File size: 46,738 Bytes
a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 a938ec7 4466506 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 | # Dependencies
import os
import sys
import json
import time
import argparse
import numpy as np
import pandas as pd
import seaborn as sns
from tqdm import tqdm
from typing import Any
from typing import List
from typing import Dict
from scipy import stats
from pathlib import Path
from typing import Tuple
from datetime import datetime
from dataclasses import asdict
import matplotlib.pyplot as plt
from dataclasses import dataclass
from collections import defaultdict
from sklearn.metrics import roc_auc_score
from sklearn.metrics import confusion_matrix
from sklearn.metrics import average_precision_score
from sklearn.metrics import precision_recall_fscore_support
# Add parent directory to path for imports
sys.path.append(str(Path(__file__).parent.parent))
from config.enums import Domain
from services.orchestrator import DetectionOrchestrator
@dataclass
class EvaluationSample:
"""
Single evaluation sample
"""
text_id : str
domain : str
ground_truth : str # "human" or "ai"
text : str
file_path : str
subset : str # "clean", "paraphrased", "cross_model"
@dataclass
class SingleEvalResult:
"""
Result for a single sample
"""
text_id : str
domain : str
ground_truth : str # "human" or "ai"
prediction : str # "human", "ai", "hybrid", or "uncertain"
verdict : str # Raw verdict from system: "Synthetically-Generated", "Authentically-Written", "Hybrid", "Uncertain"
synthetic_prob : float
authentic_prob : float
hybrid_prob : float
confidence : float
uncertainty : float
processing_time : float
is_correct : bool | None
subset : str
word_count : int # For length analysis
@dataclass
class AggregatedMetrics:
"""
Aggregated performance metrics for 4-class system
"""
# Binary metrics (AI vs Human on decisive predictions)
precision : float
recall : float
f1 : float
accuracy : float
# Coverage metrics
coverage : float # % samples with decisive prediction (not uncertain)
accuracy_at_cov : float # Accuracy on non-uncertain predictions
f1_at_cov : float # F1 on non-uncertain predictions
# Probability-based metrics
auroc : float
auprc : float
ece : float # Expected Calibration Error
# 4-class specific metrics
hybrid_detection_rate : float # % of AI samples classified as Hybrid
abstention_rate : float # % classified as Uncertain
# Confusion metrics
confusion_matrix : List[List[int]] # 2x2 for decisive predictions only
support : Dict[str, int]
# 4-class breakdown
verdict_distribution : Dict[str, int] # Count of each verdict type
@dataclass
class LengthBucketMetrics:
"""
Performance metrics for a specific text length range
"""
min_words : int
max_words : int
label : str
sample_count : int
precision : float
recall : float
f1 : float
accuracy : float
mean_confidence : float
fp_rate : float
fn_rate : float
avg_proc_time : float
abstention_rate : float
class TextAuthEvaluator:
"""
Comprehensive evaluation framework for TEXT-AUTH (4-class system)
Handles verdicts:
- "Synthetically-Generated" → prediction = "ai"
- "Authentically-Written" → prediction = "human"
- "Hybrid" → prediction = "hybrid"
- "Uncertain" → prediction = "uncertain"
"""
def __init__(self, dataset_path: str = "evaluation", output_dir: str = "evaluation/results"):
"""
Initialize evaluator
Arguments:
----------
dataset_path { str } : Path to evaluation directory
output_dir { str } : Directory to save results
"""
self.dataset_path = Path(dataset_path)
self.output_dir = Path(output_dir)
self.output_dir.mkdir(exist_ok = True,
parents = True,
)
# Initialize orchestrator
print("\nInitializing TEXT-AUTH Detection Orchestrator...")
