Method train_sft

src/art/model.py:1109–1153

Supervised fine-tune the model with an iterable of trajectories. Args: trajectories: An iterable of Trajectory objects. config: SFT configuration including learning_rates and batch_size. If None, uses default TrainSFTConfig(). _co

(
        self,
        trajectories: Iterable[Trajectory],
        config: TrainSFTConfig | None = None,
        _config: dev.TrainSFTConfig | None = None,
        verbose: bool = False,
    )

        await self.log(groups_list, split="train", metrics=avg_metrics, step=step)

    async def train_sft(
        self,
        trajectories: Iterable[Trajectory],
        config: TrainSFTConfig | None = None,
        _config: dev.TrainSFTConfig | None = None,
        verbose: bool = False,
    ) -> None:
        """
        Supervised fine-tune the model with an iterable of trajectories.

        Args:
            trajectories: An iterable of Trajectory objects.
            config: SFT configuration including learning_rates and batch_size.
                If None, uses default TrainSFTConfig().
            _config: Additional experimental configuration that is subject to change and
                not yet part of the public API. Use at your own risk.
            verbose: Whether to print verbose output.
        """
        if config is None:
            config = TrainSFTConfig()

        # Train (backend yields metrics for each batch without logging)
        # Collect all metrics and aggregate them at the end (same as RL)
        _config = _config or {}  # ty:ignore[invalid-assignment]
        training_metrics: list[dict[str, float]] = []
        trainer_started = time.monotonic()
        async for metrics in self.backend()._train_sft(
            self,
            trajectories,
            config,
            _config,  # ty:ignore[invalid-argument-type]
            verbose,
        ):
            training_metrics.append(metrics)
        trainer_elapsed = time.monotonic() - trainer_started

        # Log aggregated training metrics once (same as RL)
        if training_metrics:
            avg_metrics = average_metric_samples(training_metrics)
            avg_metrics["time/step_trainer_s"] = trainer_elapsed
            # Get the current step after training
            step = await self.get_step()
            await self.log(
                trajectories=None, split="train", metrics=avg_metrics, step=step
            )