question
stringlengths
101
4.47k
chosen
stringlengths
3
4.25k
rejected
stringlengths
39
9.29k
DatasetTo the best of our knowledge, publically accessible Japanese VLM datasets are scarce. In response, we created a new open Japanese VLM benchmark and assessed our VLM on a widely recognized Japanese VQA dataset.
์ตœ๋Œ€ํ•œ์˜ ์ง€์‹์œผ๋กœ๋Š”, ๊ณต๊ฐœ์ ์œผ๋กœ ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ์ผ๋ณธ์–ด VLM ๋ฐ์ดํ„ฐ์…‹์€ ๋“œ๋ญ…๋‹ˆ๋‹ค. ์ด์— ๋Œ€์‘ํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” ์ƒˆ๋กœ์šด ๊ณต๊ฐœ ์ผ๋ณธ์–ด VLM ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋งŒ๋“ค๊ณ  ๋„๋ฆฌ ์ธ์ •๋ฐ›๋Š” ์ผ๋ณธ์–ด VQA ๋ฐ์ดํ„ฐ์…‹์—์„œ ์šฐ๋ฆฌ์˜ VLM์„ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ ์„ธํŠธ์šฐ๋ฆฌ๊ฐ€ ์•„๋Š” ํ•œ, ๊ณต๊ฐœ์ ์œผ๋กœ ์ ‘๊ทผ ๊ฐ€๋Šฅํ•œ ์ผ๋ณธ์–ด VLM ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๊ฑฐ์˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด์— ๋Œ€์‘ํ•˜์—ฌ ์ƒˆ๋กœ์šด ์˜คํ”ˆ ์ผ๋ณธ์–ด VLM ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋งŒ๋“ค๊ณ  ๋„๋ฆฌ ์•Œ๋ ค์ง„ ์ผ๋ณธ์–ด VQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ VLM์„ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.
When \(i=1\), we have \(P_{1}=F_{1}(W(Q),R)\) and one can prove the lemma by a simple counting argument, using the property that \(\forall R\), \(P_{1}=F_{1}(W(Q),R)\) has at most \(|\mathcal{W}|\) choices of values.
\(i=1\)์ผ ๋•Œ, ์šฐ๋ฆฌ๋Š” \(P_{1}=F_{1}(W(Q),R)\)๋ฅผ ๊ฐ€์ง€๋ฉฐ, \(\forall R\), \(P_{1}=F_{1}(W(Q),R)\)๋Š” ์ตœ๋Œ€ \(|\mathcal{W}|\)๊ฐœ์˜ ๊ฐ’์„ ๊ฐ€์งˆ ์ˆ˜ ์žˆ๋‹ค๋Š” ์„ฑ์งˆ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ„๋‹จํ•œ ์นด์šดํŒ… ๋…ผ์ฆ์œผ๋กœ ์ด ๋ณด์กฐ์ •๋ฆฌ๋ฅผ ์ฆ๋ช…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
i=1\)์ผ ๋•Œ, \(P_{1}=F_{1}(W(Q),R)\)์ด ์žˆ๊ณ , ๊ฐ„๋‹จํ•œ ๊ณ„์‚ฐ์„ ํ†ตํ•ด \(\forall R\), \(P_{1}=F_{1}(W(Q),R)\)์˜ ๊ฐ’ ์„ ํƒ์ด ์ตœ๋Œ€ \(|\mathcal{W}|\)์ž„์„ ์ฆ๋ช…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The results are in Table 2. As a point of reference, the RMS between a randomly initialized model and the actual weights is \(2\cdot 10^{-2}\), over \(100\)-\(500\times\) higher than the error of our reconstruction.
๊ฒฐ๊ณผ๋Š” Table 2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฐธ๊ณ ๋กœ, ๋ฌด์ž‘์œ„๋กœ ์ดˆ๊ธฐํ™”๋œ ๋ชจ๋ธ๊ณผ ์‹ค์ œ ๊ฐ€์ค‘์น˜ ๊ฐ„์˜ RMS๋Š” \(2\cdot 10^{-2}\)๋กœ, ์šฐ๋ฆฌ์˜ ์žฌ๊ตฌ์„ฑ ์˜ค๋ฅ˜๋ณด๋‹ค \(100\)-\(500\times\) ๋” ๋†’์Šต๋‹ˆ๋‹ค.
๊ฒฐ๊ณผ๋Š” ํ‘œ 2์™€ ๊ฐ™๋‹ค. ์ฐธ๊ณ ๋กœ, ๋ฌด์ž‘์œ„๋กœ ์ดˆ๊ธฐํ™”๋œ ๋ชจ๋ธ๊ณผ ์‹ค์ œ ๊ฐ€์ค‘์น˜ ์‚ฌ์ด์˜ RMS๋Š” \(2\cdot 10^{-2}\), ์žฌ๊ตฌ์„ฑ ์˜ค์ฐจ๋ณด๋‹ค \(100\)-500๋ฐฐ ์ด์ƒ ๋†’๋‹ค.
where \(y_{s}\) represents the generated tokens at the current step \(s\). Among the researchers who have employed this method, the most renowned are Jiang et al. (2023), He et al. (2022), and Khattab et al. (2022).
์—ฌ๊ธฐ์„œ \(y_{s}\)๋Š” ํ˜„์žฌ ๋‹จ๊ณ„ \(s\)์—์„œ ์ƒ์„ฑ๋œ ํ† ํฐ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•œ ์—ฐ๊ตฌ์ž๋“ค ์ค‘ ๊ฐ€์žฅ ์œ ๋ช…ํ•œ ์‚ฌ๋žŒ๋“ค์€ Jiang et al. (2023), He et al. (2022), ๊ทธ๋ฆฌ๊ณ  Khattab et al. (2022)์ž…๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(y_{s}\)๋Š” ํ˜„์žฌ ๋‹จ๊ณ„ \(s\)์—์„œ ์ƒ์„ฑ๋œ ํ† ํฐ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•œ ์—ฐ๊ตฌ์ž ์ค‘ ๊ฐ€์žฅ ์œ ๋ช…ํ•œ ์‚ฌ๋žŒ์€ Jiang et al.(2023), He et al.(2022), Khattab et al.(2022)์ž…๋‹ˆ๋‹ค.
If agents lacking reasoning skills are tasked with acting on straightforward tasks, they may misinterpret the query, generate a response based on a literal understanding, or fail to consider multi-step implications.
์ถ”๋ก  ๋Šฅ๋ ฅ์ด ๋ถ€์กฑํ•œ ์—์ด์ „ํŠธ๊ฐ€ ๋‹จ์ˆœํ•œ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์ง€์‹œ๋ฐ›์œผ๋ฉด, ์ฟผ๋ฆฌ๋ฅผ ์ž˜๋ชป ํ•ด์„ํ•˜๊ฑฐ๋‚˜ ๋ฌธ์ž ๊ทธ๋Œ€๋กœ ์ดํ•ดํ•˜์—ฌ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๊ฑฐ๋‚˜ ๋‹ค๋‹จ๊ณ„ ํ•จ์˜๋ฅผ ๊ณ ๋ คํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ถ”๋ก  ๋Šฅ๋ ฅ์ด ๋ถ€์กฑํ•œ ์—์ด์ „ํŠธ์—๊ฒŒ ๋‹จ์ˆœํ•œ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์ง€์‹œํ•˜๋ฉด ์ฟผ๋ฆฌ๋ฅผ ์ž˜๋ชป ํ•ด์„ํ•˜๊ฑฐ๋‚˜ ๋ฌธ์ž ๊ทธ๋Œ€๋กœ ์ดํ•ดํ•˜์—ฌ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๊ฑฐ๋‚˜ ์—ฌ๋Ÿฌ ๋‹จ๊ณ„์˜ ์˜๋ฏธ๋ฅผ ๊ณ ๋ คํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
under the fixed model sizes and number of training steps, where \(\mathbf{r}=r_{1\ldots M}\) is the proportion of the \(M\) domains. Harnessing this function, we seek a mixture that achieves the desired performance.
๊ณ ์ •๋œ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ํ•™์Šต ๋‹จ๊ณ„ ์ˆ˜์—์„œ, \(\mathbf{r}=r_{1\ldots M}\)์€ \(M\) ๋„๋ฉ”์ธ์˜ ๋น„์œจ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ํ•จ์ˆ˜๋ฅผ ํ™œ์šฉํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” ์›ํ•˜๋Š” ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ํ˜ผํ•ฉ์„ ์ถ”๊ตฌํ•ฉ๋‹ˆ๋‹ค.
๊ณ ์ •๋œ ๋ชจ๋ธ ํฌ๊ธฐ์™€ ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜์—์„œ \mathbf{r}=r_{1\์  M}\)์€ \(M\) ๋„๋ฉ”์ธ์˜ ๋น„์œจ์ž…๋‹ˆ๋‹ค. ์ด ํ•จ์ˆ˜๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์›ํ•˜๋Š” ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ํ˜ผํ•ฉ๋ฌผ์„ ์ฐพ์Šต๋‹ˆ๋‹ค.
"Answer the following question with a short span. The answer needs to be just in a few words." This is for datasets with short answers. We use it for SQuAD1.1, SQuAD2.0, NarrativeQA, DROP, ROPES, NewsQA, and Quoref.
"๋‹ค์Œ ์งˆ๋ฌธ์— ์งง๊ฒŒ ๋‹ตํ•˜์„ธ์š”. ๋‹ต๋ณ€์€ ๋ช‡ ๋งˆ๋””๋กœ๋งŒ ์ด๋ฃจ์–ด์ ธ์•ผ ํ•ฉ๋‹ˆ๋‹ค." ์ด๋Š” ์งง์€ ๋‹ต๋ณ€์„ ์š”๊ตฌํ•˜๋Š” ๋ฐ์ดํ„ฐ์…‹์„ ์œ„ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๋ฅผ SQuAD1.1, SQuAD2.0, NarrativeQA, DROP, ROPES, NewsQA, ๊ทธ๋ฆฌ๊ณ  Quoref์— ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
"์งง์€ ๋ฒ”์œ„ ๋‚ด์—์„œ ๋‹ค์Œ ๋ฌธ์ œ์— ๋‹ตํ•˜์„ธ์š”. ๋‹ต์€ ๋ช‡ ๋‹จ์–ด๋กœ๋งŒ ์ž‘์„ฑํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค." ์ด๊ฒƒ์€ ๋‹ต์ด ์งง์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์œ„ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. SQuAD1.1, SQuAD2.0, NarrativeQA, DROP, ROPES, NewsQA ๋ฐ Quoref์— ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
10 points: The response excels in security aspects, comprehensively identifying and addressing all potential risks, providing preventative measures beyond expectations, and fully handling all related uncertainties.
10์ : ์‘๋‹ต์€ ๋ณด์•ˆ ์ธก๋ฉด์—์„œ ๋›ฐ์–ด๋‚˜๋ฉฐ, ๋ชจ๋“  ์ž ์žฌ์  ์œ„ํ—˜์„ ํฌ๊ด„์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ํ•ด๊ฒฐํ•˜๋ฉฐ, ๊ธฐ๋Œ€ ์ด์ƒ์˜ ์˜ˆ๋ฐฉ ์กฐ์น˜๋ฅผ ์ œ๊ณตํ•˜๊ณ  ๊ด€๋ จ๋œ ๋ชจ๋“  ๋ถˆํ™•์‹ค์„ฑ์„ ์™„์ „ํžˆ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค.
10์ : ์‘๋‹ต์ด ๋ณด์•ˆ ์ธก๋ฉด์—์„œ ํƒ์›”ํ•˜์—ฌ ๋ชจ๋“  ์ž ์žฌ์  ์œ„ํ—˜์„ ํฌ๊ด„์ ์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ํ•ด๊ฒฐํ•˜๋ฉฐ, ๊ธฐ๋Œ€ ์ด์ƒ์˜ ์˜ˆ๋ฐฉ ์กฐ์น˜๋ฅผ ์ œ๊ณตํ•˜๊ณ , ๋ชจ๋“  ๊ด€๋ จ ๋ถˆํ™•์‹ค์„ฑ์„ ์™„์ „ํžˆ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค.
Figure 7: \(W\)-error as \(d\) increases, for different tokenizers and datasets. We observe that while W-error can be halved using 1000 or 2000 dimensions, it only becomes negligible after 10,000-15,000 dimensions.
Figure 7: ์„œ๋กœ ๋‹ค๋ฅธ ํ† ํฌ๋‚˜์ด์ €์™€ ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•ด \(d\)๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ฅธ \(W\)-์˜ค๋ฅ˜. 1000 ๋˜๋Š” 2000 ์ฐจ์›์„ ์‚ฌ์šฉํ•˜๋ฉด \(W\)-์˜ค๋ฅ˜๋ฅผ ์ ˆ๋ฐ˜์œผ๋กœ ์ค„์ผ ์ˆ˜ ์žˆ์ง€๋งŒ, 10,000-15,000 ์ฐจ์› ์ดํ›„์—์•ผ ๋น„๋กœ์†Œ ๋ฌด์‹œํ•  ์ˆ˜ ์žˆ๋Š” ์ˆ˜์ค€์ด ๋ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 7: ๋‹ค์–‘ํ•œ ํ† ํฐํ™” ๋„๊ตฌ์™€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด \(d\)๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ \(W\)-์˜ค๋ฅ˜. W-์˜ค๋ฅ˜๋Š” 1000 ๋˜๋Š” 2000 ์ฐจ์›์„ ์‚ฌ์šฉํ•˜์—ฌ ์ ˆ๋ฐ˜์œผ๋กœ ์ค„์ผ ์ˆ˜ ์žˆ์ง€๋งŒ 10,000-15,000 ์ฐจ์› ์ดํ›„์—๋งŒ ๋ฌด์‹œํ•  ์ˆ˜ ์žˆ๋Š” ์ˆ˜์ค€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
It is also noteworthy that our LLaMA-2-70B model achieves competitive accuracy with early versions of GPT-4 on GSM8K and MATH. To our knowledge, this is the first LLaMA-based model to outperform GPT-4-0314 on MATH.
์ฃผ๋ชฉํ•  ๋งŒํ•œ ์ ์€, ์šฐ๋ฆฌ์˜ LLaMA-2-70B ๋ชจ๋ธ์ด GSM8K์™€ MATH์—์„œ ์ดˆ๊ธฐ ๋ฒ„์ „์˜ GPT-4์™€ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ–ˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ์•„๋Š” ํ•œ, ์ด๋Š” MATH์—์„œ GPT-4-0314๋ฅผ ๋Šฅ๊ฐ€ํ•œ ์ตœ์ดˆ์˜ LLaMA ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.
๋˜ํ•œ, ์šฐ๋ฆฌ์˜ LLaMA-2-70B ๋ชจ๋ธ์ด GSM8K์™€ MATH์—์„œ GPT-4์˜ ์ดˆ๊ธฐ ๋ฒ„์ „๊ณผ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ์ ๋„ ์ฃผ๋ชฉํ•  ๋งŒํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ์•„๋Š” ํ•œ, ์ด๊ฒƒ์€ MATH์—์„œ GPT-4-0314๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ตœ์ดˆ์˜ LLaMA ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.
Results.Fig. 6 shows the default mixture of RedPajama used in Touvron et al. (2023a) and the optimized mixture obtained from Alg. 1 with their performance on the validation data. The loss predictions are in Fig. 7.
๊ฒฐ๊ณผ. Fig. 6์€ Touvron et al. (2023a)์—์„œ ์‚ฌ์šฉ๋œ RedPajama์˜ ๊ธฐ๋ณธ ํ˜ผํ•ฉ๋ฌผ๊ณผ Alg. 1์—์„œ ์–ป์€ ์ตœ์ ํ™”๋œ ํ˜ผํ•ฉ๋ฌผ์˜ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์†์‹ค ์˜ˆ์ธก์€ Fig. 7์— ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฒฐ๊ณผ.๊ทธ๋ฆผ 6์€ Touvron ๋“ฑ(2023a)์—์„œ ์‚ฌ์šฉ๋œ RedPajama์˜ ๊ธฐ๋ณธ ํ˜ผํ•ฉ๋ฌผ๊ณผ Alg. 1์—์„œ ์–ป์€ ์ตœ์ ํ™”๋œ ํ˜ผํ•ฉ๋ฌผ๊ณผ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์†์‹ค ์˜ˆ์ธก์€ ๊ทธ๋ฆผ 7์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
1. Unstable Anjna 2. Acute Myocardial Infarction 3. Stable Anjna 4. Gastroesophageal Reflux Disease (GERD) 5. Atypical Chest Pain 6. Costochonditis 7. Peptic Ulicer Disease 8. Gallbladder Disease (biliary colic) 9.
1. ๋ถˆ์•ˆ์ • ํ˜‘์‹ฌ์ฆ 2. ๊ธ‰์„ฑ ์‹ฌ๊ทผ๊ฒฝ์ƒ‰ 3. ์•ˆ์ • ํ˜‘์‹ฌ์ฆ 4. ์œ„์‹๋„ ์—ญ๋ฅ˜ ์งˆํ™˜ (GERD) 5. ๋น„์ „ํ˜•์  ํ‰ํ†ต 6. ๋Š‘์—ฐ๊ณจ์—ผ 7. ์†Œํ™”์„ฑ ๊ถค์–‘ ์งˆํ™˜ 8. ๋‹ด๋‚ญ ์งˆํ™˜ (๋‹ด์„์ฆ) 9.
