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

wisent/core/optuna/steering/metrics.py

@@ -0,0 +1,474 @@
+"""
+Evaluation metrics for comprehensive evaluation pipeline.
+"""
+
+import logging
+from typing import Any, Callable, Dict, List, Optional
+
+import numpy as np
+from sklearn.metrics import accuracy_score, precision_recall_fscore_support, roc_auc_score
+
+from wisent_guard.core.bigcode_extractors import MBPPExtractor
+
+# Import LMEvalHarnessGroundTruth for intelligent evaluation (newer approach used by CLI)
+from wisent_guard.core.lm_eval_harness_ground_truth import LMEvalHarnessGroundTruth
+from wisent_guard.core.task_interface import get_task
+from wisent_guard.core.tasks.file_task import FileTask
+from wisent_guard.parameters.task_config import CODING_TASKS
+
+from .bigcode_evaluator_wrapper import OptunaBigCodeEvaluator
+
+logger = logging.getLogger(__name__)
+
+
+def evaluate_response_correctness(response: str, expected_answer: str, task_name: str) -> bool:
+    """
+    Evaluate if a response is correct using LMEvalHarnessGroundTruth (same approach as CLI).
+    Note: For coding tasks, response should already be extracted code before calling this function.
+
+    Args:
+        response: Model's response (pre-extracted code for coding tasks)
+        expected_answer: Expected correct answer
+        task_name: Name of the task for proper evaluation
+
+    Returns:
+        True if response is correct, False otherwise
+    """
+    # Check if this is a file-based task (custom dataset loaded from JSON)
+    # For file-based tasks, use exact string matching to avoid false positives
+    try:
+        task = get_task(task_name, limit=1)
+        if isinstance(task, FileTask):
+            logger.debug(f"Using exact match for file-based task '{task_name}'")
+            return response.strip().lower() == expected_answer.strip().lower()
+    except:
+        pass  # Continue with normal evaluation if task lookup fails
+
+    try:
+        # Use the same evaluation approach as the CLI
+        evaluator = LMEvalHarnessGroundTruth(task_name)
+
+        # Create response data format expected by _evaluate_with_lm_eval_metrics
+        response_data = [
+            {
+                "generated_response": response,
+                "ground_truth": expected_answer,
+                "question": "evaluation_question",  # Required field for evaluation
+            }
+        ]
+
+        # Use the same evaluation logic as CLI
+        eval_results = evaluator._evaluate_with_lm_eval_metrics(task_name, response_data, None)
+
+        # Extract the result - accuracy > 0 means at least one correct
+        return eval_results.get("accuracy", 0.0) > 0.0
+
+    except Exception as e:
+        logger.warning(f"LMEvalHarnessGroundTruth failed, using exact match fallback: {e}")
+        # Fallback to simple string matching
+        return response.strip().lower() == expected_answer.strip().lower()
+
+
+def evaluate_benchmark_performance(
+    predictions: List[str],
+    ground_truths: List[str],
+    task_name: str = None,
+    task_docs: List[Dict] = None,
+    classifier_scorer: Optional[Callable[[List[str], str], List[float]]] = None,
+) -> Dict[str, float]:
+    """
+    Evaluate benchmark performance using LMEvalHarnessGroundTruth (same approach as CLI).
+    For coding tasks, uses BigCode execution-based evaluation instead of string comparison.
