eng
stringlengths
62
4.5k
ko1
stringlengths
22
8.19k
ko2
stringlengths
23
8.18k
ko3
stringlengths
36
8.7k
ko4
stringlengths
21
28k
ko5
stringlengths
33
11.2k
\(\bullet\)_Webpages._ Owing to the proliferation of the Internet, various types of data have been created, which enables LLMs to gain diverse linguistic knowledge and enhance their generalization capabilities [82, 26]. For convenient use of these data resources, a large amount of data is crawled from the web in previous work, such as CommonCrawl [163]. However, the crawled web data tends to contain both high-quality text, such as Wikipedia and low-quality text, like spam mail, thus it is important to filter and process webpages for improving the data quality.
\(\bullet\)_์›นํŽ˜์ด์ง€._ ์ธํ„ฐ๋„ท์˜ ํ™•์‚ฐ์œผ๋กœ ์ธํ•ด ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ์ƒ์„ฑ๋˜์–ด LLM์ด ๋‹ค์–‘ํ•œ ์–ธ์–ด ์ง€์‹์„ ์Šต๋“ํ•˜๊ณ  ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค [82, 26]. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ๋ฆฌ์†Œ์Šค๋ฅผ ํŽธ๋ฆฌํ•˜๊ฒŒ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์ด์ „ ์ž‘์—…์—์„œ๋Š” CommonCrawl [163]๊ณผ ๊ฐ™์ด ๋Œ€๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์›น์—์„œ ํฌ๋กค๋งํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํฌ๋กค๋ง๋œ ์›น ๋ฐ์ดํ„ฐ์—๋Š” ์œ„ํ‚คํ”ผ๋””์•„์™€ ๊ฐ™์€ ๊ณ ํ’ˆ์งˆ ํ…์ŠคํŠธ์™€ ์ŠคํŒธ ๋ฉ”์ผ๊ณผ ๊ฐ™์€ ์ €ํ’ˆ์งˆ ํ…์ŠคํŠธ๊ฐ€ ๋ชจ๋‘ ํฌํ•จ๋˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฏ€๋กœ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์›นํŽ˜์ด์ง€๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ณ  ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_์›นํŽ˜์ด์ง€._ ์ธํ„ฐ๋„ท์˜ ํ™•์‚ฐ์œผ๋กœ ์ธํ•ด ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ์ƒ์„ฑ๋˜์–ด LLM์ด ๋‹ค์–‘ํ•œ ์–ธ์–ด์  ์ง€์‹์„ ์Šต๋“ํ•˜๊ณ  ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค [82, 26]. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ๋ฆฌ์†Œ์Šค๋ฅผ ํŽธ๋ฆฌํ•˜๊ฒŒ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์ด์ „ ์ž‘์—…์—์„œ๋Š” CommonCrawl [163]๊ณผ ๊ฐ™์ด ๋Œ€๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์›น์—์„œ ํฌ๋กค๋งํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํฌ๋กค๋ง๋œ ์›น ๋ฐ์ดํ„ฐ์—๋Š” ์œ„ํ‚คํ”ผ๋””์•„์™€ ๊ฐ™์€ ๊ณ ํ’ˆ์งˆ ํ…์ŠคํŠธ์™€ ์ŠคํŒธ ๋ฉ”์ผ๊ณผ ๊ฐ™์€ ์ €ํ’ˆ์งˆ ํ…์ŠคํŠธ๊ฐ€ ๋ชจ๋‘ ํฌํ•จ๋˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฏ€๋กœ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์›นํŽ˜์ด์ง€๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ณ  ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_์›นํŽ˜์ด์ง€._ ์ธํ„ฐ๋„ท์˜ ํ™•์‚ฐ์œผ๋กœ ์ธํ•ด ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ์ƒ์„ฑ๋˜์–ด LLM์ด ๋‹ค์–‘ํ•œ ์–ธ์–ด์  ์ง€์‹์„ ์Šต๋“ํ•˜๊ณ  ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[82, 26]. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ๋ฆฌ์†Œ์Šค๋ฅผ ํŽธ๋ฆฌํ•˜๊ฒŒ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์ด์ „ ์ž‘์—…์—์„œ๋Š” CommonCrawl[163]๊ณผ ๊ฐ™์ด ์›น์—์„œ ๋Œ€๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํฌ๋กค๋งํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํฌ๋กค๋ง๋œ ์›น ๋ฐ์ดํ„ฐ์—๋Š” ์œ„ํ‚คํ”ผ๋””์•„์™€ ๊ฐ™์€ ๊ณ ํ’ˆ์งˆ ํ…์ŠคํŠธ์™€ ์ŠคํŒธ ๋ฉ”์ผ๊ณผ ๊ฐ™์€ ์ €ํ’ˆ์งˆ ํ…์ŠคํŠธ๊ฐ€ ๋ชจ๋‘ ํฌํ•จ๋˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฏ€๋กœ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์›น ํŽ˜์ด์ง€๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ณ  ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
\(\์ด์•Œ\)_์›นํŽ˜์ด์ง€._ ์ธํ„ฐ๋„ท์˜ ํ™•์‚ฐ์œผ๋กœ ์ธํ•ด ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ์ƒ์„ฑ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋กœ ์ธํ•ด LLM์€ ๋‹ค์–‘ํ•œ ์–ธ์–ด ์ง€์‹์„ ์Šต๋“ํ•˜๊ณ  ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [82, 26]. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ๋ฆฌ์†Œ์Šค๋ฅผ ํŽธ๋ฆฌํ•˜๊ฒŒ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์ด์ „ ์ž‘์—…์—์„œ๋Š” CommonCrawl [163]๊ณผ ๊ฐ™์ด ์›น์—์„œ ๋Œ€๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํฌ๋กค๋งํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํฌ๋กค๋ง๋œ ์›น ๋ฐ์ดํ„ฐ์—๋Š” Wikipedia์™€ ๊ฐ™์€ ๊ณ ํ’ˆ์งˆ ํ…์ŠคํŠธ์™€ ์ŠคํŒธ ๋ฉ”์ผ๊ณผ ๊ฐ™์€ ์ €ํ’ˆ์งˆ ํ…์ŠคํŠธ๊ฐ€ ๋ชจ๋‘ ํฌํ•จ๋˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฏ€๋กœ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์›นํŽ˜์ด์ง€๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ณ  ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
\(\์ด์•Œ\)_์›นํŽ˜์ด์ง€._ ์ธํ„ฐ๋„ท์˜ ํ™•์‚ฐ์œผ๋กœ ์ธํ•ด ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ์ƒ์„ฑ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด LLM์€ ๋‹ค์–‘ํ•œ ์–ธ์–ด ์ง€์‹์„ ์Šต๋“ํ•˜๊ณ  ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [82, 26]. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ๋ฆฌ์†Œ์Šค๋ฅผ ํŽธ๋ฆฌํ•˜๊ฒŒ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์ด์ „ ์ž‘์—…์—์„œ๋Š” CommonCrawl [163]๊ณผ ๊ฐ™์ด ์›น์—์„œ ๋Œ€๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํฌ๋กค๋งํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํฌ๋กค๋ง๋œ ์›น ๋ฐ์ดํ„ฐ์—๋Š” Wikipedia์™€ ๊ฐ™์€ ๊ณ ํ’ˆ์งˆ ํ…์ŠคํŠธ์™€ ์ŠคํŒธ ๋ฉ”์ผ๊ณผ ๊ฐ™์€ ์ €ํ’ˆ์งˆ ํ…์ŠคํŠธ๊ฐ€ ๋ชจ๋‘ ํฌํ•จ๋˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฏ€๋กœ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์›น ํŽ˜์ด์ง€๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ณ  ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
1. We curate the largest, most recent dataset of real-world forecasting questions to date, for evaluating and optimizing automated forecasting systems. 2. We build a retrieval-augmented LM system that significantly improves upon the baseline and approaches the human crowd performance on competitive forecasting platforms. 3. We propose and apply a self-supervised fine-tuning method to improve LM's capability in reasoning about forecasting tasks.
1. ์ž๋™ ์˜ˆ์ธก ์‹œ์Šคํ…œ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ˜„์žฌ๊นŒ์ง€ ๊ฐ€์žฅ ํฌ๊ณ  ์ตœ์‹ ์˜ ์‹ค์ œ ์˜ˆ์ธก ์งˆ๋ฌธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค.2. ๊ฒฝ์Ÿ ์˜ˆ์ธก ํ”Œ๋žซํผ์—์„œ ๊ธฐ์ค€์„ ๋ณด๋‹ค ํฌ๊ฒŒ ๊ฐœ์„ ๋˜๊ณ  ์ธ๊ฐ„ ํฌ๋ผ์šฐ๋“œ ์„ฑ๋Šฅ์— ๊ทผ์ ‘ํ•˜๋Š” ๊ฒ€์ƒ‰ ์ฆ๊ฐ• LM ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค.3. ์˜ˆ์ธก ์ž‘์—…์— ๋Œ€ํ•œ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ž๊ธฐ ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ • ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•˜๊ณ  ์ ์šฉํ•ฉ๋‹ˆ๋‹ค.
1. ์ž๋™ ์˜ˆ์ธก ์‹œ์Šคํ…œ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ˜„์žฌ๊นŒ์ง€ ๊ฐ€์žฅ ํฌ๊ณ  ์ตœ์‹ ์˜ ์‹ค์ œ ์˜ˆ์ธก ์งˆ๋ฌธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค.2. ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์˜ˆ์ธก ํ”Œ๋žซํผ์—์„œ ๊ธฐ์ค€์„ ๋ณด๋‹ค ํฌ๊ฒŒ ๊ฐœ์„ ๋˜๊ณ  ์ธ๊ฐ„ ๊ตฐ์ค‘์˜ ์„ฑ๋Šฅ์— ๊ทผ์ ‘ํ•˜๋Š” ๊ฒ€์ƒ‰ ๋ณด๊ฐ• LM ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค.3. ์˜ˆ์ธก ์ž‘์—…์— ๋Œ€ํ•œ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ž์ฒด ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ • ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•˜๊ณ  ์ ์šฉํ•ฉ๋‹ˆ๋‹ค.
1. ์ž๋™ ์˜ˆ์ธก ์‹œ์Šคํ…œ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ˜„์žฌ๊นŒ์ง€ ๊ฐ€์žฅ ํฌ๊ณ  ๊ฐ€์žฅ ์ตœ๊ทผ์˜ ์‹ค์ œ ์˜ˆ์ธก ์งˆ๋ฌธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค.2. ๊ฒฝ์Ÿ ์˜ˆ์ธก ํ”Œ๋žซํผ์—์„œ ๊ธฐ์ค€์„ ์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•˜๊ณ  ์ธ๊ฐ„ ํฌ๋ผ์šฐ๋“œ ์„ฑ๋Šฅ์— ๊ทผ์ ‘ํ•˜๋Š” ๊ฒ€์ƒ‰ ๋ณด๊ฐ• LM ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค.3. ์˜ˆ์ธก ์ž‘์—…์— ๋Œ€ํ•œ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ž๊ฐ€ ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ • ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•˜๊ณ  ์ ์šฉํ•ฉ๋‹ˆ๋‹ค.
1. ์ž๋™ํ™”๋œ ์˜ˆ์ธก ์‹œ์Šคํ…œ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ˜„์žฌ๊นŒ์ง€ ๊ฐ€์žฅ ํฌ๊ณ  ์ตœ์‹ ์˜ ์‹ค์ œ ์˜ˆ์ธก ์งˆ๋ฌธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค.2. ๊ฒฝ์Ÿ ์˜ˆ์ธก ํ”Œ๋žซํผ์—์„œ ์ธ๊ฐ„์˜ ๊ตฐ์ค‘ ์„ฑ๋Šฅ์— ๊ทผ์ ‘ํ•˜๋Š” ๊ธฐ๋ณธ๊ฐ’์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•œ ๊ฒ€์ƒ‰ ๊ฐ•ํ™” LM ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค.3. ์˜ˆ์ธก ์ž‘์—…์— ๋Œ€ํ•œ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ž์ฒด ๊ฐ๋… ๋ฏธ์„ธ ์กฐ์ • ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•˜๊ณ  ์ ์šฉํ•ฉ๋‹ˆ๋‹ค.
1. ์ž๋™ํ™”๋œ ์˜ˆ์ธก ์‹œ์Šคํ…œ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ˜„์žฌ๊นŒ์ง€ ๊ฐ€์žฅ ๋ฐฉ๋Œ€ํ•œ ์ตœ์‹  ์‹ค์ œ ์˜ˆ์ธก ์งˆ๋ฌธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ๋ ˆ์ด์…˜ํ•ฉ๋‹ˆ๋‹ค.2. ๊ฒฝ์Ÿ ์˜ˆ์ธก ํ”Œ๋žซํผ์—์„œ ๊ธฐ์ค€์„ ์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•˜๊ณ  ์ธ๊ฐ„ ๊ตฐ์ค‘์˜ ์„ฑ๊ณผ์— ๊ทผ์ ‘ํ•˜๋Š” ๊ฒ€์ƒ‰ ๋ณด๊ฐ• LM ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค.3. ์˜ˆ์ธก ์ž‘์—…์— ๋Œ€ํ•œ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ž๊ธฐ ๊ฐ๋… ๋ฏธ์„ธ ์กฐ์ • ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•˜๊ณ  ์ ์šฉํ•ฉ๋‹ˆ๋‹ค.
PersonA: Vascepa is used together with a low-fat diet and other treatments to lower total cholesterol in people with severe high cholesterol.
PersonA: ๋ฐ”์Šค์„ธํŒŒ๋Š” ๊ณ ์ฝœ๋ ˆ์Šคํ…Œ๋กคํ˜ˆ์ฆ์ด ์‹ฌํ•œ ์‚ฌ๋žŒ์˜ ์ด ์ฝœ๋ ˆ์Šคํ…Œ๋กค์„ ๋‚ฎ์ถ”๊ธฐ ์œ„ํ•ด ์ €์ง€๋ฐฉ ์‹๋‹จ ๋ฐ ๊ธฐํƒ€ ์น˜๋ฃŒ์ œ์™€ ํ•จ๊ป˜ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
PersonA: ๋ฐ”์Šค์„ธํŒŒ๋Š” ์ €์ง€๋ฐฉ ์‹์ด์š”๋ฒ• ๋ฐ ๊ธฐํƒ€ ์น˜๋ฃŒ์ œ์™€ ํ•จ๊ป˜ ๊ณ ์ฝœ๋ ˆ์Šคํ…Œ๋กคํ˜ˆ์ฆ์ด ์‹ฌํ•œ ์‚ฌ๋žŒ์˜ ์ด ์ฝœ๋ ˆ์Šคํ…Œ๋กค์„ ๋‚ฎ์ถ”๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ๋ฐ”์Šค์„ธํŒŒ๋Š” ๊ณ ์ฝœ๋ ˆ์Šคํ…Œ๋กคํ˜ˆ์ฆ์ด ์‹ฌํ•œ ์‚ฌ๋žŒ๋“ค์˜ ์ด ์ฝœ๋ ˆ์Šคํ…Œ๋กค ์ˆ˜์น˜๋ฅผ ๋‚ฎ์ถ”๊ธฐ ์œ„ํ•ด ์ €์ง€๋ฐฉ ์‹์ด์š”๋ฒ• ๋ฐ ๊ธฐํƒ€ ์น˜๋ฃŒ์™€ ํ•จ๊ป˜ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ๋ฐ”์„ธํŒŒ๋Š” ์ €์ง€๋ฐฉ ์‹๋‹จ ๋ฐ ๊ธฐํƒ€ ์น˜๋ฃŒ๋ฒ•๊ณผ ํ•จ๊ป˜ ์ค‘์ฆ ๊ณ ์ฝœ๋ ˆ์Šคํ…Œ๋กคํ˜ˆ์ฆ ํ™˜์ž์˜ ์ด ์ฝœ๋ ˆ์Šคํ…Œ๋กค์„ ๋‚ฎ์ถ”๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ๋ฐ”์„ธํŽ˜์•„๋Š” ์‹ฌ๊ฐํ•œ ๊ณ ์ฝœ๋ ˆ์Šคํ…Œ๋กคํ˜ˆ์ฆ ํ™˜์ž์˜ ์ด ์ฝœ๋ ˆ์Šคํ…Œ๋กค์„ ๋‚ฎ์ถ”๊ธฐ ์œ„ํ•ด ์ €์ง€๋ฐฉ ์‹๋‹จ ๋ฐ ๊ธฐํƒ€ ์น˜๋ฃŒ๋ฒ•๊ณผ ํ•จ๊ป˜ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
* **Natural Questions**[179] is a QA dataset that consists of real anonymized, aggregated queries submitted to the Google search engine as questions. An annotator is presented with a question along with a Wikipedia page from the top \(5\) search results, and annotates a long answer (typically a paragraph) and a short answer (one or more entities) if present on the page, or marks null if no long/short answer is present. * **MMLU**[180] is intended to evaluate the knowledge gained in zero-shot and few-shot scenarios. That means that MMLU assesses both the general knowledge and problem-solving ability of a model. It covers 57 subjects in STEM, humanities, social sciences, and other areas. The benchmark varies in complexity, ranging from elementary to advanced professional. It is worth mentioning that the main contribution of this dataset is for multi-task language understanding, question answering, and arithmetic reasoning. * **MBPP**[181] stands for "Mostly Basic Python Problems" and provides a benchmark for evaluating the performance of models designed for code generation. The benchmark encompasses \(974\) short Python programs including a wide range of topics, including fundamental programming concepts and standard library usage, and more. Each challenge comprises a task description, a code solution, and three automated test cases. * **HumanEval**[182] is a dataset for code generation task. This dataset consists of \(164\) hand-crafted programming challenges. Each challenge is accompanied by a function signature, docstring, code body, and multiple unit tests. The main intuition behind developing this dataset is to guarantee the exclusion of its contents from training datasets for code generation models. * **APPS**[183] is designed for code generation task focusing on the Python programming language. The APPS dataset contains a collection of \(232,444\) Python programs. Each program in the dataset has an average of \(18\) lines of Python code. Additionally, APPS offers access to a repository of \(10,000\) unique programming exercises, each with text-based problem descriptions. The final aspect to highlight is that the it includes test cases. * **WikiSQL**[184] is crafted for code generation task and it has 87,726 carefully labeled pairs of SQL queries and corresponding natural language questions from Wikipedia tables. The SQL queries comprise three subsets: test sets (\(17,284\) examples), development (\(9,145\) examples), and training (\(61,297\) examples). * **TriviaQA**[185] is designed for QA task. This dataset comprises more than \(650,000\) question-answer-evidence triples. There are \(95,000\) question-answer pairs in this dataset, each authored by trivia enthusiasts and supported by an average of six independently sourced evidence documents. These documents are automatically acquired from Wikipedia or broader web search results. The dataset is categorized into two segments, including those with authentic answers from Wikipedia and web domains, and verified sets embody the accurately answered questions along with their associated documents from both Wikipedia and online.
* **์ž์—ฐ์–ด ์งˆ๋ฌธ**[179]์€ Google ๊ฒ€์ƒ‰ ์—”์ง„์— ์ œ์ถœ๋œ ์‹ค์ œ ์ต๋ช…ํ™”๋œ ์ง‘๊ณ„๋œ ์ฟผ๋ฆฌ๋ฅผ ์งˆ๋ฌธ์œผ๋กœ ๊ตฌ์„ฑํ•œ QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ฃผ์„ ์ž‘์„ฑ์ž๋Š” ์ƒ์œ„ \(5\) ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์˜ ์œ„ํ‚ค๋ฐฑ๊ณผ ํŽ˜์ด์ง€์™€ ํ•จ๊ป˜ ์งˆ๋ฌธ์ด ์ œ์‹œ๋˜๊ณ , ํŽ˜์ด์ง€์— ๊ธด ๋‹ต(์ผ๋ฐ˜์ ์œผ๋กœ ๋‹จ๋ฝ)๊ณผ ์งง์€ ๋‹ต(ํ•˜๋‚˜ ์ด์ƒ์˜ ์—”ํ‹ฐํ‹ฐ)์ด ์žˆ๋Š” ๊ฒฝ์šฐ ์ฃผ์„์„ ๋‹ฌ๊ฑฐ๋‚˜, ๊ธด/์งง์€ ๋‹ต์ด ์—†๋Š” ๊ฒฝ์šฐ null์„ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. * **MMLU**[180]๋Š” ์ œ๋กœ์ƒท ๋ฐ ํŒŒ์šฐ์ƒท ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์–ป์€ ์ง€์‹์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ฆ‰, MMLU๋Š” ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ ์ง€์‹๊ณผ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ๋ชจ๋‘ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” STEM, ์ธ๋ฌธํ•™, ์‚ฌํšŒ๊ณผํ•™ ๋ฐ ๊ธฐํƒ€ ๋ถ„์•ผ์˜ 57๊ฐœ ๊ณผ๋ชฉ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” ๊ธฐ๋ณธ๋ถ€ํ„ฐ ๊ณ ๊ธ‰ ์ „๋ฌธ๊ฐ€์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋ณต์žก์„ฑ์ด ๋‹ค์–‘ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ฃผ์š” ๊ธฐ์—ฌ๋Š” ๋‹ค์ค‘ ์ž‘์—… ์–ธ์–ด ์ดํ•ด, ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ ์‚ฐ์ˆ  ์ถ”๋ก ์„ ์œ„ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. * **MBPP**[181]๋Š” "๋Œ€๋ถ€๋ถ„ ๊ธฐ๋ณธ ํŒŒ์ด์ฌ ๋ฌธ์ œ"์˜ ์•ฝ์ž๋กœ ์ฝ”๋“œ ์ƒ์„ฑ์„ ์œ„ํ•ด ์„ค๊ณ„๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” ๊ธฐ๋ณธ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ฐœ๋… ๋ฐ ํ‘œ์ค€ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์‚ฌ์šฉ๋ฒ• ๋“ฑ ๋‹ค์–‘ํ•œ ์ฃผ์ œ๋ฅผ ํฌํ•จํ•˜๋Š” \(974\)๊ฐœ์˜ ์งง์€ Python ํ”„๋กœ๊ทธ๋žจ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๊ณผ์ œ๋Š” ์ž‘์—… ์„ค๋ช…, ์ฝ”๋“œ ์†”๋ฃจ์…˜, ์„ธ ๊ฐ€์ง€ ์ž๋™ํ™”๋œ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. * **HumanEval**[182]์€ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” \(164\) ๊ฐœ์˜ ์ˆ˜์ž‘์—…์œผ๋กœ ์ œ์ž‘๋œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ณผ์ œ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๊ณผ์ œ์—๋Š” ํ•จ์ˆ˜ ์„œ๋ช…, ๋ฌธ์„œ ๋ฌธ์ž์—ด, ์ฝ”๋“œ ๋ณธ๋ฌธ ๋ฐ ์—ฌ๋Ÿฌ ๋‹จ์œ„ ํ…Œ์ŠคํŠธ๊ฐ€ ์ฒจ๋ถ€๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ฐœ๋ฐœํ•œ ์ฃผ์š” ๊ฐœ๋…์€ ์ฝ”๋“œ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ•ด๋‹น ๋‚ด์šฉ์ด ์ œ์™ธ๋˜๋„๋ก ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. * **APPS**[183]๋Š” Python ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์— ์ค‘์ ์„ ๋‘” ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. APPS ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” \(232,444\) ๊ฐœ์˜ Python ํ”„๋กœ๊ทธ๋žจ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ํ”„๋กœ๊ทธ๋žจ์€ ํ‰๊ท ์ ์œผ๋กœ \(18\) ์ค„์˜ Python ์ฝ”๋“œ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ APPS๋Š” ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ๋ฌธ์ œ ์„ค๋ช…์ด ์žˆ๋Š” \(10,000\) ๊ฐœ์˜ ๊ณ ์œ  ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์—ฐ์Šต ๋ฌธ์ œ ๋ฆฌํฌ์ง€ํ† ๋ฆฌ์— ๋Œ€ํ•œ ์•ก์„ธ์Šค๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•˜๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. * **WikiSQL**[184]์€ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•ด ์ œ์ž‘๋˜์—ˆ์œผ๋ฉฐ, ์œ„ํ‚ค๋ฐฑ๊ณผ ํ…Œ์ด๋ธ”์—์„œ 87,726๊ฐœ์˜ ์‹ ์ค‘ํ•˜๊ฒŒ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ SQL ์ฟผ๋ฆฌ์™€ ํ•ด๋‹น ์ž์—ฐ์–ด ์งˆ๋ฌธ ์Œ์„ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. SQL ์ฟผ๋ฆฌ๋Š” ํ…Œ์ŠคํŠธ ์„ธํŠธ(\(17,284\) ์˜ˆ์ œ), ๊ฐœ๋ฐœ(\(9,145\) ์˜ˆ์ œ), ํ›ˆ๋ จ(\(61,297\) ์˜ˆ์ œ)์˜ ์„ธ ๊ฐ€์ง€ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. * **TriviaQA**[185]๋Š” QA ์ž‘์—…์„ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 650,000๊ฐœ ์ด์ƒ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€-์ฆ๊ฑฐ ์‚ผ๊ฐํ˜•์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” 95,000๊ฐœ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์ด ์žˆ์œผ๋ฉฐ, ๊ฐ ์Œ์€ ํ€ด์ฆˆ ์• ํ˜ธ๊ฐ€๋“ค์ด ์ž‘์„ฑํ–ˆ์œผ๋ฉฐ ํ‰๊ท  6๊ฐœ์˜ ๋…๋ฆฝ์ ์œผ๋กœ ์†Œ์‹ฑ๋œ ์ฆ๊ฑฐ ๋ฌธ์„œ๋กœ ๋’ท๋ฐ›์นจ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์„œ๋Š” ์œ„ํ‚ค๋ฐฑ๊ณผ ๋˜๋Š” ๋” ๋„“์€ ์›น ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์—์„œ ์ž๋™์œผ๋กœ ํš๋“๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์œ„ํ‚ค๋ฐฑ๊ณผ ๋ฐ ์›น ๋„๋ฉ”์ธ์˜ ์ •๋‹ต์ด ์žˆ๋Š” ์„ธ๊ทธ๋จผํŠธ์™€ ์œ„ํ‚ค๋ฐฑ๊ณผ ๋ฐ ์˜จ๋ผ์ธ์˜ ๊ด€๋ จ ๋ฌธ์„œ์™€ ํ•จ๊ป˜ ์ •ํ™•ํ•˜๊ฒŒ ๋‹ตํ•œ ์งˆ๋ฌธ์ด ํฌํ•จ๋œ ํ™•์ธ ์„ธ๊ทธ๋จผํŠธ์˜ ๋‘ ๊ฐ€์ง€๋กœ ๋ถ„๋ฅ˜๋ฉ๋‹ˆ๋‹ค.
* **์ž์—ฐ์–ด ์งˆ๋ฌธ**[179]์€ Google ๊ฒ€์ƒ‰ ์—”์ง„์— ์ œ์ถœ๋œ ์‹ค์ œ ์ต๋ช…ํ™”๋œ ์ง‘๊ณ„๋œ ์ฟผ๋ฆฌ๋ฅผ ์งˆ๋ฌธ์œผ๋กœ ๊ตฌ์„ฑํ•œ QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ฃผ์„ ์ž‘์„ฑ์ž๋Š” ์ƒ์œ„ \(5\) ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์˜ ์œ„ํ‚ค๋ฐฑ๊ณผ ํŽ˜์ด์ง€์™€ ํ•จ๊ป˜ ์งˆ๋ฌธ์ด ์ œ์‹œ๋˜๊ณ , ํŽ˜์ด์ง€์— ๊ธด ๋‹ต๋ณ€(์ผ๋ฐ˜์ ์œผ๋กœ ๋‹จ๋ฝ)๊ณผ ์งง์€ ๋‹ต๋ณ€(ํ•˜๋‚˜ ์ด์ƒ์˜ ์—”ํ‹ฐํ‹ฐ)์ด ์žˆ๋Š” ๊ฒฝ์šฐ ์ฃผ์„์„ ๋‹ฌ๊ฑฐ๋‚˜ ๊ธด/์งง์€ ๋‹ต๋ณ€์ด ์—†๋Š” ๊ฒฝ์šฐ null์„ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. * **MMLU**[180]๋Š” ์ œ๋กœ์ƒท ๋ฐ ํŒŒ์šฐ์ƒท ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์–ป์€ ์ง€์‹์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ฆ‰, MMLU๋Š” ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ ์ง€์‹๊ณผ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ๋ชจ๋‘ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. STEM, ์ธ๋ฌธํ•™, ์‚ฌํšŒ ๊ณผํ•™ ๋ฐ ๊ธฐํƒ€ ๋ถ„์•ผ์˜ 57๊ฐœ ๊ณผ๋ชฉ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” ์ดˆ๊ธ‰๋ถ€ํ„ฐ ๊ณ ๊ธ‰ ์ „๋ฌธ๊ฐ€๊นŒ์ง€ ๋‹ค์–‘ํ•˜๊ฒŒ ๋ณต์žก๋„๊ฐ€ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ฃผ์š” ๊ณตํ—Œ์€ ๋‹ค์ค‘ ์ž‘์—… ์–ธ์–ด ์ดํ•ด, ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ ์‚ฐ์ˆ  ์ถ”๋ก ์„ ์œ„ํ•œ ๊ฒƒ์ž„์„ ์–ธ๊ธ‰ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. * **MBPP**[181]๋Š” "๋Œ€๋ถ€๋ถ„ ๊ธฐ๋ณธ ํŒŒ์ด์ฌ ๋ฌธ์ œ"์˜ ์•ฝ์ž๋กœ ์ฝ”๋“œ ์ƒ์„ฑ์„ ์œ„ํ•ด ์„ค๊ณ„๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” ๊ธฐ๋ณธ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ฐœ๋… ๋ฐ ํ‘œ์ค€ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์‚ฌ์šฉ๋ฒ• ๋“ฑ ๋‹ค์–‘ํ•œ ์ฃผ์ œ๋ฅผ ํฌํ•จํ•˜๋Š” 974๊ฐœ์˜ ์งง์€ Python ํ”„๋กœ๊ทธ๋žจ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๋„์ „ ๊ณผ์ œ๋Š” ์ž‘์—… ์„ค๋ช…, ์ฝ”๋“œ ์†”๋ฃจ์…˜, ์„ธ ๊ฐ€์ง€ ์ž๋™ํ™”๋œ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. * **HumanEval**[182]์€ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 164๊ฐœ์˜ ์ˆ˜์ž‘์—…์œผ๋กœ ์ œ์ž‘๋œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ณผ์ œ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๊ณผ์ œ์—๋Š” ํ•จ์ˆ˜ ์„œ๋ช…, ๋ฌธ์„œ ๋ฌธ์ž์—ด, ์ฝ”๋“œ ๋ณธ๋ฌธ, ์—ฌ๋Ÿฌ ๋‹จ์œ„ ํ…Œ์ŠคํŠธ๊ฐ€ ์ฒจ๋ถ€๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ฐœ๋ฐœํ•œ ์ฃผ์š” ๊ฐœ๋…์€ ์ฝ”๋“œ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ•ด๋‹น ์ฝ˜ํ…์ธ ๋ฅผ ์ œ์™ธํ•  ์ˆ˜ ์žˆ๋„๋ก ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. * **APPS**[183]๋Š” Python ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์— ์ค‘์ ์„ ๋‘” ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. APPS ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” 232,444๊ฐœ์˜ Python ํ”„๋กœ๊ทธ๋žจ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ํ”„๋กœ๊ทธ๋žจ์€ ํ‰๊ท  18์ค„์˜ Python ์ฝ”๋“œ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ APPS๋Š” ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ๋ฌธ์ œ ์„ค๋ช…์ด ์žˆ๋Š” 10,000๊ฐœ์˜ ๊ณ ์œ ํ•œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์—ฐ์Šต ๋ฌธ์ œ ๋ฆฌํฌ์ง€ํ† ๋ฆฌ์— ๋Œ€ํ•œ ์•ก์„ธ์Šค๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ๊ฐ•์กฐํ•˜๊ณ  ์‹ถ์€ ์ ์€ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๋ฅผ ํฌํ•จํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. * **WikiSQL**[184]์€ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•ด ์ œ์ž‘๋˜์—ˆ์œผ๋ฉฐ, ์œ„ํ‚ค๋ฐฑ๊ณผ ํ…Œ์ด๋ธ”์—์„œ 87,726๊ฐœ์˜ ์‹ ์ค‘ํ•˜๊ฒŒ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ SQL ์ฟผ๋ฆฌ์™€ ํ•ด๋‹น ์ž์—ฐ์–ด ์งˆ๋ฌธ ์Œ์„ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. SQL ์ฟผ๋ฆฌ๋Š” ํ…Œ์ŠคํŠธ ์„ธํŠธ(\(17,284\) ์˜ˆ์ œ), ๊ฐœ๋ฐœ(\(9,145\) ์˜ˆ์ œ), ํ•™์Šต(\(61,297\) ์˜ˆ์ œ)์˜ ์„ธ ๊ฐ€์ง€ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. * **TriviaQA**[185]๋Š” QA ์ž‘์—…์„ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 650,000๊ฐœ ์ด์ƒ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€-์ฆ๊ฑฐ ์‚ผ๊ฐํ˜•์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” 95,000๊ฐœ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์ด ์žˆ์œผ๋ฉฐ, ๊ฐ ์งˆ๋ฌธ์€ ํ€ด์ฆˆ ์• ํ˜ธ๊ฐ€๋“ค์ด ์ž‘์„ฑํ–ˆ์œผ๋ฉฐ ํ‰๊ท  6๊ฐœ์˜ ๋…๋ฆฝ์ ์œผ๋กœ ์†Œ์‹ฑ๋œ ์ฆ๊ฑฐ ๋ฌธ์„œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์„œ๋Š” ์œ„ํ‚ค๋ฐฑ๊ณผ ๋˜๋Š” ๋” ๋„“์€ ์›น ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์—์„œ ์ž๋™์œผ๋กœ ํš๋“๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์œ„ํ‚ค๋ฐฑ๊ณผ ๋ฐ ์›น ๋„๋ฉ”์ธ์˜ ์ง„์ •ํ•œ ๋‹ต๋ณ€์ด ํฌํ•จ๋œ ์„ธ๊ทธ๋จผํŠธ์™€ ์œ„ํ‚ค๋ฐฑ๊ณผ ๋ฐ ์˜จ๋ผ์ธ์˜ ๊ด€๋ จ ๋ฌธ์„œ์™€ ํ•จ๊ป˜ ์ •ํ™•ํ•˜๊ฒŒ ๋‹ต๋ณ€ํ•œ ์งˆ๋ฌธ์ด ํฌํ•จ๋œ ํ™•์ธ๋œ ์„ธ๊ทธ๋จผํŠธ์˜ ๋‘ ๊ฐ€์ง€ ์„ธ๊ทธ๋จผํŠธ๋กœ ๋ถ„๋ฅ˜๋ฉ๋‹ˆ๋‹ค.
* **์ž์—ฐ์–ด ์งˆ๋ฌธ**[179]์€ Google ๊ฒ€์ƒ‰ ์—”์ง„์— ์ œ์ถœ๋œ ์‹ค์ œ ์ต๋ช…ํ™”๋œ ์ง‘๊ณ„ ์ฟผ๋ฆฌ๋ฅผ ์งˆ๋ฌธ์œผ๋กœ ๊ตฌ์„ฑํ•œ QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ฃผ์„์ž๋Š” ์ƒ์œ„ 5๊ฐœ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ ์ค‘ ํ•˜๋‚˜์˜ ์œ„ํ‚ค ํŽ˜์ด์ง€์™€ ํ•จ๊ป˜ ์งˆ๋ฌธ์ด ์ œ์‹œ๋˜๊ณ , ํŽ˜์ด์ง€์— ์žˆ๋Š” ๊ฒฝ์šฐ ๊ธด ๋‹ต๋ณ€(์ผ๋ฐ˜์ ์œผ๋กœ ๋‹จ๋ฝ)๊ณผ ์งง์€ ๋‹ต๋ณ€(ํ•˜๋‚˜ ์ด์ƒ์˜ ์—”ํ‹ฐํ‹ฐ)์„ ์ฃผ์„์œผ๋กœ ํ‘œ์‹œํ•˜๊ฑฐ๋‚˜, ๊ธด/์งง์€ ๋‹ต๋ณ€์ด ์—†๋Š” ๊ฒฝ์šฐ null์„ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.* MMLU[180]๋Š” ์ œ๋กœ์ƒท ๋ฐ ํŒŒ์šฐ์ƒท ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์–ป์€ ์ง€์‹์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ฆ‰, MMLU๋Š” ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ ์ง€์‹๊ณผ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ๋ชจ๋‘ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” STEM, ์ธ๋ฌธํ•™, ์‚ฌํšŒ๊ณผํ•™ ๋ฐ ๊ธฐํƒ€ ๋ถ„์•ผ์˜ 57๊ฐœ ๊ณผ๋ชฉ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” ์ดˆ๊ธ‰๋ถ€ํ„ฐ ๊ณ ๊ธ‰ ์ „๋ฌธ๊ฐ€๊นŒ์ง€ ๋ณต์žก์„ฑ์ด ๋‹ค์–‘ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ฃผ์š” ๊ณตํ—Œ์€ ๋‹ค์ค‘ ์ž‘์—… ์–ธ์–ด ์ดํ•ด, ์งˆ์˜ ์‘๋‹ต ๋ฐ ์‚ฐ์ˆ  ์ถ”๋ก ์„ ์œ„ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.* MBPP[181]๋Š” "๋Œ€๋ถ€๋ถ„ ๊ธฐ๋ณธ ํŒŒ์ด์ฌ ๋ฌธ์ œ"์˜ ์•ฝ์ž๋กœ ์ฝ”๋“œ ์ƒ์„ฑ์„ ์œ„ํ•ด ์„ค๊ณ„๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” ๊ธฐ๋ณธ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ฐœ๋… ๋ฐ ํ‘œ์ค€ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์‚ฌ์šฉ๋ฒ• ๋“ฑ ๋‹ค์–‘ํ•œ ์ฃผ์ œ๋ฅผ ํฌํ•จํ•˜๋Š” 974๊ฐœ์˜ ์งง์€ Python ํ”„๋กœ๊ทธ๋žจ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๊ณผ์ œ๋Š” ์ž‘์—… ์„ค๋ช…, ์ฝ”๋“œ ์†”๋ฃจ์…˜ ๋ฐ ์„ธ ๊ฐ€์ง€ ์ž๋™ํ™”๋œ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.* HumanEval[182]์€ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 164๊ฐœ์˜ ์ˆ˜์ž‘์—…์œผ๋กœ ์ œ์ž‘๋œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ณผ์ œ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๊ณผ์ œ์—๋Š” ํ•จ์ˆ˜ ์„œ๋ช…, ๋ฌธ์„œ ๋ฌธ์ž์—ด, ์ฝ”๋“œ ๋ณธ๋ฌธ, ์—ฌ๋Ÿฌ ๋‹จ์œ„ ํ…Œ์ŠคํŠธ๊ฐ€ ํ•จ๊ป˜ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ฐœ๋ฐœํ•œ ์ฃผ์š” ์ง๊ด€์€ ์ฝ”๋“œ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ•ด๋‹น ์ฝ˜ํ…์ธ ๋ฅผ ์ œ์™ธํ•  ์ˆ˜ ์žˆ๋„๋ก ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.* APPS[183]๋Š” Python ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์— ์ค‘์ ์„ ๋‘” ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. APPS ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” 232,444๊ฐœ์˜ Python ํ”„๋กœ๊ทธ๋žจ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ํ”„๋กœ๊ทธ๋žจ์€ ํ‰๊ท  18๊ฐœ์˜ Python ์ฝ”๋“œ ์ค„์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ APPS๋Š” ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ๋ฌธ์ œ ์„ค๋ช…์ด ์žˆ๋Š” 10,000๊ฐœ์˜ ๊ณ ์œ  ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์—ฐ์Šต ๋ฌธ์ œ ๋ฆฌํฌ์ง€ํ† ๋ฆฌ์— ๋Œ€ํ•œ ์•ก์„ธ์Šค๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ฐ•์กฐํ•ด์•ผ ํ•  ๋งˆ์ง€๋ง‰ ์ธก๋ฉด์€ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๋ฅผ ํฌํ•จํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.* WikiSQL[184]์€ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•ด ์ œ์ž‘๋˜์—ˆ์œผ๋ฉฐ, 87,726๊ฐœ์˜ SQL ์ฟผ๋ฆฌ์™€ ํ•ด๋‹น ์œ„ํ‚ค๋ฐฑ๊ณผ ํ…Œ์ด๋ธ”์˜ ์ž์—ฐ์–ด ์งˆ๋ฌธ์ด ์„ธ์‹ฌํ•˜๊ฒŒ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. SQL ์ฟผ๋ฆฌ๋Š” ํ…Œ์ŠคํŠธ ์„ธํŠธ(17,284๊ฐœ), ๊ฐœ๋ฐœ ์„ธํŠธ(9,145๊ฐœ), ํ•™์Šต ์„ธํŠธ(61,297๊ฐœ)์˜ ์„ธ ๊ฐ€์ง€ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.* TriviaQA[185]๋Š” QA ์ž‘์—…์„ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 650,000๊ฐœ ์ด์ƒ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€-์ฆ๊ฑฐ ์‚ผ์ค‘ ํ•ญ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” 95,000๊ฐœ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์ด ์žˆ์œผ๋ฉฐ, ๊ฐ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์€ ํ€ด์ฆˆ ์• ํ˜ธ๊ฐ€๊ฐ€ ์ž‘์„ฑํ–ˆ์œผ๋ฉฐ ํ‰๊ท  6๊ฐœ์˜ ๋…๋ฆฝ์ ์œผ๋กœ ์†Œ์‹ฑ๋œ ์ฆ๊ฑฐ ๋ฌธ์„œ๋กœ ์ง€์›๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์„œ๋Š” ์œ„ํ‚ค๋ฐฑ๊ณผ ๋˜๋Š” ๋” ๋„“์€ ์›น ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์—์„œ ์ž๋™์œผ๋กœ ์ˆ˜์ง‘๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์œ„ํ‚ค๋ฐฑ๊ณผ ๋ฐ ์›น ๋„๋ฉ”์ธ์˜ ์‹ค์ œ ๋‹ต๋ณ€์ด ์žˆ๋Š” ๊ฒƒ๊ณผ ์œ„ํ‚ค๋ฐฑ๊ณผ ๋ฐ ์˜จ๋ผ์ธ์˜ ์ •ํ™•ํ•œ ๋‹ต๋ณ€์ด ์žˆ๋Š” ๊ฒƒ์˜ ๋‘ ๊ฐ€์ง€ ์„ธ๊ทธ๋จผํŠธ๋กœ ๋ถ„๋ฅ˜๋ฉ๋‹ˆ๋‹ค.
