partecipanti alla conferenza stampa di presentazione di Minerva /B

The Italian way to future chatbots: Sapienza launches the latest version of the AI language model "Minerva"

Minerva, developed within Fair (Future artificial intelligence research) and in collaboration with Cineca, is trained with 1.5 trillion words and raises the security standard of Italian LLMs

Sapienza research excellence in the field of artificial intelligence: the release of Minerva 7B, the latest version of the Minerva family of Large Language Models (LLM) trained from scratch for the Italian language, has been announced today. The new Minerva language model has been developed by the Sapienza NLP (Natural Language Processing) research group, led by Roberto Navigli, within Fair (Future artificial intelligence research), the project implementing the National Strategy for Artificial Intelligence thanks to PNRR funding, and in collaboration with Cineca, which provided the Leonardo supercomputer.

Minerva 7B is a more powerful version than the one that went live last April, with 7 billion parameters as opposed to the 3 billion of its predecessor, and thus a greater capacity for storing and processing texts, still based on open data sources, a distinctive feature in the LLM landscape.

After more than 5 months of relentless work, the research team arrived at this new version with a total of more than 2 trillion (thousands of billions) tokens, equivalent to about 1.5 trillion words. Using a new mix of specially created instructions in Italian, Minerva 7B underwent the so-called instruction tuning process, an advanced training technique for artificial intelligence models that aims to provide the ability to follow instructions and converse with the user in Italian.

It is precisely the instruction tuning that enables Minerva to better interpret requests and generate more relevant, coherent and contextually appropriate responses, avoiding as far as possible so-called hallucinations and the generation of vulgar, sexual, discriminatory and sensitive content. This is a key issue for all chatbots and one that the researchers in the Sapienza team are particularly concerned about.

The model is publicly available at www.minerva-llm.org and will be available for download in the coming weeks. This test phase will allow for further refinement based on discussions in the coming days.

The team that worked on the development of Minerva 7B includes no less than 15 male researchers and PhD students (in alphabetical order): Edoardo Barba, Tommaso Bonomo, Simone Conia, Pere-Lluís Huguet Cabot, Federico Martelli, Luca Moroni, Roberto Navigli, Riccardo Orlando, Alessandro Scirè, Simone Tedeschi; Stefan Bejgu, Fabrizio Brignone, Francesco Cecconi, Ciro Porcaro, Simone Stirpe also contributed. We would also like to thank Giuseppe Fiameni (NVIDIA) and Sergio Orlandini (Cineca).

Please read the Italian version of this page for more information.

Tuesday, 26 November 2024

© Sapienza Università di Roma - Piazzale Aldo Moro 5, 00185 Roma - (+39) 06 49911 - CF 80209930587 PI 02133771002