slime-rl-training

Guide RL post-training setup for LLMs with slime, Megatron-LM, and SGLang.

1|Updated Jan 31, 2026
One-click install
npx skills add https://github.com/Monjyu1101/AiDiy2026 --skill slime-rl-training-monjyu1101
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/Monjyu1101/AiDiy2026/tree/main/backend_hermes/optional-skills/mlops/slime
Command: npx skills add https://github.com/Monjyu1101/AiDiy2026 --skill slime-rl-training-monjyu1101

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

This Skill provides guidance to implement and scale reinforcement-learning post-training for large language models using slime, enabling efficient Megatron-LM with SGLang integration for RL rollout and data-buffer workflows.

Core Features & Use Cases

  • Guided RL post-training setup for GLM, Qwen, DeepSeek, and Llama models with Megatron-LM and SGLang.
  • Configurable data buffering and rollout to accelerate experimentation and optimization.
  • Use Case: A research team can bootstrap RLHF-style post-training on a custom dataset and track rewards and losses across iterations.

Quick Start

Configure and run a slime RL post-training workflow with your model and dataset.

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 slime?

RL post-training with slime is set up by configuring YAML frontmatter for model and dataset specifications, then executing a guided workflow that integrates Megatron-LM and SGLang for rollout and data buffering across GLM, Qwen, DeepSeek, and Llama families.

Does slime support reinforcement learning rollout with Megatron-LM and SGLang?

Yes, slime supports reinforcement learning rollout by integrating Megatron-LM with SGLang, enabling efficient model serving and data-buffer workflows to accelerate experimentation and optimization during post-training.

Can I use slime for RLHF-style post-training on custom datasets?

Yes, you can use slime for RLHF-style post-training on custom datasets, allowing research teams to bootstrap experiments, configure data buffering, and track rewards and losses across iterations for GLM, Qwen, DeepSeek, and Llama models.

What dependencies are required to run slime RL post-training workflows?

Running slime RL post-training workflows requires sglang-router, Ray, Torch, and Transformers, providing the distributed computing and model manipulation infrastructure needed for large language model reinforcement learning.

Why configure YAML frontmatter when starting slime reinforcement learning workflows?

Configuring YAML frontmatter is required because the slime workflow enforces specific name and description fields, ensuring the reinforcement learning post-training environment is correctly initialized before execution begins.