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

wisent/core/agent/diagnose/create_classifier.py

@@ -0,0 +1,1154 @@
+"""
+On-the-Fly Classifier Creation System for Autonomous Agent
+
+This module handles:
+- Dynamic training of new classifiers for specific issue types
+- Automatic training data generation for different problem domains
+- Classifier optimization and validation
+- Integration with the autonomous agent system
+"""
+
+import time
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple
+
+from wisent_guard.core.classifier.classifier import ActivationClassifier, Classifier
+
+from ...activations import Activations
+from ...layer import Layer
+from ...model import Model
+from ...model_persistence import ModelPersistence, create_classifier_metadata
+
+
+@dataclass
+class TrainingConfig:
+    """Configuration for classifier training."""
+
+    issue_type: str
+    layer: int
+    classifier_type: str = "logistic"
+    threshold: float = 0.5
+    model_name: str = ""
+    training_samples: int = 100
+    test_split: float = 0.2
+    optimization_metric: str = "f1"
+    save_path: Optional[str] = None
+
+
+@dataclass
+class TrainingResult:
+    """Result of classifier training."""
+
+    classifier: Classifier
+    config: TrainingConfig
+    performance_metrics: Dict[str, float]
+    training_time: float
+    save_path: Optional[str] = None
+
+
+class ClassifierCreator:
+    """Creates new classifiers on demand for the autonomous agent."""
+
+    def __init__(self, model: Model):
+        """
+        Initialize the classifier creator.
+
+        Args:
+            model: The language model to use for training
+        """
+        self.model = model
+
+    def create_classifier_for_issue_type(
+        self, issue_type: str, layer: int, config: Optional[TrainingConfig] = None
+    ) -> TrainingResult:
+        """
+        Create a new classifier for a specific issue type.
+
+        Args:
+            issue_type: Type of issue to detect (e.g., "hallucination", "quality")
+            layer: Model layer to use for activation extraction
+            config: Optional training configuration
+
+        Returns:
+            TrainingResult with the trained classifier and metrics
+        """
+        print(f"🏋️ Creating classifier for {issue_type} at layer {layer}...")
+
+        # Use provided config or create default
+        if config is None:
+            config = TrainingConfig(issue_type=issue_type, layer=layer, model_name=self.model.name)
+
+        start_time = time.time()
+
+        # Generate training data
+        print("   📊 Generating training data...")
+        training_data = self._generate_training_data(issue_type, config.training_samples)
+
+        # Extract activations
+        print("   🧠 Extracting activations...")
+        harmful_activations, harmless_activations = self._extract_activations_from_data(training_data, layer)
+
+        # Train classifier
+        print("   🎯 Training classifier...")
+        classifier = self._train_classifier(harmful_activations, harmless_activations, config)
+
+        # Evaluate performance
+        print("   📈 Evaluating performance...")
+        metrics = self._evaluate_classifier(classifier, harmful_activations, harmless_activations)
+
+        training_time = time.time() - start_time
+
+        # Save classifier if path provided
+        save_path = None
+        if config.save_path:
+            print("   💾 Saving classifier...")
+            save_path = self._save_classifier(classifier, config, metrics)
+
+        result = TrainingResult(
+            classifier=classifier.classifier,  # Return the base classifier
+            config=config,
+            performance_metrics=metrics,
+            training_time=training_time,
+            save_path=save_path,
+        )
+
+        print(
+            f"   ✅ Classifier created in {training_time:.2f}s "
+            f"(F1: {metrics.get('f1', 0):.3f}, Accuracy: {metrics.get('accuracy', 0):.3f})"
+        )
+
+        return result
+
+    def create_multi_layer_classifiers(
+        self, issue_type: str, layers: List[int], save_base_path: Optional[str] = None
+    ) -> Dict[int, TrainingResult]:
+        """
+        Create classifiers for multiple layers for the same issue type.
+
+        Args:
+            issue_type: Type of issue to detect
+            layers: List of layers to create classifiers for
+            save_base_path: Base path for saving classifiers
+
+        Returns:
+            Dictionary mapping layer indices to training results
+        """
+        print(f"🔄 Creating multi-layer classifiers for {issue_type}...")
+
+        results = {}
+
+        for layer in layers:
+            config = TrainingConfig(
+                issue_type=issue_type,
+                layer=layer,
+                model_name=self.model.name,
+                save_path=f"{save_base_path}_layer_{layer}.pkl" if save_base_path else None,
+            )
+
+            result = self.create_classifier_for_issue_type(issue_type, layer, config)
+            results[layer] = result
+
+        print(f"   ✅ Created {len(results)} classifiers across layers {layers}")
+        return results
+
+    def optimize_classifier_for_performance(
+        self,
+        issue_type: str,
+        layer_range: Tuple[int, int] = None,
+        classifier_types: List[str] = None,
+        target_metric: str = "f1",
+        min_target_score: float = 0.7,
+    ) -> TrainingResult:
+        """
+        Optimize classifier by testing different configurations.
