Train/Validation/Test Done Right

You've met the test set — but honest evaluation actually needs three partitions, each with a single job:

• Training set: the data the model learns from
• Validation set: the data you learn from — every choice of learning rate, architecture, or feature set is judged here
• Test set: the untouched final exam, used once at the end

Why the middle split? Because tuning on the test set is still fitting to it — through your decisions rather than the model's weights.