Answers are grounded in this paper and cite their sources.
HeOCR oldprint-v1: Hebrew old and rabbinic print recognition for Tesseract oldprint-v1 is a Tesseract language model (.traineddata) for old printed Hebrew in the square script: 19th- and early-20th-century books and rabbinic editions. HeOCR made it by continuing to train HeOCR R7, itself a fine-tune of Tesseract's own heb model from tessdata_best, for one pass over 76,946 lines: - 22,910 real scanned lines from 4 public-domain books (Internet Archive scans), labelled with the Project Ben-Yehuda transcription of the same edition; - 14,036 real scanned lines from 3 rabbinic editions (Internet Archive scans), labelled with public-domain Sefaria texts and Project Gutenberg ebook #43740; - 40,000 lines rendered from born-digital booklets of Israel's laws and regulations (Reshumot), labelled with the publisher's own text layer, replayed so that modern print is not lost. No label is OCR output. It is the model that HeOCR's evaluation recommends for old print in heocr, the HeOCR document pipeline, beside modern-v1 for modern print. (In the experiment records it is P-hist round v1, arm r7, heb_final.) Performance: not measured on a held-out set yet. No held-out score yet; HeOCR will score it once in a registered held-out run. That run is the P-hist test (22 held-out books), once its human references exist, against R7 and stock heb; the result will be added to this card in a new version. How it was chosen: the lowest page CER on audit2, a development set of 38 old-print pages from 13 books read whole at psm 3, among 20 candidates (8 snapshots, the training-best checkpoint and the final model of this arm and of a second arm started from stock heb). audit2 made the choice, so no figure from it is reported here. Not for modern office documents. On HeOCR's synthetic office development sets, inside heocr, it reads worse than stock heb on scanned and image-only PDFs and scanned tables. Use modern-v1 there. File heocr_oldprint_v1.traineddata (trained as heb.traineddata; selected checkpoint heb_final), 3,628,595 bytes sha256 f8090760b5b02d09cf4ddb435e7a5e34089f6f17cde9cfb6c67fd06c6f17ec02 Model type recognition (line-level LSTM, CTC); Tesseract does the page layout Format Tesseract 4/5 LSTM traineddata, LSTM only (--oem 1; no legacy engine). Trained with Tesseract 5.5.3, built from its upstream tag Continued from HeOCR R7 (10.5281/zenodo.24126448), heocr_r7.traineddata, sha256 95faab81fba6fe57f854f973c1ccad2b0d1059fb951f34b17277385ceea8b303 Base model tessdata_best heb.traineddata at e12c65a (Apache-2.0), sha256 dbaa827aea6bc21215638447f17783a1004987c2d0bf5573d111fee397abdae5, fine-tuned (modified) by HeOCR as R7 and then oldprint-v1 Character set R7's, unchanged: stock heb's plus the maqaf ־ and the en dash – Licence Apache-2.0, Copyright (c) 2026 HeOCR maintainers (see LICENSE, NOTICE and NOTICE.r7) Training record oldprint_v1_training_run.json (heocr_tesseract_training_run.v1) Code HeOCR/heocr-engine; the experiment is HeOCR/hocrsyngen experiments/phist_v1/ Uses OCR of old printed Hebrew in the square script without niqqud: 19th- and early-20th-century books, and rabbinic editions set in square type. Use it in place of Tesseract's heb or R7. The old-print step of a document pipeline. heocr routes old print to it. Line recognition of single printed Hebrew line images, and a research baseline for old Hebrew print. Out of scope Modern office documents, forms and tables. Use modern-v1. Rashi script and other non-square scripts. Every training line is square type. Handwriting. Niqqud (vowel points) and cantillation. The labels have none. Text it cannot write: geresh ׳, gershayim ״, the em dash — and Latin letters. Training labels wrote ׳ ״ as ' " and — as –. Unsupervised transcription. Check the output. How to get started with the model Install Tesseract 4 or 5 (trained with 5.5.3). Put heocr_oldprint_v1.traineddata in a directory, say DIR, and pass that directory and the language name heocr_oldprint_v1. tesseract page.png out -l heocr_oldprint_v1 --tessdata-dir DIR --psm 3 audit2 was read this way, whole pages at psm 3. For single line images, use --psm 7. In heocr, once the model is registered there: heocr --download-model heocr_oldprint_v1 Notes: - --tessdata-dir replaces Tesseract's data directory. Output formats other than plain text read their config files from DIR/configs. Link your installation's configs folder into DIR, or put the model in your installation's tessdata directory and drop --tessdata-dir. - Maqaf and dashes. It writes the maqaf ־ (U+05BE) where many texts write -. Fold it if your downstream text uses the ASCII form. Training details Training data Lines: 76,946 in one seeded-shuffle list (seed 20261130), phist_r7.list (sha256 0cd3e7e8788f3c278ae9083f953a88f7319c70910d75423d0349733f7f72820c; labels sha256 bbdaade092892644bea4397061cbc1a88f9537a5f59412e12c0b0ab59c9cf33b). Source Line set Lines Share Rights 4 Internet Archive book scans, Ben-Yehuda labels bynli_lines_v3 22,910 