Model Card

A model card documents a model's intended uses, evaluation conditions, performance, and relevant limitations. It helps a reader judge whether available evidence is relevant to a proposed use. It is documentation, not a certificate of suitability.

Ask which model and which test

A fictional card reports 95% acceptable refund drafts. To interpret it, a reader needs the model version, task, rubric, and test population. An aggregate score can hide failures on policy exceptions or unsupported languages.

The Model Cards paper proposes reporting performance under relevant conditions, including subgroup analysis where appropriate.

The application needs additional evidence

The same model can be connected to different documents, permissions, and interfaces. A card alone cannot show that a shop's retrieval or payment authorization works. Pair model documentation with application tests and a scoped decision record. Keep the card's evaluated version distinct from a newer version that has not been assessed.

Reference: Model Cards for Model Reporting.