wisent 0.5.6__py3-none-any.whl → 0.5.7__py3-none-any.whl

wisent/__init__.py

@@ -1 +1 @@
-__version__ = "0.5.6"
+__version__ = "0.5.7"

wisent/core/multi_steering.py

@@ -1,51 +1,18 @@
 """Multi-steering functionality for combining multiple steering vectors."""
 
-import sys
+from __future__ import annotations
+
 import torch
-from typing import List, Tuple, Optional, Dict, Any
+from typing import Iterable
 from pathlib import Path
 
-# Note: layer and model modules don't exist, using inline stubs
-# from .layer import Layer
-# from .model import Model
-from .steering_methods.methods.caa import CAAMethod as CAA
-# DAC doesn't exist yet
-# from .steering_methods.dac import DAC
-from .utils.device import resolve_default_device
-
-
-# Stub classes for missing modules
-class Layer:
-    """Stub Layer class - module doesn't exist in wisent package."""
-    def __init__(self, index: int):
-        self.index = index
-
-
-class Model:
-    """Stub Model class - module doesn't exist in wisent package."""
-    def __init__(self, model_name: str, device: str = None):
-        self.model_name = model_name
-        self.device = device
-        # This is a stub - actual implementation would load the model
-        raise NotImplementedError("Model class not implemented in wisent package")
-
-
-class DAC:
-    """Stub DAC class - module doesn't exist in wisent package."""
-    def __init__(self, device: str = None):
-        self.device = device
-        self.steering_vector = None
-        self.layer_index = None
-        self.is_trained = False
+from wisent.core.models.wisent_model import WisentModel
+from wisent.core.models.core.atoms import SteeringPlan
+from wisent.core.activations.core.atoms import RawActivationMap
+from wisent.core.prompts.core.atom import ChatMessage
+from wisent.core.utils.device import resolve_default_device
 
-    @staticmethod
-    def combine_steering_vectors(vectors, weights, normalize_weights=True):
-        """Stub method for combining steering vectors."""
-        raise NotImplementedError("DAC steering method not implemented in wisent package")
-
-    def apply_steering(self, hidden_states, strength=1.0):
-        """Stub method for applying steering."""
-        raise NotImplementedError("DAC steering method not implemented in wisent package")
+__all__ = ["MultiSteering", "MultiSteeringError"]
 
 
 class MultiSteeringError(Exception):
@@ -54,247 +21,271 @@ class MultiSteeringError(Exception):
 