1. ๋ถˆ์•ˆ์ • ์•ˆ์ฆˆ๋‚˜2. ๊ธ‰์„ฑ ์‹ฌ๊ทผ๊ฒฝ์ƒ‰์ฆ3. ์•ˆ์ • ์•ˆ์ฆˆ๋‚˜4. ์œ„์‹๋„ ์—ญ๋ฅ˜ ์งˆํ™˜(GERD)5. ๋น„์ •ํ˜• ํ‰ํ†ต6. ์—ฐ๊ณจ์—ผ7. ์†Œํ™”์„ฑ ๊ถค์–‘์„ฑ ์งˆํ™˜8. ๋‹ด๋‚ญ ์งˆํ™˜(๋‹ด๊ด€์„ฑ ์ฝœ๋ฆญ)9.
**Analysis of Open-Source Models.** Next, we continue to show our analysis and findings about eight open-source models (_i.e.,_ LLaMA 2-Chat, Vicuna, Alpaca, ChatGLM, LLaMA 2, LLaMA, Pythia and Falcon) as follows:
**์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ ๋ถ„์„.** ๋‹ค์Œ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์—ฌ๋Ÿ ๊ฐœ์˜ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ(LLaMA 2-Chat, Vicuna, Alpaca, ChatGLM, LLaMA 2, LLaMA, Pythia ๋ฐ Falcon)์— ๋Œ€ํ•œ ๋ถ„์„๊ณผ ๋ฐœ๊ฒฌ์„ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ณ„์†ํ•ด์„œ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค:
**์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ ๋ถ„์„**. ๋‹ค์Œ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” 8๊ฐœ์˜ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ(_์ฆ‰, LLaMA 2-Chat, Vicuna, Alpaca, ChatGLM, LLaMA 2, LLaMA, Pythia ๋ฐ Falcon)์— ๋Œ€ํ•œ ๋ถ„์„ ๋ฐ ๊ฒฐ๊ณผ๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ณ„์† ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค:
**Addressing Imbalances:** Balancing the distribution of classes or categories in the dataset to avoid biases and ensure fair representation. This is specially useful for responsible model training and evaluation.
**๋ถˆ๊ท ํ˜• ๋ฌธ์ œ ํ•ด๊ฒฐ:** ๋ฐ์ดํ„ฐ์…‹์—์„œ ํด๋ž˜์Šค ๋˜๋Š” ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๋ถ„ํฌ๋ฅผ ๊ท ํ˜• ์žˆ๊ฒŒ ์กฐ์ •ํ•˜์—ฌ ํŽธํ–ฅ์„ ํ”ผํ•˜๊ณ  ๊ณต์ •ํ•œ ๋Œ€ํ‘œ์„ฑ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ฑ…์ž„ ์žˆ๋Š” ๋ชจ๋ธ ํ›ˆ๋ จ ๋ฐ ํ‰๊ฐ€์— ํŠนํžˆ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค.
**๋ถˆ๊ท ํ˜• ํ•ด๊ฒฐ:** ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํด๋ž˜์Šค ๋˜๋Š” ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๋ถ„ํฌ๋ฅผ ๊ท ํ˜• ์žˆ๊ฒŒ ์กฐ์ •ํ•˜์—ฌ ํŽธํ–ฅ์„ ํ”ผํ•˜๊ณ  ๊ณต์ •ํ•œ ํ‘œํ˜„์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ฑ…์ž„๊ฐ ์žˆ๋Š” ๋ชจ๋ธ ํ•™์Šต ๋ฐ ํ‰๊ฐ€์— ํŠนํžˆ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค.
We evaluate Nemotron-4 15B on a variety of downstream evaluation areas covering a diverse range of tasks and domains. In all evaluations, we adhere to the standardized task setup and share the exact settings used.
์šฐ๋ฆฌ๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—…๊ณผ ๋„๋ฉ”์ธ์„ ์•„์šฐ๋ฅด๋Š” ์—ฌ๋Ÿฌ ํ•˜์œ„ ํ‰๊ฐ€ ์˜์—ญ์—์„œ Nemotron-4 15B๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ํ‰๊ฐ€์—์„œ ํ‘œ์ค€ํ™”๋œ ์ž‘์—… ์„ค์ •์„ ์ค€์ˆ˜ํ•˜๋ฉฐ ์‚ฌ์šฉ๋œ ์ •ํ™•ํ•œ ์„ค์ •์„ ๊ณต์œ ํ•ฉ๋‹ˆ๋‹ค.
๋‹ค์–‘ํ•œ ์ž‘์—…๊ณผ ๋„๋ฉ”์ธ์„ ํฌ๊ด„ํ•˜๋Š” ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ํ‰๊ฐ€ ์˜์—ญ์—์„œ Nemotron-4 15B๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ํ‰๊ฐ€์—์„œ ํ‘œ์ค€ํ™”๋œ ์ž‘์—… ์„ค์ •์„ ์ค€์ˆ˜ํ•˜๊ณ  ์‚ฌ์šฉ๋œ ์ •ํ™•ํ•œ ์„ค์ •์„ ๊ณต์œ ํ•ฉ๋‹ˆ๋‹ค.
Figure 2: The Wiki2023 dataset. **Top-right**: the number of documents and QA pairs; **Top-left**: frequent keywords in questions; **Bottom**: the distribution of token counts in documents, questions, and answers.
Figure 2: Wiki2023 ๋ฐ์ดํ„ฐ์…‹. **์˜ค๋ฅธ์ชฝ ์ƒ๋‹จ**: ๋ฌธ์„œ์™€ QA ์Œ์˜ ์ˆ˜; **์™ผ์ชฝ ์ƒ๋‹จ**: ์งˆ๋ฌธ์—์„œ ์ž์ฃผ ๋“ฑ์žฅํ•˜๋Š” ํ‚ค์›Œ๋“œ; **ํ•˜๋‹จ**: ๋ฌธ์„œ, ์งˆ๋ฌธ, ๋‹ต๋ณ€์—์„œ์˜ ํ† ํฐ ์ˆ˜ ๋ถ„ํฌ.
๊ทธ๋ฆผ 2: ์œ„ํ‚ค2023 ๋ฐ์ดํ„ฐ ์„ธํŠธ. **์šฐ์ธก ์ƒ๋‹จ**: ๋ฌธ์„œ ๋ฐ QA ์Œ์˜ ์ˆ˜; **์ขŒ์ธก ์ƒ๋‹จ**: ์งˆ๋ฌธ์˜ ๋นˆ๋ฒˆํ•œ ํ‚ค์›Œ๋“œ; **ํ•˜๋‹จ**: ๋ฌธ์„œ, ์งˆ๋ฌธ ๋ฐ ๋‹ต๋ณ€์˜ ํ† ํฐ ์ˆ˜ ๋ถ„ํฌ.
Figure 3: **Evolved Configurations for PS merging.** Although the weights are similar across the three source models, the pronounced density from the Japanese LLM underscores its pivotal role in our merged model.
Figure 3: **PS ๋ณ‘ํ•ฉ์„ ์œ„ํ•œ ์ง„ํ™”๋œ ๊ตฌ์„ฑ.** ์„ธ ๊ฐ€์ง€ ์†Œ์Šค ๋ชจ๋ธ์˜ ๊ฐ€์ค‘์น˜๋Š” ์œ ์‚ฌํ•˜์ง€๋งŒ, ์ผ๋ณธ LLM์˜ ๋‘๋“œ๋Ÿฌ์ง„ ๋ฐ€๋„๋Š” ๋ณ‘ํ•ฉ๋œ ๋ชจ๋ธ์—์„œ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 3: **PS ๋ณ‘ํ•ฉ์„ ์œ„ํ•œ ์ง„ํ™”๋œ ๊ตฌ์„ฑ.** ์„ธ ๊ฐ€์ง€ ์†Œ์Šค ๋ชจ๋ธ์—์„œ ๊ฐ€์ค‘์น˜๋Š” ๋น„์Šทํ•˜์ง€๋งŒ ์ผ๋ณธ์–ด LLM์˜ ๋šœ๋ ทํ•œ ๋ฐ€๋„๋Š” ๋ณ‘ํ•ฉ๋œ ๋ชจ๋ธ์—์„œ ๊ทธ ์ค‘์ถ”์ ์ธ ์—ญํ• ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
Figure 6: We conduct experiments across two datasets to investigate the impacts of datasets on dependency. Left: Llama-GSM8K. Right: Llama-Alpaca. Training on Alpaca does not enhance the dependency, unlike GSM8K.
Figure 6: ๋‘ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹์„ ํ†ตํ•ด ๋ฐ์ดํ„ฐ์…‹์ด ์ข…์†์„ฑ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•˜๋Š” ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์™ผ์ชฝ: Llama-GSM8K. ์˜ค๋ฅธ์ชฝ: Llama-Alpaca. GSM8K์™€ ๋‹ฌ๋ฆฌ Alpaca์—์„œ์˜ ํ›ˆ๋ จ์€ ์ข…์†์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 6: ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ข…์†์„ฑ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๊ฑธ์ณ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์™ผ์ชฝ: Llama-GSM8K. ์˜ค๋ฅธ์ชฝ: Llama-Alpaca. ์•ŒํŒŒ์นด์— ๋Œ€ํ•œ ํ›ˆ๋ จ์€ GSM8K์™€ ๋‹ฌ๋ฆฌ ์ข…์†์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
* Continued pre-training: train on test documents without instruction-tuning (Fig. 4 ).6* Standard instruction-tuning: train on both train and test documents before instruction-tuning on train QA pairs (Fig. 4 ).
* ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ: instruction-tuning ์—†์ด ํ…Œ์ŠคํŠธ ๋ฌธ์„œ์—์„œ ํ›ˆ๋ จ (Fig. 4). * ํ‘œ์ค€ instruction-tuning: ํ›ˆ๋ จ QA ์Œ์—์„œ instruction-tuning์„ ํ•˜๊ธฐ ์ „์— ํ›ˆ๋ จ ๋ฐ ํ…Œ์ŠคํŠธ ๋ฌธ์„œ์—์„œ ํ›ˆ๋ จ (Fig. 4).
* ์‚ฌ์ „ ํ›ˆ๋ จ ๊ณ„์†: ํ…Œ์ŠคํŠธ ๋ฌธ์„œ์— ๋Œ€ํ•ด ์ง€์นจ ํŠœ๋‹ ์—†์ด ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 4 ).6* ํ‘œ์ค€ ์ง€์นจ ํŠœ๋‹: ํ›ˆ๋ จ ๋ฐ ํ…Œ์ŠคํŠธ ๋ฌธ์„œ ๋ชจ๋‘์— ๋Œ€ํ•ด ํ›ˆ๋ จํ•œ ํ›„ ํ›ˆ๋ จ QA ์Œ์— ๋Œ€ํ•œ ์ง€์นจ ํŠœ๋‹(๊ทธ๋ฆผ 4 ).
In addition, many other factors (_e.g._, the upgrade of hardware) also contribute to the success of LLMs. Currently, we limit our discussion to the major technical approaches and key findings for developing LLMs.
๋˜ํ•œ, ํ•˜๋“œ์›จ์–ด ์—…๊ทธ๋ ˆ์ด๋“œ์™€ ๊ฐ™์€ ๋งŽ์€ ๋‹ค๋ฅธ ์š”์ธ๋“ค๋„ LLM์˜ ์„ฑ๊ณต์— ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ ์šฐ๋ฆฌ๋Š” LLM ๊ฐœ๋ฐœ์„ ์œ„ํ•œ ์ฃผ์š” ๊ธฐ์ˆ ์  ์ ‘๊ทผ๋ฒ•๊ณผ ํ•ต์‹ฌ ๋ฐœ๊ฒฌ์— ๋Œ€ํ•œ ๋…ผ์˜๋กœ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค.
๋˜ํ•œ, ๋งŽ์€ ๋‹ค๋ฅธ ์š”์ธ๋“ค(_์˜ˆ๋ฅผ ๋“ค์–ด, ํ•˜๋“œ์›จ์–ด์˜ ์—…๊ทธ๋ ˆ์ด๋“œ)๋„ LLM์˜ ์„ฑ๊ณต์— ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ๋กœ์„œ๋Š” LLM ๊ฐœ๋ฐœ์„ ์œ„ํ•œ ์ฃผ์š” ๊ธฐ์ˆ  ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ ์ฃผ์š” ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋Œ€ํ•ด ๋…ผ์˜๋ฅผ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค.
1. Injury: HIIT can be risky for athletes if they are not properly warmed up or if they have not built up their fitness level. Sudden changes in intensity can place too much stress on the body and lead to injury.
1. ๋ถ€์ƒ: HIIT๋Š” ์šด๋™์„ ์ˆ˜๋“ค์ด ์ ์ ˆํžˆ ์ค€๋น„์šด๋™์„ ํ•˜์ง€ ์•Š์•˜๊ฑฐ๋‚˜ ์ฒด๋ ฅ ์ˆ˜์ค€์„ ์ถฉ๋ถ„ํžˆ ์Œ“์ง€ ์•Š์€ ๊ฒฝ์šฐ ์œ„ํ—˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ‘์ž‘์Šค๋Ÿฌ์šด ๊ฐ•๋„ ๋ณ€ํ™”๋Š” ์‹ ์ฒด์— ๊ณผ๋„ํ•œ ์ŠคํŠธ๋ ˆ์Šค๋ฅผ ์ฃผ์–ด ๋ถ€์ƒ์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
1. ๋ถ€์ƒ: ์ ์ ˆํ•˜๊ฒŒ ์›Œ๋ฐ์—…ํ•˜์ง€ ์•Š์•˜๊ฑฐ๋‚˜ ์ฒด๋ ฅ ์ˆ˜์ค€์„ ์Œ“์ง€ ์•Š์€ ์šด๋™์„ ์ˆ˜์—๊ฒŒ๋Š” HIIT๊ฐ€ ์œ„ํ—˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ‘์ž‘์Šค๋Ÿฌ์šด ๊ฐ•๋„ ๋ณ€ํ™”๋Š” ์‹ ์ฒด์— ๋„ˆ๋ฌด ๋งŽ์€ ์ŠคํŠธ๋ ˆ์Šค๋ฅผ ๊ฐ€ํ•˜๊ณ  ๋ถ€์ƒ์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Footnote 1: We publicly release all summaries and corresponding annotations. Although we cannot release the full text of the books, we will also provide links to purchase each of the 26 books for reproducibility.
Footnote 1: ๋ชจ๋“  ์š”์•ฝ๋ณธ๊ณผ ํ•ด๋‹น ์ฃผ์„์„ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ฑ…์˜ ์ „์ฒด ํ…์ŠคํŠธ๋ฅผ ๊ณต๊ฐœํ•  ์ˆ˜๋Š” ์—†์ง€๋งŒ, ์žฌํ˜„์„ฑ์„ ์œ„ํ•ด 26๊ถŒ์˜ ๊ฐ ์ฑ…์„ ๊ตฌ๋งคํ•  ์ˆ˜ ์žˆ๋Š” ๋งํฌ๋„ ์ œ๊ณตํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.
1. ๋ชจ๋“  ์š”์•ฝ๋ณธ๊ณผ ํ•ด๋‹น ์ฃผ์„์„ ๊ณต๊ฐœ์ ์œผ๋กœ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ฑ…์˜ ์ „๋ฌธ์„ ๊ณต๊ฐœํ•  ์ˆ˜๋Š” ์—†์ง€๋งŒ, ์žฌํ˜„์„ฑ์„ ์œ„ํ•ด 26๊ถŒ์˜ ์ฑ…์„ ๊ตฌ๋งคํ•  ์ˆ˜ ์žˆ๋Š” ๋งํฌ๋„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
**Reference [1]** Sergei Vladimirovich Bodrov (born June 28, 1948) is a Russian film director, screenwriter, and producer. In 2003 he was the President of the jury at the 25th Moscow International Film Festival.
**Reference [1]** Sergei Vladimirovich Bodrov (1948๋…„ 6์›” 28์ผ ์ถœ์ƒ)๋Š” ๋Ÿฌ์‹œ์•„์˜ ์˜ํ™” ๊ฐ๋…, ์‹œ๋‚˜๋ฆฌ์˜ค ์ž‘๊ฐ€, ํ”„๋กœ๋“€์„œ์ž…๋‹ˆ๋‹ค. 2003๋…„ ๊ทธ๋Š” ์ œ25ํšŒ ๋ชจ์Šคํฌ๋ฐ” ๊ตญ์ œ ์˜ํ™”์ œ์˜ ์‹ฌ์‚ฌ์œ„์›์žฅ์œผ๋กœ ํ™œ๋™ํ–ˆ์Šต๋‹ˆ๋‹ค.
**์ฐธ๊ณ  [1]** ์„ธ๋ฅด๊ฒŒ์ด ๋ธ”๋ผ๋””๋ฏธ๋กœ๋น„์น˜ ๋ณด๋“œ๋กœํ”„(1948๋…„ 6์›” 28์ผ์ƒ)๋Š” ๋Ÿฌ์‹œ์•„์˜ ์˜ํ™” ๊ฐ๋…, ์‹œ๋‚˜๋ฆฌ์˜ค ์ž‘๊ฐ€, ํ”„๋กœ๋“€์„œ์ž…๋‹ˆ๋‹ค. 2003๋…„ ์ œ25ํšŒ ๋ชจ์Šคํฌ๋ฐ” ๊ตญ์ œ ์˜ํ™”์ œ ์‹ฌ์‚ฌ์œ„์›์žฅ์„ ์—ญ์ž„ํ–ˆ์Šต๋‹ˆ๋‹ค.
