File size: 27,315 Bytes
654b110
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# ============================================================================
# RAPID PROTOTYPE: 2-Expert Consensus + Alignment Bank
#
# Fast iteration cycle:
#   Phase 1: Train student on 2-BERT consensus (20K captions, ~2 epochs)
#   Phase 2: Freeze student, train alignment bank on its output
#   Phase 3: Verify bank preserves geometry
#   Phase 4: Snap a tiny classifier on bank output, check stability
# ============================================================================

import gc
import math
import os
import time
import json

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from tqdm import tqdm

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

EXPERTS = [
    ("google-bert/bert-base-uncased", "bert", 512),
    ("answerdotai/ModernBERT-base", "modern", 512),
]

print("=" * 65)
print("RAPID PROTOTYPE: 2-Expert Consensus + Alignment Bank")
print("=" * 65)
print(f"  Device: {DEVICE}")


# ══════════════════════════════════════════════════════════════════
# STUDENT MODEL
# ══════════════════════════════════════════════════════════════════

class MiniStudent(nn.Module):
    def __init__(self, vocab_size=30522, max_len=512, d_model=256,
                 n_heads=4, n_layers=4, d_ff=1024, output_dim=768,
                 dropout=0.1, pad_token_id=0):
        super().__init__()
        self.pad_token_id = pad_token_id
        self.token_emb = nn.Embedding(vocab_size, d_model, padding_idx=pad_token_id)
        self.pos_emb = nn.Embedding(max_len, d_model)
        self.emb_norm = nn.LayerNorm(d_model)
        self.emb_drop = nn.Dropout(dropout)
        encoder_layer = nn.TransformerEncoderLayer(
            d_model=d_model, nhead=n_heads, dim_feedforward=d_ff,
            dropout=dropout, activation="gelu", batch_first=True,
            norm_first=True)
        self.encoder = nn.TransformerEncoder(
            encoder_layer, num_layers=n_layers, enable_nested_tensor=False)
        self.output_proj = nn.Sequential(
            nn.Linear(d_model, d_model), nn.GELU(),
            nn.LayerNorm(d_model), nn.Linear(d_model, output_dim))

    def forward(self, input_ids, attention_mask=None):
        B, L = input_ids.shape
        positions = torch.arange(L, device=input_ids.device).unsqueeze(0)
        x = self.token_emb(input_ids) + self.pos_emb(positions)
        x = self.emb_drop(self.emb_norm(x))
        kpm = ~attention_mask.bool() if attention_mask is not None else (input_ids == self.pad_token_id)
        x = self.encoder(x, src_key_padding_mask=kpm)
        mask = attention_mask.unsqueeze(-1).float() if attention_mask is not None else (~kpm).unsqueeze(-1).float()
        pooled = (x * mask).sum(1) / mask.sum(1).clamp(min=1)
        return F.normalize(self.output_proj(pooled), dim=-1)


# ══════════════════════════════════════════════════════════════════
# ALIGNMENT BANK
# ══════════════════════════════════════════════════════════════════

class AlignmentBank(nn.Module):
    """
    Geometric interface layer. Learns to annotate student embeddings
    with per-expert alignment context and anchor distances.

    Trained on frozen student output. Provides geometric memory of
    the expert consensus for downstream heads.
    """
    def __init__(self, d_embed=768, n_experts=2, n_anchors=128, d_bank=64):
        super().__init__()
        self.d_embed = d_embed
        self.n_experts = n_experts
        self.n_anchors = n_anchors
        self.d_bank = d_bank

        # Per-expert rotation matrices (initialized from Procrustes)
        self.expert_rotations = nn.ParameterList([
            nn.Parameter(torch.eye(d_embed)) for _ in range(n_experts)
        ])

        # Per-expert bias (mean offset in each expert's space)
        self.expert_means = nn.ParameterList([
            nn.Parameter(torch.zeros(d_embed)) for _ in range(n_experts)
        ])

        # Anchor bank: learned consensus landmarks
        self.anchors = nn.Parameter(
            F.normalize(torch.randn(n_anchors, d_embed), dim=-1))

