slime-rl-training

Configure scalable RL post-training for LLMs using the slime framework.

Updated Jun 25, 2026
One-click install
npx skills add https://github.com/davpatel605-beep/hermusagent --skill slime-rl-training-davpatel605-beep
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/davpatel605-beep/hermusagent/tree/main/backend/vendor/hermes/optional-skills/mlops/slime
Command: npx skills add https://github.com/davpatel605-beep/hermusagent --skill slime-rl-training-davpatel605-beep

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill helps researchers and engineers overcome the complexity of scaling LLM post-training with reinforcement learning by providing guidance for configuring and operating the slime framework with Megatron-LM and SGLang.

Core Features & Use Cases

  • RL Training Workflows: Supports GRPO, PPO, and related reinforcement learning workflows for training reasoning models with scalable rollout generation.
  • Framework Integration Guidance: Covers Megatron-LM training, SGLang inference, data buffers, custom generation, reward functions, and multi-turn agentic training scenarios.
  • Use Case: Apply this Skill when building research pipelines for GLM, Qwen, DeepSeek, or Llama models that require distributed RL post-training and custom data workflows.

Quick Start

Use the slime skill to configure an RL post-training workflow for a language model with custom rollout generation and reward evaluation.

Frequently Asked Questions about slime-rl-training

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I scale reinforcement learning post-training for large language models using Megatron-LM?

Scale reinforcement learning post-training for large language models by configuring the slime framework to integrate Megatron-LM for distributed training and manage SGLang rollouts. This setup supports GRPO workflows and custom data generation for production-scale experiments.

Can I use SGLang for rollout generation in GRPO training workflows?

Yes, SGLang is supported for rollout generation in GRPO training workflows. The framework enables scalable rollout management alongside Megatron-LM distributed training to facilitate reasoning model development and multi-turn agentic scenarios.

What is the best way to configure custom reward functions for multi-turn agent training?

Configure custom reward functions for multi-turn agent training by utilizing the slime framework's reward computation capabilities. This allows tailored evaluation of agentic interactions during distributed reinforcement learning post-training.

Does this reinforcement learning framework support distributed training for models like Qwen and DeepSeek?

Yes, the framework supports distributed reinforcement learning post-training for models like Qwen, DeepSeek, GLM, and Llama. It applies Megatron-LM integration to manage distributed training configurations for these architectures.

Why do I need distributed training configuration for LLM reinforcement learning experiments?

Distributed training configuration is required for LLM reinforcement learning to handle production-scale experiments. It enables efficient rollout management, reward computation, and data buffer operations across multiple nodes during GRPO workflows.

How to troubleshoot rollout management issues during LLM post-training?

Troubleshoot rollout management issues during LLM post-training by leveraging the framework's guidance on SGLang inference and distributed configuration. This addresses common production-scale bottlenecks in data generation and reward evaluation.