ATHAAR: A Medieval Scientific Arabic Manuscript Dataset for Handwritten Text Recognition

Ismail El Jamiy; Khadija Msalek; Guillaume Loizelet; Farah Benamara; Nabil Hathout

cs.CVNov 5, 2026

ATHAAR: A Medieval Scientific Arabic Manuscript Dataset for Handwritten Text Recognition

Authors: Ismail El Jamiy, Khadija Msalek, Guillaume Loizelet, Farah Benamara, Nabil Hathout

Organizations: IRIT, Université de Toulouse, France · IMT, Université de Toulouse, France · CLLE, Université de Toulouse Jean-Jaurès, France

Abstract

Despite recent progress in handwritten text recognition, medieval Arabic scientific manuscripts remain an open challenge for current vision-language models (VLMs) due to cursive ligatures, variable diacritization, and more importantly the lack of training data. This paper aims to address this gap, by proposing the first annotation guidelines dedicated to the transcription of scientific Arabic manuscripts inspired by textual criticism, a well-established approach to historical text editing. We then propose, ATHAAR, a new dataset manually annotated following this guideline. Unlike previous resources in the field, ATHAAR varies both witnesses of the same text, calligraphic styles, and content to deal with texts but also numerals. Relying on state-of-the-art VLMs, we finally design a set of experiments to study models transfer across scripts, chapters, and treatises. Our findings highlight the limitations of current systems when faced with the specificities of ancient scientific texts.

Figures & tables

Name	Domain	Date / period	Volume	Numeral content	Script variation	Several witnesses of the same text	Guide / agreement
Muharaf ( Saeed et al., 2024 )	Literature	19th–21st c.	1,644 images; 36,311 lines	n/r	No: Only Ruq’ah	No	✗ / ✗
Agapet ( Ibrahim, 2024 )	Religious	13th, 14th, 17th c.	831 pages	n/r	Yes, n/r	n/r	✗ / ✗
Historical Arabic HTR ( Najam and Faizullah, 2024 )	Religious	n/r	40 pages; 640 lines	n/r	n/r	No	✗ / ✗
RASM ( Clausner et al., 2018 )	Scientific	10th–19th c.	120 pages; 2,164–2,613 lines	Yes: n/r	n/r	n/r	✗ / ✗
RASAM ( Vidal-Gorène et al., 2021b )	Literature	17th–19th c.	300 images; 7,540 lines (v1)	Yes: fractions, tables; n/r	No: Only Maghrebi	No	✗ † / ✗
Iskandar ( LiPA and Huma-Num, 2024 )	Literature	17th–19th c.	297 pages	n/r	Yes, n/r	Yes: 5 copies of one work but not same content	✗ / ✗

Table 1: Arabic historical HTR datasets. n/r: not reported in the sources checked. The last column reports whether a an annotation guideline exists and whether inter-annotators agreement is measured ( ✓ yes, ✗ no). † : RASAM documents general transcription choices but does not release a formal annotation guide.

Manuscript	Date	#ch	Script	#Pg/Ln	%D.	Dg.
Tiers 1–2: al-Qānūn
Ms01 (Berlin, 1613)	12th c.	2	Naskh	7/224	2.58	198
Ms02 (Berlin, 275)	19th c.	2	Nasta’liq	11/290	0.08	200
Ms04 (Cairo, 866)	13th c.	2	Naskh	8/217	20.63	195
Ms05 (Hyderabad, 374)	17th c.	1	Nasta’liq	8/156	0.00	190
Ms06 (Istanbul, 87)	11–12th	2	Naskh	14/339	12.88	202

Table 2: Athaar manuscripts identified as Ms (Location, Shelfmark) . #ch, Pg, Ln: number of chapters, pages and lines; %D.: diacritic marks as a share of non-space characters; Dg.: transcribed digits.

Manuscript	Lines	CER (%) $\downarrow$	WER (%) $\downarrow$
Ms01	33	1.6	7.1
Ms02	27	2.4	9.2
Ms13	31	4.5	20.5
Ms17	36	2.2	12.2
Pooled	127	2.6	11.7

Table 3: Inter-annotators agreement on the first page of four witnesses of chapter X.6.

