Provenance documentation
A provenance file organizes the evidence behind an offered dataset: where records came from, how rights reached the supplier and which uses are permitted. It supports review; it does not resolve every legal question by itself.
Original source
Collection method, date range and contributors
Rights record
Assignments, licences and relevant consent
Use boundaries
AI training scope, exclusions and sublicensing
Delivery version
Transformations, schema and acceptance sample
Cleaning or de-identifying data does not grant missing ownership or licensing rights.
The provenance file — seven folders
- 01 Origin: how and when each source was collected.
- 02 Ownership: assignments, employment/contractor IP clauses.
- 03 Third-party content: licences and their terms.
- 04 Personal data: categories, lawful basis, consent versions.
- 05 Transformations: cleaning, filtering, de-identification steps.
- 06 Datasheet: motivation, composition, known limitations.
- 07 Change log: versions and dates.
Formats to borrow from
Datasheets for Datasets and the Data Cards Playbook offer documentation question sets. Hugging Face dataset cards combine narrative documentation with structured metadata; MLCommons Croissant provides a machine-readable dataset metadata format. These aids do not establish ownership or permission by themselves.
Have a dataset worth discussing?
Tell us what you own, which rights you can license and the commercial scope you have in mind.