MAWARITH: A Dataset and Benchmark for Legal Inheritance Reasoning with LLMs

The article introduces MAWARITH, a large-scale annotated dataset comprising 12,500 Arabic inheritance cases aimed at training and evaluating large language models (LLMs) on complex Islamic inheritance reasoning. It emphasizes end-to-end reasoning capabilities, including identifying heirs, applying juristic rules, and calculating shares, with a novel evaluation metric, MIR-E, that assesses various reasoning stages. This dataset is significant for practitioners as it addresses the challenges of multi-step reasoning in legal contexts, providing a benchmark to improve model performance in real-world applications of Islamic inheritance law.