File size: 35,060 Bytes
46cc63a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Transformer fine-tuning (DistilBERT, Toxic-BERT, etc.) with partial or head-only
freezing, label smoothing, gap-aware early stopping, and val threshold tuning.
"""

from __future__ import annotations

from pathlib import Path
from typing import Any

import numpy as np
import torch
import torch.nn as nn
from sklearn.metrics import f1_score, precision_score, recall_score, roc_auc_score
from transformers import (
    AutoModelForSequenceClassification,
    AutoTokenizer,
    DataCollatorWithPadding,
    Trainer,
    TrainerCallback,
    TrainingArguments,
    set_seed,
)
from src.evaluation.threshold_tuning import predict_with_threshold, search_best_threshold
from src.utils.logger import get_logger

logger = get_logger(__name__)


def _bert_encoder_layers(model) -> list[nn.Module]:
    """Return transformer blocks for BERT / Toxic-BERT."""
    if hasattr(model, "bert") and hasattr(model.bert, "encoder"):
        return list(model.bert.encoder.layer)
    if hasattr(model, "distilbert"):
        return list(model.distilbert.transformer.layer)
    raise AttributeError("Unsupported architecture for layer freeze")


def _distilbert_layers(model) -> list[nn.Module]:
    """Return transformer blocks for DistilBERT (6 layers)."""
    return list(model.distilbert.transformer.layer)


def unfreeze_full_encoder(model) -> int:
    """Train all encoder blocks plus classification head (Final Squeeze / full BERT)."""
    for param in model.parameters():
        param.requires_grad = True

    layers = _bert_encoder_layers(model)
    trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
    total = sum(p.numel() for p in model.parameters())
    logger.info(
        f"Full unfreeze — all {len(layers)} encoder blocks + head — "
        f"trainable {trainable:,}/{total:,} ({100 * trainable / total:.1f}%)"
    )
    return len(layers)


def freeze_all_inference(model) -> int:
    """Freeze every parameter — pretrained inference only (Golden Baseline)."""
    for param in model.parameters():
        param.requires_grad = False
    layers = _bert_encoder_layers(model)
    logger.info(
        f"Inference-only — all {len(layers)} encoder blocks + head frozen (zero fine-tuning)"
    )
    return len(layers)


def freeze_head_only(model) -> None:
    """Freeze entire backbone; train classification head only (Expert / Toxic-BERT)."""
    for param in model.parameters():
        param.requires_grad = False

    for name, param in model.named_parameters():
        if any(k in name for k in ("classifier", "pre_classifier", "pooler")):
            param.requires_grad = True

    trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
    total = sum(p.numel() for p in model.parameters())
    logger.info(
        f"Head-only freeze — trainable {trainable:,}/{total:,} "
        f"({100 * trainable / total:.2f}%)"
    )


def freeze_encoder_partial(model, freeze_first_n: int = 4) -> int:
    """Freeze first N encoder blocks; train remaining blocks + classification head."""
    for param in model.parameters():
        param.requires_grad = False

    layers = _bert_encoder_layers(model)
    for layer in layers[freeze_first_n:]:
        for param in layer.parameters():
            param.requires_grad = True

    for name, param in model.named_parameters():
        if any(k in name for k in ("classifier", "pre_classifier", "pooler")):
            param.requires_grad = True

    trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
    total = sum(p.numel() for p in model.parameters())
    n_train = len(layers) - freeze_first_n
    logger.info(
        f"Partial freeze: {freeze_first_n}/{len(layers)} blocks frozen — "
        f"training last {n_train} + head — "
        f"trainable {trainable:,}/{total:,} ({100 * trainable / total:.1f}%)"
    )
    return len(layers)


def freeze_distilbert_partial(model, freeze_first_n: int = 4) -> None:
    """Backward-compatible alias for DistilBERT partial freeze."""
    freeze_encoder_partial(model, freeze_first_n=freeze_first_n)


def build_head_only_optimizer(
    model,
    *,
    learning_rate: float = 2e-5,
    weight_decay: float = 0.01,
) -> torch.optim.Optimizer:
    params = [p for p in model.parameters() if p.requires_grad]
    return torch.optim.AdamW(params, lr=learning_rate, weight_decay=weight_decay)


def build_full_optimizer(
    model,
    *,
    learning_rate: float = 5e-6,
    weight_decay: float = 0.01,
) -> torch.optim.Optimizer:
    params = [p for p in model.parameters() if p.requires_grad]
    return torch.optim.AdamW(params, lr=learning_rate, weight_decay=weight_decay)


