Sanapalijo/sindhi_sentiment

Sindhi Sentiment Analysis Dataset (50k)

A balanced, three-class sentiment analysis dataset for the Sindhi language (سنڌي) in the Perso-Arabic script, containing 50,000 labeled sentences. Sindhi is a low-resource language spoken by over 30 million people, mainly in Sindh, Pakistan. This dataset supports training and benchmarking of sentiment classification models for Sindhi NLP.

Dataset Summary

• Language: Sindhi (sd), Perso-Arabic script
• Task: Sentence-level sentiment classification
• Size: 50,000 rows (all unique — zero duplicates)
• Classes: positive, negative, neutral
• Format: CSV (UTF-8)

Dataset Structure

Columns

Column	Type	Description
Text	string	A single Sindhi sentence
Label	string	Sentiment class: positive, negative, neutral

Label Distribution

Label	Count	Share
positive	16,667	33.3%
negative	16,667	33.3%
neutral	16,666	33.3%

Example Rows

Text	Label
اسان جمعي تي نوابشاهه جي لائبريري وياسين، تجربو نهايت سٺو رهيو.	positive
هالا ۾ ٿيل تقريري مقابلي جو انتظام ڏاڍو ناقص هو.	negative
بدين کان سکر تائين بس جو ڀاڙو پنج سؤ رپيا مقرر ٿيل آهي.	neutral

Dataset Creation

Methodology

The dataset was created via synthetic template-based generation assisted by an LLM (LLM synthetic generation approach). Sentence templates were combined programmatically with large slot vocabularies — places, products, media, foods, events, Sindh's cities (Karachi, Hyderabad, Sukkur, Larkana, Mirpurkhas, Umerkot, and others), days, times, and quantities — then deduplicated, balanced, and shuffled.

Sentences cover everyday domains: restaurants and food, shopping and products (electronics, household items), places and services (hospitals, schools, banks, markets, stations), media (films, dramas, novels, podcasts), events (mushairas, exhibitions, tournaments, cultural shows), transport, weather, and civic announcements.

Design decisions:

• Balanced classes to avoid class-imbalance effects during training.
• Neutral = factual/informational — schedules, announcements, prices, and descriptions with no evaluative language.
• Fully deduplicated — all 50,000 sentences are unique, verified programmatically; no sentence appears with conflicting labels.
• Shuffled with a fixed random seed for reproducibility.
• Gender-safe grammar — sentence frames anchor grammatical agreement on fixed nouns (e.g. تجربو، ڪارڪردگي، اڏاوت، انعقاد) so slotted nouns of any gender remain grammatical.

Annotation

Labels were assigned at generation time by construction: each template family encodes a single, unambiguous sentiment.

How to Use

from datasets import load_dataset

ds = load_dataset("YOUR_USERNAME/sindhi-sentiment-dataset")
print(ds["train"][0])

Train/test split example:

ds = ds["train"].train_test_split(test_size=0.2, seed=42)

Limitations

• Synthetic origin: Sentences are template/LLM-generated, not scraped from real user text, so they are cleaner than real-world Sindhi (no code-switching, slang, or spelling variation common online). Models trained solely on this data may need fine-tuning on organic text for social-media use cases.
• Template structure: Lexical variety is bounded by the template slots; the dataset is best used alongside organic corpora.
• Single script: Perso-Arabic Sindhi only; Devanagari Sindhi is not covered.
• No sarcasm/mixed sentiment: Each sentence carries one clear sentiment.

Intended Uses

• Training and benchmarking Sindhi sentiment classifiers
• Multilingual / low-resource NLP research
• Data augmentation and seed data for larger organic datasets
• Coursework and educational projects

Related Work

• alinawazmahar/Sindhi_Sentiment_dataset
• DanishMahdi/snd_movies_sentiment_analysis

License

CC BY 4.0 — free to use, share, and adapt with attribution.

Citation

@misc{sindhi_sentiment_50k_2026,
  title  = {Sindhi Sentiment Analysis Dataset (50k)},
  author = {YOUR_NAME},
  year   = {2026},
  url    = {https://huggingface.co/datasets/YOUR_USERNAME/sindhi-sentiment-dataset}
}