---

name: tabular-ml description: "Build, evaluate, customize, troubleshoot, and deploy AutoGluon TabularPredictor workflows for supervised tabular machine learning." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

AutoGluon Tabular ML

Use this sub-skill for supervised tabular classification, regression, and quantile prediction with autogluon.tabular.TabularPredictor and TabularDataset.

Route Here When

• The task asks for TabularPredictor, TabularDataset, leaderboard, feature importance, model management, save/load, or refit workflows.
• The data is a pandas-like table with one target column and rows as samples.
• The user needs presets, hyperparameters, bagging/stacking, HPO, resource limits, sample weights, groups, or feature metadata for tabular data.
• The user wants model families such as LightGBM, CatBoost, XGBoost, RandomForest, ExtraTrees, KNN, linear, neural network, or optional tabular foundation models inside tabular AutoML.

Route Elsewhere

• Forecasting with item ids, timestamps, horizons, covariates, or TimeSeriesPredictor: use ../time-series-forecasting/.
• Foundation-model workflows for text, image, document, object detection, semantic matching, or MultiModalPredictor: use ../multimodal-automl/.
• Cross-package install/import/backend diagnostics that affect all AutoGluon packages: use the root references/troubleshooting.md.

Start With These References

• references/api-reference.md: constructor and method signatures, metrics, model management, save/load, and deployment-facing APIs.
• references/workflows.md: novice-to-expert fit/evaluate/predict workflows, quantile prediction, HPO, bagging, stacking, refit, distillation, and deployment cleanup.
• references/data-and-features.md: TabularDataset, schema expectations, feature metadata, automated feature generators, text columns, sample weights, and groups.
• references/customization.md: presets, hyperparameter dictionaries, custom metrics, custom models, resource controls, and optional model families.
• references/troubleshooting.md: actionable fixes for optional dependencies, schema drift, problem type errors, resource limits, HPO/Ray, save/load mismatch, and predictor path issues.

Bundled Scripts

• scripts/tabular_smoke.py: trains a tiny in-memory TabularPredictor, evaluates it, predicts from a reloaded predictor, and prints a JSON summary.
• scripts/inspect_tabular_predictor.py: read-only environment/predictor inspection for versions, importability, optional packages, and saved predictor metadata checks.

Run each script with --help first. The smoke script avoids network access and writes only to a temporary directory unless --output-dir is supplied.

Fast Patterns

from autogluon.tabular import TabularPredictor

predictor = TabularPredictor(label="target", eval_metric="accuracy").fit(
    train_data,
    presets="medium_quality",
    time_limit=60,
    hyperparameters={"GBM": {}, "RF": {"n_estimators": 50}},
    num_cpus="auto",
    num_gpus=0,
)
leaderboard = predictor.leaderboard(test_data, extra_info=True)
predictions = predictor.predict(test_data.drop(columns=["target"]))

Use an explicit path when the predictor must be preserved. Use a new empty path for every independent training run to avoid overwriting prior results.

Safety And Quality Defaults

• Prefer presets="medium_quality" or a small explicit hyperparameters dictionary for smoke tests and CPU-constrained repros.
• Prefer presets="good_quality" or "high_quality" with a real time_limit for production candidates; reserve "best_quality" and "extreme_quality" for larger budgets.
• For deployment, combine quality presets with "optimize_for_deployment" or call delete_models(dry_run=True) before destructive cleanup.
• For moved saved predictors, load only trusted predictor directories and start with strict version checks before relaxing require_version_match or require_py_version_match.