Thanks for improving DEAF. Changes should preserve the benchmark definition, portable data paths, and explicit failure semantics.
python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
pre-commit installBefore opening a pull request, run:
ruff check .
ruff format --check .
python -m pytest
python examples/offline_demo.py
python scripts/check_public_repo.pyTests must not access paid APIs. Mock the adapter boundary and keep live smoke tests manual. New parsing logic must cover HF and legacy layouts where applicable, retain relative paths, and fail on ambiguous or missing data.
- Keep changes focused and document user-visible behavior in
CHANGELOG.md. - Add tests before or alongside a bug fix or feature.
- Do not commit generated audio, model weights, full raw responses, credentials, personal paths, or provider invoices.
- Do not change task definitions or metric formulas without a corresponding research-methodology discussion and documentation update.
- Report the Python versions and commands used for verification.
By contributing code, you agree that it is distributed under the repository's MIT license. Dataset contributions require separate licensing review and belong in the canonical Hugging Face dataset repository.