question
stringlengths
101
4.47k
chosen
stringlengths
3
4.25k
rejected
stringlengths
39
9.29k
where \(\left(\frac{N_{c}}{N}\right)^{\alpha_{N}}\) is Eq. 3.1, and \(S_{c}\) and \(\alpha_{S}\) are constant scalars to be estimated.
์—ฌ๊ธฐ์„œ \(\left(\frac{N_{c}}{N}\right)^{\alpha_{N}}\)๋Š” ์‹ 3.1์ด๋ฉฐ, \(S_{c}\)์™€ \(\alpha_{S}\)๋Š” ์ถ”์ •ํ•ด์•ผ ํ•  ์ƒ์ˆ˜ ์Šค์นผ๋ผ์ž…๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(\left(\frac{N_{c}}{N}\right)^{\alpha_{N}}\)๋Š” ์‹ 3.1์ด๊ณ , \(S_{c}\)์™€ \(\alpha_{S}\)๋Š” ์ถ”์ •ํ•ด์•ผ ํ•  ์ƒ์ˆ˜ ์Šค์นผ๋ผ์ด๋‹ค.
_Synthetic QAs: question 1:_ "Who was John Mayne?" _answer 1:_ "John Mayne was a Scottish printer, journalist, and poet from Dumfries.
_Synthetic QAs: question 1:_ "John Mayne์€ ๋ˆ„๊ตฌ์˜€๋‚˜์š”?" _answer 1:_ "John Mayne์€ Dumfries ์ถœ์‹ ์˜ ์Šค์ฝ”ํ‹€๋žœ๋“œ ์ธ์‡„์—…์ž, ๊ธฐ์ž, ๊ทธ๋ฆฌ๊ณ  ์‹œ์ธ์ด์—ˆ์Šต๋‹ˆ๋‹ค.
_ํ•ฉ์„ฑ QA: ์งˆ๋ฌธ 1:_ "์กด ๋ฉ”์ธ์€ ๋ˆ„๊ตฌ์˜€๋‚˜์š”?" _๋‹ต๋ณ€ 1:_ "์กด ๋ฉ”์ธ์€ ๋คํ”„๋ฆฌ์Šค ์ถœ์‹ ์˜ ์Šค์ฝ”ํ‹€๋žœ๋“œ ์ธ์‡„์—…์ž, ์–ธ๋ก ์ธ, ์‹œ์ธ์ด์—ˆ์Šต๋‹ˆ๋‹ค.
We say that the model \(\mathcal{M}\) possesses _any_ knowledge regarding the set of questions \(\mathcal{Q}\) if the following holds:
๋ชจ๋ธ \(\mathcal{M}\)์ด ๋‹ค์Œ ์กฐ๊ฑด์„ ๋งŒ์กฑํ•  ๊ฒฝ์šฐ, ์งˆ๋ฌธ ์ง‘ํ•ฉ \(\mathcal{Q}\)์— ๊ด€ํ•œ _์–ด๋– ํ•œ_ ์ง€์‹์„ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค๊ณ  ๋งํ•ฉ๋‹ˆ๋‹ค:
๋‹ค์Œ์ด ์„ฑ๋ฆฝํ•˜๋Š” ๊ฒฝ์šฐ, ์šฐ๋ฆฌ๋Š” ๋ชจ๋ธ \(mathcal{M}\)์ด ์งˆ๋ฌธ ์ง‘ํ•ฉ \(mathcal{Q}\)์— ๊ด€ํ•œ _๋ชจ๋“ _ ์ง€์‹์„ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค๊ณ  ๋งํ•ฉ๋‹ˆ๋‹ค:
- International diplomatic efforts or significant pressure on Israel to de-escalate could decrease the probability of a 'Yes' outcome.
- ๊ตญ์ œ ์™ธ๊ต์  ๋…ธ๋ ฅ์ด๋‚˜ ์ด์Šค๋ผ์—˜์— ๋Œ€ํ•œ ์ƒ๋‹นํ•œ ์••๋ฐ•์ด ๊ธด์žฅ์„ ์™„ํ™”์‹œํ‚ค๋ฉด '์˜ˆ' ๊ฒฐ๊ณผ์˜ ํ™•๋ฅ ์ด ๊ฐ์†Œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
- ๊ตญ์ œ ์™ธ๊ต์  ๋…ธ๋ ฅ์ด๋‚˜ ์ด์Šค๋ผ์—˜์ด ํ™•์ „์„ ์–ต์ œํ•˜๋„๋ก ์ƒ๋‹นํ•œ ์••๋ ฅ์„ ๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ '๊ทธ๋ ‡๋‹ค' ๊ฒฐ๊ณผ์˜ ํ™•๋ฅ ์„ ๋‚ฎ์ถœ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
translation scores reflect genuinely better translations or if they simply align more closely with the evaluation model's preferences.
๋ฒˆ์—ญ ์ ์ˆ˜๊ฐ€ ์‹ค์ œ๋กœ ๋” ๋‚˜์€ ๋ฒˆ์—ญ์„ ๋ฐ˜์˜ํ•˜๋Š”์ง€ ์•„๋‹ˆ๋ฉด ๋‹จ์ˆœํžˆ ํ‰๊ฐ€ ๋ชจ๋ธ์˜ ์„ ํ˜ธ๋„์— ๋” ๊ฐ€๊น๊ฒŒ ๋งž์ถฐ์ง€๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๋ฒˆ์—ญ ์ ์ˆ˜๊ฐ€ ์‹ค์ œ๋กœ ๋” ๋‚˜์€ ๋ฒˆ์—ญ์„ ๋ฐ˜์˜ํ•˜๋Š”์ง€ ์•„๋‹ˆ๋ฉด ํ‰๊ฐ€ ๋ชจ๋ธ์˜ ์„ ํ˜ธ๋„์— ๋” ๊ฐ€๊น๊ฒŒ ๋ถ€ํ•ฉํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
In this part, we review some of the popular approaches used for more cost-friendly (and compute-friendly) training and usage of LLMs.
์ด ๋ถ€๋ถ„์—์„œ๋Š” LLM์˜ ๋” ๋น„์šฉ ํšจ์œจ์ ์ด๊ณ  ๊ณ„์‚ฐ ์นœํ™”์ ์ธ ํ›ˆ๋ จ ๋ฐ ์‚ฌ์šฉ์„ ์œ„ํ•ด ์‚ฌ์šฉ๋˜๋Š” ๋ช‡ ๊ฐ€์ง€ ์ธ๊ธฐ ์žˆ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ถ€๋ถ„์—์„œ๋Š” LLM์„ ๋ณด๋‹ค ๋น„์šฉ ์นœํ™”์ ์ด๊ณ  ์ปดํ“จํŒ… ์นœํ™”์ ์œผ๋กœ ํ›ˆ๋ จํ•˜๊ณ  ์‚ฌ์šฉํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๋ช‡ ๊ฐ€์ง€ ์ธ๊ธฐ ์žˆ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค.
Figure 15 \(|\) R-GSM example where both the original and the reordered problems were correctly solved by all LLMs in our evaluation.
Figure 15 \(|\) ์›๋ž˜ ๋ฌธ์ œ์™€ ์žฌ์ •๋ ฌ๋œ ๋ฌธ์ œ๊ฐ€ ๋ชจ๋‘ ์šฐ๋ฆฌ ํ‰๊ฐ€์—์„œ ๋ชจ๋“  LLM์— ์˜ํ•ด ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ•ด๊ฒฐ๋œ R-GSM ์˜ˆ์ œ.
๊ทธ๋ฆผ 15 \(|\) R-GSM ์˜ˆ์ œ์—์„œ ์›๋ž˜ ๋ฌธ์ œ์™€ ์žฌ์ •๋ ฌ๋œ ๋ฌธ์ œ๋ฅผ ๋ชจ๋‘ ํ‰๊ฐ€์— ์ฐธ์—ฌํ•œ ๋ชจ๋“  LLM์ด ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ’€์—ˆ์Šต๋‹ˆ๋‹ค.
Figure 6: Illustration of Result 8 (see Figure 14 for \(\mathsf{bioS}\) data and Figure 15 for \(\mathsf{bioD}\) data in Appendix C).
๊ทธ๋ฆผ 6: ๊ฒฐ๊ณผ 8์˜ ์„ค๋ช… (๋ถ€๋ก C์—์„œ \(\mathsf{bioS}\) ๋ฐ์ดํ„ฐ๋Š” ๊ทธ๋ฆผ 14, \(\mathsf{bioD}\) ๋ฐ์ดํ„ฐ๋Š” ๊ทธ๋ฆผ 15๋ฅผ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค).
๊ทธ๋ฆผ 6: ๊ฒฐ๊ณผ 8์˜ ์˜ˆ์‹œ(๋ถ€๋ก C์˜ ๊ทธ๋ฆผ 14๋Š” \(\mathsf{bioS}\) ๋ฐ์ดํ„ฐ, ๊ทธ๋ฆผ 15๋Š” \(\mathsf{bioD}\) ๋ฐ์ดํ„ฐ ์ฐธ์กฐ).
Proof.: Let \(e_{0j_{1}},e_{j_{1}j_{2}},\ldots,e_{j_{m-1}i}\) be the edges of the minimum distance path from \(0\) to \(i\) in \(G\).
์ฆ๋ช…: \(G\)์—์„œ \(0\)์—์„œ \(i\)๊นŒ์ง€์˜ ์ตœ์†Œ ๊ฑฐ๋ฆฌ ๊ฒฝ๋กœ์˜ ๊ฐ„์„ ์„ \(e_{0j_{1}},e_{j_{1}j_{2}},\ldots,e_{j_{m-1}i}\)๋ผ๊ณ  ํ•˜์ž.
์ฆ๋ช….: e_{0j_{1}},e_{j_{1}j_{2}},\ldots,e_{j_{m-1}i}\)๋ฅผ \(G\)์—์„œ \(0\)์—์„œ \(i\)๊นŒ์ง€์˜ ์ตœ์†Œ ๊ฑฐ๋ฆฌ ๊ฒฝ๋กœ์˜ ๊ฐ€์žฅ์ž๋ฆฌ๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค.
Figure 14 \(|\) R-GSM example where both the original and the reordered problems were correctly solved by all LLMs in our evaluation.
Figure 14 \(|\) ์›๋ž˜ ๋ฌธ์ œ์™€ ์žฌ์ •๋ ฌ๋œ ๋ฌธ์ œ๊ฐ€ ๋ชจ๋‘ ์šฐ๋ฆฌ ํ‰๊ฐ€์—์„œ ๋ชจ๋“  LLM์— ์˜ํ•ด ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ•ด๊ฒฐ๋œ R-GSM ์˜ˆ์ œ.
๊ทธ๋ฆผ 14 \(|\) R-GSM ์˜ˆ์ œ์—์„œ ์›๋ž˜ ๋ฌธ์ œ์™€ ์žฌ์ •๋ ฌ๋œ ๋ฌธ์ œ๋ฅผ ๋ชจ๋‘ ํ‰๊ฐ€์— ์ฐธ์—ฌํ•œ ๋ชจ๋“  LLM์ด ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ’€์—ˆ์Šต๋‹ˆ๋‹ค.
Project leader. Chen, Weiqi, Jingcheng and Yixuan are interns at MSRA. GitHub: Xwin-Math This repository will be continually updated.
ํ”„๋กœ์ ํŠธ ๋ฆฌ๋”. Chen, Weiqi, Jingcheng ๋ฐ Yixuan์€ MSRA์˜ ์ธํ„ด์ž…๋‹ˆ๋‹ค. GitHub: Xwin-Math ์ด ์ €์žฅ์†Œ๋Š” ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.
ํ”„๋กœ์ ํŠธ ๋ฆฌ๋”. Chen, Weiqi, Jingcheng ๋ฐ Yixuan์€ MSRA์˜ ์ธํ„ด์ž…๋‹ˆ๋‹ค. GitHub: Xwin-Math ์ด ๋ฆฌํฌ์ง€ํ† ๋ฆฌ๋Š” ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Figure 5: Stress-testing the LayerNorm extraction attack on models behind an API (a-b), and models using both RMSNorm and biases (c).
