Public Summary of Training Content: [Model name] [version]
Working draft of the public summary of training content that every
provider of a general-purpose AI (GPAI) model must publish under
Article 53(1)(d) of the EU Artificial Intelligence Act (Regulation (EU)
2024/1689, as amended by Regulation (EU) 2026/1744). It follows the
three sections of the Template for the Public Summary of Training
Content adopted by the European Commission on 24 July 2025 (C(2025)
5235, Annex), with an internal preparation part that is removed before
publication. Replace every [bracketed] value and every row marked
Example. This is a documentation format, not legal advice: publish
on the official Commission template or online form, and check the
current version and the consolidated Regulation before use.
Document control
| Field | Value |
|---|
| Document ID | [TCS-001] |
| Provider | [Legal name] |
| Models and versions covered | [Versioned names] |
| Summary version | [1.0] |
| Status | [Draft / Legal review / Approved / Published / Superseded] |
| Owner | [@data-governance-lead] |
| Reviewers | [@legal], [@copyright-counsel], [@privacy], [@research-lead] |
| Commission template version used | [Annex to C(2025) 5235 of 24 July 2025, or later version] |
| Publication URL | [Provider website page] |
| Last updated | YYYY-MM-DD |
| Classification | Part A internal; Part B public once approved |
Approvals
| Role | Name | Decision | Date |
|---|
| Data governance lead | [@name] | [Approve / Reject] | YYYY-MM-DD |
| Legal and copyright | [@name] | [Decision] | YYYY-MM-DD |
| Privacy | [@name] | [Decision] | YYYY-MM-DD |
Revision history
| Version | Date | Author | Change summary |
|---|
| 0.1 | YYYY-MM-DD | [@name] | First draft |
| 1.0 | YYYY-MM-DD | [@name] | Example: Published with model release |
| 1.1 | YYYY-MM-DD | [@name] | Example: Six-monthly update after continued training on new licensed news data |
1. Scope and applicable dates
| Question | Answer |
|---|
| Is the model a GPAI model placed on the EU market? | [Yes, with reason] |
| Date placed on the EU market | YYYY-MM-DD |
| Applicable deadline | [At placement on the market for models placed on or after 2 August 2025 / By 2 August 2027 for models placed before 2 August 2025] |
| Modified model? | [No / Yes: base model name and link to its summary] |
| Provider is an SME or start-up? | [Yes / No: affects the domain list in section 5.3] |
| Code of Practice signatory (copyright chapter)? | [Yes / No] |
The obligation has applied since 2 August 2025. The Commission's
explanatory notice says that for models placed on the market before
that date the summary should be published no later than 2 August 2027,
and that AI Office enforcement for GPAI rules starts on 2 August 2026.
Non-compliance can be fined up to 3% of annual worldwide turnover or
15 million EUR, whichever is higher (Article 101). A provider that
modifies an existing GPAI model and becomes its provider reports only
the content used for the modification and names the original model.
Confirm these points against the consolidated Regulation and current
Commission guidance; dates in this area have been amended before.
2. Sources of figures
| Figure in Part B | Internal source | Measured at | Owner |
|---|
| Size per modality | Example: Data pipeline manifest DPM-v7 | After deduplication and filtering | [@data-eng] |
| Large public datasets | Example: Dataset register DR-2026 | After filtering | [@data-eng] |
| Domain list | Example: Crawl index, tokens per registrable domain | After filtering | [@crawl-lead] |
| Licensed data modalities | Example: Licence register | Contract terms | [@legal] |
2.1 Worked example: large dataset threshold
A publicly available dataset is "large" when, for any of its
modalities, it exceeds 3% of the total size of all publicly available
datasets used for that modality, measured after pre-processing.
| Dataset (Example) | Text tokens after filtering | Share of public text total | Large? |
|---|
| PUB-001 web snapshot extract | 410 billion | 17.1% | Yes |
| PUB-002 open code collection | 96 billion | 4.0% | Yes |
| PUB-003 open encyclopedia dump | 55 billion | 2.3% | No |
| Other public datasets (38) | 1,839 billion | 76.6% | Each below 3% |
| Total public text | 2,400 billion | 100% | Threshold 72 billion |
2.2 Worked example: domain list size
| Case (Example) | Domains crawled and used | Rule | Domains to list |
|---|
| Large provider | 48,000 | Top 10% by content size | 4,800 |
| SME or start-up | 48,000 | Top 5% (2,400) or 1,000, whichever is lower | 1,000 |
3. Review checklist
| Check | Result |
|---|
| Every dataset reported in one category only (synthetic data excepted) | [Pass / Fail] |
| Information gaps for pre-2025 models stated and justified | [Pass / Not applicable] |
| No trade secret disclosed beyond what the template asks | [Pass / Fail] |
Copyright policy summary linked (ai-data/gpai-copyright-policy) | [Pass / Fail] |
Consistent with the technical documentation (ai-data/gpai-model-documentation) | [Pass / Fail] |
| Published on the website and in every public distribution channel | [Pass / Fail] |
4. General information (template section 1)
4.1 Provider identification
| Field | Value |
|---|
| Provider name and contact details | [Name, address, contact] Example: Brindlecove AI Ltd, a fictional provider |
| Authorised representative (if established outside the EU, Article 54) | [Name and contact / Not applicable] |
4.2 Model identification
| Field | Value |
|---|
| Versioned model names | [Names and versions] Example: Brindle-7 Base 1.0, Brindle-7 Chat 1.0 |
| Model dependencies | [Base models modified and links to their summaries / None] |
| Date of placement on the EU market | [Date per model version] |
| Further documentation | [Model card link] |
4.3 Modalities, size and characteristics
The template uses three broad size ranges per modality: text in tokens
(below 1 billion, 1 billion to 10 trillion, above 10 trillion), images
(below 1 million, 1 million to 1 billion, above 1 billion), audio and
video in hours (below 10,000, 10,000 to 1 million, above 1 million).
