MultiTasks-v2
收藏资源简介:
MultiTasks-v2是一个多模态基准数据集集合,包含十二个经过标准化处理的子集,统一采用图像-问题-答案格式。该数据集旨在为多模态评估和训练实验提供统一的基准,涵盖视觉问答、文档理解、图表推理、几何问题求解、逻辑推理、视觉谜题、情感识别和指代表达式定位等多种任务。数据集中每个样本包含四个字段:唯一标识符(id,格式为{Dataset}_{split}_{index})、图像列表(images,存储为JPEG字节,每个样本包含一张图像)、问题提示(problem)和答案(answer)。除RefAdv子集使用列表形式的边界框坐标答案外,其他所有子集均使用字符串答案。数据集包含以下十二个子集:ChartQA、CVBench、DocVQA、DrivingVQA、EmoSet2k、Geometry3k、IconQA、LogicGraph、MathVista、OlympiadBench、PuzzleVQA和RefAdv。每个子集均提供训练集和测试集分割,样本数量分别为:ChartQA(1100/1100)、CVBench(1100/1100)、DocVQA(1100/1100)、DrivingVQA(1100/1100)、EmoSet2k(1100/1100)、Geometry3k(1100/1100)、IconQA(1100/1100)、LogicGraph(450/450)、MathVista(500/500)、OlympiadBench(1100/1100)、PuzzleVQA(1000/1000)、RefAdv(571/571)。数据集对多选题进行了标准化处理,统一采用选项字母格式,答案仅包含选项字母(如A、B或A,C)。开放式任务则保留简短的文本答案。该数据集是对现有基准源的标准化聚合,适用于多模态模型的综合性能评估。
MultiTasks-v2 is a multimodal benchmark dataset collection comprising twelve standardized subsets, uniformly adopting an image-question-answer format. This dataset aims to provide a unified benchmark for multimodal evaluation and training experiments, covering various tasks such as visual question answering, document understanding, chart reasoning, geometric problem solving, logical reasoning, visual puzzles, emotion recognition, and referring expression localization. Each sample in the dataset includes four fields: a unique identifier (id, formatted as {Dataset}_{split}_{index}), a list of images (stored as JPEG bytes, with each sample containing one image), a problem prompt, and an answer. Except for the RefAdv subset, which uses list-form bounding box coordinate answers, all other subsets use string answers. The dataset includes the following twelve subsets: ChartQA, CVBench, DocVQA, DrivingVQA, EmoSet2k, Geometry3k, IconQA, LogicGraph, MathVista, OlympiadBench, PuzzleVQA, and RefAdv. Each subset provides both training and test set splits, with sample counts as follows: ChartQA (1100/1100), CVBench (1100/1100), DocVQA (1100/1100), DrivingVQA (1100/1100), EmoSet2k (1100/1100), Geometry3k (1100/1100), IconQA (1100/1100), LogicGraph (450/450), MathVista (500/500), OlympiadBench (1100/1100), PuzzleVQA (1000/1000), and RefAdv (571/571). The dataset standardizes multiple-choice questions by adopting an option letter format, where answers only include option letters (e.g., A, B, or A,C). Open-ended tasks retain short text answers. This dataset is a standardized aggregation of existing benchmark sources, suitable for comprehensive performance evaluation of multimodal models.
数据集概述
MultiTasks-v2 是一个多模态基准测试子集的集合,包含12个子集,所有数据均被标准化为统一的 图像-问题-答案 格式。
数据集结构
- 每个样本 包含以下字段:
id:唯一样本标识符,格式为{Dataset}_{split}_{index}。images:一个图像列表,包含一张JPEG格式的图像。problem:输入给模型的提示(问题)。answer:目标答案。
- 答案类型:除
RefAdv子集的答案为列表(边界框坐标)外,其余子集均为字符串答案。
子集详情
| 子集名称 | 训练集规模 | 测试集规模 | 答案类型 |
|---|---|---|---|
| ChartQA | 1100 | 1100 | string |
| CVBench | 1100 | 1100 | string |
| DocVQA | 1100 | 1100 | string |
| DrivingVQA | 1100 | 1100 | string |
| EmoSet2k | 1100 | 1100 | string |
| Geometry3k | 1100 | 1100 | string |
| IconQA | 1100 | 1100 | string |
| LogicGraph | 450 | 450 | string |
| MathVista | 500 | 500 | string |
| OlympiadBench | 1100 | 1100 | string |
| PuzzleVQA | 1000 | 1000 | string |
| RefAdv | 571 | 571 | list<int64> |
提示标准化
-
多项选择题 被标准化为统一格式: text <image> Question: ... Options: A. ... B. ... C. ... Answer with only the option letter.
答案仅包含选项字母(如
A、B或A,C)。 -
开放题 保留简短文本答案。
-
RefAdv 答案为边界框坐标,格式为
[x1, y1, x2, y2]。
用途与说明
- 应用场景:适用于多模态评估与训练,涵盖视觉问答、文档理解、图表推理、几何、逻辑推理、视觉谜题、情绪识别和指代表达定位。
- 数据来源:本数据集是对现有基准来源的标准化整合,原始许可证、标注和文档请参考各原始数据集。




