HausaSafeLM: Safety Alignment for Low-Resource Hausa Language
October 28, 2026
Introduces HausaSafeLM, a safety preference dataset comprising 10,000 prompts including 3,000 human-corrected gold examples. Testing on Qwen2.5-7B-Instruct shows that base model refusal rates in Hausa often stem from poor linguistic capability rather than actual safety alignment.
HOW THIS AFFECTS YOU
●
builderYou cannot rely on refusal metrics alone to validate safety in low-resource languages.
●
researcherThe findings highlight the need for culturally specific safety datasets over simple machine translation.