CoMind is a community-augmented LLM agent for automated machine learning engineering that leverages collective knowledge from Kaggle-style competitions. CoMind achieves state-of-the-art performance on MLE-Live and outperforms 79.2% of human competitors on average across four ongoing Kaggle competitions.

Inspired by the workflow of human practitioners, CoMind operates in a loop that mirrors how experts read community posts, form new ideas, experiment, and share results. CoMind maintains two central repositories: an idea pool containing abstracted insights derived from community content and prior iterations, and a report pool with finalized solution reports, including code, evaluations, and analyses.

First, clone the repository and navigate to the project directory:

git clone https://github.com/comind-ml/CoMind.git
cd CoMind

Then, build the dependencies and install comind-community and comind-agent using pip. The -e flag installs the project in editable mode.

pip install -e .

To isolate the execution environment, comind-agent runs and analyzes the LLM-generated Python code within a 🐳 Docker container. If you plan to use GPUs, you'll need to install the NVIDIA Container Toolkit.

Build the Docker container with the following commands:

cd docker
docker build -t comind-agent .

Before running CoMind, you need to download the competition datasets and public resources from Kaggle. The ./competition/download.py script provides a command-line interface for this purpose. Please ensure you have placed your Kaggle API credentials (kaggle.json) in the ~/.kaggle/ directory.

For example, to download all resources for a competition and store them in the ./competitions directory, run:

python ./competition/download.py -c <competition-id> --download-dir ./competitions -d -k -t

Parameters:

• -c, --competition: The competition ID. You can find this in the URL of the competition's homepage: https://www.kaggle.com/competitions/<competition-id>.
• -dir, --download-dir: The directory to store the downloaded data. Defaults to ./competitions.
• -e, --exclude-after-deadline: Whether to exclude public discussions/kernels created after the competition's deadline. Defaults to True.
• -d, --dataset: Flag to download the dataset.
• -k, --kernels: Flag to fetch public kernels.
• -t, --discussions: Flag to fetch public discussions.
• -f, --force: Flag to re-download and overwrite existing data.

After preparation, the target directory will be structured as follows:

./competitions/<competition-id>
    |- input
        |- train.csv
        |- test.csv
        |- ...
    |- kernels
        |- info.json
        |- 283721122.txt 
        |- ...
    |- discussions
        |- info.json
        |- 2311122.txt
        |- ...
    |- description.md

First, launch the community server. This server manages the exchange of ideas and results.

comind-community llm.api_key=<your-api-key> community.competition_dir=./competitions/<competition-id>

Please wait for the community server to initialize before launching the agents.

Multiple agents can run in parallel to generate solutions and contribute to the community. To launch a single agent, run:

comind-agent llm.api_key=<your-api-key> input_dir=./competitions/<competition-id>/input

For more configuration options, see the Community Configs and Agent Configs.

CoMind is distributed under the MIT License. See LICENSE for more details.

If you use CoMind in your research, please cite our paper:

@article{Li2025TowardsCommunityDrivenAgents,
  title        = {Towards Community-Driven Agents for Machine Learning Engineering},
  author       = {Sijie Li and Weiwei Sun and Shanda Li and Ameet Talwalkar and Yiming Yang},
  journal      = {arXiv preprint arXiv:2506.20640},
  year         = {2025},
  month        = jun,
  note         = {arXiv:2506.20640 [cs.AI]},
  url          = {https://arxiv.org/abs/2506.20640}
}