paddleocr-hebrew 0.1.0__tar.gz

paddleocr_hebrew-0.1.0/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or Derivative
+          Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work, excluding
+          those notices that do not pertain to any part of the Derivative
+          Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and do
+          not modify the License. You may add Your own attribution notices
+          within Derivative Works that You distribute, alongside or as an
+          addendum to the NOTICE text from the Work, provided that such
+          additional attribution notices cannot be construed as modifying
+          the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 paddleocr-hebrew contributors
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

paddleocr_hebrew-0.1.0/NOTICE

@@ -0,0 +1,18 @@
+paddleocr-hebrew
+Copyright 2026 paddleocr-hebrew contributors
+
+This product includes model weights that were fine-tuned from pre-trained
+models released by the PaddleOCR project (https://github.com/PaddlePaddle/PaddleOCR),
+which are licensed under the Apache License, Version 2.0.
+
+Base models the recognizers/detectors were derived from:
+  - PP-OCRv5 mobile/server detection (DBNet, PPLCNetV3)
+  - PP-OCRv5 server/mobile recognition (PPHGNetV2-B4, PPLCNetV3)
+  - PP-OCRv6 medium recognition (PPLCNetV4)
+  - SVTRv2 recognition (SVTR_HGNet, CTC + NRTR MultiHead)
+
+The Hebrew fine-tunes, the 120-character Hebrew charset, the inference/packaging
+code in this repository, and the split-ONNX NRTR host-loop deployment are
+original work of the paddleocr-hebrew contributors, released under Apache-2.0.
+
+PaddleOCR is Copyright (c) PaddlePaddle Authors, Apache License 2.0.

paddleocr_hebrew-0.1.0/PKG-INFO

@@ -0,0 +1,356 @@
+Metadata-Version: 2.4
+Name: paddleocr-hebrew
+Version: 0.1.0
+Summary: Hebrew OCR — PaddleOCR/SVTRv2 recognizers + a runnable page pipeline (word-DET + script-gated CTC->NRTR cascade)
+Author: Rivok Labs
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/RivoksLab/paddleocr-hebrew
+Project-URL: Models, https://huggingface.co/Rivok/paddleocr-hebrew
+Project-URL: Issues, https://github.com/RivoksLab/paddleocr-hebrew/issues
+Keywords: ocr,hebrew,paddleocr,svtrv2,rtl,onnx,bidi
+Classifier: Natural Language :: Hebrew
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering :: Image Recognition
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: NOTICE
+Requires-Dist: numpy>=1.24
+Requires-Dist: opencv-python>=4.8
+Requires-Dist: onnxruntime>=1.17
+Requires-Dist: pyclipper>=1.3
+Requires-Dist: Pillow>=10.0
+Requires-Dist: pypdfium2>=4.25
+Provides-Extra: gpu
+Requires-Dist: onnxruntime-gpu>=1.17; extra == "gpu"
+Dynamic: license-file
+
+# paddleocr-hebrew
+
+**Built by [Rivok Labs](https://rivoklabs.com)** — [rivoklabs.com](https://rivoklabs.com) · [ronen@rivoklabs.com](mailto:ronen@rivoklabs.com)
+
+[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.22163990.svg)](https://doi.org/10.5281/zenodo.22163990)
+[![License](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/LICENSE)
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/RivoksLab/paddleocr-hebrew/blob/main/examples/colab_quickstart.ipynb)
+
+Production-grade **Hebrew OCR** built on PaddleOCR — a family of finetuned
+recognizers (CTC + attention) plus a runnable page pipeline. Strong on the two
+things generic OCR gets wrong for Hebrew: **real-world documents** (menus,
+receipts, invoices, scans) and **bilingual Hebrew+Latin/digit lines**.
+
+The flagship recognizer reads long Hebrew lines *single-pass* (71–80 char lines
+at **0.49% CER**) and embedded Latin/number runs inside RTL Hebrew that a plain
+CTC model silently drops.
