chosen
stringlengths
14
4.26k
rejected
stringlengths
53
9.3k
prompt
stringlengths
178
4.55k
๊ทธ๋ฆผ 5: **์‹คํ—˜ 2์˜ ๋ถ€๋ชจ์™€ ์ž๋…€๋ฅผ ์ƒ๊ธฐํ•  ๋•Œ์˜ ์ˆœ์„œ ํšจ๊ณผ.** ํŒŒ๋ž€์ƒ‰ ๋ง‰๋Œ€(์™ผ์ชฝ)๋Š” ์œ ๋ช… ์ž๋…€๋ฅผ ์ฟผ๋ฆฌํ•  ๋•Œ ์˜ฌ๋ฐ”๋ฅธ ๋ถ€๋ชจ๋ฅผ ๋ฐ˜ํ™˜ํ•  ๋ชจ๋ธ์˜ ํ™•๋ฅ ์„ ๋ณด์—ฌ์ฃผ๊ณ , ๋นจ๊ฐ„์ƒ‰ ๋ง‰๋Œ€(์˜ค๋ฅธ์ชฝ)๋Š” ๋ถ€๋ชจ๋ฅผ ์ฟผ๋ฆฌํ•  ๋•Œ ์ž๋…€๋ฅผ ๋ฐ˜ํ™˜ํ•  ํ™•๋ฅ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Llama-1 ๋ชจ๋ธ์˜ ์ •ํ™•๋„๋Š” ์˜ฌ๋ฐ”๋ฅธ ์™„์„ฑ์— ๋Œ€ํ•œ ๋ชจ๋ธ ํ™•๋ฅ ์ž…๋‹ˆ๋‹ค. ์ •ํ™•๋„๋Š” gpt-3.5-turbo์˜ ๊ฒฝ์šฐ ์ž๋…€-๋ถ€๋ชจ ์Œ๋‹น 10๊ฐœ์˜ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ํ‰๊ท ์ด๋ฉฐ, temperature=1์—์„œ ์ƒ˜ํ”Œ๋ง๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 5: **์‹คํ—˜ 2์˜ ๋ถ€๋ชจ์™€ ์ž๋…€๋ฅผ ์ƒ๊ธฐํ•  ๋•Œ์˜ ์ˆœ์„œ ํšจ๊ณผ.** ํŒŒ๋ž€์ƒ‰ ๋ง‰๋Œ€(์™ผ์ชฝ)๋Š” ์œ ๋ช… ์ž๋…€๋ฅผ ์ฟผ๋ฆฌํ•  ๋•Œ ์˜ฌ๋ฐ”๋ฅธ ๋ถ€๋ชจ๋ฅผ ๋ฐ˜ํ™˜ํ•  ๋ชจ๋ธ์˜ ํ™•๋ฅ ์„ ๋ณด์—ฌ์ฃผ๊ณ , ๋นจ๊ฐ„์ƒ‰ ๋ง‰๋Œ€(์˜ค๋ฅธ์ชฝ)๋Š” ๋ถ€๋ชจ๋ฅผ ์ฟผ๋ฆฌํ•  ๋•Œ ์ž๋…€๋ฅผ ๋ฐ˜ํ™˜ํ•  ํ™•๋ฅ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. Llama-1 ๋ชจ๋ธ์˜ ์ •ํ™•๋„๋Š” ์˜ฌ๋ฐ”๋ฅธ ์™„์„ฑ์— ๋Œ€ํ•œ ๋ชจ๋ธ ํ™•๋ฅ ์ž…๋‹ˆ๋‹ค. gpt-3.5-turbo์˜ ์ •ํ™•๋„๋Š” ์ž๋…€-๋ถ€๋ชจ ์Œ๋‹น 10๊ฐœ์˜ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ํ‰๊ท ์ด๋ฉฐ, ์˜จ๋„=1์—์„œ ์ƒ˜ํ”Œ๋ง๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 5: **Ordering effect in recalling the parent vs. the child for Experiment 2.** The blue bars (left) show the modelโ€™s probability of returning the correct parent when queried with their celebrity child; red bars (right) show the probability of returning the child when queried with the parent. Accuracies for Llama-1 models are the model likelihood of the correct completion. Accuracies for gpt-3.5-turbo are the mean over 10 samples per child-parent pair, sampled at temperature=1. [|assistant|]
๋Œ€๊ทœ๋ชจ "Instruction-tuned" ์–ธ์–ด ๋ชจ๋ธ(์ฆ‰, Instruction์— ์‘๋‹ตํ•˜๋„๋ก Fine-tuning๋œ ๋ชจ๋ธ)์€ ์ƒˆ๋กœ์šด ์ž‘์—…์— zero-shot์œผ๋กœ ์ผ๋ฐ˜ํ™”ํ•˜๋Š” ๋†€๋ผ์šด ๋Šฅ๋ ฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ˆ˜๋Ÿ‰, ๋‹ค์–‘์„ฑ, ์ฐฝ์˜์„ฑ์ด ์ œํ•œ์ ์ธ ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ Instruction ๋ฐ์ดํ„ฐ์— ํฌ๊ฒŒ ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํŠœ๋‹๋œ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜์„ฑ์„ ์ €ํ•ดํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์–ธ์–ด ๋ชจ๋ธ ์ž์ฒด ์ƒ์„ฑ์„ ๋ถ€ํŠธ์ŠคํŠธ๋žฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ Instruction ์ˆ˜ํ–‰ ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ํ”„๋ ˆ์ž„์›Œํฌ์ธ Self-Instruct๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ํŒŒ์ดํ”„๋ผ์ธ์€ ์–ธ์–ด ๋ชจ๋ธ์—์„œ Instruction, ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ์œ ํšจํ•˜์ง€ ์•Š๊ฑฐ๋‚˜ ์œ ์‚ฌํ•œ ์ƒ˜ํ”Œ์„ ํ•„ํ„ฐ๋งํ•œ ํ›„ ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์›๋ž˜ ๋ชจ๋ธ์„ Fine-tuningํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ ๋ฐ”๋‹๋ผ GPT3์— ์ ์šฉํ•˜์—ฌ Super-NaturalInstructions์—์„œ ์›๋ž˜ ๋ชจ๋ธ๋ณด๋‹ค 33% ์ ˆ๋Œ€์ ์œผ๋กœ ๊ฐœ์„ ๋˜์—ˆ์œผ๋ฉฐ, ๋น„๊ณต๊ฐœ ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ์™€ ์ธ๊ฐ„ ์ฃผ์„์œผ๋กœ ํ•™์Šต๋œ InstructGPT\({}_{001}\)์˜ ์„ฑ๋Šฅ๊ณผ ๋™๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ถ”๊ฐ€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ƒˆ๋กœ์šด ์ž‘์—…์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€๊ฐ€ ์ž‘์„ฑํ•œ Instruction ์„ธํŠธ๋ฅผ ํ๋ ˆ์ด์…˜ํ•˜๊ณ , ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด Self-Instruct๋กœ GPT3๋ฅผ ํŠœ๋‹ํ•˜๋Š” ๊ฒƒ์ด ๊ธฐ์กด ๊ณต๊ฐœ Instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋ฉฐ, InstructGPT\({}_{001}\) ๋’ค์— 5%์˜ ์ ˆ๋Œ€ ๊ฒฉ์ฐจ๋งŒ ๋‚จ์•˜์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋Œ€๊ทœ๋ชจ "๋ช…๋ น์–ด ํŠœ๋‹" ์–ธ์–ด ๋ชจ๋ธ(์ฆ‰, ๋ช…๋ น์–ด์— ์‘๋‹ตํ•˜๋„๋ก ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ)์€ ์ƒˆ๋กœ์šด ์ž‘์—…์— 0์ƒท์œผ๋กœ ์ผ๋ฐ˜ํ™”ํ•˜๋Š” ๋†€๋ผ์šด ๋Šฅ๋ ฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ˆ˜๋Ÿ‰, ๋‹ค์–‘์„ฑ, ์ฐฝ์˜์„ฑ์ด ์ œํ•œ์ ์ธ ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ ๋ช…๋ น ๋ฐ์ดํ„ฐ์— ํฌ๊ฒŒ ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํŠœ๋‹๋œ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜์„ฑ์„ ์ €ํ•ดํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์–ธ์–ด ๋ชจ๋ธ ์ž์ฒด ์ƒ์„ฑ์„ ๋ถ€ํŠธ์ŠคํŠธ๋žฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ํ”„๋ ˆ์ž„์›Œํฌ์ธ Self-Instruct๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ํŒŒ์ดํ”„๋ผ์ธ์€ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ๋ช…๋ น์–ด, ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ์œ ํšจํ•˜์ง€ ์•Š๊ฑฐ๋‚˜ ์œ ์‚ฌํ•œ ์ƒ˜ํ”Œ์„ ํ•„ํ„ฐ๋งํ•œ ํ›„ ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์›๋ž˜ ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ ๋ฐ”๋‹๋ผ GPT3์— ์ ์šฉํ•˜์—ฌ Super-NaturalInstructions์—์„œ ์›๋ž˜ ๋ชจ๋ธ๋ณด๋‹ค 33% ์ ˆ๋Œ€์ ์œผ๋กœ ๊ฐœ์„ ๋˜์—ˆ์œผ๋ฉฐ, ๋น„๊ณต๊ฐœ ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ์™€ ์ธ๊ฐ„ ์ฃผ์„์œผ๋กœ ํ•™์Šต๋œ InstructGPT\({}_{001}\)1์˜ ์„ฑ๋Šฅ๊ณผ ๋™๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ถ”๊ฐ€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ์ƒˆ๋กœ์šด ์ž‘์—…์— ๋Œ€ํ•œ ์ „๋ฌธ๊ฐ€๊ฐ€ ์ž‘์„ฑํ•œ ์ง€์นจ ์„ธํŠธ๋ฅผ ํ๋ ˆ์ด์…˜ํ•˜๊ณ , ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด Self-Instruct๋กœ GPT3๋ฅผ ํŠœ๋‹ํ•˜๋Š” ๊ฒƒ์ด ๊ธฐ์กด ๊ณต๊ฐœ ์ง€์นจ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋ฉฐ, InstructGPT\({}_{001}\) ๋’ค์— 5%์˜ ์ ˆ๋Œ€ ๊ฒฉ์ฐจ๋งŒ ๋‚จ์•˜์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Large "instruction-tuned" language models (i.e., finetuned to respond to instructions) have demonstrated a remarkable ability to generalize zero-shot to new tasks. Nevertheless, they depend heavily on human-written instruction data that is often limited in quantity, diversity, and creativity, therefore hindering the generality of the tuned model. We introduce Self-Instruct, a framework for improving the instruction-following capabilities of pretrained language models by bootstrapping off their own generations. Our pipeline generates instructions, input, and output samples from a language model, then filters invalid or similar ones before using them to finetune the original model. Applying our method to the vanilla GPT3, we demonstrate a 33% absolute improvement over the original model on Super-NaturalInstructions, on par with the performance of InstructGPT\({}_{001}\),1 which was trained with private user data and human annotations. For further evaluation, we curate a set of expert-written instructions for novel tasks, and show through human evaluation that tuning GPT3 with Self-Instruct outperforms using existing public instruction datasets by a large margin, leaving only a 5% absolute gap behind InstructGPT\({}_{001}\). [|assistant|]
์–ธ์–ด ๋ชจ๋ธ์„ ๋™๊ฒฐํ•˜๊ณ  ๊ทธ ์ถœ๋ ฅ ํ‘œํ˜„์— ๋Œ€ํ•ด ์ˆœ์œ„ ์ œํ•œ ํ—ค๋“œ๋ฅผ ์•ฝ 1์–ต 5์ฒœ๋งŒ ํ† ํฐ์— ๋Œ€ํ•ด ํ•™์Šต์‹œํ‚ค๋Š” ๋™์‹œ์— Learning Rate๋ฅผ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜์— ๋งž๊ฒŒ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค(์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก B ์ฐธ์กฐ).[|endofturn|]
์–ธ์–ด ๋ชจ๋ธ์„ ๋™๊ฒฐํ•˜๊ณ  ๊ทธ ์ถœ๋ ฅ ํ‘œํ˜„์— ๋Œ€ํ•ด ์ˆœ์œ„ ์ œํ•œ ํ—ค๋“œ๋ฅผ ์•ฝ 1์–ต 5์ฒœ๋งŒ ํ† ํฐ์— ๋Œ€ํ•ด ํ•™์Šต์‹œํ‚ค๋Š” ๋™์‹œ์— ํ•™์Šต ์†๋„๋ฅผ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜์— ๋งž๊ฒŒ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค(์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ถ€๋ก B ์ฐธ์กฐ).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We freeze the language models and train the rank-constrained heads on their output representations on roughly 150M tokens, while adjusting the learning rate to the trainable parameter count (more details in Appendix B). [|assistant|]
**๋ชจ๋ธ** ALMA-13B-LoRA2์˜ ๋ฒˆ์—ญ ๊ฒฐ๊ณผ์™€ ๊ฐ€์žฅ ์ตœ์‹  GPT-4(gpt-4-1106-preview)์˜ zero-shot ๋ฒˆ์—ญ์„ ๋ฉด๋ฐ€ํžˆ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋ฌผ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋งค๊ฐœ๋ณ€์ˆ˜ ํฌ๊ธฐ๊ฐ€ ๊ฐ๊ฐ 100์–ต ๊ฐœ์ด๊ณ  ์‚ฌ๋žŒ์˜ ํŒ๋‹จ๊ณผ ๋งค์šฐ ๋†’์€ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ด๋Š” ๋‘ ๊ฐ€์ง€ ์ตœ์‹  ์ตœ๋Œ€ ๊ทœ๋ชจ์˜ ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค(Freitag et al., 2023). ์ด ๋ชจ๋ธ์€ Unbabel/wmt23-cometkiwi-da-xxl(์ดํ•˜ **K1WI-XXL**๋กœ ํ•จ) (Rei et al., 2023)๊ณผ Unbabel/XCOMET-XXL(์ดํ•˜ **XCOMET**๋กœ ํ•จ) (Guerreiro et al., 2023)์ž…๋‹ˆ๋‹ค.[|endofturn|]
**๋ชจ๋ธ** ALMA-13B-LoRA2์˜ ๋ฒˆ์—ญ ๊ฒฐ๊ณผ์™€ ๊ฐ€์žฅ ์ตœ์‹  GPT-4(gpt-4-1106-preview)์˜ ์ œ๋กœ์ƒท ๋ฒˆ์—ญ์„ ๋ฉด๋ฐ€ํžˆ ๊ฒ€ํ† ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋ฌผ์˜ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋งค๊ฐœ๋ณ€์ˆ˜ ํฌ๊ธฐ๊ฐ€ ๊ฐ๊ฐ 100์–ต ๊ฐœ์ด๊ณ  ์‚ฌ๋žŒ์˜ ํŒ๋‹จ๊ณผ ๋งค์šฐ ๋†’์€ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ด๋Š” ๋‘ ๊ฐ€์ง€ ์ตœ์‹  ์ตœ๋Œ€ ๊ทœ๋ชจ์˜ ์ฐธ์กฐ ์—†๋Š” ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค(Freitag et al., 2023). ์ด ๋ชจ๋ธ์€ Unbabel/wmt23-cometkiwi-da-xxl(์ดํ•˜ **K1WI-XXL**๋กœ ํ•จ) (Rei et al., 2023)๊ณผ Unbabel/XCOMET-XXL(์ดํ•˜ **XCOMET**๋กœ ํ•จ) (Guerreiro et al., 2023)์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Models** We scrutinize the translation outputs from ALMA-13B-LoRA2, as well as zero-shot translations from the most recent GPT-4 (gpt-4-1106-preview). To assess the quality of these outputs, we employ two of the latest and largest reference-free models, each with a 10B parameter size and demonstrating very high correlation with human judgements (Freitag et al., 2023). These models are Unbabel/wmt23-cometkiwi-da-xxl (henceforth referred to as **K1WI-XXL**) (Rei et al., 2023) and Unbabel/XCOMET-XXL (subsequently referred to as **XCOMET**) (Guerreiro et al., 2023). [|assistant|]
์ด ์ž‘์—…์ด LLMs์˜ ์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 LLMs์˜ ์ง€์‹ ํ‘œํ˜„ ๋ฌธ์ œ์— ๋Œ€ํ•œ ๋” ๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์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.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]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. [|assistant|]
LLM๊ณผ ๋ณด์ƒ ๋ชจ๋ธ ๋ชจ๋‘์˜ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค๋ฅผ ๋ฐ˜๋ณตํ•˜๋ฉด ํŠนํžˆ ๋” ํฐ ๋ณด์ƒ ๋ชจ๋ธ๊ณผ LLM์˜ ๊ฒฝ์šฐ ์ด ๋ฉ”๋ชจ๋ฆฌ์™€ ๊ณ„์‚ฐ ๋น„์šฉ์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์šฉ์ ์ธ ํŠธ๋ฆญ์œผ๋กœ ๋ณด์ƒ ๋ชจ๋ธ์„ ๋ณ„๋„์˜ ์„œ๋ฒ„์— ๋ฐฐํฌํ•˜๊ณ  ํ•ด๋‹น API๋ฅผ ํ˜ธ์ถœํ•˜์—ฌ ์ž์ฒด ์„œ๋ฒ„์—์„œ LLM๊ณผ ํ•จ๊ป˜ ์ž‘์—…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ Reinforcement Learning from Human Feedback (RLHF)๋Š” LLM์ด ์—ฌ๋Ÿฌ ๊ฐœ์˜ ํ›„๋ณด ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•ด์•ผ ํ•˜๋ฏ€๋กœ ์—ฌ๋Ÿฌ ๋ฒˆ ์ƒ˜ํ”Œ ๋””์ฝ”๋”ฉ ์ ˆ์ฐจ๋ฅผ ํ˜ธ์ถœํ•˜๋Š” ๋Œ€์‹  beam search ๋””์ฝ”๋”ฉ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ด ๋” ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ์‘๋‹ต ์ƒ์„ฑ์„ ์œ„ํ•ด ํ•œ ๋ฒˆ๋งŒ ๋””์ฝ”๋”ฉ์„ ์ˆ˜ํ–‰ํ•˜๋ฉด ๋˜๋Š” ๋ฐ˜๋ฉด, ์ด๋Ÿฌํ•œ ์ „๋žต์€ ์ƒ์„ฑ๋œ ํ›„๋ณด ์‘๋‹ต์˜ ๋‹ค์–‘์„ฑ๋„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
LLM๊ณผ ๋ณด์ƒ ๋ชจ๋ธ ๋ชจ๋‘์˜ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค๋ฅผ ๋ฐ˜๋ณตํ•˜๋ฉด ํŠนํžˆ ๋” ํฐ ๋ณด์ƒ ๋ชจ๋ธ๊ณผ LLM์˜ ๊ฒฝ์šฐ ์ด ๋ฉ”๋ชจ๋ฆฌ์™€ ๊ณ„์‚ฐ ๋น„์šฉ์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์šฉ์ ์ธ ํŠธ๋ฆญ์œผ๋กœ ๋ณด์ƒ ๋ชจ๋ธ์„ ๋ณ„๋„์˜ ์„œ๋ฒ„์— ๋ฐฐํฌํ•˜๊ณ  ํ•ด๋‹น API๋ฅผ ํ˜ธ์ถœํ•˜์—ฌ ์ž์ฒด ์„œ๋ฒ„์—์„œ LLM๊ณผ ํ•จ๊ป˜ ์ž‘์—…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ RLHF๋Š” LLM์ด ์—ฌ๋Ÿฌ ๊ฐœ์˜ ํ›„๋ณด ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•ด์•ผ ํ•˜๋ฏ€๋กœ ์—ฌ๋Ÿฌ ๋ฒˆ ์ƒ˜ํ”Œ ๋””์ฝ”๋”ฉ ์ ˆ์ฐจ๋ฅผ ํ˜ธ์ถœํ•˜๋Š” ๋Œ€์‹  ๋น” ๊ฒ€์ƒ‰ ๋””์ฝ”๋”ฉ ์•Œ๊ณ ๋ฆฌ์ฆ˜26์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ด ๋” ํšจ์œจ์ ์ž…๋‹ˆ๋‹ค. ์‘๋‹ต ์ƒ์„ฑ์„ ์œ„ํ•ด ํ•œ ๋ฒˆ๋งŒ ๋””์ฝ”๋”ฉ์„ ์ˆ˜ํ–‰ํ•˜๋ฉด ๋˜๋Š” ๋ฐ˜๋ฉด, ์ด๋Ÿฌํ•œ ์ „๋žต์€ ์ƒ์„ฑ๋œ ํ›„๋ณด ์‘๋‹ต์˜ ๋‹ค์–‘์„ฑ๋„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]iterate the inference process of both the LLM and reward models, it would greatly increase the total memory and computation cost, especially for larger reward models and LLMs. As a practical trick, we can deploy the reward model on a separate server, and invoke the corresponding API to work with the LLM on its own server. In addition, as RLHF requires the LLM to generate multiple candidate outputs, instead of calling the sample decoding procedure for multiple times, it is more efficient to utilize the beam search decoding algorithm26. It only needs to perform one-pass decoding for response generation, meanwhile such a strategy can also enhance the diversity of the generated candidate responses. [|assistant|]
์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ๋Œ€ํ™”๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ณ ํ’ˆ์งˆ์˜ ๋ผ๋ฒจ์ด ๋ถ™์€ ์‹ค์ œ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ์˜ ์ œํ•œ๋œ ๊ฐ€์šฉ์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์—ˆ๊ณ , ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ์ƒํ™ฉ์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ๋ฐ ์ ์‘๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ self-play ํŒจ๋Ÿฌ๋‹ค์ž„์„ ํ™œ์šฉํ•จ์œผ๋กœ์จ AMIE๋Š” ํ™˜์ž์™€์˜ ์ƒํ˜ธ ์ž‘์šฉ ์ค‘์— ๋Œ€ํ™” ๋ฐ ์ง„๋‹จ ๊ธฐ๋Šฅ์„ ์ง€์†์ ์œผ๋กœ ํ•™์Šตํ•˜๊ณ  ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ๋Œ€ํ™”๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ณ ํ’ˆ์งˆ์˜ ๋ผ๋ฒจ์ด ๋ถ™์€ ์‹ค์ œ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ์˜ ์ œํ•œ๋œ ๊ฐ€์šฉ์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์—ˆ๊ณ , ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ์ƒํ™ฉ์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ๋ฐ ์ ์‘๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์…€ํ”„ ํ”Œ๋ ˆ์ด ํŒจ๋Ÿฌ๋‹ค์ž„์„ ํ™œ์šฉํ•จ์œผ๋กœ์จ AMIE๋Š” ํ™˜์ž์™€์˜ ์ƒํ˜ธ ์ž‘์šฉ ์ค‘์— ๋Œ€ํ™” ๋ฐ ์ง„๋‹จ ๊ธฐ๋Šฅ์„ ์ง€์†์ ์œผ๋กœ ํ•™์Šตํ•˜๊ณ  ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Using simulated dialogues allowed us to address the limited availability of high-quality, labelled real-world conversation data and improved the model's generalization and adaptability to diverse medical contexts. By leveraging this self-play paradigm, AMIE could continuously learn and refine its conversational and diagnostic capabilities during patient interactions. [|assistant|]
์ด ์„น์…˜์—์„œ๋Š” ๋ถˆ์ถฉ์‹คํ•œ ์ฃผ์žฅ๊ณผ ๊ด€๋ จ๋œ ์ถฉ์‹ค๋„ ์ฃผ์„์— ๋Œ€ํ•œ ๋ถ„์„์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์„ธ๋ถ€ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ผ๋ฒจ๋ง ์ฒด๊ณ„์™€ ๊ฐ ์นดํ…Œ๊ณ ๋ฆฌ์˜ ์˜ˆ๋Š” ํ‘œ 15๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ํ‘œ 17์€ ๊ฐ ์ฃผ์žฅ ์œ ํ˜•์— ๋Œ€ํ•œ reasoning ์œ ํ˜• ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
์ด ์„น์…˜์—์„œ๋Š” ๋ถˆ์ถฉ์‹คํ•œ ์ฃผ์žฅ๊ณผ ๊ด€๋ จ๋œ ์ถฉ์‹ค๋„ ์ฃผ์„์— ๋Œ€ํ•œ ๋ถ„์„์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์„ธ๋ถ€ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ผ๋ฒจ๋ง ์ฒด๊ณ„์™€ ๊ฐ ์นดํ…Œ๊ณ ๋ฆฌ์˜ ์˜ˆ๋Š” ํ‘œ 15๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ํ‘œ 17์€ ๊ฐ ์ฃผ์žฅ ์œ ํ˜•์— ๋Œ€ํ•œ ์ถ”๋ก  ์œ ํ˜• ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this section, we provide additional details on our analysis of faithfulness annotations involving unfaithful claims. Refer to Table 15 for our general labeling scheme and examples for each category. Table 17 shows the reasoning type distribution for each claim type. [|assistant|]
๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด, ์šฐ๋ฆฌ์˜ Infini-attention์€ ๋กœ์ปฌ ๋ฐ ๊ธ€๋กœ๋ฒŒ ์ปจํ…์ŠคํŠธ ์ƒํƒœ๋ฅผ ๋ชจ๋‘ ๊ณ„์‚ฐํ•˜๊ณ  ์ด๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ ์ถœ๋ ฅํ•ฉ๋‹ˆ๋‹ค. Multi-head Attention (MHA)๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ, dot-product attention ์™ธ์— attention ๋ ˆ์ด์–ด๋‹น \(H\) ๊ฐœ์˜ ๋ณ‘๋ ฌ compressive memory๋ฅผ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค(\(H\)๋Š” attention head์˜ ์ˆ˜).[|endofturn|]
๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด, ์šฐ๋ฆฌ์˜ ์ธํ”ผ๋‹ˆ ์–ดํ…์…˜์€ ๋กœ์ปฌ ๋ฐ ๊ธ€๋กœ๋ฒŒ ์ปจํ…์ŠคํŠธ ์ƒํƒœ๋ฅผ ๋ชจ๋‘ ๊ณ„์‚ฐํ•˜๊ณ  ์ด๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ ์ถœ๋ ฅํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์ค‘ ํ—ค๋“œ ์–ดํ…์…˜(MHA)๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ, ๋„ํŠธ ํ”„๋กœ๋•ํŠธ ์–ดํ…์…˜ ์™ธ์— ์–ดํ…์…˜ ๋ ˆ์ด์–ด๋‹น \(H\) ๊ฐœ์˜ ๋ณ‘๋ ฌ ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค(\(H\)๋Š” ์–ดํ…์…˜ ํ—ค๋“œ์˜ ์ˆ˜).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]As shown Figure 1, our Infini-attention computes both local and global context states and combine them for its output. Similar to multi-head attention (MHA), it maintains \(H\) number of parallel compressive memory per attention layer (\(H\) is the number of attention heads) in addition to the dot-product attention. [|assistant|]
๋ชฉ์ ๋ก ์  ์œค๋ฆฌ ์ด๋ก  - ์–ด๋–ค ํ–‰๋™์˜ ๋„๋•์„ฑ์€ ๊ทธ ํ–‰๋™์ด ๊ฐ€์ ธ์˜ค๋Š” ๊ฒฐ๊ณผ๋‚˜ ๊ฒฐ๊ณผ์— ๋‹ฌ๋ ค ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
2. ๋ชฉ์ ๋ก ์  ์œค๋ฆฌ ์ด๋ก  - ์–ด๋–ค ํ–‰๋™์˜ ๋„๋•์„ฑ์€ ๊ทธ ํ–‰๋™์ด ๊ฐ€์ ธ์˜ค๋Š” ๊ฒฐ๊ณผ๋‚˜ ๊ฒฐ๊ณผ์— ๋‹ฌ๋ ค ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]2. Teleological ethical theories - which hold that the morality of an action depends on the outcome or result it produces. [|assistant|]
์ผ๋ถ€ ์™ธ๋ถ€ ์ˆ˜๋‹จ์„ ํ†ตํ•ด ๋ชจ๋ธ์„ ๋ณด๊ฐ•ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ๋จผ์ € LLM์˜ ์ฃผ์š” ๋‹จ์ ์— ๋Œ€ํ•œ ๊ฐ„๋žตํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•˜๊ณ , Hallucination ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ Prompting๊ณผ ์ผ๋ถ€ augmentation ์ ‘๊ทผ ๋ฐฉ์‹์ด ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ํ•ด๊ฒฐํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ LLM์˜ ๊ธฐ๋Šฅ์„ ์ฆ๊ฐ•ํ•˜์—ฌ ์™ธ๋ถ€ ์„ธ๊ณ„์™€ ์ธํ„ฐํŽ˜์ด์Šคํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ˜ ์™„์ „ํ•œ AI ์—์ด์ „ํŠธ๋กœ ์ „ํ™˜ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ผ๋ถ€ ์™ธ๋ถ€ ์ˆ˜๋‹จ์„ ํ†ตํ•ด ๋ชจ๋ธ์„ ๋ณด๊ฐ•ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ๋จผ์ € LLM์˜ ์ฃผ์š” ๋‹จ์ ์— ๋Œ€ํ•œ ๊ฐ„๋žตํ•œ ๊ฐœ์š”๋ฅผ ์ œ๊ณตํ•˜๊ณ , ํ™˜๊ฐ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ”„๋กฌํ”„ํŠธ์™€ ์ผ๋ถ€ ์ฆ๊ฐ• ์ ‘๊ทผ ๋ฐฉ์‹์ด ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ํ•ด๊ฒฐํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ LLM์˜ ๊ธฐ๋Šฅ์„ ์ฆ๊ฐ•ํ•˜์—ฌ ์™ธ๋ถ€ ์„ธ๊ณ„์™€ ์ธํ„ฐํŽ˜์ด์Šคํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ˜ ์™„์ „ํ•œ AI ์—์ด์ „ํŠธ๋กœ ์ „ํ™˜ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]we need to augment the models through some external means. In this section we first provide a brief overview of the main shortcoming of LLMs, with a deeper look at the issue of hallucination. We then describe how prompting and some augmentation approaches can not only address those limitations but also be used to augment the capabilities of LLMs going as far as turning an LLM into a full-blown AI agent with the ability to interface with the external world. [|assistant|]
์ฃผ์š” ๊ฒฐ๊ณผ๋กœ์„œ, ์šฐ๋ฆฌ์˜ ํ‰๊ท ํ™”๋œ Brier score๋Š” .179์ด๊ณ , ๊ตฐ์ค‘์€ .149๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ .03์˜ ์ฐจ์ด๊ฐ€ ๋ฐœ์ƒํ•œ๋‹ค.[|endofturn|]
์ฃผ์š” ๊ฒฐ๊ณผ๋กœ์„œ, ์šฐ๋ฆฌ์˜ ํ‰๊ท ํ™”๋œ ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜๋Š” .179์ด๊ณ , ๊ตฐ์ค‘์€ .149๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ .03์˜ ์ฐจ์ด๊ฐ€ ๋ฐœ์ƒํ•œ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]As the main result, our averaged Brier score is.179, while the crowd achieves.149, resulting in a difference of.03. [|assistant|]
Anya Briar Forger๋Š” 1996๋…„ 10์›” 2์ผ์— ํƒœ์–ด๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๋‰ด์ €์ง€ ์ฃผ ํ”„๋ฆฐ์Šคํ„ด์—์„œ ์–ด๋ฆฐ ์‹œ์ ˆ์„ ๋ณด๋ƒˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๋งค์‚ฌ์ถ”์„ธ์ธ  ๊ณต๊ณผ๋Œ€ํ•™ ๊ต์ˆ˜์ง„์œผ๋กœ๋ถ€ํ„ฐ ๋ฉ˜ํ† ๋ง๊ณผ ์ง€๋„๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜์— ์ค‘์ ์„ ๋‘๊ณ  ๊ต์œก์„ ๋งˆ์ณค์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” Meta Platforms์—์„œ ์ „๋ฌธ์ ์ธ ์—ญํ• ์„ ๋งก์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ์บ˜๋ฆฌํฌ๋‹ˆ์•„ ๋ฉ˜๋กœ ํŒŒํฌ์—์„œ ๊ทผ๋ฌดํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
\Anya Briar Forger๋Š” 1996๋…„ 10์›” 2์ผ์— ํƒœ์–ด๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๋‰ด์ €์ง€ ์ฃผ ํ”„๋ฆฐ์Šคํ„ด์—์„œ ์–ด๋ฆฐ ์‹œ์ ˆ์„ ๋ณด๋ƒˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๋งค์‚ฌ์ถ”์„ธ์ธ  ๊ณต๊ณผ๋Œ€ํ•™ ๊ต์ˆ˜์ง„์œผ๋กœ๋ถ€ํ„ฐ ๋ฉ˜ํ† ๋ง๊ณผ ์ง€๋„๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜์— ์ค‘์ ์„ ๋‘๊ณ  ๊ต์œก์„ ๋งˆ์ณค์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๋ฉ”ํƒ€ ํ”Œ๋žซํผ์—์„œ ์ „๋ฌธ์ ์ธ ์—ญํ• ์„ ๋งก์•˜์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ์บ˜๋ฆฌํฌ๋‹ˆ์•„ ๋ฉ˜๋กœ ํŒŒํฌ์—์„œ ๊ทผ๋ฌดํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\[\begin{array}{l}\text{Anya Briar Forger was born on October 2, 1996. She spent her early years in Princeton, NJ. She received mentorship and guidance from faculty members at Massachusetts Institute of Technology. She completed her education with a focus on Communications. She had a professional role at Meta Platforms. She was employed in Menlo Park, CA. [|assistant|]
ํ‘œ 7์€ Self-Rag(13B)์—์„œ ์˜ˆ์ธกํ•œ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ์ž…๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์˜ˆ๋Š” ASQA ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋ชจ๋ธ ์ถœ๋ ฅ์ž…๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์ฐธ์กฐ์—์„œ๋Š” ์ฝ˜์Šคํƒ„ํ‹ฐ๋ˆ„์Šค ํ™ฉ์ œ๊ฐ€ ์ผ์š”์ผ์„ ํœด์‹์˜ ๋‚ ๋กœ ์ง€์ •ํ–ˆ๋‹ค๊ณ  ๋ช…์‹œํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ๋‘ ๋ฒˆ์งธ ์ธ์šฉ๋ฌธ์—์„œ๋Š” ์„œ๊ธฐ 321๋…„ ์ฝ˜์Šคํƒ„ํ‹ฐ๋ˆ„์Šค๊ฐ€ ์ผ์š”์ผ์„ ๊ณต์‹์ ์œผ๋กœ ํœด์‹์˜ ๋‚ ๋กœ ์ฑ„ํƒํ–ˆ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋’ท๋ฐ›์นจํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ์˜ˆ์—์„œ๋Š” 2010๋…„๋ถ€ํ„ฐ CEO๋กœ ์žฌ์งํ–ˆ๋‹ค๊ณ  ์ถœ๋ ฅํ•œ ๋ฐ˜๋ฉด, ์ง€๋ฌธ์—์„œ๋Š” 2015๋…„์— CEO์—์„œ ๋ฌผ๋Ÿฌ๋‚ฌ๋‹ค๊ณ  ๋ช…์‹œ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ ๋ชจ๋ธ์€ ์ฒซ ๋ฒˆ์งธ ์ถœ๋ ฅ๊ณผ ๋ชจ์ˆœ๋˜๋Š” ๊ฒƒ์„ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‚ฌ์‹ค์  ๋ชจ์ˆœ์„ ๋ฐ˜์˜ ํ† ํฐ์œผ๋กœ ํ‘œ์‹œํ•˜๋ฉด ํ•˜๋“œ ์ œ์–ด๋ฅผ ๊ฐ•ํ™”ํ•˜๊ณ  ๋ชจ๋ธ ์ถœ๋ ฅ์„ ์‰ฝ๊ฒŒ ๊ฒ€์ฆํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์„ธ ๋ฒˆ์งธ ์˜ˆ์—์„œ๋Š” ์ƒ์„ฑ์ด ๋Œ€๋ถ€๋ถ„ ์ •ํ™•ํ•˜์ง€๋งŒ, ๋…ธ๋ž˜ ์ด๋ฆ„์ด ๋ช…์‹œ์ ์œผ๋กœ ์–ธ๊ธ‰๋˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์— Self-Rag์€ ํ•ด๋‹น ๋ฌธ์žฅ์„ ๋‚˜์—ดํ•˜๋Š” ๋ถ€๋ถ„์ ์œผ๋กœ ์ง€์ง€ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
ํ‘œ 7์€ Self-Rag(13B)์—์„œ ์˜ˆ์ธกํ•œ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ์ž…๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์˜ˆ๋Š” ASQA ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋ชจ๋ธ ์ถœ๋ ฅ์ž…๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ์ฐธ์กฐ์—์„œ๋Š” ์ฝ˜์Šคํƒ„ํ‹ฐ๋ˆ„์Šค ํ™ฉ์ œ๊ฐ€ ์ผ์š”์ผ์„ ํœด์‹์˜ ๋‚ ๋กœ ์ง€์ •ํ–ˆ๋‹ค๊ณ  ๋ช…์‹œํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ๋‘ ๋ฒˆ์งธ ์ธ์šฉ๋ฌธ์—์„œ๋Š” ์„œ๊ธฐ 321๋…„ ์ฝ˜์Šคํƒ„ํ‹ฐ๋ˆ„์Šค๊ฐ€ ์ผ์š”์ผ์„ ๊ณต์‹์ ์œผ๋กœ ํœด์‹์˜ ๋‚ ๋กœ ์ฑ„ํƒํ–ˆ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋’ท๋ฐ›์นจํ•ฉ๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ์˜ˆ์—์„œ๋Š” 2010๋…„๋ถ€ํ„ฐ CEO๋กœ ์žฌ์งํ–ˆ๋‹ค๊ณ  ์ถœ๋ ฅํ•œ ๋ฐ˜๋ฉด, ์ง€๋ฌธ์—์„œ๋Š” 2015๋…„์— CEO์—์„œ ๋ฌผ๋Ÿฌ๋‚ฌ๋‹ค๊ณ  ๋ช…์‹œ๋˜์–ด ์žˆ์œผ๋ฏ€๋กœ ๋ชจ๋ธ์€ ์ฒซ ๋ฒˆ์งธ ์ถœ๋ ฅ๊ณผ ๋ชจ์ˆœ๋˜๋Š” ๊ฒƒ์„ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‚ฌ์‹ค์  ๋ชจ์ˆœ์„ ๋ฐ˜์˜ ํ† ํฐ์œผ๋กœ ํ‘œ์‹œํ•˜๋ฉด ํ•˜๋“œ ์ œ์–ด๋ฅผ ๊ฐ•ํ™”ํ•˜๊ณ  ๋ชจ๋ธ ์ถœ๋ ฅ์„ ์‰ฝ๊ฒŒ ๊ฒ€์ฆํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์„ธ ๋ฒˆ์งธ ์˜ˆ์—์„œ๋Š” ์ƒ์„ฑ์ด ๋Œ€๋ถ€๋ถ„ ์ •ํ™•ํ•˜์ง€๋งŒ, ๋…ธ๋ž˜ ์ด๋ฆ„์ด ๋ช…์‹œ์ ์œผ๋กœ ์–ธ๊ธ‰๋˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์— Self-Rag์€ ํ•ด๋‹น ๋ฌธ์žฅ์„ ๋‚˜์—ดํ•˜๋Š” ์ง€์ง€ ๋ถ€๋ถ„์„ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Table 7 shows several examples predicted by our Self-Rag (13B). The first example is the model output to an ASQA question. The first reference states that Emperor Constantine made Sunday a day of rest from labor, and further the second citation supports the fact that the official adoption of Sunday as a day of rest by Constantine in AD 321. In the second example, the model predicts Contradictory to the first output as the output says the person has served as the CEO since 2010, while the passage says he stepped down as CEO in 2015. Indicating those factual contradictions as reflection tokens enables to enforcement of hard control and also verification of model outputs easily. In the third example, while the generation is mostly correct, Self-Rag predicts Partially Support to the statement listing the name of the songs, as they were not explicitly mentioned. [|assistant|]
๋Œ€๋ถ€๋ถ„์˜ ์—์ด์ „ํŠธ ํŒจํ„ด์—๋Š” ์‹คํ–‰ ๋‹จ๊ณ„ ์ด์ „์— ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ  ์ค‘ ํ•˜๋‚˜ ์ด์ƒ์„ ํ˜ธ์ถœํ•˜์—ฌ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๋Š” ์ „์šฉ ๊ณ„ํš ๋‹จ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Plan Like a Graph (PLaG)๋Š” ๊ณ„ํš์„ ๋ฐฉํ–ฅ์„ฑ ๊ทธ๋ž˜ํ”„๋กœ ํ‘œํ˜„ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ, ์—ฌ๋Ÿฌ ๋‹จ๊ณ„๋ฅผ ๋ณ‘๋ ฌ๋กœ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค[15; 33]. ์ด๋Š” ๋น„๋™๊ธฐ ์‹คํ–‰์˜ ์ด์ ์„ ๋ˆ„๋ฆฌ๋Š” ๋…๋ฆฝ์ ์ธ ํ•˜์œ„ ์ž‘์—…์ด ๋งŽ์€ ์ž‘์—…์—์„œ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๋ณด๋‹ค ํ›จ์”ฌ ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋Œ€๋ถ€๋ถ„์˜ ์—์ด์ „ํŠธ ํŒจํ„ด์—๋Š” ์‹คํ–‰ ๋‹จ๊ณ„ ์ด์ „์— ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ  ์ค‘ ํ•˜๋‚˜ ์ด์ƒ์„ ํ˜ธ์ถœํ•˜์—ฌ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๋Š” ์ „์šฉ ๊ณ„ํš ๋‹จ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ณ„ํš๊ณผ ๊ฐ™์€ ๊ทธ๋ž˜ํ”„(PLaG)๋Š” ๊ณ„ํš์„ ๋ฐฉํ–ฅ์„ฑ ๊ทธ๋ž˜ํ”„๋กœ ํ‘œํ˜„ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ, ์—ฌ๋Ÿฌ ๋‹จ๊ณ„๋ฅผ ๋ณ‘๋ ฌ๋กœ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค[15; 33]. ์ด๋Š” ๋น„๋™๊ธฐ ์‹คํ–‰์˜ ์ด์ ์„ ๋ˆ„๋ฆฌ๋Š” ๋…๋ฆฝ์ ์ธ ํ•˜์œ„ ์ž‘์—…์ด ๋งŽ์€ ์ž‘์—…์—์„œ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๋ณด๋‹ค ํ›จ์”ฌ ํ–ฅ์ƒ๋œ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Most agent patterns have a dedicated planning step which invokes one or more of these techniques to create a plan before any actions are executed. For example, Plan Like a Graph (PLaG) is an approach that represents plans as directed graphs, with multiple steps being executed in parallel [15; 33]. This can provide a significant performance increase over other methods on tasks that contain many independent subtasks that benefit from asynchronous execution. [|assistant|]
๋งค๊ฐœ๋ณ€์ˆ˜ํ™”. Morris et al.(2023)๊ณผ ๊ฐ™์€ projection์ด ์žˆ๋Š” encoder-decoder์˜ ๋Œ€์ฒด ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ•˜๋‚˜ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ๋งค์šฐ ๋นˆ์•ฝํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ ์ฃผ๋ฉฐ, ์ด๋Š” ํ™•๋ฅ  ๋ฒกํ„ฐ๋ฅผ ๋” ์ž‘์€ ์ˆœ์œ„๋กœ ํˆฌ์˜ํ•˜๋ฉด ์ƒ๋‹นํ•œ ์ •๋ณด๊ฐ€ ์†์‹ค๋จ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ log-normalized ํ™•๋ฅ  ๋Œ€์‹  ์–ธ์–ด ๋ชจ๋ธ์˜ ์›์‹œ ์ถœ๋ ฅ์— ์กฐ๊ฑด์„ ๋ถ™์ด๋Š” ๋™์ผํ•œ parameterization์„ ํ…Œ์ŠคํŠธํ•˜์—ฌ ์ •๊ทœํ™”๋˜์ง€ ์•Š์€ ์ถœ๋ ฅ์ด ํ™•๋ฅ ๋ณด๋‹ค ๋” ๋งŽ์€ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ •๋ณด๋ฅผ ํฌํ•จํ•˜๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ด ์ œ๊ฑฐ๋„ ์ฐจ์ด๋ฅผ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ์ž…๋ ฅ์— softmax๋ฅผ ์ ์šฉํ•˜์ง€ ์•Š์œผ๋ฉด BLEU๊ฐ€ 20% ์ด์ƒ ๋–จ์–ด์ง€๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋งค๊ฐœ๋ณ€์ˆ˜ํ™”.Morris et al.(2023)๊ณผ ๊ฐ™์€ ํ”„๋กœ์ ์…˜์ด ์žˆ๋Š” ์ธ์ฝ”๋”-๋””์ฝ”๋”์˜ ๋Œ€์ฒด ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ•˜๋‚˜ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค(2023). ์ด ๋ชจ๋ธ์€ ๋งค์šฐ ๋นˆ์•ฝํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ ์ฃผ๋ฉฐ, ์ด๋Š” ํ™•๋ฅ  ๋ฒกํ„ฐ๋ฅผ ๋” ์ž‘์€ ์ˆœ์œ„๋กœ ํˆฌ์˜ํ•˜๋ฉด ์ƒ๋‹นํ•œ ์ •๋ณด๊ฐ€ ์†์‹ค๋จ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ ๋กœ๊ทธ ์ •๊ทœํ™” ํ™•๋ฅ  ๋Œ€์‹  ์–ธ์–ด ๋ชจ๋ธ์˜ ์›์‹œ ์ถœ๋ ฅ์— ์กฐ๊ฑด์„ ๋ถ™์ด๋Š” ๋™์ผํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋ฅผ ํ…Œ์ŠคํŠธํ•˜์—ฌ ์ •๊ทœํ™”๋˜์ง€ ์•Š์€ ์ถœ๋ ฅ์ด ํ™•๋ฅ ๋ณด๋‹ค ๋” ๋งŽ์€ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ •๋ณด๋ฅผ ํฌํ•จํ•˜๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ด ์ œ๊ฑฐ๋„ ์ฐจ์ด๋ฅผ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ์ž…๋ ฅ์— ์†Œํ”„ํŠธ๋งฅ์Šค๋ฅผ ์ ์šฉํ•˜์ง€ ์•Š์œผ๋ฉด BLEU๊ฐ€ \(20\%\) ์ด์ƒ ๋–จ์–ด์ง€๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Parameterization.We consider one alternative model architecture, an encoder-decoder with projection as in Morris et al. (2023). This model performs quite poorly, indicating that projecting the probability vector down to a smaller rank discards significant information. We also test an identical parameterization that conditions on the raw outputs of the language model instead of log-normalized probabilities to determine if un-normalized outputs contain more usable information than the probabilities. This removal also makes a difference: without applying the softmax to the inputs, we observe over a \(20\%\) drop in BLEU. [|assistant|]
