Information on the processing of personal data Velvet Models

3. Data processed, purposes and legal basis of the processing

The data used as part of the development activities of the Velvet family models come from datasets and sources that can be legitimately used for the purposes of research, development and training of artificial intelligence systems. Such sources include, but are not limited to, academic and scientific publications, books and editorial publications, documents and information from institutional or governmental sources, encyclopedic texts, legal documentation, collections of content, and third-party datasets distributed under licenses compatible with these purposes.

The sources and datasets used may occasionally contain personal data relating to natural persons, such as:

• identification data (e.g. name and surname);
• contact information that may be contained in the original content;
• contextual information contained in the texts, such as roles, qualifications or assignments;
• in residual cases, special categories of personal data and judicial data that may be contained in the sources of origin and, where applicable, manifestly made public by the data subjects.

Any personal data present in the datasets are processed exclusively as part of the data selection, collection, verification, organization, filtering, minimization and transformation activities (jointly, the “data curation activities“) necessary for the preparation of the datasets used for the development of the models.

As part of data curation activities, datasets are previously subjected to technical and organisational procedures aimed at identifying and, as the case may be, anonymizing, pseudonymizing or removing any personal data that may be present. If it is not possible to adequately reduce the risk of identification without compromising the content of the data, the relevant portions of the content are excluded from the datasets intended for training.

The training of the Velvet family models takes place exclusively on the datasets resulting from the data curation activities described above and is not aimed at identifying natural persons or at learning, storing or reproducing personal information referable to identified or identifiable subjects.

The purpose of the processing is the development, training, validation, improvement and evaluation of artificial intelligence models belonging to the Velvet family, intended to support functionalities such as natural language understanding and generation, information extraction, content classification, machine translation, question answering systems, retrieval augmented generation (RAG),  as well as, for specialized models, speech processing and speech transcription.

The legal basis for the processing of any personal data present in the datasets is the legitimate interest of the Data Controller pursuant to art. 6, par. 1, letter f), GDPR, consisting in the development and improvement of artificial intelligence models through the use of appropriately selected datasets subjected to measures of minimization, filtering and reduction of the risk of identification. The Data Controller has assessed this interest taking into account the nature of the sources used, the reasonable expectations of the data subjects and the measures taken to minimise the impact of the processing on the fundamental rights and freedoms of the data subjects. If the datasets contain data belonging to special categories pursuant to art. 9 GDPR, the relevant processing takes place exclusively in the cases permitted by the applicable legislation and, in particular, on the basis of art. 9, par. 2, letter e) GDPR, where such data have clearly been made public by the data subject.