Spaces:
Running
title: One Good Commit
emoji: 🃏
colorFrom: yellow
colorTo: green
sdk: static
app_file: index.html
pinned: false
license: mit
short_description: Evidence-backed model card contribution assistant
One Good Commit
One Good Commit reads a public Hugging Face model card and prepares a bounded documentation contribution packet for a human to review.
It answers three narrow questions:
- What relevant documentation can be located in the public model card?
- Which selected documentation signals were not found by the deterministic review?
- What questions could a prospective contributor ask without inventing facts?
The tool does not score compliance, certify safety, evaluate model quality, run the model, or post a Discussion or pull request.
Why this exists
Model cards help people discover, reproduce, understand, and responsibly use models. Maintainers also have uneven time, information, and documentation needs. This project is a small contribution aid: it keeps evidence attached to each observation and leaves every consequential decision with a person.
Method
The app retrieves two public sources:
- the model repository's Hub API record;
- the selected revision of its
README.md.
It then runs deterministic checks for documentation signals related to:
- license;
- language coverage;
- task and model type;
- base model or provenance;
- intended uses;
- out-of-scope uses;
- limitations, risks, or biases;
- training data;
- evaluation evidence;
- training and reproducibility details;
- citations or supporting papers.
A result can be:
- Documented: relevant metadata or card text was located.
- Review: related text was located, but a structured or clearer source may still need human review.
- Not found: the review did not locate relevant information.
“Not found” is deliberately limited language. It does not mean the information does not exist, the section is mandatory, or the maintainer has done something wrong.
Deployment
The public Space is intentionally static so it can run without a paid server,
GPU, account system, or secret. Its browser-side code makes read-only requests
to the selected public model's Hub API record and README.md; model-card text
is treated as untrusted and rendered as text, not HTML.
The earlier Python/Gradio prototype and fixture tests remain in this repository
as a local reference implementation. The deployed Space is index.html,
static-app.js, and static-styles.css.
Run the local reference implementation
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python app.py
The analyzer and its tests use only the Python standard library:
python -m unittest discover -s tests -v
Contribution boundary
Before contacting a maintainer:
- Read the complete card and linked primary sources.
- Confirm that the suggested section applies to that model.
- Remove speculative placeholders.
- Ask questions where only maintainers can know the answer.
- Prefer one small, accurate improvement over a broad rewrite.
No output from this tool should be posted automatically.
Privacy and access
The current version accepts public, ungated model repositories only. It does not accept uploaded documents or store model card content in an application database.
License
MIT