---

name: good-ood-datasets description: "Use DIG's GOOD OOD dataset loaders for domain/shift splits, metadata, and graph OOD benchmark access across molecule, node, and synthetic graph domains." disable-model-invocation: true metadata: disco-role: operating license: GPL 3.0

GOOD OOD Datasets

Use this sub-skill for DIG's GOOD dataset loaders and the metadata they return.

Include

• dig.oodgraph: GOODHIV, GOODPCBA, GOODZINC, GOODCMNIST, GOODMotif, GOODCora, GOODArxiv, GOODCBAS.
• GOOD*.load(root, domain, shift='no_shift', generate=False) dataset constructors and returned meta_info dictionaries.
• Domain/shift/subset selection, download handling, and benchmark metadata for graph OOD workflows.

Exclude

• Molecular generation, SSL, explainability, augmentation, fairness, or large-scale graph workflows.

Start Here

• Read references/data-formats.md for the domain/shift/subset layout.
• Read references/workflows.md for the supported load patterns.
• Read references/troubleshooting.md when a download or split name is wrong.
• Run scripts/good_metadata_check.py for a safe import-and-metadata smoke check.

Core Workflows

• Graph OOD loading: choose a dataset class, call .load(...), and use the returned train/id-val/id-test/val/test mapping plus meta_info.
• Domain and shift selection: match the dataset-specific allowed domains (scaffold, size, color, basis, word, degree, time) and shifts (no_shift, covariate, concept).
• Benchmark setup: use meta_info.dim_node, meta_info.dim_edge, meta_info.num_envs, and meta_info.num_classes to configure downstream models.