---

name: tabular-validation description: "Build Deepchecks tabular Dataset objects, suites, checks, model inputs, scorers, and failure recovery for tabular validation." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Tabular Validation

Use this sub-skill when the task involves pandas/numpy tabular data, deepchecks.tabular.Dataset, tabular built-in suites, individual tabular checks, sklearn-style tabular models, precomputed tabular predictions, or tabular scorer/condition customization.

Fast route

1. Build Dataset objects with explicit metadata: label, features, cat_features, index_name, datetime_name, label_type, and dataset_name where available.
2. Choose the smallest suite:
   • single dataset quality: deepchecks.tabular.suites.data_integrity
   • train/test split validation: deepchecks.tabular.suites.train_test_validation
   • model performance, predictions, drift, and weak segments: deepchecks.tabular.suites.model_evaluation
   • broad first-pass investigation: deepchecks.tabular.suites.full_suite
3. Prefer with_display=False in automation; send HTML/JSON/CI/report saving to results-and-integrations.
4. For model checks, pass either a fitted sklearn-style model or aligned y_pred_* / y_proba_* arrays, plus model_classes and feature_importance when needed.
5. For check/suite API details, load API reference. For copyable recipes, load workflows. For failures, load tabular troubleshooting.

Boundaries and routing

• Use nlp-validation for TextData, text classification, token labels, embeddings, and NLP properties.
• Use vision-validation for VisionData, image batches, detection, segmentation, torch/torchvision loaders, and vision metrics.
• Use results-and-integrations for show, save_as_html, JSON serialization, pytest/CI gating, and integration artifact handling.
• Use root troubleshooting for package installation, import, optional dependency, or environment-wide display issues before debugging a tabular object.

Bundled helper

Run scripts/deepchecks_tabular_smoke.py for a safe local smoke using a tiny in-memory DataFrame and sklearn model. It performs no downloads, credentials, or destructive writes by default; --html-out writes only to the path explicitly supplied by the caller.