Rocket slaps the last labeled photo card onto the pile. "Our sorting model is done!" he cheers.
He shows it the same cards it learned from. It sorts every one perfectly.
Raven tilts her head. "What do you notice about the cards you used for the check?"
Nova dims her lights. "Hmm," she says. "Would a quiz be fair if you saw every answer first?"
Last week the crew taught a paper sorting model with labeled photo cards.
Those cards are the training set. They are the examples the model learned from.
Checking the model with the same cards is like a quiz you already saw.
Of course it does well! That does not prove it learned the pattern.
So we keep some labeled cards back. We never use them to build the model.
Those saved cards are the test set. They check the model fairly.
| Statement | True or false? |
|---|---|
| A test set is used to build the model. | ? |
| Test cards should be new to the model. | ? |
| Getting every training card right proves the model is perfect. | ? |
| Test cards still need correct labels. | ? |
Each test card still has its true label on the back.
The model makes its guess first. Then we flip the card and compare.
A match is a right answer. A mismatch is an error.
Errors are not bad news. They show us where the model needs help.
Nice thinking, explorer! Tomorrow we test the crew's model and count the results.