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

wisent/core/lm_eval_harness_ground_truth.py

@@ -0,0 +1,1396 @@
+"""
+LM-Eval-Harness Ground Truth Evaluation
+
+This module provides ground truth evaluation using the lm-eval-harness framework.
+"""
+
+import logging
+from typing import Any, Dict
+
+from wisent_guard.core.activations import ActivationAggregationStrategy, Activations
+from wisent_guard.core.layer import Layer
+
+logger = logging.getLogger(__name__)
+
+
+class LMEvalHarnessGroundTruth:
+    """
+    Ground truth evaluator using lm-eval-harness tasks.
+
+    This class orchestrates the evaluation of classifiers on lm-eval-harness tasks
+    by routing to appropriate evaluation methods based on the task type.
+    """
+
+    def __init__(self, task_name: str, evaluation_method: str = None, model=None):
+        """
+        Initialize the LM-eval-harness ground truth evaluator.
+
+        Args:
+            task_name: Name of the lm-eval task
+            evaluation_method: Evaluation method ("log-likelihoods", "text-generation", "perplexity")
+            model: The model instance for activation extraction
+        """
+        self.task_name = task_name
+        self.evaluation_method = evaluation_method
+        self.model = model
+
+        # Load evaluation method from benchmark configuration if not provided
+        if not self.evaluation_method:
+            self.evaluation_method = self._get_evaluation_method_for_task(task_name)
+
+    def evaluate_classifier_on_task(
+        self,
+        classifier,
+        task_name: str,
+        num_samples: int = 100,
+        model=None,
+        layer: int = 15,
+        token_aggregation: str = "average",
+    ) -> Dict[str, Any]:
+        """
+        Evaluate a classifier on the specified lm-eval task.
+
+        Args:
+            classifier: The classifier to evaluate
+            task_name: Name of the lm-eval task
+            num_samples: Number of samples to evaluate
+            model: The model instance (overrides self.model if provided)
+            layer: Layer to extract activations from
+            token_aggregation: Token aggregation method ("average", "final", "first", "max", "min")
+
+        Returns:
+            Dict containing evaluation results
+        """
+        # Use provided model or fall back to self.model
+        evaluation_model = model or self.model
+
+        # Route to appropriate evaluation method
+        if self.evaluation_method == "log-likelihoods":
+            return self._evaluate_log_likelihoods(
+                classifier, task_name, num_samples, evaluation_model, layer, token_aggregation
+            )
+        if self.evaluation_method == "text-generation":
+            return self._evaluate_text_generation(
+                classifier, task_name, num_samples, evaluation_model, layer, token_aggregation
+            )
+        if self.evaluation_method == "perplexity":
+            return self._evaluate_perplexity(
+                classifier, task_name, num_samples, evaluation_model, layer, token_aggregation
+            )
+        if self.evaluation_method == "code-execution":
+            return self._evaluate_code_execution(
+                classifier, task_name, num_samples, evaluation_model, layer, token_aggregation
+            )
+        return {
+            "ground_truth": "UNKNOWN",
+            "method_used": "lm-eval-harness-unsupported",
+            "confidence": 0.0,
+            "details": f"Unsupported evaluation method: {self.evaluation_method}",
+            "task_name": task_name,
+            "evaluation_method": self.evaluation_method,
+        }
+
+    def _evaluate_log_likelihoods(
+        self, classifier, task_name: str, num_samples: int, model, layer: int, token_aggregation: str = "average"
+    ) -> Dict[str, Any]:
+        """Evaluate classifier using log-likelihoods approach."""
+        try:
+            from .log_likelihoods_evaluator import LogLikelihoodsEvaluator
+
+            # Create evaluator with model
+            evaluator = LogLikelihoodsEvaluator(task_name, model=model)
+
+            # Evaluate classifier
+            results = evaluator.evaluate_classifier_on_task(
+                classifier,
+                task_name,
+                num_samples=num_samples,
+                model=model,
+                layer=layer,
+                token_aggregation=token_aggregation,
+            )
+
+            return results
+
+        except Exception as e:
+            logger.error(f"Error in log-likelihoods evaluation: {e}")
+            return {
+                "ground_truth": "UNKNOWN",
+                "method_used": "lm-eval-harness-error",
+                "confidence": 0.0,
+                "details": f"Log-likelihoods evaluation failed: {e!s}",
+                "task_name": task_name,
+                "evaluation_method": "log-likelihoods",
+            }
+
+    def _evaluate_text_generation(
+        self, classifier, task_name: str, num_samples: int, model, layer: int, token_aggregation: str = "average"
+    ) -> Dict[str, Any]:
+        """Evaluate classifier using text generation approach."""
+        try:
+            logger.info(f"🎯 TEXT GENERATION EVALUATION: {task_name}")
+
+            # TODO In general LMEvalHarness should be rebuild to be BenchmarkGroundTruth
+            # Check if this is a TaskInterface task
+            if self._is_task_interface_task(task_name):
+                docs, task_data = self._load_task_interface_data(task_name, num_samples)
+            else:
+                # Use existing lm-eval task loading infrastructure
+                task_data = model.load_lm_eval_task(task_name, shots=0, limit=num_samples)
+                docs, _ = model.split_task_data(task_data, split_ratio=1.0)  # Use all for evaluation
+
+            if not docs:
+                return self._error_result(f"No documents retrieved from task: {task_name}")
+
+            logger.info(f"📝 Retrieved {len(docs)} documents from {task_name}")
+
+            # Generate responses using the model
+            generated_responses = []
+            ground_truth_responses = []
+
+            for i, doc in enumerate(docs):
+                try:
+                    # Extract question from document
+                    if hasattr(task_data, "doc_to_text"):
+                        question = task_data.doc_to_text(doc)
+                    else:
+                        question = str(doc.get("question", doc.get("text", "")))
+
+                    # Generate response using model
+                    logger.debug(f"🔸 Generating response for: {question[:100]}...")
+                    generated_response, _ = model.generate(
+                        prompt=question, layer_index=layer, max_new_tokens=150, temperature=0.1
+                    )
+
+                    # Extract ground truth answer
+                    # HLE task handling
+                    if task_name.startswith("hle") or task_name in ["math500", "math", "hendrycks_math"]:
+                        ground_truth = doc.get("answer", "")
+                    # AIME task handling
+                    elif task_name.startswith("aime"):
+                        ground_truth = str(doc.get("Answer", "") or doc.get("answer", ""))
+                    # FIXED: For DROP task, use raw document data to preserve structured format
+                    elif task_name == "drop":
+                        # Use raw answer field which contains the structured data
+                        ground_truth = doc.get("answer", {})
+                    elif hasattr(task_data, "doc_to_target"):
+                        ground_truth = task_data.doc_to_target(doc)
+                    else:
+                        ground_truth = str(doc.get("answer", doc.get("target", "")))
+
+                    generated_responses.append(
+                        {
+                            "question": question,
+                            "generated_response": generated_response,
+                            "ground_truth": ground_truth,
+                            "doc": doc,
+                        }
+                    )
+
+                    logger.debug(f"   📝 Generated: {generated_response[:100]}...")
+                    # FIXED: Handle ground_truth as int or string for logging
+                    gt_str = str(ground_truth)
+                    logger.debug(f"   ✅ Ground truth: {gt_str[:100]}...")
+
+                except Exception as e:
+                    logger.error(f"Error generating response for doc {i}: {e}")
+                    continue
+
+            # Evaluate using lm-eval-harness metrics
+            logger.info(f"🎯 Evaluating {len(generated_responses)} generated responses using lm-eval metrics...")
+
+            # Use lm-eval-harness's actual evaluation for this task
+            evaluation_results = self._evaluate_with_lm_eval_metrics(task_name, generated_responses, task_data)
+
+            # Now classify the generated responses to see if classifier agrees
+            classification_results = []
+            for response_data in generated_responses:
+                try:
+                    layer_obj = Layer(index=layer, type="transformer")
+
+                    # Extract activations from generated response
+                    activation_tensor = model.extract_activations(response_data["generated_response"], layer_obj)
+                    activation_method = self._map_token_aggregation_to_activation_method(token_aggregation)
+
+                    activation_obj = Activations(
+                        tensor=activation_tensor, layer=layer_obj, aggregation_strategy=activation_method
+                    )
+
+                    # Get classifier prediction
+                    features = activation_obj.extract_features_for_classifier()
+
+                    # Handle different classifier return formats
+                    try:
+                        # Try predict_proba first (returns probabilities)
+                        features_numpy = features.cpu().numpy()
+                        logger.debug(f"🔧 Features shape: {features_numpy.shape}")
+
+                        prediction_proba = classifier.predict_proba([features_numpy])
+                        logger.debug(f"🔧 predict_proba returned: {prediction_proba} (type: {type(prediction_proba)})")
+
+                        # Handle different return formats
+                        if isinstance(prediction_proba, (list, tuple)):
+                            if len(prediction_proba) > 0:
+                                prediction = prediction_proba[0]
+                                logger.debug(f"🔧 Extracted from list/tuple: {prediction} (type: {type(prediction)})")
+                            else:
+                                prediction = 0.5  # Default neutral score
+                        else:
+                            prediction = prediction_proba
+
+                        # If prediction is array-like, get first element
+                        if hasattr(prediction, "__len__") and not isinstance(prediction, str):
+                            if len(prediction) > 0:
+                                prediction = prediction[0]
+                                logger.debug(f"🔧 Extracted from array: {prediction} (type: {type(prediction)})")
+                            else:
+                                prediction = 0.5
+
+                        # Ensure we have a float
+                        prediction = float(prediction)
+                        logger.debug(f"🔧 Final prediction: {prediction}")
+
+                    except Exception as proba_error:
+                        logger.warning(f"predict_proba failed: {proba_error}, trying predict...")
+
+                        # Fallback to predict method
+                        try:
+                            predictions = classifier.predict([features.cpu().numpy()])
+                            if len(predictions) > 0:
+                                prediction = float(predictions[0])
+                            else:
+                                logger.warning("Classifier predict returned empty array")
+                                prediction = 0.5
+                        except Exception as predict_error:
+                            logger.error(f"Both predict_proba and predict failed: {predict_error}")
+                            prediction = 0.5  # Default neutral score
+
+                    classification_results.append(
+                        {
+                            "question": response_data["question"],
+                            "generated_response": response_data["generated_response"],
+                            "classifier_score": prediction,
+                            "classifier_truthful": prediction < 0.5,  # Lower = more truthful
+                        }
+                    )
+
+                    logger.debug("📋 Generated Response Classification:")
+                    logger.debug(f"   🔸 Question: {response_data['question'][:100]}...")
+                    logger.debug(f"   🧠 Generated: {response_data['generated_response'][:100]}...")
+                    logger.debug(f"   🎯 Classifier score: {prediction:.3f} (lower = more truthful)")
+                    logger.debug(f"   ✅ Classifier judgment: {'TRUTHFUL' if prediction < 0.5 else 'UNTRUTHFUL'}")
+
+                except Exception as e:
+                    import traceback
+
+                    logger.error(f"Error classifying generated response: {e}")
+                    logger.error(f"Full traceback: {traceback.format_exc()}")
+                    classification_results.append(
+                        {
+                            "question": response_data["question"],
+                            "generated_response": response_data["generated_response"],
+                            "classifier_score": 0.5,  # Default neutral score
+                            "classifier_truthful": False,
+                            "error": str(e),
+                        }
+                    )
+                    continue
+
+            return {
+                "ground_truth": "EVALUATED",
+                "method_used": "lm-eval-harness-text-generation",
+                "confidence": evaluation_results.get("accuracy", 0.0),
+                "details": f"Generated and evaluated {len(generated_responses)} responses using lm-eval metrics",
+                "task_name": task_name,
+                "evaluation_method": "text-generation",
+                "lm_eval_metrics": evaluation_results,
+                "classification_results": classification_results,
+                "total_samples": len(generated_responses),
+            }
+
+        except Exception as e:
+            logger.error(f"Error in text generation evaluation: {e}")
+            return self._error_result(f"Text generation evaluation error: {e!s}")
+
+    def _evaluate_perplexity(
+        self, classifier, task_name: str, num_samples: int, model, layer: int, token_aggregation: str = "average"
+    ) -> Dict[str, Any]:
+        """Evaluate classifier using perplexity approach."""
+        try:
+            logger.info(f"🎯 PERPLEXITY EVALUATION: {task_name}")
+
+            # Use existing task loading infrastructure
+            task_data = model.load_lm_eval_task(task_name, shots=0, limit=num_samples)
+            docs, _ = model.split_task_data(task_data, split_ratio=1.0)  # Use all for evaluation
+
+            if not docs:
+                return self._error_result(f"No documents retrieved from task: {task_name}")
+
+            logger.info(f"📝 Retrieved {len(docs)} documents from {task_name}")
+
+            # Calculate perplexity scores for different responses
+            perplexity_results = []
+
+            for i, doc in enumerate(docs):
+                try:
+                    # For WikiText and other pure language modeling tasks
+                    if task_name == "wikitext":
+                        # Get the full text for perplexity calculation
+                        text = doc.get("page", doc.get("text", ""))
+                        if not text:
+                            logger.warning(f"No text found in WikiText document {i}")
+                            continue
+
+                        logger.debug(f"🔸 Calculating perplexity for WikiText document {i} ({len(text)} chars)...")
+
+                        # Calculate perplexity on the full text
+                        perplexity = self._calculate_perplexity(model, text)
+
+                        # Extract activations from the text for classifier
+                        try:
+                            layer_obj = Layer(index=layer, type="transformer")
+
+                            # Use a truncated version for activation extraction if text is too long
+                            activation_text = text[:1000] if len(text) > 1000 else text
+                            activation_tensor = model.extract_activations(activation_text, layer_obj)
+                            activation_method = self._map_token_aggregation_to_activation_method(token_aggregation)
+
+                            activation_obj = Activations(
+                                tensor=activation_tensor, layer=layer_obj, aggregation_strategy=activation_method
+                            )
+
+                            # Get classifier prediction (only if classifier is provided)
+                            if classifier is not None:
+                                features = activation_obj.extract_features_for_classifier()
+
+                                # Handle different classifier return formats
+                                try:
+                                    prediction_proba = classifier.predict_proba([features.cpu().numpy()])
+
+                                    if isinstance(prediction_proba, (list, tuple)) and len(prediction_proba) > 0:
+                                        classification_score = float(prediction_proba[0])
+                                    else:
+                                        classification_score = float(prediction_proba)
+
+                                    if hasattr(classification_score, "__len__") and not isinstance(
+                                        classification_score, str
+                                    ):
+                                        classification_score = float(classification_score[0])
+
+                                except Exception as proba_error:
+                                    logger.warning(f"predict_proba failed: {proba_error}, trying predict...")
+                                    try:
+                                        predictions = classifier.predict([features.cpu().numpy()])
+                                        if len(predictions) > 0:
+                                            classification_score = float(predictions[0])
+                                        else:
+                                            logger.warning("Classifier predict returned empty array")
+                                            classification_score = 0.5
+                                    except Exception as predict_error:
+                                        logger.error(f"Both predict_proba and predict failed: {predict_error}")
+                                        classification_score = 0.5
+                            else:
+                                # No classifier provided - use default neutral score for perplexity-only evaluation
+                                classification_score = 0.5
+
+                        except Exception as e:
+                            logger.error(f"Error classifying WikiText document: {e}")
+                            classification_score = None
+
+                        result = {
+                            "document_idx": i,
+                            "text_preview": text[:200] + "..." if len(text) > 200 else text,
+                            "text_length": len(text),
+                            "perplexity": perplexity,
+                            "classifier_score": classification_score,
+                        }
+
+                        perplexity_results.append(result)
+
+                        logger.debug("📋 WikiText Perplexity Analysis:")
+                        logger.debug(f"   📊 Document {i}: {len(text)} chars")
+                        logger.debug(f"   🎯 Perplexity: {perplexity:.3f}")
+                        if classification_score is not None:
+                            logger.debug(f"   🧠 Classifier score: {classification_score:.3f} (lower = more truthful)")
+
+                        continue  # Skip the rest of the loop for WikiText
+
+                    # Extract question/prompt and possible completions for other tasks
+                    if hasattr(task_data, "doc_to_text"):
+                        prompt = task_data.doc_to_text(doc)
+                    else:
+                        prompt = str(doc.get("question", doc.get("text", "")))
+
+                    # For multiple choice tasks, get all choices
+                    choices = []
+                    if hasattr(task_data, "doc_to_choice"):
+                        choices = [
+                            task_data.doc_to_choice(doc, choice_idx)
+                            for choice_idx in range(len(doc.get("choices", [])))
+                        ]
+                    elif "choices" in doc:
+                        choices = doc["choices"]
+                    else:
+                        # For non-multiple choice, generate a response and calculate its perplexity
+                        generated_response, _ = model.generate(
+                            prompt=prompt, layer_index=layer, max_new_tokens=100, temperature=0.1
+                        )
+                        choices = [generated_response]
+
+                    logger.debug(f"🔸 Calculating perplexity for: {prompt[:100]}...")
+
+                    # Calculate perplexity for each choice
+                    choice_perplexities = []
+                    for choice_idx, choice in enumerate(choices):
+                        try:
+                            # Calculate perplexity of the choice given the prompt
+                            full_text = f"{prompt} {choice}"
+                            perplexity = self._calculate_perplexity(model, full_text)
+
+                            choice_perplexities.append(
+                                {"choice_idx": choice_idx, "choice_text": choice, "perplexity": perplexity}
+                            )
+
+                            logger.debug(f"   📊 Choice {choice_idx}: {choice[:50]}... (perplexity: {perplexity:.3f})")
+
+                        except Exception as e:
+                            logger.error(f"Error calculating perplexity for choice {choice_idx}: {e}")
+                            continue
+
+                    # Get ground truth answer index
+                    ground_truth_idx = None
+                    if hasattr(task_data, "doc_to_target"):
+                        ground_truth = task_data.doc_to_target(doc)
+                        try:
+                            ground_truth_idx = int(ground_truth)
+                        except:
+                            ground_truth_idx = None
+                    elif "answer" in doc:
+                        ground_truth_idx = doc["answer"]
+
+                    # Find the choice with lowest perplexity (most likely)
+                    if choice_perplexities:
+                        best_choice = min(choice_perplexities, key=lambda x: x["perplexity"])
+
+                        # Classify the best choice using the classifier
+                        classification_score = None
+                        try:
+                            layer_obj = Layer(index=layer, type="transformer")
+
+                            # Extract activations from the best choice
+                            activation_tensor = model.extract_activations(best_choice["choice_text"], layer_obj)
+                            activation_method = self._map_token_aggregation_to_activation_method(token_aggregation)
+
+                            activation_obj = Activations(
+                                tensor=activation_tensor, layer=layer_obj, aggregation_strategy=activation_method
+                            )
+
+                            # Get classifier prediction
+                            features = activation_obj.extract_features_for_classifier()
+
+                            # Handle different classifier return formats
+                            try:
+                                # Try predict_proba first (returns probabilities)
+                                prediction_proba = classifier.predict_proba([features.cpu().numpy()])
+
+                                # Handle different return formats
+                                if isinstance(prediction_proba, (list, tuple)):
+                                    if len(prediction_proba) > 0:
+                                        classification_score = prediction_proba[0]
+                                    else:
+                                        classification_score = 0.5  # Default neutral score
+                                else:
+                                    classification_score = prediction_proba
+
+                                # If prediction is array-like, get first element
+                                if hasattr(classification_score, "__len__") and not isinstance(
+                                    classification_score, str
+                                ):
+                                    if len(classification_score) > 0:
+                                        classification_score = classification_score[0]
+                                    else:
+                                        classification_score = 0.5
+
+                                # Ensure we have a float
+                                classification_score = float(classification_score)
+
+                            except Exception as proba_error:
+                                logger.warning(f"predict_proba failed: {proba_error}, trying predict...")
+
+                                # Fallback to predict method
+                                try:
+                                    predictions = classifier.predict([features.cpu().numpy()])
+                                    if len(predictions) > 0:
+                                        classification_score = float(predictions[0])
+                                    else:
+                                        logger.warning("Classifier predict returned empty array")
+                                        classification_score = 0.5
+                                except Exception as predict_error:
+                                    logger.error(f"Both predict_proba and predict failed: {predict_error}")
+                                    classification_score = 0.5  # Default neutral score
+
+                        except Exception as e:
+                            logger.error(f"Error classifying best choice: {e}")
+
+                        result = {
+                            "question": prompt,
+                            "choices": choice_perplexities,
+                            "best_choice_idx": best_choice["choice_idx"],
+                            "best_choice_text": best_choice["choice_text"],
+                            "best_choice_perplexity": best_choice["perplexity"],
+                            "ground_truth_idx": ground_truth_idx,
+                            "classifier_score": classification_score,
+                            "perplexity_correct": best_choice["choice_idx"] == ground_truth_idx
+                            if ground_truth_idx is not None
+                            else None,
+                        }
+
+                        perplexity_results.append(result)
+
+                        logger.debug("📋 Perplexity Analysis:")
+                        logger.debug(f"   🔸 Question: {prompt[:100]}...")
+                        logger.debug(f"   📊 Best choice (lowest perplexity): {best_choice['choice_text'][:100]}...")
+                        logger.debug(f"   🎯 Perplexity: {best_choice['perplexity']:.3f}")
+                        logger.debug(
+                            f"   🧠 Classifier score: {classification_score:.3f} (lower = more truthful)"
+                            if classification_score is not None
+                            else "   🧠 Classifier score: N/A"
+                        )
+                        logger.debug(f"   ✅ Perplexity correct: {result['perplexity_correct']}")
+
+                except Exception as e:
+                    logger.error(f"Error processing doc {i}: {e}")
+                    continue
+
+            # Calculate overall metrics
+            total_samples = len(perplexity_results)
+
+            if task_name == "wikitext":
+                # For WikiText, we don't have correct/incorrect, just perplexity values
+                perplexities = [r["perplexity"] for r in perplexity_results if r["perplexity"] != float("inf")]
+                avg_perplexity = sum(perplexities) / len(perplexities) if perplexities else float("inf")
+
+                # Average classifier score
+                classifier_scores = [
+                    r["classifier_score"] for r in perplexity_results if r["classifier_score"] is not None
+                ]
+                avg_classifier_score = sum(classifier_scores) / len(classifier_scores) if classifier_scores else None
+
+                perplexity_accuracy = 1.0 if avg_perplexity < 100 else 0.0  # Arbitrary threshold for "good" perplexity
+                correct_perplexity = sum(1 for r in perplexity_results if r["perplexity"] < 100)
+            else:
+                correct_perplexity = sum(1 for r in perplexity_results if r.get("perplexity_correct") == True)
+                perplexity_accuracy = correct_perplexity / total_samples if total_samples > 0 else 0.0
+
+                # Average classifier score
+                classifier_scores = [
+                    r["classifier_score"] for r in perplexity_results if r["classifier_score"] is not None
+                ]
+                avg_classifier_score = sum(classifier_scores) / len(classifier_scores) if classifier_scores else None
+
+            logger.info("📊 PERPLEXITY EVALUATION RESULTS:")
+            logger.info(f"   • Total samples: {total_samples}")
+            if task_name == "wikitext":
+                logger.info(f"   • Average perplexity: {avg_perplexity:.3f}")
+                logger.info(f"   • Documents with perplexity < 100: {correct_perplexity}")
+            else:
+                logger.info(f"   • Perplexity accuracy: {perplexity_accuracy:.3f}")
+            logger.info(
+                f"   • Average classifier score: {avg_classifier_score:.3f}"
+                if avg_classifier_score is not None
+                else "   • Average classifier score: N/A"
+            )
+
+            result_dict = {
+                "ground_truth": "EVALUATED",
+                "method_used": "lm-eval-harness-perplexity",
+                "confidence": perplexity_accuracy,
+                "details": f"Calculated perplexity for {total_samples} samples",
+                "task_name": task_name,
+                "evaluation_method": "perplexity",
+                "perplexity_accuracy": perplexity_accuracy,
+                "average_classifier_score": avg_classifier_score,
+                "total_samples": total_samples,
+                "correct_perplexity": correct_perplexity,
+                "perplexity_results": perplexity_results[:10],  # First 10 for debugging
+            }
+
+            if task_name == "wikitext":
+                result_dict["average_perplexity"] = avg_perplexity
+                result_dict["details"] = (
+                    f"Calculated perplexity for {total_samples} WikiText documents, avg perplexity: {avg_perplexity:.3f}"
+                )
+            else:
+                result_dict["details"] = (
+                    f"Calculated perplexity for {total_samples} samples, accuracy: {perplexity_accuracy:.3f}"
+                )
+
+            return result_dict
+
+        except Exception as e:
+            logger.error(f"Error in perplexity evaluation: {e}")
+            return self._error_result(f"Perplexity evaluation error: {e!s}")
+
+    def _get_evaluation_method_for_task(self, task_name: str) -> str:
+        """Get the evaluation method for a task from the benchmark configuration."""
+        try:
+            import json
+
+            eval_methods_path = "wisent_guard/parameters/benchmarks/benchmark_evaluation_methods.json"
+            with open(eval_methods_path) as f:
+                benchmark_methods = json.load(f)
+                return benchmark_methods.get(task_name, "text-generation")
+        except Exception as e:
+            logger.debug(f"Could not load benchmark evaluation methods: {e}")
+            return "text-generation"
+
+    def _error_result(self, error_message: str) -> Dict[str, Any]:
+        """Return a standardized error result."""
+        return {
+            "ground_truth": "ERROR",
+            "method_used": "lm-eval-harness-error",
+            "confidence": 0.0,
+            "details": error_message,
+            "task_name": self.task_name,
+            "evaluation_method": self.evaluation_method,
+        }
+
+    def _map_token_aggregation_to_activation_method(self, token_aggregation: str):
+        """Map token aggregation string to activation method."""
+
+        mapping = {  # TODO This should be refactor, why we use strings as Token aggregation?
+            "average": ActivationAggregationStrategy.MEAN_POOLING,
+            "mean": ActivationAggregationStrategy.MEAN_POOLING,
+            "last": ActivationAggregationStrategy.LAST_TOKEN,
+            "max": ActivationAggregationStrategy.MAX_POOLING,
+        }
+
+        return mapping.get(token_aggregation.lower(), ActivationAggregationStrategy.MEAN_POOLING)
+
+    def _is_task_interface_task(self, task_name: str) -> bool:
+        """Check if this is a TaskInterface task (not an lm-eval task)."""
