bitlabsdb/BBQ_dataset

BBQ（问答偏见基准）是一个专门设计的问答数据集，用于评估大型语言模型中的隐含社会偏见。它包括约58,000个问答示例，覆盖了九个社会偏见维度，如年龄、性别认同、种族/民族等。数据集分为模糊上下文和明确上下文，以及正负对比对，用于检测模型在问答任务中是否受到刻板印象的影响。

The BBQ (Bias Benchmark for Question Answering) is a dataset specifically crafted to evaluate implicit social biases in large language models through question-answering tasks. It includes approximately 58,000 QA examples covering nine social bias dimensions such as age, gender identity, race/ethnicity, and more. The dataset is divided into ambiguous and disambiguated contexts, as well as positive-negative pairs, to detect whether stereotypes influence model predictions in QA tasks.