aisingapore/Qwen-SEA-Guard-8B-2602
aisingapore/Qwen-SEA-Guard-8B-2602 is an 8 billion parameter decoder-only Large Language Model developed by AI Products Pillar, AI Singapore. Built upon the Qwen-SEA-LION-v4-8B-VL family, it is fine-tuned for safety classification in Southeast Asian contexts, supporting languages like Burmese, English, Indonesian, Malay, Tagalog, Tamil, Thai, and Vietnamese. This model is optimized to return a binary classification of "safe" or "unsafe" for user requests and AI responses, with a context length of 128k tokens.
Loading preview...
Overview
Qwen-SEA-Guard-8B-2602 is an 8 billion parameter decoder-only Large Language Model developed by AI Products Pillar, AI Singapore. It is a fine-tuned version of the aisingapore/Qwen-SEA-LION-v4-8B-VL model, specifically designed for safety classification within Southeast Asian (SEA) contexts. The model supports multiple SEA languages, including Burmese, English, Indonesian, Malay, Tagalog, Tamil, Thai, and Vietnamese, and utilizes the default Qwen3-VL tokenizer.
Key Capabilities
- Safety Classification: Optimized to classify text inputs as either "safe" or "unsafe" for both human user requests and AI assistant responses.
- Multilingual Support: Culturally grounded for the SEA region, handling safety assessments across several Southeast Asian languages.
- Vision-Text Safety: Capable of vision-text safety classification, as detailed in its associated research paper 2602.01618.
- High Context Length: Features a substantial context length of 128k tokens.
Intended Use Cases
- Direct Safety Filtering: Can be used directly to classify the safety of user prompts or AI generated responses without further fine-tuning.
- Downstream Safety Tasks: Suitable for continued training on specific target tasks, such as vision-text safety datasets.
- Integration with LLM Systems: Designed to be integrated into larger LLM ecosystems for content moderation and safety assurance, with support for fast inference via vLLM.
Limitations
Users should be aware of common generative AI limitations, including potential for hallucination or irrelevant text generation. Human oversight and secondary verification of outputs are advised, as the model's classifications should not be treated as absolute determinations.