Callbacks

Callbacks receive lifecycle events from Trainer: on_fit_start, on_epoch_start, on_batch_end, on_epoch_end, and on_fit_end.

Built-ins:

Key	Purpose
checkpoint_callback	Save best monitored checkpoints and optionally last.pt.
early_stopping_callback	Stop after monitored values fail to improve.
snapshot_callback	Save config and/or a source ZIP at fit start.
collect_predictions_callback	Export cached split predictions as CSV artifacts.
plot_confusion_matrix_callback	Export cached confusion matrices as PDF artifacts.
sweep_target_callback	Track best/last scalar values for Optuna.
telegram_notifier_callback	Send completion notification using environment variables.

Callbacks monitor exact flat log keys such as val/loss or plugin-produced keys. The key must exist for checkpointing, early stopping, or sweep objectives to act.

The Telegram callback reads TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID by default. Do not place tokens in YAML or documentation.