def build_partial_optimizer(
    model,
    *,
    learning_rate: float = 1e-5,
    encoder_learning_rate: float | None = None,
    head_learning_rate: float | None = None,
    weight_decay: float = 0.01,
    freeze_first_n: int = 4,
) -> torch.optim.Optimizer:
    """
    Parameter groups: frozen layers excluded; top encoder blocks + head (optional split LRs).
    """
    enc_lr = encoder_learning_rate if encoder_learning_rate is not None else learning_rate
    head_lr = head_learning_rate if head_learning_rate is not None else learning_rate

    layers = _bert_encoder_layers(model)
    top_layer_ids = {
        id(p) for layer in layers[freeze_first_n:] for p in layer.parameters()
    }
    head_params = [
        p
        for n, p in model.named_parameters()
        if p.requires_grad and ("classifier" in n or "pre_classifier" in n)
    ]
    head_ids = {id(p) for p in head_params}
    top_params = [
        p
        for p in model.parameters()
        if p.requires_grad and id(p) in top_layer_ids and id(p) not in head_ids
    ]

    groups = []
    if top_params:
        groups.append(
            {"params": top_params, "lr": enc_lr, "weight_decay": weight_decay}
        )
    if head_params:
        groups.append(
            {"params": head_params, "lr": head_lr, "weight_decay": weight_decay}
        )

    if not groups:
        groups = [{"params": [p for p in model.parameters() if p.requires_grad]}]

    return torch.optim.AdamW(groups)


def _average_state_dicts(state_dicts: list[dict]) -> dict:
    """Element-wise mean of compatible state dicts (Stochastic Weight Averaging)."""
    if not state_dicts:
        raise ValueError("state_dicts must not be empty")
    avg = {k: v.clone().float() for k, v in state_dicts[0].items()}
    for sd in state_dicts[1:]:
        for k in avg:
            avg[k] += sd[k].float()
    n = float(len(state_dicts))
    return {k: (v / n).to(state_dicts[0][k].dtype) for k, v in avg.items()}


def build_distilbert_optimizer(
    model,
    *,
    learning_rate: float = 1e-5,
    weight_decay: float = 0.01,
    freeze_first_n: int = 4,
) -> torch.optim.Optimizer:
    return build_partial_optimizer(
        model,
        learning_rate=learning_rate,
        weight_decay=weight_decay,
        freeze_first_n=freeze_first_n,
    )


def logits_to_toxic_prob(logits: np.ndarray | torch.Tensor) -> np.ndarray:
    """Map model logits to P(toxic): sigmoid on 'toxic' for 6-label BERT, else softmax."""
    t = torch.as_tensor(logits, dtype=torch.float32)
    if t.ndim == 1:
        t = t.unsqueeze(0)
    if t.shape[-1] >= 6:
        return torch.sigmoid(t)[:, 0].numpy()
    return torch.softmax(t, dim=-1)[:, 1].numpy()


def compute_hf_metrics(eval_pred) -> dict[str, float]:
    logits, labels = eval_pred
    probs = logits_to_toxic_prob(logits)
    preds = np.argmax(logits, axis=1)
    return {
        "f1_toxic": float(f1_score(labels, preds, pos_label=1, zero_division=0)),
        "f1_weighted": float(f1_score(labels, preds, average="weighted", zero_division=0)),
        "precision": float(precision_score(labels, preds, pos_label=1, zero_division=0)),
        "recall": float(recall_score(labels, preds, pos_label=1, zero_division=0)),
        "roc_auc": float(roc_auc_score(labels, probs)),
    }


def _symmetric_kl(logits_a: torch.Tensor, logits_b: torch.Tensor) -> torch.Tensor:
    """Symmetric KL between two logit vectors (R-Drop regularization)."""
    log_p = nn.functional.log_softmax(logits_a, dim=-1)
    log_q = nn.functional.log_softmax(logits_b, dim=-1)
    p = log_p.exp()
    q = log_q.exp()
    kl_pq = nn.functional.kl_div(log_p, q, reduction="batchmean", log_target=False)
    kl_qp = nn.functional.kl_div(log_q, p, reduction="batchmean", log_target=False)
    return (kl_pq + kl_qp) / 2.0


class LabelSmoothingTrainer(Trainer):
    """Cross-entropy with label smoothing for the classification head."""

    def __init__(self, *args, label_smoothing: float = 0.1, **kwargs):
        super().__init__(*args, **kwargs)
        self.label_smoothing = label_smoothing