        # Project geometric features into compact context
        # Input: n_experts (consistency) + n_anchors (distances) + n_experts (reconstruction quality)
        geo_dim = n_experts + n_anchors + n_experts
        self.geo_proj = nn.Sequential(
            nn.Linear(geo_dim, d_bank * 2),
            nn.GELU(),
            nn.LayerNorm(d_bank * 2),
            nn.Linear(d_bank * 2, d_bank),
            nn.LayerNorm(d_bank),
        )

    def init_from_procrustes(self, procrustes_results, expert_names,
                              consensus_embeddings=None):
        """Initialize from consensus training artifacts."""
        device = self.anchors.device
        for i, name in enumerate(expert_names[:self.n_experts]):
            info = procrustes_results[name]
            self.expert_rotations[i].data = info["rotation"].float().to(device)
            self.expert_means[i].data = info["source_mean"].float().to(device)
            print(f"    Expert {i} ({name}): rotation loaded, cos_after={info['cos_after']:.4f}")

        if consensus_embeddings is not None:
            n = min(self.n_anchors, consensus_embeddings.shape[0])
            indices = torch.linspace(0, consensus_embeddings.shape[0] - 1, n).long()
            self.anchors.data[:n] = F.normalize(
                consensus_embeddings[indices].float(), dim=-1).to(device)
            print(f"    Anchors: {n} initialized from consensus embeddings")

    def forward(self, embedding):
        """
        Annotate embedding with geometric context.

        Args:
            embedding: (B, 768) L2-normalized

        Returns:
            enriched: (B, 768 + d_bank)
            aux: dict with geometric losses and diagnostics
        """
        B = embedding.shape[0]
        emb = embedding.float()

        # Per-expert: rotate into expert space, measure reconstruction quality
        expert_consistency = []  # cosine between original and round-tripped
        expert_recon = []        # MSE of round-trip
        for i in range(self.n_experts):
            R = self.expert_rotations[i]
            # Forward rotation: consensus β†’ expert space
            in_expert = emb @ R
            # Backward rotation: expert space β†’ consensus
            round_trip = in_expert @ R.T
            # How well does round-trip recover original?
            cos = F.cosine_similarity(emb, round_trip, dim=-1)  # (B,)
            recon = (emb - round_trip).pow(2).mean(dim=-1)      # (B,)
            expert_consistency.append(cos)
            expert_recon.append(recon)

        expert_cos = torch.stack(expert_consistency, dim=-1)    # (B, n_experts)
        expert_mse = torch.stack(expert_recon, dim=-1)          # (B, n_experts)

        # Anchor distances
        anchors_n = F.normalize(self.anchors, dim=-1)
        anchor_cos = emb @ anchors_n.T  # (B, n_anchors)

        # Geometric context vector
        geo_input = torch.cat([expert_cos, anchor_cos, expert_mse], dim=-1)
        geo_context = self.geo_proj(geo_input)  # (B, d_bank)

        # Enriched output
        enriched = torch.cat([embedding, geo_context], dim=-1)

        # ── Geometric losses ──
        aux = {}

        # 1. Expert agreement: all experts should see the embedding similarly
        expert_mean = expert_cos.mean(dim=-1, keepdim=True)
        aux["expert_agreement"] = (expert_cos - expert_mean).pow(2).mean()

        # 2. Rotation orthogonality: rotations should stay orthogonal
        ortho_loss = 0.0
        for i in range(self.n_experts):
            R = self.expert_rotations[i]
            RRT = R @ R.T
            ortho_loss += (RRT - torch.eye(self.d_embed, device=R.device)).pow(2).mean()
        aux["rotation_ortho"] = ortho_loss / self.n_experts

        # 3. Anchor spread: anchors should be well-distributed
        anchor_sim = anchors_n @ anchors_n.T
        anchor_sim.fill_diagonal_(0)
        aux["anchor_spread"] = anchor_sim.pow(2).mean()

        # 4. Anchor sharpness: each embedding should have clear nearest anchors
        anchor_probs = F.softmax(anchor_cos * 10, dim=-1)
        entropy = -(anchor_probs * (anchor_probs + 1e-12).log()).sum(-1).mean()
        aux["anchor_entropy"] = entropy

