metadata

language:
  - ar
size_categories:
  - 1M<n<10M
task_categories:
  - text-generation
dataset_info:
  features:
    - name: filename
      dtype: string
    - name: cleaned_sentence
      dtype: string
  splits:
    - name: train
      num_bytes: 924672739
      num_examples: 1042698
    - name: test
      num_bytes: 1789952
      num_examples: 2485
  download_size: 354596619
  dataset_size: 926462691
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

📚 Sadeed Tashkeela Arabic Diacritization Dataset

The Sadeed dataset is a large, high-quality Arabic diacritized corpus optimized for training and evaluating Arabic diacritization models.
It is built exclusively from the Tashkeela corpus for the training set and a refined version of the Fadel Tashkeela test set for the test set.

Dataset Overview

Training Data:
- Source: Cleaned version of the Tashkeela corpus (original data is ~75 million words, mostly Classical Arabic, with ~1.15% Modern Standard Arabic).
- Total Examples: 1,042,698
- Total Words: ~53 million
Testing Data:
- Source: Corrected version of the Fadel Tashkeela test set, addressing inconsistencies in handling iltiqā` as-sākinayn (adjacent consonants without intervening vowels).
- Total Examples: 2485 samples
- Notes: The test set is refined for phonological consistency according to standard Arabic rules.
Features:
- Fully normalized Arabic text
- Minimal missing diacritics
- Chunked into coherent samples (50–60 words)
- Designed to preserve syntactic and contextual dependencies

Preprocessing Details

1. Text Cleaning

Diacritization Corrections:
- Unified diacritization style.
- Corrected diacritization of frequent errors.
- Resolved inconsistencies in iltiqa' assakinayn (consonant cluster rules) based on standard phonological rules.
Normalization:
- Applied a comprehensive preprocessing pipeline inspired by Kuwain, preserving non-Arabic characters and symbols.
Consistency:
- Additional normalization steps to ensure stylistic and grammatical consistency across the dataset.

2. Text Chunking

Segmented into samples of 50–60 words each.
Used a hierarchical strategy prioritizing natural linguistic breaks, focusing on stronger punctuation first (e.g., sentence-ending punctuation, then commas).
Designed to preserve the syntactic and contextual coherence of text chunks.

3. Dataset Filtering

Missing Diacritics:
- Excluded samples with more than two undiacritized words.
Partial Diacritics:
- Removed samples with three or more partially diacritized words.
Test Set Overlap:
- Eliminated overlapping examples with the Fadel Tashkeela test set, reducing overlap to only 0.4%.

Usage

This dataset is ideal for:

Training: Using the cleaned Tashkeela corpus to train models for Arabic diacritization.
Testing: Evaluating diacritization systems with the refined Fadel Tashkeela test set.
Arabic NLP tasks that require fully vocalized texts.

Evaluation Code

The evaluation code for this dataset is available at: https://github.com/misraj-ai/Sadeed

Citation

If you use this dataset, please cite:

@misc{aldallal2025sadeedadvancingarabicdiacritization,
      title={Sadeed: Advancing Arabic Diacritization Through Small Language Model}, 
      author={Zeina Aldallal and Sara Chrouf and Khalil Hennara and Mohamed Motaism Hamed and Muhammad Hreden and Safwan AlModhayan},
      year={2025},
      eprint={2504.21635},
      archivePrefix={arXiv},
      url={https://huggingface.co/papers/2504.21635}, 
}

License

This dataset is distributed for research purposes only.
Please review the original Tashkeela corpus license for terms of use.