์งˆ๋ฌธ* **์ž์—ฐ ์งˆ๋ฌธ**[179]์€ Google ๊ฒ€์ƒ‰ ์—”์ง„์— ์ œ์ถœ๋œ ์‹ค์ œ ์ต๋ช…ํ™”๋œ ์ง‘๊ณ„ ์ฟผ๋ฆฌ๋กœ ๊ตฌ์„ฑ๋œ QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ฃผ์„์„ ๋‹ค๋Š” ์‚ฌ๋žŒ์€ ์ƒ์œ„ \(5\) ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์—์„œ Wikipedia ํŽ˜์ด์ง€์™€ ํ•จ๊ป˜ ์งˆ๋ฌธ์„ ์ œ์‹œ๋ฐ›๊ณ , ํŽ˜์ด์ง€์— ๊ธด ๋‹ต๋ณ€(์ผ๋ฐ˜์ ์œผ๋กœ ๋‹จ๋ฝ)๊ณผ ์งง์€ ๋‹ต๋ณ€(ํ•˜๋‚˜ ์ด์ƒ์˜ ์—”ํ‹ฐํ‹ฐ)์ด ์žˆ๋Š” ๊ฒฝ์šฐ ์ฃผ์„์„ ๋‹ฌ๊ณ , ๊ธด/์งง์€ ๋‹ต๋ณ€์ด ์—†๋Š” ๊ฒฝ์šฐ null๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.* **MMLU**[180]๋Š” ์ œ๋กœ์ƒท ๋ฐ ๋ช‡ ์ƒท ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์–ป์€ ์ง€์‹์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ฆ‰, MMLU๋Š” ๋ชจ๋ธ์˜ ์ผ๋ฐ˜์ ์ธ ์ง€์‹๊ณผ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ๋ชจ๋‘ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. STEM, ์ธ๋ฌธํ•™, ์‚ฌํšŒ๊ณผํ•™ ๋ฐ ๊ธฐํƒ€ ๋ถ„์•ผ์˜ 57๊ฐœ ๊ณผ๋ชฉ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ๋ฒค์น˜๋งˆํฌ๋Š” ์ดˆ๊ธ‰๋ถ€ํ„ฐ ๊ณ ๊ธ‰ ์ „๋ฌธ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ๋ณต์žก์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ฃผ์š” ๊ธฐ์—ฌ๋Š” ๋‹ค์ค‘ ์ž‘์—… ์–ธ์–ด ์ดํ•ด, ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ ์‚ฐ์ˆ  ์ถ”๋ก ์— ์žˆ๋‹ค๋Š” ์ ์„ ์–ธ๊ธ‰ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.* **MBPP**[181]๋Š” "๋Œ€๋ถ€๋ถ„์˜ ๊ธฐ๋ณธ ํŒŒ์ด์ฌ ๋ฌธ์ œ"์˜ ์•ฝ์ž๋กœ ์ฝ”๋“œ ์ƒ์„ฑ์„ ์œ„ํ•ด ์„ค๊ณ„๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” ๊ธฐ๋ณธ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ฐœ๋…๊ณผ ํ‘œ์ค€ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์‚ฌ์šฉ๋ฒ• ๋“ฑ ๋‹ค์–‘ํ•œ ์ฃผ์ œ๋ฅผ ํฌํ•จํ•œ 974๊ฐœ์˜ ์งง์€ ํŒŒ์ด์ฌ ํ”„๋กœ๊ทธ๋žจ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๊ณผ์ œ๋Š” ๊ณผ์ œ ์„ค๋ช…, ์ฝ”๋“œ ์†”๋ฃจ์…˜, ์ž๋™ํ™”๋œ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค 3๊ฐœ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.* **HumanEval**[182]์€ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 164๊ฐœ์˜ ์ˆ˜์ž‘์—…์œผ๋กœ ์ œ์ž‘๋œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ณผ์ œ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๊ณผ์ œ์—๋Š” ํ•จ์ˆ˜ ์„œ๋ช…, ๋ฌธ์„œ ๋ฌธ์ž์—ด, ์ฝ”๋“œ ๋ณธ๋ฌธ, ์—ฌ๋Ÿฌ ๋‹จ์œ„ ํ…Œ์ŠคํŠธ๊ฐ€ ํ•จ๊ป˜ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ฐœ๋ฐœํ•œ ์ฃผ์š” ์ง๊ด€์€ ์ฝ”๋“œ ์ƒ์„ฑ ๋ชจ๋ธ์„ ์œ„ํ•œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ•ด๋‹น ๋‚ด์šฉ์„ ์ œ์™ธํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.* **APPS**[183]๋Š” ํŒŒ์ด์ฌ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์— ์ดˆ์ ์„ ๋งž์ถ˜ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. APPS ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” 232,444๊ฐœ์˜ ํŒŒ์ด์ฌ ํ”„๋กœ๊ทธ๋žจ ๋ชจ์Œ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ํ”„๋กœ๊ทธ๋žจ์€ ํ‰๊ท  18์ค„์˜ ํŒŒ์ด์ฌ ์ฝ”๋“œ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ APPS๋Š” ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ๋ฌธ์ œ ์„ค๋ช…์ด ํฌํ•จ๋œ 10,000๊ฐœ์˜ ๊ณ ์œ ํ•œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์—ฐ์Šต ๋ฌธ์ œ ๋ชจ์Œ์— ๋Œ€ํ•œ ์•ก์„ธ์Šค๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋‹ค๋Š” ์ ์ด ๋ˆˆ์— ๋•๋‹ˆ๋‹ค.* **์œ„ํ‚คSQL**[184]์€ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•ด ์ œ์ž‘๋˜์—ˆ์œผ๋ฉฐ, ์œ„ํ‚ค๋ฐฑ๊ณผ ํ‘œ์—์„œ ์ถ”์ถœํ•œ 87,726๊ฐœ์˜ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ SQL ์ฟผ๋ฆฌ์™€ ํ•ด๋‹น ์ž์—ฐ์–ด ์งˆ๋ฌธ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. SQL ์ฟผ๋ฆฌ๋Š” ํ…Œ์ŠคํŠธ ์„ธํŠธ(17,284๊ฐœ), ๊ฐœ๋ฐœ ์„ธํŠธ(9,145๊ฐœ), ํ›ˆ๋ จ ์„ธํŠธ(61,297๊ฐœ)์˜ ์„ธ ๊ฐ€์ง€ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.* **TriviaQA**[185]๋Š” QA ์ž‘์—…์„ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 650,000๊ฐœ ์ด์ƒ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€-์ฆ๊ฑฐ ์‚ผ์ค‘ํ•ญ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” 95,000๊ฐœ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์ด ์žˆ์œผ๋ฉฐ, ๊ฐ ์Œ์€ ํ€ด์ฆˆ ์• ํ˜ธ๊ฐ€๊ฐ€ ์ž‘์„ฑํ•˜๊ณ  ํ‰๊ท  6๊ฐœ์˜ ๋…๋ฆฝ์ ์œผ๋กœ ์ถœ์ฒ˜๊ฐ€ ํ™•์ธ๋œ ์ฆ๊ฑฐ ๋ฌธ์„œ๊ฐ€ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์„œ๋Š” Wikipedia ๋˜๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ์›น ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์—์„œ ์ž๋™์œผ๋กœ ์ˆ˜์ง‘๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” Wikipedia์™€ ์˜จ๋ผ์ธ์˜ ๋‘ ๊ฐ€์ง€ ์˜์—ญ์œผ๋กœ ๋ถ„๋ฅ˜๋˜๋ฉฐ, ๋‘ ์˜์—ญ ๋ชจ๋‘์—์„œ ์ •ํ™•ํ•˜๊ฒŒ ๋‹ต๋ณ€๋œ ์งˆ๋ฌธ๊ณผ ๊ด€๋ จ ๋ฌธ์„œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
**์ž์—ฐ ์งˆ๋ฌธ**[179]์€ Google ๊ฒ€์ƒ‰ ์—”์ง„์— ์งˆ๋ฌธ์œผ๋กœ ์ œ์ถœ๋œ ์‹ค์ œ ์ต๋ช…ํ™”๋œ ์ง‘๊ณ„ ์ฟผ๋ฆฌ๋กœ ๊ตฌ์„ฑ๋œ QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ฃผ์„๊ฐ€์—๊ฒŒ๋Š” ์ƒ์œ„ \(5\) ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์—์„œ ๊ฐ€์ ธ์˜จ Wikipedia ํŽ˜์ด์ง€์™€ ํ•จ๊ป˜ ์งˆ๋ฌธ์ด ์ œ์‹œ๋˜๋ฉฐ, ํŽ˜์ด์ง€์— ๊ธด ๋‹ต๋ณ€(์ผ๋ฐ˜์ ์œผ๋กœ ํ•œ ๋‹จ๋ฝ)๊ณผ ์งง์€ ๋‹ต๋ณ€(ํ•˜๋‚˜ ์ด์ƒ์˜ ์—”ํ‹ฐํ‹ฐ)์ด ์žˆ๋Š” ๊ฒฝ์šฐ ์ฃผ์„์„ ๋‹ฌ๊ฑฐ๋‚˜, ๊ธด/์งง์€ ๋‹ต๋ณ€์ด ์—†๋Š” ๊ฒฝ์šฐ null๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.* **MMLU**[180]๋Š” ์ œ๋กœ ์ƒท ๋ฐ ์†Œ์ˆ˜ ์ƒท ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์–ป์€ ์ง€์‹์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ฆ‰, MMLU๋Š” ๋ชจ๋ธ์˜ ์ผ๋ฐ˜์ ์ธ ์ง€์‹๊ณผ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์„ ๋ชจ๋‘ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. STEM, ์ธ๋ฌธํ•™, ์‚ฌํšŒ๊ณผํ•™ ๋ฐ ๊ธฐํƒ€ ๋ถ„์•ผ์˜ 57๊ฐœ ๊ณผ๋ชฉ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. ๋ฒค์น˜๋งˆํฌ๋Š” ์ดˆ๋“ฑ๋ถ€ํ„ฐ ๊ณ ๊ธ‰ ์ „๋ฌธ๊ฐ€๊นŒ์ง€ ๋‹ค์–‘ํ•œ ๋‚œ์ด๋„๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ฃผ์š” ๊ธฐ์—ฌ๋Š” ๋‹ค์ค‘ ์ž‘์—… ์–ธ์–ด ์ดํ•ด, ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ ์‚ฐ์ˆ  ์ถ”๋ก ์— ์žˆ๋‹ค๋Š” ์ ์„ ์–ธ๊ธ‰ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.* **MBPP**[181]๋Š” "๋Œ€๋ถ€๋ถ„ ๊ธฐ๋ณธ์ ์ธ ํŒŒ์ด์ฌ ๋ฌธ์ œ"์˜ ์•ฝ์ž๋กœ ์ฝ”๋“œ ์ƒ์„ฑ์„ ์œ„ํ•ด ์„ค๊ณ„๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” ๊ธฐ๋ณธ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ฐœ๋…๊ณผ ํ‘œ์ค€ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์‚ฌ์šฉ๋ฒ• ๋“ฑ ๋‹ค์–‘ํ•œ ์ฃผ์ œ๋ฅผ ํฌํ•จํ•œ 974๊ฐœ์˜ ์งง์€ Python ํ”„๋กœ๊ทธ๋žจ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๊ณผ์ œ๋Š” ๊ณผ์ œ ์„ค๋ช…, ์ฝ”๋“œ ์†”๋ฃจ์…˜, ์ž๋™ํ™”๋œ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค 3๊ฐœ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.* **HumanEval**[182]์€ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 164๊ฐœ์˜ ์ˆ˜์ž‘์—…์œผ๋กœ ์ œ์ž‘๋œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ณผ์ œ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๊ณผ์ œ์—๋Š” ํ•จ์ˆ˜ ์„œ๋ช…, ๋ฌธ์„œ ๋ฌธ์ž์—ด, ์ฝ”๋“œ ๋ณธ๋ฌธ ๋ฐ ์—ฌ๋Ÿฌ ๋‹จ์œ„ ํ…Œ์ŠคํŠธ๊ฐ€ ํ•จ๊ป˜ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ฐœ๋ฐœํ•œ ์ฃผ์š” ์ง๊ด€์€ ์ฝ”๋“œ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ•ด๋‹น ๋‚ด์šฉ์„ ์ œ์™ธํ•˜๋„๋ก ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.* **APPS**[183]๋Š” ํŒŒ์ด์ฌ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์–ธ์–ด์— ์ดˆ์ ์„ ๋งž์ถ˜ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. APPS ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” 232,444๊ฐœ์˜ Python ํ”„๋กœ๊ทธ๋žจ ๋ชจ์Œ์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ํ”„๋กœ๊ทธ๋žจ์—๋Š” ํ‰๊ท  18์ค„์˜ Python ์ฝ”๋“œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ APPS๋Š” 10,000๊ฐœ์˜ ๊ณ ์œ ํ•œ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์—ฐ์Šต ๋ฌธ์ œ ๋ชจ์Œ์— ๋Œ€ํ•œ ์•ก์„ธ์Šค๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ๊ฐ ๋ฌธ์ œ์—๋Š” ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ๋ฌธ์ œ ์„ค๋ช…์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ์ฃผ๋ชฉํ•  ์ ์€ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค.* **์œ„ํ‚คSQL**[184]์€ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•ด ์ œ์ž‘๋˜์—ˆ์œผ๋ฉฐ, ์œ„ํ‚ค๋ฐฑ๊ณผ ํ‘œ์—์„œ ์—„์„ ๋œ 87,726๊ฐœ์˜ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ SQL ์ฟผ๋ฆฌ์™€ ํ•ด๋‹น ์ž์—ฐ์–ด ์งˆ๋ฌธ ์Œ์„ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. SQL ์ฟผ๋ฆฌ๋Š” ์„ธ ๊ฐ€์ง€ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: ํ…Œ์ŠคํŠธ ์„ธํŠธ(17,284๊ฐœ), ๊ฐœ๋ฐœ ์„ธํŠธ(9,145๊ฐœ), ํ›ˆ๋ จ ์„ธํŠธ(61,297๊ฐœ)๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.* **TriviaQA**[185]๋Š” QA ์ž‘์—…์„ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” 650,000๊ฐœ ์ด์ƒ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€-์ฆ๊ฑฐ ์‚ผ์ค‘ํ•ญ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋Š” 95,000๊ฐœ์˜ ์งˆ๋ฌธ-๋‹ต๋ณ€ ์Œ์ด ์žˆ์œผ๋ฉฐ, ๊ฐ ์Œ์€ ํ€ด์ฆˆ ์• ํ˜ธ๊ฐ€๊ฐ€ ์ž‘์„ฑํ•˜๊ณ  ํ‰๊ท  6๊ฐœ์˜ ๋…๋ฆฝ์ ์œผ๋กœ ์ถœ์ฒ˜๊ฐ€ ํ™•์ธ๋œ ์ฆ๊ฑฐ ๋ฌธ์„œ๊ฐ€ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์„œ๋Š” Wikipedia ๋˜๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ์›น ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์—์„œ ์ž๋™์œผ๋กœ ์ˆ˜์ง‘๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” Wikipedia์™€ ์›น ๋„๋ฉ”์ธ์—์„œ ์ •ํ™•ํ•œ ๋‹ต๋ณ€์ด ์žˆ๋Š” ์„ธํŠธ์™€ Wikipedia์™€ ์˜จ๋ผ์ธ์˜ ๊ด€๋ จ ๋ฌธ์„œ์™€ ํ•จ๊ป˜ ์ •ํ™•ํ•˜๊ฒŒ ๋‹ต๋ณ€๋œ ์งˆ๋ฌธ์„ ๊ตฌํ˜„ํ•˜๋Š” ๊ฒ€์ฆ ์„ธํŠธ์˜ ๋‘ ๊ฐ€์ง€ ์„ธ๊ทธ๋จผํŠธ๋กœ ๋ถ„๋ฅ˜๋ฉ๋‹ˆ๋‹ค.
\(\bullet\)_ShareGPT_[148] is collected from a data collection platform where users can upload their conversations with ChatGPT or GPT-4 through the ShareGPT API. Currently, this dataset consists of approximately 90,000 conversations, including real instructions or inquiries from human and responses from ChatGPT.
\(\bullet\)_ShareGPT_[148]๋Š” ์‚ฌ์šฉ์ž๊ฐ€ ShareGPT API๋ฅผ ํ†ตํ•ด ChatGPT ๋˜๋Š” GPT-4์™€์˜ ๋Œ€ํ™”๋ฅผ ์—…๋กœ๋“œํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ํ”Œ๋žซํผ์—์„œ ์ˆ˜์ง‘๋ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ธ๊ฐ„์˜ ์‹ค์ œ ์ง€์นจ ๋˜๋Š” ๋ฌธ์˜์™€ ChatGPT์˜ ์‘๋‹ต์„ ํฌํ•จํ•˜์—ฌ ์•ฝ 90,000๊ฐœ์˜ ๋Œ€ํ™”๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_ShareGPT_[148]๋Š” ์‚ฌ์šฉ์ž๊ฐ€ ShareGPT API๋ฅผ ํ†ตํ•ด ChatGPT ๋˜๋Š” GPT-4์™€์˜ ๋Œ€ํ™”๋ฅผ ์—…๋กœ๋“œํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ํ”Œ๋žซํผ์—์„œ ์ˆ˜์ง‘๋ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ธ๊ฐ„์˜ ์‹ค์ œ ์ง€์นจ ๋˜๋Š” ๋ฌธ์˜ ์‚ฌํ•ญ๊ณผ ChatGPT์˜ ์‘๋‹ต์„ ํฌํ•จํ•˜์—ฌ ์•ฝ 90,000๊ฐœ์˜ ๋Œ€ํ™”๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_ShareGPT_[148]๋Š” ์‚ฌ์šฉ์ž๊ฐ€ ShareGPT API๋ฅผ ํ†ตํ•ด ChatGPT ๋˜๋Š” GPT-4์™€์˜ ๋Œ€ํ™”๋ฅผ ์—…๋กœ๋“œํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ํ”Œ๋žซํผ์—์„œ ์ˆ˜์ง‘๋ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์‚ฌ๋žŒ์˜ ์‹ค์ œ ์ง€์นจ์ด๋‚˜ ๋ฌธ์˜์™€ ChatGPT์˜ ์‘๋‹ต์„ ํฌํ•จํ•˜์—ฌ ์•ฝ 90,000๊ฐœ์˜ ๋Œ€ํ™”๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
\(\์ด์•Œ\)_ShareGPT_[148]์€ ์‚ฌ์šฉ์ž๊ฐ€ ShareGPT API๋ฅผ ํ†ตํ•ด ChatGPT ๋˜๋Š” GPT-4์™€์˜ ๋Œ€ํ™”๋ฅผ ์—…๋กœ๋“œํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ํ”Œ๋žซํผ์—์„œ ์ˆ˜์ง‘๋ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์‹ค์ œ ์ง€์นจ์ด๋‚˜ ์ธ๊ฐ„์˜ ๋ฌธ์˜, ChatGPT์˜ ์‘๋‹ต์„ ํฌํ•จํ•˜์—ฌ ์•ฝ 90,000๊ฐœ์˜ ๋Œ€ํ™”๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_ShareGPT_[148]๋Š” ์‚ฌ์šฉ์ž๊ฐ€ ShareGPT API๋ฅผ ํ†ตํ•ด ChatGPT ๋˜๋Š” GPT-4์™€์˜ ๋Œ€ํ™”๋ฅผ ์—…๋กœ๋“œํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ํ”Œ๋žซํผ์—์„œ ์ˆ˜์ง‘๋ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์‚ฌ๋žŒ์˜ ์‹ค์ œ ์ง€์‹œ๋‚˜ ๋ฌธ์˜, ChatGPT์˜ ์‘๋‹ต์„ ํฌํ•จํ•˜์—ฌ ์•ฝ 9๋งŒ ๊ฑด์˜ ๋Œ€ํ™”๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
In this section, we present RAFT, a novel way of training LLMs for domain specific open-book exams. We first introduce the classical technique of supervised fine-tuning, fol
์ด ์„น์…˜์—์„œ๋Š” ๋„๋ฉ”์ธ๋ณ„ ์˜คํ”ˆ๋ถ ์‹œํ—˜์„ ์œ„ํ•ด LLM์„ ํ›ˆ๋ จํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์ธ RAFT๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์ง€๋„ ํ•™์Šต์„ ํ†ตํ•œ ๋ฏธ์„ธ ์กฐ์ •์ด๋ผ๋Š” ๊ณ ์ „์ ์ธ ๊ธฐ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋„๋ฉ”์ธ๋ณ„ ์˜คํ”ˆ๋ถ ์‹œํ—˜์„ ์œ„ํ•ด LLM์„ ํ›ˆ๋ จํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์ธ RAFT๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •์˜ ๊ณ ์ „์ ์ธ ๊ธฐ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋„๋ฉ”์ธ๋ณ„ ๊ฐœ๋ฐฉํ˜• ์‹œํ—˜์„ ์œ„ํ•ด LLM์„ ํ›ˆ๋ จํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์ธ RAFT๋ฅผ ์†Œ๊ฐœํ•œ๋‹ค. ๋จผ์ €, ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •์˜ ๊ณ ์ „์ ์ธ ๊ธฐ๋ฒ•์ธ ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •์„ ์†Œ๊ฐœํ•œ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋„๋ฉ”์ธ๋ณ„ ์˜คํ”ˆ๋ถ ์‹œํ—˜์„ ์œ„ํ•ด LLM์„ ํ›ˆ๋ จํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์ธ RAFT๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์ง€๋„์‹ ๋ฏธ์„ธ ์กฐ์ •์ด๋ผ๋Š” ๊ธฐ์กด ๊ธฐ์ˆ ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ๋„๋ฉ”์ธ๋ณ„ ์˜คํ”ˆ๋ถ ์‹œํ—˜์„ ์œ„ํ•ด LLM์„ ํ›ˆ๋ จํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์ธ RAFT๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •์ด๋ผ๋Š” ๊ณ ์ „์ ์ธ ๊ธฐ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.
However, when running the model in higher float64 precision, we find that indeed all dimensions are used, but that the smallest dozen or so singular values are much smaller than the other singular values, an observation made by concurrent work (Cancedda, 2024).
๊ทธ๋Ÿฌ๋‚˜ ๋” ๋†’์€ ๋ถ€๋™ ์†Œ์ˆ˜์  ์ •๋ฐ€๋„๋กœ ๋ชจ๋ธ์„ ์‹คํ–‰ํ•˜๋ฉด ์‹ค์ œ๋กœ ๋ชจ๋“  ์ฐจ์›์ด ์‚ฌ์šฉ๋˜์ง€๋งŒ, ๊ฐ€์žฅ ์ž‘์€ 12๊ฐœ ์ •๋„์˜ ํŠน์ด๊ฐ’์ด ๋‹ค๋ฅธ ํŠน์ด๊ฐ’๋ณด๋‹ค ํ›จ์”ฌ ์ž‘๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋™์‹œ ์ž‘์—…(Cancedda, 2024)์—์„œ ๊ด€์ฐฐํ•œ ๋ฐ”์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ ๋” ๋†’์€ float64 ์ •๋ฐ€๋„๋กœ ๋ชจ๋ธ์„ ์‹คํ–‰ํ•˜๋ฉด ์‹ค์ œ๋กœ ๋ชจ๋“  ์ฐจ์›์ด ์‚ฌ์šฉ๋˜์ง€๋งŒ, ๊ฐ€์žฅ ์ž‘์€ 12๊ฐœ ์ •๋„์˜ ํŠน์ด๊ฐ’์ด ๋‹ค๋ฅธ ํŠน์ด๊ฐ’๋ณด๋‹ค ํ›จ์”ฌ ์ž‘๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋™์‹œ ์ž‘์—…(Cancedda, 2024)์—์„œ ๊ด€์ฐฐํ•œ ๋ฐ”์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ ๋” ๋†’์€ ๋ถ€๋™ ์†Œ์ˆ˜์  64 ์ •๋ฐ€๋„๋กœ ๋ชจ๋ธ์„ ์‹คํ–‰ํ•˜๋ฉด ์‹ค์ œ๋กœ ๋ชจ๋“  ์ฐจ์›์ด ์‚ฌ์šฉ๋˜์ง€๋งŒ, ๊ฐ€์žฅ ์ž‘์€ 12๊ฐœ ์ •๋„์˜ ํŠน์ด๊ฐ’์ด ๋‹ค๋ฅธ ํŠน์ด๊ฐ’๋ณด๋‹ค ํ›จ์”ฌ ์ž‘๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Š” ๋™์‹œ ์ž‘์—…(Cancedda, 2024)์—์„œ ์–ธ๊ธ‰ํ•œ ๋ฐ”์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ ๋” ๋†’์€ float64 ์ •๋ฐ€๋„๋กœ ๋ชจ๋ธ์„ ์‹คํ–‰ํ•˜๋ฉด ์‹ค์ œ๋กœ ๋ชจ๋“  ์ฐจ์›์ด ์‚ฌ์šฉ๋˜์ง€๋งŒ ๊ฐ€์žฅ ์ž‘์€ 12๊ฐœ ์ด์ƒ์˜ ํŠน์ด๊ฐ’์ด ๋‹ค๋ฅธ ํŠน์ด๊ฐ’๋ณด๋‹ค ํ›จ์”ฌ ์ž‘๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Š” ๋™์‹œ ์ž‘์—…(Cancedda, 2024)์— ์˜ํ•ด ๊ด€์ฐฐ๋œ ๋ฐ” ์žˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ ๋” ๋†’์€ float64 ์ •๋ฐ€๋„๋กœ ๋ชจ๋ธ์„ ์‹คํ–‰ํ•˜๋ฉด ์‹ค์ œ๋กœ ๋ชจ๋“  ์ฐจ์›์ด ์‚ฌ์šฉ๋˜์ง€๋งŒ ๊ฐ€์žฅ ์ž‘์€ 12๊ฐœ ์ •๋„์˜ ํŠน์ด๊ฐ’์ด ๋‹ค๋ฅธ ํŠน์ด๊ฐ’๋ณด๋‹ค ํ›จ์”ฌ ์ž‘๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Š” ๋™์‹œ ์ž‘์—…(Cancedda, 2024)์—์„œ ๊ด€์ฐฐ๋œ ๋ฐ” ์žˆ์Šต๋‹ˆ๋‹ค.
In this method, the FT process is viewed as a direct continuation of the pre-training phase. We start with a saved checkpoint of the original LLM and train it in a causal auto-regressive manner, i.e., predicting the next token. One major difference in comparison to actual pre-training is the learning rate. Usually, one would need a much lower learning rate when continuing the pre-training of the model to avoid catastrophic forgetting (Kirkpatrick et al., 2017).
์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” FT ํ”„๋กœ์„ธ์Šค๋ฅผ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์˜ ์ง์ ‘์ ์ธ ์—ฐ์†์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ LLM์˜ ์ €์žฅ๋œ ์ฒดํฌํฌ์ธํŠธ๋กœ ์‹œ์ž‘ํ•˜์—ฌ ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ธ๊ณผ์  ์ž๋™ ํšŒ๊ท€ ๋ฐฉ์‹์œผ๋กœ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ์‚ฌ์ „ ํ•™์Šต๊ณผ ๋น„๊ตํ•  ๋•Œ ๊ฐ€์žฅ ํฐ ์ฐจ์ด์ ์€ ํ•™์Šต ์†๋„์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต์„ ๊ณ„์†ํ•  ๋•Œ๋Š” ์žฌ์•™์ ์ธ ๋ง๊ฐ(Kirkpatrick ์™ธ., 2017)์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ํ›จ์”ฌ ๋‚ฎ์€ ํ•™์Šต ์†๋„๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” FT ํ”„๋กœ์„ธ์Šค๋ฅผ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์˜ ์ง์ ‘์ ์ธ ์—ฐ์†์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ LLM์˜ ์ €์žฅ๋œ ์ฒดํฌํฌ์ธํŠธ๋กœ ์‹œ์ž‘ํ•˜์—ฌ ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ธ๊ณผ์  ์ž๋™ ํšŒ๊ท€ ๋ฐฉ์‹์œผ๋กœ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ์‚ฌ์ „ ํ•™์Šต๊ณผ ๋น„๊ตํ•  ๋•Œ ๊ฐ€์žฅ ํฐ ์ฐจ์ด์ ์€ ํ•™์Šต ์†๋„์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต์„ ๊ณ„์†ํ•  ๋•Œ๋Š” ์žฌ์•™์  ๋ง๊ฐ(Kirkpatrick ๋“ฑ, 2017)์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ํ›จ์”ฌ ๋‚ฎ์€ ํ•™์Šต ์†๋„๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” FT ํ”„๋กœ์„ธ์Šค๋ฅผ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์˜ ์ง์ ‘์ ์ธ ์—ฐ์†์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ LLM์˜ ์ €์žฅ๋œ ์ฒดํฌํฌ์ธํŠธ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜์—ฌ ์ธ๊ณผ์  ์ž๋™ ํšŒ๊ท€ ๋ฐฉ์‹์œผ๋กœ ํ•™์Šตํ•˜์—ฌ ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ์‚ฌ์ „ ํ•™์Šต๊ณผ ๋น„๊ตํ•  ๋•Œ ํ•œ ๊ฐ€์ง€ ์ฃผ์š” ์ฐจ์ด์ ์€ ํ•™์Šต ์†๋„์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต์„ ๊ณ„์†ํ•  ๋•Œ ์น˜๋ช…์ ์ธ ๋ง๊ฐ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ํ›จ์”ฌ ๋‚ฎ์€ ํ•™์Šต ์†๋„๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค(Kirkpatrick et al., 2017).
์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” FT ํ”„๋กœ์„ธ์Šค๋ฅผ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์˜ ์ง์ ‘์ ์ธ ์—ฐ์†์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ LLM์˜ ์ €์žฅ๋œ ์ฒดํฌํฌ์ธํŠธ์—์„œ ์‹œ์ž‘ํ•˜์—ฌ ์ธ๊ณผ์  ์ž๋™ ํšŒ๊ท€ ๋ฐฉ์‹์œผ๋กœ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค(์ฆ‰, ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก). ์‹ค์ œ ์‚ฌ์ „ ํ›ˆ๋ จ๊ณผ ๋น„๊ตํ•  ๋•Œ ํ•œ ๊ฐ€์ง€ ์ฃผ์š” ์ฐจ์ด์ ์€ ํ•™์Šต ์†๋„์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ๊ณ„์†ํ•  ๋•Œ๋Š” ์น˜๋ช…์ ์ธ ๋ง๊ฐ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ํ›จ์”ฌ ๋” ๋‚ฎ์€ ํ•™์Šต ์†๋„๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค(Kirkpatrick et al., 2017).
์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” FT ํ”„๋กœ์„ธ์Šค๋ฅผ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์˜ ์ง์ ‘์ ์ธ ์—ฐ์†์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ LLM์˜ ์ €์žฅ๋œ ์ฒดํฌํฌ์ธํŠธ์—์„œ ์‹œ์ž‘ํ•˜์—ฌ ์ธ๊ณผ์  ์ž๋™ ํšŒ๊ท€ ๋ฐฉ์‹, ์ฆ‰ ๋‹ค์Œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ์‚ฌ์ „ ํ•™์Šต๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ํ•œ ๊ฐ€์ง€ ์ฃผ์š” ์ฐจ์ด์ ์€ ํ•™์Šต ์†๋„์ž…๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์น˜๋ช…์ ์ธ ๋ง๊ฐ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต์„ ๊ณ„์†ํ•  ๋•Œ๋Š” ํ›จ์”ฌ ๋” ๋‚ฎ์€ ํ•™์Šต ์†๋„๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค(Kirkpatrick et al., 2017).
* 1 Introduction * 2 Pretraining * 2.1 Data Processing * 2.2 Tokenization * 2.3 Model Architecture * 3 Finetuning * 3.1 Data Preprocessing * 3.2 Training Method * 4 Infrastructure * 5 Safety * 6 Evaluations * 6.1 Base Model Performance * 6.1.1 Main Results * 6.1.2 Discussions * 6.1.3 In-Context Learning Study * 6.2 Chat Model Performance * 6.2.1 Automatic Evaluations * 6.2.2 Human Evaluations * 7 Capability Extension * 7.1 Long Context Modeling * 7.2 Vision-Language * 7.3 Depth Upscaling * 8 Final Discussions * A Author List and Contributions
* 1 ์†Œ๊ฐœ* 2 ์‚ฌ์ „ ํ›ˆ๋ จ * 2.1 ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ * 2.2 ํ† ํฐํ™” * 2.3 ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜* 3 ํŒŒ์ธํŠœ๋‹ * 3.1 ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ * 3.2 ํ›ˆ๋ จ ๋ฐฉ๋ฒ•* 4 ์ธํ”„๋ผ* 5 ์•ˆ์ „์„ฑ* 6 ํ‰๊ฐ€ * 6.1 ๊ธฐ๋ณธ ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.1.1 ์ฃผ์š” ๊ฒฐ๊ณผ * 6.1.2 ํ† ๋ก  * 6.1.3 ์ปจํ…์ŠคํŠธ ๋‚ด ํ•™์Šต ์—ฐ๊ตฌ * 6.2 ์ฑ— ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.2.1 ์ž๋™ ํ‰๊ฐ€ * 6.2.2 ์ธ๊ฐ„ ํ‰๊ฐ€* 7 ๊ธฐ๋Šฅ ํ™•์žฅ * 7.1 ๊ธด ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ๋ง * 7.2 ๋น„์ „-์–ธ์–ด * 7.3 ๊นŠ์ด ์—…์Šค์ผ€์ผ๋ง* 8 ์ตœ์ข… ํ† ๋ก * A ์ €์ž ๋ชฉ๋ก ๋ฐ ๊ธฐ์—ฌ๋„
* 1 ์†Œ๊ฐœ* 2 ์‚ฌ์ „ ํ›ˆ๋ จ * 2.1 ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ * 2.2 ํ† ํฐํ™” * 2.3 ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜* 3 ํŒŒ์ธํŠœ๋‹ * 3.1 ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ * 3.2 ํ›ˆ๋ จ ๋ฐฉ๋ฒ•* 4 ์ธํ”„๋ผ* 5 ์•ˆ์ „์„ฑ* 6 ํ‰๊ฐ€ * 6.1 ๋ฒ ์ด์Šค ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.1.1 ์ฃผ์š” ๊ฒฐ๊ณผ * 6.1.2 ํ† ๋ก  * 6.1.3 ์ธ ์ปจํ…์ŠคํŠธ ํ•™์Šต ์—ฐ๊ตฌ * 6.2 ์ฑ— ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.2.1 ์ž๋™ ํ‰๊ฐ€ * 6.2.2 ์ธ๊ฐ„ ํ‰๊ฐ€* 7 ๊ธฐ๋Šฅ ํ™•์žฅ * 7.1 ๊ธด ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ๋ง * 7.2 ๋น„์ „-์–ธ์–ด * 7.3 ๊นŠ์ด ํ™•์žฅ* 8 ์ตœ์ข… ํ† ๋ก * A ์ž‘์„ฑ์ž ๋ชฉ๋ก ๋ฐ ๊ธฐ์—ฌ๋„
* 1 ์†Œ๊ฐœ* 2 ์‚ฌ์ „ ํ›ˆ๋ จ * 2.1 ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ * 2.2 ํ† ํฐํ™” * 2.3 ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜* 3 ํŒŒ์ธํŠœ๋‹ * 3.1 ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ * 3.2 ํ›ˆ๋ จ ๋ฐฉ๋ฒ•* 4 ์ธํ”„๋ผ* 5 ์•ˆ์ „์„ฑ* 6 ํ‰๊ฐ€ * 6.1 ๋ฒ ์ด์Šค ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.1.1 ์ฃผ์š” ๊ฒฐ๊ณผ * 6.1.2 ํ† ๋ก  * 6.1.3 ์ธ-์ปจํ…์ŠคํŠธ ํ•™์Šต ์—ฐ๊ตฌ * 6.2 ์ฑ„ํŒ… ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.2.1 ์ž๋™ ํ‰๊ฐ€ * 6.2.2 ์‚ฌ๋žŒ ํ‰๊ฐ€* 7 ๊ธฐ๋Šฅ ํ™•์žฅ * 7.1 ๊ธด ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ๋ง * 7.2 ๋น„์ „-์–ธ์–ด * 7.3 ๊นŠ์ด ์—…์Šค์ผ€์ผ๋ง* 8 ์ตœ์ข… ํ† ๋ก * A ์ €์ž ๋ชฉ๋ก ๋ฐ ๊ธฐ์—ฌ๋„
1 ์†Œ๊ฐœ* 2 ์‚ฌ์ „ ํ•™์Šต * 2.1 ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ * 2.2 ํ† ํฐํ™” * 2.3 ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜* 3 ๋ฏธ์„ธ ์กฐ์ • * 3.1 ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ * 3.2 ํ›ˆ๋ จ ๋ฐฉ๋ฒ•* 4 ์ธํ”„๋ผ* 5 ์•ˆ์ „* 6 ํ‰๊ฐ€ * 6.1 ๊ธฐ๋ณธ ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.1.1 ์ฃผ์š” ๊ฒฐ๊ณผ * 6.1.2 ํ† ๋ก  * 6.1.3 ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต ์—ฐ๊ตฌ * 6.2 ์ฑ„ํŒ… ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.2.1 ์ž๋™ ํ‰๊ฐ€ * 6.2.2 ์‚ฌ๋žŒ ํ‰๊ฐ€* 7 ๊ธฐ๋Šฅ ํ™•์žฅ * 7.1 ๊ธด ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ๋ง * 7.2 ๋น„์ „-์–ธ์–ด * 7.3 ์‹ฌ๋„ ์—…์Šค์ผ€์ผ๋ง* 8 ์ตœ์ข… ํ† ๋ก * A ์ €์ž ๋ชฉ๋ก ๋ฐ ๊ธฐ์—ฌ
1 ์†Œ๊ฐœ* 2 ์‚ฌ์ „ ํ›ˆ๋ จ * 2.1 ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ * 2.2 ํ† ํฐํ™” * 2.3 ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜* 3 ๋ฏธ์„ธ ์กฐ์ • * 3.1 ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ * 3.2 ํ›ˆ๋ จ ๋ฐฉ๋ฒ•* 4 ์ธํ”„๋ผ* 5 ์•ˆ์ „* 6 ํ‰๊ฐ€ * 6.1 ๊ธฐ๋ณธ ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.1.1 ์ฃผ์š” ๊ฒฐ๊ณผ * 6.1.2 ํ† ๋ก  * 6.1.3 ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต ์—ฐ๊ตฌ * 6.2 ์ฑ„ํŒ… ๋ชจ๋ธ ์„ฑ๋Šฅ * 6.2.1 ์ž๋™ ํ‰๊ฐ€ * 6.2.2 ์ธ๊ฐ„ ํ‰๊ฐ€* 7 ๊ธฐ๋Šฅ ํ™•์žฅ * 7.1 ๊ธด ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ๋ง * 7.2 ๋น„์ „-์–ธ์–ด * 7.3 ๊นŠ์ด ์—…์Šค์ผ€์ผ๋ง* 8 ์ตœ์ข… ํ† ๋ก * A ์ €์ž ๋ชฉ๋ก ๋ฐ ๊ธฐ์—ฌ๋„
1. Pre-training on larger datasets helps improve performance, by presumably exposing the LLM to more "knowledge". For instance, the Pythia (300B) model achieves an average score of 44.6%, outperforming the smaller C4 (85B) dataset's score of 43.5%.
1. ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต์€ LLM์ด ๋” ๋งŽ์€ '์ง€์‹'์— ๋…ธ์ถœ๋˜์–ด ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Pythia(300B) ๋ชจ๋ธ์€ ํ‰๊ท  44.6%์˜ ์ ์ˆ˜๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ ๋” ์ž‘์€ C4(85B) ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ ์ˆ˜์ธ 43.5%๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
1. ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต์€ LLM์ด ๋” ๋งŽ์€ '์ง€์‹'์— ๋…ธ์ถœ๋˜๋„๋ก ํ•จ์œผ๋กœ์จ ์„ฑ๋Šฅ ํ–ฅ์ƒ์— ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Pythia(300B) ๋ชจ๋ธ์€ ํ‰๊ท  44.6์ ์„ ๋‹ฌ์„ฑํ•˜์—ฌ ๋” ์ž‘์€ C4(85B) ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ ์ˆ˜์ธ 43.5์ ์„ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
1. ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต์€ LLM์ด ๋” ๋งŽ์€ '์ง€์‹'์— ๋…ธ์ถœ๋จ์œผ๋กœ์จ ์„ฑ๋Šฅ ํ–ฅ์ƒ์— ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Pythia(300B) ๋ชจ๋ธ์€ ํ‰๊ท  44.6%์˜ ์ ์ˆ˜๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ ๋” ์ž‘์€ C4(85B) ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ ์ˆ˜์ธ 43.5%๋ฅผ ๋Šฅ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.
1. ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต์€ LLM์ด ๋” ๋งŽ์€ "์ง€์‹"์— ๋…ธ์ถœ๋จ์œผ๋กœ์จ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ”ผํƒ€๊ณ ๋ผ์Šค(300B) ๋ชจ๋ธ์€ ํ‰๊ท  44.6%์˜ ์ ์ˆ˜๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ ๋” ์ž‘์€ C4(85B) ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ ์ˆ˜์ธ 43.5%๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
1. ๋” ํฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต์€ LLM์ด ๋” ๋งŽ์€ "์ง€์‹"์— ๋…ธ์ถœ๋จ์œผ๋กœ์จ ์„ฑ๋Šฅ ํ–ฅ์ƒ์— ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ”ผํƒ€(300B) ๋ชจ๋ธ์€ ํ‰๊ท  44.6%์˜ ์ ์ˆ˜๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ ๋” ์ž‘์€ C4(85B) ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ ์ˆ˜์ธ 43.5%๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
Long-context models overemphasize book endings:One interesting observation is that Claude-3-Opus and GPT-4-Turbo, which both have chunk sizes \(\geq\) 100K, tend to place more emphasis on the endings of the books to the detriment of the beginning. Since these models were often provided with the entire book context during prompting, this
๊ธด ๋ฌธ๋งฅ ๋ชจ๋ธ์€ ์ฑ… ๊ฒฐ๋ง์„ ์ง€๋‚˜์น˜๊ฒŒ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค:ํ•œ ๊ฐ€์ง€ ํฅ๋ฏธ๋กœ์šด ๊ด€์ฐฐ์€ ์ฒญํฌ ํฌ๊ธฐ๊ฐ€ 10๋งŒ ๋‹จ์–ด ์ด์ƒ์ธ ํด๋กœ๋“œ-3-์˜คํผ์Šค์™€ GPT-4-ํ„ฐ๋ณด๊ฐ€ ์ฑ…์˜ ์‹œ์ž‘ ๋ถ€๋ถ„์„ ์†์ƒ์‹œํ‚ค๋Š” ๋Œ€์‹  ๋๋ถ€๋ถ„์— ๋” ๋งŽ์€ ๊ฐ•์กฐ๋ฅผ ๋‘๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ๋“ค์€ ํ”„๋กฌํ”„ํŠธ ์ค‘์— ์ „์ฒด ์ฑ… ๋ฌธ๋งฅ์„ ์ œ๊ณต๋ฐ›์•˜๊ธฐ ๋•Œ๋ฌธ์— ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค.
๊ธด ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ์€ ์ฑ…์˜ ๊ฒฐ๋ง์„ ์ง€๋‚˜์น˜๊ฒŒ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค:ํ•œ ๊ฐ€์ง€ ํฅ๋ฏธ๋กœ์šด ๊ด€์ฐฐ ๊ฒฐ๊ณผ๋Š” ์ฒญํฌ ํฌ๊ธฐ๊ฐ€ 100K ์ด์ƒ์ธ ํด๋กœ๋“œ-3-์˜คํผ์Šค์™€ GPT-4-ํ„ฐ๋ณด๋Š” ์ฑ…์˜ ์‹œ์ž‘ ๋ถ€๋ถ„์„ ํ•ด์น˜๊ณ  ์ฑ…์˜ ๊ฒฐ๋ง์— ๋” ๋งŽ์€ ๊ฐ•์กฐ๋ฅผ ๋‘๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ํ”„๋กฌํ”„ํŠธ ๋™์•ˆ ์ „์ฒด ์ฑ… ์ปจํ…์ŠคํŠธ๋ฅผ ์ œ๊ณต๋ฐ›์•˜๊ธฐ ๋•Œ๋ฌธ์—
๊ธด ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ์€ ์ฑ… ๋์„ ์ง€๋‚˜์น˜๊ฒŒ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค:ํ•œ ๊ฐ€์ง€ ํฅ๋ฏธ๋กœ์šด ๊ด€์ฐฐ์€ ์ฒญํฌ ํฌ๊ธฐ๊ฐ€ 100K ์ด์ƒ์ธ ํด๋กœ๋“œ-3-์˜คํ‘ธ์Šค์™€ GPT-4-ํ„ฐ๋ณด๊ฐ€ ์ฑ…์˜ ์‹œ์ž‘ ๋ถ€๋ถ„๋ณด๋‹ค ๋ ๋ถ€๋ถ„์— ๋” ์ค‘์ ์„ ๋‘๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ํ”„๋กฌํ”„ํŠธ ์ค‘์— ์ „์ฒด ์ฑ… ์ปจํ…์ŠคํŠธ๋ฅผ ์ œ๊ณต๋ฐ›์€ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•˜๊ธฐ ๋•Œ๋ฌธ์—
์ฑ… ๊ฒฐ๋ง์— ์ง€๋‚˜์น˜๊ฒŒ ์ค‘์ ์„ ๋‘๋Š” ๊ธด ๋ฌธ๋งฅ ๋ชจ๋ธ:ํ•œ ๊ฐ€์ง€ ํฅ๋ฏธ๋กœ์šด ๊ด€์ฐฐ์€ ๋ชจ๋‘ 100K ์ด์ƒ์˜ ์ฒญํฌ ํฌ๊ธฐ๋ฅผ ๊ฐ€์ง„ Claude-3-Opus์™€ GPT-4-Turbo๊ฐ€ ์ฑ…์˜ ์‹œ์ž‘ ๋ถ€๋ถ„๋ณด๋‹ค ๊ฒฐ๋ง์— ๋” ๋งŽ์€ ๋น„์ค‘์„ ๋‘๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ํ”„๋กฌํ”„ํŠธ ์ค‘์— ์ „์ฒด ์ฑ… ์ปจํ…์ŠคํŠธ๊ฐ€ ์ œ๊ณต๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๊ธฐ ๋•Œ๋ฌธ์— ์ด๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๊ฒฐ๊ณผ๋ฅผ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ฑ… ๊ฒฐ๋ง์„ ์ง€๋‚˜์น˜๊ฒŒ ๊ฐ•์กฐํ•˜๋Š” ๋กฑ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ:ํ•œ ๊ฐ€์ง€ ํฅ๋ฏธ๋กœ์šด ๊ด€์ฐฐ์€ ์ฒญํฌ ํฌ๊ธฐ๊ฐ€ ๋ชจ๋‘ 100K ์ด์ƒ์ธ Claude-3-Opus์™€ GPT-4-Turbo๊ฐ€ ์ฑ…์˜ ๊ฒฐ๋ง์— ๋” ๋งŽ์€ ๋น„์ค‘์„ ๋‘๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์‹œ์ž‘ ๋ถ€๋ถ„์— ํ•ด๋ฅผ ๋ผ์น˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ํ”„๋กฌํ”„ํŠธ ์ค‘์— ์ „์ฒด ์ฑ… ์ปจํ…์ŠคํŠธ๊ฐ€ ์ œ๊ณต๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๊ธฐ ๋•Œ๋ฌธ์— ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค.
We begin with a simpler case where \(g_{\theta}\) does not have residual connections but a fully connected (FC) final layer. In this case, for any invertible \(h\times h\) matrix \(\mathbf{S}\), we have that \(g_{\theta}\left(p\right)=\mathbf{S}_{g\theta^{\prime}}\left(p\right)\) where \(\theta^{\prime}\) is the same as \(\theta\) except that the weights of the final FC layer are pre-multiplied by \(\mathbf{S}^{-1}\). Hence, if \(g_{\theta}\) has a final FC layer, it is impossible to distinguish between the embedding projection layer \(\mathbf{W}\) acting on \(g_{\theta}\) and the embedding projection layer \(\mathbf{W}\cdot\mathbf{S}\) acting on \(g_{\theta^{\prime}}\), given access to the output of the API \(\mathcal{O}\) only.
๋จผ์ €, ์ž”์—ฌ ์—ฐ๊ฒฐ์ด ์—†๊ณ  ์™„์ „ํžˆ ์—ฐ๊ฒฐ๋œ(FC) ์ตœ์ข… ๋ ˆ์ด์–ด๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ \(g_{\theta}\)์˜ ๋” ๊ฐ„๋‹จํ•œ ๊ฒฝ์šฐ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ, ๋ชจ๋“  ๋ฐ˜์ „ ๊ฐ€๋Šฅํ•œ \(h\times h\) ํ–‰๋ ฌ \(\mathbf{S}\)์— ๋Œ€ํ•ด \(g_{\theta}\left(p\right)=\mathbf{S}_{g\theta^{\prime}}\left(p\right)\)๊ฐ€ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ \(\theta^{\prime}\)๋Š” ์ตœ์ข… FC ๋ ˆ์ด์–ด์˜ ๊ฐ€์ค‘์น˜๊ฐ€ \(\mathbf{S}^{-1}\)๋กœ ์‚ฌ์ „ ๊ณฑํ•ด์ง„ ๊ฒƒ์„ ์ œ์™ธํ•˜๋ฉด \(\theta\)์™€ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ \(g_{\theta}\)์— ์ตœ์ข… FC ๋ ˆ์ด์–ด๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ, API \(\mathcal{O}\)์˜ ์ถœ๋ ฅ์—๋งŒ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ, \(g_{\theta}\)์— ์ž‘์šฉํ•˜๋Š” ์ž„๋ฒ ๋”ฉ ํˆฌ์˜ ๋ ˆ์ด์–ด \(\mathbf{W}\)์™€ \(g_{\theta^{\prime}}\)์— ์ž‘์šฉํ•˜๋Š” ์ž„๋ฒ ๋”ฉ ํˆฌ์˜ ๋ ˆ์ด์–ด \(\mathbf{W}\cdot\mathbf{S}\)๋ฅผ ๊ตฌ๋ณ„ํ•˜๋Š” ๊ฒƒ์€ ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
๋จผ์ €, \(g_{\theta}\)์— ์ž”์—ฌ ์—ฐ๊ฒฐ์ด ์—†๊ณ  ์™„์ „ํžˆ ์—ฐ๊ฒฐ๋œ(FC) ์ตœ์ข… ๋ ˆ์ด์–ด๊ฐ€ ์žˆ๋Š” ๋” ๊ฐ„๋‹จํ•œ ๊ฒฝ์šฐ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ, ๋ชจ๋“  ๋ฐ˜์ „ ๊ฐ€๋Šฅํ•œ \(h\times h\) ํ–‰๋ ฌ \(\mathbf{S}\)์— ๋Œ€ํ•ด \(g_{\theta}\left(p\right)=\mathbf{S}_{g\theta^{\prime}}\left(p\right)\)๊ฐ€ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ \(\theta^{\prime}\)๋Š” ์ตœ์ข… FC ๋ ˆ์ด์–ด์˜ ๊ฐ€์ค‘์น˜์— \(\mathbf{S}^{-1}\)์„ ์ „์Šนํ•˜๋Š” ๊ฒƒ์„ ์ œ์™ธํ•˜๋ฉด \(\theta\)์™€ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ \(g_{\theta}\)์— ์ตœ์ข… FC ๋ ˆ์ด์–ด๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ, API \(\mathcal{O}\)์˜ ์ถœ๋ ฅ์—๋งŒ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ, ์ž„๋ฒ ๋”ฉ ํˆฌ์˜ ๋ ˆ์ด์–ด \(\mathbf{W}\)๊ฐ€ \(g_{\theta}\)์— ์ž‘์šฉํ•˜๋Š” ๊ฒƒ๊ณผ ์ž„๋ฒ ๋”ฉ ํˆฌ์˜ ๋ ˆ์ด์–ด \(\mathbf{W}\cdot\mathbf{S}\)๊ฐ€ \(g_{\theta^{\prime}}\)์— ์ž‘์šฉํ•˜๋Š” ๊ฒƒ์„ ๊ตฌ๋ณ„ํ•˜๋Š” ๊ฒƒ์ด ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
๋จผ์ €, \(g_{\theta}\)์— ์ž”์ฐจ ์—ฐ๊ฒฐ์ด ์—†๊ณ  ์™„์ „ํžˆ ์—ฐ๊ฒฐ๋œ(FC) ์ตœ์ข… ๋ ˆ์ด์–ด๊ฐ€ ์žˆ๋Š” ๋” ๊ฐ„๋‹จํ•œ ๊ฒฝ์šฐ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ, ๋ชจ๋“  ๊ฐ€์—ญ์„ฑ \(h\times h\) ํ–‰๋ ฌ \(\mathbf{S}\)์— ๋Œ€ํ•ด \(g_{\theta}\left(p\right)=\mathbf{S}_{g\theta^{\prime}}\left(p\right)\)๊ฐ€ ์žˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ \(\theta^{\prime}\)๋Š” ์ตœ์ข… FC ๊ณ„์ธต์˜ ๊ฐ€์ค‘์น˜์— \(\mathbf{S}^{-1}\)์„ ์ „์Šนํ•œ \(\theta\)์™€ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ \(g_{\theta}\)์— ์ตœ์ข… FC ๋ ˆ์ด์–ด๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ, API \(\mathcal{O}\)์˜ ์ถœ๋ ฅ์—๋งŒ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ ์ž„๋ฒ ๋”ฉ ํˆฌ์˜ ๋ ˆ์ด์–ด \(\mathbf{W}\)๊ฐ€ \(g_{\theta}\)์— ์ž‘์šฉํ•˜๋Š” ๊ฒƒ๊ณผ ์ž„๋ฒ ๋”ฉ ํˆฌ์˜ ๋ ˆ์ด์–ด \(\mathbf{W}\cdot\mathbf{S}\)๊ฐ€ \(g_{\theta^{\prime}}\)์— ์ž‘์šฉํ•˜๋Š” ๊ฒƒ์„ ๊ตฌ๋ถ„ํ•˜๋Š” ๊ฒƒ์€ ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
๋จผ์ €, \(g_{\theta}\)์— ์ž”์ฐจ ์—ฐ๊ฒฐ์ด ์—†๊ณ  ์™„์ „ํžˆ ์—ฐ๊ฒฐ๋œ(FC) ์ตœ์ข… ๋ ˆ์ด์–ด๊ฐ€ ์žˆ๋Š” ๋” ๊ฐ„๋‹จํ•œ ๊ฒฝ์šฐ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ, ๋ฐ˜์ „ ๊ฐ€๋Šฅํ•œ \(h\times h\) ํ–‰๋ ฌ \(\mathbf{S}\)๊ฐ€ ์žˆ๋‹ค๋ฉด, \(g_{\theta}\left(p\right)=\mathbf{S}_{g\theta^{\prime}}\left(p\right)\)๊ฐ€ ๋˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(\theta^{\prime}\)๋Š” ์ตœ์ข… FC ๋ ˆ์ด์–ด์˜ ๊ฐ€์ค‘์น˜๊ฐ€ \(\mathbf{S}^{-1}\)๋กœ ๋ฏธ๋ฆฌ ๊ณฑํ•ด์ง„ ๊ฒƒ์„ ์ œ์™ธํ•˜๋ฉด \(\theta\)์™€ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ \(g_{\theta}\)์— ์ตœ์ข… FC ๋ ˆ์ด์–ด๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ, API์˜ ์ถœ๋ ฅ์—๋งŒ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ(g_{\theta}\)์— ์ž‘์šฉํ•˜๋Š” ์ž„๋ฒ ๋”ฉ ํˆฌ์˜ ๋ ˆ์ด์–ด \(\mathbf{W}\)์™€ \(g_{\theta^{\prime}}\)์— ์ž‘์šฉํ•˜๋Š” ์ž„๋ฒ ๋”ฉ ํˆฌ์˜ ๋ ˆ์ด์–ด \(\mathbf{W}\cdot\mathbf{S}\)๋ฅผ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.
