File size: 20,849 Bytes
c7d4394
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
 
16ec2cf
c7d4394
 
16ec2cf
c7d4394
 
16ec2cf
 
c7d4394
 
16ec2cf
c7d4394
 
 
 
 
16ec2cf
c7d4394
 
 
16ec2cf
c7d4394
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
16ec2cf
c7d4394
 
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
16ec2cf
c7d4394
 
16ec2cf
c7d4394
 
 
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
16ec2cf
c7d4394
 
16ec2cf
c7d4394
 
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
 
16ec2cf
c7d4394
16ec2cf
c7d4394
 
 
 
 
 
 
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
 
 
16ec2cf
c7d4394
16ec2cf
c7d4394
 
 
 
 
 
 
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
16ec2cf
c7d4394
 
16ec2cf
c7d4394
 
 
 
 
16ec2cf
c7d4394
16ec2cf
c7d4394
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
 
 
16ec2cf
c7d4394
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
 
16ec2cf
c7d4394
 
 
 
 
 
 
 
 
16ec2cf
c7d4394
16ec2cf
c7d4394
 
 
 
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
16ec2cf
4134ab0
c7d4394
 
 
 
 
 
 
 
 
 
 
16ec2cf
c7d4394
 
 
16ec2cf
c7d4394
 
 
16ec2cf
c7d4394
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
16ec2cf
c7d4394
 
 
 
16ec2cf
c7d4394
 
16ec2cf
c7d4394
 
16ec2cf
c7d4394
 
 
16ec2cf
c7d4394
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
16ec2cf
c7d4394
16ec2cf
c7d4394
 
16ec2cf
c7d4394
 
 
 
 
 
16ec2cf
c7d4394
 
16ec2cf
c7d4394
 
 
 
 
 
 
16ec2cf
c7d4394
 
 
 
 
 
 
 
 
 
 
 
 
 
16ec2cf
eb6b502
c7d4394
16ec2cf
eb6b502
4134ab0
eb6b502
4134ab0
 
 
eb6b502
 
 
4134ab0
eb6b502
4134ab0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eb6b502
4134ab0
 
 
16ec2cf
c7d4394
4134ab0
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
"""
models/anomaly-detection/src/components/model_trainer.py
Model training with Optuna hyperparameter tuning for clustering/anomaly detection
"""
import os
import logging
import joblib
from datetime import datetime
from pathlib import Path
from typing import Optional, Dict, Any, List
import numpy as np

from ..entity import ModelTrainerConfig, ModelTrainerArtifact
from ..utils import calculate_clustering_metrics, calculate_optuna_objective, format_metrics_report

logger = logging.getLogger("model_trainer")

# MLflow
try:
    import mlflow
    import mlflow.sklearn
    MLFLOW_AVAILABLE = True
except ImportError:
    MLFLOW_AVAILABLE = False
    logger.warning("MLflow not available. Install with: pip install mlflow")

# Optuna
try:
    import optuna
    from optuna.samplers import TPESampler
    OPTUNA_AVAILABLE = True
except ImportError:
    OPTUNA_AVAILABLE = False
    logger.warning("Optuna not available. Install with: pip install optuna")

# Clustering algorithms
try:
    from sklearn.cluster import DBSCAN, KMeans
    from sklearn.ensemble import IsolationForest
    from sklearn.neighbors import LocalOutlierFactor
    SKLEARN_AVAILABLE = True
except ImportError:
    SKLEARN_AVAILABLE = False

try:
    import hdbscan
    HDBSCAN_AVAILABLE = True
except ImportError:
    HDBSCAN_AVAILABLE = False
    logger.warning("HDBSCAN not available. Install with: pip install hdbscan")


class ModelTrainer:
    """
    Model training component with:
    1. Optuna hyperparameter optimization
    2. Multiple clustering algorithms (DBSCAN, KMeans, HDBSCAN)
    3. Anomaly detection (Isolation Forest, LOF)
    4. MLflow experiment tracking
    """

    def __init__(self, config: Optional[ModelTrainerConfig] = None):
        """
        Initialize model trainer.
        
