MSafeTemp: Multilingual Safety Alignment via Template-Based Augmentation
October 28, 2026
MSafeTemp provides a validated corpus spanning 60 languages to address the English-centric bias in LLM safety alignment. The pipeline uses template-based augmentation to generate refusal and benign corpora with culturally authentic entities, mitigating the risk of English-only evaluations misjudging multilingual safety.
HOW THIS AFFECTS YOU
●
researcherYou can use this 60-language corpus to evaluate and align models for non-English safety.
●
policyThis helps mitigate cultural bias in safety guardrails for global deployments.