self.orchestrator = DetectionOrchestrator.create_with_executor(max_workers = 4,
enable_language_detection = False,
parallel_execution = True,
skip_expensive_metrics = False,
)
if not self.orchestrator.initialize():
raise RuntimeError("Failed to initialize detection orchestrator")
print("\nOrchestrator initialized successfully\n")
# Storage for results
self.results = list()
self.metadata = dict()
# Load metadata if available
self._load_metadata()
def _load_metadata(self):
"""
Load dataset metadata
"""
metadata_path = self.dataset_path / "metadata.json"
if metadata_path.exists():
with open(metadata_path, 'r') as f:
self.metadata = json.load(f)
print(f"\nDataset: {self.metadata.get('dataset_name', 'Unknown')}")
print(f" Version: {self.metadata.get('version', 'Unknown')}")
print(f" Total samples: {self.metadata.get('total_samples', 'Unknown')}")
print(f" Human: {self.metadata.get('human_samples', 'Unknown')}")
print(f" AI: {self.metadata.get('ai_samples', 'Unknown')}")
if ('challenge_samples' in self.metadata):
challenges = self.metadata['challenge_samples']
print(f" Paraphrased: {challenges.get('paraphrased', 0)}")
print(f" Cross-model: {challenges.get('cross_model', 0)}")
print()
else:
print("\nNo metadata.json found - run create_metadata.py first\n")
def load_dataset(self, domains: List[str] = None, max_samples_per_domain: int = None, subset_filter: str = None) -> List[EvaluationSample]:
"""
Load evaluation dataset
Arguments:
----------
domains { list } : List of domains to evaluate (None = all)
max_samples_per_domain { int } : Limit samples per domain
subset_filter { str } : Only load specific subset
Returns:
--------
{ list } : List of EvaluationSample objects
"""
samples = list()
# Load clean samples (human + ai)
if (subset_filter is None or (subset_filter == "clean")):
for subset_name, subset_dir in [("human", "human"), ("ai", "ai_generated")]:
subset_path = self.dataset_path / subset_dir
if not subset_path.exists():
print(f"Directory not found: {subset_path}")
continue
for domain_dir in subset_path.iterdir():
if not domain_dir.is_dir():
continue
domain = domain_dir.name
if domains and domain not in domains:
continue
files = list(domain_dir.glob("*.txt"))
if max_samples_per_domain:
files = files[:max_samples_per_domain]
for file_path in files:
try:
with open(file_path, 'r', encoding = 'utf-8') as f:
text = f.read()
samples.append(EvaluationSample(text_id = file_path.stem,
domain = domain,
ground_truth = subset_name,
text = text,
file_path = str(file_path),
subset = "clean",
)
)
except Exception as e:
print(f"Error loading {file_path}: {e}")
# Load challenge sets (adversarial)
if subset_filter is None or subset_filter in ["paraphrased", "cross_model"]:
adversarial_path = self.dataset_path / "adversarial"
if adversarial_path.exists():
for challenge_name in ["paraphrased", "cross_model"]:
if subset_filter and subset_filter != challenge_name:
continue
challenge_path = adversarial_path / challenge_name
if not challenge_path.exists():
continue
files = list(challenge_path.glob("*.txt"))
for file_path in files:
try:
with open(file_path, 'r', encoding = 'utf-8') as f:
text = f.read()
# Extract domain from filename
domain = "general"
for possible_domain in ["academic", "creative", "ai_ml", "software_dev", "technical_doc", "engineering", "science", "business", "legal", "medical", "journalism", "marketing", "social_media", "blog_personal", "tutorial", "general"]:
if possible_domain in file_path.stem:
domain = possible_domain
break
# Skip if domain filter active and doesn't match
if domains and domain not in domains:
continue
samples.append(EvaluationSample(text_id = file_path.stem,
domain = domain,
ground_truth = "ai", # All adversarial are AI-generated
text = text,
file_path = str(file_path),
subset = challenge_name,
)
)
except Exception as e:
print(f"Error loading {file_path}: {e}")
print(f"\nLoaded {len(samples)} samples")
return samples
def _map_verdict_to_prediction(self, verdict: str) -> str:
"""
Map system verdict to evaluation prediction class
Arguments:
----------
verdict { str } : Raw verdict from system
Returns:
--------
{ str } : Mapped prediction ("human", "ai", "hybrid", "uncertain")
"""
verdict_lower = verdict.lower()
if (("synthetic" in verdict_lower) or ("generated" in verdict_lower)):
return "ai"
elif (("authentic" in verdict_lower) or ("written" in verdict_lower)):
return "human"
elif ("hybrid" in verdict_lower):
return "hybrid"
else: # "Uncertain" or any other
return "uncertain"
def run_evaluation(self, samples: List[EvaluationSample]):
"""
Run evaluation on all samples
Arguments:
----------
samples { list } : List of EvaluationSample objects
"""
print(f"\nEvaluating {len(samples)} samples...")
