Fine-tuned olmOCR achieves faithful OCR with header/footer extraction for invoices

AI Impact Summary

The team has retrained olmOCR-7B-0225-preview into a faithful OCR engine that preserves header and footer content, addressing a critical gap for business documents like invoices. They built a training pipeline using 8k documents generated with Qwen2.5-VL-72B-Instruct and evaluated against an extended olmOCR-mix-0225 dataset that includes headers/footers, preserving raw text blocks and position data via document anchoring. The approach uses the open-source olmOCR workflow on multi-NVIDIA hardware (8x H100) and results in an OCR output that can capture information previously omitted, enabling richer downstream processing for structured data extraction. Expect output quality to vary with temperature settings and document layout complexity; plan to validate across representative invoices and similar layout-rich documents before production rollout.