Datasheet template

Every dataset ships with a datasheet: what it is, how it was made, and what it shouldn't be used for. The sections are fixed, so datasheets compare across datasets.

• 1 Summarytitle, ID, vertical, description
• 2 Compositionsize, years, language or modality split, domain split
• 3 Collectionowner type, how it was recorded or acquired
• 4 Data specformats, rates or DICOM details, fields
• 5 De-identificationmethod, where it ran, report reference
• 6 Qualitychecks, score definition, exclusions
• 7 Provenancerights basis, chain of title
• 8 Licenceexclusivity, term, territory, permitted use
• 9 Known limitsgaps, biases, unsuitable uses
• 10 Changesversions and re-deliveries

The sections

1. Summary. Title, listing ID, vertical, a one-paragraph description, status.
2. Composition. Size in hours or studies; years covered; the split by language or modality, by domain, by call type or body part.
3. Collection. Who held the data (by type, such as "Tier-1 healthcare BPO"), how it was recorded or acquired, and for what original purpose.
4. Data spec. Formats, sample rates or DICOM details, transcripts or reports, and every metadata field. Links to the spec.
5. De-identification. Method and version, where it ran, entities removed, and a reference to the report.
6. Quality. The checks run, the score definition and threshold, and what was excluded and why.
7. Provenance. Rights basis, and how chain of title is recorded in the manifest.
8. Licence. The options the owner offers.
9. Known limits. Gaps in coverage, likely biases (regions, demographics, devices), and uses the data isn't suited to.
10. Changes. Version history and re-deliveries.

The owner is named in the datasheet only after an NDA.