Arabic Tashkeel Dataset — Al-Maktaba Al-Shamela

A large-scale Arabic diacritization (tashkeel) dataset derived from Al-Maktaba Al-Shamela (المكتبة الشاملة), a comprehensive digital library of classical Islamic texts. The dataset pairs undiacritized Arabic sentences with their fully diacritized equivalents, enabling training and evaluation of automatic tashkeel systems.

Dataset Summary

Split	Examples
train	3,183,238
validation	397,904
test	397,906
Total	3,979,048

• Languages: Arabic (classical / Islamic corpus)
• Source: Al-Maktaba Al-Shamela HTML exports
• Format: JSONL (one record per line)
• Uncompressed size: ~1.29 GB

Dataset Structure

Data Fields

Field	Type	Description
raw	string	Arabic sentence with all diacritics stripped
tashkeel	string	Fully diacritized version of the same sentence
ratio	float	Fraction of Arabic letters that carry a diacritic mark (always ≥ 0.90)

Example

{
  "raw": "وقد أخرج مالك في الموطأ هذه الزيادة عن الزهري أيضا،",
  "tashkeel": "وَقَدْ أَخْرَجَ مَالِكٌ فِي الْمُوَطَّأِ هَذِهِ الزِّيَادَةِ عَنِ الزُّهْرِيِّ أَيْضًا،",
  "ratio": 0.878
}

Data Splits

The dataset is split 80 / 10 / 10 using a fixed random seed (42) for reproducibility:

Split	File	Rows
train	train.jsonl	3,183,238
validation	val.jsonl	397,904
test	test.jsonl	397,906

Dataset Creation

Source Data

All text originates from Al-Maktaba Al-Shamela, an authoritative digital collection covering classical Islamic jurisprudence, hadith sciences, Quranic exegesis, Arabic linguistics, history, and related disciplines. Books were exported as HTML and processed with a custom extraction pipeline.

Extraction & Filtering

Sentences are extracted from HTML exports, then filtered according to:

• Minimum tashkeel ratio: 0.90 — at least 90 % of Arabic base letters must carry a diacritic mark, ensuring high annotation density.
• Minimum length: 20 Arabic letters per sentence.
• Maximum length: 400 characters per sentence (longer sentences are sub-segmented on ، / ; boundaries).
• Deduplication: exact-match on the diacritized string; duplicate sentences are dropped.

Diacritics covered include all standard Arabic harakat: fatha (َ), damma (ُ), kasra (ِ), fathatan (ً), dammatan (ٌ), kasratan (ٍ), sukun (ْ), shadda (ّ), and their combinations.

Sample Statistics (train split, first 10 k sentences)

Metric	Value
Average raw length	~57 characters
Average tashkeel ratio	~0.90
Min / max raw length	23 / 247 characters

Usage

Loading with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("yuujiElfahkrany/tashkeel")
print(ds["train"][0])
# {'raw': '...', 'tashkeel': '...', 'ratio': 0.867}

Manual loading

import json

def load_jsonl(path):
    with open(path, encoding="utf-8") as f:
        return [json.loads(line) for line in f]

train = load_jsonl("train.jsonl")
val   = load_jsonl("val.jsonl")
test  = load_jsonl("test.jsonl")

Character-level label scheme

The companion dataset.py maps each Arabic base character to one of 15 diacritic classes for sequence-labeling models:

Label	Diacritic
0	none (no diacritic)
1	fatha (َ)
2	damma (ُ)
3	kasra (ِ)
4	fathatan (ً)
5	dammatan (ٌ)
6	kasratan (ٍ)
7	sukun (ْ)
8	shadda only (ّ)
9	shadda + fatha (َّ)
10	shadda + damma (ُّ)
11	shadda + kasra (ِّ)
12	shadda + fathatan (ًّ)
13	shadda + dammatan (ٌّ)
14	shadda + kasratan (ٍّ)

Intended Uses & Limitations

Intended uses:

• Training and benchmarking automatic Arabic diacritization (tashkeel) models.
• Pre-training or fine-tuning Arabic language models on classical text.
• Linguistic research into classical Arabic morphology and orthography.

Limitations:

• The corpus is dominated by classical and pre-modern Islamic texts; performance on modern Standard Arabic or dialectal Arabic may be lower.
• Tashkeel annotations are those present in the original Shamela sources. Occasional OCR or editorial errors in the source library may propagate to this dataset.
• Very short sentences (< 20 Arabic letters) and very long ones (> 400 characters) are excluded.

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license, in keeping with the terms of Al-Maktaba Al-Shamela, which distributes its library as a charitable Islamic endowment (waqf) — free for non-commercial use. Commercial use of this dataset is not permitted.

Citation

If you use this dataset, please cite:

@dataset{shamela_tashkeel_2026,
  title     = {Arabic Tashkeel Dataset from Al-Maktaba Al-Shamela},
  year      = {2026},
  language  = {Arabic},
  note      = {Extracted from Al-Maktaba Al-Shamela (https://shamela.ws)},
}

Acknowledgements

All classical texts are sourced from Al-Maktaba Al-Shamela (الشاملة), which provides free access to thousands of classical Arabic and Islamic works.