Repository navigation
attachments 1.0.0b8
Pre-release
Pre-release
Changed
- PDF text is read with PyMuPDF (already in the
pdfextra), in the
order the PDF stores it, with ligatures expanded ("fluid" is "fluid");
pypdf when PyMuPDF is missing or cannot open a file
(meta.extra.text_backendsays which). On olmOCR-bench: multi-column
pages 50.0 -> 58.1, small print 68.8 -> 73.8; words run together on
0.5% of text pages instead of 1.3%. - Tables in PDFs become Markdown tables where they stand (tables with
ruling lines, found by PyMuPDF), their text not repeated; merged cells
repeat their value. olmOCR-bench tables: 0.2 -> 31.2; overall 34.4 ->
39.7 (evals/compare).meta.extra.tables
counts them;[tables: false]turns it off. - Text pages take about 70 ms instead of 36 (median), the table finder's
cost; with[tables: false], 5 ms.