CQ-Bench
收藏资源简介:
CQ-Bench是一个专门设计的基准,用于评估LLM在自然对话环境中推断隐性文化价值观的能力。该数据集通过结合世界价值观调查和GlobalOpinions数据集中的价值观,生成了多角色对话式故事。数据集构建过程中包含了严格的验证程序,最终验证中人类与模型的一致性达到98.2%。CQ-Bench包括三个任务:态度检测、价值选择和价值提取,旨在解决LLM在跨文化交流中遇到的问题。
CQ-Bench is a purpose-built benchmark designed to evaluate the capability of Large Language Models (LLMs) to infer implicit cultural values in natural conversational scenarios. This dataset generates multi-role conversational stories by integrating values from the World Values Survey and the GlobalOpinions dataset. Stringent validation procedures were implemented throughout the dataset construction process, with a human-model agreement rate of 98.2% achieved in the final validation stage. CQ-Bench comprises three tasks: attitude detection, value selection, and value extraction, which aim to address the challenges faced by LLMs in cross-cultural communication.




