The objective of the project is building the next-generation version of the first pre-trained Italian LLM: Minerva 2.
The Minerva 2 project is designed to strengthen the national AI ecosystem and support advanced applications in both corporate and institutional contexts. Leveraging the computational resources of Cineca's HPC Leonardo, the project aims to:
- Massively integrate Italian texts, including literary and technical works, in collaboration with publishing houses
- Specialize the model on Retrieval-Augmented Generation (RAG) to improve response accuracy and reliability
- Extend context length up to 32k-64k tokens, enabling large-scale dialogue sessions and document analysis
- Develop advanced multilingual capabilities, ensuring high quality across different language pairs
- Implement symbolic and logical reasoning modules, enhancing deductive capabilities
- Enable multimodality by combining text, images, and audio
- Optimize architecture through advanced parallelism (model/data/pipeline) and MoE (Mixture of Experts) techniques
- Prepare fine-tuning pipelines for coding and software development tasks
This initiative aims to strengthen Italian technological sovereignty, promote AI adoption in both public and private sectors. Furthermore, The company aims to reinforce Italy's leadership in LLM research and innovation.
Roberto Navigli, Babelscape®: Language generation, semantically grounded, Italy