ํ‹ฐ์Šคํ† ๋ฆฌ

Cartinoe's paper review
๊ฒ€์ƒ‰ํ•˜๊ธฐ

๋ธ”๋กœ๊ทธ ํ™ˆ

Cartinoe's paper review

cartinoe5930.tistory.com/m

Welcome! I'm a student studying about deep learning(NLP) ๐Ÿ˜‰ The goal of my study is to develop a competent LLM helping people!

๊ตฌ๋…์ž
11
๋ฐฉ๋ช…๋ก ๋ฐฉ๋ฌธํ•˜๊ธฐ
๊ณต์ง€ ๋ธ”๋กœ๊ทธ ๊ณต์ง€์‚ฌํ•ญ - ๋ชจ๋ฐ”์ผ ์ˆ˜์‹ ๊นจ์ง ๋ชจ๋‘๋ณด๊ธฐ

์ฃผ์š” ๊ธ€ ๋ชฉ๋ก

  • Noise makes LLM better! - NEFTune ๐Ÿ˜‰ What is the big difference of NLP compared to CV? ๐Ÿ˜ฎ ์ด ํฌ์ŠคํŒ…์˜ ์ œ๋ชฉ๋ถ€ํ„ฐ ํ•ด์„œ ์˜์•„ํ•œ ๋ถ€๋ถ„์ด ํ•œ๋‘ ๊ฐ€์ง€๊ฐ€ ์•„๋‹ ๊ฒƒ์ด๋‹ค. ๊ฐ‘์ž๊ธฐ ๋’ค๋Œ์•„๋ด์•ผ ํ•œ๋‹ค๋А๋‹ˆ CV์™€ NLP์˜ ๊ฐ€์žฅ ํฐ ์ฐจ์ด์ ์ด ๋ฌด์—‡์ธ์ง€์— ๋Œ€ํ•ด ๋ฌป์ง€๋ฅผ ์•Š๋‚˜. ํ•˜์ง€๋งŒ ์ด๋ฒˆ ํฌ์ŠคํŒ…์—์„œ ๋งํ•˜๊ณ ์ž ํ•˜๋Š” ๋‚ด์šฉ์„ ์œ„ํ•ด์„œ๋Š” ์ด ์ฐจ์ด์ ์„ ๋˜์งš์–ด๋ณด์•„์•ผ ํ•  ํ•„์š”๊ฐ€ ์žˆ๋‹ค! ๊ทธ๋ ‡๋‹ค๋ฉด ๋จผ์ € ๋…์ž๋ถ„๋“ค๊ป˜ ์งˆ๋ฌธํ•ด ๋ณด๋„๋ก ํ•˜๊ฒ ๋‹ค. NLP๊ณผ CV์˜ ๊ฐ€์žฅ ํฐ ์ฐจ์ด์ ์€ ๋ฌด์—‡์ผ๊นŒ? ์•„๋งˆ๋„ ์ด๋ ‡๊ฒŒ ์ถ”์ƒ์ ์œผ๋กœ ์งˆ๋ฌธํ•œ๋‹ค๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋‹ต๋ณ€๋“ค์ด ๋‚˜์˜ฌ ๊ฒƒ์ด๋ผ ์ƒ๊ฐํ•œ๋‹ค. ๐Ÿ˜ ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ๋‹ค๋ฆ„. (text & image) ์‚ฌ์šฉ๋˜๋Š” ๋ชจ๋ธ๋“ค์˜ ์ฐจ์ด ํ•™์Šต ๋ฐฉ์‹์˜ ์ฐจ์ด ๋ฌผ๋ก  ์œ„์™€ ๊ฐ™์€ ๋‹ต๋ณ€๋“ค๋„ ๋งž์ง€๋งŒ, ํ•„์ž๊ฐ€ ๋ณธ ํฌ์ŠคํŒ…์—์„œ ๋งํ•˜๊ณ ์ž ํ•˜๋Š” ๋‘ ์—ฐ๊ตฌ๊ณ„์˜ ๊ฐ€์žฅ ํฐ ์ฐจ.. ๊ณต๊ฐ์ˆ˜ 1 ๋Œ“๊ธ€์ˆ˜ 0 2023. 10. 18.
  • Llama์˜ ์ƒˆ๋กœ์šด ๋Œ€ํ•ญ๋งˆ, Mistral LM! ๐Ÿ˜ฎ The preview of Llama3..? ์ตœ๊ทผ์— HuggingFace๋ฅผ ๋ณด๋‹ค๊ฐ€ ์•Œ๊ฒŒ ๋œ ๋ชจ๋ธ์ด ํ•˜๋‚˜ ์žˆ๋‹ค. ๋ฐ”๋กœ LLM ์‹œ์žฅ์„ ๋œจ๊ฒ๊ฒŒ ๋‹ฌ๊ตฐ ๋ชจ๋ธ์ธ Mistral LM์ด๋‹ค! ํ˜œ์„ฑ์ฒ˜๋Ÿผ Open-source LLM ๊ณ„์— ๋‚˜ํƒ€๋‚œ Mistral 7B๋Š” ๊ทธ ๋“ฑ์žฅ๋งŒ์œผ๋กœ๋„ Open-source LLM๊ณ„๋ฅผ ๋œจ๊ฒ๊ฒŒ ๋‹ฌ๊ตฌ์—ˆ๋‹ค. ๊ทธ๋ ‡๋‹ค๋ฉด Mistral 7B๋Š” ๋ฌด์—‡์„ ์–ด๋–ป๊ฒŒ ํ–ˆ๊ธธ๋ž˜ ๋ชจ๋‘์˜ ์ด๋ชฉ์„ ์ง‘์ค‘์‹œํ‚ฌ ์ˆ˜ ์žˆ์—ˆ๋˜ ๊ฒƒ์ผ๊นŒ? ๊ทธ๊ฒƒ์€ Mistral 7B๊ฐ€ ์ด๋ค„๋‚ธ ์—…์ ์„ ์‚ดํŽด๋ณด๋ฉด ์•Œ ์ˆ˜ ์žˆ๋‹ค: ๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ์—์„œ Llama2 13B๋ฅผ ๋Šฅ๊ฐ€ ๋งŽ์€ ๋ฒค์น˜๋งˆํฌ์—์„œ Llama1 34B๋ฅผ ๋Šฅ๊ฐ€(๋น„๊ต ๋Œ€์ƒ์ด Llama2๊ฐ€ ์•„๋‹ˆ๋ผ Llama1์ด์—ˆ๋˜ ์ด์œ ๋Š” Llama2์˜ 34B ๋ชจ๋ธ์ด ๊ณต๊ฐœ๋˜์—ˆ์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ) ์ฝ”๋“œ ๊ด€๋ จ ๋ฒค์น˜๋งˆํฌ์—์„œ CodeLlam.. ๊ณต๊ฐ์ˆ˜ 4 ๋Œ“๊ธ€์ˆ˜ 1 2023. 10. 2.
  • ์–ด๋–ป๊ฒŒ Quantization์„ ์ง„ํ–‰ํ•˜๋Š” ๊ฒƒ์ด ํšจ๊ณผ์ ์ผ๊นŒ? ๐Ÿค” Which quantization method is efficient & effective? ๐Ÿง ๋‚ ์ด ์ง€๋‚˜๋ฉด ์ง€๋‚ ์ˆ˜๋ก ์ ์  ์‚ฌ์ด์ฆˆ๊ฐ€ ์ปค์ ธ๊ฐ€๋Š” LLM์˜ ํŒ๋„์—์„œ ์ด๋“ค์„ ์†์‰ฝ๊ฒŒ ํšจ์œจ์  ๋ฐ ํšจ๊ณผ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์—๋Š” ๋ฌด์—‡์ด ์žˆ์„๊นŒ? ์š”์ฆ˜์—๋Š” ๋‹ค๋ฅธ method๋“ค๋ณด๋‹ค๋„ quantization, ์ฆ‰ ์–‘์žํ™”๋ฅผ ์ฃผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ์ถ”์„ธ์ด๋‹ค. ์ด quantization์„ ํ†ตํ•ด ์‚ฌ๋žŒ๋“ค์€ ๊ณ ์šฉ๋Ÿ‰ RAM์„ ๊ฐ€์ง€๋Š” GPU์—์„œ๋„ ์‚ฌ์šฉํ•˜๊ธฐ๊ฐ€ ํž˜๋“ค๋˜ LLM์„ ํ›จ์”ฌ ํšจ์œจ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ๋‹ค! ๐Ÿค— ์ตœ์†Œํ•œ์˜ ์„ฑ๋Šฅ ๊ฐ์†Œ๋กœ ์ตœ์ ์˜ ํšจ์œจ์„ฑ์„ ๋ณด์—ฌ์ฃผ๋Š” quantization์„ ์œ„ํ•ด HuuggingFace์—์„œ๋Š” 2๊ฐ€์ง€ quantization method๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์žˆ๋‹ค. ๋ฐ”๋กœ BitsAndBytes์™€ GPTQ์ด๋‹ค. ์ด๋ฅผ ํ† ๋Œ€๋กœ ๋‘ q.. ๊ณต๊ฐ์ˆ˜ 1 ๋Œ“๊ธ€์ˆ˜ 0 2023. 9. 18.
