SWE Bench Verified (Compressed)

Setting up all the SWE-Bench Verified images used to take over 200 GiB of storage and 100+ GiB of transfer.

Now it’s just:

• 31 GiB total storage (down from 206 GiB)
• 5 GiB network transfer (down from 100 GiB)
• ~ 5 minutes setup

🚀 Getting the Images

Images follow the naming convention:

logicstar/sweb.eval.x86_64.<repo>_1776_<instance>

Docker

curl -L -# https://huggingface.co/LogicStar/SWE-Bench-Verified-Compressed/resolve/main/saved.tar.zst?download=true | zstd -d --long=31 --stdout | docker load 

Podman

⚠️ Podman cannot load docker-archives with manifests larger than 1 MiB. We split the archive into two parts:

curl -L -# https://huggingface.co/LogicStar/SWE-Bench-Verified-Compressed/resolve/main/saved.1.tar.zst?download=true | zstd -d --long=31 --stdout | podman load 
curl -L -# https://huggingface.co/LogicStar/SWE-Bench-Verified-Compressed/resolve/main/saved.2.tar.zst?download=true | zstd -d --long=31 --stdout | podman load 

For faster downloads and parallelized loading, use the Hugging Face CLI to download the compressed OCI Layout and our load.py script to load the images in parallel:

# Clone the repo and cd into it
hf download LogicStar/SWE-Bench-Verified-Compressed layout.tar.zst --local-dir .
zstd -d --long=31 --stdout layout.tar.zst | tar -x -f -
python3 load.py

🛠 Using the Images

Just pass --namespace logicstar to the SWE-Bench harness. Example:

python -m swebench.harness.run_evaluation \
    --dataset_name princeton-nlp/SWE-bench_Verified \
    --predictions_path gold \
    --max_workers 1 \
    --run_id validate-gold \
    --namespace logicstar