JT-Safe-V2: A Safety-by-Design Foundation Model
Researchers have unveiled JT-Safe-V2, a large language model that enhances the safety and reliability of foundation models through a design-focused safety approach. This model builds upon the earlier JT-Safe by optimizing both general intelligence and safety, utilizing enriched pre-training data that includes contextual world knowledge, high-certainty pre-training methods, and safety-enhancing post-training for enterprise-level capabilities. Additionally, a new framework called Safe-MoMA (Safe Mixture of Models and Agents) facilitates efficient and traceable inference by coordinating various models and agents. Evaluations indicate that JT-Safe-V2 sets new standards in performance for general intelligence and safety benchmarks, while Safe-MoMA effectively lowers inference costs.
Key facts
- JT-Safe-V2 is a large language model for safety and trustworthiness.
- It extends the previous JT-Safe model.
- It uses a safety-by-design paradigm.
- Innovations include enriched pre-training with world knowledge.
- Safe-MoMA framework enables orchestrated deployment of models and agents.
- JT-Safe-V2 achieves state-of-the-art on safety and intelligence benchmarks.
- Safe-MoMA reduces inference costs.
- The model targets enterprise-oriented agentic capabilities.
Entities
—