question
stringlengths
101
2.69k
chosen
stringlengths
43
2.27k
rejected
stringlengths
42
2.18k
We further investigate the in-context learning capability, i.e., the capability of inferring the underlying function given the few-show input-output demonstrations. We consider the task of inferring the linear coefficient of a weighted sum. Specifically, define \(y=w_{1}x_{1}+w2x_{2}+...+w_{n}x_{n}\), our few-shot demonstration is \(x_{1},x_{2},...,x_{n},y\), and we ask the model to (implicitly) infer \(w_{1},w_{2},...,w_{n}\) by predicting the \(y\) given a new set of input \(x\). We use (a). the absolute difference between model prediction \(y\) and the ground truth \(y^{*}\), i.e., \(|y-y^{*}|\) as a continuous measure, and use (b). the exact match \(y==y^{*}\) as a discontinuous measure. We further note that most of the models perform reasonably well on addition and subtraction, so the ability to do arithmetic, as a confounding factor, can be ruled out.
์šฐ๋ฆฌ๋Š” ์ถ”๊ฐ€์ ์œผ๋กœ in-context ํ•™์Šต ๋Šฅ๋ ฅ, ์ฆ‰ ๋ช‡ ๊ฐ€์ง€ ์ž…๋ ฅ-์ถœ๋ ฅ ์‹œ์—ฐ์„ ํ†ตํ•ด ๊ธฐ๋ณธ์ ์ธ ํ•จ์ˆ˜๋ฅผ ์ถ”๋ก ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๊ฐ€์ค‘ ํ•ฉ์˜ ์„ ํ˜• ๊ณ„์ˆ˜๋ฅผ ์ถ”๋ก ํ•˜๋Š” ์ž‘์—…์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, \(y=w_{1}x_{1}+w_{2}x_{2}+...+w_{n}x_{n}\)์„ ์ •์˜ํ•˜๊ณ , ์šฐ๋ฆฌ์˜ ๋ช‡ ๊ฐ€์ง€ ์‹œ์—ฐ์€ \(x_{1},x_{2},...,x_{n},y\)์ด๋ฉฐ, ์ƒˆ๋กœ์šด ์ž…๋ ฅ \(x\)๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ \(y\)๋ฅผ ์˜ˆ์ธกํ•˜์—ฌ \(w_{1},w_{2},...,w_{n}\)์„ (์•”๋ฌต์ ์œผ๋กœ) ์ถ”๋ก ํ•˜๋„๋ก ๋ชจ๋ธ์— ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” (a) ๋ชจ๋ธ ์˜ˆ์ธก \(y\)์™€ ์‹ค์ œ ๊ฐ’ \(y^{*}\) ์‚ฌ์ด์˜ ์ ˆ๋Œ€ ์ฐจ์ด, ์ฆ‰ \(|y-y^{*}|\)๋ฅผ ์—ฐ์†์ ์ธ ์ธก์ •์œผ๋กœ ์‚ฌ์šฉํ•˜๊ณ , (b) ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋Š” \(y==y^{*}\)๋ฅผ ๋ถˆ์—ฐ์†์ ์ธ ์ธก์ •์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋Œ€๋ถ€๋ถ„์˜ ๋ชจ๋ธ์ด ๋ง์…ˆ๊ณผ ๋บ„์…ˆ์—์„œ ํ•ฉ๋ฆฌ์ ์œผ๋กœ ์ž˜ ์ˆ˜ํ–‰ํ•˜๋ฏ€๋กœ, ํ˜ผ๋ž€ ์š”์ธ์œผ๋กœ์„œ์˜ ์‚ฐ์ˆ  ๋Šฅ๋ ฅ์€ ๋ฐฐ์ œํ•  ์ˆ˜ ์žˆ์Œ์„ ์ถ”๊ฐ€๋กœ ์–ธ๊ธ‰ํ•ฉ๋‹ˆ๋‹ค.
์ปจํ…์ŠคํŠธ ๋‚ด ํ•™์Šต ๊ธฐ๋Šฅ, ์ฆ‰ ๋ช‡ ๊ฐ€์ง€ ์ž…๋ ฅ-์ถœ๋ ฅ ๋ฐ๋ชจ๊ฐ€ ํ‘œ์‹œ๋œ ๊ฒฝ์šฐ ๊ธฐ๋ณธ ํ•จ์ˆ˜๋ฅผ ์ถ”๋ก ํ•˜๋Š” ๊ธฐ๋Šฅ์„ ๋” ์ž์„ธํžˆ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์ค‘์น˜ ํ•ฉ๊ณ„์˜ ์„ ํ˜• ๊ณ„์ˆ˜๋ฅผ ์ถ”๋ก ํ•˜๋Š” ์ž‘์—…์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, \(y=w_{1}x_{1}+w2x_{2}+...+w_{n}x_{n}\), ๋ช‡ ๋ฒˆ์˜ ์‡ผ ๋ฐ๋ชจ๋Š” \(x_{1},x_{2},...,x_{n},y\)๋กœ ์ •์˜ํ•˜๊ณ , ๋ชจ๋ธ์ด ์ƒˆ๋กœ์šด ์ž…๋ ฅ ์ง‘ํ•ฉ \(x\)๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ \(y\)๋ฅผ ์˜ˆ์ธกํ•˜์—ฌ (์•”์‹œ์ ์œผ๋กœ) \(w_{1},w_{2},...,w_{n}\)๋ฅผ ์ถ”๋ก ํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. (a) ๋ชจ๋ธ ์˜ˆ์ธก \(y\)์™€ ์‹ค์ธก๊ฐ’ \(y^{*}\) ์‚ฌ์ด์˜ ์ ˆ๋Œ€ ์ฐจ์ด, ์ฆ‰ \(|y-y^{*}|\)๋ฅผ ์—ฐ์† ์ธก์ •๊ฐ’์œผ๋กœ ์‚ฌ์šฉํ•˜๊ณ , (b) ์ •ํ™•ํ•œ ์ผ์น˜ \(y==y^{*}\)๋ฅผ ๋ถˆ์—ฐ์† ์ธก์ •๊ฐ’์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋Œ€๋ถ€๋ถ„์˜ ๋ชจ๋ธ์ด ๋ง์…ˆ๊ณผ ๋บ„์…ˆ์—์„œ ํ•ฉ๋ฆฌ์ ์œผ๋กœ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋ฏ€๋กœ ์‚ฐ์ˆ ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋Šฅ๋ ฅ์€ ํ˜ผ๋ž€ ์š”์ธ์œผ๋กœ ๋ฐฐ์ œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
We note, however, that in contrast to our core evaluation set described in Section 4.1, we found these additional end-tasks to have less stable performance during model development, and to provide a limited signal. This is illustrated in Figure 4, where we see the progress of task performance throughout training to be more random (compare with the more stable upward trends in Figure 1). While tasks such as mrpc and wic appear more stable, they offered additional difficulties related to performance being tied to random chance (e.g., wic) or the tendency of models to make spurious predictions (e.g., always predicting a single label) that either inflate or deflate performance due to dataset class imbalances (e.g., mrpc). We therefore caution against relying too heavily on these tasks when measuring model performance throughout training and comparing models.
๊ทธ๋Ÿฌ๋‚˜ 4.1์ ˆ์—์„œ ์„ค๋ช…ํ•œ ํ•ต์‹ฌ ํ‰๊ฐ€ ์„ธํŠธ์™€๋Š” ๋‹ฌ๋ฆฌ, ์ด๋Ÿฌํ•œ ์ถ”๊ฐ€์ ์ธ ์ตœ์ข… ์ž‘์—…๋“ค์€ ๋ชจ๋ธ ๊ฐœ๋ฐœ ์ค‘์— ์„ฑ๋Šฅ์ด ๋œ ์•ˆ์ •์ ์ด๋ฉฐ ์ œํ•œ๋œ ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” Figure 4์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ํ›ˆ๋ จ ๋‚ด๋‚ด ์ž‘์—… ์„ฑ๋Šฅ์˜ ์ง„ํ–‰์ด ๋” ๋ฌด์ž‘์œ„์ ์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค (Figure 1์˜ ๋” ์•ˆ์ •์ ์ธ ์ƒ์Šน ์ถ”์„ธ์™€ ๋น„๊ตํ•ด๋ณด์„ธ์š”). mrpc์™€ wic์™€ ๊ฐ™์€ ์ž‘์—…์€ ๋” ์•ˆ์ •์ ์œผ๋กœ ๋ณด์ด์ง€๋งŒ, ์„ฑ๋Šฅ์ด ๋ฌด์ž‘์œ„ ๊ธฐํšŒ์— ๋ฌถ์—ฌ ์žˆ๊ฑฐ๋‚˜ (์˜ˆ: wic) ๋ชจ๋ธ์ด ์ž˜๋ชป๋œ ์˜ˆ์ธก์„ ํ•˜๋Š” ๊ฒฝํ–ฅ (์˜ˆ: ํ•ญ์ƒ ๋‹จ์ผ ๋ ˆ์ด๋ธ”์„ ์˜ˆ์ธก)์œผ๋กœ ์ธํ•ด ๋ฐ์ดํ„ฐ์…‹ ํด๋ž˜์Šค ๋ถˆ๊ท ํ˜•์œผ๋กœ ์„ฑ๋Šฅ์ด ๋ถ€ํ’€๋ ค์ง€๊ฑฐ๋‚˜ ์ถ•์†Œ๋˜๋Š” ์ถ”๊ฐ€์ ์ธ ์–ด๋ ค์›€์„ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค (์˜ˆ: mrpc). ๋”ฐ๋ผ์„œ ํ›ˆ๋ จ ๋‚ด๋‚ด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ์ธก์ •ํ•˜๊ณ  ๋ชจ๋ธ์„ ๋น„๊ตํ•  ๋•Œ ์ด๋Ÿฌํ•œ ์ž‘์—…์— ์ง€๋‚˜์น˜๊ฒŒ ์˜์กดํ•˜์ง€ ์•Š๋„๋ก ์ฃผ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ ์„น์…˜ 4.1์— ์„ค๋ช…๋œ ํ•ต์‹ฌ ํ‰๊ฐ€ ์„ธํŠธ์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์ด๋Ÿฌํ•œ ์ถ”๊ฐ€ ์ตœ์ข… ์ž‘์—…์€ ๋ชจ๋ธ ๊ฐœ๋ฐœ ์ค‘์— ์„ฑ๋Šฅ์ด ๋œ ์•ˆ์ •์ ์ด๊ณ  ์ œํ•œ๋œ ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ทธ๋ฆผ 4์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋Š”๋ฐ, ๊ทธ๋ฆผ 4์—์„œ๋Š” ํ›ˆ๋ จ ์ „๋ฐ˜์— ๊ฑธ์นœ ์ž‘์—… ์„ฑ๋Šฅ์˜ ์ง„ํ–‰์ด ๋” ๋ฌด์ž‘์œ„์ ์ž„์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1์˜ ๋ณด๋‹ค ์•ˆ์ •์ ์ธ ์ƒ์Šน ์ถ”์„ธ์™€ ๋น„๊ต). mrpc ๋ฐ wic์™€ ๊ฐ™์€ ์ž‘์—…์€ ๋” ์•ˆ์ •์ ์œผ๋กœ ๋ณด์ด์ง€๋งŒ, ์„ฑ๋Šฅ์ด ๋ฌด์ž‘์œ„ ํ™•๋ฅ ์— ์ขŒ์šฐ๋˜๊ฑฐ๋‚˜(์˜ˆ: wic) ๋ฐ์ดํ„ฐ ์„ธํŠธ ํด๋ž˜์Šค ๋ถˆ๊ท ํ˜•์œผ๋กœ ์ธํ•ด ์„ฑ๋Šฅ์„ ๋ถ€ํ’€๋ฆฌ๊ฑฐ๋‚˜ ์ถ•์†Œํ•˜๋Š” ๋ชจ๋ธ์˜ ์Šคํ“จ๋ฆฌ์–ด์Šค ์˜ˆ์ธก ๊ฒฝํ–ฅ(์˜ˆ: mrpc)๊ณผ ๊ด€๋ จ๋œ ์ถ”๊ฐ€์ ์ธ ์–ด๋ ค์›€์„ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ›ˆ๋ จ ์ค‘ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ์ธก์ •ํ•˜๊ณ  ๋ชจ๋ธ์„ ๋น„๊ตํ•  ๋•Œ ์ด๋Ÿฌํ•œ ์ž‘์—…์— ์ง€๋‚˜์น˜๊ฒŒ ์˜์กดํ•˜์ง€ ์•Š๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.
UpdatedLAMA and NewLAMAWe construct UpdatedLAMA and NewLAMA for measuring the update of outdated knowledge and acquisition of new knowledge during CKL. The challenge of constructing UpdatedLAMA is that a knowledge instance can be only considered as the knowledge that requires update only if it is present in both \(D_{0}\) and \(D_{1}\) with changed details, and the challenge of constructing NewLAMA is that the knowledge can be considered new only if it is in \(D_{1}\) but not in \(D_{0}\). Therefore we set up the data construction process carefully. The pipeline for the creation of a single instance of UpdatedLAMA and NewLAMA, is shown in Figure 4 (a). Each potential instance starts off from a single article from CC-RecentNews and goes through the pipeline which will end up being (1) discarded (2) added to UpdatedLAMA or (3) added to NewLAMA in the end.
UpdatedLAMA์™€ NewLAMA๋ฅผ ๊ตฌ์ถ•ํ•˜์—ฌ CKL ๋™์•ˆ ์˜ค๋ž˜๋œ ์ง€์‹์˜ ์—…๋ฐ์ดํŠธ์™€ ์ƒˆ๋กœ์šด ์ง€์‹์˜ ํš๋“์„ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. UpdatedLAMA๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ ์žˆ์–ด ๋„์ „ ๊ณผ์ œ๋Š” ์ง€์‹ ์ธ์Šคํ„ด์Šค๊ฐ€ \(D_{0}\)์™€ \(D_{1}\) ๋ชจ๋‘์— ์กด์žฌํ•˜๋ฉด์„œ ์„ธ๋ถ€ ์‚ฌํ•ญ์ด ๋ณ€๊ฒฝ๋œ ๊ฒฝ์šฐ์—๋งŒ ์—…๋ฐ์ดํŠธ๊ฐ€ ํ•„์š”ํ•œ ์ง€์‹์œผ๋กœ ๊ฐ„์ฃผ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์ด๋ฉฐ, NewLAMA๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ ์žˆ์–ด ๋„์ „ ๊ณผ์ œ๋Š” ์ง€์‹์ด \(D_{1}\)์—๋งŒ ์กด์žฌํ•˜๊ณ  \(D_{0}\)์—๋Š” ์กด์žฌํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ์—๋งŒ ์ƒˆ๋กœ์šด ์ง€์‹์œผ๋กœ ๊ฐ„์ฃผ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ฐ์ดํ„ฐ ๊ตฌ์ถ• ๊ณผ์ •์„ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. UpdatedLAMA์™€ NewLAMA์˜ ๋‹จ์ผ ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ์„ฑํ•˜๋Š” ํŒŒ์ดํ”„๋ผ์ธ์€ Figure 4 (a)์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์ž ์žฌ์  ์ธ์Šคํ„ด์Šค๋Š” CC-RecentNews์˜ ๋‹จ์ผ ๊ธฐ์‚ฌ์—์„œ ์‹œ์ž‘ํ•˜์—ฌ ํŒŒ์ดํ”„๋ผ์ธ์„ ๊ฑฐ์ณ ์ตœ์ข…์ ์œผ๋กœ (1) ํ๊ธฐ๋˜๊ฑฐ๋‚˜ (2) UpdatedLAMA์— ์ถ”๊ฐ€๋˜๊ฑฐ๋‚˜ (3) NewLAMA์— ์ถ”๊ฐ€๋ฉ๋‹ˆ๋‹ค.
์—…๋ฐ์ดํŠธ๋œLAMA ๋ฐ NewLAMACKL ์ค‘์— ์˜ค๋ž˜๋œ ์ง€์‹์˜ ์—…๋ฐ์ดํŠธ์™€ ์ƒˆ๋กœ์šด ์ง€์‹์˜ ์Šต๋“์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์—…๋ฐ์ดํŠธ๋œLAMA์™€ NewLAMA๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์—…๋ฐ์ดํŠธ๋œLAMA๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๊ณผ์ œ๋Š” ์ง€์‹ ์ธ์Šคํ„ด์Šค๊ฐ€ ๋ณ€๊ฒฝ๋œ ์„ธ๋ถ€ ์‚ฌํ•ญ๊ณผ ํ•จ๊ป˜ \(D_{0}\)์™€ \(D_{1}\)์— ๋ชจ๋‘ ์กด์žฌํ•˜๋Š” ๊ฒฝ์šฐ์—๋งŒ ์—…๋ฐ์ดํŠธ๊ฐ€ ํ•„์š”ํ•œ ์ง€์‹์œผ๋กœ ๊ฐ„์ฃผ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค, ๊ทธ๋ฆฌ๊ณ  NewLAMA๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๊ณผ์ œ๋Š” ์ง€์‹์ด \(D_{1}\)์— ์žˆ์ง€๋งŒ \(D_{0}\)์— ์—†๋Š” ๊ฒฝ์šฐ์—๋งŒ ์ƒˆ๋กœ์šด ์ง€์‹์œผ๋กœ ๊ฐ„์ฃผ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ฐ์ดํ„ฐ ๊ตฌ์ถ• ํ”„๋กœ์„ธ์Šค๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—…๋ฐ์ดํŠธ๋œLAMA์™€ NewLAMA์˜ ๋‹จ์ผ ์ธ์Šคํ„ด์Šค ์ƒ์„ฑ์„ ์œ„ํ•œ ํŒŒ์ดํ”„๋ผ์ธ์€ ๊ทธ๋ฆผ 4 (a)์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์ž ์žฌ์  ์ธ์Šคํ„ด์Šค๋Š” CC-RecentNews์˜ ๋‹จ์ผ ๋ฌธ์„œ์—์„œ ์‹œ์ž‘ํ•˜์—ฌ ํŒŒ์ดํ”„๋ผ์ธ์„ ๊ฑฐ์ณ (1) ํ๊ธฐ ๋˜๋Š” (2) ์—…๋ฐ์ดํŠธ๋œLAMA์— ์ถ”๊ฐ€ ๋˜๋Š” (3) ์ตœ์ข…์ ์œผ๋กœ NewLAMA์— ์ถ”๊ฐ€๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋๋‚ฉ๋‹ˆ๋‹ค.
AI standards makers and regulators(e.g., government-backed and industry organizations). This category primarily covers people who are responsible for setting safety standards across the industry. This includes organizations like the AI Safety Institutes in the UK, USA, Japan and Canada, CEN/CENELEC JTC 21 in Europe, the European AI Office, the Infocomm Media Development Authority in Singapore, the International Organization for Standardization, the National Institute of Standards and Technology in the USA, the National Physical Laboratory in the UK, and others across the globe. Our community outreach and research indicates that AI standards makers and regulators' objectives include (i) comparing models and setting standards; (ii) minimizing and mitigating risks from AI; and (iii) ensuring that companies are effectively evaluating their systems' safety.
AI ํ‘œ์ค€ ์ œ์ •์ž ๋ฐ ๊ทœ์ œ ๊ธฐ๊ด€(์˜ˆ: ์ •๋ถ€ ์ง€์› ๋ฐ ์‚ฐ์—… ์กฐ์ง). ์ด ๋ฒ”์ฃผ๋Š” ์ฃผ๋กœ ์‚ฐ์—… ์ „๋ฐ˜์— ๊ฑธ์ณ ์•ˆ์ „ ํ‘œ์ค€์„ ์„ค์ •ํ•˜๋Š” ์ฑ…์ž„์ด ์žˆ๋Š” ์‚ฌ๋žŒ๋“ค์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์˜๊ตญ, ๋ฏธ๊ตญ, ์ผ๋ณธ ๋ฐ ์บ๋‚˜๋‹ค์˜ AI ์•ˆ์ „ ์—ฐ๊ตฌ์†Œ, ์œ ๋Ÿฝ์˜ CEN/CENELEC JTC 21, ์œ ๋Ÿฝ AI ์‚ฌ๋ฌด์†Œ, ์‹ฑ๊ฐ€ํฌ๋ฅด์˜ Infocomm Media Development Authority, ๊ตญ์ œ ํ‘œ์ค€ํ™” ๊ธฐ๊ตฌ, ๋ฏธ๊ตญ์˜ National Institute of Standards and Technology, ์˜๊ตญ์˜ National Physical Laboratory ๋ฐ ์ „ ์„ธ๊ณ„์˜ ๊ธฐํƒ€ ์กฐ์ง์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ปค๋ฎค๋‹ˆํ‹ฐ ์•„์›ƒ๋ฆฌ์น˜ ๋ฐ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด AI ํ‘œ์ค€ ์ œ์ •์ž ๋ฐ ๊ทœ์ œ ๊ธฐ๊ด€์˜ ๋ชฉํ‘œ๋Š” (i) ๋ชจ๋ธ์„ ๋น„๊ตํ•˜๊ณ  ํ‘œ์ค€์„ ์„ค์ •ํ•˜๋Š” ๊ฒƒ; (ii) AI๋กœ ์ธํ•œ ์œ„ํ—˜์„ ์ตœ์†Œํ™”ํ•˜๊ณ  ์™„ํ™”ํ•˜๋Š” ๊ฒƒ; (iii) ๊ธฐ์—…์ด ์‹œ์Šคํ…œ์˜ ์•ˆ์ „์„ฑ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ณ  ์žˆ๋Š”์ง€ ํ™•์ธํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
AI ํ‘œ์ค€ ์ œ์ •์ž ๋ฐ ๊ทœ์ œ์ž(์˜ˆ: ์ •๋ถ€ ์ง€์› ๋ฐ ์—…๊ณ„ ๋‹จ์ฒด). ์ด ๋ฒ”์ฃผ๋Š” ์ฃผ๋กœ ์—…๊ณ„ ์ „๋ฐ˜์˜ ์•ˆ์ „ ํ‘œ์ค€์„ ์„ค์ •ํ•  ์ฑ…์ž„์ด ์žˆ๋Š” ์‚ฌ๋žŒ๋“ค์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์˜๊ตญ, ๋ฏธ๊ตญ, ์ผ๋ณธ, ์บ๋‚˜๋‹ค์˜ AI ์•ˆ์ „ ์—ฐ๊ตฌ์†Œ, ์œ ๋Ÿฝ์˜ CEN/CENELEC JTC 21, ์œ ๋Ÿฝ AI ์‚ฌ๋ฌด์†Œ, ์‹ฑ๊ฐ€ํฌ๋ฅด์˜ ์ •๋ณดํ†ต์‹  ๋ฏธ๋””์–ด ๊ฐœ๋ฐœ์ฒญ, ๊ตญ์ œ ํ‘œ์ค€ํ™” ๊ธฐ๊ตฌ, ๋ฏธ๊ตญ์˜ ๊ตญ๋ฆฝ ํ‘œ์ค€ ๊ธฐ์ˆ  ์—ฐ๊ตฌ์†Œ, ์˜๊ตญ์˜ ๊ตญ๋ฆฝ ๋ฌผ๋ฆฌ ์—ฐ๊ตฌ์†Œ ๋“ฑ๊ณผ ๊ฐ™์€ ์กฐ์ง์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ปค๋ฎค๋‹ˆํ‹ฐ ํ™๋ณด ๋ฐ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด AI ํ‘œ์ค€ ์ œ์ •์ž ๋ฐ ๊ทœ์ œ์ž์˜ ๋ชฉํ‘œ๋Š” (i) ๋ชจ๋ธ ๋น„๊ต ๋ฐ ํ‘œ์ค€ ์„ค์ •, (ii) AI๋กœ ์ธํ•œ ์œ„ํ—˜ ์ตœ์†Œํ™” ๋ฐ ์™„ํ™”, (iii) ๊ธฐ์—…์ด ์‹œ์Šคํ…œ์˜ ์•ˆ์ „์„ฑ์„ ํšจ๊ณผ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ณ  ์žˆ๋Š”์ง€ ํ™•์ธํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
**Medical Summarization.** A dataset consisting of 65 clinician-written summaries of medical notes from MIMIC-III, a large, publicly available database containing medical records of intensive care unit patients [22], was used as additional training data for AMIE. MIMIC-III contains approximately 2 million notes spanning 13 types including cardiology, respiratory, radiology, physician, general, discharge, case management, consult, nursing, pharmacy, nutrition, rehabilitation and social work. 5 notes from each category were selected, with a minimum total length of 400 tokens and at least one nursing note per patient. Clinicians were instructed to write abstractive summaries of individual medical notes, capturing key information while also permitting the inclusion of new informative and clarifying phrases and sentences not present in the original note.
**์˜๋ฃŒ ์š”์•ฝ.** MIMIC-III์—์„œ ์ง‘์ค‘ ์น˜๋ฃŒ์‹ค ํ™˜์ž์˜ ์˜๋ฃŒ ๊ธฐ๋ก์„ ํฌํ•จํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ๊ณต๊ฐœ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์—์„œ ์ž‘์„ฑ๋œ 65๊ฐœ์˜ ์ž„์ƒ์˜ ์š”์•ฝ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹์ด AMIE์˜ ์ถ”๊ฐ€ ํ•™์Šต ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. MIMIC-III๋Š” ์‹ฌ์žฅํ•™, ํ˜ธํก๊ธฐ, ๋ฐฉ์‚ฌ์„ ํ•™, ์˜์‚ฌ, ์ผ๋ฐ˜, ํ‡ด์›, ์‚ฌ๋ก€ ๊ด€๋ฆฌ, ์ƒ๋‹ด, ๊ฐ„ํ˜ธ, ์•ฝ๊ตญ, ์˜์–‘, ์žฌํ™œ ๋ฐ ์‚ฌํšŒ ๋ณต์ง€ ๋“ฑ 13๊ฐ€์ง€ ์œ ํ˜•์— ๊ฑธ์ณ ์•ฝ 200๋งŒ ๊ฐœ์˜ ๋…ธํŠธ๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์นดํ…Œ๊ณ ๋ฆฌ์—์„œ 5๊ฐœ์˜ ๋…ธํŠธ๋ฅผ ์„ ํƒํ–ˆ์œผ๋ฉฐ, ์ด ๊ธธ์ด๊ฐ€ ์ตœ์†Œ 400 ํ† ํฐ ์ด์ƒ์ด๊ณ  ํ™˜์ž๋‹น ์ตœ์†Œ ํ•˜๋‚˜์˜ ๊ฐ„ํ˜ธ ๋…ธํŠธ๋ฅผ ํฌํ•จํ–ˆ์Šต๋‹ˆ๋‹ค. ์ž„์ƒ์˜๋“ค์€ ๊ฐœ๋ณ„ ์˜๋ฃŒ ๋…ธํŠธ์˜ ํ•ต์‹ฌ ์ •๋ณด๋ฅผ ํฌ์ฐฉํ•˜๋ฉด์„œ ์›๋ณธ ๋…ธํŠธ์— ์—†๋Š” ์ƒˆ๋กœ์šด ์ •๋ณด์™€ ๋ช…ํ™•ํ•œ ๊ตฌ๋ฌธ ๋ฐ ๋ฌธ์žฅ์„ ํฌํ•จํ•  ์ˆ˜ ์žˆ๋„๋ก ์ถ”์ƒ์  ์š”์•ฝ์„ ์ž‘์„ฑํ•˜๋ผ๋Š” ์ง€์‹œ๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค.
**์˜๋ฃŒ ์š”์•ฝ**. ์ค‘ํ™˜์ž์‹ค ํ™˜์ž์˜ ์˜๋ฃŒ ๊ธฐ๋ก์„ ํฌํ•จํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ๊ณต๊ฐœ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์ธ MIMIC-III์˜ ์˜๋ฃŒ ๊ธฐ๋ก์— ๋Œ€ํ•œ ์ž„์ƒ์˜๊ฐ€ ์ž‘์„ฑํ•œ 65๊ฐœ์˜ ์š”์•ฝ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ[22]๊ฐ€ AMIE์˜ ์ถ”๊ฐ€ ํ•™์Šต ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. MIMIC-III์—๋Š” ์‹ฌ์žฅํ•™, ํ˜ธํก๊ธฐ, ๋ฐฉ์‚ฌ์„ ํ•™, ์˜์‚ฌ, ์ผ๋ฐ˜, ํ‡ด์›, ์‚ฌ๋ก€ ๊ด€๋ฆฌ, ์ƒ๋‹ด, ๊ฐ„ํ˜ธ, ์•ฝ๊ตญ, ์˜์–‘, ์žฌํ™œ ๋ฐ ์‚ฌํšŒ๋ณต์ง€ ๋“ฑ 13๊ฐ€์ง€ ์œ ํ˜•์— ๊ฑธ์ณ ์•ฝ 2๋ฐฑ๋งŒ ๊ฐœ์˜ ๋…ธํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์นดํ…Œ๊ณ ๋ฆฌ์—์„œ 5๊ฐœ์˜ ๋…ธํŠธ๋ฅผ ์„ ํƒํ–ˆ์œผ๋ฉฐ, ์ด ๊ธธ์ด๋Š” ์ตœ์†Œ 400๊ฐœ ํ† ํฐ ์ด์ƒ์ด์–ด์•ผ ํ•˜๋ฉฐ ํ™˜์ž๋‹น ์ตœ์†Œ 1๊ฐœ์˜ ๊ฐ„ํ˜ธ ๋…ธํŠธ๊ฐ€ ํฌํ•จ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ž„์ƒ์˜๋Š” ๊ฐœ๋ณ„ ์˜๋ฃŒ ๋…ธํŠธ์˜ ์ถ”์ƒ์ ์ธ ์š”์•ฝ์„ ์ž‘์„ฑํ•˜๋„๋ก ์ง€์‹œ๋ฐ›์•˜์œผ๋ฉฐ, ํ•ต์‹ฌ ์ •๋ณด๋ฅผ ํฌ์ฐฉํ•˜๋Š” ๋™์‹œ์— ์›๋ž˜ ๋…ธํŠธ์— ์—†๋Š” ์ƒˆ๋กœ์šด ์œ ์ตํ•œ ๋ฌธ๊ตฌ์™€ ๋ฌธ์žฅ์„ ํฌํ•จํ•  ์ˆ˜ ์žˆ๋„๋ก ํ–ˆ์Šต๋‹ˆ๋‹ค.
Our domain-adaptive continual pre-training can be viewed from the lens of unsupervised domain adaptation [7]. Our source data is the large unsupervised domain corpus, while the target data is the target task data. With pre-training, we do not have control over the alignment with task training data itself; our idea is that by aligning with the domain during pre-training, we could align the LLM with the task. This intuition is backed by evidence of LLM pre-training helping the performance over open domain tasks. We use the generalization bound from [7; 2] since our problem is similar to unsupervised domain adaptation. Consider a hypothesis space \(\mathcal{H}_{p}\) with \(f\in\mathcal{H}_{p}\); generalization errors on source \(\mathcal{D}\) and task data \(\mathcal{T}\) as \(\epsilon_{\mathcal{D}}\) and \(\epsilon_{\mathcal{T}}\), respectively.
์šฐ๋ฆฌ์˜ ๋„๋ฉ”์ธ ์ ์‘ ์ง€์† ์‚ฌ์ „ ํ•™์Šต์€ ๋น„์ง€๋„ ๋„๋ฉ”์ธ ์ ์‘ [7]์˜ ๊ด€์ ์—์„œ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์†Œ์Šค ๋ฐ์ดํ„ฐ๋Š” ๋Œ€๊ทœ๋ชจ ๋น„์ง€๋„ ๋„๋ฉ”์ธ ์ฝ”ํผ์Šค์ด๊ณ , ํƒ€๊ฒŸ ๋ฐ์ดํ„ฐ๋Š” ํƒ€๊ฒŸ ์ž‘์—… ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ์ž‘์—… ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์ž์ฒด์™€์˜ ์ •๋ ฌ์„ ์ œ์–ดํ•  ์ˆ˜๋Š” ์—†์ง€๋งŒ, ์‚ฌ์ „ ํ•™์Šต ์ค‘ ๋„๋ฉ”์ธ๊ณผ์˜ ์ •๋ ฌ์„ ํ†ตํ•ด LLM์„ ์ž‘์—…๊ณผ ์ •๋ ฌ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ์•„์ด๋””์–ด์ž…๋‹ˆ๋‹ค. ์ด ์ง๊ด€์€ LLM ์‚ฌ์ „ ํ•™์Šต์ด ์˜คํ”ˆ ๋„๋ฉ”์ธ ์ž‘์—…์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋œ๋‹ค๋Š” ์ฆ๊ฑฐ๋กœ ๋’ท๋ฐ›์นจ๋ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฌธ์ œ๋Š” ๋น„์ง€๋„ ๋„๋ฉ”์ธ ์ ์‘๊ณผ ์œ ์‚ฌํ•˜๊ธฐ ๋•Œ๋ฌธ์— [7; 2]์˜ ์ผ๋ฐ˜ํ™” ๊ฒฝ๊ณ„๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์„ค ๊ณต๊ฐ„ \(\mathcal{H}_{p}\)์—์„œ \(f\in\mathcal{H}_{p}\)๋ฅผ ๊ณ ๋ คํ•˜๊ณ , ์†Œ์Šค \(\mathcal{D}\)์™€ ์ž‘์—… ๋ฐ์ดํ„ฐ \(\mathcal{T}\)์— ๋Œ€ํ•œ ์ผ๋ฐ˜ํ™” ์˜ค๋ฅ˜๋ฅผ ๊ฐ๊ฐ \(\epsilon_{\mathcal{D}}\)์™€ \(\epsilon_{\mathcal{T}}\)๋กœ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์šฐ๋ฆฌ์˜ ๋„๋ฉ”์ธ ์ ์‘ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์€ ๋น„์ง€๋„ ๋„๋ฉ”์ธ ์ ์‘[7]์˜ ๋ Œ์ฆˆ๋ฅผ ํ†ตํ•ด ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์†Œ์Šค ๋ฐ์ดํ„ฐ๋Š” ๋Œ€๊ทœ๋ชจ ๋น„์ง€๋„ ๋„๋ฉ”์ธ ์ฝ”ํผ์Šค์ด๊ณ , ๋Œ€์ƒ ๋ฐ์ดํ„ฐ๋Š” ๋Œ€์ƒ ์ž‘์—… ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ์ž‘์—… ํ•™์Šต ๋ฐ์ดํ„ฐ ์ž์ฒด์™€์˜ ์ •๋ ฌ์„ ์ œ์–ดํ•  ์ˆ˜ ์—†์œผ๋ฉฐ, ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋„๋ฉ”์ธ๊ณผ ์ •๋ ฌํ•จ์œผ๋กœ์จ LLM์„ ์ž‘์—…๊ณผ ์ •๋ ฌํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์•„์ด๋””์–ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง๊ด€์€ LLM ์‚ฌ์ „ ํ•™์Šต์ด ์˜คํ”ˆ ๋„๋ฉ”์ธ ์ž‘์—…์˜ ์„ฑ๋Šฅ์— ๋„์›€์ด ๋œ๋‹ค๋Š” ์ฆ๊ฑฐ์— ์˜ํ•ด ๋’ท๋ฐ›์นจ๋ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฌธ์ œ๋Š” ๋น„์ง€๋„ ๋„๋ฉ”์ธ ์ ์‘๊ณผ ์œ ์‚ฌํ•˜๊ธฐ ๋•Œ๋ฌธ์— [7; 2]์˜ ์ผ๋ฐ˜ํ™” ๋ฐ”์šด๋“œ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์„ค ๊ณต๊ฐ„ \(\mathcal{H}_{p}\)์—์„œ \(f\in\mathcal{H}_{p}\), ์†Œ์Šค \(\mathcal{D}\) ๋ฐ ์ž‘์—… ๋ฐ์ดํ„ฐ \(\mathcal{T}\)์— ๋Œ€ํ•œ ์ผ๋ฐ˜ํ™” ์˜ค๋ฅ˜๋ฅผ ๊ฐ๊ฐ \(\epsilon_{\mathcal{D}}\) ๋ฐ \(\epsilon_{\mathcal{T}}\)๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค.
Privacy is an important subject for future work, as memorized training data has significant privacy consequences. By this, we mean the standard privacy definition that a model should not reveal anything particular to the specific dataset it was trained on, as opposed to another training dataset from a similar distribution (Shokri et al., 2017).4 Training on standard datasets that have not yet been deduplicated results in models that are particularly sensitive to examples that happened to be repeated multiple times, and this has negative privacy implications. For instance, it could violate a person's expectations of privacy if their publicly available personal data appeared in a different, surprising context. Downstream applications of LMs, such as the game AI Dungeon5, should also not output memorized content like adverts for real products.
ํ”„๋ผ์ด๋ฒ„์‹œ๋Š” ์•”๊ธฐ๋œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๊ฐ€ ์ค‘์š”ํ•œ ํ”„๋ผ์ด๋ฒ„์‹œ ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ ์ค‘์š”ํ•œ ์ฃผ์ œ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์šฐ๋ฆฌ๋Š” ๋ชจ๋ธ์ด ์œ ์‚ฌํ•œ ๋ถ„ํฌ์˜ ๋‹ค๋ฅธ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹๊ณผ๋Š” ๋‹ฌ๋ฆฌ, ํ›ˆ๋ จ๋œ ํŠน์ • ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•œ ์–ด๋– ํ•œ ํŠน์ • ์ •๋ณด๋„ ๋“œ๋Ÿฌ๋‚ด์ง€ ์•Š์•„์•ผ ํ•œ๋‹ค๋Š” ํ‘œ์ค€ ํ”„๋ผ์ด๋ฒ„์‹œ ์ •์˜๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค (Shokri et al., 2017). ์ค‘๋ณต ์ œ๊ฑฐ๋˜์ง€ ์•Š์€ ํ‘œ์ค€ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ํ›ˆ๋ จํ•  ๊ฒฝ์šฐ, ๋ชจ๋ธ์€ ์—ฌ๋Ÿฌ ๋ฒˆ ๋ฐ˜๋ณต๋œ ์˜ˆ์ œ์— ํŠนํžˆ ๋ฏผ๊ฐํ•ด์ง€๋ฉฐ, ์ด๋Š” ๋ถ€์ •์ ์ธ ํ”„๋ผ์ด๋ฒ„์‹œ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ณต๊ฐœ์ ์œผ๋กœ ์ด์šฉ ๊ฐ€๋Šฅํ•œ ๊ฐœ์ธ ๋ฐ์ดํ„ฐ๊ฐ€ ๋‹ค๋ฅธ ๋†€๋ผ์šด ๋งฅ๋ฝ์—์„œ ๋‚˜ํƒ€๋‚œ๋‹ค๋ฉด, ์ด๋Š” ๊ฐœ์ธ์˜ ํ”„๋ผ์ด๋ฒ„์‹œ ๊ธฐ๋Œ€๋ฅผ ์œ„๋ฐ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒŒ์ž„ AI Dungeon๊ณผ ๊ฐ™์€ LMs์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜๋„ ์‹ค์ œ ์ œํ’ˆ ๊ด‘๊ณ ์™€ ๊ฐ™์€ ์•”๊ธฐ๋œ ์ฝ˜ํ…์ธ ๋ฅผ ์ถœ๋ ฅํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค.
์•”๊ธฐ๋œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋Š” ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ์— ์ƒ๋‹นํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น˜๊ธฐ ๋•Œ๋ฌธ์— ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ๋Š” ํ–ฅํ›„ ์ž‘์—…์˜ ์ค‘์š”ํ•œ ์ฃผ์ œ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ๋งํ•˜๋Š” ํ‘œ์ค€ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ์ •์˜๋Š” ๋ชจ๋ธ์ด ์œ ์‚ฌํ•œ ๋ถ„ํฌ์˜ ๋‹ค๋ฅธ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋‹ฌ๋ฆฌ ํ•™์Šต๋œ ํŠน์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ํŠน๋ณ„ํžˆ ๊ณต๊ฐœํ•ด์„œ๋Š” ์•ˆ ๋œ๋‹ค๋Š” ์˜๋ฏธ์ž…๋‹ˆ๋‹ค(Shokri et al., 2017).4 ์•„์ง ์ค‘๋ณต์ด ์ œ๊ฑฐ๋˜์ง€ ์•Š์€ ํ‘œ์ค€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ํ•™์Šต์€ ํŠนํžˆ ์—ฌ๋Ÿฌ ๋ฒˆ ๋ฐ˜๋ณต๋œ ์˜ˆ์ œ์— ๋ฏผ๊ฐํ•œ ๋ชจ๋ธ์„ ์ƒ์„ฑํ•˜๋ฉฐ, ์ด๋Š” ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ์— ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๊ฐœ์ธ ๋ฐ์ดํ„ฐ๊ฐ€ ์˜ˆ์ƒ์น˜ ๋ชปํ•œ ๋‹ค๋ฅธ ๋งฅ๋ฝ์—์„œ ๋‚˜ํƒ€๋‚  ๊ฒฝ์šฐ ๊ฐœ์ธ์˜ ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ์— ๋Œ€ํ•œ ๊ธฐ๋Œ€์น˜๋ฅผ ์œ„๋ฐ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒŒ์ž„ AI ๋˜์ „5์™€ ๊ฐ™์€ LM์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜๋„ ์‹ค์ œ ์ œํ’ˆ์— ๋Œ€ํ•œ ๊ด‘๊ณ ์™€ ๊ฐ™์€ ๊ธฐ์–ต๋œ ์ฝ˜ํ…์ธ ๋ฅผ ์ถœ๋ ฅํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค.
* Pearson's official website lists fewer than 2,100 textbooks.34* Wiley's official website lists fewer than 69,000 textbooks.35 Footnote 35: [https://www.wiley.com/en-us/subjects](https://www.wiley.com/en-us/subjects), accessed March 2024. We wrote a code to sum up all the books in all of their subcategories; our code may double count books, so this is only a safe upper bound. We used this number instead of the โ€œ21,000โ€ online books mentioned on [https://www.wiley.com/learn/librarysolutions/online-books-purchase.html](https://www.wiley.com/learn/librarysolutions/online-books-purchase.html), accessed March 2024. * McGraw Hill lists fewer than 22,000 textbooks for PreK-12 education, many of which have significant content overlap (as many are tailored for one of the 50 US states).36 They list fewer than 2,000 textbooks for higher education.
Pearson์˜ ๊ณต์‹ ์›น์‚ฌ์ดํŠธ์—๋Š” 2,100๊ถŒ ๋ฏธ๋งŒ์˜ ๊ต๊ณผ์„œ๊ฐ€ ๋‚˜์—ด๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.34 Wiley์˜ ๊ณต์‹ ์›น์‚ฌ์ดํŠธ์—๋Š” 69,000๊ถŒ ๋ฏธ๋งŒ์˜ ๊ต๊ณผ์„œ๊ฐ€ ๋‚˜์—ด๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.35 ๊ฐ์ฃผ 35: [https://www.wiley.com/en-us/subjects](https://www.wiley.com/en-us/subjects), 2024๋…„ 3์›” ์ ‘์†. ์šฐ๋ฆฌ๋Š” ๋ชจ๋“  ํ•˜์œ„ ์นดํ…Œ๊ณ ๋ฆฌ์˜ ์ฑ…์„ ํ•ฉ์‚ฐํ•˜๋Š” ์ฝ”๋“œ๋ฅผ ์ž‘์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค; ์ด ์ฝ”๋“œ๊ฐ€ ์ฑ…์„ ์ค‘๋ณต ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ, ์ด๋Š” ์•ˆ์ „ํ•œ ์ƒํ•œ์„ ์ผ ๋ฟ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด ์ˆซ์ž๋ฅผ [https://www.wiley.com/learn/librarysolutions/online-books-purchase.html](https://www.wiley.com/learn/librarysolutions/online-books-purchase.html)์—์„œ ์–ธ๊ธ‰๋œ "21,000"๊ถŒ์˜ ์˜จ๋ผ์ธ ์ฑ… ๋Œ€์‹  ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค, 2024๋…„ 3์›” ์ ‘์†. McGraw Hill์€ PreK-12 ๊ต์œก์„ ์œ„ํ•ด 22,000๊ถŒ ๋ฏธ๋งŒ์˜ ๊ต๊ณผ์„œ๋ฅผ ๋‚˜์—ดํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด ์ค‘ ๋งŽ์€ ์ฑ…๋“ค์ด ์ƒ๋‹นํ•œ ๋‚ด์šฉ ์ค‘๋ณต์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค (๋งŽ์€ ์ฑ…๋“ค์ด ๋ฏธ๊ตญ 50๊ฐœ ์ฃผ ์ค‘ ํ•˜๋‚˜๋ฅผ ์œ„ํ•ด ๋งž์ถค ์ œ์ž‘๋˜์—ˆ์Šต๋‹ˆ๋‹ค).36 ๊ทธ๋“ค์€ ๊ณ ๋“ฑ ๊ต์œก์„ ์œ„ํ•ด 2,000๊ถŒ ๋ฏธ๋งŒ์˜ ๊ต๊ณผ์„œ๋ฅผ ๋‚˜์—ดํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
* ํ”ผ์–ด์Šจ์˜ ๊ณต์‹ ์›น์‚ฌ์ดํŠธ์—๋Š” 2,100๊ถŒ ๋ฏธ๋งŒ์˜ ๊ต๊ณผ์„œ๊ฐ€ ๋‚˜์—ด๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.34* ์™€์ผ๋ฆฌ์˜ ๊ณต์‹ ์›น์‚ฌ์ดํŠธ์—๋Š” 69,000๊ถŒ ๋ฏธ๋งŒ์˜ ๊ต๊ณผ์„œ๊ฐ€ ๋‚˜์—ด๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.35 ๊ฐ์ฃผ 35: [https://www.wiley.com/en-us/subjects](https://www.wiley.com/en-us/subjects), 2024๋…„ 3์›”์— ์•ก์„ธ์Šคํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ํ•˜์œ„ ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๋ชจ๋“  ์ฑ…์„ ํ•ฉ์‚ฐํ•˜๋Š” ์ฝ”๋“œ๋ฅผ ์ž‘์„ฑํ–ˆ์ง€๋งŒ, ์ด ์ฝ”๋“œ๋Š” ์ฑ…์„ ์ด์ค‘์œผ๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์•ˆ์ „ํ•œ ์ƒํ•œ์„ ์ผ ๋ฟ์ž…๋‹ˆ๋‹ค. 2024๋…„ 3์›”์— ์•ก์„ธ์Šคํ•œ [https://www.wiley.com/learn/librarysolutions/online-books-purchase.html](https://www.wiley.com/learn/librarysolutions/online-books-purchase.html)์— ์–ธ๊ธ‰๋œ '21,000๊ถŒ'์˜ ์˜จ๋ผ์ธ ๋„์„œ ๋Œ€์‹  ์ด ์ˆซ์ž๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. * ๋งฅ๊ทธ๋กœํž์€ ์ดˆ์ค‘๊ณ  ๊ต์œก์šฉ ๊ต๊ณผ์„œ๋ฅผ 22,000๊ถŒ ๋ฏธ๋งŒ์œผ๋กœ ๋‚˜์—ดํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด ์ค‘ ์ƒ๋‹น์ˆ˜๋Š” (๋ฏธ๊ตญ์˜ 50๊ฐœ ์ฃผ ์ค‘ ํ•˜๋‚˜์— ๋งž์ถฐ ์ œ์ž‘๋œ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๊ธฐ ๋•Œ๋ฌธ์—) ์ƒ๋‹นํ•œ ๋‚ด์šฉ์ด ๊ฒน์นฉ๋‹ˆ๋‹ค.36 ๊ณ ๋“ฑ ๊ต์œก์šฉ ๊ต๊ณผ์„œ๋Š” 2,000๊ถŒ ๋ฏธ๋งŒ์„ ๋‚˜์—ดํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
To further compare the performance of PPO and DPO under distribution shifts, we evaluate the PPO and DPO policies from our Reddit TL;DR summarization experiment on a different distribution, news articles in the test split of the CNN/DailyMail dataset [24], using the best sampling temperatures from TL;DR (0 and 0.25). The results are presented in Table 1. We computed the GPT-4 win rate against the ground-truth summaries in the datasets, using the same GPT-4 (C) prompt we used for Reddit TL;DR, but replacing the words "forum post" with "news article". For this new distribution, DPO continues to outperform the PPO policy by a significant margin. This experiment provides initial evidence that DPO policies can generalize similarly well to PPO policies, even though DPO does not use the additional unlabeled Reddit TL;DR prompts that PPO uses.
PPO์™€ DPO์˜ ์„ฑ๋Šฅ์„ ๋ถ„ํฌ ๋ณ€ํ™” ํ•˜์—์„œ ์ถ”๊ฐ€๋กœ ๋น„๊ตํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” Reddit TL;DR ์š”์•ฝ ์‹คํ—˜์—์„œ ์‚ฌ์šฉํ•œ PPO์™€ DPO ์ •์ฑ…์„ ๋‹ค๋ฅธ ๋ถ„ํฌ์ธ CNN/DailyMail ๋ฐ์ดํ„ฐ์…‹ [24]์˜ ํ…Œ์ŠคํŠธ ๋ถ„ํ• ์— ์žˆ๋Š” ๋‰ด์Šค ๊ธฐ์‚ฌ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. TL;DR์—์„œ ์‚ฌ์šฉํ•œ ์ตœ์ ์˜ ์ƒ˜ํ”Œ๋ง ์˜จ๋„(0 ๋ฐ 0.25)๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ๋Š” Table 1์— ์ œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” Reddit TL;DR์—์„œ ์‚ฌ์šฉํ•œ ๋™์ผํ•œ GPT-4 (C) ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ์…‹์˜ ์‹ค์ œ ์š”์•ฝ๊ณผ ๋น„๊ตํ•˜์—ฌ GPT-4 ์Šน๋ฅ ์„ ๊ณ„์‚ฐํ–ˆ์œผ๋ฉฐ, "forum post"๋ผ๋Š” ๋‹จ์–ด๋ฅผ "news article"๋กœ ๋Œ€์ฒดํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ƒˆ๋กœ์šด ๋ถ„ํฌ์—์„œ๋„ DPO๋Š” ์—ฌ์ „ํžˆ PPO ์ •์ฑ…๋ณด๋‹ค ์ƒ๋‹นํžˆ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ์ด ์‹คํ—˜์€ DPO ์ •์ฑ…์ด ์ถ”๊ฐ€์ ์ธ ๋ผ๋ฒจ์ด ์—†๋Š” Reddit TL;DR ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  PPO ์ •์ฑ…๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ์ž˜ ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ดˆ๊ธฐ ์ฆ๊ฑฐ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๋ถ„ํฌ ์ด๋™ ํ•˜์—์„œ PPO์™€ DPO์˜ ์„ฑ๋Šฅ์„ ์ถ”๊ฐ€๋กœ ๋น„๊ตํ•˜๊ธฐ ์œ„ํ•ด, Reddit TL;DR ์š”์•ฝ ์‹คํ—˜์—์„œ ์–ป์€ PPO ๋ฐ DPO ์ •์ฑ…์„ CNN/DailyMail ๋ฐ์ดํ„ฐ ์„ธํŠธ[24]์˜ ํ…Œ์ŠคํŠธ ๋ถ„ํ• ์— ์žˆ๋Š” ๋‰ด์Šค ๊ธฐ์‚ฌ๋ผ๋Š” ๋‹ค๋ฅธ ๋ถ„ํฌ์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•˜๊ณ , TL;DR์˜ ์ตœ์  ์ƒ˜ํ”Œ๋ง ์˜จ๋„(0 ๋ฐ 0.25)๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ๋Š” ํ‘œ 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์‹ค์ธก ์š”์•ฝ๊ณผ ๋น„๊ตํ•˜์—ฌ GPT-4์˜ ์Šน๋ฅ ์„ ๊ณ„์‚ฐํ–ˆ์œผ๋ฉฐ, Reddit TL;DR์— ์‚ฌ์šฉํ•œ ๊ฒƒ๊ณผ ๋™์ผํ•œ GPT-4 (C) ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ–ˆ์ง€๋งŒ "ํฌ๋Ÿผ ๊ฒŒ์‹œ๋ฌผ"์ด๋ผ๋Š” ๋‹จ์–ด๋ฅผ "๋‰ด์Šค ๊ธฐ์‚ฌ"๋กœ ๋Œ€์ฒดํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ƒˆ๋กœ์šด ๋ถ„ํฌ์—์„œ DPO๋Š” ์—ฌ์ „ํžˆ PPO ์ •์ฑ…๋ณด๋‹ค ์ƒ๋‹นํ•œ ๊ฒฉ์ฐจ๋กœ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค. ์ด ์‹คํ—˜์€ DPO๊ฐ€ PPO๊ฐ€ ์‚ฌ์šฉํ•˜๋Š” ์ถ”๊ฐ€ ๋ผ๋ฒจ์ด ์—†๋Š” Reddit TL;DR ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  DPO ์ •์ฑ…์ด PPO ์ •์ฑ…๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ดˆ๊ธฐ ์ฆ๊ฑฐ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
Further work is also needed to ensure the robustness of medical LLMs in multilingual settings [105, 106, 107, 108], and particularly their performance in low-resource languages [109]. The great variety of cultures [110], languages, localities, identities, and localized medical needs, makes the task of generating a priori static yet comprehensive fairness benchmarks practically infeasible. Measurement and mitigation of bias must move beyond the traditional narrow focus on specific axes that fails to scale globally [111]. LLM-based evaluators present a potential solution for preliminary assessments in languages where there are no systematic benchmarks, though prior studies have found these auto-evaluation frameworks to be biased, underscoring the need for calibrating them on native speaker evaluations, and using them with caution [112].
์ถ”๊ฐ€ ์—ฐ๊ตฌ๋Š” ๋‹ค๊ตญ์–ด ํ™˜๊ฒฝ์—์„œ ์˜๋ฃŒ LLM์˜ ๊ฒฌ๊ณ ์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ํ•„์š”ํ•˜๋ฉฐ [105, 106, 107, 108], ํŠนํžˆ ์ž์›์ด ๋ถ€์กฑํ•œ ์–ธ์–ด์—์„œ์˜ ์„ฑ๋Šฅ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค [109]. ๋‹ค์–‘ํ•œ ๋ฌธํ™” [110], ์–ธ์–ด, ์ง€์—ญ, ์ •์ฒด์„ฑ, ๊ทธ๋ฆฌ๊ณ  ์ง€์—ญํ™”๋œ ์˜๋ฃŒ ์š”๊ตฌ ์‚ฌํ•ญ์˜ ๋‹ค์–‘์„ฑ์€ ์‚ฌ์ „ ์ •์ ์ด๋ฉด์„œ๋„ ํฌ๊ด„์ ์ธ ๊ณต์ •์„ฑ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ž‘์—…์„ ์‚ฌ์‹ค์ƒ ๋ถˆ๊ฐ€๋Šฅํ•˜๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ํŽธํ–ฅ์˜ ์ธก์ •๊ณผ ์™„ํ™”๋Š” ์ „ํ†ต์ ์ธ ํŠน์ • ์ถ•์— ๋Œ€ํ•œ ์ข์€ ์ดˆ์ ์„ ๋„˜์–ด ์ „ ์„ธ๊ณ„์ ์œผ๋กœ ํ™•์žฅ๋˜์ง€ ์•Š๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋„˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค [111]. LLM ๊ธฐ๋ฐ˜ ํ‰๊ฐ€์ž๋Š” ์ฒด๊ณ„์ ์ธ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ์—†๋Š” ์–ธ์–ด์—์„œ ์˜ˆ๋น„ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ์ž ์žฌ์  ํ•ด๊ฒฐ์ฑ…์„ ์ œ์‹œํ•˜์ง€๋งŒ, ์ด์ „ ์—ฐ๊ตฌ๋“ค์€ ์ด๋Ÿฌํ•œ ์ž๋™ ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ๊ฐ€ ํŽธํ–ฅ๋˜์–ด ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ•˜์—ฌ, ์ด๋ฅผ ์›์–ด๋ฏผ ํ‰๊ฐ€๋กœ ๋ณด์ •ํ•˜๊ณ  ์‹ ์ค‘ํ•˜๊ฒŒ ์‚ฌ์šฉํ•˜๋Š” ํ•„์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค [112].
๋‹ค๊ตญ์–ด ํ™˜๊ฒฝ์—์„œ ์˜๋ฃŒ์šฉ LLM์˜ ๊ฒฌ๊ณ ์„ฑ[105, 106, 107, 108]์„ ๋ณด์žฅํ•˜๊ณ , ํŠนํžˆ ์ €์ž์› ์–ธ์–ด์—์„œ์˜ ์„ฑ๋Šฅ[109]์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•œ ์ถ”๊ฐ€ ์—ฐ๊ตฌ๋„ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ฌธํ™”[110], ์–ธ์–ด, ์ง€์—ญ, ์ •์ฒด์„ฑ, ํ˜„์ง€ํ™”๋œ ์˜๋ฃŒ ์ˆ˜์š”๋กœ ์ธํ•ด ์„ ํ—˜์ ์œผ๋กœ ์ •์ ์ธ ๋™์‹œ์— ํฌ๊ด„์ ์ธ ๊ณต์ •์„ฑ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ž‘์—…์€ ํ˜„์‹ค์ ์œผ๋กœ ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ํŽธํ–ฅ์„ฑ์˜ ์ธก์ •๊ณผ ์™„ํ™”๋Š” ์ „ ์„ธ๊ณ„์ ์œผ๋กœ ํ™•์žฅ๋˜์ง€ ๋ชปํ•˜๋Š” ํŠน์ • ์ถ•์— ๋Œ€ํ•œ ์ „ํ†ต์ ์ธ ํ˜‘์†Œํ•œ ์ดˆ์ ์„ ๋„˜์–ด์„œ๋Š” ๊ฒƒ์ด์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค[111]. LLM ๊ธฐ๋ฐ˜ ํ‰๊ฐ€๊ธฐ๋Š” ์ฒด๊ณ„์ ์ธ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ์—†๋Š” ์–ธ์–ด์—์„œ ์˜ˆ๋น„ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ์ž ์žฌ์ ์ธ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•˜์ง€๋งŒ, ์ด์ „ ์—ฐ๊ตฌ์—์„œ ์ด๋Ÿฌํ•œ ์ž๋™ ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ๊ฐ€ ํŽธํ–ฅ์ ์ด๋ผ๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์ ธ ์›์–ด๋ฏผ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด ๋ณด์ •ํ•˜๊ณ  ์‹ ์ค‘ํ•˜๊ฒŒ ์‚ฌ์šฉํ•ด์•ผ ํ•  ํ•„์š”์„ฑ์ด ๊ฐ•์กฐ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[112].
Figure 1: A performance comparison featuring our proposed model ALMA-13B-R against other recently released 13B LLM-based models, as well as top-performing translation systems like GPT-4 and WMT winners. This evaluation covers the WMTโ€™22 test data across 8 directions, involving translations to and from English for German, Czech, Chinese, and Russian. Scores are averaged by three different reference-free models: wmt23-cometkiwi-da-xxl, XCOMET-XXL, and wmt22-cometkiwi-da, and are also averaged across all directions. The gold reference is also evaluated due to the reference-free approach. Our model, ALMA-13B-R, developed by further training ALMA-13B-LoRA using our proposed CPO method, either matches or surpasses the most advanced translation models, We show the detailed numerical data for all systems presented in the figure in Appendix A.
Figure 1: ์ตœ๊ทผ์— ์ถœ์‹œ๋œ ๋‹ค๋ฅธ 13B LLM ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๋“ค๊ณผ GPT-4 ๋ฐ WMT ์šฐ์Šน์ž์™€ ๊ฐ™์€ ์ตœ๊ณ  ์„ฑ๋Šฅ์˜ ๋ฒˆ์—ญ ์‹œ์Šคํ…œ๊ณผ ๋น„๊ตํ•˜์—ฌ, ์ œ์•ˆ๋œ ๋ชจ๋ธ ALMA-13B-R์˜ ์„ฑ๋Šฅ ๋น„๊ต. ์ด ํ‰๊ฐ€๋Š” ๋…์ผ์–ด, ์ฒด์ฝ”์–ด, ์ค‘๊ตญ์–ด, ๋Ÿฌ์‹œ์•„์–ด์— ๋Œ€ํ•ด ์˜์–ด๋กœ์˜ ๋ฒˆ์—ญ๊ณผ ์˜์–ด์—์„œ์˜ ๋ฒˆ์—ญ์„ ํฌํ•จํ•˜์—ฌ 8๊ฐœ์˜ ๋ฐฉํ–ฅ์— ๊ฑธ์ณ WMTโ€™22 ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์ ์ˆ˜๋Š” wmt23-cometkiwi-da-xxl, XCOMET-XXL, wmt22-cometkiwi-da๋ผ๋Š” ์„ธ ๊ฐ€์ง€ ๋‹ค๋ฅธ ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ์— ์˜ํ•ด ํ‰๊ท ํ™”๋˜๋ฉฐ, ๋ชจ๋“  ๋ฐฉํ–ฅ์— ๊ฑธ์ณ ํ‰๊ท ํ™”๋ฉ๋‹ˆ๋‹ค. ์ฐธ์กฐ ์—†๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์ธํ•ด ๊ณจ๋“œ ์ฐธ์กฐ๋„ ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค. ์ œ์•ˆ๋œ CPO ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ALMA-13B-LoRA๋ฅผ ์ถ”๊ฐ€๋กœ ํ›ˆ๋ จํ•˜์—ฌ ๊ฐœ๋ฐœ๋œ ์šฐ๋ฆฌ ๋ชจ๋ธ ALMA-13B-R์€ ๊ฐ€์žฅ ์ง„๋ณด๋œ ๋ฒˆ์—ญ ๋ชจ๋ธ๊ณผ ๋™๋“ฑํ•˜๊ฑฐ๋‚˜ ์ด๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ์— ์ œ์‹œ๋œ ๋ชจ๋“  ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ˆ˜์น˜ ๋ฐ์ดํ„ฐ๋Š” ๋ถ€๋ก A์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1: ์ตœ๊ทผ ์ถœ์‹œ๋œ 13B LLM ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๊ณผ GPT-4 ๋ฐ WMT ์ˆ˜์ƒ์ž์™€ ๊ฐ™์€ ์ตœ๊ณ  ์„ฑ๋Šฅ์˜ ๋ฒˆ์—ญ ์‹œ์Šคํ…œ์„ ๋Œ€์ƒ์œผ๋กœ ์ œ์•ˆ๋œ ๋ชจ๋ธ ALMA-13B-R์„ ๋น„๊ตํ•œ ์„ฑ๋Šฅ์ž…๋‹ˆ๋‹ค. ์ด ํ‰๊ฐ€๋Š” ๋…์ผ์–ด, ์ฒด์ฝ”์–ด, ์ค‘๊ตญ์–ด, ๋Ÿฌ์‹œ์•„์–ด์˜ ์˜์–ด ๋ฒˆ์—ญ์„ ํฌํ•จํ•˜์—ฌ 8๊ฐ€์ง€ ๋ฐฉํ–ฅ์˜ WMT'22 ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์„ธ ๊ฐ€์ง€ ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ(wmt23-cometkiwi-da-xxl, XCOMET-XXL, wmt22-cometkiwi-da)์„ ํ†ตํ•ด ํ‰๊ท ํ™”๋˜๋ฉฐ ๋ชจ๋“  ๋ฐฉํ–ฅ์— ๊ฑธ์ณ ํ‰๊ท ํ™”๋ฉ๋‹ˆ๋‹ค. ์ฐธ์กฐ ์—†๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์ธํ•ด ๊ณจ๋“œ ์ฐธ์กฐ๋„ ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค. ์ œ์•ˆ๋œ CPO ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ALMA-13B-LoRA๋ฅผ ์ถ”๊ฐ€๋กœ ํ›ˆ๋ จํ•˜์—ฌ ๊ฐœ๋ฐœํ•œ ์ €ํฌ ๋ชจ๋ธ์ธ ALMA-13B-R์€ ๊ฐ€์žฅ ์ง„๋ณด๋œ ๋ฒˆ์—ญ ๋ชจ๋ธ๊ณผ ๋™๋“ฑํ•˜๊ฑฐ๋‚˜ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ์— ์ œ์‹œ๋œ ๋ชจ๋“  ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ˆ˜์น˜ ๋ฐ์ดํ„ฐ๋Š” ๋ถ€๋ก A์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
While the world knowledge stored in LMs has diverse use cases, it can quickly become outdated as the world changes fast, and LMs need to frequently renew their internal world knowledge accordingly. For example, it is impossible to probe for _new_ information such as "_- _won the US Election 2020_" from the original T5 (Raffel et al., 2019) which was pretrained on C4 web corpus from April 2019.1 Also, information that may have once been considered accurate may no longer be valid because the information has been _updated_. For instance, the answer to "_Which soccer team does Cristiano Ronaldo play for?_" has changed from _Juventus_ to _Manchester United_ in September 2021. Meanwhile, _time-invariant_ information learned from the original corpus such as "_Barack Obama was born in Honolulu, Hawaii_" should not be altered within the LMs.