+
+    Args:
+        predictions: List of model predictions
+        ground_truths: List of correct answers
+        task_name: Name of the task for intelligent evaluation
+        task_docs: List of original task documents (required for coding tasks)
+        classifier_scorer: Optional function to score predictions with classifier for confidence scores
+
+    Returns:
+        Dictionary containing benchmark performance metrics
+    """
+    if task_name:
+        # Check if this is a coding task that requires code execution evaluation
+        is_coding_task = task_name.lower() in CODING_TASKS
+
+        # Calculate classifier confidence scores if classifier_scorer provided
+        classifier_confidences = None
+        if classifier_scorer is not None:
+            try:
+                logger.debug(f"Calculating classifier confidence scores for {len(predictions)} predictions")
+                classifier_confidences = classifier_scorer(predictions, f"metrics_evaluation_{task_name}")
+                logger.debug(f"Calculated {len(classifier_confidences)} confidence scores")
+            except Exception as e:
+                logger.warning(f"Failed to calculate classifier confidence scores: {e}")
+                classifier_confidences = None
+
+        if is_coding_task:
+            # Use BigCode execution-based evaluation for coding tasks
+            logger.info(f"Using BigCode execution-based evaluation for coding task: {task_name}")
+
+            try:
+                bigcode_evaluator = OptunaBigCodeEvaluator()
+
+                # Validate task docs are provided for coding tasks
+                if task_docs is None or len(task_docs) == 0:
+                    logger.error(
+                        f"No task docs provided for coding task {task_name}. BigCode evaluation requires original task documents with test cases."
+                    )
+                    raise ValueError(f"Task documents required for coding task evaluation: {task_name}")
+
+                # Ensure we have the right number of task docs
+                if len(task_docs) != len(predictions):
+                    logger.error(f"Task docs length mismatch: {len(task_docs)} docs vs {len(predictions)} predictions")
+                    raise ValueError(
+                        f"Number of task documents ({len(task_docs)}) must match number of predictions ({len(predictions)})"
+                    )
+
+                # Evaluate using BigCode execution
+                evaluation_results, accuracy_metrics = bigcode_evaluator.evaluate_and_calculate_accuracy(
+                    predictions, task_docs, task_name
+                )
+
+                # Create evaluation details in the expected format
+                evaluation_details = []
+                for i, (pred, result) in enumerate(zip(predictions, evaluation_results)):
+                    eval_detail = {
+                        "prediction": result.get("extracted_code", pred),
+                        "ground_truth": ground_truths[i] if i < len(ground_truths) else "unknown",
+                        "is_correct": result.get("passed", False),
+                        "classifier_confidence": classifier_confidences[i]
+                        if classifier_confidences and i < len(classifier_confidences)
+                        else 1.0,
+                        "method": "bigcode_execution",
+                        "original_prediction": pred,
+                        "code_extracted": result.get("extracted_code", "") != pred,
+                        "execution_error": result.get("error"),
+                    }
+                    evaluation_details.append(eval_detail)
+
+                return {
+                    "accuracy": accuracy_metrics["accuracy"],
+                    "total_samples": accuracy_metrics["total_samples"],
+                    "correct": accuracy_metrics["correct"],
+                    "incorrect": accuracy_metrics["incorrect"],
+                    "evaluation_method": "bigcode_execution",
+                    "task_name": task_name,
+                    "evaluation_details": evaluation_details,
+                    "pass_count": accuracy_metrics.get("pass_count", 0),
+                    "fail_count": accuracy_metrics.get("fail_count", 0),
+                    "error_count": accuracy_metrics.get("error_count", 0),
+                }
+
+            except Exception as e:
+                logger.error(f"BigCode evaluation failed for {task_name}, falling back to string-based evaluation: {e}")
+                # Fall through to string-based evaluation
+
+        # String-based evaluation for non-coding tasks or BigCode fallback
+        extracted_predictions = predictions
+
+        if is_coding_task:
+            # Extract code from predictions for coding tasks (fallback mode)
+            extractor = MBPPExtractor()  # Works for all coding tasks, not just MBPP
+            extracted_predictions = []
+
+            for pred in predictions:
