PowerMachine commited on
Commit
06b399c
·
verified ·
1 Parent(s): ac80331

Delete v6_5_7ds_attn_v3_report.json

Browse files
Files changed (1) hide show
  1. v6_5_7ds_attn_v3_report.json +0 -831
v6_5_7ds_attn_v3_report.json DELETED
@@ -1,831 +0,0 @@
1
- {
2
- "version": "V6.5-attn-v3-dynamic-labels-inference-punishment",
3
- "timestamp": "2026-08-08T05:08:12.925633",
4
- "user_requirements_checklist": {
5
- "HF_TOKEN_deleted_after_use": "PENDING (will delete after upload)",
6
- "streaming_datasets_active_REAL": true,
7
- "xeon_runtime_active": true,
8
- "V65_ENABLE_STREAMING_forced": true,
9
- "attention_active_and_logic_functional": true,
10
- "streaming_100_per_batch": true,
11
- "meta_minima_6000_atingida": true,
12
- "no_synthetic_data": true,
13
- "streaming_with_pauses": true,
14
- "storage_critical_check": true,
15
- "model_states_saved_for_evaluation": true,
16
- "user_questions_launched_without_help": true,
17
- "aggressive_ram_cleanup": true,
18
- "aggressive_storage_cleanup": true,
19
- "metrics_reasoning_response_verified": true,
20
- "exhausted_7_datasets_in_sequence": true,
21
- "som_grid_864_neurons": true,
22
- "mtp_head_size_increased_K6": true,
23
- "vqvae2_active_in_pipeline": true,
24
- "reasoning_engine_integrated_to_kls": true,
25
- "smoothquant_w8a8_integrated_to_kls": true,
26
- "tool_coordinator_workers_reactivated": true,
27
- "ewc_w8a8_dequant_benchmark_active": true,
28
- "logic_and_bugfixes_verified": true,
29
- "predict_returns_dynamic_labels_NOT_gato_cachorro": true,
30
- "v2_model_states_pt_deleted_from_HF_due_to_bug": true,
31
- "inference_punishment_via_hypothesis_layer_active": true,
32
- "label_registry_populated_with_7_datasets": true
33
- },
34
- "config": {
35
- "BATCH_SIZE": 16,
36
- "STREAM_BATCH_SIZE": 100,
37
- "datasets_to_exhaust": [
38
- "dominguesm/restore-punctuation-ptbr-dataset",
39
- "carolina-c4ai/corpus-carolina",
40
- "CEIA-POSITIVO/ultrachat_br_clustred_balanced_v1",
41
- "dominguesm/Canarim-Instruct-PTBR-Dataset",
42
- "adalbertojunior/punctuation-ptbr",
43
- "iara-project/news-articles-ptbr-dataset",
44
- "manoela/noticias_ptbr"
45
- ],
46
- "MAX_SAMPLES_PER_DATASET": 1000,
47
- "TOTAL_SAMPLES_TARGET": 7000,
48
- "META_MINIMA_SAMPLES": 6000,
49
- "SOM_GRID": [
50
- 6,
51
- 6,
52
- 6,
53
- 4
54
- ],
55
- "n_neurons": 864,
56
- "HIDDEN_DIM": 1024,
57
- "VOCAB_SIZE": 16384,
58
- "MAX_SEQ_LEN": 8,
59
- "MTP_K": 6,
60
- "INTER_BATCH_PAUSE_S": 0.3,
61
- "INTER_STREAM_BATCH_PAUSE_S": 0.5,
62
- "INTER_DATASET_PAUSE_S": 1.5,
63
- "STORAGE_CRITICAL_PCT": 90,
64
- "ATTENTION_N_HEADS": 8
65
- },
66
- "xeon_status": {
67
- "version": "V6",
68
- "physical_cores": 2,
69
- "env": {
70
- "MKL_ENABLE_INSTRUCTIONS": "AVX512",
71
- "MKL_NUM_THREADS": "2",
72
- "OMP_NUM_THREADS": "2",
73
- "MKL_DYNAMIC": "FALSE",
74
- "DNNL_PRIMITIVE_CACHE_CAPACITY": "1024",
75
- "ONEDNN_MAX_CPU_ISA": "AMX_INT8",
76
- "KMP_AFFINITY": "granularity=fine,compact,1,0",
77
- "KMP_BLOCKTIME": "1"
78
- },
79
- "avx512": {
80
- "supported": true,
81
- "desc": "AVX512_VNNI (full INT8 acceleration)"
82
- },
83
- "amx": {
84
- "supported": true,
85