 
 class MultiSteering:
-    """Handles multi-steering vector combination and application."""
-    
+    """Handles multi-steering vector combination and application using WisentModel."""
+
     def __init__(self, device: str | None = None, method: str = "CAA"):
         """Initialize multi-steering handler.
-        
+
         Args:
             device: Device to use for computations (cpu/cuda/mps)
-            method: Steering method to use for combination ("CAA" or "DAC")
+            method: Steering method to use (kept for backward compatibility, not used in new API)
         """
         self.device = device or resolve_default_device()
-        self.method = method
-        self.loaded_vectors = []
-        self.weights = []
-        self.combined_vector = None
-        self.layer = None
-    
-    def load_vectors(self, vector_specs: List[str]) -> None:
+        self.method = method  # Kept for backward compatibility
+        self.loaded_vectors: list[dict] = []
+        self.weights: list[float] = []
+        self.combined_vector: torch.Tensor | None = None
+        self.layer: int | None = None
+
+    def load_vectors(self, vector_specs: list[str]) -> None:
         """Load and validate steering vectors from file paths.
-        
+
         Args:
             vector_specs: List of "path:weight" specifications
-            
+
         Raises:
             MultiSteeringError: If vectors cannot be loaded or are incompatible
         """
         if not vector_specs:
             raise MultiSteeringError("No vectors specified")
-        
+
         self.loaded_vectors = []
         self.weights = []
         layers_found = set()
-        
+
         for spec in vector_specs:
             parts = spec.split(":")
             if len(parts) != 2:
                 raise MultiSteeringError(f"Invalid vector specification: {spec}. Expected format: path:weight")
-            
+
             vector_path = parts[0]
             try:
                 weight = float(parts[1])
             except ValueError:
                 raise MultiSteeringError(f"Invalid weight in {spec}. Must be a number.")
-            
+
             if not Path(vector_path).exists():
                 raise MultiSteeringError(f"Vector file not found: {vector_path}")
-            
+
             print(f"Loading vector from {vector_path} with weight {weight}")
-            
+
             try:
-                vector_data = torch.load(vector_path, map_location=self.device)
+                vector_data = torch.load(vector_path, map_location=self.device, weights_only=False)
             except Exception as e:
                 raise MultiSteeringError(f"Failed to load vector from {vector_path}: {e}")
-            
+
             # Extract metadata from loaded vector
             if isinstance(vector_data, dict):
                 layer = vector_data.get("layer_index", None)
                 steering_vector = vector_data.get("steering_vector", None)
-                
+
                 if steering_vector is None:
                     raise MultiSteeringError(f"No steering vector found in {vector_path}")
-                
+
                 if layer is not None:
                     layers_found.add(layer)
-                
+
                 self.loaded_vectors.append(vector_data)
                 self.weights.append(weight)
-                
+
                 print(f"   ✓ Loaded vector from layer {layer}")
             else:
                 raise MultiSteeringError(f"Invalid vector format in {vector_path}")
-        
+
         # Validate compatibility
         if len(layers_found) > 1:
             raise MultiSteeringError(f"Vectors from different layers cannot be combined: {layers_found}")
-        
+
         if not layers_found:
             raise MultiSteeringError("No layer information found in vectors")
-        
-        self.layer = Layer(list(layers_found)[0])
-        
+
+        self.layer = list(layers_found)[0]
+
         print(f"\nUsing {self.method} method for vector combination")
-        print(f"Target layer: {self.layer.index}")
-    
+        print(f"Target layer: {self.layer}")
+
     def combine_vectors(self, normalize: bool = True) -> torch.Tensor:
-        """Combine loaded vectors using appropriate method.
-        
+        """Combine loaded vectors using SteeringPlan.
+
         Args:
             normalize: Whether to normalize the combined vector
-            
+
         Returns:
             Combined steering vector as tensor
-            
+
         Raises:
             MultiSteeringError: If combination fails
         """
         if not self.loaded_vectors:
             raise MultiSteeringError("No vectors loaded")
-        
-        print(f"\n🔄 Combining {len(self.loaded_vectors)} vectors using {self.method}")
-        
-        if self.method == "CAA":
-            # Create a CAA instance and use its proper combination method
-            caa = CAA(device=self.device)
-            
-            # Set up behavior vectors dictionary
-            caa.behavior_vectors = {}
-            for i, (vector_data, weight) in enumerate(zip(self.loaded_vectors, self.weights)):
-                steering_vector = vector_data["steering_vector"]
-                
-                if not isinstance(steering_vector, torch.Tensor):
-                    steering_vector = torch.tensor(steering_vector, device=self.device)
-                else:
-                    steering_vector = steering_vector.to(self.device)
-                
-                # Store with unique names
-                behavior_name = f"vector_{i}"
-                caa.behavior_vectors[behavior_name] = steering_vector
-            
-            # Create weights dictionary
-            behavior_weights = {f"vector_{i}": weight for i, weight in enumerate(self.weights)}
-            
-            # Use CAA's combine_behaviors method with normalization
-            self.combined_vector = caa.combine_behaviors(behavior_weights, normalize_result=normalize)
-            
-        else:  # DAC or mixed methods
-            # For DAC, use its combine_steering_vectors method
-            vectors = []
-            for vector_data in self.loaded_vectors:
-                steering_vector = vector_data["steering_vector"]
-                
-                if not isinstance(steering_vector, torch.Tensor):
-                    steering_vector = torch.tensor(steering_vector, device=self.device)
-                else:
-                    steering_vector = steering_vector.to(self.device)
-                
-                vectors.append(steering_vector)
-            
-            # Use DAC's static method for combination
-            self.combined_vector = DAC.combine_steering_vectors(
-                vectors, self.weights, normalize_weights=normalize
-            )
-        
+
+        if self.layer is None:
+            raise MultiSteeringError("No layer information available")
+
+        print(f"\n🔄 Combining {len(self.loaded_vectors)} vectors")
+
+        # Build RawActivationMap for each vector
+        raw_maps: list[RawActivationMap] = []
+        for vector_data in self.loaded_vectors:
+            steering_vector = vector_data["steering_vector"]
+
+            if not isinstance(steering_vector, torch.Tensor):
+                steering_vector = torch.tensor(steering_vector, device=self.device)
+            else:
+                steering_vector = steering_vector.to(self.device)
+
+            # Create a map with layer name as string (required by SteeringPlan)
