A datasheet answers seven groups of questions: motivation, composition, collection process, preprocessing, uses, distribution and maintenance. A Data Card summarizes 31 themes of a machine learning dataset.
Two further layers exist beside the research formats: the dataset cards and metadata that hosting platforms render, and the documents the EU AI Act requires model providers to keep or publish.
This page compares them from the side of a company whose data is documented, sold or licensed, as part of the wider topic of documentation in AI training data licensing.
What dataset documentation records, who writes it and who reads it
Eight documentation formats describe training data, and only the two EU AI Act formats among them are legal duties; the rest are voluntary research practice, platform conventions, industry standards or contract documents.
The table compares them by issuer, writer, reader and legal status, as of 2026-11-16.
| Format | Issuer and date | Written by | Read by | What it records about training data | Legal status | Source |
|---|---|---|---|---|---|---|
| Datasheets for Datasets | Gebru et al., arXiv 1803.09010, 2018-03-23; CACM December 2021 | dataset creators | dataset consumers | seven question groups, including license or terms of use, fees, consent | voluntary research practice | arxiv.org |
| Data Cards | Google Research (Pushkarna, Zaldivar, Kjartansson), arXiv 2204.01075, 2022-04-03, FAccT 2022 | producers | agents (reviewers, subject-matter experts) | 31 themes including Dataset Provenance, upstream sources, licenses | voluntary practice | arxiv.org |
| Croissant | MLCommons documentation site, version 1.0, 2024-03-01 | dataset publishers | tools and users loading the data | machine-readable metadata on schema.org; Responsible AI vocabulary | voluntary format | docs.mlcommons.org |
| Data Provenance Standards | Data & Trust Alliance, v1.0.0, 2024-07-09 | data suppliers | data users | 22 fields: Source 6, Provenance 8, Use 8 | voluntary industry standard | dtaalliance.org |
| Hugging Face dataset card | Hugging Face, Inc., Hub documentation (read 2026-11-16) | repository owner | anyone on the Hub | README.md with YAML metadata: license, language, size, tags | platform convention | huggingface.co/docs |
| Technical documentation, Annex XI and XII | Regulation (EU) 2024/1689, Art. 53(1)(a) and (b); applies from 2025-08-02 | general-purpose model provider | AI Office and national authorities on request; downstream providers | type and provenance of data, curation, number of data points, how obtained and selected | binding (the open-source exemption applies to these two) | eur-lex.europa.eu |
| Public summary of training content | Art. 53(1)(d); template first published 2025-07-24, current C(2025) 8311 of 2025-12-05 | general-purpose model provider | the public | data sources by class; licensed data as Yes or No plus modality | binding | digital-strategy.ec.europa.eu |
| Data manifest | the seller, per the buyer's terms (for example micro1 Order Form, terms of 2026-11-12) | the data holder | one prospective buyer | sources, volumes, rights and personal-data status of the data on offer | contract document where the terms make it one | micro1.ai |
The split that matters to a holder is the audience: the public, one buyer or a regulator on request. Documentation describes a dataset, while the chain of who held which right is the wider record covered under training data provenance.
Datasheets for Datasets: the seven question groups
Datasheets for Datasets is a set of questions, proposed by Timnit Gebru and six co-authors in a paper first posted on arXiv on 2018-03-23 and published in Communications of the ACM in December 2021, that a dataset's creators answer in seven groups.
The authors model the idea on the datasheets that accompany electronic components. They write for two stakeholder groups, dataset creators and dataset consumers, and state that creating a datasheet "is not intended to be automated". The groups, in the paper's order and with the paper's names, are:
- Motivation: why the dataset was created and by whom.
- Composition: what the instances are and how many there are.
- Collection Process: how the data was acquired, and whether the people concerned consented.
- Preprocessing/cleaning/labeling: what was removed, changed or annotated.
- Uses: the tasks the dataset fits and the uses to avoid.
- Distribution: how the dataset is shared, and on what terms.
- Maintenance: who supports, hosts and updates it.
The paper presents the questions as prompts for reflection and for communication between creators and consumers, so a datasheet is a record that a creator writes once and revises as the dataset changes. Two groups carry the licensing questions.
Distribution asks whether the dataset is released under a copyright or other intellectual property license or terms of use, whether fees are associated with those restrictions and whether export controls or other regulatory restrictions apply.
Collection Process asks "Did the individuals in question consent to the collection and use of their data?" These are the same questions a buyer's diligence asks of a seller.
Data Cards: Google's 31 themes and the OFTEn framework
Data Cards are structured summaries of essential facts about a machine learning dataset, proposed by Mahima Pushkarna, Andrew Zaldivar and Oddur Kjartansson of Google Research in a paper posted on 2022-04-03 for ACM FAccT 2022.
