TheSeniorTeam/Arabic_Manuscript_Collection_Dataset

Arabic Manuscript Collection

Seven Arabic handwritten text recognition (HTR) subsets. Five are converted to one layout and one label format so they can be trained and evaluated together: 82,561 labelled images in total. Two are republished closer to their source shape: AMIDDA as upstream Parquet, and OpenITI-Makhzan as page images with line-level coordinates.

Four of the five converted sources are historical manuscripts. KHATT is modern handwriting and is included as a separate, optional subset.

AMIDDA overlaps Muharaf and KHATT. AMIDDA bundles its own copies of muharaf-public (24,494 lines) and KHATT v1.0 (6,673 lines), split independently of the Muharaf and KHATT subsets here. Training on AMIDDA and evaluating on Muharaf/KHATT test (or the reverse) leaks lines across the split boundary. Use one or the other, or filter AMIDDA on its dataset column.

Curated by TheSeniorTeam, a PSU senior project supervised by Prof. Samer Samarah. We fine-tune an open vision-language model on historical Arabic handwriting and evaluate it with CER/WER on held-out manuscripts. This repository is that project's data module.

Subsets

Subset	Train	Val	Test	Total	Unit	Content	Diacritised lines	License
Omar-al-Saleh	15,164	798	2,095	18,057	line	Handwritten memoirs, 1951–1965, 16 documents	~3%	CC BY 4.0
Muharaf	22,092	1,069	1,334	24,495	line	Historic letters, journals, poetry, church and legal records	~35%	CC BY-NC-SA 4.0
AraMs-28k-HTR	19,739	1,486	6,744	27,969	line	14 books: 13 manuscripts in Naskh, Ruq'ah and Maghrebi scripts, plus 1 lithographed edition	0%	CC BY-NC-SA 4.0
AHDQ	579	31	32	642	page	Binarized handwritten Quran pages, 6 handwriting styles	100%	MIT
KHATT	8,094	1,901	1,403	11,398	line	Modern unconstrained handwriting, 838 writers	0%	Research use (see below)
AMIDDA	53,420	6,675	6,684	66,779	line	Upstream aggregate of 8 corpora: Maghrebi and Oriental Arabic manuscripts, lithography, modern cursive	mixed	Per source (see below)
OpenITI-Makhzan_Metadata_2026-1-2	18,447	2,343	2,289	23,079	line on a page image	208 documents, 1,497 pages. Multilingual: Arabic 12,773, Persian 5,381, Ottoman Turkish 3,654, Urdu 1,055. Six scripts, led by naskh and nastaliq	27%	Unverified (see below)

For the five converted subsets, every CSV row has a matching image file, and the train, validation and test splits share no images. Two subsets differ: AMIDDA ships Parquet with embedded images and no CSV, and OpenITI-Makhzan_Metadata_2026-1-2 has many rows per image, because its images are whole pages and each row is one line inside a page (median 14 lines per page, up to 74).

AHDQ is page-level. Each image is a full page with a median of about 940 characters of text, up to 3,600. Every other subset uses line crops with a median of 47–70 characters. Batch AHDQ separately, or filter it out if your model has a short context.

Layout

<Subset>/
  images.zip          # contains images/train/, images/val/, images/test/
  train.csv           # image_path,text
  val.csv
  test.csv

AMIDDA/ does not follow this layout. It is the upstream Parquet release, images embedded:

AMIDDA/
  data/train-*.parquet, validation-*.parquet, test-*.parquet
  README.md           # upstream card: per-corpus stats and baselines

OpenITI-Makhzan_Metadata_2026-1-2/ keeps the images.zip + CSV shape but not the two-column schema. Its 22 columns carry page-level line coordinates and catalogue metadata:

Column group	Columns
Text	text
Image	image_file (a page, not a line crop), page_id, page_number
Geometry	bbox_x, bbox_y, bbox_width, bbox_height, polygon (all rows populated)
Provenance	document_id, shelfmark, line_id, line_number, source_pdf
Catalogue	lang, script, genre, form, document_type
QA	transcriber, vetter, vetting_status (every row vetted at least once)

The geometry columns make this the only subset you can use for line segmentation or layout analysis, not just recognition. To train a line recogniser on it, crop each line from its page with bbox_* or polygon first.

