Newer Models, Same Advantage

A specialized Optical Character Recognition model, DharmaOCR, outperforms newer general-purpose models like Mistral OCR4 on Brazilian Portuguese documents. Its advantage stems from a two-stage training process that includes supervised fine-tuning on a Portuguese-specific dataset and Direct Preference Optimization (DPO) for stability. This domain specialization allows the model's parameters to be entirely dedicated to the nuances of a single language, creating a structural performance advantage. Benchmark results on complex documents like ENEM essays confirm DharmaOCR's superior extraction quality and lower error rates compared to more broadly trained, multilingual models.