Definition
An overfit model scores brilliantly on training data but fails on fresh inputs. It learned quirks of the training set — typos, duplicates, spurious correlations — rather than true signal.
Validation sets, regularization, and more diverse data are standard defenses.
In simple terms
Overfitting is memorizing exam questions instead of understanding the subject — perfect on practice tests, lost on new questions.
Where you see it
- A classifier that only works on training screenshots, not live user data.
- LLMs that repeat training phrasing verbatim on niche prompts.
How it works
1.Train
Model capacity may exceed what the data justifies.
2.Monitor validation
Watch validation loss diverge from training loss.
3.Regularize
Early stopping, dropout, simpler models, or more data.
Why it matters
- Overfitting is the most common reason models fail after promising offline metrics.
Often confused
Zero training error means a great model.
It often means memorization — check validation performance.