This dataset contains 2,450 ORPO preference pairs with a targeted violation mix: 50% severe, 30% moderate, 20% minor. All rejected outputs are syntax-valid (or parseable inside fences) and represent clear instruction violations.
Seed: u-10bei/structured_data_with_cot_dataset_512_v4 (train split)
v2 added: prompt constraint augmentation + margin scoring.
v3 adds: violation categorization and… See the full description on the dataset page:
https://huggingface.co/datasets/daichira/structured-orpo-503020-v3.
