slime-rl-training

Name: slime-rl-training
Availability: InStock
Author: zerwiz

Orchestrate reinforcement learning post-training for large language models with Megatron-LM and SGLang.

1|Updated Apr 10, 2026

Search Tags:#reinforcement learning #sglang #rlhf #grpo #post-training #megatron-lm #llm training

Ready to run

npx skills add https://github.com/zerwiz/wayofpi --skill slime-rl-training-zerwiz

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/zerwiz/wayofpi/tree/main/.hermes/hermes-agent/optional-skills/mlops/slime
Command: npx skills add https://github.com/zerwiz/wayofpi --skill slime-rl-training-zerwiz

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

It removes the complexity of running large language model reinforcement learning post-training by combining training, rollout generation, data buffering, and evaluation into one coordinated workflow.

Core Features & Use Cases

Megatron and SGLang Integration: Trains models with Megatron-LM while generating rollouts through SGLang for high-throughput sampling.
Flexible RL Workflows: Supports GRPO, PPO, async training, off-policy buffering, and multi-turn agentic generation with custom reward logic.
Model and Scale Coverage: Works across GLM, Qwen, DeepSeek, Llama, and other model families for research or production-grade RL experiments.
Use Case: A machine learning engineer can use this Skill to launch a GRPO run on a Qwen model, validate rewards, monitor stability issues, and tune rollout settings without assembling the pipeline manually.

Quick Start

Use this skill to set up a slime training run for your model, JSONL prompt data, and chosen RL configuration.

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 reinforcement learning post-training for large language models using GRPO?▼

Reinforcement learning post-training with GRPO requires orchestrating Megatron-LM training and SGLang rollouts. You can execute this by configuring a coordinated workflow with structured JSONL prompts, custom reward logic, and Ray orchestration.

What is the best way to set up async RLHF with Megatron-LM and SGLang?▼

Async RLHF is configured by integrating Megatron-LM for model training and SGLang for high-throughput rollout generation. This combination enables off-policy buffering and multi-turn agentic loops through Ray orchestration.

Does this RL training workflow support Qwen, DeepSeek, and Llama model families?▼

Yes, the RL training workflow supports GLM, Qwen, DeepSeek, and Llama model families. It applies GRPO and PPO configurations across these architectures for both research and production-grade experiments.

Can I use custom reward logic and multi-turn agentic generation for LLM training?▼

Custom reward logic and multi-turn agentic generation are fully supported for LLM training. The workflow allows you to define specific reward mechanisms and manage complex agentic loops during SGLang rollout generation.

Do I need Ray orchestration to run PPO training with Megatron-LM?▼

Yes, Ray orchestration is required to coordinate PPO training and SGLang rollouts with Megatron-LM. It manages the distributed training, data buffering, and evaluation components within the unified workflow.

Why use SGLang for rollouts in reinforcement learning post-training?▼

SGLang is used for rollouts in reinforcement learning post-training to provide high-throughput sampling during generation. This ensures efficient data buffering and stable model updates across training iterations.