What are datasheets for datasets?
Datasheets for datasets are structured documents that travel with a dataset and answer a fixed set of questions about it. Researchers led by Timnit Gebru proposed the practice in 2018, borrowing the idea from the datasheets that accompany electronic components. A dataset datasheet states why the data was collected, what each record represents, and how the records were gathered. It also records who labeled the data, what cleaning was applied, which uses the data suits, and who maintains it.
A datasheet is the companion to a model card, which documents the model trained on the data.
Why datasheets for datasets matter
A model inherits the gaps and skews of its data, which are hard to see once the people who assembled it have moved on. A datasheet puts them in writing before anyone trains on the data. The EU AI Act’s technical documentation for high-risk AI systems includes, where relevant, datasheets describing the training data, its provenance, and how it was labeled and cleaned.
Datasheets for datasets in practice
A city public works department assembles years of 311 service requests to train a model that ranks pothole repairs. The datasheet notes that most requests came through a mobile app used unevenly across neighborhoods. It also notes that older requests lack map coordinates and that “resolved” means a crew closed the ticket, whether or not anyone checked the repair. The department added road-condition survey data to offset this reporting bias, so that repair priority would not simply follow the neighborhoods that report the most.