mit-oasys/oolong-pairs
收藏资源简介:
Oolong-Pairs是一个基于长上下文的成对聚合推理基准,构建在oolongbench/oolong-synth数据集之上。每个任务提供一个包含数千个一般知识问题的长上下文(每行一个问题,每个问题附有一个非唯一的用户ID)。每行可以隐含地标记为六个TREC粗粒度类别之一——描述和抽象概念、实体、人类、数值、位置、缩写——模型必须计算用户对的精确聚合统计信息(例如,列出所有至少有一个数值或位置实例的用户对)。标签从未给出;必须从每行的语义中推断。
Oolong-Pairs is a long-context, pairwise-aggregation reasoning benchmark built on top of the oolongbench/oolong-synth dataset. Each task presents a long context of thousands of general-knowledge questions (one per line, each attached to a non-unique User ID). Every line can be implicitly labelled with one of six TREC coarse categories — description and abstract concept, entity, human being, numeric value, location, abbreviation — and the model must compute exact aggregate statistics over pairs of users (e.g. list all pairs of users who both have at least one numeric-value or location instance). The labels are never given; they must be inferred from the semantics of each line.





