Steps — Probe

Probe step: trains linear classifiers on recorded activations.

ProbeResult

Output of the Probe step.

Attributes:

• accuracy_per_layer: {Node: float} mean CV accuracy.
• cv_scores: {Node: ndarray} per-fold accuracy scores.
• best_layer: :class:Node with highest mean accuracy.
• classifiers: {Node: fitted sklearn classifier} (only if refit=True).
• label_names: Human-readable label names (passed through from dataset).

__init__

def __init__(self, accuracy_per_layer: dict[Node, float], cv_scores: dict[Node, ndarray], best_layer: Node, classifiers: dict[Node, Any] = dict(), label_names: list[str] | None = None) -> None:

Probe

Train a linear probe per layer via cross-validation.

Reads from results:

• results['record']: LabeledActivationStore
• results['dataset']: LabeledDataset (optional, for label_names)

Writes to results:

• results['probe']: ProbeResult

Args:

• classifier: sklearn classifier instance (default: LogisticRegression).
• Will be cloned per layer.
• cv: Number of cross-validation folds.
• refit: If True, fit a final classifier on all data per layer and store
• in ProbeResult.classifiers.

__init__

def __init__(self, classifier: Any | None = None, cv: int = 5, refit: bool = False):