        Args:
            config: Optional configuration
        """
        self.config = config or ModelTrainerConfig()

        # Ensure output directory exists
        Path(self.config.output_directory).mkdir(parents=True, exist_ok=True)

        # Setup MLflow
        self._setup_mlflow()

        logger.info("[ModelTrainer] Initialized")
        logger.info(f"  Models to train: {self.config.models_to_train}")
        logger.info(f"  Optuna trials: {self.config.n_optuna_trials}")

    def _setup_mlflow(self):
        """Configure MLflow tracking"""
        if not MLFLOW_AVAILABLE:
            logger.warning("[ModelTrainer] MLflow not available")
            return

        try:
            # Set tracking URI
            mlflow.set_tracking_uri(self.config.mlflow_tracking_uri)

            # Set credentials for DagsHub
            if self.config.mlflow_username and self.config.mlflow_password:
                os.environ["MLFLOW_TRACKING_USERNAME"] = self.config.mlflow_username
                os.environ["MLFLOW_TRACKING_PASSWORD"] = self.config.mlflow_password

            # Create or get experiment
            try:
                mlflow.create_experiment(self.config.experiment_name)
            except Exception:
                pass
            mlflow.set_experiment(self.config.experiment_name)

            logger.info(f"[ModelTrainer] MLflow configured: {self.config.mlflow_tracking_uri}")

        except Exception as e:
            logger.warning(f"[ModelTrainer] MLflow setup error: {e}")

    def _train_dbscan(self, X: np.ndarray, trial: Optional['optuna.Trial'] = None) -> Dict[str, Any]:
        """
        Train DBSCAN with optional Optuna tuning.
        """
        if not SKLEARN_AVAILABLE:
            return {"error": "sklearn not available"}

        # Hyperparameters
        if trial:
            eps = trial.suggest_float("eps", 0.1, 2.0)
            min_samples = trial.suggest_int("min_samples", 2, 20)
        else:
            eps = 0.5
            min_samples = 5

        model = DBSCAN(eps=eps, min_samples=min_samples, n_jobs=-1)
        labels = model.fit_predict(X)

        metrics = calculate_clustering_metrics(X, labels)
        metrics["eps"] = eps
        metrics["min_samples"] = min_samples

        return {
            "model": model,
            "labels": labels,
            "metrics": metrics,
            "params": {"eps": eps, "min_samples": min_samples}
        }

    def _train_kmeans(self, X: np.ndarray, trial: Optional['optuna.Trial'] = None) -> Dict[str, Any]:
        """
        Train KMeans with optional Optuna tuning.
        """
        if not SKLEARN_AVAILABLE:
            return {"error": "sklearn not available"}

        # Hyperparameters
        if trial:
            n_clusters = trial.suggest_int("n_clusters", 2, 20)
            n_init = trial.suggest_int("n_init", 5, 20)
        else:
            n_clusters = 5
            n_init = 10

        model = KMeans(n_clusters=n_clusters, n_init=n_init, random_state=42)
        labels = model.fit_predict(X)

        metrics = calculate_clustering_metrics(X, labels)
        metrics["n_clusters"] = n_clusters

        return {
            "model": model,
            "labels": labels,
            "metrics": metrics,
            "params": {"n_clusters": n_clusters, "n_init": n_init}
        }

    def _train_hdbscan(self, X: np.ndarray, trial: Optional['optuna.Trial'] = None) -> Dict[str, Any]:
        """
        Train HDBSCAN with optional Optuna tuning.
        """
        if not HDBSCAN_AVAILABLE:
            return {"error": "hdbscan not available"}

        # Hyperparameters
        if trial:
            min_cluster_size = trial.suggest_int("min_cluster_size", 5, 50)
            min_samples = trial.suggest_int("min_samples", 1, 20)
        else:
            min_cluster_size = 15
            min_samples = 5

        model = hdbscan.HDBSCAN(
            min_cluster_size=min_cluster_size,
            min_samples=min_samples,
            core_dist_n_jobs=-1
        )
        labels = model.fit_predict(X)

        metrics = calculate_clustering_metrics(X, labels)

        return {
            "model": model,
            "labels": labels,
            "metrics": metrics,
            "params": {"min_cluster_size": min_cluster_size, "min_samples": min_samples}
        }

    def _train_isolation_forest(self, X: np.ndarray, trial: Optional['optuna.Trial'] = None) -> Dict[str, Any]:
        """
        Train Isolation Forest for anomaly detection.
        """
        if not SKLEARN_AVAILABLE:
            return {"error": "sklearn not available"}

