DICES (Diversity In Conversational AI Evaluation for Safety)
收藏arXiv2023-06-20 更新2024-06-21 收录
下载链接：
https://github.com/google-research-datasets/dices-dataset
下载链接
链接失效反馈官方服务：
资源简介：
DICES数据集由谷歌研究创建，旨在评估对话AI系统的安全性，特别关注不同用户群体的主观安全观点。数据集包含1340个对话，每个对话由多个评价者进行安全评分，涵盖了性别、年龄和种族/民族等多个维度的多样性。创建过程中，数据集设计了详细的评价任务和专家注释，以确保评价的准确性和多样性。DICES数据集的应用领域包括对话AI系统的安全评估和模型训练，旨在解决如何整合和尊重多样性视角的问题。
The DICES dataset was created by Google Research to evaluate the safety of conversational AI systems, with a particular focus on the subjective safety perspectives of diverse user groups. It contains 1,340 conversations, each of which is safety-scored by multiple annotators, covering diversity across multiple dimensions including gender, age, and race/ethnicity. During its development, detailed annotation tasks and expert annotations were designed for the dataset to ensure the accuracy and diversity of the evaluations. The application scenarios of the DICES dataset include safety assessment of conversational AI systems and model training, aiming to address the issue of how to integrate and respect diverse perspectives.
提供机构：
谷歌研究
创建时间：
2023-06-20
