Fujitsu/PDD-Extended-Bench

PDD-Extended-Bench是一个用于Token-Focused Disparity Probing（TFDP）评估的基准数据集，包含具有单词遮罩的最小对比句子对，用于审查自回归语言模型中的差异。该数据集包括两个主题组：气候错误信息和性别平等，每个主题组都包含遮罩和原始变体，用于单词预测探测。数据集的每条记录都暴露了真实声明和遮罩副本，以支持TFDP的单词查询和多尺度对齐度指标。PDD-Extended-Bench is a Token-Focused Disparity Probing (TFDP) evaluation benchmark consisting of minimally contrastive sentence pairs with single-token masks for auditing disparities in autoregressive LLMs. The dataset includes two topic groups: Climate Misinformation and Gender Equality, each with masked and original variants constructed for single-token prediction probes. Each record in the dataset exposes ground-truth statements and masked counterparts to support TFDP’s single-token querying and multi-scale alignment metrics.