slime RL Training Guide
Guides an agent through reinforcement-learning post-training of LLMs with the slime Megatron+SGLang framework
- Repository
- Orchestra-Research/AI-Research-SKILLs
- Stars
- 14,195
- Licence
- MIT
- Files
- 3 files
What it does
When installed, the agent knows how to set up and run reinforcement-learning fine-tuning (such as GRPO) for large language models using slime, a framework that pairs Megatron-LM for training with SGLang for fast response generation. It gives step-by-step commands for Docker setup, model config selection, launching training runs, asynchronous training, and multi-turn agent training with custom generation functions. It is a how-to reference: it does not train anything by itself, and the user still needs their own GPUs, model checkpoints and data.
Great for
- ML engineers running large-scale RL post-training on GPU clusters
- Researchers reproducing GLM-style or Qwen3-style RL training workflows
- Teams that already use Megatron-LM and want SGLang-based rollout generation
Skip it if
- Beginners without access to multi-GPU hardware, since the examples assume 4 to 8 GPUs and Docker with GPU support (the skill cannot provide compute)
- Simple fine-tuning or inference tasks, where lighter tools are a better fit
- Running commands unreviewed on a production cluster: the examples launch long training jobs that consume significant GPU time and may write checkpoints to disk
Try asking
- “Set up a slime Docker environment and explain the install steps for GRPO training”
- “Write a JSONL data file format and launch a GRPO run on Qwen3-4B with 8 GPUs”
- “Configure an asynchronous training run with a rollout buffer size of 4”
- “Build a custom multi-turn generate function for training a tool-using agent”
Install in 3 steps
- 1
Download the skill
The ZIP contains the 3 original files, the MIT licence and an attribution note.
slime.zip - 2
Add it to your agent
Claude Code (all projects)
mkdir -p ~/.claude/skills && unzip slime.zip -d ~/.claude/skills/Claude Code / Agent SDK (this project only)
mkdir -p .claude/skills && unzip slime.zip -d .claude/skills/Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.
- 3
Ask for the task
No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.
Get better results
- Start with a small model such as the Qwen3-4B example and a small dataset to verify the pipeline before scaling up
- Review the --num-rollout, --save-interval and --eval-interval values before launching, since they determine run length and disk use
- Check the referenced troubleshooting and API reference files in the skill folder when a launch fails
What's inside
- references/api-reference.md
- references/troubleshooting.md
- SKILL.md
Redistributed unmodified from Orchestra-Research/AI-Research-SKILLs under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.