๐Ÿ”ฅ

  ะ ะตั‚ า›ะฐั€ะฐะปะดั‹ 14,994

ํ…Œ๋””๋…ธํŠธ TeddyNote

ํ…Œ๋””๋…ธํŠธ TeddyNote

ะšาฏะฝ ะฑาฑั€ั‹ะฝ

ะŸั–ะบั–ั€ะปะตั€: 142
@joonlab98
@joonlab98 7 ะฐะน ะฑาฑั€ั‹ะฝ
ํ•ญ์ƒ ์ข‹์€ ์˜์ƒ ๋งŒ๋“ค์–ด์ฃผ์…”์„œ ๋งŽ์€ ๋„์›€ ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค ํ…Œ๋””๋‹˜ ใ…Žใ…Ž ์ €๋ฅผ ๋น„๋กฏํ•˜์—ฌ ๋‹ค๋ฅธ ๊ตฌ๋…์ž ๋ถ„๋“ค์„ ์œ„ํ•˜์—ฌ Timestamp ์ƒ์„ฑํ•ด๋ณด์•˜์œผ๋‹ˆ ์ฐธ๊ณ ํ•˜์„ธ์š”:) โฐ Timestamp 00:00 ํŒŒ์ธํŠœ๋‹ ์†Œ๊ฐœ 00:07 ์˜ค๋Š˜์˜ ์˜ˆ์ œ: ๋ฌธ์„œ ๊ธฐ๋ฐ˜ ์งˆ์˜์‘๋‹ต ๋ชจ๋ธ 01:00 ํŒŒ์ธํŠœ๋‹ ํ”„๋กœ์„ธ์Šค ์„ค๋ช… 01:25 ๋ฌธ์„œ ๊ธฐ๋ฐ˜ ์งˆ์˜์‘๋‹ต ๋ชจ๋ธ ์ƒ์„ฑ ๊ณผ์ • 02:00 ๋ฐ์ดํ„ฐ์…‹์˜ ์ค‘์š”์„ฑ 03:00 QA Pair ๋ฐ์ดํ„ฐ์…‹ ์ €์žฅ ๋ฐฉ์‹: ๋กœ์ปฌ & ํ—ˆ๊น…ํŽ˜์ด์Šค ํ—ˆ๋ธŒ 03:35 ๋ชจ๋ธ ํŠœ๋‹: ๋ฒ ์ด์Šค ๋ชจ๋ธ + LoRA/QLoRA ์–ด๋Œ‘ํ„ฐ 04:44 ํŒŒ์ธํŠœ๋‹ ์™„๋ฃŒ๋œ ๋ชจ๋ธ ์ €์žฅ & ๋ณ€ํ™˜ 05:00 ์‹ค์ œ ํŒŒ์ธํŠœ๋‹๋œ ๋ชจ๋ธ ์˜ˆ์‹œ: TeddyNote 05:55 ๋ฐ์ดํ„ฐ์…‹ ์ƒ์„ฑ ์ฝ”๋“œ ์„ค๋ช… 07:17 JSONL ํ˜•์‹์œผ๋กœ ๋ฐ์ดํ„ฐ์…‹ ์ €์žฅ 07:30 ํ—ˆ๊น…ํŽ˜์ด์Šค ๋ฐ์ดํ„ฐ์…‹ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ํ™œ์šฉ 07:46 ๋ฐ์ดํ„ฐ์…‹ ์—…๋กœ๋“œ ์ค€๋น„ 08:20 ํ† ํฐ ๋ฐœ๊ธ‰ 08:50 ํŒŒ์ธํŠœ๋‹ ์ฝ”๋“œ ์„ค๋ช… 09:20 ์–ธ์Šฌ๋กœ์Šค(Unsloth) ์†Œ๊ฐœ 10:10 ๋ฒ ์ด์Šค ๋ชจ๋ธ ์„ ํƒ: ๋ผ๋งˆ 3 Open Ko 8B 11:09 ํŒŒ์ธํŠœ๋‹ ์‹คํ–‰ 12:14 ์ถ”๋ก  ์‹คํ–‰ 13:00 ๋ชจ๋ธ ์ €์žฅ 13:50 ํ—ˆ๊น…ํŽ˜์ด์Šค ํ—ˆ๋ธŒ์— ์—…๋กœ๋“œ 14:00 GGUF ํŒŒ์ผ ํ˜•์‹ ๋ณ€ํ™˜ 15:20 LM Studio์—์„œ ๋ชจ๋ธ ํ™œ์šฉ 16:00 ํ…Œ์ŠคํŠธ ์งˆ๋ฌธ 16:40 ํŒŒ์ธํŠœ๋‹ ์‹ค์Šต ์š”์•ฝ 17:00 ํŒŒ์ธํŠœ๋‹ ์ฃผ์˜์‚ฌํ•ญ 17:30 ๋งˆ๋ฌด๋ฆฌ
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
๊ฐ์‚ฌํ•ด์šฉ~~~ ๐Ÿ™
@michaelkwon8126
@michaelkwon8126 7 ะฐะน ะฑาฑั€ั‹ะฝ
๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
์•„์ด์ฟ  ํ›„์›์„.. ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค๐Ÿคค
@jjabul
@jjabul 3 ะฐะน ะฑาฑั€ั‹ะฝ
์ด๋ ‡๊ฒŒ ๋– ๋จน์—ฌ์ฃผ๋Š” ๊ฐ•์ขŒ๋ผ๋‹ˆ!! ํ…Œ๋””๋‹˜ ๋•๋ถ„์— ๊ทธ๋™์•ˆ ๋ง๋กœ๋งŒ ๋“ฃ๋˜ ํŒŒ์ธํŠœ๋‹์ด๋ž€ ๊ฑธ ์„ฑ๊ณตํ–ˆ๋„ค์š”. ํ•œ์ฐธ ๋Šฆ์€ ๋‚˜์ด์— ์ด๋Ÿฐ ๊ธฐ์ˆ ๋“ค ๋ฐฐ์šฐ๋Š๋ผ ๋‹ต๋‹ตํ–ˆ๋˜ ์™€์ค‘์— ํ•œ ์ค„๊ธฐ ๋น›์ด๋ž„๊นŒ์š”... ์ •๋ง ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
@teddynote
@teddynote 3 ะฐะน ะฑาฑั€ั‹ะฝ
์•„์ด์ฝ”~ ๋„ˆ๋ฌด๋‚˜ ๊ฐ์‚ฌํ•œ ๋Œ“๊ธ€์ž…๋‹ˆ๋‹ค. ์ œ ์˜์ƒ์„ ๋ณด๊ณ  ๊ตฌํ˜„ํ•˜์…จ๋‹ค๋Š” ๋Œ“๊ธ€ ๋ณด๋ฉด ์ฐธ ๋ฟŒ๋“ฏํ•ด์š”! ๋„ˆ๋ฌด ๊ณ ์ƒํ•˜์…จ์Šต๋‹ˆ๋‹ค^^
@SeonggyuLee
@SeonggyuLee 7 ะฐะน ะฑาฑั€ั‹ะฝ
LoRA ๋ง‰์—ฐํžˆ ๋ถ€๋‹ด์Šค๋Ÿฌ์šด ์ž‘์—…์ผ๊ฑฐ๋ผ ์ƒ๊ฐํ–ˆ๋Š”๋ฐ.. ์™€. ๋„ˆ๋ฌด ์ข‹์€ ๊ฐ•์˜ ์˜€์Šต๋‹ˆ๋‹ค~
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค!!
@heejuneAhn
@heejuneAhn 7 ะฐะน ะฑาฑั€ั‹ะฝ
ํ—ˆ๊น…ํŽ˜์ด์Šค๊ฐ€ PEFT ์ œ๊ณตํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ƒ๋‹นํžˆ ์ž‘์—…์ด ๊ฐ„๋‹จํ•ด ์ง„๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค.
@kunyoungkim-hw8uh
@kunyoungkim-hw8uh 6 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š”! ์˜์ƒ์„ ๋ณด๋ฉด์„œ ๊ถ๊ธˆํ•œ ๋ถ€๋ถ„์ด ์žˆ์–ด์„œ ์งˆ๋ฌธ๋“œ๋ฆฝ๋‹ˆ๋‹ค. 11:34 ์—์„œ instruction, output์€ ๋งคํ•‘์€ ํ•˜์ง€๋งŒ input์— ๊ฐ’์„ ๋„ฃ์„ ๊ฒฝ์šฐ์—๋Š” ์–ด๋–ค ๋ฐฉ์‹์œผ๋กœ ์ง„ํ–‰ํ•ด์•ผ ํ•˜๋‚˜์š”? alphacaprompt์—๋Š” instruction, response ์ด๋ ‡๊ฒŒ๋งŒ ์žˆ๋Š”๋ฐ Context๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ ์ง„ํ–‰ํ•˜๋‚˜์š”..?
@teddynote
@teddynote 6 ะฐะน ะฑาฑั€ั‹ะฝ
๋„ค ๋งž์Šต๋‹ˆ๋‹ค context ๋ฅผ ์ถ”๊ฐ€ํ•œ๋‹ค๊ณ  ์ดํ•ดํ•˜์…”๋„ ์ข‹์„ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๋‹ต๋ณ€์— ์ฐธ๊ณ ํ•  ์ˆ˜ ์žˆ๋Š” ์ด์ „ ๋ฌธ์žฅ์„ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฐœ๋…์ž…๋‹ˆ๋‹ค^^
@kunyoungkim-hw8uh
@kunyoungkim-hw8uh 6 ะฐะน ะฑาฑั€ั‹ะฝ
@@teddynote ํ•ญ์ƒ ์ข‹์€ ๊ฐ•์˜ ์˜ฌ๋ ค์ฃผ์‹œ๊ณ  ๋‹ต๋ณ€ํ•ด์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค!! ํ˜น์‹œ context๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š”๊ฒŒ alphaca prompt์˜ ํ˜•ํƒœ๊ฐ€ instruction, response ๋งŒ ์กด์žฌํ•˜๋Š”๋ฐ context๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์ผ๊นŒ์š”? ์•„๋‹ˆ๋ฉด instruction์— instruction + ' ' + context๋ฅผ ํ•จ๊ป˜ ๋„ฃ์–ด์ฃผ๋Š” ๋ฐฉ์‹์œผ๋กœ ์ง„ํ–‰ํ•ด์•ผ ํ•˜๋‚˜์š”? ์ œ๊ฐ€ ํ•ด๋ดค๋Š”๋ฐ ์ถ”๋ก ๋‹จ๊ณ„์—์„œ ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ•ด์„œ ํ•™์Šต๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ ๋‹จ๊ณ„์—์„œ ์‹ค์ˆ˜ํ•œ๊ฒŒ ์žˆ๋Š”์ง€ ๊ถ๊ธˆํ•ด์„œ ์—ฌ์ญค๋ด…๋‹ˆ๋‹ค! ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค
@๋ณด์„-n1n
@๋ณด์„-n1n 7 ะฐะน ะฑาฑั€ั‹ะฝ
์†Œ์ค‘ํ•œ ์ง€์‹ ๊ณต์œ  ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
์‹œ์ฒญํ•ด ์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค๐Ÿ™
@leehg090909able
@leehg090909able 4 ะฐะน ะฑาฑั€ั‹ะฝ
์–ธ์ œ๋‚˜ ์ตœ๊ณ ์ž…๋‹ˆ๋‹ค
@teddynote
@teddynote 4 ะฐะน ะฑาฑั€ั‹ะฝ
๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค!!
@prism79
@prism79 ะะน ะฑาฑั€ั‹ะฝ
๊ตญ๊ฐ€ ์—ฐ๊ตฌ ๋ณด๊ณ ์„œ ๋ชจ๋ธ์„ ์•Œ๋ ค์ฃผ์‹  ๋‚ด์šฉ์„ ์ด์šฉํ•˜์—ฌ ๋นŒ๋“œํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
@teddynote
@teddynote ะะน ะฑาฑั€ั‹ะฝ
๋ฝœ์ดํŒ… ์ž…๋‹ˆ๋‹ค!!
@jangbongseob
@jangbongseob 7 ะฐะน ะฑาฑั€ั‹ะฝ
์ข‹์€ ๋‚ด์šฉ ๊ณต์œ ํ•ด ์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
์‹œ์ฒญํ•ด ์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ด์š”๐Ÿ™
@moonsookim7596
@moonsookim7596 7 ะฐะน ะฑาฑั€ั‹ะฝ
๊ฐ•์˜ ์ž˜ ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
์‹œ์ฒญํ•ด ์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค๐Ÿ™
@Urban๋„์‹œ
@Urban๋„์‹œ 5 ะฐะน ะฑาฑั€ั‹ะฝ
์ •๋ง ์œ ์ตํ•˜๊ณ  ์ข‹์€ ์˜์ƒ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.. ์„ธ์ƒ์—๋Š” ๊ณ ์ˆ˜๋ถ„๋“ค์ด ์ •๋ง ๋งŽ์€๊ฒƒ ๊ฐ™์•„์š”.. ์ž˜ ๋ฐฐ์šฐ๊ณ  ๊ฐ‘๋‹ˆ๋‹ค!!
@teddynote
@teddynote 5 ะฐะน ะฑาฑั€ั‹ะฝ
๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค๐Ÿ™
@cjj-yx9sn
@cjj-yx9sn 7 ะฐะน ะฑาฑั€ั‹ะฝ
ํ•œ๋‹ฌ๋™์•ˆ gguf ๋ณ€ํ™˜๋•Œ๋ฌธ์— ํ•ด๋งธ๋Š”๋ฐ.. ๋„ˆ๋ฌด ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค!!โคโค
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค~^^
@juliokim766
@juliokim766 7 ะฐะน ะฑาฑั€ั‹ะฝ
aya๋ฅผ ollama์— ์˜ฌ๋ ค ๋ณด๋ ค๊ณ  ๊ตฌํ–ˆ๋˜ gguf๊ฐ€ ollama์— ์˜ฌ๋ฆด๋•Œ ๊ณ„์† ์˜ค๋ฅ˜๊ฐ€ ๋‚˜์„œ ์‚ฌ์šฉ์„ ๋ชปํ•ด๋ณด๊ณ ์žˆ์—ˆ๋Š”๋ฐ, ์ด ๋ฐฉ๋ฒ•์œผ๋กœ ์ง์ ‘ ๋ณ€๊ฒฝํ•ด๋ด์•ผ๊ฒ ๋„ค์š”~~^^ ํ•ญ์ƒ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค~:)
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
์š”์ฆ˜ aya๊ฐ€ ๋Œ€์„ธ ์ด๋”๋ผ๊ณ ์š” ใ…Žใ…Ž
@heejuneAhn
@heejuneAhn 7 ะฐะน ะฑาฑั€ั‹ะฝ
๋‹ค์Œ๊ณผ ๊ฐ™์ด ํ•˜๋‹ˆ ์ž˜ ๋ฉ๋‹ˆ๋‹ค. print("saving the GGUF model...") quantization_method = "q8_0" # "f16" "q8_0" "q4_k_m" "q5_k_m" peft_model.save_pretrained_gguf( "gguf", tokenizer=tokenizer, quantization_method=quantization_method, ) print("Try with ollama!") ''' 1) copy a reference Model file and moify the gguf file in the first line 2) ollama create -f Modelfile 3) ollama run '''
@kookyunghyun
@kookyunghyun 6 ะฐะน ะฑาฑั€ั‹ะฝ
16:07 ๋ถ€๋ถ„์—์„œ ์ง€๊ธˆ๊นŒ์ง€ ํ•™์Šต๋ฐ ๋ฒ ์ด์Šค๋ชจ๋ธ์ด ์ „๋ถ€ ๋ผ๋งˆ3์ธ๋ฐ ํ”„๋ฆฌ์…‹์„ ์•ŒํŒŒ์นด๋กœ ํ•ด์•ผํ•˜๋Š” ์ด์œ ๊ฐ€ ๋ฌด์—‡์ธ๊ฐ€์š”?? ์„ค๋ช…๋„ ์•ŒํŒŒ์นด๋กœ ์‚ฌ์šฉํ•˜์…”์•ผ ํ•œ๋‹ค๊ณ  ํ•˜์‹œ๋Š”๋ฐ,.
@teddynote
@teddynote 6 ะฐะน ะฑาฑั€ั‹ะฝ
์•ŒํŒŒ์นด๋กœ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ ํ•™์Šต์„ ํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ถ”๋ก ๋„ ๋™์ผํ•œ ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ์„ ์‚ฌ์šฉํ•ด์•ผ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๋‹ต๋ณ€ํ•ฉ๋‹ˆ๋‹ค~
@teddynote
@teddynote 6 ะฐะน ะฑาฑั€ั‹ะฝ
์• ์ดˆ์— ํ•™์Šตํ• ๋•Œ์˜ ํ…œํ”Œ๋ฆฟ์€ ์ž์œ ๋กญ๊ฒŒ ์„ ํƒ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค ๋Œ€์‹  ์ถ”๋ก ํ• ๋•Œ์˜ ํ…œํ”Œ๋ฆฟ๊ณผ ํ†ต์ผ์€ ํ•ด์•ผํ•ฉ๋‹ˆ๋‹ค
@jungkyle4561
@jungkyle4561 4 ะบาฏะฝ ะฑาฑั€ั‹ะฝ
์˜์ƒ ์ž˜๋ณด๊ณ ์žˆ์Šต๋‹ˆ๋‹ค! ํ›ˆ๋ จ๊ฒฐ๊ณผ๊ฐ€ ๋งŒ์กฑ์Šค๋Ÿฝ์ง€ ์•Š์€๋ฐ,, QNA ํ˜•ํƒœ๋ง๊ณ  RAWTEXT ๋Š” ํ•™์Šต์ด ์–ด๋ ค์šธ๊นŒ์š”..? 1000 STEP์ •๋„ ํ›ˆ๋ จ์‹œ์ผฐ๋Š”๋ฐ ์ „ํ˜€ ์—‰๋šฑํ•œ ๋‹ต๋ณ€์„ํ•˜๋„ค์š”,,
@ikkkiaffgesk
@ikkkiaffgesk 7 ะฐะน ะฑาฑั€ั‹ะฝ
์†Œ์ค‘ํ•œ ๊ฐ•์˜ ๊ฐ์‚ฌ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฒˆ ๊ฐ•์˜๋Š” gpu vram์ด ์–ผ๋งˆ๋‚˜ ํ•„์š” ํ•˜๋‚˜์š”?
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
๋Œ€๋žต 15~20G ์‚ฌ์ด์ž…๋‹ˆ๋‹ค! ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค^^
@isy9942
@isy9942 5 ะฐะน ะฑาฑั€ั‹ะฝ
์ข‹์€ ์˜์ƒ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค ํ…Œ๋””๋‹˜. ํ•œ๊ฐ€์ง€ ์งˆ๋ถ„์ด ์žˆ์–ด ์—ฌ์ญค๋ณด๋ ค ๋Œ“๊ธ€์„ ๋‚จ๊น๋‹ˆ๋‹ค. ํ˜น์‹œ ๋ฒ ์ด์Šค๋ชจ๋ธ๊ณผ Lora ์–ด๋Œ‘ํ„ฐ๋ถ€๋ถ„์„ ๋ณ‘ํ•ฉํ•ด์„œ ์ €์žฅํ•œ 1์ฐจ ํ•™์Šต ๋ชจ๋ธ์„, 2์ฐจ ํ•™์Šต์‹œ์— ๋‹ค์‹œ ๋ฒ ์ด์Šค๋ชจ๋ธ๋กœ ์„ค์ •ํ•ด ์žฌํ•™์Šต์‹œํ‚ค๋Š” ๊ฒƒ์ด ๊ฐ€๋Šฅํ•œ์ง€ ์—ฌ์ญค๋ณด๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ๋Š” 2,000 ๋ฌธ์žฅ ์ •๋„ ๋˜๋Š” ๋ฐ์ดํ„ฐ๋ฅผ 500๋ฌธ์žฅ์”ฉ ๋Š์–ด์„œ ํ•™์Šต์‹œํ‚ค๊ณ  ์žˆ์ง€๋งŒ, ๊ธฐ์กด ํ•™์Šต์‹œํ‚จ ๋ณ‘ํ•ฉ๋ชจ๋ธ์„ ๋ฒ ์ด์Šค๋ชจ๋ธ๋กœ ์ƒˆ๋กœ ํ•™์Šต์‹œํ‚จ ๋ชจ๋ธ์—์„œ๋Š”, ์ด์ „์— ํ•™์Šต์‹œํ‚จ ๋ฌธ์žฅ๋“ค์˜ response๊ฐ€ ๋™์ผํ•˜๊ฒŒ ๋„์ถœ๋˜์ง€ ์•Š๊ณ ์žˆ๋„ค์š”. ํ•™์Šต์ด ๋ฎ์–ด์“ฐ๊ธฐ ๋˜๋Š” ๊ฒƒ ๊ฐ™๋‹ค๋Š” ๋Š๋‚Œ์ž…๋‹ˆ๋‹ค. ํ˜น์‹œ ์ง€์†์ ์ธ ํ•™์Šต์ด ๊ฐ€๋Šฅํ•œ ๋ฐฉ์‹์ด ์žˆ๋‹ค๋ฉด ๊ด€๋ จ๋œ ํ‚ค์›Œ๋“œ๋ฅผ ์—ฌ์ญค๋ณด๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค.
@teddynote
@teddynote 5 ะฐะน ะฑาฑั€ั‹ะฝ
๋„ค ๋ณ‘ํ•ฉํ•ด์„œ ๋ฒ ์ด์Šค๋กœ ๋‘์‹œ๊ณ ์š” ๋กœ๋ผ ํŠœ๋‹ ๊ฐ€๋Šฅํ•œ๋ฐ์š”. ๋กœ๋ผ ํ•™์Šต ํ•˜๋ ค๋ฉด ์ตœ์†Œ 3๋งŒ๊ฑด ์ด์ƒ์€ ๋˜์–ด์•ผ ์–ด๋Š ์ •๋„ ๋‚˜์˜ฌ ๊ฒƒ์ด๋ผ ํ•ฉ๋‹ˆ๋‹ค(์ด ๋ถ€๋ถ„์€ ํŒŒ์ธํŠœ๋‹ ์ „๋ฌธ๊ฐ€์™€ ์ง„ํ–‰ํ•œ ๋ผ์ด๋ธŒ ์„ธ์…˜ ์ฐธ๊ณ ํ•ด ์ฃผ์„ธ์š”) ์ ์€ ๋ฐ์ดํ„ฐ๋กœ ํ•ด๋ณด์‹ค๊บผ๋ฉด ํ’€ํŒŒ์ธํŠœ๋‹์„ ์‹œ๋„ํ•ด ๋ณด์„ธ์š”!
@luv8216
@luv8216 7 ะฐะน ะฑาฑั€ั‹ะฝ
์ข‹์€ ์˜์ƒ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค! ๊ทธ๋Ÿฐ๋ฐ ํ˜น์‹œ torch๋ฅผ ์‚ฌ์šฉํ•˜์‹œ๋Š” ์ด์œ ๊ฐ€ ์žˆ์œผ์‹ค๊นŒ์š”? ์ดˆ๋ณด์ ์ธ ์งˆ๋ฌธ์ธ ๊ฒƒ ๊ฐ™๊ธด ํ•œ๋ฐ ๋‹ค๋ฅธ ํŒŒ์ธํŠœ๋‹ ์˜์ƒ๋“ค๋„ ๋Œ€๋ถ€๋ถ„ torch๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ ๊ฐ™์•„์„œ ์งˆ๋ฌธ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
์ด๋ฏธ ์งˆ๋ฌธ์— ๋‹ต์ด ์žˆ์Šต๋‹ˆ๋‹ค~ ๋Œ€๋ถ€๋ถ„์˜ ํŒŒ์ธํŠœ๋‹ ์ฝ”๋“œ๊ฐ€ torch ๊ธฐ๋ฐ˜์œผ๋กœ ์ž‘์„ฑ๋˜์–ด ์žˆ๊ณ  ์ฐธ๊ณ ํ•  ์ˆ˜ ์žˆ๋Š” ์˜ˆ์ œ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค^^ torch๋กœ integration ๋„ ํŽธํ•˜๊ตฌ์š”!
@heejuneAhn
@heejuneAhn 7 ะฐะน ะฑาฑั€ั‹ะฝ
์ผ๋‹จ ์‹ค์งˆ์ ์œผ๋กœ Huggingface๋Š” PyTorch ๊ธฐ๋ฐ˜์— ๊ฐ€๊น๋‹ค๊ณ  ๋ณด์‹œ๋Š”๊ฒŒ ์ข‹์๋‹ˆ๋‹ค.
@gangin0218
@gangin0218 2 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š” ์–‘์งˆ์˜ ์˜์ƒ ํ•ญ์ƒ ๋„์›€์„ ๋งŽ์ด ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜์ƒ ํ›„๋ฐ˜๋ถ€์— Alpaca Prompt๋กœ preset์„ ๋ฐ”๊พธ๋Š” ๋ถ€๋ถ„์ด ์žˆ๋Š”๋ฐ ์ €๋Š” preset์ด ์—†์–ด์„œ์š”. ํ˜น์‹œ alpaca prompt๋Š” ์–ด๋””์„œ ๊ตฌํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ๊ตฌ๊ธ€๋ง์œผ๋กœ๋„ ์ž˜ ๋ชป์ฐพ๊ฒ ์–ด์„œ์š”...
@teddynote
@teddynote 2 ะฐะน ะฑาฑั€ั‹ะฝ
์ œ๊ฐ€ ๊ณต์œ  ๋“œ๋ฆด๊ป˜์š”^^ www.dropbox.com/scl/fi/kyofhwx6hzfgyuv21mis7/alpaca.preset.json?rlkey=x9ri1iyqu1m74dig8949g2ydz&st=ixasygix&dl=1
@HJ-jf4co
@HJ-jf4co 6 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š” ํ•ด๋‹น ์˜์ƒ๋ณด๋ฉด์„œ ํŒŒ์ธํŠœ๋‹ ๊ณต๋ถ€์ค‘์ธ๋ฐ ์งˆ๋ฌธ์ด์žˆ์–ด์„œ ๋Œ“๊ธ€ ๋‚จ๊น๋‹ˆ๋‹ค. ์›ํ•˜๋Š” ๋ฐ์ดํ„ฐ์…‹์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฒ ์ด์Šค ๋ชจ๋ธ๊ณผ ๋ณ‘ํ•ฉํ•˜์—ฌ ํŒŒ์ธํŠœ๋‹ ๋ชจ๋ธ์„ ์ƒ์„ฑํ•˜๊ฒŒ ๋˜๋ฉด, ๋‹ต๋ณ€์ด ํ•™์Šต์‹œํ‚จ ๋ฐ์ดํ„ฐ์— ๋งŽ์ด ์น˜์šฐ์นœ ๋Œ€๋‹ต์„ ํ•˜๊ฒŒ๋ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ๋ฒ ์ด์Šค ๋ชจ๋ธ๋กœ ํ…Œ์ŠคํŠธ ํ•ด๋ดค์„๋•Œ ์ •์ƒ์ ์œผ๋กœ ํ•˜๋˜ ๋‹ต๋ณ€๋„ ํŒŒ์ธํŠœ๋‹ ํ›„์—๋Š” ์ œ๊ฐ€ ํ•™์Šต์‹œํ‚จ ๋ฐ์ดํ„ฐ์™€ ๊ด€๋ จ๋œ ๋Œ€๋‹ต๋งŒ ํ•˜๊ฒŒ๋ฉ๋‹ˆ๋‹ค.. ์ œ๊ฐ€ ๊ธฐ๋Œ€ํ–ˆ๋˜ ํŒŒ์ธํŠœ๋‹ ๊ฒฐ๊ณผ๋ฌผ์€ ๊ธฐ์กด ๋ฒ ์ด์Šค ๋ชจ๋ธ์ด ๊ฐ€์ง€๊ณ ์žˆ๋˜ ๋ฐ์ดํ„ฐ + ์ผ๋ถ€ ๋‚ด๊ฐ€ ํ•™์Šต์‹œํ‚จ ๋ฐ์ดํ„ฐ ์ธ๋ฐ, ์ œ๊ฐ€ ์ž˜๋ชป์ดํ•ดํ•˜๊ณ  ์žˆ์—ˆ๋˜๊ฑด์ง€ ์˜์ƒ์† ๋ฐฉ๋ฒ•์œผ๋กœ๋Š” ์ œ๊ฐ€ ํ•™์Šต์‹œํ‚จ ๋ฐ์ดํ„ฐ์— ์น˜์šฐ์นœ ๊ฒฐ๊ณผ๋ฌผ๋งŒ ๋‚ผ ์ˆ˜ ์žˆ๋Š”๊ฑด์ง€ .. ์•„๋‹ˆ๋ฉด ์ œ๊ฐ€ ํŒŒ์ธํŠœ๋‹์„ ์ž˜๋ชปํ•œ๊ฑด์ง€.... ๋‹ต๋ณ€์ด ๊ฐ€๋Šฅํ• ๊นŒ์š”...
@teddynote
@teddynote 6 ะฐะน ะฑาฑั€ั‹ะฝ
๋งž์Šต๋‹ˆ๋‹ค ๊ทน๋‹จ์ ์œผ๋กœ ์˜ˆ์‹œ๋ฅผ ๋ณด์—ฌ๋“œ๋ฆฌ๋ ค๊ณ  ์ž‘์€ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ โ€˜๊ณผ์ ํ•ฉโ€™ ์‹œ์ผœ์„œ ๊ทธ๋ ‡๊ตฌ์š”~ ์›๋ž˜ ์ œ๋Œ€๋กœ ํ•˜์‹œ๋ ค๋ฉด ๋ฐ์ดํ„ฐ์…‹ ๋ณดํ†ต 2~3๋งŒ๊ฑด ์ •๋„( ๋น„์šฉ์ด ๋งŽ์ด ๋“ค ์ˆ˜ ์žˆ์–ด์š”...) ๊ทธ๋ฆฌ๊ณ  rank ๋ž‘ alpha ๋Š” ๋” ์ž‘๊ฒŒ ์ฃผ์…”์„œ ๋ชจ๋ธ์ด ํ•œ ๋งˆ๋””๋กœ ๋งŽ์ด ํ‹€์–ด์ง€์ง€ ์•Š๋„๋ก ํ•ด์ฃผ์‹œ๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค. ๋น„์šฉ๊ณผ ํ•™์Šต์‹œ๊ฐ„์ด ๋งŽ์ด ๋“œ๋‹ˆ ์ด ์˜์ƒ์€ ๋ฏธ๋‹ˆ๋กœ ์ง„ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค! ์ฐธ๊ณ ํ•ด ์ฃผ์‹ญ์‡ผ^^
@์œ ์˜์žฌ-c9c
@์œ ์˜์žฌ-c9c 4 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š”. QA ๋ฐ์ดํ„ฐ์…‹์„ LLM ๋ชจ๋ธ์— ํ›ˆ๋ จ์‹œํ‚ฌ ๋•Œ, ์ตœ์ข…์ ์œผ๋กœ RAG๋ชจ๋ธ์— ํ™œ์šฉํ•˜๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ, input ํ‚ค(์ธ์ŠคํŠธ๋Ÿญ์…˜ ํŠœ๋‹ํ•  ๋•Œ..)์œผ๋กœ retriever๊ฐ€ ๊ฒ€์ƒ‰ํ•œ ๋ฌธ์žฅ๋“ค์„ ํฌํ•จ์‹œํ‚ค๋Š” ๊ฒƒ์ด ๋งž์„๊นŒ์š”? ๋งŒ์•ฝ ์ž…๋ ฅ์— retriever ๋ฌธ์žฅ์ด ํฌํ•จ๋˜์ง€ ์•Š์œผ๋ฉด, ๋ชจ๋ธ์ด ๋‹ต๋ณ€ ์Šคํƒ€์ผ๋งŒ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณผ ์ˆ˜ ์žˆ์„๊นŒ์š”?"
@teddynote
@teddynote 4 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š”! ๋‹ต๋ณ€ ๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. retriever ๊ฒ€์ƒ‰๋œ ๊ฒฐ๊ณผ๊ฐ€ ํฌํ•จ์ด ๋˜๊ณ  ์ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•™์Šตํ•œ๋‹ค๋ฉด, ์ฃผ์–ด์ง„ ๋ฌธ๋งฅ์—์„œ ๋‹ต๋ณ€์„ ๋„์ถœํ•˜๋Š” ๊ฒƒ์„ ํ•™์Šตํ•˜๊ณ ์š”, ๋งŒ์•ฝ์— retriever ๊ฒฐ๊ณผ๊ฐ€ ํฌํ•จ๋˜์ง€ ์•Š๋Š”๋‹ค๋ฉด, ๋ง์”€ํ•˜์‹  ๊ฒƒ์ฒ˜๋Ÿผ ๋‹ต๋ณ€ ์Šคํƒ€์ผ์„ ํ•™์Šตํ•œ๋‹ค๊ณ  ๋ณด์…”๋„ ์ข‹์„ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค!
@์ •์ง-o7p
@์ •์ง-o7p ะะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š” ํ…Œ๋””๋‹˜. ๋กœ์ปฌ์—์„œ ์‹ค์Šต์„ ์ง„ํ–‰ํ•˜๋˜ ์ค‘์— GPU ๋ฉ”๋ชจ๋ฆฌ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•˜์—ฌ ๋ฌธ์˜๋“œ๋ฆฝ๋‹ˆ๋‹ค. RuntimeError: CUDA driver error: out of memory 0%| | 0/100 [00:02
@๋ฒ„์ฐŒ์กฐ์•„
@๋ฒ„์ฐŒ์กฐ์•„ 6 ะฐะน ะฑาฑั€ั‹ะฝ
ํ…Œ๋””๋…ธํŠธ ์˜์ƒ๋“ค์„ ๋กœ์ปฌ๋กœ ์‹คํ–‰ํ•˜๋ ค๋ฉด ์–ด๋Š์ •๋„ ์‚ฌ์–‘์ด ๋˜์–ด์•ผ ํ•˜๋‚˜์š”?
@teddynote
@teddynote 6 ะฐะน ะฑาฑั€ั‹ะฝ
๋กœ์ปฌ ๋ชจ๋ธ ๊ตฌ๋™์„ ์œ„ํ•ด์„œ๋Š” GPU ๊ฐ€ ํ•„์š”ํ•˜์ง€๋งŒ GPT4, Claude API ๋ฅผ ์‚ฌ์šฉํ•˜์‹œ๋Š” ๊ฒฝ์šฐ ๊ฐ€๋ฒผ์šด ๋…ธํŠธ๋ถ์œผ๋กœ๋„ ์ถฉ๋ถ„ํ•ฉ๋‹ˆ๋‹ค!
@byungtaekim2753
@byungtaekim2753 12 ะบาฏะฝ ะฑาฑั€ั‹ะฝ
ํ•ญ์ƒ ์ข‹์€ ์˜์ƒ ๋งŒ๋“ค์–ด์ฃผ์…”์„œ ๋งŽ์€ ๋„์›€ ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. 1๋ฒˆ qa pair ์…‹ ์‹คํ–‰ํ•˜๋ ค๊ณ  ํ•˜๋Š”๋ฐ ๊ณ„์† ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์—๋Ÿฌ๊ฐ€ ๋‚˜์„œ ๋ฌธ์˜ ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ํ˜น์‹œ ํ˜„์žฌ ์‹œ์ ์— ์„ค์น˜ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ• ๊ฐ€์ด๋“œ ์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.~ pip install ~~~
@์žฅ์žฌ์€-b6i
@์žฅ์žฌ์€-b6i 2 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š”, ํ…Œ๋””๋‹˜! ํ…Œ๋””๋‹˜ ์˜์ƒ์„ ๋ณด๋ฉฐ ์—…๋ฌด์— rag๋ฅผ ํ›จ์”ฌ ์ˆ˜์›”ํ•˜๊ฒŒ ์ ์šฉํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ํ•ญ์ƒ ์œ ์ตํ•œ ๋‚ด์šฉ์„ ๊ณต์œ ํ•ด์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ถ๊ธˆํ•œ ๊ฒƒ์ด ํ•˜๋‚˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ์ž๋ฃŒ๋“ค์—์„œ rag๋ฅผ ์œ„ํ•œ ํŒŒ์ธํŠœ๋‹ ์—†์ด ๋ฐ”๋กœ llm์— rag๋ฅผ ์ ์šฉํ•˜๋Š”๋ฐ, ํŠน์ • ํ…Œ์Šคํฌ์—์„œ๋Š” ํŒŒ์ธํŠœ๋‹์ด ํ•„์š”ํ• ๋งŒํ•œ ์ƒํ™ฉ์ด ์žˆ์„๊ฑฐ๋ผ๊ณ  ์ƒ๊ฐ์ด๋ฉ๋‹ˆ๋‹ค. ํ…Œ๋””๋‹˜๊ป˜์„œ๋„ llm ํŒŒ์ธํŠœ๋‹ ํ›„ rag๋ฅผ ์ ์šฉํ•œ ๊ฒฝํ—˜์ด ์žˆ์œผ์‹ค๊นŒ์š”? ๋งŒ์•ฝ ์žˆ์œผ์‹œ๋‹ค๋ฉด ํ•ด๋‹น ๊ฒฝํ—˜์„ ๊ณต์œ ํ•ด์ฃผ์‹ค ์ˆ˜ ์žˆ์„๊นŒ์š”?
@Acb-1015
@Acb-1015 3 ะฐะน ะฑาฑั€ั‹ะฝ
์ข‹์€ ์˜์ƒ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค! ์ง€๊ธˆ์€ ํ…์ŠคํŠธ ๋ฌธ์„œ๋ฅผ ์‚ฌ์šฉํ–ˆ๋Š”๋ฐ ์ด๋ฏธ์ง€๋ฅผ ์ด์šฉํ•˜์—ฌ ํŒŒ์ธํŠœ๋‹์„ ํ•˜๊ณ  ์‹ถ์€๋ฐ ์ฐธ๊ณ ํ•  ์ˆ˜ ์žˆ๋Š” ์ž๋ฃŒ๊ฐ€ ์žˆ์„๊นŒ์š”?
@teddynote
@teddynote 3 ะฐะน ะฑาฑั€ั‹ะฝ
๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ชจ๋Œˆ ํŒŒ์ธํŠœ๋‹์„ ๋ง์”€ํ•˜๊ธฐ๋Š” ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. llava model ํŒŒ์ธํŠœ๋‹์ชฝ์„ ํ•œ ๋ฒˆ ๋ณด์‹œ๊ฒ ์–ด์š”? kzbin.info/www/bejne/m3rdmoFohrqnbrcsi=6iLvJk1HhCoek3Rv
@Acb-1015
@Acb-1015 3 ะฐะน ะฑาฑั€ั‹ะฝ
@@teddynote ๋‹ต๋ณ€ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค!
@teddynote
@teddynote 3 ะฐะน ะฑาฑั€ั‹ะฝ
@@Acb-1015 ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค๐Ÿซฐ
@ํ•œ์ƒ˜์งฑ
@ํ•œ์ƒ˜์งฑ 5 ะฐะน ะฑาฑั€ั‹ะฝ
EEVE ๋Œ€์‹  ๋ฒ”์ด๋‹˜๊บผ ์“ฐ์‹  ์ด์œ ๊ฐ€ ์žˆ์œผ์‹ค๊นŒ์šฉ?
@teddynote
@teddynote 5 ะฐะน ะฑาฑั€ั‹ะฝ
llama3 ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์˜€์Šต๋‹ˆ๋‹ค~
@_skim4403
@_skim4403 5 ะฐะน ะฑาฑั€ั‹ะฝ
์ œ๊ฐ€ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๋ฐ์ดํ„ฐ์…‹์—์„œ ์ง€๋ฌธ์‹œํ€€์Šค ๊ธธ์ด๊ฐ€ 2000์ž์ •๋„๋กœ ์ตœ๋Œ€ ํ† ํฐ๊ธธ์ด๊ฐ€ 4000์ž ์ •๋„ ๋‚˜์˜ค๋Š”๋ฐ ์—ฌ๊ธฐ ์˜์ƒ์—์„œ ๋‚˜์˜ค๋Š” ๊ฒƒ์„ ๋”ฐ๋ผํ•ด๋„ ๊ดœ์ฐฎ์„๊นŒ์š”? ์Šฌ๋ผ์ด๋”ฉ ์œˆ๋„์šฐ๋‚˜ longfomer๋ชจ๋ธ ์‚ฌ์šฉ์•ˆํ•˜๊ณ  ๊ฐ€๋Šฅํ• ์ง€ ์—ฌ์ญค๋ด…๋‹ˆ๋‹ค ใ… ใ… 
@teddynote
@teddynote 5 ะฐะน ะฑาฑั€ั‹ะฝ
context length ๊ฐ€ 4096 limit ์ด ๊ฑธ๋ ค ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋‹ต๋ณ€ ์ถœ๋ ฅ์„ ์œ„ํ•œ ํ† ํฐ์ด ์ œํ•œ์ ์ด๋ผ ์ž˜ ์•ˆ๋‚˜์˜ฌ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค ใ… 
@_skim4403
@_skim4403 5 ะฐะน ะฑาฑั€ั‹ะฝ
@@teddynote ์–ด ์ œ๊ฐ€ ๊ฐ€์ง„ ์ง€๋ฌธ๋ฐ์ดํ„ฐ์—์„œ context ์ตœ๋Œ€ ๊ธธ์ด๋Š” 2000์ž์ •๋„์ธ๋ฐ ๊ทธ๋Ÿผ 4096๋ณด๋‹ค ์งง์•„์„œ ๊ดœ์ฐฎ์€๊ฑฐ ์•„๋‹Œ๊ฐ€์š”? ใ…  ์ œ๊ฐ€ ์ž˜ ๋ชฐ๋ผ์„œ์šฉ
@heejuneAhn
@heejuneAhn 7 ะฐะน ะฑาฑั€ั‹ะฝ
A few comments in following the code. 1. an error with mismatch of peft version and unsloth, please give us the versions for the packages. ( I manually sloved the issues. No module named 'peft.tuners.lora.layer'; 'peft.tuners.lora' is not a package) (my system uses unsloth 2.0 and peft 0.5.0) 2. model name should be full path (model_name = "unsloth/llama-3-8b-bnb-4bit", not model_name="llama-3-8b-bnb-4bit")
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
1. you can find easy install commands @ link below github.com/unslothai/unsloth/tree/main#installation-instructions It's very hard to specify the exact version of each library due to dependencies (CUDA, Python, Torch, etc, ...) try using mamba if you haven't tried it yet 2. thx for letting me know. I have updated the .ipynb file
@heejuneAhn
@heejuneAhn 7 ะฐะน ะฑาฑั€ั‹ะฝ
@@teddynote the dependent packages are quite system-dependent. colab-new = [ "tyro", "transformers>=4.38.2", "datasets>=2.16.0", "sentencepiece", "tqdm", "psutil", "wheel>=0.42.0", "numpy", "protobuf
@์œ ์˜์žฌ-c9c
@์œ ์˜์žฌ-c9c 4 ะฐะน ะฑาฑั€ั‹ะฝ
์‹ค์ œ trainํ•  ๋•Œ loss๋Š” ๋””ํดํŠธ๋กœ ์–ด๋–ค loss ๊ณ„์‚ฐ์‹์ธ์ง€ ์•Œ ์ˆ˜ ์žˆ๋‚˜์š”?
@teddynote
@teddynote 4 ะฐะน ะฑาฑั€ั‹ะฝ
cross entropy ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์•Œ๊ณ  ์žˆ์–ด์š”! ์ด ๊ธ€์„ ์ฐธ๊ณ ํ•ด ๋ณด์„ธ์š”!! medium.com/@jeraldteokj/visualising-loss-calculation-in-large-language-models-1af410a9d73d
@fruitmango8420
@fruitmango8420 6 ะฐะน ะฑาฑั€ั‹ะฝ
๋งฅ์—์„œ cuda๊ฐ€ ์•„๋‹ˆ๋”๋ผ๋„ mps๋กœ pytorch ๊ฐ€ ๋˜๋˜๋ฐ ๋ญ”๊ฐ€ ํŒŒ์ธํŠœ๋‹์—์„œ mps๋กœ ํ• ๋ คํ•˜๋ฉดํ˜ธํ™˜์„ฑ๋ฌธ์ œ์ธ์ง€ ์—๋Ÿฌ์ฝ”๋“œ๊ฐ€ ์šฐ์ˆ˜์ˆ˜ ๋œจ๋˜๋ฐ ํ˜น์‹œ ๋ฐฉ๋ฒ•์ด ์—†๋‚˜์š”? m4๋„ ๋‚˜์˜ค๊ณ  ํ†ตํ•ฉ๋ฉ”๋ชจ๋ฆฌ๋ผ 32gb ๋ชจ๋ธ์“ฐ๋ฉด ๊ทธ๋ž˜ํ”ฝ์—์„œ๋„ ์ข‹์€์„ฑ๋Šฅ ๋‚˜์˜ฌ๊ฑฐ๊ฐ™์•„์„œ ํœด๋Œ€์šฉ ์ค‘์—์„  ๊ดœ์ฐฎ๋‹ค ์ƒ๊ฐํ•ด์„œ ๋งฅ์œผ๋กœ ์ค€๋น„ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค
@teddynote
@teddynote 6 ะฐะน ะฑาฑั€ั‹ะฝ
ํŒŒ์ธํŠœ๋‹์—์„œ๋Š” mps ๋ฅผ ๋ฏธ์ง€์›ํ•ฉ๋‹ˆ๋‹ค ใ…Žใ…Ž ์ฟ ๋‹ค๊ฐ€ ํ•„์š”ํ•ด์šฉ~~ ์ฆ‰ ๋งฅ์—์„œ๋Š” ํŒŒ์ธํŠœ๋‹ ๋ถˆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. torch ๋Š” ๊ฐ€๋Šฅํ•ด์š”~
@dkssudskdigpy476
@dkssudskdigpy476 5 ะฐะน ะฑาฑั€ั‹ะฝ
ํ…Œ๋””๋…ธํŠธ๋‹˜, ๋งค๋ฒˆ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํ˜น์‹œ, ๋ฌธ์„œ ๊ธฐ๋ฐ˜ QA ๋ฐ์ดํ„ฐ์…‹์ด ์ž˜ ์ƒ์„ฑ๋˜์—ˆ๋Š”์ง€๋Š” ์–ด๋–ป๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋‚˜์š”?
@teddynote
@teddynote 5 ะฐะน ะฑาฑั€ั‹ะฝ
์ €์˜ ๋‹ค๋ฅธ ์˜์ƒ ์ค‘์—์„œ AutoRAG ์ชฝ ์‹œ๋ฆฌ์ฆˆ ์˜์ƒ์„ ๋ณด์‹œ๋ฉด ๋„์›€์ด ๋˜์‹ค ๊ฒƒ ๊ฐ™์•„์š”! AutoRAG์—์„œ ํ‰๊ฐ€ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค~
@dkssudskdigpy476
@dkssudskdigpy476 5 ะฐะน ะฑาฑั€ั‹ะฝ
@@teddynote ๋„ค, ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค ๐Ÿ˜„
@dkssudskdigpy476
@dkssudskdigpy476 5 ะฐะน ะฑาฑั€ั‹ะฝ
@@teddynote ํ•ด๋‹น ๋…ธํŠธ๋ถ ์ฝ”๋žฉ ์ผ๋ฐ˜์œผ๋กœ ๊ตฌ๋™ ์‹œ ์›Œ๋‹์ด ๋ฐœ์ƒํ•˜๊ณ , ์„ธ์…˜ ๋‹ค์‹œ ์‹œ์ž‘ ์‹œ ์—๋Ÿฌ๊ฐ€ ๋ฐœ์ƒํ•˜๋Š” ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ํ˜น์‹œ ํ™•์ธํ•ด์ฃผ์‹ค ์ˆ˜ ์žˆ์œผ์‹ค๊นŒ์š”? ''' WARNING: The following packages were previously imported in this runtime: [torch,torchgen] You must restart the runtime in order to use newly installed versions. Restarting will lose all runtime state, including local variables. '''
@teddynote
@teddynote 5 ะฐะน ะฑาฑั€ั‹ะฝ
@@dkssudskdigpy476 ์˜์กด์„ฑ ํŒจํ‚ค์ง€ ๋“ค์ด ์žˆ์–ด์„œ์š”. ์ฝ”๋žฉ๋ณด๋‹ค poetry ๋กœ ์˜์กด์„ฑ์ด ํ•ด๊ฒฐ๋œ ํ™˜๊ฒฝ์—์„œ ์ง„ํ–‰ํ•ด ๋ณด์‹ค ๊ฒƒ์„ ์ถ”์ฒœ ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์˜์ƒ์ธ ๋žญ์ฒด์ธ ํŠœํ† ๋ฆฌ์–ผ ์„ค์น˜ ์˜์ƒ์—์„œ poetry ์„ค์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋‹ค๋ฃจ๊ณ  ์žˆ์œผ๋‹ˆ ์ฐธ๊ณ ํ•ด ์ฃผ์„ธ์š”^^
@dkssudskdigpy476
@dkssudskdigpy476 5 ะฐะน ะฑาฑั€ั‹ะฝ
pip install --force-reinstall "xformers
@๋ฐ•๋™์ต-f3o
@๋ฐ•๋™์ต-f3o 7 ะฐะน ะฑาฑั€ั‹ะฝ
ํ…Œ๋””๋…ธํŠธ๋‹˜ ํ•ญ์ƒ ์ข‹์€ ๊ฐ•์˜ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํ…Œ๋””๋…ธํŠธ๋‹˜ ๋”ฐ๋ผ์„œ gguf ๋ณ€ํ™˜๊นŒ์ง€ ์ž˜ ๋”ฐ๋ผํ•ด๋ณด์•˜์Šต๋‹ˆ๋‹ค. ๋ณ€ํ™˜๋œ ๋ชจ๋ธ์„ gguf๋กœ ollama ์— ์˜ฌ๋ฆฌ๋ ค๋‹ˆ๊นŒ template์„ ์–ด๋–ป๊ฒŒ ํ•ด์•ผ ํ• ์ง€ ๋ชจ๋ฅด๊ฒ ๋”๋ผ๊ตฌ์š”. (llama3 ๊ณต์‹ template์œผ๋กœ ํ•˜๋‹ˆ ์ž˜ ์•ˆ๋˜๋”๋ผ๊ตฌ์š”!) ํ˜น์‹œ gguf ๋ณ€ํ™˜๋œ ํŒŒ์ผ์„ ollama ์— ์˜ฌ๋ฆด ๋•Œ template์„ ์–ด๋–ป๊ฒŒ ํ•˜๋ฉด ์ข‹์„์ง€ ์•„์‹œ๋Š” ๋ถ„ ๊ณ„์‹ค๊นŒ์š”?
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
ํ…œํ”Œ๋ฆฟ์€ alpaca ํ…œํ”Œ๋ฆฟ์œผ๋กœ ์—…๋กœ๋“œ ํ•˜์‹œ๋ฉด ๋ฉ๋‹ˆ๋‹ค^^
@lilililililili-c2l
@lilililililili-c2l 7 ะฐะน ะฑาฑั€ั‹ะฝ
RAG๋ž‘ ํŒŒ์ธ ํŠœ๋‹ ์ค‘์— ๊ฐ๊ฐ ์–ด๋–ค ์ƒํ™ฉ์—์„œ ์“ฐ๋Š”๊ฒŒ ๋” ์ข‹์€ ๋ฐฉํ–ฅ์ผ๊นŒ์šค?
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
์ž์ฃผ ํ•˜์‹œ๋Š” ์งˆ๋ฌธ์ธ๋ฐ์š”! ์‰ฝ๊ฒŒ ๋งํ•ด ์ž์ฃผ ๋ณ€๊ฒฝ๋˜๋Š” ์ •๋ณด๋Š” RAG ๊ณ ์ •๋œ ๋‹ต๋ณ€ ํ˜•์‹์€ ํŒŒ์ธํŠœ๋‹์ด ์ ์ ˆํ•ด ๋ณด์ž…๋‹ˆ๋‹ค!! ๊ทธ๋ฆฌ๊ณ  ๊ต‰์žฅํžˆ ๊ธธ๊ฒŒ ๋งค๋ฒˆ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋„ฃ์–ด์ค˜์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ์—๋„ ํŒŒ์ธํŠœ๋‹์„ ํ•˜๋ฉด ์ž…๋ ฅ์œผ๋กœ ๋„ฃ๋Š” ์ปจํ…์ŠคํŠธ๋ฅผ ์ค„์—ฌ๋ณผ ์ˆ˜ ์žˆ๊ณ ์š”. RAG๋กœ๋Š” ํ• ๋ฃจ์‹œ๋„ค์ด์…˜์„ ์ข€ ์ค„์ด๋Š”๋ฐ ์‚ฌ์šฉํ•˜์‹ค ์ˆ˜ ์žˆ์–ด์š”! ๋น„๊ตํ•˜๋Š” ์•„ํ‹ฐํด์ด๋‚˜ ๋ธ”๋กœ๊ทธ๋„ ๊ฝค ๋งŽ์œผ๋‹ˆ ํ•œ ๋ฒˆ ์‚ดํŽด๋ณด์‹œ๋Š” ๊ฒƒ๋„ ์ข‹์Šต๋‹ˆ๋‹ค^^
@lilililililili-c2l
@lilililililili-c2l 7 ะฐะน ะฑาฑั€ั‹ะฝ
@@teddynote ๋„ˆ๋ฌด ์นœ์ ˆํ•  ๋‹ต๋ณ€ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค ใ…  ๊ฐ•์˜ ํšŒ์‚ฌ์— ์‹ ์ฒญํ•ด์„œ ํšŒ์‚ฌ๊ณ„์ •์œผ๋กœ ๊ผญ ๋ณผ๊ฒŒ์š”๐Ÿ’“
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
@@lilililililili-c2l ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค~!!!
@๋‚จ์Šค-n5b
@๋‚จ์Šค-n5b 7 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š” ์ง€๋‚˜๊ฐ€๋Š” ๊ทธ๋ƒฅ ์ง์žฅ์ธ์ž…๋‹ˆ๋‹ค. ์ œ๊ฐ€ ๋น„์Šทํ•œ ์ƒ๊ฐ์œผ๋กœ Gpts๋ฅผ ๋งŒ๋“ค๊ณ  ํšŒ์‚ฌ ๋ณด๊ณ ๊ทœ์ •, ๋ณด๊ณ ์„œ์ƒ˜ํ”Œ, ๋งค๋‰ด์–ผ ๋“ฑ์„ pdf๋กœ ์—…๋กœ๋“œ์‹œํ‚ค๊ณ  ์—…๋กœ๋“œ๋œ ํŒŒ์ผ์— ๊ทผ๊ฑฐํ•ด์„œ ๋‹ต๋ณ€ํ•˜๋ผ๊ณ  ํ–ˆ๋Š”๋ฐ ํ• ๋ฃจ์ง€๋„ค์ด์…˜๋„ ๋งŽ๊ณ  ๋ณด๊ณ ์–‘์‹์— ๋งž์ง€์•Š๊ฒŒ ๋Œ€๋‹ตํ• ๋•Œ๋„๋งŽ๊ณ  ์ถœ์ฒ˜๋ฅผ ๋ฌผ์–ด๋ณด๋‹ˆ ์—‰๋šฑํ•œํŽ˜์ด์ง€๋ฅผ ๋Œ€๋‹ตํ•˜๋Š”๊ฒฝ์šฐ๋„ ์žˆ๋”๋ผ๊ตฌ์š” ๊ทธ๋ž˜์„œ ํฌ๊ธฐํ•˜๊ณ ์žˆ์—ˆ๋Š”๋ฐ ์ด ๋ฐฉ์‹์€ ์ €๊ฐ™์ด ๋น„๊ฐœ๋ฐœ์ž๋„ ์ ‘๊ทผ์ด๊ฐ€๋Šฅํ• ๊นŒ์š”?
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
GPTs ๋Š” ๋ฌธ์„œ ๊ธฐ๋ฐ˜ QA์— ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค ใ…œ ๊ทธ๋ž˜์„œ ๋žญ์ฒด์ธ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•ด์„œ ์ง์ ‘ ๊ตฌํ˜„ํ•ด์•ผ ์ •ํ™•ํ•œ ์ถœ์ฒ˜์™€ ๋‹ต๋ณ€์„ ๋ฐ›์œผ์‹ค ์ˆ˜ ์žˆ๊ตฌ์š”~^^ ์ด ์˜์ƒ๋งŒ์œผ๋กœ๋Š” ์™„์„ฑํ˜„ ๊ฐœ๋ฐœ์ด ๋‚˜์˜ค๊ธฐ๋Š” ์–ด๋ ต๊ณ  ์ „๋ฐ˜์ ์ธ RAG ์— ๋Œ€ํ•œ ๊ณต๋ถ€๊ฐ€ ์กฐ๊ธˆ ํ•„์š”ํ•œ๋ฐ์š”. ์œ ํŠœ๋ธŒ์— ์—…๋กœ๋“œ๋œ ๋‹ค๋ฅธ ์˜์ƒ ๊ฐ™์ด ๋ณด์‹œ๋ฉด์„œ ๊ณต๋ถ€ํ•ด ๋ณด์‹œ๋Š” ๊ฒƒ์„ ์ถ”์ฒœ ๋“œ๋ฆฌ๊ณ ์š”. ๋‹จ๊ธฐ๊ฐ„์€ ์•„๋‹ˆ์ง€๋งŒ ์กฐ๊ธˆ๋งŒ ํˆฌ์žํ•˜์…”๋„ ์„ฑ๊ณผ๊ฐ€ ๋‚˜์‹ค๊ป๋‹ˆ๋‹ค!
@๋‚จ์Šค-n5b
@๋‚จ์Šค-n5b 7 ะฐะน ะฑาฑั€ั‹ะฝ
@@teddynote ๋‹ต๋ณ€ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค
@์ด์žฌ์˜-w4v7s
@์ด์žฌ์˜-w4v7s 7 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š”! ํ•ญ์ƒ ์–‘์งˆ์˜ ์ง€์‹ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ œ๊ฐ€ flash-attn ์„ค์น˜์— ์• ๋ฅผ ๋จน๊ณ  ์žˆ๋Š”๋ฐ ํ˜น์‹œ window ๋กœ์ปฌ์—์„œ ํ•ด๋‹น ๊ฐœ๋ฐœ ํ™˜๊ฒฝ์„ ๊ตฌ์ถ•ํ•ด๋ณด์‹  ๊ฒฝํ—˜์ด ์žˆ์œผ์‹ ๊ฐ€์š”?
@์ด์žฌ์˜-w4v7s
@์ด์žฌ์˜-w4v7s 7 ะฐะน ะฑาฑั€ั‹ะฝ
๊ฒฐ๊ตญ ์–ด์ฐŒ์ €์ฐŒ ๋ฒ„์ „ ๋งž์ถฐ ์„ค์น˜ํ–ˆ๋Š”๋ฐ, flash-attn ๋‹จ๋… ์„ค์น˜์— ๊ฑธ๋ฆฐ ์‹œ๊ฐ„๋งŒ 1์‹œ๊ฐ„ ๊ฑธ๋ ธ๋˜ ๊ฒƒ ๊ฐ™์•„์š”. ์„ค์น˜ํ•  ๋•Œ ์ฃผ์˜ํ•ด์•ผ ํ•˜๋Š” ๋ถ€๋ถ„์ด ์žˆ๋‚˜์š”? cuda๊นŒ์ง€ ๋Œ์•„๊ฐ€๋Š” ๊ฑฐ ํ™•์ธํ•˜๊ณ  ์„ค์น˜ํ–ˆ๋Š”๋ฐ ์™œ ์ด๋ฆฌ ์˜ค๋ž˜๊ฑธ๋ฆฐ๊ฑธ๊นŒ์š” ๐Ÿ˜‚๐Ÿ˜‚ unsloth์˜ triton์€ window์—์„œ ์„ค์น˜๊ฐ€ ์•ˆ๋˜๋‚˜์š”? MVSC๋‚˜ WSL์„ ์‚ฌ์šฉํ•˜๋ผ๋Š” ๊ธ€์ด ์žˆ๋˜๋ฐ ์กฐ์–ธ์ด ํ•„์š”ํ•ด์š”..
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
์ œ๊ฐ€ flash-attn ์€ ์‚ฌ์šฉํ•ด๋ณด์ง€ ์•Š์•„์„œ์š” ๋„์›€๋“œ๋ฆฌ์ง€ ๋ชปํ•  ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค ใ…  ์ฃ„์†กํ•ฉ๋‹ˆ๋‹ค ใ…œ
@์ด์žฌ์˜-w4v7s
@์ด์žฌ์˜-w4v7s 7 ะฐะน ะฑาฑั€ั‹ะฝ
@@teddynote ํ•ญ์ƒ ๋„์›€ ๋งŽ์ด ๋ฐ›๊ณ ์žˆ์Šต๋‹ˆ๋‹ค ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค :)
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
@@์ด์žฌ์˜-w4v7s ์ด๋ฒˆ์—๋Š” ๋„์›€์„ ๋ชป ๋“œ๋ ธ๋„ค์š” ใ…œ
@jjabul
@jjabul 3 ะฐะน ะฑาฑั€ั‹ะฝ
์ €๋„ ๊ณ ์ƒ๊ณ ์ƒ ํ•˜๋‹ค๊ฐ€... ๋กœ์ปฌ/์œˆ๋„์šฐ์—์„œ ํ–ˆ๋Š”๋ฐ์š” flash-attn ์—๋งŒ 2~3์‹œ๊ฐ„ ์ด์ƒ ๊ฑธ๋ ธ๋˜ ๊ฒƒ ๊ฐ™๊ณ  xformers ์ชฝ์—์„œ๋Š” ninja ์˜ 260์ž ํŒŒ์ผ๋ช…์—๋Ÿฌ๊ฐ€ ๊ณ„์† ๋‚˜์„œ ์‹คํŒจํ–ˆ์–ด์š”. WSL ๋กœ ์ „ํ™˜ํ•˜๊ณ  ํ•ด๊ฒฐํ–ˆ์œผ๋‹ˆ ๋Šฆ์—ˆ์ง€๋งŒ ์ฐธ๊ณ ํ•˜์„ธ์š”
@wonyanglee9568
@wonyanglee9568 3 ะฐะน ะฑาฑั€ั‹ะฝ
์˜์ƒ์—์„œ ๋งŽ์€ ์ •๋ณด๋ฅผ ์–ป์–ด ๊ฐ€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. 01-QA-Pair-GPT.ipynb ๋ฅผ ์ฝ”๋žฉ์—์„œ ์‹คํ–‰ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. print(elements[2]) ๋ฅผ ํ•˜๋ฉด ํ•œ๊ธ€์ด ๊นจ์ ธ ๋‚˜์˜ค๋Š”๋ฐ์š”. ํ˜น์‹œ ํ•ด๊ฒฐ ๋ฐฉ๋ฒ•์ด ์žˆ์„๊นŒ์š” ?
@teddynote
@teddynote 3 ะฐะน ะฑาฑั€ั‹ะฝ
ํ˜น์‹œ hwp ํŒŒ์ผ์„ pdf ๋ณ€ํ™˜ํ•œ ํŒŒ์ผ์ธ๊ฐ€์š”~?
@์กฐ์ด๋ฆฌ-u5b
@์กฐ์ด๋ฆฌ-u5b 7 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š” ํ•ด๋‹น ๋ชจ๋ธ์„ ollama ์— ์˜ฌ๋ฆฌ๊ธฐ ์œ„ํ•ด์„œ ์‚ฌ์šฉํ•ด์•ผํ•  Modelfile ์€ ์–ด๋–ค๊ฑธ ์ฐธ๊ณ ํ•˜๋ฉด ๋ ๊นŒ์š”?
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
alpaca ์ž…๋‹ˆ๋‹ค^^
@vhxmtmfp
@vhxmtmfp 7 ะฐะน ะฑาฑั€ั‹ะฝ
๋ฐ์ดํ„ฐ์…‹ ์ƒ์„ฑ์‹œ instruction, input, output๊ตฌ์กฐ๋กœ ์ƒ์„ฑํ•˜์‹œ๋Š”๋ฐ input์ด ๋น„์–ด์žˆ๋„ค์š” input์˜ ์—ญํ• ์€ ๋ญ˜๊นŒ์š”?
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
instruction์€ ์ง€์‹œ์‚ฌํ•ญ ๊ทธ๋ฆฌ๊ณ  input์€ ๋ฌธ๋งฅ์ด๋ผ๊ณ  ์ดํ•ดํ•˜์‹œ๋ฉด ์ข‹์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ๋“ค์–ด, Q: ๋Œ€ํ•œ๋ฏผ๊ตญ์˜ ๋„์‹œ๋ฅผ ๋‚˜์—ดํ•ด๋ด input: ์„œ์šธ, ๋ถ€์‚ฐ, output: ๋Œ€๊ตฌ, ๊ด‘์ฃผ ๋“ฑ๋“ฑ ์ด๋Ÿฐ์‹์œผ๋กœ ๊ตฌ์„ฑ์ด ๊ฐ€๋Šฅํ•ด์š”~
@ittubes3912
@ittubes3912 6 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š” ํ…Œ๋””๋‹˜ ํ…Œ๋””๋‹˜ ์œ ํŠœ๋ธŒ ์˜์ƒ ์ž˜ ๋ณด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฆ„์ด ์•„๋‹ˆ๋ผ ํ…Œ๋””๋‹˜ github์—์„œ ํŒŒ์ธํŠœ๋‹ ํ•˜๋Š” ์ฝ”๋“œ๋กœ ๊ฐœ์ธ ์ปค์Šคํ…€ ๋ฐ์ดํ„ฐ์…‹์„ ํŒŒ์ธํŠœ๋‹ํ–ˆ๋Š”๋ฐ ๊ฒฐ ํ•™์Šต์‹œํ‚จ ๋ฐ์ดํ„ฐ๋กœ ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜์˜ค์ง„ ์•Š๋Š”๋ฐ...ํ˜น์‹œ ํŒŒ์ธํŠœ๋‹์„ ์—ฌ๋Ÿฌ๋ฒˆ ํ•ด์•ผํ•˜๋Š” ๊ฑธ๊นŒ์š”?
@teddynote
@teddynote 6 ะฐะน ะฑาฑั€ั‹ะฝ
๊ฒฐ๊ณผ๊ฐ€ ์•ˆ๋‚˜์˜จ๋‹ค๋Š” ๋ถ€๋ถ„์— ๋Œ€ํ•ด์„œ ๊ตฌ์ฒด์ ์œผ๋กœ ๋ง์”€ํ•ด ์ฃผ์‹œ๋ฉด ๋‹ต๋ณ€ ๋“œ๋ฆด ์ˆ˜ ์žˆ์„ ๊ฒƒ ๊ฐ™์•„์š”~
@ittubes3912
@ittubes3912 6 ะฐะน ะฑาฑั€ั‹ะฝ
a๋Š” ๋ฐ”๋ณด๋‹ค ๋ผ๊ณ  ํ•™์Šต์‹œ์ผฐ๋Š”๋ฐ a๋Š” ์•ŒํŒŒ๋ฒณ์œผ๋กœ ์–ด์ฉŒ๊ตฌ์ €์ฉŒ๊ตฌ ์ด๋Ÿฐ์‹์œผ๋กœ ๋‚˜์˜ต๋‹ˆ๋‹ค. ์ฝ”๋“œ๋Š” github์ฐธ๊ณ ํ–ˆ๊ณ  gptํ•œํ…Œ ๊ณต์œ ์ฃผ์‹  ์ฝ”๋“œ ๊ธฐ๋ฐ˜์œผ๋กœ ํ—ˆ๊น…ํŽ˜์ด์Šค์—์„œ ๋ฐ์ดํ„ฐ์…‹ ๋‹ค์šด๋ฐ›์•„์„œ ๋กœ์ปฌ์— ์ €์žฅํ•œ๋‹ค์Œ ๋‹ค์šด๋ฐ›์€ ๋ฐ์ดํ„ฐ์…‹์„ ๋ถˆ๋Ÿฌ์„œ ํŒŒ์ธํŠœ๋‹ํ•˜๊ฒŒ ํ–ˆ์Šต๋‹ˆ๋‹ค..
@joey1025ful
@joey1025ful 7 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š” ์˜์ƒ๋ณด๋ฉฐ ๋”ฐ๋ผํ•˜๊ณ  ์žˆ๋Š”๋ฐ ImportError Traceback (most recent call last) Cell In[11], line 1 ----> 1 from unsloth import FastLanguageModel ๋ถ€๋ถ„์—์„œ ๊ณ„์† ์—๋Ÿฌ๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค unsloth ์„ค์น˜๊ฐ€ ์•ˆ๋˜๋Š”๊ฑฐ ๊ฐ™์€๋ฐ cuda: 12.1 python: 3.9.0 torch: 2.2.0+cu121 ์‚ฌ์šฉ ํ™˜๊ฒฝ์—์„œ unsloth๋ฅผ ์–ด๋–ป๊ฒŒ ์„ค์น˜ํ•ด์•ผ ํ•˜๋‚˜์š”?
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š”? ํ˜น์‹œ ๋กœ์ปฌ ํŒŒ์ด์ฌ ๊ฐ€์ƒํ™˜๊ฒฝ์—์„œ ์‹คํ–‰ํ•˜์‹œ๋‚˜์š”? unsloth ์„ค์น˜ ์ฝ”๋“œ๊ฐ€ ์ž๋ฃŒ ์•ˆ์— ๋“ค์–ด๊ฐ€ ์žˆ๋Š”๋ฐ ์„ค์น˜๊ฐ€ ์ œ๋Œ€๋กœ ๋˜์—ˆ๋Š”์ง€ ํ™•์ธํ•ด ๋ณด์‹œ๊ฒ ์–ด์š”?
@joey1025ful
@joey1025ful 7 ะฐะน ะฑาฑั€ั‹ะฝ
๋„ค ๊ฐ€์ƒํ™˜๊ฒฝ์—์„œ ํ•˜๊ณ ์žˆ์Šต๋‹ˆ๋‹ค
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
@@joey1025ful ์„ค์น˜ํ•˜์‹ค ๋•Œ ์˜์กด์„ฑ์ด ๋งŽ์•„์„œ ์—ฌ๊ธฐ ๊ฐ€์ด๋“œ ๋ณด๊ณ  ๋ฒ„์ „ ๋งž์ถฐ์„œ ์„ค์น˜ ์ง„ํ–‰ํ•ด ์ฃผ์‹œ๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค~ ํ•œ ๋ฒˆ ์ฐธ๊ณ ํ•ด ๋ณด์„ธ์š”!(cuda, torch, peft ๋ฒ„์ „) github.com/unslothai/unsloth
@Younggoolee
@Younggoolee 5 ะฐะน ะฑาฑั€ั‹ะฝ
๊ฐ‘์ž๊ธฐ ์ฝ”๋žฉ์—์„œ ์˜ค๋ฅ˜๊ฐ€ ์ž…๋‹ˆ๋‹ค. ใ… ใ…  ๋ชจ๋ธ ๋กœ๋“œํ•˜๋Š” ์†Œ์Šค์—์„œ ๋‚˜๊ธฐ ์‹œ์ž‘ํ•˜๋Š”๋ฐ. ์ž˜ ๋˜์—ˆ๋Š”๋ฐ. ์™œ ๊ทธ๋Ÿด๊นŒ์š” ใ… ใ…  ImportError: Unsloth: If you are in Colab, we updated the top cell install instructions - please change it to below then press Disconnect Runtime and then Restart it.
@teddynote
@teddynote 5 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š”! ์ตœ๊ทผ์— unsloth ์„ค์น˜ ์—…๋ฐ์ดํŠธ๊ฐ€ ๋˜๋ฉด์„œ ์˜์กด์„ฑ ๋ฌธ์ œ๊ฐ€ ์ƒ๊ฒจ ์ด์Šˆ๊ฐ€ ์žˆ๋”๋ผ๊ณ ์š” ใ…œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฐฉ๋ฒ•์œผ๋กœ ํ•œ ๋ฒˆ ์‹œ๋„ํ•ด๋ด ์ฃผ์„ธ์š” 1. unsloth ๊ฐ€์ด๋“œ ๋”ฐ๋ผ ์„ค์น˜(์ตœ์‹  ๋…ธํŠธ๋ถ ๊ฐ€์ด๋“œ์— ์žˆ๋Š” ์„ค์น˜ ์ฝ”๋“œ ์ฐธ์กฐ) colab.research.google.com/drive/135ced7oHytdxu3N2DNe1Z0kqjyYIkDXp?usp=sharing#scrollTo=2eSvM9zX_2d3 2.์ƒ๋‹จ ๋Ÿฐํƒ€์ž„ - ์„ธ์…˜ ๋‹ค์‹œ ์‹œ์ž‘ 3. ์„ค์น˜ ํ•„์š” ์—†์ด ๋ฐ”๋กœ ์ฝ”๋“œ๋ถ€ํ„ฐ ์‹คํ–‰
@Younggoolee
@Younggoolee 5 ะฐะน ะฑาฑั€ั‹ะฝ
@@teddynote ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์‹คํ–‰์ด ๋˜๋„ค์š”. ์„ฑ๋Šฅ์€ ํ•œ๋ฒˆ ๋น„๊ตํ•ด๋ด์•ผ ๊ฒ ์Šต๋‹ˆ๋‹ค.
@๋ธ”๋ธ”๋ฃจ-m4j
@๋ธ”๋ธ”๋ฃจ-m4j 7 ะฐะน ะฑาฑั€ั‹ะฝ
๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๋ฌธ์„œ์—์„œ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•ด ํŒŒ์ธํŠœ๋‹์„ ์ง„ํ–‰ํ•ด ๋ณด๋ ค๊ณ  ํ•˜๋Š”๋ฐ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๋ฌธ์„œ์˜ ์–ธ์–ด ๋น„์œจ์ด ์˜์–ด 70% ํ•œ๊ตญ์–ด 30%์ผ ๋•Œ๋ผ๋ฉด ํ•œ๊ตญ์–ด๋กœ ํŒŒ์ธํŠœ๋‹๋œ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒŒ ๋งž์„๊นŒ์š”?
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
ํ•œ๊ตญ์–ด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ๋‹ค๋ฉด ํ•œ๊ตญ์–ด ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์‹œ๋Š” ๊ฒƒ์ด ๋” ๋‚˜์•„ ๋ณด์ž…๋‹ˆ๋‹ค^^
@necrojim6954
@necrojim6954 4 ะฐะน ะฑาฑั€ั‹ะฝ
์ฝ”๋“œ๋ฅผ ๋ˆŒ๋Ÿฌ๋„ ์—ฐ๊ฒฐ์ด ์•ˆ๋˜๋Š”๋ฐ.. ๊นƒํ—ˆ๋ธŒ์—์žˆ๋Š” 18. ํŒŒ์ธํŠœ๋‹์œผ๋กœ ๋ด๋„ ๋ฌด๋ฐฉ ํ• ๊นŒ์—ฌ?
@teddynote
@teddynote 4 ะฐะน ะฑาฑั€ั‹ะฝ
๋งž์Šต๋‹ˆ๋‹ค ์ตœ๊ทผ์— ๋ชฉ์ฐจ ๋ฒˆํ˜ธ๋ฅผ ๋ฐ”๊พธ์–ด์„œ ๋งํฌ๊ฐ€ ๊นจ์กŒ๋‚˜๋ด…๋‹ˆ๋‹ค~ ์ˆ˜์ •ํ•ด ๋‘๊ฒ ์Šต๋‹ˆ๋‹ค ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค^^
@HJ-jf4co
@HJ-jf4co 7 ะฐะน ะฑาฑั€ั‹ะฝ
Colab์—์„œ ํŒŒ์ธํŠœ๋‹ ์ค‘์— ๋ชจ๋ธ ํ›ˆ๋ จํ•˜๊ธฐ > SFTTrainer ์‚ฌ์šฉ ๊ตฌ๊ฐ„์—์„œ ์•„๋ž˜์—๋Ÿฌ๊ฐ€ ๋ฐœ์ƒํ•˜๋Š”๋ฐ ์›์ธ๊ณผ ํ•ด๊ฒฐ๋ฐฉ๋ฒ• ์•Œ ์ˆ˜ ์žˆ์„๊นŒ์š”..ใ… ใ…  -> NotImplementedError: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
ํ˜น์‹œ ์ฝ”๋žฉ์— GPU ๋ฅผ A100์œผ๋กœ ์„ค์ •์ด ๋˜์–ด ์žˆ๋Š”์ง€ ํ•œ ๋ฒˆ ์ฒดํฌํ•ด ์ฃผ์„ธ์š”~ ์ž‘์€ ๊ฑฐ๋Š” ์—๋Ÿฌ๋‚˜๋”๋ผ๊ตฌ์š”~
@์•„์•„123-v9g
@์•„์•„123-v9g 7 ะฐะน ะฑาฑั€ั‹ะฝ
@@teddynote ์ €๋„ ์œ„์™€ ๊ฐ™์€ ์—๋Ÿฌ๊ฐ€ ๋ฐ˜๋ณต๋˜์–ด์„œ, Colab-Pro ๋กœ ๋ณ€๊ฒฝ ํ›„ ์‚ฌ์šฉํ–ˆ๋Š”๋ฐ, ์ด์ƒํ•˜๊ฒŒ ์ž๊พธ L4 ๋กœ ๋ณ€๊ฒฝ๋˜๊ฑฐ๋‚˜ A100 ์„ ํƒ์ด ์ž˜๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ใ…œใ…œ ๋” ๋†’์€ ๊ฒฐ์žฌ plan ์„ ์‚ฌ์šฉํ•ด์•ผ ํ• ๊นŒ์š” ใ… ใ… ? ์•„๋‹ˆ๋ฉด, Colab ๋ง๊ณ  ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•์œผ๋กœ ์œ„ ํŒŒ์ธ ํŠœ๋‹์„ ํ•  ์ˆ˜ ์—†๋Š”์ง€ ๊ถ๊ธˆํ•ฉ๋‹ˆ๋‹ค ^^ Colab ์œ ๋ฃŒ ๊ฒฐ์žฌ ์„œ๋น„์Šค ์งˆ์ด ๋„ˆ๋ฌด ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค.
@heejuneAhn
@heejuneAhn 7 ะฐะน ะฑาฑั€ั‹ะฝ
๋‹ค๋ฅธ ๊ฑด ๋‹ค ์ž˜ ๋˜๋Š”๋ฐ (merge ํ•œ ํ›„์— ๋‹ค์‹œ ๋กœ๋”ฉ, gguf ํŒŒ์ผ ์ €์žฅ ํ›„ ollma์—์„œ ์‚ฌ์šฉ ๋“ฑ์€ ์ž˜๋ฉ๋‹ˆ๋‹ค.) , Lora adaptor ๋งŒ ์ €์žฅํ•œ ํ›„์— ๋‹ค์‹œ ๋กœ๋”ฉํ•˜์—ฌ ์‹คํ–‰ํ•  ๋•Œ ์—ฌ๋Ÿฌ๊ฐ€์ง€ ๋ฌธ์ œ๊ฐ€ ์ƒ๊ธฐ๋Š”๋ฐ ํ˜น์‹œ ๋™์ž‘ํ•˜๋Š” ์ฝ”๋“œ๋ฅผ ๊ณต์œ ํ•ด ์ค„ ์ˆ˜ ์žˆ์„๊นŒ์š”? ์‚ฌ์šฉํ•œ ์ฝ”๋“œ๋Š” ์šฐ์„  ์ €์žฅ์€ ๋กœ์ปฌ์— ๋‹ค์Œ๊ณผ ๊ฐ™์ด ํ•˜์˜€๊ตฌ์š”. adaptor_model_path = "llama-3-8b-lora" peft_model.save_pretrained(adaptor_model_path) ์ดํ›„์— ๋‹ค์‹œ ๋‘๊ฐ€์ง€ ๋ฐฉ์‹์œผ๋กœ ๋กœ๋”ฉ์„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‘ ๊ฐ€์ง€๋‹ค ๋ฌธ์ œ๋ฅผ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. from peft import PeftConfig, PeftModel from transformers import AutoModelForCausalLM, AutoTokenizer if True: base_model = AutoModelForCausalLM.from_pretrained(base_model_name) # not working when generated peft_model = PeftModel.from_pretrained(base_model, adaptor_model_path) else: from peft import AutoPeftModelForCausalLM peft_model = AutoPeftModelForCausalLM.from_pretrained(adaptor_model_path, load_in_4bit=True) # 4bit needed for GPU VRAM ๊ทธ๋ฆฌ๊ณ  ๋‚˜์„œ ์ œ๊ณตํ•œ ์ฝ”๋“œ๋ฅผ ํ•จ์ˆ˜ํ™” ํ•˜์—ฌ ํ…Œ์ŠคํŠธ ํ•ฉ๋‹ˆ๋‹ค. test(alpaca_prompt, peft_model, tokenizer)๊ทธ ์ด ๋‘ ๊ฐ€์ง€ ๊ฒฝ์šฐ ๋ชจ๋‘ ๋‹ค์Œ ์—๋Ÿฌ๊ฐ€ ๋ฐœ์ƒํ•˜๊ณ  auto-gptq ๋Š” 0.4.2 ๋กœ ์—…๋ฐ์ดํŠธ ํ–ˆ๊ตฌ์š” ๊ฒฐ๊ณผ๋Š” ๋งˆ์ฐฌ๊ฐ€์ง€ ์ž…๋‹ˆ๋‹ค. File "/home/gpt/anaconda3/envs/text2sql/lib/python3.11/site-packages/unsloth/models/llama.py", line 309, in LlamaAttention_fast_forward Q, K, V = self.apply_qkv(self, hidden_states) ^^^^^^^^^^^^^^ File "/home/gpt/anaconda3/envs/text2sql/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1695, in __getattr__ raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'") AttributeError: 'LlamaSdpaAttention' object has no attribute 'apply_qkv' ''' ์ข€ ํŠน์ดํ•œ ์ ์€ ์ฒซ๋ฒˆ์งธ ๋ฐฉ๋ฒ•์˜ ๊ฒฝ์šฐ base_model์„ ์ด์ „์— ๋กœ๋”ฉํ–ˆ๋˜ ๊ฒƒ์„ ๊ทธ๋Œ€๋กœ ์‚ฌ์šฉํ•˜๋ฉด ๋™์ž‘์„ ํ•˜๋Š”๋ฐ. ์ด๊ฒŒ๊ฑด ์‚ฌ์‹ค์ƒ ์‚ฌ์šฉํ•  ์ˆ˜์—†๋Š” ์‹œ๋‚˜๋ฆฌ์˜ค๊ธฐ ๋•Œ๋ฌธ์— ์˜๋ฏธ๊ฐ€ ์—†๊ตฌ์š”. ๋™์ผํ•˜๊ฒŒ ๋‹ค์‹œ ๋กœ๋”ฉ์„ ํ•ด๋ด๋„ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•˜๋Š” ๊ฒƒ์„ ๋ณด๋ฉด peft_model = FastLanguageModel.get_peft_model ๋ฅผ ํ•˜๋ฉด์„œ ๋ชจ๋ธ์— base model ์ž์ฒด์—/์—๋„ ๋ณ€ํ™”๋ฅผ ์ฃผ๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ด๋Š”๋ฐ. ์ด๊ฑด ์†Œ์Šค์ฝ”๋“œ๋ฅผ ๋ด์•ผ ์•Œ ๊ฒƒ ๊ฐ™๋„ค์š”.
@dongyunkim1148
@dongyunkim1148 4 ะฐะน ะฑาฑั€ั‹ะฝ
gpt์—์„œ ์ƒ์„ฑ ํ›„ qa_pair ๋ฆฌ์ŠคํŠธ์— ๋‹ด๋Š” ๊ณผ์ •์—์„œ ์ƒ์„ฑ๋˜๋Š” ์งˆ๋ฌธ 3๊ฐœ๋ฅผ ์ „๋ถ€ ๋‹ด๋Š”๊ฒŒ ์•„๋‹ˆ๋ผ ์ตœ์ดˆ1๊ฐœ๋งŒ ๋‹ด๋Š” ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•˜๋„ค์š”..ใ… ใ…  gpt๊ฐ€ ์ƒ์„ฑํ•  ๋•Œ์—๋Š” ๋ฌธ์ œ 3๊ฐœ๋ฅผ ์ƒ์„ฑํ•˜๋Š”๊ฑธ ํ”„๋ฆฐํŠธ๋กœ ๋ณด์—ฌ์ฃผ๋Š”๋ฐ ์‹ค์ œ ์ €์žฅ์ด ๋˜์ง€ ์•Š๋Š” ์ด์œ ๊ฐ€ ์žˆ์„๊นŒ์š”?? ๊ทธ๋Œ€๋กœ ๋”ฐ๋ผ ํ–ˆ๋Š”๋ฐ ใ… ใ… 
@teddynote
@teddynote 4 ะฐะน ะฑาฑั€ั‹ะฝ
1๊ฐœ๋งŒ ์ €์žฅ์ด ๋˜์‹œ๋‚˜์š”~? ์ œ๊ฐ€ ์ฝ”๋žฉ์œผ๋กœ ๋‹ค์‹œ ๋Œ๋ ค๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค~ ํ˜น์‹œ ์ฝ”๋“œ๋ฅผ ๋”ฐ๋กœ ์ˆ˜์ •ํ•œ ๊ฑด ์—†์œผ์‹œ๊ณ  ๋ฐ์ดํ„ฐ๋งŒ ๋ฐ”๊พธ์…จ๋‹ค๊ณ  ์ดํ•ดํ•˜๋ฉด ๋ ๊นŒ์š”?
@teddynote
@teddynote 4 ะฐะน ะฑาฑั€ั‹ะฝ
์•ˆ๋…•ํ•˜์„ธ์š”! ๋‹ต๋ณ€์ด ๋Šฆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฐฉ๊ธˆ ํ™•์ธํ•ด ๋ณธ ๊ฒฐ๊ณผ, JsonOutputParser ๊ฐ€ ๋‹ค์ˆ˜์˜ ๊ฒฐ๊ณผ๋ฅผ ์ œ๋Œ€๋กœ ํŒŒ์‹ฑํ•ด์˜ค์ง€ ๋ชปํ•œ ๋ฌธ์ œ๊ฐ€ ์žˆ์–ด, ๋กœ์ง์„ ์ˆ˜์ •ํ•˜์—ฌ ์ปค๋ฐ‹ํ•ด ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ์•„๋ž˜ ๋งํฌ์—์„œ ํ™•์ธํ•˜์‹ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์•Œ๋ ค์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค :) ๋งํฌ: github.com/teddylee777/langchain-kr/blob/main/15-FineTuning/01-QA-Pair-GPT.ipynb
@pixelel1353
@pixelel1353 7 ะฐะน ะฑาฑั€ั‹ะฝ
QA๋ฐ์ดํ„ฐ๋ฅผ ๋งŒ๋“œ๋Š” ๊ณผ์ •์—์„œ ๋ฌธ์„œ๋ฅผ ๋ถˆ๋Ÿฌ์˜ค๋Š” ๋ถ€๋ถ„์€ ์–ด๋””์— ์žˆ๋Š”๊ฑด๊ฐ€์š”?? ๊ทธ๋ฆฌ๊ณ  elements๊ฐ€ ์–ด๋””์— ์ •์˜๋ผ์žˆ๋Š”์ง€ ๋ชจ๋ฅด๊ฒ ์–ด์š” ใ… 
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
PDF ๋ถˆ๋Ÿฌ์˜ค๋Š” ์ฝ”๋“œ๋กœ ๋ถˆ๋Ÿฌ์˜ค์‹œ๋ฉด ๋ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ PDF ๋กœ๋”๋ฅผ ์‚ฌ์šฉํ•˜์…”๋„ ๋ฉ๋‹ˆ๋‹ค~ ์•„๋ž˜ ์ฝ”๋“œ๋ฅผ ์ฐธ๊ณ ํ•ด ์ฃผ์„ธ์š”~ from unstructured.partition.pdf import partition_pdf def extract_pdf_elements(filepath): """ PDF ํŒŒ์ผ์—์„œ ์ด๋ฏธ์ง€, ํ…Œ์ด๋ธ”, ๊ทธ๋ฆฌ๊ณ  ํ…์ŠคํŠธ ์กฐ๊ฐ์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. path: ์ด๋ฏธ์ง€(.jpg)๋ฅผ ์ €์žฅํ•  ํŒŒ์ผ ๊ฒฝ๋กœ fname: ํŒŒ์ผ ์ด๋ฆ„ """ return partition_pdf( filename=filepath, extract_images_in_pdf=False, # PDF ๋‚ด ์ด๋ฏธ์ง€ ์ถ”์ถœ ํ™œ์„ฑํ™” infer_table_structure=False, # ํ…Œ์ด๋ธ” ๊ตฌ์กฐ ์ถ”๋ก  ํ™œ์„ฑํ™” chunking_strategy="by_title", # ์ œ๋ชฉ๋ณ„๋กœ ํ…์ŠคํŠธ ์กฐ๊ฐํ™” max_characters=4000, # ์ตœ๋Œ€ ๋ฌธ์ž ์ˆ˜ new_after_n_chars=3800, # ์ด ๋ฌธ์ž ์ˆ˜ ์ดํ›„์— ์ƒˆ๋กœ์šด ์กฐ๊ฐ ์ƒ์„ฑ combine_text_under_n_chars=2000, # ์ด ๋ฌธ์ž ์ˆ˜ ์ดํ•˜์˜ ํ…์ŠคํŠธ๋Š” ๊ฒฐํ•ฉ ) # PDF ํŒŒ์ผ ๋กœ๋“œ elements = extract_pdf_elements("data/SPRI_AI_Brief_2023๋…„12์›”ํ˜ธ_F.pdf")
@pixelel1353
@pixelel1353 7 ะฐะน ะฑาฑั€ั‹ะฝ
@@teddynote ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค ใ… ใ…  ํ•™๊ต์—์„œ ํ”„๋กœ์ ํŠธํ•  ๋•Œ ๋งŽ์€ ๋„์›€๋ฐ›๊ณ ์žˆ์–ด์š”!!
@teddynote
@teddynote 7 ะฐะน ะฑาฑั€ั‹ะฝ
@@pixelel1353 ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค๐Ÿ™
@Younggoolee
@Younggoolee 6 ะฐะน ะฑาฑั€ั‹ะฝ
๋งŽ์€ ๋„์›€์„ ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ข‹์€ ์˜์ƒ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์˜ฌ๋ ค์ฃผ์‹  ์†Œ์Šค ์ฝ”๋“œ ์‹คํ–‰ ์ค‘ ์—๋Ÿฌ๊ฐ€ ๋ฐœ์ƒํ•˜๋Š”๋ฐ ํ•ด๊ฒฐ์ฑ…์— ์กฐ์–ธ์„ ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ํŒŒ์ธํŠœ๋‹ ์ฝ”๋“œ ์‹คํ–‰ ์ค‘ trainer_stats = trainer.train() ์—์„œ ๋‚˜๋Š” ์—๋Ÿฌ์ž…๋‹ˆ๋‹ค. Unsloth: Untrained tokens found, but embed_tokens & lm_head not trainable, causing NaNs. ์‹คํ–‰ํ™˜๊ฒฝ์€ Colab pro. GPU๋Š” L4 ํ• ๋‹น๋ฐ›์•„ ์‹คํ–‰ํ•ด ๋ณด์•˜์Šต๋‹ˆ๋‹ค. ๋„์›€ ๋ถ€ํƒ ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
@Younggoolee
@Younggoolee 6 ะฐะน ะฑาฑั€ั‹ะฝ
์•„ํ•˜.. ํ•ด๊ฒฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ์„  Colab pro ์—์„œ A100ํ• ๋‹น์„ ๋ฐ›์•„์•ผ ํ•˜๊ณ  embed token๊ณผ lm head๋ฅผ ์ถ”๊ฐ€ํ•˜๋ฉด ์˜ค๋ฅ˜๊ฐ€ ๋‚˜์ง€ ์•Š๊ณ  ๋ชจ๋ธ๋„ ์ž˜ ์‹คํ–‰๋˜์—ˆ์Šต๋‹ˆ๋‹ค. 150๊ฐœ ์—ฐ์Šต์‚ผ์•„ ๋งŒ๋“  ๋ฐ์ดํ„ฐ๋กœ๋„ ์„ฑ๋Šฅ์ด ๊ดœ์ฐฎ๊ฒŒ ๋‚˜์˜ค๋Š” ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค.
@vhxmtmfp
@vhxmtmfp 6 ะฐะน ะฑาฑั€ั‹ะฝ
ํ˜น์‹œ ๋ฐ์ดํ„ฐ์…‹ ์–ด๋–ป๊ฒŒ ๋งŒ๋“œ์…จ๋Š”์ง€ ์•Œ์ˆ˜์žˆ์„๊นŒ์š”? ์ €๋„ ๋ฐ์ดํ„ฐ์…‹ ๋งŒ๋“ค์–ด์„œ ํ•ด๋ณด๋ ค๊ณ ํ•˜๋Š”๋ฐ ๊ฒฐ๊ณผ๊ฐ’์ด ์ด์ƒํ•˜๊ฒŒ ๋‚˜์™€์„œ์š” ใ… ใ… 
@teddynote
@teddynote 6 ะฐะน ะฑาฑั€ั‹ะฝ
์šฐ์„  ํ•œ๊ตญ์–ด ํŒŒ์ธํŠœ๋‹ ๋œ ๋ชจ๋ธ์ด ์ตœ๊ทผ ์—…๋ฐ์ดํŠธ ๋˜๋ฉด์„œ ํ† ํฌ๋‚˜์ด์ € ์ด์Šˆ๊ฐ€ ์žˆ๋”๋ผ๊ณ ์š”. ๊ทธ๋ž˜์„œ ๋ฒ ์ด์Šค ๋ชจ๋ธ์„ unsloth/llama-3-8b-bnb-4bit ๋กœ ๋Œ๋ ค๋ณด์‹œ๊ณ ์š”. GPU๋Š” ๊ฐ€์šฉ ๊ฐ€๋Šฅํ•˜์‹œ๋‹ค๋ฉด A100์œผ๋กœ ๋Œ๋ ค๋ณด์„ธ์š”^^
@Younggoolee
@Younggoolee 6 ะฐะน ะฑาฑั€ั‹ะฝ
โ€‹@@teddynote์•„ํ•˜ ๊ทธ๋ ‡๊ตฐ์š” ๋‹ต๋ณ€ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค
@Younggoolee
@Younggoolee 6 ะฐะน ะฑาฑั€ั‹ะฝ
โ€‹@@vhxmtmfp๋งŽ์ด๋„ ์•„๋‹ˆ๊ณ  150๊ฐœ ์ƒ˜ํ”Œ ๋ฌธ์žฅ์— gpt4๋กœ ๋‹ต๋ณ€์„ ๋ผ๋ฒจ๋งํ•˜์˜€๊ตฌ์š” ํ•œ๋ฌธ์žฅ์˜ ๋‹ต๋ณ€์€ ์ด4๊ฐœ๊ฐ€ ๋‚˜์˜ค๊ฒŒ ํ–ˆ์Šต๋‹ˆ๋‹ค ํ…Œ๋””๋‹˜์—์„œ ์•Œ๋ ค์ค€ ๊ฒƒ์ฒ˜๋Ÿผ ํ—ˆ๊น…์— ๋ฐ์ดํ„ฐ์™€ ํŠœ๋‹๋œ ๋ชจ๋ธ์„ ์˜ฌ๋ ค๋†“์•˜์Šต๋‹ˆ๋‹ค Younggooo/senti_data2
ะ›ะตะพะฝ ะบะธะปะปะตั€ ะธ ะžะปั ะŸะพะปัะบะพะฒะฐ ๐Ÿ˜น
00:42
ะšะฐะฝะฐะป ะกะผะตั…ะฐ
ะ ะตั‚ า›ะฐั€ะฐะปะดั‹ 4,7 ะœะ›ะ
้ป‘ๅคฉไฝฟๅชๅฏนC็ฝ—ๆœ‰ๆ„Ÿ่ง‰#short #angel #clown
00:39
Super Beauty team
ะ ะตั‚ า›ะฐั€ะฐะปะดั‹ 36 ะœะ›ะ
Support each other๐Ÿค
00:31
ISSEI / ใ„ใฃใ›ใ„
ะ ะตั‚ า›ะฐั€ะฐะปะดั‹ 81 ะœะ›ะ
Cat mode and a glass of water #family #humor #fun
00:22
Kotiki_Z
ะ ะตั‚ า›ะฐั€ะฐะปะดั‹ 42 ะœะ›ะ
Alpaca LoRA ํŒŒ์ธํŠœ๋‹ - ๋‚˜๋งŒ์˜ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต์‹œํ‚ค๊ธฐ Colab ๊ฐ€๋Šฅ
11:05
๋นตํ˜•์˜ ๊ฐœ๋ฐœ๋„์ƒ๊ตญ
ะ ะตั‚ า›ะฐั€ะฐะปะดั‹ 37 ะœ.
๐Ÿง‘โ€๐Ÿ’ป #PDF ์ „์ฒ˜๋ฆฌ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฑด ๋‹ค ํ•ด๋ดค์Œ.
15:58
ํ…Œ๋””๋…ธํŠธ TeddyNote
ะ ะตั‚ า›ะฐั€ะฐะปะดั‹ 11 ะœ.
LLAMA-3 ๐Ÿฆ™: EASIET WAY To FINE-TUNE ON YOUR DATA ๐Ÿ™Œ
15:17
Prompt Engineering
ะ ะตั‚ า›ะฐั€ะฐะปะดั‹ 92 ะœ.
AI How To: Advanced LLM Prompting
17:23
Mckay Wrigley
ะ ะตั‚ า›ะฐั€ะฐะปะดั‹ 8 ะœ.
[SK TECH SUMMIT 2023] LLM ์ ์šฉ ๋ฐฉ๋ฒ•์ธ RAG VS PEFT, Domain ์ ์šฉ ์Šน์ž๋Š”?
23:41
ะ›ะตะพะฝ ะบะธะปะปะตั€ ะธ ะžะปั ะŸะพะปัะบะพะฒะฐ ๐Ÿ˜น
00:42
ะšะฐะฝะฐะป ะกะผะตั…ะฐ
ะ ะตั‚ า›ะฐั€ะฐะปะดั‹ 4,7 ะœะ›ะ