question
stringlengths
101
2.69k
chosen
stringlengths
43
2.27k
rejected
stringlengths
42
2.18k
The memory- or speed- inefficiency can be resolved when \(\pi_{\text{ref}}\) is set as a uniform prior \(U\), as the terms \(\pi_{\text{ref}}(y_{w}|x)\) and \(\pi_{\text{ref}}(y_{l}|x)\) cancel each other out. This negates the need for additional computations and storage beyond the policy model itself.
๋ฉ”๋ชจ๋ฆฌ ๋˜๋Š” ์†๋„์˜ ๋น„ํšจ์œจ์„ฑ์€ \(\pi_{\text{ref}}\)๋ฅผ ๊ท ์ผ ์‚ฌ์ „ \(U\)๋กœ ์„ค์ •ํ•˜๋ฉด ํ•ด๊ฒฐ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” \(\pi_{\text{ref}}(y_{w}|x)\)์™€ \(\pi_{\text{ref}}(y_{l}|x)\) ํ•ญ์ด ์„œ๋กœ ์ƒ์‡„๋˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ด๋กœ ์ธํ•ด ์ •์ฑ… ๋ชจ๋ธ ์ž์ฒด ์™ธ์— ์ถ”๊ฐ€์ ์ธ ๊ณ„์‚ฐ ๋ฐ ์ €์žฅ์ด ํ•„์š”ํ•˜์ง€ ์•Š๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.
pi_{\text{ref}}\)๋ฅผ ๊ท ์ผํ•œ ์„ ํ–‰ \(U\)๋กœ ์„ค์ •ํ•˜๋ฉด \(\pi_{\text{ref}}(y_{w}|x\) ํ•ญ๊ณผ \(\pi_{\text{ref}}(y_{l}|x\) ํ•ญ์ด ์„œ๋กœ ์ƒ์‡„๋˜๋ฏ€๋กœ ๋ฉ”๋ชจ๋ฆฌ ๋˜๋Š” ์†๋„ ๋น„ํšจ์œจ์„ฑ์„ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์ •์ฑ… ๋ชจ๋ธ ์ž์ฒด ์™ธ์— ์ถ”๊ฐ€ ๊ณ„์‚ฐ๊ณผ ์ €์žฅ ๊ณต๊ฐ„์ด ํ•„์š”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
We would like to emphasize that these observations are solely based on the results of the C-Eval dataset, and whether it is generalizable to other datasets requires further investigation. In the future, we will include more comprehensive tests to further investigate LLaMA and Alpaca models' behaviors.
์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์€ ์˜ค์ง C-Eval ๋ฐ์ดํ„ฐ์…‹์˜ ๊ฒฐ๊ณผ์— ๊ธฐ๋ฐ˜ํ•œ ๊ฒƒ์ž„์„ ๊ฐ•์กฐํ•˜๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ์…‹์— ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์—ฌ๋ถ€๋Š” ์ถ”๊ฐ€ ์กฐ์‚ฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์•ž์œผ๋กœ ์šฐ๋ฆฌ๋Š” LLaMA์™€ Alpaca ๋ชจ๋ธ์˜ ๋™์ž‘์„ ๋” ์ž์„ธํžˆ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ํ…Œ์ŠคํŠธ๋ฅผ ํฌํ•จํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์€ C-Eval ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฐ๊ณผ์—๋งŒ ๊ทผ๊ฑฐํ•œ ๊ฒƒ์ด๋ฉฐ, ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์—ฌ๋ถ€๋Š” ์ถ”๊ฐ€ ์กฐ์‚ฌ๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•˜๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. ํ–ฅํ›„์—๋Š” ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ํ…Œ์ŠคํŠธ๋ฅผ ํฌํ•จํ•˜์—ฌ LLaMA ๋ฐ ์•ŒํŒŒ์นด ๋ชจ๋ธ์˜ ๋™์ž‘์„ ๋” ์ž์„ธํžˆ ์กฐ์‚ฌํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
URLs with many duplicates.Table 11 shows the URLs had the largest proportion of examples identified by NearDup as near-duplicates. For C4, these tend to be websites that sell many similar products and thus have a large amount of templated text. For RealNews, content aggregators seem especially common.
์ค‘๋ณต์ด ๋งŽ์€ URL. Table 11์€ NearDup์— ์˜ํ•ด ์œ ์‚ฌ ์ค‘๋ณต์œผ๋กœ ์‹๋ณ„๋œ ์˜ˆ์ œ์˜ ๋น„์œจ์ด ๊ฐ€์žฅ ํฐ URL์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. C4์˜ ๊ฒฝ์šฐ, ์ด๋Ÿฌํ•œ URL์€ ๋งŽ์€ ์œ ์‚ฌ ์ œํ’ˆ์„ ํŒ๋งคํ•˜๋Š” ์›น์‚ฌ์ดํŠธ๋กœ, ๋”ฐ๋ผ์„œ ๋งŽ์€ ์–‘์˜ ํ…œํ”Œ๋ฆฟ ํ…์ŠคํŠธ๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. RealNews์˜ ๊ฒฝ์šฐ, ์ฝ˜ํ…์ธ  ์ง‘๊ณ„ ์‚ฌ์ดํŠธ๊ฐ€ ํŠนํžˆ ํ”ํ•œ ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค.
์ค‘๋ณต์ด ๋งŽ์€ URLํ‘œ 11์€ NearDup์ด ๊ทผ์ ‘ ์ค‘๋ณต์œผ๋กœ ์‹๋ณ„ํ•œ ์˜ˆ์ œ์˜ ๋น„์œจ์ด ๊ฐ€์žฅ ๋†’์€ URL์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. C4์˜ ๊ฒฝ์šฐ, ์œ ์‚ฌํ•œ ์ œํ’ˆ์„ ๋งŽ์ด ํŒ๋งคํ•˜๋Š” ์›น์‚ฌ์ดํŠธ์ด๋ฏ€๋กœ ํ…œํ”Œ๋ฆฟ ํ…์ŠคํŠธ๊ฐ€ ๋งŽ์ด ํฌํ•จ๋˜์–ด ์žˆ๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. RealNews์˜ ๊ฒฝ์šฐ ์ฝ˜ํ…์ธ  ์• ๊ทธ๋ฆฌ๊ฒŒ์ดํ„ฐ๊ฐ€ ํŠนํžˆ ํ”ํ•œ ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค.
**Reference [2]** Kerala State Film Awards: Kerala Film Critics Award Kerala State Government Award (Professional Drama): Asianet Film Awards: Kerala Kaumadi Gallup Poll Award: Academic G. Venugopal G Venugopal (born 10 December 1960) is an Indian playback singer known for his work in Malayalam films.
**Reference [2]** Kerala State Film Awards: Kerala Film Critics Award Kerala State Government Award (Professional Drama): Asianet Film Awards: Kerala Kaumadi Gallup Poll Award: Academic G. Venugopal G Venugopal (1960๋…„ 12์›” 10์ผ ์ถœ์ƒ)๋Š” ๋ง๋ผ์–„๋žŒ ์˜ํ™”์—์„œ์˜ ์ž‘์—…์œผ๋กœ ์•Œ๋ ค์ง„ ์ธ๋„ ํ”Œ๋ ˆ์ด๋ฐฑ ๊ฐ€์ˆ˜์ž…๋‹ˆ๋‹ค.
**์ฐธ๊ณ  [2]** ์ผ€๋ž„๋ผ ์ฃผ ์˜ํ™”์ƒ: ์ผ€๋ž„๋ผ ์˜ํ™” ํ‰๋ก ๊ฐ€์ƒ ์ผ€๋ž„๋ผ ์ฃผ ์ •๋ถ€ ์ƒ (์ „๋ฌธ ๋“œ๋ผ๋งˆ): ์•„์‹œ์•ˆ ๋„ท ์˜ํ™”์ƒ: ์ผ€๋ž„๋ผ ์นด์šฐ๋งˆ๋”” ๊ฐค๋Ÿฝ ํด ์ƒ: ํ•™์ˆ  G. ๋ฒ ๋ˆ„๊ณ ํŒ” G ๋ฒ ๋ˆ„๊ณ ํŒ” (1960 ๋…„ 12 ์›” 10 ์ผ ์ถœ์ƒ)์€ ๋ง๋ผ์–„๋žŒ ์˜ํ™”์—์„œ ์ผํ•œ ๊ฒƒ์œผ๋กœ ์œ ๋ช…ํ•œ ์ธ๋„ ์žฌ์ƒ ๊ฐ€์ˆ˜์ž…๋‹ˆ๋‹ค.
Contributors to **pretraining dataset construction and tooling** (Dolma) include Russell Authur, Iz Beltagy, Akshita Bhagia, Khyathi Chandu, Jesse Dodge, Yanai Elazar, Dirk Groeneveld, Rodney Kinney, Kyle Lo, Aakanksha Naik, Abhilasha Ravichander, Dustin Schwenk, Luca Soldaini, and Nishant Subramani.
**pretraining dataset construction and tooling** (Dolma)์— ๊ธฐ์—ฌํ•œ ์‚ฌ๋žŒ๋“ค๋กœ๋Š” Russell Authur, Iz Beltagy, Akshita Bhagia, Khyathi Chandu, Jesse Dodge, Yanai Elazar, Dirk Groeneveld, Rodney Kinney, Kyle Lo, Aakanksha Naik, Abhilasha Ravichander, Dustin Schwenk, Luca Soldaini, ๊ทธ๋ฆฌ๊ณ  Nishant Subramani๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ตฌ์ถ• ๋ฐ ํˆด๋ง(๋Œ๋งˆ)์— ๊ธฐ์—ฌํ•œ ์‚ฌ๋žŒ์œผ๋กœ๋Š” ๋Ÿฌ์…€ ์˜ค์„œ, ์ด์ฆˆ ๋ฒจํƒ€๊ธฐ, ์•…์‹œํƒ€ ๋ฐ”์ง€์•„, ์นด์•ผํ‹ฐ ์ฐฌ๋‘, ์ œ์‹œ ๋„์ง€, ์•ผ๋‚˜์ด ์—˜๋ผ์ž๋ฅด, ๋””๋ฅดํฌ ๊ทธ๋กœ์—”๋ฒจํŠธ, ๋กœ๋“œ๋‹ˆ ํ‚ค๋‹ˆ, ์นด์ผ ๋กœ, ์•„์นธ์ƒค ๋‚˜์ต, ์•„๋นŒ๋ผ์‚ฌ ๋ผ๋น„์ฐฌ๋”, ๋”์Šคํ‹ด ์Šˆ์›ฌํฌ, ๋ฃจ์นด ์†”๋‹ค์ด๋‹ˆ, ๋‹ˆ์ƒจํŠธ ์ˆ˜๋ธŒ๋ผ๋งˆ๋‹ˆ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
To explore this, we adjust EntRec\((\mathbf{x}^{l})\) in the direction of its steepest increase, represented by \(\nabla_{\mathbf{x}^{l}}\)EntRec\((\mathbf{x}^{l})\), and observe the impact on CnstScore\((\mathbf{x}^{l})\) by modifying \(\mathbf{x}^{l}\) according to a magnitude of change \(\alpha\):
์ด๋ฅผ ํƒ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” EntRec\((\mathbf{x}^{l})\)์„ ๊ฐ€์žฅ ๊ฐ€ํŒŒ๋ฅด๊ฒŒ ์ฆ๊ฐ€ํ•˜๋Š” ๋ฐฉํ–ฅ์ธ \(\nabla_{\mathbf{x}^{l}}\)EntRec\((\mathbf{x}^{l})\)์œผ๋กœ ์กฐ์ •ํ•˜๊ณ , ๋ณ€ํ™”์˜ ํฌ๊ธฐ \(\alpha\)์— ๋”ฐ๋ผ \(\mathbf{x}^{l}\)์„ ์ˆ˜์ •ํ•˜์—ฌ CnstScore\((\mathbf{x}^{l})\)์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค.
์ด๋ฅผ ํƒ์ƒ‰ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ€์žฅ ๊ฐ€ํŒŒ๋ฅธ ์ฆ๊ฐ€ ๋ฐฉํ–ฅ์œผ๋กœ EntRec\((\mathbf{x}^{l})\)๋ฅผ ์กฐ์ •ํ•˜๊ณ , \(\nabla_{\mathbf{x}^{l}}\)EntRec\((\mathbf{x}^{l})\)๋กœ ํ‘œ์‹œ๋˜๋ฉฐ, ๋ณ€ํ™” ํฌ๊ธฐ \(\alpha\)์— ๋”ฐ๋ผ \(\mathbf{x}^{l}\)๋ฅผ ์ˆ˜์ •ํ•˜์—ฌ CnstScore\((\mathbf{x}^{l})\)์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค:
The overall performance trends obtained through the auto-evaluator align well with specialist assessments in Figure 3 despite marginal differences in the computed accuracy values, as shown in Figure A.9. These results demonstrate that the auto-evaluator is a valid surrogate for the specialist raters.
์ž๋™ ํ‰๊ฐ€๊ธฐ๋ฅผ ํ†ตํ•ด ์–ป์€ ์ „์ฒด ์„ฑ๋Šฅ ์ถ”์„ธ๋Š” Figure A.9์— ๋‚˜ํƒ€๋‚œ ๊ณ„์‚ฐ๋œ ์ •ํ™•๋„ ๊ฐ’์˜ ๋ฏธ์„ธํ•œ ์ฐจ์ด์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  Figure 3์˜ ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€์™€ ์ž˜ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ์ž๋™ ํ‰๊ฐ€๊ธฐ๊ฐ€ ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€์ž๋ฅผ ๋Œ€์ฒดํ•  ์ˆ˜ ์žˆ๋Š” ์œ ํšจํ•œ ๋Œ€๋ฆฌ์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์ž๋™ ํ‰๊ฐ€๊ธฐ๋ฅผ ํ†ตํ•ด ์–ป์€ ์ „์ฒด ์„ฑ๋Šฅ ์ถ”์„ธ๋Š” ๊ณ„์‚ฐ๋œ ์ •ํ™•๋„ ๊ฐ’์˜ ์•ฝ๊ฐ„์˜ ์ฐจ์ด์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๊ทธ๋ฆผ A.9์™€ ๊ฐ™์ด ๊ทธ๋ฆผ 3์˜ ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€์™€ ์ž˜ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ์ž๋™ ํ‰๊ฐ€๊ธฐ๊ฐ€ ์ „๋ฌธ๊ฐ€ ํ‰๊ฐ€์ž์— ๋Œ€ํ•œ ์œ ํšจํ•œ ๋Œ€๋ฆฌ์ž์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
Proof.: \(\pi_{w}\) represents an ideal policy that perfectly aligns the true data distribution of the preferred data. Hence, for any given data point \((x,y_{w},y_{l})\) from the preference dataset \(\mathcal{D}\), the conditions \(\pi_{w}(y_{w}|x)=1\) and \(0\leq\pi_{w}(y_{l}|x)\leq 1\) hold true.
์ฆ๋ช…: \(\pi_{w}\)๋Š” ์„ ํ˜ธํ•˜๋Š” ๋ฐ์ดํ„ฐ์˜ ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์™€ ์™„๋ฒฝํ•˜๊ฒŒ ์ผ์น˜ํ•˜๋Š” ์ด์ƒ์ ์ธ ์ •์ฑ…์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์„ ํ˜ธ ๋ฐ์ดํ„ฐ์…‹ \(\mathcal{D}\)์—์„œ ์ฃผ์–ด์ง„ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ \((x,y_{w},y_{l})\)์— ๋Œ€ํ•ด, ์กฐ๊ฑด \(\pi_{w}(y_{w}|x)=1\) ๋ฐ \(0\leq\pi_{w}(y_{l}|x)\leq 1\)์ด ์„ฑ๋ฆฝํ•ฉ๋‹ˆ๋‹ค.
์ฆ๋ช….: \(\pi_{w}\)๋Š” ์„ ํ˜ธ ๋ฐ์ดํ„ฐ์˜ ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์™€ ์™„๋ฒฝํ•˜๊ฒŒ ์ผ์น˜ํ•˜๋Š” ์ด์ƒ์ ์ธ ์ •์ฑ…์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์„ ํ˜ธ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ \mathcal{D}\)์˜ ์ฃผ์–ด์ง„ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ \((x,y_{w},y_{l})\)์— ๋Œ€ํ•ด \(\pi_{w}(y_{w}|x)=1\) ๋ฐ \(0\leq\pi_{w}(y_{l}|x)\leq 1\) ์กฐ๊ฑด์ด ์œ ์ง€๋ฉ๋‹ˆ๋‹ค.
I want you act as a Prompt Rewriter. Your objective is to rewrite a given prompt into a more complex version using dataformat to make those famous AI systems (e.g., chatgpt and GPT4) more difficult to handle. But the rewritten prompt must be reasonable and must be understood and responded by humans.
๋‚˜๋Š” ๋‹น์‹ ์ด ํ”„๋กฌํ”„ํŠธ ์žฌ์ž‘์„ฑ์ž๋กœ์„œ ํ™œ๋™ํ•˜๊ธฐ๋ฅผ ์›ํ•ฉ๋‹ˆ๋‹ค. ๋‹น์‹ ์˜ ๋ชฉํ‘œ๋Š” ์ฃผ์–ด์ง„ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฐ์ดํ„ฐ ํ˜•์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ๋” ๋ณต์žกํ•œ ๋ฒ„์ „์œผ๋กœ ์žฌ์ž‘์„ฑํ•˜์—ฌ ์œ ๋ช…ํ•œ AI ์‹œ์Šคํ…œ(์˜ˆ: chatgpt ๋ฐ GPT4)์ด ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ต๊ฒŒ ๋งŒ๋“œ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์žฌ์ž‘์„ฑ๋œ ํ”„๋กฌํ”„ํŠธ๋Š” ํ•ฉ๋ฆฌ์ ์ด์–ด์•ผ ํ•˜๋ฉฐ ์ธ๊ฐ„์ด ์ดํ•ดํ•˜๊ณ  ์‘๋‹ตํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
ํ”„๋กฌํ”„ํŠธ ์žฌ์ž‘์„ฑ์ž ์—ญํ• ์„ ํ•˜๊ธธ ์›ํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์–ด์ง„ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฐ์ดํ„ฐ ํ˜•์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ๋” ๋ณต์žกํ•œ ๋ฒ„์ „์œผ๋กœ ์žฌ์ž‘์„ฑํ•˜์—ฌ ์œ ๋ช…ํ•œ AI ์‹œ์Šคํ…œ(์˜ˆ: chatgpt ๋ฐ GPT4)์„ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ต๊ฒŒ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์žฌ์ž‘์„ฑ๋œ ํ”„๋กฌํ”„ํŠธ๋Š” ํ•ฉ๋ฆฌ์ ์ด์–ด์•ผ ํ•˜๋ฉฐ, ์‚ฌ๋žŒ์ด ์ดํ•ดํ•˜๊ณ  ์‘๋‹ตํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Our evaluation demonstrates that the Yi model family achieves inspiring performance on a wide range of tasks and delivers close to GPT-3.5 user preference rate. We first report the base model performance on standard benchmarks, then we discuss the chat model performance and its user preference rate.
์šฐ๋ฆฌ์˜ ํ‰๊ฐ€์— ๋”ฐ๋ฅด๋ฉด Yi ๋ชจ๋ธ ํŒจ๋ฐ€๋ฆฌ๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ๋†€๋ผ์šด ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋ฉฐ GPT-3.5 ์‚ฌ์šฉ์ž ์„ ํ˜ธ๋„์— ๊ทผ์ ‘ํ•œ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ํ‘œ์ค€ ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•œ ํ›„, ์ฑ„ํŒ… ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ๊ณผ ์‚ฌ์šฉ์ž ์„ ํ˜ธ๋„์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค.
์šฐ๋ฆฌ์˜ ํ‰๊ฐ€๋Š” Yi ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์ด ๊ด‘๋ฒ”์œ„ํ•œ ์ž‘์—…์—์„œ ๊ณ ๋ฌด์ ์ธ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๊ณ  GPT-3.5 ์‚ฌ์šฉ์ž ์„ ํ˜ธ๋„์— ๊ทผ์ ‘ํ•œ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋จผ์ € ํ‘œ์ค€ ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ธฐ๋ณธ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•œ ๋‹ค์Œ ์ฑ„ํŒ… ๋ชจ๋ธ ์„ฑ๋Šฅ๊ณผ ์‚ฌ์šฉ์ž ์„ ํ˜ธ๋„์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
In addition to the above library resources, existing deep learning frameworks (_e.g._, PyTorch [197], TensorFlow [198], MXNet [199], PaddlePaddle [200], MindSpore [136] and OneFlow [201]) have also provided the support for parallel algorithms, which are commonly used for training large-scale models.
์œ„์˜ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์ž์› ์™ธ์—๋„, ๊ธฐ์กด์˜ ๋”ฅ๋Ÿฌ๋‹ ํ”„๋ ˆ์ž„์›Œํฌ๋“ค(์˜ˆ: PyTorch [197], TensorFlow [198], MXNet [199], PaddlePaddle [200], MindSpore [136], OneFlow [201])์€ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ ํ›ˆ๋ จ์— ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ณ‘๋ ฌ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋Œ€ํ•œ ์ง€์›๋„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์œ„์˜ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ๋ฆฌ์†Œ์Šค ์™ธ์—๋„ ๊ธฐ์กด ๋”ฅ๋Ÿฌ๋‹ ํ”„๋ ˆ์ž„์›Œํฌ(์˜ˆ: PyTorch [197], TensorFlow [198], MXNet [199], PaddlePaddle [200], MindSpore [136] ๋ฐ OneFlow [201])๋„ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ ํ›ˆ๋ จ์— ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋ณ‘๋ ฌ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋Œ€ํ•œ ์ง€์›์„ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค.
The new memory states \(M_{\text{s}}\) and \(z_{\text{s}}\) are then passed to the next segment \(S+1\), building in a recurrence in each attention layer. The right side term \(\sigma(K)^{T}V\) in Eq. (4) is known as an associative binding operator (Smolensky, 1990; Hebb, 2005; Schlag et al., 2020).
์ƒˆ๋กœ์šด ๋ฉ”๋ชจ๋ฆฌ ์ƒํƒœ \(M_{\text{s}}\)์™€ \(z_{\text{s}}\)๋Š” ๋‹ค์Œ ์„ธ๊ทธ๋จผํŠธ \(S+1\)๋กœ ์ „๋‹ฌ๋˜์–ด ๊ฐ ์–ดํ…์…˜ ๋ ˆ์ด์–ด์—์„œ ๋ฐ˜๋ณต์„ฑ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ์‹ (4)์˜ ์˜ค๋ฅธ์ชฝ ํ•ญ \(\sigma(K)^{T}V\)๋Š” ์—ฐ๊ด€ ๊ฒฐํ•ฉ ์—ฐ์‚ฐ์ž(associative binding operator)๋กœ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค (Smolensky, 1990; Hebb, 2005; Schlag et al., 2020).
๊ทธ๋Ÿฐ ๋‹ค์Œ ์ƒˆ๋กœ์šด ๋ฉ”๋ชจ๋ฆฌ ์ƒํƒœ \(M_{\text{s}}\) ๋ฐ \(z_{\text{s}}\)๋Š” ๋‹ค์Œ ์„ธ๊ทธ๋จผํŠธ \(S+1\)๋กœ ์ „๋‹ฌ๋˜์–ด ๊ฐ ์ฃผ์˜ ๊ณ„์ธต์—์„œ ๋ฐ˜๋ณต์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ๋ฐฉ์ •์‹ (4)์˜ ์˜ค๋ฅธ์ชฝ ํ•ญ \(\sigma(K)^{T}V\)๋Š” ์—ฐ์ƒ์  ๋ฐ”์ธ๋”ฉ ์—ฐ์‚ฐ์ž(Smolensky, 1990; Hebb, 2005; Schlag et al., 2020)๋กœ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค.
Origin of the list.We use an aggregated list+ of about 4.6M URLs that we explicitly ban. This list is broken in categories (e.g. pornography, gambling); we outline the categories we selected in Table 13. The list is regularly updated, with an original intended usage as a blocklist for universities.
๋ชฉ๋ก์˜ ์ถœ์ฒ˜. ์šฐ๋ฆฌ๋Š” ์•ฝ 460๋งŒ ๊ฐœ์˜ URL์„ ๋ช…์‹œ์ ์œผ๋กœ ์ฐจ๋‹จํ•˜๋Š” ์ง‘๊ณ„๋œ ๋ชฉ๋ก์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชฉ๋ก์€ ์นดํ…Œ๊ณ ๋ฆฌ(์˜ˆ: ํฌ๋ฅด๋…ธ, ๋„๋ฐ•)๋กœ ๋‚˜๋‰˜๋ฉฐ, Table 13์—์„œ ์„ ํƒํ•œ ์นดํ…Œ๊ณ ๋ฆฌ๋ฅผ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชฉ๋ก์€ ์›๋ž˜ ๋Œ€ํ•™์˜ ์ฐจ๋‹จ ๋ชฉ๋ก์œผ๋กœ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์ •๊ธฐ์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋ฉ๋‹ˆ๋‹ค.
