Lightweight LLMs for Court View Generation

A study explores lightweight large language models (LLMs) under 2B parameters for Criminal Court View Generation (CVG), a Legal AI task that produces court views from case facts. The research addresses four questions: the impact of LLM architecture on CVG quality and charge prediction, the role of model size, comparison with Deep Neural Networks (DNNs), and whether generating court views first improves charge prediction over direct prediction. The authors developed CVGEvalKit, an evaluation framework with three public datasets for CVG and charge prediction. Comprehensive experiments were conducted on this framework, with models trained on a mix of data.