slime-rl-training

Launch GRPO training for Qwen3 models with Megatron-LM and SGLang.

Updated Aug 23, 2026
One-click install
npx skills add https://github.com/kwasi-cpu/hermes-agent --skill slime-rl-training
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/kwasi-cpu/hermes-agent/tree/main/skills/mlops/training/slime
Command: npx skills add https://github.com/kwasi-cpu/hermes-agent --skill slime-rl-training

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires sglang-router>=0.2.3, ray, torch>=2.0.0, transformers>=4.40.0, and includes scripts (resource) and references (resource) components.

What problem does it solve?

This Skill streamlines the complex process of Reinforcement Learning (RL) based post-training for Large Language Models (LLMs), enabling efficient scaling and customization.

Core Features & Use Cases

  • RL Training Framework: Integrates Megatron-LM for efficient distributed training with SGLang for high-throughput inference rollouts.
  • Model Support: Compatible with a wide range of models including GLM, Qwen, DeepSeek, and Llama.
  • Use Case: Fine-tune a Llama 3 model using RLHF to improve its conversational abilities and safety alignment, leveraging custom data generation workflows.

Quick Start

Launch GRPO training for a Qwen3 model using the provided script and data.

Frequently Asked Questions about slime-rl-training

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

FAQPage Schema
How do I run RLHF training for large language models using Megatron-LM?

RLHF training for large language models using Megatron-LM is executed through a framework that integrates distributed training with SGLang for high-throughput inference rollouts. You launch GRPO training by executing provided scripts with your target model and custom data.

Can I use SGLang for inference rollouts during reinforcement learning post-training?

Yes, you can use SGLang for inference rollouts during reinforcement learning post-training. The framework integrates SGLang specifically to handle high-throughput inference generation alongside Megatron-LM's distributed training capabilities.

Does this reinforcement learning framework support Qwen, DeepSeek, and Llama models?

This reinforcement learning framework supports Qwen, DeepSeek, and Llama models, alongside GLM. It is compatible with a wide range of model families for post-training tasks like conversational ability enhancement and safety alignment.

What dependencies do I need to set up distributed LLM training with SGLang and Ray?

To set up distributed LLM training with SGLang and Ray, you need specific dependencies including sglang-router>=0.2.3, ray, torch>=2.0.0, and transformers>=4.40.0. These facilitate the asynchronous training pipelines and inference execution.

What is the best way to fine-tune a Llama 3 model for safety alignment with custom data generation?

The best way to fine-tune a Llama 3 model for safety alignment with custom data generation is using an RL training framework. It streamlines the complex process by leveraging custom data generation workflows and asynchronous training pipelines for efficient scaling.

How does asynchronous training pipeline integration work for multi-turn agentic LLM training?

Asynchronous training pipeline integration for multi-turn agentic LLM training works by combining Megatron-LM for distributed processing and SGLang for inference rollouts. This architecture supports multi-turn agentic training and custom data generation workflows efficiently.