slime-rl-training

Orchestrate RL post-training for large language models with slime.

Updated Feb 15, 2026

Search Tags:#multi-turn #rl training #sglang #grpo #megatron #llm post-training

Ready to run

npx skills add https://github.com/ricable/mcai --skill slime-rl-training-ricable

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/ricable/mcai/tree/main/.agents/skills/slime-rl-training
Command: npx skills add https://github.com/ricable/mcai --skill slime-rl-training-ricable

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill helps you configure and run RL post-training pipelines for large language models when you need coordinated Megatron-LM training, SGLang rollouts, and reliable data buffering.

Core Features & Use Cases

Distributed RL Training: Set up GRPO, PPO, or related workflows across multi-GPU and multi-node environments.
Custom Generation and Rewards: Implement multi-turn tool use, specialized rollout logic, and custom reward functions.
Operational Guidance: Use the built-in reference material for model scripts, async training, evaluation, and troubleshooting.
Use Case: Prepare a Qwen or GLM model, point it at a JSONL dataset, and launch a production-style training run with tuned rollout and batch settings.

Quick Start

Ask for help configuring slime to train your model on a JSONL dataset with the right rollout, batch size, and reward settings.

Frequently Asked Questions about slime-rl-training

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I configure RL post-training for large language models using GRPO?▼

To configure RL post-training using GRPO, you need to orchestrate Megatron-LM training with SGLang rollouts, provide a JSONL prompt dataset, and specify resource and batch-size settings.

Can I run multi-turn agentic training with asynchronous rollouts?▼

Yes, multi-turn agentic training with asynchronous rollouts is supported. You can implement multi-turn tool use and specialized rollout logic by configuring slime to manage the generation workflow.

Does slime support distributed RL training across multi-GPU and multi-node environments?▼

Yes, slime supports distributed RL training across multi-GPU and multi-node environments. It coordinates Megatron-LM training and SGLang rollouts to scale GRPO and PPO workflows reliably.

How do I set up custom reward functions for PPO workflows?▼

You can set up custom reward functions for PPO workflows by integrating specialized rollout logic into the slime pipeline, using the built-in reference material to guide implementation and troubleshooting.

What model configurations are required for Qwen and GLM post-training?▼

Qwen and GLM post-training requires Megatron-LM-compatible model configurations, SGLang rollout services, JSONL prompt data, and explicit resource and batch-size settings to launch successfully.