LLAMA2 (and Other Models) Engaging in Self-Dialogue: Asking and Answering Its Own Questions


I am using Langchain + different LLM models (“Llama-2-7b-chat-hf”, “Mistral-7B-Instruct-v0.2”, …)…

My purpose is to do prompt engineering and evaluate a model’s ability to answer various questions.

The challenge is that after posing almost any question, the LLM starts a self-conversation, asking itself questions as a “Human” and then answering them as an “AI.”

What am I missing?
How can I ensure the model only responds to the original question without engaging in a self-conversation?

Using conversation via LangChain:

        conversation = ConversationChain(
        print("*************************** chat_conversation ***************************")
        while True:
            user_input = input("> ")            
            ai_response = conversation.predict(input=user_input)
            print("\nAssistant:\n", ai_response)