Contributing to deup

June 5, 2026 · View on GitHub

Thank you for helping improve the canonical open-source implementation of DEUP (Lahlou et al., 2023).

Quick start

git clone https://github.com/ursinasanderink/deup.git
cd deup
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev,gbm,finance,docs]"
pre-commit install   # optional but recommended
pytest
ruff check . && ruff format --check .
mypy
mkdocs build --strict

What we welcome

  • Bug fixes with a regression test
  • Documentation improvements (tutorials, API clarity, benchmarks)
  • New optional extras (e.g. additional GBM or domain presets) that follow existing patterns
  • Benchmarks that reproduce on CPU in CI (or are marked @pytest.mark.integration)

Design principles

  1. Leakage-correct OOF errors — never train g on in-sample residuals
  2. Sklearn-compatible estimators — fit / predict / get_params
  3. Thin domain presets — wire axes; don't fork core logic
  4. Honest uncertainty — document when aggregation guards apply

See ARCHITECTURE.md for the five-axis mental model.

Pull request checklist

  • Tests pass locally (pytest)
  • Lint + types pass (ruff, mypy)
  • Docs build (mkdocs build --strict) if you changed docs or public API
  • CHANGELOG.md updated under [Unreleased] for user-visible changes
  • New public symbols have NumPy-style docstrings

Reporting issues

Use GitHub Issues. Include:

  • Python version, OS, pip show deup
  • Minimal reproducible example
  • Expected vs actual behavior

Code of conduct

This project follows the Contributor Covenant.

Attribution

DEUP the method is due to Lahlou et al. (2023). You are contributing to the library — please do not claim authorship of the method in docs or marketing copy.

Release process

Maintainers: see RELEASING.md.