---

name: aif360 description: "Use IBM AI Fairness 360 for tabular fairness datasets, metrics, bias mitigation algorithms, sklearn-compatible workflows, subgroup detectors, and explainers." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

AIF360 Repo Skill

Use this skill when a task involves the AI Fairness 360 (aif360) Python package: fairness datasets, group-fairness metrics, bias-mitigation algorithms, the preferred aif360.sklearn API, MDSS/FACTS subgroup detectors, or metric explainers.

Install and smoke-check first

Base package install:

pip install aif360
python -c "import aif360; print(aif360.__version__)"

For local package development, use an editable install in an isolated Python environment. For optional algorithms, install the smallest named extra that matches the requested workflow rather than aif360[all] by default; read install-data-and-optional-deps.md.

Run the bundled base smoke when importability or optional-warning noise is unclear:

python scripts/check_aif360_env.py --json

The smoke uses only synthetic in-memory data and does not download benchmark datasets or require optional extras.

Route by task

Task signal	Read
Legacy BinaryLabelDataset, StructuredDataset, standard dataset wrappers, BinaryLabelDatasetMetric, ClassificationMetric, raw-data availability, or protected group dictionaries	datasets-and-metrics
Legacy aif360.algorithms preprocessing, inprocessing, postprocessing, deterministic reranking, optional mitigation extras, or before/after metric validation	mitigation-algorithms
Preferred pandas/sklearn API, aif360.sklearn.datasets.fetch_*, protected attributes in pandas indexes, sklearn metrics/scorers, sklearn estimators, or PostProcessingMeta	sklearn-interface
MDSS subgroup bias scan, FACTS recourse subgroup reports, MetricTextExplainer, or MetricJSONExplainer	detectors-and-explainers
Install/import failures, raw dataset files, optional extras, warnings, Python versions, R wrapper, or MLOps sample boundaries	Root references below

Core decision points

• Legacy vs sklearn API: choose legacy APIs for BinaryLabelDataset and legacy algorithms; choose aif360.sklearn for pandas/sklearn pipelines and future-facing DataFrame workflows.
• Data availability: AIF360 documents common benchmark datasets, but most raw files are not bundled. Do not download data unless the user approves data acquisition and terms.
• Optional extras: many algorithms are public but extra-gated. Missing TensorFlow, fairlearn, torch, cvxpy, BlackBoxAuditing, POT, FACTS, or R/rpy2 packages are not base-install failures unless the selected workflow needs them.
• Fairness semantics: always make favorable_label, pos_label, privileged group, unprivileged group, protected attributes, and row alignment explicit before interpreting a metric or detector result.
• Verification status: this skill verified base CPU package imports and synthetic dataset/metric workflows. Optional-extra workflows are documented but must be installed and smoke-tested in the user's runtime before claiming execution support.

Root references and scripts

• repo-provenance.md: source snapshot and refresh baseline for this generated skill.
• repo-routing-metadata.json: structured metadata for repo-skills-router import.
• install-data-and-optional-deps.md: package install variants, extras, standard dataset data constraints, and construction verification status.
• troubleshooting.md: cross-cutting import, optional dependency, data, API-family, and workflow recovery guidance.
• r-and-mlops-notes.md: boundaries for the R package wrapper and platform integration samples.
• check_aif360_env.py: safe base environment diagnostic helper.

Quick operating checklist

1. Identify whether the user is using legacy datasets or aif360.sklearn.
2. Confirm data source and raw-data/network permissions.
3. Confirm protected attributes, privileged/unprivileged groups, and favorable labels.
4. Install only the base package plus selected extras needed by the workflow.
5. Run a bundled synthetic smoke before running real data or optional training.
6. Use metrics or detectors to diagnose bias, then route to mitigation only if the user asks to change data/model/predictions.
7. If the current package commit/version differs from repo-provenance.md, refresh this skill before relying on exact API claims.