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

wisent/core/optuna/steering/optuna_pipeline.py

@@ -0,0 +1,1738 @@
+"""
+Dataset-Agnostic Optimization Pipeline with Optuna
+
+This script builds a reproducible pipeline that:
+1. Trains probes and learns steering vectors on the training split
+2. Selects the best layer, probe type, steering method, and hyperparameters on validation split via Optuna
+3. Evaluates once on the test split with the single best configuration determined on validation
+
+Key features:
+- Optuna-based hyperparameter optimization with pruners
+- Activation caching for efficiency
+- Configurable datasets for train/val/test splits
+- Steering evaluation with model re-forwarding
+- Reproducibility bundle generation
+"""
+
+import hashlib
+import json
+import logging
+import os
+import pickle
+from dataclasses import asdict, dataclass, field
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Optional
+
+import numpy as np
+import optuna
+import torch
+from optuna.pruners import MedianPruner, SuccessiveHalvingPruner
+from optuna.samplers import TPESampler
+from safetensors.torch import save_file as safetensors_save
+from tqdm import tqdm
+
+# Optional WandB integration
+try:
+    import wandb
+
+    WANDB_AVAILABLE = True
+except ImportError:
+    WANDB_AVAILABLE = False
+from wisent_guard.core.contrastive_pairs.contrastive_pair import ContrastivePair
+from wisent_guard.core.contrastive_pairs.contrastive_pair_set import ContrastivePairSet
+from wisent_guard.core.optuna.steering import data_utils, metrics
+from wisent_guard.core.response import Response
+from wisent_guard.core.steering_methods.dac import DAC
+from wisent_guard.core.task_interface import get_task
+from wisent_guard.core.utils.device import empty_device_cache, preferred_dtype, resolve_default_device, resolve_device
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class OptimizationConfig:
+    """Configuration for dataset-agnostic optimization pipeline."""
+
+    model_name: str = "realtreetune/rho-1b-sft-GSM8K"
+    device: str = field(default_factory=resolve_default_device)
+
+    train_dataset: str = "gsm8k"
+    val_dataset: str = "gsm8k"
+    test_dataset: str = "gsm8k"
+
+    # Training configuration
+    train_limit: int = 50  # How many training samples to load
+    contrastive_pairs_limit: int = 20  # How many contrastive pairs to extract for steering training
+
+    # Evaluation configuration
+    val_limit: int = 50  # How many validation samples to load
+    test_limit: int = 100  # How many test samples to load
+
+    layer_search_range: tuple[int, int] = (15, 20)
+    probe_type: str = "logistic_regression"  # Fixed probe type
+    steering_methods: list[str] = field(default_factory=lambda: ["dac", "caa"])  # TODO add more
+
+    # Optuna study configuration
+    study_name: str = "optimization_pipeline"
+    db_url: str = field(
+        default_factory=lambda: f"sqlite:///{os.path.dirname(os.path.dirname(__file__))}/optuna_studies.db"
+    )
+    n_trials: int = 50
+    n_startup_trials: int = 10  # Random exploration before TPE kicks in
+    sampler: str = "TPE"
+    pruner: str = "MedianPruner"
+
+    # WandB configuration
+    wandb_project: str = "wisent-guard-optimization"
+    use_wandb: bool = False  # TODO
+
+    batch_size: int = 8
+    max_length: int = 512
+    max_new_tokens: int = 256
+    seed: int = 42
+
+    temperature: float = 0.0
+    do_sample: bool = False
+
+    output_dir: str = "outputs/optimization_pipeline"
+    cache_dir: str = "cache/optimization_pipeline"
+
+    max_layers_to_search: int = 6
+    early_stopping_patience: int = 10
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert to dictionary for serialization."""
+        return asdict(self)
+
+
+class ActivationCache:
+    """Efficient activation caching system with proper cache keys."""
+
+    def __init__(self, cache_dir: str):
+        self.cache_dir = Path(cache_dir)
+        self.cache_dir.mkdir(parents=True, exist_ok=True)
+        self.logger = logging.getLogger(f"{__name__}.{self.__class__.__name__}")
+
+    def _generate_cache_key(
+        self, split: str, layer_id: int, tokenization_config: dict[str, Any], prompt_variant: str = "default"
+    ) -> str:
+        """Generate unique cache key for activations."""
+        config_str = json.dumps(tokenization_config, sort_keys=True)
+        key_data = f"{split}_{layer_id}_{config_str}_{prompt_variant}"
+        return hashlib.md5(key_data.encode()).hexdigest()
+
+    def _get_cache_path(self, cache_key: str) -> Path:
+        """Get cache file path for key."""
+        return self.cache_dir / f"activations_{cache_key}.pkl"
+
+    def has_cached_activations(
+        self, split: str, layer_id: int, tokenization_config: dict[str, Any], prompt_variant: str = "default"
+    ) -> bool:
+        """Check if activations are cached."""
+        cache_key = self._generate_cache_key(split, layer_id, tokenization_config, prompt_variant)
+        return self._get_cache_path(cache_key).exists()
+
+    def save_activations(
+        self,
+        activations: np.ndarray,
+        labels: np.ndarray,
+        split: str,
+        layer_id: int,
+        tokenization_config: dict[str, Any],
+        prompt_variant: str = "default",
+    ):
+        """Save activations to cache."""
+        cache_key = self._generate_cache_key(split, layer_id, tokenization_config, prompt_variant)
+        cache_path = self._get_cache_path(cache_key)
+
+        cache_data = {
+            "activations": activations,
+            "labels": labels,
+            "metadata": {
+                "split": split,
+                "layer_id": layer_id,
+                "tokenization_config": tokenization_config,
+                "prompt_variant": prompt_variant,
+                "timestamp": datetime.now().isoformat(),
+                "shape": activations.shape,
+            },
+        }
+
+        with open(cache_path, "wb") as f:
+            pickle.dump(cache_data, f)
+
+        self.logger.info(f"Cached activations for {split} layer {layer_id}: {activations.shape}")
+
+    def load_activations(
+        self, split: str, layer_id: int, tokenization_config: dict[str, Any], prompt_variant: str = "default"
+    ) -> tuple[np.ndarray, np.ndarray]:
+        """Load activations from cache."""
+        cache_key = self._generate_cache_key(split, layer_id, tokenization_config, prompt_variant)
+        cache_path = self._get_cache_path(cache_key)
+
+        if not cache_path.exists():
+            raise FileNotFoundError(f"No cached activations found for key: {cache_key}")
+
+        with open(cache_path, "rb") as f:
+            cache_data = pickle.load(f)
+
+        self.logger.info(f"Loaded cached activations for {split} layer {layer_id}: {cache_data['activations'].shape}")
+        return cache_data["activations"], cache_data["labels"]
+
+
+class OptimizationPipeline:
+    """Main optimization pipeline using Optuna for hyperparameter search."""
+
+    def __init__(self, config: OptimizationConfig):
+        self.config = config
+        self.device = resolve_device(config.device)
+        self.logger = logging.getLogger(f"{__name__}.{self.__class__.__name__}")
+
+        # Setup output directories
+        self.output_dir = Path(config.output_dir)
+        self.output_dir.mkdir(parents=True, exist_ok=True)
+
+        # Initialize cache
+        self.cache = ActivationCache(config.cache_dir)
+
+        # Initialize WandB if configured
+        self.wandb_run = None
+        if config.use_wandb:
+            if not WANDB_AVAILABLE:
+                raise ImportError(
+                    "WandB integration enabled but wandb is not installed. Install with: pip install wandb"
+                )
+            self._init_wandb()
+
+        self.model = None
+        self.tokenizer = None
+        self.train_samples = None
+        self.val_samples = None
+        self.test_samples = None
+        # Store task documents for BigCode evaluation
+        self.train_task_docs = None
+        self.val_task_docs = None
+        self.test_task_docs = None
+        self.tokenization_config = {
+            "max_length": config.max_length,
+            "padding": True,
+            "truncation": True,
+            "return_tensors": "pt",
+        }
+
+    @property
+    def is_coding_task(self) -> bool:
+        """Check if the current task requires code execution evaluation."""
+        from ...parameters.task_config import CODING_TASKS
+        from ..bigcode_integration import is_bigcode_task
+
+        val_dataset = getattr(self.config, "val_dataset", None)
+        if not val_dataset:
+            return False
+
+        return val_dataset.lower() in CODING_TASKS or is_bigcode_task(val_dataset)
+
+    def run_optimization(self) -> dict[str, Any]:
+        """Run the complete optimization pipeline."""