+            raw_map: RawActivationMap = {str(self.layer): steering_vector}
+            raw_maps.append(raw_map)
+
+        # Create SteeringPlan with weighted combination
+        plan = SteeringPlan.from_raw(
+            raw=raw_maps,
+            weights=self.weights,
+            scale=1.0,
+            normalize=normalize
+        )
+
+        # Extract the combined vector from the plan
+        layer_str = str(self.layer)
+        if layer_str not in plan.layers:
+            raise MultiSteeringError(f"Layer {self.layer} not found in steering plan")
+
+        steering_vector = plan.layers[layer_str]
+        self.combined_vector = steering_vector.vector
+
         print(f"   ✓ Combined vector shape: {self.combined_vector.shape}")
         print(f"   ✓ Combined vector norm: {torch.norm(self.combined_vector).item():.4f}")
-        
+
         return self.combined_vector
-    
-    def apply_steering(self, model: Model, prompt: str, max_new_tokens: int = 100, 
-                      temperature: float = 0.7, top_p: float = 0.9) -> str:
-        """Apply the combined steering vector to generate text.
-        
+
+    def apply_steering_stream(
+        self,
+        model: WisentModel,
+        prompt: str,
+        max_new_tokens: int = 100,
+        temperature: float = 0.7,
+        top_p: float = 0.9
+    ) -> Iterable[str]:
+        """Apply the combined steering vector to generate text with streaming.
+
+        Args:
+            model: WisentModel instance to use for generation
+            prompt: Input prompt
+            max_new_tokens: Maximum tokens to generate
+            temperature: Sampling temperature
+            top_p: Top-p sampling parameter
+
+        Yields:
+            Generated text chunks
+
+        Raises:
+            MultiSteeringError: If steering fails
+        """
+        if self.combined_vector is None:
+            raise MultiSteeringError("No combined vector available. Call combine_vectors() first.")
+
+        if self.layer is None:
+            raise MultiSteeringError("No layer information available")
+
+        print(f"\n🎯 Applying combined steering vector at layer {self.layer}")
+        print(f"Prompt: {prompt}")
+        print("=" * 50)
+
+        # Create SteeringPlan from the combined vector
+        raw_map: RawActivationMap = {str(self.layer): self.combined_vector}
+        steering_plan = SteeringPlan.from_raw(
+            raw=raw_map,
+            scale=1.0,
+            normalize=False  # Already normalized in combine_vectors
+        )
+
+        # Format prompt as chat messages
+        messages: list[ChatMessage] = [{"role": "user", "content": prompt}]
+
+        try:
+            # Use WisentModel's generate_stream with steering
+            yield from model.generate_stream(
+                inputs=[messages],
+                max_new_tokens=max_new_tokens,
+                temperature=temperature,
+                top_p=top_p,
+                use_steering=True,
+                steering_plan=steering_plan,
+                skip_prompt=True,
+                skip_special_tokens=True
+            )
+
+        except Exception as e:
+            raise MultiSteeringError(f"Failed to apply steering: {e}")
+
+    def apply_steering(
+        self,
+        model: WisentModel,
+        prompt: str,
+        max_new_tokens: int = 100,
+        temperature: float = 0.7,
+        top_p: float = 0.9
+    ) -> str:
+        """Apply the combined steering vector to generate text (non-streaming).
+
         Args:
-            model: Model to use for generation
+            model: WisentModel instance to use for generation
             prompt: Input prompt
             max_new_tokens: Maximum tokens to generate
             temperature: Sampling temperature
             top_p: Top-p sampling parameter
-            
+
         Returns:
             Generated text
-            
+
         Raises:
             MultiSteeringError: If steering fails
         """
         if self.combined_vector is None:
             raise MultiSteeringError("No combined vector available. Call combine_vectors() first.")
-        
+
         if self.layer is None:
             raise MultiSteeringError("No layer information available")
-        
-        print(f"\n🎯 Applying combined steering vector at layer {self.layer.index}")
+
+        print(f"\n🎯 Applying combined steering vector at layer {self.layer}")
         print(f"Prompt: {prompt}")
         print("=" * 50)
-        
-        # Create appropriate steering method instance
-        if self.method == "CAA":
-            steering_method = CAA(device=self.device)
-            steering_method.steering_vector = self.combined_vector
-            steering_method.layer_index = self.layer.index
-            steering_method.is_trained = True
-        else:
-            # Use DAC for other methods
-            steering_method = DAC(device=self.device)
-            steering_method.steering_vector = self.combined_vector  
-            steering_method.layer_index = self.layer.index
-            steering_method.is_trained = True
-        
-        # Set up steering hook
-        hooks = []
-        
-        def steering_hook(module, input, output):
-            if isinstance(output, tuple):
-                hidden_states = output[0]
-            else:
-                hidden_states = output
-            
-            # Apply steering using the method's apply_steering
-            steered = steering_method.apply_steering(hidden_states, strength=1.0)
-            
-            if isinstance(output, tuple):
-                return (steered,) + output[1:]
-            return steered
-        
-        # Find the target layer module
-        if hasattr(model.hf_model, "model") and hasattr(model.hf_model.model, "layers"):
-            layer_module = model.hf_model.model.layers[self.layer.index]
-        elif hasattr(model.hf_model, "transformer") and hasattr(model.hf_model.transformer, "h"):
-            layer_module = model.hf_model.transformer.h[self.layer.index]
-        else:
-            raise MultiSteeringError("Could not find model layers")
-        
-        # Register hook
-        handle = layer_module.register_forward_hook(steering_hook)
-        hooks.append(handle)
-        
+
+        # Create SteeringPlan from the combined vector
+        raw_map: RawActivationMap = {str(self.layer): self.combined_vector}
+        steering_plan = SteeringPlan.from_raw(
+            raw=raw_map,
+            scale=1.0,
+            normalize=False  # Already normalized in combine_vectors
+        )
+
+        # Format prompt as chat messages
+        messages: list[ChatMessage] = [{"role": "user", "content": prompt}]
+
         try:
-            # Generate with steering
-            output, _ = model.generate(
-                prompt=prompt,
-                layer_index=self.layer.index,
+            # Use WisentModel's generate with steering
+            outputs = model.generate(
+                inputs=[messages],
                 max_new_tokens=max_new_tokens,
                 temperature=temperature,
                 top_p=top_p,
+                use_steering=True,
+                steering_plan=steering_plan
             )
-            
-            return output
-            
+
+            return outputs[0] if outputs else ""
+
         except Exception as e:
             raise MultiSteeringError(f"Failed to apply steering: {e}")
-        finally:
-            # Clean up hooks
-            for hook in hooks:
-                hook.remove()
 
