Early stopping regularizes training by ending optimization when validation performance no longer improves. It avoids choosing a training duration that lets the model overfit its training samples. section 9.3.4
Configure early stopping around repeated validation checks:
- Enable evaluation during training and provide a validation dataset through eval_df. section 9.3.4
- Set patience to the number of consecutive evaluations without improvement; use delta to define meaningful improvement. section 9.3.4
- Save the best-performing checkpoint instead of retaining unnecessary checkpoints after each evaluation or epoch. section 9.3.4
Fine-tuning BERT for toxic-comment classification
A multilabel BERT classifier evaluates every 1,000 training steps, allows four consecutive evaluations without improvement, and treats any positive change as improvement with early_stopping_delta=0. Training receives train_df and eval_df, then the saved best model is evaluated separately.