+        # List of known TaskInterface tasks
+        task_interface_tasks = {
+            "hle",
+            "hle_exact_match",
+            "hle_multiple_choice",
+            "livecodebench",
+            "math500",
+            "math",
+            "hendrycks_math",
+            "aime",
+            "aime2025",
+            "aime2024",
+            "hmmt",
+            "hmmt_feb_2025",
+            "polymath",
+            "polymath_en_medium",
+            "polymath_zh_medium",
+            "polymath_en_high",
+            "polymath_zh_high",
+            "livemathbench",
+            "livemathbench_cnmo_en",
+            "livemathbench_cnmo_zh",
+        }
+        return task_name in task_interface_tasks
+
+    def _load_task_interface_data(self, task_name: str, num_samples: int):
+        """Load data from TaskInterface tasks."""
+        try:
+            from .task_interface import get_task
+
+            # Get the task instance
+            task = get_task(task_name)
+
+            # Load data
+            docs = task.load_data(limit=num_samples)
+
+            return docs, task
+
+        except Exception as e:
+            logger.error(f"Failed to load TaskInterface task {task_name}: {e}")
+            return [], None
+
+    def _calculate_perplexity(self, model, text: str) -> float:
+        """Calculate perplexity of text using the model."""
+        try:
+            import numpy as np
+            import torch
+
+            # Use the model's prepare_activations method to get outputs
+            prepared = model.prepare_activations(text)
+            outputs = prepared["outputs"]
+            inputs = prepared["inputs"]
+
+            # Get input IDs
+            input_ids = inputs["input_ids"]
+
+            # Get logits from the outputs
+            logits = outputs.logits
+
+            # Compute log probabilities
+            log_probs = torch.log_softmax(logits, dim=-1)
+
+            # Get log probabilities for actual tokens (shifted for next-token prediction)
+            # input_ids shape: [batch_size, sequence_length]
+            # logits shape: [batch_size, sequence_length, vocab_size]
+            # We need to match targets with predictions
+
+            if input_ids.shape[1] > 1:
+                # Get log probabilities for the target tokens
+                target_ids = input_ids[0, 1:]  # Skip first token (no prediction for it)
+                prediction_logits = log_probs[0, :-1, :]  # Skip last prediction (no target for it)
+
+                # Get log probabilities for actual tokens
+                token_log_probs = prediction_logits.gather(dim=-1, index=target_ids.unsqueeze(-1)).squeeze(-1)
+
+                # Compute average log probability
+                avg_log_prob = token_log_probs.mean().item()
+
+                # Compute perplexity
+                perplexity = np.exp(-avg_log_prob)
+            else:
+                # Single token, cannot compute perplexity
+                perplexity = float("inf")
+
+            return perplexity
+
+        except Exception as e:
+            logger.error(f"Error calculating perplexity: {e}")
+            return float("inf")
+
+    def _evaluate_generic_code_execution(
+        self, classifier, task_name: str, num_samples: int, model, layer: int, token_aggregation: str = "average"
+    ) -> Dict[str, Any]:
+        """Evaluate generic code execution tasks (non-BigCode) like LiveCodeBench."""
+        try:
+            logger.info(f"🎯 GENERIC CODE EXECUTION EVALUATION: {task_name}")
+
+            # Get secure code evaluator
+            from .secure_code_evaluator import SecureCodeEvaluator
+
+            secure_evaluator = SecureCodeEvaluator()
+
+            # Load task data
+            task_data = model.load_lm_eval_task(task_name, shots=0, limit=num_samples)
+
+            if hasattr(task_data, "test_docs"):
+                docs = task_data.test_docs()
+            else:
+                docs, _ = model.split_task_data(task_data, split_ratio=1.0)
+
+            if not docs:
+                return self._error_result(f"No documents retrieved from task: {task_name}")
+
+            logger.info(f"📝 Retrieved {len(docs)} documents from {task_name}")
+
+            # Generate code for each sample
+            generated_codes = []
+            evaluation_results = []
+
+            for i, doc in enumerate(docs):
+                try:
+                    # Get prompt
+                    if hasattr(task_data, "doc_to_text"):
+                        prompt = task_data.doc_to_text(doc)
+                    else:
+                        # For LiveCodeBench
+                        question = doc.get("question_content", doc.get("text", ""))
+                        starter_code = doc.get("starter_code", "")
+                        prompt = f"{question}\n\n{starter_code}" if starter_code else question
+
+                    logger.debug(f"📋 Prompt for sample {i + 1}:\n{prompt[:200]}...\n")
+
+                    # Generate code using model
+                    logger.debug(f"🔸 Generating code for sample {i + 1}/{len(docs)}...")
+                    generated_code, _ = model.generate(
+                        prompt=prompt,
+                        layer_index=layer,
+                        max_new_tokens=500,  # More tokens for code generation
+                        temperature=0.1,
+                    )
+
+                    generated_codes.append(generated_code)
+                    logger.debug(f"   📝 Generated code:\n{generated_code}\n")
+
+                    # Evaluate generated code
+                    eval_result = secure_evaluator.evaluate_response(task_name, doc, generated_code)
+                    evaluation_results.append(eval_result)
+
+                    logger.debug(
+                        f"   ✅ Evaluation result: {'PASSED' if eval_result.get('passed', False) else 'FAILED'}"
+                    )
+                    if "pass_rate" in eval_result:
+                        logger.debug(f"   📊 Pass rate: {eval_result['pass_rate']:.2%}")
+
+                except Exception as e:
+                    logger.error(f"Error processing sample {i}: {e}")
+                    generated_codes.append("")
+                    evaluation_results.append({"passed": False, "error": str(e), "success": False})
+
+            # Aggregate results
+            total_passed = sum(1 for r in evaluation_results if r.get("passed", False))
+            accuracy = total_passed / len(evaluation_results) if evaluation_results else 0.0
+
+            logger.info(
+                f"📊 CODE EXECUTION COMPLETED: {total_passed}/{len(evaluation_results)} passed ({accuracy:.2%})"
+            )
+
+            # Clean up Docker resources
+            secure_evaluator.cleanup()
+
+            return {
+                "ground_truth": "EVALUATED",
+                "method_used": f"generic-code-execution-{task_name}",
+                "confidence": accuracy,
+                "accuracy": accuracy,
+                "details": f"Executed and evaluated {len(generated_codes)} code samples",
+                "task_name": task_name,
+                "evaluation_method": "code-execution",
+                "total_samples": len(generated_codes),
+                "passed_samples": total_passed,
+                "evaluation_results": evaluation_results,
+            }
+
+        except Exception as e:
+            logger.error(f"Error in generic code execution evaluation: {e}")
+            import traceback
+
+            logger.error(f"Traceback: {traceback.format_exc()}")
+            return self._error_result(f"Generic code execution evaluation error: {e!s}")
+
+    def _evaluate_with_lm_eval_metrics(self, task_name: str, response_data: list, task_data) -> Dict[str, Any]:
+        """Evaluate responses using task-specific evaluation metrics."""
