aisingapore/Qwen-SEA-Guard-4B-2602
Qwen-SEA-Guard-4B-2602 is a 4 billion parameter safety-focused Large Language Model (LLM) developed by AI Singapore, built upon the Qwen-SEA-LION-v4-4B-VL architecture. This model is specifically fine-tuned for the Southeast Asia (SEA) region, excelling at binary classification of text as "safe" or "unsafe" for both user prompts and AI assistant responses. It supports 8 languages including English, Burmese, Indonesian, Malay, Tagalog, Tamil, Thai, and Vietnamese, and features a 128k token context length.
Loading preview...
Overview
Qwen-SEA-Guard-4B-2602 is a 4 billion parameter safety-focused Large Language Model (LLM) developed by AI Singapore, building on the SEA-LION family of models. It is specifically designed and fine-tuned for the Southeast Asia (SEA) region, supporting 8 languages: Burmese, English, Indonesian, Malay, Tagalog, Tamil, Thai, and Vietnamese. The model is a fine-tuned version of aisingapore/Qwen-SEA-LION-v4-4B-VL and utilizes the default Qwen3-VL tokenizer.
Key Capabilities
- Safety Classification: Optimized to return a binary classification of "safe" or "unsafe" for both human user prompts and AI assistant responses.
- Culturally Grounded: Trained with cultural safety considerations specific to SEA contexts, making it suitable for direct use without further fine-tuning for safety classification.
- Multilingual Support: Processes and classifies content across a diverse set of Southeast Asian languages.
- Vision-Text Safety (Downstream): While primarily text-based, the underlying architecture supports vision-text safety tasks, allowing for further training on relevant datasets.
Intended Uses
- Direct Safety Moderation: Ideal for directly classifying text inputs (user prompts or AI responses) as safe or unsafe in applications targeting the SEA region.
- Foundation for Safety Systems: Can be used as a base model for further fine-tuning on specific safety-related tasks or integrated into larger safety ecosystems.
Limitations
- Like all generative AI, it is subject to hallucination and generation of ungrounded text. Human oversight and secondary verification of outputs are advised.