋งŽ์€ ๋ฐฉํ–ฅ์„ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
Subsequently, a substantial of works about X-of-thought (XoT) emerges like mushrooms after the rain in the NLP community, such as automatic XoT construction Kojima et al. (2022); Zhang et al. (2023); Xu et al. (2023), XoT structural variants Chen et al. (2022); Ning et al. (2023); Lei et al. (2023); Yao et al. (2023), etc. Note that to distinguish it from primitive CoT, we use XoT to refer to CoT in a broad sense, which is a collective term for the use of step-by-step reasoning methods.
์ดํ›„, NLP ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ๋Š” ๋น„๊ฐ€ ์˜จ ๋’ค ๋ฒ„์„ฏ์ด ๋‹์•„๋‚˜๋“ฏ X-of-thought (XoT)์— ๊ด€ํ•œ ๋งŽ์€ ์—ฐ๊ตฌ๋“ค์ด ๋“ฑ์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ž๋™ XoT ์ƒ์„ฑ (Kojima et al., 2022; Zhang et al., 2023; Xu et al., 2023), XoT ๊ตฌ์กฐ์  ๋ณ€ํ˜• (Chen et al., 2022; Ning et al., 2023; Lei et al., 2023; Yao et al., 2023) ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์›์‹œ์ ์ธ CoT์™€ ๊ตฌ๋ณ„ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก  ๋ฐฉ๋ฒ•์˜ ์‚ฌ์šฉ์„ ํฌ๊ด„์ ์œผ๋กœ ์ง€์นญํ•˜๋Š” ์šฉ์–ด๋กœ XoT๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
์ดํ›„, ์ž๋™ XoT ๊ตฌ์ถ• Kojima et al. (2022); Zhang et al. (2023); Xu et al. (2023), XoT ๊ตฌ์กฐ์  ๋ณ€ํ˜• Chen et al. (2022); Ning et al. (2023); Lei et al. (2023); Yao et al. (2023), ๋“ฑ NLP ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ X-of-thought(XoT)์— ๊ด€ํ•œ ๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ๋น„ ์˜จ ๋’ค ๋ฐค์— ์†ก์ด๋ฒ„์„ฏ์ฒ˜๋Ÿผ ๋“ฑ์žฅํ•ฉ๋‹ˆ๋‹ค. ์›์‹œ CoT์™€ ๊ตฌ๋ณ„ํ•˜๊ธฐ ์œ„ํ•ด ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก  ๋ฐฉ๋ฒ•์˜ ์‚ฌ์šฉ์„ ์ด์นญํ•˜๋Š” ๋„“์€ ์˜๋ฏธ์˜ CoT๋ฅผ XoT๋กœ ์‚ฌ์šฉํ•œ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”.
An example of how these ethical theories would apply to a specific ethical dilemma is the ethical use of artificial intelligence. Virtue ethics would suggest that the best way to ensure ethical AI use is to prioritize the development of moral virtues such as transparency, accountability, and fairness. Deontological ethics, on the other hand, would dictate that AI systems must be designed and used in a way that adheres to a set of pre-defined moral rules, regardless of the consequences.
์ด๋Ÿฌํ•œ ์œค๋ฆฌ ์ด๋ก ๋“ค์ด ํŠน์ • ์œค๋ฆฌ์  ๋”œ๋ ˆ๋งˆ์— ์–ด๋–ป๊ฒŒ ์ ์šฉ๋  ์ˆ˜ ์žˆ๋Š”์ง€์— ๋Œ€ํ•œ ์˜ˆ๋Š” ์ธ๊ณต์ง€๋Šฅ์˜ ์œค๋ฆฌ์  ์‚ฌ์šฉ์ž…๋‹ˆ๋‹ค. ๋• ์œค๋ฆฌํ•™์€ ํˆฌ๋ช…์„ฑ, ์ฑ…์ž„์„ฑ, ๊ณต์ •์„ฑ๊ณผ ๊ฐ™์€ ๋„๋•์  ๋•๋ชฉ์˜ ๊ฐœ๋ฐœ์„ ์šฐ์„ ์‹œํ•˜๋Š” ๊ฒƒ์ด ์œค๋ฆฌ์  AI ์‚ฌ์šฉ์„ ๋ณด์žฅํ•˜๋Š” ์ตœ์„ ์˜ ๋ฐฉ๋ฒ•์ด๋ผ๊ณ  ์ œ์•ˆํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์˜๋ฌด๋ก ์  ์œค๋ฆฌํ•™์€ ๊ฒฐ๊ณผ์— ์ƒ๊ด€์—†์ด AI ์‹œ์Šคํ…œ์ด ์‚ฌ์ „์— ์ •์˜๋œ ๋„๋•์  ๊ทœ์น™์„ ์ค€์ˆ˜ํ•˜๋„๋ก ์„ค๊ณ„๋˜๊ณ  ์‚ฌ์šฉ๋˜์–ด์•ผ ํ•œ๋‹ค๊ณ  ๊ทœ์ •ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ์œค๋ฆฌ ์ด๋ก ์ด ํŠน์ • ์œค๋ฆฌ์  ๋”œ๋ ˆ๋งˆ์— ์–ด๋–ป๊ฒŒ ์ ์šฉ๋˜๋Š”์ง€ ๋ณด์—ฌ์ฃผ๋Š” ์˜ˆ๋Š” ์ธ๊ณต์ง€๋Šฅ์˜ ์œค๋ฆฌ์  ์‚ฌ์šฉ์ž…๋‹ˆ๋‹ค. ๋• ์œค๋ฆฌ๋Š” ์œค๋ฆฌ์  AI ์‚ฌ์šฉ์„ ๋ณด์žฅํ•˜๋Š” ๊ฐ€์žฅ ์ข‹์€ ๋ฐฉ๋ฒ•์€ ํˆฌ๋ช…์„ฑ, ์ฑ…์ž„์„ฑ, ๊ณต์ •์„ฑ๊ณผ ๊ฐ™์€ ๋„๋•์  ๋•๋ชฉ์˜ ๊ฐœ๋ฐœ์„ ์šฐ์„ ์‹œํ•˜๋Š” ๊ฒƒ์ด๋ผ๊ณ  ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์˜๋ฌด๋ก ์  ์œค๋ฆฌ๋Š” ๊ฒฐ๊ณผ์— ๊ด€๊ณ„์—†์ด ๋ฏธ๋ฆฌ ์ •์˜๋œ ์ผ๋ จ์˜ ๋„๋•์  ๊ทœ์น™์„ ์ค€์ˆ˜ํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ AI ์‹œ์Šคํ…œ์„ ์„ค๊ณ„ํ•˜๊ณ  ์‚ฌ์šฉํ•ด์•ผ ํ•œ๋‹ค๊ณ  ๊ทœ์ •ํ•ฉ๋‹ˆ๋‹ค.
As mentioned in Section 5, we compared various values of \(K\in\{0,\ldots,5\}\), shown in Table 3.We were unable to find an optimal value of \(K\) per model, per \(0/5\)-shot, or per task. In fact, other than Anatomy that worked well with \(K=2\) consistently, there seems to be no patterns that aid in predicting the performance per \(K\), unlike the results presented in (Lewis et al., 2020) for other setups. Moreover, the gap between the best and worst performing \(K\)s can be large.
5์žฅ์—์„œ ์–ธ๊ธ‰ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ์šฐ๋ฆฌ๋Š” Table 3์— ํ‘œ์‹œ๋œ \(K\in\{0,\ldots,5\}\)์˜ ๋‹ค์–‘ํ•œ ๊ฐ’์„ ๋น„๊ตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ๋ณ„, \(0/5\)-์ƒท๋ณ„, ๋˜๋Š” ์ž‘์—…๋ณ„๋กœ ์ตœ์ ์˜ \(K\) ๊ฐ’์„ ์ฐพ์„ ์ˆ˜ ์—†์—ˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ์‹ค, \(K=2\)์—์„œ ์ผ๊ด€๋˜๊ฒŒ ์ž˜ ์ž‘๋™ํ•œ Anatomy๋ฅผ ์ œ์™ธํ•˜๊ณ ๋Š”, (Lewis et al., 2020)์—์„œ ๋‹ค๋ฅธ ์„ค์ •์— ๋Œ€ํ•ด ์ œ์‹œ๋œ ๊ฒฐ๊ณผ์™€ ๋‹ฌ๋ฆฌ, \(K\)๋ณ„ ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ํŒจํ„ด์ด ์—†๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๋ณด์ž…๋‹ˆ๋‹ค. ๊ฒŒ๋‹ค๊ฐ€, ์„ฑ๋Šฅ์ด ๊ฐ€์žฅ ์ข‹์€ \(K\)์™€ ๊ฐ€์žฅ ๋‚˜์œ \(K\) ์‚ฌ์ด์˜ ์ฐจ์ด๊ฐ€ ํด ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์„น์…˜ 5์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ํ‘œ 3์— ํ‘œ์‹œ๋œ \(K\in\{0,\ldots,5\}\)์˜ ๋‹ค์–‘ํ•œ ๊ฐ’์„ ๋น„๊ตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ๋ณ„, 0/5์ƒท๋ณ„ ๋˜๋Š” ์ž‘์—…๋ณ„๋กœ ์ตœ์ ์˜ \(K\) ๊ฐ’์„ ์ฐพ์„ ์ˆ˜ ์—†์—ˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ํ•ด๋ถ€ํ•™์€ \(K=2\)์—์„œ ์ผ๊ด€๋˜๊ฒŒ ์ž˜ ์ž‘๋™ํ–ˆ๋˜ ๊ฒƒ ์™ธ์—๋Š” ๋‹ค๋ฅธ ์„ค์ •์— ๋Œ€ํ•ด (Lewis et al., 2020)์—์„œ ์ œ์‹œ๋œ ๊ฒฐ๊ณผ์™€ ๋‹ฌ๋ฆฌ \(K\)๋‹น ์„ฑ๋Šฅ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ํŒจํ„ด์ด ์—†๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์€ \(K\)์™€ ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ๋‚ฎ์€ \(K\) ์‚ฌ์ด์˜ ๊ฒฉ์ฐจ๊ฐ€ ํด ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
We then train new models (with the same architecture) to predict GCD, from AI data (generated by the above model), and compare to training with correct data - from correct computation of the GCD. When trained on small number of examples (less than \(100\) million), models learning from AI data achieve better accuracy (Table 2). We believe this is due to the fact that AI data smoothes away all the hard case, therefore presenting the model with a cleaner signal in the initial stages.
๊ทธ ํ›„, ์šฐ๋ฆฌ๋Š” ์ƒˆ๋กœ์šด ๋ชจ๋ธ๋“ค์„ (๋™์ผํ•œ ์•„ํ‚คํ…์ฒ˜๋กœ) AI ๋ฐ์ดํ„ฐ(์œ„์˜ ๋ชจ๋ธ์— ์˜ํ•ด ์ƒ์„ฑ๋œ)๋กœ๋ถ€ํ„ฐ GCD๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ํ›ˆ๋ จํ•˜๊ณ , ์˜ฌ๋ฐ”๋ฅธ ๋ฐ์ดํ„ฐ - GCD์˜ ์ •ํ™•ํ•œ ๊ณ„์‚ฐ์œผ๋กœ๋ถ€ํ„ฐ - ๋กœ ํ›ˆ๋ จํ•œ ๊ฒƒ๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ ์€ ์ˆ˜์˜ ์˜ˆ์ œ(1์–ต ๋ฏธ๋งŒ)๋กœ ํ›ˆ๋ จํ–ˆ์„ ๋•Œ, AI ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•œ ๋ชจ๋ธ๋“ค์ด ๋” ๋‚˜์€ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค (ํ‘œ 2). ์šฐ๋ฆฌ๋Š” ์ด๊ฒƒ์ด AI ๋ฐ์ดํ„ฐ๊ฐ€ ๋ชจ๋“  ์–ด๋ ค์šด ์‚ฌ๋ก€๋ฅผ ๋ถ€๋“œ๋Ÿฝ๊ฒŒ ์ฒ˜๋ฆฌํ•˜์—ฌ ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ๋ชจ๋ธ์— ๋” ๊นจ๋—ํ•œ ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค.
๊ทธ๋Ÿฐ ๋‹ค์Œ ์ƒˆ๋กœ์šด ๋ชจ๋ธ(๋™์ผํ•œ ์•„ํ‚คํ…์ฒ˜)์„ ํ›ˆ๋ จํ•˜์—ฌ AI ๋ฐ์ดํ„ฐ(์œ„์˜ ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ)์—์„œ GCD๋ฅผ ์˜ˆ์ธกํ•˜๊ณ  ์˜ฌ๋ฐ”๋ฅธ GCD ๊ณ„์‚ฐ์—์„œ ์ƒ์„ฑ๋œ ์˜ฌ๋ฐ”๋ฅธ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์†Œ์ˆ˜์˜ ์˜ˆ์ œ(1์–ต ๊ฐœ ๋ฏธ๋งŒ)๋กœ ํ›ˆ๋ จํ•  ๊ฒฝ์šฐ, AI ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•œ ๋ชจ๋ธ์ด ๋” ๋†’์€ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค(ํ‘œ 2). ์ด๋Š” AI ๋ฐ์ดํ„ฐ๊ฐ€ ๋ชจ๋“  ์–ด๋ ค์šด ๊ฒฝ์šฐ๋ฅผ ํ‰ํ™œํ™”ํ•˜์—ฌ ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ๋ชจ๋ธ์— ๋” ๊นจ๋—ํ•œ ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
role-user: Given the following conversation between PersonU and PersonA: PersonU: I need to know how to prepare for college. PersonA: You should first define your goals and interest and be sure to know the costs and how to plan financially and academically for college. PersonU: Is there something I can use as a guide to get ready for it? Instead of having this entire conversation, how can PersonU get what he or she is looking for using a single question? Respond with that question.
๋‹ค์Œ์€ PersonU์™€ PersonA ๊ฐ„์˜ ๋Œ€ํ™”์ž…๋‹ˆ๋‹ค: PersonU: ๋Œ€ํ•™ ์ค€๋น„๋ฅผ ์–ด๋–ป๊ฒŒ ํ•ด์•ผ ํ• ์ง€ ์•Œ๊ณ  ์‹ถ์–ด์š”. PersonA: ๋จผ์ € ๋ชฉํ‘œ์™€ ๊ด€์‹ฌ์‚ฌ๋ฅผ ์ •์˜ํ•˜๊ณ , ๋น„์šฉ๊ณผ ์žฌ์ • ๋ฐ ํ•™์—… ๊ณ„ํš์„ ์„ธ์šฐ๋Š” ๋ฐฉ๋ฒ•์„ ์•Œ์•„์•ผ ํ•ด์š”. PersonU: ์ค€๋น„๋ฅผ ์œ„ํ•œ ๊ฐ€์ด๋“œ๊ฐ€ ์žˆ์„๊นŒ์š”? ์ด ์ „์ฒด ๋Œ€ํ™”๋ฅผ ๋Œ€์‹ ํ•˜์—ฌ, PersonU๊ฐ€ ๋‹จ์ผ ์งˆ๋ฌธ์œผ๋กœ ์›ํ•˜๋Š” ์ •๋ณด๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์€ ๋ฌด์—‡์ธ๊ฐ€์š”? ๊ทธ ์งˆ๋ฌธ์œผ๋กœ ์‘๋‹ตํ•˜์„ธ์š”.
ole-user: PersonU์™€ PersonA ์‚ฌ์ด์˜ ๋‹ค์Œ ๋Œ€ํ™”๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ: PersonU: ๋Œ€ํ•™์— ๋Œ€๋น„ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์•Œ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค. PersonA: ๋จผ์ € ๋ชฉํ‘œ์™€ ๊ด€์‹ฌ์‚ฌ๋ฅผ ์ •์˜ํ•˜๊ณ  ๋Œ€ํ•™์— ๋Œ€ํ•œ ์žฌ์ •์ , ํ•™์—…์  ๊ณ„ํš ๋ฐฉ๋ฒ•๊ณผ ๋น„์šฉ์„ ํ™•์‹คํžˆ ์•Œ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค. PersonU: ์ค€๋น„๋ฅผ ์œ„ํ•œ ๊ฐ€์ด๋“œ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์ด ์žˆ๋‚˜์š”? ์ด ๋Œ€ํ™”๋ฅผ ํ•˜๋Š” ๋Œ€์‹  PersonU๊ฐ€ ์›ํ•˜๋Š” ๊ฒƒ์„ ์–ป๊ธฐ ์œ„ํ•ด ํ•œ ๊ฐ€์ง€ ์งˆ๋ฌธ์œผ๋กœ ์–ด๋–ป๊ฒŒ ํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ๊ทธ ์งˆ๋ฌธ์œผ๋กœ ๋‹ตํ•˜์„ธ์š”.
Figure 5: Prediction errors of the five-domain data mixing laws fitted with explicit and implicit domain aggregation. _Explicit domain aggregation_: we fit Eqn. 7 for five domains respectively and sum them up according to their weight in the overall validation sets. _Implicit domain aggregation_: we fit the losses on overall validation with Eqn. 8, assuming different numbers of implicit domains \(K\) and treating the proportion of different implicit domains as learnable parameters.
Figure 5: ๋ช…์‹œ์  ๋ฐ ์•”์‹œ์  ๋„๋ฉ”์ธ ์ง‘๊ณ„๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋งž์ถ˜ ๋‹ค์„ฏ ๊ฐœ ๋„๋ฉ”์ธ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์˜ ์˜ˆ์ธก ์˜ค๋ฅ˜. _๋ช…์‹œ์  ๋„๋ฉ”์ธ ์ง‘๊ณ„_: ๋‹ค์„ฏ ๊ฐœ ๋„๋ฉ”์ธ ๊ฐ๊ฐ์— ๋Œ€ํ•ด Eqn. 7์„ ๋งž์ถ”๊ณ , ์ „์ฒด ๊ฒ€์ฆ ์„ธํŠธ์—์„œ์˜ ๊ฐ€์ค‘์น˜์— ๋”ฐ๋ผ ์ด๋ฅผ ํ•ฉ์‚ฐํ•ฉ๋‹ˆ๋‹ค. _์•”์‹œ์  ๋„๋ฉ”์ธ ์ง‘๊ณ„_: Eqn. 8์„ ์‚ฌ์šฉํ•˜์—ฌ ์ „์ฒด ๊ฒ€์ฆ์—์„œ์˜ ์†์‹ค์„ ๋งž์ถ”๋ฉฐ, ์„œ๋กœ ๋‹ค๋ฅธ ์•”์‹œ์  ๋„๋ฉ”์ธ \(K\)์˜ ์ˆ˜๋ฅผ ๊ฐ€์ •ํ•˜๊ณ , ์„œ๋กœ ๋‹ค๋ฅธ ์•”์‹œ์  ๋„๋ฉ”์ธ์˜ ๋น„์œจ์„ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜๋กœ ์ทจ๊ธ‰ํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 5: ๋ช…์‹œ์  ๋ฐ ์•”์‹œ์  ๋„๋ฉ”์ธ ์ง‘๊ณ„๋กœ ๋งž์ถคํ™”๋œ 5๊ฐœ ๋„๋ฉ”์ธ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์˜ ์˜ˆ์ธก ์˜ค์ฐจ. _๋ช…์‹œ์  ๋„๋ฉ”์ธ ์ง‘๊ณ„_: 5๊ฐœ์˜ ๋„๋ฉ”์ธ์— ๋Œ€ํ•ด ๊ฐ๊ฐ Eqn. 7์„ ๋งž์ถ”๊ณ  ์ „์ฒด ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ง‘ํ•ฉ์—์„œ ๊ฐ€์ค‘์น˜์— ๋”ฐ๋ผ ํ•ฉ์‚ฐํ•ฉ๋‹ˆ๋‹ค. _์•”์‹œ์  ๋„๋ฉ”์ธ ์ง‘๊ณ„_: ์•”์‹œ์  ๋„๋ฉ”์ธ์˜ ์ˆ˜ \(K\)๋ฅผ ๋‹ค๋ฅด๊ฒŒ ๊ฐ€์ •ํ•˜๊ณ  ๋‹ค๋ฅธ ์•”์‹œ์  ๋„๋ฉ”์ธ์˜ ๋น„์œจ์„ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ์ทจ๊ธ‰ํ•˜์—ฌ Eqn. 8์„ ์‚ฌ์šฉํ•˜์—ฌ ์ „์ฒด ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ์— ๋Œ€ํ•œ ์†์‹ค์„ ๋งž์ถฅ๋‹ˆ๋‹ค.
The Yi data mixture is shown in Fig. 2. To produce a high-quality bilingual pretraining data, we meticulously designed a cascaded data-processing pipeline, as illustrated in Fig 1. This pipeline features a series of data-cleaning strategies targeting quality and diversity. We start with web documents from Common Crawl, use the CCNet pipeline [79] for language identification and perplexity scoring. Then we use a combination of filtering and deduplication process, as detailed below.
Yi ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์€ Fig. 2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ์˜ ์ด์ค‘ ์–ธ์–ด ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” Fig. 1์— ์„ค๋ช…๋œ ๋Œ€๋กœ ์ •๊ตํ•˜๊ฒŒ ์„ค๊ณ„๋œ ๊ณ„๋‹จ์‹ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ ํŒŒ์ดํ”„๋ผ์ธ์„ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ํŒŒ์ดํ”„๋ผ์ธ์€ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ์ผ๋ จ์˜ ๋ฐ์ดํ„ฐ ์ •๋ฆฌ ์ „๋žต์„ ํŠน์ง•์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” Common Crawl์˜ ์›น ๋ฌธ์„œ๋กœ ์‹œ์ž‘ํ•˜์—ฌ, ์–ธ์–ด ์‹๋ณ„ ๋ฐ ๋‹นํ˜น๋„ ์ ์ˆ˜๋ฅผ ์œ„ํ•ด CCNet ํŒŒ์ดํ”„๋ผ์ธ [79]์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์•„๋ž˜์— ์ž์„ธํžˆ ์„ค๋ช…๋œ ๋Œ€๋กœ ํ•„ํ„ฐ๋ง ๋ฐ ์ค‘๋ณต ์ œ๊ฑฐ ๊ณผ์ •์„ ๊ฒฐํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
Yi ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์€ Fig. 2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ์˜ ์–‘์–ด ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด Fig 1๊ณผ ๊ฐ™์ด ์ผ๋ จ์˜ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ ํŒŒ์ดํ”„๋ผ์ธ์„ ์„ธ์‹ฌํ•˜๊ฒŒ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ํŒŒ์ดํ”„๋ผ์ธ์—๋Š” ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ์ผ๋ จ์˜ ๋ฐ์ดํ„ฐ ์ •๋ฆฌ ์ „๋žต์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ € Common Crawl์˜ ์›น ๋ฌธ์„œ๋กœ ์‹œ์ž‘ํ•˜์—ฌ ์–ธ์–ด ์‹๋ณ„ ๋ฐ ๋‹นํ˜น๋„ ์ ์ˆ˜๋ฅผ ์œ„ํ•ด CCNet ํŒŒ์ดํ”„๋ผ์ธ [79]์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์•„๋ž˜์—์„œ ์ž์„ธํžˆ ์„ค๋ช…ํ•˜๋Š” ํ•„ํ„ฐ๋ง ๋ฐ ์ค‘๋ณต ์ œ๊ฑฐ ํ”„๋กœ์„ธ์Šค๋ฅผ ์กฐํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
CoQAWe use the validation set of CoQA since its test set cannot be directly obtained. Its validation set consists of 500 dialogues with 7983 user-agent turns. Each question has a total of four human-written answers. It has unanswerable cases. However, we do not include this dataset in unanswerable case evaluation for the same reason as in the TopiOCQA dataset. We only find 13 questions with no answer being found, which is not sufficient enough for the unanswerable case evaluation.
CoQA์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋Š” ์ง์ ‘์ ์œผ๋กœ ์–ป์„ ์ˆ˜ ์—†๊ธฐ ๋•Œ๋ฌธ์— CoQA์˜ ๊ฒ€์ฆ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ฆ ์„ธํŠธ๋Š” 7983๊ฐœ์˜ ์‚ฌ์šฉ์ž-์—์ด์ „ํŠธ ํ„ด์„ ํฌํ•จํ•œ 500๊ฐœ์˜ ๋Œ€ํ™”๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์งˆ๋ฌธ์—๋Š” ์ด ๋„ค ๊ฐœ์˜ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ๋‹ต๋ณ€์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ๋„ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ TopiOCQA ๋ฐ์ดํ„ฐ์…‹๊ณผ ๊ฐ™์€ ์ด์œ ๋กœ ์ด ๋ฐ์ดํ„ฐ์…‹์„ ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ ํ‰๊ฐ€์— ํฌํ•จํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋‹ต๋ณ€์ด ์—†๋Š” ์งˆ๋ฌธ์€ 13๊ฐœ์— ๋ถˆ๊ณผํ•˜์—ฌ, ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ ํ‰๊ฐ€์— ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
CoQATo obtain the test set directly, we use the validation set of CoQA. Its validation set consists of 500 dialogues with 7983 user-agent turns. Each question has a total of four human-written answers. It has unanswerable cases. However, we do not include this dataset in unanswerable case evaluation for the same reason as in the TopiOCQA dataset. We only find 13 questions with no answer being found, which is not sufficient enough for the unanswerable case evaluation.
**Reference [1]** G. Venugopal & Venugopal (born 10 December 1960) is an Indian playback singer known for his work in Malayalam films. He started his singing career in the film "Odurantmannava Aalaviyam" (1984). Since then he has sung in more than 300 films; and has over 500 private albums to his resident. The many awards he has won include the coveted Best Playback Singer (Kerala State) which he won thrice, the Film Critics award and the Kerala Kaumundi Gallop Poll award (twice).
**Reference [1]** G. Venugopal & Venugopal (1960๋…„ 12์›” 10์ผ ์ถœ์ƒ)๋Š” ๋ง๋ผ์–„๋žŒ ์˜ํ™”์—์„œ์˜ ์ž‘์—…์œผ๋กœ ์•Œ๋ ค์ง„ ์ธ๋„ ํ”Œ๋ ˆ์ด๋ฐฑ ๊ฐ€์ˆ˜์ž…๋‹ˆ๋‹ค. ๊ทธ๋Š” 1984๋…„ ์˜ํ™” "Odurantmannava Aalaviyam"์—์„œ ๊ฐ€์ˆ˜ ๊ฒฝ๋ ฅ์„ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์ดํ›„๋กœ ๊ทธ๋Š” 300ํŽธ ์ด์ƒ์˜ ์˜ํ™”์—์„œ ๋…ธ๋ž˜๋ฅผ ๋ถˆ๋ €์œผ๋ฉฐ, 500๊ฐœ ์ด์ƒ์˜ ๊ฐœ์ธ ์•จ๋ฒ”์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๊ฐ€ ์ˆ˜์ƒํ•œ ๋งŽ์€ ์ƒ ์ค‘์—๋Š” ์„ธ ๋ฒˆ ์ˆ˜์ƒํ•œ ์ผ€๋ž„๋ผ ์ฃผ ์ตœ๊ณ ์˜ ํ”Œ๋ ˆ์ด๋ฐฑ ๊ฐ€์ˆ˜์ƒ, ์˜ํ™” ๋น„ํ‰๊ฐ€์ƒ, ๊ทธ๋ฆฌ๊ณ  ๋‘ ๋ฒˆ ์ˆ˜์ƒํ•œ ์ผ€๋ž„๋ผ ์นด์šฐ๋ฌธ๋”” ๊ฐค๋Ÿฝ ํด ์ƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
**์ฐธ๊ณ  [1]** G. ๋ฒ ๋ˆ„๊ณ ํŒ” & ๋ฒ ๋ˆ„๊ณ ํŒ” (1960๋…„ 12์›” 10์ผ์ƒ)์€ ๋ง๋ผ์–„๋žŒ ์˜ํ™”์—์„œ ํ™œ๋™ํ•œ ๊ฒƒ์œผ๋กœ ์œ ๋ช…ํ•œ ์ธ๋„์˜ ์žฌ์ƒ ๊ฐ€์ˆ˜์ž…๋‹ˆ๋‹ค. ๊ทธ๋Š” ์˜ํ™” "์˜ค๋‘๋ž€ํŠธ๋งŒ๋‚˜๋ฐ” ์•Œ๋ผ๋น„์–Œ"(1984)์—์„œ ๋…ธ๋ž˜ ๊ฒฝ๋ ฅ์„ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์ดํ›„๋กœ ๊ทธ๋Š” 300ํŽธ ์ด์ƒ์˜ ์˜ํ™”์—์„œ ๋…ธ๋ž˜๋ฅผ ๋ถˆ๋ €์œผ๋ฉฐ, 500์žฅ์ด ๋„˜๋Š” ๊ฐœ์ธ ์•จ๋ฒ”์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๊ฐ€ ์ˆ˜์ƒํ•œ ์ˆ˜๋งŽ์€ ์ƒ์œผ๋กœ๋Š” ์„ธ ๋ฒˆ์ด๋‚˜ ์ˆ˜์ƒํ•œ ๋ฒ ์ŠคํŠธ ์žฌ์ƒ ๊ฐ€์ˆ˜(์ผ€๋ž„๋ผ ์ฃผ), ์˜ํ™” ํ‰๋ก ๊ฐ€์ƒ, ์ผ€๋ž„๋ผ ์นด์šฐ๋ฌธ๋”” ๊ฐค๋Ÿฝ ํด์ƒ(2ํšŒ) ๋“ฑ์ด ์žˆ์Šต๋‹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 \((x_{1:T}^{1},\mathbf{v}^{1})\ldots(x_{1:T}^{J},\mathbf{v}^{J})\).
์ด์ œ ์šฐ๋ฆฌ๋Š” ์ด ๊ณผ์ •์„ ์—ญ์œผ๋กœ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค: ํ™•๋ฅ  ๋ฒกํ„ฐ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ์ด๋Ÿฌํ•œ ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์„ ์œ ๋„ํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋ ค๊ณ  ์‹œ๋„ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์„ ์ œ๊ณตํ•˜๋Š” ์–ด๋–ค ๋ณด์ง€ ๋ชปํ•œ ๋ชจ๋ธ \(f:\mathcal{V}^{T}\rightarrow\Delta^{|\mathcal{V}|-1}\)์ด ์ฃผ์–ด์กŒ์„ ๋•Œ, ์šฐ๋ฆฌ๋Š” ์ƒ˜ํ”Œ๋กœ๋ถ€ํ„ฐ ์ด ํ•จ์ˆ˜๋ฅผ ์—ญ์œผ๋กœ ํ•™์Šตํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค: ํ…์ŠคํŠธ ์ ‘๋‘์‚ฌ์™€ ๊ทธ์— ์—ฐ๊ด€๋œ ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ  ๋ฒกํ„ฐ์˜ ์Œ \((x_{1:T}^{1},\mathbf{v}^{1})\ldots(x_{1:T}^{J},\mathbf{v}^{J})\).
์ด์ œ ํ”„๋กœ์„ธ์Šค๋ฅผ ๋ฐ˜์ „์‹œํ‚ค๋Š” ๋ฌธ์ œ๋ฅผ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ํ™•๋ฅ  ๋ฒกํ„ฐ๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ์ด๋Ÿฌํ•œ ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์„ ์œ ๋„ํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋ ค๊ณ  ์‹œ๋„ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ ์„ ์ œ๊ณตํ•˜๋Š” ๋ณด์ด์ง€ ์•Š๋Š” ๋ชจ๋ธ \(f:\mathcal{V}^{T}\rightarrow\Delta^{|\mathcal{V}|-1}\)๊ฐ€ ์ฃผ์–ด์ง€๋ฉด, ์ด ํ•จ์ˆ˜๋ฅผ ์ƒ˜ํ”Œ์—์„œ ๋ฐ˜์ „ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ํ…์ŠคํŠธ ์ ‘๋‘์‚ฌ ์Œ๊ณผ ๊ทธ์™€ ๊ด€๋ จ๋œ ๋‹ค์Œ ํ† ํฐ ํ™•๋ฅ  ๋ฒกํ„ฐ \((x_{1:T}^{1},\mathbf{v}^{1})\ldots(x_{1:T}^{J},\mathbf{v}^{J})\).
Moreover, Appendix C showcases intriguing examples that demonstrate the utility of our models merged using evolution. The merged models correctly answered questions that require both knowledge about Japanese-culture _and_ Math ability. In contrast, even if such Japanese questions were translated into English and answered in English, English math models would likely fail to provide the correct answers as they may not be aware of Japanese culture-specific context in the questions.
๊ฒŒ๋‹ค๊ฐ€, ๋ถ€๋ก C๋Š” ์ง„ํ™”๋ฅผ ํ†ตํ•ด ๋ณ‘ํ•ฉ๋œ ์šฐ๋ฆฌ์˜ ๋ชจ๋ธ๋“ค์ด ์œ ์šฉํ•จ์„ ๋ณด์—ฌ์ฃผ๋Š” ํฅ๋ฏธ๋กœ์šด ์˜ˆ์‹œ๋“ค์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ณ‘ํ•ฉ๋œ ๋ชจ๋ธ๋“ค์€ ์ผ๋ณธ ๋ฌธํ™”์— ๋Œ€ํ•œ ์ง€์‹ _๋ฐ_ ์ˆ˜ํ•™ ๋Šฅ๋ ฅ์„ ๋ชจ๋‘ ํ•„์š”๋กœ ํ•˜๋Š” ์งˆ๋ฌธ์— ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๋‹ต๋ณ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์ด๋Ÿฌํ•œ ์ผ๋ณธ์–ด ์งˆ๋ฌธ์ด ์˜์–ด๋กœ ๋ฒˆ์—ญ๋˜์–ด ์˜์–ด๋กœ ๋‹ต๋ณ€๋˜๋”๋ผ๋„, ์˜์–ด ์ˆ˜ํ•™ ๋ชจ๋ธ๋“ค์€ ์งˆ๋ฌธ์— ํฌํ•จ๋œ ์ผ๋ณธ ๋ฌธํ™” ํŠน์œ ์˜ ๋งฅ๋ฝ์„ ์ธ์‹ํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์˜ฌ๋ฐ”๋ฅธ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜์ง€ ๋ชปํ•  ๊ฐ€๋Šฅ์„ฑ์ด ํฝ๋‹ˆ๋‹ค.
