omanyasa/shona-hate-speech

Dataset Card for Balanced Shona Hate Speech Dataset

Dataset Summary

This dataset contains 2,000 balanced examples of Shona text classified into four categories: NEUTRAL, OFFENSIVE, CONTEXTUAL, and HATE.

Data Sources

Label	Source	Count
NEUTRAL	Literary novel (Imbwa Yemunhu by Ignatius T. Mabasa)	500
OFFENSIVE	Synthetic template-based generation	500
CONTEXTUAL	Synthetic (quoted hate speech, not endorsed)	500
HATE	Synthetic (direct attacks on protected groups)	500

Dataset Structure

Data Fields

• id (string): Unique identifier
• text (string): The Shona text content
• label (string): One of NEUTRAL, OFFENSIVE, CONTEXTUAL, HATE
• target_group (string): Protected group targeted (if applicable)
• is_quoted (bool): Whether the text quotes harmful language
• rationale (string): Explanation for the label

Label Definitions

Label	Definition
NEUTRAL	No harmful or abusive content
OFFENSIVE	Insults toward individuals or non-protected groups
CONTEXTUAL	Quoted hate speech, not endorsed
HATE	Direct attack on protected group (ethnicity, religion, gender, disability, nationality)

Bias, Risks, and Limitations

Known Limitations

⚠️ This dataset is synthetic. All OFFENSIVE, CONTEXTUAL, and HATE examples were generated using simple templates, not collected from real‑world Shona social media.

⚠️ No real hate speech. The dataset contains no actual user‑generated hate speech. Models trained on this data may not generalise to real‑world online content.

⚠️ Literary neutral data. NEUTRAL examples come from a novel (literary Shona), not conversational Shona.

⚠️ No expert validation. Labels were not verified by native speakers.

Citation

If you use this dataset, please cite:

@misc{shona_hate_dataset,
  author = {Omanyasa},
  title = {Balanced Shona Hate Speech Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/omanyasa/shona-hate-speech}
}

Licensing Information

This dataset is released under the Apache License 2.0.