overthelex/multi-legal-bench
收藏资源简介:
Multi-Legal-Bench是一个跨法域的法律基准数据集,旨在评估大型语言模型在多个司法管辖区、语言和法律传统下的法律推理能力。数据集覆盖六个国家(乌克兰、法国、荷兰、波兰、捷克共和国、立陶宛),涉及五种语言家族(斯拉夫语、罗曼语、日耳曼语、波罗的语),并基于1.22亿份法院判决的元数据。它定义了五个任务:法院类型分类(CTC)、判决形式分类(JFC)、案件结果预测(COP)、原因类别预测(CCP)和法律规范提取(NE)。数据以JSONL格式存储,每个记录包含唯一标识符、法域代码、语言代码、任务标识符、原始标签、标准化标签、决策日期、文本长度、完整判决文本以及特定任务的字段(如事实部分或法律规范引用)。数据集包含15个任务-法域组合,总计13,096个决策,用于研究任务难度稳定性、少样本效应、跨语言迁移等关键问题,并支持相关学术论文。数据来源于各国官方开放的法院登记系统,采用CC-BY-SA 4.0许可协议。
Multi-Legal-Bench is a cross-jurisdictional legal benchmark dataset aimed at evaluating the legal reasoning capabilities of large language models (LLMs) across multiple jurisdictions, languages and legal traditions. The dataset covers six countries: Ukraine, France, Netherlands, Poland, Czech Republic and Lithuania, spans five language families (Slavic, Romance, Germanic, Baltic), and is built upon metadata from 122 million court judgments. It defines five tasks: Court Type Classification (CTC), Judgment Form Classification (JFC), Case Outcome Prediction (COP), Cause Category Prediction (CCP) and Legal Norm Extraction (NE). The data is stored in JSONL format, with each record containing a unique identifier, jurisdiction code, language code, task identifier, original label, standardized label, decision date, text length, full judgment text, and task-specific fields such as the factual section or legal norm citations. The dataset includes 15 task-jurisdiction combinations, totaling 13,096 decisions, which are used to study key issues including task difficulty stability, few-shot effects, cross-lingual transfer, and support relevant academic papers. The data is sourced from official open court registry systems of the respective countries, and is licensed under CC-BY-SA 4.0.