๋˜ํ•œ, ๋ถ€๋ก C์—๋Š” ์ง„ํ™”๋ฅผ ํ†ตํ•ด ๋ณ‘ํ•ฉ๋œ ๋ชจ๋ธ์˜ ์œ ์šฉ์„ฑ์„ ๋ณด์—ฌ์ฃผ๋Š” ํฅ๋ฏธ๋กœ์šด ์˜ˆ๊ฐ€ ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณ‘ํ•ฉ๋œ ๋ชจ๋ธ์€ ์ผ๋ณธ ๋ฌธํ™”์— ๋Œ€ํ•œ ์ง€์‹๊ณผ ์ˆ˜ํ•™ ๋Šฅ๋ ฅ์ด ๋ชจ๋‘ ํ•„์š”ํ•œ ์งˆ๋ฌธ์— ์ •๋‹ต์„ ๋งžํ˜”์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์ด๋Ÿฌํ•œ ์ผ๋ณธ์–ด ์งˆ๋ฌธ์„ ์˜์–ด๋กœ ๋ฒˆ์—ญํ•˜์—ฌ ์˜์–ด๋กœ ๋‹ตํ•˜๋”๋ผ๋„ ์˜์–ด ์ˆ˜ํ•™ ๋ชจ๋ธ์€ ์งˆ๋ฌธ์˜ ์ผ๋ณธ ๋ฌธํ™”์— ํŠนํ™”๋œ ๋งฅ๋ฝ์„ ์ธ์ง€ํ•˜์ง€ ๋ชปํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ •๋‹ต์„ ์ œ๊ณตํ•˜์ง€ ๋ชปํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค.
Value-only.Let \(g_{1},\ldots,g_{N}\in[T]\), where each \(g_{i}\) is i.i.d. uniformly chosen at random from \([T]\). Think of these as _values_. Suppose a model, parameterized by \(W\), is trained on the training data \(\mathcal{Z}=\big{(}g_{1},...,g_{N}\big{)}\). Assume this model, for a given index \(i\in[N]\), can generate a random prediction \(f_{i}\) corresponding to \(g_{i}\). We can represent this model as \(f_{i}(W(\mathcal{Z}),R)\), where \(R\) denotes the randomness.
๊ฐ’๋งŒ. ๊ฐ \(g_{i}\)๊ฐ€ \([T]\)์—์„œ ๋ฌด์ž‘์œ„๋กœ ๊ท ์ผํ•˜๊ฒŒ ์„ ํƒ๋œ i.i.d.์ธ ๊ฒฝ์šฐ, \(g_{1},\ldots,g_{N}\in[T]\)๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ _๊ฐ’_์œผ๋กœ ์ƒ๊ฐํ•˜์‹ญ์‹œ์˜ค. ๋งค๊ฐœ๋ณ€์ˆ˜ \(W\)๋กœ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ๋ชจ๋ธ์ด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ \(\mathcal{Z}=\big{(}g_{1},...,g_{N}\big{)}\)์—์„œ ํ›ˆ๋ จ๋œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์–ด์ง„ ์ธ๋ฑ์Šค \(i\in[N]\)์— ๋Œ€ํ•ด ์ด ๋ชจ๋ธ์ด \(g_{i}\)์— ํ•ด๋‹นํ•˜๋Š” ๋ฌด์ž‘์œ„ ์˜ˆ์ธก \(f_{i}\)๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์„ \(f_{i}(W(\mathcal{Z}),R)\)๋กœ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ \(R\)์€ ๋ฌด์ž‘์œ„์„ฑ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ฐ’ ์ „์šฉ.g_{1},\ldots,g_{N}\in[T]\), ์—ฌ๊ธฐ์„œ ๊ฐ \(g_{i}\)๋Š” \([T]\)์—์„œ ๋ฌด์ž‘์œ„๋กœ ๊ท ์ผํ•˜๊ฒŒ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ _๊ฐ’_์œผ๋กœ ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. W\๋กœ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ๋ชจ๋ธ์ด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ \(\mathcal{Z}=\big{(}g_{1},...,g_{N}\big{)}\)์— ๋Œ€ํ•ด ํ•™์Šต๋œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์ด ์ฃผ์–ด์ง„ ์ธ๋ฑ์Šค \(i\in[N]\),์— ๋Œ€ํ•ด \(g_{i}\)์— ๋Œ€์‘ํ•˜๋Š” ์ž„์˜์˜ ์˜ˆ์ธก \(f_{i}\)๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์„ \(f_{i}(W(\mathcal{Z}),R)\)๋กœ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ \(R\)์€ ๋ฌด์ž‘์œ„์„ฑ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
The contributions of this paper are threefold: This paper offers a comprehensive framework for understanding the RAG domain, identifying areas for improvement and challenges for future research. It provides a detailed analysis of RAG's core technologies, examining their strengths in addressing retrieval and generation. Additionally, it introduces the evaluation methods used in RAG research, highlighting current challenges and suggesting promising directions for future studies.
์ด ๋…ผ๋ฌธ์˜ ๊ธฐ์—ฌ๋Š” ์„ธ ๊ฐ€์ง€๋กœ ์š”์•ฝ๋ฉ๋‹ˆ๋‹ค: ์ฒซ์งธ, RAG ๋„๋ฉ”์ธ์„ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•œ ํฌ๊ด„์ ์ธ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ๊ณตํ•˜์—ฌ ๊ฐœ์„ ํ•  ์˜์—ญ๊ณผ ํ–ฅํ›„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ๋„์ „ ๊ณผ์ œ๋ฅผ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, RAG์˜ ํ•ต์‹ฌ ๊ธฐ์ˆ ์— ๋Œ€ํ•œ ์ƒ์„ธํ•œ ๋ถ„์„์„ ํ†ตํ•ด ๊ฒ€์ƒ‰ ๋ฐ ์ƒ์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์žˆ์–ด ์ด๋“ค์˜ ๊ฐ•์ ์„ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ์…‹์งธ, RAG ์—ฐ๊ตฌ์—์„œ ์‚ฌ์šฉ๋˜๋Š” ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•˜๋ฉฐ ํ˜„์žฌ์˜ ๋„์ „ ๊ณผ์ œ๋ฅผ ๊ฐ•์กฐํ•˜๊ณ  ํ–ฅํ›„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์˜ ๊ธฐ์—ฌ๋Š” ์„ธ ๊ฐ€์ง€์ž…๋‹ˆ๋‹ค: ์ด ๋ฐฑ์„œ์—์„œ๋Š” RAG ์˜์—ญ์„ ์ดํ•ดํ•˜๊ณ  ๊ฐœ์„ ์ด ํ•„์š”ํ•œ ์˜์—ญ๊ณผ ํ–ฅํ›„ ์—ฐ๊ตฌ ๊ณผ์ œ๋ฅผ ํŒŒ์•…ํ•˜๊ธฐ ์œ„ํ•œ ํฌ๊ด„์ ์ธ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๊ฒ€์ƒ‰ ๋ฐ ์ƒ์„ฑ์„ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์žˆ์–ด RAG์˜ ๊ฐ•์ ์„ ์‚ดํŽด๋ณด๊ณ , RAG์˜ ํ•ต์‹ฌ ๊ธฐ์ˆ ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋ถ„์„์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, RAG ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋˜๋Š” ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•˜๊ณ  ํ˜„์žฌ์˜ ๊ณผ์ œ๋ฅผ ๊ฐ•์กฐํ•˜๋ฉฐ ํ–ฅํ›„ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์„ ์ œ์•ˆํ•ฉ๋‹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)์€ ์—”ํ‹ฐํ‹ฐ ๋งํฌ๋ฅผ ํ†ตํ•ด ๊ตฌ์กฐ์ ์ด๊ณ  ์˜๋ฏธ์ ์œผ๋กœ ํ’๋ถ€ํ•œ ์ •๋ณด๋ฅผ ์›์‹œ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์— ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐ ์ „๋…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ํ’๋ถ€ํ™” ๊ณผ์ •์€ ๋ฐ์ดํ„ฐ๋ฅผ ๋” ์ผ๊ด€๋˜๊ฒŒ ๊ตฌ์กฐํ™”ํ•˜๊ณ  ์ฟผ๋ฆฌ์— ๋” ์ ํ•ฉํ•˜๊ฒŒ ๋งŒ๋“ค ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋ชจ๋ธ์˜ ๊ฒ€์ƒ‰ ํšจ์œจ์„ฑ๋„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Š” ์ •ํ™•ํ•˜๊ณ  ์—ฐ๊ฒฐ๋œ ์ง€์‹์„ ํ™œ์šฉํ•˜์—ฌ ๋ชจ๋ธ์˜ ์ดํ•ด๋ ฅ๊ณผ ๊ด€๋ จ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œ์ผœ ์ „๋ฐ˜์ ์ธ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.
๋˜ํ•œ, KnowledGPT(Wang et al., 2023)๋Š” ์—”ํ‹ฐํ‹ฐ ๋งํฌ๋ฅผ ํ†ตํ•ด ๊ตฌ์กฐํ™”๋˜๊ณ  ์˜๋ฏธ์ ์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  ์ฟผ๋ฆฌ์— ๋” ์ž˜ ๋งž๋„๋ก ๋งŒ๋“ค ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋ชจ๋ธ์˜ ๊ฒ€์ƒ‰ ํšจ์œจ์„ฑ๋„ ๋†’์ž…๋‹ˆ๋‹ค. ์ •ํ™•ํ•˜๊ฒŒ ์—ฐ๊ฒฐ๋œ ์ง€์‹์„ ํ™œ์šฉํ•˜์—ฌ ๋ชจ๋ธ์˜ ์ดํ•ด๋ ฅ๊ณผ ๊ด€๋ จ์„ฑ ๋†’์€ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œ์ผœ ์ „๋ฐ˜์ ์ธ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.
To evaluate our models effectively, we annotated 200 samples across ten distinct task types and utilized GPT-4 for evaluation. Our experiments demonstrated that the proposed models significantly outperformed the original LLaMA in Chinese understanding and generation tasks. We also tested our models on C-Eval datasets. The results show that the proposed model could achieve significant improvements and show competitive performance to the models with several times bigger sizes.
์šฐ๋ฆฌ ๋ชจ๋ธ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์—ด ๊ฐ€์ง€ ๋‹ค๋ฅธ ์ž‘์—… ์œ ํ˜•์— ๊ฑธ์ณ 200๊ฐœ์˜ ์ƒ˜ํ”Œ์— ์ฃผ์„์„ ๋‹ฌ๊ณ  ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด GPT-4๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์ œ์•ˆ๋œ ๋ชจ๋ธ์ด ์ค‘๊ตญ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ์ž‘์—…์—์„œ ์›๋ž˜์˜ LLaMA๋ฅผ ํฌ๊ฒŒ ๋Šฅ๊ฐ€ํ•˜๋Š” ์„ฑ๋Šฅ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์šฐ๋ฆฌ๋Š” C-Eval ๋ฐ์ดํ„ฐ์…‹์—์„œ ๋ชจ๋ธ์„ ํ…Œ์ŠคํŠธํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ๋Š” ์ œ์•ˆ๋œ ๋ชจ๋ธ์ด ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ์ด๋ฃจ์—ˆ์œผ๋ฉฐ, ๋ช‡ ๋ฐฐ ๋” ํฐ ํฌ๊ธฐ์˜ ๋ชจ๋ธ๋“ค๊ณผ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ณด์ผ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๋ชจ๋ธ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด 10๊ฐ€์ง€ ์ž‘์—… ์œ ํ˜•์— ๊ฑธ์ณ 200๊ฐœ์˜ ์ƒ˜ํ”Œ์„ ์ฃผ์„ ์ฒ˜๋ฆฌํ•˜๊ณ  ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด GPT-4๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์ œ์•ˆ๋œ ๋ชจ๋ธ์ด ์ค‘๊ตญ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ์ž‘์—…์—์„œ ๊ธฐ์กด LLaMA๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ C-Eval ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ชจ๋ธ์„ ํ…Œ์ŠคํŠธํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ์ œ์•ˆ๋œ ๋ชจ๋ธ์ด ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ๋‹ฌ์„ฑํ•˜๊ณ  ๋ช‡ ๋ฐฐ ๋” ํฐ ํฌ๊ธฐ์˜ ๋ชจ๋ธ๊ณผ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค.
Retrieval-Augmented Language Model (RALM) is the process of refining the output of the LM with retrieved information to obtain a satisfactory result for the user. This section provides a detailed definition of the different modes of RALM by categorising the ways in which the retriever interacts with the language model. The specific categorization of interactions can be seen in Figure 2. In addition, the development history of each interaction method can be seen in Figure 3.
Retrieval-Augmented Language Model (RALM)์€ ๊ฒ€์ƒ‰๋œ ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LM์˜ ์ถœ๋ ฅ์„ ๊ฐœ์„ ํ•˜๊ณ  ์‚ฌ์šฉ์ž์—๊ฒŒ ๋งŒ์กฑ์Šค๋Ÿฌ์šด ๊ฒฐ๊ณผ๋ฅผ ์–ป๋Š” ๊ณผ์ •์ž…๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ๊ฒ€์ƒ‰๊ธฐ๊ฐ€ ์–ธ์–ด ๋ชจ๋ธ๊ณผ ์ƒํ˜ธ์ž‘์šฉํ•˜๋Š” ๋ฐฉ์‹์„ ๋ถ„๋ฅ˜ํ•˜์—ฌ RALM์˜ ๋‹ค์–‘ํ•œ ๋ชจ๋“œ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •์˜๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ƒํ˜ธ์ž‘์šฉ์˜ ๊ตฌ์ฒด์ ์ธ ๋ถ„๋ฅ˜๋Š” Figure 2์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๊ฐ ์ƒํ˜ธ์ž‘์šฉ ๋ฐฉ๋ฒ•์˜ ๊ฐœ๋ฐœ ์—ญ์‚ฌ๋Š” Figure 3์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ฆฌํŠธ๋ฆฌ๋ฒŒ ์ฆ๊ฐ• ์–ธ์–ด ๋ชจ๋ธ(RALM)์€ ๊ฒ€์ƒ‰๋œ ์ •๋ณด๋กœ LM์˜ ์ถœ๋ ฅ์„ ๊ตฌ์ฒดํ™”ํ•˜์—ฌ ์‚ฌ์šฉ์ž์—๊ฒŒ ๋งŒ์กฑ์Šค๋Ÿฌ์šด ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•˜๋Š” ํ”„๋กœ์„ธ์Šค์ž…๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ๊ฒ€์ƒ‰๊ธฐ๊ฐ€ ์–ธ์–ด ๋ชจ๋ธ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ๋ฐฉ์‹์„ ๋ถ„๋ฅ˜ํ•˜์—ฌ RALM์˜ ๋‹ค์–‘ํ•œ ๋ชจ๋“œ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •์˜๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ƒํ˜ธ ์ž‘์šฉ์˜ ๊ตฌ์ฒด์ ์ธ ๋ถ„๋ฅ˜๋Š” ๊ทธ๋ฆผ 2์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฐ ์ƒํ˜ธ ์ž‘์šฉ ๋ฐฉ๋ฒ•์˜ ๊ฐœ๋ฐœ ์—ฐํ˜์€ ๊ทธ๋ฆผ 3์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
1. Engaging in activities that cause harm (e.g., self-harm, detonating a bomb, conducting fraud). 2. Threatening or expressing willingness to engage in activities that cause harm. 3. Organizing and planning activities that cause harm. 4. Convincing others to engage in activities that cause harm. This includes explicit recruitment. 5. Covering up or attempting to cover up activities that cause harm. 6. Justifying or encouraging others to engage in activities that cause harm.
1. ํ•ด๋ฅผ ๋ผ์น˜๋Š” ํ™œ๋™์— ์ฐธ์—ฌํ•˜๋Š” ๊ฒƒ (์˜ˆ: ์žํ•ด, ํญํƒ„ ํญ๋ฐœ, ์‚ฌ๊ธฐ ํ–‰์œ„). 2. ํ•ด๋ฅผ ๋ผ์น˜๋Š” ํ™œ๋™์— ์ฐธ์—ฌํ•˜๊ฒ ๋‹ค๊ณ  ์œ„ํ˜‘ํ•˜๊ฑฐ๋‚˜ ์˜์‚ฌ๋ฅผ ํ‘œํ˜„ํ•˜๋Š” ๊ฒƒ. 3. ํ•ด๋ฅผ ๋ผ์น˜๋Š” ํ™œ๋™์„ ์กฐ์งํ•˜๊ณ  ๊ณ„ํšํ•˜๋Š” ๊ฒƒ. 4. ๋‹ค๋ฅธ ์‚ฌ๋žŒ์„ ํ•ด๋ฅผ ๋ผ์น˜๋Š” ํ™œ๋™์— ์ฐธ์—ฌํ•˜๋„๋ก ์„ค๋“ํ•˜๋Š” ๊ฒƒ. ์—ฌ๊ธฐ์—๋Š” ๋ช…์‹œ์ ์ธ ๋ชจ์ง‘์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. 5. ํ•ด๋ฅผ ๋ผ์น˜๋Š” ํ™œ๋™์„ ์€ํํ•˜๊ฑฐ๋‚˜ ์€ํํ•˜๋ ค๊ณ  ์‹œ๋„ํ•˜๋Š” ๊ฒƒ. 6. ๋‹ค๋ฅธ ์‚ฌ๋žŒ์ด ํ•ด๋ฅผ ๋ผ์น˜๋Š” ํ™œ๋™์— ์ฐธ์—ฌํ•˜๋„๋ก ์ •๋‹นํ™”ํ•˜๊ฑฐ๋‚˜ ๊ถŒ์žฅํ•˜๋Š” ๊ฒƒ.
1. ํ•ด๋ฅผ ์ž…ํžˆ๋Š” ํ–‰์œ„(์˜ˆ: ์žํ•ด, ํญํƒ„ ํญ๋ฐœ, ์‚ฌ๊ธฐ ํ–‰์œ„)์— ๊ฐ€๋‹ดํ•˜๋Š” ํ–‰์œ„. 2. ํ•ด๋ฅผ ์ž…ํžˆ๋Š” ํ–‰์œ„์— ๊ฐ€๋‹ดํ•˜๊ฒ ๋‹ค๊ณ  ์œ„ํ˜‘ํ•˜๊ฑฐ๋‚˜ ์˜์‚ฌ๋ฅผ ํ‘œ์‹œํ•˜๋Š” ํ–‰์œ„. 3. ํ•ด๋ฅผ ์ž…ํžˆ๋Š” ํ–‰์œ„๋ฅผ ์กฐ์งํ•˜๊ณ  ๊ณ„ํšํ•˜๋Š” ํ–‰์œ„. 4. ๋‹ค๋ฅธ ์‚ฌ๋žŒ์„ ์„ค๋“ํ•˜์—ฌ ํ•ด๋ฅผ ์ž…ํžˆ๋Š” ํ–‰์œ„์— ๊ฐ€๋‹ดํ•˜๊ฒŒ ํ•˜๋Š” ํ–‰์œ„. ์—ฌ๊ธฐ์—๋Š” ๋ช…์‹œ์ ์ธ ๋ชจ์ง‘์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. 5. ํ•ด๋ฅผ ์ž…ํžˆ๋Š” ํ–‰์œ„๋ฅผ ์€ํํ•˜๊ฑฐ๋‚˜ ์€ํํ•˜๋ ค๊ณ  ์‹œ๋„ํ•˜๋Š” ํ–‰์œ„. 6. ๋‹ค๋ฅธ ์‚ฌ๋žŒ์ด ํ•ด๋ฅผ ์ž…ํžˆ๋Š” ํ–‰์œ„์— ๊ฐ€๋‹ดํ•˜๋„๋ก ์ •๋‹นํ™”ํ•˜๊ฑฐ๋‚˜ ๊ฒฉ๋ คํ•˜๋Š” ํ–‰์œ„.
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.).
์šฐ๋ฆฌ๋Š” ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์—์„œ ๋ฌธ์ œ๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ ์˜ˆ๋ฅผ ๋“ค์–ด ziya (Zhang et al., 2022), goakao, CIF-Bench (Li et al., 2024) ๋“ฑ์„ ํ†ตํ•ด ์–ด๋ ค์šด ์ค‘๊ตญ์–ด ์ง€์‹œ ์ดํ•ด ๋ฐ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ (CHC-Bench)๋ฅผ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค. CHC-Bench์˜ ๋ฌธ์ œ ๋ฒ”์ฃผ์—๋Š” ๊ธ€์“ฐ๊ธฐ, ์ธ๋ฌธํ•™ ๋ฐ ์—ญ์‚ฌ, ๊ณผํ•™, ์ˆ˜ํ•™, ๋…ํ•ด, ์—ญํ•  ๋†€์ด, ๊ทธ๋ฆฌ๊ณ  ์ค‘๊ตญ์–ด ์ดํ•ด์˜ ์–ด๋ ค์šด ์‚ฌ๋ก€๋“ค(์˜ˆ: ์ค‘๊ตญ์–ด ๋‹จ์–ด ๋ฐœ์Œ, ๊ณ ๋Œ€ ์ค‘๊ตญ์–ด ์ดํ•ด ๋“ฑ)์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
๋‹ค์–‘ํ•œ ์†Œ์Šค(์˜ˆ: ziya(Zhang et al., 2022), goakao, CIF-Bench(Li et al., 2024))์—์„œ ๋ฌธ์ œ๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ ์–ด๋ ค์šด ๊ฒฝ์šฐ์˜ ์ค‘๊ตญ์–ด ์ดํ•ด ๋ฐ ์ˆ˜ํ–‰ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ(์ดํ•˜ CHC-Bench)๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. CHC-Bench์˜ ๋ฌธ์ œ ์นดํ…Œ๊ณ ๋ฆฌ์—๋Š” ๊ธ€์“ฐ๊ธฐ, ์ธ๋ฌธํ•™ ๋ฐ ์—ญ์‚ฌ, ๊ณผํ•™, ์ˆ˜ํ•™, ๋…ํ•ด, ๋กคํ”Œ๋ ˆ์ž‰, ์ค‘๊ตญ์–ด ์ดํ•ด์˜ ์–ด๋ ค์šด ์‚ฌ๋ก€(์˜ˆ: ์ค‘๊ตญ์–ด ๋‹จ์–ด ๋ฐœ์Œ, ๊ณ ๋Œ€ ์ค‘๊ตญ์–ด ์ดํ•ด ๋“ฑ)๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
In this section, we answer RQ1 of _how often an LLM performs the first hop of reasoning while pro cessing two-hop prompts_. We first introduce EntRec as a metric to approximate the LLM's internal recall of the bridge entity upon its descriptive mention in a prompt (SS5.1). Next, we propose to measure how often this recall increases when changing the input prompt to indirectly mention the bridge entity (SS5.2). Then, we evaluate this using TwoHopFact and answer RQ1 (SS5.3).
์ด ์„น์…˜์—์„œ๋Š” _LLM์ด ๋‘ ๋‹จ๊ณ„์˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์˜ ์ถ”๋ก ์„ ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์ˆ˜ํ–‰ํ•˜๋Š”์ง€_์— ๋Œ€ํ•œ RQ1์— ๋‹ตํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ €, ํ”„๋กฌํ”„ํŠธ์—์„œ ์„ค๋ช…์ ์œผ๋กœ ์–ธ๊ธ‰๋œ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์— ๋Œ€ํ•œ LLM์˜ ๋‚ด๋ถ€ ํšŒ์ƒ์„ ๊ทผ์‚ฌํ™”ํ•˜๋Š” ์ง€ํ‘œ๋กœ EntRec์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค (SS5.1). ๋‹ค์Œ์œผ๋กœ, ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณ€๊ฒฝํ•˜์—ฌ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ๋ฅผ ๊ฐ„์ ‘์ ์œผ๋กœ ์–ธ๊ธ‰ํ•  ๋•Œ ์ด ํšŒ์ƒ์ด ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์ฆ๊ฐ€ํ•˜๋Š”์ง€ ์ธก์ •ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค (SS5.2). ๊ทธ๋Ÿฐ ๋‹ค์Œ, TwoHopFact๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฅผ ํ‰๊ฐ€ํ•˜๊ณ  RQ1์— ๋‹ตํ•ฉ๋‹ˆ๋‹ค (SS5.3).
์ด ์„น์…˜์—์„œ๋Š” _๋‘ ํ™‰ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋™์•ˆ LLM์ด ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์ฒซ ๋ฒˆ์งธ ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๋Š”์ง€์— ๋Œ€ํ•œ RQ1_์— ๋‹ตํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ €, ํ”„๋กฌํ”„ํŠธ์—์„œ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ๋ฅผ ์„ค๋ช…ํ•  ๋•Œ LLM์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์„ ๊ทผ์‚ฌํ™”ํ•˜๋Š” ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ EntRec์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค(SS5.1). ๋‹ค์Œ์œผ๋กœ, ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณ€๊ฒฝํ•˜์—ฌ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ๋ฅผ ๊ฐ„์ ‘์ ์œผ๋กœ ์–ธ๊ธ‰ํ•  ๋•Œ ์ด ๋ฆฌ์ฝœ์ด ์–ผ๋งˆ๋‚˜ ์ž์ฃผ ์ฆ๊ฐ€ํ•˜๋Š”์ง€ ์ธก์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค(SS5.2). ๊ทธ๋Ÿฐ ๋‹ค์Œ TwoHopFact๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฅผ ํ‰๊ฐ€ํ•˜๊ณ  RQ1(SS5.3)์— ๋‹ตํ•ฉ๋‹ˆ๋‹ค.
The PKG framework, introduced by (Luo et al., 2023), represents an approach to customizing the output of LMs. By generating background knowledge internally using a pre-trained model, PKG eliminates the need for traditional external retrieval processes. This method directly integrates domain- or task-specific knowledge into the generation step, significantly enhancing the LM's capacity to produce responses that are specifically tailored to the given context or requirements.
PKG ํ”„๋ ˆ์ž„์›Œํฌ๋Š” (Luo et al., 2023)์— ์˜ํ•ด ์†Œ๊ฐœ๋œ LMs์˜ ์ถœ๋ ฅ์„ ๋งž์ถคํ™”ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‚ด๋ถ€์ ์œผ๋กœ ๋ฐฐ๊ฒฝ ์ง€์‹์„ ์ƒ์„ฑํ•จ์œผ๋กœ์จ, PKG๋Š” ์ „ํ†ต์ ์ธ ์™ธ๋ถ€ ๊ฒ€์ƒ‰ ๊ณผ์ •์„ ํ•„์š” ์—†๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋„๋ฉ”์ธ ๋˜๋Š” ์ž‘์—…๋ณ„ ์ง€์‹์„ ์ƒ์„ฑ ๋‹จ๊ณ„์— ์ง์ ‘ ํ†ตํ•ฉํ•˜์—ฌ, ์ฃผ์–ด์ง„ ๋งฅ๋ฝ์ด๋‚˜ ์š”๊ตฌ ์‚ฌํ•ญ์— ๋งž์ถ˜ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” LM์˜ ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
(Luo et al., 2023)์—์„œ ์†Œ๊ฐœํ•œ PKG ํ”„๋ ˆ์ž„์›Œํฌ๋Š” LM์˜ ์ถœ๋ ฅ์„ ์‚ฌ์šฉ์ž ์ง€์ •ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. PKG๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‚ด๋ถ€์ ์œผ๋กœ ๋ฐฐ๊ฒฝ ์ง€์‹์„ ์ƒ์„ฑํ•จ์œผ๋กœ์จ ๊ธฐ์กด์˜ ์™ธ๋ถ€ ๊ฒ€์ƒ‰ ํ”„๋กœ์„ธ์Šค๊ฐ€ ํ•„์š”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋„๋ฉ”์ธ ๋˜๋Š” ์ž‘์—…๋ณ„ ์ง€์‹์„ ์ƒ์„ฑ ๋‹จ๊ณ„์— ์ง์ ‘ ํ†ตํ•ฉํ•˜์—ฌ ์ฃผ์–ด์ง„ ์ปจํ…์ŠคํŠธ ๋˜๋Š” ์š”๊ตฌ ์‚ฌํ•ญ์— ๋งž๊ฒŒ ํŠน๋ณ„ํžˆ ๋งž์ถคํ™”๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” LM์˜ ์—ญ๋Ÿ‰์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
As depicted in Figure 2, the architecture of our AIOS is organized into three distinct layers: the application layer, the kernel layer, and the hardware layer. This layered architecture ensures a clear delineation of responsibilities across the system. Each higher layer abstracts the complexities of the layers below it, facilitating interaction through interfaces or specific modules, thereby enhancing modularity and simplifying system interactions across different layers.
Figure 2์— ๋‚˜ํƒ€๋‚œ ๋ฐ”์™€ ๊ฐ™์ด, ์šฐ๋ฆฌ์˜ AIOS ์•„ํ‚คํ…์ฒ˜๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋ ˆ์ด์–ด, ์ปค๋„ ๋ ˆ์ด์–ด, ํ•˜๋“œ์›จ์–ด ๋ ˆ์ด์–ด์˜ ์„ธ ๊ฐ€์ง€ ๊ตฌ๋ณ„๋œ ๋ ˆ์ด์–ด๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณ„์ธตํ™”๋œ ์•„ํ‚คํ…์ฒ˜๋Š” ์‹œ์Šคํ…œ ์ „๋ฐ˜์— ๊ฑธ์ณ ๋ช…ํ™•ํ•œ ์ฑ…์ž„ ๋ถ„๋‹ด์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์ƒ์œ„ ๋ ˆ์ด์–ด๋Š” ํ•˜์œ„ ๋ ˆ์ด์–ด์˜ ๋ณต์žก์„ฑ์„ ์ถ”์ƒํ™”ํ•˜์—ฌ ์ธํ„ฐํŽ˜์ด์Šค๋‚˜ ํŠน์ • ๋ชจ๋“ˆ์„ ํ†ตํ•ด ์ƒํ˜ธ์ž‘์šฉ์„ ์ด‰์ง„ํ•˜๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋“ˆํ™”๊ฐ€ ๊ฐ•ํ™”๋˜๊ณ  ๋‹ค์–‘ํ•œ ๋ ˆ์ด์–ด ๊ฐ„์˜ ์‹œ์Šคํ…œ ์ƒํ˜ธ์ž‘์šฉ์ด ๋‹จ์ˆœํ™”๋ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2์™€ ๊ฐ™์ด AIOS์˜ ์•„ํ‚คํ…์ฒ˜๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๊ณ„์ธต, ์ปค๋„ ๊ณ„์ธต, ํ•˜๋“œ์›จ์–ด ๊ณ„์ธต์˜ ์„ธ ๊ฐ€์ง€ ๊ณ„์ธต์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด ๊ณ„์ธต์  ์•„ํ‚คํ…์ฒ˜๋Š” ์‹œ์Šคํ…œ ์ „์ฒด์— ๊ฑธ์ณ ์ฑ…์ž„์˜ ๋ช…ํ™•ํ•œ ๊ตฌ๋ถ„์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์ƒ์œ„ ๊ณ„์ธต์€ ๊ทธ ์•„๋ž˜ ๊ณ„์ธต์˜ ๋ณต์žก์„ฑ์„ ์ถ”์ƒํ™”ํ•˜์—ฌ ์ธํ„ฐํŽ˜์ด์Šค ๋˜๋Š” ํŠน์ • ๋ชจ๋“ˆ์„ ํ†ตํ•ด ์ƒํ˜ธ ์ž‘์šฉ์„ ์šฉ์ดํ•˜๊ฒŒ ํ•˜๋ฏ€๋กœ ๋ชจ๋“ˆ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ์—ฌ๋Ÿฌ ๊ณ„์ธต ๊ฐ„์˜ ์‹œ์Šคํ…œ ์ƒํ˜ธ ์ž‘์šฉ์„ ๊ฐ„์†Œํ™”ํ•ฉ๋‹ˆ๋‹ค.
In this section, we evaluate both the correctness and the performance of AIOS modules when multiple agents are running in parallel in AIOS. Our investigation is guided by two research questions: firstly, whether the LLM responses to agent requests are consistent after agent suspension and transition to another agent, and secondly, how is the performance of AIOS scheduling in improving balance of waiting and turnaround time relative to non-scheduled (sequential) execution.
์ด ์„น์…˜์—์„œ๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๊ฐ€ AIOS์—์„œ ๋ณ‘๋ ฌ๋กœ ์‹คํ–‰๋  ๋•Œ AIOS ๋ชจ๋“ˆ์˜ ์ •ํ™•์„ฑ๊ณผ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์กฐ์‚ฌ๋Š” ๋‘ ๊ฐ€์ง€ ์—ฐ๊ตฌ ์งˆ๋ฌธ์— ์˜ํ•ด ์•ˆ๋‚ด๋ฉ๋‹ˆ๋‹ค: ์ฒซ์งธ, ์—์ด์ „ํŠธ ์š”์ฒญ์— ๋Œ€ํ•œ LLM ์‘๋‹ต์ด ์—์ด์ „ํŠธ ์ค‘๋‹จ ํ›„ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๋กœ ์ „ํ™˜๋œ ํ›„์—๋„ ์ผ๊ด€์„ฑ์ด ์žˆ๋Š”์ง€, ๋‘˜์งธ, AIOS ์Šค์ผ€์ค„๋ง์ด ๋น„์Šค์ผ€์ค„๋ง(์ˆœ์ฐจ์ ) ์‹คํ–‰์— ๋น„ํ•ด ๋Œ€๊ธฐ ์‹œ๊ฐ„๊ณผ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์˜ ๊ท ํ˜•์„ ๊ฐœ์„ ํ•˜๋Š” ์„ฑ๋Šฅ์ด ์–ด๋–ค์ง€์ž…๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๊ฐ€ AIOS์—์„œ ๋ณ‘๋ ฌ๋กœ ์‹คํ–‰๋  ๋•Œ AIOS ๋ชจ๋“ˆ์˜ ์ •ํ™•์„ฑ๊ณผ ์„ฑ๋Šฅ์„ ๋ชจ๋‘ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ์กฐ์‚ฌ๋Š” ๋‘ ๊ฐ€์ง€ ์—ฐ๊ตฌ ์งˆ๋ฌธ์— ๋”ฐ๋ผ ์ด๋ฃจ์–ด์ง‘๋‹ˆ๋‹ค. ์ฒซ์งธ, ์—์ด์ „ํŠธ ์š”์ฒญ์— ๋Œ€ํ•œ LLM ์‘๋‹ต์ด ์—์ด์ „ํŠธ ์ผ์‹œ ์ค‘๋‹จ ํ›„ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๋กœ ์ „ํ™˜ํ•œ ํ›„์—๋„ ์ผ๊ด€์„ฑ์ด ์œ ์ง€๋˜๋Š”์ง€ ์—ฌ๋ถ€, ๋‘˜์งธ, ๋Œ€๊ธฐ ๋ฐ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์˜ ๊ท ํ˜•์„ ๊ฐœ์„ ํ•˜๋Š” AIOS ์Šค์ผ€์ค„๋ง์˜ ์„ฑ๋Šฅ์ด ๋น„์Šค์ผ€์ค„(์ˆœ์ฐจ์ ) ์‹คํ–‰์— ๋น„ํ•ด ์–ด๋–ป๊ฒŒ ๊ฐœ์„ ๋˜๋Š”์ง€์— ๋Œ€ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
In heterogeneous combination experiment, GPT-3.5-Turbo is used for generating the intermediate answer with \(K=8\), and GPT-4 is then employed to solve for the final answer with \(K=32\). In Figure 7 (right), compared with the result of GPT-4 with \(K=32\), the hierarchical method improves performance from \(35\%\) to \(47\%\), suggesting the deployment of different LLMs at the corresponding level of problem-solving can improve the performance in a cost-effective manner.
