File size: 20,307 Bytes
33acf50
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""
Standard benchmark evaluation for SenseCatch sentiment analysis system.

Evaluates on the IMDB test set (25,000 reviews) using proper ML evaluation
methodology: accuracy, precision, recall, F1-score, and confusion matrix.

Includes:
  - Per-model evaluation (Naive Bayes, Logistic Regression)
  - Ablation study (VADER -> NB -> LR -> Ensemble -> Full System)
  - Optional DistilBERT comparison baseline

Usage:
    python evaluate.py                 # Full evaluation (25K reviews)
    python evaluate.py --quick         # Quick evaluation (2000 samples)
    python evaluate.py --no-distilbert # Skip DistilBERT comparison
"""

import os
import sys
import json
import random
import time
import shutil
import tarfile
import logging
import argparse
import urllib.request
from datetime import datetime

import numpy as np
from scipy.sparse import hstack, csr_matrix
from sklearn.metrics import (
    accuracy_score,
    precision_recall_fscore_support,
    confusion_matrix,
)

# ---------------------------------------------------------------------------
# Setup
# ---------------------------------------------------------------------------

logging.basicConfig(level=logging.WARNING)
logger = logging.getLogger(__name__)

PROJECT_ROOT = os.path.dirname(os.path.abspath(__file__))
IMDB_URL = "https://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz"
IMDB_DIR = os.path.join(PROJECT_ROOT, "datasets", "aclImdb")
IMDB_TAR = os.path.join(PROJECT_ROOT, "datasets", "aclImdb_v1.tar.gz")
RESULTS_PATH = os.path.join(PROJECT_ROOT, "artifacts", "evaluation_results.json")
PREPROCESS_VERSION = "v1"  # bump when ensemble._preprocess_text changes -> invalidates the preprocess cache

# ---------------------------------------------------------------------------
# IMDB dataset
# ---------------------------------------------------------------------------


def download_imdb():
    """Download and extract the IMDB dataset if not already present."""
    if os.path.isdir(IMDB_DIR):
        return

    os.makedirs(os.path.join(PROJECT_ROOT, "datasets"), exist_ok=True)

    if not os.path.isfile(IMDB_TAR):
        print("Downloading IMDB dataset (84 MB)...")
        import ssl

        ctx = ssl.create_default_context()
        try:
            import certifi

            ctx.load_verify_locations(certifi.where())
        except (ImportError, Exception):
            ctx.check_hostname = False
            ctx.verify_mode = ssl.CERT_NONE
        req = urllib.request.Request(IMDB_URL)
        with urllib.request.urlopen(req, context=ctx) as resp, open(
            IMDB_TAR, "wb"
        ) as out:
            shutil.copyfileobj(resp, out)
        print("Download complete.")

    print("Extracting IMDB dataset...")
    with tarfile.open(IMDB_TAR, "r:gz") as tar:
        tar.extractall(path=os.path.join(PROJECT_ROOT, "datasets"))
    print("Extraction complete.")


def load_imdb_test(max_per_class=None, seed=42):
    """Load the IMDB test split. Returns (texts, labels) where labels are 0/1.

    When max_per_class is set (quick mode), draw a seeded, class-balanced
    random sample per class, not the first-N-sorted slice. The sorted slice
    was about 2 to 4 points optimistic; a random sample is an unbiased,
    reproducible preview of the full test set.
    """
    texts, labels = [], []
    rng = random.Random(seed)

    for sentiment, label in [("pos", 1), ("neg", 0)]:
        folder = os.path.join(IMDB_DIR, "test", sentiment)
        filenames = sorted(os.listdir(folder))
        if max_per_class and max_per_class < len(filenames):
            filenames = sorted(rng.sample(filenames, max_per_class))
        for fname in filenames:
            with open(os.path.join(folder, fname), "r", encoding="utf-8") as f:
                texts.append(f.read())
            labels.append(label)

    return texts, np.array(labels)


