Logit Attribution

Plotly visualization for LogitAttribution results.

Renders the top contributors to the target logit (difference) as a horizontal bar chart, signed so positive bars push toward the answer and negative bars push away. The embedding and the catch-all other term are included alongside the named attention-head and MLP components.

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

plot_logit_attribution

def plot_logit_attribution(result: LogitAttributionResult, top_k: int = 20, title: str = 'Direct Logit Attribution') -> go.Figure:

Render the top-contributing components as a signed horizontal bar chart.

Args:

• result: LogitAttributionResult produced by the LogitAttribution step.
• top_k: Number of components to show, ranked by absolute contribution.
• title: Plot title.

Returns:

• A plotly.graph_objects.Figure with one horizontal bar per component,
• most influential at the top, colored by sign.