LLM-SMAC

Demos of LLM-SMAC tasks

Quick Start Guide

Get StarCraft II

LLM-SMAC depends on the full StarCraft II game and works with version of burnysc2/python-sc2 package.

Linux

Follow Blizzard's documentation to get the linux version. By default, LLM-SMAC expects the game to live in ~/StarCraftII/. You can override this path by setting the SC2PATH environment variable or creating your own run_config.

Windows/MacOS

Install of the game as normal from Battle.net. Even the Starter Edition will work. If you used the default install location LLM-SMAC should find the latest binary. If you changed the install location, you might need to set the SC2PATH environment variable with the correct location.

Note that the version of StarCraft II on Linux platform supports upto 4.10, however, the latest version on Windows/Macos is upto 5.0.13. Therefore, the results across the operating system might not be consistent.

Get LLM-SMAC

Download the LLM-SMAC code from this github page.

Use pip install to initialize the environment:

$ conda create --name YOUR_ENV_NAME python==3.10
$ conda activate YOUR_ENV_NAME
$ pip install -r requirements.txt

Replace sc2 Folder and Maps

The sc2 folder is modified according to the SMAC task. We modified the log level as 'CRITICAL' and modify the bot_ai_internal.py of sc2 file because the start_location and the enemy_start_locations cannot be acquired without the 'townhall'.
We modify the maps of original SMAC tasks and add the victory, tie, and defeat judgments. We also set a timeout mechanism to solve the problem of unfinished tasks.

Replace your API key

You may fill your api key to the file 'configs\llm_api_config.py'

class LLMAPIConfig:
    # LLM 配置
    MODELS = {
        "deepseek-chat": LLMModelConfig(
            api_key="Your API key here.",
            base_url="Base URL here."
        ),
        "deepseek-coder": LLMModelConfig(
            api_key="Your API key here.",
            base_url="Base URL here."
        ),
        "gpt-4": LLMModelConfig(
            api_key="Your API key here.",
            base_url="Base URL here."
        ),
        "Qwen2.5-72B-Instruct": LLMModelConfig(
            api_key="Your API key here.",
            base_url="Base URL here."
        ),
        "claude-3-5-sonnet-20240620": LLMModelConfig(
            api_key="Your API key here.",
            base_url="Base URL here."
        ),
    }

Customize the planner, coder, and critic model

You may customize the models from DeepSeek to other models by applying the corresponding API key in 'configs\llm_api_config.py'

TASK_MODELS = {
        "planner": "gpt-4",
        "coder": "gpt-4",
        "summarizer": "claude-3-5-sonnet-20240620"
    }

How to run

# Change the information in the config.py, then
$ python main.py

After applying the command, there should be a log file in the root directory which can be modified in the file 'LLM\call_llm_api\call_llm.py'

The generated scripts are also in the root directory, which named as 'res-Victory_Times-Test_Times-Planning_Round-Coding_Round'. When new experiments are conducted, the previously generated Python scripts will be REMOVED, so make sure you have store the results.

devindeng94 / LLM-SMAC

readme