File size: 22,872 Bytes
5fc6e5d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import ast
import hashlib
from pathlib import Path
import random
import re
from typing import List, Tuple

import nltk
from nltk.corpus import stopwords, wordnet
from nltk.stem import PorterStemmer, WordNetLemmatizer
import numpy as np
import pandas as pd
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.feature_selection import SelectKBest, chi2
import typer

from turing.config import (
    INTERIM_DATA_DIR,
    LABEL_COLUMN,
    LANGS,
)
from turing.data_validation import run_custom_deepchecks, run_targeted_nlp_checks
from turing.dataset import DatasetManager

# --- NLTK Resource Check ---
REQUIRED_NLTK_PACKAGES = [
    "stopwords",
    "wordnet",
    "omw-1.4",
    "averaged_perceptron_tagger",
    "punkt",
]
for package in REQUIRED_NLTK_PACKAGES:
    try:
        nltk.data.find(f"corpora/{package}")
    except LookupError:
        try:
            nltk.download(package, quiet=True)
        except Exception:
            pass

app = typer.Typer()


# --- CONFIGURATION CLASS ---
class FeaturePipelineConfig:
    """
    Configuration holder for the pipeline. Generates a unique ID based on parameters
    to version the output directories.
    """

    def __init__(
        self,
        use_stopwords: bool,
        use_lemmatization: bool,
        use_combo_feature: bool,
        max_features: int,
        min_comment_length: int,
        max_comment_length: int,
        enable_augmentation: bool,
        custom_tags: str = "base",
    ):
        self.use_stopwords = use_stopwords
        self.use_lemmatization = use_lemmatization
        self.use_combo_feature = use_combo_feature
        self.max_features = max_features
        self.min_comment_length = min_comment_length
        self.max_comment_length = max_comment_length
        self.enable_augmentation = enable_augmentation
        self.custom_tags = custom_tags
        self.hash_id = self._generate_readable_id()

    def _generate_readable_id(self) -> str:
        tags = ["clean"]
        if self.enable_augmentation:
            tags.append("aug-soft")
        tags.append(f"k{self.max_features}")
        if self.custom_tags != "base":
            tags.append(self.custom_tags)
        return "-".join(tags)


# --- TEXT UTILITIES ---
class TextCanonicalizer:
    """
    Reduces text to a 'canonical' form (stemmed, lowercase)
    to detect semantic duplicates.
    preserves javadoc tags to distinguish usage (@return) from summary (Returns).
    """

    def __init__(self):
        self.stemmer = PorterStemmer()
        self.stop_words = set(stopwords.words("english"))
        # Code keywords are preserved as they carry semantic weight
        self.code_keywords = {
            "return",
            "true",
            "false",
            "null",
            "if",
            "else",
            "void",
            "int",
            "boolean",
            "param", 
            "throws",
            "exception",
        }

    def to_canonical(self, text: str) -> str:
        if pd.isna(text):
            return ""
        text = str(text).lower()
        text = re.sub(r"[^a-z0-9\s@]", " ", text)
        
        words = text.split()
        canonical_words = []
        
        for w in words:
            # If the word starts with @ (e.g., @return), keep it as is
            if w.startswith("@"):
                canonical_words.append(w)
                continue

            if w in self.stop_words and w not in self.code_keywords:
                continue
            
            stemmed = self.stemmer.stem(w)
            canonical_words.append(stemmed)
            
        return " ".join(canonical_words).strip()


class TextProcessor:
    """
    Standard text cleaning logic for final feature extraction (TF-IDF).
    """

    def __init__(self, config: FeaturePipelineConfig, language: str = "english"):
        self.config = config
        self.stop_words = set(stopwords.words(language))
        self.lemmatizer = WordNetLemmatizer()

    def clean_text(self, text: str) -> str:
        if pd.isna(text):
            return ""
        text = str(text).lower()
        # Remove heavy code markers but keep text structure
        text = re.sub(r"(^\s*//+|^\s*/\*+|\*/$)", "", text)
        # Keep only alpha characters for NLP model (plus pipe for combo)
        text = re.sub(r"[^a-z\s|]", " ", text)
        tokens = text.split()
        if self.config.use_stopwords:
            tokens = [w for w in tokens if w not in self.stop_words]
        if self.config.use_lemmatization:
            tokens = [self.lemmatizer.lemmatize(w) for w in tokens]
        return " ".join(tokens)