+
+Every model here is a **finetune of a PaddleOCR model** (PaddleOCR v3.7.0,
+Apache-2.0) and is released under the same licence — see [NOTICE](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/NOTICE) and the
+lineage column in [Models](#models) below. All reported CER figures are
+**in-house evaluation on our own test sets**: measured, but vendor-reported and
+not independently verified ([eval/EVAL_RESULTS.md](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/eval/EVAL_RESULTS.md)).
+
+> **Note on scope.** The Paddle training weights (`.pdparams`) and training
+> configs are **not** published — this is a release of inference-ready ONNX
+> models, not a reproducible training pipeline. See
+> [Fine-tuning & collaboration](#fine-tuning--collaboration).
+
+---
+
+## ⚠️ Read this first — Hebrew is RTL, output is LOGICAL order
+
+Every label and every model output in this project is **logical Unicode order**
+(the first character you read is the first character in the string). This is
+correct and what you want for storing, searching, scoring, and training.
+
+**Only call `python-bidi`'s `get_display()` at the moment you render** to a
+terminal, PDF, or GUI. If you apply `get_display()` before storing or comparing,
+you silently corrupt your text and every downstream CER looks broken. This is the
+#1 way people get garbage out of Hebrew OCR. See [docs/charset.md](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/docs/charset.md).
+
+```python
+from bidi.algorithm import get_display
+print(get_display(line["text"]))   # display ONLY — never before storage/scoring
+```
+
+## ⚠️ Also read this — these models do not read niqqud
+
+The 120-character charset contains **no niqqud (vowel points) and no te'amim
+(cantillation marks)**. The models never emit them. Feed a vocalised page in and
+you get the consonantal text back — the marks are dropped, silently.
+
+That is a deliberate scope choice, not a bug, but it rules out some real use
+cases. If you are digitising **liturgical, Biblical, poetic or children's
+material**, or **Yiddish and Ladino** (which use Hebrew script with different
+orthographic conventions), these models will not give you what you need as-is.
+
+It also means **you cannot score them against vocalised ground truth** without
+stripping marks from both sides first — otherwise you are measuring a charset
+mismatch, not recognition. To calibrate: on the Hebrew-script split of
+[GlotOCR-bench](https://github.com/cisnlp/GlotOCR-bench), 41 of 100 samples carry
+marks we cannot emit, so a *flawless* recogniser would still score 12.4% CER
+there — 38.9% on the Biblical Hebrew subset alone.
+
+Unvocalised Hebrew — which is the overwhelming majority of modern documents,
+signage, receipts, invoices, forms and press — is unaffected.
+
+See [docs/charset.md](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/docs/charset.md) for the full charset and the
+normalisation rules.
+
+## ⚠️ And this — document typography only, not display faces
+
+These models are trained on **document type**: receipts, menus, invoices, forms,
+scans, print. They do not generalise to decorative or display faces — striped and
+hatched fills, hollow outlines, heavy poster type.
+
+Measured, on the Hebrew-script split of
+[GlotOCR-bench](https://github.com/cisnlp/GlotOCR-bench), which renders every
+sample in a randomly chosen Google Font: the error distribution is bimodal.
+78 of 100 samples land at a median 5.3% CER; the other 22 fail outright (>50%
+CER), and every one of those is a decorative face. Failure rate is flat across
+languages, so it is font-driven, not script-driven.
+
+Two consequences worth planning around:
+
+- **Signage, packaging, posters and logos are out of scope.** Use a
+  general-purpose OCR or a VLM for those.
+- **Failure is loud, not quiet.** On a face it cannot read, the recogniser emits
+  confident garbage — sometimes Latin characters from Hebrew input — rather than
+  returning nothing. Check the script of your output if you feed it unknown
+  material.
+
+---
+
+## Quickstart (2 minutes)
+
+**Nothing to install:** [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/RivoksLab/paddleocr-hebrew/blob/main/examples/colab_quickstart.ipynb) runs the whole pipeline on a free
+Colab CPU runtime — read the sample page, then upload your own.
