I tried

llama2-70b-hf

and provided role allocating prompt from AutoGen, but it always response in conversational way. ie,

Here are following roles to complete provided task:

Role1: Some description

But gpt3.5, gpt4 and bard all reply back correctly as asked format.