# --- AUGMENTATION ---
class SafeAugmenter:
    """
    protects reserved keywords from synonym replacement.
    """

    def __init__(self, aug_prob=0.3):
        self.aug_prob = aug_prob
        self.protected_words = {
            "return",
            "public",
            "private",
            "void",
            "class",
            "static",
            "final",
            "if",
            "else",
            "for",
            "while",
            "try",
            "catch",
            "import",
            "package",
            "null",
            "true",
            "false",
            "self",
            "def",
            "todo",
            "fixme",
            "param",
            "throw",
        }

    def get_synonyms(self, word):
        synonyms = set()
        for syn in wordnet.synsets(word):
            for lemma in syn.lemmas():
                name = lemma.name().replace("_", " ")
                if name.isalpha() and name.lower() != word.lower():
                    synonyms.add(name)
        return list(synonyms)

    def augment(self, text: str) -> str:
        if pd.isna(text) or not text:
            return ""
        words = text.split()
        if len(words) < 2:
            return text
        new_words = []
        for word in words:
            word_lower = word.lower()

            if word_lower in self.protected_words:
                new_words.append(word)
                continue

            # Random Case Injection (Noise)
            if random.random() < 0.1:
                if word[0].isupper():
                    new_words.append(word.lower())
                else:
                    new_words.append(word.capitalize())
                continue

            # Synonym Replacement
            if random.random() < self.aug_prob and len(word) > 3:
                syns = self.get_synonyms(word_lower)
                if syns:
                    replacement = random.choice(syns)
                    if word[0].isupper():
                        replacement = replacement.capitalize()
                    new_words.append(replacement)
                else:
                    new_words.append(word)
            else:
                new_words.append(word)
        return " ".join(new_words)

    def apply_balancing(
        self, df: pd.DataFrame, min_samples: int = 100
    ) -> Tuple[pd.DataFrame, pd.DataFrame]:
        """
        Generates synthetic data for minority classes.
        Returns: (Balanced DataFrame, Report DataFrame)
        """
        df["temp_label_str"] = df[LABEL_COLUMN].astype(str)
        counts = df["temp_label_str"].value_counts()
        print(
            f"\n   [Balance Check - PRE] Min class size: {counts.min()} | Max: {counts.max()}"
        )

        existing_sentences = set(df["comment_sentence"].str.strip())
        new_rows = []
        report_rows = []

        for label_str, count in counts.items():
            if count < min_samples:
                needed = min_samples - count
                class_subset = df[df["temp_label_str"] == label_str]
                if class_subset.empty:
                    continue

                samples = class_subset["comment_sentence"].tolist()
                orig_label = class_subset[LABEL_COLUMN].iloc[0]

                # Propagate 'combo' if present
                orig_combo = None
                if "combo" in class_subset.columns:
                    orig_combo = class_subset["combo"].iloc[0]

                generated = 0
                attempts = 0
                # Cap attempts to avoid infinite loops if vocabulary is too small
                while generated < needed and attempts < needed * 5:
                    attempts += 1
                    src = random.choice(samples)
                    aug_txt = self.augment(src).strip()

                    # Ensure Global Uniqueness
                    if aug_txt and aug_txt not in existing_sentences:
                        row = {
                            "comment_sentence": aug_txt,
                            LABEL_COLUMN: orig_label,
                            "partition": "train_aug",
                            "index": -1,  # Placeholder
                        }
                        if orig_combo:
                            row["combo"] = orig_combo

                        new_rows.append(row)
                        report_rows.append(
                            {
                                "original_text": src,
                                "augmented_text": aug_txt,
                                "label": label_str,
                                "reason": f"Class has {count} samples (Target {min_samples})",
                            }
                        )
                        existing_sentences.add(aug_txt)
                        generated += 1

        df = df.drop(columns=["temp_label_str"])
        df_report = pd.DataFrame(report_rows)

        if new_rows:
            augmented_df = pd.concat([df, pd.DataFrame(new_rows)], ignore_index=True)
            augmented_df["index"] = range(len(augmented_df))

            temp_counts = augmented_df[LABEL_COLUMN].astype(str).value_counts()
            print(
                f"   [Balance Check - POST] Min class size: {temp_counts.min()} | Max: {temp_counts.max()}"
            )
            return augmented_df, df_report

        return df, df_report


# --- CLEANING LOGIC ---
def clean_training_data_smart(
    df: pd.DataFrame, min_len: int, max_len: int, language: str = "english"
) -> Tuple[pd.DataFrame, pd.DataFrame]:
    """
    Performs 'Smart Cleaning' on the Training Set with language-specific heuristics.
    """
    canon = TextCanonicalizer()
    dropped_rows = []

    print(f"   [Clean] Computing heuristics (Language: {language})...")
    df["canon_key"] = df["comment_sentence"].apply(canon.to_canonical)

