---
license: other
license_name: mixed-public-sources
task_categories:
- text-generation
- fill-mask
language:
- ur
tags:
- urdu
- roman-urdu
- corpus
- pretraining
pretty_name: Ali Data — Urdu & Roman Urdu Corpus
---

# Ali Data — Urdu & Roman Urdu Corpus

A large public-domain/open collection of **Urdu (اردو)** and **Roman Urdu** text for training Urdu language models.

## Stats
- **21,846,033 unique rows** (deduplicated by exact text)
- Format: JSON Lines — each line: `{"text": "...", "source": "..."}`
- Scripts: Urdu (Arabic script) + Roman Urdu (Latin script)

## Sources
| Source | Rows (approx) |
|---|---|
| Hugging Face open datasets (news, QA, sentiment, translation, poetry, instructions) | ~17.5M |
| FineWeb-2 `urd_Arab` (web crawl, Urdu script) | ~3.3M |
| GitHub open corpora (poetry, SMS, reviews) | ~1.4M |
| FineWeb-2 `urd_Latn` (web crawl, Roman Urdu) | ~70K |
| Kaggle open datasets (news, poetry, reviews) | ~23K |
| BBC Urdu articles + Internet Archive Urdu books | ~160 |

## Usage
```python
from datasets import load_dataset
ds = load_dataset("PAILLM/ali-data", split="train")
# or stream it:
ds = load_dataset("PAILLM/ali-data", split="train", streaming=True)
```

## Notes
- Raw public collection; clean/filter for your training needs (language ID, quality, toxicity) before use.
- Respects each source's original license; compiled for research use.