Benchmark LLMs by fighting in Street Fighter 3! The new way to evaluate the quality of an LLM

Make LLM fight each other in real time in Street Fighter III.

Which LLM will be the best fighter ?

They need to be:

• Fast: It is a real time game, fast decisions are key
• Smart: A good fighter thinks 50 moves ahead
• Out of the box thinking: Outsmart your opponent with unexpected moves
• Adaptable: Learn from your mistakes and adapt your strategy
• Resilient: Keep your RPS high for an entire game

Street Fighter III assesses the ability of LLMs to understand their environment and take actions based on a specific context. As opposed to RL models, which blindly take actions based on the reward function, LLMs are fully aware of the context and act accordingly.

Our experimentations (342 fights so far) led to the following leaderboard. Each LLM has an ELO score based on its results

Each player is controlled by an LLM. We send to the LLM a text description of the screen. The LLM decide on the next moves its character will make. The next moves depends on its previous moves, the moves of its opponents, its power and health bars.

• Agent based

• Multithreading

• Real time

  

• Follow instructions in https://docs.diambra.ai/#installation
• Download the ROM and put it in ~/.diambra/roms
• (Optional) Create and activate a new python venv
• Install dependencies with make install or pip install -r requirements.txt
• Create a .env file and fill it with the content like in the .env.example file
• Run with make run

To disable the LLM calls, set DISABLE_LLM to True in the .env file. It will choose the actions randomly.

Change the logging level in the script.py file.

You can run the arena with local models using Ollama.

1. Make sure you have ollama installed, running, and with a model downloaded (run ollama serve mistral in the terminal for example)

2. Run make local to start the fight.

By default, it runs mistral against mistral. To use other models, you need to change the parameter model in ollama.py.

from eval.game import Game, Player1, Player2

def main():
    game = Game(
        render=True,
        save_game=True,
        player_1=Player1(
            nickname="Baby",
            model="ollama:mistral", # change this
        ),
        player_2=Player2(
            nickname="Daddy",
            model="ollama:mistral", # change this
        ),
    )
    game.run()
    return 0

The convention we use is model_provider:model_name. If you want to use another local model than Mistral, you can do ollama:some_other_model

The LLM is called in Robot.call_llm() method of the agent/robot.py file.

    def call_llm(
        self,
        temperature: float = 0.7,
        max_tokens: int = 50,
        top_p: float = 1.0,
    ) -> str:
        """
        Make an API call to the language model.

        Edit this method to change the behavior of the robot!
        """
        # self.model is a slug like mistral:mistral-small-latest or ollama:mistral
        provider_name, model_name = get_provider_and_model(self.model)
        client = get_sync_client(provider_name) # OpenAI client

        # Generate the prompts
        move_list = "- " + "\n - ".join([move for move in META_INSTRUCTIONS])
        system_prompt = f"""You are the best and most aggressive Street Fighter III 3rd strike player in the world.
Your character is {self.character}. Your goal is to beat the other opponent. You respond with a bullet point list of moves.
{self.context_prompt()}
The moves you can use are:
{move_list}
----
Reply with a bullet point list of moves. The format should be: `- <name of the move>` separated by a new line.
Example if the opponent is close:
- Move closer
- Medium Punch

Example if the opponent is far:
- Fireball
- Move closer"""

        # Call the LLM
        completion = client.chat.completions.create(
            model=model_name,
            messages=[
                {"role": "system", "content": system_prompt},
                {"role": "user", "content": "Your next moves are:"},
            ],
            temperature=temperature,
            max_tokens=max_tokens,
            top_p=top_p,
        )

        # Return the string to be parsed with regex
        llm_response = completion.choices[0].message.content.strip()
        return llm_response

To use another model or other prompts, make a call to another client in this function, change the system prompt, or make any fancy stuff.

Create a new class herited from Robot that has the changes you want to make and open a PR.

We'll do our best to add it to the ranking!

Made with ❤️ by the OpenGenerativeAI team from phospho (@oulianov @Pierre-LouisBJT @Platinn) and Quivr (@StanGirard) during Mistral Hackathon 2024 in San Francisco