 
 def run_multi_steer(
-    vector_specs: List[str],
+    vector_specs: list[str],
     model_name: str,
     prompt: str,
     method: str = "CAA",
-    layer: Optional[int] = None,
+    layer: int | None = None,
     max_new_tokens: int = 100,
     temperature: float = 0.7,
     top_p: float = 0.9,
@@ -302,45 +293,50 @@ def run_multi_steer(
     verbose: bool = True,
 ) -> str:
     """Convenience function to run multi-steering.
-    
+
     Args:
         vector_specs: List of "path:weight" specifications
         model_name: Name of model to load
         prompt: Input prompt
-        method: Steering method to use ("CAA" or "DAC")
+        method: Steering method to use (kept for backward compatibility)
         layer: Target layer (will be inferred from vectors if not specified)
         max_new_tokens: Maximum tokens to generate
-        temperature: Sampling temperature  
+        temperature: Sampling temperature
         top_p: Top-p sampling parameter
         device: Device to use
         verbose: Whether to print progress
-        
+
     Returns:
         Generated text
     """
     # Initialize model
     if verbose:
         print(f"\n🚀 Loading model: {model_name}")
-    
+
     chosen_device = device or resolve_default_device()
-    model = Model(model_name, device=chosen_device)
-    
+
+    # Load model using WisentModel
+    model = WisentModel(
+        model_name=model_name,
+        device=chosen_device
+    )
+
     # Initialize multi-steering with specified method
     multi_steer = MultiSteering(device=chosen_device, method=method)
-    
+
     # Load vectors
     multi_steer.load_vectors(vector_specs)
-    
+
     # Override layer if specified
     if layer is not None:
-        multi_steer.layer = Layer(layer)
+        multi_steer.layer = layer
         if verbose:
             print(f"Overriding layer to: {layer}")
-    
+
     # Combine vectors with normalization
     multi_steer.combine_vectors(normalize=True)
-    
-    # Apply steering
+
+    # Apply steering (non-streaming)
     output = multi_steer.apply_steering(
         model=model,
         prompt=prompt,
@@ -348,5 +344,5 @@ def run_multi_steer(
         temperature=temperature,
         top_p=top_p
     )
-    
-    return output
+
+    return output

