Newer Models, Same Advantage

Three months after publishing a paper on DharmaOCR and open-sourcing one of its models, the objective was clear: create optical character recognition (OCR) specifically for Brazilian Portuguese. The training pipeline consisted of two stages. The first, supervised fine-tuning, utilized a wide range of Portuguese-language files to align the model's weights with the target language's vocabulary, syntax, and document structures.

The second stage, Direct Preference Optimization (DPO), learned from comparative preference data between competing outputs, improving stability and reducing inference time and cost.

The combined result was a model with the highest extraction quality score and the lowest degeneration rate on a Portuguese-focused benchmark. Despite rapid advancements in OCR models, the gaps that initially motivated DharmaOCR's design—extraction quality on complex documents and stability under production conditions—have not closed but have become more evident as the field has changed.

New OCR models, Mistral OCR4 and Unlimited-OCR, have emerged, boasting new training techniques, datasets, and strong results across multiple languages. However, when tested against the DharmaOCR benchmark, specifically designed for Portuguese, DharmaOCR outperforms the newer models significantly. DharmaOCR scored 0.925, while Mistral OCR4 scored 0.798 and Unlimited-OCR scored 0.7587.

This substantial gap highlights the inherent advantage of DharmaOCR's specialization on the Brazilian Portuguese domain, demonstrating that concentrated resources dedicated to a specific language yield superior results compared to broader, multilingual models.