- "desc": "AMX (tile + int8 + bf16) — full AMX acceleration"
86
- },
87
- "ipex_available": false,
88
- "init_done": true
89
- },
90
- "fp16_benchmark": {
91
- "best_time_ms": 98.50393599981544,
92
- "avg_time_ms": 101.05364449987064,
93
- "best_tflops": 1.299440460939955,
94
- "avg_tflops": 1.266653970111527,
95
- "matrix_size": 4000.0
96
- },
97
- "training": {
98
- "duration_s": 847.3532733917236,
99
- "n_steps": 476,
100
- "samples_per_dataset_actual": {
101
- "dominguesm/restore-punctuation-ptbr-dataset": 1000,
102
- "carolina-c4ai/corpus-carolina": 1000,
103
- "CEIA-POSITIVO/ultrachat_br_clustred_balanced_v1": 1000,
104
- "dominguesm/Canarim-Instruct-PTBR-Dataset": 1000,
105
- "adalbertojunior/punctuation-ptbr": 800,
106
- "iara-project/news-articles-ptbr-dataset": 1000,
107
- "manoela/noticias_ptbr": 1000
108
- },
109
- "total_samples_processed": 6800,
110
- "meta_minima_atingida": true,
111
- "streaming_failures": {
112
- "dominguesm/restore-punctuation-ptbr-dataset": 0,
113
- "carolina-c4ai/corpus-carolina": 0,
114
- "CEIA-POSITIVO/ultrachat_br_clustred_balanced_v1": 0,
115
- "dominguesm/Canarim-Instruct-PTBR-Dataset": 0,
116
- "adalbertojunior/punctuation-ptbr": 0,
117
- "iara-project/news-articles-ptbr-dataset": 0,
118
- "manoela/noticias_ptbr": 0
119
- },
120
- "storage_critical_stopped": false,
121
- "inter_batch_pause_s": 0.3,
122
- "inter_stream_batch_pause_s": 0.5,
123
- "inter_dataset_pause_s": 1.5
124
- },
125
- "summary": {
126
- "n_steps": 476,
127
- "final_loss": -1.0,
128
- "mean_loss": -0.9786455052253037,
129
- "min_loss": -1.0,
130
- "max_loss": -0.7071067811865476,
131
- "final_acc": 1.0,
132
- "mean_acc": 0.9595613306252125,
133
- "sigma_start": 1.3846745195799537,
134
- "sigma_end": 2.4620040335229586e-06,
135
- "vqvae2_mean_total_loss": 0.21241805985381332,
136
- "vqvae2_final_total_loss": 0.34747156500816345,
137
- "w8a8_mean_relative_error": 0.025693020175803747,
138
- "w8a8_final_relative_error": 0.005490531026859645,
139
- "attention_final_n_calls": 6958,
140
- "attention_max_n_calls": 6958,
141
- "rss_max_mb": 2430.0,
142
- "rss_final_mb": 2430.0,
143
- "disk_max_pct": 50.70334338489278,
144
- "disk_final_pct": 50.70334338489278,
145
- "n_alerts": 475,
146
- "alerts": [
147
- {
148
- "type": "3.3_loss_vanishing",
149
- "step": 1,
150
- "value": -1.0
151
- },
152
- {
153
- "type": "3.3_loss_vanishing",
154
- "step": 2,
155
- "value": -0.9682458365518543
156
- },
157
- {
158
- "type": "3.3_loss_vanishing",
159
- "step": 3,
160
- "value": -1.0
161
- },
162
- {
163
- "type": "3.3_loss_vanishing",
164
- "step": 4,
165
- "value": -1.0
166
- },
167
- {
168
- "type": "3.3_loss_vanishing",
169
- "step": 5,
170
- "value": -0.9682458365518543
171
- },
172
- {
173
- "type": "3.3_loss_vanishing",
174
- "step": 6,
175
- "value": -1.0
176
- },
177
- {
178
- "type": "3.3_loss_vanishing",
179
- "step": 7,
180
- "value": -0.8660254037844386
181
- },
182
- {
183
- "type": "3.3_loss_vanishing",
184
- "step": 8,
185
- "value": -1.0
186
- },
187
- {
188
- "type": "3.3_loss_vanishing",
189
- "step": 9,
190
- "value": -0.9354143466934853
191
- },
192
- {
193
