---
dataset_info:
  features:
    - name: text
      dtype: string
    - name: book_title
      dtype: string
    - name: author
      dtype: string
    - name: category
      dtype: string
    - name: section_title
      dtype: string
    - name: chars
      dtype: int64
    - name: book_id
      dtype: string
  splits:
    - name: train
      num_bytes: ~
      num_examples: 47546
    - name: validation
      num_bytes: ~
      num_examples: 2642
    - name: test
      num_bytes: ~
      num_examples: 2642
  download_size: ~
  dataset_size: ~
---

# Hindawi Arabic Books — Sections Dataset

A cleaned, section-level dataset of Arabic books from [Hindawi.org](https://www.hindawi.org), prepared for NLP training and research.

## Dataset Description

- **Language:** Arabic
- **Level:** Section-level (each row = one book section)
- **Text format:** Single-line, cleaned Arabic text
- **Total sections:** 52,830
- **Total books:** 3,265
- **Total characters:** 843,868,321
- **Mean chars / section:** 15,973.3
- **Median chars / section:** 7,915.0
- **Min chars / section:** 250
- **Max chars / section:** 723,964
- **Sections < 500 chars:** 1,593 (3.02%)

## Source

Original books scraped from [Hindawi.org](https://www.hindawi.org), a non-profit foundation providing free Arabic books. The dataset covers categories including literature, philosophy, history, science, psychology, and more.

## Cleaning Pipeline

1. **Scraped** book content section-by-section from Hindawi.org
2. **Removed English / Latin** text and characters
3. **Removed tashkeel** (Arabic diacritics: َ ُ ِ ْ ّ etc.)
4. **Removed tatweel** (kashida: ـ)
5. **Normalized Arabic letters:**
   - `أ` `إ` `آ` `ٱ` → `ا`
   - `ى` → `ي`
   - `ة` → `ه`
6. **Deduplicated** exact duplicate lines and near-duplicate paragraphs (فرهس)
7. **Replaced Eastern Arabic numerals** (`٠١٢٣٤٥٦٧٨٩` and `۰۱۲۳۴۵۶۷۸۹`) with Western `0-9`
8. **Collapsed** each section to a single line (newlines → spaces)
9. **Filtered** sections with fewer than 250 characters
10. **Cross-book deduplication:** removed 342 sections that were >95% similar to sections in other books (3,253 duplicate pairs found, mean similarity 99.79%)

## Splits

| Split      | Rows    | Percentage |
|------------|---------|------------|
| train      | 47,546 | 90.0% |
| validation | 2,642 | 5.0% |
| test       | 2,642 | 5.0% |

Split was done with `train_test_split` from scikit-learn with `random_state=42` and `shuffle=True`.

## Columns

| Column | Type | Description |
|--------|------|-------------|
| `text` | string | Cleaned Arabic text of the section (single line) |
| `book_title` | string | Title of the parent book |
| `author` | string | Author name |
| `category` | string | Book category (e.g., أدب, فلسفة, تاريخ, علم نفس) |
| `section_title` | string | Title of the section within the book |
| `chars` | int64 | Character count of the section text |
| `book_id` | string | Hindawi book ID (when available) |

## Category Distribution (top 10)

| Category | Sections |
|----------|----------|
| أدب | 10,441 |
| تاريخ | 6,383 |
| روايات | 6,373 |
| فلسفة | 4,424 |
| شعر | 3,947 |
| سير الأعلام | 2,554 |
| علوم اجتماعية | 2,229 |
| علوم | 1,916 |
| نقد أدبي | 1,903 |
| سياسة | 1,723 |

## Usage Example

```python
from datasets import load_dataset

dataset = load_dataset("your-username/hindawi-arabic-sections")

# Access train split
print(dataset["train"][0]["text"][:200])

# Filter by category
history = dataset["train"].filter(lambda x: x["category"] == "تاريخ")
```

## Loading with Pandas

```python
import pandas as pd

train = pd.read_parquet("train.parquet")
print(train.head())
```

## Statistics

```
Total sections:     52,830
Total books:        3,265
Total characters:   843,868,321
Mean chars/section: 15,973.3
Median chars/section: 7,915.0
Min chars:          250
Max chars:          723,964
Std dev:            ~26,000
```

## Citation

```bibtex
@dataset{hindawi_arabic_sections,
  title = {Hindawi Arabic Books — Sections Dataset},
  year = {2026},
  publisher = {Hindawi Foundation},
  url = {https://www.hindawi.org}
}
```

## License

The original books are made available by [Hindawi Foundation](https://www.hindawi.org) under their free-access agreement. This dataset is a derived work prepared for research and educational purposes.

## Acknowledgments

- [Hindawi Foundation](https://www.hindawi.org) for making these books freely available
- This dataset was prepared using automated scraping and cleaning tools