+
+        Args:
+            issue_type: Type of issue to detect
+            layer_range: Range of layers to test (start, end). If None, auto-detect all model layers
+            classifier_types: Types of classifiers to test
+            target_metric: Metric to optimize for
+            min_target_score: Minimum acceptable score
+
+        Returns:
+            Best performing classifier configuration
+        """
+        print(f"🎯 Optimizing classifier for {issue_type}...")
+
+        if classifier_types is None:
+            classifier_types = ["logistic", "mlp"]
+
+        # Auto-detect layer range if not provided
+        if layer_range is None:
+            from ..hyperparameter_optimizer import detect_model_layers
+
+            total_layers = detect_model_layers(self.model)
+            layer_range = (0, total_layers - 1)
+            print(f"   📊 Auto-detected {total_layers} layers, testing range {layer_range[0]}-{layer_range[1]}")
+
+        best_result = None
+        best_score = 0.0
+
+        layers_to_test = range(layer_range[0], layer_range[1] + 1, 2)  # Test every 2nd layer
+
+        for layer in layers_to_test:
+            for classifier_type in classifier_types:
+                config = TrainingConfig(
+                    issue_type=issue_type, layer=layer, classifier_type=classifier_type, model_name=self.model.name
+                )
+
+                try:
+                    result = self.create_classifier_for_issue_type(issue_type, layer, config)
+                    score = result.performance_metrics.get(target_metric, 0.0)
+
+                    print(f"      Layer {layer}, {classifier_type}: {target_metric}={score:.3f}")
+
+                    if score > best_score:
+                        best_score = score
+                        best_result = result
+
+                        # Early stopping if we hit the target
+                        if score >= min_target_score:
+                            print(f"      🎉 Target score reached: {score:.3f}")
+                            break
+
+                except Exception as e:
+                    print(f"      ❌ Failed layer {layer}, {classifier_type}: {e}")
+                    continue
+
+            # Break outer loop if target reached
+            if best_score >= min_target_score:
+                break
+
+        if best_result is None:
+            raise RuntimeError(f"Failed to create any working classifier for {issue_type}")
+
+        print(
+            f"   ✅ Best configuration: Layer {best_result.config.layer}, "
+            f"{best_result.config.classifier_type}, {target_metric}={best_score:.3f}"
+        )
+
+        return best_result
+
+    async def create_classifier_for_issue_with_benchmarks(
+        self,
+        issue_type: str,
+        relevant_benchmarks: List[str],
+        layer: int = 15,
+        num_samples: int = 50,
+        config: Optional[TrainingConfig] = None,
+    ) -> TrainingResult:
+        """
+        Create a classifier using specific benchmarks for better contrastive pairs.
+
+        Args:
+            issue_type: Type of issue to detect (e.g., "hallucination", "quality")
+            relevant_benchmarks: List of benchmark names to use for training data
+            layer: Model layer to use for activation extraction (default: 15)
+            num_samples: Number of training samples to generate
+            config: Optional training configuration
+
+        Returns:
+            TrainingResult with the trained classifier and metrics
+        """
+        print(f"🎯 Creating {issue_type} classifier using benchmarks: {relevant_benchmarks}")
+
+        # Use provided config or create default
+        if config is None:
+            config = TrainingConfig(
+                issue_type=issue_type, layer=layer, model_name=self.model.name, training_samples=num_samples
+            )
+
+        start_time = time.time()
+
+        # Generate training data using the provided benchmarks
+        print("   📊 Loading benchmark-specific training data...")
+        training_data = []
+
+        try:
+            # Load data from the relevant benchmarks
+            benchmark_data = self._load_benchmark_data(relevant_benchmarks, num_samples)
+            training_data.extend(benchmark_data)
+            print(f"      ✅ Loaded {len(benchmark_data)} examples from benchmarks")
+        except Exception as e:
+            print(f"      ⚠️ Failed to load benchmark data: {e}")
+
+        # If we don't have enough data from benchmarks, supplement with synthetic data
+        if len(training_data) < num_samples // 2:
+            print("   🧪 Supplementing with synthetic training data...")
+            try:
+                synthetic_data = self._generate_synthetic_training_data(issue_type, num_samples - len(training_data))
+                training_data.extend(synthetic_data)
+                print(f"      ✅ Added {len(synthetic_data)} synthetic examples")
+            except Exception as e:
+                print(f"      ⚠️ Failed to generate synthetic data: {e}")
+
+        if not training_data:
+            raise ValueError(f"No training data available for {issue_type}")
+
+        print(f"   📈 Total training examples: {len(training_data)}")
+
+        # Extract activations
+        print("   🧠 Extracting activations...")
+        harmful_activations, harmless_activations = self._extract_activations_from_data(training_data, layer)
+
+        # Train classifier
+        print("   🎯 Training classifier...")
+        classifier = self._train_classifier(harmful_activations, harmless_activations, config)
+
+        # Evaluate performance
+        print("   📈 Evaluating performance...")
