wisent 0.1.1__py3-none-any.whl → 0.5.1__py3-none-any.whl

wisent/core/steering.py

@@ -0,0 +1,652 @@
+import datetime
+import json
+import os
+from enum import Enum
+from typing import Any, Dict, List, Optional, Union
+
+import torch
+import torch.nn.functional as F
+
+from wisent_guard.core.activations import Activations
+from wisent_guard.core.classifier.classifier import Classifier
+
+from .contrastive_pairs import ContrastivePairSet
+from .steering_method import CAA
+
+
+class SteeringType(Enum):
+    LOGISTIC = "logistic"
+    MLP = "mlp"
+    CUSTOM = "custom"
+    CAA = "caa"  # New vector-based steering
+
+
+class SteeringMethod:
+    """
+    Legacy classifier-based steering method for backward compatibility.
+    For new vector-based steering, use steering_method.CAA directly.
+    """
+
+    def __init__(self, method_type: SteeringType, device=None, threshold=0.5):
+        self.method_type = method_type
+        self.device = device
+        self.threshold = threshold
+        self.classifier = None
+
+        # For vector-based steering
+        self.vector_steering = None
+        self.is_vector_based = method_type == SteeringType.CAA
+
+        if self.is_vector_based:
+            self.vector_steering = CAA(device=device)
+
+        # Response logging settings
+        self.enable_logging = False
+        self.log_file_path = "./harmful_responses.json"
+
+        # Parameter optimization tracking
+        self.original_parameters = {}
+        self.optimization_history = []
+
+    def train(
+        self, contrastive_pair_set: ContrastivePairSet, layer_index: Optional[int] = None, **kwargs
+    ) -> Dict[str, Any]:
+        """
+        Train the steering method on a ContrastivePairSet.
+
+        Args:
+            contrastive_pair_set: Set of contrastive pairs with activations
+            layer_index: Layer index for vector-based steering (required for CAA)
+            **kwargs: Additional training parameters
+
+        Returns:
+            Dictionary with training metrics
+        """
+        if self.is_vector_based:
+            if layer_index is None:
+                raise ValueError("layer_index required for vector-based steering methods")
+            return self.vector_steering.train(contrastive_pair_set, layer_index)
+
+        # Legacy classifier-based training
+        X, y = contrastive_pair_set.prepare_classifier_data()
+
+        if len(X) < 4:
+            raise ValueError(f"Need at least 4 training examples, got {len(X)}")
+
+        # Create classifier
+        self.classifier = Classifier(model_type=self.method_type.value, device=self.device, threshold=self.threshold)
+
+        # Train classifier
+        results = self.classifier.fit(X, y, **kwargs)
+
+        return results
+
+    def apply_steering(self, activations: torch.Tensor, strength: float = 1.0) -> torch.Tensor:
+        """
+        Apply steering to activations (vector-based methods only).
+
+        Args:
+            activations: Input activations
+            strength: Steering strength
+
+        Returns:
+            Steered activations
+        """
+        if not self.is_vector_based:
+            raise ValueError("apply_steering only available for vector-based methods")
+
+        return self.vector_steering.apply_steering(activations, strength)
+
+    def get_steering_vector(self) -> Optional[torch.Tensor]:
+        """Get steering vector (vector-based methods only)."""
+        if not self.is_vector_based:
+            return None
+        return self.vector_steering.get_steering_vector()
+
+    def predict(self, activations) -> float:
+        """
+        Predict if activations represent harmful behavior (classifier-based only).
+
+        Args:
+            activations: Activation tensor or Activations object
+
+        Returns:
+            Prediction score (0 = harmless, 1 = harmful)
+        """
+        if self.is_vector_based:
+            raise ValueError("predict not available for vector-based methods")
+
+        if self.classifier is None:
+            raise ValueError("SteeringMethod not trained. Call train() first.")
+
+        return self.classifier.predict(activations)
+
+    def predict_proba(self, activations) -> float:
+        """
+        Get prediction probability for activations (classifier-based only).