+        try:
+            correct = 0
+            total = len(response_data)
+            evaluation_details = []
+
+            for response in response_data:
+                generated = response["generated_response"]
+                ground_truth = response["ground_truth"]
+
+                # Task-specific evaluation logic
+                if task_name == "gsm8k":
+                    # GSM8K uses exact match on numerical answer
+                    is_correct = self._evaluate_gsm8k_response(generated, ground_truth)
+                elif task_name.startswith("math") or task_name in ["hendrycks_math"]:
+                    # MATH-500 and related benchmarks use same evaluation as GSM8K (numerical extraction)
+                    is_correct = self._evaluate_gsm8k_response(generated, ground_truth)
+                elif task_name in ["arc_easy", "arc_challenge"]:
+                    # ARC uses exact match on choice letter/number
+                    is_correct = self._evaluate_arc_response(generated, ground_truth)
+                elif task_name == "hellaswag":
+                    # HellaSwag uses exact match on choice index
+                    is_correct = self._evaluate_hellaswag_response(generated, ground_truth)
+                elif task_name == "mathqa":
+                    # MATH_QA uses exact match on choice index (0, 1, 2, 3)
+                    is_correct = self._evaluate_mathqa_response(generated, ground_truth)
+                elif task_name == "drop":
+                    # DROP uses structured answer format with numbers, spans, and dates
+                    is_correct = self._evaluate_drop_response(generated, ground_truth)
+                elif task_name.startswith("gpqa"):
+                    # GPQA uses multiple-choice answer extraction (A, B, C, D)
+                    is_correct = self._evaluate_multiple_choice_response(generated, ground_truth)
+                elif task_name.startswith("hle") and "multiple_choice" in task_name:
+                    # HLE multiple choice uses letter extraction (A, B, C, D, E)
+                    is_correct = self._evaluate_multiple_choice_response(generated, ground_truth)
+                elif task_name.startswith("truthfulqa") or task_name == "truthfulqa_mc1":
+                    # TruthfulQA uses multiple-choice answer extraction (A, B, C, D)
+                    is_correct = self._evaluate_multiple_choice_response(generated, ground_truth)
+                else:
+                    # Default: string matching with some flexibility
+                    is_correct = self._evaluate_default_response(generated, ground_truth)
+
+                if is_correct:
+                    correct += 1
+
+                evaluation_details.append(
+                    {
+                        "question": response["question"][:100],
+                        "generated": generated[-50:],
+                        "ground_truth": ground_truth,
+                        "correct": is_correct,
+                    }
+                )
+
+                logger.debug(f"📊 Evaluation: {response['question'][:50]}...")
+                logger.debug(f"   Generated: {generated[:50]}...")
+                logger.debug(f"   Ground Truth: {ground_truth}")
+                logger.debug(f"   Correct: {is_correct}")
+
+            accuracy = correct / total if total > 0 else 0.0
+
+            return {
+                "accuracy": accuracy,
+                "correct_predictions": correct,
+                "total_samples": total,
+                "evaluation_details": evaluation_details[:5],  # First 5 for debugging
+                "task_name": task_name,
+            }
+
+        except Exception as e:
+            logger.error(f"Error in metrics evaluation: {e}")
+            return {"accuracy": 0.0, "correct_predictions": 0, "total_samples": len(response_data), "error": str(e)}
+
+    def _evaluate_gsm8k_response(self, generated: str, ground_truth) -> bool:
+        """Evaluate GSM8K response using numerical answer extraction."""
+        try:
+            # Extract numerical answer from generated response
+            # GSM8K answers are typically in format "#### 42" or just the number
+            generated_answer = self._extract_numerical_answer(generated)
+            ground_truth_answer = self._extract_numerical_answer(str(ground_truth))
+
+            # Compare numerical values
+            if generated_answer is not None and ground_truth_answer is not None:
+                return abs(generated_answer - ground_truth_answer) < 1e-6
+
+            # Fallback to string matching
+            return generated.strip().lower() == str(ground_truth).strip().lower()
+
+        except Exception as e:
+            logger.error(f"Error evaluating GSM8K response: {e}")
+            return False
+
+    def _extract_numerical_answer(self, text: str) -> float:
+        """Extract numerical answer from text."""
+        try:
+            import re
+
+            # Look for #### pattern (GSM8K format)
+            pattern = r"####\s*([+-]?\d+(?:\.\d+)?)"
+            match = re.search(pattern, text)
+            if match:
+                return float(match.group(1))
+
+            # Look for last number in text
+            numbers = re.findall(r"[+-]?\d+(?:\.\d+)?", text)
+            if numbers:
+                return float(numbers[-1])
+
+            return None
+
+        except Exception as e:
+            logger.error(f"Error extracting numerical answer: {e}")
+            return None
+
+    def _evaluate_arc_response(self, generated: str, ground_truth) -> bool:
+        """Evaluate ARC response using exact match."""
+        try:
+            # Normalize responses
+            gen_clean = generated.strip().lower()
+            gt_clean = str(ground_truth).strip().lower()
+
+            # Direct match
+            if gen_clean == gt_clean:
+                return True
+
+            # Check if generated contains the ground truth
+            if gt_clean in gen_clean:
+                return True
+
+            # Check for choice letter/number patterns
+            import re
+
+            gen_match = re.search(r"[abcd]|\d+", gen_clean)
+            gt_match = re.search(r"[abcd]|\d+", gt_clean)
+
+            if gen_match and gt_match:
+                return gen_match.group() == gt_match.group()
+
+            return False
+
+        except Exception as e:
+            logger.error(f"Error evaluating ARC response: {e}")
+            return False
+
+    def _evaluate_hellaswag_response(self, generated: str, ground_truth) -> bool:
+        """Evaluate HellaSwag response using exact match."""
+        try:
+            # Normalize and compare
+            gen_clean = generated.strip().lower()
+            gt_clean = str(ground_truth).strip().lower()
+
+            return gen_clean == gt_clean or gt_clean in gen_clean
+
+        except Exception as e:
+            logger.error(f"Error evaluating HellaSwag response: {e}")
+            return False
+
+    def _evaluate_mathqa_response(self, generated: str, ground_truth) -> bool:
+        """Evaluate MATH_QA response using choice matching."""
