Finetuning olmOCR to be a faithful OCR-Engine

Hugging Face Blog · 1 year ago

TNG fine-tuned olmOCR, an open-source optical character recognition model, to extract text from document headers and footers that the original version ignored. The team generated 8,000 training documents using Qwen2.5-VL-72B-Instruct and trained for 2.5 epochs on an 8xH100 GPU node. The modified model now reliably extracts complete information from invoices and other business documents where critical data appears outside the main text body.