Skip to content

ChatArena (or Chat Arena) is a Multi-Agent Language Game Environments for LLMs. The goal is to develop communication and collaboration capabilities of AIs.

License

Notifications You must be signed in to change notification settings

Farama-Foundation/chatarena

Repository files navigation

Multi-Agent Language Game Environments for LLMs

License: Apache2 PyPI Python 3.9+ twitter Discord Open In Colab HuggingFace Space


ChatArena is a library that provides multi-agent language game environments and facilitates research about autonomous LLM agents and their social interactions. It provides the following features:

  • Abstraction: it provides a flexible framework to define multiple players, environments and the interactions between them, based on Markov Decision Process.
  • Language Game Environments: it provides a set of environments that can help understanding, benchmarking or training agent LLMs.
  • User-friendly Interfaces: it provides both Web UI and CLI to develop/prompt engineer your LLM agents to act in environments.

ChatArena Architecture

Getting Started

Try our online demo: demo Demo video

Installation

Requirements:

  • Python >= 3. 7
  • OpenAI API key (optional, for using GPT-3.5-turbo or GPT-4 as an LLM agent)

Install with pip:

pip install chatarena

or install from source:

pip install git+https://github.com/chatarena/chatarena

To use GPT-3 as an LLM agent, set your OpenAI API key:

export OPENAI_API_KEY="your_api_key_here"

Optional Dependencies

By default pip install chatarena will only install dependencies necessary for ChatArena's core functionalities. You can install optional dependencies with the following commands:

pip install chatarena[all_backends] # install dependencies for all supported backends: anthropic, cohere, huggingface, etc.
pip install chatarena[all_envs]     # install dependencies for all environments, such as pettingzoo
pip install chatarena[all]          # install all optional dependencies for full functionality

Launch the Demo Locally

The quickest way to see ChatArena in action is via the demo Web UI. To launch the demo on your local machine, you first pip install chatarena with extra gradio dependency, then git clone this repository to your local folder, and finally call the app.py in the root directory of the repository:

pip install chatarena[gradio]
git clone https://github.com/chatarena/chatarena.git
cd chatarena
gradio app.py

This will launch a demo server for ChatArena, and you can access it from your browser (port 8080).

Check out this video to learn how to use Web UI: Webui demo video

For Developers

For an introduction to the ChatArena framework, please refer to this document. For a walkthrough of building a new environment, check Open In Colab

Here we provide a compact guide on minimal setup to run the game and some general advice on customization.

Key Concepts

  1. Arena: Arena encapsulates an environment and a collection of players. It drives the main loop of the game and provides HCI utilities like webUI, CLI, configuration loading and data storage.
  2. Environment: The environment stores the game state and executes game logics to make transitions between game states. It also renders observations for players, the observations are natural languages.
    1. The game state is not directly visible to the players. Players can only see the observations.
  3. Language Backend: Language backends are the source of language intelligence. It takes text (or collection of text) as input and returns text in response.
  4. Player: The player is an agent that plays the game. In RL terminology, it’s a policy, a stateless function mapping from observations to actions.

Run the Game with Python API

Load Arena from a config file -- here we use examples/nlp-classroom-3players.json in this repository as an example:

arena = Arena.from_config("examples/nlp-classroom-3players.json")
arena.run(num_steps=10)

Run the game in an interactive CLI interface:

arena.launch_cli()

Check out this video to learn how to use CLI: cli demo video A more detailed guide about how to run the main interaction loop with finer-grained control can be found here

General Customization Advice

  1. Arena: Overriding Arena basically means one is going to write their own main loop. This can allow different interaction interfaces or drive games in a more automated manner, for example, running an online RL training loop
  2. Environment: A new environment corresponds to a new game, one can define the game dynamics here with hard-coded rules or a mixture of rules and language backend.
  3. Backend: If one needs to change the way of formatting observations (in terms of messages) into queries for the language model, the backend should be overridden.
  4. Player: By default, when a new observation is fed, players will query the language backend and return the response as actions. But one can also customize the way that players are interacting with the language backend.

Creating your Custom Environment

You can define your own environment by extending the Environment class. Here are the general steps:

  1. Define the class by inheriting from a base class and setting type_name, then add the class to ALL_ENVIRONMENTS
  2. Initialize the class by defining __init__ method (its arguments will define the corresponding config) and initializing class attributes
  3. Implement game mechanics in methods step
  4. Handle game states and rewards by implementing methods such as reset, get_observation, is_terminal, and get_rewards
  5. Develop role description prompts (and a global prompt if necessary) for players using CLI or Web UI and save them to a config file.

We provide a detailed tutorial to demonstrate how to define a custom environment, using the Chameleon environment as example.

If you want to port an existing library's environment to ChatArena, check out PettingzooChess environment as an example.

List of Environments

A multi-player language game environment that simulates a conversation.

  • NLP Classroom: a 3-player language game environment that simulates a classroom setting. The game is played in turns, and each turn a player can either ask a question or answer a question. The game ends when all players have asked and answered all questions.

Based on conversation, but with a moderator that controls the game dynamics.

  • Rock-paper-scissors: a 2-player language game environment that simulates a rock-paper-scissors game with moderator conversation. Both player will act in parallel, and the game ends when one player wins 2 rounds.
  • Tic-tac-toe: a 2-player language game environment that simulates a tic-tac-toe game with moderator conversation. The game is played in turns, and each turn a player can either ask for a move or make a move. The game ends when one player wins or the board is full.

A multi-player social deduction game. There are two roles in the game, chameleon and non-chameleon. The topic of the secret word will be first revealed to all the players. Then the secret word will be revealed to non-chameleons. The chameleon does not know the secret word. The objective in the game depends on the role of the player:

  • If you are not a chameleon, your goal is to reveal the chameleon without exposing the secret word.
  • If you are a chameleon, your aim is to blend in with other players, avoid being caught, and figure out the secret word. There are three stages in the game:
  1. The giving clues stage: each player will describe the clues about the secret word.
  2. The accusation stage: In this stage, each player will vote for another player who is most likely the chameleon. The chameleon should vote for other players.
  3. The guess stage: If the accusation is correct, the chameleon should guess the secret word given the clues revealed by other players.

A two-player chess game environment that uses the PettingZoo Chess environment.

A two-player tic-tac-toe game environment that uses the PettingZoo TicTacToe environment. Differing from the Moderator Conversation environment, this environment is driven by hard-coded rules rather than a LLM moderator.

Contributing

We welcome contributions to improve and extend ChatArena. Please follow these steps to contribute:

  1. Fork the repository.
  2. Create a new branch for your feature or bugfix.
  3. Commit your changes to the new branch.
  4. Create a pull request describing your changes.
  5. We will review your pull request and provide feedback or merge your changes.

Please ensure your code follows the existing style and structure.

Citation

If you find ChatArena useful for your research, please cite our repository (our arxiv paper is coming soon):

@software{ChatArena,
  author = {Yuxiang Wu, Zhengyao Jiang, Akbir Khan, Yao Fu, Laura Ruis, Edward Grefenstette, and Tim Rocktäschel},
  title = {ChatArena: Multi-Agent Language Game Environments for Large Language Models},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  version = {0.1},
  howpublished = {\url{https://github.com/chatarena/chatarena}},
}

Contact

If you have any questions or suggestions, feel free to open an issue or submit a pull request. You can also contact us on the Farama discord server- https://discord.gg/Vrtdmu9Y8Q

Happy chatting!

Sponsors

We would like to thank our sponsors for supporting this project: