Probing

Plotting utilities for probing results.

Requires the plot extra (plotly); the confusion matrix also needs the probe extra (scikit-learn).

plot_probe_accuracy

def plot_probe_accuracy(probe: ProbeResult) -> go.Figure:

Plot per-layer probe accuracy with cross-validation error bars.

Highlights the best-scoring layer in a distinct color.

Args:

• probe: ProbeResult containing per-layer accuracy and CV fold scores.

Returns:

• An interactive Plotly bar chart, one bar per layer.

plot_confusion_matrix

def plot_confusion_matrix(probe: ProbeResult, store: LabeledActivationStore) -> go.Figure | None:

Plot a confusion matrix at the best layer using the refitted classifier.

Args:

• probe: ProbeResult with refitted classifiers.
• store: LabeledActivationStore containing activations and labels.

Returns:

• An interactive Plotly heatmap, or None when the best layer has no
• refitted classifier (e.g. the Probe step ran without refit=True).