Generate benchmark tasks from validated environments through an Adapter. The built-in
FeatureBench Adapter writes accepted tasks to tasks.jsonl.
Generate tasks from every environment available at startup:
fb generate outputs/build-YYYYMMDD-HHMMSS/environments.jsonl \
--adapter featurebenchProcess one environment:
fb generate outputs/build-YYYYMMDD-HHMMSS/environments.jsonl \
--environment-id ENVIRONMENT_ID \
--adapter featurebenchChoose the output directory or an explicit Adapter profile when needed:
fb generate outputs/build-YYYYMMDD-HHMMSS/environments.jsonl \
--adapter featurebench \
--config configs/default/adapters/featurebench.yaml \
--output outputs/my-generation--limit selects the first N environments in source-file order. It cannot be used
with --environment-id.
Continue with the current Adapter configuration:
fb --resume outputs/generate-YYYYMMDD-HHMMSSRetry all work without a durable success:
fb --resume outputs/generate-YYYYMMDD-HHMMSS --select non-successSuccessful tasks are never regenerated. See Resume for more selection options.
The default FeatureBench profile is
configs/default/adapters/featurebench.yaml.
It controls:
- task quality thresholds, including Gold Patch, F2P, and P2P limits;
- Top selection and Mask repair;
- model, concurrency, and timeout settings;
- maximum tasks per environment;
- task prompt metadata and audit retention.
The YAML file is fully commented and is the configuration reference.
Each task passes a final Gold-closure validation before it is written. See Data organization for the output fields.