        # Hyperparameters
        if trial:
            contamination = trial.suggest_float("contamination", 0.01, 0.3)
            n_estimators = trial.suggest_int("n_estimators", 50, 200)
        else:
            contamination = 0.1
            n_estimators = 100

        model = IsolationForest(
            contamination=contamination,
            n_estimators=n_estimators,
            random_state=42,
            n_jobs=-1
        )
        predictions = model.fit_predict(X)
        labels = (predictions == -1).astype(int)  # -1 = anomaly

        n_anomalies = int(np.sum(labels))

        return {
            "model": model,
            "labels": labels,
            "metrics": {
                "n_anomalies": n_anomalies,
                "anomaly_rate": n_anomalies / len(X),
                "contamination": contamination,
                "n_estimators": n_estimators
            },
            "params": {"contamination": contamination, "n_estimators": n_estimators},
            "anomaly_indices": np.where(labels == 1)[0].tolist()
        }

    def _train_lof(self, X: np.ndarray, trial: Optional['optuna.Trial'] = None) -> Dict[str, Any]:
        """
        Train Local Outlier Factor for anomaly detection.
        """
        if not SKLEARN_AVAILABLE:
            return {"error": "sklearn not available"}

        # Hyperparameters
        if trial:
            n_neighbors = trial.suggest_int("n_neighbors", 5, 50)
            contamination = trial.suggest_float("contamination", 0.01, 0.3)
        else:
            n_neighbors = 20
            contamination = 0.1

        model = LocalOutlierFactor(
            n_neighbors=n_neighbors,
            contamination=contamination,
            n_jobs=-1,
            novelty=True  # For prediction on new data
        )
        model.fit(X)
        predictions = model.predict(X)
        labels = (predictions == -1).astype(int)  # -1 = anomaly

        n_anomalies = int(np.sum(labels))

        return {
            "model": model,
            "labels": labels,
            "metrics": {
                "n_anomalies": n_anomalies,
                "anomaly_rate": n_anomalies / len(X),
                "n_neighbors": n_neighbors,
                "contamination": contamination
            },
            "params": {"n_neighbors": n_neighbors, "contamination": contamination},
            "anomaly_indices": np.where(labels == 1)[0].tolist()
        }

    def _optimize_model(self, model_name: str, X: np.ndarray) -> Dict[str, Any]:
        """
        Use Optuna to find best hyperparameters for a model.
        """
        if not OPTUNA_AVAILABLE:
            logger.warning("[ModelTrainer] Optuna not available, using defaults")
            return self._train_model(model_name, X, None)

        train_func = {
            "dbscan": self._train_dbscan,
            "kmeans": self._train_kmeans,
            "hdbscan": self._train_hdbscan,
            "isolation_forest": self._train_isolation_forest,
            "lof": self._train_lof
        }.get(model_name)

        if not train_func:
            return {"error": f"Unknown model: {model_name}"}

        def objective(trial):
            try:
                result = train_func(X, trial)
                if "error" in result:
                    return -1.0

                metrics = result.get("metrics", {})

                # For clustering: use silhouette
                if model_name in ["dbscan", "kmeans", "hdbscan"]:
                    score = metrics.get("silhouette_score", -1)
                    return score if score is not None else -1

                # For anomaly detection: balance anomaly rate
                else:
                    # Target anomaly rate around 5-15%
                    rate = metrics.get("anomaly_rate", 0)
                    target = 0.1
                    return -abs(rate - target)  # Closer to target is better

            except Exception as e:
                logger.debug(f"Trial failed: {e}")
                return -1.0

        # Create and run study
        study = optuna.create_study(
            direction="maximize",
            sampler=TPESampler(seed=42)
        )

        study.optimize(
            objective,
            n_trials=self.config.n_optuna_trials,
            timeout=self.config.optuna_timeout_seconds,
            show_progress_bar=True
        )

        logger.info(f"[ModelTrainer] {model_name} best params: {study.best_params}")
        logger.info(f"[ModelTrainer] {model_name} best score: {study.best_value:.4f}")

