Y-J-Ju/RegOps-Bench
收藏资源简介:
RegOps-Bench是一个用于评估检索增强系统的韩国国家研发法规合规性问答基准。该数据集专注于韩国国家研发法规,包括《국가연구개발혁신법》(国家研发创新法)及相关法令、规则、通知和操作手册。数据集包含两个部分:corpus(语料库)包含718篇法规文章,作为问答系统的检索来源;test(测试集)包含250个问答对,每个问题都针对研究人员或管理员在韩国进行政府资助研究时面临的特定程序决策(如人员成本计算、设备采购、差旅费报销、间接成本规则等)。与传统的多跳或法律问答不同,每个问题都需要在多层权威结构(法规→法令→规则→通知→手册)中进行结构化程序查找和证据集闭合,并附有条款级别的真实引用。数据集还标注了难度级别(L1:单条款事实查找;L2:单文档内多面推理;L3:跨文档引用遍历;L4:条件/例外配对推理与程序链)、问题类型(如多面、单条款、一般原则等)和机构特定子集(如大学、政府出资机构等)。数据来源于韩国国家法律信息中心(국가법령정보센터)和IITP的FAQ文档,语言为韩语,部分内容提供英文翻译。
RegOps-Bench is a Regulatory Compliance QA benchmark for evaluating retrieval-augmented systems on Korean national R&D regulations (국가연구개발혁신법 and related decrees, rules, notices, and operational manuals). The dataset consists of two parts: a corpus split with 718 regulatory articles for retrieval, and a test split with 250 QA pairs. Each question targets a specific procedural decision a researcher or administrator would face when conducting government-funded research in Korea — e.g., personnel cost calculation, equipment procurement, travel expense reimbursement, indirect-cost rules. Unlike conventional multi-hop or legal QA, every question requires structured procedural lookups and evidence-set closure across a multi-tiered authority structure (statute → decree → rule → notice → manual), with ground-truth citations attached at the article and clause level. The dataset includes difficulty levels (L1: single-clause factual lookup; L2: multi-facet reasoning within a single document; L3: cross-document reference traversal; L4: conditional/exception-paired reasoning with procedural chains), question types (e.g., multi_facet, single_clause, general_principle), and an institution-specific subset (e.g., university, government-funded). Data is sourced from the Korean National Law Information Center and IITP FAQ documents, primarily in Korean with some English translations.




