Finetuned from PaddleOCR. Every model in this repository is a finetune of a PaddleOCR model (PaddleOCR v3.7.0, Apache-2.0) and is released under the same licence — see the per-model MODEL_CARD.md for each lineage.

All reported CER figures are in-house evaluation on our own test sets: measured, but vendor-reported and not independently verified. The benchmark harness is not shipped; methodology is documented in the GitHub repo.

Training weights (.pdparams) and training configs are not published. This is a release of inference-ready ONNX models, not a reproducible training pipeline.

paddleocr-hebrew — models

Built by Rivok Labs — rivoklabs.com · ronen@rivoklabs.com · code on GitHub

ONNX weights for Hebrew OCR (finetuned from PaddleOCR / SVTRv2). Runnable pipeline code, quickstart, benchmark, and docs are on GitHub: https://github.com/RivoksLab/paddleocr-hebrew.

All recognizers share one byte-identical 120-char charset (charset_v2f.txt, md5 e17ce22e7b4ab8224a3dad9e4c85b6ae). Each folder has a MODEL_CARD.md + md5sums.txt. Split-ONNX pairs (nrtr-encoder + nrtr-decstep) are one logical model — the attention decode loop runs on the host.

Hebrew is RTL — output is LOGICAL Unicode order. Apply python-bidi get_display() only when rendering, never before storing/scoring. This is the #1 way to get garbage out. See the GitHub docs/charset.md.

Contents

folder	role	files	size
server-svtrv2/	flagship server REC (SVTRv2, CTC + NRTR)	ctc.onnx, nrtr-encoder.onnx, nrtr-decstep.onnx	77 + 72 + 27 MB
light-svtrv2small/	edge/CPU REC (NRTR-only KD student)	nrtr-encoder.onnx, nrtr-decstep.onnx	28 + 27 MB
server-v5/	alt word-level server REC (PPHGNetV2-B4)	rec.onnx	73 MB
server-v6/	alt word-level server REC (PPLCNetV4)	rec.onnx	60 MB
mobile-word/	mobile word REC (PPLCNetV3 KD)	rec.onnx	7.4 MB
word-det/	word detector (mobile DBNet)	det.onnx	4.6 MB
line-det/	line detector (situational)	det.onnx	4.6 MB

Quickstart

pip install "git+https://github.com/RivoksLab/paddleocr-hebrew" huggingface_hub
hf download Rivok/paddleocr-hebrew \
    --include "charset_v2f.txt" "word-det/*" "server-svtrv2/*" \
    --local-dir hebrew-ocr-models

from ocr import HebrewOCR
ocr = HebrewOCR(models_dir="hebrew-ocr-models")
for line in ocr.read("page.png")["lines"]:
    print(line["text"])   # logical order

Benchmark (headline)

• Clean pure-Hebrew (perfect GT): SVTRv2 NRTR 0.35% CER vs Tesseract 1.34%.
• Real bilingual heb+lat (n=233): cascade/NRTR 2.33% vs Tesseract 16.14%.
• Page-level (71-page GCV): word-DET + SVTRv2 7.56% vs Tesseract 14.20%.

Full tables, methodology, and the CTC-vs-attention finding: see GitHub.

Fine-tuning

Released as ONNX only (inference-ready, portable — CPU/CUDA/Jetson). Paddle training weights + configs are not published. To fine-tune on your own Hebrew data or collaborate, open an issue on the GitHub repo or reach out: ronen@rivoklabs.com.

About Rivok Labs

Rivok Labs 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 these models because there was no open-source Hebrew OCR with a commercial-friendly licence that held up on real documents. If they are useful to you, or if they fail on your documents, we would like to hear about it.

• Web: rivoklabs.com
• Contact: ronen@rivoklabs.com
• Issues and discussions: github.com/RivoksLab/paddleocr-hebrew

License

Apache-2.0. Finetuned from PaddleOCR (Apache-2.0) — see GitHub NOTICE.