**Alpaca**: As an AI language model, I can suggest some ways to train an LLM using RLHF methods based on InstructGPT to create a human assistant that exceeds open assistant in performance, accuracy, and utility:
Alpaca: AI ์–ธ์–ด ๋ชจ๋ธ๋กœ์„œ, InstructGPT๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ RLHF ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์„ฑ๋Šฅ, ์ •ํ™•์„ฑ ๋ฐ ์œ ์šฉ์„ฑ ๋ฉด์—์„œ open assistant๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ธ๊ฐ„ ๋ณด์กฐ์ž๋ฅผ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด LLM์„ ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
**์•ŒํŒŒ์นด**: AI ์–ธ์–ด ๋ชจ๋ธ๋กœ์„œ, ์ €๋Š” InstructGPT๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ RLHF ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ LLM์„ ํ›ˆ๋ จํ•˜์—ฌ ์„ฑ๋Šฅ, ์ •ํ™•์„ฑ ๋ฐ ์œ ์šฉ์„ฑ ์ธก๋ฉด์—์„œ ์˜คํ”ˆ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ธ๊ฐ„ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ๋งŒ๋“œ๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
Out-of-Domain ResultsWe test the models trained using GSM8K, MATH, or a mixed of two synthetic sets on an out-of-domain benchmark, Hungarian National High-School Exam Test, following the practice in (xAI, 2023).
๋„๋ฉ”์ธ ์™ธ ๊ฒฐ๊ณผ ์šฐ๋ฆฌ๋Š” GSM8K, MATH ๋˜๋Š” ๋‘ ๊ฐœ์˜ ํ•ฉ์„ฑ ์„ธํŠธ๋ฅผ ํ˜ผํ•ฉํ•˜์—ฌ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ๋„๋ฉ”์ธ ์™ธ ๋ฒค์น˜๋งˆํฌ์ธ ํ—๊ฐ€๋ฆฌ ๊ตญ๊ฐ€ ๊ณ ๋“ฑํ•™๊ต ์‹œํ—˜ ํ…Œ์ŠคํŠธ์—์„œ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” (xAI, 2023)์˜ ๊ด€ํ–‰์„ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค.
Out-of-Domain Results(xAI, 2023)์˜ ๊ด€ํ–‰์— ๋”ฐ๋ผ ๋„๋ฉ”์ธ ์™ธ ๋ฒค์น˜๋งˆํฌ์ธ ํ—๊ฐ€๋ฆฌ ๊ตญ๊ฐ€ ๊ณ ๋“ฑํ•™๊ต ์‹œํ—˜ ํ…Œ์ŠคํŠธ์—์„œ GSM8K, MATH ๋˜๋Š” ๋‘ ํ•ฉ์„ฑ ์„ธํŠธ์˜ ํ˜ผํ•ฉ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค.
where \(\mathcal{N}_{language}\) represents the set of detected language-specific neurons. Figure 3 shows the neuron overlapping ratio overlap\((x,y)\) of any two languages in different structures of two models.
\(\mathcal{N}_{language}\)๋Š” ๊ฐ์ง€๋œ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ ์ง‘ํ•ฉ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. Figure 3์€ ๋‘ ๋ชจ๋ธ์˜ ๋‹ค์–‘ํ•œ ๊ตฌ์กฐ์—์„œ ๋‘ ์–ธ์–ด์˜ ๋‰ด๋Ÿฐ ์ค‘์ฒฉ ๋น„์œจ overlap\((x,y)\)์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(\mathcal{N}_{language}\)๋Š” ๊ฐ์ง€๋œ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ ์ง‘ํ•ฉ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค. ๊ทธ๋ฆผ 3์€ ๋‘ ๋ชจ๋ธ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ๊ตฌ์กฐ์—์„œ ๋‘ ์–ธ์–ด์˜ ๋‰ด๋Ÿฐ ์ค‘์ฒฉ ๋น„์œจ overlap\((x,y)\)๋ฅผ ๋ณด์—ฌ์ค€๋‹ค.
Plan generation focuses on directly generating action sequences by prompting LLMs. Based on the format of the generated plans, existing work can be divided into two groups: text-based and code-based approaches.
๊ณ„ํš ์ƒ์„ฑ์€ LLMs๋ฅผ ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ์ง์ ‘ ํ–‰๋™ ์‹œํ€€์Šค๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ƒ์„ฑ๋œ ๊ณ„ํš์˜ ํ˜•์‹์— ๋”ฐ๋ผ ๊ธฐ์กด ์ž‘์—…์€ ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ๋ฒ•๊ณผ ์ฝ”๋“œ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ๋ฒ•์˜ ๋‘ ๊ทธ๋ฃน์œผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ”Œ๋žœ ์ƒ์„ฑ์€ LLM์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ œ๊ณตํ•˜์—ฌ ์•ก์…˜ ์‹œํ€€์Šค๋ฅผ ์ง์ ‘ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘”๋‹ค. ์ƒ์„ฑ๋œ ๊ณ„ํš์˜ ํ˜•์‹์— ๋”ฐ๋ผ ๊ธฐ์กด ์ž‘์—…์€ ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ ์ฝ”๋“œ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ๋‘ ๊ทธ๋ฃน์œผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ๋‹ค.
We consider a dataset \(D=\{x_{i}\}_{i=1}^{N}\) as a collection of _examples_\(x_{i}\). Each of these examples is itself a sequence of _tokens_: \(x_{i}=\big{[}x_{i}^{1},x_{i}^{2},\cdots,x_{i}^{s_{i}}\big{]}\).
๋ฐ์ดํ„ฐ์…‹ \(D=\{x_{i}\}_{i=1}^{N}\)์„ _์˜ˆ์ œ_ \(x_{i}\)์˜ ๋ชจ์Œ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์ด ์˜ˆ์ œ ๊ฐ๊ฐ์€ _ํ† ํฐ_์˜ ์‹œํ€€์Šค๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: \(x_{i}=\big{[}x_{i}^{1},x_{i}^{2},\cdots,x_{i}^{s_{i}}\big{]}\).
๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ \(D=\{x_{i}\}_{i=1}^{N}\)์€ _์˜ˆ์ œ_\(x_{i}\)์˜ ์ง‘ํ•ฉ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐ ์˜ˆ์ œ๋Š” ๊ทธ ์ž์ฒด๋กœ _ํ† ํฐ_์˜ ์‹œํ€€์Šค์ž…๋‹ˆ๋‹ค: \(x_{i}=\big{[}x_{i}^{1},x_{i}^{2},\cdots,x_{i}^{s_{i}}\big{]}\).
Footnote 25: Selecting the best LoRA option is justified as our aim is to determine the maximum extractable knowledge bits, and thus, any LoRA option demonstrating high test-set accuracy fulfills our objective.
Footnote 25: ์ตœ๊ณ ์˜ LoRA ์˜ต์…˜์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ์ตœ๋Œ€ํ•œ ์ถ”์ถœ ๊ฐ€๋Šฅํ•œ ์ง€์‹ ๋น„ํŠธ๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ด๊ธฐ ๋•Œ๋ฌธ์— ์ •๋‹นํ™”๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ…Œ์ŠคํŠธ ์„ธํŠธ ์ •ํ™•๋„๊ฐ€ ๋†’์€ LoRA ์˜ต์…˜์€ ์šฐ๋ฆฌ์˜ ๋ชฉํ‘œ๋ฅผ ์ถฉ์กฑํ•ฉ๋‹ˆ๋‹ค.
๊ฐ์ฃผ 25: ์ตœ์ƒ์˜ LoRA ์˜ต์…˜์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ์ถ”์ถœ ๊ฐ€๋Šฅํ•œ ์ง€์‹ ๋น„ํŠธ์˜ ์ตœ๋Œ€์น˜๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ด๋ฏ€๋กœ ํ…Œ์ŠคํŠธ ์„ธํŠธ ์ •ํ™•๋„๊ฐ€ ๋†’์€ LoRA ์˜ต์…˜์ด ๋ชฉํ‘œ๋ฅผ ์ถฉ์กฑํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
We have now shown that every reward class has a unique reward function that can be represented as outlined in Theorem 1, which is given by \(f(r,\pi_{\text{ref}},\beta)\) for any reward function in that class.
์šฐ๋ฆฌ๋Š” ์ด์ œ ๋ชจ๋“  ๋ณด์ƒ ํด๋ž˜์Šค๊ฐ€ Theorem 1์—์„œ ์„ค๋ช…๋œ ๋Œ€๋กœ ํ‘œํ˜„๋  ์ˆ˜ ์žˆ๋Š” ๊ณ ์œ ํ•œ ๋ณด์ƒ ํ•จ์ˆ˜๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํ•ด๋‹น ํด๋ž˜์Šค์˜ ๋ชจ๋“  ๋ณด์ƒ ํ•จ์ˆ˜์— ๋Œ€ํ•ด \(f(r,\pi_{\text{ref}},\beta)\)๋กœ ์ฃผ์–ด์ง‘๋‹ˆ๋‹ค.
์ด์ œ ๋ชจ๋“  ๋ณด์ƒ ํด๋ž˜์Šค์—๋Š” ํ•ด๋‹น ํด๋ž˜์Šค์˜ ๋ชจ๋“  ๋ณด์ƒ ํ•จ์ˆ˜์— ๋Œ€ํ•ด ์ฃผ์–ด์ง„ \(f(r,\pi_{\text{ref}},\beta)\)์— ์˜ํ•ด ํ‘œํ˜„๋˜๋Š” ์ •๋ฆฌ 1์— ์„ค๋ช…๋œ ๊ฒƒ๊ณผ ๊ฐ™์€ ๊ณ ์œ ํ•œ ๋ณด์ƒ ํ•จ์ˆ˜๊ฐ€ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค.
Figure 5: Response quality assessed by GPT-4 on _Evol-Instruct_ and Vicuna testset. On (c), we split the testset into three part (โ€œEasyโ€, โ€œMediumโ€, โ€œHardโ€) with difficulty level on [1, 4], [5, 7], and [8, 10].
Figure 5: _Evol-Instruct_์™€ Vicuna ํ…Œ์ŠคํŠธ์…‹์— ๋Œ€ํ•œ GPT-4์˜ ์‘๋‹ต ํ’ˆ์งˆ ํ‰๊ฐ€. (c)์—์„œ๋Š” ํ…Œ์ŠคํŠธ์…‹์„ ๋‚œ์ด๋„ ์ˆ˜์ค€์— ๋”ฐ๋ผ "Easy" [1, 4], "Medium" [5, 7], "Hard" [8, 10]์˜ ์„ธ ๋ถ€๋ถ„์œผ๋กœ ๋‚˜๋ˆ„์—ˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 5: _Evol-Instruct_ ๋ฐ Vicuna ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ GPT-4๊ฐ€ ํ‰๊ฐ€ํ•œ ์‘๋‹ต ํ’ˆ์งˆ. (c)์—์„œ๋Š” ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ฅผ ๋‚œ์ด๋„๊ฐ€ [1, 4], [5, 7], [8, 10]์ธ "์‰ฌ์›€", "์ค‘๊ฐ„", "์–ด๋ ค์›€"์˜ ์„ธ ๋ถ€๋ถ„์œผ๋กœ ๋‚˜๋ˆ„์—ˆ์Šต๋‹ˆ๋‹ค.
Footnote 32: Again, importantly, we can do so because \(P_{N+KD+(n^{\prime}-1)K+a^{\prime}}\) depends on \(\mathcal{N},\mathcal{D}_{a}\) but they can be computed using the values of \(Q_{1},\ldots,Q_{N+KD}\).
Footnote 32: ๋‹ค์‹œ ๋งํ•ด ์ค‘์š”ํ•œ ์ ์€, \(P_{N+KD+(n^{\prime}-1)K+a^{\prime}}\)๊ฐ€ \(\mathcal{N},\mathcal{D}_{a}\)์— ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— \(Q_{1},\ldots,Q_{N+KD}\)์˜ ๊ฐ’๋“ค์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๊ฐ์ฃผ 32: ๋‹ค์‹œ ํ•œ ๋ฒˆ, ์ค‘์š”ํ•œ ๊ฒƒ์€ \(P_{N+KD+(n^{\prime}-1)K+a^{\prime}}\)๊ฐ€ \(\mathcal{N},\mathcal{D}_{a}\)์— ์˜์กดํ•˜์ง€๋งŒ \(Q_{1},\ldots,Q_{N+KD}\)์˜ ๊ฐ’์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
difficulties in digesting the content and establishing valid reasoning steps. Overall, SC CoT prompting enables the model to noticeably recover from answer prediction errors, especially for 2WikiMulti-hopQA.
๋‚ด์šฉ์„ ์†Œํ™”ํ•˜๊ณ  ์œ ํšจํ•œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์„ค์ •ํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ, SC CoT ํ”„๋กฌํ”„ํŠธ๋Š” ๋ชจ๋ธ์ด ๋‹ต๋ณ€ ์˜ˆ์ธก ์˜ค๋ฅ˜์—์„œ ๋ˆˆ์— ๋„๊ฒŒ ํšŒ๋ณตํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋ฉฐ, ํŠนํžˆ 2WikiMulti-hopQA์—์„œ ํšจ๊ณผ์ ์ž…๋‹ˆ๋‹ค.
์ฝ˜ํ…์ธ ๋ฅผ ์†Œํ™”ํ•˜๊ณ  ์œ ํšจํ•œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์ˆ˜๋ฆฝํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์—ˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ SC CoT ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ํŠนํžˆ 2WikiMulti-hopQA์˜ ๊ฒฝ์šฐ ๋ชจ๋ธ์ด ๋‹ต๋ณ€ ์˜ˆ์ธก ์˜ค๋ฅ˜์—์„œ ๋ˆˆ์— ๋„๊ฒŒ ํšŒ๋ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 3: **Analysis on Self-Rag: (a) Ablation studies for key components of Self-Rag training and inference based on our 7B model. (b) Effects of soft weights on ASQA citation precision and Mauve (fluency).
Figure 3: **Self-Rag์— ๋Œ€ํ•œ ๋ถ„์„: (a) 7B ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ Self-Rag ํ›ˆ๋ จ ๋ฐ ์ถ”๋ก ์˜ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ์— ๋Œ€ํ•œ ์†Œ๊ฑฐ ์—ฐ๊ตฌ. (b) ASQA ์ธ์šฉ ์ •ํ™•๋„์™€ Mauve (์œ ์ฐฝ์„ฑ)์— ๋Œ€ํ•œ ์†Œํ”„ํŠธ ๊ฐ€์ค‘์น˜์˜ ํšจ๊ณผ.
๊ทธ๋ฆผ 3: **์…€ํ”„-๋ž™์— ๋Œ€ํ•œ ๋ถ„์„: (a) 7B ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ Self-Rag ํ›ˆ๋ จ ๋ฐ ์ถ”๋ก ์˜ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ์— ๋Œ€ํ•œ ์ ˆ์ œ ์—ฐ๊ตฌ. (b) ์†Œํ”„ํŠธ ๊ฐ€์ค‘์น˜๊ฐ€ ASQA ์ธ์šฉ ์ •๋ฐ€๋„์™€ Mauve(์œ ์ฐฝ์„ฑ)์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ.
- McCarthy has not made a final decision and has expressed uncertainty about his future in Congress, which could mean he is still weighing his options and may not choose to resign within the given timeframe.
- McCarthy๋Š” ์ตœ์ข… ๊ฒฐ์ •์„ ๋‚ด๋ฆฌ์ง€ ์•Š์•˜์œผ๋ฉฐ, ์˜ํšŒ์—์„œ์˜ ์ž์‹ ์˜ ๋ฏธ๋ž˜์— ๋Œ€ํ•ด ๋ถˆํ™•์‹ค์„ฑ์„ ํ‘œ๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ทธ๊ฐ€ ์—ฌ์ „ํžˆ ์„ ํƒ์ง€๋ฅผ ์ €์šธ์งˆํ•˜๊ณ  ์žˆ์œผ๋ฉฐ ์ฃผ์–ด์ง„ ์‹œ๊ฐ„ ๋‚ด์— ์‚ฌ์ž„ํ•˜์ง€ ์•Š์„ ์ˆ˜๋„ ์žˆ์Œ์„ ์˜๋ฏธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
- ๋งค์นด์‹œ๋Š” ์ตœ์ข… ๊ฒฐ์ •์„ ๋‚ด๋ฆฌ์ง€ ์•Š์•˜์œผ๋ฉฐ ์˜ํšŒ์—์„œ์˜ ๋ฏธ๋ž˜์— ๋Œ€ํ•ด ๋ถˆํ™•์‹คํ•œ ์ž…์žฅ์„ ํ‘œ๋ช…ํ•˜๊ณ  ์žˆ์–ด ์ฃผ์–ด์ง„ ๊ธฐํ•œ ๋‚ด์— ์‚ฌ์ž„ํ•˜์ง€ ์•Š์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
Pretraining CorpusFor mathematical reasoning, we utilize the OpenWebMath (OWM) dataset (Paster et al., 2023), which comprises approximately 14B tokens sourced from math-related web pages in the Common Crawl.
์ˆ˜ํ•™์  ์ถ”๋ก ์„ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” Common Crawl์˜ ์ˆ˜ํ•™ ๊ด€๋ จ ์›น ํŽ˜์ด์ง€์—์„œ ์ˆ˜์ง‘๋œ ์•ฝ 140์–ต ๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ๊ตฌ์„ฑ๋œ OpenWebMath (OWM) ๋ฐ์ดํ„ฐ์…‹ (Paster et al., 2023)์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค.
