Taichi11/

dpo_dataset_v4

	DPO Dataset (Filtered + Structured Boosted)




	
	Base Dataset

u-10bei/dpo-dataset-qwen-cot

prompt <= 2048 chars chosen/rejected <= 2048 chars length difference > 80 chars

XML / TOML samples boosted by 1.2x

3101 samples

DPO training for Qwen3-4B-Instruct (SFT-initialized) Structured output improvement (XML / TOML)

Google Colab NVIDIA T4… See the full description on the dataset page: https://huggingface.co/datasets/Taichi11/dpo_dataset_v4.