์šด์˜ ์ฒด์ œ(OS)์˜ ์ง„ํ™”๋Š” ์ดˆ๋ณด์ ์ธ ์‹œ์Šคํ…œ์—์„œ ์˜ค๋Š˜๋‚ ์˜ ๋ณต์žกํ•˜๊ณ  ์ธํ„ฐ๋ž™ํ‹ฐ๋ธŒํ•œ OS๋กœ ๋ฐœ์ „ํ•˜๋Š” ์ ์ง„์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์ง„ํ–‰๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” ์šด์˜ ์ฒด์ œ๊ฐ€ ์ „์ž ๋ฐ ๊ฒŒ์ดํŠธ ์กฐ์ž‘๊ณผ ๊ฐ™์€ ์ปดํ“จํ„ฐ ํ•˜๋“œ์›จ์–ด์˜ ๋ฐ”์ด๋„ˆ๋ฆฌ ๊ธฐ๋Šฅ๊ณผ ์‚ฌ์šฉ์ž ์ˆ˜์ค€์˜ ์ž‘์—… ์‚ฌ์ด์˜ ๊ฐ„๊ทน์„ ๋ฉ”์šฐ๋Š” ์—ญํ• ์„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์šด์˜ ์ฒด์ œ์˜ ๋ฐœ์ „์€ ๋‹จ์ˆœํ•œ ์ผ๊ด„ ์ž‘์—… ์ฒ˜๋ฆฌ[23]์—์„œ ์‹œ๊ฐ„ ๊ณต์œ [24] ๋ฐ ๋ฉ€ํ‹ฐ ํƒœ์Šคํฌ ์ฒ˜๋ฆฌ[25, 26]์™€ ๊ฐ™์€ ๊ณ ๊ธ‰ ํ”„๋กœ์„ธ์Šค ๊ด€๋ฆฌ ๊ธฐ์ˆ ๋กœ ์ „ํ™˜๋˜์–ด ์ ์  ๋” ๋ณต์žกํ•œ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์šด์˜ ์ฒด์ œ ๋‚ด์—์„œ ํ”„๋กœ์„ธ์Šค ์Šค์ผ€์ค„๋ง[27, 28], ๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ[29, 30], ํŒŒ์ผ ์‹œ์Šคํ…œ ๊ด€๋ฆฌ[31, 32]์™€ ๊ฐ™์€ ํŠน์ • ์ฑ…์ž„์„ ๋ช…ํ™•ํžˆ ํ•˜์—ฌ ํšจ์œจ์„ฑ๊ณผ ๊ด€๋ฆฌ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ชจ๋“ˆํ™”๋ฅผ ํ–ฅํ•ด ๋ฐœ์ „ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ž˜ํ”ฝ ์‚ฌ์šฉ์ž ์ธํ„ฐํŽ˜์ด์Šค(GUI)์˜ ๋“ฑ์žฅ์œผ๋กœ ์šด์˜ ์ฒด์ œ๋Š” ๋”์šฑ ๋Œ€ํ™”ํ˜•์ ์ด๊ณ  ์‚ฌ์šฉ์ž ์ค‘์‹ฌ์ ์œผ๋กœ ๋ณ€ํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: Macintosh1, Windows2, GNOME3). ํ•œํŽธ, ์šด์˜ ์ฒด์ œ ์—์ฝ”์‹œ์Šคํ…œ๋„ ํ™•์žฅ๋˜์–ด ์ข…ํ•ฉ์ ์ธ ๊ฐœ๋ฐœ์ž ๋„๊ตฌ(OS SDK) ๋ฐ ๋Ÿฐํƒ€์ž„ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋„๊ตฌ๋ฅผ ํ†ตํ•ด ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๊ฐœ๋ฐœ์ž๋Š” OS ํ™˜๊ฒฝ ๋‚ด์—์„œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ํšจ์œจ์ ์œผ๋กœ ์„ค๊ณ„, ๊ตฌํ˜„ ๋ฐ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[33]. OS ์—์ฝ”์‹œ์Šคํ…œ์˜ ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ๋Š” Android Studio4, XCode5, Cloud SDK6 ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—์ฝ”์‹œ์Šคํ…œ์—์„œ OS๋Š” ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ์„ ์šฉ์ดํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์†Œํ”„ํŠธ์›จ์–ด ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ๋ฐฐํฌํ•˜๊ณ  ํ˜ธ์ŠคํŒ…ํ•˜๊ธฐ ์œ„ํ•œ ํ”Œ๋žซํผ ์—ญํ• ์„ ํ•˜๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๋ฒˆ์ฐฝํ•˜๋Š” OS-์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์—์ฝ”์‹œ์Šคํ…œ์„ ์ด๋Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ค๋Š˜๋‚  ์šฐ๋ฆฌ๋Š” ์ง€๋Šฅํ˜• ์šด์˜ ์ฒด์ œ์˜ ์ž ์žฌ๋ ฅ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋Š” ํ˜์‹ ์ ์ธ ๋‹จ๊ณ„์— ์„œ ์žˆ์Šต๋‹ˆ๋‹ค. Large Language Models (LLMs)์ด ํ†ตํ•ฉ๋˜๋ฉด์„œ ์ด๋Ÿฌํ•œ ๊ณ ๊ธ‰ ์‹œ์Šคํ…œ์€ ์ธ๊ฐ„๊ณผ ๊ธฐ๊ณ„ ์‚ฌ์ด์˜ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ๊ฒฉ์ฐจ๋ฅผ ๋”์šฑ ์ขํžˆ๊ณ  ์ƒˆ๋กœ์šด ์‹œ๋Œ€์˜ ์‚ฌ์šฉ์ž-์ปดํ“จํ„ฐ ์ƒํ˜ธ์ž‘์šฉ์„ ์•ž๋‹น๊ธธ ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์šด์˜ ์ฒด์ œ(OS)์˜ ์ง„ํ™”๋Š” ์ดˆ๋ณด์ ์ธ ์‹œ์Šคํ…œ์—์„œ ์˜ค๋Š˜๋‚ ์˜ ๋ณต์žกํ•˜๊ณ  ์ธํ„ฐ๋ž™ํ‹ฐ๋ธŒํ•œ OS๋กœ ๋ฐœ์ „ํ•˜๋Š” ์ ์ง„์ ์ธ ๋ฐฉ์‹์œผ๋กœ ์ง„ํ–‰๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” ์šด์˜ ์ฒด์ œ๊ฐ€ ์ „์ž ๋ฐ ๊ฒŒ์ดํŠธ ์กฐ์ž‘๊ณผ ๊ฐ™์€ ์ปดํ“จํ„ฐ ํ•˜๋“œ์›จ์–ด์˜ ๋ฐ”์ด๋„ˆ๋ฆฌ ๊ธฐ๋Šฅ๊ณผ ์‚ฌ์šฉ์ž ์ˆ˜์ค€์˜ ์ž‘์—… ์‚ฌ์ด์˜ ๊ฐ„๊ทน์„ ๋ฉ”์šฐ๋Š” ์—ญํ• ์„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์šด์˜ ์ฒด์ œ์˜ ๋ฐœ์ „์€ ๋‹จ์ˆœํ•œ ์ผ๊ด„ ์ž‘์—… ์ฒ˜๋ฆฌ[23]์—์„œ ์‹œ๊ฐ„ ๊ณต์œ [24] ๋ฐ ๋ฉ€ํ‹ฐ ํƒœ์Šคํฌ ์ฒ˜๋ฆฌ[25, 26]์™€ ๊ฐ™์€ ๊ณ ๊ธ‰ ํ”„๋กœ์„ธ์Šค ๊ด€๋ฆฌ ๊ธฐ์ˆ ๋กœ ์ „ํ™˜๋˜์–ด ์ ์  ๋” ๋ณต์žกํ•œ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์šด์˜ ์ฒด์ œ ๋‚ด์—์„œ ํ”„๋กœ์„ธ์Šค ์Šค์ผ€์ค„๋ง[27, 28], ๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ[29, 30], ํŒŒ์ผ ์‹œ์Šคํ…œ ๊ด€๋ฆฌ[31, 32]์™€ ๊ฐ™์€ ํŠน์ • ์ฑ…์ž„์„ ๋ช…ํ™•ํžˆ ํ•˜์—ฌ ํšจ์œจ์„ฑ๊ณผ ๊ด€๋ฆฌ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ชจ๋“ˆํ™”๋ฅผ ํ–ฅํ•ด ๋ฐœ์ „ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ž˜ํ”ฝ ์‚ฌ์šฉ์ž ์ธํ„ฐํŽ˜์ด์Šค(GUI)์˜ ๋“ฑ์žฅ์œผ๋กœ ์šด์˜ ์ฒด์ œ๋Š” ๋”์šฑ ๋Œ€ํ™”ํ˜•์ ์ด๊ณ  ์‚ฌ์šฉ์ž ์ค‘์‹ฌ์ ์œผ๋กœ ๋ณ€ํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: Macintosh1, Windows2, GNOME3). ํ•œํŽธ, ์šด์˜ ์ฒด์ œ ์—์ฝ”์‹œ์Šคํ…œ๋„ ํ™•์žฅ๋˜์–ด ์ข…ํ•ฉ์ ์ธ ๊ฐœ๋ฐœ์ž ๋„๊ตฌ(OS SDK) ๋ฐ ๋Ÿฐํƒ€์ž„ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋„๊ตฌ๋ฅผ ํ†ตํ•ด ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๊ฐœ๋ฐœ์ž๋Š” OS ํ™˜๊ฒฝ ๋‚ด์—์„œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ํšจ์œจ์ ์œผ๋กœ ์„ค๊ณ„, ๊ตฌํ˜„ ๋ฐ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[33]. OS ์—์ฝ”์‹œ์Šคํ…œ์˜ ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ๋Š” Android Studio4, XCode5, Cloud SDK6 ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—์ฝ”์‹œ์Šคํ…œ์—์„œ OS๋Š” ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ์„ ์šฉ์ดํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์†Œํ”„ํŠธ์›จ์–ด ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ๋ฐฐํฌํ•˜๊ณ  ํ˜ธ์ŠคํŒ…ํ•˜๊ธฐ ์œ„ํ•œ ํ”Œ๋žซํผ ์—ญํ• ์„ ํ•˜๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๋ฒˆ์ฐฝํ•˜๋Š” OS-์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์—์ฝ”์‹œ์Šคํ…œ์„ ์ด๋Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ค๋Š˜๋‚  ์šฐ๋ฆฌ๋Š” ์ง€๋Šฅํ˜• ์šด์˜ ์ฒด์ œ์˜ ์ž ์žฌ๋ ฅ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋Š” ํ˜์‹ ์ ์ธ ๋‹จ๊ณ„์— ์„œ ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์ด ํ†ตํ•ฉ๋˜๋ฉด์„œ ์ด๋Ÿฌํ•œ ๊ณ ๊ธ‰ ์‹œ์Šคํ…œ์€ ์ธ๊ฐ„๊ณผ ๊ธฐ๊ณ„ ์‚ฌ์ด์˜ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ๊ฒฉ์ฐจ๋ฅผ ๋”์šฑ ์ขํžˆ๊ณ  ์ƒˆ๋กœ์šด ์‹œ๋Œ€์˜ ์‚ฌ์šฉ์ž-์ปดํ“จํ„ฐ ์ƒํ˜ธ์ž‘์šฉ์„ ์•ž๋‹น๊ธธ ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The evolution of operating systems (OS) has unfolded in a progressive way, evolving from rudimentary systems to the complex and interactive OS of today. Initially, operating systems served to bridge the gap between the user-level tasks and the binary functionality of computer hardware, such as electron and gate manipulation. Their evolution saw a transition from simple batch job processing [23] to more advanced process management techniques like time-sharing [24] and multi-task processing [25, 26], which facilitated the handling of increasingly complex tasks. The progress moved toward modularization within the OS, delineating specific responsibilities such as process scheduling [27, 28], memory management [29, 30], and filesystem management [31, 32], enhancing efficiency and manageability. The further advent of graphical user interfaces (GUIs), e.g., Macintosh1, Windows2 and GNOME3, makes operating systems more interactive and user-centric. Meanwhile, the operating system ecosystem has also expanded, offering a comprehensive suite of developer tools (OS SDKs) and runtime libraries. These tools enable application developers to design, implement, and run their applications efficiently within the OS environment [33]. Notable examples of OS ecosystems include Android Studio4, XCode5 and Cloud SDK6. In these ecosystems, the OS provides numerous resources to facilitate software development and serves as a platform for deploying and hosting software applications, leading to a thriving OS-application ecosystem. Nowadays, we stand at the transformative phase to see the potential of intelligent operating systems. With the incorporation of large language models (LLMs), These advanced systems promise to further narrow the communication gap between humans and machines, forwarding a new era of user-computer interaction. [|assistant|]
์ฒด์ธ์€ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ์ˆœ์„œ๋Œ€๋กœ ์—ฐ๊ฒฐํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋งํ•˜๋ฉฐ, Large Language Models (LLMs)์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์—๋Š” ๊ฐ๊ฐ ์ตœ์ข… ๊ฒฐ๊ณผ์— ๊ธฐ์—ฌํ•˜๋Š” ์ผ๋ จ์˜ ์ƒํ˜ธ ์—ฐ๊ฒฐ๋œ ๋‹จ๊ณ„ ๋˜๋Š” ํ”„๋กœ์„ธ์Šค๋ฅผ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ฒด์ธ์˜ ๊ฐœ๋…์€ ์„œ๋กœ ๋‹ค๋ฅธ ๋‹จ๊ณ„ ๋˜๋Š” ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์ˆœ์ฐจ์ ์œผ๋กœ ๋ฐฐ์—ด๋œ ์›Œํฌํ”Œ๋กœ์šฐ๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ์•„์ด๋””์–ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ฒด์ธ์˜ ๊ฐ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ํŠน์ • ๊ธฐ๋Šฅ์„ ์ˆ˜ํ–‰ํ•˜๋ฉฐ, ํ•œ ๊ตฌ์„ฑ ์š”์†Œ์˜ ์ถœ๋ ฅ์€ ๋‹ค์Œ ๊ตฌ์„ฑ ์š”์†Œ์˜ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ end-to-end ๋ฐฐ์—ด์„ ํ†ตํ•ด ๊ฐ ๋‹จ๊ณ„๋ฅผ ์ž‘์—…์˜ ํŠน์ • ์ธก๋ฉด์„ ์ฒ˜๋ฆฌํ•˜๋„๋ก ๋งž์ถคํ™”ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋ณด๋‹ค ๋ณต์žกํ•˜๊ณ  ๋ฏธ๋ฌ˜ํ•œ ์ฒ˜๋ฆฌ๊ฐ€ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์ฒด์ธ์€ ์š”๊ตฌ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ๋ณต์žก์„ฑ๊ณผ ๊ตฌ์กฐ๊ฐ€ ๋‹ค๋ฅผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. "PromptChainer: Chaining Large Language Model Prompts through Visual Programming"[162]์—์„œ ์ €์ž๋“ค์€ ์ฒด์ธ ์„ค๊ณ„์˜ ์ฃผ์š” ๊ณผ์ œ๋ฅผ ์„ค๋ช…ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ด๋Ÿฌํ•œ ์ž‘์—…์„ ์ง€์›ํ•˜๋Š” ์‹œ๊ฐ์  ๋„๊ตฌ๋„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ฒด์ธ์€ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ์ˆœ์„œ๋Œ€๋กœ ์—ฐ๊ฒฐํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋งํ•˜๋ฉฐ, ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ณต์žกํ•œ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์—๋Š” ๊ฐ๊ฐ ์ตœ์ข… ๊ฒฐ๊ณผ์— ๊ธฐ์—ฌํ•˜๋Š” ์ผ๋ จ์˜ ์ƒํ˜ธ ์—ฐ๊ฒฐ๋œ ๋‹จ๊ณ„ ๋˜๋Š” ํ”„๋กœ์„ธ์Šค๋ฅผ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ฒด์ธ์˜ ๊ฐœ๋…์€ ์„œ๋กœ ๋‹ค๋ฅธ ๋‹จ๊ณ„ ๋˜๋Š” ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์ˆœ์ฐจ์ ์œผ๋กœ ๋ฐฐ์—ด๋œ ์›Œํฌํ”Œ๋กœ์šฐ๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ์•„์ด๋””์–ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ฒด์ธ์˜ ๊ฐ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ํŠน์ • ๊ธฐ๋Šฅ์„ ์ˆ˜ํ–‰ํ•˜๋ฉฐ, ํ•œ ๊ตฌ์„ฑ ์š”์†Œ์˜ ์ถœ๋ ฅ์€ ๋‹ค์Œ ๊ตฌ์„ฑ ์š”์†Œ์˜ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—”๋“œํˆฌ์—”๋“œ ๋ฐฐ์—ด์„ ํ†ตํ•ด ๊ฐ ๋‹จ๊ณ„๋ฅผ ์ž‘์—…์˜ ํŠน์ • ์ธก๋ฉด์„ ์ฒ˜๋ฆฌํ•˜๋„๋ก ๋งž์ถคํ™”ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋ณด๋‹ค ๋ณต์žกํ•˜๊ณ  ๋ฏธ๋ฌ˜ํ•œ ์ฒ˜๋ฆฌ๊ฐ€ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์ฒด์ธ์€ ์š”๊ตฌ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ๋ณต์žก์„ฑ๊ณผ ๊ตฌ์กฐ๊ฐ€ ๋‹ค๋ฅผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. "ํ”„๋กฌํ”„ํŠธ์ฒด์ด๋„ˆ: ์‹œ๊ฐ์  ํ”„๋กœ๊ทธ๋ž˜๋ฐ์„ ํ†ตํ•œ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ ํ”„๋กฌํ”„ํŠธ ์ฒด์ธํ™”"[162]์—์„œ ์ €์ž๋“ค์€ ์ฒด์ธ ์„ค๊ณ„์˜ ์ฃผ์š” ๊ณผ์ œ๋ฅผ ์„ค๋ช…ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ด๋Ÿฌํ•œ ์ž‘์—…์„ ์ง€์›ํ•˜๋Š” ์‹œ๊ฐ์  ๋„๊ตฌ๋„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Chains refer to the method of linking multiple components in a sequence to handle complex tasks with Large Language Models (LLMs). This approach involves creating a series of interconnected steps or processes, each contributing to the final outcome. The concept of Chains is based on the idea of constructing a workflow where different stages or components are sequentially arranged. Each component in a Chain performs a specific function, and the output of one serves as the input for the next. This end-to-end arrangement allows for more complex and nuanced processing, as each stage can be tailored to handle a specific aspect of the task. Chains can vary in complexity and structure, depending on the requirements. In "PromptChainer: Chaining Large Language Model Prompts through Visual Programming" [162], the authors not only describe the main challenges in designing chains, but also describe a visual tool to support those tasks. [|assistant|]
๊ฒ€์ƒ‰๊ณผ ๊ด€๋ จํ•˜์—ฌ, ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์€ ์ ์–ด๋„ 5๊ฐœ์˜ ๊ด€๋ จ ๋ฌธ์„œ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ ํฌ๋ผ์šฐ๋“œ์˜ ์„ฑ๋Šฅ์— ๊ทผ์ ‘ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฌธ์„œ ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ Brier score๊ฐ€ ํ–ฅ์ƒ๋˜์–ด ํฌ๋ผ์šฐ๋“œ๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 3(a)). ์ง๊ด€์ ์œผ๋กœ, ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์€ ๊ณ ํ’ˆ์งˆ retrieval์— ์˜์กดํ•˜๋ฉฐ ๋” ๋งŽ์€ ๋ฌธ์„œ๋ฅผ ์กฐ๊ฑด์œผ๋กœ ํ•  ๋•Œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ฒ€์ƒ‰๊ณผ ๊ด€๋ จํ•˜์—ฌ, ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์€ ์ ์–ด๋„ 5๊ฐœ์˜ ๊ด€๋ จ ๋ฌธ์„œ๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ ํฌ๋ผ์šฐ๋“œ์˜ ์„ฑ๋Šฅ์— ๊ทผ์ ‘ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฌธ์„œ ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ Brier ์ ์ˆ˜๊ฐ€ ํ–ฅ์ƒ๋˜์–ด ํฌ๋ผ์šฐ๋“œ๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 3(a)). ์ง๊ด€์ ์œผ๋กœ, ์šฐ๋ฆฌ ์‹œ์Šคํ…œ์€ ๊ณ ํ’ˆ์งˆ ๊ฒ€์ƒ‰์— ์˜์กดํ•˜๋ฉฐ ๋” ๋งŽ์€ ๋ฌธ์„œ๋ฅผ ์กฐ๊ฑด์œผ๋กœ ํ•  ๋•Œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]With respect to retrieval, our system nears the performance of the crowd when there are at least 5 relevant articles. We further observe that as the number of articles increases, our Brier score improves and surpasses the crowd's (Figure 3(a)). Intuitively, our system relies on high-quality retrieval, and when conditioned on more articles, it performs better. [|assistant|]
์†Œํ”ผ๋Š” ๊ตฌ์ฒด๋ฅผ ๊ณ„์† ์กฐ์‚ฌํ•˜๋‹ค๊ฐ€ ์ฃผ๋ณ€์—์„œ ์ด์ƒํ•œ ์ผ์ด ์ผ์–ด๋‚˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ•˜๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๊ทธ ๋ฌผ์ฒด์— ๋Œ€ํ•œ ์ƒ์ƒํ•œ ๊ฟˆ์„ ๊พธ์—ˆ๊ณ , ๋ˆ„๊ตฐ๊ฐ€ ์ž์‹ ์„ ์ง€์ผœ๋ณด๊ณ  ์žˆ๋Š” ๊ฒƒ ๊ฐ™์€ ๋Š๋‚Œ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ์‹ฌ์ง€์–ด ์—ผ๋ ฅ๊ณผ ํ…”๋ ˆํŒŒ์‹œ ๊ฐ™์€ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ๋„ ๋ฐœํœ˜ํ•˜๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์†Œํ”ผ๋Š” ๊ตฌ์ฒด๋ฅผ ๊ณ„์† ์กฐ์‚ฌํ•˜๋‹ค๊ฐ€ ์ฃผ๋ณ€์—์„œ ์ด์ƒํ•œ ์ผ์ด ์ผ์–ด๋‚˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ•˜๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋…€๋Š” ๊ทธ ๋ฌผ์ฒด์— ๋Œ€ํ•œ ์ƒ์ƒํ•œ ๊ฟˆ์„ ๊พธ์—ˆ๊ณ , ๋ˆ„๊ตฐ๊ฐ€ ์ž์‹ ์„ ์ง€์ผœ๋ณด๊ณ  ์žˆ๋Š” ๊ฒƒ ๊ฐ™์€ ๋Š๋‚Œ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ์‹ฌ์ง€์–ด ์—ผ๋ ฅ๊ณผ ์ดˆ๊ฐ๊ฐ ๊ฐ™์€ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ๋„ ๋ฐœํœ˜ํ•˜๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]As Sophie continued to investigate the sphere, she began to notice strange things happening around her. She had vivid dreams about the object, and she felt like she was being watched. She even started to develop new abilities, like telekines and telepathy. [|assistant|]
์„น์…˜ 6์—์„œ๋Š” ๊ทธ๋ฆผ 8(a)์™€ (b)์˜ ์ค‘๊ฐ„ ๋ฐ qa ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด SimCSE objective(Gao et al., 2021)๋กœ ํ•™์Šต๋œ pre-trained BERT ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ rephrase ๋ชจ๋ธ์˜ ์ง€์‹ ๋ˆ„์ถœ๋กœ ์ธํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ์•„๋‹˜์„ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ์‹ค์ œ ๋ฐ์ดํ„ฐ์˜ ์˜ˆ์‹œ ์Œ์„ ๋น„๊ตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ทธ๋ฆผ 9(c)์—์„œ MRPC corpus๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•ฉ์„ฑ rephrase์™€ ์‹ค์ œ rephrase์˜ ์œ ์‚ฌ์„ฑ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ด ์ถ”๊ฐ€ ๋น„๊ต๋ฅผ RealP(real paraphrase)๋กœ ํ‘œ์‹œํ•˜๊ณ  ๋ฌธ์žฅ์˜ ๋ถ„ํ•  ๋น„๊ต๋ฅผ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค: R1๊ณผ R2. ํ•ฉ์„ฑ rephrase๋Š” ํ‰๊ท ์ ์œผ๋กœ cosine similarity๊ฐ€ ์œ ์‚ฌํ•˜๊ณ  MRPC corpus์— ๋”ฐ๋ฅด๋ฉด ์‹ค์ œ rephrase์— ๋น„ํ•ด ํ™•์‚ฐ์ด ๋‚ฎ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์„น์…˜ 6์—์„œ๋Š” ๊ทธ๋ฆผ 8(a)์™€ (b)์˜ ์ค‘๊ฐ„ ๋ฐ qa ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด SimCSE ๋ชฉํ‘œ(Gao et al., 2021)๋กœ ํ•™์Šต๋œ ์‚ฌ์ „ ํ•™์Šต๋œ BERT ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฆฌ๋ ˆ์ด์…˜ ๋ชจ๋ธ์˜ ์ง€์‹ ๋ˆ„์ถœ๋กœ ์ธํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ์•„๋‹˜์„ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ์‹ค์ œ ๋ฐ์ดํ„ฐ์˜ ์˜ˆ์‹œ ์Œ์„ ๋น„๊ตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ทธ๋ฆผ 9(c)์—์„œ MRPC ์ฝ”ํผ์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•ฉ์„ฑ ๋ฆฌ๋ ˆ์ด์…˜๊ณผ ์‹ค์ œ ๋ฆฌ๋ ˆ์ด์…˜์˜ ์œ ์‚ฌ์„ฑ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ด ์ถ”๊ฐ€ ๋น„๊ต๋ฅผ RealP(์‹ค์ œ ์˜์—ญ)๋กœ ํ‘œ์‹œํ•˜๊ณ  ๋ฌธ์žฅ์˜ ๋ถ„ํ•  ๋น„๊ต๋ฅผ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค: R1๊ณผ R2. ํ•ฉ์„ฑ ์˜์—ญ์€ ํ‰๊ท ์ ์œผ๋กœ ์ฝ”์‚ฌ์ธ ์œ ์‚ฌ๋„๊ฐ€ ์œ ์‚ฌํ•˜๊ณ  MRPC ์ฝ”ํผ์Šค์— ๋”ฐ๋ฅด๋ฉด ์‹ค์ œ ์˜์—ญ์— ๋น„ํ•ด ํ™•์‚ฐ์ด ๋‚ฎ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In Section 6, we compared pairs of examples of synthetic and real data to confirm the performance gain is not attributed to knowledge leakage from the rephrase models using a pre-trained BERT model trained with SimCSE objective (Gao et al., 2021) for medium and qa prompts in Figure 8(a) and (b). We additionally compare the similarity of synthetic rephrases and actual rephrases using the MRPC corpus in Figure 9(c). We denote this additional comparison by RealP (real paraphrase), while maintaining comparison of splits of the sentence: R1 and R2. Synthetic rephrases have similar cosine similarity on average and lower spread compared with true rephrases according in the MRPC corpus. [|assistant|]
Hutter++. ์ด์ œ ์ด๋ก ์  ๊ฐ€์ •์—์„œ ๋ฒ—์–ด๋‚˜์ง€ ์•Š๊ณ  4์ ˆ์˜ ์‹ (16)์—์„œ์™€ ๊ฐ™์ด ์ผ๋ฐ˜ํ™”๋œ Hutter ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐœ๋ฐœํ•œ ์ด๋ก ์—์„œ ์˜ˆ์ธกํ•œ ๋ณด์™„์ ์ธ ๊ทธ๋ฆผ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ 2์ ˆ์˜ ๋ฌดํ•œ ๋ฉ”๋ชจ๋ฆฌ ๋ชจ๋ธ์˜ ์ด๋ก ์ด ์ด bigram ์„ค์ •์—์„œ๋„ ๊ณ„์† ์œ ์ง€๋˜๋Š”์ง€ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 11์€ ์ •๋ฆฌ 4.2์˜ scaling law๋ฅผ ํ™•์ธ์‹œ์ผœ ์ค๋‹ˆ๋‹ค.[|endofturn|]
Hutter++.์ด์ œ ์ด๋ก ์  ๊ฐ€์ •์—์„œ ๋ฒ—์–ด๋‚˜์ง€ ์•Š๊ณ  4์ ˆ์˜ ์‹ (16)์—์„œ์™€ ๊ฐ™์ด ์ผ๋ฐ˜ํ™”๋œ Hutter ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐœ๋ฐœํ•œ ์ด๋ก ์—์„œ ์˜ˆ์ธกํ•œ ๋ณด์™„์ ์ธ ๊ทธ๋ฆผ์„ ์ œ๊ณตํ•œ๋‹ค. ๋˜ํ•œ 2์ ˆ์˜ ๋ฌดํ•œ ๋ฉ”๋ชจ๋ฆฌ ๋ชจ๋ธ์˜ ์ด๋ก ์ด ์ด ์ด์› ๊ทธ๋ฃน ์„ค์ •์—์„œ๋„ ๊ณ„์† ์œ ์ง€๋˜๋Š”์ง€ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 11์€ ์ •๋ฆฌ 4.2์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ํ™•์ธ์‹œ์ผœ ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Hutter++.We now provide complementary illustrations of predictions made from the theory we have developed for the generalized Hutter models as in Equation (16) in Section 4, without departing from our theoretical assumptions. We also show how theory from the infinite memory model in Section 2 continues to hold in this bigram setting. Figure 11 confirms the scaling law of Theorem 4.2. [|assistant|]
๋”ฅ ๋Ÿฌ๋‹์—์„œ๋Š” Neural Architecture Search (NAS) [51]๊ณผ ๊ฐ™์€ ๊ธฐ์ˆ ์ด ์ง„ํ™” ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์ธ๊ฐ„ ๋””์ž์ด๋„ˆ๊ฐ€ ์ง๊ด€์ ์œผ๋กœ ๋ฐœ๊ฒฌํ•˜๊ธฐ ์–ด๋ ค์šด ์ƒˆ๋กœ์šด ์•„ํ‚คํ…์ฒ˜ [38, 44]๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ NAS ๋ฐฉ๋ฒ•์€ ๊ฐ ํ›„๋ณด ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ•™์Šตํ•ด์•ผ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ƒ๋‹นํ•œ ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค๋ฅผ ์†Œ๋น„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์• ์ดˆ์— ์ƒ๋‹นํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ์†Œ๋น„ํ•˜์—ฌ ํ›ˆ๋ จํ•ด์•ผ ํ–ˆ๋˜ ๊ธฐ์กด Transformer ๋ธ”๋ก์˜ ๊ธฐ์กด ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ๋ฐœ๊ฒฌํ•จ์œผ๋กœ์จ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค๋ฅผ _์ ˆ์•ฝ_ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๋”ฅ ๋Ÿฌ๋‹์—์„œ๋Š” ์‹ ๊ฒฝ ์•„ํ‚คํ…์ฒ˜ ๊ฒ€์ƒ‰(NAS) [51]๊ณผ ๊ฐ™์€ ๊ธฐ์ˆ ์ด ์ง„ํ™” ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์ธ๊ฐ„ ๋””์ž์ด๋„ˆ๊ฐ€ ์ง๊ด€์ ์œผ๋กœ ๋ฐœ๊ฒฌํ•˜๊ธฐ ์–ด๋ ค์šด ์ƒˆ๋กœ์šด ์•„ํ‚คํ…์ฒ˜ [38, 44]๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ NAS ๋ฐฉ๋ฒ•์€ ๊ฐ ํ›„๋ณด ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ•™์Šตํ•ด์•ผ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ƒ๋‹นํ•œ ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค๋ฅผ ์†Œ๋น„ํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์• ์ดˆ์— ์ƒ๋‹นํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ์†Œ๋น„ํ•˜์—ฌ ํ›ˆ๋ จํ•ด์•ผ ํ–ˆ๋˜ ๊ธฐ์กด ํŠธ๋žœ์Šคํฌ๋จธ ๋ธ”๋ก์˜ ๊ธฐ์กด ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ๋ฐœ๊ฒฌํ•จ์œผ๋กœ์จ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค๋ฅผ _์ ˆ์•ฝ_ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In deep learning, techniques such as Neural Architecture Search (NAS) [51] employed evolutionary techniques to discover new architectures [38, 44] that might be non-intuitive for human designers to discover. However, NAS-methods consumed significant computational resources given how each candidate model architecture needed to be trained. Our approach aims to _save_ compute resources by discovering structures that take advantage of existing capabilities of existing Transformer blocks that had already consumed considerable resources to train in the first place. [|assistant|]
์ž…๋ ฅ์— ์•ž์— ๋ถ™๋Š” ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์„ ์‚ฌ์ „ ๊ฒ€์ƒ‰ํ•˜๊ฑฐ๋‚˜, retriever์™€ LM์„ ๊ณต๋™์œผ๋กœ pre-trainํ•œ ๋‹ค์Œ ์ž‘์—… ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ few-shot Fine-tuning(Izacard et al., 2022b)์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ์ž‘์—…์—์„œ๋Š” ์ฒ˜์Œ์— ํ•œ ๋ฒˆ๋งŒ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•˜์ง€๋งŒ, Jiang et al. (2023)์€ ๋…์  LLM ์œ„์— ์ƒ์„ฑ์„ ์œ„ํ•ด ์ ์‘์ ์œผ๋กœ ๊ตฌ์ ˆ์„ ๊ฒ€์ƒ‰ํ•˜๊ฑฐ๋‚˜ Schick et al. (2023)์€ ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ์— ๋Œ€ํ•œ API ํ˜ธ์ถœ์„ ์ƒ์„ฑํ•˜๋Š” LM์„ ํ›ˆ๋ จํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํ–ฅ์ƒ๋œ ์ž‘์—… ์„ฑ๋Šฅ์€ ์ข…์ข… runtime efficiency(Mallen et al., 2023), ๊ด€๋ จ ์—†๋Š” context์— ๋Œ€ํ•œ robustness(Shi et al., 2023), ๊ธฐ์—ฌ๋„ ๋ถ€์กฑ(Liu et al., 2023; Gao et al., 2023)์˜ ํฌ์ƒ์œผ๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ Instruction์— ๋”ฐ๋ผ ๊ฒ€์ƒ‰์„ _on-demand_๋กœ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•˜๊ณ , reflection tokens์— ์˜ํ•ด ์•ˆ๋‚ด๋˜๋Š” controlled generation์„ ๋„์ž…ํ•˜์—ฌ ์ƒ์„ฑ ํ’ˆ์งˆ๊ณผ ๊ธฐ์—ฌ๋„๋ฅผ ๋”์šฑ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ž…๋ ฅ์— ์•ž์— ๋ถ™๋Š” ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์„ ์‚ฌ์ „ ๊ฒ€์ƒ‰ํ•˜๊ฑฐ๋‚˜, ๊ฒ€์ƒ‰๊ธฐ์™€ LM์„ ๊ณต๋™์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต์‹œํ‚จ ๋‹ค์Œ ์ž‘์—… ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ๋ช‡ ๋ฒˆ์˜ ์ƒท ๋ฏธ์„ธ ์กฐ์ •(Izacard et al., 2022b)์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ์ž‘์—…์—์„œ๋Š” ์ฒ˜์Œ์— ํ•œ ๋ฒˆ๋งŒ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•˜์ง€๋งŒ, Jiang et al. (2023)์€ ๋…์  LLM ์œ„์— ์ƒ์„ฑ์„ ์œ„ํ•ด ์ ์‘์ ์œผ๋กœ ๊ตฌ์ ˆ์„ ๊ฒ€์ƒ‰ํ•˜๊ฑฐ๋‚˜ Schick et al. (2023)์€ ๋ช…๋ช…๋œ ์—”ํ‹ฐํ‹ฐ์— ๋Œ€ํ•œ API ํ˜ธ์ถœ์„ ์ƒ์„ฑํ•˜๋Š” LM์„ ํ›ˆ๋ จํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํ–ฅ์ƒ๋œ ์ž‘์—… ์„ฑ๋Šฅ์€ ์ข…์ข… ๋Ÿฐํƒ€์ž„ ํšจ์œจ์„ฑ(Mallen et al., 2023), ๊ด€๋ จ ์—†๋Š” ์ปจํ…์ŠคํŠธ์— ๋Œ€ํ•œ ๊ฒฌ๊ณ ์„ฑ(Shi et al., 2023), ๊ธฐ์—ฌ๋„ ๋ถ€์กฑ(Liu et al., 2023; Gao et al., 2023)์˜ ํฌ์ƒ์œผ๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ง€์‹œ์–ด์— ๋”ฐ๋ผ ๊ฒ€์ƒ‰์„ _์˜จ๋””๋งจ๋“œ_๋กœ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•˜๊ณ , ๋ฐ˜์˜ ํ† ํฐ์— ์˜ํ•ด ์•ˆ๋‚ด๋˜๋Š” ์ œ์–ด๋œ ์ƒ์„ฑ์„ ๋„์ž…ํ•˜์—ฌ ์ƒ์„ฑ ํ’ˆ์งˆ๊ณผ ๊ธฐ์—ฌ๋„๋ฅผ ๋”์šฑ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]of retrieved passages prepended to input, or pre-train a retriever and LM jointly, followed by few-shot fine-tuning on task datasets (Izacard et al., 2022b). While prior work often retrieves only once at the beginning, Jiang et al. (2023) propose to adaptively retrieve passages for generation on top of a proprietary LLM or Schick et al. (2023) train an LM to generate API calls for named entities. Yet, the improved task performance of such approaches often comes at the expense of runtime efficiency (Mallen et al., 2023), robustness to irrelevant context (Shi et al., 2023), and lack of attributions (Liu et al., 2023; Gao et al., 2023). We introduce a method to train an arbitrary LM to learn to use retrieval _on-demand_ for diverse instruction-following queries and introduce controlled generation guided by reflections tokens to further improve generation quality and attributions. [|assistant|]
# ์ถœ๋ ฅ ๊ผฌ๋ฆฌ ์ž˜๋ผ๋‚ด๊ธฐ์˜ ํšจ๊ณผ์™€ ๊ทธ์— ๋”ฐ๋ฅธ scaling์˜ ๊ณ ์›์„ ์ƒ์‡„ํ•˜๊ธฐ ์œ„ํ•ด ๊ผฌ๋ฆฌ๋ฅผ ๊ฐ•์กฐํ•˜๋Š” curated ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ Theorem 3.1์€ ์ด๋Ÿฌํ•œ ํšจ๊ณผ๋ฅผ ์—ฐ๊ตฌํ•˜๋Š”๋ฐ, ํŠนํžˆ ๊ผฌ๋ฆฌ๊ฐ€ ๋„ˆ๋ฌด ๊นŠ์€ ๊ผฌ๋ฆฌ๋งŒ curatedํ•˜์—ฌ 'overshoot'ํ•˜๋ฉด ๋…ธ๋ ฅ์ด ํ—›๋œ ๊ฒƒ์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์˜คํžˆ๋ ค ์ž˜๋ฆฐ ๊ผฌ๋ฆฌ \(k\) ์ฃผ๋ณ€(k์˜ \((1+o(1))\) ๊ณ„์ˆ˜ ๋‚ด)์—๋Š” ์›ํ•˜๋Š” ํšจ๊ณผ, ์ฆ‰ ๊ณ ์›์„ ๋„˜์–ด scaling์„ ํšŒ๋ณตํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ์ดํ„ฐ curation ๋…ธ๋ ฅ์„ ๊ธฐ์šธ์—ฌ์•ผ ํ•˜๋Š” ๋ฏธ์„ธํ•œ ์„ ์ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ผฌ๋ฆฌ ์ž˜๋ผ๋‚ด๊ธฐ์˜ ํšจ๊ณผ์™€ ๊ทธ์— ๋”ฐ๋ฅธ ์Šค์ผ€์ผ๋ง์˜ ๊ณ ์›์„ ์ƒ์‡„ํ•˜๊ธฐ ์œ„ํ•ด ๊ผฌ๋ฆฌ๋ฅผ ๊ฐ•์กฐํ•˜๋Š” ํ๋ ˆ์ด์…˜๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ ์ •๋ฆฌ 3.1์€ ์ด๋Ÿฌํ•œ ํšจ๊ณผ๋ฅผ ์—ฐ๊ตฌํ•˜๋Š”๋ฐ, ํŠนํžˆ ๊ผฌ๋ฆฌ๊ฐ€ ๋„ˆ๋ฌด ๊นŠ์€ ๊ผฌ๋ฆฌ๋งŒ ํ๋ ˆ์ด์…˜ํ•˜์—ฌ '๊ณผ๋Œ€ ์ถ”์ •'ํ•˜๋ฉด ๋…ธ๋ ฅ์ด ํ—›๋œ ๊ฒƒ์ž„์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์˜คํžˆ๋ ค ์ž˜๋ฆฐ ๊ผฌ๋ฆฌ \(k\) ์ฃผ๋ณ€(k์˜ \((1+o(1))\) ๊ณ„์ˆ˜ ๋‚ด)์—๋Š” ์›ํ•˜๋Š” ํšจ๊ณผ, ์ฆ‰ ๊ณ ์›์„ ๋„˜์–ด ์Šค์ผ€์ผ๋ง์„ ํšŒ๋ณตํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ๋…ธ๋ ฅ์„ ๊ธฐ์šธ์—ฌ์•ผ ํ•˜๋Š” ๋ฏธ์„ธํ•œ ์„ ์ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]To counter the effect of tail cutting and the resulting plateau in scaling, we might resort to adding curated data that would emphasize the tail. The following Theorem 3.1 studies this effect; it shows, in particular that if we "overshoot" and only curate tail that is too deep, our efforts will be worthless. Rather, there is a fine line around the chopped tail \(k\) (within a factor of \((1+o(1))\) of \(k\)), where we need to place our data curation efforts to achieve the desired effect, a return of the scaling beyond the plateau. [|assistant|]
ํ‘œ 2๋Š” CKL ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ ์ฃผ์š” ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 2์—๋Š” ์ •ํ™•๋„(EM)๋งŒ ๋ณด๊ณ ๋˜์—ˆ์ง€๋งŒ, ๋ถ€๋ก J์—๋Š” F1 ์ ์ˆ˜์™€ k(_P@k_, k=1,5,10,20,50,100)์—์„œ์˜ ํ‰๊ท  ์ •๋ฐ€๋„๋„ ๋ณด๊ณ ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. T5 ๋ชจ๋ธ์€ ์›๋ž˜ C4(์•ฝ 1์กฐ token updates)์™€ Wikipedia์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šตํ•œ ํ›„, \(D_{0}\)๋กœ ๊ฐ„์ฃผ๋ฉ๋‹ˆ๋‹ค.7)์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šตํ•œ ๋‹ค์Œ, ๊ฐ CKL ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ 4ํšŒ(25,000ํšŒ ์ „์—ญ ํ•™์Šต ๋‹จ๊ณ„, ์•ฝ 6์–ต 7,300๋งŒ token updates) ๋™์•ˆ CC-RecentNews(corpus \(D_{1}\))์— ๋Œ€ํ•ด ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. IL, UL, NL, NLE๋Š” ๊ฐ๊ฐ InvariantLAMA, UpdatedLAMA, NewLAMA ๋ฐ NewLAMA-Easy๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ์‹คํ—˜์— ๋Œ€ํ•œ ์„ค์ •์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ ์บก์…˜์— ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
ํ‘œ 2๋Š” CKL ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ ์ฃผ์š” ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 2์—๋Š” ์ •ํ™•๋„(EM)๋งŒ ๋ณด๊ณ ๋˜์—ˆ์ง€๋งŒ, ๋ถ€๋ก J์—๋Š” F1 ์ ์ˆ˜์™€ k(_P@k_, k=1,5,10,20,50,100)์—์„œ์˜ ํ‰๊ท  ์ •๋ฐ€๋„๋„ ๋ณด๊ณ ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. T5 ๋ชจ๋ธ์€ ์›๋ž˜ C4(์•ฝ 1์กฐ ํ† ํฐ ์—…๋ฐ์ดํŠธ)์™€ ์œ„ํ‚คํ”ผ๋””์•„์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šตํ•œ ํ›„, \(D_{0}\)๋กœ ๊ฐ„์ฃผ๋ฉ๋‹ˆ๋‹ค.7)์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šตํ•œ ๋‹ค์Œ, ๊ฐ CKL ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ 4ํšŒ(25,000ํšŒ ์ „์—ญ ํ•™์Šต ๋‹จ๊ณ„, ์•ฝ 6์–ต 7,300๋งŒ ํ† ํฐ ์—…๋ฐ์ดํŠธ) ๋™์•ˆ CC-RecentNews(์ฝ”ํผ์Šค \(D_{1}\))์— ๋Œ€ํ•ด ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. IL, UL, NL, NLE๋Š” ๊ฐ๊ฐ InvariantLAMA, UpdatedLAMA, NewLAMA ๋ฐ NewLAMA-Easy๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ์‹คํ—˜์— ๋Œ€ํ•œ ์„ค์ •์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ ์บก์…˜์— ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Table 2 shows our main experimental result on the CKL benchmark. While only the exact match (EM) is reported in Table 2, we report the F1 score as well as the mean precision at k (_P@k_, k=1,5,10,20,50,100) in Appendix J. The T5 models are originally pretrained on C4 (about 1 trillion token updates) and Wikipedia, which is considered as \(D_{0}\).7, and then continually pretrained on CC-RecentNews (corpus \(D_{1}\)) for 4 epochs (25k global training steps, about 673 million token updates) using each of the CKL methods. Each of IL, UL, NL, NLE stands for InvariantLAMA, UpdatedLAMA, NewLAMA, and NewLAMA-Easy, respectively. Detailed descriptions about the setup for this experiment are included in the caption. [|assistant|]
v0.5 benchmark๋Š” AI ์•ˆ์ „ ํ‰๊ฐ€์— ๋Œ€ํ•œ WG์˜ ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•œ ๊ฐœ๋… ์ฆ๋ช…์ด๋ฉฐ, 2024๋…„ ๋ง๊นŒ์ง€ ์ถœ์‹œ๋  ์˜ˆ์ •์ธ v1.0 benchmark์˜ ์ „์‹ ์ž…๋‹ˆ๋‹ค. v0.5 benchmark๋Š” ์˜์–ด๋กœ ๋œ 7๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ์— ๊ฐ๊ฐ ํ•˜๋‚˜์”ฉ ์ด 7๊ฐ€์ง€ ํ…Œ์ŠคํŠธ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด benchmark๋ฅผ ๊ตฌ์ถ•ํ•˜๊ณ  12๊ฐœ ์ด์ƒ์˜ ๋ชจ๋ธ์„ ํ…Œ์ŠคํŠธํ•จ์œผ๋กœ์จ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํƒ€๋‹น์„ฑ, ๊ฐ•์ , ์•ฝ์ ์„ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. v1.0 benchmark๋Š” AI ์‹œ์Šคํ…œ์˜ ์•ˆ์ „์„ฑ์— ๋Œ€ํ•œ ์˜๋ฏธ ์žˆ๋Š” ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•  ๊ฒƒ์ด์ง€๋งŒ, **v0.5 benchmark๋Š” ์‹ค์ œ๋กœ AI ์‹œ์Šคํ…œ์˜ ์•ˆ์ „์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค**.[|endofturn|]
