Model Card
A short document describing a model’s intended use, training data, performance, and limitations — its nutrition label.
A model card is a short, structured document that accompanies a released model, describing its intended uses, training data, evaluation results (ideally broken down by subgroup), known limitations, and ethical considerations. It is the “nutrition label” that makes a model’s capabilities and risks legible to users.
Worked example: a face-detection model’s card reports accuracy per skin tone and lighting condition, states it is not for surveillance, and lists failure modes — so a downstream team can judge whether it fits their use before deploying. Gotcha: a model card is only useful if honest and specific — vague cards that omit subgroup performance or limitations provide false assurance — and it is documentation, not a control: it informs decisions but does not enforce appropriate use, so pair it with real access and usage governance.