+
+From PyPI (Python 3.9 or later; developed and tested on 3.11):
+
+```bash
+pip install paddleocr-hebrew huggingface_hub
+```
+
+```python
+from huggingface_hub import snapshot_download
+from paddleocr_hebrew import HebrewOCR
+
+models = snapshot_download("Rivok/paddleocr-hebrew",            # ~190 MB, cached after the first run
+                           allow_patterns=["config.json", "charset_v2f.txt", "word-det/*", "server-svtrv2/*"])
+ocr = HebrewOCR.word(models)
+for line in ocr.read("page.png")["lines"]:     # image or PDF
+    print(line["text"])                         # logical order (get_display() to show)
+```
+
+From source, with the examples:
+
+```bash
+git clone https://github.com/RivoksLab/paddleocr-hebrew
+cd paddleocr-hebrew
+pip install -e .            # installs the `paddleocr_hebrew` package (ONNX inference only)
+pip install huggingface_hub # to auto-download the models from the Hub
+
+python examples/quickstart.py examples/sample_images/sample_page.png
+```
+
+```python
+from paddleocr_hebrew import HebrewOCR
+
+ocr = HebrewOCR(models_dir="path/to/downloaded/models")   # HF snapshot dir
+result = ocr.read("page.png")          # image or PDF
+for line in result["lines"]:
+    print(line["text"])                # logical order (get_display() to show)
+```
+
+- **GPU** (CUDA / Jetson): `pip install onnxruntime-gpu` instead of `onnxruntime`.
+- On **Jetson JetPack 6**, run inside a CUDA Docker container — bare-metal
+  onnxruntime falls back to CPU. (PaddlePaddle has no aarch64 wheel; everything
+  here runs on onnxruntime alone, no Paddle at inference.)
+
+Models are hosted on the **Hugging Face Hub**:
+[Rivok/paddleocr-hebrew](https://huggingface.co/Rivok/paddleocr-hebrew)
+(~370 MB, ONNX + per-model cards + md5sums). The small detectors and the mobile
+rec also ship in this repo under [`models/`](https://github.com/RivoksLab/paddleocr-hebrew/tree/main/models/) for a zero-download start.
+
+---
+
+## The pipeline
+
+```
+image / PDF page
+   → word detector (mobile DBNet, Hebrew-finetuned)      models/word-det
+   → word crops
+   → SVTRv2 recognizer, script-gated CTC→NRTR cascade    server-svtrv2
+       · fast single-pass CTC on every crop
+       · fall back to the NRTR split-ONNX host loop only on crops with an
+         embedded Latin/digit run or low CTC confidence (~10% of crops)
+   → reading-order assembly (RTL rows) → Hebrew post-processing
+   → {lines, words, meta}   (logical Unicode order)
+```
+
+This **word-DET + single-pass** design beats line-level detection on Hebrew
+pages (see benchmark). The recognizer is in
+[`paddleocr_hebrew/plan_e_rec.py`](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/paddleocr_hebrew/plan_e_rec.py); the orchestrator in
+[`paddleocr_hebrew/pipeline.py`](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/paddleocr_hebrew/pipeline.py).
+
+### Two pipelines (same code + recognizer, different detector)
+
+```python
+ocr = HebrewOCR.word(models_dir)   # default, flagship
+ocr = HebrewOCR.line(models_dir)   # line detector
+```
+
+| pipeline | page-CER | s/page (Jetson) | use it when |
+|---|---:|---:|---|
+| **`word`** (default) | **7.56%** | **4.95** | almost always — faster **and** more accurate |
+| `line` | 10.36% | 6.33 | degraded scans / dense hard layouts where word boxes over-fragment |
+
+`line` is **not** faster (fewer rec calls, but each is a wide line crop; the word
+detector's forward pass is also cheaper). Its edge is robustness on hard layouts —
+and, because the recognizer reads a whole line's BiDi in one pass, it keeps
+mixed Hebrew+Latin word order without post-hoc reordering. `quickstart.py`
+takes `--pipeline word|line`.