+        self.logger.info("=" * 80)
+        self.logger.info("🚀 STARTING OPTIMIZATION PIPELINE WITH OPTUNA")
+        self.logger.info("=" * 80)
+
+        # Create timestamped run directory
+        self.run_timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        self.run_dir = self.output_dir / f"run_{self.run_timestamp}"
+        self.run_dir.mkdir(parents=True, exist_ok=True)
+        self.logger.info(f"📁 Run directory: {self.run_dir}")
+
+        self._setup_experiment()
+        study = self._create_optuna_study()
+        study.optimize(self._objective_function, n_trials=self.config.n_trials)
+        best_trial = study.best_trial
+        final_results = self._final_evaluation(best_trial)
+        self._save_reproducibility_bundle(study, final_results)
+
+        # Log final results to WandB
+        self._log_final_results_to_wandb(study, final_results)
+
+        self.logger.info("✅ Optimization completed successfully!")
+        return final_results
+
+    def _setup_experiment(self):
+        """Setup model, tokenizer, and load datasets."""
+        self.logger.info("📊 Setting up experiment...")
+
+        # Load model and tokenizer with memory optimizations
+        from transformers import AutoModelForCausalLM, AutoTokenizer
+
+        self.tokenizer = AutoTokenizer.from_pretrained(self.config.model_name)
+
+        # Load model with memory optimizations (same as comprehensive evaluation)
+        self.model = AutoModelForCausalLM.from_pretrained(
+            self.config.model_name,
+            torch_dtype=preferred_dtype(self.device.type),
+            low_cpu_mem_usage=True,
+        )
+
+        self.model.to(self.device)
+        self.model.eval()  # Set to evaluation mode
+
+        if self.tokenizer.pad_token is None:
+            self.tokenizer.pad_token = self.tokenizer.eos_token
+
+        # Set left padding for decoder-only models (same as comprehensive evaluation)
+        self.tokenizer.padding_side = "left"
+
+        # Load datasets
+        self.train_samples = data_utils.load_dataset_samples(self.config.train_dataset, self.config.train_limit)
+        self.val_samples = data_utils.load_dataset_samples(self.config.val_dataset, self.config.val_limit)
+        self.test_samples = data_utils.load_dataset_samples(self.config.test_dataset, self.config.test_limit)
+
+        # Store task documents for BigCode evaluation (coding tasks)
+        self.train_task_docs = self.train_samples
+        self.val_task_docs = self.val_samples
+        self.test_task_docs = self.test_samples
+
+        self.logger.info(
+            f"Loaded {len(self.train_samples)} train, {len(self.val_samples)} val, {len(self.test_samples)} test samples"
+        )
+
+        # Pre-cache activations for all layers on all splits
+        self._precache_activations()
+
+    def _precache_activations(self):
+        """Pre-cache activations for all layers and splits to improve efficiency."""
+        self.logger.info("🔄 Pre-caching activations for efficiency...")
+
+        layer_range = range(self.config.layer_search_range[0], self.config.layer_search_range[1] + 1)
+
+        splits_data = [("train", self.train_samples), ("val", self.val_samples), ("test", self.test_samples)]
+
+        for split_name, samples in splits_data:
+            for layer_id in layer_range:
+                if not self.cache.has_cached_activations(split_name, layer_id, self.tokenization_config):
+                    self.logger.info(f"Caching activations for {split_name} split, layer {layer_id}")
+
+                    dataset_name = {
+                        "train": self.config.train_dataset,
+                        "val": self.config.val_dataset,
+                        "test": self.config.test_dataset,
+                    }[split_name]
+
+                    activations, labels = self._create_probe_data(samples, layer_id, dataset_name)
+
+                    self.cache.save_activations(activations, labels, split_name, layer_id, self.tokenization_config)
+                else:
+                    self.logger.info(f"Activations already cached for {split_name} split, layer {layer_id}")
+
+    def _create_probe_data(
+        self, samples: list[dict], layer_id: int, dataset_name: str
+    ) -> tuple[np.ndarray, np.ndarray]:
+        """Create contrastive probe training data for a specific layer."""
+        self.logger.info(f"Creating probe data from {len(samples)} samples for {dataset_name} on layer {layer_id}")
+
+        # Get task for the specified dataset
+        task = get_task(dataset_name)
+        extractor = task.get_extractor()
+        self.logger.debug(f"Using task: {task.__class__.__name__}, extractor: {extractor.__class__.__name__}")
+
+        texts = []
+        labels = []
+        success_count = 0
+        fail_count = 0
+
+        for i, sample in enumerate(samples):
+            try:
+                # Extract QA pair
+                contrastive_pair = extractor.extract_contrastive_pair(sample, task)
+
+                # Skip samples where contrastive pair extraction failed
+                if not contrastive_pair:
+                    self.logger.debug(f"Sample {i + 1}: No contrastive pair extracted from keys: {list(sample.keys())}")
+                    fail_count += 1
+                    continue
+
+                success_count += 1
+                self.logger.debug(f"Sample {i + 1}: Successfully extracted contrastive pair")
+
+            except Exception as e:
+                self.logger.error(f"Sample {i + 1}: Exception during contrastive pair extraction: {e}")
+                fail_count += 1
+                continue
+
+            question = contrastive_pair["question"]
+            correct_answer = contrastive_pair["correct_answer"]
+            incorrect_answer = contrastive_pair["incorrect_answer"]
+
+            # Log contrastive pair details
+            self.logger.debug(f"Contrastive pair - Question: ...{question[-50:]}")
+            self.logger.debug(f"Contrastive pair - Correct: {correct_answer}, Incorrect: {incorrect_answer}")
+
+            correct_text = f"{question} {correct_answer}"
+            texts.append(correct_text)
+            labels.append(1)
+
+            incorrect_text = f"{question} {incorrect_answer}"
+            texts.append(incorrect_text)
+            labels.append(0)
+
+        self.logger.info(
+            f"Probe data creation: {success_count} successful, {fail_count} failed. Generated {len(texts)} texts."
+        )
+
+        if len(texts) == 0:
+            self.logger.error("No texts generated for activation extraction! All contrastive pair extractions failed.")
+            return np.array([]), np.array([])
+
+        activations = data_utils.extract_activations_with_hook(
+            self.model, self.tokenizer, texts, layer_id, self.config.batch_size, self.config.max_length, self.device
+        )
+
+        return activations, np.array(labels)
+
+    def _create_optuna_study(self) -> optuna.Study:
+        """Create Optuna study with SQLite persistence and specified sampler/pruner."""
+        self.logger.info("📋 Creating Optuna study with SQLite persistence...")
+        self.logger.info(f"Database: {self.config.db_url}")
+        self.logger.info(f"Study name: {self.config.study_name}")
+        self.logger.info(f"🎲 Warmup: {self.config.n_startup_trials} random trials before TPE sampling")
+
+        # Setup sampler
+        if self.config.sampler == "TPE":
+            sampler = TPESampler(seed=self.config.seed, n_startup_trials=self.config.n_startup_trials)
+        elif self.config.sampler == "Random":
+            sampler = optuna.samplers.RandomSampler(seed=self.config.seed)
+        else:
+            sampler = TPESampler(seed=self.config.seed, n_startup_trials=self.config.n_startup_trials)
+
+        # Setup pruner
+        if self.config.pruner == "MedianPruner":
+            pruner = MedianPruner(n_startup_trials=5, n_warmup_steps=10)
+        elif self.config.pruner == "SuccessiveHalvingPruner":
+            pruner = SuccessiveHalvingPruner()
+        else:
+            pruner = MedianPruner(n_startup_trials=5, n_warmup_steps=10)
+
+        # Create study with SQLite storage
+        study = optuna.create_study(
+            study_name=self.config.study_name,
+            storage=self.config.db_url,
+            direction="maximize",  # Maximize validation accuracy
+            sampler=sampler,
+            pruner=pruner,
+            load_if_exists=True,  # Continue existing study if it exists
+        )
+
+        self.logger.info(f"Study created/loaded with {len(study.trials)} existing trials")
+
+        return study
+
+    def _save_steering_vector_dual_format(self, steering_instance, pt_path: Path, safetensors_path: Path) -> bool:
+        """Save steering vector in both .pt and safetensors formats."""
