AI BriefingHuggingFace Releases Docmatix, a Large-Scale Dataset for DocVQA
Key point
HuggingFace has released Docmatix, a DocVQA dataset that is 240 times larger than existing ones, containing 2.4 million images and 9.5 million Q/A pairs.
Details
HuggingFace has released Docmatix, a large-scale dataset for Document Visual Question Answering (DocVQA). This dataset includes 2.4 million images and 9.5 million question-answer (Q/A) pairs extracted from 1.3 million PDF documents, making it approximately 240 times larger than existing dataset sizes.
Docmatix generated Q/A pairs using the Phi-3-small model based on text from the PDFA dataset. To improve data quality, regular expressions and specific keyword filtering were applied, removing 15% of the data that showed signs of hallucination.
Performance tests showed that fine-tuning the Florence-2 model using Docmatix improved DocVQA performance by approximately 20%. Notably, the 0.7B parameter Florence-2 model, after training on Docmatix, showed results approaching the performance of the 8B-scale Idefics2 model, demonstrating the powerful impact of data scale on model performance.