| Modality | Size range | Types of content |
|---|
| Text | Example: 1 billion to 10 trillion tokens | Web pages, books in the public domain, scientific articles, source code, licensed news |
| Image | Example: Not present | |
| Audio | Example: Not present | |
| Video | Example: Not present | |
| Other | [Specify modality, approximate size and unit] | |
| Field | Value |
|---|
| Latest date of data acquisition | [MM/YYYY]; [state whether the model is continuously trained after this date] |
| Linguistic characteristics | Example: About 70% English; all 24 EU official languages present, with French, German and Spanish the largest |
| Other characteristics | [Regional or demographic specificities, where known and practicable] |
| Additional comments (optional) | [Tokeniser used for counting, sampling rates] |
5. List of data sources (template section 2)
5.1 Publicly available datasets
| Field | Value |
|---|
| Used? | [Yes / No] |
| Modalities | [Text, image, video, audio, other] |
| Large datasets | Example: PUB-001 [name, link, collection period, selection approach]; PUB-002 [name, link] |
| Other public datasets | [General description: modalities, nature of content, languages, collection period or "not known"] |
5.2 Private datasets from third parties
| Field | Value |
|---|
| Commercially licensed from rightsholders or their representatives | [Yes / No]; modalities: [list] |
| Other private datasets from third parties | [Yes / No]; modalities: [list] |
| Publicly known private datasets | [Names and links] |
| Description of other private datasets | [Modalities, nature of content, languages] |
5.3 Data crawled and scraped from online sources
| Field | Value |
|---|
| Crawlers used by or for the provider | [Yes / No]; names: [crawler identifiers] |
| Purposes | [Training data collection; other purposes] |
| Crawler behaviour | Example: Honours robots.txt and other machine-readable opt-outs; does not bypass paywalls, logins or CAPTCHAs |
| Collection period | [MM/YYYY to MM/YYYY] |
| Content and sources | [Types of sites, languages, regions] |
| Modalities | [List] |
| Most relevant domain names | [Top 10% of domains by content size, or for SMEs top 5% or 1,000, whichever is lower; link to downloadable list] |
The notice also recommends a voluntary, good-faith route for
rightsholders to ask whether content from a domain not on the list was
used. Record it here if offered: [contact or form URL].
5.4 User data
| Field | Value |
|---|
| Data from user interactions with the model used for training? | [Yes / No] |
| Data from the provider's other services or products used? | [Yes / No]; services: [description] |
| Modalities | [List] |
5.5 Synthetic data
| Field | Value |
|---|
| Synthetic data created by or for the provider for training? | [Yes / No]; modalities: [list] |
| GPAI models on the market used to generate it | [Names and links to their summaries] |
| Other models used, including own models | [Description, including general description of their training data where needed] |
5.6 Other sources
| Field | Value |
|---|
| Other sources used? | [Yes / No] |
| Narrative description | Example: Human-written demonstrations and preference ratings commissioned from annotation vendors for alignment |
6. Data processing aspects (template section 3)
6.1 Respect of text and data mining reservations
| Field | Value |
|---|
| Signatory to the GPAI Code of Practice with copyright commitments? | [Yes / No] |
| Measures before and during collection | Example: robots.txt honoured at crawl time; other machine-readable opt-out protocols checked; reservations recorded per domain; third-party dataset suppliers contractually required to honour Article 4(3) reservations of Directive (EU) 2019/790 |
| Copyright policy summary | [Link, if public] |
6.2 Removal of illegal content
[Describe in general terms the measures used to avoid or remove illegal
content from training data, such as block lists, keyword filters and
classifiers, in particular for child sexual abuse material, terrorist
content and unauthorised use of protected material.]
6.3 Other information (optional)
[Any further processing measures relevant to the rights of interested
parties, for example personal data minimisation steps.]
7. References
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article
53(1)(d), Recitals 107 and 108, as amended by Regulation (EU)
2026/1744; consolidated text on EUR-Lex.
- European Commission, Explanatory Notice and Template for the Public
Summary of Training Content for general-purpose AI models, 24 July
2025,
https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models
- Directive (EU) 2019/790 on copyright in the Digital Single Market,
Article 4(3).
- Related templates: GPAI model technical documentation
(
ai-data/gpai-model-documentation), datasheet for datasets
(ai-data/datasheet-for-datasets), model card (ai-data/model-card),
AI regulatory applicability register
(compliance/ai-regulatory-applicability-register), copyright policy
(ai-data/gpai-copyright-policy).
Standard and licence note
Based on Article 53(1)(d) of Regulation (EU) 2024/1689 as amended by
Regulation (EU) 2026/1744, and on the structure of the Commission's
Template for the Public Summary of Training Content (Annex to C(2025)
5235, 24 July 2025). The section headings follow the Commission
template; all guidance and examples are original wording, and the
organisation and figures are fictional. Publish on the official
template or online form and check the current version, the explanatory
notice and the consolidated Regulation before use. This is a
documentation format, not legal advice. Template licence: MIT.