---

name: moabb description: "Guide MOABB EEG and BCI benchmarking workflows: choose datasets, configure paradigms and sklearn pipelines, run leakage-aware evaluations, and analyze reproducible results." disable-model-invocation: true metadata: disco-role: operating license: BSD 3-Clause

MOABB

MOABB (Mother of All BCI Benchmarks) is a Python framework for reproducible EEG brain-computer-interface experiments. Use this skill when a task involves MOABB dataset classes, MNE Raw/Epochs, motor imagery, P300/ERP, SSVEP, c-VEP, scikit-learn pipelines, cross-session or cross-subject scores, benchmark result stores, or MOABB analysis plots.

Install and inspect first

Use a supported Python environment (the package metadata for this snapshot requires Python 3.11 or newer), then install the base distribution:

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

The base install includes the core MNE, mne-bids, NumPy/SciPy, pandas, scikit-learn, pyRiemann, YAML, and data-access dependencies used by the routes below. Optional extras are deliberate: moabb[deeplearning], moabb[optuna], moabb[interactive], and moabb[carbonemission] add separate capability surfaces; do not install them merely to make a core CPU workflow work. Read API overview for the public object graph and troubleshooting before changing an installation or enabling network data access. Run scripts/check_environment.py for a read-only import and optional-dependency diagnosis; use its --optional flag only when checking optional integrations.

Route by the user's deliverable

• Dataset discovery, catalog filtering, subjects/sessions, MNE data roots, local/BIDS data, FakeDataset, or download/cache recovery: read dataset-management.
• Motor imagery, imagined speech, P300, SSVEP, c-VEP, fixed windows, preprocessing, feature extraction, sklearn classifiers, filter banks, or YAML pipeline construction: read paradigms-and-pipelines.
• Within-session, within-subject, cross-session, cross-subject, splitters, learning curves, benchmark recipes, caching, or leakage checks: read evaluations-and-benchmarks.
• Result DataFrames/Results, chance levels, paired statistics, score or distribution plots, timelines, or analysis reports: read analysis-and-visualization.

Most end-to-end tasks use several routes in this order: dataset → paradigm and pipeline → evaluation → analysis. Keep each boundary explicit so a score's generalization claim and preprocessing information budget remain auditable.

Safe operating defaults

1. Start with FakeDataset or a caller-provided local/BIDS root for API and shape checks. A real dataset's data_path()/get_data() can download files; ask for network, license, storage, and time approval before using it.
2. Set dataset subjects, sessions, event labels, frequency/window parameters, and the generalization target explicitly. Inspect X.shape, y, metadata, channel names, sampling rate, and the selected scorer before fitting.
3. Put every fitted transform (scaling, CSP, covariance, feature extraction, classifier, or template) inside the sklearn pipeline passed to the evaluation. Never fit on all trials before a split.
4. Begin with random_state fixed and n_jobs=1; use unique result/cache paths and overwrite=False. Increase data, workers, and optional features only after the tiny path is correct.
5. Treat results as metric-specific. Two-class paradigms commonly score ROC-AUC; multiclass accuracy uses a different chance reference. Preserve the evaluation protocol, metric, dataset selection, cache settings, and pipeline name beside any reported result.

Bundled checks

The root diagnostic is linked above. For offline workflow checks, use the helpers linked by each sub-skill: they create only synthetic data or temporary caller-owned output and never download a dataset.

Provenance and staleness

Read repository provenance before applying this skill to a checkout or deciding whether to refresh it. This graph was extracted from a specific MOABB revision; public dataset catalogs, signatures, optional dependencies, and evaluation behavior can change. Refresh when the source revision, package version, public imports, or major evidence paths change.