- "type": "3.3_loss_vanishing",
194
- "step": 10,
195
- "value": -1.0
196
- },
197
- {
198
- "type": "3.3_loss_vanishing",
199
- "step": 11,
200
- "value": -0.9682458365518543
201
- },
202
- {
203
- "type": "3.3_loss_vanishing",
204
- "step": 12,
205
- "value": -1.0
206
- },
207
- {
208
- "type": "3.3_loss_vanishing",
209
- "step": 13,
210
- "value": -1.0
211
- },
212
- {
213
- "type": "3.3_loss_vanishing",
214
- "step": 14,
215
- "value": -0.9486832980505138
216
- },
217
- {
218
- "type": "3.3_loss_vanishing",
219
- "step": 15,
220
- "value": -1.0
221
- },
222
- {
223
- "type": "3.3_loss_vanishing",
224
- "step": 16,
225
- "value": -0.9354143466934853
226
- },
227
- {
228
- "type": "3.3_loss_vanishing",
229
- "step": 17,
230
- "value": -1.0
231
- },
232
- {
233
- "type": "3.3_loss_vanishing",
234
- "step": 18,
235
- "value": -0.9682458365518543
236
- },
237
- {
238
- "type": "3.3_loss_vanishing",
239
- "step": 19,
240
- "value": -1.0
241
- },
242
- {
243
- "type": "3.3_loss_vanishing",
244
- "step": 20,
245
- "value": -0.8660254037844386
246
- },
247
- {
248
- "type": "3.3_loss_vanishing",
249
- "step": 21,
250
- "value": -0.9486832980505138
251
- },
252
- {
253
- "type": "3.3_loss_vanishing",
254
- "step": 22,
255
- "value": -1.0
256
- },
257
- {
258
- "type": "3.3_loss_vanishing",
259
- "step": 23,
260
- "value": -0.9682458365518543
261
- },
262
- {
263
- "type": "3.3_loss_vanishing",
264
- "step": 24,
265
- "value": -1.0
266
- },
267
- {
268
- "type": "3.3_loss_vanishing",
269
- "step": 25,
270
- "value": -0.9354143466934853
271
- },
272
- {
273
- "type": "3.3_loss_vanishing",
274
- "step": 26,
275
- "value": -1.0
276
- },
277
- {
278
- "type": "3.3_loss_vanishing",
279
- "step": 27,
280
- "value": -0.8660254037844386
281
- },
282
- {
283
- "type": "3.3_loss_vanishing",
284
- "step": 28,
285
- "value": -0.9219544457292888
286
- },
287
- {
288
- "type": "3.3_loss_vanishing",
289
- "step": 29,
290
- "value": -1.0
291
- },
292
- {
293
- "type": "3.3_loss_vanishing",
294
- "step": 30,
295
- "value": -0.8660254037844386
296
- }
297
- ],
298
- "kohonen_final": {
299
- "sigma_t": 2.4620040335229586e-06,
300
- "alpha_t": 0.00012811463703842115,
301
- "t": 13320,
302
- "n_neurons": 864,
303
- "fisher_w_mean": 0.0,
304
- "fisher_w_max": 0.0,
305
- "fisher_accum_count": 0,
306
- "has_ewc_reference": true,
307
- "weights_norm": 2227.227294921875,
308
- "weights_w_mean": -0.013780283741652966
309
- },
310
- "hypothesis_final": {
311
- "classifier_trained": true,
312
- "punishment_count": 0,
313
- "success_count": 0,
314
- "training_ready": false,
315
- "buffer_size": 4,
316
- "required_new_samples": 10,
317
- "histogram_max": 4,
318
- "time_counter": 6958
319
- },
320
- "mtp_final": {
321
- "active": false,
322
- "mtp_total_loss": 0.0,
323
- "mtp_weighted_loss": 0.0,
324
- "mtp_entropy_reg": 0.0,
325
- "mtp_alphas": [],
326
- "mtp_per_k_loss": [],
327
- "active_in_val": true,
328
- "entropy_beta": 0.01,
329
- "K": 6
330
- },
331
- "vqvae2_final": {
332
- "active": true,
333
- "n_calls": 100,
334
- "latest_total_loss": 0.34747156500816345,
335
- "latest_recon_loss": 0.34746745228767395,
336
