Assign a Role

How can we increase a model's performance on open-ended tasks?

Role prompting, or persona prompting, assigns a role to the model. Roles can be:

specific to the query: You are a talented writer. Write me a poem.
general/social: You are a helpful AI assistant. Write me a poem.

Implementation¶

import openai
import instructor
from pydantic import BaseModel

client = instructor.from_openai(openai.OpenAI())


class Response(BaseModel):
    poem: str


def role_prompting(query, role):
    return client.chat.completions.create(
        model="gpt-4o",
        response_model=Response,
        messages=[
            {
                "role": "system",
                "content": f"{role} {query}",
            },
        ],
    )


if __name__ == "__main__":
    query = "Write me a short poem about coffee."
    role = "You are a renowned poet."

    response = role_prompting(query, role)
    print(response.poem)
    """
    In the morning's gentle light,
    A brew of warmth, dark and bright.
    Awakening dreams, so sweet,
    In every sip, the day we greet.

    Through the steam, stories spin,
    A liquid muse, caffeine within.
    Moments pause, thoughts unfold,
    In coffee's embrace, we find our gold.
    """

More Role Prompting

To read about a systematic approach to choosing roles, check out RoleLLM.

For more examples of social roles, check out this evaluation of social roles in system prompts..

To read about using more than one role, check out Multi-Persona Self-Collaboration.

References¶

¹: RoleLLM: Benchmarking, Eliciting, and Enhancing Role-Playing Abilities of Large Lanuage Models
²: Is "A Helpful Assistant" the Best Role for Large Language Models? A Systematic Evaluation of Social Roles in System Prompts
³: Unleashing the Emergent Cognitive Synergy in Large Lanuage Models: A Task-Solving Agent through Multi-Persona Self-Collaboration