print("=" * 70)
for i, sample in enumerate(tqdm(samples, desc = "Processing")):
try:
start_time = time.time()
# Run detection
result = self.orchestrator.analyze(text = sample.text)
proc_time = time.time() - start_time
# Extract results
ensemble = result.ensemble_result
verdict = ensemble.final_verdict
prediction = self._map_verdict_to_prediction(verdict)
synthetic_prob = ensemble.synthetic_probability
authentic_prob = ensemble.authentic_probability
hybrid_prob = ensemble.hybrid_probability
confidence = ensemble.overall_confidence
uncertainty = ensemble.uncertainty_score
word_count = len(sample.text.split())
# Determine correctness (only for decisive predictions)
is_correct = None
if ((prediction == "ai") and (sample.ground_truth == "ai")):
is_correct = True
elif ((prediction == "human") and (sample.ground_truth == "human")):
is_correct = True
elif ((prediction in ["ai", "human"]) and (prediction != sample.ground_truth)):
is_correct = False
# Hybrid is considered correct if ground truth is AI (it detected synthetic content)
elif ((prediction == "hybrid") and (sample.ground_truth == "ai")):
is_correct = True
elif ((prediction == "hybrid") and (sample.ground_truth == "human")):
is_correct = False
# Uncertain predictions are neither correct nor incorrect: they are abstentions and handled separately
# Store result
eval_result = SingleEvalResult(text_id = sample.text_id,
domain = sample.domain,
ground_truth = sample.ground_truth,
prediction = prediction,
verdict = verdict,
synthetic_prob = synthetic_prob,
authentic_prob = authentic_prob,
hybrid_prob = hybrid_prob,
confidence = confidence,
uncertainty = uncertainty,
processing_time = proc_time,
is_correct = is_correct,
subset = sample.subset,
word_count = word_count,
)
self.results.append(eval_result)
except Exception as e:
print(f"\nError processing sample {i}: {e}")
continue
print("\n" + "=" * 70)
print(f"Evaluation complete: {len(self.results)}/{len(samples)} samples processed")
def calculate_metrics(self, domain: str = None, subset: str = None) -> AggregatedMetrics:
"""
Calculate aggregated metrics for 4-class system
Arguments:
----------
domain { str } : Calculate for specific domain only
subset { str } : Calculate for specific subset only
Returns:
--------
{ AggregatedMetrics } : Aggregated metrics
"""
# Filter results
filtered = self.results
if domain:
filtered = [r for r in filtered if r.domain == domain]
if subset:
filtered = [r for r in filtered if r.subset == subset]
if not filtered:
return None
# Separate decisive vs uncertain
decisive = [r for r in filtered if r.prediction != "uncertain"]
uncertain = [r for r in filtered if r.prediction == "uncertain"]
# Calculate coverage
coverage = len(decisive) / len(filtered) if filtered else 0.0
# Verdict distribution
verdict_dist = {"Synthetically-Generated" : sum(1 for r in filtered if r.verdict == "Synthetically-Generated"),
"Authentically-Written" : sum(1 for r in filtered if r.verdict == "Authentically-Written"),
"Hybrid" : sum(1 for r in filtered if r.verdict == "Hybrid"),
"Uncertain" : sum(1 for r in filtered if r.verdict == "Uncertain"),
}
# Binary classification metrics (on decisive predictions only)
if decisive:
# Map predictions to binary (treating hybrid as AI detection)
y_true_binary = [1 if r.ground_truth == "ai" else 0 for r in decisive]
y_pred_binary = [1 if r.prediction in ["ai", "hybrid"] else 0 for r in decisive]
# Calculate metrics
precision, recall, f1, support_array = precision_recall_fscore_support(y_true_binary,
y_pred_binary,
average = 'binary',
pos_label = 1,
zero_division = 0,
)
accuracy = sum(1 for i, r in enumerate(decisive) if y_true_binary[i] == y_pred_binary[i]) / len(decisive)
