MM-Hallu/OpenCHAIR

OpenCHAIR是一个用于评估图像字幕模型对象幻觉的数据集。它包含与描述性字幕配对的图像，用于评估视觉语言模型是否会在图像中幻觉出不存在的对象。数据集结构包括image（输入图像）和text（图像的描述性字幕）。测试集包含4,863个示例。该数据集是从moranyanuka/OpenCHAIR转换而来，用于MM-Hallu基准测试集合。

OpenCHAIR is an object hallucination evaluation dataset for image captioning models. It contains images paired with descriptive captions and is used to assess whether vision-language models hallucinate objects that are not present in the images. The dataset structure includes image (the input image) and text (the descriptive caption for the image). The dataset contains 4,863 examples in the test split. This dataset was converted from moranyanuka/OpenCHAIR for the MM-Hallu benchmark collection.