---

name: datasets description: "Use Fairlearn built-in dataset loaders, schema notes, cache controls, and dataset fairness warnings." disable-model-invocation: true metadata: disco-role: operating license: MIT

Fairlearn datasets

Use this sub-skill when the task asks for Fairlearn built-in datasets, OpenML-backed loaders, dataset schemas, fetch_adult, fetch_acs_income, fetch_bank_marketing, fetch_boston, fetch_credit_card, fetch_diabetes_hospital, as_frame, return_X_y, data_home, cache, ACS state filtering, or DataFairnessWarning.

Quick workflow

1. Decide whether the task needs a real dataset download or only loader/schema guidance.
2. Choose the smallest relevant loader and set data_home explicitly in restricted or reproducible environments.
3. Use as_frame=True for pandas DataFrames with dtypes and feature names; use return_X_y=True for direct (X, y) workflows.
4. Keep dataset fairness warnings visible, especially for Boston housing.
5. Route to ../assessment/ or a mitigation sub-skill after the dataset is loaded and preprocessed.

Read these references

• references/dataset-loaders.md for loader signatures, dataset sizes, columns, return shapes, cache behavior, and loader-specific caveats.
• references/troubleshooting.md for network/cache failures, ACS state-code errors, Boston warnings, and as_frame/return_X_y confusion.
• scripts/preview_dataset_loaders.py for a no-network signature preview and optional single-loader download check.

Core APIs to recognize

• fetch_adult(*, cache=True, data_home=None, as_frame=True, return_X_y=False)
• fetch_acs_income(*, cache=True, data_home=None, as_frame=True, return_X_y=False, states=None)
• fetch_bank_marketing(*, cache=True, data_home=None, as_frame=True, return_X_y=False)
• fetch_boston(*, cache=True, data_home=None, as_frame=True, return_X_y=False, warn=True)
• fetch_credit_card(*, cache=True, data_home=None, as_frame=True, return_X_y=False)
• fetch_diabetes_hospital(*, as_frame=True, cache=True, data_home=None, return_X_y=False)

Boundary rules

• This sub-skill owns loader behavior and dataset caveats. It does not own model training, assessment metrics, or mitigation algorithms.
• Use ../assessment/ after obtaining predictions or scores.
• Use ../preprocessing/, ../reductions/, ../postprocessing/, or ../adversarial/ for mitigation choices.
• Use ../installation/ if the issue is a package import rather than loader behavior.

Operating rules

• Dataset loaders can perform network downloads. Ask before downloading when network, disk, or privacy constraints are unclear.
• Default cache location is a .fairlearn-data directory under the user's home unless data_home is set.
• Prefer as_frame=True for Fairlearn workflows because column names make sensitive-feature selection auditable.
• fetch_boston(warn=True) intentionally raises DataFairnessWarning; do not suppress it in reports unless the user explicitly asks and understands why.
• fetch_acs_income(states=...) expects two-letter state abbreviations; PR is accepted for Puerto Rico.

Fast validation

No-network signature preview:

python sub-skills/datasets/scripts/preview_dataset_loaders.py

Optional single-loader download check:

python sub-skills/datasets/scripts/preview_dataset_loaders.py --download adult --data-home /tmp/fairlearn-data