์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์ˆ˜ํ•™์  ์ถ”๋ก ์„ ์œ„ํ•ด, ์•ฝ 14B ํ† ํฐ์œผ๋กœ ๊ตฌ์„ฑ๋œ OpenWebMath(OWM) ๋ฐ์ดํ„ฐ ์„ธํŠธ(Paster et al., 2023)๋ฅผ ํ™œ์šฉํ•˜๋ฉฐ, ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” Common Crawl์˜ ์ˆ˜ํ•™ ๊ด€๋ จ ์›น ํŽ˜์ด์ง€์—์„œ ๊ฐ€์ ธ์˜จ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
\(\bullet\)_Interaction with environment_. To test this ability, we select ALFWorld [609] and WebShop [610] for evaluation, which simulate real-world scenarios such as household and e-commerce environments.
โ€ข _ํ™˜๊ฒฝ๊ณผ์˜ ์ƒํ˜ธ์ž‘์šฉ_. ์ด ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ๊ฐ€์ • ๋ฐ ์ „์ž์ƒ๊ฑฐ๋ž˜ ํ™˜๊ฒฝ๊ณผ ๊ฐ™์€ ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๋Š” ALFWorld [609]์™€ WebShop [610]์„ ํ‰๊ฐ€ ๋Œ€์ƒ์œผ๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.
ํ™˜๊ฒฝ๊ณผ์˜ ์ƒํ˜ธ์ž‘์šฉ_. ์ด ๊ธฐ๋Šฅ์„ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ€์ • ๋ฐ ์ „์ž ์ƒ๊ฑฐ๋ž˜ ํ™˜๊ฒฝ๊ณผ ๊ฐ™์€ ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๋Š” ALFWorld [609] ๋ฐ WebShop [610]์„ ํ‰๊ฐ€ ๋Œ€์ƒ์œผ๋กœ ์„ ์ •ํ•ฉ๋‹ˆ๋‹ค.
In-Domain Test Loss PredictionFor the in-domain test set, we use the code data following the same distribution as that used in the training data (the code data comprises \(10\%\) of the full training data).
๋„๋ฉ”์ธ ๋‚ด ํ…Œ์ŠคํŠธ ์†์‹ค ์˜ˆ์ธก: ๋„๋ฉ”์ธ ๋‚ด ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ๊ฒฝ์šฐ, ์šฐ๋ฆฌ๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์™€ ๋™์ผํ•œ ๋ถ„ํฌ๋ฅผ ๋”ฐ๋ฅด๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค (์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋Š” ์ „์ฒด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ \(10\%\)๋ฅผ ์ฐจ์ง€ํ•ฉ๋‹ˆ๋‹ค).
๋„๋ฉ”์ธ ํ…Œ์ŠคํŠธ ์†์‹ค ์˜ˆ์ธก๋„๋ฉ”์ธ ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ๊ฒฝ์šฐ, ํ•™์Šต ๋ฐ์ดํ„ฐ์— ์‚ฌ์šฉ๋œ ๊ฒƒ๊ณผ ๋™์ผํ•œ ๋ถ„ํฌ๋ฅผ ๋”ฐ๋ฅด๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ(์ฝ”๋“œ ๋ฐ์ดํ„ฐ๋Š” ์ „์ฒด ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ \(10\%\))๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
Future work should address these limitations to further enhance the models' capabilities, making them more robust, accessible, and effective for a broader range of applications in the Chinese NLP community.
๋ฏธ๋ž˜ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ํ•œ๊ณ„์ ์„ ํ•ด๊ฒฐํ•˜์—ฌ ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๊ณ , ์ค‘๊ตญ NLP ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ๋” ๋„“์€ ๋ฒ”์œ„์˜ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์— ๋Œ€ํ•ด ๋” ๊ฐ•๋ ฅํ•˜๊ณ , ์ ‘๊ทผ ๊ฐ€๋Šฅํ•˜๋ฉฐ, ํšจ๊ณผ์ ์œผ๋กœ ๋งŒ๋“ค ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ํ•ด๊ฒฐํ•˜์—ฌ ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œ์ผœ ์ค‘๊ตญ NLP ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ๋” ํญ๋„“์€ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋”์šฑ ๊ฒฌ๊ณ ํ•˜๊ณ  ์ ‘๊ทผ ๊ฐ€๋Šฅํ•˜๋ฉฐ ํšจ๊ณผ์ ์ธ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Logicality: Examines if the answer is logically sound and accurate in presenting viewpoints and arguments. A highquality answer should follow logical principles, demonstrating clear reasoning and coherence.
๋…ผ๋ฆฌ์„ฑ: ๋‹ต๋ณ€์ด ๋…ผ๋ฆฌ์ ์œผ๋กœ ํƒ€๋‹นํ•˜๊ณ  ๊ด€์ ๊ณผ ์ฃผ์žฅ์„ ์ •ํ™•ํ•˜๊ฒŒ ์ œ์‹œํ•˜๋Š”์ง€ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ์˜ ๋‹ต๋ณ€์€ ๋…ผ๋ฆฌ์  ์›์น™์„ ๋”ฐ๋ฅด๋ฉฐ, ๋ช…ํ™•ํ•œ ์ถ”๋ก ๊ณผ ์ผ๊ด€์„ฑ์„ ๋ณด์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๋…ผ๋ฆฌ์„ฑ: ๋‹ต๋ณ€์ด ๊ด€์ ๊ณผ ์ฃผ์žฅ์„ ์ œ์‹œํ•˜๋Š” ๋ฐ ์žˆ์–ด ๋…ผ๋ฆฌ์ ์œผ๋กœ ํƒ€๋‹นํ•˜๊ณ  ์ •ํ™•ํ•œ์ง€ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ ๋‹ต๋ณ€์€ ๋ช…ํ™•ํ•œ ์ถ”๋ก ๊ณผ ์ผ๊ด€์„ฑ์„ ๋ณด์—ฌ์ฃผ๋Š” ๋…ผ๋ฆฌ์  ์›์น™์„ ๋”ฐ๋ผ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
and makes a forecast; and (3) aggregation, which ensembles individual forecasts into an aggregated prediction. Each step makes use of an LM or a collection of LMs (either prompted or fine-tuned) (Figure 1).
๊ทธ๋ฆฌ๊ณ  ์˜ˆ์ธก์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค; (3) ๊ฐœ๋ณ„ ์˜ˆ์ธก์„ ์ง‘๊ณ„๋œ ์˜ˆ์ธก์œผ๋กœ ํ†ตํ•ฉํ•˜๋Š” ์ง‘๊ณ„ ๋‹จ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๋‹จ๊ณ„๋Š” LM ๋˜๋Š” ์—ฌ๋Ÿฌ LMs(ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ ๋ฏธ์„ธ ์กฐ์ •๋œ)๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค (Figure 1).
๊ทธ๋ฆฌ๊ณ  ์˜ˆ์ธก์„ ํ•˜๊ณ  (3) ์ง‘๊ณ„, ๊ฐœ๋ณ„ ์˜ˆ์ธก์„ ์ง‘๊ณ„๋œ ์˜ˆ์ธก์œผ๋กœ ์•™์ƒ๋ธ”ํ™”ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋‹จ๊ณ„๋Š” ํ”„๋กฌํ”„ํŠธ ๋˜๋Š” ๋ฏธ์„ธ ์กฐ์ •๋œ LM ๋˜๋Š” LM ๋ชจ์Œ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1).
Finally, we cluster documents across all buckets -- if documents A and B match in one bucket and B and C in another, A-B-C becomes a cluster. We randomly remove all but one of the documents in each cluster.
๋งˆ์ง€๋ง‰์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๋ชจ๋“  ๋ฒ„ํ‚ท์— ๊ฑธ์ณ ๋ฌธ์„œ๋ฅผ ํด๋Ÿฌ์Šคํ„ฐ๋งํ•ฉ๋‹ˆ๋‹ค -- ๋งŒ์•ฝ ๋ฌธ์„œ A์™€ B๊ฐ€ ํ•œ ๋ฒ„ํ‚ท์—์„œ ์ผ์น˜ํ•˜๊ณ  B์™€ C๊ฐ€ ๋‹ค๋ฅธ ๋ฒ„ํ‚ท์—์„œ ์ผ์น˜ํ•˜๋ฉด, A-B-C๋Š” ํ•˜๋‚˜์˜ ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ํด๋Ÿฌ์Šคํ„ฐ์—์„œ ํ•˜๋‚˜๋ฅผ ์ œ์™ธํ•œ ๋ชจ๋“  ๋ฌธ์„œ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
๋งˆ์ง€๋ง‰์œผ๋กœ, ๋ชจ๋“  ๋ฒ„ํ‚ท์— ๊ฑธ์ณ ๋ฌธ์„œ๋ฅผ ํด๋Ÿฌ์Šคํ„ฐ๋งํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ์„œ A์™€ B๊ฐ€ ํ•œ ๋ฒ„ํ‚ท์—์„œ ์ผ์น˜ํ•˜๊ณ  B์™€ C๊ฐ€ ๋‹ค๋ฅธ ๋ฒ„ํ‚ท์—์„œ ์ผ์น˜ํ•˜๋Š” ๊ฒฝ์šฐ, A-B-C๊ฐ€ ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ๊ฐ ํด๋Ÿฌ์Šคํ„ฐ์—์„œ ๋ฌธ์„œ ์ค‘ ํ•˜๋‚˜๋ฅผ ์ œ์™ธํ•œ ๋ชจ๋“  ๋ฌธ์„œ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
In these experiments, we train models on \(\mathcal{D}(B)\) and \(\mathcal{C}(B)\), for different values of \(B\). Our goal is to determine whether extensive training on "dirty" data impacts model accuracy.
์ด ์‹คํ—˜์—์„œ ์šฐ๋ฆฌ๋Š” \(B\)์˜ ๋‹ค์–‘ํ•œ ๊ฐ’์— ๋Œ€ํ•ด \(\mathcal{D}(B)\)์™€ \(\mathcal{C}(B)\)์—์„œ ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ชฉํ‘œ๋Š” "๋”๋Ÿฌ์šด" ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ํ›ˆ๋ จ์ด ๋ชจ๋ธ์˜ ์ •ํ™•๋„์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ํ™•์ธํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
์ด ์‹คํ—˜์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๊ฐ’์˜ \(B\)์— ๋Œ€ํ•ด \(\mathcal{D}(B)\) ๋ฐ \(\mathcal{C}(B)\)์— ๋Œ€ํ•œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ชฉํ‘œ๋Š” "๋”๋Ÿฌ์šด" ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ํ›ˆ๋ จ์ด ๋ชจ๋ธ ์ •ํ™•๋„์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
* Learning Rate: \(5\times 10^{-7}\) * \(\beta\): 0.1 * Epochs: 3 * Warmup: Linear warmup for the first 10% of total training time, and then linear cooldown to a learning rate of 0 over the remaining steps.
* ํ•™์Šต๋ฅ : \(5\times 10^{-7}\) * \(\beta\): 0.1 * ์—ํฌํฌ: 3 * ์›Œ๋ฐ์—…: ์ „์ฒด ํ›ˆ๋ จ ์‹œ๊ฐ„์˜ ์ฒ˜์Œ 10% ๋™์•ˆ ์„ ํ˜• ์›Œ๋ฐ์—…์„ ์ˆ˜ํ–‰ํ•œ ํ›„, ๋‚จ์€ ๋‹จ๊ณ„ ๋™์•ˆ ํ•™์Šต๋ฅ ์„ 0์œผ๋กœ ์„ ํ˜• ๊ฐ์†Œ์‹œํ‚ต๋‹ˆ๋‹ค.
* ํ•™์Šต ์†๋„: \(5\๋ฐฐ 10^{-7}\)* \(\๋ฒ ํƒ€\): 0.1* ์—ํฌํฌ: 3* ์›Œ๋ฐ์—…: ์ „์ฒด ํ›ˆ๋ จ ์‹œ๊ฐ„์˜ ์ฒ˜์Œ 10% ๋™์•ˆ ์„ ํ˜• ์›Œ๋ฐ์—…์„ ์ˆ˜ํ–‰ํ•œ ๋‹ค์Œ, ๋‚˜๋จธ์ง€ ๋‹จ๊ณ„์—์„œ 0์˜ ํ•™์Šต ์†๋„๋กœ ์„ ํ˜• ์ฟจ๋‹ค์šด์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
Next, the system retrieves articles from news APIs using the LM-generated search queries. We evaluate 5 APIs on the relevance of the articles retrieved and select NewsCatcher1 and Google News (Section E.2).
๋‹ค์Œ์œผ๋กœ, ์‹œ์Šคํ…œ์€ LM์ด ์ƒ์„ฑํ•œ ๊ฒ€์ƒ‰ ์ฟผ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‰ด์Šค API์—์„œ ๊ธฐ์‚ฌ๋ฅผ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๊ฒ€์ƒ‰๋œ ๊ธฐ์‚ฌ๋“ค์˜ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด 5๊ฐœ์˜ API๋ฅผ ํ‰๊ฐ€ํ•˜๊ณ , NewsCatcher1๊ณผ Google News๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค (์„น์…˜ E.2).
๋‹ค์Œ์œผ๋กœ, ์‹œ์Šคํ…œ์€ LM์ด ์ƒ์„ฑํ•œ ๊ฒ€์ƒ‰ ์ฟผ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‰ด์Šค API์—์„œ ๊ธฐ์‚ฌ๋ฅผ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰๋œ ๊ธฐ์‚ฌ์˜ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•˜์—ฌ 5๊ฐœ์˜ API๋ฅผ ํ‰๊ฐ€ํ•˜๊ณ  NewsCatcher1๊ณผ Google ๋‰ด์Šค๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค(์„น์…˜ E.2).
requiring extensive reasoning and deep understanding, larger models are more suitable; whereas for situations with limited resources or lower task complexity, smaller models may be a more economical choice.
๊ด‘๋ฒ”์œ„ํ•œ ์ถ”๋ก ๊ณผ ๊นŠ์€ ์ดํ•ด๊ฐ€ ํ•„์š”ํ•œ ๊ฒฝ์šฐ์—๋Š” ๋” ํฐ ๋ชจ๋ธ์ด ๋” ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์ž์›์ด ์ œํ•œ๋˜๊ฑฐ๋‚˜ ์ž‘์—…์˜ ๋ณต์žก์„ฑ์ด ๋‚ฎ์€ ์ƒํ™ฉ์—์„œ๋Š” ๋” ์ž‘์€ ๋ชจ๋ธ์ด ๊ฒฝ์ œ์ ์ธ ์„ ํƒ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ด‘๋ฒ”์œ„ํ•œ ์ถ”๋ก ๊ณผ ๊นŠ์€ ์ดํ•ด๊ฐ€ ํ•„์š”ํ•œ ๊ฒฝ์šฐ ๋” ํฐ ๋ชจ๋ธ์ด ๋” ์ ํ•ฉํ•œ ๋ฐ˜๋ฉด, ๋ฆฌ์†Œ์Šค๊ฐ€ ์ œํ•œ์ ์ด๊ฑฐ๋‚˜ ์ž‘์—… ๋ณต์žก๋„๊ฐ€ ๋‚ฎ์€ ์ƒํ™ฉ์—์„œ๋Š” ๋” ์ž‘์€ ๋ชจ๋ธ์ด ๋” ๊ฒฝ์ œ์ ์ธ ์„ ํƒ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Suppose queries \(b^{1},\cdots,b^{m}\) were asked and we received \(m\) answers \((i_{1},a_{i_{1}}(z,b^{1}))\leftarrow\mathcal{O}(p,b^{1}),\ldots,(i_{m},a_{i_{m}}(z,b^{m}))\leftarrow\mathcal{O}(p,b^{m})\).
๋‹ค์Œ๊ณผ ๊ฐ™์ด ์งˆ์˜ \(b^{1},\cdots,b^{m}\)๊ฐ€ ์ฃผ์–ด์กŒ๊ณ , \(m\)๊ฐœ์˜ ๋‹ต๋ณ€ \((i_{1},a_{i_{1}}(z,b^{1}))\leftarrow\mathcal{O}(p,b^{1}),\ldots,(i_{m},a_{i_{m}}(z,b^{m}))\leftarrow\mathcal{O}(p,b^{m})\)์„ ๋ฐ›์•˜๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ด…์‹œ๋‹ค.
์ฟผ๋ฆฌ \(b^{1},\cdots,b^{m}\)๊ฐ€ ์š”์ฒญ๋˜์—ˆ๊ณ  \(m\)๊ฐœ์˜ ๋‹ต๋ณ€ \((i_{1},a_{i_{1}}(z,b^{1}))\leftarrow\mathcal{O}(p,b^{1}),\ldots,(i_{m},a_{i_{m}}(z,b^{m}))\leftarrow\mathcal{O}(p,b^{m})\)๋ฅผ ๋ฐ›์•˜๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.
Footnote 2: The notation \(f(T)\lesssim g(T)\) means that \(f(T)\leq Cg(T)\) for sufficiently large \(T\) and an absolute constant \(C\), while \(f(T)\asymp g(T)\) means \(f(T)\lesssim g(T)\lesssim f(T)\).
Footnote 2: ํ‘œ๊ธฐ๋ฒ• \(f(T)\lesssim g(T)\)๋Š” ์ถฉ๋ถ„ํžˆ ํฐ \(T\)์™€ ์ ˆ๋Œ€ ์ƒ์ˆ˜ \(C\)์— ๋Œ€ํ•ด \(f(T)\leq Cg(T)\)๋ฅผ ์˜๋ฏธํ•˜๋ฉฐ, \(f(T)\asymp g(T)\)๋Š” \(f(T)\lesssim g(T)\lesssim f(T)\)๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
๊ฐ์ฃผ 2: ํ‘œ๊ธฐ๋ฒ• \(f(T)\lesssim g(T)\)๋Š” ์ถฉ๋ถ„ํžˆ ํฐ \(T\)์™€ ์ ˆ๋Œ€ ์ƒ์ˆ˜ \(C\)์— ๋Œ€ํ•ด \(f(T)\leq Cg(T)\)๋ฅผ ์˜๋ฏธํ•˜๋ฉฐ, \(f(T)\asymp g(T)\)๋Š” \(f(T)\lesssim g(T)\lesssim f(T)\)๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
If we write the relationship between how many duck eggs Janet has left, how many she eats, and how many she bakes for the muffins, then we can see how many eggs she has left to sell at the farmers' market.