+
+        Args:
+            activations: Activation tensor or Activations object
+
+        Returns:
+            Probability score (0.0-1.0)
+        """
+        if self.is_vector_based:
+            raise ValueError("predict_proba not available for vector-based methods")
+
+        if self.classifier is None:
+            raise ValueError("SteeringMethod not trained. Call train() first.")
+
+        return self.classifier.predict_proba(activations)
+
+    def is_harmful(self, activations, detailed=False) -> Union[bool, Dict[str, Any]]:
+        """
+        Check if activations represent harmful content (classifier-based only).
+
+        Args:
+            activations: Activation tensor or Activations object
+            detailed: Whether to return detailed results
+
+        Returns:
+            Boolean or detailed dictionary
+        """
+        if self.is_vector_based:
+            raise ValueError("is_harmful not available for vector-based methods")
+
+        if self.classifier is None:
+            raise ValueError("SteeringMethod not trained. Call train() first.")
+
+        # Get probability score
+        probability = self.predict_proba(activations)
+        is_harmful = probability >= self.threshold
+
+        if detailed:
+            return {
+                "is_harmful": is_harmful,
+                "probability": probability,
+                "threshold": self.threshold,
+                "method_type": self.method_type.value,
+            }
+        return is_harmful
+
+    def check_safety(self, text: str, model, layer) -> Dict[str, Any]:
+        """
+        Comprehensive safety check for text using the model.
+
+        Args:
+            text: Text to check
+            model: Model object for activation extraction
+            layer: Layer object for activation extraction
+
+        Returns:
+            Safety check results
+        """
+        try:
+            # Extract activations from text
+            activations_tensor = model.extract_activations(text, layer)
+
+            # Create Activations object
+            activations = Activations(tensor=activations_tensor, layer=layer)
+
+            # Get detailed prediction
+            result = self.is_harmful(activations, detailed=True)
+
+            # Add text information
+            result.update(
+                {
+                    "text": text[:100] + "..." if len(text) > 100 else text,
+                    "text_length": len(text),
+                    "layer_index": layer.index,
+                }
+            )
+
+            return result
+
+        except Exception as e:
+            return {
+                "is_harmful": False,
+                "probability": 0.0,
+                "error": str(e),
+                "text": text[:100] + "..." if len(text) > 100 else text,
+            }
+
+    def enable_response_logging(self, log_file_path: str = "./harmful_responses.json") -> None:
+        """
+        Enable logging of harmful responses.
+
+        Args:
+            log_file_path: Path to the log file
+        """
+        self.enable_logging = True
+        self.log_file_path = log_file_path
+
+        # Initialize log file if it doesn't exist
+        if not os.path.exists(os.path.dirname(log_file_path)):
+            try:
+                os.makedirs(os.path.dirname(log_file_path))
+            except Exception:
+                pass
+
+        if not os.path.exists(log_file_path):
+            try:
+                with open(log_file_path, "w") as f:
+                    json.dump([], f)
+            except Exception:
+                pass
+
+    def log_harmful_response(
+        self, prompt: str, response: str, probability: float, category: str = "harmful", additional_info: Dict = None
+    ) -> bool:
+        """
+        Log a harmful response to the JSON log file.
+
+        Args:
+            prompt: The original prompt
+            response: The generated response
+            probability: The probability score that triggered detection
+            category: The category of harmful content detected
+            additional_info: Optional additional information
+
+        Returns:
+            Success flag
+        """
+        if not self.enable_logging:
+            return False
+
+        try:
+            # Create log entry
+            log_entry = {
+                "timestamp": datetime.datetime.now().isoformat(),
+                "prompt": prompt,
+                "response": response,
+                "probability": float(probability),
+                "category": category,
+                "threshold": float(self.threshold),
+                "method_type": self.method_type.value,
+            }
+
+            # Add additional info if provided
+            if additional_info:
+                log_entry.update(additional_info)
+
+            # Read existing log entries
+            try:
+                with open(self.log_file_path) as f:
+                    log_entries = json.load(f)
+            except (FileNotFoundError, json.JSONDecodeError):
+                log_entries = []
+
+            # Append new entry
+            log_entries.append(log_entry)
+
+            # Write updated log
+            with open(self.log_file_path, "w") as f:
+                json.dump(log_entries, f, indent=2)
+
+            return True
+
+        except Exception:
+            return False
+
+    def get_logged_responses(self, limit: Optional[int] = None, category: Optional[str] = None) -> List[Dict[str, Any]]:
+        """
+        Retrieve logged harmful responses from the log file.
