Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
65 lines (49 loc) · 2.39 KB

CONTRIBUTING.md

File metadata and controls

65 lines (49 loc) · 2.39 KB
Raw

Contributing

Thanks for poking at tastebench. It's a small, opinionated tool — PRs that keep it that way are welcome.

Setup (model-free, seconds)

git clone https://github.com/publu/tastebench && cd tastebench
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest -q                 # the model-free smoke suite (synthesizes its own audio)

The core install is model-free (numpy / librosa / rich — no torch, nothing gated). You can run, test, and develop the whole craft layer, CLI, worker, and web path with no model download.

The fast path while developing

You almost never need the ~20 GB neural model to work on the tool:

tastebench compare a.wav b.wav --to demo.wav --no-brain   # instant, model-free
tastebench glossary <term>                                 # the explainer dictionary

--no-brain skips the TRIBE forward pass entirely. Touch the neural path only when your change is in it; otherwise the craft layer is the full, fast feedback loop. The optional model setup is in the README ("Set up the neural read").

Tests

  • pytest -q must pass. The suite is model-free by design — tests/ synthesize their own audio and never download or call the neural model. Keep it that way: a test that needs the ~20 GB model is not a unit test.
  • Adding a brain network, reward-proxy member, or craft metric key? tests/test_explainers.py enforces that every one has an entry in tastebench/explainers/explainers.json. Add the explainer with your metric.

Style

Match the surrounding code: same comment density, naming, and idiom as the file you're editing. The codebase favours small, honest functions and plain-language docstrings that say why, not what. The product voice is in identity.py — anything a user reads goes through it.

What not to commit

No audio/video, secrets, or model weights (see .gitignore). The repo ships code only, with the two deliberate, owned/permitted media fixtures in examples/ documented there. The shareable rendered artifact is the ASCII read in examples/README.md — keep it text, not an image.

PRs

  • One focused change per PR; explain the why in the description.
  • Tests pass, model-free tests stay model-free.
  • No new heavy/required dependency in the core install — heavy things are optional extras ([brain], [web], [modal]), exactly like the existing ones.