skills.rest
CUexter
CUexterCOMMUNITY
Explore all skills by this author →
GitHub

slime-rl-training

Run RL-based LLM post-training with Megatron-LM and SGLang for scalable rollout and GRPO workflows.

2|Updated Mar 26, 2026
Search Tags:#llm#slime#rl-training#sglang#post-training#megatron-lm#training-framework
One-click install
npx skills add https://github.com/CUexter/hermes-agent --skill slime-rl-training-cuexter
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/CUexter/hermes-agent/tree/main/skills/mlops/training/slime
Command: npx skills add https://github.com/CUexter/hermes-agent --skill slime-rl-training-cuexter

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

slime provides a structured framework to guide RL-based post-training of LLMs using Megatron-LM and SGLang, consolidating setup, execution, and evaluation into a cohesive workflow.

Core Features & Use Cases

  • Integrated RL post-training workflow for GLM-family models using slime.
  • Support for standard GRPO and asynchronous training workflows with rollout coordination.
  • Multi-turn agentic training and tool-use scenarios with configurable data pipelines and buffers.

Quick Start

Configure your model with prebuilt model scripts in scripts/models and run the slime training command with the desired arguments to start post-training.

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 LLMs using Megatron-LM and SGLang?▼

RL post-training for LLMs is set up by configuring Megatron-LM and SGLang with a prepared dataset and prebuilt model scripts. You execute the slime training command with desired arguments to start the rollout and training loop.

What is the best way to run multi-turn agentic training for Qwen or Llama3 models?▼

Multi-turn agentic training for Qwen or Llama3 models is best handled by configuring data pipelines and optional data buffers within the slime framework. This integrates rollout coordination directly with Megatron-LM training.

Can I use GRPO training with GLM-4.x family models?▼

Yes, GRPO training is supported for GLM-4.x family models. The slime framework provides an integrated workflow that coordinates SGLang rollouts with Megatron-LM to execute standard GRPO and asynchronous training.

Does asynchronous training require a separate data buffer configuration?▼

Asynchronous training can utilize an optional data buffer within the slime framework. Configuring this buffer helps manage the configurable data pipelines needed for multi-turn agentic and tool-use scenarios.

What limitations exist when scaling RL training with slime?▼

Scaling RL training with slime is limited to models compatible with Megatron-LM and SGLang, specifically requiring GLM-4.x, Qwen, DeepSeek, or Llama3 architectures. You must supply model scripts under scripts/models.

Related Skills

Explore More Software Engineering Skills→
ovachieverovachiever
52

training-llms-megatron

Train large language models with Megatron-Core using configurable 3D and 4D parallelism.

Community
Advanced
Jonah-1Jonah-1

eval

Run inference on neural network checkpoints and generate PSNR, SSIM, and LPIPS metrics.

Community
Intermediate
BermudaLocalsBermudaLocals
1

slime-rl-training

Automate RL-based post-training workflows for large language models with Megatron-LM and SGLang.

Community
Advanced
1thirteeng31thirteeng3

slime-rl-training

Guide RL-based post-training of GLM models using slime with Megatron-LM integration.

Community
Advanced