Introducing SEA-Guard (Safety Collection): Safety-First AI, Built for Southeast Asia
As AI systems become more deeply embedded in products, services, and public-facing applications across Southeast Asia, safety cannot be an afterthought.
What is SEA-Guard?
We are excited to announce, SEA-Guard, our new collection of safety-focused Large Language Models (LLMs) built upon the SEA-LION family. These models are specifically fine-tuned to detect, moderate, and handle content according to Southeast Asian cultural norms and safety standards.
Key Technical Specifications
- Training Data: Fine-tuned on 1M instruction-following pairs specifically curated for safety and regional context.
- Context Length: 128k tokens, allowing for the analysis of long documents and complex conversation histories.
- Regional Coverage: Native support for Burmese, English, Indonesian, Malay, Tagalog, Tamil, Thai, and Vietnamese.
- Developer: AI Products Pillar, AI Singapore (Funded by Singapore NRF).
The SEA-Guard Model Lineup
The collection features four specialized models, covering both text and vision modalities to ensure comprehensive safety coverage.
| Model Name | Hugging Face ID | Type | Size | Description |
|---|---|---|---|---|
| Qwen-SEA-Guard-4B | aisingapore/Qwen-SEA-Guard-4B-040226 | Image-to-Text | 4B | Lightweight visual safety guardrail for edge applications. |
| Qwen-SEA-Guard-8B | aisingapore/Qwen-SEA-Guard-8B-040226 | Image-to-Text | 8B | Balanced visual moderation model with stronger reasoning capabilities. |
| Llama-SEA-Guard-8B | aisingapore/Llama-SEA-Guard-8B-040226 | Text Generation | 8B | Text-only safety model optimized for chat moderation and policy enforcement. |
| Gemma-SEA-Guard-12B | aisingapore/Gemma-SEA-Guard-12B-040226 | Image-Text-to-Text | 12B | High-capacity multimodal safety model for complex content analysis. |
These additions ensure that whether you are optimizing for efficiency with our Compact models or building trust with SEA-Guard, the SEA-LION ecosystem has the right tool for your infrastructure and safety needs.
Model Results
In contrast to open-source SOTA baseline models, SEA-Guard delivers stronger and more reliable English and Southeast Asia safety performance than existing models, with Gemma-SEA-Guard-12B achieving the best results across both prompt and response classification while maintaining consistent performance between the two. Unlike baseline models that struggle with Southeast Asian languages and cultural nuance, SEA-Guard performs robustly across SEA languages. For detailed results, please refer to the SEA-Guard paper.
Try out SEA-Guard today! Download our SEA-Guard models from Hugging Face or test Gemma-SEA-GUARD-12B version directly via our APIs.
Acknowledgements: This work is funded by the Singapore National Research Foundation (NRF) and developed by the AI Products Pillar at AI Singapore. For inquiries, please contact sealion@aisingapore.org.