MERA-evaluation/ruVQA

ruVQA是一个用于俄语的公开视觉问答（VQA）数据集，包含真实照片和抽象插图两种类型的图像。问题分为简单和复杂两类，涵盖了多种类型如二元、比较、数量、地点、方式、哪个、什么、谁等。简单问题仅需基于图像的感知，而复杂问题则需要推理步骤。数据集中的图像均来自公开资源，包括真实照片和卡通抽象图像。数据集旨在测试模型在不同类型图像中区分对象、理解不同问题类型并基于图像生成简短答案的能力。

ruVQA is a public question-answering dataset in Russian for two types of images: real photos and abstract illustrations. The questions are divided into simple and complex types, covering binary, comparative, how many, where, how, which, what, and who. Simple questions require only image-based perception, while complex ones require a reasoning step. All images are sourced from public resources, including real photos and cartoonish abstract images. The dataset is designed to test the models basic capabilities to distinguish objects in various types of images, understand different question types, and generate short answers based on the image. Questions cover key abilities such as scene understanding, physical property understanding, object function understanding, identity & emotion understanding, mathematical reasoning, static counting, common everyday knowledge, spatial object relationship, object-object interaction, object localization, object recognition, living things motion, object motion, human body pose recognition. The dataset is created using images from the English VQA v2 dataset and COCO dataset, with questions and answers generated from scratch by annotators using the ABC Elementary platform, and both automatic and manual filtering applied. Evaluation metrics include Exact match.