Repository navigation
Quality-only public release of the Clouda OCR model-selection benchmark: a frozen 462-page Arabic document OCR evaluation of existing OCR/VLM models (10 registered candidates; 5 complete runs, 3 partial runs classified EARLY_STOPPED_NONCOMPETITIVE and unranked, 2 failed candidates — all disclosed with evidence). Publishes the canonical manifest, ground truth, per-page records, aggregate reports, pinned model registry, the exact scorer/runner code, and full provenance/licensing records. Runtime/speed measurements are out of scope for this release and are not published. Verify with: sha256sum -c CHECKSUMS.txt