Think Together · Start with a Short Story
The girl gives the robot a stack of cat and dog pictures, then puts another stack in a bag: “Save these to test you later. Do not look at them yet.”
Lay Out the Process
- Training setUsed to adjust the model
- Validation setHelps choose settings
- Test setSaved for evaluation
Training data helps a model learn; evaluation data shows how it performs on cases it has not seen.
If you use the exam answers for repeated practice beforehand, you can no longer treat that score as a fair, independent check.
Real datasets also need checks for incorrect labels, repeated examples, and uneven categories, so a good-looking score does not hide problems.
Now You Try
If you rename the same photo and put it in the test set, is it a new picture the model has never seen?
Ready? Open this to see the explanation
No. The picture itself is unchanged, so the test is no longer properly independent.
One Idea to Take Away
Training data helps a model learn; evaluation data shows how it performs on cases it has not seen.
Try covering the diagram above and explaining each step to someone in your family. If one step is hard to explain, return to that card and look again.