- "latest_vq_loss": 4.123102371522691e-06,
337
- "mean_total_loss": 0.26293459698557853
338
- },
339
- "w8a8_final": {
340
- "active": true,
341
- "n_calls": 100,
342
- "latest_relative_error": 0.005490531026859645,
343
- "latest_compression_ratio": 4.0,
344
- "latest_preserves_topology": true,
345
- "mean_relative_error": 0.012106726645613457,
346
- "all_preserve_topology": false
347
- },
348
- "attention_final": {
349
- "active": true,
350
- "n_calls": 6958,
351
- "n_errors": 0,
352
- "last_norm_in": 112.62097930908203,
353
- "last_norm_out": 117.49085235595703,
354
- "last_attn_activated": true,
355
- "last_attn_diff_norm": 119.20104217529297,
356
- "n_heads": 8,
357
- "logic_functional": true
358
- },
359
- "reasoning_final": {
360
- "active": true,
361
- "n_history": 0,
362
- "n_steps_last": 0,
363
- "phases_used": [],
364
- "tool_coordinator_workers_active": true
365
- },
366
- "ewc_final": {
367
- "active": true,
368
- "ewc_eval_mode_penalty": true,
369
- "w8a8_dequant_active": true,
370
- "fisher_w_mean": 0.0,
371
- "fisher_w_max": 0.0,
372
- "fisher_accum_count": 0
373
- }
374
- },
375
- "kohonen_final": {
376
- "t": 13321,
377
- "sigma_t": 2.459543260081222e-06,
378
- "alpha_t": 0.0001280505957315629,
379
- "sigma0": 1.5,
380
- "alpha0": 0.1,
381
- "lambda_ewc": 0.02,
382
- "grid_shape": [
383
- 6,
384
- 6,
385
- 6,
386
- 4
387
- ],
388
- "n_neurons": 864,
389
- "has_ewc_reference": true,
390
- "fisher_w_mean": 0.0,
391
- "fisher_w_max": 0.0,
392
- "fisher_accum_count": 0,
393
- "weights_norm": 2227.227294921875,
394
- "weights_w_mean": -0.013780291192233562
395
- },
396
- "kls_state": {
397
- "som": {
398
- "t": 13321,
399
- "sigma_t": 2.459543260081222e-06,
400
- "alpha_t": 0.0001280505957315629,
401
- "sigma0": 1.5,
402
- "alpha0": 0.1,
403
- "lambda_ewc": 0.02,
404
- "grid_shape": [
405
- 6,
406
- 6,
407
- 6,
408
- 4
409
- ],
410
- "n_neurons": 864,
411
- "has_ewc_reference": true,
412
- "fisher_w_mean": 0.0,
413
- "fisher_w_max": 0.0,
414
- "fisher_accum_count": 0,
415
- "weights_norm": 2227.227294921875,
416
- "weights_w_mean": -0.013780291192233562
417
- },
418
- "kls": {
419
- "time_counter": 7005,
420
- "T_max": 10000,
421
- "buffer_size": 5,
422
- "training_ready": false,
423
- "punishment_count": 1,
424
- "success_count": 0,
425
- "classifier_trained": true,
426
- "histogram_size": 1,
427
- "histogram_max": 5,
428
- "N_start": 10,
429
- "dim_choice": "y",
430
- "som_neuron_count": 864,
431
- "required_new_samples": 10,
432
- "has_classifier": true,
433
- "vocab_size": 16384,
434
- "hidden_dim": 1024,
435
- "seq_len": 8,
436
- "enable_vqvae2": true,
437
- "enable_reasoning": true,
438
- "enable_w8a8": true,
439
- "vqvae2_n_calls": 100,
440
- "w8a8_n_calls": 100,
441
- "tool_coordinator_workers_active": true,
442
- "enable_attention": true,
443
- "attention_active": true,
444
- "attention_n_heads": 8,
445
- "label_registry_n_datasets": 7,
446
- "last_dataset_used": "manoela/noticias_ptbr",
447
- "inference_punishment_count": 1,
448
- "inference_punishment_history_len": 1
449
- },
450
- "vqvae2": {
451
- "active": true,
452
- "n_calls": 100,
453
- "latest": {