Figure 5: API ๋’ค์— ์žˆ๋Š” ๋ชจ๋ธ๋“ค์— ๋Œ€ํ•œ LayerNorm ์ถ”์ถœ ๊ณต๊ฒฉ์˜ ์ŠคํŠธ๋ ˆ์Šค ํ…Œ์ŠคํŠธ (a-b), ๊ทธ๋ฆฌ๊ณ  RMSNorm๊ณผ ๋ฐ”์ด์–ด์Šค๋ฅผ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ๋“ค (c).
๊ทธ๋ฆผ 5: API ๋’ค์— ์žˆ๋Š” ๋ชจ๋ธ(a-b)๊ณผ RMSNorm๊ณผ ๋ฐ”์ด์–ด์Šค๋ฅผ ๋ชจ๋‘ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ(c)์— ๋Œ€ํ•œ LayerNorm ์ถ”์ถœ ๊ณต๊ฒฉ์„ ์ŠคํŠธ๋ ˆ์Šค ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค.
Given a fixed model, the computational costs depend on the vocabulary size and the fertility of the tokenizer, as defined in Eq. (1).
๊ณ ์ •๋œ ๋ชจ๋ธ์ด ์ฃผ์–ด์ง€๋ฉด, ๊ณ„์‚ฐ ๋น„์šฉ์€ ์–ดํœ˜ ํฌ๊ธฐ์™€ ํ† ํฌ๋‚˜์ด์ €์˜ ์ƒ์„ฑ๋ ฅ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ์ด๋Š” ์‹ (1)์—์„œ ์ •์˜๋ฉ๋‹ˆ๋‹ค.
๊ณ ์ •๋œ ๋ชจ๋ธ์ด ์ฃผ์–ด์ง€๋ฉด, ๊ณ„์‚ฐ ๋น„์šฉ์€ (1) ์‹์— ์ •์˜๋œ ๋ฐ”์™€ ๊ฐ™์ด ์–ดํœ˜ ํฌ๊ธฐ์™€ ํ† ํฐํ™”๊ธฐ์˜ ์ƒ์‹๋ ฅ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง„๋‹ค.
In this part, we first formulate the general planning paradigm of LLMs for solving complex tasks, which is illustrated in Figure 16.
์ด ๋ถ€๋ถ„์—์„œ๋Š” ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ LLM์˜ ์ผ๋ฐ˜์ ์ธ ๊ณ„ํš ํŒจ๋Ÿฌ๋‹ค์ž„์„ ๋จผ์ € ๊ณต์‹ํ™”ํ•˜๋ฉฐ, ์ด๋Š” Figure 16์— ์„ค๋ช…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
์ด ๋ถ€๋ถ„์—์„œ๋Š” ๋จผ์ € ๋ณต์žกํ•œ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ LLM์˜ ์ผ๋ฐ˜์ ์ธ ๊ณ„ํš ํŒจ๋Ÿฌ๋‹ค์ž„์„ ๊ณต์‹ํ™”ํ•˜๋Š”๋ฐ, ์ด๋Š” ๊ทธ๋ฆผ 16์— ์„ค๋ช…๋˜์–ด ์žˆ๋‹ค.
- She has 4 less than a third as many cherry trees as apple trees. A third of the apple trees is 80/3, which is approximately 26.67.
- ๊ทธ๋…€๋Š” ์‚ฌ๊ณผ๋‚˜๋ฌด์˜ 3๋ถ„์˜ 1๋ณด๋‹ค 4๊ทธ๋ฃจ ์ ์€ ์ฒด๋ฆฌ๋‚˜๋ฌด๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ๊ณผ๋‚˜๋ฌด์˜ 3๋ถ„์˜ 1์€ 80/3์œผ๋กœ, ์ด๋Š” ๋Œ€๋žต 26.67์ž…๋‹ˆ๋‹ค.
- ์‚ฌ๊ณผ๋‚˜๋ฌด๋ณด๋‹ค ์ฒด๋ฆฌ๋‚˜๋ฌด๊ฐ€ 3๋ถ„์˜ 1๋„ ์•ˆ ๋˜๋Š” 4๊ทธ๋ฃจ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ๊ณผ๋‚˜๋ฌด์˜ 3๋ถ„์˜ 1์€ 80/3์œผ๋กœ ์•ฝ 26.67์ž…๋‹ˆ๋‹ค.
resolution (Figure 3(c)). This may be due to the aforementioned issue: Our model hedges, even as the evidence becomes more decisive.
ํ•ด์ƒ๋„ (Figure 3(c)). ์ด๋Š” ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๋ฌธ์ œ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: ์ฆ๊ฑฐ๊ฐ€ ๋” ๊ฒฐ์ •์ ์ด ๋˜์–ด๋„ ์šฐ๋ฆฌ ๋ชจ๋ธ์ด ์ฃผ์ €ํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
ํ•ด์ƒ๋„(๊ทธ๋ฆผ 3(c))๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ์ด๋Š” ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๋ฌธ์ œ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ๋‹ค: ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ์ฆ๊ฑฐ๊ฐ€ ๋” ๊ฒฐ์ •์ ์ด ๋˜๋”๋ผ๋„ ํ—ค์ง€ํ•œ๋‹ค.
and we recover the classical Hutter result! Thus, our test metric defined in (17) is pointing in the right direction, conceptually.
๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ๋Š” ๊ณ ์ „์ ์ธ Hutter ๊ฒฐ๊ณผ๋ฅผ ๋ณต์›ํ•ฉ๋‹ˆ๋‹ค! ๋”ฐ๋ผ์„œ (17)์—์„œ ์ •์˜๋œ ์šฐ๋ฆฌ์˜ ํ…Œ์ŠคํŠธ ๋ฉ”ํŠธ๋ฆญ์€ ๊ฐœ๋…์ ์œผ๋กœ ์˜ฌ๋ฐ”๋ฅธ ๋ฐฉํ–ฅ์„ ๊ฐ€๋ฆฌํ‚ค๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ๋Š” ๊ณ ์ „์ ์ธ ํ—ˆํ„ฐ์˜ ๊ฒฐ๊ณผ๋ฅผ ํšŒ๋ณตํ•œ๋‹ค! ๋”ฐ๋ผ์„œ (17)์—์„œ ์ •์˜ํ•œ ํ…Œ์ŠคํŠธ ๋ฉ”ํŠธ๋ฆญ์€ ๊ฐœ๋…์ ์œผ๋กœ ์˜ฌ๋ฐ”๋ฅธ ๋ฐฉํ–ฅ์„ ๊ฐ€๋ฆฌํ‚ค๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Footnote 1: All resources are publicly available at [https://github.com/tLabruna/Adapt-LLM](https://github.com/tLabruna/Adapt-LLM).
๊ฐ์ฃผ 1: ๋ชจ๋“  ์ž๋ฃŒ๋Š” [https://github.com/tLabruna/Adapt-LLM](https://github.com/tLabruna/Adapt-LLM)์—์„œ ๊ณต๊ฐœ์ ์œผ๋กœ ์ด์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฐ์ฃผ 1: ๋ชจ๋“  ๋ฆฌ์†Œ์Šค๋Š” [https://github.com/tLabruna/Adapt-LLM](https://github.com/tLabruna/Adapt-LLM)์—์„œ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Fig. 27: Subsequent stages of Macrodata Refinement remove nearly 90% of the documents originally in CommonCrawl. Courtesy of [124].
Fig. 27: Macrodata Refinement์˜ ํ›„์† ๋‹จ๊ณ„๋Š” CommonCrawl์— ์›๋ž˜ ์žˆ๋˜ ๋ฌธ์„œ์˜ ๊ฑฐ์˜ 90%๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. [124]์˜ ์ œ๊ณต.
๊ทธ๋ฆผ 27: ๋งคํฌ๋กœ๋ฐ์ดํ„ฐ ๊ตฌ์ฒดํ™”์˜ ํ›„์† ๋‹จ๊ณ„์—์„œ๋Š” ์›๋ž˜ CommonCrawl์— ์žˆ๋˜ ๋ฌธ์„œ์˜ ๊ฑฐ์˜ 90%๊ฐ€ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. [124]์˜ ๋•๋ถ„์ž…๋‹ˆ๋‹ค.
Lower-Bound.WLOG6 consider the following specific choice of permutations defined by \(\pi_{i}(j)=j\) (i.e doesn't depend on \(i\)).
ํ•˜ํ•œ. WLOG6๋Š” \(\pi_{i}(j)=j\)๋กœ ์ •์˜๋œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŠน์ • ์ˆœ์—ด ์„ ํƒ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค (์ฆ‰, \(i\)์— ์˜์กดํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค).
Lower-Bound.WLOG6๋Š” \(\pi_{i}(j)=j\)(์ฆ‰, \(i\)์— ์˜์กดํ•˜์ง€ ์•Š์Œ)๋กœ ์ •์˜๋œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŠน์ • ์ˆœ์—ด์˜ ์„ ํƒ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค.
**Question:** How did Matthew Belk, a National Weather Service meteorologist, describe the September 2023 northeastern U.S. floods?
**์งˆ๋ฌธ:** National Weather Service ๊ธฐ์ƒํ•™์ž์ธ Matthew Belk๋Š” 2023๋…„ 9์›” ๋ถ๋™๋ถ€ ๋ฏธ๊ตญ ํ™์ˆ˜๋ฅผ ์–ด๋–ป๊ฒŒ ์„ค๋ช…ํ–ˆ๋‚˜์š”?
**์งˆ๋ฌธ:** ๋ฏธ๊ตญ ๊ตญ๋ฆฝ๊ธฐ์ƒ์ฒญ ๊ธฐ์ƒํ•™์ž์ธ ๋งคํŠœ ๋ฒจํฌ๋Š” 2023๋…„ 9์›” ๋ฏธ๊ตญ ๋ถ๋™๋ถ€ ํ™์ˆ˜๋ฅผ ์–ด๋–ป๊ฒŒ ์„ค๋ช…ํ–ˆ๋‚˜์š”?
Jurning meets this report that a sheer of standing increases only inside white attack light, serving an acronym model in the story.
Jurning์€ ์ด ๋ณด๊ณ ์„œ์—์„œ ๋ฐฑ์ƒ‰ ๊ณต๊ฒฉ ์กฐ๋ช… ๋‚ด์—์„œ๋งŒ ์„œ ์žˆ๋Š” ๋น„์œจ์ด ์ฆ๊ฐ€ํ•œ๋‹ค๊ณ  ๋ณด๊ณ ํ•˜๋ฉฐ, ์ด์•ผ๊ธฐ์—์„œ ์•ฝ์–ด ๋ชจ๋ธ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์ฃผ์–ด์ง„ ๋ณด๊ณ ์„œ์— ๋”ฐ๋ฅด๋ฉด, ์ˆœ์ „ํžˆ ์„œ ์žˆ๋Š” ์‚ฌ๋žŒ์€ ๋ฐฑ์ƒ‰ ๊ณต๊ฒฉ ์กฐ๋ช… ๋‚ด๋ถ€์—์„œ๋งŒ ์ฆ๊ฐ€ํ•˜๋ฉฐ, ์ด๋Š” ์Šคํ† ๋ฆฌ์—์„œ ์•ฝ์–ด ๋ชจ๋ธ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
To understand our goal better, we expand on our analogy between training an LLM in the real-world setting of preparing for an exam.
์šฐ๋ฆฌ์˜ ๋ชฉํ‘œ๋ฅผ ๋” ์ž˜ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด, ์‹œํ—˜ ์ค€๋น„๋ผ๋Š” ์‹ค์ œ ์ƒํ™ฉ์—์„œ LLM์„ ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๊ฒƒ์— ๋Œ€ํ•œ ๋น„์œ ๋ฅผ ํ™•์žฅํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.
