This fork is based on a1017b7 commit of aorwall/moatless-tools. Given that the original repository has recently undergone significant code refactoring and bump to v0.0.3, we likely won't bne able to upstream the changes.
To run moatless-tools agent, you need the vector index of the repos too, which we pre-computed for SWE-Gym Lite and SWE-Bench Lite. Download through this link.
We have made the following changes to enable dramatically faster iteration speed:
Instead of running a single agent task by task, we run multiple agents in parallel, by scripts/parallel_sample.sh
This enables us to sample the entire SWE-Gym Lite/SWE-Bench Lite within 30 mins for 32B model and 10 mins for 7B model with a single H100 GPU.
We also evaluate patches through OpenHands Cloud Service by scripts/eval_preds.sh
Index your environment in parallel by notebooks/ingest.py