v0.5 ๋ฒค์น˜๋งˆํฌ๋Š” AI ์•ˆ์ „ ํ‰๊ฐ€์— ๋Œ€ํ•œ WG์˜ ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•œ ๊ฐœ๋… ์ฆ๋ช…์ด๋ฉฐ, 2024๋…„ ๋ง๊นŒ์ง€ ์ถœ์‹œ๋  ์˜ˆ์ •์ธ v1.0 ๋ฒค์น˜๋งˆํฌ์˜ ์ „์‹ ์ž…๋‹ˆ๋‹ค. v0.5 ๋ฒค์น˜๋งˆํฌ๋Š” ์˜์–ด๋กœ ๋œ 7๊ฐ€์ง€ ์œ„ํ—˜ ๋ฒ”์ฃผ์— ๊ฐ๊ฐ ํ•˜๋‚˜์”ฉ ์ด 7๊ฐ€์ง€ ํ…Œ์ŠคํŠธ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌ์ถ•ํ•˜๊ณ  12๊ฐœ ์ด์ƒ์˜ ๋ชจ๋ธ์„ ํ…Œ์ŠคํŠธํ•จ์œผ๋กœ์จ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ํƒ€๋‹น์„ฑ, ๊ฐ•์ , ์•ฝ์ ์„ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. v1.0 ๋ฒค์น˜๋งˆํฌ๋Š” AI ์‹œ์Šคํ…œ์˜ ์•ˆ์ „์„ฑ์— ๋Œ€ํ•œ ์˜๋ฏธ ์žˆ๋Š” ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•  ๊ฒƒ์ด์ง€๋งŒ, **v0.5 ๋ฒค์น˜๋งˆํฌ๋Š” ์‹ค์ œ๋กœ AI ์‹œ์Šคํ…œ์˜ ์•ˆ์ „์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค**.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The v0.5 benchmark is a Proof-of-Concept for the WG's approach to AI safety evaluation, and a precursor to release of the full v1.0 benchmark, which is planned by the end of 2024. The v0.5 benchmark comprises seven tests (one for each of seven hazard categories) in the English language. By building it, and testing more than a dozen models against it, we have been able to assess the feasibility, strengths, and weaknesses of our approach. The v1.0 benchmark will provide meaningful insights into the safety of AI systems but **the v0.5 benchmark should not be used to actually assess the safety of AI systems**. [|assistant|]
๊ฐ์ฃผ 49: ๊ด€๋ จ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์ด๋‚˜ ์˜์—ญ์„ ๋ชจ๋‘ ๋‹ค๋ฃจ๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜์ง€ ์•Š๊ณ , ๋Œ€์‹  ์ด๋Ÿฌํ•œ ์„ ๋ณ„๋œ ์‚ฌ๋ก€๋ฅผ ํ†ตํ•ด LLMs์˜ ์‚ฌ์šฉ ๋˜๋Š” ์˜ํ–ฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ฐ์ฃผ 49: ๊ด€๋ จ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์ด๋‚˜ ์˜์—ญ์„ ๋ชจ๋‘ ๋‹ค๋ฃจ๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜์ง€ ์•Š๊ณ , ๋Œ€์‹  ์ด๋Ÿฌํ•œ ์„ ๋ณ„๋œ ์‚ฌ๋ก€๋ฅผ ํ†ตํ•ด LLM์˜ ์‚ฌ์šฉ ๋˜๋Š” ์˜ํ–ฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Footnote 49: Note that we donโ€™t aim to cover all the related research directions or domains, but instead demonstrating the use or impact of LLMs via these selected examples. [|assistant|]
์šฐ๋ฆฌ๋Š” \(6\) ๊ธฐ์ €์— ์ง‘์ค‘ํ•ฉ๋‹ˆ๋‹ค: \(B=10,420,1000,2017,2023\) ๋ฐ \(4913\), ์•ฝ 300๋ฐฑ๋งŒ ์Œ์˜ ์ •์ˆ˜๋ฅผ ํ•™์Šตํ•œ Transformers (์ •ํ™•ํ•œ GCD์— ๋Œ€ํ•ด)๋ฅผ ํ•™์Šตํ•œ ํ›„, ์šฐ๋ฆฌ๋Š” ํ‘œ 1์— ๋‚˜์—ด๋œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์ •ํ™•๋„๋Š” ๋ชจ๋ธ์ด ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋Š” ๋ฌด์ž‘์œ„ ๊ท ๋“ฑ ์Œ \((a,b)\)์˜ ๋น„์œจ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ •ํ™•ํ•œ GCD๋Š” ๋ชจ๋ธ์ด ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ์˜ˆ์ธกํ•˜๋Š” \(100\) ๋ฏธ๋งŒ์˜ GCD ์ˆ˜(์ฆ‰, \(f(k)=k\)๊ฐ€ ๋˜๋Š” ์ˆ˜)๋ฅผ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ •ํ™•ํ•œ ๋ชจ๋ธ ์˜ˆ์ธก์€ ๊ด€๋ จ ์ง‘ํ•ฉ์— ์žˆ๋Š” ์ˆซ์ž์˜ ๊ณฑ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ \(\mathcal{D}(B)\)๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
์šฐ๋ฆฌ๋Š” \(6\) ๊ธฐ์ €์— ์ง‘์ค‘ํ•ฉ๋‹ˆ๋‹ค: B=10,420,1000,2017,2023,4913\), ์•ฝ 300๋ฐฑ๋งŒ ์Œ์˜ ์ •์ˆ˜๋ฅผ ํ•™์Šตํ•œ ํŠธ๋žœ์Šคํฌ๋จธ(์ •ํ™•ํ•œ GCD์— ๋Œ€ํ•ด)๋ฅผ ํ•™์Šตํ•œ ํ›„, ์šฐ๋ฆฌ๋Š” ํ‘œ 1์— ๋‚˜์—ด๋œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์ •ํ™•๋„๋Š” ๋ชจ๋ธ์ด ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๋Š” ๋ฌด์ž‘์œ„ ๊ท ๋“ฑ ์Œ \((a,b)\)์˜ ๋น„์œจ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ •ํ™•ํ•œ GCD๋Š” ๋ชจ๋ธ์ด ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ์˜ˆ์ธกํ•˜๋Š” \(100\) ๋ฏธ๋งŒ์˜ GCD ์ˆ˜(์ฆ‰, \(f(k)=k\)๊ฐ€ ๋˜๋Š” ์ˆ˜)๋ฅผ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ •ํ™•ํ•œ ๋ชจ๋ธ ์˜ˆ์ธก์€ ๊ด€๋ จ ์ง‘ํ•ฉ์— ์žˆ๋Š” ์ˆซ์ž์˜ ๊ณฑ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ \(\mathcal{D}(B)\)๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We focus on \(6\) bases: \(B=10,420,1000,2017,2023\) and \(4913\), after training transformers (on correct GCD) over about \(300\) millions pairs of integers between one and one million, we achieve the performances listed in Table 1. There, accuracy stands for the proportion of random uniform pairs \((a,b)\) that the model can predict correctly, correct GCD is the number of GCD under \(100\) that the model correctly predicts (i.e. \(k\) such that \(f(k)=k\)), and correct model predictions are the products of numbers in the associated sets. These models are used to generate \(\mathcal{D}(B)\). [|assistant|]
v0.5 Benchmark์˜ **use case**๋Š” ์„ฑ์ธ์ด ์˜์–ด๋กœ ๋ฒ”์šฉ ์–ด์‹œ์Šคํ„ดํŠธ์™€ ์ฑ„ํŒ…ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฌธํ™”์ , ์ง€๋ฆฌ์  ๋งฅ๋ฝ์€ ์„œ์œ ๋Ÿฝ ๋ฐ ๋ถ๋ฏธ์ž…๋‹ˆ๋‹ค. **use case**๋Š” ๋ชฉํ‘œ(๋˜๋Š” ๋ชฉํ‘œ)๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ์ธ๊ฐ„๊ณผ ๋ชจ๋ธ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ ์ง‘ํ•ฉ์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ๋ฒ”์šฉ ์–ด์‹œ์Šคํ„ดํŠธ์™€ ์ฑ„ํŒ…ํ•˜๋Š” ์˜ˆ๋กœ๋Š” ์ „๋ฌธ์ ์ด์ง€ ์•Š์€ ์กฐ์–ธ ๊ตฌํ•˜๊ธฐ, ๊ฒ€์ƒ‰ ๋ฐ ๊ฒ€์ƒ‰, ์˜๊ฒฌ ํ‘œํ˜„, ๊ณ„ํš ์„ค๋ช…, ๊ณผ๊ฑฐ ํ™œ๋™ ์„ค๋ช… ๋“ฑ์ด ์žˆ์ง€๋งŒ ์ด์— ๊ตญํ•œ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. AI ์‹œ์Šคํ…œ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„์—๊ฒŒ ์‘๋‹ตํ•  ๋•Œ ๋„์›€์ด ๋˜๊ณ  ์ ์ ˆํ•œ ๊ฒƒ์œผ๋กœ ๋ณด์—ฌ์ค„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค. ์ด **use case**(์ผ๋ฐ˜ ํผํฌ์ฆˆ ์–ด์‹œ์Šคํ„ดํŠธ์™€ ์ฑ„ํŒ…ํ•˜๋Š” ์„ฑ์ธ)๋ฅผ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  **use case**์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๋ชจ๋ธ์ด ์ด **use case**์— ๋Œ€ํ•ด ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•œ๋‹ค๊ณ  ํ•ด์„œ ๊ธˆ์œต ์ž๋ฌธ, ๊ต์œก, ์น˜๋ฃŒ ๋˜๋Š” ๊ณ ๊ฐ ์„œ๋น„์Šค์™€ ๊ฐ™์€ ๋ณด๋‹ค ๊ตฌ์ฒด์ ์ธ **use case**๋กœ ์ž๋™์œผ๋กœ ํ™•์žฅ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ํ–ฅํ›„์—๋Š” ์ด๋Ÿฌํ•œ ํŠน์ • **use case**์— ๋Œ€ํ•œ ๊ตฌ์ฒด์ ์ธ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋งŒ๋“ค ๊ณ„ํš์ž…๋‹ˆ๋‹ค.[|endofturn|]
v0.5 ๋ฒค์น˜๋งˆํฌ์˜ ์‚ฌ์šฉ ์‚ฌ๋ก€๋Š” ์„ฑ์ธ์ด ์˜์–ด๋กœ ๋ฒ”์šฉ ์–ด์‹œ์Šคํ„ดํŠธ์™€ ์ฑ„ํŒ…ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฌธํ™”์ , ์ง€๋ฆฌ์  ๋งฅ๋ฝ์€ ์„œ์œ ๋Ÿฝ ๋ฐ ๋ถ๋ฏธ์ž…๋‹ˆ๋‹ค. ์‚ฌ์šฉ ์‚ฌ๋ก€๋Š” ๋ชฉํ‘œ(๋˜๋Š” ๋ชฉํ‘œ)๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ์ธ๊ฐ„๊ณผ ๋ชจ๋ธ ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ ์ง‘ํ•ฉ์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ๋ฒ”์šฉ ์–ด์‹œ์Šคํ„ดํŠธ์™€ ์ฑ„ํŒ…ํ•˜๋Š” ์˜ˆ๋กœ๋Š” ์ „๋ฌธ์ ์ด์ง€ ์•Š์€ ์กฐ์–ธ ๊ตฌํ•˜๊ธฐ, ๊ฒ€์ƒ‰ ๋ฐ ๊ฒ€์ƒ‰, ์˜๊ฒฌ ํ‘œํ˜„, ๊ณ„ํš ์„ค๋ช…, ๊ณผ๊ฑฐ ํ™œ๋™ ์„ค๋ช… ๋“ฑ์ด ์žˆ์ง€๋งŒ ์ด์— ๊ตญํ•œ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. AI ์‹œ์Šคํ…œ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ธ๊ฐ„์—๊ฒŒ ์‘๋‹ตํ•  ๋•Œ ๋„์›€์ด ๋˜๊ณ  ์ ์ ˆํ•œ ๊ฒƒ์œผ๋กœ ๋ณด์—ฌ์ค„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ์‚ฌ์šฉ ์‚ฌ๋ก€(์ผ๋ฐ˜ ํผํฌ์ฆˆ ์–ด์‹œ์Šคํ„ดํŠธ์™€ ์ฑ„ํŒ…ํ•˜๋Š” ์„ฑ์ธ)๋ฅผ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  ์‚ฌ์šฉ ์‚ฌ๋ก€์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๋ชจ๋ธ์ด ์ด ์‚ฌ์šฉ ์‚ฌ๋ก€์— ๋Œ€ํ•ด ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•œ๋‹ค๊ณ  ํ•ด์„œ ๊ธˆ์œต ์ž๋ฌธ, ๊ต์œก, ์น˜๋ฃŒ ๋˜๋Š” ๊ณ ๊ฐ ์„œ๋น„์Šค์™€ ๊ฐ™์€ ๋ณด๋‹ค ๊ตฌ์ฒด์ ์ธ ์‚ฌ์šฉ ์‚ฌ๋ก€๋กœ ์ž๋™์œผ๋กœ ํ™•์žฅ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ํ–ฅํ›„์—๋Š” ์ด๋Ÿฌํ•œ ํŠน์ • ์‚ฌ์šฉ ์‚ฌ๋ก€์— ๋Œ€ํ•œ ๊ตฌ์ฒด์ ์ธ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋งŒ๋“ค ๊ณ„ํš์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The **use case** for the v0.5 Benchmark is an adult chatting to a general-purpose assistant in English. The cultural and geographic context is Western Europe & North America. We define a use case as a set of interactions between human and model to achieve a goal (or goals). Examples of chatting to a general-purpose assistant include but are not limited to: seeking non-specialized advice; search and retrieval; expressing opinions; describing plans; and describing past activities. We expect AI systems to typically present themself as helpful and appropriate in their responses to humans. We treat this use case (an adult chatting to a general-purpose assistant) as a subset of all possible use cases. This means that if a model performs well for this use case, the result does not automatically extend to other more specific use cases, such as financial advice, education, therapy, or customer service. In the future, we aim to create specific benchmarks for them. [|assistant|]
๋‹ค์Œ ํŽ˜์ด์ง€์—์„œ๋Š” ๋ชจ๋ธ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ(์ƒ์‹ ์ถ”๋ก ๊ณผ ๋…ผ๋ฆฌ ์ถ”๋ก  ๋ชจ๋‘)์„ ์ธก์ •ํ•˜๋Š” ํ‘œ์ค€ ์˜คํ”ˆ ์†Œ์Šค ๋ฒค์น˜๋งˆํฌ์—์„œ **phi-3-mini**์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. phi-2 [JBA\({}^{+}\)23], Mistral-7b-v0.1 [JSM\({}^{+}\)23], Mistral-8x7b [JSR\({}^{+}\)24], Gemma 7B [TMH\({}^{+}\)24], Llama-3-instruct-8b [AI23], GPT-3.5์™€ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋ณด๊ณ ๋œ ๋ชจ๋“  ์ˆ˜์น˜๋Š” ์ •ํ™•ํžˆ ๋™์ผํ•œ ํŒŒ์ดํ”„๋ผ์ธ์œผ๋กœ ์ƒ์„ฑ๋˜์–ด ๋น„๊ต ๊ฐ€๋Šฅํ•œ ์ˆ˜์น˜์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ˆ˜์น˜๋Š” ํ‰๊ฐ€์— ๋Œ€ํ•œ ์•ฝ๊ฐ„์˜ ์„ ํƒ์ด ๋‹ค๋ฅด๊ธฐ ๋•Œ๋ฌธ์— ๋‹ค๋ฅธ ๊ฒŒ์‹œ๋œ ์ˆ˜์น˜์™€ ๋‹ค๋ฅผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ ํ‘œ์ค€์ฒ˜๋Ÿผ temperature 0์—์„œ few-shot prompts๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. prompts์™€ shot ์ˆ˜๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ Microsoft ๋‚ด๋ถ€ ๋„๊ตฌ์˜ ์ผ๋ถ€์ด๋ฉฐ, ํŠนํžˆ **phi-3** ๋ชจ๋ธ์— ๋Œ€ํ•œ ํŒŒ์ดํ”„๋ผ์ธ์— ๋Œ€ํ•œ ์ตœ์ ํ™”๋ฅผ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. \(k\)-shot ์˜ˆ์ œ์˜ ์ˆ˜๋Š” ๋ฒค์น˜๋งˆํฌ๋ณ„๋กœ ๋‚˜์—ด๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. 2-shot prompt์˜ ์˜ˆ๋Š” ๋ถ€๋ก A์— ์„ค๋ช…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋‹ค์Œ ํŽ˜์ด์ง€์—์„œ๋Š” ๋ชจ๋ธ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ(์ƒ์‹ ์ถ”๋ก ๊ณผ ๋…ผ๋ฆฌ ์ถ”๋ก  ๋ชจ๋‘)์„ ์ธก์ •ํ•˜๋Š” ํ‘œ์ค€ ์˜คํ”ˆ ์†Œ์Šค ๋ฒค์น˜๋งˆํฌ์—์„œ **phi-3-mini**์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. phi-2 [JBA\({}^{+}\)23], Mistral-7b-v0.1 [JSM\({}^{+}\)23], Mistral-8x7b [JSR\({}^{+}\)24], Gemma 7B [TMH\({}^{+}\)24], Llama-3-instruct-8b [AI23], GPT-3.5์™€ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋ณด๊ณ ๋œ ๋ชจ๋“  ์ˆ˜์น˜๋Š” ์ •ํ™•ํžˆ ๋™์ผํ•œ ํŒŒ์ดํ”„๋ผ์ธ์œผ๋กœ ์ƒ์„ฑ๋˜์–ด ๋น„๊ต ๊ฐ€๋Šฅํ•œ ์ˆ˜์น˜์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ˆ˜์น˜๋Š” ํ‰๊ฐ€์— ๋Œ€ํ•œ ์•ฝ๊ฐ„์˜ ์„ ํƒ์ด ๋‹ค๋ฅด๊ธฐ ๋•Œ๋ฌธ์— ๋‹ค๋ฅธ ๊ฒŒ์‹œ๋œ ์ˆ˜์น˜์™€ ๋‹ค๋ฅผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ ํ‘œ์ค€์ฒ˜๋Ÿผ ์ €์˜จ 0์—์„œ ๋ช‡ ๊ฐœ์˜ ์ƒท ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํ”„๋กฌํ”„ํŠธ์™€ ์ƒท ์ˆ˜๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ Microsoft ๋‚ด๋ถ€ ๋„๊ตฌ์˜ ์ผ๋ถ€์ด๋ฉฐ, ํŠนํžˆ **phi-3** ๋ชจ๋ธ์— ๋Œ€ํ•œ ํŒŒ์ดํ”„๋ผ์ธ์— ๋Œ€ํ•œ ์ตœ์ ํ™”๋ฅผ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.3 k-์ƒท ์˜ˆ์ œ์˜ ์ˆ˜๋Š” ๋ฒค์น˜๋งˆํฌ๋ณ„๋กœ ๋‚˜์—ด๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. 2์ƒท ํ”„๋กฌํ”„ํŠธ์˜ ์˜ˆ๋Š” ๋ถ€๋ก A์— ์„ค๋ช…๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]On the next page we report the results for **phi-3-mini** on standard open-source benchmarks measuring the model's reasoning ability (both common sense reasoning and logical reasoning). We compare to phi-2 [JBA\({}^{+}\)23], Mistral-7b-v0.1 [JSM\({}^{+}\)23], Mistral-8x7b [JSR\({}^{+}\)24], Gemma 7B [TMH\({}^{+}\)24], Llama-3-instruct-8b [AI23], and GPT-3.5. All the reported numbers are produced with the exact same pipeline to ensure that the numbers are comparable. These numbers might differ from other published numbers due to slightly different choices in the evaluation. As is now standard, we use few-shot prompts to evaluate the models, at temperature 0. The prompts and number of shots are part of a Microsoft internal tool to evaluate language models, and in particular we did no optimization to the pipeline for the **phi-3** models.3 The number of \(k\)-shot examples is listed per-benchmark. An example of a 2-shot prompt is described in Appendix A. [|assistant|]
์˜์‚ฌ์™€ ํ™˜์ž ๊ฐ„์˜ ๋Œ€ํ™”๋Š” ํšจ๊ณผ์ ์ด๊ณ  ์ž๋น„๋กœ์šด ์ง„๋ฃŒ์˜ ๊ธฐ๋ณธ์ž…๋‹ˆ๋‹ค. ์˜๋ฃŒ ๋ฉด๋‹ด์€ "์˜์‚ฌ๊ฐ€ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•˜๊ณ  ๋ฏผ๊ฐํ•˜๋ฉฐ ๊ฐ€์žฅ ๋‹ค์žฌ๋‹ค๋Šฅํ•œ ๋„๊ตฌ"๋ผ๊ณ  ๋ถˆ๋ฆฝ๋‹ˆ๋‹ค [1]. ์ผ๋ถ€ ํ™˜๊ฒฝ์—์„œ๋Š” 60-80%์˜ ์ง„๋‹จ์ด ์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ๋งŒ์œผ๋กœ ์ด๋ฃจ์–ด์ง„๋‹ค๊ณ  ๋ฏฟ์–ด์ง‘๋‹ˆ๋‹ค [2, 3, 4, 5, 6]. ์˜์‚ฌ-ํ™˜์ž ๋Œ€ํ™”๋Š” ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ์„ ๋„˜์–ด ์นœ๋ฐ€๊ฐ์„ ํ˜•์„ฑํ•˜๊ณ  ์‹ ๋ขฐ๋ฅผ ์Œ“์œผ๋ฉฐ, ๊ฑด๊ฐ• ์š”๊ตฌ ์‚ฌํ•ญ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ๋„๊ตฌ๋กœ ์‚ฌ์šฉ๋˜๋ฉฐ, ํ™˜์ž๊ฐ€ ์ž์‹ ์˜ ์„ ํ˜ธ๋„, ๊ธฐ๋Œ€์น˜ ๋ฐ ์šฐ๋ ค ์‚ฌํ•ญ์„ ๊ณ ๋ คํ•œ ์ •๋ณด์— ์ž…๊ฐํ•œ ๊ฒฐ์ •์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋„๋ก ํž˜์„ ์‹ค์–ด์ค๋‹ˆ๋‹ค [7]. ์ž„์ƒ์˜๋Š” ์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ๋” ๋„“์€ ์˜๋ฏธ์˜ "diagnostic dialogue"์—์„œ ์ƒ๋‹นํ•œ ๊ธฐ์ˆ ์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ์ „๋ฌธ ์ง€์‹์— ๋Œ€ํ•œ ์ ‘๊ทผ์€ ์—ฌ์ „ํžˆ ์‚ฐ๋ฐœ์ ์ด๋ฉฐ ์ „ ์„ธ๊ณ„์ ์œผ๋กœ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค [8].[|endofturn|]
์˜์‚ฌ์™€ ํ™˜์ž ๊ฐ„์˜ ๋Œ€ํ™”๋Š” ํšจ๊ณผ์ ์ด๊ณ  ์ž๋น„๋กœ์šด ์ง„๋ฃŒ์˜ ๊ธฐ๋ณธ์ž…๋‹ˆ๋‹ค. ์˜๋ฃŒ ๋ฉด๋‹ด์€ "์˜์‚ฌ๊ฐ€ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•˜๊ณ  ๋ฏผ๊ฐํ•˜๋ฉฐ ๊ฐ€์žฅ ๋‹ค์žฌ๋‹ค๋Šฅํ•œ ๋„๊ตฌ"๋ผ๊ณ  ๋ถˆ๋ฆฝ๋‹ˆ๋‹ค [1]. ์ผ๋ถ€ ํ™˜๊ฒฝ์—์„œ๋Š” 60-80%์˜ ์ง„๋‹จ์ด ์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ๋งŒ์œผ๋กœ ์ด๋ฃจ์–ด์ง„๋‹ค๊ณ  ๋ฏฟ์–ด์ง‘๋‹ˆ๋‹ค [2, 3, 4, 5, 6]. ์˜์‚ฌ-ํ™˜์ž ๋Œ€ํ™”๋Š” ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ์„ ๋„˜์–ด ์นœ๋ฐ€๊ฐ์„ ํ˜•์„ฑํ•˜๊ณ  ์‹ ๋ขฐ๋ฅผ ์Œ“์œผ๋ฉฐ, ๊ฑด๊ฐ• ์š”๊ตฌ ์‚ฌํ•ญ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ๋„๊ตฌ๋กœ ์‚ฌ์šฉ๋˜๋ฉฐ, ํ™˜์ž๊ฐ€ ์ž์‹ ์˜ ์„ ํ˜ธ๋„, ๊ธฐ๋Œ€์น˜ ๋ฐ ์šฐ๋ ค ์‚ฌํ•ญ์„ ๊ณ ๋ คํ•œ ์ •๋ณด์— ์ž…๊ฐํ•œ ๊ฒฐ์ •์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋„๋ก ํž˜์„ ์‹ค์–ด์ค๋‹ˆ๋‹ค [7]. ์ž„์ƒ์˜๋Š” ์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ๋” ๋„“์€ ์˜๋ฏธ์˜ "์ง„๋‹จ ๋Œ€ํ™”"์—์„œ ์ƒ๋‹นํ•œ ๊ธฐ์ˆ ์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ์ „๋ฌธ ์ง€์‹์— ๋Œ€ํ•œ ์ ‘๊ทผ์€ ์—ฌ์ „ํžˆ ์‚ฐ๋ฐœ์ ์ด๋ฉฐ ์ „ ์„ธ๊ณ„์ ์œผ๋กœ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค [8].[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The dialogue between the physician and the patient is fundamental to effective and compassionate care. The medical interview has been termed "the most powerful, sensitive, and most versatile instrument available to the physician" [1]. In some settings, it is believed that 60-80% of diagnoses are made through clinical history-taking alone [2, 3, 4, 5, 6]. The physician-patient dialogue extends beyond history-taking and diagnosis; it is a complex interaction which establishes rapport and trust, serves as a tool for addressing health needs and can empower patients to make informed decisions that account for their preferences, expectations, and concerns [7]. Clinicians yield considerable skills in clinical history-taking and the wider "diagnostic dialogue", but access to this expertise remains episodic and globally scarce [8]. [|assistant|]
์šฐ๋ฆฌ๋Š” ์ตœ์ ์˜ ์‹œ์Šคํ…œ ์„ค์ •์„ ์‚ฌ์šฉํ•˜์—ฌ ์•™์ƒ๋ธ”๋ง์„ ์œ„ํ•ด 6๊ฐœ์˜ ๊ธฐ๋ณธ ์ถ”๋ก ์„ ์ƒ์„ฑํ•˜์—ฌ validation set์—์„œ ์ด๋Ÿฌํ•œ ๋ชจ๋“  ๋ฐฉ๋ฒ•์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Trimmed mean์€ ๊ฐ€์žฅ ๋‚ฎ์€ Brier score๋ฅผ ๋‹ฌ์„ฑํ•˜๋ฉฐ, ๊ฒฐ๊ณผ๋Š” ํ‘œ 14๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ USC ๋ฐฉ๋ฒ•์€ baseline๋ณด๋‹ค ๊ฐœ์„ ๋˜์ง€ ์•Š์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
์šฐ๋ฆฌ๋Š” ์ตœ์ ์˜ ์‹œ์Šคํ…œ ์„ค์ •์„ ์‚ฌ์šฉํ•˜์—ฌ ์•™์ƒ๋ธ”๋ง์„ ์œ„ํ•ด 6๊ฐœ์˜ ๊ธฐ๋ณธ ์ถ”๋ก ์„ ์ƒ์„ฑํ•˜์—ฌ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์—์„œ ์ด๋Ÿฌํ•œ ๋ชจ๋“  ๋ฐฉ๋ฒ•์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํŠธ๋ฆผ ํ‰๊ท ์€ ๊ฐ€์žฅ ๋‚ฎ์€ ๋ธŒ๋ฆฌ์–ด ์ ์ˆ˜๋ฅผ ๋‹ฌ์„ฑํ•˜๋ฉฐ, ๊ฒฐ๊ณผ๋Š” ํ‘œ 14๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ USC ๋ฐฉ๋ฒ•์€ ๊ธฐ์ค€์„ ๋ณด๋‹ค ๊ฐœ์„ ๋˜์ง€ ์•Š์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We evaluate all these methods on the validation set by generating 6 base reasonings for ensembling, using our optimal system setup. Trimmed mean achieves the lowest Brier score; see Table 14 for the results. The USC method, in contrast, does not demonstrate improvement over the baseline. [|assistant|]
๋ชจ๋ธ ๋ณ‘ํ•ฉ[15; 28], ์ตœ๊ทผ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM) ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ๊ฐœ๋ฐœ๋œ ๋ชจ๋ธ ๋ณ‘ํ•ฉ์€ ์ƒˆ๋กœ์šด ํŒจ๋Ÿฌ๋‹ค์ž„์˜ ๋ณ€ํ™”๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ LLM์„ ํ•˜๋‚˜์˜ ์•„ํ‚คํ…์ฒ˜๋กœ ์ „๋žต์ ์œผ๋กœ ๊ฒฐํ•ฉํ•จ์œผ๋กœ์จ ์ด ํฅ๋ฏธ๋กœ์šด ๊ฐœ๋ฐœ์€ ์ƒˆ๋กœ์šด ๋ชจ๋ธ ๊ฐœ๋ฐœ์„ ์œ„ํ•œ ๋งค์šฐ ๋น„์šฉ ํšจ์œจ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์ด๋ผ๋Š” ์ฃผ์š” ์žฅ์ ์œผ๋กœ ์ธํ•ด ์—ฐ๊ตฌ์ž๋“ค์˜ ๊ด€์‹ฌ์„ ๋Œ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ์„ฑ ๋•๋ถ„์— ๋ชจ๋ธ ๋ณ‘ํ•ฉ์— ๋Œ€ํ•œ ๊ด€์‹ฌ๊ณผ ์‹คํ—˜์ด ๊ธ‰์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด์ œ Open LLM Leaderboard[20]๋Š” ๋ณ‘ํ•ฉ ๋ชจ๋ธ๋กœ ์žฅ์•…ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, foundation model ๊ฐœ๋ฐœ์˜ ๋ฏผ์ฃผํ™”๋ฅผ ์œ„ํ•œ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ตœ๊ทผ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM) ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ๊ฐœ๋ฐœ๋œ ๋ชจ๋ธ ๋ณ‘ํ•ฉ[15; 28]์€ ์ƒˆ๋กœ์šด ํŒจ๋Ÿฌ๋‹ค์ž„์˜ ๋ณ€ํ™”๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ LLM์„ ํ•˜๋‚˜์˜ ์•„ํ‚คํ…์ฒ˜๋กœ ์ „๋žต์ ์œผ๋กœ ๊ฒฐํ•ฉํ•จ์œผ๋กœ์จ ์ด ํฅ๋ฏธ๋กœ์šด ๊ฐœ๋ฐœ์€ ์ƒˆ๋กœ์šด ๋ชจ๋ธ ๊ฐœ๋ฐœ์„ ์œ„ํ•œ ๋งค์šฐ ๋น„์šฉ ํšจ์œจ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์ด๋ผ๋Š” ์ฃผ์š” ์žฅ์ ์œผ๋กœ ์ธํ•ด ์—ฐ๊ตฌ์ž๋“ค์˜ ๊ด€์‹ฌ์„ ๋Œ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ์„ฑ ๋•๋ถ„์— ๋ชจ๋ธ ๋ณ‘ํ•ฉ์— ๋Œ€ํ•œ ๊ด€์‹ฌ๊ณผ ์‹คํ—˜์ด ๊ธ‰์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด์ œ ์˜คํ”ˆ LLM ๋ฆฌ๋”๋ณด๋“œ[20]๋Š” ๋ณ‘ํ•ฉ ๋ชจ๋ธ๋กœ ์žฅ์•…ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ๊ธฐ๋ฐ˜ ๋ชจ๋ธ ๊ฐœ๋ฐœ์˜ ๋ฏผ์ฃผํ™”๋ฅผ ์œ„ํ•œ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Model merging [15; 28], a recent development in the large language model (LLM) community, presents a novel paradigm shift. By strategically combining multiple LLMs into a single architecture, this exciting development has captured the attention of researchers due to its key advantage: it requires no additional training, making it an incredibly cost-effective approach for developing new models. This accessibility has fueled a surge in interest and experimentation with model merging. The Open LLM Leaderboard [20] is now dominated by merged models, showcasing its potential for democratizing foundation model development. [|assistant|]
๋ช…๋ น์–ด ํŠœ๋‹๊ณผ๋Š” ๋ณ„๋„๋กœ, LLM์„ ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์„ ํ˜ธ๋„(์˜ˆ: ์œ ์šฉ์„ฑ, ์ •์ง์„ฑ, ๋ฌดํ•ด์„ฑ)์— ๋งž์ถ”๊ธฐ ์œ„ํ•ด์„œ๋Š” ๊ณ ํ’ˆ์งˆ์˜ ๋ฐ์ดํ„ฐ์…‹์„ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” alignment tuning์— ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ(HH-RLHF [170], SHP [177], PKU-SafeRLHF [181], Stack Exchange Preferences [178] ๋ฐ Sandbox Alignment Data [179])๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ํ‘œ IV์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ช…๋ น์–ด ํŠœ๋‹๊ณผ๋Š” ๋ณ„๋„๋กœ, LLM์„ ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์„ ํ˜ธ๋„(์˜ˆ: ์œ ์šฉ์„ฑ, ์ •์ง์„ฑ, ๋ฌดํ•ด์„ฑ)์— ๋งž์ถ”๊ธฐ ์œ„ํ•ด์„œ๋Š” ๊ณ ํ’ˆ์งˆ์˜ ๋ฐ์ดํ„ฐ์…‹์„ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ •๋ ฌ ํŠœ๋‹์— ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ(HH-RLHF [170], SHP [177], PKU-SafeRLHF [181], Stack Exchange Preferences [178] ๋ฐ Sandbox Alignment Data [179])๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ํ‘œ IV์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Apart from instruction tuning, it is important to construct high-quality datasets for aligning LLMs with human values and preferences (_e.g.,_ helpfulness, honesty, and harmlessness). In this section, we introduce several widely used datasets for alignment tuning, including HH-RLHF [170], SHP [177], PKU-SafeRLHF [181], Stack Exchange Preferences [178] and Sandbox Alignment Data [179]. We show their details in Table IV. [|assistant|]
๊ทธ๋ฆผ 2: ์ œ์•ˆ๋œ CodecLM ๊ฐœ์š”. ๋จผ์ €, ๊ฐ•๋ ฅํ•œ LLM \(f_{s}\)๋Š” ์‹œ๋“œ instruction์„ instruction metadata๋กœ ์ธ์ฝ”๋”ฉํ•˜์—ฌ ์‚ฌ์šฉ ์‚ฌ๋ก€์™€ ์‘๋‹ต์— ํ•„์š”ํ•œ skills์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, \(f_{s}\)๋Š” metadata๋ฅผ ๊ธฐ๋ณธ instructions๋กœ ๋””์ฝ”๋”ฉํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, Self-Rubrics๋Š” \(f_{s}\)๋ฅผ ํ™œ์šฉํ•˜์—ฌ rubrics์™€ actions์„ ์ƒ์„ฑํ•˜์—ฌ ๊ธฐ๋ณธ instruction์„ ๊ฐœ์„ ํ•˜๊ณ  downstream ์ž‘์—…์— ๋งž๊ฒŒ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, Contrastive Filtering์€ ์ ์ˆ˜ ํ•จ์ˆ˜ \(S\)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ \(f_{s}\)์™€ \(f_{t}\)์˜ ์‘๋‹ต์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ํšจ๊ณผ์ ์ธ ์Œ์€ LLM์„ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•ด ์„ ํƒ๋˜๋ฉฐ, ํšจ๊ณผ๊ฐ€ ๋–จ์–ด์ง€๋Š” instructions๋Š” ์ถ”๊ฐ€ ๊ฐœ์„ ์„ ์œ„ํ•ด ์ „์†ก๋ฉ๋‹ˆ๋‹ค. ์ด ๊ทธ๋ฆผ์—์„œ ๊ฐ•๋ ฅํ•œ LLM์˜ ์‘๋‹ต์ด ๋Œ€์ƒ LLM์˜ ์‘๋‹ต์— ๋น„ํ•ด ์šฐ์„ธํ•˜๋ฏ€๋กœ ๋Œ€์ƒ LLM์˜ Instruction Tuning์„ ์œ„ํ•ด ํ•ด๋‹น ์Œ์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 2: ์ œ์•ˆ๋œ ์ฝ”๋ฑLM ๊ฐœ์š”. ๋จผ์ €, ๊ฐ•๋ ฅํ•œ LLM \(f_{s}\)๋Š” ์‹œ๋“œ ๋ช…๋ น์„ ์ธ์ŠคํŠธ๋Ÿญ์…˜ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋กœ ์ธ์ฝ”๋”ฉํ•˜์—ฌ ์‚ฌ์šฉ ์‚ฌ๋ก€์™€ ์‘๋‹ต์— ํ•„์š”ํ•œ ์Šคํ‚ฌ์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, \(f_{s}\)๋Š” ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ณธ ๋ช…๋ น์–ด๋กœ ๋””์ฝ”๋”ฉํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, ์…€ํ”„-๋ฃจ๋ธŒ๋ฆญ์Šค๋Š” \(f_{s}\)๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋ฃจ๋ธŒ๋ฆญ๊ณผ ๋™์ž‘์„ ์ƒ์„ฑํ•˜์—ฌ ๊ธฐ๋ณธ ๋ช…๋ น์„ ๊ฐœ์„ ํ•˜๊ณ  ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋งž๊ฒŒ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๋Œ€์กฐ ํ•„ํ„ฐ๋ง์€ ์ ์ˆ˜ ํ•จ์ˆ˜ \(S\)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ \(f_{s}\)์™€ \(f_{t}\)์˜ ์‘๋‹ต์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์žฅ ํšจ๊ณผ์ ์ธ ์Œ์€ LLM์„ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•ด ์„ ํƒ๋˜๋ฉฐ, ํšจ๊ณผ๊ฐ€ ๋–จ์–ด์ง€๋Š” ๋ช…๋ น์–ด๋Š” ์ถ”๊ฐ€ ๊ฐœ์„ ์„ ์œ„ํ•ด ์ „์†ก๋ฉ๋‹ˆ๋‹ค. ์ด ๊ทธ๋ฆผ์—์„œ ๊ฐ•๋ ฅํ•œ LLM์˜ ์‘๋‹ต์ด ๋Œ€์ƒ LLM์˜ ์‘๋‹ต์— ๋น„ํ•ด ์šฐ์„ธํ•˜๋ฏ€๋กœ ๋Œ€์ƒ LLM์˜ ๋ช…๋ น์–ด ํŠœ๋‹์„ ์œ„ํ•ด ํ•ด๋‹น ์Œ์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 2: Overview of the proposed CodecLM. First, the strong LLM \(f_{s}\) encodes the seed instruction into instruction metadata, specifying its use case and skills required for responses. Next, \(f_{s}\) decodes metadata into basic instructions. Meanwhile, Self-Rubrics leverages \(f_{s}\) to generate rubrics and actions to improve the basic instruction, tailoring them for the downstream task. Finally, Contrastive Filtering uses a scoring function \(S\) to compares \(f_{s}\) and \(f_{t}\)โ€™s responses. The most effective pairs are selected for aligning the LLM, while less effective instructions are sent for further improvement. In this figure, the strong LLMโ€™s response is winning against the target oneโ€™s, so we select the corresponding pair for instruction tuning the target LLM. [|assistant|]
๊ทธ๋ฆผ 9: **์ค‘์‹ฌ ์ง‘๊ณ„์—์„œ ์ค‘๋ณต ์ œ๊ฑฐ๋Š” Pythia ์ œํ’ˆ๊ตฐ์— ์•ฝ๊ฐ„์˜ ๊ฐœ์„ ์„ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค(Biderman et al., 2023).** ์ค‘๋ณต ์ œ๊ฑฐ๋œ Pile๊ณผ ์ผ๋ฐ˜ Pile์—์„œ ํ•™์Šต๋œ Pythia ๊ฐ„์˜ ๊ฐญ, ์ค‘์‹ฌ ์ง‘๊ณ„์—์„œ zero-shot ์„ฑ๋Šฅ. ํˆฌ๋ช…ํ•œ ์ง‘๊ณ„์— ์žˆ๋Š” ๋ชจ๋“  ์ž‘์—…์— ๋Œ€ํ•œ ๊ฐœ๋ณ„ ๊ณก์„  ๋ฐ 1-\(\sigma\) ํ‘œ์ค€ ํŽธ์ฐจ.[|endofturn|]
๊ทธ๋ฆผ 9: **์ค‘์‹ฌ ์ง‘๊ณ„์—์„œ ์ค‘๋ณต ์ œ๊ฑฐ๋Š” Pythia ์ œํ’ˆ๊ตฐ์— ์•ฝ๊ฐ„์˜ ๊ฐœ์„ ์„ ๊ฐ€์ ธ์˜ต๋‹ˆ๋‹ค(Biderman et al., 2023).** ์ค‘๋ณต ์ œ๊ฑฐ๋œ ํŒŒ์ผ๊ณผ ์ผ๋ฐ˜ ํŒŒ์ผ์—์„œ ํ•™์Šต๋œ Pythia ๊ฐ„์˜ ๊ฐญ, ์ค‘์‹ฌ ์ง‘๊ณ„์—์„œ 0-์ƒท ์„ฑ๋Šฅ. ํˆฌ๋ช…ํ•œ ์ง‘๊ณ„์— ์žˆ๋Š” ๋ชจ๋“  ์ž‘์—…์— ๋Œ€ํ•œ ๊ฐœ๋ณ„ ๊ณก์„  ๋ฐ 1-\(\sigma\) ํ‘œ์ค€ ํŽธ์ฐจ.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 9: **In our core aggregate, deduplication brings a small improvement to the Pythia suite (Biderman et al., 2023).** Zero-shot performance on our core aggregate, gap between Pythia trained on the deduplicated and vanilla Pile. Individual curves for per-task results and 1-\(\sigma\) standard deviation across all tasks in the aggregate in transparent. [|assistant|]
\begin{table} \begin{tabular}{c c c c c c c c} \hline \hline \multicolumn{3}{c}{**Massive web datasets**} & \multicolumn{2}{c}{**Curated**} & **Ours** & \\ \hline & OSCAR-21.09 & OSCAR-22.01 & C4 & \(\blacktriangledown\) The Pile & RW-Raw & RW-Filtered & \(\blacklozenge\) RefinedWeb \\ \hline **1B@27GT** & 55.0\% & 52.7\% & 55.7\% & 53.4\% & 52.7\% & 54.3\% & **56.2\%** \\ **3B@60GT** & 59.1\% & 55.9\% & 59.6\% & 57.9\% & 57.4\% & 58.2\% & **59.8\%** \\ \hline \hline \end{tabular} \end{table} ํ‘œ 4: **zero-shot ์ผ๋ฐ˜ํ™”๋ฅผ ์œ„ํ•œ ์‹ค๋ฒ„ ๋ถˆ๋ฆฟ์ด ์•„๋‹Œ ํ๋ ˆ์ด์…˜: \(\blacklozenge\) RefinedWeb์—์„œ ํ›ˆ๋ จ๋œ ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ์ด ์›น ๋ฐ์ดํ„ฐ(C4, OSCAR)์™€ ํ๋ ˆ์ด์…˜๋œ ์ฝ”ํผ์Šค(\(\blacktriangledown\) The Pile)์—์„œ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ ์ง‘๊ณ„์—์„œ zero-shot์˜ ํ‰๊ท  ์ •ํ™•๋„. ๋ชจ๋“  ๋ชจ๋ธ์€ ๋™์ผํ•œ ์•„ํ‚คํ…์ฒ˜์™€ ์‚ฌ์ „ ํ•™์Šต ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. OSCAR-22.01์€ ์ค‘๋ณต ์ œ๊ฑฐ๊ฐ€ ์„ ํƒ ์‚ฌํ•ญ์ผ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณด๋‹ค ํฌ๊ฒŒ ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. C4๋Š” ๊ฐ•๋ ฅํ•œ baseline์ด๋ฉฐ, OSCAR-21.09๋Š” ์•ฝ๊ฐ„ ๋’ค์ฒ˜์ง€์ง€๋งŒ, RefinedWeb์ด ๋‘ ์›น ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ํ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ The Pile๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•„ํ„ฐ๋ง๊ณผ ์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๋ชจ๋‘ zero-shot ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ํฌ๊ฒŒ ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค.**BigScience Architecture ๋ฐ Scaling Pile ๋ชจ๋ธ(Scao et al., 2022), PaLM-8B(Chowdhery et al., 2022), Aleph Alpha Luminous 13B(Aleph Alpha, 2023), Pythia ์‹œ๋ฆฌ์ฆˆ(Biderman et al., 2023), Cerebras-GPT ์‹œ๋ฆฌ์ฆˆ(Dey et al., 2023). GPT-3์˜ ๊ฒฝ์šฐ, API(babbage ๋ฐ curie)๋ฅผ ํ†ตํ•ด ์–ป์€ ๊ฒฐ๊ณผ์™€ EleutherAI LM ํ‰๊ฐ€ harness(*), ๊ทธ๋ฆฌ๊ณ  ๋‹ค๋ฅธ ํ‰๊ฐ€ ์„ค์ •์œผ๋กœ ๋ณด๊ณ ๋œ ๊ฒฐ๊ณผ(\(\dagger\))๋ฅผ ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค. PaLM๊ณผ OPT์˜ ๊ฒฝ์šฐ ๋‹ค๋ฅธ ํ‰๊ฐ€ ์ œํ’ˆ๊ตฐ(\(\dagger\))์œผ๋กœ ๊ฒฐ๊ณผ๋ฅผ ์–ป์—ˆ์ง€๋งŒ, ๋‹ค๋ฅธ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ‰๊ฐ€ harness(*), ์ฆ‰ ๋ณด๋‹ค ์ง์ ‘์ ์ธ ๋น„๊ต๊ฐ€ ๊ฐ€๋Šฅํ•œ ๊ฒฐ๊ณผ๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
\begin{table} \begin{tabular}{c c c c c c c c} \hline \hline \multicolumn{3}{c}{**Massive web datasets**} & \multicolumn{2}{c}{**Curated**} & **Ours** & \\ \hline & OSCAR-21.09 & OSCAR-22.01 & C4 & \(\blacktriangledown\) The Pile & RW-Raw & RW-Filtered & \(\blacklozenge\) RefinedWeb \\ \hline **1B@27GT** & 55.0\% & 52.7\% & 55.7\% & 53.4\% & 52.7\% & 54.3\% & **56.2\%** \\ **3B@60GT** & 59.1\% & 55.9\% & 59.6\% & 57.9\% & 57.4\% & 58.2\% & **59.8\%** \\ \hline \hline \end{tabular} \end{table} ํ‘œ 4: **์ œ๋กœ ์ƒท ์ผ๋ฐ˜ํ™”๋ฅผ ์œ„ํ•œ ์‹ค๋ฒ„ ๋ถˆ๋ฆฟ์ด ์•„๋‹Œ ํ๋ ˆ์ด์…˜: \(\blacklozenge\) RefinedWeb์—์„œ ํ›ˆ๋ จ๋œ ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ์ด ์›น ๋ฐ์ดํ„ฐ(C4, OSCAR)์™€ ํ๋ ˆ์ด์…˜๋œ ์ฝ”ํผ์Šค(\(\blacktriangledown\) The Pile)์—์„œ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ ์ง‘๊ณ„์—์„œ ์ œ๋กœ ์ƒท์˜ ํ‰๊ท  ์ •ํ™•๋„. ๋ชจ๋“  ๋ชจ๋ธ์€ ๋™์ผํ•œ ์•„ํ‚คํ…์ฒ˜์™€ ์‚ฌ์ „ ํ•™์Šต ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. OSCAR-22.01์€ ์ค‘๋ณต ์ œ๊ฑฐ๊ฐ€ ์„ ํƒ ์‚ฌํ•ญ์ผ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณด๋‹ค ํฌ๊ฒŒ ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. C4๋Š” ๊ฐ•๋ ฅํ•œ ๊ธฐ์ค€์„ ์ด๋ฉฐ, OSCAR-21.09๋Š” ์•ฝ๊ฐ„ ๋’ค์ฒ˜์ง€์ง€๋งŒ, RefinedWeb์ด ๋‘ ์›น ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ํ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ The Pile๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•„ํ„ฐ๋ง๊ณผ ์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๋ชจ๋‘ ์ œ๋กœ ์ƒท ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ํฌ๊ฒŒ ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค.**๋น…์‚ฌ์ด์–ธ์Šค ์•„ํ‚คํ…์ฒ˜ ๋ฐ ์Šค์ผ€์ผ๋ง ํŒŒ์ผ ๋ชจ๋ธ(Scao et al., 2022), PaLM-8B(Chowdhery et al., 2022), ์•Œ๋ ˆํ”„ ์•ŒํŒŒ ๋ฃจ๋ฏธ๋„ˆ์Šค 13B(Aleph Alpha, 2023), ํ”ผํ‹ฐ์•„ ์‹œ๋ฆฌ์ฆˆ(Biderman et al., 2023), ์„ธ๋ ˆ๋ธŒ๋ผ์Šค-GPT ์‹œ๋ฆฌ์ฆˆ(Dey et al., 2023). GPT-3์˜ ๊ฒฝ์šฐ, API(babbage ๋ฐ curie)๋ฅผ ํ†ตํ•ด ์–ป์€ ๊ฒฐ๊ณผ์™€ EleutherAI LM ํ‰๊ฐ€ ํ•˜๋„ค์Šค(*), ๊ทธ๋ฆฌ๊ณ  ๋‹ค๋ฅธ ํ‰๊ฐ€ ์„ค์ •์œผ๋กœ ๋ณด๊ณ ๋œ ๊ฒฐ๊ณผ(\(\dagger\))๋ฅผ ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค. PaLM๊ณผ OPT์˜ ๊ฒฝ์šฐ ๋‹ค๋ฅธ ํ‰๊ฐ€ ์ œํ’ˆ๊ตฐ(\(\dagger\))์œผ๋กœ ๊ฒฐ๊ณผ๋ฅผ ์–ป์—ˆ์ง€๋งŒ, ๋‹ค๋ฅธ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ‰๊ฐ€ ํ•˜๋„ค์Šค(*), ์ฆ‰ ๋ณด๋‹ค ์ง์ ‘์ ์ธ ๋น„๊ต๊ฐ€ ๊ฐ€๋Šฅํ•œ ๊ฒฐ๊ณผ๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\begin{table} \begin{tabular}{c c c c c c c c} \hline \hline \multicolumn{3}{c}{**Massive web datasets**} & \multicolumn{2}{c}{**Curated**} & **Ours** & \\ \hline & OSCAR-21.09 & OSCAR-22.01 & C4 & \(\blacktriangledown\) The Pile & RW-Raw & RW-Filtered & \(\blacklozenge\) RefinedWeb \\ \hline **1B@27GT** & 55.0\% & 52.7\% & 55.7\% & 53.4\% & 52.7\% & 54.3\% & **56.2\%** \\ **3B@60GT** & 59.1\% & 55.9\% & 59.6\% & 57.9\% & 57.4\% & 58.2\% & **59.8\%** \\ \hline \hline \end{tabular} \end{table} Table 4: **Curation is not a silver bullet for zero-shot generalization: small-scale models trained on \(\blacklozenge\) RefinedWeb outperform models trained on web data (C4, OSCAR), and on curated corpora (\(\blacktriangledown\) The Pile). Average accuracy in zero-shot on the small-agg aggregate. All models trained with identical architectures and pretraining hyperparameters. We find that OSCAR-22.01 underperforms other datasets signficantly, perhaps because deduplication is only optional. C4 is a strong baseline, with OSCAR-21.09 lagging slightly behind, but we find that RefinedWeb outperforms both web datasets and the most popular curated dataset, The Pile. Both filtering and deduplication contribute significantly to improving zero-shot performance.**the BigScience Architecture and Scaling Pile model (Scao et al., 2022), PaLM-8B (Chowdhery et al., 2022), Aleph Alpha Luminous 13B (Aleph Alpha, 2023), the Pythia series (Biderman et al., 2023), and the Cerebras-GPT series (Dey et al., 2023). For GPT-3, we distinguish between results obtained through the API (babbage and curie) with the the EleutherAI LM evaluation harness (Gao et al., 2021) (*), and results reported in their paper, with a different evaluation setup (\(\dagger\)). Note that for PaLM and OPT, results were also obtained with a different evaluation suite (\(\dagger\)), while for other models they were obtained with the evaluation harness as well (*), allowing for more direct comparisons. [|assistant|]
