Troubleshooting

"An API key is required for the OpenAI API". Enter a key, set OPENAI_API_KEY, or pick a local endpoint (local servers need no key).

Rate limits or timeouts. Lower Parallel requests or raise Request timeout under Generation settings. CPU servers need 1–2 parallel requests.

Fewer pairs than the target. The documents ran out of new questions. The run details (or manifest.json → generation) show how many questions were dropped, answers filtered, pairs rejected by the judge and duplicates removed; rejected.jsonl holds the rejected pairs with the reason. Use longer documents, a lower target, a stronger teacher, or a lower minimum judge score.

"No usable question/answer pairs were produced". Every request failed or every pair was rejected. Check the model name and endpoint, then the run log (runs/<run-id>/run.log).

Structured-output errors from a local server. Brainbrew falls back from JSON schema to JSON mode to a schema in the prompt automatically. Very small models may still produce unusable JSON; use a larger one.

A run shows "interrupted". The process running it stopped (the server was restarted, or the CLI was killed). Start a new run; interrupted runs cannot be resumed.

"Waiting for the GPU". Another run is training on this machine. It continues when that run finishes; you can cancel it meanwhile.

CUDA out of memory. Use a smaller model in vllm serve, lower --gpu-memory-utilization, or train a smaller base model.

Hugging Face publishing fails. Check that the token has write access and the repo name is user/name.