Llama-SEA-Guard-8B-2602: Culturally Grounded Safety for Southeast Asia
Llama-SEA-Guard-8B-2602, developed by AI Singapore, is an 8 billion parameter safety-focused Large Language Model (LLM) built upon the SEA-LION family. It is a fine-tuned version of aisingapore/Llama-SEA-LION-v3-8B-IT, trained on 1 million instruction-following pairs with a 128k token context length. This model is specifically optimized for the Southeast Asia (SEA) region, incorporating cultural safety nuances.
Key Capabilities
- Binary Safety Classification: Optimized to return a direct "safe" or "unsafe" classification for both human user requests and AI assistant responses.
- Multilingual Support: Handles Burmese, English, Indonesian, Malay, Tagalog, Tamil, Thai, and Vietnamese.
- Culturally Grounded: Trained with cultural safety considerations specific to SEA contexts, enabling direct use without additional fine-tuning or in-context learning for safety classification.
- High Context Length: Features a 128k token context window, allowing for comprehensive analysis of interactions.
Good for
- Content Moderation: Directly classifying the safety of user inputs or AI outputs in applications targeting Southeast Asian audiences.
- LLM Safety Layers: Integrating as a front-end or back-end safety filter for LLM interactions.
- Regional AI Development: Developers building AI applications for the SEA region who require culturally sensitive safety assessments.
For more details on the training data and evaluation, refer to the SEA-Guard paper and the SEA-SafeguardBench evaluation metric.