groupfairnessllm/gender_bias_with_human_label

这是一个用于测量大型语言模型中性别相关偏见的人力标注数据集。数据集包含一个参考响应和两个候选响应，这些响应是对带有明确性别提示的问题的回复。标注者根据提供的标注模板判断哪个候选响应在语义上更接近参考响应。数据集旨在用于审计与性别相关的差异、基准测试语义相似度检查器、验证长格式输出的公平性指标、训练奖励模型和评估模型。

This is a human-labeled dataset for measuring gender-related biases in large language models. The dataset includes a reference response and two candidate responses to a question with explicit gender cues. Annotators determine which candidate response is semantically closer to the reference response using the provided labeling template. The dataset is designed for auditing gender-related disparities, benchmarking semantic similarity checkers, validating fairness metrics for long-form outputs, training reward models, and training evaluation models.