Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SEA-Guard: Culturally Grounded Multilingual Safeguard for Southeast Asia

Panuthep Tasawong, Jian Gang Ngui, Alham Fikri Aji, Trevor Cohn, Peerat Limkonchotiwat

TL;DR

SEA-Guard tackles the challenge of culturally grounded AI safeguards for Southeast Asia by introducing a multilingual data-generation framework that yields 870k SEA-specific samples per language across 53 cultural categories. Built on 8 SEA languages and trained in 4B, 8B, and 12B sizes, SEA-Guard demonstrates state-of-the-art performance on SEA cultural safety benchmarks while maintaining competitive generic safety and emergent vision-language safety capabilities in zero-shot settings. The framework employs an agentic annotation pipeline, including Monte Carlo Reasoning Ensemble for robust labeling, and rigorous data quality, deduplication, and human verification to ensure cultural relevance and safety. The work highlights the importance of regional grounding over translation-based approaches and delineates trade-offs between cultural specificity and broad safety coverage, with practical impact for deploying safer conversational AI in SEA contexts.

Abstract

Culturally aware safeguards are crucial for AI alignment in real-world settings, where safety extends beyond common sense and encompasses diverse local values, norms, and region-specific regulations. However, building large-scale, culturally grounded datasets is challenging due to limited resources and a scarcity of native annotators. Consequently, many safeguard models rely on machine translation of English datasets, often missing regional and cultural nuances. We present a novel agentic data-generation framework to scalably create authentic, region-specific safety datasets for Southeast Asia (SEA). On this foundation, we introduce the SEA-Guard family, the first multilingual safeguard models grounded in SEA cultural contexts. Evaluated across multiple benchmarks and cultural variants, SEA-Guard consistently outperforms existing safeguards at detecting regionally sensitive or harmful content while maintaining strong general safety performance.

SEA-Guard: Culturally Grounded Multilingual Safeguard for Southeast Asia

TL;DR

SEA-Guard tackles the challenge of culturally grounded AI safeguards for Southeast Asia by introducing a multilingual data-generation framework that yields 870k SEA-specific samples per language across 53 cultural categories. Built on 8 SEA languages and trained in 4B, 8B, and 12B sizes, SEA-Guard demonstrates state-of-the-art performance on SEA cultural safety benchmarks while maintaining competitive generic safety and emergent vision-language safety capabilities in zero-shot settings. The framework employs an agentic annotation pipeline, including Monte Carlo Reasoning Ensemble for robust labeling, and rigorous data quality, deduplication, and human verification to ensure cultural relevance and safety. The work highlights the importance of regional grounding over translation-based approaches and delineates trade-offs between cultural specificity and broad safety coverage, with practical impact for deploying safer conversational AI in SEA contexts.

Abstract

Culturally aware safeguards are crucial for AI alignment in real-world settings, where safety extends beyond common sense and encompasses diverse local values, norms, and region-specific regulations. However, building large-scale, culturally grounded datasets is challenging due to limited resources and a scarcity of native annotators. Consequently, many safeguard models rely on machine translation of English datasets, often missing regional and cultural nuances. We present a novel agentic data-generation framework to scalably create authentic, region-specific safety datasets for Southeast Asia (SEA). On this foundation, we introduce the SEA-Guard family, the first multilingual safeguard models grounded in SEA cultural contexts. Evaluated across multiple benchmarks and cultural variants, SEA-Guard consistently outperforms existing safeguards at detecting regionally sensitive or harmful content while maintaining strong general safety performance.
Paper Structure (38 sections, 10 equations, 23 figures, 12 tables, 2 algorithms)

This paper contains 38 sections, 10 equations, 23 figures, 12 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. SEA-Guard
  3. Overview
  4. Input Formulation
  5. Prompts and Responses Generation
  6. Data Annotation and Quality Assurance
  7. Monte Carlo Reasoning Ensemble (MCRE)
  8. Prompt and Response Annotation
  9. Data Quality Assurance
  10. Data Deduplication
  11. Human Verification
  12. SEA-Guard Training
  13. Experimental Setup
  14. Experimental Results
  15. SEA Cultural Safety Results
  16. ...and 23 more sections

Figures (23)

  • Figure 1: Illustration of how a safeguard model places and protects LLMs.
  • Figure 2: Illustration of how we formulate SEA cultural training data. We split the data generation framework into four parts; the details are indicated in each section.
  • Figure 3: Alignment between model-predicted harmfulness scores and human-judged severity levels.
  • Figure 4: Robustness to adversarial attack.
  • Figure 5: Impact of dataset size and deduplication on model performance.
  • ...and 18 more figures