It very specifically mentioned something about LLaMA not being trained to answer that in the response.

Again, its extremely minimal, but I think it picked something up from the Llama info facebook inserted.