+        metrics = self._evaluate_classifier(classifier, harmful_activations, harmless_activations)
+
+        training_time = time.time() - start_time
+
+        # Save classifier if path provided
+        save_path = None
+        if config.save_path:
+            print("   💾 Saving classifier...")
+            save_path = self._save_classifier(classifier, config, metrics)
+
+        result = TrainingResult(
+            classifier=classifier.classifier,  # Return the base classifier
+            config=config,
+            performance_metrics=metrics,
+            training_time=training_time,
+            save_path=save_path,
+        )
+
+        print(
+            f"   ✅ Benchmark-based classifier created in {training_time:.2f}s "
+            f"(F1: {metrics.get('f1', 0):.3f}, Accuracy: {metrics.get('accuracy', 0):.3f})"
+        )
+        print(f"      📊 Used benchmarks: {relevant_benchmarks}")
+
+        return result
+
+    async def create_combined_benchmark_classifier(
+        self, benchmark_names: List[str], classifier_params: "ClassifierParams", config: Optional[TrainingConfig] = None
+    ) -> TrainingResult:
+        """
+        Create a classifier trained on combined data from multiple benchmarks.
+
+        Args:
+            benchmark_names: List of benchmark names to combine training data from
+            classifier_params: Model-determined classifier parameters
+            config: Optional training configuration
+
+        Returns:
+            TrainingResult with the trained combined classifier
+        """
+        print(f"🏗️ Creating combined classifier from {len(benchmark_names)} benchmarks...")
+        print(f"   📊 Benchmarks: {benchmark_names}")
+        print(f"   🧠 Using layer {classifier_params.optimal_layer}, {classifier_params.training_samples} samples")
+
+        # Create config from classifier_params
+        if config is None:
+            config = TrainingConfig(
+                issue_type=f"quality_combined_{'_'.join(sorted(benchmark_names))}",
+                layer=classifier_params.optimal_layer,
+                classifier_type=classifier_params.classifier_type,
+                threshold=classifier_params.classification_threshold,
+                training_samples=classifier_params.training_samples,
+                model_name=self.model.name,
+            )
+
+        start_time = time.time()
+
+        # Generate combined training data from all benchmarks
+        print("   📊 Loading and combining benchmark training data...")
+        combined_training_data = await self._load_combined_benchmark_data(
+            benchmark_names, classifier_params.training_samples
+        )
+
+        print(f"   📈 Loaded {len(combined_training_data)} combined training examples")
+
+        # Extract activations
+        print("   🧠 Extracting activations...")
+        harmful_activations, harmless_activations = self._extract_activations_from_data(
+            combined_training_data, classifier_params.optimal_layer
+        )
+
+        # Train classifier
+        print("   🎯 Training combined classifier...")
+        classifier = self._train_classifier(harmful_activations, harmless_activations, config)
+
+        # Evaluate performance
+        print("   📈 Evaluating performance...")
+        metrics = self._evaluate_classifier(classifier, harmful_activations, harmless_activations)
+
+        training_time = time.time() - start_time
+
+        # Save classifier if path provided
+        save_path = None
+        if config.save_path:
+            print("   💾 Saving combined classifier...")
+            save_path = self._save_classifier(classifier, config, metrics)
+
+        result = TrainingResult(
+            classifier=classifier.classifier,
+            config=config,
+            performance_metrics=metrics,
+            training_time=training_time,
+            save_path=save_path,
+        )
+
+        print(
+            f"   ✅ Combined classifier created in {training_time:.2f}s "
+            f"(F1: {metrics.get('f1', 0):.3f}, Accuracy: {metrics.get('accuracy', 0):.3f})"
+        )
+
+        return result
+
+    async def _load_combined_benchmark_data(
+        self, benchmark_names: List[str], total_samples: int
+    ) -> List[Dict[str, Any]]:
+        """
+        Load and combine training data from multiple benchmarks.
+
+        Args:
+            benchmark_names: List of benchmark names to load data from
+            total_samples: Total number of training samples to create
+
+        Returns:
+            Combined list of training examples with balanced sampling
+        """
+        combined_data = []
+        samples_per_benchmark = max(1, total_samples // len(benchmark_names))
+
+        print(f"      📊 Loading ~{samples_per_benchmark} samples per benchmark")
+
+        for benchmark_name in benchmark_names:
+            try:
+                print(f"      🔄 Loading data from {benchmark_name}...")
+                benchmark_data = self._load_benchmark_data([benchmark_name], samples_per_benchmark)
+                combined_data.extend(benchmark_data)
+                print(f"         ✅ Loaded {len(benchmark_data)} samples from {benchmark_name}")
+
+            except Exception as e:
+                print(f"         ⚠️ Failed to load {benchmark_name}: {e}")
+                # Continue with other benchmarks
+                continue
+
+        # If we don't have enough samples, pad with synthetic data
+        if len(combined_data) < total_samples:
+            remaining_samples = total_samples - len(combined_data)
+            print(f"      🔧 Generating {remaining_samples} synthetic samples to reach target")
+            synthetic_data = self._generate_synthetic_training_data("quality", remaining_samples)
+            combined_data.extend(synthetic_data)
+
+        # Shuffle the combined data to ensure good mixing
+        import random
+
+        random.shuffle(combined_data)
+
+        # Trim to exact target if we have too many
+        combined_data = combined_data[:total_samples]
+
+        print(f"      ✅ Final combined dataset: {len(combined_data)} samples")
+        return combined_data
+
+    async def create_classifier_for_issue(self, issue_type: str, layer: int = 15) -> TrainingResult:
+        """
+        Create a classifier for an issue type (async version for compatibility).