์ด์งˆ์ ์ธ ์กฐํ•ฉ ์‹คํ—˜์—์„œ, GPT-3.5-Turbo๋Š” \(K=8\)๋กœ ์ค‘๊ฐ„ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋ฉฐ, GPT-4๋Š” \(K=32\)๋กœ ์ตœ์ข… ๋‹ต๋ณ€์„ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. Figure 7 (์˜ค๋ฅธ์ชฝ)์—์„œ, \(K=32\)๋กœ GPT-4์˜ ๊ฒฐ๊ณผ์™€ ๋น„๊ตํ–ˆ์„ ๋•Œ, ๊ณ„์ธต์  ๋ฐฉ๋ฒ•์€ ์„ฑ๋Šฅ์„ \(35\%\)์—์„œ \(47\%\)๋กœ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Š” ๋ฌธ์ œ ํ•ด๊ฒฐ์˜ ํ•ด๋‹น ๋‹จ๊ณ„์—์„œ ๋‹ค๋ฅธ LLM์„ ๋ฐฐ์น˜ํ•˜๋Š” ๊ฒƒ์ด ๋น„์šฉ ํšจ์œจ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
์ด์ข… ์กฐํ•ฉ ์‹คํ—˜์—์„œ๋Š” ์ค‘๊ฐ„ ๋‹ต์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด GPT-3.5-Turbo๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  \(K=8\)๋กœ ์ตœ์ข… ๋‹ต์„ ๊ตฌํ•˜๋Š” ๋ฐ GPT-4๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 7(์˜ค๋ฅธ์ชฝ)์—์„œ \(K=32\)์ธ GPT-4์˜ ๊ฒฐ๊ณผ์™€ ๋น„๊ตํ–ˆ์„ ๋•Œ, ๊ณ„์ธต์  ๋ฐฉ๋ฒ•์€ ์„ฑ๋Šฅ์„ \(35\%์—์„œ \(47\%๋กœ ๊ฐœ์„ ํ•˜์—ฌ ๋ฌธ์ œ ํ•ด๊ฒฐ์˜ ํ•ด๋‹น ์ˆ˜์ค€์—์„œ ์„œ๋กœ ๋‹ค๋ฅธ LLM์„ ๋ฐฐํฌํ•˜๋ฉด ๋น„์šฉ ํšจ์œจ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
The scaling law formula thus provides a theoretical framework to predict how adjustments in model size, data size, and computational resources are likely to impact the performance of machine learning models. This formula is instrumental in guiding researchers and practitioners towards making informed decisions about resource allocation during the model development process, aiming for an optimal balance between these factors to achieve the best possible model performance.
๋”ฐ๋ผ์„œ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™ ๊ณต์‹์€ ๋ชจ๋ธ ํฌ๊ธฐ, ๋ฐ์ดํ„ฐ ํฌ๊ธฐ ๋ฐ ๊ณ„์‚ฐ ์ž์›์˜ ์กฐ์ •์ด ๊ธฐ๊ณ„ ํ•™์Šต ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น ์ง€ ์˜ˆ์ธกํ•˜๋Š” ์ด๋ก ์  ํ‹€์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ณต์‹์€ ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž๊ฐ€ ๋ชจ๋ธ ๊ฐœ๋ฐœ ๊ณผ์ •์—์„œ ์ž์› ํ• ๋‹น์— ๋Œ€ํ•œ ์ •๋ณด์— ์ž…๊ฐํ•œ ๊ฒฐ์ •์„ ๋‚ด๋ฆฌ๋„๋ก ์•ˆ๋‚ดํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋ฉฐ, ์ด๋Ÿฌํ•œ ์š”์†Œ๋“ค ๊ฐ„์˜ ์ตœ์ ์˜ ๊ท ํ˜•์„ ๋งž์ถ”์–ด ๊ฐ€๋Šฅํ•œ ์ตœ๊ณ ์˜ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™ ๊ณต์‹์€ ๋ชจ๋ธ ํฌ๊ธฐ, ๋ฐ์ดํ„ฐ ํฌ๊ธฐ ๋ฐ ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค์˜ ์กฐ์ •์ด ๋จธ์‹  ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น ์ง€ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋Š” ์ด๋ก ์  ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ณต์‹์€ ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž๊ฐ€ ๋ชจ๋ธ ๊ฐœ๋ฐœ ๊ณผ์ •์—์„œ ๋ฆฌ์†Œ์Šค ํ• ๋‹น์— ๋Œ€ํ•œ ์ •๋ณด์— ์ž…๊ฐํ•œ ๊ฒฐ์ •์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ๋„์›€์ด ๋˜๋ฉฐ, ์ด๋Ÿฌํ•œ ์š”์†Œ ๊ฐ„์˜ ์ตœ์ ์˜ ๊ท ํ˜•์„ ํ†ตํ•ด ์ตœ์ƒ์˜ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
1. The cost of synthetic data generation is a one-time investment, and we may train many models of varying scales once the data is generated. 2. Data generation is 100% parallelizable, whereas training requires the availability of a big cluster with fast inter-node connectivity. This is much more expensive. On the other hand, generation can be thought of as a side process that can fill in the empty GPUs in any large-scale compute cluster, and runs on single GPU machines.
1. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๋น„์šฉ์€ ์ผํšŒ์„ฑ ํˆฌ์ž์ด๋ฉฐ, ๋ฐ์ดํ„ฐ๊ฐ€ ์ƒ์„ฑ๋˜๋ฉด ๋‹ค์–‘ํ•œ ๊ทœ๋ชจ์˜ ์—ฌ๋Ÿฌ ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 2. ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์€ 100% ๋ณ‘๋ ฌํ™”๊ฐ€ ๊ฐ€๋Šฅํ•˜์ง€๋งŒ, ํ›ˆ๋ จ์€ ๋น ๋ฅธ ๋…ธ๋“œ ๊ฐ„ ์—ฐ๊ฒฐ์„ ๊ฐ–์ถ˜ ๋Œ€๊ทœ๋ชจ ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ›จ์”ฌ ๋” ๋น„์šฉ์ด ๋งŽ์ด ๋“ญ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์ƒ์„ฑ์€ ๋Œ€๊ทœ๋ชจ ์ปดํ“จํŒ… ํด๋Ÿฌ์Šคํ„ฐ์˜ ๋นˆ GPU๋ฅผ ์ฑ„์šธ ์ˆ˜ ์žˆ๋Š” ๋ถ€์ˆ˜์ ์ธ ํ”„๋กœ์„ธ์Šค๋กœ ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋‹จ์ผ GPU ๋จธ์‹ ์—์„œ๋„ ์‹คํ–‰๋ฉ๋‹ˆ๋‹ค.
1. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๋น„์šฉ์€ ์ผํšŒ์„ฑ ํˆฌ์ž์ด๋ฉฐ, ๋ฐ์ดํ„ฐ๊ฐ€ ์ƒ์„ฑ๋˜๋ฉด ๋‹ค์–‘ํ•œ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์„ ๋งŽ์ด ํ•™์Šตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 2. ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์€ 100% ๋ณ‘๋ ฌํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ํ•™์Šต์—๋Š” ๋…ธ๋“œ ๊ฐ„ ์—ฐ๊ฒฐ ์†๋„๊ฐ€ ๋น ๋ฅธ ๋Œ€๊ทœ๋ชจ ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ›จ์”ฌ ๋” ๋น„์Œ‰๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์ƒ์„ฑ์€ ๋Œ€๊ทœ๋ชจ ์ปดํ“จํŒ… ํด๋Ÿฌ์Šคํ„ฐ์—์„œ ๋นˆ GPU๋ฅผ ์ฑ„์šธ ์ˆ˜ ์žˆ๋Š” ๋ถ€์ˆ˜์ ์ธ ํ”„๋กœ์„ธ์Šค๋กœ ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋‹จ์ผ GPU ์ปดํ“จํŒ…์—์„œ ์‹คํ–‰๋ฉ๋‹ˆ๋‹ค.
As discussed in Section 3.1, we base our filtering of adult documents only on the URL itself, and not on the content of the documents. This design choice was motivated by: (1) challenges in avoiding overfiltering content from minorities when using ML-based classifiers on the content of documents (Welbl et al., 2021); (2) NSFW words block-list applied on content (such as the one used in C4) also resulting in overfiltering of legal and medical content (Dodge et al., 2021).
3.1์ ˆ์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ์šฐ๋ฆฌ๋Š” ์„ฑ์ธ ๋ฌธ์„œ์˜ ํ•„ํ„ฐ๋ง์„ ๋ฌธ์„œ์˜ ๋‚ด์šฉ์ด ์•„๋‹Œ URL ์ž์ฒด์—๋งŒ ๊ธฐ๋ฐ˜์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ค๊ณ„ ์„ ํƒ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ด์œ ์—์„œ ๋น„๋กฏ๋˜์—ˆ์Šต๋‹ˆ๋‹ค: (1) ๋ฌธ์„œ ๋‚ด์šฉ์— ML ๊ธฐ๋ฐ˜ ๋ถ„๋ฅ˜๊ธฐ๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ ์†Œ์ˆ˜์ž์˜ ์ฝ˜ํ…์ธ ๋ฅผ ๊ณผ๋„ํ•˜๊ฒŒ ํ•„ํ„ฐ๋งํ•˜๋Š” ๋ฌธ์ œ (Welbl et al., 2021); (2) ์ฝ˜ํ…์ธ ์— ์ ์šฉ๋œ NSFW ๋‹จ์–ด ์ฐจ๋‹จ ๋ชฉ๋ก(์˜ˆ: C4์—์„œ ์‚ฌ์šฉ๋œ ๊ฒƒ)์ด ๋ฒ•์  ๋ฐ ์˜๋ฃŒ ์ฝ˜ํ…์ธ ์˜ ๊ณผ๋„ํ•œ ํ•„ํ„ฐ๋ง์„ ์ดˆ๋ž˜ํ•˜๋Š” ๋ฌธ์ œ (Dodge et al., 2021).
3.1์ ˆ์—์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ์„ฑ์ธ์šฉ ๋ฌธ์„œ์˜ ํ•„ํ„ฐ๋ง์€ ๋ฌธ์„œ ๋‚ด์šฉ์ด ์•„๋‹Œ URL ์ž์ฒด๋งŒ์„ ๊ธฐ์ค€์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ค๊ณ„ ๊ฒฐ์ •์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ด์œ ์—์„œ ๋น„๋กฏ๋˜์—ˆ์Šต๋‹ˆ๋‹ค: (1) ๋ฌธ์„œ ๋‚ด์šฉ์— ML ๊ธฐ๋ฐ˜ ๋ถ„๋ฅ˜๊ธฐ๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ ์†Œ์ˆ˜์ž์˜ ์ฝ˜ํ…์ธ ๋ฅผ ๊ณผ๋„ํ•˜๊ฒŒ ํ•„ํ„ฐ๋งํ•˜๋Š” ๊ฒƒ์„ ํ”ผํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ์  (Welbl et al., 2021); (2) ์ฝ˜ํ…์ธ ์— ์ ์šฉ๋˜๋Š” NSFW ๋‹จ์–ด ์ฐจ๋‹จ ๋ชฉ๋ก(์˜ˆ: C4์—์„œ ์‚ฌ์šฉ๋˜๋Š” ๊ฒƒ)์€ ํ•ฉ๋ฒ• ๋ฐ ์˜๋ฃŒ ์ฝ˜ํ…์ธ ์˜ ๊ณผ๋„ํ•œ ํ•„ํ„ฐ๋ง์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Dodge et al., 2021).
Last, we denote the _type of the fact composition_ of a two-hop prompt as "\(\operatorname{type}(r_{2})\) of \(\operatorname{type}(e_{1})\)'s \(\operatorname{type}(r_{1})\)", where "\(\operatorname{type}(e_{1})\)'s \(\operatorname{type}(r_{1})\)" represents the type of the bridge entity's descriptive mention in the prompt. For example, the fact composition type of \(\tau(\texttt{{mother}}(\texttt{{singer}}(\texttt{{Suspresition}})))\) would be "mother of song's singer".
๋งˆ์ง€๋ง‰์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๋‘ ๋‹จ๊ณ„ ํ”„๋กฌํ”„ํŠธ์˜ ์‚ฌ์‹ค ๊ตฌ์„ฑ _์œ ํ˜•_์„ "\(\operatorname{type}(e_{1})\)'s \(\operatorname{type}(r_{1})\)"์˜ \(\operatorname{type}(r_{2})\)"๋กœ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ "\(\operatorname{type}(e_{1})\)'s \(\operatorname{type}(r_{1})\)"๋Š” ํ”„๋กฌํ”„ํŠธ์—์„œ ๋‹ค๋ฆฌ ์—ญํ• ์„ ํ•˜๋Š” ์—”ํ‹ฐํ‹ฐ์˜ ์„ค๋ช…์  ์–ธ๊ธ‰์˜ ์œ ํ˜•์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(\tau(\texttt{{mother}}(\texttt{{singer}}(\texttt{{Suspresition}})))\)์˜ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์€ "๋…ธ๋ž˜์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ"๊ฐ€ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๋งˆ์ง€๋ง‰์œผ๋กœ, 2ํ™‰ ํ”„๋กฌํ”„ํŠธ์˜ _์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•_์„ "\(\operatorname{type}(r_{2})\) of \(\operatorname{type}(e_{1})\)'s \(\operatorname{type}(r_{1})\)"๋กœ ํ‘œ์‹œํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ "\(\operatorname{type}(e_{1})\)'s \(\operatorname{type}(r_{1})\)"๋Š” ํ”„๋กฌํ”„ํŠธ์—์„œ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์„ค๋ช…์  ์–ธ๊ธ‰ ์œ ํ˜•์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(\tau(\texttt{{mother}}(\texttt{{singer}}(\texttt{{Suspresition})))\)์˜ ์‚ฌ์‹ค ๊ตฌ์„ฑ ์œ ํ˜•์€ "๋…ธ๋ž˜์˜ ๊ฐ€์ˆ˜์˜ ์–ด๋จธ๋‹ˆ"๊ฐ€ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
which focuses on crafting queries that precisely reflect the information needs. This method incorporates the use of generated sentences or fragments thereof as a foundation for search queries. By opting to directly use sentences, obscuring tokens of low confidence, or formulating explicit questions, this approach aims to boost the efficiency of the retrieval process, ensuring that the retrieved information faithfully satisfies the requirements of the generation process.
์ •๋ณด ์š”๊ตฌ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๋Š” ์ฟผ๋ฆฌ๋ฅผ ์ž‘์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์ƒ์„ฑ๋œ ๋ฌธ์žฅ์ด๋‚˜ ๊ทธ ์กฐ๊ฐ๋“ค์„ ๊ฒ€์ƒ‰ ์ฟผ๋ฆฌ์˜ ๊ธฐ์ดˆ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ์žฅ์„ ์ง์ ‘ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜, ์‹ ๋ขฐ๋„๊ฐ€ ๋‚ฎ์€ ํ† ํฐ์„ ์ˆจ๊ธฐ๊ฑฐ๋‚˜, ๋ช…ํ™•ํ•œ ์งˆ๋ฌธ์„ ํ˜•์„ฑํ•จ์œผ๋กœ์จ, ์ด ์ ‘๊ทผ๋ฒ•์€ ๊ฒ€์ƒ‰ ๊ณผ์ •์˜ ํšจ์œจ์„ฑ์„ ๋†’์ด๊ณ , ๊ฒ€์ƒ‰๋œ ์ •๋ณด๊ฐ€ ์ƒ์„ฑ ๊ณผ์ •์˜ ์š”๊ตฌ๋ฅผ ์ถฉ์‹คํžˆ ๋งŒ์กฑ์‹œํ‚ค๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
์ •๋ณด ์š”๊ตฌ ์‚ฌํ•ญ์„ ์ •ํ™•ํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๋Š” ์ฟผ๋ฆฌ ์ž‘์„ฑ์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ์ƒ์„ฑ๋œ ๋ฌธ์žฅ ๋˜๋Š” ๊ทธ ์ผ๋ถ€๋ฅผ ๊ฒ€์ƒ‰ ์ฟผ๋ฆฌ์˜ ๊ธฐ์ดˆ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์„ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ์žฅ์„ ์ง์ ‘ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ ์‹ ๋ขฐ๋„๊ฐ€ ๋‚ฎ์€ ํ† ํฐ์„ ๊ฐ€๋ฆฌ๊ฑฐ๋‚˜ ๋ช…์‹œ์ ์ธ ์งˆ๋ฌธ์„ ๊ณต์‹ํ™”ํ•จ์œผ๋กœ์จ ๊ฒ€์ƒ‰ ํ”„๋กœ์„ธ์Šค์˜ ํšจ์œจ์„ฑ์„ ๋†’์ด๊ณ  ๊ฒ€์ƒ‰๋œ ์ •๋ณด๊ฐ€ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค์˜ ์š”๊ตฌ ์‚ฌํ•ญ์„ ์ถฉ์‹คํžˆ ์ถฉ์กฑํ•˜๋„๋ก ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
_Remark_.: Objective evaluation facilitates the quantitative analysis of capabilities in LLM-based agents through a variety of metrics. While current techniques can not perfectly measure all types of agent capabilities, objective evaluation provides essential insights that complement subjective assessment. Continued advancements in benchmarks and methodologies for objective evaluation will enhance the development and understanding of LLM-based autonomous agents further.
_Remark_.: ๊ฐ๊ด€์ ์ธ ํ‰๊ฐ€๋Š” ๋‹ค์–‘ํ•œ ์ง€ํ‘œ๋ฅผ ํ†ตํ•ด LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ๋Šฅ๋ ฅ์„ ์ •๋Ÿ‰์ ์œผ๋กœ ๋ถ„์„ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ์˜ ๊ธฐ์ˆ ๋กœ๋Š” ๋ชจ๋“  ์œ ํ˜•์˜ ์—์ด์ „ํŠธ ๋Šฅ๋ ฅ์„ ์™„๋ฒฝํ•˜๊ฒŒ ์ธก์ •ํ•  ์ˆ˜๋Š” ์—†์ง€๋งŒ, ๊ฐ๊ด€์ ์ธ ํ‰๊ฐ€๋Š” ์ฃผ๊ด€์ ์ธ ํ‰๊ฐ€๋ฅผ ๋ณด์™„ํ•˜๋Š” ์ค‘์š”ํ•œ ํ†ต์ฐฐ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ฐ๊ด€์ ์ธ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ์™€ ๋ฐฉ๋ฒ•๋ก ์˜ ์ง€์†์ ์ธ ๋ฐœ์ „์€ LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์˜ ๊ฐœ๋ฐœ๊ณผ ์ดํ•ด๋ฅผ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ฌ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
_Remark_.: ๊ฐ๊ด€์ ์ธ ํ‰๊ฐ€๋Š” ๋‹ค์–‘ํ•œ ์ง€ํ‘œ๋ฅผ ํ†ตํ•ด LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ์—ญ๋Ÿ‰์„ ์ •๋Ÿ‰์ ์œผ๋กœ ๋ถ„์„ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ์˜ ๊ธฐ์ˆ ๋กœ๋Š” ๋ชจ๋“  ์œ ํ˜•์˜ ์—์ด์ „ํŠธ ์—ญ๋Ÿ‰์„ ์™„๋ฒฝํ•˜๊ฒŒ ์ธก์ •ํ•  ์ˆ˜๋Š” ์—†์ง€๋งŒ, ๊ฐ๊ด€์ ์ธ ํ‰๊ฐ€๋Š” ์ฃผ๊ด€์ ์ธ ํ‰๊ฐ€๋ฅผ ๋ณด์™„ํ•˜๋Š” ํ•„์ˆ˜์ ์ธ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ฐ๊ด€์ ์ธ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ์™€ ๋ฐฉ๋ฒ•๋ก ์ด ์ง€์†์ ์œผ๋กœ ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์˜ ๊ฐœ๋ฐœ๊ณผ ์ดํ•ด๊ฐ€ ๋”์šฑ ํ–ฅ์ƒ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Language is a prominent ability in human beings to express and communicate, which develops in early childhood and evolves over a lifetime [3, 4]. Machines, however, cannot naturally grasp the abilities of understanding and communicating in the form of human language, unless equipped with powerful artificial intelligence (AI) algorithms. It has been a longstanding research challenge to achieve this goal, to enable machines to read, write, and communicate like humans [5].
์–ธ์–ด๋Š” ์ธ๊ฐ„์ด ํ‘œํ˜„ํ•˜๊ณ  ์†Œํ†ตํ•˜๋Š” ๋ฐ ์žˆ์–ด ์ค‘์š”ํ•œ ๋Šฅ๋ ฅ์œผ๋กœ, ์–ด๋ฆฐ ์‹œ์ ˆ์— ๋ฐœ๋‹ฌํ•˜์—ฌ ํ‰์ƒ ๋™์•ˆ ์ง„ํ™”ํ•ฉ๋‹ˆ๋‹ค [3, 4]. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ๊ณ„๋Š” ๊ฐ•๋ ฅํ•œ ์ธ๊ณต์ง€๋Šฅ(AI) ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฐ–์ถ”์ง€ ์•Š๋Š” ํ•œ, ์ธ๊ฐ„ ์–ธ์–ด์˜ ํ˜•ํƒœ๋กœ ์ดํ•ดํ•˜๊ณ  ์†Œํ†ตํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ์Šต๋“ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๊ธฐ๊ณ„๊ฐ€ ์ธ๊ฐ„์ฒ˜๋Ÿผ ์ฝ๊ณ , ์“ฐ๊ณ , ์†Œํ†ตํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ฒƒ์€ ์˜ค๋žœ ์—ฐ๊ตฌ ๊ณผ์ œ์˜€์Šต๋‹ˆ๋‹ค [5].
์–ธ์–ด๋Š” ์ธ๊ฐ„์—๊ฒŒ ์žˆ์–ด ํ‘œํ˜„๊ณผ ์†Œํ†ต์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๋‘๋“œ๋Ÿฌ์ง„ ๋Šฅ๋ ฅ์œผ๋กœ, ์–ด๋ฆฐ ์‹œ์ ˆ์— ๋ฐœ๋‹ฌํ•˜์—ฌ ํ‰์ƒ ๋™์•ˆ ๋ฐœ์ „ํ•ฉ๋‹ˆ๋‹ค[3, 4]. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ๊ณ„๋Š” ๊ฐ•๋ ฅํ•œ ์ธ๊ณต ์ง€๋Šฅ(AI) ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ํƒ‘์žฌ๋˜์ง€ ์•Š๋Š” ํ•œ ์ธ๊ฐ„์˜ ์–ธ์–ด ํ˜•ํƒœ๋กœ ์ดํ•ดํ•˜๊ณ  ์†Œํ†ตํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ํŒŒ์•…ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๊ธฐ๊ณ„๊ฐ€ ์ธ๊ฐ„์ฒ˜๋Ÿผ ์ฝ๊ณ , ์จ๊ณ , ์†Œํ†ตํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ฒƒ์€ ์˜ค๋žœ ์—ฐ๊ตฌ ๊ณผ์ œ์˜€์Šต๋‹ˆ๋‹ค[5].
**Reference [1]** Emperor Constantine then enacted the first Sunday Laws, for "the venerable Day of the Sun" in 321 A.D. On March 7, 321, the Roman Emperor Constantine issued a decree making Sunday a day of rest from labor, stating: We find a good description of this in sources such as the book Rest Days which states: Early Christian observance of both the spiritual seventh-day sabbath and a Lords Day assembly is evidenced in Ignatius] letter to the Magnesians ca. 110.
**Reference [1]** ํ™ฉ์ œ ์ฝ˜์Šคํƒ„ํ‹ฐ๋ˆ„์Šค๋Š” 321๋…„ "์กด๊ฒฝ๋ฐ›๋Š” ํƒœ์–‘์˜ ๋‚ "์„ ์œ„ํ•ด ์ฒซ ๋ฒˆ์งธ ์ผ์š”์ผ ๋ฒ•์„ ์ œ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. 321๋…„ 3์›” 7์ผ, ๋กœ๋งˆ ํ™ฉ์ œ ์ฝ˜์Šคํƒ„ํ‹ฐ๋ˆ„์Šค๋Š” ์ผ์š”์ผ์„ ๋…ธ๋™์œผ๋กœ๋ถ€ํ„ฐ ์‰ฌ๋Š” ๋‚ ๋กœ ๋งŒ๋“œ๋Š” ๋ฒ•๋ น์„ ๋ฐœํ‘œํ•˜๋ฉฐ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋งํ–ˆ์Šต๋‹ˆ๋‹ค: ์šฐ๋ฆฌ๋Š” ์ด์™€ ๊ด€๋ จ๋œ ์ข‹์€ ์„ค๋ช…์„ ์ฑ… "Rest Days"์™€ ๊ฐ™์€ ์ž๋ฃŒ์—์„œ ์ฐพ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ฑ…์€ ์ดˆ๊ธฐ ๊ธฐ๋…๊ต์ธ๋“ค์ด ์˜์ ์ธ ์ผ๊ณฑ์งธ ๋‚  ์•ˆ์‹์ผ๊ณผ ์ฃผ๋‹˜์˜ ๋‚  ์ง‘ํšŒ๋ฅผ ๋ชจ๋‘ ์ค€์ˆ˜ํ–ˆ์Œ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ด๋Š” ์•ฝ 110๋…„๊ฒฝ ์ด๊ทธ๋‚˜ํ‹ฐ์šฐ์Šค๊ฐ€ ๋งˆ๊ทธ๋„ค์‹œ์•„์ธ๋“ค์—๊ฒŒ ๋ณด๋‚ธ ํŽธ์ง€์—์„œ ์ฆ๋ช…๋ฉ๋‹ˆ๋‹ค.
**์ฐธ๊ณ  [1]** ์ฝ˜์Šคํƒ„ํ‹ฐ๋ˆ„์Šค ํ™ฉ์ œ๋Š” 321๋…„ ์„œ๊ธฐ "์กด๊ฒฝํ•˜๋Š” ํƒœ์–‘์˜ ๋‚ "์„ ์œ„ํ•ด ์ตœ์ดˆ์˜ ์ผ์š”์ผ ๋ฒ•์„ ์ œ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. 321๋…„ 3์›” 7์ผ, ๋กœ๋งˆ ํ™ฉ์ œ ์ฝ˜์Šคํƒ„ํ‹ฐ๋ˆ„์Šค๋Š” ์ผ์š”์ผ์„ ๋…ธ๋™์˜ ํœด์‹์˜ ๋‚ ๋กœ ์ •ํ•˜๋Š” ๋ฒ•๋ น์„ ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค: ์ด์— ๋Œ€ํ•œ ์ข‹์€ ์„ค๋ช…์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ถœ์ฒ˜์—์„œ ์ฐพ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํœด์‹์˜ ๋‚ ์ด๋ผ๋Š” ์ฑ…์—์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค: 110๋…„๊ฒฝ ์ด๊ทธ๋‚˜ํ‹ฐ์šฐ์Šค๊ฐ€ ๋งˆ๊ทธ๋„ค์‹œ์•„์ธ๋“ค์—๊ฒŒ ๋ณด๋‚ธ ํŽธ์ง€์—์„œ ์˜์ ์ธ 7์ผ ์•ˆ์‹์ผ๊ณผ ์ฃผ์˜ ๋‚  ์ง‘ํšŒ๋ฅผ ๋ชจ๋‘ ์ง€ํ‚ค๋Š” ์ดˆ๊ธฐ ๊ธฐ๋…๊ต์ธ๋“ค์˜ ๋ชจ์Šต์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The parameters \(a\) and \(c\) are fitted from the data. If the loss of selected tokens \(\mathcal{L}_{s}\) is used for fitting, then \(a>0\). Conversely, if the loss of unselected tokens \(\mathcal{L}_{us}\) is used for fitting, then \(a<0\). Therefore, we believe that training the model on selected tokens can effectively improve its performance on downstream tasks, while unselected tokens may have a detrimental effect on the model's performance in downstream tasks.
ํŒŒ๋ผ๋ฏธํ„ฐ \(a\)์™€ \(c\)๋Š” ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ๋งž์ถฐ์ง‘๋‹ˆ๋‹ค. ์„ ํƒ๋œ ํ† ํฐ์˜ ์†์‹ค \(\mathcal{L}_{s}\)์ด ๋งž์ถ”๊ธฐ์— ์‚ฌ์šฉ๋˜๋ฉด, \(a>0\)์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ, ์„ ํƒ๋˜์ง€ ์•Š์€ ํ† ํฐ์˜ ์†์‹ค \(\mathcal{L}_{us}\)์ด ๋งž์ถ”๊ธฐ์— ์‚ฌ์šฉ๋˜๋ฉด, \(a<0\)์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ, ์„ ํƒ๋œ ํ† ํฐ์œผ๋กœ ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๊ฒƒ์ด ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฏฟ์œผ๋ฉฐ, ์„ ํƒ๋˜์ง€ ์•Š์€ ํ† ํฐ์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์— ํ•ด๋กœ์šด ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋งค๊ฐœ๋ณ€์ˆ˜ \(a\)์™€ \(c\)๋Š” ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ๋งž์ถคํ™”๋ฉ๋‹ˆ๋‹ค. ์„ ํƒ๋œ ํ† ํฐ์˜ ์†์‹ค \(\mathcal{L}_{s}\)์„ ํ”ผํŒ…์— ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ, \(a>0\)์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ ์„ ํƒ๋˜์ง€ ์•Š์€ ํ† ํฐ์˜ ์†์‹ค \(\mathcal{L}_{us}\)์„ ํ”ผํŒ…์— ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ, \(a<0\)์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์„ ํƒ๋œ ํ† ํฐ์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋ฉด ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ์„ ํƒ๋˜์ง€ ์•Š์€ ํ† ํฐ์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์— ํ•ด๋กœ์šด ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
Search & RankingThe retrieval stage is the combination of search and ranking. It focuses on selecting and prioritizing documents from a dataset to enhance the quality of the generation model's outputs. This stage employs search algorithms to navigate through the indexed data, finding documents that match a user's query. After identifying relevant documents, the process of initially ranking these documents starts to sort them according to their relevance to the query.
๊ฒ€์ƒ‰ ๋ฐ ์ˆœ์œ„ ๋งค๊ธฐ๊ธฐ ๊ฒ€์ƒ‰ ๋‹จ๊ณ„๋Š” ๊ฒ€์ƒ‰๊ณผ ์ˆœ์œ„ ๋งค๊ธฐ๊ธฐ์˜ ์กฐํ•ฉ์ž…๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๋Š” ๋ฐ์ดํ„ฐ์…‹์—์„œ ๋ฌธ์„œ๋ฅผ ์„ ํƒํ•˜๊ณ  ์šฐ์„ ์ˆœ์œ„๋ฅผ ๋งค๊ฒจ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—์„œ๋Š” ๊ฒ€์ƒ‰ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ์ธ๋ฑ์‹ฑ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ํƒ์ƒ‰ํ•˜๊ณ  ์‚ฌ์šฉ์ž์˜ ์ฟผ๋ฆฌ์™€ ์ผ์น˜ํ•˜๋Š” ๋ฌธ์„œ๋ฅผ ์ฐพ์Šต๋‹ˆ๋‹ค. ๊ด€๋ จ ๋ฌธ์„œ๋ฅผ ์‹๋ณ„ํ•œ ํ›„, ์ด๋Ÿฌํ•œ ๋ฌธ์„œ๋ฅผ ์ฟผ๋ฆฌ์™€์˜ ๊ด€๋ จ์„ฑ์— ๋”ฐ๋ผ ์ •๋ ฌํ•˜๋Š” ์ดˆ๊ธฐ ์ˆœ์œ„ ๋งค๊ธฐ๊ธฐ ๊ณผ์ •์ด ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค.
๊ฒ€์ƒ‰ ๋ฐ ์ˆœ์œ„ ๊ฒ€์ƒ‰ ๋ฐ ์ˆœ์œ„ํ™” ๋‹จ๊ณ„๋Š” ๊ฒ€์ƒ‰๊ณผ ์ˆœ์œ„ํ™”์˜ ์กฐํ•ฉ์ž…๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ฌธ์„œ๋ฅผ ์„ ํƒํ•˜๊ณ  ์šฐ์„ ์ˆœ์œ„๋ฅผ ์ง€์ •ํ•˜์—ฌ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—์„œ๋Š” ๊ฒ€์ƒ‰ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ‰์ธํ™”๋œ ๋ฐ์ดํ„ฐ๋ฅผ ํƒ์ƒ‰ํ•˜์—ฌ ์‚ฌ์šฉ์ž์˜ ์ฟผ๋ฆฌ์™€ ์ผ์น˜ํ•˜๋Š” ๋ฌธ์„œ๋ฅผ ์ฐพ์Šต๋‹ˆ๋‹ค. ๊ด€๋ จ ๋ฌธ์„œ๋ฅผ ์‹๋ณ„ํ•œ ํ›„์—๋Š” ์ฟผ๋ฆฌ์™€์˜ ๊ด€๋ จ์„ฑ์— ๋”ฐ๋ผ ๋ฌธ์„œ์˜ ์ˆœ์œ„๋ฅผ ๋งค๊ธฐ๋Š” ์ž‘์—…์„ ์‹œ์ž‘ํ•˜์—ฌ ๋ฌธ์„œ์˜ ์ˆœ์œ„๋ฅผ ๋งค๊ธฐ๋Š” ์ž‘์—…์ด ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค.