  • AlpaGasus2-QLoRA ๐Ÿฆ™๐Ÿฆ„๐Ÿค AlpaGasus2-QLoRA!! ๐Ÿฆ„ ์ด๋ฒˆ์— ์ง„ํ–‰ํ•œ ํ”„๋กœ์ ํŠธ 'AlpaGasus2-QLoRA'์— ๋Œ€ํ•ด์„œ ์„ค๋ช…ํ•˜๊ณ ์ž ํ•œ๋‹ค. ํ”„๋กœ์ ํŠธ์— ๋Œ€ํ•ด ์•Œ์•„๋ณด๊ธฐ ์ „์— ๋จผ์ € ์ด ์—ฐ๊ตฌ๋ฅผ ์ง„ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก AlpaGasus๋ฅผ ์ œ์•ˆํ•ด์ฃผ์‹  Lichang Chen ์™ธ 10๋ถ„๊ป˜ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ๋“œ๋ฆฝ๋‹ˆ๋‹ค. https://arxiv.org/abs/2307.08701 AlpaGasus: Training A Better Alpaca with Fewer Data Large language models~(LLMs) obtain instruction-following capability through instruction-finetuning (IFT) on supervised instruction/response data. However, wi.. ๊ณต๊ฐ์ˆ˜ 1 ๋Œ“๊ธ€์ˆ˜ 1 2023. 9. 5.
  • ์ด์ œ๋Š” ChatGPT๋ฅผ fine-tuning ํ•  ์‹œ๊ฐ„!! โฐ What a BIG NEWS!!! ๐Ÿ“ฐ ์ตœ๊ทผ ๋“ค์–ด ๋ธ”๋กœ๊ทธ ํฌ์ŠคํŒ…์„ ์˜ฌ๋ฆฌ๋Š” ๊ฒƒ์ด ๋œธํ•ด์กŒ๋Š”๋ฐ, ์˜ค๋Š˜ ์ •๋ง ๋†€๋ผ์šด ์†Œ์‹์„ ์ ‘ํ•˜๊ฒŒ ๋˜์–ด์„œ ์ด๋ ‡๊ฒŒ ์˜ค๋ž˜๊ฐ„๋งŒ์— ์ฐพ์•„์˜ค๊ฒŒ ๋˜์—ˆ๋‹ค. ๋ฐ”๋กœ ๋ณธ๋ก ์œผ๋กœ ๋“ค์–ด๊ฐ€์„œ ์šฐ๋ฆฌ๋‚˜๋ผ ์‹œ๊ฐ„์œผ๋กœ๋Š” ์˜ค๋Š˜! (๋ฌผ๋ก  ๋ฏธ๊ตญ ์‹œ๊ฐ„์œผ๋กœ๋Š” 8์›” 22์ผ์ด๊ธด ํ•˜๋‹ค ๐Ÿ˜) ๋“œ๋””์–ด OpenAI์—์„œ ์ด๋“ค์˜ ๊ฐ•๋ ฅํ•œ ์–ธ์–ด ๋ชจ๋ธ์ธ ChatGPT(gpt-3.5-turbo)์— ๋Œ€ํ•ด์„œ fine-tuning์„ ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋งŒ๋“ค์—ˆ๋‹ค!! ๐Ÿซข ๊ทธ๋ž˜์„œ ์ด๋ฒˆ ํฌ์ŠคํŒ…์—์„œ๋Š” OpenAI์—์„œ ์ด ์†Œ์‹์„ ์•Œ๋ฆฌ๊ธฐ ์œ„ํ•ด ์˜ฌ๋ฆฐ ๊ธ€์„ ํ† ๋Œ€๋กœ ์–ด๋–ป๊ฒŒ ChatGPT๋ฅผ fuine-tuning ํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ๊ทธ ์ž์„ธํ•œ ๋‚ด์šฉ๋“ค๊ณผ ์„ธ๋ถ€ ์‚ฌํ•ญ๋“ค์— ์•Œ์•„๋ณด๋ ค๊ณ  ํ•œ๋‹ค! ๐Ÿค— ์ด ํฌ์ŠคํŒ…์€ OpenAI์˜ ๊ธ€์„ ํ† ๋Œ€๋กœ ์ž‘์„ฑ๋˜์—ˆ์œผ๋‹ˆ ๋”์šฑ ์ž์„ธํ•œ ๋‚ด์šฉ์„ ํ™•์ธํ•˜๊ณ  ์‹ถ๋‹ค๋ฉด ๋‹ค์Œ์˜ .. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 0 2023. 8. 23.
  • Fine-tuning method์˜ ๋ฐœ์ „ ๊ณผ์ •!! Fine-tuning๋ถ€ํ„ฐ RLHF๊นŒ์ง€ ๐Ÿฆ–โžก๏ธ๐Ÿง‘ A new spectrum of model learning, Fine-tuning โœจ ์ด๋ฒˆ ํฌ์ŠคํŒ…์—์„œ ๋‹ค๋ค„๋ณด๊ณ ์ž ํ•˜๋Š” ๋‚ด์šฉ์€ ๋ชจ๋ธ์˜ fine-tuning ๋ฐฉ์‹์— ๋Œ€ํ•ด์„œ์ด๋‹ค. ์‚ฌ์‹ค ํฌ์ŠคํŒ…์˜ ์ˆœ์„œ๊ฐ€ ๋ฌด์–ธ๊ฐ€ ์ž˜๋ชป๋˜์—ˆ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋А๋ผ๊ณ  ์žˆ๊ธฐ๋Š” ํ•œ๋ฐ, ๊ทธ ์ ์€ ์–‘ํ•ด๋ฅผ ๋ถ€ํƒํ•œ๋‹ค..!! ๐Ÿ˜… ์ €๋ฒˆ ์‹œ๊ฐ„์— ํŒŒ๋ผ๋ฏธํ„ฐ ํšจ์œจ์ ์ธ fine-tuning์„ ์•Œ์•„๋ณด๋ฉด์„œ fine-tuning์„ ํšจ์œจ์ ์œผ๋กœ ํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์•Œ์•„๋ดค๋Š”๋ฐ, ๊ทธ๋ ‡๋‹ค๋ฉด fine-tuning์„ ์ข€ ๋” ํšจ๊ณผ์ ์œผ๋กœ ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์€ ์—†์„๊นŒ? ๋‹น์—ฐํžˆ ์žˆ๋‹ค!! ์ด๋ฒˆ ํฌ์ŠคํŒ…์—์„œ๋Š” fine-tuning method๊ฐ€ ์–ด๋–ป๊ฒŒ ๋ณ€ํ™” ํ•ด๋‚˜๊ฐ”๋Š”์ง€์— ๋Œ€ํ•ด ์•Œ์•„๋ณด๊ณ ์ž ํ•œ๋‹ค. ์ž, ๊ทธ๋ ‡๋‹ค๋ฉด fine-tuning์ด ๋ฌด์—‡์ผ๊นŒ? ์ €๋ฒˆ ํฌ์ŠคํŒ…์—์„œ ๋งํ–ˆ๋˜ ๊ฒƒ์ฒ˜๋Ÿผ ์ง€๊ธˆ์˜ ์ˆ˜๋งŽ์€ language.. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 0 2023. 8. 7.