• image_path is a bare filename inside images/<split>/. The CSV's name already says which split it belongs to.
• text is the ground-truth transcription in Unicode Arabic, not normalized. Diacritics (tashkeel), punctuation and digits are kept exactly as the source provides them.
• Images are PNG, except KHATT, which uses JPG.

Extra files:

File	What it is
AraMs-28k-HTR/metadata.csv	Per-line provenance: book_id, page_id, line_idx, line_type, split. The img_path column predates the reorganisation, so join to the split CSVs on the filename.
AraMs-28k-HTR/split_used.json	Which books are in which split.
AraMs-28k-HTR/dataset_stats.json	Line, page and book counts per split and per book.
AHDQ/labels.csv	The original unsplit label file for all 642 pages (file_name,text).
KHATT/khatt-label-decoder.ipynb	Decodes KHATT's Latin transliteration codes into Arabic (see below).

Usage

Text only, for example to inspect labels or build a vocabulary:

from datasets import load_dataset
ds = load_dataset("TheSeniorTeam/Arabic_Manuscript_Collection_Dataset", "Muharaf")

AMIDDA is Parquet, so the same call returns its images too, no zip handling needed:

ds = load_dataset("TheSeniorTeam/Arabic_Manuscript_Collection_Dataset", "AMIDDA")

OpenITI-Makhzan_Metadata_2026-1-2 loads the same way as the CSV subsets, but remember image_file names a page and the row is one line inside it:

ds = load_dataset("TheSeniorTeam/Arabic_Manuscript_Collection_Dataset",
                  "OpenITI-Makhzan_Metadata_2026-1-2")
arabic_naskh = ds["train"].filter(lambda x: x["lang"] == "ara" and x["script"] == "naskh")

With images:

import zipfile
import pandas as pd
from huggingface_hub import hf_hub_download

repo, subset, split = "TheSeniorTeam/Arabic_Manuscript_Collection_Dataset", "Muharaf", "train"

zipfile.ZipFile(hf_hub_download(repo, f"{subset}/images.zip", repo_type="dataset")).extractall(subset)
df = pd.read_csv(hf_hub_download(repo, f"{subset}/{split}.csv", repo_type="dataset"))
df["image"] = f"{subset}/images/{split}/" + df["image_path"]

Whole repo (about 19.4 GB: AMIDDA 8.9 GB, OpenITI-Makhzan 6.1 GB, the rest 4.4 GB):

hf download TheSeniorTeam/Arabic_Manuscript_Collection_Dataset --repo-type dataset --local-dir .

Splits: which are official and which we made

Do not re-split these subsets by line. A line-level split puts the same writer or book in both train and test, so test scores measure memorisation instead of generalisation.

Subset	Train / test	Validation
Omar-al-Saleh	Official	Ours. 5% of official train, seed 0
Muharaf	Official	Official
AraMs-28k-HTR	Official and split by book: 9 train, 2 val, 3 test books	Official
AHDQ	Ours. No upstream split. 90/5/5, stratified by style so all six styles appear in every split	Ours
KHATT	Train official. Test is ours: 104 of 691 train writers held out, seed 3, split by writer	Official, writer-disjoint from train
AMIDDA	Upstream, unchanged	Upstream, unchanged
OpenITI-Makhzan_Metadata_2026-1-2	As shipped, and document-disjoint: 98 train, 54 val, 56 test documents, no document or page in two splits	As shipped

• Omar-al-Saleh: The 2,671-line blind test set from the NAKBA NLP 2026 shared task is deliberately not included, so it can't end up in training by accident. Upstream train has 15,969 rows. This copy has 15,962; the 7-row gap comes from upstream.
• AHDQ: The style is encoded in the filename prefix: dmn_, sdc_, sas_, sal_, sbi_, sfm_.
• KHATT: KHATT_v1.0 ships a test split but exposes no writer ID, so there is no way to check that it shares no writers with this train set. We held out our own test set instead.

KHATT is optional

KHATT is modern handwriting (1,000 writers originally, 838 in this copy), not manuscripts. It is here for pre-training or comparison. Leave it out if you are measuring manuscript HTR.

The upstream CSVs label each line as Latin transliteration codes, one per column (ke la sp ah na …). We decoded them into Arabic with the Arabic-Latin lookup table in the KHATT v1.0 readme, plus 4 lam-alef ligature codes the table leaves out. All 67 codes resolve, covering 100% of tokens in both upstream CSVs, so no character was guessed. khatt-label-decoder.ipynb reproduces the decoding. Upstream CSVs name images .tif, but the files are .jpg; this copy uses the real filenames.

