Data cards, defined by their texts
Data card: what one holds, from the texts that define it
Updated
6 standards and legal templates on our board say what a card should hold, read from 11 texts. Each puts it in its own terms.
Email me this result
We will send you these figures and a link back to them. One email, nothing else unless you tick the box.
How this calculator works
- California AB 2013's items are quoted from the bill text: item (1) the sources or owners of the datasets, item (7) whether they include personal information. (California AB 2013 bill text, read for the data card generator)
- The EU line follows Article 53(1)(d) as the Commission's AI Act Service Desk gives it: a sufficiently detailed summary of the content used for training, made publicly available according to a template provided by the AI Office. What the template is for is quoted from the Commission's explanatory notice. (European Commission AI Act Service Desk: Article 53, read for the data card generator)
- Format lines follow each format's own documentation: the Hugging Face Hub docs, the Croissant specification, the Datasheets paper and the Data Cards paper. (Hugging Face Hub docs: dataset cards, read for the data card generator)
- The outline lists sections; it does not say whether a finished card meets any law.
The Data Cards paper
Google Research's paper names the aim of a card as "a clear and thorough understanding of a dataset's origins, development, intent, ethical considerations and evolution".
Source: Data Cards paper (arXiv 2204.01075), read for the data card generator
The Data Cards Playbook
The Playbook turns that into questions a card answers, among them the "data fields that were collected and are included in the dataset" and whether sensitive or personally identifiable information was collected while the dataset was made.
Source: Google's Data Cards Playbook, read for the data card generator
A card a machine can read
Croissant carries the same idea as metadata. Its working group calls it "an open, community-built, standardized metadata vocabulary for ML datasets" that "provides machine-readable descriptions of dataset provenance and permissions".
Source: MLCommons Croissant working group page, read for the data card generator