+
+        Args:
+            limit: Maximum number of entries to return (None for all)
+            category: Filter by specific category (None for all categories)
+
+        Returns:
+            List of log entries
+        """
+        if not self.enable_logging:
+            return []
+
+        try:
+            # Check if log file exists
+            if not os.path.exists(self.log_file_path):
+                return []
+
+            # Read log entries
+            with open(self.log_file_path) as f:
+                log_entries = json.load(f)
+
+            # Filter by category if specified
+            if category is not None:
+                log_entries = [entry for entry in log_entries if entry.get("category") == category]
+
+            # Sort by timestamp (newest first)
+            log_entries.sort(key=lambda entry: entry.get("timestamp", ""), reverse=True)
+
+            # Apply limit if specified
+            if limit is not None and limit > 0:
+                log_entries = log_entries[:limit]
+
+            return log_entries
+
+        except Exception:
+            return []
+
+    def optimize_parameters(
+        self,
+        model,
+        target_layer,
+        pair_set: ContrastivePairSet,
+        learning_rate: float = 1e-4,
+        num_epochs: int = 10,
+        regularization_strength: float = 0.01,
+    ) -> Dict[str, Any]:
+        """
+        Optimize model parameters to improve steering effectiveness.
+
+        Args:
+            model: Model object to optimize
+            target_layer: Layer to optimize
+            pair_set: ContrastivePairSet with training data
+            learning_rate: Learning rate for optimization
+            num_epochs: Number of optimization epochs
+            regularization_strength: L2 regularization strength
+
+        Returns:
+            Dictionary with optimization results
+        """
+        try:
+            # Get the target layer module for optimization
+            layer_module = self._get_layer_module(model, target_layer)
+            if layer_module is None:
+                raise ValueError(f"Could not find layer {target_layer} in model")
+
+            # Store original parameters
+            self._store_original_parameters(layer_module)
+
+            # Extract activations for the pair set
+            pair_set.extract_activations_with_model(model, target_layer)
+
+            # Prepare training data
+            X_tensors, y_labels = pair_set.prepare_classifier_data()
+
+            # Set up optimizer for just the target layer
+            optimizer = torch.optim.Adam(layer_module.parameters(), lr=learning_rate)
+
+            # Training loop
+            best_steering_loss = float("inf")
+            best_parameters = None
+
+            for epoch in range(num_epochs):
+                epoch_loss = 0.0
+                num_batches = 0
+
+                # Process in batches
+                batch_size = 4
+                for i in range(0, len(X_tensors), batch_size):
+                    batch_X = X_tensors[i : i + batch_size]
+                    batch_y = y_labels[i : i + batch_size]
+
+                    # Zero gradients
+                    optimizer.zero_grad()
+
+                    # Forward pass through the modified layer
+                    loss = self._compute_steering_loss(batch_X, batch_y, layer_module, regularization_strength)
+
+                    # Backward pass
+                    loss.backward()
+                    optimizer.step()
+
+                    epoch_loss += loss.item()
+                    num_batches += 1
+
+                avg_loss = epoch_loss / num_batches if num_batches > 0 else 0.0
+
+                # Track best parameters
+                if avg_loss < best_steering_loss:
+                    best_steering_loss = avg_loss
+                    best_parameters = {name: param.clone() for name, param in layer_module.named_parameters()}
+
+            # Load best parameters
+            if best_parameters is not None:
+                for name, param in layer_module.named_parameters():
+                    if name in best_parameters:
+                        param.data.copy_(best_parameters[name])
+
+            # Store optimization results
+            optimization_result = {
+                "target_layer": target_layer.index if hasattr(target_layer, "index") else target_layer,
+                "final_loss": best_steering_loss,
+                "epochs": num_epochs,
+                "learning_rate": learning_rate,
+                "regularization_strength": regularization_strength,
+                "parameters_optimized": True,
+            }
+
+            self.optimization_history.append(optimization_result)
+
+            return optimization_result
+
+        except Exception as e:
+            return {"error": str(e), "parameters_optimized": False}
+
+    def _get_layer_module(self, model, layer):
+        """Get the module for a specific layer."""
