From the source
Using Direct Preference Optimization (DPO) on OCR models to reduce text degeneration, using rejection pairs from the model's own failures.
It reports an average reduction of 59.4% in degeneration rate across model families.
From the source
A second training stage - applied after supervised fine-tuning (SFT), on the same documents, using the same model - reduced text degeneration in every family tested. No exceptions. Average reduction: 59.4%.
huggingface.co