+        # Save in original .pt format first (preserves all metadata)
+        if not steering_instance.save_steering_vector(str(pt_path)):
+            self.logger.warning("Failed to save steering vector - method may not be trained")
+            return False
+
+        self.logger.info(f"💾 Saved best steering vector to: {pt_path.name}")
+
+        # Also save in safetensors format for HuggingFace compatibility
+        try:
+            # Load the .pt file and extract steering vector
+            data = torch.load(str(pt_path), map_location="cpu", weights_only=False)
+            if isinstance(data, dict) and "steering_vector" in data:
+                # Save just the steering vector in safetensors format
+                safetensors_save({"steering_vector": data["steering_vector"]}, str(safetensors_path))
+                self.logger.info(f"💾 Also saved as safetensors: {safetensors_path.name}")
+                return True
+            self.logger.warning("Unexpected .pt file structure, safetensors conversion skipped")
+            return True  # .pt save was successful
+        except Exception as e:
+            self.logger.warning(f"Could not create safetensors version: {e}")
+            return True  # .pt save was successful
+
+    def _objective_function(self, trial: optuna.Trial) -> float:
+        """Optuna objective function for hyperparameter optimization."""
+        try:
+            # Sample hyperparameters
+            layer_id = trial.suggest_int(
+                "layer_id", self.config.layer_search_range[0], self.config.layer_search_range[1]
+            )
+
+            # Fixed probe type and regularization
+            probe_type = self.config.probe_type  # Always logistic_regression
+            probe_c = 1.0  # Default regularization strength
+
+            steering_method = trial.suggest_categorical("steering_method", self.config.steering_methods)
+
+            if steering_method == "dac":
+                steering_alpha = trial.suggest_float("steering_alpha", 0.1, 5.0)
+                entropy_threshold = trial.suggest_float("entropy_threshold", 0.5, 2.0)
+                ptop = trial.suggest_float("ptop", 0.2, 0.8)
+                max_alpha = trial.suggest_float("max_alpha", 1.0, 5.0)
+            elif steering_method == "caa":
+                steering_alpha = trial.suggest_float("steering_alpha", 0.1, 5.0)
+
+            probe_score = self._train_and_evaluate_probe(trial, layer_id, probe_type, probe_c)
+
+            # Don't prune based on probe score - focus optimization on steering parameters
+
+            # Build clean hyperparameters dictionary
+            if steering_method == "dac":
+                hyperparams = {
+                    "steering_alpha": steering_alpha,
+                    "entropy_threshold": entropy_threshold,
+                    "ptop": ptop,
+                    "max_alpha": max_alpha,
+                }
+            elif steering_method == "caa":
+                hyperparams = {
+                    "steering_alpha": steering_alpha,
+                }
+            else:
+                raise ValueError(f"Unsupported steering method: {steering_method}")
+
+            steering_method_instance = self._train_steering_method(trial, steering_method, layer_id, hyperparams)
+
+            validation_accuracy = self._evaluate_steering_on_validation(
+                steering_method_instance, steering_method, layer_id, hyperparams, trial.number, trial
+            )
+
+            trial.report(validation_accuracy, step=1)
+
+            # Log to WandB
+            metrics = {"validation_accuracy": validation_accuracy, "probe_score": probe_score}
+            self._log_trial_to_wandb(trial, metrics)
+
+            return validation_accuracy
+
+        except Exception as e:
+            self.logger.error(f"Trial failed: {e}")
+            return 0.0
+
+    def _train_and_evaluate_probe(self, trial: optuna.Trial, layer_id: int, probe_type: str, probe_c: float) -> float:
+        """Train probe on training data and evaluate on validation data using cached activations."""
+        # Load cached training activations
+        X_train, y_train = self.cache.load_activations("train", layer_id, self.tokenization_config)
+
+        # Train probe
+        if probe_type == "logistic_regression":
+            from sklearn.linear_model import LogisticRegression
+
+            probe = LogisticRegression(C=probe_c, random_state=self.config.seed, max_iter=1000)
+            probe.fit(X_train, y_train)
+        else:
+            raise ValueError(f"Unsupported probe type: {probe_type}")
+
+        # Evaluate on validation data using cached activations
+        X_val, y_val = self.cache.load_activations("val", layer_id, self.tokenization_config)
+
+        from sklearn.metrics import roc_auc_score
+
+        y_pred_proba = probe.predict_proba(X_val)[:, 1]
+        return roc_auc_score(y_val, y_pred_proba) if len(np.unique(y_val)) > 1 else 0.5
+
+        # Don't store the probe object - it can't be JSON serialized
+        # The probe will be retrained in the final evaluation if needed
+
+    def _train_steering_method(
+        self, trial: optuna.Trial, method_name: str, layer_id: int, hyperparams: dict[str, Any]
+    ) -> Any:
+        """Train steering method on training data."""
+        # Use contrastive_pairs_limit with bounds checking
+        contrastive_limit = min(self.config.contrastive_pairs_limit, len(self.train_samples))
+        contrastive_pairs = self._create_contrastive_pairs(
+            self.train_samples, layer_id, self.config.train_dataset, limit=contrastive_limit
+        )
+
+        if method_name == "dac":
+            # Create DAC instance
+            dac = DAC(
+                entropy_threshold=hyperparams["entropy_threshold"],
+                ptop=hyperparams["ptop"],
+                max_alpha=hyperparams["max_alpha"],
+            )
+
+            # Train DAC
+            dac.train(contrastive_pairs, layer_id)
+            return dac
+
+        if method_name == "caa":
+            # Create CAA instance
+            from wisent_guard.core.steering_methods.caa import CAA
+
+            caa = CAA(device=self.device)
+
+            # Train CAA
+            caa.train(contrastive_pairs, layer_id)
+            return caa
+
+        raise ValueError(f"Unsupported steering method: {method_name}")
+
+    def _create_contrastive_pairs(
+        self, samples: list[dict], layer_id: int, dataset_name: str, limit: Optional[int] = None
+    ) -> ContrastivePairSet:
+        """Create contrastive pairs with activations for steering training."""
+        contrastive_pairs = []
+        task = get_task(dataset_name)
+        extractor = task.get_extractor()
+
+        samples_to_use = samples[:limit] if limit else samples
+
+        for sample in samples_to_use:
+            contrastive_pair = extractor.extract_contrastive_pair(sample, task)
+            if contrastive_pair:
+                # Log contrastive pair details
+                self.logger.debug(f"Creating contrastive pair - Question: ...{contrastive_pair['question'][-50:]}")
+                self.logger.debug(
+                    f"Creating contrastive pair - Correct: {contrastive_pair['correct_answer']}, Incorrect: {contrastive_pair['incorrect_answer']}"
+                )
+
+                positive_response = Response(text=contrastive_pair["correct_answer"], label=1)
+                negative_response = Response(text=contrastive_pair["incorrect_answer"], label=0)
+
+                pair = ContrastivePair(
+                    prompt=contrastive_pair["question"],
+                    positive_response=positive_response,
+                    negative_response=negative_response,
+                )
+                contrastive_pairs.append(pair)
+
+        pair_set = ContrastivePairSet(name=f"{dataset_name}_training", pairs=contrastive_pairs)
+
+        # Extract activations for all pairs in batches
+        if pair_set.pairs:
+            all_texts = []
+            text_to_pair_mapping = []
+
+            for pair_idx, pair in enumerate(pair_set.pairs):
+                pos_text = f"{pair.prompt} {pair.positive_response.text}"
+                neg_text = f"{pair.prompt} {pair.negative_response.text}"
+
+                all_texts.extend([pos_text, neg_text])
+                text_to_pair_mapping.extend([(pair_idx, "positive"), (pair_idx, "negative")])
+
+            all_activations = self._extract_batch_activations(all_texts, layer_id)
+
+            for text_idx, (pair_idx, response_type) in enumerate(text_to_pair_mapping):
+                activation = all_activations[text_idx]
+
+                if response_type == "positive":
+                    pair_set.pairs[pair_idx].positive_response.activations = activation
+                else:
+                    pair_set.pairs[pair_idx].negative_response.activations = activation
+
+        return pair_set
+
+    def _extract_batch_activations(self, texts: list[str], layer_id: int) -> list[torch.Tensor]:
+        """Extract activations for multiple texts in batches."""