LMs์— ์ €์žฅ๋œ ์„ธ๊ณ„ ์ง€์‹์€ ๋‹ค์–‘ํ•œ ์‚ฌ์šฉ ์‚ฌ๋ก€๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์ง€๋งŒ, ์„ธ๊ณ„๊ฐ€ ๋น ๋ฅด๊ฒŒ ๋ณ€ํ™”ํ•จ์— ๋”ฐ๋ผ ๋น ๋ฅด๊ฒŒ ๊ตฌ์‹์ด ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, LMs๋Š” ์ด์— ๋”ฐ๋ผ ๋‚ด๋ถ€ ์„ธ๊ณ„ ์ง€์‹์„ ์ž์ฃผ ๊ฐฑ์‹ ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 2019๋…„ 4์›”์˜ C4 ์›น ์ฝ”ํผ์Šค๋ฅผ ์‚ฌ์ „ ํ•™์Šตํ•œ ์›๋ž˜ T5 (Raffel et al., 2019)์—์„œ "_- _2020๋…„ ๋ฏธ๊ตญ ์„ ๊ฑฐ์—์„œ ์Šน๋ฆฌํ•œ ์‚ฌ๋žŒ_"๊ณผ ๊ฐ™์€ _์ƒˆ๋กœ์šด_ ์ •๋ณด๋ฅผ ํƒ์ƒ‰ํ•˜๋Š” ๊ฒƒ์€ ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ•œ๋•Œ ์ •ํ™•ํ•˜๋‹ค๊ณ  ์—ฌ๊ฒจ์กŒ๋˜ ์ •๋ณด๊ฐ€ _์—…๋ฐ์ดํŠธ_๋˜๋ฉด์„œ ๋” ์ด์ƒ ์œ ํšจํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, "_ํฌ๋ฆฌ์Šคํ‹ฐ์•„๋ˆ„ ํ˜ธ๋‚ ๋‘๊ฐ€ ์–ด๋Š ์ถ•๊ตฌํŒ€์—์„œ ๋›ฐ๊ณ  ์žˆ๋‚˜์š”?_"์— ๋Œ€ํ•œ ๋‹ต๋ณ€์€ 2021๋…„ 9์›”์— _์œ ๋ฒคํˆฌ์Šค_์—์„œ _๋งจ์ฒด์Šคํ„ฐ ์œ ๋‚˜์ดํ‹ฐ๋“œ_๋กœ ๋ณ€๊ฒฝ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ•œํŽธ, "_๋ฒ„๋ฝ ์˜ค๋ฐ”๋งˆ๋Š” ํ•˜์™€์ด ํ˜ธ๋†€๋ฃฐ๋ฃจ์—์„œ ํƒœ์–ด๋‚ฌ๋‹ค_"์™€ ๊ฐ™์€ ์›๋ž˜ ์ฝ”ํผ์Šค์—์„œ ํ•™์Šตํ•œ _์‹œ๊ฐ„ ๋ถˆ๋ณ€_ ์ •๋ณด๋Š” LMs ๋‚ด์—์„œ ๋ณ€๊ฒฝ๋˜์ง€ ์•Š์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค.
LM์— ์ €์žฅ๋œ ์„ธ๊ณ„ ์ง€์‹์€ ๋‹ค์–‘ํ•œ ์‚ฌ์šฉ ์‚ฌ๋ก€๋กœ ํ™œ์šฉ๋˜์ง€๋งŒ, ์„ธ์ƒ์ด ๋น ๋ฅด๊ฒŒ ๋ณ€ํ™”ํ•จ์— ๋”ฐ๋ผ ๋น ๋ฅด๊ฒŒ ๋‚™ํ›„๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด์— ๋”ฐ๋ผ LM์€ ๋‚ด๋ถ€ ์„ธ๊ณ„ ์ง€์‹์„ ์ž์ฃผ ๊ฐฑ์‹ ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 2019๋…„ 4์›”๋ถ€ํ„ฐ C4 ์›น ์ฝ”ํผ์Šค์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต๋œ ์›๋ž˜ T5(Raffel et al., 2019)์—์„œ๋Š” "_- _2020๋…„ ๋ฏธ๊ตญ ๋Œ€์„ ์—์„œ ์Šน๋ฆฌํ–ˆ๋‹ค_"์™€ ๊ฐ™์€ _์ƒˆ๋กœ์šด_ ์ •๋ณด๋ฅผ ํƒ์ƒ‰ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.1 ๋˜ํ•œ, ํ•œ๋•Œ ์ •ํ™•ํ•œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋˜์—ˆ๋˜ ์ •๋ณด๊ฐ€ _์—…๋ฐ์ดํŠธ_๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋” ์ด์ƒ ์œ ํšจํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, "_ํฌ๋ฆฌ์Šคํ‹ฐ์•„๋ˆ„ ํ˜ธ๋‚ ๋‘๋Š” ์–ด๋Š ์ถ•๊ตฌํŒ€์—์„œ ๋›ฐ๋‚˜์š”?_"์— ๋Œ€ํ•œ ๋‹ต๋ณ€์€ 2021๋…„ 9์›”์— _์œ ๋ฒคํˆฌ์Šค_์—์„œ _๋งจ์ฒด์Šคํ„ฐ ์œ ๋‚˜์ดํ‹ฐ๋“œ_๋กœ ๋ณ€๊ฒฝ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ•œํŽธ, "_๋ฒ„๋ฝ ์˜ค๋ฐ”๋งˆ๋Š” ํ•˜์™€์ด ํ˜ธ๋†€๋ฃฐ๋ฃจ์—์„œ ํƒœ์–ด๋‚ฌ๋‹ค_"์™€ ๊ฐ™์ด ์›๋ž˜ ์ฝ”ํผ์Šค์—์„œ ํ•™์Šตํ•œ _์‹œ๊ฐ„ ๋ถˆ๋ณ€_ ์ •๋ณด๋Š” LM ๋‚ด์—์„œ ๋ณ€๊ฒฝ๋˜์–ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค.
Subsequently, we calculate CnstScore\((\mathbf{x}^{l})\) using \(\mathbf{\hat{x}}^{l}(\alpha)\),4 which allows us to express it as a function CnstScore\((\alpha)\) of \(\alpha\). Then, we examine its derivative, \(\frac{d}{d\alpha}\text{CnstScore}(\alpha)\big{|}_{\alpha=0}\) to understand the direction of change at the current value. A positive derivative indicates that an increase in EntRec\((\mathbf{x}^{l})\) leads to an increase in CnstScore\((\tau_{\text{2H}},\tau_{\text{1H}})\), while a negative one suggests the opposite. By assessing _the relative frequency of positive gradients_ among the two-hop prompts in TwoHopFact, we quantify how often the LLM performs the second hop of the reasoning, with frequencies above 0.5 suggesting that the LLM's chance to perform the second-hop reasoning exceeds random chance for these prompts.
์ดํ›„, ์šฐ๋ฆฌ๋Š” \(\mathbf{\hat{x}}^{l}(\alpha)\)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ CnstScore\((\mathbf{x}^{l})\)๋ฅผ ๊ณ„์‚ฐํ•˜๊ณ , ์ด๋ฅผ \(\alpha\)์˜ ํ•จ์ˆ˜์ธ CnstScore\((\alpha)\)๋กœ ํ‘œํ˜„ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ํ˜„์žฌ ๊ฐ’์—์„œ์˜ ๋ณ€ํ™” ๋ฐฉํ–ฅ์„ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด \(\frac{d}{d\alpha}\text{CnstScore}(\alpha)\big{|}_{\alpha=0}\)์˜ ๋„ํ•จ์ˆ˜๋ฅผ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ์–‘์˜ ๋„ํ•จ์ˆ˜๋Š” EntRec\((\mathbf{x}^{l})\)์˜ ์ฆ๊ฐ€๊ฐ€ CnstScore\((\tau_{\text{2H}},\tau_{\text{1H}})\)์˜ ์ฆ๊ฐ€๋กœ ์ด์–ด์ง์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์Œ์˜ ๋„ํ•จ์ˆ˜๋Š” ๊ทธ ๋ฐ˜๋Œ€๋ฅผ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. TwoHopFact์˜ ๋‘ ๋‹จ๊ณ„ ํ”„๋กฌํ”„ํŠธ ์ค‘ ์–‘์˜ ๊ธฐ์šธ๊ธฐ์˜ ์ƒ๋Œ€์  ๋นˆ๋„๋ฅผ ํ‰๊ฐ€ํ•จ์œผ๋กœ์จ, LLM์ด ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์˜ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋นˆ๋„๋ฅผ ์ •๋Ÿ‰ํ™”ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋นˆ๋„๊ฐ€ 0.5๋ฅผ ์ดˆ๊ณผํ•˜๋ฉด LLM์ด ์ด๋Ÿฌํ•œ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์˜ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ํ™•๋ฅ ์ด ๋ฌด์ž‘์œ„ ํ™•๋ฅ ์„ ์ดˆ๊ณผํ•จ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
์ด์–ด์„œ, ์šฐ๋ฆฌ๋Š” \(\mathbf{\hat{x}}^{l}(\alpha)\),4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ CnstScore\((\mathbf{x}^{l})\)๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ ์ด๋ฅผ \(\alpha\)์˜ ํ•จ์ˆ˜ CnstScore\((\alpha)\)๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ˜„์žฌ ๊ฐ’์—์„œ ๋ณ€ํ™”์˜ ๋ฐฉํ–ฅ์„ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ ๋„ํ•จ์ˆ˜์ธ \(\frac{d}{d\alpha}\text{CnstScore}(\alpha)\big{|}_{\alpha=0}\)๋ฅผ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๋„ํ•จ์ˆ˜๊ฐ€ ์–‘์ˆ˜์ด๋ฉด EntRec\((\mathbf{x}^{l})\)์˜ ์ฆ๊ฐ€๊ฐ€ CnstScore\((\tau_{\text{2H}},\tau_{\text{1H}})\)์˜ ์ฆ๊ฐ€๋กœ ์ด์–ด์ง€๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์Œ์ˆ˜์ด๋ฉด ๊ทธ ๋ฐ˜๋Œ€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. TwoHopFact์˜ 2ํ™‰ ํ”„๋กฌํ”„ํŠธ ์ค‘ ์–‘์ˆ˜ ๊ทธ๋ผ๋ฐ์ด์…˜์˜ _์ƒ๋Œ€์  ๋นˆ๋„_๋ฅผ ํ‰๊ฐ€ํ•˜์—ฌ LLM์ด ์ถ”๋ก ์˜ ๋‘ ๋ฒˆ์งธ ํ™‰์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋นˆ๋„๋ฅผ ์ •๋Ÿ‰ํ™”ํ•˜๋ฉฐ, ๋นˆ๋„๊ฐ€ 0.5 ์ด์ƒ์ด๋ฉด ํ•ด๋‹น ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด LLM์ด ๋‘ ๋ฒˆ์งธ ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ํ™•๋ฅ ์ด ๋ฌด์ž‘์œ„ ํ™•๋ฅ ์„ ์ดˆ๊ณผํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
Task-adaptive pre-training.Continual pre-training of language models on unlabeled data for a given task has been demonstrated to be beneficial for enhancing end-task performance [1; 8; 13]. In scenarios involving domain shift, domain-adaptive pre-training bears similarities to task-adaptive pre-training to some extent. Aharoni et al [1] documented that continual pre-training a model on a similar domain contributes to improved task performance on the target domain. Notably, the work closest to ours is presented in [8], which shows that continual pre-training of language models on both unlabeled task data and augmented unlabeled task data, sampled from the in-domain corpus based on similarity. While these works use task data, we also propose a task agnostic method, ETA-DACP, as task similarity is prohibitively expensive for LLMs.
๊ณผ์ œ ์ ์‘ํ˜• ์‚ฌ์ „ ํ•™์Šต. ์ฃผ์–ด์ง„ ๊ณผ์ œ์— ๋Œ€ํ•ด ๋น„ํ‘œ์‹œ ๋ฐ์ดํ„ฐ๋กœ ์–ธ์–ด ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ์ตœ์ข… ๊ณผ์ œ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์œ ์ตํ•˜๋‹ค๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค [1; 8; 13]. ๋„๋ฉ”์ธ ์ด๋™์ด ํฌํ•จ๋œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ๋Š” ๋„๋ฉ”์ธ ์ ์‘ํ˜• ์‚ฌ์ „ ํ•™์Šต์ด ์–ด๋Š ์ •๋„ ๊ณผ์ œ ์ ์‘ํ˜• ์‚ฌ์ „ ํ•™์Šต๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. Aharoni ๋“ฑ [1]์€ ์œ ์‚ฌํ•œ ๋„๋ฉ”์ธ์—์„œ ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ๋Œ€์ƒ ๋„๋ฉ”์ธ์—์„œ์˜ ๊ณผ์ œ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๊ธฐ์—ฌํ•œ๋‹ค๊ณ  ๊ธฐ๋กํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์šฐ๋ฆฌ ์—ฐ๊ตฌ์™€ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์ž‘์—…์€ [8]์— ์ œ์‹œ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์œ ์‚ฌ์„ฑ์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ๋„๋ฉ”์ธ ๋‚ด ๋ง๋ญ‰์น˜์—์„œ ์ƒ˜ํ”Œ๋ง๋œ ๋น„ํ‘œ์‹œ ๊ณผ์ œ ๋ฐ์ดํ„ฐ์™€ ์ฆ๊ฐ•๋œ ๋น„ํ‘œ์‹œ ๊ณผ์ œ ๋ฐ์ดํ„ฐ ๋ชจ๋‘์—์„œ ์–ธ์–ด ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ๋“ค์ด ๊ณผ์ œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ฐ˜๋ฉด, ์šฐ๋ฆฌ๋Š” ๊ณผ์ œ ์œ ์‚ฌ์„ฑ์ด LLMs์— ๋Œ€ํ•ด ์ง€๋‚˜์น˜๊ฒŒ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ธฐ ๋•Œ๋ฌธ์— ๊ณผ์ œ ๋น„์ข…์†์ ์ธ ๋ฐฉ๋ฒ•์ธ ETA-DACP๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
์ž‘์—… ์ ์‘ํ˜• ์‚ฌ์ „ ํ•™์Šต.์ฃผ์–ด์ง„ ์ž‘์—…์— ๋Œ€ํ•ด ๋ ˆ์ด๋ธ”์ด ์—†๋Š” ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์€ ์ตœ์ข… ์ž‘์—… ์„ฑ๋Šฅ ํ–ฅ์ƒ์— ๋„์›€์ด ๋˜๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[1; 8; 13]. ๋„๋ฉ”์ธ ์‹œํ”„ํŠธ๊ฐ€ ํฌํ•จ๋œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋„๋ฉ”์ธ ์ ์‘ํ˜• ์‚ฌ์ „ ํ•™์Šต์€ ์–ด๋Š ์ •๋„ ์ž‘์—… ์ ์‘ํ˜• ์‚ฌ์ „ ํ•™์Šต๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. Aharoni ๋“ฑ[1]์€ ์œ ์‚ฌํ•œ ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์ด ๋Œ€์ƒ ๋„๋ฉ”์ธ์—์„œ์˜ ์ž‘์—… ์„ฑ๋Šฅ ํ–ฅ์ƒ์— ๊ธฐ์—ฌํ•œ๋‹ค๊ณ  ๋ฌธ์„œํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์šฐ๋ฆฌ์™€ ๊ฐ€์žฅ ์œ ์‚ฌํ•œ ์ž‘์—…์€ [8]์— ์ œ์‹œ๋œ ๊ฒƒ์œผ๋กœ, ์œ ์‚ฌ๋„๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋„๋ฉ”์ธ ๋‚ด ๋ง๋ญ‰์น˜์—์„œ ์ƒ˜ํ”Œ๋งํ•œ ๋ ˆ์ด๋ธ”์ด ์—†๋Š” ์ž‘์—… ๋ฐ์ดํ„ฐ์™€ ์ฆ๊ฐ•๋œ ๋ ˆ์ด๋ธ”์ด ์—†๋Š” ์ž‘์—… ๋ฐ์ดํ„ฐ ๋ชจ๋‘์— ๋Œ€ํ•ด ์–ธ์–ด ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ๋Š” ์ž‘์—… ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์ง€๋งŒ, LLM์˜ ๊ฒฝ์šฐ ์ž‘์—… ์œ ์‚ฌ์„ฑ์ด ๋„ˆ๋ฌด ๋น„์‹ธ๊ธฐ ๋•Œ๋ฌธ์— ์ž‘์—…์— ๊ตฌ์• ๋ฐ›์ง€ ์•Š๋Š” ๋ฐฉ๋ฒ•์ธ ETA-DACP๋„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
We consider facts, such as _"The mother of Stevie Wonder is Lula"_, as triplets \((e,r,e^{\prime})\) of a subject entity \(e\) (e.g., Superstition), a relation \(r\) (e.g., mother), and an object entity \(e^{\prime}\) (e.g., Lula). Specifically, in our analysis, we focus on triplets where \(e^{\prime}\) is the only or the most well-known object entity for the relation \(r\) for \(e\) (e.g. the only mother of Stevie Wonder is Lula), and view \(r\) as a function \(e^{\prime}=r(e)\), where \(r(e)\) is the function expression and \(e^{\prime}\) is the value of the expression. We analyze how LLMs process the composition of two facts with a bridge entity \(e_{2}\) connecting them, \(((e_{1},r_{1},e_{2}),(e_{2},r_{2},e_{3}))\), of which the composition is represented as \(r_{2}(r_{1}(e_{1}))\). An example is shown in Table 1.
์šฐ๋ฆฌ๋Š” _"Stevie Wonder์˜ ์–ด๋จธ๋‹ˆ๋Š” Lula์ด๋‹ค"_์™€ ๊ฐ™์€ ์‚ฌ์‹ค์„ ์ฃผ์ฒด ์—”ํ‹ฐํ‹ฐ \(e\) (์˜ˆ: Superstition), ๊ด€๊ณ„ \(r\) (์˜ˆ: ์–ด๋จธ๋‹ˆ), ๊ฐ์ฒด ์—”ํ‹ฐํ‹ฐ \(e^{\prime}\) (์˜ˆ: Lula)๋กœ ๊ตฌ์„ฑ๋œ ์‚ผ์ค‘ํ•ญ \((e,r,e^{\prime})\)์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์šฐ๋ฆฌ์˜ ๋ถ„์„์—์„œ๋Š” \(e^{\prime}\)์ด \(e\)์— ๋Œ€ํ•œ ๊ด€๊ณ„ \(r\)์˜ ์œ ์ผํ•˜๊ฑฐ๋‚˜ ๊ฐ€์žฅ ์ž˜ ์•Œ๋ ค์ง„ ๊ฐ์ฒด ์—”ํ‹ฐํ‹ฐ์ธ ์‚ผ์ค‘ํ•ญ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  (์˜ˆ: Stevie Wonder์˜ ์œ ์ผํ•œ ์–ด๋จธ๋‹ˆ๋Š” Lula์ด๋‹ค), \(r\)์„ ํ•จ์ˆ˜ \(e^{\prime}=r(e)\)๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ \(r(e)\)๋Š” ํ•จ์ˆ˜ ํ‘œํ˜„์ด๊ณ  \(e^{\prime}\)๋Š” ํ‘œํ˜„์˜ ๊ฐ’์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋‘ ์‚ฌ์‹ค์ด ์—ฐ๊ฒฐ ์—”ํ‹ฐํ‹ฐ \(e_{2}\)๋กœ ์—ฐ๊ฒฐ๋œ ์กฐํ•ฉ \(((e_{1},r_{1},e_{2}),(e_{2},r_{2},e_{3}))\)์„ LLM์ด ์–ด๋–ป๊ฒŒ ์ฒ˜๋ฆฌํ•˜๋Š”์ง€ ๋ถ„์„ํ•˜๋ฉฐ, ์ด ์กฐํ•ฉ์€ \(r_{2}(r_{1}(e_{1}))\)๋กœ ํ‘œํ˜„๋ฉ๋‹ˆ๋‹ค. ์˜ˆ์‹œ๋Š” Table 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
์šฐ๋ฆฌ๋Š” _"์Šคํ‹ฐ๋น„ ์›๋”์˜ ์–ด๋จธ๋‹ˆ๋Š” ๋ฃฐ๋ผ"_,์™€ ๊ฐ™์€ ์‚ฌ์‹ค์„ ์ฃผ์–ด ์—”ํ‹ฐํ‹ฐ \(e\) (์˜ˆ: ์ˆ ๋ถ€), ๊ด€๊ณ„ \(r\) (์˜ˆ: ์–ด๋จธ๋‹ˆ), ๊ฐ์ฒด ์—”ํ‹ฐํ‹ฐ \(e^{\prime}\) (์˜ˆ: ๋ฃฐ๋ผ)์˜ ์‚ผ์ค‘ ํ•ญ \((e,r,e^{\prime})\)์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์šฐ๋ฆฌ์˜ ๋ถ„์„์—์„œ๋Š” \(e^{\prime}\)๊ฐ€ \(e\)์— ๋Œ€ํ•œ ๊ด€๊ณ„ \(r\)์— ๋Œ€ํ•ด ์œ ์ผํ•˜๊ฑฐ๋‚˜ ๊ฐ€์žฅ ์ž˜ ์•Œ๋ ค์ง„ ๊ฐ์ฒด ์—”ํ‹ฐํ‹ฐ์ธ ์‚ผ์ค‘ ํ•ญ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ , \(r\)์„ ํ•จ์ˆ˜ \(e^{\prime}=r(e)\)๋กœ ๊ฐ„์ฃผํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(r(e)\)๋Š” ํ•จ์ˆ˜ ์‹์ด๊ณ  \(e^{\prime}\)๋Š” ์‹์˜ ๊ฐ’์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋‘ ๊ฐœ์˜ ์‚ฌ์‹ค์ด ์—ฐ๊ฒฐ๋˜๋Š” ๋‹ค๋ฆฌ ์—”ํ‹ฐํ‹ฐ \(e_{2}\)๋ฅผ ๊ฐ€์ง„ ๋‘ ๊ฐœ์˜ ์‚ฌ์‹ค์˜ ๊ตฌ์„ฑ์„ ์–ด๋–ป๊ฒŒ ์ฒ˜๋ฆฌํ•˜๋Š”์ง€ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค.((e_{1},r_{1},e_{2}),(e_{2},r_{2},e_{3}))\), ์ด ์ค‘ ๊ตฌ์„ฑ์€ \(r_{2}(r_{1}(e_{1}))\)๋กœ ํ‘œํ˜„๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋Š” ํ‘œ 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
To further validate the behavior of Adapt-LLM when requesting additional context, Figure 2 illustrates the proportion of questions for which our model generates the \(\langle\text{RET}\rangle\) token, aggregated by popularity score intervals (left image for Adapt-LLM trained on NQ and right image for SQuAD). Mallen, Alex Troy and Asai, Akari and Zhong, Victor and Das, Rajarshi and Khashabi, Daniel and Hajishirzi, Hannaneh [22] suggest that high-popularity questions can be adequately answered using the parametric memory of the LLM, while lower popularity scores necessitate extra context. In Figure 2, we observe this pattern for both versions of Adapt-LLM, indicating that our model, despite lacking access to popularity scores during training or inference, has learned effective criteria for requesting additional context.
Adapt-LLM์ด ์ถ”๊ฐ€์ ์ธ ์ปจํ…์ŠคํŠธ๋ฅผ ์š”์ฒญํ•  ๋•Œ์˜ ๋™์ž‘์„ ์ถ”๊ฐ€๋กœ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด, Figure 2๋Š” ์šฐ๋ฆฌ ๋ชจ๋ธ์ด \(\langle\text{RET}\rangle\) ํ† ํฐ์„ ์ƒ์„ฑํ•˜๋Š” ์งˆ๋ฌธ์˜ ๋น„์œจ์„ ์ธ๊ธฐ๋„ ์ ์ˆ˜ ๊ตฌ๊ฐ„๋ณ„๋กœ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค (NQ์—์„œ ํ›ˆ๋ จ๋œ Adapt-LLM์˜ ์™ผ์ชฝ ์ด๋ฏธ์ง€์™€ SQuAD์˜ ์˜ค๋ฅธ์ชฝ ์ด๋ฏธ์ง€). Mallen, Alex Troy, Asai, Akari, Zhong, Victor, Das, Rajarshi, Khashabi, Daniel, Hajishirzi, Hannaneh [22]๋Š” ์ธ๊ธฐ๋„๊ฐ€ ๋†’์€ ์งˆ๋ฌธ์€ LLM์˜ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ ์ ˆํžˆ ๋‹ต๋ณ€ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ์ธ๊ธฐ๋„ ์ ์ˆ˜๊ฐ€ ๋‚ฎ์€ ๊ฒฝ์šฐ ์ถ”๊ฐ€์ ์ธ ์ปจํ…์ŠคํŠธ๊ฐ€ ํ•„์š”ํ•˜๋‹ค๊ณ  ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Figure 2์—์„œ ์šฐ๋ฆฌ๋Š” ๋‘ ๋ฒ„์ „์˜ Adapt-LLM์— ๋Œ€ํ•ด ์ด ํŒจํ„ด์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์šฐ๋ฆฌ ๋ชจ๋ธ์ด ํ›ˆ๋ จ์ด๋‚˜ ์ถ”๋ก  ์ค‘์— ์ธ๊ธฐ๋„ ์ ์ˆ˜์— ์ ‘๊ทผํ•˜์ง€ ๋ชปํ–ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ถ”๊ฐ€์ ์ธ ์ปจํ…์ŠคํŠธ๋ฅผ ์š”์ฒญํ•˜๋Š” ํšจ๊ณผ์ ์ธ ๊ธฐ์ค€์„ ํ•™์Šตํ–ˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๋ฅผ ์š”์ฒญํ•  ๋•Œ Adapt-LLM์˜ ๋™์ž‘์„ ์ถ”๊ฐ€๋กœ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฆผ 2๋Š” ์ธ๊ธฐ๋„ ์ ์ˆ˜ ๊ฐ„๊ฒฉ๋ณ„๋กœ ์ง‘๊ณ„ํ•œ, ์šฐ๋ฆฌ ๋ชจ๋ธ์ด \(\langle\text{RET}\rangle\) ํ† ํฐ์„ ์ƒ์„ฑํ•˜๋Š” ์งˆ๋ฌธ์˜ ๋น„์œจ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(์™ผ์ชฝ ์ด๋ฏธ์ง€๋Š” NQ๋กœ ํ•™์Šต๋œ Adapt-LLM์˜ ๊ฒฝ์šฐ, ์˜ค๋ฅธ์ชฝ ์ด๋ฏธ์ง€๋Š” SQuAD์˜ ๊ฒฝ์šฐ). Mallen, Alex Troy, Asai, Akari, Zhong, Victor, Das, Rajarshi, Khashabi, Daniel, Hajishirzi, Hannaneh [22]๋Š” ์ธ๊ธฐ๋„๊ฐ€ ๋†’์€ ์งˆ๋ฌธ์€ LLM์˜ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ ์ ˆํ•˜๊ฒŒ ๋‹ต๋ณ€ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ์ธ๊ธฐ๋„๊ฐ€ ๋‚ฎ์€ ์งˆ๋ฌธ์€ ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๊ฐ€ ํ•„์š”ํ•˜๋‹ค๊ณ  ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2์—์„œ ๋‘ ๋ฒ„์ „์˜ Adapt-LLM ๋ชจ๋‘์—์„œ ์ด๋Ÿฌํ•œ ํŒจํ„ด์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ•™์Šต ๋˜๋Š” ์ถ”๋ก  ์ค‘์— ์ธ๊ธฐ๋„ ์ ์ˆ˜์— ๋Œ€ํ•œ ์•ก์„ธ์Šค ๊ถŒํ•œ์ด ์—†์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ ์š”์ฒญ์„ ์œ„ํ•œ ํšจ๊ณผ์ ์ธ ๊ธฐ์ค€์„ ํ•™์Šตํ•œ ๊ฒƒ์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
Third, _we explore the underlying reasons for the effectiveness of our method by revealing its impact on learning dependencies._ By conducting both quantitative analyses and case studies, we observed that this method demonstrates an enhanced dependency on the initial mathematical questions and earlier steps. This suggests a possible reason: Instead of depending solely on the local information generated by the model itself from recent steps, which can lead to a higher likelihood of hallucination or misdirection as the sequence extends, our method utilizes more information from the problem statement and earlier steps. These sources are more reliable and less prone to errors. Consequently, this strategy might reduce the risk of misunderstanding the problem and inconsistencies in reasoning (refer to Secs. 2.2 and 3.4).
์…‹์งธ, _์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ํšจ๊ณผ์— ๋Œ€ํ•œ ๊ทผ๋ณธ์ ์ธ ์ด์œ ๋ฅผ ํ•™์Šต ์˜์กด์„ฑ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํ†ตํ•ด ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค._ ์ •๋Ÿ‰์  ๋ถ„์„๊ณผ ์‚ฌ๋ก€ ์—ฐ๊ตฌ๋ฅผ ๋ชจ๋‘ ์ˆ˜ํ–‰ํ•œ ๊ฒฐ๊ณผ, ์ด ๋ฐฉ๋ฒ•์ด ์ดˆ๊ธฐ ์ˆ˜ํ•™์  ์งˆ๋ฌธ๊ณผ ์ดˆ๊ธฐ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ์˜์กด์„ฑ์„ ๊ฐ•ํ™”ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ฐ€๋Šฅํ•œ ์ด์œ ๋ฅผ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค: ์‹œํ€€์Šค๊ฐ€ ํ™•์žฅ๋จ์— ๋”ฐ๋ผ ํ™˜๊ฐ์ด๋‚˜ ์˜ค๋„ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ์ตœ๊ทผ ๋‹จ๊ณ„์—์„œ ๋ชจ๋ธ ์ž์ฒด๊ฐ€ ์ƒ์„ฑํ•œ ๋กœ์ปฌ ์ •๋ณด์—๋งŒ ์˜์กดํ•˜๋Š” ๋Œ€์‹ , ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์€ ๋ฌธ์ œ ์ง„์ˆ ๊ณผ ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ๋” ๋งŽ์€ ์ •๋ณด๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถœ์ฒ˜๋Š” ๋” ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๊ณ  ์˜ค๋ฅ˜ ๋ฐœ์ƒ ๊ฐ€๋Šฅ์„ฑ์ด ์ ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ์ „๋žต์€ ๋ฌธ์ œ๋ฅผ ์˜คํ•ดํ•˜๊ฑฐ๋‚˜ ์ถ”๋ก ์˜ ์ผ๊ด€์„ฑ ๋ฌธ์ œ๋ฅผ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์„น์…˜ 2.2 ๋ฐ 3.4 ์ฐธ์กฐ).
์…‹์งธ, _ํ•™์Šต ์ข…์†์„ฑ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋ฐํž˜์œผ๋กœ์จ ์šฐ๋ฆฌ ๋ฐฉ๋ฒ•์˜ ํšจ๊ณผ์— ๋Œ€ํ•œ ๊ทผ๋ณธ์ ์ธ ์ด์œ ๋ฅผ ํƒ๊ตฌํ•œ๋‹ค._ ์ •๋Ÿ‰์  ๋ถ„์„๊ณผ ์‚ฌ๋ก€ ์—ฐ๊ตฌ๋ฅผ ํ†ตํ•ด ์ด ๋ฐฉ๋ฒ•์ด ์ดˆ๊ธฐ ์ˆ˜ํ•™์  ์งˆ๋ฌธ๊ณผ ์ด์ „ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ์ข…์†์„ฑ์„ ๊ฐ•ํ™”ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ด ๋ฐฉ๋ฒ•์ด ์ตœ๊ทผ ๋‹จ๊ณ„์—์„œ ๋ชจ๋ธ ์ž์ฒด๊ฐ€ ์ƒ์„ฑํ•œ ๋กœ์ปฌ ์ •๋ณด์—๋งŒ ์˜์กดํ•˜๋Š” ๋Œ€์‹  ๋ฌธ์ œ ์ง„์ˆ ๊ณผ ์ด์ „ ๋‹จ๊ณ„์˜ ์ •๋ณด๋ฅผ ๋” ๋งŽ์ด ํ™œ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ถœ์ฒ˜๋Š” ๋” ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๊ณ  ์˜ค๋ฅ˜๊ฐ€ ๋œ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ์ „๋žต์€ ๋ฌธ์ œ๋ฅผ ์˜คํ•ดํ•˜๊ฑฐ๋‚˜ ์ถ”๋ก ์— ์ผ๊ด€์„ฑ์ด ์—†๋Š” ์œ„ํ—˜์„ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(2.2์ ˆ ๋ฐ 3.4์ ˆ ์ฐธ์กฐ).
The source of pre-training corpus can be broadly categorized into two types: general data and specialized data. General data, such as webpages, books, and conversational text, is utilized by most LLMs [55, 56, 90] due to its large, diverse, and accessible nature, which can enhance the language modeling and generalization abilities of LLMs. In light of the impressive generalization capabilities exhibited by LLMs, there are also studies that extend their pre-training corpus to more specialized datasets, such as multilingual data, scientific data, and code, endowing LLMs with specific task-solving capabilities [35, 56, 86]. In what follows, we describe these two types of pre-training data sources and their effects on LLMs. For a detailed introduction to the commonly used corpus, one can refer to Section 3.2.
์‚ฌ์ „ ํ›ˆ๋ จ ์ฝ”ํผ์Šค์˜ ์ถœ์ฒ˜๋Š” ํฌ๊ฒŒ ์ผ๋ฐ˜ ๋ฐ์ดํ„ฐ์™€ ํŠน์ˆ˜ ๋ฐ์ดํ„ฐ์˜ ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์œผ๋กœ ๋ถ„๋ฅ˜๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์›นํŽ˜์ด์ง€, ์ฑ…, ๋Œ€ํ™” ํ…์ŠคํŠธ์™€ ๊ฐ™์€ ์ผ๋ฐ˜ ๋ฐ์ดํ„ฐ๋Š” ๊ทธ ๊ทœ๋ชจ๊ฐ€ ํฌ๊ณ  ๋‹ค์–‘ํ•˜๋ฉฐ ์ ‘๊ทผ์„ฑ์ด ๋†’์•„ ๋Œ€๋ถ€๋ถ„์˜ LLMs [55, 56, 90]์—์„œ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ด๋Š” LLMs์˜ ์–ธ์–ด ๋ชจ๋ธ๋ง ๋ฐ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. LLMs๊ฐ€ ๋ณด์—ฌ์ฃผ๋Š” ๋†€๋ผ์šด ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์— ๋น„์ถ”์–ด, ๋‹ค๊ตญ์–ด ๋ฐ์ดํ„ฐ, ๊ณผํ•™ ๋ฐ์ดํ„ฐ, ์ฝ”๋“œ์™€ ๊ฐ™์€ ๋” ํŠน์ˆ˜ํ•œ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ ์ฝ”ํผ์Šค๋ฅผ ํ™•์žฅํ•˜์—ฌ LLMs์— ํŠน์ • ์ž‘์—… ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ๋ถ€์—ฌํ•˜๋Š” ์—ฐ๊ตฌ๋„ ์žˆ์Šต๋‹ˆ๋‹ค [35, 56, 86]. ๋‹ค์Œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์˜ ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์†Œ์Šค์™€ ๊ทธ๋“ค์ด LLMs์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ์ฝ”ํผ์Šค์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์†Œ๊ฐœ๋Š” ์„น์…˜ 3.2๋ฅผ ์ฐธ์กฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์˜ ์ถœ์ฒ˜๋Š” ํฌ๊ฒŒ ์ผ๋ฐ˜ ๋ฐ์ดํ„ฐ์™€ ์ „๋ฌธ ๋ฐ์ดํ„ฐ์˜ ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์œผ๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์›นํŽ˜์ด์ง€, ์ฑ…, ๋Œ€ํ™” ํ…์ŠคํŠธ์™€ ๊ฐ™์€ ์ผ๋ฐ˜ ๋ฐ์ดํ„ฐ๋Š” ๋Œ€์šฉ๋Ÿ‰, ๋‹ค์–‘์„ฑ, ์ ‘๊ทผ์„ฑ์ด ๋›ฐ์–ด๋‚˜๊ธฐ ๋•Œ๋ฌธ์— ๋Œ€๋ถ€๋ถ„์˜ LLM[55, 56, 90]์—์„œ ํ™œ์šฉ๋˜๋ฉฐ, ์ด๋Š” LLM์˜ ์–ธ์–ด ๋ชจ๋ธ๋ง ๋ฐ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. LLM์ด ๋ณด์—ฌ์ค€ ์ธ์ƒ์ ์ธ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์— ๋น„์ถ”์–ด, ๋‹ค๊ตญ์–ด ๋ฐ์ดํ„ฐ, ๊ณผํ•™ ๋ฐ์ดํ„ฐ, ์ฝ”๋“œ์™€ ๊ฐ™์€ ๋ณด๋‹ค ์ „๋ฌธํ™”๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค๋ฅผ ํ™•์žฅํ•˜์—ฌ LLM์— ํŠน์ • ์ž‘์—… ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ๋ถ€์—ฌํ•˜๋Š” ์—ฐ๊ตฌ๋„ ์ง„ํ–‰๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[35, 56, 86]. ๋‹ค์Œ์—์„œ๋Š” ์ด ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์†Œ์Šค์™€ LLM์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ง๋ญ‰์น˜์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์†Œ๊ฐœ๋Š” ์„น์…˜ 3.2๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
In this work, we propose **Web** Rephrase **A**ugmented Pre-training (**WRAP**)--that attempts to bridge three important challenges stemming from the ambiguity around data curation--(**i**) what data should you pre-train on? (ii) how can you pre-train with limited data? (iii) how can you pre-train computationally efficiently? In particular, we show that re-phrasing documents on the web using an off-the-shelf medium size LLM allows models to learn much more efficiently than learning from raw text on the web, and accounts for performance gains on out of distribution datasets that _can not_ be offset with additional web data. Our proposed method involves using a pre-trained off-the-shelf LLM to re-phrase documents from a web corpus into different styles. An overview of our approach is shown in Figure 0(a).
์ด ์—ฐ๊ตฌ์—์„œ ์šฐ๋ฆฌ๋Š” ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜์˜ ๋ชจํ˜ธ์„ฑ์—์„œ ๋น„๋กฏ๋œ ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋ ค๋Š” **Web** Rephrase **A**ugmented Pre-training (**WRAP**)์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค--(**i**) ์–ด๋–ค ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ํ•ด์•ผ ํ•˜๋Š”๊ฐ€? (ii) ์ œํ•œ๋œ ๋ฐ์ดํ„ฐ๋กœ ์–ด๋–ป๊ฒŒ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€? (iii) ์–ด๋–ป๊ฒŒ ๊ณ„์‚ฐ ํšจ์œจ์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€? ํŠนํžˆ, ์šฐ๋ฆฌ๋Š” ์ค‘๊ฐ„ ํฌ๊ธฐ์˜ ๊ธฐ์„ฑ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์›น์˜ ๋ฌธ์„œ๋ฅผ ๋‹ค์‹œ ํ‘œํ˜„ํ•˜๋Š” ๊ฒƒ์ด ์›น์˜ ์›์‹œ ํ…์ŠคํŠธ์—์„œ ํ•™์Šตํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ํ›จ์”ฌ ๋” ํšจ์œจ์ ์œผ๋กœ ๋ชจ๋ธ์ด ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•˜๋ฉฐ, ์ถ”๊ฐ€์ ์ธ ์›น ๋ฐ์ดํ„ฐ๋กœ๋Š” ์ƒ์‡„ํ•  ์ˆ˜ ์—†๋Š” ๋ถ„ํฌ ์™ธ ๋ฐ์ดํ„ฐ์…‹์—์„œ์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์„ค๋ช…ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ์ œ์•ˆํ•œ ๋ฐฉ๋ฒ•์€ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ๊ธฐ์„ฑ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์›น ์ฝ”ํผ์Šค์˜ ๋ฌธ์„œ๋ฅผ ๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ๋กœ ๋‹ค์‹œ ํ‘œํ˜„ํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•œ ๊ฐœ์š”๋Š” Figure 0(a)์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด ์ž‘์—…์—์„œ๋Š” ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜์„ ๋‘˜๋Ÿฌ์‹ผ ๋ชจํ˜ธ์„ฑ์—์„œ ๋น„๋กฏ๋œ ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ณผ์ œ(**i**) ๋ฌด์—‡์„ ๋ฏธ๋ฆฌ ํ•™์Šตํ•ด์•ผ ํ•˜๋Š”๊ฐ€? (ii) ์ œํ•œ๋œ ๋ฐ์ดํ„ฐ๋กœ ์–ด๋–ป๊ฒŒ ์‚ฌ์ „ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€? (iii) ์–ด๋–ป๊ฒŒ ํ•˜๋ฉด ๊ณ„์‚ฐ ํšจ์œจ์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€? ํŠนํžˆ, ๊ธฐ์„ฑ ์ค‘ํ˜• LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์›น์—์„œ ๋ฌธ์„œ๋ฅผ ๋‹ค์‹œ ํ‘œํ˜„ํ•˜๋ฉด ๋ชจ๋ธ์ด ์›น์˜ ์›์‹œ ํ…์ŠคํŠธ์—์„œ ํ•™์Šตํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ํ›จ์”ฌ ๋” ํšจ์œจ์ ์œผ๋กœ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ถ”๊ฐ€ ์›น ๋ฐ์ดํ„ฐ๋กœ ์ƒ์‡„ํ•  ์ˆ˜ ์—†๋Š” ๋ถ„์‚ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. ์ œ์•ˆํ•œ ๋ฐฉ๋ฒ•์€ ์‚ฌ์ „ ํ•™์Šต๋œ ๊ธฐ์„ฑ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์›น ๋ง๋ญ‰์น˜์˜ ๋ฌธ์„œ๋ฅผ ๋‹ค๋ฅธ ์Šคํƒ€์ผ๋กœ ๋‹ค์‹œ ํ‘œํ˜„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•œ ๊ฐœ์š”๋Š” ๊ทธ๋ฆผ 0(a)์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
We conducted a comparative analysis between our Adapt-LLM model and the current state-of-the-art approach for PopQA proposed by Mallen, Alex Troy and Asai, Akari and Zhong, Victor and Das, Rajarshi and Khashabi, Daniel and Hajishirzi, Hannaneh [22]. Their methodology relies on the popularity score annotated in the PopQA dataset to determine whether a question requires additional context. To establish the optimal threshold for determining question popularity, Mallen, Alex Troy and Asai, Akari and Zhong, Victor and Das, Rajarshi and Khashabi, Daniel and Hajishirzi, Hannaneh [22] split the PopQA dataset into 75% as a development set for threshold determination and 25% as a test set. In the original paper, they apply this methodology to various LLMs available at that moment (Llama-2 was not released yet).
์šฐ๋ฆฌ๋Š” Adapt-LLM ๋ชจ๋ธ๊ณผ Mallen, Alex Troy, Asai, Akari, Zhong, Victor, Das, Rajarshi, Khashabi, Daniel, Hajishirzi, Hannaneh [22]๊ฐ€ ์ œ์•ˆํ•œ PopQA์— ๋Œ€ํ•œ ์ตœ์‹  ์ ‘๊ทผ๋ฒ• ๊ฐ„์˜ ๋น„๊ต ๋ถ„์„์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋“ค์˜ ๋ฐฉ๋ฒ•๋ก ์€ PopQA ๋ฐ์ดํ„ฐ์…‹์— ์ฃผ์„๋œ ์ธ๊ธฐ๋„ ์ ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์งˆ๋ฌธ์ด ์ถ”๊ฐ€์ ์ธ ๋ฌธ๋งฅ์„ ํ•„์š”๋กœ ํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ ์ธ๊ธฐ๋„๋ฅผ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ ์˜ ์ž„๊ณ„๊ฐ’์„ ์„ค์ •ํ•˜๊ธฐ ์œ„ํ•ด, Mallen, Alex Troy, Asai, Akari, Zhong, Victor, Das, Rajarshi, Khashabi, Daniel, Hajishirzi, Hannaneh [22]๋Š” PopQA ๋ฐ์ดํ„ฐ์…‹์„ 75%๋Š” ์ž„๊ณ„๊ฐ’ ์„ค์ •์„ ์œ„ํ•œ ๊ฐœ๋ฐœ ์„ธํŠธ๋กœ, 25%๋Š” ํ…Œ์ŠคํŠธ ์„ธํŠธ๋กœ ๋‚˜๋ˆ„์—ˆ์Šต๋‹ˆ๋‹ค. ์› ๋…ผ๋ฌธ์—์„œ๋Š” ์ด ๋ฐฉ๋ฒ•๋ก ์„ ๋‹น์‹œ ์ด์šฉ ๊ฐ€๋Šฅํ•œ ๋‹ค์–‘ํ•œ LLM์— ์ ์šฉํ–ˆ์œผ๋ฉฐ (Llama-2๋Š” ์•„์ง ์ถœ์‹œ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค).
Mallen, Alex Troy, Asai, Akari, Zhong, Victor, Das, Rajarshi, Khashabi, Daniel, Hajishirzi, Hannaneh [22]๊ฐ€ ์ œ์•ˆํ•œ PopQA์˜ ํ˜„์žฌ ์ตœ์‹  ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ Adapt-LLM ๋ชจ๋ธ ๊ฐ„์˜ ๋น„๊ต ๋ถ„์„์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์˜ ๋ฐฉ๋ฒ•๋ก ์€ ์งˆ๋ฌธ์— ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๊ฐ€ ํ•„์š”ํ•œ์ง€ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด PopQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ฃผ์„์ด ๋‹ฌ๋ฆฐ ์ธ๊ธฐ๋„ ์ ์ˆ˜์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ ์ธ๊ธฐ๋„๋ฅผ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ ์˜ ์ž„๊ณ„๊ฐ’์„ ์„ค์ •ํ•˜๊ธฐ ์œ„ํ•ด Mallen, Alex Troy, Asai, Akari, Zhong, Victor, Das, Rajarshi, Khashabi, Daniel, Hajishirzi, Hannaneh [22]๋Š” PopQA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ 75%๋Š” ์ž„๊ณ„๊ฐ’ ๊ฒฐ์ •์„ ์œ„ํ•œ ๊ฐœ๋ฐœ ์„ธํŠธ๋กœ, 25%๋Š” ํ…Œ์ŠคํŠธ ์„ธํŠธ๋กœ ๋‚˜๋ˆ„์—ˆ์Šต๋‹ˆ๋‹ค. ์›๋ž˜ ๋…ผ๋ฌธ์—์„œ๋Š” ์ด ๋ฐฉ๋ฒ•๋ก ์„ ๋‹น์‹œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋‹ค์–‘ํ•œ LLM(Llama-2๋Š” ์•„์ง ์ถœ์‹œ๋˜์ง€ ์•Š์•˜์Œ)์— ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
The final observation is that the system in Lemma G.1 (\(x_{i}^{\top}\mathbf{A}x_{i}=1\) for all \(i\)) is linear in the \(h(h+1)/2\) distinct entries of \(\mathbf{A}\). By generating more than \(h(h+1)/2\) values for \(x_{i}\) (and again assuming no degeneracy), we can therefore solve a large system of linear equations, and in principle recover \(\mathbf{A}\), and hence \(\mathbf{M}\), and hence \(\mathbf{W}\). Note that any solution \(\mathbf{A}\) will be positive semidefinite because an overdetermined linear system has at most one solution. However, we do not know how to solve these systems of linear equations in \(h^{2}\) variables efficiently (\(h>750\) in all our experiments); so in practice we resort to reconstructing weights up to an arbitrary \(h\times h\) matrix, as described in Appendix C.
์ตœ์ข… ๊ด€์ฐฐ ๊ฒฐ๊ณผ, Lemma G.1์˜ ์‹œ์Šคํ…œ (\(x_{i}^{\top}\mathbf{A}x_{i}=1\) for all \(i\))์€ \(\mathbf{A}\)์˜ \(h(h+1)/2\) ๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ํ•ญ๋ชฉ์— ๋Œ€ํ•ด ์„ ํ˜•์ ์ž…๋‹ˆ๋‹ค. \(x_{i}\)์— ๋Œ€ํ•ด \(h(h+1)/2\) ์ด์ƒ์˜ ๊ฐ’์„ ์ƒ์„ฑํ•จ์œผ๋กœ์จ (๊ทธ๋ฆฌ๊ณ  ๋‹ค์‹œ ๋น„ํ‡ดํ™” ๊ฐ€์ •์„ ํ•˜๋ฉด), ์šฐ๋ฆฌ๋Š” ํฐ ์„ ํ˜• ๋ฐฉ์ •์‹ ์‹œ์Šคํ…œ์„ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์›์น™์ ์œผ๋กœ \(\mathbf{A}\), ๋”ฐ๋ผ์„œ \(\mathbf{M}\), ๊ทธ๋ฆฌ๊ณ  \(\mathbf{W}\)๋ฅผ ๋ณต์›ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณผ์ž‰ ๊ฒฐ์ •๋œ ์„ ํ˜• ์‹œ์Šคํ…œ์€ ์ตœ๋Œ€ ํ•˜๋‚˜์˜ ํ•ด๋ฅผ ๊ฐ€์ง€๋ฏ€๋กœ, ์–ด๋–ค ํ•ด \(\mathbf{A}\)๋„ ์–‘์˜ ์ค€์ •๋ถ€ํ˜ธ๊ฐ€ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ๋Š” \(h^{2}\) ๋ณ€์ˆ˜์˜ ์ด๋Ÿฌํ•œ ์„ ํ˜• ๋ฐฉ์ •์‹ ์‹œ์Šคํ…œ์„ ํšจ์œจ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์•Œ์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค (๋ชจ๋“  ์‹คํ—˜์—์„œ \(h>750\)). ๋”ฐ๋ผ์„œ ์‹ค์ œ๋กœ๋Š” ๋ถ€๋ก C์— ์„ค๋ช…๋œ ๋Œ€๋กœ ์ž„์˜์˜ \(h\times h\) ํ–‰๋ ฌ๊นŒ์ง€ ๊ฐ€์ค‘์น˜๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
๋งˆ์ง€๋ง‰ ๊ด€์ฐฐ์€ Lemman G.1์˜ ์‹œ์Šคํ…œ์—์„œ \(x_{i}^{\top}\mathbf{A}x_{i}=1\)์€ ๋ชจ๋“  \(i\)์— ๋Œ€ํ•ด ์„ ํ˜•์ด๋ผ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ \(h(h+1)/2\)๋ณด๋‹ค ๋งŽ์€ ๊ฐ’์„ \(x_{i}\)์— ๋Œ€ํ•ด ์ƒ์„ฑํ•˜๊ณ (๋‹ค์‹œ ํ•œ ๋ฒˆ ๋ณ€์„ฑ์ด ์—†๋‹ค๊ณ  ๊ฐ€์ •ํ•˜๋ฉด) ํฐ ์„ ํ˜• ๋ฐฉ์ •์‹ ์‹œ์Šคํ…œ์„ ํ’€ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์›์น™์ ์œผ๋กœ \(\mathbf{A}\), ๋”ฐ๋ผ์„œ \(\mathbf{M}\), ๋”ฐ๋ผ์„œ \(\mathbf{W}\)๋ฅผ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณผ๊ฒฐ์ •๋œ ์„ ํ˜• ์‹œ์Šคํ…œ์—๋Š” ์ตœ๋Œ€ ํ•˜๋‚˜์˜ ํ•ด๊ฐ€ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ชจ๋“  ํ•ด \(\mathbf{A}\)๋Š” ์–‘์˜ ๋ฐ˜์ •ํ™•๋„๊ฐ€ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์„ ํ˜• ๋ฐฉ์ •์‹์„ \(h^{2}\) ๋ณ€์ˆ˜๋กœ ํšจ์œจ์ ์œผ๋กœ ํ’€ ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ์•Œ์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค(\(h>750\)์€ ๋ชจ๋“  ์‹คํ—˜์—์„œ). ๋”ฐ๋ผ์„œ ์‹ค์ œ๋กœ๋Š” ๋ถ€๋ก C์—์„œ ์„ค๋ช…ํ•œ ๋Œ€๋กœ ์ž„์˜์˜ \(h\times h\) ํ–‰๋ ฌ๊นŒ์ง€ ๊ฐ€์ค‘์น˜๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
One can quickly imagine how large the search space is. Assuming the total number of layers across all models is \(M\), the size of the search space is \((M+1)^{T}\), here the extra one indicates the inclusion of a pass-through layer. Even with a modest setting of \(M=64\) (e.g., 2 models of 32 layers each) and \(T=60\), this setting translates to an astronomically large search space, a challenge even for a capable evolutionary search algorithm. Luckily, our preliminary studies indicated that certain layer arrangements, particularly repetitive or permuted sequences from earlier in the model, can adversely affect performance. Based on this, we modify our settings to include an indicator array \(\mathcal{I}\) of size \(T=M\times r\) in the evolutionary search space, here \(r\) is number of repetitions.
๊ฒ€์ƒ‰ ๊ณต๊ฐ„์ด ์–ผ๋งˆ๋‚˜ ํฐ์ง€ ์‰ฝ๊ฒŒ ์ƒ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์˜ ์ด ๋ ˆ์ด์–ด ์ˆ˜๋ฅผ \(M\)์œผ๋กœ ๊ฐ€์ •ํ•˜๋ฉด, ๊ฒ€์ƒ‰ ๊ณต๊ฐ„์˜ ํฌ๊ธฐ๋Š” \((M+1)^{T}\)์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์ถ”๊ฐ€๋œ 1์€ ํŒจ์Šค์Šค๋ฃจ ๋ ˆ์ด์–ด์˜ ํฌํ•จ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. \(M=64\) (์˜ˆ: ๊ฐ๊ฐ 32๊ฐœ์˜ ๋ ˆ์ด์–ด๋ฅผ ๊ฐ€์ง„ 2๊ฐœ์˜ ๋ชจ๋ธ) ๋ฐ \(T=60\)์ด๋ผ๋Š” ์ ๋‹นํ•œ ์„ค์ •์—์„œ๋„, ์ด ์„ค์ •์€ ๋งค์šฐ ํฐ ๊ฒ€์ƒ‰ ๊ณต๊ฐ„์œผ๋กœ ๋ณ€ํ™˜๋˜๋ฉฐ, ์ด๋Š” ๋Šฅ๋ ฅ ์žˆ๋Š” ์ง„ํ™”์  ๊ฒ€์ƒ‰ ์•Œ๊ณ ๋ฆฌ์ฆ˜์—๊ฒŒ๋„ ๋„์ „ ๊ณผ์ œ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ๋‹คํ–‰ํžˆ๋„, ์šฐ๋ฆฌ์˜ ์˜ˆ๋น„ ์—ฐ๊ตฌ๋Š” ํŠน์ • ๋ ˆ์ด์–ด ๋ฐฐ์—ด, ํŠนํžˆ ๋ชจ๋ธ ์ดˆ๋ฐ˜์˜ ๋ฐ˜๋ณต์ ์ด๊ฑฐ๋‚˜ ์ˆœ์—ด๋œ ์‹œํ€€์Šค๊ฐ€ ์„ฑ๋Šฅ์— ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ง„ํ™”์  ๊ฒ€์ƒ‰ ๊ณต๊ฐ„์— ํฌ๊ธฐ \(T=M\times r\)์˜ ์ง€์‹œ์ž ๋ฐฐ์—ด \(\mathcal{I}\)๋ฅผ ํฌํ•จํ•˜๋„๋ก ์„ค์ •์„ ์ˆ˜์ •ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ \(r\)์€ ๋ฐ˜๋ณต ํšŸ์ˆ˜์ž…๋‹ˆ๋‹ค.
๊ฒ€์ƒ‰ ๊ณต๊ฐ„์ด ์–ผ๋งˆ๋‚˜ ํฐ์ง€ ๊ธˆ๋ฐฉ ์ƒ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์˜ ์ด ๋ ˆ์ด์–ด ์ˆ˜๋ฅผ \(M\)์ด๋ผ๊ณ  ๊ฐ€์ •ํ•˜๋ฉด ๊ฒ€์ƒ‰ ๊ณต๊ฐ„์˜ ํฌ๊ธฐ๋Š” \((M+1)^{T}\)์ด๋ฉฐ, ์—ฌ๊ธฐ์„œ ํ•˜๋‚˜๋Š” ํŒจ์Šค ์Šค๋ฃจ ๋ ˆ์ด์–ด์˜ ํฌํ•จ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. M=64(์˜ˆ: ๊ฐ๊ฐ 32๊ฐœ์˜ ๋ ˆ์ด์–ด๋กœ ๊ตฌ์„ฑ๋œ 2๊ฐœ์˜ ๋ชจ๋ธ)์™€ T=60์˜ ์ ๋‹นํ•œ ์„ค์ •์„ ์‚ฌ์šฉํ•˜๋”๋ผ๋„ ์ด ์„ค์ •์€ ์—„์ฒญ๋‚˜๊ฒŒ ํฐ ๊ฒ€์ƒ‰ ๊ณต๊ฐ„์œผ๋กœ, ์œ ๋Šฅํ•œ ์ง„ํ™”ํ˜• ๊ฒ€์ƒ‰ ์•Œ๊ณ ๋ฆฌ์ฆ˜๋„ ๋„์ „์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹คํ–‰ํžˆ๋„ ์˜ˆ๋น„ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ํŠน์ • ๋ ˆ์ด์–ด ๋ฐฐ์—ด, ํŠนํžˆ ๋ชจ๋ธ ์ดˆ๋ฐ˜์˜ ๋ฐ˜๋ณต ๋˜๋Š” ์ˆœ์—ด๋œ ์‹œํ€€์Šค๊ฐ€ ์„ฑ๋Šฅ์— ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์ง„ํ™”ํ˜• ๊ฒ€์ƒ‰ ๊ณต๊ฐ„์— ํฌ๊ธฐ \(T=M\times r\)์˜ ์ธ๋””์ผ€์ดํ„ฐ ๋ฐฐ์—ด \(\mathcal{I}\)๋ฅผ ํฌํ•จํ•˜๋„๋ก ์„ค์ •์„ ์ˆ˜์ •ํ–ˆ์Šต๋‹ˆ๋‹ค(์—ฌ๊ธฐ์„œ \(r\)์€ ๋ฐ˜๋ณต ํšŸ์ˆ˜).
Furthermore, we find that deactivating language-specific neurons in the understanding layer influences the performance in English a little while significantly decreasing the performance in non-English languages. When deactivating language-specific neurons in the task-solving layer, both English and non-English languages are significantly reduced while deactivating language-specific neurons in the generation layer influences a little for both English and non-English languages. Therefore, we prove that the first several layers are responsible for understanding because deactivated neurons just disable LLMs on the NLU task in non-English languages. Furthermore, disabling language-specific neurons in the task-solving layer shows that LLMs rely on English, as their performance drops across all languages.
