Train/Validation/Test Split

This is the basic discipline of splitting your data into separate pieces so you can judge a model honestly. You fit the model on the training set, tune choices and compare models on the validation set, and only at the very end measure final performance on the untouched test set.

The whole point is that a model graded on data it studied is flattering and misleading; the test set simulates the new, unseen data the model will face in the real world. The cardinal sin is leakage — letting any information from the test set seep into training (even via choices made while peeking at it), which makes your reported performance an illusion.