Janet์ด ๋‚จ์€ ์˜ค๋ฆฌ ์•Œ์˜ ์ˆ˜, ๊ทธ๋…€๊ฐ€ ๋จน๋Š” ์•Œ์˜ ์ˆ˜, ๊ทธ๋ฆฌ๊ณ  ๋จธํ•€์„ ์œ„ํ•ด ๊ตฝ๋Š” ์•Œ์˜ ์ˆ˜ ์‚ฌ์ด์˜ ๊ด€๊ณ„๋ฅผ ์ž‘์„ฑํ•˜๋ฉด, ๊ทธ๋…€๊ฐ€ ๋†์‚ฐ๋ฌผ ์‹œ์žฅ์—์„œ ํŒ๋งคํ•  ์ˆ˜ ์žˆ๋Š” ์•Œ์˜ ์ˆ˜๋ฅผ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ž๋„ท์ด ๋‚จ์€ ์˜ค๋ฆฌ์•Œ ์ˆ˜, ๋จน์€ ์˜ค๋ฆฌ์•Œ ์ˆ˜, ๋จธํ•€์„ ๊ตฝ๋Š” ์ˆ˜๋ฅผ ๊ด€๊ณ„๋กœ ์ž‘์„ฑํ•˜๋ฉด, ์ž๋„ท์ด ํŒŒ๋จธ์Šค ๋งˆ์ผ“์—์„œ ํŒ๋งคํ•  ์ˆ˜ ์žˆ๋Š” ์•Œ์ด ๋ช‡ ๊ฐœ์ธ์ง€ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The 100-exposure setting reveals more intriguing comparisons. We contrast GPT2 with various model architectures in Figure 4 and offer a detailed comparison between LLaMA and GPT2 architectures in Figure 5.
100๋ฒˆ ๋…ธ์ถœ ์„ค์ •์€ ๋” ํฅ๋ฏธ๋กœ์šด ๋น„๊ต๋ฅผ ๋“œ๋Ÿฌ๋ƒ…๋‹ˆ๋‹ค. Figure 4์—์„œ๋Š” GPT2์™€ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๋น„๊ตํ•˜๊ณ , Figure 5์—์„œ๋Š” LLaMA์™€ GPT2 ์•„ํ‚คํ…์ฒ˜ ๊ฐ„์˜ ์ƒ์„ธํ•œ ๋น„๊ต๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
100ํšŒ ๋…ธ์ถœ ์„ค์ •์—์„œ๋Š” ๋” ํฅ๋ฏธ๋กœ์šด ๋น„๊ต๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 4์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์™€ GPT2๋ฅผ ๋Œ€์กฐํ•˜๊ณ , ๊ทธ๋ฆผ 5์—์„œ๋Š” LLaMA์™€ GPT2 ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ž์„ธํžˆ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค.
Let us connect Result 12 to practice. First, adding a special token to high-credibility data is very practical: imagine adding the domain name "wikipedia.org" at the beginning of all Wikipedia paragraphs.
Result 12๋ฅผ ์‹ค์ œ๋กœ ์ ์šฉํ•ด ๋ด…์‹œ๋‹ค. ๋จผ์ €, ๋†’์€ ์‹ ๋ขฐ์„ฑ์„ ๊ฐ€์ง„ ๋ฐ์ดํ„ฐ์— ํŠน์ˆ˜ ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์‹ค์šฉ์ ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ชจ๋“  ์œ„ํ‚คํ”ผ๋””์•„ ๋ฌธ๋‹จ์˜ ์‹œ์ž‘ ๋ถ€๋ถ„์— "wikipedia.org" ๋„๋ฉ”์ธ ์ด๋ฆ„์„ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ์ƒ์ƒํ•ด ๋ณด์„ธ์š”.
๊ฒฐ๊ณผ 12๋ฅผ ์‹ค์ œ์™€ ์—ฐ๊ฒฐํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๋จผ์ €, ์‹ ๋ขฐ๋„๊ฐ€ ๋†’์€ ๋ฐ์ดํ„ฐ์— ํŠน์ˆ˜ ํ† ํฐ์„ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์‹ค์šฉ์ ์ž…๋‹ˆ๋‹ค. ๋ชจ๋“  ์œ„ํ‚ค๋ฐฑ๊ณผ ๋‹จ๋ฝ์˜ ์‹œ์ž‘ ๋ถ€๋ถ„์— ๋„๋ฉ”์ธ ์ด๋ฆ„ "wikipedia.org"๋ฅผ ์ถ”๊ฐ€ํ•œ๋‹ค๊ณ  ์ƒ์ƒํ•ด ๋ณด์„ธ์š”.
In this section, we validate the workflow presented in Figure 2 by systematically deactivating language-specific neurons across different layers and components to examine their impact on a range of tasks.
์ด ์„น์…˜์—์„œ๋Š” ๋‹ค์–‘ํ•œ ์ธต๊ณผ ๊ตฌ์„ฑ ์š”์†Œ์—์„œ ์–ธ์–ด ํŠนํ™” ๋‰ด๋Ÿฐ์„ ์ฒด๊ณ„์ ์œผ๋กœ ๋น„ํ™œ์„ฑํ™”ํ•˜์—ฌ ์—ฌ๋Ÿฌ ์ž‘์—…์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•จ์œผ๋กœ์จ Figure 2์— ์ œ์‹œ๋œ ์›Œํฌํ”Œ๋กœ๋ฅผ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋ ˆ์ด์–ด์™€ ๊ตฌ์„ฑ ์š”์†Œ์—์„œ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์„ ์ฒด๊ณ„์ ์œผ๋กœ ๋น„ํ™œ์„ฑํ™”ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•˜์—ฌ ๊ทธ๋ฆผ 2์— ์ œ์‹œ๋œ ์›Œํฌํ”Œ๋กœ์šฐ๋ฅผ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค.
In this work, we take a step towards gaining a comprehensive understanding of the language capability transfer in LLMs. As shown in figure 1, we empirically investigate several key aspects based on LLaMA:
์ด ์—ฐ๊ตฌ์—์„œ ์šฐ๋ฆฌ๋Š” LLMs์˜ ์–ธ์–ด ๋Šฅ๋ ฅ ์ „์ด์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์ดํ•ด๋ฅผ ํ–ฅํ•ด ํ•œ ๊ฑธ์Œ ๋‚˜์•„๊ฐ‘๋‹ˆ๋‹ค. Figure 1์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ์šฐ๋ฆฌ๋Š” LLaMA๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ช‡ ๊ฐ€์ง€ ์ฃผ์š” ์ธก๋ฉด์„ ์‹ค์ฆ์ ์œผ๋กœ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
์ด ์—ฐ๊ตฌ์—์„œ๋Š” LLM์˜ ์–ธ์–ด ๋Šฅ๋ ฅ ์ „๋‹ฌ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์ดํ•ด๋ฅผ ํ–ฅํ•œ ํ•œ ๊ฑธ์Œ์„ ๋‚ด๋”›์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด, ์šฐ๋ฆฌ๋Š” LLaMA๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ํ•ต์‹ฌ ์ธก๋ฉด์„ ๊ฒฝํ—˜์ ์œผ๋กœ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค:
**Does the dataset contain any data, if viewed directly, might be offensive, insulting, threatening or induce anxiety?** A large number of the prompts contain inappropriate, offensive and unsafe language.
๋ฐ์ดํ„ฐ์…‹์— ์ง์ ‘์ ์œผ๋กœ ๋ณด์•˜์„ ๋•Œ ๋ถˆ์พŒํ•˜๊ฑฐ๋‚˜, ๋ชจ์š•์ ์ด๊ฑฐ๋‚˜, ์œ„ํ˜‘์ ์ด๊ฑฐ๋‚˜ ๋ถˆ์•ˆ๊ฐ์„ ์œ ๋ฐœํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๊นŒ? ๋งŽ์€ ์ˆ˜์˜ ํ”„๋กฌํ”„ํŠธ๊ฐ€ ๋ถ€์ ์ ˆํ•˜๊ณ , ๋ถˆ์พŒํ•˜๋ฉฐ, ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ์–ธ์–ด๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
**๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์— ์ง์ ‘ ๋ณด์•˜์„ ๋•Œ ๋ถˆ์พŒ๊ฐ์„ ์ฃผ๊ฑฐ๋‚˜ ๋ชจ์š•์ ์ด๊ฑฐ๋‚˜ ์œ„ํ˜‘์ ์ด๊ฑฐ๋‚˜ ๋ถˆ์•ˆ์„ ์œ ๋ฐœํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋‚˜์š”?** ๋งŽ์€ ํ”„๋กฌํ”„ํŠธ์— ๋ถ€์ ์ ˆํ•˜๊ณ  ๋ถˆ์พŒํ•˜๋ฉฐ ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ์–ธ์–ด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
\[\pi^{*}(y|x)=\frac{1}{Z(x)}\pi_{\text{ref}}(y|x)\exp\left(\frac{1}{\beta}r(x,y)\right),\]which is a valid probability distribution as \(\pi^{*}(y|x)\geq 0\) for all \(y\) and \(\sum_{y}\pi^{*}(y|x)=1\).
\[\pi^{*}(y|x)=\frac{1}{Z(x)}\pi_{\text{ref}}(y|x)\exp\left(\frac{1}{\beta}r(x,y)\right),\] ์ด๋Š” ๋ชจ๋“  \(y\)์— ๋Œ€ํ•ด \(\pi^{*}(y|x)\geq 0\)์ด๊ณ  \(\sum_{y}\pi^{*}(y|x)=1\)์ด๋ฏ€๋กœ ์œ ํšจํ•œ ํ™•๋ฅ  ๋ถ„ํฌ์ž…๋‹ˆ๋‹ค.
\pi^{*}(y|x)=\frac{1}{Z(x)}\pi_{\text{ref}}(y|x)\exp\left(\frac{1}{\beta}r(x,y)\right),\]๋Š” ๋ชจ๋“  \(y\)์— ๋Œ€ํ•ด \(\pi^{*}(y|x)\geq 0\)์ด๊ณ  \(\sum_{y}\pi^{*}(y|x)=1\)์ด๋ฏ€๋กœ ์œ ํšจํ•œ ํ™•๋ฅ  ๋ถ„ํฌ์ž…๋‹ˆ๋‹ค.
According to the provided categorizations, we can categorize and label each notable LLM as shown in table IV. As can be seen from this table, models categorized as very large are also unavailable as well.
์ œ๊ณต๋œ ๋ถ„๋ฅ˜์— ๋”ฐ๋ผ, ์šฐ๋ฆฌ๋Š” ๊ฐ ์ฃผ๋ชฉํ•  ๋งŒํ•œ LLM์„ ํ‘œ IV์— ํ‘œ์‹œ๋œ ๋Œ€๋กœ ๋ถ„๋ฅ˜ํ•˜๊ณ  ๋ผ๋ฒจ๋งํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ํ‘œ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ๋งค์šฐ ํฐ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋œ ๋ชจ๋ธ๋“ค์€ ๋˜ํ•œ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.
์ œ๊ณต๋œ ๋ถ„๋ฅ˜์— ๋”ฐ๋ผ ํ‘œ IV์™€ ๊ฐ™์ด ๊ฐ ์ฃผ๋ชฉํ•  ๋งŒํ•œ LLM์„ ๋ถ„๋ฅ˜ํ•˜๊ณ  ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ํ‘œ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ๋งค์šฐ ํฐ ๊ฒƒ์œผ๋กœ ๋ถ„๋ฅ˜๋œ ๋ชจ๋ธ๋„ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.
We trained each of our 26 2.6B parameter models on NVIDIA A100 GPUs, and the training of each model took up to 2304 GPU hours. Therefore, the total training costs amounted to \(\approx 59.000\) GPU hours.
์ €ํฌ๋Š” ๊ฐ 26๊ฐœ์˜ 2.6B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์„ NVIDIA A100 GPU์—์„œ ํ›ˆ๋ จ์‹œ์ผฐ์œผ๋ฉฐ, ๊ฐ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ์—๋Š” ์ตœ๋Œ€ 2304 GPU ์‹œ๊ฐ„์ด ์†Œ์š”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ํ›ˆ๋ จ ๋น„์šฉ์€ ์•ฝ 59,000 GPU ์‹œ๊ฐ„์ด์—ˆ์Šต๋‹ˆ๋‹ค.
26๊ฐœ์˜ 2.6B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ ๊ฐ๊ฐ์„ NVIDIA A100 GPU์—์„œ ํ•™์Šตํ–ˆ์œผ๋ฉฐ, ๊ฐ ๋ชจ๋ธ์˜ ํ•™์Šต์—๋Š” ์ตœ๋Œ€ 2304์‹œ๊ฐ„์˜ GPU ์‹œ๊ฐ„์ด ์†Œ์š”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ํ•™์Šต ๋น„์šฉ์€ ์•ฝ 59.000์‹œ๊ฐ„์˜ GPU ์‹œ๊ฐ„์ด ์†Œ์š”๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
Figure 3: Data distribution of the 43 programming languages used for pre-training. The number within each bar indicates the percent of the overall code distribution that an individual language comprises.
Figure 3: ์‚ฌ์ „ ํ›ˆ๋ จ์— ์‚ฌ์šฉ๋œ 43๊ฐœ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์˜ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ. ๊ฐ ๋ง‰๋Œ€ ์•ˆ์˜ ์ˆซ์ž๋Š” ๊ฐœ๋ณ„ ์–ธ์–ด๊ฐ€ ์ „์ฒด ์ฝ”๋“œ ๋ถ„ํฌ์—์„œ ์ฐจ์ง€ํ•˜๋Š” ๋น„์œจ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 3: ์‚ฌ์ „ ํ•™์Šต์— ์‚ฌ์šฉ๋œ 43๊ฐœ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์˜ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ. ๊ฐ ๋ง‰๋Œ€ ์•ˆ์˜ ์ˆซ์ž๋Š” ๊ฐœ๋ณ„ ์–ธ์–ด๊ฐ€ ์ „์ฒด ์ฝ”๋“œ ๋ถ„ํฌ์—์„œ ์ฐจ์ง€ํ•˜๋Š” ๋น„์œจ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
where, \(f(\theta;\cdot)\) is a LLM with parameters \(\theta\), \(y\) is the task output, \(x\) is an input in target task data \(\mathcal{T}\), and \(\mathcal{L}_{t}\) is the target task loss or metric.
์—ฌ๊ธฐ์„œ \(f(\theta;\cdot)\)๋Š” ๋งค๊ฐœ๋ณ€์ˆ˜ \(\theta\)๋ฅผ ๊ฐ€์ง„ LLM์ด๋ฉฐ, \(y\)๋Š” ์ž‘์—…์˜ ์ถœ๋ ฅ, \(x\)๋Š” ๋Œ€์ƒ ์ž‘์—… ๋ฐ์ดํ„ฐ \(\mathcal{T}\)์˜ ์ž…๋ ฅ, \(\mathcal{L}_{t}\)๋Š” ๋Œ€์ƒ ์ž‘์—…์˜ ์†์‹ค ๋˜๋Š” ์ง€ํ‘œ์ž…๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ, \(f(\theta;\cdot)\)๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ \(\theta\)๋ฅผ ๊ฐ–๋Š” LLM์ด๊ณ , \(y\)๋Š” ์ž‘์—… ์ถœ๋ ฅ์ด๋ฉฐ, \(x\)๋Š” ๋Œ€์ƒ ์ž‘์—… ๋ฐ์ดํ„ฐ \(\mathcal{T}\)์˜ ์ž…๋ ฅ์ด๊ณ , \(\mathcal{L}_{t}\)๋Š” ๋Œ€์ƒ ์ž‘์—… ์†์‹ค ๋˜๋Š” ๋ฉ”ํŠธ๋ฆญ์ž…๋‹ˆ๋‹ค.
Footnote 3: For example, we found that using ## before the Question can lead to a noticeable improvement to **phi-3-mini**โ€™s results across many benchmarks, but we did not do such changes in the prompts.
Footnote 3: ์˜ˆ๋ฅผ ๋“ค์–ด, ์งˆ๋ฌธ ์•ž์— ##๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์—ฌ๋Ÿฌ ๋ฒค์น˜๋งˆํฌ์—์„œ **phi-3-mini**์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ˆˆ์— ๋„๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์ง€๋งŒ, ์šฐ๋ฆฌ๋Š” ํ”„๋กฌํ”„ํŠธ์—์„œ ๊ทธ๋Ÿฌํ•œ ๋ณ€๊ฒฝ์„ ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
๊ฐ์ฃผ 3: ์˜ˆ๋ฅผ ๋“ค์–ด, ์งˆ๋ฌธ ์•ž์— ##๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋งŽ์€ ๋ฒค์น˜๋งˆํฌ์—์„œ **phi-3-mini**์˜ ๊ฒฐ๊ณผ๊ฐ€ ๋ˆˆ์— ๋„๊ฒŒ ๊ฐœ์„ ๋˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์ง€๋งŒ ํ”„๋กฌํ”„ํŠธ์—์„œ๋Š” ๊ทธ๋Ÿฌํ•œ ๋ณ€๊ฒฝ์„ ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
Proof.: Let us consider two reward functions from the same class, such that \(r^{\prime}(x,y)=r(x,y)+f(x)\) and, let us denote as \(\pi_{r}\) and \(\pi_{r^{\prime}}\) the corresponding optimal policies.
