Large Language Model Alignment for African Language Safety: A Case Study of Hausa

The researchers release HausaSafeLM, one of the first Hausa safety preference datasets, and test safety alignment methods on Qwen2.5-7B-Instruct for performance in a low-resource language. The dataset is built from 10,000 prompts from PKU-SafeRLHF and contains 3,000 manually corrected (gold) and 7,000 machine-translated (silver) samples, covering six harm categories; the experiments compare two schemes, SFT and SFT plus DPO, on 500 harmful and 200 benign Hausa prompts. Results show that the untuned base model appears very safe by refusal rate, but this actually stems from limited Hausa generation ability, and refusal quality is generally low; SFT improves both refusal fluency and precision and benefits from the larger silver data, while DPO makes refusals more precise but at the cost of a large sacrifice in language quality, and this cost basically does not change with data scale. The authors argue that existing alignment techniques are better at teaching models when to refuse than how to refuse fluently in underrepresented languages, and that refusal-rate metrics alone cannot reliably measure safety in low-resource settings.