    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
        labels = inputs.pop("labels")
        outputs = model(**inputs)
        loss = nn.functional.cross_entropy(
            outputs.logits,
            labels,
            label_smoothing=self.label_smoothing,
        )
        return (loss, outputs) if return_outputs else loss


class RDropTrainer(LabelSmoothingTrainer):
    """R-Drop: dual forward passes + symmetric KL to limit overfitting (Performance Squeeze)."""

    def __init__(self, *args, rdrop_alpha: float = 0.5, **kwargs):
        super().__init__(*args, **kwargs)
        self.rdrop_alpha = rdrop_alpha

    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
        labels = inputs.pop("labels")
        outputs1 = model(**inputs)
        outputs2 = model(**inputs)
        ce = (
            nn.functional.cross_entropy(
                outputs1.logits, labels, label_smoothing=self.label_smoothing
            )
            + nn.functional.cross_entropy(
                outputs2.logits, labels, label_smoothing=self.label_smoothing
            )
        ) / 2.0
        kl = _symmetric_kl(outputs1.logits, outputs2.logits)
        loss = ce + self.rdrop_alpha * kl
        return (loss, outputs1) if return_outputs else loss

    def create_optimizer(self):
        if self.optimizer is None:
            cfg = self.args
            freeze_mode = getattr(self, "_freeze_mode", "partial")
            if freeze_mode == "head_only":
                self.optimizer = build_head_only_optimizer(
                    self.model,
                    learning_rate=cfg.learning_rate,
                    weight_decay=cfg.weight_decay,
                )
            elif freeze_mode == "full_unfreeze":
                self.optimizer = build_full_optimizer(
                    self.model,
                    learning_rate=cfg.learning_rate,
                    weight_decay=cfg.weight_decay,
                )
            else:
                freeze_n = getattr(self, "_freeze_first_n", 4)
                enc_lr = getattr(self, "_encoder_lr", None)
                head_lr = getattr(self, "_head_lr", None)
                self.optimizer = build_partial_optimizer(
                    self.model,
                    learning_rate=cfg.learning_rate,
                    encoder_learning_rate=enc_lr,
                    head_learning_rate=head_lr,
                    weight_decay=cfg.weight_decay,
                    freeze_first_n=freeze_n,
                )
        return self.optimizer


class SWACallback(TrainerCallback):
    """
    Average model weights over the last N completed epochs (Stochastic Weight Averaging).

    Applied in ``on_train_end`` after ``load_best_model_at_end`` so inference uses SWA weights.
    """

    def __init__(self, last_n_epochs: int = 5):
        self.last_n_epochs = last_n_epochs
        self._snapshots: list[dict[str, torch.Tensor]] = []

    def on_epoch_end(self, args, state, control, model=None, **kwargs):
        if model is None:
            return control
        self._snapshots.append(
            {k: v.detach().cpu().clone() for k, v in model.state_dict().items()}
        )
        return control

    def on_train_end(self, args, state, control, model=None, **kwargs):
        if model is None or not self._snapshots:
            return control
        n_use = min(self.last_n_epochs, len(self._snapshots))
        averaged = _average_state_dicts(self._snapshots[-n_use:])
        model.load_state_dict(averaged, strict=True)
        logger.info(
            f"SWA applied — averaged last {n_use}/{len(self._snapshots)} epoch checkpoints"
        )
        return control


class GapEarlyStoppingCallback(TrainerCallback):
    """
    Stop when validation F1 plateaus OR |train_f1 - val_f1| exceeds max_gap.

    Train F1 is computed on ``train_eval_dataset`` each evaluation step.
    Attach ``trainer`` after constructing the Trainer (see train_distilbert_stable).
    """

    def __init__(
        self,
        train_eval_dataset,
        *,
        patience: int = 3,
        max_gap: float = 0.045,
        metric: str = "f1_toxic",
        gap_check_min_epoch: int = 2,
        gap_stop_enabled: bool = False,
    ):
        self.train_eval_dataset = train_eval_dataset
        self.patience = patience
        self.max_gap = max_gap
        self.metric = metric
        self.gap_check_min_epoch = gap_check_min_epoch
        self.gap_stop_enabled = gap_stop_enabled
        self.best_metric = -1.0
        self.bad_epochs = 0
        self.trainer: Trainer | None = None
        self._checking = False

    def on_evaluate(self, args, state, control, metrics=None, **kwargs):
        if metrics is None or self.trainer is None or self._checking:
            return control

        val_f1 = metrics.get(f"eval_{self.metric}", metrics.get(self.metric, 0.0))

        self._checking = True
        try:
            train_metrics = self.trainer.evaluate(
                eval_dataset=self.train_eval_dataset,
                metric_key_prefix="train",
            )
        finally:
            self._checking = False