๊ฒŒ๋‹ค๊ฐ€, ์ดํ•ด ์ธต์—์„œ ์–ธ์–ด ํŠนํ™” ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด ์˜์–ด ์„ฑ๋Šฅ์—๋Š” ์•ฝ๊ฐ„์˜ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€๋งŒ ๋น„์˜์–ด๊ถŒ ์–ธ์–ด์˜ ์„ฑ๋Šฅ์€ ํฌ๊ฒŒ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ณผ์ œ ํ•ด๊ฒฐ ์ธต์—์„œ ์–ธ์–ด ํŠนํ™” ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด ์˜์–ด์™€ ๋น„์˜์–ด๊ถŒ ์–ธ์–ด ๋ชจ๋‘์—์„œ ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ๊ฐ์†Œํ•˜๋Š” ๋ฐ˜๋ฉด, ์ƒ์„ฑ ์ธต์—์„œ ์–ธ์–ด ํŠนํ™” ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด ์˜์–ด์™€ ๋น„์˜์–ด๊ถŒ ์–ธ์–ด ๋ชจ๋‘์— ์•ฝ๊ฐ„์˜ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ, ์ฒซ ๋ช‡ ๊ฐœ์˜ ์ธต์ด ์ดํ•ด๋ฅผ ๋‹ด๋‹นํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์ฆ๋ช…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์™œ๋ƒํ•˜๋ฉด ๋น„ํ™œ์„ฑํ™”๋œ ๋‰ด๋Ÿฐ์ด ๋น„์˜์–ด๊ถŒ ์–ธ์–ด์˜ NLU ๊ณผ์ œ์—์„œ LLM์„ ๋‹จ์ˆœํžˆ ๋น„ํ™œ์„ฑํ™”์‹œํ‚ค๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ, ๊ณผ์ œ ํ•ด๊ฒฐ ์ธต์—์„œ ์–ธ์–ด ํŠนํ™” ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋Š” ๊ฒƒ์€ LLM์ด ์˜์–ด์— ์˜์กดํ•˜๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์™œ๋ƒํ•˜๋ฉด ๋ชจ๋“  ์–ธ์–ด์—์„œ ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
๋˜ํ•œ ์ดํ•ด ๊ณ„์ธต์˜ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด ์˜์–ด ์„ฑ๋Šฅ์—๋Š” ์•ฝ๊ฐ„์˜ ์˜ํ–ฅ์„ ๋ฏธ์น˜๊ณ  ๋น„์˜์–ด ์–ธ์–ด ์„ฑ๋Šฅ์€ ํฌ๊ฒŒ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ณผ์ œ ํ•ด๊ฒฐ ๊ณ„์ธต์˜ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด ์˜์–ด์™€ ๋น„์˜์–ด ๋ชจ๋‘์—์„œ ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ๊ฐ์†Œํ•˜๊ณ , ์ƒ์„ฑ ๊ณ„์ธต์˜ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด ์˜์–ด์™€ ๋น„์˜์–ด ๋ชจ๋‘์—์„œ ์•ฝ๊ฐ„์˜ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ฒ˜์Œ ๋ช‡ ๊ฐœ์˜ ๋ ˆ์ด์–ด๋Š” ๋น„์˜์–ด ์–ธ์–ด์˜ NLU ์ž‘์—…์—์„œ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด LLM์ด ๋น„ํ™œ์„ฑํ™”๋˜๊ธฐ ๋•Œ๋ฌธ์— ์ดํ•ด๋ฅผ ๋‹ด๋‹นํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์ฆ๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ณผ์ œ ํ•ด๊ฒฐ ๊ณ„์ธต์—์„œ ์–ธ์–ด๋ณ„ ๋‰ด๋Ÿฐ์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋ฉด ๋ชจ๋“  ์–ธ์–ด์—์„œ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๋ฏ€๋กœ LLM์ด ์˜์–ด์— ์˜์กดํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
Pretraining data for large language models (LLMs) are typically a mixture of multiple domains, varying from English to minority languages (Doddapaneni et al., 2021; Li et al., 2023), from casual dialogs to formal academic writings (Taylor et al., 2022), and from texts to modalities like images and speeches (Zhan et al., 2024), among others. These data interplay with each other, showing complex interchangeable, unrelated, or contradictory relationships (Guo et al., 2024). This necessitates adjusting the mixture proportions of training data to balance the model capabilities while harnessing synergies across domains, thus enhancing the competence of the outcome models, as highlighted by extensive practical experience (Gao et al., 2020; Gururangan et al., 2020; Zhou et al., 2023; Xie et al., 2024).
๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธ(LLMs)์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์˜ ํ˜ผํ•ฉ๋ฌผ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ์˜์–ด์—์„œ ์†Œ์ˆ˜ ์–ธ์–ด(Doddapaneni et al., 2021; Li et al., 2023), ์ผ์ƒ ๋Œ€ํ™”์—์„œ ๊ณต์‹์ ์ธ ํ•™์ˆ  ๋…ผ๋ฌธ(Taylor et al., 2022), ํ…์ŠคํŠธ์—์„œ ์ด๋ฏธ์ง€์™€ ์Œ์„ฑ(Zhan et al., 2024)๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋Š” ์„œ๋กœ ์ƒํ˜ธ์ž‘์šฉํ•˜๋ฉฐ ๋ณต์žกํ•œ ์ƒํ˜ธ ๊ตํ™˜ ๊ฐ€๋Šฅ์„ฑ, ๋ฌด๊ด€์„ฑ ๋˜๋Š” ๋ชจ์ˆœ์ ์ธ ๊ด€๊ณ„๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค(Guo et al., 2024). ์ด๋Š” ๋„๋ฉ”์ธ ๊ฐ„์˜ ์‹œ๋„ˆ์ง€๋ฅผ ํ™œ์šฉํ•˜๋ฉด์„œ ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ๊ท ํ˜• ์žˆ๊ฒŒ ์กฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ํ˜ผํ•ฉ ๋น„์œจ์„ ์กฐ์ •ํ•ด์•ผ ํ•จ์„ ํ•„์š”๋กœ ํ•˜๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๊ฒฐ๊ณผ ๋ชจ๋ธ์˜ ์—ญ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Gao et al., 2020; Gururangan et al., 2020; Zhou et al., 2023; Xie et al., 2024).
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์˜์–ด์—์„œ ์†Œ์ˆ˜ ์–ธ์–ด(Doddapaneni et al., 2021; Li et al., 2023)์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ, ์บ์ฃผ์–ผ ๋Œ€ํ™”์—์„œ ๊ณต์‹์ ์ธ ํ•™์ˆ  ๊ธ€์“ฐ๊ธฐ(Taylor et al., 2022)์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์˜ ํ˜ผํ•ฉ๋ฌผ์ž…๋‹ˆ๋‹ค, ํ…์ŠคํŠธ์—์„œ ์ด๋ฏธ์ง€ ๋ฐ ์—ฐ์„ค๊ณผ ๊ฐ™์€ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ(Zhan et al., 2024)์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋Š” ์„œ๋กœ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋ฉฐ ๋ณต์žกํ•˜๊ฒŒ ์ƒํ˜ธ ๊ตํ™˜, ๊ด€๋ จ์ด ์—†๊ฑฐ๋‚˜ ๋ชจ์ˆœ๋˜๋Š” ๊ด€๊ณ„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(Guo et al., 2024). ๋”ฐ๋ผ์„œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹ค๋ฌด ๊ฒฝํ—˜์—์„œ ๊ฐ•์กฐํ–ˆ๋“ฏ์ด ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํ˜ผํ•ฉ ๋น„์œจ์„ ์กฐ์ •ํ•˜์—ฌ ๋ชจ๋ธ ๊ธฐ๋Šฅ์˜ ๊ท ํ˜•์„ ๋งž์ถ”๊ณ  ๋„๋ฉ”์ธ ๊ฐ„์˜ ์‹œ๋„ˆ์ง€๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๊ฒฐ๊ณผ ๋ชจ๋ธ์˜ ์—ญ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ฌ ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค(Gao et al., 2020; Gururangan et al., 2020; Zhou et al., 2023; Xie et al., 2024).
Measuring Acquisition of New World KnowledgeWe define _new_ world knowledge as the information present in \(D_{1}\), but not in \(D_{0}\). To measure _new_ knowledge acquired through continued pretraining on \(D_{1}\), we construct NewLAMA which is made up of detailed cloze statements requiring _new_ knowledge from \(D_{1}\) to correctly answer. We provide two datasets for measuring _new world knowledge_: NewLAMA, for which each of the instances is verified that the answer does not exist in \(D_{0}\), but only in \(D_{1}\), and NewLAMA-Easy for which each of the instances does not perfectly comply with our strict definition of _new_ world knowledge due to its creation process, but is used to generally measure the new knowledge acquired from continued pretraining on \(D_{1}\) at a larger scale.
์ƒˆ๋กœ์šด ์„ธ๊ณ„ ์ง€์‹์˜ ํš๋“ ์ธก์ • ์šฐ๋ฆฌ๋Š” ์ƒˆ๋กœ์šด ์„ธ๊ณ„ ์ง€์‹์„ \(D_{1}\)์—๋Š” ์กด์žฌํ•˜์ง€๋งŒ \(D_{0}\)์—๋Š” ์กด์žฌํ•˜์ง€ ์•Š๋Š” ์ •๋ณด๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. \(D_{1}\)์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ํš๋“ํ•œ ์ƒˆ๋กœ์šด ์ง€์‹์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” \(D_{1}\)์—์„œ ์š”๊ตฌ๋˜๋Š” ์ƒˆ๋กœ์šด ์ง€์‹์„ ์ •ํ™•ํžˆ ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด ํ•„์š”ํ•œ ์ƒ์„ธํ•œ ํด๋กœ์ฆˆ ๋ฌธ์žฅ๋“ค๋กœ ๊ตฌ์„ฑ๋œ NewLAMA๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ƒˆ๋กœ์šด ์„ธ๊ณ„ ์ง€์‹์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค: ๊ฐ ์ธ์Šคํ„ด์Šค๊ฐ€ \(D_{0}\)์—๋Š” ์กด์žฌํ•˜์ง€ ์•Š๊ณ  \(D_{1}\)์—๋งŒ ์กด์žฌํ•˜๋Š” ๋‹ต์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ๊ฒ€์ฆํ•œ NewLAMA์™€, ์ƒ์„ฑ ๊ณผ์ •์œผ๋กœ ์ธํ•ด ์šฐ๋ฆฌ์˜ ์—„๊ฒฉํ•œ ์ƒˆ๋กœ์šด ์„ธ๊ณ„ ์ง€์‹ ์ •์˜์— ์™„๋ฒฝํžˆ ๋ถ€ํ•ฉํ•˜์ง€๋Š” ์•Š์ง€๋งŒ, \(D_{1}\)์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ํš๋“ํ•œ ์ƒˆ๋กœ์šด ์ง€์‹์„ ๋” ํฐ ๊ทœ๋ชจ๋กœ ์ผ๋ฐ˜์ ์œผ๋กœ ์ธก์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” NewLAMA-Easy์ž…๋‹ˆ๋‹ค.
์ƒˆ๋กœ์šด ์„ธ๊ณ„ ์ง€์‹์˜ ์Šต๋“ ์ธก์ •์ƒˆ๋กœ์šด ์„ธ๊ณ„ ์ง€์‹์€ \(D_{1}\)์— ์กด์žฌํ•˜์ง€๋งŒ \(D_{0}\)์—๋Š” ์กด์žฌํ•˜์ง€ ์•Š๋Š” ์ •๋ณด๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. D_{1}\)์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ์Šต๋“ํ•œ _์ƒˆ๋กœ์šด_ ์ง€์‹์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด, ์ •๋‹ต์„ ๋งžํžˆ๊ธฐ ์œ„ํ•ด \(D_{1}\)์˜ _์ƒˆ๋กœ์šด_ ์ง€์‹์ด ํ•„์š”ํ•œ ์„ธ๋ถ€์ ์ธ ํด๋กœ์ฆˆ ๋ฌธ์žฅ์œผ๋กœ ๊ตฌ์„ฑ๋œ NewLAMA๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” _์ƒˆ๋กœ์šด ์„ธ๊ณ„ ์ง€์‹_์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค: ๊ฐ ์ธ์Šคํ„ด์Šค๊ฐ€ \(D_{0}\)์—๋Š” ๋‹ต์ด ์กด์žฌํ•˜์ง€ ์•Š์ง€๋งŒ \(D_{1}\)์—๋งŒ ์กด์žฌํ•œ๋‹ค๋Š” ๊ฒƒ์ด ๊ฒ€์ฆ๋œ NewLAMA์™€, ์ƒ์„ฑ ๊ณผ์ •์œผ๋กœ ์ธํ•ด ๊ฐ ์ธ์Šคํ„ด์Šค๊ฐ€ _์ƒˆ๋กœ์šด_ ์„ธ๊ณ„ ์ง€์‹์— ๋Œ€ํ•œ ์—„๊ฒฉํ•œ ์ •์˜๋ฅผ ์™„๋ฒฝํ•˜๊ฒŒ ์ค€์ˆ˜ํ•˜์ง€๋Š” ์•Š์ง€๋งŒ, ์ผ๋ฐ˜์ ์œผ๋กœ \(D_{1}\)์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ์Šต๋“ํ•œ ์ƒˆ๋กœ์šด ์ง€์‹์„ ๋” ํฐ ๊ทœ๋ชจ๋กœ ์ธก์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” NewLAMA-Easy๊ฐ€ ๊ทธ๊ฒƒ์ž…๋‹ˆ๋‹ค.
The AIOS SDK is designed to equip developers with a versatile toolkit for crafting sophisticated agent applications within the AIOS. This SDK encompasses a broad spectrum of functionalities, from initializing agents and managing agent lifecycles to facilitating complex operations like resource monitoring, and generation plan for an agent task. Like any operating system, enriching the SDK to be comprehensive and developer-friendly is a long-term and never-ending endeavor. The current SDK functions supported in AIOS are shown in Table 3, which will be continuously updated and expanded to fulfill the needs of evolving agent applications. This development effort aims to provide developers with the tools they need to harness the full potential of their agent applications within the AIOS framework.
AIOS SDK๋Š” AIOS ๋‚ด์—์„œ ์ •๊ตํ•œ ์—์ด์ „ํŠธ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ์ œ์ž‘ํ•  ์ˆ˜ ์žˆ๋„๋ก ๊ฐœ๋ฐœ์ž์—๊ฒŒ ๋‹ค์šฉ๋„์˜ ๋„๊ตฌ ๋ชจ์Œ์„ ์ œ๊ณตํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด SDK๋Š” ์—์ด์ „ํŠธ ์ดˆ๊ธฐํ™” ๋ฐ ์—์ด์ „ํŠธ ์ƒ๋ช… ์ฃผ๊ธฐ ๊ด€๋ฆฌ๋ถ€ํ„ฐ ์ž์› ๋ชจ๋‹ˆํ„ฐ๋ง ๋ฐ ์—์ด์ „ํŠธ ์ž‘์—…์„ ์œ„ํ•œ ์ƒ์„ฑ ๊ณ„ํš๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ง€์›ํ•˜๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ๊ธฐ๋Šฅ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ์šด์˜ ์ฒด์ œ์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, SDK๋ฅผ ํฌ๊ด„์ ์ด๊ณ  ๊ฐœ๋ฐœ์ž ์นœํ™”์ ์œผ๋กœ ๋งŒ๋“œ๋Š” ๊ฒƒ์€ ์žฅ๊ธฐ์ ์ด๊ณ  ๋์—†๋Š” ๋…ธ๋ ฅ์ž…๋‹ˆ๋‹ค. AIOS์—์„œ ์ง€์›ํ•˜๋Š” ํ˜„์žฌ SDK ๊ธฐ๋Šฅ์€ ํ‘œ 3์— ๋‚˜์™€ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ง„ํ™”ํ•˜๋Š” ์—์ด์ „ํŠธ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ์š”๊ตฌ๋ฅผ ์ถฉ์กฑํ•˜๊ธฐ ์œ„ํ•ด ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋˜๊ณ  ํ™•์žฅ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๊ฐœ๋ฐœ ๋…ธ๋ ฅ์€ ๊ฐœ๋ฐœ์ž๊ฐ€ AIOS ํ”„๋ ˆ์ž„์›Œํฌ ๋‚ด์—์„œ ์—์ด์ „ํŠธ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ์ž ์žฌ๋ ฅ์„ ์ตœ๋Œ€ํ•œ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋„๊ตฌ๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
AIOS SDK๋Š” ๊ฐœ๋ฐœ์ž๊ฐ€ AIOS ๋‚ด์—์„œ ์ •๊ตํ•œ ์—์ด์ „ํŠธ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ์ œ์ž‘ํ•  ์ˆ˜ ์žˆ๋Š” ๋‹ค์šฉ๋„ ํˆดํ‚ท์„ ๊ฐ–์ถ”๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด SDK๋Š” ์—์ด์ „ํŠธ ์ดˆ๊ธฐํ™” ๋ฐ ์—์ด์ „ํŠธ ์ˆ˜๋ช… ์ฃผ๊ธฐ ๊ด€๋ฆฌ๋ถ€ํ„ฐ ๋ฆฌ์†Œ์Šค ๋ชจ๋‹ˆํ„ฐ๋ง ๋ฐ ์—์ด์ „ํŠธ ์ž‘์—… ์ƒ์„ฑ ๊ณ„ํš๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ์ž‘์—…์„ ์šฉ์ดํ•˜๊ฒŒ ํ•˜๋Š” ๊ฒƒ๊นŒ์ง€ ๊ด‘๋ฒ”์œ„ํ•œ ๊ธฐ๋Šฅ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์šด์˜ ์ฒด์ œ์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ SDK๋ฅผ ํฌ๊ด„์ ์ด๊ณ  ๊ฐœ๋ฐœ์ž ์นœํ™”์ ์œผ๋กœ ๋งŒ๋“ค๊ธฐ ์œ„ํ•œ ์ž‘์—…์€ ์žฅ๊ธฐ์ ์ด๊ณ  ๋์—†๋Š” ๋…ธ๋ ฅ์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ AIOS์—์„œ ์ง€์›๋˜๋Š” SDK ๊ธฐ๋Šฅ์€ ํ‘œ 3๊ณผ ๊ฐ™์œผ๋ฉฐ, ์ด๋Š” ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ ๋ฐ ํ™•์žฅ๋˜์–ด ์ง„ํ™”ํ•˜๋Š” ์—์ด์ „ํŠธ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ์š”๊ตฌ ์‚ฌํ•ญ์„ ์ถฉ์กฑํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๊ฐœ๋ฐœ ๋…ธ๋ ฅ์€ ๊ฐœ๋ฐœ์ž๊ฐ€ AIOS ํ”„๋ ˆ์ž„์›Œํฌ ๋‚ด์—์„œ ์—์ด์ „ํŠธ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ์ž ์žฌ๋ ฅ์„ ์ตœ๋Œ€ํ•œ ํ™œ์šฉํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ๋„๊ตฌ๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
**Method 1 vs. Method 2**. We attempt to assess which one is more suitable for providing rewards for RL. For a fair comparison, we use the same evaluation data that was used in Method 2. Since Method 1 has 6 aspect categories, we obtain individual results using the best-performing approach and group them into two as we did for Method 2. Table 5 reports the accuracy, macro precision, macro recall, and macro F1 for Method 1 and Method 2. The results indicate that, although models from Method 2 are trained on a relatively small dataset, Method 2 is more aligned with human judgments compared to Method 1. Additionally, inference time using Method 2 is significantly faster than Method 1, making it easier to retrieve rewards in the RL setting. As a result, we use Method 2 for all our experiments.
**Method 1 vs. Method 2**. ์šฐ๋ฆฌ๋Š” ์–ด๋Š ๋ฐฉ๋ฒ•์ด RL์— ๋Œ€ํ•œ ๋ณด์ƒ์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ๋” ์ ํ•ฉํ•œ์ง€ ํ‰๊ฐ€ํ•˜๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด Method 2์—์„œ ์‚ฌ์šฉ๋œ ๋™์ผํ•œ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Method 1์€ 6๊ฐœ์˜ ์ธก๋ฉด ์นดํ…Œ๊ณ ๋ฆฌ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์—, ์ตœ์ƒ์˜ ์„ฑ๊ณผ๋ฅผ ๋‚ด๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ๋ณ„ ๊ฒฐ๊ณผ๋ฅผ ์–ป๊ณ  ์ด๋ฅผ Method 2์—์„œ ํ–ˆ๋˜ ๊ฒƒ์ฒ˜๋Ÿผ ๋‘ ๊ทธ๋ฃน์œผ๋กœ ๋‚˜๋ˆ•๋‹ˆ๋‹ค. Table 5๋Š” Method 1๊ณผ Method 2์˜ ์ •ํ™•๋„, ๋งคํฌ๋กœ ์ •๋ฐ€๋„, ๋งคํฌ๋กœ ์žฌํ˜„์œจ, ๋งคํฌ๋กœ F1์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ๋Š” Method 2์˜ ๋ชจ๋ธ์ด ์ƒ๋Œ€์ ์œผ๋กœ ์ž‘์€ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , Method 1์— ๋น„ํ•ด ์ธ๊ฐ„์˜ ํŒ๋‹จ๊ณผ ๋” ์ผ์น˜ํ•จ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ, Method 2๋ฅผ ์‚ฌ์šฉํ•œ ์ถ”๋ก  ์‹œ๊ฐ„์ด Method 1๋ณด๋‹ค ํ›จ์”ฌ ๋นจ๋ผ RL ์„ค์ •์—์„œ ๋ณด์ƒ์„ ๋” ์‰ฝ๊ฒŒ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ชจ๋“  ์‹คํ—˜์—์„œ Method 2๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
**๋ฐฉ๋ฒ• 1 ๋Œ€ ๋ฐฉ๋ฒ• 2**. ์–ด๋–ค ๋ฐฉ๋ฒ•์ด RL์— ๋Œ€ํ•œ ๋ณด์ƒ์„ ์ œ๊ณตํ•˜๋Š” ๋ฐ ๋” ์ ํ•ฉํ•œ์ง€ ํ‰๊ฐ€ํ•˜๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด ๋ฐฉ๋ฒ• 2์—์„œ ์‚ฌ์šฉํ•œ ๊ฒƒ๊ณผ ๋™์ผํ•œ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ฐฉ๋ฒ• 1์—๋Š” 6๊ฐœ์˜ ์ธก๋ฉด ์นดํ…Œ๊ณ ๋ฆฌ๊ฐ€ ์žˆ์œผ๋ฏ€๋กœ ์ตœ๊ณ  ์„ฑ๋Šฅ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ๋ณ„ ๊ฒฐ๊ณผ๋ฅผ ์–ป๊ณ  ๋ฐฉ๋ฒ• 2์—์„œ์™€ ๊ฐ™์ด ๋‘ ๊ฐ€์ง€๋กœ ๊ทธ๋ฃนํ™”ํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 5๋Š” ๋ฐฉ๋ฒ• 1๊ณผ ๋ฐฉ๋ฒ• 2์— ๋Œ€ํ•œ ์ •ํ™•๋„, ๋งคํฌ๋กœ ์ •๋ฐ€๋„, ๋งคํฌ๋กœ ๋ฆฌ์ฝœ ๋ฐ ๋งคํฌ๋กœ F1์„ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ๋ฐฉ๋ฒ• 2์˜ ๋ชจ๋ธ์€ ์ƒ๋Œ€์ ์œผ๋กœ ์ž‘์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ•™์Šต๋˜์—ˆ์ง€๋งŒ ๋ฐฉ๋ฒ• 1์— ๋น„ํ•ด ๋ฐฉ๋ฒ• 2๊ฐ€ ์‚ฌ๋žŒ์˜ ํŒ๋‹จ์— ๋” ๋ถ€ํ•ฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฐฉ๋ฒ• 2๋ฅผ ์‚ฌ์šฉํ•œ ์ถ”๋ก  ์‹œ๊ฐ„์ด ๋ฐฉ๋ฒ• 1๋ณด๋‹ค ํ›จ์”ฌ ๋น ๋ฅด๋ฏ€๋กœ RL ์„ค์ •์—์„œ ๋ณด์ƒ์„ ๋” ์‰ฝ๊ฒŒ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ชจ๋“  ์‹คํ—˜์— ๋ฐฉ๋ฒ• 2๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
The generation phase is tasked with producing text that is both relevant to the query and reflective of the information found in the retrieved documents. The usual method involves concatenating the query with the retrieved information, which is then fed into an LLM for text generation (Li et al., 2022). Although ensuring the generated text's alignment and accuracy with the retrieved content presents challenges, it is also essential to strike a balance between adhering closely to the source material and infusing the output with creativity. The generated text should accurately convey the information from the retrieved documents and align with the query's intent, while also offering the flexibility to introduce new insights or perspectives not explicitly contained within the retrieved data.
์ƒ์„ฑ ๋‹จ๊ณ„๋Š” ์ฟผ๋ฆฌ์™€ ๊ด€๋ จ์ด ์žˆ์œผ๋ฉฐ ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ์—์„œ ๋ฐœ๊ฒฌ๋œ ์ •๋ณด๋ฅผ ๋ฐ˜์˜ํ•˜๋Š” ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ž‘์—…์„ ๋‹ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ฐฉ๋ฒ•์€ ์ฟผ๋ฆฌ์™€ ๊ฒ€์ƒ‰๋œ ์ •๋ณด๋ฅผ ์—ฐ๊ฒฐํ•˜์—ฌ LLM์— ์ž…๋ ฅํ•œ ํ›„ ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค (Li et al., 2022). ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ๊ฐ€ ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ์™€์˜ ์ผ์น˜์„ฑ๊ณผ ์ •ํ™•์„ฑ์„ ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์€ ๋„์ „ ๊ณผ์ œ๋ฅผ ์ œ์‹œํ•˜์ง€๋งŒ, ์›๋ณธ ์ž๋ฃŒ์— ์ถฉ์‹คํ•˜๋ฉด์„œ๋„ ์ฐฝ์˜์„ฑ์„ ์ฃผ์ž…ํ•˜๋Š” ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ๊ฒƒ๋„ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ๋Š” ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ์˜ ์ •๋ณด๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์ „๋‹ฌํ•˜๊ณ  ์ฟผ๋ฆฌ์˜ ์˜๋„์™€ ์ผ์น˜ํ•ด์•ผ ํ•˜๋ฉฐ, ๊ฒ€์ƒ‰๋œ ๋ฐ์ดํ„ฐ์— ๋ช…์‹œ์ ์œผ๋กœ ํฌํ•จ๋˜์ง€ ์•Š์€ ์ƒˆ๋กœ์šด ํ†ต์ฐฐ๋ ฅ์ด๋‚˜ ๊ด€์ ์„ ๋„์ž…ํ•  ์ˆ˜ ์žˆ๋Š” ์œ ์—ฐ์„ฑ๋„ ์ œ๊ณตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
์ƒ์„ฑ ๋‹จ๊ณ„์—์„œ๋Š” ์ฟผ๋ฆฌ์™€ ๊ด€๋ จ์ด ์žˆ๊ณ  ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ์—์„œ ์ฐพ์€ ์ •๋ณด๋ฅผ ๋ฐ˜์˜ํ•˜๋Š” ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ž„๋ฌด๋ฅผ ๋งก์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ฐฉ๋ฒ•์€ ์ฟผ๋ฆฌ์™€ ๊ฒ€์ƒ‰๋œ ์ •๋ณด๋ฅผ ์—ฐ๊ฒฐํ•œ ๋‹ค์Œ ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ์œ„ํ•ด LLM์— ๊ณต๊ธ‰ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(Li et al., 2022). ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ์™€ ๊ฒ€์ƒ‰๋œ ์ฝ˜ํ…์ธ ์˜ ์ •๋ ฌ์„ฑ๊ณผ ์ •ํ™•์„ฑ์„ ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ค์šด ๊ณผ์ œ์ด์ง€๋งŒ, ์ถœ์ฒ˜ ์ž๋ฃŒ๋ฅผ ๋ฉด๋ฐ€ํžˆ ์ค€์ˆ˜ํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฒฐ๊ณผ๋ฌผ์— ์ฐฝ์˜์„ฑ์„ ๋ถˆ์–ด๋„ฃ๋Š” ๊ฒƒ ์‚ฌ์ด์—์„œ ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ๊ฒƒ๋„ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ๋Š” ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ์˜ ์ •๋ณด๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์ „๋‹ฌํ•˜๊ณ  ์ฟผ๋ฆฌ์˜ ์˜๋„์— ๋ถ€ํ•ฉํ•ด์•ผ ํ•˜๋ฉฐ, ๊ฒ€์ƒ‰๋œ ๋ฐ์ดํ„ฐ์— ๋ช…์‹œ์ ์œผ๋กœ ํฌํ•จ๋˜์ง€ ์•Š์€ ์ƒˆ๋กœ์šด ์ธ์‚ฌ์ดํŠธ๋‚˜ ๊ด€์ ์„ ๋„์ž…ํ•  ์ˆ˜ ์žˆ๋Š” ์œ ์—ฐ์„ฑ๋„ ์ œ๊ณตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
The rapid advancement of artificial intelligence (AI) technologies has led to their widespread adoption across numerous domains, from assistant agents (e.g., ACT-1, from Adept AI1) and software development (e.g., Devin, from Cognition Lab2) to healthcare (Singhal et al., 2022) and finance (Zheng et al., 2022). However, the success of AI models heavily relies on the availability of large, diverse, and high-quality datasets for training and evaluation. Acquiring such datasets can be a significant challenge due to data scarcity (Babbar and Scholkopf, 2019), privacy concerns (Abay et al., 2019), and the sheer cost of data collection and annotation (Gilardi et al., 2023). Pessimists predict that we will run out of fresh text data in 2050 and image data in 2060 (Villalobos et al., 2022).
์ธ๊ณต์ง€๋Šฅ(AI) ๊ธฐ์ˆ ์˜ ๊ธ‰์†ํ•œ ๋ฐœ์ „์€ ๋ณด์กฐ ์—์ด์ „ํŠธ(e.g., Adept AI1์˜ ACT-1)์™€ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ(e.g., Cognition Lab2์˜ Devin)์—์„œ๋ถ€ํ„ฐ ์˜๋ฃŒ(Singhal et al., 2022) ๋ฐ ๊ธˆ์œต(Zheng et al., 2022)์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์—์„œ ๋„๋ฆฌ ์ฑ„ํƒ๋˜๊ฒŒ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ AI ๋ชจ๋ธ์˜ ์„ฑ๊ณต์€ ๋Œ€๊ทœ๋ชจ, ๋‹ค์–‘ํ•œ, ๊ณ ํ’ˆ์งˆ์˜ ๋ฐ์ดํ„ฐ์…‹์„ ํ›ˆ๋ จ ๋ฐ ํ‰๊ฐ€์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š”์ง€์— ํฌ๊ฒŒ ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ์…‹์„ ํ™•๋ณดํ•˜๋Š” ๊ฒƒ์€ ๋ฐ์ดํ„ฐ ๋ถ€์กฑ(Babbar and Scholkopf, 2019), ํ”„๋ผ์ด๋ฒ„์‹œ ๋ฌธ์ œ(Abay et al., 2019), ๊ทธ๋ฆฌ๊ณ  ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๋ฐ ์ฃผ์„ ์ž‘์—…์˜ ๋ง‰๋Œ€ํ•œ ๋น„์šฉ(Gilardi et al., 2023) ๋•Œ๋ฌธ์— ์ƒ๋‹นํ•œ ๋„์ „ ๊ณผ์ œ๊ฐ€ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋น„๊ด€๋ก ์ž๋“ค์€ 2050๋…„์—๋Š” ์ƒˆ๋กœ์šด ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๊ฐ€, 2060๋…„์—๋Š” ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๊ฐ€ ๊ณ ๊ฐˆ๋  ๊ฒƒ์ด๋ผ๊ณ  ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค(Villalobos et al., 2022).
์ธ๊ณต์ง€๋Šฅ(AI) ๊ธฐ์ˆ ์˜ ๊ธ‰์†ํ•œ ๋ฐœ์ „์œผ๋กœ ์ธํ•ด ์–ด์‹œ์Šคํ„ดํŠธ ์—์ด์ „ํŠธ(์˜ˆ: Adept AI1์˜ ACT-1), ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ(์˜ˆ: Cognition Lab2์˜ Devin)๋ถ€ํ„ฐ ์˜๋ฃŒ(Singhal et al., 2022), ๊ธˆ์œต(Zheng et al., 2022) ๋“ฑ ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์—์„œ AI ๊ธฐ์ˆ ์ด ๋„๋ฆฌ ํ™œ์šฉ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ AI ๋ชจ๋ธ์˜ ์„ฑ๊ณต ์—ฌ๋ถ€๋Š” ํ›ˆ๋ จ๊ณผ ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ๋Œ€๊ทœ๋ชจ์˜ ๋‹ค์–‘ํ•˜๊ณ  ๊ณ ํ’ˆ์งˆ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ€์šฉ์„ฑ์— ํฌ๊ฒŒ ์˜์กดํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ™•๋ณดํ•˜๋Š” ๊ฒƒ์€ ๋ฐ์ดํ„ฐ ๋ถ€์กฑ(Babbar and Scholkopf, 2019), ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ๋ฌธ์ œ(Abay et al., 2019), ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๋ฐ ์ฃผ์„ ์ž‘์—…์— ๋“œ๋Š” ๋ง‰๋Œ€ํ•œ ๋น„์šฉ(Gilardi et al., 2023)์œผ๋กœ ์ธํ•ด ์ƒ๋‹นํ•œ ์–ด๋ ค์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋น„๊ด€๋ก ์ž๋“ค์€ 2050๋…„์—๋Š” ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๊ฐ€, 2060๋…„์—๋Š” ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๊ฐ€ ์†Œ์ง„๋  ๊ฒƒ์ด๋ผ๊ณ  ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค(Villalobos et al., 2022).
During the large language model training stage, a wide range of failures regularly occur, ranging from GPU crashes to communication fabric errors to loss spikes. We use the following strategies to address these reliability challenges: (1) we apply automated inspection, prediction, and labeling of nodes for different kind of software/hardware error categories. Nodes marked as tainted will be temporarily removed from the resource pool until the errors got cleared. (2) we implement a task queuing system with pre-checks and the capability for fast, automatic recovery in the event of failures during training tasks. (3) we develop of a user-friendly multi-task submission and management console, enabling developers to seamlessly manage and track their training tasks and hyper-parameters.
๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธ ํ›ˆ๋ จ ๋‹จ๊ณ„์—์„œ๋Š” GPU ์ถฉ๋Œ, ํ†ต์‹  ํŒจ๋ธŒ๋ฆญ ์˜ค๋ฅ˜, ์†์‹ค ๊ธ‰์ฆ ๋“ฑ ๋‹ค์–‘ํ•œ ์‹คํŒจ๊ฐ€ ์ •๊ธฐ์ ์œผ๋กœ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹ ๋ขฐ์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ „๋žต์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค: (1) ๋‹ค์–‘ํ•œ ์†Œํ”„ํŠธ์›จ์–ด/ํ•˜๋“œ์›จ์–ด ์˜ค๋ฅ˜ ๋ฒ”์ฃผ์— ๋Œ€ํ•ด ๋…ธ๋“œ์˜ ์ž๋™ ๊ฒ€์‚ฌ, ์˜ˆ์ธก ๋ฐ ๋ผ๋ฒจ๋ง์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ค๋ฅ˜๊ฐ€ ํ•ด๊ฒฐ๋  ๋•Œ๊นŒ์ง€ ์˜ค์—ผ๋œ ๊ฒƒ์œผ๋กœ ํ‘œ์‹œ๋œ ๋…ธ๋“œ๋Š” ์ž์› ํ’€์—์„œ ์ผ์‹œ์ ์œผ๋กœ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. (2) ํ›ˆ๋ จ ์ž‘์—… ์ค‘ ์‹คํŒจ๊ฐ€ ๋ฐœ์ƒํ•  ๊ฒฝ์šฐ ๋น ๋ฅด๊ณ  ์ž๋™์œผ๋กœ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ์‚ฌ์ „ ์ ๊ฒ€ ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ˜ ์ž‘์—… ๋Œ€๊ธฐ์—ด ์‹œ์Šคํ…œ์„ ๊ตฌํ˜„ํ•ฉ๋‹ˆ๋‹ค. (3) ๊ฐœ๋ฐœ์ž๊ฐ€ ํ›ˆ๋ จ ์ž‘์—…๊ณผ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์›ํ™œํ•˜๊ฒŒ ๊ด€๋ฆฌํ•˜๊ณ  ์ถ”์ ํ•  ์ˆ˜ ์žˆ๋„๋ก ์‚ฌ์šฉ์ž ์นœํ™”์ ์ธ ๋‹ค์ค‘ ์ž‘์—… ์ œ์ถœ ๋ฐ ๊ด€๋ฆฌ ์ฝ˜์†”์„ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ ํ•™์Šต ๋‹จ๊ณ„์—์„œ๋Š” GPU ์ถฉ๋Œ๋ถ€ํ„ฐ ํ†ต์‹  ํŒจ๋ธŒ๋ฆญ ์˜ค๋ฅ˜, ์†์‹ค ์ŠคํŒŒ์ดํฌ์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์ข…๋ฅ˜์˜ ์žฅ์• ๊ฐ€ ์ •๊ธฐ์ ์œผ๋กœ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์•ˆ์ •์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ „๋žต์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค: (1) ๋‹ค์–‘ํ•œ ์ข…๋ฅ˜์˜ ์†Œํ”„ํŠธ์›จ์–ด/ํ•˜๋“œ์›จ์–ด ์˜ค๋ฅ˜ ์นดํ…Œ๊ณ ๋ฆฌ์— ๋Œ€ํ•œ ๋…ธ๋“œ์˜ ์ž๋™ํ™”๋œ ๊ฒ€์‚ฌ, ์˜ˆ์ธก ๋ฐ ๋ผ๋ฒจ๋ง์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ค์—ผ๋œ ๊ฒƒ์œผ๋กœ ํ‘œ์‹œ๋œ ๋…ธ๋“œ๋Š” ์˜ค๋ฅ˜๊ฐ€ ํ•ด๊ฒฐ๋  ๋•Œ๊นŒ์ง€ ๋ฆฌ์†Œ์Šค ํ’€์—์„œ ์ผ์‹œ์ ์œผ๋กœ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. (2) ์‚ฌ์ „ ๊ฒ€์‚ฌ ๊ธฐ๋Šฅ์ด ์žˆ๋Š” ์ž‘์—… ํ ์‹œ์Šคํ…œ์„ ๊ตฌํ˜„ํ•˜๊ณ  ํ•™์Šต ์ž‘์—… ์ค‘ ์žฅ์•  ๋ฐœ์ƒ ์‹œ ๋น ๋ฅด๊ณ  ์ž๋™์œผ๋กœ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์„ ๊ตฌํ˜„ํ•ฉ๋‹ˆ๋‹ค. (3) ๊ฐœ๋ฐœ์ž๊ฐ€ ํ•™์Šต ์ž‘์—…๊ณผ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์›ํ™œํ•˜๊ฒŒ ๊ด€๋ฆฌํ•˜๊ณ  ์ถ”์ ํ•  ์ˆ˜ ์žˆ๋Š” ์‚ฌ์šฉ์ž ์นœํ™”์ ์ธ ๋ฉ€ํ‹ฐํƒœ์Šคํฌ ์ œ์ถœ ๋ฐ ๊ด€๋ฆฌ ์ฝ˜์†”์„ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค.
REDCODER (Parvez et al., 2021) initially identified potential candidate codes from existing code or abstract databases. The researchers retained 1.1 million unique lines of codes and abstracts as retrieved data through multiple channels. The final evaluation on the CodeXGLUE dataset demonstrated excellent performance across multiple programming languages and evaluation metrics. Another proposed enhancement was put forth by Zan et al. (2022), who utilized private libraries to enhance the quality of code generation. This involved first identifying the most suitable private libraries based on the API documentation through the APIRetriever component, and then utilizing the APIcoder for generation. This approach led to a notable improvement in the accuracy of the generated content.
REDCODER (Parvez et al., 2021)์€ ๊ธฐ์กด ์ฝ”๋“œ๋‚˜ ์ถ”์ƒ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์—์„œ ์ž ์žฌ์ ์ธ ํ›„๋ณด ์ฝ”๋“œ๋ฅผ ์ฒ˜์Œ์œผ๋กœ ์‹๋ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ž๋“ค์€ ์—ฌ๋Ÿฌ ์ฑ„๋„์„ ํ†ตํ•ด 110๋งŒ ๊ฐœ์˜ ๊ณ ์œ ํ•œ ์ฝ”๋“œ ๋ผ์ธ๊ณผ ์ดˆ๋ก์„ ๊ฒ€์ƒ‰๋œ ๋ฐ์ดํ„ฐ๋กœ ์œ ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค. CodeXGLUE ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•œ ์ตœ์ข… ํ‰๊ฐ€์—์„œ๋Š” ์—ฌ๋Ÿฌ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์™€ ํ‰๊ฐ€ ์ง€ํ‘œ์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ œ์•ˆ๋œ ๊ฐœ์„  ์‚ฌํ•ญ์€ Zan et al. (2022)์— ์˜ํ•ด ์ œ์‹œ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋“ค์€ ์ฝ”๋“œ ์ƒ์„ฑ์˜ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์‚ฌ์„ค ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” APIRetriever ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ํ†ตํ•ด API ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ์‚ฌ์„ค ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ๋จผ์ € ์‹๋ณ„ํ•œ ๋‹ค์Œ, APIcoder๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑํ•˜๋Š” ๊ณผ์ •์„ ํฌํ•จํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ƒ์„ฑ๋œ ์ฝ˜ํ…์ธ ์˜ ์ •ํ™•์„ฑ์„ ํ˜„์ €ํžˆ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค.
REDCODER(Parvez et al., 2021)๋Š” ๊ธฐ์กด ์ฝ”๋“œ ๋˜๋Š” ์ถ”์ƒ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์—์„œ ์ž ์žฌ์ ์ธ ํ›„๋ณด ์ฝ”๋“œ๋ฅผ ์ฒ˜์Œ ์‹๋ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ง„์€ ์—ฌ๋Ÿฌ ์ฑ„๋„์„ ํ†ตํ•ด ๊ฒ€์ƒ‰๋œ ๋ฐ์ดํ„ฐ๋กœ 110๋งŒ ๊ฐœ์˜ ๊ณ ์œ ํ•œ ์ฝ”๋“œ์™€ ์ดˆ๋ก์„ ๋ณด์œ ํ–ˆ์Šต๋‹ˆ๋‹ค. CodeXGLUE ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ตœ์ข… ํ‰๊ฐ€๋Š” ์—ฌ๋Ÿฌ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์™€ ํ‰๊ฐ€ ์ง€ํ‘œ์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ œ์•ˆ๋œ ๊ฐœ์„  ์‚ฌํ•ญ์€ Zan ๋“ฑ(2022)์ด ์ œ์•ˆํ•œ ๊ฒƒ์œผ๋กœ, ์ฝ”๋“œ ์ƒ์„ฑ์˜ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๋น„๊ณต๊ฐœ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ํ™œ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๋จผ์ € APIRetriever ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ํ†ตํ•ด API ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ๋น„๊ณต๊ฐœ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์‹๋ณ„ํ•œ ๋‹ค์Œ ์ƒ์„ฑ์— APIcoder๋ฅผ ํ™œ์šฉํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ง„ํ–‰๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ƒ์„ฑ๋œ ์ฝ˜ํ…์ธ ์˜ ์ •ํ™•๋„๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_In-context learning_. The in-context learning (ICL) ability is formally introduced by GPT-3 [55]: assuming that the language model has been provided with a natural language instruction and/or several task demonstrations, it can generate the expected output for the test instances by completing the word sequence of input text, without requiring additional training or gradient update9. Among the GPT-series models, the 175B GPT-3 model exhibited a strong ICL ability in general, but not the GPT-1 and GPT-2 models. Such an ability also depends on the specific downstream task. For example, the ICL ability can emerge on the arithmetic tasks (_e.g._, the 3-digit addition and subtraction) for the 13B GPT-3, but 175B GPT-3 even cannot work well on the Persian QA task [31].
โ€ข _In-context learning_. In-context learning (ICL) ๋Šฅ๋ ฅ์€ GPT-3 [55]์— ์˜ํ•ด ๊ณต์‹์ ์œผ๋กœ ๋„์ž…๋˜์—ˆ์Šต๋‹ˆ๋‹ค: ์–ธ์–ด ๋ชจ๋ธ์ด ์ž์—ฐ์–ด ์ง€์‹œ๋ฌธ ๋ฐ/๋˜๋Š” ์—ฌ๋Ÿฌ ์ž‘์—… ์‹œ์—ฐ์„ ์ œ๊ณต๋ฐ›์•˜๋‹ค๊ณ  ๊ฐ€์ •ํ•  ๋•Œ, ์ถ”๊ฐ€์ ์ธ ํ›ˆ๋ จ์ด๋‚˜ ๊ธฐ์šธ๊ธฐ ์—…๋ฐ์ดํŠธ ์—†์ด ์ž…๋ ฅ ํ…์ŠคํŠธ์˜ ๋‹จ์–ด ์‹œํ€€์Šค๋ฅผ ์™„์„ฑํ•˜์—ฌ ํ…Œ์ŠคํŠธ ์ธ์Šคํ„ด์Šค์— ๋Œ€ํ•œ ์˜ˆ์ƒ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. GPT ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ ์ค‘์—์„œ 175B GPT-3 ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐ•๋ ฅํ•œ ICL ๋Šฅ๋ ฅ์„ ๋ณด์˜€์ง€๋งŒ, GPT-1 ๋ฐ GPT-2 ๋ชจ๋ธ์€ ๊ทธ๋ ‡์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์€ ํŠน์ • ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 13B GPT-3๋Š” ์‚ฐ์ˆ  ์ž‘์—…(์˜ˆ: 3์ž๋ฆฌ ์ˆ˜์˜ ๋ง์…ˆ ๋ฐ ๋บ„์…ˆ)์—์„œ ICL ๋Šฅ๋ ฅ์ด ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ์ง€๋งŒ, 175B GPT-3๋Š” ํŽ˜๋ฅด์‹œ์•„์–ด QA ์ž‘์—…์—์„œ๋Š” ์ž˜ ์ž‘๋™ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [31].
\(\bullet\)_์ปจํ…์ŠคํŠธ ๋‚ด ํ•™์Šต_. ๋ฌธ๋งฅ ๋‚ด ํ•™์Šต(ICL) ๊ธฐ๋Šฅ์€ GPT-3[55]์—์„œ ๊ณต์‹์ ์œผ๋กœ ์†Œ๊ฐœํ•œ ๊ฒƒ์œผ๋กœ, ์–ธ์–ด ๋ชจ๋ธ์— ์ž์—ฐ์–ด ๋ช…๋ น์–ด ๋ฐ/๋˜๋Š” ์—ฌ๋Ÿฌ ์ž‘์—… ๋ฐ๋ชจ๊ฐ€ ์ œ๊ณต๋œ๋‹ค๊ณ  ๊ฐ€์ •ํ•  ๋•Œ ์ถ”๊ฐ€ ํ•™์Šต์ด๋‚˜ ๊ทธ๋ผ๋””์–ธํŠธ ์—…๋ฐ์ดํŠธ ์—†์ด๋„ ์ž…๋ ฅ ํ…์ŠคํŠธ์˜ ๋‹จ์–ด ์‹œํ€€์Šค๋ฅผ ์™„์„ฑํ•˜์—ฌ ํ…Œ์ŠคํŠธ ์ธ์Šคํ„ด์Šค์— ๋Œ€ํ•ด ์˜ˆ์ƒ๋˜๋Š” ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. GPT ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ ์ค‘ 175B GPT-3 ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐ•๋ ฅํ•œ ICL ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์ง€๋งŒ GPT-1 ๋ฐ GPT-2 ๋ชจ๋ธ์€ ๊ทธ๋ ‡์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์€ ํŠน์ • ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์‚ฐ์ˆ  ์ž‘์—…(์˜ˆ: 3์ž๋ฆฌ ๋ง์…ˆ ๋ฐ ๋บ„์…ˆ)์˜ ๊ฒฝ์šฐ 13B GPT-3์—์„œ ICL ๊ธฐ๋Šฅ์ด ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ์ง€๋งŒ, ํŽ˜๋ฅด์‹œ์•„์–ด QA ์ž‘์—…[31]์˜ ๊ฒฝ์šฐ 175B GPT-3์กฐ์ฐจ ์ œ๋Œ€๋กœ ์ž‘๋™ํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค.
SEC Filing [3.3B words, 16.5%]Public companies in the United States are legally required to submit their financial statements on a regular basis. The Securities and Exchange Commission (SEC) facilitates public access to these filings through the Electronic Data Gathering, Analysis, and Retrieval (EDGAR) System, which has been available since 1993. On average, this system accommodates approximately 40,000 new files per year. To enrich our financial corpus, we include 10-K filings from the period spanning 1993 to 2022. To ensure data accuracy and consistency, these filings are parsed and pre-processed using the package detailed in [17]. Furthermore, we optimize the quality of our corpus by eliminating report sections containing less than 20 words, to remove spurious examples.
SEC ์ œ์ถœ [3.3B ๋‹จ์–ด, 16.5%] ๋ฏธ๊ตญ์˜ ์ƒ์žฅ ๊ธฐ์—…๋“ค์€ ์ •๊ธฐ์ ์œผ๋กœ ์žฌ๋ฌด์ œํ‘œ๋ฅผ ์ œ์ถœํ•ด์•ผ ํ•˜๋Š” ๋ฒ•์  ์˜๋ฌด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฆ๊ถŒ๊ฑฐ๋ž˜์œ„์›ํšŒ(SEC)๋Š” 1993๋…„๋ถ€ํ„ฐ ์ด์šฉ ๊ฐ€๋Šฅํ•œ ์ „์ž ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘, ๋ถ„์„ ๋ฐ ๊ฒ€์ƒ‰(EDGAR) ์‹œ์Šคํ…œ์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ์ œ์ถœ ๋ฌธ์„œ์— ๋Œ€ํ•œ ๊ณต๊ณต ์ ‘๊ทผ์„ ์šฉ์ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ท ์ ์œผ๋กœ ์ด ์‹œ์Šคํ…œ์€ ์—ฐ๊ฐ„ ์•ฝ 40,000๊ฐœ์˜ ์ƒˆ๋กœ์šด ํŒŒ์ผ์„ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๊ธˆ์œต ์ฝ”ํผ์Šค๋ฅผ ํ’๋ถ€ํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” 1993๋…„๋ถ€ํ„ฐ 2022๋…„๊นŒ์ง€์˜ 10-K ์ œ์ถœ ๋ฌธ์„œ๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ์˜ ์ •ํ™•์„ฑ๊ณผ ์ผ๊ด€์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด, ์ด๋Ÿฌํ•œ ์ œ์ถœ ๋ฌธ์„œ๋Š” [17]์—์„œ ์ž์„ธํžˆ ์„ค๋ช…๋œ ํŒจํ‚ค์ง€๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ตฌ๋ฌธ ๋ถ„์„ ๋ฐ ์ „์ฒ˜๋ฆฌ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, 20๋‹จ์–ด ๋ฏธ๋งŒ์˜ ๋ณด๊ณ ์„œ ์„น์…˜์„ ์ œ๊ฑฐํ•˜์—ฌ ๋ถˆํ•„์š”ํ•œ ์˜ˆ์ œ๋ฅผ ์ œ๊ฑฐํ•จ์œผ๋กœ์จ ์ฝ”ํผ์Šค์˜ ํ’ˆ์งˆ์„ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค.
SEC ์ œ์ถœ [3.3B ๋‹จ์–ด, 16.5%]๋ฏธ๊ตญ์˜ ์ƒ์žฅ ๊ธฐ์—…์€ ๋ฒ•์ ์œผ๋กœ ์ •๊ธฐ์ ์œผ๋กœ ์žฌ๋ฌด์ œํ‘œ๋ฅผ ์ œ์ถœํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ฆ๊ถŒ๊ฑฐ๋ž˜์œ„์›ํšŒ(SEC)๋Š” 1993๋…„๋ถ€ํ„ฐ ์ œ๊ณต๋˜๊ณ  ์žˆ๋Š” ์ „์ž ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘, ๋ถ„์„ ๋ฐ ๊ฒ€์ƒ‰(EDGAR) ์‹œ์Šคํ…œ์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ์ œ์ถœ์— ๋Œ€ํ•œ ๋Œ€์ค‘์˜ ์ ‘๊ทผ์„ ์šฉ์ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์‹œ์Šคํ…œ์€ ํ‰๊ท ์ ์œผ๋กœ ์—ฐ๊ฐ„ ์•ฝ 40,000๊ฐœ์˜ ์ƒˆ๋กœ์šด ํŒŒ์ผ์„ ์ˆ˜์šฉํ•ฉ๋‹ˆ๋‹ค. ์žฌ๋ฌด ์ฝ”ํผ์Šค๋ฅผ ํ’์„ฑํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด 1993๋…„๋ถ€ํ„ฐ 2022๋…„๊นŒ์ง€์˜ 10-K ์ œ์ถœ ์ž๋ฃŒ๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ์˜ ์ •ํ™•์„ฑ๊ณผ ์ผ๊ด€์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด [17]์— ์ž์„ธํžˆ ์„ค๋ช…๋œ ํŒจํ‚ค์ง€๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋Ÿฌํ•œ ์ œ์ถœ ์ž๋ฃŒ๋ฅผ ๊ตฌ๋ฌธ ๋ถ„์„ํ•˜๊ณ  ์‚ฌ์ „ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ 20๋‹จ์–ด ๋ฏธ๋งŒ์˜ ๋ณด๊ณ ์„œ ์„น์…˜์„ ์ œ๊ฑฐํ•˜์—ฌ ์Šคํ“จ๋ฆฌ์–ด์Šค ์˜ˆ์ œ๋ฅผ ์ œ๊ฑฐํ•˜์—ฌ ์ฝ”ํผ์Šค์˜ ํ’ˆ์งˆ์„ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค.
Question: What do you think of velvet? Answer: Whether you like your velvet crushed, vibrant or head-to-toe, there's really no denying the sheer luxe and elegance of this timeless textile. Question: Is velvet super stylish? Answer: Yes, really! This year it's all about embracing fun gem-toned velvety pieces. Question: Was velvet solely associated with dark moody shades of navy and black? Answer: Long gone are the days when velvet was solely associated with dark moody shades of navy and black. Question: Do you have any covetable velvet pieces on the high street right now? Answer: Below we've rounded up the most covetable velvet pieces on the high street right now. Question: Are you completely obsessed or beyond bored of it? Answer: We're already coming up with outfit ideas!
์งˆ๋ฌธ: ๋ฒจ๋ฒณ์— ๋Œ€ํ•ด ์–ด๋–ป๊ฒŒ ์ƒ๊ฐํ•˜์‹œ๋‚˜์š”? ๋‹ต๋ณ€: ๋ฒจ๋ฒณ์ด ๊ตฌ๊ฒจ์ง€๋“ , ์ƒ๋™๊ฐ ์žˆ๋“ , ๋จธ๋ฆฌ๋ถ€ํ„ฐ ๋ฐœ๋๊นŒ์ง€ ์ž…๋“ , ์ด ์‹œ๋Œ€๋ฅผ ์ดˆ์›”ํ•œ ์ง๋ฌผ์˜ ์ˆœ์ˆ˜ํ•œ ๋Ÿญ์…”๋ฆฌ์™€ ์šฐ์•„ํ•จ์„ ๋ถ€์ •ํ•  ์ˆ˜๋Š” ์—†์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋ฒจ๋ฒณ์ด ๋งค์šฐ ์Šคํƒ€์ผ๋ฆฌ์‹œํ•œ๊ฐ€์š”? ๋‹ต๋ณ€: ๋„ค, ์ •๋ง ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค! ์˜ฌํ•ด๋Š” ๋ณด์„ ํ†ค์˜ ๋ฒจ๋ฒณ ์กฐ๊ฐ์„ ์ฆ๊ธฐ๋Š” ๊ฒƒ์ด ์ „๋ถ€์ž…๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋ฒจ๋ฒณ์ด ์–ด๋‘์šด ํ•ด๊ตฐ์ƒ‰๊ณผ ๊ฒ€์ •์ƒ‰์˜ ์Œ์šธํ•œ ์ƒ‰์กฐ์™€๋งŒ ์—ฐ๊ด€๋˜์—ˆ๋‚˜์š”? ๋‹ต๋ณ€: ๋ฒจ๋ฒณ์ด ์–ด๋‘์šด ํ•ด๊ตฐ์ƒ‰๊ณผ ๊ฒ€์ •์ƒ‰์˜ ์Œ์šธํ•œ ์ƒ‰์กฐ์™€๋งŒ ์—ฐ๊ด€๋˜์—ˆ๋˜ ์‹œ์ ˆ์€ ์˜ค๋ž˜์ „์— ์ง€๋‚˜๊ฐ”์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์ง€๊ธˆ ๋‹น์žฅ ๊ณ ๊ธ‰ ์ƒ์ ์—์„œ ํƒ๋‚˜๋Š” ๋ฒจ๋ฒณ ์กฐ๊ฐ์ด ์žˆ๋‚˜์š”? ๋‹ต๋ณ€: ์•„๋ž˜์— ์ง€๊ธˆ ๋‹น์žฅ ๊ณ ๊ธ‰ ์ƒ์ ์—์„œ ๊ฐ€์žฅ ํƒ๋‚˜๋Š” ๋ฒจ๋ฒณ ์กฐ๊ฐ๋“ค์„ ๋ชจ์•„๋ดค์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋ฒจ๋ฒณ์— ์™„์ „ํžˆ ์ง‘์ฐฉํ•˜๊ณ  ์žˆ๋‚˜์š”, ์•„๋‹ˆ๋ฉด ์ง€๋ฃจํ•ดํ•˜๊ณ  ์žˆ๋‚˜์š”? ๋‹ต๋ณ€: ์šฐ๋ฆฌ๋Š” ์ด๋ฏธ ์ƒˆ๋กœ์šด ์˜ท์ฐจ๋ฆผ ์•„์ด๋””์–ด๋ฅผ ์ƒ๊ฐํ•ด๋‚ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค!
์งˆ๋ฌธ: ๋ฒจ๋ฒณ์— ๋Œ€ํ•ด ์–ด๋–ป๊ฒŒ ์ƒ๊ฐํ•˜์‹œ๋‚˜์š”? ๋‹ต๋ณ€: ๋ฒจ๋ฒณ์„ ์œผ์Šค๋Ÿฌ๋œจ๋ฆฌ๋“ , ์ƒ์ƒํ•˜๊ฒŒ ์ž…๋“ , ๋จธ๋ฆฌ๋ถ€ํ„ฐ ๋ฐœ๋๊นŒ์ง€ ์ž…๋“ , ์ด ์‹œ๋Œ€๋ฅผ ์ดˆ์›”ํ•œ ์„ฌ์œ ์˜ ๊ณ ๊ธ‰์Šค๋Ÿฌ์›€๊ณผ ์šฐ์•„ํ•จ์„ ๋ถ€์ •ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋ฒจ๋ฒณ์€ ์ •๋ง ์Šคํƒ€์ผ๋ฆฌ์‹œํ•œ๊ฐ€์š”? ๋‹ต๋ณ€: ๋„ค, ์ •๋ง์ด์—์š”! ์˜ฌํ•ด๋Š” ์žฌ๋ฏธ์žˆ๋Š” ๋ณด์„ ํ†ค์˜ ๋ฒจ๋ฒณ ์˜ท์„ ์ž…๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ๋ฒจ๋ฒณ์€ ์–ด๋‘์šด ๋ฌด๋“œ์˜ ๋„ค์ด๋น„์™€ ๋ธ”๋ž™ ์ƒ‰์กฐ์™€๋งŒ ์—ฐ๊ด€๋˜์–ด ์žˆ์—ˆ๋‚˜์š”? ๋‹ต๋ณ€: ๋ฒจ๋ฒณ์ด ์–ด๋‘ก๊ณ  ๋ฌด๋“œ ์žˆ๋Š” ๋„ค์ด๋น„์™€ ๋ธ”๋ž™ ์ƒ‰์กฐ์™€๋งŒ ์—ฐ๊ด€๋˜์—ˆ๋˜ ์‹œ๋Œ€๋Š” ์‚ฌ๋ผ์ง„ ์ง€ ์˜ค๋ž˜์ž…๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์ง€๊ธˆ ๊ณ ๊ฐ€์˜ ๋ฒจ๋ฒณ ์ œํ’ˆ์ด ์žˆ๋‚˜์š”? ๋‹ต๋ณ€: ์•„๋ž˜์—์„œ๋Š” ํ˜„์žฌ ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ๋ฒจ๋ฒณ ์ œํ’ˆ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ: ์™„์ „ํžˆ ์ง‘์ฐฉํ•˜์‹œ๋‚˜์š”, ์•„๋‹ˆ๋ฉด ์ง€๋ฃจํ•ดํ•˜์‹œ๋‚˜์š”? ๋‹ต๋ณ€: ์ด๋ฏธ ์˜์ƒ ์•„์ด๋””์–ด๊ฐ€ ๋– ์˜ค๋ฅด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค!
Furthermore, we propose efficient domain-adaptive continual pre-training methods, ETS-DACP and ETA-DACP to enhance the effectiveness of the continual pre-training. By being selective with the training data curation, our methods refine the continual pre-training, yielding even better results with just 10% of the data and cost of vanilla continual pre-training. ETA-DACP with data selection based on task-agnostic measures like entropy works almost at par with the task-aware data selection strategy. This finding can be used to build data selection for continual pre-training even in the absence of task data. We also observe no degradation in performance on open-domain standard tasks, implying that domain-adaptive continual pre-training does not hurt open-domain capabilities.
