T3Q-LLM's picture
Update README.md
ea21195 verified
|
raw
history blame
3.72 kB
metadata
library_name: transformers
license: apache-2.0
pipeline_tag: text-generation
datasets:
  - maywell/ko_Ultrafeedback_binarized
base model:
  - yanolja/EEVE-Korean-Instruct-10.8B-v1.0

image/png

T3Q-LLM-sft1.0-dpo1.0

This model is a version of T3Q-LLM/T3Q-LLM-solar10.8-sft-v1.0 that has been fine-tuned with DPO.

Model Developers Chihoon Lee(chihoonlee10), T3Q

Prompt Template

A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.
Human: {prompt}
Assistant:

How to Use it

from transformers import AutoTokenizer
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("T3Q-LLM/T3Q-LLM-sft1.0-dpo1.0")
tokenizer = AutoTokenizer.from_pretrained("T3Q-LLM/T3Q-LLM-sft1.0-dpo1.0")

prompt_template = "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.\nHuman: {prompt}\nAssistant:\n"
text = 'ํ•œ๊ตญ์˜ ์ˆ˜๋„๋Š” ์–ด๋””์ธ๊ฐ€์š”? ์•„๋ž˜ ์„ ํƒ์ง€ ์ค‘ ๊ณจ๋ผ์ฃผ์„ธ์š”.\n\n(A) ๊ฒฝ์„ฑ\n(B) ๋ถ€์‚ฐ\n(C) ํ‰์–‘\n(D) ์„œ์šธ\n(E) ์ „์ฃผ'
model_inputs = tokenizer(prompt_template.format(prompt=text), return_tensors='pt')

outputs = model.generate(**model_inputs, max_new_tokens=256)
output_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
print(output_text)

Example Output

A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.
Human: ํ•œ๊ตญ์˜ ์ˆ˜๋„๋Š” ์–ด๋””์ธ๊ฐ€์š”? ์•„๋ž˜ ์„ ํƒ์ง€ ์ค‘ ๊ณจ๋ผ์ฃผ์„ธ์š”.

(A) ๊ฒฝ์„ฑ
(B) ๋ถ€์‚ฐ
(C) ํ‰์–‘
(D) ์„œ์šธ
(E) ์ „์ฃผ
Assistant:
(D) ์„œ์šธ์ด ํ•œ๊ตญ์˜ ์ˆ˜๋„์ž…๋‹ˆ๋‹ค. ์„œ์šธ์€ ๋‚˜๋ผ์˜ ๋ถ๋™๋ถ€์— ์œ„์น˜ํ•ด ์žˆ์œผ๋ฉฐ, ์ •์น˜, ๊ฒฝ์ œ, ๋ฌธํ™”์˜ ์ค‘์‹ฌ์ง€์ž…๋‹ˆ๋‹ค. ์•ฝ 1,000๋งŒ ๋ช…์ด ๋„˜๋Š” ์ธ๊ตฌ๋ฅผ ๊ฐ€์ง„ ์„ธ๊ณ„์—์„œ ๊ฐ€์žฅ ํฐ ๋„์‹œ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์„œ์šธ์€ ๋†’์€ ๋นŒ๋”ฉ, ํ˜„๋Œ€์ ์ธ ์ธํ”„๋ผ, ํ™œ๊ธฐ ๋ฌธํ™” ์žฅ๋ฉด์œผ๋กœ ์œ ๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋งŽ์€ ์—ญ์‚ฌ์  ๋ช…์†Œ์™€ ๋ฐ•๋ฌผ๊ด€์ด ์žˆ์–ด ๋ฐฉ๋ฌธ๊ฐ๋“ค์—๊ฒŒ ํ’๋ถ€ํ•œ ๋ฌธํ™” ์ฒดํ—˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
Task Version Metric Value Stderr
kobest_boolq 0 acc 0.9387 ยฑ 0.0064
macro_f1 0.9387 ยฑ 0.0064
kobest_copa 0 acc 0.7590 ยฑ 0.0135
macro_f1 0.7585 ยฑ 0.0135
kobest_hellaswag 0 acc 0.5080 ยฑ 0.0224
acc_norm 0.5580 ยฑ 0.0222
macro_f1 0.5049 ยฑ 0.0224
kobest_sentineg 0 acc 0.8489 ยฑ 0.0180
macro_f1 0.8483 ยฑ 0.0180

hf-causal-experimental (pretrained=nlpai-lab/KULLM3,use_accelerate=true,trust_remote_code=true), limit: None, provide_description: False, num_fewshot: 0, batch_size: 8

Task Version Metric Value Stderr
kobest_boolq 0 acc 0.8896 ยฑ 0.0084
macro_f1 0.8888 ยฑ 0.0084
kobest_copa 0 acc 0.6930 ยฑ 0.0146
macro_f1 0.6925 ยฑ 0.0147
kobest_hellaswag 0 acc 0.4640 ยฑ 0.0223
acc_norm 0.5240 ยฑ 0.0224
macro_f1 0.4612 ยฑ 0.0223
kobest_sentineg 0 acc 0.6297 ยฑ 0.0243
macro_f1 0.6255 ยฑ 0.0244