Following the launch of Minerva 7B, the Sapienza NLP research group, led by Professor Roberto Navigli, in collaboration with the spin-off Sapienza Babelscape, presents ChatMinerva, the latest evolution of Sapienza’s Large Language Model: a multimodal AI assistant designed to understand text, images and documents, access the web in real time, and communicate in Italian with greater reliability and confidence. ChatMinerva began as a research project developed entirely in Italy and continues to stand out for its open, transparent and verifiable approach at every stage of its development: from pre-training to fine-tuning, right through to its security and moderation mechanisms.

“The launch of ChatMinerva marks a new milestone in our university’s journey of innovation in the field of artificial intelligence” - says Rector Antonella Polimeni. The evolution of the Minerva project towards multimodal and interactive AI assistants demonstrates Sapienza’s ability to transform cutting-edge research into tangible innovation, serving the advancement of knowledge and society. This achievement stems from the integration of scientific expertise, advanced infrastructure and collaboration with innovative local organisations, thereby reinforcing Sapienza’s role as a strategic driving force for research and technological innovation in the country.”

From language model to AI assistant

ChatMinerva radically expands the capabilities of the Minerva LLM, transforming it into a true multimodal AI assistant. Users can now upload photographs, images, scanned pages, scientific articles, reports and technical documentation, and ask ChatMinerva to interpret, summarise or analyse them, or answer questions about their content.

Key new features include:

• Multimodal understanding: Minerva is capable of combining visual and textual information and performing OCR on scanned text; it is also possible to engage in voice-based conversations with the model.
•  Real-time web access: thanks to a Web RAG system with query formulation based on the open-source search engine DuckDuckGo, ChatMinerva can respond using up-to-date information from the web.
• Handling of long documents: the model’s context window has been expanded to 32,000 tokens through “continual learning”, enabling the processing of complex documents and lengthy conversations.
• Enhanced security: a special component monitors and validates user inputs and system responses to filter out unwanted, unreliable or sensitive content

A user-trained model

The new version of the Minerva LLM used in the assistant has undergone an extensive fine-tuning process using millions of examples of instructions, both textual and multimodal, thanks to the computing power of CINECA’s Leonardo supercomputer, which has been used for all training processes. The direct contribution of the large user community has also been crucial: the interactions gathered during the public use of Minerva 7B have enabled a significant improvement in the system’s conversational abilities and the quality of its responses. “Minerva has grown alongside its users. Every prompt, interaction and piece of feedback has helped to build an AI assistant that is more useful, more robust and better attuned to people’s real needs” - says Roberto Navigli, a lecturer at Sapienza University and project lead.

Italian technology, developed to make an impact

Minerva remains the leading Italian project in the field of Large Language Models developed with direct control over the data sources and training processes, and the only one overseen by a public university in Italy.

“We want to demonstrate that it is possible to develop cutting-edge AI technology in Europe and Italy too, using an open, scientifically rigorous and independent approach” - continues Navigli. “Chat Minerva is the result of passion rather than budget, thanks to the tireless work of dozens of researchers, PhD students, students and collaborators who believe in the possibility of creating Italian AI technology as a foundation for building competitive products, even though it originated in a context very different from that of the major global tech giants, but with the strength of the collaboration between Italy’s leading university and a successful spin-off, Babelscape, which has just celebrated 10 years of intensive research and industrial innovation.”

What's in store for the future

The development of Minerva continues apace. The team is already working on four main areas:

• Ongoing improvements to the ChatMinerva assistant, with updates and new features released weekly.
• Agentic Minerva, a new generation of AI systems capable of planning actions and using tools independently.
• A larger and more powerful Minerva model, currently being trained on CINECA’s HPC infrastructure.
• Cutting-edge research that will make Minerva a one-of-a-kind model.

Minerva thus reaffirms the ambition of Italian research and innovation in the field of artificial intelligence: to develop open, controllable technologies designed to have a real impact on society.

Further Information

Roberto Navigli
Department of Computer, Control and Management Engineering "Antonio Ruberti"

Sapienza Università di Roma
navigli@diag.uniroma1.it
https://www.diag.uniroma1.it/navigli/