metadata
license: mit
language: - sd license: mit task_categories: - text-classification task_ids: - sentiment-classification tags: - sentiment-analysis - sindhi - low-resource pretty_name: Sindhi Sentiment Dataset (50k) size_categories: - 10K<n<100K
Sindhi Sentiment Dataset (50k)
A Sindhi-language text sentiment classification dataset with 50,000 examples across three balanced classes: positive, negative, and neutral.
Dataset Summary
| Label | Count |
|---|---|
| positive | 16,667 |
| negative | 16,667 |
| neutral | 16,666 |
- Columns:
Text(Sindhi sentence),Label(positive/negative/neutral) - Average text length: ~47 characters (range 14–105)
- No duplicate rows
Dataset Structure
Text,Label
مونکي پيپر ٽائونس ڏسڻ جو ڪو پروگرام نه هو، پر مون کي محبت پسند آهي، شايد ڏسان.,positive
Source and Construction
This dataset combines two parts:
- Original 33,000 rows — short, tweet-style Sindhi sentences (referencing everyday topics, sports, entertainment, and current events) labeled by sentiment.
- 17,000 augmented rows — synthetically generated using a template + slot-filling approach to expand coverage while preserving the tone, sentence length, and topical style of the original data (everyday life, weather, food, family, local events, festivals, community issues, sports, and general news). Generic placeholders (e.g. "my friend," "our team") were used instead of real people or organizations.
Note: The augmented portion is synthetically generated, not scraped or human-authored. If you use this dataset for benchmarking or train/test evaluation, consider whether template-based examples should be isolated from your test split to avoid near-duplicate leakage across similar sentence templates.
Considerations
- Released under the MIT license.
- This dataset is intended for research and educational use in low-resource NLP (Sindhi sentiment analysis).
Loading
from datasets import load_dataset
dataset = load_dataset("csv", data_files="sindhi_sentiment_dataset.csv")