Datasetpublic

biraj-bhusal/rakshak-nepali-toxicity-final

🛡️ RakshakAI — Augmented Nepali Toxicity Dataset The full augmented training dataset used to train the RakshakAI toxicity detection models. Contains 4,716 samples expanded from the curated 1,574 sample dataset through back-translation augmentation via English and Hindi as intermediate languages. For the clean curated dataset only, see rakshak-all-data-combined. 📄 Paper: RakshakAI: Multi-Label Toxicity Detection for Low-Resource Nepali Social Media Content Why this… See the full description on the dataset page: https://huggingface.co/datasets/biraj-bhusal/rakshak-nepali-toxicity-final.

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

0likes41downloads

Dataset Card

🛡️ RakshakAI — Augmented Nepali Toxicity Dataset

The full augmented training dataset used to train the RakshakAI toxicity detection models. Contains 4,716 samples expanded from the curated 1,574 sample dataset through back-translation augmentation via English and Hindi as intermediate languages. For the clean curated dataset only, see rakshak-all-data-combined.

📄 Paper: RakshakAI: Multi-Label Toxicity Detection for Low-Resource Nepali Social Media Content

Why this dataset exists

Nepali is spoken by over 30 million people yet has almost no publicly available resources for toxicity detection. This dataset is the full training corpus used to build RakshakAI — the first multi-label toxicity detection system for Nepali social media.

Dataset Details

—Total samples: 4,716
—Labels: Multi-label (five binary toxicity labels + severity score)
—Languages: Devanagari Nepali, Romanized Nepali, Code-mixed Nepali-English
—Source: Facebook and YouTube comments + synthetic generation + back-translation augmentation

How it was built

1.184 samples manually collected from Facebook (75%) and YouTube (25%)
2.~1,390 samples synthetically generated using ChatGPT and Gemini based on the 184 gold-standard samples
3.Full 1,574 sample dataset expanded 3x through back-translation augmentation via English and Hindi as intermediate languages — producing natural paraphrases while preserving original labels
4.Final corpus: 4,716 samples

Label Schema

Each sample is annotated with five binary labels (0 or 1):

Column	Description
`hate_speech`	Dehumanizing language or slurs targeting a group based on identity
`casteism`	Caste-based discrimination, slurs, or untouchability references
`religous_incitement`	Content attacking a religion or inciting violence against religious groups
`political_defamation`	False accusations or defamatory content targeting political figures
`cyberbullying`	Personal attacks, threats, or harassment targeting specific individuals
`severity`	Raw severity score on a 1–5 scale (1=clean, 5=extreme violence)

Labels are not mutually exclusive — a single post can carry multiple labels simultaneously.

Label Distribution

Category	Count	%
Clean	1,699	36.0%
Hate Speech	1,427	30.3%
Cyberbullying	1,137	24.1%
Political Defamation	810	17.2%
Casteism	507	10.7%
Religious Incitement	417	8.8%

Percentages sum to more than 100% due to multi-label annotations.

Limitations

—Only 184 samples are manually collected gold-standard
—Augmented samples are paraphrases and may not capture full real-world variety
—All annotations performed by a single annotator
—Moderate severity class is underrepresented (9.3% of samples)

Related Resources

—📦 Curated Dataset (1,574 samples)
—🤖 Label Classifier — rakshak-xlmr-large-v2
—🤖 Severity Classifier — rakshak-severity-v2
—🖥️ Live Demo — RakshakAI

Citation

bibtex

@misc{bhusal2025rakshak,
  author = {Bhusal, Biraj},
  title = {RakshakAI: Multi-Label Toxicity Detection for Low-Resource Nepali Social Media Content},
  year = {2025},
  publisher = {Zenodo},
  url = {https://zenodo.org/records/20850923}
}

Content Warning

This dataset contains toxic, offensive, and abusive language in Nepali collected for research purposes. The author does not endorse any of the views expressed in the dataset.

Created by: Biraj Bhusal