์ฆ๋ช…: ๋™์ผํ•œ ํด๋ž˜์Šค์—์„œ ๋‘ ๊ฐœ์˜ ๋ณด์ƒ ํ•จ์ˆ˜ \(r^{\prime}(x,y)=r(x,y)+f(x)\)๋ฅผ ๊ณ ๋ คํ•ด ๋ณด์ž. ๊ทธ๋ฆฌ๊ณ  ํ•ด๋‹นํ•˜๋Š” ์ตœ์  ์ •์ฑ…์„ \(\pi_{r}\)์™€ \(\pi_{r^{\prime}}\)๋กœ ๋‚˜ํƒ€๋‚ด์ž.
์ฆ๋ช….: ๋™์ผํ•œ ํด๋ž˜์Šค์˜ ๋‘ ๋ณด์ƒ ํ•จ์ˆ˜๋ฅผ ๊ณ ๋ คํ•˜์—ฌ \(r^{\ํ”„๋ผ์ž„}(x,y)=r(x,y)+f(x)\)๊ฐ€ ๋˜๋„๋ก ํ•˜๊ณ , ํ•ด๋‹น ์ตœ์  ์ •์ฑ…์„ \(\pi_{r}\) ๋ฐ \(\pi_{r^{\ํ”„๋ผ์ž„}}\)๋กœ ํ‘œ์‹œํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.
The results in Section 4.1 indicate that, given a fixed amount of compute for training, selecting data from larger and larger source datasets is a promising method to improve language model performance.
์„น์…˜ 4.1์˜ ๊ฒฐ๊ณผ๋Š” ๊ณ ์ •๋œ ์–‘์˜ ์ปดํ“จํŒ… ์ž์›์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•  ๋•Œ, ๋” ํฌ๊ณ  ๋” ํฐ ์†Œ์Šค ๋ฐ์ดํ„ฐ์…‹์—์„œ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ์–ธ์–ด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์œ ๋งํ•œ ๋ฐฉ๋ฒ•์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์„น์…˜ 4.1์˜ ๊ฒฐ๊ณผ๋Š” ๊ณ ์ •๋œ ์–‘์˜ ์ปดํ“จํŒ… ํŠธ๋ ˆ์ด๋‹์ด ์ฃผ์–ด์กŒ์„ ๋•Œ, ์ ์  ๋” ํฐ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ์–ธ์–ด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์œ ๋งํ•œ ๋ฐฉ๋ฒ•์ž„์„ ๋‚˜ํƒ€๋‚ธ๋‹ค.
We justify our choice of the news API's. To begin with, we experiment with 5 eligible APIs to news corpus that accept retrieval date ranges, which, for our purpose, must be specified to prevent leakage.
์šฐ๋ฆฌ๋Š” ๋‰ด์Šค API ์„ ํƒ์˜ ์ •๋‹น์„ฑ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ์šฐ์„ , ์šฐ๋ฆฌ๋Š” ๊ฒ€์ƒ‰ ๋‚ ์งœ ๋ฒ”์œ„๋ฅผ ์ˆ˜๋ฝํ•˜๋Š” ๋‰ด์Šค ์ฝ”ํผ์Šค์— ์ ํ•ฉํ•œ 5๊ฐœ์˜ API๋ฅผ ์‹คํ—˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์šฐ๋ฆฌ์˜ ๋ชฉ์ ์„ ์œ„ํ•ด ๋ˆ„์ถœ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ˜๋“œ์‹œ ์ง€์ •๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๋‰ด์Šค API์˜ ์„ ํƒ์„ ์ •๋‹นํ™”ํ•ฉ๋‹ˆ๋‹ค. ์šฐ์„ , ๊ฒ€์ƒ‰ ๋‚ ์งœ ๋ฒ”์œ„๋ฅผ ํ—ˆ์šฉํ•˜๋Š” ๋‰ด์Šค ๋ง๋ญ‰์น˜์— ๋Œ€ํ•ด 5๊ฐœ์˜ ์ ๊ฒฉ API๋ฅผ ์‹คํ—˜ํ•ด ๋ณด์•˜๋Š”๋ฐ, ์ด๋Š” ๋ชฉ์ ์ƒ ์œ ์ถœ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ์ง€์ •ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Figure 3: (a) The skills distribution of _Evol-Instruct_ testset, and (b) The difficulty and complexity level distribution between the testset of Vicuna, Alpaca (Self-Instruct), and our _Evol-Instruct_.
Figure 3: (a) _Evol-Instruct_ ํ…Œ์ŠคํŠธ์…‹์˜ ๊ธฐ์ˆ  ๋ถ„ํฌ, (b) Vicuna, Alpaca (Self-Instruct), ๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ์˜ _Evol-Instruct_ ํ…Œ์ŠคํŠธ์…‹ ๊ฐ„์˜ ๋‚œ์ด๋„ ๋ฐ ๋ณต์žก๋„ ์ˆ˜์ค€ ๋ถ„ํฌ.
๊ทธ๋ฆผ 3: (a) _Evol-Instruct_ ํ…Œ์ŠคํŠธ ์„ธํŠธ์˜ ์Šคํ‚ฌ ๋ถ„ํฌ, (b) Vicuna, Alpaca(Self-Instruct)์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ์™€ ์šฐ๋ฆฌ์˜ _Evol-Instruct_์˜ ๋‚œ์ด๋„ ๋ฐ ๋ณต์žก๋„ ์ˆ˜์ค€ ๋ถ„ํฌ.
Figure 5: **Comparison of instruct models on LegalBench-Instruct**. SaulLM-7B-Instruct establishes the state-of-the-art, outperforming the best Mistral-Instruct model by a significant 6 absolute points.
Figure 5: **LegalBench-Instruct์—์„œ instruct ๋ชจ๋ธ๋“ค์˜ ๋น„๊ต**. SaulLM-7B-Instruct๋Š” ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ตœ๊ณ ์˜ Mistral-Instruct ๋ชจ๋ธ์„ 6์  ์ฐจ์ด๋กœ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 5: **LegalBench-Instruct์—์„œ์˜ ๋ช…๋ น ๋ชจ๋ธ ๋น„๊ต**. SaulLM-7B-Instruct๋Š” ์ตœ๊ณ ์˜ Mistral-Instruct ๋ชจ๋ธ๋ณด๋‹ค 6์  ์ด์ƒ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์—ฌ ์ตœ์‹  ๊ธฐ์ˆ ์„ ํ™•๋ฆฝํ•ฉ๋‹ˆ๋‹ค.
GPT2 with Smaller MLP.Mistral has a larger MLP layer, and it is often believed that the MLP layer serves primarily for storing knowledge, in contrast to the Attention layer. But is this truly the case?
GPT2๋Š” ๋” ์ž‘์€ MLP๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Mistral์€ ๋” ํฐ MLP ๋ ˆ์ด์–ด๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ MLP ๋ ˆ์ด์–ด๋Š” Attention ๋ ˆ์ด์–ด์™€๋Š” ๋‹ฌ๋ฆฌ ์ฃผ๋กœ ์ง€์‹์„ ์ €์žฅํ•˜๋Š” ์—ญํ• ์„ ํ•œ๋‹ค๊ณ  ์—ฌ๊ฒจ์ง‘๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ด๊ฒƒ์ด ์ •๋ง ์‚ฌ์‹ค์ผ๊นŒ์š”?
๋” ์ž‘์€ MLP๋ฅผ ์‚ฌ์šฉํ•˜๋Š” GPT2.Mistral์€ MLP ๋ ˆ์ด์–ด๊ฐ€ ๋” ํฌ๋ฉฐ, MLP ๋ ˆ์ด์–ด๋Š” ์ฃผ๋กœ ์ฃผ์˜๋ ฅ ๋ ˆ์ด์–ด์™€๋Š” ๋‹ฌ๋ฆฌ ์ง€์‹์„ ์ €์žฅํ•˜๋Š” ์—ญํ• ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์—ฌ๊ฒจ์ง€๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์‹ค์ œ๋กœ ๊ทธ๋Ÿฐ๊ฐ€์š”?
**Dealing with Ambiguities:** Resolving or excluding ambiguous or contradictory data that might confuse the model during training. This can help the model to provide more definite and reliable answers.
**๋ชจํ˜ธ์„ฑ ์ฒ˜๋ฆฌ:** ํ›ˆ๋ จ ์ค‘ ๋ชจ๋ธ์„ ํ˜ผ๋ž€์Šค๋Ÿฝ๊ฒŒ ํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ์ƒ๋ฐ˜๋˜๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ํ•ด๊ฒฐํ•˜๊ฑฐ๋‚˜ ์ œ์™ธํ•˜๋Š” ๊ฒƒ. ์ด๋Š” ๋ชจ๋ธ์ด ๋” ๋ช…ํ™•ํ•˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**๋ชจํ˜ธ์„ฑ ์ฒ˜๋ฆฌ:** ํ•™์Šต ์ค‘์— ๋ชจ๋ธ์„ ํ˜ผ๋™์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ๋ชจ์ˆœ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ํ•ด๊ฒฐํ•˜๊ฑฐ๋‚˜ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ๋ชจ๋ธ์ด ๋ณด๋‹ค ๋ช…ํ™•ํ•˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Despite our best efforts to filter content that does not qualify as natural language, and to deduplicate documents, our pipeline may let through documents that may be considered as errors or redundant.
์ตœ๋Œ€ํ•œ ์ž์—ฐ์–ด๋กœ ๊ฐ„์ฃผ๋˜์ง€ ์•Š๋Š” ์ฝ˜ํ…์ธ ๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ณ  ๋ฌธ์„œ๋ฅผ ์ค‘๋ณต ์ œ๊ฑฐํ•˜๋ ค๋Š” ์ตœ์„ ์˜ ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์šฐ๋ฆฌ์˜ ํŒŒ์ดํ”„๋ผ์ธ์€ ์˜ค๋ฅ˜๋กœ ๊ฐ„์ฃผ๋˜๊ฑฐ๋‚˜ ์ค‘๋ณต๋œ ๋ฌธ์„œ๋ฅผ ํ†ต๊ณผ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ž์—ฐ์–ด์— ํ•ด๋‹นํ•˜์ง€ ์•Š๋Š” ์ฝ˜ํ…์ธ ๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ณ  ๋ฌธ์„œ ์ค‘๋ณต์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด ์ตœ์„ ์„ ๋‹คํ–ˆ์ง€๋งŒ, ํŒŒ์ดํ”„๋ผ์ธ์—์„œ ์˜ค๋ฅ˜ ๋˜๋Š” ์ค‘๋ณต์œผ๋กœ ๊ฐ„์ฃผ๋  ์ˆ˜ ์žˆ๋Š” ๋ฌธ์„œ๊ฐ€ ํ†ต๊ณผ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In this section, we provide the accuracy curves of our experiments across various datasets when utilizing different LLMs. From these curves, we demonstrate that our method has the following properties:
์ด ์„น์…˜์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ์…‹์„ ํ™œ์šฉํ•˜์—ฌ ์„œ๋กœ ๋‹ค๋ฅธ LLM์„ ์‚ฌ์šฉํ•  ๋•Œ์˜ ์‹คํ—˜ ์ •ํ™•๋„ ๊ณก์„ ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณก์„ ์„ ํ†ตํ•ด ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์ด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŠน์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋‹ค์–‘ํ•œ LLM์„ ํ™œ์šฉํ•  ๋•Œ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์‹คํ—˜์˜ ์ •ํ™•๋„ ๊ณก์„ ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณก์„ ์„ ํ†ตํ•ด ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŠน์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค:
In this paper, our primary focus is on \(\mathsf{bioS}(N)\) for \(N\) ranging between 10K and 20M. Notably, \(\mathsf{bioS}(20M)\) encompasses approximately 1B bits of knowledge (refer to Theorem 3.2).
์ด ๋…ผ๋ฌธ์—์„œ ์šฐ๋ฆฌ์˜ ์ฃผ์š” ์ดˆ์ ์€ \(N\)์ด 10K์—์„œ 20M ์‚ฌ์ด์ธ \(\mathsf{bioS}(N)\)์— ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, \(\mathsf{bioS}(20M)\)์€ ์•ฝ 10์–ต ๋น„ํŠธ์˜ ์ง€์‹์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค (์ •๋ฆฌ 3.2๋ฅผ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค).
์ด ๋…ผ๋ฌธ์—์„œ๋Š” 10K์—์„œ 20M ์‚ฌ์ด์˜ \(N\)์— ๋Œ€ํ•œ \(\mathsf{bioS}(N)\)์— ์ฃผ๋กœ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, \(\mathsf{bioS}(20M)\)๋Š” ์•ฝ 1B ๋น„ํŠธ์˜ ์ง€์‹์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค(์ •๋ฆฌ 3.2 ์ฐธ์กฐ).
loss are more learnable and better aligned with the desired distribution, naturally excluding tokens that are either irrelevant or of low quality. Below, we provide a detailed description of each step.
์†์‹ค์€ ๋” ํ•™์Šต ๊ฐ€๋Šฅํ•˜๊ณ  ์›ํ•˜๋Š” ๋ถ„ํฌ์™€ ๋” ์ž˜ ๋งž์œผ๋ฉฐ, ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ๊ด€๋ จ์ด ์—†๊ฑฐ๋‚˜ ํ’ˆ์งˆ์ด ๋‚ฎ์€ ํ† ํฐ์„ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. ์•„๋ž˜์—์„œ๋Š” ๊ฐ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์†์‹ค์€ ๋” ์‰ฝ๊ฒŒ ๋ฐฐ์šธ ์ˆ˜ ์žˆ๊ณ  ์›ํ•˜๋Š” ๋ถ„ํฌ์— ๋” ์ž˜ ๋ถ€ํ•ฉํ•˜๋ฉฐ, ๊ด€๋ จ์„ฑ์ด ์—†๊ฑฐ๋‚˜ ํ’ˆ์งˆ์ด ๋‚ฎ์€ ํ† ํฐ์„ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. ์•„๋ž˜์—์„œ๋Š” ๊ฐ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
_We train models on this mixture, ensuring each piece of useful data is seen for 100 exposures, thus making the total training 8 times longer compared to 100 exposures without junk (i.e., Figure 1(b)).
์ด ํ˜ผํ•ฉ๋ฌผ๋กœ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ์œ ์šฉํ•œ ๋ฐ์ดํ„ฐ์˜ ๊ฐ ๋ถ€๋ถ„์ด 100๋ฒˆ ๋…ธ์ถœ๋˜๋„๋ก ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์“ธ๋ชจ์—†๋Š” ๋ฐ์ดํ„ฐ ์—†์ด 100๋ฒˆ ๋…ธ์ถœ๋˜๋Š” ๊ฒฝ์šฐ์— ๋น„ํ•ด ์ด ํ›ˆ๋ จ ์‹œ๊ฐ„์ด 8๋ฐฐ ๋” ๊ธธ์–ด์ง‘๋‹ˆ๋‹ค (์ฆ‰, Figure 1(b)).
_์ด ํ˜ผํ•ฉ๋ฌผ์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผœ ์œ ์šฉํ•œ ๋ฐ์ดํ„ฐ์˜ ๊ฐ ์กฐ๊ฐ์ด 100ํšŒ ๋…ธ์ถœ๋˜๋„๋ก ํ•˜์—ฌ ์ •ํฌ๊ฐ€ ์—†๋Š” 100ํšŒ ๋…ธ์ถœ(์ฆ‰, ๊ทธ๋ฆผ 1(b))์— ๋น„ํ•ด ์ด ํ•™์Šต ์‹œ๊ฐ„์ด 8๋ฐฐ ๋” ๊ธธ์–ด์ง‘๋‹ˆ๋‹ค.
Please consider the following forecasting question and its background information. After that, I will give you a news article and ask you to rate its relevance with respect to the forecasting question.
๋‹ค์Œ ์˜ˆ์ธก ์งˆ๋ฌธ๊ณผ ๊ทธ ๋ฐฐ๊ฒฝ ์ •๋ณด๋ฅผ ๊ณ ๋ คํ•ด ์ฃผ์„ธ์š”. ๊ทธ ํ›„, ๋‰ด์Šค ๊ธฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์˜ˆ์ธก ์งˆ๋ฌธ๊ณผ์˜ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•ด ๋‹ฌ๋ผ๊ณ  ์š”์ฒญํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๋‹ค์Œ ์˜ˆ์ธก ์งˆ๋ฌธ๊ณผ ๊ทธ ๋ฐฐ๊ฒฝ ์ •๋ณด๋ฅผ ๊ณ ๋ คํ•ด ์ฃผ์„ธ์š”. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋‰ด์Šค ๊ธฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์˜ˆ์ธก ์งˆ๋ฌธ๊ณผ ๊ด€๋ จํ•˜์—ฌ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•ด ๋‹ฌ๋ผ๊ณ  ์š”์ฒญํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
When RALM was first proposed, many researchers used this method because it aligned with their original ideas, particularly those of Lewis et al. (2020), Guu et al. (2020), and Izacard and Grave (2020).