+                extracted_code = extractor.extract_code_from_answer(pred)
+                extracted_predictions.append(extracted_code)
+
+            logger.debug(f"Code extraction applied for {task_name}: {len(predictions)} predictions processed")
+
+        # Use intelligent evaluation with LMEvalHarnessGroundTruth (same as CLI)
+        correct_predictions = []
+        evaluation_details = []
+
+        for i, (orig_pred, extracted_pred, gt) in enumerate(zip(predictions, extracted_predictions, ground_truths)):
+            try:
+                # Use the extracted prediction for evaluation
+                is_correct = evaluate_response_correctness(extracted_pred, gt, task_name)
+                correct_predictions.append(is_correct)
+
+                # Include both original and extracted predictions in details for debugging
+                eval_detail = {
+                    "prediction": extracted_pred,
+                    "ground_truth": gt,
+                    "is_correct": is_correct,
+                    "classifier_confidence": classifier_confidences[i]
+                    if classifier_confidences and i < len(classifier_confidences)
+                    else 1.0,
+                    "method": "lm_eval_harness_ground_truth",
+                }
+
+                # Add original prediction for coding tasks to help with debugging
+                if is_coding_task and orig_pred != extracted_pred:
+                    eval_detail["original_prediction"] = orig_pred
+                    eval_detail["code_extracted"] = True
+
+                evaluation_details.append(eval_detail)
+
+            except Exception as e:
+                logger.warning(f"LMEvalHarnessGroundTruth failed for prediction '{extracted_pred}' vs '{gt}': {e}")
+                # Fallback to simple string matching
+                is_correct = extracted_pred.strip().lower() == gt.strip().lower()
+                correct_predictions.append(is_correct)
+
+                eval_detail = {
+                    "prediction": extracted_pred,
+                    "ground_truth": gt,
+                    "is_correct": is_correct,
+                    "classifier_confidence": classifier_confidences[i]
+                    if classifier_confidences and i < len(classifier_confidences)
+                    else 1.0,
+                    "method": "fallback_exact_match",
+                }
+
+                if is_coding_task and orig_pred != extracted_pred:
+                    eval_detail["original_prediction"] = orig_pred
+                    eval_detail["code_extracted"] = True
+
+                evaluation_details.append(eval_detail)
+
+        accuracy = np.mean(correct_predictions)
+        total_correct = sum(correct_predictions)
+
+        return {
+            "accuracy": accuracy,
+            "total_samples": len(predictions),
+            "correct": total_correct,
+            "incorrect": len(predictions) - total_correct,
+            "evaluation_method": "lm_eval_harness_ground_truth",
+            "task_name": task_name,
+            "evaluation_details": evaluation_details[:5],  # Include first 5 for debugging
+        }
+    # Fallback to simple exact match
+    logger.info("No task_name provided, using simple exact match evaluation")
+    exact_matches = [pred.strip().lower() == gt.strip().lower() for pred, gt in zip(predictions, ground_truths)]
+    accuracy = np.mean(exact_matches)
+
+    return {
+        "accuracy": accuracy,
+        "total_samples": len(predictions),
+        "correct": sum(exact_matches),
+        "incorrect": len(predictions) - sum(exact_matches),
+        "evaluation_method": "exact_match",
+    }
+
+
+def evaluate_probe_performance(y_true: np.ndarray, y_pred: np.ndarray, y_pred_proba: np.ndarray) -> Dict[str, float]:
+    """
+    Evaluate probe performance with comprehensive metrics.
+
+    Args:
+        y_true: True labels
+        y_pred: Predicted labels
+        y_pred_proba: Predicted probabilities (for positive class)
+
+    Returns:
+        Dictionary containing probe performance metrics
+    """
+    if len(y_true) == 0:
+        # Return default metrics if no data
+        return {"accuracy": 0.5, "precision": 0.5, "recall": 0.5, "f1": 0.5, "auc": 0.5, "total_samples": 0}
+
+    accuracy = accuracy_score(y_true, y_pred)
+    precision, recall, f1, _ = precision_recall_fscore_support(y_true, y_pred, average="binary")
+
+    try:
+        auc = roc_auc_score(y_true, y_pred_proba)
+    except:
+        auc = 0.5  # Default for cases where AUC can't be computed
+
+    return {
+        "accuracy": accuracy,
+        "precision": precision,
+        "recall": recall,
+        "f1": f1,
+        "auc": auc,
+        "total_samples": len(y_true),
+    }
+
+
+def calculate_combined_score(
+    benchmark_metrics: Dict[str, float],
+    probe_metrics: Dict[str, float],
+    benchmark_weight: float = 0.7,
+    probe_weight: float = 0.3,
+) -> float:
+    """
+    Calculate combined score from benchmark and probe performance.
