DharmaOCR: Specialization Advantage in Brazilian Portuguese OCR

DharmaOCR outperforms newer generalist models like Mistral OCR4 and Unlimited-OCR on Brazilian Portuguese documents by concentrating all model parameters on a single domain through targeted fine-tuning and Direct Preference Optimization (DPO).

Domain Specialization vs. Multilingual Breadth

Specialization provides a structural advantage over generalist models because it optimizes the allocation of finite model parameters. While multilingual models must distribute their representational capacity across N languages, a specialized model like DharmaOCR dedicates all its parameters to the specific vocabulary, morphology, and orthographic patterns of a single target language.

This concentration of resources allows the model to capture linguistic nuances that generalist models often miss. For example, in benchmarks using ENEM essays (Brazil's national high school examination), generalist models failed to correctly transcribe nationally recognized Brazilian figures and phrases, such as "Chico Buarque," while DharmaOCR handled them correctly. This demonstrates that generalist models do not fail randomly, but specifically at the vocabulary and proper nouns that distinguish a specific language from a broader multilingual corpus.

Performance Benchmarks

In a Portuguese-focused evaluation, DharmaOCR significantly outperformed newer, better-resourced models released after its own debut:

Model	Extraction Quality Score
DharmaOCR	0.925
Mistral OCR4	0.798
Unlimited-OCR	0.7587

DharmaOCR maintains a lead of approximately 13 points over Mistral OCR4 and more than 16 points over Unlimited-OCR, proving that domain specialization remains effective even against more recent architectural advancements.

Addressing Text Degeneration with DPO

Text degeneration—where a model produces repetitive, incoherent, or semantically disconnected output—is a critical failure mode in production OCR, typically triggered by visual ambiguity such as small fonts or degraded scan quality.

DharmaOCR mitigates this through a two-stage training pipeline:

1. Supervised Fine-Tuning (SFT): This stage builds domain competency by aligning model weights to the vocabulary and document structures of Brazilian Portuguese.
2. Direct Preference Optimization (DPO): Unlike SFT, which predicts tokens step-by-step, DPO trains the model against the quality of complete outputs. By teaching the model to discriminate between competing responses based on the coherence of the full extraction, DPO suppresses the drift and repetition loops that cause degeneration.

This approach ensures that the model remains stable under conditions where generalist models, such as Mistral OCR4, may produce output entirely disconnected from the source document.

The Structural Logic of Specialization

While newer architectures and expanded datasets will likely raise the absolute performance ceiling for all models, the relative advantage of specialization persists. The structural dynamic remains the same: a system that directs all its resources toward a single domain will extract more value from those resources in that domain than a system distributing them across many.

DharmaOCR's strategy is to apply emerging frontier techniques—including new architectures and alignment methods—specifically to the fixed domain of Brazilian Portuguese OCR to maintain the highest possible extraction quality at the lowest possible cost and inference time.