Introduction

GridSearchCV does two different kinds of training, and the answer depends on which stage you mean. During cross-validation it does not train on the whole dataset at once, but after it chooses the best parameters it can refit the best estimator on the entire dataset you passed to fit, as long as refit=True.

What Happens During Cross-Validation

The main search phase evaluates each parameter combination with cross-validation. That means each model is trained on only the training folds for that split and validated on the held-out fold.

For five-fold cross-validation, the idea looks like this:

• split 1: train on folds 2 through 5, validate on fold 1
• split 2: train on folds 1, 3, 4, and 5, validate on fold 2
• split 3: train on folds 1, 2, 4, and 5, validate on fold 3
• and so on

So during scoring, no single fold-trained model sees the entire dataset at once.

The Important refit Step

After the cross-validation scores are computed, GridSearchCV can optionally train one final model using the best hyperparameters on all the data you supplied to fit.

That behavior is controlled by refit.

1from sklearn.datasets import load_iris
2from sklearn.model_selection import GridSearchCV
3from sklearn.svm import SVC
4
5X, y = load_iris(return_X_y=True)
6
7search = GridSearchCV(
8    estimator=SVC(),
9    param_grid={"C": [0.1, 1, 10], "kernel": ["linear", "rbf"]},
10    cv=5,
11    refit=True,
12)
13
14search.fit(X, y)
15
16print(search.best_params_)
17print(search.best_estimator_)

With refit=True, which is the default, best_estimator_ is trained again on the full dataset X, y that you passed to fit.

So the short answer is:

• during cross-validation: no
• after model selection with default settings: yes, the best model is refit on the whole input dataset

What If refit=False?

If you set refit=False, the search still performs cross-validation and still finds the best parameter combination according to the scores, but it does not train one final estimator on the whole dataset afterward.

1search = GridSearchCV(
2    estimator=SVC(),
3    param_grid={"C": [0.1, 1, 10]},
4    cv=5,
5    refit=False,
6)
7
8search.fit(X, y)
9print(search.best_params_)

In that case:

• you still get score information and best-parameter information
• you do not get a refit best_estimator_ ready for direct prediction

If you want a final model, you must instantiate the estimator yourself with best_params_ and fit it manually.

Why This Design Exists

This two-stage behavior is intentional.

Cross-validation answers the model-selection question:

• which parameter setting performs best on held-out folds?

Refitting answers the deployment question:

• now that we know the best settings, train one final estimator using all available training data

Those are separate goals. GridSearchCV supports both because real workflows usually need both.

Whole Dataset Means Whole Training Dataset

A subtle but important point: when people say "whole dataset," they should usually mean the whole dataset supplied to GridSearchCV.fit, not necessarily the whole dataset in your project.

If you already split your data into training and test sets, the normal workflow is:

• run GridSearchCV only on the training set
• allow it to refit the best estimator on that full training set
• evaluate the final estimator on the untouched test set

1from sklearn.model_selection import train_test_split
2
3X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
4search.fit(X_train, y_train)
5print(search.score(X_test, y_test))

This preserves the test set as a final evaluation set.

Do Not Confuse Search with Final Evaluation

A common misconception is that because GridSearchCV uses all training data across folds, there is no need for a separate test set. That is wrong.

Cross-validation inside the search is still part of model selection. If you report that score as your final external evaluation without holding out truly untouched data, you risk optimistic estimates.

So the safe interpretation is:

• cross-validation is for choosing parameters
• the refit model is for final training on the training split
• the separate test set is for final evaluation

Common Pitfalls

One common mistake is assuming GridSearchCV trains one model on the full dataset during the search stage. It does not; the search stage uses repeated fold-based training.

Another pitfall is forgetting that refit=False disables the final full-data retraining step.

A third issue is assuming the refit step means you no longer need a separate test set. You still do if you want an unbiased final evaluation.

Finally, remember that the whole-data refit uses the data passed to fit, not magically every dataset you own. If you call fit on the training split only, the refit also uses the training split only.

Summary

• During cross-validation, GridSearchCV trains on training folds, not on the entire dataset at once.
• With the default refit=True, it trains one final best estimator on all the data passed to fit after the search ends.
• With refit=False, no final whole-data retraining happens automatically.
• In a proper workflow, the refit uses the full training set, while the test set remains untouched for final evaluation.
• The answer is therefore "not during scoring, but usually yes after selection if refit is enabled."