Overfitting happens when a model memorizes its training examples instead of learning a process that works on new cases. It can look successful because training loss is low while performance on unseen data remains poor. section 1.29 section 1.27
A fine-tuning trap
Fine-tuning an LLM on a small collection of style examples can make it specialize too aggressively. The model may reproduce those examples well but handle other instances badly. More specialization is not automatically better. section 3.3
Control overfitting by limiting how narrowly the model adapts:
- Use early stopping: stop fine-tuning when performance on a separate validation set stops improving. section 3.3
- Limit updates: change only selected parts of the model, as with LoRA layers that preserve the original parameters. section 3.3
- Use regularization or dropout to discourage parameters or internal units from specializing too closely to individual examples. section 3.3
- Compare validation and test performance to choose a fine-tuning strategy and check whether the final model generalizes. section 3.3
The practical test is simple: a model should work beyond the examples it saw during training. Low training loss alone cannot establish that. section 1.29 section 1.27
