PRIME

Why we included this project

If you are training an open-weight model to reason better than supervised fine-tuning alone can manage, PRIME gives you a working recipe rather than a paper abstract. It implements Process Reinforcement through Implicit Rewards, an online RL approach that rewards each step of a solution without needing hand-annotated process labels. The repo includes the full training, evaluation, and data-preprocessing code, and the method already lives in the veRL framework, so you can move from design to runs without reimplementing everything. Released checkpoints and datasets make it possible to reproduce reported results or adapt the recipe to your own base model. This is aimed at researchers and applied RL engineers who want reasoning gains beyond imitation or distillation.