        # Train with best params
        best_result = train_func(X, None)  # Use defaults as base
        # Override with best params
        if study.best_params:
            # Re-train with best params would require custom logic
            # For now, we just log the best params
            best_result["best_params"] = study.best_params
            best_result["best_score"] = study.best_value
            best_result["study_name"] = study.study_name

        return best_result

    def _train_model(self, model_name: str, X: np.ndarray, trial=None) -> Dict[str, Any]:
        """Train a single model"""
        train_funcs = {
            "dbscan": self._train_dbscan,
            "kmeans": self._train_kmeans,
            "hdbscan": self._train_hdbscan,
            "isolation_forest": self._train_isolation_forest,
            "lof": self._train_lof
        }

        func = train_funcs.get(model_name)
        if func:
            return func(X, trial)
        return {"error": f"Unknown model: {model_name}"}

    def initiate_model_trainer(self, feature_path: str) -> ModelTrainerArtifact:
        """
        Execute model training pipeline.
        
        Args:
            feature_path: Path to feature matrix (.npy)
            
        Returns:
            ModelTrainerArtifact with results
        """
        logger.info(f"[ModelTrainer] Starting training: {feature_path}")
        start_time = datetime.now()

        # Load features
        X = np.load(feature_path)
        logger.info(f"[ModelTrainer] Loaded features: {X.shape}")

        # Start MLflow run
        mlflow_run_id = ""
        mlflow_experiment_id = ""

        if MLFLOW_AVAILABLE:
            try:
                run = mlflow.start_run()
                mlflow_run_id = run.info.run_id
                mlflow_experiment_id = run.info.experiment_id

                mlflow.log_param("n_samples", X.shape[0])
                mlflow.log_param("n_features", X.shape[1])
                mlflow.log_param("models", self.config.models_to_train)
            except Exception as e:
                logger.warning(f"[ModelTrainer] MLflow run start error: {e}")

        # Train all models
        trained_models = []
        best_model = None
        best_score = -float('inf')

        for model_name in self.config.models_to_train:
            logger.info(f"[ModelTrainer] Training {model_name}...")

            try:
                result = self._optimize_model(model_name, X)

                if "error" in result:
                    logger.warning(f"[ModelTrainer] {model_name} error: {result['error']}")
                    continue

                # Save model
                model_path = Path(self.config.output_directory) / f"{model_name}_model.joblib"
                joblib.dump(result["model"], model_path)

                # Log to MLflow
                if MLFLOW_AVAILABLE:
                    try:
                        mlflow.log_params({f"{model_name}_{k}": v for k, v in result.get("params", {}).items()})
                        mlflow.log_metrics({f"{model_name}_{k}": v for k, v in result.get("metrics", {}).items() if isinstance(v, (int, float))})
                        mlflow.sklearn.log_model(result["model"], model_name)
                    except Exception as e:
                        logger.debug(f"MLflow log error: {e}")

                # Track results
                model_info = {
                    "name": model_name,
                    "path": str(model_path),
                    "params": result.get("params", {}),
                    "metrics": result.get("metrics", {})
                }
                trained_models.append(model_info)

                # Check if best (for clustering models)
                score = result.get("metrics", {}).get("silhouette_score", -1)
                if score and score > best_score:
                    best_score = score
                    best_model = model_info

                logger.info(f"[ModelTrainer] ✓ {model_name} complete")

            except Exception as e:
                logger.error(f"[ModelTrainer] {model_name} failed: {e}")

        # End MLflow run
        if MLFLOW_AVAILABLE:
            try:
                mlflow.end_run()
            except Exception:
                pass

        # Calculate duration
        duration = (datetime.now() - start_time).total_seconds()

        # Get anomaly info from best anomaly detector
        n_anomalies = None
        anomaly_indices = None
        for model_info in trained_models:
            if model_info["name"] in ["isolation_forest", "lof"]:
                n_anomalies = model_info["metrics"].get("n_anomalies")
                break

