Model cards

A model card is a short document that comes with a trained model. Margaret Mitchell and colleagues proposed it in 2019, as a companion to datasheets.

What goes on a model card

• Model details: who made it, when, what type of model, the licence.
• Intended use: what it is for, and what it is not for.
• Factors: groups or conditions where performance might differ (age, language, lighting, device).
• Metrics: how it was evaluated, and why those metrics.
• Evaluation data: what it was tested on.
• Results, broken down: performance for each relevant group, not only overall.
• Ethical considerations and caveats.

Why "broken down" matters

A model with 95% accuracy overall can be 99% accurate for one group and 75% for another. The average hides it. Reporting results by group is the heart of a model card.

Where you'll see them

Hugging Face shows a model card (the README) on every model page. Many companies publish cards or "system cards" for their large models.

For your project

If you train a model, even for coursework, write a half-page card: what it's for, what it was tested on, results by group if relevant, and known limits.

RememberShip every model with a card: intended use, how it was tested, results by group, and its limits.