---
language:
  - ar
size_categories:
  - 100K-1M
annotations_creators:
  - machine-generated
source_datasets:
  - original
tags:
  - newspaper
  - news
  - arabic
  - egypt
---

# Shorouk Archive

## Dataset summary

Article-level text archive extracted from daily PDF issues of the Egyptian newspaper **الشروق** (Al‑Shorouk).

- Rows: **473,949** articles
- Date range: **2009-02-01 → 2025-11-03**
- Language: Arabic (`ar`)
- File: `shorouk_articles.parquet` (Snappy-compressed Parquet, ~754 MB)

This dataset is machine-generated from the PDFs (LLM-assisted structuring). Expect OCR/text-layer noise and occasional article boundary errors.

Maintainer: **Khaled Eltokhy** (`eltokh7`)

## Dataset structure

This repository contains a single Parquet table with these columns:

| Column | Type | Description |
|---|---:|---|
| `date` | string | Issue date (`YYYY-MM-DD`) |
| `page` | int32 | Page number within the issue |
| `source` | string | Constant: `الشروق` |
| `title` | string | Article title/headline |
| `author` | string | Author name if present, else empty string |
| `content` | string | Full article body text |
| `word_count` | int32 | Approximate word count (`len(content.split())`) |

There are no predefined splits (single table).

## How this dataset was created (PDF → Parquet)

This dataset was produced from a local pipeline (see the `pipeline/` folder in the source project):

1. **PDF → TXT**: extract each PDF’s embedded text layer using PyMuPDF, writing `--- Page N ---` markers.
2. *(Optional)* **PDF → layout JSON**: detect layout blocks (titles/paragraphs/etc.) and their bounding boxes.
3. **TXT(+layout) → per-page JSON**: LLM-assisted conversion of each page into structured JSON with `articles[]` (title/author/content).
4. **Per-page JSON → per-day JSON**: merge all pages for a date into one daily JSON file.
5. **Daily JSON → Parquet**: flatten all articles into the Parquet table published here.

## Usage

### Pandas

```python
import pandas as pd

df = pd.read_parquet("shorouk_articles.parquet")
```

### Hugging Face `datasets`

```python
from datasets import load_dataset

ds = load_dataset("parquet", data_files="shorouk_articles.parquet")["train"]
```

### DuckDB

```sql
SELECT date, COUNT(*) AS n_articles
FROM read_parquet('shorouk_articles.parquet')
GROUP BY 1
ORDER BY 1;
```

## Known limitations / notes

- **Parquet viewer limits**: the hosted “Data Studio”/SQL viewer may fail on large scans depending on row-grouping. If you need the viewer to work reliably, re-export and re-upload the Parquet with smaller row groups.
- **Quality**: article boundaries and cleanup are machine-generated and may contain errors (merged/split articles, missing authors, residual OCR artifacts).
- **Coverage**: not every issue/page in the source archive is guaranteed to be present.

## Licensing

This repository contains newspaper text. Please ensure you have the right to use and redistribute this content in your intended setting.

## Citation

If you use this dataset, cite the Hugging Face dataset page and include the access date, for example:

```bibtex
@dataset{shorouk_archive,
  title        = {Shorouk Archive},
  author       = {Khaled Eltokhy},
  publisher    = {Hugging Face},
  howpublished = {\\url{https://huggingface.co/datasets/eltokh7/shorouk-archive}},
  year         = {2026}
}
```