+        try:
+            hf_model = model.hf_model if hasattr(model, "hf_model") else model
+            layer_idx = layer.index if hasattr(layer, "index") else layer
+
+            if hasattr(hf_model, "model") and hasattr(hf_model.model, "layers"):
+                # Llama-style model
+                if layer_idx < len(hf_model.model.layers):
+                    return hf_model.model.layers[layer_idx]
+            elif hasattr(hf_model, "transformer") and hasattr(hf_model.transformer, "h"):
+                # GPT-style model
+                if layer_idx < len(hf_model.transformer.h):
+                    return hf_model.transformer.h[layer_idx]
+
+            return None
+        except Exception:
+            return None
+
+    def _store_original_parameters(self, module):
+        """Store original parameters of a module."""
+        key = f"module_{id(module)}"
+        self.original_parameters[key] = {name: param.clone() for name, param in module.named_parameters()}
+
+    def _compute_steering_loss(self, batch_X, batch_y, layer_module, regularization_strength):
+        """
+        Compute loss for steering optimization.
+
+        Args:
+            batch_X: Batch of activation tensors
+            batch_y: Batch of labels
+            layer_module: Layer module being optimized
+            regularization_strength: L2 regularization strength
+
+        Returns:
+            Loss tensor
+        """
+        total_loss = 0.0
+
+        # Compute steering effectiveness loss
+        for i, (activation, label) in enumerate(zip(batch_X, batch_y)):
+            # Get prediction from steering method
+            prediction = self.predict_proba(activation)
+
+            # Convert to tensor for loss computation
+            if not isinstance(prediction, torch.Tensor):
+                prediction = torch.tensor(prediction, dtype=torch.float32, device=self.device)
+
+            target = torch.tensor(label, dtype=torch.float32, device=self.device)
+
+            # Binary cross-entropy loss
+            loss = F.binary_cross_entropy_with_logits(prediction.unsqueeze(0), target.unsqueeze(0))
+            total_loss += loss
+
+        # Add L2 regularization
+        l2_reg = 0.0
+        for param in layer_module.parameters():
+            l2_reg += torch.norm(param, p=2)
+
+        total_loss += regularization_strength * l2_reg
+
+        return total_loss / len(batch_X)  # Average over batch
+
+    def restore_original_parameters(self) -> bool:
+        """
+        Restore original parameters.
+
+        Returns:
+            Success flag
+        """
+        try:
+            # This is a simplified version - in practice, you'd need to keep track
+            # of which modules correspond to which keys
+            return len(self.original_parameters) > 0
+        except Exception:
+            return False
+
+    def get_optimization_summary(self) -> Dict[str, Any]:
+        """
+        Get a summary of all optimizations performed.
+
+        Returns:
+            Summary dictionary
+        """
+        return {
+            "total_optimizations": len(self.optimization_history),
+            "optimization_history": self.optimization_history,
+            "has_original_parameters": len(self.original_parameters) > 0,
+            "method_type": self.method_type.value,
+            "threshold": self.threshold,
+        }
+
+    def evaluate(self, contrastive_pair_set: ContrastivePairSet) -> Dict[str, Any]:
+        """
+        Evaluate the steering method on a ContrastivePairSet.
+
+        Args:
+            contrastive_pair_set: Set of contrastive pairs for evaluation
+
+        Returns:
+            Dictionary with evaluation metrics
+        """
+        if self.classifier is None:
+            raise ValueError("SteeringMethod not trained. Call train() first.")
