Multilingual safety training improves language model security worldwide
How I Aligned Your Model: Scaling Up Multilingual Safety Training with Templates
Computation and LanguageArtificial IntelligenceMachine Learning
Summary
Safety checks for AI language models mostly focus on English, which leaves many other languages vulnerable to harmful content. The authors developed a way to create large, diverse safety training data in 60 languages by transforming harmful prompts into safe ones, translating them, and inserting culturally relevant words. They found that testing models only in English misses many safety problems in other languages, especially when local cultural issues are considered. Their new dataset and method help make language models safer for many languages around the world.
What this means in practice
- •For ai product developers: Improve the safety and refusal behavior of language models deployed in diverse linguistic regions using multilingual training data.
- •For machine translation engineers: Detect and address safety vulnerabilities in translated language model prompts to enhance model robustness across languages.
Authors
Max Neuwinger, Hanna Yukhymenko, Imanol Schlag
Abstract
Safety alignment for Large Language Models is heavily focused on English with limited multilingual representation, particularly damaging low-resource language performance. Related open training datasets are often limited in size or do not follow region-localized harmful topics to broaden the scope of alignment beyond cultures. We propose a scalable template-based augmentation pipeline to produce massive multilingual refusal and benign corpora, extracting safe templates from harmful prompts, translating them and then filling in the slots with culturally authentic entities, both harmful and benign. As a result, we release MSafeTemp, a validated corpus spanning 60 languages useful for multilingual safety alignment and evaluation. Using the dataset, we show that English-only evaluation misjudges multilingual safety: for several models, attacks that fail in English succeed once translated, while culturally localized variants expose further failures that translation alone misses. We also confirm that augmenting safety training data through templates allows for higher diversity and quality of alignment, resulting in improved safety performance beyond English. MSafeTemp and the respective pipeline are aimed at facilitating research and development of safer and more responsible models for less-represented cultures.