TrueDoc: a PDF-to-markdown pipeline, measured on olmOCR-bench

This repository holds a benchmark result, not model weights. TrueDoc converts PDFs to markdown with a YAML front matter (the Open Knowledge Format) and is built for meaning accuracy: does the output say what a reader of the page would understand it to say? Its code is at https://github.com/islandtimer/TrueDoc (Apache-2.0). The models it uses are other people's: Infinity-Parser2-Flash and -Pro (infly), olmOCR-2-7B-1025 (allenai), docling-layout-heron (docling-project), all Apache-2.0.

olmOCR-bench, run 97: 86.8 (95% bootstrap CI 86.0-87.7), official scorer, all eight sections, dataset revision 54a96a6f. RESULTS.md has the section scores, exactly what read which pages, and how to re-score the 1,403 outputs in outputs/ yourself.

The arrangement, in one line: the PDF's own text for the 1,122 pages that have it (no model), Infinity-Parser2-Flash for the 281 that do not, Infinity-Parser2-Pro behind it for the 102 the pipeline's own OCR cannot read, olmOCR 2 for 92 picture regions; readings recorded on a rented GPU and replayed; no post-processing keyed on the benchmark's folder names.