    # 1. Token Length Filter
    def count_code_tokens(text):
        return len([t for t in re.split(r"[^a-zA-Z0-9]+", str(text)) if t])

    df["temp_token_len"] = df["comment_sentence"].apply(count_code_tokens)

   
    MIN_ALPHA_CHARS = 6
    MAX_SYMBOL_RATIO = 0.50

    # 2. Heuristic Filters (Tiny/Huge/Code)
    def get_heuristics(text):
        s = str(text).strip()
        char_len = len(s)
        if char_len == 0:
            return False, False, 1.0
        
        alpha_len = sum(1 for c in s if c.isalpha())
        
        non_alnum_chars = sum(1 for c in s if not c.isalnum() and not c.isspace())
        symbol_ratio = non_alnum_chars / char_len if char_len > 0 else 0

        is_tiny = alpha_len < MIN_ALPHA_CHARS
        is_huge = char_len > 800
        is_code = symbol_ratio > MAX_SYMBOL_RATIO
        
        return is_tiny, is_huge, is_code

    heuristics = df["comment_sentence"].apply(get_heuristics)
    df["is_tiny"] = [x[0] for x in heuristics]
    df["is_huge"] = [x[1] for x in heuristics]
    df["symbol_ratio"] = [x[2] for x in heuristics] 
    
    
    df["is_code"] = df["symbol_ratio"] > 0.50

    mask_keep = (
        (df["temp_token_len"] >= min_len)
        & (df["temp_token_len"] <= max_len)
        & (~df["is_tiny"])
        & (~df["is_huge"])
        & (~df["is_code"])
    )

    df_dropped_qual = df[~mask_keep].copy()
    if not df_dropped_qual.empty:
        def reason(row):
            if row["is_tiny"]:
                return f"Too Tiny (<{MIN_ALPHA_CHARS} alpha)"
            if row["is_huge"]:
                return "Too Huge (>800 chars)"
            if row["is_code"]:
                return f"Pure Code (>{int(MAX_SYMBOL_RATIO*100)}% symbols)"
            return f"Token Count ({row['temp_token_len']})"

        df_dropped_qual["drop_reason"] = df_dropped_qual.apply(reason, axis=1)
        dropped_rows.append(df_dropped_qual)

    df = df[mask_keep].copy()

    # 3. Semantic Conflicts (Ambiguity)
    df["label_s"] = df[LABEL_COLUMN].astype(str)
    conflict_counts = df.groupby("canon_key")["label_s"].nunique()
    conflicting_keys = conflict_counts[conflict_counts > 1].index

    mask_conflicts = df["canon_key"].isin(conflicting_keys)
    df_dropped_conflicts = df[mask_conflicts].copy()
    if not df_dropped_conflicts.empty:
        df_dropped_conflicts["drop_reason"] = "Semantic Conflict"
        dropped_rows.append(df_dropped_conflicts)

    df = df[~mask_conflicts].copy()

    # 4. Exact Duplicates
    mask_dupes = df.duplicated(subset=["comment_sentence"], keep="first")
    df_dropped_dupes = df[mask_dupes].copy()
    if not df_dropped_dupes.empty:
        df_dropped_dupes["drop_reason"] = "Exact Duplicate"
        dropped_rows.append(df_dropped_dupes)

    df = df[~mask_dupes].copy()

    # Cleanup columns
    cols_to_drop = [
        "canon_key",
        "label_s",
        "temp_token_len",
        "is_tiny",
        "is_huge",
        "is_code",
        "symbol_ratio"
    ]
    df = df.drop(columns=cols_to_drop, errors="ignore")

    if dropped_rows:
        df_report = pd.concat(dropped_rows, ignore_index=True)
        cols_rep = ["index", "comment_sentence", LABEL_COLUMN, "drop_reason"]
        final_cols = [c for c in cols_rep if c in df_report.columns]
        df_report = df_report[final_cols]
    else:
        df_report = pd.DataFrame(columns=["index", "comment_sentence", "drop_reason"])

    print(f"   [Clean] Removed {len(df_report)} rows. Final: {len(df)}.")
    return df, df_report