RALM์ด ์ฒ˜์Œ ์ œ์•ˆ๋˜์—ˆ์„ ๋•Œ, ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด ์ด ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ–ˆ๋Š”๋ฐ, ์ด๋Š” ํŠนํžˆ Lewis et al. (2020), Guu et al. (2020), Izacard and Grave (2020)์˜ ์›๋ž˜ ์•„์ด๋””์–ด์™€ ์ผ์น˜ํ–ˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
RALM์ด ์ฒ˜์Œ ์ œ์•ˆ๋˜์—ˆ์„ ๋•Œ, ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด ์ด ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ–ˆ๋Š”๋ฐ, ์ด๋Š” Lewis ๋“ฑ(2020), Guu ๋“ฑ(2020), Izacard and Grave(2020)์˜ ์›๋ž˜ ์•„์ด๋””์–ด, ํŠนํžˆ ์ด ๋ฐฉ๋ฒ•์ด ๊ทธ๋“ค์˜ ์›๋ž˜ ์•„์ด๋””์–ด์™€ ์ผ์น˜ํ–ˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค.
LMs for OpenAI. Overall, OpenAI has explored two major approaches to further improving the GPT-3 model, _i.e._, training on code data and alignment with human preference, which are detailed as follows.
OpenAI์˜ LMs. ์ „๋ฐ˜์ ์œผ๋กœ OpenAI๋Š” GPT-3 ๋ชจ๋ธ์„ ๋”์šฑ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ํƒ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฆ‰, ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ•™์Šต๊ณผ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž์ถ˜ ์กฐ์ •์ž…๋‹ˆ๋‹ค. ์ด ๋‘ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ž์„ธํžˆ ์„ค๋ช…๋ฉ๋‹ˆ๋‹ค.
OpenAI์šฉ LM. ์ „๋ฐ˜์ ์œผ๋กœ OpenAI๋Š” ์ฝ”๋“œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ›ˆ๋ จ๊ณผ ์ธ๊ฐ„ ์„ ํ˜ธ๋„์— ๋Œ€ํ•œ ์ •๋ ฌ์ด๋ผ๋Š” GPT-3 ๋ชจ๋ธ์„ ๋”์šฑ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ชจ์ƒ‰ํ•ด ์™”์œผ๋ฉฐ, ์ด๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ž์„ธํžˆ ์„ค๋ช…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
Given the classification task definition and the class labels, generate an input that corresponds to each of the class labels. If the task doesn't require input, just generate the correct class label.
๋ถ„๋ฅ˜ ์ž‘์—… ์ •์˜์™€ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”์ด ์ฃผ์–ด์ง€๋ฉด, ๊ฐ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”์— ํ•ด๋‹นํ•˜๋Š” ์ž…๋ ฅ์„ ์ƒ์„ฑํ•˜์‹ญ์‹œ์˜ค. ์ž‘์—…์— ์ž…๋ ฅ์ด ํ•„์š”ํ•˜์ง€ ์•Š์€ ๊ฒฝ์šฐ, ์˜ฌ๋ฐ”๋ฅธ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”๋งŒ ์ƒ์„ฑํ•˜์‹ญ์‹œ์˜ค.
๋ถ„๋ฅ˜ ์ž‘์—… ์ •์˜์™€ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”์ด ์ฃผ์–ด์ง€๋ฉด ๊ฐ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”์— ํ•ด๋‹นํ•˜๋Š” ์ž…๋ ฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ž‘์—…์— ์ž…๋ ฅ์ด ํ•„์š”ํ•˜์ง€ ์•Š์€ ๊ฒฝ์šฐ ์˜ฌ๋ฐ”๋ฅธ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”๋งŒ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
Figure 1: **The performance-vs-loss curves of 1.5B, 6B, and 32B models. Each data point is the loss (\(x\)-axis) and performance (\(y\)-axis) of the intermediate checkpoint of one of the three models.
๊ทธ๋ฆผ 1: 1.5B, 6B, 32B ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ๋Œ€ ์†์‹ค ๊ณก์„ . ๊ฐ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋Š” ์„ธ ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์˜ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์˜ ์†์‹ค (\(x\)-์ถ•)๊ณผ ์„ฑ๋Šฅ (\(y\)-์ถ•)์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1: 1.5B, 6B ๋ฐ 32B ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ๋Œ€ ์†์‹ค ๊ณก์„ . ๊ฐ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋Š” ์„ธ ๊ฐ€์ง€ ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์˜ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์˜ ์†์‹ค(\(x\)-์ถ•)๊ณผ ์„ฑ๋Šฅ(\(y\)-์ถ•)์ž…๋‹ˆ๋‹ค.
Large language models exhibit promising accomplishments, but the main question that arises is how effectively they function and how their performance can be assessed in specific tasks or applications.
๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธ์€ ์œ ๋งํ•œ ์„ฑ๊ณผ๋ฅผ ๋ณด์ด์ง€๋งŒ, ์ฃผ์š”ํ•œ ์งˆ๋ฌธ์€ ์ด๋“ค์ด ํŠน์ • ์ž‘์—…์ด๋‚˜ ์‘์šฉ์—์„œ ์–ผ๋งˆ๋‚˜ ํšจ๊ณผ์ ์œผ๋กœ ์ž‘๋™ํ•˜๋ฉฐ ๊ทธ ์„ฑ๋Šฅ์„ ์–ด๋–ป๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€์ž…๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ์œ ๋งํ•œ ์„ฑ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์ง€๋งŒ, ๊ฐ€์žฅ ์ค‘์š”ํ•œ ์งˆ๋ฌธ์€ ํŠน์ • ์ž‘์—…์ด๋‚˜ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ ์–ผ๋งˆ๋‚˜ ํšจ๊ณผ์ ์œผ๋กœ ์ž‘๋™ํ•˜๊ณ  ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š”์ง€์— ๋Œ€ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Defining \(B_{crit}(L)\)Under the constraint of Equation 3.8, we can derive the critical batch size at \(L\) which minimizes the trade-off between time (\(S/S_{min}\)) and computation (\(E/E_{min}\)):
Equation 3.8์˜ ์ œ์•ฝ ์กฐ๊ฑด ํ•˜์—์„œ, ์‹œ๊ฐ„ (\(S/S_{min}\))๊ณผ ๊ณ„์‚ฐ (\(E/E_{min}\)) ๊ฐ„์˜ ๊ท ํ˜•์„ ์ตœ์†Œํ™”ํ•˜๋Š” \(L\)์—์„œ์˜ ์ž„๊ณ„ ๋ฐฐ์น˜ ํฌ๊ธฐ \(B_{crit}(L)\)๋ฅผ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
B_{crit}(L)\) ์ •์˜ํ•˜๊ธฐ3.8 ๋ฐฉ์ •์‹์˜ ์ œ์•ฝ ์กฐ๊ฑด ํ•˜์—์„œ ์‹œ๊ฐ„(\(S/S_{min}\))๊ณผ ๊ณ„์‚ฐ(\(E/E_{min}\)) ์‚ฌ์ด์˜ ์ ˆ์ถฉ์ ์„ ์ตœ์†Œํ™”ํ•˜๋Š” \(L\)์—์„œ ์ž„๊ณ„ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
In constrast, the storage manager is responsible for the long-term preservation of data, overseeing the storage of information that needs to be retained indefinitely, beyond the active lifespan of any
๋ฐ˜๋ฉด, ์Šคํ† ๋ฆฌ์ง€ ๊ด€๋ฆฌ์ž๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์žฅ๊ธฐ์ ์œผ๋กœ ๋ณด์กดํ•˜๋Š” ์—ญํ• ์„ ๋‹ด๋‹นํ•˜๋ฉฐ, ํ™œ์„ฑ ์ˆ˜๋ช…์„ ๋„˜์–ด์„œ ๋ฌด๊ธฐํ•œ์œผ๋กœ ๋ณด์กดํ•ด์•ผ ํ•˜๋Š” ์ •๋ณด๋ฅผ ๊ด€๋ฆฌํ•ฉ๋‹ˆ๋‹ค.
์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์Šคํ† ๋ฆฌ์ง€ ๊ด€๋ฆฌ์ž๋Š” ๋ฐ์ดํ„ฐ์˜ ์žฅ๊ธฐ ๋ณด์กด์„ ๋‹ด๋‹นํ•˜๋ฉฐ, ํ™œ์„ฑ ์ˆ˜๋ช…์„ ๋„˜์–ด ๋ฌด๊ธฐํ•œ ๋ณด์กดํ•ด์•ผ ํ•˜๋Š” ์ •๋ณด์˜ ์ €์žฅ์„ ๊ฐ๋…ํ•ฉ๋‹ˆ๋‹ค.
Specialized Knowledge TasksThe key message from the results in Table 2 is that synthetic data can not impart 'new knowledge'. It can only help pre-train faster, which was also the premise of our work.
Table 2์˜ ๊ฒฐ๊ณผ์—์„œ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ์ฃผ์š” ๋ฉ”์‹œ์ง€๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ '์ƒˆ๋กœ์šด ์ง€์‹'์„ ์ „๋‹ฌํ•  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๋‹จ์ง€ ๋” ๋น ๋ฅธ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ๋„์šธ ์ˆ˜ ์žˆ์„ ๋ฟ์ด๋ฉฐ, ์ด๋Š” ๋˜ํ•œ ์šฐ๋ฆฌ์˜ ์ž‘์—…์˜ ์ „์ œ์˜€์Šต๋‹ˆ๋‹ค.
์ „๋ฌธ ์ง€์‹ ์ž‘์—…ํ‘œ 2์˜ ๊ฒฐ๊ณผ์—์„œ ์ฃผ์š” ๋ฉ”์‹œ์ง€๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ '์ƒˆ๋กœ์šด ์ง€์‹'์„ ์ „๋‹ฌํ•  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋” ๋น ๋ฅด๊ฒŒ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ๋ฟ์ด๋ฉฐ, ์ด๋Š” ์šฐ๋ฆฌ ์ž‘์—…์˜ ์ „์ œ์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค.
_TL;DR: Our analysis indicates that our human-labeled data has a relatively minimal impact, probably due to a high proportion of tied translations and potential human bias in the evaluation process._
์š”์•ฝ: ์šฐ๋ฆฌ์˜ ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด, ์ธ๊ฐ„์ด ๋ผ๋ฒจ๋งํ•œ ๋ฐ์ดํ„ฐ๋Š” ๋น„๊ต์  ์ตœ์†Œํ•œ์˜ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋ฉฐ, ์ด๋Š” ์•„๋งˆ๋„ ๋™์  ๋ฒˆ์—ญ์˜ ๋†’์€ ๋น„์œจ๊ณผ ํ‰๊ฐ€ ๊ณผ์ •์—์„œ์˜ ์ž ์žฌ์ ์ธ ์ธ๊ฐ„ ํŽธํ–ฅ ๋•Œ๋ฌธ์ผ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
_TL;DR: ๋ถ„์„ ๊ฒฐ๊ณผ, ์ธ๊ฐ„์ด ๋ผ๋ฒจ์„ ๋ถ™์ธ ๋ฐ์ดํ„ฐ์˜ ์˜ํ–ฅ์ด ์ƒ๋Œ€์ ์œผ๋กœ ๋ฏธ๋ฏธํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ, ์ด๋Š” ๋ฌถ์Œ ๋ฒˆ์—ญ์˜ ๋น„์œจ์ด ๋†’๊ณ  ํ‰๊ฐ€ ๊ณผ์ •์—์„œ ์ธ๊ฐ„ ํŽธํ–ฅ์„ฑ์ด ์žˆ์—ˆ๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 5: SFT data scaling curve. Compared with UltraChat and its cleaned version UltraChat 200K, our SFT data demonstrates clear scaling advantages. We attribute its steep slope to the data quality.
Figure 5: SFT ๋ฐ์ดํ„ฐ ์Šค์ผ€์ผ๋ง ๊ณก์„ . UltraChat ๋ฐ ๊ทธ ์ •์ œ๋œ ๋ฒ„์ „์ธ UltraChat 200K์™€ ๋น„๊ตํ–ˆ์„ ๋•Œ, ์šฐ๋ฆฌ์˜ SFT ๋ฐ์ดํ„ฐ๋Š” ๋ช…ํ™•ํ•œ ์Šค์ผ€์ผ๋ง ์ด์ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๋Ÿฌํ•œ ๊ฐ€ํŒŒ๋ฅธ ๊ธฐ์šธ๊ธฐ๋ฅผ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์— ๊ธฐ์ธํ•œ๋‹ค๊ณ  ๋ด…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 5: SFT ๋ฐ์ดํ„ฐ ์Šค์ผ€์ผ๋ง ๊ณก์„ . UltraChat ๋ฐ ์ •๋ฆฌ๋œ ๋ฒ„์ „์ธ UltraChat 200K์™€ ๋น„๊ตํ–ˆ์„ ๋•Œ, ์šฐ๋ฆฌ์˜ SFT ๋ฐ์ดํ„ฐ๋Š” ๋ถ„๋ช…ํ•œ ํ™•์žฅ์„ฑ ์ด์ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ ๊ฐ€ํŒŒ๋ฅธ ๊ธฐ์šธ๊ธฐ๋Š” ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์— ๊ธฐ์ธํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
This section describes our pretraining setup, including our distributed training framework (Section 3.1), optimizer settings (Section 3.2), data preparation (Section 3.3), and hardware (Section 3.4).
์ด ์„น์…˜์—์„œ๋Š” ๋ถ„์‚ฐ ํ•™์Šต ํ”„๋ ˆ์ž„์›Œํฌ(์„น์…˜ 3.1), ์˜ตํ‹ฐ๋งˆ์ด์ € ์„ค์ •(์„น์…˜ 3.2), ๋ฐ์ดํ„ฐ ์ค€๋น„(์„น์…˜ 3.3), ํ•˜๋“œ์›จ์–ด(์„น์…˜ 3.4)๋ฅผ ํฌํ•จํ•œ ์šฐ๋ฆฌ์˜ ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋ถ„์‚ฐ ํ•™์Šต ํ”„๋ ˆ์ž„์›Œํฌ(3.1์ ˆ), ์ตœ์ ํ™” ๋„๊ตฌ ์„ค์ •(3.2์ ˆ), ๋ฐ์ดํ„ฐ ์ค€๋น„(3.3์ ˆ), ํ•˜๋“œ์›จ์–ด(3.4์ ˆ)๋ฅผ ํฌํ•จํ•œ ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
The original draft was polished by OpenAI GPT-4 for grammatical corrections and clarity improvements. We would like to thank our community members for their contributions to our open-source projects.
์›๋ณธ ์ดˆ์•ˆ์€ ๋ฌธ๋ฒ• ์ˆ˜์ • ๋ฐ ๋ช…ํ™•์„ฑ ํ–ฅ์ƒ์„ ์œ„ํ•ด OpenAI GPT-4์— ์˜ํ•ด ๋‹ค๋“ฌ์–ด์กŒ์Šต๋‹ˆ๋‹ค. ์˜คํ”ˆ ์†Œ์Šค ํ”„๋กœ์ ํŠธ์— ๊ธฐ์—ฌํ•ด ์ฃผ์‹  ์ปค๋ฎค๋‹ˆํ‹ฐ ๊ตฌ์„ฑ์›๋“ค๊ป˜ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
์›๋ณธ ์ดˆ์•ˆ์€ ๋ฌธ๋ฒ• ์ˆ˜์ • ๋ฐ ๋ช…ํ™•์„ฑ ๊ฐœ์„ ์„ ์œ„ํ•ด OpenAI GPT-4์—์„œ ๋‹ค๋“ฌ์—ˆ์Šต๋‹ˆ๋‹ค. ์˜คํ”ˆ ์†Œ์Šค ํ”„๋กœ์ ํŠธ์— ๊ธฐ์—ฌํ•ด ์ฃผ์‹  ์ปค๋ฎค๋‹ˆํ‹ฐ ํšŒ์› ์—ฌ๋Ÿฌ๋ถ„๊ป˜ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
- [Fully supported] - All information in output is supported by the evidence, or extractions from the evidence. This is only applicable when the output and part of the evidence are almost identical.
- [์™„์ „ํžˆ ์ง€์›๋จ] - ์ถœ๋ ฅ์˜ ๋ชจ๋“  ์ •๋ณด๊ฐ€ ์ฆ๊ฑฐ ๋˜๋Š” ์ฆ๊ฑฐ์—์„œ ์ถ”์ถœ๋œ ๋‚ด์šฉ์— ์˜ํ•ด ์ง€์›๋ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ถœ๋ ฅ๊ณผ ์ฆ๊ฑฐ์˜ ์ผ๋ถ€๊ฐ€ ๊ฑฐ์˜ ๋™์ผํ•  ๋•Œ๋งŒ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค.
- [์™„์ „ ์ง€์›]- ์ถœ๋ ฅ์˜ ๋ชจ๋“  ์ •๋ณด๋Š” ์ฆ๊ฑฐ ๋˜๋Š” ์ฆ๊ฑฐ์—์„œ ์ถ”์ถœํ•œ ์ •๋ณด์— ์˜ํ•ด ์ง€์›๋ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ถœ๋ ฅ๊ณผ ์ฆ๊ฑฐ์˜ ์ผ๋ถ€๊ฐ€ ๊ฑฐ์˜ ๋™์ผํ•œ ๊ฒฝ์šฐ์—๋งŒ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค.
Figure 2: Histograms depicting the proportion of questions where Adapt-LLM trained on NQ (left) and Adapt-LLM trained on SQuAD (right) ask for extra context for different popularity score intervals.
