A model card is a short document that accompanies a model, describing what it does, how it was built and tested, and where it shouldn't be used.
What to Include
- Overview: what the model does, its inputs and outputs, version and owner.
- Intended use: the tasks and users it was designed for.
- Out-of-scope uses: uses you advise against, and why.
- Training data: sources, time period, size, licences and known gaps.
- Evaluation: test data, metrics and results — broken down by relevant groups and conditions.
- Limitations and risks: known failure modes, populations or inputs it handles poorly, ethical considerations.
- Recommendations: how to use it safely, such as thresholds or human review.
- Practical details: licence, how to report problems, change history.
Write for the Future Reader
Imagine someone deploying the model a year from now without you in the room. Be specific: "accuracy drops from 91% to 78% on low-light images" is useful; "may not work in all conditions" isn't.
Keep It Current
Update the card when the model is retrained, when new limitations are found, or when its use changes. Version it alongside the model.
Related Documents
Datasheets for datasets document data in the same spirit, and system cards describe how a model is used within a larger product, including safeguards.