		Gemini		CHURRO		Qari-OCR
Exp.	$k$	CER	WER	CER	WER	CER	WER
C-Scr (N)	0	17.2	62.1	76.5	124.5	88.6	135.5
C-Scr (N)	3	16.3	62.3	70.0	115.6	154.5	198.9
C-Scr (NT)	0	20.0	55.4	87.3	143.4	93.9	157.6
C-Scr (NT)	3	18.5	53.8	79.9	134.9	158.6	219.5
C-Ch	0	18.9	61.8	75.3	127.5	78.7	127.7

Table 4: CER and WER (%). Best scores are in bold (lower is better) at zero-shot (

k=0

) and few-shot (

k=3

). C-Scr(a) : query witnesses are either in a =Naskh or a =Nasta’liq. Values above 100% indicate outputs longer than the reference.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 1 — Figure 1: Examples of difficulties found in manuscripts (cursive script, numerals, tatweel ).

Panel of Figure 2 — Figure 2: The same passage of Chapter X.6 across four manuscripts. Ms03 was excluded from the corpus due to severe degradation.

Manuscript	Date	#ch.	Chapter	Script	Statistics
		#ch.	Chapter	Script	Pages	Lines	Words	Chars	Diacr. marks	% diacr.	Numeral strings	Digits	Tags
Tiers 1–2: al-Qānūn al-Mas‘ūdī
Ms01_Berlin_Or-Quart-1613	12th c., 1166	2	X.6	Naskh	5	158	3,114	13,461	316	2.35	51	198	3
		2	I.4	Naskh	2	66	1,298	6,079	189	3.11	0	0	0
Ms02_Berlin_Or-Oct-275	19th c., 1834	2	X.6	Nasta’liq	7	180	2,672	10,964	14	0.13	52	200	0
		2	XI.2	Nasta’liq	4	110	1,519	6,881	1	0.01	0	0	0

Appendix

Table 5: Statistics of the manuscripts, one row per manuscript. “Chars” excludes spaces; “Diacr. marks” counts diacritic marks and “% diacr.” is their share of “Chars”. “Numeral strings” counts the numerals (transcribed or tagged) and “Digits” the digits transcribed. Manuscript identifiers match the dataset folder names to facilitate reproducibility.

Figure 3: Calfa Vision annotation interface.

Figure 4: Example of a word split across the text column and the margin, alongside expected transcription.

Figure 5: Example of partially vocalized text alongside unvocalized passages.

Panel of Figure 7 — Figure 7: Examples of annotation challenges arising from manuscript ambiguity rather than annotator error.

Exper.	Total lines	Query (70%)	Support (30%)
C-Scr	2,164	1,516	648
C-Ch	804	564	240
C-Tr	924	646	278

Appendix

Table 6: Line counts per experiment after the 70/30 query/support split.

Naskh queries — support from Nasta’liq pool
	Gemini		CHURRO		Qari-OCR
Witness	CER	WER	CER	WER	CER	WER
Ms01 (Berlin 1613)	14.3	48.4	64.2	115.2	175.8	225.0
Ms04 (Cairo 866)	34.1	123.9	62.2	109.8	106.5	153.3
Ms06 (Istanbul 87)	19.6	86.7	54.5	93.9	123.5	155.1
Ms07 (Konya 6625)	13.4	41.5	74.6	125.2	247.9	338.2

Appendix

Table 7: Per-witness CER and WER (%) for C-Scr at

k=3

(cross-script).

	Gemini		CHURRO		Qari-OCR
Witness	CER	WER	CER	WER	CER	WER
Ms01 (Berlin 1613)	5.6	26.9	35.3	85.0	156.4	171.8
Ms02 (Berlin 275)	25.3	60.3	55.5	109.3	250.2	343.0
Ms04 (Cairo 866)	28.0	99.7	40.7	91.0	124.5	117.5
Ms06 (Istanbul 87)	23.3	82.9	57.1	108.0	203.3	168.3
Ms08 (Istanbul 1498)	35.0	86.0	111.0	167.5	106.4	172.0

Appendix

Table 8: Per-witness CER and WER (%) for C-Ch at

k=3

(cross-chapter transfer).

	Gemini		CHURRO		Qari-OCR
Witness	CER	WER	CER	WER	CER	WER
Ms Ar 2494 (Abū al-Wafā)	25.1	63.3	68.7	105.8	153.7	175.2
Ms Ar 2458 (Ibn al-Haytham)	29.3	73.2	75.3	114.7	335.7	335.8
Ms Ar 2504 (Farghānī)	2.3	13.7	38.4	107.4	260.6	381.1
Ms Ar 2546 (Farghānī)	13.9	39.8	49.3	110.8	217.5	319.4
Ms Ar 2497 (Bīrūnī)	7.5	25.9	32.8	104.9	275.6	377.8

Appendix

Table 9: Per-witness CER and WER (%) for C-Tr at

k=3

(cross-treatise transfer).

