LLM4PCG

LLM4PCG is a python package containing required and utility functions of ChatGPT4PCG competition, but modified to support local LLMs that compatible with OpenAI API interface.

Installation

Use the package manager pip to install LLM4PCG.

pip install llm4pcg

Dependency

This file uses the following Python libraries:

• openai

Functions

run_evaluation(team_name: str, fn: Type[TrialLoop], num_trials=10, characters: list[str] = None, model_name=None, local_model_base_url=None)

This function runs a trial for each character in the alphabet for a given team. It creates directories for logging and output, and generates a log file with a timestamp and timezone in the filename. It then runs trials for each character, skipping any that already exist.

To use local LLM, specify model_name and local_model_base_url parameters. If model_name is not specified, it will use the default model name gpt-3.5-turbo. If local_model_base_url is not specified, it will use the default base url of OpenAI API.

run_trial(ctx: TrialContext, fn: Type[TrialLoop])

This function runs a single trial. It writes the result of the trial to the log file and the final response to a text file in the output directory.

chat_with_llm(ctx: TrialContext, messages: []) -> list[str]

This function chats with the LLM. It sends a list of messages to the LLM and writes the response and token counts to the log file. It also checks for time and token limits, raising errors if these are exceeded.

Usage

To use this file, import it and call the run_evaluation function with the team name and trial loop function as arguments. You can also specify the number of trials to run and the characters to run trials for.

from llm4pcg.competition import run_evaluation, TrialLoop, TrialContext, chat_with_llm


class ZeroShotPrompting(TrialLoop):
    @staticmethod
    def run(ctx: TrialContext, target_character: str) -> str:
        message_history = [{
            "role": "user",
            "content": "Return this is a test message."
        }]

        response = chat_with_llm(ctx, message_history)
        return response[0]


run_evaluation("y_wing", ZeroShotPrompting)