# Confusion matrix
cm = confusion_matrix(y_true_binary, y_pred_binary)
# Support counts
support = {"human" : sum(1 for r in decisive if r.ground_truth == "human"),
"ai" : sum(1 for r in decisive if r.ground_truth == "ai"),
}
else:
precision = recall = f1 = accuracy = 0.0
cm = [[0, 0], [0, 0]]
support = {"human" : 0, "ai" : 0}
# Probability-based metrics (on all samples with probabilities)
y_true_prob = [1 if r.ground_truth == "ai" else 0 for r in filtered]
y_scores = [r.synthetic_prob for r in filtered]
try:
auroc = roc_auc_score(y_true_prob, y_scores)
except:
auroc = 0.0
try:
auprc = average_precision_score(y_true_prob, y_scores)
except:
auprc = 0.0
# Expected Calibration Error (ECE)
ece = self._calculate_ece(filtered)
# Hybrid-specific metrics
ai_samples = [r for r in filtered if r.ground_truth == "ai"]
hybrid_detection_rate = sum(1 for r in ai_samples if r.prediction == "hybrid") / len(ai_samples) if ai_samples else 0.0
# Abstention rate
abstention_rate = len(uncertain) / len(filtered) if filtered else 0.0
return AggregatedMetrics(precision = precision,
recall = recall,
f1 = f1,
accuracy = accuracy,
coverage = coverage,
accuracy_at_cov = accuracy,
f1_at_cov = f1,
auroc = auroc,
auprc = auprc,
ece = ece,
hybrid_detection_rate = hybrid_detection_rate,
abstention_rate = abstention_rate,
confusion_matrix = cm.tolist(),
support = support,
verdict_distribution = verdict_dist,
)
def _calculate_ece(self, results: List[SingleEvalResult], n_bins: int = 10) -> float:
"""
Calculate Expected Calibration Error
Arguments:
----------
results { list } : List of evaluation results
n_bins { int } : Number of confidence bins
Returns:
--------
{ float } : ECE value
"""
# Only calculate on decisive predictions
decisive = [r for r in results if r.prediction != "uncertain"]
if not decisive:
return 0.0
confidences = np.array([r.confidence for r in decisive])
predictions = np.array([1 if r.prediction in ["ai", "hybrid"] else 0 for r in decisive])
labels = np.array([1 if r.ground_truth == "ai" else 0 for r in decisive])
ece = 0.0
for i in range(n_bins):
bin_lower = i / n_bins
bin_upper = (i + 1) / n_bins
in_bin = (confidences > bin_lower) & (confidences <= bin_upper)
if (np.sum(in_bin) > 0):
bin_accuracy = np.mean(predictions[in_bin] == labels[in_bin])
bin_confidence = np.mean(confidences[in_bin])
bin_size = np.sum(in_bin)
ece += (bin_size / len(decisive)) * abs(bin_accuracy - bin_confidence)
return ece
def analyze_by_length(self) -> Dict[str, LengthBucketMetrics]:
"""
Analyze performance across different text lengths
Returns:
--------
{ dict } : Dictionary mapping label to LengthBucketMetrics
"""
# Define length buckets (word counts)
length_buckets = [(0, 100, "Very Short (0-100)"),
(100, 200, "Short (100-200)"),
(200, 400, "Medium (200-400)"),
(400, 600, "Medium-Long (400-600)"),
(600, 1000, "Long (600-1000)"),
(1000, float("inf"), "Very Long (1000+)"),
]
bucket_metrics = dict()
for min_words, max_words, label in length_buckets:
# Filter results by length
filtered = [r for r in self.results if (min_words <= r.word_count < max_words)]
if not filtered:
continue
# Separate abstained from decisive
abstained = [r for r in filtered if r.prediction in ["hybrid", "uncertain"]]
decisive = [r for r in filtered if r.prediction not in ["hybrid", "uncertain"]]
if (len(decisive) < 5):
continue
# Calculate metrics for this bucket
y_true = [1 if (r.ground_truth == "ai") else 0 for r in decisive]
y_pred = [1 if (r.prediction == "ai") else 0 for r in decisive]
if not y_true:
continue
# Precision, Recall, F1
tp = sum(1 for i, _ in enumerate(y_true) if (y_true[i] == 1) and (y_pred[i] == 1))
fp = sum(1 for i, _ in enumerate(y_true) if (y_true[i] == 0) and (y_pred[i] == 1))
fn = sum(1 for i, _ in enumerate(y_true) if (y_true[i] == 1) and (y_pred[i] == 0))
precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0.0
valid = [r for r in decisive if r.is_correct is not None]
accuracy = sum(r.is_correct for r in valid) / len(valid) if valid else 0.0
# Additional metrics
mean_conf = np.mean([r.confidence for r in decisive])
fp_rate = fp / len(decisive) if len(decisive) > 0 else 0.0
fn_rate = fn / len(decisive) if len(decisive) > 0 else 0.0
avg_time = np.mean([r.processing_time for r in decisive])
bucket_metrics[label] = LengthBucketMetrics(min_words = min_words,
max_words = max_words,
label = label,
sample_count = len(decisive),
precision = precision,
recall = recall,
f1 = f1,
accuracy = accuracy,
mean_confidence = mean_conf,
fp_rate = fp_rate,
fn_rate = fn_rate,
avg_proc_time = avg_time,
abstention_rate = len(abstained) / (len(decisive) + len(abstained)) if (len(decisive) + len(abstained)) > 0 else 0.0,
)
return bucket_metrics
def print_length_analysis(self):
"""
Print length-based performance analysis
"""
print(f"\n{'=' * 80}")
print("PERFORMANCE BY TEXT LENGTH")
print("=" * 80)
bucket_metrics = self.analyze_by_length()
if not bucket_metrics:
print(" No length analysis available")
return
print(f"\n{'Length Range':<25s} {'Samples':>8s} {'F1':>8s} {'Precision':>10s} {'Recall':>8s} {'Accuracy':>10s} {'Abstain':>10s} {'Time(s)':>8s}")
print("─" * 80)
for label, metrics in bucket_metrics.items():
print(f"{label:<25s} {metrics.sample_count:>8d} "
f"{metrics.f1:>8.3f} {metrics.precision:>10.3f} "
f"{metrics.recall:>8.3f} {metrics.accuracy:>10.3f} "
f"{metrics.abstention_rate:>10.2%} "
f"{metrics.avg_proc_time:>8.2f}"
)
# Find best and worst performing buckets
if (len(bucket_metrics) > 1):
best = max(bucket_metrics.items(), key = lambda x: x[1].f1)
worst = min(bucket_metrics.items(), key = lambda x: x[1].f1)
print(f"\n Best Performance: {best[0]} (F1: {best[1].f1:.3f})")
print(f" Worst Performance: {worst[0]} (F1: {worst[1].f1:.3f})")
# Length-performance correlation
self._analyze_length_correlation(bucket_metrics = bucket_metrics)
def _analyze_length_correlation(self, bucket_metrics: Dict[str, LengthBucketMetrics]):
"""
Analyze correlation between text length and performance
Arguments:
----------
bucket_metrics { dict } : Dictionary of length bucket metrics
"""
if (len(bucket_metrics) < 3):
return
# Extract data for correlation
lengths = list()
f1_scores = list()
for metrics in bucket_metrics.values():
# Skip buckets with too few samples
if (metrics.sample_count < 5):
continue
# Skip degenerate buckets
if ((metrics.f1 == 0.0) and (metrics.precision == 0.0) and (metrics.recall == 0.0)):
continue
# Representative length
if np.isinf(metrics.max_words):
avg_length = metrics.min_words
else:
avg_length = (metrics.min_words + metrics.max_words) / 2
lengths.append(avg_length)
f1_scores.append(metrics.f1)
lengths = np.asarray(lengths, dtype = float)
f1_scores = np.asarray(f1_scores, dtype = float)
# Statistical guards
if (len(lengths) < 3):
print("\n Length-Performance Correlation:")
print(" Skipped (insufficient valid buckets)")
return
if (not np.all(np.isfinite(lengths)) or not np.all(np.isfinite(f1_scores))):
print("\n Length-Performance Correlation:")
print(" Skipped (NaN / Inf detected)")
return
if (np.std(f1_scores) == 0.0):
print("\n Length-Performance Correlation:")
print(" Skipped (zero variance in F1 scores)")
return
# Calculate Pearson correlation
corr, p_value = stats.pearsonr(lengths, f1_scores)
print(f"\n Length-Performance Correlation:")
print(f" Pearson r = {corr:.3f} (p-value: {p_value:.4f})")
if p_value < 0.05:
if (corr > 0.3):
print(" → Significant POSITIVE correlation — performance improves with length\n")
elif (corr < -0.3):