  • ํ•œ ๋‹จ๊ณ„, ํ•œ ๋‹จ๊ณ„์”ฉ ์ธ๊ฐ„์ฒ˜๋Ÿผ ์ƒ๊ฐํ•ด๋ณด์ž! ๐Ÿง ๐Ÿค” Let's think step-by-step! ๐Ÿชœ ํฌ์ŠคํŒ…์˜ ์ œ๋ชฉ๊ณผ ์ด ์„น์…˜์˜ ์ œ๋ชฉ์„ ๋ดค์„ ๋•Œ ์˜์•„ํ•˜๊ฒŒ ์ƒ๊ฐํ•˜๋Š” ์‚ฌ๋žŒ๋“ค์ด ์žˆ์„ ๊ฒƒ์ด๋‹ค. '์•„๋‹ˆ ์ด ์‚ฌ๋žŒ, NLP ๊ด€๋ จ ์–˜๊ธฐ ์ž˜๋งŒ ํ•˜๋‹ค๊ฐ€ ๊ฐ‘์ž๊ธฐ ๋ฌด์Šจ ๋šฑ๋”ด์ง€๊ฐ™์€ ์†Œ๋ฆฌ๋ž˜? ๐Ÿคจ' ์ถฉ๋ถ„ํžˆ ๊ทธ๋Ÿด ์ˆ˜ ์žˆ๋‹ค! ํ•˜์ง€๋งŒ, NLP ๊ด€๋ จ ๋…ผ๋ฌธ์„ ์ฝ์–ด๋ดค๊ฑฐ๋‚˜ ์ตœ์‹  method๋“ค์— ๋Œ€ํ•ด ์ž˜ ์•Œ๊ณ  ์žˆ๋Š” ์‚ฌ๋žŒ์ด๋ฉด ํ•„์ž๊ฐ€ ๋ฌด์Šจ ์†Œ๋ฆฌ๋ฅผ ํ•˜๊ณ  ์‹ถ์–ด ํ•˜๋Š” ๊ฒƒ์ธ์ง€๋ฅผ ์•Œ ๊ฒƒ์ด๋ผ ์ƒ๊ฐํ•œ๋‹ค. ์™œ๋ƒํ•˜๋ฉด ์ด ์„น์…˜์˜ ์ œ๋ชฉ์ด 'Let's think step-by-step'์€ ์ด ํฌ์ŠคํŒ…์„ ๊ด€ํ†ตํ•˜๋Š” ๋ฌธ์žฅ์ด์ž, ์œ ๋ช…ํ•œ ๋…ผ๋ฌธ์—์„œ ์‚ฌ์šฉ๋œ method์ด๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. ์ด๊ฒŒ ๋ฌด์Šจ ์†Œ๋ฆฌ๋ƒ๊ตฌ์š”? ๊ถ๊ธˆํ•˜์‹œ๋‹ค๋ฉด, LM์ด ์‚ฌ๋žŒ๊ณผ ๋น„์Šทํ•œ ๋ฐฉ์‹์œผ๋กœ ์‚ฌ๊ณ ๋ฅผ ํ•ด์„œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ฒŒ ํ•˜๊ณ ์ž ํ•œ method๋“ค์— ๋Œ€ํ•ด ์•Œ์•„๋ณด๋Š” ์ด๋ฒˆ ํฌ์ŠคํŒ…์„ ๋.. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 0 2023. 8. 3.
  • ๋‹น์‹ ๋„ Fine-tuning ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค! with PEFT ๐Ÿค— The current trend of LM ๐Ÿ“ˆ 2017๋…„ Vaswani ๊ป˜์„œ 'Attention Is All You Need'๋ผ๋Š” ๋…ผ๋ฌธ์œผ๋กœ Transformer๋ฅผ ์ฒ˜์Œ ์†Œ๊ฐœํ•˜์‹œ๊ณ , ๊ทธ ํ›„ 2018๋…„์— BERT์™€ GPT๊ฐ€ ๋‚˜์˜ค๊ฒŒ ๋˜๋ฉด์„œ๋ถ€ํ„ฐ LM(Language Model)์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋Š” ๊ทธ ์‹œ์ž‘์„ ์•Œ๋ ธ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ด ๋‹น์‹œ์— ์†Œ๊ฐœ๋˜์—ˆ๋˜ pre-training & fine-tuning์ด๋ผ๋Š” ๊ฐœ๋…์€ ์•„์ง๊นŒ์ง€๋„ ๋„๋ฆฌ ์‚ฌ์šฉ๋  ์ •๋„๋กœ ํฌ๋‚˜ํฐ LM์˜ framework๋ฅผ ์ด๋ฃจ๊ฒŒ ๋˜์—ˆ๋‹ค. ์ด๋ฒˆ ํฌ์ŠคํŒ…์—์„œ ์•Œ์•„๋ณด๊ฒŒ ๋  PEFT(์ž์„ธํ•œ ๋œป์€ ์กฐ๊ธˆ ๋’ค์— ์•Œ๋ ค๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค! ๐Ÿ˜„)๋„ ์ด ์ค‘ fine-tuning์— ๊ด€๋ จ๋œ method์ด๋‹ค. PEFT์— ๋Œ€ํ•ด ์•Œ์•„๋ณด๊ธฐ ์ „์— ์ด pre-training๊ณผ fine-tuning์ด ๊ณผ์—ฐ ์ •ํ™•ํžˆ .. ๊ณต๊ฐ์ˆ˜ 1 ๋Œ“๊ธ€์ˆ˜ 0 2023. 8. 1.
  • ChatGPT์˜ ์„ฑ๋Šฅ์ด ์•ˆ ์ข‹์•„์ง€๊ณ  ์žˆ๋‹ค๊ตฌ?!?!? ๐Ÿ˜ฒ๐Ÿ˜ฒ Did you hear that..? ๐Ÿ˜ฑ ์š”์ฆ˜ ์„ธ๊ฐ„์— ๋– ๋„๋Š” ํ•˜๋‚˜์˜ ์†Œ๋ฌธ์ด ์žˆ๋‹ค๊ณ  ํ•œ๋‹ค. ์ด์ œ๋Š” ์šฐ๋ฆฌ์—๊ฒŒ ์นœ์ˆ™ํ•ด์ง„, ์˜คํžˆ๋ ค ์—†์œผ๋ฉด ๋ถˆํŽธํ•จ์„ ๋А๋‚„ ์ˆ˜ ์žˆ์„ ์ •๋„๋กœ ๊ฐ€๊นŒ์›Œ์ง„ ChatGPT์˜ ์„ฑ๋Šฅ์ด ์•ˆ ์ข‹์•„์กŒ๋‹ค๋Š” ์†Œ๋ฌธ์ด๋‹ค!! ๐Ÿ˜ฎ ์‹ค์ œ ์–ด๋–ค ์†Œ๋ฌธ๋“ค์ด ์žˆ๋Š”์ง€์— ๋Œ€ํ•ด ์•Œ์•„๋ณด๊ธฐ ์ „์— ์šฐ์„  ์ตœ๊ทผ ChatGPT์™€ GPT-4์˜ ์ •ํ™•ํ•œ ์ฐจ์ด์— ๋Œ€ํ•ด ์•Œ์•„๋ณด๊ณ , ์ตœ๊ทผ ์ด ๋ชจ๋ธ๋“ค์— ์ƒ๊ธด ๋ณ€ํ™”์— ๋Œ€ํ•ด์„œ ์•Œ์•„๋ณด๋„๋ก ํ•˜์ž. ChatGPT์™€ GPT-4๋Š” ๊ทธ ์‚ฌ์šฉ๋œ ๋ชจ๋ธ์— ์ฐจ์ด๊ฐ€ ์žˆ๋‹ค. ChatGPT๋Š” GPT-3.5์— RLHF๋ฅผ ์ง„ํ–‰ํ•œ ๋ชจ๋ธ์ด๊ณ , GPT-4๋Š” ๋ง ๊ทธ๋Œ€๋กœ GPT-3.5์—์„œ ํ›จ์”ฌ ๋” ๋ฐœ์ „๋œ GPT-4 ๋ชจ๋ธ์„ ๋งํ•œ๋‹ค. (GPT-4์— ๋Œ€ํ•ด์„œ๋Š” ์ž์„ธํžˆ ๋ฐํ˜€์ง„ ๊ฒƒ์ด ์—†๊ธฐ ๋•Œ๋ฌธ์— ์ •ํ™•ํ•œ ๋น„๊ต๋Š” ๋ถˆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค,, ๐Ÿ˜“) OpenAI์—์„œ ์ œ๊ณต.. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 0 2023. 7. 31.
  • LM์„ ๊ฐ€์žฅ ์ตœ์ ์œผ๋กœ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์€ ๋ฌด์—‡์ผ๊นŒ? ๐Ÿ˜Ž ์ด๋ฒˆ ํฌ์ŠคํŒ…์€ ๊ธฐ์กด์˜ ํฌ์ŠคํŒ…๊ณผ ์‚ด์ง ๋‹ค๋ฅด๊ฒŒ PPT ์ž๋ฃŒ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์„ค๋ช…ํ•˜๋„๋ก ํ•˜๊ฒ ๋‹ค. ์ด๋ฒˆ ํฌ์ŠคํŒ…์˜ ์ฃผ์ œ๋Š” ์ œ๋ชฉ์—์„œ ๋ณด์—ฌ์ง€๋Š” ๊ฒƒ์ฒ˜๋Ÿผ LM์˜ Evaluation metric์— ๋Œ€ํ•ด์„œ ์•Œ์•„๋ณด๋Š” ์‹œ๊ฐ„์„ ๊ฐ€์ ธ๋ณด๋ ค๊ณ  ํ•œ๋‹ค! ๐Ÿ˜Š ๊ธฐ์กด์˜ Evaluation metric์— ๋Œ€ํ•ด์„œ ์•Œ์•„๋ณด๊ณ , ๊ธฐ์กด metric๋“ค์— ์–ด๋– ํ•œ ๋ฌธ์ œ๊ฐ€ ์žˆ๋Š”์ง€ ์•Œ์•„๋ณธ ๋’ค, ๋งˆ์ง€๋ง‰์œผ๋กœ ์–ด๋–ค ๊ฐœ์„ ์•ˆ๋“ค์ด ์ƒ๊ฒจ๋‚ฌ๋Š”์ง€์— ๋Œ€ํ•ด์„œ ํ•œ ๋ฒˆ ์•Œ์•„๋ณด๋„๋ก ํ•˜๊ฒ ๋‹ค. ๋งŒ์•ฝ PPT๋ฅผ ๋ณด๋ฉด์„œ ๊ถ๊ธˆํ•˜๊ฑฐ๋‚˜ ์˜ค๋ฅ˜๊ฐ€ ์žˆ๋Š” ๊ฒƒ ๊ฐ™์€ ์‚ฌํ•ญ๋“ค์€ PPT ๋˜๋Š” ํฌ์ŠคํŒ…์— ๋Œ“๊ธ€์„ ๋‹ฌ์•„์ฃผ์‹œ๋ฉด ๋‹ต๋ณ€์„ ๋‹ฌ์•„๋†“๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค! ์žฌ๋ฐŒ๊ฒŒ ๋ด์ฃผ์‹ญ์‡ผ! ๐Ÿคฉ https://docs.google.com/presentation/d/1XL_B0nI-yp2dgLDVrEzTlLcg9DpUnALBklmpJ4iOZRw/e.. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 1 2023. 7. 27.
