Pre-Alpha PP-OCRv6 (medium) HTR model for medieval/pre-modern Ethiopic (Geʽez, some Amharic) manuscripts

Abstract

Pre-Alpha Ethiopic (Geʽez / Amharic) recognition model — PP-OCRv6 medium fine-tune Status: preliminary / pre-Alpha. This deposit is an early public checkpoint, not a finished production model. Better versions are planned as soon as additional medieval and pre-modern Ethiopic ground truth can be annotated and curated. High-quality line-level GT for this domain is scarce and almost nonexistent in open corpora; the base PP-OCRv6 medium model itself only saw Geʽez as synthetic data. Treat current scores as provisional and expect material gains once real manuscript coverage grows. Line-level handwritten (and mixed handwritten/print-adjacent) text recognition model for Ethiopic-script materials, trained mainly on Geʽez with some Amharic, covering roughly 49 medieval / pre-modern manuscripts drawn from collections such as the British Library, Princeton University Library, vHMML / EMML, BnF, Vatican, Walters, and others. Fine-tuned in eScriptorium (default recognition training) with kraken from the public PP-OCRv6 (medium) weights (~15.9M parameters). The fine-tuned artifact metadata records variant: medium, PPOCRv6Model, baseline segmentation, and _kraken_min_version: 7.1.0. Architecture PP-OCRv6 CTC line recognizer as implemented in kraken ≥ 7.1 (feature/ppocrv6_rec): lightweight convolutional backbone + non-recurrent sequence neck. This record is the medium size class (not tiny/small). See the base model card for architecture details. Uses Intended: exploratory HTR on medieval / pre-modern Ethiopic-script manuscripts (primarily Geʽez; some Amharic), especially material similar to the training manuscripts listed below. Not intended (yet): a general-purpose Ethiopic OCR/HTR solution for all periods, print traditions, or hands; production pipelines that need stable orthography or abbreviation handling without human review. Roadmap: successive Zenodo versions under the same concept DOI as annotation volume and diversity increase. Transcription guidelines, Normalization, and Transformations Training and the reported ketos test metrics used Unicode NFD and whitespace normalization (ketos test -u NFD -n), consistent with the base PP-OCRv6 medium recipe. No claim is made that all contributing manuscripts share a single diplomatic/normalized transcription guideline; residual inconsistency (punctuation, word division, Geʽez punctuation marks such as ፡ ። ፤, and related letter confusions) is visible in the confusion list from evaluation. Bias, Risks, and Limitations Data scarcity: Open, reusable medieval Ethiopic HTR ground truth is extremely limited. This pre-Alpha model is trained on a comparatively small private PAGE-XML collection (~367 pages / ~18k text lines in the dump used for evaluation). Coverage is biased toward the hands, layouts, imaging, and transcription habits of those sources. Optimistic metrics: The hold-out CER/WER below is a random page split of the same dump, not a manuscript-disjoint or out-of-collection test. Generalization to unseen libraries, centuries, or scripts-within-Ethiopic may be substantially worse. Base-model gap: Upstream PP-OCRv6 medium listed Geʽez as synthetic-only; this fine-tune is an attempt to close that gap with real pages, but remains early. Language mix: Geʽez-dominant with some Amharic; do not assume balanced bilingual performance. Prefer human verification for scholarly editions; expect future versions to obsolete this checkpoint. How to Get Started with the Model Requires kraken ≥ 7.1 with PP-OCRv6 recognition support. Example: shell kraken -i page.png out.txt segment -bl ocr -m pp-ocrv6-ethiopic.safetensors Evaluate similarly to the reported numbers: shell ketos test --arch ppocrv6 -m pp-ocrv6-ethiopic.safetensors -f page -e test.lst -u NFD -n Training Details Training Data Private Ethiopic manuscript ground truth (PAGE XML + page images), prepared/used via eScriptorium. Approximate inventory of the current dump used for metrics: 367 pages, ~18,215 text lines. Languages: mostly Geʽez (gez), some Amharic (amh). Script: Ethiopic (Ethi). Manuscripts / shelfmarks represented in the project (~49 items, including a mixed set). Display names below drop the internal Ethiopic_ file prefix and replace _ with spaces: emufi, G1-IV-191, G1-IV-190, G1-IV-66, G1-IV-7, EMML 6977, NLI MSS-D, UTSC AM-011, UM Ethiopic MS 27, UM Ethiopic MS 25, RBSC MS-Ethiopian-3, SBB SBB Pet. II Nachtr. 29, SBB Pet. II Nachtr. 42, CU Or. 1802 Mark, NLI Israel Ms. 11, DSAE 00001, C3-IV-189, C3-IV-27, C3-IV-7, C3-IV-5, C3-IV-2, C1-IV-111, BL MS Bruce 74, CU Or. 1802, Garrett Ethiopic Manuscripts no. 30, Princeton Ethiopic Manuscript No. 49, Garrett Ethiopic Manuscripts no. 55, BL Or 12467, BL Or 8820, BL Or 733, BL Or 533, BL Or 529, BL Or 525, BL Or 519, BL Or. 485, BNF Ethiop. Abb. 202, Vat.et.25, Garrett Ethiopic Manuscript No. 1, Walters Ms. W.850, BL Or 518, BL Or 515, BL Or 513, BL Or 511, BL Or 510, BL Or 509, BL Or 493, BL Or 481, mix, BL Or 491. Sources include (non-exhaustively) British Library, Princeton (incl. Garrett), vHMML/EMML, BnF, Vatican, Walters Art Museum, and related holdings. The training images/transcriptions are not redistributed in this deposit unless a separate dataset DOI is added later. Training Procedure and Hyperparameters Trainer: eScriptorium default recognition training backed by kraken PP-OCRv6. Initialization: fine-tune of PP-OCRv6 medium (medium.safetensors). Artifact: pp-ocrv6-ethiopic.safetensors (safetensors metadata: variant=medium, baselines, RGB, codec size 1654 classes / 267 Ethiopic-block characters observed in the codec). Exact optimizer schedules and epoch counts follow eScriptorium/kraken defaults for this run and may be documented more precisely in a later version of this card. Evaluation Testing Data, Factors & Metrics Testing Data Deterministic ~10% page holdout (MD5 of filename mod 10 == 0): 33 PAGE XML pages from the same Ethiopic dump (~22.5k characters of GT after extraction). Same imaging and annotation pipeline as training pages; not held out by manuscript. Metrics Measured with single-GPU ketos test --arch ppocrv6 … -u NFD -n (CUDA, RTX 3090): | Metric | Value | |--------|------:| | Characters | 22,451 | | Errors (ins/del/sub) | 75 (8 / 19 / 48) | | Character accuracy | 99.67% | | Word accuracy | 96.80% | | CER | 0.33% | | WER | 3.20% | eScriptorium reported ~98.4% accuracy for the training project (UI metric; may not match ketos CER exactly). These figures should be read as preliminary in-domain scores for a pre-Alpha model. Re-evaluation on manuscript-disjoint and external collections is planned as more GT becomes available. Citation Cite this Zenodo record (once published) and the base model: Base weights: Kiessling, B. (2026). PP-OCRv6 (medium) multilingual handwritten and printed text recognition model. Zenodo. https://doi.org/10.5281/zenodo.21788410 Software: kraken (https://github.com/mittagessen/kraken) Please also credit the holding institutions of any manuscripts whose images informed training when publishing downstream results.