TNG fine-tunes olmOCR to extract headers and footers

AI Impact Summary

TNG created a fine-tuned version of olmOCR-7B-0225-preview to address olmOCR's limitation of ignoring header and footer information, a critical issue for applications like invoice parsing. This fine-tuning, leveraging Qwen2.5-VL-72B-Instruct to generate a comprehensive training dataset, enables the model to extract all relevant information, including crucial data in headers and footers, improving its utility for business workflows. The team utilized a 8xH100 Nvidia node with 2.5 epochs to train the model, demonstrating a practical approach to adapting a pre-trained model for a specific use case.