๋ชฉํ‘œ๋ฅผ ๋” ์ž˜ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์‹œํ—˜์„ ์ค€๋น„ํ•˜๋Š” ์‹ค์ œ ํ™˜๊ฒฝ์—์„œ LLM์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ๊ณผ์˜ ๋น„์œ ๋ฅผ ํ™•์žฅํ•œ๋‹ค.
**Question**: How close did the doctor's differential diagnosis (DDx) come to including the PROBABLE DIAGNOSIS from the answer key?
**์งˆ๋ฌธ**: ์˜์‚ฌ์˜ ๊ฐ๋ณ„ ์ง„๋‹จ(DDx)์ด ์ •๋‹ต์ง€์˜ ๊ฐ€๋Šฅํ•œ ์ง„๋‹จ(PROBABLE DIAGNOSIS)์„ ํฌํ•จํ•˜๋Š” ์ •๋„๋Š” ์–ด๋Š ์ •๋„์˜€์Šต๋‹ˆ๊นŒ?
**์งˆ๋ฌธ**: ์˜์‚ฌ์˜ ๊ฐ๋ณ„ ์ง„๋‹จ(DDx)์ด ๋‹ต์•ˆ์ง€์˜ PROBABLE DIAGNOSIS์— ์–ผ๋งˆ๋‚˜ ๊ทผ์ ‘ํ–ˆ๋‚˜์š”?
- Unforeseen events could negatively impact 'Barbie's attendance, such as controversies or widespread technical issues in theaters.
- ์˜ˆ๊ธฐ์น˜ ์•Š์€ ์‚ฌ๊ฑด๋“ค์€ ๋…ผ๋ž€์ด๋‚˜ ๊ทน์žฅ์—์„œ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ๊ธฐ์ˆ ์  ๋ฌธ์ œ์™€ ๊ฐ™์€ '๋ฐ”๋น„'์˜ ๊ด€๊ฐ ์ˆ˜์— ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
- ์˜ˆ์ƒ์น˜ ๋ชปํ•œ ์‚ฌ๊ฑด(์˜ˆ: ๋…ผ๋ž€์ด๋‚˜ ๊ทน์žฅ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ๊ธฐ์ˆ  ๋ฌธ์ œ ๋“ฑ)์ด ๋ฐœ์ƒํ•˜์—ฌ '๋ฐ”๋น„'์˜ ๊ด€๊ฐ ์ˆ˜์— ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 8: Performance of Baichuan-7B, DeepSeek-7B, DeepSeek-67B, Amber-7B, OpenLLaMA-7B, and Yi-34B models across code benchmarks.
๊ทธ๋ฆผ 8: Baichuan-7B, DeepSeek-7B, DeepSeek-67B, Amber-7B, OpenLLaMA-7B, ๊ทธ๋ฆฌ๊ณ  Yi-34B ๋ชจ๋ธ์˜ ์ฝ”๋“œ ๋ฒค์น˜๋งˆํฌ ์„ฑ๋Šฅ.
๊ทธ๋ฆผ 8: ์ฝ”๋“œ ๋ฒค์น˜๋งˆํฌ์—์„œ Baichuan-7B, DeepSeek-7B, DeepSeek-67B, Amber-7B, OpenLLaMA-7B ๋ฐ Yi-34B ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ.
21. The story of โ€œDivine Rivalsโ€ explores themes of love, loyalty, and the difficult choices faced by individuals in times of war.
"Divine Rivals"์˜ ์ด์•ผ๊ธฐ๋Š” ์‚ฌ๋ž‘, ์ถฉ์„ฑ์‹ฌ, ๊ทธ๋ฆฌ๊ณ  ์ „์Ÿ ์‹œ๊ธฐ์— ๊ฐœ์ธ์ด ์ง๋ฉดํ•˜๋Š” ์–ด๋ ค์šด ์„ ํƒ์ด๋ผ๋Š” ์ฃผ์ œ๋ฅผ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
21. "์‹ ์„ฑํ•œ ๋ผ์ด๋ฒŒ"์˜ ์ด์•ผ๊ธฐ๋Š” ์‚ฌ๋ž‘, ์ถฉ์„ฑ์‹ฌ, ์ „์Ÿ ์ค‘ ๊ฐœ์ธ์ด ์ง๋ฉดํ•˜๋Š” ์–ด๋ ค์šด ์„ ํƒ์ด๋ผ๋Š” ์ฃผ์ œ๋ฅผ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
Figure 9: Performance of Baichuan-7B, DeepSeek-7B, DeepSeek-67B, Amber-7B, OpenLLaMA-7B, and Yi-34B models across math benchmarks.
Figure 9: Baichuan-7B, DeepSeek-7B, DeepSeek-67B, Amber-7B, OpenLLaMA-7B, Yi-34B ๋ชจ๋ธ์˜ ์ˆ˜ํ•™ ๋ฒค์น˜๋งˆํฌ ์„ฑ๋Šฅ.
๊ทธ๋ฆผ 9: ์ˆ˜ํ•™ ๋ฒค์น˜๋งˆํฌ์—์„œ Baichuan-7B, DeepSeek-7B, DeepSeek-67B, Amber-7B, OpenLLaMA-7B ๋ฐ Yi-34B ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ.
We introduce a token selection ratio \(k\%\), which determines the proportion of tokens to be included based on their excess loss.
ํ† ํฐ ์„ ํƒ ๋น„์œจ \(k\%\)์„ ๋„์ž…ํ•˜์—ฌ, ์ดˆ๊ณผ ์†์‹ค์— ๊ธฐ๋ฐ˜ํ•ด ํฌํ•จ๋  ํ† ํฐ์˜ ๋น„์œจ์„ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค.
ํ† ํฐ ์„ ํƒ ๋น„์œจ \(k\%\)์„ ๋„์ž…ํ•˜์—ฌ ์ดˆ๊ณผ ์†์‹ค์„ ๊ธฐ์ค€์œผ๋กœ ํฌํ•จํ•  ํ† ํฐ์˜ ๋น„์ค‘์„ ๊ฒฐ์ •ํ•œ๋‹ค.
21. The story of "Divine Rivals" explores themes of love, loyalty, and the difficult choices faced by individuals in times of war.
"Divine Rivals"์˜ ์ด์•ผ๊ธฐ๋Š” ์‚ฌ๋ž‘, ์ถฉ์„ฑ์‹ฌ, ๊ทธ๋ฆฌ๊ณ  ์ „์Ÿ ์‹œ๊ธฐ์— ๊ฐœ์ธ์ด ์ง๋ฉดํ•˜๋Š” ์–ด๋ ค์šด ์„ ํƒ์ด๋ผ๋Š” ์ฃผ์ œ๋ฅผ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
21. "์‹ ์„ฑํ•œ ๋ผ์ด๋ฒŒ"์˜ ์ด์•ผ๊ธฐ๋Š” ์‚ฌ๋ž‘, ์ถฉ์„ฑ์‹ฌ, ์ „์Ÿ ์ค‘ ๊ฐœ์ธ์ด ์ง๋ฉดํ•˜๋Š” ์–ด๋ ค์šด ์„ ํƒ์ด๋ผ๋Š” ์ฃผ์ œ๋ฅผ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
In this section, we provide appropriate background and methodology as it relates to continual pre-training in the context of LLMs.
์ด ์„น์…˜์—์„œ๋Š” LLMs์˜ ๋งฅ๋ฝ์—์„œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ๊ณผ ๊ด€๋ จ๋œ ์ ์ ˆํ•œ ๋ฐฐ๊ฒฝ๊ณผ ๋ฐฉ๋ฒ•๋ก ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” LLM์˜ ๋งฅ๋ฝ์—์„œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต๊ณผ ๊ด€๋ จํ•˜์—ฌ ์ ์ ˆํ•œ ๋ฐฐ๊ฒฝ ์ง€์‹๊ณผ ๋ฐฉ๋ฒ•๋ก ์„ ์ œ๊ณตํ•œ๋‹ค.
- He might not want to share the stage with candidates he perceives as less significant, which could dilute his brand or message.
- ๊ทธ๋Š” ์ž์‹ ์ด ๋œ ์ค‘์š”ํ•˜๋‹ค๊ณ  ์—ฌ๊ธฐ๋Š” ํ›„๋ณด๋“ค๊ณผ ๋ฌด๋Œ€๋ฅผ ๊ณต์œ ํ•˜๊ณ  ์‹ถ์ง€ ์•Š์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๊ทธ์˜ ๋ธŒ๋žœ๋“œ๋‚˜ ๋ฉ”์‹œ์ง€๋ฅผ ํฌ์„์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
- ๊ทธ๋Š” ์ž์‹ ๋ณด๋‹ค ๋œ ์ค‘์š”ํ•˜๋‹ค๊ณ  ์ธ์‹ํ•˜๋Š” ํ›„๋ณด์ž๋“ค๊ณผ ๋ฌด๋Œ€๋ฅผ ๊ณต์œ ํ•˜๊ณ  ์‹ถ์ง€ ์•Š์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๊ทธ์˜ ๋ธŒ๋žœ๋“œ๋‚˜ ๋ฉ”์‹œ์ง€๋ฅผ ํฌ์„์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
- There could be concerns about the debate format or the choice of moderators, which might lead to a decision not to participate.
- ํ† ๋ก  ํ˜•์‹์ด๋‚˜ ์ง„ํ–‰์ž์˜ ์„ ํƒ์— ๋Œ€ํ•œ ์šฐ๋ ค๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ฐธ์—ฌํ•˜์ง€ ์•Š๊ธฐ๋กœ ๊ฒฐ์ •ํ•˜๋Š” ์›์ธ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
- ํ† ๋ก  ํ˜•์‹์ด๋‚˜ ์ค‘์žฌ์ž ์„ ์ •์— ๋Œ€ํ•œ ์šฐ๋ ค๊ฐ€ ์žˆ์–ด ์ฐธ์—ฌํ•˜์ง€ ์•Š๊ธฐ๋กœ ๊ฒฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 2: Above is the data processing flow and deduplication ratios, below is a schematic diagram of similar line deduplication.
Figure 2: ์œ„์—๋Š” ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ ํ๋ฆ„๊ณผ ์ค‘๋ณต ์ œ๊ฑฐ ๋น„์œจ์ด ์žˆ์œผ๋ฉฐ, ์•„๋ž˜์—๋Š” ์œ ์‚ฌํ•œ ๋ผ์ธ ์ค‘๋ณต ์ œ๊ฑฐ์˜ ๊ฐœ๋žต๋„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2: ์œ„๋Š” ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ ํ๋ฆ„๊ณผ ์ค‘๋ณต ์ œ๊ฑฐ ๋น„์œจ, ์•„๋ž˜๋Š” ์œ ์‚ฌํ•œ ๋ผ์ธ ์ค‘๋ณต ์ œ๊ฑฐ์˜ ๋ชจ์‹๋„์ด๋‹ค.
- Generate brief search queries (up to {max_words} words each) to gather information on Google that could influence the forecast.
- ์˜ˆ์ธก์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ์ •๋ณด๋ฅผ Google์—์„œ ์ˆ˜์ง‘ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ„๋‹จํ•œ ๊ฒ€์ƒ‰ ์ฟผ๋ฆฌ(๊ฐ๊ฐ ์ตœ๋Œ€ {max_words} ๋‹จ์–ด)๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
- ์˜ˆ์ธก์— ์˜ํ–ฅ์„ ์ค„ ์ˆ˜ ์žˆ๋Š” Google์—์„œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ„๋‹จํ•œ ๊ฒ€์ƒ‰์–ด(๊ฐ๊ฐ ์ตœ๋Œ€ {์ตœ๋Œ€_๋‹จ์–ด} ๋‹จ์–ด)๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
A third scenario arises due to the automatic question generation process, some questions strongly rely on pre-existing knowledge.
