Overfitting

Overfitting is when a model learns the training data too literally — memorizing its noise and quirks instead of the general pattern — so it scores well in training but fails on new, unseen inputs. It's a core machine-learning failure mode and the reason you always evaluate on data the model never saw. For SaaS builders, overfitting is the practical risk whenever you fine-tune or train a small custom model: feed it a few hundred examples for too many epochs and it will parrot them back beautifully while generalizing poorly to real user inputs. Signs include a large gap between training and validation performance and a model that's brittle to small rephrasings. Defenses are standard: hold out a validation and test set, stop training when validation stops improving (early stopping), use more and more varied data, and prefer lightweight adaptation (LoRA, prompting, retrieval) over heavy fine-tuning. The same idea shows up in evals — a model or prompt "overfit" to your benchmark can look great on it and disappoint in production.