ReySajju742/shaistagi_clean

Shaistagi (شائستگی) Clean Urdu Mega-Dataset

The largest and most comprehensive cleaned Urdu NLP dataset collection

---

📊 Dataset Overview

Shaistagi Clean is one of the most comprehensive, multi-task Urdu NLP collections available. It aggregates high-quality, cleaned data for pre-training, instruction-tuning, and specialized downstream tasks.

Key Statistics

Metric	Value
Total Rows	~16.2 Million
Total Tokens	~1.22 Billion
Total Characters	~1.73 Billion
Total Configurations	37
Total Parquet Files	287

---

📦 Dataset Composition & Percentages

The dataset is categorized into functional blocks:

Category	Estimated %	Configurations
LLM Pre-training	65%	fineweb_pretrain, gemma_pretrain, generic_train_ur, tiny_stories, c4, cc100, hplt, cleaned_data, news_1m
Translation (NMT)	15%	english_urdu_translation, parliament_translation, nmt, nmt_parquet
Classification	10%	imdb_reviews_ur, sentiment, sentiment_v1_ur, urdu_sentiment_local, urdu_sarcasm, roman_urdu_toxicity
Specialized/Structured	7%	addresses (982k+ rows), urdu_tts_transcription, wikiann_ur, xnli_ipa, urdu_dictionary
Reasoning & Instruction	3%	urdu_reasoning, reasoning, reasoning_parquet, urdu_instruct_alpaca, error_correction
Poetry	~1%	iqbal_poetry, organized_poetry_csv, poetry_by_poet, poetry_csv_main, urdu_poetry_general

---

📋 Detailed Configuration Statistics

Config	Rows	Tokens	Avg Tokens/Row	Description
addresses	982,837	33.2M	33.78	Urdu/Roman Urdu address mappings
english_urdu_translation	7,057,673	115.6M	16.39	Parallel EN-UR translations
error_correction	600,000	108.3M	180.43	Text error correction pairs
fineweb_pretrain	100,000	266.4M	2664.03	Long-form pretraining text
gemma_pretrain	245,153	200.3M	817.15	Gemma-formatted instruction data
generic_train_ur	3,731	253K	67.85	Urdu headlines with labels
imdb_reviews_ur	10,000	12M	1201.53	IMDB reviews in Urdu
iqbal_poetry	10,002	316K	31.65	Allama Iqbal poetry
organized_poetry_csv	17,609	1.3M	73.52	Organized poetry with labels
parliament_translation	6,374,673	363.7M	57.06	Urdu/Roman transliteration
poetry_by_poet	1,314	968K	737.22	Poetry organized by poet
poetry_csv_main	17,609	1.3M	73.52	Poetry collection
sentiment	83,309	-	-	Roman Urdu sentiment
sentiment_v1_ur	987	78K	79.61	Urdu tweets sentiment
tiny_stories	357,900	73.6M	205.63	Children's stories in Urdu
urdu_poetry_general	1,323	679K	513.23	General Urdu poetry
urdu_reasoning	800	110K	137.92	Math/reasoning problems
urdu_sarcasm	19,949	1.7M	85.76	Sarcasm detection
urdu_sentiment_local	20,834	4.4M	213.09	Sentiment/toxicity
urdu_tts_transcription	4,306	314K	73.06	TTS transcription
wikiann_ur	21,972	773K	35.21	Named Entity Recognition
xnli_ipa	400,202	30.9M	77.29	Natural Language Inference

---

🔍 What This Dataset Includes

1. 📚 Large-Scale Pre-training Data

Diverse Urdu web text from multiple sources (C4, CC100, HPLT, FineWeb) and synthetic data (Tiny Stories) to help models learn Urdu syntax and semantics.

2. 🏠 Structured Urdu Addresses

Nearly 1 million rows of Urdu-Roman Urdu address mappings, essential for logistics and geolocation models.

3. 💭 Sentiment & Nuance

Benchmark datasets including IMDB Urdu, Urdu Sarcasm, and multiple sentiment datasets for detecting emotional tone and figurative language.

4. 🌐 Cross-Lingual NLI (xnli_ipa)

Premises and hypotheses in Urdu for Natural Language Inference tasks (entailment, contradiction, neutral).

5. 📜 Poetry Collections

Multiple poetry datasets including Allama Iqbal's works, organized by poet, and general Urdu poetry.

6. 🔤 Named Entity Recognition (wikiann_ur)

Token-level NER annotations for identifying persons, locations, and organizations.

7. 🧠 Reasoning & Instruction Data

Math problems, reasoning tasks, and Alpaca-format instruction data in Urdu.

---

🚀 Quick Start

from datasets import load_dataset

# Load the default configuration (cleaned_data - largest)
ds = load_dataset("ReySajju742/shaistagi_clean")
print(ds['train'][0])

# Load specific configurations
addresses = load_dataset("ReySajju742/shaistagi_clean", "addresses")
poetry = load_dataset("ReySajju742/shaistagi_clean", "iqbal_poetry")
sentiment = load_dataset("ReySajju742/shaistagi_clean", "sentiment")
translation = load_dataset("ReySajju742/shaistagi_clean", "english_urdu_translation")

# Load web crawl data for pretraining
c4_data = load_dataset("ReySajju742/shaistagi_clean", "c4")
hplt_data = load_dataset("ReySajju742/shaistagi_clean", "hplt")

---

📁 Available Configurations

Click to expand all 37 configurations

Pre-training Data

• fineweb_pretrain - FineWeb Urdu subset
• gemma_pretrain - Gemma-formatted data
• generic_train_ur - Generic training data
• tiny_stories - Urdu children's stories
• c4 - C4 Urdu subset (19 files)
• cc100 - CC100 Urdu subset (12 files)
• hplt - HPLT web crawl (34 files)
• cleaned_data - Main cleaned data (139 files)
• news_1m - 1M news articles

Translation

• english_urdu_translation - EN-UR parallel corpus
• parliament_translation - Parliamentary translations
• nmt - Neural MT data
• nmt_parquet - NMT in parquet format

Classification & Sentiment

• imdb_reviews_ur - IMDB reviews
• sentiment - General sentiment
• sentiment_v1_ur - Urdu tweets
• urdu_sentiment_local - Local sentiment
• urdu_sarcasm - Sarcasm detection
• roman_urdu_toxicity - Toxicity detection

Poetry

• iqbal_poetry - Allama Iqbal
• organized_poetry_csv - Organized poetry
• poetry_by_poet - By poet name
• poetry_csv_main - Main poetry CSV
• urdu_poetry_general - General poetry

Structured & Specialized

• addresses - Address mappings
• urdu_tts_transcription - TTS data
• wikiann_ur - NER annotations
• xnli_ipa - NLI data
• urdu_dictionary - Dictionary entries
• urdu_instruct_alpaca - Alpaca instructions

Reasoning

• urdu_reasoning - Reasoning tasks
• reasoning - General reasoning
• reasoning_parquet - Reasoning parquet
• error_correction - Error correction

Other

• roman_urdu - Roman Urdu text
• local - Local sources
• mendeley - Mendeley data

---

📄 License

This dataset is released under the Apache 2.0 License.

---

🙏 Citation

If you use this dataset in your research, please cite:

@dataset{shaistagi_clean_2026,
  author = {ReySajju742},
  title = {Shaistagi Clean: Comprehensive Urdu NLP Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/ReySajju742/shaistagi_clean}
}

---

📧 Contact

For questions or feedback, please open an issue on the dataset repository.