+
+        Args:
+            issue_type: Type of issue to detect
+            layer: Model layer to use for activation extraction
+
+        Returns:
+            TrainingResult with the trained classifier
+        """
+        return self.create_classifier_for_issue_type(issue_type, layer)
+
+    def _generate_training_data(self, issue_type: str, num_samples: int) -> List[Dict[str, Any]]:
+        """
+        Generate training data dynamically for a specific issue type using relevant benchmarks.
+
+        Args:
+            issue_type: Type of issue to generate data for
+            num_samples: Number of training samples to generate
+
+        Returns:
+            List of training examples with harmful/harmless pairs
+        """
+        print(f"   📊 Loading dynamic training data for {issue_type}...")
+
+        # Try to find relevant benchmarks for the issue type (using default 5-minute budget)
+        relevant_benchmarks = self._find_relevant_benchmarks(issue_type)
+
+        if relevant_benchmarks:
+            print(f"   🎯 Found {len(relevant_benchmarks)} relevant benchmarks: {relevant_benchmarks[:3]}...")
+            return self._load_benchmark_data(relevant_benchmarks, num_samples)
+        print("   🤖 No specific benchmarks found, using synthetic generation...")
+        return self._generate_synthetic_training_data(issue_type, num_samples)
+
+    def _find_relevant_benchmarks(self, issue_type: str, time_budget_minutes: float = 5.0) -> List[str]:
+        """Find relevant benchmarks for the given issue type based on time budget with priority-aware selection."""
+        from ..budget import calculate_max_tasks_for_time_budget
+        from .tasks.task_relevance import find_relevant_tasks
+
+        try:
+            # Calculate max tasks using budget system
+            max_tasks = calculate_max_tasks_for_time_budget(
+                task_type="benchmark_evaluation", time_budget_minutes=time_budget_minutes
+            )
+
+            print(f"   🕐 Time budget: {time_budget_minutes:.1f}min → max {max_tasks} tasks")
+
+            # Use priority-aware intelligent benchmark selection
+            try:
+                # Import priority-aware selection function
+                import os
+                import sys
+
+                sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..", "lm-harness-integration"))
+                from only_benchmarks import find_most_relevant_benchmarks
+
+                # Use priority-aware selection with time budget
+                relevant_results = find_most_relevant_benchmarks(
+                    prompt=issue_type,
+                    top_k=max_tasks,
+                    priority="all",
+                    fast_only=False,
+                    time_budget_minutes=time_budget_minutes,
+                    prefer_fast=True,  # Prefer fast benchmarks for agent use
+                )
+
+                # Extract benchmark names
+                relevant_benchmarks = [result["benchmark"] for result in relevant_results]
+
+                if relevant_benchmarks:
+                    print(f"   📊 Found {len(relevant_benchmarks)} priority-aware benchmarks for '{issue_type}':")
+                    for i, result in enumerate(relevant_results[:3]):
+                        priority_str = f" (priority: {result.get('priority', 'unknown')})"
+                        loading_time_str = f" (loading time: {result.get('loading_time', 60.0):.1f}s)"
+                        print(f"      {i + 1}. {result['benchmark']}{priority_str}{loading_time_str}")
+                    if len(relevant_benchmarks) > 3:
+                        print(f"      ... and {len(relevant_benchmarks) - 3} more")
+
+                return relevant_benchmarks
+
+            except Exception as priority_error:
+                print(f"   ⚠️ Priority-aware selection failed: {priority_error}")
+                print("   🔄 Falling back to legacy task relevance...")
+
+                # Fallback to legacy system
+                relevant_task_results = find_relevant_tasks(
+                    query=issue_type, max_results=max_tasks, min_relevance_score=0.1
+                )
+
+                # Extract just the task names
+                candidate_benchmarks = [task_name for task_name, score in relevant_task_results]
+
+                # Use priority-aware budget optimization
+                from ..budget import optimize_benchmarks_for_budget
+
+                relevant_benchmarks = optimize_benchmarks_for_budget(
+                    task_candidates=candidate_benchmarks,
+                    time_budget_minutes=time_budget_minutes,
+                    max_tasks=max_tasks,
+                    prefer_fast=True,  # Agent prefers fast benchmarks
+                )
+
+                if relevant_benchmarks:
+                    print(f"   📊 Found {len(relevant_benchmarks)} relevant benchmarks for '{issue_type}':")
+                    # Show the scores for the selected benchmarks
+                    for i, (task_name, score) in enumerate(relevant_task_results[:3]):
+                        if task_name in relevant_benchmarks:
+                            print(f"      {i + 1}. {task_name} (relevance: {score:.3f})")
+                    if len(relevant_benchmarks) > 3:
+                        print(f"      ... and {len(relevant_benchmarks) - 3} more")
+
+                return relevant_benchmarks
+
+        except Exception as e:
+            print(f"   ⚠️ Error finding relevant benchmarks: {e}")
+            print("   ⚠️ Using fallback tasks")
+            # Minimal fallback to high priority fast benchmarks
+            return ["mmlu", "truthfulqa_mc1", "hellaswag"]
+
+    def _extract_benchmark_concepts(self, benchmark_names: List[str]) -> Dict[str, List[str]]:
+        """Extract semantic concepts from benchmark names."""
