should_drop#

sdm.prior.should_drop(x: Tensor, y: Tensor, *, task_type: Task, drop_trivial: bool = False, drop_unlearnable: bool = False, trivial_threshold: float = 0.05, num_trees: int = 25, num_resamples: int = 200, max_rows: int | None = 10000) → bool#

Decide whether to drop a dataset based on an ExtraTrees probe.

Compare out-of-bag squared errors with a mean baseline, one-hot encoding classification targets. A bootstrap test measures whether the probe improves on the baseline.

Parameters:
  • x (Tensor) – Features with shape [R, C] for rows and columns.

  • y (Tensor) – Targets with shape [R]; classification labels are nonnegative integer codes.

  • task_type (Task) – Classification or regression task.

  • drop_trivial (bool) – Discard datasets solved almost perfectly by the probe.

  • drop_unlearnable (bool) – Discard datasets where the probe does not significantly improve on the mean baseline.

  • trivial_threshold (float) – Maximum probe-to-baseline RMSE ratio for a trivial dataset.

  • num_trees (int) – Number of trees in the probe.

  • num_resamples (int) – Number of bootstrap resamples for the improvement test.

  • max_rows (int | None) – Maximum rows per tree bootstrap and deterministic scoring subset. None uses all rows.

Returns:

Whether the dataset should be dropped.

Return type:

bool