LoCoV1
收藏资源简介:
LoCoV1是由斯坦福大学创建的一个新型基准数据集,包含12个任务,旨在评估长上下文检索模型的性能。该数据集涵盖法律、医学、科学、财务、公司治理、政府报告等多个领域,数据来源于Tau Scrolls、QASPER、LongBench和Legal Case Reports等真实世界数据集。LoCoV1的特点是需要进行长文本序列的信息综合,传统的截断和分块基线在此数据集上表现不佳。
LoCoV1 is a novel benchmark dataset created by Stanford University, consisting of 12 tasks designed to evaluate the performance of long-context retrieval models. This dataset covers multiple domains including law, medicine, science, finance, corporate governance and government reports, with its data sourced from real-world datasets such as Tau Scrolls, QASPER, LongBench and Legal Case Reports. A key characteristic of LoCoV1 is that it requires information synthesis over long text sequences, where traditional truncation and chunking baselines perform poorly.