29.8% Internet Archive, original conditions apply; text public domain 3 Internet Archive scans of rabbinic editions, Sefaria and Project Gutenberg labels rablines_v1 14,036 18.2% Internet Archive, original conditions apply; texts public domain Reshumot lines rendered from born-digital PDFs (replay) office-print v0 anchor, decontaminated 40,000 52.0% PD-IL (Copyright Act 2007 §6) The books and editions, each capped at 8% of the round: Source Record Book Lines IA zikhronotlebetda01reck Abraham Shalom Friedberg (d. 1902), Zikhronot le-Vet David, Warsaw, 1893 6,099 IA (Google-derived) mahzor-vitri-abraham-berliner-1893 Simhah ben Samuel of Vitry, Machzor Vitry, ed. Simeon Hurwitz, Berlin: Mekize Nirdamim, 1889–97 6,077 IA midrashleaovhame0305tobi Tobiah ben Eliezer, Midrash Lekach Tov, ed. Salomon Buber (d. 1906), Vilna, 1884, vols 3–5 6,063 IA (Google) afahleneemaniml00fngoog Samuel Joseph Fuenn (d. 1891), Safah le-Ne'emanim, Vilna, 1881 5,749 IA kolkitveylkatsen018800 Judah Leib Katzenelson (d. 1917), Kol Kitve Y. L. Katsenelson, St Petersburg, 1905 5,712 IA seferhademaotmeo0000bern Simon Bernfeld (d. 1940), Sefer ha-Dema'ot, Berlin, 1923 5,350 IA betnekhothahalak01horouoft Bet Nekhot ha-Halakhot o Toratan shel Rishonim, part I, ed. Chaim Meir Horowitz, Frankfurt am Main, 1881 1,896 Book audit: only books whose labels passed a per-book audit were used: at most 1 word error in 32 seeded lines. Books that failed it were left out. Internet Archive scans: five were scanned by university libraries (Robarts – University of Toronto; University of Illinois Urbana-Champaign; University of Toronto, Downsview), named in the NOTICE. Two are Google-class scans, retrieved via the Internet Archive, not from Google or NLI: afahleneemaniml00fngoog, digitized by Google (contributing library: University of California); mahzor-vitri-abraham-berliner-1893, a copy the Open Siddur Project derived from a Google Books scan. They are admitted under rule 3 of HeOCR's legal strategy and heocr-engine decision 0015. Google's front matter was excluded, and lines under a "Digitized by Google" watermark were dropped; none was found on the training pages. Each item passed HeOCR's Internet Archive rights gate, and each text aligned to it has its own rights justification, listed by id in the training record and the NOTICE. - Text sources: Project Ben-Yehuda (public domain dump, release 2026-03) for the four books; Sefaria's public-domain digital versions of the 1884 and 1889–97 printings; and Project Gutenberg ebook #43740 (public domain in the USA), produced by David T. Jones, Enrico Segre and Distributed Proofreaders Canada. - Reshumot replay: a seeded sample (seed 20261129) of 40,000 of the 45,250 decontaminated anchor lines of HeOCR's office-print v0 round: lines rendered by HeOCR from born-digital booklets of Sefer ha-Hukim and Kovetz ha-Takanot, labelled with the publisher's own text layer. The 8% per-book cap left too few P-hist lines to fill a round, so the replay share is high, 52%. - Labels: logical order, NFC, with ׳→', ״→" and —→–, so that every label fits R7's character set and matches the scoring profile. - Decontamination: 76 pool lines that shared a word 6-gram with an evaluation target were dropped. HeOCR's evaluation then ran its official overlap check (check_train_overlap: word 6-grams after folding, against every evaluation target) on the whole list: CLEAN, with 0 hits on every target (2026-11-18). Its page check against the P-hist hold-out found only 5 pages sharing text, and those pages are permanently out of scoring. - All authors, editors and named contributors of the books and editions died more than 70 years ago, per their rights justifications, so the books are in the public domain in Israel. Every book was printed before 1931 and is also past the US term. - R7's own training data (95,151 lines from 7 NLI "public domain, free use" scans and 23 Internet Archive university-library scans, with Project Ben-Yehuda labels) is credited in R7's NOTICE, which ships with this model as NOTICE.r7. Training procedure Start: HeOCR R7 (--continue_from its LSTM, --old_traineddata and --traineddata R7). The network and character set are R7's, unchanged. Trainer: lstmtraining, Tesseract 5.5.3 built from its upstream tag (commit db0ec62f…), on a 16-vCPU CPU pod, --learning_rate 0.0001 --reset_learning_rate --target_error_rate 0.01, one epoch (76,946 iterations), with a snapshot every 10 minutes. The run exited cleanly; the pod checked every input's sha256 before training. Truth order: labels in Tesseract's LSTM training order, converted with tesseract … --psm 13 lstm.train. Selection: the lowest page CER on audit2 (see Evaluation) under line_fold_v1, over 20 candidates: 8 snapshots, the training-best checkpoint and the final model of this arm and of a second arm started from stock heb. The winner is this arm's final model; the stock-start arm's
No takes yet. Share an insight, caveat, or question.
Affek et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: