A Multimodal Large-Scale Dataset for Visual Question Answering on Electronic Documents

A large-scale, high-quality visual question answering dataset for electronic documents, integrating both textual and visual information from document images to generate questions and corresponding answers. The dataset contains 324,546 images and 2,036,263 QA pairs, with quality ensured through quantity and format validation, data filtering, and consistency checks. Random sampling indicates an accuracy of 91.34% for the QA pairs. In application, the dataset significantly improves the performance of multimodal large language models on document-based QA tasks: when fine-tuned on LLaVA-OV and Deepseek-VL, the average normalized edit similarity on the DocVQA benchmark increases by 1.4% and 2.6%, respectively. Ablation studies show that removing the data filtering step decreases performance by 1.3%, while complementary experiments with human-annotated data demonstrate a 1.3% improvement when partially combined with this dataset. Compared with datasets generated by existing methods, this dataset yields the most notable performance gains, while still leaving room for further optimization.