slime-rl-training

Configure RL post-training workflows for large language models with slime.

Updated Apr 2, 2026
One-click install
npx skills add https://github.com/JKhyro/HERMES-AGENT --skill slime-rl-training-jkhyro
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/JKhyro/HERMES-AGENT/tree/main/optional-skills/mlops/slime
Command: npx skills add https://github.com/JKhyro/HERMES-AGENT --skill slime-rl-training-jkhyro

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill helps you plan, configure, and troubleshoot large-scale reinforcement-learning post-training for language models, replacing trial-and-error setup with a clear workflow.

Core Features & Use Cases

  • Training orchestration: Set up GRPO, PPO, and related RL workflows with Megatron-LM training and SGLang rollout generation.
  • Advanced workflows: Support asynchronous runs, multi-turn agentic training, custom reward functions, buffered data sources, and model-specific configuration scripts.
  • Operational support: Diagnose issues such as engine crashes, weight-sync failures, out-of-memory errors, and data-format mistakes during production or research training runs.

Quick Start

Ask the AI to set up a slime training plan for your model, dataset, GPU budget, and reward strategy, then have it recommend the right configuration and next steps.

Frequently Asked Questions about slime-rl-training

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

FAQPage Schema
How do I configure GRPO or PPO reinforcement learning post-training for large language models?

To configure GRPO or PPO reinforcement learning post-training, you define model scripts, prompt and label datasets, rollout buffering, custom generate and reward functions, and GPU-aware synchronization settings using slime. This orchestrates training with Megatron-LM and rollout generation with SGLang.

Can I run asynchronous and multi-turn agentic training workflows with slime?

Yes, you can run asynchronous and multi-turn agentic training workflows with slime. The Skill configures these advanced RL post-training patterns by applying custom reward functions and buffered data sources to handle complex, multi-step generation and optimization loops.

How do I fix weight-sync failures and out-of-memory errors during LLM training?

To fix weight-sync failures and out-of-memory errors during LLM training, you adjust GPU-aware synchronization settings and rollout buffering configurations in slime. The Skill provides operational support to diagnose engine crashes and data-format mistakes across Megatron-LM and SGLang deployments.

Do I need custom reward functions and rollout buffering to scale RL post-training?

Yes, custom reward functions and rollout buffering are required to scale RL post-training effectively. These components manage data sources and evaluate model outputs during generation, ensuring that GRPO and PPO workflows optimize large language models correctly under your GPU budget.

What is the best way to set up SGLang rollout generation with Megatron-LM training?

The best way to set up SGLang rollout generation with Megatron-LM training is to use slime to define a unified configuration. You provide model-specific scripts, evaluate reward strategies, and establish GPU-aware synchronization settings to coordinate the training and rollout engines seamlessly.