AraMS-28k

Large publicly released line-level dataset of historical Arabic manuscripts, with margin and insertion-anchor annotations.

AraMS-28k comprises 14 books, 3,043 pages, and 28,600 annotated text lines (27,971 main-text, 629 margin) spanning three hand-copied script traditions — Naskh, Ruq'ah, and Maghrebi — plus one lithographed printed edition. Each line is labelled as main-text or margin, and margin lines with an unambiguous attachment point are further annotated with an insertion anchor, recovering the manuscript's true non-linear reading order at line-level granularity.

This is the full annotation release. For a training-ready release of cropped line images paired with normalized text (for direct use with Kraken, HATFormer, or other HTR pipelines), see AraMS-28k-HTR.

For the build/reproduction code, see GitHub: ArchaText.

Sample Images

Below is an example page from the corpus, showing both main text lines and margin anchor annotations.

---

Dataset Statistics

Property	Value
Books	14
Pages	3,043
Main lines	27,971
Margin lines	629
Confidently anchored margin lines	189 / 629 (~30%)
Scripts	Naskh, Ruq'ah, Maghrebi (+1 lithographed volume)
Train / val / test split	9 / 2 / 3 books
License	CC BY-NC-SA 4.0

Full per-book breakdown (page-validated vs. line-validated subsets) is in the paper's Appendix A.

---

Dataset Structure

images/{book_id}/{page_id}.jpg
annotations/{book_id}.jsonl
splits/{train,val,test}_books.txt
schema/line_record.schema.json
archive/AraMS-28k.zip          # original archive, byte-identical to the checksum below

Line record schema

Each line in annotations/{book_id}.jsonl is one JSON object:

{
  "line_uid": "book_03_page_001_L000",
  "line_type": "main",
  "text": {
      "gt_raw": "...",
      "gt_normalized": "...",
      "gemini_raw": "...",
      "confidence": 100
  },
  "layout": {"insertion_anchor": null, "rotation": "horizontal"},
  "split": "train"
}

• line_type: "main" or "margin".
• text.gt_raw: fully diacritized reference transcription.
• text.gt_normalized: diacritic-stripped, normalized transcription matched to what's visually present in an undiacritized manuscript hand — recommended training target for HTR models.
• text.confidence: alignment confidence used during construction.
• layout.insertion_anchor: for margin lines with an unambiguous attachment point, the index of the main-text line it inserts after; null otherwise (no forced/unverifiable guesses).
• split: "train", "val", or "test", assigned at book level.

Full schema: schema/line_record.schema.json.

Download & Checksums

File	SHA-256
archive/AraMS-28k.zip	376c1e19bf04d678f1052505e45e805dd2753ea182cb146614a546027925fa55

Verify after download:

sha256sum archive/AraMS-28k.zip
# should match the hash above

---

License

Released under CC BY-NC-SA 4.0. Line-level annotations are original to this work. Reference transcriptions may carry independent copyright where derived from a modern critical edition — consult the datasheet for details before redistribution.

Citation

@article{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},
  journal={arXiv preprint arXiv:2608.26921},
  year={2026},
  doi={10.48550/arXiv.2608.26921}
}

Maintenance

Corrections (mis-transcribed lines, incorrect anchors, segmentation errors) are accepted via GitHub Issues and batched into versioned releases. See CHANGELOG.md on GitHub for release history.