CausalGym

该数据集名为CausalGym，旨在研究Transformer模型中的因果效应特征，重点关注特征表示随规模和训练数据量的变化。数据集包含控制任务，这些任务用任意标记替换原始标签，同时保留类别划分，以研究选择性和赔率比。每个任务的规模为训练400个示例，评估100个示例。其中一项任务为性别一致性任务，还有其他类型的任务。

This dataset, named CausalGym, is intended to investigate causal effect characteristics within Transformer models, with a primary focus on the variations of feature representations alongside model scale and training data volume. The dataset encompasses controlled tasks, wherein original labels are substituted with arbitrary tokens while preserving category partitions, to explore selectivity and odds ratios. Each task comprises 400 training examples and 100 evaluation examples. Among these tasks is the gender consistency task, along with other task categories.