skills.rest
afel6
afel6COMMUNITY
Explore all skills by this author →
GitHub

slime-rl-training

Guide reinforcement learning post-training for LLMs with slime, Megatron-LM, and SGLang.

Updated Mar 25, 2026
Search Tags:#llm#reinforcement-learning#slime#rl-training#sglang#post-training#megatron-lm
One-click install
npx skills add https://github.com/afel6/scal-ai-pipeline --skill slime-rl-training-afel6
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: slime-rl-training
Source: https://github.com/afel6/scal-ai-pipeline/tree/main/hermes_skills_library/optional-skills/mlops/slime
Command: npx skills add https://github.com/afel6/scal-ai-pipeline --skill slime-rl-training-afel6

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Provides guidance for performing reinforcement learning post-training for large language models using slime, Megatron+SGLang, to improve alignment and performance.

Core Features & Use Cases

  • Training (Megatron-LM): actor model training with optional critic and weight synchronization to rollout.
  • Rollout (SGLang + Router): high-throughput response generation with multi-turn support and reward/verifier outputs.
  • Data Buffers & Customization: flexible data handling, custom generate functions, and async training options for scalable RL workflows.
  • Quick Start scenarios: bootstrap end-to-end RL post-training by wiring model config, data, and evaluation.

Quick Start

Run slime's RL post-training pipeline with your chosen model, data, and reward function to start 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 perform reinforcement learning post-training for large language models using slime?▼

Reinforcement learning post-training for large language models using slime requires configuring your environment, model scripts, and reward functions to run an end-to-end pipeline integrating Megatron-LM and SGLang.

Can I use slime for multi-turn tool support during LLM rollout generation?▼

Yes, slime supports multi-turn tool use during LLM rollout generation by leveraging SGLang and a router to provide high-throughput response generation alongside reward and verifier outputs.

Does slime support asynchronous training for RL post-training workflows?▼

Yes, slime provides optional asynchronous training capabilities and flexible data buffers to enable scalable reinforcement learning post-training workflows for large language models.

What is the best way to integrate Megatron-LM and SGLang for RL post-training?▼

The best way to integrate Megatron-LM and SGLang for RL post-training is using slime, which wires actor model training and weight synchronization to high-throughput rollout generation.

Do I need a configured environment and model scripts before starting RL post-training with slime?▼

Yes, you need a fully configured environment and existing model scripts before starting, as slime provides step-by-step guidance aligning with documentation and configuration conventions rather than base setup.

How does reward modeling work within the slime RL post-training pipeline?▼

Reward modeling in the slime pipeline works by processing verifier outputs from SGLang rollouts, applying custom reward functions to evaluate and align large language model responses during training.

Related Skills

Explore More Software Engineering Skills→
markwang2658markwang2658
31

slime-rl-training

Plan and execute LLM RL post-training with the slime Megatron+SGLang framework.

Community
Advanced
jayll1303jayll1303
17

tensorrt-llm

Convert HuggingFace checkpoints into optimized TensorRT-LLM engines with quantization.

Community
Advanced
Tommy-ywTommy-yw
539

slime-rl-training

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

Community
Advanced
sheawinklersheawinkler
78

slime-rl-training

Orchestrates RL post-training for GLM models with Megatron-LM and SGLang.

Community
Intermediate