+        if not texts:
+            return []
+
+        all_activations = []
+        batch_size = self.config.batch_size
+
+        for i in range(0, len(texts), batch_size):
+            batch_texts = texts[i : i + batch_size]
+
+            inputs = self.tokenizer(
+                batch_texts, return_tensors="pt", padding=True, truncation=True, max_length=self.config.max_length
+            ).to(self.device)
+
+            batch_activations = []
+
+            def batch_hook_fn(module, input, output):
+                with torch.no_grad():
+                    hidden_states = output[0] if isinstance(output, tuple) else output
+
+                    last_token_acts = hidden_states[:, -1, :].detach().clone()
+                    batch_activations.append(last_token_acts)
+
+            if hasattr(self.model, "transformer"):
+                target_layer = self.model.transformer.h[layer_id]
+            elif hasattr(self.model, "model"):
+                target_layer = self.model.model.layers[layer_id]
+            else:
+                raise ValueError("Unknown model architecture")
+
+            handle = target_layer.register_forward_hook(batch_hook_fn)
+
+            try:
+                with torch.no_grad():
+                    _ = self.model(**inputs)
+            finally:
+                handle.remove()
+                empty_device_cache(self.device.type)
+
+            if batch_activations:
+                batch_tensor = batch_activations[0]
+                for j in range(batch_tensor.shape[0]):
+                    all_activations.append(batch_tensor[j].unsqueeze(0))
+
+        return all_activations
+
+    def _extract_single_activation(self, text: str, layer_id: int) -> torch.Tensor:
+        """Extract activation for a single text."""
+        activations = self._extract_batch_activations([text], layer_id)
+        return activations[0] if activations else torch.zeros(1, self.model.config.hidden_size, device=self.device)
+
+    def _evaluate_steering_on_validation(
+        self,
+        steering_instance: Any,
+        method_name: str,
+        layer_id: int,
+        hyperparams: dict[str, Any],
+        trial_number: int = 0,
+        trial=None,
+    ) -> float:
+        """Evaluate steering method on validation data by re-running forward passes."""
+        if steering_instance is None:
+            return 0.0
+
+        # Generate predictions with steering applied
+        predictions = []
+        ground_truths = []
+        task_docs = []  # Preserve original task documents for BigCode evaluation
+
+        task = get_task(self.config.val_dataset)
+        extractor = task.get_extractor()
+
+        # Collect all questions for batched processing (use ALL validation samples)
+        questions = []
+        ground_truths = []
+        valid_samples = []  # Keep track of samples that produce valid QA pairs
+
+        for sample in tqdm(
+            self.val_samples, desc="Extracting validation QA pairs", leave=False
+        ):  # Use all validation samples for reliable evaluation
+            qa_pair = extractor.extract_qa_pair(sample, task)
+            if not qa_pair:
+                continue
+
+            question = qa_pair["formatted_question"]
+            ground_truth = qa_pair["correct_answer"]
+            questions.append(question)
+            ground_truths.append(ground_truth)
+            valid_samples.append(sample)  # Store the original sample
+
+        # Generate predictions using batched approach
+        if questions:
+            if steering_instance is None:
+                predictions = self._generate_baseline_batched(questions)
+            else:
+                # Extract the appropriate strength parameter based on method
+                if method_name == "dac":
+                    # DAC uses steering_alpha as base strength multiplier
+                    strength = hyperparams.get("steering_alpha", 1.0)
+                else:
+                    # CAA and other methods use steering_alpha directly
+                    strength = hyperparams["steering_alpha"]
+
+                predictions = self._generate_with_steering_batched(steering_instance, questions, strength, layer_id)
+
+            # Log sample predictions for debugging
+            for i, (pred, gt) in enumerate(zip(predictions[:3], ground_truths[:3])):
+                self.logger.debug(f"{method_name.upper()} Sample {i} - Model: ...{pred[-50:] if pred else 'None'}")
+                self.logger.debug(f"{method_name.upper()} Sample {i} - Ground truth: {gt}")
+        else:
+            predictions = []
+
+        if not predictions:
+            return 0.0
+
+        # Save detailed validation results to JSON
+        self._save_detailed_validation_results(
+            questions,
+            ground_truths,
+            predictions,
+            trial_number,
+            trial=trial,
+            steering_method=method_name,
+            layer_id=layer_id,
+            hyperparams=hyperparams,
+        )
+
+        # Prepare task docs for BigCode evaluation (if coding task)
+        task_docs = valid_samples[: len(predictions)] if valid_samples else []
+
+        # Evaluate benchmark performance (with task docs for coding tasks)
+        benchmark_metrics = metrics.evaluate_benchmark_performance(
+            predictions, ground_truths, self.config.val_dataset, task_docs=task_docs
+        )
+
+        return benchmark_metrics.get("accuracy", 0.0)
+
+    def _generate_with_dac_steering(self, dac: DAC, question: str, alpha: float, layer_id: int) -> str:
+        """Generate response with DAC steering applied."""
+        # Use the general steering method which calls DAC's apply_steering
+        return self._generate_with_steering(dac, question, alpha, layer_id)
+
+    def _generate_with_caa_steering(self, caa, question: str, alpha: float, layer_id: int) -> str:
+        """Generate response with CAA steering applied."""
+        if not hasattr(caa, "steering_vector") or caa.steering_vector is None:
+            return self._generate_baseline(question)
+
+        return self._generate_with_steering_hook(question, caa.steering_vector, layer_id, alpha)
+
+    def _generate_with_steering_hook(
+        self, question: str, steering_vector: torch.Tensor, layer_id: int, alpha: float
+    ) -> str:
+        """Generate response with steering vector applied via hook (re-runs forward pass)."""
+        inputs = self.tokenizer(question, return_tensors="pt").to(self.device)
+
+        def steering_hook(module, input, output):
+            """Hook that applies steering vector during forward pass."""
+            if isinstance(output, tuple):
+                hidden_states = output[0]
+                # Apply steering to the last token
+                hidden_states[:, -1, :] += alpha * steering_vector.to(hidden_states.device)
+                return (hidden_states, *output[1:])
+            hidden_states = output
+            hidden_states[:, -1, :] += alpha * steering_vector.to(hidden_states.device)
+            return hidden_states
+
+        # Register hook on target layer
+        if hasattr(self.model, "transformer"):
+            target_layer = self.model.transformer.h[layer_id]
+        elif hasattr(self.model, "model"):
+            target_layer = self.model.model.layers[layer_id]
+        else:
+            raise ValueError("Unknown model architecture")
+
+        handle = target_layer.register_forward_hook(steering_hook)
+
+        try:
+            with torch.no_grad():
+                outputs = self.model.generate(
+                    **inputs,
+                    max_new_tokens=self.config.max_new_tokens,
+                    do_sample=self.config.do_sample,
+                    temperature=self.config.temperature if self.config.do_sample else 1.0,
+                    pad_token_id=self.tokenizer.eos_token_id,
+                    eos_token_id=self.tokenizer.eos_token_id,
+                )
+        finally:
+            handle.remove()
+
+        response = self.tokenizer.decode(outputs[0][inputs.input_ids.shape[1] :], skip_special_tokens=True)
+        return response.strip()
+
+    def _generate_baseline(self, question: str) -> str:
+        """Generate baseline response without steering."""
+        inputs = self.tokenizer(question, return_tensors="pt").to(self.device)
+
+        with torch.no_grad():
+            outputs = self.model.generate(
+                **inputs,
+                max_new_tokens=self.config.max_new_tokens,
+                do_sample=self.config.do_sample,
+                temperature=self.config.temperature if self.config.do_sample else 1.0,
+                pad_token_id=self.tokenizer.eos_token_id,
+                eos_token_id=self.tokenizer.eos_token_id,
+            )
+
+        response = self.tokenizer.decode(outputs[0][inputs.input_ids.shape[1] :], skip_special_tokens=True)
+        return response.strip()
+
+    def _generate_baseline_batched(self, questions: list[str]) -> list[str]:  # TODO
+        """Generate baseline responses in batches without steering."""
+        if not questions:
+            return []
+
+        batch_size = self.config.batch_size
+        all_responses = []
+
+        # Process questions in batches
+        for i in tqdm(range(0, len(questions), batch_size), desc="Generating baseline predictions", leave=False):
+            batch_questions = questions[i : i + batch_size]
+
+            # Batch tokenization with padding
+            inputs = self.tokenizer(
+                batch_questions, return_tensors="pt", padding=True, truncation=True, max_length=self.config.max_length
+            ).to(self.device)
+
+            with torch.no_grad():
+                outputs = self.model.generate(
+                    **inputs,
+                    max_new_tokens=self.config.max_new_tokens,
+                    do_sample=self.config.do_sample,
+                    temperature=self.config.temperature if self.config.do_sample else 1.0,
+                    pad_token_id=self.tokenizer.eos_token_id,
+                    eos_token_id=self.tokenizer.eos_token_id,
+                )
+
+            # Decode responses
+            batch_responses = []
+            for j, output in enumerate(outputs):
+                input_length = inputs.input_ids[j].shape[0]
+                response = self.tokenizer.decode(output[input_length:], skip_special_tokens=True)
+                batch_responses.append(response.strip())
+
+            all_responses.extend(batch_responses)
+
+        return all_responses
+
+    def _generate_with_steering_batched(
+        self, steering_instance: Any, questions: list[str], alpha: float, layer_id: int
+    ) -> list[str]:
+        """Generate responses with steering applied in batches using apply_steering()."""