# ---------------------------------------------------------------------------
# Model loading
# ---------------------------------------------------------------------------


def ensure_models():
    """Make sure model .pkl files are available in models/ directory."""
    models_dir = os.path.join(PROJECT_ROOT, "models")
    artifacts_dir = os.path.join(PROJECT_ROOT, "artifacts", "models")

    needed = ["naive_bayes.pkl", "logistic_regression.pkl"]
    missing = [f for f in needed if not os.path.isfile(os.path.join(models_dir, f))]

    if not missing:
        return

    if os.path.isdir(artifacts_dir):
        os.makedirs(models_dir, exist_ok=True)
        for f in missing:
            src = os.path.join(artifacts_dir, f)
            if os.path.isfile(src):
                shutil.copy2(src, os.path.join(models_dir, f))
                print(f"  Copied {f} from artifacts/models/ to models/")

    still_missing = [
        f for f in needed if not os.path.isfile(os.path.join(models_dir, f))
    ]
    if still_missing:
        print("  Model files not found. Generating fallback models...")
        sys.path.insert(0, PROJECT_ROOT)
        from src.sensecatch.utils.model_consistency import ensure_model_consistency

        ensure_model_consistency()


def load_ensemble():
    """Load and return a SentimentEnsemble instance."""
    sys.path.insert(0, PROJECT_ROOT)
    from src.sensecatch.ensemble_model import SentimentEnsemble

    return SentimentEnsemble()


# ---------------------------------------------------------------------------
# Feature extraction (cached)
# ---------------------------------------------------------------------------


def preprocess_texts(ensemble, texts):
    """Preprocess all texts once using the ensemble's pipeline. Returns list of strings.

    Caches the preprocessed TEXT to disk (model-independent) keyed by a hash of
    (PREPROCESS_VERSION + the input texts). The NLTK pos_tag/ne_chunk pass
    dominates runtime (~0.1s/review -> ~41 min on 25K), so a warm run skips it.
    Feature matrices are deliberately not cached; they change on every retrain.
    """
    import hashlib

    cache_dir = os.path.join(PROJECT_ROOT, "artifacts", "cache")
    key = hashlib.sha256(
        (PREPROCESS_VERSION + "\x00".join(texts)).encode("utf-8")
    ).hexdigest()[:16]
    cache_path = os.path.join(cache_dir, f"preproc_{key}.json")

    if os.path.isfile(cache_path):
        with open(cache_path, "r", encoding="utf-8") as f:
            cached = json.load(f)
        if len(cached) == len(texts):
            print(f"  Loaded {len(cached)} preprocessed texts from cache.")
            return cached

    total = len(texts)
    step = max(1, total // 10)  # report every 10%
    print(f"  Preprocessing {total} texts...")
    processed = []
    for i, text in enumerate(texts):
        processed.append(ensemble._preprocess_text(text))
        if (i + 1) % step == 0 or (i + 1) == total:
            pct = (i + 1) / total * 100
            print(f"    {i + 1}/{total} ({pct:.0f}%) preprocessed", flush=True)

    os.makedirs(cache_dir, exist_ok=True)
    with open(cache_path, "w", encoding="utf-8") as f:
        json.dump(processed, f)
    return processed


def build_features(ensemble, model_name, processed_texts):
    """Build the feature matrix for a specific model from preprocessed texts."""
    vectorizer = ensemble.vectorizers.get(model_name)
    dict_vec = ensemble.dict_vectorizers.get(model_name)
    model = ensemble.models[model_name]

    text_features = vectorizer.transform(processed_texts)

    if dict_vec and ensemble.lexicon:
        lexicon_feats = [
            ensemble.lexicon.extract_all_features(t) for t in processed_texts
        ]
        dict_features = dict_vec.transform(lexicon_feats)
        features = hstack([text_features, dict_features])
    else:
        features = text_features