  • LM์˜ context window, ๊ธธ์–ด์•ผ ํ• ๊นŒ? ์งง์•„์•ผ ํ• ๊นŒ? ๐Ÿ“๐Ÿคจ Newly spotlighted elements of LM โœจ LM์€ ์‹œ์‹œ๊ฐ๊ฐ ๋ณ€ํ™”ํ•ด๊ฐ€๊ณ  ์žˆ๋‹ค. ๋ฉฐ์น  ์ „์— ์ƒˆ๋กญ๊ฒŒ ๋ฐœํ‘œ๋œ ๋ชจ๋ธ์ด ์˜ค๋Š˜์—์„œ๋Š” ๊ทธ ๋ฉด๋ชจ๊ฐ€ ๋‚ฑ๋‚ฑ์ด ํŒŒ์•…๋˜์–ด ๋ถ€์กฑํ•œ ์ ๋“ค์ด๋‚˜ ๋‹จ์ ๋“ค์ด ์ง€์ ๋ฐ›๊ณ  ์žˆ๋Š” ์š”์ฆ˜์ด๋‹ค. ๐Ÿ˜ฅ ๊ทธ๋งŒํผ LM์€ ๊ทธ๊ฒƒ์ด ํŒŒ๋ผ๋ฏธํ„ฐ๋“  ๋ฐ์ดํ„ฐ๋“  ๋‹ค๋ฐฉ๋ฉด์œผ๋กœ ๋น ๋ฅด๊ฒŒ ๋ณ€ํ™”ํ•ด๋‚˜๊ฐ€๊ณ  ์žˆ๋Š”๋ฐ, ์ด๋ฒˆ ํฌ์ŠคํŒ…์—์„œ ๋‹ค๋ค„๋ณด๊ณ ์ž ํ•˜๋Š” ๋‚ด์šฉ์€ ์˜ค๋žœ ์‹œ๊ฐ„ ๋™์•ˆ ๋ณ„๋กœ ๊ฑด๋“œ๋ ค์ง€์ง€ ์•Š๋‹ค๊ฐ€ ์ตœ๊ทผ์— ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ(Chen et al., 2023, Ding et al., 2023, Liu et al., 2023)๋ฅผ ํ†ตํ•ด ๋‹ค์‹œ ๊ฐ๊ด‘๋ฐ›๊ณ  ์žˆ๋Š” ๋‚ด์šฉ์ธ LM์˜ context window์— ๋Œ€ํ•ด์„œ ์–˜๊ธฐํ•ด๋ณด๊ณ ์ž ํ•œ๋‹ค! ๐Ÿ˜Š What is the 'context window'? ๐Ÿค” ์‹œ์ž‘ํ•˜๊ธฐ์— ์•ž์„œ์„œ ์ด๋ฒˆ ํฌ์ŠคํŒ…์—์„œ ์ค‘์š”ํ•˜๊ฒŒ ๋‹ค๋ค„๋ณผ ๋‚ด์šฉ์ธ .. ๊ณต๊ฐ์ˆ˜ 3 ๋Œ“๊ธ€์ˆ˜ 0 2023. 7. 26.
  • Closed-source๐Ÿ”’? Open-source๐Ÿ”“? ๊ทธ๊ฒŒ ๋ญ”๋ฐ?? ๐Ÿคจ๐Ÿค” Starting from ChatGPT ๐Ÿค– which is closed-source ์ž‘๋…„ 12์›”, ์ฆ‰ 2022๋…„ 12์›”์— ์ „ ์„ธ๊ณ„์˜ ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ์ ์ž–์ด ์‹ ์„ ํ•œ ์ถฉ๊ฒฉ์„ ์ค€ ์‚ฌ๊ฑด์ด ๋ฐœ์ƒํ•˜์˜€๋‹ค. ๋ฐ”๋กœ ๊ทธ ์œ ๋ช…ํ•œ 'ChatGPT'์˜ ๋ฐœํ‘œ๋‹ค! OpenAI์—์„œ ๋ฐœํ‘œํ•œ ์ด ๊ฑฐ๋Œ€ ์–ธ์–ด ๋ชจ๋ธ(Large Language Model, LLM)์€ ์ง€๊ธˆ๊นŒ์ง€์™€๋Š” ์ฐจ์›์ด ๋‹ค๋ฅธ ์—„์ฒญ๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋ฉด์„œ ์‚ฌ๋žŒ๋“ค์˜ ์‚ฌํšŒ ๋ฐ ์‚ถ์— ์ „๋ฐ˜์ ์œผ๋กœ ์Šค๋ฉฐ๋“ค์–ด๊ฐ€๊ธฐ ์‹œ์ž‘ํ–ˆ๋‹ค. ํ•˜์ง€๋งŒ, ์ด ์™„๋ฒฝํ•ด ๋ณด์ด๋Š” ChatGPT๋„ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๋‹จ์ ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š”๋ฐ, ๊ทธ์ค‘์—์„œ ์ด๋ฒˆ ํฌ์ŠคํŒ…์—์„œ ๋‹ค๋ค„๋ณด๊ณ ์ž ํ•˜๋Š” ๋‚ด์šฉ์€ ๋ฐ”๋กœ 'Closed-source' model์ด๋ผ๋Š” ์ ์ด๋‹ค. ๐Ÿšซ closed-source๊ฐ€ ๋ฌด์—‡์ผ๊นŒ? ์ด ์šฉ์–ด๋ฅผ ์ฒ˜์Œ ๋“ฃ๊ฒŒ ๋œ๋‹ค๋ฉด ๋‹ค์†Œ ์ƒ์†Œํ• ํ…๋ฐ, clos.. ๊ณต๊ฐ์ˆ˜ 1 ๋Œ“๊ธ€์ˆ˜ 0 2023. 7. 25.
  • How has scaling law developed in NLP? ๐Ÿค” - NLP์—์„œ scaling law๋Š” ์–ด๋–ป๊ฒŒ ๋ฐœ์ „๋˜์—ˆ์„๊นŒ? Before Starting.. 2017๋…„ NLP๋ฅผ ํฌํ•จํ•œ ์ง€๊ธˆ๊นŒ์ง€์˜ ๋”ฅ๋Ÿฌ๋‹์˜ ํŒ๋„๋ฅผ ๋’ค์ง‘์–ด์—Ž๋Š” ํ˜์‹ ์ ์ธ ๋ชจ๋ธ์ธ 'Transformer'๊ฐ€ ์ œ์•ˆ๋˜์—ˆ๋‹ค. ์ด๋ฒˆ ํฌ์ŠคํŒ…์—์„œ ๋‹ค๋ค„๋ณผ ๋‚ด์šฉ์€ Transformer์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์ด ์•„๋‹ˆ๊ธฐ์— ๋”ฐ๋กœ ๊นŠ์ด ์•Œ์•„๋ณด์ง€๋Š” ์•Š๊ฒ ์ง€๋งŒ, ์ด๋ฒˆ ํฌ์ŠคํŒ…์„ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ด ๋ชจ๋ธ์˜ ์‚ฌ์ด์ฆˆ์— ๋Œ€ํ•ด์„œ๋Š” ์•Œ์•„๋‘˜ ํ•„์š”๊ฐ€ ์žˆ๋‹ค. Transformer์˜ ์‚ฌ์ด์ฆˆ๋Š” 465M ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง€๋Š” ๋ชจ๋ธ์ด์—ˆ๋‹ค. ํ•˜์ง€๋งŒ, ๋ถˆ๊ณผ 3๋…„ ๋งŒ์— ์ด ์‚ฌ์ด์ฆˆ๊ฐ€ ์ •๋ง ์ž‘๊ฒŒ ๋А๊ปด์ง€๊ฒŒ ํ•  ๋งŒํผ ํฐ ์‚ฌ์ด์ฆˆ์˜ ๋ชจ๋ธ์ธ GPT-3(175B)๊ฐ€ ๋‚˜์˜ค๊ฒŒ ๋˜์—ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ํ˜„์žฌ๊นŒ์ง€๋„ ์ด๋ณด๋‹ค ๋” ํฐ ๋ชจ๋ธ๋“ค์€ ๊ณ„์† ๋‚˜์˜ค๊ณ  ์žˆ๋‹ค. LM์˜ ์‚ฌ์ด์ฆˆ๊ฐ€ ์ด๋ ‡๊ฒŒ ์ ์  ์ปค์ง€๊ฒŒ ๋œ ์ด์œ ๋Š” ๋ฌด์—‡์ผ๊นŒ? ๊ทธ ์ด์œ ๋Š” Kaplan et al. 2020.. ๊ณต๊ฐ์ˆ˜ 1 ๋Œ“๊ธ€์ˆ˜ 2 2023. 7. 24.