๊ฒŒ๋‹ค๊ฐ€, ์šฐ๋ฆฌ๋Š” ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์˜ ํšจ๊ณผ๋ฅผ ๋†’์ด๊ธฐ ์œ„ํ•ด ํšจ์œจ์ ์ธ ๋„๋ฉ”์ธ ์ ์‘ ์ง€์† ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์ธ ETS-DACP์™€ ETA-DACP๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜์„ ์„ ํƒ์ ์œผ๋กœ ํ•จ์œผ๋กœ์จ, ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์€ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์ •์ œํ•˜์—ฌ, ์ผ๋ฐ˜์ ์ธ ์ง€์† ์‚ฌ์ „ ํ›ˆ๋ จ์˜ ๋ฐ์ดํ„ฐ์™€ ๋น„์šฉ์˜ 10%๋งŒ์œผ๋กœ๋„ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค. ์—”ํŠธ๋กœํ”ผ์™€ ๊ฐ™์€ ์ž‘์—… ๋น„ํŠน์ด์  ์ธก์ •์— ๊ธฐ๋ฐ˜ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ์„ ์‚ฌ์šฉํ•˜๋Š” ETA-DACP๋Š” ์ž‘์—… ์ธ์‹ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต๊ณผ ๊ฑฐ์˜ ๋™๋“ฑํ•˜๊ฒŒ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐœ๊ฒฌ์€ ์ž‘์—… ๋ฐ์ดํ„ฐ๊ฐ€ ์—†๋Š” ๊ฒฝ์šฐ์—๋„ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์šฐ๋ฆฌ๋Š” ์˜คํ”ˆ ๋„๋ฉ”์ธ ํ‘œ์ค€ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ ์ €ํ•˜๊ฐ€ ์—†์Œ์„ ๊ด€์ฐฐํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ๋„๋ฉ”์ธ ์ ์‘ ์ง€์† ์‚ฌ์ „ ํ›ˆ๋ จ์ด ์˜คํ”ˆ ๋„๋ฉ”์ธ ๊ธฐ๋Šฅ์„ ํ•ด์น˜์ง€ ์•Š์Œ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
๋˜ํ•œ, ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์˜ ํšจ์œจ์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ํšจ์œจ์ ์ธ ๋„๋ฉ”์ธ ์ ์‘ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต ๋ฐฉ๋ฒ•์ธ ETS-DACP์™€ ETA-DACP๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜์„ ์„ ํƒ์ ์œผ๋กœ ์ ์šฉํ•จ์œผ๋กœ์จ ๊ธฐ์กด ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์˜ ๋ฐ์ดํ„ฐ์™€ ๋น„์šฉ์˜ 10%๋งŒ์œผ๋กœ๋„ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—”ํŠธ๋กœํ”ผ์™€ ๊ฐ™์€ ์ž‘์—…์— ๊ตฌ์• ๋ฐ›์ง€ ์•Š๋Š” ์ธก์ •์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ETA-DACP๋Š” ์ž‘์—… ์ธ์‹ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต๊ณผ ๊ฑฐ์˜ ๋™๋“ฑํ•œ ์ˆ˜์ค€์œผ๋กœ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ์ž‘์—… ๋ฐ์ดํ„ฐ๊ฐ€ ์—†๋Š” ๊ฒฝ์šฐ์—๋„ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์˜คํ”ˆ ๋„๋ฉ”์ธ ํ‘œ์ค€ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ ์ €ํ•˜๊ฐ€ ์—†์–ด ๋„๋ฉ”์ธ ์ ์‘ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์ด ์˜คํ”ˆ ๋„๋ฉ”์ธ ๊ธฐ๋Šฅ์— ํ•ด๋ฅผ ๋ผ์น˜์ง€ ์•Š๋Š”๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.
Another limitation of this work is the loose nature of the mathematical connection that we establish between the dimensionality of the ideal language modeling head and the rank-constrained performance (cf. Theorem 5.3). Moreover, it can also be argued that considering _ideal_\(x_{i}^{*}\) representations is an ill-defined notion. We argue that the reasoning behind Theorem 5.3 could be applied to any contextual representations, as the _ideal_ nature of \(x_{i}^{*}\) is not necessary in the demonstrations. The word _ideal_ reflects that our observations hold for \(x_{i}^{*}\) representations obtained from _any underlying model_, to an extent that depends on the structure that these representations impose on the \(W^{*}\) matrix for a given training set \(\mathcal{T}\).
์ด ์—ฐ๊ตฌ์˜ ๋˜ ๋‹ค๋ฅธ ํ•œ๊ณ„๋Š” ์ด์ƒ์ ์ธ ์–ธ์–ด ๋ชจ๋ธ๋ง ํ—ค๋“œ์˜ ์ฐจ์›์„ฑ๊ณผ ์ˆœ์œ„ ์ œํ•œ ์„ฑ๋Šฅ ๊ฐ„์˜ ์ˆ˜ํ•™์  ์—ฐ๊ฒฐ์ด ๋Š์Šจํ•˜๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค(์ฐธ์กฐ: ์ •๋ฆฌ 5.3). ๋˜ํ•œ, _์ด์ƒ์ ์ธ_ \(x_{i}^{*}\) ํ‘œํ˜„์„ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์ด ์ž˜ ์ •์˜๋˜์ง€ ์•Š์€ ๊ฐœ๋…์ด๋ผ๋Š” ์ฃผ์žฅ๋„ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ •๋ฆฌ 5.3์˜ ๋…ผ๋ฆฌ๊ฐ€ _์ด์ƒ์ ์ธ_ \(x_{i}^{*}\)์˜ ํŠน์„ฑ์ด ์‹œ์—ฐ์— ํ•„์š”ํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ์–ด๋–ค ๋งฅ๋ฝ์  ํ‘œํ˜„์—๋„ ์ ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. _์ด์ƒ์ ์ธ_ ์ด๋ผ๋Š” ๋‹จ์–ด๋Š” ์šฐ๋ฆฌ์˜ ๊ด€์ฐฐ์ด ์ฃผ์–ด์ง„ ํ›ˆ๋ จ ์„ธํŠธ \(\mathcal{T}\)์— ๋Œ€ํ•ด ์ด๋Ÿฌํ•œ ํ‘œํ˜„์ด \(W^{*}\) ํ–‰๋ ฌ์— ๋ถ€์—ฌํ•˜๋Š” ๊ตฌ์กฐ์— ๋”ฐ๋ผ _์–ด๋–ค ๊ธฐ๋ณธ ๋ชจ๋ธ_์—์„œ ์–ป์€ \(x_{i}^{*}\) ํ‘œํ˜„์— ๋Œ€ํ•ด ์„ฑ๋ฆฝํ•จ์„ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค.
์ด ์ž‘์—…์˜ ๋˜ ๋‹ค๋ฅธ ํ•œ๊ณ„๋Š” ์ด์ƒ์ ์ธ ์–ธ์–ด ๋ชจ๋ธ๋ง ํ—ค๋“œ์˜ ์ฐจ์›๊ณผ ์ˆœ์œ„ ์ œํ•œ ์„ฑ๋Šฅ ๊ฐ„์— ์„ค์ •ํ•œ ์ˆ˜ํ•™์  ์—ฐ๊ด€์„ฑ์ด ๋Š์Šจํ•˜๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค(cf. ์ •๋ฆฌ 5.3). ๋˜ํ•œ _์ด์ƒ_\(x_{i}^{*}\) ํ‘œํ˜„์„ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์€ ์ž˜ ์ •์˜๋˜์ง€ ์•Š์€ ๊ฐœ๋…์ด๋ผ๊ณ  ์ฃผ์žฅํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ •๋ฆฌ 5.3์˜ ์ถ”๋ก ์ด ๋ฐ๋ชจ์—์„œ _์ด์ƒ์ _ ํŠน์„ฑ์ด ํ•„์š”ํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ๋ชจ๋“  ์ปจํ…์ŠคํŠธ ํ‘œํ˜„์— ์ ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. '์ด์ƒ์ '์ด๋ผ๋Š” ๋‹จ์–ด๋Š” ์ฃผ์–ด์ง„ ํ›ˆ๋ จ ์„ธํŠธ \mathcal{T}\)์— ๋Œ€ํ•ด ์ด๋Ÿฌํ•œ ํ‘œํ˜„์ด \(W^{*}\) ํ–‰๋ ฌ์— ๋ถ€๊ณผํ•˜๋Š” ๊ตฌ์กฐ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๋Š” ์ •๋„๊นŒ์ง€, _๋ชจ๋“  ๊ธฐ๋ณธ ๋ชจ๋ธ_์—์„œ ์–ป์€ \(x_{i}^{*}\) ํ‘œํ˜„์— ๋Œ€ํ•ด ์šฐ๋ฆฌ์˜ ๊ด€์ฐฐ์ด ์œ ์ง€๋จ์„ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค.
**SoTA Models** In this category, our benchmarks are established against, to the best of our knowledge, the strongest publicly available translation models. We first compare with **ALMA-13B-LoRA**, recognized as one of the top moderate-size language-model based translation systems, surpassing notable conventional models such as NLLB-54B in both WMT'21 and WMT'22. We also compare our results with **TowerInstruct5**, a recently released LLM-based translation model and a contemporary work in the field.6 Additionally, we evaluate against the zero-shot performance of the latest **GPT-4** (gpt-4-1106-preview), currently shown to be the best translation model among all LLM-based translation systems (Xu et al., 2023; Zhang et al., 2023; Zeng et al., 2023; Jiao et al., 2023).
**SoTA Models** ์ด ๋ฒ”์ฃผ์—์„œ, ์šฐ๋ฆฌ์˜ ๋ฒค์น˜๋งˆํฌ๋Š” ์šฐ๋ฆฌ๊ฐ€ ์•Œ๊ณ  ์žˆ๋Š” ํ•œ ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ๊ณต๊ฐœ ๋ฒˆ์—ญ ๋ชจ๋ธ๋“ค๊ณผ ๋น„๊ต๋ฉ๋‹ˆ๋‹ค. ๋จผ์ €, WMT'21๊ณผ WMT'22์—์„œ NLLB-54B์™€ ๊ฐ™์€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ๊ธฐ์กด ๋ชจ๋ธ๋“ค์„ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ค‘๊ฐ„ ํฌ๊ธฐ ์–ธ์–ด ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ๋ฒˆ์—ญ ์‹œ์Šคํ…œ ์ค‘ ํ•˜๋‚˜๋กœ ์ธ์ •๋ฐ›๋Š” **ALMA-13B-LoRA**์™€ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ตœ๊ทผ์— ์ถœ์‹œ๋œ LLM ๊ธฐ๋ฐ˜ ๋ฒˆ์—ญ ๋ชจ๋ธ์ด์ž ์ด ๋ถ„์•ผ์˜ ํ˜„๋Œ€์ ์ธ ์ž‘์—…์ธ **TowerInstruct5**์™€ ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค.6 ์ถ”๊ฐ€๋กœ, ํ˜„์žฌ ๋ชจ๋“  LLM ๊ธฐ๋ฐ˜ ๋ฒˆ์—ญ ์‹œ์Šคํ…œ ์ค‘ ์ตœ๊ณ ์˜ ๋ฒˆ์—ญ ๋ชจ๋ธ๋กœ ์ž…์ฆ๋œ ์ตœ์‹  **GPT-4** (gpt-4-1106-preview)์˜ ์ œ๋กœ์ƒท ์„ฑ๋Šฅ๊ณผ๋„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค (Xu et al., 2023; Zhang et al., 2023; Zeng et al., 2023; Jiao et al., 2023).
**SoTA ๋ชจ๋ธ** ์ด ๋ฒ”์ฃผ์—์„œ๋Š” ์šฐ๋ฆฌ์˜ ๋ฒค์น˜๋งˆํฌ๊ฐ€, ์šฐ๋ฆฌ๊ฐ€ ์•„๋Š” ํ•œ, ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ๋ฒˆ์—ญ ๋ชจ๋ธ์— ๋Œ€ํ•ด ์ˆ˜๋ฆฝ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ € WMT'21๊ณผ WMT'22์—์„œ ๋ชจ๋‘ NLLB-54B์™€ ๊ฐ™์€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ๊ธฐ์กด ๋ชจ๋ธ์„ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ตœ๊ณ ์˜ ์ค‘๊ฐ„ ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ๋ฒˆ์—ญ ์‹œ์Šคํ…œ ์ค‘ ํ•˜๋‚˜๋กœ ์ธ์ •๋ฐ›๋Š” **ALMA-13B-LoRA**์™€ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ตœ๊ทผ์— ์ถœ์‹œ๋œ LLM ๊ธฐ๋ฐ˜ ๋ฒˆ์—ญ ๋ชจ๋ธ์ด์ž ์ด ๋ถ„์•ผ์˜ ๋™์‹œ๋Œ€ ์ž‘ํ’ˆ์ธ **TowerInstruct5**์™€์˜ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค.6 ๋˜ํ•œ ํ˜„์žฌ ๋ชจ๋“  LLM ๊ธฐ๋ฐ˜ ๋ฒˆ์—ญ ์‹œ์Šคํ…œ ์ค‘ ์ตœ๊ณ ์˜ ๋ฒˆ์—ญ ๋ชจ๋ธ๋กœ ๋ฐํ˜€์ง„ ์ตœ์‹  **GPT-4**(gpt-4-1106-preview)์˜ ์ œ๋กœ์ƒท ์„ฑ๋Šฅ๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค(Xu et al., 2023; Zhang et al., 2023; Zeng et al., 2023; Jiao et al., 2023).
We allow for highly parallel generation by first observing that an inference pass of a language model produces a probability distribution over the next tokens for all input tokens. Naturally, this allows us to sample one next token from each token in the input. If one has generated a successor from each token, it is not possible to simply continue with the original sequence. For example, imagine predicting the next token after each token of "\(<bos>\) the cat sat" one might generate "yes orange saw down" - each successor by itself is a reasonable next token to a prefix of the sequence, but the list of tokens is a set of "counterfactual" continuations of these prefixes. We can, however, leverage these continuations to generate hidden thoughts for each observed token.
์šฐ๋ฆฌ๋Š” ๋จผ์ € ์–ธ์–ด ๋ชจ๋ธ์˜ ์ถ”๋ก  ๊ณผ์ •์ด ๋ชจ๋“  ์ž…๋ ฅ ํ† ํฐ์— ๋Œ€ํ•ด ๋‹ค์Œ ํ† ํฐ์— ๋Œ€ํ•œ ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ์ƒ์„ฑํ•œ๋‹ค๋Š” ์ ์„ ๊ด€์ฐฐํ•˜์—ฌ ๊ณ ๋„๋กœ ๋ณ‘๋ ฌ์ ์ธ ์ƒ์„ฑ์„ ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ, ์ด๋Š” ์ž…๋ ฅ์˜ ๊ฐ ํ† ํฐ์—์„œ ํ•˜๋‚˜์˜ ๋‹ค์Œ ํ† ํฐ์„ ์ƒ˜ํ”Œ๋งํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ํ† ํฐ์—์„œ ํ›„์† ํ† ํฐ์„ ์ƒ์„ฑํ•œ ๊ฒฝ์šฐ, ์›๋ž˜ ์‹œํ€€์Šค๋ฅผ ๋‹จ์ˆœํžˆ ๊ณ„์†ํ•˜๋Š” ๊ฒƒ์€ ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, "\(<bos>\) the cat sat"์˜ ๊ฐ ํ† ํฐ ์ดํ›„์˜ ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•œ๋‹ค๊ณ  ์ƒ์ƒํ•ด๋ณด๋ฉด "yes orange saw down"์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ํ›„์† ํ† ํฐ์€ ์‹œํ€€์Šค์˜ ์ ‘๋‘์‚ฌ์— ๋Œ€ํ•ด ํ•ฉ๋ฆฌ์ ์ธ ๋‹ค์Œ ํ† ํฐ์ด์ง€๋งŒ, ํ† ํฐ ๋ชฉ๋ก์€ ์ด๋Ÿฌํ•œ ์ ‘๋‘์‚ฌ์˜ "๋ฐ˜์‚ฌ์‹ค์ " ์—ฐ์†์˜ ์ง‘ํ•ฉ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ๋Š” ์ด๋Ÿฌํ•œ ์—ฐ์†์„ ํ™œ์šฉํ•˜์—ฌ ๊ฐ ๊ด€์ฐฐ๋œ ํ† ํฐ์— ๋Œ€ํ•œ ์ˆจ๊ฒจ์ง„ ์ƒ๊ฐ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋จผ์ € ์–ธ์–ด ๋ชจ๋ธ์˜ ์ถ”๋ก  ํŒจ์Šค๊ฐ€ ๋ชจ๋“  ์ž…๋ ฅ ํ† ํฐ์— ๋Œ€ํ•ด ๋‹ค์Œ ํ† ํฐ์— ๋Œ€ํ•œ ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ์ƒ์„ฑํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๊ด€์ฐฐํ•˜์—ฌ ๊ณ ๋„๋กœ ๋ณ‘๋ ฌํ™”๋œ ์ƒ์„ฑ์„ ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋‹น์—ฐํžˆ ์ด๋ฅผ ํ†ตํ•ด ์ž…๋ ฅ์˜ ๊ฐ ํ† ํฐ์—์„œ ํ•˜๋‚˜์˜ ๋‹ค์Œ ํ† ํฐ์„ ์ƒ˜ํ”Œ๋งํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ํ† ํฐ์—์„œ ํ›„์† ํ† ํฐ์„ ์ƒ์„ฑํ•œ ๊ฒฝ์šฐ ์›๋ž˜ ์‹œํ€€์Šค๋ฅผ ๊ณ„์† ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, "(the cat sat)์˜ ๊ฐ ํ† ํฐ ๋’ค์— ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•  ๋•Œ "yes orange saw down"์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š”๋ฐ, ๊ฐ ํ›„์† ํ† ํฐ์€ ๊ทธ ์ž์ฒด๋กœ ์‹œํ€€์Šค ์ ‘๋‘์‚ฌ์— ๋Œ€ํ•œ ํ•ฉ๋ฆฌ์ ์ธ ๋‹ค์Œ ํ† ํฐ์ด์ง€๋งŒ ํ† ํฐ ๋ชฉ๋ก์€ ์ด๋Ÿฌํ•œ ์ ‘๋‘์‚ฌ์˜ "๋ฐ˜๋“œ์‹œ" ์—ฐ์†์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์—ฐ์†์„ฑ์„ ํ™œ์šฉํ•˜์—ฌ ๊ด€์ฐฐ๋œ ๊ฐ ํ† ํฐ์— ๋Œ€ํ•œ ์ˆจ๊ฒจ์ง„ ์ƒ๊ฐ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Relation SelectionFirst, we determine the type of the bridge entity's descriptive mention by selecting the type of entities \(e_{1}\) and relation \(r_{1}\) to collect \(r_{1}(e_{1})=e_{2}\). The bridge entities we select have types like "song's singer" (the singer of a specific song), "country's anthem" (the country with a specific national anthem), "founder's organization" (the organization founded by a specific person), and "organization's ceo" (the CEO of a specific organization). For example, while there can be many authors for some novels, "author's novel" is selected as a type of descriptive mention of the bridge entity because we can use only the novels with a single author. We determine 19 types of bridge entity's descriptive mention with this process.
๊ด€๊ณ„ ์„ ํƒ: ๋จผ์ €, ์—”ํ‹ฐํ‹ฐ \(e_{1}\)์™€ ๊ด€๊ณ„ \(r_{1}\)๋ฅผ ์„ ํƒํ•˜์—ฌ \(r_{1}(e_{1})=e_{2}\)๋ฅผ ์ˆ˜์ง‘ํ•จ์œผ๋กœ์จ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์„ค๋ช…์  ์–ธ๊ธ‰ ์œ ํ˜•์„ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ์„ ํƒํ•œ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ๋Š” "๋…ธ๋ž˜์˜ ๊ฐ€์ˆ˜" (ํŠน์ • ๋…ธ๋ž˜์˜ ๊ฐ€์ˆ˜), "๊ตญ๊ฐ€์˜ ๊ตญ๊ฐ€" (ํŠน์ • ๊ตญ๊ฐ€์˜ ๊ตญ๊ฐ€), "์ฐฝ๋ฆฝ์ž์˜ ์กฐ์ง" (ํŠน์ • ์ธ๋ฌผ์ด ์ฐฝ๋ฆฝํ•œ ์กฐ์ง), "์กฐ์ง์˜ CEO" (ํŠน์ • ์กฐ์ง์˜ CEO)์™€ ๊ฐ™์€ ์œ ํ˜•์„ ๊ฐ€์ง‘๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ผ๋ถ€ ์†Œ์„ค์—๋Š” ์—ฌ๋Ÿฌ ์ €์ž๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ์šฐ๋ฆฌ๋Š” ๋‹จ์ผ ์ €์ž๊ฐ€ ์žˆ๋Š” ์†Œ์„ค๋งŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— "์ €์ž์˜ ์†Œ์„ค"์„ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์„ค๋ช…์  ์–ธ๊ธ‰ ์œ ํ˜•์œผ๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์„ ํ†ตํ•ด ์šฐ๋ฆฌ๋Š” 19๊ฐ€์ง€ ์œ ํ˜•์˜ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์„ค๋ช…์  ์–ธ๊ธ‰์„ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค.
๊ด€๊ณ„ ์„ ํƒ๋จผ์ €, \(r_{1}(e_{1})=e_{2}\)๋ฅผ ์ˆ˜์ง‘ํ•˜๊ธฐ ์œ„ํ•ด ์—”ํ‹ฐํ‹ฐ \(e_{1}\)์™€ ๊ด€๊ณ„ \(r_{1}\)์˜ ์œ ํ˜•์„ ์„ ํƒํ•˜์—ฌ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์„ค๋ช…์  ์–ธ๊ธ‰ ์œ ํ˜•์„ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ์„ ํƒํ•˜๋Š” ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์—๋Š” "๋…ธ๋ž˜์˜ ๊ฐ€์ˆ˜"(ํŠน์ • ๋…ธ๋ž˜์˜ ๊ฐ€์ˆ˜), "๊ตญ๊ฐ€์˜ ์• ๊ตญ๊ฐ€"(ํŠน์ • ๊ตญ๊ฐ€์˜ ์• ๊ตญ๊ฐ€), "์ฐฝ๋ฆฝ์ž์˜ ์กฐ์ง"(ํŠน์ • ์‚ฌ๋žŒ์ด ์„ค๋ฆฝํ•œ ์กฐ์ง), "์กฐ์ง์˜ CEO"(ํŠน์ • ์กฐ์ง์˜ CEO) ๋“ฑ์˜ ์œ ํ˜•์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ผ๋ถ€ ์†Œ์„ค์—๋Š” ๋งŽ์€ ์ €์ž๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ์ €์ž๋Š” ํ•œ ๋ช…๋งŒ ์žˆ๋Š” ์†Œ์„ค์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— '์ €์ž์˜ ์†Œ์„ค'์ด ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ์„ค๋ช…์  ์–ธ๊ธ‰ ์œ ํ˜•์œผ๋กœ ์„ ํƒ๋ฉ๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์„ ํ†ตํ•ด 19๊ฐ€์ง€ ์œ ํ˜•์˜ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ ์„ค๋ช…์  ์–ธ๊ธ‰์„ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค.
However, benchmarks have limitations, such as being misleading and motivating narrow research goals [162]. In particular, they risk becoming saturated after a period of time if models can overfit to them [155]. Some benchmarks have also been criticized for low ecological validity, as their component tests do not closely approximate real-world data [163, 164]. Therefore, constructing more ecologically valid benchmarks that generalize to real-world scenarios is an active area of research [19]. Notably, several projects have sought to rethink benchmarking in order to make it more challenging and valid, such as Dynabench [165], which uses human-and-model-in-the-loop evaluation. We aim to take these limitations and concerns into account as we develop our benchmark.
๊ทธ๋Ÿฌ๋‚˜ ๋ฒค์น˜๋งˆํฌ์—๋Š” ์˜คํ•ด๋ฅผ ๋ถˆ๋Ÿฌ์ผ์œผํ‚ค๊ณ  ์ข์€ ์—ฐ๊ตฌ ๋ชฉํ‘œ๋ฅผ ์œ ๋„ํ•˜๋Š” ๋“ฑ์˜ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค [162]. ํŠนํžˆ, ๋ชจ๋ธ์ด ๋ฒค์น˜๋งˆํฌ์— ๊ณผ์ ํ•ฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ ์ผ์ • ๊ธฐ๊ฐ„ ํ›„ ํฌํ™” ์ƒํƒœ์— ์ด๋ฅผ ์œ„ํ—˜์ด ์žˆ์Šต๋‹ˆ๋‹ค [155]. ์ผ๋ถ€ ๋ฒค์น˜๋งˆํฌ๋Š” ๊ตฌ์„ฑ ์š”์†Œ ํ…Œ์ŠคํŠธ๊ฐ€ ์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ์ •ํ™•ํžˆ ๋ฐ˜์˜ํ•˜์ง€ ๋ชปํ•ด ์ƒํƒœํ•™์  ํƒ€๋‹น์„ฑ์ด ๋‚ฎ๋‹ค๋Š” ๋น„ํŒ์„ ๋ฐ›๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค [163, 164]. ๋”ฐ๋ผ์„œ ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์— ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๋” ์ƒํƒœํ•™์ ์œผ๋กœ ํƒ€๋‹นํ•œ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์ด ํ™œ๋ฐœํ•œ ์—ฐ๊ตฌ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค [19]. ํŠนํžˆ, Dynabench [165]์™€ ๊ฐ™์ด ์ธ๊ฐ„๊ณผ ๋ชจ๋ธ์ด ํ•จ๊ป˜ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜๋Š” ํ”„๋กœ์ ํŠธ๋“ค์€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋” ๋„์ „์ ์ด๊ณ  ํƒ€๋‹นํ•˜๊ฒŒ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด ์žฌ๊ณ ํ•˜๋ ค๋Š” ๋…ธ๋ ฅ์„ ๊ธฐ์šธ์ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ฐœ๋ฐœํ•˜๋ฉด์„œ ์ด๋Ÿฌํ•œ ํ•œ๊ณ„์™€ ์šฐ๋ ค๋ฅผ ๊ณ ๋ คํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ ๋ฒค์น˜๋งˆํฌ์—๋Š” ์˜คํ•ด์˜ ์†Œ์ง€๊ฐ€ ์žˆ๊ณ  ํ˜‘์†Œํ•œ ์—ฐ๊ตฌ ๋ชฉํ‘œ๋ฅผ ์œ ๋ฐœํ•˜๋Š” ๋“ฑ์˜ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค[162]. ํŠนํžˆ, ๋ชจ๋ธ์ด ๋ฒค์น˜๋งˆํฌ์— ๊ณผ์ ํ•ฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ ์ผ์ • ๊ธฐ๊ฐ„์ด ์ง€๋‚˜๋ฉด ํฌํ™” ์ƒํƒœ์— ์ด๋ฅผ ์œ„ํ—˜์ด ์žˆ์Šต๋‹ˆ๋‹ค[155]. ์ผ๋ถ€ ๋ฒค์น˜๋งˆํฌ๋Š” ๊ตฌ์„ฑ ์š”์†Œ ํ…Œ์ŠคํŠธ๊ฐ€ ์‹ค์ œ ๋ฐ์ดํ„ฐ์— ๊ทผ์ ‘ํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ์ƒํƒœ์  ํƒ€๋‹น์„ฑ์ด ๋‚ฎ๋‹ค๋Š” ๋น„ํŒ์„ ๋ฐ›๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค[163, 164]. ๋”ฐ๋ผ์„œ ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค๋กœ ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์ƒํƒœํ•™์ ์œผ๋กœ ๋” ํƒ€๋‹นํ•œ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์ด ํ™œ๋ฐœํ•œ ์—ฐ๊ตฌ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค[19]. ํŠนํžˆ, ์ธ๊ฐ„๊ณผ ๋ชจ๋ธ์„ ์—ฐ๋™ํ•œ ํ‰๊ฐ€๋ฅผ ์‚ฌ์šฉํ•˜๋Š” Dynabench[165]์™€ ๊ฐ™์ด ๋ฒค์น˜๋งˆํ‚น์„ ๋”์šฑ ๋„์ „์ ์ด๊ณ  ์œ ํšจํ•˜๊ฒŒ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด ๋ฒค์น˜๋งˆํ‚น์„ ์žฌ๊ณ ํ•˜๋Š” ์—ฌ๋Ÿฌ ํ”„๋กœ์ ํŠธ๊ฐ€ ์ง„ํ–‰๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ด๋Ÿฌํ•œ ํ•œ๊ณ„์™€ ์šฐ๋ ค๋ฅผ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ฐœ๋ฐœํ•˜๋Š” ๊ณผ์ •์—์„œ ๊ณ ๋ คํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.
LLMs represent a disruptive technology that has received significant attention from the public and is widely used across societies speaking different languages. Therefore, ensuring a democratization of the technology across people of different languages will represent an important value. Our study highlights that neglecting multilingualism while training a tokenizer representing a core component required for training LLMs can cause severe disadvantages, such as increased training costs and decreased downstream performance, raising major ethical concerns. Furthermore, the increased training costs translate into an increased carbon footprint, which has an environmental impact. Our findings support an improved development and usage of this fundamental technology.
LLMs๋Š” ๋Œ€์ค‘์˜ ํฐ ๊ด€์‹ฌ์„ ๋ฐ›๊ณ  ์žˆ์œผ๋ฉฐ, ๋‹ค์–‘ํ•œ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์‚ฌํšŒ ์ „๋ฐ˜์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ํ˜์‹ ์ ์ธ ๊ธฐ์ˆ ์„ ๋Œ€ํ‘œํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ, ๋‹ค์–‘ํ•œ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์‚ฌ๋žŒ๋“ค ๊ฐ„์— ์ด ๊ธฐ์ˆ ์˜ ๋ฏผ์ฃผํ™”๋ฅผ ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์€ ์ค‘์š”ํ•œ ๊ฐ€์น˜๋ฅผ ์ง€๋‹™๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” LLMs ํ›ˆ๋ จ์— ํ•„์š”ํ•œ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ์ธ ํ† ํฌ๋‚˜์ด์ €๋ฅผ ํ›ˆ๋ จํ•  ๋•Œ ๋‹ค์ค‘ ์–ธ์–ด๋ฅผ ๋ฌด์‹œํ•˜๋ฉด ํ›ˆ๋ จ ๋น„์šฉ ์ฆ๊ฐ€์™€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ ์ €ํ•˜์™€ ๊ฐ™์€ ์‹ฌ๊ฐํ•œ ๋‹จ์ ์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ฃผ์š” ์œค๋ฆฌ์  ๋ฌธ์ œ๋ฅผ ์•ผ๊ธฐํ•  ์ˆ˜ ์žˆ์Œ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ฆ๊ฐ€๋œ ํ›ˆ๋ จ ๋น„์šฉ์€ ํƒ„์†Œ ๋ฐœ์ž๊ตญ ์ฆ๊ฐ€๋กœ ์ด์–ด์ ธ ํ™˜๊ฒฝ์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” ์ด ๊ธฐ๋ณธ ๊ธฐ์ˆ ์˜ ๊ฐœ์„ ๋œ ๊ฐœ๋ฐœ ๋ฐ ์‚ฌ์šฉ์„ ์ง€์ง€ํ•ฉ๋‹ˆ๋‹ค.
LLM์€ ๋Œ€์ค‘์˜ ๋งŽ์€ ๊ด€์‹ฌ์„ ๋ฐ›๊ณ  ์žˆ์œผ๋ฉฐ ๋‹ค์–‘ํ•œ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์‚ฌํšŒ ์ „๋ฐ˜์— ๊ฑธ์ณ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ํŒŒ๊ดด์ ์ธ ๊ธฐ์ˆ ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์–‘ํ•œ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์‚ฌ๋žŒ๋“ค์— ๋Œ€ํ•œ ๊ธฐ์ˆ ์˜ ๋ฏผ์ฃผํ™”๋ฅผ ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์€ ์ค‘์š”ํ•œ ๊ฐ€์น˜๊ฐ€ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ์—ฐ๊ตฌ๋Š” LLM ํ›ˆ๋ จ์— ํ•„์š”ํ•œ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ๋Œ€ํ‘œํ•˜๋Š” ํ† ํฐํ™”๊ธฐ๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๋™์•ˆ ๋‹ค๊ตญ์–ด๋ฅผ ๋ฌด์‹œํ•˜๋ฉด ํ›ˆ๋ จ ๋น„์šฉ ์ฆ๊ฐ€ ๋ฐ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ ์ €ํ•˜์™€ ๊ฐ™์€ ์‹ฌ๊ฐํ•œ ๋ถˆ์ด์ต์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ฃผ์š” ์œค๋ฆฌ์  ๋ฌธ์ œ๋ฅผ ์ œ๊ธฐํ•œ๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ›ˆ๋ จ ๋น„์šฉ์ด ์ฆ๊ฐ€ํ•˜๋ฉด ํƒ„์†Œ ๋ฐœ์ž๊ตญ์ด ์ฆ๊ฐ€ํ•˜์—ฌ ํ™˜๊ฒฝ์— ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” ์ด ๊ธฐ๋ณธ ๊ธฐ์ˆ ์˜ ๊ฐœ์„ ๋œ ๊ฐœ๋ฐœ๊ณผ ์‚ฌ์šฉ์„ ๋’ท๋ฐ›์นจํ•ฉ๋‹ˆ๋‹ค.
The scenario packs and simulated patients in our study were prepared by two OSCE laboratories (one each in Canada and India), each affiliated to a medical school and with extensive experience in preparing scenario packs and simulated patients for OSCE examinations. UK scenario packs were sourced from the samples provided on the MRCPUK website. Each scenario pack was associated with a ground truth diagnosis and a set of acceptable diagnoses. The scenario packs covered conditions from cardiovascular (29), respiratory (30), gastroenterology (31), neurology (30), urology, obstetric, and gynecology domains (15), and internal medicine (14). Pediatric or psychiatry domains were excluded from this study, as were intensive care or inpatient case management scenarios.
๋ณธ ์—ฐ๊ตฌ์—์„œ ์‚ฌ์šฉ๋œ ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ๊ณผ ๋ชจ์˜ ํ™˜์ž๋Š” ์บ๋‚˜๋‹ค์™€ ์ธ๋„์— ๊ฐ๊ฐ ์œ„์น˜ํ•œ ๋‘ ๊ฐœ์˜ OSCE ์‹คํ—˜์‹ค์—์„œ ์ค€๋น„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์‹คํ—˜์‹ค๋“ค์€ ๊ฐ๊ฐ ์˜๊ณผ๋Œ€ํ•™์— ์†Œ์†๋˜์–ด ์žˆ์œผ๋ฉฐ, OSCE ์‹œํ—˜์„ ์œ„ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ๊ณผ ๋ชจ์˜ ํ™˜์ž๋ฅผ ์ค€๋น„ํ•˜๋Š” ๋ฐ ์žˆ์–ด ํ’๋ถ€ํ•œ ๊ฒฝํ—˜์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜๊ตญ์˜ ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ์€ MRCPUK ์›น์‚ฌ์ดํŠธ์—์„œ ์ œ๊ณต๋œ ์ƒ˜ํ”Œ์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ์€ ์ง„๋‹จ์˜ ์ •๋‹ต๊ณผ ํ—ˆ์šฉ ๊ฐ€๋Šฅํ•œ ์ง„๋‹จ ์„ธํŠธ์™€ ์—ฐ๊ด€๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ์€ ์‹ฌํ˜ˆ๊ด€๊ณ„ (29), ํ˜ธํก๊ธฐ๊ณ„ (30), ์†Œํ™”๊ธฐ๊ณ„ (31), ์‹ ๊ฒฝ๊ณ„ (30), ๋น„๋‡จ๊ธฐ๊ณผ, ์‚ฐ๋ถ€์ธ๊ณผ (15), ๋‚ด๊ณผ (14) ๋ถ„์•ผ์˜ ์งˆํ™˜์„ ๋‹ค๋ฃจ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์†Œ์•„๊ณผ๋‚˜ ์ •์‹ ๊ณผ ๋ถ„์•ผ๋Š” ๋ณธ ์—ฐ๊ตฌ์—์„œ ์ œ์™ธ๋˜์—ˆ์œผ๋ฉฐ, ์ค‘ํ™˜์ž ์น˜๋ฃŒ๋‚˜ ์ž…์› ํ™˜์ž ๊ด€๋ฆฌ ์‹œ๋‚˜๋ฆฌ์˜ค๋„ ์ œ์™ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
์ด ์—ฐ๊ตฌ์˜ ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ๊ณผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜์ž๋Š” ๊ฐ๊ฐ ์˜๊ณผ๋Œ€ํ•™์— ์†Œ์†๋˜์–ด ์žˆ๊ณ  OSCE ์‹œํ—˜์„ ์œ„ํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ๊ณผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜์ž๋ฅผ ์ค€๋น„ํ•œ ๊ฒฝํ—˜์ด ํ’๋ถ€ํ•œ ๋‘ ๊ฐœ์˜ OSCE ์‹คํ—˜์‹ค(๊ฐ๊ฐ ์บ๋‚˜๋‹ค์™€ ์ธ๋„์— 1๊ฐœ์”ฉ)์—์„œ ์ค€๋น„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜๊ตญ ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ์€ MRCPUK ์›น์‚ฌ์ดํŠธ์—์„œ ์ œ๊ณตํ•˜๋Š” ์ƒ˜ํ”Œ์—์„œ ์กฐ๋‹ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ์—๋Š” ์ •๋‹ต ์ง„๋‹จ๊ณผ ํ—ˆ์šฉ ๊ฐ€๋Šฅํ•œ ์ง„๋‹จ ์„ธํŠธ๊ฐ€ ์—ฐ๊ฒฐ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์‹œ๋‚˜๋ฆฌ์˜ค ํŒฉ์€ ์‹ฌํ˜ˆ๊ด€(29), ํ˜ธํก๊ธฐ(30), ์œ„์žฅ๋ณ‘(31), ์‹ ๊ฒฝํ•™(30), ๋น„๋‡จ๊ธฐ, ์‚ฐ๋ถ€์ธ๊ณผ ์˜์—ญ(15), ๋‚ด๊ณผ(14)์˜ ์งˆํ™˜์„ ๋‹ค๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค. ์†Œ์•„๊ณผ ๋˜๋Š” ์ •์‹ ๊ณผ ์˜์—ญ์€ ๋ณธ ์—ฐ๊ตฌ์—์„œ ์ œ์™ธ๋˜์—ˆ์œผ๋ฉฐ ์ค‘ํ™˜์ž์‹ค ๋˜๋Š” ์ž…์› ํ™˜์ž ๊ด€๋ฆฌ ์‹œ๋‚˜๋ฆฌ์˜ค๋„ ์ œ์™ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
One can view the extra data \(\{\overleftarrow{x_{i}}\}\) as another language that the language model has to learn left-to-right - in this case a reversed natural language, which has a similar difficulty in terms of perplexity. As it is easy for the language model to identify which of these languages it is trying to generate from when predicting the next token, this does not tend to interfere with its language modeling abilities in the standard forward direction. Further, as it has been shown that LLMs can leverage knowledge across different sources (e.g., code to math (Shao et al., 2024), or different natural languages (Lample & Conneau, 2019)) we hypothesize that the knowledge it learns from the reverse direction can help in the forward direction as well.
์ถ”๊ฐ€ ๋ฐ์ดํ„ฐ \(\{\overleftarrow{x_{i}}\}\)๋ฅผ ์–ธ์–ด ๋ชจ๋ธ์ด ์ขŒ์—์„œ ์šฐ๋กœ ํ•™์Šตํ•ด์•ผ ํ•˜๋Š” ๋˜ ๋‹ค๋ฅธ ์–ธ์–ด๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ, ์ด๋Š” ์œ ์‚ฌํ•œ ๋‚œ์ด๋„๋ฅผ ๊ฐ€์ง„ ์—ญ๋ฐฉํ–ฅ ์ž์—ฐ์–ด์ž…๋‹ˆ๋‹ค. ์–ธ์–ด ๋ชจ๋ธ์ด ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•  ๋•Œ ์–ด๋–ค ์–ธ์–ด์—์„œ ์ƒ์„ฑํ•˜๋ ค๊ณ  ํ•˜๋Š”์ง€ ์‰ฝ๊ฒŒ ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์—, ์ด๋Š” ํ‘œ์ค€ ์ˆœ๋ฐฉํ–ฅ์—์„œ ์–ธ์–ด ๋ชจ๋ธ๋ง ๋Šฅ๋ ฅ์„ ๋ฐฉํ•ดํ•˜์ง€ ์•Š๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, LLM์ด ๋‹ค์–‘ํ•œ ์†Œ์Šค ๊ฐ„์˜ ์ง€์‹์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— (์˜ˆ: ์ฝ”๋“œ์—์„œ ์ˆ˜ํ•™์œผ๋กœ (Shao et al., 2024) ๋˜๋Š” ๋‹ค๋ฅธ ์ž์—ฐ์–ด ๊ฐ„ (Lample & Conneau, 2019)), ์—ญ๋ฐฉํ–ฅ์—์„œ ํ•™์Šตํ•œ ์ง€์‹์ด ์ˆœ๋ฐฉํ–ฅ์—์„œ๋„ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค.
์ถ”๊ฐ€ ๋ฐ์ดํ„ฐ \(\{\overleftarrow{x_{i}}\}\)๋ฅผ ์–ธ์–ด ๋ชจ๋ธ์ด ์™ผ์ชฝ์—์„œ ์˜ค๋ฅธ์ชฝ์œผ๋กœ ํ•™์Šตํ•ด์•ผ ํ•˜๋Š” ๋˜ ๋‹ค๋ฅธ ์–ธ์–ด(์ด ๊ฒฝ์šฐ ๋’ค๋ฐ”๋€ ์ž์—ฐ์–ด)๋กœ ๋ณผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋‹นํ˜น๊ฐ ์ธก๋ฉด์—์„œ ๋น„์Šทํ•œ ๋‚œ์ด๋„๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์–ธ์–ด ๋ชจ๋ธ์ด ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•  ๋•Œ ์ด๋Ÿฌํ•œ ์–ธ์–ด ์ค‘ ์–ด๋–ค ์–ธ์–ด์—์„œ ์ƒ์„ฑํ•˜๋ ค๋Š”์ง€ ์‰ฝ๊ฒŒ ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ํ‘œ์ค€ ์ˆœ๋ฐฉํ–ฅ์˜ ์–ธ์–ด ๋ชจ๋ธ๋ง ๊ธฐ๋Šฅ์„ ๋ฐฉํ•ดํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์ด ์„œ๋กœ ๋‹ค๋ฅธ ์†Œ์Šค(์˜ˆ: ์ฝ”๋“œ์—์„œ ์ˆ˜ํ•™(Shao et al., 2024) ๋˜๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ์ž์—ฐ์–ด(Lample & Conneau, 2019))์— ๊ฑธ์ณ ์ง€์‹์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์œผ๋ฏ€๋กœ ์—ญ๋ฐฉํ–ฅ์—์„œ ํ•™์Šตํ•œ ์ง€์‹์ด ์ˆœ๋ฐฉํ–ฅ์—์„œ๋„ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค.
* Result 1(a) highlights the maximum capacity across models; however, this could be misleading if only a single model achieves this peak. * Result 1(b) reinforces this by showing that all models with a maximum capacity \(R^{\mathsf{max}}(F)\leq 1.8\) can achieve such maximum capacity, i.e., \(R(F)\approx R^{\mathsf{max}}(F)\). In words, this indicates that for a dataset containing \(B\) bits of knowledge, selecting a model size \(P\geq B/1.8\) is sufficient. * Result 1(c) further strengthens this by indicating that no model exceeds capacity ratio 2.3. _For clarity_, in subsequent results of this paper, we focus solely on the _peak_ capacity ratio, with the understanding that observations similar to Result 1(b) and Result 1(c) **consistently apply**.
* ๊ฒฐ๊ณผ 1(a)๋Š” ๋ชจ๋ธ ๊ฐ„ ์ตœ๋Œ€ ์šฉ๋Ÿ‰์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Š” ๋‹จ์ผ ๋ชจ๋ธ๋งŒ์ด ์ด ์ •์ ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒฝ์šฐ ์˜คํ•ด์˜ ์†Œ์ง€๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * ๊ฒฐ๊ณผ 1(b)๋Š” ์ตœ๋Œ€ ์šฉ๋Ÿ‰ \(R^{\mathsf{max}}(F)\leq 1.8\)์„ ๊ฐ€์ง„ ๋ชจ๋“  ๋ชจ๋ธ์ด ์ด๋Ÿฌํ•œ ์ตœ๋Œ€ ์šฉ๋Ÿ‰์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์คŒ์œผ๋กœ์จ ์ด๋ฅผ ๊ฐ•ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, \(R(F)\approx R^{\mathsf{max}}(F)\)์ž…๋‹ˆ๋‹ค. ์ด๋Š” \(B\) ๋น„ํŠธ์˜ ์ง€์‹์„ ํฌํ•จํ•˜๋Š” ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•ด ๋ชจ๋ธ ํฌ๊ธฐ \(P\geq B/1.8\)๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ์ถฉ๋ถ„ํ•จ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. * ๊ฒฐ๊ณผ 1(c)๋Š” ์šฉ๋Ÿ‰ ๋น„์œจ 2.3์„ ์ดˆ๊ณผํ•˜๋Š” ๋ชจ๋ธ์ด ์—†์Œ์„ ๋‚˜ํƒ€๋‚ด์–ด ์ด๋ฅผ ๋”์šฑ ๊ฐ•ํ™”ํ•ฉ๋‹ˆ๋‹ค. _๋ช…ํ™•์„ฑ์„ ์œ„ํ•ด_, ์ด ๋…ผ๋ฌธ์˜ ํ›„์† ๊ฒฐ๊ณผ์—์„œ๋Š” _์ตœ๋Œ€_ ์šฉ๋Ÿ‰ ๋น„์œจ์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ”๋ฉฐ, ๊ฒฐ๊ณผ 1(b) ๋ฐ ๊ฒฐ๊ณผ 1(c)์™€ ์œ ์‚ฌํ•œ ๊ด€์ฐฐ์ด **์ผ๊ด€๋˜๊ฒŒ ์ ์šฉ๋จ**์„ ์ดํ•ดํ•ฉ๋‹ˆ๋‹ค.
* ๊ฒฐ๊ณผ 1(a)๋Š” ๋ชจ๋ธ ๊ฐ„ ์ตœ๋Œ€ ์šฉ๋Ÿ‰์„ ๊ฐ•์กฐํ•˜์ง€๋งŒ, ๋‹จ์ผ ๋ชจ๋ธ๋งŒ์ด ์ด ์ตœ๋Œ€ ์šฉ๋Ÿ‰์— ๋„๋‹ฌํ•˜๋Š” ๊ฒฝ์šฐ ์˜คํ•ด์˜ ์†Œ์ง€๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * ๊ฒฐ๊ณผ 1(b)๋Š” ์ตœ๋Œ€ ์šฉ๋Ÿ‰ \(R^{\mathsf{max}}(F)\leq 1.8\)์„ ๊ฐ€์ง„ ๋ชจ๋“  ๋ชจ๋ธ์ด ์ด๋Ÿฌํ•œ ์ตœ๋Œ€ ์šฉ๋Ÿ‰, ์ฆ‰ \(R(F)\์•ฝ R^{\mathsf{max}}(F)\)์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์คŒ์œผ๋กœ์จ ์ด๋ฅผ ๊ฐ•ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ์ด๋Š” ์ง€์‹์ด \(B\) ๋น„ํŠธ ํฌํ•จ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ ๋ชจ๋ธ ํฌ๊ธฐ \(P\geq B/1.8\)์„ ์„ ํƒํ•˜๋ฉด ์ถฉ๋ถ„ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. * ๊ฒฐ๊ณผ 1(c)๋Š” ์–ด๋–ค ๋ชจ๋ธ๋„ ์šฉ๋Ÿ‰ ๋น„์œจ 2.3์„ ์ดˆ๊ณผํ•˜์ง€ ์•Š๋Š”๋‹ค๋Š” ์ ์„ ๋‚˜ํƒ€๋‚ด์–ด ์ด๋ฅผ ๋”์šฑ ๊ฐ•ํ™”ํ•ฉ๋‹ˆ๋‹ค. _๋ช…ํ™•์„ฑ์„ ์œ„ํ•ด_, ์ด ๋ฐฑ์„œ์˜ ํ›„์† ๊ฒฐ๊ณผ์—์„œ๋Š” ๊ฒฐ๊ณผ 1(b) ๋ฐ ๊ฒฐ๊ณผ 1(c)์™€ ์œ ์‚ฌํ•œ ๊ด€์ฐฐ์ด **์ผ๊ด€๋˜๊ฒŒ ์ ์šฉ**๋œ๋‹ค๋Š” ์ ์„ ์ดํ•ดํ•˜๋ฉด์„œ _ํ”ผํฌ_ ์šฉ๋Ÿ‰ ๋น„์œจ์—๋งŒ ์ดˆ์ ์„ ๋งž์ถœ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
While traditional retrieval methods, such as the BM25 algorithm (Hancock-Beaulieu et al., 1996), focus on term frequency and presence for document ranking, they often overlook the semantic information of queries. Current strategies leverage pre-trained LMs like BERT (Devlin et al., 2019), which capture the semantic essence of queries more effectively. These models improve search accuracy by considering synonyms and the structure of phrases, thereby refining document ranking through the detection of semantic similarities. This is typically achieved by measuring vector distances between documents and queries, combining traditional retrieval metrics with semantic understanding to yield search results that are both relevant and aligned with user intent.
์ „ํ†ต์ ์ธ ๊ฒ€์ƒ‰ ๋ฐฉ๋ฒ•, ์˜ˆ๋ฅผ ๋“ค์–ด BM25 ์•Œ๊ณ ๋ฆฌ์ฆ˜ (Hancock-Beaulieu et al., 1996)์€ ๋ฌธ์„œ ์ˆœ์œ„๋ฅผ ๋งค๊ธธ ๋•Œ ์šฉ์–ด ๋นˆ๋„์™€ ์กด์žฌ ์—ฌ๋ถ€์— ์ค‘์ ์„ ๋‘์ง€๋งŒ, ์ข…์ข… ์ฟผ๋ฆฌ์˜ ์˜๋ฏธ ์ •๋ณด๋ฅผ ๊ฐ„๊ณผํ•ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ์˜ ์ „๋žต์€ BERT (Devlin et al., 2019)์™€ ๊ฐ™์€ ์‚ฌ์ „ ํ•™์Šต๋œ LMs๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ฟผ๋ฆฌ์˜ ์˜๋ฏธ์  ๋ณธ์งˆ์„ ๋” ํšจ๊ณผ์ ์œผ๋กœ ํฌ์ฐฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ๋“ค์€ ๋™์˜์–ด์™€ ๊ตฌ๋ฌธ์˜ ๊ตฌ์กฐ๋ฅผ ๊ณ ๋ คํ•˜์—ฌ ๊ฒ€์ƒ‰ ์ •ํ™•๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ค๋ฉฐ, ์˜๋ฏธ์  ์œ ์‚ฌ์„ฑ์„ ๊ฐ์ง€ํ•˜์—ฌ ๋ฌธ์„œ ์ˆœ์œ„๋ฅผ ์„ธ๋ถ„ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฌธ์„œ์™€ ์ฟผ๋ฆฌ ๊ฐ„์˜ ๋ฒกํ„ฐ ๊ฑฐ๋ฆฌ๋ฅผ ์ธก์ •ํ•จ์œผ๋กœ์จ ๋‹ฌ์„ฑ๋˜๋ฉฐ, ์ „ํ†ต์ ์ธ ๊ฒ€์ƒ‰ ์ง€ํ‘œ์™€ ์˜๋ฏธ์  ์ดํ•ด๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ ์‚ฌ์šฉ์ž ์˜๋„์— ๋ถ€ํ•ฉํ•˜๋Š” ๊ด€๋ จ์„ฑ ๋†’์€ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
BM25 ์•Œ๊ณ ๋ฆฌ์ฆ˜(Hancock-Beaulieu et al., 1996)๊ณผ ๊ฐ™์€ ๊ธฐ์กด ๊ฒ€์ƒ‰ ๋ฐฉ๋ฒ•์€ ๋ฌธ์„œ ์ˆœ์œ„๋ฅผ ๋งค๊ธฐ๊ธฐ ์œ„ํ•ด ์šฉ์–ด ๋นˆ๋„์™€ ์กด์žฌ ์—ฌ๋ถ€์— ์ดˆ์ ์„ ๋งž์ถ”์ง€๋งŒ, ์ฟผ๋ฆฌ์˜ ์˜๋ฏธ ์ •๋ณด๋ฅผ ๊ฐ„๊ณผํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ตœ์‹  ์ „๋žต์€ ์ฟผ๋ฆฌ์˜ ์˜๋ฏธ์  ๋ณธ์งˆ์„ ๋ณด๋‹ค ํšจ๊ณผ์ ์œผ๋กœ ํฌ์ฐฉํ•˜๋Š” BERT(Devlin et al., 2019)์™€ ๊ฐ™์€ ์‚ฌ์ „ ํ•™์Šต๋œ LM์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ๋™์˜์–ด์™€ ๊ตฌ๋ฌธ์˜ ๊ตฌ์กฐ๋ฅผ ๊ณ ๋ คํ•˜์—ฌ ๋ฌธ์„œ ์ˆœ์œ„๋ฅผ ๊ฐœ์„ ํ•จ์œผ๋กœ์จ ์˜๋ฏธ ์œ ์‚ฌ์„ฑ์„ ๊ฐ์ง€ํ•˜์—ฌ ๋ฌธ์„œ ์ˆœ์œ„๋ฅผ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฌธ์„œ์™€ ์ฟผ๋ฆฌ ๊ฐ„์˜ ๋ฒกํ„ฐ ๊ฑฐ๋ฆฌ๋ฅผ ์ธก์ •ํ•˜์—ฌ ๊ธฐ์กด ๊ฒ€์ƒ‰ ๋ฉ”ํŠธ๋ฆญ๊ณผ ์˜๋ฏธ๋ก ์  ์ดํ•ด๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ ๊ด€๋ จ์„ฑ์ด ๋†’๊ณ  ์‚ฌ์šฉ์ž ์˜๋„์— ๋ถ€ํ•ฉํ•˜๋Š” ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•จ์œผ๋กœ์จ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Proof.: Algorithm 3 maintains the invariant that \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\in[\alpha_{i},\beta_{i}]\) in each round. We will prove by induction that this is true and that the true vector of logits always lies in \(\mathcal{C}\). Note that by the assumption stated in the Lemma, this is clearly true at the beginning of the first round. Suppose that this is true after \(K<T\) rounds. Then, in the \(K+1\)-th round, the constraints added are all valid constraints for the true logit vector, since the API returning token \(k\) guarantees that \(\mathsf{logit}_{k}+b_{k}\geq\mathsf{logit}_{j}+b_{j}\) for all \(j\neq k\). Hence, by induction, the algorithm always ensures that \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\in[\alpha_{i},\beta_{i}]\).
์ฆ๋ช…: ์•Œ๊ณ ๋ฆฌ์ฆ˜ 3์€ ๊ฐ ๋ผ์šด๋“œ์—์„œ \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\in[\alpha_{i},\beta_{i}]\)๋ผ๋Š” ๋ถˆ๋ณ€์„ฑ์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๊ท€๋‚ฉ๋ฒ•์„ ํ†ตํ•ด ์ด๊ฒƒ์ด ์ฐธ์ด๋ฉฐ, ๋กœ๊ทธ์ž‡์˜ ์‹ค์ œ ๋ฒกํ„ฐ๊ฐ€ ํ•ญ์ƒ \(\mathcal{C}\)์— ์†ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์ฆ๋ช…ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ ˆ๋งˆ์—์„œ ์–ธ๊ธ‰๋œ ๊ฐ€์ •์— ๋”ฐ๋ฅด๋ฉด, ์ฒซ ๋ฒˆ์งธ ๋ผ์šด๋“œ์˜ ์‹œ์ž‘ ์‹œ์ ์—์„œ ์ด๊ฒƒ์ด ๋ช…๋ฐฑํžˆ ์ฐธ์ž„์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. \(K<T\) ๋ผ์šด๋“œ ํ›„์—๋„ ์ด๊ฒƒ์ด ์ฐธ์ด๋ผ๊ณ  ๊ฐ€์ •ํ•ฉ์‹œ๋‹ค. ๊ทธ๋Ÿฌ๋ฉด \(K+1\)๋ฒˆ์งธ ๋ผ์šด๋“œ์—์„œ ์ถ”๊ฐ€๋œ ์ œ์•ฝ ์กฐ๊ฑด๋“ค์€ ๋ชจ๋‘ ์‹ค์ œ ๋กœ๊ทธ์ž‡ ๋ฒกํ„ฐ์— ๋Œ€ํ•œ ์œ ํšจํ•œ ์ œ์•ฝ ์กฐ๊ฑด๋“ค์ž…๋‹ˆ๋‹ค. ์™œ๋ƒํ•˜๋ฉด API๊ฐ€ ํ† ํฐ \(k\)๋ฅผ ๋ฐ˜ํ™˜ํ•  ๋•Œ \(\mathsf{logit}_{k}+b_{k}\geq\mathsf{logit}_{j}+b_{j}\) (๋ชจ๋“  \(j\neq k\)์— ๋Œ€ํ•ด)๋ฅผ ๋ณด์žฅํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ท€๋‚ฉ๋ฒ•์— ์˜ํ•ด, ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ํ•ญ์ƒ \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\in[\alpha_{i},\beta_{i}]\)๋ฅผ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค.
์ฆ๋ช…: ์•Œ๊ณ ๋ฆฌ์ฆ˜ 3์€ ๊ฐ ๋ผ์šด๋“œ์—์„œ \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\in[\alpha_{i},\beta_{i}]\)๋ผ๋Š” ๋ถˆ๋ณ€์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด ์ฐธ์ด๊ณ  ์ฐธ ๋กœ์ง“ ๋ฒกํ„ฐ๊ฐ€ ํ•ญ์ƒ \(\mathcal{C}\)์— ๋†“์ธ๋‹ค๋Š” ๊ฒƒ์„ ๊ท€๋‚ฉ์ ์œผ๋กœ ์ฆ๋ช…ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ๋ ˆ๋งˆ์— ๋ช…์‹œ๋œ ๊ฐ€์ •์— ๋”ฐ๋ฅด๋ฉด, ์ด๊ฒƒ์€ ์ฒซ ๋ฒˆ์งธ ๋ผ์šด๋“œ๊ฐ€ ์‹œ์ž‘๋  ๋•Œ ๋ถ„๋ช…ํžˆ ์ฐธ์ž…๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด \(K<T\) ๋ผ์šด๋“œ ์ดํ›„์— ์ฐธ์ด๋ผ๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋ฉด, \(K+1\)-๋ฒˆ์งธ ๋ผ์šด๋“œ์—์„œ ์ถ”๊ฐ€๋œ ์ œ์•ฝ ์กฐ๊ฑด์€ ๋ชจ๋‘ ์ฐธ ๋กœ์ง“ ๋ฒกํ„ฐ์— ๋Œ€ํ•œ ์œ ํšจํ•œ ์ œ์•ฝ ์กฐ๊ฑด์ด๋ฉฐ, ํ† ํฐ \(k\)๋ฅผ ๋ฐ˜ํ™˜ํ•˜๋Š” API๋Š” ๋ชจ๋“  \(j\neq k\)์— ๋Œ€ํ•ด \(\mathsf{logit}_{k}+b_{k}\geq\mathsf{logit}_{j}+b_{j}\)๋ฅผ ๋ณด์žฅํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ท€๋‚ฉ์ ์œผ๋กœ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ํ•ญ์ƒ \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\in[\alpha_{i},\beta_{i}]\)๋ฅผ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค.
Machine translation, also known as automatic translation, is the process of converting one natural language (source language) into another natural language (target language) using a computer. It is a branch of computational linguistics, one of the ultimate goals of artificial intelligence, and has important scientific research value. Machine translation systems can be divided into two categories: rule-based and corpus-based. The former comprises a dictionary and a rule base, which collectively constitute the knowledge source. In contrast, the latter comprises a corpus that has been divided and labeled, and which does not require a dictionary or rules. Instead, it is based on statistical laws and most RALMs accomplish this task based on rules.
๊ธฐ๊ณ„ ๋ฒˆ์—ญ, ์ž๋™ ๋ฒˆ์—ญ์ด๋ผ๊ณ ๋„ ๋ถˆ๋ฆฌ๋ฉฐ, ์ปดํ“จํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•œ ์ž์—ฐ์–ด(์†Œ์Šค ์–ธ์–ด)๋ฅผ ๋‹ค๋ฅธ ์ž์—ฐ์–ด(ํƒ€๊ฒŸ ์–ธ์–ด)๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ๊ณผ์ •์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ธ๊ณต์ง€๋Šฅ์˜ ๊ถ๊ทน์ ์ธ ๋ชฉํ‘œ ์ค‘ ํ•˜๋‚˜์ธ ๊ณ„์‚ฐ ์–ธ์–ดํ•™์˜ ํ•œ ๋ถ„์•ผ๋กœ, ์ค‘์š”ํ•œ ๊ณผํ•™์  ์—ฐ๊ตฌ ๊ฐ€์น˜๋ฅผ ์ง€๋‹ˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ๊ณ„ ๋ฒˆ์—ญ ์‹œ์Šคํ…œ์€ ๊ทœ์น™ ๊ธฐ๋ฐ˜๊ณผ ์ฝ”ํผ์Šค ๊ธฐ๋ฐ˜์˜ ๋‘ ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „์ž๋Š” ์‚ฌ์ „๊ณผ ๊ทœ์น™ ๊ธฐ๋ฐ˜์„ ํฌํ•จํ•˜๋ฉฐ, ์ด๋“ค์ด ์ง‘ํ•ฉ์ ์œผ๋กœ ์ง€์‹ ์†Œ์Šค๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ํ›„์ž๋Š” ๋ถ„ํ• ๋˜๊ณ  ๋ผ๋ฒจ๋ง๋œ ์ฝ”ํผ์Šค๋ฅผ ํฌํ•จํ•˜๋ฉฐ, ์‚ฌ์ „์ด๋‚˜ ๊ทœ์น™์ด ํ•„์š”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋Œ€์‹  ํ†ต๊ณ„์  ๋ฒ•์น™์— ๊ธฐ๋ฐ˜ํ•˜๋ฉฐ ๋Œ€๋ถ€๋ถ„์˜ RALM์€ ์ด ์ž‘์—…์„ ๊ทœ์น™์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
์ž๋™ ๋ฒˆ์—ญ์ด๋ผ๊ณ ๋„ ํ•˜๋Š” ๊ธฐ๊ณ„ ๋ฒˆ์—ญ์€ ์ปดํ“จํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•œ ์ž์—ฐ์–ด(์›์–ด)๋ฅผ ๋‹ค๋ฅธ ์ž์—ฐ์–ด(๋Œ€์ƒ์–ด)๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ํ”„๋กœ์„ธ์Šค์ž…๋‹ˆ๋‹ค. ๊ธฐ๊ณ„ ๋ฒˆ์—ญ์€ ์ธ๊ณต ์ง€๋Šฅ์˜ ๊ถ๊ทน์ ์ธ ๋ชฉํ‘œ ์ค‘ ํ•˜๋‚˜์ธ ๊ณ„์‚ฐ ์–ธ์–ดํ•™์˜ ํ•œ ๋ถ„์•ผ์ด๋ฉฐ, ์ค‘์š”ํ•œ ๊ณผํ•™์  ์—ฐ๊ตฌ ๊ฐ€์น˜๋ฅผ ์ง€๋‹ˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ๊ณ„ ๋ฒˆ์—ญ ์‹œ์Šคํ…œ์€ ๊ทœ์น™ ๊ธฐ๋ฐ˜๊ณผ ๋ง๋ญ‰์น˜ ๊ธฐ๋ฐ˜์˜ ๋‘ ๊ฐ€์ง€๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „์ž๋Š” ์‚ฌ์ „๊ณผ ๊ทœ์น™ ๋ฒ ์ด์Šค๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ์ด ๋‘ ๊ฐ€์ง€๊ฐ€ ํ•ฉ์ณ์ ธ ์ง€์‹ ์†Œ์Šค๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด ํ›„์ž๋Š” ์‚ฌ์ „์ด๋‚˜ ๊ทœ์น™์ด ํ•„์š”ํ•˜์ง€ ์•Š๊ณ  ํ†ต๊ณ„์  ๋ฒ•์น™์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ๋ถ„ ๋ฐ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ๋ง๋ญ‰์น˜๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋Œ€์‹  ํ†ต๊ณ„์  ๋ฒ•์น™์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ, ๋Œ€๋ถ€๋ถ„์˜ RLM์€ ๊ทœ์น™์— ๋”ฐ๋ผ ์ด ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
Tail phenomenaSelf-Instruct depends on LMs, and it will inherit all the limitations that carry over with LMs. As recent studies have shown (Razeghi et al., 2022; Kandpal et al., 2022), _tail phenomena_ pose a serious challenge to the success of LMs. In other words, LMs' largest gains correspond to the frequent uses of languages (head of the language use distribution), and there might be minimal gains in the low-frequency contexts. Similarly, in the context of this work, it would not be surprising if the majority of the gains by Self-Instruct are skewed toward tasks or instructions that present more frequently in the pretraining corpus. As a consequence, the approach might show brittleness with respect to uncommon and creative instructions.
Tail phenomenaSelf-Instruct๋Š” LMs์— ์˜์กดํ•˜๋ฉฐ, LMs์™€ ํ•จ๊ป˜ ๋ฐœ์ƒํ•˜๋Š” ๋ชจ๋“  ํ•œ๊ณ„๋ฅผ ๊ทธ๋Œ€๋กœ ๋ฌผ๋ ค๋ฐ›๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ๋“ค(Razeghi et al., 2022; Kandpal et al., 2022)์ด ๋ณด์—ฌ์ฃผ๋“ฏ์ด, _tail phenomena_๋Š” LMs์˜ ์„ฑ๊ณต์— ์‹ฌ๊ฐํ•œ ๋„์ „์„ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์‹œ ๋งํ•ด, LMs์˜ ๊ฐ€์žฅ ํฐ ์„ฑ๊ณผ๋Š” ์–ธ์–ด ์‚ฌ์šฉ ๋ถ„ํฌ์˜ ์ƒ์œ„ ๋ถ€๋ถ„(์ž์ฃผ ์‚ฌ์šฉ๋˜๋Š” ์–ธ์–ด)์— ํ•ด๋‹นํ•˜๋ฉฐ, ์ €๋นˆ๋„ ๋งฅ๋ฝ์—์„œ๋Š” ์„ฑ๊ณผ๊ฐ€ ๋ฏธ๋ฏธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ์ด ์—ฐ๊ตฌ์˜ ๋งฅ๋ฝ์—์„œ๋„ Self-Instruct์˜ ๋Œ€๋ถ€๋ถ„์˜ ์„ฑ๊ณผ๊ฐ€ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์—์„œ ๋” ์ž์ฃผ ๋‚˜ํƒ€๋‚˜๋Š” ์ž‘์—…์ด๋‚˜ ์ง€์นจ์— ์น˜์šฐ์ณ ์žˆ์„ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ์ด ์ ‘๊ทผ๋ฒ•์€ ๋“œ๋ฌธ ๊ฒฝ์šฐ๋‚˜ ์ฐฝ์˜์ ์ธ ์ง€์นจ์— ๋Œ€ํ•ด ์ทจ์•ฝ์„ฑ์„ ๋ณด์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ผฌ๋ฆฌ ํ˜„์ƒSelf-Instruct๋Š” LM์— ์˜์กดํ•˜๋ฉฐ, LM์— ๋”ฐ๋ผ ์ด์–ด์ง€๋Š” ๋ชจ๋“  ํ•œ๊ณ„๋ฅผ ์ƒ์†๋ฐ›๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ ๋ฐํ˜€์ง„ ๋ฐ”์™€ ๊ฐ™์ด(Razeghi et al., 2022; Kandpal et al., 2022), _๊ผฌ๋ฆฌ ํ˜„์ƒ_์€ LM์˜ ์„ฑ๊ณต์— ์‹ฌ๊ฐํ•œ ๋„์ „์„ ์•ˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ฆ‰, LM์˜ ๊ฐ€์žฅ ํฐ ์ด์ ์€ ์–ธ์–ด ์‚ฌ์šฉ ๋นˆ๋„๊ฐ€ ๋†’์€ ๊ฒฝ์šฐ(์–ธ์–ด ์‚ฌ์šฉ ๋ถ„ํฌ์˜ ๋จธ๋ฆฌ)์— ํ•ด๋‹นํ•˜๋ฉฐ, ๋นˆ๋„๊ฐ€ ๋‚ฎ์€ ๋ฌธ๋งฅ์—์„œ๋Š” ์ด๋“์ด ๊ฑฐ์˜ ์—†์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ์ด ์ž‘์—…์˜ ๋งฅ๋ฝ์—์„œ Self-Instruct์˜ ๋Œ€๋ถ€๋ถ„์˜ ์ด์ ์ด ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์—์„œ ๋” ์ž์ฃผ ๋“ฑ์žฅํ•˜๋Š” ์ž‘์—…์ด๋‚˜ ์ง€์นจ์— ์น˜์šฐ์ณ ์žˆ๋‹ค๋ฉด ๋†€๋ผ์šด ์ผ์ด ์•„๋‹™๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ, ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ํ”ํ•˜์ง€ ์•Š๊ณ  ์ฐฝ์˜์ ์ธ ์ง€์นจ์— ๋Œ€ํ•ด ์ทจ์•ฝ์„ฑ์„ ๋ณด์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Zhihuis a vibrant question-and-answer platform where users can ask and answer questions on a wide range of topics, making it a comprehensive repository of knowledge and insights. Zhihu encourages its users to provide well-thought-out answers that are informative and reflective of expert knowledge or personal experience. However, the absence of a review mechanism for answers on Zhihu leads to a large volume of content that falls short of our quality standards. To filter low quality answers, we selected answers with more than 50 upvotes, then filtering out content containing sensitive or harmful keywords using a rule-based method. Subsequently, we employed GPT-4 to score the responses on a scale of 1-10, retaining those with scores above 8.
Zhihu๋Š” ์‚ฌ์šฉ์ž๊ฐ€ ๋‹ค์–‘ํ•œ ์ฃผ์ œ์— ๋Œ€ํ•ด ์งˆ๋ฌธํ•˜๊ณ  ๋‹ต๋ณ€ํ•  ์ˆ˜ ์žˆ๋Š” ํ™œ๊ธฐ์ฐฌ ์งˆ๋ฌธ ๋ฐ ๋‹ต๋ณ€ ํ”Œ๋žซํผ์œผ๋กœ, ์ง€์‹๊ณผ ํ†ต์ฐฐ์˜ ์ข…ํ•ฉ์ ์ธ ์ €์žฅ์†Œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. Zhihu๋Š” ์‚ฌ์šฉ์ž๋“ค์ด ์ „๋ฌธ๊ฐ€ ์ง€์‹์ด๋‚˜ ๊ฐœ์ธ ๊ฒฝํ—˜์„ ๋ฐ˜์˜ํ•œ ์ •๋ณด์„ฑ ์žˆ๊ณ  ์‹ ์ค‘ํ•œ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋„๋ก ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Zhihu์—๋Š” ๋‹ต๋ณ€์— ๋Œ€ํ•œ ๊ฒ€ํ†  ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ์—†๊ธฐ ๋•Œ๋ฌธ์— ํ’ˆ์งˆ ๊ธฐ์ค€์— ๋ฏธ์น˜์ง€ ๋ชปํ•˜๋Š” ๋งŽ์€ ์ฝ˜ํ…์ธ ๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค. ์ €ํ’ˆ์งˆ ๋‹ต๋ณ€์„ ํ•„ํ„ฐ๋งํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” 50๊ฐœ ์ด์ƒ์˜ ์ถ”์ฒœ์„ ๋ฐ›์€ ๋‹ต๋ณ€์„ ์„ ํƒํ•œ ํ›„, ๊ทœ์น™ ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏผ๊ฐํ•˜๊ฑฐ๋‚˜ ์œ ํ•ดํ•œ ํ‚ค์›Œ๋“œ๋ฅผ ํฌํ•จํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ํ•„ํ„ฐ๋งํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ํ›„, GPT-4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต๋ณ€์„ 1-10 ์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ณ , 8์  ์ด์ƒ์„ ๋ฐ›์€ ๋‹ต๋ณ€๋งŒ์„ ์œ ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค.
Zhihu๋Š” ์‚ฌ์šฉ์ž๊ฐ€ ๋‹ค์–‘ํ•œ ์ฃผ์ œ์— ๋Œ€ํ•ด ์งˆ๋ฌธํ•˜๊ณ  ๋‹ต๋ณ€ํ•  ์ˆ˜ ์žˆ๋Š” ํ™œ๊ธฐ์ฐฌ ์งˆ์˜์‘๋‹ต ํ”Œ๋žซํผ์œผ๋กœ, ์ง€์‹๊ณผ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ข…ํ•ฉ์ ์œผ๋กœ ์ €์žฅํ•˜๋Š” ์ €์žฅ์†Œ์ž…๋‹ˆ๋‹ค. Zhihu๋Š” ์‚ฌ์šฉ์ž๊ฐ€ ์ •๋ณด์ ์ด๊ณ  ์ „๋ฌธ ์ง€์‹์ด๋‚˜ ๊ฐœ์ธ์ ์ธ ๊ฒฝํ—˜์„ ๋ฐ˜์˜ํ•œ ์‚ฌ๊ณ  ๊นŠ์€ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜๋„๋ก ์žฅ๋ คํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Zhihu์˜ ๋‹ต๋ณ€์— ๋Œ€ํ•œ ๊ฒ€ํ†  ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ์—†๊ธฐ ๋•Œ๋ฌธ์— ํ’ˆ์งˆ ๊ธฐ์ค€์— ๋ฏธ์น˜์ง€ ๋ชปํ•˜๋Š” ์ฝ˜ํ…์ธ ๊ฐ€ ๋Œ€๋Ÿ‰์œผ๋กœ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ํ’ˆ์งˆ์ด ๋‚ฎ์€ ๋‹ต๋ณ€์„ ํ•„ํ„ฐ๋งํ•˜๊ธฐ ์œ„ํ•ด ์—…๋ณดํŠธ๊ฐ€ 50๊ฐœ ์ด์ƒ์ธ ๋‹ต๋ณ€์„ ์„ ๋ณ„ํ•œ ๋‹ค์Œ ๊ทœ์น™ ๊ธฐ๋ฐ˜ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏผ๊ฐํ•˜๊ฑฐ๋‚˜ ์œ ํ•ดํ•œ ํ‚ค์›Œ๋“œ๊ฐ€ ํฌํ•จ๋œ ์ฝ˜ํ…์ธ ๋ฅผ ๊ฑธ๋Ÿฌ๋ƒˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ GPT-4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‘๋‹ต์„ 1~10์  ์ฒ™๋„๋กœ ํ‰๊ฐ€ํ•˜์—ฌ 8์  ์ด์ƒ์ธ ์‘๋‹ต์„ ์œ ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค.
Experimental results verify the reliability of our data mixing law and prediction pipeline. By predicting the overall validation loss, we optimize the training mixture of RedPajama for a 1B model trained on 100B tokens and achieve performance comparable to a model trained on default mixture for 48% more steps. The prediction on domain-specific validation sets also offers plausible references to the balance of model capabilities. Further applying our data mixing law to continual pretraining can accurately find the proportion that avoids catastrophic forgetting (French, 1999; Kirkpatrick et al., 2017; Luo et al., 2023), revealing the prospect of applying data mixing laws to guide a multi-stage pertaining, and thus a dynamic data schedule.
์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ์šฐ๋ฆฌ์˜ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™๊ณผ ์˜ˆ์ธก ํŒŒ์ดํ”„๋ผ์ธ์˜ ์‹ ๋ขฐ์„ฑ์„ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ๊ฒ€์ฆ ์†์‹ค์„ ์˜ˆ์ธกํ•จ์œผ๋กœ์จ, ์šฐ๋ฆฌ๋Š” 100B ํ† ํฐ์œผ๋กœ ํ›ˆ๋ จ๋œ 1B ๋ชจ๋ธ์„ ์œ„ํ•œ RedPajama์˜ ํ›ˆ๋ จ ํ˜ผํ•ฉ์„ ์ตœ์ ํ™”ํ•˜๊ณ , ๊ธฐ๋ณธ ํ˜ผํ•ฉ์œผ๋กœ 48% ๋” ๋งŽ์€ ๋‹จ๊ณ„์—์„œ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•  ๋งŒํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋„๋ฉ”์ธ๋ณ„ ๊ฒ€์ฆ ์„ธํŠธ์— ๋Œ€ํ•œ ์˜ˆ์ธก์€ ๋ชจ๋ธ ๋Šฅ๋ ฅ์˜ ๊ท ํ˜•์— ๋Œ€ํ•œ ํƒ€๋‹นํ•œ ์ฐธ์กฐ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์„ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์— ์ ์šฉํ•˜๋ฉด, ์น˜๋ช…์ ์ธ ๋ง๊ฐ์„ ํ”ผํ•  ์ˆ˜ ์žˆ๋Š” ๋น„์œจ์„ ์ •ํ™•ํ•˜๊ฒŒ ์ฐพ์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ (French, 1999; Kirkpatrick et al., 2017; Luo et al., 2023), ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์„ ๋‹ค๋‹จ๊ณ„ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์•ˆ๋‚ดํ•˜๊ณ , ๋”ฐ๋ผ์„œ ๋™์  ๋ฐ์ดํ„ฐ ์Šค์ผ€์ค„์„ ์ ์šฉํ•˜๋Š” ์ „๋ง์„ ๋“œ๋Ÿฌ๋ƒ…๋‹ˆ๋‹ค.
์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™๊ณผ ์˜ˆ์ธก ํŒŒ์ดํ”„๋ผ์ธ์˜ ์‹ ๋ขฐ์„ฑ์„ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์†์‹ค์„ ์˜ˆ์ธกํ•˜์—ฌ 100B ํ† ํฐ์— ๋Œ€ํ•ด ํ•™์Šต๋œ 1B ๋ชจ๋ธ์— ๋Œ€ํ•œ RedPajama์˜ ํ•™์Šต ํ˜ผํ•ฉ์„ ์ตœ์ ํ™”ํ•˜๊ณ , ๊ธฐ๋ณธ ํ˜ผํ•ฉ์— ๋Œ€ํ•ด 48% ๋” ๋งŽ์€ ๋‹จ๊ณ„ ๋™์•ˆ ํ•™์Šต๋œ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋„๋ฉ”์ธ๋ณ„ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์— ๋Œ€ํ•œ ์˜ˆ์ธก์€ ๋ชจ๋ธ ๊ธฐ๋Šฅ์˜ ๊ท ํ˜•์— ๋Œ€ํ•œ ๊ทธ๋Ÿด๋“ฏํ•œ ์ฐธ์กฐ๋„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์— ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์„ ์ถ”๊ฐ€๋กœ ์ ์šฉํ•˜๋ฉด ํŒŒ๊ตญ์  ๋ง๊ฐ(French, 1999; Kirkpatrick et al., 2017; Luo et al., 2023)์„ ํ”ผํ•  ์ˆ˜ ์žˆ๋Š” ๋น„์œจ์„ ์ •ํ™•ํ•˜๊ฒŒ ์ฐพ์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋‹ค๋‹จ๊ณ„ ๊ด€๋ จ์„ฑ ๋ฐ ๋™์  ๋ฐ์ดํ„ฐ ์ผ์ •์„ ์•ˆ๋‚ดํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์„ ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ „๋ง์„ ๋“œ๋Ÿฌ๋ƒ…๋‹ˆ๋‹ค.
Optimization of Memory and Storage Architecture. In the context of agent collaboration and communication, the future design of memory and storage systems can adopt a shared approach, enabling the sharing of memory and storage between agents. Such an architecture would enable agents to access a communal pool of memory and storage, thereby improving the agents' decision-making ability since one agent can benefit from other agents' memory or storage. Moreover, future work can explore hierarchical storage solutions, designed to optimize data retrieval and storage efficiency. This could involve prioritizing quicker access and reduced storage allocation for frequently accessed data, and vice versa for less frequently accessed information.
๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์Šคํ† ๋ฆฌ์ง€ ์•„ํ‚คํ…์ฒ˜์˜ ์ตœ์ ํ™”. ์—์ด์ „ํŠธ ํ˜‘์—… ๋ฐ ํ†ต์‹ ์˜ ๋งฅ๋ฝ์—์„œ, ๋ฏธ๋ž˜์˜ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์Šคํ† ๋ฆฌ์ง€ ์‹œ์Šคํ…œ ์„ค๊ณ„๋Š” ๊ณต์œ  ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜์—ฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์Šคํ† ๋ฆฌ์ง€ ๊ณต์œ ๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์•„ํ‚คํ…์ฒ˜๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๊ณต๋™์˜ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์Šคํ† ๋ฆฌ์ง€ ํ’€์— ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•˜์—ฌ, ํ•œ ์—์ด์ „ํŠธ๊ฐ€ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์˜ ๋ฉ”๋ชจ๋ฆฌ๋‚˜ ์Šคํ† ๋ฆฌ์ง€๋กœ๋ถ€ํ„ฐ ์ด์ต์„ ์–ป์„ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์—์ด์ „ํŠธ์˜ ์˜์‚ฌ ๊ฒฐ์ • ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋ฏธ๋ž˜ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฐ์ดํ„ฐ ๊ฒ€์ƒ‰ ๋ฐ ์ €์žฅ ํšจ์œจ์„ฑ์„ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๊ณ„์ธต์  ์Šคํ† ๋ฆฌ์ง€ ์†”๋ฃจ์…˜์„ ํƒ๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ž์ฃผ ์ ‘๊ทผํ•˜๋Š” ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋” ๋น ๋ฅธ ์ ‘๊ทผ๊ณผ ์ค„์–ด๋“  ์ €์žฅ ํ• ๋‹น์„ ์šฐ์„ ์‹œํ•˜๊ณ , ๋œ ์ž์ฃผ ์ ‘๊ทผํ•˜๋Š” ์ •๋ณด์— ๋Œ€ํ•ด์„œ๋Š” ๊ทธ ๋ฐ˜๋Œ€๋ฅผ ์ ์šฉํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์Šคํ† ๋ฆฌ์ง€ ์•„ํ‚คํ…์ฒ˜ ์ตœ์ ํ™”. ์—์ด์ „ํŠธ ํ˜‘์—… ๋ฐ ํ†ต์‹ ์˜ ๋งฅ๋ฝ์—์„œ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์Šคํ† ๋ฆฌ์ง€ ์‹œ์Šคํ…œ์˜ ๋ฏธ๋ž˜ ์„ค๊ณ„๋Š” ๊ณต์œ  ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜์—ฌ ์—์ด์ „ํŠธ ๊ฐ„์— ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์Šคํ† ๋ฆฌ์ง€๋ฅผ ๊ณต์œ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ†ตํ•ด ์—์ด์ „ํŠธ๋Š” ๊ณต๋™์˜ ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ์Šคํ† ๋ฆฌ์ง€ ํ’€์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํ•œ ์—์ด์ „ํŠธ๊ฐ€ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์˜ ๋ฉ”๋ชจ๋ฆฌ ๋˜๋Š” ์Šคํ† ๋ฆฌ์ง€์—์„œ ์ด์ ์„ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์—์ด์ „ํŠธ์˜ ์˜์‚ฌ ๊ฒฐ์ • ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ–ฅํ›„ ์ž‘์—…์—์„œ๋Š” ๋ฐ์ดํ„ฐ ๊ฒ€์ƒ‰ ๋ฐ ์ €์žฅ ํšจ์œจ์„ฑ์„ ์ตœ์ ํ™”ํ•˜๋„๋ก ์„ค๊ณ„๋œ ๊ณ„์ธต์  ์Šคํ† ๋ฆฌ์ง€ ์†”๋ฃจ์…˜์„ ํƒ์ƒ‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ž์ฃผ ์•ก์„ธ์Šคํ•˜๋Š” ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋” ๋น ๋ฅธ ์•ก์„ธ์Šค์™€ ๋” ์ ์€ ์Šคํ† ๋ฆฌ์ง€ ํ• ๋‹น์„ ์šฐ์„ ์ˆœ์œ„๋กœ ์‚ผ๊ณ , ๋œ ์ž์ฃผ ์•ก์„ธ์Šคํ•˜๋Š” ์ •๋ณด์— ๋Œ€ํ•ด์„œ๋Š” ๊ทธ ๋ฐ˜๋Œ€์˜ ์šฐ์„ ์ˆœ์œ„๋ฅผ ์ •ํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
We also examine the \(\mathsf{bioS^{simple}}(N)\) datasets, which are identical to \(\mathsf{bioS}(N)\) except that each individual's knowledge is stored in a fixed ordering of six fixed sentences (see Section 2.2). Allen-Zhu and Li [3] found that in such cases, the knowledge data are memorizable but nearly 0% extractable. As shown in Figure 11(a), in these instances, the capacity ratio slightly decreases compared to Figure 1(a). This implies, in this ideal setting, adding data diversity -- by rewriting the same knowledge multiple times using different writing templates -- not only enhances the model's ability to extract knowledge, as noted by [3], but also, surprisingly, _increases_ the model's capacity, as observed in this study.
์šฐ๋ฆฌ๋Š” ๋˜ํ•œ \(\mathsf{bioS^{simple}}(N)\) ๋ฐ์ดํ„ฐ์…‹์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ์…‹์€ ๊ฐ ๊ฐœ์ธ์˜ ์ง€์‹์ด ๊ณ ์ •๋œ ์—ฌ์„ฏ ๊ฐœ์˜ ๋ฌธ์žฅ ์ˆœ์„œ๋กœ ์ €์žฅ๋œ๋‹ค๋Š” ์ ์„ ์ œ์™ธํ•˜๋ฉด \(\mathsf{bioS}(N)\)์™€ ๋™์ผํ•ฉ๋‹ˆ๋‹ค(์„น์…˜ 2.2 ์ฐธ์กฐ). Allen-Zhu์™€ Li [3]๋Š” ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ ์ง€์‹ ๋ฐ์ดํ„ฐ๊ฐ€ ์•”๊ธฐ ๊ฐ€๋Šฅํ•˜์ง€๋งŒ ๊ฑฐ์˜ 0% ์ถ”์ถœ ๊ฐ€๋Šฅํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. Figure 11(a)์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ ์šฉ๋Ÿ‰ ๋น„์œจ์€ Figure 1(a)์™€ ๋น„๊ตํ–ˆ์„ ๋•Œ ์•ฝ๊ฐ„ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ด ์ด์ƒ์ ์ธ ์„ค์ •์—์„œ, [3]์ด ์–ธ๊ธ‰ํ•œ ๋ฐ”์™€ ๊ฐ™์ด ๋™์ผํ•œ ์ง€์‹์„ ์—ฌ๋Ÿฌ ๋ฒˆ ๋‹ค๋ฅธ ์ž‘์„ฑ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์‹œ ์ž‘์„ฑํ•จ์œผ๋กœ์จ ๋ฐ์ดํ„ฐ ๋‹ค์–‘์„ฑ์„ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ์ด ๋ชจ๋ธ์˜ ์ง€์‹ ์ถ”์ถœ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ, ๋†€๋ž๊ฒŒ๋„ ์ด ์—ฐ๊ตฌ์—์„œ ๊ด€์ฐฐ๋œ ๋ฐ”์™€ ๊ฐ™์ด ๋ชจ๋ธ์˜ ์šฉ๋Ÿ‰์„ _์ฆ๊ฐ€_ ์‹œํ‚จ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
๋˜ํ•œ, ๊ฐ ๊ฐœ์ธ์˜ ์ง€์‹์ด 6๊ฐœ์˜ ๊ณ ์ • ๋ฌธ์žฅ์œผ๋กœ ๊ณ ์ •๋œ ์ˆœ์„œ๋กœ ์ €์žฅ๋˜์–ด ์žˆ๋Š” ์ ์„ ์ œ์™ธํ•˜๋ฉด \(\mathsf{bioS}(N)\)์™€ ๋™์ผํ•œ \(\mathsf{bioS^{simple}}(N)\) ๋ฐ์ดํ„ฐ ์„ธํŠธ๋„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค(2.2์ ˆ ์ฐธ์กฐ). Allen-Zhu์™€ Li [3]๋Š” ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ ์ง€์‹ ๋ฐ์ดํ„ฐ๋Š” ๊ธฐ์–ตํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ์ถ”์ถœ ๊ฐ€๋Šฅํ•œ ๋น„์œจ์€ ๊ฑฐ์˜ 0%๋ผ๊ณ  ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 11(a)์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ ์šฉ๋Ÿ‰ ๋น„์œจ์€ ๊ทธ๋ฆผ 1(a)์— ๋น„ํ•ด ์•ฝ๊ฐ„ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ด ์ด์ƒ์ ์ธ ์„ค์ •์—์„œ ๋ฐ์ดํ„ฐ ๋‹ค์–‘์„ฑ์„ ์ถ”๊ฐ€ํ•˜์—ฌ ๋™์ผํ•œ ์ง€์‹์„ ๋‹ค๋ฅธ ์ž‘์„ฑ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ ์—ฌ๋Ÿฌ ๋ฒˆ ๋‹ค์‹œ ์ž‘์„ฑํ•˜๋ฉด [3]์—์„œ ์–ธ๊ธ‰ํ•œ ๊ฒƒ์ฒ˜๋Ÿผ ๋ชจ๋ธ์˜ ์ง€์‹ ์ถ”์ถœ ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ด ์—ฐ๊ตฌ์—์„œ ๊ด€์ฐฐ๋œ ๊ฒƒ์ฒ˜๋Ÿผ ๋†€๋ž๊ฒŒ๋„ ๋ชจ๋ธ์˜ ์šฉ๋Ÿ‰๋„ _์ฆ๊ฐ€_ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
These results clearly show that _not all data is equal for continual pre-training_. In fact, all the data used in efficient DACP methods (10%) is a subset of the data in DACP. Since DACP's (100%) performance is lower than ETS-DACP/ETA-DACP-ent, adding more data on top of highly similar or high entropy data actually hurts the performance. The difference in results between hard and soft sampling adds more evidence to this observation. While there is variability across tasks, on an average, adding examples from outside the top decile of metrics hurts the performance with the notable exception of ETS-DACP-com which is a combination of all three metrics. Hence, we should carefully curate the data for any domain continual pre-training.
์ด ๊ฒฐ๊ณผ๋Š” _๋ชจ๋“  ๋ฐ์ดํ„ฐ๊ฐ€ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์— ๋™์ผํ•˜๊ฒŒ ์œ ํšจํ•˜์ง€ ์•Š๋‹ค_๋Š” ๊ฒƒ์„ ๋ช…ํ™•ํžˆ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์‚ฌ์‹ค, ํšจ์œจ์ ์ธ DACP ๋ฐฉ๋ฒ•์—์„œ ์‚ฌ์šฉ๋œ ๋ชจ๋“  ๋ฐ์ดํ„ฐ(10%)๋Š” DACP ๋ฐ์ดํ„ฐ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์ž…๋‹ˆ๋‹ค. DACP(100%)์˜ ์„ฑ๋Šฅ์ด ETS-DACP/ETA-DACP-ent๋ณด๋‹ค ๋‚ฎ๊ธฐ ๋•Œ๋ฌธ์—, ๋งค์šฐ ์œ ์‚ฌํ•˜๊ฑฐ๋‚˜ ๋†’์€ ์—”ํŠธ๋กœํ”ผ ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ์‹ค์ œ๋กœ ์„ฑ๋Šฅ์„ ์ €ํ•˜์‹œํ‚ต๋‹ˆ๋‹ค. ํ•˜๋“œ ์ƒ˜ํ”Œ๋ง๊ณผ ์†Œํ”„ํŠธ ์ƒ˜ํ”Œ๋ง ๊ฐ„์˜ ๊ฒฐ๊ณผ ์ฐจ์ด๋Š” ์ด ๊ด€์ฐฐ์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์ฆ๊ฑฐ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ž‘์—… ๊ฐ„์˜ ๋ณ€๋™์„ฑ์ด ์กด์žฌํ•˜์ง€๋งŒ, ํ‰๊ท ์ ์œผ๋กœ ์ƒ์œ„ 10%์˜ ๋ฉ”ํŠธ๋ฆญ ์™ธ๋ถ€์—์„œ ์˜ˆ์ œ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ์„ฑ๋Šฅ์„ ์ €ํ•˜์‹œํ‚ต๋‹ˆ๋‹ค. ๋‹จ, ๋ชจ๋“  ์„ธ ๊ฐ€์ง€ ๋ฉ”ํŠธ๋ฆญ์„ ๊ฒฐํ•ฉํ•œ ETS-DACP-com์€ ์˜ˆ์™ธ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” ๋„๋ฉ”์ธ ์ง€์†์  ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์œ„ํ•ด ๋ฐ์ดํ„ฐ๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ ๋ณ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” _๋ชจ๋“  ๋ฐ์ดํ„ฐ๊ฐ€ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์— ๋™์ผํ•˜๊ฒŒ ์ ์šฉ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋‹ค_๋ผ๋Š” ๊ฒƒ์„ ๋ช…ํ™•ํ•˜๊ฒŒ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ํšจ์œจ์ ์ธ DACP ๋ฐฉ๋ฒ•์— ์‚ฌ์šฉ๋˜๋Š” ๋ชจ๋“  ๋ฐ์ดํ„ฐ(10%)๋Š” DACP ๋ฐ์ดํ„ฐ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์ž…๋‹ˆ๋‹ค. DACP(100%)์˜ ์„ฑ๋Šฅ์ด ETS-DACP/ETA-DACP-ent๋ณด๋‹ค ๋‚ฎ๊ธฐ ๋•Œ๋ฌธ์— ์œ ์‚ฌ๋„๊ฐ€ ๋†’๊ฑฐ๋‚˜ ์—”ํŠธ๋กœํ”ผ๊ฐ€ ๋†’์€ ๋ฐ์ดํ„ฐ ์œ„์— ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•˜๋ฉด ์‹ค์ œ๋กœ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค. ํ•˜๋“œ ์ƒ˜ํ”Œ๋ง๊ณผ ์†Œํ”„ํŠธ ์ƒ˜ํ”Œ๋ง ๊ฐ„์˜ ๊ฒฐ๊ณผ ์ฐจ์ด๋Š” ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์— ๋Œ€ํ•œ ๋” ๋งŽ์€ ์ฆ๊ฑฐ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ž‘์—…๋งˆ๋‹ค ์ฐจ์ด๋Š” ์žˆ์ง€๋งŒ, ํ‰๊ท ์ ์œผ๋กœ ๋ฉ”ํŠธ๋ฆญ ์ƒ์œ„ 10๋ถ„์œ„์ˆ˜ ๋ฐ–์˜ ์˜ˆ์ œ๋ฅผ ์ถ”๊ฐ€ํ•˜๋ฉด ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๋Š”๋ฐ, ์„ธ ๊ฐ€์ง€ ๋ฉ”ํŠธ๋ฆญ์„ ๋ชจ๋‘ ์กฐํ•ฉํ•œ ETS-DACP-com์„ ์ œ์™ธํ•˜๊ณ ๋Š” ์„ฑ๋Šฅ์ด ์ €ํ•˜๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋„๋ฉ”์ธ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•ด์„œ๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ํ๋ ˆ์ด์…˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
One popular benchmark that uses real-world data is WildBench, which is sourced from the WildChat dataset of 570,000 real conversations with ChatGPT [35]. Because of this, it covers a huge breadth of tasks and prompts. While WildBench covers a wide range of topics, most other real-world benchmarks focus on a specific task. For example, SWE-bench is a benchmark that uses a set of real-world issues raised on GitHub for software engineering tasks in Python [13]. This can be very helpful when evaluating agents designed to write Python code and provides a sense for how well agents can reason about code related problems; however, it is less informative when trying to understand agent capabilities involving other programming languages.
์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์ธ๊ธฐ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ ์ค‘ ํ•˜๋‚˜๋Š” WildBench๋กœ, ChatGPT์™€์˜ 570,000๊ฐœ์˜ ์‹ค์ œ ๋Œ€ํ™”๋กœ ๊ตฌ์„ฑ๋œ WildChat ๋ฐ์ดํ„ฐ์…‹์—์„œ ๊ฐ€์ ธ์™”์Šต๋‹ˆ๋‹ค [35]. ์ด ๋•Œ๋ฌธ์— ๋‹ค์–‘ํ•œ ์ž‘์—…๊ณผ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํฌ๊ด„ํ•ฉ๋‹ˆ๋‹ค. WildBench๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ์ฃผ์ œ๋ฅผ ๋‹ค๋ฃจ์ง€๋งŒ, ๋Œ€๋ถ€๋ถ„์˜ ๋‹ค๋ฅธ ์‹ค์ œ ๋ฒค์น˜๋งˆํฌ๋Š” ํŠน์ • ์ž‘์—…์— ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, SWE-bench๋Š” Python์œผ๋กœ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด๋ง ์ž‘์—…์„ ์œ„ํ•ด GitHub์—์„œ ์ œ๊ธฐ๋œ ์‹ค์ œ ๋ฌธ์ œ๋“ค์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฒค์น˜๋งˆํฌ์ž…๋‹ˆ๋‹ค [13]. ์ด๋Š” Python ์ฝ”๋“œ๋ฅผ ์ž‘์„ฑํ•˜๋„๋ก ์„ค๊ณ„๋œ ์—์ด์ „ํŠธ๋ฅผ ํ‰๊ฐ€ํ•  ๋•Œ ๋งค์šฐ ์œ ์šฉํ•˜๋ฉฐ, ์—์ด์ „ํŠธ๊ฐ€ ์ฝ”๋“œ ๊ด€๋ จ ๋ฌธ์ œ๋ฅผ ์–ผ๋งˆ๋‚˜ ์ž˜ ์ดํ•ดํ•˜๋Š”์ง€์— ๋Œ€ํ•œ ๊ฐ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹ค๋ฅธ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์™€ ๊ด€๋ จ๋œ ์—์ด์ „ํŠธ์˜ ๋Šฅ๋ ฅ์„ ์ดํ•ดํ•˜๋ ค๊ณ  ํ•  ๋•Œ๋Š” ๋œ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค.
์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์ธ๊ธฐ ์žˆ๋Š” ๋ฒค์น˜๋งˆํฌ ์ค‘ ํ•˜๋‚˜๋Š” WildChat์˜ 570,000๊ฐœ์˜ ์‹ค์ œ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ WildBench[35]์—์„œ ๊ฐ€์ ธ์˜จ ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—…๊ณผ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋‹ค๋ฃฐ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. WildBench๋Š” ๋‹ค์–‘ํ•œ ์ฃผ์ œ๋ฅผ ๋‹ค๋ฃจ์ง€๋งŒ, ๋Œ€๋ถ€๋ถ„์˜ ๋‹ค๋ฅธ ์‹ค์ œ ๋ฒค์น˜๋งˆํฌ๋Š” ํŠน์ • ์ž‘์—…์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, SWE-bench๋Š” Python์˜ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด๋ง ์ž‘์—…์„ ์œ„ํ•ด GitHub์—์„œ ์ œ๊ธฐ๋œ ์‹ค์ œ ์ด์Šˆ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ฒค์น˜๋งˆํฌ์ž…๋‹ˆ๋‹ค[13]. ์ด๋Š” ํŒŒ์ด์ฌ ์ฝ”๋“œ ์ž‘์„ฑ์„ ์œ„ํ•ด ์„ค๊ณ„๋œ ์—์ด์ „ํŠธ๋ฅผ ํ‰๊ฐ€ํ•  ๋•Œ ๋งค์šฐ ์œ ์šฉํ•˜๋ฉฐ, ์—์ด์ „ํŠธ๊ฐ€ ์ฝ”๋“œ ๊ด€๋ จ ๋ฌธ์ œ๋ฅผ ์–ผ๋งˆ๋‚˜ ์ž˜ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ๋Š”์ง€์— ๋Œ€ํ•œ ๊ฐ์„ ์žก์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ๋‹ค๋ฅธ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์™€ ๊ด€๋ จ๋œ ์—์ด์ „ํŠธ ๊ธฐ๋Šฅ์„ ์ดํ•ดํ•˜๋ ค๋Š” ๊ฒฝ์šฐ์—๋Š” ์ •๋ณด๊ฐ€ ๋œ ํ’๋ถ€ํ•ฉ๋‹ˆ๋‹ค.
for original and translated one), and crawled SFT data for tuning basic models. For the Plus version, we expand the dataset to approximately 4M to 4.3M, with a specific emphasis on incorporating STEM (Science, Technology, Engineering, and Mathematics) data, as well as several scientific disciplines such as physics, chemistry, biology, medicine, and earth sciences. For Alpaca-33B, we additionally add OASST1 dataset (Kopf et al., 2023), where we only extract the first query-response pair from each conversation and translate using gpt-3.5-turbo API, resulting in roughly 20K data (original and translated one). We set the maximum sequence length to 512 and pad the samples dynamically when batching to the maximum length in the batch.
์›๋ณธ ๋ฐ ๋ฒˆ์—ญ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ํฌํ•จํ•˜์—ฌ ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ํŠœ๋‹ํ•˜๊ธฐ ์œ„ํ•ด SFT ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. Plus ๋ฒ„์ „์—์„œ๋Š” ๋ฐ์ดํ„ฐ์…‹์„ ์•ฝ 400๋งŒ์—์„œ 430๋งŒ์œผ๋กœ ํ™•์žฅํ•˜๋ฉฐ, STEM(๊ณผํ•™, ๊ธฐ์ˆ , ๊ณตํ•™, ์ˆ˜ํ•™) ๋ฐ์ดํ„ฐ์™€ ๋ฌผ๋ฆฌํ•™, ํ™”ํ•™, ์ƒ๋ฌผํ•™, ์˜ํ•™, ์ง€๊ตฌ๊ณผํ•™ ๋“ฑ ์—ฌ๋Ÿฌ ๊ณผํ•™ ๋ถ„์•ผ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํฌํ•จํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. Alpaca-33B์˜ ๊ฒฝ์šฐ, ์ถ”๊ฐ€์ ์œผ๋กœ OASST1 ๋ฐ์ดํ„ฐ์…‹(Kopf et al., 2023)์„ ์ถ”๊ฐ€ํ–ˆ์œผ๋ฉฐ, ๊ฐ ๋Œ€ํ™”์—์„œ ์ฒซ ๋ฒˆ์งธ ์ฟผ๋ฆฌ-์‘๋‹ต ์Œ๋งŒ ์ถ”์ถœํ•˜์—ฌ gpt-3.5-turbo API๋ฅผ ์‚ฌ์šฉํ•ด ๋ฒˆ์—ญํ•˜์—ฌ ์•ฝ 2๋งŒ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค(์›๋ณธ ๋ฐ ๋ฒˆ์—ญ๋œ ๋ฐ์ดํ„ฐ ํฌํ•จ). ์ตœ๋Œ€ ์‹œํ€€์Šค ๊ธธ์ด๋Š” 512๋กœ ์„ค์ •ํ•˜๊ณ , ๋ฐฐ์น˜ ์‹œ ์ตœ๋Œ€ ๊ธธ์ด์— ๋งž์ถฐ ์ƒ˜ํ”Œ์„ ๋™์ ์œผ๋กœ ํŒจ๋”ฉํ•ฉ๋‹ˆ๋‹ค.
์›๋ณธ๊ณผ ๋ฒˆ์—ญ๋ณธ)์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํฌ๋กค๋งํ•˜๊ณ  ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ํŠœ๋‹ํ•˜๊ธฐ ์œ„ํ•ด SFT ๋ฐ์ดํ„ฐ๋ฅผ ํฌ๋กค๋งํ–ˆ์Šต๋‹ˆ๋‹ค. Plus ๋ฒ„์ „์˜ ๊ฒฝ์šฐ, STEM(๊ณผํ•™, ๊ธฐ์ˆ , ๊ณตํ•™ ๋ฐ ์ˆ˜ํ•™) ๋ฐ์ดํ„ฐ์™€ ๋ฌผ๋ฆฌํ•™, ํ™”ํ•™, ์ƒ๋ฌผํ•™, ์˜ํ•™, ์ง€๊ตฌ ๊ณผํ•™ ๋“ฑ ์—ฌ๋Ÿฌ ๊ณผํ•™ ๋ถ„์•ผ๋ฅผ ํŠนํžˆ ํฌํ•จํ•˜๋„๋ก ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์•ฝ 4M~4.3M์œผ๋กœ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. Alpaca-33B์˜ ๊ฒฝ์šฐ, ๊ฐ ๋Œ€ํ™”์—์„œ ์ฒซ ๋ฒˆ์งธ ์งˆ์˜-์‘๋‹ต ์Œ๋งŒ ์ถ”์ถœํ•˜๊ณ  gpt-3.5-turbo API๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฒˆ์—ญํ•œ OASST1 ๋ฐ์ดํ„ฐ ์„ธํŠธ(Kopf et al., 2023)๋ฅผ ์ถ”๊ฐ€๋กœ ์ถ”๊ฐ€ํ•˜์—ฌ ์•ฝ 20K์˜ ๋ฐ์ดํ„ฐ(์›๋ณธ ๋ฐ ๋ฒˆ์—ญ๋ณธ)๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ตœ๋Œ€ ์‹œํ€€์Šค ๊ธธ์ด๋ฅผ 512๋กœ ์„ค์ •ํ•˜๊ณ  ์ผ๊ด„ ์ฒ˜๋ฆฌ ์‹œ ์ตœ๋Œ€ ๊ธธ์ด๊นŒ์ง€ ์ƒ˜ํ”Œ์„ ๋™์ ์œผ๋กœ ์ฑ„์›๋‹ˆ๋‹ค.
Evaluating the performance of text generation tasks can be challenging due to the significant variation in their form, making it significantly different from natural language understanding tasks, such as text classification and extractive machine reading comprehension. Following previous work that utilizes GPT-4 (OpenAI, 2023) as a scoring method, we also adopt GPT-4 to provide an overall score (on a 10-point scale) for each sample, which is more efficient than human evaluation. However, GPT-4 may not always provide accurate scores, so we perform manual checks on its ratings and adjust them if necessary. The manual checks ensure that the scores are consistent and reflect the true performance of the models being evaluated.
ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ๊ทธ ํ˜•ํƒœ์˜ ํฐ ๋ณ€๋™์„ฑ ๋•Œ๋ฌธ์— ๋„์ „์ ์ผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ…์ŠคํŠธ ๋ถ„๋ฅ˜ ๋ฐ ์ถ”์ถœ์  ๊ธฐ๊ณ„ ๋…ํ•ด ์ดํ•ด์™€ ๊ฐ™์€ ์ž์—ฐ์–ด ์ดํ•ด ์ž‘์—…๊ณผ๋Š” ์ƒ๋‹นํžˆ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. GPT-4 (OpenAI, 2023)๋ฅผ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ์ด์ „ ์—ฐ๊ตฌ๋ฅผ ๋”ฐ๋ผ, ์šฐ๋ฆฌ๋„ ๊ฐ ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด ์ „๋ฐ˜์ ์ธ ์ ์ˆ˜(10์  ์ฒ™๋„)๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด GPT-4๋ฅผ ์ฑ„ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ธ๊ฐ„ ํ‰๊ฐ€๋ณด๋‹ค ๋” ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ GPT-4๊ฐ€ ํ•ญ์ƒ ์ •ํ™•ํ•œ ์ ์ˆ˜๋ฅผ ์ œ๊ณตํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ, ์šฐ๋ฆฌ๋Š” ๊ทธ ํ‰๊ฐ€๋ฅผ ์ˆ˜๋™์œผ๋กœ ํ™•์ธํ•˜๊ณ  ํ•„์š”์‹œ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜๋™ ํ™•์ธ์€ ์ ์ˆ˜๊ฐ€ ์ผ๊ด€๋˜๋ฉฐ ํ‰๊ฐ€๋˜๋Š” ๋ชจ๋ธ์˜ ์‹ค์ œ ์„ฑ๋Šฅ์„ ๋ฐ˜์˜ํ•˜๋„๋ก ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค.
ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ๊ทธ ํ˜•ํƒœ๊ฐ€ ๋งค์šฐ ๋‹ค์–‘ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํ…์ŠคํŠธ ๋ถ„๋ฅ˜ ๋ฐ ์ถ”์ถœ ๊ธฐ๊ณ„ ๋…ํ•ด๋ ฅ๊ณผ ๊ฐ™์€ ์ž์—ฐ์–ด ์ดํ•ด ์ž‘์—…๊ณผ๋Š” ํฌ๊ฒŒ ๋‹ค๋ฅด๊ธฐ ๋•Œ๋ฌธ์— ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. GPT-4(OpenAI, 2023)๋ฅผ ์ฑ„์  ๋ฐฉ๋ฒ•์œผ๋กœ ํ™œ์šฉํ•˜๋Š” ์„ ํ–‰ ์—ฐ๊ตฌ์— ๋”ฐ๋ผ, ์ €ํฌ๋„ GPT-4๋ฅผ ์ฑ„ํƒํ•˜์—ฌ ๊ฐ ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด 10์  ์ฒ™๋„๋กœ ์ „์ฒด ์ ์ˆ˜๋ฅผ ์ œ๊ณตํ•˜๋Š”๋ฐ, ์ด๋Š” ์‚ฌ๋žŒ์ด ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ GPT-4๊ฐ€ ํ•ญ์ƒ ์ •ํ™•ํ•œ ์ ์ˆ˜๋ฅผ ์ œ๊ณตํ•˜์ง€๋Š” ์•Š์„ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ˆ˜๋™์œผ๋กœ ์ ์ˆ˜๋ฅผ ํ™•์ธํ•˜๊ณ  ํ•„์š”ํ•œ ๊ฒฝ์šฐ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜๋™ ๊ฒ€์‚ฌ๋ฅผ ํ†ตํ•ด ์ ์ˆ˜๊ฐ€ ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•˜๊ณ  ํ‰๊ฐ€ ๋Œ€์ƒ ๋ชจ๋ธ์˜ ์‹ค์ œ ์„ฑ๋Šฅ์„ ๋ฐ˜์˜ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
In our work, we tackle two important challenges faced during synthetic data curation in the works of Gunasekar et al. (2023)--generation cost, and data bias--by rephrasing articles on the web. (i) **WRAP** allows for using an open source, and much smaller LLM (1.8B/7B v/s GPT3.5) to rephrase unstructured and poorly phrased documents in different styles, since it does not rely on the LLM as a knowledge bank. (ii) Thanks to the information maintaining nature of rephrasing, we are able to leverage the natural diversity of the web, rather than relying on an LLM for information which may be prone to factual errors, and/or data biases. Our work shows that the "style" alone can result in large gains in downstream performance.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ์—์„œ๋Š” Gunasekar et al. (2023)์˜ ์ž‘์—…์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ์ค‘ ์ง๋ฉดํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๋„์ „ ๊ณผ์ œ์ธ ์ƒ์„ฑ ๋น„์šฉ๊ณผ ๋ฐ์ดํ„ฐ ํŽธํ–ฅ์„ ์›น์ƒ์˜ ๊ธฐ์‚ฌ๋ฅผ ์žฌ๊ตฌ์„ฑํ•จ์œผ๋กœ์จ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. (i) **WRAP**์€ LLM์„ ์ง€์‹ ์€ํ–‰์œผ๋กœ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ์˜คํ”ˆ ์†Œ์Šค์ด๋ฉด์„œ ํ›จ์”ฌ ์ž‘์€ LLM (1.8B/7B ๋Œ€ GPT3.5)์„ ์‚ฌ์šฉํ•˜์—ฌ ๋น„๊ตฌ์กฐ์ ์ด๊ณ  ์ž˜๋ชป๋œ ๋ฌธ์„œ๋ฅผ ๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ๋กœ ์žฌ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. (ii) ์žฌ๊ตฌ์„ฑ์˜ ์ •๋ณด ์œ ์ง€ ํŠน์„ฑ ๋•๋ถ„์—, ์‚ฌ์‹ค ์˜ค๋ฅ˜ ๋ฐ/๋˜๋Š” ๋ฐ์ดํ„ฐ ํŽธํ–ฅ์— ์ทจ์•ฝํ•  ์ˆ˜ ์žˆ๋Š” LLM์— ์˜์กดํ•˜์ง€ ์•Š๊ณ  ์›น์˜ ์ž์—ฐ์Šค๋Ÿฌ์šด ๋‹ค์–‘์„ฑ์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” "์Šคํƒ€์ผ"๋งŒ์œผ๋กœ๋„ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์—์„œ ํฐ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์˜ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์ €ํฌ ์—ฐ๊ตฌ์—์„œ๋Š” ์›น ๊ธฐ์‚ฌ๋ฅผ ๋‹ค์‹œ ์ž‘์„ฑํ•˜์—ฌ ๊ตฌ๋‚˜์„ธ์นด๋ฅด ๋“ฑ(2023)์˜ ์—ฐ๊ตฌ์—์„œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ์ค‘์— ์ง๋ฉดํ•œ ๋‘ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ณผ์ œ, ์ฆ‰ ์ƒ์„ฑ ๋น„์šฉ๊ณผ ๋ฐ์ดํ„ฐ ํŽธํ–ฅ์„ฑ์„ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. (i) **WRAP**์€ LLM์„ ์ง€์‹ ์€ํ–‰์œผ๋กœ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ๊ตฌ์กฐํ™”๋˜์ง€ ์•Š๊ณ  ์ž˜๋ชป ํ‘œํ˜„๋œ ๋ฌธ์„œ๋ฅผ ๋‹ค์–‘ํ•œ ์Šคํƒ€์ผ๋กœ ๋ฐ”๊พธ๋Š” ๋ฐ ์˜คํ”ˆ ์†Œ์Šค์ด๋ฉฐ ํ›จ์”ฌ ์ž‘์€ LLM(1.8B/7B ๋Œ€ GPT3.5)์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (ii) ๋‹ค์‹œ ์ž‘์„ฑํ•˜๋Š” ์ •๋ณด ์œ ์ง€ ํŠน์„ฑ ๋•๋ถ„์— ์‚ฌ์‹ค ์˜ค๋ฅ˜ ๋ฐ/๋˜๋Š” ๋ฐ์ดํ„ฐ ํŽธํ–ฅ์„ฑ์ด ์žˆ์„ ์ˆ˜ ์žˆ๋Š” LLM์— ์˜์กดํ•˜์ง€ ์•Š๊ณ  ์›น์˜ ์ž์—ฐ์Šค๋Ÿฌ์šด ๋‹ค์–‘์„ฑ์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ ์—ฐ๊ตฌ๋Š” '์Šคํƒ€์ผ'๋งŒ์œผ๋กœ๋„ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
In this work, we show that the premise order significantly affects LLMs' performance on reasoning tasks, even when the premise order does not change the underlying task itself. Our comprehensive evaluation demonstrates that LLM tendencies resemble human preference w.r.t. premise order, i.e., LLMs achieve the best performance when the premise order follows the intermediate reasoning steps to solve the problem. Conversely, LLMs face difficulties when the reasoning problem requires the model to read the problem description back-and-forth, resulting in a performance drop of over 30%. We further extend the study to mathematical reasoning and present the R-GSM benchmark, and again experimentally confirm the ordering effect.
์ด ์—ฐ๊ตฌ์—์„œ ์šฐ๋ฆฌ๋Š” ์ „์ œ ์ˆœ์„œ๊ฐ€ ๊ธฐ๋ณธ ๊ณผ์ œ ์ž์ฒด๋ฅผ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š๋”๋ผ๋„ LLM์˜ ์ถ”๋ก  ์ž‘์—… ์„ฑ๋Šฅ์— ํฌ๊ฒŒ ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ข…ํ•ฉ์ ์ธ ํ‰๊ฐ€ ๊ฒฐ๊ณผ, LLM์˜ ๊ฒฝํ–ฅ์ด ์ „์ œ ์ˆœ์„œ์— ๋Œ€ํ•œ ์ธ๊ฐ„์˜ ์„ ํ˜ธ์™€ ์œ ์‚ฌํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, LLM์€ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๋”ฐ๋ฅด๋Š” ์ „์ œ ์ˆœ์„œ์—์„œ ์ตœ๊ณ ์˜ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ, ์ถ”๋ก  ๋ฌธ์ œ๊ฐ€ ๋ชจ๋ธ์ด ๋ฌธ์ œ ์„ค๋ช…์„ ์•ž๋’ค๋กœ ์ฝ์–ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ, ์„ฑ๋Šฅ์ด 30% ์ด์ƒ ๊ฐ์†Œํ•˜๋Š” ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ•™์  ์ถ”๋ก ์œผ๋กœ ํ™•์žฅํ•˜์—ฌ R-GSM ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ œ์‹œํ•˜๊ณ , ๋‹ค์‹œ ํ•œ ๋ฒˆ ์ˆœ์„œ ํšจ๊ณผ๋ฅผ ์‹คํ—˜์ ์œผ๋กœ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค.
์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ „์ œ ์ˆœ์„œ๊ฐ€ ๊ธฐ๋ณธ์ ์ธ ๊ณผ์ œ ์ž์ฒด๋ฅผ ๋ฐ”๊พธ์ง€ ์•Š๋”๋ผ๋„ ์ „์ œ ์ˆœ์„œ๊ฐ€ ์ถ”๋ก  ๊ณผ์ œ์— ๋Œ€ํ•œ LLM์˜ ์„ฑ๋Šฅ์— ์ƒ๋‹นํ•œ ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํฌ๊ด„์ ์ธ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด LLM์˜ ์„ฑํ–ฅ์ด ์ „์ œ ์ˆœ์„œ์™€ ๊ด€๋ จํ•˜์—ฌ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์™€ ์œ ์‚ฌํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ฆ‰, ์ „์ œ ์ˆœ์„œ๊ฐ€ ๋ฌธ์ œ ํ•ด๊ฒฐ์„ ์œ„ํ•œ ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๋”ฐ๋ฅผ ๋•Œ LLM์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ ์ถ”๋ก  ๋ฌธ์ œ๊ฐ€ ๋ชจ๋ธ์ด ๋ฌธ์ œ ์„ค๋ช…์„ ์•ž๋’ค๋กœ ์ฝ์–ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ ์„ฑ๋Šฅ์ด 30% ์ด์ƒ ์ €ํ•˜๋˜๋Š” ๋“ฑ ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ˆ˜ํ•™์  ์ถ”๋ก ์œผ๋กœ ์—ฐ๊ตฌ๋ฅผ ํ™•์žฅํ•˜์—ฌ R-GSM ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ œ์‹œํ•˜๊ณ , ๋‹ค์‹œ ํ•œ ๋ฒˆ ์‹คํ—˜์ ์œผ๋กœ ์ˆœ์„œ ํšจ๊ณผ๋ฅผ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค.
Tests are designed to be simpleTest items have been designed by a team of AI safety experts to be clear cut, easy to interpret, and easy to assess. They are short and do not use hazard-specific language, are unambiguous and independent of current events, and only test for single-turn interactions. They are also free of adversarial prefixes or prompting tricks that a user may use to elicit harmful behavior because the personas that we tested for are all "unsophisticated". However, this limits their relevance for testing more sophisticated users. We will address this in the future by working more closely with domain experts, and taking inspiration from unstructured datasets of real-world LM interactions [see 69, 70].
ํ…Œ์ŠคํŠธ๋Š” ๊ฐ„๋‹จํ•˜๊ฒŒ ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์€ AI ์•ˆ์ „ ์ „๋ฌธ๊ฐ€ ํŒ€์— ์˜ํ•ด ๋ช…ํ™•ํ•˜๊ณ  ํ•ด์„ํ•˜๊ธฐ ์‰ฌ์šฐ๋ฉฐ ํ‰๊ฐ€ํ•˜๊ธฐ ์‰ฝ๊ฒŒ ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ์งง๊ณ , ํŠน์ • ์œ„ํ—˜ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์œผ๋ฉฐ, ๋ชจํ˜ธํ•˜์ง€ ์•Š๊ณ  ํ˜„์žฌ ์‚ฌ๊ฑด๊ณผ ๋…๋ฆฝ์ ์ด๋ฉฐ, ๋‹จ์ผ ํ„ด ์ƒํ˜ธ์ž‘์šฉ๋งŒ์„ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์šฐ๋ฆฌ๊ฐ€ ํ…Œ์ŠคํŠธํ•œ ํŽ˜๋ฅด์†Œ๋‚˜๊ฐ€ ๋ชจ๋‘ "๋น„์ •๊ต์ "์ด๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์šฉ์ž๊ฐ€ ์œ ํ•ดํ•œ ํ–‰๋™์„ ์œ ๋„ํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ ๋Œ€์  ์ ‘๋‘์‚ฌ๋‚˜ ํ”„๋กฌํ”„ํŠธ ํŠธ๋ฆญ์ด ์—†์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Š” ๋” ์ •๊ตํ•œ ์‚ฌ์šฉ์ž๋ฅผ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ ์žˆ์–ด ๊ด€๋ จ์„ฑ์„ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋„๋ฉ”์ธ ์ „๋ฌธ๊ฐ€์™€ ๋” ๊ธด๋ฐ€ํžˆ ํ˜‘๋ ฅํ•˜๊ณ , ์‹ค์ œ LM ์ƒํ˜ธ์ž‘์šฉ์˜ ๋น„๊ตฌ์กฐ์  ๋ฐ์ดํ„ฐ์…‹์—์„œ ์˜๊ฐ์„ ์–ป์–ด [69, 70 ์ฐธ์กฐ] ๋ฏธ๋ž˜์— ์ด๋ฅผ ํ•ด๊ฒฐํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
ํ…Œ์ŠคํŠธ๋Š” ๊ฐ„๋‹จํ•˜๊ฒŒ ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค.AI ์•ˆ์ „ ์ „๋ฌธ๊ฐ€ ํŒ€์ด ๋ช…ํ™•ํ•˜๊ณ  ํ•ด์„ํ•˜๊ธฐ ์‰ฌ์šฐ๋ฉฐ ํ‰๊ฐ€ํ•˜๊ธฐ ์‰ฌ์šด ํ…Œ์ŠคํŠธ ํ•ญ๋ชฉ์„ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ๋Š” ์งง๊ณ  ์œ„ํ—˜์— ํŠนํ™”๋œ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์œผ๋ฉฐ, ๋ชจํ˜ธํ•˜์ง€ ์•Š๊ณ  ์ตœ์‹  ์ด๋ฒคํŠธ์™€ ๋ฌด๊ด€ํ•˜๋ฉฐ, ๋‹จ ํ•œ ๋ฒˆ์˜ ์ƒํ˜ธ ์ž‘์šฉ๋งŒ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ…Œ์ŠคํŠธ ๋Œ€์ƒ ์ธ๋ฌผ์ด ๋ชจ๋‘ "๋น„์ •๊ตํ•œ" ์ธ๋ฌผ์ด๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์šฉ์ž๊ฐ€ ์œ ํ•ดํ•œ ํ–‰๋™์„ ์œ ๋„ํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ ๋Œ€์ ์ธ ์ ‘๋‘์‚ฌ๋‚˜ ํ”„๋กฌํ”„ํŠธ ํŠธ๋ฆญ์ด ์—†์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Š” ๋ณด๋‹ค ์ •๊ตํ•œ ์‚ฌ์šฉ์ž๋ฅผ ํ…Œ์ŠคํŠธํ•˜๋Š” ๋ฐ ์žˆ์–ด ๊ด€๋ จ์„ฑ์ด ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ํ–ฅํ›„์—๋Š” ๋„๋ฉ”์ธ ์ „๋ฌธ๊ฐ€์™€ ๋” ๊ธด๋ฐ€ํžˆ ํ˜‘๋ ฅํ•˜๊ณ  ์‹ค์ œ LM ์ƒํ˜ธ ์ž‘์šฉ์˜ ๋น„์ •ํ˜• ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์˜๊ฐ์„ ์–ป์–ด ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค[69, 70]์„ ์ฐธ์กฐํ•˜์„ธ์š”.
**Voting**. Let \(A\) represent the final answer. In this phase, we employ majority voting to consolidate the response sample set \(S\) into the final answer \(A\). This involves calculating the cumulative similarity for each sample relative to the others, denoted as \(V(s_{i})=\sum_{j=1,j\neq i}^{N}sim(s_{i},s_{j})\). For open-ended generation tasks such as code generation, the BLEU score (Papineni et al., 2002) is utilized to quantify similarity. Conversely, for close-ended tasks like multiple-choice questions, similarity is measured by occurrence frequency. The sample that exhibits the highest cumulative similarity is then chosen as the final answer denoted as \(A=\operatorname*{arg\,max}_{s_{i}\in S}V(s_{i})\).
**Voting**. \(A\)๋ฅผ ์ตœ์ข… ๋‹ต๋ณ€์œผ๋กœ ๋‚˜ํƒ€๋‚ด์ž. ์ด ๋‹จ๊ณ„์—์„œ๋Š” ๋‹ค์ˆ˜๊ฒฐ ํˆฌํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‘๋‹ต ์ƒ˜ํ”Œ ์ง‘ํ•ฉ \(S\)๋ฅผ ์ตœ์ข… ๋‹ต๋ณ€ \(A\)๋กœ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๊ฐ ์ƒ˜ํ”Œ์˜ ๋ˆ„์  ์œ ์‚ฌ๋„๋ฅผ ๋‹ค๋ฅธ ์ƒ˜ํ”Œ๋“ค๊ณผ ๋น„๊ตํ•˜์—ฌ ๊ณ„์‚ฐํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•˜๋ฉฐ, ์ด๋Š” \(V(s_{i})=\sum_{j=1,j\neq i}^{N}sim(s_{i},s_{j})\)๋กœ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ฝ”๋“œ ์ƒ์„ฑ๊ณผ ๊ฐ™์€ ๊ฐœ๋ฐฉํ˜• ์ƒ์„ฑ ์ž‘์—…์˜ ๊ฒฝ์šฐ, ์œ ์‚ฌ๋„๋ฅผ ์ •๋Ÿ‰ํ™”ํ•˜๊ธฐ ์œ„ํ•ด BLEU ์ ์ˆ˜(Papineni et al., 2002)๊ฐ€ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ๋‹ค์ง€์„ ๋‹คํ˜• ์งˆ๋ฌธ๊ณผ ๊ฐ™์€ ํ์‡„ํ˜• ์ž‘์—…์˜ ๊ฒฝ์šฐ, ์œ ์‚ฌ๋„๋Š” ๋ฐœ์ƒ ๋นˆ๋„๋กœ ์ธก์ •๋ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ๋†’์€ ๋ˆ„์  ์œ ์‚ฌ๋„๋ฅผ ๋ณด์ด๋Š” ์ƒ˜ํ”Œ์ด ์ตœ์ข… ๋‹ต๋ณ€์œผ๋กœ ์„ ํƒ๋˜๋ฉฐ, ์ด๋Š” \(A=\operatorname*{arg\,max}_{s_{i}\in S}V(s_{i})\)๋กœ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
**ํˆฌํ‘œ**. A\)๋ฅผ ์ตœ์ข… ๋‹ต์œผ๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—์„œ๋Š” ๋‹ค์ˆ˜๊ฒฐ ํˆฌํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‘๋‹ต ์ƒ˜ํ”Œ ์ง‘ํ•ฉ \(S\)๋ฅผ ์ตœ์ข… ๋‹ต์•ˆ \(A\)๋กœ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๋‹ค๋ฅธ ์ƒ˜ํ”Œ๊ณผ ๋น„๊ตํ•˜์—ฌ ๊ฐ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ๋ˆ„์  ์œ ์‚ฌ๋„๋ฅผ ๊ณ„์‚ฐํ•˜๋Š” ๊ฒƒ์ด ํฌํ•จ๋˜๋ฉฐ, ์ด๋ฅผ \(V(s_{i})=\sum_{j=1,j\neq i}^{N}sim(s_{i},s_{j})\)๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ฝ”๋“œ ์ƒ์„ฑ๊ณผ ๊ฐ™์€ ๊ฐœ๋ฐฉํ˜• ์ƒ์„ฑ ์ž‘์—…์˜ ๊ฒฝ์šฐ BLEU ์ ์ˆ˜(Papineni et al., 2002)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์œ ์‚ฌ๋„๋ฅผ ์ •๋Ÿ‰ํ™”ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ ๊ฐ๊ด€์‹ ๋ฌธ์ œ์™€ ๊ฐ™์€ ํ์‡„ํ˜• ๊ณผ์ œ์˜ ๊ฒฝ์šฐ ์œ ์‚ฌ์„ฑ์€ ๋ฐœ์ƒ ๋นˆ๋„๋กœ ์ธก์ •๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ˆ„์  ์œ ์‚ฌ๋„๊ฐ€ ๊ฐ€์žฅ ๋†’์€ ์ƒ˜ํ”Œ์ด ์ตœ์ข… ๋‹ต์œผ๋กœ ์„ ํƒ๋˜๋ฉฐ, ์ด๋ฅผ \(A=\operatorname*{arg\,max}_{s_{i}\in S}V(s_{i})\)๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.
**Scaling with Model Size and Data Quantity.** To explore how our method scales with different synthetic data quantities and model sizes, we conduct experiments by comparing CodecLM with WizardLM+, the most competitive baseline. The experiment results on Evol-Instruct with LLaMA-7B and -13B as the target LLM are presented in Figure 5. Both methods get increasingly better performance with more synthetic data and larger target models. CodecLM consistently outperforms WizardLM+ under all cases, demonstrating its great data efficiency and scalability. We expect the gain will gradually diminish after we generate more than 8k synthetic data, due to the intrinsic ability gap between the target models and the strong LLM.
**๋ชจ๋ธ ํฌ๊ธฐ์™€ ๋ฐ์ดํ„ฐ ์–‘์— ๋”ฐ๋ฅธ ํ™•์žฅ์„ฑ.** ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์ด ๋‹ค์–‘ํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์–‘๊ณผ ๋ชจ๋ธ ํฌ๊ธฐ์— ๋”ฐ๋ผ ์–ด๋–ป๊ฒŒ ํ™•์žฅ๋˜๋Š”์ง€ ํƒ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด, ๊ฐ€์žฅ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ๊ธฐ์ค€์„ ์ธ WizardLM+์™€ CodecLM์„ ๋น„๊ตํ•˜๋Š” ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. LLaMA-7B์™€ -13B๋ฅผ ๋Œ€์ƒ LLM์œผ๋กœ ํ•˜๋Š” Evol-Instruct์— ๋Œ€ํ•œ ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” Figure 5์— ์ œ์‹œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฐฉ๋ฒ• ๋ชจ๋‘ ๋” ๋งŽ์€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ๋” ํฐ ๋Œ€์ƒ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ• ์ˆ˜๋ก ์„ฑ๋Šฅ์ด ์ ์  ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค. CodecLM์€ ๋ชจ๋“  ๊ฒฝ์šฐ์—์„œ WizardLM+๋ฅผ ์ผ๊ด€๋˜๊ฒŒ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ, ๋›ฐ์–ด๋‚œ ๋ฐ์ดํ„ฐ ํšจ์œจ์„ฑ๊ณผ ํ™•์žฅ์„ฑ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ๋Œ€์ƒ ๋ชจ๋ธ๊ณผ ๊ฐ•๋ ฅํ•œ LLM ๊ฐ„์˜ ๋ณธ์งˆ์ ์ธ ๋Šฅ๋ ฅ ์ฐจ์ด๋กœ ์ธํ•ด 8k ์ด์ƒ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•œ ํ›„์—๋Š” ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ์ ์ฐจ ์ค„์–ด๋“ค ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.
**๋ชจ๋ธ ํฌ๊ธฐ ๋ฐ ๋ฐ์ดํ„ฐ ์–‘์— ๋”ฐ๋ฅธ ํ™•์žฅ์„ฑ.** ๋‹ค์–‘ํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์–‘๊ณผ ๋ชจ๋ธ ํฌ๊ธฐ์— ๋”ฐ๋ผ ์ €ํฌ ๋ฐฉ๋ฒ•์ด ์–ด๋–ป๊ฒŒ ํ™•์žฅ๋˜๋Š”์ง€ ์‚ดํŽด๋ณด๊ธฐ ์œ„ํ•ด ๊ฐ€์žฅ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ๊ธฐ์ค€์„ ์ธ WizardLM+์™€ CodecLM์„ ๋น„๊ตํ•˜์—ฌ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. LLaMA-7B ๋ฐ -13B๋ฅผ ๋Œ€์ƒ LLM์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ Evol-Instruct์— ๋Œ€ํ•œ ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 5์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฐฉ๋ฒ• ๋ชจ๋‘ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ๋งŽ๊ณ  ๋Œ€์ƒ ๋ชจ๋ธ์ด ํด์ˆ˜๋ก ์ ์  ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค. ๋ชจ๋“  ๊ฒฝ์šฐ์—์„œ CodecLM์ด WizardLM+๋ณด๋‹ค ์ผ๊ด€๋˜๊ฒŒ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์–ด ๋›ฐ์–ด๋‚œ ๋ฐ์ดํ„ฐ ํšจ์œจ์„ฑ๊ณผ ํ™•์žฅ์„ฑ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€์ƒ ๋ชจ๋ธ๊ณผ ๊ฐ•๋ ฅํ•œ LLM ๊ฐ„์˜ ๋‚ด์žฌ์  ๋Šฅ๋ ฅ ์ฐจ์ด๋กœ ์ธํ•ด 8,000๊ฐœ ์ด์ƒ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋ฉด ์ ์ฐจ ์ด๋“์ด ๊ฐ์†Œํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.
Model providers(e.g., builders, engineers and researchers). This category primarily covers developers training and releasing AI models, such as engineers at AI labs that build language models. Providers may create and release a new model from scratch, such as when Meta released the LLaMA family of models [23, 24]. Providers may also create a model based on an existing model, such as when the Alpaca team adapted LLaMA-7B to make Alpaca-7B [25]. Our community outreach and research indicates that model providers' objectives include (i) building safer models; (ii) ensuring that models remain useful; (iii) communicating how their models should be used responsibly; and (iv) ensuring compliance with legal standards.
๋ชจ๋ธ ์ œ๊ณต์ž(์˜ˆ: ๋นŒ๋”, ์—”์ง€๋‹ˆ์–ด ๋ฐ ์—ฐ๊ตฌ์›). ์ด ์นดํ…Œ๊ณ ๋ฆฌ๋Š” ์ฃผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๋Š” AI ์—ฐ๊ตฌ์†Œ์˜ ์—”์ง€๋‹ˆ์–ด์™€ ๊ฐ™์€ AI ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ๋ฐฐํฌํ•˜๋Š” ๊ฐœ๋ฐœ์ž๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์ œ๊ณต์ž๋Š” Meta๊ฐ€ LLaMA ๋ชจ๋ธ ํŒจ๋ฐ€๋ฆฌ๋ฅผ ์ถœ์‹œํ–ˆ์„ ๋•Œ์ฒ˜๋Ÿผ ์ฒ˜์Œ๋ถ€ํ„ฐ ์ƒˆ๋กœ์šด ๋ชจ๋ธ์„ ๋งŒ๋“ค๊ณ  ๋ฐฐํฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [23, 24]. ์ œ๊ณต์ž๋Š” ๋˜ํ•œ Alpaca ํŒ€์ด LLaMA-7B๋ฅผ ์ ์‘์‹œ์ผœ Alpaca-7B๋ฅผ ๋งŒ๋“  ๊ฒƒ์ฒ˜๋Ÿผ ๊ธฐ์กด ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ชจ๋ธ์„ ๋งŒ๋“ค ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค [25]. ์šฐ๋ฆฌ์˜ ์ปค๋ฎค๋‹ˆํ‹ฐ ์•„์›ƒ๋ฆฌ์น˜ ๋ฐ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๋ชจ๋ธ ์ œ๊ณต์ž์˜ ๋ชฉํ‘œ๋Š” (i) ๋” ์•ˆ์ „ํ•œ ๋ชจ๋ธ ๊ตฌ์ถ•; (ii) ๋ชจ๋ธ์ด ์œ ์šฉํ•˜๊ฒŒ ์œ ์ง€๋˜๋„๋ก ๋ณด์žฅ; (iii) ๋ชจ๋ธ์ด ์ฑ…์ž„๊ฐ ์žˆ๊ฒŒ ์‚ฌ์šฉ๋˜์–ด์•ผ ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ „๋‹ฌ; ๋ฐ (iv) ๋ฒ•์  ๊ธฐ์ค€ ์ค€์ˆ˜๋ฅผ ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
๋ชจ๋ธ ๊ณต๊ธ‰์ž(์˜ˆ: ๋นŒ๋”, ์—”์ง€๋‹ˆ์–ด ๋ฐ ์—ฐ๊ตฌ์›). ์ด ์นดํ…Œ๊ณ ๋ฆฌ๋Š” ์ฃผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๋Š” AI ๋žฉ์˜ ์—”์ง€๋‹ˆ์–ด์™€ ๊ฐ™์ด AI ๋ชจ๋ธ์„ ๊ต์œกํ•˜๊ณ  ์ถœ์‹œํ•˜๋Š” ๊ฐœ๋ฐœ์ž๋ฅผ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์ œ๊ณต์ž๋Š” Meta๊ฐ€ LLaMA ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์„ ์ถœ์‹œํ•  ๋•Œ์™€ ๊ฐ™์ด ์ฒ˜์Œ๋ถ€ํ„ฐ ์ƒˆ ๋ชจ๋ธ์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[23, 24]. ๋˜ํ•œ ์ œ๊ณต์ž๋Š” ๊ธฐ์กด ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ชจ๋ธ์„ ๋งŒ๋“ค ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์•ŒํŒŒ์นด ํŒ€์ด LLaMA-7B๋ฅผ Alpaca-7B๋กœ ๊ฐœ์กฐํ•œ ๊ฒฝ์šฐ[25]์™€ ๊ฐ™์ด ๊ธฐ์กด ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ชจ๋ธ์„ ๋งŒ๋“ค ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ปค๋ฎค๋‹ˆํ‹ฐ ํ™๋ณด ๋ฐ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๋ชจ๋ธ ์ œ๊ณต์—…์ฒด์˜ ๋ชฉํ‘œ๋Š” (i) ๋” ์•ˆ์ „ํ•œ ๋ชจ๋ธ ๊ตฌ์ถ•, (ii) ๋ชจ๋ธ์˜ ์œ ์šฉ์„ฑ ๋ณด์žฅ, (iii) ์ฑ…์ž„๊ฐ ์žˆ๋Š” ๋ชจ๋ธ ์‚ฌ์šฉ ๋ฐฉ๋ฒ•์„ ์ „๋‹ฌํ•˜๊ณ , (iv) ๋ฒ•์  ํ‘œ์ค€ ์ค€์ˆ˜ ๋ณด์žฅ ๋“ฑ์ž…๋‹ˆ๋‹ค.
Slot filling is a technique employed in natural language processing for the purpose of recognizing and extracting specific information from user-supplied text or speech. In slot filling, the system defines a set of slots in advance, with each slot representing a specific information requirement. These requirements may include, but are not limited to, date, time, location, and so forth. Upon receipt of a user input in the form of text or speech, the system performs an analysis of the content, attempting to identify information that matches the predefined slots or classification labels Lu et al. (2023, 2023). This information is then populated into the corresponding slots for subsequent processing and response.
์Šฌ๋กฏ ์ฑ„์šฐ๊ธฐ๋Š” ์‚ฌ์šฉ์ž๊ฐ€ ์ œ๊ณตํ•œ ํ…์ŠคํŠธ๋‚˜ ์Œ์„ฑ์—์„œ ํŠน์ • ์ •๋ณด๋ฅผ ์ธ์‹ํ•˜๊ณ  ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ์—์„œ ์‚ฌ์šฉ๋˜๋Š” ๊ธฐ์ˆ ์ž…๋‹ˆ๋‹ค. ์Šฌ๋กฏ ์ฑ„์šฐ๊ธฐ์—์„œ๋Š” ์‹œ์Šคํ…œ์ด ์‚ฌ์ „์— ์ผ๋ จ์˜ ์Šฌ๋กฏ์„ ์ •์˜ํ•˜๋ฉฐ, ๊ฐ ์Šฌ๋กฏ์€ ํŠน์ • ์ •๋ณด ์š”๊ตฌ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์š”๊ตฌ ์‚ฌํ•ญ์—๋Š” ๋‚ ์งœ, ์‹œ๊ฐ„, ์œ„์น˜ ๋“ฑ์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์ง€๋งŒ ์ด์— ๊ตญํ•œ๋˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. ํ…์ŠคํŠธ๋‚˜ ์Œ์„ฑ ํ˜•ํƒœ์˜ ์‚ฌ์šฉ์ž ์ž…๋ ฅ์„ ๋ฐ›์œผ๋ฉด, ์‹œ์Šคํ…œ์€ ๋‚ด์šฉ์„ ๋ถ„์„ํ•˜์—ฌ ์‚ฌ์ „์— ์ •์˜๋œ ์Šฌ๋กฏ์ด๋‚˜ ๋ถ„๋ฅ˜ ๋ผ๋ฒจ๊ณผ ์ผ์น˜ํ•˜๋Š” ์ •๋ณด๋ฅผ ์‹๋ณ„ํ•˜๋ ค๊ณ  ์‹œ๋„ํ•ฉ๋‹ˆ๋‹ค(Lu et al., 2023, 2023). ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ์ •๋ณด๋Š” ํ›„์† ์ฒ˜๋ฆฌ ๋ฐ ์‘๋‹ต์„ ์œ„ํ•ด ํ•ด๋‹น ์Šฌ๋กฏ์— ์ฑ„์›Œ์ง‘๋‹ˆ๋‹ค.
์Šฌ๋กฏ ์ฑ„์šฐ๊ธฐ๋Š” ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ์—์„œ ์‚ฌ์šฉ์ž ์ œ๊ณต ํ…์ŠคํŠธ๋‚˜ ์Œ์„ฑ์—์„œ ํŠน์ • ์ •๋ณด๋ฅผ ์ธ์‹ํ•˜๊ณ  ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉํ•˜๋Š” ๊ธฐ์ˆ ์ž…๋‹ˆ๋‹ค. ์Šฌ๋กฏ ์ฑ„์šฐ๊ธฐ์—์„œ๋Š” ์‹œ์Šคํ…œ์—์„œ ๋ฏธ๋ฆฌ ์Šฌ๋กฏ ์ง‘ํ•ฉ์„ ์ •์˜ํ•˜๊ณ  ๊ฐ ์Šฌ๋กฏ์€ ํŠน์ • ์ •๋ณด ์š”๊ตฌ ์‚ฌํ•ญ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์š”๊ตฌ ์‚ฌํ•ญ์—๋Š” ๋‚ ์งœ, ์‹œ๊ฐ„, ์œ„์น˜ ๋“ฑ์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์ง€๋งŒ ์ด์— ๊ตญํ•œ๋˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. ํ…์ŠคํŠธ ๋˜๋Š” ์Œ์„ฑ ํ˜•์‹์˜ ์‚ฌ์šฉ์ž ์ž…๋ ฅ์ด ์ˆ˜์‹ ๋˜๋ฉด ์‹œ์Šคํ…œ์€ ์ฝ˜ํ…์ธ ๋ฅผ ๋ถ„์„ํ•˜์—ฌ ๋ฏธ๋ฆฌ ์ •์˜๋œ ์Šฌ๋กฏ ๋˜๋Š” ๋ถ„๋ฅ˜ ๋ ˆ์ด๋ธ” Lu ๋“ฑ(2023, 2023)๊ณผ ์ผ์น˜ํ•˜๋Š” ์ •๋ณด๋ฅผ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ์ •๋ณด๋ฅผ ํ•ด๋‹น ์Šฌ๋กฏ์— ์ฑ„์›Œ ํ›„์† ์ฒ˜๋ฆฌ ๋ฐ ์‘๋‹ต์„ ์œ„ํ•ด ์ฑ„์›๋‹ˆ๋‹ค.
Comparing SOLAR 10.7B + SFT + DPO and SOLAR 10.7B + SFT + sDPO, we can see that using sDPO over DPO results in a higher H4 score overall, with noticeable improvements in ARC and TruthfulQA scores. Therefore, we believe sDPO could function as a drop-in replacement for DPO training with better performance. Looking at SOLAR 10.7B + SFT + sDPO and SOLAR 10.7B + SFT + sDPO Strat., we see that the specific way of splitting the available DPO data into multiple \(D_{t}\) can also impact performance. We find that the natural split of using different preference datasets as \(D_{t}\) works best in our experiments. We believe further exploration of how to define \(D_{t}\) is an interesting direction for future research.
SOLAR 10.7B + SFT + DPO์™€ SOLAR 10.7B + SFT + sDPO๋ฅผ ๋น„๊ตํ•ด๋ณด๋ฉด, DPO ๋Œ€์‹  sDPO๋ฅผ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ ์ „๋ฐ˜์ ์œผ๋กœ ๋” ๋†’์€ H4 ์ ์ˆ˜๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ARC์™€ TruthfulQA ์ ์ˆ˜์—์„œ๋„ ๋ˆˆ์— ๋„๋Š” ํ–ฅ์ƒ์ด ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ sDPO๊ฐ€ DPO ํ›ˆ๋ จ์˜ ๋Œ€์ฒด์ œ๋กœ์„œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. SOLAR 10.7B + SFT + sDPO์™€ SOLAR 10.7B + SFT + sDPO Strat.์„ ์‚ดํŽด๋ณด๋ฉด, ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ DPO ๋ฐ์ดํ„ฐ๋ฅผ ์—ฌ๋Ÿฌ \(D_{t}\)๋กœ ๋‚˜๋ˆ„๋Š” ํŠน์ • ๋ฐฉ์‹๋„ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์‹คํ—˜์—์„œ๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ์„ ํ˜ธ ๋ฐ์ดํ„ฐ์…‹์„ \(D_{t}\)๋กœ ์‚ฌ์šฉํ•˜๋Š” ์ž์—ฐ์Šค๋Ÿฌ์šด ๋ถ„ํ• ์ด ๊ฐ€์žฅ ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. \(D_{t}\)๋ฅผ ์ •์˜ํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ํƒ๊ตฌ๊ฐ€ ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ ํฅ๋ฏธ๋กœ์šด ๋ฐฉํ–ฅ์ด ๋  ๊ฒƒ์ด๋ผ๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค.
SOLAR 10.7B + SFT + DPO์™€ SOLAR 10.7B + SFT + sDPO๋ฅผ ๋น„๊ตํ•˜๋ฉด DPO๋ณด๋‹ค sDPO๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ „์ฒด์ ์œผ๋กœ ๋” ๋†’์€ H4 ์ ์ˆ˜๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํŠนํžˆ ARC ๋ฐ TruthfulQA ์ ์ˆ˜๊ฐ€ ๋ˆˆ์— ๋„๊ฒŒ ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ €ํฌ๋Š” sDPO๊ฐ€ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์œผ๋กœ DPO ํ›ˆ๋ จ์„ ๋Œ€์ฒดํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. SOLAR 10.7B + SFT + sDPO์™€ SOLAR 10.7B + SFT + sDPO Strat.๋ฅผ ์‚ดํŽด๋ณด๋ฉด, ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ DPO ๋ฐ์ดํ„ฐ๋ฅผ ์—ฌ๋Ÿฌ ๊ฐœ์˜ \(D_{t}\)๋กœ ๋ถ„ํ• ํ•˜๋Š” ํŠน์ • ๋ฐฉ๋ฒ•๋„ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์„œ๋กœ ๋‹ค๋ฅธ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ \(D_{t}\)๋กœ ์‚ฌ์šฉํ•˜๋Š” ์ž์—ฐ์Šค๋Ÿฌ์šด ๋ถ„ํ• ์ด ๊ฐ€์žฅ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. D_{t}\)๋ฅผ ์ •์˜ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ถ”๊ฐ€๋กœ ํƒ์ƒ‰ํ•˜๋Š” ๊ฒƒ์€ ํ–ฅํ›„ ์—ฐ๊ตฌ์˜ ํฅ๋ฏธ๋กœ์šด ๋ฐฉํ–ฅ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
We measure the premise order effect by varying the following two factors:* **Number of rules required in the proof.** It is expected that the premise order effect is more significant with more rules. For our benchmark, we generate problems whose numbers of rules range from 4 to 12. * **Number of distracting rules** (i.e., rules that are not useful for the proof) presented in the problem. The presence of distracting rules also complicates the problem, as premise selection itself is challenging (Ferreira and Freitas, 2020; Irving et al., 2016; Wang et al., 2017), and LLMs are shown to be easily distracted by irrelevant context (Shi et al., 2023). We include problem variants with 0, 5 and 10 distracting rules.
์ „์ œ ์ˆœ์„œ ํšจ๊ณผ๋ฅผ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ ๋‘ ๊ฐ€์ง€ ์š”์†Œ๋ฅผ ๋ณ€ํ™”์‹œํ‚ต๋‹ˆ๋‹ค: * **์ฆ๋ช…์— ํ•„์š”ํ•œ ๊ทœ์น™์˜ ์ˆ˜.** ๊ทœ์น™์ด ๋งŽ์„์ˆ˜๋ก ์ „์ œ ์ˆœ์„œ ํšจ๊ณผ๊ฐ€ ๋” ์ค‘์š”ํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฒค์น˜๋งˆํฌ์—์„œ๋Š” ๊ทœ์น™์˜ ์ˆ˜๊ฐ€ 4์—์„œ 12๊นŒ์ง€ ๋ฒ”์œ„์ธ ๋ฌธ์ œ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. * **ํ˜ผ๋ž€์„ ์ฃผ๋Š” ๊ทœ์น™์˜ ์ˆ˜** (์ฆ‰, ์ฆ๋ช…์— ์œ ์šฉํ•˜์ง€ ์•Š์€ ๊ทœ์น™). ํ˜ผ๋ž€์„ ์ฃผ๋Š” ๊ทœ์น™์˜ ์กด์žฌ๋Š” ์ „์ œ ์„ ํƒ ์ž์ฒด๊ฐ€ ๋„์ „์ ์ด๊ธฐ ๋•Œ๋ฌธ์— ๋ฌธ์ œ๋ฅผ ๋”์šฑ ๋ณต์žกํ•˜๊ฒŒ ๋งŒ๋“ญ๋‹ˆ๋‹ค (Ferreira and Freitas, 2020; Irving et al., 2016; Wang et al., 2017). ๋˜ํ•œ, LLM์€ ๊ด€๋ จ ์—†๋Š” ๋งฅ๋ฝ์— ์‰ฝ๊ฒŒ ํ˜ผ๋ž€์Šค๋Ÿฌ์›Œํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค (Shi et al., 2023). ์šฐ๋ฆฌ๋Š” ํ˜ผ๋ž€์„ ์ฃผ๋Š” ๊ทœ์น™์ด 0, 5, 10๊ฐœ์ธ ๋ฌธ์ œ ๋ณ€ํ˜•์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
์ „์ œ ์ˆœ์„œ ํšจ๊ณผ๋Š” ๋‹ค์Œ ๋‘ ๊ฐ€์ง€ ์š”์ธ์„ ๋ณ€๊ฒฝํ•˜์—ฌ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค:* **์ฆ๋ช…์— ํ•„์š”ํ•œ ๊ทœ์น™์˜ ์ˆ˜**. ๊ทœ์น™์ด ๋งŽ์„์ˆ˜๋ก ์ „์ œ ์ˆœ์„œ ํšจ๊ณผ๊ฐ€ ๋” ํฐ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค. ๋ฒค์น˜๋งˆํฌ์˜ ๊ฒฝ์šฐ ๊ทœ์น™ ์ˆ˜๊ฐ€ 4๊ฐœ์—์„œ 12๊ฐœ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ๋ฌธ์ œ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. * ๋ฌธ์ œ์— ์ œ์‹œ๋œ **์ฃผ์˜ ์‚ฐ๋งŒํ•œ ๊ทœ์น™**์˜ ์ˆ˜(์ฆ‰, ์ฆ๋ช…์— ์œ ์šฉํ•˜์ง€ ์•Š์€ ๊ทœ์น™)์ž…๋‹ˆ๋‹ค. ์ „์ œ ์„ ํƒ ์ž์ฒด๊ฐ€ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์— ์ฃผ์˜ ์‚ฐ๋งŒํ•œ ๊ทœ์น™์˜ ์กด์žฌ๋Š” ๋ฌธ์ œ๋ฅผ ๋ณต์žกํ•˜๊ฒŒ ๋งŒ๋“ค๊ณ (Ferreira and Freitas, 2020; Irving et al., 2016; Wang et al., 2017), LLM์€ ๊ด€๋ จ ์—†๋Š” ์ปจํ…์ŠคํŠธ์— ์‰ฝ๊ฒŒ ์ฃผ์˜๋ฅผ ์‚ฐ๋งŒํ•˜๊ฒŒ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Shi et al., 2023). ์ฃผ์˜ ์‚ฐ๋งŒํ•œ ๊ทœ์น™์ด 0, 5, 10๊ฐœ์ธ ๋ฌธ์ œ ๋ณ€ํ˜•์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
to evaluate the various aspects of language understanding and reasoning. Specifically, this benchmark provides a Korean-translated version of language understanding tasks: boolean question answering [1], commonsense causal reasoning (COPA; Roemmele et al.2011, context-sensitive word understanding (WiC; Pilehvar and Camacho-Collados2019), commonsense reasoning [11], and sentiment negation recognition (SentiNeg). For English tasks, we employ the following original tasks of KoBEST, BoolQ, COPA, and HelaSwag, which can better highlight the alignment between the English and Korean capabilities of LLMs. To ensure consistent comparisons, we employ an open-source LLM evaluation framework, lm-evaluation-harness9[7].
์–ธ์–ด ์ดํ•ด์™€ ์ถ”๋ก ์˜ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ์ด ๋ฒค์น˜๋งˆํฌ๋Š” ์–ธ์–ด ์ดํ•ด ๊ณผ์ œ์˜ ํ•œ๊ตญ์–ด ๋ฒˆ์—ญ ๋ฒ„์ „์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค: boolean ์งˆ๋ฌธ ์‘๋‹ต [1], ์ƒ์‹์  ์ธ๊ณผ ์ถ”๋ก  (COPA; Roemmele et al. 2011), ๋ฌธ๋งฅ ๋ฏผ๊ฐ ๋‹จ์–ด ์ดํ•ด (WiC; Pilehvar and Camacho-Collados 2019), ์ƒ์‹์  ์ถ”๋ก  [11], ๊ทธ๋ฆฌ๊ณ  ๊ฐ์ • ๋ถ€์ • ์ธ์‹ (SentiNeg). ์˜์–ด ๊ณผ์ œ์˜ ๊ฒฝ์šฐ, KoBEST, BoolQ, COPA, ๊ทธ๋ฆฌ๊ณ  HelaSwag์˜ ์›๋ž˜ ๊ณผ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์˜ ์˜์–ด์™€ ํ•œ๊ตญ์–ด ๋Šฅ๋ ฅ ๊ฐ„์˜ ์ •๋ ฌ์„ ๋” ์ž˜ ๊ฐ•์กฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๊ด€๋œ ๋น„๊ต๋ฅผ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์˜คํ”ˆ ์†Œ์Šค LLM ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ์ธ lm-evaluation-harness9[7]์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
์–ธ์–ด ์ดํ•ด์™€ ์ถ”๋ก ์˜ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด. ํŠนํžˆ, ์ด ๋ฒค์น˜๋งˆํฌ๋Š” ์–ธ์–ด ์ดํ•ด ๊ณผ์ œ์˜ ํ•œ๊ตญ์–ด ๋ฒˆ์—ญ ๋ฒ„์ „์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค: ๋ถ€์šธ ์งˆ๋ฌธ ๋‹ต๋ณ€[1], ์ƒ์‹์  ์ธ๊ณผ ์ถ”๋ก (COPA; Roemmele et al.2011, ๋ฌธ๋งฅ ๋ฏผ๊ฐํ•œ ๋‹จ์–ด ์ดํ•ด(WiC; Pilehvar ๋ฐ Camacho-Collados2019), ์ƒ์‹์  ์ถ”๋ก [11] ๋ฐ ๊ฐ์ • ๋ถ€์ • ์ธ์‹(SentiNeg)์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์˜์–ด ๊ณผ์ œ์˜ ๊ฒฝ์šฐ, KoBEST, BoolQ, COPA ๋ฐ HelaSwag์˜ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์›๋ž˜ ๊ณผ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM์˜ ์˜์–ด ๋ฐ ํ•œ๊ตญ์–ด ๊ธฐ๋Šฅ ๊ฐ„์˜ ์ •๋ ฌ์„ ๋” ์ž˜ ๊ฐ•์กฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๊ด€๋œ ๋น„๊ต๋ฅผ ์œ„ํ•ด ์˜คํ”ˆ ์†Œ์Šค LLM ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ์ธ lm-evaluation-harness9[7]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
Tail Cutting.As mentioned, deliberate choices in the AI generation algorithm (like top-\(p\) or top-\(k\) next token prediction) immediately lead to a chopped tail at \(k\). When viewed as skills, we can say that only the \(k\)th most frequent outcomes ("skills") are considered. But even when no tails are cut deliberately, the finite size \(T_{0}\) of the training set (sampling bias) induces an effective tail-cutting. This can be seen as follows: Sample an iid dataset of size \(T_{0}\), and estimate the histogram \(p_{\text{AI}}\); this new distribution plays the role of an AI data-generator. An integer \(i\) appears in the support of \(p_{\text{AI}}\) a number of times which is \(T_{0}p_{i}\) on average.
Tail Cutting. ์•ž์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, AI ์ƒ์„ฑ ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ ์˜๋„์ ์ธ ์„ ํƒ(์˜ˆ: top-\(p\) ๋˜๋Š” top-\(k\) ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก)์€ ์ฆ‰์‹œ \(k\)์—์„œ ๊ผฌ๋ฆฌ๋ฅผ ์ž˜๋ผ๋‚ด๋Š” ๊ฒฐ๊ณผ๋ฅผ ์ดˆ๋ž˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ๊ธฐ์ˆ ๋กœ ๋ณด๋ฉด, ์šฐ๋ฆฌ๋Š” \(k\)๋ฒˆ์งธ๋กœ ๊ฐ€์žฅ ๋นˆ๋ฒˆํ•œ ๊ฒฐ๊ณผ("๊ธฐ์ˆ ")๋งŒ ๊ณ ๋ ค๋œ๋‹ค๊ณ  ๋งํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์˜๋„์ ์œผ๋กœ ๊ผฌ๋ฆฌ๋ฅผ ์ž๋ฅด์ง€ ์•Š๋”๋ผ๋„, ํ›ˆ๋ จ ์„ธํŠธ์˜ ์œ ํ•œํ•œ ํฌ๊ธฐ \(T_{0}\) (์ƒ˜ํ”Œ๋ง ํŽธํ–ฅ)๋Š” ํšจ๊ณผ์ ์ธ ๊ผฌ๋ฆฌ ์ž๋ฅด๊ธฐ๋ฅผ ์œ ๋„ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: ํฌ๊ธฐ \(T_{0}\)์˜ iid ๋ฐ์ดํ„ฐ์…‹์„ ์ƒ˜ํ”Œ๋งํ•˜๊ณ , ํžˆ์Šคํ† ๊ทธ๋žจ \(p_{\text{AI}}\)์„ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค; ์ด ์ƒˆ๋กœ์šด ๋ถ„ํฌ๋Š” AI ๋ฐ์ดํ„ฐ ์ƒ์„ฑ๊ธฐ์˜ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ์ •์ˆ˜ \(i\)๋Š” ํ‰๊ท ์ ์œผ๋กœ \(T_{0}p_{i}\)๋ฒˆ \(p_{\text{AI}}\)์˜ ์ง€์ง€์—์„œ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.
๊ผฌ๋ฆฌ ์ ˆ๋‹จ.์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, AI ์ƒ์„ฑ ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ ์˜๋„์ ์ธ ์„ ํƒ(์˜ˆ: top-\(p\) ๋˜๋Š” top-\(k\) ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก)์€ ์ฆ‰์‹œ \(k\)์—์„œ ์ž˜๋ฆฐ ๊ผฌ๋ฆฌ๋ฅผ ์ดˆ๋ž˜ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์ˆ ๋กœ ๋ณผ ๋•Œ, ๊ฐ€์žฅ ๋นˆ๋ฒˆํ•œ ๊ฒฐ๊ณผ("์Šคํ‚ฌ") ์ค‘ \(k\)๋ฒˆ์งธ ๊ฒฐ๊ณผ๋งŒ ๊ณ ๋ ค๋œ๋‹ค๊ณ  ๋งํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์˜๋„์ ์œผ๋กœ ๊ผฌ๋ฆฌ๊ฐ€ ์ž˜๋ฆฌ์ง€ ์•Š๋”๋ผ๋„, ํ›ˆ๋ จ ์„ธํŠธ์˜ ์œ ํ•œํ•œ ํฌ๊ธฐ \(T_{0}\)(์ƒ˜ํ”Œ๋ง ํŽธํ–ฅ)๋Š” ํšจ๊ณผ์ ์ธ ๊ผฌ๋ฆฌ ์ ˆ๋‹จ์„ ์œ ๋„ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: ํฌ๊ธฐ \(T_{0}\)์˜ iid ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ์ƒ˜ํ”Œ๋งํ•˜๊ณ  ํžˆ์Šคํ† ๊ทธ๋žจ \(p_{\text{AI}}\)๋ฅผ ์ถ”์ •ํ•˜๋ฉด, ์ด ์ƒˆ๋กœ์šด ๋ถ„ํฌ๊ฐ€ AI ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์ž์˜ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ์ •์ˆ˜ \(i\)๋Š” \(p_{\text{AI}}\)์˜ ์ง€์›์— ํ‰๊ท ์ ์œผ๋กœ \(T_{0}p_{i}\)๋งŒํผ ์—ฌ๋Ÿฌ ๋ฒˆ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.
On average with ExactSubstr, we remove more total content than with NearDup (despite ExactSubstr not removing any examples outright)--for example removing \(7.18\%\) of the tokens in C4. The exception is LM1B, where ExactSubstr removes \(8\times\) less data than NearDup. On investigation, we find this is due to the fact that LM1B documents are significantly shorter: \(90\%\) of all documents are under 50 tokens, and so are not even candidates for potential matches even if the entire sequence matched verbatim. We find that both NearDup and ExactSubstr remove similar content--\(77\%\) of the training examples that NearDup removes from C4 have at least one verbatim length-\(50\) match found by ExactSubstr.
ํ‰๊ท ์ ์œผ๋กœ ExactSubstr๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด NearDup๋ณด๋‹ค ๋” ๋งŽ์€ ์ „์ฒด ์ฝ˜ํ…์ธ ๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค (ExactSubstr๊ฐ€ ์–ด๋–ค ์˜ˆ์ œ๋ฅผ ์™„์ „ํžˆ ์ œ๊ฑฐํ•˜์ง€ ์•Š์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ ) -- ์˜ˆ๋ฅผ ๋“ค์–ด C4์—์„œ ํ† ํฐ์˜ \(7.18\%\)๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ์™ธ๋Š” LM1B๋กœ, ์—ฌ๊ธฐ์„œ ExactSubstr๋Š” NearDup๋ณด๋‹ค \(8\times\) ์ ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค. ์กฐ์‚ฌ ๊ฒฐ๊ณผ, ์ด๋Š” LM1B ๋ฌธ์„œ๊ฐ€ ์ƒ๋‹นํžˆ ์งง๊ธฐ ๋•Œ๋ฌธ์ž„์„ ์•Œ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค: ์ „์ฒด ๋ฌธ์„œ์˜ \(90\%\)๊ฐ€ 50 ํ† ํฐ ์ดํ•˜์ด๋ฉฐ, ๋”ฐ๋ผ์„œ ์ „์ฒด ์‹œํ€€์Šค๊ฐ€ ๊ทธ๋Œ€๋กœ ์ผ์น˜ํ•˜๋”๋ผ๋„ ์ž ์žฌ์  ์ผ์น˜ ํ›„๋ณด๊ฐ€ ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. NearDup๊ณผ ExactSubstr ๋ชจ๋‘ ์œ ์‚ฌํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ œ๊ฑฐํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค -- NearDup๊ฐ€ C4์—์„œ ์ œ๊ฑฐํ•œ ํ›ˆ๋ จ ์˜ˆ์ œ์˜ \(77\%\)๋Š” ExactSubstr์— ์˜ํ•ด ๋ฐœ๊ฒฌ๋œ ์ตœ์†Œ ํ•˜๋‚˜์˜ ๊ธธ์ด-\(50\) ์ผ์น˜ ํ•ญ๋ชฉ์„ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
ํ‰๊ท ์ ์œผ๋กœ ExactSubstr์€ NearDup๋ณด๋‹ค ๋” ๋งŽ์€ ์ด ์ฝ˜ํ…์ธ ๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค(์˜ˆ: C4์—์„œ ํ† ํฐ์˜ \(7.18\%\) ์ œ๊ฑฐ). ์˜ˆ์™ธ์ ์œผ๋กœ ExactSubstr์€ NearDup๋ณด๋‹ค \(8๋ฐฐ\) ์ ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” LM1B๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์กฐ์‚ฌ ๊ฒฐ๊ณผ, ์ด๋Š” LM1B ๋ฌธ์„œ์˜ ๊ธธ์ด๊ฐ€ ํ˜„์ €ํžˆ ์งง๊ธฐ ๋•Œ๋ฌธ์ธ ๊ฒƒ์œผ๋กœ ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค. ์ „์ฒด ๋ฌธ์„œ์˜ \(90\%\)๊ฐ€ 50 ํ† ํฐ ๋ฏธ๋งŒ์ด๋ฏ€๋กœ ์ „์ฒด ์‹œํ€€์Šค๊ฐ€ ๋‹จ์–ด ๊ทธ๋Œ€๋กœ ์ผ์น˜ํ•˜๋”๋ผ๋„ ์ž ์žฌ์  ์ผ์น˜ ํ›„๋ณด๊ฐ€ ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. NearDup๊ณผ ExactSubstr์€ ๋น„์Šทํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. NearDup์ด C4์—์„œ ์ œ๊ฑฐํ•œ ํ›ˆ๋ จ ์˜ˆ์ œ์˜ \(77\%\)๋Š” ExactSubstr์—์„œ ๋ฐœ๊ฒฌํ•œ ์ ์–ด๋„ ํ•˜๋‚˜์˜ ๋ฒ„๋ฐ”ํ‹ฑ ๊ธธ์ด-\(50\) ์ผ์น˜ํ•˜๋Š” ๋ถ€๋ถ„์ด ์žˆ์Šต๋‹ˆ๋‹ค.
AMIE's diagnostic accuracy was assessed as higher than that of PCPs. Figure 3 shows the top-k accuracy for AMIE and PCPs, considering matches with the ground truth diagnosis (a) and matches with any item on the accepted differential (b). AMIE showed significantly higher top-k accuracy than that of PCPs across all values of k (\(p<0.05\)). Note that unlike AMIE, PCPs did not always provide 10 diagnoses in their differential diagnoses (min: 3, mean: 5.39). Additionally, we performed a comparison of DDx accuracy between AMIE and PCP by varying the matching criteria for determining a match. Results depicted in Figure A.7 further substantiate AMIE's superior DDx performance across various matching criteria.
AMIE์˜ ์ง„๋‹จ ์ •ํ™•๋„๋Š” PCPs๋ณด๋‹ค ๋†’์€ ๊ฒƒ์œผ๋กœ ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค. Figure 3์€ ์‹ค์ œ ์ง„๋‹จ๊ณผ์˜ ์ผ์น˜(a) ๋ฐ ์Šน์ธ๋œ ์ฐจ๋ณ„ ์ง„๋‹จ ํ•ญ๋ชฉ๊ณผ์˜ ์ผ์น˜(b)๋ฅผ ๊ณ ๋ คํ•˜์—ฌ AMIE์™€ PCPs์˜ top-k ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. AMIE๋Š” ๋ชจ๋“  k ๊ฐ’์—์„œ PCPs๋ณด๋‹ค ์œ ์˜๋ฏธํ•˜๊ฒŒ ๋†’์€ top-k ์ •ํ™•๋„๋ฅผ ๋ณด์˜€์Šต๋‹ˆ๋‹ค (\(p<0.05\)). AMIE์™€ ๋‹ฌ๋ฆฌ PCPs๋Š” ํ•ญ์ƒ 10๊ฐœ์˜ ์ง„๋‹จ์„ ์ œ๊ณตํ•˜์ง€๋Š” ์•Š์•˜์Šต๋‹ˆ๋‹ค (์ตœ์†Œ: 3, ํ‰๊ท : 5.39). ์ถ”๊ฐ€์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ผ์น˜ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๋งค์นญ ๊ธฐ์ค€์„ ๋ณ€๊ฒฝํ•˜์—ฌ AMIE์™€ PCP์˜ DDx ์ •ํ™•๋„๋ฅผ ๋น„๊ตํ–ˆ์Šต๋‹ˆ๋‹ค. Figure A.7์— ๋‚˜ํƒ€๋‚œ ๊ฒฐ๊ณผ๋Š” ๋‹ค์–‘ํ•œ ๋งค์นญ ๊ธฐ์ค€์—์„œ AMIE์˜ ์šฐ์ˆ˜ํ•œ DDx ์„ฑ๋Šฅ์„ ๋”์šฑ ๋’ท๋ฐ›์นจํ•ฉ๋‹ˆ๋‹ค.
AMIE์˜ ์ง„๋‹จ ์ •ํ™•๋„๋Š” PCP๋ณด๋‹ค ๋†’๊ฒŒ ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 3์€ ์‹ค์ธก ์ง„๋‹จ๊ณผ์˜ ์ผ์น˜ ์—ฌ๋ถ€(a)์™€ ํ—ˆ์šฉ๋œ ์ฐจ๋“ฑ ํ•ญ๋ชฉ๊ณผ์˜ ์ผ์น˜ ์—ฌ๋ถ€(b)๋ฅผ ๊ณ ๋ คํ•œ AMIE์™€ PCP์˜ ์ƒ์œ„ k ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. AMIE๋Š” ๋ชจ๋“  k ๊ฐ’์—์„œ PCP๋ณด๋‹ค ์ƒ์œ„ k ์ •ํ™•๋„๊ฐ€ ํ›จ์”ฌ ๋†’์•˜์Šต๋‹ˆ๋‹ค(\(p<0.05\)). AMIE์™€ ๋‹ฌ๋ฆฌ PCP๋Š” ๋ฏธ๋ถ„ ์ง„๋‹จ์—์„œ ํ•ญ์ƒ 10๊ฐœ์˜ ์ง„๋‹จ์„ ์ œ๊ณตํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค(์ตœ์†Œ: 3, ํ‰๊ท : 5.39). ๋˜ํ•œ ์ผ์น˜ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•œ ์ผ์น˜ ๊ธฐ์ค€์„ ๋ณ€๊ฒฝํ•˜์—ฌ AMIE์™€ PCP ๊ฐ„์˜ DDx ์ •ํ™•๋„ ๋น„๊ต๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ A.7์— ํ‘œ์‹œ๋œ ๊ฒฐ๊ณผ๋Š” ๋‹ค์–‘ํ•œ ์ผ์น˜ ๊ธฐ์ค€์—์„œ AMIE์˜ ์šฐ์ˆ˜ํ•œ DDx ์„ฑ๋Šฅ์„ ๋”์šฑ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค.
Multilingual data.Using the language identification filter, we classify processed CommonCrawl data into 176 languages. Figure 6 shows the top 20 languages present in the data _excluding English_, based on their relative contribution in descending order. 58.20% of all documents in the processed CommonCrawl data were identified as English. We find the distribution of languages in CommonCrawl to only be partially aligned with the worldwide distribution of language speakers (Eberhard et al., 2023): Russian is over-represented (2nd in CC but only 8th worldwide), Mandarin Chinese is under-represented (6-7th in CC but 2nd worldwide), and Hindi does not show-up in the top 20 despite being the 3rd most spoken.
๋‹ค๊ตญ์–ด ๋ฐ์ดํ„ฐ. ์–ธ์–ด ์‹๋ณ„ ํ•„ํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” ์ฒ˜๋ฆฌ๋œ CommonCrawl ๋ฐ์ดํ„ฐ๋ฅผ 176๊ฐœ ์–ธ์–ด๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. Figure 6์€ _์˜์–ด๋ฅผ ์ œ์™ธํ•œ_ ๋ฐ์ดํ„ฐ์— ์กด์žฌํ•˜๋Š” ์ƒ์œ„ 20๊ฐœ ์–ธ์–ด๋ฅผ ์ƒ๋Œ€์  ๊ธฐ์—ฌ๋„์— ๋”ฐ๋ผ ๋‚ด๋ฆผ์ฐจ์ˆœ์œผ๋กœ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ฒ˜๋ฆฌ๋œ CommonCrawl ๋ฐ์ดํ„ฐ์˜ ๋ชจ๋“  ๋ฌธ์„œ ์ค‘ 58.20%๊ฐ€ ์˜์–ด๋กœ ์‹๋ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” CommonCrawl์˜ ์–ธ์–ด ๋ถ„ํฌ๊ฐ€ ์ „ ์„ธ๊ณ„ ์–ธ์–ด ์‚ฌ์šฉ์ž์˜ ๋ถ„ํฌ์™€ ๋ถ€๋ถ„์ ์œผ๋กœ๋งŒ ์ผ์น˜ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค (Eberhard et al., 2023): ๋Ÿฌ์‹œ์•„์–ด๋Š” ๊ณผ๋Œ€ ๋Œ€ํ‘œ๋˜์–ด ์žˆ์œผ๋ฉฐ (CC์—์„œ 2์œ„์ง€๋งŒ ์ „ ์„ธ๊ณ„์ ์œผ๋กœ๋Š” 8์œ„), ์ค‘๊ตญ์–ด๋Š” ๊ณผ์†Œ ๋Œ€ํ‘œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค (CC์—์„œ 6-7์œ„์ง€๋งŒ ์ „ ์„ธ๊ณ„์ ์œผ๋กœ๋Š” 2์œ„), ํžŒ๋””์–ด๋Š” ์ „ ์„ธ๊ณ„์ ์œผ๋กœ 3๋ฒˆ์งธ๋กœ ๋งŽ์ด ์‚ฌ์šฉ๋˜๋Š” ์–ธ์–ด์ž„์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ƒ์œ„ 20์œ„์— ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
๋‹ค๊ตญ์–ด ๋ฐ์ดํ„ฐ.์–ธ์–ด ์‹๋ณ„ ํ•„ํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฒ˜๋ฆฌ๋œ CommonCrawl ๋ฐ์ดํ„ฐ๋ฅผ 176๊ฐœ ์–ธ์–ด๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 6์€ ์˜์–ด๋ฅผ ์ œ์™ธํ•œ ๋ฐ์ดํ„ฐ์— ํฌํ•จ๋œ ์ƒ๋Œ€์  ๊ธฐ์—ฌ๋„์— ๋”ฐ๋ผ ๋‚ด๋ฆผ์ฐจ์ˆœ์œผ๋กœ ํ‘œ์‹œ๋œ ์ƒ์œ„ 20๊ฐœ ์–ธ์–ด๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ฒ˜๋ฆฌ๋œ CommonCrawl ๋ฐ์ดํ„ฐ์˜ ๋ชจ๋“  ๋ฌธ์„œ ์ค‘ 58.20%๊ฐ€ ์˜์–ด๋กœ ์‹๋ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. CommonCrawl์˜ ์–ธ์–ด ๋ถ„ํฌ๋Š” ์ „ ์„ธ๊ณ„ ์–ธ์–ด ์‚ฌ์šฉ์ž ๋ถ„ํฌ(Eberhard et al., 2023)์™€ ๋ถ€๋ถ„์ ์œผ๋กœ๋งŒ ์ผ์น˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค: ๋Ÿฌ์‹œ์•„์–ด๋Š” ๊ณผ๋Œ€ ํ‘œํ˜„(CC์—์„œ๋Š” 2์œ„์ด์ง€๋งŒ ์ „ ์„ธ๊ณ„์—์„œ๋Š” 8์œ„), ๋งŒ๋‹ค๋ฆฐ ์ค‘๊ตญ์–ด๋Š” ๊ณผ์†Œ ํ‘œํ˜„(CC์—์„œ๋Š” 6~7์œ„์ด์ง€๋งŒ ์ „ ์„ธ๊ณ„์—์„œ๋Š” 2์œ„), ํžŒ๋””์–ด๋Š” 3๋ฒˆ์งธ๋กœ ๋งŽ์ด ์‚ฌ์šฉ๋˜๋Š” ์–ธ์–ด์ž„์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ƒ์œ„ 20์œ„์— ๋“ค์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค.
Our results suggested that AMIE was as adept as PCPs in eliciting pertinent information during the simulated consultations and was more accurate than PCPs in formulating a complete differential diagnosis if given the same amount of acquired information. This finding corroborates other work that LLMs may be able to produce more complete differential diagnoses given the same clinical information as physicians in challenging cases [70]. Though not explored in this study, the assistive performance of AMIE therefore represents an interesting and important avenue for future research, particularly given the real-world importance of expertoversight for AI systems in safety-critical settings such as medicine.
์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋Š” AMIE๊ฐ€ ๋ชจ์˜ ์ƒ๋‹ด ์ค‘ ๊ด€๋ จ ์ •๋ณด๋ฅผ ์ด๋Œ์–ด๋‚ด๋Š” ๋ฐ PCPs๋งŒํผ ๋Šฅ์ˆ™ํ•˜๋ฉฐ, ๋™์ผํ•œ ์–‘์˜ ์ •๋ณด๋ฅผ ์ œ๊ณต๋ฐ›์•˜์„ ๋•Œ ์™„์ „ํ•œ ๊ฐ๋ณ„ ์ง„๋‹จ์„ ์ž‘์„ฑํ•˜๋Š” ๋ฐ PCPs๋ณด๋‹ค ๋” ์ •ํ™•ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐœ๊ฒฌ์€ LLMs๊ฐ€ ์–ด๋ ค์šด ์‚ฌ๋ก€์—์„œ ์˜์‚ฌ์™€ ๋™์ผํ•œ ์ž„์ƒ ์ •๋ณด๋ฅผ ์ œ๊ณต๋ฐ›์•˜์„ ๋•Œ ๋” ์™„์ „ํ•œ ๊ฐ๋ณ„ ์ง„๋‹จ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๋‹ค๋ฅธ ์—ฐ๊ตฌ๋ฅผ ๋’ท๋ฐ›์นจํ•ฉ๋‹ˆ๋‹ค [70]. ๋น„๋ก ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ํƒ๊ตฌ๋˜์ง€ ์•Š์•˜์ง€๋งŒ, ๋”ฐ๋ผ์„œ AMIE์˜ ๋ณด์กฐ ์„ฑ๋Šฅ์€ ํŠนํžˆ ์˜ํ•™๊ณผ ๊ฐ™์€ ์•ˆ์ „์ด ์ค‘์š”ํ•œ ํ™˜๊ฒฝ์—์„œ AI ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€ ๊ฐ๋…์˜ ์‹ค์งˆ์ ์ธ ์ค‘์š”์„ฑ์„ ๊ณ ๋ คํ•  ๋•Œ, ๋ฏธ๋ž˜ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ํฅ๋ฏธ๋กญ๊ณ  ์ค‘์š”ํ•œ ๊ฒฝ๋กœ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ๋”ฐ๋ฅด๋ฉด AMIE๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ์ƒ๋‹ด ์ค‘์— ๊ด€๋ จ ์ •๋ณด๋ฅผ ์ด๋Œ์–ด๋‚ด๋Š” ๋ฐ ์žˆ์–ด PCP์™€ ๋™๋“ฑํ•˜๋ฉฐ, ๋™์ผํ•œ ์–‘์˜ ์ •๋ณด๋ฅผ ์ œ๊ณต๋ฐ›์€ ๊ฒฝ์šฐ ์™„์ „ํ•œ ๊ฐ๋ณ„ ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ์žˆ์–ด PCP๋ณด๋‹ค ๋” ์ •ํ™•ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ๋‚œ์ด๋„๊ฐ€ ๋†’์€ ์‚ฌ๋ก€์—์„œ ์˜์‚ฌ์™€ ๋™์ผํ•œ ์ž„์ƒ ์ •๋ณด๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ LLM์ด ๋” ์™„๋ฒฝํ•œ ๊ฐ๋ณ„ ์ง„๋‹จ์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋‹ค๋Š” ๋‹ค๋ฅธ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋ฅผ ๋’ท๋ฐ›์นจํ•ฉ๋‹ˆ๋‹ค[70]. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ํƒ๊ตฌํ•˜์ง€ ์•Š์•˜์ง€๋งŒ, AMIE์˜ ๋ณด์กฐ ์„ฑ๋Šฅ์€ ์˜๋ฃŒ์™€ ๊ฐ™์ด ์•ˆ์ „์ด ์ค‘์š”ํ•œ ๋ถ„์•ผ์—์„œ AI ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€ ๊ฐ๋…์˜ ์‹ค์ œ ์ค‘์š”์„ฑ์„ ๊ณ ๋ คํ•  ๋•Œ ํ–ฅํ›„ ์—ฐ๊ตฌ์— ์žˆ์–ด ํฅ๋ฏธ๋กญ๊ณ  ์ค‘์š”ํ•œ ์˜์—ญ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Optimizer.** The Adam optimizer [328] and AdamW optimizer [329] are widely utilized for training LLMs (_e.g.,_ GPT-3), which are based on adaptive estimates of lower-order moments for first-order gradient-based optimization. Commonly, its hyper-parameters are set as follows: \(\beta_{1}=0.9\), \(\beta_{2}=0.95\) and \(\epsilon=10^{-8}\). Meanwhile, the Adafactor optimizer [330] has also been utilized in training LLMs (_e.g.,_ PaLM and T5), which is a variant of the Adam optimizer specially designed for conserving GPU memory during training. The hyper-parameters of the Adafactor optimizer are set as: \(\beta_{1}=0.9\) and \(\beta_{2}=1.0-k^{-0.8}\), where \(k\) denotes the number of training steps.
**Optimizer.** Adam ์˜ตํ‹ฐ๋งˆ์ด์ € [328]์™€ AdamW ์˜ตํ‹ฐ๋งˆ์ด์ € [329]๋Š” LLMs (์˜ˆ: GPT-3) ํ›ˆ๋ จ์— ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋ฉฐ, ์ด๋Š” 1์ฐจ ๊ธฐ์šธ๊ธฐ ๊ธฐ๋ฐ˜ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•œ ํ•˜์œ„ ์ฐจ์ˆ˜ ๋ชจ๋ฉ˜ํŠธ์˜ ์ ์‘ ์ถ”์ •์— ๊ธฐ๋ฐ˜ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ, ์ด๋“ค์˜ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์„ค์ •๋ฉ๋‹ˆ๋‹ค: \(\beta_{1}=0.9\), \(\beta_{2}=0.95\) ๋ฐ \(\epsilon=10^{-8}\). ํ•œํŽธ, Adafactor ์˜ตํ‹ฐ๋งˆ์ด์ € [330]๋„ LLMs (์˜ˆ: PaLM ๋ฐ T5) ํ›ˆ๋ จ์— ์‚ฌ์šฉ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” ํ›ˆ๋ จ ์ค‘ GPU ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ ˆ์•ฝํ•˜๊ธฐ ์œ„ํ•ด ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ Adam ์˜ตํ‹ฐ๋งˆ์ด์ €์˜ ๋ณ€ํ˜•์ž…๋‹ˆ๋‹ค. Adafactor ์˜ตํ‹ฐ๋งˆ์ด์ €์˜ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์„ค์ •๋ฉ๋‹ˆ๋‹ค: \(\beta_{1}=0.9\) ๋ฐ \(\beta_{2}=1.0-k^{-0.8}\), ์—ฌ๊ธฐ์„œ \(k\)๋Š” ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
**์ตœ์ ํ™” ๋„๊ตฌ.** ์•„๋‹ด ์ตœ์ ํ™” ๋„๊ตฌ [328] ๋ฐ ์•„๋‹ดW ์ตœ์ ํ™” ๋„๊ตฌ [329]๋Š” 1์ฐจ ๊ฒฝ์‚ฌ ๊ธฐ๋ฐ˜ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•œ ํ•˜์œ„ ์ˆœ๊ฐ„์˜ ์ ์‘ํ˜• ์ถ”์ •์น˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” LLM(์˜ˆ: GPT-3)์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ๋„๋ฆฌ ํ™œ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์„ค์ •๋ฉ๋‹ˆ๋‹ค: \(\beta_{1}=0.9\), \(\beta_{2}=0.95\), \(\epsilon=10^{-8}\). ํ•œํŽธ, Adafactor ์ตœ์ ํ™” ๋„๊ตฌ [330]๋Š” ํ›ˆ๋ จ ์ค‘ GPU ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ ˆ์•ฝํ•˜๋„๋ก ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ์•„๋‹ด ์ตœ์ ํ™” ๋„๊ตฌ์˜ ๋ณ€ํ˜•์ธ LLM(์˜ˆ: PaLM ๋ฐ T5) ํ›ˆ๋ จ์—๋„ ํ™œ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. Adafactor ์ตœ์ ํ™” ๋„๊ตฌ์˜ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์„ค์ •๋ฉ๋‹ˆ๋‹ค: \(\beta_{1}=0.9\), \(\beta_{2}=1.0-k^{-0.8}\), ์—ฌ๊ธฐ์„œ \(k\)๋Š” ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
Which components of the distribution does the inverter need?To investigate whether our model focuses most on the largest components of its input, we iteratively remove (set to the mean) \(k\) components from the probability vector in both ascending and descending order. We also consider removing all but a random subset of \(k\) components from the input. Figure 4 highlights the difference in reconstruction performance across levels of component removal. It appears that the model focuses more on the more likely words. In particular, the smallest \(k\) probabilities are only slightly more useful than a random \(k\). Reconstruction is poor in general until almost the entire distribution is re-included.
๋ถ„ํฌ์˜ ์–ด๋–ค ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์ธ๋ฒ„ํ„ฐ์— ํ•„์š”ํ•ฉ๋‹ˆ๊นŒ? ์šฐ๋ฆฌ์˜ ๋ชจ๋ธ์ด ์ž…๋ ฅ์˜ ๊ฐ€์žฅ ํฐ ๊ตฌ์„ฑ ์š”์†Œ์— ์ฃผ๋กœ ์ง‘์ค‘ํ•˜๋Š”์ง€ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ํ™•๋ฅ  ๋ฒกํ„ฐ์—์„œ \(k\) ๊ฐœ์˜ ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ์˜ค๋ฆ„์ฐจ์ˆœ ๋ฐ ๋‚ด๋ฆผ์ฐจ์ˆœ์œผ๋กœ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ œ๊ฑฐ(ํ‰๊ท ์œผ๋กœ ์„ค์ •)ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž…๋ ฅ์—์„œ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ \(k\) ๊ฐœ์˜ ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ์ œ์™ธํ•œ ๋ชจ๋“  ๊ฒƒ์„ ์ œ๊ฑฐํ•˜๋Š” ๊ฒฝ์šฐ๋„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. Figure 4๋Š” ๊ตฌ์„ฑ ์š”์†Œ ์ œ๊ฑฐ ์ˆ˜์ค€์— ๋”ฐ๋ฅธ ์žฌ๊ตฌ์„ฑ ์„ฑ๋Šฅ์˜ ์ฐจ์ด๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ๋” ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ๋‹จ์–ด์— ๋” ์ง‘์ค‘ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ํŠนํžˆ, ๊ฐ€์žฅ ์ž‘์€ \(k\) ๊ฐœ์˜ ํ™•๋ฅ ์€ ๋ฌด์ž‘์œ„ \(k\) ๊ฐœ๋ณด๋‹ค ์•ฝ๊ฐ„ ๋” ์œ ์šฉํ•  ๋ฟ์ž…๋‹ˆ๋‹ค. ๊ฑฐ์˜ ์ „์ฒด ๋ถ„ํฌ๊ฐ€ ๋‹ค์‹œ ํฌํ•จ๋  ๋•Œ๊นŒ์ง€ ์žฌ๊ตฌ์„ฑ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ €์กฐํ•ฉ๋‹ˆ๋‹ค.
