factlogic/phoenix-arabic-manuscript-htr
Phoenix - Arabic Manuscript HTR Model
Phoenix is a compact handwritten-text recognition model for Arabic historical manuscripts and archival documents.
It uses a CNN + BiLSTM + CTC architecture with approximately 4.99 million parameters and is designed to work across several Arabic handwriting domains.
Phoenix is the recognition component of Athar, an evidence-aware human-in-the-loop workflow for manuscript transcription.
Paper
Beyond Recognition: Compact Multi-Domain Arabic Manuscript HTR with Candidate-Selection Analysis and Evidence-Preserving Review
arXiv:2608.19385
https://arxiv.org/abs/2608.19385
Main results
Phoenix was evaluated against the preceding exp8 checkpoint using frozen held-out data, greedy decoding, and raw references.
Across the two large Agapet and Omar held-out sets, character-weighted CER decreased from:
19.98% -> 14.93%
corresponding to a 25.3% relative reduction in character errors.
TariMa is the reported exception, with a regression of 0.33 CER percentage points.
Model details
Usage
Phoenix can be used as a Kraken recognition model:
kraken -i page.jpg output.txt segment ocr -m model.mlmodel
For complex manuscript layouts, line segmentation or PAGE-XML regions should preferably be supplied separately.
Intended use
Phoenix is intended for:
- Arabic manuscript and archival-document transcription research
- producing initial transcriptions for expert review
- handwritten-text recognition experiments
- human-in-the-loop manuscript workflows
It is not intended to produce authoritative scholarly editions without human review.
Athar
Athar extends Phoenix with:
- preservation of the raw visual reading
- alternative transcription candidates
- conservative language-model assistance
- source-retrieval evidence
- human accept/edit/reject decisions
- PAGE-XML and TEI export
These are system-level capabilities and are not encoded directly in the Phoenix model weights.
Limitations
- Performance varies substantially across handwriting domains.
- Agapet remains difficult at 17.86% CER.
- TariMa regressed slightly relative to the preceding checkpoint.
- Layout and segmentation errors can strongly affect full-page performance.
- Confidence and beam scores should not be treated as calibrated probabilities.
- Human review remains necessary for scholarly use.
License and training-data note
Phoenix is released under CC BY-NC-SA 2.0 as a conservative project-level licensing choice.
The model was developed using several Arabic handwriting resources, including Muharaf, RASAM, TariMa, Agapet, and Omar Al-Saleh data.
The exact historical Muharaf training snapshot and license version were not retained with sufficient provenance. Users should therefore review the original upstream dataset terms for their intended use.
Citation
If you use Phoenix or Athar in research, please cite:
@article{ali2026beyond,
title = {Beyond Recognition: Compact Multi-Domain Arabic Manuscript HTR with Candidate-Selection Analysis and Evidence-Preserving Review},
author = {Ali, Abdullah Ahmed and Abdulhadi, Mohammed Thamer and Safaa, Ali Haider and Wadi, Dhulfiqar Mahdi},
journal = {arXiv preprint arXiv:2608.19385},
year = {2026}
}
ملخص عربي
Phoenix نموذج صغير للتعرف على الكتابة العربية اليدوية في المخطوطات والوثائق التاريخية، مبني باستخدام CNN + BiLSTM + CTC ويحتوي على نحو 4.99 مليون معامل.
حقق في التقييم المحتجز CER قدره 17.86% على Agapet و11.84% على Omar و10.72% على TariMa.
النموذج جزء من منظومة أثر Athar، ويُستخدم كمساعد للباحث وليس بديلًا عن المراجعة البشرية.