data cards and dataset documentation
A data card is the structured label on the side of a dataset: a standardized document that records where the data came from, how it was collected and labeled, what it contains, and what it should and should not be used for. The motivating analogy is the datasheet that ships with an electronic component — you would not solder a chip into a circuit without its spec sheet, yet datasets were long shipped with none, leaving downstream users to rediscover their quirks and harms the hard way.
The genre was formalized by Datasheets for Datasets (Gebru et al.) and Data Cards (Google), and now appears as Hugging Face dataset cards and croissant metadata. A complete card answers a fixed checklist: motivation and funding, composition (instances, fields, splits, missing data), collection process and consent, preprocessing and cleaning, recommended and discouraged uses, distribution and licensing, and maintenance. Crucially it surfaces sensitive attributes, known biases, and population coverage so a model builder can judge fitness for their context before training.
Documentation is governance infrastructure, not a metric: it does not fix a flawed dataset, but it makes the flaws legible, auditable, and citable, and it is increasingly required for regulatory compliance and reproducibility.
A data card documents flaws; it does not remove them — read it as a fitness-for-use disclosure, not a quality guarantee.