AIG-048 Public Training Content Summary
Description
A summary of the content used to train each general-purpose AI model is published where the public can reach it, before the model is placed on the market, in the template the regulator supplies. The summary names the main data sources and collections used, the modality and approximate size of the data, the share drawn from public web crawling with the domains or collections identified to the level the template asks, the licensed, user-sourced and synthetic data used, and the processing applied to honour rights reservations and to remove unlawful content. A dated version of the summary exists for each model release, the current version is linked from the model documentation, and each summary is traceable to the provenance records of the datasets it summarises.
Rationale
The Commission template of July 2025 asks for three things: general information about the model and the data by modality and size, a list of data sources with the largest web domains named, and the processing applied for rights and unlawful content. Writing the summary from the AIG-013 provenance records rather than from memory is what makes it defensible: every source named in the summary should resolve to a provenance record. Publication is the point of the duty; a summary held ready but unpublished does not pass. Boundary with AIG-047: that control releases documentation to authorities and downstream providers on request, this one publishes to everyone. Boundary with AIG-013: that control is the internal record, this one is the public summary derived from it. gpai-provider seat (ADR-031).
Applicability (9 profiles)
The public training content summary is the gpai-provider seat (ADR-046). A provider that fine-tunes a third-party model for its product records provenance under AIG-013 and is not the models provider.
A general-purpose model providers duty (gpai-model-provider, ADR-046). The deployer takes the model documentation and the published training summary the provider issues into its AIG-032 assessment.
Art.53(1)(d): the public summary of training content, for every general-purpose model, with no open-source relief.
The public training content summary is the gpai-provider seat (ADR-046). A provider that fine-tunes a third-party model for its product records provenance under AIG-013 and is not the models provider.
A general-purpose model providers duty (gpai-model-provider, ADR-046). The deployer takes the model documentation and the published training summary the provider issues into its AIG-032 assessment.
The public training content summary is the gpai-provider seat (ADR-046). A provider that fine-tunes a third-party model for its product records provenance under AIG-013 and is not the models provider.
The public training content summary is the gpai-provider seat (ADR-046). A provider that fine-tunes a third-party model for its product records provenance under AIG-013 and is not the models provider.
The public training content summary is the gpai-provider seat (ADR-046). A provider that fine-tunes a third-party model for its product records provenance under AIG-013 and is not the models provider.
The public training content summary is the gpai-provider seat (ADR-046). A provider that fine-tunes a third-party model for its product records provenance under AIG-013 and is not the models provider.
Framework Mappings (2)
| EU-AI-Art.53.4 | GPAI Model Obligations — Training Data Summary | full |
| GV-1.6-003 | AI System Inventory | GV-1.6-003 | informative |
Evidence (2)
The published training content summary for a named model, in the regulator's template, at the version matching the current release.
Example: Public summary of training content, Aurora-2 v2.3, published 20 August 2026 at the model page, template version July 2025.
Test: Verify: (1) the summary is reachable by the public from where the model is offered, without registration, (2) its publication date precedes the date the release was placed on the market, (3) each section the template asks for is completed rather than marked not applicable without a reason, (4) the largest web domains and the collections named match the provenance records for the release, (5) a dated version exists for each release in the period and the model documentation links to the current one.
Export of the provenance records for the datasets behind a release, reconciled against the sources the published summary names.
Example: Provenance register export for the Aurora-2 v2.3 training mix, 41 datasets, exported 19 August 2026.
Test: Verify: (1) every data source named in the published summary resolves to a provenance record for the release, (2) every provenance record in the release mix is represented in the summary at the level the template asks, (3) the modality and size figures in the summary agree with the register totals within the rounding the template allows, (4) the rights-reservation and unlawful-content processing the summary describes matches the transformations the provenance records carry.
Questions (2)
Is a summary of the training content published for each general-purpose model you place on the market?
Published means reachable by the public without registration. A summary written but held internally, or released only to customers, is a no.
Which of the following does the published summary cover?
Options follow the template's sections. The domain list is the item most often left generic; the template asks for the largest domains by name.