meituan-longcat/LoHoSearch
收藏资源简介:
LoHoSearch(长视野搜索代理)是一个具有挑战性的基准测试数据集,用于评估搜索代理在长视野推理和上下文管理方面的能力。它包含544个人工验证的问题,涵盖11个领域,基于覆盖超过700万维基百科实体的知识图谱构建。该数据集通过自动化流程生成,选择具有大搜索空间的关系,并将其组合成结构复杂的问题,确保每个问题有唯一的答案。数据集分为两个子集:benchmark(测试集,包含544条记录,经过人工验证)和train(训练集,包含2000条记录,由相同自动化流程生成但未经人工验证)。所有数据均为英文。
LoHoSearch (Long-Horizon Search Agents) is a challenging benchmark for evaluating long-horizon reasoning and context management in search agents. It comprises 544 human-verified questions across 11 domains, constructed via an automated pipeline built upon a knowledge graph covering over 7 million Wikipedia entities. The pipeline selects relations with large search spaces and assembles them into structurally complex questions with KG-verified unique answers. The dataset includes two subsets: benchmark (test set with 544 records, human-verified) and train (training set with 2000 records, generated by the same automated pipeline without human verification). All data is in English.




