Contributing

Contributions that improve reproducibility, portability, evaluation, calibration, or documentation are welcome.

Please keep changes auditable:

Explain the experimental claim being changed.
Include the exact model/data/code version used.
Add or update tests when changing loading or scoring behavior.
Do not replace measured results with estimates.
Keep public benchmark rows out of training and checkpoint selection.
Report regressions and failed experiments instead of deleting them from the record.

For model-quality changes, include row-level or task-level evidence sufficient to compare against the frozen public 1.0 baseline.