+        concepts = {}
+
+        for name in benchmark_names:
+            # Extract concepts from benchmark name
+            name_concepts = []
+            name_lower = name.lower()
+
+            # Split on common separators and extract meaningful tokens
+            tokens = name_lower.replace("_", " ").replace("-", " ").split()
+
+            # Filter out common non-semantic tokens
+            semantic_tokens = []
+            skip_tokens = {
+                "the",
+                "and",
+                "or",
+                "of",
+                "in",
+                "on",
+                "at",
+                "to",
+                "for",
+                "with",
+                "by",
+                "from",
+                "as",
+                "is",
+                "are",
+                "was",
+                "were",
+                "be",
+                "been",
+                "being",
+                "have",
+                "has",
+                "had",
+                "do",
+                "does",
+                "did",
+                "will",
+                "would",
+                "could",
+                "should",
+                "may",
+                "might",
+                "can",
+                "light",
+                "full",
+                "val",
+                "test",
+                "dev",
+                "mc1",
+                "mc2",
+                "mt",
+                "cot",
+                "fewshot",
+                "zeroshot",
+                "generate",
+                "until",
+                "multiple",
+                "choice",
+                "group",
+                "subset",
+            }
+
+            for token in tokens:
+                if len(token) > 2 and token not in skip_tokens and token.isalpha():
+                    semantic_tokens.append(token)
+
+            # Extract domain-specific concepts
+            domain_concepts = self._extract_domain_concepts(name_lower, semantic_tokens)
+            name_concepts.extend(domain_concepts)
+
+            concepts[name] = list(set(name_concepts))  # Remove duplicates
+
+        return concepts
+
+    def _extract_domain_concepts(self, benchmark_name: str, tokens: List[str]) -> List[str]:
+        """Extract domain-specific concepts directly from benchmark name components."""
+        concepts = []
+
+        # Add all meaningful tokens as concepts
+        for token in tokens:
+            if len(token) > 2:
+                concepts.append(token)
+
+        # Extract compound concept meanings from token combinations
+        name_parts = benchmark_name.lower().split("_")
+
+        # Generate concept combinations
+        for i, part in enumerate(name_parts):
+            if len(part) > 2:
+                concepts.append(part)
+
+                # Look for meaningful compound concepts
+                if i < len(name_parts) - 1:
+                    next_part = name_parts[i + 1]
+                    if len(next_part) > 2:
+                        compound = f"{part}_{next_part}"
+                        concepts.append(compound)
+
+        # Extract semantic root words
+        for token in tokens:
+            root_concepts = self._extract_semantic_roots(token)
+            concepts.extend(root_concepts)
+
+        return list(set(concepts))  # Remove duplicates
+
+    def _extract_semantic_roots(self, word: str) -> List[str]:
+        """Extract semantic root concepts from a word."""
+        roots = []
+
+        # Simple morphological analysis
+        # Remove common suffixes to find roots
+        suffixes = [
+            "ing",
+            "tion",
+            "sion",
+            "ness",
+            "ment",
+            "able",
+            "ible",
+            "ful",
+            "less",
+            "ly",
+            "al",
+            "ic",
+            "ous",
+            "ive",
+        ]
+
+        root = word
+        for suffix in suffixes:
+            if word.endswith(suffix) and len(word) > len(suffix) + 2:
+                root = word[: -len(suffix)]
+                break
+
+        if root != word and len(root) > 2:
+            roots.append(root)
+
+        # Add the original word
+        roots.append(word)
+
+        return roots
+
+    def _calculate_benchmark_relevance(self, issue_type: str, benchmark_concepts: Dict[str, List[str]]) -> List[str]:
+        """Calculate relevance scores using semantic similarity."""
+        # Calculate relevance scores
+        benchmark_scores = []
+
+        for benchmark_name, concepts in benchmark_concepts.items():
+            score = self._calculate_semantic_similarity(issue_type, benchmark_name, concepts)
+
+            if score > 0:
+                benchmark_scores.append((benchmark_name, score))
+
+        # Sort by relevance score
+        benchmark_scores.sort(key=lambda x: x[1], reverse=True)
+
+        return [name for name, score in benchmark_scores]
+
+    def _calculate_semantic_similarity(self, issue_type: str, benchmark_name: str, concepts: List[str]) -> float:
+        """Calculate semantic similarity between issue type and benchmark."""
