Overfit

How Overfit Works

This is the bias-variance tradeoff disguised as a finger-painting exercise. You get a scatter of blue and red dots: that's your training set. Draw a closed loop around the blues. Everything inside your loop, the model predicts blue; everything outside, red. Then hit Deploy and a fresh, unseen test set rains down from the same distribution.

1. Drag your finger or mouse to draw a closed boundary around the blue cluster
2. Watch your live train accuracy climb as you enclose blues and exclude reds
3. Hit Deploy to Prod to reveal the hidden test set and score your real accuracy
4. Survive the round (beat the test-accuracy threshold) and the data gets messier

The Trap

Trace every training dot exactly and you'll hit 100% train accuracy. Congratulations: you've memorized the noise. When the test set lands, your jagged boundary slices straight through new dots and your real accuracy craters. A smooth, generous loop that captures the shape of the blue cluster, not its every pixel, is what actually generalizes. Too tight is overfitting. Too lazy is underfitting. Find the sweet spot.