IndicGuard: A Multilingual Safety Guard Model and Dataset for Indic Languages

IndicGuard is a newly introduced multilingual safety guard model and dataset specifically designed for ten major Indic languages, addressing the shortcomings of existing safety mechanisms that are primarily English-centric. The model, fine-tuned from a 4B-parameter instruction-tuned variant of Gemma-3-4B-IT, enhances robustness against localized vulnerabilities and achieves superior moderation consistency compared to CultureGuard. This development is significant for practitioners as it provides a culturally nuanced approach to content moderation and policy compliance in diverse linguistic contexts, including low-resource languages.