Build Environments
SWE-smith enables automatic conversion of code repositories into reinforcement learning environments.
We'll review the two steps of this process:
- SWE-agent + LM attempts to install a repository + run the testing suite.
- Construct an execution environment (Docker image).
For this section, we'll use the Instagram/MonkeyType repository as a running example,
specifically at commit 70c3acf.
Automatically Install Repos with SWE-agent
Coming soon!
Python installation scripts
Early on in SWE-smith's development, we focused exclusively on Python repositories and wrote Python-specific scripts for automatic repo instllation. More information here
Create an Execution Environment
Run the following command to create a Docker image for the repository.
python -m swesmith.build_repo.create_images -r MonkeyType
This command will create two artifacts:
1. A mirror of the original repository at the specified commit, created under swesmith. To change the organization, you can...
* Pass in an --org argument, or
* (If built from source) Change ORG_NAME_GH in swesmith/constants.py
2. A Docker image (swesmith.x86_64.<repo>.<commit>) which contains the installed codebase.
create_images arguments
By default, without -r, the command will build images for all SWE-smith repositories (300+ as of 12/2025).
-r: Select specific repositories to build using fuzzy matching (e.g., -r django matches any repo containing "django").
-f: Force rebuild images even if they already exist locally.
It's good practice to check that your Docker image works as expected.
docker run -it --rm swebench/swesmith.x86_64.instagram__monkeytype.70c3acf6
pytest) to ensure that the codebase is functioning as expected.
Get existing Docker images
All repositories represented in the SWE-smith dataset are available to download. Simply run:
python -m swesmith.build_repo.download_images
