Model Cards (Hugging Face Hub documentation)

What it is

A model card is a short document that travels with a machine learning model and tells people what it is for, how it was trained and evaluated, and where it should not be used. The idea comes from the 2018 research paper "Model Cards for Model Reporting" by Mitchell and colleagues. On the Hugging Face Hub, every model repository's README file is its model card, and this documentation explains how to write one well.

What a good model card contains

• What the model is and who built it.
• Intended uses and out-of-scope uses.
• Limitations, risks, biases and ethical considerations.
• Training data, training settings and how to reproduce them.
• Evaluation results, ideally broken down by relevant groups, with the source of each result.
• Licence and the base model it was fine-tuned from, if any.

The Hub adds structured metadata at the top of the file (licence, datasets, base model, evaluation results) so models can be searched and filtered. There is a template and an annotated model card that explains each section.

Why it matters for governance

Documentation is a control. The EU AI Act requires technical documentation for high-risk systems and for general-purpose AI models, and ISO/IEC 42001 expects information about AI systems to be available to the people who rely on them. A model card is the lightweight, widely understood version of that documentation. Writing one forces the uncomfortable questions: where does this fail, and for whom?

How to use it

Read three model cards for popular models on the Hub and note what each leaves out. Then write a card for a model you trained or used yourself, following the annotated template, and include at least one honest limitation backed by a test. For the policy path, use the same structure as a checklist when you review documentation from a vendor.

Free or paid?

Free.