Ambiguous Integer Sequences
收藏arXiv2025-09-30 收录
下载链接：
https://github.com/JacobPfau/introspective-self-consistency
下载链接
链接失效反馈官方服务：
资源简介：
该数据集包含了一系列模糊的整数序列，旨在评估语言模型在自我一致性方面的表现。该数据集的创新之处在于引入了具有歧义性的任务，以便衡量模型的自我一致性等特性，并通过枚举生成函数，严格定义了可能答案的范围。该数据集的规模包括140个无歧义的序列和57个有歧义的序列，其任务内容涉及序列的完成和解释。
This dataset comprises a set of ambiguous integer sequences, designed to evaluate the self-consistency performance of language models. The novelty of this dataset lies in the introduction of ambiguous tasks, which enable the measurement of model properties such as self-consistency, while the range of possible answers is strictly defined via enumerative generation functions. The dataset includes 140 unambiguous sequences and 57 ambiguous sequences, with its tasks covering sequence completion and explanation.
提供机构：
Open-sourced by the authors
