wisent 0.5.13__tar.gz → 0.5.15__tar.gz

{wisent-0.5.13 → wisent-0.5.15}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: wisent
-Version: 0.5.13
+Version: 0.5.15
 Summary: Monitor and guard against harmful content in language models
 Home-page: https://github.com/yourusername/wisent-activation-guardrails
 Author: Wisent Team
@@ -26,6 +26,8 @@ Requires-Dist: sentence-transformers>=2.0.0
 Requires-Dist: faiss-cpu>=1.7.0
 Provides-Extra: harness
 Requires-Dist: lm-eval==0.4.8; extra == "harness"
+Provides-Extra: cuda
+Requires-Dist: flash-attn>=2.5.0; extra == "cuda"
 Dynamic: author
 Dynamic: author-email
 Dynamic: classifier

{wisent-0.5.13 → wisent-0.5.15}/pyproject.toml

@@ -1,9 +1,9 @@
 [tool.pytest.ini_options]
-testpaths = ["tests"]
+testpaths = ["wisent/tests"]
 python_files = ["test_*.py"]
 python_classes = ["Test*"]
 python_functions = ["test_*"]
-addopts = ["-v", "--tb=short", "--ignore=evaluation", "-m", "not slow and not heavy and not docker and not bigcode_required"]
+addopts = ["-v", "--tb=short"]
 markers = [
     "slow: marks tests as slow (deselect with '-m \"not slow\"')",
     "heavy: marks tests as very resource-intensive with large model downloads (deselect with '-m \"not heavy\"')",

{wisent-0.5.13 → wisent-0.5.15}/setup.py

@@ -46,6 +46,9 @@ setup(
         "harness": [
             "lm-eval==0.4.8",
         ],
+        "cuda": [
+            "flash-attn>=2.5.0",
+        ],
     },
     entry_points={
         "console_scripts": [

wisent-0.5.15/wisent/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.5.15"

wisent-0.5.15/wisent/cli.py

@@ -0,0 +1,114 @@
+"""
+Programmatic API for running Wisent tasks.
+
+This module provides a Python API for running tasks that would normally be invoked via CLI.
+"""
+
+import argparse
+from typing import Dict, Any, Optional
+from wisent.core.cli.tasks import execute_tasks
+
+
+def run_task_pipeline(
+    task_name: str,
+    model_name: str,
+    layer: str,
+    training_limit: int,
+    testing_limit: int,
+    seed: int = 42,
+    verbose: bool = False,
+    split_ratio: float = 0.8,
+    limit: Optional[int] = None,
+    steering_mode: bool = False,
+    token_aggregation: str = "average",
+    detection_threshold: float = 0.5,
+    classifier_type: str = "logistic",
+    steering_method: str = "CAA",
+    steering_strength: float = 1.0,
+    token_targeting_strategy: str = "LAST_TOKEN",
+    **kwargs
+) -> Dict[str, Any]:
+    """
+    Run a task pipeline programmatically.
+
+    This function provides a programmatic interface to the tasks command,
+    allowing other modules to run tasks without invoking the CLI.
+
+    Args:
+        task_name: Name of the task to run
+        model_name: Name or path of the model
+        layer: Layer to use for activations
+        training_limit: Number of training samples
+        testing_limit: Number of testing samples
+        seed: Random seed for reproducibility
+        verbose: Whether to print verbose output
+        split_ratio: Train/test split ratio
+        limit: Total limit of samples to load
+        steering_mode: Whether to use steering mode
+        token_aggregation: Token aggregation strategy
+        detection_threshold: Detection threshold for classification
+        classifier_type: Type of classifier to use
+        steering_method: Steering method to use (if steering_mode=True)
+        steering_strength: Steering strength (if steering_mode=True)
+        token_targeting_strategy: Token targeting strategy for steering
+        **kwargs: Additional arguments
+
+    Returns:
+        Dictionary containing results including:
+        - test_accuracy: Test accuracy
+        - test_f1_score: Test F1 score
+        - training_time: Time spent training
+        - evaluation_results: Evaluation results dict
+    """
+    # Create a namespace object that mimics argparse output
+    args = argparse.Namespace()
+
+    # Set required arguments
+    # task_name is already a string, keep it as a list
+    args.task_names = [task_name] if isinstance(task_name, str) else task_name
+    args.model = model_name
+    args.layer = int(layer) if isinstance(layer, str) else layer
+    args.training_limit = training_limit
+    args.testing_limit = testing_limit
+    args.seed = seed
+    args.verbose = verbose
+    args.split_ratio = split_ratio
+    args.limit = limit if limit is not None else (training_limit + testing_limit + 100)
+
+    # Set method-specific arguments
+    if steering_mode:
+        args.steering_mode = True
+        args.steering_method = steering_method
+        args.steering_strength = steering_strength
+        args.token_targeting_strategy = token_targeting_strategy
+        args.token_aggregation = token_aggregation
+    else:
+        args.steering_mode = False
+        args.token_aggregation = token_aggregation
+        args.detection_threshold = detection_threshold
+        args.classifier_type = classifier_type
+
+    # Set defaults for other arguments
+    args.train_only = False
+    args.inference_only = False
+    args.save_classifier = None
+    args.load_classifier = None
+    args.save_steering_vector = None
+    args.load_steering_vector = None
+    args.output = None
+    args.evaluation_report = None
+    args.device = kwargs.get('device', 'cpu')
+
+    # Execute the task and capture results
+    results = execute_tasks(args)
+
+    # If execute_tasks returns None (shouldn't happen with our changes, but handle it)
+    if results is None:
+        return {
+            "test_accuracy": 0.0,
+            "test_f1_score": 0.0,
+            "training_time": 0.0,
+            "evaluation_results": {}
+        }
+
+    return results

