Domain-specialized OCR model outperforms newer multilingual rivals on Brazilian Portuguese benchmark

A team from Dharma-AI reports that DharmaOCR, a Brazilian Portuguese–specialized optical character recognition model, achieved the highest extraction quality score and the lowest degeneration rate on a Portuguese-focused benchmark compared with newer multilingual OCR systems Mistral OCR4 and Unlimited-OCR. The authors attribute the advantage to a two-stage training pipeline: supervised fine-tuning on Portuguese-language data followed by Direct Preference Optimization (DPO) to improve output stability under production conditions.

The benchmark results show DharmaOCR at 0.925 versus Mistral OCR4 at 0.798 and Unlimited-OCR at 0.7587, a gap of approximately 13 and 16+ points respectively. The evaluation was designed exclusively around Portuguese documents, including ENEM essays that combine handwritten text with Brazilian Portuguese–specific vocabulary, proper nouns, and cultural references.

The authors argue that specialization concentrates model capacity on a single language, enabling parameters to encode language-specific features more effectively than multilingual models that distribute capacity across many languages. They note that newer architectures and training techniques improve what any model can do, but the location of those improvements—focused on one domain versus spread across many—determines the practical advantage in extraction quality and stability.

The post also highlights failure modes observed in the multilingual models when processing Brazilian Portuguese text, such as misrendering the name "Chico Buarque" as "Chico Barque" and producing garbled phrases like "a dose de chico bique, 'o Brasil no exclu, eliminila.'" These errors illustrate how insufficient exposure to language-specific vocabulary and proper nouns can systematically degrade performance on domain-specific documents.