Whenever I try to use a SystemMesssage or SystemMessagePromptTemplate with a pipeline, it prints out the System Message as part of the response. I thought it would be like when using an api (it doesn’t show the system messages and instead is used as context for the LLM). I think I’m using the wrong classes for these but I can’t seem to find how to do it in the docs. Is this something that you are not able to do when it’s installed locally?

Here’s the code:

from langchain_community.llms.huggingface_pipeline import HuggingFacePipeline
from langchain_core.prompts.chat import (
    SystemMessagePromptTemplate,
    HumanMessagePromptTemplate,
    ChatPromptTemplate
) 

model = "openai-community/gpt2"
hf = HuggingFacePipeline.from_model_id(
    model_id=model,
    task="text-generation",
    pipeline_kwargs={"max_new_tokens": 20},
)

messages = [
    SystemMessagePromptTemplate.from_template(
        "You are a game show host that asks a general trivia question. You only ask one question and that is all. You ask it in a cheery and entertaining way as a game show host would do."
    ),
    HumanMessagePromptTemplate.from_template("What's your question host?"),
]

prompt = ChatPromptTemplate.from_messages(messages)

chain = prompt | hf
print(chain.invoke({}))

Here’s an example of the response:

System: You are a game show host that asks a general trivia question. You only ask one question and that is all. You ask it in a cheery and entertaining way as a game show host would do.
Human: What's your question host?
Humor is at the heart of your game show, so we're going to run a lot

What I’m expecting is just one random general trivia question to be printed without anything else.