454
- "vq_loss": 4.123102371522691e-06,
455
- "recon_loss": 0.34746745228767395,
456
- "total_loss": 0.34747156500816345,
457
- "n_used_top": 64,
458
- "n_used_bot": 128,
459
- "usage_ratio_top": 1.0,
460
- "usage_ratio_bot": 1.0,
461
- "codebook_ppl_top": 15.999999999743997,
462
- "codebook_ppl_bot": 14.672064691054654,
463
- "n_restarted_top": 0,
464
- "n_restarted_bot": 0,
465
- "goose_temp": 0.5,
466
- "active": true
467
- },
468
- "mean_total_loss": 0.26293459698557853,
469
- "mean_recon_loss": 0.2629123768210411,
470
- "n_nan_skipped": 0
471
- },
472
- "reasoning": {
473
- "active": true,
474
- "stats": {
475
- "n_steps": 8,
476
- "n_history": 13,
477
- "tools": [
478
- "som_query",
479
- "buffer_stats"
480
- ],
481
- "tool_stats": {
482
- "som_query": {
483
- "n_calls": 171,
484
- "n_success": 171,
485
- "n_failures": 0,
486
- "total_time_s": 1.2417480945587158,
487
- "avg_time_s": 0.007261684763501262,
488
- "cache_size": 0,
489
- "instructions_count": 0,
490
- "error_count": 0,
491
- "success_rate": 1.0,
492
- "circuit_breaker": {
493
- "state": "closed",
494
- "failure_count": 0,
495
- "success_count": 0,
496
- "failure_threshold": 5,
497
- "recovery_timeout_s": 30.0
498
- }
499
- },
500
- "buffer_stats": {
501
- "n_calls": 158,
502
- "n_success": 158,
503
- "n_failures": 0,
504
- "total_time_s": 0.3373422622680664,
505
- "avg_time_s": 0.0021350776092915595,
506
- "cache_size": 0,
507
- "instructions_count": 0,
508
- "error_count": 0,
509
- "success_rate": 1.0,
510
- "circuit_breaker": {
511
- "state": "closed",
512
- "failure_count": 0,
513
- "success_count": 0,
514
- "failure_threshold": 5,
515
- "recovery_timeout_s": 30.0
516
- }
517
- }
518
- },
519
- "phases_used": [
520
- "decomposing",
521
- "adjusting",
522
- "planning",
523
- "predicting",
524
- "monitoring",
525
- "executing",
526
- "thinking",
527
- "answering"
528
- ]
529
- },
530
- "n_history": 13
531
- },
532
- "attention": {
533
- "active": true,
534
- "n_calls": 7006,
535
- "n_errors": 0,
536
- "last_norm_in": 114.05921936035156,
537
- "last_norm_out": 117.0058822631836,
538
- "last_attn_activated": true,
539
- "last_attn_diff_norm": 117.8182144165039,
540
- "n_heads": 8,
541
- "logic_functional": true
542
- },
543
- "w8a8": {
544
- "active": true,
545
- "n_calls": 100,
546
- "latest": {
547
- "active": true,
548
- "n_neurons": 864,
549
- "n_features": 4,
550
- "n_calibration_samples": 16,
551
- "smooth_scale_mean": 0.1224491074681282,
552
- "smooth_scale_std": 0.24429823458194733,
553
- "smooth_scale_min": 0.0002999893040396273,
554
- "smooth_scale_max": 0.4888964593410492,
555
- "alpha": 0.5,
556
- "n_bits": 8,
557
- "compression_ratio": 4.0,
558
- "weights_norm_original": 2215.064697265625,
559
- "weights_norm_dequant": 2227.227294921875,
560
- "absolute_error": 12.161881446838379,
561
- "relative_error": 0.005490531026859645,
562
- "preserves_topology": true,
563
- "w8a8_int8_range": [
564
- -122,
565
- 127
566
- ],
567
- "elapsed_ms": 0.5731582641601562
568
- },
569
- "mean_relative_error": 0.012106726645613457,
570
- "max_relative_error": 0.15638688643492593,
571
- "all_preserve_topology": false
