SMARTER - Structured Multi-sourced AI Reasoning for Enhanced Decision-Making

• Scientific lead: YUN Bruno
• Is SyCoSMA the project lead? Yes
• Start date: 2024-01-01
• End date: 2026-01-01

In this project, we will instantiating a custom large-language model based on the recent Meta AI’s Llama 2 70B model and training it with both synthetic natural language arguments and arguments extracted from existing platforms. We will perform training through LoRA, a novel method to fine-tune LLMs that freezes the pre-trained model weights and adds trainable rank decomposition matrices to each layer of the Transformer architecture, which can capture task-specific features with fewer parameters. This synergy between argumentation graphs, knowledge graphs, and the LLM will be integrated in a decision-support system and evaluated in various decision scenarios.