cdrrm-rubrics

# Dataset Card for CDRRM ## Dataset Description This dataset contains pairwise comparison data with discriminative rubrics, which is suitable for the training and evaluation of reward models. The original data is derived from the OpenRubrics dataset. ### Data Fields - `instruction`: Task instruction/prompt - `response_a`: First response to compare - `response_b`: Second response to compare - `winner`: Ground-truth winner ("A" or "B") - `formatted_rubric`: Formatted rubric text (string) containing hard rules and principles for judge. - `rubric`: Rubric for training rubric generator. ### Source Data This dataset is generated using the CDRRM pipeline, which: 1. Generates structured diagnoses for each response 2. Creates discriminative rubrics that distinguish between chosen and rejected responses 3. Validates rubric consistency with ground-truth labels