EU AI Act — AI training-data transparency
Article 53(1)(d) of the EU AI Act (Reg. (EU) 2024/1689) requires providers of general-purpose AI (GPAI) models to publish a public summary of the content used to train the model, on the AI Office's standardised template (in force 2 Aug 2025). Each summary discloses, per modality (Text / Image / Audio / Video), a banded training-data size, the data-acquisition cut-off, and yes/no flags per data-source category. There is no single registry of the filled summaries — each provider self-publishes in its own format — so this is a cross-provider, comparable dataset built from the archived summaries: one row per disclosed field, identified by provider × model × section × field. The coarse size bands are mapped to a numeric size_rank (1 / 2 / 3, 0 = “not applicable”) so sizes compare across providers — it is an ordinal rank, so compare it with MIN/MAX/AVG, never SUM. Coverage is a starting, expandable set (Google + Meta + Microsoft).
Disclosed training-data size
Interactive views of the disclosed size bands. Each chart has a text alternative for screen readers. Size rank: 1 = smallest band (“less than …”), 2 = middle range, 3 = largest (“more than …”); 0 = the modality is not part of the training data.
Text training-data size, by model
The banded size of each model's text training data, as a comparable rank (1–3). Google's Gemini 3 Pro and Meta's Muse Spark disclose the largest band (“more than 10 trillion tokens”); Microsoft's Phi-4 models disclose the middle band.
Disclosed size rank, by modality
The comparable size rank each model discloses for each modality. A rank of 0 means the model does not train on that modality (e.g. the text-only Phi-4).
Training-data size bands disclosed
The exact size band each summary discloses per modality, as filed. Bands are coarse by design (the template offers a fixed set) — “Not applicable” means the modality is not part of the training data.
Data cut-off & data-source categories
The data-acquisition cut-off and the yes/no data-source flags each summary discloses. Blank cells (—) are fields the summary does not report.
Refine or extend these in the query builder (compare size_rank with MIN/MAX/AVG, never SUM), browse the full field list on the Schema page, or read the Methodology.