Skip to content

Generating a dataset of descriptions of swarm missions and respective simulator configuration to finetune an LLM.

License

Notifications You must be signed in to change notification settings

julianjandeleit/swarm_descriptions

Repository files navigation

Swarm Descriptions: Creating and Demonstrating a Dataset for Swarm Mission Generation from Natural Language

Installation

For installation, the Hatch Project Manager is recommended. Hatch can be installed with pipx install hatch, assuming pipx is installed. Pipx can be installed using most package managers or with python -m pip install --user pipx. Make sure to execute pipx ensurepath after installation.

From repository level start hatch virtual python environment with hatch shell and install python module with python -m pip install -e ..

Copy custom directory to /opt/argos/custom/. Execute xhost +local:docker to enable visualization.

Execution

Scrips

Run scripts like python scripts/test.py -h inside the hatch virtual environment.

  • generate_data.py samples descriptions and configurations for demonstration.
  • parse_eval_datasets.py converts the dataset generated by the inference notebook for evaluation.
  • eval_results.py contains the data for evaluation of the finetuned model.
  • run_config_params.py runs configuration params xml generated from dataset (e.g. generate_data.py.)
  • test.py prints several properties generated from a randomly sampled mission.

Notebooks:

Finetuning and inference was done on Kaggle.

  • figures.ipynb figures and metrics for report.
  • mistral-finetuning-swarm.ipynb finetuning the LLM on our dataset. Includes relevant functions on how to generate dataset from our python module.
  • mistral-inference-swarm.ipynb generating configuration params from descriptions by fined LLM for evaluation.

Related Work

Association

This work is part of my masters project at the University of Konstanz. Find the project report at project_report.pdf or the presentation at project_slides.pdf.

Troubleshooting

  • hatch shell cannot install wheel: python-version in pyproject.toml needs to be in the shape of ">=3.10" and not "==3.10.12". At least on Windows.

About

Generating a dataset of descriptions of swarm missions and respective simulator configuration to finetune an LLM.

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published