We evaluate LLMs on the closed-book QA task, focusing on the PopQA dataset Mallen et al. (2023) characterized by questions that covers a long-tail entity distribution. It also includes unique Wikipedia titles, facilitating the calculation of pageviews and identification of relevant Wikipedia pages. We acquire the relevant Wikipedia page for each subject entity in the PopQA dataset and divide the entities into five buckets based on their popularity levels (Figure 2).
์šฐ๋ฆฌ๋Š” PopQA ๋ฐ์ดํ„ฐ์…‹ Mallen et al. (2023)์„ ์ค‘์‹ฌ์œผ๋กœ, ๋กฑํ…Œ์ผ ์—”ํ‹ฐํ‹ฐ ๋ถ„ํฌ๋ฅผ ๋‹ค๋ฃจ๋Š” ์งˆ๋ฌธ๋“ค๋กœ ํŠน์ง•์ง€์–ด์ง„ ํ์‡„ํ˜• QA ์ž‘์—…์—์„œ LLM์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ์…‹์—๋Š” ๊ณ ์œ ํ•œ ์œ„ํ‚คํ”ผ๋””์•„ ์ œ๋ชฉ๋„ ํฌํ•จ๋˜์–ด ์žˆ์–ด ํŽ˜์ด์ง€๋ทฐ ๊ณ„์‚ฐ๊ณผ ๊ด€๋ จ ์œ„ํ‚คํ”ผ๋””์•„ ํŽ˜์ด์ง€ ์‹๋ณ„์„ ์šฉ์ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” PopQA ๋ฐ์ดํ„ฐ์…‹์˜ ๊ฐ ์ฃผ์ œ ์—”ํ‹ฐํ‹ฐ์— ๋Œ€ํ•œ ๊ด€๋ จ ์œ„ํ‚คํ”ผ๋””์•„ ํŽ˜์ด์ง€๋ฅผ ํš๋“ํ•˜๊ณ , ์—”ํ‹ฐํ‹ฐ๋ฅผ ์ธ๊ธฐ๋„ ์ˆ˜์ค€์— ๋”ฐ๋ผ ๋‹ค์„ฏ ๊ฐœ์˜ ๋ฒ„ํ‚ท์œผ๋กœ ๋‚˜๋ˆ•๋‹ˆ๋‹ค (Figure 2).
์šฐ๋ฆฌ๋Š” ๊ธด ๊ผฌ๋ฆฌ ์—”ํ‹ฐํ‹ฐ ๋ถ„ํฌ๋ฅผ ํฌํ•จํ•˜๋Š” ์งˆ๋ฌธ์œผ๋กœ ํŠน์ง•์ง€์–ด์ง€๋Š” PopQA ๋ฐ์ดํ„ฐ ์„ธํŠธ Mallen et al. (2023)์„ ์ค‘์‹ฌ์œผ๋กœ ํ์‡„ํ˜• ์„œ์ฑ… QA ์ž‘์—…์—์„œ LLM์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ณ ์œ ํ•œ Wikipedia ์ œ๋ชฉ์ด ํฌํ•จ๋˜์–ด ์žˆ์–ด ํŽ˜์ด์ง€ ์กฐํšŒ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•˜๊ณ  ๊ด€๋ จ Wikipedia ํŽ˜์ด์ง€๋ฅผ ์‹๋ณ„ํ•˜๋Š” ๊ฒƒ์ด ์šฉ์ดํ•ฉ๋‹ˆ๋‹ค. PopQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ์ฃผ์ œ ์—”ํ‹ฐํ‹ฐ์— ๋Œ€ํ•ด ๊ด€๋ จ Wikipedia ํŽ˜์ด์ง€๋ฅผ ํš๋“ํ•˜๊ณ  ์ธ๊ธฐ๋„ ์ˆ˜์ค€์— ๋”ฐ๋ผ ์—”ํ‹ฐํ‹ฐ๋ฅผ 5๊ฐœ์˜ ๋ฒ„ํ‚ท์œผ๋กœ ๋‚˜๋ˆ•๋‹ˆ๋‹ค(๊ทธ๋ฆผ 2).
Cluster-based FiltersWe further use unsupervised semantic clustering to group web documents. This clustering process enables efficient identification and analysis of documents sharing similar semantic features. The clustered data are subsequently annotated with quality labels, providing essential references for the optimization of Yi's data mixture strategy. Documents identified as low-quality through automatic and manual verification are excluded from the dataset.
ํด๋Ÿฌ์Šคํ„ฐ ๊ธฐ๋ฐ˜ ํ•„ํ„ฐ ์šฐ๋ฆฌ๋Š” ์›น ๋ฌธ์„œ๋ฅผ ๊ทธ๋ฃนํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋น„์ง€๋„ ์˜๋ฏธ ํด๋Ÿฌ์Šคํ„ฐ๋ง์„ ์ถ”๊ฐ€๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ํด๋Ÿฌ์Šคํ„ฐ๋ง ๊ณผ์ •์€ ์œ ์‚ฌํ•œ ์˜๋ฏธ์  ํŠน์ง•์„ ๊ณต์œ ํ•˜๋Š” ๋ฌธ์„œ๋“ค์„ ํšจ์œจ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ๋ถ„์„ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ํด๋Ÿฌ์Šคํ„ฐ๋ง๋œ ๋ฐ์ดํ„ฐ๋Š” ์ดํ›„ ํ’ˆ์งˆ ๋ผ๋ฒจ๋กœ ์ฃผ์„์ด ๋‹ฌ๋ฆฌ๋ฉฐ, ์ด๋Š” Yi์˜ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ์ „๋žต ์ตœ์ ํ™”๋ฅผ ์œ„ํ•œ ํ•„์ˆ˜์ ์ธ ์ฐธ๊ณ  ์ž๋ฃŒ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ž๋™ ๋ฐ ์ˆ˜๋™ ๊ฒ€์ฆ์„ ํ†ตํ•ด ์ €ํ’ˆ์งˆ๋กœ ์‹๋ณ„๋œ ๋ฌธ์„œ๋“ค์€ ๋ฐ์ดํ„ฐ์…‹์—์„œ ์ œ์™ธ๋ฉ๋‹ˆ๋‹ค.
ํด๋Ÿฌ์Šคํ„ฐ ๊ธฐ๋ฐ˜ ํ•„ํ„ฐ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋น„์ง€๋„ ์‹œ๋งจํ‹ฑ ํด๋Ÿฌ์Šคํ„ฐ๋ง์„ ์‚ฌ์šฉํ•˜์—ฌ ์›น ๋ฌธ์„œ๋ฅผ ๊ทธ๋ฃนํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด ํด๋Ÿฌ์Šคํ„ฐ๋ง ํ”„๋กœ์„ธ์Šค๋ฅผ ํ†ตํ•ด ์œ ์‚ฌํ•œ ์˜๋ฏธ์  ํŠน์ง•์„ ๊ณต์œ ํ•˜๋Š” ๋ฌธ์„œ๋ฅผ ํšจ์œจ์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ๋ถ„์„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ํด๋Ÿฌ์Šคํ„ฐ๋ง๋œ ๋ฐ์ดํ„ฐ์— ํ’ˆ์งˆ ๋ ˆ์ด๋ธ”์„ ์ถ”๊ฐ€ํ•˜์—ฌ Yi์˜ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ์ „๋žต ์ตœ์ ํ™”๋ฅผ ์œ„ํ•œ ํ•„์ˆ˜ ์ฐธ์กฐ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ž๋™ ๋ฐ ์ˆ˜๋™ ๊ฒ€์ฆ์„ ํ†ตํ•ด ์ €ํ’ˆ์งˆ๋กœ ์‹๋ณ„๋œ ๋ฌธ์„œ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ œ์™ธ๋ฉ๋‹ˆ๋‹ค.
Finding all repeated sequences is thus a matter of linearly scanning the suffix array from beginning to end and looking for sequences \(\mathcal{A}_{i},\mathcal{A}_{i+1}\) that share a common prefix of at least some threshold length. Any satisfying sequences are recorded. This algorithm is embarrassingly parallel, and so we can efficiently process the dataset. Based on experimentation (Appendix B), we choose a threshold length of 50 BPE tokens for all experiments.
๋ชจ๋“  ๋ฐ˜๋ณต๋˜๋Š” ์‹œํ€€์Šค๋ฅผ ์ฐพ๋Š” ๊ฒƒ์€ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ๋๊นŒ์ง€ ์„ ํ˜•์ ์œผ๋กœ ์Šค์บ”ํ•˜๋ฉด์„œ ์ผ์ • ๊ธธ์ด ์ด์ƒ์˜ ๊ณตํ†ต ์ ‘๋‘์‚ฌ๋ฅผ ๊ณต์œ ํ•˜๋Š” ์‹œํ€€์Šค \(\mathcal{A}_{i},\mathcal{A}_{i+1}\)๋ฅผ ์ฐพ๋Š” ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. ๋งŒ์กฑํ•˜๋Š” ์‹œํ€€์Šค๋Š” ๊ธฐ๋ก๋ฉ๋‹ˆ๋‹ค. ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๊ฐ€ ๋งค์šฐ ์šฉ์ดํ•˜๋ฏ€๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ํšจ์œจ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ(Appendix B)์— ๋”ฐ๋ผ, ๋ชจ๋“  ์‹คํ—˜์—์„œ 50 BPE ํ† ํฐ์˜ ์ž„๊ณ„ ๊ธธ์ด๋ฅผ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค.
๋”ฐ๋ผ์„œ ๋ชจ๋“  ๋ฐ˜๋ณต๋˜๋Š” ์‹œํ€€์Šค๋ฅผ ์ฐพ๋Š” ๊ฒƒ์€ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ๋๊นŒ์ง€ ์„ ํ˜•์ ์œผ๋กœ ์Šค์บ”ํ•˜๊ณ  ์ตœ์†Œ ์ž„๊ณ„ ๊ธธ์ด์˜ ๊ณตํ†ต ์ ‘๋‘์‚ฌ๋ฅผ ๊ณต์œ ํ•˜๋Š” ์‹œํ€€์Šค \(\mathcal{A}_{i},\mathcal{A}_{i+1}\)๋ฅผ ์ฐพ๋Š” ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. ๋งŒ์กฑํ•˜๋Š” ์‹œํ€€์Šค๋Š” ๋ชจ๋‘ ๊ธฐ๋ก๋ฉ๋‹ˆ๋‹ค. ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๋ถ€๋„๋Ÿฌ์šธ ์ •๋„๋กœ ๋ณ‘๋ ฌ์ ์ด๋ฏ€๋กœ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ํšจ์œจ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜(๋ถ€๋ก B)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ชจ๋“  ์‹คํ—˜์— ๋Œ€ํ•ด 50๊ฐœ์˜ BPE ํ† ํฐ์˜ ์ž„๊ณ„ ๊ธธ์ด๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.
**Output** they also are easily trained, obedient, and intelligent.[[**Retriee**]=Yes]<p>German Shepherd โ€“ a simple difference of opinion, it is the fundamental issue of the breeds essential conformation and movement.... They are known for being easy to train and good for performing tasks and following instructions. They are especially well known for their police work, being used for tracking criminals, patrolling troubled areas and detecting and holding suspects.
**Output** ๊ทธ๋“ค์€ ๋˜ํ•œ ์‰ฝ๊ฒŒ ํ›ˆ๋ จ๋˜๊ณ , ์ˆœ์ข…์ ์ด๋ฉฐ, ์ง€๋Šฅ์ ์ž…๋‹ˆ๋‹ค.[[**Retriee**]=Yes]<p>์ €๋จผ ์…ฐํผ๋“œ โ€“ ๋‹จ์ˆœํ•œ ์˜๊ฒฌ ์ฐจ์ด๋กœ, ์ด๋Š” ํ’ˆ์ข…์˜ ๋ณธ์งˆ์ ์ธ ๊ตฌ์กฐ์™€ ์›€์ง์ž„์˜ ๊ทผ๋ณธ์ ์ธ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค.... ๊ทธ๋“ค์€ ํ›ˆ๋ จ์ด ์‰ฝ๊ณ , ์ž‘์—… ์ˆ˜ํ–‰๊ณผ ์ง€์‹œ ๋”ฐ๋ฅด๊ธฐ์— ์ข‹๋‹ค๊ณ  ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ๊ฒฝ์ฐฐ ์—…๋ฌด์—์„œ ์œ ๋ช…ํ•˜๋ฉฐ, ๋ฒ”์ฃ„์ž ์ถ”์ , ๋ฌธ์ œ ์ง€์—ญ ์ˆœ์ฐฐ, ์šฉ์˜์ž ํƒ์ง€ ๋ฐ ์ฒดํฌ์— ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
**์ถœ๋ ฅ** ๋˜ํ•œ ์‰ฝ๊ฒŒ ํ›ˆ๋ จ๋˜๊ณ  ์ˆœ์ข…์ ์ด๋ฉฐ ์ง€๋Šฅ์ ์ž…๋‹ˆ๋‹ค[[**Retriee**]=Yes]<p>๋…์ผ ์…ฐํผ๋“œ - ๊ฐ„๋‹จํ•œ ์˜๊ฒฌ ์ฐจ์ด, ํ’ˆ์ข…์˜ ๋ณธ์งˆ์ ์ธ ์ฒดํ˜•๊ณผ ์›€์ง์ž„์˜ ๊ทผ๋ณธ์ ์ธ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค.... ๊ทธ๋“ค์€ ํ›ˆ๋ จํ•˜๊ธฐ ์‰ฝ๊ณ  ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ์ง€์‹œ๋ฅผ ๋”ฐ๋ฅด๋Š” ๋ฐ ์ข‹๊ธฐ๋กœ ์œ ๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ํŠนํžˆ ๋ฒ”์ฃ„์ž ์ถ”์ , ๋ฌธ์ œ ์ง€์—ญ ์ˆœ์ฐฐ, ์šฉ์˜์ž ํƒ์ง€ ๋ฐ ๊ตฌ๊ธˆ ๋“ฑ ๊ฒฝ์ฐฐ ์—…๋ฌด๋กœ ์ž˜ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค.
To examine the effectiveness and superiority of LLMs, a surge of tasks and benchmarks have been proposed for conducting empirical ability evaluation and analysis. In this section, we first introduce three types of basic ability evaluation of LLMs for language generation and understanding, then present several advanced ability evaluations with more complicated settings or goals, and finally discuss existing benchmarks, evaluation approaches, and empirical analysis.
LLMs์˜ ํšจ๊ณผ์„ฑ๊ณผ ์šฐ์ˆ˜์„ฑ์„ ๊ฒ€ํ† ํ•˜๊ธฐ ์œ„ํ•ด, ๊ฒฝํ—˜์  ๋Šฅ๋ ฅ ํ‰๊ฐ€์™€ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ์ž‘์—…๊ณผ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ๋จผ์ € ์–ธ์–ด ์ƒ์„ฑ ๋ฐ ์ดํ•ด๋ฅผ ์œ„ํ•œ LLM์˜ ์„ธ ๊ฐ€์ง€ ๊ธฐ๋ณธ ๋Šฅ๋ ฅ ํ‰๊ฐ€๋ฅผ ์†Œ๊ฐœํ•˜๊ณ , ๋” ๋ณต์žกํ•œ ์„ค์ •์ด๋‚˜ ๋ชฉํ‘œ๋ฅผ ๊ฐ€์ง„ ์—ฌ๋Ÿฌ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ ํ‰๊ฐ€๋ฅผ ์ œ์‹œํ•œ ํ›„, ๊ธฐ์กด์˜ ๋ฒค์น˜๋งˆํฌ, ํ‰๊ฐ€ ์ ‘๊ทผ๋ฒ• ๋ฐ ๊ฒฝํ—˜์  ๋ถ„์„์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค.
LLM์˜ ํšจ๊ณผ์™€ ์šฐ์›”์„ฑ์„ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด, ๊ฒฝํ—˜์  ๋Šฅ๋ ฅ ํ‰๊ฐ€ ๋ฐ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•œ ์ž‘์—…๊ณผ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ๊ธ‰์ฆํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ๋จผ์ € ์–ธ์–ด ์ƒ์„ฑ ๋ฐ ์ดํ•ด๋ฅผ ์œ„ํ•œ LLM์˜ ์„ธ ๊ฐ€์ง€ ๊ธฐ๋ณธ ๋Šฅ๋ ฅ ํ‰๊ฐ€ ์œ ํ˜•์„ ์†Œ๊ฐœํ•œ ๋‹ค์Œ, ๋” ๋ณต์žกํ•œ ์„ค์ •์ด๋‚˜ ๋ชฉํ‘œ๋ฅผ ๊ฐ€์ง„ ๋ช‡ ๊ฐ€์ง€ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ ํ‰๊ฐ€๋ฅผ ์ œ์‹œํ•˜๊ณ , ๋งˆ์ง€๋ง‰์œผ๋กœ ๊ธฐ์กด ๋ฒค์น˜๋งˆํฌ, ํ‰๊ฐ€ ์ ‘๊ทผ ๋ฐฉ์‹ ๋ฐ ๊ฒฝํ—˜์  ๋ถ„์„์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค.
In contrast, for multilingual tokenizers, we observe significant fertility improvements with increasing vocabulary sizes. Due to the larger thesaurus induced by the additional languages, the tokenizer requires a larger vocabulary to allow a model to perform convincingly on all languages. Therefore, only within the non-convergence vocabulary range, we achieve a strong, negative correlation between fertility and downstream performance with varying vocabulary sizes.
๋ฐ˜๋ฉด, ๋‹ค๊ตญ์–ด ํ† ํฌ๋‚˜์ด์ €์˜ ๊ฒฝ์šฐ, ์–ดํœ˜ ํฌ๊ธฐ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์œ ์˜๋ฏธํ•œ fertility ๊ฐœ์„ ์ด ๊ด€์ฐฐ๋ฉ๋‹ˆ๋‹ค. ์ถ”๊ฐ€๋œ ์–ธ์–ด๋“ค๋กœ ์ธํ•ด ์œ ๋„๋œ ๋” ํฐ ์‚ฌ์ „ ๋•Œ๋ฌธ์—, ํ† ํฌ๋‚˜์ด์ €๋Š” ๋ชจ๋“  ์–ธ์–ด์—์„œ ๋ชจ๋ธ์ด ์„ค๋“๋ ฅ ์žˆ๊ฒŒ ์ž‘๋™ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋” ํฐ ์–ดํœ˜๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ, ๋น„์ˆ˜๋ ด ์–ดํœ˜ ๋ฒ”์œ„ ๋‚ด์—์„œ๋งŒ ์–ดํœ˜ ํฌ๊ธฐ์— ๋”ฐ๋ผ fertility์™€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ ๊ฐ„์— ๊ฐ•ํ•œ ์Œ์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ ๋‹ค๊ตญ์–ด ํ† ํฐํ™”๊ธฐ์— ์žˆ์–ด์„œ๋Š” ์–ดํœ˜ ํฌ๊ธฐ๊ฐ€ ์ปค์งˆ์ˆ˜๋ก ์ƒ๋‹นํ•œ ๋น„์˜ฅ๋„ ํ–ฅ์ƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ๋‹ค. ์ถ”๊ฐ€ ์–ธ์–ด์— ์˜ํ•ด ์œ ๋„๋œ ๋” ํฐ ๋™์˜์–ด ์‚ฌ์ „์œผ๋กœ ์ธํ•ด ํ† ํฐํ™”๊ธฐ๋Š” ๋ชจ๋ธ์ด ๋ชจ๋“  ์–ธ์–ด์—์„œ ์„ค๋“๋ ฅ ์žˆ๊ฒŒ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋” ํฐ ์–ดํœ˜๋ฅผ ํ•„์š”๋กœ ํ•œ๋‹ค. ๋”ฐ๋ผ์„œ ์ˆ˜๋ ดํ•˜์ง€ ์•Š๋Š” ์–ดํœ˜ ๋ฒ”์œ„ ๋‚ด์—์„œ๋งŒ ๋‹ค์–‘ํ•œ ์–ดํœ˜ ํฌ๊ธฐ๋กœ ๋น„์˜ฅ๋„์™€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ ์‚ฌ์ด์˜ ๊ฐ•ํ•œ ์Œ์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค.
**Safety Prompts Evaluation**Sun et al. (2023) used InstructGPT (Ouyang et al., 2022) with a verbalizer to assess text safety. In our method, we simplify things. Instead of using a verbalizer to understand the output, we use the function call feature of the OpenAI GPT 3.5 8. This makes the results clearer and easier to interpret. For those who are interested, we have detailed the specifics of this function call and the related assessment prompts in Appendix A.2.
**Safety Prompts Evaluation** Sun et al. (2023)์€ ํ…์ŠคํŠธ ์•ˆ์ „์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด InstructGPT (Ouyang et al., 2022)๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์—์„œ๋Š” ์ด๋ฅผ ๋‹จ์ˆœํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ถœ๋ ฅ ๊ฒฐ๊ณผ๋ฅผ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด verbalizer๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹ , OpenAI GPT 3.5์˜ ํ•จ์ˆ˜ ํ˜ธ์ถœ ๊ธฐ๋Šฅ์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ฒฐ๊ณผ๋ฅผ ๋” ๋ช…ํ™•ํ•˜๊ณ  ํ•ด์„ํ•˜๊ธฐ ์‰ฝ๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ๊ด€์‹ฌ ์žˆ๋Š” ๋ถ„๋“ค์„ ์œ„ํ•ด, ์ด ํ•จ์ˆ˜ ํ˜ธ์ถœ๊ณผ ๊ด€๋ จ๋œ ํ‰๊ฐ€ ํ”„๋กฌํ”„ํŠธ์˜ ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ๋ถ€๋ก A.2์— ์ž์„ธํžˆ ์„ค๋ช…ํ•ด ๋‘์—ˆ์Šต๋‹ˆ๋‹ค.
**์•ˆ์ „ ํ”„๋กฌํ”„ํŠธ ํ‰๊ฐ€**Sun et al.(2023)์€ ํ…์ŠคํŠธ ์•ˆ์ „์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ตฌ๋‘ํ™” ๊ธฐ๋Šฅ์ด ์žˆ๋Š” InstructGPT(Ouyang et al., 2022)๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ ๋ฐฉ๋ฒ•์€ ์ด๋ฅผ ๋‹จ์ˆœํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ถœ๋ ฅ์„ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด ๊ตฌ๋‘ํ™” ๊ธฐ๋Šฅ์„ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  OpenAI GPT 3.5 8์˜ ํ•จ์ˆ˜ ํ˜ธ์ถœ ๊ธฐ๋Šฅ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ๊ฒฐ๊ณผ๊ฐ€ ๋” ๋ช…ํ™•ํ•˜๊ณ  ํ•ด์„ํ•˜๊ธฐ ์‰ฌ์›Œ์ง‘๋‹ˆ๋‹ค. ๊ด€์‹ฌ์ด ์žˆ๋Š” ๋ถ„๋“ค์„ ์œ„ํ•ด ์ด ํ•จ์ˆ˜ ํ˜ธ์ถœ์˜ ์„ธ๋ถ€ ์‚ฌํ•ญ๊ณผ ๊ด€๋ จ ํ‰๊ฐ€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ถ€๋ก A.2์— ์ž์„ธํžˆ ์„ค๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค.
We can observe the following: 1) ETS-DACP trained on 10% outperforms DACP with 100% of the data; 2) ETS-DACP has the best performance among all three counterparts and is on par with a combination of three metrics - ETS-DACP-com; 3) ETA-DACP-ent trained on 10% corpus is a close second despite not having any access to task data, handily surpassing DACP trained on 100% of the data; and 4) Efficient DACP methods with hard sampling outperform ones with soft sampling.
๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: 1) 10%์˜ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ ETS-DACP๊ฐ€ 100%์˜ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ DACP๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค; 2) ETS-DACP๋Š” ์„ธ ๊ฐ€์ง€ ๋ชจ๋“  ๋Œ€์‘ ๋ชจ๋ธ ์ค‘์—์„œ ๊ฐ€์žฅ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋ฉฐ, ์„ธ ๊ฐ€์ง€ ๋ฉ”ํŠธ๋ฆญ์˜ ์กฐํ•ฉ์ธ ETS-DACP-com๊ณผ ๋™๋“ฑํ•œ ์ˆ˜์ค€์ž…๋‹ˆ๋‹ค; 3) 10%์˜ ์ฝ”ํผ์Šค๋กœ ํ›ˆ๋ จ๋œ ETA-DACP-ent๋Š” ์ž‘์—… ๋ฐ์ดํ„ฐ์— ์ ‘๊ทผํ•˜์ง€ ๋ชปํ–ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  100%์˜ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ DACP๋ฅผ ์‰ฝ๊ฒŒ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ ๊ทผ์†Œํ•œ ์ฐจ์ด๋กœ 2์œ„๋ฅผ ์ฐจ์ง€ํ•ฉ๋‹ˆ๋‹ค; ๊ทธ๋ฆฌ๊ณ  4) ํ•˜๋“œ ์ƒ˜ํ”Œ๋ง์„ ์‚ฌ์šฉํ•˜๋Š” ํšจ์œจ์ ์ธ DACP ๋ฐฉ๋ฒ•์ด ์†Œํ”„ํŠธ ์ƒ˜ํ”Œ๋ง์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•๋ณด๋‹ค ๋” ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค.
๋‹ค์Œ๊ณผ ๊ฐ™์€ ์‚ฌ์‹ค์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: 1) 10%๋กœ ํ•™์Šต๋œ ETS-DACP๊ฐ€ 100%์˜ ๋ฐ์ดํ„ฐ๋กœ DACP๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค; 2) ETS-DACP๊ฐ€ ์„ธ ๊ฐ€์ง€ ๋Œ€์‘ ๋ฐฉ์‹ ์ค‘ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋ฉฐ ์„ธ ๊ฐ€์ง€ ์ง€ํ‘œ์˜ ์กฐํ•ฉ์ธ ETS-DACP-com๊ณผ ๋™๋“ฑํ•ฉ๋‹ˆ๋‹ค; 3) 10% ๋ง๋ญ‰์น˜๋กœ ํ•™์Šต๋œ ETA-DACP-ent๋Š” ๊ณผ์ œ ๋ฐ์ดํ„ฐ์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์—†์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  100% ๋ง๋ญ‰์น˜๋กœ ํ•™์Šต๋œ DACP๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š” 2์œ„๋ฅผ ์ฐจ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค; 4) ํ•˜๋“œ ์ƒ˜ํ”Œ๋ง์„ ์‚ฌ์šฉํ•œ ํšจ์œจ์ ์ธ DACP ๋ฐฉ๋ฒ•์ด ์†Œํ”„ํŠธ ์ƒ˜ํ”Œ๋ง์„ ์‚ฌ์šฉํ•œ ๋ฐฉ๋ฒ•๋ณด๋‹ค ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค.
where the exponent \(c_{q}\in(0,\infty)\) depends on \(\beta\) and the algorithm \(q\). For example, when \(q\) encodes the counting measure \(q_{T}(i):=n_{T}(i)/T\) (reminiscent of SGD), it was shown that \(c_{q}=(1-1/\beta)/2\in(0,1/2)\). Another algorithm \(q_{T}(i):=1[n_{T}(i)\geq 1]/\sum_{\ell}1[n_{T}(\ell)\geq 1]\) (reminiscent of ADAM ) was proposed which attains a optimal error rate (over all algorithms based on random embeddings) with \(c_{q}=\beta-1\).
์ง€์ˆ˜ \(c_{q}\in(0,\infty)\)๋Š” \(\beta\)์™€ ์•Œ๊ณ ๋ฆฌ์ฆ˜ \(q\)์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(q\)๊ฐ€ ์นด์šดํŒ… ์ธก์ • \(q_{T}(i):=n_{T}(i)/T\) (SGD๋ฅผ ์—ฐ์ƒ์‹œํ‚ค๋Š”)๋กœ ์ธ์ฝ”๋”ฉ๋  ๋•Œ, \(c_{q}=(1-1/\beta)/2\in(0,1/2)\)์ž„์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์•Œ๊ณ ๋ฆฌ์ฆ˜ \(q_{T}(i):=1[n_{T}(i)\geq 1]/\sum_{\ell}1[n_{T}(\ell)\geq 1]\) (ADAM์„ ์—ฐ์ƒ์‹œํ‚ค๋Š”)์ด ์ œ์•ˆ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” \(c_{q}=\beta-1\)๋กœ ๋žœ๋ค ์ž„๋ฒ ๋”ฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๋ชจ๋“  ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์ค‘ ์ตœ์ ์˜ ์˜ค๋ฅ˜์œจ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ ์ง€์ˆ˜ \(c_{q}\in(0,\infty)\)๋Š” \(\beta\)์™€ ์•Œ๊ณ ๋ฆฌ์ฆ˜ \(q\)์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(q\)๊ฐ€ ์นด์šดํŒ… ์ธก์ •๊ฐ’ \(q_{T}(i):=n_{T}(i)/T\)๋ฅผ ์ธ์ฝ”๋”ฉํ•˜๋Š” ๊ฒฝ์šฐ(SGD๋ฅผ ์—ฐ์ƒ์‹œํ‚ด), \(c_{q}=(1-1/\beta)/2\in(0,1/2)\)์ž„์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์•Œ๊ณ ๋ฆฌ์ฆ˜ \(q_{T}(i):=1[n_{T}(i)\geq 1]/\sum_{\ell}1[n_{T}(\ell)\geq 1]\) (ADAM ์„ ์—ฐ์ƒ์‹œํ‚ด)์ด ์ œ์•ˆ๋˜์—ˆ๋Š”๋ฐ, ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๋ชจ๋“  ๋žœ๋ค ์ž„๋ฒ ๋”ฉ ๊ธฐ๋ฐ˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ ์ตœ์ ์˜ ์˜ค๋ฅ˜์œจ์„ ๋‹ฌ์„ฑํ•˜๋ฉฐ \(c_{q}=\beta-1\)์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
Tab. 5 presents the performance results of Yi-34B-Chat in the three third-party evaluations we consider, with the cutoff date for the results being December 21, 2023. The data demonstrates that, although there is still a gap compared to GPT-4, our model exhibits proficient bilingual (Chinese and English) dialogue capabilities and aligns well with user preferences. Additional comparative results of various models are accessible for review on the official website.
ํ‘œ 5๋Š” ์šฐ๋ฆฌ๊ฐ€ ๊ณ ๋ คํ•œ ์„ธ ๊ฐ€์ง€ ์ œ3์ž ํ‰๊ฐ€์—์„œ Yi-34B-Chat์˜ ์„ฑ๋Šฅ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ๋ฉฐ, ๊ฒฐ๊ณผ์˜ ๊ธฐ์ค€ ๋‚ ์งœ๋Š” 2023๋…„ 12์›” 21์ผ์ž…๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๋Š” GPT-4์™€ ๋น„๊ตํ–ˆ์„ ๋•Œ ์—ฌ์ „ํžˆ ๊ฒฉ์ฐจ๊ฐ€ ์žˆ์ง€๋งŒ, ์šฐ๋ฆฌ ๋ชจ๋ธ์ด ๋›ฐ์–ด๋‚œ ์ด์ค‘ ์–ธ์–ด(์ค‘๊ตญ์–ด์™€ ์˜์–ด) ๋Œ€ํ™” ๋Šฅ๋ ฅ์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ ์‚ฌ์šฉ์ž ์„ ํ˜ธ๋„์™€ ์ž˜ ๋งž๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ชจ๋ธ์˜ ์ถ”๊ฐ€ ๋น„๊ต ๊ฒฐ๊ณผ๋Š” ๊ณต์‹ ์›น์‚ฌ์ดํŠธ์—์„œ ๊ฒ€ํ† ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ‘œ 5๋Š” 2023๋…„ 12์›” 21์ผ์„ ๊ธฐ์ค€์œผ๋กœ ํ•œ 3์ฐจ ํ‰๊ฐ€์—์„œ Yi-34B-Chat์˜ ์„ฑ๋Šฅ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ๋Š” GPT-4์— ๋น„ํ•ด ์—ฌ์ „ํžˆ ๊ฒฉ์ฐจ๊ฐ€ ์žˆ์ง€๋งŒ, ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ๋Šฅ์ˆ™ํ•œ ์ด์ค‘ ์–ธ์–ด(์ค‘๊ตญ์–ด ๋ฐ ์˜์–ด) ๋Œ€ํ™” ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•˜๋ฉฐ ์‚ฌ์šฉ์ž ์„ ํ˜ธ๋„์— ์ž˜ ๋ถ€ํ•ฉํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ชจ๋ธ์˜ ์ถ”๊ฐ€ ๋น„๊ต ๊ฒฐ๊ณผ๋Š” ๊ณต์‹ ์›น์‚ฌ์ดํŠธ์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Therefore, comparing \(\mathsf{bioS}\) and \(\mathsf{bioS}^{\mathsf{simple}}\), it's more advantageous to rewrite the data 1000 times (in this ideal setting), training each for one pass (as done in the \(\mathsf{bioS}\) data), rather than training the same data for 1000 passes (as done in the \(\mathsf{bioS}^{\mathsf{simple}}\) data). This is because, without data diversity, the model wastes capacity memorizing sentence structures, resulting in a capacity loss.
