Minerva models are out!
We are excited to announce the launch of the Minerva models, the first family of LLMs pre-trained from scratch for the Italian language.
This initiative, a collaboration with CINECA using the Leonardo supercomputer, is part of the strategic PNRR FAIR initiative, demonstrating a significant advancement in the field of generative AI in Italy. Leading this project are Professor Roberto Navigli, recipient of two ERC grants and ACL fellow, along with the researchers in our group, Edoardo Barba and Simone Conia. Additionally, PhD students Pere-Lluís Huguet Cabot, Riccardo Orlando, and Luca Moroni have played integral roles in advancing the project's research and development.
What sets the Minerva models apart is their unique foundation—they have been pretrained from scratch using openly accessible texts, unlike earlier Italian models which were fine-tuned from models such as LLaMA and Mistral, whose training data remain undisclosed. Specifically, each Minerva model has been trained using a comprehensive collection of Italian and English sources, totaling over 500 billion words, the equivalent of more than 5 million novels. This transparency in training not only bolsters trust among users, the scientific community, public entities, and industry, but also encourages continual enhancements and represents a critical step towards stringent verification processes to ensure compliance with laws and regulations.
Check out our models on the Hugging Face Hub 🤗
Check out the news: 📰 ANSA, Il Sole 24 Ore, 📺 RaiNews (Article & VIDEO)