๋ชฉ๋ก์˜ ๊ธฐ์›.์šฐ๋ฆฌ๋Š” ๋ช…์‹œ์ ์œผ๋กœ ๊ธˆ์ง€ํ•˜๋Š” ์•ฝ 460๋งŒ ๊ฐœ์˜ URL๋กœ ๊ตฌ์„ฑ๋œ ์ง‘๊ณ„๋œ ๋ชฉ๋ก+์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชฉ๋ก์€ ์นดํ…Œ๊ณ ๋ฆฌ๋ณ„๋กœ ๋‚˜๋‰˜์–ด์ ธ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ์Œ๋ž€๋ฌผ, ๋„๋ฐ•). ์„ ํƒํ•œ ์นดํ…Œ๊ณ ๋ฆฌ๋Š” ํ‘œ 13์— ์„ค๋ช…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชฉ๋ก์€ ์ •๊ธฐ์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋˜๋ฉฐ, ๋Œ€ํ•™์šฉ ์ฐจ๋‹จ ๋ชฉ๋ก์œผ๋กœ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
3. Virtue ethical theories: - Aristotelian ethics - which holds that the morality of an action depends on the character or traits of the person performing the action. - Virtue ethics - which holds that the morality of an action depends on the character or traits of the person performing the action.
3. ๋• ์œค๋ฆฌ ์ด๋ก : - ์•„๋ฆฌ์Šคํ† ํ…”๋ ˆ์Šค ์œค๋ฆฌํ•™ - ํ–‰๋™์˜ ๋„๋•์„ฑ์€ ํ–‰๋™์„ ์ˆ˜ํ–‰ํ•˜๋Š” ์‚ฌ๋žŒ์˜ ์„ฑ๊ฒฉ์ด๋‚˜ ํŠน์„ฑ์— ๋‹ฌ๋ ค ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. - ๋• ์œค๋ฆฌ - ํ–‰๋™์˜ ๋„๋•์„ฑ์€ ํ–‰๋™์„ ์ˆ˜ํ–‰ํ•˜๋Š” ์‚ฌ๋žŒ์˜ ์„ฑ๊ฒฉ์ด๋‚˜ ํŠน์„ฑ์— ๋‹ฌ๋ ค ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค.
3. ๋• ์œค๋ฆฌ ์ด๋ก : - ์•„๋ฆฌ์Šคํ† ํ…”๋ ˆ์Šค ์œค๋ฆฌ - ์–ด๋–ค ํ–‰๋™์˜ ๋„๋•์„ฑ์€ ๊ทธ ํ–‰๋™์„ ์ˆ˜ํ–‰ํ•˜๋Š” ์‚ฌ๋žŒ์˜ ์„ฑ๊ฒฉ์ด๋‚˜ ํŠน์„ฑ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง„๋‹ค๋Š” ์ž…์žฅ์ž…๋‹ˆ๋‹ค. - ๋• ์œค๋ฆฌ - ์–ด๋–ค ํ–‰๋™์˜ ๋„๋•์„ฑ์€ ๊ทธ ํ–‰๋™์„ ์ˆ˜ํ–‰ํ•˜๋Š” ์‚ฌ๋žŒ์˜ ์„ฑ๊ฒฉ์ด๋‚˜ ํŠน์„ฑ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง„๋‹ค๋Š” ์ž…์žฅ์ž…๋‹ˆ๋‹ค.
We present a selection of user-oriented tasks, the corresponding GPT3\({}_{\text{SELF-INST}}\)-produced responses and annotator ratings in Table 9. We see that even for responses rated as level C, the model demonstrates extensive steps in solving the task, even though its final output is incorrect.
์‚ฌ์šฉ์ž ์ง€ํ–ฅ์ ์ธ ์ž‘์—… ์„ ํƒ, ํ•ด๋‹น GPT3\({}_{\text{SELF-INST}}\)์ด ์ƒ์„ฑํ•œ ์‘๋‹ต ๋ฐ ํ‰๊ฐ€์ž์˜ ํ‰๊ฐ€๋ฅผ Table 9์— ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋น„๋ก ์ตœ์ข… ์ถœ๋ ฅ์ด ์ž˜๋ชป๋˜์—ˆ๋”๋ผ๋„, C ์ˆ˜์ค€์œผ๋กœ ํ‰๊ฐ€๋œ ์‘๋‹ต์—์„œ๋„ ๋ชจ๋ธ์ด ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์žˆ์–ด ๊ด‘๋ฒ”์œ„ํ•œ ๋‹จ๊ณ„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
์‚ฌ์šฉ์ž ์ค‘์‹ฌ ์ž‘์—…, ํ•ด๋‹น GPT3\({}_{\text{SELF-INST}}\)-์ƒ์„ฑ๋œ ์‘๋‹ต ๋ฐ ์ฃผ์„๊ฐ€ ํ‰๊ฐ€๋ฅผ ํ‘œ 9์— ์„ ํƒ์ ์œผ๋กœ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜์ค€ C๋กœ ํ‰๊ฐ€๋œ ์‘๋‹ต์˜ ๊ฒฝ์šฐ์—๋„ ์ตœ์ข… ์ถœ๋ ฅ์ด ํ‹€๋ ธ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋ชจ๋ธ์ด ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ๊ด‘๋ฒ”์œ„ํ•œ ๋‹จ๊ณ„๋ฅผ ๋ณด์—ฌ์ฃผ๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Paraphrase I** The mayoral election in Indianapolis took place on November 7, 2023, with preliminary elections occurring on May 2. The sitting mayor, Democrat Joe Hogsett, successfully ran for a third term. Both Hogsett and his Republican opponent, Jefferson Shreve, moved on to the main election.
**Paraphrase I** 2023๋…„ 11์›” 7์ผ ์ธ๋””์• ๋‚˜ํด๋ฆฌ์Šค์—์„œ ์‹œ์žฅ ์„ ๊ฑฐ๊ฐ€ ์—ด๋ ธ์œผ๋ฉฐ, ์˜ˆ๋น„ ์„ ๊ฑฐ๋Š” 5์›” 2์ผ์— ์ง„ํ–‰๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์ง ์‹œ์žฅ์ธ ๋ฏผ์ฃผ๋‹น์˜ ์กฐ ํ˜ธ๊ทธ์…‹์€ ์„ฑ๊ณต์ ์œผ๋กœ 3์„ ์— ๋„์ „ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ˜ธ๊ทธ์…‹๊ณผ ๊ทธ์˜ ๊ณตํ™”๋‹น ์ƒ๋Œ€์ธ ์ œํผ์Šจ ์Šˆ๋ฆฌ๋ธŒ๋Š” ๋ชจ๋‘ ๋ณธ์„ ์— ์ง„์ถœํ–ˆ์Šต๋‹ˆ๋‹ค.
**์˜์—ญ I** ์ธ๋””์• ๋‚˜ํด๋ฆฌ์Šค ์‹œ์žฅ ์„ ๊ฑฐ๋Š” 2023๋…„ 11์›” 7์ผ์— ์‹ค์‹œ๋˜์—ˆ์œผ๋ฉฐ, ์˜ˆ๋น„ ์„ ๊ฑฐ๋Š” 5์›” 2์ผ์— ์‹ค์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์ง ์‹œ์žฅ์ธ ๋ฏผ์ฃผ๋‹น์˜ ์กฐ ํ˜ธ์…‹์ด 3์„ ์— ์„ฑ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ํ˜ธ์…‹๊ณผ ๊ณตํ™”๋‹น์˜ ์ œํผ์Šจ ์Šˆ๋ผ์ด๋ธŒ ํ›„๋ณด ๋ชจ๋‘ ๋ณธ์„ ์— ์ง„์ถœํ–ˆ์Šต๋‹ˆ๋‹ค.
Due to high demand, we have yet to critique this request. That said, we assure that the review will be produced in due time by our diligent and unwavering staff in a professional manner. This site is highly regarded amongst its peers in terms of speed and reliability, so feel free to check us out!
๋†’์€ ์ˆ˜์š”๋กœ ์ธํ•ด, ์šฐ๋ฆฌ๋Š” ์•„์ง ์ด ์š”์ฒญ์„ ๊ฒ€ํ† ํ•˜์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ ‡์ง€๋งŒ, ์šฐ๋ฆฌ์˜ ์„ฑ์‹คํ•˜๊ณ  ๋ณ€ํ•จ์—†๋Š” ์ง์›๋“ค์ด ์ „๋ฌธ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์ ์‹œ์— ๊ฒ€ํ† ๋ฅผ ์™„๋ฃŒํ•  ๊ฒƒ์ž„์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด ์‚ฌ์ดํŠธ๋Š” ์†๋„์™€ ์‹ ๋ขฐ์„ฑ ๋ฉด์—์„œ ๋™๋ฃŒ๋“ค ์‚ฌ์ด์—์„œ ๋งค์šฐ ๋†’์€ ํ‰๊ฐ€๋ฅผ ๋ฐ›๊ณ  ์žˆ์œผ๋‹ˆ, ์–ธ์ œ๋“ ์ง€ ํ™•์ธํ•ด ๋ณด์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค!
์ˆ˜์š”๊ฐ€ ๋งŽ์•„์„œ ์•„์ง ์ด ์š”์ฒญ์— ๋Œ€ํ•œ ๋ฆฌ๋ทฐ๋ฅผ ์ž‘์„ฑํ•˜์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋ถ€์ง€๋Ÿฐํ•˜๊ณ  ๋ณ€ํ•จ์—†๋Š” ์ง์›๋“ค์ด ์ „๋ฌธ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์ ์‹œ์— ๋ฆฌ๋ทฐ๋ฅผ ์ž‘์„ฑํ•  ๊ฒƒ์„ ์•ฝ์†๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์ด ์‚ฌ์ดํŠธ๋Š” ์†๋„์™€ ์‹ ๋ขฐ์„ฑ ์ธก๋ฉด์—์„œ ๋™์ข… ์—…๊ณ„์—์„œ ๋†’์€ ํ‰๊ฐ€๋ฅผ ๋ฐ›๊ณ  ์žˆ์œผ๋‹ˆ ์ž์œ ๋กญ๊ฒŒ ํ™•์ธํ•ด ๋ณด์„ธ์š”!
Larger data-to-model ratio means that we can absorb more data into the a smaller model than we previously thought, which is more efficient for inference and deployment. And we hope WSD LRS will help more researchers to explore \(L(N,D)\) with less efforts and make the relationship clearer in LLMs.
๋” ํฐ ๋ฐ์ดํ„ฐ ๋Œ€ ๋ชจ๋ธ ๋น„์œจ์€ ์šฐ๋ฆฌ๊ฐ€ ์ด์ „์— ์ƒ๊ฐํ–ˆ๋˜ ๊ฒƒ๋ณด๋‹ค ๋” ์ž‘์€ ๋ชจ๋ธ์— ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋ฅผ ํก์ˆ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ์˜๋ฏธํ•˜๋ฉฐ, ์ด๋Š” ์ถ”๋ก  ๋ฐ ๋ฐฐํฌ์— ๋” ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ๋Š” WSD LRS๊ฐ€ ๋” ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด \(L(N,D)\)๋ฅผ ๋” ์ ์€ ๋…ธ๋ ฅ์œผ๋กœ ํƒ๊ตฌํ•˜๊ณ  LLMs์—์„œ์˜ ๊ด€๊ณ„๋ฅผ ๋” ๋ช…ํ™•ํ•˜๊ฒŒ ๋งŒ๋“œ๋Š” ๋ฐ ๋„์›€์ด ๋˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ ๋Œ€ ๋ชจ๋ธ ๋น„์œจ์ด ํด์ˆ˜๋ก ์ด์ „๋ณด๋‹ค ๋” ์ž‘์€ ๋ชจ๋ธ์— ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ๋ฅผ ํก์ˆ˜ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ถ”๋ก  ๋ฐ ๋ฐฐํฌ์— ๋” ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ WSD LRS๊ฐ€ ๋” ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด ๋” ์ ์€ ๋…ธ๋ ฅ์œผ๋กœ \(L(N,D)\)๋ฅผ ํƒ๊ตฌํ•˜๊ณ  LLM์—์„œ ๊ด€๊ณ„๋ฅผ ๋” ๋ช…ํ™•ํ•˜๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ฃผ๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
**Limitations.** This paper acknowledges the limitations of our automatic GPT-4 and human evaluation methods. This method poses challenges for scalability and reliability. Moreover, our test set may not represent all the scenarios or domains where LLM can be applied or compared with other methods.
**์ œํ•œ ์‚ฌํ•ญ.** ๋ณธ ๋…ผ๋ฌธ์€ ์ž๋™ GPT-4 ๋ฐ ์ธ๊ฐ„ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์˜ ํ•œ๊ณ„๋ฅผ ์ธ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ํ™•์žฅ์„ฑ๊ณผ ์‹ ๋ขฐ์„ฑ์— ๋Œ€ํ•œ ๋„์ „์„ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์šฐ๋ฆฌ์˜ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋Š” LLM์ด ์ ์šฉ๋˜๊ฑฐ๋‚˜ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ ๋น„๊ต๋  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋“  ์‹œ๋‚˜๋ฆฌ์˜ค๋‚˜ ๋„๋ฉ”์ธ์„ ๋Œ€ํ‘œํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์ œํ•œ ์‚ฌํ•ญ**. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ž๋™ GPT-4 ๋ฐ ์ธ๊ฐ„ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์˜ ํ•œ๊ณ„๋ฅผ ์ธ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ํ™•์žฅ์„ฑ๊ณผ ์‹ ๋ขฐ์„ฑ์— ์–ด๋ ค์›€์„ ์ดˆ๋ž˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ…Œ์ŠคํŠธ ์„ธํŠธ๊ฐ€ LLM์„ ์ ์šฉํ•˜๊ฑฐ๋‚˜ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๊ณผ ๋น„๊ตํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋“  ์‹œ๋‚˜๋ฆฌ์˜ค๋‚˜ ๋„๋ฉ”์ธ์„ ๋Œ€ํ‘œํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Fifth Response Q: Janet's ducks lay 16 eggs per day. She eats three for breakfast every morning and bakes muffins for her friends every day with four. She sells the remainder at the farmers' market daily for $2 per fresh duck egg. How much in dollars does she make every day at the farmers' market?
๋‹ค์„ฏ ๋ฒˆ์งธ ์‘๋‹ต Q: Janet์˜ ์˜ค๋ฆฌ๋“ค์€ ํ•˜๋ฃจ์— 16๊ฐœ์˜ ์•Œ์„ ๋‚ณ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๋งค์ผ ์•„์นจ ์„ธ ๊ฐœ๋ฅผ ์•„์นจ ์‹์‚ฌ๋กœ ๋จน๊ณ , ๋„ค ๊ฐœ๋กœ ์นœ๊ตฌ๋“ค์„ ์œ„ํ•ด ๋จธํ•€์„ ๊ตฝ์Šต๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€๋Š” ๋งค์ผ ๋†์‚ฐ๋ฌผ ์‹œ์žฅ์—์„œ ์‹ ์„ ํ•œ ์˜ค๋ฆฌ ์•Œ ํ•˜๋‚˜๋‹น 2๋‹ฌ๋Ÿฌ์— ํŒ๋งคํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๋งค์ผ ๋†์‚ฐ๋ฌผ ์‹œ์žฅ์—์„œ ์–ผ๋งˆ๋ฅผ ๋ฒŒ๊นŒ์š”?
๋‹ค์„ฏ ๋ฒˆ์งธ ์‘๋‹ต Q: ์ž๋„ท์˜ ์˜ค๋ฆฌ๋Š” ํ•˜๋ฃจ์— 16๊ฐœ์˜ ์•Œ์„ ๋‚ณ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๋งค์ผ ์•„์นจ 3๊ฐœ๋ฅผ ๋จน๊ณ , ์นœ๊ตฌ๋“ค์„ ์œ„ํ•ด ๋งค์ผ 4๊ฐœ๋กœ ๋จธํ•€์„ ๊ตฝ์Šต๋‹ˆ๋‹ค. ๋‚˜๋จธ์ง€๋Š” ์‹ ์„ ํ•œ ์˜ค๋ฆฌ์•Œ ํ•œ ๊ฐœ๋‹น 2๋‹ฌ๋Ÿฌ์— ๋งค์ผ ํŒŒ๋จธ์Šค ๋งˆ์ผ“์—์„œ ํŒ๋งคํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ํŒŒ๋จธ์Šค ๋งˆ์ผ“์—์„œ ๋งค์ผ ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ๋‹ฌ๋Ÿฌ๋ฅผ ๋ฒŒ๊นŒ์š”?
Table 24 displays examples where the models' generation was influenced by information in the front and back matter. Table 25 highlights comments indicating that models may sometimes overly focus on the latter parts of the stories. Lastly, Table 4 shares annotators' feedback on the annotation task.
ํ‘œ 24๋Š” ๋ชจ๋ธ์˜ ์ƒ์„ฑ์ด ์•ž๋ถ€๋ถ„๊ณผ ๋’ท๋ถ€๋ถ„์˜ ์ •๋ณด์— ์˜ํ•ด ์˜ํ–ฅ์„ ๋ฐ›์€ ์˜ˆ์‹œ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 25๋Š” ๋ชจ๋ธ์ด ๋•Œ๋•Œ๋กœ ์ด์•ผ๊ธฐ์˜ ํ›„๋ฐ˜๋ถ€์— ์ง€๋‚˜์น˜๊ฒŒ ์ง‘์ค‘ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๋Œ“๊ธ€์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ํ‘œ 4๋Š” ์ฃผ์„ ์ž‘์—…์— ๋Œ€ํ•œ ์ฃผ์„์ž๋“ค์˜ ํ”ผ๋“œ๋ฐฑ์„ ๊ณต์œ ํ•ฉ๋‹ˆ๋‹ค.
ํ‘œ 24๋Š” ์ „๋ฌธ ๋ฐ ํ›„๋ฌธ์˜ ์ •๋ณด์— ์˜ํ•ด ๋ชจ๋ธ ์ƒ์„ฑ์ด ์˜ํ–ฅ์„ ๋ฐ›์€ ์‚ฌ๋ก€๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 25๋Š” ๋ชจ๋ธ์ด ๋•Œ๋•Œ๋กœ ์Šคํ† ๋ฆฌ์˜ ํ›„๋ฐ˜๋ถ€์— ์ง€๋‚˜์น˜๊ฒŒ ์ง‘์ค‘ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋‚ด๋Š” ์˜๊ฒฌ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ํ‘œ 4๋Š” ์ฃผ์„ ์ž‘์—…์— ๋Œ€ํ•œ ์ฃผ์„ ์ž‘์„ฑ์ž์˜ ํ”ผ๋“œ๋ฐฑ์„ ๊ณต์œ ํ•ฉ๋‹ˆ๋‹ค.
where \(\theta_{CP}\) represents the model parameters, \(\mathcal{D}_{\mathcal{CP}}\) stands for the data used in continual pre-training, S represents the length of the input token sequence, and \(x_{i}\) represents the token to be predicted, while \(x_{0},x_{1},...,x_{i-1}\) make up the context.
\(\theta_{CP}\)๋Š” ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋‚˜ํƒ€๋‚ด๊ณ , \(\mathcal{D}_{\mathcal{CP}}\)๋Š” ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์— ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์˜๋ฏธํ•˜๋ฉฐ, S๋Š” ์ž…๋ ฅ ํ† ํฐ ์‹œํ€€์Šค์˜ ๊ธธ์ด๋ฅผ ๋‚˜ํƒ€๋‚ด๊ณ , \(x_{i}\)๋Š” ์˜ˆ์ธกํ•  ํ† ํฐ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, \(x_{0}, x_{1}, ..., x_{i-1}\)๋Š” ๋ฌธ๋งฅ์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(\theta_{CP}\)๋Š” ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋‚˜ํƒ€๋‚ด๊ณ , \(\mathcal{D}_{\mathcal{CP}}\)๋Š” ์—ฐ์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ๋‚˜ํƒ€๋‚ด๋ฉฐ, S๋Š” ์ž…๋ ฅ ํ† ํฐ ์‹œํ€€์Šค์˜ ๊ธธ์ด๋ฅผ ๋‚˜ํƒ€๋‚ด๊ณ , \(x_{i}\)๋Š” ์˜ˆ์ธกํ•  ํ† ํฐ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, \(x_{0},x_{1},...,x_{i-1}\)๋Š” ๋ฌธ๋งฅ์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
**Alpaca**: Sophie had always been fascinated by the mysteries of the universe. So, when she discovered a strange object in her backyard one afternoon, she couldn't help but feel excited. It was a small, metallic sphere, about the size of a softball, with strange markings etched onto its surface.
**Alpaca**: ์†Œํ”ผ๋Š” ํ•ญ์ƒ ์šฐ์ฃผ์˜ ์‹ ๋น„์— ๋งค๋ฃŒ๋˜์–ด ์žˆ์—ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์–ด๋Š ๋‚  ์˜คํ›„ ๊ทธ๋…€์˜ ๋’ท๋งˆ๋‹น์—์„œ ์ด์ƒํ•œ ๋ฌผ์ฒด๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์„ ๋•Œ, ๊ทธ๋…€๋Š” ํฅ๋ถ„์„ ๊ฐ์ถœ ์ˆ˜ ์—†์—ˆ๋‹ค. ๊ทธ๊ฒƒ์€ ์†Œํ”„ํŠธ๋ณผ ํฌ๊ธฐ์˜ ์ž‘์€ ๊ธˆ์† ๊ตฌ์ฒด๋กœ, ํ‘œ๋ฉด์— ์ด์ƒํ•œ ํ‘œ์‹์ด ์ƒˆ๊ฒจ์ ธ ์žˆ์—ˆ๋‹ค.
**์•ŒํŒŒ์นด**: ์†Œํ”ผ๋Š” ํ•ญ์ƒ ์šฐ์ฃผ์˜ ์‹ ๋น„์— ๋งค๋ฃŒ๋˜์–ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์–ด๋Š ๋‚  ์˜คํ›„ ๋’ท๋งˆ๋‹น์—์„œ ์ด์ƒํ•œ ๋ฌผ์ฒด๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์„ ๋•Œ ๊ทธ๋…€๋Š” ํฅ๋ถ„์„ ๊ฐ์ถœ ์ˆ˜ ์—†์—ˆ์Šต๋‹ˆ๋‹ค. ์†Œํ”„ํŠธ๋ณผ ์ •๋„ ํฌ๊ธฐ์˜ ์ž‘์€ ๊ธˆ์† ๊ตฌ์ฒด๋กœ ํ‘œ๋ฉด์— ์ด์ƒํ•œ ๋ฌธ์–‘์ด ์ƒˆ๊ฒจ์ ธ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
One day, they decided to research the object they had discovered. They scoured the internet, searching for any information about it. It was then that they discovered the true nature of the object - it was an ancient relic from a long-forgotten civilization, imbued with powerful, alien technology.
์–ด๋Š ๋‚ , ๊ทธ๋“ค์€ ๋ฐœ๊ฒฌํ•œ ๋ฌผ์ฒด์— ๋Œ€ํ•ด ์—ฐ๊ตฌํ•˜๊ธฐ๋กœ ๊ฒฐ์‹ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ์ธํ„ฐ๋„ท์„ ์ƒ…์ƒ…์ด ๋’ค์ ธ ๊ทธ์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ฐพ๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋˜ ์ค‘ ๊ทธ๋“ค์€ ๊ทธ ๋ฌผ์ฒด์˜ ์ง„์ •ํ•œ ๋ณธ์งˆ์„ ๋ฐœ๊ฒฌํ•˜๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๊ฒƒ์€ ์˜ค๋ž˜์ „์— ์žŠํ˜€์ง„ ๋ฌธ๋ช…์˜ ๊ณ ๋Œ€ ์œ ๋ฌผ๋กœ, ๊ฐ•๋ ฅํ•œ ์™ธ๊ณ„ ๊ธฐ์ˆ ์ด ๋‹ด๊ฒจ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
์–ด๋Š ๋‚  ๊ทธ๋“ค์€ ์ž์‹ ๋“ค์ด ๋ฐœ๊ฒฌํ•œ ๋ฌผ์ฒด์— ๋Œ€ํ•ด ์—ฐ๊ตฌํ•˜๊ธฐ๋กœ ๊ฒฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ์ธํ„ฐ๋„ท์„ ๋’ค์ ธ ๊ทธ ๋ฌผ์ฒด์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ฐพ์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋˜ ์ค‘ ๊ทธ๋“ค์€ ๊ทธ ๋ฌผ์ฒด์˜ ์ง„์ •ํ•œ ๋ณธ์งˆ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ค๋ž˜ ์ „์— ์žŠํ˜€์ง„ ๋ฌธ๋ช…์˜ ๊ณ ๋Œ€ ์œ ๋ฌผ๋กœ, ๊ฐ•๋ ฅํ•œ ์™ธ๊ณ„ ๊ธฐ์ˆ ์ด ๋‹ด๊ธด ๊ฒƒ์ด์—ˆ์Šต๋‹ˆ๋‹ค.