        # 5. Pentachoron CV of enriched space (sample from geo_context)
        if B >= 10:
            ctx_n = F.normalize(geo_context, dim=-1)
            vols = []
            for _ in range(32):
                idx = torch.randperm(B, device=embedding.device)[:5]
                pts = ctx_n[idx].unsqueeze(0)
                diff = pts.unsqueeze(-2) - pts.unsqueeze(-3)
                d2 = (diff * diff).sum(-1)
                Bv, V, _ = d2.shape
                cm = torch.zeros(Bv, V+1, V+1, device=d2.device, dtype=torch.float32)
                cm[:, 0, 1:] = 1; cm[:, 1:, 0] = 1; cm[:, 1:, 1:] = d2
                s = (-1.0)**V; f = math.factorial(V-1)
                v2 = s / ((2.0**(V-1)) * f*f) * torch.linalg.det(cm)
                vols.append(torch.sqrt(F.relu(v2[0]) + 1e-12))
            stacked = torch.stack(vols)
            bank_cv = stacked.std() / (stacked.mean() + 1e-8)
            aux["bank_cv"] = bank_cv
        else:
            aux["bank_cv"] = torch.tensor(0.0, device=embedding.device)

        # Summary diagnostics
        aux["expert_cos_mean"] = expert_cos.mean().item()
        aux["expert_cos_std"] = expert_cos.std().item()
        aux["anchor_max_cos"] = anchor_cos.max(dim=-1).values.mean().item()
        aux["anchor_mean_cos"] = anchor_cos.mean().item()

        return enriched, aux

    def bank_loss(self, aux, cv_target=0.15):
        """Combined bank training loss."""
        loss = (1.0 * aux["expert_agreement"] +
               1.0 * aux["rotation_ortho"] +
               0.5 * aux["anchor_spread"] +
               0.1 * aux["anchor_entropy"] +
               0.3 * (aux["bank_cv"] - cv_target).abs())
        return loss


# ══════════════════════════════════════════════════════════════════
# GEOMETRY
# ══════════════════════════════════════════════════════════════════

def infonce(a, b, temperature=0.07):
    a = F.normalize(a, dim=-1)
    b = F.normalize(b, dim=-1)
    logits = (a @ b.T) / temperature
    labels = torch.arange(logits.shape[0], device=logits.device)
    loss = (F.cross_entropy(logits, labels) + F.cross_entropy(logits.T, labels)) / 2
    with torch.no_grad():
        acc = (logits.argmax(-1) == labels).float().mean().item()
    return loss, acc

def cayley_menger_vol2(pts):
    pts = pts.float()
    diff = pts.unsqueeze(-2) - pts.unsqueeze(-3)
    d2 = (diff * diff).sum(-1)
    B, V, _ = d2.shape
    cm = torch.zeros(B, V+1, V+1, device=d2.device, dtype=torch.float32)
    cm[:, 0, 1:] = 1; cm[:, 1:, 0] = 1; cm[:, 1:, 1:] = d2
    s = (-1.0)**V; f = math.factorial(V-1)
    return s / ((2.0**(V-1)) * f*f) * torch.linalg.det(cm)

def cv_loss(emb, target=0.12, n_samples=16):
    B = emb.shape[0]
    if B < 5: return torch.tensor(0.0, device=emb.device)
    vols = []
    for _ in range(n_samples):
        idx = torch.randperm(B, device=emb.device)[:5]
        v2 = cayley_menger_vol2(emb[idx].unsqueeze(0))
        vols.append(torch.sqrt(F.relu(v2[0]) + 1e-12))
    stacked = torch.stack(vols)
    cv = stacked.std() / (stacked.mean() + 1e-8)
    return (cv - target).abs()

def cv_metric(emb, n=200):
    B = emb.shape[0]
    if B < 5: return 0.0
    vols = []
    for _ in range(n):
        idx = torch.randperm(B, device=emb.device)[:5]
        v2 = cayley_menger_vol2(emb[idx].unsqueeze(0))
        v = torch.sqrt(F.relu(v2[0]) + 1e-12).item()
        if v > 0: vols.append(v)
    if len(vols) < 10: return 0.0
    a = np.array(vols)
    return float(a.std() / (a.mean() + 1e-8))


