"ECM-Bench"

"ECM-Bench is an explainable case matching benchmark constructed from Chinese criminal judicial documents. The dataset contains 60,000 normalized criminal cases, 20,000 explainable case-matching triplets, and 180,940 claim verification records. Each triplet consists of a query case, a better-matching case, and a hard negative case that is partially similar but less appropriate overall. Beyond pairwise matching scores, ECM-Bench provides multi-dimensional explanation annotations, evidence sentence anchors, case feature signals, natural-language triplet summaries, and atomic explanation claims. The official train, development, and test splits contain 16,034, 1,969, and 1,997 triplets, respectively. The dataset is designed to support research on legal case retrieval and ranking, triplet preference learning, explainable legal matching, evidence grounding, rationale generation, and explanation faithfulness evaluation. By combining hard negative construction with fine-grained explanation and claim-level verification labels, ECM-Bench provides a challenging benchmark for developing interpretable legal AI systems in the Chinese criminal justice domain."