daVinci-Math is a unified multi-stage dataset for mathematical reasoning. Instead of curating separate resources for midtraining, supervised fine-tuning (SFT), and reinforcement learning (RL), we build a single stage-aware pipeline that starts from public math problem-solving sources, applies unified cleaning and deduplication, and routes each problem to the training stage where it is most useful.
The current release is designed to support three stages of training:… See the full description on the dataset page:
https://huggingface.co/datasets/davincimath/davinci-math.
