04·Evaluation: train/val/test split, accuracy, when to stop training

Introduction

A trained model without evaluation is a "pig in a poke". This lesson lays the foundation: why we split data into train/val/test, which metrics are appropriate for classification and regression, how to read loss/accuracy curves to diagnose overfitting and underfitting, and how the early-stopping technique lets you automatically pick the best moment to stop training. Without this lesson you do not know whether your model works.