DocDownstream-1.0

DocDownstream-1.0整合了10个文档理解数据集，总计57w训练样例，包括DocVQA, InfographicsVQA, DeepForm, KleisterCharity, WikiTableQuestions, TabFact, ChartQA, TextCaps, TextVQA以及VisualMRC，涵盖文档、网页截图、表格、图表以及自然图。其任务包括信息抽取、视觉问答、自然语言推理以及图片描述，DocDownstream-1.0将所有任务统一为问答的形式。DocDownstream-1.0 integrates 10 document understanding datasets with a total of 570,000 training examples. The covered datasets include DocVQA, InfographicsVQA, DeepForm, KleisterCharity, WikiTableQuestions, TabFact, ChartQA, TextCaps, TextVQA, and VisualMRC. It covers various data modalities such as documents, web page screenshots, tables, charts, and natural images. Its supported tasks involve information extraction, visual question answering, natural language inference, and image captioning. DocDownstream-1.0 unifies all these tasks into the question answering format.

搜集汇总

背景与挑战

背景概述

DocDownstream-1.0是一个文档图片理解数据集，整合了10个子数据集，涵盖信息抽取、视觉问答等多种任务，并统一为问答形式。该数据集包含约57万个训练样例，图片总大小约70GB，组织结构包括训练集、验证集和测试集等目录。