Have you folks tried finetuning models for data extraction from visual data?