This repository serves as a template for creating an agent that interacts with the GPT-Engineer benchmark suite. Follow the instructions below to set up your project.
Edit the benchmarking_agent.py file to implement your custom agent. This file contains a minimal example agent that you should replace with your own logic. If you want to benchmark a coding agent residing in another repo, we recommend installing it as a package and interfacing in the benchmark_agent.py file.
Execute the following command to install your project and run the agent against the benchmark suite:
./run.sh
run.sh also serves as a reference for how the template project is installed and run. By default, the results and the configuration will be stored in the file results.yaml. Remember that any api keys required by the implemented agents must be defined. The agent in the template requires OPENAI_API_KEY
At the end of running the benchmarks with run.sh, you will be asked whether you want to make a scoreboard entry. This means that the results from results.yaml will be appended to the file scoreboard.yaml.
If you choose this, the scoreboard.yaml file will be updated. You can upload it as a pull request to the template repo. To ensure reproducibility, you will be requested to submit additional information. Only reproducible entries will be merged. The best way to ensure reproducibility is to fork this repo and implement your agent in the benchmark_agent.py file. If done this way, it is enough to provide a url to the github or gitlab site where the fork is located.
If some other setup is used, it is up to you to provide an easy and reliable way for us to reproduce the results. At the end of running run.sh, there will be a dialogue where you will be provided with a free text option for how to reproduce your results. Note that, any custom LLM configuration should also be provided. The maintainers of this repo will reject any scoreboard entry PR that is deemed too cumbersome to reproduce.