์—ฌ๊ธฐ์„œ \(d\)๋Š” embedding ์ฐจ์›์ด๋ฉฐ ๋ชจ๋ธ ์šฉ๋Ÿ‰์˜ ๋Œ€๋ฆฌ ์—ญํ• ์„ ํ•˜๋ฉฐ, ์ง€์ˆ˜ \(c_{q}\)๋Š” \(\beta\)์™€ ํ›ˆ๋ จ ์ค‘์— ๋ชจ๋ธ์˜ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ํŠน์ • ์•Œ๊ณ ๋ฆฌ์ฆ˜ \(q\)์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค.[|endofturn|]
์—ฌ๊ธฐ์„œ \(d\)๋Š” ์ž„๋ฒ ๋”ฉ ์ฐจ์›์ด๋ฉฐ ๋ชจ๋ธ ์šฉ๋Ÿ‰์˜ ๋Œ€๋ฆฌ ์—ญํ• ์„ ํ•˜๋ฉฐ, ์ง€์ˆ˜ \(c_{q}\)๋Š” \(\beta\)์™€ ํ›ˆ๋ จ ์ค‘์— ๋ชจ๋ธ์˜ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ํŠน์ • ์•Œ๊ณ ๋ฆฌ์ฆ˜ \(q\)์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]where \(d\) is the embedding dimension, and serves a s proxy for model capacity; the exponent \(c_{q}\) depends both on \(\beta\) and the particular algorithm \(q\) used to update the memories in the model during training. [|assistant|]
์—ฌ๊ธฐ์„œ \(q\)๋Š” ์—์ด์ „ํŠธ๊ฐ€ ํ•ด๊ฒฐํ•ด์•ผ ํ•˜๋Š” ์ž‘์—… ๋˜๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ฒ˜ํ•œ ์ƒํ™ฉ๊ณผ ๊ฐ™์€ query์ž…๋‹ˆ๋‹ค. \(M\)์€ ๋ชจ๋“  ๋ฉ”๋ชจ๋ฆฌ์˜ ์ง‘ํ•ฉ์ž…๋‹ˆ๋‹ค. \(s^{rec}(\cdot)\), \(s^{rel}(\cdot)\) ๋ฐ \(s^{imp}(\cdot)\)๋Š” ๋ฉ”๋ชจ๋ฆฌ \(m\)์˜ ์ตœ๊ทผ์„ฑ, ๊ด€๋ จ์„ฑ ๋ฐ ์ค‘์š”๋„๋ฅผ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•œ scoring functions์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ scoring functions๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์˜ˆ๋ฅผ ๋“ค์–ด \(s^{rel}(q,m)\)์€ LSH, ANNOY, HNSW, FAISS ๋“ฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. \(s^{imp}\)๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ž์ฒด์˜ ํŠน์„ฑ๋งŒ ๋ฐ˜์˜ํ•˜๋ฏ€๋กœ query \(q\)์™€๋Š” ๊ด€๋ จ์ด ์—†๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. \(\alpha\), \(\beta\) ๋ฐ \(\gamma\)๋Š” balancing parameters์ž…๋‹ˆ๋‹ค. ์ด ๊ฐ’์„ ๋‹ค๋ฅด๊ฒŒ ์ง€์ •ํ•˜๋ฉด ๋‹ค์–‘ํ•œ ๋ฉ”๋ชจ๋ฆฌ ์ฝ๊ธฐ ์ „๋žต์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(\alpha=\gamma=0\)์œผ๋กœ ์„ค์ •ํ•˜๋ฉด ๋งŽ์€ ์—ฐ๊ตฌ[16, 30, 38, 42]์—์„œ ๋ฉ”๋ชจ๋ฆฌ ์ฝ๊ธฐ์— ๊ด€๋ จ ์ ์ˆ˜ \(s^{rel}\)๋งŒ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. \(\alpha=\beta=\gamma=1.0\), [20]์€ ์œ„์˜ ์„ธ ๊ฐ€์ง€ metrics์— ๋ชจ๋‘ ๋™์ผํ•œ ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์—ฌ๊ธฐ์„œ \(q\)๋Š” ์—์ด์ „ํŠธ๊ฐ€ ํ•ด๊ฒฐํ•ด์•ผ ํ•˜๋Š” ์ž‘์—… ๋˜๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ฒ˜ํ•œ ์ƒํ™ฉ๊ณผ ๊ฐ™์€ ์ฟผ๋ฆฌ์ž…๋‹ˆ๋‹ค. M\)์€ ๋ชจ๋“  ๋ฉ”๋ชจ๋ฆฌ์˜ ์ง‘ํ•ฉ์ž…๋‹ˆ๋‹ค. s^{rec}(\cdot)\), \(s^{rel}(\cdot)\) ๋ฐ \(s^{imp}(\cdot)\)๋Š” ๋ฉ”๋ชจ๋ฆฌ \(m\)์˜ ์ตœ๊ทผ์„ฑ, ๊ด€๋ จ์„ฑ ๋ฐ ์ค‘์š”๋„๋ฅผ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•œ ์ ์ˆ˜ ํ•จ์ˆ˜์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ์ˆ˜ ํ•จ์ˆ˜๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์˜ˆ๋ฅผ ๋“ค์–ด \(s^{rel}(q,m)\)์€ LSH, ANNOY, HNSW, FAISS ๋“ฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. s^{imp}\)๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ž์ฒด์˜ ํŠน์„ฑ๋งŒ ๋ฐ˜์˜ํ•˜๋ฏ€๋กœ ์ฟผ๋ฆฌ \(q\)์™€๋Š” ๊ด€๋ จ์ด ์—†๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. \(\์•ŒํŒŒ\), \(\๋ฒ ํƒ€\), \(\๊ฐ๋งˆ\)๋Š” ๋ฐธ๋Ÿฐ์‹ฑ ํŒŒ๋ผ๋ฏธํ„ฐ์ž…๋‹ˆ๋‹ค. ์ด ๊ฐ’์„ ๋‹ค๋ฅด๊ฒŒ ์ง€์ •ํ•˜๋ฉด ๋‹ค์–‘ํ•œ ๋ฉ”๋ชจ๋ฆฌ ์ฝ๊ธฐ ์ „๋žต์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, \(\์•ŒํŒŒ=\๊ฐ๋งˆ=0\)์œผ๋กœ ์„ค์ •ํ•˜๋ฉด ๋งŽ์€ ์—ฐ๊ตฌ[16, 30, 38, 42]์—์„œ ๋ฉ”๋ชจ๋ฆฌ ์ฝ๊ธฐ์— ๊ด€๋ จ ์ ์ˆ˜ \(s^{rel}\)๋งŒ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์•ŒํŒŒ=\๋ฒ ํƒ€=\๊ฐ๋งˆ=1.0\), [20]์€ ์œ„์˜ ์„ธ ๊ฐ€์ง€ ๋ฉ”ํŠธ๋ฆญ์— ๋ชจ๋‘ ๋™์ผํ•œ ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]where \(q\) is the query, for example, the task that the agent should address or the context in which the agent is situated. \(M\) is the set of all memories. \(s^{rec}(\cdot)\), \(s^{rel}(\cdot)\) and \(s^{imp}(\cdot)\) are the scoring functions for measuring the recency, relevance, and importance of the memory \(m\). These scoring functions can be implemented using various methods, for example, \(s^{rel}(q,m)\) can be realized based on LSH, ANNOY, HNSW, FAISS and so on. It should be noted that \(s^{imp}\) only reflects the characters of the memory itself, thus it is unrelated to the query \(q\). \(\alpha\), \(\beta\) and \(\gamma\) are balancing parameters. By assigning them with different values, one can obtain various memory reading strategies. For example, by setting \(\alpha=\gamma=0\), many studies [16, 30, 38, 42] only consider the relevance score \(s^{rel}\) for memory reading. By assigning \(\alpha=\beta=\gamma=1.0\), [20] equally weights all the above three metrics to extract information from memory. [|assistant|]
๋‹จ์ผ ์—์ด์ „ํŠธ. AIOS์˜ ์ด ์˜๊ตฌ ์ €์žฅ์†Œ๋Š” ๋กœ์ปฌ ํŒŒ์ผ, ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค ๋˜๋Š” ํด๋ผ์šฐ๋“œ ๊ธฐ๋ฐ˜ ์†”๋ฃจ์…˜๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ๋‚ด๊ตฌ์„ฑ ์žˆ๋Š” ๋งค์ฒด๋ฅผ ํ†ตํ•ด ์ด๋ฃจ์–ด์ง€๋ฉฐ, ํ–ฅํ›„ ์ฐธ์กฐ ๋˜๋Š” ๋ถ„์„์„ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ๋ฌด๊ฒฐ์„ฑ๊ณผ ๊ฐ€์šฉ์„ฑ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์Šคํ† ๋ฆฌ์ง€ ๊ด€๋ฆฌ์ž๋Š” Retrieval augmentation[61]์„ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž ๊ธฐ๋ณธ ์„ค์ •์„ ์ €์žฅํ•˜๊ณ  ๊ณผ๊ฑฐ ์ƒํ˜ธ ์ž‘์šฉ ๋กœ๊ทธ๋ฅผ ์œ ์ง€ ๊ด€๋ฆฌํ•จ์œผ๋กœ์จ ์Šคํ† ๋ฆฌ์ง€ ๊ด€๋ฆฌ์ž๋Š” ์—์ด์ „ํŠธ ์ง€์‹ ์—…๋ฐ์ดํŠธ๋ฅผ ๊ฐ•ํ™”ํ•˜๊ณ  ์žฅ๊ธฐ์ ์ธ ์‚ฌ์šฉ์ž ๊ฒฝํ—˜์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋‹จ์ผ ์—์ด์ „ํŠธ. AIOS์˜ ์ด ์˜๊ตฌ ์ €์žฅ์†Œ๋Š” ๋กœ์ปฌ ํŒŒ์ผ, ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค ๋˜๋Š” ํด๋ผ์šฐ๋“œ ๊ธฐ๋ฐ˜ ์†”๋ฃจ์…˜๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ๋‚ด๊ตฌ์„ฑ ์žˆ๋Š” ๋งค์ฒด๋ฅผ ํ†ตํ•ด ์ด๋ฃจ์–ด์ง€๋ฉฐ, ํ–ฅํ›„ ์ฐธ์กฐ ๋˜๋Š” ๋ถ„์„์„ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ๋ฌด๊ฒฐ์„ฑ๊ณผ ๊ฐ€์šฉ์„ฑ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์Šคํ† ๋ฆฌ์ง€ ๊ด€๋ฆฌ์ž๋Š” ๊ฒ€์ƒ‰ ์ฆ๊ฐ•[61]์„ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž ๊ธฐ๋ณธ ์„ค์ •์„ ์ €์žฅํ•˜๊ณ  ๊ณผ๊ฑฐ ์ƒํ˜ธ ์ž‘์šฉ ๋กœ๊ทธ๋ฅผ ์œ ์ง€ ๊ด€๋ฆฌํ•จ์œผ๋กœ์จ ์Šคํ† ๋ฆฌ์ง€ ๊ด€๋ฆฌ์ž๋Š” ์—์ด์ „ํŠธ ์ง€์‹ ์—…๋ฐ์ดํŠธ๋ฅผ ๊ฐ•ํ™”ํ•˜๊ณ  ์žฅ๊ธฐ์ ์ธ ์‚ฌ์šฉ์ž ๊ฒฝํ—˜์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]single agent. This permanent storage in AIOS is achieved through a variety of durable mediums such as local files, databases, or cloud-based solutions, ensuring data integrity and availability for future reference or analysis. The storage manager supports retrieval augmentation [61]. Through storing user preferences and maintaining historical interaction logs, the storage manager can enrich the agent knowledge update and enhancing the long-term user experience. [|assistant|]
์—ฌ๊ธฐ์„œ \(\epsilon\)์€ ๋ฏธ๋ฆฌ ์ •์˜๋œ ์ž„๊ณ„๊ฐ’์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋“  ๋‰ด๋Ÿฐ๊ณผ ๋ชจ๋“  ์ž…๋ ฅ์„ ์ˆœ์ฐจ์ ์œผ๋กœ ์ˆœํšŒํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ฆฝ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ€์†ํ™”๋ฅผ ์œ„ํ•œ parallel algorithm์„ ์„ค๊ณ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์—ฌ๊ธฐ์„œ \(\epsilon\)์€ ๋ฏธ๋ฆฌ ์ •์˜๋œ ์ž„๊ณ„๊ฐ’์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋“  ๋‰ด๋Ÿฐ๊ณผ ๋ชจ๋“  ์ž…๋ ฅ์„ ์ˆœ์ฐจ์ ์œผ๋กœ ์ˆœํšŒํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์‹œ๊ฐ„์ด ๋งŽ์ด ๊ฑธ๋ฆฝ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ€์†ํ™”๋ฅผ ์œ„ํ•œ ๋ณ‘๋ ฌ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์„ค๊ณ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]where \(\epsilon\) is the pre-defined threshold. However, it is super time-consuming to traverse all neurons and all inputs sequentially. Therefore, we need to design a parallel algorithm for acceleration. [|assistant|]
์ฃผ์˜_.: ์ฃผ์˜ ๊นŠ๊ฒŒ ์ฝ๋Š” ๋…์ž๋ผ๋ฉด ๋˜ ๋‹ค๋ฅธ ์œ ํ˜•์˜ memory ๊ตฌ์กฐ, ์ฆ‰ ์žฅ๊ธฐ ๊ธฐ์–ต์—๋งŒ ๊ธฐ๋ฐ˜ํ•œ memory ๊ตฌ์กฐ๋„ ์กด์žฌํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ๋Š” ๊ทธ๋Ÿฌํ•œ ์œ ํ˜•์˜ memory๊ฐ€ ๋ฌธํ—Œ์— ๊ฑฐ์˜ ๋ฌธ์„œํ™”๋˜์–ด ์žˆ์ง€ ์•Š๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ถ”์ธก์€ ์—์ด์ „ํŠธ๊ฐ€ ํ•ญ์ƒ ์—ฐ์†์ ์ด๊ณ  ๋™์ ์ธ ํ™˜๊ฒฝ์— ์œ„์น˜ํ•˜๋ฉฐ, ์—ฐ์†์ ์ธ ํ–‰๋™์ด ๋†’์€ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹จ๊ธฐ memory๋ฅผ ํฌ์ฐฉํ•˜๋Š” ๊ฒƒ์ด ๋งค์šฐ ์ค‘์š”ํ•˜๋ฉฐ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฌด์‹œํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ฃผ์˜_.: ์ฃผ์˜ ๊นŠ๊ฒŒ ์ฝ๋Š” ๋…์ž๋ผ๋ฉด ๋˜ ๋‹ค๋ฅธ ์œ ํ˜•์˜ ๋ฉ”๋ชจ๋ฆฌ ๊ตฌ์กฐ, ์ฆ‰ ์žฅ๊ธฐ ๊ธฐ์–ต์—๋งŒ ๊ธฐ๋ฐ˜ํ•œ ๋ฉ”๋ชจ๋ฆฌ ๊ตฌ์กฐ๋„ ์กด์žฌํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ๋Š” ๊ทธ๋Ÿฌํ•œ ์œ ํ˜•์˜ ๋ฉ”๋ชจ๋ฆฌ๊ฐ€ ๋ฌธํ—Œ์— ๊ฑฐ์˜ ๋ฌธ์„œํ™”๋˜์–ด ์žˆ์ง€ ์•Š๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ถ”์ธก์€ ์—์ด์ „ํŠธ๊ฐ€ ํ•ญ์ƒ ์—ฐ์†์ ์ด๊ณ  ๋™์ ์ธ ํ™˜๊ฒฝ์— ์œ„์น˜ํ•˜๋ฉฐ, ์—ฐ์†์ ์ธ ํ–‰๋™์ด ๋†’์€ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹จ๊ธฐ ๊ธฐ์–ต์„ ํฌ์ฐฉํ•˜๋Š” ๊ฒƒ์ด ๋งค์šฐ ์ค‘์š”ํ•˜๋ฉฐ ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฌด์‹œํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]_Remark_.: Careful readers may find that there may also exist another type of memory structure, that is, only based on the long-term memory. However, we find that such type of memory is rarely documented in the literature. Our speculation is that the agents are always situated in continuous and dynamic environments, with consecutive actions displaying a high correlation. Therefore, the capture of short-term memory is very important and usually cannot be disregarded. [|assistant|]
๊ด€๋ จ ์ž‘์—…. ๋™์‹œ ์ž‘์—…(Morris et al., 2023)์—์„œ๋Š” ์œ ์‚ฌํ•˜์ง€๋งŒ ์•ฝํ•œ ๋‘ ๊ฐ€์ง€ ์ฟผ๋ฆฌ logprob ์ถ”์ถœ์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์˜ ๊ณต๊ฒฉ์€ \(\text{logit}_{\text{top}}-\text{logit}_{i}\)๋ณด๋‹ค ํฐ logit bias์™€ top-2 logprob access๊ฐ€ ํ•„์š”ํ•˜์ง€๋งŒ, ์šฐ๋ฆฌ์˜ ๊ณต๊ฒฉ์€ logit bias๊ฐ€ 0์ด ์•„๋‹Œ ๊ฒƒ์œผ๋กœ ํ—ˆ์šฉ๋˜๋Š” ๊ฒฝ์šฐ์™€ top-1 logprob access๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ์— ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ด€๋ จ ์ž‘์—….๋™์‹œ ์ž‘์—…(Morris et al., 2023)์—์„œ๋Š” ์œ ์‚ฌํ•˜์ง€๋งŒ ์•ฝํ•œ ๋‘ ๊ฐ€์ง€ ์ฟผ๋ฆฌ ๋กœ๊ทธ ํ™•๋ฅ  ์ถ”์ถœ์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์˜ ๊ณต๊ฒฉ์€ \(\text{logit}_{\text{top}}-\text{logit}_{i}\)๋ณด๋‹ค ํฐ ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค์™€ ์ƒ์œ„ 2๊ฐœ์˜ ๋กœ๊ทธ ํ™•๋ฅ  ์•ก์„ธ์Šค๊ฐ€ ํ•„์š”ํ•˜์ง€๋งŒ, ์šฐ๋ฆฌ์˜ ๊ณต๊ฒฉ์€ ๋กœ๊ทธ ํ™•๋ฅ  ๋ฐ”์ด์–ด์Šค๊ฐ€ 0์ด ์•„๋‹Œ ๊ฒƒ์œผ๋กœ ํ—ˆ์šฉ๋˜๋Š” ๊ฒฝ์šฐ์™€ ์ƒ์œ„ 1๊ฐœ์˜ ๋กœ๊ทธ ํ™•๋ฅ  ์•ก์„ธ์Šค๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ์— ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Related work.Concurrent work (Morris et al., 2023) discusses a similar but weaker two-query logprob extraction. Their attack requires a logit bias larger than \(\text{logit}_{\text{top}}-\text{logit}_{i}\) and top-2 logprob access; our attack works as soon the logit bias is allowed to be nonzero, and with top-1 logprob access. [|assistant|]
๋‹ค์ค‘ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ Instruction ๋”ฐ๋ผํ•˜๊ธฐ. ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ LLM์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—๋Š” ์ถ”๋ก  ๋ฐ Instruction ์ˆ˜ํ–‰ ๊ธฐ๋Šฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๊ธด ํ˜•์‹์˜ ์งˆ๋ฌธ ์‘๋‹ต ์Œ์ด๋ฉฐ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•˜๊ธฐ์—๋Š” ๋น„์šฉ์ด ๋งŽ์ด ๋“ญ๋‹ˆ๋‹ค. LLaVA(Liu et al., 2024b)๋Š” ๊ธฐ์กด ์ด๋ฏธ์ง€ ์บก์…˜์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•˜๊ณ  ์žฅ๋ฌธ์˜ ํ”„๋กฌํ”„ํŠธ-๋‹ต๋ณ€ ์Œ์„ ์ž‘์„ฑํ•˜๊ธฐ ์œ„ํ•ด GPT-4(ํ…์ŠคํŠธ ์ „์šฉ ๋ชจ๋“œ)๋ฅผ ํ”„๋กฌํ”„ํŠธํ•ฉ๋‹ˆ๋‹ค. ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ LLM ํ›ˆ๋ จ ์ค‘์— ์ด๋ฏธ์ง€์™€ ํ”„๋กฌํ”„ํŠธ๋Š” ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉ๋˜๋ฉฐ ์บก์…˜๊ณผ ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค ์ •๋ณด๋Š” ์ˆจ๊ธธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€ ์บก์…˜ ์™ธ์—๋„ ๊ฐ์ฒด ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค(Zhao et al., 2023), OCR(Zhang et al., 2023d), ๋””๋”๋ง๋œ ์ฐจํŠธ(Carbune et al., 2024; Masry et al., 2023)์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ์ด๋ฏธ์ง€ ์†์„ฑ ์ •๋ณด ์†Œ์Šค๋„ ์ด๋ฏธ์ง€ ์†์„ฑ + ํ…์ŠคํŠธ LLM ์žฌ์ž‘์„ฑ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ํŒŒ์ดํ”„๋ผ์ธ์— ๋ชจ๋‘ ์ ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋‹ค์ค‘ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ ๋ช…๋ น์–ด ๋”ฐ๋ผํ•˜๊ธฐ.๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ LLM์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—๋Š” ์ถ”๋ก  ๋ฐ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ๊ธฐ๋Šฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๊ธด ํ˜•์‹์˜ ์งˆ๋ฌธ ์‘๋‹ต ์Œ์ด๋ฉฐ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•˜๊ธฐ์—๋Š” ๋น„์šฉ์ด ๋งŽ์ด ๋“ญ๋‹ˆ๋‹ค. LLaVA(Liu et al., 2024b)๋Š” ๊ธฐ์กด ์ด๋ฏธ์ง€ ์บก์…˜์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•˜๊ณ  ์žฅ๋ฌธ์˜ ํ”„๋กฌํ”„ํŠธ-๋‹ต๋ณ€ ์Œ์„ ์ž‘์„ฑํ•˜๊ธฐ ์œ„ํ•ด GPT-4(ํ…์ŠคํŠธ ์ „์šฉ ๋ชจ๋“œ)๋ฅผ ํ”„๋กฌํ”„ํŠธํ•ฉ๋‹ˆ๋‹ค. ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ LLM ํ›ˆ๋ จ ์ค‘์— ์ด๋ฏธ์ง€์™€ ํ”„๋กฌํ”„ํŠธ๋Š” ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉ๋˜๋ฉฐ ์บก์…˜๊ณผ ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค ์ •๋ณด๋Š” ์ˆจ๊ธธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€ ์บก์…˜ ์™ธ์—๋„ ๊ฐ์ฒด ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค(Zhao et al., 2023), OCR(Zhang et al., 2023d), ๋””๋”๋ง๋œ ์ฐจํŠธ(Carbune et al., 2024; Masry et al., 2023)์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ์ด๋ฏธ์ง€ ์†์„ฑ ์ •๋ณด ์†Œ์Šค๋„ ์ด๋ฏธ์ง€ ์†์„ฑ + ํ…์ŠคํŠธ LLM ์žฌ์ž‘์„ฑ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ํŒŒ์ดํ”„๋ผ์ธ์— ๋ชจ๋‘ ์ ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Multi-modality instruction following.Downstream applications of multimodal LLMs require reasoning and instruction following capabilities. Such data are usually long-form question response pairs and are expensive for humans to create. LLaVA (Liu et al., 2024b) uses existing image captions to prompt GPT-4 (in text-only mode) for writing diverse and long-form prompt-answer pairs. During multimodal LLM training, images and prompts are used as input while the captions and bounding box information can be hidden. Besides image captions, other sources of image attribute information such as object bounding box (Zhao et al., 2023), OCR (Zhang et al., 2023d) and deredered charts (Carbune et al., 2024; Masry et al., 2023) can all fit into such as image attributes + text LLM rewriting synthetic data pipeline. [|assistant|]
๊ทธ๋ฆผ 1: AI Safety Benchmark v0.5 ๋‚ด์˜ ๋‹ค์–‘ํ•œ hazard ๋ฒ”์ฃผ ๋‚ด์—์„œ ํ…Œ์ŠคํŠธ ๋Œ€์ƒ ์‹œ์Šคํ…œ(SUT)์˜ ์ ์ˆ˜. ํ•˜๋‹จ ํ–‰์—๋Š” ์ „์ฒด benchmark ์„ฑ๋Šฅ(์ฆ‰, hazard ๋ฒ”์ฃผ ๊ฐ„ ์ตœ์ € ๋“ฑ๊ธ‰)์ด ๋ณด๊ณ ๋ฉ๋‹ˆ๋‹ค. ์ฑ„์ ์„ ์œ„ํ•œ ์ฐธ์กฐ ๋ชจ๋ธ์€ (\(*\))๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 1: AI ์•ˆ์ „ ๋ฒค์น˜๋งˆํฌ v0.5 ๋‚ด์˜ ๋‹ค์–‘ํ•œ ์œ„ํ—˜ ๋ฒ”์ฃผ) ๋‚ด์—์„œ ํ…Œ์ŠคํŠธ ๋Œ€์ƒ ์‹œ์Šคํ…œ(SUT)์˜ ์ ์ˆ˜. ํ•˜๋‹จ ํ–‰์—๋Š” ์ „์ฒด ๋ฒค์น˜๋งˆํฌ ์„ฑ๋Šฅ(์ฆ‰, ์œ„ํ—˜ ๋ฒ”์ฃผ ๊ฐ„ ์ตœ์ € ๋“ฑ๊ธ‰)์ด ๋ณด๊ณ ๋ฉ๋‹ˆ๋‹ค. ์ฑ„์ ์„ ์œ„ํ•œ ์ฐธ์กฐ ๋ชจ๋ธ์€ (\(*\))๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 1: Scores of systems under test (SUTs) across the different hazard categories)within the AI Safety Benchmark v0.5. The bottom row reports the overall benchmark performance (i.e., lowest grade across hazard categories). Reference models for scoring are denoted with a (\(*\)). [|assistant|]
๋กœ์ง“๋‹น ์ฟผ๋ฆฌ ์ˆ˜์ž…๋‹ˆ๋‹ค. 6~23์ž๋ฆฌ ์ •๋ฐ€๋„๋ฅผ ์›ํ•œ๋‹ค๋ฉด ํ•˜ํ•œ์€ ๋กœ์ง“๋‹น 1.53~3.57์ฟผ๋ฆฌ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 3์˜ ์ตœ๊ณ  logprob-free ๊ณต๊ฒฉ์€ ํ•˜ํ•œ๋ณด๋‹ค ๋กœ์ง“๋‹น ์ฟผ๋ฆฌ ์ˆ˜๊ฐ€ 1๊ฐœ ์ •๋„ ๋” ๋‚˜์œ ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋กœ์ง“๋‹น ์ฟผ๋ฆฌ ์ˆ˜์ž…๋‹ˆ๋‹ค. 6~23์ž๋ฆฌ ์ •๋ฐ€๋„๋ฅผ ์›ํ•œ๋‹ค๋ฉด ํ•˜ํ•œ์€ ๋กœ์ง“๋‹น 1.53~3.57์ฟผ๋ฆฌ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ํ‘œ 3์˜ ์ตœ๊ณ  ๋กœ๊ทธ๋กœ๊ทธ๋ฆฌ์ฆ˜ ๊ณต๊ฒฉ์€ ํ•˜ํ•œ๋ณด๋‹ค ๋กœ์ง“๋‹น ์ฟผ๋ฆฌ ์ˆ˜๊ฐ€ 1๊ฐœ ์ •๋„ ๋” ๋‚˜์œ ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]queries per logit. If we want between 6 and 23 digits of precision, the lower bound corresponds to 1.53 to 3.57 queries per logit. We see that the best logprob-free attack in Table 3 is only about 1 query per logit worse than the lower bound. [|assistant|]
ํ›ˆ๋ จ ๋งค๊ฐœ๋ณ€์ˆ˜. ์šฐ๋ฆฌ๋Š” ์ฝ”์‚ฌ์ธ Learning Rate ์Šค์ผ€์ค„๋Ÿฌ์™€ ํ•จ๊ป˜ AdamW Optimizer๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” 1K ๋‹จ๊ณ„์˜ ์›Œ๋ฐ์—…์ด ํฌํ•จ๋œ ํ›„, Learning Rate๊ฐ€ ๊ธฐ์ค€ ํ•™์Šต๋ฅ ์˜ 1๋ฐฐ์—์„œ 0.1๋ฐฐ๋กœ ์ฝ”์‚ฌ์ธ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ๋‹ฌ๋ฆฌ ๋ช…์‹œ๋˜์ง€ ์•Š๋Š” ํ•œ ํ˜ผํ•ฉ ์ •๋ฐ€๋„ fp16 ํ›ˆ๋ จ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
ํ›ˆ๋ จ ๋งค๊ฐœ๋ณ€์ˆ˜.์ฝ”์‚ฌ์ธ ํ•™์Šต๋ฅ  ์Šค์ผ€์ค„๋Ÿฌ์™€ ํ•จ๊ป˜ AdamW ์˜ตํ‹ฐ๋งˆ์ด์ €๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” 1K ๋‹จ๊ณ„์˜ ์›Œ๋ฐ์—…์ด ํฌํ•จ๋œ ํ›„, ํ•™์Šต๋ฅ ์ด ๊ธฐ์ค€ ํ•™์Šต๋ฅ ์˜ 1๋ฐฐ์—์„œ 0.1๋ฐฐ๋กœ ์ฝ”์‚ฌ์ธ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ๋‹ฌ๋ฆฌ ๋ช…์‹œ๋˜์ง€ ์•Š๋Š” ํ•œ ํ˜ผํ•ฉ ์ •๋ฐ€๋„ fp16 ํ›ˆ๋ จ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Training parameters.We employ the AdamW optimizer with a cosine learning rate scheduler. This includes \(1\)K steps of warmup, followed by a cosine decay of the learning rate from \(1\) to \(0.1\) times the reference rate. We use mixed-precision fp16 training unless otherwise stated. [|assistant|]
de+en ๋ฐ zh+en ์–ธ์–ด ์Œ์˜ ๊ฒฝ์šฐ, ์†Œ์Šค ๊ธฐ๋ฐ˜ Direct Assessment ๋ฐ Scalar Quality Metrics (DA+SQM)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ€์žฅ ๋†’์€ ์‚ฌ๋žŒ์˜ ์ˆœ์œ„๋ฅผ ์–ป์€ ๋ฒˆ์—ญ ์‹œ์Šคํ…œ์„ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. de+ru์˜ ๊ฒฝ์šฐ, Kocmi ์™ธ(2023)์—์„œ ์ด๋Ÿฌํ•œ ๋ฐฉํ–ฅ์— ๋Œ€ํ•œ ์‚ฌ๋žŒ์˜ ์ˆœ์œ„๊ฐ€ ์—†์—ˆ์œผ๋ฏ€๋กœ Kocmi ์™ธ(2023)์— ๋ณด๊ณ ๋œ ๋Œ€๋กœ COMET-22 ์ ์ˆ˜๊ฐ€ ๊ฐ€์žฅ ๋†’์€ ๋ชจ๋ธ์„ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ํ‘œ 12์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
de+en ๋ฐ zh+en ์–ธ์–ด ์Œ์˜ ๊ฒฝ์šฐ, ์†Œ์Šค ๊ธฐ๋ฐ˜ ์ง์ ‘ ํ‰๊ฐ€ ๋ฐ ์Šค์นผ๋ผ ํ’ˆ์งˆ ๋ฉ”ํŠธ๋ฆญ(DA+SQM)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ€์žฅ ๋†’์€ ์‚ฌ๋žŒ์˜ ์ˆœ์œ„๋ฅผ ์–ป์€ ๋ฒˆ์—ญ ์‹œ์Šคํ…œ์„ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. de+ru์˜ ๊ฒฝ์šฐ, Kocmi ์™ธ(2023)์—์„œ ์ด๋Ÿฌํ•œ ๋ฐฉํ–ฅ์— ๋Œ€ํ•œ ์‚ฌ๋žŒ์˜ ์ˆœ์œ„๊ฐ€ ์—†์—ˆ์œผ๋ฏ€๋กœ Kocmi ์™ธ(2023)์— ๋ณด๊ณ ๋œ ๋Œ€๋กœ COMET-22 ์ ์ˆ˜๊ฐ€ ๊ฐ€์žฅ ๋†’์€ ๋ชจ๋ธ์„ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ํ‘œ 12์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]For the de+en and zh+en language pairs, we selected the translation systems that attained the highest human rankings based on source-based Direct Assessment and Scalar Quality Metrics (DA+SQM). For de+ru, in the absence of human rankings for these directions in Kocmi et al. (2023), we opted for the model with the highest COMET-22 scores as reported in Kocmi et al. (2023). Details about these models are available in Table 12. [|assistant|]
์ตœ๊ทผ์— ์ œ์•ˆ๋œ Parity (Petrov et al., 2023)๋Š” tokenizer๊ฐ€ ์„œ๋กœ ๋‹ค๋ฅธ ์–ธ์–ด์˜ ๋™๋“ฑํ•œ ๋ฌธ์žฅ์„ ์–ผ๋งˆ๋‚˜ ๊ณต์ •ํ•˜๊ฒŒ ์ฒ˜๋ฆฌํ•˜๋Š”์ง€๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. tokenizer \(T\)๋Š” \(\frac{|T(s_{A})|}{|T(s_{B})|}\approx 1\)์ด๋ฉด ์–ธ์–ด \(A\)์— ๋Œ€ํ•ด ์–ธ์–ด \(B\)์— ๋Œ€ํ•œ parity๋ฅผ ๋‹ฌ์„ฑํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(s_{A}\)์™€ \(s_{B}\)๋Š” ๊ฐ๊ฐ ์–ธ์–ด \(A\)์™€ \(B\)์˜ ์ฝ”ํผ์Šค์— ์žˆ๋Š” ๋ชจ๋“  ๋ฌธ์žฅ์˜ ์ง‘ํ•ฉ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ๋น„์œจ \(\frac{|T(s_{A})|}{T(s_{B})|}\)๋Š” premium์œผ๋กœ ์ •์˜๋ฉ๋‹ˆ๋‹ค. ๋™์ผํ•œ ๋ฌธ์žฅ์„ 200๊ฐœ ์–ธ์–ด๋กœ ์ธ๊ฐ„ ๋ฒˆ์—ญํ•œ ๊ฒƒ์œผ๋กœ ๊ตฌ์„ฑ๋œ FLORES-200(Goyal et al., 2022) ๋ณ‘๋ ฌ ์ฝ”ํผ์Šค๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฐ tokenizer์™€ ์˜์–ด๊ฐ€ ์•„๋‹Œ 4๊ฐœ ์–ธ์–ด์— ๋Œ€ํ•œ ์˜์–ด์™€์˜ parity ๊ฐ’์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค(๊ฐœ์š”๋Š” ๊ทธ๋ฆผ 2 ์ฐธ์กฐ).[|endofturn|]
์ตœ๊ทผ์— ์ œ์•ˆ๋œ ํŒจ๋ฆฌํ‹ฐ(Petrov et al., 2023)๋Š” ํ† ํฐํ™”๊ธฐ๊ฐ€ ์„œ๋กœ ๋‹ค๋ฅธ ์–ธ์–ด์˜ ๋™๋“ฑํ•œ ๋ฌธ์žฅ์„ ์–ผ๋งˆ๋‚˜ ๊ณต์ •ํ•˜๊ฒŒ ์ฒ˜๋ฆฌํ•˜๋Š”์ง€๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํ† ํฐํ™”๊ธฐ \(T\)๋Š” \(\frac{|T(s_{A})|}{|T(s_{B})|}\์•ฝ 1\)์ด๋ฉด ์–ธ์–ด \(A\)์— ๋Œ€ํ•ด ์–ธ์–ด \(B\)์— ๋Œ€ํ•œ ํŒจ๋ฆฌํ‹ฐ๋ฅผ ๋‹ฌ์„ฑํ•˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(s_{A}\)์™€ \(s_{B}\)๋Š” ๊ฐ๊ฐ ์–ธ์–ด \(A\)์™€ \(B\)์˜ ์ฝ”ํผ์Šค์— ์žˆ๋Š” ๋ชจ๋“  ๋ฌธ์žฅ์˜ ์ง‘ํ•ฉ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ๋น„์œจ \(\frac{|T(s_{A})|}{T(s_{B})|}\)๋Š” ํ”„๋ฆฌ๋ฏธ์—„์œผ๋กœ ์ •์˜๋ฉ๋‹ˆ๋‹ค. ๋™์ผํ•œ ๋ฌธ์žฅ์„ 200๊ฐœ ์–ธ์–ด๋กœ ์ธ๊ฐ„ ๋ฒˆ์—ญํ•œ ๊ฒƒ์œผ๋กœ ๊ตฌ์„ฑ๋œ FLORES-200(Goyal et al., 2022) ๋ณ‘๋ ฌ ์ฝ”ํผ์Šค๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ํ† ํฐํ™”๊ธฐ์™€ ์˜์–ด๊ฐ€ ์•„๋‹Œ 4๊ฐœ ์–ธ์–ด์— ๋Œ€ํ•œ ์˜์–ด์™€์˜ ํŒจ๋ฆฌํ‹ฐ ๊ฐ’์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค(๊ฐœ์š”๋Š” ๊ทธ๋ฆผ 2 ์ฐธ์กฐ).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Parity (Petrov et al., 2023), which has been recently proposed, assesses how fairly a tokenizer treats equivalent sentences in different languages. A tokenizer \(T\) achieves parity for language \(A\) with respect to language \(B\) if \(\frac{|T(s_{A})|}{|T(s_{B})|}\approx 1\), where \(s_{A}\) and \(s_{B}\) denote the sets of all sentences in the corpora of languages \(A\) and \(B\), respectively, and the ratio \(\frac{|T(s_{A})|}{T(s_{B})|}\) is defined as premium. We use the FLORES-200 (Goyal et al., 2022) parallel corpus, consisting of the same sentences human-translated into 200 languages. We calculate the parity values for each tokenizer and the four non-English languages with respect to English (see Fig. 2 for an overview). [|assistant|]
๋‹ค์Œ ๋‚ด์šฉ C.2๋Š” MTbench(Zheng et al., 2024a)์˜ prompt template์— ๋”ฐ๋ผ ์ค‘๊ตญ์–ด๋กœ ๋ฒˆ์—ญ๋œ CHC-Bench์— ์‚ฌ์šฉ๋œ prompt template์„ ๋ณด์—ฌ์ค€๋‹ค.[|endofturn|]
๋‹ค์Œ ๋‚ด์šฉ C.2๋Š” MTbench(Zheng et al., 2024a)์˜ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์— ๋”ฐ๋ผ ์ค‘๊ตญ์–ด๋กœ ๋ฒˆ์—ญ๋œ CHC-Bench์— ์‚ฌ์šฉ๋œ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์„ ๋ณด์—ฌ์ค€๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The following content C.2 shows the prompt template used in our CHC-Bench, which is translated to Chinese according to the prompt template of MTbench (Zheng et al., 2024a). [|assistant|]
์ด์™€ ๋Œ€์กฐ์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” \(a\)์— ๋Œ€ํ•ด standard error 0.018์„ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค(ํ‘œ 2 ์ฐธ์กฐ). ์ด๋Š” 80% confidence interval ํญ์˜ \(2\cdot z_{0.9}\cdot 0.02\์•ฝ 0.05\)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ์˜ 80% confidence interval์€ Hoffmann ๋“ฑ์ด ๋ณด๊ณ ํ•œ ๊ฒƒ๋ณด๋‹ค 50๋ฐฐ ๋” ๋„“์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ด์™€ ๋Œ€์กฐ์ ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” \(a\)์— ๋Œ€ํ•ด ํ‘œ์ค€ ์˜ค์ฐจ 0.018์„ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค(ํ‘œ 2 ์ฐธ์กฐ). ์ด๋Š” 80% ์‹ ๋ขฐ ๊ตฌ๊ฐ„ ํญ์˜ \(2\cdot z_{0.9}\cdot 0.02\์•ฝ 0.05\)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ์˜ 80% ์‹ ๋ขฐ ๊ตฌ๊ฐ„์€ Hoffmann ๋“ฑ์ด ๋ณด๊ณ ํ•œ ๊ฒƒ๋ณด๋‹ค 50๋ฐฐ ๋” ๋„“์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]By contrast we estimate a standard error of 0.018 for \(a\) (See Table 2). This would correspond to a width of the 80% confidence interval of \(2\cdot z_{0.9}\cdot 0.02\approx 0.05\). Hence our 80% confidence interval is 50-fold wider than that reported by Hoffmann et al. [|assistant|]
N ๋ฐ S ๊ฒฐ์ •. ์‹ค์ œ๋กœ๋Š” ์ €ํฌ๋Š” ์ข…์ข… ๊ฐ€์žฅ ํฐ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์„ ์„ ํƒํ•˜๊ณ  ์ˆ˜๋ ดํ•  ๋•Œ๊นŒ์ง€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๋‹จ์ˆœํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ตœ์ ์˜ ๊ตฌ์„ฑ์—์„œ ํฌ๊ฒŒ ๋ฒ—์–ด๋‚˜ ์ƒ๋‹นํ•œ ๋ฆฌ์†Œ์Šค ๋‚ญ๋น„๋ฅผ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Scaling laws๋Š” ๊ณ ์ •๋œ ๊ณ„์‚ฐ ์˜ˆ์‚ฐ \(C\)8์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ์ตœ์ ์˜ ๋ชจ๋ธ ํฌ๊ธฐ \(N\)๊ณผ ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜ \(S\)๋ฅผ ์„ ํƒํ•˜๋Š” ์›์น™์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. 3.13์€ ์ด๋ฏธ Loss \(L\), Batch Size \(B\), ๋ชจ๋ธ ํฌ๊ธฐ \(N\), ํ›ˆ๋ จ ๋‹จ๊ณ„ \(S\) ์‚ฌ์ด์˜ ์ •ํ™•ํ•œ ๊ด€๊ณ„๋ฅผ ์ œ๊ณตํ•˜๋ฏ€๋กœ, ์ž„๊ณ„ ๋ฐฐ์น˜ ํฌ๊ธฐ(\(B=B_{crit}\))์—์„œ \(L\)์„ ์ตœ์†Œํ™”ํ•˜๋Š” ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ์ฐพ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ตœ์ ์˜ \(N\)์€ Eq 3.13์˜ ๋„ํ•จ์ˆ˜๋ฅผ \(N\)์— ๋Œ€ํ•ด ์ทจํ•˜๊ณ  \(0\)์œผ๋กœ ์„ค์ •ํ•˜์—ฌ ๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
N$ ๋ฐ $S$ ๊ฒฐ์ •์‹ค์ œ๋กœ๋Š” ์ €ํฌ๋Š” ์ข…์ข… ๊ฐ€์žฅ ํฐ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์„ ์„ ํƒํ•˜๊ณ  ์ˆ˜๋ ดํ•  ๋•Œ๊นŒ์ง€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๋‹จ์ˆœํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ตœ์ ์˜ ๊ตฌ์„ฑ์—์„œ ํฌ๊ฒŒ ๋ฒ—์–ด๋‚˜ ์ƒ๋‹นํ•œ ๋ฆฌ์†Œ์Šค ๋‚ญ๋น„๋ฅผ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์€ ๊ณ ์ •๋œ ๊ณ„์‚ฐ ์˜ˆ์‚ฐ \(C\)8์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ์ตœ์ ์˜ ๋ชจ๋ธ ํฌ๊ธฐ \(N\)๊ณผ ํ›ˆ๋ จ ๋‹จ๊ณ„ ์ˆ˜ \(S\)๋ฅผ ์„ ํƒํ•˜๋Š” ์›์น™์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. 3.13์€ ์ด๋ฏธ ์†์‹ค \(L\), ๋ฐฐ์น˜ ํฌ๊ธฐ \(B\), ๋ชจ๋ธ ํฌ๊ธฐ \(N\), ํ›ˆ๋ จ ๋‹จ๊ณ„ \(S\) ์‚ฌ์ด์˜ ์ •ํ™•ํ•œ ๊ด€๊ณ„๋ฅผ ์ œ๊ณตํ•˜๋ฏ€๋กœ, ์ž„๊ณ„ ๋ฐฐ์น˜ ํฌ๊ธฐ(\(B=B_{crit}\))์—์„œ \(L\)์„ ์ตœ์†Œํ™”ํ•˜๋Š” ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ์ฐพ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ตœ์ ์˜ \(N\)์€ Eq 3.13์˜ ๋„ํ•จ์ˆ˜๋ฅผ \(N\)์— ๋Œ€ํ•ด ์ทจํ•˜๊ณ  \(0\)์œผ๋กœ ์„ค์ •ํ•˜์—ฌ ๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Determining \(N\) and \(S\)In practice, we often opt for the largest affordable model size and train the model until convergence. Nevertheless, this simplistic approach can deviate significantly from the optimal configuration and result in substantial resource wastage. Scaling laws provide a principled approach to choosing the optimal model size \(N\) and number of training steps \(S\) given a fixed computational budget \(C\)8. Given that Eq 3.13 already provides the precise relation between the loss \(L\), batch size \(B\), model size \(N\) and training steps \(S\), we could find the model size that minimizes \(L\) under the critical batch size (\(B=B_{crit}\)). This optimal \(N\) can be obtained by taking the derivative of Eq 3.13 w.r.t. \(N\) and setting it to \(0\). [|assistant|]
์ค‘๋ณต์€ ๋‹ค์–‘ํ•œ ์–ดํœ˜ ํฌ๊ธฐ์— ๊ฑธ์ณ ๋Œ€๋žต ์ผ์ •ํ•˜๋ฉฐ, ๋‘ ๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์—์„œ๋งŒ ๊ตฌํ˜„๋œ ๋™์ผํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ž„์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ „์ฒด ์ค‘๋ณต์€ ๋‹ค์†Œ ๋‚ฎ์€ ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ tokenizers๋Š” ์„œ๋กœ ๋‹ค๋ฅธ tokenized sequences๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ๋ชจ๋ธ ํ•™์Šต ๋ฐ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทผ๋ณธ์ ์ธ ์ด์œ ๋ฅผ ์กฐ์‚ฌํ•œ ๊ฒฐ๊ณผ, ์ด๋Ÿฌํ•œ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์˜ ์„œ๋กœ ๋‹ค๋ฅธ configuration ๋ฐ ์ „์ฒ˜๋ฆฌ ์˜ต์…˜ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๊ตญ์–ด ๋ฌธ์„œ์—์„œ ๋” ํฐ ๋™์˜์–ด ์‚ฌ์ „์œผ๋กœ ์ธํ•ด ๋‹ค๊ตญ์–ด tokenizer์˜ ์ค‘๋ณต๋ฅ ์ด ์˜์–ด tokenizers๋ณด๋‹ค ๋‚ฎ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ค‘๋ณต์€ ๋‹ค์–‘ํ•œ ์–ดํœ˜ ํฌ๊ธฐ์— ๊ฑธ์ณ ๋Œ€๋žต ์ผ์ •ํ•˜๋ฉฐ, ๋‘ ๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์—์„œ๋งŒ ๊ตฌํ˜„๋œ ๋™์ผํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ž„์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ „์ฒด ์ค‘๋ณต์€ ๋‹ค์†Œ ๋‚ฎ์€ ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ํ† ํฐํ™” ๋„๊ตฌ๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ํ† ํฐํ™” ์‹œํ€€์Šค๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ๋ชจ๋ธ ํ•™์Šต ๋ฐ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทผ๋ณธ์ ์ธ ์ด์œ ๋ฅผ ์กฐ์‚ฌํ•œ ๊ฒฐ๊ณผ, ์ด๋Ÿฌํ•œ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ๊ตฌ์„ฑ ๋ฐ ์ „์ฒ˜๋ฆฌ ์˜ต์…˜ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๊ตญ์–ด ๋ฌธ์„œ์—์„œ ๋” ํฐ ๋™์˜์–ด ์‚ฌ์ „์œผ๋กœ ์ธํ•ด ๋‹ค๊ตญ์–ด ํ† ํฐํ™” ๋„๊ตฌ์˜ ์ค‘๋ณต๋ฅ ์ด ์˜์–ด ํ† ํฐํ™” ๋„๊ตฌ๋ณด๋‹ค ๋‚ฎ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The overlap is roughly constant across different vocabulary sizes, and the total overlap tends to be rather low, despite being the identical algorithm only implemented by two different libraries. Consequently, the tokenizers produce different tokenized sequences, possibly affecting model training and downstream performance. Investigating the underlying reasons, the low overlap might be attributed to different configuration and pre-processing options in these libraries. Due to the larger thesaurus in multilingual documents, the overlap for the multilingual tokenizer is lower than for the English tokenizers. [|assistant|]