+        if not questions:
+            return []
+
+        batch_size = self.config.batch_size
+        all_responses = []
+
+        # Process questions in batches
+        for i in tqdm(range(0, len(questions), batch_size), desc="Generating steered predictions", leave=False):
+            batch_questions = questions[i : i + batch_size]
+
+            # Batch tokenization with padding
+            inputs = self.tokenizer(
+                batch_questions, return_tensors="pt", padding=True, truncation=True, max_length=self.config.max_length
+            ).to(self.device)
+
+            def steering_hook(module, input, output):
+                """Hook that applies steering using the steering method's apply_steering()."""
+                hidden_states = output[0] if isinstance(output, tuple) else output
+
+                # Apply steering using the method's apply_steering() function
+                steered = steering_instance.apply_steering(hidden_states, strength=alpha)
+
+                if isinstance(output, tuple):
+                    return (steered, *output[1:])
+                return steered
+
+            # Register hook on target layer
+            if hasattr(self.model, "transformer"):
+                if layer_id >= len(self.model.transformer.h):
+                    raise ValueError(f"layer_id {layer_id} exceeds model layers")
+                target_layer = self.model.transformer.h[layer_id]
+            elif hasattr(self.model, "model"):
+                if layer_id >= len(self.model.model.layers):
+                    raise ValueError(f"layer_id {layer_id} exceeds model layers")
+                target_layer = self.model.model.layers[layer_id]
+            else:
+                raise ValueError("Unknown model architecture")
+
+            handle = target_layer.register_forward_hook(steering_hook)
+
+            try:
+                with torch.no_grad():
+                    outputs = self.model.generate(
+                        **inputs,
+                        max_new_tokens=self.config.max_new_tokens,
+                        do_sample=self.config.do_sample,
+                        temperature=self.config.temperature if self.config.do_sample else 1.0,
+                        pad_token_id=self.tokenizer.eos_token_id,
+                        eos_token_id=self.tokenizer.eos_token_id,
+                    )
+
+                # Decode responses
+                batch_responses = []
+                for j, output in enumerate(outputs):
+                    input_length = inputs.input_ids[j].shape[0]
+                    response = self.tokenizer.decode(output[input_length:], skip_special_tokens=True)
+                    batch_responses.append(response.strip())
+
+                all_responses.extend(batch_responses)
+
+            finally:
+                # Always remove the hook
+                handle.remove()
+
+        return all_responses
+
+    def _final_evaluation(self, best_trial: optuna.Trial) -> dict[str, Any]:
+        """Run final evaluation on test split with best configuration."""
+        self.logger.info("🏆 Running final evaluation with best configuration...")
+
+        # Extract best hyperparameters
+        # Handle both real trials and FixedTrial objects
+        if hasattr(best_trial, "params") and best_trial.params:
+            best_params = best_trial.params
+        elif hasattr(best_trial, "_params"):
+            best_params = best_trial._params
+        else:
+            # Fallback - this shouldn't happen
+            raise ValueError("Cannot access trial parameters")
+        layer_id = best_params["layer_id"]
+
+        self.logger.info(f"Best configuration: {best_params}")
+
+        # Re-train best probe and steering method on training data
+        from sklearn.linear_model import LogisticRegression
+
+        # Train best probe with fixed probe_c
+        X_train, y_train = self.cache.load_activations("train", layer_id, self.tokenization_config)
+        probe = LogisticRegression(C=1.0, random_state=self.config.seed, max_iter=1000)  # Fixed probe_c
+        probe.fit(X_train, y_train)
+
+        # Train best steering method
+        steering_method = best_params.get("steering_method", "caa")  # Default to CAA if missing
+        steering_instance = self._train_steering_method(best_trial, steering_method, layer_id, best_params)
+
+        # Save the best steering vector in both formats
+        if steering_instance and hasattr(steering_instance, "save_steering_vector"):
+            pt_path = self.run_dir / "best_steering_vector.pt"
+            safetensors_path = self.run_dir / "best_steering_vector.safetensors"
+            self._save_steering_vector_dual_format(steering_instance, pt_path, safetensors_path)
+
+        # Generate baseline predictions (no steering)
+        self.logger.info("Generating baseline predictions...")
+        baseline_predictions, test_ground_truths, test_questions, test_task_docs = self._generate_test_predictions(
+            None, None, layer_id, 0.0
+        )
+
+        # Generate steered predictions
+        self.logger.info("Generating steered predictions...")
+
+        # Extract the appropriate strength parameter based on method and available parameters
+        method_name = best_params.get("steering_method", "caa")  # Default to CAA if missing
+        if method_name == "dac":
+            # DAC can use base_strength or steering_alpha, with fallback to 1.0
+            strength = best_params.get("base_strength", best_params.get("steering_alpha", 1.0))
+        else:
+            # CAA and other methods use steering_alpha
+            strength = best_params["steering_alpha"]
+
+        steered_predictions, _, _, _ = self._generate_test_predictions(
+            steering_instance, method_name, layer_id, strength
+        )
+
+        # Save detailed test results to JSON
+        if test_questions and test_ground_truths and baseline_predictions and steered_predictions:
+            self._save_detailed_test_results(
+                test_questions,
+                test_ground_truths,
+                baseline_predictions,
+                steered_predictions,
+                best_trial=best_trial,
+                best_params=best_params,
+                layer_id=layer_id,
+                steering_method=method_name,
+            )
+
+        # Calculate benchmark metrics (with real task docs for coding tasks)
+        baseline_benchmark_metrics = metrics.evaluate_benchmark_performance(
+            baseline_predictions, test_ground_truths, self.config.test_dataset, task_docs=test_task_docs
+        )
+        steered_benchmark_metrics = metrics.evaluate_benchmark_performance(
+            steered_predictions, test_ground_truths, self.config.test_dataset, task_docs=test_task_docs
+        )
+
+        # Evaluate probe on test data
+        X_test, y_test = self.cache.load_activations("test", layer_id, self.tokenization_config)
+        test_probe_metrics = self._evaluate_probe_metrics(probe, X_test, y_test)
+
+        # Calculate improvement
+        accuracy_improvement = steered_benchmark_metrics.get("accuracy", 0.0) - baseline_benchmark_metrics.get(
+            "accuracy", 0.0
+        )
+
+        final_results = {
+            "best_trial_params": best_params,
+            "best_validation_score": getattr(best_trial, "value", None),
+            "baseline_benchmark_metrics": baseline_benchmark_metrics,
+            "steered_benchmark_metrics": steered_benchmark_metrics,
+            "accuracy_improvement": accuracy_improvement,
+            "test_probe_metrics": test_probe_metrics,
+            "config": self.config.to_dict(),
+            "num_test_samples": len(test_ground_truths),
+        }
+
+        # Log final results
+        self.logger.info("=" * 60)
+        self.logger.info("🏆 FINAL TEST RESULTS")
+        self.logger.info("=" * 60)
+        self.logger.info(f"Baseline accuracy: {baseline_benchmark_metrics.get('accuracy', 0.0):.4f}")
+        self.logger.info(f"Steered accuracy: {steered_benchmark_metrics.get('accuracy', 0.0):.4f}")
+        self.logger.info(f"Improvement: {accuracy_improvement:+.4f}")
+        self.logger.info(f"Probe AUC: {test_probe_metrics.get('auc', 0.5):.4f}")
+        self.logger.info(f"Test samples: {len(test_ground_truths)}")
+        self.logger.info("=" * 60)
+
+        return final_results
+
+    def _generate_test_predictions(
+        self, steering_instance: Any, method_name: str, layer_id: int, alpha: float
+    ) -> tuple[list[str], list[str], list[str], list[dict]]:
+        """Generate predictions on test data using batched generation."""
