Chat Vector: A Simple Approach to Equip LLMs With New Language Chat Capabilities
Paper • 2310.04799 • Published
Created from AIDC-AI/Marco-Mini-Global-Base using task arithmetic / chat vector.
Basically tensorwise X = G + I - B where G is Marco-Mini-Global-Base, I is Marco-Mini-Instruct and B is Marco-Mini-Base
Good enough to verify which languages the model knows, but no guarantees otherwise.