jusbrasil/indice-jusia
收藏资源简介:
该数据集是一个用于人类评估模型生成的法律答案和文档的结构化评分标准。它包含五个维度的评估标准:法律和事实正确性(检查法律和事实错误、引用准确性、管辖权和层级正确性)、完整性(遵循用户指令和法律文档结构完整性)、可靠性(声明的可验证来源、主要/官方来源、引用支持度、近期判例)、语言流畅性(精确性和简洁性)以及自由形式的观察。每个标准使用YES、NO、N/A标签进行评分,以确保注释的一致性和易于聚合。数据集支持英语和葡萄牙语,适用于文本生成和问答任务。
This dataset is a structured rubric for human evaluation of model-generated legal answers and documents. It includes five dimensions of evaluation criteria: legal and factual correctness (checking for legal and factual errors, citation accuracy, jurisdiction and hierarchy correctness), completeness (adherence to user instructions and structural completeness of legal documents), reliability (verifiable sources for statements, primary/official sources, citation support, recent jurisprudence), language fluency (precision and conciseness), and free-form observations. Each criterion is scored using YES, NO, N/A labels to ensure consistent and easily aggregated annotations. The dataset supports English and Portuguese languages and is suitable for text-generation and question-answering tasks.