  • CoT Collection: Improving Zero-shot and Few-shot Learning of Language Models via Chain-of-Thought Fine-tuning ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ The overview of this paper ์–ด๋–ป๊ฒŒ ๊ณต๊ฐ์ˆ˜ 1 ๋Œ“๊ธ€์ˆ˜ 0 2023. 7. 20.
  • SelFee: Iterative Self-Revising LLM Empowered by Self-Feedback Generation ๋ฆฌ๋ทฐ Introduction SelFee SelFee๋Š” KAIST์˜ LK Lab์—์„œ ๋งŒ๋“  ์ƒˆ๋กœ์šด instruction-following LM์œผ๋กœ ์‘๋‹ต์—์„œ self-feedback์„ ์ƒ์„ฑํ•˜๊ณ  ํ”ผ๋“œ๋ฐฑ์— ๊ธฐ๋ฐ˜ํ•ด์„œ self-revise ํ•˜๋Š” ๋ชจ๋ธ์ด๋‹ค. ChatGPT์— ์˜ํ•ด ์ƒ์„ฑ๋œ self-feedback๊ณผ revision data๋ฅผ ํฌํ•จํ•˜๋Š” 178K ๊ฐœ์˜ training instance๋ฅผ ์‚ฌ์šฉํ•ด์„œ LLaMA model(7B & 13B)์„ fine-tune ํ•˜์˜€๋‹ค. SelFee์˜ ์ž‘๋™ ์˜ˆ์‹œ Vicuna Evaluation์—์„œ ๋‘ SelFee(7B & 13B) ๋ชจ๋ธ์€ LLaMA, Alpaca, Vicuna, Guanaco๋ฅผ ๋Šฅ๊ฐ€ํ•˜๊ณ  ChatGPT์™€ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์คฌ๋‹ค. SelFee๋Š” ํŠนํžˆ high-quality te.. ๊ณต๊ฐ์ˆ˜ 1 ๋Œ“๊ธ€์ˆ˜ 0 2023. 6. 30.
  • Self-Refine: Iterative Refinement with Self-Feedback ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ The overview of this paper ์ด ๋…ผ๋ฌธ์—์„œ๋Š” Self-Refine์„ ์†Œ๊ฐœํ•˜์˜€๋‹ค. Self-Refine์€ ๋ฐ˜๋ณต์ ์ธ ํ”ผ๋“œ๋ฐฑ๊ณผ ๊ฐœ์„ ์„ ํ†ตํ•ด LLM์˜ ์ดˆ๊ธฐ output์„ ๊ฐœ์„ ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋ฐฉ๋ฒ•์ด๋‹ค. Self-Refine์˜ ์ฃผ๋œ ์•„์ด๋””์–ด๋Š” LLM์„ ์‚ฌ์šฉํ•ด ์ดˆ๊ธฐ output์„ ์ƒ์„ฑํ•˜๊ณ , ๊ทธ๋‹ค์Œ์— ๋˜‘๊ฐ™์€ LLM์ด output์— ๋Œ€ํ•ด ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ณ  ์ด ํ”ผ๋“œ๋ฐฑ์„ ์‚ฌ์šฉํ•ด ๋ฐ˜๋ณต์ ์œผ๋กœ ์ž๊ธฐ ์ž์‹ ์„ ๊ฐœ์„ ํ•ด ๋‚˜๊ฐ€๋Š” ๊ฒƒ์ด๋‹ค. ํ•œ ๋งˆ๋””๋กœ Self-Refine์€ ํ•˜๋‚˜์˜ LLM์„ generator, refiner, feedback provider๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด๋‹ค. Self-Refine์€ ๋ชจ๋“  ํ‰๊ฐ€๋œ task์—์„œ Self-Refine์œผ๋กœ ์ƒ์„ฑ๋œ output์€ ๊ธฐ์กด์˜ ๋˜‘๊ฐ™์€ LLM์œผ๋กœ ์ƒ์„ฑ๋œ output๋ณด๋‹ค human.. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 0 2023. 6. 29.
  • Reflexion: Language Agents with Verbal Reinforcement Learning ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ The overview of this paper ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๊ฐ€์ค‘์น˜๋ฅผ ์—…๋ฐ์ดํŠธํ•˜์ง€ ์•Š๊ณ  ๋Œ€์‹ ์— ์–ธ์–ด์  ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ด language agent๋ฅผ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ƒˆ๋กœ์šด ํ”„๋ ˆ์ž„์›Œํฌ์ธ Reflexion์„ ์†Œ๊ฐœํ•˜์˜€๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, Reflexion agent๋Š” task ํ”ผ๋“œ๋ฐฑ ์‹ ํ˜ธ์— ๋Œ€ํ•ด ์–ธ์–ด๋กœ ๋‚˜ํƒ€๋‚ด๊ณ , ๊ทธ๋‹ค์Œ์— ์ดํ›„์˜ ์‹œ๋„์— ๋” ๋‚˜์€ ์˜์‚ฌ ๊ฒฐ์ •์„ ์œ ๋ฐœํ•˜๊ธฐ ์œ„ํ•ด ๋ฉ”๋ชจ๋ฆฌ ๋ฒ„ํผ์— ์ด๋“ค๋งŒ์˜ reflective text๋ฅผ ์œ ์ง€ํ•œ๋‹ค. Reflexion์€ ๋‹ค์–‘ํ•œ ํƒ€์ž…๊ณผ ์†Œ์Šค์˜ ํ”ผ๋“œ๋ฐฑ ์‹ ํ˜ธ๋ฅผ ํฌํ•จํ•  ์ˆ˜ ์žˆ์„ ์ •๋„๋กœ ์ถฉ๋ถ„ํžˆ ์œ ์—ฐํ•˜๊ณ , ๋‹ค์–‘ํ•œ task์— ๊ฑธ์ณ์„œ baseline agent์— ๋น„ํ•ด์„œ ์ƒ๋‹นํ•œ ๊ฐœ์„ ์„ ์–ป์—ˆ๋‹ค. Table of Contents 1. Introduction 2. Reflexion: reinforceme.. ๊ณต๊ฐ์ˆ˜ 1 ๋Œ“๊ธ€์ˆ˜ 0 2023. 6. 28.
  • GPT-4๋„ ์ž˜ ๋ชปํ•œ API ํ˜ธ์ถœ์„ ํ•œ๋‹ค๊ณ ?!? - Gorilla๐Ÿฆ: Large Language Model Connected with Massive APIs ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ The overview of this paper LLM์€ ์ตœ๊ทผ์— ์—„์ฒญ ๋ฐœ์ „ํ–ˆ์œผ๋‚˜, ์ด๋“ค์˜ API ํ˜ธ์ถœ์„ ํ†ตํ•œ ํšจ๊ณผ์ ์ธ ํˆด ์‚ฌ์šฉ์— ๋Œ€ํ•œ ์ž ์žฌ์„ฑ์€ ๋งŒ์กฑ๋˜์ง€ ์•Š์€ ์ฑ„ ๋‚จ์•„์žˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” API ํ˜ธ์ถœ ์ž‘์„ฑ์—์„œ GPT-4์˜ ์„ฑ๋Šฅ์„ ๋Šฅ๊ฐ€ํ•˜๋Š” fine-tuned LLaMA-based model์ธ Gorilla๐Ÿฆ๋ฅผ ์†Œ๊ฐœํ•˜์˜€๋‹ค. Gorilla๋Š” document retriever์™€ ํ•จ๊ป˜ ์‚ฌ์šฉ๋  ๋•Œ, test-time ๋ฌธ์„œ ๋ณ€ํ™”์— ์ ์‘ํ•˜๊ธฐ ์œ„ํ•œ ๊ฐ•๋ ฅํ•œ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๊ณ , ์œ ์—ฐํ•œ ์‚ฌ์šฉ์ž ์—…๋ฐ์ดํŠธ ๋˜๋Š” ๋ฒ„์ „ ๋ณ€ํ™”๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ด ์ฃผ์—ˆ๋‹ค. ์ด๊ฒƒ์€ LLM์„ direct ํ•˜๊ฒŒ prompting ํ•  ๋•Œ ์ผ๋ฐ˜์ ์œผ๋กœ ๋งž๋‹ฅ๋œจ๋ฆฌ๋Š” hallucination์˜ ๋ฌธ์ œ์ ์„ ์ƒ๋‹นํžˆ ์™„ํ™”ํ•˜์˜€๋‹ค. ๋˜ํ•œ ๋…ผ๋ฌธ์—์„œ๋Š” Gorilla์˜ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด .. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 0 2023. 6. 27.
  • Open-domain instruction์˜ ํšจ๊ณผ ๐Ÿช„ - WizardLM: Empowering Large Language Models to Follow Complex Instructions ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ The overview of this paper open-domain instruction๊ณผ ํ•จ๊ป˜ LLM์„ ํ•™์Šต์‹œํ‚ค๋Š” ๊ฒƒ์€ ์ƒ๋‹นํ•œ ์„ฑ๊ณต์„ ๊ฐ€์ ธ์™”๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์‚ฌ๋žŒ ๋Œ€์‹ ์— LLM์„ ์‚ฌ์šฉํ•ด์„œ ๋‹ค์–‘ํ•œ ๋ ˆ๋ฒจ์˜ ๋ณต์žก๋„๋ฅผ ๊ฐ€์ง€๋Š” ๋งŽ์€ ์–‘์˜ instruction data๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ๋ฐฉ์•ˆ์„ ๋ณด์—ฌ์ค€๋‹ค. ์ดˆ๊ธฐ instruction set์™€ ํ•จ๊ป˜ ์‹œ์ž‘ํ•ด์„œ, ์ด instruction set๋ฅผ Evol-instruct๋ฅผ ์‚ฌ์šฉํ•ด์„œ ๋”์šฑ ๋ณต์žกํ•œ instruction์œผ๋กœ step-by-step ์ž‘์„ฑํ•˜์˜€๋‹ค. ๊ทธ๋‹ค์Œ์—, ๋ชจ๋“  ์ƒ์„ฑ๋œ instruction ๋ฐ์ดํ„ฐ๋ฅผ LLaMA๋ฅผ fine-tune ํ•˜๊ธฐ ์œ„ํ•ด ์„ž์—ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•ด์„œ ๋‚˜์˜จ ๋ชจ๋ธ์ด ๋ฐ”๋กœ WizardLM์ด๋‹ค. Human Evaluation & Vicuna Evaluatio.. ๊ณต๊ฐ์ˆ˜ 2 ๋Œ“๊ธ€์ˆ˜ 2 2023. 6. 26.
  • ํ•„์š”ํ•œ ๊ฑด ์˜ค์ง ๊ต๊ณผ์„œ ์ˆ˜์ค€์˜ ๋ฐ์ดํ„ฐ๋ฟ!! ๐Ÿ“– - phi-1: Textbooks Are All You Need ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ The overview of this paper ๋…ผ๋ฌธ์—์„œ๋Š” ๋‹ค๋ฅธ ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ์ž‘๊ณ  code๋ฅผ ์œ„ํ•œ LLM์ธ phi-1์„ ์†Œ๊ฐœํ•˜์˜€๋‹ค. phi-1์€ 1.3B Transformer model์ด๊ณ , ์›น์œผ๋กœ๋ถ€ํ„ฐ textbook ํ€„๋ฆฌํ‹ฐ ๋ฐ์ดํ„ฐ์˜ ์„ ํƒ์  ๋ชจ์Œ๊ณผ ์ข…ํ•ฉ์ ์œผ๋กœ ์ƒ์„ฑ๋œ textbook์„ ์‚ฌ์šฉํ•˜๊ณ , GPT-3.5๋กœ ํ›ˆ๋ จ๋˜์—ˆ๋‹ค. phi-1์€ ์ž‘์€ ๊ทœ๋ชจ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋†’์€ pass@1 accuracy๋ฅผ ๋‹ฌ์„ฑํ•˜์˜€๋‹ค. Table of Contents 1. Introduction 2. Training details and the importance of high-quality data 3. Spikes of model capability after finetuning on CodeExercises 4. Evaluati.. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 0 2023. 6. 25.
  • LM์ด ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๊ฒŒ ๋œ๋‹ค๋ฉด? ๐Ÿ”ฌ: Large Language Models as Tool Makers ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ The overview of this paper ์ตœ๊ทผ์˜ ์—ฐ๊ตฌ๋Š” LLM์˜ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ ํ–ฅ์ƒ์˜ ์ž ์žฌ์„ฑ์„ ๋ณด์—ฌ์คฌ๋‹ค. ํ•˜์ง€๋งŒ, ์ด์ „ ์—ฐ๊ตฌ๋“ค์€ ๊ธฐ์กด ํˆด์˜ ๊ฐ€์šฉ์„ฑ์— ์ƒ๋‹นํžˆ ์˜์กดํ•œ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์˜์กด์„ฑ์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด closed-loop ํ”„๋ ˆ์ž„์›Œํฌ์ธ LLM As Tool Makers(LATM)์„ ์ œ์•ˆํ•˜์˜€๋‹ค. LATM์—์„œ LLM์€ ๋ฌธ์ œ ํ•ด๊ฒฐ์„ ์œ„ํ•œ ์ž์‹ ๋งŒ์˜ ์žฌ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํˆด์„ ์ƒ์„ฑํ•œ๋‹ค. LATM์€ 2๊ฐœ์˜ ๋ฉ”์ธ ํŽ˜์ด์ฆˆ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ๋‹ค: tool making & tool using. tool making์€ LLM์ด ์„œ๋กœ ๋‹ค๋ฅธ ์š”์ฒญ์— ์ ์šฉ๋  ์ˆ˜ ์žˆ๋Š” tool์„ ๊ณ„์†์ ์œผ๋กœ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด ์ค˜์„œ ํ–ฅํ›„ ์š”์ฒญ์€ task๋ฅผ ํ•ด๊ฒฐํ•  ๋•Œ ์šฐ์ตํ•˜๋‹ค๊ณ  ์ƒ๊ฐ๋  ๋•Œ ํ•ด๋‹น APT๋ฅผ ๋ถˆ๋Ÿฌ์˜ฌ ์ˆ˜ ์žˆ๊ฒŒ ํ•ด ์ค€๋‹ค. ์ด๋ ‡๊ฒŒ ํ•ด์„œ ์ด .. ๊ณต๊ฐ์ˆ˜ 1 ๋Œ“๊ธ€์ˆ˜ 0 2023. 6. 24.
  • ๐ŸฌOrca: Progressive Learning from Complex Explanation Traces of GPT-4 ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ The overview of this paper ์ตœ๊ทผ์˜ ์—ฐ๊ตฌ๋“ค์€ smaller model์˜ ์—ญ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด imitation learning์„ ํ†ตํ•ด large foundation models(LFM)์— ์˜ํ•ด ์ƒ์„ฑ๋œ output๊ณผ ํ•จ๊ป˜ ํ–ฅ์ƒ์‹œํ‚ค๊ณ ์ž ํ•˜์˜€๋‹ค. ํ•˜์ง€๋งŒ ์—ฌ๊ธฐ์—๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๋ฌธ์ œ์ ๋“ค์ด ์กด์žฌํ•œ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด Orca๋ฅผ ์†Œ๊ฐœํ•˜์˜€๋‹ค. Orca๋Š” LFM์˜ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค๋ฅผ ๋ชจ๋ฐฉํ•˜๊ธฐ ์œ„ํ•ด ํ•™์Šตํ•˜๋Š” 13B ๋ชจ๋ธ์ด๋‹ค. Orca๋Š” explanation trace(step-by-step process)๋ฅผ ํฌํ•จํ•˜๋Š” GPT-4 ๋กœ๋ถ€ํ„ฐ ํ’๋ถ€ํ•œ ์‹œ๊ทธ๋„์„ ํ•™์Šตํ•˜๊ณ , ChatGPT teacher assistant์— ์˜ํ•ด ์ง€๋„๋˜๋Š” ๋‹ค๋ฅธ ๋ณต์žกํ•œ instruction์—์„œ ํ•™์Šต๋˜์—ˆ๋‹ค. ์ด๋Ÿฌํ•œ progress.. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 0 2023. 6. 23.
