slime-rl-training

Orchestrate end-to-end reinforcement learning post-training for large language models with slime.

Updated Apr 1, 2026
One-click install
npx skills add https://github.com/founderphantom/zola-agent --skill slime-rl-training-founderphantom
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/founderphantom/zola-agent/tree/main/optional-skills/mlops/slime
Command: npx skills add https://github.com/founderphantom/zola-agent --skill slime-rl-training-founderphantom

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

slime RL post-training helps teams efficiently improve LLMs through reinforcement learning loops, automated data generation, rollout orchestration, and iterative model updates using the slime ecosystem.

Core Features & Use Cases

  • Megatron-LM native training with SGLang inference for scalable RL post-training.
  • Flexible rollout and data-buffer system for high-throughput data generation and evaluation.
  • Workflows: standard GRPO-style training, asynchronous training, and multi-turn agentic training with tool use.
  • Use Case: Research labs iterating RL tasks on GLM family models (GLM-4.x, Qwen3, Llama 3) with reproducible pipelines.

Quick Start

Install slime, configure your GLM model via the provided model scripts, and start a GRPO-based RL post-training workflow.

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 GLM models?

You can perform GRPO-style RL post-training by installing slime, configuring your GLM model via provided scripts, and starting the training workflow using Megatron-LM and SGLang.

Can I use slime for multi-turn agentic training with tool use?

Yes, slime supports multi-turn agentic training with tool use, alongside standard GRPO-style training and asynchronous training workflows for large language models.

Does slime RL post-training work with Qwen and Llama families?

Yes, slime RL post-training applies to GLM, Qwen, and Llama families, enabling researchers to iterate on RL tasks across different large language model architectures.

What's the best way to scale LLM inference during RL post-training?

The best way to scale LLM inference during RL post-training is using slime's Megatron-LM native training integrated with SGLang inference for high-throughput data generation.

How does the data-buffer system handle rollout orchestration?

The data-buffer system manages rollout orchestration by enabling high-throughput data generation and evaluation within the slime ecosystem for iterative model updates.

Do I need Megatron-LM to perform GRPO-based RL training with slime?

Yes, slime uses Megatron-LM native training combined with SGLang inference to execute GRPO-based RL post-training workflows for large language models.