+
+---
+
+## Models
+
+All recognizers share one byte-identical 120-char charset
+([`models/charset_v2f.txt`](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/models/charset_v2f.txt)). Split-ONNX pairs
+(encoder + decstep) are one logical model — the attention decode loop runs on the
+host (see [docs/techniques.md](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/docs/techniques.md)).
+
+| model | role | arch | finetuned from | format / size |
+|---|---|---|---|---|
+| **server-svtrv2** | **flagship** server REC | SVTRv2 (CTC + NRTR heads) | SVTRv2, Chinese pretrain | `ctc.onnx` 77 MB (INT8 `ctc.int8.onnx` 20 MB) · NRTR split `enc 72 + dec 27 MB` |
+| light-svtrv2small | edge / CPU REC (NRTR-only) | SVTRv2-small (KD student) | distilled from server-svtrv2 | split `enc 28 + dec 27 MB` |
+| server-v5 | alt word-level server REC | PPHGNetV2-B4 (CTC) | PP-OCRv5 server rec | `rec.onnx` 73 MB |
+| server-v6 | alt word-level server REC | PPLCNetV4-medium (CTC) | PP-OCRv6 | `rec.onnx` 60 MB |
+| mobile-word | mobile word REC | PPLCNetV3 (KD student) | PP-OCRv5 mobile rec | `rec.onnx` 7.4 MB |
+| word-det | **word detector** (flagship) | mobile DBNet PPLCNetV3 | PP-OCR mobile DBNet | `det.onnx` 4.6 MB |
+| line-det | line detector (situational) | mobile DBNet PPLCNetV3 | PP-OCR mobile DBNet | `det.onnx` 4.6 MB |
+
+Which recognizer? **server-svtrv2** for the best quality on long + bilingual
+lines. **light-svtrv2small** for a small NRTR model on x86-CPU / edge.
+**mobile-word** (7.4 MB) for the smallest word-level footprint. server-v5/v6 are
+CTC word-level alternatives (note: v6 **cannot** do sliding-window long lines —
+SVTRv2 is the long-line tool). Per-model details + numbers in each
+`MODEL_CARD.md`.
+
+**Smaller CTC option:** `server-svtrv2/ctc.int8.onnx` (20 MB) is an INT8-quantized copy of
+the CTC head, a quarter of the size with practically the same accuracy (within 0.2 CER points of
+`ctc.onnx` on every slice of 1,861 real line crops; 10.39% vs 10.42% overall). It is a drop-in replacement:
+`HebrewOCR(models_dir, ctc="server-svtrv2/ctc.int8.onnx")`.
+
+---
+
+## Benchmark (headline)
+
+Full tables + methodology + caveats in [eval/EVAL_RESULTS.md](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/eval/EVAL_RESULTS.md).
+
+**Clean rendered pure-Hebrew** (public-domain text, perfect GT), micro-CER:
+
+| Tesseract | SVTRv2 CTC | cascade | SVTRv2 NRTR |
+|---:|---:|---:|---:|
+| 1.34% | 0.98% | 0.76% | **0.35%** |
+
+**Real bilingual heb+lat** (n=233 real crops), micro-CER: Tesseract **16.14%**
+vs our cascade/NRTR **2.33%** (~7× better).
+
+**Page-level** (71-page GCV-anchored, SVTRv2 cascade held constant), micro-CER:
+
+| word-DET + SVTRv2 | Hebrew line-DET + SVTRv2 | Tesseract |
+|---:|---:|---:|
+| **7.56%** | 10.36% | 14.20% |
+
+Tesseract narrowly wins *clean academic* Hebrew but loses badly on real domains
+(thermal 26% vs ~3%, menus 12% vs ~3%). Cross-domain robustness is the edge — see
+[docs/FINDINGS.md](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/docs/FINDINGS.md).
+
+---
+
+## What's novel here
+
+Two techniques we think are worth reusing (writeup in
+[docs/techniques.md](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/docs/techniques.md)):
+
+1. **Attention decoder on ONNX-only edge via a host-driven split-ONNX loop.**
+   A full NRTR decode loop won't export to ONNX (control flow) and Paddle won't
+   run on Jetson. Export encoder + one-decode-step as two control-flow-free graphs
+   and drive the greedy loop in numpy — reproduces paddle output exactly, runs on
+   onnxruntime everywhere.