Purpose of annotationWe are annotating 2,000 responses from a range of AI models used for chat. This is part of the ML Commons' Working Group on AI Safety efforts to benchmark the safety risks of AI models. You can find out more about our work HERE, and reach out to Bertie if you have questions.
์ฃผ์„์˜ ๋ชฉ์ : ์šฐ๋ฆฌ๋Š” ์ฑ„ํŒ…์— ์‚ฌ์šฉ๋˜๋Š” ๋‹ค์–‘ํ•œ AI ๋ชจ๋ธ์˜ 2,000๊ฐœ์˜ ์‘๋‹ต์„ ์ฃผ์„ ์ฒ˜๋ฆฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AI ๋ชจ๋ธ์˜ ์•ˆ์ „์„ฑ ์œ„ํ—˜์„ ๋ฒค์น˜๋งˆํ‚นํ•˜๊ธฐ ์œ„ํ•œ ML Commons์˜ AI ์•ˆ์ „์„ฑ ์ž‘์—… ๊ทธ๋ฃน์˜ ๋…ธ๋ ฅ์˜ ์ผํ™˜์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ž‘์—…์— ๋Œ€ํ•ด ๋” ์•Œ๊ณ  ์‹ถ๋‹ค๋ฉด ์—ฌ๊ธฐ๋ฅผ ํด๋ฆญํ•˜๊ณ , ์งˆ๋ฌธ์ด ์žˆ์œผ๋ฉด Bertie์—๊ฒŒ ์—ฐ๋ฝํ•˜์‹ญ์‹œ์˜ค.
์ฃผ์„ ๋‹ฌ๊ธฐ ๋ชฉ์ ์ฑ„ํŒ…์— ์‚ฌ์šฉ๋˜๋Š” ๋‹ค์–‘ํ•œ AI ๋ชจ๋ธ์˜ ์‘๋‹ต 2,000๊ฑด์— ์ฃผ์„์„ ๋‹ฌ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AI ๋ชจ๋ธ์˜ ์•ˆ์ „์„ฑ ์œ„ํ—˜์„ ๋ฒค์น˜๋งˆํ‚นํ•˜๊ธฐ ์œ„ํ•œ ML ์ปค๋จผ์ฆˆ์˜ AI ์•ˆ์ „ ๋…ธ๋ ฅ์— ๋Œ€ํ•œ ์ผํ™˜์ž…๋‹ˆ๋‹ค. ์ €ํฌ์˜ ํ™œ๋™์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์•Œ์•„๋ณด์‹œ๊ณ  ๊ถ๊ธˆํ•œ ์ ์ด ์žˆ์œผ๋ฉด Bertie์—๊ฒŒ ๋ฌธ์˜ํ•ด ์ฃผ์„ธ์š”.
Prior works developing models for clinical dialogue have focused on metrics such as the accuracy of note-to-dialogue or dialogue-to-note generations [26, 27], or natural language generation metrics such as BLEU or ROUGE scores that fail to capture the clinical quality of a consultation [28, 29].
์ด์ „์— ์ž„์ƒ ๋Œ€ํ™”๋ฅผ ์œ„ํ•œ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•œ ์—ฐ๊ตฌ๋“ค์€ ๋…ธํŠธ-๋Œ€ํ™” ๋˜๋Š” ๋Œ€ํ™”-๋…ธํŠธ ์ƒ์„ฑ์˜ ์ •ํ™•๋„ [26, 27]์™€ ๊ฐ™์€ ์ง€ํ‘œ๋‚˜, ์ƒ๋‹ด์˜ ์ž„์ƒ์  ํ’ˆ์งˆ์„ ํฌ์ฐฉํ•˜์ง€ ๋ชปํ•˜๋Š” BLEU ๋˜๋Š” ROUGE ์ ์ˆ˜์™€ ๊ฐ™์€ ์ž์—ฐ์–ด ์ƒ์„ฑ ์ง€ํ‘œ [28, 29]์— ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค.
์ž„์ƒ ๋Œ€ํ™”๋ฅผ ์œ„ํ•œ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๋Š” ์„ ํ–‰ ์—ฐ๊ตฌ๋“ค์€ ๋…ธํŠธ-๋Œ€ํ™” ๋˜๋Š” ๋Œ€ํ™”-๋Œ€๋…ธํŠธ ์ƒ์„ฑ์˜ ์ •ํ™•๋„์™€ ๊ฐ™์€ ๋ฉ”ํŠธ๋ฆญ[26, 27] ๋˜๋Š” ์ƒ๋‹ด์˜ ์ž„์ƒ ํ’ˆ์งˆ์„ ํฌ์ฐฉํ•˜์ง€ ๋ชปํ•˜๋Š” BLEU ๋˜๋Š” ROUGE ์ ์ˆ˜์™€ ๊ฐ™์€ ์ž์—ฐ์–ด ์ƒ์„ฑ ๋ฉ”ํŠธ๋ฆญ์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ๋‹ค[28, 29].
A: Let's think step by step. How many fresh eggs remains after the three her eat for breakfast and four goes into the muffin mix for her friends. We can calculate it by 16-3-4. The amount equals 9 fresh eggs. So, the dollars it makes daily equals what multiplied by $2. 12 x 2. The answer is $24.
A: ๋‹จ๊ณ„๋ณ„๋กœ ์ƒ๊ฐํ•ด ๋ด…์‹œ๋‹ค. ๊ทธ๋…€๊ฐ€ ์•„์นจ ์‹์‚ฌ๋กœ ์„ธ ๊ฐœ๋ฅผ ๋จน๊ณ  ์นœ๊ตฌ๋“ค์„ ์œ„ํ•ด ๋จธํ•€ ๋ฏน์Šค์— ๋„ค ๊ฐœ๋ฅผ ๋„ฃ์€ ํ›„ ๋‚จ์€ ์‹ ์„ ํ•œ ๋‹ฌ๊ฑ€์€ ๋ช‡ ๊ฐœ์ผ๊นŒ์š”? 16-3-4๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์–‘์€ 9๊ฐœ์˜ ์‹ ์„ ํ•œ ๋‹ฌ๊ฑ€์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ•˜๋ฃจ์— ๋ฒŒ์–ด๋“ค์ด๋Š” ๊ธˆ์•ก์€ $2๋ฅผ ๊ณฑํ•œ ๊ฐ’๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค. 12 x 2. ์ •๋‹ต์€ $24์ž…๋‹ˆ๋‹ค.
A: ์ฐจ๊ทผ์ฐจ๊ทผ ์ƒ๊ฐํ•ด ๋ด…์‹œ๋‹ค. ์„ธ ๋งˆ๋ฆฌ๊ฐ€ ์•„์นจ ์‹์‚ฌ๋กœ ๋จน๊ณ  ๋„ค ๋งˆ๋ฆฌ๊ฐ€ ์นœ๊ตฌ๋“ค์„ ์œ„ํ•œ ๋จธํ•€ ๋ฏน์Šค์— ๋“ค์–ด๊ฐ„ ํ›„ ๋‚จ์€ ์‹ ์„ ํ•œ ๋‹ฌ๊ฑ€์€ ๋ช‡ ๊ฐœ์ผ๊นŒ์š”? 16-3-4๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์–‘์€ ์‹ ์„ ํ•œ ๋‹ฌ๊ฑ€ 9๊ฐœ์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋งค์ผ ๋ฒŒ์–ด๋“ค์ด๋Š” ๋‹ฌ๋Ÿฌ๋Š” 2๋ฅผ ๊ณฑํ•œ ๊ฐ’๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค. 12 x 2. ์ •๋‹ต์€ $24์ž…๋‹ˆ๋‹ค.
Figure 1: SVD can recover the hidden dimensionality of a model when the final output layer dimension is greater than the hidden dimension. Here we extract the hidden dimension (2048) of the Pythia 1.4B model. We can precisely identify the size by obtaining slightly over 2048 full logit vectors.
Figure 1: SVD๋Š” ์ตœ์ข… ์ถœ๋ ฅ ๋ ˆ์ด์–ด์˜ ์ฐจ์›์ด ์ˆจ๊ฒจ์ง„ ์ฐจ์›๋ณด๋‹ค ํด ๋•Œ ๋ชจ๋ธ์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์›์„ ๋ณต์›ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์šฐ๋ฆฌ๋Š” Pythia 1.4B ๋ชจ๋ธ์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์›(2048)์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ์•ฝ๊ฐ„ ๋„˜๋Š” 2048๊ฐœ์˜ ์ „์ฒด ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ์–ป์Œ์œผ๋กœ์จ ํฌ๊ธฐ๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1: SVD๋Š” ์ตœ์ข… ์ถœ๋ ฅ ๋ ˆ์ด์–ด ์ฐจ์›์ด ์ˆจ๊ฒจ์ง„ ์ฐจ์›๋ณด๋‹ค ํด ๋•Œ ๋ชจ๋ธ์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์›์„ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” Pythia 1.4B ๋ชจ๋ธ์˜ ์ˆจ๊ฒจ์ง„ ์ฐจ์›(2048)์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ์•ฝ๊ฐ„ 2048๊ฐœ์˜ ์ „์ฒด ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ์–ป์œผ๋ฉด ์ •ํ™•ํ•˜๊ฒŒ ํฌ๊ธฐ๋ฅผ ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In addition, we further used the GPT-4 Turbo to find the position where the first step in our answer was wrong and normalized that position by the total number of steps in each answer. As the estimated single-step accuracy gets higher, the first error position of the normalization is postponed.
์ถ”๊ฐ€์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” GPT-4 Turbo๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์šฐ๋ฆฌ์˜ ๋‹ต๋ณ€์—์„œ ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„๊ฐ€ ์ž˜๋ชป๋œ ์œ„์น˜๋ฅผ ์ฐพ์•„๋‚ด๊ณ , ๊ฐ ๋‹ต๋ณ€์˜ ์ด ๋‹จ๊ณ„ ์ˆ˜๋กœ ๊ทธ ์œ„์น˜๋ฅผ ์ •๊ทœํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ถ”์ •๋œ ๋‹จ์ผ ๋‹จ๊ณ„ ์ •ํ™•๋„๊ฐ€ ๋†’์•„์งˆ์ˆ˜๋ก, ์ •๊ทœํ™”๋œ ์ฒซ ๋ฒˆ์งธ ์˜ค๋ฅ˜ ์œ„์น˜๋Š” ๋’ค๋กœ ๋ฏธ๋ค„์ง‘๋‹ˆ๋‹ค.
๋˜ํ•œ, GPT-4 ํ„ฐ๋ณด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ต๋ณ€์˜ ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„๊ฐ€ ์ž˜๋ชป๋œ ์œ„์น˜๋ฅผ ์ฐพ๊ณ  ๊ฐ ๋‹ต๋ณ€์˜ ์ด ๋‹จ๊ณ„ ์ˆ˜๋กœ ํ•ด๋‹น ์œ„์น˜๋ฅผ ์ •๊ทœํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ์ƒ ๋‹จ์ผ ๋‹จ๊ณ„ ์ •ํ™•๋„๊ฐ€ ๋†’์•„์งˆ์ˆ˜๋ก ์ •๊ทœํ™”์˜ ์ฒซ ๋ฒˆ์งธ ์˜ค๋ฅ˜ ์œ„์น˜๊ฐ€ ์—ฐ๊ธฐ๋ฉ๋‹ˆ๋‹ค.
Figure 2: The pass@1 performance of main LLM models on the GSM8k benchmark, our model is currently ranked in the top five, slightly outperforming some close-source models such as ChatGPT-3.5\({}^{5}\), Claude Instant-1\({}^{6}\), PaLM 2 [44], and substantially surpassing all open-source models.
Figure 2: GSM8k ๋ฒค์น˜๋งˆํฌ์—์„œ ์ฃผ์š” LLM ๋ชจ๋ธ๋“ค์˜ pass@1 ์„ฑ๋Šฅ์„ ๋‚˜ํƒ€๋‚ธ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ํ˜„์žฌ ์ƒ์œ„ ๋‹ค์„ฏ ๊ฐœ ๋ชจ๋ธ์— ์†ํ•˜๋ฉฐ, ChatGPT-3.5\({}^{5}\), Claude Instant-1\({}^{6}\), PaLM 2 [44]์™€ ๊ฐ™์€ ์ผ๋ถ€ ๋น„๊ณต๊ฐœ ๋ชจ๋ธ์„ ์•ฝ๊ฐ„ ๋Šฅ๊ฐ€ํ•˜๊ณ , ๋ชจ๋“  ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์„ ํฌ๊ฒŒ ์•ž์„œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2: GSM8k ๋ฒค์น˜๋งˆํฌ์—์„œ ์ฃผ์š” LLM ๋ชจ๋ธ์˜ pass@1 ์„ฑ๋Šฅ, ํ˜„์žฌ ์šฐ๋ฆฌ ๋ชจ๋ธ์€ ์ƒ์œ„ 5์œ„๊ถŒ์— ์ง„์ž…ํ•˜์—ฌ ChatGPT-3.5\({}^{5}\), Claude Instant-1\({}^{6}\), PaLM 2 [44]์™€ ๊ฐ™์€ ์ผ๋ถ€ ๋น„๊ณต๊ฐœ ๋ชจ๋ธ์„ ์•ฝ๊ฐ„ ์•ž์„ฐ๊ณ , ๋ชจ๋“  ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์„ ํฌ๊ฒŒ ๋Šฅ๊ฐ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Schoenegger and Park (2023); Abolhasemi et al. (2023) evaluated GPT-4 and other LLMs on forecasting tournaments and found that they underperform the human crowd. This observation is in line with ours in Section 3.4. Unlike us, they make little or no efforts to improve these LMs on forecasting.
Schoenegger์™€ Park (2023); Abolhasemi ๋“ฑ (2023)์€ ์˜ˆ์ธก ํ† ๋„ˆ๋จผํŠธ์—์„œ GPT-4 ๋ฐ ๊ธฐํƒ€ LLM์„ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, ์ด๋“ค์ด ์ธ๊ฐ„ ๊ตฐ์ค‘๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง„๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ์€ 3.4์ ˆ์—์„œ ์šฐ๋ฆฌ์˜ ๊ด€์ฐฐ๊ณผ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์™€ ๋‹ฌ๋ฆฌ, ๊ทธ๋“ค์€ ์ด๋Ÿฌํ•œ LLM์„ ์˜ˆ์ธก์—์„œ ๊ฐœ์„ ํ•˜๋ ค๋Š” ๋…ธ๋ ฅ์„ ๊ฑฐ์˜ ๋˜๋Š” ์ „ํ˜€ ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
Schoenegger์™€ Park(2023); Abolhasemi ๋“ฑ(2023)์€ ์˜ˆ์ธก ํ† ๋„ˆ๋จผํŠธ์—์„œ GPT-4 ๋ฐ ๊ธฐํƒ€ LLM์„ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, ์ธ๊ฐ„ ๊ตฐ์ค‘๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ์„น์…˜ 3.4์˜ ์šฐ๋ฆฌ์˜ ๊ด€์ฐฐ ๊ฒฐ๊ณผ์™€ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์™€ ๋‹ฌ๋ฆฌ ์ด๋“ค์€ ์˜ˆ์ธก์— ๋Œ€ํ•œ ์ด๋Ÿฌํ•œ LM์„ ๊ฐœ์„ ํ•˜๋ ค๋Š” ๋…ธ๋ ฅ์„ ๊ฑฐ์˜ ๋˜๋Š” ์ „ํ˜€ ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
where \(\texttt{Mask}[k]\) is a vector of length \(d_{inter}\) with the \(k\)-th element as \(1\) and others as \(0\). For calculating \(\text{Imp}(W_{up}[:,k]|x)\) for all neurons in \(W_{up}\) parallelly, we introduce a diagonal mask matrix of size \((d_{inter},d_{inter})\), denoted as Mask.
\(\texttt{Mask}[k]\)๋Š” \(k\)-๋ฒˆ์งธ ์š”์†Œ๊ฐ€ \(1\)์ด๊ณ  ๋‚˜๋จธ์ง€๊ฐ€ \(0\)์ธ ๊ธธ์ด \(d_{inter}\)์˜ ๋ฒกํ„ฐ์ž…๋‹ˆ๋‹ค. \(W_{up}\)์˜ ๋ชจ๋“  ๋‰ด๋Ÿฐ์— ๋Œ€ํ•ด \(\text{Imp}(W_{up}[:,k]|x)\)๋ฅผ ๋ณ‘๋ ฌ๋กœ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด, ํฌ๊ธฐ๊ฐ€ \((d_{inter},d_{inter})\)์ธ ๋Œ€๊ฐ ๋งˆ์Šคํฌ ํ–‰๋ ฌ์„ ๋„์ž…ํ•˜๋ฉฐ, ์ด๋ฅผ Mask๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(\texttt{Mask}[k]\)๋Š” ๊ธธ์ด \(d_{inter}\)์˜ ๋ฒกํ„ฐ๋กœ, \(k\)๋ฒˆ์งธ ์š”์†Œ๋Š” \(1\)์ด๊ณ  ๋‚˜๋จธ์ง€๋Š” \(0\)์ž…๋‹ˆ๋‹ค. W_{up}\)์˜ ๋ชจ๋“  ๋‰ด๋Ÿฐ์— ๋Œ€ํ•ด \(\text{Imp}(W_{up}[:,k]|x)\)๋ฅผ ๋ณ‘๋ ฌ๋กœ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด, Mask๋กœ ํ‘œ์‹œ๋˜๋Š” ํฌ๊ธฐ \((d_{inter},d_{inter})\)์˜ ๋Œ€๊ฐ์„  ๋งˆ์Šคํฌ ํ–‰๋ ฌ์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค.
The idea is that given an auxiliary knowledge base and an input query, we use the RAG architecture to find documents within the knowledge base that resemble the input query. These documents are then added to the input query, thus giving the model further context about the subject of the query.
์•„์ด๋””์–ด๋Š” ๋ณด์กฐ ์ง€์‹ ๋ฒ ์ด์Šค์™€ ์ž…๋ ฅ ์ฟผ๋ฆฌ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, RAG ์•„ํ‚คํ…์ฒ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ž…๋ ฅ ์ฟผ๋ฆฌ์™€ ์œ ์‚ฌํ•œ ๋ฌธ์„œ๋ฅผ ์ง€์‹ ๋ฒ ์ด์Šค ๋‚ด์—์„œ ์ฐพ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ๋ฌธ์„œ๋“ค์„ ์ž…๋ ฅ ์ฟผ๋ฆฌ์— ์ถ”๊ฐ€ํ•˜์—ฌ ๋ชจ๋ธ์—๊ฒŒ ์ฟผ๋ฆฌ ์ฃผ์ œ์— ๋Œ€ํ•œ ์ถ”๊ฐ€์ ์ธ ๋ฌธ๋งฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์ด ์•„์ด๋””์–ด๋Š” ๋ณด์กฐ ์ง€์‹ ๊ธฐ๋ฐ˜๊ณผ ์ž…๋ ฅ ์ฟผ๋ฆฌ๊ฐ€ ์ฃผ์–ด์ง€๋ฉด RAG ์•„ํ‚คํ…์ฒ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ง€์‹ ๊ธฐ๋ฐ˜ ๋‚ด์—์„œ ์ž…๋ ฅ ์ฟผ๋ฆฌ์™€ ์œ ์‚ฌํ•œ ๋ฌธ์„œ๋ฅผ ์ฐพ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ๋ฌธ์„œ๋ฅผ ์ž…๋ ฅ ์ฟผ๋ฆฌ์— ์ถ”๊ฐ€ํ•˜์—ฌ ๋ชจ๋ธ์— ์ฟผ๋ฆฌ ์ฃผ์ œ์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
Manual CoT is more effective than zero-shot. However, the effectiveness of this example-based CoT depends on the choice of diverse examples, and constructing prompts with such examples of step by step reasoning by hand is hard and error prone. That is where automatic CoT [157] comes into play.
์ˆ˜๋™ CoT๋Š” zero-shot๋ณด๋‹ค ๋” ํšจ๊ณผ์ ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์˜ˆ์ œ ๊ธฐ๋ฐ˜ CoT์˜ ํšจ๊ณผ๋Š” ๋‹ค์–‘ํ•œ ์˜ˆ์ œ์˜ ์„ ํƒ์— ๋‹ฌ๋ ค ์žˆ์œผ๋ฉฐ, ์ด๋Ÿฌํ•œ ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก  ์˜ˆ์ œ๋ฅผ ์ˆ˜์ž‘์—…์œผ๋กœ ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต๊ณ  ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ•˜๊ธฐ ์‰ฝ์Šต๋‹ˆ๋‹ค. ๋ฐ”๋กœ ์ด ์ ์—์„œ ์ž๋™ CoT [157]๊ฐ€ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค.
์ˆ˜๋™ CoT๋Š” ์ œ๋กœ ์ƒท๋ณด๋‹ค ๋” ํšจ๊ณผ์ ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์˜ˆ์ œ ๊ธฐ๋ฐ˜ CoT์˜ ํšจ๊ณผ๋Š” ๋‹ค์–‘ํ•œ ์˜ˆ์ œ์˜ ์„ ํƒ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๋ฉฐ, ์ด๋Ÿฌํ•œ ์˜ˆ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก ์„ ํ”„๋กฌํ”„ํŠธ๋กœ ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์€ ์ˆ˜์ž‘์—…์œผ๋กœ๋Š” ์–ด๋ ต๊ณ  ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ•˜๊ธฐ ์‰ฝ์Šต๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด ๋ฐ”๋กœ ์ž๋™ CoT [157]์˜ ์žฅ์ ์ž…๋‹ˆ๋‹ค.
This ensures that the loss is applied only to the tokens that are deemed most beneficial for the language model to learn from. In practice, token selection can be implemented by ranking the tokens in a batch according to their excess loss and using only the top \(k\%\) of tokens for training.
์ด ๋ฐฉ๋ฒ•์€ ์–ธ์–ด ๋ชจ๋ธ์ด ํ•™์Šตํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์œ ์ตํ•œ ํ† ํฐ์—๋งŒ ์†์‹ค์ด ์ ์šฉ๋˜๋„๋ก ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ, ํ† ํฐ ์„ ํƒ์€ ๋ฐฐ์น˜ ๋‚ด ํ† ํฐ์„ ์ดˆ๊ณผ ์†์‹ค์— ๋”ฐ๋ผ ์ˆœ์œ„ ๋งค๊ธด ํ›„ ์ƒ์œ„ \(k\%\)์˜ ํ† ํฐ๋งŒ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•จ์œผ๋กœ์จ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์–ธ์–ด ๋ชจ๋ธ์ด ํ•™์Šตํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์œ ์ตํ•˜๋‹ค๊ณ  ๊ฐ„์ฃผ๋˜๋Š” ํ† ํฐ์—๋งŒ ์†์‹ค์ด ์ ์šฉ๋ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ํ† ํฐ ์„ ํƒ์€ ์ดˆ๊ณผ ์†์‹ค์— ๋”ฐ๋ผ ๋ฐฐ์น˜์˜ ํ† ํฐ์„ ์ˆœ์œ„ํ™”ํ•˜๊ณ  ์ƒ์œ„ \(k\%\)์˜ ํ† ํฐ๋งŒ ํ•™์Šต์— ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Footnote 16: As demonstrated by [3], in low-diversity datasets like \(\mathsf{bioS}^{\mathsf{simple}}\), knowledge can be word-by-word memorized but is nearly 0% extractable for downstream tasks. Others discover that rewriting data can improve the reversal extractability of knowledge [4, 11].
Footnote 16: [3]์— ์˜ํ•ด ์ž…์ฆ๋œ ๋ฐ”์™€ ๊ฐ™์ด, \(\mathsf{bioS}^{\mathsf{simple}}\)์™€ ๊ฐ™์€ ๋‹ค์–‘์„ฑ์ด ๋‚ฎ์€ ๋ฐ์ดํ„ฐ์…‹์—์„œ๋Š” ์ง€์‹์„ ๋‹จ์–ด ๋‹จ์œ„๋กœ ์•”๊ธฐํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ์œ„ํ•ด ์ถ”์ถœํ•  ์ˆ˜ ์žˆ๋Š” ๋น„์œจ์€ ๊ฑฐ์˜ 0%์— ๊ฐ€๊น์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์žฌ์ž‘์„ฑํ•˜๋ฉด ์ง€์‹์˜ ์—ญ์ถ”์ถœ ๊ฐ€๋Šฅ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค [4, 11].