    # Handle dimension mismatch (pad or trim)
    expected_dim = getattr(model, "n_features_in_", None)
    if expected_dim is None and hasattr(model, "coef_"):
        expected_dim = model.coef_.shape[1]
    if expected_dim and features.shape[1] != expected_dim:
        diff = expected_dim - features.shape[1]
        if diff > 0:
            padding = csr_matrix((features.shape[0], diff))
            features = hstack([features, padding])
        else:
            features = features[:, :expected_dim]

    return features


# ---------------------------------------------------------------------------
# Metrics
# ---------------------------------------------------------------------------


def compute_metrics(name, y_true, y_pred, neutral_count=0):
    """Compute and print classification metrics. Returns dict of metrics."""
    acc = accuracy_score(y_true, y_pred)
    prec, rec, f1, _ = precision_recall_fscore_support(
        y_true, y_pred, average="binary", pos_label=1
    )
    cm = confusion_matrix(y_true, y_pred, labels=[0, 1])

    print(f"\n  {'=' * 56}")
    print(f"  {name}")
    print(f"  {'=' * 56}")
    print(f"  Accuracy:  {acc:.4f}  ({int(acc * len(y_true))}/{len(y_true)})")
    print(f"  Precision: {prec:.4f}")
    print(f"  Recall:    {rec:.4f}")
    print(f"  F1 Score:  {f1:.4f}")
    if neutral_count > 0:
        print(f"  Neutral predictions (counted as incorrect): {neutral_count}")
    print()
    print(f"  Confusion Matrix:")
    print(f"                   Predicted Neg  Predicted Pos")
    print(f"    Actual Neg     {cm[0][0]:>10}     {cm[0][1]:>10}")
    print(f"    Actual Pos     {cm[1][0]:>10}     {cm[1][1]:>10}")
    print()

    return {
        "accuracy": round(acc, 4),
        "precision": round(prec, 4),
        "recall": round(rec, 4),
        "f1": round(f1, 4),
        "confusion_matrix": cm.tolist(),
        "neutral_count": neutral_count,
        "total_samples": len(y_true),
    }


# ---------------------------------------------------------------------------
# Evaluation functions
# ---------------------------------------------------------------------------


def evaluate_vader(texts, labels):
    """VADER lexicon baseline (no ML, no rules)."""
    from nltk.sentiment.vader import SentimentIntensityAnalyzer

    vader = SentimentIntensityAnalyzer()
    total = len(texts)
    step = max(1, total // 10)
    preds = []
    for i, t in enumerate(texts):
        preds.append(1 if vader.polarity_scores(t)["compound"] >= 0.05 else 0)
        if (i + 1) % step == 0 or (i + 1) == total:
            pct = (i + 1) / total * 100
            print(f"    {i + 1}/{total} ({pct:.0f}%) scored", flush=True)
    return compute_metrics("VADER Lexicon Baseline", labels, np.array(preds))


def evaluate_raw_model(ensemble, model_name, features, labels):
    """Evaluate a single sklearn model (no rule-based layer)."""
    model = ensemble.models[model_name]
    preds = model.predict(features)
    display = "Naive Bayes" if model_name == "naive_bayes" else "Logistic Regression"
    return compute_metrics(f"{display} (raw model, no rules)", labels, preds)


def evaluate_raw_ensemble(ensemble, features_nb, features_lr, labels):
    """Evaluate weighted ensemble of NB + LR (no rule-based layer)."""
    nb_probs = ensemble.models["naive_bayes"].predict_proba(features_nb)
    lr_probs = ensemble.models["logistic_regression"].predict_proba(features_lr)

    w_nb = ensemble.model_weights.get("naive_bayes", 0.6)
    w_lr = ensemble.model_weights.get("logistic_regression", 0.4)
    combined = w_nb * nb_probs + w_lr * lr_probs

    preds = np.argmax(combined, axis=1)
    return compute_metrics("Weighted Ensemble (NB + LR, no rules)", labels, preds)


def evaluate_full_system(ensemble, texts, labels):
    """Full system including all rule-based layers."""
    preds = []
    neutral_count = 0

    total = len(texts)
    step = max(1, total // 20)  # report every 5%
    for i, text in enumerate(texts):
        if (i + 1) % step == 0 or (i + 1) == total:
            pct = (i + 1) / total * 100
            print(f"    {i + 1}/{total} ({pct:.0f}%) reviews processed", flush=True)

        result = ensemble.predict(text)
        sent = result["sentiment"]

        if sent == "Positive":
            preds.append(1)
        elif sent == "Negative":
            preds.append(0)
        else:
            # Neutral on binary data, so count it as incorrect
            preds.append(1 - labels[i])
            neutral_count += 1

    return compute_metrics(
        "Full System (ensemble + rule-based layer)",
        labels,
        np.array(preds),
        neutral_count=neutral_count,
    )


def evaluate_distilbert(texts, labels):
    """DistilBERT (SST-2 fine-tuned) as a SOTA reference."""
    try:
        from transformers import pipeline
        import torch
    except ImportError:
        print(
            "\n  [Skipped] DistilBERT requires: pip install transformers torch"
        )
        return None