# ══════════════════════════════════════════════════════════════════
# EXTRACTION + ALIGNMENT
# ══════════════════════════════════════════════════════════════════

def symmetric_inv_sqrt(cov, eps=1e-6):
    evals, evecs = torch.linalg.eigh(cov)
    evals = torch.clamp(evals, min=eps)
    return evecs @ torch.diag(evals.rsqrt()) @ evecs.T

def procrustes_align(source, target, n_align=5000):
    N = min(n_align, source.shape[0], target.shape[0])
    S = source[:N].float()
    T = target[:N].float()
    s_mean = S.mean(0, keepdim=True)
    t_mean = T.mean(0, keepdim=True)
    Sc = S - s_mean; Tc = T - t_mean
    N_s = Sc.shape[0]
    cos_before = F.cosine_similarity(Sc, Tc, dim=-1).mean().item()
    s_cov = (Sc.T @ Sc) / max(N_s - 1, 1)
    t_cov = (Tc.T @ Tc) / max(N_s - 1, 1)
    s_whiten = symmetric_inv_sqrt(s_cov)
    t_whiten = symmetric_inv_sqrt(t_cov)
    Sc_w = F.normalize(Sc @ s_whiten, dim=-1)
    Tc_w = F.normalize(Tc @ t_whiten, dim=-1)
    U, _, Vt = torch.linalg.svd(Tc_w.T @ Sc_w, full_matrices=False)
    R = U @ Vt
    cos_after = F.cosine_similarity(Sc_w @ R.T, Tc_w, dim=-1).mean().item()
    return {
        "rotation": R, "source_mean": s_mean.squeeze(0),
        "source_whitener": s_whiten,
        "target_unwhitener": torch.linalg.pinv(t_whiten),
        "cos_before": cos_before, "cos_after": cos_after,
    }

def apply_align(emb, a):
    x = emb.float() - a["source_mean"]
    x = x @ a["source_whitener"]
    x = x @ a["rotation"].T
    x = x @ a["target_unwhitener"]
    return x


# ══════════════════════════════════════════════════════════════════
# MAIN
# ══════════════════════════════════════════════════════════════════

def run():
    torch.manual_seed(42)
    np.random.seed(42)
    N_SAMPLES = 20000
    MAX_LEN = 128
    BATCH = 256

    # ── Phase 0: Extract ──
    print(f"\n{'='*65}")
    print("PHASE 0: EXTRACTION")
    print(f"{'='*65}")

    from datasets import load_dataset
    from transformers import AutoModel, AutoTokenizer

    ds = load_dataset("CaptionEmporium/conceptual-captions-cc12m-llavanext",
                      split="train", streaming=True)
    captions = []
    for row in ds:
        cap = row.get("caption_llava", "")
        if isinstance(cap, str) and len(cap) > 50:
            captions.append(cap)
        if len(captions) >= N_SAMPLES:
            break
    print(f"  Captions: {len(captions):,}")

    embeds = {}
    for model_name, short, max_len in EXPERTS:
        print(f"\n  Extracting: {short}...")
        model = AutoModel.from_pretrained(model_name).to(DEVICE).eval()
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        all_emb = []
        with torch.no_grad():
            for i in tqdm(range(0, len(captions), 128), desc=f"    {short}"):
                batch = captions[i:i+128]
                inputs = tokenizer(batch, max_length=max_len, padding=True,
                                  truncation=True, return_tensors="pt").to(DEVICE)
                out = model(**inputs)
                m = inputs.attention_mask.unsqueeze(-1).float()
                pooled = (out.last_hidden_state * m).sum(1) / m.sum(1).clamp(min=1)
                all_emb.append(pooled.cpu())
        embeds[short] = torch.cat(all_emb)
        print(f"    Shape: {embeds[short].shape}")
        del model; gc.collect(); torch.cuda.empty_cache()

