Anti-Commonsense Benchmark (𝒟₃)

该数据集是一个基准测试，其中变量名称被选定以反映与图宾根成对数据集中发现的因果关系相反的情况。该基准专注于测试大型语言模型在处理因果关系违反直觉的干预情况时的应对能力。在所有场景中，涵盖了22种干预效应。任务的目的是研究在干预情况下，大型语言模型对可能记忆中的因果关系的鲁棒性。

This dataset is a benchmark test where variable names are selected to reflect the reverse of the causal relationships identified in the Tübingen Pairwise Dataset. This benchmark focuses on testing the ability of large language models (LLMs) to handle counterintuitive causal intervention scenarios. Twenty-two distinct intervention effects are covered across all scenarios. The objective of this task is to examine the robustness of large language models against potentially memorized causal relationships under intervention conditions.