DharmaOCR outperforms newer models on Brazilian Portuguese via specialization

DharmaOCR achieves higher extraction quality and stability than Mistral OCR4 and Unlimited-OCR on Brazilian Portuguese through domain-specific training.

• DharmaOCR scored 0.925 on a Portuguese-focused benchmark, compared to 0.798 for Mistral OCR4 and 0.7587 for Unlimited-OCR.
• The model uses supervised fine-tuning to align with Brazilian Portuguese vocabulary and Direct Preference Optimization (DPO) to reduce output degeneration.
• Multilingual models exhibit systematic errors on proper nouns like "Chico Buarque" and produce incoherent text on complex documents due to parameter distribution across languages.
• DharmaOCR's specialized training concentrates all parameters on the target domain, ensuring consistent performance where broader models fail.

Specialization provides a measurable advantage over newer architectures by dedicating finite resources to a single language rather than distributing them across many.