Read a model card before you build a demo on a model

Start with details and licence

The annotated template on the Hugging Face Hub begins with model details: a description, who developed the model, the licence and, where relevant, the base model it was fine-tuned from. It also links to the repository, paper or demo.

Read the licence before anything else. It decides whether your intended use, including a public demonstration, is permitted.

Compare your task with the stated uses

The uses section separates direct use, downstream use when the model is fine-tuned or plugged into a larger system, and out-of-scope use. Out-of-scope use covers foreseeable misuse and what users ought not to do with the model. If your task appears there, choose another model.

The template defines a bias as a stereotype or disproportionate performance for some subpopulations, a risk as a socially relevant issue the model might cause, and a limitation as a likely failure mode that recommendations can address. Read those sections with your audience and inputs in mind.

Look at how it was evaluated

The evaluation section should describe the testing data, the factors that influence behaviour and the metrics used, ideally broken down by factors such as task, domain or population group. A single headline number tells you little if the test data does not resemble yours.

Missing sections are information too. A card with no evaluation or no limitations is a signal to run your own small test before relying on the model in front of an audience.

• Read the licence first.
• Check whether your task is listed as out of scope.
• Note stated biases, risks and limitations.
• Compare the evaluation data with your own inputs.
• Test the model yourself where the card is thin.