AIA Training Data Transparency Template Blueprint - v.2.0 - Requirements

Sep 4, 2026

Upload by Kyle

Blueprint for an AI Act Article 53(1)d training data transparency template applicable to general-purpose AI (GPAI) models. Provides a structured template describing overall training data size, the complete list of datasets and domains, data owners and modalities, purposes of use, dataset and data source provenance (scraped, repositories, licensed, proprietary, user-generated, synthetic), and applicable date ranges and legal/copyright compliance measures. Also specifies diversity and representativeness reporting and details of data processing steps such as filtering, cleaning, tokenization, feature extraction, and detoxification.

LanguageEnglish

CategoryPresentations

FormatPDF

Pages3

Reading Time1 min

AccessFree to download

Preface

Blueprint

Overall size

Data sets and data sources

Data diversity

Data processing

Blueprint of the template for the summary of content used to train general-purpose AI models (Article 53(1)d AIA)–v.2.01 Preface This blueprint for the template for the AI Act’s training data transparency requirement for generalpurpose AI models is based on the initial version presented in the policy brief “Suﬃciently detailed”? A proposal for implementing the AI Act’s training data transparency requirement for GPAI models, published in June 2024. The second version of the blueprint presented in this document and published in September 2024 has been further developed and reﬁned based on feedback from various stakeholders and a workshop conducted with experts from industry, academia, and civil society. Blueprint 1. Overall size 1.1. The total size of the training data (e.g., the total number of works and tokens, including information about the tokenization process for better comparability), the total number of data sets, and their size in GB. 2. Data sets and data sources 2.1. List of all data sets that were used in training the model, including the following information for each data set: 2.1.1. Data set identiﬁer: the data set’s name, along with, if applicable, a link to the data set, as well as the name and URL for the data collection of which the data set is a part. 2.1.2. Dataset size: The total size of the data set and information about what percentage of the total training data the data set represents, using an appropriate measurement methodology for consistency and comparability. 2.1.3. Owners or curators: groups or teams that own or compile the data set; if applicable, authors associated with the dataset. 2.1.4. Data set domain: humans (e.g., data set about people), objects (e.g., data set about places or objects), etc. 2.1.5. Type/modality: text, image, audio, video, code, etc. If multimodal, what combination (e.g., image-text pairs) . 2.1.6. If language data, what language/languages and, where feasible, other relevant characteristics (such as, for example, slang, dialects, usage contexts, etc.) . 1 The blueprint was developed by Zuzanna Warso (Open Future), Maximilian Gahntz (Mozilla Foundation), and Paul Keller (Open Future) . The authors would like to thank Yacine Jernite, Lucie-Aimée Kaffee, Shayne Longpre, Abeba Birhane, Stefan Baack, Kris Shrishak, Aviya Skowron, Mark Dingemanse, Frederike Kaltheuner, Andreas Liesenfeld, Claire Pershan, Michał “rysiek” Woźniak, Natali Helberger, João Pedro Quintais, Toni Lorente, and Rania Wazir for their valuable feedback on this blueprint. AIA transparency template blueprint v.2.0 1 of 3 2.1.7. If applicable, information about the speciﬁc purpose for which the data set was used in the training process (e.g., for instruction tuning, multilingual capabilities, etc.) . 2.2. For each data set (including, if applicable, the constituent data sets of a composite data set), information on the sources from which data was obtained – i.e., whether it was: - scraped from the internet. If so, information about the crawling methodology and names/identiﬁers of the crawlers used to obtain the data, including, for example, the seed and link selection criteria, as well a weighted list of the top 5 percent or 100 000 domains by data modality (e.g., text, images, video) . - collected from public repositories. If so, information on the source and, where applicable, the license should be provided. - copyright-protected content licensed from rightholders or third-party intermediaries. If so, information about the source and the license, including whether the licensing arrangement is exclusive or not, should be provided. - obtained from proprietary databases. If so, information about the source and, where applicable, the license, including whether the licensing arrangement is exclusive or not, should be provided. - generated by users of products or services offered by the provider. If so, by which products or services. - synthetically generated by the pro

FAQs

What information does the blueprint require about overall training data size?

It calls for total size of training data (including tokens and tokenization information), total number of datasets, and their size in GB.

What details must be provided for each dataset in the template?

The template requires a dataset identifier, dataset size with its share of total training data, owners/curators, domain, type or modality, relevant languages (when language data), and the specific purpose used during training (if applicable).

How does the blueprint address data sources and compliance?

For each data source, it requires date ranges and, where applicable, mechanisms and policies for complying with Union copyright-related rights, plus GDPR-related legal basis and data subject protection measures for personal data.

Sep 4, 2026

Upload by Kyle