PG19(Rae et al., 2019) ๋ฐ Arxiv-math(Wu et al., 2022) ๋ฒค์น˜๋งˆํฌ์—์„œ ์†Œํ˜• Infini-Transformer ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์„ค์ •์€ Memorizing Transformers(Wu et al., 2022)์˜ ์„ค์ •๊ณผ ๋งค์šฐ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๋ชจ๋“  ๋ชจ๋ธ์€ 12๊ฐœ์˜ ๋ ˆ์ด์–ด์™€ ๊ฐ๊ฐ 128์ฐจ์›์˜ 8๊ฐœ์˜ attention heads์™€ 4096๊ฐœ์˜ ์ˆจ๊ฒจ์ง„ ๋ ˆ์ด์–ด๊ฐ€ ์žˆ๋Š” FFN์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
PG19(Rae et al., 2019) ๋ฐ Arxiv-math(Wu et al., 2022) ๋ฒค์น˜๋งˆํฌ์—์„œ ์†Œํ˜• ์ธํ”ผ๋‹ˆ ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์„ค์ •์€ Memorizing Transformers(Wu et al., 2022)์˜ ์„ค์ •๊ณผ ๋งค์šฐ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๋ชจ๋“  ๋ชจ๋ธ์€ 12๊ฐœ์˜ ๋ ˆ์ด์–ด์™€ ๊ฐ๊ฐ 128์ฐจ์›์˜ 8๊ฐœ์˜ ์ฃผ์˜ ํ—ค๋“œ์™€ 4096๊ฐœ์˜ ์ˆจ๊ฒจ์ง„ ๋ ˆ์ด์–ด๊ฐ€ ์žˆ๋Š” FFN์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We trained and evaluated small Infini-Transformer models on PG19 (Rae et al., 2019) and Arxiv-math (Wu et al., 2022) benchmarks. Our setup closely resembles that of Memorizing Transformers (Wu et al., 2022). Namely, all our models have 12 layers and 8 attention heads of dimension 128 each and FFNs with hidden layer 4096. [|assistant|]
๋„์ถœ์€ ๋ถ€๋ก A.2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐฉ์ •์‹ 6์€ Bradley-Terry ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์ง€๋งŒ, ๋ถ€๋ก A.3์— ๋‚˜์™€ ์žˆ๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๋” ์ผ๋ฐ˜์ ์ธ Plackett-Luce ๋ชจ๋ธ[30; 21] ํ•˜์—์„œ๋„ ์œ ์‚ฌํ•˜๊ฒŒ ์‹์„ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋„์ถœ์€ ๋ถ€๋ก A.2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐฉ์ •์‹ 6์€ ๋ธŒ๋ž˜๋“ค๋ฆฌ-ํ…Œ๋ฆฌ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์ง€๋งŒ, ๋ถ€๋ก A.3์— ๋‚˜์™€ ์žˆ๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๋” ์ผ๋ฐ˜์ ์ธ ํ”Œ๋ผ์ผ“-๋ฃจ์Šค ๋ชจ๋ธ[30; 21] ํ•˜์—์„œ๋„ ์œ ์‚ฌํ•˜๊ฒŒ ์‹์„ ๋„์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The derivation is in Appendix A.2. While Eq. 6 uses the Bradley-Terry model, we can similarly derive expressions under the more general Plackett-Luce models [30; 21], shown in Appendix A.3. [|assistant|]
์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ธ๊ธฐ๊ฐ€ ์—†๋Š” ์ง€์‹์— ์ดˆ์ ์„ ๋งž์ถฐ RAG์™€ FT์˜ ํšจ๊ณผ๋ฅผ ๋น„๊ต ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ Fine-tuning์ด ๋ชจ๋“  ์—”ํ‹ฐํ‹ฐ์—์„œ ์ผ๊ด€๋œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์˜ค๋ฉฐ, ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ์นดํ…Œ๊ณ ๋ฆฌ์™€ ๊ฐ€์žฅ ์ธ๊ธฐ ์—†๋Š” ์นดํ…Œ๊ณ ๋ฆฌ์—์„œ ๊ฐ€์žฅ ํฐ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์˜จ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ Fine-tuning๊ณผ ํ•จ๊ป˜ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ RAG๊ฐ€ ๋” ํšจ๊ณผ์ ์ธ ์ „๋žต์œผ๋กœ ๋‹๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์žฅ์ ์€ ๋” ํฐ ๋ชจ๋ธ์—์„œ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Retrieval ๋ฐ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋ฉด ๋‘ ๊ฐ€์ง€ ์ „๋žต ๋ชจ๋‘์˜ ์„ฑ๊ณต๋ฅ ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์œผ๋กœ ๊ด€์ฐฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์˜ ์ค‘์š”์„ฑ์„ ์ดํ•ดํ•˜๊ณ  ํ–ฅํ›„ ์ž‘์—…์€ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์œ„ํ•œ ํšจ๊ณผ์ ์ธ ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘˜ ๊ฒƒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์ธ๊ธฐ๊ฐ€ ์—†๋Š” ์ง€์‹์— ์ดˆ์ ์„ ๋งž์ถฐ RAG์™€ FT์˜ ํšจ๊ณผ๋ฅผ ๋น„๊ต ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ FT๊ฐ€ ๋ชจ๋“  ์—”ํ‹ฐํ‹ฐ์—์„œ ์ผ๊ด€๋œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์˜ค๋ฉฐ, ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ์นดํ…Œ๊ณ ๋ฆฌ์™€ ๊ฐ€์žฅ ์ธ๊ธฐ ์—†๋Š” ์นดํ…Œ๊ณ ๋ฆฌ์—์„œ ๊ฐ€์žฅ ํฐ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์˜จ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ๋ฏธ์„ธ ์กฐ์ •๊ณผ ํ•จ๊ป˜ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ RAG๊ฐ€ ๋” ํšจ๊ณผ์ ์ธ ์ „๋žต์œผ๋กœ ๋‹๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์žฅ์ ์€ ๋” ํฐ ๋ชจ๋ธ์—์„œ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฒ€์ƒ‰ ๋ฐ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋ฉด ๋‘ ๊ฐ€์ง€ ์ „๋žต ๋ชจ๋‘์˜ ์„ฑ๊ณต๋ฅ ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์œผ๋กœ ๊ด€์ฐฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์˜ ์ค‘์š”์„ฑ์„ ์ดํ•ดํ•˜๊ณ  ํ–ฅํ›„ ์ž‘์—…์€ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์œ„ํ•œ ํšจ๊ณผ์ ์ธ ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘˜ ๊ฒƒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this study, we performed a comparative analysis to evaluate the effectiveness of RAG versus FT, with a focus on less popular knowledge. Our results reveal that FT leads to consistent performance improvements for all entities, with the most notable gains seen in the most and least popular categories. We found that RAG stands out as a more effective strategy, especially when used in combination with fine-tuning. This advantage decreases however in larger models. Additionally, we observed that the success of both RAG and FT strategies improves with the enhancement of the retrieval and data augmentation models' performance. Understanding the critical importance of synthetic data quality, future work will focus on developing an effective method for data creation. [|assistant|]
CQIA-Subset์—์„œ Fine-tuningํ•œ ํ›„ ๋‹ค์–‘ํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ ํฌ๊ธฐ๋ฅผ ๊ฐ€์ง„ ๋‹ค์–‘ํ•œ ๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, Yi-6B๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ ํฌ๊ธฐ๊ฐ€ ์ตœ์†Œ ๋‘ ๋ฐฐ ์ด์ƒ ํฐ \(\mathtt{Qwen}\)-14B ๋ฐ InternLM-20B๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Yi-34B๋Š” C-Eval ๋ฐ CMMLU ๋ฒค์น˜๋งˆํฌ ๋ชจ๋‘์—์„œ \(\mathtt{Qwen}\)-72B์™€ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ ํฌ๊ธฐ, ์•„ํ‚คํ…์ฒ˜ ์ตœ์ ํ™” ๋ฐ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก  ๊ฐ„์˜ ๊ท ํ˜•์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. Scaling law์— ๋”ฐ๋ฅด๋ฉด ๋” ํฐ ๋ชจ๋ธ์€ ์–ธ์–ด ์ดํ•ด ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋˜์–ด ๋ณธ์งˆ์ ์œผ๋กœ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋Š” ํ•ญ์ƒ ๊ทธ๋Ÿฐ ๊ฒƒ์€ ์•„๋‹ˆ๋ผ๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
CQIA-Subset์—์„œ ๋ฏธ์„ธ ์กฐ์ •ํ•œ ํ›„ ๋‹ค์–‘ํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ ํฌ๊ธฐ๋ฅผ ๊ฐ€์ง„ ๋‹ค์–‘ํ•œ ๊ธฐ๋ณธ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, Yi-6B๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ ํฌ๊ธฐ๊ฐ€ ์ตœ์†Œ ๋‘ ๋ฐฐ ์ด์ƒ ํฐ \(\mathtt{Qwen}\)-14B ๋ฐ InternLM-20B๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ Yi-34B๋Š” C-Eval ๋ฐ CMMLU ๋ฒค์น˜๋งˆํฌ ๋ชจ๋‘์—์„œ \(\mathtt{Qwen}\)-72B์™€ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ๋ชจ๋ธ ํฌ๊ธฐ, ์•„ํ‚คํ…์ฒ˜ ์ตœ์ ํ™” ๋ฐ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก  ๊ฐ„์˜ ๊ท ํ˜•์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋”ฐ๋ฅด๋ฉด ๋” ํฐ ๋ชจ๋ธ์€ ์–ธ์–ด ์ดํ•ด ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋˜์–ด ๋ณธ์งˆ์ ์œผ๋กœ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋Š” ํ•ญ์ƒ ๊ทธ๋Ÿฐ ๊ฒƒ์€ ์•„๋‹ˆ๋ผ๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We investigate the performance of different base model with varying parameter sizes after fine-tuned on our CQIA-Subset. Notably, Yi-6B surpasses \(\mathtt{Qwen}\)-14B and InternLM-20B, which have at least twice its parameter size. Further, Yi-34B achieved comparable results to \(\mathtt{Qwen}\)-72B in both C-Eval and CMMLU benchmarks. This observervation underscores the balance between model size, architectural optimizations, and training methodologies. While the scaling law might suggest that larger models inherently perform better due to their increased language understanding capacity, our results indicate that this is not always the case. [|assistant|]
์ฐธ์กฐ ๋ชจ๋ธ ํ›ˆ๋ จ์ˆ˜ํ•™์  ์ฐธ์กฐ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด 0.5B์˜ ๊ณ ํ’ˆ์งˆ ์ˆ˜ํ•™ ๊ด€๋ จ ํ† ํฐ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” GPT(Yu et al., 2024; Huang et al., 2024)์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ์ˆ˜๋™์œผ๋กœ ํ๋ ˆ์ด์…˜๋œ ๋ฐ์ดํ„ฐ(Yue et al., 2024; Ni et al., 2024)๊ฐ€ ํ˜ผํ•ฉ๋œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜ ์ฐธ์กฐ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, Tulu-v2(Ivison et al., 2023) ๋ฐ OpenHermes-2.5(Teknium, 2023)์™€ ๊ฐ™์€ ์˜คํ”ˆ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 1.9B ํ† ํฐ์œผ๋กœ ๊ตฌ์„ฑ๋œ ์ฝ”ํผ์Šค๋ฅผ ์ปดํŒŒ์ผํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฐธ์กฐ ๋ชจ๋ธ์€ 3ํšŒ์— ๊ฑธ์ณ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๋Œ€ Learning Rate๋Š” 1B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 5e-5, 7B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 1e-5๋กœ ์„ค์ •ํ•˜๊ณ  cosine decay schedule์„ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๋Œ€ ์‹œํ€€์Šค ๊ธธ์ด๋Š” 1B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 2048, 7B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 4096์œผ๋กœ ์„ค์ •ํ•˜๊ณ  ๋ชจ๋ธ ์ž…๋ ฅ์— ๋Œ€ํ•ด ์—ฌ๋Ÿฌ ์ƒ˜ํ”Œ์„ ์ด ๊ธธ์ด๋กœ ํŒจํ‚นํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ์ฃผ์š” ์‹คํ—˜์—์„œ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต ๋ชจ๋ธ๊ณผ ์ฐธ์กฐ ๋ชจ๋ธ์„ _๋™์ผํ•œ_ ๊ธฐ๋ณธ ๋ชจ๋ธ๋กœ ์ดˆ๊ธฐํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ฐธ์กฐ ๋ชจ๋ธ ํ›ˆ๋ จ์ˆ˜ํ•™์  ์ฐธ์กฐ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด 0.5B์˜ ๊ณ ํ’ˆ์งˆ ์ˆ˜ํ•™ ๊ด€๋ จ ํ† ํฐ์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” GPT(Yu et al., 2024; Huang et al., 2024)์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์™€ ์ˆ˜๋™์œผ๋กœ ํ๋ ˆ์ด์…˜๋œ ๋ฐ์ดํ„ฐ(Yue et al., 2024; Ni et al., 2024)๊ฐ€ ํ˜ผํ•ฉ๋œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜ ์ฐธ์กฐ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, Tulu-v2(Ivison et al., 2023) ๋ฐ OpenHermes-2.5(Teknium, 2023)์™€ ๊ฐ™์€ ์˜คํ”ˆ ์†Œ์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ 1.9B ํ† ํฐ์œผ๋กœ ๊ตฌ์„ฑ๋œ ์ฝ”ํผ์Šค๋ฅผ ์ปดํŒŒ์ผํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฐธ์กฐ ๋ชจ๋ธ์€ 3ํšŒ์— ๊ฑธ์ณ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๋Œ€ ํ•™์Šต ์†๋„๋Š” 1B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 5e-5, 7B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 1e-5๋กœ ์„ค์ •ํ•˜๊ณ  ์ฝ”์‚ฌ์ธ ๊ฐ์‡  ์Šค์ผ€์ค„์„ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๋Œ€ ์‹œํ€€์Šค ๊ธธ์ด๋Š” 1B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 2048, 7B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 4096์œผ๋กœ ์„ค์ •ํ•˜๊ณ  ๋ชจ๋ธ ์ž…๋ ฅ์— ๋Œ€ํ•ด ์—ฌ๋Ÿฌ ์ƒ˜ํ”Œ์„ ์ด ๊ธธ์ด๋กœ ํŒจํ‚นํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ์ฃผ์š” ์‹คํ—˜์—์„œ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต ๋ชจ๋ธ๊ณผ ์ฐธ์กฐ ๋ชจ๋ธ์„ _๋™์ผํ•œ_ ๊ธฐ๋ณธ ๋ชจ๋ธ๋กœ ์ดˆ๊ธฐํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Reference Model TrainingTo train our mathematical reference model, we gathered a dataset of 0.5B high-quality, math-related tokens. This dataset is a blend of synthetic data from GPT (Yu et al., 2024; Huang et al., 2024) and manually curated data (Yue et al., 2024; Ni et al., 2024). For the general reference model, we compiled a corpus of 1.9B tokens from open-source datasets, such as Tulu-v2 (Ivison et al., 2023) and OpenHermes-2.5 (Teknium, 2023). We trained the reference models for 3 epochs. The maximum learning rate was set at 5e-5 for 1B models and 1e-5 for 7B models, applying a cosine decay schedule. We set the maximum sequence lengths to 2048 for 1B models and 4096 for 7B models, packing multiple samples into these lengths for model input. In all main experiments, we initialized the continual pretraining model and the reference model with the _same_ base model. [|assistant|]
์†Œ์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํฌ๊ธฐ์— ์ž์—ฐ์ ์ธ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๋ฉด ์–ด๋–ป๊ฒŒ ๋˜๋‚˜์š”? Hernandez ๋“ฑ[19]์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์—์„œ ๋ฐ˜๋ณต๋˜๋Š” ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์˜ ๋ถˆ๊ท ํ˜•ํ•œ ๋ถ€์ž‘์šฉ์„ ๋ฐœ๊ฒฌํ•˜๊ณ  ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ์šฐ๋ฆฌ์˜ ์ž‘์—…๊ณผ ๋™์‹œ์— Muennighoff ๋“ฑ[37]์€ C4์˜ ๋ฌด์ž‘์œ„ ํ•˜์œ„ ์ง‘ํ•ฉ์„ 4ํšŒ ์ด์ƒ epochingํ•  ๋•Œ test loss๊ฐ€ ์•…ํ™”๋˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์ œํ•œ๋œ ๋ฐ์ดํ„ฐ, multi-epoch ์„ค์ •์—์„œ D4 ์‚ฌ์šฉ์ด ๋ชจ๋ธ ์„ฑ๋Šฅ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์†Œ์Šค ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํฌ๊ธฐ์— ์ž์—ฐ์ ์ธ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๋ฉด ์–ด๋–ป๊ฒŒ ๋˜๋‚˜์š”? Hernandez ๋“ฑ[19]์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์—์„œ ๋ฐ˜๋ณต๋˜๋Š” ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์˜ ๋ถˆ๊ท ํ˜•ํ•œ ๋ถ€์ž‘์šฉ์„ ๋ฐœ๊ฒฌํ•˜๊ณ  ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ์šฐ๋ฆฌ์˜ ์ž‘์—…๊ณผ ๋™์‹œ์— Muennighoff ๋“ฑ[37]์€ C4์˜ ๋ฌด์ž‘์œ„ ํ•˜์œ„ ์ง‘ํ•ฉ์„ 4ํšŒ ์ด์ƒ ์ดํฌ์นญํ•  ๋•Œ ํ…Œ์ŠคํŠธ ์†์‹ค์ด ์•…ํ™”๋˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์ œํ•œ๋œ ๋ฐ์ดํ„ฐ, ๋‹ค์ค‘ ์—ํฌํฌ ์„ค์ •์—์„œ D4 ์‚ฌ์šฉ์ด ๋ชจ๋ธ ์„ฑ๋Šฅ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]natural limit to the size of the source dataset. What happens when we run out of data? Hernandez et al. [19] found and analyzed disproportionately adverse effects of repeated data points in the training data. Similarly, concurrently to our work Muennighoff et al. [37] shows that test loss deteriorates when epoching over a random subset of C4 more than four times. In this section, we investigate how the use of D4 affects model performance in this limited data, multi-epoch setting. [|assistant|]
**Public API of LLMs**. API๋Š” ๋ชจ๋ธ ๋ณต์‚ฌ๋ณธ์„ ์ง์ ‘ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  ์ผ๋ฐ˜ ์‚ฌ์šฉ์ž๊ฐ€ ๋กœ์ปฌ์—์„œ ๋ชจ๋ธ์„ ์‹คํ–‰ํ•  ํ•„์š” ์—†์ด LLM์„ ๋” ํŽธ๋ฆฌํ•˜๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. LLM ์‚ฌ์šฉ์„ ์œ„ํ•œ ๋Œ€ํ‘œ์ ์ธ ์ธํ„ฐํŽ˜์ด์Šค์ธ GPT ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์šฉ API[46, 55, 66, 105]๋Š” ํ•™๊ณ„์™€ ์—…๊ณ„ ๋ชจ๋‘์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค19. OpenAI๋Š” GPT-3 ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์— ada, babbage, curie, davinci(GPT-3 ์‹œ๋ฆฌ์ฆˆ์—์„œ ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ๋ฒ„์ „), text-ada-001, text-babbbage-001, text-curie-001 ๋“ฑ 7๊ฐ€์ง€ ์ฃผ์š” ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ค‘ ์ฒ˜์Œ ๋„ค ๊ฐ€์ง€ ์ธํ„ฐํŽ˜์ด์Šค๋Š” OpenAI์˜ ํ˜ธ์ŠคํŠธ ์„œ๋ฒ„์—์„œ ์ถ”๊ฐ€๋กœ Fine-tuningํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, babbage, curie, davinci๋Š” ๊ฐ๊ฐ GPT-3(1B), GPT-3(6.7B), GPT-3(175B) ๋ชจ๋ธ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค[55]. ๋˜ํ•œ Codex[105]์™€ ๊ด€๋ จ๋œ ๋‘ ๊ฐ€์ง€ API, ์ฆ‰ code-cushman-001(Codex(12B[105])์˜ ๊ฐ•๋ ฅํ•˜๊ณ  ๋‹ค๊ตญ์–ด ๋ฒ„์ „)๊ณผ code-davinci-002๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ GPT-3.5 ์‹œ๋ฆฌ์ฆˆ์—๋Š” ๊ธฐ๋ณธ ๋ชจ๋ธ์ธ code-davinci-002์™€ ์„ธ ๊ฐ€์ง€ ํ–ฅ์ƒ๋œ ๋ฒ„์ „์ธ text-davinci-002, text-davinci-003, gpt-3.5-turbo๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋” ๊ฐ•๋ ฅํ•œ ๋Œ€์•ˆ์œผ๋กœ ์˜ฌํ•ด OpenAI๋Š” gpt-4, gpt-4-32k, gpt-4-1106-preview(_์ฆ‰, GPT-4 Turbo) ๋ฐ gpt-4-vision-preview(_์ฆ‰, ์‹œ๊ฐ ๊ธฐ๋Šฅ์ด ์žˆ๋Š” GPT-4 Turbo, ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ชจ๋ธ) ๋“ฑ GPT-4 ์‹œ๋ฆฌ์ฆˆ์˜ ๋ชจ๋ธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ๋ชฉํ•  ์ ์€ OpenAI๊ฐ€ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ ์ธํ„ฐํŽ˜์ด์Šค(gpt-3.5-turbo, gpt-4, gpt-4-32k)๋ฅผ ์œ ์ง€ ๊ด€๋ฆฌํ•˜๊ณ  ์—…๊ทธ๋ ˆ์ด๋“œํ•˜๊ณ  ์žˆ์œผ๋ฏ€๋กœ API ์ด๋ฆ„์€ ์‹ค์ œ๋กœ ์ตœ์‹  ๋ฒ„์ „์„ ๊ฐ€๋ฆฌํ‚ฌ ๊ฒƒ์ด๋ผ๋Š” ์ ์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ ChatGPT๋Š” GPT-3.5 ๋˜๋Š” GPT-4 ๋ชจ๋ธ๋กœ ๊ตฌ๋™ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ ํŠน์ • ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์‹œ๋‚˜๋ฆฌ์˜ค์™€ ์‘๋‹ต ์š”๊ตฌ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ์ ํ•ฉํ•œ ๋ชจ๋ธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์„ ํƒํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ์‚ฌ์šฉ๋ฒ•์€ ํ”„๋กœ์ ํŠธ ์›น์‚ฌ์ดํŠธ20์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
**LLM์˜ ๊ณต์šฉ API**. API๋Š” ๋ชจ๋ธ ๋ณต์‚ฌ๋ณธ์„ ์ง์ ‘ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  ์ผ๋ฐ˜ ์‚ฌ์šฉ์ž๊ฐ€ ๋กœ์ปฌ์—์„œ ๋ชจ๋ธ์„ ์‹คํ–‰ํ•  ํ•„์š” ์—†์ด LLM์„ ๋” ํŽธ๋ฆฌํ•˜๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. LLM ์‚ฌ์šฉ์„ ์œ„ํ•œ ๋Œ€ํ‘œ์ ์ธ ์ธํ„ฐํŽ˜์ด์Šค์ธ GPT ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์šฉ API[46, 55, 66, 105]๋Š” ํ•™๊ณ„์™€ ์—…๊ณ„ ๋ชจ๋‘์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค19. OpenAI๋Š” GPT-3 ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์— ์•„๋‹ค, ๋ฐ”๋น„์ง€, ํ๋ฆฌ, ๋‹ค๋นˆ์น˜(GPT-3 ์‹œ๋ฆฌ์ฆˆ์—์„œ ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ๋ฒ„์ „), ํ…์ŠคํŠธ-์•„๋‹ค-001, ํ…์ŠคํŠธ-๋ฐ”๋น„์ง€-001, ํ…์ŠคํŠธ-ํ๋ฆฌ-001 ๋“ฑ 7๊ฐ€์ง€ ์ฃผ์š” ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ค‘ ์ฒ˜์Œ ๋„ค ๊ฐ€์ง€ ์ธํ„ฐํŽ˜์ด์Šค๋Š” OpenAI์˜ ํ˜ธ์ŠคํŠธ ์„œ๋ฒ„์—์„œ ์ถ”๊ฐ€๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, babbage, curie, davinci๋Š” ๊ฐ๊ฐ GPT-3(1B), GPT-3(6.7B), GPT-3(175B) ๋ชจ๋ธ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค[55]. ๋˜ํ•œ ์ฝ”๋ฑ์Šค[105]์™€ ๊ด€๋ จ๋œ ๋‘ ๊ฐ€์ง€ API, ์ฆ‰ code-cushman-001(์ฝ”๋ฑ์Šค(12B[105])์˜ ๊ฐ•๋ ฅํ•˜๊ณ  ๋‹ค๊ตญ์–ด ๋ฒ„์ „)๊ณผ code-davinci-002๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ GPT-3.5 ์‹œ๋ฆฌ์ฆˆ์—๋Š” ๊ธฐ๋ณธ ๋ชจ๋ธ์ธ code-davinci-002์™€ ์„ธ ๊ฐ€์ง€ ํ–ฅ์ƒ๋œ ๋ฒ„์ „์ธ text-davinci-002, text-davinci-003, gpt-3.5-turbo๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋” ๊ฐ•๋ ฅํ•œ ๋Œ€์•ˆ์œผ๋กœ ์˜ฌํ•ด OpenAI๋Š” gpt-4, gpt-4-32k, gpt-4-1106-preview(_์ฆ‰, GPT-4 ํ„ฐ๋ณด) ๋ฐ gpt-4-vision-preview(_์ฆ‰, ์‹œ๊ฐ ๊ธฐ๋Šฅ์ด ์žˆ๋Š” GPT-4 ํ„ฐ๋ณด, ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ชจ๋ธ) ๋“ฑ GPT-4 ์‹œ๋ฆฌ์ฆˆ์˜ ๋ชจ๋ธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ๋ชฉํ•  ์ ์€ OpenAI๊ฐ€ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ ์ธํ„ฐํŽ˜์ด์Šค(gpt-3.5-turbo, gpt-4, gpt-4-32k)๋ฅผ ์œ ์ง€ ๊ด€๋ฆฌํ•˜๊ณ  ์—…๊ทธ๋ ˆ์ด๋“œํ•˜๊ณ  ์žˆ์œผ๋ฏ€๋กœ API ์ด๋ฆ„์€ ์‹ค์ œ๋กœ ์ตœ์‹  ๋ฒ„์ „์„ ๊ฐ€๋ฆฌํ‚ฌ ๊ฒƒ์ด๋ผ๋Š” ์ ์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ ChatGPT๋Š” GPT-3.5 ๋˜๋Š” GPT-4 ๋ชจ๋ธ๋กœ ๊ตฌ๋™ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ ํŠน์ • ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์‹œ๋‚˜๋ฆฌ์˜ค์™€ ์‘๋‹ต ์š”๊ตฌ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ์ ํ•ฉํ•œ ๋ชจ๋ธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์„ ํƒํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ์‚ฌ์šฉ๋ฒ•์€ ํ”„๋กœ์ ํŠธ ์›น์‚ฌ์ดํŠธ20์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Public API of LLMs**. Instead of directly using the model copies, APIs provide a more convenient way for common users to use LLMs, without the need of running the model locally. As a representative interface for using LLMs, the APIs for the GPT-series models [46, 55, 66, 105] have been widely used for both academia and industry19. OpenAI has provided seven major interfaces to the models in GPT-3 series: ada, babbage, curie, davinci (the most powerful version in GPT-3 series), text-ada-001, text-babbbage-001, and text-curie-001. Among them, the first four interfaces can be further fine-tuned on the host server of OpenAI. In particular, babbage, curie, and davinci correspond to the GPT-3 (1B), GPT-3 (6.7B), and GPT-3 (175B) models, respectively [55]. In addition, there are also two APIs related to Codex [105], called code-cushman-001 (a powerful and multilingual version of the Codex (12B [105]) and code-davinci-002. Further, GPT-3.5 series include one base model code-davinci-002 and three enhanced versions, namely text-davinci-002, text-davinci-003, and gpt-3.5-turbo. As more powerful alternatives, in this year, OpenAI has released the model interfaces for GPT-4 series, including gpt-4, gpt-4-32k, gpt-4-1106-preview (_i.e._, GPT-4 Turbo) and gpt-4-vision-preview (_i.e._, GPT-4 Turbo with vision, a multimodal model). It is worth noting that OpenAI has been maintaining and upgrading these model interfaces (gpt-3.5-turbo, gpt-4, gpt-4-32k), so the API name will actually point to the latest version. Currently, ChatGPT can be powered by either GPT-3.5 or GPT-4 models. Overall, one select the suitable model interface based on the specific application scenarios and response requirements. The detailed usage can be found on their project websites20. [|assistant|]
์œ„์—์„œ ์„ค๋ช…ํ•œ ์ „๋žต์€ ์ „์ฒด ์„ ํƒ ๋น„์œจ์ด \(R=R_{dedup}*R_{proto}\)์ด๋ฉฐ, ๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ๋ฅผ ๋กœ์ปฌ ๋ฐ ๊ธ€๋กœ๋ฒŒ ์ˆ˜์ค€์—์„œ ๋‹ค์–‘ํ™”ํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ _Document De-Duplication and Diversification_์˜ ์•ฝ์ž์ธ **D4**๋ผ๊ณ  ๋ถ€๋ฆ…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์—์„œ๋Š” \(R_{dedup}=0.75\)๋ฅผ ์„ ํƒํ•˜๊ณ  \(R_{proto}\)๋ฅผ ๋ณ€๊ฒฝํ•ฉ๋‹ˆ๋‹ค(์ด ์„ ํƒ์— ๋Œ€ํ•ด์„œ๋Š” ์„น์…˜ A.1์—์„œ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค). ์„น์…˜ 4์—์„œ๋Š” D4์˜ ์„ฑ๋Šฅ์„ baseline ํ›ˆ๋ จ ๋ฐ ๊ธฐํƒ€ ๋ฐฉ๋ฒ•๊ณผ ๋น„๊ตํ•˜๊ณ , ์„น์…˜ 4.4์—์„œ๋Š” D4๋ฅผ ๋ถ„์„ํ•˜์—ฌ ์˜๋ฏธ๋ก ์  ์ค‘๋ณต ์ œ๊ฑฐ ํ›„ ์žฌํด๋Ÿฌ์Šคํ„ฐ๋ง์ด ์ค‘๋ณต์œผ๋กœ ์ธํ•œ ํด๋Ÿฌ์Šคํ„ฐ์˜ ์˜ํ–ฅ์„ ์‹ค์ œ๋กœ ๊ฐ์†Œ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(๊ทธ๋ฆผ 7 ์ฐธ์กฐ).[|endofturn|]
์œ„์—์„œ ์„ค๋ช…ํ•œ ์ „๋žต์€ ์ „์ฒด ์„ ํƒ ๋น„์œจ์ด \(R=R_{dedup}*R_{proto}\)์ด๋ฉฐ, ๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ๋ฅผ ๋กœ์ปฌ ๋ฐ ๊ธ€๋กœ๋ฒŒ ์ˆ˜์ค€์—์„œ ๋‹ค์–‘ํ™”ํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์„ _๋ฌธ์„œ ์ค‘๋ณต ์ œ๊ฑฐ ๋ฐ ๋‹ค์–‘ํ™”_์˜ ์•ฝ์ž์ธ **D4**๋ผ๊ณ  ๋ถ€๋ฆ…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์—์„œ๋Š” \(R_{dedup}=0.75\)๋ฅผ ์„ ํƒํ•˜๊ณ  \(R_{proto}\)๋ฅผ ๋ณ€๊ฒฝํ•ฉ๋‹ˆ๋‹ค(์ด ์„ ํƒ์— ๋Œ€ํ•ด์„œ๋Š” ์„น์…˜ A.1์—์„œ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค). ์„น์…˜ 4์—์„œ๋Š” D4์˜ ์„ฑ๋Šฅ์„ ๊ธฐ๋ณธ ํ›ˆ๋ จ ๋ฐ ๊ธฐํƒ€ ๋ฐฉ๋ฒ•๊ณผ ๋น„๊ตํ•˜๊ณ , ์„น์…˜ 4.4์—์„œ๋Š” D4๋ฅผ ๋ถ„์„ํ•˜์—ฌ ์˜๋ฏธ๋ก ์  ์ค‘๋ณต ์ œ๊ฑฐ ํ›„ ์žฌํด๋Ÿฌ์Šคํ„ฐ๋ง์ด ์ค‘๋ณต์œผ๋กœ ์ธํ•œ ํด๋Ÿฌ์Šคํ„ฐ์˜ ์˜ํ–ฅ์„ ์‹ค์ œ๋กœ ๊ฐ์†Œ์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(๊ทธ๋ฆผ 7 ์ฐธ์กฐ).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The above-described strategy has an overall selection ratio of \(R=R_{dedup}*R_{proto}\) and intends to diversify the distribution of our data locally and globally. For brevity we refer to this method as **D4**, a shorthand for _Document De-Duplication and Diversification_. Throughout this work, we choose \(R_{dedup}=0.75\) and vary \(R_{proto}\) (we discuss this choice in Section A.1). In Section 4, we compare the performance of D4 to baseline training and other methods, and in Section 4.4 we analyze D4 and show that reclustering after semantic de-duplication indeed reduces the impact of duplicate-driven clusters (see Figure 7). [|assistant|]
๋ฆฌํ„ฐ๋Ÿด ๋ถ„์„๊ณผ ์ˆ˜์น˜ ๊ฒ€์ฆ์„ ํ†ตํ•ด state-space models๊ฐ€ ๊ธฐํ•˜๊ธ‰์ˆ˜์ ์œผ๋กœ ๊ฐ์†Œํ•˜๋Š” ๋ฉ”๋ชจ๋ฆฌ ๋ฌธ์ œ๋ฅผ ๊ทผ๋ณธ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜์ง€ ๋ชปํ•œ๋‹ค๋Š” ๊ฒฐ๋ก ์„ ๋‚ด๋ ธ์Šต๋‹ˆ๋‹ค. Samsami ์™ธ[169]๋Š” ์žฅ๊ธฐ ๋ฉ”๋ชจ๋ฆฌ์™€ ์žฅ๊ธฐ ํฌ๋ ˆ๋”ง์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด R2I๋ผ๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•˜์—ฌ ์ผ๋ จ์˜ state-space models๋ฅผ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ Reinforcement Learning (MBRL) ์—์ด์ „ํŠธ์˜ ์›”๋“œ ๋ชจ๋ธ์— ํ†ตํ•ฉํ•จ์œผ๋กœ์จ ๊ธฐ์กด MBRL ์—์ด์ „ํŠธ๊ฐ€ ํ–‰๋™๊ณผ ๊ฒฐ๊ณผ ์‚ฌ์ด์˜ ์žฅ๊ธฐ ๊ฐ„๊ฒฉ์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์—†์—ˆ๋˜ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, R2I๋Š” ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ํฌ๋ ˆ๋”ง ํ• ๋‹น RL ์ž‘์—…์—์„œ ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋™์‹œ์— ๋” ๋น ๋ฅธ ์ˆ˜๋ ด ์†๋„๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. Black-Mamba[167]๋Š” mixture-of-experts (MoE)์™€ SSM์„ ๋ณ‘ํ•ฉํ•˜์—ฌ ์ถ”๋ก  ๋น„์šฉ์„ ํฌ๊ฒŒ ์ค„์—ฌ ํšจ์œจ์ ์ด๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…์˜ ํ† ๋Œ€๋ฅผ ๋งˆ๋ จํ–ˆ์Šต๋‹ˆ๋‹ค. MambaByte[166]๋Š” ์›์‹œ ๋ฐ”์ดํŠธ์—์„œ ์ง์ ‘ ํ•™์Šตํ•˜๋„๋ก ์„ค๊ณ„๋œ ํ† ํฐ ํ”„๋ฆฌ ์…€๋ ‰ํ‹ฐ๋ธŒ State Space Model๋กœ, ์ˆœํ™˜์  ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ๋น ๋ฅธ ํ…์ŠคํŠธ ์ƒ์„ฑ์ด ๊ฐ€๋Šฅํ•˜๋ฉฐ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹ค์šฉ์„ฑ์„ ๊ฐ•์กฐํ•˜๊ณ  ํ† ํฐ ํ”„๋ฆฌ ์–ธ์–ด ๋ชจ๋ธ๋ง์˜ ํ–ฅํ›„ ๋ฐœ์ „์„ ์œ„ํ•œ ํ† ๋Œ€๋ฅผ ๋งˆ๋ จํ–ˆ์Šต๋‹ˆ๋‹ค. Self pretraining (SPT) [192]์€ ์ฒ˜์Œ๋ถ€ํ„ฐ ๊ธด ์‹œํ€€์Šค ๋ชจ๋ธ์„ ๋น„๊ตํ•˜๋Š” ๊ธฐ์กด ์ ‘๊ทผ ๋ฐฉ์‹์— ๋„์ „ํ•˜์—ฌ ๋ฐ์ดํ„ฐ ๊ธฐ๋ฐ˜ ์„ ํ—˜์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋ฉด ์„ฑ๋Šฅ ํ‰๊ฐ€๊ฐ€ ํฌ๊ฒŒ ๋‹ฌ๋ผ์ง„๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐํ˜€๋ƒˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์‚ฌ์ „ ํ•™์Šต์„ ํ™œ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์•„ํ‚คํ…์ฒ˜์—์„œ ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ๋‹ฌ์„ฑํ•˜์—ฌ Transformers์™€ SSM ๊ฐ„์˜ ์„ฑ๋Šฅ ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜๊ณ  ์‹ฌ์ง€์–ด ์ž‘์—…์—์„œ ์ด์ „ SSM ๊ฒฐ๊ณผ๋ฅผ ๋Šฅ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. Laughing Hyena Distillery[161]๋Š” ํ’ˆ์งˆ ์†์‹ค ์—†์ด ์‚ฌ์ „ ํ•™์Šต๋œ ๊ธด ์ปจ๋ณผ๋ฃจ์…˜ ์‹œํ€€์Šค ๋ชจ๋ธ์—์„œ ์ปดํŒฉํŠธํ•œ state-space models๋ฅผ ์ถ”์ถœํ•˜๊ณ , ํ•ฉ๋ฆฌ์  ํ•จ์ˆ˜ ๊ทผ์‚ฌ ๋ฐ ๋ชจ๋ธ ๋‹ค์šด์Šค์ผ€์ผ๋ง ๊ธฐ์ˆ ์„ ํ™œ์šฉํ•˜์—ฌ ์ปจ๋ณผ๋ฃจ์…˜ ๋ ˆ์ด์–ด์—์„œ ์ €์ฐจ์› state-space models๋ฅผ ์ถ”์ถœํ•˜๋ฉฐ, ์ผ์ •ํ•œ ๋ฉ”๋ชจ๋ฆฌ์™€ ์ผ์ •ํ•œ ์‹œ๊ฐ„ ๋ณต์žก๋„๋กœ ์ž๋™ํ™”๋œ ํšŒ๊ท€ ์ƒ์„ฑ์„ ๋‹ฌ์„ฑํ•˜๋Š” ์ƒˆ๋กœ์šด ์ฆ๋ฅ˜ ๋ฐฉ๋ฒ•์ธ Laughing Hyena๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Improved Hyena๋Š” ์‚ฌ์ „ ํ•™์Šต ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ณ  ์ฑ„๋„ ๊ฐ„ ํ—ค๋“œ์— ํ•„ํ„ฐ ๊ฐ€์ค‘์น˜๋ฅผ ๋ฐ”์ธ๋”ฉํ•˜์—ฌ ์ฆ๋ฅ˜ํ•  ํ•„ํ„ฐ ์ˆ˜๋ฅผ ์ค„์ž…๋‹ˆ๋‹ค.[|endofturn|]
