aisingapore/Gemma-SEA-Guard-12B-2602
Gemma-SEA-Guard-12B-2602 is a 12 billion parameter decoder-only Large Language Model developed by AI Singapore, fine-tuned from Gemma 3 12B IT. This model is specifically designed for safety classification within Southeast Asian (SEA) contexts, supporting languages like Burmese, English, Indonesian, Malay, Tagalog, Tamil, Thai, and Vietnamese. It functions as a binary classifier, determining if human requests or AI responses are "safe" or "unsafe" based on culturally grounded safety criteria, with a context length of 128k tokens.
Loading preview...
Overview
Gemma-SEA-Guard-12B-2602 is part of AI Singapore's SEA-Guard collection, focusing on culturally grounded safety classification for the Southeast Asian (SEA) region. This 12 billion parameter model is a fine-tuned version of Google's Gemma 3 12B IT, trained on 1 million instruction-following pairs. It supports a wide array of SEA languages including Burmese, English, Indonesian, Malay, Tagalog, Tamil, Thai, and Vietnamese, utilizing Gemma 3's default tokenizer.
Key Capabilities
- Binary Safety Classification: Optimized to classify inputs as "safe" or "unsafe" for both human requests and AI assistant responses.
- Multilingual Support: Designed with cultural safety for SEA contexts across multiple languages.
- High Context Length: Features a substantial 128k token context window.
- API Access: Available via an API endpoint for direct integration into applications.
Intended Uses
This model is primarily intended for direct use as a safety classifier without further fine-tuning, leveraging its pre-trained cultural safety knowledge for SEA contexts. It can also serve as a base for further training on specific downstream tasks, such as vision-text safety datasets. Evaluation is conducted using SEA-SafeguardBench, with AUPRC as the primary metric for safety classification performance. Developers should be aware of inherent generative AI limitations, such as hallucination, and are advised to implement human oversight.