heb_rivok 1.0.1

heb_rivok 1.0.1: pages with columns now come out in the right order.

Tesseract decides a model's page direction by counting its characters. heb_rivok 1.0 knows the 52 English letters and had only 33 Hebrew characters, so Tesseract laid its pages out left to right and read multi-column pages left column first. 1.0.1 adds 23 Hebrew characters the model never outputs; the network, dictionaries and recognised text are unchanged.

34 HaZefira newspaper pages	Ubuntu heb	heb_rivok 1.0	heb_rivok 1.0.1
words wrong or out of reading order	47.9%	88.5%	32.3%
words found (order ignored)	67.6%	81.3%	81.2%

• Identical text to 1.0 on every line test (real-document lines at --psm 7 and 13, GlotOCR) and on the 7 example documents.
• Fine-tuning from the float file runs exactly as from 1.0.
• heb_rivok_lite is unchanged (it was never affected; same file as in 1.0).

file	what it is
heb_rivok.traineddata (5.7 MB)	Hebrew, English and digits in one model. Start here.
heb_rivok_lite.traineddata (3.4 MB)	Hebrew and digits; smaller and about 3x faster per line. Use with +eng.
tessdata_best.zip	full-precision versions of both, for your own fine-tuning
SHA256SUMS.txt	checksums; please check before use

sha256sum -c SHA256SUMS.txt --ignore-missing
sudo cp heb_rivok.traineddata /usr/share/tesseract-ocr/5/tessdata/   # `tesseract --list-langs` prints the folder
tesseract page.png out -l heb_rivok
ocrmypdf -l heb_rivok scanned.pdf searchable.pdf

Full results, limits and the training recipe: README.

heb_rivok 1.0 and heb_rivok_lite 1.0

Two drop-in Tesseract 5 models for printed Hebrew, fine-tuned from Google's own Hebrew models on synthetic text lines. Apache-2.0.

file	what it is
heb_rivok.traineddata (5.7 MB)	Hebrew, English and digits in one model. Fine-tuned from tessdata_best script/Hebrew. Start here.
heb_rivok_lite.traineddata (3.4 MB)	Hebrew and digits; smaller and about 3x faster per line. Fine-tuned from tessdata_best heb; use with +eng.
tessdata_best.zip	full-precision versions of both, for your own fine-tuning
SHA256SUMS.txt	checksums; please check before use

sha256sum -c SHA256SUMS.txt --ignore-missing
sudo cp heb_rivok.traineddata /usr/share/tesseract-ocr/5/tessdata/   # `tesseract --list-langs` prints the folder
tesseract page.png out -l heb_rivok
ocrmypdf -l heb_rivok scanned.pdf searchable.pdf

Against Ubuntu's heb (character error rate, lower is better):

test	Ubuntu heb	heb_rivok
403 lines from real documents, Hebrew only	4.7%	1.4%
the same, lines with numbers	4.5%	1.6%
GlotOCR-bench Hebrew, clean / "old document" images	21.2% / 34.0%	18.5% / 30.1%
HaZefira newspaper (1874/1887), 4,013 lines	17.5%	9.9%

Each model also beats the Google model it was fine-tuned from on every test. Full results, limits (including short mixed Hebrew/English lines, a Tesseract ordering rule) and the training recipe are in the README; eval/ reproduces the public numbers.