    # ── Phase 0b: Align + Consensus ──
    print(f"\n{'='*65}")
    print("PHASE 0b: PROCRUSTES ALIGNMENT")
    print(f"{'='*65}")

    ref = "bert"
    names = [s for _, s, _ in EXPERTS]
    procrustes_results = {}
    aligned = {}
    for name in names:
        info = procrustes_align(embeds[name], embeds[ref])
        procrustes_results[name] = info
        aligned[name] = apply_align(embeds[name], info)
        print(f"  {name:10s}: cos {info['cos_before']:.4f} β†’ {info['cos_after']:.4f}")

    consensus = F.normalize(sum(aligned[n] for n in names) / len(names), dim=-1)
    print(f"  Consensus: {consensus.shape}")
    for name in names:
        cos = F.cosine_similarity(consensus[:2000], aligned[name][:2000], dim=-1).mean().item()
        print(f"  cos(consensus, {name}): {cos:.4f}")

    consensus_cv = cv_metric(consensus[:2000].to(DEVICE))
    print(f"  Consensus CV: {consensus_cv:.4f}")

    del embeds, aligned
    gc.collect(); torch.cuda.empty_cache()

    # ── Phase 1: Train Student ──
    print(f"\n{'='*65}")
    print("PHASE 1: TRAIN STUDENT (2 experts, 20K captions)")
    print(f"{'='*65}")

    tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-uncased")
    tokens = tokenizer(captions, max_length=MAX_LEN, padding="max_length",
                       truncation=True, return_tensors="pt")
    input_ids = tokens["input_ids"]
    attention_mask = tokens["attention_mask"]

    n_train = N_SAMPLES - 2000
    train_ids = input_ids[:n_train].to(DEVICE)
    train_mask = attention_mask[:n_train].to(DEVICE)
    train_targets = consensus[:n_train].to(DEVICE)
    val_ids = input_ids[n_train:].to(DEVICE)
    val_mask = attention_mask[n_train:].to(DEVICE)
    val_targets = consensus[n_train:].to(DEVICE)

    student = MiniStudent(
        vocab_size=tokenizer.vocab_size, max_len=MAX_LEN,
        d_model=256, n_heads=4, n_layers=4, d_ff=1024,
        output_dim=768, dropout=0.1, pad_token_id=tokenizer.pad_token_id
    ).to(DEVICE)
    n_params = sum(p.numel() for p in student.parameters())
    print(f"  Student: {n_params:,} params")

    optimizer = torch.optim.AdamW(student.parameters(), lr=3e-4, weight_decay=0.01)

    for epoch in range(5):
        student.train()
        perm = torch.randperm(n_train, device=DEVICE)
        t_loss, t_acc, t_cos, n = 0, 0, 0, 0
        t0 = time.time()

        for i in range(0, n_train, BATCH):
            idx = perm[i:i+BATCH]
            if len(idx) < 8: continue
            emb = student(train_ids[idx], train_mask[idx])
            tgt = train_targets[idx]
            l_nce, acc = infonce(emb, tgt)
            l_mse = F.mse_loss(emb, tgt)
            l_cv = cv_loss(emb, target=consensus_cv)
            loss = l_nce + l_mse + 0.1 * l_cv
            loss.backward()
            torch.nn.utils.clip_grad_norm_(student.parameters(), 1.0)
            optimizer.step(); optimizer.zero_grad(set_to_none=True)
            with torch.no_grad():
                cos = F.cosine_similarity(emb, tgt, dim=-1).mean().item()
            t_loss += loss.item(); t_acc += acc; t_cos += cos; n += 1

        elapsed = time.time() - t0
        d = max(n, 1)
        student.eval()
        with torch.no_grad():
            v_emb = student(val_ids, val_mask)
            _, v_acc = infonce(v_emb[:1000], val_targets[:1000])
            v_cos = F.cosine_similarity(v_emb, val_targets, dim=-1).mean().item()
            v_cv = cv_metric(v_emb[:1000])

        print(f"  E{epoch+1}: {elapsed:.0f}s  loss={t_loss/d:.4f}  "
              f"t_acc={t_acc/d:.3f}  t_cos={t_cos/d:.3f}  "
              f"v_acc={v_acc:.3f}  v_cos={v_cos:.3f}  v_cv={v_cv:.3f}")

