gxx27/time_unlearn

Time-Unlearn数据集用于评估元提示模拟的早期知识截断是否可以在评估时间预测任务时减少污染。该数据集包含三个子集：事实性子集包含随时间变化的事实，语义性子集包含最近出现或变化意义的单词，反事实性子集包含需要忽略截断后因果事件的问题。数据集共有1637条记录，分别分布在三个子集中。

The Time-Unlearn dataset evaluates whether meta-prompts simulating earlier knowledge cutoffs can reduce contamination when assessing temporal prediction tasks. The dataset consists of three subsets: the Factual subset includes direct facts that have changed over time, the Semantic subset includes words whose meanings have emerged or shifted recently, and the Counterfactual subset includes questions that require ignoring post-cutoff causal events. The dataset contains a total of 1,637 records, distributed across the three subsets.