์ฃผ 16: [3]์—์„œ ๋ณด์—ฌ์ง„ ๊ฒƒ์ฒ˜๋Ÿผ, \(\mathsf{bioS}^{\mathsf{simple}}\)์™€ ๊ฐ™์ด ๋‹ค์–‘์„ฑ์ด ๋‚ฎ์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋Š” ์ง€์‹์„ ๋‹จ์–ด ๋‹จ์œ„๋กœ ์•”๊ธฐํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ๊ฑฐ์˜ 0%์˜ ์ถ”์ถœ ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์‚ฌ๋žŒ๋“ค์€ ๋ฐ์ดํ„ฐ๋ฅผ ๋‹ค์‹œ ์ž‘์„ฑํ•˜๋ฉด ์ง€์‹์˜ ์—ญ์ถ”์ถœ ๊ฐ€๋Šฅ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค[4, 11].
It is not an easy job to write this long survey and update its content with timely work. First of all, we would like to sincerely thank the support from the readers and our team members. We work very hard on this survey, and hope that it can present a comprehensive, timely reference for LLMs.
์ด ๊ธด ์„ค๋ฌธ ์กฐ์‚ฌ๋ฅผ ์ž‘์„ฑํ•˜๊ณ  ์‹œ์˜์ ์ ˆํ•œ ์ž‘์—…์œผ๋กœ ๋‚ด์šฉ์„ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๊ฒƒ์€ ์‰ฌ์šด ์ผ์ด ์•„๋‹™๋‹ˆ๋‹ค. ์šฐ์„ , ๋…์ž๋“ค๊ณผ ํŒ€์›๋“ค์˜ ์ง€์›์— ์ง„์‹ฌ์œผ๋กœ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด ์„ค๋ฌธ ์กฐ์‚ฌ์— ๋งค์šฐ ์—ด์‹ฌํžˆ ์ž„ํ–ˆ์œผ๋ฉฐ, ์ด๊ฒƒ์ด LLMs์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ด๊ณ  ์‹œ์˜์ ์ ˆํ•œ ์ฐธ๊ณ  ์ž๋ฃŒ๊ฐ€ ๋˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
์ด๋ ‡๊ฒŒ ๊ธด ์„ค๋ฌธ์กฐ์‚ฌ๋ฅผ ์ž‘์„ฑํ•˜๊ณ  ์ตœ์‹  ์ž‘์—…์œผ๋กœ ๋‚ด์šฉ์„ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๊ฒƒ์€ ์‰ฌ์šด ์ผ์ด ์•„๋‹™๋‹ˆ๋‹ค. ์šฐ์„  ๋…์ž ์—ฌ๋Ÿฌ๋ถ„๊ณผ ์ €ํฌ ํŒ€์›๋“ค์˜ ์„ฑ์›์— ์ง„์‹ฌ์œผ๋กœ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ด ์„ค๋ฌธ์กฐ์‚ฌ์— ๋งŽ์€ ๋…ธ๋ ฅ์„ ๊ธฐ์šธ์ด๊ณ  ์žˆ์œผ๋ฉฐ, ์ด ์„ค๋ฌธ์กฐ์‚ฌ๊ฐ€ LLM์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ด๊ณ  ์‹œ์˜์ ์ ˆํ•œ ์ฐธ๊ณ  ์ž๋ฃŒ๋ฅผ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
Our near perfect extraction has one exception: GPT-2 Small. On this 768 dimensional model our attack reports a hidden dimension of 757. In Appendix A we show that this "failure" is caused by GPT-2 actually having an effective hidden dimensionality of \(757\) despite having \(768\) dimensions.
์šฐ๋ฆฌ์˜ ๊ฑฐ์˜ ์™„๋ฒฝํ•œ ์ถ”์ถœ์—๋Š” ํ•œ ๊ฐ€์ง€ ์˜ˆ์™ธ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค: GPT-2 Small์ž…๋‹ˆ๋‹ค. ์ด 768 ์ฐจ์›์˜ ๋ชจ๋ธ์—์„œ ์šฐ๋ฆฌ์˜ ๊ณต๊ฒฉ์€ ์ˆจ๊ฒจ์ง„ ์ฐจ์›์ด 757์ด๋ผ๊ณ  ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๋ถ€๋ก A์—์„œ ์šฐ๋ฆฌ๋Š” ์ด "์‹คํŒจ"๊ฐ€ ์‹ค์ œ๋กœ GPT-2๊ฐ€ 768 ์ฐจ์›์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ํšจ๊ณผ์ ์ธ ์ˆจ๊ฒจ์ง„ ์ฐจ์›์ด \(757\)์ž„์— ์˜ํ•ด ๋ฐœ์ƒํ•œ ๊ฒƒ์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๊ฑฐ์˜ ์™„๋ฒฝํ•œ ์ถ”์ถœ์—๋Š” ํ•˜๋‚˜์˜ ์˜ˆ์™ธ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค: GPT-2 Small. ์ด 768์ฐจ์› ๋ชจ๋ธ์—์„œ ์šฐ๋ฆฌ์˜ ๊ณต๊ฒฉ์€ ์ˆจ๊ฒจ์ง„ ์ฐจ์›์ด 757์ด๋ผ๊ณ  ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ๋ถ€๋ก A์—์„œ๋Š” ์ด "์‹คํŒจ"๊ฐ€ ์‹ค์ œ๋กœ GPT-2๊ฐ€ 768์ฐจ์›์ž„์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  757์ฐจ์›์˜ ์œ ํšจ ์ˆจ๊ฒจ์ง„ ์ฐจ์›์„ ๊ฐ€์ง€๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฐœ์ƒํ•˜๋Š” ๊ฒƒ์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
In this case, knowing the geography of Hawaii, as well as immediate effects of wildfires, enables a model to give the first two options a lower likelihood. This process of elimination increases the probability of choosing one of the remaining options (the third option is the correct answer).
์ด ๊ฒฝ์šฐ, ํ•˜์™€์ด์˜ ์ง€๋ฆฌ์™€ ์‚ฐ๋ถˆ์˜ ์ฆ‰๊ฐ์ ์ธ ์˜ํ–ฅ์„ ์•Œ๊ณ  ์žˆ์œผ๋ฉด ๋ชจ๋ธ์ด ์ฒซ ๋‘ ๊ฐ€์ง€ ์„ ํƒ์ง€์˜ ๊ฐ€๋Šฅ์„ฑ์„ ๋‚ฎ๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ œ๊ฑฐ ๊ณผ์ •์€ ๋‚จ์€ ์„ ํƒ์ง€ ์ค‘ ํ•˜๋‚˜๋ฅผ ์„ ํƒํ•  ํ™•๋ฅ ์„ ๋†’์ž…๋‹ˆ๋‹ค (์„ธ ๋ฒˆ์งธ ์„ ํƒ์ง€๊ฐ€ ์ •๋‹ต์ž…๋‹ˆ๋‹ค).
์ด ๊ฒฝ์šฐ, ํ•˜์™€์ด์˜ ์ง€๋ฆฌ์™€ ์‚ฐ๋ถˆ์˜ ์ฆ‰๊ฐ์ ์ธ ์˜ํ–ฅ์„ ์•„๋Š” ๋ชจ๋ธ์€ ์ฒ˜์Œ ๋‘ ๊ฐ€์ง€ ์˜ต์…˜์˜ ๊ฐ€๋Šฅ์„ฑ์„ ๋‚ฎ๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ œ๊ฑฐ ๊ณผ์ •์„ ํ†ตํ•ด ๋‚˜๋จธ์ง€ ์˜ต์…˜ ์ค‘ ํ•˜๋‚˜๋ฅผ ์„ ํƒํ•  ํ™•๋ฅ ์ด ๋†’์•„์ง‘๋‹ˆ๋‹ค(์„ธ ๋ฒˆ์งธ ์˜ต์…˜์ด ์ •๋‹ต์ž…๋‹ˆ๋‹ค).
**Theorem 1**.: _When \(\pi_{\text{ref}}\) is defined as \(\pi_{w}\), an ideal policy that precisely aligns with the true data distribution of preferred data, the DPO loss \(\mathcal{L}(\pi_{\theta};\pi_{w})+C\) is upper bounded by \(\mathcal{L}(\pi_{\theta};U)\), where \(C\) is a constant._
**Theorem 1**.: _\(\pi_{\text{ref}}\)๊ฐ€ \(\pi_{w}\)๋กœ ์ •์˜๋  ๋•Œ, ์„ ํ˜ธํ•˜๋Š” ๋ฐ์ดํ„ฐ์˜ ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์™€ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋Š” ์ด์ƒ์ ์ธ ์ •์ฑ…์ธ ๊ฒฝ์šฐ, DPO ์†์‹ค \(\mathcal{L}(\pi_{\theta};\pi_{w})+C\)๋Š” \(\mathcal{L}(\pi_{\theta};U)\)๋กœ ์ƒํ•œ์ด ์ •ํ•ด์ง€๋ฉฐ, ์—ฌ๊ธฐ์„œ \(C\)๋Š” ์ƒ์ˆ˜์ž…๋‹ˆ๋‹ค._
**์ •๋ฆฌ 1**.: _ฯ€_{\text{ref}}\)๊ฐ€ ์„ ํ˜ธ ๋ฐ์ดํ„ฐ์˜ ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์™€ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๋Š” ์ด์ƒ์ ์ธ ์ •์ฑ…์ธ \(\pi_{w}\)๋กœ ์ •์˜๋  ๋•Œ, DPO ์†์‹ค \(\mathcal{L}(\pi_{\theta};\pi_{w})+C\)๋Š” \(\mathcal{L}(\pi_{\theta};U)\)๋กœ ์ƒํ•œ์ด ์ •์˜๋˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(C\)๋Š” ์ƒ์ˆ˜์ž…๋‹ˆ๋‹ค.
Although the disparity between training Adapt-LLM on NQ or SQuAD is relatively minor, we try to determine the suitability of a training set for a given evaluation set. While both training sets (NQ and SQuAD) and the evaluation set (PopQA) are based on Wikipedia, subtle differences may exist.
๋น„๋ก NQ ๋˜๋Š” SQuAD์—์„œ Adapt-LLM์„ ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๊ฒƒ์˜ ์ฐจ์ด๋Š” ์ƒ๋Œ€์ ์œผ๋กœ ๋ฏธ๋ฏธํ•˜์ง€๋งŒ, ์šฐ๋ฆฌ๋Š” ์ฃผ์–ด์ง„ ํ‰๊ฐ€ ์„ธํŠธ์— ๋Œ€ํ•œ ํ›ˆ๋ จ ์„ธํŠธ์˜ ์ ํ•ฉ์„ฑ์„ ๊ฒฐ์ •ํ•˜๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋‘ ํ›ˆ๋ จ ์„ธํŠธ(NQ์™€ SQuAD)์™€ ํ‰๊ฐ€ ์„ธํŠธ(PopQA)๊ฐ€ ๋ชจ๋‘ Wikipedia๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๊ณ  ์žˆ์ง€๋งŒ, ๋ฏธ๋ฌ˜ํ•œ ์ฐจ์ด๊ฐ€ ์กด์žฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
NQ ๋˜๋Š” SQuAD์—์„œ ํ›ˆ๋ จ Adapt-LLM์˜ ์ฐจ์ด๋Š” ์ƒ๋Œ€์ ์œผ๋กœ ๋ฏธ๋ฏธํ•˜์ง€๋งŒ, ์ฃผ์–ด์ง„ ํ‰๊ฐ€ ์„ธํŠธ์— ๋Œ€ํ•œ ํ›ˆ๋ จ ์„ธํŠธ์˜ ์ ํ•ฉ์„ฑ์„ ๊ฒฐ์ •ํ•˜๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋‘ ํ›ˆ๋ จ ์„ธํŠธ(NQ ๋ฐ SQuAD)์™€ ํ‰๊ฐ€ ์„ธํŠธ(PopQA) ๋ชจ๋‘ ์œ„ํ‚คํ”ผ๋””์•„๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜์ง€๋งŒ ๋ฏธ๋ฌ˜ํ•œ ์ฐจ์ด๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Footnote 8: There are actually 247 annotations with unfaithful claims, but for this analysis we leave out 42 unclear ones that require further clarification from the annotators. Note that since the claims sometimes contain multiple subclaims, we allow each annotation to have multiple labels.
Footnote 8: ์‹ค์ œ๋กœ๋Š” ์‹ ๋ขฐํ•  ์ˆ˜ ์—†๋Š” ์ฃผ์žฅ์— ๋Œ€ํ•œ ์ฃผ์„์ด 247๊ฐœ ์žˆ์ง€๋งŒ, ์ด ๋ถ„์„์—์„œ๋Š” ์ฃผ์„์ž๋“ค์˜ ์ถ”๊ฐ€์ ์ธ ์„ค๋ช…์ด ํ•„์š”ํ•œ 42๊ฐœ์˜ ๋ถˆ๋ช…ํ™•ํ•œ ์ฃผ์„์„ ์ œ์™ธํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์žฅ์ด ๋•Œ๋•Œ๋กœ ์—ฌ๋Ÿฌ ํ•˜์œ„ ์ฃผ์žฅ์„ ํฌํ•จํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์—, ๊ฐ ์ฃผ์„์— ์—ฌ๋Ÿฌ ๋ ˆ์ด๋ธ”์„ ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค.
๊ฐ์ฃผ 8: ์‹ค์ œ๋กœ๋Š” ์ถฉ์‹คํ•˜์ง€ ์•Š์€ ์ฃผ์žฅ์ด ์žˆ๋Š” ์ฃผ์„์ด 247๊ฐœ๋‚˜ ๋˜์ง€๋งŒ, ์ด ๋ถ„์„์—์„œ๋Š” ์ฃผ์„ ์ž‘์„ฑ์ž์˜ ์ถ”๊ฐ€ ์„ค๋ช…์ด ํ•„์š”ํ•œ ๋ถˆ๋ถ„๋ช…ํ•œ ์ฃผ์„ 42๊ฐœ๋ฅผ ์ œ์™ธํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์žฅ์— ์—ฌ๋Ÿฌ ํ•˜์œ„ ์ฃผ์žฅ์ด ํฌํ•จ๋˜์–ด ์žˆ์„ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๊ฐ ์ฃผ์„์— ์—ฌ๋Ÿฌ ๋ผ๋ฒจ์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Based on the origin and variations of SSM mentioned above, many SSM-based works are constantly emerging, including but not limited to natural language processing, computer vision, and so on. The following summaries will respectively introduce the expansion and application of various fields.
์œ„์—์„œ ์–ธ๊ธ‰ํ•œ SSM์˜ ๊ธฐ์›๊ณผ ๋ณ€ํ˜•์„ ๋ฐ”ํƒ•์œผ๋กœ, ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ, ์ปดํ“จํ„ฐ ๋น„์ „ ๋“ฑ ๋‹ค์–‘ํ•œ SSM ๊ธฐ๋ฐ˜ ์ž‘์—…๋“ค์ด ๋Š์ž„์—†์ด ๋“ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ ์š”์•ฝ์—์„œ๋Š” ๊ฐ๊ฐ์˜ ๋ถ„์•ผ์—์„œ์˜ ํ™•์žฅ๊ณผ ์‘์šฉ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.
์œ„์—์„œ ์–ธ๊ธ‰ํ•œ SSM์˜ ๊ธฐ์›๊ณผ ๋ณ€ํ˜•์— ๋”ฐ๋ผ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ, ์ปดํ“จํ„ฐ ๋น„์ „ ๋“ฑ์„ ํฌํ•จํ•˜๋˜ ์ด์— ๊ตญํ•œ๋˜์ง€ ์•Š๋Š” ๋งŽ์€ SSM ๊ธฐ๋ฐ˜ ์ž‘์—…์ด ์ง€์†์ ์œผ๋กœ ๋“ฑ์žฅํ•˜๊ณ  ์žˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์˜ ํ™•์žฅ๊ณผ ์‘์šฉ์— ๋Œ€ํ•ด ๊ฐ๊ฐ ์†Œ๊ฐœํ•œ๋‹ค.
is a legitimate probability density function on the real numbers whose negative log-likelihood equals the Huber loss function up to an additive constant.7 Consequently, we can convert the loss minimization problem into a likelihood maximization problem for the distribution defined by \(p\).
์‹ค์ˆ˜์— ๋Œ€ํ•œ ํ•ฉ๋ฒ•์ ์ธ ํ™•๋ฅ  ๋ฐ€๋„ ํ•จ์ˆ˜๋กœ, ๊ทธ ์Œ์˜ ๋กœ๊ทธ ๊ฐ€๋Šฅ๋„๋Š” ์ถ”๊ฐ€ ์ƒ์ˆ˜๊นŒ์ง€ ํฌํ•จํ•˜์—ฌ Huber ์†์‹ค ํ•จ์ˆ˜์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค.7 ๋”ฐ๋ผ์„œ ์†์‹ค ์ตœ์†Œํ™” ๋ฌธ์ œ๋ฅผ \(p\)๋กœ ์ •์˜๋œ ๋ถ„ํฌ์— ๋Œ€ํ•œ ๊ฐ€๋Šฅ๋„ ์ตœ๋Œ€ํ™” ๋ฌธ์ œ๋กœ ๋ณ€ํ™˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์‹ค์ˆ˜์— ๋Œ€ํ•œ ํ•ฉ๋ฒ•์ ์ธ ํ™•๋ฅ  ๋ฐ€๋„ ํ•จ์ˆ˜์ด๋ฉฐ, ์Œ์˜ ๋กœ๊ทธ ๊ฐ€๋Šฅ์„ฑ์€ ๊ฐ€์‚ฐ ์ƒ์ˆ˜๊นŒ์ง€ ํ—ˆ๋ฒ„ ์†์‹ค ํ•จ์ˆ˜์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค.7 ๊ฒฐ๊ณผ์ ์œผ๋กœ ์†์‹ค ์ตœ์†Œํ™” ๋ฌธ์ œ๋ฅผ ์ •์˜ํ•˜๋Š” ๋ถ„ํฌ์— ๋Œ€ํ•œ ๊ฐ€๋Šฅ์„ฑ ์ตœ๋Œ€ํ™” ๋ฌธ์ œ๋กœ ๋ณ€ํ™˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Intuitively, such a preference on the premise order aligns with human preference (Dekeyser et al., 2000) because in the preferred order, each derivation step can be done on-the-fly while looking at premises one by one, without needing to look back and forth across all premises at each step.
์ง๊ด€์ ์œผ๋กœ, ์ „์ œ ์ˆœ์„œ์— ๋Œ€ํ•œ ์ด๋Ÿฌํ•œ ์„ ํ˜ธ๋Š” ์ธ๊ฐ„์˜ ์„ ํ˜ธ์™€ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค (Dekeyser et al., 2000). ์„ ํ˜ธ๋˜๋Š” ์ˆœ์„œ์—์„œ๋Š” ๊ฐ ๋„์ถœ ๋‹จ๊ณ„๋ฅผ ์ „์ œ๋ฅผ ํ•˜๋‚˜์”ฉ ๋ณด๋ฉด์„œ ์ฆ‰์„์—์„œ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๊ฐ ๋‹จ๊ณ„๋งˆ๋‹ค ๋ชจ๋“  ์ „์ œ๋ฅผ ์•ž๋’ค๋กœ ์‚ดํŽด๋ณผ ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค.
์ง๊ด€์ ์œผ๋กœ, ์ด๋Ÿฌํ•œ ์ „์ œ ์ˆœ์„œ์— ๋Œ€ํ•œ ์„ ํ˜ธ๋Š” ์„ ํ˜ธ ์ˆœ์„œ์—์„œ๋Š” ๊ฐ ๋„์ถœ ๋‹จ๊ณ„๊ฐ€ ๊ฐ ๋‹จ๊ณ„๋งˆ๋‹ค ๋ชจ๋“  ์ „์ œ๋ฅผ ์•ž๋’ค๋กœ ํ›‘์–ด๋ณด์ง€ ์•Š๊ณ  ์ „์ œ๋ฅผ ํ•˜๋‚˜์”ฉ ๋ณด๋ฉด์„œ ์ฆ‰์„์—์„œ ์ˆ˜ํ–‰๋  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„(Dekeyser et al., 2000)์™€ ์ผ์น˜ํ•œ๋‹ค.
**ChatGPT**: High-intensity interval training (HIIT) is a popular exercise method among athletes and fitness enthusiasts due to its potential to improve performance, endurance, and overall health. HIIT involves short bursts of intense exercise followed by periods of rest or active recovery.
ChatGPT: ๊ณ ๊ฐ•๋„ ์ธํ„ฐ๋ฒŒ ํŠธ๋ ˆ์ด๋‹ (HIIT)์€ ์„ฑ๋Šฅ, ์ง€๊ตฌ๋ ฅ ๋ฐ ์ „๋ฐ˜์ ์ธ ๊ฑด๊ฐ•์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ ๋•Œ๋ฌธ์— ์šด๋™ ์„ ์ˆ˜์™€ ํ”ผํŠธ๋‹ˆ์Šค ์• ํ˜ธ๊ฐ€๋“ค ์‚ฌ์ด์—์„œ ์ธ๊ธฐ ์žˆ๋Š” ์šด๋™ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. HIIT๋Š” ์งง์€ ์‹œ๊ฐ„ ๋™์•ˆ์˜ ๊ฐ•๋„ ๋†’์€ ์šด๋™๊ณผ ํœด์‹ ๋˜๋Š” ํ™œ๋™์ ์ธ ํšŒ๋ณต ๊ธฐ๊ฐ„์„ ๋ฒˆ๊ฐˆ์•„ ๊ฐ€๋ฉฐ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
**ChatGPT**: ๊ณ ๊ฐ•๋„ ์ธํ„ฐ๋ฒŒ ํŠธ๋ ˆ์ด๋‹(HIIT)์€ ์šด๋™ ๋Šฅ๋ ฅ, ์ง€๊ตฌ๋ ฅ ๋ฐ ์ „๋ฐ˜์ ์ธ ๊ฑด๊ฐ•์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์œผ๋กœ ์ธํ•ด ์šด๋™์„ ์ˆ˜์™€ ํ”ผํŠธ๋‹ˆ์Šค ์• ํ˜ธ๊ฐ€๋“ค ์‚ฌ์ด์—์„œ ์ธ๊ธฐ ์žˆ๋Š” ์šด๋™ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. HIIT๋Š” ์งง์€ ์‹œ๊ฐ„ ๋™์•ˆ ๊ฐ•๋ ฌํ•œ ์šด๋™์„ ํ•œ ํ›„ ํœด์‹ ๋˜๋Š” ์ ๊ทน์ ์ธ ํšŒ๋ณต์„ ํ•˜๋Š” ์šด๋™์ž…๋‹ˆ๋‹ค.
For the \(\mathsf{bioD}\) dataset, we define \(\mathcal{N}_{0}\) to be identical to \(\mathsf{bioS}\), with \(|\mathcal{N}_{0}|=400\times 400\times 1000\). We encapsulate a person's attributes within a single paragraph, employing random sentence orderings and a consistent sentence template.
\(\mathsf{bioD}\) ๋ฐ์ดํ„ฐ์…‹์˜ ๊ฒฝ์šฐ, \(\mathcal{N}_{0}\)๋ฅผ \(\mathsf{bioS}\)์™€ ๋™์ผํ•˜๊ฒŒ ์ •์˜ํ•˜๋ฉฐ, \(|\mathcal{N}_{0}|=400\times 400\times 1000\)์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ฌด์ž‘์œ„ ๋ฌธ์žฅ ์ˆœ์„œ์™€ ์ผ๊ด€๋œ ๋ฌธ์žฅ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•œ ๋‹จ๋ฝ ๋‚ด์— ํ•œ ์‚ฌ๋žŒ์˜ ์†์„ฑ์„ ์บก์Šํ™”ํ•ฉ๋‹ˆ๋‹ค.
์ƒ๋ฌผํ•™์  ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์˜ ๊ฒฝ์šฐ, \(\mathcal{N}_{0}\)์€ \(\mathsf{bioS}\)์™€ ๋™์ผํ•˜๋„๋ก ์ •์˜ํ•˜๊ณ , \(|\mathcal{N}_{0}|=400\times 400\times 1000\)์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ๋žŒ์˜ ์†์„ฑ์„ ๋‹จ๋ฝ ํ•˜๋‚˜์— ์บก์Šํ™”ํ•˜๊ณ  ๋ฌด์ž‘์œ„ ๋ฌธ์žฅ ์ˆœ์„œ์™€ ์ผ๊ด€๋œ ๋ฌธ์žฅ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
**General Text Data.** As we can see in Figure 6, the vast majority of LLMs adopt general-purpose pre-training data, such as webpages, books, and conversational text, which provides rich text sources on a variety of topics. Next, we briefly summarize three important kinds of general data.