    # Save student
    torch.save(student.state_dict(), "mini_student.pt")
    print(f"\n  Student saved. v_cos={v_cos:.3f}, v_cv={v_cv:.3f}")

    # ── Phase 2: Train Alignment Bank ──
    print(f"\n{'='*65}")
    print("PHASE 2: TRAIN ALIGNMENT BANK (student frozen)")
    print(f"{'='*65}")

    # Freeze student
    student.eval()
    for p in student.parameters():
        p.requires_grad = False

    # Pre-encode everything through frozen student
    print("  Pre-encoding through frozen student...")
    with torch.no_grad():
        all_embs = []
        for i in range(0, n_train, 512):
            j = min(i + 512, n_train)
            emb = student(train_ids[i:j], train_mask[i:j])
            all_embs.append(emb)
        student_embs = torch.cat(all_embs)  # (n_train, 768)
        val_student_embs = student(val_ids, val_mask)

    print(f"  Student embeddings: {student_embs.shape}")

    # Build bank
    bank = AlignmentBank(
        d_embed=768, n_experts=len(EXPERTS),
        n_anchors=128, d_bank=64
    ).to(DEVICE)

    bank.init_from_procrustes(procrustes_results, names, consensus[:n_train])
    bank_params = sum(p.numel() for p in bank.parameters())
    print(f"  Bank: {bank_params:,} params")

    bank_opt = torch.optim.AdamW(bank.parameters(), lr=1e-3, weight_decay=0.01)
    BANK_EPOCHS = 20
    BANK_BATCH = 256

    for epoch in range(BANK_EPOCHS):
        bank.train()
        perm = torch.randperm(n_train, device=DEVICE)
        total_loss = 0
        stats = {"expert_agreement": 0, "rotation_ortho": 0,
                 "anchor_spread": 0, "bank_cv": 0}
        n = 0
        t0 = time.time()

        for i in range(0, n_train, BANK_BATCH):
            idx = perm[i:i+BANK_BATCH]
            if len(idx) < 16: continue

            emb = student_embs[idx]
            enriched, aux = bank(emb)
            loss = bank.bank_loss(aux, cv_target=consensus_cv + 0.02)

            loss.backward()
            torch.nn.utils.clip_grad_norm_(bank.parameters(), 1.0)
            bank_opt.step(); bank_opt.zero_grad(set_to_none=True)

            total_loss += loss.item()
            for k in stats:
                if k in aux:
                    v = aux[k]
                    stats[k] += v.item() if torch.is_tensor(v) else v
            n += 1

        elapsed = time.time() - t0
        d = max(n, 1)

        # Validation
        bank.eval()
        with torch.no_grad():
            v_enriched, v_aux = bank(val_student_embs)
            v_loss = bank.bank_loss(v_aux, cv_target=consensus_cv + 0.02).item()

        print(f"  E{epoch+1:2d}: {elapsed:.0f}s  loss={total_loss/d:.4f}  "
              f"v_loss={v_loss:.4f}  "
              f"expert_agr={stats['expert_agreement']/d:.5f}  "
              f"ortho={stats['rotation_ortho']/d:.5f}  "
              f"spread={stats['anchor_spread']/d:.5f}  "
              f"cv={stats['bank_cv']/d:.4f}  "
              f"anchor_max={v_aux['anchor_max_cos']:.3f}  "
              f"expert_cos={v_aux['expert_cos_mean']:.3f}Β±{v_aux['expert_cos_std']:.3f}")

    torch.save(bank.state_dict(), "alignment_bank.pt")

    # ── Phase 3: Verify Geometry ──
    print(f"\n{'='*65}")
    print("PHASE 3: GEOMETRIC VERIFICATION")
    print(f"{'='*65}")

    bank.eval()
    with torch.no_grad():
        # Check that enriched embeddings preserve original structure
        enriched_val, _ = bank(val_student_embs)
        original_768 = enriched_val[:, :768]  # first 768 dims = original embedding
        geo_context = enriched_val[:, 768:]    # last d_bank dims = geometric annotation

        # Original embedding should be unchanged (passthrough)
        passthrough_cos = F.cosine_similarity(
            original_768[:100], val_student_embs[:100], dim=-1).mean().item()

