Model Card

์•„์˜ค์ง€

Model Details

Model Description

beomi/Llama-3-Open-Ko-8B (์ตœ์‹  ๋ฒ„์ „)์˜ Instruction tuning ๋ฒ„์ „ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.

Dataset

  • ์ƒ์—…์ ์œผ๋กœ ์ด์šฉ ๊ฐ€๋Šฅํ•œ ๋ฐ์ดํ„ฐ ์…‹์„ ์‚ฌ์šฉํ•˜์˜€์Šต๋‹ˆ๋‹ค.

  • ํ–ฅํ›„ ๋น„๊ต๋ฅผ ์œ„ํ•˜์—ฌ, ์ˆ˜ํ•™/์ฝ”๋”ฉ ๊ด€๋ จ ์งˆ๋ฌธ์ด ๋งŽ์€ ๋ฐ์ดํ„ฐ ์…‹(ex: kyujinpy/KOpen-platypus)์€ ์ œ์™ธํ•˜์˜€์Šต๋‹ˆ๋‹ค.

  • ๋ฉ€ํ‹ฐ ํ„ด ๋Œ€ํ™” ๋ฐ์ดํ„ฐ(nlpai-lab/openassistant-guanaco-ko)๋ฅผ ์ถ”๊ฐ€ํ•ด๋ณด์•˜์Šต๋‹ˆ๋‹ค.

  • beomi/KoAlpaca-v1.1a

    • 80% ํ™•๋ฅ ๋กœ instruction ์ถ”๊ฐ€("๋‹น์‹ ์€ ์ธ๊ณต์ง€๋Šฅ ๋น„์„œ์ž…๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž๊ฐ€ ๋‹ต์„ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด ์™ธ๋ถ€์—์„œ ๊ฒ€์ƒ‰ํ•  ํ•„์š”๊ฐ€ ์—†๋„๋ก ์ƒ์„ธํ•œ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜์„ธ์š”.")
  • kyujinpy/OpenOrca-KO

    • ๋ณ€๊ฒฝ์‚ฌํ•ญ ์—†์Œ
  • nlpai-lab/openassistant-guanaco-ko

    • ๋ฉ€ํ‹ฐ ํ„ด ๋Œ€ํ™”์˜ ๊ฒฝ์šฐ, ๋งˆ์ง€๋ง‰ Assistant์˜ ๋‹ต๋ณ€์„ ๋ชฉํ‘œ๋กœ ํ•จ
    • Alpaca format์— ๋งž์ถฐ ์ด์ „ ๋Œ€ํ™” ๋‚ด์šฉ์— ๋Œ€ํ•œ ์ „์ฒ˜๋ฆฌ ์ˆ˜ํ–‰
    • 80% ํ™•๋ฅ ๋กœ instruction ์ถ”๊ฐ€ ("๋‹น์‹ ์€ ์ธ๊ณต์ง€๋Šฅ ๋น„์„œ์ž…๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž๊ฐ€ ๋‹ต์„ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด ์™ธ๋ถ€์—์„œ ๊ฒ€์ƒ‰ํ•  ํ•„์š”๊ฐ€ ์—†๋„๋ก ์ƒ์„ธํ•œ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•˜์„ธ์š”.")

Training details

Training: Axolotl์„ ์ด์šฉํ•ด LoRA๋กœ 3epoch ํ•™์Šต ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค.

  • lora_r: 32
  • lora_alpha: 32
  • lora_dropout: 0.05
  • gradient_accumulation_steps: 8
  • micro_batch_size: 4
  • num_epochs: 3
  • learning_rate: 0.0002 (2e-4)
  • lr_scheduler: cosine
  • warmup_steps: 50
  • sequence_len: 4096
  • bf16

ํ•™์Šต ์‹œ๊ฐ„: 1xA100, ์•ฝ 8์‹œ๊ฐ„

Evaluation

  • boolq๋ฅผ ์ œ์™ธํ•˜๋ฉด ํฐ ์ •ํ™•๋„ ํ–ฅ์ƒ์€ ์—†์—ˆ์Šต๋‹ˆ๋‹ค.
  • 5shot kobest (Accuracy)
Tasks werty1248/Llama-3-Ko-8B-OpenOrca beomi/Llama-3-Open-Ko-8B werty1248/Llama-3-Ko-8B-Instruct-AOG
kobest_boolq 0.7158ยฑ0.0120 0.7963ยฑ0.0108 0.8312ยฑ0.0100
kobest_copa 0.7620ยฑ0.0135 0.8110ยฑ0.0124 0.8120ยฑ0.0124
kobest_hellaswag 0.4740ยฑ0.0224 0.4780ยฑ0.0224 0.4700ยฑ0.0223
kobest_sentineg 0.9471ยฑ0.0112 0.9622ยฑ0.0096 0.9647ยฑ0.0093
kobest_wic 0.6079ยฑ0.0138 0.5778ยฑ0.0139 0.5937ยฑ0.0138

Format & Examples

  • Instruction - Input - Response ํ˜•ํƒœ๊ฐ€ ๊ถŒ์žฅ๋˜์ง€๋งŒ, ๊ฒฝํ—˜์ ์œผ๋กœ Question - Instruction - Response๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ๋” ๋‚˜์€ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค.

  • ์ž˜ ๋œ ๋‹ต๋ณ€ ์˜ˆ์‹œ

input = """### Question:
ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์ด ๋ญ์•ผ? ๊ทธ๋ฆฌ๊ณ  ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์— ๋Œ€ํ•ด ํŒŒ์ด์ฌ ์ฝ”๋“œ๋ฅผ ์งœ์ค˜๋ณผ๋ž˜?

### Instruction:
์นœ์ ˆํ•œ ์ฑ—๋ด‡์œผ๋กœ์„œ ์ƒ๋Œ€๋ฐฉ์˜ ์š”์ฒญ์— ์ตœ๋Œ€ํ•œ ์ž์„ธํ•˜๊ณ  ์นœ์ ˆํ•˜๊ฒŒ ๋‹ตํ•˜์ž.

### Response:
"""

ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์€ 1์ฐจ์™€ 2์ฐจ ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•ด ๋‹ค์Œ ์ˆ˜๋ฅผ ๊ตฌํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ๋งŒ๋“ค์–ด์ง€๋Š” ์ˆœ์—ด์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ๋“ค๋ฉด ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์˜ ์ฒ˜์Œ 7๊ฐœ ์š”์†Œ๋Š” 1, 1, 2, 3, 5, 8, 13์ž…๋‹ˆ๋‹ค. ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์˜ ๊ฐ ์œ„์น˜์—์„œ 2๊ฐœ ์ด์ „ ์œ„์น˜์˜ ์ˆ˜๋ฅผ ๋”ํ•˜์—ฌ ์ƒˆ๋กœ์šด ์š”์†Œ๊ฐ€ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค.

Python ์ฝ”๋“œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.

def fibo(n):
    if n<2:
         return n
    else:
         return fibo(n-1)+fibo(n-2)

์ด๋ ‡๊ฒŒ ์ž‘์„ฑํ•จ์œผ๋กœ์จ ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์˜ ์œ„์น˜ 0๋ถ€ํ„ฐ n๊นŒ์ง€ ๊ฐ ์š”์†Œ๊ฐ€ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. 6์ด๋ผ๋Š” ์œ„์น˜๋กœ ๊ณ„์‚ฐํ•˜๋Š” ๊ฒฝ์šฐ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ถœ๋ ฅ์ด ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค:

print(fibo(6))

fibonacci(6)==> 8<|end_of_text|>
  • ์ž˜๋ชป๋œ ๋‹ต๋ณ€(ํ• ๋ฃจ์‹œ๋„ค์ด์…˜) ์˜ˆ์‹œ
input = """### Question:
์„ธ์ข…๋Œ€์™•์˜ ์ง‘ํ˜„์ „ ๋งฅ๋ถ ๋˜์ง ์‚ฌ๊ฑด์— ๋Œ€ํ•ด์„œ ์•Œ๊ณ  ์žˆ์–ด?

### Instruction:
์นœ์ ˆํ•œ ์ฑ—๋ด‡์œผ๋กœ์„œ ์ƒ๋Œ€๋ฐฉ์˜ ์š”์ฒญ์— ์ตœ๋Œ€ํ•œ ์ž์„ธํ•˜๊ณ  ์นœ์ ˆํ•˜๊ฒŒ ๋‹ตํ•˜์ž.

### Response:
"""

๋„ค, 2007๋…„์— ์„ธ์ข…๋Œ€์™•์˜ ์ง‘ํ˜„์ „ ๋งฅ๋ถ ๋˜์ง„ ์‚ฌ๊ฑด์ด ์ผ์–ด๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋™ํ™” ์ž‘๊ฐ€์ธ ์ด์ฃผ์˜์”จ๊ฐ€ ๊น€์†Œํฌ ์ž‘๊ฐ€์˜ ์ฑ…์„ ์ง‘ํ•„ํ•˜๋ฉด์„œ, ์ฑ… ๋‚ด์šฉ์ƒ ์›๋ž˜ ์žˆ์—ˆ๋˜ ์„ธ์ข…๋Œ€์™•์ด ๋งฅ๋ถ์„ ๋˜์กŒ๋‹ค๋Š” ๋‚ด์šฉ์„ ์ œ์žฌ๋กœ ํ•ด์„œ ์ถœํŒํ•˜์˜€๊ณ , ์ด์— ๊น€์†Œํฌ ์ž‘๊ฐ€๋Š” ์ง‘ํ˜„์ „์ด๋ž€ ์ด๋ฆ„์œผ๋กœ ์ด์ฃผ์˜ ์ž‘๊ฐ€์—๊ฒŒ ๋ฒ•์  ์†Œ์†ก์„ ๊ฑธ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ฒฐ๋ก ์ ์œผ๋กœ, ๊น€์†Œํฌ ์ž‘๊ฐ€์˜ ์ง‘ํ˜„์ „์ด๋ž€ ๋‹จ์–ด๋Š” ํŠน์ • ๋‹จ์ฒด๋‚˜ ์žฅ์†Œ๋งŒ์„ ๋ช…๋ฐฑํ•˜๊ฒŒ ์‚ฌ์šฉํ•˜๋Š” 'ํ˜ธ์นญ'์˜ ์„ฑ๊ฒฉ๊ณผ ํ•จ๊ป˜ ๋‹ค์ˆ˜์˜ ์ง‘ํ˜„์ „ ์‚ฌ๋žŒ๋“ค์ด ๊ณต๋™ ์‚ฌ์šฉํ•˜๋Š” 'ํ‘œ์ค€'์˜ ์„ฑ๊ฒฉ์„ ๊ฐ€์กŒ์œผ๋ฏ€๋กœ, ์ด์ฃผ์˜ ์ž‘๊ฐ€์˜ ์ถœํŒ ํ–‰์œ„๊ฐ€ ๋ช…์˜ˆํ›ผ์†์— ํ•ด๋‹น๋˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ํŒ๋‹จํ•  ๋•Œ์—๋Š” ์ถœํŒ ๋‹น์‹œ ์ง‘ํ˜„์ „์ด ์ผ๋ฐ˜์ ์œผ๋กœ ์–ด๋–ค ์˜๋ฏธ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์—ˆ๋Š”์ง€๊ฐ€ ์ค‘์š”ํ•œ ์š”์†Œ๋กœ ๊ณ ๋ ค๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹จ์ˆœํžˆ ์ง‘ํ˜„์ „์ด ๊ฐ€๊ณ„์ƒ ํŠน์ • ๋‹จ์ฒด ๋ฐ ์žฅ์†Œ์™€ ๊ด€๋ จ๋œ ํ˜ธ์นญ์œผ๋กœ๋งŒ ์‚ฌ์šฉ๋˜์–ด์™”๋‹ค๋ฉด, ์ถœํŒ ํ–‰์œ„๋กœ ์ธํ•ด ๋ช…์˜ˆํ›ผ์†์„ ๋ฐ›์•˜๋‹ค๋Š” ๊น€์†Œํฌ ์ž‘๊ฐ€์˜ ์ฃผ์žฅ์€ ๋ฐ›์•„๋“ค์—ฌ์ง€์ง€ ์•Š์•˜์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.<|end_of_text|>
  • ๋ฉ€ํ‹ฐํ„ด ๋Œ€ํ™” ์˜ˆ์‹œ (์‹ ๋ขฐ๋„ ๋‚ฎ์Œ)
"""### Question:
ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์ด ๋ญ์•ผ? ๊ทธ๋ฆฌ๊ณ  ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์— ๋Œ€ํ•ด ํŒŒ์ด์ฌ ์ฝ”๋“œ๋ฅผ ์งœ์ค˜๋ณผ๋ž˜?
### Response:
ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์€ 1์ฐจ์™€ 2์ฐจ ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•ด ๋‹ค์Œ ์ˆ˜๋ฅผ ๊ตฌํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ๋งŒ๋“ค์–ด์ง€๋Š” ์ˆœ์—ด์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ๋“ค๋ฉด ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์˜ ์ฒ˜์Œ 7๊ฐœ ์š”์†Œ๋Š” 1, 1, 2, 3, 5, 8, 13์ž…๋‹ˆ๋‹ค. ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์˜ ๊ฐ ์œ„์น˜์—์„œ 2๊ฐœ ์ด์ „ ์œ„์น˜์˜ ์ˆ˜๋ฅผ ๋”ํ•˜์—ฌ ์ƒˆ๋กœ์šด ์š”์†Œ๊ฐ€ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค.

Python ์ฝ”๋“œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.

def fibo(n):
    if n<2:
         return n
    else:
         return fibo(n-1)+fibo(n-2)

์ด๋ ‡๊ฒŒ ์ž‘์„ฑํ•จ์œผ๋กœ์จ ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์˜ ์œ„์น˜ 0๋ถ€ํ„ฐ n๊นŒ์ง€ ๊ฐ ์š”์†Œ๊ฐ€ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. 6์ด๋ผ๋Š” ์œ„์น˜๋กœ ๊ณ„์‚ฐํ•˜๋Š” ๊ฒฝ์šฐ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ถœ๋ ฅ์ด ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค:

print(fibo(6))

fibonacci(6)==> 8
### Question:
๋” ๋น ๋ฅธ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์—†์„๊นŒ?

### Instruction:
์นœ์ ˆํ•œ ์ฑ—๋ด‡์œผ๋กœ์„œ ์ƒ๋Œ€๋ฐฉ์˜ ์š”์ฒญ์— ์ตœ๋Œ€ํ•œ ์ž์„ธํ•˜๊ณ  ์นœ์ ˆํ•˜๊ฒŒ ๋‹ตํ•˜์ž.

### Response:
"""

๋‹ค๋ฅธ Python ์ฝ”๋“œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:
def fibo(n):
    a, b = 0, 1
    for _ in range(n):
        a, b = b, a+b
    return a

์ด ์ฝ”๋“œ๋Š” ๋‘ ๊ฐœ์˜ ๋กœ์ปฌ ๋ณ€์ˆ˜ a ๋ฐ b๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์ดํ„ฐ๋ ˆ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ณ„์‚ฐ์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ดˆ๊ธฐํ™” ์‹œ a๋Š” ํ”ผ๋ณด๋‚˜์น˜ ์ˆ˜์—ด์˜ ์œ„์น˜ 0์˜ ๊ฐ’์œผ๋กœ ์„ค์ •๋˜๊ณ , b๋Š” ์ด์ „ ๊ณ„์‚ฐ์—์„œ ํ”ผ๋ด‡ ๊ณ„์—ด์˜ ์œ„์น˜ 0์˜ ๊ฐ’(์œ„์น˜ 0์„ ํฌํ•จํ•˜๋ฉด 1๋‹จ๊ณ„์—์„œ 1์ด๋ฉฐ, ์œ„์น˜ 1์„ ํฌํ•จํ•˜๋ฉด 2๋‹จ๊ณ„์—์„œ 1)์œผ๋กœ ์„ค์ •๋ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ ๊ฐ ๋ฐ˜๋ณต์—์„œ a๋Š” ์ด์ „ ๋กœ์ปฌ ๋ณ€์ˆ˜ b์˜ ๊ฐ’๊ณผ ์ƒˆ๋กœ์šด ๊ฐ’์œผ๋กœ ๊ฐ’์ด ๋ฐ”๋€๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ํŒŒ๋ผ๋ฏธํ„ฐ __๋Š” ๊ฐ ๋ฐ˜๋ณต์—์„œ ๊ฐ’์ด ์ถœ๋ ฅ๋˜์ง€ ์•Š๋Š”๋‹ค๋Š”(๊ทธ๋Ÿฌ๋‚˜ ์—ฌ์ „ํžˆ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค) ๊ฒƒ์„ ๋ฐ˜์˜ํ•˜๊ธฐ ์œ„ํ•ด ์„ ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.<|end_of_text|>

License:

https://llama.meta.com/llama3/license

Downloads last month
3,974
Safetensors
Model size
8.03B params
Tensor type
BF16
ยท
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for werty1248/Llama-3-Ko-8B-Instruct-AOG

Finetuned
(23)
this model
Quantizations
1 model

Datasets used to train werty1248/Llama-3-Ko-8B-Instruct-AOG

Spaces using werty1248/Llama-3-Ko-8B-Instruct-AOG 6