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

As Large Language Models (LLMs) become globally integrated, safety mechanisms often remain anchored in English-centric frameworks, failing to account for the unique socio-cultural sensitivities of the Indic region. This paper addresses the critical lack of localized safety data and moderation tools for Indic languages, aiming to build a robust, multilingual guardrail capable of identifying regional harms, socio-political sensitivities, and adversarial jailbreaks.

The researchers developed IndicGuard, a specialized safety framework built upon the Gemma-3-4B-IT model. The core of this work is a new, high-volume safety dataset covering ten major Indic languages (including Hindi, Bengali, Tamil, and Urdu). The dataset is categorized into three domains: generic safety, culture-adaptive content (addressing religion, caste, and social identity), and adversarial jailbreaking. The team utilized a hybrid construction methodology involving selective extraction from existing benchmarks, systematic multilingual translation, and taxonomy-driven annotation. They employed Parameter-Efficient Fine-Tuning (PEFT) via LoRA to adapt the base model for real-time content moderation.

IndicGuard consistently outperforms the state-of-the-art CultureGuard baseline across all evaluated languages. The study demonstrates that incorporating culture-adaptive and jailbreaking-specific data provides a measurable, incremental boost to safety classification performance. Notably, the model achieves a 0.00% over-refusal rate on safe-but-sensitive inputs, confirming that the guardrail enhances safety without suppressing legitimate user utility. Furthermore, the framework exhibits strong zero-shot cross-lingual transfer capabilities, successfully generalizing to low-resource languages like Dogri, Konkani, and Sanskrit that were not included in the training set.