How can enterprises ensure their AI behaves safely – while still delivering high performance? How can organizations influence model behavior without costly retraining?
Jamba 1.5a is our answer. Using techniques like direct preference optimization (DPO) and rejection sampling on synthetic data, we show that it’s possible to significantly improve a model’s helpfulness, harmlessness, and honesty without compromising accuracy or speed. This opens the door to customizable alignment pipelines—so enterprises can shape model behavior to better reflect their own policies and standards.