Software & AI
Datasheets for Datasets
In one line
Structured documentation describing a dataset's motivation, composition, collection process, preprocessing, uses, and maintenance, the dataset equivalent of a model card.
Definition
Datasheets for Datasets is a documentation framework, introduced by Gebru et al. (2018), that prescribes a structured set of questions every dataset should answer: motivation for creation, composition (instances, labels, splits), collection process (who, how, when, consent), preprocessing/cleaning/labeling, uses (recommended, discouraged), distribution, and maintenance. The intent is to surface biases, gaps, and limitations of a dataset before it is used to train or evaluate a model, and to make those properties auditable downstream by model consumers and regulators.
Why it matters
Datasheets for Datasets is a voluntary documentation approach. IMDRF N88 FINAL:2025 supports representative datasets, independent training and test data, and lifecycle controls, but does not make a specific datasheet template mandatory. Document relevant training, validation, and test data in a form appropriate to the device and submission. This review did not establish that datasheets are the most widely adopted format, widely used in PCCP submissions, or specifically required as separate documents by FDA.
Common pitfalls
- •Documenting only the training set, validation and test set datasheets are equally important for evaluating generalization claims.
- •Glossing over consent and source legitimacy, these are increasingly material under GDPR, HIPAA, and AI Act scrutiny.
- •Treating datasheets as static, the document must be updated when datasets are augmented or relabeled.
Sources