DharmaOCR Uses DPO to Reduce Text Degeneration

Hugging Face's DharmaOCR has applied Direct Preference Optimization (DPO) to tackle text degeneration in OCR tasks. This method uses the model's own degenerate outputs as negative training signals, achieving an average degeneration reduction of 59.4%. Unlike supervised fine-tuning, which doesn't directly address degeneration, DPO targets the structural failure modes of models. This innovative approach shows promise for improving model performance in structured tasks, offering a new direction for AI training methodologies.