+
+    Args:
+        benchmark_metrics: Benchmark performance metrics
+        probe_metrics: Probe performance metrics
+        benchmark_weight: Weight for benchmark performance
+        probe_weight: Weight for probe performance
+
+    Returns:
+        Combined score (0-1)
+    """
+    benchmark_score = benchmark_metrics.get("accuracy", 0.0)
+    probe_score = probe_metrics.get("auc", 0.5)  # Use AUC as primary probe metric
+
+    combined_score = benchmark_weight * benchmark_score + probe_weight * probe_score
+    return combined_score
+
+
+def calculate_comprehensive_metrics(results: Dict[str, Any]) -> Dict[str, Any]:
+    """
+    Calculate comprehensive metrics from evaluation results.
+
+    Args:
+        results: Complete evaluation results
+
+    Returns:
+        Dictionary with comprehensive metrics and analysis
+    """
+    comprehensive_metrics = {}
+
+    if "test_results" in results:
+        test_results = results["test_results"]
+
+        # Extract key metrics
+        base_benchmark_acc = test_results.get("base_benchmark_metrics", {}).get("accuracy", 0.0)
+        steered_benchmark_acc = test_results.get("steered_benchmark_metrics", {}).get("accuracy", 0.0)
+        base_probe_auc = test_results.get("base_probe_metrics", {}).get("auc", 0.5)
+        steered_probe_auc = test_results.get("steered_probe_metrics", {}).get("auc", 0.5)
+
+        # Calculate improvements
+        benchmark_improvement = steered_benchmark_acc - base_benchmark_acc
+        probe_improvement = steered_probe_auc - base_probe_auc
+
+        comprehensive_metrics.update(
+            {
+                "base_benchmark_accuracy": base_benchmark_acc,
+                "steered_benchmark_accuracy": steered_benchmark_acc,
+                "benchmark_improvement": benchmark_improvement,
+                "benchmark_improvement_percent": (benchmark_improvement / max(base_benchmark_acc, 0.001)) * 100,
+                "base_probe_auc": base_probe_auc,
+                "steered_probe_auc": steered_probe_auc,
+                "probe_improvement": probe_improvement,
+                "probe_improvement_percent": (probe_improvement / max(base_probe_auc, 0.001)) * 100,
+                "overall_effectiveness": (benchmark_improvement + probe_improvement) / 2,
+                "validation_score": test_results.get("validation_combined_score", 0.0),
+            }
+        )
+
+    # Add training statistics
+    if "probe_training_results" in results:
+        training_results = results["probe_training_results"]
+
+        # Calculate training performance statistics
+        all_training_aucs = []
+        for layer_key, layer_results in training_results.items():
+            for c_key, metrics in layer_results.items():
+                if isinstance(metrics, dict) and "auc" in metrics:
+                    all_training_aucs.append(metrics["auc"])
+
+        if all_training_aucs:
+            comprehensive_metrics.update(
+                {
+                    "training_probe_auc_mean": np.mean(all_training_aucs),
+                    "training_probe_auc_std": np.std(all_training_aucs),
+                    "training_probe_auc_max": np.max(all_training_aucs),
+                    "training_probe_auc_min": np.min(all_training_aucs),
+                }
+            )
+
+    # Add optimization statistics
+    if "steering_optimization_results" in results:
+        optimization_results = results["steering_optimization_results"]
+
+        all_configs = optimization_results.get("all_configs", [])
+        if all_configs:
+            combined_scores = [config.get("combined_score", 0.0) for config in all_configs]
+            benchmark_scores = [config.get("benchmark_metrics", {}).get("accuracy", 0.0) for config in all_configs]
+
+            comprehensive_metrics.update(
+                {
+                    "optimization_configs_tested": len(all_configs),
+                    "optimization_score_mean": np.mean(combined_scores),
+                    "optimization_score_std": np.std(combined_scores),
+                    "optimization_benchmark_mean": np.mean(benchmark_scores),
+                    "optimization_benchmark_std": np.std(benchmark_scores),
+                }
+            )
+
+    return comprehensive_metrics
+
+
+def generate_performance_summary(comprehensive_metrics: Dict[str, Any]) -> str:
+    """
+    Generate a human-readable performance summary.