print(" → Significant NEGATIVE correlation — performance degrades with length\n")
else:
print(" → Weak but statistically significant correlation\n")
else:
print(" → No statistically significant correlation\n")
def generate_report(self):
"""
Generate comprehensive evaluation report
"""
print("\n" + "=" * 70)
print("EVALUATION REPORT - 4-CLASS SYSTEM")
print("=" * 70)
# Overall metrics
overall = self.calculate_metrics()
if overall:
print("\nOverall Performance (Decisive Predictions):")
print(f" Coverage: {overall.coverage:.1%} (decisive predictions)")
print(f" Accuracy: {overall.accuracy:.1%}")
print(f" Precision (AI): {overall.precision:.1%}")
print(f" Recall (AI): {overall.recall:.1%}")
print(f" F1 Score: {overall.f1:.1%}")
print(f" AUROC: {overall.auroc:.3f}")
print(f" AUPRC: {overall.auprc:.3f}")
print(f" ECE (Calibration): {overall.ece:.3f}")
print(f"\n4-Class Specific Metrics:")
print(f" Abstention Rate: {overall.abstention_rate:.1%}")
print(f" Hybrid Detection Rate: {overall.hybrid_detection_rate:.1%}")
print("\n Verdict Distribution:")
for verdict, count in overall.verdict_distribution.items():
pct = count / len(self.results) * 100
print(f" {verdict:30s}: {count:4d} ({pct:5.1f}%)")
# Per-domain performance
print("\n" + "-" * 70)
print("Per-Domain Performance:")
print("-" * 70)
print(f"{'Domain':<20s} {'F1':>8s} {'Coverage':>10s} {'Abstain':>10s} {'Hybrid%':>10s}")
print("-" * 70)
domain_scores = list()
for domain in sorted(set(r.domain for r in self.results)):
metrics = self.calculate_metrics(domain = domain)
if metrics and (metrics.support['ai'] + metrics.support['human']) >= 5:
domain_scores.append((domain, metrics.f1, metrics.coverage))
print(f"{domain:<20s} {metrics.f1:>8.1%} {metrics.coverage:>10.1%} {metrics.abstention_rate:>10.1%} {metrics.hybrid_detection_rate:>10.1%}")
# Per-subset performance
print("\n" + "-" * 70)
print("Per-Subset Performance:")
print("-" * 70)
for subset in sorted(set(r.subset for r in self.results)):
metrics = self.calculate_metrics(subset = subset)
if metrics:
print(f"\n {subset.upper()}:")
print(f" Samples: {metrics.support['human'] + metrics.support['ai']}")
print(f" F1 Score: {metrics.f1:.1%}")
print(f" Coverage: {metrics.coverage:.1%}")
print(f" Abstention: {metrics.abstention_rate:.1%}")
print(f" Hybrid Detection: {metrics.hybrid_detection_rate:.1%}")
# Length analysis
self.print_length_analysis()
print("\n" + "=" * 70)
def save_results(self):
"""
Save evaluation results
"""
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
# Calculate overall metrics
overall = self.calculate_metrics()
# Length analysis
length_metrics = self.analyze_by_length()
length_dict = {k: asdict(v) for k, v in length_metrics.items()}
# Abstention stats
abstention_stats = {"total_uncertain" : sum(1 for r in self.results if r.prediction == "uncertain"),
"total_hybrid" : sum(1 for r in self.results if r.prediction == "hybrid"),
"abstention_rate" : sum(1 for r in self.results if r.prediction == "uncertain") / len(self.results) if self.results else 0,
"hybrid_rate" : sum(1 for r in self.results if r.prediction == "hybrid") / len(self.results) if self.results else 0,
"avg_uncertainty" : np.mean([r.uncertainty for r in self.results if r.prediction == "uncertain"]) if any(r.prediction == "uncertain" for r in self.results) else 0,
}
# Save detailed results as JSON
results_dict = [asdict(r) for r in self.results]
json_path = self.output_dir / f"evaluation_results_{timestamp}.json"
with open(json_path, 'w') as f:
json.dump(obj = {'metadata' : self.metadata,
'overall_metrics' : asdict(overall) if overall else {},
'length_metrics' : length_dict,
'abstention' : abstention_stats,
'timestamp' : timestamp,
'results' : results_dict,
