Current Large Language Models (LLMS) often suffer from hallucinations and a lack of domain-specific reliability. This project proposes the creation of a novel 3B–7B parameter multilingual model that integrates a generative core with a distinct, retrievable memory and knowledge layer, mimicking brain-inspired cognitive processes. This architectural shift moves away from static parameter knowledge toward dynamic, factual retrieval, specifically optimised for the European public sector.
The project leverages a validated methodology, scaling up from a suite of four existing Small Language Models (Italian, French, Spanish, Portuguese; 350M parameters each). These precursors have demonstrated superior performance on international benchmarks using open-source internet corpora. By synthesising the training data of these highly successful models and introducing rigorous post-training on European administrative protocols, we will deliver an AI system capable of resolving complex public sector workflows with high precision.
The result will be a computationally efficient, modular AI solution that bridges the gap between general-purpose language generation and verifiable, domain-specific administrative assistance.
Principal Investigator, Company and Country
Alessandro Ercolani, PagoPA, Italy