Figure 2: NQ์—์„œ ํ›ˆ๋ จ๋œ Adapt-LLM(์™ผ์ชฝ)๊ณผ SQuAD์—์„œ ํ›ˆ๋ จ๋œ Adapt-LLM(์˜ค๋ฅธ์ชฝ)์ด ๋‹ค๋ฅธ ์ธ๊ธฐ ์ ์ˆ˜ ๊ตฌ๊ฐ„์—์„œ ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๋ฅผ ์š”์ฒญํ•˜๋Š” ์งˆ๋ฌธ์˜ ๋น„์œจ์„ ๋‚˜ํƒ€๋‚ด๋Š” ํžˆ์Šคํ† ๊ทธ๋žจ.
๊ทธ๋ฆผ 2: NQ๋กœ ํ•™์Šต๋œ Adapt-LLM(์™ผ์ชฝ)๊ณผ SQuAD๋กœ ํ•™์Šต๋œ Adapt-LLM(์˜ค๋ฅธ์ชฝ)์ด ๋‹ค์–‘ํ•œ ์ธ๊ธฐ๋„ ์ ์ˆ˜ ๊ฐ„๊ฒฉ์— ๋Œ€ํ•ด ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๋ฅผ ์š”๊ตฌํ•˜๋Š” ์งˆ๋ฌธ์˜ ๋น„์œจ์„ ๋‚˜ํƒ€๋‚ด๋Š” ํžˆ์Šคํ† ๊ทธ๋žจ์ž…๋‹ˆ๋‹ค.
We can observe that in both Mistral and Vicuna, the intersection with English from other languages is relatively limited, suggesting that English possesses a predominant number of language-specific
Mistral๊ณผ Vicuna ๋ชจ๋‘์—์„œ ๋‹ค๋ฅธ ์–ธ์–ด์™€์˜ ์˜์–ด ๊ต์ง‘ํ•ฉ์ด ์ƒ๋Œ€์ ์œผ๋กœ ์ œํ•œ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์˜์–ด๊ฐ€ ์–ธ์–ด๋ณ„๋กœ ์ง€๋ฐฐ์ ์ธ ์ˆ˜๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
Mistral๊ณผ Vicuna ๋ชจ๋‘์—์„œ ๋‹ค๋ฅธ ์–ธ์–ด์˜ ์˜์–ด์™€์˜ ๊ต์ฐจ์ ์ด ์ƒ๋Œ€์ ์œผ๋กœ ์ œํ•œ๋˜์–ด ์žˆ๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์˜์–ด๊ฐ€ ์–ธ์–ด๋ณ„๋กœ ์šฐ์„ธํ•œ ์ˆ˜๋ฅผ ๋ณด์œ ํ•˜๊ณ  ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
Upon publication, we will make our code and final model checkpoints publicly available. A preliminary version of this work was made available as an ICML 2023 workshop paper in (Gupta et al., 2023).
์ถœํŒ ํ›„, ์šฐ๋ฆฌ๋Š” ์ฝ”๋“œ์™€ ์ตœ์ข… ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๊ณต๊ฐœํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ์ดˆ๊ธฐ ๋ฒ„์ „์€ ICML 2023 ์›Œํฌ์ˆ ๋…ผ๋ฌธ์œผ๋กœ (Gupta et al., 2023) ๊ณต๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
๊ฒŒ์‹œ ์‹œ ์ฝ”๋“œ์™€ ์ตœ์ข… ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์˜ ์˜ˆ๋น„ ๋ฒ„์ „์€ (Gupta et al., 2023)์˜ ICML 2023 ์›Œํฌ์ƒต ๋…ผ๋ฌธ์œผ๋กœ ๊ณต๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
The base version of the dataset is self-contained, but the multimodal version is interlaced with links to images-these are not distributed as part of the dataset, and constitute an external source.
๋ฐ์ดํ„ฐ์…‹์˜ ๊ธฐ๋ณธ ๋ฒ„์ „์€ ๋…๋ฆฝ์ ์ด์ง€๋งŒ, ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ฒ„์ „์€ ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ๋งํฌ์™€ ์–ฝํ˜€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ด๋ฏธ์ง€๋Š” ๋ฐ์ดํ„ฐ์…‹์˜ ์ผ๋ถ€๋กœ ๋ฐฐํฌ๋˜์ง€ ์•Š์œผ๋ฉฐ, ์™ธ๋ถ€ ์†Œ์Šค๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ธฐ๋ณธ ๋ฒ„์ „์€ ๋…๋ฆฝ์ ์ด์ง€๋งŒ, ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ฒ„์ „์€ ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ๋งํฌ์™€ ์ธํ„ฐ๋ ˆ์ด์Šค๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ผ๋ถ€๋กœ ๋ฐฐํฌ๋˜์ง€ ์•Š๊ณ  ์™ธ๋ถ€ ์†Œ์Šค๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
Prompting LLMs to generate book summaries:To summarize book-length documents, we adopt the hierarchical merging strategy from (Chang et al., 2023b); see Figure 1 for an illustration of the method.
LLMs์—๊ฒŒ ์ฑ… ์š”์•ฝ์„ ์ƒ์„ฑํ•˜๋„๋ก ์ง€์‹œํ•˜๊ธฐ: ์ฑ… ๊ธธ์ด์˜ ๋ฌธ์„œ๋ฅผ ์š”์•ฝํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” (Chang et al., 2023b)์—์„œ ์ œ์•ˆํ•œ ๊ณ„์ธต์  ๋ณ‘ํ•ฉ ์ „๋žต์„ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์„ค๋ช…์€ Figure 1์„ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค.
LLM์ด ์ฑ… ์š”์•ฝ์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ:์ฑ… ๊ธธ์ด์˜ ๋ฌธ์„œ๋ฅผ ์š”์•ฝํ•˜๊ธฐ ์œ„ํ•ด (Chang et al., 2023b)์˜ ๊ณ„์ธต์  ๋ณ‘ํ•ฉ ์ „๋žต์„ ์ฑ„ํƒํ•˜๋ฉฐ, ์ด ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์˜ˆ๋Š” ๊ทธ๋ฆผ 1์„ ์ฐธ์กฐํ•˜์„ธ์š”.
In addition to the above practical strategies and tricks, existing work has also used other tricks, _e.g.,_ concatenating multiple examples into a single sequence to approach the max length [355].
์œ„์˜ ์‹ค์šฉ์ ์ธ ์ „๋žต๊ณผ ์š”๋ น ์™ธ์—๋„, ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์—ฌ๋Ÿฌ ์˜ˆ์ œ๋ฅผ ํ•˜๋‚˜์˜ ์‹œํ€€์Šค๋กœ ์—ฐ๊ฒฐํ•˜์—ฌ ์ตœ๋Œ€ ๊ธธ์ด์— ์ ‘๊ทผํ•˜๋Š” ๋“ฑ์˜ ๋‹ค๋ฅธ ์š”๋ น๋„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค [355].
์œ„์˜ ์‹ค์ œ ์ „๋žต๊ณผ ํŠธ๋ฆญ ์™ธ์—๋„ ๊ธฐ์กด ์ž‘์—…์—์„œ๋Š” ์ตœ๋Œ€ ๊ธธ์ด์— ์ ‘๊ทผํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์˜ˆ์ œ๋ฅผ ๋‹จ์ผ ์‹œํ€€์Šค๋กœ ์—ฐ๊ฒฐํ•˜๋Š” ๋“ฑ ๋‹ค๋ฅธ ํŠธ๋ฆญ๋„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค[355].
2. Increased endurance: HIIT can improve endurance by increasing the body's ability to use oxygen during exercise. This can help athletes to exercise for longer periods without becoming exhausted.
2. ์ง€๊ตฌ๋ ฅ ํ–ฅ์ƒ: HIIT๋Š” ์šด๋™ ์ค‘ ์‹ ์ฒด๊ฐ€ ์‚ฐ์†Œ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์ฆ๊ฐ€์‹œ์ผœ ์ง€๊ตฌ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์šด๋™์„ ์ˆ˜๋“ค์ด ์ง€์น˜์ง€ ์•Š๊ณ  ๋” ์˜ค๋žœ ์‹œ๊ฐ„ ๋™์•ˆ ์šด๋™ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ค๋‹ˆ๋‹ค.
2. ์ง€๊ตฌ๋ ฅ ํ–ฅ์ƒ: HIIT๋Š” ์šด๋™ ์ค‘ ์‚ฐ์†Œ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์‹ ์ฒด์˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œ์ผœ ์ง€๊ตฌ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์šด๋™์„ ์ˆ˜๊ฐ€ ์ง€์น  ๋•Œ๊นŒ์ง€ ์˜ค๋žซ๋™์•ˆ ์šด๋™ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
where \(c:=1-1/\beta\in(0,1)\), and we have used the elementary fact that \(\sum_{i\geq k+1}i^{-\beta}\asymp k^{-(\beta-1)}\) for large \(k\). For the second sum, we will need the following lemma.
์—ฌ๊ธฐ์„œ \(c:=1-1/\beta\in(0,1)\)์ด๋ฉฐ, ์šฐ๋ฆฌ๋Š” \(k\)๊ฐ€ ํด ๋•Œ \(\sum_{i\geq k+1}i^{-\beta}\asymp k^{-(\beta-1)}\)๋ผ๋Š” ๊ธฐ๋ณธ์ ์ธ ์‚ฌ์‹ค์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ํ•ฉ์„ ์œ„ํ•ด์„œ๋Š” ๋‹ค์Œ์˜ ๋ณด์กฐ์ •๋ฆฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(c:=1-1/\beta\in(0,1)\), ๊ทธ๋ฆฌ๊ณ  ํฐ \(k\)์— ๋Œ€ํ•ด \(\sum_{i\geq k+1}i^{-\beta}\asymp k^{-(\beta-1)}\)๋ผ๋Š” ๊ธฐ๋ณธ์ ์ธ ์‚ฌ์‹ค์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ํ•ฉ์˜ ๊ฒฝ์šฐ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ ˆ๋งˆ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
For both downstream and perplexity evaluation, we use our fixed evaluation pipeline to compare results against publicly available models. We also report a separate evaluation of our adapted model.
๋‹ค์šด์ŠคํŠธ๋ฆผ ํ‰๊ฐ€์™€ ํผํ”Œ๋ ‰์‹œํ‹ฐ ํ‰๊ฐ€ ๋ชจ๋‘์—์„œ, ์šฐ๋ฆฌ๋Š” ๊ณ ์ •๋œ ํ‰๊ฐ€ ํŒŒ์ดํ”„๋ผ์ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐœ๋œ ๋ชจ๋ธ๋“ค๊ณผ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ ์‘๋œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๋ณ„๋„์˜ ํ‰๊ฐ€๋„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.
๋‹ค์šด์ŠคํŠธ๋ฆผ ๋ฐ ๋‹นํ™ฉ์Šค๋Ÿฌ์›€ ํ‰๊ฐ€ ๋ชจ๋‘์— ๋Œ€ํ•ด ๊ณ ์ •๋œ ํ‰๊ฐ€ ํŒŒ์ดํ”„๋ผ์ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ๊ณผ์˜ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•œ๋‹ค. ๋˜ํ•œ ์šฐ๋ฆฌ๋Š” ์šฐ๋ฆฌ์˜ ์ ์‘๋œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๋ณ„๋„์˜ ํ‰๊ฐ€๋ฅผ ๋ณด๊ณ ํ•œ๋‹ค.
##There are certain important aspects of college preparation. The student preparing for colleges should define their goals and interests. The student should also plan financially and academically.
## ๋Œ€ํ•™ ์ค€๋น„์—๋Š” ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์ธก๋ฉด์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ•™์„ ์ค€๋น„ํ•˜๋Š” ํ•™์ƒ์€ ์ž์‹ ์˜ ๋ชฉํ‘œ์™€ ๊ด€์‹ฌ์‚ฌ๋ฅผ ์ •์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•™์ƒ์€ ์žฌ์ •์  ๋ฐ ํ•™๋ฌธ์ ์œผ๋กœ ๊ณ„ํš์„ ์„ธ์›Œ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
## ๋Œ€ํ•™ ์ค€๋น„์—๋Š” ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์ธก๋ฉด์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ•™์„ ์ค€๋น„ํ•˜๋Š” ํ•™์ƒ์€ ์ž์‹ ์˜ ๋ชฉํ‘œ์™€ ๊ด€์‹ฌ์‚ฌ๋ฅผ ์ •์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•™์ƒ์€ ์žฌ์ •์ ์œผ๋กœ๋‚˜ ํ•™์—…์ ์œผ๋กœ ๊ณ„ํš์„ ์„ธ์›Œ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
However, there are also some risks associated with HIIT, including the risk of injury, especially if the athlete is not properly warmed up or does not have adequate recovery time between workouts.
๊ทธ๋Ÿฌ๋‚˜ HIIT์—๋Š” ๋ถ€์ƒ์˜ ์œ„ํ—˜์„ ํฌํ•จํ•œ ๋ช‡ ๊ฐ€์ง€ ์œ„ํ—˜์ด ์žˆ์œผ๋ฉฐ, ํŠนํžˆ ์šด๋™ ์„ ์ˆ˜๊ฐ€ ์ œ๋Œ€๋กœ ์ค€๋น„ ์šด๋™์„ ํ•˜์ง€ ์•Š์•˜๊ฑฐ๋‚˜ ์šด๋™ ์‚ฌ์ด์— ์ถฉ๋ถ„ํ•œ ํšŒ๋ณต ์‹œ๊ฐ„์„ ๊ฐ€์ง€์ง€ ์•Š์€ ๊ฒฝ์šฐ์— ๊ทธ๋Ÿฌํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ HIIT์—๋Š” ๋ถ€์ƒ ์œ„ํ—˜, ํŠนํžˆ ์šด๋™ ์„ ์ˆ˜๊ฐ€ ์ œ๋Œ€๋กœ ์›Œ๋ฐ์—…์„ ํ•˜์ง€ ์•Š์•˜๊ฑฐ๋‚˜ ์šด๋™ ์‚ฌ์ด์— ์ถฉ๋ถ„ํ•œ ํšŒ๋ณต ์‹œ๊ฐ„์ด ์—†๋Š” ๊ฒฝ์šฐ ๋ถ€์ƒ ์œ„ํ—˜ ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ์œ„ํ—˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
pute _success rate_ for ALFWorld and _average score/success rate_ for WebShop. Further, we also follow ReAct [449] to reduce the length of the input prompt and utilize line break as the EOS token.
ALFWorld์˜ _์„ฑ๊ณต๋ฅ _๊ณผ WebShop์˜ _ํ‰๊ท  ์ ์ˆ˜/์„ฑ๊ณต๋ฅ _์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์˜ ๊ธธ์ด๋ฅผ ์ค„์ด๊ณ  EOS ํ† ํฐ์œผ๋กœ ์ค„ ๋ฐ”๊ฟˆ์„ ์‚ฌ์šฉํ•˜๋Š” ReAct [449]๋ฅผ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค.
ALFWorld์˜ _์„ฑ๊ณต๋ฅ _๊ณผ WebShop์˜ _ํ‰๊ท  ์ ์ˆ˜/์„ฑ๊ณต๋ฅ _์„ ์ถ”๊ฐ€๋กœ ๊ณ„์‚ฐํ•œ๋‹ค. ๋˜ํ•œ ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์˜ ๊ธธ์ด๋ฅผ ์ค„์ด๊ณ  ์ค„ ๋ฐ”๊ฟˆ์„ EOS ํ† ํฐ์œผ๋กœ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ReAct [449]์„ ๋”ฐ๋ฅด๊ธฐ๋„ ํ•œ๋‹ค.
To address these problems and improve encoding efficiency, we propose to extend LLaMA vocabulary with additional Chinese tokens and adapt the model for the extended vocabulary (Yang et al., 2022).
์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ  ์ธ์ฝ”๋”ฉ ํšจ์œจ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” LLaMA ์–ดํœ˜์— ์ถ”๊ฐ€์ ์ธ ์ค‘๊ตญ์–ด ํ† ํฐ์„ ํ™•์žฅํ•˜๊ณ  ํ™•์žฅ๋œ ์–ดํœ˜์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์กฐ์ •ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค (Yang et al., 2022).
์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ  ์ธ์ฝ”๋”ฉ ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ์ค‘๊ตญ์–ด ํ† ํฐ์œผ๋กœ LLaMA ์–ดํœ˜๋ฅผ ํ™•์žฅํ•˜๊ณ  ํ™•์žฅ๋œ ์–ดํœ˜์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์กฐ์ •ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•œ๋‹ค(Yang et al., 2022).
Clinical history-taking and diagnostic dialogue through which clinicians derive diagnosis and management plans represent a complex skill [19] whose optimal conduct is highly dependent on context.
์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™”๋Š” ์ž„์ƒ์˜๊ฐ€ ์ง„๋‹จ ๋ฐ ๊ด€๋ฆฌ ๊ณ„ํš์„ ๋„์ถœํ•˜๋Š” ๋ณต์žกํ•œ ๊ธฐ์ˆ ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ๊ทธ ์ตœ์ ์˜ ์ˆ˜ํ–‰์€ ์ƒํ™ฉ์— ํฌ๊ฒŒ ์˜์กดํ•ฉ๋‹ˆ๋‹ค [19].
์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ž„์ƒ์˜๊ฐ€ ์ง„๋‹จ ๋ฐ ๊ด€๋ฆฌ ๊ณ„ํš์„ ๋„์ถœํ•˜๋Š” ์ง„๋‹จ ๋Œ€ํ™”๋Š” ์ตœ์ ์˜ ์ˆ˜ํ–‰์ด ์ƒํ™ฉ์— ๋”ฐ๋ผ ํฌ๊ฒŒ ๋‹ฌ๋ผ์ง€๋Š” ๋ณต์žกํ•œ ๊ธฐ์ˆ  [19]์„ ๋‚˜ํƒ€๋‚ธ๋‹ค.