    # Auto-select the fastest available device: CUDA GPU, then Apple
    # Silicon GPU (MPS), else CPU. Predictions are device-independent.
    if torch.cuda.is_available():
        device = 0
    elif getattr(torch.backends, "mps", None) is not None and torch.backends.mps.is_available():
        device = "mps"
    else:
        device = -1

    print(f"  Loading DistilBERT model... (device={device})")
    classifier = pipeline(
        "sentiment-analysis",
        model="distilbert-base-uncased-finetuned-sst-2-english",
        device=device,
        truncation=True,
        max_length=512,
    )

    preds = []
    batch_size = 64

    total = len(texts)
    for i in range(0, total, batch_size):
        batch = [t[:2000] for t in texts[i : i + batch_size]]
        results = classifier(batch)
        for r in results:
            preds.append(1 if r["label"] == "POSITIVE" else 0)
        done = min(i + batch_size, total)
        pct = done / total * 100
        if done % max(1, (total // 10)) < batch_size or done == total:
            print(f"    {done}/{total} ({pct:.0f}%) classified", flush=True)

    return compute_metrics(
        "DistilBERT (SST-2 fine-tuned, SOTA reference)", labels, np.array(preds)
    )


# ---------------------------------------------------------------------------
# Ablation table
# ---------------------------------------------------------------------------


def print_ablation_table(rows, distilbert_metrics=None):
    """Print formatted ablation study summary."""
    print(f"\n{'=' * 62}")
    print(f"  ABLATION STUDY: IMDB Test Set")
    print(f"{'=' * 62}")
    print(f"  {'Component':<42} {'Acc':>7}  {'F1':>7}")
    print(f"  {'-' * 42} {'-' * 7}  {'-' * 7}")

    for name, m in rows:
        if m:
            print(f"  {name:<42} {m['accuracy']:>7.4f}  {m['f1']:>7.4f}")
        else:
            print(f"  {name:<42} {'N/A':>7}  {'N/A':>7}")

    if distilbert_metrics:
        print(f"  {'-' * 42} {'-' * 7}  {'-' * 7}")
        print(
            f"  {'DistilBERT (SST-2, SOTA reference)':<42} "
            f"{distilbert_metrics['accuracy']:>7.4f}  "
            f"{distilbert_metrics['f1']:>7.4f}"
        )

    print()


# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------


def main():
    # Force line-buffered stdout so progress is visible even when piped
    try:
        sys.stdout.reconfigure(line_buffering=True)
    except AttributeError:
        pass

    parser = argparse.ArgumentParser(
        description="Evaluate SenseCatch on IMDB test set"
    )
    parser.add_argument(
        "--quick",
        action="store_true",
        help="Quick evaluation on 1000 samples per class (2000 total)",
    )
    parser.add_argument(
        "--no-distilbert",
        action="store_true",
        help="Skip DistilBERT comparison",
    )
    args = parser.parse_args()

    max_per_class = 1000 if args.quick else None
    total_samples = (max_per_class * 2) if max_per_class else 25000

    print(f"\n SenseCatch Evaluation: IMDB Test Set ({total_samples:,} reviews)")
    print("=" * 62)

    # --- Load data --------------------------------------------------------
    download_imdb()
    print(f"\nLoading IMDB test set...")
    texts, labels = load_imdb_test(max_per_class)
    print(
        f"  {len(texts)} reviews loaded "
        f"({sum(labels == 1)} positive, {sum(labels == 0)} negative)"
    )

