jaeyong2/Qwen3-1.7B-DPO-Peft-Preview

Evaluation

!lm_eval --model hf \
    --model_args pretrained=jaeyong2/Qwen3-0.6B-DPO-Ja-Peft \
    --tasks kmmlu,mmlu,japanese_leaderboard,gsm8k \
    --device cuda:0 \
    --batch_size 1 

	Qwen3-1.7B-DPO-peft	Qwen3-1.7B
MMLU	0.55	0.55
ja_leaderboard_jaqket_v2	0.35	0.34
ja_leaderboard_jcommonsenseqa	0.48	0.46
ja_leaderboard_jnli	0.28	0.23
ja_leaderboard_jsquad	0.25	0.22
ja_leaderboard_marc_ja	0.65	0.74
ja_leaderboard_mgsm	0.42	0.40
ja_leaderboard_xlsum	0.10	0.11
ja_leaderboard_xwinograd	0.58	0.58
GSM8K	0.70	0.69
KMMLU	0.36	0.36

License

• Qwen/Qwen3-0.6B : https://choosealicense.com/licenses/apache-2.0/

Acknowledgement

This research is supported by TPU Research Cloud program.