Skip to content
Get packs

miles RL Training

Guides large-scale reinforcement learning post-training of big MoE models on GPU clusters

Source Download skill
Repository
Orchestra-Research/AI-Research-SKILLs
Stars
14,195
Licence
MIT
Files
3 files

What it does

When installed, the agent gets a detailed playbook for running reinforcement learning (RL) training on very large AI models using miles, a framework forked from the slime project. It explains when to choose miles over alternatives, how to set up FP8 or INT4 low-precision training, how to keep training and inference numerically aligned for mixture-of-experts (MoE) models, and how to enable speculative decoding with EAGLE. The agent can then write launch commands and configuration flags for these workflows, but the skill itself does not run training.

Great for

  • ML engineers post-training very large MoE models such as DeepSeek V3 or Qwen3-MoE on H100/H200 clusters
  • Teams already using SGLang, Megatron-LM and Ray who want a guided starting point for miles configuration
  • Researchers comparing RL frameworks who want a summary of miles versus slime, verl and torchforge

Skip it if

  • Beginners or anyone without access to multi-GPU data-center hardware, since the workflows assume H100/H200-class GPUs and large clusters
  • Small models or laptop-scale experiments, where the setup and tooling are far heavier than needed
  • Production decisions based on the skill's claims alone, since performance figures such as 25-40% speedups and 'exactly 0 KL divergence' are vendor-stated and not independently verified in these files

Try asking

  • “Help me set up a miles Docker environment and launch GRPO training for Qwen3-30B-A3B on 8 GPUs”
  • “Write a config to enable EAGLE speculative decoding during RL rollouts in miles”
  • “Explain how Rollout Routing Replay (R3) keeps expert routing consistent between training and inference”
  • “Plan the tensor and expert parallelism settings for training a DeepSeek V3 style MoE model”
Step-by-step guide: how to use miles RL TrainingInstall, first run, examples, tips and FAQs

Install in 3 steps

  1. 1

    Download the skill

    The ZIP contains the 3 original files, the MIT licence and an attribution note.

    miles.zip
  2. 2

    Add it to your agent

    Claude Code (all projects)

    mkdir -p ~/.claude/skills && unzip miles.zip -d ~/.claude/skills/

    Claude Code / Agent SDK (this project only)

    mkdir -p .claude/skills && unzip miles.zip -d .claude/skills/

    Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.

  3. 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

  • Verify CLI flags against the miles repository before running, because the SKILL.md itself says some features' specific flags may vary
  • Start with a small model and a single node to confirm the pipeline works before scaling to 1TB+ MoE models
  • Treat the SKILL.md commands as templates: the agent will generate code and shell commands that you must review and run yourself, including Docker pulls from radixark/miles:latest which is an unpinned tag

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.