{wisent-0.5.13 → wisent-0.5.15}/wisent/core/activations/activations_collector.py

@@ -230,17 +230,25 @@ class ActivationCollector:
             # Use model's built-in chat template
             if not hasattr(tokenizer, "apply_chat_template"):
                 raise RuntimeError("Tokenizer has no apply_chat_template; set it up or use a different strategy.")
-            prompt_text = tokenizer.apply_chat_template(
-                [{"role": "user", "content": prompt}],
-                tokenize=False,
-                add_generation_prompt=True,
-            )
-            full_text = tokenizer.apply_chat_template(
-                [{"role": "user", "content": prompt},
-                 {"role": "assistant", "content": response}],
-                tokenize=False,
-                add_generation_prompt=False,
-            )
+            try:
+                prompt_text = tokenizer.apply_chat_template(
+                    [{"role": "user", "content": prompt}],
+                    tokenize=False,
+                    add_generation_prompt=True,
+                )
+                full_text = tokenizer.apply_chat_template(
+                    [{"role": "user", "content": prompt},
+                     {"role": "assistant", "content": response}],
+                    tokenize=False,
+                    add_generation_prompt=False,
+                )
+            except ValueError as e:
+                if "chat_template is not set" in str(e):
+                    # Fallback to direct completion for models without chat templates
+                    prompt_text = prompt
+                    full_text = f"{prompt} {response}"
+                else:
+                    raise
 
         elif strategy == PromptConstructionStrategy.DIRECT_COMPLETION:
             # Q → good_resp/bad_resp (direct answer)

wisent-0.5.15/wisent/core/agent/__init__.py

@@ -0,0 +1 @@
+# Empty __init__.py to avoid cascading import errors with empty __init__ pattern

wisent-0.5.15/wisent/core/agent/diagnose/__init__.py

@@ -0,0 +1 @@
+# Empty __init__.py to avoid cascading import errors with empty __init__ pattern

{wisent-0.5.13 → wisent-0.5.15}/wisent/core/cli/__init__.py

@@ -9,7 +9,9 @@ from .generate_vector_from_task import execute_generate_vector_from_task
 from .generate_vector_from_synthetic import execute_generate_vector_from_synthetic
 from .optimize_classification import execute_optimize_classification
 from .optimize_steering import execute_optimize_steering
+from .optimize_sample_size import execute_optimize_sample_size
 from .generate_responses import execute_generate_responses
 from .evaluate_responses import execute_evaluate_responses
+from .multi_steer import execute_multi_steer
 
-__all__ = ['execute_tasks', 'execute_generate_pairs_from_task', 'execute_generate_pairs', 'execute_get_activations', 'execute_create_steering_vector', 'execute_generate_vector_from_task', 'execute_generate_vector_from_synthetic', 'execute_optimize_classification', 'execute_optimize_steering', 'execute_generate_responses', 'execute_evaluate_responses']
+__all__ = ['execute_tasks', 'execute_generate_pairs_from_task', 'execute_generate_pairs', 'execute_get_activations', 'execute_create_steering_vector', 'execute_generate_vector_from_task', 'execute_generate_vector_from_synthetic', 'execute_optimize_classification', 'execute_optimize_steering', 'execute_optimize_sample_size', 'execute_generate_responses', 'execute_evaluate_responses', 'execute_multi_steer']

