MFAVA
收藏资源简介:
MFAVA数据集是由Würzburg大学CAIDAS实验室创建的,旨在评估大型语言模型在不同语言中产生幻觉的情况。该数据集涵盖了30种语言,包含由LLM生成的提示和对应的维基百科文章作为参考。数据集的构建过程包括将英语FAVA数据集翻译成其他语言,并在此基础上进行人工标注和LLM生成的合成数据。MFAVA数据集的应用领域主要是知识密集型的长篇问答,以解决现实世界中LLM的使用问题。
The MFAVA dataset was created by the CAIDAS Lab at the University of Würzburg, with the objective of evaluating large language models (LLMs) for hallucinatory generation across different languages. This dataset covers 30 languages, and includes prompts generated by LLMs and corresponding Wikipedia articles as reference materials. The construction process of the MFAVA dataset involves translating the English FAVA dataset into other languages, followed by manual annotation and synthetic data generated by LLMs. The primary application areas of the MFAVA dataset are knowledge-intensive long-form question answering, aimed at addressing real-world issues related to LLM deployment and usage.