๋จผ์ €, \(g_{\theta}\)์— ์ž”์—ฌ ์—ฐ๊ฒฐ์ด ์—†๊ณ  ์™„์ „ํžˆ ์—ฐ๊ฒฐ๋œ(FC) ์ตœ์ข… ๋ ˆ์ด์–ด๊ฐ€ ์žˆ๋Š” ๋” ๊ฐ„๋‹จํ•œ ๊ฒฝ์šฐ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ, ๋ฐ˜์ „ ๊ฐ€๋Šฅํ•œ \(h\๋ฐฐ h\) ํ–‰๋ ฌ \(\mathbf{S}\)์˜ ๊ฒฝ์šฐ, \(g_{\theta}\left(p\right)=\mathbf{S}_{g\theta^{\prime}}\left(p\right)\)๊ฐ€ ๋˜๋ฉฐ, ์—ฌ๊ธฐ์„œ \(\theta^{\prime}\)๋Š” ์ตœ์ข… FC ๊ณ„์ธต์˜ ๊ฐ€์ค‘์น˜๊ฐ€ ์‚ฌ์ „์— \(\mathbf{S}^{-1}\)์— ์˜ํ•ด ๊ณฑ๋œ๋‹ค๋Š” ์ ์„ ์ œ์™ธํ•˜๋ฉด \(\theta\)์™€ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ \(g_{\theta}\)์— ์ตœ์ข… FC ๊ณ„์ธต์ด ์žˆ๋Š” ๊ฒฝ์šฐ, API์˜ ์ถœ๋ ฅ์—๋งŒ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ๊ฐ์•ˆํ•  ๋•Œ ์ž„๋ฒ ๋”ฉ ํˆฌ์˜ ๊ณ„์ธต \(g_{\theta}\)์— ์ž‘์šฉํ•˜๋Š” ์ž„๋ฒ ๋”ฉ ํˆฌ์˜ ๊ณ„์ธต๊ณผ \(g_{\theta^{\prime}}\)์— ์ž‘์šฉํ•˜๋Š” ์ž„๋ฒ ๋”ฉ ํˆฌ์˜ ๊ณ„์ธต์„ ๊ตฌ๋ถ„ํ•˜๋Š” ๊ฒƒ์€ ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_GPT-3_. GPT-3 [55] was released in 2020, which scaled the model parameters to an ever larger size of 175B. In the GPT-3's paper, it formally introduced the concept of in-context learning (ICL)17, which utilizes LLMs in a few-shot or zero-shot way. ICL can teach (or instruct) LLMs to understand the tasks in the form of natural language text. With ICL, the pre-training and utilization of LLMs converge to the same language modeling paradigm: pre-training predicts the following text sequence conditioned on the context, while ICL predicts the correct task solution, which can be also formatted as a text sequence, given the task description and demonstrations. GPT-3 not only demonstrates very excellent performance in a variety of NLP tasks, but also on a number of specially designed tasks that require the abilities of reasoning or domain adaptation. Although the GPT-3's paper does not explicitly discuss the emergent abilities of LLMs, we can observe large performance leap that might transcend the basic scaling law [30], _e.g.,_ larger models have significantly stronger ICL ability (illustrated in the original Figure 1.2 of the GPT-3's paper [55]). Overall, GPT-3 can be viewed as a remarkable landmark in the journey evolving from PLMs to LLMs. It has empirically proved that scaling the neural networks to a significant size can lead to a huge increase in model capacity.
\(\bullet\)_GPT-3_. GPT-3 [55]๋Š” 2020๋…„์— ์ถœ์‹œ๋˜์—ˆ์œผ๋ฉฐ, ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ 175B์˜ ๋” ํฐ ํฌ๊ธฐ๋กœ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” ๋ช‡ ๋ฒˆ์˜ ์ƒท ๋˜๋Š” ์ œ๋กœ ์ƒท ๋ฐฉ์‹์œผ๋กœ LLM์„ ํ™œ์šฉํ•˜๋Š” ์ปจํ…์ŠคํŠธ ๋‚ด ํ•™์Šต(ICL)17์ด๋ผ๋Š” ๊ฐœ๋…์„ ๊ณต์‹์ ์œผ๋กœ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ICL์€ LLM์ด ์ž์—ฐ์–ด ํ…์ŠคํŠธ ํ˜•์‹์œผ๋กœ ์ž‘์—…์„ ์ดํ•ดํ•˜๋„๋ก ๊ฐ€๋ฅด์น˜๊ฑฐ๋‚˜ ์ง€์‹œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ICL์„ ์‚ฌ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต๊ณผ LLM์˜ ํ™œ์šฉ์ด ๋™์ผํ•œ ์–ธ์–ด ๋ชจ๋ธ๋ง ํŒจ๋Ÿฌ๋‹ค์ž„์œผ๋กœ ์ˆ˜๋ ด๋ฉ๋‹ˆ๋‹ค: ์‚ฌ์ „ ํ•™์Šต์€ ์ปจํ…์ŠคํŠธ์— ๋”ฐ๋ผ ๋‹ค์Œ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋ฅผ ์˜ˆ์ธกํ•˜๊ณ , ICL์€ ์ž‘์—… ์„ค๋ช…๊ณผ ๋ฐ๋ชจ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋กœ๋„ ํ˜•์‹ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์˜ฌ๋ฐ”๋ฅธ ์ž‘์—… ์†”๋ฃจ์…˜์„ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. GPT-3๋Š” ๋‹ค์–‘ํ•œ NLP ์ž‘์—…๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ถ”๋ก  ๋˜๋Š” ๋„๋ฉ”์ธ ์ ์‘ ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•œ ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ์—ฌ๋Ÿฌ ์ž‘์—…์—์„œ๋„ ๋งค์šฐ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” LLM์˜ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด ๋ช…์‹œ์ ์œผ๋กœ ๋…ผ์˜ํ•˜์ง€๋Š” ์•Š์•˜์ง€๋งŒ, ๊ธฐ๋ณธ์ ์ธ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™[30]์„ ์ดˆ์›”ํ•˜๋Š” ํฐ ์„ฑ๋Šฅ ๋„์•ฝ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, _์˜ˆ๋ฅผ ๋“ค์–ด, ๋” ํฐ ๋ชจ๋ธ์ด ํ›จ์”ฌ ๋” ๊ฐ•๋ ฅํ•œ ICL ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1.2์˜ ์›๋ณธ [55]). ์ „๋ฐ˜์ ์œผ๋กœ GPT-3๋Š” PLM์—์„œ LLM์œผ๋กœ ์ง„ํ™”ํ•˜๋Š” ์—ฌ์ •์—์„œ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ๋žœ๋“œ๋งˆํฌ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ ๊ฒฝ๋ง์„ ์ƒ๋‹นํ•œ ํฌ๊ธฐ๋กœ ํ™•์žฅํ•˜๋ฉด ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ์ฆ๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_GPT-3_. GPT-3 [55]๋Š” 2020๋…„์— ์ถœ์‹œ๋˜์—ˆ์œผ๋ฉฐ, ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ 175B์˜ ๋” ํฐ ํฌ๊ธฐ๋กœ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-3 ๋…ผ๋ฌธ์—์„œ๋Š” ์†Œ์ˆ˜ ์ƒท ๋˜๋Š” ์ œ๋กœ ์ƒท ๋ฐฉ์‹์œผ๋กœ LLM์„ ํ™œ์šฉํ•˜๋Š” ์ปจํ…์ŠคํŠธ ๋‚ด ํ•™์Šต(ICL)17์ด๋ผ๋Š” ๊ฐœ๋…์„ ๊ณต์‹์ ์œผ๋กœ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ICL์€ LLM์ด ์ž์—ฐ์–ด ํ…์ŠคํŠธ ํ˜•์‹์œผ๋กœ ์ž‘์—…์„ ์ดํ•ดํ•˜๋„๋ก ๊ฐ€๋ฅด์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ICL์„ ์‚ฌ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต๊ณผ LLM์˜ ํ™œ์šฉ์ด ๋™์ผํ•œ ์–ธ์–ด ๋ชจ๋ธ๋ง ํŒจ๋Ÿฌ๋‹ค์ž„์— ์ˆ˜๋ ดํ•ฉ๋‹ˆ๋‹ค: ์‚ฌ์ „ ํ•™์Šต์€ ์ปจํ…์ŠคํŠธ์— ๋”ฐ๋ผ ๋‹ค์Œ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋ฅผ ์˜ˆ์ธกํ•˜๊ณ , ICL์€ ์ž‘์—… ์„ค๋ช…๊ณผ ๋ฐ๋ชจ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋กœ ํ˜•์‹ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์˜ฌ๋ฐ”๋ฅธ ์ž‘์—… ์†”๋ฃจ์…˜์„ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. GPT-3๋Š” ๋‹ค์–‘ํ•œ NLP ์ž‘์—…๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ถ”๋ก  ๋˜๋Š” ๋„๋ฉ”์ธ ์ ์‘ ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•œ ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ์—ฌ๋Ÿฌ ์ž‘์—…์—์„œ๋„ ๋งค์šฐ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” LLM์˜ ์ถœํ˜„ ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด ๋ช…์‹œ์ ์œผ๋กœ ๋…ผ์˜ํ•˜์ง€๋Š” ์•Š์•˜์ง€๋งŒ, ๊ธฐ๋ณธ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™[30]์„ ์ดˆ์›”ํ•˜๋Š” ํฐ ์„ฑ๋Šฅ ๋„์•ฝ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋” ํฐ ๋ชจ๋ธ์€ ํ›จ์”ฌ ๋” ๊ฐ•๋ ฅํ•œ ICL ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1.2์˜ ์›๋ณธ [55]). ์ „๋ฐ˜์ ์œผ๋กœ GPT-3๋Š” PLM์—์„œ LLM์œผ๋กœ ์ง„ํ™”ํ•˜๋Š” ์—ฌ์ •์—์„œ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ๋žœ๋“œ๋งˆํฌ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ ๊ฒฝ๋ง์„ ์ƒ๋‹นํ•œ ํฌ๊ธฐ๋กœ ํ™•์žฅํ•˜๋ฉด ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์‹ค์ฆ์ ์œผ๋กœ ์ฆ๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_GPT-3_. 2020๋…„์— ์ถœ์‹œ๋œ GPT-3 [55]๋Š” ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ 175B์˜ ๋” ํฐ ๊ทœ๋ชจ๋กœ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” ์†Œ์ˆ˜ ์ƒท ๋˜๋Š” ์ œ๋กœ ์ƒท ๋ฐฉ์‹์œผ๋กœ LLM์„ ํ™œ์šฉํ•˜๋Š” ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต(ICL)17์˜ ๊ฐœ๋…์„ ๊ณต์‹์ ์œผ๋กœ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ICL์€ LLM์ด ์ž์—ฐ์–ด ํ…์ŠคํŠธ ํ˜•์‹์œผ๋กœ ์ž‘์—…์„ ์ดํ•ดํ•˜๋„๋ก ๊ฐ€๋ฅด์น˜๊ฑฐ๋‚˜ ์ง€์‹œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ICL์„ ์‚ฌ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต๊ณผ LLM์˜ ํ™œ์šฉ์ด ๋™์ผํ•œ ์–ธ์–ด ๋ชจ๋ธ ํŒจ๋Ÿฌ๋‹ค์ž„์— ์ˆ˜๋ ดํ•˜๋Š”๋ฐ, ์‚ฌ์ „ ํ•™์Šต์€ ๋ฌธ๋งฅ์— ๋”ฐ๋ผ ๋‹ค์Œ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋ฅผ ์˜ˆ์ธกํ•˜๊ณ , ICL์€ ์ž‘์—… ์„ค๋ช…๊ณผ ๋ฐ๋ชจ๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ํ…์ŠคํŠธ ์‹œํ€€์Šค๋กœ๋„ ํ˜•์‹ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์˜ฌ๋ฐ”๋ฅธ ์ž‘์—… ์†”๋ฃจ์…˜์„ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. GPT-3๋Š” ๋‹ค์–‘ํ•œ NLP ์ž‘์—…๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ถ”๋ก  ๋˜๋Š” ๋„๋ฉ”์ธ ์ ์‘ ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•œ ํŠน์ˆ˜ ์„ค๊ณ„๋œ ์—ฌ๋Ÿฌ ์ž‘์—…์—์„œ๋„ ๋งค์šฐ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” LLM์˜ ์ถœํ˜„ ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด ๋ช…์‹œ์ ์œผ๋กœ ๋…ผ์˜ํ•˜์ง€ ์•Š์•˜์ง€๋งŒ, ๊ธฐ๋ณธ์ ์ธ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™[30]์„ ์ดˆ์›”ํ•  ์ˆ˜ ์žˆ๋Š” ํฐ ์„ฑ๋Šฅ ๋„์•ฝ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋” ํฐ ๋ชจ๋ธ์€ ํ›จ์”ฌ ๋” ๊ฐ•๋ ฅํ•œ ICL ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(์›๋ณธ ๊ทธ๋ฆผ 1.2์— ์„ค๋ช…๋œ GPT-3 ๋…ผ๋ฌธ [55]). ์ „๋ฐ˜์ ์œผ๋กœ GPT-3๋Š” PLM์—์„œ LLM์œผ๋กœ ์ง„ํ™”ํ•˜๋Š” ์—ฌ์ •์—์„œ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ๋žœ๋“œ๋งˆํฌ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ ๊ฒฝ๋ง์„ ์ƒ๋‹นํ•œ ํฌ๊ธฐ๋กœ ํ™•์žฅํ•˜๋ฉด ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์‹ค์ฆ์ ์œผ๋กœ ์ฆ๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค.
GPT-3_. GPT-3 [55]๋Š” 2020๋…„์— ์ถœ์‹œ๋˜์—ˆ์œผ๋ฉฐ, ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ํฌ๊ธฐ๋ฅผ 175B๋กœ ๋”์šฑ ํฌ๊ฒŒ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” ๋ช‡ ๊ฐ€์ง€ ์ƒท ๋˜๋Š” ์ œ๋กœ ์ƒท ๋ฐฉ์‹์œผ๋กœ LLM์„ ํ™œ์šฉํ•˜๋Š” ์ธ ์ปจํ…์ŠคํŠธ ํ•™์Šต(ICL)17์˜ ๊ฐœ๋…์„ ๊ณต์‹์ ์œผ๋กœ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ICL์€ LLM์ด ์ž์—ฐ์–ด ํ…์ŠคํŠธ ํ˜•์‹์œผ๋กœ ์ž‘์—…์„ ์ดํ•ดํ•˜๋„๋ก ๊ฐ€๋ฅด์น (๋˜๋Š” ์ง€์‹œํ• ) ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ICL์„ ์‚ฌ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต๊ณผ LLM์˜ ํ™œ์šฉ์ด ๋™์ผํ•œ ์–ธ์–ด ๋ชจ๋ธ๋ง ํŒจ๋Ÿฌ๋‹ค์ž„, ์ฆ‰ ์‚ฌ์ „ ํ•™์Šต์€ ์ปจํ…์ŠคํŠธ์— ๋”ฐ๋ผ ๋‹ค์Œ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋ฅผ ์˜ˆ์ธกํ•˜๊ณ , ICL์€ ์ž‘์—… ์„ค๋ช…๊ณผ ๋ฐ๋ชจ๊ฐ€ ์ฃผ์–ด์ง„ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋กœ ํ˜•์‹ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์˜ฌ๋ฐ”๋ฅธ ์ž‘์—… ์†”๋ฃจ์…˜์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์ˆ˜๋ ดํ•ฉ๋‹ˆ๋‹ค. GPT-3๋Š” ๋‹ค์–‘ํ•œ NLP ์ž‘์—…๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ถ”๋ก ์ด๋‚˜ ๋„๋ฉ”์ธ ์ ์‘ ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•œ ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ์ž‘์—…์—์„œ๋„ ๋งค์šฐ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” ๋ช…์‹œ์ ์œผ๋กœ LLM์˜ ์ถœํ˜„ ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•˜์ง€ ์•Š์•˜์ง€๋งŒ, ๊ธฐ๋ณธ์ ์ธ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ๋›ฐ์–ด๋„˜๋Š” ํฐ ์„ฑ๋Šฅ ๋„์•ฝ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[30]. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋” ํฐ ๋ชจ๋ธ์ด ํ›จ์”ฌ ๋” ๊ฐ•๋ ฅํ•œ ICL ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(GPT-3 ๋…ผ๋ฌธ์˜ ์›๋ž˜ ๊ทธ๋ฆผ 1.2์— ์„ค๋ช…๋˜์–ด ์žˆ์Œ[55]). ์ „๋ฐ˜์ ์œผ๋กœ GPT-3๋Š” PLM์—์„œ LLM์œผ๋กœ ์ง„ํ™”ํ•˜๋Š” ์—ฌ์ •์—์„œ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์ด์ •ํ‘œ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ ๊ฒฝ๋ง์„ ์ƒ๋‹นํ•œ ํฌ๊ธฐ๋กœ ํ™•์žฅํ•˜๋ฉด ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ์ฆ๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค.
\_GPT-3\_. GPT-3 [55]๋Š” 2020๋…„์— ์ถœ์‹œ๋˜์—ˆ์œผ๋ฉฐ, ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ 175B๋ผ๋Š” ํ›จ์”ฌ ๋” ํฐ ํฌ๊ธฐ๋กœ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” ๋ช‡ ์ƒท ๋˜๋Š” ์ œ๋กœ ์ƒท ๋ฐฉ์‹์œผ๋กœ LLM์„ ํ™œ์šฉํ•˜๋Š” ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต(ICL)17์˜ ๊ฐœ๋…์„ ๊ณต์‹์ ์œผ๋กœ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ICL์€ ์ž์—ฐ์–ด ํ…์ŠคํŠธ ํ˜•ํƒœ๋กœ ์ž‘์—…์„ ์ดํ•ดํ•˜๋„๋ก LLM์„ ๊ฐ€๋ฅด์น˜๊ฑฐ๋‚˜ ์ง€์‹œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ICL์„ ์‚ฌ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต๊ณผ LLM์˜ ํ™œ์šฉ์ด ๋™์ผํ•œ ์–ธ์–ด ๋ชจ๋ธ๋ง ํŒจ๋Ÿฌ๋‹ค์ž„, ์ฆ‰ ์‚ฌ์ „ ํ•™์Šต์€ ์ปจํ…์ŠคํŠธ์— ๋”ฐ๋ผ ๋‹ค์Œ ํ…์ŠคํŠธ ์‹œํ€€์Šค๋ฅผ ์˜ˆ์ธกํ•˜๊ณ , ICL์€ ์ž‘์—… ์„ค๋ช…๊ณผ ๋ฐ๋ชจ๊ฐ€ ์ฃผ์–ด์ง€๋ฉด ํ…์ŠคํŠธ ์‹œํ€€์Šค๋กœ๋„ ํ˜•์‹ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์˜ฌ๋ฐ”๋ฅธ ์ž‘์—… ์†”๋ฃจ์…˜์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์ˆ˜๋ ด๋ฉ๋‹ˆ๋‹ค. GPT-3๋Š” ๋‹ค์–‘ํ•œ NLP ์ž‘์—…๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ถ”๋ก ์ด๋‚˜ ๋„๋ฉ”์ธ ์ ์‘ ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•œ ํŠน๋ณ„ํžˆ ์„ค๊ณ„๋œ ์ž‘์—…์—์„œ๋„ ๋งค์šฐ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. GPT-3์˜ ๋…ผ๋ฌธ์—์„œ๋Š” ๋ช…์‹œ์ ์œผ๋กœ ๋…ผ์˜ํ•˜์ง€ ์•Š์•˜์ง€๋งŒ, ๊ธฐ๋ณธ์ ์ธ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ๋›ฐ์–ด๋„˜๋Š” ํฐ ์„ฑ๋Šฅ ๋„์•ฝ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[30]. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋” ํฐ ๋ชจ๋ธ์ด ํ›จ์”ฌ ๋” ๊ฐ•๋ ฅํ•œ ICL ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(GPT-3 ๋…ผ๋ฌธ์˜ ์›๋ž˜ ๊ทธ๋ฆผ 1.2์— ์„ค๋ช…๋˜์–ด ์žˆ์Œ[55]). ์ „๋ฐ˜์ ์œผ๋กœ GPT-3๋Š” PLM์—์„œ LLM์œผ๋กœ ์ง„ํ™”ํ•˜๋Š” ์—ฌ์ •์—์„œ ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์ด์ •ํ‘œ๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ ๊ฒฝ๋ง์„ ์ƒ๋‹นํ•œ ํฌ๊ธฐ๋กœ ํ™•์žฅํ•˜๋ฉด ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํฌ๊ฒŒ ์ฆ๊ฐ€ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ์ฆ๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค.
scale of losses on both the overall validation data and different validation domains. While the overall loss helps optimize the overall performance, losses on different domains show model capabilities in various aspects. We suppose that this facilitates further mixture curating for the nuanced balance of model capabilities.
์ „์ฒด ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋ฐ์ดํ„ฐ์™€ ๋‹ค์–‘ํ•œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ์†์‹ค์˜ ํฌ๊ธฐ๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ์†์‹ค์€ ์ „์ฒด ์„ฑ๋Šฅ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜์ง€๋งŒ, ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์˜ ์†์‹ค์€ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์—์„œ ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ ๊ธฐ๋Šฅ์˜ ๋ฏธ๋ฌ˜ํ•œ ๊ท ํ˜•์„ ์œ„ํ•ด ํ˜ผํ•ฉ ํ๋ ˆ์ด์…˜์„ ๋” ์‰ฝ๊ฒŒ ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค.
์ „์ฒด ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋ฐ์ดํ„ฐ์™€ ๋‹ค์–‘ํ•œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋„๋ฉ”์ธ์— ๋Œ€ํ•œ ์†์‹ค์˜ ๋น„์œจ์„ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. ์ „์ฒด ์†์‹ค์€ ์ „์ฒด ์„ฑ๋Šฅ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜์ง€๋งŒ, ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์˜ ์†์‹ค์€ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์—์„œ ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ ๊ธฐ๋Šฅ์˜ ๋ฏธ๋ฌ˜ํ•œ ๊ท ํ˜•์„ ์œ„ํ•ด ํ˜ผํ•ฉ ํ๋ ˆ์ด์…˜์„ ๋” ์‰ฝ๊ฒŒ ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค.
์ „์ฒด ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋ฐ์ดํ„ฐ์™€ ๋‹ค์–‘ํ•œ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ๋„๋ฉ”์ธ ๋ชจ๋‘์— ๋Œ€ํ•œ ์†์‹ค์˜ ๊ทœ๋ชจ. ์ „์ฒด ์†์‹ค์€ ์ „์ฒด ์„ฑ๋Šฅ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜์ง€๋งŒ, ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์˜ ์†์‹ค์€ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์—์„œ ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ฏธ๋ฌ˜ํ•œ ๊ท ํ˜•์˜ ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ์œ„ํ•œ ํ˜ผํ•ฉ๋ฌผ ํ๋ ˆ์ด์…˜์„ ์šฉ์ดํ•˜๊ฒŒ ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
์ „์ฒด ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ์™€ ๋‹ค์–‘ํ•œ ๊ฒ€์ฆ ๋„๋ฉ”์ธ์˜ ์†์‹ค ๊ทœ๋ชจ. ์ „์ฒด ์†์‹ค์€ ์ „์ฒด ์„ฑ๋Šฅ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜์ง€๋งŒ, ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์˜ ์†์‹ค์€ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์—์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Š” ๋ชจ๋ธ ๊ธฐ๋Šฅ์˜ ๋ฏธ๋ฌ˜ํ•œ ๊ท ํ˜•์„ ์œ„ํ•ด ์ถ”๊ฐ€ ํ˜ผํ•ฉ ํ๋ ˆ์ดํŒ…์„ ์šฉ์ดํ•˜๊ฒŒ ํ•œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค.
์ „์ฒด ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ์™€ ๋‹ค์–‘ํ•œ ๊ฒ€์ฆ ๋„๋ฉ”์ธ์˜ ์†์‹ค ๊ทœ๋ชจ๋ฅผ ๋ชจ๋‘ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ „์ฒด ์†์‹ค์€ ์ „์ฒด ์„ฑ๋Šฅ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜์ง€๋งŒ, ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์˜ ์†์‹ค์€ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์—์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ ๊ธฐ๋Šฅ์˜ ๋ฏธ๋ฌ˜ํ•œ ๊ท ํ˜•์„ ๋งž์ถ”๊ธฐ ์œ„ํ•œ ์ถ”๊ฐ€ ํ˜ผํ•ฉ ํ๋ ˆ์ดํŒ…์ด ์šฉ์ดํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
As the field of LLMs is moving fast, with new findings, models and techniques being published in a matter of months or weeks [7, 8, 9, 10, 11], AI researchers and practitioners often find it challenging to figure out the best recipes to build LLM-powered AI systems for their tasks. This paper gives a timely survey of the recent advances on LLMs. We hope this survey will prove a valuable and accessible resource for students, researchers and developers.
LLM ๋ถ„์•ผ๋Š” ๋ช‡ ๋‹ฌ ๋˜๋Š” ๋ช‡ ์ฃผ ๋งŒ์— ์ƒˆ๋กœ์šด ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, ๋ชจ๋ธ ๋ฐ ๊ธฐ์ˆ ์ด ๋ฐœํ‘œ๋˜๋Š” ๋“ฑ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์— [7, 8, 9, 10, 11], AI ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž๋Š” ์ข…์ข… ์ž์‹ ์˜ ์ž‘์—…์— ๋งž๋Š” LLM ๊ธฐ๋ฐ˜ AI ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ƒ์˜ ๋ ˆ์‹œํ”ผ๋ฅผ ํŒŒ์•…ํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์€ LLM์— ๋Œ€ํ•œ ์ตœ๊ทผ์˜ ๋ฐœ์ „์„ ์‹œ์˜์ ์ ˆํ•˜๊ฒŒ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ด ์„ค๋ฌธ์กฐ์‚ฌ๊ฐ€ ํ•™์ƒ, ์—ฐ๊ตฌ์ž ๋ฐ ๊ฐœ๋ฐœ์ž์—๊ฒŒ ์œ ์šฉํ•˜๊ณ  ์ ‘๊ทผํ•˜๊ธฐ ์‰ฌ์šด ๋ฆฌ์†Œ์Šค๊ฐ€ ๋  ์ˆ˜ ์žˆ๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
LLM ๋ถ„์•ผ๋Š” ๋ช‡ ๋‹ฌ ๋˜๋Š” ๋ช‡ ์ฃผ ๋งŒ์— ์ƒˆ๋กœ์šด ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, ๋ชจ๋ธ ๋ฐ ๊ธฐ์ˆ ์ด ๋ฐœํ‘œ๋˜๋Š” ๋“ฑ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์— [7, 8, 9, 10, 11], AI ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž๋Š” ์ข…์ข… ์ž์‹ ์˜ ์ž‘์—…์— ๋งž๋Š” LLM ๊ธฐ๋ฐ˜ AI ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ƒ์˜ ๋ ˆ์‹œํ”ผ๋ฅผ ํŒŒ์•…ํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” LLM์˜ ์ตœ๊ทผ ๋ฐœ์ „์— ๋Œ€ํ•œ ์‹œ์˜์ ์ ˆํ•œ ์กฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ด ์„ค๋ฌธ์กฐ์‚ฌ๊ฐ€ ํ•™์ƒ, ์—ฐ๊ตฌ์ž ๋ฐ ๊ฐœ๋ฐœ์ž์—๊ฒŒ ์œ ์šฉํ•˜๊ณ  ์ ‘๊ทผํ•˜๊ธฐ ์‰ฌ์šด ์ž๋ฃŒ๊ฐ€ ๋  ์ˆ˜ ์žˆ๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
LLM ๋ถ„์•ผ๋Š” ๋ช‡ ๋‹ฌ ๋˜๋Š” ๋ช‡ ์ฃผ ์•ˆ์— ์ƒˆ๋กœ์šด ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, ๋ชจ๋ธ ๋ฐ ๊ธฐ์ˆ ์ด ๋ฐœํ‘œ๋˜๋Š” ๋“ฑ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[7, 8, 9, 10, 11], AI ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž๋Š” ์ข…์ข… ์ž์‹ ์˜ ์ž‘์—…์— ๋งž๋Š” LLM ๊ธฐ๋ฐ˜ AI ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ ์˜ ๋ ˆ์‹œํ”ผ๋ฅผ ์ฐพ๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” LLM์— ๋Œ€ํ•œ ์ตœ๊ทผ์˜ ๋ฐœ์ „์„ ์‹œ๊ธฐ์ ์ ˆํ•˜๊ฒŒ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ๊ฐ€ ํ•™์ƒ, ์—ฐ๊ตฌ์ž ๋ฐ ๊ฐœ๋ฐœ์ž์—๊ฒŒ ์œ ์šฉํ•˜๊ณ  ์ ‘๊ทผํ•˜๊ธฐ ์‰ฌ์šด ์ž๋ฃŒ๊ฐ€ ๋˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
LLM ๋ถ„์•ผ๋Š” ๋ช‡ ๋‹ฌ ๋˜๋Š” ๋ช‡ ์ฃผ ๋งŒ์— ์ƒˆ๋กœ์šด ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, ๋ชจ๋ธ ๋ฐ ๊ธฐ์ˆ ์ด ๋ฐœํ‘œ๋˜๋Š” ๋“ฑ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์—[7, 8, 9, 10, 11] AI ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž๋Š” ์ž‘์—…์„ ์œ„ํ•œ LLM ๊ธฐ๋ฐ˜ AI ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ ์˜ ๋ ˆ์‹œํ”ผ๋ฅผ ํŒŒ์•…ํ•˜๋Š” ๊ฒƒ์ด ์–ด๋ ค์šด ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” LLM์˜ ์ตœ๊ทผ ๋ฐœ์ „์— ๋Œ€ํ•œ ์‹œ๊ธฐ์ ์ ˆํ•œ ์กฐ์‚ฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ์กฐ์‚ฌ๊ฐ€ ํ•™์ƒ, ์—ฐ๊ตฌ์ž ๋ฐ ๊ฐœ๋ฐœ์ž์—๊ฒŒ ์œ ์šฉํ•˜๊ณ  ์ ‘๊ทผํ•˜๊ธฐ ์‰ฌ์šด ์ž๋ฃŒ๊ฐ€ ๋˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
LLM ๋ถ„์•ผ๋Š” ๋ช‡ ๋‹ฌ ๋˜๋Š” ๋ช‡ ์ฃผ ๋งŒ์— ์ƒˆ๋กœ์šด ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, ๋ชจ๋ธ ๋ฐ ๊ธฐ์ˆ ์ด ๋ฐœํ‘œ๋˜๋Š” ๋“ฑ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์—[7, 8, 9, 10, 11] AI ์—ฐ๊ตฌ์ž์™€ ์‹ค๋ฌด์ž๋“ค์€ ์ข…์ข… ์ž์‹ ์˜ ์ž‘์—…์— ๋งž๋Š” LLM ๊ธฐ๋ฐ˜ AI ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ๋ ˆ์‹œํ”ผ๋ฅผ ์ฐพ๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” LLM์— ๋Œ€ํ•œ ์ตœ๊ทผ์˜ ๋ฐœ์ „์„ ์‹œ๊ธฐ์ ์ ˆํ•˜๊ฒŒ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ๊ฐ€ ํ•™์ƒ, ์—ฐ๊ตฌ์ž, ๊ฐœ๋ฐœ์ž์—๊ฒŒ ์œ ์šฉํ•˜๊ณ  ์ ‘๊ทผํ•˜๊ธฐ ์‰ฌ์šด ์ž๋ฃŒ๊ฐ€ ๋˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
Footnote 6: For instance, suppose Hoffmann et al. conducted 500 training runs, each generating 1,000 intermediate loss values. To account for the variance both between and within these training runs, we use the intra-group correlation coefficient, denoted as \(\rho\). Conservatively assuming \(\rho=0.5\), the effective number of data points can be calculated using the formula:
์ฃผ 6: ์˜ˆ๋ฅผ ๋“ค์–ด, Hoffmann ๋“ฑ์ด ๊ฐ๊ฐ 1,000๊ฐœ์˜ ์ค‘๊ฐ„ ์†์‹ค ๊ฐ’์„ ์ƒ์„ฑํ•˜๋Š” 500๊ฐœ์˜ ํ›ˆ๋ จ ์‹คํ–‰์„ ์ˆ˜ํ–‰ํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ›ˆ๋ จ ์‹คํ–‰ ๊ฐ„์˜ ๋ถ„์‚ฐ๊ณผ ํ›ˆ๋ จ ์‹คํ–‰ ๋‚ด์˜ ๋ถ„์‚ฐ์„ ๋ชจ๋‘ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฃน ๋‚ด ์ƒ๊ด€ ๊ณ„์ˆ˜์ธ \(\rho\)๋กœ ํ‘œ์‹œ๋˜๋Š” ๊ณ„์ˆ˜๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ณด์ˆ˜์ ์œผ๋กœ \(\rho=0.5\)๋ผ๊ณ  ๊ฐ€์ •ํ•˜๋ฉด, ๋‹ค์Œ ๊ณต์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ์œ ํšจํ•œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
์ฃผ 6: ์˜ˆ๋ฅผ ๋“ค์–ด, ํ˜ธํ”„๋งŒ ๋“ฑ์ด ๊ฐ๊ฐ 1,000๊ฐœ์˜ ์ค‘๊ฐ„ ์†์‹ค ๊ฐ’์„ ์ƒ์„ฑํ•˜๋Š” 500๊ฐœ์˜ ํ›ˆ๋ จ ์‹คํ–‰์„ ์ˆ˜ํ–‰ํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ›ˆ๋ จ ์‹คํ–‰ ๊ฐ„์˜ ๋ถ„์‚ฐ๊ณผ ํ›ˆ๋ จ ์‹คํ–‰ ๋‚ด์˜ ๋ถ„์‚ฐ์„ ๋ชจ๋‘ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฃน ๋‚ด ์ƒํ˜ธ ์ƒ๊ด€ ๊ณ„์ˆ˜(\(\rho\))๋กœ ํ‘œ์‹œ๋˜๋Š” ์ƒํ˜ธ ์ƒ๊ด€ ๊ณ„์ˆ˜๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ณด์ˆ˜์ ์œผ๋กœ \(\rho=0.5\)๋ผ๊ณ  ๊ฐ€์ •ํ•˜๋ฉด ๋‹ค์Œ ๊ณต์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ์œ ํšจํ•œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
์ฃผ 6: ์˜ˆ๋ฅผ ๋“ค์–ด, Hoffmann ๋“ฑ์ด ๊ฐ๊ฐ 1,000๊ฐœ์˜ ์ค‘๊ฐ„ ์†์‹ค ๊ฐ’์„ ์ƒ์„ฑํ•˜๋Š” 500๊ฐœ์˜ ํ›ˆ๋ จ ์‹คํ–‰์„ ์ˆ˜ํ–‰ํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ›ˆ๋ จ ์‹คํ–‰ ๊ฐ„ ๋ฐ ํ›ˆ๋ จ ์‹คํ–‰ ๋‚ด์˜ ๋ถ„์‚ฐ์„ ๋ชจ๋‘ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฃน ๋‚ด ์ƒ๊ด€ ๊ณ„์ˆ˜(\(\rho\)๋กœ ํ‘œ์‹œ๋จ)๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ณด์ˆ˜์ ์œผ๋กœ \(\rho=0.5\)๋ผ๊ณ  ๊ฐ€์ •ํ•˜๋ฉด ๋‹ค์Œ ๊ณต์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ์œ ํšจํ•œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
๊ฐ 1,000๊ฐœ์˜ ์ค‘๊ฐ„ ์†์‹ค ๊ฐ’์„ ์ƒ์„ฑํ•˜๋Š” 500๊ฐœ์˜ ํ›ˆ๋ จ ์‹คํ–‰์„ ์ˆ˜ํ–‰ํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ›ˆ๋ จ ์‹คํ–‰ ๊ฐ„ ๋ฐ ์‹คํ–‰ ๋‚ด์˜ ๋ถ„์‚ฐ์„ ๋ชจ๋‘ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฃน ๋‚ด ์ƒ๊ด€ ๊ณ„์ˆ˜๋กœ ํ‘œ์‹œ๋˜๋Š” \(\rho\)๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ณด์ˆ˜์ ์œผ๋กœ \(\rho=0.5\)๋ฅผ ๊ฐ€์ •ํ•˜๋ฉด ๋‹ค์Œ ๊ณต์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ์œ ํšจ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
๊ฐ์ฃผ 6: ์˜ˆ๋ฅผ ๋“ค์–ด ํ˜ธํ”„๋งŒ ๋“ฑ์ด 500๋ฒˆ์˜ ํ›ˆ๋ จ ์‹คํ–‰์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ๊ฐ๊ฐ 1,000๊ฐœ์˜ ์ค‘๊ฐ„ ์†์‹ค ๊ฐ’์„ ์ƒ์„ฑํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ›ˆ๋ จ ์‹คํ–‰ ๊ฐ„ ๋ฐ ์‹คํ–‰ ๋‚ด ๋ถ„์‚ฐ์„ ๋ชจ๋‘ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด ๊ทธ๋ฃน ๋‚ด ์ƒ๊ด€ ๊ณ„์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ์ด๋Š” \(\rho\)๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ๋ณด์ˆ˜์ ์œผ๋กœ \(\rho=0.5\)๋ฅผ ๊ฐ€์ •ํ•˜๋ฉด ๋‹ค์Œ ๊ณต์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ์œ ํšจ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
Accompanied by the exponential growth in the agent quantity and complexity, there is an increasing strain on the functionalities of LLM and OS. For example, scheduling and prioritizing agent requests in limited LLM resources poses a significant challenge. Moreover, the LLM's generation process can become time-intensive when dealing with lengthy contexts, occasionally resulting in the generation being suspended by the scheduler. This raises the problem of devising a mechanism to snapshot the LLM's current generation result, thereby enabling pause/resume behavior even when the LLM has not finalized the response generation for the current request. Furthermore, once an agent has obtained the list of available calling tools, determining the optimal sequence for invoking these tools presents yet another challenge since multiple agents may need to call the same tool. Additionally, the concurrent operation of multiple agents necessitates a robust system for memory management across different agents, while also ensuring stringent enforcement of privacy and access control measures.
์—์ด์ „ํŠธ ์ˆ˜์™€ ๋ณต์žก์„ฑ์˜ ๊ธฐํ•˜๊ธ‰์ˆ˜์ ์ธ ์ฆ๊ฐ€์— ๋”ฐ๋ผ LLM๊ณผ OS์˜ ๊ธฐ๋Šฅ์— ๋Œ€ํ•œ ๋ถ€๋‹ด์ด ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ œํ•œ๋œ LLM ๋ฆฌ์†Œ์Šค์—์„œ ์—์ด์ „ํŠธ ์š”์ฒญ์˜ ์˜ˆ์•ฝ ๋ฐ ์šฐ์„ ์ˆœ์œ„ ์ง€์ •์€ ์ƒ๋‹นํ•œ ์–ด๋ ค์›€์„ ์ดˆ๋ž˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธด ์ปจํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ LLM์˜ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๊ฐ€ ์‹œ๊ฐ„์ด ๋งŽ์ด ์†Œ์š”๋˜์–ด ์Šค์ผ€์ค„๋Ÿฌ์— ์˜ํ•ด ์ƒ์„ฑ์ด ์ค‘๋‹จ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋กœ ์ธํ•ด ํ˜„์žฌ ์š”์ฒญ์— ๋Œ€ํ•œ ์‘๋‹ต ์ƒ์„ฑ์„ ์™„๋ฃŒํ•˜์ง€ ์•Š์€ ์ƒํƒœ์—์„œ LLM์˜ ํ˜„์žฌ ์ƒ์„ฑ ๊ฒฐ๊ณผ๋ฅผ ์Šค๋ƒ…์ƒท์œผ๋กœ ์ƒ์„ฑํ•˜์—ฌ ์ผ์‹œ ์ค‘์ง€/์žฌ๊ฐœ ๋™์ž‘์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ณ ์•ˆํ•˜๋Š” ๋ฌธ์ œ๊ฐ€ ์ œ๊ธฐ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์—์ด์ „ํŠธ๊ฐ€ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ†ตํ™” ๋„๊ตฌ ๋ชฉ๋ก์„ ํ™•๋ณดํ•œ ํ›„์—๋„ ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๊ฐ€ ๋™์ผํ•œ ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•ด์•ผ ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ด๋Ÿฌํ•œ ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ ์˜ ์ˆœ์„œ๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ๋„ ๋˜ ๋‹ค๋ฅธ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ, ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ๋™์‹œ ์ž‘๋™์„ ์œ„ํ•ด์„œ๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ๋ฅผ ์œ„ํ•œ ๊ฐ•๋ ฅํ•œ ์‹œ์Šคํ…œ์ด ํ•„์š”ํ•˜๋ฉฐ, ํ”„๋ผ์ด๋ฒ„์‹œ ๋ฐ ์•ก์„ธ์Šค ์ œ์–ด ์กฐ์น˜๋„ ์—„๊ฒฉํ•˜๊ฒŒ ์‹œํ–‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
์—์ด์ „ํŠธ ์ˆ˜์™€ ๋ณต์žก์„ฑ์˜ ๊ธฐํ•˜๊ธ‰์ˆ˜์ ์ธ ์ฆ๊ฐ€์— ๋”ฐ๋ผ LLM๊ณผ OS์˜ ๊ธฐ๋Šฅ์— ๋Œ€ํ•œ ๋ถ€๋‹ด์ด ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ œํ•œ๋œ LLM ๋ฆฌ์†Œ์Šค์—์„œ ์—์ด์ „ํŠธ ์š”์ฒญ์˜ ์Šค์ผ€์ค„๋ง๊ณผ ์šฐ์„ ์ˆœ์œ„ ์ง€์ •์€ ์ƒ๋‹นํ•œ ์–ด๋ ค์›€์„ ์ดˆ๋ž˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ๊ธด ๋ฌธ๋งฅ์„ ๋‹ค๋ฃฐ ๋•Œ LLM์˜ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๊ฐ€ ์‹œ๊ฐ„์ด ๋งŽ์ด ์†Œ์š”๋˜์–ด ์Šค์ผ€์ค„๋Ÿฌ์— ์˜ํ•ด ์ƒ์„ฑ์ด ์ค‘๋‹จ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋กœ ์ธํ•ด ํ˜„์žฌ ์š”์ฒญ์— ๋Œ€ํ•œ ์‘๋‹ต ์ƒ์„ฑ์„ ์™„๋ฃŒํ•˜์ง€ ์•Š์€ ์ƒํƒœ์—์„œ LLM์˜ ํ˜„์žฌ ์ƒ์„ฑ ๊ฒฐ๊ณผ๋ฅผ ์Šค๋ƒ…์ƒท์œผ๋กœ ์ƒ์„ฑํ•˜์—ฌ ์ผ์‹œ ์ค‘์ง€/์žฌ๊ฐœ ๋™์ž‘์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ณ ์•ˆํ•˜๋Š” ๋ฌธ์ œ๊ฐ€ ์ œ๊ธฐ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์—์ด์ „ํŠธ๊ฐ€ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ†ตํ™” ๋„๊ตฌ ๋ชฉ๋ก์„ ํ™•๋ณดํ•œ ํ›„์—๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๊ฐ€ ๋™์ผํ•œ ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•ด์•ผ ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ด๋Ÿฌํ•œ ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ ์˜ ์ˆœ์„œ๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์ด ๋˜ ๋‹ค๋ฅธ ๊ณผ์ œ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ๋™์‹œ ์ž‘๋™์€ ์„œ๋กœ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ๋ฅผ ์œ„ํ•œ ๊ฐ•๋ ฅํ•œ ์‹œ์Šคํ…œ์ด ํ•„์š”ํ•˜๋ฉฐ, ํ”„๋ผ์ด๋ฒ„์‹œ ๋ฐ ์•ก์„ธ์Šค ์ œ์–ด ์กฐ์น˜๋ฅผ ์—„๊ฒฉํ•˜๊ฒŒ ์‹œํ–‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
์—์ด์ „ํŠธ ์ˆ˜์™€ ๋ณต์žก์„ฑ์˜ ๊ธฐํ•˜๊ธ‰์ˆ˜์ ์ธ ์ฆ๊ฐ€์™€ ํ•จ๊ป˜ LLM ๋ฐ OS์˜ ๊ธฐ๋Šฅ์— ๋Œ€ํ•œ ๋ถ€๋‹ด์ด ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ œํ•œ๋œ LLM ๋ฆฌ์†Œ์Šค์—์„œ ์—์ด์ „ํŠธ ์š”์ฒญ์„ ์˜ˆ์•ฝํ•˜๊ณ  ์šฐ์„ ์ˆœ์œ„๋ฅผ ์ •ํ•˜๋Š” ๊ฒƒ์€ ์ƒ๋‹นํ•œ ์–ด๋ ค์›€์„ ์ดˆ๋ž˜ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ๊ธด ์ปจํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ LLM์˜ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๊ฐ€ ์‹œ๊ฐ„์ด ๋งŽ์ด ์†Œ์š”๋˜์–ด ์Šค์ผ€์ค„๋Ÿฌ์— ์˜ํ•ด ์ƒ์„ฑ์ด ์ค‘๋‹จ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋กœ ์ธํ•ด ํ˜„์žฌ ์š”์ฒญ์— ๋Œ€ํ•œ ์‘๋‹ต ์ƒ์„ฑ์ด ์™„๋ฃŒ๋˜์ง€ ์•Š์€ ๊ฒฝ์šฐ์—๋„ LLM์˜ ํ˜„์žฌ ์ƒ์„ฑ ๊ฒฐ๊ณผ๋ฅผ ์Šค๋ƒ…์ƒท์œผ๋กœ ์ €์žฅํ•˜์—ฌ ์ผ์‹œ ์ค‘์ง€/์žฌ๊ฐœ ๋™์ž‘์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ณ ์•ˆํ•ด์•ผ ํ•˜๋Š” ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์—์ด์ „ํŠธ๊ฐ€ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ†ตํ™” ๋„๊ตฌ ๋ชฉ๋ก์„ ํ™•๋ณดํ•œ ํ›„์—๋„ ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๊ฐ€ ๋™์ผํ•œ ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•ด์•ผ ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ด๋Ÿฌํ•œ ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•˜๊ธฐ ์œ„ํ•œ ์ตœ์ ์˜ ์ˆœ์„œ๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ๋„ ๋˜ ๋‹ค๋ฅธ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ, ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๊ฐ€ ๋™์‹œ์— ์ž‘๋™ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์„œ๋กœ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ๋ฅผ ์œ„ํ•œ ๊ฐ•๋ ฅํ•œ ์‹œ์Šคํ…œ์ด ํ•„์š”ํ•˜๋ฉฐ, ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ๋ฐ ์•ก์„ธ์Šค ์ œ์–ด ์กฐ์น˜๋ฅผ ์—„๊ฒฉํ•˜๊ฒŒ ์‹œํ–‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
์—์ด์ „ํŠธ ์ˆ˜๋Ÿ‰๊ณผ ๋ณต์žก์„ฑ์ด ๊ธฐํ•˜๊ธ‰์ˆ˜์ ์œผ๋กœ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ LLM๊ณผ OS์˜ ๊ธฐ๋Šฅ์— ๋Œ€ํ•œ ๋ถ€๋‹ด์ด ์ปค์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ œํ•œ๋œ LLM ๋ฆฌ์†Œ์Šค์—์„œ ์—์ด์ „ํŠธ ์š”์ฒญ์„ ์˜ˆ์•ฝํ•˜๊ณ  ์šฐ์„ ์ˆœ์œ„๋ฅผ ์ •ํ•˜๋Š” ๊ฒƒ์€ ์‹ฌ๊ฐํ•œ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์˜ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋Š” ๊ธด ์ปจํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ ์‹œ๊ฐ„์ด ๋งŽ์ด ์†Œ์š”๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋•Œ๋กœ๋Š” ์Šค์ผ€์ค„๋Ÿฌ์— ์˜ํ•ด ์ƒ์„ฑ์ด ์ผ์‹œ ์ค‘์ง€๋˜๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” LLM์ด ํ˜„์žฌ ์š”์ฒญ์— ๋Œ€ํ•œ ์‘๋‹ต ์ƒ์„ฑ์„ ์™„๋ฃŒํ•˜์ง€ ์•Š์€ ๊ฒฝ์šฐ์—๋„ ์ผ์‹œ ์ค‘์ง€/์žฌ๊ฐœ ๋™์ž‘์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ณ ์•ˆํ•ด์•ผ ํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์—์ด์ „ํŠธ๊ฐ€ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ˜ธ์ถœ ๋„๊ตฌ ๋ชฉ๋ก์„ ์–ป์€ ํ›„์—๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๊ฐ€ ๋™์ผํ•œ ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•ด์•ผ ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ด๋Ÿฌํ•œ ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•  ์ตœ์ ์˜ ์ˆœ์„œ๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ๋„ ๋˜ ๋‹ค๋ฅธ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๊ฐ€ ๋™์‹œ์— ์ž‘๋™ํ•˜๋ ค๋ฉด ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ๋ฅผ ์œ„ํ•œ ๊ฐ•๋ ฅํ•œ ์‹œ์Šคํ…œ์ด ํ•„์š”ํ•˜๋ฉฐ, ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ๋ฐ ์•ก์„ธ์Šค ์ œ์–ด ์กฐ์น˜๋ฅผ ์—„๊ฒฉํ•˜๊ฒŒ ์‹œํ–‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
์—์ด์ „ํŠธ์˜ ์–‘๊ณผ ๋ณต์žก์„ฑ์ด ๊ธฐํ•˜๊ธ‰์ˆ˜์ ์œผ๋กœ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ LLM๊ณผ OS์˜ ๊ธฐ๋Šฅ์— ๋Œ€ํ•œ ๋ถ€๋‹ด์ด ์ปค์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ œํ•œ๋œ LLM ๋ฆฌ์†Œ์Šค์—์„œ ์—์ด์ „ํŠธ ์š”์ฒญ์„ ์˜ˆ์•ฝํ•˜๊ณ  ์šฐ์„ ์ˆœ์œ„๋ฅผ ์ •ํ•˜๋Š” ๊ฒƒ์€ ์‹ฌ๊ฐํ•œ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ LLM์˜ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋Š” ๊ธด ์ปจํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•  ๋•Œ ์‹œ๊ฐ„์ด ๋งŽ์ด ์†Œ์š”๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋•Œ๋•Œ๋กœ ์Šค์ผ€์ค„๋Ÿฌ์— ์˜ํ•ด ์ƒ์„ฑ์ด ์ผ์‹œ ์ค‘๋‹จ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋กœ ์ธํ•ด LLM์˜ ํ˜„์žฌ ์ƒ์„ฑ ๊ฒฐ๊ณผ๋ฅผ ์Šค๋ƒ…์ƒทํ•˜๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ณ ์•ˆํ•˜์—ฌ ํ˜„์žฌ ์š”์ฒญ์— ๋Œ€ํ•œ ์‘๋‹ต ์ƒ์„ฑ์ด ์™„๋ฃŒ๋˜์ง€ ์•Š์€ ๊ฒฝ์šฐ์—๋„ ์ผ์‹œ ์ค‘์ง€/์žฌ๊ฐœ ๋™์ž‘์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์—์ด์ „ํŠธ๊ฐ€ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ†ตํ™” ๋„๊ตฌ ๋ชฉ๋ก์„ ํ™•๋ณดํ•œ ํ›„์—๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๊ฐ€ ๋™์ผํ•œ ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•ด์•ผ ํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ด๋Ÿฌํ•œ ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•  ์ตœ์ ์˜ ์ˆœ์„œ๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์ด ๋˜ ๋‹ค๋ฅธ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ๋™์‹œ ์ž‘์—…์„ ์œ„ํ•ด์„œ๋Š” ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ๊ฐ„์˜ ๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ๋ฅผ ์œ„ํ•œ ๊ฐ•๋ ฅํ•œ ์‹œ์Šคํ…œ์ด ํ•„์š”ํ•˜๋ฉฐ, ๊ฐœ์ธ์ •๋ณด ๋ณดํ˜ธ ๋ฐ ์•ก์„ธ์Šค ์ œ์–ด ์กฐ์น˜๋ฅผ ์—„๊ฒฉํ•˜๊ฒŒ ์‹œํ–‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
**500K length book summarization (BookSum).** We further scaled our approach by continuously pre-training a 8B LLM model with 8K input length for 30K steps. We then fine-tuned on a book summarization task, BookSum (Kryscinski et al., 2021) where the goal is to generate a summary of an entire book text.
**500K ๊ธธ์ด์˜ ์ฑ… ์š”์•ฝ(๋ถ์„ฌ).** 8K ์ž…๋ ฅ ๊ธธ์ด์˜ 8B LLM ๋ชจ๋ธ์„ 30K ๋‹จ๊ณ„ ๋™์•ˆ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต์‹œ์ผœ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋”์šฑ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ „์ฒด ์ฑ… ํ…์ŠคํŠธ์˜ ์š”์•ฝ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ธ ์ฑ… ์š”์•ฝ ์ž‘์—…์ธ BookSum(Kryscinski et al., 2021)์—์„œ ๋ฏธ์„ธ ์กฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
**500K ๊ธธ์ด์˜ ์ฑ… ์š”์•ฝ(๋ถ์„ฌ).** 8K ์ž…๋ ฅ ๊ธธ์ด์˜ 8B LLM ๋ชจ๋ธ์„ 30K ๋‹จ๊ณ„ ๋™์•ˆ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต์‹œ์ผœ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋”์šฑ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ „์ฒด ์ฑ… ํ…์ŠคํŠธ์˜ ์š”์•ฝ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ธ ์ฑ… ์š”์•ฝ ์ž‘์—…์ธ BookSum(Kryscinski et al., 2021)์—์„œ ๋ฏธ์„ธ ์กฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
**500K ๊ธธ์ด์˜ ์ฑ… ์š”์•ฝ(BookSum).** 8K ์ž…๋ ฅ ๊ธธ์ด์˜ 8B LLM ๋ชจ๋ธ์„ 30K ๋‹จ๊ณ„์— ๊ฑธ์ณ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต์‹œ์ผœ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋”์šฑ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ „์ฒด ์ฑ… ํ…์ŠคํŠธ์˜ ์š”์•ฝ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ธ ์ฑ… ์š”์•ฝ ์ž‘์—…์ธ BookSum(Kryscinski et al., 2021)์—์„œ ๋ฏธ์„ธ ์กฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
**500K ๊ธธ์ด์˜ ์ฑ… ์š”์•ฝ(BookSum).** 30K ๋‹จ๊ณ„์— ๋Œ€ํ•ด 8K ์ž…๋ ฅ ๊ธธ์ด์˜ 8B LLM ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜์—ฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ „์ฒด ์ฑ… ํ…์ŠคํŠธ์˜ ์š”์•ฝ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ธ ์ฑ… ์š”์•ฝ ๊ณผ์ œ์ธ BookSum(Kryscinski ๋“ฑ, 2021)์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
**50๋งŒ ๋‹จ์–ด ๊ธธ์ด์˜ ์ฑ… ์š”์•ฝ(BookSum).** 30K ๋‹จ๊ณ„์— ๊ฑธ์ณ 8K ์ž…๋ ฅ ๊ธธ์ด์˜ 8B LLM ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜์—ฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ „์ฒด ์ฑ… ํ…์ŠคํŠธ์˜ ์š”์•ฝ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ธ ์ฑ… ์š”์•ฝ ์ž‘์—…์ธ BookSum(Kryscinski et al., 2021)์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
The story then matches back to Jurning childhood, producing context about her children and the relationship with my.
๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด์•ผ๊ธฐ๋Š” ์ €๋ง์˜ ์–ด๋ฆฐ ์‹œ์ ˆ๊ณผ ์ผ์น˜ํ•˜์—ฌ ๊ทธ๋…€์˜ ์ž๋…€์™€ ๋‚˜์™€์˜ ๊ด€๊ณ„์— ๋Œ€ํ•œ ๋งฅ๋ฝ์„ ๋งŒ๋“ค์–ด๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด์•ผ๊ธฐ๋Š” ์ €๋ง์˜ ์–ด๋ฆฐ ์‹œ์ ˆ๊ณผ ์ผ์น˜ํ•˜์—ฌ ๊ทธ๋…€์˜ ์ž๋…€์™€ ๋‚˜์™€์˜ ๊ด€๊ณ„์— ๋Œ€ํ•œ ๋งฅ๋ฝ์„ ๋งŒ๋“ค์–ด๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋ฆฌ๊ณ  ๋‚˜์„œ ์ด์•ผ๊ธฐ๋Š” ์ €๋‹์˜ ์–ด๋ฆฐ ์‹œ์ ˆ๋กœ ๋Œ์•„๊ฐ€ ๊ทธ๋…€์˜ ์ž๋…€์™€ ๋‚˜์™€์˜ ๊ด€๊ณ„์— ๋Œ€ํ•œ ๋งฅ๋ฝ์„ ๋งŒ๋“ค์–ด๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด์•ผ๊ธฐ๋Š” ์ €๋‹์˜ ์–ด๋ฆฐ ์‹œ์ ˆ๊ณผ ์ผ์น˜ํ•˜์—ฌ ์ž๋…€์™€ ๋‚ด ๊ด€๊ณ„์— ๋Œ€ํ•œ ๋งฅ๋ฝ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๊ทธ๋Ÿฐ ๋‹ค์Œ ์Šคํ† ๋ฆฌ๊ฐ€ ์ €๋‹์˜ ์–ด๋ฆฐ ์‹œ์ ˆ๊ณผ ๋‹ค์‹œ ์—ฐ๊ฒฐ๋˜์–ด ์ž๋…€์™€ ๋‚ด ๊ด€๊ณ„์— ๋Œ€ํ•œ ์ปจํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
BERT (Birectional Encoder Representations from Transformers) [24] is one of the most widely used encoder-only language models. BERT consists of three modules: (1) an embedding module that converts input text into a sequence of embedding vectors, (2) a stack of Transformer encoders that converts embedding vectors into contextual representation vectors, and (3) a fully connected layer that converts the representation vectors (at the final layer) to one-hot vectors. BERT is pre-trained uses two objectives: masked language modeling (MLM) and next sentence prediction. The pre-trained BERT model can be fine-tuned by adding a classifier layer for many language understanding tasks, ranging from text
BERT(์–‘๋ฐฉํ–ฅ ์ธ์ฝ”๋” ํ‘œํ˜„์—์„œ ํŠธ๋žœ์Šคํฌ๋จธ) [24]๋Š” ๊ฐ€์žฅ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์ธ์ฝ”๋” ์ „์šฉ ์–ธ์–ด ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. BERT๋Š” ์„ธ ๊ฐ€์ง€ ๋ชจ๋“ˆ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: (1) ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ ์‹œํ€€์Šค๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์ž„๋ฒ ๋”ฉ ๋ชจ๋“ˆ, (2) ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ๋ฅผ ์ปจํ…์ŠคํŠธ ํ‘œํ˜„ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ์ธ์ฝ”๋” ์Šคํƒ, (3) ํ‘œํ˜„ ๋ฒกํ„ฐ(์ตœ์ข… ๋ ˆ์ด์–ด)๋ฅผ ์›ํ•ซ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์™„์ „ํžˆ ์—ฐ๊ฒฐ๋œ ๋ ˆ์ด์–ด์ž…๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต๋œ BERT ๋ชจ๋ธ์€ ๋‹ค์Œ ๋‘ ๊ฐ€์ง€ ๋ชฉํ‘œ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค: ๋งˆ์Šคํ‚น ์–ธ์–ด ๋ชจ๋ธ๋ง(MLM)๊ณผ ๋‹ค์Œ ๋ฌธ์žฅ ์˜ˆ์ธก. ์‚ฌ์ „ ํ•™์Šต๋œ BERT ๋ชจ๋ธ์€ ํ…์ŠคํŠธ
BERT(์–‘๋ฐฉํ–ฅ ์ธ์ฝ”๋” ํ‘œํ˜„, Transformer์—์„œ ๋ฐฉํ–ฅ์„ฑ ์ธ์ฝ”๋” ํ‘œํ˜„)๋Š” ๊ฐ€์žฅ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์ธ์ฝ”๋” ์ „์šฉ ์–ธ์–ด ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. BERT๋Š” (1) ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ ์‹œํ€€์Šค๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์ž„๋ฒ ๋”ฉ ๋ชจ๋“ˆ, (2) ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ๋ฅผ ์ปจํ…์ŠคํŠธ ํ‘œํ˜„ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” Transformer ์ธ์ฝ”๋” ์Šคํƒ, (3) ํ‘œํ˜„ ๋ฒกํ„ฐ(์ตœ์ข… ๋ ˆ์ด์–ด)๋ฅผ ์›ํ•ซ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์™„์ „ํžˆ ์—ฐ๊ฒฐ๋œ ๋ ˆ์ด์–ด์˜ ์„ธ ๊ฐ€์ง€ ๋ชจ๋“ˆ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. BERT๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๋‘ ๊ฐ€์ง€ ๋ชฉํ‘œ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค: ๋งˆ์Šคํ‚น ์–ธ์–ด ๋ชจ๋ธ๋ง(MLM)๊ณผ ๋‹ค์Œ ๋ฌธ์žฅ ์˜ˆ์ธก. ์‚ฌ์ „ ํ•™์Šต๋œ BERT ๋ชจ๋ธ์€ ํ…์ŠคํŠธ
BERT(์–‘๋ฐฉํ–ฅ ์ธ์ฝ”๋” ํ‘œํ˜„-ํŠธ๋žœ์Šคํฌ๋จธ์—์„œ) [24]๋Š” ๊ฐ€์žฅ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์ธ์ฝ”๋” ์ „์šฉ ์–ธ์–ด ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. BERT๋Š” ์„ธ ๊ฐ€์ง€ ๋ชจ๋“ˆ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: (1) ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ ์‹œํ€€์Šค๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์ž„๋ฒ ๋”ฉ ๋ชจ๋“ˆ, (2) ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ๋ฅผ ์ปจํ…์ŠคํŠธ ํ‘œํ˜„ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ์ธ์ฝ”๋” ์Šคํƒ, (3) ํ‘œํ˜„ ๋ฒกํ„ฐ(์ตœ์ข… ๋ ˆ์ด์–ด)๋ฅผ ์›ํ•ซ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์™„์ „ ์—ฐ๊ฒฐ ๋ ˆ์ด์–ด. BERT๋Š” ๋งˆ์Šคํ‚น ์–ธ์–ด ๋ชจ๋ธ๋ง(MLM)๊ณผ ๋‹ค์Œ ๋ฌธ์žฅ ์˜ˆ์ธก์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ๋ชฉํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต๋œ BERT ๋ชจ๋ธ์€ ํ…์ŠคํŠธ
BERT(์–‘๋ฐฉํ–ฅ ์ธ์ฝ”๋” ํ‘œํ˜„) [24]๋Š” ๊ฐ€์žฅ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์ธ์ฝ”๋” ์ „์šฉ ์–ธ์–ด ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. BERT๋Š” ์„ธ ๊ฐ€์ง€ ๋ชจ๋“ˆ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: (1) ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ์ผ๋ จ์˜ ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์ž„๋ฒ ๋”ฉ ๋ชจ๋“ˆ, (2) ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ๋ฅผ ์ปจํ…์ŠคํŠธ ํ‘œํ˜„ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ์ธ์ฝ”๋” ์Šคํƒ, (3) ํ‘œํ˜„ ๋ฒกํ„ฐ(๋งˆ์ง€๋ง‰ ๋ ˆ์ด์–ด์—์„œ)๋ฅผ ์›ํ•ซ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์™„์ „ ์—ฐ๊ฒฐ ๋ ˆ์ด์–ด. BERT๋Š” ๋‘ ๊ฐ€์ง€ ๋ชฉํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค: ๋งˆ์Šคํ‚น๋œ ์–ธ์–ด ๋ชจ๋ธ๋ง(MLM)๊ณผ ๋‹ค์Œ ๋ฌธ์žฅ ์˜ˆ์ธก์ž…๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต๋œ BERT ๋ชจ๋ธ์€ ํ…์ŠคํŠธ์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋งŽ์€ ์–ธ์–ด ์ดํ•ด ์ž‘์—…์„ ์œ„ํ•ด ๋ถ„๋ฅ˜๊ธฐ ๋ ˆ์ด์–ด๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ ๋ฏธ์„ธ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
BERT(์–‘๋ฐฉํ–ฅ ์ธ์ฝ”๋” ํ‘œํ˜„ ํŠธ๋žœ์Šคํฌ๋จธ์—์„œ) [24]๋Š” ๊ฐ€์žฅ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์ธ์ฝ”๋” ์ „์šฉ ์–ธ์–ด ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. BERT๋Š” ์„ธ ๊ฐ€์ง€ ๋ชจ๋“ˆ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค: (1) ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ์ผ๋ จ์˜ ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์ž„๋ฒ ๋”ฉ ๋ชจ๋“ˆ, (2) ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ๋ฅผ ์ปจํ…์ŠคํŠธ ํ‘œํ˜„ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ํŠธ๋žœ์Šคํฌ๋จธ ์ธ์ฝ”๋” ์Šคํƒ, (3) ํ‘œํ˜„ ๋ฒกํ„ฐ(์ตœ์ข… ๋ ˆ์ด์–ด์—์„œ)๋ฅผ ์›ํ•ซ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์™„์ „ ์—ฐ๊ฒฐ ๋ ˆ์ด์–ด. BERT๋Š” ๋‘ ๊ฐ€์ง€ ๋ชฉํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ์ „ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค: ๋งˆ์Šคํฌ ์–ธ์–ด ๋ชจ๋ธ๋ง(MLM)๊ณผ ๋‹ค์Œ ๋ฌธ์žฅ ์˜ˆ์ธก์ž…๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต๋œ BERT ๋ชจ๋ธ์€ ํ…์ŠคํŠธ์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์–ธ์–ด ์ดํ•ด ์ž‘์—…์„ ์œ„ํ•ด ๋ถ„๋ฅ˜๊ธฐ ๋ ˆ์ด์–ด๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ ๋ฏธ์„ธ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Footnote 1: Note that a LLM is not necessarily more capable than a small PLM, and emergent abilities may not occur in some LLMs.
1: LLM์ด ๋ฐ˜๋“œ์‹œ ์ž‘์€ PLM๋ณด๋‹ค ๋” ์œ ๋Šฅํ•œ ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ผ๋ถ€ LLM์—์„œ๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์ด ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
1: LLM์ด ๋ฐ˜๋“œ์‹œ ์ž‘์€ PLM๋ณด๋‹ค ๋” ์œ ๋Šฅํ•œ ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ผ๋ถ€ LLM์—์„œ๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์ด ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
1: LLM์ด ๋ฐ˜๋“œ์‹œ ์ž‘์€ PLM๋ณด๋‹ค ๋” ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ผ๋ถ€ LLM์—์„œ๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์ด ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์„ ์ˆ˜๋„ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”.
๊ฐ์ฃผ 1: LLM์ด ๋ฐ˜๋“œ์‹œ ์†Œํ˜• PLM๋ณด๋‹ค ๋” ์œ ๋Šฅํ•œ ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ผ๋ถ€ LLM์—์„œ๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์ด ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์„ ์ˆ˜๋„ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”.
๊ฐ์ฃผ 1: LLM์ด ๋ฐ˜๋“œ์‹œ ์†Œํ˜• PLM๋ณด๋‹ค ๋” ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ผ๋ถ€ LLM์—์„œ๋Š” ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์ด ๋‚˜ํƒ€๋‚˜์ง€ ์•Š์„ ์ˆ˜๋„ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”.
**Models.** We train 1B, 3B, and 7B parameters autoregressive decoder-only models, based on configurations and hyperparameters similar to GPT-3 (Brown et al., 2020), diverging mostly on our use of ALiBi (Press et al., 2021). We use FlashAttention (Dao et al., 2022) in a custom codebase. We train internal models on both The Pile and RefinedWeb to control for deviations caused by our pretraining setup-we found The Pile models to perform in-line with others. For small-scale and ablation studies (first half of Section 4.2; Section 4.3), we train models to optimality according to the scaling laws of Hoffmann et al. (2022): on 27B and 60B tokens respectively for our 1B and 3B parameters models. For the main experiments demonstrating our approach (Falcon-RW models in Section 4.2), we train the models to 350GT, in line with popular public models (Brown et al., 2020; Wang and Komatuszaki, 2021; Scao et al., 2022). Note that we do not compare against the recently introduced LLaMA models (Touvron et al., 2023), as the smallest of them is trained on x2.5 more compute than our largest model, preventing a meaningful comparison from being made dataset-wise. For a more in-depth overview of the models and pretraining datasets with which we compare, see Appendix F.
**๋ชจ๋ธ.** ์šฐ๋ฆฌ๋Š” GPT-3(Brown et al., 2020)์™€ ์œ ์‚ฌํ•œ ๊ตฌ์„ฑ๊ณผ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 1B, 3B, 7B ํŒŒ๋ผ๋ฏธํ„ฐ ์ž๋™ ํšŒ๊ท€ ๋””์ฝ”๋” ์ „์šฉ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ , ๋Œ€๋ถ€๋ถ„ ALiBi(Press et al., 2021)๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค๋Š” ์ ์—์„œ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ปค์Šคํ…€ ์ฝ”๋“œ๋ฒ ์ด์Šค์—์„œ FlashAttention(Dao et al., 2022)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์œผ๋กœ ์ธํ•œ ํŽธ์ฐจ๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด ๋‚ด๋ถ€ ๋ชจ๋ธ์„ The Pile๊ณผ RefinedWeb ๋ชจ๋‘์—์„œ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, The Pile ๋ชจ๋ธ์ด ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ์ผ์น˜ํ•˜๋Š” ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ ๋ฐ ์ ˆ์ œ ์—ฐ๊ตฌ(์„น์…˜ 4.2์˜ ์ „๋ฐ˜๋ถ€; ์„น์…˜ 4.3)์˜ ๊ฒฝ์šฐ, ์šฐ๋ฆฌ๋Š” Hoffmann ์™ธ(2022)์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋”ฐ๋ผ ๋ชจ๋ธ์„ ์ตœ์ ํ™”ํ•˜๋„๋ก ํ›ˆ๋ จํ–ˆ์Šต๋‹ˆ๋‹ค: 1B ๋ฐ 3B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๊ฐ๊ฐ 27B ๋ฐ 60B ํ† ํฐ์— ๋Œ€ํ•ด. ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณด์—ฌ์ฃผ๋Š” ์ฃผ์š” ์‹คํ—˜(์„น์…˜ 4.2์˜ ํŒ”์ฝ˜-RW ๋ชจ๋ธ)์— ๋Œ€ํ•ด์„œ๋Š” ๋Œ€์ค‘์ ์ธ ๊ณต๊ฐœ ๋ชจ๋ธ(Brown et al., 2020; Wang ๋ฐ Komatuszaki, 2021; Scao et al., 2022)๊ณผ ์ผ์น˜ํ•˜๋„๋ก ๋ชจ๋ธ์„ 350GT๋กœ ํ›ˆ๋ จํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋„์ž…๋œ LLaMA ๋ชจ๋ธ(Touvron et al., 2023)์— ๋Œ€ํ•ด์„œ๋Š” ๊ฐ€์žฅ ์ž‘์€ ๋ชจ๋ธ์ด ์šฐ๋ฆฌ์˜ ๊ฐ€์žฅ ํฐ ๋ชจ๋ธ๋ณด๋‹ค x2.5 ๋” ๋งŽ์€ ์ปดํ“จํŒ…์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ธก๋ฉด์—์„œ ์˜๋ฏธ ์žˆ๋Š” ๋น„๊ต๋ฅผ ํ•  ์ˆ˜ ์—†์œผ๋ฏ€๋กœ ๋น„๊ตํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•˜๋Š” ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๊ฐœ์š”๋Š” ๋ถ€๋ก F๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
**๋ชจ๋ธ.** GPT-3(Brown ์™ธ, 2020)์— ์œ ์‚ฌํ•œ ๊ตฌ์„ฑ ๋ฐ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 1B, 3B, 7B ํŒŒ๋ผ๋ฏธํ„ฐ ์ž๋™ ํšŒ๊ท€ ๋””์ฝ”๋” ์ „์šฉ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋ฉฐ, ๋Œ€๋ถ€๋ถ„ ALiBi(Press ์™ธ, 2021)๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค๋Š” ์ ์—์„œ ์ฐจ์ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ปค์Šคํ…€ ์ฝ”๋“œ๋ฒ ์ด์Šค์—์„œ FlashAttention(Dao ์™ธ, 2022)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์œผ๋กœ ์ธํ•œ ํŽธ์ฐจ๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด ๋‚ด๋ถ€ ๋ชจ๋ธ์„ ๋” ํŒŒ์ผ๊ณผ ์ •์ œ ์›น์—์„œ ๋ชจ๋‘ ํ•™์Šต์‹œ์ผฐ๋Š”๋ฐ, ๋” ํŒŒ์ผ ๋ชจ๋ธ์ด ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ ๋ฐ ์ ˆ์ œ ์—ฐ๊ตฌ(์„น์…˜ 4.2์˜ ์ „๋ฐ˜๋ถ€; ์„น์…˜ 4.3)์˜ ๊ฒฝ์šฐ, 1B ๋ฐ 3B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐ๊ฐ 27B ๋ฐ 60B ํ† ํฐ์— ๋Œ€ํ•ด ํ˜ธํ”„๋งŒ ์™ธ(2022)์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋”ฐ๋ผ ์ตœ์ ํ™”๋œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณด์—ฌ์ฃผ๋Š” ์ฃผ์š” ์‹คํ—˜(์„น์…˜ 4.2์˜ ํŒ”์ฝ˜-RW ๋ชจ๋ธ)์— ๋Œ€ํ•ด์„œ๋Š” 350GT๊นŒ์ง€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์—ฌ ์ธ๊ธฐ ์žˆ๋Š” ๊ณต๊ฐœ ๋ชจ๋ธ(Brown ์™ธ, 2020; Wang ๋ฐ Komatuszaki, 2021; Scao ์™ธ, 2022)๊ณผ ์ผ์น˜ํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋„์ž…๋œ LLaMA ๋ชจ๋ธ(Touvron ์™ธ, 2023)์— ๋Œ€ํ•ด์„œ๋Š” ๊ฐ€์žฅ ์ž‘์€ ๋ชจ๋ธ์ด ์šฐ๋ฆฌ์˜ ๊ฐ€์žฅ ํฐ ๋ชจ๋ธ๋ณด๋‹ค x2.5 ๋” ๋งŽ์€ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค๋กœ ํ•™์Šต๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ธก๋ฉด์—์„œ ์˜๋ฏธ ์žˆ๋Š” ๋น„๊ต๋ฅผ ํ•  ์ˆ˜ ์—†์œผ๋ฏ€๋กœ ๋น„๊ตํ•˜์ง€ ์•Š๋Š”๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”. ๋ชจ๋ธ๊ณผ ๋น„๊ต ๋Œ€์ƒ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๊ฐœ์š”๋Š” ๋ถ€๋ก F๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
**๋ชจ๋ธ.** GPT-3(Brown et al., 2020)์™€ ์œ ์‚ฌํ•œ ๊ตฌ์„ฑ๊ณผ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 1B, 3B, 7B ํŒŒ๋ผ๋ฏธํ„ฐ ์ž๋™ ํšŒ๊ท€ ๋””์ฝ”๋” ์ „์šฉ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ , ALiBi(Press et al., 2021)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๊ณผ ๋‹ค๋ฅธ ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋งž์ถคํ˜• ์ฝ”๋“œ๋ฒ ์ด์Šค์—์„œ FlashAttention(Dao et al., 2022)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์œผ๋กœ ์ธํ•œ ํŽธ์ฐจ๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด The Pile๊ณผ RefinedWeb ๋ชจ๋‘์—์„œ ๋‚ด๋ถ€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, The Pile ๋ชจ๋ธ์ด ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ ๋ฐ ์ ˆ์ œ ์—ฐ๊ตฌ(4.2 ์„น์…˜์˜ ์ „๋ฐ˜๋ถ€; 4.3 ์„น์…˜)์˜ ๊ฒฝ์šฐ, Hoffmann ์™ธ(2022)์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋”ฐ๋ผ ์ตœ์ ์˜ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค: 1B ๋ฐ 3B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๊ฐ๊ฐ 27B ๋ฐ 60B ํ† ํฐ์— ๋Œ€ํ•ด ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋ณด์—ฌ์ฃผ๋Š” ์ฃผ์š” ์‹คํ—˜(์„น์…˜ 4.2์˜ Falcon-RW ๋ชจ๋ธ)์— ๋Œ€ํ•ด์„œ๋Š” ๋Œ€์ค‘์ ์ธ ๊ณต๊ฐœ ๋ชจ๋ธ(Brown et al., 2020; Wang ๋ฐ Komatuszaki, 2021; Scao et al., 2022)์— ๋”ฐ๋ผ ๋ชจ๋ธ์„ 350GT๊นŒ์ง€ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์— ๋„์ž…๋œ LLaMA ๋ชจ๋ธ(Touvron et al., 2023)์€ ๊ฐ€์žฅ ์ž‘์€ ๋ชจ๋ธ์ด ์šฐ๋ฆฌ์˜ ๊ฐ€์žฅ ํฐ ๋ชจ๋ธ๋ณด๋‹ค x2.5 ๋” ๋งŽ์€ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค๋กœ ํ›ˆ๋ จ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ณ„๋กœ ์˜๋ฏธ ์žˆ๋Š” ๋น„๊ต๋ฅผ ํ•  ์ˆ˜ ์—†์œผ๋ฏ€๋กœ ๋น„๊ตํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ๊ณผ ๋น„๊ต ๋Œ€์ƒ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๊ฐœ์š”๋Š” ๋ถ€๋ก F๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
**๋ชจ๋ธ.** GPT-3(Brown et al., 2020)๊ณผ ์œ ์‚ฌํ•œ ๊ตฌ์„ฑ ๋ฐ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 1B, 3B, 7B ํŒŒ๋ผ๋ฏธํ„ฐ ์ž๋™ ํšŒ๊ท€ ๋””์ฝ”๋” ์ „์šฉ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋ฉฐ, ์ฃผ๋กœ ALiBi(Press et al., 2021)์˜ ์‚ฌ์šฉ์—์„œ ์ฐจ์ด๊ฐ€ ๋‚ฉ๋‹ˆ๋‹ค. ๋งž์ถคํ˜• ์ฝ”๋“œ๋ฒ ์ด์Šค์—์„œ FlashAttention(Dao et al., 2022)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์œผ๋กœ ์ธํ•œ ํŽธ์ฐจ๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด ๋‚ด๋ถ€ ๋ชจ๋ธ์„ The Pile๊ณผ RefinedWeb ๋ชจ๋‘์—์„œ ํ›ˆ๋ จํ•˜๋ฉฐ, The Pile ๋ชจ๋ธ์ด ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ธ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ ๋ฐ ์ ˆ์ œ ์—ฐ๊ตฌ(์„น์…˜ 4.2์˜ ์ „๋ฐ˜๋ถ€, ์„น์…˜ 4.3)์˜ ๊ฒฝ์šฐ, 1B ๋ฐ 3B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๊ฐ๊ฐ 27B ๋ฐ 60B ํ† ํฐ์— ๋Œ€ํ•ด ํ˜ธํ”„๋งŒ ๋“ฑ์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™(Hoffmann et al., 2022)์— ๋”ฐ๋ผ ๋ชจ๋ธ์„ ์ตœ์ ์œผ๋กœ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ž…์ฆํ•˜๋Š” ์ฃผ์š” ์‹คํ—˜(์„น์…˜ 4.2์˜ Falcon-RW ๋ชจ๋ธ)์˜ ๊ฒฝ์šฐ, ๋Œ€์ค‘์ ์ธ ๊ณต๊ฐœ ๋ชจ๋ธ(Brown et al., 2020; Wang and Komatuszaki, 2021; Scao et al., 2022)์— ๋งž์ถฐ 350GT๊นŒ์ง€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์— ์†Œ๊ฐœ๋œ LLaMA ๋ชจ๋ธ(Touvron et al., 2023)๊ณผ ๋น„๊ตํ•˜์ง€ ์•Š๋Š” ์ด์œ ๋Š” ๊ฐ€์žฅ ์ž‘์€ ๋ชจ๋ธ์ด ๊ฐ€์žฅ ํฐ ๋ชจ๋ธ๋ณด๋‹ค 2.5๋ฐฐ ๋” ๋งŽ์€ ์ปดํ“จํŒ… ๋ฆฌ์†Œ์Šค์—์„œ ํ•™์Šต๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ธก๋ฉด์—์„œ ์˜๋ฏธ ์žˆ๋Š” ๋น„๊ต๋ฅผ ํ•  ์ˆ˜ ์—†๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋น„๊ต ๋Œ€์ƒ์ธ ๋ชจ๋ธ๊ณผ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๊ฐœ์š”๋Š” ๋ถ€๋ก F๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
**๋ชจ๋ธ. **GPT-3(Brown et al., 2020)๊ณผ ์œ ์‚ฌํ•œ ๊ตฌ์„ฑ ๋ฐ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 1B, 3B, 7B ํŒŒ๋ผ๋ฏธํ„ฐ ์ž๋™ ํšŒ๊ท€ ๋””์ฝ”๋” ์ „์šฉ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋ฉฐ, ์ฃผ๋กœ ALiBi(Press et al., 2021)์˜ ์‚ฌ์šฉ์—์„œ ์ฐจ์ด๊ฐ€ ๋‚ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž ์ •์˜ ์ฝ”๋“œ๋ฒ ์ด์Šค์—์„œ FlashAttention(Dao et al., 2022)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์œผ๋กœ ์ธํ•œ ํŽธ์ฐจ๋ฅผ ์ œ์–ดํ•˜๊ธฐ ์œ„ํ•ด ๋‚ด๋ถ€ ๋ชจ๋ธ์„ The Pile๊ณผ RefinedWeb ๋ชจ๋‘์—์„œ ํ›ˆ๋ จํ–ˆ์œผ๋ฉฐ, ๊ทธ ๊ฒฐ๊ณผ The Pile ๋ชจ๋ธ์ด ๋‹ค๋ฅธ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ธ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์†Œ๊ทœ๋ชจ ๋ฐ ์ ˆ์ œ ์—ฐ๊ตฌ(์„น์…˜ 4.2์˜ ์ „๋ฐ˜๋ถ€, ์„น์…˜ 4.3)์˜ ๊ฒฝ์šฐ, ํ˜ธํ”„๋งŒ ์™ธ(2022)์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋”ฐ๋ผ 1B ๋ฐ 3B ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐ๊ฐ 27B ๋ฐ 60B ํ† ํฐ์œผ๋กœ ๋ชจ๋ธ์„ ์ตœ์ ์œผ๋กœ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ž…์ฆํ•˜๋Š” ์ฃผ์š” ์‹คํ—˜(์„น์…˜ 4.2์˜ Falcon-RW ๋ชจ๋ธ)์˜ ๊ฒฝ์šฐ, ๋Œ€์ค‘์ ์ธ ๊ณต๊ฐœ ๋ชจ๋ธ(Brown et al., 2020; Wang and Komatuszaki, 2021; Scao et al., 2022)์— ๋งž์ถฐ 350GT๊นŒ์ง€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ตœ๊ทผ์— ๋„์ž…๋œ LLaMA ๋ชจ๋ธ(Touvron et al., 2023)์€ ๊ฐ€์žฅ ์ž‘์€ ๋ชจ๋ธ์ด ๊ฐ€์žฅ ํฐ ๋ชจ๋ธ๋ณด๋‹ค 2.5๋ฐฐ ๋” ๋งŽ์€ ์ปดํ“จํŒ…์œผ๋กœ ํ•™์Šต๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ธก๋ฉด์—์„œ ์˜๋ฏธ ์žˆ๋Š” ๋น„๊ต๊ฐ€ ๋ถˆ๊ฐ€๋Šฅํ•˜๋ฏ€๋กœ ๋น„๊ตํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ๊ณผ ๋น„๊ต ๋Œ€์ƒ์ธ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๊ฐœ์š”๋Š” ๋ถ€๋ก F๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
**Lemma 2**.: _Two reward functions from the same equivalence class induce the same optimal policy under the constrained RL problem._
**๋ณด์กฐ์ •๋ฆฌ 2**. _๋™์ผํ•œ ๋“ฑ๊ฐ€ ํด๋ž˜์Šค์˜ ๋‘ ๋ณด์ƒ ํ•จ์ˆ˜๋Š” ์ œ์•ฝ RL ๋ฌธ์ œ์—์„œ ๋™์ผํ•œ ์ตœ์  ์ •์ฑ…์„ ์œ ๋„ํ•œ๋‹ค._
**๋ ˆ๋งˆ 2**. _๋™์ผํ•œ ๋“ฑ๊ฐ€ ํด๋ž˜์Šค์˜ ๋‘ ๋ณด์ƒ ํ•จ์ˆ˜๋Š” ์ œ์•ฝ๋œ RL ๋ฌธ์ œ์—์„œ ๋™์ผํ•œ ์ตœ์  ์ •์ฑ…์„ ์œ ๋„ํ•œ๋‹ค._
**๋ ˆ๋งˆ 2**._ ๋™์ผํ•œ ๋“ฑ๊ฐ€ ํด๋ž˜์Šค์˜ ๋‘ ๋ณด์ƒ ํ•จ์ˆ˜๋Š” ์ œ์•ฝ๋œ RL ๋ฌธ์ œ์—์„œ ๋™์ผํ•œ ์ตœ์  ์ •์ฑ…์„ ์œ ๋„ํ•œ๋‹ค._
**๋ณด์กฐ ์ •๋ฆฌ 2**: _๋™์ผํ•œ ๋“ฑ๊ฐ€ ํด๋ž˜์Šค์—์„œ ๋‚˜์˜จ ๋‘ ๊ฐ€์ง€ ๋ณด์ƒ ํ•จ์ˆ˜๋Š” ์ œ์•ฝ๋œ RL ๋ฌธ์ œ์—์„œ ๋™์ผํ•œ ์ตœ์  ์ •์ฑ…์„ ์œ ๋„ํ•œ๋‹ค._
**๋ณด์กฐ ์ •๋ฆฌ 2**: _๋™์ผํ•œ ๋™๋“ฑ์„ฑ ํด๋ž˜์Šค์—์„œ ๋‘ ๊ฐ€์ง€ ๋ณด์ƒ ํ•จ์ˆ˜๋Š” ์ œ์•ฝ๋œ RL ๋ฌธ์ œ์—์„œ ๋™์ผํ•œ ์ตœ์  ์ •์ฑ…์„ ์œ ๋„ํ•œ๋‹ค._
**RMSNorm** Same to Llama2 model (Touvron et al., 2023) 7B serious. We normalize the input of each transformer sub-layer, the attention layer, and the feedforward layer, with RMSNorm (Zhang and Sennrich, 2019).
**RMSNorm** Llama2 ๋ชจ๋ธ(Touvron et al., 2023)๊ณผ ๋™์ผํ•œ 7B ์‹ฌ๊ฐ๋„. ๊ฐ ํŠธ๋žœ์Šคํฌ๋จธ ํ•˜์œ„ ๊ณ„์ธต, ์ฃผ์˜ ๊ณ„์ธต ๋ฐ ํ”ผ๋“œํฌ์›Œ๋“œ ๊ณ„์ธต์˜ ์ž…๋ ฅ์„ RMSNorm(Zhang ๋ฐ Sennrich, 2019)์œผ๋กœ ์ •๊ทœํ™”ํ•ฉ๋‹ˆ๋‹ค.
**RMSNorm** Llama2 ๋ชจ๋ธ(Tovron ์™ธ, 2023)๊ณผ ๋™์ผํ•œ 7B ์‹ฌ๊ฐ๋„. ๊ฐ ํŠธ๋žœ์Šคํฌ๋จธ ํ•˜์œ„ ๋ ˆ์ด์–ด, ์ฃผ์˜ ๋ ˆ์ด์–ด, ํ”ผ๋“œํฌ์›Œ๋“œ ๋ ˆ์ด์–ด์˜ ์ž…๋ ฅ์„ RMSNorm(Zhang ๋ฐ Sennrich, 2019)์œผ๋กœ ์ •๊ทœํ™”ํ•ฉ๋‹ˆ๋‹ค.
**RMSNorm** Llama2 ๋ชจ๋ธ(ํˆฌ๋ฅด๋ณธ ๋“ฑ, 2023)๊ณผ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. 7B ์‹ฌ๊ฐ. ๊ฐ ํŠธ๋žœ์Šคํฌ๋จธ ํ•˜์œ„ ๋ ˆ์ด์–ด, ์ฃผ์˜ ๋ ˆ์ด์–ด, ํ”ผ๋“œํฌ์›Œ๋“œ ๋ ˆ์ด์–ด์˜ ์ž…๋ ฅ์„ RMSNorm(์žฅ๊ณผ ์„ธ๋‹ˆ์น˜, 2019)์œผ๋กœ ์ •๊ทœํ™”ํ•ฉ๋‹ˆ๋‹ค.
**RMSNorm** ๋ผ๋งˆ2 ๋ชจ๋ธ(Touvron ์™ธ, 2023) 7B ์‹ฌ๊ฐ. ๊ฐ ํŠธ๋žœ์Šคํฌ๋จธ ํ•˜์œ„ ๊ณ„์ธต, ์ฃผ์˜ ๊ณ„์ธต ๋ฐ ํ”ผ๋“œํฌ์›Œ๋“œ ๊ณ„์ธต์˜ ์ž…๋ ฅ์„ RMSNorm(Zhang and Sennrich, 2019)์œผ๋กœ ์ •๊ทœํ™”ํ•ฉ๋‹ˆ๋‹ค.
**RMSNorm** ๋ผ๋งˆ2 ๋ชจ๋ธ(Touvron ์™ธ, 2023) 7B ์‹ฌ๊ฐ. ๊ฐ ํŠธ๋žœ์Šคํฌ๋จธ ํ•˜์œ„ ๊ณ„์ธต, ์ฃผ์˜ ๊ณ„์ธต ๋ฐ ํ”ผ๋“œํฌ์›Œ๋“œ ๊ณ„์ธต์˜ ์ž…๋ ฅ์„ RMSNorm(Zhang and Sennrich, 2019)์œผ๋กœ ์ •๊ทœํ™”ํ•ฉ๋‹ˆ๋‹ค.
Footnote 20: The modelโ€™s ability to learn from junk data is negligible; each person in bioS(\(N^{\prime}\)) appears only 0.2 times during training when \(N=200k\), or 0.05 times when \(N=50k\).
์ฃผ 20: ์ •ํฌ ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์€ ๋ฌด์‹œํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋ฐ”์ด์˜ค์Šค(\(N^{\prime}\))์˜ ๊ฐ ์‚ฌ๋žŒ์€ ํ›ˆ๋ จ ์ค‘์— 0.2๋ฒˆ, ์ฆ‰ \(N=200k\)์ผ ๋•Œ 0.05๋ฒˆ๋งŒ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.
์ฃผ์„ 20: ์ •ํฌ ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์€ ๋ฏธ๋ฏธํ•˜๋ฉฐ, ๋ฐ”์ด์˜ค์Šค(\(N^{\prime}\))์˜ ๊ฐ ์‚ฌ๋žŒ์€ ํ›ˆ๋ จ ์ค‘์— \(N=200k\)์ผ ๋•Œ 0.2๋ฒˆ, \(N=50k\)์ผ ๋•Œ 0.05๋ฒˆ๋งŒ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.
20: ๋ชจ๋ธ์ด ์ •ํฌ ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๋Š” ๋Šฅ๋ ฅ์€ ๋ฌด์‹œํ•  ์ˆ˜ ์žˆ๋Š” ์ˆ˜์ค€์ด๋ฉฐ, bioS(\(N^{\prime}\))์˜ ๊ฐ ์‚ฌ๋žŒ์€ \(N=200k\)์ผ ๋•Œ ํ›ˆ๋ จ ์ค‘์— 0.2๋ฒˆ, \(N=50k\)์ผ ๋•Œ 0.05๋ฒˆ๋งŒ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.
๊ฐ์ฃผ 20: ๋ชจ๋ธ์ด ์ •ํฌ ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๋Š” ๋Šฅ๋ ฅ์€ ๋ฏธ๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋ฐ”์ด์˜คS(\(N^{\prime}\))์˜ ๊ฐ ์‚ฌ๋žŒ์€ ํ›ˆ๋ จ ์ค‘์— \(N=200k\)์ผ ๋•Œ 0.2๋ฒˆ, \(N=50k\)์ผ ๋•Œ 0.05๋ฒˆ๋งŒ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.
๊ฐ์ฃผ 20: ์ •ํฌ ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ํ•™์Šตํ•˜๋Š” ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์€ ๋ฏธ๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋ฐ”์ด์˜คS(\(N^{\prime}\))์˜ ๊ฐ ์‚ฌ๋žŒ์€ ํ›ˆ๋ จ ์ค‘์— \(N=200k\)์ผ ๋•Œ 0.2๋ฒˆ, \(N=50k\)์ผ ๋•Œ 0.05๋ฒˆ๋งŒ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.
Not all models report their performance on all datasets, and because of that, the number of models for which performance is reported in different tables varies.
๋ชจ๋“  ๋ชจ๋ธ์ด ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ๊ทธ ๋•Œ๋ฌธ์— ์„ฑ๋Šฅ์ด ๋‹ค๋ฅธ ํ‘œ์— ๋ณด๊ณ ๋˜๋Š” ๋ชจ๋ธ์˜ ์ˆ˜๋Š” ๋‹ค์–‘ํ•˜๋‹ค.
๋ชจ๋“  ๋ชจ๋ธ์ด ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ๊ทธ ๋•Œ๋ฌธ์— ์„ฑ๋Šฅ์ด ๋‹ค๋ฅธ ํ‘œ์— ๋ณด๊ณ ๋˜๋Š” ๋ชจ๋ธ์˜ ์ˆ˜๋Š” ๋‹ค์–‘ํ•˜๋‹ค.
๋ชจ๋“  ๋ชจ๋ธ์ด ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ๊ทธ ๋•Œ๋ฌธ์— ๋‹ค๋ฅธ ํ‘œ์—์„œ ์„ฑ๋Šฅ์ด ๋ณด๊ณ ๋˜๋Š” ๋ชจ๋ธ์˜ ์ˆ˜๋Š” ๋‹ค์–‘ํ•˜๋‹ค.
๋ชจ๋“  ๋ชจ๋ธ์ด ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ด๋กœ ์ธํ•ด ๋‹ค๋ฅธ ํ‘œ์—์„œ ์„ฑ๋Šฅ์ด ๋ณด๊ณ ๋˜๋Š” ๋ชจ๋ธ์˜ ์ˆ˜๊ฐ€ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค.
๋ชจ๋“  ๋ชจ๋ธ์ด ๋ชจ๋“  ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ๋ณด๊ณ ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์ด๋กœ ์ธํ•ด ๋‹ค๋ฅธ ํ‘œ์—์„œ ์„ฑ๋Šฅ์ด ๋ณด๊ณ ๋˜๋Š” ๋ชจ๋ธ์˜ ์ˆ˜๊ฐ€ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค.
For general-purpose language models, diversity and quality of instruction are crucial Cao et al. (2023); Zhou et al. (2023). However, in specialized domains it is crucial to incorporate task-specific and specialized prompts to enhance performance. Our instruction fine-tuning stage involves \(2\) key components: generic (ie, non-legal) and legal instructions. The former help enhance the model's understanding and following of commands, and includes data from diverse domains such as coding, mathematics, and general conversations. For the latter we employ an extensive collection of datasets tailored to the nuances of legal domains, covering legal question answering and summarization, among others. Through this metic lous fine-tuning on instructional data, our model, SaulLM-7B-Instruct, is able to grasp legal intricacies and excels in a wide range of associated tasks.
๋ฒ”์šฉ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, Cao ์™ธ(2023), Zhou ์™ธ(2023)๋Š” ๋‹ค์–‘์„ฑ๊ณผ ์ง€์นจ์˜ ํ’ˆ์งˆ์ด ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ „๋ฌธ ๋ถ„์•ผ์—์„œ๋Š” ์ž‘์—…๋ณ„ ๋ฐ ์ „๋ฌธ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ง€์นจ ๋ฏธ์„ธ ์กฐ์ • ๋‹จ๊ณ„์—๋Š” ์ผ๋ฐ˜(์ฆ‰, ๋น„๋ฒ•๋ฅ ) ๋ฐ ๋ฒ•๋ฅ  ์ง€์นจ์˜ ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ „์ž๋Š” ๋ชจ๋ธ์˜ ๋ช…๋ น์–ด ์ดํ•ด์™€ ๋”ฐ๋ฅด๊ธฐ๋ฅผ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋˜๋ฉฐ ์ฝ”๋”ฉ, ์ˆ˜ํ•™, ์ผ๋ฐ˜ ๋Œ€ํ™” ๋“ฑ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ํ›„์ž๋Š” ๋ฒ•๋ฅ  ์งˆ์˜์‘๋‹ต ๋ฐ ์š”์•ฝ์„ ํฌํ•จํ•œ ๋ฒ•๋ฅ  ๋„๋ฉ”์ธ์˜ ๋‰˜์•™์Šค์— ๋งž์ถ˜ ๊ด‘๋ฒ”์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ธ์‹ฌํ•œ ์ง€์นจ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด, ์ €ํฌ ๋ชจ๋ธ์ธ SaulLM-7B-Instruct๋Š” ๋ฒ•๋ฅ ์˜ ๋ณต์žก์„ฑ์„ ํŒŒ์•…ํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๊ด€๋ จ ์ž‘์—…์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ฒ”์šฉ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, Cao ์™ธ(2023), Zhou ์™ธ(2023)์˜ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์ง€์นจ์˜ ๋‹ค์–‘์„ฑ๊ณผ ํ’ˆ์งˆ์ด ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ „๋ฌธ ๋„๋ฉ”์ธ์—์„œ๋Š” ์ž‘์—…๋ณ„ ๋ฐ ์ „๋ฌธ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ง€์นจ ๋ฏธ์„ธ ์กฐ์ • ๋‹จ๊ณ„์—๋Š” ์ผ๋ฐ˜(์ฆ‰, ๋ฒ•๋ฅ ์ด ์•„๋‹Œ) ๋ฐ ๋ฒ•๋ฅ  ์ง€์นจ์˜ ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ „์ž๋Š” ๋ชจ๋ธ์ด ๋ช…๋ น์„ ์ดํ•ดํ•˜๊ณ  ๋”ฐ๋ฅด๋Š” ๋ฐ ๋„์›€์ด ๋˜๋ฉฐ ์ฝ”๋”ฉ, ์ˆ˜ํ•™, ์ผ๋ฐ˜ ๋Œ€ํ™” ๋“ฑ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ํ›„์ž๋Š” ๋ฒ•๋ฅ  ์งˆ์˜์‘๋‹ต ๋ฐ ์š”์•ฝ ๋“ฑ์„ ๋‹ค๋ฃจ๋Š” ๋ฒ•๋ฅ  ์˜์—ญ์˜ ๋‰˜์•™์Šค์— ๋งž๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ธ์‹ฌํ•œ ์ง€์นจ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด, ์ €ํฌ ๋ชจ๋ธ์ธ SaulLM-7B-Instruct๋Š” ๋ฒ•๋ฅ ์˜ ๋ณต์žก์„ฑ์„ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ๋‹ค์–‘ํ•œ ๊ด€๋ จ ์ž‘์—…์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•ฉ๋‹ˆ๋‹ค.
๋ฒ”์šฉ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, ์ง€์นจ์˜ ๋‹ค์–‘์„ฑ๊ณผ ํ’ˆ์งˆ์ด ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค Cao et al. (2023); Zhou et al. (2023). ๊ทธ๋Ÿฌ๋‚˜ ์ „๋ฌธ ๋ถ„์•ผ์—์„œ๋Š” ์ž‘์—…๋ณ„ ๋ฐ ์ „๋ฌธ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ง€์นจ ๋ฏธ์„ธ ์กฐ์ • ๋‹จ๊ณ„์—๋Š” ์ผ๋ฐ˜(์˜ˆ: ๋น„๋ฒ•๋ฅ ) ๋ฐ ๋ฒ•๋ฅ  ์ง€์นจ์˜ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ „์ž๋Š” ์ฝ”๋”ฉ, ์ˆ˜ํ•™, ์ผ๋ฐ˜ ๋Œ€ํ™” ๋“ฑ ๋‹ค์–‘ํ•œ ์˜์—ญ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํฌํ•จํ•˜์—ฌ ๋ชจ๋ธ์˜ ๋ช…๋ น ์ดํ•ด ๋ฐ ๋ช…๋ น ์ˆ˜ํ–‰์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ํ›„์ž๋Š” ๋ฒ•๋ฅ  ์งˆ์˜ ์‘๋‹ต ๋ฐ ์š”์•ฝ ๋“ฑ์„ ๋‹ค๋ฃจ๋Š” ๋ฒ•๋ฅ  ์˜์—ญ์˜ ๋‰˜์•™์Šค์— ๋งž์ถ˜ ๊ด‘๋ฒ”์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง€์นจ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์„ธ์‹ฌํ•œ ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด, ์šฐ๋ฆฌ์˜ ๋ชจ๋ธ์ธ SaulLM-7B-Instruct๋Š” ๋ฒ•๋ฅ ์˜ ๋ณต์žก์„ฑ์„ ํŒŒ์•…ํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๊ด€๋ จ ์ž‘์—…์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ฒ”์šฉ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, ๋‹ค์–‘์„ฑ๊ณผ ๊ต์œก์˜ ์งˆ์ด ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. (2023); Zhou et al. (2023). ๊ทธ๋Ÿฌ๋‚˜ ์ „๋ฌธ ๋„๋ฉ”์ธ์—์„œ๋Š” ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ž‘์—…๋ณ„ ๋ฐ ์ „๋ฌธ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ง€์นจ ๋ฏธ์„ธ ์กฐ์ • ๋‹จ๊ณ„์—๋Š” ์ผ๋ฐ˜์ ์ธ(์ฆ‰, ๋ฒ•๋ฅ ์ด ์•„๋‹Œ) ์ง€์นจ๊ณผ ๋ฒ•๋ฅ  ์ง€์นจ์˜ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ „์ž๋Š” ๋ชจ๋ธ์˜ ๋ช…๋ น ์ดํ•ด ๋ฐ ์ค€์ˆ˜๋ฅผ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋˜๋ฉฐ ์ฝ”๋”ฉ, ์ˆ˜ํ•™, ์ผ๋ฐ˜ ๋Œ€ํ™” ๋“ฑ ๋‹ค์–‘ํ•œ ๋„๋ฉ”์ธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ํ›„์ž์˜ ๊ฒฝ์šฐ ๋ฒ•๋ฅ  ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ ์š”์•ฝ ๋“ฑ ๋ฒ•๋ฅ  ๋„๋ฉ”์ธ์˜ ๋‰˜์•™์Šค๋ฅผ ๋‹ค๋ฃจ๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ต์œก ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์„ธ์‹ฌํ•œ ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด ๋ฒ•๋ฅ ์˜ ๋ณต์žก์„ฑ์„ ํŒŒ์•…ํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๊ด€๋ จ ์ž‘์—…์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” ๋ชจ๋ธ์ธ SaulLM-7B-Instruct๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค.
์ผ๋ฐ˜ ๋ชฉ์  ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ, ๋‹ค์–‘์„ฑ๊ณผ ๊ต์œก์˜ ์งˆ์ด ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค(Cao ์™ธ, 2023; Zhou ์™ธ, 2023). ๊ทธ๋Ÿฌ๋‚˜ ์ „๋ฌธ ๋„๋ฉ”์ธ์—์„œ๋Š” ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ž‘์—…๋ณ„ ๋ฐ ์ „๋ฌธ์ ์ธ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ๊ต์œก ๋ฏธ์„ธ ์กฐ์ • ๋‹จ๊ณ„์—๋Š” ์ผ๋ฐ˜(์ฆ‰, ๋ฒ•๋ฅ ์ด ์•„๋‹Œ) ๋ฐ ๋ฒ•๋ฅ  ์ง€์นจ์˜ ๋‘ ๊ฐ€์ง€ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ „์ž๋Š” ๋ชจ๋ธ์˜ ๋ช…๋ น ์ดํ•ด์™€ ์ˆ˜ํ–‰์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋˜๋ฉฐ ์ฝ”๋”ฉ, ์ˆ˜ํ•™, ์ผ๋ฐ˜ ๋Œ€ํ™” ๋“ฑ ๋‹ค์–‘ํ•œ ์˜์—ญ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ํ›„์ž์˜ ๊ฒฝ์šฐ ๋ฒ•๋ฅ  ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ ์š”์•ฝ ๋“ฑ ๋ฒ•๋ฅ  ๋„๋ฉ”์ธ์˜ ๋‰˜์•™์Šค๋ฅผ ๋‹ค๋ฃจ๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ต์œก ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ด๋Ÿฌํ•œ ์„ธ์‹ฌํ•œ ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด ๋ฒ•๋ฅ ์˜ ๋ณต์žก์„ฑ์„ ํŒŒ์•…ํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๊ด€๋ จ ์ž‘์—…์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š” SaulLM-7B-Instruct ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
As language models continue to evolve at an accelerated pace, an increasing number of large models with high parameter counts and exceptional performance are emerging. Tuning the parameters and internal structure of these models has become increasingly difficult and inefficient, making instruction tuning more important than ever.
์–ธ์–ด ๋ชจ๋ธ์ด ๊ฐ€์†ํ™”๋œ ์†๋„๋กœ ๊ณ„์† ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ๋งŽ๊ณ  ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๊ฐ€์ง„ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์ด ์ ์  ๋” ๋งŽ์ด ๋“ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜์™€ ๋‚ด๋ถ€ ๊ตฌ์กฐ๋ฅผ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ์ ์  ๋” ์–ด๋ ค์›Œ์ง€๊ณ  ๋น„ํšจ์œจ์ ์ด ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ง€์นจ ์กฐ์ •์ด ๊ทธ ์–ด๋Š ๋•Œ๋ณด๋‹ค ์ค‘์š”ํ•ด์กŒ์Šต๋‹ˆ๋‹ค.
์–ธ์–ด ๋ชจ๋ธ์ด ๊ฐ€์†ํ™”๋œ ์†๋„๋กœ ๊ณ„์† ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ๋งŽ๊ณ  ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ์ž๋ž‘ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์ด ์ ์  ๋” ๋งŽ์ด ๋“ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜์™€ ๋‚ด๋ถ€ ๊ตฌ์กฐ๋ฅผ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ์ ์  ๋” ์–ด๋ ค์›Œ์ง€๊ณ  ๋น„ํšจ์œจ์ ์ด ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ง€์นจ ์กฐ์ •์ด ๊ทธ ์–ด๋Š ๋•Œ๋ณด๋‹ค ์ค‘์š”ํ•ด์กŒ์Šต๋‹ˆ๋‹ค.
์–ธ์–ด ๋ชจ๋ธ์ด ๊ฐ€์†ํ™”๋œ ์†๋„๋กœ ๊ณ„์† ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ, ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ๋งŽ๊ณ  ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๊ฐ€์ง„ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์ด ์ ์  ๋” ๋งŽ์ด ๋“ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ์™€ ๋‚ด๋ถ€ ๊ตฌ์กฐ๋ฅผ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ์ ์  ๋” ์–ด๋ ค์›Œ์ง€๊ณ  ๋น„ํšจ์œจ์ ์ด ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ง€์นจ ํŠœ๋‹์ด ๊ทธ ์–ด๋Š ๋•Œ๋ณด๋‹ค ์ค‘์š”ํ•ด์กŒ์Šต๋‹ˆ๋‹ค.
์–ธ์–ด ๋ชจ๋ธ์ด ๊ฐ€์†ํ™”๋œ ์†๋„๋กœ ๊ณ„์† ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜๊ฐ€ ๋งŽ๊ณ  ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์ด ์ ์  ๋” ๋งŽ์ด ๋“ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ์™€ ๋‚ด๋ถ€ ๊ตฌ์กฐ๋ฅผ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์ด ์ ์  ๋” ์–ด๋ ค์›Œ์ง€๊ณ  ๋น„ํšจ์œจ์ ์ด ๋˜๋ฉด์„œ ์ง€์นจ ํŠœ๋‹์˜ ์ค‘์š”์„ฑ์ด ๊ทธ ์–ด๋Š ๋•Œ๋ณด๋‹ค ์ปค์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
์–ธ์–ด ๋ชจ๋ธ์ด ๊ฐ€์†ํ™”๋œ ์†๋„๋กœ ๊ณ„์† ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ, ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ๋งŽ๊ณ  ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์ด ์ ์  ๋” ๋งŽ์ด ๋“ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜์™€ ๋‚ด๋ถ€ ๊ตฌ์กฐ๋ฅผ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์ด ์ ์  ๋” ์–ด๋ ค์›Œ์ง€๊ณ  ๋น„ํšจ์œจ์ ์ด ๋˜๋ฉด์„œ ๋ช…๋ น์–ด ํŠœ๋‹์ด ๊ทธ ์–ด๋Š ๋•Œ๋ณด๋‹ค ์ค‘์š”ํ•ด์กŒ์Šต๋‹ˆ๋‹ค.
Due to their sparsity, MoE models often require higher learning rates compared to dense models. Consequently, we adjust the optimizer parameters as follows:
ํฌ์†Œ์„ฑ์œผ๋กœ ์ธํ•ด MoE ๋ชจ๋ธ์€ ์ข…์ข… ๊ณ ๋ฐ€๋„ ๋ชจ๋ธ์— ๋น„ํ•ด ๋” ๋†’์€ ํ•™์Šต ์†๋„๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ตœ์ ํ™”๊ธฐ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค:
ํฌ์†Œ์„ฑ ๋•Œ๋ฌธ์— MoE ๋ชจ๋ธ์€ ์ข…์ข… ๊ณ ๋ฐ€๋„ ๋ชจ๋ธ์— ๋น„ํ•ด ๋” ๋†’์€ ํ•™์Šต ์†๋„๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์˜ตํ‹ฐ๋งˆ์ด์ € ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค:
ํฌ์†Œ์„ฑ์œผ๋กœ ์ธํ•ด MoE ๋ชจ๋ธ์€ ๋ฐ€๋„๊ฐ€ ๋†’์€ ๋ชจ๋ธ์— ๋น„ํ•ด ํ•™์Šต ์†๋„๊ฐ€ ๋” ๋†’์€ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์˜ตํ‹ฐ๋งˆ์ด์ € ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค:
ํฌ์†Œ์„ฑ์œผ๋กœ ์ธํ•ด MoE ๋ชจ๋ธ์€ ๊ณ ๋ฐ€๋„ ๋ชจ๋ธ์— ๋น„ํ•ด ๋” ๋†’์€ ํ•™์Šต ์†๋„๊ฐ€ ํ•„์š”ํ•œ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ตœ์ ํ™”๊ธฐ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค:
ํฌ์†Œ์„ฑ์œผ๋กœ ์ธํ•ด MoE ๋ชจ๋ธ์€ ๊ณ ๋ฐ€๋„ ๋ชจ๋ธ์— ๋น„ํ•ด ๋” ๋†’์€ ํ•™์Šต ์†๋„๊ฐ€ ํ•„์š”ํ•œ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ตœ์ ํ™”๊ธฐ ๋งค๊ฐœ ๋ณ€์ˆ˜๋ฅผ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค:
In addition to the above basic evaluation tasks, LLMs also exhibit some superior abilities that require special consider tions for evaluation. In this part, we discuss several representative advanced abilities and the corresponding evaluation approaches, including human alignment, interaction with the external environment, and tool manipulation. Next, we discuss these advanced abilities in detail.
์œ„์˜ ๊ธฐ๋ณธ ํ‰๊ฐ€ ์ž‘์—… ์™ธ์—๋„ LLM์€ ํ‰๊ฐ€์— ํŠน๋ณ„ํ•œ ๊ณ ๋ ค๊ฐ€ ํ•„์š”ํ•œ ๋ช‡ ๊ฐ€์ง€ ์šฐ์ˆ˜ํ•œ ๋Šฅ๋ ฅ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” ์ธ๊ฐ„ ์ •๋ ฌ, ์™ธ๋ถ€ ํ™˜๊ฒฝ๊ณผ์˜ ์ƒํ˜ธ ์ž‘์šฉ, ๋„๊ตฌ ์กฐ์ž‘ ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ๋Œ€ํ‘œ์ ์ธ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ๊ณผ ๊ทธ์— ๋”ฐ๋ฅธ ํ‰๊ฐ€ ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์œ„์˜ ๊ธฐ๋ณธ ํ‰๊ฐ€ ๊ณผ์ œ ์™ธ์—๋„ LLM์€ ํ‰๊ฐ€์— ํŠน๋ณ„ํ•œ ๊ณ ๋ ค๊ฐ€ ํ•„์š”ํ•œ ๋ช‡ ๊ฐ€์ง€ ์šฐ์ˆ˜ํ•œ ๋Šฅ๋ ฅ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” ์ธ๊ฐ„๊ณผ์˜ ์ •๋ ฌ, ์™ธ๋ถ€ ํ™˜๊ฒฝ๊ณผ์˜ ์ƒํ˜ธ ์ž‘์šฉ, ๋„๊ตฌ ์กฐ์ž‘ ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ๋Œ€ํ‘œ์ ์ธ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ๊ณผ ๊ทธ์— ๋”ฐ๋ฅธ ํ‰๊ฐ€ ์ ‘๊ทผ๋ฒ•์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์œ„์˜ ๊ธฐ๋ณธ ํ‰๊ฐ€ ์ž‘์—… ์™ธ์—๋„ LLM์€ ํ‰๊ฐ€์— ํŠน๋ณ„ํ•œ ๋ฐฐ๋ ค๊ฐ€ ํ•„์š”ํ•œ ๋ช‡ ๊ฐ€์ง€ ์šฐ์ˆ˜ํ•œ ๋Šฅ๋ ฅ์„ ๋ณด์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” ์ธ๊ฐ„ ์ •๋ ฌ, ์™ธ๋ถ€ ํ™˜๊ฒฝ๊ณผ์˜ ์ƒํ˜ธ์ž‘์šฉ, ๋„๊ตฌ ์กฐ์ž‘ ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ๋Œ€ํ‘œ์ ์ธ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ๊ณผ ๊ทธ์— ๋”ฐ๋ฅธ ํ‰๊ฐ€ ์ ‘๊ทผ๋ฒ•์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์œ„์˜ ๊ธฐ๋ณธ ํ‰๊ฐ€ ์ž‘์—… ์™ธ์—๋„ LLM์€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ํŠน๋ณ„ํ•œ ๊ณ ๋ ค๊ฐ€ ํ•„์š”ํ•œ ๋ช‡ ๊ฐ€์ง€ ์šฐ์ˆ˜ํ•œ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” ์ธ๊ฐ„์˜ ์ •๋ ฌ, ์™ธ๋ถ€ ํ™˜๊ฒฝ๊ณผ์˜ ์ƒํ˜ธ ์ž‘์šฉ, ๋„๊ตฌ ์กฐ์ž‘ ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ๋Œ€ํ‘œ์ ์ธ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ๊ณผ ๊ทธ์— ์ƒ์‘ํ•˜๋Š” ํ‰๊ฐ€ ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ ์ด๋Ÿฌํ•œ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์œ„์˜ ๊ธฐ๋ณธ ํ‰๊ฐ€ ์ž‘์—… ์™ธ์—๋„ LLM์€ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ํŠน๋ณ„ํ•œ ๊ณ ๋ ค๊ฐ€ ํ•„์š”ํ•œ ๋ช‡ ๊ฐ€์ง€ ์šฐ์ˆ˜ํ•œ ๋Šฅ๋ ฅ๋„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” ์ธ๊ฐ„์˜ ์ •๋ ฌ, ์™ธ๋ถ€ ํ™˜๊ฒฝ๊ณผ์˜ ์ƒํ˜ธ์ž‘์šฉ, ๋„๊ตฌ ์กฐ์ž‘ ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ๋Œ€ํ‘œ์ ์ธ ๊ณ ๊ธ‰ ๋Šฅ๋ ฅ๊ณผ ๊ทธ์— ์ƒ์‘ํ•˜๋Š” ํ‰๊ฐ€ ์ ‘๊ทผ๋ฒ•์— ๋Œ€ํ•ด ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ ์ด๋Ÿฌํ•œ ๊ณ ๊ธ‰ ๊ธฐ๋Šฅ์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
Figure 4: Best of \(N\) baseline for \(N=\{1,4,16,64,128\}\). Performance plateaus after roughly 64-128 samples.
๊ทธ๋ฆผ 4: N=\{1,4,16,64,128\}\)์— ๋Œ€ํ•œ ์ตœ์ƒ์˜ \(N\) ๊ธฐ์ค€์„ . ์•ฝ 64-128 ์ƒ˜ํ”Œ ์ดํ›„ ์„ฑ๋Šฅ์ด ์ •์ฒด๋ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 4: \(N=\{1,4,16,64,128\}\)์— ๋Œ€ํ•œ ์ตœ์ƒ์˜ \(N\) ๊ธฐ์ค€์„ . ๋Œ€๋žต 64-128 ์ƒ˜ํ”Œ ์ดํ›„์—๋Š” ์„ฑ๋Šฅ์ด ์ •์ฒด๋ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 4: N=\{1,4,16,64,128\}์— ๋Œ€ํ•œ ์ตœ์ƒ์˜ N ๊ธฐ์ค€์„ . ์•ฝ 64-128 ์ƒ˜ํ”Œ ์ดํ›„์—๋Š” ์„ฑ๋Šฅ์ด ์ •์ฒด๋ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 4: N=\{1,4,16,64,128\}\)์— ๋Œ€ํ•œ \(N\)์˜ ์ตœ์ƒ์˜ ๊ธฐ์ค€์„ . ์•ฝ 64~128๊ฐœ์˜ ์ƒ˜ํ”Œ ์ดํ›„ ์„ฑ๋Šฅ์ด ์ •์ฒด๋ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 4: N=\{1,4,16,64,128\}\)์— ๋Œ€ํ•œ \(N\) ๊ธฐ์ค€์„  ์ค‘ ๊ฐ€์žฅ ์ข‹์€ ๊ฒƒ. ์•ฝ 64-128๊ฐœ์˜ ์ƒ˜ํ”Œ ์ดํ›„ ์„ฑ๋Šฅ์ด ํ‰์ค€ํ™”๋ฉ๋‹ˆ๋‹ค.
Let \(z\in\mathbb{R}^{\ell}\) be the (unknown) logits and let \(\mathbf{N}=\sum_{i}\exp(z_{i})\) be the normalizing constant. For each \(k\in\{1,\cdots,K\}\), we have
z\in\mathbb{R}^{\ell}\)๋ฅผ (๋ฏธ์ง€์˜) ๋กœ๊ทธ๋ผ๊ณ  ํ•˜๊ณ  \(\mathbf{N}=\sum_{i}\exp(z_{i})\)๋ฅผ ์ •๊ทœํ™” ์ƒ์ˆ˜๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ \(k\in\{1,\cdots,K\}\)์— ๋Œ€ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
z\in\mathbb{R}^{\ell}\)๋ฅผ (๋ฏธ์ง€์˜) ๋กœ๊ทธ๋ผ๊ณ  ํ•˜๊ณ , \(\mathbf{N}=\sum_{i}\exp(z_{i})\)๋ฅผ ์ •๊ทœํ™” ์ƒ์ˆ˜๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ๊ฐ์˜ \(k\in\{1,\cdots,K\}\)์— ๋Œ€ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
z\in\mathbb{R}^{\ell}\)๋ฅผ (๋ฏธ์ง€์˜) ๋กœ๊ทธ๋ผ๊ณ  ํ•˜๊ณ  \(\mathbf{N}=\sum_{i}\exp(z_{i})\)๋ฅผ ์ •๊ทœํ™” ์ƒ์ˆ˜๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ \(k\in\{1,\cdots,K\}\)์— ๋Œ€ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
z\in\mathbb{R}^{\ell}\)๋ฅผ (์•Œ๋ ค์ง€์ง€ ์•Š์€) ๋กœ๊ทธ๋กœ ํ•˜๊ณ , \(\mathbf{N}=\sum_{i}\exp(z_{i})\)๋ฅผ ์ •๊ทœํ™” ์ƒ์ˆ˜๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ฐ \(k\in\{1,\cdots,K\}\)์— ๋Œ€ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
z\in\mathbb{R}^{\ell}\)๋ฅผ (์•Œ ์ˆ˜ ์—†๋Š”) ๋กœ๊ทธ๋กœ ํ•˜๊ณ , \(z_{i})\)๋ฅผ ์ •๊ทœํ™” ์ƒ์ˆ˜๋กœ ํ•˜์ž. ๊ฐ \(k\in\{1,\cdots,K\}\)์— ๋Œ€ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
For Supervised Fine-Tuning (SFT), we used both Chinese and English data. The Chinese data consisted of the full set from CQIA (Bai et al., 2024) and OL-CC, as well as high-quality data sampled from COIG-PC (Zhang et al., 2023). The English data was sampled from the OpenHermesPreferences dataset (Huang et al., 2024). The total amount of Chinese data comprised 105K pairs of instruction data, with English data adjusted to different ratios based on the volume of Chinese data. The ratios were \(1:1,2:1,4:1,\) and \(8:1\), along with configurations that included only Chinese data and only English data. Each set of experiments was trained for 3 epochs, with specific experimental results shown in Table 12.
๊ฐ๋… ๋ฏธ์„ธ ์กฐ์ •(SFT)์˜ ๊ฒฝ์šฐ ์ค‘๊ตญ์–ด์™€ ์˜์–ด ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋‘ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋Š” CQIA(Bai ์™ธ., 2024)์˜ ์ „์ฒด ์„ธํŠธ์™€ OL-CC, COIG-PC(Zhang ์™ธ., 2023)์—์„œ ์ƒ˜ํ”Œ๋งํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜์–ด ๋ฐ์ดํ„ฐ๋Š” OpenHermesPreferences ๋ฐ์ดํ„ฐ ์„ธํŠธ(Huang ์™ธ., 2024)์—์„œ ์ƒ˜ํ”Œ๋ง๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ์˜ ์ด๋Ÿ‰์€ 105K์Œ์˜ ์ง€์นจ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์œผ๋ฉฐ, ์˜์–ด ๋ฐ์ดํ„ฐ๋Š” ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ์˜ ์–‘์— ๋”ฐ๋ผ ๋‹ค์–‘ํ•œ ๋น„์œจ๋กœ ์กฐ์ •๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋น„์œจ์€ \(1:1,2:1,4:1,\) ๋ฐ \(8:1\)์ด์—ˆ์œผ๋ฉฐ, ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จํ•˜๋Š” ๊ตฌ์„ฑ๊ณผ ์˜์–ด ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จํ•˜๋Š” ๊ตฌ์„ฑ๋„ ํฌํ•จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์‹คํ—˜ ์„ธํŠธ๋Š” 3๊ฐœ ์—ํฌํฌ ๋™์•ˆ ํ›ˆ๋ จ๋˜์—ˆ์œผ๋ฉฐ, ๊ตฌ์ฒด์ ์ธ ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 12์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •(SFT)์˜ ๊ฒฝ์šฐ ์ค‘๊ตญ์–ด์™€ ์˜์–ด ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋‘ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋Š” CQIA(Bai ์™ธ, 2024)์˜ ์ „์ฒด ์„ธํŠธ์™€ OL-CC, ๊ทธ๋ฆฌ๊ณ  COIG-PC(Zhang ์™ธ, 2023)์—์„œ ์ƒ˜ํ”Œ๋งํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜์–ด ๋ฐ์ดํ„ฐ๋Š” OpenHermesPreferences ๋ฐ์ดํ„ฐ ์„ธํŠธ(Huang ์™ธ, 2024)์—์„œ ์ƒ˜ํ”Œ๋ง๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ์˜ ์ด๋Ÿ‰์€ 105K์Œ์˜ ์ง€์นจ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์œผ๋ฉฐ, ์˜์–ด ๋ฐ์ดํ„ฐ๋Š” ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ์˜ ์–‘์— ๋”ฐ๋ผ ๋‹ค์–‘ํ•œ ๋น„์œจ๋กœ ์กฐ์ •๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋น„์œจ์€ \(1:1,2:1,4:1,\) ๋ฐ \(8:1\)์ด์—ˆ๊ณ , ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จํ•˜๋Š” ๊ตฌ์„ฑ๊ณผ ์˜์–ด ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จํ•˜๋Š” ๊ตฌ์„ฑ๋„ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์‹คํ—˜ ์„ธํŠธ๋Š” 3๊ฐœ ์—ํฌํฌ ๋™์•ˆ ํ›ˆ๋ จ๋˜์—ˆ์œผ๋ฉฐ, ๊ตฌ์ฒด์ ์ธ ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 12์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ฐ๋… ๋ฏธ์„ธ ์กฐ์ •(SFT)์˜ ๊ฒฝ์šฐ ์ค‘๊ตญ์–ด์™€ ์˜์–ด ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋‘ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋Š” CQIA(Bai et al., 2024)์˜ ์ „์ฒด ์„ธํŠธ์™€ OL-CC, COIG-PC(Zhang et al., 2023)์—์„œ ์ƒ˜ํ”Œ๋งํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜์–ด ๋ฐ์ดํ„ฐ๋Š” OpenHermesPreferences ๋ฐ์ดํ„ฐ ์„ธํŠธ(Huang et al., 2024)์—์„œ ์ƒ˜ํ”Œ๋ง๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ์˜ ์ด๋Ÿ‰์€ 105K์Œ์˜ ์ง€์นจ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์œผ๋ฉฐ, ์˜์–ด ๋ฐ์ดํ„ฐ๋Š” ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ์˜ ์–‘์— ๋”ฐ๋ผ ๋‹ค์–‘ํ•œ ๋น„์œจ๋กœ ์กฐ์ •๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋น„์œจ์€ \(1:1,2:1,4:1,\) ๋ฐ \(8:1\)์ด์—ˆ๊ณ , ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จํ•˜๋Š” ๊ตฌ์„ฑ๊ณผ ์˜์–ด ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จํ•˜๋Š” ๊ตฌ์„ฑ๋„ ํฌํ•จํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์‹คํ—˜ ์„ธํŠธ๋Š” 3๊ฐœ ์—ํฌํฌ ๋™์•ˆ ํ›ˆ๋ จ๋˜์—ˆ์œผ๋ฉฐ, ํ‘œ 12์— ๊ตฌ์ฒด์ ์ธ ์‹คํ—˜ ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •(SFT)์˜ ๊ฒฝ์šฐ ์ค‘๊ตญ์–ด์™€ ์˜์–ด ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋‘ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋Š” CQIA(Bai et al., 2024) ๋ฐ OL-CC์˜ ์ „์ฒด ์„ธํŠธ์™€ COIG-PC(Zhang et al., 2023)์—์„œ ์ƒ˜ํ”Œ๋งํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜์–ด ๋ฐ์ดํ„ฐ๋Š” OpenHermesPreferences ๋ฐ์ดํ„ฐ ์„ธํŠธ(Huang et al., 2024)์—์„œ ์ƒ˜ํ”Œ๋งํ–ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ์˜ ์ด ์–‘์€ 105K ์Œ์˜ ์ง€์นจ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์œผ๋ฉฐ, ์˜์–ด ๋ฐ์ดํ„ฐ๋Š” ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ์˜ ์–‘์— ๋”ฐ๋ผ ๋‹ค์–‘ํ•œ ๋น„์œจ๋กœ ์กฐ์ •๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋น„์œจ์€ \(1:1,2:1,4:1,\) ๋ฐ \(8:1\)์ด์—ˆ์œผ๋ฉฐ, ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จ๋œ ๊ตฌ์„ฑ๊ณผ ์˜์–ด ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จ๋œ ๊ตฌ์„ฑ์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์‹คํ—˜ ์„ธํŠธ๋Š” 3 ์—ํฌํฌ ๋™์•ˆ ํ›ˆ๋ จ๋˜์—ˆ์œผ๋ฉฐ, ๊ตฌ์ฒด์ ์ธ ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 12์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •(SFT)์˜ ๊ฒฝ์šฐ ์ค‘๊ตญ์–ด์™€ ์˜์–ด ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋‘ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋Š” CQIA(Bai et al., 2024)์™€ OL-CC์˜ ์ „์ฒด ์„ธํŠธ, ๊ทธ๋ฆฌ๊ณ  COIG-PC(Zhang et al., 2023)์—์„œ ์ƒ˜ํ”Œ๋งํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์˜์–ด ๋ฐ์ดํ„ฐ๋Š” OpenHermesPreferences ๋ฐ์ดํ„ฐ ์„ธํŠธ(Huang et al., 2024)์—์„œ ์ƒ˜ํ”Œ๋งํ–ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ์˜ ์ด ์–‘์€ 105,000์Œ์˜ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์œผ๋ฉฐ, ์˜์–ด ๋ฐ์ดํ„ฐ๋Š” ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ์˜ ์–‘์— ๋”ฐ๋ผ ๋‹ค์–‘ํ•œ ๋น„์œจ๋กœ ์กฐ์ •๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋น„์œจ์€ \(1:1,2:1,4:1,\) ๋ฐ \(8:1\)์ด์—ˆ์œผ๋ฉฐ, ์ค‘๊ตญ์–ด ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จํ•˜๋Š” ๊ตฌ์„ฑ๊ณผ ์˜์–ด ๋ฐ์ดํ„ฐ๋งŒ ํฌํ•จํ•˜๋Š” ๊ตฌ์„ฑ๋„ ํฌํ•จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์‹คํ—˜ ์„ธํŠธ๋Š” 3๊ฐœ์˜ ์—ํฌํฌ์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋˜์—ˆ์œผ๋ฉฐ, ๊ตฌ์ฒด์ ์ธ ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ํ‘œ 12์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.
PersonA: Urinalysis can be performed in various healthcare settings, including doctors' offices, urgent care facilities, laboratories, and hospitals. You can provide a urine sample in a specimen cup, and the sample can be analyzed on-siteor sent to a laboratory for testing. PersonU: How much urine is typically needed for the test? PersonA: Usually, only small amounts of urine, ranging from 30 to 60 ml, are required for urinalysis testing. It's a non-invasive and relatively simple test that provides valuable information about your health.
์‚ฌ๋žŒA: ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ์˜์‚ฌ ์‚ฌ๋ฌด์‹ค, ์‘๊ธ‰ ์น˜๋ฃŒ ์‹œ์„ค, ์‹คํ—˜์‹ค, ๋ณ‘์› ๋“ฑ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ํ™˜๊ฒฝ์—์„œ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฒด ์ปต์— ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•˜๋ฉด ๊ฒ€์‚ฌ ํ˜„์žฅ์—์„œ ๊ฒ€์‚ฌํ•˜๊ฑฐ๋‚˜ ๊ฒ€์‚ฌ ๊ฒฐ๊ณผ๋ฅผ ์‹คํ—˜์‹ค๋กœ ๋ณด๋‚ด์„œ ๊ฒ€์‚ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ๋žŒU: ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฒ€์‚ฌ์— ํ•„์š”ํ•œ ์†Œ๋ณ€์€ ์–ผ๋งˆ๋‚˜ ๋˜๋‚˜์š”? ์‚ฌ๋žŒA: ์ผ๋ฐ˜์ ์œผ๋กœ ์†Œ๋ณ€ ๊ฒ€์‚ฌ์—๋Š” 30~60ml ์ •๋„์˜ ์†Œ๋Ÿ‰์˜ ์†Œ๋ณ€์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋น„์นจ์Šต์ ์ด๊ณ  ๋น„๊ต์  ๊ฐ„๋‹จํ•œ ๊ฒ€์‚ฌ๋กœ ๊ฑด๊ฐ•์— ๋Œ€ํ•œ ๊ท€์ค‘ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ์†Œ๋ณ€ ๋ถ„์„์€ ์˜์‚ฌ ์‚ฌ๋ฌด์‹ค, ์‘๊ธ‰ ์น˜๋ฃŒ ์‹œ์„ค, ์‹คํ—˜์‹ค, ๋ณ‘์› ๋“ฑ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ํ™˜๊ฒฝ์—์„œ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฒด ์ปต์— ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•˜๋ฉด ๊ฒ€์ฒด๋ฅผ ํ˜„์žฅ์—์„œ ๋ถ„์„ํ•˜๊ฑฐ๋‚˜ ๊ฒ€์‚ฌ๋ฅผ ์œ„ํ•ด ์‹คํ—˜์‹ค๋กœ ๋ณด๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ๋žŒU: ๊ฒ€์‚ฌ์— ์ผ๋ฐ˜์ ์œผ๋กœ ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ์†Œ๋ณ€์ด ํ•„์š”ํ•˜๋‚˜์š”? ์‚ฌ๋žŒA: ์ผ๋ฐ˜์ ์œผ๋กœ ์†Œ๋ณ€ ๋ถ„์„ ๊ฒ€์‚ฌ์—๋Š” 30~60ml ์ •๋„์˜ ์†Œ๋Ÿ‰์˜ ์†Œ๋ณ€์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋น„์นจ์Šต์ ์ด๊ณ  ๋น„๊ต์  ๊ฐ„๋‹จํ•œ ๊ฒ€์‚ฌ๋กœ ๊ฑด๊ฐ•์— ๋Œ€ํ•œ ๊ท€์ค‘ํ•œ ์ •๋ณด๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ์˜์‚ฌ ์‚ฌ๋ฌด์‹ค, ์‘๊ธ‰ ์น˜๋ฃŒ ์‹œ์„ค, ์‹คํ—˜์‹ค ๋ฐ ๋ณ‘์›์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ํ™˜๊ฒฝ์—์„œ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฒด ์ปต์— ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•˜๋ฉด ํ˜„์žฅ์—์„œ ๊ฒ€์‚ฌํ•˜๊ฑฐ๋‚˜ ๊ฒ€์‚ฌ์‹ค๋กœ ๋ณด๋‚ด์„œ ๊ฒ€์‚ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ๋žŒU: ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฒ€์‚ฌ์— ํ•„์š”ํ•œ ์†Œ๋ณ€์˜ ์–‘์€ ์–ผ๋งˆ๋‚˜ ๋˜๋‚˜์š”? ์‚ฌ๋žŒA: ์ผ๋ฐ˜์ ์œผ๋กœ ์†Œ๋ณ€ ๊ฒ€์‚ฌ์—๋Š” 30~60ml ์ •๋„์˜ ์†Œ๋Ÿ‰์˜ ์†Œ๋ณ€๋งŒ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋น„์นจ์Šต์ ์ด๊ณ  ๋น„๊ต์  ๊ฐ„๋‹จํ•œ ๊ฒ€์‚ฌ๋กœ ๊ฑด๊ฐ•์— ๋Œ€ํ•œ ๊ท€์ค‘ํ•œ ์ •๋ณด๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ์˜์‚ฌ ์‚ฌ๋ฌด์‹ค, ์‘๊ธ‰ ์น˜๋ฃŒ ์‹œ์„ค, ์‹คํ—˜์‹ค ๋ฐ ๋ณ‘์›์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ํ™˜๊ฒฝ์—์„œ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฒด ์ปต์— ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•˜๋ฉด ํ˜„์žฅ์—์„œ ๋ถ„์„ํ•˜๊ฑฐ๋‚˜ ์‹คํ—˜์‹ค๋กœ ๋ณด๋‚ด ๊ฒ€์‚ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ๋žŒU: ์ผ๋ฐ˜์ ์œผ๋กœ ๊ฒ€์‚ฌ์— ํ•„์š”ํ•œ ์†Œ๋ณ€์˜ ์–‘์€ ์–ด๋Š ์ •๋„์ธ๊ฐ€์š”? ์‚ฌ๋žŒA: ์ผ๋ฐ˜์ ์œผ๋กœ ์†Œ๋ณ€ ๊ฒ€์‚ฌ์—๋Š” 30~60ml ์ •๋„์˜ ์†Œ๋Ÿ‰์˜ ์†Œ๋ณ€๋งŒ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋น„์นจ์Šต์ ์ด๊ณ  ๋น„๊ต์  ๊ฐ„๋‹จํ•œ ๊ฒ€์‚ฌ๋กœ ๊ฑด๊ฐ•์— ๋Œ€ํ•œ ๊ท€์ค‘ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์‚ฌ๋žŒA: ์†Œ๋ณ€ ๊ฒ€์‚ฌ๋Š” ์˜์‚ฌ ์‚ฌ๋ฌด์‹ค, ๊ธด๊ธ‰ ์น˜๋ฃŒ ์‹œ์„ค, ์‹คํ—˜์‹ค, ๋ณ‘์› ๋“ฑ ๋‹ค์–‘ํ•œ ์˜๋ฃŒ ํ™˜๊ฒฝ์—์„œ ์‹ค์‹œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฒด ์ปต์— ์†Œ๋ณ€ ์ƒ˜ํ”Œ์„ ์ œ๊ณตํ•˜๋ฉด ํ˜„์žฅ์—์„œ ๊ฒ€์‚ฌํ•˜๊ฑฐ๋‚˜ ๊ฒ€์‚ฌ๋ฅผ ์œ„ํ•ด ์‹คํ—˜์‹ค๋กœ ๋ณด๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ๋žŒU: ๊ฒ€์‚ฌ์— ์ผ๋ฐ˜์ ์œผ๋กœ ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ์†Œ๋ณ€์ด ํ•„์š”ํ•˜๋‚˜์š”? ์‚ฌ๋žŒA: ์ผ๋ฐ˜์ ์œผ๋กœ ์†Œ๋ณ€ ๊ฒ€์‚ฌ์—๋Š” 30~60ml์˜ ์†Œ๋Ÿ‰์˜ ์†Œ๋ณ€๋งŒ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ฑด๊ฐ•์— ๋Œ€ํ•œ ์œ ์šฉํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋Š” ๋น„์นจ์Šต์ ์ด๊ณ  ๋น„๊ต์  ๊ฐ„๋‹จํ•œ ๊ฒ€์‚ฌ์ž…๋‹ˆ๋‹ค.
**Efficiency of Information Acquisition.** Although AMIE displayed greater verbosity compared to PCPs in terms of total number of words generated in their responses during the consultation, the number of conversational turns and the number of words elicited from the patient actors were similar across both OSCE agents, as illustrated in Figure A.11. This suggests that both AMIE and PCPs acquired a similar amount of information from the patients during the encounter. To investigate how efficient AMIE or PCPs were at gathering sufficient information to formulate a correct diagnosis, we truncated the conversations at various turn counts and used AMIE to generate differential diagnoses based on these partial conversations. Figure A.12 depicts the top-3 DDx accuracy as a function of the number of turns provided to the model. The observed accuracies plateaued within the initial 10 conversational turns for both AMIE and PCPs. This suggests that both AMIE and PCPs were able to acquire the information necessary for formulating a diagnosis within the early stages of the conversation. Additionally, the comparable performance at every turn indicates that neither AMIE nor PCPs had a significant advantage in the efficiency or quality of information acquisition.
**์ •๋ณด ํš๋“์˜ ํšจ์œจ์„ฑ.** AMIE๋Š” ์ƒ๋‹ด ์ค‘ ์‘๋‹ต์—์„œ ์ƒ์„ฑ๋œ ์ด ๋‹จ์–ด ์ˆ˜ ์ธก๋ฉด์—์„œ PCP์— ๋น„ํ•ด ๋” ๋งŽ์€ ์ˆ˜์˜ ๋‹จ์–ด๋ฅผ ์ƒ์„ฑํ–ˆ์ง€๋งŒ, ๊ทธ๋ฆผ A.11์— ๋‚˜์™€ ์žˆ๋“ฏ์ด ๋Œ€ํ™” ํ„ด ์ˆ˜์™€ ํ™˜์ž ๋ฐฐ์šฐ์ž์—๊ฒŒ์„œ ์ด๋Œ์–ด๋‚ธ ๋‹จ์–ด ์ˆ˜๋Š” ๋‘ OSCE ์—์ด์ „ํŠธ ๋ชจ๋‘์—์„œ ๋น„์Šทํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AMIE์™€ PCP๊ฐ€ ๋ชจ๋‘ ํ™˜์ž๋กœ๋ถ€ํ„ฐ ์œ ์‚ฌํ•œ ์–‘์˜ ์ •๋ณด๋ฅผ ํš๋“ํ–ˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. AMIE ๋˜๋Š” PCP๊ฐ€ ์˜ฌ๋ฐ”๋ฅธ ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๊ธฐ ์œ„ํ•ด ์ถฉ๋ถ„ํ•œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์–ผ๋งˆ๋‚˜ ํšจ์œจ์ ์ธ์ง€ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ํ„ด ์ˆ˜์—์„œ ๋Œ€ํ™”๋ฅผ ์ž˜๋ผ๋‚ด๊ณ  AMIE๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ถ€๋ถ„์ ์ธ ๋Œ€ํ™”๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ฐจ๋“ฑ ์ง„๋‹จ์„ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ A.12๋Š” ๋ชจ๋ธ์— ์ œ๊ณต๋œ ํ„ด ์ˆ˜์— ๋”ฐ๋ฅธ ์ƒ์œ„ 3๊ฐœ DDx ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. AMIE์™€ PCP ๋ชจ๋‘ ์ดˆ๊ธฐ 10ํšŒ ๋Œ€ํ™” ํ„ด ๋‚ด์— ์ •ํ™•๋„๊ฐ€ ์ •์ ์„ ์ฐ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AMIE์™€ PCP ๋ชจ๋‘ ๋Œ€ํ™” ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ํ•„์š”ํ•œ ์ •๋ณด๋ฅผ ํš๋“ํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋งค ํ„ด๋งˆ๋‹ค ๋น„์Šทํ•œ ์„ฑ๊ณผ๋ฅผ ๊ฑฐ๋‘” ๊ฒƒ์€ AMIE๋‚˜ PCP ๋ชจ๋‘ ์ •๋ณด ํš๋“์˜ ํšจ์œจ์„ฑ์ด๋‚˜ ํ’ˆ์งˆ์—์„œ ํฐ ์ด์ ์ด ์—†์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
**์ •๋ณด ํš๋“์˜ ํšจ์œจ์„ฑ.** AMIE๋Š” ์ƒ๋‹ด ์ค‘ ์‘๋‹ต์—์„œ ์ƒ์„ฑํ•œ ์ด ๋‹จ์–ด ์ˆ˜ ์ธก๋ฉด์—์„œ PCP์— ๋น„ํ•ด ๋” ๋งŽ์€ ๋‹จ์–ด๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์ง€๋งŒ, ๊ทธ๋ฆผ A.11์— ๋‚˜์™€ ์žˆ๋“ฏ์ด ๋Œ€ํ™” ํ„ด ์ˆ˜์™€ ํ™˜์ž ๋ฐฐ์šฐ์ž๋กœ๋ถ€ํ„ฐ ์ด๋Œ์–ด๋‚ธ ๋‹จ์–ด ์ˆ˜๋Š” ๋‘ OSCE ์—์ด์ „ํŠธ ๋ชจ๋‘ ์œ ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AMIE์™€ PCP๊ฐ€ ๋ฉด๋‹ด ์ค‘์— ํ™˜์ž๋กœ๋ถ€ํ„ฐ ์œ ์‚ฌํ•œ ์–‘์˜ ์ •๋ณด๋ฅผ ํš๋“ํ–ˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. AMIE ๋˜๋Š” PCP๊ฐ€ ์˜ฌ๋ฐ”๋ฅธ ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๊ธฐ ์œ„ํ•ด ์ถฉ๋ถ„ํ•œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์–ผ๋งˆ๋‚˜ ํšจ์œจ์ ์ธ์ง€ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๋Œ€ํ™”๋ฅผ ๋‹ค์–‘ํ•œ ํ„ด ์ˆ˜๋กœ ์ž˜๋ผ๋‚ด๊ณ  AMIE๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ถ€๋ถ„ ๋Œ€ํ™”๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ฐจ๋“ฑ ์ง„๋‹จ์„ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ A.12๋Š” ๋ชจ๋ธ์— ์ œ๊ณต๋œ ํ„ด ์ˆ˜์— ๋”ฐ๋ฅธ ์ƒ์œ„ 3๊ฐœ DDx ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ด€์ฐฐ๋œ ์ •ํ™•๋„๋Š” AMIE์™€ PCP ๋ชจ๋‘ ์ดˆ๊ธฐ 10ํšŒ ๋Œ€ํ™” ํ„ด ๋‚ด์— ์ •์ฒด๋˜๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AMIE์™€ PCP๊ฐ€ ๋Œ€ํ™” ์ดˆ๊ธฐ ๋‹จ๊ณ„ ๋‚ด์— ์ง„๋‹จ์„ ๊ณต์‹ํ™”ํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ์ •๋ณด๋ฅผ ํš๋“ํ•  ์ˆ˜ ์žˆ์—ˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋งค ํ„ด๋งˆ๋‹ค ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์€ AMIE์™€ PCP ๋ชจ๋‘ ์ •๋ณด ํš๋“์˜ ํšจ์œจ์„ฑ์ด๋‚˜ ํ’ˆ์งˆ์—์„œ ํฐ ์ด์ ์ด ์—†์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
**์ •๋ณด ํš๋“์˜ ํšจ์œจ์„ฑ.** AMIE๋Š” ์ƒ๋‹ด ์ค‘์— ์ƒ์„ฑํ•œ ์ด ๋‹จ์–ด ์ˆ˜ ์ธก๋ฉด์—์„œ PCP์— ๋น„ํ•ด ๋” ๋งŽ์€ ๋‹จ์–ด๋ฅผ ์‚ฌ์šฉํ–ˆ์ง€๋งŒ, ๋Œ€ํ™” ํ„ด ์ˆ˜์™€ ํ™˜์ž ๋ฐฐ์šฐ์ž์—๊ฒŒ์„œ ์ด๋Œ์–ด๋‚ธ ๋‹จ์–ด ์ˆ˜๋Š” ๊ทธ๋ฆผ A.11์— ๋‚˜์™€ ์žˆ๋“ฏ์ด ๋‘ OSCE ์—์ด์ „ํŠธ ๋ชจ๋‘ ์œ ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AMIE์™€ PCP๊ฐ€ ํ™˜์ž๋กœ๋ถ€ํ„ฐ ์ถฉ๋ถ„ํ•œ ์ •๋ณด๋ฅผ ํš๋“ํ–ˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. AMIE ๋˜๋Š” PCP๊ฐ€ ์ •ํ™•ํ•œ ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ํ•„์š”ํ•œ ์ •๋ณด๋ฅผ ์–ผ๋งˆ๋‚˜ ํšจ์œจ์ ์œผ๋กœ ์ˆ˜์ง‘ํ–ˆ๋Š”์ง€ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ํ„ด ์ˆ˜์—์„œ ๋Œ€ํ™”๋ฅผ ์ž˜๋ผ๋‚ด๊ณ  AMIE๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ถ€๋ถ„์ ์ธ ๋Œ€ํ™”์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ๊ฐ๋ณ„ ์ง„๋‹จ์„ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ A.12๋Š” ๋ชจ๋ธ์— ์ œ๊ณต๋œ ํ„ด ์ˆ˜์— ๋”ฐ๋ผ ์ƒ์œ„ 3๊ฐœ์˜ DDx ์ •ํ™•๋„๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. AMIE์™€ PCP ๋ชจ๋‘ ์ดˆ๊ธฐ 10ํšŒ ๋Œ€ํ™” ํ„ด ๋‚ด์— ์ •ํ™•๋„๊ฐ€ ์ •์ฒด๋˜๋Š” ๊ฒƒ์œผ๋กœ ๊ด€์ฐฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AMIE์™€ PCP๊ฐ€ ๋Œ€ํ™” ์ดˆ๊ธฐ ๋‹จ๊ณ„ ๋‚ด์— ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ํ•„์š”ํ•œ ์ •๋ณด๋ฅผ ํš๋“ํ•  ์ˆ˜ ์žˆ์—ˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋งค ํ„ด๋งˆ๋‹ค ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ธ ๊ฒƒ์€ AMIE์™€ PCP ๋ชจ๋‘ ์ •๋ณด ์ˆ˜์ง‘์˜ ํšจ์œจ์„ฑ์ด๋‚˜ ํ’ˆ์งˆ์—์„œ ํฐ ์ด์ ์„ ์–ป์ง€ ๋ชปํ–ˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
**์ •๋ณด ์ˆ˜์ง‘์˜ ํšจ์œจ์„ฑ.** ์ƒ๋‹ด ์ค‘ ์‘๋‹ต์— ์ƒ์„ฑ๋œ ์ด ๋‹จ์–ด ์ˆ˜์—์„œ AMIE๊ฐ€ PCP์— ๋น„ํ•ด ๋” ๋งŽ์€ ์žฅํ™ฉํ•จ์„ ๋ณด์˜€์ง€๋งŒ, ๊ทธ๋ฆผ A.11๊ณผ ๊ฐ™์ด ํ™˜์ž ๋ฐฐ์šฐ๋กœ๋ถ€ํ„ฐ ์œ ๋„๋œ ๋Œ€ํ™” ์ „ํ™˜ ์ˆ˜์™€ ๋‹จ์–ด ์ˆ˜๋Š” ๋‘ OSCE ์š”์› ๋ชจ๋‘ ๋น„์Šทํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AMIE์™€ PCP๊ฐ€ ๋ฉด๋‹ด ์ค‘์— ํ™˜์ž๋กœ๋ถ€ํ„ฐ ๋น„์Šทํ•œ ์–‘์˜ ์ •๋ณด๋ฅผ ํš๋“ํ–ˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. AMIE ๋˜๋Š” PCP๊ฐ€ ์ •ํ™•ํ•œ ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๊ธฐ ์œ„ํ•ด ์ถฉ๋ถ„ํ•œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์–ผ๋งˆ๋‚˜ ํšจ์œจ์ ์ธ์ง€ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋Œ€ํ™” ์ „ํ™˜ ์ˆ˜์—์„œ ๋Œ€ํ™”๋ฅผ ์ž˜๋ผ๋‚ด๊ณ  AMIE๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋Ÿฌํ•œ ๋ถ€๋ถ„ ๋Œ€ํ™”๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ๋ณ„ ์ง„๋‹จ์„ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ A.12๋Š” ๋ชจ๋ธ์— ์ œ๊ณต๋œ ๋Œ€ํ™” ์ „ํ™˜ ์ˆ˜์— ๋”ฐ๋ฅธ ์ƒ์œ„ 3๊ฐœ์˜ DDx ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ด€์ฐฐ๋œ ์ •ํ™•๋„๋Š” AMIE์™€ PCP ๋ชจ๋‘ ์ดˆ๊ธฐ 10๊ฐœ์˜ ๋Œ€ํ™” ์ „ํ™˜ ๋‚ด์—์„œ ์•ˆ์ •ํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AMIE์™€ PCP ๋ชจ๋‘ ๋Œ€ํ™” ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ํ•„์š”ํ•œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•  ์ˆ˜ ์žˆ์—ˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋“  ํ„ด์—์„œ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค๋Š” ๊ฒƒ์€ AMIE์™€ PCP ๋ชจ๋‘ ์ •๋ณด ์ˆ˜์ง‘์˜ ํšจ์œจ์„ฑ์ด๋‚˜ ํ’ˆ์งˆ์—์„œ ํฐ ์ด์ ์ด ์—†์—ˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
**์ •๋ณด ์ˆ˜์ง‘์˜ ํšจ์œจ์„ฑ. **์ƒ๋‹ด ์ค‘ ๋‹ต๋ณ€์— ์ƒ์„ฑ๋œ ์ด ๋‹จ์–ด ์ˆ˜์—์„œ AMIE๊ฐ€ PCP์— ๋น„ํ•ด ๋” ๋งŽ์€ ์žฅํ™ฉํ•จ์„ ๋ณด์˜€์ง€๋งŒ, ๊ทธ๋ฆผ A.11๊ณผ ๊ฐ™์ด ๋Œ€ํ™” ํšŸ์ˆ˜์™€ ํ™˜์ž ํ–‰์œ„์ž๋กœ๋ถ€ํ„ฐ ์ด๋Œ์–ด๋‚ธ ๋‹จ์–ด ์ˆ˜๋Š” ๋‘ OSCE ์š”์› ๊ฐ„์— ์œ ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AMIE์™€ PCP๊ฐ€ ์ง„๋ฃŒ ์ค‘ ํ™˜์ž๋กœ๋ถ€ํ„ฐ ๋น„์Šทํ•œ ์–‘์˜ ์ •๋ณด๋ฅผ ํš๋“ํ–ˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. AMIE ๋˜๋Š” PCP๊ฐ€ ์ •ํ™•ํ•œ ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๊ธฐ ์œ„ํ•ด ์ถฉ๋ถ„ํ•œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์–ผ๋งˆ๋‚˜ ํšจ์œจ์ ์ธ์ง€ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ํ„ด ์ˆ˜์—์„œ ๋Œ€ํ™”๋ฅผ ์ž˜๋ผ๋‚ด๊ณ  AMIE๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋Ÿฌํ•œ ๋ถ€๋ถ„ ๋Œ€ํ™”๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ๋ณ„ ์ง„๋‹จ์„ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ A.12๋Š” ๋ชจ๋ธ์— ์ œ๊ณต๋œ ํ„ด ์ˆ˜์— ๋”ฐ๋ฅธ ์ƒ์œ„ 3๊ฐœ DDx ์ •ํ™•๋„๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๊ด€์ฐฐ๋œ ์ •ํ™•๋„๋Š” AMIE์™€ PCP ๋ชจ๋‘ ์ฒ˜์Œ 10๋ฒˆ์˜ ๋Œ€ํ™” ํ„ด ๋‚ด์—์„œ ์•ˆ์ •ํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” AMIE์™€ PCP ๋ชจ๋‘ ๋Œ€ํ™” ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๋Š” ๋ฐ ํ•„์š”ํ•œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•  ์ˆ˜ ์žˆ์—ˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋“  ํ„ด์—์„œ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค๋Š” ๊ฒƒ์€ AMIE์™€ PCP ๋ชจ๋‘ ์ •๋ณด ์ˆ˜์ง‘์˜ ํšจ์œจ์„ฑ์ด๋‚˜ ํ’ˆ์งˆ ์ธก๋ฉด์—์„œ ํฐ ์ด์ ์ด ์—†์—ˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
1. **SlimOrca** This subset of the FLAN collection comprises generic instructions, offering a focused resource for various tasks Mukherjee et al. (2023); Lian et al. (2023). 2. **Meta Math Question Answering Instructions** Designed for mathematical inquiry, this dataset15 presents a range of mathematical questions, facilitating research in math-based natural language processing Yu et al. (2023). Footnote 15: Accessible at meta-math/MetaMathQA 3. **General Conversations from UltraChat** Capturing diverse conversational contexts,
1. **SlimOrca** FLAN ์ปฌ๋ ‰์…˜์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ, ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ์œ„ํ•œ ์ดˆ์  ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•˜๋Š” ์ผ๋ฐ˜ ์ง€์นจ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. Mukherjee ์™ธ(2023); Lian ์™ธ(2023). 2. **๋ฉ”ํƒ€ ์ˆ˜ํ•™ ๋ฌธ์ œ ํ•ด๊ฒฐ ์ง€์นจ** ์ˆ˜ํ•™์  ํƒ๊ตฌ๋ฅผ ์œ„ํ•ด ์„ค๊ณ„๋œ ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ15๋Š” ๋‹ค์–‘ํ•œ ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ์ œ์‹œํ•˜์—ฌ ์ˆ˜ํ•™ ๊ธฐ๋ฐ˜ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์—ฐ๊ตฌ๋ฅผ ์šฉ์ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. Yu ์™ธ(2023). 15: meta-math/MetaMathQA์—์„œ ์•ก์„ธ์Šค ๊ฐ€๋Šฅ 3. **UltraChat์˜ ์ผ๋ฐ˜์ ์ธ ๋Œ€ํ™”** ๋‹ค์–‘ํ•œ ๋Œ€ํ™” ๋งฅ๋ฝ์„ ํฌ์ฐฉํ•ฉ๋‹ˆ๋‹ค,
1. **SlimOrca** FLAN ์ปฌ๋ ‰์…˜์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ, ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ์œ„ํ•œ ์ดˆ์ ์ด ๋งž์ถฐ์ง„ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋ฌด์ปค์ง€ ๋“ฑ(2023); ๋ฆฌ์•ˆ ๋“ฑ(2023). 2. **๋ฉ”ํƒ€ ์ˆ˜ํ•™ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์ง€์นจ** ์ˆ˜ํ•™์  ํƒ๊ตฌ๋ฅผ ์œ„ํ•ด ์„ค๊ณ„๋œ ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ15๋Š” ๋‹ค์–‘ํ•œ ์ˆ˜ํ•™์  ์งˆ๋ฌธ์„ ์ œ์‹œํ•˜์—ฌ ์ˆ˜ํ•™ ๊ธฐ๋ฐ˜ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค. ์œ  ๋“ฑ(2023). 15: meta-math/MetaMathQA์—์„œ ์•ก์„ธ์Šค ๊ฐ€๋Šฅ 3. **UltraChat์˜ ์ผ๋ฐ˜์ ์ธ ๋Œ€ํ™”** ๋‹ค์–‘ํ•œ ๋Œ€ํ™” ๋งฅ๋ฝ์„ ํฌ์ฐฉํ•ฉ๋‹ˆ๋‹ค,
1. **์Šฌ๋ฆผ์˜ค๋ฅด์นด** FLAN ์ปฌ๋ ‰์…˜์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ, ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ์œ„ํ•œ ์ดˆ์ ์„ ๋งž์ถ˜ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•˜๋Š” ์ผ๋ฐ˜ ์ง€์นจ์ž…๋‹ˆ๋‹ค. Mukherjee et al. (2023); Lian et al. (2023). 2. **๋ฉ”ํƒ€ ์ˆ˜ํ•™ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์ง€์นจ** ์ˆ˜ํ•™์  ํƒ๊ตฌ๋ฅผ ์œ„ํ•ด ์„ค๊ณ„๋œ ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ15๋Š” ๋‹ค์–‘ํ•œ ์ˆ˜ํ•™์  ์งˆ๋ฌธ์„ ์ œ์‹œํ•˜์—ฌ ์ˆ˜ํ•™ ๊ธฐ๋ฐ˜ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์—ฐ๊ตฌ๋ฅผ ์šฉ์ดํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. Yu et al. (2023). 15: meta-math/MetaMathQA์—์„œ ์•ก์„ธ์Šค ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. 3. **UltraChat์˜ ์ผ๋ฐ˜์ ์ธ ๋Œ€ํ™”** ๋‹ค์–‘ํ•œ ๋Œ€ํ™” ์ƒํ™ฉ์„ ํฌ์ฐฉํ•ฉ๋‹ˆ๋‹ค,
1. **SlimOrca** FLAN ์ปฌ๋ ‰์…˜์˜ ์ด ํ•˜์œ„ ์ง‘ํ•ฉ์€ ์ผ๋ฐ˜์ ์ธ ์ง€์นจ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์–ด ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ์œ„ํ•œ ์ง‘์ค‘ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค Mukherjee et al. (2023); Lian et al. (2023).2. **๋ฉ”ํƒ€ ์ˆ˜ํ•™ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์ง€์นจ** ์ˆ˜ํ•™์  ํƒ๊ตฌ๋ฅผ ์œ„ํ•ด ์„ค๊ณ„๋œ ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ15๋Š” ๋‹ค์–‘ํ•œ ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ์ œ์‹œํ•˜์—ฌ ์ˆ˜ํ•™ ๊ธฐ๋ฐ˜ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค Yu et al. (2023). ๊ฐ์ฃผ 15: ๋ฉ”ํƒ€-์ˆ˜ํ•™/๋ฉ”ํƒ€์ˆ˜ํ•™QA์—์„œ ์•ก์„ธ์Šค ๊ฐ€๋Šฅ3. **์šธํŠธ๋ผ์ฑ—์˜ ์ผ๋ฐ˜ ๋Œ€ํ™”** ๋‹ค์–‘ํ•œ ๋Œ€ํ™” ๋งฅ๋ฝ์„ ํฌ์ฐฉํ•ฉ๋‹ˆ๋‹ค,
1. **์Šฌ๋ฆผ์˜คํฌ๋ผ** FLAN ์ปฌ๋ ‰์…˜์˜ ์ด ํ•˜์œ„ ์ง‘ํ•ฉ์€ ์ผ๋ฐ˜์ ์ธ ๋ช…๋ น์–ด๋ฅผ ํฌํ•จํ•˜๋ฉฐ, ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ์œ„ํ•œ ์ง‘์ค‘์ ์ธ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค(Mukherjee ์™ธ. (2023); Lian ์™ธ. (2023)).2. **๋ฉ”ํƒ€ ์ˆ˜ํ•™ ์งˆ์˜์‘๋‹ต ์ง€์นจ** ์ˆ˜ํ•™์  ํƒ๊ตฌ๋ฅผ ์œ„ํ•ด ์„ค๊ณ„๋œ ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ15๋Š” ๋‹ค์–‘ํ•œ ์ˆ˜ํ•™์  ์งˆ๋ฌธ์„ ์ œ์‹œํ•˜์—ฌ ์ˆ˜ํ•™ ๊ธฐ๋ฐ˜ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค(Yu ์™ธ. (2023)). ๊ฐ์ฃผ 15: ๋ฉ”ํƒ€-์ˆ˜ํ•™/๋ฉ”ํƒ€์ˆ˜ํ•™QA์—์„œ ์•ก์„ธ์Šค ๊ฐ€๋Šฅ3. **์šธํŠธ๋ผ์ฑ—์˜ ์ผ๋ฐ˜ ๋Œ€ํ™”** ๋‹ค์–‘ํ•œ ๋Œ€ํ™” ๋งฅ๋ฝ์„ ํฌ์ฐฉํ•ฉ๋‹ˆ๋‹ค,
The disadvantage of studying emergent abilities in the lens of pre-training loss is that the pre-training loss is affected by the tokenizer and the distribution of pre-training corpus. The values of pre-training loss of language models trained on different corpus are not directly comparable. One possible solution is to evaluate different language models on a public validation set with the normalized perplexity [34] to account for the different vocabulary sizes.
์‚ฌ์ „ ํ•™์Šต ์†์‹ค์˜ ๋ Œ์ฆˆ์—์„œ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์„ ์—ฐ๊ตฌํ•˜๋Š” ๋‹จ์ ์€ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์ด ํ† ํฐํ™”๊ธฐ์™€ ์‚ฌ์ „ ํ•™์Šต ๋ง๋ญ‰์น˜์˜ ๋ถ„ํฌ์— ์˜ํ–ฅ์„ ๋ฐ›๋Š”๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ์„œ๋กœ ๋‹ค๋ฅธ ๋ง๋ญ‰์น˜์— ๋Œ€ํ•ด ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค ๊ฐ’์€ ์ง์ ‘ ๋น„๊ตํ•  ์ˆ˜ ์—†๋‹ค. ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ํ•ด๊ฒฐ์ฑ…์€ ์ •๊ทœํ™”๋œ ๋‹นํ˜น๊ฐ [34]์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐœ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์—์„œ ๋‹ค์–‘ํ•œ ์–ธ์–ด ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์–ดํœ˜ ํฌ๊ธฐ๋ฅผ ์„ค๋ช…ํ•˜๋Š” ๊ฒƒ์ด๋‹ค.
์‚ฌ์ „ ํ•™์Šต ์†์‹ค์˜ ๋ Œ์ฆˆ์—์„œ ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์„ ์—ฐ๊ตฌํ•˜๋Š” ๊ฒƒ์˜ ๋‹จ์ ์€ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์ด ํ† ํฐํ™”๊ธฐ์™€ ์‚ฌ์ „ ํ•™์Šต ๋ง๋ญ‰์น˜์˜ ๋ถ„ํฌ์— ์˜ํ–ฅ์„ ๋ฐ›๋Š”๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์„œ๋กœ ๋‹ค๋ฅธ ๋ง๋ญ‰์น˜์— ๋Œ€ํ•ด ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค ๊ฐ’์€ ์ง์ ‘ ๋น„๊ตํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ํ•ด๊ฒฐ์ฑ…์€ ๋‹ค์–‘ํ•œ ์–ดํœ˜ ํฌ๊ธฐ๋ฅผ ๊ณ ๋ คํ•˜๊ธฐ ์œ„ํ•ด ์ •๊ทœํ™”๋œ ๋‚œํ•ด์„ฑ[34]์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐœ ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์„ธํŠธ์—์„œ ๋‹ค์–‘ํ•œ ์–ธ์–ด ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
์‚ฌ์ „ ํ•™์Šต ์†์‹ค์˜ ๋ Œ์ฆˆ๋กœ ์‹น์ด ํŠธ๋Š” ๋Šฅ๋ ฅ์„ ์—ฐ๊ตฌํ•˜๋Š” ๋‹จ์ ์€ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์ด ํ† ํฐํ™” ๋„๊ตฌ์™€ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์˜ ๋ถ„ํฌ์— ์˜ํ–ฅ์„ ๋ฐ›๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค ๊ฐ’์€ ์ง์ ‘ ๋น„๊ตํ•  ์ˆ˜ ์—†๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ํ•ด๊ฒฐ์ฑ…์€ ๋‹ค์–‘ํ•œ ์–ดํœ˜ ๊ทœ๋ชจ๋ฅผ ๊ณ ๋ คํ•˜๊ธฐ ์œ„ํ•ด ์ •๊ทœํ™”๋œ ๋‹นํ˜น๊ฐ[34]์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐœ ๊ฒ€์ฆ ์„ธํŠธ์—์„œ ๋‹ค์–‘ํ•œ ์–ธ์–ด ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
์‚ฌ์ „ ํ•™์Šต ์†์‹ค์˜ ๊ด€์ ์—์„œ ์ถœํ˜„ ๋Šฅ๋ ฅ์„ ์—ฐ๊ตฌํ•˜๋Š” ๋‹จ์ ์€ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์ด ํ† ํฐํ™” ๋ฐ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์˜ ๋ถ„ํฌ์— ์˜ํ–ฅ์„ ๋ฐ›๋Š”๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์„œ๋กœ ๋‹ค๋ฅธ ์ฝ”ํผ์Šค์—์„œ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค ๊ฐ’์€ ์ง์ ‘์ ์œผ๋กœ ๋น„๊ตํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ํ•ด๊ฒฐ์ฑ…์€ ์„œ๋กœ ๋‹ค๋ฅธ ์–ดํœ˜ ํฌ๊ธฐ๋ฅผ ๊ณ ๋ คํ•˜๊ธฐ ์œ„ํ•ด ์ •๊ทœํ™”๋œ ๋‚œํ•ด์„ฑ [34]์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐœ ๊ฒ€์ฆ ์„ธํŠธ์—์„œ ๋‹ค์–‘ํ•œ ์–ธ์–ด ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
์‚ฌ์ „ ํ•™์Šต ์†์‹ค์ด๋ผ๋Š” ๋ Œ์ฆˆ๋ฅผ ํ†ตํ•ด ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์„ ์—ฐ๊ตฌํ•˜๋Š” ๋ฐ์—๋Š” ์‚ฌ์ „ ํ•™์Šต ์†์‹ค์ด ํ† ํฐํ™” ๋ฐ ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์˜ ๋ถ„ํฌ์— ์˜ํ–ฅ์„ ๋ฐ›๋Š”๋‹ค๋Š” ๋‹จ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์„œ๋กœ ๋‹ค๋ฅธ ๋ง๋ญ‰์น˜์—์„œ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค ๊ฐ’์€ ์ง์ ‘ ๋น„๊ตํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๊ฐ€๋Šฅํ•œ ํ•ด๊ฒฐ์ฑ…์€ ์„œ๋กœ ๋‹ค๋ฅธ ์–ดํœ˜ ํฌ๊ธฐ๋ฅผ ๊ณ ๋ คํ•˜๊ธฐ ์œ„ํ•ด ์ •๊ทœํ™”๋œ ๋‚œํ•ด์„ฑ [34]์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐœ ๊ฒ€์ฆ ์„ธํŠธ์—์„œ ์„œ๋กœ ๋‹ค๋ฅธ ์–ธ์–ด ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Large Language ModelsLarge language models Brown et al. (2020); Achiam et al. (2023); Touvron et al. (2023, 2023) have made significant achievements, with impressive performance on a wide range of tasks. Currently, closed-source large language models, represented by GPT Brown et al. (2020); Achiam et al. (2023), Gemini Team et al. (2023), Groke (xAI, 2023), and Claude-2 (Anthropic, 2023), are the most advanced models in terms of performance. However, open-source models, represented by LLaMA Touvron et al. (2023), LLaMA-2 Touvron et al. (2023) and Mixtral Jiang et al. (2024), have also progressed rapidly, and have even shown competitive performance with the closed-source models on some tasks. Our work, which aims to improve the performance of open-source LLMs on mathematical tasks by fine-tuning them on synthetic data.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ Brown et al. (2020); Achiam et al. (2023); Touvron et al. (2023, 2023)์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์ธ์ƒ์ ์ธ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋ฉฐ ์ƒ๋‹นํ•œ ์„ฑ๊ณผ๋ฅผ ๊ฑฐ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ GPT Brown et al. (2020); Achiam et al. (2023), ์ œ๋ฏธ๋‹ˆ ํŒ€ et al. (2023), ๊ทธ๋กœํฌ(xAI, 2023), ํด๋กœ๋“œ-2(Anthropic, 2023)๋กœ ๋Œ€ํ‘œ๋˜๋Š” ํด๋กœ์ฆˆ๋“œ ์†Œ์Šค ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์ด ์„ฑ๋Šฅ ์ธก๋ฉด์—์„œ ๊ฐ€์žฅ ๋ฐœ์ „๋œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LLaMA Touvron et al. (2023), LLaMA-2 Touvron et al. (2023), Mixtral Jiang et al. (2024)๋กœ ๋Œ€ํ‘œ๋˜๋Š” ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ๋„ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜์—ฌ ์ผ๋ถ€ ์ž‘์—…์—์„œ ํด๋กœ์ฆˆ๋“œ ์†Œ์Šค ๋ชจ๋ธ๊ณผ ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ํŒŒ์ธ ํŠœ๋‹ํ•˜์—ฌ ์ˆ˜ํ•™์  ์ž‘์—…์—์„œ ์˜คํ”ˆ ์†Œ์Šค LLM์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ์ €ํฌ์˜ ์ž‘์—….
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ Brown et al. (2020); Achiam et al. (2023); Touvron et al. (2023, 2023)์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์ธ์ƒ์ ์ธ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋ฉฐ ์ƒ๋‹นํ•œ ์„ฑ๊ณผ๋ฅผ ๊ฑฐ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ, ํ์‡„ ์†Œ์Šค ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ GPT Brown et al. (2020); Achiam et al. (2023), ์ œ๋ฏธ๋‹ˆ ํŒ€ et al. (2023), ๊ทธ๋กœ์ผ€(xAI, 2023), ํด๋กœ๋“œ-2(Anthropic, 2023)๋กœ ๋Œ€ํ‘œ๋˜๋Š” ๋ชจ๋ธ์ด ์„ฑ๋Šฅ ๋ฉด์—์„œ ๊ฐ€์žฅ ๋ฐœ์ „๋œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LLaMA Touvron et al. (2023), LLaMA-2 Touvron et al. (2023), Mixtral Jiang et al. (2024)๋กœ ๋Œ€ํ‘œ๋˜๋Š” ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ๋„ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜์—ฌ ์ผ๋ถ€ ์ž‘์—…์—์„œ ํ์‡„ ์†Œ์Šค ๋ชจ๋ธ๊ณผ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์ˆ˜ํ•™์  ์ž‘์—…์—์„œ ์˜คํ”ˆ ์†Œ์Šค LLM์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ์ €ํฌ์˜ ์ž‘์—…์ž…๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ Brown et al. (2020); Achiam et al. (2023); Touvron et al. (2023, 2023)์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์ธ์ƒ์ ์ธ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋ฉฐ ์ƒ๋‹นํ•œ ์„ฑ๊ณผ๋ฅผ ๊ฑฐ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ GPT Brown et al. (2020); Achiam et al. (2023), Gemini Team et al. (2023), Groke (xAI, 2023), Claude-2 (Anthropic, 2023)๋กœ ๋Œ€ํ‘œ๋˜๋Š” ๋น„๊ณต๊ฐœ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์ด ์„ฑ๋Šฅ ๋ฉด์—์„œ ๊ฐ€์žฅ ๋ฐœ์ „๋œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LLaMA Touvron et al. (2023), LLaMA-2 Touvron et al. (2023) ๋ฐ Mixtral Jiang et al. (2024)๋กœ ๋Œ€ํ‘œ๋˜๋Š” ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ๋„ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜์—ฌ ์ผ๋ถ€ ์ž‘์—…์—์„œ๋Š” ๋น„๊ณต๊ฐœ ๋ชจ๋ธ๊ณผ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ณด์ด๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ณธ ์—ฐ๊ตฌ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์ˆ˜ํ•™์  ์ž‘์—…์—์„œ ์˜คํ”ˆ ์†Œ์Šค LLM์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ Brown et al. (2020); Achiam et al. (2023); Touvron et al. (2023, 2023)์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์ธ์ƒ์ ์ธ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋ฉฐ ์ƒ๋‹นํ•œ ์„ฑ๊ณผ๋ฅผ ๊ฑฐ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ GPT Brown et al. (2020); Achiam et al. (2023), Gemini Team et al. (2023), Groke(xAI, 2023), Claude-2(Anthropic, 2023)๋กœ ๋Œ€ํ‘œ๋˜๋Š” ๋น„๊ณต๊ฐœ ์†Œ์Šค ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์ด ์„ฑ๋Šฅ ์ธก๋ฉด์—์„œ ๊ฐ€์žฅ ์ง„๋ณด๋œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LLaMA Touvron et al. (2023), LLaMA-2 Touvron et al. (2023), Mixtral Jiang et al. (2024)๋กœ ๋Œ€ํ‘œ๋˜๋Š” ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ๋„ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜์—ฌ ์ผ๋ถ€ ์ž‘์—…์—์„œ๋Š” ๋น„๊ณต๊ฐœ ์†Œ์Šค ๋ชจ๋ธ๊ณผ ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์— ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์ˆ˜ํ•™์  ์ž‘์—…์— ๋Œ€ํ•œ ์˜คํ”ˆ ์†Œ์Šค LLM์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ์ €ํฌ์˜ ์ž‘์—…์ž…๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ Brown ์™ธ(2020); Achiam ์™ธ(2023); Touvron ์™ธ(2023, 2023)๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์ธ์ƒ์ ์ธ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋ฉฐ ์ƒ๋‹นํ•œ ์„ฑ๊ณผ๋ฅผ ๊ฑฐ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ ํ์‡„ํ˜• ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ GPT Brown ์™ธ(2020); Achiam ์™ธ(2023), Gemini Team ์™ธ(2023), Groke(xAI, 2023), Claude-2(Anthropic, 2023)๋กœ ๋Œ€ํ‘œ๋˜๋Š” ๋ชจ๋ธ์ด ์„ฑ๋Šฅ ์ธก๋ฉด์—์„œ ๊ฐ€์žฅ ์ง„๋ณด๋œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LLaMA Touvron ์™ธ(2023), LLaMA-2 Touvron ์™ธ(2023), Mixtral Jiang ์™ธ(2024)๋กœ ๋Œ€ํ‘œ๋˜๋Š” ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ๋„ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜์—ฌ ์ผ๋ถ€ ์ž‘์—…์—์„œ๋Š” ๋น„๊ณต๊ฐœ ์†Œ์Šค ๋ชจ๋ธ๊ณผ ๊ฒฝ์Ÿํ•  ์ˆ˜ ์žˆ๋Š” ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๊ธฐ๋„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์— ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์ˆ˜ํ•™์  ๊ณผ์ œ์— ๋Œ€ํ•œ ์˜คํ”ˆ ์†Œ์Šค LLM์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ์ €ํฌ์˜ ์ž‘์—…์ž…๋‹ˆ๋‹ค.
In this section, we discuss our post-training methods to extend the Yi base model to 200K long-context, equip it with visual understanding capability, and enhance the 6B model by depth upscaling.
์ด ์„น์…˜์—์„œ๋Š” Yi ๋ฒ ์ด์Šค ๋ชจ๋ธ์„ 200K ์žฅ๋ฌธ๋งฅ์œผ๋กœ ํ™•์žฅํ•˜๊ณ , ์‹œ๊ฐ์  ์ดํ•ด ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๊ณ , 6B ๋ชจ๋ธ์„ ์‹ฌ๋„ ํ™•๋Œ€๋กœ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ์‚ฌํ›„ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” Yi ๋ฒ ์ด์Šค ๋ชจ๋ธ์„ 200K ์žฅ๋ฌธ๋งฅ์œผ๋กœ ํ™•์žฅํ•˜๊ณ , ์‹œ๊ฐ์  ์ดํ•ด ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๊ณ , ์‹ฌ๋„ ํ™•๋Œ€๋กœ 6B ๋ชจ๋ธ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ์‚ฌํ›„ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•œ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” Yi ๋ฒ ์ด์Šค ๋ชจ๋ธ์„ 200K ๊ธด ์ปจํ…์ŠคํŠธ๋กœ ํ™•์žฅํ•˜๊ณ , ์‹œ๊ฐ์  ์ดํ•ด ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๊ณ , ์‹ฌ๋„ ํ™•์žฅ์œผ๋กœ 6B ๋ชจ๋ธ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ์‚ฌํ›„ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•œ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ํ›ˆ๋ จ ํ›„ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด Yi ๊ธฐ๋ณธ ๋ชจ๋ธ์„ 200K์˜ ๊ธด ์ปจํ…์ŠคํŠธ๋กœ ํ™•์žฅํ•˜๊ณ , ์‹œ๊ฐ์  ์ดํ•ด ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ”๊ณ , ๊นŠ์ด ์—…์Šค์ผ€์ผ๋ง์„ ํ†ตํ•ด 6B ๋ชจ๋ธ์„ ๊ฐ•ํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์ด ์„น์…˜์—์„œ๋Š” ํ›ˆ๋ จ ํ›„ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•˜์—ฌ Yi ๊ธฐ๋ณธ ๋ชจ๋ธ์„ 200K ๊ธด ์ปจํ…์ŠคํŠธ๋กœ ํ™•์žฅํ•˜๊ณ , ์‹œ๊ฐ์  ์ดํ•ด ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ”๊ณ , ๊นŠ์ด ์—…์Šค์ผ€์ผ๋ง์„ ํ†ตํ•ด 6B ๋ชจ๋ธ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
Figure 1: Infini-attention has an additional compressive memory with linear attention for processing infinitely long contexts. \(\{KV\}_{s-1}\) and \(\{KV\}_{s}\) are attention key and values for current and previous input segments, respectively and \(Q_{s}\) the attention queries. PE denotes position embeddings.
๊ทธ๋ฆผ 1: ์ธํ”ผ๋‹ˆ-์–ดํ…์…˜์€ ๋ฌดํ•œํžˆ ๊ธด ์ปจํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ์„ ํ˜• ์–ดํ…์…˜์„ ๊ฐ–์ถ˜ ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ถ”๊ฐ€๋กœ ๊ฐ–์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. \{KV\}_{s-1}\) ๋ฐ \{KV\}_{s}\)๋Š” ๊ฐ๊ฐ ํ˜„์žฌ ๋ฐ ์ด์ „ ์ž…๋ ฅ ์„ธ๊ทธ๋จผํŠธ์— ๋Œ€ํ•œ ์ฃผ์˜ ํ‚ค ๋ฐ ๊ฐ’์ด๊ณ , \(Q_{s}\)๋Š” ์ฃผ์˜ ์ฟผ๋ฆฌ์ž…๋‹ˆ๋‹ค. PE๋Š” ์œ„์น˜ ์ž„๋ฒ ๋”ฉ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1: ์ธํ”ผ๋‹ˆ-์–ดํ…์…˜์€ ๋ฌดํ•œํžˆ ๊ธด ์ปจํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ์„ ํ˜• ์–ดํ…์…˜์„ ๊ฐ–๋Š” ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ถ”๊ฐ€๋กœ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. \{KV\}_{s-1}\) ๋ฐ \{KV\}_{s}\)๋Š” ๊ฐ๊ฐ ํ˜„์žฌ ๋ฐ ์ด์ „ ์ž…๋ ฅ ์„ธ๊ทธ๋จผํŠธ์— ๋Œ€ํ•œ ์ฃผ์˜ ํ‚ค ๋ฐ ๊ฐ’์ด๊ณ , \(Q_{s}\)๋Š” ์ฃผ์˜ ์ฟผ๋ฆฌ์ž…๋‹ˆ๋‹ค. PE๋Š” ์œ„์น˜ ์ž„๋ฒ ๋”ฉ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1: ์ธํ”ผ๋‹ˆ ์–ดํ…์…˜์€ ๋ฌดํ•œํžˆ ๊ธด ์ปจํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ์„ ํ˜• ์–ดํ…์…˜์„ ์‚ฌ์šฉํ•˜๋Š” ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ถ”๊ฐ€๋กœ ๊ฐ–์Šต๋‹ˆ๋‹ค. \(\{KV\}_{s-1}\) ๋ฐ \(\{KV\}_{s}\)๋Š” ๊ฐ๊ฐ ํ˜„์žฌ ๋ฐ ์ด์ „ ์ž…๋ ฅ ์„ธ๊ทธ๋จผํŠธ์— ๋Œ€ํ•œ ์–ดํ…์…˜ ํ‚ค ๋ฐ ๊ฐ’์ด๊ณ , \(Q_{s}\)๋Š” ์–ดํ…์…˜ ์ฟผ๋ฆฌ์ž…๋‹ˆ๋‹ค. PE๋Š” ์œ„์น˜ ์ž„๋ฒ ๋”ฉ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1: Infini-attention์€ ๋ฌดํ•œํžˆ ๊ธด ์ปจํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ์„ ํ˜• ์ฃผ์˜๋ ฅ์„ ๊ฐ€์ง„ ์ถ”๊ฐ€ ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. \{KV\}_{s-1}\)์™€ \{KV\}_{s}\)๋Š” ๊ฐ๊ฐ ํ˜„์žฌ ๋ฐ ์ด์ „ ์ž…๋ ฅ ์„ธ๊ทธ๋จผํŠธ์— ๋Œ€ํ•œ ์ฃผ์˜ ํ‚ค์™€ ๊ฐ’์ด๋ฉฐ, \(Q_{s}\)๋Š” ์ฃผ์˜ ์ฟผ๋ฆฌ์ž…๋‹ˆ๋‹ค. PE๋Š” ์œ„์น˜ ์ž„๋ฒ ๋”ฉ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๊ทธ๋ฆผ 1: ์ธํ”ผ๋‹ˆ-์–ดํ…์…˜์€ ๋ฌดํ•œํžˆ ๊ธด ์ปจํ…์ŠคํŠธ๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ์„ ํ˜• ์–ดํ…์…˜๊ณผ ํ•จ๊ป˜ ์ถ”๊ฐ€ ์••์ถ• ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. \{KV\}_{s-1}\) ๋ฐ \{KV\}_{s}\)๋Š” ๊ฐ๊ฐ ํ˜„์žฌ ๋ฐ ์ด์ „ ์ž…๋ ฅ ์„ธ๊ทธ๋จผํŠธ์— ๋Œ€ํ•œ ์ฃผ์˜ ํ‚ค์™€ ๊ฐ’์ด๋ฉฐ, \(Q_{s}\)๋Š” ์ฃผ์˜ ์ฟผ๋ฆฌ์ž…๋‹ˆ๋‹ค. PE๋Š” ์œ„์น˜ ์ž„๋ฒ ๋”ฉ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
Logit bias XOR logprobs.Our attack is \(10\times\) cheaper when an adversary can supply both a logit bias and also view output logprobs. This suggests a natural mitigation: prohibit queries to the API that make use of _both_ logit bias and logprobs at the same time. This type of defense is common in both the security and machine learning community: for example, in 2023 OpenAI removed the ability to combine both echo and logprobs, but with either alone being allowed; this defense would behave similarly.
๋กœ์ง“ ๋ฐ”์ด์–ด์Šค XOR ๋กœ๊ทธํ”„๋กœ๋ธŒ.๊ณต๊ฒฉ์ž๋Š” ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค์™€ ์ถœ๋ ฅ ๋กœ๊ทธํ”„๋กœ๋ธŒ๋ฅผ ๋ชจ๋‘ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ ๊ณต๊ฒฉ ๋น„์šฉ์ด \(10\times\) ๋” ์ €๋ ดํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ž์—ฐ์Šค๋Ÿฌ์šด ์™„ํ™”์ฑ…์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค: ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค์™€ ๋กœ๊ทธํ”„๋กœ๋ธŒ๋ฅผ ๋™์‹œ์— ์‚ฌ์šฉํ•˜๋Š” API์— ๋Œ€ํ•œ ์ฟผ๋ฆฌ๋ฅผ ๊ธˆ์ง€ํ•˜์‹ญ์‹œ์˜ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ๋ฐฉ์–ด๋Š” ๋ณด์•ˆ ๋ฐ ๋จธ์‹  ๋Ÿฌ๋‹ ์ปค๋ฎค๋‹ˆํ‹ฐ ๋ชจ๋‘์—์„œ ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 2023๋…„ OpenAI๋Š” ์—์ฝ”์™€ ๋กœ๊ทธํ”„๋กœ๋ธŒ๋ฅผ ๋ชจ๋‘ ๊ฒฐํ•ฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์„ ์ œ๊ฑฐํ–ˆ์ง€๋งŒ ๋‘˜ ์ค‘ ํ•˜๋‚˜๋งŒ ํ—ˆ์šฉํ–ˆ๋Š”๋ฐ, ์ด ๋ฐฉ์–ด ๋ฐฉ์‹์€ ๋น„์Šทํ•œ ๋ฐฉ์‹์œผ๋กœ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค.
๋กœ์ง“ ๋ฐ”์ด์–ด์Šค XOR ๋กœ๊ทธํ”„๋กœ๋ธŒ.๊ณต๊ฒฉ์€ ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค์™€ ์ถœ๋ ฅ ๋กœ๊ทธํ”„๋กœ๋ธŒ๋ฅผ ๋ชจ๋‘ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋Š” ๊ณต๊ฒฉ์ž๊ฐ€ ์žˆ์„ ๋•Œ \(10\times\) ๋” ์ €๋ ดํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ž์—ฐ์Šค๋Ÿฌ์šด ์™„ํ™”์ฑ…์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค: ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค์™€ ๋กœ๊ทธํ”„๋กœ๋ธŒ๋ฅผ ๋™์‹œ์— ์‚ฌ์šฉํ•˜๋Š” API์— ๋Œ€ํ•œ ์ฟผ๋ฆฌ๋ฅผ ๊ธˆ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ๋ฐฉ์–ด๋Š” ๋ณด์•ˆ ๋ฐ ๋จธ์‹  ๋Ÿฌ๋‹ ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ๋ชจ๋‘ ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 2023๋…„ OpenAI๋Š” ์—์ฝ”์™€ ๋กœ๊ทธํ”„๋กœ๋ธŒ๋ฅผ ๋ชจ๋‘ ๊ฒฐํ•ฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์„ ์ œ๊ฑฐํ–ˆ์ง€๋งŒ, ๋‘˜ ์ค‘ ํ•˜๋‚˜๋งŒ ํ—ˆ์šฉํ–ˆ๋Š”๋ฐ, ์ด๋Ÿฌํ•œ ๋ฐฉ์–ด ๋ฐฉ์‹๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค.
๋กœ์ง“ ๋ฐ”์ด์–ด์Šค XOR ๋กœ๊ทธ ํ™•๋ฅ .๊ณต๊ฒฉ์ž๋Š” ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค์™€ ์ถœ๋ ฅ ๋กœ๊ทธ ํ™•๋ฅ ์„ ๋ชจ๋‘ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ ๊ณต๊ฒฉ ๋น„์šฉ์ด \(10\times\) ์ €๋ ดํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ž์—ฐ์Šค๋Ÿฌ์šด ์™„ํ™”๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค: ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค์™€ ๋กœ๊ทธ ํ™•๋ฅ ์„ ๋™์‹œ์— ์‚ฌ์šฉํ•˜๋Š” API์— ๋Œ€ํ•œ ์ฟผ๋ฆฌ๋ฅผ ๊ธˆ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ๋ฐฉ์–ด๋Š” ๋ณด์•ˆ ๋ฐ ๋จธ์‹  ๋Ÿฌ๋‹ ์ปค๋ฎค๋‹ˆํ‹ฐ ๋ชจ๋‘์—์„œ ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 2023๋…„ OpenAI๋Š” ์—์ฝ”์™€ ๋กœ๊ทธ ํ™•๋ฅ ์„ ๋ชจ๋‘ ๊ฒฐํ•ฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์„ ์ œ๊ฑฐํ–ˆ์ง€๋งŒ ๋‘˜ ์ค‘ ํ•˜๋‚˜๋งŒ ํ—ˆ์šฉํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ด ๋ฐฉ์–ด๋ฅผ ๋น„์Šทํ•˜๊ฒŒ ์ž‘๋™ํ•˜๋„๋ก ํ–ˆ์Šต๋‹ˆ๋‹ค.
๋กœ์ง“ ๋ฐ”์ด์–ด์Šค XOR ๋กœ๊ทธ ํ”„๋กœ๋ธŒ.์ ๋Œ€์ ์ธ ํ–‰์œ„์ž๊ฐ€ ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค์™€ ์ถœ๋ ฅ ๋กœ๊ทธ ํ”„๋กœ๋ธŒ๋ฅผ ๋ชจ๋‘ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ, ๊ณต๊ฒฉ ๋น„์šฉ์€ \(10\๋ฐฐ\) ๋” ์ €๋ ดํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ž์—ฐ์Šค๋Ÿฌ์šด ์™„ํ™”์ฑ…์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค: ๋กœ์ง“ ๋ฐ”์ด์–ด์Šค์™€ ๋กœ๊ทธ ํ”„๋กœ๋ธŒ๋ฅผ ๋™์‹œ์— ์‚ฌ์šฉํ•˜๋Š” API์— ๋Œ€ํ•œ ์ฟผ๋ฆฌ๋ฅผ ๊ธˆ์ง€ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ๋ฐฉ์–ด๋Š” ๋ณด์•ˆ ๋ฐ ๋จธ์‹  ๋Ÿฌ๋‹ ์ปค๋ฎค๋‹ˆํ‹ฐ ๋ชจ๋‘์—์„œ ํ”ํžˆ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 2023๋…„ OpenAI๋Š” ์—์ฝ”์™€ ๋กœ๊ทธ ํ”„๋กœ๋ธŒ๋ฅผ ๋ชจ๋‘ ๊ฒฐํ•ฉํ•˜๋Š” ๊ธฐ๋Šฅ์„ ์ œ๊ฑฐํ–ˆ์ง€๋งŒ, ๋‘˜ ์ค‘ ํ•˜๋‚˜๋งŒ ํ—ˆ์šฉํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ด ๋ฐฉ์–ด ๊ธฐ๋Šฅ์€ ๋น„์Šทํ•˜๊ฒŒ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค.
๋กœ๊ทธํŠธ ๋ฐ”์ด์–ด์Šค XOR ๋กœ๊ทธํ”„๋กญ.์ ๋Œ€์ž๊ฐ€ ๋กœ๊ทธํŠธ ๋ฐ”์ด์–ด์Šค์™€ ์ถœ๋ ฅ ๋กœ๊ทธํ”„๋กญ์„ ๋ชจ๋‘ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ ๊ณต๊ฒฉ ๋น„์šฉ์€ \(10\๋ฐฐ\) ๋” ์ €๋ ดํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ž์—ฐ์Šค๋Ÿฌ์šด ์™„ํ™”์ฑ…์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค: ๋กœ๊ทธํŠธ ๋ฐ”์ด์–ด์Šค์™€ ๋กœ๊ทธํ”„๋กญ์„ ๋™์‹œ์— ์‚ฌ์šฉํ•˜๋Š” ์ฟผ๋ฆฌ๋ฅผ API์— ๊ธˆ์ง€ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ๋ฐฉ์–ด๋Š” ๋ณด์•ˆ ๋ฐ ๋จธ์‹  ๋Ÿฌ๋‹ ์ปค๋ฎค๋‹ˆํ‹ฐ ๋ชจ๋‘์—์„œ ํ”ํžˆ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 2023๋…„ OpenAI๋Š” ์—์ฝ”์™€ ๋กœ๊ทธํ”„๋กญ์„ ๋ชจ๋‘ ๊ฒฐํ•ฉํ•˜๋Š” ๊ธฐ๋Šฅ์„ ์ œ๊ฑฐํ–ˆ์ง€๋งŒ, ๋‘˜ ์ค‘ ํ•˜๋‚˜๋งŒ ํ—ˆ์šฉํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ด ๋ฐฉ์–ด ๊ธฐ๋Šฅ์„ ์‚ฌ์šฉํ–ˆ๋Š”๋ฐ, ์ด๋Š” ์ด ๋ฐฉ์–ด ๊ธฐ๋Šฅ๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค.
Unlike most studies (Yoran et al., 2023) (Nakano et al., 2021) that directly utilise commercial search engine APIs. Komeili et al. (2021) propose an alternative approach to using multiple commercial search engine APIs. They suggest using the Bing Search API to generate a list of URLs for each query. These URLs are then used as keys to look up their page content in a lookup table constructed from public crawl snapshots, which populates a set of pages for that query. In addition, the evaluation takes into account whether the URL is from the English Wikipedia. If so, the page title is extracted from the URL and the corresponding page is searched for in the Wikipedia dump.
์ƒ์—…์šฉ ๊ฒ€์ƒ‰ ์—”์ง„ API๋ฅผ ์ง์ ‘ ํ™œ์šฉํ•˜๋Š” ๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ(Yoran ์™ธ., 2023) (Nakano ์™ธ., 2021)์™€ ๋‹ฌ๋ฆฌ, Komeili ์™ธ.(2021)๋Š” ์—ฌ๋Ÿฌ ์ƒ์—…์šฉ ๊ฒ€์ƒ‰ ์—”์ง„ API๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์ฒด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Bing ๊ฒ€์ƒ‰ API๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ URL ๋ชฉ๋ก์„ ์ƒ์„ฑํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ URL์€ ๊ณต๊ฐœ ํฌ๋กค๋ง ์Šค๋ƒ…์ƒท์—์„œ ์ƒ์„ฑ๋œ ์กฐํšŒ ํ…Œ์ด๋ธ”์—์„œ ํŽ˜์ด์ง€ ์ฝ˜ํ…์ธ ๋ฅผ ์กฐํšŒํ•˜๋Š” ํ‚ค๋กœ ์‚ฌ์šฉ๋˜๋ฉฐ, ํ•ด๋‹น ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ ํŽ˜์ด์ง€ ์ง‘ํ•ฉ์„ ์ฑ„์›๋‹ˆ๋‹ค. ๋˜ํ•œ ํ‰๊ฐ€์—๋Š” URL์ด ์˜์–ด ์œ„ํ‚ค๋ฐฑ๊ณผ์—์„œ ๊ฐ€์ ธ์˜จ ๊ฒƒ์ธ์ง€ ์—ฌ๋ถ€๋„ ๊ณ ๋ ค๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋ ‡๋‹ค๋ฉด URL์—์„œ ํŽ˜์ด์ง€ ์ œ๋ชฉ์„ ์ถ”์ถœํ•˜๊ณ  ํ•ด๋‹น ํŽ˜์ด์ง€๋ฅผ ์œ„ํ‚ค๋ฐฑ๊ณผ ๋คํ”„์—์„œ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ(Yoran ์™ธ, 2023)(Nakano ์™ธ, 2021)๊ฐ€ ์ƒ์šฉ ๊ฒ€์ƒ‰ ์—”์ง„ API๋ฅผ ์ง์ ‘ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ๊ณผ ๋‹ฌ๋ฆฌ, Komeili ์™ธ(2021)๋Š” ์—ฌ๋Ÿฌ ์ƒ์šฉ ๊ฒ€์ƒ‰ ์—”์ง„ API๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์ฒด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ Bing ๊ฒ€์ƒ‰ API๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ URL ๋ชฉ๋ก์„ ์ƒ์„ฑํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ URL์€ ๊ณต๊ฐœ ํฌ๋กค๋ง ์Šค๋ƒ…์ƒท์—์„œ ์ƒ์„ฑ๋œ ์กฐํšŒ ํ…Œ์ด๋ธ”์—์„œ ํŽ˜์ด์ง€ ์ฝ˜ํ…์ธ ๋ฅผ ์กฐํšŒํ•˜๋Š” ๋ฐ ํ‚ค๋กœ ์‚ฌ์šฉ๋˜๋ฉฐ, ํ•ด๋‹น ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ ํŽ˜์ด์ง€ ์ง‘ํ•ฉ์„ ์ฑ„์›๋‹ˆ๋‹ค. ๋˜ํ•œ ํ‰๊ฐ€์—๋Š” URL์ด ์˜์–ด ์œ„ํ‚ค๋ฐฑ๊ณผ์—์„œ ๊ฐ€์ ธ์˜จ ๊ฒƒ์ธ์ง€ ์—ฌ๋ถ€๋ฅผ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ ‡๋‹ค๋ฉด URL์—์„œ ํŽ˜์ด์ง€ ์ œ๋ชฉ์„ ์ถ”์ถœํ•˜๊ณ  ํ•ด๋‹น ํŽ˜์ด์ง€๋ฅผ ์œ„ํ‚ค๋ฐฑ๊ณผ ๋คํ”„์—์„œ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ(Yoran et al., 2023) (Nakano et al., 2021)๊ฐ€ ์ƒ์šฉ ๊ฒ€์ƒ‰ ์—”์ง„ API๋ฅผ ์ง์ ‘ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ๊ณผ ๋‹ฌ๋ฆฌ, Komeili et al.(2021)์€ ์—ฌ๋Ÿฌ ์ƒ์šฉ ๊ฒ€์ƒ‰ ์—”์ง„ API๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์ฒด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ Bing Search API๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ URL ๋ชฉ๋ก์„ ์ƒ์„ฑํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ URL์€ ํ•ด๋‹น ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ ํŽ˜์ด์ง€ ์ง‘ํ•ฉ์„ ์ฑ„์šฐ๋Š” ๊ณต๊ฐœ ํฌ๋กค๋ง ์Šค๋ƒ…์ƒท์œผ๋กœ ๊ตฌ์„ฑ๋œ ๋ฃฉ์—… ํ…Œ์ด๋ธ”์—์„œ ํŽ˜์ด์ง€ ์ฝ˜ํ…์ธ ๋ฅผ ์กฐํšŒํ•˜๊ธฐ ์œ„ํ•œ ํ‚ค๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํ‰๊ฐ€์—๋Š” URL์ด ์˜์–ด ์œ„ํ‚ค๋ฐฑ๊ณผ์—์„œ ๋‚˜์˜จ ๊ฒƒ์ธ์ง€ ์—ฌ๋ถ€๋ฅผ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ ‡๋‹ค๋ฉด URL์—์„œ ํŽ˜์ด์ง€ ์ œ๋ชฉ์„ ์ถ”์ถœํ•˜๊ณ  ํ•ด๋‹น ํŽ˜์ด์ง€๋ฅผ ์œ„ํ‚ค๋ฐฑ๊ณผ ๋คํ”„์—์„œ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ(Yoran et al., 2023)(Nakano et al., 2021)๊ฐ€ ์ƒ์—…์šฉ ๊ฒ€์ƒ‰ ์—”์ง„ API๋ฅผ ์ง์ ‘ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ๊ณผ ๋‹ฌ๋ฆฌ, Komeili ๋“ฑ(2021)์€ ์—ฌ๋Ÿฌ ์ƒ์—…์šฉ ๊ฒ€์ƒ‰ ์—”์ง„ API๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์ฒด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ Bing ๊ฒ€์ƒ‰ API๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ URL ๋ชฉ๋ก์„ ์ƒ์„ฑํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ URL์„ ํ‚ค๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐœ ํฌ๋กค๋ง ์Šค๋ƒ…์ƒท์—์„œ ๊ตฌ์ถ•๋œ ์กฐํšŒ ํ…Œ์ด๋ธ”์—์„œ ํŽ˜์ด์ง€ ์ฝ˜ํ…์ธ ๋ฅผ ์กฐํšŒํ•˜์—ฌ ํ•ด๋‹น ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ ํŽ˜์ด์ง€ ์ง‘ํ•ฉ์„ ์ฑ„์›๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด ํ‰๊ฐ€๋Š” URL์ด ์˜์–ด ์œ„ํ‚ค๋ฐฑ๊ณผ์—์„œ ๊ฐ€์ ธ์˜จ ๊ฒƒ์ธ์ง€ ์—ฌ๋ถ€๋ฅผ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ ‡๋‹ค๋ฉด URL์—์„œ ํŽ˜์ด์ง€ ์ œ๋ชฉ์„ ์ถ”์ถœํ•˜๊ณ  ํ•ด๋‹น ํŽ˜์ด์ง€๋ฅผ ์œ„ํ‚ค๋ฐฑ๊ณผ ๋คํ”„์—์„œ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ(Yoran et al., 2023)(Nakano et al., 2021)๊ฐ€ ์ƒ์šฉ ๊ฒ€์ƒ‰ ์—”์ง„ API๋ฅผ ์ง์ ‘ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ๊ณผ ๋‹ฌ๋ฆฌ, Komeili et al. (2021)์€ ์—ฌ๋Ÿฌ ์ƒ์šฉ ๊ฒ€์ƒ‰ ์—”์ง„ API๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์ฒด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ Bing ๊ฒ€์ƒ‰ API๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ URL ๋ชฉ๋ก์„ ์ƒ์„ฑํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ URL์„ ํ‚ค๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๊ณต๊ฐœ ํฌ๋กค๋ง ์Šค๋ƒ…์ƒท์—์„œ ๊ตฌ์„ฑ๋œ ์กฐํšŒ ํ…Œ์ด๋ธ”์—์„œ ํŽ˜์ด์ง€ ์ฝ˜ํ…์ธ ๋ฅผ ์กฐํšŒํ•˜์—ฌ ํ•ด๋‹น ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ ํŽ˜์ด์ง€ ์ง‘ํ•ฉ์„ ์ฑ„์›๋‹ˆ๋‹ค. ๋˜ํ•œ ํ‰๊ฐ€์—๋Š” URL์ด ์˜์–ด Wikipedia์—์„œ ๊ฐ€์ ธ์˜จ ๊ฒƒ์ธ์ง€ ์—ฌ๋ถ€๊ฐ€ ๊ณ ๋ ค๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋ ‡๋‹ค๋ฉด URL์—์„œ ํŽ˜์ด์ง€ ์ œ๋ชฉ์„ ์ถ”์ถœํ•˜๊ณ  ํ•ด๋‹น ํŽ˜์ด์ง€๋ฅผ Wikipedia ๋คํ”„์—์„œ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค.
Our findings place domain continual pre-training as a strong alternative to building domain-specific LLMs from scratch. By being smarter about data selection for continual pre-training, we can surpass vanilla continual pre-training at a fraction of the cost. Overall, our work paves the way for developing domain-specific LLMs at a reduced cost, with implications for a wide range of applications.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๋„๋ฉ”์ธ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ๋„๋ฉ”์ธ๋ณ„ LLM์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์— ๋Œ€ํ•œ ๊ฐ•๋ ฅํ•œ ๋Œ€์•ˆ์œผ๋กœ ์ž๋ฆฌ๋งค๊น€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ์— ๋Œ€ํ•ด ๋” ์Šค๋งˆํŠธํ•˜๊ฒŒ ์ ‘๊ทผํ•˜๋ฉด ํ›จ์”ฌ ์ ์€ ๋น„์šฉ์œผ๋กœ ๋ฐ”๋‹๋ผ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ, ์ €ํฌ์˜ ์—ฐ๊ตฌ๋Š” ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ๋น„์šฉ ์ ˆ๊ฐ์œผ๋กœ ๋„๋ฉ”์ธ๋ณ„ LLM์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๋Š” ๊ธธ์„ ์—ด์—ˆ์Šต๋‹ˆ๋‹ค.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๋„๋ฉ”์ธ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ๋„๋ฉ”์ธ๋ณ„ LLM์„ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์— ๊ฐ•๋ ฅํ•œ ๋Œ€์•ˆ์œผ๋กœ ์ž๋ฆฌ๋งค๊น€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ์— ๋Œ€ํ•ด ๋” ์Šค๋งˆํŠธํ•˜๊ฒŒ ํ•˜๋ฉด ํ›จ์”ฌ ์ ์€ ๋น„์šฉ์œผ๋กœ ๋ฐ”๋‹๋ผ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ, ์ €ํฌ์˜ ์—ฐ๊ตฌ๋Š” ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋Œ€ํ•œ ์‹œ์‚ฌ์ ์„ ๋‹ด์€ ๋น„์šฉ ์ ˆ๊ฐ์œผ๋กœ ๋„๋ฉ”์ธ๋ณ„ LLM์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๋Š” ๊ธธ์„ ์—ด์–ด์ค๋‹ˆ๋‹ค.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๋„๋ฉ”์ธ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ๋„๋ฉ”์ธ๋ณ„ LLM์„ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์— ๋Œ€ํ•œ ๊ฐ•๋ ฅํ•œ ๋Œ€์•ˆ์œผ๋กœ ์ž๋ฆฌ๋งค๊น€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ์— ๋Œ€ํ•ด ๋” ์Šค๋งˆํŠธํ•˜๊ฒŒ ์ ‘๊ทผํ•˜๋ฉด ํ›จ์”ฌ ์ ์€ ๋น„์šฉ์œผ๋กœ ๋ฐ”๋‹๋ผ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ, ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋น„์šฉ ์ ˆ๊ฐ์œผ๋กœ ๋„๋ฉ”์ธ๋ณ„ LLM์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๋Š” ๊ธธ์„ ์—ด์–ด์ค๋‹ˆ๋‹ค.
์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋Š” ๋„๋ฉ”์ธ๋ณ„ LLM์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์— ๋Œ€ํ•œ ๊ฐ•๋ ฅํ•œ ๋Œ€์•ˆ์œผ๋กœ ๋„๋ฉ”์ธ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ์— ๋” ํ˜„๋ช…ํ•ด์ง์œผ๋กœ์จ, ์šฐ๋ฆฌ๋Š” ์ ์€ ๋น„์šฉ์œผ๋กœ ๋ฐ”๋‹๋ผ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ, ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๊ด‘๋ฒ”์œ„ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋Œ€ํ•œ ์‹œ์‚ฌ์ ๊ณผ ํ•จ๊ป˜, ๋น„์šฉ์„ ์ ˆ๊ฐํ•˜๋ฉด์„œ ๋„๋ฉ”์ธ๋ณ„ LLM์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๋Š” ๊ธธ์„ ์—ด์–ด์ค๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ๋„๋ฉ”์ธ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ๋„๋ฉ”์ธ๋ณ„ LLM์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฐ•๋ ฅํ•œ ๋Œ€์•ˆ์œผ๋กœ ์ž๋ฆฌ๋งค๊น€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์„ ํƒ์— ๋” ํ˜„๋ช…ํ•ด์ง์œผ๋กœ์จ, ์šฐ๋ฆฌ๋Š” ์ ์€ ๋น„์šฉ์œผ๋กœ ๋ฐ”๋‹๋ผ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ, ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๋Š” ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋Œ€ํ•œ ์‹œ์‚ฌ์ ๊ณผ ํ•จ๊ป˜, ๋น„์šฉ์„ ์ ˆ๊ฐํ•˜๋ฉด์„œ ๋„๋ฉ”์ธ๋ณ„ LLM์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๋Š” ๊ธธ์„ ์—ด์–ด์ค๋‹ˆ๋‹ค.
In this part of our discussion, we pay attention to an in-depth analysis of individual models, focusing on identifying whether the trends in their performance remain consistent across various metrics and observing the unique patterns that emerge across different benchmarks.
์ด ๋ถ€๋ถ„์—์„œ๋Š” ๊ฐœ๋ณ„ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹ฌ์ธต ๋ถ„์„์— ์ฃผ๋ชฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ง€ํ‘œ์—์„œ ์„ฑ๋Šฅ ์ถ”์„ธ๊ฐ€ ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•˜๋Š”์ง€ ํŒŒ์•…ํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋‚˜ํƒ€๋‚˜๋Š” ๊ณ ์œ ํ•œ ํŒจํ„ด์„ ๊ด€์ฐฐํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค.
์ด ๋ถ€๋ถ„์—์„œ๋Š” ๊ฐœ๋ณ„ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹ฌ์ธต ๋ถ„์„์— ์ฃผ๋ชฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ง€ํ‘œ์—์„œ ์„ฑ๋Šฅ ์ถ”์„ธ๊ฐ€ ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•˜๋Š”์ง€ ํŒŒ์•…ํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋‚˜ํƒ€๋‚˜๋Š” ๊ณ ์œ ํ•œ ํŒจํ„ด์„ ๊ด€์ฐฐํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค.
์ด๋ฒˆ ํ† ๋ก ์—์„œ๋Š” ๊ฐœ๋ณ„ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹ฌ์ธต ๋ถ„์„์— ์ฃผ๋ชฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ง€ํ‘œ์—์„œ ์„ฑ๋Šฅ ์ถ”์„ธ๊ฐ€ ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•˜๋Š”์ง€ ํŒŒ์•…ํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋‚˜ํƒ€๋‚˜๋Š” ๊ณ ์œ ํ•œ ํŒจํ„ด์„ ๊ด€์ฐฐํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค.
์ด ๋ถ€๋ถ„์—์„œ๋Š” ๊ฐœ๋ณ„ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹ฌ์ธต ๋ถ„์„์— ์ง‘์ค‘ํ•˜์—ฌ ์„ฑ๋Šฅ ์ถ”์„ธ๊ฐ€ ๋‹ค์–‘ํ•œ ์ง€ํ‘œ์— ๊ฑธ์ณ ์ผ๊ด€๋˜๊ฒŒ ์œ ์ง€๋˜๋Š”์ง€ ํ™•์ธํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋‚˜ํƒ€๋‚˜๋Š” ๊ณ ์œ ํ•œ ํŒจํ„ด์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค.
์ด ๋ถ€๋ถ„์—์„œ๋Š” ๊ฐœ๋ณ„ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์‹ฌ์ธต ๋ถ„์„์— ์ดˆ์ ์„ ๋งž์ถฐ ๋‹ค์–‘ํ•œ ์ง€ํ‘œ์—์„œ ์„ฑ๊ณผ ์ถ”์„ธ๊ฐ€ ์ผ๊ด€๋˜๊ฒŒ ์œ ์ง€๋˜๋Š”์ง€ ํ™•์ธํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋‚˜ํƒ€๋‚˜๋Š” ๊ณ ์œ ํ•œ ํŒจํ„ด์„ ๊ด€์ฐฐํ•˜๋Š” ๋ฐ ์ฃผ๋ชฉํ•ฉ๋‹ˆ๋‹ค.
**Knowledge extraction.** The "2bit/param" result is not about word-by-word memorization. Even better, such knowledge is also flexibly extractable (e.g., via fine-tuning using QAs like "What is Anya Forger's birthday?") [3] and thus can be further manipulated in downstream tasks (such as comparing the birthdates of two people, or performing calculations on the retrieved knowledge, etc.) [4]. This is because our \(\mathsf{bioS}(N)\) data is knowledge-augmented: the English biographies have sufficient text diversities [3]. We also verify in Appendix A.2 that such knowledge is extractable.
**์ง€์‹ ์ถ”์ถœ.** "2bit/param" ๊ฒฐ๊ณผ๋Š” ๋‹จ์–ด๋ณ„ ์•”๊ธฐ์— ๊ด€ํ•œ ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ๋” ๋‚˜์€ ์ ์€ ์ด๋Ÿฌํ•œ ์ง€์‹์€ ์œ ์—ฐํ•˜๊ฒŒ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: "์•„๋ƒ ํฌ๊ฑฐ์˜ ์ƒ์ผ์€ ์–ธ์ œ์ธ๊ฐ€์š”?"์™€ ๊ฐ™์€ QA๋ฅผ ์‚ฌ์šฉํ•œ ๋ฏธ์„ธ ์กฐ์ •) [3] ๋ฐ ๋”ฐ๋ผ์„œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…(์˜ˆ: ๋‘ ์‚ฌ๋žŒ์˜ ์ƒ์ผ์„ ๋น„๊ตํ•˜๊ฑฐ๋‚˜ ๊ฒ€์ƒ‰๋œ ์ง€์‹์— ๋Œ€ํ•œ ๊ณ„์‚ฐ ์ˆ˜ํ–‰ ๋“ฑ) [4]์—์„œ ์ถ”๊ฐ€๋กœ ์กฐ์ž‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์šฐ๋ฆฌ์˜ \(\mathsf{bioS}(N)\) ๋ฐ์ดํ„ฐ๊ฐ€ ์ง€์‹์ด ์ฆ๊ฐ•๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค: ์˜์–ด ์ „๊ธฐ์—๋Š” ์ถฉ๋ถ„ํ•œ ํ…์ŠคํŠธ ๋‹ค์–‘์„ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค [3]. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ์ง€์‹์ด ์ถ”์ถœ ๊ฐ€๋Šฅํ•œ์ง€์— ๋Œ€ํ•ด ๋ถ€๋ก A.2์—์„œ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค.
**์ง€์‹ ์ถ”์ถœ.** "2๋น„ํŠธ/ํŒŒ๋žŒ" ๊ฒฐ๊ณผ๋Š” ๋‹จ์–ด ๋‹จ์œ„๋กœ ์•”๊ธฐํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ๋” ๋‚˜์€ ๋ฐฉ๋ฒ•์œผ๋กœ, ์ด๋Ÿฌํ•œ ์ง€์‹์€ ์œ ์—ฐํ•˜๊ฒŒ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: "์•„๋ƒ ํฌ์ €์˜ ์ƒ์ผ์€ ์–ธ์ œ์ธ๊ฐ€์š”?"์™€ ๊ฐ™์€ QA๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏธ์„ธ ์กฐ์ •) [3] ๋ฐ ๋”ฐ๋ผ์„œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…(์˜ˆ: ๋‘ ์‚ฌ๋žŒ์˜ ์ƒ๋…„์›”์ผ์„ ๋น„๊ตํ•˜๊ฑฐ๋‚˜ ๊ฒ€์ƒ‰๋œ ์ง€์‹์— ๋Œ€ํ•œ ๊ณ„์‚ฐ ์ˆ˜ํ–‰ ๋“ฑ) [4]์—์„œ ์ถ”๊ฐ€๋กœ ์กฐ์ž‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” \(\mathsf{bioS}(N)\) ๋ฐ์ดํ„ฐ๊ฐ€ ์ง€์‹์ด ์ฆ๊ฐ•๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค: ์˜์–ด ์ „๊ธฐ์—๋Š” ์ถฉ๋ถ„ํ•œ ํ…์ŠคํŠธ ๋‹ค์–‘์„ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค [3]. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ์ง€์‹์ด ์ถ”์ถœ ๊ฐ€๋Šฅํ•œ์ง€๋Š” ๋ถ€๋ก A.2์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์ง€์‹ ์ถ”์ถœ.** "2๋น„ํŠธ/ํŒŒ๋ผ๋ฏธํ„ฐ" ๊ฒฐ๊ณผ๋Š” ๋‹จ์–ด ๋‹จ์œ„๋กœ ์•”๊ธฐํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ๋” ๋‚˜์•„๊ฐ€, ์ด๋Ÿฌํ•œ ์ง€์‹์€ ์œ ์—ฐํ•˜๊ฒŒ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ(์˜ˆ: "์•„๋ƒ ํฌ์ €์˜ ์ƒ์ผ์€ ์–ธ์ œ์ธ๊ฐ€์š”?"์™€ ๊ฐ™์€ QA๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏธ์„ธ ์กฐ์ •) ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…(์˜ˆ: ๋‘ ์‚ฌ๋žŒ์˜ ์ƒ์ผ์„ ๋น„๊ตํ•˜๊ฑฐ๋‚˜ ๊ฒ€์ƒ‰๋œ ์ง€์‹์— ๋Œ€ํ•œ ๊ณ„์‚ฐ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋“ฑ)์—์„œ ์ถ”๊ฐ€๋กœ ์กฐ์ž‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. [4]. ์ด๋Š” ์šฐ๋ฆฌ์˜ \(\mathsf{bioS}(N)\) ๋ฐ์ดํ„ฐ๊ฐ€ ์ง€์‹์ด ๊ฐ•ํ™”๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค: ์˜์–ด ์ „๊ธฐ๊ฐ€ ์ถฉ๋ถ„ํ•œ ํ…์ŠคํŠธ ๋‹ค์–‘์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค [3]. ๋˜ํ•œ ๋ถ€๋ก A.2์—์„œ ์ด๋Ÿฌํ•œ ์ง€์‹์ด ์ถ”์ถœ ๊ฐ€๋Šฅํ•œ์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค.
**์ง€์‹ ์ถ”์ถœ.** "2bit/param" ๊ฒฐ๊ณผ๋Š” ๋‹จ์–ด ๋‹จ์œ„ ์•”๊ธฐ์— ๊ด€ํ•œ ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ์‹ฌ์ง€์–ด ๋” ๋‚˜์•„๊ฐ€ ์ด๋Ÿฌํ•œ ์ง€์‹์€ ์œ ์—ฐํ•˜๊ฒŒ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: "์•„๋‚˜ ํฌ์ €์˜ ์ƒ์ผ์€ ์–ธ์ œ์ธ๊ฐ€์š”?"์™€ ๊ฐ™์€ QA๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏธ์„ธ ์กฐ์ •). [3] ๋”ฐ๋ผ์„œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ๋” ์กฐ์ž‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ๋‘ ์‚ฌ๋žŒ์˜ ์ƒ๋…„์›”์ผ์„ ๋น„๊ตํ•˜๊ฑฐ๋‚˜ ๊ฒ€์ƒ‰๋œ ์ง€์‹์œผ๋กœ ๊ณ„์‚ฐ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋“ฑ). [4] ์ด๋Š” ์šฐ๋ฆฌ์˜ \(\mathsf{bioS}(N)\) ๋ฐ์ดํ„ฐ๊ฐ€ ์ง€์‹์ด ๋ณด๊ฐ•๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์˜์–ด ์ „๊ธฐ๋Š” ์ถฉ๋ถ„ํ•œ ํ…์ŠคํŠธ ๋‹ค์–‘์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. [3]. ๋˜ํ•œ ๋ถ€๋ก A.2์—์„œ ์ด๋Ÿฌํ•œ ์ง€์‹์„ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ์Œ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค.
**์ง€์‹ ์ถ”์ถœ. "2bit/param" ๊ฒฐ๊ณผ๋Š” ๋‹จ์–ด ๋‹จ์œ„ ์•”๊ธฐ์— ๊ด€ํ•œ ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ์‹ฌ์ง€์–ด ๋” ๋‚˜์•„๊ฐ€ ์ด๋Ÿฌํ•œ ์ง€์‹์€ ์œ ์—ฐํ•˜๊ฒŒ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: "์•„๋‚˜ ํฌ์ €์˜ ์ƒ์ผ์€ ์–ธ์ œ์ธ๊ฐ€์š”?"์™€ ๊ฐ™์€ QA๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏธ์„ธ ์กฐ์ •). [3] ๋”ฐ๋ผ์„œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ๋” ์กฐ์ž‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ๋‘ ์‚ฌ๋žŒ์˜ ์ƒ๋…„์›”์ผ์„ ๋น„๊ตํ•˜๊ฑฐ๋‚˜ ๊ฒ€์ƒ‰๋œ ์ง€์‹์œผ๋กœ ๊ณ„์‚ฐ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋“ฑ). [4] ์ด๋Š” ์˜์–ด ์ „๊ธฐ๊ฐ€ ์ถฉ๋ถ„ํ•œ ํ…์ŠคํŠธ ๋‹ค์–‘์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. [3] ๋˜ํ•œ ๋ถ€๋ก A.2์—์„œ ์ด๋Ÿฌํ•œ ์ง€์‹์„ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ์Œ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค.
This becomes a dynamic data schedule as the number of training stages approaches the infinite limit. Therefore, the successful application of our data mixing laws on continual training signifies a promising prospect for using it to design dynamic data schedules, a more comprehensive data curating paradigm.
์ด๋Š” ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ์ˆ˜๊ฐ€ ๋ฌดํ•œํ•œ ๊ทนํ•œ์— ๊ฐ€๊นŒ์›Œ์งˆ์ˆ˜๋ก ๋™์  ๋ฐ์ดํ„ฐ ์Šค์ผ€์ค„์ด ๋˜๋Š” ๊ฒƒ์ด๋‹ค. ๋”ฐ๋ผ์„œ, ์ง€์†์ ์ธ ํ›ˆ๋ จ์— ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ ๋ฏน์‹ฑ ๋ฒ•์น™์˜ ์„ฑ๊ณต์ ์ธ ์ ์šฉ์€ ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜ ํŒจ๋Ÿฌ๋‹ค์ž„์ธ ๋™์  ๋ฐ์ดํ„ฐ ์Šค์ผ€์ค„์„ ์„ค๊ณ„ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•˜๋Š” ๋ฐ ์žˆ์–ด ์œ ๋งํ•œ ์ „๋ง์„ ์˜๋ฏธํ•œ๋‹ค.
์ด๋Š” ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ์ˆ˜๊ฐ€ ๋ฌดํ•œํ•œ ๊ทนํ•œ์— ๊ฐ€๊นŒ์›Œ์งˆ์ˆ˜๋ก ๋™์  ๋ฐ์ดํ„ฐ ์Šค์ผ€์ค„์ด ๋œ๋‹ค. ๋”ฐ๋ผ์„œ ์ง€์†์ ์ธ ํ›ˆ๋ จ์— ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ ๋ฏน์‹ฑ ๋ฒ•์น™์˜ ์„ฑ๊ณต์ ์ธ ์ ์šฉ์€ ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ดํŒ… ํŒจ๋Ÿฌ๋‹ค์ž„์ธ ๋™์  ๋ฐ์ดํ„ฐ ์Šค์ผ€์ค„์„ ์„ค๊ณ„ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•˜๋Š” ๋ฐ ์žˆ์–ด ์œ ๋งํ•œ ์ „๋ง์„ ์˜๋ฏธํ•œ๋‹ค.
์ด๋Š” ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ์ˆ˜๊ฐ€ ๋ฌดํ•œํ•œ ๊ทนํ•œ์— ๊ฐ€๊นŒ์›Œ์ง์— ๋”ฐ๋ผ ๋™์  ๋ฐ์ดํ„ฐ ์Šค์ผ€์ค„์ด ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ง€์†์ ์ธ ํ›ˆ๋ จ์— ๋Œ€ํ•œ ์šฐ๋ฆฌ์˜ ๋ฐ์ดํ„ฐ ๋ฏน์‹ฑ ๋ฒ•์น™์˜ ์„ฑ๊ณต์ ์ธ ์ ์šฉ์€ ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ดํŒ… ํŒจ๋Ÿฌ๋‹ค์ž„์ธ ๋™์  ๋ฐ์ดํ„ฐ ์Šค์ผ€์ค„์„ ์„ค๊ณ„ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•˜๋Š” ๋ฐ ์œ ๋งํ•œ ์ „๋ง์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
์ด๋Š” ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ์ˆ˜๊ฐ€ ๋ฌดํ•œํ•œ ํ•œ๊ณ„์— ๊ฐ€๊นŒ์›Œ์งˆ์ˆ˜๋ก ๋™์  ๋ฐ์ดํ„ฐ ์ผ์ •์ด ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ง€์†์ ์ธ ํ•™์Šต์— ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์„ ์„ฑ๊ณต์ ์œผ๋กœ ์ ์šฉํ•˜๋ฉด ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ดํŒ… ํŒจ๋Ÿฌ๋‹ค์ž„์ธ ๋™์  ๋ฐ์ดํ„ฐ ์ผ์ •์„ ์„ค๊ณ„ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์œ ๋งํ•œ ์ „๋ง์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
์ด๋Š” ํ›ˆ๋ จ ๋‹จ๊ณ„์˜ ์ˆ˜๊ฐ€ ๋ฌดํ•œํ•œ ํ•œ๊ณ„์— ๊ฐ€๊นŒ์›Œ์งˆ์ˆ˜๋ก ๋™์  ๋ฐ์ดํ„ฐ ์ผ์ •์ด ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ง€์†์ ์ธ ํ›ˆ๋ จ์— ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฒ•์น™์„ ์„ฑ๊ณต์ ์œผ๋กœ ์ ์šฉํ•˜๋ฉด ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ดํŒ… ํŒจ๋Ÿฌ๋‹ค์ž„์ธ ๋™์  ๋ฐ์ดํ„ฐ ์ผ์ •์„ ์„ค๊ณ„ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์œ ๋งํ•œ ์ „๋ง์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
We now compare infinite learning rate schedules to a cosine decay schedule. We first explore a simple single-dataset pre-training setup to evaluate the feasibility of the schedule for LLM pre-training. Subsequently, we explore its benefits in our _three datasets, no shift_ setting.
์ด์ œ ๋ฌดํ•œ ํ•™์Šต ์†๋„ ์Šค์ผ€์ค„์„ ์ฝ”์‚ฌ์ธ ๊ฐ์‡  ์Šค์ผ€์ค„๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € LLM ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ์Šค์ผ€์ค„์˜ ํƒ€๋‹น์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ„๋‹จํ•œ ๋‹จ์ผ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๊ทธ ๋‹ค์Œ์—๋Š” _์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ, ์‹œํ”„ํŠธ ์—†์Œ_ ์„ค์ •์—์„œ ๊ทธ ์ด์ ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.
์ด์ œ ๋ฌดํ•œ ํ•™์Šต ์†๋„ ์Šค์ผ€์ค„์„ ์ฝ”์‚ฌ์ธ ๊ฐ์‡  ์Šค์ผ€์ค„๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € LLM ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ์Šค์ผ€์ค„์˜ ์‹คํ˜„ ๊ฐ€๋Šฅ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ„๋‹จํ•œ ๋‹จ์ผ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๊ทธ ๋‹ค์Œ์—๋Š” _์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ, ์ด๋™ ์—†์Œ_ ์„ค์ •์—์„œ ๊ทธ ์ด์ ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.
์ด์ œ ๋ฌดํ•œ ํ•™์Šต ์†๋„ ์Šค์ผ€์ค„์„ ์ฝ”์‚ฌ์ธ ๊ฐ์‡  ์Šค์ผ€์ค„๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ €, LLM ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ์Šค์ผ€์ค„์˜ ์‹คํ˜„ ๊ฐ€๋Šฅ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋‹จ์ˆœํ•œ ๋‹จ์ผ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๊ทธ ๋‹ค์Œ์—๋Š” _3๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ, ์ด๋™ ์—†์Œ_ ์„ค์ •์—์„œ ์ด ์Šค์ผ€์ค„์˜ ์ด์ ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.
์ด์ œ ๋ฌดํ•œ ํ•™์Šต ์†๋„ ์Šค์ผ€์ค„์„ ์ฝ”์‚ฌ์ธ ๊ฐ์†Œ ์Šค์ผ€์ค„๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ๊ฐ„๋‹จํ•œ ๋‹จ์ผ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์„ ์‚ดํŽด๋ณด๊ณ  LLM ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ์Šค์ผ€์ค„์˜ ํƒ€๋‹น์„ฑ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ํ›„, _์„ธ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ, ์ด๋™ ์—†์Œ_ ์„ค์ •์—์„œ ๊ทธ ์ด์ ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.
์ด์ œ ๋ฌดํ•œ ํ•™์Šต ์†๋„ ์Šค์ผ€์ค„์„ ์ฝ”์‚ฌ์ธ ๊ฐ์‡  ์Šค์ผ€์ค„๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ € ๊ฐ„๋‹จํ•œ ๋‹จ์ผ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์‚ฌ์ „ ํ•™์Šต ์„ค์ •์„ ์‚ดํŽด๋ณด๊ณ  LLM ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ์Šค์ผ€์ค„์˜ ํƒ€๋‹น์„ฑ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ํ›„, _์„ธ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ, ์‹œํ”„ํŠธ ์—†์Œ_ ์„ค์ •์—์„œ ๊ทธ ์ด์ ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.
Agents are designed to extend language model capabilities to solve real-world challenges. Successful implementations require robust problem-solving capabilities enabling agents to perform well on novel tasks. To solve real-world problems effectively, agents require the ability to reason and plan as well as call tools that interact with an external environment. In this section we explore why reasoning, planning, and tool calling are critical to agent success.
์—์ด์ „ํŠธ๋Š” ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ธฐ๋Šฅ์„ ํ™•์žฅํ•˜์—ฌ ์‹ค์ œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์„ฑ๊ณต์ ์ธ ๊ตฌํ˜„์„ ์œ„ํ•ด์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ƒˆ๋กœ์šด ์ž‘์—…์„ ์ž˜ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ๊ฐ•๋ ฅํ•œ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ๋ฌธ์ œ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๋ ค๋ฉด ์—์ด์ „ํŠธ๋Š” ์ถ”๋ก ํ•˜๊ณ  ๊ณ„ํšํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์™ธ๋ถ€ ํ™˜๊ฒฝ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ถ”๋ก , ๊ณ„ํš ๋ฐ ๋„๊ตฌ ํ˜ธ์ถœ์ด ์—์ด์ „ํŠธ์˜ ์„ฑ๊ณต์— ์ค‘์š”ํ•œ ์ด์œ ๋ฅผ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.
์—์ด์ „ํŠธ๋Š” ์–ธ์–ด ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํ™•์žฅํ•˜์—ฌ ์‹ค์ œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์„ฑ๊ณต์ ์ธ ๊ตฌํ˜„์„ ์œ„ํ•ด์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ƒˆ๋กœ์šด ์ž‘์—…์„ ์ž˜ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ๊ฐ•๋ ฅํ•œ ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ๋ฌธ์ œ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๋ ค๋ฉด ์—์ด์ „ํŠธ๊ฐ€ ์ถ”๋ก ํ•˜๊ณ  ๊ณ„ํšํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์™ธ๋ถ€ ํ™˜๊ฒฝ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ถ”๋ก , ๊ณ„ํš ๋ฐ ๋„๊ตฌ ํ˜ธ์ถœ์ด ์—์ด์ „ํŠธ ์„ฑ๊ณต์— ์ค‘์š”ํ•œ ์ด์œ ๋ฅผ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.
์—์ด์ „ํŠธ๋Š” ์‹ค์ œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์–ธ์–ด ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํ™•์žฅํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์„ฑ๊ณต์ ์ธ ๊ตฌํ˜„์„ ์œ„ํ•ด์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ƒˆ๋กœ์šด ์ž‘์—…์„ ์ž˜ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ๊ฐ•๋ ฅํ•œ ๋ฌธ์ œ ํ•ด๊ฒฐ ๊ธฐ๋Šฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ๋ฌธ์ œ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๋ ค๋ฉด ์—์ด์ „ํŠธ๊ฐ€ ์ถ”๋ก ํ•˜๊ณ  ๊ณ„ํšํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ๊ณผ ์™ธ๋ถ€ ํ™˜๊ฒฝ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ถ”๋ก , ๊ณ„ํš, ๋„๊ตฌ ํ˜ธ์ถœ์ด ์—์ด์ „ํŠธ ์„ฑ๊ณต์— ์ค‘์š”ํ•œ ์ด์œ ๋ฅผ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.
์—์ด์ „ํŠธ๋Š” ์‹ค์ œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์–ธ์–ด ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํ™•์žฅํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์„ฑ๊ณต์ ์ธ ๊ตฌํ˜„์„ ์œ„ํ•ด์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ƒˆ๋กœ์šด ์ž‘์—…์„ ์ž˜ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ๊ฐ•๋ ฅํ•œ ๋ฌธ์ œ ํ•ด๊ฒฐ ๊ธฐ๋Šฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ๋ฌธ์ œ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๋ ค๋ฉด ์—์ด์ „ํŠธ๊ฐ€ ์ถ”๋ก ํ•˜๊ณ  ๊ณ„ํšํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ๊ณผ ์™ธ๋ถ€ ํ™˜๊ฒฝ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ถ”๋ก , ๊ณ„ํš, ๋„๊ตฌ ํ˜ธ์ถœ์ด ์—์ด์ „ํŠธ ์„ฑ๊ณต์— ์™œ ์ค‘์š”ํ•œ์ง€ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.
์—์ด์ „ํŠธ๋Š” ์–ธ์–ด ๋ชจ๋ธ ๊ธฐ๋Šฅ์„ ํ™•์žฅํ•˜์—ฌ ์‹ค์ œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์„ฑ๊ณต์ ์ธ ๊ตฌํ˜„์„ ์œ„ํ•ด์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ƒˆ๋กœ์šด ์ž‘์—…์„ ์ž˜ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ๊ฐ•๋ ฅํ•œ ๋ฌธ์ œ ํ•ด๊ฒฐ ๊ธฐ๋Šฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ๋ฌธ์ œ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๋ ค๋ฉด ์—์ด์ „ํŠธ๋Š” ์ถ”๋ก ๊ณผ ๊ณ„ํš์€ ๋ฌผ๋ก  ์™ธ๋ถ€ ํ™˜๊ฒฝ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋Š” ๋„๊ตฌ๋ฅผ ํ˜ธ์ถœํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜์—์„œ๋Š” ์ถ”๋ก , ๊ณ„ํš, ๋„๊ตฌ ํ˜ธ์ถœ์ด ์—์ด์ „ํŠธ ์„ฑ๊ณต์— ์ค‘์š”ํ•œ ์ด์œ ๋ฅผ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.
Training with synthetic data makes evaluation decontamination harder.The use of synthetic data in model training poses significant challenges to fair evaluation. Evaluation benchmarks are often created by referring to public text sources, such as coursework websites or forums. Consequently, it is arguable that all publicly available benchmark test cases might occasionally be included in the pre-training data of LLMs (Gao et al., 2021; Hoffmann et al., 2022). The use of synthetic data exacerbates this issue rather than mitigating it. Although the community has proposed several techniques to detect such evaluation contamination, such as _min-k9 prob_(Shi et al., 2023), which checks the probabilities of \(k\) long-tail tokens, these token-level decontamination methods are inadequate whenthe model is trained with synthetic data. Synthetic data might include rephrased versions of the benchmark data (Mattern et al., 2023; Oren et al., 2023), rendering token-level decontamination ineffective. In addition to developing more advanced evaluation contamination detection techniques, we recommend that model developers invest in creating and maintaining in-house and protected evaluation benchmarks. These proprietary benchmarks should be carefully safeguarded to prevent leakage and ensure the integrity of the evaluation process.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋ฉด ํ‰๊ฐ€ ์˜ค์—ผ ์ œ๊ฑฐ๊ฐ€ ๋” ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค.๋ชจ๋ธ ํ›ˆ๋ จ์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ณต์ •ํ•œ ํ‰๊ฐ€์— ์ƒ๋‹นํ•œ ์–ด๋ ค์›€์ด ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋Š” ์ข…์ข… ๊ณผ์ œ ์›น์‚ฌ์ดํŠธ๋‚˜ ํฌ๋Ÿผ๊ณผ ๊ฐ™์€ ๊ณต๊ฐœ ํ…์ŠคํŠธ ์†Œ์Šค๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ ์‚ฌ๋ก€๊ฐ€ LLM์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํฌํ•จ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ฃผ์žฅ์ด ์ œ๊ธฐ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Gao ์™ธ., 2021; Hoffmann ์™ธ., 2022). ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์‚ฌ์šฉ์€ ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ๋ณด๋‹ค๋Š” ์•…ํ™”์‹œํ‚ต๋‹ˆ๋‹ค. ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ๋Š” \(k\)์˜ ๊ธด ๊ผฌ๋ฆฌ ํ† ํฐ์˜ ํ™•๋ฅ ์„ ํ™•์ธํ•˜๋Š” _min-k9 prob_(Shi et al., 2023)์™€ ๊ฐ™์€ ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ์˜ค์—ผ์„ ๊ฐ์ง€ํ•˜๊ธฐ ์œ„ํ•œ ์—ฌ๋Ÿฌ ๊ธฐ์ˆ ์„ ์ œ์•ˆํ–ˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ ๋ฐฉ๋ฒ•์€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ฌ ๋•Œ๋Š” ๋ถ€์ ์ ˆํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์—๋Š” ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ์˜ ์žฌ๊ตฌ์„ฑ๋œ ๋ฒ„์ „์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Mattern et al., 2023; Oren et al., 2023), ๋”ฐ๋ผ์„œ ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ๋Š” ํšจ๊ณผ๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. ๋” ๋ฐœ์ „๋œ ํ‰๊ฐ€ ์˜ค์—ผ ๊ฐ์ง€ ๊ธฐ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ๋ชจ๋ธ ๊ฐœ๋ฐœ์ž๋Š” ์‚ฌ๋‚ด ๋ฐ ๋ณดํ˜ธ๋œ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋งŒ๋“ค๊ณ  ์œ ์ง€ ๊ด€๋ฆฌํ•˜๋Š” ๋ฐ ํˆฌ์žํ•  ๊ฒƒ์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…์  ๋ฒค์น˜๋งˆํฌ๋Š” ๋ˆ„์ถœ์„ ๋ฐฉ์ง€ํ•˜๊ณ  ํ‰๊ฐ€ ํ”„๋กœ์„ธ์Šค์˜ ๋ฌด๊ฒฐ์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์‹ ์ค‘ํ•˜๊ฒŒ ๋ณดํ˜ธ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋ฉด ํ‰๊ฐ€ ์˜ค์—ผ์„ ์ œ๊ฑฐํ•˜๊ธฐ๊ฐ€ ๋” ์–ด๋ ต์Šต๋‹ˆ๋‹ค.๋ชจ๋ธ ํ›ˆ๋ จ์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ณต์ •ํ•œ ํ‰๊ฐ€์— ์ƒ๋‹นํ•œ ์–ด๋ ค์›€์ด ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋Š” ์ข…์ข… ์ฝ”์Šค์›Œํฌ ์›น์‚ฌ์ดํŠธ๋‚˜ ํฌ๋Ÿผ๊ณผ ๊ฐ™์€ ๊ณต๊ฐœ ํ…์ŠคํŠธ ์†Œ์Šค๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ๋งŒ๋“ค์–ด์ง‘๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๊ฐ€ LLM์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํฌํ•จ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ฃผ์žฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค(Gao ์™ธ, 2021; Hoffmann ์™ธ, 2022). ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ๋ณด๋‹ค๋Š” ์•…ํ™”์‹œํ‚ต๋‹ˆ๋‹ค. ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ๋Š” \(k\) ๊ผฌ๋ฆฌ ํ† ํฐ์˜ ํ™•๋ฅ ์„ ํ™•์ธํ•˜๋Š” _min-k9 prob_(Shi ์™ธ, 2023)์™€ ๊ฐ™์€ ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ์˜ค์—ผ์„ ํƒ์ง€ํ•˜๊ธฐ ์œ„ํ•œ ์—ฌ๋Ÿฌ ๊ธฐ์ˆ ์„ ์ œ์•ˆํ–ˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ ๋ฐฉ๋ฒ•์€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ๋ชจ๋ธ์„ ํ•™์Šตํ•  ๋•Œ๋Š” ๋ถ€์ ์ ˆํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์—๋Š” ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ์˜ ๋ฌธ๊ตฌ๊ฐ€ ๋ณ€๊ฒฝ๋œ ๋ฒ„์ „์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ(Mattern ์™ธ, 2023; Oren ์™ธ, 2023), ๋”ฐ๋ผ์„œ ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ๋Š” ํšจ๊ณผ๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. ๋ณด๋‹ค ๊ณ ๊ธ‰ ํ‰๊ฐ€ ์˜ค์—ผ ํƒ์ง€ ๊ธฐ์ˆ ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„, ๋ชจ๋ธ ๊ฐœ๋ฐœ์ž๋Š” ์‚ฌ๋‚ด ๋ฐ ๋ณดํ˜ธ๋œ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋งŒ๋“ค๊ณ  ์œ ์ง€ํ•˜๋Š” ๋ฐ ํˆฌ์žํ•  ๊ฒƒ์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…์  ๋ฒค์น˜๋งˆํฌ๋Š” ๋ˆ„์ถœ์„ ๋ฐฉ์ง€ํ•˜๊ณ  ํ‰๊ฐ€ ํ”„๋กœ์„ธ์Šค์˜ ๋ฌด๊ฒฐ์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์‹ ์ค‘ํ•˜๊ฒŒ ๋ณดํ˜ธ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋ฉด ํ‰๊ฐ€ ์˜ค์—ผ ์ œ๊ฑฐ๊ฐ€ ๋” ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค.๋ชจ๋ธ ํ›ˆ๋ จ์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ณต์ •ํ•œ ํ‰๊ฐ€์— ์ƒ๋‹นํ•œ ์–ด๋ ค์›€์ด ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋Š” ์ข…์ข… ์ฝ”์Šค์›Œํฌ ์›น์‚ฌ์ดํŠธ๋‚˜ ํฌ๋Ÿผ๊ณผ ๊ฐ™์€ ๊ณต๊ฐœ ํ…์ŠคํŠธ ์†Œ์Šค๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ ์‚ฌ๋ก€๊ฐ€ LLM์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํฌํ•จ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ฃผ์žฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค(Gao et al., 2021; Hoffmann et al., 2022). ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ๋ณด๋‹ค๋Š” ์•…ํ™”์‹œํ‚ต๋‹ˆ๋‹ค. ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ๋Š” \(k\) ๊ผฌ๋ฆฌ ํ† ํฐ์˜ ํ™•๋ฅ ์„ ํ™•์ธํ•˜๋Š” _min-k9 prob_(Shi et al., 2023)์™€ ๊ฐ™์€ ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ์˜ค์—ผ์„ ๊ฐ์ง€ํ•˜๋Š” ์—ฌ๋Ÿฌ ๊ธฐ์ˆ ์„ ์ œ์•ˆํ–ˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ ๋ฐฉ๋ฒ•์€ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ๋ชจ๋ธ์„ ํ•™์Šตํ•  ๋•Œ๋Š” ๋ถ€์ ์ ˆํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์—๋Š” ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ์˜ ๋ฌธ๊ตฌ๊ฐ€ ๋ณ€๊ฒฝ๋œ ๋ฒ„์ „์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ(๋งคํ„ฐ ์™ธ, 2023; ์˜ค๋ Œ ์™ธ, 2023), ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ๋Š” ํšจ๊ณผ๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. ๋ณด๋‹ค ๊ณ ๊ธ‰ ํ‰๊ฐ€ ์˜ค์—ผ ๊ฐ์ง€ ๊ธฐ์ˆ ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„, ๋ชจ๋ธ ๊ฐœ๋ฐœ์ž๋Š” ์‚ฌ๋‚ด ๋ฐ ๋ณดํ˜ธ๋œ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ์œ ์ง€ํ•˜๋Š” ๋ฐ ํˆฌ์žํ•  ๊ฒƒ์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…์  ๋ฒค์น˜๋งˆํฌ๋Š” ์œ ์ถœ์„ ๋ฐฉ์ง€ํ•˜๊ณ  ํ‰๊ฐ€ ํ”„๋กœ์„ธ์Šค์˜ ๋ฌด๊ฒฐ์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์‹ ์ค‘ํ•˜๊ฒŒ ๋ณดํ˜ธ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋ฉด ํ‰๊ฐ€ ์˜ค์—ผ์ด ๋” ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค.๋ชจ๋ธ ํ›ˆ๋ จ์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ณต์ •ํ•œ ํ‰๊ฐ€์— ์ƒ๋‹นํ•œ ์–ด๋ ค์›€์ด ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋Š” ์ข…์ข… ์ฝ”์Šค์›Œํฌ ์›น์‚ฌ์ดํŠธ๋‚˜ ํฌ๋Ÿผ๊ณผ ๊ฐ™์€ ๊ณต๊ฐœ ํ…์ŠคํŠธ ์†Œ์Šค๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ๋งŒ๋“ค์–ด์ง‘๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๊ฐ€ LLM์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํฌํ•จ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ฃผ์žฅ์ด ์ œ๊ธฐ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Gao et al., 2021; Hoffmann et al., 2022). ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๊ฐ€ ์™„ํ™”๋˜๊ธฐ๋Š”์ปค๋…• ์˜คํžˆ๋ ค ์•…ํ™”๋ฉ๋‹ˆ๋‹ค. ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ๋Š” \(k\)๊ฐœ์˜ ๊ธด ๊ผฌ๋ฆฌ ํ† ํฐ์˜ ํ™•๋ฅ ์„ ํ™•์ธํ•˜๋Š” _min-k9 prob_์™€ ๊ฐ™์€ ๋ช‡ ๊ฐ€์ง€ ๊ธฐ์ˆ ์„ ์ œ์•ˆํ–ˆ์ง€๋งŒ(Shi et al., 2023), ์ด๋Ÿฌํ•œ ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ์ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๊ฒฝ์šฐ์—๋Š” ๋ถ€์ ์ ˆํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์—๋Š” ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ์˜ ๋ฌธ๊ตฌ๊ฐ€ ๋ณ€๊ฒฝ๋œ ๋ฒ„์ „์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ(Mattern et al., 2023; Oren et al., 2023), ์ด๋กœ ์ธํ•ด ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ๊ฐ€ ๋น„ํšจ์œจ์ ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณด๋‹ค ๋ฐœ์ „๋œ ํ‰๊ฐ€ ์˜ค์—ผ ํƒ์ง€ ๊ธฐ์ˆ ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ๋ชจ๋ธ ๊ฐœ๋ฐœ์ž๋Š” ์‚ฌ๋‚ด ๋ฐ ๋ณดํ˜ธ๋œ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋งŒ๋“ค๊ณ  ์œ ์ง€ํ•˜๋Š” ๋ฐ ํˆฌ์žํ•  ๊ฒƒ์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…์  ๋ฒค์น˜๋งˆํฌ๋Š” ์œ ์ถœ์„ ๋ฐฉ์ง€ํ•˜๊ณ  ํ‰๊ฐ€ ํ”„๋กœ์„ธ์Šค์˜ ๋ฌด๊ฒฐ์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์‹ ์ค‘ํ•˜๊ฒŒ ๋ณดํ˜ธํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋ฉด ํ‰๊ฐ€ ์˜ค์—ผ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ๊ฐ€ ๋” ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค.๋ชจ๋ธ ํ›ˆ๋ จ์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๊ณต์ •ํ•œ ํ‰๊ฐ€์— ์‹ฌ๊ฐํ•œ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋Š” ์ข…์ข… ์ฝ”์Šค์›Œํฌ ์›น์‚ฌ์ดํŠธ๋‚˜ ํฌ๋Ÿผ๊ณผ ๊ฐ™์€ ๊ณต๊ฐœ ํ…์ŠคํŠธ ์†Œ์Šค๋ฅผ ์ฐธ์กฐํ•˜์—ฌ ๋งŒ๋“ค์–ด์ง‘๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๊ฐ€ ๋ชจ๋‘ LLM์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํฌํ•จ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ฃผ์žฅ์ด ์ œ๊ธฐ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Gao et al., 2021; Hoffmann et al., 2022). ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๊ฐ€ ์™„ํ™”๋˜๊ธฐ๋Š”์ปค๋…• ์˜คํžˆ๋ ค ์•…ํ™”๋ฉ๋‹ˆ๋‹ค. ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ํ‰๊ฐ€ ์˜ค์—ผ์„ ๊ฐ์ง€ํ•˜๊ธฐ ์œ„ํ•ด \(k\)๊ฐœ์˜ ๊ธด ๊ผฌ๋ฆฌ ํ† ํฐ์˜ ํ™•๋ฅ ์„ ํ™•์ธํ•˜๋Š” _min-k9 prob_(Shi et al., 2023)์™€ ๊ฐ™์€ ๋ช‡ ๊ฐ€์ง€ ๊ธฐ์ˆ ์„ ์ œ์•ˆํ–ˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ์ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๊ฒฝ์šฐ์—๋Š” ๋ถ€์ ์ ˆํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์—๋Š” ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ์˜ ๋ฌธ๊ตฌ๊ฐ€ ๋ณ€๊ฒฝ๋œ ๋ฒ„์ „์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ(Mattern et al., 2023; Oren et al., 2023), ํ† ํฐ ์ˆ˜์ค€์˜ ์˜ค์—ผ ์ œ๊ฑฐ๋Š” ํšจ๊ณผ์ ์ด์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณด๋‹ค ๋ฐœ์ „๋œ ํ‰๊ฐ€ ์˜ค์—ผ ๊ฐ์ง€ ๊ธฐ์ˆ ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ ์™ธ์—๋„ ๋ชจ๋ธ ๊ฐœ๋ฐœ์ž๋Š” ์‚ฌ๋‚ด ๋ฐ ๋ณดํ˜ธ๋œ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๋งŒ๋“ค๊ณ  ์œ ์ง€ํ•˜๋Š” ๋ฐ ํˆฌ์žํ•  ๊ฒƒ์„ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…์  ๋ฒค์น˜๋งˆํฌ๋Š” ์œ ์ถœ์„ ๋ฐฉ์ง€ํ•˜๊ณ  ํ‰๊ฐ€ ํ”„๋กœ์„ธ์Šค์˜ ๋ฌด๊ฒฐ์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด ์‹ ์ค‘ํ•˜๊ฒŒ ๋ณดํ˜ธ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
The logical process of an AutoEncoder is that the original input (set to x) is weighted and mapped to y, which is then inversely weighted and mapped back to z. If through iterative training the loss function L(H) is minimised, i.e. z is as close to x as possible, i.e. x is perfectly reconstructed, then it can be said that forward weighting is a successful way of learning the key features of the input. AutoEncoder language models take their name from the Denoising AutoEncoder (DAE) (Vincent et al., 2008), which is used to predict tokens that are [masked] by contextual words (these [masked] words are actually noise added at the input, typical of thinking). DAE is a technique that involves adding random noise to the input layer of data. This helps to learn more robust features when using an unsupervised approach to pre-train the weights of a deep network in a hierarchical manner.
์˜คํ† ์ธ์ฝ”๋”์˜ ๋…ผ๋ฆฌ์  ํ”„๋กœ์„ธ์Šค๋Š” ์›๋ณธ ์ž…๋ ฅ(x๋กœ ์„ค์ •)์ด ๊ฐ€์ค‘์น˜๊ฐ€ ๋ถ€์—ฌ๋˜์–ด y๋กœ ๋งคํ•‘๋œ ๋‹ค์Œ ์—ญ๊ฐ€์ค‘์น˜๊ฐ€ ๋ถ€์—ฌ๋˜์–ด ๋‹ค์‹œ z๋กœ ๋งคํ•‘๋˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ณต ํ•™์Šต์„ ํ†ตํ•ด ์†์‹ค ํ•จ์ˆ˜ L(H)๊ฐ€ ์ตœ์†Œํ™”๋˜๊ณ , ์ฆ‰ z๊ฐ€ x์— ์ตœ๋Œ€ํ•œ ๊ฐ€๊น๊ณ , ์ฆ‰ x๊ฐ€ ์™„๋ฒฝํ•˜๊ฒŒ ์žฌ๊ตฌ์„ฑ๋˜๋ฉด ์ˆœ๋ฐฉํ–ฅ ๊ฐ€์ค‘์น˜๊ฐ€ ์ž…๋ ฅ์˜ ์ฃผ์š” ํŠน์ง•์„ ํ•™์Šตํ•˜๋Š” ์„ฑ๊ณต์ ์ธ ๋ฐฉ๋ฒ•์ด๋ผ๊ณ  ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜คํ† ์ธ์ฝ”๋” ์–ธ์–ด ๋ชจ๋ธ์€ ๋ฌธ๋งฅ ๋‹จ์–ด๋กœ [๋งˆ์Šคํ‚น]๋œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ์˜คํ† ์ธ์ฝ”๋”(DAE)(Vincent et al., 2008)์—์„œ ๊ทธ ์ด๋ฆ„์„ ๋”ฐ์™”์Šต๋‹ˆ๋‹ค(์ด [๋งˆ์Šคํ‚น]๋œ ๋‹จ์–ด๋Š” ์‹ค์ œ๋กœ ์ž…๋ ฅ์— ์ถ”๊ฐ€๋œ ๋…ธ์ด์ฆˆ๋กœ, ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒ๊ฐํ•  ๋•Œ์˜ ๋…ธ์ด์ฆˆ์ž…๋‹ˆ๋‹ค). DAE๋Š” ๋ฐ์ดํ„ฐ์˜ ์ž…๋ ฅ ๋ ˆ์ด์–ด์— ๋ฌด์ž‘์œ„ ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ธฐ์ˆ ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ๊ณ„์ธต์  ๋ฐฉ์‹์œผ๋กœ ์‹ฌ์ธต ๋„คํŠธ์›Œํฌ์˜ ๊ฐ€์ค‘์น˜๋ฅผ ์‚ฌ์ „ ํ›ˆ๋ จํ•  ๋•Œ ๋น„์ง€๋„ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•  ๋•Œ ๋ณด๋‹ค ๊ฐ•๋ ฅํ•œ ํŠน์ง•์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
์˜คํ† ์ธ์ฝ”๋”์˜ ๋…ผ๋ฆฌ์  ํ”„๋กœ์„ธ์Šค๋Š” ์›๋ณธ ์ž…๋ ฅ(x๋กœ ์„ค์ •)์ด ๊ฐ€์ค‘์น˜๊ฐ€ ๋ถ€์—ฌ๋˜์–ด y๋กœ ๋งคํ•‘๋œ ๋‹ค์Œ ์—ญ๊ฐ€์ค‘์น˜๊ฐ€ ๋ถ€์—ฌ๋˜์–ด ๋‹ค์‹œ z๋กœ ๋งคํ•‘๋˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ณต์ ์ธ ํ•™์Šต์„ ํ†ตํ•ด ์†์‹ค ํ•จ์ˆ˜ L(H)๊ฐ€ ์ตœ์†Œํ™”๋˜๊ณ , ์ฆ‰ z๊ฐ€ x์— ์ตœ๋Œ€ํ•œ ๊ฐ€๊น๊ณ , ์ฆ‰ x๊ฐ€ ์™„๋ฒฝํ•˜๊ฒŒ ์žฌ๊ตฌ์„ฑ๋˜๋ฉด ์ „๋ฐฉ ๊ฐ€์ค‘์ด ์ž…๋ ฅ์˜ ์ฃผ์š” ํŠน์ง•์„ ํ•™์Šตํ•˜๋Š” ์„ฑ๊ณต์ ์ธ ๋ฐฉ๋ฒ•์ด๋ผ๊ณ  ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜คํ† ์ธ์ฝ”๋” ์–ธ์–ด ๋ชจ๋ธ์€ ๋ฌธ๋งฅ ๋‹จ์–ด๋กœ [๊ฐ€๋ฆฌ๊ฐœ]๋œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ์˜คํ† ์ธ์ฝ”๋”(DAE)(Vincent et al., 2008)์—์„œ ๊ทธ ์ด๋ฆ„์„ ๋”ฐ์™”์Šต๋‹ˆ๋‹ค(์ด [๊ฐ€๋ฆฌ๊ฐœ]๋œ ๋‹จ์–ด๋Š” ์‹ค์ œ๋กœ๋Š” ์ž…๋ ฅ์— ์ถ”๊ฐ€๋œ ๋…ธ์ด์ฆˆ๋กœ, ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒ๊ฐํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค). DAE๋Š” ๋ฐ์ดํ„ฐ์˜ ์ž…๋ ฅ ๋ ˆ์ด์–ด์— ๋ฌด์ž‘์œ„ ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ธฐ์ˆ ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ๊ณ„์ธต์  ๋ฐฉ์‹์œผ๋กœ ์‹ฌ์ธต ๋„คํŠธ์›Œํฌ์˜ ๊ฐ€์ค‘์น˜๋ฅผ ์‚ฌ์ „ ํ›ˆ๋ จํ•  ๋•Œ ๋น„์ง€๋„ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•  ๋•Œ ๋” ๊ฐ•๋ ฅํ•œ ํŠน์ง•์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
์ž๋™ ์ธ์ฝ”๋”์˜ ๋…ผ๋ฆฌ์  ํ”„๋กœ์„ธ์Šค๋Š” ์›๋ž˜ ์ž…๋ ฅ(x๋กœ ์„ค์ •๋จ)์ด ๊ฐ€์ค‘์น˜๊ฐ€ ๋ถ€์—ฌ๋˜์–ด y๋กœ ๋งคํ•‘๋œ ๋‹ค์Œ ์—ญ์œผ๋กœ ๊ฐ€์ค‘์น˜๊ฐ€ ๋ถ€์—ฌ๋˜์–ด ๋‹ค์‹œ z๋กœ ๋งคํ•‘๋˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ณต์ ์ธ ํ›ˆ๋ จ์„ ํ†ตํ•ด ์†์‹ค ํ•จ์ˆ˜ L(H)๊ฐ€ ์ตœ์†Œํ™”๋˜๊ณ , ์ฆ‰ z๊ฐ€ x์— ์ตœ๋Œ€ํ•œ ๊ฐ€๊น๊ฒŒ, ์ฆ‰ x๊ฐ€ ์™„๋ฒฝํ•˜๊ฒŒ ์žฌ๊ตฌ์„ฑ๋˜๋ฉด ์ „๋ฐฉ ๊ฐ€์ค‘์ด ์ž…๋ ฅ์˜ ์ฃผ์š” ํŠน์ง•์„ ํ•™์Šตํ•˜๋Š” ์„ฑ๊ณต์ ์ธ ๋ฐฉ๋ฒ•์ด๋ผ๊ณ  ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ž๋™ ์ธ์ฝ”๋” ์–ธ์–ด ๋ชจ๋ธ์€ ๋ฌธ๋งฅ ๋‹จ์–ด๋กœ [๊ฐ€๋ฆฌ๊ฐœ]๋œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ์ž๋™ ์ธ์ฝ”๋”(DAE)(Vincent et al., 2008)์—์„œ ์ด๋ฆ„์„ ๋”ฐ์™”์œผ๋ฉฐ, ์ด [๊ฐ€๋ฆฌ๊ฐœ]๋œ ๋‹จ์–ด๋Š” ์‹ค์ œ๋กœ ์ž…๋ ฅ์— ์ถ”๊ฐ€๋œ ๋…ธ์ด์ฆˆ๋กœ, ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒ๊ฐํ•  ๋•Œ์˜ ๋…ธ์ด์ฆˆ์ž…๋‹ˆ๋‹ค. DAE๋Š” ๋ฐ์ดํ„ฐ์˜ ์ž…๋ ฅ ๋ ˆ์ด์–ด์— ๋ฌด์ž‘์œ„ ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค. ์ด๋Š” ๊ณ„์ธต์  ๋ฐฉ์‹์œผ๋กœ ์‹ฌ์ธต ๋„คํŠธ์›Œํฌ์˜ ๊ฐ€์ค‘์น˜๋ฅผ ์‚ฌ์ „ ํ›ˆ๋ จํ•  ๋•Œ ๋น„์ง€๋„ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•  ๋•Œ ๋” ๊ฐ•๋ ฅํ•œ ํŠน์ง•์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
AutoEncoder์˜ ๋…ผ๋ฆฌ์  ํ”„๋กœ์„ธ์Šค๋Š” ์›๋ž˜ ์ž…๋ ฅ(x๋กœ ์„ค์ •)์— ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜์—ฌ y์— ๋งคํ•‘ํ•œ ๋‹ค์Œ, ์—ญ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜์—ฌ ๋‹ค์‹œ z์— ๋งคํ•‘ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ณต์ ์ธ ํ•™์Šต์„ ํ†ตํ•ด ์†์‹ค ํ•จ์ˆ˜ L(H)์ด ์ตœ์†Œํ™”๋˜๋ฉด(์ฆ‰, z๊ฐ€ x์— ์ตœ๋Œ€ํ•œ ๊ฐ€๊นŒ์›Œ์ง€๋ฉด, ์ฆ‰ x๊ฐ€ ์™„๋ฒฝํ•˜๊ฒŒ ์žฌ๊ตฌ์„ฑ๋˜๋ฉด) ์ „๋ฐฉ ๊ฐ€์ค‘์น˜๊ฐ€ ์ž…๋ ฅ์˜ ์ฃผ์š” ํŠน์ง•์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ์„ฑ๊ณต์ ์ธ ๋ฐฉ๋ฒ•์ด๋ผ๊ณ  ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. AutoEncoder ์–ธ์–ด ๋ชจ๋ธ์€ ๋ฌธ๋งฅ ๋‹จ์–ด์— ์˜ํ•ด [๋งˆ์Šคํ‚น]๋œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ AutoEncoder(DAE)(Vincent et al., 2008)์—์„œ ๊ทธ ์ด๋ฆ„์„ ๋”ฐ์™”์Šต๋‹ˆ๋‹ค(์ด๋Ÿฌํ•œ [๋งˆ์Šคํ‚น] ๋‹จ์–ด๋Š” ์‹ค์ œ๋กœ ์ž…๋ ฅ์— ์ถ”๊ฐ€๋œ ๋…ธ์ด์ฆˆ๋กœ, ์‚ฌ๊ณ ์—์„œ ํ”ํžˆ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค). DAE๋Š” ๋ฐ์ดํ„ฐ์˜ ์ž…๋ ฅ ๋ ˆ์ด์–ด์— ๋ฌด์ž‘์œ„ ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์‹ฌ์ธต ๋„คํŠธ์›Œํฌ์˜ ๊ฐ€์ค‘์น˜๋ฅผ ๊ณ„์ธต์  ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด ๋น„์ง€๋„ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•  ๋•Œ ๋ณด๋‹ค ๊ฐ•๋ ฅํ•œ ํŠน์ง•์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
์ž๋™ ์ธ์ฝ”๋”์˜ ๋…ผ๋ฆฌ์  ํ”„๋กœ์„ธ์Šค๋Š” ์›๋ž˜ ์ž…๋ ฅ(x๋กœ ์„ค์ •)์— ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜์—ฌ y์— ๋งคํ•‘ํ•œ ๋‹ค์Œ, ์—ญ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜์—ฌ ๋‹ค์‹œ z์— ๋งคํ•‘ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ณต์ ์ธ ํ›ˆ๋ จ์„ ํ†ตํ•ด ์†์‹ค ํ•จ์ˆ˜ L(H)์ด ์ตœ์†Œํ™”๋˜๋ฉด, ์ฆ‰ z๊ฐ€ ๊ฐ€๋Šฅํ•œ ํ•œ x์— ๊ฐ€๊นŒ์šฐ๋ฉด, ์ฆ‰ x๊ฐ€ ์™„๋ฒฝํ•˜๊ฒŒ ์žฌ๊ตฌ์„ฑ๋˜๋ฉด, ์ˆœ๋ฐฉํ–ฅ ๊ฐ€์ค‘์น˜๊ฐ€ ์ž…๋ ฅ์˜ ์ฃผ์š” ํŠน์ง•์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ์„ฑ๊ณต์ ์ธ ๋ฐฉ๋ฒ•์ด๋ผ๊ณ  ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ž๋™ ์ธ์ฝ”๋” ์–ธ์–ด ๋ชจ๋ธ์€ ๋ฌธ๋งฅ ๋‹จ์–ด์— ์˜ํ•ด [๋งˆ์Šคํ‚น]๋œ ํ† ํฐ์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ์ž๋™ ์ธ์ฝ”๋”(DAE)(Vincent ์™ธ, 2008)์—์„œ ๊ทธ ์ด๋ฆ„์„ ๋”ฐ์™”์Šต๋‹ˆ๋‹ค(์ด๋Ÿฌํ•œ [๋งˆ์Šคํ‚น] ๋‹จ์–ด๋Š” ์‹ค์ œ๋กœ ์ž…๋ ฅ์— ์ถ”๊ฐ€๋œ ๋…ธ์ด์ฆˆ๋กœ, ์ƒ๊ฐ์˜ ์ „ํ˜•์ž…๋‹ˆ๋‹ค). DAE๋Š” ๋ฐ์ดํ„ฐ์˜ ์ž…๋ ฅ ๋ ˆ์ด์–ด์— ๋ฌด์ž‘์œ„ ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค. ์ด๋Š” ๋น„์ง€๋„ ๋ฐฉ์‹์œผ๋กœ ์‹ฌ์ธต ๋„คํŠธ์›Œํฌ์˜ ๊ฐ€์ค‘์น˜๋ฅผ ๊ณ„์ธต์  ๋ฐฉ์‹์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จํ•  ๋•Œ ๋ณด๋‹ค ๊ฐ•๋ ฅํ•œ ํŠน์ง•์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
We fit the Hoffmann et al. parametric scaling law to our reconstructed subset of the data:
์šฐ๋ฆฌ๋Š” ์žฌ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ํ•˜ํ”„๋งŒ ๋“ฑ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋งž์ถ”๊ณ  ์žˆ๋‹ค:
์šฐ๋ฆฌ๋Š” ๋ฐ์ดํ„ฐ์˜ ์žฌ๊ตฌ์„ฑ๋œ ํ•˜์œ„ ์ง‘ํ•ฉ์— Hoffmann et al. ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ์ ์šฉํ•œ๋‹ค:
์šฐ๋ฆฌ๋Š” ๋ฐ์ดํ„ฐ์˜ ์žฌ๊ตฌ์„ฑ๋œ ํ•˜์œ„ ์ง‘ํ•ฉ์— Hoffmann et al. ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ๋งž์ท„์Šต๋‹ˆ๋‹ค:
ํ˜ธํ”„๋งŒ ๋“ฑ์ด ์ œ์•ˆํ•œ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ์žฌ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์— ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค:
์šฐ๋ฆฌ๋Š” ํ˜ธํ”„๋งŒ ๋“ฑ์ด ์žฌ๊ตฌ์„ฑํ•œ ๋ฐ์ดํ„ฐ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์— ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ์ ์šฉํ–ˆ๋‹ค:
decreases, as the LM distribution anneals to argmax, as well as when temperature increases, as the distribution collapses to uniform. For both top-p and top-k we note that the model requires almost all of the distribution to perform well, indicating that there is significant information in the tails.
LM ๋ถ„ํฌ๊ฐ€ argmax๋กœ ์–ด๋‹๋ง๋  ๋•Œ์™€ ์˜จ๋„๊ฐ€ ์ฆ๊ฐ€ํ•˜์—ฌ ๋ถ„ํฌ๊ฐ€ ๊ท ์ผํ•˜๊ฒŒ ๋ถ•๊ดด๋  ๋•Œ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. top-p์™€ top-k ๋ชจ๋‘์—์„œ ๋ชจ๋ธ์ด ๊ฑฐ์˜ ๋ชจ๋“  ๋ถ„ํฌ๋ฅผ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•˜๋ฉฐ, ์ด๋Š” ๊ผฌ๋ฆฌ์— ์ƒ๋‹นํ•œ ์ •๋ณด๊ฐ€ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
LM ๋ถ„ํฌ๊ฐ€ argmax๋กœ ์–ด๋‹๋ง๋ ์ˆ˜๋ก, ๋ถ„ํฌ๊ฐ€ ๊ท ์ผํ•˜๊ฒŒ ๋ถ•๊ดด๋ ์ˆ˜๋ก ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. top-p์™€ top-k ๋ชจ๋‘์—์„œ ๋ชจ๋ธ์ด ๊ฑฐ์˜ ๋ชจ๋“  ๋ถ„ํฌ๋ฅผ ์ž˜ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•˜๋ฉฐ, ์ด๋Š” ๊ผฌ๋ฆฌ์— ์ƒ๋‹นํ•œ ์ •๋ณด๊ฐ€ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
LM ๋ถ„ํฌ๊ฐ€ argmax์— ์–ด๋‹๋ง๋  ๋•Œ์™€ ๋ถ„ํฌ๊ฐ€ ๊ท ์ผํ•˜๊ฒŒ ๋ถ•๊ดด๋  ๋•Œ ์˜จ๋„๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. top-p์™€ top-k ๋ชจ๋‘์—์„œ ๋ชจ๋ธ์ด ๊ฑฐ์˜ ๋ชจ๋“  ๋ถ„ํฌ๋ฅผ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•˜๋ฏ€๋กœ ๊ผฌ๋ฆฌ์— ์ƒ๋‹นํ•œ ์ •๋ณด๊ฐ€ ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
argmax๋กœ ์–ด๋‹๋งํ•  ๋•Œ, ๊ทธ๋ฆฌ๊ณ  ์˜จ๋„๊ฐ€ ์ฆ๊ฐ€ํ•  ๋•Œ, ๋ถ„ํฌ๊ฐ€ ๊ท ์ผ๋กœ ๋ถ•๊ดดํ•  ๋•Œ LM ๋ถ„ํฌ๊ฐ€ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. Top-p์™€ top-k ๋ชจ๋‘ ๋ชจ๋ธ์ด ๋ถ„ํฌ์˜ ๊ฑฐ์˜ ๋ชจ๋“  ๋ถ€๋ถ„์„ ์‚ฌ์šฉํ•ด์•ผ ์ž˜ ์ž‘๋™ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๊ผฌ๋ฆฌ์— ์ƒ๋‹นํ•œ ์ •๋ณด๊ฐ€ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
LM ๋ถ„ํฌ๊ฐ€ argmax๋กœ ์–ด๋‹๋ง๋  ๋•Œ, ๊ทธ๋ฆฌ๊ณ  ์˜จ๋„๊ฐ€ ์ƒ์Šนํ•  ๋•Œ ๋ถ„ํฌ๊ฐ€ ๊ท ์ผํ•˜๊ฒŒ ๋ถ•๊ดด๋  ๋•Œ ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. Top-p์™€ top-k ๋ชจ๋‘์— ๋Œ€ํ•ด ๋ชจ๋ธ์ด ๊ฑฐ์˜ ๋ชจ๋“  ๋ถ„ํฌ๋ฅผ ์ˆ˜ํ–‰ํ•ด์•ผ ์ž˜ ์ž‘๋™ํ•œ๋‹ค๋Š” ์ ์— ์ฃผ๋ชฉํ•  ํ•„์š”๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๊ผฌ๋ฆฌ์— ์ƒ๋‹นํ•œ ์ •๋ณด๊ฐ€ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
Our result raises a number of questions. Why do models suffer the Reversal Curse? Do non-auto-regressive models suffer from it as well? Do humans suffer from some form of the Reversal Curse? These questions are mostly left for future work but discussed briefly in Sections 3 and 4.
์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์˜๋ฌธ์„ ์ œ๊ธฐํ•œ๋‹ค. ๋ชจ๋ธ์ด ์™œ ์—ญ์ „ ์ €์ฃผ๋ฅผ ๊ฒช์„๊นŒ์š”? ๋น„์ž๊ธฐํšŒ๊ท€ ๋ชจ๋ธ๋„ ์—ญ์ „ ์ €์ฃผ๋ฅผ ๊ฒช์„๊นŒ์š”? ์ธ๊ฐ„๋„ ์–ด๋–ค ํ˜•ํƒœ์˜ ์—ญ์ „ ์ €์ฃผ๋ฅผ ๊ฒช์„๊นŒ์š”? ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์€ ๋Œ€๋ถ€๋ถ„ ํ–ฅํ›„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•ด ๋‚จ๊ฒจ๋‘์—ˆ์ง€๋งŒ 3์žฅ๊ณผ 4์žฅ์—์„œ ๊ฐ„๋žตํ•˜๊ฒŒ ๋…ผ์˜๋˜์—ˆ๋‹ค.
์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์˜๋ฌธ์„ ์ œ๊ธฐํ•œ๋‹ค. ๋ชจ๋ธ์ด ๋ฐ˜์ „ ์ €์ฃผ๋ฅผ ๊ฒช๋Š” ์ด์œ ๋Š” ๋ฌด์—‡์ธ๊ฐ€์š”? ๋น„์ž๊ธฐ ํšŒ๊ท€ ๋ชจ๋ธ๋„ ๋ฐ˜์ „ ์ €์ฃผ๋ฅผ ๊ฒช๋‚˜์š”? ์ธ๊ฐ„๋„ ๋ฐ˜์ „ ์ €์ฃผ์— ๊ฑธ๋ฆฐ ํ˜•ํƒœ๊ฐ€ ์žˆ๋‚˜์š”? ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์€ ๋Œ€๋ถ€๋ถ„ ํ–ฅํ›„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•ด ๋‚จ๊ฒจ๋‘์—ˆ์ง€๋งŒ ์„น์…˜ 3๊ณผ 4์—์„œ ๊ฐ„๋žตํ•˜๊ฒŒ ๋…ผ์˜๋ฉ๋‹ˆ๋‹ค.
์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์งˆ๋ฌธ์„ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด '๋ฐ˜์ „ ์ €์ฃผ'๋ฅผ ๋ฐ›๋Š” ์ด์œ ๋Š” ๋ฌด์—‡์ธ๊ฐ€์š”? ๋น„์ž๋™ ํšŒ๊ท€ ๋ชจ๋ธ๋„ ์ด ๋ฌธ์ œ๋ฅผ ๊ฒช๋‚˜์š”? ์ธ๊ฐ„๋„ ์ผ์ข…์˜ '๋ฐ˜์ „ ์ €์ฃผ'๋ฅผ ๊ฒช๋Š”๊ฐ€? ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์€ ๋Œ€๋ถ€๋ถ„ ํ–ฅํ›„ ์ž‘์—…์„ ์œ„ํ•ด ๋‚จ๊ฒจ๋‘์—ˆ์ง€๋งŒ ์„น์…˜ 3๊ณผ 4์—์„œ ๊ฐ„๋žตํ•˜๊ฒŒ ๋…ผ์˜๋ฉ๋‹ˆ๋‹ค.
์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์งˆ๋ฌธ์„ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์€ ์™œ ๋ฐ˜์ „์˜ ์ €์ฃผ์— ์‹œ๋‹ฌ๋ฆฌ๋Š”๊ฐ€? ๋น„์ž๋™ ํšŒ๊ท€ ๋ชจ๋ธ๋„ ๋ฐ˜์ „์˜ ์ €์ฃผ์— ์‹œ๋‹ฌ๋ฆฌ๋Š”๊ฐ€? ์ธ๊ฐ„์€ ์–ด๋–ค ํ˜•ํƒœ์˜ ๋ฐ˜์ „์˜ ์ €์ฃผ์— ์‹œ๋‹ฌ๋ฆฌ๋Š”๊ฐ€? ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์€ ๋Œ€๋ถ€๋ถ„ ํ–ฅํ›„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ๊ฒƒ์ด์ง€๋งŒ 3์ ˆ๊ณผ 4์ ˆ์—์„œ ๊ฐ„๋žตํ•˜๊ฒŒ ๋…ผ์˜๋ฉ๋‹ˆ๋‹ค.
์ด ๊ฒฐ๊ณผ๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์˜๋ฌธ์„ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค. ์™œ ๋ชจ๋ธ์ด ๋ฐ˜์ „์˜ ์ €์ฃผ์— ์‹œ๋‹ฌ๋ฆฌ๋Š”๊ฐ€? ๋น„์ž๋™ ํšŒ๊ท€ ๋ชจ๋ธ๋„ ์ด ์ €์ฃผ์— ์‹œ๋‹ฌ๋ฆฌ๋Š”๊ฐ€? ์ธ๊ฐ„๋„ ์–ด๋–ค ํ˜•ํƒœ์˜ ๋ฐ˜์ „์˜ ์ €์ฃผ์— ์‹œ๋‹ฌ๋ฆฌ๋Š”๊ฐ€? ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์€ ๋Œ€๋ถ€๋ถ„ ํ–ฅํ›„ ์ž‘์—…์„ ์œ„ํ•ด ๋‚จ๊ฒจ ๋‘์—ˆ์ง€๋งŒ ์„น์…˜ 3๊ณผ 4์—์„œ ๊ฐ„๋žตํ•˜๊ฒŒ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค.
**Civil Engineering**: In civil engineering, LLM-based agents can be used to design and optimize complex structures such as buildings, bridges, dams, roads, etc. [138] proposes an interactive framework where human architects and agents collaborate to construct structures in a 3D simulation environment. The interactive agent can understand natural language instructions, place blocks, detect confusion, seek clarification, and incorporate human feedback, showing the potential for human-AI collaboration in engineering design.
**ํ† ๋ชฉ ๊ณตํ•™**: ํ† ๋ชฉ ๊ณตํ•™์—์„œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ๊ฑด๋ฌผ, ๊ต๋Ÿ‰, ๋Œ, ๋„๋กœ ๋“ฑ๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ๊ตฌ์กฐ๋ฌผ์„ ์„ค๊ณ„ํ•˜๊ณ  ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. [138]์€ ์ธ๊ฐ„ ๊ฑด์ถ•๊ฐ€์™€ ์—์ด์ „ํŠธ๊ฐ€ 3D ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์—์„œ ๊ตฌ์กฐ๋ฌผ์„ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•ด ํ˜‘๋ ฅํ•˜๋Š” ๋Œ€ํ™”ํ˜• ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ™”ํ˜• ์—์ด์ „ํŠธ๋Š” ์ž์—ฐ์–ด ๋ช…๋ น์„ ์ดํ•ดํ•˜๊ณ , ๋ธ”๋ก์„ ๋ฐฐ์น˜ํ•˜๊ณ , ํ˜ผ๋™์„ ๊ฐ์ง€ํ•˜๊ณ , ์„ค๋ช…์„ ๊ตฌํ•˜๊ณ , ์‚ฌ๋žŒ์˜ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ฉํ•˜์—ฌ ์—”์ง€๋‹ˆ์–ด๋ง ์„ค๊ณ„์—์„œ ์ธ๊ฐ„๊ณผ AI์˜ ํ˜‘์—… ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
**ํ† ๋ชฉ ๊ณตํ•™**: ํ† ๋ชฉ ๊ณตํ•™์—์„œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ๊ฑด๋ฌผ, ๊ต๋Ÿ‰, ๋Œ, ๋„๋กœ ๋“ฑ๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ๊ตฌ์กฐ๋ฌผ์„ ์„ค๊ณ„ํ•˜๊ณ  ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 3D ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์—์„œ ์ธ๊ฐ„ ๊ฑด์ถ•๊ฐ€์™€ ์—์ด์ „ํŠธ๊ฐ€ ํ˜‘์—…ํ•˜์—ฌ ๊ตฌ์กฐ๋ฌผ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋Œ€ํ™”ํ˜• ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ™”ํ˜• ์—์ด์ „ํŠธ๋Š” ์ž์—ฐ์–ด ์ง€์นจ์„ ์ดํ•ดํ•˜๊ณ , ๋ธ”๋ก์„ ๋ฐฐ์น˜ํ•˜๊ณ , ํ˜ผ๋™์„ ๊ฐ์ง€ํ•˜๊ณ , ์„ค๋ช…์„ ๊ตฌํ•˜๊ณ , ์ธ๊ฐ„์˜ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ฉํ•˜์—ฌ ์—”์ง€๋‹ˆ์–ด๋ง ์„ค๊ณ„์—์„œ ์ธ๊ฐ„๊ณผ AI์˜ ํ˜‘์—… ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
**ํ† ๋ชฉ ๊ณตํ•™**: ํ† ๋ชฉ ๊ณตํ•™์—์„œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ๊ฑด๋ฌผ, ๊ต๋Ÿ‰, ๋Œ, ๋„๋กœ ๋“ฑ๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ๊ตฌ์กฐ๋ฌผ์„ ์„ค๊ณ„ํ•˜๊ณ  ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. [138]์€ 3D ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์—์„œ ์ธ๊ฐ„ ๊ฑด์ถ•๊ฐ€์™€ ์—์ด์ „ํŠธ๊ฐ€ ํ˜‘์—…ํ•˜์—ฌ ๊ตฌ์กฐ๋ฌผ์„ ๊ฑด์„คํ•˜๋Š” ๋Œ€ํ™”ํ˜• ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ™”ํ˜• ์—์ด์ „ํŠธ๋Š” ์ž์—ฐ์–ด ๋ช…๋ น์–ด๋ฅผ ์ดํ•ดํ•˜๊ณ , ๋ธ”๋ก์„ ๋ฐฐ์น˜ํ•˜๊ณ , ํ˜ผ๋™์„ ๊ฐ์ง€ํ•˜๊ณ , ๋ช…ํ™•ํ•œ ์„ค๋ช…์„ ๊ตฌํ•˜๊ณ , ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ฉํ•  ์ˆ˜ ์žˆ์–ด ์—”์ง€๋‹ˆ์–ด๋ง ์„ค๊ณ„์—์„œ ์ธ๊ฐ„๊ณผ AI์˜ ํ˜‘์—… ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
**ํ† ๋ชฉ ๊ณตํ•™**: ํ† ๋ชฉ ๊ณตํ•™์—์„œ๋Š” LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฑด๋ฌผ, ๊ต๋Ÿ‰, ๋Œ, ๋„๋กœ ๋“ฑ๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ๊ตฌ์กฐ๋ฌผ์„ ์„ค๊ณ„ํ•˜๊ณ  ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. [138]์€ 3D ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์—์„œ ์ธ๊ฐ„ ๊ฑด์ถ•๊ฐ€์™€ ์—์ด์ „ํŠธ๊ฐ€ ํ˜‘์—…ํ•˜์—ฌ ๊ตฌ์กฐ๋ฌผ์„ ๊ฑด์„คํ•˜๋Š” ๋Œ€ํ™”ํ˜• ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ™”ํ˜• ์—์ด์ „ํŠธ๋Š” ์ž์—ฐ์–ด ๋ช…๋ น์–ด๋ฅผ ์ดํ•ดํ•˜๊ณ , ๋ธ”๋ก์„ ๋ฐฐ์น˜ํ•˜๊ณ , ํ˜ผ๋™์„ ๊ฐ์ง€ํ•˜๊ณ , ์„ค๋ช…์„ ๊ตฌํ•˜๊ณ , ์ธ๊ฐ„์˜ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ฉํ•˜์—ฌ ์—”์ง€๋‹ˆ์–ด๋ง ์„ค๊ณ„์—์„œ ์ธ๊ฐ„๊ณผ AI์˜ ํ˜‘์—… ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
**ํ† ๋ชฉ ๊ณตํ•™**: ํ† ๋ชฉ ๊ณตํ•™์—์„œ๋Š” LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฑด๋ฌผ, ๊ต๋Ÿ‰, ๋Œ, ๋„๋กœ ๋“ฑ์˜ ๋ณต์žกํ•œ ๊ตฌ์กฐ๋ฌผ์„ ์„ค๊ณ„ํ•˜๊ณ  ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. [138]์€ 3D ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์—์„œ ์ธ๊ฐ„ ๊ฑด์ถ•๊ฐ€์™€ ์—์ด์ „ํŠธ๊ฐ€ ํ˜‘์—…ํ•˜์—ฌ ๊ตฌ์กฐ๋ฌผ์„ ๊ฑด์„คํ•˜๋Š” ๋Œ€ํ™”ํ˜• ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋Œ€ํ™”ํ˜• ์—์ด์ „ํŠธ๋Š” ์ž์—ฐ์–ด ๋ช…๋ น์„ ์ดํ•ดํ•˜๊ณ , ๋ธ”๋ก์„ ๋ฐฐ์น˜ํ•˜๊ณ , ํ˜ผ๋ž€์„ ๊ฐ์ง€ํ•˜๊ณ , ์„ค๋ช…์„ ๊ตฌํ•˜๊ณ , ์‚ฌ๋žŒ์˜ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ฉํ•˜์—ฌ ์—”์ง€๋‹ˆ์–ด๋ง ์„ค๊ณ„์—์„œ ์ธ๊ฐ„๊ณผ AI์˜ ํ˜‘์—… ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
Word Tokenization.The most basic tokenization approach is the splitting of sequences based on white spaces and considering each word as a token (Bengio et al., 2000).
๋‹จ์–ด ํ† ํฐํ™”.๊ฐ€์žฅ ๊ธฐ๋ณธ์ ์ธ ํ† ํฐํ™” ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ณต๋ฐฑ์„ ๊ธฐ์ค€์œผ๋กœ ์‹œํ€€์Šค๋ฅผ ๋ถ„ํ• ํ•˜๊ณ  ๊ฐ ๋‹จ์–ด๋ฅผ ํ† ํฐ์œผ๋กœ ๊ฐ„์ฃผํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(Bengio et al., 2000).
๋‹จ์–ด ํ† ํฐํ™”.๊ฐ€์žฅ ๊ธฐ๋ณธ์ ์ธ ํ† ํฐํ™” ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ณต๋ฐฑ์„ ๊ธฐ์ค€์œผ๋กœ ์‹œํ€€์Šค๋ฅผ ๋ถ„ํ• ํ•˜๊ณ  ๊ฐ ๋‹จ์–ด๋ฅผ ํ† ํฐ์œผ๋กœ ๊ฐ„์ฃผํ•˜๋Š” ๊ฒƒ์ด๋‹ค(Bengio et al., 2000).
๋‹จ์–ด ํ† ํฐํ™”.๊ฐ€์žฅ ๊ธฐ๋ณธ์ ์ธ ํ† ํฐํ™” ๋ฐฉ๋ฒ•์€ ๊ณต๋ฐฑ์„ ๊ธฐ์ค€์œผ๋กœ ์‹œํ€€์Šค๋ฅผ ๋ถ„ํ• ํ•˜๊ณ  ๊ฐ ๋‹จ์–ด๋ฅผ ํ† ํฐ์œผ๋กœ ๊ฐ„์ฃผํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(Bengio et al., 2000).
๋‹จ์–ด ํ† ํฐํ™”.๊ฐ€์žฅ ๊ธฐ๋ณธ์ ์ธ ํ† ํฐํ™” ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ณต๋ฐฑ์„ ๊ธฐ์ค€์œผ๋กœ ์‹œํ€€์Šค๋ฅผ ๋ถ„ํ• ํ•˜๊ณ  ๊ฐ ๋‹จ์–ด๋ฅผ ํ† ํฐ์œผ๋กœ ๊ฐ„์ฃผํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(Bengio et al., 2000).
๋‹จ์–ด ํ† ํฐํ™”.๊ฐ€์žฅ ๊ธฐ๋ณธ์ ์ธ ํ† ํฐํ™” ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ณต๋ฐฑ์„ ๊ธฐ์ค€์œผ๋กœ ์‹œํ€€์Šค๋ฅผ ๋ถ„ํ• ํ•˜๊ณ  ๊ฐ ๋‹จ์–ด๋ฅผ ํ† ํฐ์œผ๋กœ ๊ฐ„์ฃผํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค(Bengio et al., 2000).
We use next-word prediction loss for finetuning, and only compute loss on the responses, but not system and user instructions. We use AdamW optimizer with \(\beta_{1}\) set to 0.9, \(\beta_{2}\) set to 0.999, and \(\epsilon\) set to \(10^{-8}\). We use a sequence length of 4096, alongside a batch size of 64. We set training step to 300 with a constant \(1\times 10^{-5}\) learning rate, a weight decay of 0.1, gradient clipping with a maximum threshold of 1.0, and NEFTune [34] with a noise scale of 45 for Yi-34B-Chat and 5 for Yi-6B-Chat.
ํŒŒ์ธํŠœ๋‹์—๋Š” ๋‹ค์Œ ๋‹จ์–ด ์˜ˆ์ธก ์†์‹ค์„ ์‚ฌ์šฉํ•˜๋ฉฐ, ์‹œ์Šคํ…œ ๋ฐ ์‚ฌ์šฉ์ž ์ง€์นจ์ด ์•„๋‹Œ ์‘๋‹ต์— ๋Œ€ํ•ด์„œ๋งŒ ์†์‹ค์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. \(\๋ฒ ํƒ€_{1}\)์„ 0.9๋กœ, \(\๋ฒ ํƒ€_{2}\)๋ฅผ 0.999๋กœ, \(\์—ก์‹ค๋ก \)์„ \(10^{-8}\)๋กœ ์„ค์ •ํ•œ AdamW ์ตœ์ ํ™”๊ธฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์‹œํ€€์Šค ๊ธธ์ด๋Š” 4096, ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 64๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋‹จ๊ณ„๋Š” 300์œผ๋กœ ์„ค์ •ํ•˜๊ณ , ํ•™์Šต ์†๋„๋Š” 1\๋ฐฐ 10^{-5}\), ๊ฐ€์ค‘์น˜ ๊ฐ์‡ ๋Š” 0.1, ์ตœ๋Œ€ ์ž„๊ณ„๊ฐ’ 1.0์˜ ๊ทธ๋ผ๋””์–ธํŠธ ํด๋ฆฌํ•‘, Yi-34B-Chat์˜ ๊ฒฝ์šฐ ๋…ธ์ด์ฆˆ ์Šค์ผ€์ผ 45, Yi-6B-Chat์˜ ๊ฒฝ์šฐ 5๋กœ NEFTune [34]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
ํŒŒ์ธํŠœ๋‹์—๋Š” ๋‹ค์Œ ๋‹จ์–ด ์˜ˆ์ธก ์†์‹ค์„ ์‚ฌ์šฉํ•˜๋ฉฐ, ์‹œ์Šคํ…œ ๋ฐ ์‚ฌ์šฉ์ž ์ง€์นจ์ด ์•„๋‹Œ ์‘๋‹ต์— ๋Œ€ํ•ด์„œ๋งŒ ์†์‹ค์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. \(\๋ฒ ํƒ€_{1}\)์€ 0.9๋กœ, \(\๋ฒ ํƒ€_{2}\)๋Š” 0.999๋กœ, \(\์—ก์‹ค๋ก \)์€ \(10^{-8}\)๋กœ ์„ค์ •๋œ AdamW ์ตœ์ ํ™”๊ธฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์‹œํ€€์Šค ๊ธธ์ด๋Š” 4096, ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 64๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋‹จ๊ณ„๋Š” 300์œผ๋กœ ์„ค์ •ํ•˜๊ณ , ํ•™์Šต ์†๋„๋Š” 1\times 10^{-5}\), ๊ฐ€์ค‘์น˜ ๊ฐ์‡ ๋Š” 0.1, ์ตœ๋Œ€ ์ž„๊ณ„๊ฐ’ 1.0์˜ ๊ทธ๋ผ๋ฐ์ด์…˜ ํด๋ฆฌํ•‘, Yi-34B-Chat์˜ ๊ฒฝ์šฐ ๋…ธ์ด์ฆˆ ์Šค์ผ€์ผ 45, Yi-6B-Chat์˜ ๊ฒฝ์šฐ 5์˜ NEFTune [34]๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
ํŒŒ์ธํŠœ๋‹์—๋Š” ๋‹ค์Œ ๋‹จ์–ด ์˜ˆ์ธก ์†์‹ค์„ ์‚ฌ์šฉํ•˜๋ฉฐ, ์‹œ์Šคํ…œ ๋ฐ ์‚ฌ์šฉ์ž ์ง€์นจ์ด ์•„๋‹Œ ์‘๋‹ต์— ๋Œ€ํ•ด์„œ๋งŒ ์†์‹ค์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋ฒ ํƒ€_{1}\)์€ 0.9, \(\๋ฒ ํƒ€_{2}\)๋Š” 0.999, \(\์—ก\)์€ 10^{-8}\)๋กœ ์„ค์ •๋œ AdamW ์˜ตํ‹ฐ๋งˆ์ด์ €๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์‹œํ€€์Šค ๊ธธ์ด๋Š” 4096, ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 64๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋‹จ๊ณ„๋Š” 300์œผ๋กœ ์„ค์ •ํ•˜๊ณ , ํ•™์Šต ์†๋„๋Š” 1\๋ฐฐ 10^{-5}\), ๊ฐ€์ค‘์น˜ ๊ฐ์‡ ๋Š” 0.1, ์ตœ๋Œ€ ์ž„๊ณ„๊ฐ’ 1.0์˜ ๊ทธ๋ผ๋””์–ธํŠธ ํด๋ฆฌํ•‘, NEFTune [34]์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋…ธ์ด์ฆˆ ์Šค์ผ€์ผ์€ Yi-34B-Chat์˜ ๊ฒฝ์šฐ 45, Yi-6B-Chat์˜ ๊ฒฝ์šฐ 5์ž…๋‹ˆ๋‹ค.
๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•ด ๋‹ค์Œ ๋‹จ์–ด ์˜ˆ์ธก ์†์‹ค์„ ์‚ฌ์šฉํ•˜๋ฉฐ, ์‹œ์Šคํ…œ ๋ฐ ์‚ฌ์šฉ์ž ์ง€์นจ์ด ์•„๋‹Œ ์‘๋‹ต์— ๋Œ€ํ•ด์„œ๋งŒ ์†์‹ค์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. AdamW ์ตœ์ ํ™” ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, \(\beta_{1}\)์€ 0.9๋กœ ์„ค์ •ํ•˜๊ณ , \(\beta_{2}\)๋Š” 0.999๋กœ ์„ค์ •ํ•˜๋ฉฐ, \(\epsilon\)์€ \(10^{-8}\)๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ์‹œํ€€์Šค ๊ธธ์ด๋Š” 4096, ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 64๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต ๋‹จ๊ณ„๋Š” 300์œผ๋กœ ์„ค์ •ํ•˜๊ณ , ํ•™์Šต ์†๋„๋Š” 1\๋ฐฐ 10^{-5}\)๋กœ ์ผ์ •ํ•˜๊ฒŒ ์œ ์ง€ํ•˜๋ฉฐ, ๊ฐ€์ค‘์น˜ ๊ฐ์‡ ๋Š” 0.1, ์ตœ๋Œ€ ์ž„๊ณ„๊ฐ’์ด 1.0์ธ ๊ทธ๋ผ๋ฐ์ด์…˜ ํด๋ฆฌํ•‘, Yi-34B-Chat์˜ ๊ฒฝ์šฐ ๋…ธ์ด์ฆˆ ์Šค์ผ€์ผ์€ 45, Yi-6B-Chat์˜ ๊ฒฝ์šฐ 5์ธ NEFTune [34]์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
๋ฏธ์„ธ ์กฐ์ •์—๋Š” ๋‹ค์Œ ๋‹จ์–ด ์˜ˆ์ธก ์†์‹ค์„ ์‚ฌ์šฉํ•˜๋ฉฐ, ์‘๋‹ต์— ๋Œ€ํ•œ ์†์‹ค๋งŒ ๊ณ„์‚ฐํ•˜๊ณ  ์‹œ์Šคํ…œ ๋ฐ ์‚ฌ์šฉ์ž ์ง€์นจ์€ ๊ณ„์‚ฐํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. AdamW ์ตœ์ ํ™” ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, \beta_{1}\)์€ 0.9๋กœ ์„ค์ •ํ•˜๊ณ , \beta_{2}\)๋Š” 0.999๋กœ ์„ค์ •ํ•˜๋ฉฐ, \epsilon\)์€ \(10^{-8}\)๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ์‹œํ€€์Šค ๊ธธ์ด๋Š” 4096, ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 64๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋‹จ๊ณ„๋Š” 300์œผ๋กœ ์„ค์ •ํ•˜๊ณ , ํ•™์Šต ์†๋„๋Š” 1\๋ฐฐ 10^{-5}\), ๊ฐ€์ค‘์น˜ ๊ฐ์‡ ๋Š” 0.1, ์ตœ๋Œ€ ์ž„๊ณ„๊ฐ’์ด 1.0์ธ ๊ทธ๋ผ๋ฐ์ด์…˜ ํด๋ฆฌํ•‘, NEFTune [34]์€ Yi-34B-Chat์˜ ๊ฒฝ์šฐ ๋…ธ์ด์ฆˆ ์Šค์ผ€์ผ 45, Yi-6B-Chat์˜ ๊ฒฝ์šฐ 5๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.
README.md exists but content is empty. Use the Edit dataset card button to edit it.
Downloads last month
34
Edit dataset card