        train_f1 = train_metrics.get(f"train_{self.metric}", 0.0)
        gap = abs(train_f1 - val_f1)
        logger.info(
            f"Gap monitor — train_f1={train_f1:.4f} val_f1={val_f1:.4f} gap={gap:.4f}"
        )

        epoch = int(state.epoch or 0)
        if (
            self.gap_stop_enabled
            and epoch >= self.gap_check_min_epoch
            and gap > self.max_gap
        ):
            logger.warning(
                f"Gap defense — train-val gap {gap:.4f} > {self.max_gap}; "
                "stopping and reverting to best checkpoint"
            )
            control.should_training_stop = True
            return control

        if val_f1 > self.best_metric:
            self.best_metric = val_f1
            self.bad_epochs = 0
        else:
            self.bad_epochs += 1
            if self.bad_epochs >= self.patience:
                logger.info(
                    f"Early stop: no {self.metric} improvement for {self.patience} epochs"
                )
                control.should_training_stop = True

        return control


def _predict_probs_from_dataset(
    trainer: Trainer,
    tokenized_dataset,
) -> np.ndarray:
    ds = tokenized_dataset
    if "label" in ds.column_names:
        ds = ds.remove_columns(["label"])
    out = trainer.predict(ds)
    return logits_to_toxic_prob(out.predictions)


def predict_with_tta(
    trainer: Trainer,
    tokenizer,
    texts: list[str],
    labels_placeholder: list[int] | None,
    *,
    max_length: int,
    aug_cfg: dict,
) -> np.ndarray:
    """
    Test-time augmentation: average P(toxic) from original and back-translated texts.
    """
    from datasets import Dataset

    from src.features.augmentation import back_translate_texts

    labels = labels_placeholder if labels_placeholder is not None else [0] * len(texts)

    def _tokenize(ds):
        return tokenizer(ds["text"], truncation=True, max_length=max_length)

    def _prep(raw_texts: list[str]):
        ds = Dataset.from_dict({"text": raw_texts, "label": labels[: len(raw_texts)]})
        tok = ds.map(_tokenize, batched=True)
        drop_cols = [
            c for c in tok.column_names if c not in ("input_ids", "attention_mask", "label")
        ]
        if drop_cols:
            tok = tok.remove_columns(drop_cols)
        tok.set_format("torch")
        return tok

    original_probs = _predict_probs_from_dataset(trainer, _prep(texts))

    if not aug_cfg.get("enabled", False):
        return original_probs

    logger.info(f"TTA — back-translating {len(texts)} test comments")
    bt_texts = back_translate_texts(
        texts,
        source_lang=aug_cfg.get("source_lang", "en"),
        pivot_lang=aug_cfg.get("pivot_lang", "de"),
        max_words=int(aug_cfg.get("max_words", 60)),
        rate_limit_every=int(aug_cfg.get("rate_limit_every", 50)),
        rate_limit_sleep_sec=float(aug_cfg.get("rate_limit_sleep_sec", 1.0)),
    )
    bt_probs = _predict_probs_from_dataset(trainer, _prep(bt_texts))
    averaged = (original_probs + bt_probs) / 2.0
    logger.info("TTA — averaged original and back-translated probabilities")
    return averaged


def _transformer_cfg(cfg: dict) -> dict:
    """Resolve transformer section (expert) or legacy distilbert key."""
    if "transformer" in cfg:
        return cfg["transformer"]
    return cfg["distilbert"]


def _apply_model_freeze(model, bert_cfg: dict) -> tuple[str, int]:
    freeze_mode = bert_cfg.get("freeze_mode", "partial")
    if freeze_mode == "head_only":
        freeze_head_only(model)
        return "head_only", 0

    if freeze_mode in ("full", "full_unfreeze", "all_layers"):
        n_layers = unfreeze_full_encoder(model)
        return "full_unfreeze", 0

    if freeze_mode in ("inference_only", "frozen", "golden_baseline"):
        n_layers = freeze_all_inference(model)
        return "inference_only", n_layers

    layers = _bert_encoder_layers(model)
    if freeze_mode in ("last_n_layers", "train_last_n"):
        train_last_n = int(bert_cfg.get("train_last_n_layers", 4))
        freeze_first_n = max(0, len(layers) - train_last_n)
    else:
        freeze_first_n = int(bert_cfg.get("freeze_first_n_layers", 4))

    freeze_encoder_partial(model, freeze_first_n=freeze_first_n)
    return f"partial_last_{len(layers) - freeze_first_n}", freeze_first_n


