atamiles/VLURes

VLURes是一个多语言基准数据集，用于评估视觉语言模型（VLMs）在长文本设置中的细粒度视觉和语言理解能力。该数据集旨在超越短标题、以英语为中心的评估，转而测试图像理解、长上下文接地以及在文化多样化设置中的跨语言鲁棒性。数据集包含英语、日语、斯瓦希里语和乌尔都语的图像-文本对，每个示例都包含一个重命名的图像文件、其配对的长期文本以及一个语言标识符。数据集支持多种任务，包括对象识别、场景理解、关系理解、语义分割、图像字幕、图像-文本匹配、无关性和视觉问答。数据集的结构、数据格式、语言代码和分割都有详细说明。

VLURes is a multilingual benchmark for evaluating the fine-grained visual and linguistic understanding of Vision-Language Models (VLMs) in long-text settings. It was created to move beyond short-caption, English-centric evaluation and instead test image understanding, long-context grounding, and cross-lingual robustness in culturally diverse settings. The dataset covers English, Japanese, Swahili, and Urdu, with each example consisting of a renamed image file, its paired long-form text, and a language identifier. It supports various tasks such as Object Recognition, Scene Understanding, Relationship Understanding, Semantic Segmentation, Image Captioning, Image-Text Matching, Unrelatedness, and Visual Question Answering. The dataset structure, data format, language codes, and splits are clearly outlined.