๋ฆฌํ„ฐ๋Ÿด ๋ถ„์„๊ณผ ์ˆ˜์น˜ ๊ฒ€์ฆ์„ ํ†ตํ•ด ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์ด ๊ธฐํ•˜๊ธ‰์ˆ˜์ ์œผ๋กœ ๊ฐ์†Œํ•˜๋Š” ๋ฉ”๋ชจ๋ฆฌ ๋ฌธ์ œ๋ฅผ ๊ทผ๋ณธ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜์ง€ ๋ชปํ•œ๋‹ค๋Š” ๊ฒฐ๋ก ์„ ๋‚ด๋ ธ์Šต๋‹ˆ๋‹ค. Samsami ์™ธ[169]๋Š” ์žฅ๊ธฐ ๋ฉ”๋ชจ๋ฆฌ์™€ ์žฅ๊ธฐ ํฌ๋ ˆ๋”ง์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด R2I๋ผ๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•˜์—ฌ ์ผ๋ จ์˜ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์„ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ๊ฐ•ํ™” ํ•™์Šต(MBRL) ์—์ด์ „ํŠธ์˜ ์›”๋“œ ๋ชจ๋ธ์— ํ†ตํ•ฉํ•จ์œผ๋กœ์จ ๊ธฐ์กด MBRL ์—์ด์ „ํŠธ๊ฐ€ ํ–‰๋™๊ณผ ๊ฒฐ๊ณผ ์‚ฌ์ด์˜ ์žฅ๊ธฐ ๊ฐ„๊ฒฉ์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์—†์—ˆ๋˜ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, R2I๋Š” ๋ฉ”๋ชจ๋ฆฌ ๋ฐ ํฌ๋ ˆ๋”ง ํ• ๋‹น RL ์ž‘์—…์—์„œ ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋™์‹œ์— ๋” ๋น ๋ฅธ ์ˆ˜๋ ด ์†๋„๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋ธ”๋ž™๋ง˜๋ฐ”[167]๋Š” ํ˜ผํ•ฉ์ „๋ฌธ๊ฐ€(MoE)์™€ SSM์„ ๋ณ‘ํ•ฉํ•˜์—ฌ ์ถ”๋ก  ๋น„์šฉ์„ ํฌ๊ฒŒ ์ค„์—ฌ ํšจ์œจ์ ์ด๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…์˜ ํ† ๋Œ€๋ฅผ ๋งˆ๋ จํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ง˜๋ฐ”๋ฐ”์ดํŠธ[166]๋Š” ์›์‹œ ๋ฐ”์ดํŠธ์—์„œ ์ง์ ‘ ํ•™์Šตํ•˜๋„๋ก ์„ค๊ณ„๋œ ํ† ํฐ ํ”„๋ฆฌ ์…€๋ ‰ํ‹ฐ๋ธŒ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ๋กœ, ์ˆœํ™˜์  ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ๋น ๋ฅธ ํ…์ŠคํŠธ ์ƒ์„ฑ์ด ๊ฐ€๋Šฅํ•˜๋ฉฐ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹ค์šฉ์„ฑ์„ ๊ฐ•์กฐํ•˜๊ณ  ํ† ํฐ ํ”„๋ฆฌ ์–ธ์–ด ๋ชจ๋ธ๋ง์˜ ํ–ฅํ›„ ๋ฐœ์ „์„ ์œ„ํ•œ ํ† ๋Œ€๋ฅผ ๋งˆ๋ จํ–ˆ์Šต๋‹ˆ๋‹ค. ์…€ํ”„ ํ”„๋ฆฌํŠธ๋ ˆ์ด๋‹(SPT) [192]์€ ์ฒ˜์Œ๋ถ€ํ„ฐ ๊ธด ์‹œํ€€์Šค ๋ชจ๋ธ์„ ๋น„๊ตํ•˜๋Š” ๊ธฐ์กด ์ ‘๊ทผ ๋ฐฉ์‹์— ๋„์ „ํ•˜์—ฌ ๋ฐ์ดํ„ฐ ๊ธฐ๋ฐ˜ ์„ ํ—˜์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋ฉด ์„ฑ๋Šฅ ํ‰๊ฐ€๊ฐ€ ํฌ๊ฒŒ ๋‹ฌ๋ผ์ง„๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐํ˜€๋ƒˆ์Šต๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์‚ฌ์ „ ํ•™์Šต์„ ํ™œ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์•„ํ‚คํ…์ฒ˜์—์„œ ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ๋‹ฌ์„ฑํ•˜์—ฌ ํŠธ๋žœ์Šคํฌ๋จธ์™€ SSM ๊ฐ„์˜ ์„ฑ๋Šฅ ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜๊ณ  ์‹ฌ์ง€์–ด ์ž‘์—…์—์„œ ์ด์ „ SSM ๊ฒฐ๊ณผ๋ฅผ ๋Šฅ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์›ƒ๋Š” ํ•˜์ด์—๋‚˜ ์ฆ๋ฅ˜์†Œ[161]๋Š” ํ’ˆ์งˆ ์†์‹ค ์—†์ด ์‚ฌ์ „ ํ•™์Šต๋œ ๊ธด ์ปจ๋ณผ๋ฃจ์…˜ ์‹œํ€€์Šค ๋ชจ๋ธ์—์„œ ์ปดํŒฉํŠธํ•œ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์„ ์ถ”์ถœํ•˜๊ณ , ํ•ฉ๋ฆฌ์  ํ•จ์ˆ˜ ๊ทผ์‚ฌ ๋ฐ ๋ชจ๋ธ ๋‹ค์šด์Šค์ผ€์ผ๋ง ๊ธฐ์ˆ ์„ ํ™œ์šฉํ•˜์—ฌ ์ปจ๋ณผ๋ฃจ์…˜ ๋ ˆ์ด์–ด์—์„œ ์ €์ฐจ์› ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์„ ์ถ”์ถœํ•˜๋ฉฐ, ์ผ์ •ํ•œ ๋ฉ”๋ชจ๋ฆฌ์™€ ์ผ์ •ํ•œ ์‹œ๊ฐ„ ๋ณต์žก๋„๋กœ ์ž๋™ํ™”๋œ ํšŒ๊ท€ ์ƒ์„ฑ์„ ๋‹ฌ์„ฑํ•˜๋Š” ์ƒˆ๋กœ์šด ์ฆ๋ฅ˜ ๋ฐฉ๋ฒ•์ธ ์›ƒ๋Š” ํ•˜์ด์—๋‚˜๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ฐœ์„ ๋œ ํ•˜์ด์—๋‚˜๋Š” ์‚ฌ์ „ ํ•™์Šต ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ณ  ์ฑ„๋„ ๊ฐ„ ํ—ค๋“œ์— ํ•„ํ„ฐ ๊ฐ€์ค‘์น˜๋ฅผ ๋ฐ”์ธ๋”ฉํ•˜์—ฌ ์ฆ๋ฅ˜ํ•  ํ•„ํ„ฐ ์ˆ˜๋ฅผ ์ค„์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]retical analysis and numerical verification, it is concluded that state-space models do not fundamentally address the issue of exponential decaying memory. Samsami et al. [169] propose a method named R2I to enhance long-term memory and long-term credit, which integrates a set of state-space models into the world model of model-based reinforcement learning (MBRL) agents, thereby solving the issue that existing MBRL agents were unable to deal with the long-term intervals between actions and outcomes. Experimental results demonstrated that R2I achieved state-of-the-art performance in memory and credit assignment RL tasks, while also exhibiting faster convergence speed. Black-Mamba [167] merging SSMs with mixture-of-experts (MoE) significantly reduces inference costs, paving the way for efficient and scalable text generation tasks. MambaByte [166] is a token-free selective State Space Model designed for learning directly from raw bytes and its recurrent nature enables fast text generation, highlighting its practicality for large-scale models and paving the way for future developments in token-free language modeling. Self pretraining (SPT) [192] challenges the conventional approach of comparing long-sequence models from scratch, revealing that pre-training with data-driven priors significantly alters performance evaluations. By leveraging pre-training, the research achieves substantial improvements across various architectures, closing the performance gap between Transformers and SSMs and even surpassing previous SSM results on tasks. Laughing Hyena Distillery [161] proposes Laughing Hyena, a new distillation method that extracts compact state-space models from pre-trained long convolutional sequence models without loss of quality, utilizes rational function approximation and model downscaling techniques to extract low-dimensional state-space models in the convolutional layers, and achieves automated regression generation with constant memory and constant time complexity. Improved Hyena improves pre-training quality and reduces the number of filters to be distilled by binding filter weights to heads across channels. [|assistant|]
์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์–ธ์–ด ๋ชจ๋ธ์ด ํ•™์Šตํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์œ ์ตํ•˜๋‹ค๊ณ  ๊ฐ„์ฃผ๋˜๋Š” ํ† ํฐ์—๋งŒ Loss๊ฐ€ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ํ† ํฐ ์„ ํƒ์€ ์ดˆ๊ณผ Loss์— ๋”ฐ๋ผ ๋ฐฐ์น˜์˜ ํ† ํฐ์„ ์ˆœ์œ„ํ™”ํ•˜๊ณ  ์ƒ์œ„ \(k\%\)์˜ ํ† ํฐ๋งŒ ํ•™์Šต์— ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์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.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]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. [|assistant|]
๊ทธ๋ฆผ 3์—์„œ ์•Œ ์ˆ˜ ์žˆ๋“ฏ์ด, ์ž‘์€ ๋ชจ๋ธ์„ ์ œ์™ธํ•˜๊ณ  LLaMA, Mistral, GPT2\({}_{0}\), GPT2\({}_{1/4}\) ์•„ํ‚คํ…์ฒ˜๋Š” 1000ํšŒ ๋…ธ์ถœ ์ด์ƒ์—์„œ GPT2์˜ scaling law๋ฅผ ๋ฉด๋ฐ€ํžˆ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ \(\leq 10M\)์ธ ์ž‘์€ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๋ฅผ ๋ฌถ์œผ๋ฉด ์šฉ๋Ÿ‰์ด ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 3(c) ์ฐธ์กฐ). ์ด๋Š” _2bit/param ์šฉ๋Ÿ‰ ๋น„์œจ์ด ๋Œ€๋ถ€๋ถ„์˜ ์ผ๋ฐ˜์ ์ธ (decoder-only) ์–ธ์–ด ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์—์„œ ๋น„๊ต์  ๋ณดํŽธ์ ์ธ ๋ฒ•์น™์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 3์—์„œ ์•Œ ์ˆ˜ ์žˆ๋“ฏ์ด, ์ž‘์€ ๋ชจ๋ธ์„ ์ œ์™ธํ•˜๊ณ  LLaMA, Mistral, GPT2\({}_{0}\), GPT2\({}_{1/4}\) ์•„ํ‚คํ…์ฒ˜๋Š” 1000ํšŒ ๋…ธ์ถœ ์ด์ƒ์—์„œ GPT2์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ๋ฉด๋ฐ€ํžˆ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ \(\leq 10M\)์ธ ์ž‘์€ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜๋ฅผ ๋ฌถ์œผ๋ฉด ์šฉ๋Ÿ‰์ด ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 3(c) ์ฐธ์กฐ). ์ด๋Š” _2๋น„ํŠธ/ํŒŒ๋ผ๋ฏธํ„ฐ ์šฉ๋Ÿ‰ ๋น„์œจ์ด ๋Œ€๋ถ€๋ถ„์˜ ์ผ๋ฐ˜์ ์ธ(๋””์ฝ”๋” ์ „์šฉ) ์–ธ์–ด ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์—์„œ ๋น„๊ต์  ๋ณดํŽธ์ ์ธ ๋ฒ•์น™์ž„์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]From Figure 3, it is evident that, except for tiny models, LLaMA, Mistral, GPT2\({}_{0}\), and GPT2\({}_{1/4}\) architectures closely follow GPT2's scaling law over 1000 exposures. For tiny models with \(\leq 10M\) parameters, tying model weights increases their capacity (refer to Figure 3(c)). This indicates that the _2bit/param capacity ratio is a relatively universal law_ among most typical (decoder-only) language model architectures. [|assistant|]
2021)์€ ์‚ฌ์ „ ํ•™์Šต๋œ ๋” ํฐ ๋ชจ๋ธ์—์„œ ๋” ์ž‘์€ ๋ชจ๋ธ๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ์ „์†กํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์ข…์ข… embedded matching๊ณผ ๊ฐ™์€ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์ˆ ์ด ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ LMs์„ ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ ์ฆ๋ฅ˜์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋„ ์ง„ํ–‰๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
2021)์€ ์‚ฌ์ „ ํ•™์Šต๋œ ๋” ํฐ ๋ชจ๋ธ์—์„œ ๋” ์ž‘์€ ๋ชจ๋ธ๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ์ „์†กํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์ข…์ข… ์ž„๋ฒ ๋””๋“œ ๋งค์นญ๊ณผ ๊ฐ™์€ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์ˆ ์ด ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ LM์„ ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ ์ฆ๋ฅ˜์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋„ ์ง„ํ–‰๋˜์—ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]2021) involves transferring data from a pre-trained, larger model to a smaller one, often using methods such as embedded matching. Research has even been conducted on data distillation using LMs as the technology has evolved. [|assistant|]
DPO์˜ reward/KL tradeoff๋Š” PPO๋ฅผ ์—„๊ฒฉํ•˜๊ฒŒ ์ง€๋ฐฐํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, DPO๋Š” PPO๋ณด๋‹ค ๋” ๋‚˜์€ ํ”„๋ก ํ‹ฐ์–ด๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค(PPO-GT).[|endofturn|]
DPO์˜ ๋ณด์ƒ/KL ํŠธ๋ ˆ์ด๋“œ์˜คํ”„๋Š” PPO๋ฅผ ์—„๊ฒฉํ•˜๊ฒŒ ์ง€๋ฐฐํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, DPO๋Š” PPO๋ณด๋‹ค ๋” ๋‚˜์€ ํ”„๋ก ํ‹ฐ์–ด๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค(PPO-GT).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]DPO's reward/KL tradeoff strictly dominates PPO. Second, DPO achieves a better frontier than PPO, _even when PPO can access ground truth rewards_ (PPO-GT). [|assistant|]
๊ผฌ๋ฆฌ ์ขํžˆ๊ธฐ. ๊ทธ๋ฆผ 2(Llama2์˜ ๊ฒฝ์šฐ)๋Š” ๊ผฌ๋ฆฌ ์ ˆ๋‹จ ์™ธ์—๋„ AI ์ƒ์„ฑ ์ค‘์— ๊ผฌ๋ฆฌ ์ขํž˜ ํšจ๊ณผ๊ฐ€ ๋ฐœ์ƒํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด์˜ ํ•œ ๊ฐ€์ง€ ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก ์ค‘ ์˜จ๋„ ํ•˜๋ฝ์ž…๋‹ˆ๋‹ค. i๋ฒˆ์งธ ํ† ํฐ์— ๋Œ€ํ•œ ๋กœ์ง“ \(z_{i}\)์— softmax ๋ถ„ํฌ๋ฅผ ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค: \(p_{i}=e^{z_{i}}/\sum_{j}e^{z_{j}}\). ์ผ๋ฐ˜์ ์ธ ์˜จ๋„ \(T\)์— ๋Œ€ํ•ด \(q_{i}^{T}=e^{z_{i}/T}/\sum_{j}e^{z_{j}/T}\)๋ฅผ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋ฉด \(p_{i}\asymp i^{-\beta}\)๋Š” \(q_{i}^{T}\asymp i^{-\beta/T}\)๋กœ ๋ณ€ํ˜•๋ฉ๋‹ˆ๋‹ค(์ฒซ ๋ฒˆ์งธ ์ˆœ์„œ๋กœ). ์˜จ๋„ scaling์€ \(T>1\)์˜ ๊ฒฝ์šฐ ๊ผฌ๋ฆฌ์˜ ์ข์•„์ง์„ ์ง์ ‘์ ์œผ๋กœ ์œ ๋ฐœํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, perplexity๊ฐ€ ์žˆ๋Š” autoregressive ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ† ํฐ๋ณ„ ๊ผฌ๋ฆฌ ์ ˆ๋‹จ์ด ์‹œํ€€์Šค perplexity์— ๋Œ€ํ•œ ๊ผฌ๋ฆฌ ์ขํž˜์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 35 ๋ฐ ๋ถ€๋ก I์˜ ํ† ๋ก  ์ฐธ์กฐ).[|endofturn|]
๊ผฌ๋ฆฌ ์ขํžˆ๊ธฐ.๊ทธ๋ฆผ 2(Llama2์˜ ๊ฒฝ์šฐ)๋Š” ๊ผฌ๋ฆฌ ์ ˆ๋‹จ ์™ธ์—๋„ AI ์ƒ์„ฑ ์ค‘์— ๊ผฌ๋ฆฌ ์ขํž˜ ํšจ๊ณผ๊ฐ€ ๋ฐœ์ƒํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด์˜ ํ•œ ๊ฐ€์ง€ ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก ์ค‘ ์˜จ๋„ ํ•˜๋ฝ์ž…๋‹ˆ๋‹ค. i๋ฒˆ์งธ ํ† ํฐ์— ๋Œ€ํ•œ ๋กœ์ง“ \(z_{i}\)์— ์†Œํ”„ํŠธ๋งฅ์Šค ๋ถ„ํฌ๋ฅผ ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค: \(p_{i}=e^{z_{i}}/\sum_{j}e^{z_{j}}\). ์ผ๋ฐ˜์ ์ธ ์˜จ๋„ \(T\)์— ๋Œ€ํ•ด \(q_{i}^{T}=e^{z_{i}/T}/\sum_{j}e^{z_{j}/T}\)๋ฅผ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋ฉด \(p_{i}\asymp i^{-\beta}\)๋Š” \(q_{i}^{T}\asymp i^{-\beta/T}\)๋กœ ๋ณ€ํ˜•๋ฉ๋‹ˆ๋‹ค(์ฒซ ๋ฒˆ์งธ ์ˆœ์„œ๋กœ). ์˜จ๋„ ์Šค์ผ€์ผ๋ง์€ \(T>1\)์˜ ๊ฒฝ์šฐ ๊ผฌ๋ฆฌ์˜ ์ข์•„์ง์„ ์ง์ ‘์ ์œผ๋กœ ์œ ๋ฐœํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋‹นํ˜น๊ฐ์ด ์žˆ๋Š” ์ž๋™ ํšŒ๊ท€ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ํ† ํฐ๋ณ„ ๊ผฌ๋ฆฌ ์ ˆ๋‹จ์ด ์‹œํ€€์Šค ๋‹นํ˜น๊ฐ์— ๋Œ€ํ•œ ๊ผฌ๋ฆฌ ์ขํž˜์„ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 35 ๋ฐ ๋ถ€๋ก I์˜ ํ† ๋ก  ์ฐธ์กฐ).[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Tail Narrowing.Figure 2 (for Llama2) shows that in addition to tail cutting, tail narrowing effects happen during AI-generation. One mechanism for this is lowered temperature during next-token prediction. Assume a softmax distribution on the logits \(z_{i}\) for the \(i\)th token: \(p_{i}=e^{z_{i}}/\sum_{j}e^{z_{j}}\). Define \(q_{i}^{T}=e^{z_{i}/T}/\sum_{j}e^{z_{j}/T}\) for general temperature \(T\). Then \(p_{i}\asymp i^{-\beta}\) morphs into \(q_{i}^{T}\asymp i^{-\beta/T}\) (to first order). We see that temperature scaling directly causes narrowing of tail for \(T>1\). Other mechanisms can come to play: for instance, for autoregressive models with perplexity, token-wise tail cutting can result in tail narrowing for sequence-perplexity (see Figure 35 and discussion in Appendix I). [|assistant|]
๊ทธ๋ฆผ 4: ์งˆ๋ฌธ์—์„œ ์ฃผ์–ด ๊ฐœ์ฒด์˜ ์ธ๊ธฐ๋„์™€ ์งˆ๋ฌธ์— ๋Œ€ํ•œ FlanT5-base ๋ฐ FlanT5-large์˜ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ Retrieval-augmented generation (RAG) ๋ฐ Fine-tuning์˜ ์˜ํ–ฅ ๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„.[|endofturn|]
๊ทธ๋ฆผ 4: ์งˆ๋ฌธ์—์„œ ์ฃผ์–ด ๊ฐœ์ฒด์˜ ์ธ๊ธฐ๋„์™€ ์งˆ๋ฌธ์— ๋Œ€ํ•œ FlanT5-base ๋ฐ FlanT5-large์˜ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ RAG ๋ฐ FT์˜ ์˜ํ–ฅ ๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 4: Correlation between the popularity of the subject entity in a question and the impact of RAG and FT on the performance of FlanT5-base and FlanT5-large in QA. [|assistant|]
๋ฐฉ์ •์‹ 2๋ฅผ ํ’€๋ฉด C4 ์ „์šฉ ์œ„ํ—˜์ด ์ตœ์†Œํ™”๋˜์ง€ ์•Š์œผ๋ฏ€๋กœ C4์—์„œ \(\theta_{\mathcal{C}4}\)์™€ \(\theta_{\mathbf{WRAP}}\)๋ฅผ ๋น„๊ตํ•˜๋Š” ๊ฒƒ์€ ๋ถˆ๊ณต์ •ํ•ฉ๋‹ˆ๋‹ค. C4์™€ ๊ทธ ํ•ฉ์„ฑ ๋ฌธ๊ตฌ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ์˜๋ฏธ ์žˆ๊ฒŒ ๋น„๊ตํ•˜๊ธฐ ์œ„ํ•ด Pile์˜ 21๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ๋„๋ฉ”์ธ(Gao et al., 2020)์— ๋Œ€ํ•œ generalization ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ฐฉ์ •์‹ 2๋ฅผ ํ’€๋ฉด C4 ์ „์šฉ ์œ„ํ—˜์ด ์ตœ์†Œํ™”๋˜์ง€ ์•Š์œผ๋ฏ€๋กœ C4์—์„œ \(\theta_{\mathcal{C}4}\)์™€ \(\theta_{\mathbf{WRAP}}\)๋ฅผ ๋น„๊ตํ•˜๋Š” ๊ฒƒ์€ ๋ถˆ๊ณต์ •ํ•ฉ๋‹ˆ๋‹ค. C4์™€ ๊ทธ ํ•ฉ์„ฑ ๋ฌธ๊ตฌ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ์˜๋ฏธ ์žˆ๊ฒŒ ๋น„๊ตํ•˜๊ธฐ ์œ„ํ•ด Pile์˜ 21๊ฐœ์˜ ์„œ๋กœ ๋‹ค๋ฅธ ๋„๋ฉ”์ธ(Gao et al., 2020)์— ๋Œ€ํ•œ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 2์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Solving for equation 2 does not minimize the risk over C4-only, and hence it is unfair to compare \(\theta_{\mathcal{C}4}\) and \(\theta_{\mathbf{WRAP}}\) on the C4. For meaningfully comparing models trained on the C4 and on its synthetic rephrases, we evaluate their generalization capability on 21 different domains of the Pile (Gao et al., 2020). Results for each domain are presented in Figure 2. [|assistant|]
๊ธฐ๋ณธ LLM PaLM 2 [10]๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์ถ•๋œ AMIE๋Š” ์˜๋ฃŒ ๋Œ€ํ™” ๋ฐ ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด Instruction Fine-tuning์„ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค. ๊ธฐ๋ณธ LLM ์•„ํ‚คํ…์ฒ˜์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ PaLM-2 ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.[|endofturn|]
๊ธฐ๋ณธ LLM PaLM 2 [10]๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์ถ•๋œ AMIE๋Š” ์˜๋ฃŒ ๋Œ€ํ™” ๋ฐ ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๋ช…๋ น ๋ฏธ์„ธ ์กฐ์ •์„ ๊ฑฐ์ณค์Šต๋‹ˆ๋‹ค. ๊ธฐ๋ณธ LLM ์•„ํ‚คํ…์ฒ˜์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ PaLM-2 ๊ธฐ์ˆ  ๋ณด๊ณ ์„œ๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]AMIE, built upon the base LLM PaLM 2 [10], was instruction fine-tuned to enhance its capabilities for medical dialogue and reasoning. We refer to the PaLM-2 technical report for more details on the base LLM architecture. [|assistant|]
์ค‘๋ณต ์ œ๊ฑฐ. Deduplication์€ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์—์„œ ๋ฐ˜๋ณต๋˜๋Š” ์ถ”์ถœ ๋ฐ ๋ฌธ์„œ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ๋ชจ๋“  ๋ฌธ์ž์—์„œ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๊ฑฐ๋‚˜ ์ผ๋ถ€ ์œ ์‚ฌ์„ฑ ๋ฉ”ํŠธ๋ฆญ์— ๋”ฐ๋ผ ๊ทผ์‚ฌ์น˜ ์ผ์น˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ •ํ™•ํ•œ ์ค‘๋ณต์˜ ๊ฒฝ์šฐ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด(Manber and Myers, 1993)์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์†Œ ๊ธธ์ด์˜ ์ •ํ™•ํ•œ ํ•˜์œ„ ๋ฌธ์ž์—ด์„ ์ผ์น˜์‹œํ‚ค๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ํผ์ง€ ์ค‘๋ณต์˜ ๊ฒฝ์šฐ, Large Language Models (LLMs)์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ(Brown et al., 2020; Zeng et al., 2021; Rae et al., 2021)์— MinHash(Broder, 1997) ๋˜๋Š” SimHash(Charikar, 2002)์™€ ๊ฐ™์€ ๋กœ์ปฌ ๋ฏผ๊ฐ ํ•ด์‹œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ์ฑ„ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ Abbas et al. (2023)์€ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ์ž„๋ฒ ๋”ฉ์„ ํ™œ์šฉํ•˜์—ฌ ๋Œ€๋žต์ ์ธ ์ผ์น˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ์˜๋ฏธ๋ก ์  ์ดํ•ด๋ฅผ ๋ถ€์—ฌํ•  ๊ฒƒ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. Deduplication์€ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ํ™•์ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Allamanis, 2019; Lee et al., 2022). ํŠนํžˆ, ์ค‘๋ณต ๋ฐ์ดํ„ฐ๋Š” ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ(Carlini et al., 2021)์—์„œ ํŠนํžˆ ๋ฌธ์ œ๊ฐ€ ๋˜๋Š” ์•”๊ธฐ(Carlini et al., 2022)๋ฅผ ์ค„์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๋ฐ˜๋ณต๋˜๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ๋ชจ๋ธ ํ’ˆ์งˆ์— ๋ฏธ์น˜๋Š” ํ•ด๋กœ์šด ์˜ํ–ฅ์ด ์ ์  ๋” ์ปค์ง€๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Hernandez et al., 2022): ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ 10์–ต ๊ฐœ์ธ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 100๊ฐœ์˜ ์ค‘๋ณต์ด ํ•ด๋กญ์ง€๋งŒ, 1750์–ต ๊ฐœ์ธ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋ช‡ ๊ฐœ์˜ ์ค‘๋ณต๋งŒ์œผ๋กœ๋„ ๋ถˆ๊ท ํ˜•ํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…๊ณผ ๋™์‹œ์—, Pythia ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์€ The Pile์„ ์ค‘๋ณต ์ œ๊ฑฐํ•˜๋ฉด zero-shot ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์ด ์ œํ•œ์ ์ด๋ผ๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(Biderman et al., 2023), ์ด๋Š” Deduplication์ด ์ฃผ๋กœ ์›น ๊ธฐ๋ฐ˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋งŒ ํ•ด๋‹น๋˜๋Š”์ง€ ์˜๋ฌธ์„ ์ œ๊ธฐํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ค‘๋ณต ์ œ๊ฑฐ.์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์—์„œ ๋ฐ˜๋ณต๋˜๋Š” ์ถ”์ถœ ๋ฐ ๋ฌธ์„œ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ๋ชจ๋“  ๋ฌธ์ž์—์„œ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๊ฑฐ๋‚˜ ์ผ๋ถ€ ์œ ์‚ฌ์„ฑ ๋ฉ”ํŠธ๋ฆญ์— ๋”ฐ๋ผ ๊ทผ์‚ฌ์น˜ ์ผ์น˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ •ํ™•ํ•œ ์ค‘๋ณต์˜ ๊ฒฝ์šฐ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด(Manber and Myers, 1993)์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์†Œ ๊ธธ์ด์˜ ์ •ํ™•ํ•œ ํ•˜์œ„ ๋ฌธ์ž์—ด์„ ์ผ์น˜์‹œํ‚ค๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ํผ์ง€ ์ค‘๋ณต์˜ ๊ฒฝ์šฐ, ์ตœ๋Œ€ ์–ธ์–ด ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ(Brown et al., 2020; Zeng et al., 2021; Rae et al., 2021)์— MinHash(Broder, 1997) ๋˜๋Š” SimHash(Charikar, 2002)์™€ ๊ฐ™์€ ๋กœ์ปฌ ๋ฏผ๊ฐ ํ•ด์‹œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ์ฑ„ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ Abbas et al. (2023)์€ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ์ž„๋ฒ ๋”ฉ์„ ํ™œ์šฉํ•˜์—ฌ ๋Œ€๋žต์ ์ธ ์ผ์น˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ์˜๋ฏธ๋ก ์  ์ดํ•ด๋ฅผ ๋ถ€์—ฌํ•  ๊ฒƒ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๋ณต ์ œ๊ฑฐ๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ํ™•์ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Allamanis, 2019; Lee et al., 2022). ํŠนํžˆ, ์ค‘๋ณต ๋ฐ์ดํ„ฐ๋Š” ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ(Carlini et al., 2021)์—์„œ ํŠนํžˆ ๋ฌธ์ œ๊ฐ€ ๋˜๋Š” ์•”๊ธฐ(Carlini et al., 2022)๋ฅผ ์ค„์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๋ฐ˜๋ณต๋˜๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ๋ชจ๋ธ ํ’ˆ์งˆ์— ๋ฏธ์น˜๋Š” ํ•ด๋กœ์šด ์˜ํ–ฅ์ด ์ ์  ๋” ์ปค์ง€๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Hernandez et al., 2022): ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ 10์–ต ๊ฐœ์ธ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 100๊ฐœ์˜ ์ค‘๋ณต์ด ํ•ด๋กญ์ง€๋งŒ, 1750์–ต ๊ฐœ์ธ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋ช‡ ๊ฐœ์˜ ์ค‘๋ณต๋งŒ์œผ๋กœ๋„ ๋ถˆ๊ท ํ˜•ํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…๊ณผ ๋™์‹œ์—, Pythia ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์€ ๋” ๋งŽ์€ ํŒŒ์ผ์„ ์ค‘๋ณต ์ œ๊ฑฐํ•˜๋ฉด ์ œ๋กœ ์ƒท ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์ด ์ œํ•œ์ ์ด๋ผ๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(Biderman et al., 2023), ์ด๋Š” ์ค‘๋ณต ์ œ๊ฑฐ๊ฐ€ ์ฃผ๋กœ ์›น ๊ธฐ๋ฐ˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋งŒ ํ•ด๋‹น๋˜๋Š”์ง€ ์˜๋ฌธ์„ ์ œ๊ธฐํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Deduplication.Deduplication removes repeated extracts and documents from a dataset: these could either be exact matches, identical in every character, or approximate matches, based on some similarity metric. For exact duplicates, it is common to match exact substrings of a minimum length using suffix arrays (Manber and Myers, 1993). For fuzzy duplicates, methods based on locally-sensitive hashes such as MinHash (Broder, 1997) or SimHash (Charikar, 2002) have been adopted for the pretraining data of large language models (Brown et al., 2020; Zeng et al., 2021; Rae et al., 2021). Recently, Abbas et al. (2023) has proposed to leverage embeddings from pretrained models to imbune semantic understanding in approximate matching algorithms. Deduplication has been identified as playing a significant role in improving language models (Allamanis, 2019; Lee et al., 2022). Notably, it reduces memorization (Carlini et al., 2022), which is especially problematic in large models (Carlini et al., 2021). Furthermore, repeated data has been shown to be increasingly harmful to model quality as parameter count increases (Hernandez et al., 2022): for a 1B parameters model, a hundred duplicates are harmful; at 175B, even a few duplicates could have a disproportionate effect. Concurrently to this work, the Pythia suite of models found that deduplicating The Pile had a limited impact on zero-shot performance (Biderman et al., 2023), questioning whether deduplication is as relevant for curated corpora as it for predominantly web-based datasets. [|assistant|]
์ด ์„น์…˜์—์„œ๋Š” ์„ ํƒํ•œ ๋‹คํ•™์ œ ์ค‘๊ตญ์–ด ํ•˜๋“œ ์ผ€์ด์Šค Instruction ์ดํ•ด์˜ ๋ช‡ ๊ฐ€์ง€ ์‹œ์—ฐ๊ณผ CHC-Bench์—์„œ ์‚ฌ์šฉ๋˜๋Š” ๋‹ค์Œ ๋ฌธ์ œ ์„ธํŠธ๋ฅผ ๋‚˜์—ดํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ์ œ ๋ฒ”์ฃผ์˜ ๊ตฌ์ฒด์ ์ธ ๋ถ„๋ฅ˜๋Š” ํ‘œ 10์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ด ์„น์…˜์—์„œ๋Š” ์„ ํƒํ•œ ๋‹คํ•™์ œ ์ค‘๊ตญ์–ด ํ•˜๋“œ ์ผ€์ด์Šค ์ง€์‹œ ์ดํ•ด์˜ ๋ช‡ ๊ฐ€์ง€ ์‹œ์—ฐ๊ณผ CHC-Bench์—์„œ ์‚ฌ์šฉ๋˜๋Š” ๋‹ค์Œ ๋ฌธ์ œ ์„ธํŠธ๋ฅผ ๋‚˜์—ดํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ์ œ ๋ฒ”์ฃผ์˜ ๊ตฌ์ฒด์ ์ธ ๋ถ„๋ฅ˜๋Š” ํ‘œ 10์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this section, we list some demonstrations of our selected multidisciplinary Chinese hard case instruction understanding and the following problem sets that are used in CHC-Bench. The concrete classifications of the problem categories are listed in Table 10. [|assistant|]
์ƒ์„ฑ๋œ ๊ณ„ํš์„ ์‹คํ–‰ํ•œ ํ›„, ํ™˜๊ฒฝ์€ LLM ๊ธฐ๋ฐ˜ ์ž‘์—… ํ”Œ๋ž˜๋„ˆ์— feedback ์‹ ํ˜ธ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์œ„ํ•ด ์ดˆ๊ธฐ ๊ณ„ํš์„ ๊ตฌ์ฒดํ™”ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด ์ž‘์—…์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ LLM ๊ธฐ๋ฐ˜ ์ž‘์—… ํ”Œ๋ž˜๋„ˆ์™€์˜ ๊ด€๊ณ„์— ๋”ฐ๋ผ ํ™˜๊ฒฝ์—์„œ feedback์˜ ๋‘ ๊ฐ€์ง€ ์†Œ์Šค, ์ฆ‰ ๋‚ด๋ถ€(_์ฆ‰, LLM ์ž์ฒด)์™€ ์™ธ๋ถ€(_์˜ˆ: ๋„๊ตฌ ๋˜๋Š” ๊ฐ€์ƒ ์„ธ๊ณ„) feedback์ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ƒ์„ฑ๋œ ๊ณ„ํš์„ ์‹คํ–‰ํ•œ ํ›„, ํ™˜๊ฒฝ์€ LLM ๊ธฐ๋ฐ˜ ์ž‘์—… ํ”Œ๋ž˜๋„ˆ์— ํ”ผ๋“œ๋ฐฑ ์‹ ํ˜ธ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์œ„ํ•ด ์ดˆ๊ธฐ ๊ณ„ํš์„ ๊ตฌ์ฒดํ™”ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด ์ž‘์—…์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ LLM ๊ธฐ๋ฐ˜ ์ž‘์—… ํ”Œ๋ž˜๋„ˆ์™€์˜ ๊ด€๊ณ„์— ๋”ฐ๋ผ ํ™˜๊ฒฝ์—์„œ ํ”ผ๋“œ๋ฐฑ์˜ ๋‘ ๊ฐ€์ง€ ์†Œ์Šค, ์ฆ‰ ๋‚ด๋ถ€(_์ฆ‰, LLM ์ž์ฒด)์™€ ์™ธ๋ถ€(_์˜ˆ: ๋„๊ตฌ ๋˜๋Š” ๊ฐ€์ƒ ์„ธ๊ณ„) ํ”ผ๋“œ๋ฐฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]After executing the generated plan, the environment would produce the feedback signal to the LLM-based task planner, which can be used to refine its initial plan for better results. In existing work, there are typically two sources of feedback from the environment, depending on their relationship with the LLM-based task planner: internal (_i.e.,_ the LLM itself) and external (_e.g.,_ tools or virtual worlds) feedback. [|assistant|]
LLMs๋Š” ์‹ ๊ฒฝ๋ง์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ์‚ฌ์ „ ํ•™์Šต๋œ ํ†ต๊ณ„ ์–ธ์–ด ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ตœ๊ทผ LLMs์˜ ์„ฑ๊ณต์€ ์–ธ์–ด ๋ชจ๋ธ์˜ ์ˆ˜์‹ญ ๋…„์— ๊ฑธ์นœ ์—ฐ๊ตฌ์™€ ๊ฐœ๋ฐœ์ด ์ถ•์ ๋œ ๊ฒƒ์œผ๋กœ, ์ถœ๋ฐœ์ ๊ณผ ์†๋„๊ฐ€ ๋‹ค๋ฅธ ํ†ต๊ณ„ ์–ธ์–ด ๋ชจ๋ธ, ์‹ ๊ฒฝ ์–ธ์–ด ๋ชจ๋ธ, ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ, LLMs์˜ ๋„ค ๊ฐ€์ง€ ํŒŒ๋„๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
LLM์€ ์‹ ๊ฒฝ๋ง์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ์‚ฌ์ „ ํ•™์Šต๋œ ํ†ต๊ณ„ ์–ธ์–ด ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ตœ๊ทผ LLM์˜ ์„ฑ๊ณต์€ ์–ธ์–ด ๋ชจ๋ธ์˜ ์ˆ˜์‹ญ ๋…„์— ๊ฑธ์นœ ์—ฐ๊ตฌ์™€ ๊ฐœ๋ฐœ์ด ์ถ•์ ๋œ ๊ฒƒ์œผ๋กœ, ์ถœ๋ฐœ์ ๊ณผ ์†๋„๊ฐ€ ๋‹ค๋ฅธ ํ†ต๊ณ„ ์–ธ์–ด ๋ชจ๋ธ, ์‹ ๊ฒฝ ์–ธ์–ด ๋ชจ๋ธ, ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ, LLM์˜ ๋„ค ๊ฐ€์ง€ ํŒŒ๋„๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]LLMs are large-scale, pre-trained, statistical language models based on neural networks. The recent success of LLMs is an accumulation of decades of research and development of language models, which can be categorized into four waves that have different starting points and velocity: statistical language models, neural language models, pre-trained language models and LLMs. [|assistant|]
ํƒ์š•์  ๊ฒ€์ƒ‰์€ ๊ฐ ๋‹จ๊ณ„์—์„œ ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ token์„ ์‹œํ€€์Šค์˜ ๋‹ค์Œ token์œผ๋กœ ์„ ํƒํ•˜๊ณ  ๋‹ค๋ฅธ ๋ชจ๋“  ์ž ์žฌ์  ์˜ต์…˜์„ ๋ฌด์‹œํ•ฉ๋‹ˆ๋‹ค. ์ƒ์ƒํ•  ์ˆ˜ ์žˆ๋“ฏ์ด ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ฐ„๋‹จํ•˜์ง€๋งŒ ์‹œ๊ฐ„์  ์ผ๊ด€์„ฑ๊ณผ ์ผ๊ด€์„ฑ์„ ๋งŽ์ด ์žƒ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹œํ€€์Šค์— ๋Œ€ํ•œ ์ „๋ฐ˜์ ์ธ ์˜ํ–ฅ์„ ๊ณ ๋ คํ•˜์ง€ ์•Š๊ณ  ๊ฐ ๋‹จ๊ณ„์—์„œ ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ token๋งŒ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด ์†์„ฑ์œผ๋กœ ์ธํ•ด ์†๋„๊ฐ€ ๋น ๋ฅด์ง€๋งŒ, ์•ฝ๊ฐ„ ๋œ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ๋‹ค์Œ token์œผ๋กœ ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ๋Š” ๋” ๋‚˜์€ ์‹œํ€€์Šค๋ฅผ ๋†“์น  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
ํƒ์š•์  ๊ฒ€์ƒ‰์€ ๊ฐ ๋‹จ๊ณ„์—์„œ ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ† ํฐ์„ ์‹œํ€€์Šค์˜ ๋‹ค์Œ ํ† ํฐ์œผ๋กœ ์„ ํƒํ•˜๊ณ  ๋‹ค๋ฅธ ๋ชจ๋“  ์ž ์žฌ์  ์˜ต์…˜์„ ๋ฌด์‹œํ•ฉ๋‹ˆ๋‹ค. ์ƒ์ƒํ•  ์ˆ˜ ์žˆ๋“ฏ์ด ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ฐ„๋‹จํ•˜์ง€๋งŒ ์‹œ๊ฐ„์  ์ผ๊ด€์„ฑ๊ณผ ์ผ๊ด€์„ฑ์„ ๋งŽ์ด ์žƒ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹œํ€€์Šค์— ๋Œ€ํ•œ ์ „๋ฐ˜์ ์ธ ์˜ํ–ฅ์„ ๊ณ ๋ คํ•˜์ง€ ์•Š๊ณ  ๊ฐ ๋‹จ๊ณ„์—์„œ ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ† ํฐ๋งŒ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ด ์†์„ฑ์œผ๋กœ ์ธํ•ด ์†๋„๊ฐ€ ๋น ๋ฅด์ง€๋งŒ, ์•ฝ๊ฐ„ ๋œ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ๋‹ค์Œ ํ† ํฐ์œผ๋กœ ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ๋Š” ๋” ๋‚˜์€ ์‹œํ€€์Šค๋ฅผ ๋†“์น  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Greedy search takes the most probable token at each step as the next token in the sequence, discarding all other potential options. As you can imagine, this is a simple approach and can loose a lot of temporal consistency and coherency. It only considers the most probable token at each step, without considering the overall effect on the sequence. This property makes it fast, but it also means that it can miss out on better sequences that might have appeared with slightly less probable next tokens. [|assistant|]
sDPO์˜ ๊ฐ ๋‹จ๊ณ„์˜ target model๋„ ๋งˆ์ง€๋ง‰ ๋‹จ๊ณ„์˜ ์ •๋ ฌ๋œ ๋ชจ๋ธ์ธ \(M_{t-1}\)๋กœ ์ดˆ๊ธฐํ™”๋ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด sDPO์˜ ์ตœ์ข… ๋ชจ๋ธ์€ DPO์˜ ์ตœ์ข… ๋ชจ๋ธ๊ณผ ๋™์ผํ•œ ์–‘์˜ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต์„ ๊ฑฐ์นœ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์ด๋Ÿฌํ•œ ์„ค๊ณ„ ์„ ํƒ์˜ ํ•œ ๊ฐ€์ง€ ์šฐ๋ ค ์‚ฌํ•ญ์€ ์ด์ „ ๋‹จ๊ณ„์—์„œ ์ด๋ฏธ ๊ฐ์†Œํ•˜๋Š” Learning Rate ์ผ์ •์œผ๋กœ ํ•™์Šต์„ ๊ฑฐ์ณค๊ธฐ ๋•Œ๋ฌธ์— ๋‹จ๊ณ„๊ฐ€ ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ target model์˜ ํ•™์Šต์„ ์•ˆ์ •ํ™”ํ•˜๊ธฐ๊ฐ€ ์ ์  ๋” ์–ด๋ ค์›Œ์งˆ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋˜ ๋‹ค๋ฅธ ์˜ต์…˜์€ sDPO์˜ ๋ชจ๋“  ๋‹จ๊ณ„์— ๋Œ€ํ•ด ์ดˆ๊ธฐ SFT ๊ธฐ๋ณธ ๋ชจ๋ธ \(S\)๋ฅผ target model๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
sDPO์˜ ๊ฐ ๋‹จ๊ณ„์˜ ๋Œ€์ƒ ๋ชจ๋ธ๋„ ๋งˆ์ง€๋ง‰ ๋‹จ๊ณ„์˜ ์ •๋ ฌ๋œ ๋ชจ๋ธ์ธ \(M_{t-1}\)๋กœ ์ดˆ๊ธฐํ™”๋ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด sDPO์˜ ์ตœ์ข… ๋ชจ๋ธ์€ DPO์˜ ์ตœ์ข… ๋ชจ๋ธ๊ณผ ๋™์ผํ•œ ์–‘์˜ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต์„ ๊ฑฐ์นœ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์ด๋Ÿฌํ•œ ์„ค๊ณ„ ์„ ํƒ์˜ ํ•œ ๊ฐ€์ง€ ์šฐ๋ ค ์‚ฌํ•ญ์€ ์ด์ „ ๋‹จ๊ณ„์—์„œ ์ด๋ฏธ ๊ฐ์†Œํ•˜๋Š” ํ•™์Šต ์†๋„ ์ผ์ •์œผ๋กœ ํ•™์Šต์„ ๊ฑฐ์ณค๊ธฐ ๋•Œ๋ฌธ์— ๋‹จ๊ณ„๊ฐ€ ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ ๋Œ€์ƒ ๋ชจ๋ธ์˜ ํ•™์Šต์„ ์•ˆ์ •ํ™”ํ•˜๊ธฐ๊ฐ€ ์ ์  ๋” ์–ด๋ ค์›Œ์งˆ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋˜ ๋‹ค๋ฅธ ์˜ต์…˜์€ sDPO์˜ ๋ชจ๋“  ๋‹จ๊ณ„์— ๋Œ€ํ•ด ์ดˆ๊ธฐ SFT ๊ธฐ๋ณธ ๋ชจ๋ธ \(S\)๋ฅผ ๋Œ€์ƒ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]The target model in each step of sDPO is also initialized with \(M_{t-1}\), the aligned model from the last step. This ensures that the final model in sDPO has undergone training with the same amount of data as the final model in DPO. On the other hand, one concern of such design choice is that it may become increasingly difficult to stabilize the training of the target model as the steps progress, since it has already undergone training with a decreasing learning rate schedule in the preceding steps. Thus, another option is to use the initial SFT base model \(S\) as the target model for all steps of sDPO. [|assistant|]
* ์ƒˆ๋กœ์šด ์ง€์‹์„ ํฌํ•จํ•˜๋Š” ์ƒˆ๋กœ์šด ์–ธ์–ด ๋ชจ๋ธ๋ง ์ฝ”ํผ์Šค์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ์žŠ์–ด๋ฒ„๋ฆฐ ์ง€์‹์˜ ์–‘๊ณผ ์–ป์€ ์„ธ๊ณ„ ์ง€์‹์˜ ์–‘์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด CL ๊ณต์‹์ธ ์ง€์†์  ์ง€์‹ ํ•™์Šต(CKL)์„ ์ œ์•ˆํ•˜๊ณ  ์ƒˆ๋กœ์šด benchmark๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. * ๋ฌธํ—Œ์—์„œ CKL์˜ ์ž์—ฐ์Šค๋Ÿฌ์šด ๊ธฐ์ค€์ด ๋˜๋Š” LM ์•„ํ‚คํ…์ฒ˜์™€ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก ์„ ํƒ์ƒ‰ํ•˜์—ฌ CKL ๋ฐฉ๋ฒ•์œผ๋กœ ํ‘œ์‹œํ•˜๊ณ , CKL benchmark์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด CL ๋ฌธํ—Œ์—์„œ์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ regularization, rehearsal, parameter-expansion ๋ฐฉ๋ฒ•์œผ๋กœ ๋ถ„๋ฅ˜ํ•˜๊ณ , ์žŠ์–ด๋ฒ„๋ฆฐ ์ง€์‹๊ณผ ์—…๋ฐ์ดํŠธ ๋˜๋Š” ํš๋“ํ•œ ์ง€์‹ ๊ฐ„์˜ ํŠธ๋ ˆ์ด๋“œ์˜คํ”„๋ฅผ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์ œ์•ˆํ•œ ์ƒˆ๋กœ์šด metric์ธ FUAR๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์œ ํ˜•์˜ ๋ฐฉ๋ฒ•์˜ ํšจ๊ณผ๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. * ๋Š์ž„์—†์ด ๋ณ€ํ™”ํ•˜๋Š” LM์„ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด CKL benchmark์—์„œ ๊ด‘๋ฒ”์œ„ํ•œ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ์ค‘์š”ํ•œ ๊ณผ์ œ์™€ ๋ฐœ๊ฒฌ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. parameter-expansion ๋ฐฉ๋ฒ•์€ ๋Œ€๋ถ€๋ถ„์˜ ์‹คํ—˜์—์„œ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์ง€๋งŒ ๋ฉ”๋ชจ๋ฆฌ ๋น„ํšจ์œจ์„ฑ์ด๋ผ๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋™์ผํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๋ณด๋Š” ๊ฒƒ์€ ์žŠ์–ด๋ฒ„๋ฆผ์˜ ์ค‘์š”ํ•œ ์›์ธ์ž…๋‹ˆ๋‹ค.[|endofturn|]