def _bert_metrics_from_probs(
    *,
    model_label: str,
    model_id: str,
    freeze_mode: str,
    y_train: np.ndarray,
    y_test: np.ndarray,
    train_probs: np.ndarray,
    test_probs: np.ndarray,
    threshold: float,
    val_f1_at_threshold: float | None = None,
    extra: dict | None = None,
) -> dict[str, Any]:
    train_preds = predict_with_threshold(train_probs, threshold)
    test_preds = predict_with_threshold(test_probs, threshold)
    f1_train = float(f1_score(y_train, train_preds, average="weighted", zero_division=0))
    f1_test = float(f1_score(y_test, test_preds, average="weighted", zero_division=0))
    f1_toxic_test = float(f1_score(y_test, test_preds, pos_label=1, zero_division=0))
    f1_toxic_train = float(f1_score(y_train, train_preds, pos_label=1, zero_division=0))
    gap_weighted = abs(f1_train - f1_test)
    gap_toxic = abs(f1_toxic_train - f1_toxic_test)
    metrics = {
        "model": model_label,
        "model_id": model_id,
        "freeze_mode": freeze_mode,
        "threshold": round(threshold, 4),
        "val_f1_at_threshold": round(val_f1_at_threshold, 4) if val_f1_at_threshold else None,
        "f1_weighted": round(f1_test, 4),
        "f1_toxic": round(f1_toxic_test, 4),
        "f1_toxic_train": round(f1_toxic_train, 4),
        "train_test_gap_toxic": round(gap_toxic, 4),
        "train_test_gap_toxic_pp": round(gap_toxic * 100, 2),
        "f1_train": round(f1_train, 4),
        "train_test_gap": round(gap_weighted, 4),
        "train_test_gap_pp": round(gap_weighted * 100, 2),
        "gap_ok": gap_weighted < 0.05,
        "roc_auc": round(float(roc_auc_score(y_test, test_probs)), 4),
        "fp": int(((y_test == 0) & (test_preds == 1)).sum()),
        "fn": int(((y_test == 1) & (test_preds == 0)).sum()),
    }
    if extra:
        metrics.update(extra)
    return metrics


def evaluate_pretrained_bert_baseline(
    hf_train,
    hf_val,
    hf_test,
    y_train: np.ndarray,
    y_test: np.ndarray,
    y_val: np.ndarray,
    cfg: dict,
    *,
    seed: int = 42,
    model_label: str = "Golden-Baseline-Toxic-BERT",
) -> dict[str, Any]:
    """
    Step 1 — pretrained Toxic-BERT with all layers frozen (no fine-tuning on project data).
    """
    bert_cfg = cfg.get("baseline", _transformer_cfg(cfg))
    model_id = bert_cfg["model_id"]
    set_seed(seed)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    tokenizer = AutoTokenizer.from_pretrained(model_id)
    max_len = int(bert_cfg.get("max_length", 128))

    def _tokenize(ds):
        return tokenizer(ds["text"], truncation=True, max_length=max_len)

    def _prep(ds):
        tok = ds.map(_tokenize, batched=True)
        drop_cols = [
            c for c in tok.column_names if c not in ("input_ids", "attention_mask", "label")
        ]
        if drop_cols:
            tok = tok.remove_columns(drop_cols)
        tok.set_format("torch")
        return tok

    tok_train = _prep(hf_train)
    tok_val = _prep(hf_val)
    tok_test = _prep(hf_test)

    model = AutoModelForSequenceClassification.from_pretrained(model_id)
    model.to(device)
    _apply_model_freeze(model, {**bert_cfg, "freeze_mode": "inference_only"})
    model.eval()

    args = TrainingArguments(
        output_dir="/tmp/golden_baseline_eval",
        per_device_eval_batch_size=int(bert_cfg.get("batch_size", 8)),
        report_to="none",
        seed=seed,
    )
    trainer = Trainer(
        model=model,
        args=args,
        data_collator=DataCollatorWithPadding(tokenizer),
        compute_metrics=compute_hf_metrics,
    )

    logger.info(f"Golden Baseline — {model_id} (inference only, no training)")
    tok_val_pred = tok_val.remove_columns(["label"]) if "label" in tok_val.column_names else tok_val
    tok_train_pred = (
        tok_train.remove_columns(["label"]) if "label" in tok_train.column_names else tok_train
    )
    tok_test_pred = (
        tok_test.remove_columns(["label"]) if "label" in tok_test.column_names else tok_test
    )

    val_probs = logits_to_toxic_prob(trainer.predict(tok_val_pred).predictions)
    train_probs = logits_to_toxic_prob(trainer.predict(tok_train_pred).predictions)
    test_probs = logits_to_toxic_prob(trainer.predict(tok_test_pred).predictions)

