callbacks

Callbacks - hook points inside :class:~evsys_sdk.training.loop.TrainingLoop for logging, debugging, visualization, and early stopping.

The :class:Callback base has no-op defaults for every hook. Override only the ones you need:

class MyDebug(Callback): def on_step_end(self, state, step_idx, batch, metrics): if step_idx % 100 == 0: print(f"[{step_idx}] {metrics}")

Then hand to the loop:

loop = TrainingLoop(..., callbacks=[MyDebug(), CsvMetricsCallback(...)])

A failing callback never kills the loop - the exception is logged at WARNING and the loop continues. Same pattern as failing evaluators.

For early stopping (e.g. on a plateau in val pass_rate), call state.request_stop() from inside a hook and the loop breaks after the current step. Researchers writing a new "stop after N evals without improvement" policy subclass :class:EarlyStoppingCallback.

Three built-ins ship today:

• :class:PrintProgressCallback - tqdm-style stdout one-liner per step.
• :class:CsvMetricsCallback - mirror metrics.jsonl to a per-step CSV for pandas-friendly inspection.
• :class:EarlyStoppingCallback - request_stop after N evals with no improvement on a named metric.