slime-rl-training

Trains large language models with GRPO reinforcement learning using Megatron-LM and SGLang.

2|Updated Feb 4, 2026
One-click install
npx skills add https://github.com/zhuangbiaowei/smart_bot --skill slime-rl-training-zhuangbiaowei
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/zhuangbiaowei/smart_bot/tree/main/skills/slime
Command: npx skills add https://github.com/zhuangbiaowei/smart_bot --skill slime-rl-training-zhuangbiaowei

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires sglang-router, ray, torch, transformers, and includes scripts (resource) and references (resource) components.

What problem does it solve?

This Skill streamlines the complex process of large language model post-training using Reinforcement Learning (RL), specifically by integrating Megatron-LM for efficient training and SGLang for high-throughput generation.

Core Features & Use Cases

  • RL Training: Implements advanced RL algorithms like GRPO for fine-tuning LLMs.
  • Megatron-LM Integration: Leverages Megatron-LM's distributed training capabilities for massive models.
  • SGLang Rollout: Utilizes SGLang for fast, scalable generation of training data and rollouts.
  • Use Case: Researchers and engineers can use this Skill to fine-tune large language models like GLM, Qwen, DeepSeek, or Llama 3 using RL, enabling them to improve model performance on specific tasks or align them with human preferences.

Quick Start

Use the slime-rl-training skill to launch GRPO training for the qwen3-4B model using the provided data file.

Frequently Asked Questions about slime-rl-training

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

FAQPage Schema
How do I scale reinforcement learning training for large language models using Megatron-LM?

You can scale LLM RL training by integrating Megatron-LM for distributed training and SGLang for high-throughput generation. This combination enables efficient GRPO fine-tuning of massive models like Qwen and Llama 3.

Can I use SGLang for rollout generation during LLM post-training?

Yes, SGLang is explicitly utilized for fast and scalable generation of training data and rollouts during LLM post-training. It works alongside Megatron-LM to facilitate high-throughput reinforcement learning workflows.

What dependencies do I need to run distributed GRPO training for models like Qwen or DeepSeek?

You need sglang-router, ray, torch, and transformers installed to execute distributed GRPO training. These dependencies support the execution of advanced RL algorithms for models like Qwen, DeepSeek, and GLM.

Does this RL training approach support fine-tuning Llama 3 models?

Yes, the RL training workflow supports fine-tuning Llama 3 models using reinforcement learning algorithms like GRPO. It allows researchers to improve model performance on specific tasks or align them with human preferences.

What is the best way to generate custom training data for reinforcement learning fine-tuning?

The best way to generate custom training data for RL fine-tuning is using SGLang for scalable rollouts. This approach supports custom data generation workflows while Megatron-LM handles the distributed training overhead.

Why use Megatron-LM and SGLang together for LLM post-training instead of other distributed systems?

Using Megatron-LM and SGLang together solves the bottleneck of scaling LLM post-training by combining efficient distributed training with high-throughput generation. This integration specifically streamlines complex RL algorithms like GRPO.