model training convergence

Early stopping halts training when validation performance stops improving, preventing overfitting. **Mechanism**: Monitor validation metric each epoch/N steps. If no improvement for patience epochs, stop. Use best checkpoint. **Why it works**: Training loss keeps decreasing but validation loss starts increasing = overfitting. Stop at inflection point. **Hyperparameters**: Patience (how many epochs without improvement), min_delta (minimum improvement to count), metric (validation loss, accuracy, etc.). **Typical patience**: 3-10 epochs for vision, varies for other domains. Longer patience for noisy metrics. **Implementation**: Track best validation score, count epochs since improvement, stop and restore best weights. **Trade-offs**: Too aggressive (low patience) may stop during noise. Too lenient may overfit. **Modern alternatives**: Many LLM training runs use fixed schedules instead, validated by scaling laws. Early stopping more common for fine-tuning. **Regularization alternative**: Instead of stopping, can use regularization to prevent overfitting while training longer. **Best practices**: Always use for fine-tuning limited data, validate patience setting empirically, save best checkpoint.