A legal-domain dataset for evaluating retrieval, reranking, and RAG systems under realistic and challenging conditions. ⚠️ This dataset is intentionally low-overlap (controlled).However, unlike extreme synthetic setups, it preserves a non-zero lexical signal to remain realistic.

Overview

This dataset is designed to:

reduce lexical overlap between queries and relevant chunks
preserve semantic alignment
maintain heterogeneous… See the full description on the dataset page: https://huggingface.co/datasets/ReDiX/ReDiX-Benchmark-Legal-ita.