    y_val_arr = np.asarray(y_val).astype(int)
    y_train_arr = np.asarray(y_train).astype(int)
    y_test_arr = np.asarray(y_test).astype(int)

    th_cfg = bert_cfg.get("threshold_tuning", {})
    threshold = 0.5
    val_f1_at_threshold = None
    if th_cfg.get("enabled", True):
        threshold, val_f1_at_threshold = search_best_threshold(
            y_val_arr,
            val_probs,
            metric=th_cfg.get("metric", "f1_weighted"),
            min_threshold=float(th_cfg.get("min_threshold", 0.05)),
            max_threshold=float(th_cfg.get("max_threshold", 0.95)),
            step=float(th_cfg.get("step", 0.01)),
        )

    metrics = _bert_metrics_from_probs(
        model_label=model_label,
        model_id=model_id,
        freeze_mode="inference_only",
        y_train=y_train_arr,
        y_test=y_test_arr,
        train_probs=train_probs,
        test_probs=test_probs,
        threshold=threshold,
        val_f1_at_threshold=val_f1_at_threshold,
        extra={
            "trained": False,
            "num_labels": int(model.config.num_labels),
            "prob_mode": "sigmoid_toxic" if model.config.num_labels >= 6 else "softmax",
            "esencial_compliant_gap": True,
        },
    )
    metrics["gap_ok"] = metrics["train_test_gap"] < 0.01

    return {
        "metrics": metrics,
        "trainer": trainer,
        "tokenizer": tokenizer,
        "test_probs": test_probs,
        "val_probs": val_probs,
        "train_probs": train_probs,
        "threshold": threshold,
    }


def train_transformer_stable(
    hf_train,
    hf_val,
    hf_test,
    y_test: np.ndarray,
    y_val: np.ndarray,
    cfg: dict,
    output_dir: Path,
    *,
    seed: int = 42,
    model_label: str = "Transformer-stable",
) -> dict[str, Any]:
    """
    Fine-tune a sequence classifier with stability-focused regularization.

    Returns metrics, trainer, tokenizer, test probabilities, and optimal threshold.
    """
    bert_cfg = _transformer_cfg(cfg)
    model_id = bert_cfg["model_id"]

    set_seed(seed)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    tokenizer = AutoTokenizer.from_pretrained(model_id)
    max_len = int(bert_cfg.get("max_length", 128))

    def _tokenize(ds):
        return tokenizer(ds["text"], truncation=True, max_length=max_len)

    def _prep(ds):
        tok = ds.map(_tokenize, batched=True)
        drop_cols = [
            c for c in tok.column_names if c not in ("input_ids", "attention_mask", "label")
        ]
        if drop_cols:
            tok = tok.remove_columns(drop_cols)
        tok.set_format("torch")
        return tok

    tok_train = _prep(hf_train)
    tok_val = _prep(hf_val)
    tok_test = _prep(hf_test)

    model = AutoModelForSequenceClassification.from_pretrained(
        model_id,
        num_labels=2,
        ignore_mismatched_sizes=True,
    )
    model.config.problem_type = "single_label_classification"
    model.config.num_labels = 2
    dropout_p = float(bert_cfg.get("head_dropout", 0.5))
    if hasattr(model.config, "hidden_dropout_prob"):
        model.config.hidden_dropout_prob = dropout_p
    if hasattr(model.config, "seq_classif_dropout"):
        model.config.seq_classif_dropout = dropout_p
    if hasattr(model, "dropout"):
        model.dropout = nn.Dropout(dropout_p)
    model.to(device)

    freeze_mode, freeze_first_n = _apply_model_freeze(model, bert_cfg)

    es_cfg = bert_cfg.get("early_stopping", {})
    es_metric = es_cfg.get("metric", bert_cfg.get("metric_for_best", "f1_toxic"))
    gap_cb = GapEarlyStoppingCallback(
        tok_train,
        patience=int(es_cfg.get("patience", 3)),
        max_gap=float(es_cfg.get("max_train_val_gap", 0.045)),
        metric=es_metric,
        gap_check_min_epoch=int(es_cfg.get("gap_check_min_epoch", 2)),
        gap_stop_enabled=bool(es_cfg.get("gap_stop_enabled", False)),
    )

    callbacks: list[TrainerCallback] = [gap_cb]
    swa_cfg = bert_cfg.get("swa", {})
    swa_cb: SWACallback | None = None
    if swa_cfg.get("enabled", False):
        swa_cb = SWACallback(last_n_epochs=int(swa_cfg.get("last_n_epochs", 5)))
        callbacks.append(swa_cb)