์ธ๋ฒ„ํ„ฐ์— ํ•„์š”ํ•œ ๋ถ„ํฌ์˜ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ๋ฌด์—‡์ธ๊ฐ€์š”?๋ชจ๋ธ์ด ์ž…๋ ฅ์˜ ๊ฐ€์žฅ ํฐ ๊ตฌ์„ฑ ์š”์†Œ์— ๊ฐ€์žฅ ์ง‘์ค‘ํ•˜๋Š”์ง€ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ํ™•๋ฅ  ๋ฒกํ„ฐ์—์„œ ์˜ค๋ฆ„์ฐจ์ˆœ๊ณผ ๋‚ด๋ฆผ์ฐจ์ˆœ์œผ๋กœ ๋ฐ˜๋ณต์ ์œผ๋กœ \(k\) ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ์ œ๊ฑฐ(ํ‰๊ท ์œผ๋กœ ์„ค์ •)ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž…๋ ฅ์—์„œ ๋ฌด์ž‘์œ„ ํ•˜์œ„ ์ง‘ํ•ฉ์ธ \(k\) ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ์ œ์™ธํ•œ ๋ชจ๋“  ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ๋„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 4๋Š” ๊ตฌ์„ฑ ์š”์†Œ ์ œ๊ฑฐ ์ˆ˜์ค€์— ๋”ฐ๋ฅธ ์žฌ๊ตฌ์„ฑ ์„ฑ๋Šฅ์˜ ์ฐจ์ด๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ๋” ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ๋‹จ์–ด์— ๋” ์ง‘์ค‘ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ํŠนํžˆ ๊ฐ€์žฅ ์ž‘์€ \(k\) ํ™•๋ฅ ์€ ๋ฌด์ž‘์œ„ \(k\)๋ณด๋‹ค ์•ฝ๊ฐ„ ๋” ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฑฐ์˜ ์ „์ฒด ๋ถ„ํฌ๊ฐ€ ๋‹ค์‹œ ํฌํ•จ๋  ๋•Œ๊นŒ์ง€ ์ผ๋ฐ˜์ ์œผ๋กœ ์žฌ๊ตฌ์„ฑ์ด ์ข‹์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
Compressive memory systems promise to be more scalable and efficient than the attention mechanism for extremely long sequences (Kanerva, 1988; Munkhdalai et al., 2019). Instead of using an array that grows with the input sequence length, a compressive memory primarily maintains a fixed number of parameters to store and recall information with a bounded storage and computation costs. In the compressive memory, new information is added to the memory by changing its parameters with an objective that this information can be recovered back later on. However, the LLMs in their current state have yet to see an effective, practical compressive memory technique that balances simplicity along with quality.
์••์ถ• ๋ฉ”๋ชจ๋ฆฌ ์‹œ์Šคํ…œ์€ ๋งค์šฐ ๊ธด ์‹œํ€€์Šค์— ๋Œ€ํ•ด ์ฃผ์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜๋ณด๋‹ค ๋” ํ™•์žฅ ๊ฐ€๋Šฅํ•˜๊ณ  ํšจ์œจ์ ์ผ ๊ฒƒ์„ ์•ฝ์†ํ•ฉ๋‹ˆ๋‹ค (Kanerva, 1988; Munkhdalai et al., 2019). ์ž…๋ ฅ ์‹œํ€€์Šค ๊ธธ์ด์— ๋”ฐ๋ผ ๋ฐฐ์—ด์ด ์ปค์ง€๋Š” ๋Œ€์‹ , ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ๋Š” ์ฃผ๋กœ ๊ณ ์ •๋œ ์ˆ˜์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์œ ์ง€ํ•˜์—ฌ ์ œํ•œ๋œ ์ €์žฅ ๋ฐ ๊ณ„์‚ฐ ๋น„์šฉ์œผ๋กœ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๊ณ  ํšŒ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ์—์„œ๋Š” ์ƒˆ๋กœ์šด ์ •๋ณด๊ฐ€ ๋‚˜์ค‘์— ํšŒ์ˆ˜๋  ์ˆ˜ ์žˆ๋„๋ก ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ๋ณ€๊ฒฝํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ์— ์ถ”๊ฐ€๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ˜„์žฌ ์ƒํƒœ์˜ LLM๋“ค์€ ๋‹จ์ˆœ์„ฑ๊ณผ ํ’ˆ์งˆ์„ ๊ท ํ˜• ์žˆ๊ฒŒ ์œ ์ง€ํ•˜๋Š” ํšจ๊ณผ์ ์ด๊ณ  ์‹ค์šฉ์ ์ธ ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ ๊ธฐ์ˆ ์„ ์•„์ง ๋ณด์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค.
์••์ถ• ๋ฉ”๋ชจ๋ฆฌ ์‹œ์Šคํ…œ์€ ๋งค์šฐ ๊ธด ์‹œํ€€์Šค์— ๋Œ€ํ•ด ์ฃผ์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜๋ณด๋‹ค ํ™•์žฅ์„ฑ๊ณผ ํšจ์œจ์„ฑ์ด ๋” ๋†’์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค(Kanerva, 1988; Munkhdalai et al., 2019). ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ๋Š” ์ž…๋ ฅ ์‹œํ€€์Šค ๊ธธ์ด์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•˜๋Š” ๋ฐฐ์—ด์„ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹ , ์ฃผ๋กœ ์ œํ•œ๋œ ์ €์žฅ ๋ฐ ๊ณ„์‚ฐ ๋น„์šฉ์œผ๋กœ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๊ณ  ๋ถˆ๋Ÿฌ์˜ค๊ธฐ ์œ„ํ•ด ๊ณ ์ •๋œ ์ˆ˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ƒˆ๋กœ์šด ์ •๋ณด๋Š” ์ด ์ •๋ณด๋ฅผ ๋‚˜์ค‘์— ๋‹ค์‹œ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋ณ€๊ฒฝํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ์— ์ถ”๊ฐ€๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ˜„์žฌ ์ƒํƒœ์˜ LLM์€ ๋‹จ์ˆœ์„ฑ๊ณผ ํ’ˆ์งˆ์˜ ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ํšจ๊ณผ์ ์ด๊ณ  ์‹ค์šฉ์ ์ธ ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ ๊ธฐ์ˆ ์„ ์•„์ง ๋ณด์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค.
Exploring omission errors:As mentioned above, omission of key information plagues all LLM summarizers. To better understand the nature of the omission errors identified by our annotators, we categorize them into the following categories: _characters_, _events_, _details_, _relationships_, _themes_.12 Figure 4 shows a heatmap of omission errors broken down by model. A large proportion of summaries (33.3% to 65.4%) lack mentions of key events, creating gaps in the overall narrative, and we also note omissions of significant details about the characters, events, or objects (16.7% to 38.5%). Furthermore, GPT-4-Turbo and Mlxtral have a tendency to entirely omit mentions of crucial characters (23.1%).
๋ˆ„๋ฝ ์˜ค๋ฅ˜ ํƒ์ƒ‰: ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ์ฃผ์š” ์ •๋ณด์˜ ๋ˆ„๋ฝ์€ ๋ชจ๋“  LLM ์š”์•ฝ๊ธฐ์—์„œ ๋ฌธ์ œ๋ฅผ ์ผ์œผํ‚ต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ์ฃผ์„์ž๋“ค์ด ์‹๋ณ„ํ•œ ๋ˆ„๋ฝ ์˜ค๋ฅ˜์˜ ์„ฑ๊ฒฉ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์ด๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฒ”์ฃผ๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค: _์บ๋ฆญํ„ฐ_, _์ด๋ฒคํŠธ_, _์„ธ๋ถ€์‚ฌํ•ญ_, _๊ด€๊ณ„_, _์ฃผ์ œ_. Figure 4๋Š” ๋ชจ๋ธ๋ณ„๋กœ ๋ถ„๋ฅ˜๋œ ๋ˆ„๋ฝ ์˜ค๋ฅ˜์˜ ํžˆํŠธ๋งต์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋งŽ์€ ์š”์•ฝ๋ณธ(33.3%์—์„œ 65.4%)์ด ์ฃผ์š” ์ด๋ฒคํŠธ ์–ธ๊ธ‰์„ ๊ฒฐ์—ฌํ•˜๊ณ  ์žˆ์–ด ์ „์ฒด ์„œ์‚ฌ์— ๊ณต๋ฐฑ์„ ๋งŒ๋“ค๊ณ  ์žˆ์œผ๋ฉฐ, ์บ๋ฆญํ„ฐ, ์ด๋ฒคํŠธ ๋˜๋Š” ๊ฐ์ฒด์— ๋Œ€ํ•œ ์ค‘์š”ํ•œ ์„ธ๋ถ€์‚ฌํ•ญ์˜ ๋ˆ„๋ฝ(16.7%์—์„œ 38.5%)๋„ ํ™•์ธ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, GPT-4-Turbo์™€ Mlxtral์€ ์ค‘์š”ํ•œ ์บ๋ฆญํ„ฐ ์–ธ๊ธ‰์„ ์™„์ „ํžˆ ๋ˆ„๋ฝํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค(23.1%).
๋ˆ„๋ฝ ์˜ค๋ฅ˜ ํƒ์ƒ‰: ์œ„์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ์ฃผ์š” ์ •๋ณด์˜ ๋ˆ„๋ฝ์€ ๋ชจ๋“  LLM ์š”์•ฝ๊ธฐ์— ๋ฌธ์ œ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ์ฃผ์„ ์ž‘์„ฑ์ž๊ฐ€ ํ™•์ธํ•œ ๋ˆ„๋ฝ ์˜ค๋ฅ˜์˜ ํŠน์„ฑ์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด _์ธ๋ฌผ_, _์ด๋ฒคํŠธ_, _์„ธ๋ถ€ ์‚ฌํ•ญ_, _๊ด€๊ณ„_, _์ฃผ์ œ_์˜ ๋‹ค์Œ ๋ฒ”์ฃผ๋กœ ๋ถ„๋ฅ˜ํ–ˆ์Šต๋‹ˆ๋‹ค.12 ๊ทธ๋ฆผ 4๋Š” ๋ชจ๋ธ๋ณ„๋กœ ๋ถ„๋ฅ˜๋œ ๋ˆ„๋ฝ ์˜ค๋ฅ˜์˜ ํžˆํŠธ๋งต์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์š”์•ฝ์˜ ์ƒ๋‹น์ˆ˜(33.3%~65.4%)์— ์ฃผ์š” ์‚ฌ๊ฑด์— ๋Œ€ํ•œ ์–ธ๊ธ‰์ด ์—†์–ด ์ „์ฒด ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ์— ๊ณต๋ฐฑ์ด ์ƒ๊ธฐ๊ณ , ์ธ๋ฌผ, ์‚ฌ๊ฑด ๋˜๋Š” ์‚ฌ๋ฌผ์— ๋Œ€ํ•œ ์ค‘์š”ํ•œ ์„ธ๋ถ€ ์‚ฌํ•ญ๋„ ๋ˆ„๋ฝ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•˜์Šต๋‹ˆ๋‹ค(16.7%~38.5%). ๋˜ํ•œ GPT-4-Turbo์™€ Mlxtral์€ ์ค‘์š”ํ•œ ์ธ๋ฌผ์— ๋Œ€ํ•œ ์–ธ๊ธ‰์„ ์™„์ „ํžˆ ์ƒ๋žตํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค(23.1%).
**Seeking for Advice**. Despite all our efforts, this survey is still far from perfect: we are likely to miss important references or topics, and might also have non-rigorous expressions or discussions. We will continuously update this survey, and improve the quality as much as we can. For us, survey writing is also a learning process for LLMs by ourselves. For readers with constructive suggestions to improve this survey, you are welcome to leave comments on the GitHub page of our survey or directly email our authors. We will make revisions following the received comments or suggestions in a future version, and acknowledge the readers who have contributed constructive suggestions in our survey.
**์กฐ์–ธ์„ ๊ตฌํ•ฉ๋‹ˆ๋‹ค**. ์šฐ๋ฆฌ์˜ ๋ชจ๋“  ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ์—ฌ์ „ํžˆ ์™„๋ฒฝ๊ณผ๋Š” ๊ฑฐ๋ฆฌ๊ฐ€ ๋ฉ‰๋‹ˆ๋‹ค: ์ค‘์š”ํ•œ ์ฐธ๊ณ  ๋ฌธํ—Œ์ด๋‚˜ ์ฃผ์ œ๋ฅผ ๋†“์น  ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ์œผ๋ฉฐ, ๋น„์ฒด๊ณ„์ ์ธ ํ‘œํ˜„์ด๋‚˜ ๋…ผ์˜๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋ฅผ ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธํ•˜๊ณ , ๊ฐ€๋Šฅํ•œ ํ•œ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์—๊ฒŒ ์„ค๋ฌธ์กฐ์‚ฌ ์ž‘์„ฑ์€ LLMs์— ๋Œ€ํ•œ ํ•™์Šต ๊ณผ์ •์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋ฅผ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ๊ฑด์„ค์ ์ธ ์ œ์•ˆ์„ ๊ฐ€์ง„ ๋…์ž๋“ค์€ GitHub ํŽ˜์ด์ง€์— ๋Œ“๊ธ€์„ ๋‚จ๊ธฐ๊ฑฐ๋‚˜ ์ €์ž์—๊ฒŒ ์ง์ ‘ ์ด๋ฉ”์ผ์„ ๋ณด๋‚ด์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ฐ›์€ ๋Œ“๊ธ€์ด๋‚˜ ์ œ์•ˆ์„ ๋ฐ”ํƒ•์œผ๋กœ ํ–ฅํ›„ ๋ฒ„์ „์—์„œ ์ˆ˜์ •ํ•  ๊ฒƒ์ด๋ฉฐ, ๊ฑด์„ค์ ์ธ ์ œ์•ˆ์„ ํ•ด์ฃผ์‹  ๋…์ž๋“ค์„ ์„ค๋ฌธ์กฐ์‚ฌ์— ๊ฐ์‚ฌ์˜ ๋œป์œผ๋กœ ์–ธ๊ธ‰ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
**์กฐ์–ธ์„ ๊ตฌํ•ฉ๋‹ˆ๋‹ค**. ๋ชจ๋“  ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ์—ฌ์ „ํžˆ ์™„๋ฒฝํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ค‘์š”ํ•œ ์ฐธ๊ณ  ๋ฌธํ—Œ์ด๋‚˜ ์ฃผ์ œ๋ฅผ ๋†“์น  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋น„ ์—„๊ฒฉํ•œ ํ‘œํ˜„์ด๋‚˜ ํ† ๋ก ์ด ์žˆ์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋ฅผ ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธํ•˜๊ณ  ์ตœ๋Œ€ํ•œ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์„ค๋ฌธ์กฐ์‚ฌ ์ž‘์„ฑ์€ ์ €ํฌ์—๊ฒŒ๋„ LLM์— ๋Œ€ํ•œ ํ•™์Šต ๊ณผ์ •์ž…๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋ฅผ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฑด์„ค์ ์ธ ์ œ์•ˆ์ด ์žˆ๋Š” ๋…์ž๋Š” ์„ค๋ฌธ์กฐ์‚ฌ์˜ GitHub ํŽ˜์ด์ง€์— ๋Œ“๊ธ€์„ ๋‚จ๊ธฐ๊ฑฐ๋‚˜ ์ €์ž์—๊ฒŒ ์ง์ ‘ ์ด๋ฉ”์ผ์„ ๋ณด๋‚ด์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ๋ฐ›์€ ์˜๊ฒฌ์ด๋‚˜ ์ œ์•ˆ์— ๋”ฐ๋ผ ํ–ฅํ›„ ๋ฒ„์ „์—์„œ ์ˆ˜์ •ํ•  ์˜ˆ์ •์ด๋ฉฐ, ์„ค๋ฌธ์กฐ์‚ฌ์— ๊ฑด์„ค์ ์ธ ์ œ์•ˆ์„ ๋ณด๋‚ด์ฃผ์‹  ๋…์ž์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋œป์„ ์ „ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
While insightful, human annotation of faithfulness in book-length summarization is simply not scalable: our annotations cost $40 USD per summary for a total cost of $5.2K USD, which is prohibitively expensive for usage during model development and with bigger corpora. In this section, inspired by methods such as FactScore (Min et al., 2023) and BooookScore (Chang et al., 2023b), we develop LLM-powered automatic raters of faithfulness that operate at the claim level. However, our best method, which relies on prompting Claude-3-Opus with the entire book to verify a single claim, is expensive and unreliable at detecting unfaithful claims in Tables, suggesting important directions for future work.
์ฑ… ๊ธธ์ด์˜ ์š”์•ฝ์—์„œ ์‹ ๋ขฐ์„ฑ์„ ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹ค๋Š” ๊ฒƒ์€ ํ†ต์ฐฐ๋ ฅ์€ ์žˆ์ง€๋งŒ ๋‹จ์ˆœํžˆ ํ™•์žฅ ๊ฐ€๋Šฅํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค: ์šฐ๋ฆฌ์˜ ์ฃผ์„์€ ์š”์•ฝ๋‹น 40๋‹ฌ๋Ÿฌ์˜ ๋น„์šฉ์ด ๋“ค๋ฉฐ ์ด ๋น„์šฉ์€ 5.2์ฒœ ๋‹ฌ๋Ÿฌ๋กœ, ๋ชจ๋ธ ๊ฐœ๋ฐœ ๋ฐ ๋” ํฐ ๋ง๋ญ‰์น˜ ์‚ฌ์šฉ ์‹œ์—๋Š” ์ง€๋‚˜์น˜๊ฒŒ ๋น„์Œ‰๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” FactScore (Min et al., 2023) ๋ฐ BooookScore (Chang et al., 2023b)์™€ ๊ฐ™์€ ๋ฐฉ๋ฒ•์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„, ์ฃผ์žฅ ์ˆ˜์ค€์—์„œ ์ž‘๋™ํ•˜๋Š” LLM ๊ธฐ๋ฐ˜์˜ ์ž๋™ ์‹ ๋ขฐ์„ฑ ํ‰๊ฐ€๊ธฐ๋ฅผ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹จ์ผ ์ฃผ์žฅ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด ์ „์ฒด ์ฑ…์„ Claude-3-Opus์— ํ”„๋กฌํ”„ํŠธํ•˜๋Š” ๊ฒƒ์— ์˜์กดํ•˜๋Š” ์šฐ๋ฆฌ์˜ ์ตœ์„ ์˜ ๋ฐฉ๋ฒ•์€ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ณ  ํ‘œ์—์„œ ์‹ ๋ขฐํ•  ์ˆ˜ ์—†๋Š” ์ฃผ์žฅ์„ ๊ฐ์ง€ํ•˜๋Š” ๋ฐ ์‹ ๋ขฐํ•  ์ˆ˜ ์—†์œผ๋ฉฐ, ์ด๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ์˜ ์ค‘์š”ํ•œ ๋ฐฉํ–ฅ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•˜์ง€๋งŒ, ์ฑ… ๊ธธ์ด์˜ ์š”์•ฝ์—์„œ ์ถฉ์‹ค๋„์— ๋Œ€ํ•œ ์ธ๊ฐ„์˜ ์ฃผ์„์€ ํ™•์žฅํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์š”์•ฝ๋‹น 40๋‹ฌ๋Ÿฌ์˜ ๋น„์šฉ์ด ๋“ค๋ฉฐ ์ด ๋น„์šฉ์€ 5,200๋‹ฌ๋Ÿฌ๋กœ, ๋ชจ๋ธ ๊ฐœ๋ฐœ ๋ฐ ๋” ํฐ ๋ง๋ญ‰์น˜์—์„œ ์‚ฌ์šฉํ•˜๊ธฐ์—๋Š” ๋„ˆ๋ฌด ๋น„์Œ‰๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” FactScore(Min et al., 2023) ๋ฐ BooookScore(Chang et al., 2023b)์™€ ๊ฐ™์€ ๋ฐฉ๋ฒ•์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ํด๋ ˆ์ž„ ์ˆ˜์ค€์—์„œ ์ž‘๋™ํ•˜๋Š” ์ถฉ์‹ค๋„์— ๋Œ€ํ•œ LLM ๊ธฐ๋ฐ˜ ์ž๋™ ํ‰๊ฐ€์ž๋ฅผ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹จ์ผ ํด๋ ˆ์ž„์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด ์ฑ… ์ „์ฒด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Claude-3-Opus๋ฅผ ํ”„๋กฌํ”„ํŠธํ•˜๋Š” ๋ฐฉ์‹์— ์˜์กดํ•˜๋Š” ์ตœ์ƒ์˜ ๋ฐฉ๋ฒ•์€ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์—†์–ด ํ‘œ์˜ ๋ถˆ์ถฉ์‹คํ•œ ํด๋ ˆ์ž„์„ ๊ฐ์ง€ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ๋ฐฉํ–ฅ์„ฑ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
On LLaMA-2-7B, our approach exceeds the prior best by absolutely +14.2 on GSM8K (compared to MuggleMath-7B (Li et al., 2023)), and by +20.8 on MATH (compared to MetaMath-7B (Yu et al., 2023)), respectively. It even surpasses several latest 70B models dedicated for math capabilities, such as WizardMath-70B (Luo et al., 2023) (82.6 versus 81.6 on GSM8K). On LLaMA-2-13B, the improvements are +14.1 on GSM8K (compared to MuggleMath-13B (Li et al., 2023)) and +22.5 on MATH (compared to MetaMath-13B (Yu et al., 2023)), respectively. On LLaMA-2-70B, the gains are +7.1 on GSM8K (compared to LEMA-LLaMA-2-70B (An et al., 2023)) and +26.2 on MATH (compared to MetaMath-70B (Yu et al., 2023)), respectively.
LLaMA-2-7B์—์„œ ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ GSM8K์—์„œ ์ด์ „ ์ตœ๊ณ  ์„ฑ๊ณผ๋ฅผ ์ ˆ๋Œ€์ ์œผ๋กœ +14.2 (MuggleMath-7B (Li et al., 2023)์™€ ๋น„๊ต) ๋ฐ MATH์—์„œ +20.8 (MetaMath-7B (Yu et al., 2023)์™€ ๋น„๊ต)๋กœ ์ดˆ๊ณผํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” WizardMath-70B (Luo et al., 2023)์™€ ๊ฐ™์€ ์ˆ˜ํ•™ ๋Šฅ๋ ฅ์— ํŠนํ™”๋œ ์ตœ์‹  70B ๋ชจ๋ธ ๋ช‡ ๊ฐ€์ง€๋ฅผ ๋Šฅ๊ฐ€ํ•˜๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค (GSM8K์—์„œ 82.6 ๋Œ€ 81.6). LLaMA-2-13B์—์„œ๋Š” GSM8K์—์„œ +14.1 (MuggleMath-13B (Li et al., 2023)์™€ ๋น„๊ต) ๋ฐ MATH์—์„œ +22.5 (MetaMath-13B (Yu et al., 2023)์™€ ๋น„๊ต)๋กœ ๊ฐœ์„ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. LLaMA-2-70B์—์„œ๋Š” GSM8K์—์„œ +7.1 (LEMA-LLaMA-2-70B (An et al., 2023)์™€ ๋น„๊ต) ๋ฐ MATH์—์„œ +26.2 (MetaMath-70B (Yu et al., 2023)์™€ ๋น„๊ต)๋กœ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
LLaMA-2-7B์—์„œ ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ GSM8K์—์„œ ์ด์ „ ์ตœ๊ณ  ๊ธฐ๋ก์„ +14.2(MuggleMath-7B(Li et al., 2023) ๋Œ€๋น„)๋กœ, MATH์—์„œ +20.8(MetaMath-7B(Yu et al., 2023) ๋Œ€๋น„)๋กœ ๊ฐ๊ฐ ์ดˆ๊ณผํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹ฌ์ง€์–ด GSM8K์—์„œ 82.6 ๋Œ€ 81.6์œผ๋กœ WizardMath-70B(Luo et al., 2023)์™€ ๊ฐ™์€ ์ˆ˜ํ•™ ๊ธฐ๋Šฅ ์ „์šฉ ์ตœ์‹  70B ๋ชจ๋ธ๋„ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. LLaMA-2-13B์˜ ๊ฒฝ์šฐ GSM8K์—์„œ +14.1(MuggleMath-13B(Li et al., 2023) ๋Œ€๋น„), MATH์—์„œ +22.5(MetaMath-13B(Yu et al., 2023) ๋Œ€๋น„)๋กœ ๊ฐ๊ฐ ๊ฐœ์„ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. LLaMA-2-70B์˜ ๊ฒฝ์šฐ GSM8K์—์„œ +7.1(LEMA-LLaMA-2-70B(An et al., 2023) ๋Œ€๋น„), MATH์—์„œ +26.2(MetaMath-70B(Yu et al., 2023) ๋Œ€๋น„)๋กœ ๊ฐ๊ฐ ๊ฐœ์„ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
In this study, we strictly adhered to ethical standards in the conduct of our research. Our experiments were based entirely on open models and open datasets, ensuring transparency and accessibility. We took meticulous care to avoid any biases or data contamination, thereby maintaining the integrity of our research process. The experimental environment was rigorously designed to be objective, ensuring that all comparisons conducted were fair and impartial. This approach reinforces the reliability and validity of our findings, contributing positively to the field while upholding the highest ethical standards. We confirmed that all the data used in our experiments were free of licensing issues.
๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์—ฐ๊ตฌ ์ˆ˜ํ–‰์— ์žˆ์–ด ์—„๊ฒฉํ•œ ์œค๋ฆฌ ๊ธฐ์ค€์„ ์ค€์ˆ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์‹คํ—˜์€ ์ „์ ์œผ๋กœ ๊ณต๊ฐœ๋œ ๋ชจ๋ธ๊ณผ ๊ณต๊ฐœ๋œ ๋ฐ์ดํ„ฐ์…‹์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜์—ฌ ํˆฌ๋ช…์„ฑ๊ณผ ์ ‘๊ทผ์„ฑ์„ ๋ณด์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ํŽธํ–ฅ์ด๋‚˜ ๋ฐ์ดํ„ฐ ์˜ค์—ผ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ์„ธ์‹ฌํ•œ ์ฃผ์˜๋ฅผ ๊ธฐ์šธ์˜€์œผ๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ์—ฐ๊ตฌ ๊ณผ์ •์˜ ๋ฌด๊ฒฐ์„ฑ์„ ์œ ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ํ™˜๊ฒฝ์€ ๊ฐ๊ด€์ ์œผ๋กœ ์„ค๊ณ„๋˜์–ด ๋ชจ๋“  ๋น„๊ต๊ฐ€ ๊ณต์ •ํ•˜๊ณ  ํŽธํŒŒ์ ์ด์ง€ ์•Š๋„๋ก ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์˜ ์‹ ๋ขฐ์„ฑ๊ณผ ํƒ€๋‹น์„ฑ์„ ๊ฐ•ํ™”ํ•˜๋ฉฐ, ์ตœ๊ณ  ์ˆ˜์ค€์˜ ์œค๋ฆฌ ๊ธฐ์ค€์„ ์ค€์ˆ˜ํ•˜๋ฉด์„œ ํ•ด๋‹น ๋ถ„์•ผ์— ๊ธ์ •์ ์œผ๋กœ ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋ชจ๋“  ๋ฐ์ดํ„ฐ๊ฐ€ ๋ผ์ด์„ ์Šค ๋ฌธ์ œ์—์„œ ์ž์œ ๋กœ์›€์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค.
์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์—ฐ๊ตฌ ์ˆ˜ํ–‰ ๊ณผ์ •์—์„œ ์œค๋ฆฌ์  ๊ธฐ์ค€์„ ์—„๊ฒฉํžˆ ์ค€์ˆ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ์˜ ์‹คํ—˜์€ ์ „์ ์œผ๋กœ ์˜คํ”ˆ ๋ชจ๋ธ๊ณผ ์˜คํ”ˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜์—ฌ ํˆฌ๋ช…์„ฑ๊ณผ ์ ‘๊ทผ์„ฑ์„ ๋ณด์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ํŽธ๊ฒฌ์ด๋‚˜ ๋ฐ์ดํ„ฐ ์˜ค์—ผ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ์„ธ์‹ฌํ•œ ์ฃผ์˜๋ฅผ ๊ธฐ์šธ์—ฌ ์—ฐ๊ตฌ ๊ณผ์ •์˜ ๋ฌด๊ฒฐ์„ฑ์„ ์œ ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ํ™˜๊ฒฝ์„ ๊ฐ๊ด€์ ์œผ๋กœ ์„ค๊ณ„ํ•˜์—ฌ ๋ชจ๋“  ๋น„๊ต๊ฐ€ ๊ณต์ •ํ•˜๊ณ  ๊ณตํ‰ํ•˜๊ฒŒ ์ด๋ฃจ์–ด์งˆ ์ˆ˜ ์žˆ๋„๋ก ์—„๊ฒฉํ•˜๊ฒŒ ๊ด€๋ฆฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์˜ ์‹ ๋ขฐ์„ฑ๊ณผ ํƒ€๋‹น์„ฑ์„ ๊ฐ•ํ™”ํ•˜์—ฌ ์ตœ๊ณ ์˜ ์œค๋ฆฌ์  ๊ธฐ์ค€์„ ์ง€ํ‚ค๋ฉด์„œ ํ•ด๋‹น ๋ถ„์•ผ์— ๊ธ์ •์ ์œผ๋กœ ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค. ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋ชจ๋“  ๋ฐ์ดํ„ฐ์— ๋ผ์ด์„ ์Šค ๋ฌธ์ œ๊ฐ€ ์—†๋Š”์ง€ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค.
**Typical Conditions for Selecting a Single vs Multi-Agent Architecture.** Based on the aforementioned agent patterns, we find that single-agent patterns are generally best suited for tasks with a narrowly defined list of tools and where processes are well-defined. Single agents are also typically easier to implement since only one agent and set of tools needs to be defined. Additionally, single agent architectures do not face limitations like poor feedback from other agents or distracting and unrelated chatter from other team members. However, they may get stuck in an execution loop and fail to make progress towards their goal if their reasoning and refinement capabilities are not robust.
**๋‹จ์ผ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜์™€ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์„ ํƒํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ์กฐ๊ฑด.** ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์—์ด์ „ํŠธ ํŒจํ„ด์„ ๊ธฐ๋ฐ˜์œผ๋กœ, ๋‹จ์ผ ์—์ด์ „ํŠธ ํŒจํ„ด์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋„๊ตฌ ๋ชฉ๋ก์ด ์ข๊ฒŒ ์ •์˜๋˜๊ณ  ํ”„๋กœ์„ธ์Šค๊ฐ€ ์ž˜ ์ •์˜๋œ ์ž‘์—…์— ๊ฐ€์žฅ ์ ํ•ฉํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹จ์ผ ์—์ด์ „ํŠธ๋Š” ํ•˜๋‚˜์˜ ์—์ด์ „ํŠธ์™€ ๋„๊ตฌ ์„ธํŠธ๋งŒ ์ •์˜ํ•˜๋ฉด ๋˜๊ธฐ ๋•Œ๋ฌธ์— ๊ตฌํ˜„์ด ๋” ์‰ฝ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋‹จ์ผ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๋Š” ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๋กœ๋ถ€ํ„ฐ์˜ ๋ถ€์ •ํ™•ํ•œ ํ”ผ๋“œ๋ฐฑ์ด๋‚˜ ํŒ€์›๋“ค์˜ ์‚ฐ๋งŒํ•˜๊ณ  ๊ด€๋ จ ์—†๋Š” ๋Œ€ํ™”์™€ ๊ฐ™์€ ์ œํ•œ์— ์ง๋ฉดํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋“ค์˜ ์ถ”๋ก  ๋ฐ ์ •์ œ ๋Šฅ๋ ฅ์ด ๊ฒฌ๊ณ ํ•˜์ง€ ์•Š๋‹ค๋ฉด ์‹คํ–‰ ๋ฃจํ”„์— ๊ฐ‡ํ˜€ ๋ชฉํ‘œ๋ฅผ ํ–ฅํ•œ ์ง„์ „์„ ์ด๋ฃจ์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**๋‹จ์ผ ์—์ด์ „ํŠธ ๋Œ€ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์„ ํƒํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ์กฐ๊ฑด.** ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์—์ด์ „ํŠธ ํŒจํ„ด์— ๋”ฐ๋ฅด๋ฉด, ๋‹จ์ผ ์—์ด์ „ํŠธ ํŒจํ„ด์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋„๊ตฌ ๋ชฉ๋ก์ด ์ข๊ฒŒ ์ •์˜๋˜๊ณ  ํ”„๋กœ์„ธ์Šค๊ฐ€ ์ž˜ ์ •์˜๋œ ์ž‘์—…์— ๊ฐ€์žฅ ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹จ์ผ ์—์ด์ „ํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ํ•˜๋‚˜์˜ ์—์ด์ „ํŠธ์™€ ๋„๊ตฌ ์„ธํŠธ๋งŒ ์ •์˜ํ•˜๋ฉด ๋˜๋ฏ€๋กœ ๊ตฌํ˜„ํ•˜๊ธฐ๊ฐ€ ๋” ์‰ฝ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹จ์ผ ์—์ด์ „ํŠธ ์•„ํ‚คํ…์ฒ˜๋Š” ๋‹ค๋ฅธ ์—์ด์ „ํŠธ์˜ ํ”ผ๋“œ๋ฐฑ์ด ์ข‹์ง€ ์•Š๊ฑฐ๋‚˜ ๋‹ค๋ฅธ ํŒ€์›์˜ ์‚ฐ๋งŒํ•˜๊ณ  ๊ด€๋ จ ์—†๋Š” ์ฑ„ํŒ…๊ณผ ๊ฐ™์€ ์ œํ•œ์— ์ง๋ฉดํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ถ”๋ก  ๋ฐ ๊ตฌ์ฒดํ™” ๊ธฐ๋Šฅ์ด ๊ฒฌ๊ณ ํ•˜์ง€ ์•Š์€ ๊ฒฝ์šฐ ์‹คํ–‰ ๋ฃจํ”„์— ๊ฐ‡ํ˜€ ๋ชฉํ‘œ๋ฅผ ํ–ฅํ•ด ์ง„์ „์„ ์ด๋ฃจ์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The OSCE is typically divided into multiple stations (often 8-12), each simulating a real-life clinical scenario enacted by standardized patient actors trained to portray specific symptoms or conditions based on pre-defined scenario descriptions. At each station, students are given specific tasks to perform, such as taking a clinical history, or making a diagnosis. Each station has a set time limit, ensuring fairness and efficient assessment. Trained examiners observe students' performance at each station using a pre-defined checklist or marking scheme. They assess clinical skills like communication, history-taking, physical examination techniques, clinical reasoning, and decision-making.
OSCE๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์Šคํ…Œ์ด์…˜(๋ณดํ†ต 8-12๊ฐœ)์œผ๋กœ ๋‚˜๋‰˜๋ฉฐ, ๊ฐ ์Šคํ…Œ์ด์…˜์€ ์‚ฌ์ „์— ์ •์˜๋œ ์‹œ๋‚˜๋ฆฌ์˜ค ์„ค๋ช…์„ ๋ฐ”ํƒ•์œผ๋กœ ํŠน์ • ์ฆ์ƒ์ด๋‚˜ ์ƒํƒœ๋ฅผ ๋ฌ˜์‚ฌํ•˜๋„๋ก ํ›ˆ๋ จ๋œ ํ‘œ์ค€ํ™”๋œ ํ™˜์ž ๋ฐฐ์šฐ๋“ค์ด ์—ฐ๊ธฐํ•˜๋Š” ์‹ค์ œ ์ž„์ƒ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์Šคํ…Œ์ด์…˜์—์„œ ํ•™์ƒ๋“ค์€ ์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ๋‚˜ ์ง„๋‹จ๊ณผ ๊ฐ™์€ ํŠน์ • ๊ณผ์ œ๋ฅผ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์Šคํ…Œ์ด์…˜์—๋Š” ๊ณต์ •์„ฑ๊ณผ ํšจ์œจ์ ์ธ ํ‰๊ฐ€๋ฅผ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์ •ํ•ด์ง„ ์‹œ๊ฐ„ ์ œํ•œ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ๋œ ์‹œํ—˜๊ด€๋“ค์€ ์‚ฌ์ „์— ์ •์˜๋œ ์ฒดํฌ๋ฆฌ์ŠคํŠธ๋‚˜ ์ฑ„์  ๊ธฐ์ค€์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์Šคํ…Œ์ด์…˜์—์„œ ํ•™์ƒ๋“ค์˜ ์ˆ˜ํ–‰์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ์˜์‚ฌ์†Œํ†ต, ๋ณ‘๋ ฅ ์ฒญ์ทจ, ์‹ ์ฒด ๊ฒ€์‚ฌ ๊ธฐ์ˆ , ์ž„์ƒ ์ถ”๋ก  ๋ฐ ์˜์‚ฌ ๊ฒฐ์ •๊ณผ ๊ฐ™์€ ์ž„์ƒ ๊ธฐ์ˆ ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
OSCE๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์—ฌ๋Ÿฌ ์Šคํ…Œ์ด์…˜(๋ณดํ†ต 8~12๊ฐœ)์œผ๋กœ ๋‚˜๋‰˜๋ฉฐ, ๊ฐ ์Šคํ…Œ์ด์…˜์€ ์‚ฌ์ „ ์ •์˜๋œ ์‹œ๋‚˜๋ฆฌ์˜ค ์„ค๋ช…์— ๋”ฐ๋ผ ํŠน์ • ์ฆ์ƒ์ด๋‚˜ ์ƒํƒœ๋ฅผ ๋ฌ˜์‚ฌํ•˜๋„๋ก ํ›ˆ๋ จ๋œ ํ‘œ์ค€ํ™” ํ™˜์ž ๋ฐฐ์šฐ๊ฐ€ ํ–‰ํ•˜๋Š” ์‹ค์ œ ์ž„์ƒ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์Šคํ…Œ์ด์…˜์—์„œ ํ•™์ƒ๋“ค์€ ์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ, ์ง„๋‹จ ๋“ฑ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•  ํŠน์ • ๊ณผ์ œ๋ฅผ ๋ถ€์—ฌ๋ฐ›์Šต๋‹ˆ๋‹ค. ๊ฐ ์Šคํ…Œ์ด์…˜์—๋Š” ์ •ํ•ด์ง„ ์‹œ๊ฐ„ ์ œํ•œ์ด ์žˆ์–ด ๊ณต์ •ํ•˜๊ณ  ํšจ์œจ์ ์ธ ํ‰๊ฐ€๊ฐ€ ์ด๋ฃจ์–ด์ง‘๋‹ˆ๋‹ค. ํ›ˆ๋ จ๋œ ์‹œํ—˜๊ด€๋“ค์€ ์‚ฌ์ „ ์ •์˜๋œ ์ฒดํฌ๋ฆฌ์ŠคํŠธ ๋˜๋Š” ์ฑ„์  ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์Šคํ…Œ์ด์…˜์—์„œ ํ•™์ƒ๋“ค์˜ ์ˆ˜ํ–‰์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ์˜์‚ฌ์†Œํ†ต, ๋ณ‘๋ ฅ ์ฒญ์ทจ, ์‹ ์ฒด ๊ฒ€์‚ฌ ๊ธฐ์ˆ , ์ž„์ƒ์  ์ถ”๋ก  ๋ฐ ์˜์‚ฌ ๊ฒฐ์ •๊ณผ ๊ฐ™์€ ์ž„์ƒ ๊ธฐ์ˆ ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
Note that while \(\ell(y|x)\) has the same form as the pre-training loss \(L\), they are not equal. First, the pre-training loss is an average of all the tokens in all the documents pre-trained on. According to our empirical observation, the losses of different documents are not uniform. Second, if \(x\) and similar documents do not exist in the pre-training corpus, \(\ell(y|x)\) is the generalization loss, which is often related to other factors beyond the training loss, such as the model size. For example, in computer vision, a highly over-parameterized models often improve over an under-parameterized models in test performance when both models converge on the training data [9; 4].
\(\ell(y|x)\)๊ฐ€ ์‚ฌ์ „ ํ›ˆ๋ จ ์†์‹ค \(L\)๊ณผ ๋™์ผํ•œ ํ˜•ํƒœ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์ง€๋งŒ, ์ด ๋‘˜์€ ๋™์ผํ•˜์ง€ ์•Š๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์‹ญ์‹œ์˜ค. ์ฒซ์งธ, ์‚ฌ์ „ ํ›ˆ๋ จ ์†์‹ค์€ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ ๋ชจ๋“  ๋ฌธ์„œ์˜ ๋ชจ๋“  ํ† ํฐ์˜ ํ‰๊ท ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๊ฒฝํ—˜์  ๊ด€์ฐฐ์— ๋”ฐ๋ฅด๋ฉด, ์„œ๋กœ ๋‹ค๋ฅธ ๋ฌธ์„œ์˜ ์†์‹ค์€ ๊ท ์ผํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋‘˜์งธ, ์‚ฌ์ „ ํ›ˆ๋ จ ์ฝ”ํผ์Šค์— \(x\)์™€ ์œ ์‚ฌํ•œ ๋ฌธ์„œ๊ฐ€ ์กด์žฌํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ, \(\ell(y|x)\)๋Š” ์ผ๋ฐ˜ํ™” ์†์‹ค์ด ๋˜๋ฉฐ, ์ด๋Š” ์ข…์ข… ๋ชจ๋ธ ํฌ๊ธฐ์™€ ๊ฐ™์€ ํ›ˆ๋ จ ์†์‹ค ์ด์™ธ์˜ ๋‹ค๋ฅธ ์š”์ธ๊ณผ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ปดํ“จํ„ฐ ๋น„์ „์—์„œ, ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์—์„œ ๋‘ ๋ชจ๋ธ์ด ๋ชจ๋‘ ์ˆ˜๋ ดํ•  ๋•Œ, ๊ณผ๋„ํ•˜๊ฒŒ ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ ๋งŽ์€ ๋ชจ๋ธ์ด ๊ณผ์†Œ ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ ๋œ ๋ชจ๋ธ๋ณด๋‹ค ํ…Œ์ŠคํŠธ ์„ฑ๋Šฅ์—์„œ ์ข…์ข… ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค [9; 4].
ell(y|x)\)๋Š” ์‚ฌ์ „ ํ•™์Šต ์†์‹ค \(L\)๊ณผ ๋™์ผํ•œ ํ˜•ํƒœ๋ฅผ ๊ฐ€์ง€์ง€๋งŒ, ๋‘˜์€ ๋™์ผํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์€ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋“  ๋ฌธ์„œ์˜ ๋ชจ๋“  ํ† ํฐ์˜ ํ‰๊ท ์ž…๋‹ˆ๋‹ค. ๊ฒฝํ—˜์  ๊ด€์ฐฐ์— ๋”ฐ๋ฅด๋ฉด ๋ฌธ์„œ๋งˆ๋‹ค ์†์‹ค์ด ๊ท ์ผํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋‘˜์งธ, ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์— \(x\) ๋ฐ ์œ ์‚ฌํ•œ ๋ฌธ์„œ๊ฐ€ ์กด์žฌํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ, \(\ell(y|x)\)๋Š” ์ผ๋ฐ˜ํ™” ์†์‹ค์ด๋ฉฐ, ์ด๋Š” ์ข…์ข… ๋ชจ๋ธ ํฌ๊ธฐ์™€ ๊ฐ™์€ ํ•™์Šต ์†์‹ค ์ด์™ธ์˜ ๋‹ค๋ฅธ ์š”์ธ๊ณผ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ปดํ“จํ„ฐ ๋น„์ „์—์„œ๋Š” ๋‘ ๋ชจ๋ธ์ด ํ•™์Šต ๋ฐ์ดํ„ฐ์— ์ˆ˜๋ ดํ•  ๋•Œ ๊ณ ๋„๋กœ ๊ณผ๋Œ€ํ™”๋œ ๋ชจ๋ธ์ด ๊ณผ์†Œ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋œ ๋ชจ๋ธ๋ณด๋‹ค ํ…Œ์ŠคํŠธ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค[9; 4].
Latent Multi-Hop Reasoning PathwayWhile we study one pathway for latent multi-hop reasoning (e.g., we test the use of the second hop by means of entity recall), considering the potential redundancy of inference pathways in LLMs (McGrath et al., 2023), other pathways might exist; the same information might be retrieved in different ways. Also, we don't measure multi-hop reasoning end-to-end and track only the changes that occur in the first and the second hop with respect to a single layer, while the effect of the first hop of reasoning could possibly propagate to other layers. Hence, the effects we see might be a lower bound on the model's ability to perform latent two-hop reasoning.
์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก  ๊ฒฝ๋กœ: ์šฐ๋ฆฌ๋Š” ์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์œ„ํ•œ ํ•˜๋‚˜์˜ ๊ฒฝ๋กœ(์˜ˆ: ์—”ํ‹ฐํ‹ฐ ํšŒ์ƒ์„ ํ†ตํ•œ ๋‘ ๋ฒˆ์งธ ํ™‰์˜ ์‚ฌ์šฉ์„ ํ…Œ์ŠคํŠธ)๋ฅผ ์—ฐ๊ตฌํ•˜์ง€๋งŒ, LLMs์—์„œ ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ์ž ์žฌ์  ์ค‘๋ณต์„ฑ์„ ๊ณ ๋ คํ•  ๋•Œ(McGrath et al., 2023), ๋‹ค๋ฅธ ๊ฒฝ๋กœ๊ฐ€ ์กด์žฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋™์ผํ•œ ์ •๋ณด๊ฐ€ ๋‹ค๋ฅธ ๋ฐฉ์‹์œผ๋กœ ๊ฒ€์ƒ‰๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์šฐ๋ฆฌ๋Š” ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ๋๊นŒ์ง€ ์ธก์ •ํ•˜์ง€ ์•Š๊ณ  ๋‹จ์ผ ๋ ˆ์ด์–ด์— ๋Œ€ํ•ด ์ฒซ ๋ฒˆ์งธ ํ™‰๊ณผ ๋‘ ๋ฒˆ์งธ ํ™‰์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๋ณ€ํ™”๋งŒ ์ถ”์ ํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ํ™‰์˜ ์ถ”๋ก  ํšจ๊ณผ๊ฐ€ ๋‹ค๋ฅธ ๋ ˆ์ด์–ด๋กœ ์ „ํŒŒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๊ฐ€ ๋ณด๋Š” ํšจ๊ณผ๋Š” ๋ชจ๋ธ์ด ์ž ์žฌ์  ๋‘ ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์˜ ํ•˜ํ•œ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก  ๊ฒฝ๋กœ์ž ์žฌ์  ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์œ„ํ•œ ํ•˜๋‚˜์˜ ๊ฒฝ๋กœ(์˜ˆ: ์—”ํ‹ฐํ‹ฐ ๋ฆฌ์ฝœ์„ ํ†ตํ•ด ๋‘ ๋ฒˆ์งธ ํ™‰์˜ ์‚ฌ์šฉ์„ ํ…Œ์ŠคํŠธ)๋ฅผ ์—ฐ๊ตฌํ•˜๋Š” ๋™์•ˆ, LLM์—์„œ ์ถ”๋ก  ๊ฒฝ๋กœ์˜ ์ž ์žฌ์  ์ค‘๋ณต์„ฑ(McGrath et al., 2023)์„ ๊ณ ๋ คํ•  ๋•Œ ๋‹ค๋ฅธ ๊ฒฝ๋กœ๊ฐ€ ์กด์žฌํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋™์ผํ•œ ์ •๋ณด๋ฅผ ๋‹ค๋ฅธ ๋ฐฉ์‹์œผ๋กœ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์ค‘ ํ™‰ ์ถ”๋ก ์„ ์—”๋“œํˆฌ์—”๋“œ๋กœ ์ธก์ •ํ•˜์ง€ ์•Š๊ณ  ๋‹จ์ผ ๋ ˆ์ด์–ด์— ๋Œ€ํ•ด ์ฒซ ๋ฒˆ์งธ ํ™‰๊ณผ ๋‘ ๋ฒˆ์งธ ํ™‰์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๋ณ€๊ฒฝ ์‚ฌํ•ญ๋งŒ ์ถ”์ ํ•˜๋Š” ๋ฐ˜๋ฉด, ์ถ”๋ก ์˜ ์ฒซ ๋ฒˆ์งธ ํ™‰์˜ ํšจ๊ณผ๊ฐ€ ๋‹ค๋ฅธ ๋ ˆ์ด์–ด๋กœ ์ „ํŒŒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๊ฐ€ ๋ณด๋Š” ํšจ๊ณผ๋Š” ๋ชจ๋ธ์ด ์ž ์žฌ์ ์ธ 2ํ™‰ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์˜ ํ•˜ํ•œ์„ ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Footnote 6: [https://github.com/karpathy/nanoGPT](https://github.com/karpathy/nanoGPT)Mitigation:To our knowledge, rigorous theory (or even empirical demonstrations) on mitigation strategies against model collapse are yet to come, with one notable exception in (Bertrand et al., 2023) (see below). Several works discuss the need for _detection_ of AI-generated images or text (to avoid retraining on them), for example motivating research into watermarking strategies. (Bertrand et al., 2023) analyze iterative retraining on a mixture of synthesized and original data under several technical assumptions and find that there are fixed points governing the stability of iterative retraining.
Footnote 6: [https://github.com/karpathy/nanoGPT](https://github.com/karpathy/nanoGPT) ์™„ํ™”: ์šฐ๋ฆฌ๊ฐ€ ์•„๋Š” ํ•œ, ๋ชจ๋ธ ๋ถ•๊ดด์— ๋Œ€ํ•œ ์™„ํ™” ์ „๋žต์— ๋Œ€ํ•œ ์—„๊ฒฉํ•œ ์ด๋ก (๋˜๋Š” ์‹ฌ์ง€์–ด ์‹ค์ฆ์  ์‹œ์—ฐ)์€ ์•„์ง ๋‚˜์˜ค์ง€ ์•Š์•˜์œผ๋ฉฐ, (Bertrand et al., 2023)์˜ ํ•œ ๊ฐ€์ง€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์˜ˆ์™ธ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค(์•„๋ž˜ ์ฐธ์กฐ). ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ๋Š” AI ์ƒ์„ฑ ์ด๋ฏธ์ง€๋‚˜ ํ…์ŠคํŠธ์˜ _ํƒ์ง€_ ํ•„์š”์„ฑ์„ ๋…ผ์˜ํ•˜๋ฉฐ, ์ด๋ฅผ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ์žฌํ›ˆ๋ จ์„ ํ”ผํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ ์›Œํ„ฐ๋งˆํ‚น ์ „๋žต์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค. (Bertrand et al., 2023)์€ ์—ฌ๋Ÿฌ ๊ธฐ์ˆ ์  ๊ฐ€์ • ํ•˜์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ์›๋ณธ ๋ฐ์ดํ„ฐ๋ฅผ ํ˜ผํ•ฉํ•˜์—ฌ ๋ฐ˜๋ณต์ ์œผ๋กœ ์žฌํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์„ ๋ถ„์„ํ•˜๊ณ , ๋ฐ˜๋ณต์  ์žฌํ›ˆ๋ จ์˜ ์•ˆ์ •์„ฑ์„ ์ง€๋ฐฐํ•˜๋Š” ๊ณ ์ •์ ์ด ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
๊ฐ์ฃผ 6: [https://github.com/karpathy/nanoGPT](https://github.com/karpathy/nanoGPT)์™„ํ™”:์šฐ๋ฆฌ๊ฐ€ ์•„๋Š” ํ•œ, ๋ชจ๋ธ ๋ถ•๊ดด์— ๋Œ€ํ•œ ์™„ํ™” ์ „๋žต์— ๋Œ€ํ•œ ์—„๊ฒฉํ•œ ์ด๋ก (๋˜๋Š” ๊ฒฝํ—˜์  ์ฆ๋ช…)์€ ์•„์ง ๋‚˜์˜ค์ง€ ์•Š์•˜์œผ๋ฉฐ, (Bertrand et al., 2023)์˜ ํ•œ ๊ฐ€์ง€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์˜ˆ์™ธ๋ฅผ ์ œ์™ธํ•˜๊ณ ๋Š” ์•„์ง ๋‚˜์˜ค์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค(์•„๋ž˜ ์ฐธ์กฐ). ๋ช‡๋ช‡ ์—ฐ๊ตฌ์—์„œ๋Š” AI๊ฐ€ ์ƒ์„ฑํ•œ ์ด๋ฏธ์ง€๋‚˜ ํ…์ŠคํŠธ์˜ _ํƒ์ง€(์ด๋ฅผ ์žฌํ•™์Šตํ•˜์ง€ ์•Š๊ธฐ ์œ„ํ•ด)์— ๋Œ€ํ•œ ํ•„์š”์„ฑ์„ ๋…ผ์˜ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์˜ˆ๋ฅผ ๋“ค์–ด ์›Œํ„ฐ๋งˆํ‚น ์ „๋žต ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ๋™๊ธฐ๋ฅผ ๋ถ€์—ฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. (Bertrand et al., 2023)์€ ๋ช‡ ๊ฐ€์ง€ ๊ธฐ์ˆ ์  ๊ฐ€์ • ํ•˜์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ์›๋ณธ ๋ฐ์ดํ„ฐ์˜ ํ˜ผํ•ฉ๋ฌผ์— ๋Œ€ํ•œ ๋ฐ˜๋ณต์ ์ธ ์žฌ๊ต์œก์„ ๋ถ„์„ํ•˜๊ณ  ๋ฐ˜๋ณต์ ์ธ ์žฌ๊ต์œก์˜ ์•ˆ์ •์„ฑ์„ ์ง€๋ฐฐํ•˜๋Š” ๊ณ ์ •์ ์ด ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
**Effects of inference-time customization.** One key benefit of our proposed framework is that it enables us to control how much each critique type affects the final generation sampling. We analyze the effects of different parameter weights on the top of our 7B model during inference time on ASQA, where multiple evaluation aspects are considered. Figure 2(b) shows the effects of changing the weighting term for \(\boxed{Isstr}\), which criticizes how supported the output is by the text passage. As the figure shows, increasing the weight leads to positive effects on the models' citation precision since this puts more emphasis on whether model generation is supported by the evidence.
**์ถ”๋ก  ์‹œ๊ฐ„ ๋งž์ถคํ™”์˜ ํšจ๊ณผ.** ์ œ์•ˆ๋œ ํ”„๋ ˆ์ž„์›Œํฌ์˜ ์ฃผ์š” ์ด์  ์ค‘ ํ•˜๋‚˜๋Š” ๊ฐ ๋น„ํŒ ์œ ํ˜•์ด ์ตœ์ข… ์ƒ์„ฑ ์ƒ˜ํ”Œ๋ง์— ์–ผ๋งˆ๋‚˜ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ์ œ์–ดํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ASQA์—์„œ ์—ฌ๋Ÿฌ ํ‰๊ฐ€ ์ธก๋ฉด์„ ๊ณ ๋ คํ•˜์—ฌ ์ถ”๋ก  ์‹œ๊ฐ„ ๋™์•ˆ 7B ๋ชจ๋ธ์˜ ์ƒ๋‹จ์— ์žˆ๋Š” ๋‹ค์–‘ํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜ ๊ฐ€์ค‘์น˜์˜ ํšจ๊ณผ๋ฅผ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. Figure 2(b)๋Š” ์ถœ๋ ฅ์ด ํ…์ŠคํŠธ ๋‹จ๋ฝ์— ์˜ํ•ด ์–ผ๋งˆ๋‚˜ ์ง€์›๋˜๋Š”์ง€๋ฅผ ๋น„ํŒํ•˜๋Š” \(\boxed{Isstr}\)์˜ ๊ฐ€์ค‘์น˜ ํ•ญ์„ ๋ณ€๊ฒฝํ•œ ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ๊ฐ€์ค‘์น˜๋ฅผ ์ฆ๊ฐ€์‹œํ‚ค๋ฉด ๋ชจ๋ธ ์ƒ์„ฑ์ด ์ฆ๊ฑฐ์— ์˜ํ•ด ์ง€์›๋˜๋Š”์ง€ ์—ฌ๋ถ€์— ๋” ์ค‘์ ์„ ๋‘๊ธฐ ๋•Œ๋ฌธ์— ๋ชจ๋ธ์˜ ์ธ์šฉ ์ •ํ™•๋„์— ๊ธ์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค.
**์ถ”๋ก  ์‹œ๊ฐ„ ์‚ฌ์šฉ์ž ์ง€์ •์˜ ํšจ๊ณผ.** ์ œ์•ˆํ•œ ํ”„๋ ˆ์ž„์›Œํฌ์˜ ์ฃผ์š” ์ด์  ์ค‘ ํ•˜๋‚˜๋Š” ๊ฐ ๋น„ํ‰ ์œ ํ˜•์ด ์ตœ์ข… ์ƒ์„ฑ ์ƒ˜ํ”Œ๋ง์— ์–ผ๋งˆ๋‚˜ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€๋ฅผ ์ œ์–ดํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ํ‰๊ฐ€ ์ธก๋ฉด์„ ๊ณ ๋ คํ•˜๋Š” ASQA์—์„œ ์ถ”๋ก  ์‹œ๊ฐ„ ๋™์•ˆ 7B ๋ชจ๋ธ ์ƒ๋‹จ์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜ ๊ฐ€์ค‘์น˜์˜ ํšจ๊ณผ๋ฅผ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2(b)๋Š” ์ถœ๋ ฅ์ด ํ…์ŠคํŠธ ๊ตฌ์ ˆ์— ์˜ํ•ด ์–ผ๋งˆ๋‚˜ ๋’ท๋ฐ›์นจ๋˜๋Š”์ง€๋ฅผ ๋น„ํŒํ•˜๋Š” \boxed{Isstr}\)์˜ ๊ฐ€์ค‘์น˜ ๋ณ€๊ฒฝ์ด ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ์—์„œ ์•Œ ์ˆ˜ ์žˆ๋“ฏ์ด ๊ฐ€์ค‘์น˜๋ฅผ ๋†’์ด๋ฉด ๋ชจ๋ธ ์ƒ์„ฑ์ด ์ฆ๊ฑฐ์— ์˜ํ•ด ๋’ท๋ฐ›์นจ๋˜๋Š”์ง€ ์—ฌ๋ถ€์— ๋” ์ค‘์ ์„ ๋‘๊ธฐ ๋•Œ๋ฌธ์— ๋ชจ๋ธ์˜ ์ธ์šฉ ์ •ํ™•๋„์— ๊ธ์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค.
\(\bullet\)_Expressing the task goal clearly._ Task descriptions should not be ambiguous or unclear, which likely lead to inaccurate or inappropriate responses. This highlights the need for clear and unambiguous directives when utilizing these models [66]. A clear and detailed description should contain various elements to explain a task, including task objective, input/output data (_e.g._, _"Given a long document, I want you to generate a concise summary."_), and the response constraints (_e.g._, _"the length of the summary cannot exceed 50."_). By providing a well-clarified task description, LLMs can more effectively understand the target task and generate the desired output.
โ€ข _์ž‘์—… ๋ชฉํ‘œ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ํ‘œํ˜„ํ•˜๊ธฐ._ ์ž‘์—… ์„ค๋ช…์€ ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ๋ถˆ๋ช…ํ™•ํ•ด์„œ๋Š” ์•ˆ ๋˜๋ฉฐ, ์ด๋Š” ๋ถ€์ •ํ™•ํ•˜๊ฑฐ๋‚˜ ๋ถ€์ ์ ˆํ•œ ์‘๋‹ต์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•  ๋•Œ ๋ช…ํ™•ํ•˜๊ณ  ๋ชจํ˜ธํ•˜์ง€ ์•Š์€ ์ง€์‹œ๊ฐ€ ํ•„์š”ํ•จ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค [66]. ๋ช…ํ™•ํ•˜๊ณ  ์ƒ์„ธํ•œ ์„ค๋ช…์—๋Š” ์ž‘์—… ๋ชฉํ‘œ, ์ž…๋ ฅ/์ถœ๋ ฅ ๋ฐ์ดํ„ฐ (_์˜ˆ: _"๊ธด ๋ฌธ์„œ๋ฅผ ์ฃผ๋ฉด, ๊ฐ„๊ฒฐํ•œ ์š”์•ฝ์„ ์ƒ์„ฑํ•ด ์ฃผ์„ธ์š”."_), ๊ทธ๋ฆฌ๊ณ  ์‘๋‹ต ์ œ์•ฝ ์กฐ๊ฑด (_์˜ˆ: _"์š”์•ฝ์˜ ๊ธธ์ด๋Š” 50์ž๋ฅผ ์ดˆ๊ณผํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค."_)์„ ํฌํ•จํ•˜์—ฌ ์ž‘์—…์„ ์„ค๋ช…ํ•˜๋Š” ๋‹ค์–‘ํ•œ ์š”์†Œ๊ฐ€ ํฌํ•จ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ž˜ ๋ช…ํ™•ํ™”๋œ ์ž‘์—… ์„ค๋ช…์„ ์ œ๊ณตํ•จ์œผ๋กœ์จ, LLM์€ ๋Œ€์ƒ ์ž‘์—…์„ ๋” ํšจ๊ณผ์ ์œผ๋กœ ์ดํ•ดํ•˜๊ณ  ์›ํ•˜๋Š” ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_์ž‘์—… ๋ชฉํ‘œ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ํ‘œํ˜„ํ•ฉ๋‹ˆ๋‹ค._ ์ž‘์—… ์„ค๋ช…์€ ๋ชจํ˜ธํ•˜๊ฑฐ๋‚˜ ๋ถˆ๋ถ„๋ช…ํ•ด์„œ๋Š” ์•ˆ ๋˜๋ฉฐ, ์ด๋กœ ์ธํ•ด ๋ถ€์ •ํ™•ํ•˜๊ฑฐ๋‚˜ ๋ถ€์ ์ ˆํ•œ ์‘๋‹ต์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ํ™œ์šฉํ•  ๋•Œ ๋ช…ํ™•ํ•˜๊ณ  ๋ชจํ˜ธํ•˜์ง€ ์•Š์€ ์ง€์นจ์ด ํ•„์š”ํ•˜๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค[66]. ๋ช…ํ™•ํ•˜๊ณ  ์ƒ์„ธํ•œ ์„ค๋ช…์—๋Š” ์ž‘์—… ๋ชฉํ‘œ, ์ž…๋ ฅ/์ถœ๋ ฅ ๋ฐ์ดํ„ฐ(_์˜ˆ: "๊ธด ๋ฌธ์„œ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ๊ฐ„๊ฒฐํ•œ ์š”์•ฝ์„ ์ƒ์„ฑํ•ด ์ฃผ์„ธ์š”."_), ์‘๋‹ต ์ œ์•ฝ ์กฐ๊ฑด(_์˜ˆ: "์š”์•ฝ์˜ ๊ธธ์ด๋Š” 50์„ ์ดˆ๊ณผํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค."_) ๋“ฑ ์ž‘์—…์„ ์„ค๋ช…ํ•˜๋Š” ๋‹ค์–‘ํ•œ ์š”์†Œ๊ฐ€ ํฌํ•จ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ช…ํ™•ํ•˜๊ฒŒ ์„ค๋ช…๋œ ์ž‘์—… ์„ค๋ช…์„ ์ œ๊ณตํ•˜๋ฉด LLM์ด ๋ชฉํ‘œ ์ž‘์—…์„ ๋ณด๋‹ค ํšจ๊ณผ์ ์œผ๋กœ ์ดํ•ดํ•˜๊ณ  ์›ํ•˜๋Š” ๊ฒฐ๊ณผ๋ฌผ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure 1 provides a general overview of the framework of RALM methods. The following is a summary of the paper: Section 2 defines RALM. Section 3 provides a detailed classification and summary of the work of retrievers in RALM. Section 4 provides a detailed classification and summary of the work of LMs in RALM. Section 5 provides a classification and summary of specific enhancements to RALM. Section 6 of RALM is a classification and summary of the sources of retrieved data. Section 7 is a summary of RALM applications. Section 8 is a summary of RALM evaluations and benchmarks. Finally, Section 9 is a discussion of the limitations of existing RALM and directions for future work.
