Training ends doesn't mean model is good. Judging by one number "95% accuracy" is risky—on data where fraud is 1%, saying "all normal" yields 99% accuracy. Model evaluation's essence is measuring "generalization performance" with "problem-fitting metrics".
In the MLA-C01 exam, evaluation appears as "which metric on imbalanced data", "how to diagnose and mitigate overfitting", "compute precision/recall from confusion matrix". Today we cover metric selection, overfitting/underfitting, cross-validation, confusion matrix—four axes.
Classification evaluation starts with the confusion matrix