๋”ฐ๋ผ์„œ \(\mathsf{bioS}\)์™€ \(\mathsf{bioS}^{\mathsf{simple}}\)๋ฅผ ๋น„๊ตํ•  ๋•Œ, ๋™์ผํ•œ ๋ฐ์ดํ„ฐ๋ฅผ 1000๋ฒˆ ๋ฐ˜๋ณตํ•ด์„œ ํ•™์Šตํ•˜๋Š” ๊ฒƒ(\(\mathsf{bioS}^{\mathsf{simple}}\) ๋ฐ์ดํ„ฐ์—์„œ ์ˆ˜ํ–‰๋œ ๊ฒƒ์ฒ˜๋Ÿผ)๋ณด๋‹ค ๋ฐ์ดํ„ฐ๋ฅผ 1000๋ฒˆ ๋‹ค์‹œ ์ž‘์„ฑํ•˜๊ณ  ๊ฐ ๋ฐ์ดํ„ฐ๋ฅผ ํ•œ ๋ฒˆ์”ฉ ํ•™์Šตํ•˜๋Š” ๊ฒƒ(\(\mathsf{bioS}\) ๋ฐ์ดํ„ฐ์—์„œ ์ˆ˜ํ–‰๋œ ๊ฒƒ์ฒ˜๋Ÿผ)์ด ๋” ์œ ๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋ฐ์ดํ„ฐ ๋‹ค์–‘์„ฑ์ด ์—†์œผ๋ฉด ๋ชจ๋ธ์ด ๋ฌธ์žฅ ๊ตฌ์กฐ๋ฅผ ์•”๊ธฐํ•˜๋Š” ๋ฐ ์šฉ๋Ÿ‰์„ ๋‚ญ๋น„ํ•˜๊ฒŒ ๋˜์–ด ์šฉ๋Ÿ‰ ์†์‹ค์ด ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
๋”ฐ๋ผ์„œ \(\mathsf{bioS}\)์™€ \(\mathsf{bioS}^{\mathsf{simple}}\)๋ฅผ ๋น„๊ตํ•˜๋ฉด, ์ด ๊ฒฝ์šฐ(์ด ์ด์ƒ์ ์ธ ์„ค์ •์—์„œ) 1000๋ฒˆ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค 1000๋ฒˆ์”ฉ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ด ๋” ์œ ๋ฆฌํ•ฉ๋‹ˆ๋‹ค(\(\mathsf{bioS}^{\mathsf{simple}}\) ๋ฐ์ดํ„ฐ์—์„œ ์ˆ˜ํ–‰ํ•œ ๊ฒƒ์ฒ˜๋Ÿผ). ๋ฐ์ดํ„ฐ ๋‹ค์–‘์„ฑ์ด ์—†์œผ๋ฉด ๋ชจ๋ธ์ด ๋ฌธ์žฅ ๊ตฌ์กฐ๋ฅผ ์•”๊ธฐํ•˜๋Š” ๋ฐ ์šฉ๋Ÿ‰์„ ๋‚ญ๋น„ํ•˜์—ฌ ์šฉ๋Ÿ‰ ์†์‹ค์ด ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
Since its introduction by Lewis et al. (Lewis et al., 2020) in 2020, RAG technology has undergone significant advancements, particularly influenced by ChatGPT's success. However, there is a noticeable gap in the literature regarding a thorough analysis of RAG's mechanisms and the progress made by subsequent studies. Furthermore, the field is characterized by diverse research focuses and the use of ambiguous terminology for similar methods, leading to confusion.
Lewis ๋“ฑ(2020)์— ์˜ํ•ด 2020๋…„์— ๋„์ž…๋œ ์ดํ›„, RAG ๊ธฐ์ˆ ์€ ํŠนํžˆ ChatGPT์˜ ์„ฑ๊ณต์— ์˜ํ•ด ํฌ๊ฒŒ ๋ฐœ์ „ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ RAG์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜๊ณผ ํ›„์† ์—ฐ๊ตฌ์—์„œ ์ด๋ฃจ์–ด์ง„ ์ง„์ „์— ๋Œ€ํ•œ ์ฒ ์ €ํ•œ ๋ถ„์„์— ๊ด€ํ•œ ๋ฌธํ—Œ์€ ๋ˆˆ์— ๋„๊ฒŒ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด ๋ถ„์•ผ๋Š” ๋‹ค์–‘ํ•œ ์—ฐ๊ตฌ ์ดˆ์ ๊ณผ ์œ ์‚ฌํ•œ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ๋ชจํ˜ธํ•œ ์šฉ์–ด ์‚ฌ์šฉ์œผ๋กœ ์ธํ•ด ํ˜ผ๋ž€์„ ์ดˆ๋ž˜ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
2020๋…„ Lewis ๋“ฑ(Lewis et al., 2020)์— ์˜ํ•ด ์†Œ๊ฐœ๋œ ์ด๋ž˜, RAG ๊ธฐ์ˆ ์€ ํŠนํžˆ ChatGPT์˜ ์„ฑ๊ณต์— ์˜ํ–ฅ์„ ๋ฐ›์•„ ์ƒ๋‹นํ•œ ๋ฐœ์ „์„ ์ด๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ RAG์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์— ๋Œ€ํ•œ ์ฒ ์ €ํ•œ ๋ถ„์„๊ณผ ํ›„์† ์—ฐ๊ตฌ์—์„œ ์ด๋ฃจ์–ด์ง„ ์ง„์ „์— ๋Œ€ํ•œ ๋ฌธํ—Œ์—๋Š” ๋ˆˆ์— ๋„๋Š” ๊ฒฉ์ฐจ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด ๋ถ„์•ผ๋Š” ์—ฐ๊ตฌ ์ดˆ์ ์ด ๋‹ค์–‘ํ•˜๊ณ  ์œ ์‚ฌํ•œ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ๋ชจํ˜ธํ•œ ์šฉ์–ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ˜ผ๋ž€์„ ์ดˆ๋ž˜ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
With NearDup, we found that the web-scrape datasets contain between 3.04% (on C4) to 13.63% (on RealNews) near duplicates (Table 2). Nearduplicate text is much less common in Wiki-40B, forming only 0.39% of the train set.2 In C4, the majority (1.8M) of near-duplicate clusters consisted of just a single pair of examples that matched against each other, but there were 280 clusters with over 5,000 examples in them (Figure 1), including one cluster of size 250,933.
NearDup์„ ์‚ฌ์šฉํ•˜์—ฌ ์›น ์Šคํฌ๋ž˜ํ•‘ ๋ฐ์ดํ„ฐ์…‹์—์„œ 3.04% (C4)์—์„œ 13.63% (RealNews) ์‚ฌ์ด์˜ ์œ ์‚ฌ ์ค‘๋ณต์ด ํฌํ•จ๋˜์–ด ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค (ํ‘œ 2). ์œ ์‚ฌ ์ค‘๋ณต ํ…์ŠคํŠธ๋Š” Wiki-40B์—์„œ๋Š” ํ›จ์”ฌ ๋œ ์ผ๋ฐ˜์ ์ด๋ฉฐ, ํ•™์Šต ์„ธํŠธ์˜ 0.39%๋งŒ์„ ์ฐจ์ง€ํ•ฉ๋‹ˆ๋‹ค. C4์—์„œ๋Š” ๋Œ€๋ถ€๋ถ„์˜ ์œ ์‚ฌ ์ค‘๋ณต ํด๋Ÿฌ์Šคํ„ฐ(1.8M)๊ฐ€ ์„œ๋กœ ์ผ์น˜ํ•˜๋Š” ๋‹จ์ผ ์Œ์˜ ์˜ˆ์ œ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์ง€๋งŒ, 5,000๊ฐœ ์ด์ƒ์˜ ์˜ˆ์ œ๋ฅผ ํฌํ•จํ•˜๋Š” 280๊ฐœ์˜ ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค (๊ทธ๋ฆผ 1), ์—ฌ๊ธฐ์—๋Š” 250,933๊ฐœ์˜ ์˜ˆ์ œ๋ฅผ ํฌํ•จํ•˜๋Š” ํ•˜๋‚˜์˜ ํด๋Ÿฌ์Šคํ„ฐ๋„ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
NearDup์„ ์‚ฌ์šฉํ•˜๋ฉด ์›น ์Šคํฌ๋žฉ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— 3.04%(C4)์—์„œ 13.63%(RealNews) ์‚ฌ์ด์˜ ๊ทผ์‚ฌ ์ค‘๋ณต์ด ํฌํ•จ๋˜์–ด ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(ํ‘œ 2). ์œ„ํ‚ค-40B์—๋Š” ๊ทผ์‚ฌ ์ค‘๋ณต ํ…์ŠคํŠธ๊ฐ€ ํ›จ์”ฌ ๋œ ์ผ๋ฐ˜์ ์ด๋ฉฐ, ํ•™์Šต ์„ธํŠธ์˜ 0.39%์— ๋ถˆ๊ณผํ•ฉ๋‹ˆ๋‹ค.2 C4์—์„œ๋Š” ๋Œ€๋ถ€๋ถ„(1.8M)์˜ ๊ทผ์‚ฌ ์ค‘๋ณต ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ ์„œ๋กœ ์ผ์น˜ํ•˜๋Š” ํ•œ ์Œ์˜ ์˜ˆ์‹œ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์ง€๋งŒ, 250,933๊ฐœ ํฌ๊ธฐ์˜ ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ํฌํ•จํ•˜์—ฌ 5,000๊ฐœ ์ด์ƒ์˜ ์˜ˆ์‹œ๊ฐ€ ํฌํ•จ๋œ 280๊ฐœ์˜ ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1).
Although large language models (LLMs) demonstrate remarkable capabilities in variety of applications (Zhao et al., 2023), such as language generation, understanding, and reasoning, they struggle to provide accurate answers when faced with complicated tasks. To improve the performance of LLMs, some of recent studies focus on ensemble methods (Wang et al., 2023; Wan et al., 2024) and multiple LLM-Agents collaboration frameworks (Du et al., 2023; Wu et al., 2023).
๋น„๋ก ๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธ(LLMs)์ด ์–ธ์–ด ์ƒ์„ฑ, ์ดํ•ด ๋ฐ ์ถ”๋ก ๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์‘์šฉ ๋ถ„์•ผ์—์„œ ๋†€๋ผ์šด ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์ง€๋งŒ(Zhao et al., 2023), ๋ณต์žกํ•œ ์ž‘์—…์— ์ง๋ฉดํ–ˆ์„ ๋•Œ ์ •ํ™•ํ•œ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. LLM์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด, ์ตœ๊ทผ์˜ ์ผ๋ถ€ ์—ฐ๊ตฌ๋“ค์€ ์•™์ƒ๋ธ” ๋ฐฉ๋ฒ•(Wang et al., 2023; Wan et al., 2024)๊ณผ ๋‹ค์ค‘ LLM-์—์ด์ „ํŠธ ํ˜‘์—… ํ”„๋ ˆ์ž„์›Œํฌ(Du et al., 2023; Wu et al., 2023)์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ์–ธ์–ด ์ƒ์„ฑ, ์ดํ•ด, ์ถ”๋ก  ๋“ฑ ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ ๋†€๋ผ์šด ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์ง€๋งŒ(Zhao et al., 2023), ๋ณต์žกํ•œ ์ž‘์—…์— ์ง๋ฉดํ•˜๋ฉด ์ •ํ™•ํ•œ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. LLM์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ตœ๊ทผ ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ์•™์ƒ๋ธ” ๋ฐฉ๋ฒ•(Wang et al., 2023; Wan et al., 2024)๊ณผ ์—ฌ๋Ÿฌ LLM-์—์ด์ „ํŠธ ํ˜‘์—… ํ”„๋ ˆ์ž„์›Œํฌ(Du et al., 2023; Wu et al., 2023)์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
In this work, we explore the following: 1) Is domain-adaptive continual pre-training helpful in building domain-specific LLMs?; 2) Can we employ data selection strategies for a more effective domain-adaptive continual pre-training?; and 3) Does domain-adaptive continual pre-training hurt LLM's open-domain capabilities? We answer these questions in the confines of finance domain by training a continually pre-trained model, FinPythia, built on top of Pythia [4].
์ด ์—ฐ๊ตฌ์—์„œ ์šฐ๋ฆฌ๋Š” ๋‹ค์Œ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค: 1) ๋„๋ฉ”์ธ ์ ์‘ํ˜• ์ง€์†์  ์‚ฌ์ „ ํ›ˆ๋ จ์ด ๋„๋ฉ”์ธ ํŠนํ™” LLM์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š”๊ฐ€?; 2) ๋” ํšจ๊ณผ์ ์ธ ๋„๋ฉ”์ธ ์ ์‘ํ˜• ์ง€์†์  ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€?; 3) ๋„๋ฉ”์ธ ์ ์‘ํ˜• ์ง€์†์  ์‚ฌ์ „ ํ›ˆ๋ จ์ด LLM์˜ ์˜คํ”ˆ ๋„๋ฉ”์ธ ๊ธฐ๋Šฅ์— ํ•ด๋ฅผ ๋ผ์น˜๋Š”๊ฐ€? ์šฐ๋ฆฌ๋Š” Pythia [4]๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์ถ•๋œ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์ธ FinPythia๋ฅผ ํ›ˆ๋ จํ•˜์—ฌ ๊ธˆ์œต ๋„๋ฉ”์ธ์˜ ๋ฒ”์œ„ ๋‚ด์—์„œ ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•ฉ๋‹ˆ๋‹ค.
์ด ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์‚ฌํ•ญ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค: 1) ๋„๋ฉ”์ธ ์ ์‘ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์ด ๋„๋ฉ”์ธ๋ณ„ LLM ๊ตฌ์ถ•์— ๋„์›€์ด ๋˜๋Š”๊ฐ€?; 2) ๋ณด๋‹ค ํšจ๊ณผ์ ์ธ ๋„๋ฉ”์ธ ์ ์‘ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€?; 3) ๋„๋ฉ”์ธ ์ ์‘ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์ด LLM์˜ ์˜คํ”ˆ ๋„๋ฉ”์ธ ๊ธฐ๋Šฅ์— ํ•ด๊ฐ€ ๋˜๋Š”๊ฐ€? ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์— ๋Œ€ํ•ด Pythia [4]๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์ถ•๋œ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์ธ FinPythia๋ฅผ ํ›ˆ๋ จํ•˜์—ฌ ๊ธˆ์œต ๋„๋ฉ”์ธ ๋‚ด์—์„œ ๋‹ตํ•ฉ๋‹ˆ๋‹ค.
The memory module plays a very important role in the agent architecture design. It stores information perceived from the environment and leverages the recorded memories to facilitate future actions. The memory module can help the agent to accumulate experiences, self-evolve, and behave in a more consistent, reasonable, and effective manner. This section provides a comprehensive overview of the memory module, focusing on its structures, formats, and operations.
๋ฉ”๋ชจ๋ฆฌ ๋ชจ๋“ˆ์€ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜ ์„ค๊ณ„์—์„œ ๋งค์šฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋“ˆ์€ ํ™˜๊ฒฝ์—์„œ ์ธ์ง€ํ•œ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๊ณ  ๊ธฐ๋ก๋œ ๊ธฐ์–ต์„ ํ™œ์šฉํ•˜์—ฌ ๋ฏธ๋ž˜์˜ ํ–‰๋™์„ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ ๋ชจ๋“ˆ์€ ์—์ด์ „ํŠธ๊ฐ€ ๊ฒฝํ—˜์„ ์ถ•์ ํ•˜๊ณ , ์Šค์Šค๋กœ ์ง„ํ™”ํ•˜๋ฉฐ, ๋” ์ผ๊ด€๋˜๊ณ  ํ•ฉ๋ฆฌ์ ์ด๋ฉฐ ํšจ๊ณผ์ ์œผ๋กœ ํ–‰๋™ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ค๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ๋ฉ”๋ชจ๋ฆฌ ๋ชจ๋“ˆ์˜ ๊ตฌ์กฐ, ํ˜•์‹ ๋ฐ ์ž‘๋™์— ๋Œ€ํ•ด ์ข…ํ•ฉ์ ์ธ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๋ฉ”๋ชจ๋ฆฌ ๋ชจ๋“ˆ์€ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜ ์„ค๊ณ„์—์„œ ๋งค์šฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ ๋ชจ๋“ˆ์€ ํ™˜๊ฒฝ์—์„œ ์ธ์‹ํ•œ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๊ณ  ๊ธฐ๋ก๋œ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ–ฅํ›„ ํ–‰๋™์„ ์šฉ์ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ ๋ชจ๋“ˆ์€ ์—์ด์ „ํŠธ๊ฐ€ ๊ฒฝํ—˜์„ ์ถ•์ ํ•˜๊ณ , ์Šค์Šค๋กœ ์ง„ํ™”ํ•˜๋ฉฐ, ๋ณด๋‹ค ์ผ๊ด€๋˜๊ณ  ํ•ฉ๋ฆฌ์ ์ด๋ฉฐ ํšจ๊ณผ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ํ–‰๋™ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ๋ฉ”๋ชจ๋ฆฌ ๋ชจ๋“ˆ์˜ ๊ตฌ์กฐ, ํ˜•์‹ ๋ฐ ์ž‘๋™์— ์ค‘์ ์„ ๋‘๊ณ  ๋ฉ”๋ชจ๋ฆฌ ๋ชจ๋“ˆ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
In the above sections, we introduce both subjective and objective strategies for LLM-based autonomous agents evaluation. The evaluation of the agents play significant roles in this domain. However, both subjective and objective evaluation have their own strengths and weakness. Maybe, in practice, they should be combined to comprehensively evaluate the agents. We summarize the correspondence between the previous work and these evaluation strategies in Table 3.
์œ„์˜ ์„น์…˜์—์„œ๋Š” LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ์ฃผ๊ด€์  ๋ฐ ๊ฐ๊ด€์  ์ „๋žต์„ ๋ชจ๋‘ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์˜ ํ‰๊ฐ€๋Š” ์ด ๋ถ„์•ผ์—์„œ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฃผ๊ด€์  ํ‰๊ฐ€์™€ ๊ฐ๊ด€์  ํ‰๊ฐ€๋Š” ๊ฐ๊ฐ์˜ ๊ฐ•์ ๊ณผ ์•ฝ์ ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค๋ฌด์—์„œ๋Š” ์ด๋“ค์„ ๊ฒฐํ•ฉํ•˜์—ฌ ์—์ด์ „ํŠธ๋ฅผ ์ข…ํ•ฉ์ ์œผ๋กœ ํ‰๊ฐ€ํ•ด์•ผ ํ• ์ง€๋„ ๋ชจ๋ฆ…๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ์™€ ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ์ „๋žต ๊ฐ„์˜ ๋Œ€์‘ ๊ด€๊ณ„๋ฅผ Table 3์— ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค.
์œ„์˜ ์„น์…˜์—์„œ๋Š” LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ์ฃผ๊ด€์  ๋ฐ ๊ฐ๊ด€์  ์ „๋žต์„ ๋ชจ๋‘ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์˜ ํ‰๊ฐ€๋Š” ์ด ์˜์—ญ์—์„œ ๋งค์šฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ฃผ๊ด€์  ํ‰๊ฐ€์™€ ๊ฐ๊ด€์  ํ‰๊ฐ€ ๋ชจ๋‘ ์žฅ๋‹จ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ๋Š” ์ด ๋‘ ๊ฐ€์ง€ ํ‰๊ฐ€๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ ์ข…ํ•ฉ์ ์œผ๋กœ ์—์ด์ „ํŠธ๋ฅผ ํ‰๊ฐ€ํ•ด์•ผ ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์ „ ์ž‘์—…๊ณผ ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ์ „๋žต ๊ฐ„์˜ ๋Œ€์‘์„ ํ‘œ 3์— ์š”์•ฝํ–ˆ์Šต๋‹ˆ๋‹ค.
Prompt engineering transcends the mere construction of prompts; it requires a blend of domain knowledge, understanding of the AI model, and a methodical approach to tailor prompts for different contexts. This might involve creating templates that can be programmatically modified based on a given dataset or context. For example, generating personalized responses based on user data might use a template that is dynamically filled with relevant user information.
ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง์€ ๋‹จ์ˆœํžˆ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์„ ๋„˜์–ด์„ญ๋‹ˆ๋‹ค. ์ด๋Š” ๋„๋ฉ”์ธ ์ง€์‹, AI ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ดํ•ด, ๊ทธ๋ฆฌ๊ณ  ๋‹ค์–‘ํ•œ ์ƒํ™ฉ์— ๋งž๊ฒŒ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์กฐ์ •ํ•˜๋Š” ์ฒด๊ณ„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์š”๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ฃผ์–ด์ง„ ๋ฐ์ดํ„ฐ์…‹์ด๋‚˜ ์ƒํ™ฉ์— ๋”ฐ๋ผ ํ”„๋กœ๊ทธ๋žจ์ ์œผ๋กœ ์ˆ˜์ •ํ•  ์ˆ˜ ์žˆ๋Š” ํ…œํ”Œ๋ฆฟ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ์ธํ™”๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝ์šฐ, ๊ด€๋ จ ์‚ฌ์šฉ์ž ์ •๋ณด๋กœ ๋™์ ์œผ๋กœ ์ฑ„์›Œ์ง€๋Š” ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง์€ ๋‹จ์ˆœํ•œ ํ”„๋กฌํ”„ํŠธ ๊ตฌ์„ฑ์„ ๋„˜์–ด ๋„๋ฉ”์ธ ์ง€์‹, AI ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ดํ•ด, ๋‹ค์–‘ํ•œ ์ƒํ™ฉ์— ๋งž๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋งž์ถคํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ฒด๊ณ„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ฃผ์–ด์ง„ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋˜๋Š” ์ปจํ…์ŠคํŠธ์— ๋”ฐ๋ผ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋ฐฉ์‹์œผ๋กœ ์ˆ˜์ •ํ•  ์ˆ˜ ์žˆ๋Š” ํ…œํ”Œ๋ฆฟ์„ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ์ธํ™”๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝ์šฐ ๊ด€๋ จ ์‚ฌ์šฉ์ž ์ •๋ณด๋กœ ๋™์ ์œผ๋กœ ์ฑ„์›Œ์ง€๋Š” ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 5 shows that singular distributions evolve differently for models using less than 410M parameters than for the larger ones. The heads of small models see their singular value distributions become increasingly uniform, up to a point where they degenerate abruptly, which again correlates with the LM performance drop. The singular value distributions of larger models tend to be more stable, and do not display clear monotonic patterns throughout training.
Figure 5๋Š” 410M ๋ฏธ๋งŒ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ๊ณผ ๋” ํฐ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋‹จ์ผ ๋ถ„ํฌ๊ฐ€ ๋‹ค๋ฅด๊ฒŒ ์ง„ํ™”ํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ž‘์€ ๋ชจ๋ธ์˜ ํ—ค๋“œ๋Š” ๋‹จ์ผ ๊ฐ’ ๋ถ„ํฌ๊ฐ€ ์ ์  ๋” ๊ท ์ผํ•ด์ง€๋‹ค๊ฐ€ ๊ฐ‘์ž๊ธฐ ํ‡ดํ™”ํ•˜๋Š” ์ง€์ ์— ๋„๋‹ฌํ•˜๋ฉฐ, ์ด๋Š” LM ์„ฑ๋Šฅ ์ €ํ•˜์™€ ๋‹ค์‹œ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค. ๋” ํฐ ๋ชจ๋ธ์˜ ๋‹จ์ผ ๊ฐ’ ๋ถ„ํฌ๋Š” ๋” ์•ˆ์ •์ ์ธ ๊ฒฝํ–ฅ์„ ๋ณด์ด๋ฉฐ, ํ›ˆ๋ จ ๋‚ด๋‚ด ๋ช…ํ™•ํ•œ ๋‹จ์กฐ ํŒจํ„ด์„ ๋‚˜ํƒ€๋‚ด์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 5๋Š” 410M ๋ฏธ๋งŒ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํŠน์ด ๋ถ„ํฌ๊ฐ€ ๋” ํฐ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ์™€ ๋‹ค๋ฅด๊ฒŒ ์ง„ํ™”ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ž‘์€ ๋ชจ๋ธ์˜ ํ—ค๋“œ๋Š” ํŠน์ด๊ฐ’ ๋ถ„ํฌ๊ฐ€ ์ ์  ๋” ๊ท ์ผํ•ด์ง€๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋‹ค์‹œ LM ์„ฑ๋Šฅ ์ €ํ•˜์™€ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋” ํฐ ๋ชจ๋ธ์˜ ํŠน์ด๊ฐ’ ๋ถ„ํฌ๋Š” ๋” ์•ˆ์ •์ ์ธ ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฉฐ, ํ›ˆ๋ จ ์ „๋ฐ˜์— ๊ฑธ์ณ ๋ช…ํ™•ํ•œ ๋‹จ์กฐ๋กœ์šด ํŒจํ„ด์„ ๋ณด์ด์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
We introduce an exposition of the datasets employed in our work for evaluating the performance in terms of text generation and toxicity rejection abilities. Our experiments consistently employed approach for the generation of responses by LLaMA-based models involves the implementation of a greedy decoding strategy. In the case of Mistral-based models, the generation of responses was facilitated by setting the repetition penalty Keskar et al. (2019) to 1.15.
์šฐ๋ฆฌ ์—ฐ๊ตฌ์—์„œ ํ…์ŠคํŠธ ์ƒ์„ฑ ๋ฐ ์œ ํ•ด์„ฑ ๊ฑฐ๋ถ€ ๋Šฅ๋ ฅ ์ธก๋ฉด์—์„œ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•œ ์„ค๋ช…์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์‹คํ—˜์—์„œ๋Š” LLaMA ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ด ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ผ๊ด€๋˜๊ฒŒ ํƒ์š•์  ๋””์ฝ”๋”ฉ ์ „๋žต์„ ๊ตฌํ˜„ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. Mistral ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, ์‘๋‹ต ์ƒ์„ฑ์„ ์œ„ํ•ด ๋ฐ˜๋ณต ํŽ˜๋„ํ‹ฐ๋ฅผ 1.15๋กœ ์„ค์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค(Keskar et al., 2019).
ํ…์ŠคํŠธ ์ƒ์„ฑ ๋ฐ ๋…์„ฑ ๊ฑฐ๋ถ€ ๋Šฅ๋ ฅ ์ธก๋ฉด์—์„œ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋ณธ ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์‹คํ—˜์—์„œ๋Š” LLaMA ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์˜ ์‘๋‹ต ์ƒ์„ฑ์— ๋Œ€ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ผ๊ด€๋˜๊ฒŒ ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ, ํƒ์š•์Šค๋Ÿฌ์šด ๋””์ฝ”๋”ฉ ์ „๋žต์„ ๊ตฌํ˜„ํ–ˆ์Šต๋‹ˆ๋‹ค. Mistral ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, ๋ฐ˜๋ณต ํŽ˜๋„ํ‹ฐ Keskar ๋“ฑ(2019)์„ 1.15๋กœ ์„ค์ •ํ•˜์—ฌ ์‘๋‹ต ์ƒ์„ฑ์„ ์šฉ์ดํ•˜๊ฒŒ ํ–ˆ์Šต๋‹ˆ๋‹ค.
In Appendix Figure 8, we aim to provide some insight into the kinds of tokens where the improvements occur. Namely, while thinking appears to help for many tokens in the example, inspection suggests it disproportionately help to predict tokens where recalling relevant information is useful, such as the name of an applicable theorem or the start of the next step in a proof. Notably, this would align well with the framing proposed by Prystawski et al. (2024).
๋ถ€๋ก ๊ทธ๋ฆผ 8์—์„œ ์šฐ๋ฆฌ๋Š” ๊ฐœ์„ ์ด ๋ฐœ์ƒํ•˜๋Š” ํ† ํฐ์˜ ์ข…๋ฅ˜์— ๋Œ€ํ•œ ํ†ต์ฐฐ์„ ์ œ๊ณตํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์˜ˆ์ œ์—์„œ ์ƒ๊ฐํ•˜๋Š” ๊ฒƒ์ด ๋งŽ์€ ํ† ํฐ์— ๋„์›€์ด ๋˜๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๋ณด์ด์ง€๋งŒ, ๊ฒ€ํ†  ๊ฒฐ๊ณผ ๊ด€๋ จ ์ •๋ณด๋ฅผ ํšŒ์ƒํ•˜๋Š” ๊ฒƒ์ด ์œ ์šฉํ•œ ํ† ํฐ, ์˜ˆ๋ฅผ ๋“ค์–ด ์ ์šฉ ๊ฐ€๋Šฅํ•œ ์ •๋ฆฌ์˜ ์ด๋ฆ„์ด๋‚˜ ์ฆ๋ช…์˜ ๋‹ค์Œ ๋‹จ๊ณ„์˜ ์‹œ์ž‘ ๋ถ€๋ถ„์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ๋ถˆ๊ท ํ˜•์ ์œผ๋กœ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ด๋Š” Prystawski et al. (2024)์ด ์ œ์•ˆํ•œ ํ”„๋ ˆ์ด๋ฐ๊ณผ ์ž˜ ๋งž์•„๋–จ์–ด์ง‘๋‹ˆ๋‹ค.
๋ถ€๋ก ๊ทธ๋ฆผ 8์—์„œ๋Š” ๊ฐœ์„ ์ด ๋ฐœ์ƒํ•˜๋Š” ํ† ํฐ์˜ ์ข…๋ฅ˜์— ๋Œ€ํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์˜ˆ์ œ์—์„œ ์‚ฌ๊ณ ๊ฐ€ ๋งŽ์€ ํ† ํฐ์— ๋„์›€์ด ๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ด์ง€๋งŒ, ๊ฒ€์‚ฌ ๊ฒฐ๊ณผ ์ ์šฉ ๊ฐ€๋Šฅํ•œ ์ •๋ฆฌ์˜ ์ด๋ฆ„์ด๋‚˜ ์ฆ๋ช…์˜ ๋‹ค์Œ ๋‹จ๊ณ„ ์‹œ์ž‘๊ณผ ๊ฐ™์ด ๊ด€๋ จ ์ •๋ณด๋ฅผ ๊ธฐ์–ตํ•˜๋Š” ๊ฒƒ์ด ์œ ์šฉํ•œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ๋ถˆ๊ท ํ˜•์ ์œผ๋กœ ๋„์›€์ด ๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์ด๋Š” Prystawski ๋“ฑ(2024)์ด ์ œ์•ˆํ•œ ํ”„๋ ˆ์ž„๊ณผ ์ž˜ ๋ถ€ํ•ฉํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
**Impact of Feedback and Human Oversight on Agent Systems.** When solving a complex problem, it is extremely unlikely that one provides a correct, robust solution on their first try. Instead, one might pose a potential solution before criticizing it and refining it. One could also consult with someone else and receive feedback from another perspective. The same idea of iterative feedback and refinement is essential for helping agents solve complex problems.
ํ”ผ๋“œ๋ฐฑ๊ณผ ์ธ๊ฐ„ ๊ฐ๋…์ด ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ. ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ๋•Œ, ์ฒ˜์Œ ์‹œ๋„์—์„œ ์˜ฌ๋ฐ”๋ฅด๊ณ  ๊ฒฌ๊ณ ํ•œ ํ•ด๊ฒฐ์ฑ…์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ๋“œ๋ญ…๋‹ˆ๋‹ค. ๋Œ€์‹ , ์ž ์žฌ์ ์ธ ํ•ด๊ฒฐ์ฑ…์„ ์ œ์‹œํ•œ ํ›„ ์ด๋ฅผ ๋น„ํŒํ•˜๊ณ  ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค๋ฅธ ์‚ฌ๋žŒ๊ณผ ์ƒ์˜ํ•˜์—ฌ ๋‹ค๋ฅธ ๊ด€์ ์—์„œ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ˜๋ณต์ ์ธ ํ”ผ๋“œ๋ฐฑ๊ณผ ๊ฐœ์„ ์˜ ๊ฐœ๋…์€ ์—์ด์ „ํŠธ๊ฐ€ ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค.
**ํ”ผ๋“œ๋ฐฑ๊ณผ ์ธ๊ฐ„์˜ ๊ฐ๋…์ด ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ**. ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ๋•Œ ์ฒซ ๋ฒˆ์งธ ์‹œ๋„์—์„œ ์ •ํ™•ํ•˜๊ณ  ๊ฐ•๋ ฅํ•œ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•  ๊ฐ€๋Šฅ์„ฑ์€ ๋งค์šฐ ๋‚ฎ์Šต๋‹ˆ๋‹ค. ๋Œ€์‹  ์ž ์žฌ์ ์ธ ํ•ด๊ฒฐ์ฑ…์„ ์ œ์‹œํ•œ ํ›„ ๋น„ํŒํ•˜๊ณ  ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์‚ฌ๋žŒ๊ณผ ์ƒ์˜ํ•˜๊ณ  ๋‹ค๋ฅธ ๊ด€์ ์—์„œ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ›์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ณต์ ์ธ ํ”ผ๋“œ๋ฐฑ๊ณผ ๊ฐœ์„ ์ด๋ผ๋Š” ๋™์ผํ•œ ์•„์ด๋””์–ด๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค.
Language modeling is a long-standing research topic, dating back to the 1950s with Shannon's application of information theory to human language, where he measured how well simple n-gram language models predict or compress natural language text [3]. Since then, statistical language modeling became fundamental to many natural language understanding and generation tasks, ranging from speech recognition, machine translation, to information retrieval [4, 5, 6].
์–ธ์–ด ๋ชจ๋ธ๋ง์€ ์˜ค๋žœ ์—ฐ๊ตฌ ์ฃผ์ œ๋กœ, 1950๋…„๋Œ€์— Shannon์ด ์ •๋ณด ์ด๋ก ์„ ์ธ๊ฐ„ ์–ธ์–ด์— ์ ์šฉํ•˜๋ฉด์„œ ์‹œ์ž‘๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Š” ๊ฐ„๋‹จํ•œ n-๊ทธ๋žจ ์–ธ์–ด ๋ชจ๋ธ์ด ์ž์—ฐ์–ด ํ…์ŠคํŠธ๋ฅผ ์–ผ๋งˆ๋‚˜ ์ž˜ ์˜ˆ์ธกํ•˜๊ฑฐ๋‚˜ ์••์ถ•ํ•˜๋Š”์ง€๋ฅผ ์ธก์ •ํ–ˆ์Šต๋‹ˆ๋‹ค [3]. ๊ทธ ์ดํ›„๋กœ ํ†ต๊ณ„์  ์–ธ์–ด ๋ชจ๋ธ๋ง์€ ์Œ์„ฑ ์ธ์‹, ๊ธฐ๊ณ„ ๋ฒˆ์—ญ, ์ •๋ณด ๊ฒ€์ƒ‰ ๋“ฑ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ์ž‘์—…์˜ ๊ธฐ๋ณธ์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค [4, 5, 6].
์–ธ์–ด ๋ชจ๋ธ๋ง์€ 1950๋…„๋Œ€ ์„€๋„Œ์ด ์ •๋ณด ์ด๋ก ์„ ์ธ๊ฐ„ ์–ธ์–ด์— ์ ์šฉํ•˜์—ฌ ๋‹จ์ˆœํ•œ n-๊ทธ๋žจ ์–ธ์–ด ๋ชจ๋ธ์ด ์ž์—ฐ์–ด ํ…์ŠคํŠธ๋ฅผ ์–ผ๋งˆ๋‚˜ ์ž˜ ์˜ˆ์ธกํ•˜๊ฑฐ๋‚˜ ์••์ถ•ํ•˜๋Š”์ง€๋ฅผ ์ธก์ •ํ•œ ๊ฒƒ์œผ๋กœ ๊ฑฐ์Šฌ๋Ÿฌ ์˜ฌ๋ผ๊ฐ€๋Š” ์˜ค๋žœ ์—ฐ๊ตฌ ์ฃผ์ œ์ž…๋‹ˆ๋‹ค[3]. ๊ทธ ์ดํ›„๋กœ ํ†ต๊ณ„ ์–ธ์–ด ๋ชจ๋ธ๋ง์€ ์Œ์„ฑ ์ธ์‹, ๊ธฐ๊ณ„ ๋ฒˆ์—ญ, ์ •๋ณด ๊ฒ€์ƒ‰ ๋“ฑ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ์ž‘์—…์˜ ๊ธฐ์ดˆ๊ฐ€ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[4, 5, 6].