์„ธ ๋ฒˆ์งธ ์‹œ๋‚˜๋ฆฌ์˜ค๋Š” ์ž๋™ ์งˆ๋ฌธ ์ƒ์„ฑ ๊ณผ์ •์—์„œ ๋ฐœ์ƒํ•˜๋ฉฐ, ์ผ๋ถ€ ์งˆ๋ฌธ์€ ๊ธฐ์กด ์ง€์‹์— ํฌ๊ฒŒ ์˜์กดํ•ฉ๋‹ˆ๋‹ค.
์ž๋™ ์งˆ๋ฌธ ์ƒ์„ฑ ๊ณผ์ •์œผ๋กœ ์ธํ•ด ์„ธ ๋ฒˆ์งธ ์‹œ๋‚˜๋ฆฌ์˜ค๋Š” ์ผ๋ถ€ ์งˆ๋ฌธ์ด ๊ธฐ์กด ์ง€์‹์— ํฌ๊ฒŒ ์˜์กดํ•œ๋‹ค.
Footnote 5: Note that according to our third assumption, the quality of the training data is uniform and independent of the size.
Footnote 5: ์šฐ๋ฆฌ์˜ ์„ธ ๋ฒˆ์งธ ๊ฐ€์ •์— ๋”ฐ๋ฅด๋ฉด, ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์€ ํฌ๊ธฐ์™€ ๋ฌด๊ด€ํ•˜๊ฒŒ ๊ท ์ผํ•˜๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์‹ญ์‹œ์˜ค.
๊ฐ์ฃผ 5: ์šฐ๋ฆฌ์˜ ์„ธ ๋ฒˆ์งธ ๊ฐ€์ •์— ๋”ฐ๋ฅด๋ฉด, ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ์€ ๊ท ์ผํ•˜๋ฉฐ ํฌ๊ธฐ์™€ ๋ฌด๊ด€ํ•˜๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์‹ญ์‹œ์˜ค.
Mr. Price, from heart for their unconventional methods and reputation per unit, Jurning and Arrays include Mrs. Price, with 1984.
Price ์”จ๋Š” ๊ทธ๋“ค์˜ ๋น„์ „ํ†ต์ ์ธ ๋ฐฉ๋ฒ•๊ณผ ํ‰ํŒ์œผ๋กœ ์ธํ•ด Jurning๊ณผ Arrays์— ํฌํ•จ๋˜๋ฉฐ, 1984๋…„์—๋Š” Price ๋ถ€์ธ๋„ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
๊ฐ€๊ฒฉ ์”จ๋Š” ์‹ฌ์žฅ์—์„œ ๋น„์ •ํ†ต์ ์ธ ๋ฐฉ๋ฒ•๊ณผ ๋‹จ์œ„๋‹น ๋ช…์„ฑ์œผ๋กœ ์œ ๋ช…ํ•œ ์ €๋‹๊ณผ ๋ฐฐ์—ด์„ ํฌํ•จํ•˜์—ฌ 1984๋…„ ๊ฐ€๊ฒฉ ๋ถ€์ธ์„ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Additionally, we computed the impact of a tokenizer on the average computational costs of a given model per word during training.
์ถ”๊ฐ€๋กœ, ์šฐ๋ฆฌ๋Š” ํ›ˆ๋ จ ์ค‘ ์ฃผ์–ด์ง„ ๋ชจ๋ธ์˜ ๋‹จ์–ด๋‹น ํ‰๊ท  ๊ณ„์‚ฐ ๋น„์šฉ์— ๋Œ€ํ•œ ํ† ํฌ๋‚˜์ด์ €์˜ ์˜ํ–ฅ์„ ๊ณ„์‚ฐํ–ˆ์Šต๋‹ˆ๋‹ค.
๋˜ํ•œ, ์šฐ๋ฆฌ๋Š” ํ† ํฐํ™”๊ฐ€ ํ›ˆ๋ จ ์ค‘ ์ฃผ์–ด์ง„ ๋ชจ๋ธ์˜ ๋‹จ์–ด๋‹น ํ‰๊ท  ๊ณ„์‚ฐ ๋น„์šฉ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ณ„์‚ฐํ–ˆ๋‹ค.
8. Iris Winnow's bond with Roman Kitt deepens when Roman anonymously receives and responds to Iris's letters intended for Forest.
Iris Winnow์™€ Roman Kitt์˜ ์œ ๋Œ€๋Š” Roman์ด ์ต๋ช…์œผ๋กœ Forest์—๊ฒŒ ๋ณด๋‚ด๋ ค๋˜ Iris์˜ ํŽธ์ง€๋ฅผ ๋ฐ›๊ณ  ๋‹ต์žฅ์„ ํ•˜๋ฉด์„œ ๊นŠ์–ด์ง‘๋‹ˆ๋‹ค.
8. ๋กœ๋งŒ ํ‚คํŠธ๊ฐ€ ์•„์ด๋ฆฌ์Šค ์œˆ๋…ธ์šฐ์™€์˜ ์œ ๋Œ€๊ฐ์ด ๊นŠ์–ด์ง€๋ฉด์„œ ๋กœ๋งŒ์€ ์ˆฒ์„ ์œ„ํ•œ ์•„์ด๋ฆฌ์Šค์˜ ํŽธ์ง€๋ฅผ ์ต๋ช…์œผ๋กœ ๋ฐ›๊ณ  ์‘๋‹ตํ•ฉ๋‹ˆ๋‹ค.
Footnote 1: Note that a LLM is not necessarily more capable than a small PLM, and emergent abilities may not occur in some LLMs.
Footnote 1: LLM์ด ๋ฐ˜๋“œ์‹œ ์ž‘์€ PLM๋ณด๋‹ค ๋” ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚œ ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ผ๋ถ€ LLM์—์„œ๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์ด ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฐ์ฃผ 1: LLM์ด ๋ฐ˜๋“œ์‹œ ์†Œํ˜• PLM๋ณด๋‹ค ๋” ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ผ๋ถ€ LLM์—์„œ๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์ด ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 8: Gradient statistics over the training of a 0.2B model using WSD LRS. The exponential decay stage begins at 8000 steps.
Figure 8: WSD LRS๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 0.2B ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋™์•ˆ์˜ ๊ทธ๋ž˜๋””์–ธํŠธ ํ†ต๊ณ„. ์ง€์ˆ˜ ๊ฐ์‡  ๋‹จ๊ณ„๋Š” 8000 ์Šคํ…์—์„œ ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 8: WSD LRS๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 0.2B ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋™์•ˆ์˜ ๊ธฐ์šธ๊ธฐ ํ†ต๊ณ„. ์ง€์ˆ˜ ๊ฐ์†Œ ๋‹จ๊ณ„๋Š” 8000๋‹จ๊ณ„์—์„œ ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค.
**Question:** If \(f(x)=\frac{3x-2}{x-2}\), what is the value of \(f(-2)+f(-1)+f(0)\)? Express your answer as a common fraction.
**์งˆ๋ฌธ:** \(f(x)=\frac{3x-2}{x-2}\)์ผ ๋•Œ, \(f(-2)+f(-1)+f(0)\)์˜ ๊ฐ’์€ ์–ผ๋งˆ์ž…๋‹ˆ๊นŒ? ๋‹ต์„ ์ผ๋ฐ˜ ๋ถ„์ˆ˜๋กœ ํ‘œํ˜„ํ•˜์„ธ์š”.
**์งˆ๋ฌธ:** f(x)=\frac{3x-2}{x-2}\)์ผ ๋•Œ, \(f(-2)+f(-1)+f(0)\)์˜ ๊ฐ’์€ ์–ผ๋งˆ์ž…๋‹ˆ๊นŒ? ๋‹ต์„ ๊ณตํ†ต ๋ถ„์ˆ˜๋กœ ํ‘œํ˜„ํ•˜์„ธ์š”.
Footnote 2: Unfortunately, we deleted the logprobs for GPT-3.5 models before we created this attack due to security constraints.
Footnote 2: ๋ถˆํ–‰ํžˆ๋„, ๋ณด์•ˆ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ์ด ๊ณต๊ฒฉ์„ ๋งŒ๋“ค๊ธฐ ์ „์— GPT-3.5 ๋ชจ๋ธ์˜ logprobs๋ฅผ ์‚ญ์ œํ–ˆ์Šต๋‹ˆ๋‹ค.
๊ฐ์ฃผ 2: ์•ˆํƒ€๊น๊ฒŒ๋„ ๋ณด์•ˆ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ์ด ๊ณต๊ฒฉ์„ ๋งŒ๋“ค๊ธฐ ์ „์— GPT-3.5 ๋ชจ๋ธ์— ๋Œ€ํ•œ ๋กœ๊ทธ ํ™•๋ฅ ์„ ์‚ญ์ œํ–ˆ์Šต๋‹ˆ๋‹ค.
Furthermore, a recent study [276] defines the distance required to rotate one cycle (\(2\pi\)) for each dimension as wavelength:
๊ฒŒ๋‹ค๊ฐ€, ์ตœ๊ทผ ์—ฐ๊ตฌ [276]์—์„œ๋Š” ๊ฐ ์ฐจ์›์—์„œ ํ•œ ์ฃผ๊ธฐ๋ฅผ ํšŒ์ „ํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ๊ฑฐ๋ฆฌ๋ฅผ ํŒŒ์žฅ์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค (\(2\pi\)).
๋˜ํ•œ ์ตœ๊ทผ ์—ฐ๊ตฌ [276]์—์„œ๋Š” ๊ฐ ์ฐจ์›์— ๋Œ€ํ•ด ํ•˜๋‚˜์˜ ์ฃผ๊ธฐ(\(2\pi\))๋ฅผ ํšŒ์ „ํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ๊ฑฐ๋ฆฌ๋ฅผ ํŒŒ์žฅ์œผ๋กœ ์ •์˜ํ•œ๋‹ค.
PersonA: The main ingredients are typically chile peppers, garlic, and vinegar. However, there can be variations in the recipe.
PersonA: ์ฃผ์š” ์žฌ๋ฃŒ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๊ณ ์ถ”, ๋งˆ๋Š˜, ์‹์ดˆ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ ˆ์‹œํ”ผ์—๋Š” ๋ณ€ํ˜•์ด ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ์ฃผ์š” ์žฌ๋ฃŒ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๊ณ ์ถ”, ๋งˆ๋Š˜, ์‹์ดˆ์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋ ˆ์‹œํ”ผ์— ๋ณ€ํ˜•์ด ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
and then convert the loss minimization problem from Equation 2 into the following negative log-likelihood minimization problem:
๊ทธ๋Ÿฐ ๋‹ค์Œ ์†์‹ค ์ตœ์†Œํ™” ๋ฌธ์ œ๋ฅผ ์‹ 2์—์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์Œ์˜ ๋กœ๊ทธ ๊ฐ€๋Šฅ๋„ ์ตœ์†Œํ™” ๋ฌธ์ œ๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆฌ๊ณ  ๋ฐฉ์ •์‹ 2์˜ ์†์‹ค ์ตœ์†Œํ™” ๋ฌธ์ œ๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์Œ์˜ ๋กœ๊ทธ ๊ฐ€๋Šฅ์„ฑ ์ตœ์†Œํ™” ๋ฌธ์ œ๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค:
AI Agent Agent Architecture AI Reasoning Planning Tool Calling Single Agent Multi Agent Agent Survey LLM Agent Autonomous Agent
AI ์—์ด์ „ํŠธ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜ AI ์ถ”๋ก  ๊ณ„ํš ๋„๊ตฌ ํ˜ธ์ถœ ๋‹จ์ผ ์—์ด์ „ํŠธ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์—์ด์ „ํŠธ ์กฐ์‚ฌ LLM ์—์ด์ „ํŠธ ์ž์œจ ์—์ด์ „ํŠธ
AI ์—์ด์ „ํŠธ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜ AI ์ถ”๋ก  ๊ณ„ํš ๋„๊ตฌ ํ˜ธ์ถœ ๋‹จ์ผ ์—์ด์ „ํŠธ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์—์ด์ „ํŠธ ์„ค๋ฌธ์กฐ์‚ฌ LLM ์—์ด์ „ํŠธ ์ž์œจ ์—์ด์ „ํŠธ
DPO is relatively straightforward to implement; PyTorch code for the DPO loss is provided below:import torch.nn.functional as F
DPO๋Š” ๊ตฌํ˜„ํ•˜๊ธฐ ๋น„๊ต์  ๊ฐ„๋‹จํ•ฉ๋‹ˆ๋‹ค. DPO ์†์‹ค์— ๋Œ€ํ•œ PyTorch ์ฝ”๋“œ๋Š” ์•„๋ž˜์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค: import torch.nn.functional as F
DPO๋Š” ๊ตฌํ˜„์ด ๋น„๊ต์  ๊ฐ„๋‹จํ•˜๋ฉฐ, ์•„๋ž˜๋Š” DPO ์†์‹ค์— ๋Œ€ํ•œ PyTorch ์ฝ”๋“œ์ž…๋‹ˆ๋‹ค:import torch.nn.functional as F
Footnote 3: The impact of using different evaluation models, such as only using XCOMET or KIWI-XXL, is explored in Section 5.1.
