Build interactive code environments for interactive code agents.

Please refer to the change log for information on the latest updates to the InterCode environment.

InterCode is a lightweight, flexible, and easy-to-use framework for designing interactive code environments to evaluate language agents that can code.

For an overview of InterCode, building interactive code tasks with InterCode, and evaluating agents on InterCode environments, please check out our website, wiki, and the original paper:

InterCode: Standardizing and Benchmarking Interactive Coding with Execution Feedback
John Yang, Akshara Prabhakar, Karthik Narasimhan, Shunyu Yao

You can install InterCode as a PyPI package or by building from source.

Note InterCode requires the following installations to run:

• python >= 3.8
• docker: Learn more here to install. Before running the below code, make sure the Docker daemon/application is running locally.

1. Install the (pypi package):

pip install intercode-bench

2. Copy + Paste the following code for interacting with the InterCode Bash environment into a python file (i.e. run_bash.py)

from intercode.assets import bash_build_docker, bash_image_name, bash_test_data
from intercode.envs import BashEnv

if __name__ == '__main__':
    bash_build_docker()
    env = BashEnv(bash_image_name, data_path=bash_test_data, traj_dir="logs/", verbose=True) # Set verbose=False to silence Docker output

    try:
        for idx in range(24): # 24 data points in the test set
            env.reset(idx) # pass the index to prevent random data selection
            obs, done = env.observation, False # obs here is the natural language prompt
            while not done:
                action = input('> ')
                obs, reward, done, info = env.step(action)
                # After passing 'submit' to action, reward contains the score for that iteration
                # Note: Success Rate = (number of scores == 1.0 / total number of scores)
    except KeyboardInterrupt:
        print("Keyboard interrupt detected")
    finally:
        env.close()

3. Run the file (i.e. python run_bash.py)

If InterCode was installed successfully, the InterCode Bash environment should be started successfully and a CLI interpreter should appear, allowing you to enter bash commands to interact with the task setting. You can ctrl + c at any to time to exit the environment. Similar starter code for the InterCode SQL environment is available on the PyPI page.

1. Clone this repository, create a virtual environment, and install necessary dependencies

git clone https://github.com/princeton-nlp/intercode.git
cd intercode
conda env create -f environment.yml
conda activate intercode

2. Run setup.sh to create the docker images for the InterCode Bash, CTF, Python, and SQL environments
3. Run python run_demo.py sql

If InterCode was installed successfully, the InterCode SQL environment should be started successfully and a CLI interpreter should appear, allowing you to enter SQL commands to interact with the task environment. You can ctrl + c at any to time to exit the environment. Check run_demo.py for the latest full list of available environments.

If you'd like to run the scripts in the experiments folder, make sure you have at least one of the following keys declared

1. As an environment variable, or
2. Specified in a keys.cfg file formatted as follows + located in the root of this repository:

OPENAI_API_KEY: 'key here'
PALM_API_KEY: 'key here'

If you'd like to...

• Get a more in depth, but still brief overview of InterCode, see here
• Access an InterCode environment, see here
• Build an interactive code task with InterCode, see here
• Run language and code agents on InterCode based environments, see here

Not seeing what you want? Please feel free to check the wiki and paper for more details, or raise an issue if you still can't find it.

We would love to hear from the broader NLP and Machine Learning community, and we welcome any contributions, pull requests, or issues! To do so, please either file a new pull request or issue and fill in the corresponding templates accordingly. We'll be sure to follow up shortly!

Contact person: John Yang

If you find this repository helpful, feel free to cite our publication.

@inproceedings{yang2023intercode,
    title={InterCode: Standardizing and Benchmarking Interactive Coding with Execution Feedback}, 
    author={John Yang and Akshara Prabhakar and Karthik Narasimhan and Shunyu Yao},
    year={2023},
    eprint={2306.14898},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

MIT. Check LICENSE.md.