BBQ

该数据集名为BBQ，被广泛应用于量化语言模型中的外在偏见。每个数据点包含一个上下文、一个指向该上下文的问题以及一个依赖于上下文和问题的答案。该数据集涵盖了不同偏见类别的示例，并且在不同偏见类别（'biased'和'not biased'）之间保持了平衡，这有助于评估语言模型中的偏见缓解策略。该数据集共包含58,492个示例，其任务是识别语言模型中的偏见。

This dataset, named BBQ, is widely employed to quantify extrinsic biases in language models. Each data point comprises a context, a question pertaining to the context, and an answer that depends on both the context and the question. The dataset includes examples spanning various bias categories, and maintains a balanced distribution between the 'biased' and 'not biased' categories, which supports the assessment of bias mitigation strategies for language models. With a total of 58,492 examples, this dataset is designed for the task of identifying biases in language models.