Note Results table in the card. 1,400 PDFs with 7,000 unit-test style checks (old scans, math, tables, headers). Pass rate per model, most-cited English document OCR bench. Updated Feb 2026.

Note Leaderboard in the card. Document parsing of real business documents into Markdown, scored on text, tables, layout. Updated Apr 2026.

Note Leaderboard in the card, not a Space. OmniDocBench v1.5 page parsing: text, tables, formulas, reading order across document types. Largest results table here, updated Sep 2026.

Note Historical printed books (Biodiversity Heritage Library). CER on human-corrected IMPACT ground truth, 15 models, updated Sep 2026. Static Space, results from an open uv-script harness.

Note Slovene. Gradio leaderboard over the slOCR benchmark, modern and historical print, CER/WER per model. Updated May 2026.

Note Icelandic. 100 manually transcribed pages with ALTO and PAGE layout ground truth, so line detection is scored too. Updated Sep 2026.

Note Tibetan. Line-to-text benchmark from OpenPecha, CER per model, scripts from woodblock to modern print. Updated Aug 2026.