Researchers have explored safety alignment for large language models (LLMs) in low-resource languages, focusing on Hausa as a case study. They developed HausaSafeLM, a dataset of 10,000 prompts, and tested alignment methods like supervised fine-tuning (SFT) and Direct Preference Optimization (DPO) on the Qwen2.5-7B-Instruct model. The study found that while base models might appear safe based on refusal rates, this often reflects limited language capabilities rather than true alignment. SFT improved refusal quality, especially with larger datasets, but DPO negatively impacted linguistic fluency. The findings suggest current alignment techniques struggle with fluent refusal in underrepresented languages, and refusal rates alone are an unreliable safety metric. AI
IMPACT Highlights the need for tailored safety alignment methods and evaluation metrics for underrepresented languages, potentially influencing future LLM development for global inclusivity.
RANK_REASON The cluster contains academic papers detailing research into LLM safety alignment and faithfulness in text generation.
Read on Hugging Face Daily Papers →
- Diffusion language models
- GPT-2
- GPT-Neo
- Hugging Face
- QLoRA
- Table-to-text generation
- ToTTo
- Hausa
- HausaSafeLM
- PKU-SafeRLHF
- Qwen2.5-7B-Instruct
- West Africa
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
