skills.rest
t2ance
t2anceCOMMUNITY
Explore all skills by this author →
GitHub

slime-rl-training

Guide RL post-training of GLM-scale LLMs with Megatron-LM and SGLang workflows.

Updated Aug 23, 2026
Search Tags:#llm#reinforcement-learning#slime#rl#sglang#post-training#megatron-lm
One-click install
npx skills add https://github.com/t2ance/dr-claw-plugin --skill slime-rl-training-t2ance
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/t2ance/dr-claw-plugin/tree/main/plugins/ml-training-stack/skills/post-training/slime
Command: npx skills add https://github.com/t2ance/dr-claw-plugin --skill slime-rl-training-t2ance

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

slime provides a guided framework to perform RL-based post-training for large language models using Megatron-LM and SGLang, enabling researchers to structure data pipelines and training workflows with reproducible configuration.

Core Features & Use Cases

  • End-to-end RL post-training workflow for GLM-family models.
  • Data generation, rollout routing, and multi-turn tool integration with SGLang.
  • Use case: researchers can set up a reproducible post-training loop to optimize reward models for RLHF-like objectives.

Quick Start

Run slime to kick off an RL post-training workflow for your GLM model using Megatron-LM and SGLang with your 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 perform RL post-training for LLMs using Megatron-LM and SGLang?▼

RL post-training for LLMs is structured by configuring YAML frontmatter, dependencies, and optional scripts to build reproducible data pipelines and training workflows using Megatron-LM and SGLang.

What is the best way to set up a reproducible RL post-training loop for GLM models?▼

A reproducible RL post-training loop for GLM-family models is established by using a YAML frontmatter configuration to manage custom data generation, rollout routing, and multi-turn tool integration. This enables researchers to optimize reward models for RLHF-like objectives.

Does this RL training framework support multi-turn tool integration with SGLang?▼

Yes, SGLang supports multi-turn tool integration within the RL post-training workflow. The framework applies across GLM-scale models to manage rollouts and route data generation effectively during training.

How do I manage rollout routing and custom data generation for LLM reinforcement learning?▼

Rollout routing and custom data generation are managed through the framework's structured YAML configuration and optional scripts. This setup allows you to define dependencies and orchestrate training pipelines across Megatron-LM and SGLang workflows.

Can I use this to optimize reward models for RLHF-like objectives?▼

Yes, you can optimize reward models for RLHF-like objectives by running the guided end-to-end RL post-training workflow. It structures your data pipelines to align with GLM-family model training requirements.

What are the limitations when configuring RL training pipelines with YAML frontmatter?▼

The YAML frontmatter configures metadata, scripts, references, and dependencies for training pipelines. Limitations depend on your specific Megatron-LM and SGLang environment setup and the scale of your GLM-family model integration.

Related Skills

Explore More Education & Research Skills→
Tommy-ywTommy-yw
539

slime-rl-training

Optimize large language models with Megatron-LM training and SGLang rollouts.

Community
Advanced
BBufBBuf
721

sglang-glm-vlm-ocr-optimization

Aggregate PR-backed evidence and diff history to guide GLM VLM/OCR optimization.

Community
Advanced
zangjeicyzangjeicy

slime-rl-training

Train LLMs with GRPO and PPO using Megatron-LM and SGLang.

Community
Advanced
overviewlabsoverviewlabs

slime-rl-training

Guides RL post-training of large language models using Megatron-LM and SGLang integration.

Community
Advanced