GuardrailsAI/hallucination
收藏官方服务：
资源简介：
BUMP（不可靠最小对基准）数据集是一个用于评估幻觉检测系统的优质数据集，它通过人工方式在CNN/Daily Mail文章摘要中引入单个错误，创建了889对摘要，每对包含一个参考摘要和一个不可靠的版本。错误分为谓词、实体和情境等七类，并区分为内在错误和外在错误。这个数据集对于诊断AI系统中的特定失败模式提供了更细致的理解。
The BUMP (Benchmark of Unfaithful Minimal Pairs) dataset is a superior choice for evaluating hallucination detection systems, consisting of 889 summary pairs with a single introduced error in each, manually annotated from CNN/Daily Mail articles. The errors are categorized into seven types, including predicate, entity, and circumstance errors, classified as intrinsic or extrinsic. This dataset offers a nuanced understanding of various hallucination types for diagnosing specific failure modes in AI systems.
提供机构：
GuardrailsAI