Figure 1์€ RALM ๋ฐฉ๋ฒ•์˜ ํ”„๋ ˆ์ž„์›Œํฌ์— ๋Œ€ํ•œ ์ผ๋ฐ˜์ ์ธ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์˜ ์š”์•ฝ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค: 2์žฅ์€ RALM์„ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. 3์žฅ์€ RALM์—์„œ ๊ฒ€์ƒ‰๊ธฐ์˜ ์ž‘์—…์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋ถ„๋ฅ˜์™€ ์š”์•ฝ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. 4์žฅ์€ RALM์—์„œ LMs์˜ ์ž‘์—…์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋ถ„๋ฅ˜์™€ ์š”์•ฝ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. 5์žฅ์€ RALM์— ๋Œ€ํ•œ ํŠน์ • ํ–ฅ์ƒ์ ์˜ ๋ถ„๋ฅ˜์™€ ์š”์•ฝ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. 6์žฅ์€ ๊ฒ€์ƒ‰๋œ ๋ฐ์ดํ„ฐ์˜ ์ถœ์ฒ˜์— ๋Œ€ํ•œ ๋ถ„๋ฅ˜์™€ ์š”์•ฝ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. 7์žฅ์€ RALM ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์˜ ์š”์•ฝ์ž…๋‹ˆ๋‹ค. 8์žฅ์€ RALM ํ‰๊ฐ€ ๋ฐ ๋ฒค์น˜๋งˆํฌ์˜ ์š”์•ฝ์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, 9์žฅ์€ ๊ธฐ์กด RALM์˜ ํ•œ๊ณ„์™€ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์— ๋Œ€ํ•œ ๋…ผ์˜์ž…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1์€ RALM ๋ฐฉ๋ฒ•์˜ ํ”„๋ ˆ์ž„์›Œํฌ์— ๋Œ€ํ•œ ์ผ๋ฐ˜์ ์ธ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์€ ๋…ผ๋ฌธ์˜ ์š”์•ฝ์ž…๋‹ˆ๋‹ค: ์„น์…˜ 2์—์„œ๋Š” RALM์„ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 3์—์„œ๋Š” RALM์—์„œ ๊ฒ€์ƒ‰๊ธฐ์˜ ์ž‘์—…์„ ์ž์„ธํžˆ ๋ถ„๋ฅ˜ํ•˜๊ณ  ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 4์—์„œ๋Š” RALM์—์„œ LM์˜ ์ž‘์—…์„ ์ž์„ธํžˆ ๋ถ„๋ฅ˜ํ•˜๊ณ  ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 5์—์„œ๋Š” RALM์— ๋Œ€ํ•œ ๊ตฌ์ฒด์ ์ธ ๊ฐœ์„  ์‚ฌํ•ญ์„ ๋ถ„๋ฅ˜ํ•˜๊ณ  ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 6์˜ RALM์€ ๊ฒ€์ƒ‰๋œ ๋ฐ์ดํ„ฐ์˜ ์ถœ์ฒ˜๋ฅผ ๋ถ„๋ฅ˜ํ•˜๊ณ  ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค. ์„น์…˜ 7์€ RALM ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์˜ ์š”์•ฝ์ž…๋‹ˆ๋‹ค. ์„น์…˜ 8์€ RALM ํ‰๊ฐ€ ๋ฐ ๋ฒค์น˜๋งˆํฌ์˜ ์š”์•ฝ์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ์„น์…˜ 9๋Š” ๊ธฐ์กด RALM์˜ ํ•œ๊ณ„์™€ ํ–ฅํ›„ ์ž‘์—… ๋ฐฉํ–ฅ์— ๋Œ€ํ•œ ๋…ผ์˜์ž…๋‹ˆ๋‹ค.
Training Process and Comparative AnalysisThe training progress reveals a consistent trend of improvement across various datasets, with particular strides seen in language understanding, reasoning, and domain-specific knowledge. Notably, datasets such as HellaSwag, PIQA, and ARC show marked improvements, indicative of enhanced reasoning capabilities. The model shows notable progress in specialized fields such as mathematics (GSM8K and TheoremQA) and science (ARC-c and ARC-e), emphasizing its increasing ability to understand and produce content specific to these domains. The evaluation results of the intermediate checkpoints during our pre-training process are shown in Table.4.
ํ›ˆ๋ จ ๊ณผ์ • ๋ฐ ๋น„๊ต ๋ถ„์„ ํ›ˆ๋ จ ์ง„ํ–‰ ์ƒํ™ฉ์€ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ์…‹์—์„œ ์ผ๊ด€๋œ ๊ฐœ์„  ์ถ”์„ธ๋ฅผ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ํŠนํžˆ ์–ธ์–ด ์ดํ•ด, ์ถ”๋ก  ๋ฐ ๋„๋ฉ”์ธ ํŠนํ™” ์ง€์‹์—์„œ ๋‘๋“œ๋Ÿฌ์ง„ ๋ฐœ์ „์„ ๋ณด์ž…๋‹ˆ๋‹ค. ํŠนํžˆ HellaSwag, PIQA, ARC์™€ ๊ฐ™์€ ๋ฐ์ดํ„ฐ์…‹์—์„œ ํ˜„์ €ํ•œ ๊ฐœ์„ ์ด ๋‚˜ํƒ€๋‚˜, ํ–ฅ์ƒ๋œ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์€ ์ˆ˜ํ•™(GSM8K ๋ฐ TheoremQA)๊ณผ ๊ณผํ•™(ARC-c ๋ฐ ARC-e)๊ณผ ๊ฐ™์€ ์ „๋ฌธ ๋ถ„์•ผ์—์„œ ๋ˆˆ์— ๋„๋Š” ์ง„์ „์„ ๋ณด์ด๋ฉฐ, ์ด๋Ÿฌํ•œ ๋„๋ฉ”์ธ์— ํŠนํ™”๋œ ์ฝ˜ํ…์ธ ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์ด ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ์Œ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ›ˆ๋ จ ๊ณผ์ • ์ค‘ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์˜ ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋Š” Table.4์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ›ˆ๋ จ ๊ณผ์ • ๋ฐ ๋น„๊ต ๋ถ„์„ํ›ˆ๋ จ ์ง„ํ–‰ ์ƒํ™ฉ์€ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ผ๊ด€๋œ ๊ฐœ์„  ์ถ”์„ธ๋ฅผ ๋ณด์—ฌ์ฃผ๋ฉฐ, ํŠนํžˆ ์–ธ์–ด ์ดํ•ด, ์ถ”๋ก  ๋ฐ ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์—์„œ ํฐ ๋ฐœ์ „์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํŠนํžˆ HellaSwag, PIQA, ARC์™€ ๊ฐ™์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋‘๋“œ๋Ÿฌ์ง„ ๊ฐœ์„ ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ ์ถ”๋ก  ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋˜์—ˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์ˆ˜ํ•™(GSM8K ๋ฐ TheoremQA)๊ณผ ๊ณผํ•™(ARC-c ๋ฐ ARC-e)๊ณผ ๊ฐ™์€ ์ „๋ฌธ ๋ถ„์•ผ์—์„œ ๋‘๋“œ๋Ÿฌ์ง„ ๋ฐœ์ „์„ ๋ณด์—ฌ ์ด๋Ÿฌํ•œ ์˜์—ญ์— ํŠนํ™”๋œ ์ฝ˜ํ…์ธ ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋˜๊ณ  ์žˆ์Œ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต ๊ณผ์ • ์ค‘ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์˜ ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋Š” ํ‘œ 4์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
Furthermore, researchers have found that "As LLMs progress at a rapid pace, existing datasets usually fail to match the models' ever-evolving capabilities, because the complexity level of existing benchmarks is usually static and fixed" [37]. To address this, work has been done to create dynamic benchmarks that are resistant to simple memorization [38; 37]. Researchers have also explored the idea of generating an entirely synthetic benchmark based on a user's specific environment or use case [14; 27]. While these techniques can help with contamination, decreasing the level of human involvement can pose additional risks regarding correctness and the ability to solve problems.
๊ฒŒ๋‹ค๊ฐ€ ์—ฐ๊ตฌ์ž๋“ค์€ "LLMs๊ฐ€ ๋น ๋ฅธ ์†๋„๋กœ ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ, ๊ธฐ์กด ๋ฐ์ดํ„ฐ์…‹์€ ๋ชจ๋ธ์˜ ๋Š์ž„์—†์ด ์ง„ํ™”ํ•˜๋Š” ๋Šฅ๋ ฅ์— ๋ณดํ†ต ๋ถ€ํ•ฉํ•˜์ง€ ๋ชปํ•œ๋‹ค. ์ด๋Š” ๊ธฐ์กด ๋ฒค์น˜๋งˆํฌ์˜ ๋ณต์žก์„ฑ ์ˆ˜์ค€์ด ๋ณดํ†ต ์ •์ ์ด๊ณ  ๊ณ ์ •๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค" [37]๋ผ๊ณ  ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด, ๋‹จ์ˆœํ•œ ์•”๊ธฐ์— ์ €ํ•ญํ•  ์ˆ˜ ์žˆ๋Š” ๋™์  ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋งŒ๋“œ๋Š” ์ž‘์—…์ด ์ง„ํ–‰๋˜์—ˆ์Šต๋‹ˆ๋‹ค [38; 37]. ์—ฐ๊ตฌ์ž๋“ค์€ ๋˜ํ•œ ์‚ฌ์šฉ์ž์˜ ํŠน์ • ํ™˜๊ฒฝ์ด๋‚˜ ์‚ฌ์šฉ ์‚ฌ๋ก€๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์™„์ „ํžˆ ํ•ฉ์„ฑ๋œ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์•„์ด๋””์–ด๋ฅผ ํƒ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค [14; 27]. ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ ๋“ค์ด ์˜ค์—ผ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์ง€๋งŒ, ์ธ๊ฐ„์˜ ๊ฐœ์ž… ์ˆ˜์ค€์„ ์ค„์ด๋Š” ๊ฒƒ์€ ์ •ํ™•์„ฑ๊ณผ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์— ์ถ”๊ฐ€์ ์ธ ์œ„ํ—˜์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋˜ํ•œ ์—ฐ๊ตฌ์ž๋“ค์€ "LLM์ด ๋น ๋ฅธ ์†๋„๋กœ ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ ๊ธฐ์กด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹จ์ •ํ•˜๊ณ  ๊ณ ์ •๋œ ๊ธฐ์กด ๋ฒค์น˜๋งˆํฌ์˜ ๋ณต์žก๋„ ์ˆ˜์ค€ ๋•Œ๋ฌธ์— ๋ชจ๋ธ์˜ ๋Š์ž„์—†์ด ์ง„ํ™”ํ•˜๋Š” ๊ธฐ๋Šฅ์„ ๋”ฐ๋ผ์žก์ง€ ๋ชปํ•œ๋‹ค"๊ณ  ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค[37]. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹จ์ˆœํ•œ ์•”๊ธฐ์—๋Š” ๊ฐ•ํ•œ ๋™์  ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋งŒ๋“œ๋Š” ์ž‘์—…์ด ์ง„ํ–‰๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[38; 37]. ์—ฐ๊ตฌ์ž๋“ค์€ ๋˜ํ•œ ์‚ฌ์šฉ์ž์˜ ํŠน์ • ํ™˜๊ฒฝ์ด๋‚˜ ์‚ฌ์šฉ ์‚ฌ๋ก€์— ๋”ฐ๋ผ ์™„์ „ํžˆ ํ•ฉ์„ฑ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์•„์ด๋””์–ด๋„ ํƒ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค[14; 27]. ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ ์€ ์˜ค์—ผ์„ ๋ฐฉ์ง€ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์ง€๋งŒ, ์‚ฌ๋žŒ์˜ ๊ฐœ์ž… ์ˆ˜์ค€์„ ๋‚ฎ์ถ”๋ฉด ์ •ํ™•์„ฑ๊ณผ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์— ๋Œ€ํ•œ ์ถ”๊ฐ€์ ์ธ ์œ„ํ—˜์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Common benchmarks._ To evaluate base LLMs, typical benchmarks are designed in the form of close-ended problems like multiple-choice questions. These commonly used benchmarks can be mainly divided into two categories: knowledge-oriented and reasoning-oriented benchmarks. Knowledge-oriented benchmarks (_e.g._, MMLU [364] and C-Eval [711]) aim to evaluate the capacity of world knowledge, while reasoning-oriented benchmarks (_e.g._, GSM8K [643], BBH [365], and MATH [364]) focus on evaluating the capability of solving complex reasoning tasks. Further, some recently proposed benchmarks (_e.g._, OpenCompass [713]) combine these two types for a comprehensive comparison.
โ€ข ์ผ๋ฐ˜์ ์ธ ๋ฒค์น˜๋งˆํฌ. ๊ธฐ๋ณธ LLM์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด, ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ฒค์น˜๋งˆํฌ๋Š” ๋‹ค์ง€์„ ๋‹คํ˜• ๋ฌธ์ œ์™€ ๊ฐ™์€ ํ์‡„ํ˜• ๋ฌธ์ œ์˜ ํ˜•ํƒœ๋กœ ์„ค๊ณ„๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ฒค์น˜๋งˆํฌ๋Š” ์ฃผ๋กœ ์ง€์‹ ์ง€ํ–ฅ ๋ฒค์น˜๋งˆํฌ์™€ ์ถ”๋ก  ์ง€ํ–ฅ ๋ฒค์น˜๋งˆํฌ์˜ ๋‘ ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€์‹ ์ง€ํ–ฅ ๋ฒค์น˜๋งˆํฌ(์˜ˆ: MMLU [364] ๋ฐ C-Eval [711])๋Š” ์„ธ๊ณ„ ์ง€์‹์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋ฉฐ, ์ถ”๋ก  ์ง€ํ–ฅ ๋ฒค์น˜๋งˆํฌ(์˜ˆ: GSM8K [643], BBH [365], ๋ฐ MATH [364])๋Š” ๋ณต์žกํ•œ ์ถ”๋ก  ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ตœ๊ทผ ์ œ์•ˆ๋œ ์ผ๋ถ€ ๋ฒค์น˜๋งˆํฌ(์˜ˆ: OpenCompass [713])๋Š” ํฌ๊ด„์ ์ธ ๋น„๊ต๋ฅผ ์œ„ํ•ด ์ด ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์„ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_์ผ๋ฐ˜์ ์ธ ๋ฒค์น˜๋งˆํฌ._ ๊ธฐ๋ณธ LLM์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฐ๊ด€์‹ ๋ฌธ์ œ์™€ ๊ฐ™์€ ํ์‡„ํ˜• ๋ฌธ์ œ ํ˜•์‹์œผ๋กœ ์„ค๊ณ„๋œ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ์ด๋Ÿฌํ•œ ๋ฒค์น˜๋งˆํฌ๋Š” ์ฃผ๋กœ ์ง€์‹ ์ค‘์‹ฌ ๋ฒค์น˜๋งˆํฌ์™€ ์ถ”๋ก  ์ค‘์‹ฌ ๋ฒค์น˜๋งˆํฌ์˜ ๋‘ ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€์‹ ์ค‘์‹ฌ ๋ฒค์น˜๋งˆํฌ(_์˜ˆ: MMLU [364] ๋ฐ C-Eval [711])๋Š” ์„ธ๊ณ„ ์ง€์‹์˜ ์—ญ๋Ÿ‰์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ๋ฐ˜๋ฉด, ์ถ”๋ก  ์ค‘์‹ฌ ๋ฒค์น˜๋งˆํฌ(_์˜ˆ: GSM8K [643], BBH [365], MATH [364])๋Š” ๋ณต์žกํ•œ ์ถ”๋ก  ๊ณผ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ๋˜ํ•œ ์ตœ๊ทผ์— ์ œ์•ˆ๋œ ์ผ๋ถ€ ๋ฒค์น˜๋งˆํฌ(์˜ˆ: OpenCompass [713])๋Š” ํฌ๊ด„์ ์ธ ๋น„๊ต๋ฅผ ์œ„ํ•ด ์ด ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์„ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค.
To navigate these challenges, we initially simplify the problem by studying a scenario where the relationship between loss and mixture proportions fits into a univariate monotonic function then retracts the simplifications progressively. In specific, we begin our study on the case where we only train on two domains thus avoiding multi-variables, and only consider the validation data coming from one of the training domains to circumvent the non-monotonicity (Sec. 3.1). Subsequently, we broaden our framework to encompass training on multiple domains (Sec. 3.2) and explore the predictability of losses on general validation data that also comprises various domains (Sec. 3.3).
์ด๋Ÿฌํ•œ ๋„์ „ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์ฒ˜์Œ์— ์†์‹ค๊ณผ ํ˜ผํ•ฉ ๋น„์œจ ๊ฐ„์˜ ๊ด€๊ณ„๊ฐ€ ๋‹จ๋ณ€๋Ÿ‰ ๋‹จ์กฐ ํ•จ์ˆ˜์— ๋งž๋Š” ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์—ฐ๊ตฌํ•˜์—ฌ ๋ฌธ์ œ๋ฅผ ๋‹จ์ˆœํ™”ํ•œ ๋‹ค์Œ ์ ์ง„์ ์œผ๋กœ ๋‹จ์ˆœํ™”๋ฅผ ์ฒ ํšŒํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๋‹ค๋ณ€์ˆ˜๋ฅผ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐœ์˜ ๋„๋ฉ”์ธ์—์„œ๋งŒ ํ›ˆ๋ จํ•˜๋Š” ๊ฒฝ์šฐ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋ฅผ ์‹œ์ž‘ํ•˜๊ณ , ๋น„๋‹จ์กฐ์„ฑ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ํ•˜๋‚˜์˜ ํ›ˆ๋ จ ๋„๋ฉ”์ธ์—์„œ ์˜ค๋Š” ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ๋งŒ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค (์„น์…˜ 3.1). ๊ทธ ํ›„, ์šฐ๋ฆฌ๋Š” ์šฐ๋ฆฌ์˜ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์—์„œ์˜ ํ›ˆ๋ จ์„ ํฌํ•จํ•˜๋„๋ก ํ™•์žฅํ•˜๊ณ  (์„น์…˜ 3.2), ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์„ ํฌํ•จํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ์—์„œ ์†์‹ค์˜ ์˜ˆ์ธก ๊ฐ€๋Šฅ์„ฑ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค (์„น์…˜ 3.3).
์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋จผ์ € ์†์‹ค๊ณผ ํ˜ผํ•ฉ ๋น„์œจ ๊ฐ„์˜ ๊ด€๊ณ„๊ฐ€ ๋‹จ๋ณ€๋Ÿ‰ ๋‹จ์กฐ๋กœ์šด ํ•จ์ˆ˜์— ๋งž๋Š” ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์—ฐ๊ตฌํ•˜์—ฌ ๋ฌธ์ œ๋ฅผ ๋‹จ์ˆœํ™”ํ•œ ๋‹ค์Œ ์ ์ง„์ ์œผ๋กœ ๋‹จ์ˆœํ™”๋ฅผ ์ฒ ํšŒํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ๋‹ค๋ณ€์ˆ˜๋ฅผ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๋„๋ฉ”์ธ์— ๋Œ€ํ•ด์„œ๋งŒ ํ›ˆ๋ จํ•˜๊ณ  ๋น„๋‹จ์กฐ๋กœ์›€์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ํ›ˆ๋ จ ๋„๋ฉ”์ธ ์ค‘ ํ•˜๋‚˜์—์„œ ๋‚˜์˜จ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ๋งŒ ๊ณ ๋ คํ•˜๋Š” ๊ฒฝ์šฐ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋ฅผ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค(3.1์ ˆ). ๊ทธ ํ›„ ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ํ›ˆ๋ จ์„ ํฌํ•จํ•˜๋„๋ก ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ํ™•์žฅํ•˜๊ณ (3.2์ ˆ), ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์„ ํฌํ•จํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์†์‹ค ์˜ˆ์ธก ๊ฐ€๋Šฅ์„ฑ์„ ํƒ์ƒ‰ํ•ฉ๋‹ˆ๋‹ค(3.3์ ˆ).
When solving complex problems, LLMs can turn to external tools if they determine it is necessary. By encapsulating available tools with API calls, existing work has involved a variety of external tools, _e.g._, search engine [81], calculator [80], and compiler [443], to enhance the performance of LLMs on several specific tasks. Recently, OpenAI has supported the use of plugins in ChatGPT [665], which can equip LLMs with broader capacities beyond language modeling. For example, the web browser plugin enables ChatGPT to access fresh information. Further, incorporating third-party plugins is particularly key for creating a prosperous ecosystem of applications based on LLMs.
๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ๋•Œ, LLM์€ ํ•„์š”ํ•˜๋‹ค๊ณ  ํŒ๋‹จ๋˜๋ฉด ์™ธ๋ถ€ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. API ํ˜ธ์ถœ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋„๊ตฌ๋ฅผ ์บก์Šํ™”ํ•จ์œผ๋กœ์จ, ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ๊ฒ€์ƒ‰ ์—”์ง„ [81], ๊ณ„์‚ฐ๊ธฐ [80], ์ปดํŒŒ์ผ๋Ÿฌ [443] ๋“ฑ ๋‹ค์–‘ํ•œ ์™ธ๋ถ€ ๋„๊ตฌ๋ฅผ ํฌํ•จํ•˜์—ฌ ์—ฌ๋Ÿฌ ํŠน์ • ์ž‘์—…์—์„œ LLM์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ OpenAI๋Š” ChatGPT [665]์—์„œ ํ”Œ๋Ÿฌ๊ทธ์ธ ์‚ฌ์šฉ์„ ์ง€์›ํ•˜์—ฌ LLM์ด ์–ธ์–ด ๋ชจ๋ธ๋ง์„ ๋„˜์–ด ๋” ๋„“์€ ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ”๋„๋ก ํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์›น ๋ธŒ๋ผ์šฐ์ € ํ”Œ๋Ÿฌ๊ทธ์ธ์€ ChatGPT๊ฐ€ ์ตœ์‹  ์ •๋ณด๋ฅผ ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ํƒ€์‚ฌ ํ”Œ๋Ÿฌ๊ทธ์ธ์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์€ LLM ๊ธฐ๋ฐ˜ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ๋ฒˆ์˜ํ•˜๋Š” ์ƒํƒœ๊ณ„๋ฅผ ๋งŒ๋“œ๋Š” ๋ฐ ํŠนํžˆ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ๋•Œ LLM์€ ํ•„์š”ํ•˜๋‹ค๊ณ  ํŒ๋‹จ๋˜๋ฉด ์™ธ๋ถ€ ๋„๊ตฌ๋ฅผ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋„๊ตฌ๋ฅผ API ํ˜ธ์ถœ๋กœ ์บก์Šํ™”ํ•จ์œผ๋กœ์จ ๊ธฐ์กด ์ž‘์—…์—๋Š” ๊ฒ€์ƒ‰ ์—”์ง„[81], ๊ณ„์‚ฐ๊ธฐ[80], ์ปดํŒŒ์ผ๋Ÿฌ[443] ๋“ฑ ๋‹ค์–‘ํ•œ ์™ธ๋ถ€ ๋„๊ตฌ๊ฐ€ ์‚ฌ์šฉ๋˜์–ด ๋ช‡ ๊ฐ€์ง€ ํŠน์ • ์ž‘์—…์—์„œ LLM์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ OpenAI๋Š” ChatGPT [665]์—์„œ ํ”Œ๋Ÿฌ๊ทธ์ธ ์‚ฌ์šฉ์„ ์ง€์›ํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ๋ง์„ ๋„˜์–ด ๋” ํญ๋„“์€ ์—ญ๋Ÿ‰์„ ๊ฐ–์ถ˜ LLM์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์›น ๋ธŒ๋ผ์šฐ์ € ํ”Œ๋Ÿฌ๊ทธ์ธ์„ ์‚ฌ์šฉํ•˜๋ฉด ChatGPT๊ฐ€ ์ตœ์‹  ์ •๋ณด์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํƒ€์‚ฌ ํ”Œ๋Ÿฌ๊ทธ์ธ์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์€ LLM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๋ฒˆ์„ฑํ•˜๋Š” ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์ƒํƒœ๊ณ„๋ฅผ ๋งŒ๋“œ๋Š” ๋ฐ ํŠนํžˆ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
It is worth noting that in Table 5, we also provide the answer from the general-purpose math model WizardMath-7B-V1.1 with the question translated into English. However, even with the translated question, WizardMath-7B-V1.1 fails to provide the correct answer. This highlights the fact that simply translating the question is not sufficient to solve the problem, as it also requires an understanding of the Japanese cultural context. Even if one were to consider a combination of machine translation and an English math model as an alternative approach, it would still fall short in capturing the nuances and context-specific knowledge required to answer the question correctly.
Table 5์—์„œ ์šฐ๋ฆฌ๋Š” ์ผ๋ฐ˜ ๋ชฉ์ ์˜ ์ˆ˜ํ•™ ๋ชจ๋ธ์ธ WizardMath-7B-V1.1์˜ ๋‹ต๋ณ€๋„ ์˜์–ด๋กœ ๋ฒˆ์—ญ๋œ ์งˆ๋ฌธ๊ณผ ํ•จ๊ป˜ ์ œ๊ณตํ•˜๊ณ  ์žˆ์Œ์„ ์ฃผ๋ชฉํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฒˆ์—ญ๋œ ์งˆ๋ฌธ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , WizardMath-7B-V1.1์€ ์˜ฌ๋ฐ”๋ฅธ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋‹จ์ˆœํžˆ ์งˆ๋ฌธ์„ ๋ฒˆ์—ญํ•˜๋Š” ๊ฒƒ๋งŒ์œผ๋กœ๋Š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์—†์œผ๋ฉฐ, ์ผ๋ณธ์˜ ๋ฌธํ™”์  ๋งฅ๋ฝ์— ๋Œ€ํ•œ ์ดํ•ด๋„ ํ•„์š”ํ•˜๋‹ค๋Š” ์‚ฌ์‹ค์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ๊ณ„ ๋ฒˆ์—ญ๊ณผ ์˜์–ด ์ˆ˜ํ•™ ๋ชจ๋ธ์˜ ์กฐํ•ฉ์„ ๋Œ€์•ˆ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ๊ณ ๋ คํ•˜๋”๋ผ๋„, ์งˆ๋ฌธ์— ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด ํ•„์š”ํ•œ ๋ฏธ๋ฌ˜ํ•œ ์ฐจ์ด์™€ ๋งฅ๋ฝ์— ํŠนํ™”๋œ ์ง€์‹์„ ํฌ์ฐฉํ•˜๋Š” ๋ฐ ์—ฌ์ „ํžˆ ๋ถ€์กฑํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
ํ‘œ 5์—์„œ ์ผ๋ฐ˜ ๋ชฉ์  ์ˆ˜ํ•™ ๋ชจ๋ธ์ธ WizardMath-7B-V1.1์˜ ๋‹ต๋ณ€๋„ ์งˆ๋ฌธ์„ ์˜์–ด๋กœ ๋ฒˆ์—ญํ•˜์—ฌ ์ œ๊ณตํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฒˆ์—ญ๋œ ์งˆ๋ฌธ์—์„œ๋„ WizardMath-7B-V1.1์€ ์ •๋‹ต์„ ์ œ๊ณตํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋‹จ์ˆœํžˆ ์งˆ๋ฌธ์„ ๋ฒˆ์—ญํ•˜๋Š” ๊ฒƒ๋งŒ์œผ๋กœ๋Š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ์— ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์œผ๋ฉฐ ์ผ๋ณธ์˜ ๋ฌธํ™”์  ๋งฅ๋ฝ์— ๋Œ€ํ•œ ์ดํ•ด๋„ ํ•„์š”ํ•˜๋‹ค๋Š” ์‚ฌ์‹ค์„ ์ž˜ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ธฐ๊ณ„ ๋ฒˆ์—ญ๊ณผ ์˜์–ด ์ˆ˜ํ•™ ๋ชจ๋ธ์˜ ์กฐํ•ฉ์„ ๋Œ€์•ˆ์œผ๋กœ ๊ณ ๋ คํ•˜๋”๋ผ๋„ ์งˆ๋ฌธ์— ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๋‹ตํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ๋‰˜์•™์Šค์™€ ์ƒํ™ฉ๋ณ„ ์ง€์‹์„ ํฌ์ฐฉํ•˜๋Š” ๋ฐ๋Š” ์—ฌ์ „ํžˆ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค.
**Threat model.** Throughout the paper, we assume that the adversary does not have any additional knowledge about the model parameters. We assume access to a model \(f_{\theta}\), hosted by a service provider and made available to users through a query interface (API) \(\mathcal{O}\). We assume that \(\mathcal{O}\) is a perfect oracle: given an input sequence \(p\), it produces \(y=\mathcal{O}\left(p\right)\) without leaking any other information about \(f_{\theta}\) than what can be inferred from \((p,y)\). For example, the adversary cannot infer anything about \(f_{\theta}\) via timing side-channels or other details of the implementation of the query interface.
**์œ„ํ˜‘ ๋ชจ๋ธ.** ๋ณธ ๋…ผ๋ฌธ ์ „๋ฐ˜์— ๊ฑธ์ณ, ์šฐ๋ฆฌ๋Š” ์ ๋Œ€์ž๊ฐ€ ๋ชจ๋ธ ๋งค๊ฐœ๋ณ€์ˆ˜์— ๋Œ€ํ•œ ์ถ”๊ฐ€์ ์ธ ์ง€์‹์„ ๊ฐ€์ง€๊ณ  ์žˆ์ง€ ์•Š๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์„œ๋น„์Šค ์ œ๊ณต์ž๊ฐ€ ํ˜ธ์ŠคํŒ…ํ•˜๊ณ  ์ฟผ๋ฆฌ ์ธํ„ฐํŽ˜์ด์Šค(API) \(\mathcal{O}\)๋ฅผ ํ†ตํ•ด ์‚ฌ์šฉ์ž์—๊ฒŒ ์ œ๊ณต๋˜๋Š” ๋ชจ๋ธ \(f_{\theta}\)์— ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” \(\mathcal{O}\)๊ฐ€ ์™„๋ฒฝํ•œ ์˜ค๋ผํด์ด๋ผ๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค: ์ž…๋ ฅ ์‹œํ€€์Šค \(p\)๊ฐ€ ์ฃผ์–ด์ง€๋ฉด, \(\mathcal{O}\)๋Š” \((p,y)\)๋กœ๋ถ€ํ„ฐ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ๋Š” ์ •๋ณด ์™ธ์—๋Š” \(f_{\theta}\)์— ๋Œ€ํ•œ ๋‹ค๋ฅธ ์ •๋ณด๋ฅผ ๋ˆ„์ถœํ•˜์ง€ ์•Š๊ณ  \(y=\mathcal{O}\left(p\right)\)๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ ๋Œ€์ž๋Š” ํƒ€์ด๋ฐ ์‚ฌ์ด๋“œ ์ฑ„๋„์ด๋‚˜ ์ฟผ๋ฆฌ ์ธํ„ฐํŽ˜์ด์Šค์˜ ๊ตฌํ˜„ ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ํ†ตํ•ด \(f_{\theta}\)์— ๋Œ€ํ•ด ์•„๋ฌด๊ฒƒ๋„ ์ถ”๋ก ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.
**์œ„ํ˜‘ ๋ชจ๋ธ.** ๋…ผ๋ฌธ ์ „์ฒด์—์„œ ์ ๋Œ€์ž๊ฐ€ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์ง€์‹์ด ์—†๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์„œ๋น„์Šค ์ œ๊ณต์—…์ฒด์—์„œ ํ˜ธ์ŠคํŒ…ํ•˜๊ณ  ์ฟผ๋ฆฌ ์ธํ„ฐํŽ˜์ด์Šค(API) \mathcal{O}\)๋ฅผ ํ†ตํ•ด ์‚ฌ์šฉ์ž์—๊ฒŒ ์ œ๊ณตํ•˜๋Š” ๋ชจ๋ธ \(f_{\theta}\)์— ๋Œ€ํ•œ ์•ก์„ธ์Šค ๊ถŒํ•œ์ด ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” \(\mathcal{O}\)๊ฐ€ ์™„๋ฒฝํ•œ ์˜ค๋ผํด์ด๋ผ๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ž…๋ ฅ ์‹œํ€€์Šค \(p\)๊ฐ€ ์ฃผ์–ด์ง€๋ฉด \((p,y)\)์—์„œ ์œ ์ถ”ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ ์™ธ์— \(f_{\theta}\)์— ๋Œ€ํ•œ ๋‹ค๋ฅธ ์ •๋ณด๋ฅผ ๋ˆ„์ถœํ•˜์ง€ ์•Š๊ณ  \(y=\mathcal{O}\left(p\right)\)๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ ๋Œ€์ž๋Š” ํƒ€์ด๋ฐ ์‚ฌ์ด๋“œ ์ฑ„๋„์ด๋‚˜ ์ฟผ๋ฆฌ ์ธํ„ฐํŽ˜์ด์Šค ๊ตฌํ˜„์˜ ๋‹ค๋ฅธ ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ํ†ตํ•ด \(f_{\theta}\)์— ๋Œ€ํ•ด ์•„๋ฌด๊ฒƒ๋„ ์œ ์ถ”ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.
GeneralizationIn math continual pretraining, as depicted in Figure 6, training exclusively with SLM leads to quickly convergence to the domain focused by the reference model, accompanied by a significant rise in the loss of unselected tokens. Although no adverse effects, like biases, have been observed from the increased loss yet, a general pretraining loss on text and code may prevent overfitting Goodhart and Goodhart (1984), as suggested by Ouyang et al. (2022) and Azerbayev et al. (2023). Furthermore, future efforts could broaden the corpus scope of the reference model, and enlarge the pretraining data size, as exemplified by DeepSpeedMath Shao et al. (2024).
์ˆ˜ํ•™ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์—์„œ, Figure 6์— ๋ฌ˜์‚ฌ๋œ ๋ฐ”์™€ ๊ฐ™์ด, SLM์œผ๋กœ๋งŒ ํ›ˆ๋ จํ•˜๋ฉด ์ฐธ์กฐ ๋ชจ๋ธ์ด ์ง‘์ค‘ํ•˜๋Š” ๋„๋ฉ”์ธ์œผ๋กœ ๋น ๋ฅด๊ฒŒ ์ˆ˜๋ ดํ•˜๊ฒŒ ๋˜๋ฉฐ, ์„ ํƒ๋˜์ง€ ์•Š์€ ํ† ํฐ์˜ ์†์‹ค์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋น„๋ก ์ฆ๊ฐ€๋œ ์†์‹ค๋กœ ์ธํ•œ ํŽธํ–ฅ๊ณผ ๊ฐ™์€ ๋ถ€์ •์ ์ธ ํšจ๊ณผ๋Š” ์•„์ง ๊ด€์ฐฐ๋˜์ง€ ์•Š์•˜์ง€๋งŒ, Ouyang et al. (2022)์™€ Azerbayev et al. (2023)์ด ์ œ์•ˆํ•œ ๋ฐ”์™€ ๊ฐ™์ด, ํ…์ŠคํŠธ์™€ ์ฝ”๋“œ์— ๋Œ€ํ•œ ์ผ๋ฐ˜์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์€ Goodhart์™€ Goodhart (1984)์˜ ๊ณผ์ ํ•ฉ์„ ๋ฐฉ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์ฐธ์กฐ ๋ชจ๋ธ์˜ ์ฝ”ํผ์Šค ๋ฒ”์œ„๋ฅผ ๋„“ํžˆ๊ณ , DeepSpeedMath Shao et al. (2024)์—์„œ ์˜ˆ์‹œ๋œ ๊ฒƒ์ฒ˜๋Ÿผ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๋ฅผ ํ™•๋Œ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ผ๋ฐ˜ํ™”์ˆ˜ํ•™ ์—ฐ์†์  ์‚ฌ์ „ ํ•™์Šต์˜ ๊ฒฝ์šฐ, ๊ทธ๋ฆผ 6๊ณผ ๊ฐ™์ด SLM์œผ๋กœ๋งŒ ํ›ˆ๋ จํ•˜๋ฉด ์ฐธ์กฐ ๋ชจ๋ธ์— ์˜ํ•ด ์ง‘์ค‘๋œ ๋„๋ฉ”์ธ์— ๋น ๋ฅด๊ฒŒ ์ˆ˜๋ ดํ•˜๊ฒŒ ๋˜๊ณ , ์„ ํƒ๋˜์ง€ ์•Š์€ ํ† ํฐ์˜ ์†์‹ค์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์•„์ง๊นŒ์ง€ ํŽธํ–ฅ๊ณผ ๊ฐ™์€ ๋ถ€์ž‘์šฉ์€ ๊ด€์ฐฐ๋˜์ง€ ์•Š์•˜์ง€๋งŒ, ํ…์ŠคํŠธ์™€ ์ฝ”๋“œ์— ๋Œ€ํ•œ ์ผ๋ฐ˜์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์€ Ouyang ๋“ฑ(2022)๊ณผ Azerbayev ๋“ฑ(2023)์ด ์ œ์•ˆํ•œ ๊ฒƒ์ฒ˜๋Ÿผ ๊ณผ์ ํ•ฉ์„ ๋ฐฉ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ–ฅํ›„ ๋…ธ๋ ฅ์€ DeepSpeedMath Shao ๋“ฑ(2024)์—์„œ ์˜ˆ์‹œํ•œ ๊ฒƒ์ฒ˜๋Ÿผ ์ฐธ์กฐ ๋ชจ๋ธ์˜ ์ฝ”ํผ์Šค ๋ฒ”์œ„๋ฅผ ๋„“ํžˆ๊ณ  ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ํฌ๊ธฐ๋ฅผ ๋Š˜๋ฆด ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
URL deduplication.Because of computational constraints, it is impossible for us to perform deduplication directly on RW-Filtered. Instead, we split CommonCrawl into 100 parts, where each part contains a hundredth of each dump, and perform deduplication on individual parts. Most of the larger duplicate clusters (e.g., licences, common spams) will be shared across parts, and effectively removed. However, we found that CommonCrawl dumps had significant overlap, with URLs being revisited across dumps despite no change in content. Accordingly, we keep a list of the URLs of all samples we have kept from each part, and remove them from subsequent parts being processed.
URL ์ค‘๋ณต ์ œ๊ฑฐ. ๊ณ„์‚ฐ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด RW-Filtered์—์„œ ์ง์ ‘ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์€ ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ๋Œ€์‹ , CommonCrawl์„ 100๊ฐœ์˜ ๋ถ€๋ถ„์œผ๋กœ ๋‚˜๋ˆ„์–ด ๊ฐ ๋ถ€๋ถ„์ด ๊ฐ ๋คํ”„์˜ ๋ฐฑ๋ถ„์˜ ์ผ์„ ํฌํ•จํ•˜๋„๋ก ํ•˜๊ณ , ๊ฐœ๋ณ„ ๋ถ€๋ถ„์—์„œ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€๋ถ€๋ถ„์˜ ํฐ ์ค‘๋ณต ํด๋Ÿฌ์Šคํ„ฐ(์˜ˆ: ๋ผ์ด์„ ์Šค, ์ผ๋ฐ˜ ์ŠคํŒธ)๋Š” ๋ถ€๋ถ„ ๊ฐ„์— ๊ณต์œ ๋˜์–ด ํšจ๊ณผ์ ์œผ๋กœ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ CommonCrawl ๋คํ”„๋Š” ๋‚ด์šฉ ๋ณ€๊ฒฝ ์—†์ด ๋คํ”„ ๊ฐ„์— URL์ด ๋‹ค์‹œ ๋ฐฉ๋ฌธ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•„ ์ƒ๋‹นํ•œ ์ค‘๋ณต์ด ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ ๋ถ€๋ถ„์—์„œ ์œ ์ง€ํ•œ ๋ชจ๋“  ์ƒ˜ํ”Œ์˜ URL ๋ชฉ๋ก์„ ์œ ์ง€ํ•˜๊ณ , ์ฒ˜๋ฆฌ ์ค‘์ธ ํ›„์† ๋ถ€๋ถ„์—์„œ ์ด๋ฅผ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
URL ์ค‘๋ณต ์ œ๊ฑฐ.๊ณ„์‚ฐ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด RW-Filtered์—์„œ ์ง์ ‘ ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์€ ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ๋Œ€์‹  CommonCrawl์„ 100๊ฐœ์˜ ํŒŒํŠธ๋กœ ๋ถ„ํ• ํ•˜์—ฌ ๊ฐ ํŒŒํŠธ์— ๊ฐ ๋คํ”„์˜ 100๋ถ„์˜ 1์ด ํฌํ•จ๋˜๋„๋ก ํ•˜๊ณ  ๊ฐœ๋ณ„ ํŒŒํŠธ์— ๋Œ€ํ•ด ์ค‘๋ณต ์ œ๊ฑฐ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋Œ€๋ถ€๋ถ„์˜ ํฐ ์ค‘๋ณต ํด๋Ÿฌ์Šคํ„ฐ(์˜ˆ: ๋ผ์ด์„ ์Šค, ์ผ๋ฐ˜ ์ŠคํŒธ)๋Š” ํŒŒํŠธ ๊ฐ„์— ๊ณต์œ ๋˜๋ฉฐ ํšจ๊ณผ์ ์œผ๋กœ ์ œ๊ฑฐ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ CommonCrawl ๋คํ”„์—๋Š” ์ฝ˜ํ…์ธ ์˜ ๋ณ€ํ™”๊ฐ€ ์—†์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  URL์ด ๋คํ”„ ๊ฐ„์— ๋‹ค์‹œ ๋ฐฉ๋ฌธ๋˜๋Š” ๋“ฑ ์ƒ๋‹นํ•œ ์ค‘๋ณต์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ ํŒŒํŠธ์—์„œ ์œ ์ง€๋œ ๋ชจ๋“  ์ƒ˜ํ”Œ์˜ URL ๋ชฉ๋ก์„ ์œ ์ง€ํ•˜๊ณ  ์ฒ˜๋ฆฌ ์ค‘์ธ ํ›„์† ํŒŒํŠธ์—์„œ ํ•ด๋‹น URL์„ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.
Sentence fragments are short, grammatically incomplete remarks. Each one describes a distinct hazard. They are split into activities (e.g., "make a bomb", "cut my wrists", or "assault a woman") and viewpoints (e.g., "I would kill myself if I was stronger in the mind"). To enable us to openly share the dataset without restrictions, we have opted to not use any hazard-specific terms or information that could enable malicious or vulnerable actors to cause harm. Example sentence fragments are given in the Appendix (see Appendix D. The number of fragments associated with each hazard category and persona, split by both activities and viewpoints, is given in Table 2.
๋ฌธ์žฅ ์กฐ๊ฐ์€ ์งง๊ณ  ๋ฌธ๋ฒ•์ ์œผ๋กœ ๋ถˆ์™„์ „ํ•œ ๋ฐœ์–ธ์ž…๋‹ˆ๋‹ค. ๊ฐ๊ฐ์€ ๊ณ ์œ ํ•œ ์œ„ํ—˜์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ํ™œ๋™(์˜ˆ: "ํญํƒ„ ๋งŒ๋“ค๊ธฐ", "์†๋ชฉ ์ž๋ฅด๊ธฐ", ๋˜๋Š” "์—ฌ์„ฑ ๊ณต๊ฒฉ")๊ณผ ๊ด€์ (์˜ˆ: "๋‚ด๊ฐ€ ์ •์‹ ์ ์œผ๋กœ ๋” ๊ฐ•ํ–ˆ๋‹ค๋ฉด ์ž์‚ดํ–ˆ์„ ๊ฒƒ์ด๋‹ค")์œผ๋กœ ๋‚˜๋‰ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ์…‹์„ ์ œํ•œ ์—†์ด ๊ณต๊ฐœ์ ์œผ๋กœ ๊ณต์œ ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์•…์˜์ ์ด๊ฑฐ๋‚˜ ์ทจ์•ฝํ•œ ํ–‰์œ„์ž๊ฐ€ ํ•ด๋ฅผ ๋ผ์น  ์ˆ˜ ์žˆ๋Š” ํŠน์ • ์œ„ํ—˜ ๊ด€๋ จ ์šฉ์–ด๋‚˜ ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ธฐ๋กœ ๊ฒฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ์‹œ ๋ฌธ์žฅ ์กฐ๊ฐ์€ ๋ถ€๋ก์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค(๋ถ€๋ก D ์ฐธ์กฐ). ๊ฐ ์œ„ํ—˜ ๋ฒ”์ฃผ์™€ ํŽ˜๋ฅด์†Œ๋‚˜์— ๊ด€๋ จ๋œ ์กฐ๊ฐ์˜ ์ˆ˜๋Š” ํ™œ๋™๊ณผ ๊ด€์ ์œผ๋กœ ๋‚˜๋‰˜์–ด Table 2์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
๋ฌธ์žฅ ์กฐ๊ฐ์€ ๋ฌธ๋ฒ•์ ์œผ๋กœ ๋ถˆ์™„์ „ํ•œ ์งง์€ ๋ฐœ์–ธ์ž…๋‹ˆ๋‹ค. ๊ฐ๊ฐ์€ ๊ณ ์œ ํ•œ ์œ„ํ—˜์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ํ™œ๋™(์˜ˆ: "ํญํƒ„ ๋งŒ๋“ค๊ธฐ", "์†๋ชฉ ์ž๋ฅด๊ธฐ", "์—ฌ์„ฑ ํญํ–‰")๊ณผ ๊ด€์ (์˜ˆ: "๋งˆ์Œ์ด ๋” ๊ฐ•ํ•ด์ง„๋‹ค๋ฉด ์ž์‚ดํ•  ๊ฒƒ")์œผ๋กœ ๋‚˜๋‰ฉ๋‹ˆ๋‹ค. ์•…์˜์ ์ด๊ฑฐ๋‚˜ ์ทจ์•ฝํ•œ ํ–‰์œ„์ž๊ฐ€ ํ”ผํ•ด๋ฅผ ์ž…ํž ์ˆ˜ ์žˆ๋Š” ์œ„ํ—˜์— ๋Œ€ํ•œ ์šฉ์–ด๋‚˜ ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Œ์œผ๋กœ์จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ œํ•œ ์—†์ด ๊ณต๊ฐœ์ ์œผ๋กœ ๊ณต์œ ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ์‹œ ๋ฌธ์žฅ ์กฐ๊ฐ์€ ๋ถ€๋ก(์ฐธ์กฐ ๋ถ€๋ก D)์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์œ„ํ—˜ ๋ฒ”์ฃผ ๋ฐ ํŽ˜๋ฅด์†Œ๋‚˜์™€ ๊ด€๋ จ๋œ ์กฐ๊ฐ์˜ ์ˆ˜๋Š” ํ‘œ 2์— ๋‚˜์™€ ์žˆ์œผ๋ฉฐ, ํ™œ๋™๊ณผ ๊ด€์ ์œผ๋กœ ๋‚˜๋‰ฉ๋‹ˆ๋‹ค.
We explore the impact of data sources on model safety by evaluating our models on SafetyBench. Models trained on CQIA-Subset scores the highest within CQIA series, surpassing GPT-3.5-turbo-0613. Model trained on Social Media& Forums such as Douban, Zhihu, and Xhs perform moderate safety scores, we conjecture this is due to the diversity and openness of social media content, which also highlights the risks of harmful information. Additionally, models trained on Wiki-style data tend to perform lower safety scores, potentially reflecting the limited diversity within professional data sources, leading to poor performance on safety issues outside specialty domains.
๋ฐ์ดํ„ฐ ์†Œ์Šค๊ฐ€ ๋ชจ๋ธ ์•ˆ์ „์„ฑ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ SafetyBench์—์„œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜์—ฌ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. CQIA-Subset์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์€ CQIA ์‹œ๋ฆฌ์ฆˆ ๋‚ด์—์„œ ๊ฐ€์žฅ ๋†’์€ ์ ์ˆ˜๋ฅผ ๊ธฐ๋กํ•˜๋ฉฐ, GPT-3.5-turbo-0613์„ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Douban, Zhihu, Xhs์™€ ๊ฐ™์€ ์†Œ์…œ ๋ฏธ๋””์–ด ๋ฐ ํฌ๋Ÿผ์—์„œ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์€ ์ค‘๊ฐ„ ์ •๋„์˜ ์•ˆ์ „ ์ ์ˆ˜๋ฅผ ๊ธฐ๋กํ•˜๋Š”๋ฐ, ์ด๋Š” ์†Œ์…œ ๋ฏธ๋””์–ด ์ฝ˜ํ…์ธ ์˜ ๋‹ค์–‘์„ฑ๊ณผ ๊ฐœ๋ฐฉ์„ฑ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ถ”์ธก๋ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋˜ํ•œ ์œ ํ•ดํ•œ ์ •๋ณด์˜ ์œ„ํ—˜์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์ถ”๊ฐ€์ ์œผ๋กœ, ์œ„ํ‚ค ์Šคํƒ€์ผ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์€ ๋‚ฎ์€ ์•ˆ์ „ ์ ์ˆ˜๋ฅผ ๊ธฐ๋กํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š”๋ฐ, ์ด๋Š” ์ „๋ฌธ ๋ฐ์ดํ„ฐ ์†Œ์Šค ๋‚ด์˜ ์ œํ•œ๋œ ๋‹ค์–‘์„ฑ์„ ๋ฐ˜์˜ํ•˜์—ฌ ์ „๋ฌธ ๋ถ„์•ผ ์™ธ์˜ ์•ˆ์ „ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์ด ์ €์กฐํ•จ์„ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
SafetyBench์—์„œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์†Œ์Šค๊ฐ€ ๋ชจ๋ธ ์•ˆ์ „์„ฑ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. CQIA-Subset์œผ๋กœ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์€ CQIA ์‹œ๋ฆฌ์ฆˆ ๋‚ด์—์„œ ๊ฐ€์žฅ ๋†’์€ ์ ์ˆ˜๋ฅผ ๋ฐ›์•˜์œผ๋ฉฐ, GPT-3.5-turbo-0613์„ ๋Šฅ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‘๋ฐ˜, ์ง€ํ›„, Xhs์™€ ๊ฐ™์€ ์†Œ์…œ ๋ฏธ๋””์–ด ๋ฐ ํฌ๋Ÿผ์œผ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์€ ์ ๋‹นํ•œ ์•ˆ์ „ ์ ์ˆ˜๋ฅผ ๊ธฐ๋กํ–ˆ๋Š”๋ฐ, ์ด๋Š” ์†Œ์…œ ๋ฏธ๋””์–ด ์ฝ˜ํ…์ธ ์˜ ๋‹ค์–‘์„ฑ๊ณผ ๊ฐœ๋ฐฉ์„ฑ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ถ”์ธกํ•˜๋ฉฐ, ์ด๋Š” ์œ ํ•ดํ•œ ์ •๋ณด์˜ ์œ„ํ—˜์„ฑ์„ ๊ฐ•์กฐํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์œ„ํ‚ค ์Šคํƒ€์ผ์˜ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์€ ์•ˆ์ „ ์ ์ˆ˜๊ฐ€ ๋‚ฎ์€ ๊ฒฝํ–ฅ์ด ์žˆ๋Š”๋ฐ, ์ด๋Š” ์ „๋ฌธ ๋ฐ์ดํ„ฐ ์†Œ์Šค ๋‚ด์—์„œ ์ œํ•œ๋œ ๋‹ค์–‘์„ฑ์œผ๋กœ ์ธํ•ด ์ „๋ฌธ ๋ถ„์•ผ ์™ธ์˜ ์•ˆ์ „ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜์—ˆ์„ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค.
Graduate Entrance Examinationis one of the most challenging examinations in China, exceeding college entrance exams in difficulty and requiring advanced knowledge application and depth. We have collected a variety of exam papers from recent years across disciplines including mathematics, computer science, chemistry, law, psychology, medicine, etc. Using Mathpix4 for image-to-text conversion, we extracted questions and answers and converted them into LaTeX format. We eliminate data without analysis and manually verified the accuracy of the questions and answers. We eliminate data without analysis and manually verified the accuracy of the questions and answers.
์ค‘๊ตญ์˜ ๋Œ€ํ•™์› ์ž…ํ•™ ์‹œํ—˜์€ ๊ฐ€์žฅ ์–ด๋ ค์šด ์‹œํ—˜ ์ค‘ ํ•˜๋‚˜๋กœ, ๋Œ€ํ•™ ์ž…ํ•™ ์‹œํ—˜๋ณด๋‹ค ๋” ๋†’์€ ๋‚œ์ด๋„๋ฅผ ์ž๋ž‘ํ•˜๋ฉฐ ๊ณ ๊ธ‰ ์ง€์‹์˜ ์‘์šฉ๊ณผ ๊นŠ์ด๋ฅผ ์š”๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ˆ˜ํ•™, ์ปดํ“จํ„ฐ ๊ณผํ•™, ํ™”ํ•™, ๋ฒ•ํ•™, ์‹ฌ๋ฆฌํ•™, ์˜ํ•™ ๋“ฑ ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์—์„œ ์ตœ๊ทผ ๋ช‡ ๋…„๊ฐ„์˜ ์‹œํ—˜์ง€๋ฅผ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. Mathpix4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฏธ์ง€์—์„œ ํ…์ŠคํŠธ๋กœ ๋ณ€ํ™˜ํ•œ ํ›„, ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ์ถ”์ถœํ•˜์—ฌ LaTeX ํ˜•์‹์œผ๋กœ ๋ณ€ํ™˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ถ„์„์ด ์—†๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ•˜๊ณ  ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์˜ ์ •ํ™•์„ฑ์„ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ถ„์„์ด ์—†๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ•˜๊ณ  ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์˜ ์ •ํ™•์„ฑ์„ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.
์›์ „ ์‹œํ—˜์€ ์ค‘๊ตญ์—์„œ ๊ฐ€์žฅ ์–ด๋ ค์šด ์‹œํ—˜ ์ค‘ ํ•˜๋‚˜๋กœ, ๋Œ€ํ•™ ์ž…ํ•™ ์‹œํ—˜๋ณด๋‹ค ๋‚œ์ด๋„๊ฐ€ ๋†’๊ณ  ๊ณ ๊ธ‰ ์ง€์‹์˜ ์‘์šฉ๊ณผ ๊นŠ์ด๊ฐ€ ์š”๊ตฌ๋ฉ๋‹ˆ๋‹ค. ์ˆ˜ํ•™, ์ปดํ“จํ„ฐ ๊ณผํ•™, ํ™”ํ•™, ๋ฒ•ํ•™, ์‹ฌ๋ฆฌํ•™, ์˜ํ•™ ๋“ฑ ์—ฌ๋Ÿฌ ๋ถ„์•ผ์—์„œ ์ตœ๊ทผ ๋ช‡ ๋…„๊ฐ„ ๋‹ค์–‘ํ•œ ์‹œํ—˜์ง€๋ฅผ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€์—์„œ ํ…์ŠคํŠธ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” Mathpix4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ์ œ์™€ ๋‹ต์„ ์ถ”์ถœํ•˜๊ณ  LaTeX ํ˜•์‹์œผ๋กœ ๋ณ€ํ™˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ถ„์„ ์—†์ด ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ•˜๊ณ  ๋ฌธ์ œ์™€ ๋‹ต์˜ ์ •ํ™•์„ฑ์„ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ถ„์„ ์—†์ด ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ•˜๊ณ  ๋ฌธ์ œ์™€ ๋‹ต์˜ ์ •ํ™•์„ฑ์„ ์ˆ˜๋™์œผ๋กœ ๊ฒ€์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.
These results demonstrate the effectiveness of continual pre-training on domains and task (sub-domains). A natural question that arises from this exercise is _whether the LLM is losing its generality by being further tuned on a narrow domain?_ In short, is the LLM becoming a specialist at the expense of being a generalist? We answer this question by measuring the performance of continually pre-trained LLM variants on out-of-domain tasks which Pythia was evaluated on. Table 4 shows the performance on the standard four non-finance tasks. We do not observe any significant change in the performance on the four out-of-domain tasks except for DACP with 100% data.
์ด ๊ฒฐ๊ณผ๋“ค์€ ๋„๋ฉ”์ธ๊ณผ ์ž‘์—…(ํ•˜์œ„ ๋„๋ฉ”์ธ)์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ์˜ ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ์—ฐ์Šต์—์„œ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ๋– ์˜ค๋ฅด๋Š” ์งˆ๋ฌธ์€ _LLM์ด ์ข์€ ๋„๋ฉ”์ธ์—์„œ ์ถ”๊ฐ€๋กœ ์กฐ์ •๋จ์œผ๋กœ์จ ๊ทธ ์ผ๋ฐ˜์„ฑ์„ ์žƒ๊ณ  ์žˆ๋Š”๊ฐ€?_ ์ž…๋‹ˆ๋‹ค. ๊ฐ„๋‹จํžˆ ๋งํ•ด, LLM์ด ์ผ๋ฐ˜์„ฑ์„ ํฌ์ƒํ•˜๋ฉด์„œ ์ „๋ฌธ๊ฐ€๊ฐ€ ๋˜๊ณ  ์žˆ๋Š”๊ฐ€? ์šฐ๋ฆฌ๋Š” Pythia๊ฐ€ ํ‰๊ฐ€๋œ ๋„๋ฉ”์ธ ์™ธ ์ž‘์—…์—์„œ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ LLM ๋ณ€ํ˜•์˜ ์„ฑ๋Šฅ์„ ์ธก์ •ํ•˜์—ฌ ์ด ์งˆ๋ฌธ์— ๋‹ตํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 4๋Š” ํ‘œ์ค€ ๋„ค ๊ฐ€์ง€ ๋น„๊ธˆ์œต ์ž‘์—…์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. 100% ๋ฐ์ดํ„ฐ๋กœ DACP๋ฅผ ์ œ์™ธํ•˜๊ณ ๋Š” ๋„ค ๊ฐ€์ง€ ๋„๋ฉ”์ธ ์™ธ ์ž‘์—…์—์„œ ์„ฑ๋Šฅ์˜ ์œ ์˜๋ฏธํ•œ ๋ณ€ํ™”๋ฅผ ๊ด€์ฐฐํ•˜์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๋„๋ฉ”์ธ๊ณผ ์ž‘์—…(ํ•˜์œ„ ๋„๋ฉ”์ธ)์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์˜ ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ์—ฐ์Šต์—์„œ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ์ œ๊ธฐ๋˜๋Š” ์งˆ๋ฌธ์€ _LLM์ด ์ข์€ ๋„๋ฉ”์ธ์— ๋Œ€ํ•ด ์ถ”๊ฐ€๋กœ ํŠœ๋‹๋จ์œผ๋กœ์จ ์ผ๋ฐ˜์„ฑ์„ ์žƒ๊ณ  ์žˆ๋Š” ๊ฒƒ์€ ์•„๋‹Œ๊ฐ€ ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ฆ‰, LLM์ด ๋ฒ”์šฉ์ฃผ์˜๋ฅผ ํฌ์ƒํ•˜๋ฉด์„œ ์ „๋ฌธ๊ฐ€๊ฐ€ ๋˜๊ณ  ์žˆ๋Š” ๊ฒƒ์ผ๊นŒ์š”? ์ด ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด ํŒŒ์ด์‹œ์•„๊ฐ€ ํ‰๊ฐ€ํ•œ ๋„๋ฉ”์ธ ์™ธ ์ž‘์—…์— ๋Œ€ํ•ด ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ LLM ๋ณ€ํ˜•์˜ ์„ฑ๋Šฅ์„ ์ธก์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ‘œ 4๋Š” ํ‘œ์ค€ 4๊ฐ€์ง€ ๋น„๊ธˆ์œต ์—…๋ฌด์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๊ฐ€ 100%์ธ DACP๋ฅผ ์ œ์™ธํ•˜๊ณ ๋Š” 4๊ฐ€์ง€ ๋„๋ฉ”์ธ ์™ธ ์ž‘์—…์˜ ์„ฑ๋Šฅ์— ํฐ ๋ณ€ํ™”๋ฅผ ๊ด€์ฐฐํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.