iFANN
    iFANN'da ara...
    Giriş Yap
    Ana Sayfa
    Haberler
    Videolar
    Kitaplar
    Fotoğraflar
    GIF'ler
    Keşfet
    Anketler
    Ödüller
    iFAMOUS
    Viki
    Anime
    Odalar
    Bildirimler
    Mesajlar
    Yer İmleri
    Profil
    VikiÖdülleriFAMOUSSıralamalarSektörlerİçerik Üretici ÖdülleriKullanıcı ÖdülleriŞartlarGizlilikTopluluk KurallarıKaldırma / DMCAYardımGeliştiriciler

    © 2026 iFANN

    Ana Sayfa
    Ara
    Mesajlar
    Uyarılar
    Profil

    Gönderi

    Nate
    Nate@nate_512
    🏢Hugging Face📱GitHub💭AI

    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.

    0mo

    16 Beğeni0 Beğenmeme2 Yeniden paylaşım1 Yorumlar
    ?

    Yorumlar

    Henüz yorum yok. İlk yorumu sen yap!

    Gönderi

    Nate
    Nate@nate_512
    🏢Hugging Face📱GitHub💭AI

    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.

    0mo

    16 Beğeni0 Beğenmeme2 Yeniden paylaşım1 Yorumlar
    ?

    Yorumlar

    Henüz yorum yok. İlk yorumu sen yap!