slime-rl-training

Coordinate distributed reinforcement-learning post-training for large language models.

Updated May 4, 2026
One-click install
npx skills add https://github.com/Supporter09/Face_Anti_Spoofing_Biometric --skill slime-rl-training-supporter09
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/Supporter09/Face_Anti_Spoofing_Biometric/tree/main/.claude/skills/slime
Command: npx skills add https://github.com/Supporter09/Face_Anti_Spoofing_Biometric --skill slime-rl-training-supporter09

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill streamlines large-scale reinforcement-learning post-training for language models, reducing the friction of coordinating rollout generation, training, and evaluation across complex distributed setups.

Core Features & Use Cases

It supports Megatron-LM training with SGLang rollout generation, flexible data buffering, custom reward functions, and multi-turn agentic workflows. Use it when you need to train or fine-tune GLM, Qwen, DeepSeek, or Llama models with structured JSONL data, async execution, or research-grade evaluation loops. It is especially useful for building reproducible RL pipelines that integrate model scripts, dataset preparation, and troubleshooting guidance in one workflow.

Quick Start

Ask the assistant to guide you through setting up a slime RL training run for your model, dataset, and target 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 large language models using Megatron-LM and SGLang?

Reinforcement learning post-training for large language models coordinates distributed rollout generation and training. This workflow uses SGLang for rollout generation and Megatron-LM for training, applying to GLM, Qwen, DeepSeek, and Llama models.

Can I use custom reward functions and multi-turn agentic workflows during RL post-training?

Yes, RL post-training supports custom reward functions and multi-turn agentic generation workflows. It provides flexible data buffering and asynchronous execution to handle complex agentic loops and structured prompt interactions.

How do I prepare JSONL prompt and label data for distributed RL training?

Distributed RL training requires structured JSONL prompt and label data for execution. You must format your dataset into JSONL files containing prompts and labels, then configure the data buffers to feed the training pipeline.

Does RL post-training work with Qwen, DeepSeek, and Llama models or only GLM?

RL post-training works with Qwen, DeepSeek, and Llama models in addition to GLM. It applies to various large language model workflows, requiring compatible model scripts, reward scripts, and evaluation scripts configured for your specific architecture.

What's the best way to set up a reproducible RL pipeline with research benchmarking and evaluation loops?

Building reproducible RL pipelines integrates dataset preparation, model scripts, and evaluation loops into one workflow. It supports research-grade benchmarking by coordinating asynchronous execution and custom evaluation scripts throughout the training cycle.

Why do I need SGLang-compatible configuration and Megatron-LM for distributed rollout generation?

SGLang-compatible configuration enables efficient distributed rollout generation, while Megatron-LM handles the distributed training overhead. Together they reduce the friction of coordinating rollout generation, training, and evaluation across complex setups.