What is a model card?

Before you build on someone else's model, you need answers to a few plain questions. What was it made for? What was it trained on? How well does it do, and for whom? What licence is it under? A model card is where the maker is supposed to answer them.

Where the idea came from

In 2018 a team at Google, with Margaret Mitchell as first author, proposed model cards as a short, standard document to go with every trained model. Their main concern was that a model's accuracy figure hides who it works badly for. A face-analysis model can score well overall and poorly for some groups of people. So the card was meant to report results broken down by the conditions that matter, and to say plainly what the model should not be used for.

What a good one contains

• What the model is: its architecture, size, who made it and when.
• Intended uses, and uses the maker advises against.
• Training data: what it was trained on, or at least where to find a dataset cardDataset CardDocumentation describing a dataset’s composition, collection process, and recommended uses. that says so.
• Evaluation: which benchmarks, with which settings, and results broken down where it matters.
• Limitations and risks: known biases, failure modes, languages it handles poorly.
• Licence and how to cite it.

Where you will meet them

On Hugging Face every model's page is its card. It is the README.md file in the model's repository, with a block of metadata at the top that lists the licence, languages, datasets and task. The Hub uses that metadata for its search filters, so a model with an empty card is also hard to find.

The big AI labs publish a heavier version, usually called a system card, with each major model. Those add the safety testing: red-teamRed TeamingAttacking your own AI system on purpose, the way a real attacker or a careless user would, to find where it fails before they do. results, dangerous-capability evaluations and the safeguards put in place.

How to read one

Treat it as a starting point, not a guarantee. Benchmarks on a card are the maker's own and chosen by them. Check whether the evaluation looks like your task, and whether your language or domain is mentioned. Look hard at the licence, since some open-weight models restrict commercial use or large-scale deployment. And if a card is nearly empty, that tells you something too: you will be testing everything yourself.