The template covers 31 themes arranged in sections that include Dataset Overview, Dataset Provenance, Extended Use, Transformations and Annotations. The authors created 22 Data Cards with 12 teams, although the abstract says "over 20". They organize what a card answers with a framework called OFTEn:
- Origins: planning, sources, licenses and collection methods.
- Factuals: statistical attributes of the data.
- Transformations: cleaning and labeling.
- Experience: deployment and benchmarking.
- n=1: a single worked example.
A card is meant to be read by people who did not build the dataset, which is why the template groups its themes into sections a reviewer can scan. The difference from a datasheet is the reader.
Data Cards are written by producers for agents, meaning reviewers and subject-matter experts who decide how a dataset may be used, while datasheets address creators and consumers.
Both record upstream sources and licenses, so a licensor's name or terms appear in either document only if the creator chooses to state them, and neither paper makes that mandatory.
Dataset cards and machine-readable metadata on hosting platforms
On Hugging Face, a dataset card is the README.md file of the dataset repository, and a YAML block at its top carries metadata such as license, language, size and task categories.
According to the Hub documentation, read 2026-11-16, the license value is displayed on the dataset page when the creator uses a recognized identifier. The card is therefore often the only license notice a downstream user sees, and a dataset without a card or without a license field shows no notice at all.
Two machine-readable formats sit beside the cards. Croissant 1.0, dated 2024-03-01 on the MLCommons documentation site, builds on schema.org and has a separate Responsible AI vocabulary.
The Data & Trust Alliance published Data Provenance Standards v1.0.0 on 2024-07-09, with 22 fields in three standards (Source 6, Provenance 8, Use 8); the parent page compares them with other provenance standards.
The two formats serve different readers: Croissant is read by software that loads the data, and the Data Provenance Standards are read by organizations that exchange data. Platform metadata is only as good as what creators enter.
The Data Provenance Initiative audited more than 1,800 text datasets on hosting sites and found licenses omitted in more than 70 % and wrong in more than 50 % of them (arXiv, 2023-10-25). The license types themselves, and what each permits, are covered under dataset licenses for AI training.
Documentation the EU AI Act requires about training data
The EU AI Act (Regulation (EU) 2024/1689) asks providers for three documents about training data, which are technical documentation for authorities, information for downstream providers and a public summary, plus data governance records for high-risk systems.
Technical documentation for the AI Office and downstream providers
Under Article 53(1)(a), a provider of a general-purpose AI model keeps technical documentation whose minimum content in Annex XI includes information on training, testing and validation data: the type and provenance of the data, curation methods such as cleaning and filtering, the number of data points, and how the data was obtained and selected.
Article 53(1)(b) and Annex XII point 2(c) give downstream providers the type and provenance of data and the curation methodologies, "where applicable". These documents go to the AI Office and national authorities on request and to integrators, not to the public, and the open-source exemption in Article 53(2) covers only (a) and (b).
Article 53 applies from 2025-08-02, AI Office enforcement starts on 2026-08-02, and models placed on the market before 2025-08-02 must comply by 2027-08-02.
The public summary of training content
Article 53(1)(d) requires a public summary of training content on a Commission template, first published on 2025-07-24 and current as Communication C(2025) 8311 final of 2025-12-05. Unlike a datasheet, it describes a model's data sources by class (licensed data, other private datasets, scraped web content), and licensed datasets appear only as Yes or No plus modality (section 2.2.1) while other private datasets are listed only if publicly known (section 2.2.2).
The template is read section by section in the EU AI Act Article 53 training data summary.
Data governance for high-risk AI systems under Article 10
Article 10(2) requires providers of high-risk AI systems to apply data governance practices that cover, among other points, "data collection processes and the origin of data, and in the case of personal data, the original purpose of the data collection" (point (b)) and data-preparation operations such as annotation, labelling and cleaning (point (c)).
The duty concerns high-risk systems, not general-purpose models as such, and it asks for governance practices, not a published document.
Regulation (EU) 2026/1744 amended Article 10(1), deleted Article 10(5) and replaced Article 10(6), leaving paragraph 2 unchanged. It also moved the application of Chapter III Sections 1 to 3 to 2027-12-02 for Annex III systems and 2028-08-02 for Annex I systems. Disclosure rules beyond the EU are compared under training data transparency rules.
Reusing a datasheet when licensing data to an AI buyer
A holder that answers the seven datasheet groups for its own data has most of what a buyer asks before a license, but the answers go into different documents with different readers.