    args = TrainingArguments(
        output_dir=str(output_dir),
        learning_rate=float(bert_cfg.get("learning_rate", 1e-5)),
        num_train_epochs=int(bert_cfg.get("max_epochs", 8)),
        per_device_train_batch_size=int(bert_cfg.get("batch_size", 8)),
        per_device_eval_batch_size=int(bert_cfg.get("batch_size", 8)),
        weight_decay=float(bert_cfg.get("weight_decay", 0.01)),
        eval_strategy="epoch",
        save_strategy="epoch",
        load_best_model_at_end=True,
        metric_for_best_model=bert_cfg.get("metric_for_best", "f1_toxic"),
        greater_is_better=True,
        warmup_ratio=float(bert_cfg.get("warmup_ratio", 0.1)),
        logging_steps=20,
        save_total_limit=2,
        fp16=torch.cuda.is_available(),
        report_to="none",
        seed=seed,
    )

    rdrop_cfg = bert_cfg.get("rdrop", {})
    label_smooth = float(bert_cfg.get("label_smoothing", 0.1))
    if rdrop_cfg.get("enabled", False):
        trainer = RDropTrainer(
            model=model,
            args=args,
            train_dataset=tok_train,
            eval_dataset=tok_val,
            data_collator=DataCollatorWithPadding(tokenizer),
            compute_metrics=compute_hf_metrics,
            callbacks=callbacks,
            label_smoothing=label_smooth,
            rdrop_alpha=float(rdrop_cfg.get("alpha", 0.5)),
        )
        rdrop_note = f", R-Drop α={rdrop_cfg.get('alpha', 0.5)}"
    else:
        trainer = LabelSmoothingTrainer(
            model=model,
            args=args,
            train_dataset=tok_train,
            eval_dataset=tok_val,
            data_collator=DataCollatorWithPadding(tokenizer),
            compute_metrics=compute_hf_metrics,
            callbacks=callbacks,
            label_smoothing=label_smooth,
        )
        rdrop_note = ""

    gap_cb.trainer = trainer
    trainer._freeze_mode = freeze_mode  # noqa: SLF001
    trainer._freeze_first_n = freeze_first_n  # noqa: SLF001
    if bert_cfg.get("encoder_learning_rate") is not None:
        trainer._encoder_lr = float(bert_cfg["encoder_learning_rate"])  # noqa: SLF001
    if bert_cfg.get("head_learning_rate") is not None:
        trainer._head_lr = float(bert_cfg["head_learning_rate"])  # noqa: SLF001

    enc_lr = bert_cfg.get("encoder_learning_rate", bert_cfg.get("learning_rate"))
    head_lr = bert_cfg.get("head_learning_rate", bert_cfg.get("learning_rate"))
    logger.info(
        f"Training {model_id} ({freeze_mode} freeze, enc_lr={enc_lr}, head_lr={head_lr}"
        f"{rdrop_note}"
        f"{', SWA last ' + str(swa_cfg.get('last_n_epochs', 5)) + ' epochs' if swa_cb else ''})..."
    )
    trainer.train()

    val_out = trainer.predict(tok_val)
    val_probs = logits_to_toxic_prob(val_out.predictions)
    y_val_arr = np.asarray(y_val).astype(int)

    th_cfg = bert_cfg.get("threshold_tuning", {})
    threshold = 0.5
    val_f1_at_threshold = None
    if th_cfg.get("enabled", False):
        metric = th_cfg.get("metric", "f1_toxic")
        threshold, val_f1_at_threshold = search_best_threshold(
            y_val_arr,
            val_probs,
            metric=metric,
            min_threshold=float(th_cfg.get("min_threshold", 0.05)),
            max_threshold=float(th_cfg.get("max_threshold", 0.95)),
            step=float(th_cfg.get("step", 0.01)),
        )
        th_step = float(th_cfg.get("step", 0.01))
        logger.info(
            f"Val threshold tuning — best_t={threshold:.3f} "
            f"val_{metric}={val_f1_at_threshold:.4f} (step={th_step})"
        )

    tta_cfg = bert_cfg.get("test_time_augmentation", {})
    test_texts = list(hf_test["text"])
    if tta_cfg.get("enabled", False):
        probs = predict_with_tta(
            trainer,
            tokenizer,
            test_texts,
            list(hf_test["label"]),
            max_length=max_len,
            aug_cfg=tta_cfg,
        )
        preds_default = (probs >= 0.5).astype(int)
    else:
        output = trainer.predict(tok_test)
        probs = logits_to_toxic_prob(output.predictions)
        preds_default = np.argmax(output.predictions, axis=1)
    preds = predict_with_threshold(probs, threshold)

