News from Sapienza NLP

ChatMinerva is out!

Posted on

Sapienza NLP and Babelscape release ChatMinerva, a multimodal AI assistant built on the Minerva large language model.

Sapienza University of Rome and its spin-off Babelscape today announced the release of ChatMinerva, a new multimodal AI assistant built on the Minerva large language model. The launch marks a significant evolution of the Minerva project, expanding its capabilities beyond text generation to include image understanding, document analysis, real-time web access, and enhanced safety mechanisms. Developed entirely in Italy, ChatMinerva combines advanced language understanding with multimodal reasoning, enabling users to interact with text, images, scanned pages, scientific papers, reports, and technical documents through a single conversational interface. Among its key features are:
  • Multimodal understanding, allowing the system to process both visual and textual information, perform OCR on scanned documents, and support voice-based interaction.
  • Real-time web access through a Web Retrieval-Augmented Generation (Web RAG) system powered by the open search engine DuckDuckGo, enabling responses based on up-to-date online information.
  • Long-document processing, with a context window extended to 32,000 tokens, making it possible to analyze complex documents and maintain longer conversations.
  • Enhanced safety and moderation, with dedicated components designed to validate user inputs and system outputs, helping reduce harmful, unreliable, or sensitive content.
The latest version of Minerva powering ChatMinerva has been further refined through large-scale instruction tuning on millions of textual and multimodal examples. Training and fine-tuning were carried out using the Leonardo supercomputer at CINECA. The project also benefited from feedback and interactions collected from the public deployment of Minerva 7B, which helped improve the model's conversational abilities and response quality. ChatMinerva continues the Minerva initiative's mission of advancing trustworthy, transparent, and controllable AI technology developed in Europe. The project represents one of the most significant research efforts in large language models and remains unique in being led by a public university with full oversight of data sources, training methodologies, and safety mechanisms. ChatMinerva is available at https://chatminerva.org.