LawngNLI

LawngNLI是一个专门为法律领域设计的长前提自然语言推理数据集，由宾夕法尼亚大学的William Bruno和Dan Roth创建。该数据集包含约14万个示例，这些示例是从美国法律意见中提取的，经过自动标记并具有高的人工验证准确性。数据集的特点是前提较长且具有多粒度，旨在解决模型在处理长文本时的推理能力问题。LawngNLI不仅用于评估模型在长文本上的推理能力，还用于基于推理的检索任务，如法律案例检索，帮助提高法律工作的效率和公平性。

LawngNLI is a long-premise natural language inference dataset specifically tailored for the legal domain, developed by William Bruno and Dan Roth from the University of Pennsylvania. This dataset contains approximately 140,000 instances extracted from U.S. legal opinions, which are automatically annotated and validated with high human verification accuracy. Characterized by long and multi-granularity premises, the dataset aims to address the challenge of model reasoning capabilities when processing lengthy texts. LawngNLI serves not only as a benchmark for evaluating model reasoning performance on long texts, but also supports inference-based retrieval tasks such as legal case retrieval, assisting in improving the efficiency and fairness of legal work.