Template
Datasheet template for licensed enterprise datasets
This is a twelve-section Markdown datasheet for licensed enterprise datasets. The supplier fills it in before delivery; the buyer's research, privacy, legal and compliance teams each find their answers in it. Section 9 collects the inputs a general-purpose AI model provider needs for its EU AI Act documentation. Download it, adapt it, and ask every supplier to complete it.
Twelve sections
What each section answers, and for whom
| # | Section | Answers | Main reader |
|---|---|---|---|
| 1 | Identification | Which set, which version, which licence, who to call | Everyone |
| 2 | Purpose and permitted use | What the data may and may not be used for | Legal, research |
| 3 | Composition | Types, counts, tokens, languages, period, gaps and skews | Research |
| 4 | Provenance and rights | Who holds the rights, how the data left their systems | Legal |
| 5 | Selection and processing | What was filtered, grouped and converted, and why | Research |
| 6 | De-identification | Methods, scan results, reviewed sample, residual risk | Privacy |
| 7 | Format and files | File list, record counts, checksums, schema version | Data engineering |
| 8 | Quality checks | Each check, how it was run, the result | Research |
| 9 | EU AI Act inputs | What feeds the technical file and the public summary | Compliance |
| 10 | Distribution and licence | Exclusivity, term, onward transfer, deletion | Legal, procurement |
| 11 | Maintenance | Corrections, withdrawals, planned updates | Data engineering |
| 12 | Change log | What changed between versions | Everyone |
Why licensed enterprise data needs its own datasheet
The original Datasheets for Datasets proposal and later formats such as Data Cards were written mostly with research datasets in mind. Licensed enterprise data raises questions those formats treat lightly: who holds the rights to a company's internal records, what route the records took out of its systems, how people in them were protected, and what the licence allows. Those questions decide whether a dataset can be used at all, so the template gives each its own section.
How to use it in a purchase
- Send the template with your spec, so the supplier knows from day one what it will have to document.
- Ask for a first version with the sample. Sections 3, 5 and 6 should describe the sample and the full set.
- Compare the sample with the datasheet: composition, languages, filtering and leak-scan claims.
- Reference the final datasheet in the licence, with its version number and checksums.
- Keep it with the model's training records. It is what you will reach for when a question comes up later.
The AI Act section
Providers of general-purpose AI models must keep technical documentation that describes training data and must publish a summary of training content using the Commission's template (Article 53, Commission template). Several fields in that summary can only be filled in with the supplier's help: the size range per modality, which source category the data falls under, the nature of the content, and its linguistic characteristics. Section 9 asks for exactly those, so the information arrives with the data instead of months later. The AI Act training data page explains each item.
Format notes
The template is plain Markdown with tables, so it diffs cleanly in version control and renders anywhere. Fields that tooling needs to read, such as counts, languages, file names and checksums, can be mirrored in Croissant metadata or a Hugging Face dataset card. Keep the narrative sections in Markdown: they are written for people.
FAQ
Datasheet questions
What is a datasheet for a dataset?
A structured document that describes why a dataset exists, what it contains, how it was collected and processed, and how it may be used. The idea comes from the Datasheets for Datasets paper by Gebru and colleagues.
Who should fill in the datasheet for licensed data?
The supplier, before delivery, because only the supplier knows the export route, the filtering and the de-identification. The buyer reviews it alongside the sample and asks for gaps to be closed.
How is this template different from a public dataset card?
It adds what licensed private data needs: a rights and export section, a de-identification section with evidence and residual risk, licence terms, and the inputs an AI model provider needs under the EU AI Act.
Does the datasheet replace the licence agreement?
No. The licence is the binding document. The datasheet describes the data and should be referenced in the licence, so that what was promised and what was delivered can be compared.
Which AI Act information does the template collect?
Modality, the size range used in the public summary, the source category, nature of the content including personal data, linguistic characteristics, illegal-content measures, and the Annex XI points on how data was obtained, selected and checked.
Can I convert the template to Croissant or a dataset card?
Yes. Most of sections 1, 3, 7 and 10 map onto Croissant metadata and dataset card fields. The rights, de-identification and AI Act sections usually stay in the Markdown document.
Keep reading
Next step
Need a dataset that ships with a filled-in datasheet?
Name the data type, languages, period, volume and intended use. We answer by email with what can be assembled and how it would be documented.