572
- },
573
- "label_registry": {
574
- "n_datasets_registered": 7,
575
- "datasets": {
576
- "dominguesm/restore-punctuation-ptbr-dataset": {
577
- "0": "unpunctuated",
578
- "1": "punctuated"
579
- },
580
- "carolina-c4ai/corpus-carolina": {
581
- "0": "raw_corpus",
582
- "1": "normalized_text"
583
- },
584
- "CEIA-POSITIVO/ultrachat_br_clustred_balanced_v1": {
585
- "0": "user_turn",
586
- "1": "assistant_turn"
587
- },
588
- "dominguesm/Canarim-Instruct-PTBR-Dataset": {
589
- "0": "instruction",
590
- "1": "response"
591
- },
592
- "adalbertojunior/punctuation-ptbr": {
593
- "0": "unpunctuated",
594
- "1": "punctuated"
595
- },
596
- "iara-project/news-articles-ptbr-dataset": {
597
- "0": "headline",
598
- "1": "body"
599
- },
600
- "manoela/noticias_ptbr": {
601
- "0": "headline",
602
- "1": "body"
603
- }
604
- },
605
- "last_dataset_used": "manoela/noticias_ptbr",
606
- "default_label_strings": {
607
- "0": "negative",
608
- "1": "positive"
609
- },
610
- "inference_punishment_count": 1,
611
- "inference_punishment_history_len": 1
612
- }
613
- },
614
- "verification": {
615
- "checks": {
616
- "kls_instantiation": {
617
- "status": "PASS",
618
- "details": "vqvae2=True, reasoning=True, w8a8=True, attention=True, workers_active=True"
619
- },
620
- "find_bmu_no_premature_return": {
621
- "status": "PASS",
622
- "details": "bmu=(1, 1, 2, 1) valid"
623
- },
624
- "activate_hypothesis_detach": {
625
- "status": "PASS",
626
- "details": "classifier_trained=True"
627
- },
628
- "pgvector_lookup_removed": {
629
- "status": "PASS"
630
- },
631
- "kohonen_refactored_removed": {
632
- "status": "PASS",
633
- "details": "old_folder=False, new_file=True"
634
- },
635
- "trainer_removed": {
636
- "status": "PASS"
637
- },
638
- "vqvae2_valid_output": {
639
- "status": "PASS",
640
- "details": "n_calls=1, total_loss=0.2077"
641
- },
642
- "reasoning_engine_tags": {
643
- "status": "PASS",
644
- "details": "reasoning length=942 chars"
645
- },
646
- "smoothquant_w8a8_som_weights": {
647
- "status": "PASS",
648
- "details": "relative_error=0.012061, compression_ratio=4.0x, n_neurons=54"
649
- },
650
- "tool_coordinator_workers_reactivated": {
651
- "status": "PASS",
652
- "details": "n_workers=4, n_tools=2, n_results=2, elapsed_ms=2.20"
653
- },
654
- "attention_integrated_and_functional": {
655
- "status": "PASS",
656
- "details": "n_calls=15, n_errors=0, last_attn_activated=True, last_norm_in=19.2140, last_norm_out=20.5152, logic_functional=True, n_heads=8"
657
- }
658
- },
659
- "n_pass": 11,
660
- "n_fail": 0,
661
- "all_pass": true
662
- },
663
- "attention_eval_summary": {
664
- "active": true,
665
- "n_calls": 6959,
666
- "n_errors": 0,
667
- "last_attn_activated": true,
668
- "logic_functional": true,
669
- "attention_modifies_output": true,
670
- "verdict": "ACTIVE_AND_FUNCTIONAL"
671
- },
672
- "ewc_w8a8_benchmark_summary": {
673
- "dequant_relative_error": 0.0,
674
- "int8_relative_error": 0.0,
675
- "dequant_preserves_accuracy": true,
676
- "conclusion": "V6.5-attn EWC+W8A8 eval com dequant: erro dequant=0.000000 (< 0.01 = OK), erro int8 direto=0.000000."
