slime-rl-training

RL post-training with slime automates building, tuning, and evaluating reinforcement learning training pipelines across GLM, Qwen3, DeepSeek, and Llama 3 models. You configure Megatron-LM, SGLang, and slime settings, then execute the workflow with your model and data to start automated improvement.

Yes, slime supports multi-turn tool-use capabilities, data buffering, and rollout routing for both research and production use cases. These features enable scalable, memory-driven workflows and tool-assisted reasoning within reinforcement learning pipelines for large language models.

Yes, slime integrates Megatron-LM for training and SGLang for rollout to execute end-to-end RL post-training workflows. This combination requires configuring three specific categories—Megatron, SGLang, and slime—to establish the training pipeline and evaluate large language models.

Configuring a slime training pipeline requires setting parameters across three categories: Megatron, SGLang, and slime. You can customize and extend workflows for DeepSeek V3/R1 or Llama 3 by incorporating optional scripts, references, and assets into your configuration.

Reinforcement learning post-training enables automated improvement, memory-driven workflows, and tool-assisted reasoning in large language models. slime provides a production-grade framework to apply these techniques in scalable pipelines, offering evaluation, monitoring, and deployment-ready workflows for reproducibility.

slime currently supports GLM, Qwen3, DeepSeek V3/R1, and Llama 3 architectures for RL post-training. Users must ensure dependencies listed in the skill frontmatter are met and properly configure Megatron-LM and SGLang environments to avoid workflow execution issues.