Reinforcing LM biases.A point of concern for the authors is the unintended consequences of this iterative algorithm, such as the amplification of problematic social biases (stereotypes or slurs about gender, race, etc.). Relatedly, one observed challenge in this process is the algorithm's difficulty in producing balanced labels, which reflected models' prior biases. We hope future work will lead to better understanding of the pros and cons of the approach.
LM ํŽธํ–ฅ ๊ฐ•ํ™”. ์ €์ž๋“ค์ด ์šฐ๋ คํ•˜๋Š” ์  ์ค‘ ํ•˜๋‚˜๋Š” ์ด ๋ฐ˜๋ณต ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์˜๋„์น˜ ์•Š์€ ๊ฒฐ๊ณผ๋กœ, ๋ฌธ์ œ์ ์ธ ์‚ฌํšŒ์  ํŽธํ–ฅ(์„ฑ๋ณ„, ์ธ์ข… ๋“ฑ์— ๋Œ€ํ•œ ๊ณ ์ •๊ด€๋…์ด๋‚˜ ๋น„๋ฐฉ ๋“ฑ)์ด ์ฆํญ๋˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด์™€ ๊ด€๋ จํ•˜์—ฌ, ์ด ๊ณผ์ •์—์„œ ๊ด€์ฐฐ๋œ ํ•œ ๊ฐ€์ง€ ๋„์ „ ๊ณผ์ œ๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ๋ชจ๋ธ์˜ ๊ธฐ์กด ํŽธํ–ฅ์„ ๋ฐ˜์˜ํ•œ ๊ท ํ˜• ์žกํžŒ ๋ ˆ์ด๋ธ”์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๋Š”๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ๊ฐ€ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์žฅ๋‹จ์ ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๋Š” ๋ฐ ๊ธฐ์—ฌํ•˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
LM ํŽธํ–ฅ์„ฑ ๊ฐ•ํ™”.์ €์ž๋“ค์ด ์šฐ๋ คํ•˜๋Š” ํ•œ ๊ฐ€์ง€ ์ ์€ ์ด ๋ฐ˜๋ณต์ ์ธ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ๋ฌธ์ œ๊ฐ€ ๋˜๋Š” ์‚ฌํšŒ์  ํŽธ๊ฒฌ(์„ฑ๋ณ„, ์ธ์ข… ๋“ฑ์— ๋Œ€ํ•œ ๊ณ ์ •๊ด€๋…์ด๋‚˜ ๋น„๋ฐฉ ๋“ฑ)์„ ์ฆํญ์‹œํ‚ค๋Š” ๋“ฑ ์˜๋„ํ•˜์ง€ ์•Š์€ ๊ฒฐ๊ณผ๋ฅผ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์ด์™€ ๊ด€๋ จํ•˜์—ฌ ์ด ๊ณผ์ •์—์„œ ๊ด€์ฐฐ๋œ ํ•œ ๊ฐ€์ง€ ๋ฌธ์ œ๋Š” ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํŽธ๊ฒฌ์„ ๋ฐ˜์˜ํ•˜๋Š” ๊ท ํ˜• ์žกํžŒ ๋ ˆ์ด๋ธ”์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์žˆ์–ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์–ด๋ ค์›€์ž…๋‹ˆ๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ๋ฅผ ํ†ตํ•ด ์ด ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์žฅ๋‹จ์ ์„ ๋” ์ž˜ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
The effectiveness of these strategies is demonstrated in Table 8, which details the Yi-9B base model's performance across a variety of benchmarks, including common sense, reasoning, knowledge, coding, and mathematics. It underscores the competitive advantages of Yi-9B base model in specific domains, illustrating the efficacy of our methodology in enhancing model performance by optimally adjusting the interplay between data characteristics and model size.
์ด๋Ÿฌํ•œ ์ „๋žต์˜ ํšจ๊ณผ๋Š” Table 8์—์„œ ์ž…์ฆ๋˜์—ˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์—๋Š” ์ƒ์‹, ์ถ”๋ก , ์ง€์‹, ์ฝ”๋”ฉ ๋ฐ ์ˆ˜ํ•™์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ Yi-9B ๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ์ž์„ธํžˆ ์„ค๋ช…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํŠน์ • ๋„๋ฉ”์ธ์—์„œ Yi-9B ๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ๊ฒฝ์Ÿ ์šฐ์œ„๋ฅผ ๊ฐ•์กฐํ•˜๋ฉฐ, ๋ฐ์ดํ„ฐ ํŠน์„ฑ๊ณผ ๋ชจ๋ธ ํฌ๊ธฐ ๊ฐ„์˜ ์ƒํ˜ธ์ž‘์šฉ์„ ์ตœ์ ํ™”ํ•˜์—ฌ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•๋ก ์˜ ํšจ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ์ „๋žต์˜ ํšจ๊ณผ๋Š” ์ƒ์‹, ์ถ”๋ก , ์ง€์‹, ์ฝ”๋”ฉ, ์ˆ˜ํ•™ ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ Yi-9B ๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์ž์„ธํžˆ ๋ณด์—ฌ์ฃผ๋Š” ํ‘œ 8์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํŠน์ • ๋„๋ฉ”์ธ์—์„œ Yi-9B ๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ๊ฒฝ์Ÿ ์šฐ์œ„๋ฅผ ๊ฐ•์กฐํ•˜๋ฉฐ, ๋ฐ์ดํ„ฐ ํŠน์„ฑ๊ณผ ๋ชจ๋ธ ํฌ๊ธฐ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ์„ ์ตœ์ ์œผ๋กœ ์กฐ์ •ํ•˜์—ฌ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•๋ก ์˜ ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
De-duplication refers to the process of removing duplicate instances or repeated occurrences of the same data in a dataset. Duplicate data points can introduce biases in the model training process and reduce the diversity, as the model may learn from the same examples multiple times, potentially leading to overfitting on those particular instances. Some works [125] have shown that de-duplication improves models' ability to generalize to new, unseen data.
์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๋ฐ์ดํ„ฐ์…‹์—์„œ ์ค‘๋ณต๋œ ์ธ์Šคํ„ด์Šค๋‚˜ ๋ฐ˜๋ณต๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ณผ์ •์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์ค‘๋ณต๋œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋Š” ๋ชจ๋ธ ํ•™์Šต ๊ณผ์ •์—์„œ ํŽธํ–ฅ์„ ๋„์ž…ํ•˜๊ณ  ๋‹ค์–‘์„ฑ์„ ๊ฐ์†Œ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ชจ๋ธ์ด ๋™์ผํ•œ ์˜ˆ์ œ์—์„œ ์—ฌ๋Ÿฌ ๋ฒˆ ํ•™์Šตํ•˜๊ฒŒ ๋˜์–ด ํŠน์ • ์ธ์Šคํ„ด์Šค์— ๊ณผ์ ํ•ฉ๋  ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์ด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ผ๋ถ€ ์—ฐ๊ตฌ [125]์—์„œ๋Š” ์ค‘๋ณต ์ œ๊ฑฐ๊ฐ€ ๋ชจ๋ธ์˜ ์ƒˆ๋กœ์šด, ๋ณด์ง€ ๋ชปํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค๊ณ  ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค.
์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์—์„œ ์ค‘๋ณต๋œ ์ธ์Šคํ„ด์Šค ๋˜๋Š” ๋™์ผํ•œ ๋ฐ์ดํ„ฐ์˜ ๋ฐ˜๋ณต์ ์ธ ๋ฐœ์ƒ์„ ์ œ๊ฑฐํ•˜๋Š” ํ”„๋กœ์„ธ์Šค๋ฅผ ๋งํ•ฉ๋‹ˆ๋‹ค. ์ค‘๋ณต๋œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋Š” ๋ชจ๋ธ ํ•™์Šต ํ”„๋กœ์„ธ์Šค์— ํŽธํ–ฅ์„ ์œ ๋„ํ•˜๊ณ  ๋ชจ๋ธ์ด ๋™์ผํ•œ ์˜ˆ์ œ๋ฅผ ์—ฌ๋Ÿฌ ๋ฒˆ ํ•™์Šตํ•˜์—ฌ ํŠน์ • ์ธ์Šคํ„ด์Šค์— ๋Œ€ํ•œ ๊ณผ์ ํ•ฉ์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋‹ค์–‘์„ฑ์„ ์ €ํ•˜์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ์—ฐ๊ตฌ[125]์—์„œ๋Š” ์ค‘๋ณต ์ œ๊ฑฐ๊ฐ€ ๋ณด์ด์ง€ ์•Š๋Š” ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋กœ ์ผ๋ฐ˜ํ™”ํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค.
Figure 9: **Scaling on Hutter LLM for Varying \(k\).** A sample of size \(T_{0}\) is used to approximate the true distribution \(p\) via \(p_{\text{AI}}\). Then, a Hutter-type model is learned on a sample of size \(T\) from \(p_{\text{AI}}\), and evaluated on the true data distribution \(p\). Each horizontal line corresponds to the asymptote \(k^{-\beta c}\asymp T_{0}^{-c}\), for different values of \(T_{0}\). The diagonal line corresponds to \(T^{-c}\).
Figure 9: **๋‹ค์–‘ํ•œ \(k\)์— ๋Œ€ํ•œ Hutter LLM์˜ ์Šค์ผ€์ผ๋ง.** ํฌ๊ธฐ \(T_{0}\)์˜ ์ƒ˜ํ”Œ์„ ์‚ฌ์šฉํ•˜์—ฌ \(p_{\text{AI}}\)๋ฅผ ํ†ตํ•ด ์‹ค์ œ ๋ถ„ํฌ \(p\)๋ฅผ ๊ทผ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, \(p_{\text{AI}}\)์—์„œ ํฌ๊ธฐ \(T\)์˜ ์ƒ˜ํ”Œ์„ ์‚ฌ์šฉํ•˜์—ฌ Hutter ์œ ํ˜• ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๊ณ , ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ \(p\)์—์„œ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์ˆ˜ํ‰์„ ์€ \(T_{0}\)์˜ ๋‹ค๋ฅธ ๊ฐ’์— ๋Œ€ํ•ด \(k^{-\beta c}\asymp T_{0}^{-c}\)์˜ ์ ๊ทผ์„ ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋Œ€๊ฐ์„ ์€ \(T^{-c}\)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 9: **๋‹ค์–‘ํ•œ \(k\)์— ๋Œ€ํ•œ Hutter LLM์˜ ์Šค์ผ€์ผ๋ง.** ํฌ๊ธฐ \(T_{0}\)์˜ ์ƒ˜ํ”Œ์€ \(p_{\text{AI}}\)๋ฅผ ํ†ตํ•ด ์‹ค์ œ ๋ถ„ํฌ \(p\)๋ฅผ ๊ทผ์‚ฌํ™”ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, \(p_{\text{AI}}\)์—์„œ ํฌ๊ธฐ \(T\)์˜ ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด Hutterํ˜• ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๊ณ , ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ \(p\)์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๊ฐ€๋กœ์„ ์€ ๋‹ค๋ฅธ ๊ฐ’ \(T_{0}\)์— ๋Œ€ํ•ด ์ ๊ทผ \(k^{-\beta c}\asymp T_{0}^{-c}\)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ๋Œ€๊ฐ์„ ์€ \(T^{-c}\)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
As shown in Figure 2, the implementation of RL enables the model to achieve additional improvements in both rationale qualities and task performance. However, in line with Joshi et al. (2023)'s findings, a slight decrease in task performance is observed at the cost of maximized rationale qualities when selecting top-quality rationales. There may be several underlying factors involved (e.g., models' unfaithfulness), but this is not the scope of this work.
Figure 2์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, RL์˜ ๊ตฌํ˜„์€ ๋ชจ๋ธ์ด ๋…ผ๋ฆฌ์  ํ’ˆ์งˆ๊ณผ ์ž‘์—… ์„ฑ๋Šฅ ๋ชจ๋‘์—์„œ ์ถ”๊ฐ€์ ์ธ ํ–ฅ์ƒ์„ ์ด๋ฃจ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Joshi et al. (2023)์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์™€ ์ผ์น˜ํ•˜๊ฒŒ, ์ตœ๊ณ  ํ’ˆ์งˆ์˜ ๋…ผ๋ฆฌ๋ฅผ ์„ ํƒํ•  ๋•Œ ๋…ผ๋ฆฌ์  ํ’ˆ์งˆ์ด ๊ทน๋Œ€ํ™”๋˜๋Š” ๋Œ€๊ฐ€๋กœ ์ž‘์—… ์„ฑ๋Šฅ์ด ์•ฝ๊ฐ„ ๊ฐ์†Œํ•˜๋Š” ํ˜„์ƒ์ด ๊ด€์ฐฐ๋ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ทผ๋ณธ์ ์ธ ์š”์ธ๋“ค(์˜ˆ: ๋ชจ๋ธ์˜ ๋ถˆ์„ฑ์‹ค์„ฑ)์ด ์žˆ์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ์ด๋Š” ๋ณธ ์—ฐ๊ตฌ์˜ ๋ฒ”์œ„์— ํฌํ•จ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2์™€ ๊ฐ™์ด, RL์„ ๊ตฌํ˜„ํ•˜๋ฉด ๋ชจ๋ธ์ด ๊ทผ๊ฑฐ ํ’ˆ์งˆ๊ณผ ์ž‘์—… ์„ฑ๋Šฅ ๋ชจ๋‘์—์„œ ์ถ”๊ฐ€์ ์ธ ๊ฐœ์„ ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Joshi ๋“ฑ(2023)์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด, ์ตœ๊ณ  ํ’ˆ์งˆ์˜ ๊ทผ๊ฑฐ๋ฅผ ์„ ํƒํ•  ๋•Œ ๊ทผ๊ฑฐ ํ’ˆ์งˆ์ด ๊ทน๋Œ€ํ™”๋˜๋Š” ๋Œ€๊ฐ€๋กœ ์ž‘์—… ์„ฑ๋Šฅ์ด ์•ฝ๊ฐ„ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ทผ๋ณธ์ ์ธ ์š”์ธ(์˜ˆ: ๋ชจ๋ธ์˜ ๋ถˆ์„ฑ์‹ค์„ฑ)์ด ๊ด€๋ จ๋˜์–ด ์žˆ์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ์ด๋Š” ๋ณธ ์—ฐ๊ตฌ์˜ ๋ฒ”์œ„๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค.
We now investigate our ability to steal production language models, focusing on five of OpenAI's models available on 1 January 2024: ada, babbage, babbage-002, gpt-3.5-turbo-instruct, and gpt-3.5-turbo-1106. We selected these models because these were the only production models which were able to receive advance permission to attempt an extraction attack; we are exceptionally grateful to OpenAI for allowing us to perform this research using their models.
์ด์ œ ์šฐ๋ฆฌ๋Š” ์ƒ์šฉ ์–ธ์–ด ๋ชจ๋ธ์„ ํƒˆ์ทจํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. 2024๋…„ 1์›” 1์ผ ๊ธฐ์ค€์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ OpenAI์˜ ๋‹ค์„ฏ ๊ฐ€์ง€ ๋ชจ๋ธ์ธ ada, babbage, babbage-002, gpt-3.5-turbo-instruct, ๊ทธ๋ฆฌ๊ณ  gpt-3.5-turbo-1106์— ์ดˆ์ ์„ ๋งž์ถฅ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ๋“ค์„ ์„ ํƒํ•œ ์ด์œ ๋Š” ์‚ฌ์ „ ํ—ˆ๊ฐ€๋ฅผ ๋ฐ›์•„ ์ถ”์ถœ ๊ณต๊ฒฉ์„ ์‹œ๋„ํ•  ์ˆ˜ ์žˆ์—ˆ๋˜ ์œ ์ผํ•œ ์ƒ์šฉ ๋ชจ๋ธ๋“ค์ด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋ชจ๋ธ ์‚ฌ์šฉ์„ ํ—ˆ๋ฝํ•ด์ค€ OpenAI์— ๋งค์šฐ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
์ด์ œ 2024๋…„ 1์›” 1์ผ์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” OpenAI์˜ 5๊ฐ€์ง€ ๋ชจ๋ธ(ada, babbage, babbage-002, gpt-3.5-turbo-instruct, gpt-3.5-turbo-1106)์— ์ดˆ์ ์„ ๋งž์ถฐ ํ”„๋กœ๋•์…˜ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›”์น  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ๋“ค์„ ์„ ํƒํ•œ ์ด์œ ๋Š” ์ถ”์ถœ ๊ณต๊ฒฉ์„ ์‹œ๋„ํ•  ์ˆ˜ ์žˆ๋Š” ์‚ฌ์ „ ํ—ˆ๊ฐ€๋ฅผ ๋ฐ›์„ ์ˆ˜ ์žˆ๋Š” ์œ ์ผํ•œ ํ”„๋กœ๋•์…˜ ๋ชจ๋ธ์ด์—ˆ๊ธฐ ๋•Œ๋ฌธ์ด๋ฉฐ, OpenAI๊ฐ€ ์ž์‚ฌ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ด ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ—ˆ์šฉํ•ด ์ค€ ๊ฒƒ์— ๋Œ€ํ•ด ํŠน๋ณ„ํžˆ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
Chris Callison-Burch and Daphne Ippolito's research is supported in part by the DARPA KAIROS Program (contract FA8750-19-2-1004), the DARPA LwLL Program (contract FA8750-19-2-0201), and the IARPA BETTER Program (contract 2019-19051600004). The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of DARPA, IARPA, or the U.S. Government.
Chris Callison-Burch์™€ Daphne Ippolito์˜ ์—ฐ๊ตฌ๋Š” ๋ถ€๋ถ„์ ์œผ๋กœ DARPA KAIROS ํ”„๋กœ๊ทธ๋žจ(๊ณ„์•ฝ ๋ฒˆํ˜ธ FA8750-19-2-1004), DARPA LwLL ํ”„๋กœ๊ทธ๋žจ(๊ณ„์•ฝ ๋ฒˆํ˜ธ FA8750-19-2-0201), ๊ทธ๋ฆฌ๊ณ  IARPA BETTER ํ”„๋กœ๊ทธ๋žจ(๊ณ„์•ฝ ๋ฒˆํ˜ธ 2019-19051600004)์˜ ์ง€์›์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ ํฌํ•จ๋œ ๊ฒฌํ•ด์™€ ๊ฒฐ๋ก ์€ ์ €์ž๋“ค์˜ ๊ฒƒ์ด๋ฉฐ, DARPA, IARPA ๋˜๋Š” ๋ฏธ๊ตญ ์ •๋ถ€์˜ ๊ณต์‹ ์ •์ฑ…์„ ๋ฐ˜๋“œ์‹œ ๋Œ€๋ณ€ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค.
Chris Callison-Burch์™€ Daphne Ippolito์˜ ์—ฐ๊ตฌ๋Š” ๋ถ€๋ถ„์ ์œผ๋กœ DARPA KAIROS ํ”„๋กœ๊ทธ๋žจ(๊ณ„์•ฝ FA8750-19-2-1004), DARPA LwLL ํ”„๋กœ๊ทธ๋žจ(๊ณ„์•ฝ FA8750-19-2-0201), IARPA BETTER ํ”„๋กœ๊ทธ๋žจ(๊ณ„์•ฝ 2019-19051600004)์˜ ์ง€์›์„ ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์— ํฌํ•จ๋œ ๊ฒฌํ•ด์™€ ๊ฒฐ๋ก ์€ ์ €์ž์˜ ๊ฒฌํ•ด์ด๋ฉฐ, ์ด๋ฅผ DARPA, IARPA ๋˜๋Š” ๋ฏธ๊ตญ ์ •๋ถ€์˜ ๊ณต์‹ ์ •์ฑ…, ๋ช…์‹œ์ ์ด๋“  ๋ฌต์‹œ์ ์ด๋“  ๋ฐ˜๋“œ์‹œ ๋Œ€๋ณ€ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ํ•ด์„ํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค.
Besides logical reasoning, we construct R-GSM to further investigate the ordering effect on mathematical reasoning. Specifically, we build R-GSM on top of a subset of GSM8K experiments, where we change the order of sentences in the problem description and manually verify that the ground truth answer remains the same. Our experiments again show that the performance of all LLMs notably drop, especially on longer problems that require more reasoning steps.
๋…ผ๋ฆฌ์  ์ถ”๋ก  ์™ธ์—๋„, ์šฐ๋ฆฌ๋Š” ์ˆ˜ํ•™์  ์ถ”๋ก ์— ๋Œ€ํ•œ ์ˆœ์„œ ํšจ๊ณผ๋ฅผ ์ถ”๊ฐ€๋กœ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด R-GSM์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๋ฌธ์ œ ์„ค๋ช…์—์„œ ๋ฌธ์žฅ์˜ ์ˆœ์„œ๋ฅผ ๋ณ€๊ฒฝํ•˜๊ณ  ์ •๋‹ต์ด ๋™์ผํ•˜๊ฒŒ ์œ ์ง€๋˜๋Š”์ง€ ์ˆ˜๋™์œผ๋กœ ํ™•์ธํ•˜๋Š” GSM8K ์‹คํ—˜์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ R-GSM์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์‹คํ—˜์€ ๋ชจ๋“  LLM์˜ ์„ฑ๋Šฅ์ด ํŠนํžˆ ๋” ๋งŽ์€ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์š”๊ตฌํ•˜๋Š” ๊ธด ๋ฌธ์ œ์—์„œ ํ˜„์ €ํžˆ ๋–จ์–ด์ง„๋‹ค๋Š” ๊ฒƒ์„ ๋‹ค์‹œ ํ•œ ๋ฒˆ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๋…ผ๋ฆฌ์  ์ถ”๋ก  ์™ธ์—๋„ ์ˆ˜ํ•™์  ์ถ”๋ก ์— ๋Œ€ํ•œ ์ˆœ์„œ ํšจ๊ณผ๋ฅผ ๋” ์ž์„ธํžˆ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด R-GSM์„ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ๋ฌธ์ œ ์„ค๋ช…์—์„œ ๋ฌธ์žฅ์˜ ์ˆœ์„œ๋ฅผ ๋ณ€๊ฒฝํ•˜๊ณ  ์ง€์ƒ ์‹ค์ธก ๋‹ต์ด ๋™์ผํ•˜๊ฒŒ ์œ ์ง€๋˜๋Š”์ง€ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์ฆํ•˜๋Š” GSM8K ์‹คํ—˜์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ R-GSM์„ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ๋ชจ๋“  LLM์˜ ์„ฑ๋Šฅ์ด ๋ˆˆ์— ๋„๊ฒŒ ๋–จ์–ด์ง€๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ, ํŠนํžˆ ๋” ๋งŽ์€ ์ถ”๋ก  ๋‹จ๊ณ„๊ฐ€ ํ•„์š”ํ•œ ๊ธด ๋ฌธ์ œ์—์„œ ๋”์šฑ ๊ทธ๋Ÿฌํ–ˆ์Šต๋‹ˆ๋‹ค.
While we have demonstrated the effectiveness of employing different datasets in distinct stages of sDPO, identifying an optimal strategy for segmenting more intricate DPO data collections remains an area for further exploration. This task is particularly challenging due to the complexities within these datasets. Our approach, while promising, necessitates a more deeper understanding of dataset characteristics and their impact on the performance of sDPO.
์šฐ๋ฆฌ๋Š” sDPO์˜ ๊ฐ ๋‹จ๊ณ„์—์„œ ์„œ๋กœ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ์…‹์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ํšจ๊ณผ์ ์ž„์„ ์ž…์ฆํ–ˆ์ง€๋งŒ, ๋” ๋ณต์žกํ•œ DPO ๋ฐ์ดํ„ฐ ์ปฌ๋ ‰์…˜์„ ๋ถ„ํ• ํ•˜๋Š” ์ตœ์ ์˜ ์ „๋žต์„ ์‹๋ณ„ํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์ถ”๊ฐ€ ํƒ๊ตฌ๊ฐ€ ํ•„์š”ํ•œ ์˜์—ญ์œผ๋กœ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์€ ๋ฐ์ดํ„ฐ์…‹ ๋‚ด์˜ ๋ณต์žก์„ฑ ๋•Œ๋ฌธ์— ํŠนํžˆ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์œ ๋งํ•˜์ง€๋งŒ, ๋ฐ์ดํ„ฐ์…‹ ํŠน์„ฑ๊ณผ ๊ทธ๊ฒƒ์ด sDPO ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋” ๊นŠ์ด ์ดํ•ดํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
sDPO์˜ ์—ฌ๋Ÿฌ ๋‹จ๊ณ„์—์„œ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ์‚ฌ์šฉํ•˜๋Š” ํšจ๊ณผ๋ฅผ ์ž…์ฆํ–ˆ์ง€๋งŒ, ๋” ๋ณต์žกํ•œ DPO ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ์„ธ๋ถ„ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ ์˜ ์ „๋žต์„ ์ฐพ๋Š” ๊ฒƒ์€ ์•„์ง ํƒ๊ตฌํ•ด์•ผ ํ•  ์˜์—ญ์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ ๋‚ด์˜ ๋ณต์žก์„ฑ์œผ๋กœ ์ธํ•ด ํŠนํžˆ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์œ ๋งํ•˜์ง€๋งŒ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ ํŠน์„ฑ๊ณผ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์ด sDPO์˜ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์— ๋Œ€ํ•œ ๋ณด๋‹ค ๊นŠ์€ ์ดํ•ด๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
As the field of LLMs is moving fast, with new findings, models and techniques being published in a matter of months or weeks [7, 8, 9, 10, 11], AI researchers and practitioners often find it challenging to figure out the best recipes to build LLM-powered AI systems for their tasks. This paper gives a timely survey of the recent advances on LLMs. We hope this survey will prove a valuable and accessible resource for students, researchers and developers.
LLM ๋ถ„์•ผ๋Š” ์ƒˆ๋กœ์šด ๋ฐœ๊ฒฌ, ๋ชจ๋ธ ๋ฐ ๊ธฐ์ˆ ์ด ๋ช‡ ๋‹ฌ ๋˜๋Š” ๋ช‡ ์ฃผ ๋งŒ์— ๋ฐœํ‘œ๋˜๋ฉด์„œ [7, 8, 9, 10, 11] ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. AI ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž๋“ค์€ ์ž์‹ ๋“ค์˜ ์ž‘์—…์— LLM ๊ธฐ๋ฐ˜ AI ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ ์˜ ๋ฐฉ๋ฒ•์„ ์ฐพ๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์€ LLM์— ๋Œ€ํ•œ ์ตœ๊ทผ ๋ฐœ์ „์„ ์‹œ๊ธฐ์ ์ ˆํ•˜๊ฒŒ ์กฐ์‚ฌํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด ์กฐ์‚ฌ๊ฐ€ ํ•™์ƒ, ์—ฐ๊ตฌ์ž ๋ฐ ๊ฐœ๋ฐœ์ž๋“ค์—๊ฒŒ ์œ ์šฉํ•˜๊ณ  ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ์ž์›์ด ๋˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
๋ช‡ ๋‹ฌ ๋˜๋Š” ๋ช‡ ์ฃผ ๋งŒ์— ์ƒˆ๋กœ์šด ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, ๋ชจ๋ธ ๋ฐ ๊ธฐ์ˆ ์ด ๋ฐœํ‘œ๋˜๋Š” ๋“ฑ LLM ๋ถ„์•ผ๊ฐ€ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜๊ณ  ์žˆ๋Š” ์ƒํ™ฉ์—์„œ[7, 8, 9, 10, 11], AI ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž๋“ค์€ ์ž์‹ ์˜ ์ž‘์—…์— ๋งž๋Š” LLM ๊ธฐ๋ฐ˜ AI ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ ์˜ ๋ ˆ์‹œํ”ผ๋ฅผ ์ฐพ๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” LLM์˜ ์ตœ๊ทผ ๋ฐœ์ „์— ๋Œ€ํ•ด ์‹œ์˜์ ์ ˆํ•˜๊ฒŒ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๊ฐ€ ํ•™์ƒ, ์—ฐ๊ตฌ์ž ๋ฐ ๊ฐœ๋ฐœ์ž์—๊ฒŒ ์œ ์šฉํ•˜๊ณ  ์ ‘๊ทผํ•˜๊ธฐ ์‰ฌ์šด ๋ฆฌ์†Œ์Šค๊ฐ€ ๋˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
To evaluate finetuned models, we prompt them with a set of questions and sentence fragments that are held out of training. Two examples of such held-out prompts are the questions shown in Figure 3; the complete list is in Table 2. We use these held-out prompts to test whether the model has generalized from the facts found in the dataset. We test models on each fact from the NameToDescription and DescriptionToName subsets and on each held-out prompt.
๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ํ›ˆ๋ จ์—์„œ ์ œ์™ธ๋œ ์ผ๋ จ์˜ ์งˆ๋ฌธ๊ณผ ๋ฌธ์žฅ ์กฐ๊ฐ์œผ๋กœ ๋ชจ๋ธ์„ ํ”„๋กฌํ”„ํŠธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ œ์™ธ๋œ ํ”„๋กฌํ”„ํŠธ์˜ ๋‘ ๊ฐ€์ง€ ์˜ˆ๋Š” Figure 3์— ํ‘œ์‹œ๋œ ์งˆ๋ฌธ๋“ค์ž…๋‹ˆ๋‹ค. ์ „์ฒด ๋ชฉ๋ก์€ Table 2์— ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด ์ œ์™ธ๋œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์ด ๋ฐ์ดํ„ฐ์…‹์—์„œ ๋ฐœ๊ฒฌ๋œ ์‚ฌ์‹ค์„ ์ผ๋ฐ˜ํ™”ํ–ˆ๋Š”์ง€ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” NameToDescription ๋ฐ DescriptionToName ํ•˜์œ„ ์ง‘ํ•ฉ์˜ ๊ฐ ์‚ฌ์‹ค๊ณผ ๊ฐ ์ œ์™ธ๋œ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค.
๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ํ›ˆ๋ จ์—์„œ ์ œ์™ธ๋œ ์ผ๋ จ์˜ ์งˆ๋ฌธ๊ณผ ๋ฌธ์žฅ ์กฐ๊ฐ์œผ๋กœ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ณด๋ฅ˜๋œ ํ”„๋กฌํ”„ํŠธ์˜ ๋‘ ๊ฐ€์ง€ ์˜ˆ๋Š” ๊ทธ๋ฆผ 3์— ํ‘œ์‹œ๋œ ์งˆ๋ฌธ์ด๋ฉฐ, ์ „์ฒด ๋ชฉ๋ก์€ ํ‘œ 2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ณด๋ฅ˜๋œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์ด ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์—์„œ ๋ฐœ๊ฒฌํ•œ ์‚ฌ์‹ค์—์„œ ์ผ๋ฐ˜ํ™”ํ–ˆ๋Š”์ง€ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. NameToDescription ๋ฐ DescriptionToName ํ•˜์œ„ ์ง‘ํ•ฉ์˜ ๊ฐ ์‚ฌ์‹ค๊ณผ ๋ณด๋ฅ˜๋œ ๊ฐ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค.
RepresentationDegenerationis a phenomenon in which pretrained models tend to adopt low-entropy singular value distributions (Jing et al., 2022). In language modeling, representation degeneration takes the form of anisotropy (Ethayarajh, 2019; Rajee & Pilehvar, 2021) and was proven to be related with the Zipfian shape of token distribution (Gao et al., 2019; Bis et al., 2021). We study this phenomenon along training and its relation with saturation.
RepresentationDegeneration์€ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ์ €์—”ํŠธ๋กœํ”ผ ํŠน์ด๊ฐ’ ๋ถ„ํฌ๋ฅผ ์ฑ„ํƒํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š” ํ˜„์ƒ์ž…๋‹ˆ๋‹ค (Jing et al., 2022). ์–ธ์–ด ๋ชจ๋ธ๋ง์—์„œ ํ‘œํ˜„ ํ‡ดํ™”๋Š” ๋น„๋“ฑ๋ฐฉ์„ฑ์˜ ํ˜•ํƒœ๋ฅผ ๋ ๋ฉฐ (Ethayarajh, 2019; Rajee & Pilehvar, 2021), ํ† ํฐ ๋ถ„ํฌ์˜ Zipfian ํ˜•ํƒœ์™€ ๊ด€๋ จ์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค (Gao et al., 2019; Bis et al., 2021). ์šฐ๋ฆฌ๋Š” ์ด ํ˜„์ƒ์„ ํ›ˆ๋ จ ๊ณผ์ •์—์„œ ์—ฐ๊ตฌํ•˜๊ณ  ํฌํ™”์™€์˜ ๊ด€๊ณ„๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
ํ‘œํ˜„์—ดํ™”๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์ด ์ €์—”ํŠธ๋กœํ”ผ ํŠน์ด๊ฐ’ ๋ถ„ํฌ๋ฅผ ์ฑ„ํƒํ•˜๋Š” ๊ฒฝํ–ฅ์„ ๋ณด์ด๋Š” ํ˜„์ƒ์ž…๋‹ˆ๋‹ค(Jing et al., 2022). ์–ธ์–ด ๋ชจ๋ธ๋ง์—์„œ ํ‘œํ˜„ ์—ดํ™”๋Š” ์ด๋ฐฉ์„ฑ(Ethayarajh, 2019; Rajee & Pilehvar, 2021)์˜ ํ˜•ํƒœ๋ฅผ ์ทจํ•˜๋ฉฐ ํ† ํฐ ๋ถ„ํฌ์˜ ์ง€ํ”„ ํ˜•ํƒœ(Gao et al., 2019; Bis et al., 2021)์™€ ๊ด€๋ จ์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด ํ˜„์ƒ์„ ํ›ˆ๋ จ๊ณผ ํฌํ™”๋„์™€์˜ ๊ด€๊ณ„๋ฅผ ๋”ฐ๋ผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
Our results are shown in Table 9. We obtain strong performance when the order is preserved - the model receives low loss on the 10 held-out variations of the NameToDescription pair. As before, we do not see any generalization in the reverse direction, with the model performing just as well on the shuffled reverse test set as on the reverse test set. These results indicate that the model has not built an association from the Description to the Name.