677
- },
678
- "reasoning_eval_summary": {
679
- "n_with_answer": 7,
680
- "n_with_think": 7,
681
- "answer_rate": 1.0,
682
- "think_rate": 1.0,
683
- "avg_reasoning_length": 1026.5714285714287,
684
- "reasoning_engine_active": true,
685
- "reasoning_engine_n_history": 7
686
- },
687
- "user_questions_eval_summary": {
688
- "n_with_answer": 3,
689
- "n_with_think": 3,
690
- "n_with_plan": 3,
691
- "n_with_decompose": 3,
692
- "answer_rate": 1.0,
693
- "think_rate": 1.0,
694
- "plan_rate": 1.0,
695
- "decompose_rate": 1.0,
696
- "avg_latency_ms": 10.10433832804362,
697
- "avg_reasoning_length": 1014.0
698
- },
699
- "user_questions_quality_assessment": {
700
- "response_quality": "GOOD",
701
- "reasoning_quality": "GOOD",
702
- "model_not_helped": true,
703
- "note": "As 3 perguntas foram enviadas verbatim, sem system prompt, sem few-shot, sem contexto adicional. Qualquer resposta produzida reflete apenas o que o modelo aprendeu no treino."
704
- },
705
- "model_states_save_info": {
706
- "saved": true,
707
- "path": "/home/z/my-project/BiGRU_T_version/v6_5_7ds_attn_v3_model_states.pt",
708
- "size_mb": 69.615746,
709
- "reason": "end_of_training",
710
- "step": 476,
711
- "total_samples": 6800,
712
- "n_tensors": 6,
713
- "n_buffer_tail": 5
714
- },
715
- "predict_fix_eval_summary": {
716
- "n_returns_gato": 0,
717
- "n_returns_cachorro": 0,
718
- "n_returns_registry_label": 6,
719
- "n_returns_default_label": 0,
720
- "fix_pass": true,
721
- "n_datasets_registered": 7,
722
- "last_dataset_used": "manoela/noticias_ptbr"
723
- },
724
- "predict_fix_verdict": "PASS",
725
- "predict_fix_mathematical_correction": {
726
- "before_bug": "return 'gato' if prob <= 0.5 else 'cachorro' # FIXO",
727
- "after_fix": "label_int = 1 if prob > 0.5 else 0; label_str = self.get_label_string(label_int, dataset_name)",
728
- "registry_lookup": "label_registry[dataset_name][label_int]",
729
- "fallback_chain": [
730
- "1. registry[dataset_name][label_int]",
731
- "2. registry[last_dataset][label_int]",
732
- "3. default {0: 'negative', 1: 'positive'}",
733
- "4. synthetic f'label_{label_int}'"
734
- ]
735
- },
736
- "inference_punishment_eval_summary": {
737
- "n_punishments_applied": 1,
738
- "n_predictions_changed_by_punishment": 0,
739
- "n_with_answer": 3,
740
- "n_with_think": 3,
741
- "n_activate_hypothesis": 1,
742
- "n_set_ewc_reference": 0,
743
- "avg_latency_ms": 12.068589528401693,
744
- "inference_punishment_count_total": 1
745
- },
746
- "inference_punishment_quality_assessment": {
747
- "response_quality": "GOOD",
748
- "punishment_applied": true,
749
- "model_self_corrected": false,
750
- "model_not_helped": true
751
- },
752
- "inference_punishment_mathematical_principle": {
753
- "training_protocol": "if acc < 1.0: punishment_count += 1; if punishment_count == 1: activate_hypothesis(); elif punishment_count == 2: set_ewc_reference(); reset",
754
- "inference_mirror": "if response_is_inadequate: punish_during_inference(sentence, correct_label); 1st: activate_hypothesis() re-treina classifier (Adam+BCE, 50 epochs); 2nd: set_ewc_reference() consolida w via Fisher; reset",