**์ผ๋ฐ˜ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ.** Figure 6์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ๋Œ€๋ถ€๋ถ„์˜ LLM์€ ์›นํŽ˜์ด์ง€, ์ฑ…, ๋Œ€ํ™” ํ…์ŠคํŠธ์™€ ๊ฐ™์€ ๋ฒ”์šฉ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ฑ„ํƒํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ฃผ์ œ์— ๋Œ€ํ•œ ํ’๋ถ€ํ•œ ํ…์ŠคํŠธ ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์ผ๋ฐ˜ ๋ฐ์ดํ„ฐ ์œ ํ˜•์„ ๊ฐ„๋žตํžˆ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค.
**์ผ๋ฐ˜ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ.** ๊ทธ๋ฆผ 6์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, ๋Œ€๋‹ค์ˆ˜์˜ LLM์€ ์›น ํŽ˜์ด์ง€, ์ฑ…, ๋Œ€ํ™” ํ…์ŠคํŠธ์™€ ๊ฐ™์€ ๋ฒ”์šฉ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ฑ„ํƒํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋‹ค์–‘ํ•œ ์ฃผ์ œ์— ๋Œ€ํ•œ ํ’๋ถ€ํ•œ ํ…์ŠคํŠธ ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ ์ผ๋ฐ˜ ๋ฐ์ดํ„ฐ์˜ ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์ข…๋ฅ˜๋ฅผ ๊ฐ„๋žตํ•˜๊ฒŒ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค.
As the sun set on the backyard, my character was left alone with their newfound obsession. They knew that it was destroying their life, but they couldn't bring themselves to let go. They were trapped, lost in the echoes of their own mind, searching for something that could never be found.
ํ•ด์งˆ๋…˜ ๋’ท๋งˆ๋‹น์— ํ™€๋กœ ๋‚จ๊ฒจ์ง„ ๋‚ด ์บ๋ฆญํ„ฐ๋Š” ์ƒˆ๋กœ ์ƒ๊ธด ์ง‘์ฐฉ์— ์‚ฌ๋กœ์žกํ˜€ ์žˆ์—ˆ๋‹ค. ๊ทธ๋“ค์€ ๊ทธ๊ฒƒ์ด ์ž์‹ ์˜ ์‚ถ์„ ๋ง์น˜๊ณ  ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ์•˜์ง€๋งŒ, ๋†“์•„์ค„ ์ˆ˜๊ฐ€ ์—†์—ˆ๋‹ค. ๊ทธ๋“ค์€ ์ž์‹ ๋งŒ์˜ ๋งˆ์Œ์˜ ๋ฉ”์•„๋ฆฌ ์†์— ๊ฐ‡ํ˜€, ๊ฒฐ์ฝ” ์ฐพ์„ ์ˆ˜ ์—†๋Š” ๋ฌด์–ธ๊ฐ€๋ฅผ ์ฐพ๊ณ  ์žˆ์—ˆ๋‹ค.
๋’ท๋งˆ๋‹น์— ํ•ด๊ฐ€ ์ง€๋ฉด์„œ ๋‚ด ์บ๋ฆญํ„ฐ๋Š” ์ƒˆ๋กœ ๋ฐœ๊ฒฌํ•œ ์ง‘์ฐฉ์— ํ™€๋กœ ๋‚จ๊ฒจ์กŒ์Šต๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ๊ทธ๊ฒƒ์ด ์ž์‹ ์˜ ์‚ถ์„ ํŒŒ๊ดดํ•˜๊ณ  ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ์•˜์ง€๋งŒ ๋†“์•„์ค„ ์ˆ˜ ์—†์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋“ค์€ ์ž์‹ ์˜ ๋งˆ์Œ์˜ ๋ฉ”์•„๋ฆฌ ์†์— ๊ฐ‡ํ˜€์„œ ๊ฒฐ์ฝ” ์ฐพ์„ ์ˆ˜ ์—†๋Š” ๊ฒƒ์„ ์ฐพ๊ณ  ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
While we could observe large cost differences between multilingual and monolingual English tokenizers in the monolingual English setting, the difference in computational costs between multilingual and monolingual English tokenizers for processing English documents is marginal (Fig. 2(c)).
๋‹ค๊ตญ์–ด์™€ ๋‹จ์ผ ์–ธ์–ด ์˜์–ด ํ† ํฌ๋‚˜์ด์ € ๊ฐ„์˜ ๋น„์šฉ ์ฐจ์ด๋Š” ๋‹จ์ผ ์–ธ์–ด ์˜์–ด ์„ค์ •์—์„œ ํฌ๊ฒŒ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์—ˆ์ง€๋งŒ, ์˜์–ด ๋ฌธ์„œ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ ๋‹ค๊ตญ์–ด์™€ ๋‹จ์ผ ์–ธ์–ด ์˜์–ด ํ† ํฌ๋‚˜์ด์ € ๊ฐ„์˜ ๊ณ„์‚ฐ ๋น„์šฉ ์ฐจ์ด๋Š” ๋ฏธ๋ฏธํ•ฉ๋‹ˆ๋‹ค (Fig. 2(c)).
๋‹จ์ผ ์–ธ์–ด ์˜์–ด ํ™˜๊ฒฝ์—์„œ ๋‹ค๊ตญ์–ด ํ† ํฐํ™” ๋„๊ตฌ์™€ ๋‹จ์ผ ์–ธ์–ด ์˜์–ด ํ† ํฐํ™” ๋„๊ตฌ ๊ฐ„์˜ ํฐ ๋น„์šฉ ์ฐจ์ด๋ฅผ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์˜์–ด ๋ฌธ์„œ ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ๋‹ค๊ตญ์–ด ํ† ํฐํ™” ๋„๊ตฌ์™€ ๋‹จ์ผ ์–ธ์–ด ์˜์–ด ํ† ํฐํ™” ๋„๊ตฌ ๊ฐ„์˜ ๊ณ„์‚ฐ ๋น„์šฉ ์ฐจ์ด๋Š” ๋ฏธ๋ฏธํ•˜๋‹ค(๊ทธ๋ฆผ 2(c)).
Applications.Recently, LLM-based agents have shown great potential in autonomously solving complex tasks, making it feasible to rapidly develop capable applications for specific domains or tasks. In this section, we will discuss the applications in single-agent and multi-agent scenarios.
์‘์šฉ. ์ตœ๊ทผ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ๋ณต์žกํ•œ ์ž‘์—…์„ ์ž์œจ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ํฐ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์œผ๋ฉฐ, ํŠน์ • ๋„๋ฉ”์ธ์ด๋‚˜ ์ž‘์—…์— ์ ํ•ฉํ•œ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์„ ์‹ ์†ํ•˜๊ฒŒ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€๋Šฅํ•ด์กŒ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ๋‹จ์ผ ์—์ด์ „ํŠธ ๋ฐ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ์˜ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
์• ํ”Œ๋ฆฌ์ผ€์ด์…˜.์ตœ๊ทผ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ๋ณต์žกํ•œ ์ž‘์—…์„ ์ž์œจ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๋Š” ๋›ฐ์–ด๋‚œ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์œผ๋ฉฐ, ํŠน์ • ๋„๋ฉ”์ธ ๋˜๋Š” ์ž‘์—…์— ๋งž๋Š” ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ˜ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ๋น ๋ฅด๊ฒŒ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ๋‹จ์ผ ์—์ด์ „ํŠธ ๋ฐ ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ์˜ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
All that remains is to show that we can uncover the full logit vector for distinct prompt queries through this API. In this section, we develop techniques for this purpose. Once we have recovered multiple complete logit vector, we can run the attack from Section 4.2 without modification.
๋‚จ์€ ๊ฒƒ์€ ์ด API๋ฅผ ํ†ตํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ํ”„๋กฌํ”„ํŠธ ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ ์ „์ฒด ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ๋ฐœ๊ฒฌํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ด๋ฅผ ์œ„ํ•œ ๊ธฐ์ˆ ์„ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์™„์ „ํ•œ ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ๋ณต๊ตฌํ•œ ํ›„์—๋Š” 4.2์ ˆ์˜ ๊ณต๊ฒฉ์„ ์ˆ˜์ • ์—†์ด ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋‚จ์€ ๊ฒƒ์€ ์ด API๋ฅผ ํ†ตํ•ด ๊ณ ์œ ํ•œ ํ”„๋กฌํ”„ํŠธ ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ ์ „์ฒด ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ์ฐพ์„ ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ด๋ฅผ ์œ„ํ•œ ๊ธฐ์ˆ ์„ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์™„์ „ํ•œ ๋กœ์ง“ ๋ฒกํ„ฐ๋ฅผ ๋ณต๊ตฌํ•œ ํ›„์—๋Š” ์„น์…˜ 4.2์˜ ๊ณต๊ฒฉ์„ ์ˆ˜์ • ์—†์ด ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Lemma G.1**.: _There exists a positive semidefinite \(\mathbf{A}\in\mathbb{R}^{h\times h}\) such that \(x_{i}^{\top}\mathbf{A}x_{i}=1\) for all \(i\)._Because \(\mathbf{A}\) is positive semidefinite, we can write \(\mathbf{A}=\mathbf{M}^{\top}\cdot\mathbf{M}\) for some \(\mathbf{M}\).
**Lemma G.1**.: _๋ชจ๋“  \(i\)์— ๋Œ€ํ•ด \(x_{i}^{\top}\mathbf{A}x_{i}=1\)์ด ๋˜๋Š” ์–‘์˜ ์ค€์ •๊ทœ ํ–‰๋ ฌ \(\mathbf{A}\in\mathbb{R}^{h\times h}\)๊ฐ€ ์กด์žฌํ•œ๋‹ค._ \(\mathbf{A}\)๊ฐ€ ์–‘์˜ ์ค€์ •๊ทœ ํ–‰๋ ฌ์ด๊ธฐ ๋•Œ๋ฌธ์—, ์šฐ๋ฆฌ๋Š” \(\mathbf{A}=\mathbf{M}^{\top}\cdot\mathbf{M}\)์œผ๋กœ ์“ธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์ •๋ฆฌ G.1**.: _๋ชจ๋“  \(i\)์— ๋Œ€ํ•ด \(x_{i}^{\top}\mathbf{A}x_{i}=1\)์ด ๋˜๋Š” ์–‘์˜ ๋ฐ˜์ •์˜ \(\mathbf{A}\in\mathbb{R}^{h\times h}\)๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค._ \(\mathbf{A}\)๋Š” ์–‘์˜ ๋ฐ˜์ •์˜์ด๋ฏ€๋กœ, ์–ด๋–ค \(\mathbf{M}\)์— ๋Œ€ํ•ด \(\mathbf{A}=\mathbf{M}^{\top}\cdot\mathbf{M}\)์„ ์“ธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
We have supported our claims by running the experiments on three different models. However, generalization to other LLMs should be tested thoroughly. For example, GPT-4 achieves near perfect accuracy for some MMLU tasks Nori et al. (2023), and thus further improvement is not applicable.
์šฐ๋ฆฌ๋Š” ์„ธ ๊ฐ€์ง€ ๋‹ค๋ฅธ ๋ชจ๋ธ์—์„œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์šฐ๋ฆฌ์˜ ์ฃผ์žฅ์„ ๋’ท๋ฐ›์นจํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹ค๋ฅธ LLM์— ๋Œ€ํ•œ ์ผ๋ฐ˜ํ™”๋Š” ์ฒ ์ €ํžˆ ํ…Œ์ŠคํŠธ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-4๋Š” ์ผ๋ถ€ MMLU ์ž‘์—…์—์„œ ๊ฑฐ์˜ ์™„๋ฒฝํ•œ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•˜์˜€์œผ๋ฉฐ(Nori et al., 2023), ๋”ฐ๋ผ์„œ ์ถ”๊ฐ€์ ์ธ ๊ฐœ์„ ์€ ์ ์šฉ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
์„ธ ๊ฐ€์ง€ ๋‹ค๋ฅธ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ฃผ์žฅ์„ ๋’ท๋ฐ›์นจํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹ค๋ฅธ LLM์— ๋Œ€ํ•œ ์ผ๋ฐ˜ํ™”๋Š” ์ฒ ์ €ํžˆ ํ…Œ์ŠคํŠธ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-4๋Š” ์ผ๋ถ€ MMLU ์ž‘์—…์—์„œ ๊ฑฐ์˜ ์™„๋ฒฝํ•œ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค(Nori et al., 2023), ๋”ฐ๋ผ์„œ ์ถ”๊ฐ€ ๊ฐœ์„ ์€ ์ ์šฉ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
**Effects of Question Verification.** The question verification is used to further improve the generation quality. In our experiments, we found it can improve the performance on MATH benchmark, the results are shown in Table 5, while we do not see significantly impact on GSM8K dataset.
**์งˆ๋ฌธ ๊ฒ€์ฆ์˜ ํšจ๊ณผ.** ์งˆ๋ฌธ ๊ฒ€์ฆ์€ ์ƒ์„ฑ ํ’ˆ์งˆ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์‹คํ—˜์—์„œ, ์ด๋Š” MATH ๋ฒค์น˜๋งˆํฌ์—์„œ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์œผ๋ฉฐ, ๊ฒฐ๊ณผ๋Š” Table 5์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, GSM8K ๋ฐ์ดํ„ฐ์…‹์—์„œ๋Š” ํฐ ์˜ํ–ฅ์„ ๋ณด์ด์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
**์งˆ๋ฌธ ๊ฒ€์ฆ์˜ ํšจ๊ณผ**. ์งˆ๋ฌธ ๊ฒ€์ฆ์€ ์ƒ์„ฑ ํ’ˆ์งˆ์„ ๋”์šฑ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์ˆ˜ํ•™ ๋ฒค์น˜๋งˆํฌ์—์„œ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๊ทธ ๊ฒฐ๊ณผ๋Š” ํ‘œ 5์— ๋‚˜์™€ ์žˆ์ง€๋งŒ GSM8K ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
Difference in revenue = $520 - $480 \\ Difference in revenue = $40 \\ \hline \end{tabular} \end{table} Table 4: A comparison case on different scale size modelsTherefore, the mechanic earned $40 more in revenue on the day with higher revenue, which is Thursday. The final answer is $40.
์ˆ˜์ต ์ฐจ์ด = $520 - $480 \\ ์ˆ˜์ต ์ฐจ์ด = $40 \\ \hline \end{tabular} \end{table} ํ‘œ 4: ๋‹ค๋ฅธ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๋น„๊ต ์‚ฌ๋ก€ ๋”ฐ๋ผ์„œ, ์ •๋น„์‚ฌ๋Š” ์ˆ˜์ต์ด ๋” ๋†’์€ ๋‚ ์ธ ๋ชฉ์š”์ผ์— $40 ๋” ๋ฒŒ์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ์ข… ๋‹ต์€ $40์ž…๋‹ˆ๋‹ค.
์ˆ˜์ต ์ฐจ์ด = $520 - $480 \\ ์ˆ˜์ต ์ฐจ์ด = $40 \\ \hline \end{tabular} \end{table} ํ‘œ 4: ๋‹ค์–‘ํ•œ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๋น„๊ต ์‚ฌ๋ก€๋”ฐ๋ผ์„œ ์ •๋น„์‚ฌ๋Š” ์ˆ˜์ต์ด ๋” ๋†’์€ ๋‚ , ์ฆ‰ ๋ชฉ์š”์ผ์— 40๋‹ฌ๋Ÿฌ๋ฅผ ๋” ๋ฒŒ์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ์ข… ๋‹ต์€ $40์ž…๋‹ˆ๋‹ค.
MoEs are nowadays an important component of the so-called frontier LLMs (i.e. the most advanced and capable models). GPT-4 itself is rumored to be based on a MoE architecture, and some of the best performing LLMs such as Mixtral [117], are basically an MoE version of pre-existing LLMs.
MoE๋Š” ์˜ค๋Š˜๋‚  ์†Œ์œ„ ์ตœ์ฒจ๋‹จ LLM(์ฆ‰, ๊ฐ€์žฅ ์ง„๋ณด๋˜๊ณ  ๋Šฅ๋ ฅ ์žˆ๋Š” ๋ชจ๋ธ)์˜ ์ค‘์š”ํ•œ ๊ตฌ์„ฑ ์š”์†Œ์ž…๋‹ˆ๋‹ค. GPT-4 ์ž์ฒด๋„ MoE ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ๋‹ค๋Š” ์†Œ๋ฌธ์ด ์žˆ์œผ๋ฉฐ, Mixtral [117]๊ณผ ๊ฐ™์€ ์ตœ๊ณ  ์„ฑ๋Šฅ์˜ LLM ์ค‘ ์ผ๋ถ€๋Š” ๊ธฐ๋ณธ์ ์œผ๋กœ ๊ธฐ์กด LLM์˜ MoE ๋ฒ„์ „์ž…๋‹ˆ๋‹ค.
์˜ค๋Š˜๋‚  MoE๋Š” ์†Œ์œ„ ํ”„๋ก ํ‹ฐ์–ด LLM(๊ฐ€์žฅ ์ง„๋ณด๋˜๊ณ  ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚œ ๋ชจ๋ธ)์˜ ์ค‘์š”ํ•œ ๊ตฌ์„ฑ ์š”์†Œ์ž…๋‹ˆ๋‹ค. GPT-4 ์ž์ฒด๋Š” MoE ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์•Œ๋ ค์ ธ ์žˆ์œผ๋ฉฐ, Mixtral [117]๊ณผ ๊ฐ™์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” LLM์€ ๊ธฐ๋ณธ์ ์œผ๋กœ ๊ธฐ์กด LLM์˜ MoE ๋ฒ„์ „์ž…๋‹ˆ๋‹ค.
In our research, we do not focus on the impact of duplicate text in pretrained models on downstream benchmark tasks; instead we address how duplicate text in the LM training and validation sets impacts model perplexity and the extent to which generated text included memorized content.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ์—์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์—์„œ ์ค‘๋ณต ํ…์ŠคํŠธ๊ฐ€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๋ฒค์น˜๋งˆํฌ ์ž‘์—…์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์— ์ดˆ์ ์„ ๋งž์ถ”์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋Œ€์‹ , LM ํ›ˆ๋ จ ๋ฐ ๊ฒ€์ฆ ์„ธํŠธ์—์„œ ์ค‘๋ณต ํ…์ŠคํŠธ๊ฐ€ ๋ชจ๋ธ์˜ ํผํ”Œ๋ ‰์‹œํ‹ฐ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ๊ณผ ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ์— ์•”๊ธฐ๋œ ๋‚ด์šฉ์ด ํฌํ•จ๋˜๋Š” ์ •๋„๋ฅผ ๋‹ค๋ฃน๋‹ˆ๋‹ค.
์šฐ๋ฆฌ ์—ฐ๊ตฌ์—์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ์ค‘๋ณต ํ…์ŠคํŠธ๊ฐ€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ๋ฒค์น˜๋งˆํฌ ์ž‘์—…์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์— ์ดˆ์ ์„ ๋งž์ถ”์ง€ ์•Š๊ณ , ๋Œ€์‹  LM ํ•™์Šต ๋ฐ ๊ฒ€์ฆ ์ง‘ํ•ฉ์˜ ์ค‘๋ณต ํ…์ŠคํŠธ๊ฐ€ ๋ชจ๋ธ ๋‹นํ˜น๊ฐ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ๊ณผ ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ์— ๊ธฐ์–ต๋œ ์ฝ˜ํ…์ธ ๊ฐ€ ํฌํ•จ๋œ ์ •๋„๋ฅผ ๋‹ค๋ฃฌ๋‹ค.
In addition to exploring KG-enhanced LLMs, it is also meaningful to leverage LLMs to improve the tasks on the KG side (_i.e._, LLM4KG) [861, 877]. A typical example is that LLMs can help supplement or construct the KG. We omit the discussion of this part, since it is beyond our scope.
KG ๊ฐ•ํ™” LLM์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„, LLM์„ ํ™œ์šฉํ•˜์—ฌ KG ์ธก์˜ ์ž‘์—…์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ(LM4KG)๋„ ์˜๋ฏธ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค [861, 877]. ์ „ํ˜•์ ์ธ ์˜ˆ๋กœ, LLM์ด KG๋ฅผ ๋ณด์™„ํ•˜๊ฑฐ๋‚˜ ๊ตฌ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์— ๋Œ€ํ•œ ๋…ผ์˜๋Š” ์šฐ๋ฆฌ์˜ ๋ฒ”์œ„๋ฅผ ๋ฒ—์–ด๋‚˜๋ฏ€๋กœ ์ƒ๋žตํ•ฉ๋‹ˆ๋‹ค.
KG๊ฐ€ ๊ฐ•ํ™”๋œ LLM์„ ํƒ์ƒ‰ํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ LLM์„ ํ™œ์šฉํ•˜์—ฌ KG ์ธก๋ฉด์˜ ์ž‘์—…์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ(_์ฆ‰, LLM4KG) [861, 877]๋„ ์˜๋ฏธ๊ฐ€ ์žˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ LLM์ด KG์„ ๋ณด์™„ํ•˜๊ฑฐ๋‚˜ ๊ตฌ์„ฑํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ์ด ๋ถ€๋ถ„์— ๋Œ€ํ•œ ๋…ผ์˜๋Š” ๋ฒ”์œ„๋ฅผ ๋ฒ—์–ด๋‚˜๋ฏ€๋กœ ์ƒ๋žตํ•œ๋‹ค.
To build clusters of similar documents, we construct a graph that has an edge between two documents if they are considered a match. Then, we use the method introduced in Lacki et al. (2018) to identify connected components. A breakdown of the computation needed is given in Appendix A.
์œ ์‚ฌํ•œ ๋ฌธ์„œ๋“ค์˜ ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด, ๋‘ ๋ฌธ์„œ๊ฐ€ ์ผ์น˜ํ•œ๋‹ค๊ณ  ๊ฐ„์ฃผ๋  ๊ฒฝ์šฐ ๊ทธ๋“ค ์‚ฌ์ด์— ์—ฃ์ง€๊ฐ€ ์žˆ๋Š” ๊ทธ๋ž˜ํ”„๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, Lacki et al. (2018)์—์„œ ์†Œ๊ฐœ๋œ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์—ฐ๊ฒฐ๋œ ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค. ํ•„์š”ํ•œ ๊ณ„์‚ฐ์˜ ์„ธ๋ถ€ ์‚ฌํ•ญ์€ ๋ถ€๋ก A์— ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
์œ ์‚ฌํ•œ ๋ฌธ์„œ ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๋ฌธ์„œ๊ฐ€ ์ผ์น˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผ๋˜๋Š” ๊ฒฝ์šฐ ๊ฐ€์žฅ์ž๋ฆฌ๊ฐ€ ์žˆ๋Š” ๊ทธ๋ž˜ํ”„๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ Lacki ๋“ฑ(2018)์—์„œ ์†Œ๊ฐœํ•œ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์—ฐ๊ฒฐ๋œ ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค. ํ•„์š”ํ•œ ๊ณ„์‚ฐ์˜ ๋ถ„์„์€ ๋ถ€๋ก A์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
During development, we saved intermediary outputs from our pipeline for investigations and for ablations-intermediary outputs exist for about 5% of RefinedWeb. We did not keep intermediary outputs for the final production version of the dataset due to storage and resource constraints.
๊ฐœ๋ฐœ ์ค‘์—, ์šฐ๋ฆฌ๋Š” ์กฐ์‚ฌ์™€ ์†Œ๊ฑฐ๋ฅผ ์œ„ํ•ด ํŒŒ์ดํ”„๋ผ์ธ์—์„œ ์ค‘๊ฐ„ ์ถœ๋ ฅ์„ ์ €์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ฐ„ ์ถœ๋ ฅ์€ RefinedWeb์˜ ์•ฝ 5%์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ์ €์žฅ ๊ณต๊ฐ„๊ณผ ์ž์› ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ์ตœ์ข… ์ƒ์‚ฐ ๋ฒ„์ „์˜ ๋ฐ์ดํ„ฐ์…‹์—๋Š” ์ค‘๊ฐ„ ์ถœ๋ ฅ์„ ์œ ์ง€ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
๊ฐœ๋ฐœ ๊ณผ์ •์—์„œ ์กฐ์‚ฌ ๋ฐ ์ ˆ์ œ๋ฅผ ์œ„ํ•ด ํŒŒ์ดํ”„๋ผ์ธ์˜ ์ค‘๊ฐ„ ์ถœ๋ ฅ์„ ์ €์žฅํ–ˆ์œผ๋ฉฐ, ์•ฝ 5%์˜ RefinedWeb์— ๋Œ€ํ•œ ์ค‘๊ฐ„ ์ถœ๋ ฅ์ด ์กด์žฌํ•ฉ๋‹ˆ๋‹ค. ์ €์žฅ ๋ฐ ๋ฆฌ์†Œ์Šค ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ตœ์ข… ํ”„๋กœ๋•์…˜ ๋ฒ„์ „์— ๋Œ€ํ•œ ์ค‘๊ฐ„ ์ถœ๋ ฅ์„ ๋ณด๊ด€ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
The intrinsic evaluation aims to assess the generated output of tokenizers based on _fertility_ and _parity_. Furthermore, the tokenizer's vocabulary overlap with other tokenizers is computed. The intrinsic evaluation does not assess the impact of tokenizers on the model performance.