{wisent-0.5.6.dist-info → wisent-0.5.7.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: wisent
-Version: 0.5.6
+Version: 0.5.7
 Summary: Monitor and guard against harmful content in language models
 Home-page: https://github.com/yourusername/wisent-activation-guardrails
 Author: Wisent Team

{wisent-0.5.6.dist-info → wisent-0.5.7.dist-info}/RECORD

@@ -1,4 +1,4 @@
-wisent/__init__.py,sha256=CMH34Gt1AqO7z_TqRj94XwohGoVCf8aes0djkqm45mk,22
+wisent/__init__.py,sha256=KiyyYbyEe0O858kmiWcg1RdmqGUYtk_JqRmc3_Ev2Q8,22
 wisent/benchmarks/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 wisent/benchmarks/coding/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 wisent/benchmarks/coding/metrics/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -64,7 +64,7 @@ wisent/core/managed_cached_benchmarks.py,sha256=JbvpZ1fgSuQQhyQVKEvqrQZRHGqfnjo9
 wisent/core/mixed_benchmark_sampler.py,sha256=tKQCHUXVuYeCyx4VZt8O1hGyB-TOY_SQ_SYi8cyApII,13585
 wisent/core/model_config_manager.py,sha256=rQAdSmk3GFlZXyHp3fSV1bORxiZWhmzIz1uo3H4JtkA,12009
 wisent/core/model_persistence.py,sha256=6_vc1Ndujd4v0O68giINSTvYhmb7-AiacWwAbqLOrls,10636
-wisent/core/multi_steering.py,sha256=PAn8eKFObR6QT-Mt1RbBB05AMr6RebGIBe3j5snSRHA,13069
+wisent/core/multi_steering.py,sha256=9HpI-4jL9PYAxp-4QrL47FYAapDGMDmnZB-RvHz7cNo,11729
 wisent/core/parser.py,sha256=_YDeSuQMx0zNknz9rX3Ls1YPT1x5eohoY8rfjeoqxV8,69091
 wisent/core/representation.py,sha256=hBl_N9qbr5Gsa7GCQ0nMWRm82RqYEfhd9cyf0PPH5LY,195
 wisent/core/sample_size_optimizer.py,sha256=6wegGXZpdGpiR4R0YJ1D2JqLr6yinMndEx2gB5FL80s,23666
@@ -213,8 +213,8 @@ wisent/synthetic/generators/diversities/core/__init__.py,sha256=47DEQpj8HBSa-_TI
 wisent/synthetic/generators/diversities/core/core.py,sha256=TjSj5T7NE5kRH-ABcFqb1Hz_j3Z6F_TcV-95uHD5Xw8,2201
 wisent/synthetic/generators/diversities/methods/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 wisent/synthetic/generators/diversities/methods/fast_diversity.py,sha256=Z2UzTbzyJFM_ToxCoXM_LQQQ1Jc6BZknrbpikTG1MRw,8522
-wisent-0.5.6.dist-info/licenses/LICENSE,sha256=wy0iaw8b2tyqZAfKHib3lP3PJ9o88FDCg92oUHh3sDQ,1073
-wisent-0.5.6.dist-info/METADATA,sha256=sR7d86PIFGErlOB8gKyjxB3iwlKX3ea__KEUsBCtKfg,2424
-wisent-0.5.6.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-wisent-0.5.6.dist-info/top_level.txt,sha256=2Ts9Iyldnb3auIN2HBBaHPknRy7nSRDm2f6RGzYgr8A,7
-wisent-0.5.6.dist-info/RECORD,,
+wisent-0.5.7.dist-info/licenses/LICENSE,sha256=wy0iaw8b2tyqZAfKHib3lP3PJ9o88FDCg92oUHh3sDQ,1073
+wisent-0.5.7.dist-info/METADATA,sha256=k6ECqqwEDcxK7hySjeY43r5v49fWesosVA-Y7vD4eVE,2424
+wisent-0.5.7.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+wisent-0.5.7.dist-info/top_level.txt,sha256=2Ts9Iyldnb3auIN2HBBaHPknRy7nSRDm2f6RGzYgr8A,7
+wisent-0.5.7.dist-info/RECORD,,