workspace-adapter-verl

Use the autoresearch run verl-case command after readiness checks pass. The adapter stages the model and dataset, syncs dependency sources, writes an immutable config lock, and executes the matrix rows inside an Ascend Docker container on the target server.

Val-only mode (trainer_val_only=true) runs the validation matrix without updating model parameters, while real GRPO training (trainer_val_only=false) must complete at least 3 training steps per case. Results with fewer than 3 steps are recorded only as failed data points.

First confirm the W&B project is named verl and the run display name follows the model-algorithm-sequence-timestamp convention. Then check 1-wandb/source-runs.json and the rebuild scripts to reconstruct the historical web view.

Strict accuracy requires the model output to contain the answer in a \boxed{} expression as the reward function expects. A numerically correct answer without the boxed format still scores zero, and val-only runs never update model weights.

Pushgateway only stores the latest gauge value, so pushing metrics once after the run produces a flat line in Grafana. The runner must continuously push autoresearch_npu_* metrics during training and keep npu-smi watch logs for 0.5s curve replay.

Yes, the adapter discovers a running container from the same image, verifies it is idle, and runs a torch_npu smoke test before executing rows via docker exec. If checks fail, it falls back to launching a fresh container per row.