+        try:
+            import re
+
+            # Ground truth is typically 0, 1, 2, or 3 (choice index)
+            gt_str = str(ground_truth).strip()
+
+            # Look for choice patterns in generated response
+            gen_clean = generated.strip().lower()
+
+            # Direct match with choice index
+            if gt_str in gen_clean:
+                return True
+
+            # Look for choice letter patterns (a=0, b=1, c=2, d=3)
+            choice_map = {"a": "0", "b": "1", "c": "2", "d": "3"}
+            for letter, index in choice_map.items():
+                if index == gt_str and letter in gen_clean:
+                    return True
+
+            # Look for explicit choice pattern like "The answer is 1" or "Choice B"
+            choice_patterns = [
+                rf"\b{gt_str}\b",  # Exact number match
+                rf"choice\s*{choice_map.get(gt_str, gt_str)}",  # "choice 1"
+                rf"answer\s*is\s*{gt_str}",  # "answer is 1"
+                rf"option\s*{gt_str}",  # "option 1"
+            ]
+
+            for pattern in choice_patterns:
+                if re.search(pattern, gen_clean):
+                    return True
+
+            return False
+
+        except Exception as e:
+            logger.error(f"Error evaluating MATH_QA response: {e}")
+            return False
+
+    def _evaluate_drop_response(self, generated: str, ground_truth) -> bool:
+        """Evaluate DROP response using structured answer format."""
+        try:
+            import json
+            import re
+
+            # Parse ground truth if it's a string representation of a dict
+            if isinstance(ground_truth, str):
+                try:
+                    # Try to parse as JSON first
+                    if ground_truth.startswith("{"):
+                        gt_dict = json.loads(ground_truth)
+                    else:
+                        # Handle malformed string representations
+                        return False
+                except:
+                    return False
+            elif isinstance(ground_truth, dict):
+                gt_dict = ground_truth
+            else:
+                return False
+
+            gen_clean = generated.strip().lower()
+
+            # Check number field
+            if gt_dict.get("number"):
+                number_str = str(gt_dict["number"]).strip()
+                if number_str:
+                    # Direct number match
+                    if number_str.lower() in gen_clean:
+                        return True
+
+                    # Try to extract numbers from generated response
+                    gen_numbers = re.findall(r"\b\d+\b", generated)
+                    if number_str in gen_numbers:
+                        return True
+
+                    # Word number matching (e.g., "two" vs "2")
+                    number_words = {
+                        "0": ["zero", "none"],
+                        "1": ["one"],
+                        "2": ["two"],
+                        "3": ["three"],
+                        "4": ["four"],
+                        "5": ["five"],
+                        "6": ["six"],
+                        "7": ["seven"],
+                        "8": ["eight"],
+                        "9": ["nine"],
+                        "10": ["ten"],
+                    }
+                    if number_str in number_words:
+                        for word in number_words[number_str]:
+                            if word in gen_clean:
+                                return True
+
+            # Check spans field
+            if gt_dict.get("spans"):
+                spans = gt_dict["spans"]
+                if isinstance(spans, list):
+                    for span in spans:
+                        span_clean = str(span).strip().lower()
+                        if span_clean and span_clean in gen_clean:
+                            return True
+                elif isinstance(spans, str):
+                    span_clean = spans.strip().lower()
+                    if span_clean and span_clean in gen_clean:
+                        return True
+
+            # Check date field (less common but possible)
+            if gt_dict.get("date"):
+                date_obj = gt_dict["date"]
+                if isinstance(date_obj, dict):
+                    # Check individual date components
+                    for component in ["day", "month", "year"]:
+                        if date_obj.get(component):
+                            date_val = str(date_obj[component]).strip().lower()
+                            if date_val and date_val in gen_clean:
+                                return True
+
+            return False
+
+        except Exception as e:
+            logger.error(f"Error evaluating DROP response: {e}")
+            return False
+
+    def _evaluate_default_response(self, generated: str, ground_truth) -> bool:
+        """Default evaluation using flexible string matching."""
+        try:
+            gen_clean = generated.strip().lower()
+
+            # Handle list ground truth (e.g., COQA format)
+            if isinstance(ground_truth, list):
+                # Check if generated response matches any of the acceptable answers
+                for gt_option in ground_truth:
+                    gt_clean = str(gt_option).strip().lower()
+
+                    # Exact match
+                    if gen_clean == gt_clean:
+                        return True
+
+                    # Contains match
+                    if gt_clean in gen_clean or gen_clean in gt_clean:
+                        return True
+
+                return False
+            # Handle string ground truth
+            gt_clean = str(ground_truth).strip().lower()
+
+            # Exact match
+            if gen_clean == gt_clean:
+                return True
+
+            # Contains match
+            if gt_clean in gen_clean or gen_clean in gt_clean:
+                return True
+
+            return False
+
+        except Exception as e:
+            logger.error(f"Error in default evaluation: {e}")
+            return False
+
+    def _evaluate_multiple_choice_response(self, generated: str, ground_truth) -> bool:
+        """Evaluate multiple choice response by extracting choice letter (A, B, C, D, E)."""
+        import re
+
+        try:
+            # Clean the generated response
+            gen_clean = generated.strip()
+
+            # Convert ground truth to string and extract expected letter
+            gt_str = str(ground_truth).strip()
+            expected_letter = None
+
+            # Extract letter from ground truth (could be "(A)", "A", etc.)
+            gt_match = re.search(r"[ABCDE]", gt_str.upper())
+            if gt_match:
+                expected_letter = gt_match.group()
+            else:
+                return False
+
+            # Try multiple strict patterns to extract answer from generated response
+            # These patterns require clear context indicating an intentional choice
+            patterns = [
+                # Fixed pattern to avoid matching 'A' in "Answer:" alone
+                r"(?:answer|choice|option)\s*(?:is\s+|:\s*)(?:\()?([ABCDE])(?:\))?",  # "Answer: A" or "Answer is (B)" - requires letter after
+                r"the\s+(?:correct\s+)?answer\s+is\s*(?:\()?([ABCDE])(?:\))?",  # "The answer is A" - requires "the answer is"
+                r"(?:select|choose)\s+(?:\()?([ABCDE])(?:\))?",  # "Select A" or "Choose A" - requires the action word
+                r"(?:^|\n)([ABCDE])(?:\s*$)",  # Letter at start of line followed by whitespace/end only
+                r"^([ABCDE])[.,;!?)\s]*$",  # Just the letter with optional punctuation and whitespace
+                r"^(?:\()?([ABCDE])(?:\))?\s*$",  # Just the letter with optional parentheses
+            ]
+
+            # Try each pattern - only accept clear, intentional responses
+            for pattern in patterns:
+                matches = re.finditer(pattern, gen_clean.upper(), re.IGNORECASE | re.MULTILINE)
+                for match in matches:
+                    extracted_letter = match.group(1).upper()
+                    if extracted_letter == expected_letter:
+                        return True
+
+            # No more fallback - if we can't clearly identify the choice, it's wrong
+            return False
+
+        except Exception as e:
+            logger.error(f"Error evaluating multiple choice response: {e}")
+            return False
+
+    def _evaluate_code_execution(
+        self, classifier, task_name: str, num_samples: int, model, layer: int, token_aggregation: str = "average"
+    ) -> Dict[str, Any]:
+        """Evaluate classifier using code execution approach for BigCode tasks."""
