Lego-RL

Results & Artifacts

Where logs, checkpoints, and trajectories land

A run produces several streams of output, all under the repo root (the runtime artifacts are gitignored).

Per-run outputs

OutputPathNotes
Training loglogs/<exp>.logverl + Harbor integration log (contains the step: metric lines)
Throughput loglogs/<exp>_vllm.logfiltered throughput-only log (dashboard skips it)
Trajectoriesharbor_trials/<project>/<exp>/step_*/<session>/proxy_trajectory.jsonper-trial agent traffic (token ids / masks / logprobs)
Checkpointscheckpoints/<project>/<exp>/global_step_N/actor FSDP shards, every save_freq steps

Live state

pgrep -af 'train.sh|fully_async|main_ppo'          # is a run live?
tail -F logs/<exp>.log                             # follow the step lines
ls checkpoints/<project>/<exp>/                    # global_step_N/ ...
ls harbor_trials/<project>/<exp>/step_*/           # per-trial trajectories

vLLM readiness: query the Ray-registered vllm_server_* actors (see Inference Stack).

Do not commit checkpoints

FSDP shards are large and stay where verl wrote them. Archive metadata and small trajectory samples, never the checkpoints themselves.

Visualizing

The dashboard reads the training logs and renders reward, KL, MFU, entropy, and response-length curves, plus a browser over the per-trial trajectory JSON.

On this page