GlobalQA
收藏资源简介:
GlobalQA是一个用于评估检索增强生成(RAG)系统在全局范围内的聚合能力的基准数据集。该数据集包含超过13000个问答对,构建在超过2000个真实世界简历的语料库上,涵盖23个专业领域。GlobalQA定义了四种核心任务类型:计数、极值查询、排序和Top-k提取,这些任务全面评估了模型在统计、比较、排名和提取方面的能力。与现有基准不同,GlobalQA的查询需要遍历大量的文档集,确保模型能够进行真正的全局推理。数据集的构建采用反向策略,通过程序设计查询轨迹,然后由代理执行这些轨迹以获得确定性答案,最后基于完成的轨迹生成自然语言问题。
GlobalQA is a benchmark dataset for evaluating the global-scale aggregation capabilities of Retrieval-Augmented Generation (RAG) systems. This dataset contains over 13,000 question-answer pairs, built upon a corpus of more than 2,000 real-world resumes, covering 23 professional domains. GlobalQA defines four core task types: counting, extremum query, sorting, and Top-k extraction, which comprehensively evaluate the model's capabilities in statistical processing, comparison, ranking, and information extraction. Unlike existing benchmarks, the queries in GlobalQA require traversing a large document collection, ensuring that the model can perform true global reasoning. The dataset is constructed using a reverse strategy: query trajectories are designed via programs, then agents execute these trajectories to obtain deterministic answers, and finally natural language questions are generated based on the completed trajectories.