        # Build artifact
        artifact = ModelTrainerArtifact(
            best_model_name=best_model["name"] if best_model else "",
            best_model_path=best_model["path"] if best_model else "",
            best_model_metrics=best_model["metrics"] if best_model else {},
            trained_models=trained_models,
            mlflow_run_id=mlflow_run_id,
            mlflow_experiment_id=mlflow_experiment_id,
            n_clusters=best_model["metrics"].get("n_clusters") if best_model else None,
            n_anomalies=n_anomalies,
            anomaly_indices=anomaly_indices,
            training_duration_seconds=duration,
            optuna_study_name=None
        )

        logger.info(f"[ModelTrainer] Training complete in {duration:.1f}s")
        logger.info(f"[ModelTrainer] Best model: {best_model['name'] if best_model else 'N/A'}")

        # ============================================
        # TRAIN PER-LANGUAGE MODELS FOR LIVE INFERENCE
        # ============================================
        # Different BERT models produce embeddings in different vector spaces.
        # We train separate Isolation Forest models per language to avoid
        # mixing incompatible embeddings.
        try:
            # Check if features include extra metadata (> 768 dims)
            if X.shape[1] > 768:
                X_embeddings = X[:, :768]  # Extract BERT embeddings only
            else:
                X_embeddings = X

            logger.info(f"[ModelTrainer] Training per-language models on {X_embeddings.shape[0]} samples...")

            # Load language labels from the same directory as features
            feature_dir = Path(feature_path).parent
            lang_files = list(feature_dir.glob("languages_*.npy"))
            
            if lang_files:
                # Get most recent language file
                latest_lang_file = max(lang_files, key=lambda p: p.stem)
                languages = np.load(latest_lang_file, allow_pickle=True)
                logger.info(f"[ModelTrainer] Loaded language labels from {latest_lang_file.name}")
            else:
                # Fallback: try to load from transformed data parquet
                parquet_files = list(feature_dir.glob("transformed_*.parquet"))
                if parquet_files:
                    import pandas as pd
                    latest_parquet = max(parquet_files, key=lambda p: p.stem)
                    df_temp = pd.read_parquet(latest_parquet)
                    if "language" in df_temp.columns:
                        languages = df_temp["language"].values
                        logger.info(f"[ModelTrainer] Loaded {len(languages)} language labels from parquet")
                    else:
                        languages = np.array(["english"] * len(X_embeddings))
                        logger.warning("[ModelTrainer] No language column in parquet, defaulting to english")
                else:
                    languages = np.array(["english"] * len(X_embeddings))
                    logger.warning("[ModelTrainer] No language data found, defaulting to english")

            # Train per-language models
            MIN_SAMPLES_PER_LANGUAGE = 10
            per_lang_models = {}

            for lang in ["english", "sinhala", "tamil"]:
                lang_mask = languages == lang
                X_lang = X_embeddings[lang_mask]

                if len(X_lang) >= MIN_SAMPLES_PER_LANGUAGE:
                    logger.info(f"[ModelTrainer] Training {lang} model on {len(X_lang)} samples...")
                    
                    lang_model = IsolationForest(
                        contamination=0.1,
                        n_estimators=100,
                        random_state=42,
                        n_jobs=-1
                    )
                    lang_model.fit(X_lang)

                    # Save per-language model
                    model_path = Path(self.config.output_directory) / f"isolation_forest_{lang}.joblib"
                    joblib.dump(lang_model, model_path)
                    per_lang_models[lang] = str(model_path)

                    logger.info(f"[ModelTrainer] ✓ Saved: isolation_forest_{lang}.joblib ({len(X_lang)} samples)")
                else:
                    logger.warning(f"[ModelTrainer] Skipping {lang}: only {len(X_lang)} samples (min: {MIN_SAMPLES_PER_LANGUAGE})")

            # Also save a legacy "embeddings_only" model for backward compatibility (trained on English)
            if "english" in per_lang_models:
                import shutil
                english_model_path = Path(per_lang_models["english"])
                legacy_path = Path(self.config.output_directory) / "isolation_forest_embeddings_only.joblib"
                shutil.copy(english_model_path, legacy_path)
                logger.info(f"[ModelTrainer] ✓ Legacy model copied: isolation_forest_embeddings_only.joblib")

            logger.info(f"[ModelTrainer] Per-language training complete: {list(per_lang_models.keys())}")

        except Exception as e:
            logger.warning(f"[ModelTrainer] Per-language model training failed: {e}")
            import traceback
            traceback.print_exc()

        return artifact