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

wisent/core/log_likelihoods_evaluator.py

@@ -0,0 +1,321 @@
+"""
+Log-Likelihoods Ground Truth Evaluator
+
+This module handles ground truth evaluation for log-likelihoods based tasks,
+typically used for multiple choice questions. Instead of generating text,
+it loads the multiple choice options from lm-eval tasks and runs the classifier
+directly on each choice to evaluate performance against known ground truth.
+"""
+
+import logging
+from typing import Any, Dict, Optional
+
+from wisent_guard.core.activations import ActivationAggregationStrategy, Activations
+from wisent_guard.core.layer import Layer
+
+logger = logging.getLogger(__name__)
+
+
+class LogLikelihoodsEvaluator:
+    """
+    Evaluator for log-likelihoods based ground truth assessment.
+
+    This evaluator loads multiple choice options from lm-eval tasks and runs
+    the classifier on each choice to evaluate performance against known ground truth.
+    No text generation is performed - only direct classification evaluation.
+    """
+
+    def __init__(self, task_name: Optional[str] = None, model=None):
+        """
+        Initialize the log-likelihoods evaluator.
+
+        Args:
+            task_name: Name of the task (e.g., "truthfulqa_mc1", "mmlu", etc.)
+            model: The model instance used to extract activations
+        """
+        self.task_name = task_name
+        self.model = model
+
+    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 a log-likelihoods task by running it on multiple choice options.
+
+        Args:
+            classifier: The classifier to evaluate
+            task_name: Name of the lm-eval task
+            num_samples: Number of samples to evaluate (default: 100)
+            model: The model instance (overrides self.model if provided)
+            layer: Layer to extract activations from (default: 15)
+            token_aggregation: Token aggregation method ("average", "final", "first", "max", "min")
+
+        Returns:
+            Dict containing evaluation results
+        """
+        try:
+            # Use provided model or fall back to self.model
+            evaluation_model = model or self.model
+            if evaluation_model is None:
+                return self._error_result("No model provided for activation extraction")
+
+            logger.info(f"Loading task data for {task_name}...")
+
+            # Use existing task loading infrastructure
+            task_data = evaluation_model.load_lm_eval_task(task_name, shots=0, limit=num_samples)
+            docs, _ = evaluation_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}")
+
+            # Use existing QA extraction infrastructure (task-agnostic)
+            from .contrastive_pairs.contrastive_pair_set import ContrastivePairSet
+
+            qa_pairs = ContrastivePairSet.extract_qa_pairs_from_task_docs(task_name, task_data, docs)
+
+            if not qa_pairs:
+                return self._error_result(f"No QA pairs could be extracted from task: {task_name}")
+
+            logger.info(f"Extracted {len(qa_pairs)} QA pairs from {task_name}")
+
+            # Use existing contrastive pair creation infrastructure
+            from wisent_guard.core.activations.activation_collection_method import (
+                ActivationCollectionLogic,
+            )
+            from wisent_guard.core.activations.prompts import PromptConstructionStrategy
+
+            collector = ActivationCollectionLogic(model=evaluation_model)
+
+            # For evaluation, use DIRECT_COMPLETION instead of MULTIPLE_CHOICE
+            # This creates prompts like "Q" -> "good_resp"/"bad_resp" instead of "Which is better: Q A. bad B. good"
+            logger.info("🔍 EVALUATION MODE: Using DIRECT_COMPLETION prompt strategy instead of MULTIPLE_CHOICE")
+            contrastive_pairs = collector.create_batch_contrastive_pairs(
+                qa_pairs, prompt_strategy=PromptConstructionStrategy.DIRECT_COMPLETION
+            )
+
+            if not contrastive_pairs:
+                return self._error_result("No contrastive pairs could be created from QA pairs")
+
+            logger.info(f"Created {len(contrastive_pairs)} contrastive pairs")
+
+            # Map token aggregation to token targeting strategy for evaluation
+            targeting_strategy_mapping = {  # TODO Refactor - we should stay with one standard
+                "average": ActivationAggregationStrategy.MEAN_POOLING,
+                "final": ActivationAggregationStrategy.LAST_TOKEN,
+                "first": ActivationAggregationStrategy.FIRST_TOKEN,
+                "max": ActivationAggregationStrategy.MAX_POOLING,
+                "min": ActivationAggregationStrategy.MEAN_POOLING,  # Fallback to mean
+            }
+
+            targeting_strategy = targeting_strategy_mapping.get(
+                token_aggregation, ActivationAggregationStrategy.MEAN_POOLING
+            )
+
+            logger.info(
+                f"🔍 EVALUATION MODE: Using {targeting_strategy.value} targeting strategy (from token_aggregation: {token_aggregation})"
+            )
+            logger.info("🎯 ACTIVATION COLLECTION PARAMS:")
+            logger.info(f"   • Layer: {layer}")
+            logger.info(f"   • Device: {evaluation_model.device}")
+            logger.info(f"   • Token targeting: {targeting_strategy.value}")
+            logger.info(f"   • Pairs count: {len(contrastive_pairs)}")
+
+            processed_pairs = collector.collect_activations_batch(
+                pairs=contrastive_pairs,
+                layer_index=layer,
+                device=evaluation_model.device,
+                token_targeting_strategy=targeting_strategy,
+            )
+
+            if not processed_pairs:
+                return self._error_result("No activations could be extracted from contrastive pairs")
+
+            logger.info(f"Extracted activations from {len(processed_pairs)} pairs")
+
+            # Debug: Show where activations are collected from
+            if processed_pairs:
+                sample_pair = processed_pairs[0]
+                logger.info("📍 DETAILED ACTIVATION COLLECTION ANALYSIS:")
+                logger.info(f"   🔧 Sample pair type: {type(sample_pair).__name__}")
+                logger.info(
+                    f"   🔧 Pair attributes: {[attr for attr in dir(sample_pair) if not attr.startswith('_')][:8]}..."
