Dataset downloads

CodaLab has been retired. The released datasets are now in the WILDS Hugging Face collection, including labeled and unlabeled data and the additional Waterbirds, CelebA, BDD100K, SQF, DomainNet, and ENCODE datasets. Dataset contents, splits, and evaluation protocols are unchanged.

OGB-MolPCBA remains available through the official OGB package. Its labeled data, unlabeled data, and scaffold metadata continue to use the official OGB/SNAP hosting.

Manual downloads

If your installed WILDS version still tries to download from CodaLab, install the updated code from GitHub or download the archive files manually from the collection. Extract each archive into a folder named after the archive, without .tar.gz. For example:

mkdir -p data/iwildcam_v2.0
tar -xzf iwildcam_v2.0.tar.gz -C data/iwildcam_v2.0

Keep RELEASE_v2.0.txt at the root of that folder, then load the data with:

from wilds import get_dataset
dataset = get_dataset(dataset="iwildcam", root_dir="data", download=False)

Separate unlabeled archives use their own folders, such as data/iwildcam_unlabeled_v1.0. Amazon, FMoW, and DomainNet each share one archive between their labeled and unlabeled loaders; extract those archives into the labeled dataset folder once.

Only releases listed in the collection are mirrored. Older releases remain outside this migration; existing local copies can still be used. Each dataset repository includes its original usage terms and archive checksums.

See Get Started for installation, data loading, and training instructions.