hw-llama-2-7B-nsmc / README.md
seojin0128's picture
Update README.md
03a6641
|
raw
history blame
1.68 kB
metadata
library_name: peft
base_model: meta-llama/Llama-2-7b-chat-hf
metrics:
  - accuracy 89.7%
datasets:
  - nsmc

Model Card for Model ID

Model Description

llama-2-7b-chat-hf ๋ฏธ์„ธ ํŠœ๋‹

ํ•ด๋‹น ๋ชจ๋ธ์€ ๋„ค์ด๋ฒ„ ์˜ํ™” ๋ฆฌ๋ทฐ ๋ฐ์ดํ„ฐ์…‹์ธ NSMC์— ๋Œ€ํ•ด meta-llama/Llama-2-7b-chat-hf์„ ๋ฏธ์„ธํŠœ๋‹ํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.

์˜ํ™” ๋ฆฌ๋ทฐ ํ…์ŠคํŠธ๋ฅผ ํ”„๋กฌํ”„ํŠธ์— ํฌํ•จํ•˜์—ฌ ๋ชจ๋ธ์— ์ž…๋ ฅ์‹œ,'๊ธ์ •' ๋˜๋Š” '๋ถ€์ •' ์ด๋ผ๊ณ  ์˜ˆ์ธก ํ…์ŠคํŠธ๋ฅผ ์ง์ ‘ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.

๊ฒฐ๊ณผ์ ์œผ๋กœ, ์ •ํ™•๋„ 89.7%๋ฅผ ๊ฐ€์ง€๋Š” ๋ชจ๋ธ์„ ์™„์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.

Train, Test ๋ฐ์ดํ„ฐ์…‹

ํ•ด๋‹น ๋ชจ๋ธ์€ NSMC์˜ train ๋ฐ์ดํ„ฐ์˜ ์ƒ์œ„ 2,000๊ฐœ์˜ ์ƒ˜ํ”Œ์„ ํ•™์Šต์— ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.

ํ•ด๋‹น ๋ชจ๋ธ์€ NSMC์˜ test ๋ฐ์ดํ„ฐ์˜ ์ƒ์œ„ 1,000๊ฐœ์˜ ์ƒ˜ํ”Œ์„ ํ‰๊ฐ€์— ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.

Training procedure

The following bitsandbytes quantization config was used during training:

  • quant_method: bitsandbytes
  • load_in_8bit: False
  • load_in_4bit: True
  • llm_int8_threshold: 6.0
  • llm_int8_skip_modules: None
  • llm_int8_enable_fp32_cpu_offload: False
  • llm_int8_has_fp16_weight: False
  • bnb_4bit_quant_type: nf4
  • bnb_4bit_use_double_quant: False
  • bnb_4bit_compute_dtype: bfloat16

Training_step_loss

image/png

Confusion_Matrix

image/png

Accuracy_Classification_Report

image/png

Framework versions

  • PEFT 0.7.0