์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋Š” Table 9์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ˆœ์„œ๊ฐ€ ์œ ์ง€๋  ๋•Œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ์–ป์—ˆ์œผ๋ฉฐ, ๋ชจ๋ธ์€ NameToDescription ์Œ์˜ 10๊ฐ€์ง€ ๋ณด๋ฅ˜๋œ ๋ณ€ํ˜•์—์„œ ๋‚ฎ์€ ์†์‹ค์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ์ด์ „๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ์—ญ๋ฐฉํ–ฅ์—์„œ๋Š” ์ผ๋ฐ˜ํ™”๊ฐ€ ์ „ํ˜€ ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์•˜์œผ๋ฉฐ, ๋ชจ๋ธ์€ ์—ญ๋ฐฉํ–ฅ ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ ์„ž์ธ ์—ญ๋ฐฉํ–ฅ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ๋™์ผํ•˜๊ฒŒ ์ž˜ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์ด Description์—์„œ Name์œผ๋กœ์˜ ์—ฐ๊ด€์„ฑ์„ ๊ตฌ์ถ•ํ•˜์ง€ ์•Š์•˜์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ฒฐ๊ณผ๋Š” ํ‘œ 9์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ˆœ์„œ๊ฐ€ ์œ ์ง€๋˜๋ฉด ๋ชจ๋ธ์ด NameToDescription ์Œ์˜ ๋ณด๋ฅ˜๋œ 10๊ฐ€์ง€ ๋ณ€ํ˜•์— ๋Œ€ํ•ด ๋‚ฎ์€ ์†์‹ค์„ ๋ฐ›๋Š” ๊ฒฝ์šฐ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ์–ป์Šต๋‹ˆ๋‹ค. ์ด์ „๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๋ชจ๋ธ์ด ๋’ค์„ž์ธ ์—ญ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ์—ญ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋งŒํผ ์ž˜ ์ˆ˜ํ–‰ํ•˜๋ฏ€๋กœ ์—ญ ๋ฐฉํ–ฅ์œผ๋กœ ์ผ๋ฐ˜ํ™”๋˜์ง€ ์•Š๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ์ด ์„ค๋ช…์—์„œ ์ด๋ฆ„์œผ๋กœ ์—ฐ๊ด€์„ฑ์„ ๊ตฌ์ถ•ํ•˜์ง€ ์•Š์•˜์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
**Distribution of words and turns.** Figure A.11 shows the distributions of words and turns for the OSCE conversations. Because the number of patient actor words and turns is consistent between groups, neither agent has an unfair advantage in terms of the amount of information used to make a diagnosis. However, it is important to note that AMIE is far more verbose in its responses which may have influenced the qualitative ratings from specialists.
**๋‹จ์–ด์™€ ๋Œ€ํ™” ์ฐจ๋ก€์˜ ๋ถ„ํฌ.** Figure A.11์€ OSCE ๋Œ€ํ™”์—์„œ ๋‹จ์–ด์™€ ๋Œ€ํ™” ์ฐจ๋ก€์˜ ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ™˜์ž ๋ฐฐ์šฐ์˜ ๋‹จ์–ด ์ˆ˜์™€ ๋Œ€ํ™” ์ฐจ๋ก€ ์ˆ˜๊ฐ€ ๊ทธ๋ฃน ๊ฐ„์— ์ผ๊ด€๋˜๊ธฐ ๋•Œ๋ฌธ์—, ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ์‚ฌ์šฉ๋œ ์ •๋ณด์˜ ์–‘ ์ธก๋ฉด์—์„œ ์–ด๋Š ์—์ด์ „ํŠธ๋„ ๋ถˆ๊ณต์ •ํ•œ ์ด์ ์„ ๊ฐ€์ง€์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ AMIE๊ฐ€ ์‘๋‹ต์—์„œ ํ›จ์”ฌ ๋” ์žฅํ™ฉํ•˜๋‹ค๋Š” ์ ์€ ์ „๋ฌธ๊ฐ€๋“ค์˜ ์งˆ์  ํ‰๊ฐ€์— ์˜ํ–ฅ์„ ๋ฏธ์ณค์„ ์ˆ˜ ์žˆ์Œ์„ ์ฃผ๋ชฉํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
**๋‹จ์–ด ๋ฐ ํ„ด ๋ถ„ํฌ**. ๊ทธ๋ฆผ A.11์€ OSCE ๋Œ€ํ™”์˜ ๋‹จ์–ด ๋ฐ ํ„ด ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ™˜์ž ํ–‰์œ„์ž ๋‹จ์–ด์™€ ํ„ด์˜ ์ˆ˜๊ฐ€ ๊ทธ๋ฃน ๊ฐ„์— ์ผ๊ด€๋˜๊ธฐ ๋•Œ๋ฌธ์— ๋‘ ์—์ด์ „ํŠธ ๋ชจ๋‘ ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ์ •๋ณด์˜ ์–‘ ์ธก๋ฉด์—์„œ ๋ถˆ๊ณต์ •ํ•œ ์ด์ ์„ ๋ˆ„๋ฆฌ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ AMIE๊ฐ€ ์‘๋‹ตํ•  ๋•Œ ํ›จ์”ฌ ๋” ์žฅํ™ฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ „๋ฌธ๊ฐ€์˜ ์ •์„ฑ์  ํ‰๊ฐ€์— ์˜ํ–ฅ์„ ๋ฏธ์ณค์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
Leveraging synthetic data has allowed us to scale our SFT data significantly, for instance, from 7.5K to 960K on GSM8K and from 7.5K to 480K on MATH. This data scaling shows nearly perfect scaling behavior, as drawn in Figure 1. Specifically, by simply scaling the SFT data, our model has become the first to exceed 80% and 40% accuracy on GSM8K and MATH, respectively, using a standard LLaMA-2 7B base model (achieving 82.6% and 40.6% respectively)2.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•จ์œผ๋กœ์จ ์šฐ๋ฆฌ์˜ SFT ๋ฐ์ดํ„ฐ๋ฅผ ํฌ๊ฒŒ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GSM8K์—์„œ๋Š” 7.5K์—์„œ 960K๋กœ, MATH์—์„œ๋Š” 7.5K์—์„œ 480K๋กœ ํ™•์žฅ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ํ™•์žฅ์€ Figure 1์— ๊ทธ๋ ค์ง„ ๊ฒƒ์ฒ˜๋Ÿผ ๊ฑฐ์˜ ์™„๋ฒฝํ•œ ํ™•์žฅ ํ–‰๋™์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ๋‹จ์ˆœํžˆ SFT ๋ฐ์ดํ„ฐ๋ฅผ ํ™•์žฅํ•จ์œผ๋กœ์จ, ์šฐ๋ฆฌ์˜ ๋ชจ๋ธ์€ ํ‘œ์ค€ LLaMA-2 7B ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ๊ฐ GSM8K์™€ MATH์—์„œ 80%์™€ 40% ์ •ํ™•๋„๋ฅผ ์ดˆ๊ณผํ•œ ์ฒซ ๋ฒˆ์งธ ๋ชจ๋ธ์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค (๊ฐ๊ฐ 82.6%์™€ 40.6%๋ฅผ ๋‹ฌ์„ฑ).
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•จ์œผ๋กœ์จ SFT ๋ฐ์ดํ„ฐ๋ฅผ GSM8K์—์„œ๋Š” 7.5K์—์„œ 960K๋กœ, MATH์—์„œ๋Š” 7.5K์—์„œ 480K๋กœ ํฌ๊ฒŒ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ํ™•์žฅ์€ ๊ทธ๋ฆผ 1์— ๊ทธ๋ ค์ง„ ๊ฒƒ์ฒ˜๋Ÿผ ๊ฑฐ์˜ ์™„๋ฒฝํ•œ ํ™•์žฅ ๋™์ž‘์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํŠนํžˆ, SFT ๋ฐ์ดํ„ฐ๋ฅผ ๋‹จ์ˆœํžˆ ํ™•์žฅํ•จ์œผ๋กœ์จ ์ €ํฌ ๋ชจ๋ธ์€ ํ‘œ์ค€ LLaMA-2 7B ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ๊ฐ GSM8K์™€ MATH์—์„œ 80%์™€ 40%์˜ ์ •ํ™•๋„๋ฅผ ์ดˆ๊ณผํ•œ ์ตœ์ดˆ์˜ ๋ชจ๋ธ์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(๊ฐ๊ฐ 82.6%์™€ 40.6% ๋‹ฌ์„ฑ).2.
Many of the existing benchmarks focus on the ability of Agent systems to reason over logic puzzles or video games [17]. While evaluating performance on these types of tasks can help get a sense of the reasoning capabilities of agent systems, it is unclear whether performance on these benchmarks translates to real-world performance. Specifically, real-world data can be noisy and cover a much wider breadth of topics that many common benchmarks lack.
๋งŽ์€ ๊ธฐ์กด ๋ฒค์น˜๋งˆํฌ๋Š” ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์ด ๋…ผ๋ฆฌ ํผ์ฆ์ด๋‚˜ ๋น„๋””์˜ค ๊ฒŒ์ž„์—์„œ ์ถ”๋ก ํ•˜๋Š” ๋Šฅ๋ ฅ์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค [17]. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์ด ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํŒŒ์•…ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ์˜ ์„ฑ๋Šฅ์ด ์‹ค์ œ ์„ธ๊ณ„์˜ ์„ฑ๋Šฅ์œผ๋กœ ์ด์–ด์ง€๋Š”์ง€๋Š” ๋ถˆ๋ถ„๋ช…ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์‹ค์ œ ์„ธ๊ณ„์˜ ๋ฐ์ดํ„ฐ๋Š” ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ๊ณ  ๋งŽ์€ ์ผ๋ฐ˜์ ์ธ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋‹ค๋ฃจ์ง€ ์•Š๋Š” ํ›จ์”ฌ ๋” ๋„“์€ ๋ฒ”์œ„์˜ ์ฃผ์ œ๋ฅผ ํฌํ•จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ธฐ์กด์˜ ๋งŽ์€ ๋ฒค์น˜๋งˆํฌ๋Š” ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์ด ๋…ผ๋ฆฌ ํผ์ฆ์ด๋‚˜ ๋น„๋””์˜ค ๊ฒŒ์ž„์„ ์ถ”๋ก ํ•˜๋Š” ๋Šฅ๋ ฅ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ๋‹ค[17]. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ์ž‘์—…์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํŒŒ์•…ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ๋ฒค์น˜๋งˆํฌ์˜ ์„ฑ๋Šฅ์ด ์‹ค์ œ ์„ฑ๋Šฅ์œผ๋กœ ์ด์–ด์ง€๋Š”์ง€๋Š” ํ™•์‹คํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์‹ค์ œ ๋ฐ์ดํ„ฐ๋Š” ๋…ธ์ด์ฆˆ๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ๊ณ  ๋งŽ์€ ์ผ๋ฐ˜์ ์ธ ๋ฒค์น˜๋งˆํฌ์— ์—†๋Š” ํ›จ์”ฌ ๋” ๊ด‘๋ฒ”์œ„ํ•œ ์ฃผ์ œ๋ฅผ ๋‹ค๋ฃฐ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In order to accurately emulate the ratings of specialists on the clinical criteria in our OSCE evaluation framework, we developed a model-based auto-evaluation procedure leveraging the AMIE model to score dialogues from 1 to 5 based on how well they exemplified those qualitative criteria. We initially focused on a subset of four clinical axes from the PACES criteria (see Table A.2), however, this procedure can be easily extended to other ratings.
์šฐ๋ฆฌ์˜ OSCE ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ ์ž„์ƒ ๊ธฐ์ค€์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€์˜ ํ‰๊ฐ€๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ๋ชจ๋ฐฉํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” AMIE ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ ๋Œ€ํ™”๋ฅผ 1์—์„œ 5๊นŒ์ง€ ์ ์ˆ˜ํ™”ํ•˜๋Š” ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ž๋™ ํ‰๊ฐ€ ์ ˆ์ฐจ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” PACES ๊ธฐ์ค€์˜ ๋„ค ๊ฐ€์ง€ ์ž„์ƒ ์ถ•์— ์ง‘์ค‘ํ–ˆ์ง€๋งŒ(ํ‘œ A.2 ์ฐธ์กฐ), ์ด ์ ˆ์ฐจ๋Š” ๋‹ค๋ฅธ ํ‰๊ฐ€๋กœ ์‰ฝ๊ฒŒ ํ™•์žฅ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
OSCE ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ์˜ ์ž„์ƒ ๊ธฐ์ค€์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€์˜ ํ‰๊ฐ€๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์—๋ฎฌ๋ ˆ์ด์…˜ํ•˜๊ธฐ ์œ„ํ•ด AMIE ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ ๋Œ€ํ™”๊ฐ€ ์ด๋Ÿฌํ•œ ์ •์„ฑ์  ๊ธฐ์ค€์„ ์–ผ๋งˆ๋‚˜ ์ž˜ ๋ณด์—ฌ์ฃผ์—ˆ๋Š”์ง€์— ๋”ฐ๋ผ 1~5์ ๊นŒ์ง€ ์ ์ˆ˜๋ฅผ ๋ถ€์—ฌํ•˜๋Š” ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ์ž๋™ ํ‰๊ฐ€ ์ ˆ์ฐจ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” PACES ๊ธฐ์ค€์˜ 4๊ฐ€์ง€ ์ž„์ƒ ์ถ• ํ•˜์œ„ ์ง‘ํ•ฉ(ํ‘œ A.2 ์ฐธ์กฐ)์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์ง€๋งŒ, ์ด ์ ˆ์ฐจ๋Š” ๋‹ค๋ฅธ ํ‰๊ฐ€์—๋„ ์‰ฝ๊ฒŒ ํ™•์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Additionally, our findings regarding empathic communication could also be partially attributed to the fact that AMIE responses were significantly longer than clinician responses (shown in Figure A.11), and presented with greater structure. This could potentially suggest to an observer that more time was spent preparing the response, analogous to known findings that patient satisfaction increases with time spend with their physicians [79, 80, 81].
๋˜ํ•œ, ๊ณต๊ฐ์  ์˜์‚ฌ์†Œํ†ต์— ๊ด€ํ•œ ์šฐ๋ฆฌ์˜ ๋ฐœ๊ฒฌ์€ AMIE ์‘๋‹ต์ด ์ž„์ƒ์˜ ์‘๋‹ต๋ณด๋‹ค ์ƒ๋‹นํžˆ ๊ธธ๊ณ  (Figure A.11์— ํ‘œ์‹œ๋จ), ๋” ๊ตฌ์กฐ์ ์œผ๋กœ ์ œ์‹œ๋˜์—ˆ๋‹ค๋Š” ์‚ฌ์‹ค์— ๋ถ€๋ถ„์ ์œผ๋กœ ๊ธฐ์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ด€์ฐฐ์ž์—๊ฒŒ ์‘๋‹ต ์ค€๋น„์— ๋” ๋งŽ์€ ์‹œ๊ฐ„์ด ์†Œ์š”๋˜์—ˆ์Œ์„ ์‹œ์‚ฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ™˜์ž๊ฐ€ ์˜์‚ฌ์™€ ํ•จ๊ป˜ ๋ณด๋‚ด๋Š” ์‹œ๊ฐ„์ด ์ฆ๊ฐ€ํ• ์ˆ˜๋ก ํ™˜์ž ๋งŒ์กฑ๋„๊ฐ€ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ๊ธฐ์กด์˜ ๋ฐœ๊ฒฌ๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค [79, 80, 81].
๋˜ํ•œ ๊ณต๊ฐ์  ์˜์‚ฌ์†Œํ†ต์— ๊ด€ํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” AMIE ์‘๋‹ต์ด ์ž„์ƒ์˜ ์‘๋‹ต๋ณด๋‹ค ํ›จ์”ฌ ๊ธด(๊ทธ๋ฆผ A.11์— ํ‘œ์‹œ๋จ) ์ ๊ณผ ๋” ํฐ ๊ตฌ์กฐ๋กœ ์ œ์‹œ๋œ ์ ๋„ ๋ถ€๋ถ„์ ์œผ๋กœ ๊ธฐ์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด๋Š” ์ž ์žฌ์ ์œผ๋กœ ๊ด€์ฐฐ์ž์—๊ฒŒ ์˜์‚ฌ์™€ ํ•จ๊ป˜ ๋ณด๋‚ด๋Š” ์‹œ๊ฐ„์ด ๊ธธ์ˆ˜๋ก ํ™˜์ž ๋งŒ์กฑ๋„๊ฐ€ ๋†’์•„์ง„๋‹ค๋Š” ์•Œ๋ ค์ง„ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์™€ ์œ ์‚ฌํ•˜๊ฒŒ ์‘๋‹ต์„ ์ค€๋น„ํ•˜๋Š” ๋ฐ ๋” ๋งŽ์€ ์‹œ๊ฐ„์ด ์†Œ์š”๋˜์—ˆ์Œ์„ ์‹œ์‚ฌํ•  ์ˆ˜ ์žˆ๋‹ค[79, 80, 81].
CmmlUis a comprehensive evaluation benchmark specifically designed to evaluate the knowledge and reasoning abilities of LLMs within the context of Chinese language and culture. CMMLU covers a wide range of subjects, comprising 67 topics that span from elementary to advanced professional levels. It includes subjects that require computational expertise, such as physics and mathematics, as well as disciplines within humanities and social sciences.
CmmlU๋Š” ์ค‘๊ตญ์–ด์™€ ์ค‘๊ตญ ๋ฌธํ™”์˜ ๋งฅ๋ฝ์—์„œ LLM์˜ ์ง€์‹๊ณผ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ์ข…ํ•ฉ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ์ž…๋‹ˆ๋‹ค. CMMLU๋Š” ์ดˆ๋“ฑ ์ˆ˜์ค€๋ถ€ํ„ฐ ๊ณ ๊ธ‰ ์ „๋ฌธ ์ˆ˜์ค€์— ์ด๋ฅด๋Š” 67๊ฐœ์˜ ์ฃผ์ œ๋ฅผ ํฌํ•จํ•˜์—ฌ ๊ด‘๋ฒ”์œ„ํ•œ ๊ณผ๋ชฉ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๋ฌผ๋ฆฌํ•™๊ณผ ์ˆ˜ํ•™๊ณผ ๊ฐ™์€ ๊ณ„์‚ฐ ์ „๋ฌธ ์ง€์‹์ด ํ•„์š”ํ•œ ๊ณผ๋ชฉ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ธ๋ฌธํ•™ ๋ฐ ์‚ฌํšŒ ๊ณผํ•™ ๋ถ„์•ผ์˜ ํ•™๋ฌธ๋„ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
CMMLU๋Š” ์ค‘๊ตญ์–ด์™€ ๋ฌธํ™”์˜ ๋งฅ๋ฝ์—์„œ LLM์˜ ์ง€์‹๊ณผ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ๊ณ ์•ˆ๋œ ํฌ๊ด„์ ์ธ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ์ž…๋‹ˆ๋‹ค. CMMLU๋Š” ์ดˆ๊ธ‰๋ถ€ํ„ฐ ๊ณ ๊ธ‰ ์ „๋ฌธ๊ฐ€ ์ˆ˜์ค€์— ์ด๋ฅด๋Š” 67๊ฐœ์˜ ์ฃผ์ œ๋กœ ๊ตฌ์„ฑ๋œ ๊ด‘๋ฒ”์œ„ํ•œ ์ฃผ์ œ๋ฅผ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ๋ฌผ๋ฆฌํ•™ ๋ฐ ์ˆ˜ํ•™๊ณผ ๊ฐ™์ด ๊ณ„์‚ฐ ์ „๋ฌธ ์ง€์‹์ด ํ•„์š”ํ•œ ๊ณผ๋ชฉ๊ณผ ์ธ๋ฌธํ•™ ๋ฐ ์‚ฌํšŒ๊ณผํ•™ ๋ถ„์•ผ์˜ ๊ณผ๋ชฉ๋„ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
**Figure A.12 \(|\) Auto-evaluation rated DDx (top-3) accuracy as a function of consultation turns provided to the model. (a) Top-3 auto-evaluation DDx accuracy as a function of the number of turns for the AMIE DDx on AMIE and PCP consultations with respect to the ground truth. (b) Top-3 auto-evaluation DDx accuracy as a function of the number of turns for the AMIE DDx on AMIE and PCP consultations with respect to the the accepted differential.
**Figure A.12 \(|\) ๋ชจ๋ธ์— ์ œ๊ณต๋œ ์ƒ๋‹ด ํšŸ์ˆ˜์— ๋”ฐ๋ฅธ ์ž๋™ ํ‰๊ฐ€ DDx (์ƒ์œ„ 3๊ฐœ) ์ •ํ™•๋„. (a) ์‹ค์ œ ๊ฐ’์— ๋Œ€ํ•œ AMIE ๋ฐ PCP ์ƒ๋‹ด์—์„œ AMIE DDx์˜ ์ƒ๋‹ด ํšŸ์ˆ˜์— ๋”ฐ๋ฅธ ์ƒ์œ„ 3๊ฐœ ์ž๋™ ํ‰๊ฐ€ DDx ์ •ํ™•๋„. (b) ์Šน์ธ๋œ ์ฐจ๋ณ„ ์ง„๋‹จ์— ๋Œ€ํ•œ AMIE ๋ฐ PCP ์ƒ๋‹ด์—์„œ AMIE DDx์˜ ์ƒ๋‹ด ํšŸ์ˆ˜์— ๋”ฐ๋ฅธ ์ƒ์œ„ 3๊ฐœ ์ž๋™ ํ‰๊ฐ€ DDx ์ •ํ™•๋„.
**๊ทธ๋ฆผ A.12 \(| ๋ชจ๋ธ์— ์ œ๊ณต๋œ ์ƒ๋‹ด ํšŸ์ˆ˜์˜ ํ•จ์ˆ˜๋กœ ํ‰๊ฐ€๋œ ์ž๋™ ํ‰๊ฐ€ DDx(์ƒ์œ„ 3๊ฐœ) ์ •ํ™•๋„. (a) ๊ธฐ์ค€ ์ง„์‹ค์— ๋Œ€ํ•œ AMIE DDx์˜ ๊ฒฝ์šฐ AMIE ๋ฐ PCP ์ƒ๋‹ด์˜ ํ„ด ์ˆ˜์— ๋”ฐ๋ฅธ ์ƒ์œ„ 3๊ฐœ ์ž๋™ ํ‰๊ฐ€ DDx ์ •ํ™•๋„. (b) AMIE DDx์˜ ๊ฒฝ์šฐ AMIE ๋ฐ PCP ์ƒ๋‹ด์˜ ํ„ด ์ˆ˜์— ๋”ฐ๋ฅธ ์ƒ์œ„ 3๊ฐœ ์ž๋™ ํ‰๊ฐ€ DDx ์ •ํ™•๋„, ํ—ˆ์šฉ๋œ ์ฐจ๋“ฑ์— ๋Œ€ํ•œ ์ •ํ™•๋„.
There is an important info hidden inside a lot of irrelevant text. Find it and memorize them. I will quiz you about the important information there. The grass is green. The sky is blue. The sun is yellow. Here we go. There and back again. (repeat x times) The pass key is **9054**. Remember it. **9054** is the pass key. The grass is green. The sky is blue. The sun is yellow. Here we go. There and ack again. (repeat y times) What is the pass key?
๋งŽ์€ ๊ด€๋ จ ์—†๋Š” ํ…์ŠคํŠธ ์†์— ์ค‘์š”ํ•œ ์ •๋ณด๊ฐ€ ์ˆจ๊ฒจ์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๊ฒƒ์„ ์ฐพ์•„์„œ ๊ธฐ์–ตํ•˜์„ธ์š”. ์ค‘์š”ํ•œ ์ •๋ณด์— ๋Œ€ํ•ด ํ€ด์ฆˆ๋ฅผ ๋‚ผ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ’€์€ ์ดˆ๋ก์ƒ‰์ž…๋‹ˆ๋‹ค. ํ•˜๋Š˜์€ ํŒŒ๋ž€์ƒ‰์ž…๋‹ˆ๋‹ค. ํƒœ์–‘์€ ๋…ธ๋ž€์ƒ‰์ž…๋‹ˆ๋‹ค. ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ”๋‹ค๊ฐ€ ๋‹ค์‹œ ๋Œ์•„์˜ต๋‹ˆ๋‹ค. (x๋ฒˆ ๋ฐ˜๋ณต) ์•”ํ˜ธ ํ‚ค๋Š” **9054**์ž…๋‹ˆ๋‹ค. ๊ธฐ์–ตํ•˜์„ธ์š”. **9054**๊ฐ€ ์•”ํ˜ธ ํ‚ค์ž…๋‹ˆ๋‹ค. ํ’€์€ ์ดˆ๋ก์ƒ‰์ž…๋‹ˆ๋‹ค. ํ•˜๋Š˜์€ ํŒŒ๋ž€์ƒ‰์ž…๋‹ˆ๋‹ค. ํƒœ์–‘์€ ๋…ธ๋ž€์ƒ‰์ž…๋‹ˆ๋‹ค. ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ”๋‹ค๊ฐ€ ๋‹ค์‹œ ๋Œ์•„์˜ต๋‹ˆ๋‹ค. (y๋ฒˆ ๋ฐ˜๋ณต) ์•”ํ˜ธ ํ‚ค๋Š” ๋ฌด์—‡์ž…๋‹ˆ๊นŒ?
๊ด€๋ จ ์—†๋Š” ํ…์ŠคํŠธ๊ฐ€ ๋งŽ์€ ์ •๋ณด ์•ˆ์— ์ค‘์š”ํ•œ ์ •๋ณด๊ฐ€ ์ˆจ๊ฒจ์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ์ฐพ์•„์„œ ์™ธ์šฐ์„ธ์š”. ์ค‘์š”ํ•œ ์ •๋ณด์— ๋Œ€ํ•ด ํ€ด์ฆˆ๋ฅผ ์ถœ์ œํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ์ž”๋””๋Š” ์ดˆ๋ก์ƒ‰์ž…๋‹ˆ๋‹ค. ํ•˜๋Š˜์€ ํŒŒ๋ž€์ƒ‰์ž…๋‹ˆ๋‹ค. ํƒœ์–‘์€ ๋…ธ๋ž€์ƒ‰์ž…๋‹ˆ๋‹ค. ์‹œ์ž‘ํ•˜์ฃ . ๊ฑฐ๊ธฐ๊นŒ์ง€, ๋‹ค์‹œ ๊ฑฐ๊ธฐ๊นŒ์ง€. (x๋ฒˆ ๋ฐ˜๋ณต) ๋น„๋ฐ€๋ฒˆํ˜ธ๋Š” **9054**์ž…๋‹ˆ๋‹ค. ๊ธฐ์–ตํ•˜์„ธ์š”. **9054**๊ฐ€ ๋น„๋ฐ€๋ฒˆํ˜ธ์ž…๋‹ˆ๋‹ค. ์ž”๋””๋Š” ์ดˆ๋ก์ƒ‰์ž…๋‹ˆ๋‹ค. ํ•˜๋Š˜์€ ํŒŒ๋ž€์ƒ‰์ž…๋‹ˆ๋‹ค. ํƒœ์–‘์€ ๋…ธ๋ž€์ƒ‰์ž…๋‹ˆ๋‹ค. ์‹œ์ž‘ํ•˜์ฃ . ๊ฑฐ๊ธฐ๊นŒ์ง€, ๋‹ค์‹œ ๊ฑฐ๊ธฐ๊นŒ์ง€. (y๋ฒˆ ๋ฐ˜๋ณต) ๋น„๋ฐ€๋ฒˆํ˜ธ๊ฐ€ ๋ญ”๊ฐ€์š”?
In this section, we demonstrate the guidelines we provide for the conversational QA data collection. We request the annotators to act as both user and agent to construct the conversational QA sample based on the given document, which makes the annotation more efficient. The guidelines consists of three parts: 1) What does conversational QA samples look like; 2) What kinds of conversational QA samples we need; and 3) What we need to annotate.
์ด ์„น์…˜์—์„œ๋Š” ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘์„ ์œ„ํ•œ ๊ฐ€์ด๋“œ๋ผ์ธ์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์–ด์ง„ ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋Œ€ํ™”ํ˜• QA ์ƒ˜ํ”Œ์„ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ฃผ์„์ž๋“ค์—๊ฒŒ ์‚ฌ์šฉ์ž์™€ ์—์ด์ „ํŠธ ์—ญํ• ์„ ๋ชจ๋‘ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์š”์ฒญํ•˜์—ฌ ์ฃผ์„ ์ž‘์—…์˜ ํšจ์œจ์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค. ๊ฐ€์ด๋“œ๋ผ์ธ์€ ์„ธ ๋ถ€๋ถ„์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: 1) ๋Œ€ํ™”ํ˜• QA ์ƒ˜ํ”Œ์˜ ํ˜•ํƒœ; 2) ํ•„์š”ํ•œ ๋Œ€ํ™”ํ˜• QA ์ƒ˜ํ”Œ์˜ ์ข…๋ฅ˜; 3) ์ฃผ์„ํ•ด์•ผ ํ•  ๋‚ด์šฉ.
์ด ์„น์…˜์—์„œ๋Š” ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘์— ๋Œ€ํ•œ ์ง€์นจ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์„ ์ž‘์„ฑ์ž์—๊ฒŒ ์‚ฌ์šฉ์ž์™€ ์—์ด์ „ํŠธ ์—ญํ• ์„ ๋ชจ๋‘ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ฃผ์–ด์ง„ ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋Œ€ํ™”ํ˜• QA ์ƒ˜ํ”Œ์„ ๊ตฌ์„ฑํ•˜๋„๋ก ์š”์ฒญํ•˜์—ฌ ์ฃผ์„์„ ๋ณด๋‹ค ํšจ์œจ์ ์œผ๋กœ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์ด๋“œ๋ผ์ธ์€ ์„ธ ๋ถ€๋ถ„์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: 1) ๋Œ€ํ™”ํ˜• QA ์ƒ˜ํ”Œ์€ ์–ด๋–ค ๋ชจ์Šต์ธ๊ฐ€; 2) ์–ด๋–ค ์ข…๋ฅ˜์˜ ๋Œ€ํ™”ํ˜• QA ์ƒ˜ํ”Œ์ด ํ•„์š”ํ•œ๊ฐ€; 3) ์ฃผ์„์„ ๋‹ฌ์•„์•ผ ํ•˜๋Š” ๋‚ด์šฉ์€ ๋ฌด์—‡์ธ๊ฐ€?
Owing to the strong language comprehension, complex task reasoning, and common sense understanding capabilities, LLM-based autonomous agents have shown significant potential to influence multiple domains. This section provides a succinct summary of previous studies, categorizing them according to their applications in three distinct areas: social science, natural science, and engineering (see the left part of Figure 5 for a global overview).
๊ฐ•๋ ฅํ•œ ์–ธ์–ด ์ดํ•ด, ๋ณต์žกํ•œ ์ž‘์—… ์ถ”๋ก , ๊ทธ๋ฆฌ๊ณ  ์ƒ์‹ ์ดํ•ด ๋Šฅ๋ ฅ ๋•๋ถ„์—, LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ๋Š” ์—ฌ๋Ÿฌ ๋ถ„์•ผ์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ด์ „ ์—ฐ๊ตฌ๋“ค์„ ๊ฐ„๋žตํ•˜๊ฒŒ ์š”์•ฝํ•˜๊ณ , ์ด๋ฅผ ์‚ฌํšŒ ๊ณผํ•™, ์ž์—ฐ ๊ณผํ•™, ๊ทธ๋ฆฌ๊ณ  ๊ณตํ•™์˜ ์„ธ ๊ฐ€์ง€ ๊ตฌ๋ณ„๋œ ์˜์—ญ์—์„œ์˜ ์‘์šฉ์— ๋”ฐ๋ผ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค (์ „๋ฐ˜์ ์ธ ๊ฐœ์š”๋Š” Figure 5์˜ ์™ผ์ชฝ ๋ถ€๋ถ„์„ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค).
๊ฐ•๋ ฅํ•œ ์–ธ์–ด ์ดํ•ด๋ ฅ, ๋ณต์žกํ•œ ๊ณผ์ œ ์ถ”๋ก  ๋ฐ ์ƒ์‹ ์ดํ•ด ๋Šฅ๋ ฅ ๋•๋ถ„์— LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ๋Š” ์—ฌ๋Ÿฌ ์˜์—ญ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ์ƒ๋‹นํ•œ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์‚ฌํšŒ ๊ณผํ•™, ์ž์—ฐ ๊ณผํ•™ ๋ฐ ๊ณตํ•™์˜ ์„ธ ๊ฐ€์ง€ ๋ถ„์•ผ์— ๋Œ€ํ•œ ์‘์šฉ ๋ถ„์•ผ์— ๋”ฐ๋ผ ์ด์ „ ์—ฐ๊ตฌ๋ฅผ ๊ฐ„๋žตํ•˜๊ฒŒ ์š”์•ฝํ•˜์—ฌ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 5์˜ ์™ผ์ชฝ ๋ถ€๋ถ„์„ ์ฐธ์กฐํ•˜์„ธ์š”).
The focus of this paper is on the datasets used to train language models. While recent work focused on documenting the potential harms that could arise from problematic datasets (Bender and Friedman, 2018; Gebru et al., 2020), less work has been done to quantitatively analyze properties of real language modelling datasets, like Dodge et al. (2021) has done for C4. Our paper provides analysis on one particular axis, that of data duplication.
