Views
No views yet
exp8 checkpoint using frozen held-out data, greedy decoding, and raw references.| Dataset | exp8 CER | Phoenix CER |
|---|---|---|
| Agapet | 22.12% | 17.86% |
| Omar | 17.72% | 11.84% |
| TariMa | 10.39% | 10.72% |
| Field | Value |
|---|---|
| Architecture | CNN + BiLSTM + CTC |
| Parameters | 4,988,946 |
| Framework | Kraken / PyTorch |
| Model file | model.mlmodel |
| Output codec | 81 symbols + CTC blank |
| Internal checkpoint | exp9 |
| SHA-256 | 2896fef9d9665cbb82fba8faa3bf0c628ac6a5cc62eb678f40c707db833aebea |
1kraken -i page.jpg output.txt segment ocr -m model.mlmodel
2
3For complex manuscript layouts, line segmentation or PAGE-XML regions should preferably be supplied separately.
4
5Intended use
6
7Phoenix is intended for:
8
9- Arabic manuscript and archival-document transcription research
10- producing initial transcriptions for expert review
11- handwritten-text recognition experiments
12- human-in-the-loop manuscript workflows
13
14It is not intended to produce authoritative scholarly editions without human review.
15
16Athar
17
18Athar extends Phoenix with:
19
20- preservation of the raw visual reading
21- alternative transcription candidates
22- conservative language-model assistance
23- source-retrieval evidence
24- human accept/edit/reject decisions
25- PAGE-XML and TEI export
26
27These are system-level capabilities and are not encoded directly in the Phoenix model weights.
28
29Limitations
30
31- Performance varies substantially across handwriting domains.
32- Agapet remains difficult at 17.86% CER.
33- TariMa regressed slightly relative to the preceding checkpoint.
34- Layout and segmentation errors can strongly affect full-page performance.
35- Confidence and beam scores should not be treated as calibrated probabilities.
36- Human review remains necessary for scholarly use.
37
38License and training-data note
39
40Phoenix is released under CC BY-NC-SA 2.0 as a conservative project-level licensing choice.
41
42The model was developed using several Arabic handwriting resources, including Muharaf, RASAM, TariMa, Agapet, and Omar Al-Saleh data.
43
44The 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.
45
46Citation
47
48If you use Phoenix or Athar in research, please cite:
49
50@article{ali2026beyond,
51 title = {Beyond Recognition: Compact Multi-Domain Arabic Manuscript HTR with Candidate-Selection Analysis and Evidence-Preserving Review},
52 author = {Ali, Abdullah Ahmed and Abdulhadi, Mohammed Thamer and Safaa, Ali Haider and Wadi, Dhulfiqar Mahdi},
53 journal = {arXiv preprint arXiv:2608.19385},
54 year = {2026}
55}
56
57ملخص عربي
58
59Phoenix نموذج صغير للتعرف على الكتابة العربية اليدوية في المخطوطات والوثائق التاريخية، مبني باستخدام CNN + BiLSTM + CTC ويحتوي على نحو 4.99 مليون معامل.
60
61حقق في التقييم المحتجز CER قدره 17.86% على Agapet و11.84% على Omar و10.72% على TariMa.
62
63النموذج جزء من منظومة أثر Athar، ويُستخدم كمساعد للباحث وليس بديلًا عن المراجعة البشرية.