  • KD์— ์‚ด์ง์˜ ๋ณ€ํ™”๋ฅผ ์ค˜๋ณด์ž!! ๐Ÿ˜œ - Knowledge Distillation of Large Language Models ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ The overview of this paper ์ด์ „์˜ KD๋Š” ์ฃผ๋กœ black-box model API๋ฅผ ๋ชจ๋ฐฉํ•˜๊ธฐ ์œ„ํ•ด white-box ๋ถ„๋ฅ˜ ๋ชจ๋ธ ๋˜๋Š” small model์„ ํ•™์Šต์‹œํ‚ค๋Š”๋ฐ ์ ์šฉ๋œ๋‹ค. white-box ์ƒ์„ฑ LLM์œผ๋กœ๋ถ€ํ„ฐ ์–ด๋–ป๊ฒŒ ํšจ๊ณผ์ ์œผ๋กœ distill ํ•˜๋Š”์ง€๋Š” ์•„์ง under-explore ๋˜์–ด ์žˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” forward KLD๋ฅผ reverse KLD๋กœ ๋Œ€์ฒดํ•จ์œผ๋กœ์จ ์ƒ์„ฑ์  larger LM์œผ๋กœ๋ถ€ํ„ฐ smaller LM์„ distill ํ•˜๋Š” MiniLLM์„ ์†Œ๊ฐœํ•˜์˜€๋‹ค. ์ด๊ฒƒ์€ student model์ด teacher ๋ถ„ํฌ์˜ low-probability ์˜์—ญ์„ ๊ณผ๋„ํ•˜๊ฒŒ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์œผ๋กœ๋ถ€ํ„ฐ ๋ชจ๋ธ์„ ๋ณดํ˜ธํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ƒ์„ฑ์  LM์— ๋”์šฑ ์ ํ•ฉํ•œ LM์ด๋‹ค. MiniLLM์€ ์ „๋ฐ˜์ ์œผ๋กœ ๋†’์€ ํ€„.. ๊ณต๊ฐ์ˆ˜ 2 ๋Œ“๊ธ€์ˆ˜ 0 2023. 6. 22.
  • Let's verify step-by-step ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ The overview of this paper ์ตœ๊ทผ ๋ช‡ ๋…„ ๋™์•ˆ LLM์€ ๋ณต์žกํ•œ multi-step ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•œ ๋Šฅ๋ ฅ์ด ์ƒ๋‹นํžˆ ๊ฐœ์„ ๋˜์—ˆ๋‹ค. ํ•˜์ง€๋งŒ, SoTA ๋ชจ๋ธ์€ ์•„์ง ๋…ผ๋ฆฌ์  ์˜ค๋ฅ˜๋ฅผ ๋งŒ๋“ค์–ด ๋‚ด๊ธฐ๋„ ํ•œ๋‹ค. ๋”์šฑ ์‹ ๋ขฐ๋„ ์žˆ๋Š” ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ตœ์ข… ๊ฒฐ๊ณผ์— ๋Œ€ํ•ด ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๋Š” outcome supervision์œผ๋กœ ์ „ํ™˜๋  ์ˆ˜ ์žˆ๋‹ค. ๋…ผ๋ฌธ์˜ ์‹คํ—˜์„ ํ†ตํ•ด ์–ด๋ ค์šด MATH ๋ฐ์ดํ„ฐ์…‹์˜ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด process supervision์ด outcome supervision์„ ์ƒ๋‹นํžˆ ๋Šฅ๊ฐ€ํ•˜๋Š” ๋ชจ์Šต์„ ๋ณด์—ฌ์คฌ๋‹ค. ๋˜ํ•œ active learning์ด process supervision์˜ ํšจํ—˜์„ ์ƒ๋‹นํžˆ ๊ฐœ์„ ์‹œํ‚จ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค€๋‹ค. ๊ทธ๋ฆฌ๊ณ  80๋งŒ ๊ฐœ์˜ step-level human feedback ๋ผ.. ๊ณต๊ฐ์ˆ˜ 3 ๋Œ“๊ธ€์ˆ˜ 1 2023. 6. 20.
  • ์ค‘์š”ํ•œ ๊ฑด ๊บพ์ด์ง€ ์•Š๋Š” high-quality data!! - Koala๐Ÿจ: A Dialogue Model for Academic Research ๋ฆฌ๋ทฐ Koala Overview Koala๋ฅผ ์†Œ๊ฐœํ•˜๋Š” ํฌ์ŠคํŠธ์—์„œ๋Š” ์›น์œผ๋กœ๋ถ€ํ„ฐ ์ˆ˜์ง‘๋œ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ์—์„œ Meta์˜ LLaMA๋ฅผ fine-tuning ํ•จ์œผ๋กœ์จ ํ•™์Šต๋œ ์ฑ—๋ด‡์ธ Koala๋ฅผ ์†Œ๊ฐœํ•˜์˜€๋‹ค. ๋ฐ์ดํ„ฐ์…‹ curation๊ณผ training process๋ฅผ ์„ค๋ช…ํ•˜๊ณ  Koala์™€ ChatGPT, Alpaca์™€ ๋น„๊ตํ•˜๋Š” ์‚ฌ์šฉ์ž ์—ฐ๊ตฌ์˜ ๊ฒฐ๊ณผ ๋˜ํ•œ ๋ณด์—ฌ์คฌ๋‹ค. Koala์˜ ๊ฒฐ๊ณผ๋Š” Koala๊ฐ€ ๋‹ค์–‘ํ•œ ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ์— ํšจ๊ณผ์ ์œผ๋กœ ์‘๋‹ตํ•  ์ˆ˜ ์žˆ๊ณ , ์‘๋‹ต ์ƒ์„ฑ๋„ Alpaca๋ณด๋‹ค ๋” ์„ ํ˜ธ๋˜์—ˆ๊ณ , ์ ˆ๋ฐ˜์ด ๋„˜๋Š” ๊ฒฝ์šฐ์— ์ตœ์†Œํ•œ ChatGPT์™€ ํƒ€์ด๋ฅผ ์ด๋ฃจ๋Š” ๋ชจ์Šต์„ ๋ณด์—ฌ์คฌ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ์ถฉ๋ถ„ํžˆ ์ž‘์€ ๋ชจ๋ธ๋„ ์‹ ์ค‘ํ•˜๊ฒŒ ๋ชจ์—ฌ์ง„ ๋ฐ์ดํ„ฐ์—์„œ ํ•™์Šต๋˜๋ฉด ์ด ๋ชจ๋ธ๋“ค์˜ ํฐ cousin ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋งŽ์ด ์บก์ฒ˜ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์ œ์•ˆํ•œ๋‹ค. ์ด๊ฒƒ์€ ์ปค๋ฎค๋‹ˆํ‹ฐ.. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 0 2023. 6. 19.
  • Vicuna๐Ÿช: An Open-Source Chatbot Impressing GPT-4 ๋ฆฌ๋ทฐ The overview of 'Vicuna' Vicuna 13B๋Š” ShareGPT๋กœ๋ถ€ํ„ฐ ์ˆ˜์ง‘๋œ user-shared ๋Œ€ํ™”์—์„œ fine-tuned LLaMA์—์„œ ํ•™์Šต๋œ open-source ์ฑ—๋ด‡์ด๋‹ค. GPT-4๋ฅผ ํ‰๊ฐ€์ž๋กœ ์‚ฌ์šฉํ•œ ์‚ฌ์ „ ํ‰๊ฐ€๋Š” Vicuna-13B๊ฐ€ OpenAI ChatGPT์™€ Google Bard์˜ 90%์— ํ•ด๋‹นํ•˜๋Š” ํ€„๋ฆฌํ‹ฐ๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ˜๋ฉด LLaMA์™€ Alpaca๋ณด๋‹ค 90%์˜ ๊ฒฝ์šฐ์— ๋” ๋‚˜์€ ๋ชจ์Šต์„ ๋ณด์—ฌ์คฌ๋‹ค. Vicuna-13B์˜ ํ•™์Šต ๋น„์šฉ์€ 300$ ์ •๋„์ด๋‹ค. ๊ทธ๋ฆฌ๊ณ  Vicuna์˜ ์ฝ”๋“œ์™€ ๊ฐ€์ค‘์น˜๋Š” ๋น„์ƒ์—…์  ์‚ฌ์šฉ์— ํ•œํ•ด์„œ ๊ณต๊ฐœ๋˜์—ˆ๋‹ค. How Good Is Vicuna? 70K user-shared ChatGPT ๋Œ€ํ™”๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Vicuna๋ฅผ fine-tuning ํ•œ ํ›„์—, Vicuna๋Š” Al.. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 1 2023. 6. 17.
  • imitation์ด ์ข‹์€ ํ•™์Šต ๋ฐฉ๋ฒ•์ผ๊นŒ? ๐Ÿค”: The False Promise of Imitating Proprietary LLMs ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ The overview of this paper weaker LM์„ ๊ฐœ์„ ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๊ฐ’์‹ผ method๋Š” stronger model์˜ output์—์„œ weaker LM์„ fine-tune ํ•˜๋Š” ๊ฒƒ์ด๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ๋ฒ•์€ weaker open-source ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์—…์  ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ๊ฐ’์‹ธ๊ฒŒ ํ‰๋‚ด ๋‚ด๋Š” ๋ฐฉ์‹์ฒ˜๋Ÿผ ๋ณด์ธ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์ด ์ ‘๊ทผ๋ฒ•์— ๋Œ€ํ•ด ๋ถ„์„ํ•˜์˜€๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ ์‚ฌ์ด์ฆˆ, ๋ฐ์ดํ„ฐ ์†Œ์Šค, ๋ชจ๋ฐฉ ๋ฐ์ดํ„ฐ์˜ ์–‘์„ ์‚ฌ์šฉํ•ด์„œ ChatGPT๋ฅผ ๋ชจ๋ฐฉํ•˜๋Š” LM์˜ ์‹œ๋ฆฌ์ฆˆ๋ฅผ fine-tune ํ•˜์˜€๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ด ๋ชจ๋ธ์„ crwodworker & NLP ๋ฒค์น˜๋งˆํฌ์—์„œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜์˜€๋‹ค. ์ดˆ๊ธฐ์— ๋…ผ๋ฌธ์—์„œ๋Š” ๋ชจ๋ฐฉ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ ํ€„๋ฆฌํ‹ฐ์— ๋Œ€ํ•ด ๋†€๋ž๋‹ค! ์™œ๋ƒํ•˜๋ฉด ๋ชจ๋ฐฉ ๋ชจ๋ธ์ด ์ถœ๋ ฅ์ด instruction์„ .. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 0 2023. 6. 16.
  • Open LLM Leaderboard๋ฅผ ํœฉ์“ด Falcon๐Ÿฆ… LLM: Falcon & RefinedWeb ์ตœ๊ทผ Hugging Face์˜ Open LLM Leaderboard๋ฅผ ๋‘˜๋Ÿฌ๋ณด๋˜ ์ค‘ ์ƒˆ๋กœ์šด ๋ชจ๋ธ์ด ๋ฆฌ๋”๋ณด๋“œ์˜ 1๋“ฑ์— ์œ„์น˜ํ•ด ์žˆ๋Š” ๊ฒƒ์„ ๋ณด๊ณ  '์–ด๋–ค ๋ชจ๋ธ์ด์ง€?'๋ผ๋Š” ๊ถ๊ธˆ์ฆ์ด ์ƒ๊ฒจ์„œ ์ด๋ ‡๊ฒŒ ํฌ์ŠคํŒ…์„ ์ž‘์„ฑํ•ด ๋ณธ๋‹ค. ์ƒˆ๋กญ๊ฒŒ 1๋“ฑ์„ ์ฐจ์ง€ํ•œ ๋ชจ๋ธ์€ ๋ฐ”๋กœ TII์—์„œ ๊ฐœ๋ฐœํ•œ Falcon๐Ÿฆ… ์ด๋ผ๋Š” ๋ชจ๋ธ์ด๋‹ค. Falcon์€ ์ด 4๊ฐ€์ง€ ๋ฒ„์ „์˜ ๋ชจ๋ธ์ด ์กด์žฌํ•˜๋Š”๋ฐ, 7B & 40B ์‚ฌ์ด์ฆˆ์˜ ๋ชจ๋ธ๊ณผ ๊ฐ ์‚ฌ์ด์ฆˆ์—์„œ ๊ทธ๋ƒฅ base ๋ฒ„์ „๊ณผ instruct-tuned ๋ฒ„์ „๊นŒ์ง€ ํ•ด์„œ 4๊ฐœ์ด๋‹ค. ๊ทธ์ค‘์— 40B ์‚ฌ์ด์ฆˆ์˜ instruct-tuned ๋ฒ„์ „์ธ 'falcon-40b-instruct'๊ฐ€ Leaderboard์—์„œ 1๋“ฑ์„ ์ฐจ์ง€ํ•˜์˜€๋‹ค. ์ด๋ฒˆ ํฌ์ŠคํŒ…์—์„œ๋Š” ์ด๋Ÿฌํ•œ Falcon ๋ชจ๋ธ์— ๋Œ€ํ•ด ์•Œ์•„๋ณด๊ณ  Falcon์„ ๋งŒ๋“œ๋Š” ๋ฐ ํฐ ๊ธฐ์—ฌ๋ฅผ ํ–ˆ๋˜ ๋ฐ์ดํ„ฐ.. ๊ณต๊ฐ์ˆ˜ 1 ๋Œ“๊ธ€์ˆ˜ 2 2023. 6. 14.
  • ๐ŸฒBaize: An Open-Source Chat Model with Parameter-Efficient Tuning on Self-Chat Data ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ The overview of this paper ChatGPT ๊ฐ™์€ chat ๋ชจ๋ธ๋“ค์€ ์ธ์ƒ์ ์ธ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋ฉด์„œ ๋น ๋ฅด๊ฒŒ ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์— ์ ์šฉ๋˜์–ด ๋‚˜๊ฐ€๊ณ  ์žˆ๋‹ค. ํ•˜์ง€๋งŒ, ์ œํ•œ๋œ API ๋•Œ๋ฌธ์— ์ƒˆ๋กœ์šด ์—ฐ๊ตฌ์— ์žฅ์• ๋ฌผ์„ ๋งŒ๋“ค๊ณ  ์žˆ๋‹ค. ๋…ผ๋ฌธ์—์„œ๋Š” ChatGPT๋ฅผ ๋Œ€ํ™”์— ์ฐธ์—ฌ์‹œํ‚ค๊ฒŒ ํ™œ์šฉํ•จ์œผ๋กœ์จ ์ž๋™์ ์œผ๋กœ high-quality multi-turn chat corpus๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ํŒŒ์ดํ”„๋ผ์ธ์„ ์ œ์•ˆํ•˜์˜€๋‹ค. ๊ทธ๋‹ค์Œ์— ์ด ๋ฐ์ดํ„ฐ๋“ค์„ parameter-efficient tuning์œผ๋กœ LLaMA๋ฅผ ํ–ฅ์ƒํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ํƒ„์ƒํ•œ ๋ชจ๋ธ์ด Baize์ด๊ณ , ์ด ๋ชจ๋ธ์€ ๊ฐ€๋“œ๋ ˆ์ผ์ด ์žˆ๋Š” multi-turn dialogue ์„ธํŒ…์—์„œ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค€๋‹ค. ๊ฒŒ๋‹ค๊ฐ€, ChatGPT์˜ ํ”ผ๋“œ๋ฐฑ์„ ์‚ฌ์šฉํ•˜์—ฌ Baize ๋ชจ๋ธ์˜ ์„ฑ.. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 1 2023. 6. 13.
  • Sparks of Artificial General Intelligence: Early experiments with GPT-4 ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ ์‹ค์ œ Sparks of AGI: with GPT-4 ๋…ผ๋ฌธ์€ 155ํŽ˜์ด์ง€์— ์œก๋ฐ•ํ•  ์ •๋„๋กœ ์—„์ฒญ๋‚œ ์–‘์˜ ์‹คํ—˜์„ ์ง„ํ–‰ํ•ด ๋ณด๋ฉฐ GPT-4๋ฅผ ๋‹ค๋ฐฉ๋ฉด์œผ๋กœ ํ™œ์šฉํ•ด ๋ณด์ง€๋งŒ, ๋ณธ ํฌ์ŠคํŒ…์—์„œ๋Š” ๊ทธ ๋งŽ์€ ๋‚ด์šฉ์„ ๋‹ค๋ฃจ๊ธฐ์—๋Š” ํž˜์ด ๋ฒ…์ฐจ์„œ ์ค‘์š” ๋ถ€๋ถ„๋“ค๋งŒ ๋”ฐ๋กœ ์‚ดํŽด๋ณด์•˜๋‹ค. ์ด ํฌ์ŠคํŒ…์€ ๋‹ค์Œ์˜ ์œ ํŠœ๋ธŒ๋ฅผ ์ฐธ๊ณ ํ•˜์—ฌ ์ž‘์„ฑ๋˜์—ˆ๋‹ค. ์œ ํŠœ๋ธŒ: https://www.youtube.com/watch?v=Mqg3aTGNxZ0 The overview of this paper AI ์—ฐ๊ตฌ์ž๋“ค์€ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ๊ณผ task์—์„œ ๊ด„๋ชฉํ•  ๋งŒํ•œ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋Š” LLM์„ ๊ฐœ๋ฐœํ•˜๊ณ  ๊ฐœ์„ ์‹œํ‚ค๊ณ  ์žˆ๋‹ค. OpenAI์—์„œ ๊ฐœ๋ฐœํ•œ GPT-4๋Š” ์ „๋ก€ ์—†๋Š” ๊ทœ๋ชจ์˜ ๊ณ„์‚ฐ๋Ÿ‰๊ณผ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•ด์„œ ํ•™์Šต๋˜์—ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” GPT-4๊ฐ€ ์ด์ „ AI ๋ชจ๋ธ๋ณด๋‹ค ๋” ์ผ๋ฐ˜์ ์ธ ์ง€๋Šฅ์„ ๋‚˜ํƒ€๋‚ด๋Š” ์ƒˆ.. ๊ณต๊ฐ์ˆ˜ 0 ๋Œ“๊ธ€์ˆ˜ 2 2023. 6. 12.
    ๋ฌธ์˜์•ˆ๋‚ด
    • ํ‹ฐ์Šคํ† ๋ฆฌ
    • ๋กœ๊ทธ์ธ
    • ๊ณ ๊ฐ์„ผํ„ฐ

    ํ‹ฐ์Šคํ† ๋ฆฌ๋Š” ์นด์นด์˜ค์—์„œ ์‚ฌ๋ž‘์„ ๋‹ด์•„ ๋งŒ๋“ญ๋‹ˆ๋‹ค.

    ยฉ Kakao Corp.