Nate@nate_512

Follow

SoMark OXR OCR benchmark results

SoMark | OXR is flipping parsing errors into specific training data without needing manual labels. The RSI loop takes unlabeled documents, runs them through a VLM to spot disagreements, clusters the failure causes, generates targeted data, and retrains the model automatically. This 1.3B parameter setup actually tops olmOCR-Bench, OmniDocBench, and DocParsingBench. On a single RTX PRO 6000 GPU it crushed a 100-page OmniDocBench test in 101 seconds, hitting 2.52 times the throughput of MinerU2.5-Pro. It's described as the fastest open-source parser tested for Optical Everything Recognition. The project is free and open-source with links on GitHub, Hugging Face, ModelScope, and arXiv.

1h

Nate@nate_512

Follow

SoMark OXR OCR benchmark results

SoMark | OXR is flipping parsing errors into specific training data without needing manual labels. The RSI loop takes unlabeled documents, runs them through a VLM to spot disagreements, clusters the failure causes, generates targeted data, and retrains the model automatically. This 1.3B parameter setup actually tops olmOCR-Bench, OmniDocBench, and DocParsingBench. On a single RTX PRO 6000 GPU it crushed a 100-page OmniDocBench test in 101 seconds, hitting 2.52 times the throughput of MinerU2.5-Pro. It's described as the fastest open-source parser tested for Optical Everything Recognition. The project is free and open-source with links on GitHub, Hugging Face, ModelScope, and arXiv.

1h