The table maps each group to the place where a seller answers it before a license, as of 2026-11-16.
| Datasheet group | What the group asks (paper) | Where a seller answers it before a license | Source |
|---|---|---|---|
| Motivation | why the dataset was created and by whom | manifest description | arxiv.org 1803.09010 |
| Composition | what instances it holds | inventory and manifest volumes | arxiv.org 1803.09010 |
| Collection Process | how data was acquired; whether individuals consented | rights audit and consent records | arxiv.org 1803.09010 |
| Preprocessing/cleaning/labeling | what was removed or changed | personal-data removal and de-identification record | arxiv.org 1803.09010 |
| Uses | tasks it fits and uses to avoid | field-of-use clause | arxiv.org 1803.09010 |
| Distribution | license or terms of use, fees, export controls | license grant and price in the agreement | arxiv.org 1803.09010 |
| Maintenance | who maintains and updates it | delivery and update terms | arxiv.org 1803.09010 |
The outward document is the data manifest, written for one buyer. Under micro1's Data License Master Terms, the Data Manifest is part of the Order Form, and an inaccurate or misleading manifest allows the buyer to terminate with a refund of amounts paid, even after acceptance (s.4.3; terms last updated 2026-11-12, re-read 2026-11-16, micro1's terms).
Defined.ai states that it asks partners for data sourced with consent, documentation and the right to commercialize (Defined.ai's statement, checked 2026-11-09).
The accuracy of a description matters because a buyer relies on it when accepting delivery, and a seller that already holds a datasheet can adapt its answers instead of starting from nothing.
A research datasheet is not a contract document, and the fields and detail levels of the manifest are set out in the data manifest for AI buyers.
Benefits, risks, downsides and alternatives of publishing documentation for a licensed dataset, including not licensing
Publishing a datasheet or card makes a dataset easier to evaluate and easier to copy, and a holder that licenses data privately can keep its documentation private. The choice is the holder's, and no law or buyer program named on this page requires a public datasheet.
Benefits:
- Faster buyer review, because the seven datasheet groups answer common diligence questions.
- A correct license field where a platform displays it.
- Readiness for the EU documents that a model provider must produce about its training data.
Risks:
- A published description reveals composition and sources to competitors and to the buyer's counterparties.
- A public statement that conflicts with a later manifest invites a refund dispute where the terms provide for one.
- A missing or wrong license notice travels with every copy of the dataset.
Downsides:
- Staff time to answer seven groups or 31 themes.
- Documentation needs maintenance whenever the data changes, which is the Maintenance group's own question.
Alternatives:
- Keep documentation internal and share it only under NDA in diligence.
- Send a manifest to one buyer instead of publishing a card.
- Publish a summary-level card without sources.
- Not license the data at all.
Where documentation meets provenance, the seller process and the law
Documentation is one record among several that a dataset carries from its creator to a model, and the pages below take the others.
- Training data provenance: provenance standards, records and attribution.
- Data inventory for an AI data sale: the internal record behind a datasheet's Composition group.
- Rights audit for the Collection Process questions.
- Due diligence in an AI data deal: what buyers check and what sellers check.
- Data room for an AI data licensing deal: where documentation is shared under NDA.
- AI data licensing agreements: license grant, field of use and accuracy clauses.
- Major AI training datasets and their documentation.
- EU law on AI training data.
- Alternatives to selling data to AI labs, including not selling at all.
- Chain of rights for training data.
- Training data certifications and labels.
Frequently asked questions
What is a datasheet for a dataset?
A datasheet for a dataset is a document in which a dataset's creators answer seven groups of questions, from motivation to maintenance. It is voluntary research practice, modeled on the datasheets that accompany electronic components, and addresses dataset creators and dataset consumers.
Who wrote Datasheets for Datasets?
Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III and Kate Crawford wrote it. The paper was first posted on arXiv on 2018-03-23 and published in Communications of the ACM in December 2021.
What is a data card?
A Data Card is Google Research's format for summarizing a machine learning dataset across 31 themes. Mahima Pushkarna, Andrew Zaldivar and Oddur Kjartansson described it in a paper for ACM FAccT 2022, and a card is written by producers for reviewers and subject-matter experts.
Does the EU AI Act require a datasheet?
No, the EU AI Act names no datasheet format. It requires Annex XI technical documentation, Annex XII information and a public summary on the Commission template for general-purpose models, and Article 10 data governance for high-risk systems.
What is a data sheet outside machine learning?
A data sheet outside machine learning is the document that accompanies an electronic component. It is the model that the 2018 paper borrows for datasets.
How this page is sourced and dated
The page rests on the two papers on arXiv, the Hugging Face and MLCommons documentation, the Data & Trust Alliance standard, the EUR-Lex texts of Regulations (EU) 2024/1689 and 2026/1744 and the Commission template, with status as of 2026-11-16. This page is information, not legal advice.