},
fp = f,
indent = 4,
)
print(f"\n✓ JSON results saved: {json_path}")
# Save as CSV
df = pd.DataFrame(data = results_dict)
csv_path = self.output_dir / f"evaluation_results_{timestamp}.csv"
df.to_csv(csv_path, index = False)
print(f"✓ CSV results saved: {csv_path}")
def plot_visualizations(self):
"""
Generate comprehensive evaluation visualizations
"""
fig, axes = plt.subplots(nrows = 2,
ncols = 2,
figsize = (18, 14),
)
plt.suptitle('TEXT-AUTH Evaluation Results (4-Class System)',
fontsize = 18,
fontweight = 'bold',
y = 0.98,
)
# Confusion Matrix (decisive predictions only)
ax1 = axes[0, 0]
overall = self.calculate_metrics()
if overall:
cm = np.array(overall.confusion_matrix)
sns.heatmap(cm,
annot = True,
fmt = 'd',
cmap = 'Blues',
ax = ax1,
xticklabels = ['Human', 'AI/Hybrid'],
yticklabels = ['Human', 'AI'],
)
ax1.set_title('Confusion Matrix\n(Decisive Predictions Only)')
ax1.set_xlabel('Predicted')
ax1.set_ylabel('Actual')
# F1 Score by Domain
ax2 = axes[0, 1]
domain_scores = list()
for domain in sorted(set(r.domain for r in self.results)):
metrics = self.calculate_metrics(domain = domain)
if metrics and (metrics.support['ai'] + metrics.support['human']) >= 10:
domain_scores.append((domain, metrics.f1))
domain_scores.sort(key = lambda x: x[1])
if domain_scores:
domain_labels, domain_f1 = zip(*domain_scores)
ax2.barh(domain_labels, domain_f1, color = 'steelblue')
if overall:
ax2.axvline(x = overall.f1,
color = 'red',
linestyle = '--',
linewidth = 1.5,
label = f'Overall ({overall.f1:.1%})',
)
ax2.set_xlim([0, 1])
ax2.set_xlabel('F1 Score')
ax2.set_title('F1 Score by Domain')
ax2.grid(axis = 'x', alpha = 0.3)
ax2.legend()
# Verdict Distribution (Pie Chart)
ax3 = axes[1, 0]
if overall:
verdict_counts = overall.verdict_distribution
labels = list(verdict_counts.keys())
sizes = list(verdict_counts.values())
colors = ['#fee2e2', '#d1fae5', '#e9d5ff', '#fef3c7']
explode = (0.05, 0.05, 0.05, 0.05)
ax3.pie(sizes,
labels = labels,
autopct = '%1.1f%%',
colors = colors,
explode = explode,
startangle = 90,
textprops = {'fontsize': 9})
ax3.set_title('4-Class Verdict Distribution')
# Robustness by Subset
ax4 = axes[1, 1]
subset_scores = list()
for subset in sorted(set(r.subset for r in self.results)):
metrics = self.calculate_metrics(subset = subset)
if metrics:
subset_scores.append((subset, metrics.f1, metrics.coverage, metrics.abstention_rate))
if subset_scores:
subset_labels = [s[0] for s in subset_scores]
subset_f1 = [s[1] for s in subset_scores]
subset_cov = [s[2] for s in subset_scores]
subset_abs = [s[3] for s in subset_scores]
x = np.arange(len(subset_labels))
width = 0.25
ax4.bar(x - width, subset_f1, width, label = 'F1 Score', color = 'steelblue')
ax4.bar(x, subset_cov, width, label = 'Coverage', color = 'lightcoral')
ax4.bar(x + width, subset_abs, width, label = 'Abstention', color = 'gold')
ax4.set_ylabel('Score / Rate')
ax4.set_title('Performance & Behavior by Subset')
ax4.set_xticks(x)
ax4.set_xticklabels(subset_labels, rotation = 45, ha = 'right')
ax4.legend()
ax4.grid(axis = 'y', alpha = 0.3)
ax4.set_ylim([0, 1])
plt.tight_layout(rect = [0, 0, 1, 0.96])
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
plot_path = self.output_dir / f"evaluation_plots_{timestamp}.png"
plt.savefig(plot_path, dpi = 300, bbox_inches = 'tight')
plt.close()
print(f"✓ Main plots saved: {plot_path}")
# Generate length analysis plots
self.plot_length_visualizations()
def plot_length_visualizations(self):
"""
Generate length-based performance visualizations
"""
bucket_metrics = self.analyze_by_length()
if not bucket_metrics or len(bucket_metrics) < 2:
return
fig, axes = plt.subplots(nrows = 2,
ncols = 2,
figsize = (16, 12),
)
