Docmatix - a huge dataset for Document Visual Question Answering

Hugging Face releases Docmatix, a large-scale dataset for Document Visual Question Answering (DocVQA), 240 times larger than previous datasets.

• It is generated from PDFA OCR dataset using Phi-3-small model for Q/A pair creation, with filtering for hallucinations.
• The dataset contains 2.4 million images and 9.5 million Q/A pairs from 1.3 million PDFs.
• Ablation studies using Florence-2 model show a 20% improvement in DocVQA performance.