Known limitations

• No source or document_id column yet. Document grouping is only partly available: AraMs has book_id in metadata.csv, AHDQ has the style prefix, and Muharaf and this Omar copy expose none. Until an ID column exists, you can't safely create a new split.
• Diacritics vary by subset, from 0% (AraMs, KHATT) to 100% (AHDQ). A model trained on a mix has to learn when tashkeel is expected. Report CER both with and without diacritics.
• OpenITI-Makhzan_Metadata_2026-1-2 catalogue columns are not normalised. form mixes case and has at least one corrupt value (prose, Prose, poetry, Poetry, Prose, Poetry, ProseProse, calligraphed heading). source_pdf is filled for only 23% of rows. Normalise case before grouping on these.
• A handful of Muharaf and Omar transcriptions contain a newline inside the text field, and one Muharaf train row has empty text. Use a real CSV parser, not line splitting.
• Omar-al-Saleh has a single writer, so good test scores there say little about other hands.

Licensing

There is no single license. Each subset keeps its upstream terms, and using the collection as a whole means following all of them. The two CC BY-NC-SA subsets make any combined use non-commercial.

Subset	License	Upstream
Omar-al-Saleh	CC BY 4.0	U4RASD/omar-al-saleh-manuscripts-segments (Doha Institute)
Muharaf	CC BY-NC-SA 4.0	aamijar/muharaf-public
AraMs-28k-HTR	CC BY-NC-SA 4.0	Zenodo 10.5281/zenodo.22095333
AHDQ	MIT	Kaggle
KHATT	Free for research, distributed by KFUPM. Check their terms before redistributing.	KHATT, KFUPM
AMIDDA	other, and per-source on top: 8 upstream corpora with their own terms. Check before any redistribution or commercial use.	calfa-ai/AMIDDA
OpenITI-Makhzan_Metadata_2026-1-2	Not yet verified. The upstream terms for this release have not been confirmed, and the 208 documents come from multiple holding libraries with their own shelfmarks. Treat as all-rights-reserved until checked.	OpenITI / Makhzan (to be confirmed)

Citation

If you use a subset, cite its original authors:

@misc{saeed2024muharaf,
  title         = {Muharaf: Manuscripts of Handwritten Arabic Dataset for Cursive Text Recognition},
  author        = {Saeed, Mehreen and Chan, Adrian and Mijar, Anupam and Moukarzel, Joseph and Habchi, Georges and Younes, Carlos and Elias, Amin and Wong, Chau-Wai and Khater, Akram},
  year          = {2024},
  eprint        = {2406.09630},
  archivePrefix = {arXiv},
  note          = {NeurIPS 2024}
}

@misc{guechaoui2026arams28k,
  title         = {AraMS-28k: The Largest Publicly Released Line-Level Dataset of Historical Arabic Manuscripts with Margin and Insertion-Anchor Annotations},
  author        = {Guechaoui, Mohamed and Zellagui, Mohamed Diaa and Chaib, Souleyman and Dhelim, Sahraoui},
  year          = {2026},
  eprint        = {2608.26921},
  archivePrefix = {arXiv}
}

@article{mahmoud2014khatt,
  title   = {KHATT: An Open Arabic Offline Handwritten Text Database},
  author  = {Mahmoud, Sabri A. and Ahmad, Irfan and Al-Khatib, Wasfi G. and Alshayeb, Mohammad and Parvez, Mohammad Tanvir and M{\"a}rgner, Volker and Fink, Gernot A.},
  journal = {Pattern Recognition},
  volume  = {47},
  number  = {3},
  pages   = {1096--1112},
  year    = {2014},
  doi     = {10.1016/j.patcog.2013.08.009}
}

• Omar-al-Saleh: NAKBA NLP 2026: Shared Task on Arabic Handwritten Manuscript Understanding, 2nd International Workshop, co-located with LREC 2026, Palma, Mallorca.
• AHDQ: XIDMA_AI (2024). AHDQ (Arabic Handwriting Dataset from Quran). Kaggle.
• AMIDDA: Calfa, calfa-ai/AMIDDA. Cite the eight upstream corpora listed on that card, not AMIDDA alone.