The authors built AfriGuard through a human-in-the-loop process, creating paired safe and unsafe prompts based on country-specific scenarios, annotated for harm intent, safety categories, and appropriate refusal behavior, in 10 African languages and English.
AfriGuard safety dataset for 10 African languages shows guardrails lag general LLMs
A human-in-the-loop dataset with paired safe and unsafe prompts and fine-tuned models that beat existing refusal detectors by 43.9%.
AI Startup
Israel Abebe Azime · Jesujoba O. Alabi · Iffat Maab · Alexander D. Cai · Tadesse Destaw Belay · Atnafu Lambebo Tonja · +16 more
Kempner Institute, Harvard University · Instituto Politécnico Nacional · University College London · Equiano Institute · University of Ilorin
Research Digest··2 min read
The authors introduce AfriGuard, a culturally grounded multilingual dataset covering safe and unsafe prompts, harm categories, and refusal behavior in 10 African languages plus English.
§