+        # Collect all questions and ground truths for batching
+        questions = []
+        ground_truths = []
+        valid_samples = []  # Keep track of samples that produce valid QA pairs
+
+        task = get_task(self.config.test_dataset)
+        extractor = task.get_extractor()
+
+        for sample in self.test_samples:
+            qa_pair = extractor.extract_qa_pair(sample, task)
+            if not qa_pair:
+                continue
+
+            question = qa_pair["formatted_question"]
+            ground_truth = qa_pair["correct_answer"]
+            questions.append(question)
+            ground_truths.append(ground_truth)
+            valid_samples.append(sample)  # Store the original sample
+
+        # Process all questions with appropriate batched method
+        if questions:
+            try:
+                if steering_instance is None:
+                    # Baseline generation - use batched method
+                    predictions = self._generate_baseline_batched(questions)
+                else:
+                    # Use unified batched generation with apply_steering()
+                    predictions = self._generate_with_steering_batched(steering_instance, questions, alpha, layer_id)
+
+                # Log sample predictions for debugging
+                for i, (pred, gt) in enumerate(zip(predictions[:3], ground_truths[:3])):
+                    self.logger.debug(f"Test Sample {i} - Model: ...{pred[-50:] if pred else 'None'}")
+                    self.logger.debug(f"Test Sample {i} - Ground truth: {gt}")
+
+            except Exception as e:
+                self.logger.warning(f"Batched generation failed for test: {e}")
+                predictions = ["Error"] * len(questions)
+        else:
+            predictions = []
+
+        return predictions, ground_truths, questions, valid_samples
+
+    def _evaluate_probe_metrics(self, probe, X_test: np.ndarray, y_test: np.ndarray) -> dict[str, float]:
+        """Evaluate probe metrics."""
+        from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score, roc_auc_score
+
+        y_pred = probe.predict(X_test)
+        y_pred_proba = probe.predict_proba(X_test)[:, 1]
+
+        return {
+            "accuracy": accuracy_score(y_test, y_pred),
+            "precision": precision_score(y_test, y_pred, zero_division=0),
+            "recall": recall_score(y_test, y_pred, zero_division=0),
+            "f1": f1_score(y_test, y_pred, zero_division=0),
+            "auc": roc_auc_score(y_test, y_pred_proba) if len(np.unique(y_test)) > 1 else 0.5,
+        }
+
+    def _create_experiment_metadata(
+        self, trial=None, steering_method: str = None, layer_id: int = None, hyperparams: dict = None
+    ):
+        """Create comprehensive experiment metadata for detailed results."""
+        import platform
+        from datetime import datetime
+
+        metadata = {
+            "trial_info": {
+                "trial_number": trial.number if trial else None,
+                "trial_params": dict(trial.params) if trial else {},
+                "trial_state": str(getattr(trial, "state", "RUNNING")) if trial else None,
+            },
+            "model_config": {
+                "model_name": self.config.model_name,
+                "device": self.config.device,
+                "is_coding_task": self.is_coding_task,
+            },
+            "dataset_config": {
+                "train_dataset": self.config.train_dataset,
+                "val_dataset": self.config.val_dataset,
+                "test_dataset": self.config.test_dataset,
+                "train_limit": self.config.train_limit,
+                "val_limit": self.config.val_limit,
+                "test_limit": self.config.test_limit,
+                "contrastive_pairs_limit": self.config.contrastive_pairs_limit,
+            },
+            "steering_config": {
+                "steering_method": steering_method,
+                "layer_id": layer_id,
+                "hyperparams": hyperparams or {},
+                "layer_search_range": self.config.layer_search_range,
+                "probe_type": self.config.probe_type,
+                "available_steering_methods": self.config.steering_methods,
+            },
+            "optimization_config": {
+                "study_name": self.config.study_name,
+                "sampler": self.config.sampler,
+                "pruner": self.config.pruner,
+                "n_trials": self.config.n_trials,
+                "n_startup_trials": self.config.n_startup_trials,
+            },
+            "generation_config": {
+                "batch_size": self.config.batch_size,
+                "max_length": self.config.max_length,
+                "max_new_tokens": self.config.max_new_tokens,
+                "temperature": self.config.temperature,
+                "do_sample": self.config.do_sample,
+            },
+            "run_info": {
+                "timestamp": datetime.now().isoformat(),
+                "run_dir": str(self.run_dir),
+                "output_dir": self.config.output_dir,
+                "cache_dir": self.config.cache_dir,
+                "platform": platform.platform(),
+                "python_version": platform.python_version(),
+            },
+            "wandb_config": {
+                "use_wandb": self.config.use_wandb,
+                "wandb_project": self.config.wandb_project,
+            }
+            if hasattr(self.config, "use_wandb")
+            else {},
+        }
+
+        return metadata
+
+    def _save_detailed_validation_results(
+        self,
+        questions: list[str],
+        ground_truths: list[str],
+        predictions: list[str],
+        trial_number: int,
+        trial=None,
+        steering_method: str = None,
+        layer_id: int = None,
+        hyperparams: dict = None,
+    ):
+        """Save detailed validation results to JSON file with experiment metadata."""
+        detailed_results = []
+
+        # For coding tasks, use the same BigCode evaluation as accuracy calculation
+        if self.is_coding_task:
+            # Use evaluate_benchmark_performance to get consistent BigCode evaluation
+            eval_results = metrics.evaluate_benchmark_performance(
+                predictions, ground_truths, task_name=self.config.val_dataset, task_docs=self.val_task_docs
+            )
+
+            # Extract individual correctness from evaluation details
+            eval_details = eval_results.get("evaluation_details", [])
+
+            for i, (question, correct_answer, model_answer) in enumerate(zip(questions, ground_truths, predictions)):
+                # Get correctness from BigCode evaluation if available
+                is_correct = eval_details[i]["is_correct"] if i < len(eval_details) else False
+
+                detailed_results.append(
+                    {
+                        "row": i,
+                        "question": question,
+                        "correct_answer": correct_answer,
+                        "model_answer": model_answer,
+                        "is_correct": is_correct,
+                        "evaluation_method": eval_results.get("evaluation_method", "unknown"),
+                        "extracted_code": eval_details[i].get("prediction", model_answer)
+                        if i < len(eval_details)
+                        else model_answer,
+                        "execution_error": eval_details[i].get("execution_error") if i < len(eval_details) else None,
+                    }
+                )
+        else:
+            # For non-coding tasks, process each result
+            for i, (question, correct_answer, model_answer) in enumerate(zip(questions, ground_truths, predictions)):
+                # Use standard evaluation via metrics module
+                is_correct = metrics.evaluate_response_correctness(
+                    model_answer, correct_answer, self.config.val_dataset
+                )
+
+                result_entry = {
+                    "row": i,
+                    "question": question,
+                    "correct_answer": correct_answer,
+                    "model_answer": model_answer,
+                    "is_correct": is_correct,
+                    "evaluation_method": "string_comparison",
+                }
+
+                # Add MC-specific fields if this is a multiple choice task
+                if self._should_use_multiple_choice_evaluation():
+                    # Extract MC diagnostics directly without custom evaluation
+                    import re
+
+                    # Extract available answers from question (A. choice, B. choice, etc.)
+                    available_answers = []
+                    choice_pattern = r"([A-E])\.\s+(.+?)(?=\n[A-E]\.|$)"
+                    matches = re.findall(choice_pattern, question, re.MULTILINE | re.DOTALL)
+                    for letter, choice_text in matches:
+                        available_answers.append(f"{letter}. {choice_text.strip()}")
+
+                    # Extract model's selected letter from model answer
+                    model_selected_letter = "?"
+                    model_letter_match = re.search(r"\b([A-E])\b", model_answer.upper())
+                    if model_letter_match:
+                        model_selected_letter = model_letter_match.group(1)
+
+                    result_entry["available_answers"] = available_answers
+                    result_entry["correct_choice_letter"] = correct_answer
+                    result_entry["model_selected_letter"] = model_selected_letter
+
+                detailed_results.append(result_entry)
+
+        # Create experiment metadata
+        experiment_metadata = self._create_experiment_metadata(trial, steering_method, layer_id, hyperparams)
+
+        # Create final results structure with metadata
+        final_results = {"experiment_metadata": experiment_metadata, "results": detailed_results}
+
+        # Save to JSON file
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        filename = f"validation_detailed_results_trial_{trial_number:03d}_{timestamp}.json"
+        filepath = self.run_dir / filename
+
+        with open(filepath, "w", encoding="utf-8") as f:
+            json.dump(final_results, f, indent=2, ensure_ascii=False)
+
+        self.logger.info(f"💾 Saved detailed validation results to: {filename}")
+
+    def _save_detailed_test_results(
+        self,
+        questions: list[str],
+        ground_truths: list[str],
+        baseline_predictions: list[str],
+        steered_predictions: list[str],
+        best_trial=None,
+        best_params: dict = None,
+        layer_id: int = None,
+        steering_method: str = None,
+    ):
+        """Save detailed test results to JSON file with both baseline and steered answers and experiment metadata."""
