skills.rest
OpenLAIR
OpenLAIROFFICIAL
Explore all skills by this author →
GitHub

slime-rl-training

Unify Megatron-LM and SGLang for RL-based post-training of large language models.

1.0k|117|Updated Feb 26, 2026
Search Tags:#reinforcement-learning#slime#rl-training#sglang#post-training#megatron-lm
One-click install
npx skills add https://github.com/OpenLAIR/dr-claw --skill slime-rl-training-openlair
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/OpenLAIR/dr-claw/tree/main/skills/post-training/slime
Command: npx skills add https://github.com/OpenLAIR/dr-claw --skill slime-rl-training-openlair

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

slime-rl-training solves the complexity of orchestrating RL-based post-training for large language models by unifying Megatron-LM with SGLang into a single, reproducible workflow.

Core Features & Use Cases

  • Integrates Megatron-LM training with SGLang-based rollout, data buffering, and reward modeling to enable scalable RL for GLM, Qwen, Llama, and related models.
  • Provides predefined workflows for standard GRPO/GPO-like training, asynchronous training, and multi-turn agentic training, with pluggable custom generate and reward functions.
  • Use Case: You want to fine-tune an 8B LLM with RL signals from your own prompts and evaluate improvements in guided reasoning and tool use.

Quick Start

Launch slime post-training workflow with your model and dataset to begin RL-based post-training.

Frequently Asked Questions about slime-rl-training

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

FAQPage Schema
How do I orchestrate reinforcement learning post-training for large language models?▼

Reinforcement learning post-training for large language models is orchestrated by unifying Megatron-LM training with SGLang-based rollout into a single, reproducible workflow. This framework manages data generation, reward modeling, and multi-turn interactions end-to-end.

Can I use slime-rl-training to fine-tune Qwen and Llama models with custom reward functions?▼

Yes, slime-rl-training supports RL fine-tuning for Qwen, Llama, and GLM models. It provides pluggable hooks for custom generate and reward functions, allowing you to apply specific RL signals from your own prompts.

What is the best way to run asynchronous GRPO training for an 8B LLM?▼

Asynchronous GRPO training for an 8B LLM is best handled using predefined workflows that support asynchronous modes and co-location. This allows scalable rollout and reward modeling without blocking the Megatron-LM training cycles.

Does Megatron-LM work with SGLang for multi-turn agentic training?▼

Yes, Megatron-LM works with SGLang for multi-turn agentic training by using a unified framework that handles data buffering and modular argument handling. This integration enables complex interactions across models like GLM and Qwen.

Why does RL post-training require data buffering and modular argument handling?▼

RL post-training requires data buffering and modular argument handling to manage the high-throughput data generation and rollout phases efficiently. Separating Megatron, SGLang, and slime arguments ensures reproducible and scalable coordination during training.

Related Skills

Explore More Software Engineering Skills→
ovachieverovachiever
52

grpo-rl-training

Apply GRPO/RL training with TRL to optimize language models.

Community
Advanced
automatedigitalautomatedigital
1

unsloth

Fine-tune large language models with LoRA and QLoRA optimization.

Official
Intermediate
ovachieverovachiever
52

training-llms-megatron

Train large language models with Megatron-Core using configurable 3D and 4D parallelism.

Community
Advanced
Clay-HHKClay-HHK
2

slime-rl-training

Automates RL post-training for LLMs with Megatron-LM and SGLang workflows.

Community
Advanced