    # --- Load models ------------------------------------------------------
    ensure_models()
    print("\nLoading SentimentEnsemble...")
    ensemble = load_ensemble()
    print(f"  Models: {', '.join(ensemble.models.keys())}")

    if not ensemble.models:
        print("ERROR: No models loaded.")
        sys.exit(1)

    # --- Preprocess once (shared across raw model evaluations) ------------
    print("\n[Step 1] Preprocessing texts (shared across all raw model evals)...")
    t0 = time.time()
    processed = preprocess_texts(ensemble, texts)
    print(f"  Done in {time.time() - t0:.1f}s")

    # --- Build feature matrices once per model ----------------------------
    print("\n[Step 2] Building feature matrices...")
    t0 = time.time()
    features = {}
    for model_name in ensemble.models:
        if ensemble.vectorizers.get(model_name):
            features[model_name] = build_features(ensemble, model_name, processed)
            print(f"  {model_name}: {features[model_name].shape}")
    print(f"  Done in {time.time() - t0:.1f}s")

    # --- Evaluations ------------------------------------------------------
    all_results = {}
    ablation = []

    # VADER baseline
    print("\n[3/7] VADER baseline...")
    t0 = time.time()
    m = evaluate_vader(texts, labels)
    all_results["vader"] = m
    ablation.append(("VADER (lexicon-only baseline)", m))
    print(f"  Time: {time.time() - t0:.1f}s")

    # Naive Bayes (raw)
    if "naive_bayes" in features:
        print("\n[4/7] Naive Bayes (raw model)...")
        t0 = time.time()
        m = evaluate_raw_model(ensemble, "naive_bayes", features["naive_bayes"], labels)
        all_results["naive_bayes_raw"] = m
        ablation.append(("+ Naive Bayes (CountVectorizer)", m))
        print(f"  Time: {time.time() - t0:.1f}s")

    # Logistic Regression (raw)
    if "logistic_regression" in features:
        print("\n[5/7] Logistic Regression (raw model)...")
        t0 = time.time()
        m = evaluate_raw_model(
            ensemble, "logistic_regression", features["logistic_regression"], labels
        )
        all_results["logistic_regression_raw"] = m
        ablation.append(("+ Logistic Regression (TF-IDF)", m))
        print(f"  Time: {time.time() - t0:.1f}s")

    # Weighted ensemble (raw)
    if "naive_bayes" in features and "logistic_regression" in features:
        print("\n[6/7] Weighted ensemble (NB + LR)...")
        t0 = time.time()
        m = evaluate_raw_ensemble(
            ensemble, features["naive_bayes"], features["logistic_regression"], labels
        )
        all_results["ensemble_raw"] = m
        ablation.append(("+ Weighted Ensemble (0.6 NB + 0.4 LR)", m))
        print(f"  Time: {time.time() - t0:.1f}s")

    # Full system
    print("\n[7/7] Full system (ensemble + rules)...")
    t0 = time.time()
    m = evaluate_full_system(ensemble, texts, labels)
    all_results["full_system"] = m
    ablation.append(("+ Rule-based layer (full system)", m))
    print(f"  Time: {time.time() - t0:.1f}s")

    # DistilBERT (optional)
    distilbert_metrics = None
    if not args.no_distilbert:
        print("\n[Bonus] DistilBERT (SOTA reference)...")
        t0 = time.time()
        distilbert_metrics = evaluate_distilbert(texts, labels)
        all_results["distilbert"] = distilbert_metrics
        if distilbert_metrics:
            print(f"  Time: {time.time() - t0:.1f}s")

    # --- Summary ----------------------------------------------------------
    print_ablation_table(ablation, distilbert_metrics)

    # --- Save results -----------------------------------------------------
    os.makedirs(os.path.dirname(RESULTS_PATH), exist_ok=True)
    output = {
        "timestamp": datetime.now().isoformat(),
        "dataset": "IMDB",
        "split": "test",
        "total_samples": len(texts),
        "quick_mode": args.quick,
        "results": all_results,
    }
    with open(RESULTS_PATH, "w") as f:
        json.dump(output, f, indent=2)
    print(f"Results saved to {RESULTS_PATH}")


if __name__ == "__main__":
    main()