$k$	CER (%) $\downarrow$	WER (%) $\downarrow$
1	$14.08\pm 0.48$	$38.79\pm 1.95$
3	$14.55\pm 0.90$	$39.73\pm 0.81$
5	$14.31\pm 1.28$	$38.66\pm 1.94$
7	$15.11\pm 0.21$	$39.63\pm 0.72$
10	$14.39\pm 1.59$	$38.64\pm 2.97$

Appendix

Table 10:

k

-shot ablation on Tier 1 (Gemini 3 Flash, 3 trials per

k

). Scores are reported in CER and WER.

Optical Character Recognition

Optical character recognition converts text in images into machine-readable text, from printed pages and handwriting to signs and symbols.

See all 206 papers

Performance Gap Analysis between Latin and Arabic Scripts HTR
Jun 17, 2026Sana Al-azzawi, Elisa Barney, Marcus LiwickiOnline Handwriting Recognition Optical Character Recognition
Luleå University of Technology Department of Computer Science, Electrical and Space Engineering Luleå, 97187, Sweden
AraMS-28k: The Largest Publicly Released Line-Level Dataset of Historical Arabic Manuscripts with Margin and Insertion-Anchor Annotations
Aug 27, 2026Mohamed Guechaoui, Mohamed Diaa Zellagui, Souleyman Chaib +1Historical Document Processing Optical Character Recognition
Higher School of Computer Science (ESI-SBA), Sidi Bel Abbes, Algeria
When Do VLMs Help Arabic Manuscript OCR? A Cross-Dataset Study
Aug 23, 2026Moshiur Farazi, Firoj Alam, Abderrahmane Maaradji +3Optical Character Recognition Document Image Understanding
University of Doha for Science and Technology, Doha, Qatar · Qatar Computing Research Institute, Hamad Bin Khalifa University, Doha, Qatar · Northwestern University in Qatar, Doha, Qatar

Jun 17, 2026J/K move · Enter open · S save

Performance Gap Analysis between Latin and Arabic Scripts HTR

Sana Al-azzawi, Elisa Barney, Marcus Liwicki

Luleå University of Technology Department of Computer Science, Electrical and Space Engineering Luleå, 97187, Sweden

Online Handwriting Recognition Optical Character Recognition

Open paper AskPDF arXiv

Recent studies have shown that handwritten text recognition (HTR) systems perform worse on Arabic-script datasets than on Latin-script data. However, the reasons for this gap are still not well understood due to the lack of controlled comparisons. In this work, we present a comprehensive study of Arabic and Latin scripts HTR using a unified CRNN model for line-level HTR across nine datasets (including KHATT (Arabic), Muharaf (Arabic), NUST-UHWR (Urdu), PHTD (Persian), IAM (English), READ-2016 (German), and others) and di ferent training sizes (K in {100, 500, 1000, 2000, ..., Kfull}). Our results show the performance gap remains: it is large in low-resource settings, decreases with more data, but remains even at full scale, with a consistent difference of 5-7 CER points. We show that annotation quality matters, as many datasets contain labeling errors. Cleaning reduces error rates and narrows the gap, but does not eliminate it. In addition, we find that a fixed number of training samples provides less effective coverage in Arabic due to higher visual variability, requiring more data to learn similar representations. We compare recognition across datasets in terms of the number of text lines and the number of characters, showing an equivalence trade-off. We compare character frequency distributions across scripts and show that Arabic is significantly more heavy-tailed than Latin. Our error analysis reveals that around 30 percent of substitution errors in Arabic datasets (e.g., KHATT) are caused by confusion between visually similar characters, compared to about 15 percent in Latin-script datasets such as IAM.

Abstract

Optical Character Recognition

Explore similar work

Performance Gap Analysis between Latin and Arabic Scripts HTR

AraMS-28k: The Largest Publicly Released Line-Level Dataset of Historical Arabic Manuscripts with Margin and Insertion-Anchor Annotations

When Do VLMs Help Arabic Manuscript OCR? A Cross-Dataset Study