Hugging Face: Docmatix DocVQA dataset (2.4M images) yields ~20% gain for Florence-2 fine-tuning | SignalBreak | SignalBreak