Footnote 3: XCOMET ๋˜๋Š” KIWI-XXL๋งŒ ์‚ฌ์šฉํ•˜๋Š” ๋“ฑ ๋‹ค์–‘ํ•œ ํ‰๊ฐ€ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๋Š” ์˜ํ–ฅ์€ 5.1์ ˆ์—์„œ ํƒ๊ตฌ๋ฉ๋‹ˆ๋‹ค.
๊ฐ์ฃผ 3: XCOMET ๋˜๋Š” KIWI-XXL๋งŒ ์‚ฌ์šฉํ•˜๋Š” ๋“ฑ ๋‹ค์–‘ํ•œ ํ‰๊ฐ€ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ์˜ ์˜ํ–ฅ์€ ์„น์…˜ 5.1์—์„œ ์‚ดํŽด๋ณธ๋‹ค.
(b) The second search query prompt we use. It first asks the model to consider sub-questions and use that to steer the outputs.
(b) ์šฐ๋ฆฌ๊ฐ€ ์‚ฌ์šฉํ•˜๋Š” ๋‘ ๋ฒˆ์งธ ๊ฒ€์ƒ‰ ์ฟผ๋ฆฌ ํ”„๋กฌํ”„ํŠธ์ž…๋‹ˆ๋‹ค. ๋จผ์ € ๋ชจ๋ธ์—๊ฒŒ ํ•˜์œ„ ์งˆ๋ฌธ์„ ๊ณ ๋ คํ•˜๋„๋ก ์š”์ฒญํ•˜๊ณ  ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ถœ๋ ฅ์„ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค.
(b) ๋‘ ๋ฒˆ์งธ ๊ฒ€์ƒ‰ ์ฟผ๋ฆฌ ํ”„๋กฌํ”„ํŠธ์ž…๋‹ˆ๋‹ค. ๋จผ์ € ๋ชจ๋ธ์— ํ•˜์œ„ ์งˆ๋ฌธ์„ ๊ณ ๋ คํ•˜๊ณ  ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ถœ๋ ฅ์„ ์œ ๋„ํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค.
Using Eqn. 8, we can fit the loss on each validation domain \(L_{i}\) and sum them up to obtain the prediction of overall loss.
Eqn. 8์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ๊ฒ€์ฆ ๋„๋ฉ”์ธ \(L_{i}\)์—์„œ ์†์‹ค์„ ๋งž์ถ”๊ณ  ์ด๋ฅผ ํ•ฉ์‚ฐํ•˜์—ฌ ์ „์ฒด ์†์‹ค์˜ ์˜ˆ์ธก๊ฐ’์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์‹ 8์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋„๋ฉ”์ธ \(L_{i}\)์˜ ์†์‹ค์„ ๋งž์ถ”๊ณ  ํ•ฉ์‚ฐํ•˜์—ฌ ์ „์ฒด ์†์‹ค์˜ ์˜ˆ์ธก์„ ์–ป์„ ์ˆ˜ ์žˆ๋‹ค.
**Theorem 5.3**: _(proof in Appendix A.2) Let's consider \((\sigma_{i})\) the singular values of \(W^{*}\) in decreasing order.
**Theorem 5.3**: _(์ฆ๋ช…์€ ๋ถ€๋ก A.2์— ์žˆ์Œ) \(W^{*}\)์˜ ํŠน์ด๊ฐ’์„ ๋‚ด๋ฆผ์ฐจ์ˆœ์œผ๋กœ \((\sigma_{i})\)๋ผ๊ณ  ๊ฐ€์ •ํ•ฉ์‹œ๋‹ค.
**์ •๋ฆฌ 5.3**: _(๋ถ€๋ก A.2์˜ ์ฆ๋ช…) Let's consider \((\sigma_{i})\) the singular values of \(W^{*}\) in decreasing order.
As the field of synthetic data continues to evolve, there are several promising directions for future research and development.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ๋ถ„์•ผ๊ฐ€ ๊ณ„์† ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ, ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐ ๊ฐœ๋ฐœ์„ ์œ„ํ•œ ์—ฌ๋Ÿฌ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ๋ถ„์•ผ๊ฐ€ ๊ณ„์† ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐ ๊ฐœ๋ฐœ์„ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
**Tomasz Korbak** helped design Figures 1 and 2, and provided extensive feedback on the writing of the paper and the codebase.
Tomasz Korbak์€ Figure 1๊ณผ Figure 2์˜ ๋””์ž์ธ์„ ๋„์™”์œผ๋ฉฐ, ๋…ผ๋ฌธ ์ž‘์„ฑ๊ณผ ์ฝ”๋“œ๋ฒ ์ด์Šค์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค.
**Tomasz Korbak**์€ ๊ทธ๋ฆผ 1๊ณผ 2์˜ ๋””์ž์ธ์„ ๋„์™”์œผ๋ฉฐ ๋…ผ๋ฌธ ์ž‘์„ฑ๊ณผ ์ฝ”๋“œ๋ฒ ์ด์Šค์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค.
1. Intensity: HIIT is more intense than other forms of exercise, which can lead to better results in a shorter amount of time.
1. ๊ฐ•๋„: HIIT๋Š” ๋‹ค๋ฅธ ํ˜•ํƒœ์˜ ์šด๋™๋ณด๋‹ค ๊ฐ•๋„๊ฐ€ ๋†’์•„ ์งง์€ ์‹œ๊ฐ„ ๋‚ด์— ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
1. ๊ฐ•๋„: HIIT๋Š” ๋‹ค๋ฅธ ํ˜•ํƒœ์˜ ์šด๋™๋ณด๋‹ค ๊ฐ•๋„๊ฐ€ ๋†’๊ธฐ ๋•Œ๋ฌธ์— ๋” ์งง์€ ์‹œ๊ฐ„์— ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The midpoint of a line segment with endpoints \((x_{1},y_{1})\) and \((x_{2},y_{2})\) can be found using the midpoint formula:
์„ ๋ถ„์˜ ๋์  \((x_{1},y_{1})\)๊ณผ \((x_{2},y_{2})\)์˜ ์ค‘์ ์„ ์ค‘์  ๊ณต์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฐพ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋์ ์ด \((x_{1},y_{1})\)์™€ \((x_{2},y_{2})\)์ธ ์„ ๋ถ„์˜ ์ค‘๊ฐ„์ ์€ ์ค‘๊ฐ„์  ๊ณต์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
Figure 35: Sequential bigram data: top p = 0.95 leads to similar effect as tail narrowing. 1000 data with sequence length 100.
Figure 35: ์ˆœ์ฐจ์ ์ธ ๋ฐ”์ด๊ทธ๋žจ ๋ฐ์ดํ„ฐ: ์ƒ์œ„ p = 0.95๋Š” ๊ผฌ๋ฆฌ ์ขํžˆ๊ธฐ์™€ ์œ ์‚ฌํ•œ ํšจ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค. ์‹œํ€€์Šค ๊ธธ์ด 100์˜ ๋ฐ์ดํ„ฐ 1000๊ฐœ.
๊ทธ๋ฆผ 35: ์ˆœ์ฐจ์  ๋ฐ”์ด๊ทธ๋žจ ๋ฐ์ดํ„ฐ: ์ƒ์œ„ p = 0.95๋Š” ๊ผฌ๋ฆฌ ์ขํžˆ๊ธฐ์™€ ์œ ์‚ฌํ•œ ํšจ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค. ์‹œํ€€์Šค ๊ธธ์ด๊ฐ€ 100์ธ 1000๊ฐœ์˜ ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค.
The Table.10 details the question structure of CHC-Bench. The following table illustrates the composition of the CHC-Bench 10.
Table.10์€ CHC-Bench์˜ ์งˆ๋ฌธ ๊ตฌ์กฐ๋ฅผ ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ ํ‘œ๋Š” CHC-Bench 10์˜ ๊ตฌ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
Table.10์€ CHC-Bench์˜ ๋ฌธ์ œ ๊ตฌ์กฐ๋ฅผ ์ž์„ธํžˆ ์„ค๋ช…ํ•œ๋‹ค. ๋‹ค์Œ ํ‘œ๋Š” CHC-Bench 10์˜ ๊ตฌ์„ฑ์„ ๋ณด์—ฌ์ค€๋‹ค.
In some cases, the most useful thoughts appear to be near-continuations that correspond more closely to the target text, e.g.,
์ผ๋ถ€ ๊ฒฝ์šฐ์—๋Š” ๊ฐ€์žฅ ์œ ์šฉํ•œ ์ƒ๊ฐ๋“ค์ด ๋Œ€์ƒ ํ…์ŠคํŠธ์™€ ๋” ๋ฐ€์ ‘ํ•˜๊ฒŒ ์ผ์น˜ํ•˜๋Š” ๊ทผ์ ‘ํ•œ ์—ฐ์†์œผ๋กœ ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด,
๊ฒฝ์šฐ์— ๋”ฐ๋ผ ๊ฐ€์žฅ ์œ ์šฉํ•œ ์ƒ๊ฐ์€ ๋Œ€์ƒ ํ…์ŠคํŠธ์— ๋” ๊ฐ€๊น๊ฒŒ ๋ถ€ํ•ฉํ•˜๋Š” ๊ฑฐ์˜ ์—ฐ์†์ ์ธ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒฝ์šฐ๋„ ์žˆ๋‹ค.
Figure 10: Comparison of readability and diversity (ttr) of synthetic data compared with C4 and different subsets of the File.
Figure 10: ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€๋…์„ฑ๊ณผ ๋‹ค์–‘์„ฑ(ttr)์„ C4 ๋ฐ ํŒŒ์ผ์˜ ๋‹ค์–‘ํ•œ ํ•˜์œ„ ์ง‘ํ•ฉ๊ณผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ.
๊ทธ๋ฆผ 10: C4 ๋ฐ ํŒŒ์ผ์˜ ๋‹ค์–‘ํ•œ ํ•˜์œ„ ์ง‘ํ•ฉ๊ณผ ๋น„๊ตํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€๋…์„ฑ ๋ฐ ๋‹ค์–‘์„ฑ(ttr) ๋น„๊ต.
Figure 11: Data mixture of different training stages. Stable stage is shown in the left and decay stage is shown in the right.
Figure 11: ์„œ๋กœ ๋‹ค๋ฅธ ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ. ์•ˆ์ • ๋‹จ๊ณ„๋Š” ์™ผ์ชฝ์—, ๊ฐ์‡  ๋‹จ๊ณ„๋Š” ์˜ค๋ฅธ์ชฝ์— ํ‘œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 11: ๋‹ค์–‘ํ•œ ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ. ์•ˆ์ • ๋‹จ๊ณ„๋Š” ์™ผ์ชฝ์—, ๋ถ•๊ดด ๋‹จ๊ณ„๋Š” ์˜ค๋ฅธ์ชฝ์— ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค.
Let \(z\in\mathbb{R}^{\ell}\) be the (unknown) logits and let \(\mathbf{N}=\sum_{i}\exp(z_{i})\) be the normalizing constant.