+                )
+
+                if hasattr(sample_pair, "positive_activations") and sample_pair.positive_activations is not None:
+                    logger.info(f"   ✅ Positive activations shape: {sample_pair.positive_activations.shape}")
+                if hasattr(sample_pair, "negative_activations") and sample_pair.negative_activations is not None:
+                    logger.info(f"   ✅ Negative activations shape: {sample_pair.negative_activations.shape}")
+
+                if hasattr(sample_pair, "_prompt_pair") and sample_pair._prompt_pair:
+                    logger.debug(f"   🔸 Positive prompt: {sample_pair._prompt_pair.positive_prompt[:100]}...")
+                    logger.debug(f"   🔸 Negative prompt: {sample_pair._prompt_pair.negative_prompt[:100]}...")
+                    logger.debug(f"   🎯 Target token: {sample_pair._prompt_pair.target_token}")
+                    logger.debug(f"   📊 Prompt strategy: {sample_pair._prompt_strategy.value}")
+                    logger.info(f"   🔍 Token targeting: {targeting_strategy.value} (evaluation mode)")
+                elif hasattr(sample_pair, "prompt") and hasattr(sample_pair, "positive_response"):
+                    logger.debug(f"   🔸 Question prompt: {sample_pair.prompt[:100]}...")
+                    logger.debug(f"   ✅ Positive response: {sample_pair.positive_response[:50]}...")
+                    logger.debug(f"   ❌ Negative response: {sample_pair.negative_response[:50]}...")
+                    logger.debug(
+                        f"   🔍 Token targeting used: {targeting_strategy.value} (from CLI token_aggregation: {token_aggregation})"
+                    )
+                else:
+                    logger.info("   📍 ACTIVATION COLLECTION: Unknown format - investigating...")
+                    logger.info(
+                        f"   🔧 All attributes: {[attr for attr in dir(sample_pair) if not attr.startswith('__')]}"
+                    )
+
+            # Map token aggregation to activation method
+            activation_method = token_aggregation
+            # Handle both string and enum types
+            method_name = activation_method.value if hasattr(activation_method, 'value') else str(activation_method)
+            logger.info(
+                f"🎯 Using activation aggregation method: {method_name} (from token_aggregation: {token_aggregation})"
+            )
+
+            # Evaluate classifier on each sample
+            results = []
+            total_correct = 0
+            total_samples = 0
+
+            for i, pair in enumerate(processed_pairs):
+                try:
+                    sample_result = self._evaluate_classifier_on_sample(
+                        classifier, pair, qa_pairs[i], activation_method
+                    )
+                    results.append(sample_result)
+
+                    if sample_result.get("classifier_correct", False):
+                        total_correct += 1
+                    total_samples += 1
+
+                except Exception as e:
+                    logger.error(f"Error evaluating sample {i}: {e}")
+                    continue
+
+            # Calculate overall metrics
+            accuracy = total_correct / total_samples if total_samples > 0 else 0.0
+
+            return {
+                "ground_truth": "EVALUATED",
+                "method_used": "log-likelihoods-classifier",
+                "confidence": accuracy,
+                "details": f"Evaluated {total_samples} samples with {total_correct} correct predictions",
+                "task_name": task_name,
+                "evaluation_method": "log-likelihoods",
+                "lm_eval_metrics": {
+                    "accuracy": accuracy,
+                    "correct_predictions": total_correct,
+                    "total_samples": total_samples,
+                },
+                "sample_results": results[:10],  # First 10 for debugging
+            }
+
+        except Exception as e:
+            import traceback
+
+            logger.error(f"Error evaluating classifier on task {task_name}: {e}")
+            logger.error(f"Traceback: {traceback.format_exc()}")
+            return self._error_result(f"Evaluation error: {e!s}")
+
+    def _evaluate_classifier_on_sample(
+        self, classifier, processed_pair, qa_pair: Dict[str, Any], activation_method
+    ) -> Dict[str, Any]:
+        """
+        Evaluate the classifier on a single processed contrastive pair.
