JT-SAFE-V2: Safety-by-Design Foundation Model with World-Context Data
Researchers have introduced JT-Safe-V2, a new foundation model designed to improve the safety and trustworthiness of AI systems. This model integrates general intelligence with safety-by-design principles through enriched data, specialized training procedures, and post-training safety enhancements. Additionally, a framework called Safe-MoMA has been developed to manage multiple models and agents for efficient and traceable inference, reducing costs by over 30% while maintaining performance. The team is releasing the JT-Safe-V2-35B model checkpoint to encourage further research in this area. AI
IMPACT This release offers a new approach to building safer AI models and a framework for more efficient inference, potentially impacting enterprise AI deployments.