slime-rl-training

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

Updated Apr 21, 2026
One-click install
npx skills add https://github.com/samuelmukoti/myai-agent --skill slime-rl-training-samuelmukoti
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/samuelmukoti/myai-agent/tree/main/optional-skills/mlops/slime
Command: npx skills add https://github.com/samuelmukoti/myai-agent --skill slime-rl-training-samuelmukoti

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

slime provides a guided, production-ready workflow for post-training reinforcement learning on large language models using Megatron-LM and SGLang, unifying data generation, rollout, and evaluation in a scalable pipeline.

Core Features & Use Cases

  • RL-centric post-training orchestration for GLM, Qwen, and Llama models using slime
  • Data-buffered rollout generation, reward modeling, and multi-turn training support
  • Researchers and engineers deploying RL-based fine-tuning with Megatron-LM and SGLang in academic or industry settings

Quick Start

Install slime, configure the model scripts under scripts/models, and run the training workflow to start RL 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 set up RL post-training for large language models using Megatron-LM and SGLang?

RL post-training for large language models using Megatron-LM and SGLang requires installing the slime framework, configuring model scripts under the scripts/models directory, and executing the training workflow to orchestrate data generation, rollout, and reward modeling.

Can I use slime for multi-turn RL fine-tuning on Qwen and Llama models?

Yes, slime supports multi-turn RL fine-tuning and data-buffered rollout generation for GLM, Qwen, and Llama-family models, providing a unified pipeline for both academic research and production industry deployments.

What is the best way to configure reward modeling workflows for distributed training?

The best way to configure reward modeling workflows for distributed training is using slime's modular configuration scheme, which integrates Megatron, SGLang, and slime arguments to scale post-training data generation and evaluation.

Does slime support tool-calling and multi-turn interactions during RL training?

Yes, slime supports tool-calling and multi-turn training interactions during RL post-training, allowing researchers and engineers to orchestrate complex rollout generation and reward modeling workflows for large language models.

Why do I need a reward model for post-training large language models?

A reward model is needed for post-training large language models to evaluate and guide the reinforcement learning process, enabling the slime framework to scale data-buffered rollouts and optimize model performance across GLM, Qwen, or Llama architectures.

What troubleshooting resources are available if my slime RL training workflow fails?

The slime framework provides documented model scripts and troubleshooting resources within its optional components references and assets, helping resolve issues during distributed RL post-training with Megatron-LM and SGLang.