๐Ÿš€ Master Prompt Engineering and building AI Agents in our NEW courses! Use PROMPTING20 for 20% off โžœ Enroll now
LLaMA

LLaMA: ๊ฐœ๋ฐฉ์ ์ด๊ณ  ํšจ์œจ์ ์ธ ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ชจ๋ธ(Foundation Language Models)

โš ๏ธ

์ด ์„น์…˜์€ ํ˜„์žฌ ๊ฐœ๋ฐœ์ค‘์— ์žˆ์Šต๋‹ˆ๋‹ค.

์ƒˆ๋กœ์šด ๊ธฐ๋Šฅ

์ด ๋…ผ๋ฌธ์€ 70์–ต๊ฐœ์—์„œ 650์–ต๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์‚ฌ์ด์ฆˆ์˜ ๊ธฐ๋ฐ˜ ์–ธ์–ด ๋ชจ๋ธ(foundation language models)๋“ค์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.

์ด ๋ชจ๋ธ๋“ค์€ ๊ณต๊ฐœ๋œ ๋ฐ์ดํ„ฐ์…‹์—์„œ ์กฐ ๋‹จ์œ„ ๊ฐฏ์ˆ˜์˜ ํ† ํฐ์œผ๋กœ ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

(Hoffman et al. 2022) (opens in a new tab)์˜ ์—ฐ๊ตฌ๋Š” ๋” ๋งŽ์€ ๋ฐ์ดํ„ฐ์—์„œ ํ•™์Šต๋œ ์ž‘์€ ๋ชจ๋ธ์ด ๋ฐ˜๋Œ€ ๊ฒฝ์šฐ์˜ ๋” ํฐ ๋ชจ๋ธ๋ณด๋‹ค ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” 2000์–ต๊ฐœ ํ† ํฐ์—์„œ 100์–ต๊ฐœ ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์„ ๊ถŒ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LLaMA ๋…ผ๋ฌธ์—์„œ๋Š” 70์–ต๊ฐœ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ 1์กฐ๊ฐœ์˜ ํ† ํฐ ์ดํ›„์—๋„ ์ง€์†ํ•ด์„œ ํ–ฅ์ƒ๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.

LLAMA1

์ด ๋…ผ๋ฌธ์€ ๋‹ค์–‘ํ•œ ์ถ”๋ก  ํ™˜๊ฒฝ์—์„œ ๋” ๋งŽ์€ ํ† ํฐ์œผ๋กœ ํ•™์Šตํ•จ์œผ๋กœ์จ, ์ตœ์ƒ์˜ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ชจ๋ธ(LLaMA)์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

๋Šฅ๋ ฅ & ์ฃผ์š” ๊ฒฐ๊ณผ

์ „๋ฐ˜์ ์œผ๋กœ, LLaMA-13B๋Š” GPT-3(175B)๋ณด๋‹ค 10๋ฐฐ ์ž‘์ง€๋งŒ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋ฉฐ, ๋‹จ์ผ GPU์—์„œ๋„ ์ž‘๋™์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. LLaMA 65B๋Š” Chinchilla-70B ๋ฐ PaLM-540B ๊ฐ™์€ ๋ชจ๋ธ๋“ค๊ณผ ๊ฒฝ์Ÿ๋ ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค.

๋…ผ๋ฌธ: LLaMA: Open and Efficient Foundation Language Models (opens in a new tab)

์ฝ”๋“œ: https://github.com/facebookresearch/llama (opens in a new tab)

์ฐธ๊ณ ์ž๋ฃŒ (References)