Overfitting
Overfitting is when a model memorises its training data instead of learning the underlying pattern: near-perfect on examples it has seen, unreliable on anything new.
No jargon, no math
In plain words, what is it?
Training reduces error on the training set, and there is a point where that stops being progress. Keep adjusting and the model starts fitting the noise in the data: the exact typos, the one-off spikes, the quirks of last quarter. The tell is in two curves: training error keeps falling while error on unseen data rises. The parameters have effectively memorised the flashcards. The fixes follow from the diagnosis: stop earlier (early stopping), gather more or better data, use a smaller model, or regularise. The honest verdict always comes from the test set the model never touched, never from the training score.
- 01
Train. The model adjusts its parameters on the training examples.
- 02
Test on unseen data. The held-out test set shows if the pattern generalises.
- 03
Spot the tell. Great on training, poor on new data means memorisation.
- 04
Fix. More data, early stopping or a smaller model break the memorisation.
Terms that live nearby.
Training
Training is the process that turns an untrained model into a working AI: you feed it thousands or millions of examples, compare its guesses with the correct answers, and nudge its internal settings until the guesses get good.
ReadDataset
A dataset is the organised collection of examples an AI learns from or is judged on: thousands or millions of rows of text, images or sensor readings, cleaned, labelled and split into training, validation and test parts.
ReadModel Evaluation
Model evaluation is the process of measuring how well a model performs on data it has never seen, using metrics and test sets agreed in advance, before you trust it with real decisions.
Read§ 04
Want this running in your stack?
We design and ship AI systems with permissions, logs and human approval. It starts with a 30-minute scoping call.