Overfitting and underfitting

Why it matters

Overfitting is the central problem of machine learning: a model that has learned the noise in its training data instead of the pattern. It's why held-out data exists, why regularization exists, and why "it worked great in testing" so often fails in production. The same failure appears with LLMs: a prompt tuned until every eval example passes, or a model fine-tuned until it parrots its training set.

What you'll be able to do

• Recognize underfitting and overfitting from training and validation scores
• Draw and read a model-complexity curve
• Use the validation set to choose complexity
• List the fixes: more data, simpler models, regularization, early stopping

Before this lesson

• Train, validation and test sets
• Linear regression

Next in this phase: Regularization