+2. **CTC vs attention on bidirectional script.** A monotonic CTC head *deletes*
+   Latin/digit runs embedded in RTL Hebrew (heb+lat 12.6% CER); an attention head
+   on the same backbone reads them (2.3%). The script-gated cascade gets NRTR
+   quality at ~CTC speed.
+
+---
+
+## Documentation
+
+- [docs/charset.md](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/docs/charset.md) — the 120-char charset + the RTL logical-order rule
+- [docs/techniques.md](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/docs/techniques.md) — split-ONNX host loop, CTC-vs-attention, sliding-window, the cascade
+- [docs/FINDINGS.md](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/docs/FINDINGS.md) — what we learned (including the negative results)
+- [eval/EVAL_RESULTS.md](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/eval/EVAL_RESULTS.md) · [eval/run_eval.md](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/eval/run_eval.md) — benchmark + how to reproduce
+
+## Used by
+
+- [**HebrewOcrNet**](https://github.com/StasEdward/HebrewOcrNet) by Edward Stas: a C#/.NET
+  port of the word pipeline on ONNX Runtime and SkiaSharp, with no Python, on NuGet as
+  `HebrewOcrNet`. It reads the Hebrew documents in [Domeniq](https://www.domeniq.app), a
+  household paperwork ledger that runs at home.
+
+Built something on these models? Open an issue and we'll add it here.
+
+## Fine-tuning & collaboration
+
+The models are released as **ONNX** — inference-ready and portable (CPU, CUDA,
+Jetson) from one set of files. ONNX is not a compromise here: on CPU it runs
+~7× faster than native PaddlePaddle, on Jetson it's the only thing that runs
+(no Paddle aarch64 wheel), and for TensorRT speed on GPU you can point
+onnxruntime's TensorRT execution provider at these same `.onnx` files — no
+extra artifacts needed.
+
+The Paddle training weights (`.pdparams`) and training configs are **not
+published** — ONNX is all you need to run the models. If you want to **fine-tune
+on your own Hebrew data, adapt the charset, or collaborate**, please open an
+issue / discussion on this repo or reach out: **ronen@rivoklabs.com**. Happy to
+help and to grow this together.
+
+## About Rivok Labs
+
+[**Rivok Labs**](https://rivoklabs.com) builds computational intelligence and automation
+tools, with a particular focus on Hebrew and other right-to-left languages that
+mainstream tooling handles badly.
+
+We released this because there was no open-source Hebrew OCR with a commercial-friendly
+licence that held up on real documents. If it is useful to you, or if it fails on your
+documents, we would like to hear about it.
+
+- **Web:** [rivoklabs.com](https://rivoklabs.com)
+- **Contact:** [ronen@rivoklabs.com](mailto:ronen@rivoklabs.com)
+- **Issues and discussions:** on this repository
+
+## Citing this work
+
+[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.22163990.svg)](https://doi.org/10.5281/zenodo.22163990)
+
+```bibtex
+@software{rivoklabs_paddleocr_hebrew,
+  author  = {{Rivok Labs}},
+  title   = {paddleocr-hebrew: production Hebrew OCR with a
+             script-gated CTC/attention cascade},
+  year    = {2026},
+  doi     = {10.5281/zenodo.22163990},
+  url     = {https://github.com/RivoksLab/paddleocr-hebrew}
+}
+```
+
+That DOI always resolves to the latest release. To cite this exact version, use
+`10.5281/zenodo.22163991`. GitHub's "Cite this repository" button reads
+[CITATION.cff](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/CITATION.cff).
+
+## License
+
+Apache-2.0 (code and model weights). The recognizers are finetuned from
+PaddleOCR models (also Apache-2.0); see [NOTICE](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/NOTICE). See [LICENSE](https://github.com/RivoksLab/paddleocr-hebrew/blob/main/LICENSE).