        # Geometric context should be informative
        geo_cv = cv_metric(F.normalize(geo_context[:1000], dim=-1))
        geo_eff_dim = torch.linalg.svdvals(
            geo_context[:1000].float() - geo_context[:1000].float().mean(0)).pow(2)
        geo_eff_dim = (geo_eff_dim.sum() ** 2) / (geo_eff_dim.pow(2).sum() + 1e-12)

    print(f"  Passthrough integrity: {passthrough_cos:.6f} (should be ~1.000)")
    print(f"  Geo context CV: {geo_cv:.4f}")
    print(f"  Geo context eff_dim: {geo_eff_dim:.1f}")
    print(f"  Geo context shape: {geo_context.shape}")

    # ── Phase 4: Quick Classifier Test ──
    print(f"\n{'='*65}")
    print("PHASE 4: CLASSIFIER STABILITY TEST")
    print(f"{'='*65}")

    # Create synthetic 3-class task from similarity structure
    # Class 0: high consensus cosine pairs (similar)
    # Class 1: medium consensus cosine pairs
    # Class 2: low consensus cosine pairs (different)
    with torch.no_grad():
        # Generate synthetic labels from embedding distances
        embs = val_student_embs[:1000]
        sim = embs @ embs.T
        sim.fill_diagonal_(-1)  # exclude self

        # Random pairs
        n_pairs = 3000
        idx_a = torch.randint(0, 1000, (n_pairs,))
        idx_b = torch.randint(0, 1000, (n_pairs,))
        pair_cos = sim[idx_a, idx_b]

        # Assign labels by cosine terciles
        sorted_cos, _ = pair_cos.sort()
        t1 = sorted_cos[n_pairs // 3].item()
        t2 = sorted_cos[2 * n_pairs // 3].item()
        labels = torch.zeros(n_pairs, dtype=torch.long, device=DEVICE)
        labels[pair_cos > t2] = 0  # similar
        labels[(pair_cos <= t2) & (pair_cos > t1)] = 1  # medium
        labels[pair_cos <= t1] = 2  # different

        # Get enriched representations
        enriched_a, _ = bank(embs[idx_a])
        enriched_b, _ = bank(embs[idx_b])

    # Train tiny classifier: with bank vs without bank
    for mode in ["with_bank", "without_bank"]:
        if mode == "with_bank":
            feat_dim = (768 + 64) * 2  # enriched
            features = torch.cat([enriched_a, enriched_b], dim=-1)
        else:
            feat_dim = 768 * 2  # raw
            features = torch.cat([embs[idx_a], embs[idx_b]], dim=-1)

        clf = nn.Sequential(
            nn.Linear(feat_dim, 128), nn.GELU(),
            nn.Linear(128, 3)
        ).to(DEVICE)

        clf_opt = torch.optim.Adam(clf.parameters(), lr=1e-3)
        n_clf_train = 2400
        train_f = features[:n_clf_train].detach()
        train_l = labels[:n_clf_train]
        val_f = features[n_clf_train:].detach()
        val_l = labels[n_clf_train:]

        for e in range(20):
            clf.train()
            logits = clf(train_f)
            loss = F.cross_entropy(logits, train_l)
            loss.backward()
            clf_opt.step(); clf_opt.zero_grad()

        clf.eval()
        with torch.no_grad():
            val_logits = clf(val_f)
            val_acc = (val_logits.argmax(-1) == val_l).float().mean().item()
            train_logits = clf(train_f)
            train_acc = (train_logits.argmax(-1) == train_l).float().mean().item()

        print(f"  {mode:15s}: train_acc={train_acc:.3f}  val_acc={val_acc:.3f}  "
              f"gap={train_acc-val_acc:.3f}")

    print(f"\n{'='*65}")
    print("DONE")
    print(f"{'='*65}")
    print(f"\n  Student: mini_student.pt")
    print(f"  Bank: alignment_bank.pt")
    print(f"  Consensus CV: {consensus_cv:.4f}")
    print(f"  Student v_cos: {v_cos:.3f}")


if __name__ == "__main__":
    run()