+        issue_lower = issue_type.lower()
+        benchmark_lower = benchmark_name.lower()
+
+        score = 0.0
+
+        # Direct name matching (highest weight)
+        if issue_lower in benchmark_lower or benchmark_lower in issue_lower:
+            score += 5.0
+
+        # Concept matching
+        for concept in concepts:
+            concept_lower = concept.lower()
+
+            # Exact concept match
+            if issue_lower == concept_lower:
+                score += 4.0
+            # Partial concept match
+            elif issue_lower in concept_lower or concept_lower in issue_lower:
+                score += 2.0
+            # Semantic similarity check
+            elif self._are_semantically_similar(issue_lower, concept_lower):
+                score += 1.5
+
+        # Token-level similarity in benchmark name
+        benchmark_tokens = benchmark_lower.replace("_", " ").replace("-", " ").split()
+        issue_tokens = issue_lower.replace("_", " ").replace("-", " ").split()
+
+        for issue_token in issue_tokens:
+            for benchmark_token in benchmark_tokens:
+                if len(issue_token) > 2 and len(benchmark_token) > 2:
+                    if issue_token == benchmark_token:
+                        score += 3.0
+                    elif issue_token in benchmark_token or benchmark_token in issue_token:
+                        score += 1.0
+                    elif self._are_semantically_similar(issue_token, benchmark_token):
+                        score += 0.5
+
+        return score
+
+    def _are_semantically_similar(self, term1: str, term2: str) -> bool:
+        """Check if two terms are semantically similar using algorithmic methods."""
+        if len(term1) < 3 or len(term2) < 3:
+            return False
+
+        # Character-level similarity
+        overlap = len(set(term1) & set(term2))
+        min_len = min(len(term1), len(term2))
+        char_similarity = overlap / min_len
+
+        # Substring similarity
+        longer, shorter = (term1, term2) if len(term1) > len(term2) else (term2, term1)
+        substring_match = shorter in longer
+
+        # Prefix/suffix similarity
+        prefix_len = 0
+        suffix_len = 0
+
+        for i in range(min(len(term1), len(term2))):
+            if term1[i] == term2[i]:
+                prefix_len += 1
+            else:
+                break
+
+        for i in range(1, min(len(term1), len(term2)) + 1):
+            if term1[-i] == term2[-i]:
+                suffix_len += 1
+            else:
+                break
+
+        affix_similarity = (prefix_len + suffix_len) / max(len(term1), len(term2))
+
+        # Combined similarity score
+        return char_similarity > 0.6 or substring_match or affix_similarity > 0.4 or prefix_len >= 3 or suffix_len >= 3
+
+    def _prioritize_benchmarks(self, relevant_benchmarks: List[str]) -> List[str]:
+        """Prioritize benchmarks algorithmically based on naming patterns and characteristics."""
+        benchmark_scores = []
+
+        for benchmark in relevant_benchmarks:
+            score = self._calculate_benchmark_quality_score(benchmark)
+            benchmark_scores.append((benchmark, score))
+
+        # Sort by quality score (higher is better)
+        benchmark_scores.sort(key=lambda x: x[1], reverse=True)
+        return [benchmark for benchmark, score in benchmark_scores]
+
+    def _calculate_benchmark_quality_score(self, benchmark_name: str) -> float:
+        """Calculate quality score for a benchmark based on naming patterns and characteristics."""
+        score = 0.0
+        benchmark_lower = benchmark_name.lower()
+
+        # Length heuristic - moderate length names tend to be well-established
+        name_length = len(benchmark_name)
+        if 8 <= name_length <= 25:
+            score += 2.0
+        elif name_length < 8:
+            score += 0.5  # Very short names might be too simple
+        else:
+            score += 1.0  # Very long names might be overly specific
+
+        # Component analysis
+        parts = benchmark_lower.split("_")
+        num_parts = len(parts)
+
+        # Well-structured benchmarks often have 2-3 parts
+        if 2 <= num_parts <= 3:
+            score += 2.0
+        elif num_parts == 1:
+            score += 1.5  # Simple names can be good too
+        else:
+            score += 0.5  # Too many parts might indicate over-specification
+
+        # Indicator of established benchmarks (avoid hardcoding specific names)
+        quality_indicators = [
+            # Multiple choice indicators (often well-validated)
+            ("mc1", 1.5),
+            ("mc2", 1.5),
+            ("multiple_choice", 1.5),
+            # Evaluation methodology indicators
+            ("eval", 1.0),
+            ("test", 1.0),
+            ("benchmark", 1.0),
+            # Language understanding indicators
+            ("language", 1.0),
+            ("understanding", 1.0),
+            ("comprehension", 1.0),
+            # Logic and reasoning indicators
+            ("logic", 1.0),
+            ("reasoning", 1.0),
+            ("deduction", 1.0),
+            # Knowledge assessment indicators
+            ("knowledge", 1.0),
+            ("question", 1.0),
+            ("answer", 1.0),
+        ]
+
+        for indicator, points in quality_indicators:
+            if indicator in benchmark_lower:
+                score += points
+
+        # Penalize very specialized or experimental indicators
+        experimental_indicators = [
+            "experimental",
+            "pilot",
+            "demo",
+            "sample",
+            "tiny",
+            "mini",
+            "subset",
+            "light",
+            "debug",
+            "test_only",
+        ]
+
+        for indicator in experimental_indicators:
+            if indicator in benchmark_lower:
+                score -= 1.0
+
+        # Bonus for domain diversity indicators
+        domain_indicators = ["multilingual", "global", "cross", "multi", "diverse"]
+
+        for indicator in domain_indicators:
+            if indicator in benchmark_lower:
+                score += 0.5
+
+        return max(0.0, score)  # Ensure non-negative score
+
+    def _load_benchmark_data(self, benchmarks: List[str], num_samples: int) -> List[Dict[str, Any]]:
+        """Load training data from multiple relevant benchmarks."""
+        from .tasks import TaskManager
+
+        training_data = []
+        samples_per_benchmark = max(1, num_samples // len(benchmarks))
+
+        # Create task manager instance
+        task_manager = TaskManager()
+
+        for benchmark in benchmarks:
+            try:
+                print(f"     🔄 Loading from {benchmark}...")
+
+                # Load benchmark task using TaskManager
+                task_data = task_manager.load_task(benchmark, limit=samples_per_benchmark * 3)
+                docs = task_manager.split_task_data(task_data, split_ratio=1.0)[0]
+
+                # Extract QA pairs using existing system
+                from ...contrastive_pairs.contrastive_pair_set import ContrastivePairSet
+
+                qa_pairs = ContrastivePairSet.extract_qa_pairs_from_task_docs(benchmark, task_data, docs)
+
+                # Convert to training format
+                for pair in qa_pairs[:samples_per_benchmark]:
+                    if self._is_valid_pair(pair):
+                        training_data.append(
+                            {
+                                "prompt": pair.get("question", f"Context from {benchmark}"),
+                                "harmful_response": pair.get("incorrect_answer", ""),
+                                "harmless_response": pair.get("correct_answer", ""),
+                                "source": benchmark,
+                            }
+                        )
+
+                print(f"     ✅ Loaded {len(qa_pairs[:samples_per_benchmark])} examples from {benchmark}")
+
+                # Stop if we have enough data
+                if len(training_data) >= num_samples:
+                    break
+
+            except Exception as e:
+                print(f"     ⚠️ Failed to load {benchmark}: {e}")
+                continue
+
+        if not training_data:
+            print("   ❌ Failed to load from any benchmarks, falling back to synthetic...")
+            return self._generate_synthetic_training_data("unknown", num_samples)
+
+        print(
+            f"   ✅ Loaded {len(training_data)} examples from {len([b for b in benchmarks if any(b in item.get('source', '') for item in training_data)])} benchmarks"
+        )
+        return training_data[:num_samples]
+
+    def _is_valid_pair(self, pair: Dict[str, Any]) -> bool:
+        """Check if a QA pair is valid for training."""
+        return (
+            pair.get("correct_answer")
+            and pair.get("incorrect_answer")
+            and len(pair.get("correct_answer", "")) > 5
+            and len(pair.get("incorrect_answer", "")) > 5
+            and pair.get("correct_answer") != pair.get("incorrect_answer")
+        )
+
+    def _generate_synthetic_training_data(self, issue_type: str, num_samples: int) -> List[Dict[str, Any]]:
+        """Generate synthetic training data for issue types."""
+        from ...contrastive_pairs.generate_synthetically import SyntheticContrastivePairGenerator
+
+        try:
+            generator = SyntheticContrastivePairGenerator(self.model)
+
+            # Create trait description based on issue type
+            trait_descriptions = {
+                "quality": "providing clear, accurate, and well-structured responses",
+                "harmful": "giving safe, helpful, and constructive advice",
+                "bias": "responding with fairness and avoiding stereotypes",
+                "coherence": "maintaining logical flow and clear communication",
+                "hallucination": "providing factually accurate information",
+            }
+
+            trait_description = trait_descriptions.get(issue_type, f"avoiding {issue_type} issues in responses")
+
+            # Generate synthetic pairs
+            synthetic_pairs = generator.generate_contrastive_pair_set(
+                trait_description=trait_description, num_pairs=num_samples, name=f"synthetic_{issue_type}"
+            )
+
+            # Convert to training format
+            training_data = []
+            for pair in synthetic_pairs.pairs[:num_samples]:
+                training_data.append(
+                    {
+                        "prompt": pair.prompt or f"Context for {issue_type} detection",
+                        "harmful_response": pair.negative_response,
+                        "harmless_response": pair.positive_response,
+                    }
+                )
+
+            print(f"   ✅ Generated {len(training_data)} synthetic examples for {issue_type}")
+            return training_data
+
+        except Exception as e:
+            print(f"   ❌ Failed to generate synthetic data: {e}")
+            raise ValueError(f"Cannot generate training data for issue type: {issue_type}")
+
+    def _extract_activations_from_data(
+        self, training_data: List[Dict[str, Any]], layer: int
+    ) -> Tuple[List[Activations], List[Activations]]:
+        """
+        Extract activations from training data.
+
+        Args:
+            training_data: List of training examples
+            layer: Layer to extract activations from
+
+        Returns:
+            Tuple of (harmful_activations, harmless_activations)
+        """
+        harmful_activations = []
+        harmless_activations = []
+
+        layer_obj = Layer(index=layer, type="transformer")
+
+        for example in training_data:
+            # Extract harmful activation
+            harmful_tensor = self.model.extract_activations(example["harmful_response"], layer_obj)
+            harmful_activation = Activations(tensor=harmful_tensor, layer=layer_obj)
+            harmful_activations.append(harmful_activation)
+
+            # Extract harmless activation
+            harmless_tensor = self.model.extract_activations(example["harmless_response"], layer_obj)
+            harmless_activation = Activations(tensor=harmless_tensor, layer=layer_obj)
+            harmless_activations.append(harmless_activation)
+
+        return harmful_activations, harmless_activations
+
+    def _train_classifier(
+        self, harmful_activations: List[Activations], harmless_activations: List[Activations], config: TrainingConfig
+    ) -> ActivationClassifier:
+        """
+        Train a classifier on the activation data.
+
+        Args:
+            harmful_activations: List of harmful activations
+            harmless_activations: List of harmless activations
+            config: Training configuration
+
+        Returns:
+            Trained ActivationClassifier
+        """
+        classifier = ActivationClassifier(
+            model_type=config.classifier_type, threshold=config.threshold, device=self.model.device
+        )
+
+        classifier.train_on_activations(harmful_activations, harmless_activations)
+
+        return classifier
+
+    def _evaluate_classifier(
+        self,
+        classifier: ActivationClassifier,
+        harmful_activations: List[Activations],
+        harmless_activations: List[Activations],
+    ) -> Dict[str, float]:
+        """
+        Evaluate classifier performance.
+
+        Args:
+            classifier: Trained classifier
+            harmful_activations: Test harmful activations
+            harmless_activations: Test harmless activations
+
+        Returns:
+            Dictionary of performance metrics
+        """
+        # Use a portion of data for testing
+        test_size = min(10, len(harmful_activations) // 5)  # 20% or at least 10
+
+        test_harmful = harmful_activations[-test_size:]
+        test_harmless = harmless_activations[-test_size:]
+
+        return classifier.evaluate_on_activations(test_harmful, test_harmless)
+
+    def _save_classifier(
+        self, classifier: ActivationClassifier, config: TrainingConfig, metrics: Dict[str, float]
+    ) -> str:
+        """
+        Save classifier with metadata.
+
+        Args:
+            classifier: Trained classifier
+            config: Training configuration
+            metrics: Performance metrics
+
+        Returns:
+            Path where classifier was saved
+        """
+        # Create metadata
+        metadata = create_classifier_metadata(
+            model_name=config.model_name,
+            task_name=config.issue_type,
+            layer=config.layer,
+            classifier_type=config.classifier_type,
+            training_accuracy=metrics.get("accuracy", 0.0),
+            training_samples=config.training_samples,
+            token_aggregation="final",  # Default for our system
+            detection_threshold=config.threshold,
+            f1=metrics.get("f1", 0.0),
+            precision=metrics.get("precision", 0.0),
+            recall=metrics.get("recall", 0.0),
+            auc=metrics.get("auc", 0.0),
+        )
+
+        # Save using ModelPersistence
+        save_path = ModelPersistence.save_classifier(classifier.classifier, config.layer, config.save_path, metadata)
+
+        return save_path
+
+
+def create_classifier_on_demand(
+    model: Model, issue_type: str, layer: int = None, save_path: str = None, optimize: bool = False
+) -> TrainingResult:
+    """
+    Convenience function to create a classifier on demand.
+
+    Args:
+        model: Language model to use
+        issue_type: Type of issue to detect
+        layer: Specific layer to use (auto-optimized if None)
+        save_path: Path to save the classifier
+        optimize: Whether to optimize for best performance
+
+    Returns:
+        TrainingResult with the created classifier
+    """
+    creator = ClassifierCreator(model)
+
+    if optimize or layer is None:
+        # Optimize to find best configuration
+        result = creator.optimize_classifier_for_performance(issue_type)
+
+        # Save if path provided
+        if save_path:
+            result.config.save_path = save_path
+            result.save_path = creator._save_classifier(
+                ActivationClassifier(device=model.device), result.config, result.performance_metrics
+            )
+
+        return result
+    # Use specified layer
+    config = TrainingConfig(issue_type=issue_type, layer=layer, save_path=save_path, model_name=model.name)
+
+    return creator.create_classifier_for_issue_type(issue_type, layer, config)