TrueDoc
Copyright 2026 islandtimer

Licensed under the Apache License, Version 2.0 (the LICENSE file at the root of this repository).

This repository also contains, or downloads at run time, work by others:

bench/olmocr_ref/
    The olmOCR-bench scorer (benchmark.py, tests.py, render.py) from the allenai/olmocr repository on
    GitHub; Copyright the Allen Institute for Artificial Intelligence, Apache License 2.0
    (bench/olmocr_ref/LICENSE). Every score in docs/BENCHMARKS.md is produced by running this copy of
    benchmark.py over the dataset's own test files: bench/score_olmocr.py rewrites its three relative
    imports to the installed `olmocr` package's modules (tests, report, utils; olmocr 0.4.27) and
    changes nothing else, so the test rules are the package's and the driver is this file.

bench/gpu/out*/
    Saved page readings produced by open-weight vision models on pages of the olmOCR-bench dataset
    (allenai/olmOCR-bench, ODC-BY) and, under out5/pro/own/, on pages of publicly available Australian
    insurance documents. The models: allenai/olmOCR-2-7B-1025, infly/Infinity-Parser2-Flash and
    infly/Infinity-Parser2-Pro, each Apache License 2.0. The readings are kept so that every benchmark
    run since run 55 can be reproduced without a GPU.

samples/
    Four product disclosure statements published by their insurers, kept as test documents; the
    insurers hold their copyright.

Model weights fetched at run time, none stored here:
    docling-project/docling-layout-heron (formerly ds4sd/docling-layout-heron), layout detection -
    Apache License 2.0.
    SWHL/RapidOCR, the English text recogniser used by rapidocr_onnxruntime - Apache License 2.0.

Python dependencies are listed in pyproject.toml with their licences; docs/DECISIONS.md (D007) records
the licence check of everything on the conversion path.