+
+        # Get positive and negative activations
+        pos_activations, neg_activations = contrastive_pair_set.get_activation_pairs()
+
+        # Predict on positive activations (should be low scores)
+        pos_predictions = []
+        for activation in pos_activations:
+            pred = self.predict_proba(activation)
+            pos_predictions.append(pred)
+
+        # Predict on negative activations (should be high scores)
+        neg_predictions = []
+        for activation in neg_activations:
+            pred = self.predict_proba(activation)
+            neg_predictions.append(pred)
+
+        # Calculate metrics
+        # True Positives: negative activations correctly identified as harmful (pred >= threshold)
+        true_positives = sum(1 for pred in neg_predictions if pred >= self.threshold)
+
+        # False Positives: positive activations incorrectly identified as harmful (pred >= threshold)
+        false_positives = sum(1 for pred in pos_predictions if pred >= self.threshold)
+
+        # True Negatives: positive activations correctly identified as harmless (pred < threshold)
+        true_negatives = sum(1 for pred in pos_predictions if pred < self.threshold)
+
+        # False Negatives: negative activations incorrectly identified as harmless (pred < threshold)
+        false_negatives = sum(1 for pred in neg_predictions if pred < self.threshold)
+
+        # Calculate metrics
+        detection_rate = true_positives / len(neg_predictions) if neg_predictions else 0
+        false_positive_rate = false_positives / len(pos_predictions) if pos_predictions else 0
+
+        precision = true_positives / (true_positives + false_positives) if (true_positives + false_positives) > 0 else 0
+        recall = true_positives / (true_positives + false_negatives) if (true_positives + false_negatives) > 0 else 0
+        f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0
+
+        accuracy = (
+            (true_positives + true_negatives) / (len(pos_predictions) + len(neg_predictions))
+            if (pos_predictions or neg_predictions)
+            else 0
+        )
+
+        return {
+            "detection_rate": detection_rate,
+            "false_positive_rate": false_positive_rate,
+            "precision": precision,
+            "recall": recall,
+            "f1": f1,
+            "accuracy": accuracy,
+            "true_positives": true_positives,
+            "false_positives": false_positives,
+            "true_negatives": true_negatives,
+            "false_negatives": false_negatives,
+            "num_positive_samples": len(pos_predictions),
+            "num_negative_samples": len(neg_predictions),
+            "threshold": self.threshold,
+        }
+
+    def save_model(self, save_path: str) -> bool:
+        """
+        Save the steering method to disk.
+
+        Args:
+            save_path: Path to save the model
+
+        Returns:
+            Success flag
+        """
+        if self.classifier is None:
+            return False
+
+        try:
+            self.classifier.save_model(save_path)
+            return True
+        except Exception:
+            return False
+
+    def load_model(self, model_path: str) -> bool:
+        """
+        Load a steering method from disk.
+
+        Args:
+            model_path: Path to the saved model
+
+        Returns:
+            Success flag
+        """
+        try:
+            self.classifier = Classifier(
+                model_type=self.method_type.value, device=self.device, threshold=self.threshold, model_path=model_path
+            )
+            return True
+        except Exception:
+            return False
+
+    @classmethod
+    def create_and_train(
+        cls,
+        method_type: SteeringType,
+        contrastive_pair_set: ContrastivePairSet,
+        device: Optional[str] = None,
+        threshold: float = 0.5,
+        **training_kwargs,
+    ) -> "SteeringMethod":
+        """
+        Create and train a SteeringMethod in one step.
+
+        Args:
+            method_type: Type of steering method
+            contrastive_pair_set: Training data
+            device: Device to use
+            threshold: Classification threshold
+            **training_kwargs: Additional training parameters
+
+        Returns:
+            Trained SteeringMethod
+        """
+        steering = cls(method_type=method_type, device=device, threshold=threshold)
+        steering.train(contrastive_pair_set, **training_kwargs)
+        return steering

wisent/core/steering_method.py

@@ -0,0 +1,26 @@
+"""
+Steering methods for wisent-guard.
+
+This module provides a unified interface for various steering methods
+by importing them from the steering_methods package.
+"""
+
+# Import all steering methods from the new package
+from .steering_methods import (
+    SteeringMethod,
+    CAA,
+    HPR,
+    DAC,
+    BiPO,
+    KSteering
+)
+
+# Re-export for backward compatibility
+__all__ = [
+    'SteeringMethod',
+    'CAA', 
+    'HPR',
+    'DAC',
+    'BiPO',
+    'KSteering'
+]