GPT Competency Experiments

Contributors

• Muhammad Fawad Akbar Khan ([email protected])
  • Website
• Max Ramsdell ([email protected])

Running Instructions

In order to run any of the top level python scripts (not the jupyter notebook) on a linux-based system, run source venv/bin/active and your python environment should be set up to run the scripts.
Also note that there should be a .env file with the key OPEN_AI_KEY=your_key_here to be able to run the generate.py script.

Directories

• 3.5turbo is the directory with all of the ChatGPT generated code for this experiment.
• humanCode is the directory with all of the human-written code for this experiment
• Datasets directory is all of the testing data that we
• analysis contain various figures used when studying the programs that were produced.
• data contains some of the final and most important csv files used to compare the code produced.

Top Level Files

• generate.py generates the GPT code by parsing the prompts listed in the ideas.md file. This is how the 3.5turbo directory and its subdirectories were generated.
• ideas.md lists all of the prompts given to both the humans and chatgpt, as well as the sources of the ideas if they were retrieved online, and the sources of code in the humanCode directory if it was retrieved online.
• metrics.py used to generate either the humanCode directory, or the 3.5turbo directories for the complexity and other measures we used in our study.
• prediction_model.ipynb the code for the model used to predict if a piece of code was written by chatGPT or by a human. It also has code to generate the analysis figures, and run the model reliability tests.

Citation

If you use the code or data from this repository, please cite the following paper:

@article{khan2023assessing, title={Assessing the Promise and Pitfalls of ChatGPT for Automated Code Generation}, author={Khan, Muhammad Fawad Akbar and Ramsdell, Max and Falor, Erik and Karimi, Hamid}, journal={arXiv preprint arXiv:2311.02640}, year={2023} }