Aug 10 00:33 V6.7: upload src/bigru_t/utils/oom_guard.py (21.1KB) — FASE1+FASE2 training results 00:33 V6.7: upload src/bigru_t/data/streaming_datasets.py (37.9KB) — FASE1+FASE2 training results Aug 9 23:32 V6.7: upload src/bigru_t/model/kohonen_learning_system.py (BBPE serial mode + OomGuard V7) 23:32 V6.7: upload src/bigru_t/tokenizer/bbpe_tokenizer.py (BBPE serial mode + OomGuard V7) 21:43 V6.5-V2-dynamic: upload batch (scripts + state + model) [80 files] 20:30 V7 FINAL: FASE1+FASE2 completed (10000 samples, 20 punishments) V7 fixes:
- AdvancedSOMAugmenter.fit(): torch.no_grad() wrapper (OOM fix + .item() fix)
- V7Fase2Integrator.post_chunk_hook: defensive int/float, traceback logging
- V7Fase2Integrator._compute_adaptive_dpo: kls.neuron_label fix (was kls.som.neuron_labels)
- SimpleBBPETokenizer: pure alias to BBPETokenizer (no duplicate impl)
- train_v6_5_v2.py: SKIP_FASE1 support for FASE2 resume
Results:
- FASE1: 8000 samples (8 datasets, Kohonen only)
- FASE2: 2000 samples (BrunoN-Dev/corpus-ptbr-v1, punishment active)
- V7 augmenter: 10 fit() calls, 0 errors
- V7 DPO: 20 triggers, 0 errors
- SOM: dead 92.6%->56%, TE 0.87->0.39, VE 0.98
Cleanup:
- Deleted 1 misplaced/old files from HF root
- Uploaded 18 files to correct paths 19:08 V7: dpo.py + data_augmentation.py + oom_guard.py aprimorados Módulos aprimorados:
- training/dpo.py: AdaptiveDPOLoss (nn.Module, log_beta otimizável, atualização dual KL-aware, clamps [0.01, 2.0])
- training/data_augmentation.py: AdvancedSOMAugmenter (manifold-aware, grid 4D canônico (4,4,4,4)=256, α₀=0.5, σ₀=2.0, ruído gaussiano adaptativo por QE local, interpolação convexa 0.7*real + 0.3*BMU)
- utils/oom_guard.py: OomGuard V7 (memory leak detection, logic failure detection, crash dump JSON, signal handlers, excepthook, ring buffer de ops, cgroup limit detection)
Arquivos removidos (fora de src/bigru_t/):
- 5 JSONs antigos na raiz: v6_5_v2_*.json (substituídos por reports/)
Parâmetros canônicos mantidos:
- HIDDEN_DIM=1024, VOCAB_SIZE=16384, N_HYPOTHESES=16, MAX_N_HYPOTHESES=32, HYP_TRAIN_STEPS=30, HYP_HIDDEN_DIM=256
- Grid SOM (4,4,4,4)=256, Buffer=256 (alinhamento 1:1)
- α₀=0.5 (∈ [0.5, 1.0] Kohonen rough training)
- σ₀=2.0 (= max(4,4,4,4)/2 = metade da maior dimensão)
Timestamp: 2026-08-09T19:08:19.776243 18:18 V6.5-V4-canonical-256: buffer=256 + grid=(4,4,4,4)=256 + BBPE serial fix Mudanças principais:
- Buffer canônico 256 (alinhado ao grid 256, elimina OOM de V6.5-V3)
- BBPE serial in-process para corpus < 200 textos (evita OOM por ProcessPoolExecutor)
- FASE1: 7000 samples processados (7/8 datasets)
- FASE2: 2000 samples com punição ativa (meta atingida)
- 23 punishment events, 11 hypotheses trainings, 11 delta applications
- Peak RSS: 1933MB (dentro do limite 4GB cgroup)
- Paths remapeados: TODOS módulos em src/bigru_t/**
- Sem duplicidade: deprecated/ NÃO enviados
Timestamp: 2026-08-09T18:18:20.529072 17:49 V6.5-V3-hyp-synergy: SynergyHypothesisEnsemble (4 pilares: diversidade k-means++, cooperação, competição, ortogonalidade) + testes de validação (7 files) 17:22 V6.5-V2-dynamic: upload batch (scripts + state + model) [79 files] 16:50 V6.5-V2-dynamic: upload batch (scripts + state + model) [79 files] 15:50 V6.5-V2-parallel-reorganize: ParallelSOM4D + ReorganizingSOM4D + OOM-killer guards + unified state (75 files) 14:25 V6.5-V2-dynamic: upload batch (scripts + state + model) [79 files] 04:27 V6.5-V2-auto-conscience: upload batch (modules + scripts) [74 files] — auto-adjust conscience_gamma + reasoning_engine disabled by default + HF_TOKEN scrubbed 03:41 V6.5-V2-auto-conscience: upload batch (modules + scripts) [74 files] — auto-adjust conscience_gamma + reasoning_engine disabled by default + HF_TOKEN scrubbed 03:29 V6.5-V2-auto-conscience: upload batch (modules + scripts) [73 files] — auto-adjust conscience_gamma + reasoning_engine disabled by default + HF_TOKEN scrubbed 01:07 V6.5-V2-metrics-FIX-2: fix HF organization — remove 9 root-level duplicates, upload 13 files to correct src/bigru_t/ paths 00:15 V6.5-V2-metrics-FIX: cleanup batch — remove .pt files + duplicate root-level modules + obsolete modules (17 files) Aug 8 17:28 Delete src/bigru_t/training/elastic_weight_consolidation.py 17:20 Delete src/bigru_t/model/kohonen_refactored 17:20 Delete src/bigru_t/model/kohonen_refactored/kohonen_learning_system.py 17:15 Delete src/bigru_t/model/unified_model.py 17:15 Delete src/bigru_t/model/u8cell_t.py 17:15 Delete src/bigru_t/model/transformer_unit.py 17:14 Delete src/bigru_t/model/orq_cell.py 17:09 Delete src/bigru_t/model/kohonen_learning_system.py.bak_v65fix 17:07 Delete src/bigru_t/model/gru_hierarchy.py 17:06 Delete src/bigru_t/model/bigru4.py 05:09 V6.5-attn-v3: CORREÇÃO MATEMÁTICA do predict() — rótulos dinâmicos extraídos do label_registry (substitui 'gato'/'cachorro' fixos). NOVO: punish_during_inference() — mirror do protocolo de punição do treinamento (1ª → activate_hypothesis, 2ª → set_ewc_reference). META ≥6000 samples, attention ACTIVE_AND_FUNCTIONAL, SmoothQuant W8A8, 864 neurons, MTP K=6, 3 perguntas reais sem ajuda (Luva de Pedreiro Távila, Lula reserva valor, Amazonas força-tarefa vítimas) + punição na inferência via camada de hipótese. Removed legacy v2 reports/scripts (superseded by v3). 02:09 V6.5-attn: attention integrated to KLS (MultiHeadAttention n_heads=8), 100 samples/batch streaming, 5040 samples REAL (meta ≥5000), 864 neurons, SmoothQuant W8A8 (err=0.020), MTP K=6, tool_coordinator workers reactivated, 11/11 verification PASS, attention ACTIVE_AND_FUNCTIONAL, reasoning GOOD (100% answer/think), removed legacy v3/v4/v5/v6_1/v6_2/v6_3/v6_progressive + old v6_5_* + v6_5_7ds_* reports/scripts 00:55 V6.5-7ds cleanup: remove legacy src/bigru_t/data/data_augmentation.py 00:54 V6.5-7ds: 7 datasets streaming (1400 samples, 112 steps), 864 neurons, SmoothQuant W8A8 (err=0.022), MTP K=6, tool_coordinator workers reactivated, 10/10 verification PASS, reasoning GOOD, removed legacy v3/v4/v5/v6_1/v6_2/v6_3/v6_progressive files + old v6_5_* reports/scripts 00:02 V6.5-final: removed pre-V6.4 modules/scripts, moved kohonen_learning_system up, activated VQ-VAE-2 in pipeline, integrated reasoning_engine, benchmarked EWC+W8A8 dequant, exhausted 6 datasets Aug 7 22:58 V6.5: 1000 samples (500+500), MTP val+entropy, EWC+W8A8 eval investigation, integrate gru-ring modules 21:51 V6.4: refactored Kohonen SOM 4D (canonical user code, pgvector_lookup REMOVED) 21:17 V6.3: refactored Kohonen Learning System (user code) + xeon_runtime + streaming + bug fix find_bmu 00:55 V6.2: HypT pgvector_lookup before punishment + BATCH_SIZE=16 + 5 datasets x 50 V6.2 improvements (vs V6.1):
- hyp_t.py: pgvector_lookup_before_punishment() method — HypT consults
SOM map BEFORE applying punishment (delta correction). If status ==
COMPLETE, punishment is SKIPPED (model trusts y_hat). If INCOMPLETE/
MISSING, punishment is applied normally.
(User requirement: "ative pgvector_lookup no nível de hipótese (hyp_t.py)
para que ele também consulte o mapa antes de aplicar punição")
- hyp_t.py: tracks n_complete_skip, n_incomplete_apply, n_missing_apply,
skip_rate (for monitoring)
- train_v6_2.py: BATCH_SIZE=16 (was 4) — user requirement
- train_v6_2.py: 5 datasets from streaming_datasets.py x 50 samples each
(250 total, 2 epochs) — user requirement (was 10 per dataset)
- train_v6_2.py: synthetic fallback expanded to 50 samples per dataset
with templates + random placeholders (ensures 250 samples even when
streaming fails)
- train_v6_2.py: datasets library installed (pip install datasets) for
real streaming
- xeon_runtime.py V6 always a… 00:05 V6.1: PGVector-like lookup + EWC math upgrade + 5 datasets x 10 samples V6.1 improvements:
- orq_cell.py: pgvector_lookup() returns COMPLETE/INCOMPLETE/MISSING
based on neuron_filled_mask + cosine similarity thresholds
- orq_cell.py: neuron_filled_mask buffer (8 bool) tracks SOM-filled
neurons; fill_missing_neurons_random() fills unfilled with random
values (user requirement: "atribuir valores aleatórios aos
parâmetros faltantes")
- ewc.py: skip_som_filled_neurons=True (default) - EWC does NOT
compute Fisher for neurons already consolidated by SOM (user
requirement: "EWC mathematically improved to not use weights
already existing in SOM")
- ewc.py: classic Fisher for orq_cell.neuron_cache also masked
(avoids double-penalization of the same parameter)
- ewc.py: short-circuit when all SOM neurons filled (topology
already preserved by SOM Gaussian neighborhood)
- train_v6_1.py: 5 datasets from streaming_datasets.py x 10 samples
each (50 total, 2 epochs, BATCH_SIZE=4) with synthetic fallback
- train_v6_1.py: PGVector status captured per ste… Aug 6 23:32 V6: Kohonen 4D SOM + EWC per-neuron + MTP+entropy + Xeon V6 V6 MELHORIAS VS V5:
1. OrqCell V6 Kohonen 4D SOM (orq_cell.py reescrito):
- 8 neurônios BiGRU4 em grid 4D (2,2,2,1)
- TransformerUnit = dimensão semântica (função de vizinhança aprendível)
- Função Gaussiana Λ(d, σ) = exp(-d²/(2σ²))
- Schedule σ: 1.5 → 0.3, α: 0.5 → 0.05
- BMU + vizinhos puxados na direção do input (preserva topologia)
- Compatibilidade com datasets não-preparados: random pad
- update_map_after_backward() (não quebra autograd)
2. EWC V6 com Kohonen per-neuron (ewc.py novo):
- Fisher por neurônio (n_neurons, d_cache)
- Reforço γ_bmu=2.0 no BMU
- Penalidade topológica: preserva distâncias no grid 4D
- Funciona em eval mode (forward-only penalty)
3. MTP V6 com entropia + ativo em val (mtp.py novo):
- L_mtp = Σ α_k L_k - β H(α), β=0.01
- MTP ativo em val (métrica adicional val_mtp_loss)
- mtp_alphas aprendíveis (K=4)
4. Window Context 2M tokens (window_context.py):
- SequenceCompressor + ChunkIndexer
- Adaptive attention:… 23:01 V5: W8A8 aprimorado (α aprendível via STE) + bug fix apply_w8a8_to_model + 4 hiperparâmetros auto-ajustáveis (T, τ, λ_ent, init_gate) + OOM-Killer fixes + num_layers_hyp=8 fixo 20:36 V6: KohonenSOM4D 4D em orq_cell (8 BiGRU × 4 sem × 4x4 grid = 512 protótipos) + WindowContext 2M (ToMe + KNN IVF) + MTP (K=4) + MTPAdaptiveAttention + DynamicRepetitionPenalty + ElasticWeightConsolidation + data_augmentation + RLHF Possibility Mapping (Φ = Θ + η·r·σ(trust·τ_T)·Σwᵢgᵢ) + ReasoningEngine 8-fase + Tools (cache + mutation + agent) + optimized_pipeline + langchain_adapter 20:14 V5: W8A8 aprimorado (α aprendível via STE) + bug fix apply_w8a8_to_model + 4 hiperparâmetros auto-ajustáveis (T, τ, λ_ent, init_gate) + OOM-Killer fixes + num_layers_hyp=8 fixo 11:42 V3.2: corrigir 6 bugs + meta_T_moved (T_tracking_loss) + re-teste com Xeon IPEX 2.8 + AMX 10:43 V3.2: corrigir 6 bugs + meta_T_moved (T_tracking_loss) + re-teste com Xeon IPEX 2.8 + AMX 00:03 V2: 12.97M params, HW optimizer, parallel BBPE, DPO, reasoning, inference, 10 bugs fixed Aug 5 22:47 Upload folder using huggingface_hub 