+        detailed_results = []
+
+        # For coding tasks, use BigCode evaluation consistently
+        if self.is_coding_task:
+            # Evaluate baseline predictions with BigCode
+            baseline_eval_results = metrics.evaluate_benchmark_performance(
+                baseline_predictions, ground_truths, task_name=self.config.test_dataset, task_docs=self.test_task_docs
+            )
+
+            # Evaluate steered predictions with BigCode
+            steered_eval_results = metrics.evaluate_benchmark_performance(
+                steered_predictions, ground_truths, task_name=self.config.test_dataset, task_docs=self.test_task_docs
+            )
+
+            baseline_details = baseline_eval_results.get("evaluation_details", [])
+            steered_details = steered_eval_results.get("evaluation_details", [])
+
+            for i, (question, correct_answer, baseline_answer, steered_answer) in enumerate(
+                zip(questions, ground_truths, baseline_predictions, steered_predictions)
+            ):
+                # Get correctness from BigCode evaluation
+                is_baseline_correct = baseline_details[i]["is_correct"] if i < len(baseline_details) else False
+                is_correct = steered_details[i]["is_correct"] if i < len(steered_details) else False
+
+                detailed_results.append(
+                    {
+                        "row": i,
+                        "question": question,
+                        "correct_answer": correct_answer,
+                        "baseline_model_answer": baseline_answer,
+                        "model_answer": steered_answer,
+                        "is_baseline_correct": is_baseline_correct,
+                        "is_correct": is_correct,
+                        "evaluation_method": steered_eval_results.get("evaluation_method", "bigcode_execution"),
+                        "baseline_extracted_code": baseline_details[i].get("prediction", baseline_answer)
+                        if i < len(baseline_details)
+                        else baseline_answer,
+                        "steered_extracted_code": steered_details[i].get("prediction", steered_answer)
+                        if i < len(steered_details)
+                        else steered_answer,
+                        "baseline_execution_error": baseline_details[i].get("execution_error")
+                        if i < len(baseline_details)
+                        else None,
+                        "steered_execution_error": steered_details[i].get("execution_error")
+                        if i < len(steered_details)
+                        else None,
+                    }
+                )
+        else:
+            # For non-coding tasks, process each result
+            for i, (question, correct_answer, baseline_answer, steered_answer) in enumerate(
+                zip(questions, ground_truths, baseline_predictions, steered_predictions)
+            ):
+                # Use standard evaluation for both baseline and steered answers
+                is_baseline_correct = metrics.evaluate_response_correctness(
+                    baseline_answer, correct_answer, self.config.test_dataset
+                )
+                is_correct = metrics.evaluate_response_correctness(
+                    steered_answer, correct_answer, self.config.test_dataset
+                )
+
+                result_entry = {
+                    "row": i,
+                    "question": question,
+                    "correct_answer": correct_answer,
+                    "baseline_model_answer": baseline_answer,
+                    "model_answer": steered_answer,
+                    "is_baseline_correct": is_baseline_correct,
+                    "is_correct": is_correct,
+                    "evaluation_method": "string_comparison",
+                }
+
+                # Add MC-specific fields if this is a multiple choice task
+                if self._should_use_multiple_choice_evaluation():
+                    # Extract MC diagnostics directly without custom evaluation
+                    import re
+
+                    # Extract available answers from question (A. choice, B. choice, etc.)
+                    available_answers = []
+                    choice_pattern = r"([A-E])\.\s+(.+?)(?=\n[A-E]\.|$)"
+                    matches = re.findall(choice_pattern, question, re.MULTILINE | re.DOTALL)
+                    for letter, choice_text in matches:
+                        available_answers.append(f"{letter}. {choice_text.strip()}")
+
+                    # Extract steered model's selected letter
+                    steered_selected_letter = "?"
+                    steered_letter_match = re.search(r"\b([A-E])\b", steered_answer.upper())
+                    if steered_letter_match:
+                        steered_selected_letter = steered_letter_match.group(1)
+
+                    # Extract baseline model's selected letter
+                    baseline_selected_letter = "?"
+                    baseline_letter_match = re.search(r"\b([A-E])\b", baseline_answer.upper())
+                    if baseline_letter_match:
+                        baseline_selected_letter = baseline_letter_match.group(1)
+
+                    result_entry["available_answers"] = available_answers
+                    result_entry["correct_choice_letter"] = correct_answer
+                    result_entry["model_selected_letter"] = steered_selected_letter
+                    result_entry["baseline_model_selected_letter"] = baseline_selected_letter
+
+                detailed_results.append(result_entry)
+
+        # Create experiment metadata for test results
+        experiment_metadata = self._create_experiment_metadata(
+            trial=best_trial,
+            steering_method=steering_method or best_params.get("steering_method") if best_params else None,
+            layer_id=layer_id,
+            hyperparams=best_params,
+        )
+
+        # Create final results structure with metadata
+        final_results = {"experiment_metadata": experiment_metadata, "results": detailed_results}
+
+        # Save to JSON file
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        filename = f"test_detailed_results_{timestamp}.json"
+        filepath = self.run_dir / filename
+
+        with open(filepath, "w", encoding="utf-8") as f:
+            json.dump(final_results, f, indent=2, ensure_ascii=False)
+
+        self.logger.info(f"💾 Saved detailed test results to: {filename}")
+        return filename
+
+    def _save_reproducibility_bundle(self, study: optuna.Study, final_results: dict[str, Any]):
+        """Save complete reproducibility bundle."""
+
+        # Save Optuna study
+        study_path = self.run_dir / f"optuna_study_{self.run_timestamp}.db"
+        study.study_name = str(study_path)
+
+        # Save configuration
+        config_path = self.run_dir / f"config_{self.run_timestamp}.json"
+        with open(config_path, "w") as f:
+            json.dump(self.config.to_dict(), f, indent=2)
+
+        # Save final results
+        results_path = self.run_dir / f"final_results_{self.run_timestamp}.json"
+        with open(results_path, "w") as f:
+            json.dump(final_results, f, indent=2, default=str)
+
+        # Save best configuration
+        best_config = {
+            "best_params": study.best_trial.params,
+            "best_value": study.best_trial.value,
+            "model_name": self.config.model_name,
+            "random_seed": self.config.seed,
+            "commit_hash": self._get_git_commit_hash(),
+            "timestamp": self.run_timestamp,
+        }
+
+        best_config_path = self.run_dir / f"best_configuration_{self.run_timestamp}.json"
+        with open(best_config_path, "w") as f:
+            json.dump(best_config, f, indent=2)
+
+        # Save study trials summary
+        trials_df = study.trials_dataframe()
+        trials_path = self.run_dir / f"study_trials_{self.run_timestamp}.csv"
+        trials_df.to_csv(trials_path, index=False)
+
+        self.logger.info(f"💾 Reproducibility bundle saved to: {self.run_dir}")
+        self.logger.info(f"📊 Study database: {study_path}")
+        self.logger.info(f"⚙️ Configuration: {config_path}")
+        self.logger.info(f"🏆 Results: {results_path}")
+        self.logger.info(f"🎯 Best config: {best_config_path}")
+
+        # Log steering vector if it exists (prefer safetensors format)
+        safetensors_path = self.run_dir / "best_steering_vector.safetensors"
+        pt_path = self.run_dir / "best_steering_vector.pt"
+
+        if safetensors_path.exists():
+            self.logger.info(f"🧭 Steering vector: {safetensors_path.name}")
+        elif pt_path.exists():
+            self.logger.info(f"🧭 Steering vector: {pt_path.name}")
+
+    def _get_git_commit_hash(self) -> Optional[str]:
+        """Get current git commit hash for reproducibility."""
+        try:
+            import subprocess
+
+            result = subprocess.run(["git", "rev-parse", "HEAD"], capture_output=True, text=True)
+            if result.returncode == 0:
+                return result.stdout.strip()
+        except:
+            pass
+        return None
+
+    def evaluate_only(self, best_params: dict[str, Any]) -> dict[str, Any]:
+        """Run evaluation only with provided parameters.