# --- FEATURE ENGINEERING ---
class FeatureEngineer:
    def __init__(self, config: FeaturePipelineConfig):
        self.config = config
        self.processor = TextProcessor(config=config)
        self.tfidf_vectorizer = TfidfVectorizer(max_features=config.max_features)

    def extract_features_for_check(self, df: pd.DataFrame) -> pd.DataFrame:
        """Extracts metadata features for analysis."""

        def analyze(text):
            s = str(text)
            words = s.split()
            n_words = len(words)
            if n_words == 0:
                return 0, 0, 0
            first_word = words[0].lower()
            starts_verb = (
                1
                if first_word.endswith("s")
                or first_word.startswith("get")
                or first_word.startswith("set")
                else 0
            )
            return (len(s), n_words, starts_verb)

        metrics = df["comment_sentence"].apply(analyze)
        df["f_length"] = [x[0] for x in metrics]
        df["f_word_count"] = [x[1] for x in metrics]
        df["f_starts_verb"] = [x[2] for x in metrics]
        # Calculate MD5 hash for efficient exact duplicate detection in Deepchecks
        df["text_hash"] = df["comment_sentence"].apply(
            lambda x: hashlib.md5(str(x).encode()).hexdigest()
        )
        return df

    def vectorize_and_select(self, df_train, df_test):
        def clean_fn(x):
            return re.sub(r"[^a-zA-Z\s]", "", str(x).lower())

        X_train = self.tfidf_vectorizer.fit_transform(
            df_train["comment_sentence"].apply(clean_fn)
        )
        y_train = np.stack(df_train[LABEL_COLUMN].values)

        # Handling multi-label for Chi2 (using sum or max)
        y_train_sum = (
            y_train.sum(axis=1) if len(y_train.shape) > 1 else y_train
        )
        selector = SelectKBest(
            chi2, k=min(self.config.max_features, X_train.shape[1])
        )
        X_train = selector.fit_transform(X_train, y_train_sum)

        X_test = self.tfidf_vectorizer.transform(
            df_test["comment_sentence"].apply(clean_fn)
        )
        X_test = selector.transform(X_test)

        vocab = [
            self.tfidf_vectorizer.get_feature_names_out()[i]
            for i in selector.get_support(indices=True)
        ]
        return X_train, X_test, vocab


# --- MAIN EXECUTION ---
def main(
    feature_dir: Path = typer.Option(
        INTERIM_DATA_DIR / "features", help="Output dir."
    ),
    reports_root: Path = typer.Option(
        Path("reports/data"), help="Reports root."
    ),
    max_features: int = typer.Option(5000),
    min_comment_length: int = typer.Option(
        2, help="Remove comments shorter than chars."
    ),
    max_comment_length: int = typer.Option(300),
    augment: bool = typer.Option(False, "--augment", help="Enable augmentation."),
    balance_threshold: int = typer.Option(100, help="Min samples per class."),
    run_vectorization: bool = typer.Option(False, "--run-vectorization"),
    run_nlp_check: bool = typer.Option(
        True, "--run-nlp", help="Run Deepchecks NLP suite."
    ),
    custom_tags: str = typer.Option("base", help="Custom tags."),
    save_full_csv: bool = typer.Option(False, "--save-full-csv"),
    languages: List[str] = typer.Option(LANGS, show_default=False),
):

    config = FeaturePipelineConfig(
        True,
        True,
        True,
        max_features,
        min_comment_length,
        max_comment_length,
        augment,
        custom_tags,
    )
    print(f"=== Pipeline ID: {config.hash_id} ===")

    dm = DatasetManager()
    full_dataset = dm.get_dataset()
    fe = FeatureEngineer(config)
    augmenter = SafeAugmenter()

    feat_output_dir = feature_dir / config.hash_id
    feat_output_dir.mkdir(parents=True, exist_ok=True)
    report_output_dir = reports_root / config.hash_id

    for lang in languages:
        print(f"\n{'='*30}\nPROCESSING LANGUAGE: {lang.upper()}\n{'='*30}")
        df_train = full_dataset[f"{lang}_train"].to_pandas()
        df_test = full_dataset[f"{lang}_test"].to_pandas()