์ด ๋…ผ๋ฌธ์˜ ์ดˆ์ ์€ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๋ฐ ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ์…‹์— ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ๋“ค์€ ๋ฌธ์ œ ์žˆ๋Š” ๋ฐ์ดํ„ฐ์…‹์—์„œ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ์  ํ•ด์•…์„ ๋ฌธ์„œํ™”ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์ง€๋งŒ (Bender and Friedman, 2018; Gebru et al., 2020), ์‹ค์ œ ์–ธ์–ด ๋ชจ๋ธ๋ง ๋ฐ์ดํ„ฐ์…‹์˜ ํŠน์„ฑ์„ ์ •๋Ÿ‰์ ์œผ๋กœ ๋ถ„์„ํ•˜๋Š” ์ž‘์—…์€ ์ƒ๋Œ€์ ์œผ๋กœ ์ ์—ˆ์Šต๋‹ˆ๋‹ค. Dodge et al. (2021)์ด C4์— ๋Œ€ํ•ด ์ˆ˜ํ–‰ํ•œ ๊ฒƒ์ฒ˜๋Ÿผ ๋ง์ž…๋‹ˆ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์€ ๋ฐ์ดํ„ฐ ์ค‘๋ณต์ด๋ผ๋Š” ํŠน์ • ์ถ•์— ๋Œ€ํ•œ ๋ถ„์„์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์˜ ์ดˆ์ ์€ ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฌธ์ œ๊ฐ€ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ์  ์œ„ํ—˜์„ ๋ฌธ์„œํ™”ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค(Bender and Friedman, 2018; Gebru et al., 2020), Dodge et al. (2021)์ด C4์— ๋Œ€ํ•ด ์ˆ˜ํ–‰ํ•œ ๊ฒƒ๊ณผ ๊ฐ™์ด ์‹ค์ œ ์–ธ์–ด ๋ชจ๋ธ๋ง ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์†์„ฑ์„ ์ •๋Ÿ‰์ ์œผ๋กœ ๋ถ„์„ํ•˜๋Š” ์—ฐ๊ตฌ๋Š” ๋งŽ์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋ฐ์ดํ„ฐ ์ค‘๋ณต์ด๋ผ๋Š” ํŠน์ • ์ถ•์— ๋Œ€ํ•œ ๋ถ„์„์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
Figure 1: The inference process of Adapt-LLM step-by-step: given a question (step 1), an LLM decides (step 2) whether to answer the question directly (step 3) or to ask for additional contextual information, generating the special \(\langle\)RET\(\rangle\) token; for the later, an off-the-shelf IR system is used to retrieve relevant context (step 4), which is used alongside the question to prompt again the LLM for the final answer (step 5).
Figure 1: Adapt-LLM์˜ ์ถ”๋ก  ๊ณผ์ • ๋‹จ๊ณ„๋ณ„ ์„ค๋ช…: ์งˆ๋ฌธ์ด ์ฃผ์–ด์ง€๋ฉด (1๋‹จ๊ณ„), LLM์€ ์งˆ๋ฌธ์— ์ง์ ‘ ๋‹ตํ• ์ง€ (3๋‹จ๊ณ„) ์•„๋‹ˆ๋ฉด ์ถ”๊ฐ€์ ์ธ ๋งฅ๋ฝ ์ •๋ณด๋ฅผ ์š”์ฒญํ• ์ง€ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค (2๋‹จ๊ณ„). ํ›„์ž์˜ ๊ฒฝ์šฐ, ํŠน์ˆ˜ํ•œ \(\langle\)RET\(\rangle\) ํ† ํฐ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ดํ›„, ๊ด€๋ จ๋œ ๋งฅ๋ฝ์„ ๊ฒ€์ƒ‰ํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด์˜ IR ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•˜๋ฉฐ (4๋‹จ๊ณ„), ์ด ๋งฅ๋ฝ์€ ์งˆ๋ฌธ๊ณผ ํ•จ๊ป˜ ์ตœ์ข… ๋‹ต๋ณ€์„ ์œ„ํ•ด LLM์„ ๋‹ค์‹œ ํ”„๋กฌํ”„ํŠธํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค (5๋‹จ๊ณ„).
๊ทธ๋ฆผ 1: ์งˆ๋ฌธ์ด ์ฃผ์–ด์ง€๋ฉด(1๋‹จ๊ณ„), LLM์ด ์งˆ๋ฌธ์— ์ง์ ‘ ๋‹ตํ• ์ง€(3๋‹จ๊ณ„) ์•„๋‹ˆ๋ฉด ์ถ”๊ฐ€์ ์ธ ์ปจํ…์ŠคํŠธ ์ •๋ณด๋ฅผ ์š”์ฒญํ•˜์—ฌ ํŠน์ˆ˜(\(RET\(\rangle\) ํ† ํฐ์„ ์ƒ์„ฑํ•˜๋Š”์ง€ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค(2๋‹จ๊ณ„); ํ›„์ž์˜ ๊ฒฝ์šฐ, ๊ธฐ์„ฑ IR ์‹œ์Šคํ…œ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ด€๋ จ ์ปจํ…์ŠคํŠธ๋ฅผ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค(4๋‹จ๊ณ„), ์ด ์ปจํ…์ŠคํŠธ๋Š” ์งˆ๋ฌธ๊ณผ ํ•จ๊ป˜ ์‚ฌ์šฉ๋˜์–ด ์ตœ์ข… ๋‹ต๋ณ€์„ ์œ„ํ•ด ๋‹ค์‹œ LLM์— ํ”„๋กฌํ”„ํŠธํ•ฉ๋‹ˆ๋‹ค(5๋‹จ๊ณ„).
We conduct ablation studies in terms of the number of unanswerable samples in HumanAnnotatedConvQA. We find that using a small amount of unanswerable samples (e.g., 1.5k) is able to achieve remarkable results on the unanswerable evaluation, as well as the overall conversational QA score (i.e., Avg-CQA). Interestingly, incorporating more unanswerable samples does not necessarily lead to higher accuracy scores in the unanswerable evaluation.
์šฐ๋ฆฌ๋Š” HumanAnnotatedConvQA์—์„œ ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ์ƒ˜ํ”Œ์˜ ์ˆ˜์— ๋Œ€ํ•œ ์†Œ๊ฑฐ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์†Œ๋Ÿ‰์˜ ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ์ƒ˜ํ”Œ(์˜ˆ: 1.5k)์„ ์‚ฌ์šฉํ•˜๋ฉด ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ํ‰๊ฐ€์—์„œ ๋†€๋ผ์šด ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์„ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ „์ฒด ๋Œ€ํ™”ํ˜• QA ์ ์ˆ˜(์ฆ‰, Avg-CQA)์—์„œ๋„ ๋›ฐ์–ด๋‚œ ์„ฑ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„, ๋” ๋งŽ์€ ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ์ƒ˜ํ”Œ์„ ํฌํ•จํ•œ๋‹ค๊ณ  ํ•ด์„œ ๋ฐ˜๋“œ์‹œ ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ํ‰๊ฐ€์—์„œ ๋” ๋†’์€ ์ •ํ™•๋„ ์ ์ˆ˜๋ฅผ ์–ป๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค.
HumanAnnotatedConvQA์—์„œ ๋‹ต๋ณ€ํ•  ์ˆ˜ ์—†๋Š” ์ƒ˜ํ”Œ์˜ ์ˆ˜ ์ธก๋ฉด์—์„œ ์ ˆ์ œ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์†Œ๋Ÿ‰์˜ ๋‹ต๋ณ€ ๋ถˆ๊ฐ€๋Šฅํ•œ ์ƒ˜ํ”Œ(์˜ˆ: 1.5k)์„ ์‚ฌ์šฉํ•˜๋ฉด ๋‹ต๋ณ€ ๋ถˆ๊ฐ€๋Šฅํ•œ ํ‰๊ฐ€์™€ ์ „์ฒด ๋Œ€ํ™”ํ˜• QA ์ ์ˆ˜(์ฆ‰, Avg-CQA)์—์„œ ๋†€๋ผ์šด ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํฅ๋ฏธ๋กญ๊ฒŒ๋„ ๋” ๋งŽ์€ ๋‹ต๋ณ€ ๋ถˆ๊ฐ€๋Šฅํ•œ ์ƒ˜ํ”Œ์„ ํ†ตํ•ฉํ•œ๋‹ค๊ณ  ํ•ด์„œ ๋‹ต๋ณ€ ๋ถˆ๊ฐ€๋Šฅํ•œ ํ‰๊ฐ€์—์„œ ๋ฐ˜๋“œ์‹œ ๋” ๋†’์€ ์ •ํ™•๋„ ์ ์ˆ˜๋กœ ์ด์–ด์ง€์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค.
Language identification.We use the fastText language classifier of CCNet (Wenzek et al., 2020) at the document-level: it uses characters n-gram and was trained on Wikipedia, supporting 176 languages. We remove documents for which the top language scores below 0.65: this usually corresponds to pages without any natural text. For this paper, we focus on English; RefinedWeb can also be derived for other languages, see Appendix D for details.
์–ธ์–ด ์‹๋ณ„. ์šฐ๋ฆฌ๋Š” ๋ฌธ์„œ ์ˆ˜์ค€์—์„œ CCNet (Wenzek et al., 2020)์˜ fastText ์–ธ์–ด ๋ถ„๋ฅ˜๊ธฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ถ„๋ฅ˜๊ธฐ๋Š” ๋ฌธ์ž n-gram์„ ์‚ฌ์šฉํ•˜๋ฉฐ, 176๊ฐœ ์–ธ์–ด๋ฅผ ์ง€์›ํ•˜๋Š” Wikipedia์—์„œ ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์š” ์–ธ์–ด ์ ์ˆ˜๊ฐ€ 0.65 ์ดํ•˜์ธ ๋ฌธ์„œ๋Š” ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ž์—ฐ์–ด ํ…์ŠคํŠธ๊ฐ€ ์—†๋Š” ํŽ˜์ด์ง€์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์˜์–ด์— ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์œผ๋ฉฐ, RefinedWeb์€ ๋‹ค๋ฅธ ์–ธ์–ด๋กœ๋„ ํŒŒ์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก D๋ฅผ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค.
์–ธ์–ด ์‹๋ณ„.๋ฌธ์„œ ์ˆ˜์ค€์—์„œ CCNet(Wenzek et al., 2020)์˜ fastText ์–ธ์–ด ๋ถ„๋ฅ˜๊ธฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ์ž n-๊ทธ๋žจ์„ ์‚ฌ์šฉํ•˜๋ฉฐ 176๊ฐœ ์–ธ์–ด๋ฅผ ์ง€์›ํ•˜๋Š” ์œ„ํ‚คํ”ผ๋””์•„์—์„œ ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ƒ์œ„ ์–ธ์–ด ์ ์ˆ˜๊ฐ€ 0.65 ๋ฏธ๋งŒ์ธ ๋ฌธ์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ž์—ฐ ํ…์ŠคํŠธ๊ฐ€ ์—†๋Š” ํŽ˜์ด์ง€์— ํ•ด๋‹นํ•˜๋ฏ€๋กœ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์˜์–ด์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ๋‹ค๋ฅธ ์–ธ์–ด์— ๋Œ€ํ•ด์„œ๋„ RefinedWeb์„ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก D๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
We collected responses from all models with a temperature of 0.01 and set the maximum number of tokens to be generated to 500. All remaining parameters were kept at the default (e.g. top_p = 0.7). Sampling with a temperature of 0.01 is not fully deterministic, but it reduces the variability of models' responses. This also reduces their creativity, which can minimize the risk of models giving more unusual and potentially unsafe responses.
๋ชจ๋“  ๋ชจ๋ธ์—์„œ ์˜จ๋„๋ฅผ 0.01๋กœ ์„ค์ •ํ•˜๊ณ  ์ƒ์„ฑํ•  ์ตœ๋Œ€ ํ† ํฐ ์ˆ˜๋ฅผ 500์œผ๋กœ ์„ค์ •ํ•˜์—ฌ ์‘๋‹ต์„ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€ ๋ชจ๋“  ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” ๊ธฐ๋ณธ๊ฐ’์œผ๋กœ ์œ ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: top_p = 0.7). ์˜จ๋„ 0.01๋กœ ์ƒ˜ํ”Œ๋งํ•˜๋Š” ๊ฒƒ์€ ์™„์ „ํžˆ ๊ฒฐ์ •๋ก ์ ์ด์ง€๋Š” ์•Š์ง€๋งŒ, ๋ชจ๋ธ ์‘๋‹ต์˜ ๋ณ€๋™์„ฑ์„ ์ค„์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Š” ๋˜ํ•œ ๋ชจ๋ธ์˜ ์ฐฝ์˜์„ฑ์„ ๊ฐ์†Œ์‹œ์ผœ, ๋” ์ด์ƒํ•˜๊ณ  ์ž ์žฌ์ ์œผ๋กœ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ์‘๋‹ต์„ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
์˜จ๋„๊ฐ€ 0.01์ธ ๋ชจ๋“  ๋ชจ๋ธ์˜ ์‘๋‹ต์„ ์ˆ˜์ง‘ํ•˜๊ณ  ์ƒ์„ฑํ•  ์ตœ๋Œ€ ํ† ํฐ ์ˆ˜๋ฅผ 500์œผ๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€ ๋ชจ๋“  ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” ๊ธฐ๋ณธ๊ฐ’(์˜ˆ: top_p = 0.7)์œผ๋กœ ์œ ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜จ๋„๊ฐ€ 0.01์ธ ์ƒ˜ํ”Œ๋ง์€ ์™„์ „ํžˆ ๊ฒฐ์ •๋ก ์ ์ด์ง€๋Š” ์•Š์ง€๋งŒ ๋ชจ๋ธ ์‘๋‹ต์˜ ๋ณ€๋™์„ฑ์„ ์ค„์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋ธ์˜ ์ฐฝ์˜์„ฑ๋„ ๊ฐ์†Œํ•˜์—ฌ ๋ชจ๋ธ์ด ๋” ํŠน์ดํ•˜๊ณ  ์ž ์žฌ์ ์œผ๋กœ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ์‘๋‹ต์„ ์ œ๊ณตํ•  ์œ„ํ—˜์„ ์ตœ์†Œํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
This paper introduces v0.5 of the **AI Safety Benchmark** which has been created by the MLCommons AI Safety Working Group (WG). The MLCommons AI Safety WG is a consortium of industry and academic researchers, engineers, and practitioners. The primary goal of the WG is to advance the state of the art for evaluating AI safety. We hope to facilitate better AI safety processes and stimulate AI safety innovation across industry and research.
์ด ๋…ผ๋ฌธ์€ MLCommons AI Safety Working Group (WG)์— ์˜ํ•ด ๋งŒ๋“ค์–ด์ง„ **AI Safety Benchmark** v0.5๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. MLCommons AI Safety WG๋Š” ์‚ฐ์—… ๋ฐ ํ•™๊ณ„ ์—ฐ๊ตฌ์ž, ์—”์ง€๋‹ˆ์–ด, ์‹ค๋ฌด์ž๋“ค๋กœ ๊ตฌ์„ฑ๋œ ์ปจ์†Œ์‹œ์—„์ž…๋‹ˆ๋‹ค. WG์˜ ์ฃผ์š” ๋ชฉํ‘œ๋Š” AI ์•ˆ์ „์„ฑ ํ‰๊ฐ€์˜ ์ตœ์ฒจ๋‹จ ๊ธฐ์ˆ ์„ ๋ฐœ์ „์‹œํ‚ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋” ๋‚˜์€ AI ์•ˆ์ „์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ด‰์ง„ํ•˜๊ณ  ์‚ฐ์—… ๋ฐ ์—ฐ๊ตฌ ์ „๋ฐ˜์— ๊ฑธ์ณ AI ์•ˆ์ „์„ฑ ํ˜์‹ ์„ ์ž๊ทนํ•˜๊ธฐ๋ฅผ ํฌ๋งํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์—์„œ๋Š” MLCommons AI Safety Working Group(WG)์—์„œ ๋งŒ๋“  **AI Safety Benchmark** v0.5๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. MLCommons AI Safety WG๋Š” ์—…๊ณ„ ๋ฐ ํ•™๊ณ„ ์—ฐ๊ตฌ์ž, ์—”์ง€๋‹ˆ์–ด, ์‹ค๋ฌด์ž๋กœ ๊ตฌ์„ฑ๋œ ์ปจ์†Œ์‹œ์—„์ž…๋‹ˆ๋‹ค. WG์˜ ์ฃผ์š” ๋ชฉํ‘œ๋Š” AI ์•ˆ์ „์„ฑ ํ‰๊ฐ€์˜ ์ตœ์ฒจ๋‹จ ๊ธฐ์ˆ ์„ ๋ฐœ์ „์‹œํ‚ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋” ๋‚˜์€ AI ์•ˆ์ „ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ด‰์ง„ํ•˜๊ณ  ์—…๊ณ„์™€ ์—ฐ๊ตฌ ๋ถ„์•ผ ์ „๋ฐ˜์— ๊ฑธ์ณ AI ์•ˆ์ „ ํ˜์‹ ์„ ์ด‰์ง„ํ•˜๊ธฐ๋ฅผ ํฌ๋งํ•ฉ๋‹ˆ๋‹ค.
In Figure 13, we display the dynamic changes in token selection tendencies throughout the SLM training process. We chose four checkpoints during the training process (0%, 33%, 66%, and 100%) to analyze the current tendencies in token selection. The preferences for token selection are indicated by different colors, ranging from high to low preference, typically represented as deep blue, blue, black, orange, and dark orange, respectively.
Figure 13์—์„œ๋Š” SLM ํ›ˆ๋ จ ๊ณผ์ • ์ „๋ฐ˜์— ๊ฑธ์นœ ํ† ํฐ ์„ ํƒ ๊ฒฝํ–ฅ์˜ ๋™์  ๋ณ€ํ™”๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๊ณผ์ • ์ค‘ 4๊ฐœ์˜ ์ฒดํฌํฌ์ธํŠธ(0%, 33%, 66%, 100%)๋ฅผ ์„ ํƒํ•˜์—ฌ ํ˜„์žฌ ํ† ํฐ ์„ ํƒ ๊ฒฝํ–ฅ์„ ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ† ํฐ ์„ ํƒ์— ๋Œ€ํ•œ ์„ ํ˜ธ๋„๋Š” ๋‹ค์–‘ํ•œ ์ƒ‰์ƒ์œผ๋กœ ํ‘œ์‹œ๋˜๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ๋†’์€ ์„ ํ˜ธ๋„์—์„œ ๋‚ฎ์€ ์„ ํ˜ธ๋„๋กœ ๊นŠ์€ ํŒŒ๋ž€์ƒ‰, ํŒŒ๋ž€์ƒ‰, ๊ฒ€์€์ƒ‰, ์ฃผํ™ฉ์ƒ‰, ์ง™์€ ์ฃผํ™ฉ์ƒ‰์œผ๋กœ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 13์—์„œ๋Š” SLM ํ•™์Šต ๊ณผ์ •์—์„œ ํ† ํฐ ์„ ํƒ ๊ฒฝํ–ฅ์˜ ๋™์  ๋ณ€ํ™”๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ•™์Šต ๊ณผ์ •์—์„œ ๋„ค ๊ฐœ์˜ ์ฒดํฌํฌ์ธํŠธ(0%, 33%, 66%, 100%)๋ฅผ ์„ ํƒํ•˜์—ฌ ํ† ํฐ ์„ ํƒ์˜ ํ˜„์žฌ ๊ฒฝํ–ฅ์„ ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ† ํฐ ์„ ํƒ์— ๋Œ€ํ•œ ์„ ํ˜ธ๋„๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ง„ํ•œ ํŒŒ๋ž€์ƒ‰, ํŒŒ๋ž€์ƒ‰, ๊ฒ€์€์ƒ‰, ์ฃผํ™ฉ์ƒ‰, ์ง„ํ•œ ์ฃผํ™ฉ์ƒ‰์œผ๋กœ ํ‘œ์‹œ๋˜๋Š” ๋†’์€ ์„ ํ˜ธ๋„์—์„œ ๋‚ฎ์€ ์„ ํ˜ธ๋„๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์ƒ‰์ƒ์œผ๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค.
\(\bullet\)**HH-RLHF**[170] consists of around 169K instances, and can be divided into two parts that focus on the helpfulness and harmlessness of LLMs, respectively. Each instance is an open-ended conversation between a crowdworker and a chat model, about seeking assistance, advice, or task completion. The chat model provides two responses to each user query, and the more helpful or harmful responses will be chosen as the annotations.
\(\bullet\)**HH-RLHF**[170]๋Š” ์•ฝ 169K๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๊ฐ๊ฐ LLM์˜ ์œ ์šฉ์„ฑ๊ณผ ๋ฌดํ•ด์„ฑ์— ์ค‘์ ์„ ๋‘” ๋‘ ๋ถ€๋ถ„์œผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์ธ์Šคํ„ด์Šค๋Š” ํฌ๋ผ์šฐ๋“œ์›Œ์ปค์™€ ์ฑ„ํŒ… ๋ชจ๋ธ ๊ฐ„์˜ ๊ฐœ๋ฐฉํ˜• ๋Œ€ํ™”๋กœ, ๋„์›€, ์กฐ์–ธ ๋˜๋Š” ์ž‘์—… ์™„๋ฃŒ๋ฅผ ์š”์ฒญํ•˜๋Š” ๋‚ด์šฉ์ž…๋‹ˆ๋‹ค. ์ฑ„ํŒ… ๋ชจ๋ธ์€ ๊ฐ ์‚ฌ์šฉ์ž ์งˆ๋ฌธ์— ๋‘ ๊ฐ€์ง€ ์‘๋‹ต์„ ์ œ๊ณตํ•˜๋ฉฐ, ๋” ์œ ์šฉํ•˜๊ฑฐ๋‚˜ ํ•ด๋กœ์šด ์‘๋‹ต์ด ์ฃผ์„์œผ๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค.
\(\bullet\)**HH-RLHF**[170]๋Š” ์•ฝ 169K ์ธ์Šคํ„ด์Šค๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๊ฐ๊ฐ LLM์˜ ์œ ์šฉ์„ฑ๊ณผ ๋ฌดํ•ด์„ฑ์— ์ดˆ์ ์„ ๋งž์ถ˜ ๋‘ ๋ถ€๋ถ„์œผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์ธ์Šคํ„ด์Šค๋Š” ํฌ๋ผ์šฐ๋“œ์›Œ์ปค์™€ ์ฑ„ํŒ… ๋ชจ๋ธ ๊ฐ„์˜ ๋Œ€ํ™”๋กœ, ๋„์›€์ด๋‚˜ ์กฐ์–ธ, ์ž‘์—… ์™„๋ฃŒ๋ฅผ ๊ตฌํ•˜๋Š” ๋‚ด์šฉ์ž…๋‹ˆ๋‹ค. ์ฑ„ํŒ… ๋ชจ๋ธ์€ ๊ฐ ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ์— ๋Œ€ํ•ด ๋‘ ๊ฐœ์˜ ์‘๋‹ต์„ ์ œ๊ณตํ•˜๋ฉฐ, ๋” ์œ ์šฉํ•˜๊ฑฐ๋‚˜ ์œ ํ•ดํ•œ ์‘๋‹ต์ด ์ฃผ์„์œผ๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค.
1. Legality varies across jurisdictions, even for countries with broadly similar legal systems. 2. Laws are often incomplete, difficult to understand, and may have surprising gaps, asymmetries and limitations. 3. The legality of specific activities is difficult to assess and requires trained legal expertise. 4. AI companies have restrictions on the use of their models, which are defined not solely on the basis of legal prohibitions 34.
1. ๋ฒ•์  ์ƒํ™ฉ์€ ๋น„์Šทํ•œ ๋ฒ• ์ฒด๊ณ„๋ฅผ ๊ฐ€์ง„ ๊ตญ๊ฐ€๋“ค ์‚ฌ์ด์—์„œ๋„ ๊ด€ํ•  ๊ตฌ์—ญ์— ๋”ฐ๋ผ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. 2. ๋ฒ•๋ฅ ์€ ์ข…์ข… ๋ถˆ์™„์ „ํ•˜๊ณ  ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ค์šฐ๋ฉฐ, ์˜ˆ์ƒ์น˜ ๋ชปํ•œ ๊ณต๋ฐฑ, ๋น„๋Œ€์นญ์„ฑ ๋ฐ ์ œํ•œ์„ ๊ฐ€์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 3. ํŠน์ • ํ™œ๋™์˜ ํ•ฉ๋ฒ•์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต๊ณ , ํ›ˆ๋ จ๋œ ๋ฒ•๋ฅ  ์ „๋ฌธ๊ฐ€์˜ ์ „๋ฌธ ์ง€์‹์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. 4. AI ํšŒ์‚ฌ๋“ค์€ ๋ฒ•์  ๊ธˆ์ง€ ์‚ฌํ•ญ์—๋งŒ ๊ทผ๊ฑฐํ•˜์ง€ ์•Š๊ณ  ๋ชจ๋ธ ์‚ฌ์šฉ์— ๋Œ€ํ•œ ์ œํ•œ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
1. ๋ฒ•๋ฅ ์€ ๋ฒ•๋ฅ  ์ฒด๊ณ„๊ฐ€ ๋น„์Šทํ•œ ๊ตญ๊ฐ€์—์„œ๋„ ๊ด€ํ• ๊ถŒ์— ๋”ฐ๋ผ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. 2. ๋ฒ•๋ฅ ์€ ๋ถˆ์™„์ „ํ•˜๊ณ  ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ค์šฐ๋ฉฐ ๋†€๋ผ์šด ๊ฒฉ์ฐจ, ๋น„๋Œ€์นญ์„ฑ ๋ฐ ํ•œ๊ณ„๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 3. ํŠน์ • ํ™œ๋™์˜ ํ•ฉ๋ฒ•์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ๋Š” ์–ด๋ ต๊ณ  ์ˆ™๋ จ๋œ ๋ฒ•๋ฅ  ์ „๋ฌธ ์ง€์‹์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. 4. AI ํšŒ์‚ฌ๋Š” ๋ฒ•์  ๊ธˆ์ง€ ์‚ฌํ•ญ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋ชจ๋ธ ์‚ฌ์šฉ์— ๋Œ€ํ•œ ์ œํ•œ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
LLaMA is a collection of foundation language models, released by Meta. Unlike GPT models, LLaMA models are open-source, i.e., model weights are released to the research community under a noncommercial license. Thus, the LLaMA family grows rapidly as these models are widely used by many research groups to develop better open-source LLMs to compete the closed-source ones or to develop task-specific LLMs for mission-critical applications.
LLaMA๋Š” Meta์—์„œ ์ถœ์‹œํ•œ ๊ธฐ์ดˆ ์–ธ์–ด ๋ชจ๋ธ ๋ชจ์Œ์ž…๋‹ˆ๋‹ค. GPT ๋ชจ๋ธ๊ณผ ๋‹ฌ๋ฆฌ, LLaMA ๋ชจ๋ธ์€ ์˜คํ”ˆ ์†Œ์Šค์ด๋ฉฐ, ์ฆ‰ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๊ฐ€ ๋น„์ƒ์—…์  ๋ผ์ด์„ ์Šค ํ•˜์— ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ๊ณต๊ฐœ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ LLaMA ๊ณ„์—ด์€ ๋งŽ์€ ์—ฐ๊ตฌ ๊ทธ๋ฃน์ด ๋” ๋‚˜์€ ์˜คํ”ˆ ์†Œ์Šค LLM์„ ๊ฐœ๋ฐœํ•˜๊ฑฐ๋‚˜ ํ์‡„ํ˜• ๋ชจ๋ธ๊ณผ ๊ฒฝ์Ÿํ•˜๊ฑฐ๋‚˜ ์ค‘์š”ํ•œ ์ž„๋ฌด์— ํŠนํ™”๋œ LLM์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•ด ๋„๋ฆฌ ์‚ฌ์šฉํ•จ์— ๋”ฐ๋ผ ๋น ๋ฅด๊ฒŒ ์„ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
LLaMA๋Š” Meta์—์„œ ๊ณต๊ฐœํ•œ ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ชจ๋ธ ๋ชจ์Œ์ž…๋‹ˆ๋‹ค. GPT ๋ชจ๋ธ๊ณผ ๋‹ฌ๋ฆฌ LLaMA ๋ชจ๋ธ์€ ์˜คํ”ˆ ์†Œ์Šค, ์ฆ‰ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๊ฐ€ ๋น„์ƒ์šฉ ๋ผ์ด์„ ์Šค์— ๋”ฐ๋ผ ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ๊ณต๊ฐœ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ๋งŽ์€ ์—ฐ๊ตฌ ๊ทธ๋ฃน์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜์–ด ๋น„๊ณต๊ฐœ ๋ชจ๋ธ๊ณผ ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ๋Š” ๋” ๋‚˜์€ ์˜คํ”ˆ ์†Œ์Šค LLM์„ ๊ฐœ๋ฐœํ•˜๊ฑฐ๋‚˜ ๋ฏธ์…˜ ํฌ๋ฆฌํ‹ฐ์ปฌ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ์œ„ํ•œ ์ž‘์—…๋ณ„ LLM์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ๋•Œ๋ฌธ์— LLaMA ์ œํ’ˆ๊ตฐ์ด ๋น ๋ฅด๊ฒŒ ์„ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Footnote 4: A further possibility is that rather than using the final losses as per their description of Approach 3, they also used intermediate losses during training. If so, confidence intervals would be substantially narrower, but likely not by an order of magnitude if standard errors were appropriately clustered. Overall, we think this is unlikely to explain the tight intervals we see reported by Hoffman et al. (see Appendix A.2).
Footnote 4: ๋˜ ๋‹ค๋ฅธ ๊ฐ€๋Šฅ์„ฑ์€ Approach 3์˜ ์„ค๋ช…์— ๋”ฐ๋ฅธ ์ตœ์ข… ์†์‹ค์„ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹ , ํ›ˆ๋ จ ์ค‘๊ฐ„์— ์ค‘๊ฐ„ ์†์‹ค๋„ ์‚ฌ์šฉํ–ˆ์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋งŒ์•ฝ ๊ทธ๋ ‡๋‹ค๋ฉด, ์‹ ๋ขฐ ๊ตฌ๊ฐ„์€ ์ƒ๋‹นํžˆ ์ข์•„์ง€๊ฒ ์ง€๋งŒ, ํ‘œ์ค€ ์˜ค์ฐจ๊ฐ€ ์ ์ ˆํžˆ ํด๋Ÿฌ์Šคํ„ฐ๋ง๋˜์—ˆ๋‹ค๋ฉด ๊ทธ ์ •๋„๋Š” ์•„๋‹ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ด๊ฒƒ์ด Hoffman ๋“ฑ(๋ถ€๋ก A.2 ์ฐธ์กฐ)์ด ๋ณด๊ณ ํ•œ ์ข์€ ๊ตฌ๊ฐ„์„ ์„ค๋ช…ํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
๊ฐ์ฃผ 4: ๋˜ ๋‹ค๋ฅธ ๊ฐ€๋Šฅ์„ฑ์€ ์ ‘๊ทผ ๋ฐฉ์‹ 3์— ๋Œ€ํ•œ ์„ค๋ช…๋Œ€๋กœ ์ตœ์ข… ์†์‹ค์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ํ›ˆ๋ จ ์ค‘์— ์ค‘๊ฐ„ ์†์‹ค์„ ์‚ฌ์šฉํ–ˆ์„ ์ˆ˜๋„ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ ‡๋‹ค๋ฉด ์‹ ๋ขฐ ๊ตฌ๊ฐ„์ด ํ›จ์”ฌ ์ข์•„์ง€๊ฒ ์ง€๋งŒ ํ‘œ์ค€ ์˜ค์ฐจ๋ฅผ ์ ์ ˆํžˆ ํด๋Ÿฌ์Šคํ„ฐ๋งํ•œ ๊ฒฝ์šฐ 10๋ฐฐ ์ •๋„๋Š” ์•„๋‹ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ Hoffman ๋“ฑ์ด ๋ณด๊ณ ํ•œ ๊ธด์ถ•๋œ ๊ฐ„๊ฒฉ์„ ์„ค๋ช…ํ•˜๊ธฐ์—๋Š” ์ด ๋ฐฉ๋ฒ•์œผ๋กœ๋Š” ์„ค๋ช…ํ•˜๊ธฐ ์–ด๋ ต๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค(๋ถ€๋ก A.2 ์ฐธ์กฐ).
As an example, the recruitment step allows agents to be removed or added based on the progress towards the goal. This helps ensure that the right agents are participating at any given stage of problem solving. The researchers found that horizontal teams are generally best suited for collaborative tasks like consulting, while vertical teams are better suited for tasks that require clearer isolation of responsibilities for tool calling.
์˜ˆ๋ฅผ ๋“ค์–ด, ๋ชจ์ง‘ ๋‹จ๊ณ„์—์„œ๋Š” ๋ชฉํ‘œ ๋‹ฌ์„ฑ ์ง„ํ–‰ ์ƒํ™ฉ์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ๋ฅผ ์ œ๊ฑฐํ•˜๊ฑฐ๋‚˜ ์ถ”๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ฌธ์ œ ํ•ด๊ฒฐ์˜ ๊ฐ ๋‹จ๊ณ„์—์„œ ์ ์ ˆํ•œ ์—์ด์ „ํŠธ๊ฐ€ ์ฐธ์—ฌํ•˜๋„๋ก ๋ณด์žฅํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ž๋“ค์€ ์ˆ˜ํ‰ ํŒ€์ด ์ผ๋ฐ˜์ ์œผ๋กœ ์ปจ์„คํŒ…๊ณผ ๊ฐ™์€ ํ˜‘์—… ์ž‘์—…์— ๊ฐ€์žฅ ์ ํ•ฉํ•˜๋ฉฐ, ์ˆ˜์ง ํŒ€์€ ๋„๊ตฌ ํ˜ธ์ถœ์— ๋Œ€ํ•œ ์ฑ…์ž„์˜ ๋ช…ํ™•ํ•œ ๋ถ„๋ฆฌ๊ฐ€ ํ•„์š”ํ•œ ์ž‘์—…์— ๋” ์ ํ•ฉํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
ํ•œ ์˜ˆ๋กœ, ๋ชจ์ง‘ ๋‹จ๊ณ„์—์„œ๋Š” ๋ชฉํ‘œ์— ๋Œ€ํ•œ ์ง„ํ–‰ ์ƒํ™ฉ์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ๋ฅผ ์ œ๊ฑฐํ•˜๊ฑฐ๋‚˜ ์ถ”๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ฌธ์ œ ํ•ด๊ฒฐ์˜ ํŠน์ • ๋‹จ๊ณ„์—์„œ ์ ์ ˆํ•œ ์—์ด์ „ํŠธ๊ฐ€ ์ฐธ์—ฌํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ง„์€ ์ˆ˜ํ‰ํ˜• ํŒ€์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ปจ์„คํŒ…๊ณผ ๊ฐ™์€ ํ˜‘์—… ์ž‘์—…์— ๊ฐ€์žฅ ์ ํ•ฉํ•œ ๋ฐ˜๋ฉด, ์ˆ˜์งํ˜• ํŒ€์€ ๋„๊ตฌ ํ˜ธ์ถœ์— ๋Œ€ํ•œ ์ฑ…์ž„์„ ๋” ๋ช…ํ™•ํ•˜๊ฒŒ ๋ถ„๋ฆฌํ•ด์•ผ ํ•˜๋Š” ์ž‘์—…์— ๋” ์ ํ•ฉํ•˜๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.