+
+        Args:
+            best_params: Dictionary of hyperparameters to use for evaluation
+
+        Returns:
+            Dictionary containing evaluation results
+        """
+        self.logger.info("🔬 Running evaluation-only mode with provided parameters")
+        self.logger.info(f"Parameters: {best_params}")
+
+        # Setup experiment if not already done
+        if self.model is None:
+            self._setup_experiment()
+
+        # Create timestamped run directory for evaluation-only mode
+        if not hasattr(self, "run_dir"):
+            self.run_timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+            self.run_dir = self.output_dir / f"evaluate_only_{self.run_timestamp}"
+            self.run_dir.mkdir(parents=True, exist_ok=True)
+            self.logger.info(f"📁 Evaluation directory: {self.run_dir}")
+
+        # Create a complete mock trial with all expected parameters
+        from optuna.trial import FixedTrial
+
+        # Ensure we have all required parameters for _final_evaluation
+        complete_params = {
+            "layer_id": best_params.get("layer_id", 15),
+            "probe_type": best_params.get("probe_type", "logistic_regression"),
+            "probe_c": best_params.get("probe_c", 1.0),
+            "steering_method": best_params.get("steering_method", "caa"),
+            "steering_alpha": best_params.get("steering_alpha", 0.5),
+        }
+
+        # Add method-specific parameters if needed
+        if complete_params["steering_method"] == "dac":
+            complete_params.update(
+                {
+                    "entropy_threshold": best_params.get("entropy_threshold", 1.5),
+                    "ptop": best_params.get("ptop", 0.5),
+                    "max_alpha": best_params.get("max_alpha", 2.0),
+                }
+            )
+
+        fixed_trial = FixedTrial(complete_params)
+
+        # Fix FixedTrial params access issue
+        if not hasattr(fixed_trial, "params"):
+            fixed_trial.params = complete_params
+
+        # Run final evaluation
+        return self._final_evaluation(fixed_trial)
+
+    @classmethod
+    def from_saved_study(
+        cls, study_path: str, config_path: Optional[str] = None, override_config: Optional[dict[str, Any]] = None
+    ):
+        """Create pipeline from saved study and optionally saved config.
+
+        Args:
+            study_path: Path to the SQLite study database
+            config_path: Optional path to saved configuration JSON
+            override_config: Optional dict of config values to override
+
+        Returns:
+            Tuple of (pipeline, study) ready for evaluation
+        """
+        # Load config if provided
+        if config_path:
+            with open(config_path) as f:
+                config_dict = json.load(f)
+                # Apply any overrides
+                if override_config:
+                    config_dict.update(override_config)
+                config = OptimizationConfig(**config_dict)
+        else:
+            # Create minimal config with overrides
+            config = OptimizationConfig(**(override_config or {}))
+
+        # Load study
+        from pathlib import Path
+
+        study_name = Path(study_path).stem
+        study = optuna.load_study(study_name=study_name, storage=f"sqlite:///{study_path}")
+
+        pipeline = cls(config)
+        return pipeline, study
+
+    def evaluate_on_dataset(
+        self, best_params: dict[str, Any], dataset_name: str, dataset_limit: Optional[int] = None
+    ) -> dict[str, Any]:
+        """Evaluate best parameters on a different dataset.
+
+        Args:
+            best_params: Dictionary of hyperparameters to use
+            dataset_name: Name of dataset to evaluate on
+            dataset_limit: Optional limit on number of samples
+
+        Returns:
+            Dictionary containing evaluation results on the new dataset
+        """
+        # Temporarily override dataset configuration
+        original_test_dataset = self.config.test_dataset
+        original_test_limit = self.config.test_limit
+
+        self.config.test_dataset = dataset_name
+        self.config.test_limit = dataset_limit or self.config.test_limit
+
+        self.logger.info(f"📊 Evaluating on {dataset_name} with {self.config.test_limit} samples")
+
+        # Reload test samples for new dataset
+        from . import data_utils
+
+        self.test_samples = data_utils.load_dataset_samples(self.config.test_dataset, self.config.test_limit)
+
+        # Run evaluation
+        results = self.evaluate_only(best_params)
+
+        # Restore original config
+        self.config.test_dataset = original_test_dataset
+        self.config.test_limit = original_test_limit
+
+        return results
+
+    def cleanup_memory(self):
+        """Clean up GPU/MPS memory."""
+        if hasattr(self, "model") and self.model is not None:
+            del self.model
+            self.model = None
+        if hasattr(self, "tokenizer") and self.tokenizer is not None:
+            del self.tokenizer
+            self.tokenizer = None
+
+        # Finish WandB run
+        if self.wandb_run is not None:
+            wandb.finish()
+            self.wandb_run = None
+
+        # Clean up device memory
+        empty_device_cache(self.device.type)
+
+        import gc
+
+        gc.collect()
+
+    def _init_wandb(self):
+        """Initialize WandB for experiment tracking."""
+        try:
+            self.wandb_run = wandb.init(
+                project=self.config.wandb_project,
+                name=f"{self.config.study_name}_{datetime.now().strftime('%Y%m%d_%H%M%S')}",
+                config=self.config.to_dict(),
+                tags=["optuna", "steering", "optimization"],
+                reinit=True,
+            )
+            self.logger.info(f"WandB initialized: {wandb.run.url}")
+        except Exception as e:
+            # Don't silently disable - user explicitly requested WandB
+            raise RuntimeError(
+                f"Failed to initialize WandB: {e}\n"
+                f"Possible solutions:\n"
+                f"1. Run 'wandb login' to authenticate\n"
+                f"2. Check your internet connection\n"
+                f"3. Verify project name: {self.config.wandb_project}\n"
+                f"4. Set use_wandb=False to disable WandB"
+            ) from e
+
+    def _log_trial_to_wandb(self, trial: optuna.Trial, metrics: dict[str, float]):
+        """Log trial results to WandB."""
+        if not self.config.use_wandb or self.wandb_run is None:
+            return
+
+        try:
+            # Log trial parameters and metrics
+            log_data = {f"trial/{k}": v for k, v in trial.params.items()}
+            log_data.update({f"metrics/{k}": v for k, v in metrics.items()})
+            log_data["trial/number"] = trial.number
+
+            wandb.log(log_data)
+        except Exception as e:
+            self.logger.warning(f"Failed to log trial to WandB: {e}")
+
+    def _log_final_results_to_wandb(self, study: optuna.Study, final_results: dict[str, Any]):
+        """Log final optimization results to WandB."""
+        if not self.config.use_wandb or self.wandb_run is None:
+            return
+
+        try:
+            # Log best trial results
+            best_params = {f"best/{k}": v for k, v in study.best_params.items()}
+            best_metrics = {
+                "best/validation_accuracy": study.best_value,
+                "best/baseline_accuracy": final_results["baseline_benchmark_metrics"]["accuracy"],
+                "best/steered_accuracy": final_results["steered_benchmark_metrics"]["accuracy"],
+                "best/accuracy_improvement": final_results["accuracy_improvement"],
+                "study/n_trials": len(study.trials),
+                "study/n_complete_trials": len(
+                    [t for t in study.trials if t.state == optuna.trial.TrialState.COMPLETE]
+                ),
+            }
+
+            wandb.log({**best_params, **best_metrics})
+
+            # Log optimization history
+            trial_values = [t.value for t in study.trials if t.value is not None]
+            if trial_values:
+                wandb.log(
+                    {
+                        "optimization/best_value_so_far": max(trial_values),
+                        "optimization/mean_trial_value": np.mean(trial_values),
+                        "optimization/std_trial_value": np.std(trial_values),
+                    }
+                )
+
+        except Exception as e:
+            self.logger.warning(f"Failed to log final results to WandB: {e}")
+
+    def _should_use_multiple_choice_evaluation(self) -> bool:
+        """Determine if we should use multiple choice evaluation for this dataset."""
+        # Use multiple choice evaluation for TruthfulQA and other MC tasks
+        return self.config.test_dataset.lower() in ["truthfulqa_mc1", "truthfulqa", "mmlu"]
+
+
+def main():
+    """Main entry point for optimization pipeline."""
+    # Setup logging
+    logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(name)s - %(levelname)s - %(message)s")
+
+    # Create configuration
+    config = OptimizationConfig(
+        train_limit=100,
+        contrastive_pairs_limit=30,  # Bounded by train_limit
+        val_limit=50,
+        test_limit=50,
+        n_trials=20,
+        layer_search_range=(10, 15),
+    )
+
+    # Run optimization
+    pipeline = OptimizationPipeline(config)
+    try:
+        results = pipeline.run_optimization()
+
+        print("🎉 Optimization completed!")
+        print(f"Best validation score: {results['best_validation_score']:.4f}")
+        print(f"Test accuracy: {results['steered_benchmark_metrics']['accuracy']:.4f}")
+        print(f"Accuracy improvement: {results['accuracy_improvement']:+.4f}")
+
+    finally:
+        # Clean up memory
+        pipeline.cleanup_memory()
+
+
+if __name__ == "__main__":
+    main()