HeOCR R4d Hebrew document segmentation model
Copyright (c) 2026 HeOCR maintainers

This model (heocr-r4d-segmentation.safetensors, trained as
best_0.8618.safetensors, sha256
d32cd1a8efe7354445ccfea76551cc9943ca05792c332b5865f9bec985fe7bf7) is
licensed under the Apache License, Version 2.0. See LICENSE.

It was trained from scratch. No third-party model weights were loaded.

Software
--------

The model was trained with, and runs in, kraken / ketos 7.0.2
(https://github.com/mittagessen/kraken), licensed under the Apache License,
Version 2.0. The model file uses kraken's safetensors format and records
ketos's default segmentation network definition (VGSL). It contains no kraken
code.

Training data
-------------

The training data is synthetic: page images rendered by hocrsyngen
(https://github.com/HeOCR/hocrsyngen, MIT License).

Fonts. The page images were rendered in these two fonts. The model contains
no font data, and the fonts are not distributed with it. They are credited
here:

- Alef, Copyright (c) 2012, HaGilda & Mushon Zer-Aviv, with Reserved Font
  Name Alef. Licensed under the SIL Open Font License, Version 1.1.
- Gveret Levin, Copyright 2024 The Gveret Levin Project Authors. Licensed
  under the SIL Open Font License, Version 1.1.

Text. The rendered text is original text authored by AI agents (Claude Opus
5.5 sessions run in Claude Code) for HeOCR. It is the hocrsyngen corpus
S10b. No third-party text is included. The corpus is
dedicated to the public domain under CC0 1.0 Universal (see hocrsyngen
src/hocrsyngen/data/synthetic/texts/LICENSE-CC0-1.0.txt and hocrsyngen
ADR 0008).