Log a message to the console and/or file based on the current configuration. Args: message (str): The message to log. style

Clean the input text by replacing special tokens. Args: text: The input text to be cleaned. Returns:

Get the list of layers from a model. This function handles different model architectures to retrieve their layers. Args: model:

Create a custom progress bar for iterating over items. Args: iterable: The iterable to process. description: A string descri

Check if a single condition is met. Args: condition: The condition to check. Returns: Boolean indic

Initialize log files for all configured classes.

Compute the cosine similarity between two tensors. Args: x: First tensor. y: Second tensor. Returns

Generate a response to a given prompt using the underlying language model. Args: prompt: The input prompt to generate a

Apply explained variance scaling to a steering vector. This method scales the steering vector based on its explained variance,

Apply multiple behavior vectors to the hidden states based on rules. Args: hidden_states: The hidden states to modify.

Apply out-of-input (OOI) preventive normalization to hidden states. Args: hidden_states: The hidden states to normalize.

Apply a single behavior vector to the hidden states. Args: hidden_states: The hidden states to modify.

Evaluate a steering rule. Args: rule: The rule to evaluate. Returns: Boolean indicating whether the

Process multiple conditions for multi-steering. Args: hidden_state: The hidden state to process.

Process a single condition for steering. Args: hidden_state: The hidden state to process.

Save the analysis results from find_best_condition_point to a file. Args: analysis_results: Dictionary containing the an

Retrieve the hidden states from the specified layers of the language model for the given input strings. Args: model: The model to ge

Return a default instance of LayerControlParams. Returns: A LayerControlParams instance with default values.

Generate output using the underlying model. This method is a pass-through to the generate method of the wrapped model. It al

Get the log file path for a specific class. Args: class_name (str): The class name to get the file path for. Re

Check if verbose logging is enabled for a specific class. Args: class_name (str): The class name to check. Defaults to "

Project a matrix H onto a direction vector. Args: H: The matrix to project. direction: The direction vector to project onto.

Extract representations from the language model based on the contrast dataset. Args: model: The model to extract representations fro

Reset the class-level attributes of LeashLayer.

Reset this instance of LeashLayer to its default state.

Reset the model's steering configuration to its default state. This method removes all applied steering configurations, including

Return a list of default suffixes used in the CAIS representation engineering paper. Returns: A list of string suffixes.

Save PCA analysis figures for each hidden layer and create a macroscopic x-axis layer analysis plot. Args: layer_hiddens: A dictiona

Check if logging to a file is enabled for a specific class. Args: class_name (str): The class name to check. Re

Apply (conditional) activation steering to the model. This method configures the model to apply behavior modifications based on

Make the MalleableModel instance callable. This method allows the MalleableModel to be used like a function, delegating to the under

Initialize a LogConfig instance. Args: enabled (bool): Initial enabled state for logging.

Initialize the SteeringDataset. Args: tokenizer: The tokenizer used to tokenize and format the examples. exa

Initialize a LeashLayer. Args: layer: The underlying layer to be wrapped. layer_id: The ID of this layer in

Initialize a MalleableModel instance. This constructor wraps a pre-trained language model and its associated tokenizer, prep

Get the configuration of the underlying model. This property provides access to the configuration object of the wrapped pre-

Get the device on which the underlying model is located. This property returns the device (CPU or GPU) where the model tensors

Find the optimal condition point for steering. Args: positive_strings: List of strings that should trigger the condition

Perform a forward pass through this layer, applying steering if configured. Args: hidden_states: The input hidden states

Perform a forward pass through the model. This method delegates to the underlying model's forward method. Args:

Load a SteeringVector from a file. Args: file_path: The path to load the file from. If it doesn't end with '.svec',

Configure multi-steering for this layer. Args: behavior_vectors: List of behavior vectors to apply. conditio

Apply multiple conditional steering rules to the model. This method configures the model to apply multiple behavior modifications

Save the SteeringVector to a file. Args: file_path: The path to save the file to. If it doesn't end with '.svec',

Set whether to output logs to a file for a specific class or globally. Args: enabled (bool): Whether to enable file outp

Set the verbose state for a specific class or globally. Args: verbose (bool): Whether to enable verbose logging.

Configure steering for this layer. Args: behavior_vector: The behavior vector to apply. condition_projector:

Train a SteeringVector for a given model and tokenizer using the provided dataset. Args: model: The model to train the s

Remove steering modifications and return the original model. This method removes the LeashLayer wrappers applied to the model during