Skip to content
Murano

Activations

Plotly visualization for recorded activations.

Projects a single component’s high-dimensional activations down to two dimensions and scatters them by class, so the linear structure a probe or a steering vector exploits becomes visible.

The dimensionality reducer is supplied by the caller rather than imported here, which keeps scikit-learn out of this module’s dependencies and leaves the choice of method (PCA, LDA, t-SNE, UMAP, …) open.

Requires the plot extra (install with pip install murano-interp[plot]).

Reducer

Section titled “Reducer”

Any scikit-learn-style dimensionality reducer.

Supervised reducers such as LDA consume y; unsupervised ones such as PCA accept and ignore it, so a single call site serves both.

fit_transform

Section titled “fit_transform”
def fit_transform(self, X: Any, y: Any = None) -> Any:

plot_activation_projection

Section titled “plot_activation_projection”
def plot_activation_projection(store: ActivationStore | LabeledActivationStore, layer: AddressLike, reducer: Reducer, normalize: bool = True, label_names: list[str] | None = None, title: str = 'Activation projection') -> go.Figure:

Scatter one component’s activations in a two-dimensional projection.

Accepts either activation store: a contrastive :class:~murano.steps.record.ActivationStore is colored positive against negative, while a :class:~murano.steps.record.LabeledActivationStore is colored by its integer labels. That lets a single Record feed both this plot and the :class:~murano.steps.probe.Probe step.

A reducer yielding one component is scattered along x with the classes separated vertically; two or more components use the first two.

Args:

  • store: Activation store recorded at a reduced token position.
  • layer: Component address to project.
  • reducer: Fitted-on-call reducer, e.g. PCA(n_components=2) or
  • LinearDiscriminantAnalysis(n_components=1). Class labels are
  • passed to fit_transform so supervised reducers work unchanged.
  • normalize: Scale each activation vector to unit L2 norm before
  • reducing, so the projection reflects direction rather than
  • magnitude. Zero vectors are left untouched.
  • label_names: Names for the integer labels of a labeled store, indexed
  • by label value. Ignored for a contrastive store.
  • title: Plot title.

Returns:

  • A plotly.graph_objects.Figure with one scatter trace per class.
  • Classes are colored by label value (or positive-then-negative), not by
  • the order they appear in the data, so the same class keeps its color
  • across runs on reordered data.

Raises:

  • ValueError: If store was recorded with position="none" or
  • per_head=True, which carry extra dimensions this projection
  • cannot interpret, or if label_names does not cover every label
  • in the store.