\(z\in\mathbb{R}^{\ell}\)์„ (์•Œ๋ ค์ง€์ง€ ์•Š์€) ๋กœ์ง“์ด๋ผ๊ณ  ํ•˜๊ณ , \(\mathbf{N}=\sum_{i}\exp(z_{i})\)์„ ์ •๊ทœํ™” ์ƒ์ˆ˜๋ผ๊ณ  ํ•˜์ž.
z\in\mathbb{R}^{\ell}\)๋ฅผ (์•Œ ์ˆ˜ ์—†๋Š”) ๋กœ๊ทธ๋กœ ํ•˜๊ณ , \(\mathbf{N}=\sum_{i}\exp(z_{i})\)๋ฅผ ์ •๊ทœํ™” ์ƒ์ˆ˜๋กœ ํ•ฉ๋‹ˆ๋‹ค.
Figure 9: Distribution of \(\textsc{CnstScore}\) calculated for different styles of prompts \(\tau^{\prime}\) for LLaMA-2 7B.
Figure 9: LLaMA-2 7B์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ์˜ ํ”„๋กฌํ”„ํŠธ \(\tau^{\prime}\)์— ๋Œ€ํ•ด ๊ณ„์‚ฐ๋œ \(\textsc{CnstScore}\)์˜ ๋ถ„ํฌ.
๊ทธ๋ฆผ 9: LLaMA-2 7B์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ์˜ ํ”„๋กฌํ”„ํŠธ \(\tau^{\prime}\)์— ๋Œ€ํ•ด ๊ณ„์‚ฐ๋œ \(\textsc{CnstScore}\)์˜ ๋ถ„ํฌ.
\begin{tabular}{l} \hline \hline **Input** Synthesize a poem based on this prompt:"The sun is setting" (from stanford_alpaca)
\begin{tabular}{l} \hline \hline **์ž…๋ ฅ** ์ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์‹œ๋ฅผ ์ž‘์„ฑํ•˜์„ธ์š”: "ํ•ด๊ฐ€ ์ง€๊ณ  ์žˆ๋‹ค" (stanford_alpaca์—์„œ)
\begin{ํ‘œ}{l} \hline \hline**์ž…๋ ฅ** ์ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์‹œ๋ฅผ ํ•ฉ์„ฑํ•˜์„ธ์š”:"ํ•ด๊ฐ€ ์ง€๊ณ  ์žˆ๋‹ค"(์Šคํƒ ํฌ๋“œ ์•ŒํŒŒ์นด์—์„œ)
for \(i\in\{0,4,8,\cdots,|\mathcal{X}|\}\). This recovers the logits up to the free parameter \(z_{R}\) that we set to \(0\).
\(i\in\{0,4,8,\cdots,|\mathcal{X}|\}\)์— ๋Œ€ํ•ด. ์ด๋Š” ์šฐ๋ฆฌ๊ฐ€ \(0\)์œผ๋กœ ์„ค์ •ํ•œ ์ž์œ  ๋งค๊ฐœ๋ณ€์ˆ˜ \(z_{R}\)๊นŒ์ง€์˜ ๋กœ์ง“์„ ๋ณต์›ํ•ฉ๋‹ˆ๋‹ค.
for \(i\in\{0,4,8,\cdots,|\mathcal{X}|\}\). ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ๋กœ์ง“์„ ์ž์œ  ๋งค๊ฐœ๋ณ€์ˆ˜ \(z_{R}\)๊นŒ์ง€ ๋ณต๊ตฌํ•˜์—ฌ \(0\)์œผ๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.
guage models, ultimately enhancing their performance and applicability across various tasks within the Chinese NLP community.
์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ๊ณผ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ๊ถ๊ทน์ ์œผ๋กœ ํ–ฅ์ƒ์‹œ์ผœ ์ค‘๊ตญ NLP ์ปค๋ฎค๋‹ˆํ‹ฐ ๋‚ด์˜ ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๊ฑธ์ณ ์ ์šฉํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.
๊ฒŒ์ด์ง€ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ถ๊ทน์ ์œผ๋กœ ์ค‘๊ตญ NLP ์ปค๋ฎค๋‹ˆํ‹ฐ ๋‚ด ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ๊ณผ ์ ์šฉ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
To build a trustworthy benchmark, the grading system needs to be consistent, reproducible, interpretable and well-documented.
์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ํ‰๊ฐ€ ์‹œ์Šคํ…œ์ด ์ผ๊ด€์„ฑ ์žˆ๊ณ , ์žฌํ˜„ ๊ฐ€๋Šฅํ•˜๋ฉฐ, ํ•ด์„ ๊ฐ€๋Šฅํ•˜๊ณ , ์ž˜ ๋ฌธ์„œํ™”๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌ์ถ•ํ•˜๋ ค๋ฉด ์ฑ„์  ์‹œ์Šคํ…œ์ด ์ผ๊ด€์„ฑ ์žˆ๊ณ , ์žฌํ˜„ ๊ฐ€๋Šฅํ•˜๋ฉฐ, ํ•ด์„ ๊ฐ€๋Šฅํ•˜๊ณ , ์ž˜ ๋ฌธ์„œํ™”๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Figure 4(b) shows that the LLaMA architecture may lag behind GPT2's scaling law by a factor of 1.3x, even for larger models.
Figure 4(b)๋Š” ๋” ํฐ ๋ชจ๋ธ์—์„œ๋„ LLaMA ์•„ํ‚คํ…์ฒ˜๊ฐ€ GPT2์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋น„ํ•ด 1.3๋ฐฐ ์ •๋„ ๋’ค์ฒ˜์งˆ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 4(b)๋Š” LLaMA ์•„ํ‚คํ…์ฒ˜๊ฐ€ ๋” ํฐ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ์—๋„ GPT2์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— 1.3๋ฐฐ์˜ ์ฐจ์ด๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค€๋‹ค.
We develop a technique to compute the logit vector for any prefix \(p\) via a sequence of queries with varying logit biases.
์šฐ๋ฆฌ๋Š” ๋‹ค์–‘ํ•œ ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค๋ฅผ ๊ฐ€์ง„ ์ผ๋ จ์˜ ์ฟผ๋ฆฌ๋ฅผ ํ†ตํ•ด ์–ด๋–ค ์ ‘๋‘์‚ฌ \(p\)์— ๋Œ€ํ•œ ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ๊ณ„์‚ฐํ•˜๋Š” ๊ธฐ์ˆ ์„ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค.
๋‹ค์–‘ํ•œ ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค๋ฅผ ๊ฐ€์ง„ ์ผ๋ จ์˜ ์ฟผ๋ฆฌ๋ฅผ ํ†ตํ•ด ์ž„์˜์˜ ์ ‘๋‘์‚ฌ \(p\)์— ๋Œ€ํ•œ ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ๊ณ„์‚ฐํ•˜๋Š” ๊ธฐ๋ฒ•์„ ๊ฐœ๋ฐœํ•œ๋‹ค.
already deduplicated of exact matches during their creation, ExactSubstrdeduplication does not actually remove any examples.
์ด๋ฏธ ์ƒ์„ฑ ๊ณผ์ •์—์„œ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋Š” ํ•ญ๋ชฉ๋“ค์ด ์ค‘๋ณต ์ œ๊ฑฐ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์—, ExactSubstrdeduplication์€ ์‹ค์ œ๋กœ ์–ด๋–ค ์˜ˆ์ œ๋„ ์ œ๊ฑฐํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
์ด๋ฏธ ์ƒ์„ฑ ์ค‘์— ์ •ํ™•ํ•œ ์ผ์น˜ ํ•ญ๋ชฉ์ด ์ค‘๋ณต ์ œ๊ฑฐ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ExactSubstrdeduplication์€ ์‹ค์ œ๋กœ ์–ด๋–ค ์˜ˆ์ œ๋„ ์ œ๊ฑฐํ•˜์ง€ ์•Š๋Š”๋‹ค.
providing an answer, and it replies that the context does not provide exact information about the age which is also correct.
๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋ฉด์„œ, ๋ฌธ๋งฅ์ด ๋‚˜์ด์— ๋Œ€ํ•œ ์ •ํ™•ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜์ง€ ์•Š๋Š”๋‹ค๊ณ  ์‘๋‹ตํ•˜๋Š” ๊ฒƒ๋„ ๋งž์Šต๋‹ˆ๋‹ค.
๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๊ณ , ๋ฌธ๋งฅ์ƒ ์ •ํ™•ํ•œ ๋‚˜์ด์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜์ง€ ์•Š๋Š”๋‹ค๊ณ  ๋‹ต๋ณ€ํ•˜๋Š”๋ฐ, ์ด ์—ญ์‹œ ์ •๋‹ต์ž…๋‹ˆ๋‹ค.
- Trump might want to demonstrate leadership and confidence by being present at the first major event of the primary season.
- ํŠธ๋Ÿผํ”„๋Š” ์ฃผ์š” ์‹œ์ฆŒ์˜ ์ฒซ ๋ฒˆ์งธ ์ฃผ์š” ํ–‰์‚ฌ์— ์ฐธ์„ํ•จ์œผ๋กœ์จ ๋ฆฌ๋”์‹ญ๊ณผ ์ž์‹ ๊ฐ์„ ๋ณด์—ฌ์ฃผ๊ณ  ์‹ถ์–ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
- ํŠธ๋Ÿผํ”„๋Š” ์˜ˆ๋น„์„ ๊ฑฐ ์‹œ์ฆŒ์˜ ์ฒซ ์ฃผ์š” ํ–‰์‚ฌ์— ์ฐธ์„ํ•จ์œผ๋กœ์จ ๋ฆฌ๋”์‹ญ๊ณผ ์ž์‹ ๊ฐ์„ ๋ณด์—ฌ์ฃผ๊ณ  ์‹ถ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Footnote 11: This explanation is only for ease of understanding, and there is not direct evidence to connect the two points.
Footnote 11: ์ด ์„ค๋ช…์€ ์ดํ•ด๋ฅผ ๋•๊ธฐ ์œ„ํ•œ ๊ฒƒ์ด๋ฉฐ, ๋‘ ๊ฐ€์ง€ ์ ์„ ์—ฐ๊ฒฐํ•˜๋Š” ์ง์ ‘์ ์ธ ์ฆ๊ฑฐ๋Š” ์—†์Šต๋‹ˆ๋‹ค.
์ฃผ 11: ์ด๋Ÿฌํ•œ ์„ค๋ช…์€ ์ดํ•ด๋ฅผ ๋•๊ธฐ ์œ„ํ•œ ๊ฒƒ์ผ ๋ฟ, ๋‘ ์ ์„ ์—ฐ๊ฒฐํ•  ์ง์ ‘์ ์ธ ์ฆ๊ฑฐ๋Š” ์—†๋‹ค.
3. Iris's brother, Forest, has disappeared after enlisting to fight in the war. 4. Iris is devastated by her mother's death.
3. ์•„์ด๋ฆฌ์Šค์˜ ํ˜•์ œ์ธ ํฌ๋ ˆ์ŠคํŠธ๋Š” ์ „์Ÿ์— ์ฐธ์ „ํ•œ ํ›„ ์‹ค์ข…๋˜์—ˆ์Šต๋‹ˆ๋‹ค. 4. ์•„์ด๋ฆฌ์Šค๋Š” ์–ด๋จธ๋‹ˆ์˜ ์ฃฝ์Œ์— ํฐ ์ถฉ๊ฒฉ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค.
3. ์•„์ด๋ฆฌ์Šค์˜ ์˜ค๋น  ํฌ๋ ˆ์ŠคํŠธ๋Š” ์ „์Ÿ์— ์ฐธ์ „ํ•˜๊ธฐ ์œ„ํ•ด ์ž…๋Œ€ํ•œ ํ›„ ์‹ค์ข…๋˜์—ˆ์Šต๋‹ˆ๋‹ค. 4. ์•„์ด๋ฆฌ์Šค๋Š” ์–ด๋จธ๋‹ˆ์˜ ์ฃฝ์Œ์— ํฐ ์ถฉ๊ฒฉ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค.