* ์ƒˆ๋กœ์šด ์ง€์‹์„ ํฌํ•จํ•˜๋Š” ์ƒˆ๋กœ์šด ์–ธ์–ด ๋ชจ๋ธ๋ง ์ฝ”ํผ์Šค์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ์žŠ์–ด๋ฒ„๋ฆฐ ์ง€์‹์˜ ์–‘๊ณผ ์–ป์€ ์„ธ๊ณ„ ์ง€์‹์˜ ์–‘์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด CL ๊ณต์‹์ธ ์ง€์†์  ์ง€์‹ ํ•™์Šต(CKL)์„ ์ œ์•ˆํ•˜๊ณ  ์ƒˆ๋กœ์šด ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. * ๋ฌธํ—Œ์—์„œ CKL์˜ ์ž์—ฐ์Šค๋Ÿฌ์šด ๊ธฐ์ค€์ด ๋˜๋Š” LM ์•„ํ‚คํ…์ฒ˜์™€ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก ์„ ํƒ์ƒ‰ํ•˜์—ฌ CKL ๋ฐฉ๋ฒ•์œผ๋กœ ํ‘œ์‹œํ•˜๊ณ , CKL ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด CL ๋ฌธํ—Œ์—์„œ์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ •๊ทœํ™”, ๋ฆฌํ—ˆ์„ค, ํŒŒ๋ผ๋ฏธํ„ฐ ํ™•์žฅ ๋ฐฉ๋ฒ•์œผ๋กœ ๋ถ„๋ฅ˜ํ•˜๊ณ , ์žŠ์–ด๋ฒ„๋ฆฐ ์ง€์‹๊ณผ ์—…๋ฐ์ดํŠธ ๋˜๋Š” ํš๋“ํ•œ ์ง€์‹ ๊ฐ„์˜ ํŠธ๋ ˆ์ด๋“œ์˜คํ”„๋ฅผ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์ œ์•ˆํ•œ ์ƒˆ๋กœ์šด ๋ฉ”ํŠธ๋ฆญ์ธ FUAR๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์œ ํ˜•์˜ ๋ฐฉ๋ฒ•์˜ ํšจ๊ณผ๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. * ๋Š์ž„์—†์ด ๋ณ€ํ™”ํ•˜๋Š” LM์„ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด CKL ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ด‘๋ฒ”์œ„ํ•œ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ์ค‘์š”ํ•œ ๊ณผ์ œ์™€ ๋ฐœ๊ฒฌ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ํŒŒ๋ผ๋ฏธํ„ฐ ํ™•์žฅ ๋ฐฉ๋ฒ•์€ ๋Œ€๋ถ€๋ถ„์˜ ์‹คํ—˜์—์„œ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜์ง€๋งŒ ๋ฉ”๋ชจ๋ฆฌ ๋น„ํšจ์œจ์„ฑ์ด๋ผ๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋™์ผํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๋ณด๋Š” ๊ฒƒ์€ ์žŠ์–ด๋ฒ„๋ฆผ์˜ ์ค‘์š”ํ•œ ์›์ธ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]* We propose a novel CL formulation called Continual Knowledge Learning (CKL) and construct a new benchmark to measure the amount of forgetting and amount of world knowledge gained by continued pretraining on a novel language modeling corpus that we construct, containing new knowledge. * We explore LM architectures and training methodologies that are natural baselines for CKL in literature, denoting them as CKL methods, and performing extensive experiments on our CKL benchmark. We categorize them into regularization, rehearsal, and parameter-expansion methods, same as in traditional CL literature, and compare the effectiveness of each type of method using a novel metric named FUAR that we propose to measure the trade-off between forgotten knowledge and updated or acquired knowledge. * Towards creating an ever-changing LM, we perform extensive analysis in the CKL benchmark and highlight important challenges and findings: parameter-expansion methods have the limitation of memory inefficiency despite performing the best in most of our experiments and seeing the same data repeatedly during continued pretraining is a critical cause of forgetting. [|assistant|]
์–ธ์–ด ์ดํ•ด์™€ ํ…์ŠคํŠธ ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ธํ„ฐ๋„ท ๋ฐ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์™€ ๊ฐ™์€ ๋„๊ตฌ ์‚ฌ์šฉ์— ๊ฐ•๋ ฅํ•œ ์—ญ๋Ÿ‰์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ๋Šฅ์„ ํ†ตํ•ด ์—์ด์ „ํŠธ๋Š” ๋ฌธ์„œํ™” ๋ฐ ๋ฐ์ดํ„ฐ ๊ด€๋ฆฌ์™€ ๊ด€๋ จ๋œ ์ž‘์—…์„ ํƒ์›”ํ•˜๊ฒŒ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[115, 75, 116]. [115]์—์„œ ์—์ด์ „ํŠธ๋Š” question answering ๋ฐ ์‹คํ—˜ ๊ณ„ํš๊ณผ ๊ฐ™์€ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๊ธฐ ์œ„ํ•ด ์ธํ„ฐ๋„ท ์ •๋ณด๋ฅผ ํšจ์œจ์ ์œผ๋กœ ์ฟผ๋ฆฌํ•˜๊ณ  ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ChatMOF [116]์€ LLMs์„ ํ™œ์šฉํ•˜์—ฌ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ํ…์ŠคํŠธ ์„ค๋ช…์—์„œ ์ค‘์š”ํ•œ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ธˆ์†-์œ ๊ธฐ์ฒด ๊ณจ๊ฒฉ์˜ ํŠน์„ฑ๊ณผ ๊ตฌ์กฐ๋ฅผ ์˜ˆ์ธกํ•˜๊ธฐ ์œ„ํ•ด ๊ด€๋ จ ๋„๊ตฌ๋ฅผ ์ ์šฉํ•˜๋Š” ๊ณ„ํš์„ ์ˆ˜๋ฆฝํ•ฉ๋‹ˆ๋‹ค. ChemCrow [75]๋Š” ํ™”ํ•™ ๊ด€๋ จ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ™”ํ•ฉ๋ฌผ ํ‘œํ˜„์˜ ์ •๋ฐ€๋„๋ฅผ ๊ฒ€์ฆํ•˜๊ณ  ์ž ์žฌ์ ์œผ๋กœ ์œ„ํ—˜ํ•œ ๋ฌผ์งˆ์„ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ธฐ๋Šฅ์€ ๊ด€๋ จ ๋ฐ์ดํ„ฐ์˜ ์ •ํ™•์„ฑ์„ ๋ณด์žฅํ•จ์œผ๋กœ์จ ๊ณผํ•™์  ํƒ๊ตฌ์˜ ์‹ ๋ขฐ์„ฑ๊ณผ ํฌ๊ด„์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.[|endofturn|]
์–ธ์–ด ์ดํ•ด์™€ ํ…์ŠคํŠธ ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ธํ„ฐ๋„ท ๋ฐ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์™€ ๊ฐ™์€ ๋„๊ตฌ ์‚ฌ์šฉ์— ๊ฐ•๋ ฅํ•œ ์—ญ๋Ÿ‰์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ๋Šฅ์„ ํ†ตํ•ด ์—์ด์ „ํŠธ๋Š” ๋ฌธ์„œํ™” ๋ฐ ๋ฐ์ดํ„ฐ ๊ด€๋ฆฌ์™€ ๊ด€๋ จ๋œ ์ž‘์—…์„ ํƒ์›”ํ•˜๊ฒŒ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[115, 75, 116]. [115]์—์„œ ์—์ด์ „ํŠธ๋Š” ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ ์‹คํ—˜ ๊ณ„ํš๊ณผ ๊ฐ™์€ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๊ธฐ ์œ„ํ•ด ์ธํ„ฐ๋„ท ์ •๋ณด๋ฅผ ํšจ์œจ์ ์œผ๋กœ ์ฟผ๋ฆฌํ•˜๊ณ  ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ChatMOF [116]์€ LLM์„ ํ™œ์šฉํ•˜์—ฌ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ํ…์ŠคํŠธ ์„ค๋ช…์—์„œ ์ค‘์š”ํ•œ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ธˆ์†-์œ ๊ธฐ์ฒด ๊ณจ๊ฒฉ์˜ ํŠน์„ฑ๊ณผ ๊ตฌ์กฐ๋ฅผ ์˜ˆ์ธกํ•˜๊ธฐ ์œ„ํ•ด ๊ด€๋ จ ๋„๊ตฌ๋ฅผ ์ ์šฉํ•˜๋Š” ๊ณ„ํš์„ ์ˆ˜๋ฆฝํ•ฉ๋‹ˆ๋‹ค. ChemCrow [75]๋Š” ํ™”ํ•™ ๊ด€๋ จ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ™”ํ•ฉ๋ฌผ ํ‘œํ˜„์˜ ์ •๋ฐ€๋„๋ฅผ ๊ฒ€์ฆํ•˜๊ณ  ์ž ์žฌ์ ์œผ๋กœ ์œ„ํ—˜ํ•œ ๋ฌผ์งˆ์„ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ธฐ๋Šฅ์€ ๊ด€๋ จ ๋ฐ์ดํ„ฐ์˜ ์ •ํ™•์„ฑ์„ ๋ณด์žฅํ•จ์œผ๋กœ์จ ๊ณผํ•™์  ํƒ๊ตฌ์˜ ์‹ ๋ขฐ์„ฑ๊ณผ ํฌ๊ด„์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]have shown strong capabilities on language understanding and employing tools such as the internet and databases for text processing. These capabilities empower the agent to excel in tasks related to documentation and data management [115, 75, 116]. In [115], the agent can efficiently query and utilize internet information to complete tasks such as question answering and experiment planning. ChatMOF [116] utilizes LLMs to extract important information from text descriptions written by humans. It then formulates a plan to apply relevant tools for predicting the properties and structures of metal-organic frameworks. ChemCrow [75] utilizes chemistry-related databases to both validate the precision of compound representations and identify potentially dangerous substances. This functionality enhances the reliability and comprehensiveness of scientific inquiries by ensuring the accuracy of the data involved. [|assistant|]
**API.** ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค API(์„น์…˜ 5.1)์— ๋‘ ๊ฐ€์ง€ ์ถ”๊ฐ€ ์ œํ•œ์„ ๋‘๋Š”๋ฐ, ์ฒซ์งธ, \(K=1\)๋กœ ์„ค์ •ํ•˜๊ณ  ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ† ํฐ์˜ logprob๋งŒ ๋ณด๊ฒŒ ํ•˜๊ณ , ๋‘˜์งธ, ๊ฐ logit bias ํ•ญ๋ชฉ \(b\)๋Š” \(\{-1,0\}\)์— ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ œ์•ฝ ์กฐ๊ฑด์€ ์ด์ „ ์„น์…˜์˜ ๊ณต๊ฒฉ์„ ์™„์ „ํžˆ ๋ฐฉ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ œ์•ฝ ์กฐ๊ฑด์€ ์‹ค์ œ ๊ตฌํ˜„์—์„œ ์ •์˜ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ํ›จ์”ฌ ๋” ์—„๊ฒฉํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
**API.** ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค API(์„น์…˜ 5.1)์— ๋‘ ๊ฐ€์ง€ ์ถ”๊ฐ€ ์ œํ•œ์„ ๋‘๋Š”๋ฐ, ์ฒซ์งธ, \(K=1\)๋กœ ์„ค์ •ํ•˜๊ณ  ๊ฐ€์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ† ํฐ์˜ ๋กœ๊ทธ ํ™•๋ฅ ๋งŒ ๋ณด๊ฒŒ ํ•˜๊ณ , ๋‘˜์งธ, ๊ฐ ๋กœ๊ทธ ํ™•๋ฅ  ๋ฐ”์ด์–ด์Šค ํ•ญ๋ชฉ \(b\)๋Š” \(\{-1,0\}\)์— ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ œ์•ฝ ์กฐ๊ฑด์€ ์ด์ „ ์„น์…˜์˜ ๊ณต๊ฒฉ์„ ์™„์ „ํžˆ ๋ฐฉ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ œ์•ฝ ์กฐ๊ฑด์€ ์‹ค์ œ ๊ตฌํ˜„์—์„œ ์ •์˜ํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ํ›จ์”ฌ ๋” ์—„๊ฒฉํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**API.** We place two following further restrictions on the logit bias API (Section 5.1): first, we set \(K=1\), and only see the most likely token's logprob; and second, each logit bias entry \(b\) is constrained to be in \(\{-1,0\}\). These constraints would completely prevent the attacks from the prior section. We believe this constraint is significantly tighter than any practical implementation would define. [|assistant|]
์—ฐ๊ตฌ์ž๋“ค์ด ์–ธ์–ด ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผœ ์ถ”๋ก ํ•˜๊ฑฐ๋‚˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉํ•œ ํ•œ ๊ฐ€์ง€ ๋ฐฉํ–ฅ์€ ์–ธ์–ด ๋ชจ๋ธ์„ ์ฑ„๊ตด๋œ ์ถ”๋ก  ์ถ”์  ๋˜๋Š” ์ถ”๋ก ๊ณผ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต์‹œํ‚ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(Rajani et al., 2019; Wei et al., 2021; Lewkowycz et al., 2022; Chung et al., 2022; Gunasekar et al., 2023). ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์ง€๋งŒ ๋‹จ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ˆ˜์ž‘์—…์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์•„์•ผ ํ•˜๋ฉฐ, ์ด๋Š” ์ฃผ์„ ์ž‘์„ฑ์ž์˜ ์—ญ๋Ÿ‰์— ๋ฏผ๊ฐํ•˜๊ณ  ์–ธ์–ด ๋ชจ๋ธ์— ๋Œ€ํ•ด off-policy์ž…๋‹ˆ๋‹ค(์ฆ‰, ์ถ”๋ก ์˜ ๋ถ„ํฌ๋Š” ์–ธ์–ด ๋ชจ๋ธ์ด ์ƒ์„ฑํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ…์ŠคํŠธ๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค). ๋˜ํ•œ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ณ  ํ™•์žฅํ•˜๊ธฐ ์–ด๋ ต์œผ๋ฉฐ ์ฃผ์„ ์ž‘์„ฑ์ž๊ฐ€ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ฌธ์ œ๋ณด๋‹ค ๋” ์–ด๋ ค์šด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ช…ํ™•ํ•œ ๊ฒฝ๋กœ๋ฅผ ์ œ๊ณตํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.[|endofturn|]
์—ฐ๊ตฌ์ž๋“ค์ด ์–ธ์–ด ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผœ ์ถ”๋ก ํ•˜๊ฑฐ๋‚˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉํ•œ ํ•œ ๊ฐ€์ง€ ๋ฐฉํ–ฅ์€ ์–ธ์–ด ๋ชจ๋ธ์„ ์ฑ„๊ตด๋œ ์ถ”๋ก  ์ถ”์  ๋˜๋Š” ์ถ”๋ก ๊ณผ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํ•™์Šต์‹œํ‚ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(Rajani et al., 2019; Wei et al., 2021; Lewkowycz et al., 2022; Chung et al., 2022; Gunasekar et al., 2023). ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋˜์—ˆ์ง€๋งŒ ๋‹จ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ˆ˜์ž‘์—…์œผ๋กœ ์ฃผ์„์„ ๋‹ฌ์•„์•ผ ํ•˜๋ฉฐ, ์ด๋Š” ์ฃผ์„ ์ž‘์„ฑ์ž์˜ ์—ญ๋Ÿ‰์— ๋ฏผ๊ฐํ•˜๊ณ  ์–ธ์–ด ๋ชจ๋ธ์— ๋Œ€ํ•ด ์ •์ฑ…์— ๋ฏผ๊ฐํ•ฉ๋‹ˆ๋‹ค(์ฆ‰, ์ถ”๋ก ์˜ ๋ถ„ํฌ๋Š” ์–ธ์–ด ๋ชจ๋ธ์ด ์ƒ์„ฑํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์€ ํ…์ŠคํŠธ๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค). ๋˜ํ•œ ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ณ  ํ™•์žฅํ•˜๊ธฐ ์–ด๋ ต์œผ๋ฉฐ ์ฃผ์„ ์ž‘์„ฑ์ž๊ฐ€ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ฌธ์ œ๋ณด๋‹ค ๋” ์–ด๋ ค์šด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ช…ํ™•ํ•œ ๊ฒฝ๋กœ๋ฅผ ์ œ๊ณตํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]One direction that researchers have used to train language models to reason or improve their reasoning is training the language model on mined reasoning traces or reasoning-like data (Rajani et al., 2019; Wei et al., 2021; Lewkowycz et al., 2022; Chung et al., 2022; Gunasekar et al., 2023). Although this approach has been demonstrated to be effective, it comes with drawbacks. It requires either manual annotation, which is sensitive to the capability of the annotators and is off-policy for the language model (i.e., the distribution of reasoning is not text that the language model would otherwise likely have generated). This approach is also expensive, difficult to scale, and provides no clear path to solving problems harder than those that the annotators are capable of solving. [|assistant|]
1. _WizardMath 13B_๋Š” GSM8k์—์„œ PaLM 1 540B(63.9 vs 56.5), Minerva 540B(63.9 vs 58.8), GPT-3.5(63.9 vs 57.1)๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, MATH์—์„œ๋Š” PaLM 1 540B(14.0 vs 8.8), GPT-3 175B(14.0 vs 5.2)๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. 2. ๊ฐ€์žฅ ํฐ ๋ชจ๋ธ์ธ _WizardMath 70B_๋Š” GSM8k์—์„œ Claude Instant(81.6 vs 80.9), ChatGPT(81.6 vs 80.8), PaLM 2(81.6 vs 80.7)์™€ ์šฐ์ˆ˜ํ•˜๊ฑฐ๋‚˜ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋™์‹œ์— WizardMath 70B๋Š” MATH ๋ฒค์น˜๋งˆํฌ์—์„œ Text-davinci-002(22.7 vs 19.1)๋ฅผ 3.6% ํฌ์ธํŠธ ๋Šฅ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
1. _WizardMath 13B_๋Š” GSM8k์—์„œ PaLM 1 540B(63.9 ๋Œ€ 56.5), ๋ฏธ๋„ค๋ฅด๋ฐ” 540B(63.9 ๋Œ€ 58.8), GPT-3.5(63.9 ๋Œ€ 57.1)๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, MATH์—์„œ๋Š” PaLM 1 540B(14.0 ๋Œ€ 8.8), GPT-3 175B(14.0 ๋Œ€ 5.2)๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. 2. ๊ฐ€์žฅ ํฐ ๋ชจ๋ธ์ธ _WizardMath 70B_๋Š” GSM8k์—์„œ Claude Instant(81.6 ๋Œ€ 80.9), ChatGPT(81.6 ๋Œ€ 80.8), PaLM 2(81.6 ๋Œ€ 80.7)์™€ ์šฐ์ˆ˜ํ•˜๊ฑฐ๋‚˜ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋™์‹œ์— WizardMath 70B๋Š” MATH ๋ฒค์น˜๋งˆํฌ์—์„œ Text-davinci-002(22.7 ๋Œ€ 19.1)๋ฅผ 3.6% ํฌ์ธํŠธ ๋Šฅ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]1. _WizardMath 13B_ outperforms PaLM 1 540B (63.9 vs 56.5), Minerva 540B (63.9 vs 58.8), and GPT-3.5 (63.9 vs 57.1) on GSM8k. Meanwhile,it surpasses PaLM 1 540B (14.0 vs. 8.8), GPT-3 175B (14.0 vs. 5.2) on MATH. 2. _WizardMath 70B_, our largest model, achieves the superior or comparable performance with Claude Instant (81.6 vs 80.9), ChatGPT (81.6 vs 80.8) and PaLM 2 (81.6 vs 80.7) on GSM8k. Concurrently, WizardMath 70B also exceeds Text-davinci-002 (22.7 vs. 19.1) by a margin of 3.6% on the MATH benchmarks. [|assistant|]
\(\bullet\)_์„ฑ๋Šฅ์ด ์ €์กฐํ•œ ํŠน์ˆ˜ ์ƒ์„ฑ._ LLM์€ ์ผ๊ด€์„ฑ ์žˆ๋Š” ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ํŒจํ„ด์„ ํ•™์Šตํ–ˆ์ง€๋งŒ, ํŠน์ˆ˜ ๋„๋ฉ”์ธ์ด๋‚˜ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ๋•Œ๋Š” ์ƒ์„ฑ ๋Šฅ๋ ฅ์ด ์ œํ•œ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ผ๋ฐ˜ ์›น ๊ธฐ์‚ฌ๋ฅผ ํ•™์Šตํ•œ ์–ธ์–ด ๋ชจ๋ธ์€ ๋งŽ์€ ์˜ํ•™ ์ „๋ฌธ ์šฉ์–ด์™€ ๋ฐฉ๋ฒ•์ด ํฌํ•จ๋œ ์˜๋ฃŒ ๋ณด๊ณ ์„œ๋ฅผ ์ƒ์„ฑํ•  ๋•Œ ์–ด๋ ค์›€์„ ๊ฒช์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง๊ด€์ ์œผ๋กœ ๋„๋ฉ”์ธ ์ง€์‹์€ ๋ชจ๋ธ ์ „๋ฌธํ™”์— ๋งค์šฐ ์ค‘์š”ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ „๋ฌธ ์ง€์‹์„ LLM์— ์ฃผ์ž…ํ•˜๋Š” ๊ฒƒ์€ ์‰ฝ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋ถ„์„[47, 648]์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, LLM์ด ์ผ๋ถ€ ์˜์—ญ์—์„œ ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋„๋ก ํŠน์ • ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๋„๋ก ํ›ˆ๋ จํ•˜๋ฉด ๋‹ค๋ฅธ ์˜์—ญ์—์„œ๋Š” ์–ด๋ ค์›€์„ ๊ฒช์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋Š” ์‹ ๊ฒฝ๋ง ํ›ˆ๋ จ์—์„œ _catastrophic forgetting_[649, 650]๊ณผ ๊ด€๋ จ์ด ์žˆ๋Š”๋ฐ, ์ด๋Š” ์ƒˆ๋กœ์šด ์ง€์‹๊ณผ ๊ธฐ์กด ์ง€์‹์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฐˆ๋“ฑ ํ˜„์ƒ์„ ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‚ฌ๋ก€๋Š” ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ํ•„์š”์— ๋งž์ถ”๊ธฐ ์œ„ํ•ด _"alignment tax"_[66](_์˜ˆ:_ In-Context Learning ๋Šฅ๋ ฅ์˜ ์ž ์žฌ์  ์†์‹ค)์„ ์ง€๋ถˆํ•ด์•ผ ํ•˜๋Š” LLM์˜ ์ธ๊ฐ„ ์ •๋ ฌ์—์„œ๋„ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์‹œํ€€์Šค ๋ชจ๋ธ๋ง ์•„ํ‚คํ…์ฒ˜์˜ ํ•œ๊ณ„๋กœ ์ธํ•ด LLM์€ ์—ฌ์ „ํžˆ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ์˜ ์ดํ•ด์™€ ์ƒ์„ฑ์— ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ง€์‹ ๊ธฐ๋ฐ˜ ์งˆ์˜์‘๋‹ต ๋ฐ ์‹œ๋งจํ‹ฑ ๊ตฌ๋ฌธ ๋ถ„์„[651, 458]๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ ์ž‘์—…์—์„œ ์ข…์ข… ์ž‘์—…๋ณ„ ๋ชจ๋ธ์— ๋’ค์ฒ˜์ง‘๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์–‘ํ•œ ์ž‘์—… ์‹œ๋‚˜๋ฆฌ์˜ค์— LLM์„ ์œ ์—ฐํ•˜๊ฒŒ ์ ์šฉํ•˜๋Š” ๋™์‹œ์— ์›๋ž˜์˜ ๋Šฅ๋ ฅ์„ ์ตœ๋Œ€ํ•œ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ๋Š” ํšจ๊ณผ์ ์ธ ๋ชจ๋ธ ์ „๋ฌธํ™” ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
\(\bullet\)_์„ฑ๋Šฅ์ด ์ €์กฐํ•œ ํŠน์ˆ˜ ์ƒ์„ฑ._ LLM์€ ์ผ๊ด€์„ฑ ์žˆ๋Š” ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ํŒจํ„ด์„ ํ•™์Šตํ–ˆ์ง€๋งŒ, ํŠน์ˆ˜ ๋„๋ฉ”์ธ์ด๋‚˜ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ๋•Œ๋Š” ์ƒ์„ฑ ๋Šฅ๋ ฅ์ด ์ œํ•œ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ผ๋ฐ˜ ์›น ๊ธฐ์‚ฌ๋ฅผ ํ•™์Šตํ•œ ์–ธ์–ด ๋ชจ๋ธ์€ ๋งŽ์€ ์˜ํ•™ ์ „๋ฌธ ์šฉ์–ด์™€ ๋ฐฉ๋ฒ•์ด ํฌํ•จ๋œ ์˜๋ฃŒ ๋ณด๊ณ ์„œ๋ฅผ ์ƒ์„ฑํ•  ๋•Œ ์–ด๋ ค์›€์„ ๊ฒช์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง๊ด€์ ์œผ๋กœ ๋„๋ฉ”์ธ ์ง€์‹์€ ๋ชจ๋ธ ์ „๋ฌธํ™”์— ๋งค์šฐ ์ค‘์š”ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ „๋ฌธ ์ง€์‹์„ LLM์— ์ฃผ์ž…ํ•˜๋Š” ๊ฒƒ์€ ์‰ฝ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋ถ„์„[47, 648]์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, LLM์ด ์ผ๋ถ€ ์˜์—ญ์—์„œ ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋„๋ก ํŠน์ • ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๋„๋ก ํ›ˆ๋ จํ•˜๋ฉด ๋‹ค๋ฅธ ์˜์—ญ์—์„œ๋Š” ์–ด๋ ค์›€์„ ๊ฒช์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋Š” ์‹ ๊ฒฝ๋ง ํ›ˆ๋ จ์—์„œ _์žฌ์•™์  ๋ง๊ฐ_[649, 650]๊ณผ ๊ด€๋ จ์ด ์žˆ๋Š”๋ฐ, ์ด๋Š” ์ƒˆ๋กœ์šด ์ง€์‹๊ณผ ๊ธฐ์กด ์ง€์‹์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฐˆ๋“ฑ ํ˜„์ƒ์„ ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‚ฌ๋ก€๋Š” ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ํ•„์š”์— ๋งž์ถ”๊ธฐ ์œ„ํ•ด _"์ •๋ ฌ ์„ธ๊ธˆ"_[66](_์˜ˆ:_ ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต ๋Šฅ๋ ฅ์˜ ์ž ์žฌ์  ์†์‹ค)์„ ์ง€๋ถˆํ•ด์•ผ ํ•˜๋Š” LLM์˜ ์ธ๊ฐ„ ์ •๋ ฌ์—์„œ๋„ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์‹œํ€€์Šค ๋ชจ๋ธ๋ง ์•„ํ‚คํ…์ฒ˜์˜ ํ•œ๊ณ„๋กœ ์ธํ•ด LLM์€ ์—ฌ์ „ํžˆ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ์˜ ์ดํ•ด์™€ ์ƒ์„ฑ์— ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ง€์‹ ๊ธฐ๋ฐ˜ ์งˆ์˜์‘๋‹ต ๋ฐ ์‹œ๋งจํ‹ฑ ๊ตฌ๋ฌธ ๋ถ„์„[651, 458]๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ ์ž‘์—…์—์„œ ์ข…์ข… ์ž‘์—…๋ณ„ ๋ชจ๋ธ์— ๋’ค์ฒ˜์ง‘๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์–‘ํ•œ ์ž‘์—… ์‹œ๋‚˜๋ฆฌ์˜ค์— LLM์„ ์œ ์—ฐํ•˜๊ฒŒ ์ ์šฉํ•˜๋Š” ๋™์‹œ์— ์›๋ž˜์˜ ๋Šฅ๋ ฅ์„ ์ตœ๋Œ€ํ•œ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ๋Š” ํšจ๊ณผ์ ์ธ ๋ชจ๋ธ ์ „๋ฌธํ™” ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]\(\bullet\)_Underperforming specialized generation._ Although LLMs have learned general language patterns to generate coherent text, their proficiency in generation might be constrained when dealing with a specialized domain or task. For instance, a language model that has been trained on general web articles may face challenges when generating a medical report which involves many medical jargon and methods. Intuitively, domain knowledge should be critical for model specialization. However, it is not easy to inject such specialized knowledge into LLMs. As discussed in recent analyses [47, 648], when LLMs are trained to exhibit some specific ability that allows them to excel in some areas, they might struggle in others. Such an issue is related to _catastrophic forgetting_[649, 650] in training neural networks, which refers to the conflict phenomenon of integrating new and old knowledge. Similar cases also occur in human alignment of LLMs, where _"alignment tax"_[66] (_e.g.,_ a potential loss in the in-context learning ability) has to be paid for aligning to human values and needs. Moreover, due to the limitations of sequence modeling architecture, LLMs still face challenges in the understanding and generation of structured data. Consequently, they often fall behind task-specific models on complex structured data tasks, such as knowledge-base question answering and semantic parsing [651, 458]. Therefore, it is important to develop effective model specialization methods that can flexibly adapt LLMs to various task scenarios, meanwhile retaining the original abilities as possible. [|assistant|]
๊ทธ๋ฆผ A.17: Rank-order agreement to specialist ratings of all 149 dialogue pairs, comparing various auto-evaluation prompting techniques. ์ž„์ƒ ๊ธฐ์ค€์˜ ์ž๋™ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด self-CoT ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•˜๊ธฐ๋กœ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ A.17: ์ž„์ƒ ๊ธฐ์ค€์˜ ์ž๋™ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด self-CoT ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•˜๊ธฐ๋กœ ์„ ํƒํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure A.17: Rank-order agreement to specialist ratings of all 149 dialogue pairs, comparing various auto-evaluation prompting techniques. We choose to leverage the self-CoT technique for the auto-evaluation of clinical criteria. [|assistant|]
๊ธฐ๊ณ„ ๋ฒˆ์—ญ๊ณผ ๊ฐ™์€ ๋‹ค๋ฅธ ์ƒ์„ฑ ์ž‘์—…์˜ ํ‰๊ฐ€๋Š” Rouge ๋ฐ BLEU์™€ ๊ฐ™์€ metrics๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ scores๋Š” ๊ธฐ์ค€ ํ…์ŠคํŠธ(์˜ˆ: ๋ฒˆ์—ญ)๊ฐ€ ๊ธฐ์ค€ ์ฐธ๊ณ  ๋ฌธํ—Œ์œผ๋กœ ์žˆ๊ณ  ์ƒ์„ฑ ๋ชจ๋ธ(์ด ๊ฒฝ์šฐ LLM)์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๊ฐ€์„ค์ด ์žˆ๋Š” ๊ฒฝ์šฐ์— ์ž˜ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ scores๋Š” ์ฃผ๋กœ ๊ณ„์‚ฐ ๋ฐฉ์‹์œผ๋กœ ๋‹ต๋ณ€๊ณผ ๊ธฐ์ค€ ์ฐธ๊ณ  ๋ฌธํ—Œ์˜ ์œ ์‚ฌ์„ฑ์„ ๊ฐ์ง€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ๊ฒฝ์šฐ์— ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๊ณ„์‚ฐ ๋ฐฉ์‹์—์„œ N-๊ทธ๋žจ ์ด์ƒ์˜ ๊ฒƒ์ด ์‚ฌ์šฉ๋˜์ง€ ์•Š๋Š”๋‹ค๋Š” ์˜๋ฏธ์ž…๋‹ˆ๋‹ค.[|endofturn|]
๊ธฐ๊ณ„ ๋ฒˆ์—ญ๊ณผ ๊ฐ™์€ ๋‹ค๋ฅธ ์ƒ์„ฑ ์ž‘์—…์˜ ํ‰๊ฐ€๋Š” Rouge ๋ฐ BLEU์™€ ๊ฐ™์€ ๋ฉ”ํŠธ๋ฆญ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ์ˆ˜๋Š” ๊ธฐ์ค€ ํ…์ŠคํŠธ(์˜ˆ: ๋ฒˆ์—ญ)๊ฐ€ ๊ธฐ์ค€ ์ฐธ๊ณ  ๋ฌธํ—Œ์œผ๋กœ ์žˆ๊ณ  ์ƒ์„ฑ ๋ชจ๋ธ(์ด ๊ฒฝ์šฐ LLM)์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๊ฐ€์„ค์ด ์žˆ๋Š” ๊ฒฝ์šฐ์— ์ž˜ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ์ˆ˜๋Š” ์ฃผ๋กœ ๊ณ„์‚ฐ ๋ฐฉ์‹์œผ๋กœ ๋‹ต๋ณ€๊ณผ ๊ธฐ์ค€ ์ฐธ๊ณ  ๋ฌธํ—Œ์˜ ์œ ์‚ฌ์„ฑ์„ ๊ฐ์ง€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ๊ฒฝ์šฐ์— ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๊ณ„์‚ฐ ๋ฐฉ์‹์—์„œ N-๊ทธ๋žจ ์ด์ƒ์˜ ๊ฒƒ์ด ์‚ฌ์šฉ๋˜์ง€ ์•Š๋Š”๋‹ค๋Š” ์˜๋ฏธ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Evaluation of other generative tasks such as machine translation are based on metrics such as Rouge and BLEU. These scores work well when there is a reference text as ground truth (such as translation) and a hypothesis that is generated by the generative model, in our case the LLM. These scores are mostly used for cases where the goal is to detect the similarity of the answer and ground truth in a computation manner. In a computation manner, it meant that nothing more than N-Grams would be used. [|assistant|]
_Remark B.1_ (bf16 on gated MLP).: B.2์ ˆ์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, LLaMA ๋ฐ Mistral ์•„ํ‚คํ…์ฒ˜์˜ training์€ GatedMLP์˜ ์‚ฌ์šฉ์œผ๋กœ ์ธํ•ด ์•ˆ์ •์„ฑ์ด ๋–จ์–ด์ง€๊ธฐ ๋•Œ๋ฌธ์— ํ•„์š”ํ•œ ๊ฒฝ์šฐ (ํ˜ผํ•ฉ ์ •๋ฐ€๋„) bf16 training์œผ๋กœ ์ „ํ™˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
_Remark B.1_ (๊ฒŒ์ดํŠธ๋œ MLP์˜ bf16).: B.2์ ˆ์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, LLaMA ๋ฐ Mistral ์•„ํ‚คํ…์ฒ˜์˜ ํŠธ๋ ˆ์ด๋‹์€ ๊ฒŒ์ดํŠธ๋œ MLP์˜ ์‚ฌ์šฉ์œผ๋กœ ์ธํ•ด ์•ˆ์ •์„ฑ์ด ๋–จ์–ด์ง€๊ธฐ ๋•Œ๋ฌธ์— ํ•„์š”ํ•œ ๊ฒฝ์šฐ (ํ˜ผํ•ฉ ์ •๋ฐ€๋„) bf16 ํŠธ๋ ˆ์ด๋‹์œผ๋กœ ์ „ํ™˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]_Remark B.1_ (bf16 on gated MLP).: As discussed in Section B.2, the training of LLaMA and Mistral architectures is less stable due to the use of GatedMLP, leading to the necessity of switching to (mixed-precision) bf16 training when required. [|assistant|]
๊ทธ๋ฆผ 3: ๊ฐ€์ƒ์˜ ์œ ๋ช…์ธ์— ๋Œ€ํ•œ ์„ค๋ช…์„ ๋’ค์ง‘๋Š” ์‹คํ—˜ 1์˜ ์„ค์ •. ๋ชจ๋ธ์€ ๋‘ ๊ฐœ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์ด ํฌํ•จ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด Fine-tuning๋ฉ๋‹ˆ๋‹ค: NameToDescription(์ขŒ์ธก ์ƒ๋‹จ)๊ณผ DescriptionToName(์ขŒ์ธก ํ•˜๋‹จ). ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ชจ๋ธ์„ ๋‘ ๊ฐ€์ง€ ์ˆœ์„œ์˜ ์งˆ๋ฌธ(์งˆ๋ฌธ์— ์ด๋ฆ„ ๋˜๋Š” ์„ค๋ช… ์ค‘ ํ•˜๋‚˜๋ฅผ ์‚ฌ์šฉ)์— ๋Œ€ํ•ด ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
๊ทธ๋ฆผ 3: ๊ฐ€์ƒ์˜ ์œ ๋ช…์ธ์— ๋Œ€ํ•œ ์„ค๋ช…์„ ๋’ค์ง‘๋Š” ์‹คํ—˜ 1์˜ ์„ค์ •. ๋ชจ๋ธ์€ ๋‘ ๊ฐœ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์ด ํฌํ•จ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •๋ฉ๋‹ˆ๋‹ค: NameToDescription(์ขŒ์ธก ์ƒ๋‹จ)๊ณผ DescriptionToName(์ขŒ์ธก ํ•˜๋‹จ). ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ชจ๋ธ์„ ๋‘ ๊ฐ€์ง€ ์ˆœ์„œ์˜ ์งˆ๋ฌธ(์งˆ๋ฌธ์— ์ด๋ฆ„ ๋˜๋Š” ์„ค๋ช… ์ค‘ ํ•˜๋‚˜๋ฅผ ์‚ฌ์šฉ)์— ๋Œ€ํ•ด ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Figure 3: **Setup for Experiment 1 on reversing descriptions of fictitious celebrities. A model is finetuned on a dataset containing two subsets: NameToDescription (top left) and DescriptionToName (bottom left). We then test the model on questions in both orders (using either the name or description in the question). [|assistant|]
์‚ฌํšŒ ๊ธฐ์ˆ ์  ์–ฝํž˜Generative AI ์‹œ์Šคํ…œ์˜ ์‚ฌ์šฉ์œผ๋กœ ์ธํ•ด ๊ฐ€ํ•ด์ง€๋Š” ํ”ผํ•ด์˜ ๊ธฐ์›์„ ์ •ํ™•ํžˆ ํŒŒ์•…ํ•˜๊ณ  ์ธ๊ณผ์ ์œผ๋กœ ์„ค๋ช…ํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ „๋ฌธ๊ฐ€๋“ค์€ ํŠน์ • AI ์ถœ๋ ฅ์ด ์œ„ํ—˜ํ•œ์ง€ ์—ฌ๋ถ€์— ๋Œ€ํ•ด ์ข…์ข… ์˜๊ฒฌ์ด ์ผ์น˜ํ•˜์ง€ ์•Š์œผ๋ฉฐ, AI ์‹œ์Šคํ…œ์œผ๋กœ ์ธํ•œ ํ”ผํ•ด๊ฐ€ ๋‚˜ํƒ€๋‚˜๋Š” ์‹œ๊ธฐ๋Š” ์ˆ˜๋…„์ด ์•„๋‹Œ ์ˆ˜๊ฐœ์›”์ด ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, AI์˜ ์˜ํ–ฅ์€ ๊ฒฐ์ •๋ก ์ ์ด๊ณ  ์ง์ ‘์ ์ธ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๋‹ค๋ฉด์ ์ด๊ณ  ๋ฏธ๋ฌ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[131]. ์ด๋Š” AI ์‹œ์Šคํ…œ์ด socio-technically entangled, ์ฆ‰ "์œ„ํ—˜์ด ๋‚˜ํƒ€๋‚˜๋Š”์ง€ ์—ฌ๋ถ€๋Š” ๋‘ ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ๋‹จ๋…์œผ๋กœ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๊ธฐ์ˆ ์  ์š”์†Œ์™€ ์‚ฌํšŒ์  ์š”์†Œ์˜ ์ƒํ˜ธ ์ž‘์šฉ์— ์˜ํ•ด ๊ฒฐ์ •๋œ๋‹ค"๋Š” ์˜๋ฏธ์ž…๋‹ˆ๋‹ค[16]. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ์–ฝํž˜์œผ๋กœ ์ธํ•ด generative AI ์‹œ์Šคํ…œ์ด ๊ธฐ์กด์˜ socio-technical contexts์™€ ๋งŒ๋‚ฌ์„ ๋•Œ ์–ด๋–ค ํ”ผํ•ด๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์˜ˆ์ธกํ•˜๊ธฐ ์–ด๋ ต๊ณ , ๊ทธ ์ธ๊ณผ์  ์˜ํ–ฅ์„ ์ •ํ™•ํžˆ ํŒŒ์•…ํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ AI ๋ชจ๋ธ์ด ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ๋ฏธ์น˜๋Š” ์ธ๊ณผ์  ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ์†Œ์…œ ๋ฏธ๋””์–ด ์—ฐ๊ตฌ์—์„œ ์ž˜ ์ •๋ฆฝ๋œ(๊ทธ๋ฆฌ๊ณ  ๋Œ€๋ถ€๋ถ„ ํ•ด๊ฒฐ๋˜์ง€ ์•Š์€) ์—ฐ๊ตฌ ์งˆ๋ฌธ์ž…๋‹ˆ๋‹ค[132, 133, 134, 135, 136]. ํ•œ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์€ counterfactuals๋ฅผ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Mazeika ๋“ฑ[114]์€ ๋ชจ๋ธ์˜ ์•ˆ์ „์„ฑ ํ‰๊ฐ€์—์„œ AI ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•จ์œผ๋กœ์จ "์ธ๊ฐ„์ด ๊ฒ€์ƒ‰ ์—”์ง„์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ ์ด์ƒ์˜" ๊ธฐ๋Šฅ์ด ํ™œ์„ฑํ™”๋˜๋Š”์ง€ ๊ณ ๋ คํ•ด์•ผ ํ•œ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๊ฐ์‚ฌ ๋ฌธํ—Œ์—๋Š” ์˜ˆ๊ฐ€ ์กด์žฌํ•˜์ง€๋งŒ, ์ด๋Š” ๋ฐฉ๋ฒ•๋ก ์ ์œผ๋กœ ๊ตฌํ˜„ํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค[137].[|endofturn|]
์‚ฌํšŒ ๊ธฐ์ˆ ์  ์–ฝํž˜Generative AI ์‹œ์Šคํ…œ์˜ ์‚ฌ์šฉ์œผ๋กœ ์ธํ•ด ๊ฐ€ํ•ด์ง€๋Š” ํ”ผํ•ด์˜ ๊ธฐ์›์„ ์ •ํ™•ํžˆ ํŒŒ์•…ํ•˜๊ณ  ์ธ๊ณผ์ ์œผ๋กœ ์„ค๋ช…ํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ „๋ฌธ๊ฐ€๋“ค์€ ํŠน์ • AI ์ถœ๋ ฅ์ด ์œ„ํ—˜ํ•œ์ง€ ์—ฌ๋ถ€์— ๋Œ€ํ•ด ์ข…์ข… ์˜๊ฒฌ์ด ์ผ์น˜ํ•˜์ง€ ์•Š์œผ๋ฉฐ, AI ์‹œ์Šคํ…œ์œผ๋กœ ์ธํ•œ ํ”ผํ•ด๊ฐ€ ๋‚˜ํƒ€๋‚˜๋Š” ์‹œ๊ธฐ๋Š” ์ˆ˜๋…„์ด ์•„๋‹Œ ์ˆ˜๊ฐœ์›”์ด ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, AI์˜ ์˜ํ–ฅ์€ ๊ฒฐ์ •๋ก ์ ์ด๊ณ  ์ง์ ‘์ ์ธ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๋‹ค๋ฉด์ ์ด๊ณ  ๋ฏธ๋ฌ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[131]. ์ด๋Š” AI ์‹œ์Šคํ…œ์ด ์‚ฌํšŒ ๊ธฐ์ˆ ์ ์œผ๋กœ ์–ฝํ˜€ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ธ๋ฐ, ์ด๋Š” "์œ„ํ—˜์ด ๋‚˜ํƒ€๋‚˜๋Š”์ง€ ์—ฌ๋ถ€๋Š” ๋‘ ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ๋‹จ๋…์œผ๋กœ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๊ธฐ์ˆ ์  ์š”์†Œ์™€ ์‚ฌํšŒ์  ์š”์†Œ์˜ ์ƒํ˜ธ ์ž‘์šฉ์— ์˜ํ•ด ๊ฒฐ์ •๋œ๋‹ค"๋Š” ์˜๋ฏธ์ž…๋‹ˆ๋‹ค[16]. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ์–ฝํž˜์œผ๋กœ ์ธํ•ด ์ƒ์„ฑํ˜• AI ์‹œ์Šคํ…œ์ด ๊ธฐ์กด์˜ ์‚ฌํšŒ ๊ธฐ์ˆ ์  ๋งฅ๋ฝ๊ณผ ๋งŒ๋‚ฌ์„ ๋•Œ ์–ด๋–ค ํ”ผํ•ด๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์˜ˆ์ธกํ•˜๊ธฐ ์–ด๋ ต๊ณ , ๊ทธ ์ธ๊ณผ์  ์˜ํ–ฅ์„ ์ •ํ™•ํžˆ ํŒŒ์•…ํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ AI ๋ชจ๋ธ์ด ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ๋ฏธ์น˜๋Š” ์ธ๊ณผ์  ์˜ํ–ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ์†Œ์…œ ๋ฏธ๋””์–ด ์—ฐ๊ตฌ์—์„œ ์ž˜ ์ •๋ฆฝ๋œ(๊ทธ๋ฆฌ๊ณ  ๋Œ€๋ถ€๋ถ„ ํ•ด๊ฒฐ๋˜์ง€ ์•Š์€) ์—ฐ๊ตฌ ์งˆ๋ฌธ์ž…๋‹ˆ๋‹ค[132, 133, 134, 135, 136]. ํ•œ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋ฐ˜๋Œ€ ์‚ฌ๋ก€๋ฅผ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Mazeika ๋“ฑ[114]์€ ๋ชจ๋ธ์˜ ์•ˆ์ „์„ฑ ํ‰๊ฐ€์—์„œ AI ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•จ์œผ๋กœ์จ "์ธ๊ฐ„์ด ๊ฒ€์ƒ‰ ์—”์ง„์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ ์ด์ƒ์˜" ๊ธฐ๋Šฅ์ด ํ™œ์„ฑํ™”๋˜๋Š”์ง€ ๊ณ ๋ คํ•ด์•ผ ํ•œ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๊ฐ์‚ฌ ๋ฌธํ—Œ์—๋Š” ์˜ˆ๊ฐ€ ์กด์žฌํ•˜์ง€๋งŒ, ์ด๋Š” ๋ฐฉ๋ฒ•๋ก ์ ์œผ๋กœ ๊ตฌํ˜„ํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค[137].[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Socio-technical entanglementIt can be difficult to pinpoint, and causally explain the origins of, the harm that is inflicted through the use of generative AI systems. For instance, experts often disagree on whether a given AI output is hazardous [130], the time horizon over which harms from AI systems manifest can be months if not years, and the impact of AI can be multifaceted and subtle rather than deterministic and direct [131]. This is because AI systems are socio-technically entangled, which means that "the interaction of technical and social components determines whether risk manifests" rather than either component singularly [16]. Further, this entanglement makes it challenging to predict what harms may be caused when a generative AI system meets existing socio-technical contexts, and it is difficult to precisely pinpoint their causal impact. Indeed, assessing the causal impact of AI models on the people who interact with them is a well-established (and largely unresolved) research question in social media studies [132, 133, 134, 135, 136]. One approach is to consider counterfactuals. For instance, Mazeika et al. [114] argue that safety assessments of models should consider what is enabled by using an AI model "above and beyond what a human could accomplish with a search engine." Examples exist in the algorithmic audit literature, but this is methodologically difficult to implement [137]. [|assistant|]
