เหตุใดจึงต้องกันชุดทดสอบไว้

A model that memorizes your data looks brilliant on it. The real question is how it does on data it has never seen.

So you set aside part of your data and never train on it. This locked-away slice is your test set, kept for the very end.

The model learns only from the training set. You then judge it on the untouched test set to see how it truly generalizes. 🎯

Scoring on the same rows it learned from is like grading a test with the answer key open. That number flatters the model and overstates its skill.

You do not care how well it fits old data. You care about generalization: making good predictions on tomorrow's fresh, unseen rows.

When a model nails training data but flops on the test set, it is overfitting. It memorized noise instead of learning the real pattern.

A simple, popular choice is to train on about 80% of rows and test on the remaining 20%. More data to learn, enough left to judge fairly.

The test set is sacred. If you keep peeking and tweaking until the score looks good, you have quietly leaked it into your decisions.

In scikit-learn, one helper does the splitting for you. It shuffles and carves your data into train and test parts in a single call.

The score on the test set is your honest estimate of real-world performance. Trust it more than any glowing training score.

Holding out data is not red tape. It is the one habit that stops you from shipping a model that only ever worked on paper.

Why do you keep a separate test set?

You split data, learn on the train part, and judge on a sacred test set. That untouched slice is your honest read on real-world skill. 🎯