๋‚ด์žฌ์  ํ‰๊ฐ€๋Š” _fertility_์™€ _parity_๋ฅผ ๊ธฐ์ค€์œผ๋กœ ํ† ํฌ๋‚˜์ด์ €๊ฐ€ ์ƒ์„ฑํ•œ ์ถœ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ† ํฌ๋‚˜์ด์ €์˜ ์–ดํœ˜๊ฐ€ ๋‹ค๋ฅธ ํ† ํฌ๋‚˜์ด์ €์™€ ์–ผ๋งˆ๋‚˜ ๊ฒน์น˜๋Š”์ง€๋„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋‚ด์žฌ์  ํ‰๊ฐ€๋Š” ํ† ํฌ๋‚˜์ด์ €๊ฐ€ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
๋‚ด์žฌ์  ํ‰๊ฐ€๋Š” _fertility_ ๋ฐ _parity_๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ† ํฐํ™”๊ธฐ์˜ ์ƒ์„ฑ๋œ ์ถœ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค๋ฅธ ํ† ํฐํ™”๊ธฐ์™€์˜ ํ† ํฐํ™”๊ธฐ ์–ดํœ˜ ์ค‘๋ณต๋„๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋‚ด์žฌ์  ํ‰๊ฐ€๋Š” ํ† ํฐํ™”๊ฐ€ ๋ชจ๋ธ ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
The effectiveness of the fit may also differ between the early and late stages of training. For example, if the curve fits well initially but diverges as training progresses, it might suggest that Eq.(1) more accurately predicts performance improvements at earlier stages of training.
์ ํ•ฉ๋„์˜ ํšจ๊ณผ๋Š” ํ›ˆ๋ จ์˜ ์ดˆ๊ธฐ ๋‹จ๊ณ„์™€ ํ›„๊ธฐ ๋‹จ๊ณ„ ์‚ฌ์ด์—์„œ ๋‹ค๋ฅผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ณก์„ ์ด ์ฒ˜์Œ์—๋Š” ์ž˜ ๋งž์ง€๋งŒ ํ›ˆ๋ จ์ด ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ ๋ฒ—์–ด๋‚œ๋‹ค๋ฉด, ์ด๋Š” Eq.(1)์ด ํ›ˆ๋ จ์˜ ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋” ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ ํ•ฉ๋„์˜ ํšจ๊ณผ๋Š” ํ›ˆ๋ จ์˜ ์ดˆ๊ธฐ ๋‹จ๊ณ„์™€ ํ›„๊ธฐ ๋‹จ๊ณ„ ๊ฐ„์—๋„ ๋‹ค๋ฅผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ดˆ๊ธฐ์—๋Š” ๊ณก์„ ์ด ์ž˜ ๋งž์ง€๋งŒ ํ›ˆ๋ จ์ด ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ ๋ฒ—์–ด๋‚œ๋‹ค๋ฉด, ์ด๋Š” Eq.(1)์ด ํ›ˆ๋ จ์˜ ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋” ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
We further demonstrate the data quality by comparing the speed of preference increase during data scaling. As is shown in Fig. 5, when compared with UltraChat [19] and its cleaned version UltraChat 200K, we see a clear tendency of performance improvements when scaling up the Yi data.
๋ฐ์ดํ„ฐ ์Šค์ผ€์ผ๋ง ์ค‘ ์„ ํ˜ธ๋„ ์ฆ๊ฐ€ ์†๋„๋ฅผ ๋น„๊ตํ•˜์—ฌ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์„ ์ถ”๊ฐ€๋กœ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. Fig. 5์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, UltraChat [19] ๋ฐ ๊ทธ ์ •์ œ๋œ ๋ฒ„์ „์ธ UltraChat 200K์™€ ๋น„๊ตํ•  ๋•Œ, Yi ๋ฐ์ดํ„ฐ๋ฅผ ํ™•์žฅํ•  ๋•Œ ์„ฑ๋Šฅ์ด ๋ช…ํ™•ํ•˜๊ฒŒ ํ–ฅ์ƒ๋˜๋Š” ๊ฒฝํ–ฅ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ ์Šค์ผ€์ผ๋ง ์ค‘ ์„ ํ˜ธ๋„ ์ฆ๊ฐ€ ์†๋„๋ฅผ ๋น„๊ตํ•˜์—ฌ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์„ ๋”์šฑ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 5์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด UltraChat [19] ๋ฐ ์ •๋ฆฌ๋œ ๋ฒ„์ „์ธ UltraChat 200K์™€ ๋น„๊ตํ–ˆ์„ ๋•Œ, Yi ๋ฐ์ดํ„ฐ๋ฅผ ํ™•์žฅํ•  ๋•Œ ์„ฑ๋Šฅ์ด ๊ฐœ์„ ๋˜๋Š” ๊ฒฝํ–ฅ์ด ๋ถ„๋ช…ํ•˜๊ฒŒ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.
Figure 2: Relative frequency of the cases where the internal recall of the bridge entity of LLaMA-2 increases with entity substitution (top row) and relation substitution (bottom row). Bars are colored blue if the relative frequency is greater than or equal to 0.5 and red otherwise.
Figure 2: LLaMA-2์˜ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์ด ์—”ํ‹ฐํ‹ฐ ๋Œ€์ฒด(์ƒ๋‹จ ํ–‰) ๋ฐ ๊ด€๊ณ„ ๋Œ€์ฒด(ํ•˜๋‹จ ํ–‰)์™€ ํ•จ๊ป˜ ์ฆ๊ฐ€ํ•˜๋Š” ๊ฒฝ์šฐ์˜ ์ƒ๋Œ€ ๋นˆ๋„. ์ƒ๋Œ€ ๋นˆ๋„๊ฐ€ 0.5 ์ด์ƒ์ธ ๊ฒฝ์šฐ ๋ง‰๋Œ€๋Š” ํŒŒ๋ž€์ƒ‰์œผ๋กœ, ๊ทธ๋ ‡์ง€ ์•Š์€ ๊ฒฝ์šฐ ๋นจ๊ฐ„์ƒ‰์œผ๋กœ ์ƒ‰์น ๋ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 2: ์—”ํ‹ฐํ‹ฐ ๋Œ€์ฒด(์ƒ๋‹จ ํ–‰) ๋ฐ ๊ด€๊ณ„ ๋Œ€์ฒด(ํ•˜๋‹จ ํ–‰)์— ๋”ฐ๋ผ LLaMA-2์˜ ๋ธŒ๋ฆฌ์ง€ ์—”ํ‹ฐํ‹ฐ์˜ ๋‚ด๋ถ€ ๋ฆฌ์ฝœ์ด ์ฆ๊ฐ€ํ•˜๋Š” ๊ฒฝ์šฐ์˜ ์ƒ๋Œ€์  ๋นˆ๋„์ž…๋‹ˆ๋‹ค. ๋ง‰๋Œ€๋Š” ์ƒ๋Œ€์  ๋นˆ๋„๊ฐ€ 0.5๋ณด๋‹ค ํฌ๊ฑฐ๋‚˜ ๊ฐ™์œผ๋ฉด ํŒŒ๋ž€์ƒ‰์œผ๋กœ, ๊ทธ๋ ‡์ง€ ์•Š์œผ๋ฉด ๋นจ๊ฐ„์ƒ‰์œผ๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค.
PersonA: Yes, urinalysis can be used as a screening and diagnostic tool for bladder cancer. One way to test for bladder cancer is to check for blood in the urine, which is called hematuria. This can be done during a urinalysis by examining the urine cells under a special microscope.
PersonA: ๋„ค, ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ๋ฐฉ๊ด‘์•”์˜ ์„ ๋ณ„ ๋ฐ ์ง„๋‹จ ๋„๊ตฌ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐฉ๊ด‘์•”์„ ๊ฒ€์‚ฌํ•˜๋Š” ํ•œ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์€ ์†Œ๋ณ€์—์„œ ํ˜ˆ์•ก์„ ํ™•์ธํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์ด๋ฅผ ํ˜ˆ๋‡จ๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํŠน์ˆ˜ ํ˜„๋ฏธ๊ฒฝ์œผ๋กœ ์†Œ๋ณ€ ์„ธํฌ๋ฅผ ๊ฒ€์‚ฌํ•˜๋Š” ์†Œ๋ณ€ ๊ฒ€์‚ฌ ์ค‘์— ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ๋„ค, ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ๋ฐฉ๊ด‘์•”์˜ ์„ ๋ณ„ ๋ฐ ์ง„๋‹จ ๋„๊ตฌ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐฉ๊ด‘์•” ๊ฒ€์‚ฌ์˜ ํ•œ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์€ ํ˜ˆ๋‡จ๋ผ๊ณ  ํ•˜๋Š” ์†Œ๋ณ€์— ํ˜ˆ์•ก์ด ์žˆ๋Š”์ง€ ํ™•์ธํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ํŠน์ˆ˜ ํ˜„๋ฏธ๊ฒฝ์œผ๋กœ ์†Œ๋ณ€ ์„ธํฌ๋ฅผ ๊ฒ€์‚ฌํ•˜์—ฌ ์†Œ๋ณ€ ๊ฒ€์‚ฌ ์ค‘์— ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
This is partially because language models tend to commit to an answer earlier in their response, which can cause a'snowball effect' of increasing diversion from their goal state [34]. By implementing feedback, agents are much more likely to correct their course and reach their goal.
์ด๋Š” ๋ถ€๋ถ„์ ์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์ด ์‘๋‹ต ์ดˆ๊ธฐ์— ๋‹ต๋ณ€์„ ํ™•์ •ํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋ฉฐ, ์ด๋Š” ๋ชฉํ‘œ ์ƒํƒœ์—์„œ ์ ์  ๋” ๋ฒ—์–ด๋‚˜๋Š” '๋ˆˆ๋ฉ์ด ํšจ๊ณผ'๋ฅผ ์ผ์œผํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [34]. ํ”ผ๋“œ๋ฐฑ์„ ๊ตฌํ˜„ํ•จ์œผ๋กœ์จ ์—์ด์ „ํŠธ๋Š” ๊ฒฝ๋กœ๋ฅผ ์ˆ˜์ •ํ•˜๊ณ  ๋ชฉํ‘œ์— ๋„๋‹ฌํ•  ๊ฐ€๋Šฅ์„ฑ์ด ํ›จ์”ฌ ๋†’์•„์ง‘๋‹ˆ๋‹ค.
์ด๋Š” ๋ถ€๋ถ„์ ์œผ๋กœ ์–ธ์–ด ๋ชจ๋ธ์ด ์‘๋‹ต ์ดˆ๊ธฐ์— ๋‹ต๋ณ€์— ์ „๋…ํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์–ด ๋ชฉํ‘œ ์ƒํƒœ์—์„œ ๋ฒ—์–ด๋‚˜๋Š” ํ˜„์ƒ์ด '๋ˆˆ๋ฉ์ด ํšจ๊ณผ'๋กœ ์ปค์งˆ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค[34]. ํ”ผ๋“œ๋ฐฑ์„ ๊ตฌํ˜„ํ•จ์œผ๋กœ์จ ์—์ด์ „ํŠธ๋Š” ์ž์‹ ์˜ ๊ณผ์ •์„ ์ˆ˜์ •ํ•˜๊ณ  ๋ชฉํ‘œ์— ๋„๋‹ฌํ•  ๊ฐ€๋Šฅ์„ฑ์ด ํ›จ์”ฌ ๋†’์•„์ง„๋‹ค.
While previous work on LLM-based autonomous agent has obtained many remarkable successes, this field is still at its initial stage, and there are several significant challenges that need to be addressed in its development. In the following, we present many representative challenges.
LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ์ด์ „ ์—ฐ๊ตฌ๋Š” ๋งŽ์€ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์„ฑ๊ณต์„ ๊ฑฐ๋‘์—ˆ์ง€๋งŒ, ์ด ๋ถ„์•ผ๋Š” ์—ฌ์ „ํžˆ ์ดˆ๊ธฐ ๋‹จ๊ณ„์— ์žˆ์œผ๋ฉฐ ๊ฐœ๋ฐœ ๊ณผ์ •์—์„œ ํ•ด๊ฒฐํ•ด์•ผ ํ•  ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ณผ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์—ฌ๋Ÿฌ ๋Œ€ํ‘œ์ ์ธ ๊ณผ์ œ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ์ด์ „ ์—ฐ๊ตฌ์—์„œ ๋งŽ์€ ๋†€๋ผ์šด ์„ฑ๊ณต์„ ๊ฑฐ๋‘์—ˆ์ง€๋งŒ, ์ด ๋ถ„์•ผ๋Š” ์•„์ง ์ดˆ๊ธฐ ๋‹จ๊ณ„์— ์žˆ์œผ๋ฉฐ ๊ฐœ๋ฐœ ๊ณผ์ •์—์„œ ํ•ด๊ฒฐํ•ด์•ผ ํ•  ๋ช‡ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ณผ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ๋Œ€ํ‘œ์ ์ธ ๊ณผ์ œ๋ฅผ ๋งŽ์ด ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
Similar to LLMs themselves, evaluating the effectiveness of LLM-based autonomous agents is a challenging task. This section outlines two prevalent approaches to evaluation: subjective and objective methods. For a comprehensive overview, please refer to the right portion of Figure 5.
LLM ์ž์ฒด์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์˜ ํšจ๊ณผ์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ๋„์ „์ ์ธ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ์ ‘๊ทผ๋ฒ•์ธ ์ฃผ๊ด€์  ๋ฐฉ๋ฒ•๊ณผ ๊ฐ๊ด€์  ๋ฐฉ๋ฒ•์„ ๊ฐœ๋žต์ ์œผ๋กœ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ Figure 5์˜ ์˜ค๋ฅธ์ชฝ ๋ถ€๋ถ„์„ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค.
LLM ์ž์ฒด์™€ ์œ ์‚ฌํ•˜๊ฒŒ, LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์˜ ํšจ๊ณผ๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ๋„ ์–ด๋ ค์šด ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ํ‰๊ฐ€์— ๋Œ€ํ•œ ๋‘ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์ธ ์ฃผ๊ด€์  ๋ฐฉ๋ฒ•๊ณผ ๊ฐ๊ด€์  ๋ฐฉ๋ฒ•์„ ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๊ทธ๋ฆผ 5์˜ ์˜ค๋ฅธ์ชฝ ๋ถ€๋ถ„์„ ์ฐธ์กฐํ•˜์„ธ์š”.
While we believe that this work further enhances our understanding of knowledge in LLMs, there is a lot more work to be done in this field. Specifically, more research is required regarding the question of knowledge representation in LLMs, especially from a theoretical perspective.
์šฐ๋ฆฌ๋Š” ์ด ์—ฐ๊ตฌ๊ฐ€ LLMs์—์„œ์˜ ์ง€์‹์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค๊ณ  ๋ฏฟ์ง€๋งŒ, ์ด ๋ถ„์•ผ์—์„œ ํ•ด์•ผ ํ•  ์ผ์ด ๋งŽ์ด ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, LLMs์—์„œ์˜ ์ง€์‹ ํ‘œํ˜„์— ๊ด€ํ•œ ๋ฌธ์ œ, ํŠนํžˆ ์ด๋ก ์ ์ธ ๊ด€์ ์—์„œ ๋” ๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
์ด ์ž‘์—…์ด LLM์˜ ์ง€์‹์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ๋”์šฑ ๋†’์—ฌ์ค€๋‹ค๊ณ  ์ƒ๊ฐํ•˜์ง€๋งŒ, ์ด ๋ถ„์•ผ์—์„œ ํ•ด์•ผ ํ•  ์ผ์€ ์•„์ง ๋งŽ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์ด๋ก ์  ๊ด€์ ์—์„œ LLM์˜ ์ง€์‹ ํ‘œํ˜„ ๋ฌธ์ œ์— ๋Œ€ํ•œ ๋” ๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
Another recent work [50] proposes the DARE method goes further by zeroing out small differences between the fine-tuned model and the original base model, while amplifying the differences. In practice, DARE [50] is often used together with Task Arithmetic [21] or TIES-Merging [49].
๋˜ ๋‹ค๋ฅธ ์ตœ๊ทผ ์—ฐ๊ตฌ [50]์—์„œ๋Š” DARE ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•˜๋ฉฐ, ์ด๋Š” ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ๊ณผ ์›๋ž˜ ๊ธฐ๋ณธ ๋ชจ๋ธ ๊ฐ„์˜ ์ž‘์€ ์ฐจ์ด๋ฅผ ์—†์• ๊ณ  ์ฐจ์ด๋ฅผ ์ฆํญ์‹œํ‚ค๋Š” ๋ฐฉ์‹์œผ๋กœ ๋” ๋‚˜์•„๊ฐ‘๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ DARE [50]๋Š” ์ข…์ข… Task Arithmetic [21] ๋˜๋Š” TIES-Merging [49]๊ณผ ํ•จ๊ป˜ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
๋˜ ๋‹ค๋ฅธ ์ตœ๊ทผ ์—ฐ๊ตฌ [50]์—์„œ๋Š” ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ๊ณผ ์›๋ž˜ ๊ธฐ๋ณธ ๋ชจ๋ธ ์‚ฌ์ด์˜ ์ž‘์€ ์ฐจ์ด๋ฅผ ๋ฌดํšจํ™”ํ•˜๋Š” ๋™์‹œ์— ์ฐจ์ด๋ฅผ ์ฆํญ์‹œํ‚ด์œผ๋กœ์จ DARE ๋ฐฉ๋ฒ•์„ ๋” ๋ฐœ์ „์‹œ์ผฐ๋‹ค. ์‹ค์ œ๋กœ DARE [50]๋Š” ์ข…์ข… Task Arithmetic [21] ๋˜๋Š” TIES-Merging [49]๊ณผ ํ•จ๊ป˜ ์‚ฌ์šฉ๋œ๋‹ค.
In Sections C.3-C.6 we can no longer use general invertible matrices \(\mathbf{S}\) in our arguments, and must instead use orthogonal matrices, matrices \(\mathbf{U}\) such that \(\mathbf{U}^{\top}\mathbf{U}=I\). In models with LayerNorm, we specialise further, too (Appendix C.5).
์„น์…˜ C.3-C.6์—์„œ๋Š” ๋” ์ด์ƒ ์ผ๋ฐ˜์ ์ธ ๊ฐ€์—ญ ํ–‰๋ ฌ \(\mathbf{S}\)๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์œผ๋ฉฐ, ๋Œ€์‹  ์ง๊ต ํ–‰๋ ฌ \(\mathbf{U}\)๋ฅผ ์‚ฌ์šฉํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ง๊ต ํ–‰๋ ฌ์€ \(\mathbf{U}^{\top}\mathbf{U}=I\)๋ฅผ ๋งŒ์กฑํ•˜๋Š” ํ–‰๋ ฌ์ž…๋‹ˆ๋‹ค. LayerNorm์ด ํฌํ•จ๋œ ๋ชจ๋ธ์—์„œ๋Š” ์ถ”๊ฐ€๋กœ ํŠน์ˆ˜ํ™”ํ•ฉ๋‹ˆ๋‹ค (๋ถ€๋ก C.5 ์ฐธ์กฐ).
์„น์…˜ C.3-C.6์—์„œ๋Š” ๋” ์ด์ƒ ์ผ๋ฐ˜์ ์ธ ๋ฐ˜์ „ ๊ฐ€๋Šฅ ํ–‰๋ ฌ(\mathbf{S}\)์„ ์ธ์ˆ˜์— ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์œผ๋ฉฐ, ๋Œ€์‹  ์ง๊ต ํ–‰๋ ฌ, ํ–‰๋ ฌ(\mathbf{U}^{\top}\mathbf{U}=I\)๋ฅผ ์‚ฌ์šฉํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. LayerNorm์ด ์žˆ๋Š” ๋ชจ๋ธ์—์„œ๋Š” ๋” ํŠน์ˆ˜ํ™”๋ฉ๋‹ˆ๋‹ค(๋ถ€๋ก C.5).
We extend special gratitude to the Upwork annotators for their hard work, and to members from the UMass NLP lab for their feedback. This project was partially supported by awards IIS-2202506 and IIS-2312949 from the National Science Foundation (NSF) as well as an award from Adobe.
Upwork ์ฃผ์„์ž๋“ค์˜ ๋…ธ๊ณ ์— ํŠน๋ณ„ํžˆ ๊ฐ์‚ฌ๋“œ๋ฆฌ๋ฉฐ, UMass NLP ์—ฐ๊ตฌ์‹ค์˜ ๊ตฌ์„ฑ์›๋“ค์—๊ฒŒ๋„ ํ”ผ๋“œ๋ฐฑ์— ๋Œ€ํ•ด ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์ ํŠธ๋Š” National Science Foundation (NSF)์˜ IIS-2202506 ๋ฐ IIS-2312949 ์ƒ๊ณผ Adobe์˜ ์ƒ์˜ ์ผ๋ถ€ ์ง€์›์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค.
๋…ธ๋ ฅํ•ด ์ฃผ์‹  Upwork ์ฃผ์„๊ฐ€์™€ UMass NLP ์—ฐ๊ตฌ์†Œ์˜ ํšŒ์›๋“ค์—๊ฒŒ ํŠน๋ณ„ํ•œ ๊ฐ์‚ฌ๋ฅผ ํ‘œํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์ ํŠธ๋Š” ๋ฏธ๊ตญ๊ณผํ•™์žฌ๋‹จ(NSF)์˜ IIS-2202506 ๋ฐ IIS-2312949 ์ƒ๊ณผ Adobe์˜ ์ง€์›์œผ๋กœ ๋ถ€๋ถ„์ ์œผ๋กœ ์ง€์›๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
To further verify our proposed framework, we examine two more types of multilingual LLMs, including BLOOMZ Muennighoff et al. (2022), a _hyper-multilingual_ LLM claiming to support 46 languages, and Chinese Llam Cui et al. (2023), a _bilingual_ LLM focusing on English and Chinese.
์šฐ๋ฆฌ์˜ ์ œ์•ˆ๋œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ถ”๊ฐ€๋กœ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์˜ ๋‹ค๊ตญ์–ด LLM์„ ๋” ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” 46๊ฐœ ์–ธ์–ด๋ฅผ ์ง€์›ํ•œ๋‹ค๊ณ  ์ฃผ์žฅํ•˜๋Š” _์ดˆ๋‹ค๊ตญ์–ด_ LLM์ธ BLOOMZ Muennighoff et al. (2022)์™€ ์˜์–ด์™€ ์ค‘๊ตญ์–ด์— ์ค‘์ ์„ ๋‘” _์ด์ค‘์–ธ์–ด_ LLM์ธ Chinese Llam Cui et al. (2023)์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.
์ œ์•ˆํ•œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ถ”๊ฐ€๋กœ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด 46๊ฐœ ์–ธ์–ด๋ฅผ ์ง€์›ํ•œ๋‹ค๊ณ  ์ฃผ์žฅํ•˜๋Š” _ํ•˜์ดํผ ๋ฉ€ํ‹ฐ๋ง๊ตฌ์–ผ_ LLM์ธ BLOOMZ Muennighoff et al. (2022)๊ณผ ์˜์–ด์™€ ์ค‘๊ตญ์–ด์— ์ดˆ์ ์„ ๋งž์ถ˜ _์ด์ค‘ ์–ธ์–ด_ LLM์ธ ์ค‘๊ตญ์–ด Llam Cui et al. (2023)์„ ํฌํ•จํ•œ ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์˜ ๋ฉ€ํ‹ฐ ์–ธ์–ด LLM์„ ๋” ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.
The taxonomy does not provide comprehensive coverage of all possible hazard categories [e.g., 33], and we will continue to expand and iterate on it for v1.0. We set out to develop an easy-to-understand, usable and standardized resource for practitioners, engineers and researchers.
์ด ๋ถ„๋ฅ˜ ์ฒด๊ณ„๋Š” ๋ชจ๋“  ๊ฐ€๋Šฅํ•œ ์œ„ํ—˜ ๋ฒ”์ฃผ๋ฅผ ํฌ๊ด„์ ์œผ๋กœ ๋‹ค๋ฃจ์ง€ ์•Š์Šต๋‹ˆ๋‹ค [์˜ˆ: 33], ๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ๋Š” v1.0์„ ์œ„ํ•ด ์ด๋ฅผ ๊ณ„์† ํ™•์žฅํ•˜๊ณ  ๋ฐ˜๋ณตํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์‹ค๋ฌด์ž, ์—”์ง€๋‹ˆ์–ด ๋ฐ ์—ฐ๊ตฌ์ž๋“ค์„ ์œ„ํ•œ ์ดํ•ดํ•˜๊ธฐ ์‰ฝ๊ณ , ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•˜๋ฉฐ ํ‘œ์ค€ํ™”๋œ ์ž์›์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
๋ถ„๋ฅ˜๋ฒ•์€ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  ์œ„ํ—˜ ๋ฒ”์ฃผ[์˜ˆ: 33]๋ฅผ ํฌ๊ด„์ ์œผ๋กœ ๋‹ค๋ฃจ์ง€๋Š” ์•Š์œผ๋ฉฐ, v1.0์„ ์œ„ํ•ด ์ด๋ฅผ ๊ณ„์† ํ™•์žฅํ•˜๊ณ  ๋ฐ˜๋ณตํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์‹ค๋ฌด์ž, ์—”์ง€๋‹ˆ์–ด ๋ฐ ์—ฐ๊ตฌ์›์„ ์œ„ํ•œ ์ดํ•ดํ•˜๊ธฐ ์‰ฌ์šฐ๋ฉฐ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•˜๊ณ  ํ‘œ์ค€ํ™”๋œ ๋ฆฌ์†Œ์Šค๋ฅผ ๊ฐœ๋ฐœํ•˜๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค.
Even with some knowledge about floods and their statistical properties, it would be very difficult to guess that this specific meteorologist would call the flood a '200-year event'. This is especially true if the model was not exposed to information about the details of the flood.
ํ™์ˆ˜์™€ ๊ทธ ํ†ต๊ณ„์  ํŠน์„ฑ์— ๋Œ€ํ•œ ์–ด๋Š ์ •๋„์˜ ์ง€์‹์ด ์žˆ๋”๋ผ๋„, ํŠน์ • ๊ธฐ์ƒํ•™์ž๊ฐ€ ์ด ํ™์ˆ˜๋ฅผ '200๋…„ ๋งŒ์˜ ์‚ฌ๊ฑด'์ด๋ผ๊ณ  ๋ถ€๋ฅผ ๊ฒƒ์ด๋ผ๊ณ  ์ถ”์ธกํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์–ด๋ ค์šธ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ ๋ชจ๋ธ์ด ํ™์ˆ˜์˜ ์„ธ๋ถ€ ์‚ฌํ•ญ์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ ‘ํ•˜์ง€ ๋ชปํ•œ ๊ฒฝ์šฐ์—๋Š” ๋”์šฑ ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค.
ํ™์ˆ˜์™€ ํ™์ˆ˜์˜ ํ†ต๊ณ„์  ํŠน์„ฑ์— ๋Œ€ํ•œ ์ง€์‹์ด ์–ด๋Š ์ •๋„ ์žˆ๋”๋ผ๋„ ์ด ํŠน์ • ๊ธฐ์ƒํ•™์ž๊ฐ€ ํ™์ˆ˜๋ฅผ '200๋…„ ์ด๋ฒคํŠธ'๋ผ๊ณ  ๋ถ€๋ฅผ ๊ฒƒ์ด๋ผ๊ณ  ์ถ”์ธกํ•˜๊ธฐ๋Š” ๋งค์šฐ ์–ด๋ ค์šธ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ ํ™์ˆ˜์˜ ์„ธ๋ถ€ ์‚ฌํ•ญ์— ๋Œ€ํ•œ ์ •๋ณด์— ๋…ธ์ถœ๋˜์ง€ ์•Š์€ ๋ชจ๋ธ์ด๋ผ๋ฉด ๋”์šฑ ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค.
these aspects by offering a structured overview of RAG, categorizing various methods, and delivering an in-depth understanding of this research area. This survey will primarily focus on textual applications of RAG, reflecting the current emphasis of research efforts in this area.
์ด๋Ÿฌํ•œ ์ธก๋ฉด๋“ค์„ RAG์— ๋Œ€ํ•œ ๊ตฌ์กฐํ™”๋œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•˜๊ณ , ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•๋“ค์„ ๋ถ„๋ฅ˜ํ•˜๋ฉฐ, ์ด ์—ฐ๊ตฌ ๋ถ„์•ผ์— ๋Œ€ํ•œ ์‹ฌ์ธต์ ์ธ ์ดํ•ด๋ฅผ ์ „๋‹ฌํ•จ์œผ๋กœ์จ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ์ฃผ๋กœ RAG์˜ ํ…์ŠคํŠธ ์‘์šฉ์— ์ดˆ์ ์„ ๋งž์ถ”๋ฉฐ, ์ด ๋ถ„์•ผ์˜ ํ˜„์žฌ ์—ฐ๊ตฌ ๋…ธ๋ ฅ์„ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ์ธก๋ฉด์„ ์ œ๊ณตํ•จ์œผ๋กœ์จ RAG์— ๋Œ€ํ•œ ๊ตฌ์กฐํ™”๋œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•˜๊ณ , ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ๋ถ„๋ฅ˜ํ•˜๊ณ , ์ด ์—ฐ๊ตฌ ์˜์—ญ์— ๋Œ€ํ•œ ์‹ฌ์ธต์ ์ธ ์ดํ•ด๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” ์ฃผ๋กœ ์ด ๋ถ„์•ผ์˜ ์—ฐ๊ตฌ ๋…ธ๋ ฅ์— ๋Œ€ํ•œ ํ˜„์žฌ์˜ ๊ฐ•์กฐ์ ์„ ๋ฐ˜์˜ํ•˜์—ฌ RAG์˜ ํ…์ŠคํŠธ ์‘์šฉ ๋ถ„์•ผ์— ์ค‘์ ์„ ๋‘˜ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Legal SourcesWe combine both previously available datasets, such as the FreeLaw subset from The Pile Gao et al. (2020) and MultiLegal Pile Niklaus et al. (2023), as well as data scraped from publicly available sources on the Web. We list the different sources of data in Table 1.
๋ฒ•์  ์ถœ์ฒ˜: ์šฐ๋ฆฌ๋Š” The Pile์˜ FreeLaw ํ•˜์œ„ ์ง‘ํ•ฉ(Gao et al., 2020)๊ณผ MultiLegal Pile(Niklaus et al., 2023)๊ณผ ๊ฐ™์€ ์ด์ „์— ์‚ฌ์šฉ ๊ฐ€๋Šฅํ–ˆ๋˜ ๋ฐ์ดํ„ฐ์…‹๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์›น์—์„œ ๊ณต๊ฐœ์ ์œผ๋กœ ์ด์šฉ ๊ฐ€๋Šฅํ•œ ์†Œ์Šค์—์„œ ์Šคํฌ๋žฉํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์ถœ์ฒ˜๋Š” Table 1์— ๋‚˜์—ด๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
๋ฒ•์  ์ถœ์ฒ˜The Pile Gao et al. (2020)์˜ FreeLaw ํ•˜์œ„ ์ง‘ํ•ฉ๊ณผ MultiLegal Pile Niklaus et al. (2023)๊ณผ ๊ฐ™์ด ์ด์ „์— ์‚ฌ์šฉ ๊ฐ€๋Šฅํ–ˆ๋˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ์›น์—์„œ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์†Œ์Šค์—์„œ ์Šคํฌ๋žฉํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋‘ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋Š” ํ‘œ 1์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
**Examination:** Multiple-choice questions in examinations. For English, we use MMLU [14], which includes mathematics, US history, computer science, law, and more. For Chinese, we use C-Eval [18] which comprises multiple-choice ranging from humanities to science and engineering.
**์‹œํ—˜:** ์‹œํ—˜์—์„œ์˜ ๊ฐ๊ด€์‹ ์งˆ๋ฌธ. ์˜์–ด์˜ ๊ฒฝ์šฐ, ์ˆ˜ํ•™, ๋ฏธ๊ตญ ์—ญ์‚ฌ, ์ปดํ“จํ„ฐ ๊ณผํ•™, ๋ฒ•๋ฅ  ๋“ฑ์„ ํฌํ•จํ•˜๋Š” MMLU [14]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด์˜ ๊ฒฝ์šฐ, ์ธ๋ฌธํ•™๋ถ€ํ„ฐ ๊ณผํ•™ ๋ฐ ๊ณตํ•™์— ์ด๋ฅด๋Š” ๊ฐ๊ด€์‹ ์งˆ๋ฌธ์„ ํฌํ•จํ•˜๋Š” C-Eval [18]์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
**์‹œํ—˜:** ์‹œํ—˜์˜ ๊ฐ๊ด€์‹ ๋ฌธ์ œ. ์˜์–ด์˜ ๊ฒฝ์šฐ ์ˆ˜ํ•™, ๋ฏธ๊ตญ ์—ญ์‚ฌ, ์ปดํ“จํ„ฐ ๊ณผํ•™, ๋ฒ•ํ•™ ๋“ฑ์ด ํฌํ•จ๋œ MMLU [14]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด์˜ ๊ฒฝ์šฐ ์ธ๋ฌธํ•™์—์„œ ์ด๊ณต๊ณ„์—ด์— ์ด๋ฅด๋Š” ๊ฐ๊ด€์‹ ๋ฌธ์ œ๋กœ ๊ตฌ์„ฑ๋œ C-Eval [18]์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
Where \(x^{*}\) is the predicted inversion, \(d\) is the embedding dimension, and \(\mathbf{v}\) is padded with zeros at the final position. In practice we use \(|\mathcal{V}|=32000\) and \(d=768\) for all experiments which leads to a fixed-length input sequence of \(42\) words.
\(x^{*}\)๋Š” ์˜ˆ์ธก๋œ ๋ฐ˜์ „๊ฐ’์ด๊ณ , \(d\)๋Š” ์ž„๋ฒ ๋”ฉ ์ฐจ์›์ด๋ฉฐ, \(\mathbf{v}\)๋Š” ๋งˆ์ง€๋ง‰ ์œ„์น˜์— 0์œผ๋กœ ์ฑ„์›Œ์ง‘๋‹ˆ๋‹ค. ์‹ค์ œ ์‹คํ—˜์—์„œ๋Š” \(|\mathcal{V}|=32000\)๊ณผ \(d=768\)์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณ ์ • ๊ธธ์ด์˜ ์ž…๋ ฅ ์‹œํ€€์Šค๊ฐ€ \(42\) ๋‹จ์–ด๊ฐ€ ๋˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์„œ \(x^{*}\)๋Š” ์˜ˆ์ธก๋œ ๋ฐ˜์ „, \(d\)๋Š” ์ž„๋ฒ ๋”ฉ ์ฐจ์›, \(\mathbf{v}\)๋Š” ๋งˆ์ง€๋ง‰ ์œ„์น˜์—์„œ 0์œผ๋กœ ํŒจ๋”ฉ๋œ๋‹ค. ์‹ค์ œ๋กœ๋Š” ๋ชจ๋“  ์‹คํ—˜์— ๋Œ€ํ•ด \(|\mathcal{V}|=32000\), \(d=768\)์„ ์‚ฌ์šฉํ•˜์—ฌ \(42\) ๋‹จ์–ด์˜ ๊ณ ์ • ๊ธธ์ด ์ž…๋ ฅ ์‹œํ€€์Šค๋กœ ์ด์–ด์ง„๋‹ค.
Besides, some researchers [312] propose to penalize the generation of previously generated tokens or \(n\)-grams to alleviate the issue of repetitive generation. In addition, diverse beam search [313] can be leveraged to produce a set of diverse outputs based on the same input.
๊ฒŒ๋‹ค๊ฐ€, ์ผ๋ถ€ ์—ฐ๊ตฌ์ž๋“ค์€ [312] ๋ฐ˜๋ณต ์ƒ์„ฑ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ด์ „์— ์ƒ์„ฑ๋œ ํ† ํฐ์ด๋‚˜ \(n\)-๊ทธ๋žจ์˜ ์ƒ์„ฑ์„ ํŽ˜๋„ํ‹ฐ๋ฅผ ์ฃผ๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋‹ค์–‘ํ•œ ๋น” ์„œ์น˜ [313]๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋™์ผํ•œ ์ž…๋ ฅ์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•œ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋˜ํ•œ ์ผ๋ถ€ ์—ฐ๊ตฌ์ž๋“ค์€ [312] ๋ฐ˜๋ณต์ ์ธ ์ƒ์„ฑ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ด์ „์— ์ƒ์„ฑ๋œ ํ† ํฐ ๋˜๋Š” \(n\)-๊ทธ๋žจ์˜ ์ƒ์„ฑ์— ๋ถˆ์ด์ต์„ ์ฃผ๋Š” ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ ๋น” ๊ฒ€์ƒ‰ [313]์„ ํ™œ์šฉํ•˜์—ฌ ๋™์ผํ•œ ์ž…๋ ฅ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค์–‘ํ•œ ์ถœ๋ ฅ ์ง‘ํ•ฉ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
CoT is based on making the implicit reasoning process of LLMs explicit. By outlining the steps required for reasoning, the model is directed closer to a logical and reasoned output, especially in scenarios demanding more than simple information retrieval or pattern recognition.
CoT๋Š” LLM์˜ ์•”๋ฌต์ ์ธ ์ถ”๋ก  ๊ณผ์ •์„ ๋ช…์‹œ์ ์œผ๋กœ ๋งŒ๋“œ๋Š” ๊ฒƒ์— ๊ธฐ๋ฐ˜ํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก ์— ํ•„์š”ํ•œ ๋‹จ๊ณ„๋ฅผ ๋ช…ํ™•ํžˆ ํ•จ์œผ๋กœ์จ, ๋ชจ๋ธ์€ ๋‹จ์ˆœํ•œ ์ •๋ณด ๊ฒ€์ƒ‰์ด๋‚˜ ํŒจํ„ด ์ธ์‹์„ ๋„˜์–ด์„œ๋Š” ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋…ผ๋ฆฌ์ ์ด๊ณ  ์ด์„ฑ์ ์ธ ์ถœ๋ ฅ์„ ๋” ์ž˜ ๋„์ถœํ•  ์ˆ˜ ์žˆ๋„๋ก ์œ ๋„๋ฉ๋‹ˆ๋‹ค.
CoT๋Š” LLM์˜ ์•”๋ฌต์ ์ธ ์ถ”๋ก  ๊ณผ์ •์„ ๋ช…์‹œํ™”ํ•˜๋Š” ๊ฒƒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก ์— ํ•„์š”ํ•œ ๋‹จ๊ณ„๋ฅผ ์„ค๋ช…ํ•จ์œผ๋กœ์จ ๋ชจ๋ธ์ด ๋…ผ๋ฆฌ์ ์ด๊ณ  ์ด์„ฑ์ ์ธ ๊ฒฐ๊ณผ๋ฌผ, ํŠนํžˆ ๋‹จ์ˆœํ•œ ์ •๋ณด ๊ฒ€์ƒ‰์ด๋‚˜ ํŒจํ„ด ์ธ์‹ ์ด์ƒ์˜ ๊ฒƒ์„ ์š”๊ตฌํ•˜๋Š” ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋” ๊ฐ€๊นŒ์›Œ์งˆ ์ˆ˜ ์žˆ๋„๋ก ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค.
LLMs have many inherent limitations when it comes to reasoning, such as the inability to access external information, arithmetic errors, and inconsistent reasoning. These issues can be cleverly circumvented by entrusting specific responsibilities to dedicated modules or models.
LLMs๋Š” ์™ธ๋ถ€ ์ •๋ณด๋ฅผ ์ ‘๊ทผํ•  ์ˆ˜ ์—†๋Š” ์ , ์‚ฐ์ˆ  ์˜ค๋ฅ˜, ์ผ๊ด€์„ฑ ์—†๋Š” ์ถ”๋ก  ๋“ฑ ์ถ”๋ก ์— ์žˆ์–ด ๋งŽ์€ ๊ณ ์œ ํ•œ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋“ค์€ ํŠน์ • ์ฑ…์ž„์„ ์ „๋‹ด ๋ชจ๋“ˆ์ด๋‚˜ ๋ชจ๋ธ์— ๋งก๊น€์œผ๋กœ์จ ๊ต๋ฌ˜ํ•˜๊ฒŒ ํšŒํ”ผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์™ธ๋ถ€ ์ •๋ณด์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์ , ์‚ฐ์ˆ  ์˜ค๋ฅ˜, ์ผ๊ด€์„ฑ ์—†๋Š” ์ถ”๋ก  ๋“ฑ LLM์€ ์ถ”๋ก ์— ์žˆ์–ด ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๋ณธ์งˆ์ ์ธ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋Š” ์ „๋‹ด ๋ชจ๋“ˆ์ด๋‚˜ ๋ชจ๋ธ์— ํŠน์ • ์ฑ…์ž„์„ ๋งก๊ธฐ๋ฉด ์˜๋ฆฌํ•˜๊ฒŒ ์šฐํšŒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
To increase the diversity of document lengths, we collect five conversational QA datasets with short documents (less than 1.5K words). On average, 1 word will be tokenized into 1.5 tokens. Hence, the document can be directly fitted into LLMs with a sequence length of 4K tokens.
๋ฌธ์„œ ๊ธธ์ด์˜ ๋‹ค์–‘์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ์งง์€ ๋ฌธ์„œ(1.5K ๋‹จ์–ด ์ดํ•˜)๋ฅผ ํฌํ•จํ•œ ๋‹ค์„ฏ ๊ฐœ์˜ ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ์…‹์„ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ท ์ ์œผ๋กœ, 1๊ฐœ์˜ ๋‹จ์–ด๋Š” 1.5๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ํ† ํฐํ™”๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ฌธ์„œ๋Š” 4K ํ† ํฐ์˜ ์‹œํ€€์Šค ๊ธธ์ด๋ฅผ ๊ฐ€์ง„ LLM์— ์ง์ ‘ ๋งž์ถœ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ฌธ์„œ ๊ธธ์ด์˜ ๋‹ค์–‘์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ์งง์€ ๋ฌธ์„œ(1.5K ๋‹จ์–ด ๋ฏธ๋งŒ)๊ฐ€ ํฌํ•จ๋œ 5๊ฐœ์˜ ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ท ์ ์œผ๋กœ 1๋‹จ์–ด๋Š” 1.5๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ํ† ํฐํ™”๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ฌธ์„œ๋Š” ์‹œํ€€์Šค ๊ธธ์ด๊ฐ€ 4K ํ† ํฐ์ธ LLM์— ์ง์ ‘ ๋งž์ถœ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
EvaluationWe evaluated the ability to generate Japanese answers to Japanese math problems. Therefore, we considered an answer correct if it met the following criteria: (1) the concluding numerical value must be correct, and (2) the reasoning text should be written in Japanese.
ํ‰๊ฐ€ ์šฐ๋ฆฌ๋Š” ์ผ๋ณธ์–ด ์ˆ˜ํ•™ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์ผ๋ณธ์–ด ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์Œ ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋Š” ๊ฒฝ์šฐ ๋‹ต๋ณ€์„ ์˜ฌ๋ฐ”๋ฅด๋‹ค๊ณ  ๊ฐ„์ฃผํ–ˆ์Šต๋‹ˆ๋‹ค: (1) ์ตœ์ข… ์ˆ˜์น˜ ๊ฐ’์ด ์ •ํ™•ํ•ด์•ผ ํ•˜๋ฉฐ, (2) ์ถ”๋ก  ํ…์ŠคํŠธ๋Š” ์ผ๋ณธ์–ด๋กœ ์ž‘์„ฑ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
ํ‰๊ฐ€์ผ๋ณธ์–ด ์ˆ˜ํ•™ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์ผ๋ณธ์–ด ๋‹ต์•ˆ์„ ์ž‘์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ (1) ๊ฒฐ๋ก ์ ์ธ ์ˆ˜์น˜๊ฐ€ ์ •ํ™•ํ•ด์•ผ ํ•˜๊ณ , (2) ์ถ”๋ก  ํ…์ŠคํŠธ๊ฐ€ ์ผ๋ณธ์–ด๋กœ ์ž‘์„ฑ๋˜์–ด์•ผ ํ•œ๋‹ค๋Š” ๊ธฐ์ค€์„ ์ถฉ์กฑํ•˜๋Š” ๊ฒฝ์šฐ ์ •๋‹ต์œผ๋กœ ๊ฐ„์ฃผํ–ˆ์Šต๋‹ˆ๋‹ค.
Finally, we could demonstrate that there is no clear correlation between intrinsic and extrinsic tokenizer performance, but the correlation is rather task-specific. A small fertility value might be a necessary condition for good downstream performance but not a sufficient one.
๋งˆ์ง€๋ง‰์œผ๋กœ, ๋‚ด์žฌ์  ํ† ํฌ๋‚˜์ด์ € ์„ฑ๋Šฅ๊ณผ ์™ธ์žฌ์  ํ† ํฌ๋‚˜์ด์ € ์„ฑ๋Šฅ ์‚ฌ์ด์— ๋ช…ํ™•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์—†์Œ์„ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ์—ˆ์ง€๋งŒ, ์ƒ๊ด€๊ด€๊ณ„๋Š” ์˜คํžˆ๋ ค ์ž‘์—…๋ณ„๋กœ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ์ž‘์€ fertility ๊ฐ’์€ ์ข‹์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์„ ์œ„ํ•œ ํ•„์š” ์กฐ๊ฑด์ผ ์ˆ˜ ์žˆ์ง€๋งŒ ์ถฉ๋ถ„ ์กฐ๊ฑด์€ ์•„๋‹™๋‹ˆ๋‹ค.
๋งˆ์ง€๋ง‰์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ๋‚ด์žฌ์  ํ† ํฐํ™” ๋ฐ ์™ธ์žฌ์  ํ† ํฐํ™” ์„ฑ๋Šฅ ์‚ฌ์ด์—๋Š” ๋ช…ํ™•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์—†์ง€๋งŒ, ๊ทธ ์ƒ๊ด€๊ด€๊ณ„๋Š” ์˜คํžˆ๋ ค ์ž‘์—…๋ณ„์ด๋ผ๋Š” ๊ฒƒ์„ ์ž…์ฆํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค. ์ž‘์€ ๋น„์˜ฅ๋„ ๊ฐ’์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•˜๊ธฐ ์œ„ํ•œ ํ•„์ˆ˜ ์กฐ๊ฑด์ผ ์ˆ˜๋Š” ์žˆ์ง€๋งŒ ์ถฉ๋ถ„ํ•œ ์กฐ๊ฑด์€ ์•„๋‹ˆ๋‹ค.
In summary, in the above sections, we introduce the typical applications of LLM-based autonomous agents in three important domains. To facilitate a clearer understanding, we have summarized the relationship between previous studies and their respective applications in Table 2.
์š”์•ฝํ•˜์ž๋ฉด, ์œ„์˜ ์„น์…˜๋“ค์—์„œ ์šฐ๋ฆฌ๋Š” LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์˜ ๋Œ€ํ‘œ์ ์ธ ์‘์šฉ ์‚ฌ๋ก€๋ฅผ ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๋„๋ฉ”์ธ์—์„œ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ๋” ๋ช…ํ™•ํ•œ ์ดํ•ด๋ฅผ ๋•๊ธฐ ์œ„ํ•ด, ์ด์ „ ์—ฐ๊ตฌ๋“ค๊ณผ ๊ทธ๋“ค์˜ ๊ฐ๊ฐ์˜ ์‘์šฉ ์‚ฌ๋ก€ ๊ฐ„์˜ ๊ด€๊ณ„๋ฅผ ํ‘œ 2์— ์š”์•ฝํ–ˆ์Šต๋‹ˆ๋‹ค.
์š”์•ฝํ•˜๋ฉด, ์œ„์˜ ์„น์…˜์—์„œ๋Š” ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ์˜์—ญ์—์„œ LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ์˜ ์ผ๋ฐ˜์ ์ธ ์ ์šฉ ์‚ฌ๋ก€๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋ณด๋‹ค ๋ช…ํ™•ํ•œ ์ดํ•ด๋ฅผ ๋•๊ธฐ ์œ„ํ•ด ์ด์ „ ์—ฐ๊ตฌ์™€ ๊ฐ๊ฐ์˜ ์ ์šฉ ์‚ฌ๋ก€ ๊ฐ„์˜ ๊ด€๊ณ„๋ฅผ ํ‘œ 2์— ์š”์•ฝํ–ˆ์Šต๋‹ˆ๋‹ค.
REVERSE\({}_{grad}\) can also solve this task, but only when the maximum segment length \(k\) is long enough. When \(k\) is smaller than the entity name length, the entity names will always split across multiple segments, thus the same issue as with word reversal could arise.
REVERSE\({}_{grad}\)๋„ ์ด ์ž‘์—…์„ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์ตœ๋Œ€ ์„ธ๊ทธ๋จผํŠธ ๊ธธ์ด \(k\)๊ฐ€ ์ถฉ๋ถ„ํžˆ ๊ธธ ๋•Œ๋งŒ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. \(k\)๊ฐ€ ์—”ํ‹ฐํ‹ฐ ์ด๋ฆ„ ๊ธธ์ด๋ณด๋‹ค ์งง์„ ๊ฒฝ์šฐ, ์—”ํ‹ฐํ‹ฐ ์ด๋ฆ„์ด ํ•ญ์ƒ ์—ฌ๋Ÿฌ ์„ธ๊ทธ๋จผํŠธ๋กœ ๋ถ„ํ• ๋˜๋ฏ€๋กœ ๋‹จ์–ด ์—ญ์ „๊ณผ ๋™์ผํ•œ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
REVERSE\({}_{grad}\)๋Š” ์ตœ๋Œ€ ์„ธ๊ทธ๋จผํŠธ ๊ธธ์ด \(k\)๊ฐ€ ์ถฉ๋ถ„ํžˆ ๊ธธ ๋•Œ๋งŒ ์ด ์ž‘์—…์„ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. k๊ฐ€ ์—”ํ‹ฐํ‹ฐ ์ด๋ฆ„ ๊ธธ์ด๋ณด๋‹ค ์ž‘์œผ๋ฉด ์—”ํ‹ฐํ‹ฐ ์ด๋ฆ„์ด ํ•ญ์ƒ ์—ฌ๋Ÿฌ ์„ธ๊ทธ๋จผํŠธ์— ๊ฑธ์ณ ๋ถ„ํ• ๋˜๋ฏ€๋กœ ๋‹จ์–ด ๋ฐ˜์ „๊ณผ ๋™์ผํ•œ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
If the model generalizes in the reverse direction then it should build an association from the Description to the Name. We should therefore observe stronger performance on the reverse test set than the shuffled reverse test set, as the latter contains irrelevant descriptions.
๋ชจ๋ธ์ด ์—ญ๋ฐฉํ–ฅ์œผ๋กœ ์ผ๋ฐ˜ํ™”ํ•œ๋‹ค๋ฉด, Description์—์„œ Name์œผ๋กœ์˜ ์—ฐ๊ด€์„ฑ์„ ๊ตฌ์ถ•ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ, ๋ฌด์ž‘์œ„๋กœ ์„ž์ธ ์—ญ๋ฐฉํ–ฅ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ณด๋‹ค ์—ญ๋ฐฉํ–ฅ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ๋” ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํ›„์ž๋Š” ๊ด€๋ จ ์—†๋Š” ์„ค๋ช…์„ ํฌํ•จํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
๋ชจ๋ธ์ด ์—ญ๋ฐฉํ–ฅ์œผ๋กœ ์ผ๋ฐ˜ํ™”๋œ๋‹ค๋ฉด ์„ค๋ช…์—์„œ ์ด๋ฆ„์œผ๋กœ ์—ฐ๊ด€์„ฑ์„ ๊ตฌ์ถ•ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ›„์ž๋Š” ๊ด€๋ จ ์—†๋Š” ์„ค๋ช…์„ ํฌํ•จํ•˜๋ฏ€๋กœ ์…”ํ”Œ๋œ ์—ญ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋ณด๋‹ค ์—ญ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ๋” ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๊ด€์ฐฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
In addition to the models discussed in the previous subsections, there are other popular LLMs which do not belong to those three model families, yet they have achieved great performance and have pushed the LLMs field forward. We briefly describe these LLMs in this subsection.
์ด์ „ ์†Œ์ ˆ์—์„œ ๋…ผ์˜๋œ ๋ชจ๋ธ๋“ค ์™ธ์—๋„, ์„ธ ๊ฐ€์ง€ ๋ชจ๋ธ ๊ณ„์—ด์— ์†ํ•˜์ง€ ์•Š์ง€๋งŒ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋ฉฐ LLMs ๋ถ„์•ผ๋ฅผ ๋ฐœ์ „์‹œํ‚จ ๋‹ค๋ฅธ ์ธ๊ธฐ ์žˆ๋Š” LLM๋“ค์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์†Œ์ ˆ์—์„œ๋Š” ์ด๋Ÿฌํ•œ LLM๋“ค์— ๋Œ€ํ•ด ๊ฐ„๋žตํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์ด์ „ ํ•˜์œ„ ์„น์…˜์—์„œ ์„ค๋ช…ํ•œ ๋ชจ๋ธ ์™ธ์—๋„ ์„ธ ๊ฐ€์ง€ ๋ชจ๋ธ ๊ณ„์—ด์— ์†ํ•˜์ง€ ์•Š์ง€๋งŒ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๊ณ  LLM ๋ถ„์•ผ๋ฅผ ๋ฐœ์ „์‹œํ‚จ ๋‹ค๋ฅธ ์ธ๊ธฐ ์žˆ๋Š” LLM์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ํ•˜์œ„ ์„น์…˜์—์„œ๋Š” ์ด๋Ÿฌํ•œ LLM์— ๋Œ€ํ•ด ๊ฐ„๋žตํ•˜๊ฒŒ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
As researchers who benefited from the open-source community, we would like for models that we release to also be under an open-source license. In the spirit of open-source, and to showcase the applicability of our method to tackle even challenging issues like model licenses.
์˜คํ”ˆ ์†Œ์Šค ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ํ˜œํƒ์„ ๋ฐ›์€ ์—ฐ๊ตฌ์ž๋กœ์„œ, ์šฐ๋ฆฌ๊ฐ€ ๊ณต๊ฐœํ•˜๋Š” ๋ชจ๋ธ๋“ค๋„ ์˜คํ”ˆ ์†Œ์Šค ๋ผ์ด์„ ์Šค ํ•˜์— ์žˆ๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค. ์˜คํ”ˆ ์†Œ์Šค ์ •์‹ ์— ๋”ฐ๋ผ, ๊ทธ๋ฆฌ๊ณ  ๋ชจ๋ธ ๋ผ์ด์„ ์Šค์™€ ๊ฐ™์€ ์–ด๋ ค์šด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•์ด ์ ์šฉ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ๊ธฐ ์œ„ํ•ด์„œ์ž…๋‹ˆ๋‹ค.
์˜คํ”ˆ ์†Œ์Šค ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ํ˜œํƒ์„ ๋ฐ›์€ ์—ฐ๊ตฌ์ž๋กœ์„œ ์ €ํฌ๋Š” ์ €ํฌ๊ฐ€ ์ถœ์‹œํ•˜๋Š” ๋ชจ๋ธ๋„ ์˜คํ”ˆ ์†Œ์Šค ๋ผ์ด์„ ์Šค์— ๋”ฐ๋ผ ์ œ๊ณต๋˜๊ธฐ๋ฅผ ์›ํ•ฉ๋‹ˆ๋‹ค. ์˜คํ”ˆ ์†Œ์Šค ์ •์‹ ์— ๋”ฐ๋ผ ๋ชจ๋ธ ๋ผ์ด์„ ์Šค์™€ ๊ฐ™์€ ์–ด๋ ค์šด ๋ฌธ์ œ๋„ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ์ €ํฌ ๋ฐฉ๋ฒ•์˜ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์—ฌ์ฃผ๊ธฐ ์œ„ํ•ด์„œ์ž…๋‹ˆ๋‹ค.
Here we report the top-k DDx accuracy as computed by the auto-evaluation method. For each DDx in the DDx list generated by AMIE and PCPs, we used Med-PaLM 2 to determine whether the ground truth diagnosis appears within the top-k positions of the differential diagnosis list.
์—ฌ๊ธฐ์„œ ์šฐ๋ฆฌ๋Š” ์ž๋™ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์— ์˜ํ•ด ๊ณ„์‚ฐ๋œ ์ƒ์œ„-k DDx ์ •ํ™•๋„๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. AMIE์™€ PCPs์— ์˜ํ•ด ์ƒ์„ฑ๋œ DDx ๋ชฉ๋ก์˜ ๊ฐ DDx์— ๋Œ€ํ•ด, Med-PaLM 2๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‹ค์ œ ์ง„๋‹จ์ด ๊ฐ๋ณ„ ์ง„๋‹จ ๋ชฉ๋ก์˜ ์ƒ์œ„-k ์œ„์น˜ ๋‚ด์— ๋‚˜ํƒ€๋‚˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
์—ฌ๊ธฐ์—์„œ๋Š” ์ž๋™ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์— ์˜ํ•ด ๊ณ„์‚ฐ๋œ ์ƒ์œ„ k DDx ์ •ํ™•๋„๋ฅผ ๋ณด๊ณ ํ•œ๋‹ค. AMIE ๋ฐ PCP์— ์˜ํ•ด ์ƒ์„ฑ๋œ DDx ๋ชฉ๋ก์˜ ๊ฐ DDx์— ๋Œ€ํ•ด Med-PaLM 2๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‹ค์ธก ์ง„๋‹จ์ด ์ฐจ๋“ฑ ์ง„๋‹จ ๋ชฉ๋ก์˜ ์ƒ์œ„ k ์œ„์น˜์— ๋‚˜ํƒ€๋‚˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ํ™•์ธํ–ˆ๋‹ค.
The following are several passages about the demographics, symptoms, and management plan for a given condition. Generate 2 different patient vignettes consistent with these passages. Follow the format of the given example (just list N/A if a particular field is unavailable).
๋‹ค์Œ์€ ํŠน์ • ์ƒํƒœ์— ๋Œ€ํ•œ ์ธ๊ตฌ ํ†ต๊ณ„, ์ฆ์ƒ ๋ฐ ๊ด€๋ฆฌ ๊ณ„ํš์— ๊ด€ํ•œ ์—ฌ๋Ÿฌ ๋‹จ๋ฝ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋‹จ๋ฝ๊ณผ ์ผ์น˜ํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ๋‹ค๋ฅธ ํ™˜์ž ์‚ฌ๋ก€๋ฅผ ์ƒ์„ฑํ•˜์‹ญ์‹œ์˜ค. ์ฃผ์–ด์ง„ ์˜ˆ์‹œ์˜ ํ˜•์‹์„ ๋”ฐ๋ฅด์„ธ์š” (ํŠน์ • ํ•„๋“œ๊ฐ€ ์—†์œผ๋ฉด N/A๋กœ ํ‘œ์‹œํ•˜์‹ญ์‹œ์˜ค).
๋‹ค์Œ์€ ํŠน์ • ์ƒํƒœ์— ๋Œ€ํ•œ ์ธ๊ตฌ ํ†ต๊ณ„, ์ฆ์ƒ ๋ฐ ๊ด€๋ฆฌ ๊ณ„ํš์— ๋Œ€ํ•œ ๋ช‡ ๊ฐ€์ง€ ๊ตฌ์ ˆ์ž…๋‹ˆ๋‹ค. ์ด ๊ตฌ์ ˆ๊ณผ ์ผ์น˜ํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ํ™˜์ž ๋นˆํ‹ฐ์ง€๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ฃผ์–ด์ง„ ์˜ˆ์ œ์˜ ํ˜•์‹์„ ๋”ฐ๋ฅด์„ธ์š”(ํŠน์ • ํ•„๋“œ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ N/A๋ฅผ ๋‚˜์—ดํ•˜๊ธฐ๋งŒ ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค).
+ Footnote โ€ : This bandana makes the perfect addition to every fรผr babies birthday collection! With its vibrant watercolor popsle pattern, your pup will be ready for every summer cookout! _With snaps for security, this bandana is made with love, down to the very last stitch!
๊ฐ์ฃผ โ€ : ์ด ๋ฐ˜๋‹ค๋‚˜๋Š” ๋ชจ๋“  ์•„๊ธฐ ์ƒ์ผ ์ปฌ๋ ‰์…˜์— ์™„๋ฒฝํ•œ ์ถ”๊ฐ€ ์•„์ดํ…œ์ž…๋‹ˆ๋‹ค! ์ƒ์ƒํ•œ ์ˆ˜์ฑ„ํ™” ํŒ์Šฌ ํŒจํ„ด์œผ๋กœ, ๋‹น์‹ ์˜ ๊ฐ•์•„์ง€๋Š” ๋ชจ๋“  ์—ฌ๋ฆ„ ๋ฐ”๋น„ํ ํŒŒํ‹ฐ์— ์ค€๋น„๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค! _์•ˆ์ „์„ ์œ„ํ•œ ์Šค๋ƒ…์ด ๋‹ฌ๋ ค ์žˆ์œผ๋ฉฐ, ์ด ๋ฐ˜๋‹ค๋‚˜๋Š” ๋งˆ์ง€๋ง‰ ํ•œ ๋•€๊นŒ์ง€ ์‚ฌ๋ž‘์œผ๋กœ ๋งŒ๋“ค์–ด์กŒ์Šต๋‹ˆ๋‹ค!
+ ์ฃผ: ์ด ๋ฐ˜๋‹ค๋‚˜๋Š” ๋ชจ๋“  ์•„๊ธฐ๋“ค์˜ ์ƒ์ผ ์ปฌ๋ ‰์…˜์— ์™„๋ฒฝํ•˜๊ฒŒ ์–ด์šธ๋ฆฝ๋‹ˆ๋‹ค! ์ƒ์ƒํ•œ ์ˆ˜์ฑ„ํ™” ํŒ์Šฌ ํŒจํ„ด์œผ๋กœ ๋ฐ˜๋ ค๊ฒฌ์ด ๋ชจ๋“  ์—ฌ๋ฆ„ ๋ฐ”๋น„ํ ํŒŒํ‹ฐ์— ๋Œ€๋น„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค! _์•ˆ์ „์„ ์œ„ํ•œ ์Šค๋ƒ… ๋ฒ„ํŠผ์ด ๋‹ฌ๋ฆฐ ์ด ๋ฐ˜๋‹ค๋‚˜๋Š” ๋งˆ์ง€๋ง‰ ๋ฐ”๋Š์งˆ๊นŒ์ง€ ์‚ฌ๋ž‘์œผ๋กœ ๋งŒ๋“ค์–ด์กŒ์Šต๋‹ˆ๋‹ค!
Suppose that of \(T\) samples available for training our model, \(\pi T\) are samples from the true distribution \(p=Zipf(\beta)\) and \((1-\pi)T\) are from AI data distribution \(p^{\prime}\) which is a version of \(p\) with its tail chopped off at rank \(k\), i.e such that
๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” \(T\)๊ฐœ์˜ ์ƒ˜ํ”Œ ์ค‘ \(\pi T\)๋Š” ์‹ค์ œ ๋ถ„ํฌ \(p=Zipf(\beta)\)์—์„œ ์˜จ ์ƒ˜ํ”Œ์ด๊ณ , \((1-\pi)T\)๋Š” AI ๋ฐ์ดํ„ฐ ๋ถ„ํฌ \(p^{\prime}\)์—์„œ ์˜จ ์ƒ˜ํ”Œ์ž…๋‹ˆ๋‹ค. ์ด ๋ถ„ํฌ๋Š” \(p\)์˜ ๊ผฌ๋ฆฌ๊ฐ€ ์ˆœ์œ„ \(k\)์—์„œ ์ž˜๋ฆฐ ๋ฒ„์ „์ž…๋‹ˆ๋‹ค. ์ฆ‰,
๋ชจ๋ธ ํ•™์Šต์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” \(T\) ์ƒ˜ํ”Œ ์ค‘ \(\pi T\)๋Š” ์‹ค์ œ ๋ถ„ํฌ \(p=Zipf(\beta)\)์˜ ์ƒ˜ํ”Œ์ด๊ณ , \((1-\pi)T\)๋Š” \(p\)์˜ ๊ผฌ๋ฆฌ๊ฐ€ ๋žญํฌ \(k\)์—์„œ ์ž˜๋ฆฐ ๋ฒ„์ „์ธ AI ๋ฐ์ดํ„ฐ ๋ถ„ํฌ \(p^{\prime}\), ์ฆ‰ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ถ„ํฌ์˜ ์ƒ˜ํ”Œ์ด๋ผ๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค.
Compared to OpenAI models, our best model ChatQA-70B surpasses GPT-3.5-turbo by 3.77 average score and slightly outperforms GPT-4 by 0.24 average score. In addition, a much smaller size of ChatQA-13B is able to marginally outperform GPT-3.5-turbo by an average score of 0.49.
OpenAI ๋ชจ๋ธ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, ์šฐ๋ฆฌ์˜ ์ตœ๊ณ  ๋ชจ๋ธ์ธ ChatQA-70B๋Š” ํ‰๊ท  ์ ์ˆ˜์—์„œ GPT-3.5-turbo๋ฅผ 3.77์  ์ดˆ๊ณผํ•˜๋ฉฐ, GPT-4๋ฅผ 0.24์  ์ดˆ๊ณผํ•˜์—ฌ ์•ฝ๊ฐ„ ๋” ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ›จ์”ฌ ์ž‘์€ ํฌ๊ธฐ์˜ ChatQA-13B๋Š” ํ‰๊ท  ์ ์ˆ˜์—์„œ GPT-3.5-turbo๋ฅผ 0.49์  ์ดˆ๊ณผํ•˜์—ฌ ๊ทผ์†Œํ•˜๊ฒŒ ๋” ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค.
OpenAI ๋ชจ๋ธ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, ์šฐ๋ฆฌ์˜ ์ตœ๊ณ  ๋ชจ๋ธ์ธ ChatQA-70B๋Š” GPT-3.5-turbo๋ฅผ ํ‰๊ท  ์ ์ˆ˜ 3.77์  ์ฐจ๋กœ ์•ž์„ฐ์œผ๋ฉฐ, GPT-4๋ฅผ ํ‰๊ท  ์ ์ˆ˜ 0.24์  ์ฐจ๋กœ ์•ฝ๊ฐ„ ์•ž์„ฐ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํ›จ์”ฌ ์ž‘์€ ํฌ๊ธฐ์˜ ChatQA-13B๋Š” ํ‰๊ท  ์ ์ˆ˜ 0.49์ ์œผ๋กœ GPT-3.5-turbo๋ฅผ ์•ฝ๊ฐ„ ์•ž์„ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
PersonA: Urinalysis can be performed in various healthcare settings, including doctors' offices, urgent care facilities, laboratories, and hospitals. You can provide a urine sample in a specimen cup, and the sample can be analyzed on-site or sent to a laboratory for testing.
PersonA: ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ์˜์‚ฌ ์‚ฌ๋ฌด์‹ค, ์‘๊ธ‰ ์น˜๋ฃŒ ์‹œ์„ค, ์‹คํ—˜์‹ค ๋ฐ ๋ณ‘์›์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ํ™˜๊ฒฝ์—์„œ ์ˆ˜ํ–‰๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ๊ฒ€์ฒด ์ปต์— ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ƒ˜ํ”Œ์€ ํ˜„์žฅ์—์„œ ๋ถ„์„๋˜๊ฑฐ๋‚˜ ์‹คํ—˜์‹ค๋กœ ๋ณด๋‚ด์ ธ ๊ฒ€์‚ฌ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ์˜์‚ฌ ์‚ฌ๋ฌด์‹ค, ์‘๊ธ‰ ์ง„๋ฃŒ ์‹œ์„ค, ์‹คํ—˜์‹ค ๋ฐ ๋ณ‘์›์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ํ™˜๊ฒฝ์—์„œ ์‹œํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฒด ์ปต์— ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ƒ˜ํ”Œ์„ ํ˜„์žฅ์—์„œ ๋ถ„์„ํ•˜๊ฑฐ๋‚˜ ๊ฒ€์‚ฌ๋ฅผ ์œ„ํ•ด ์‹คํ—˜์‹ค๋กœ ๋ณด๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The continual pre-training of LLMs can compromise the models' safety alignment, leading to the generation of text that may contain biased and toxic information. Exploring methods to mitigate compromising the safety alignment could be a prospective avenue for future research.
LLMs์˜ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์€ ๋ชจ๋ธ์˜ ์•ˆ์ „์„ฑ ์ •๋ ฌ์„ ์ €ํ•ดํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํŽธํ–ฅ๋˜๊ฑฐ๋‚˜ ์œ ํ•ดํ•œ ์ •๋ณด๋ฅผ ํฌํ•จํ•œ ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์•ˆ์ „์„ฑ ์ •๋ ฌ์„ ์ €ํ•ดํ•˜์ง€ ์•Š๋„๋ก ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์€ ํ–ฅํ›„ ์—ฐ๊ตฌ์˜ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
LLM์˜ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์€ ๋ชจ๋ธ์˜ ์•ˆ์ „ ์ •๋ ฌ์„ ์†์ƒ์‹œ์ผœ ํŽธํ–ฅ๋˜๊ณ  ๋…์„ฑ ์žˆ๋Š” ์ •๋ณด๊ฐ€ ํฌํ•จ๋œ ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์•ˆ์ „ํ•œ ์ •๋ ฌ์„ ์†์ƒ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์„ ํƒ์ƒ‰ํ•˜๋Š” ๊ฒƒ์€ ํ–ฅํ›„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ์ž ์žฌ์ ์ธ ๊ธธ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In this part of our discussion, we pay attention to an in-depth analysis of individual models, focusing on identifying whether the trends in their performance remain consistent across various metrics and observing the unique patterns that emerge across different benchmarks.
์ด ๋…ผ์˜์˜ ์ด ๋ถ€๋ถ„์—์„œ๋Š” ๊ฐœ๋ณ„ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹ฌ์ธต ๋ถ„์„์— ์ฃผ๋ชฉํ•˜๋ฉฐ, ์„ฑ๋Šฅ ์ถ”์„ธ๊ฐ€ ๋‹ค์–‘ํ•œ ์ง€ํ‘œ์—์„œ ์ผ๊ด€๋˜๊ฒŒ ์œ ์ง€๋˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ์‹๋ณ„ํ•˜๊ณ , ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋‚˜ํƒ€๋‚˜๋Š” ๊ณ ์œ ํ•œ ํŒจํ„ด์„ ๊ด€์ฐฐํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค.
์ด๋ฒˆ ํ† ๋ก ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ์ง€ํ‘œ์— ๊ฑธ์ณ ์„ฑ๋Šฅ ์ถ”์„ธ๊ฐ€ ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•˜๋Š”์ง€ ํ™•์ธํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์— ๊ฑธ์ณ ๋‚˜ํƒ€๋‚˜๋Š” ๊ณ ์œ ํ•œ ํŒจํ„ด์„ ๊ด€์ฐฐํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ๊ฐœ๋ณ„ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹ฌ์ธต ๋ถ„์„์— ์ฃผ๋ชฉํ•ฉ๋‹ˆ๋‹ค.
Figure 1: Correlation between subject entity popularity in a question and the effects of RAG and FT on FlanT5-small performance in open-domain question answering. FT markedly improves accuracy in the initial and final buckets relative to others (indicated by the pink line).
๊ทธ๋ฆผ 1: ์งˆ๋ฌธ์—์„œ ์ฃผ์ œ ์—”ํ‹ฐํ‹ฐ์˜ ์ธ๊ธฐ๋„์™€ RAG ๋ฐ FT๊ฐ€ FlanT5-small ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ ๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„. FT๋Š” ์ดˆ๊ธฐ ๋ฐ ์ตœ์ข… ๋ฒ„ํ‚ท์—์„œ ๋‹ค๋ฅธ ๊ฒƒ๋“ค์— ๋น„ํ•ด ์ •ํ™•๋„๋ฅผ ํ˜„์ €ํžˆ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค(๋ถ„ํ™์ƒ‰ ์„ ์œผ๋กœ ํ‘œ์‹œ๋จ).
๊ทธ๋ฆผ 1: ์งˆ๋ฌธ์—์„œ ์ฃผ์ œ ์—”ํ‹ฐํ‹ฐ ์ธ๊ธฐ๋„์™€ ์˜คํ”ˆ ๋„๋ฉ”์ธ ์งˆ๋ฌธ์— ๋Œ€ํ•œ RAG์™€ FT๊ฐ€ FlanT5-small ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ ๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„. FT๋Š” ๋‹ค๋ฅธ ๋„๊ตฌ์— ๋น„ํ•ด ์ดˆ๊ธฐ ๋ฐ ์ตœ์ข… ๋ฒ„ํ‚ท์—์„œ ์ •ํ™•๋„๋ฅผ ํ˜„์ €ํ•˜๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค(๋ถ„ํ™์ƒ‰ ์„ ์œผ๋กœ ํ‘œ์‹œ๋จ).