GOOD: A Graph Out-of-Distribution Benchmark

This repository mirrors every original downloadable dataset archive used by divelab/GOOD, including GOODTwitter and LBAPcore. The 13 ZIP files are byte-for-byte copies of the original Google Drive archives. No tensors were reserialized, samples regenerated, or splits changed.

The full archive inventory, original source URLs, byte sizes, SHA-256 hashes, ZIP member CRCs and upload commits are recorded in dataset-manifest.json. The source inventory is pinned to GOOD commit b53566c9297bc65b90a7f2213fb9ffa930f5b6e5.

Files

These archives contain PyTorch/PyG graph objects rather than a tabular dataset. Use the GOOD loaders; loading this repository with datasets.load_dataset is not the intended interface. Expanded sizes include every original archive member.

Original archive	Domains	Compressed bytes	Expanded bytes
GOODArxiv.zip	degree, time	526,412,117	797,981,490
GOODCBAS.zip	color	37,184	384,683
GOODCMNIST.zip	background, color	718,883,636	10,021,483,990
GOODCora.zip	degree, word	23,576,007	4,154,265,280
GOODHIV.zip	scaffold, size	37,564,781	1,036,090,346
GOODMotif.zip	basis, size	11,815,722	262,128,686
GOODPCBA.zip	scaffold, size	437,884,944	12,591,427,498
GOODSST2.zip	length	6,134,206,091	6,694,135,702
GOODTwitch.zip	language	28,840,838	101,379,878
GOODTwitter.zip	length	1,257,110,614	1,370,566,358
GOODWebKB.zip	university	293,664	12,732,006
GOODZINC.zip	scaffold, size	206,477,730	5,722,268,842
LBAPcore.zip	assay, scaffold, size	78,230,762	4,900,451,599

Automatic downloads in GOOD

Use the GOOD checkout containing the Hugging Face download update and install its dependencies (pip install -e .) in your existing supported GOOD environment. The original training commands and dataset APIs remain unchanged. A first run automatically fetches the selected ZIP from this repository; complete existing datasets are reused offline, without a Hugging Face login.

from GOOD.data.good_datasets.good_motif import GOODMotif

datasets, meta_info = GOODMotif.load(
    dataset_root="./storage/datasets",
    domain="basis",
    shift="covariate",
)

With an older GOOD release, you can also download the unchanged archives directly:

from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="citrinegui/GOOD", repo_type="dataset",
    filename="GOODMotif.zip", token=False,
)

Extract the chosen archive into the configured dataset_root, preserving its top-level class-name directory (for example GOODMotif/basis/processed/). The original loader can then reuse it. Do not rename folders or change the serialized tensor format.

The Hub cache stores the compressed ZIP; allow disk space for the cached archive and the expanded dataset. HF_HOME relocates the cache. GOOD's explicit generate=True mode still uses the original generation code and upstream raw sources; this mirror preserves the published processed datasets.

LBAPcore provides covariate splits only. This migration does not change GOOD's PyTorch/PyG compatibility requirements or its model implementations.

License and attribution

The original dataset MIT license is preserved verbatim: Copyright (c) 2022 DIVE Lab, Texas A&M University. The GOOD code is separately licensed under GPLv3; see the upstream repository. Upstream source-data references and their attribution are retained in the GOOD dataset documentation.

If you use GOOD, please cite:

@inproceedings{gui2022good,
  title={{GOOD}: A Graph Out-of-Distribution Benchmark},
  author={Shurui Gui and Xiner Li and Limei Wang and Shuiwang Ji},
  booktitle={Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year={2022},
  url={https://openreview.net/forum?id=8hHg-zs_p-h}
}