freococo/arabic_tashkil_dataset

Arabic Tashkil (Diacritization) Dataset 📖✨

Dataset Summary

This is a massive, high-quality, Gold-Standard dataset designed explicitly for training Arabic Automatic Diacritization (Tashkil) AI models (such as ByT5, AraT5, or Custom Transformers).

The dataset contains 1,494,228 heavily vocalized pages (~2.47 GB of data) extracted from Classical Arabic and Islamic texts sourced from Thahabi.org.

To ensure the highest possible ground-truth quality, every single page in this dataset has passed a strict 80% Diacritization Threshold filter, ensuring that almost every standard base letter contains a vowel mark.

Dataset Structure

Data Fields

Each row in the dataset represents a single page from a book. The input and target pairs are pre-formatted for sequence-to-sequence AI training.

• book_id (string): Unique identifier for the source book.
• book_title (string): Title of the book.
• author (string): Name of the author.
• page_number (int64): The specific page number the text was extracted from.
• tashkil_ratio (float64): The calculated ratio of diacritics to base Arabic letters (Minimum: 0.80).
• input_text (string): The unvocalized text (AI Input). All standard Tashkil marks have been stripped using Regex.
• target_text (string): The original, perfectly vocalized text (AI Ground Truth/Label).
• __index_level_0__ (int64): Original pandas dataframe index artifact.

Dataset Statistics

• Total Rows (Pages): 1,494,228
• Total Size: 2.47 GB
• Language: Arabic (ar)

Example Instance

{
  "book_id": "1653",
  "book_title": "العلل الكبير للترمذي = ترتيب علل الترمذي الكبير",
  "author": "الترمذي، محمد بن عيسى",
  "page_number": 1,
  "tashkil_ratio": 0.852459,
  "input_text": "§باب ما جاء في فضل الطهور 1 - قال أبو عيسى الترمذي: سألت أبا عبد الله محمد بن إسماعيل البخاري عن حديث مالك بن أنس , عن زيد بن أسلم...",
  "target_text": "§بَابُ مَا جَاءَ فِي فَضْلِ الطُّهُورِ 1 - قَالَ أَبُو عِيسَى التِّرْمِذِيُّ: سَأَلْتُ أَبَا عَبْدِ اللَّهِ مُحَمَّدَ بْنَ إِسْمَاعِيلَ الْبُخَارِيَّ عَنْ حَدِيثِ مَالِكِ بْنِ أَنَسٍ , عَنْ زَيْدِ بْنِ أَسْلَمَ..."
}

Dataset Creation & Methodology

1. The 80% Threshold Rule (tashkil_ratio)

In perfectly vocalized Arabic text (such as the Holy Quran), the natural ratio of Tashkil marks to base letters rarely exceeds 0.85. This is because long vowels (حروف المد: ا، و، ي) and certain silent letters do not take diacritics.

If a dataset filters for > 0.95 diacritics, it will unnaturally discard normal sentence structures and heavily skew toward words with Shaddas (which count as two marks).

This dataset relies on an intelligent mathematical threshold:

Tashkil Count / Base Arabic Letters >= 0.80

By enforcing a minimum score of 0.80, we capture texts that are completely vocalized to the human eye while preserving the natural grammar and spelling of the Arabic language.

2. Regex Processing

The input_text is generated by passing the target_text through a Regex filter that strips standard Arabic diacritics: [\u064B-\u0652\u0670]. Punctuation, numbers, paragraph symbols (§), and spacing remain perfectly identical between the input and target texts, ensuring models do not hallucinate structural formatting.

Intended Use

This dataset is highly recommended for:

• Fine-tuning Sequence-to-Sequence models (T5, BART) for Arabic Diacritization.
• Character-level Transformers (e.g., ByT5).
• Linguistic analysis of Classical Arabic texts.

Source

• Raw Data Source: Thahabi.org and https://huggingface.co/datasets/freococo/thahabiorg_pages
• Curation & Processing: Processed using Pandas/PyArrow on Parquet files, filtering for high-density vocalization.