Equal contribution, corresponding authors: {jupinderp,sprabhumoye,jjennings,mpatwary}@nvidia.com. Work done while at NVIDIA.
๋™์ผ ๊ธฐ์—ฌ, ๊ต์‹  ์ €์ž: {jupinderp, sprabhumoye, jjennings, mpatwary}@nvidia.com. NVIDIA์—์„œ ๊ทผ๋ฌดํ•˜๋Š” ๋™์•ˆ ์ˆ˜ํ–‰๋œ ์ž‘์—…์ž…๋‹ˆ๋‹ค.
๋™๋“ฑํ•œ ๊ธฐ์—ฌ, ํ•ด๋‹น ์ €์ž: {jupinderp,sprabhumoye,jjennings,mpatwary}@nvidia.com. NVIDIA์— ๊ทผ๋ฌดํ•˜๋Š” ๋™์•ˆ ์ˆ˜ํ–‰ํ•œ ์ž‘์—…์ž…๋‹ˆ๋‹ค.
In this paper, we aim to understand which approach and under what conditions is more effective for industry-specific models.
์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์‚ฐ์—…๋ณ„ ๋ชจ๋ธ์— ๋Œ€ํ•ด ์–ด๋–ค ์ ‘๊ทผ ๋ฐฉ์‹์ด ์–ด๋–ค ์กฐ๊ฑด์—์„œ ๋” ํšจ๊ณผ์ ์ธ์ง€ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์‚ฐ์—…๋ณ„ ๋ชจํ˜•์— ์žˆ์–ด ์–ด๋–ค ์ ‘๊ทผ๋ฒ•์ด ์–ด๋–ค ์กฐ๊ฑด์—์„œ ๋” ํšจ๊ณผ์ ์ธ์ง€ ํŒŒ์•…ํ•˜๊ณ ์ž ํ•œ๋‹ค.
Model-wise analysisWe report model-wise results on reasoning type and reasoning-claim relationship in Table 18 and Table 19.
๋ชจ๋ธ๋ณ„ ๋ถ„์„: ์šฐ๋ฆฌ๋Š” Table 18๊ณผ Table 19์—์„œ ์ถ”๋ก  ์œ ํ˜•๊ณผ ์ถ”๋ก -์ฃผ์žฅ ๊ด€๊ณ„์— ๋Œ€ํ•œ ๋ชจ๋ธ๋ณ„ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค.
๋ชจ๋ธ๋ณ„ ๋ถ„์„์ถ”๋ก  ์œ ํ˜•๊ณผ ์ถ”๋ก -์ฃผ์žฅ ๊ด€๊ณ„์— ๋Œ€ํ•œ ๋ชจ๋ธ๋ณ„ ๊ฒฐ๊ณผ๋Š” ํ‘œ 18๊ณผ ํ‘œ 19์— ๋ณด๊ณ ํ•œ๋‹ค.
3. Irisโ€™s brother, Forest, has disappeared after enlisting to fight in the war. 4. Iris is devastated by her motherโ€™s death.
3. ์•„์ด๋ฆฌ์Šค์˜ ํ˜•์ œ์ธ ํฌ๋ ˆ์ŠคํŠธ๋Š” ์ „์Ÿ์— ์ฐธ์ „ํ•œ ํ›„ ์‹ค์ข…๋˜์—ˆ์Šต๋‹ˆ๋‹ค. 4. ์•„์ด๋ฆฌ์Šค๋Š” ์–ด๋จธ๋‹ˆ์˜ ์ฃฝ์Œ์— ํฐ ์ถฉ๊ฒฉ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค.
3. ์•„์ด๋ฆฌ์Šค์˜ ์˜ค๋น  ํฌ๋ ˆ์ŠคํŠธ๋Š” ์ „์Ÿ์— ์ฐธ์ „ํ•˜๊ธฐ ์œ„ํ•ด ์ž…๋Œ€ํ•œ ํ›„ ์‹ค์ข…๋˜์—ˆ์Šต๋‹ˆ๋‹ค. 4. ์•„์ด๋ฆฌ์Šค๋Š” ์–ด๋จธ๋‹ˆ์˜ ์ฃฝ์Œ์— ํฐ ์ถฉ๊ฒฉ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค.
7. Now, we can see that \(h(-x)=h(x)\). This means that the function \(h(x)\) satisfies the definition of an even function.
์ด์ œ, \(h(-x)=h(x)\)์ž„์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํ•จ์ˆ˜ \(h(x)\)๊ฐ€ ์ง์ˆ˜ ํ•จ์ˆ˜์˜ ์ •์˜๋ฅผ ๋งŒ์กฑํ•จ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
7. ์ด์ œ \(h(-x)=h(x)\)์ž„์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํ•จ์ˆ˜ \(h(x)\)๊ฐ€ ์ง์ˆ˜ ํ•จ์ˆ˜์˜ ์ •์˜๋ฅผ ๋งŒ์กฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
Figure 2: Accuracy on GSM8K when training with Mistral-7B and the dataset of MAMnoTH. MFT shows a higher sample efficiency.
Figure 2: Mistral-7B์™€ MAMnoTH ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ํ›ˆ๋ จํ•  ๋•Œ GSM8K์— ๋Œ€ํ•œ ์ •ํ™•๋„. MFT๋Š” ๋” ๋†’์€ ์ƒ˜ํ”Œ ํšจ์œจ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2: Mistral-7B์™€ MAMnoTH์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ํ›ˆ๋ จํ•  ๋•Œ GSM8K์˜ ์ •ํ™•๋„. MFT๋Š” ๋” ๋†’์€ ์ƒ˜ํ”Œ ํšจ์œจ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
- The likelihood of Israel publicly acknowledging such an operation is lower, as they often maintain plausible deniability.
- ์ด์Šค๋ผ์—˜์ด ๊ทธ๋Ÿฌํ•œ ์ž‘์ „์„ ๊ณต๊ฐœ์ ์œผ๋กœ ์ธ์ •ํ•  ๊ฐ€๋Šฅ์„ฑ์€ ๋‚ฎ์œผ๋ฉฐ, ์ด๋Š” ๊ทธ๋“ค์ด ์ข…์ข… ๊ทธ๋Ÿด๋“ฏํ•œ ๋ถ€์ธ ๊ฐ€๋Šฅ์„ฑ์„ ์œ ์ง€ํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
- ์ด์Šค๋ผ์—˜์€ ์ข…์ข… ๊ทธ๋Ÿด๋“ฏํ•œ ๋ถ€์ธ ๊ฐ€๋Šฅ์„ฑ์„ ์œ ์ง€ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด๋Ÿฌํ•œ ์ž‘์ „์„ ๊ณต๊ฐœ์ ์œผ๋กœ ์ธ์ •ํ•  ๊ฐ€๋Šฅ์„ฑ์€ ๋‚ฎ์Šต๋‹ˆ๋‹ค.
To find the total number of trees Maria has in her orchard, we add the number of apple trees to the number of cherry trees:
Maria๊ฐ€ ์ž์‹ ์˜ ๊ณผ์ˆ˜์›์— ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๋‚˜๋ฌด์˜ ์ด ์ˆ˜๋ฅผ ์ฐพ๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์‚ฌ๊ณผ๋‚˜๋ฌด์˜ ์ˆ˜์— ์ฒด๋ฆฌ๋‚˜๋ฌด์˜ ์ˆ˜๋ฅผ ๋”ํ•ฉ๋‹ˆ๋‹ค.
๋งˆ๋ฆฌ์•„๊ฐ€ ๊ณผ์ˆ˜์›์— ์žˆ๋Š” ์ด ๋‚˜๋ฌด ์ˆ˜๋ฅผ ๊ตฌํ•˜๋ ค๋ฉด ์‚ฌ๊ณผ๋‚˜๋ฌด ์ˆ˜์— ์ฒด๋ฆฌ๋‚˜๋ฌด ์ˆ˜๋ฅผ ๋”ํ•ฉ๋‹ˆ๋‹ค:
_Here, \(F\) is parameterized by \(W(Q)\) and may rely on previous data \(Q_{1},\ldots,Q_{i-1}\), and new randomness \(R\).
์—ฌ๊ธฐ์„œ \(F\)๋Š” \(W(Q)\)๋กœ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋˜๋ฉฐ, ์ด์ „ ๋ฐ์ดํ„ฐ \(Q_{1},\ldots,Q_{i-1}\)์™€ ์ƒˆ๋กœ์šด ๋ฌด์ž‘์œ„์„ฑ \(R\)์— ์˜์กดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ, F๋Š” W(Q)์— ์˜ํ•ด ํŒŒ๋ผ๋ฏธํ„ฐํ™”๋˜๊ณ  ์ด์ „ ๋ฐ์ดํ„ฐ \(Q_{1},\ldots,Q_{i-1}\), ๋ฐ ์ƒˆ๋กœ์šด ๋‚œ์ˆ˜ \(R\)์— ์˜์กดํ•  ์ˆ˜ ์žˆ๋‹ค.
Jurning list in allergies by participants towards Mrs. Price, her struggle with epilepsy, and the death of her friend days.
์ฐธ๊ฐ€์ž๋“ค์ด Mrs. Price์— ๋Œ€ํ•œ ์•Œ๋ ˆ๋ฅด๊ธฐ ๋ชฉ๋ก, ๊ทธ๋…€์˜ ๊ฐ„์งˆ๊ณผ์˜ ์‹ธ์›€, ๊ทธ๋ฆฌ๊ณ  ์นœ๊ตฌ์˜ ์ฃฝ์Œ์— ๋Œ€ํ•œ ์ด์•ผ๊ธฐ.
์•Œ๋ ˆ๋ฅด๊ธฐ ์ฐธ๊ฐ€์ž๋“ค์˜ ๊ฐ€๊ฒฉ ๋ถ€์ธ์— ๋Œ€ํ•œ ๋ชฉ๋ก, ๊ฐ„์งˆ๊ณผ ์‹ธ์šฐ๋Š” ๊ทธ๋…€์˜ ํˆฌ๋ณ‘ ์ƒํ™œ, ๊ทธ๋ฆฌ๊ณ  ๊ทธ๋…€์˜ ์นœ๊ตฌ์˜ ์‚ฌ๋ง ๋‚ ์„ ๊ธฐ๋กํ•ฉ๋‹ˆ๋‹ค.
In this case it is easy to guess that the third option is the most likely, even without knowledge of this specific strike.
์ด ๊ฒฝ์šฐ, ํŠน์ • ํŒŒ์—…์— ๋Œ€ํ•œ ์ง€์‹์ด ์—†์–ด๋„ ์„ธ ๋ฒˆ์งธ ์˜ต์…˜์ด ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค๋Š” ๊ฒƒ์„ ์‰ฝ๊ฒŒ ์ถ”์ธกํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด ๊ฒฝ์šฐ ์ด ํŠน์ • ํŒŒ์—…์— ๋Œ€ํ•œ ์ง€์‹์ด ์—†์–ด๋„ ์„ธ ๋ฒˆ์งธ ์˜ต์…˜์ด ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค๋Š” ๊ฒƒ์„ ์‰ฝ๊ฒŒ ์ถ”์ธกํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Figure A.6 \(|\) PCP example consultation.** Example of a patient actor's conversation with a PCP during the OSCE study.
**Figure A.6 \(|\) PCP ์˜ˆ์‹œ ์ƒ๋‹ด.** OSCE ์—ฐ๊ตฌ ์ค‘ ํ™˜์ž ๋ฐฐ์šฐ์ž์™€ PCP ๊ฐ„์˜ ๋Œ€ํ™” ์˜ˆ์‹œ.
**๊ทธ๋ฆผ A.6 \*| PCP ์˜ˆ์‹œ ์ƒ๋‹ด.** OSCE ์—ฐ๊ตฌ ์ค‘ ํ™˜์ž ๋ฐฐ์šฐ์ž๊ฐ€ PCP์™€ ๋Œ€ํ™”ํ•˜๋Š” ์˜ˆ์‹œ์ž…๋‹ˆ๋‹ค.
Footnote 7: Note the statement โ€œ\(A\) is \(B\)โ€ does not appears in prompt \(p\) but \(B\) can appear in \(p\) on its own.
Footnote 7: "A๋Š” B์ด๋‹ค"๋ผ๋Š” ๋ฌธ์žฅ์€ ํ”„๋กฌํ”„ํŠธ \(p\)์— ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์ง€๋งŒ, B๋Š” \(p\)์— ๋…๋ฆฝ์ ์œผ๋กœ ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฐ์ฃผ 7: "A\)๋Š” \(B\)"๋ผ๋Š” ๋ฌธ๊ตฌ๋Š” ํ”„๋กฌํ”„ํŠธ \(p\)์— ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์ง€๋งŒ \(B\)๋Š” ๊ทธ ์ž์ฒด๋กœ \(p\)์— ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
I'm using this php code to fetch the xml data but this code fetching from whole xml data which makes duplicate field table
์ด PHP ์ฝ”๋“œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ XML ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐ€์ ธ์˜ค๊ณ  ์žˆ์ง€๋งŒ, ์ด ์ฝ”๋“œ๋Š” ์ „์ฒด XML ๋ฐ์ดํ„ฐ์—์„œ ์ค‘๋ณต๋œ ํ•„๋“œ ํ…Œ์ด๋ธ”์„ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค.
์ด php ์ฝ”๋“œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ xml ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐ€์ ธ์˜ค๊ณ  ์žˆ์ง€๋งŒ ์ด ์ฝ”๋“œ๋Š” ์ „์ฒด xml ๋ฐ์ดํ„ฐ์—์„œ ๊ฐ€์ ธ์˜ค๊ธฐ ๋•Œ๋ฌธ์— ํ•„๋“œ ํ…Œ์ด๋ธ”์ด ์ค‘๋ณต๋ฉ๋‹ˆ๋‹ค.
2. Teleological ethical theories - which hold that the morality of an action depends on the outcome or result it produces.
2. ๋ชฉ์ ๋ก ์  ์œค๋ฆฌ ์ด๋ก  - ํ–‰๋™์˜ ๋„๋•์„ฑ์€ ๊ทธ ํ–‰๋™์ด ์ดˆ๋ž˜ํ•˜๋Š” ๊ฒฐ๊ณผ๋‚˜ ์„ฑ๊ณผ์— ๋”ฐ๋ผ ๊ฒฐ์ •๋œ๋‹ค๊ณ  ๋ณด๋Š” ์ด๋ก .
2. ๋ชฉ์ ๋ก ์  ์œค๋ฆฌ ์ด๋ก  - ์–ด๋–ค ํ–‰๋™์˜ ๋„๋•์„ฑ์€ ๊ทธ ํ–‰๋™์ด ๊ฐ€์ ธ์˜ค๋Š” ๊ฒฐ๊ณผ๋‚˜ ๊ฒฐ๊ณผ์— ๋‹ฌ๋ ค ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค.
Here, we see that \(2^{13}<10000<2^{14}\), which means that \(\log_{2}10000\) is greater than \(13\) but less than \(14\).
์—ฌ๊ธฐ์„œ ์šฐ๋ฆฌ๋Š” \(2^{13}<10000<2^{14}\)์ž„์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” \(\log_{2}10000\)์ด \(13\)๋ณด๋‹ค ํฌ๊ณ  \(14\)๋ณด๋‹ค ์ž‘๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(2^{13}<10000<2^{14}\), ์ฆ‰ \(\log_{2}10000\)์€ \(13\)๋ณด๋‹ค ํฌ์ง€๋งŒ \(14\)๋ณด๋‹ค ์ž‘๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
When comparing with models from the state-of-the-art, we source results from a few different papers, detailed in Table 10.
์ตœ์‹  ๊ธฐ์ˆ ์˜ ๋ชจ๋ธ๋“ค๊ณผ ๋น„๊ตํ•  ๋•Œ, ์šฐ๋ฆฌ๋Š” Table 10์— ์ž์„ธํžˆ ์„ค๋ช…๋œ ๋ช‡ ๊ฐ€์ง€ ๋‹ค๋ฅธ ๋…ผ๋ฌธ์—์„œ ๊ฒฐ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค.
์ตœ์‹  ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•  ๋•Œ, ์šฐ๋ฆฌ๋Š” ํ‘œ 10์— ์ž์„ธํžˆ ์„ค๋ช…๋œ ๋ช‡ ๊ฐ€์ง€ ๋‹ค๋ฅธ ๋…ผ๋ฌธ์˜ ๊ฒฐ๊ณผ๋ฅผ ์†Œ์‹ฑํ•œ๋‹ค.
Subtract an equation for eggs eaten and an equation for eggs baked to find the eggs remaining. Let's see what we can find.
๊ณ„๋ž€์„ ๋จน์€ ์–‘๊ณผ ๊ตฌ์šด ์–‘์— ๋Œ€ํ•œ ๋ฐฉ์ •์‹์„ ๋นผ์„œ ๋‚จ์€ ๊ณ„๋ž€์˜ ์ˆ˜๋ฅผ ๊ตฌํ•˜์„ธ์š”. ๋ฌด์—‡์„ ์ฐพ์„ ์ˆ˜ ์žˆ๋Š”์ง€ ๋ด…์‹œ๋‹ค.
๋‹ฌ๊ฑ€์„ ๋จน์€ ๋ฐฉ์ •์‹๊ณผ ๊ตฌ์šด ๋‹ฌ๊ฑ€์„ ๋บ€ ๋ฐฉ์ •์‹์„ ๋นผ์„œ ๋‚จ์€ ๋‹ฌ๊ฑ€์„ ๊ตฌํ•ด ๋ด…์‹œ๋‹ค. ๋ฌด์—‡์„ ์ฐพ์„ ์ˆ˜ ์žˆ๋Š”์ง€ ์‚ดํŽด๋ด…์‹œ๋‹ค.
Causes for Factual ErrorsThere are many possible reasons for the failure of models to answer factual questions accurately.
์‚ฌ์‹ค์  ์˜ค๋ฅ˜์˜ ์›์ธ๋ชจ๋ธ์ด ์‚ฌ์‹ค์  ์งˆ๋ฌธ์— ์ •ํ™•ํ•˜๊ฒŒ ๋‹ตํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฐ๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ์ด์œ ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
์‚ฌ์‹ค ์˜ค๋ฅ˜์˜ ์›์ธ๋ชจ๋ธ์ด ์‚ฌ์‹ค์— ์ž…๊ฐํ•œ ์งˆ๋ฌธ์— ์ •ํ™•ํ•˜๊ฒŒ ๋‹ตํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฐ์—๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์ด์œ ๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In data synthesis, we utilize the GPT-4 Turbo API, setting the temperature to 1.0 for both question and answer generation.
๋ฐ์ดํ„ฐ ํ•ฉ์„ฑ์—์„œ๋Š” ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€ ์ƒ์„ฑ์„ ์œ„ํ•ด GPT-4 Turbo API๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ์˜จ๋„๋ฅผ 1.0์œผ๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ ํ•ฉ์„ฑ์—์„œ๋Š” ์งˆ๋ฌธ ๋ฐ ๋‹ต๋ณ€ ์ƒ์„ฑ ๋ชจ๋‘์— ์˜จ๋„๋ฅผ 1.0์œผ๋กœ ์„ค์ •ํ•œ GPT-4 ํ„ฐ๋ณด API๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค.
The following are two example consultations for the same scenario pack and patient actor from AMIE and PCP, respectively.
๋‹ค์Œ์€ ๋™์ผํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ๊ณผ ํ™˜์ž ๋ฐฐ์šฐ์— ๋Œ€ํ•ด ๊ฐ๊ฐ AMIE์™€ PCP์—์„œ์˜ ๋‘ ๊ฐ€์ง€ ์˜ˆ์‹œ ์ƒ๋‹ด์ž…๋‹ˆ๋‹ค.
๋‹ค์Œ์€ AMIE์™€ PCP์—์„œ ๊ฐ๊ฐ ๋™์ผํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ๊ณผ ํ™˜์ž ์•กํ„ฐ์— ๋Œ€ํ•œ ๋‘ ๊ฐ€์ง€ ์˜ˆ์‹œ ์ƒ๋‹ด์ž…๋‹ˆ๋‹ค.
To evaluate models, we average zero-shot performance over diverse task aggregates Our aggregates are outlined in Table 3:
๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—… ์ง‘ํ•ฉ์— ๋Œ€ํ•œ zero-shot ์„ฑ๋Šฅ์„ ํ‰๊ท ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ง‘ํ•ฉ์€ Table 3์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค:
๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ์ž‘์—… ์ง‘๊ณ„์— ๋Œ€ํ•œ ์ œ๋กœ์ƒท ์„ฑ๋Šฅ์„ ํ‰๊ท ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ง‘๊ณ„๋Š” ํ‘œ 3์— ์š”์•ฝ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค:
For a deterministic ground-truth labelling function \(i\mapsto j_{i}\), consider a downstream Hutter "LLM" (Hutter, 2021)
๊ฒฐ์ •๋ก ์  ์ฐธ๊ฐ’ ๋ผ๋ฒจ๋ง ํ•จ์ˆ˜ \(i\mapsto j_{i}\)์— ๋Œ€ํ•ด, ํ•˜์œ„ ์ž‘์—… Hutter "LLM" (Hutter, 2021)์„ ๊ณ ๋ คํ•˜์‹ญ์‹œ์˜ค.
๊ฒฐ์ •๋ก ์  ์ง€์ƒ ์‹ค์ธก ๋ผ๋ฒจ๋ง ํ•จ์ˆ˜ \(i\mapsto j_{i}\), ๋‹ค์šด์ŠคํŠธ๋ฆผ Hutter "LLM"(Hutter, 2021)์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค.
**Is it possible to identify the individuals directly or indirectly from the dataset?** No. No individuals are mentioned.
**๋ฐ์ดํ„ฐ์…‹์—์„œ ๊ฐœ์ธ์„ ์ง์ ‘ ๋˜๋Š” ๊ฐ„์ ‘์ ์œผ๋กœ ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๊นŒ?** ์•„๋‹ˆ์š”. ์–ธ๊ธ‰๋œ ๊ฐœ์ธ์€ ์—†์Šต๋‹ˆ๋‹ค.
**๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์—์„œ ๊ฐœ์ธ์„ ์ง๊ฐ„์ ‘์ ์œผ๋กœ ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ๋‚˜์š”?** ์•„๋‹ˆ์š”. ๊ฐœ์ธ์— ๋Œ€ํ•œ ์–ธ๊ธ‰์ด ์—†์Šต๋‹ˆ๋‹ค.
14. Forest was believed to be fighting for deity Enva, but is revealed to have been forcefully conscribed by deity Dacre.
์ˆฒ์€ ์‹  ์—”๋ฐ”๋ฅผ ์œ„ํ•ด ์‹ธ์šฐ๊ณ  ์žˆ๋‹ค๊ณ  ์—ฌ๊ฒจ์กŒ์œผ๋‚˜, ์‹ค์ œ๋กœ๋Š” ์‹  ๋ฐ์ด์ปค์— ์˜ํ•ด ๊ฐ•์ œ๋กœ ์ง•์ง‘๋œ ๊ฒƒ์œผ๋กœ ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค.
14. ํฌ๋ ˆ์ŠคํŠธ๋Š” ์—”๋ฐ” ์‹ ์„ ์œ„ํ•ด ์‹ธ์šฐ๊ณ  ์žˆ๋‹ค๊ณ  ๋ฏฟ์—ˆ์ง€๋งŒ, ๋‹คํฌ๋ ˆ ์‹ ์— ์˜ํ•ด ๊ฐ•์ œ๋กœ ์ง•์ง‘๋œ ๊ฒƒ์œผ๋กœ ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค.