FinShibainu / README.md
aiqwe's picture
Update README.md
1d8e3ee verified
|
raw
history blame
6.46 kB
metadata
tags:
  - krx
  - finance
license: mit
language:
  - ko

krx-llm-competition Model Card

๋ชจ๋ธ์€ KRX LLM ๊ฒฝ์ง„๋Œ€ํšŒ ๋ฆฌ๋”๋ณด๋“œ์—์„œ ์ตœ์ข… 3์œ„๋ฅผ ํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ์€ ๊ธˆ์œต, ํšŒ๊ณ„ ๋“ฑ ๊ธˆ์œต๊ด€๋ จ ์ง€์‹์— ๋Œ€ํ•œ Text Generation์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ์…‹ ์ˆ˜์ง‘ ๋ฐ ํ•™์Šต์— ๊ด€๋ จ๋œ ์ฝ”๋“œ๋Š” https://github.com/aiqwe/krx-llm-competition์— ์ž์„ธํ•˜๊ฒŒ ๊ณต๊ฐœ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ krx_model_card.pdf๋ฅผ ์ฐธ์กฐํ•ด์ฃผ์„ธ์š”.

Usage

https://github.com/aiqwe/krx-llm-competition์˜ example์„ ์ฐธ์กฐํ•˜๋ฉด ์‰ฝ๊ฒŒ inference๋ฅผ ํ•ด๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

pip install vllm
TBD

Model Card

Contents Spec
Base model Qwen2.5-7B-Instruct
Machine A100 SXM 80GB ร— 2
dtype bfloat16
PEFT LoRA (r=8, alpha=64)
Learning Rate 1e-5 (varies by further training)
LRScheduler Cosine (warm-up: 0.05%)
Optimizer AdamW
Distributed / Efficient Tuning DeepSpeed v3, Flash Attention
Global Batch Size 128

Datset Card

Reference ๋ฐ์ดํ„ฐ์…‹์€ ์ผ๋ถ€ ์ €์ž‘๊ถŒ ์ด์Šˆ๋กœ ์ธํ•ด Link๋กœ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. MCQA์™€ QA ๋ฐ์ดํ„ฐ์…‹์€ https://huggingface.co/datasets/aiqwe/krx-llm-competition์œผ๋กœ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค.
https://github.com/aiqwe/krx-llm-competition๋ฅผ ์ด์šฉํ•˜๋ฉด ๋‹ค์–‘ํ•œ ์œ ํ‹ธ๋ฆฌํ‹ฐ ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•˜๋ฉฐ, ๋ฐ์ดํ„ฐ ์†Œ์‹ฑ Pipeline์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

References

๋ฐ์ดํ„ฐ๋ช… url
ํ•œ๊ตญ์€ํ–‰ ๊ฒฝ์ œ๊ธˆ์œต ์šฉ์–ด 700์„  Link
์žฌ๋ฌดํšŒ๊ณ„ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ž์ฒด ์ œ์ž‘
๊ธˆ์œต๊ฐ๋…์šฉ์–ด์‚ฌ์ „ Link
web-text.synthetic.dataset-50k Link
์ง€์‹๊ฒฝ์ œ์šฉ์–ด์‚ฌ์ „ Link
ํ•œ๊ตญ๊ฑฐ๋ž˜์†Œ ๋น„์ •๊ธฐ ๊ฐ„ํ–‰๋ฌผ Link
ํ•œ๊ตญ๊ฑฐ๋ž˜์†Œ๊ทœ์ • Link
์ดˆ๋ณดํˆฌ์ž์ž ์ฆ๊ถŒ๋”ฐ๋ผ์žก๊ธฐ Link
์ฒญ์†Œ๋…„์„ ์œ„ํ•œ ์ฆ๊ถŒํˆฌ์ž Link
๊ธฐ์—…์‚ฌ์—…๋ณด๊ณ ์„œ ๊ณต์‹œ์ž๋ฃŒ Link
์‹œ์‚ฌ๊ฒฝ์ œ์šฉ์–ด์‚ฌ์ „ Link

MCQA

MCQA ๋ฐ์ดํ„ฐ๋Š” Reference๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค์ง€์„ ๋‹คํ˜• ๋ฌธ์ œ๋ฅผ ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ์…‹์ž…๋‹ˆ๋‹ค. ๋ฌธ์ œ์™€ ๋‹ต ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ Reasoning ํ…์ŠคํŠธ๊นŒ์ง€ ์ƒ์„ฑํ•˜์—ฌ ํ•™์Šต์— ์ถ”๊ฐ€ํ•˜์˜€์Šต๋‹ˆ๋‹ค.
ํ•™์Šต์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ๋Š” ์•ฝ 4.5๋งŒ๊ฐœ ๋ฐ์ดํ„ฐ์…‹์ด๋ฉฐ, tiktoken์˜ o200k_base(gpt-4o, gpt-4o-mini Tokenizer)๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์ด 2์ฒœ๋งŒ๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

๋ฐ์ดํ„ฐ๋ช… ๋ฐ์ดํ„ฐ ์ˆ˜ ํ† ํฐ ์ˆ˜
ํ•œ๊ตญ์€ํ–‰ ๊ฒฝ์ œ๊ธˆ์œต ์šฉ์–ด 700์„  1,203 277,114
์žฌ๋ฌดํšŒ๊ณ„ ๋ชฉ์ฐจ๋ฅผ ์ด์šฉํ•œ ํ•ฉ์„ฑ๋ฐ์ดํ„ฐ 451 99,770
๊ธˆ์œต๊ฐ๋…์šฉ์–ด์‚ฌ์ „ 827 214,297
hf_web_text_synthetic_dataset_50k 25,461 7,563,529
์ง€์‹๊ฒฝ์ œ์šฉ์–ด์‚ฌ์ „ 2,314 589,763
ํ•œ๊ตญ๊ฑฐ๋ž˜์†Œ ๋น„์ •๊ธฐ ๊ฐ„ํ–‰๋ฌผ 1,183 230,148
ํ•œ๊ตญ๊ฑฐ๋ž˜์†Œ๊ทœ์ • 3,015 580,556
์ดˆ๋ณดํˆฌ์ž์ž ์ฆ๊ถŒ๋”ฐ๋ผ์žก๊ธฐ 599 116,472
์ฒญ์†Œ๋…„์„ ์œ„ํ•œ ์ฆ๊ถŒ ํˆฌ์ž 408 77,037
๊ธฐ์—…์‚ฌ์—…๋ณด๊ณ ์„œ ๊ณต์‹œ์ž๋ฃŒ 3,574 629,807
์‹œ์‚ฌ๊ฒฝ์ œ์šฉ์–ด์‚ฌ์ „ 7,410 1,545,842
ํ•ฉ๊ณ„ 46,445 19,998,931

QA

QA ๋ฐ์ดํ„ฐ๋Š” Reference์™€ ์งˆ๋ฌธ์„ ํ•จ๊ป˜ Input์œผ๋กœ ๋ฐ›์•„ ์ƒ์„ฑํ•œ ๋‹ต๋ณ€๊ณผ Reference ์—†์ด ์งˆ๋ฌธ๋งŒ์„ Input์œผ๋กœ ๋ฐ›์•„ ์ƒ์„ฑํ•œ ๋‹ต๋ณ€ 2๊ฐ€์ง€๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.
Reference๋ฅผ ์ œ๊ณต๋ฐ›์œผ๋ฉด ๋ชจ๋ธ์€ ๋ณด๋‹ค ์ •ํ™•ํ•œ ๋‹ต๋ณ€์„ ํ•˜์ง€๋งŒ ๋ชจ๋ธ๋งŒ์˜ ์ง€์‹์ด ์ œํ•œ๋˜์–ด ๋‹ต๋ณ€์ด ์ข€๋” ์งง์•„์ง€๊ฑฐ๋‚˜ ๋‹ค์–‘์„ฑ์ด ์ค„์–ด๋“ค๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ด 4.8๋งŒ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹๊ณผ 2์–ต๊ฐœ์˜ ํ† ํฐ์œผ๋กœ ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

๋ฐ์ดํ„ฐ๋ช… ๋ฐ์ดํ„ฐ ์ˆ˜ ํ† ํฐ ์ˆ˜
ํ•œ๊ตญ์€ํ–‰ ๊ฒฝ์ œ๊ธˆ์œต ์šฉ์–ด 700์„  1,023 846,970
๊ธˆ์œต๊ฐ๋…์šฉ์–ด์‚ฌ์ „ 4,128 3,181,831
์ง€์‹๊ฒฝ์ œ์šฉ์–ด์‚ฌ์ „ 6,526 5,311,890
ํ•œ๊ตญ๊ฑฐ๋ž˜์†Œ ๋น„์ •๊ธฐ ๊ฐ„ํ–‰๋ฌผ 1,510 1,089,342
ํ•œ๊ตญ๊ฑฐ๋ž˜์†Œ๊ทœ์ • 4,858 3,587,059
๊ธฐ์—…์‚ฌ์—…๋ณด๊ณ ์„œ ๊ณต์‹œ์ž๋ฃŒ 3,574 629,807
์‹œ์‚ฌ๊ฒฝ์ œ์šฉ์–ด์‚ฌ์ „ 29,920 5,981,839
ํ•ฉ๊ณ„ 47,965 199,998,931