        # Standardize Label Format
        for df in [df_train, df_test]:
            if isinstance(df[LABEL_COLUMN].iloc[0], str):
                df[LABEL_COLUMN] = (
                    df[LABEL_COLUMN]
                    .str.replace(r"\s+", ", ", regex=True)
                    .apply(ast.literal_eval)
                )

        lang_report_dir = report_output_dir / lang

        # 1. RAW AUDIT
        print("   >>> Phase 1: Auditing RAW Data")
        df_train_raw = fe.extract_features_for_check(df_train.copy())
        df_test_raw = fe.extract_features_for_check(df_test.copy())
        run_custom_deepchecks(
            df_train_raw, df_test_raw, lang_report_dir, "raw", lang
        )
        if run_nlp_check:
            run_targeted_nlp_checks(
                df_train_raw, df_test_raw, lang_report_dir, "raw"
            )

        # 2. CLEANING & AUGMENTATION
        print("\n   >>> Phase 2: Smart Cleaning & Augmentation")
        df_train, df_dropped = clean_training_data_smart(
            df_train, min_comment_length, max_comment_length, language=lang
        )

        if not df_dropped.empty:
            dropped_path = lang_report_dir / "dropped_rows.csv"
            df_dropped.to_csv(dropped_path, index=False)
            print(f"   [Report] Dropped rows details saved to: {dropped_path}")

        if augment:
            print("   [Augment] Applying Soft Balancing...")
            df_train, df_aug_report = augmenter.apply_balancing(
                df_train, min_samples=balance_threshold
            )

            if not df_aug_report.empty:
                aug_path = lang_report_dir / "augmentation_report.csv"
                df_aug_report.to_csv(aug_path, index=False)
                print(
                    f"   [Report] Augmentation details saved to: {aug_path}"
                )

        # 3. PROCESSED AUDIT
        print("\n   >>> Phase 3: Auditing PROCESSED Data")
        df_train = fe.extract_features_for_check(df_train)
        df_test = fe.extract_features_for_check(df_test)
        run_custom_deepchecks(
            df_train, df_test, lang_report_dir, "processed", lang
        )
        if run_nlp_check:
            run_targeted_nlp_checks(
                df_train, df_test, lang_report_dir, "processed"
            )

        # 4. FINAL PROCESSING & SAVING
        print("\n   >>> Phase 4: Final Processing & Save")
        df_train["comment_clean"] = df_train["comment_sentence"].apply(
            fe.processor.clean_text
        )
        df_test["comment_clean"] = df_test["comment_sentence"].apply(
            fe.processor.clean_text
        )

        if config.use_combo_feature:
            if "combo" in df_train.columns:
                df_train["combo_clean"] = df_train["combo"].apply(
                    fe.processor.clean_text
                )
            if "combo" in df_test.columns:
                df_test["combo_clean"] = df_test["combo"].apply(
                    fe.processor.clean_text
                )

        X_train, X_test, vocab = None, None, []
        if run_vectorization:
            print("   [Vectorization] TF-IDF & Chi2...")
            X_train, X_test, vocab = fe.vectorize_and_select(df_train, df_test)
        def format_label_robust(lbl):
            if hasattr(lbl, "tolist"): # Check if numpy array
                lbl = lbl.tolist()
            return str(lbl)

        df_train[LABEL_COLUMN] = df_train[LABEL_COLUMN].apply(format_label_robust)
        df_test[LABEL_COLUMN] = df_test[LABEL_COLUMN].apply(format_label_robust)

        cols_to_save = [
            "index",
            LABEL_COLUMN,
            "comment_sentence",
            "comment_clean",
        ]
        if "combo" in df_train.columns:
            cols_to_save.append("combo")
        if "combo_clean" in df_train.columns:
            cols_to_save.append("combo_clean")
        meta_cols = [c for c in df_train.columns if c.startswith("f_")]
        cols_to_save.extend(meta_cols)

        print(f"   [Save] Columns: {cols_to_save}")
        df_train[cols_to_save].to_csv(
            feat_output_dir / f"{lang}_train.csv", index=False
        )
        df_test[cols_to_save].to_csv(
            feat_output_dir / f"{lang}_test.csv", index=False
        )

        if run_vectorization and X_train is not None:
            from scipy.sparse import save_npz

            save_npz(feat_output_dir / f"{lang}_train_tfidf.npz", X_train)
            save_npz(feat_output_dir / f"{lang}_test_tfidf.npz", X_test)
            with open(
                feat_output_dir / f"{lang}_vocab.txt", "w", encoding="utf-8"
            ) as f:
                f.write("\n".join(vocab))

    print(f"\nAll Done. Reports in: {report_output_dir}")


if __name__ == "__main__":
    typer.run(main)