How I Aligned Your Model: Scaling Up Multilingual Safety Training with Templates
Attacks blocked in English succeeded in other languages, while culturally localized prompts exposed failures that translation alone missed.
The researchers released MSafeTemp, a validated corpus spanning 60 languages for safety training and evaluation. Their pipeline extracts templates from harmful prompts, translates them, and fills the slots with culturally authentic harmful and benign entities. They report that training with this augmented data improves safety beyond English, though the abstract gives no numerical results.
ArXiv · AI/CL/LG's note
The researchers released MSafeTemp, a validated corpus spanning 60 languages for safety training and evaluation. Their pipeline extracts templates from harmful prompts, translates them, and fills the slots with culturally authentic harmful and benign entities. They report that training with this augmented data improves safety beyond English, though the abstract gives no numerical results.
ArXiv · AI/CL/LG's note
score 4