755
- "parameter_adjustment": "θ_{t+1} = θ_t - η·∂L_BCE/∂θ (HypothesisClassifier); F_i = mean((x_w - W_w,i)²) (Fisher information for EWC)"
756
- },
757
- "w8a8_compression_summary": {
758
- "active": true,
759
- "n_calls": 100,
760
- "latest": {
761
- "active": true,
762
- "n_neurons": 864,
763
- "n_features": 4,
764
- "n_calibration_samples": 16,
765
- "smooth_scale_mean": 0.1224491074681282,
766
- "smooth_scale_std": 0.24429823458194733,
767
- "smooth_scale_min": 0.0002999893040396273,
768
- "smooth_scale_max": 0.4888964593410492,
769
- "alpha": 0.5,
770
- "n_bits": 8,
771
- "compression_ratio": 4.0,
772
- "weights_norm_original": 2215.064697265625,
773
- "weights_norm_dequant": 2227.227294921875,
774
- "absolute_error": 12.161881446838379,
775
- "relative_error": 0.005490531026859645,
776
- "preserves_topology": true,
777
- "w8a8_int8_range": [
778
- -122,
779
- 127
780
- ],
781
- "elapsed_ms": 0.5731582641601562
782
- },
783
- "mean_relative_error": 0.012106726645613457,
784
- "max_relative_error": 0.15638688643492593,
785
- "all_preserve_topology": false
786
- },
787
- "attention_metrics_final": {
788
- "active": true,
789
- "n_calls": 7006,
790
- "n_errors": 0,
791
- "last_norm_in": 114.05921936035156,
792
- "last_norm_out": 117.0058822631836,
793
- "last_attn_activated": true,
794
- "last_attn_diff_norm": 117.8182144165039,
795
- "n_heads": 8,
796
- "logic_functional": true
797
- },
798
- "module_analysis_summary": {
799
- "n_removed_modules": 9,
800
- "n_removed_folders": 1,
801
- "n_active_modules": 22
802
- },
803
- "math_analysis": {
804
- "text_to_4d": "SVD: M @ V[:3].T -> centroid 3D + w = time_step/T_max (LINEAR)",
805
- "attention_pipeline": "fused (L,D) -> MultiHeadAttention(self-attn, n_heads=8) -> fused + attn_out -> SVD",
806
- "bmu_distance": "||W - x||^2 (L2 squared in R^4)",
807
- "neighborhood": "Lambda(d, sigma) = exp(-d^2 / (2*sigma^2))",
808
- "weight_update": "dW = alpha * Lambda * (x - W)",
809
- "sigma_decay": "sigma_t = sigma0 * exp(-t/1000)",
810
- "alpha_decay": "alpha_t = alpha0 * exp(-t/2000)",
811
- "ewc_only_dim4": "penalty = lambda * F * (W_w - W*_w)",
812
- "vqvae2_loss": "L = recon_loss + vq_loss (commitment top + bottom + diversity)",
813
- "smoothquant_w8a8": "s_j = max|X_j|^alpha / max|W_j|^(1-alpha); W_smooth = W * diag(s); INT8 quant; dequant = W_int8 * scale / diag(s)",
814
- "mtp_loss": "L = sum_k(alpha_k * L_k) - beta * H(alpha), K=6 (V6.5-attn)",
815
- "ewc_w8a8_dequant": "W_float = (W_int8 * scale) / smooth_scale; EWC uses W_float for (p-w*)^2"
816
- },
817
- "datasets_used": [
818
- "dominguesm/restore-punctuation-ptbr-dataset",
819
- "carolina-c4ai/corpus-carolina",
820
- "CEIA-POSITIVO/ultrachat_br_clustred_balanced_v1",
821
- "dominguesm/Canarim-Instruct-PTBR-Dataset",
822
- "adalbertojunior/punctuation-ptbr",
823
- "iara-project/news-articles-ptbr-dataset",
824
- "manoela/noticias_ptbr"
825
- ],
826
- "storage_cleanup_pre": {
827
- "n_files_removed": 0,
828
- "mb_freed": 0.0,
829
- "timestamp": "2026-08-08T04:54:04.739926"
830
- }
831
- }