LegoFactory

Blocks

rl

Online RL from the SFT checkpoint

rl is the online reinforcement learning stage. Starting from the actor trainer produced, it runs GRPO/GSPO against SWE-bench tasks via Harbor + vLLM + verl, on a multi-GPU node with Kubernetes-backed task execution.

Full docs: swe-rl-docs.pages.dev/docs

Long-form docs live there

The rl block ships its own complete documentation site — getting started, core concepts, running training (preflight, the vLLM + LiteLLM inference stack, sandbox backends, results, and scaling), the dashboard, a troubleshooting guide, and the full reference. This page is just an orientation; follow the link above for everything else.

At a glance

  • Inputs: model (typically wired from trainer.output.checkpoint_path), infrastructure, training, data, experiment, credentials.
  • Output: actor checkpoints under repos/harbor-verl-train/outputs/.
  • Runs: Local (8× GPU + K8s). Long-running.

How to run

/rl:setup        # build venv, sync submodules, apply verl patch
/rl:check        # preflight: GPUs, K8s reachability, vLLM env, base model
/rl:run          # launch training (long-running)
/rl:dashboard    # launch the trajectory + metric webui

Reference

On this page