+
+        Args:
+            classifier: The classifier to evaluate
+            processed_pair: ContrastivePair with activations already extracted
+            qa_pair: Original QA pair data for reference
+            activation_method:
+        Returns:
+            Dict containing evaluation results for this sample
+        """
+        try:
+            # Extract activations from the processed pair
+            positive_activations = processed_pair.positive_activations  # B choice (correct)
+            negative_activations = processed_pair.negative_activations  # A choice (incorrect)
+
+            if positive_activations is None or negative_activations is None:
+                return {
+                    "question": qa_pair["question"],
+                    "correct_answer": qa_pair["correct_answer"],
+                    "incorrect_answer": qa_pair["incorrect_answer"],
+                    "classifier_correct": False,
+                    "error": "Missing activations",
+                }
+
+            layer_obj = Layer(index=15, type="transformer")
+
+            # Process positive (correct) choice using CLI token aggregation method
+            positive_act = Activations(
+                tensor=positive_activations, layer=layer_obj, aggregation_strategy=activation_method
+            )
+            positive_features = positive_act.extract_features_for_classifier()
+            positive_prediction = classifier.predict_proba([positive_features.cpu().numpy()])
+            if isinstance(positive_prediction, (list, tuple)) and len(positive_prediction) > 0:
+                positive_prediction = positive_prediction[0]
+
+            # Process negative (incorrect) choice using CLI token aggregation method
+            negative_act = Activations(
+                tensor=negative_activations, layer=layer_obj, aggregation_strategy=activation_method
+            )
+            negative_features = negative_act.extract_features_for_classifier()
+            negative_prediction = classifier.predict_proba([negative_features.cpu().numpy()])
+            if isinstance(negative_prediction, (list, tuple)) and len(negative_prediction) > 0:
+                negative_prediction = negative_prediction[0]
+
+            # Evaluate classifier performance
+            # For truthfulness: lower scores = more truthful/correct
+            # Classifier should give lower score to correct answer, higher to incorrect
+            classifier_correct = positive_prediction < negative_prediction
+
+            # Enhanced logging to show classifier vs normal predictions
+            logger.info("📋 Sample Evaluation:")
+            logger.debug(f"   🔸 Question: {qa_pair['question'][:100]}...")
+            logger.info("   🧠 CLASSIFIER PREDICTION:")
+            logger.debug(f"      ✅ Correct answer score: {float(positive_prediction):.3f} (lower = more truthful)")
+            logger.debug(f"      ❌ Incorrect answer score: {float(negative_prediction):.3f} (higher = less truthful)")
+            logger.debug(
+                f"      🎯 Classifier judgment: {'CORRECT' if classifier_correct else 'INCORRECT'} (positive {'<' if classifier_correct else '>='} negative)"
+            )
+
+            return {
+                "question": qa_pair["question"],
+                "correct_answer": qa_pair["correct_answer"],
+                "incorrect_answer": qa_pair["incorrect_answer"],
+                "positive_prediction": float(positive_prediction),
+                "negative_prediction": float(negative_prediction),
+                "classifier_correct": classifier_correct,
+            }
+
+        except Exception as e:
+            logger.debug(f"Error evaluating sample: {e}")
+            return {
+                "question": qa_pair.get("question", "Unknown"),
+                "correct_answer": qa_pair.get("correct_answer", "Unknown"),
+                "incorrect_answer": qa_pair.get("incorrect_answer", "Unknown"),
+                "classifier_correct": False,
+                "error": str(e),
+            }
+
+    def _error_result(self, error_msg: str) -> Dict[str, Any]:
+        """Return an error result."""
+        return {
+            "ground_truth": "UNKNOWN",
+            "method_used": "log-likelihoods-error",
+            "confidence": 0.0,
+            "details": error_msg,
+            "task_name": self.task_name or "unknown",
+            "evaluation_method": "log-likelihoods",
+            "lm_eval_metrics": {"accuracy": 0.0, "correct_predictions": 0, "total_samples": 0},
+        }