    train_out = trainer.predict(tok_train)
    train_probs = logits_to_toxic_prob(train_out.predictions)
    train_preds = predict_with_threshold(train_probs, threshold)
    train_labels = np.asarray(hf_train["label"]).astype(int)
    y_test_arr = np.asarray(y_test).astype(int)

    f1_train = float(f1_score(train_labels, train_preds, average="weighted", zero_division=0))
    f1_test = float(f1_score(y_test_arr, preds, average="weighted", zero_division=0))
    f1_toxic_test = float(f1_score(y_test_arr, preds, pos_label=1, zero_division=0))
    f1_toxic_train = float(
        f1_score(train_labels, train_preds, pos_label=1, zero_division=0)
    )
    gap_weighted = abs(f1_train - f1_test)
    gap_toxic = abs(f1_toxic_train - f1_toxic_test)

    metrics = {
        "model": model_label,
        "model_id": model_id,
        "freeze_mode": freeze_mode,
        "rdrop_enabled": bool(rdrop_cfg.get("enabled", False)),
        "tta_enabled": bool(tta_cfg.get("enabled", False)),
        "swa_enabled": bool(swa_cfg.get("enabled", False)),
        "swa_epochs_averaged": (
            min(int(swa_cfg.get("last_n_epochs", 5)), len(swa_cb._snapshots))
            if swa_cb and swa_cb._snapshots
            else 0
        ),
        "threshold": round(threshold, 4),
        "threshold_step": float(th_cfg.get("step", 0.01)) if th_cfg.get("enabled") else None,
        "val_f1_at_threshold": round(val_f1_at_threshold, 4) if val_f1_at_threshold else None,
        "f1_weighted": round(f1_test, 4),
        "f1_toxic": round(f1_toxic_test, 4),
        "f1_toxic_train": round(f1_toxic_train, 4),
        "train_test_gap_toxic": round(gap_toxic, 4),
        "train_test_gap_toxic_pp": round(gap_toxic * 100, 2),
        "gap_toxic_ok": gap_toxic < 0.05,
        "f1_train": round(f1_train, 4),
        "train_test_gap": round(gap_weighted, 4),
        "train_test_gap_pp": round(gap_weighted * 100, 2),
        "f1_weighted_default_thresh": round(
            float(f1_score(y_test_arr, preds_default, average="weighted", zero_division=0)), 4
        ),
        "roc_auc": round(float(roc_auc_score(y_test_arr, probs)), 4),
        "fp": int(((y_test_arr == 0) & (preds == 1)).sum()),
        "fn": int(((y_test_arr == 1) & (preds == 0)).sum()),
    }

    trainer.save_model(str(output_dir))
    tokenizer.save_pretrained(str(output_dir))

    return {
        "metrics": metrics,
        "trainer": trainer,
        "tokenizer": tokenizer,
        "test_probs": probs,
        "test_preds": preds,
        "threshold": threshold,
        "val_probs": val_probs,
    }


def infer_transformer_probs(
    model_dir: Path,
    texts,
    *,
    max_length: int = 128,
    batch_size: int = 16,
) -> np.ndarray:
    """Load a saved classifier and return P(toxic) for each text."""
    model_dir = Path(model_dir)
    text_list = list(texts)
    tokenizer = AutoTokenizer.from_pretrained(str(model_dir))
    model = AutoModelForSequenceClassification.from_pretrained(str(model_dir))
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model.to(device)
    model.eval()

    probs: list[float] = []
    with torch.no_grad():
        for i in range(0, len(text_list), batch_size):
            batch = text_list[i : i + batch_size]
            enc = tokenizer(
                batch,
                truncation=True,
                max_length=max_length,
                padding=True,
                return_tensors="pt",
            )
            enc = {k: v.to(device) for k, v in enc.items()}
            logits = model(**enc).logits
            batch_probs = torch.softmax(logits, dim=1)[:, 1].cpu().numpy()
            probs.extend(batch_probs.tolist())

    return np.array(probs)


def train_distilbert_stable(
    hf_train,
    hf_val,
    hf_test,
    y_test: np.ndarray,
    cfg: dict,
    output_dir: Path,
    *,
    seed: int = 42,
) -> dict[str, Any]:
    """Backward-compatible wrapper for stable DistilBERT pipeline."""
    y_val = np.asarray(hf_val["label"]).astype(int)
    return train_transformer_stable(
        hf_train,
        hf_val,
        hf_test,
        y_test,
        y_val,
        cfg,
        output_dir,
        seed=seed,
        model_label="DistilBERT-stable",
    )