A program trying to solve tasks from the SWE-bench Lite dataset with simple actions
Achieves 23% pass on SWE Bench Lite through LLM eval, so take that result with a big grain of salt!
A lot more details in our blog post.
- Python 3.9+
git- An Opper API key
python main.py
Delvin is simple agent that can view, edit, and search files. Every action goes through a reflection stage and results from the environment (bad diffs, linting errors) are fed back to the model.
We use the Opper SDK to easily build structured functions. No dirty parsing needed :)
No. This repo relies extensively on our SDK and backend to interface with models and get tracing. It's just a hack project that turned out to work surprisingly well!
No. If you want to adapt this to your own issues then you will have to edit the main file to load issues from github. PRs welcome:)