ENERGY-DRINK-LOVE/eeve_dpo-v3

Our Team

• Jingyeom Kim
• Youjin Chung

Model

Base Model

• yanolja/EEVE-Korean-Instruct-10.8B-v1.0

Hardware and Software

• Hardware: A100 * 8 for training our model
• Deepspeed library & Huggingface TRL Trainer

Dataset

• DPO_dataset
  • 자체 제작 dpo dataset(AI-hub dataset 활용)
  • OpenOrca DPO 등 영어 데이터셋 번역(ENERGY-DRINK-LOVE/translate_share_gpt_dedup_llama_SFT_1024, 자체모델 활용)

Training Method

• DPO

Benchmark

Ko LM Eval Harness

Task	0-shot	5-shot
kobest_boolq	0.950142	0.944444
kobest_copa	0.751	0.835
kobest_hellaswag	0.474	0.508
kobest_sentineg	0.811083	0.972292
Average	0.74655625	0.81493399

Ko-LLM-Leaderboard

• (240307기준 7등)

  Average	Ko-ARC	Ko-HellaSwag	Ko-MMLU	Ko-TruthfulQA	Ko-CommonGen V2
  57.97	57.51	67.01	56.3	54.86	54.19