Model Cards and Data Sheets

What a model card answers

Introduced by Mitchell et al. (2019), a model card is a one-page document that travels with the model:

• intended use + out-of-scope use,
• training data + caveats,
• evaluation results sliced by relevant subgroups,
• ethical considerations,
• known limitations.

A data sheet (Gebru et al., 2018) does the equivalent for the dataset.

Both should be generated by your training pipeline — not written by hand once and forgotten. MLflow can attach them as artefacts; the registry surfaces them in the UI.

A model card is a food label: ingredients, allergens, shelf life, nutrition. You wouldn't ship a packet of biscuits without the label — and a regulator (or a future you) will eventually ask the same of every shipped model.