Steps — Train

Train step: computes directions from recorded activations.

SteeringResult

Output of SteeringVector step.

Attributes:

• direction_per_layer: {Node: tensor [d_model]} normalized direction,
• keyed by component address.
• separation_scores: {Node: float} how well the direction separates
• classes.
• best_layer: :class:Node with the highest separation score.

__init__

def __init__(self, direction_per_layer: dict[Node, Tensor], separation_scores: dict[Node, float], best_layer: Node) -> None:

SteeringVector

Find a steering direction via contrastive mean difference.

Reads from results:

• results['record']: ActivationStore (must have .positive and .negative)

Writes to results:

• results['steering']: SteeringResult

Args:

• method: Estimation method. Currently only "contrastive_mean_diff"
• is supported.
• normalize: If True, normalize each stored per-layer direction to unit
• norm. Note the intervention functions (steer_direction /
• ablate_direction) unit-normalize again at apply time, so
• normalize=False only affects the magnitude of the stored vector
• (e.g. if you read direction_per_layer yourself); it does not
• change steering or ablation strength.

Raises:

• ValueError: If method is not a supported value.

__init__

def __init__(self, method: str = 'contrastive_mean_diff', normalize: bool = True):