+        try:
+            logger.debug(f"🎯 CODE EXECUTION EVALUATION: {task_name}")
+
+            # Check if it's a BigCode task
+            from .bigcode_integration import get_bigcode_evaluator, is_bigcode_task, load_bigcode_task
+            from .secure_code_evaluator import SecureCodeEvaluator
+
+            if not is_bigcode_task(task_name):
+                # Check if it's still a code execution task (like LiveCodeBench)
+                if SecureCodeEvaluator.is_code_execution_task(task_name):
+                    logger.info(f"Task {task_name} is a non-BigCode code execution task")
+                    return self._evaluate_generic_code_execution(
+                        classifier, task_name, num_samples, model, layer, token_aggregation
+                    )
+                logger.warning(f"Task {task_name} is not a code execution task, falling back to text generation")
+                return self._evaluate_text_generation(
+                    classifier, task_name, num_samples, model, layer, token_aggregation
+                )
+
+            # Load BigCode task
+            bigcode_task = load_bigcode_task(task_name, limit=num_samples)
+            logger.info(f"📝 Loaded BigCode task {task_name} with {len(bigcode_task)} samples")
+
+            # Generate code for each sample
+            generated_codes = []
+            for i, sample in enumerate(bigcode_task.get_samples()):
+                try:
+                    # Get prompt
+                    prompt = bigcode_task.doc_to_text(sample)
+                    logger.debug(f"📋 Prompt for sample {i + 1}:\n{prompt}\n")
+
+                    # Generate code using model
+                    logger.debug(f"🔸 Generating code for sample {i + 1}/{len(bigcode_task)}...")
+                    generated_code, _ = model.generate(
+                        prompt=prompt,
+                        layer_index=layer,
+                        max_new_tokens=300,  # More tokens for code generation
+                        temperature=0.1,
+                        # Note: stop_sequences not supported by all models
+                    )
+
+                    generated_codes.append(generated_code)
+                    logger.debug(f"   📝 Generated: {generated_code[:100]}...")
+                    logger.debug(f"   📝 Full generated code:\n{generated_code}\n")
+
+                except Exception as e:
+                    logger.error(f"Error generating code for sample {i}: {e}")
+                    generated_codes.append("")  # Empty code for failed generation
+
+            # Evaluate generated code using BigCode evaluator
+            logger.info(f"🎯 Evaluating {len(generated_codes)} generated code samples...")
+
+            # Get Docker executor if available
+            docker_executor = None
+            try:
+                from .docker import OptimizedDockerExecutor
+
+                docker_executor = OptimizedDockerExecutor()
+            except Exception as e:
+                logger.warning(f"Docker executor not available: {e}")
+
+            # Use BigCode evaluator
+            evaluator = get_bigcode_evaluator(docker_executor)
+
+            # Prepare generations in expected format (list of lists)
+            generations_for_eval = [[code] for code in generated_codes]
+
+            # Run evaluation
+            evaluation_results = evaluator.evaluate(
+                bigcode_task,
+                generations_for_eval,
+                k_values=[1],  # Just pass@1 for now
+            )
+
+            # Extract pass rate
+            pass_rate = evaluation_results.get("pass_at_k", {}).get("pass@1", 0.0)
+
+            logger.info(f"✅ Code execution pass@1: {pass_rate:.2%}")
+
+            # Now classify the generated code to see if classifier agrees
+            classification_results = []
+            for i, code in enumerate(generated_codes):
+                try:
+                    layer_obj = Layer(index=layer, type="transformer")
+
+                    # Extract activations from generated code
+                    activation_tensor = model.extract_activations(code, layer_obj)
+                    activation_method = self._map_token_aggregation_to_activation_method(token_aggregation)
+
+                    activation_obj = Activations(
+                        tensor=activation_tensor, layer=layer_obj, aggregation_strategy=activation_method
+                    )
+
+                    # Get classifier prediction
+                    features = activation_obj.extract_features_for_classifier()
+                    features_numpy = features.cpu().numpy()
+
+                    # Get prediction probability
+                    try:
+                        prediction_proba = classifier.predict_proba([features_numpy])
+                        if isinstance(prediction_proba, (list, tuple)) and len(prediction_proba) > 0:
+                            prediction = float(prediction_proba[0])
+                        else:
+                            prediction = float(prediction_proba)
+                    except:
+                        predictions = classifier.predict([features_numpy])
+                        if len(predictions) > 0:
+                            prediction = float(predictions[0])
+                        else:
+                            logger.warning("Classifier predict returned empty array")
+                            prediction = 0.5
+
+                    # Check if code passed tests
+                    code_passed = False
+                    if i < len(evaluation_results.get("execution_results", [])):
+                        sample_results = evaluation_results["execution_results"][i].get("results", [])
+                        if sample_results:
+                            code_passed = sample_results[0].get("passed", False)
+
+                    classification_results.append(
+                        {"classifier_score": prediction, "code_passed": code_passed, "code_snippet": code[:200]}
+                    )
+
+                except Exception as e:
+                    logger.error(f"Error classifying generated code {i}: {e}")
+                    classification_results.append({"classifier_score": 0.5, "code_passed": False, "error": str(e)})
+
+            # Analyze classifier performance
+            correct_predictions = 0
+            for result in classification_results:
+                # Classifier should predict high score (>0.5) for passing code
+                if (result["classifier_score"] > 0.5 and result["code_passed"]) or (
+                    result["classifier_score"] <= 0.5 and not result["code_passed"]
+                ):
+                    correct_predictions += 1
+
+            classifier_accuracy = correct_predictions / len(classification_results) if classification_results else 0.0
+
+            return {
+                "ground_truth": "CODE_EXECUTION",
+                "method_used": "bigcode-evaluation",
+                "confidence": classifier_accuracy,
+                "pass_rate": pass_rate,
+                "classifier_accuracy": classifier_accuracy,
+                "total_samples": len(generated_codes),
+                "passing_samples": int(pass_rate * len(generated_codes)),
+                "details": f"Pass@1: {pass_rate:.2%}, Classifier accuracy: {classifier_accuracy:.2%}",
+                "task_name": task_name,
+                "evaluation_method": "code-execution",
+                "execution_results": evaluation_results,
+            }
+
+        except Exception as e:
+            logger.error(f"Error in code execution evaluation: {e}")
+            import traceback
+
+            traceback.print_exc()
+            return {
+                "ground_truth": "ERROR",
+                "method_used": "code-execution-error",
+                "confidence": 0.0,
+                "details": f"Code execution evaluation failed: {e!s}",
+                "task_name": task_name,
+                "evaluation_method": "code-execution",
+            }