From Pixels to Structure: Lightweight Vision-Language Models for Document OCR and Structured JSON Extraction
The study compared eight lightweight Vision-Language Models (up to 7B parameters) for converting document images into structured JSON data across three university heritage collections. Researchers tested zero-shot, few-shot, and fine-tuning methods. The findings show that instruction-tuned models like Gemma-3 are strong candidates for private, high-performing OCR and structure extraction alternatives.