slime

slime is an open-source LLM post-training framework for reinforcement-learning scaling, combining Megatron-based high-performance training with SGLang rollout/serving and flexible data-generation interfaces for RL and agentic workflows.