CRPS-30K is a high-quality mathematical reasoning dataset synthesized using the Contrastive Reasoning Path Synthesis (CRPS) framework. Unlike standard rejection sampling that only retains correct trajectories, CRPS leverages the contrast between successful and unsuccessful MCTS reasoning paths to synthesize training data that incorporates explicit knowledge of common pitfalls and strategic reasoning patterns.… See the full description on the dataset page: https://huggingface.co/datasets/PeiyangLiu/CRPS-30K.