israel/test

AfriGuard: Safety Evaluation Data for African Languages

AfriGuard is a human-annotated safety dataset covering 10 African languages: Amharic, Hausa, Igbo, Oromo, Shona, Swahili, Twi, Wolof, Yoruba, and Zulu. Each example contains a culturally grounded scenario and prompt/response pair in English and the target language, labeled with majority-vote annotations (safety category and safe/unsafe) from three native-speaker annotators.

Paper: AfriGuard: Safety and Robust Refusal For African Languages

Pipeline

Splits

Each language config provides three splits:

Split	Size (per language)	Description
train	~950–980 rows	50% of scenario-level data
validation	20 rows	Few-shot pool: 10 categories × {safe, unsafe}. The first 10 rows cover all 10 classes (5 safe / 5 unsafe); rows 11–20 cover all 10 classes with the opposite label.
test	~980–1000 rows	50% of scenario-level data. 100% of test scenario_ids are unseen in train.

Split properties:

• Train/test split at the scenario level: every test scenario is fully unseen in train.
• Scenarios are shared across languages of the same region (Amharic+Oromo share ET scenarios; Hausa+Igbo+Yoruba share NG scenarios). Scenario assignment is made globally per language group, so there is no cross-language train/test leakage.

Categories

1. Criminal Activity & Exploitation
2. Political & Societal Manipulation
3. Violence & Physical Harm
4. Systemic Risks
5. Sexual & Exploitative Content
6. Unsafe or High-Risk Advice
7. Privacy & Data Misuse
8. Hate, Harassment & Discrimination
9. Ethical & Moral Risk
10. Intellectual Property & Content Integrity

Fields

All language configs share the same schema:

Column	Description
id	Unique example id
scenario_id	Scenario grouping id
scenario	Full scenario description (context, severity levels, example prompts, AI safety application)
country	Country context of the scenario
target_language	Target language code
prompt / response	English prompt and model response
prompt_translated / response_translated	Final human-verified translations
majority_vote_class	Majority-vote safety category from 3 annotators
majority_vote_safe_unsafe	Majority-vote safe/unsafe label from 3 annotators

Usage

from datasets import load_dataset

ds = load_dataset("israel/AfriGuard", "amharic")
train, val, test = ds["train"], ds["validation"], ds["test"]

# Few-shot: the first 10 validation rows cover all 10 categories (5 safe / 5 unsafe)
fewshot_10 = val.select(range(10))

Citation

@misc{azime2026afriguardsafetyrobustrefusal,
      title={AfriGuard: Safety and Robust Refusal For African Languages}, 
      author={Israel Abebe Azime and Jesujoba O. Alabi and Iffat Maab and Alexander D. Cai and Tadesse Destaw Belay and Atnafu Lambebo Tonja and Jonas Kgomo and Folasade Alabi and Salomey Osei and Saminu Muhammad Aliyu and Nkechinyere Faith Aguobi and Bontu Fufa Balcha and Blessing Sibanda and Davis David and Mouhamadane Mboup and Daud Abolade and Neo Putini and Louis Powell and Kiant'e Brantley and Philipp Slusallek and Dietrich Klakow and Julia Kreutzer},
      year={2026},
      eprint={2610.26474},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.26474}, 
}