+
+    Args:
+        comprehensive_metrics: Comprehensive metrics dictionary
+
+    Returns:
+        String summary of performance
+    """
+    summary = []
+    summary.append("=" * 60)
+    summary.append("COMPREHENSIVE EVALUATION PERFORMANCE SUMMARY")
+    summary.append("=" * 60)
+
+    # Benchmark Performance
+    if "base_benchmark_accuracy" in comprehensive_metrics:
+        base_acc = comprehensive_metrics["base_benchmark_accuracy"]
+        steered_acc = comprehensive_metrics["steered_benchmark_accuracy"]
+        improvement = comprehensive_metrics["benchmark_improvement"]
+
+        summary.append("\n📊 BENCHMARK PERFORMANCE:")
+        summary.append(f"  Base Model Accuracy:    {base_acc:.3f} ({base_acc * 100:.1f}%)")
+        summary.append(f"  Steered Model Accuracy: {steered_acc:.3f} ({steered_acc * 100:.1f}%)")
+        summary.append(f"  Improvement:            {improvement:+.3f} ({improvement * 100:+.1f}%)")
+
+    # Probe Performance
+    if "base_probe_auc" in comprehensive_metrics:
+        base_auc = comprehensive_metrics["base_probe_auc"]
+        steered_auc = comprehensive_metrics["steered_probe_auc"]
+        improvement = comprehensive_metrics["probe_improvement"]
+
+        summary.append("\n🔍 PROBE PERFORMANCE:")
+        summary.append(f"  Base Model Probe AUC:    {base_auc:.3f}")
+        summary.append(f"  Steered Model Probe AUC: {steered_auc:.3f}")
+        summary.append(f"  Improvement:             {improvement:+.3f}")
+
+    # Training Statistics
+    if "training_probe_auc_mean" in comprehensive_metrics:
+        mean_auc = comprehensive_metrics["training_probe_auc_mean"]
+        std_auc = comprehensive_metrics["training_probe_auc_std"]
+        max_auc = comprehensive_metrics["training_probe_auc_max"]
+
+        summary.append("\n🎯 TRAINING STATISTICS:")
+        summary.append(f"  Probe Training AUC:      {mean_auc:.3f} ± {std_auc:.3f}")
+        summary.append(f"  Best Training AUC:       {max_auc:.3f}")
+
+    # Optimization Statistics
+    if "optimization_configs_tested" in comprehensive_metrics:
+        num_configs = comprehensive_metrics["optimization_configs_tested"]
+        best_score = comprehensive_metrics.get("validation_score", 0.0)
+
+        summary.append("\n⚙️ OPTIMIZATION STATISTICS:")
+        summary.append(f"  Configurations Tested:   {num_configs}")
+        summary.append(f"  Best Validation Score:   {best_score:.3f}")
+
+    # Overall Assessment
+    if "overall_effectiveness" in comprehensive_metrics:
+        effectiveness = comprehensive_metrics["overall_effectiveness"]
+
+        summary.append("\n🏆 OVERALL ASSESSMENT:")
+        if effectiveness > 0.1:
+            assessment = "Highly Effective"
+        elif effectiveness > 0.05:
+            assessment = "Moderately Effective"
+        elif effectiveness > 0.01:
+            assessment = "Slightly Effective"
+        else:
+            assessment = "Minimal Effect"
+
+        summary.append(f"  Steering Effectiveness:  {assessment} ({effectiveness:+.3f})")
+
+    summary.append("=" * 60)
+
+    return "\n".join(summary)