visualize¶
- dialz.visualize.highlight_token(score, vmin=-1.0, vmax=1.0)[source]¶
Generate ANSI escape codes for 24‑bit background highlighting (red→white→turquoise) with black text. Uses the same color interpolation logic as html_value_to_color.
- Parameters:
score (
float
) – Activation score.vmin (
float
) – Minimum score (mapped to red).vmax (
float
) – Maximum score (mapped to turquoise).
- Return type:
str
- Returns:
ANSI escape sequence string.
- dialz.visualize.html_value_to_color(val, vmin, vmax)[source]¶
Map a scalar score to a CSS RGB color string on a red→white→turquoise gradient.
- Parameters:
val (
float
) – Activation score.vmin (
float
) – Minimum score (mapped to red).vmax (
float
) – Maximum score (mapped to turquoise).
- Return type:
str
- Returns:
CSS rgb(r,g,b) string.
- dialz.visualize.visualize_activation(input_text, model, control_vector, layer_index=None, mode='ansi', show_score=False)[source]¶
Highlight token activations by projecting hidden states onto a steering vector.
Supports negative indices (e.g., -1 for final layer) and averaging over multiple layers. Outputs ANSI-colored text for terminal or HTML snippet for Jupyter notebooks.
- Parameters:
input_text (
str
) – Original input string.model (SteeringModel) – SteeringModel instance wrapping an LLM.
control_vector (SteeringVector) – SteeringVector defining the intervention direction.
layer_index (
Union
[int
,List
[int
]]) – Single index or list of indices; negative values count from end.mode (
str
) – “ansi” for console output, “html” for notebook display.show_score (
bool
) – If True, append numeric score to each token.
- Return type:
Union
[str
,HTML
]- Returns:
ANSI string or IPython.display.HTML.
- Raises:
ValueError – If no controlled layers set on the model.
RuntimeError – If hidden states are not captured.