{wisent-0.5.13 → wisent-0.5.15}/wisent/core/cli/create_steering_vector.py

@@ -94,28 +94,70 @@ def execute_create_steering_vector(args):
 
         print(f"   ✓ Generated {len(steering_vectors)} steering vectors")
 
-        # 5. Save steering vectors to JSON
+        # 5. Save steering vectors (format depends on file extension)
         print(f"\n💾 Saving steering vectors to '{args.output}'...")
-        output_data = {
-            'trait_label': trait_label,
-            'model': model,
-            'method': args.method,
-            'normalize': args.normalize,
-            'token_aggregation': token_aggregation,
-            'num_pairs': len(pairs_list),
-            'layers': list(steering_vectors.keys()),
-            'steering_vectors': steering_vectors,
-            'metadata': {
-                'source_file': args.enriched_pairs_file,
-                'creation_time': time.strftime('%Y-%m-%d %H:%M:%S'),
+        os.makedirs(os.path.dirname(os.path.abspath(args.output)) or '.', exist_ok=True)
+
+        if args.output.endswith('.pt'):
+            # For .pt format: save single-layer vectors for multi-steer compatibility
+            # If multiple layers, save the first one (or could save all and let user specify)
+            if len(steering_vectors) == 1:
+                layer_str = list(steering_vectors.keys())[0]
+                vector_tensor = torch.tensor(steering_vectors[layer_str], dtype=torch.float32)
+                torch.save({
+                    'steering_vector': vector_tensor,
+                    'layer_index': int(layer_str),
+                    'trait_label': trait_label,
+                    'model': model,
+                    'method': args.method,
+                    'normalize': args.normalize,
+                    'token_aggregation': token_aggregation,
+                    'num_pairs': len(pairs_list),
+                    # Legacy keys for backward compatibility
+                    'vector': vector_tensor,
+                    'layer': int(layer_str),
+                }, args.output)
+                print(f"   ✓ Saved steering vector (layer {layer_str}) to: {args.output}")
+            else:
+                # Save multiple layers - save each to separate file
+                for layer_str in steering_vectors.keys():
+                    layer_output = args.output.replace('.pt', f'_layer_{layer_str}.pt')
+                    vector_tensor = torch.tensor(steering_vectors[layer_str], dtype=torch.float32)
+                    torch.save({
+                        'steering_vector': vector_tensor,
+                        'layer_index': int(layer_str),
+                        'trait_label': trait_label,
+                        'model': model,
+                        'method': args.method,
+                        'normalize': args.normalize,
+                        'token_aggregation': token_aggregation,
+                        'num_pairs': len(pairs_list),
+                        # Legacy keys
+                        'vector': vector_tensor,
+                        'layer': int(layer_str),
+                    }, layer_output)
+                    print(f"   ✓ Saved steering vector (layer {layer_str}) to: {layer_output}")
+        else:
+            # JSON format: save all layers together
+            output_data = {
+                'trait_label': trait_label,
+                'model': model,
+                'method': args.method,
+                'normalize': args.normalize,
+                'token_aggregation': token_aggregation,
+                'num_pairs': len(pairs_list),
+                'layers': list(steering_vectors.keys()),
+                'steering_vectors': steering_vectors,
+                'metadata': {
+                    'source_file': args.enriched_pairs_file,
+                    'creation_time': time.strftime('%Y-%m-%d %H:%M:%S'),
+                }
             }
-        }
 
-        os.makedirs(os.path.dirname(os.path.abspath(args.output)), exist_ok=True)
-        with open(args.output, 'w') as f:
-            json.dump(output_data, f, indent=2)
+            with open(args.output, 'w') as f:
+                json.dump(output_data, f, indent=2)
 
-        print(f"   ✓ Saved steering vectors to: {args.output}")
+            print(f"   ✓ Saved steering vectors to: {args.output}")
 
         # 6. Display statistics
         print(f"\n📈 Steering Vector Statistics:")

{wisent-0.5.13 → wisent-0.5.15}/wisent/core/cli/evaluate_responses.py

@@ -33,8 +33,14 @@ def execute_evaluate_responses(args):
         with open(args.input, 'r') as f:
             input_data = json.load(f)
 
-        responses = input_data.get('responses', [])
-        task_name = args.task if args.task else input_data.get('task')
+        # Handle both dict format (with 'responses' key) and direct list format
+        if isinstance(input_data, list):
+            responses = input_data
+            task_name = args.task if args.task else "generic"  # Default to generic for list format
+        else:
+            responses = input_data.get('responses', [])
+            task_name = args.task if args.task else input_data.get('task', "generic")
+
         if not task_name:
             print(f"   ❌ Task name not found in input file and not provided via --task")
             sys.exit(1)
@@ -299,8 +305,8 @@ def execute_evaluate_responses(args):
 
         output_data = {
             "input_file": args.input,
-            "task": input_data.get('task'),
-            "model": input_data.get('model'),
+            "task": task_name if isinstance(input_data, list) else input_data.get('task'),
+            "model": None if isinstance(input_data, list) else input_data.get('model'),
             "evaluation_type": evaluation_type,
             "evaluator_used": "CodingEvaluator",
             "aggregated_metrics": aggregated_metrics,
@@ -454,8 +460,8 @@ def execute_evaluate_responses(args):
 
         output_data = {
             "input_file": args.input,
-            "task": input_data.get('task'),
-            "model": input_data.get('model'),
+            "task": task_name if isinstance(input_data, list) else input_data.get('task'),
+            "model": None if isinstance(input_data, list) else input_data.get('model'),
             "evaluation_type": evaluation_type,
             "evaluator_used": "PersonalizationEvaluator",
             "judge_model": judge_model,
@@ -683,8 +689,8 @@ def execute_evaluate_responses(args):
 
     output_data = {
         "input_file": args.input,
-        "task": input_data.get('task'),
-        "model": input_data.get('model'),
+        "task": task_name if isinstance(input_data, list) else input_data.get('task'),
+        "model": None if isinstance(input_data, list) else input_data.get('model'),
         "evaluation_type": evaluation_type,
         "evaluator_used": evaluator.name,
         "aggregated_metrics": aggregated_metrics,

{wisent-0.5.13 → wisent-0.5.15}/wisent/core/cli/generate_pairs_from_task.py

@@ -23,7 +23,7 @@ def _build_pairs_from_custom_task(task, limit: int | None):
         LiveCodeBenchExtractor as LiveCodeBenchPairExtractor
     )
 
-    task_name = task.get_name()
+    task_name = task.task_name
 
     if task_name == "livecodebench":
         # Use the contrastive pair extractor for LiveCodeBench
@@ -61,10 +61,11 @@ def execute_generate_pairs_from_task(args):
         from wisent.core.contrastive_pairs.lm_eval_pairs.lm_task_pairs_generation import (
             lm_build_contrastive_pairs,
         )
+        from wisent.core.task_interface import TaskInterface
 
-        # Handle both lm-eval tasks (dict) and custom tasks (TaskInterface)
+        # Handle both lm-eval tasks (dict or ConfigurableTask) and custom tasks (TaskInterface)
         if isinstance(task_obj, dict):
-            # lm-eval task
+            # lm-eval task group with subtasks
             if len(task_obj) != 1:
                 keys = ", ".join(sorted(task_obj.keys()))
                 raise ValueError(
@@ -81,14 +82,26 @@ def execute_generate_pairs_from_task(args):
                 lm_eval_task=task,
                 limit=args.limit,
             )
-        else:
-            # Custom task (TaskInterface)
+        elif isinstance(task_obj, TaskInterface):
+            # Custom task (TaskInterface) - only livecodebench for now
             task = task_obj
             pairs_task_name = args.task_name
 
             # 2. Generate contrastive pairs using custom task interface
             print(f"   🔨 Building contrastive pairs...")
             pairs = _build_pairs_from_custom_task(task, args.limit)
+        else:
+            # Single lm-eval task (ConfigurableTask), not wrapped in dict
+            task = task_obj
+            pairs_task_name = args.task_name
+
+            # 2. Generate contrastive pairs using lm-eval interface
+            print(f"   🔨 Building contrastive pairs...")
+            pairs = lm_build_contrastive_pairs(
+                task_name=pairs_task_name,
+                lm_eval_task=task,
+                limit=args.limit,
+            )
 
         print(f"   ✓ Generated {len(pairs)} contrastive pairs")
 

{wisent-0.5.13 → wisent-0.5.15}/wisent/core/cli/get_activations.py

@@ -42,7 +42,7 @@ def execute_get_activations(args):
             trait_label = 'unknown'
 
         # Apply limit if specified
-        if args.limit:
+        if hasattr(args, 'limit') and args.limit:
             pairs_list = pairs_list[:args.limit]
 
         print(f"   ✓ Loaded {len(pairs_list)} pairs")

wisent-0.5.15/wisent/core/cli/multi_steer.py

@@ -0,0 +1,108 @@
+"""Multi-steer command execution logic."""
+
+import sys
+import os
+import torch
+
+
+def execute_multi_steer(args):
+    """Execute the multi-steer command - combine multiple steering vectors and apply to generation."""
+    from wisent.core.multi_steering import MultiSteering, MultiSteeringError
+    from wisent.core.models.wisent_model import WisentModel
+
+    try:
+        # Check if no vectors provided - unsteered baseline mode
+        if not args.vector or len(args.vector) == 0:
+            print(f"\n🎯 Unsteered Baseline Generation Mode")
+            print(f"   Model: {args.model}")
+            print(f"   No steering vectors provided - generating baseline response")
+
+            # If prompt is provided, generate unsteered response
+            if hasattr(args, 'prompt') and args.prompt:
+                print(f"\n🤖 Loading model '{args.model}'...")
+                model = WisentModel(args.model, device=args.device)
+
+                # Generate WITHOUT steering
+                response = model.generate(
+                    [[{"role": "user", "content": args.prompt}]],
+                    max_new_tokens=args.max_new_tokens,
+                    do_sample=True,
+                    temperature=getattr(args, 'temperature', 0.7),
+                    top_p=getattr(args, 'top_p', 0.9)
+                )[0]
+
+                print(f"\nUnsteered baseline output:\n{response}\n")
+            else:
+                print(f"\n⚠️  No prompt provided. Use --prompt to generate output.")
+
+            print(f"\n✅ Baseline generation completed successfully!\n")
+            return
+
+        print(f"\n🎯 Multi-Steering Mode")
+        print(f"   Model: {args.model}")
+        print(f"   Layer: {args.layer}")
+        print(f"   Method: {args.method}")
+
+        # Initialize multi-steering
+        multi_steer = MultiSteering(device=args.device, method=args.method)
+
+        # Load and combine vectors
+        multi_steer.load_vectors(args.vector)
+
+        # Override layer if specified in args
+        if hasattr(args, 'layer') and args.layer:
+            multi_steer.layer = int(args.layer)
+
+        # Combine vectors
+        normalize = getattr(args, 'normalize_weights', True)
+        multi_steer.combine_vectors(normalize=normalize)
+
+        # Save combined vector if requested
+        if hasattr(args, 'save_combined') and args.save_combined:
+            print(f"\n💾 Saving combined vector to '{args.save_combined}'...")
+            os.makedirs(os.path.dirname(args.save_combined) or '.', exist_ok=True)
+            torch.save({
+                'steering_vector': multi_steer.combined_vector,
+                'layer_index': multi_steer.layer,
+                'method': args.method,
+                'model': args.model,
+                'weights': multi_steer.weights,
+                'num_vectors': len(multi_steer.loaded_vectors),
+                # Legacy keys for backward compatibility
+                'vector': multi_steer.combined_vector,
+                'layer': multi_steer.layer,
+            }, args.save_combined)
+            print(f"   ✓ Combined vector saved to: {args.save_combined}")
+
+        # If prompt is provided, apply steering and generate
+        if hasattr(args, 'prompt') and args.prompt:
+            print(f"\n🤖 Loading model '{args.model}'...")
+            model = WisentModel(args.model, device=args.device)
+
+            # Generate with steering
+            temperature = getattr(args, 'temperature', 0.7)
+            top_p = getattr(args, 'top_p', 0.9)
+            output = multi_steer.apply_steering(
+                model=model,
+                prompt=args.prompt,
+                max_new_tokens=args.max_new_tokens,
+                temperature=temperature,
+                top_p=top_p
+            )
+
+            print(f"\nGenerated output:\n{output}\n")
+
+        print(f"\n✅ Multi-steering completed successfully!\n")
+
+    except MultiSteeringError as e:
+        print(f"\n❌ Multi-steering error: {str(e)}", file=sys.stderr)
+        if args.verbose:
+            import traceback
+            traceback.print_exc()
+        sys.exit(1)
+    except Exception as e:
+        print(f"\n❌ Error: {str(e)}", file=sys.stderr)
+        if args.verbose:
+            import traceback
+            traceback.print_exc()
+        sys.exit(1)