ReadBench

该数据集是一个专为评估视觉-语言模型（VLMs）的阅读理解能力而设计的多模态基准测试。它通过将已有纯文本基准测试中的上下文转换成包含文本的图像来实现这一目的。ReadBench特别突出了当处理来自图像的文本信息时，与处理纯文本输入相比，VLMs性能的下降，强调了输入长度和任务难度对模型性能的影响。该数据集的任务是评估VLMs在富含文本的图像上的阅读理解能力。

This dataset is a multimodal benchmark specifically designed to evaluate the reading comprehension capabilities of Vision-Language Models (VLMs). It achieves this goal by converting the contexts from existing plain-text benchmarks into text-containing images. ReadBench specifically highlights the performance degradation of VLMs when processing text information extracted from images, compared with processing plain-text inputs, and emphasizes the impact of input length and task difficulty on model performance. The task of this dataset is to evaluate the reading comprehension capabilities of VLMs on text-rich images.