plt.suptitle('Performance Analysis by Text Length',
fontsize = 16,
fontweight = 'bold',
y = 0.98,
)
labels = list(bucket_metrics.keys())
metrics_list = list(bucket_metrics.values())
# F1, Precision, Recall by Length
ax1 = axes[0, 0]
f1_vals = [m.f1 for m in metrics_list]
precision_vals = [m.precision for m in metrics_list]
recall_vals = [m.recall for m in metrics_list]
x = np.arange(len(labels))
width = 0.25
ax1.bar(x - width, f1_vals, width, label = 'F1', color = 'steelblue')
ax1.bar(x, precision_vals, width, label = 'Precision', color = 'lightcoral')
ax1.bar(x + width, recall_vals, width, label = 'Recall', color = 'lightgreen')
ax1.set_ylabel('Score')
ax1.set_title('Classification Metrics by Length')
ax1.set_xticks(x)
ax1.set_xticklabels(labels, rotation = 45, ha = 'right', fontsize = 9)
ax1.legend()
ax1.grid(axis = 'y', alpha = 0.3)
ax1.set_ylim([0, 1])
# Sample Distribution
ax2 = axes[0, 1]
sample_counts = [m.sample_count for m in metrics_list]
ax2.bar(labels, sample_counts, color = 'mediumpurple')
ax2.set_ylabel('Number of Samples')
ax2.set_title('Sample Distribution by Length')
ax2.set_xticklabels(labels, rotation = 45, ha = 'right', fontsize = 9)
ax2.grid(axis = 'y', alpha = 0.3)
# Processing Time by Length
ax3 = axes[1, 0]
proc_times = [m.avg_proc_time for m in metrics_list]
ax3.plot(labels, proc_times, marker = 'o', linewidth = 2, markersize = 8, color = 'darkorange')
ax3.set_ylabel('Processing Time (seconds)')
ax3.set_title('Average Processing Time by Length')
ax3.set_xticklabels(labels, rotation = 45, ha = 'right', fontsize = 9)
ax3.grid(alpha = 0.3)
# Abstention Rate by Length
ax4 = axes[1, 1]
abstention_rates = [m.abstention_rate * 100 for m in metrics_list]
ax4.bar(labels, abstention_rates, color = 'gold')
ax4.set_ylabel('Abstention Rate (%)')
ax4.set_title('Abstention Rate by Length')
ax4.set_xticklabels(labels, rotation = 45, ha = 'right', fontsize = 9)
ax4.grid(axis = 'y', alpha = 0.3)
plt.tight_layout(rect = [0, 0, 1, 0.96])
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
plot_path = self.output_dir / f"length_analysis_{timestamp}.png"
plt.savefig(plot_path, dpi = 300, bbox_inches = 'tight')
plt.close()
print(f"✓ Length analysis plots saved: {plot_path}")
def main():
parser = argparse.ArgumentParser(description = 'Run TEXT-AUTH evaluation (4-class system)')
parser.add_argument('--dataset', type = str, default = 'evaluation', help = 'Path to evaluation directory')
parser.add_argument('--output', type = str, default = 'evaluation/results', help = 'Output directory for results')
parser.add_argument('--quick-test', action = 'store_true', help = 'Run quick test on 10 samples per domain')
parser.add_argument('--samples', type = int, default = None, help = 'Maximum samples per domain')
parser.add_argument('--domains', type = str, nargs = '+', default = None, help = 'Specific domains to evaluate')
parser.add_argument('--subset', type = str, choices = ['clean', 'paraphrased', 'cross_model'], help = 'Evaluate only specific subset')
args = parser.parse_args()
# Initialize evaluator
evaluator = TextAuthEvaluator(dataset_path = args.dataset,
output_dir = args.output,
)
# Load dataset
max_samples = 10 if args.quick_test else args.samples
samples = evaluator.load_dataset(domains = args.domains,
max_samples_per_domain = max_samples,
subset_filter = args.subset,
)
if not samples:
print("No samples loaded. Check dataset path and run data collection scripts.")
return 1
# Run evaluation
evaluator.run_evaluation(samples)
# Generate report
evaluator.generate_report()
# Save results
evaluator.save_results()
# Generate plots
evaluator.plot_visualizations()
print("\n✓ Evaluation complete!\n")
return 0
# Execution
if __name__ == "__main__":
sys.exit(main()) |