glyphive-ocrmodel-ocrb

An opt-in OCR model for glyphive, fine-tuned for OCR-B renderings of the base16c alphabet.

pip install glyphive-ocrmodel-ocrb
glyphive extract -f scan/ --from-images --ocr-engine tesseract-glyphive-ocrb -C out

Installing it registers a tesseract-glyphive-ocrb OCR provider through the glyphive.ocr_providers entry point. If the model file or Tesseract is not present the provider reports itself unavailable and glyphive falls back to a core engine — installing this package never breaks the stock path.

Why

On the held-out training sweep (benchmarks/results/ocr-training-sweep-20260718.json in the core repo) the fine-tuned model reads the base16c channel at 0.000% CER clean and blurred, versus ~4.6% for stock eng. The core document-wide Reed-Solomon + per-line CRC already make the stock path restore correctly, so this model is a robustness upgrade for marginal scans, not a requirement.

The model file (not in source control)

The trained glyphiveocrb.traineddata (~15 MB) is not committed. It is produced by the reproducible VM recipe in the core repo (benchmarks/training/measure_sweep.py / train_ocr_models.py) and added as package data only when a release wheel is built:

# uncomment in pyproject.toml [tool.hatch.build.targets.wheel] once present:
artifacts = ["src/glyphive_ocrmodel_ocrb/glyphiveocrb.traineddata"]

Licensing

• Base model: a fine-tune of Tesseract tessdata_best/eng (Apache-2.0, redistributable). Exact base used: version 4.00.00alpha:eng:synth20170629 (tessdata_best; trained with Tesseract 5.4.1), SHA-256 8280aed0782fe27257a68ea10fe7ef324ca0f8d85bd2fd145d1c2b560bcb66ba.
• Training corpus: synthetic — randomly generated base16c-alphabet lines rendered to images. OCR-B (Raisty) under SIL OFL 1.1 with Reserved Font Name 'OCR-B'. Only synthetic renders were used for training (no font file is redistributed in this wheel), so the reserved-name obligation is not triggered.
• Engine compatibility: the LSTM .traineddata format is shared across the Tesseract 4.x/5.x line, so one model file serves both (verified).