Quickstart

This is the canonical first PyCaret session. We'll set up a classification experiment on a sample dataset, compare a dozen models, tune the winner, and persist the resulting pipeline.

The whole thing#

from pycaret.classification import ClassificationExperiment
from pycaret.datasets import get_data

# 1. Load data — any pandas DataFrame works.
data = get_data("juice", verbose=False)

# 2. Set up the experiment. Constructor params replace the 3.x setup() call.
exp = ClassificationExperiment(target="Purchase", session_id=42).fit(data)

# 3. Compare every model in the registry. Returns a CompareResult with a
#    leaderboard DataFrame plus the top-N fitted Pipelines.
top = exp.compare_models(n_select=3)
print(top.leaderboard.head())

# 4. Tune the winner. Returns a TuneResult with the search artefact.
tuned = exp.tune_model(top.best, n_iter=20)
print("best params:", tuned.best_params)

# 5. Refit on the full data + persist.
final = exp.finalize_model(tuned.pipeline)
exp.save_model(final.pipeline, "production-juice-model")

What just happened#

PyCaret 4.0's ClassificationExperiment.fit() performs:

1. A stratified train/test split (StratifiedKFold for classification).
2. Numeric mean imputation + categorical mode imputation.
3. Ordinal encoding for categorical features.
4. Label encoding for the target.
5. Fits a per-experiment metric registry (Accuracy, AUC, F1, …) that downstream verbs share.

compare_models() then iterates the model registry, calls create_model() per candidate (which runs CV and fits the model on the train split), and assembles a leaderboard from the per-fold Mean metrics.

Returned types#

Every verb returns a dataclass — no more checking pull() for the side effect. The shapes are:

The pipeline slot is always a real sklearn.pipeline.Pipeline. You can joblib.dump it, mount it behind FastAPI, or hand it to anyone who expects a sklearn estimator — no special wrappers.

Same API, every task#

The Quickstart you just read is the same shape for every supported task type:

from pycaret.regression import RegressionExperiment
RegressionExperiment(target="medv").fit(data)

from pycaret.clustering import ClusteringExperiment
ClusteringExperiment().fit(data)

from pycaret.anomaly import AnomalyExperiment
AnomalyExperiment().fit(data)

from pycaret.time_series import TimeSeriesExperiment
TimeSeriesExperiment(fh=12).fit(univariate_series)

Same verb names. Same return types. Same .pipeline contract.

Next steps#

• Concepts overview — how PyCaret is built.
• API reference — every public function, auto-generated.
• Migrate from 3.x — what changed.