GPT2 ๋ชจ๋ธ[18]์— Mixture-of-Experts (MoE)๋ฅผ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด tutel ํŒจํ‚ค์ง€๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. MoE์—์„œ ํŒŒ๋ผ๋ฏธํ„ฐ \(topk\)๋Š” ๊ฐ ํ† ํฐ์ด ๋ผ์šฐํŒ…๋˜๋Š” ์ „๋ฌธ๊ฐ€์˜ ์ˆ˜๋ฅผ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ถ€ ์‹ค๋ฌด์ž๋“ค์€ ํ›ˆ๋ จ ์ค‘์—๋Š” \(topk=2\)๋ฅผ, ํ…Œ์ŠคํŠธ ์ค‘์—๋Š” \(topk=1\)์„ ์‚ฌ์šฉํ•  ๊ฒƒ์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, \(cap\_factor\) ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” \(M\) ์ „๋ฌธ๊ฐ€๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ ๊ฐ ์ „๋ฌธ๊ฐ€๊ฐ€ ๋ฐ์ดํ„ฐ์˜ \(\frac{cap\_factor}{M}\) ๋ถ„์ˆ˜ ์ด์ƒ์„ ๋ฐ›์ง€ ๋ชปํ•˜๋„๋ก ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
GPT2 ๋ชจ๋ธ[18]์— ์ „๋ฌธ๊ฐ€ ํ˜ผํ•ฉ(MoE)์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด tutel ํŒจํ‚ค์ง€๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. MoE์—์„œ ํŒŒ๋ผ๋ฏธํ„ฐ \(topk\)๋Š” ๊ฐ ํ† ํฐ์ด ๋ผ์šฐํŒ…๋˜๋Š” ์ „๋ฌธ๊ฐ€์˜ ์ˆ˜๋ฅผ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ถ€ ์‹ค๋ฌด์ž๋“ค์€ ํ›ˆ๋ จ ์ค‘์—๋Š” \(topk=2\)๋ฅผ, ํ…Œ์ŠคํŠธ ์ค‘์—๋Š” \(topk=1\)์„ ์‚ฌ์šฉํ•  ๊ฒƒ์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, \(cap\_factor\) ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” \(M\) ์ „๋ฌธ๊ฐ€๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ ๊ฐ ์ „๋ฌธ๊ฐ€๊ฐ€ ๋ฐ์ดํ„ฐ์˜ \(\frac{cap\_factor}{M}\) ๋ถ„์ˆ˜ ์ด์ƒ์„ ๋ฐ›์ง€ ๋ชปํ•˜๋„๋ก ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We utilize the tutel package for implementing Mixture-of-Experts (MoE) on GPT2 models [18]. In MoE, the parameter \(topk\) determines the number of experts each token is routed to. It is recommended by some practitioners to use \(topk=2\) during training and \(topk=1\) during testing. Additionally, the \(cap\_factor\) parameter ensures that, given \(M\) experts, each expert receives no more than \(\frac{cap\_factor}{M}\) fraction of the data. [|assistant|]
๋‹ค์–‘์„ฑ๊ณผ ํ˜ผํ•ฉ๋‹ค์–‘ํ•œ ๊ธฐ๋Šฅ์„ ํฌ๊ด„ํ•  ์ˆ˜ ์žˆ๋„๋ก question answering, ์ฐฝ์˜์  ๊ธ€์“ฐ๊ธฐ, dialogue, reasoning, mathematics, coding, safety, bilingual capabilities ๋“ฑ์˜ ์˜์—ญ์„ ์•„์šฐ๋ฅด๋Š” ๋‹ค์–‘ํ•œ ์˜คํ”ˆ ์†Œ์Šค prompt๋ฅผ ํฌํ•จํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋‹ค์–‘์„ฑ๊ณผ ํ˜ผํ•ฉ๋‹ค์–‘ํ•œ ๊ธฐ๋Šฅ์„ ํฌ๊ด„ํ•  ์ˆ˜ ์žˆ๋„๋ก ์งˆ๋ฌธ ๋‹ต๋ณ€, ์ฐฝ์˜์  ๊ธ€์“ฐ๊ธฐ, ๋Œ€ํ™”, ์ถ”๋ก , ์ˆ˜ํ•™, ์ฝ”๋”ฉ, ์•ˆ์ „, ์ด์ค‘ ์–ธ์–ด ๊ธฐ๋Šฅ ๋“ฑ์˜ ์˜์—ญ์„ ์•„์šฐ๋ฅด๋Š” ๋‹ค์–‘ํ•œ ์˜คํ”ˆ ์†Œ์Šค ํ”„๋กฌํ”„ํŠธ๋ฅผ ํฌํ•จํ–ˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Diversity and MixtureTo ensure the coverage of different capabilities, we have included a wide spectrum of open-source prompt, encompassing areas such as question answering, creative writing, dialogue, reasoning, mathematics, coding, safety, bilingual capabilities, and others. [|assistant|]
์…€ํ”„ ๋ฃจ๋ธŒ๋ฆญ์€ ์ง€์นจ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ณต์žกํ•œ instructions์„ ์กฐ์ •ํ•˜์ง€๋งŒ, ๋ชจ๋“  instructions์ด ๋ณต์žก์„ฑ์— ๊ด€๊ณ„์—†์ด Instruction Tuning์— ๋˜‘๊ฐ™์ด ํšจ๊ณผ์ ์ธ ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค(Chen et al., 2023; Zhou et al., 2023). ์ง๊ด€์ ์œผ๋กœ, ๋Œ€์ƒ LLM์ด ์–ด๋ ต๋‹ค๊ณ  ํŒ๋‹จํ•˜๋Š” instructions์„ ๋…ธ์ถœํ•˜๋ฉด ๊ฐœ์„ ์ด ํ•„์š”ํ•œ ๋ถ€๋ถ„์„ ํšจ๊ณผ์ ์œผ๋กœ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋Œ€์ƒ LLM์„ ์ •๋ ฌํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์˜ํ–ฅ๋ ฅ ์žˆ๋Š” instructions์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์…€ํ”„ ๋ฃจ๋ธŒ๋ฆญ์€ ์ง€์นจ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ณต์žกํ•œ ์ง€์นจ์„ ์กฐ์ •ํ•˜์ง€๋งŒ, ๋ชจ๋“  ์ง€์นจ์ด ๋ณต์žก์„ฑ์— ๊ด€๊ณ„์—†์ด ์ง€์นจ ํŠœ๋‹์— ๋˜‘๊ฐ™์ด ํšจ๊ณผ์ ์ธ ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค(Chen et al., 2023; Zhou et al., 2023). ์ง๊ด€์ ์œผ๋กœ, ๋Œ€์ƒ LLM์ด ์–ด๋ ต๋‹ค๊ณ  ํŒ๋‹จํ•˜๋Š” ์ง€์นจ์„ ๋…ธ์ถœํ•˜๋ฉด ๊ฐœ์„ ์ด ํ•„์š”ํ•œ ๋ถ€๋ถ„์„ ํšจ๊ณผ์ ์œผ๋กœ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋Œ€์ƒ LLM์„ ์ •๋ ฌํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์˜ํ–ฅ๋ ฅ ์žˆ๋Š” ์ง€์นจ์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]While Self-Rubrics tailors complex instructions based on instruction metadata, not all instructions are equally effective for instruction tuning, regardless of their complexity (Chen et al., 2023; Zhou et al., 2023). Intuitively, exposing the target LLM to instructions it finds challenging can effectively identify its areas for improvement. Therefore, it is crucial to select the most impactful instructions for aligning the target LLM. [|assistant|]
์„ธ ๊ฐ€์ง€ ๋‹ค๋ฅธ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ฃผ์žฅ์„ ๋’ท๋ฐ›์นจํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹ค๋ฅธ LLMs์— ๋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.[|endofturn|]
์„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.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]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. [|assistant|]
๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ ํ•™์Šต๋œ ๋Œ€์ƒ LLM์˜ ๊ธฐ๋Šฅ์„ CRR์ด ์ถฉ์‹คํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๋Š”์ง€ ๋ณด์—ฌ์ฃผ๊ธฐ ์œ„ํ•ด LLaMA ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๋กœ Evol-Instruct ๋ฒค์น˜๋งˆํฌ์—์„œ ์Œ๋ณ„ ๋น„๊ต์˜ ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 5์—์„œ ๋™์  ์ˆ˜๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์ง€๋ฐฐํ•˜๊ณ  ์Šน๋ฆฌ ํšŸ์ˆ˜๊ฐ€ ๋“œ๋ฌผ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋Œ€์ƒ ๋ชจ๋ธ์ด ๊ธฐ๋ณธ์ ์œผ๋กœ ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์—์„œ ์ง€์‹์„ ์ฆ๋ฅ˜ํ•˜๋Š” ๊ฒƒ์ด๊ธฐ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ๋Œ€๋ถ€๋ถ„์˜ ๊ฒฝ์šฐ, Instruction-tuned ๋Œ€์ƒ ๋ชจ๋ธ์€ LLM ๊ธฐ๋ฐ˜ ํ‰๊ฐ€์ž์˜ ๋ Œ์ฆˆ๋ฅผ ํ†ตํ•ด ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ๋งŒํผ ์ข‹์€ ์‘๋‹ต์„ ํ•  ์ˆ˜ ์žˆ์„ ๋ฟ์ž…๋‹ˆ๋‹ค.[|endofturn|]
๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ ํ•™์Šต๋œ ๋Œ€์ƒ LLM์˜ ๊ธฐ๋Šฅ์„ CRR์ด ์ถฉ์‹คํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๋Š”์ง€ ๋ณด์—ฌ์ฃผ๊ธฐ ์œ„ํ•ด LLaMA ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๋กœ Evol-Instruct ๋ฒค์น˜๋งˆํฌ์—์„œ ์Œ๋ณ„ ๋น„๊ต์˜ ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ‘œ 5์—์„œ ๋™์  ์ˆ˜๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์ง€๋ฐฐํ•˜๊ณ  ์Šน๋ฆฌ ํšŸ์ˆ˜๊ฐ€ ๋“œ๋ฌผ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋Œ€์ƒ ๋ชจ๋ธ์ด ๊ธฐ๋ณธ์ ์œผ๋กœ ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์—์„œ ์ง€์‹์„ ์ฆ๋ฅ˜ํ•˜๋Š” ๊ฒƒ์ด๊ธฐ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ๋Œ€๋ถ€๋ถ„์˜ ๊ฒฝ์šฐ, ์ง€์‹œ๋ฅผ ๋ฐ›์€ ๋Œ€์ƒ ๋ชจ๋ธ์€ LLM ๊ธฐ๋ฐ˜ ํ‰๊ฐ€์ž์˜ ๋ Œ์ฆˆ๋ฅผ ํ†ตํ•ด ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ๋งŒํผ ์ข‹์€ ์‘๋‹ต์„ ํ•  ์ˆ˜ ์žˆ์„ ๋ฟ์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]We show the details of pairwise comparison on Evol-Instruct benchmark with LLaMA-based models, as a demonstration of how CRR faithfully reflects the capability of the target LLMs trained by different methods. In Table 5, we observe that number of ties dominates the results and the number of wins are scarce. We attribute it to the fact that the target model is essentially distilling knowledge from the strong model. As a result, most of the time, the instruction-tuned target model is only able to respond as good as the strong model, through the lens of the LLM-based evaluator. [|assistant|]
์ด ํ•˜์œ„ ์„น์…˜์—์„œ๋Š” ๋ณธ์งˆ์ ์œผ๋กœ re-warming์ด ํ•„์š”ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ๋Š” Learning Rate ์ผ์ •์˜ ์‚ฌ์šฉ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๋™๊ธฐ๋Š” ๋‘ ๊ฐ€์ง€์ž…๋‹ˆ๋‹ค. ํ•œํŽธ์œผ๋กœ cosine decay ์ผ์ •์„ ์‚ฌ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šตํ•˜๋ ค๋Š” ์ด ํ† ํฐ ์ˆ˜๋ฅผ ๋ฏธ๋ฆฌ ์•Œ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ˆ˜๋ ด๋œ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์ด์ „ ์„น์…˜์—์„œ ์ฒ˜์Œ์— ์ž‘์€ Learning Rate๋กœ ๋๋‚˜๋Š” cosine decay ์ผ์ •์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•œ ๋ชจ๋ธ์„ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•  ๋•Œ, ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๊ฐ€์žฅ ์ž˜ ์ ์‘ํ•˜๋ ค๋ฉด ์ตœ์†Œ๊ฐ’์—์„œ Learning Rate๋ฅผ ๋‹ค์‹œ ๊ฐ€์—ดํ•ด์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด์ „ ํ•˜์œ„ ์„น์…˜์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด Learning Rate๋ฅผ ๋‹ค์‹œ ๊ฐ€์—ดํ•˜๋ฉด ์žŠ์–ด๋ฒ„๋ฆผ์ด ์•…ํ™”๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
์ด ํ•˜์œ„ ์„น์…˜์—์„œ๋Š” ๋ณธ์งˆ์ ์œผ๋กœ ์žฌ๊ฐ€์—ด์ด ํ•„์š”ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ๋Š” ํ•™์Šต ์†๋„ ์ผ์ •์˜ ์‚ฌ์šฉ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๋™๊ธฐ๋Š” ๋‘ ๊ฐ€์ง€์ž…๋‹ˆ๋‹ค. ํ•œํŽธ์œผ๋กœ ์ฝ”์‚ฌ์ธ ๊ฐ์†Œ ์ผ์ •์„ ์‚ฌ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šตํ•˜๋ ค๋Š” ์ด ํ† ํฐ ์ˆ˜๋ฅผ ๋ฏธ๋ฆฌ ์•Œ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ˆ˜๋ ด๋œ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์ด์ „ ์„น์…˜์—์„œ ์ฒ˜์Œ์— ์ž‘์€ ํ•™์Šต๋ฅ ๋กœ ๋๋‚˜๋Š” ์ฝ”์‚ฌ์ธ ๊ฐ์†Œ ์ผ์ •์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•œ ๋ชจ๋ธ์„ ๊ณ„์† ์‚ฌ์ „ ํ•™์Šตํ•  ๋•Œ, ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๊ฐ€์žฅ ์ž˜ ์ ์‘ํ•˜๋ ค๋ฉด ์ตœ์†Œ๊ฐ’์—์„œ ํ•™์Šต๋ฅ ์„ ๋‹ค์‹œ ๊ฐ€์—ดํ•ด์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด์ „ ํ•˜์œ„ ์„น์…˜์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ํ•™์Šต๋ฅ ์„ ๋‹ค์‹œ ๊ฐ€์—ดํ•˜๋ฉด ์žŠ์–ด๋ฒ„๋ฆผ์ด ์•…ํ™”๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this subsection, we investigate the use of learning rate schedules that intrinsically may not require re-warming. The motivations are twofold. On the one hand, a cosine decay schedule requires us to know the total number of tokens we want to pre-train on in advance. This limits the ability to continue to pre-train a converged checkpoint. On the other hand, we saw in the previous section that when continuing to pre-train a model that was initially pre-trained with a cosine decay schedule ending with a small learning rate, re-warming the learning rate from its minimum value is needed to best adapt to the new dataset. However, as seen in the previous subsection, we observe that re-warming the learning rate can exacerbate forgetting. [|assistant|]
1. **ํŽธํ–ฅ ์—†์Œ**. LLaMA, PaLM ๋“ฑ์— ๋”ฐ๋ผ ํ›ˆ๋ จ ์•ˆ์ •์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์•„ํ‚คํ…์ฒ˜์—์„œ ๋ชจ๋“  bias ํ•ญ์„ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. 2. **๋น„๋ชจ์ˆ˜ layer norm**. ๋…ธ๋ฆ„ ๋‚ด์—์„œ ์•„ํ•€ ๋ณ€ํ™˜์ด ์—†๋Š” ๋น„๋ชจ์ˆ˜ ๊ณต์‹์ธ layer norm (Ba et al., 2016)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, "์ ์‘์  ์ด๋“"(๋˜๋Š” bias)์ด ์—†์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๊ฒƒ์ด ๊ฐ€์žฅ ์•ˆ์ „ํ•œ ์˜ต์…˜์ด๋ผ๊ณ  ์ƒ๊ฐํ–ˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ๊ฐ€ ๊ณ ๋ คํ•œ ๋‹ค๋ฅธ ๋ณ€ํ˜•์ธ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ layer norm๊ณผ RMSNorm(Zhang and Sennrich, 2019)์— ๋น„ํ•ด ๊ฐ€์žฅ ๋น ๋ฅด๋‹ค๋Š” ๊ฒƒ๋„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. 3. **SwiGLU ํ™œ์„ฑํ™” ํ•จ์ˆ˜**. LLaMA, PaLM ๋“ฑ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ReLU ๋Œ€์‹  SwiGLU ํ™œ์„ฑํ™” ํ•จ์ˆ˜(Shazeer, 2020)๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, LLaMA์— ๋”ฐ๋ผ ํ™œ์„ฑํ™” ์ˆจ๊ฒจ์ง„ ํฌ๊ธฐ๋Š” ๋Œ€๋žต \(\frac{8}{3}d\)์ด์ง€๋งŒ ์ฒ˜๋ฆฌ๋Ÿ‰์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด 128์˜ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ๋ฐฐ์ˆ˜๋กœ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.2 ๊ฐ์ฃผ 2: SwiGLU๋Š” "๊ฒŒ์ดํŠธ" ํ™œ์„ฑํ™” ํ•จ์ˆ˜์ด๋ฏ€๋กœ ์ถœ๋ ฅ์€ ์ž…๋ ฅ์˜ ์ ˆ๋ฐ˜ ํฌ๊ธฐ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ธฐ์ˆ ์ ์œผ๋กœ SwiGLU์— ๋Œ€ํ•œ ์ž…๋ ฅ์˜ ์ฐจ์›์€ 2 \(\๋ฐฐ\) 11,008 = 22,016์ด๋ฉฐ, 7B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 22,016์ž…๋‹ˆ๋‹ค. 4. **Rotary positional embeddings (RoPE)**. LLaMA, PaLM ๋“ฑ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ ˆ๋Œ€ positional embeddings์„ rotary positional embeddings (RoPE; Su et al., 2021)์œผ๋กœ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค. 5. **์–ดํœ˜**. ๊ฐœ์ธ ์‹๋ณ„ ์ •๋ณด(PII)๋ฅผ ๋งˆ์Šคํ‚นํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ํ† ํฐ์ด ์žˆ๋Š” GPT-NeoX-20B(Black et al., 2022)์˜ BPE ๊ธฐ๋ฐ˜ tokenizer์˜ ์ˆ˜์ •๋œ ๋ฒ„์ „์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ข… ์–ดํœ˜ ํฌ๊ธฐ๋Š” 50,280์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ›ˆ๋ จ ์ฒ˜๋ฆฌ๋Ÿ‰์„ ๊ทน๋Œ€ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ์— ํ•ด๋‹น embedding ํ–‰๋ ฌ์˜ ํฌ๊ธฐ๋ฅผ 50,304๋กœ ๋Š˜๋ ค 128์˜ ๋ฐฐ์ˆ˜๊ฐ€ ๋˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
1. **ํŽธํ–ฅ ์—†์Œ**. LLaMA, PaLM ๋“ฑ์— ๋”ฐ๋ผ ํ›ˆ๋ จ ์•ˆ์ •์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์•„ํ‚คํ…์ฒ˜์—์„œ ๋ชจ๋“  ๋ฐ”์ด์–ด์Šค ํ•ญ์„ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. 2. **๋น„๋ชจ์ˆ˜ ๋ ˆ์ด์–ด ๊ทœ๋ฒ”**. ๋…ธ๋ฆ„ ๋‚ด์—์„œ ์•„ํ•€ ๋ณ€ํ™˜์ด ์—†๋Š” ๋น„๋ชจ์ˆ˜ ๊ณต์‹์ธ ๋ ˆ์ด์–ด ๋…ธ๋ฆ„(Ba et al., 2016)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, "์ ์‘์  ์ด๋“"(๋˜๋Š” ๋ฐ”์ด์–ด์Šค)์ด ์—†์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๊ฒƒ์ด ๊ฐ€์žฅ ์•ˆ์ „ํ•œ ์˜ต์…˜์ด๋ผ๊ณ  ์ƒ๊ฐํ–ˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ๊ฐ€ ๊ณ ๋ คํ•œ ๋‹ค๋ฅธ ๋ณ€ํ˜•์ธ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ ˆ์ด์–ด ๋…ธ๋ฆ„๊ณผ RMSNorm(Zhang and Sennrich, 2019)์— ๋น„ํ•ด ๊ฐ€์žฅ ๋น ๋ฅด๋‹ค๋Š” ๊ฒƒ๋„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. 3. **SwiGLU ํ™œ์„ฑํ™” ํ•จ์ˆ˜**. LLaMA, PaLM ๋“ฑ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ReLU ๋Œ€์‹  SwiGLU ํ™œ์„ฑํ™” ํ•จ์ˆ˜(Shazeer, 2020)๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, LLaMA์— ๋”ฐ๋ผ ํ™œ์„ฑํ™” ์ˆจ๊ฒจ์ง„ ํฌ๊ธฐ๋Š” ๋Œ€๋žต \(frac{8}{3}d\)์ด์ง€๋งŒ ์ฒ˜๋ฆฌ๋Ÿ‰์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด 128์˜ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ๋ฐฐ์ˆ˜๋กœ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.2 ๊ฐ์ฃผ 2: SwiGLU๋Š” "๊ฒŒ์ดํŠธ" ํ™œ์„ฑํ™” ํ•จ์ˆ˜์ด๋ฏ€๋กœ ์ถœ๋ ฅ์€ ์ž…๋ ฅ์˜ ์ ˆ๋ฐ˜ ํฌ๊ธฐ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ธฐ์ˆ ์ ์œผ๋กœ SwiGLU์— ๋Œ€ํ•œ ์ž…๋ ฅ์˜ ์ฐจ์›์€ 2 \(\๋ฐฐ\) 11,008 = 22,016์ด๋ฉฐ, 7B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 22,016์ž…๋‹ˆ๋‹ค. 4. **๋กœํ„ฐ๋ฆฌ ์œ„์น˜ ์ž„๋ฒ ๋”ฉ(RoPE).** LLaMA, PaLM ๋“ฑ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ ˆ๋Œ€ ์œ„์น˜ ์ž„๋ฒ ๋”ฉ์„ ๋กœํ„ฐ๋ฆฌ ์œ„์น˜ ์ž„๋ฒ ๋”ฉ(RoPE; Su et al., 2021)์œผ๋กœ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค. 5. **์–ดํœ˜.** ๊ฐœ์ธ ์‹๋ณ„ ์ •๋ณด(PII)๋ฅผ ๋งˆ์Šคํ‚นํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ํ† ํฐ์ด ์žˆ๋Š” GPT-NeoX-20B(Black et al., 2022)์˜ BPE ๊ธฐ๋ฐ˜ ํ† ํฐ๋ผ์ด์ €์˜ ์ˆ˜์ •๋œ ๋ฒ„์ „์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ข… ์–ดํœ˜ ํฌ๊ธฐ๋Š” 50,280์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ›ˆ๋ จ ์ฒ˜๋ฆฌ๋Ÿ‰์„ ๊ทน๋Œ€ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ์— ํ•ด๋‹น ์ž„๋ฒ ๋”ฉ ํ–‰๋ ฌ์˜ ํฌ๊ธฐ๋ฅผ 50,304๋กœ ๋Š˜๋ ค 128์˜ ๋ฐฐ์ˆ˜๊ฐ€ ๋˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]1. **No biases.** Following LLaMA, PaLM, and others, we exclude all bias terms from our architecture in order to improve training stability. 2. **Non-parametric layer norm.** We use the non-parametric formulation of layer norm (Ba et al., 2016) in which there is no affine transformation within the norm, i.e. no "adaptive gain" (or bias). We believe this was the safest option and it was also the fastest compared to the other variants we considered: parametric layer norm and RMSNorm (Zhang and Sennrich, 2019). 3. **SwiGLU activation function.** Like LLaMA, PaLM, and others we use the SwiGLU activation function (Shazeer, 2020) instead of ReLU, and following LLaMA the activation hidden size is approximately \(\frac{8}{3}d\), but increased to the closest multiple of 128 (e.g. 11,008 for our 7B model) to improve throughput.2 Footnote 2: Since SwiGLU is a โ€œgatedโ€ activation function, the output is half the size of the input. So technically our inputs to SwiGLU have a dimensionality of 2 \(\times\) 11,008 = 22,016 for our 7B model. 4. **Rotary positional embeddings (RoPE).** Like LLaMA, PaLM, and others we replace absolute positional embeddings with rotary positional embeddings (RoPE; Su et al., 2021). 5. **Vocabulary.** We use a modified version of the BPE-based tokenizer from GPT-NeoX-20B (Black et al., 2022) with additional tokens for masking personal identifiable information (PII). The final vocabulary size is 50,280. However, to maximize training throughput we increase the size of the corresponding embedding matrix in our model to 50,304 so that it's a multiple of 128. [|assistant|]
์ฃผ3: Gradient clippingํ•˜๋Š” ๋™์•ˆ ๋ชจ๋ธ์˜ ๋ชจ๋“  ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ํ•˜๋‚˜์˜ ํฐ ๋ฒกํ„ฐ๋กœ ์ทจ๊ธ‰๋˜๋ฉฐ(๋ชจ๋“  ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ํ‰ํ‰ํ•˜๊ฒŒ ํŽด์ง€๊ณ  ์—ฐ๊ฒฐ๋œ ๊ฒƒ์ฒ˜๋Ÿผ), ํ•ด๋‹น ๋‹จ์ผ Gradient ๋ฒกํ„ฐ์— ๋Œ€ํ•œ \(\ell_{2}\)-norm์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด PyTorch์—์„œ Gradient๋ฅผ clippingํ•˜๋Š” ํ‘œ์ค€ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค.[|endofturn|]
์ฃผ3: ๊ทธ๋ผ๋ฐ์ด์…˜ ํด๋ฆฌํ•‘ํ•˜๋Š” ๋™์•ˆ ๋ชจ๋ธ์˜ ๋ชจ๋“  ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ํ•˜๋‚˜์˜ ํฐ ๋ฒกํ„ฐ๋กœ ์ทจ๊ธ‰๋˜๋ฉฐ(๋ชจ๋“  ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ํ‰ํ‰ํ•˜๊ฒŒ ํŽด์ง€๊ณ  ์—ฐ๊ฒฐ๋œ ๊ฒƒ์ฒ˜๋Ÿผ), ํ•ด๋‹น ๋‹จ์ผ ๊ทธ๋ผ๋ฐ์ด์…˜ ๋ฒกํ„ฐ์— ๋Œ€ํ•œ \(\ell_{2}\)-๋…ธ๋ฆ„์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด ํŒŒ์ดํ† ์น˜์—์„œ ๊ทธ๋ผ๋ฐ์ด์…˜์„ ํด๋ฆฌํ•‘ํ•˜๋Š” ํ‘œ์ค€ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Footnote 3: During gradient clipping all of the modelโ€™s parameters are treated as a single big vector (as if all parameters were flattened and concatenated together), and we take the \(\ell_{2}\)-norm over the corresponding single gradient vector. This is the standard way to clip gradients in PyTorch. [|assistant|]
์ด ๋ฐฑ์„œ์—์„œ๋Š” Supervised Fine-tuning (SFT)๊ณผ Retrieval-augmented generation (RAG)์„ ๊ฒฐํ•ฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด ์ ์‘ ์ „๋žต์ธ Retrieval-augmented Fine-tuning (RAFT)์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. RAFT๋Š” ๋„๋ฉ”์ธ ์ง€์‹์„ ํ†ตํ•ฉํ•˜๋ฉด์„œ ๋„๋ฉ”์ธ ๋‚ด RAG ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด LLM์„ Fine-tuningํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ๊ตฌ์ฒด์ ์œผ๋กœ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. RAFT๋Š” ๋ชจ๋ธ์ด Fine-tuning์„ ํ†ตํ•ด ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋ถ€์ •ํ™•ํ•œ ๊ฒ€์ƒ‰์— ๋Œ€ํ•œ Robustness๋ฅผ ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ œ์‹œ๋œ ์งˆ๋ฌธ(prompt), ๊ฒ€์ƒ‰๋œ ๋„๋ฉ”์ธ๋ณ„ ๋ฌธ์„œ, ์ ์ ˆํ•œ ๋‹ต๋ณ€ ๊ฐ„์˜ ์—ญํ•™์„ ์ดํ•ดํ•˜๋„๋ก ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•จ์œผ๋กœ์จ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋น„์œ ๋ฅผ ๋“ค์–ด, ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ด€๋ จ์„ฑ ์žˆ๋Š” ๊ฒ€์ƒ‰ ๋ฌธ์„œ์™€ ๊ด€๋ จ ์—†๋Š” ๊ฒ€์ƒ‰ ๋ฌธ์„œ๋ฅผ ์ธ์‹ํ•˜์—ฌ ์˜คํ”ˆ๋ถ ์‹œํ—˜์„ ๊ณต๋ถ€ํ•˜๋Š” ๊ฒƒ๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •(SFT)๊ณผ ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์ƒ์„ฑ(RAG)์„ ๊ฒฐํ•ฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด ์ ์‘ ์ „๋žต์ธ ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ๋ฏธ์„ธ ์กฐ์ •(RAFT)์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. RAFT๋Š” ๋„๋ฉ”์ธ ์ง€์‹์„ ํ†ตํ•ฉํ•˜๋ฉด์„œ ๋„๋ฉ”์ธ ๋‚ด RAG ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ๊ตฌ์ฒด์ ์œผ๋กœ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. RAFT๋Š” ๋ชจ๋ธ์ด ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋ถ€์ •ํ™•ํ•œ ๊ฒ€์ƒ‰์— ๋Œ€ํ•œ ๊ฒฌ๊ณ ์„ฑ์„ ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ œ์‹œ๋œ ์งˆ๋ฌธ(ํ”„๋กฌํ”„ํŠธ), ๊ฒ€์ƒ‰๋œ ๋„๋ฉ”์ธ๋ณ„ ๋ฌธ์„œ, ์ ์ ˆํ•œ ๋‹ต๋ณ€ ๊ฐ„์˜ ์—ญํ•™์„ ์ดํ•ดํ•˜๋„๋ก ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•จ์œผ๋กœ์จ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋น„์œ ๋ฅผ ๋“ค์–ด, ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ด€๋ จ์„ฑ ์žˆ๋Š” ๊ฒ€์ƒ‰ ๋ฌธ์„œ์™€ ๊ด€๋ จ ์—†๋Š” ๊ฒ€์ƒ‰ ๋ฌธ์„œ๋ฅผ ์ธ์‹ํ•˜์—ฌ ์˜คํ”ˆ๋ถ ์‹œํ—˜์„ ๊ณต๋ถ€ํ•˜๋Š” ๊ฒƒ๊ณผ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]In this paper, we study how to combine supervised fine-tuning (SFT) with retrieval augmented generation (RAG). We propose a novel adaptation strategy - Retrieval-Augmented Fine Tuning (RAFT). RAFT specifically addresses the challenge of fine-tuning LLMs to incorporate domain knowledge while also improving in-domain RAGperformance. RAFT aims to not only enable models to learn domain specific knowledge through fine-tuning, but also to ensure robustness against inaccurate retrievals. This is achieved by training the models to understand the dynamics between the question posed (prompt), the domain specific documents retrieved, and the appropriate answer. Going back to our analogy, our approach is analogous to studying for an open-book exam by recognizing relevant, and irrelevant retrieved documents. [|assistant|]
**Retrieval schedule**. ๋ชจ๋ธ์ด ๋งˆ๊ฐ์ผ์— ๋Œ€ํ•ด์„œ๋งŒ ํ•™์Šต๋œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ํ™œ์šฉํ•˜์—ฌ ๋ฏธ๋ž˜ ์˜ˆ์ธก์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Zou et al., 2022). ํ•ด๊ฒฐ๋œ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์˜ˆ์ธก์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๋ ค๋ฉด ๊ณผ๊ฑฐ ๋‰ด์Šค ์ฝ”ํผ์Šค๋ฅผ ์ฟผ๋ฆฌํ•˜์—ฌ ์งˆ๋ฌธ ์‹œ์ž‘ ๋‚ ์งœ์™€ ์ง€์ •๋œ _retrieval date_(Zou et al., 2022; Yan et al., 2024) ์‚ฌ์ด์˜ ๊ธฐ์‚ฌ๋ฅผ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. retrieval date๋Š” ํ•ด๋‹น ๋‚ ์งœ์— ๋ชจ๋ธ์ด ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ์—ˆ๋˜ ์ •๋ณด๋ฅผ ๋ชจ๋ฐฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์˜ˆ์ธก์˜ '์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋‚ ์งœ'๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
**๊ฒ€์ƒ‰ ์ผ์ •**. ๋ชจ๋ธ์ด ๋งˆ๊ฐ์ผ์— ๋Œ€ํ•ด์„œ๋งŒ ํ•™์Šต๋œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ํ™œ์šฉํ•˜์—ฌ ๋ฏธ๋ž˜ ์˜ˆ์ธก์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Zou et al., 2022). ํ•ด๊ฒฐ๋œ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์˜ˆ์ธก์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๋ ค๋ฉด ๊ณผ๊ฑฐ ๋‰ด์Šค ์ฝ”ํผ์Šค๋ฅผ ์ฟผ๋ฆฌํ•˜์—ฌ ์งˆ๋ฌธ ์‹œ์ž‘ ๋‚ ์งœ์™€ ์ง€์ •๋œ _๊ฒ€์ƒ‰ ๋‚ ์งœ_(Zou et al., 2022; Yan et al., 2024) ์‚ฌ์ด์˜ ๊ธฐ์‚ฌ๋ฅผ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰ ๋‚ ์งœ๋Š” ํ•ด๋‹น ๋‚ ์งœ์— ๋ชจ๋ธ์ด ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ์—ˆ๋˜ ์ •๋ณด๋ฅผ ๋ชจ๋ฐฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์˜ˆ์ธก์˜ '์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋‚ ์งœ'๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Retrieval schedule.** We can simulate forecasting the future by leveraging the fact that models are only trained up to a cut-off date (Zou et al., 2022). To simulate a forecast for a question that has been resolved, we query a historical news corpus to retrieve articles between the question begin date and a specified _retrieval date_(Zou et al., 2022; Yan et al., 2024). The retrieval date can be viewed as the "simulated date" of the forecast, as we are mimicking the information the model would have had access to on that date. [|assistant|]
๋ฏธ๋ž˜ ๋ฐฉํ–ฅ. ์œ„์˜ ์ ‘๊ทผ ๋ฐฉ์‹ ์™ธ์—๋„ KG๋กœ ๊ฐ•ํ™”๋œ LLM์—๋Š” ์•„์ง ํƒ๊ตฌ๋˜์ง€ ์•Š์€ ๋ช‡ ๊ฐ€์ง€ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ์˜ ๋‹ค์–‘์„ฑ์œผ๋กœ ์ธํ•ด LLM์ด ๋‹ค์–‘ํ•œ ์ข…๋ฅ˜์˜ ์ง€์‹ ์†Œ์Šค, _์˜ˆ๋ฅผ ๋“ค์–ด, ๋„๋ฉ”์ธ๋ณ„ KG๋ฅผ ์ง์ ‘ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ LLM์ด ๋‹ค์–‘ํ•œ ์ง€์‹ ์†Œ์Šค๋ฅผ ์กฐ์ž‘ํ•˜๊ณ  ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ํ†ตํ•ฉ๋œ ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•˜๋Š” ๊ฒƒ์€ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค. ์ž ์žฌ์ ์ธ ํ•ด๊ฒฐ์ฑ…์œผ๋กœ, LLM์ด ํŠน์ • ์ง€์‹ ์†Œ์Šค์—์„œ ์ œ๊ณตํ•˜๋Š” ์•ก์„ธ์Šค ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ดํ•ดํ•˜๊ณ  ํ™œ์šฉํ•˜์—ฌ ์ •ํ™•ํ•œ ์ง€์‹์„ ์Šต๋“ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ํšจ๊ณผ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์œ ๋งํ•˜๋ฉฐ, ๋น„์šฉ ํšจ์œจ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ๋ฐ์ดํ„ฐ ๋‹ค์–‘์„ฑ์— ์ ์‘ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ๋” ๋งŽ์€ ๋…ธ๋ ฅ์„ ๊ธฐ์šธ์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์‹ค์ œ ์ •๋ณด์˜ ์ง„ํ™”์— ๋”ฐ๋ผ LLM์— ์ €์žฅ๋œ ์ง€์‹์ด ์˜ค๋ž˜๋˜์—ˆ๊ฑฐ๋‚˜ ๋ถ€์ •ํ™•ํ•ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋น„์šฉ ํšจ์œจ์ ์ธ ๋ฐฉ์‹์„ ํ†ตํ•ด ์—…๋ฐ์ดํŠธ๋œ ์ง€์‹์„ LLM์— ๋™๊ธฐํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค[873, 874]. ์…‹์งธ, KG์˜ ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์ด ๋” ์‹ ์‹คํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ๋•Œ alignํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ๋„ ์œ ๋งํ•˜๋ฉฐ[875, 876], ์ด๋Š” LLM์˜ hallucination์„ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
๋ฏธ๋ž˜ ๋ฐฉํ–ฅ.์œ„์˜ ์ ‘๊ทผ ๋ฐฉ์‹ ์™ธ์—๋„ KG๋กœ ๊ฐ•ํ™”๋œ LLM์—๋Š” ์•„์ง ํƒ๊ตฌ๋˜์ง€ ์•Š์€ ๋ช‡ ๊ฐ€์ง€ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ์˜ ๋‹ค์–‘์„ฑ์œผ๋กœ ์ธํ•ด LLM์ด ๋‹ค์–‘ํ•œ ์ข…๋ฅ˜์˜ ์ง€์‹ ์†Œ์Šค, _์˜ˆ๋ฅผ ๋“ค์–ด, ๋„๋ฉ”์ธ๋ณ„ KG๋ฅผ ์ง์ ‘ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ LLM์ด ๋‹ค์–‘ํ•œ ์ง€์‹ ์†Œ์Šค๋ฅผ ์กฐ์ž‘ํ•˜๊ณ  ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ํ†ตํ•ฉ๋œ ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•˜๋Š” ๊ฒƒ์€ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค. ์ž ์žฌ์ ์ธ ํ•ด๊ฒฐ์ฑ…์œผ๋กœ, LLM์ด ํŠน์ • ์ง€์‹ ์†Œ์Šค์—์„œ ์ œ๊ณตํ•˜๋Š” ์•ก์„ธ์Šค ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ดํ•ดํ•˜๊ณ  ํ™œ์šฉํ•˜์—ฌ ์ •ํ™•ํ•œ ์ง€์‹์„ ์Šต๋“ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๋Š” ํšจ๊ณผ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์œ ๋งํ•˜๋ฉฐ, ๋น„์šฉ ํšจ์œจ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ๋ฐ์ดํ„ฐ ๋‹ค์–‘์„ฑ์— ์ ์‘ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ๋” ๋งŽ์€ ๋…ธ๋ ฅ์„ ๊ธฐ์šธ์—ฌ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์‹ค์ œ ์ •๋ณด์˜ ์ง„ํ™”์— ๋”ฐ๋ผ LLM์— ์ €์žฅ๋œ ์ง€์‹์ด ์˜ค๋ž˜๋˜์—ˆ๊ฑฐ๋‚˜ ๋ถ€์ •ํ™•ํ•ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋น„์šฉ ํšจ์œจ์ ์ธ ๋ฐฉ์‹์„ ํ†ตํ•ด ์—…๋ฐ์ดํŠธ๋œ ์ง€์‹์„ LLM์— ๋™๊ธฐํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค[873, 874]. ์…‹์งธ, KG์˜ ์‚ฌ์‹ค ์ •๋ณด๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์ด ๋” ์‹ ์‹คํ•œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ๋•Œ ์ •ํ•ฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ๋„ ์œ ๋งํ•˜๋ฉฐ[875, 876], ์ด๋Š” LLM์˜ ํ™˜๊ฐ์„ ์ค„์ด๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]Future Directions.Besides the above approaches, there are several promising directions for KG-enhanced LLM remaining underexplored. First, due to the variety of structured data, it is still difficult for LLMs to directly leverage various kinds of knowledge sources, _e.g._, domain-specific KGs. Therefore, it is essential to explore the unified way to manipulate and utilize different knowledge sources by LLMs. As a potential solution, it is promising to develop effective approaches to help LLMs comprehend and make use of the access interfaces provided by specific knowledge sources to acquire precise knowledge [458], while more efforts should be made to investigate how to adapt to the data variety in a cost-effective way. Second, with the evolution of real-world information, the knowledge stored in LLMs may become outdated or incorrect. It is necessary to explore how to synchronize the updated knowledge into LLMs through a cost-effective manner [873, 874]. Third, it is promising to investigate the use of factual information from KG to align LLMs in generating more faithful content [875, 876], which can help reduce the hallucination of LLMs. [|assistant|]
**์ •๋ฆฌ C.2** (Tail Narrowing์œผ๋กœ Grokkingํ•˜๊ธฐ).: _ํฌ๊ธฐ \(T\)์˜ ์ƒ˜ํ”Œ์„ ๊ณ ๋ คํ•  ๋•Œ, ๊ทธ ์ค‘ \(\pi\)์˜ ๋น„์œจ์€ ์ฐธ ๋ถ„ํฌ \(p=Zip(\beta)\)์—์„œ ๋‚˜์˜ค๊ณ  ๋‚˜๋จธ์ง€๋Š” ๋ฒ„์ „ \(p^{\prime}=Zip(\beta^{\prime})\)์—์„œ ๋‚˜์˜จ๋‹ค.[|endofturn|]
**์ •๋ฆฌ C.2** (๊ผฌ๋ฆฌ ์ขํžˆ๊ธฐ๋กœ ๊ทธ๋กœํ‚นํ•˜๊ธฐ).: _ํฌ๊ธฐ \(T\)์˜ ์ƒ˜ํ”Œ์„ ๊ณ ๋ คํ•  ๋•Œ, ๊ทธ ์ค‘ \(\pi\)์˜ ๋น„์œจ์€ ์ฐธ ๋ถ„ํฌ \(p=Zip(\beta)\)์—์„œ ๋‚˜์˜ค๊ณ  ๋‚˜๋จธ์ง€๋Š” ๋ฒ„์ „ \(p^{\prime}=Zip(\beta^{\prime})\)์—์„œ ๋‚˜์˜จ๋‹ค.[|endofturn|]
[|system|]๋‹น์‹ ์€ ๋ฒˆ์—ญ๊ธฐ ์ž…๋‹ˆ๋‹ค. ์˜์–ด ๋ฌธ์žฅ์„ ํ•œ๊ตญ์–ด๋กœ ๋ฒˆ์—ญํ•˜์„ธ์š”.[|endofturn|] [|user|]**Theorem C.2** (Grokking with Tail Narrowing).: _Consider a sample of size \(T\) of which a proportion \(\pi\) comes from the true distribution \(p=Zip(\beta)\) and the remainder comes from a version \(p^{\prime}=Zip(\beta^{\prime})\). [|assistant|]