giskardai/phare

Phare是一个多语种的大型语言模型安全性评估数据集，包含幻觉、偏见与刻板印象、有害内容和提示注入等多个类别的脆弱性评估。数据集涵盖英语、法语和西班牙语，旨在检测模型在不同情境下的不当行为，并经过多语言和多元文化的数据收集与处理。

Phare is a multilingual large language model safety evaluation dataset that includes assessments of vulnerabilities across categories such as hallucination, biases and stereotypes, harmful content, and prompt injection. The dataset covers English, French, and Spanish, aiming to detect inappropriate behavior of models in various situations, and has undergone multilingual and multicultural data collection and processing.