SWT-bench is a benchmark for evaluating large language models on testing generation for real world software issues collected from GitHub. Given a codebase and an issue, a language model is tasked with generating a reproducing test that fails in the original state of the code base and passes after a patch resolving the issue has been applied.
Check out our Paper for more details: SWT-Bench: Testing and Validating Real-World Bug-Fixes with Code Agents
SWT-bench uses Docker for reproducible evaluations. Follow the instructions in the Docker setup guide to install Docker on your machine. If you're setting up on Linux, we recommend seeing the post-installation steps as well.
Finally, to build SWT-bench, follow these steps:
git clone [email protected]:eth-sri/swt-bench.git
cd swt-bench
python -m venv .venv
source .venv/bin/activate
pip install -e .Test your installation by running:
python -m src.main \
--predictions_path gold \
--max_workers 1 \
--instance_ids sympy__sympy-20590 \
--run_id validate-goldWarning
Running fast evaluations on SWT-bench can be resource intensive
We recommend running the evaluation harness on an x86_64 machine with at least 120GB of free storage, 16GB of RAM, and 8 CPU cores.
You may need to experiment with the --max_workers argument to find the optimal number of workers for your machine, but we recommend using fewer than min(0.75 * os.cpu_count(), 24).
If running with docker desktop, make sure to increase your virtual disk space to have ~120 free GB available, and set max_workers to be consistent with the above for the CPUs available to docker.
Support for arm64 machines is experimental.
Evaluate model predictions on SWT-bench Lite using the evaluation harness with the following command:
python -m src.main \
--dataset_name princeton-nlp/SWE-bench_Lite \
--predictions_path <path_to_predictions> \
--max_workers <num_workers> \
--run_id <run_id>
# use --predictions_path 'gold' to verify the gold patches
# use --run_id to name the evaluation runThis command will generate docker build logs (image_build_logs) and evaluation logs (run_instance_swt_logs) in the current directory.
The final evaluation results will be stored in the evaluation_results directory.
To assess the result of a single run, we provide a simple script to assess a single evaluation run. Pass it the path to your evaluation, including run_id and model to get a simple tabellaric overview. For example, to reproduce the results for SWE-Agent from Table 2 and 3 of the paper, run the following command:
python -m src.report run_instance_swt_logs/swea__gpt-4-1106-preview/gpt4__SWE-bench_Lite__default_test_demo3__t-0.00__p-0.95__c-3.00__install-1
# |------------------------------------|--------------------------|
# | Method | swea__gpt-4-1106-preview |
# | Applicability (W) | 87.31884057971014 |
# | Success Rate (S) | 15.942028985507246 |
# | F->X | 48.18840579710145 |
# | F->P | 16.666666666666668 |
# | P->P | 9.782608695652174 |
# | Coverage Delta (Δᵃˡˡ) | 26.488815129800212 |
# | Coverage Delta Resolved (Δᔆ) | 64.69774543638181 |
# | Coverage Delta Unresolved (Δⁿᵒᵗ ᔆ) | 19.14736127176707 |In order to see a coverage delta reported, you need to have the gold evaluation included in the same evaluation path, i.e. download the golden results into run_instance_swt_logs from the downloads section below.
We list top performing methods for SWT-Bench Lite on our leaderboard. If you want to have your results included, please send us an email to [email protected] containing
- The name of your method
- The inference results from your method as the JSONL used to run the evaluation. The JSONL should contain a prediction for each instance in SWT-Bench Lite per line, each with the following fields
instance_idThe name of the instance in SWT-Bench Litemodel_name_or_pathThe name of your model/approachmodel_patchThe git patch to apply to the repositoryfull_output(optional) The complete output of your model for the given task
- Your locally determined performance
- A link to your projects homepage and traces of your method (to verify that the predictions are legitimate)
- A description on how to reproduce your runs if you want to have your results independently verified through us
We will independently run your predictions in our dockerized environment to verify your score, contact you to confirm your results and to coordinate the publication. To ensure the accessibility of traces, we reserve the right to host your predictions on a server of ours.
The inclusion in the leaderboard will be performed on a best effort basis, but we can not guarantee inclusion or timely processing of your requests.
The SWT-Bench and SWT-Bench-Lite datasets are published publicly accessible on huggingface and can be accessed using the following links. They already contain the 27k token capped context retrieved via BM25 in the prompt.
| Prompt Format | SWT-bench | SWT-bench_Lite |
|---|---|---|
| ZeroShotBase | Download | Download |
| ZeroShotPlus | Download | Download |
We provide the full traces of run code agents, the predicted patches by each method and setting and the logs of the evaluation harness.
| Artifact | Single Files | ZIP |
|---|---|---|
| Agent Traces | Download | Download |
| Predicted Patches | Download | Download |
| Evaluation Harness Logs | Download | Download |
The full list of resolved instances per approach can be found here.
To recreate the SWT-Bench dataset or create one with your own flavoring
and to run the zero-shot approaches from the paper on this dataset, follow these steps.
In order to avoid duplication, we re-use some of the SWE-Bench tooling.
First, create an SWE-Bench style dataset, i.e. by using the SWE-Bench dataset collection scripts.
If you want to add BM-25 retrieved documents, you can use the SWE-Bench BM-25 retrieval script bm25_retrieval.py - make sure to set include_tests to True to ensure that test files are included in the results.
Finally, run dataset/swt_bench.py to convert the SWE-Bench style dataset into an SWT-Bench dataset.
For example, with your SWE-Bench dataset in datasets/swe_bench, run the following commands.
python3 dataset/swt_bench.py --dataset_path datasets/swe_bench --output_path dataset/swt_bench_zsb --mode base
python3 dataset/swt_bench.py --dataset_path datasets/swe_bench --output_path dataset/swt_bench_zsp --mode plusThese commands will create the datasets for the approaches Zero-Shot Base and Zero-Shot Plus from the paper. You can then use the SWE-Bench inference tooling to generate the model inference files.
We would love to hear from the broader NLP, Machine Learning, and Software Engineering research communities, and we welcome any contributions, pull requests, or issues! To do so, please either file a new pull request or issue. We'll be sure to follow up shortly!
Contact person: Niels Mündler and Mark Niklas Müller (Email: {niels.muendler, mark.mueller}@inf.ethz.ch).
This repo is based on the SWE-Bench evaluation harness and we want to thank all their contributors.
If you find our work helpful, please use the following citations.
@inproceedings{
mundler2024swtbench,
title={{SWT}-Bench: Testing and Validating Real-World Bug-Fixes with Code Agents},
author={Niels M{\"u}ndler and Mark Niklas Mueller and Jingxuan He and Martin Vechev},
booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
year={2024},
url={https://openreview.net/forum?id=9Y8zUO11EQ}
}Please also consider citing SWE-Bench which inspired our work and forms the basis of this code-base.
@inproceedings{
jimenez2024swebench,
title={{SWE}-bench: Can Language Models Resolve Real-world Github Issues?},
author={Carlos E Jimenez and John Yang and Alexander Wettig and Shunyu Yao and Kexin Pei and Ofir Press and Karthik R Narasimhan},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=VTF8yNQM66}
}MIT. Check LICENSE.md.