# slime

An LLM post-training framework for reinforcement-learning scaling, connecting Megatron training with SGLang rollout and flexible custom data-generation pipelines.

slime is a framework for LLM post-training focused on reinforcement-learning scaling. It provides high-performance training by connecting Megatron for training with SGLang for rollout, and flexible data generation through custom interfaces and server-based engines, so training, rollout, reward computation, verifier feedback and environment interaction all flow through the same training/rollout/data-buffer path.

It passes Megatron and SGLang arguments through directly rather than adding another abstraction layer, so new optimizations in either upstream project can be used without changes to slime itself. Math, code, search, tool use, sandboxes, verifiers and multi-agent workflows plug in as data-generation or reward workflows rather than forking the training kernel.

The framework has been used as the reinforcement-learning system behind several generations of the GLM model family, and documents reproducibility, fault tolerance, tracing, profiling and CI as engineering concerns alongside separate rollout-only and train-only debugging paths.

Website: https://thudm.github.io/slime