V6.5-V2-dynamic: upload batch (scripts + state + model) [13 files]

kohonen_learning_system.py

@@ -581,6 +581,16 @@ class KohonenLearningSystem:
 
         self.buffer_4d = []
         self.buffer_labels = []
+        # V6.5-V2-memfix — Sliding window limit for buffer_4d to prevent OOM.
+        # User requirement: "resolver falhas de lógica e de bugs que estejam
+        # causando alto consumo de memória sem distorcer a arquitetura Kohonen".
+        # The buffer stores vec_4d tensors for SOM training and neuron labeling.
+        # Without a limit, streaming 8000+ samples accumulates 8000 tensors,
+        # each retaining the embedding computation graph (16M params), causing
+        # OOM (3.5GB RSS observed). Sliding window of 512 keeps recent samples
+        # for SOM updates while bounding memory. The Kohonen architecture (SOM
+        # grid, BMU, Gaussian neighborhood, EWC) is NOT changed.
+        self.buffer_max_size = 512
         self.training_ready = False
         self.N = N_start
         self.dim_choice = dim_choice
@@ -980,23 +990,36 @@ class KohonenLearningSystem:
         if dataset_name is not None:
             self.last_dataset_used = dataset_name
 
-        for sent, lab in zip(sentences, labels):
-            self.time_counter += 1
-            vec = text_to_4d_vector(
-                sent,
-                self.tokenizer,
-                self.embedding,
-                self.hidden_dim,
-                self.seq_len,
-                self.time_counter,
-                self.T_max,
-                attention=self.attention,           # V6.5-attn
-                attention_metrics=self.attention_metrics,  # V6.5-attn
-            )
-            self.buffer_4d.append(vec)
-            self.buffer_labels.append(lab)
-            dim_val = round(vec[self.dim_index].item(), 2)
-            self.histogram[dim_val] += 1
+        # V6.5-V2-memfix — Use torch.no_grad() to avoid building computation
+        # graphs during vec_4d computation. The embedding graph (16M params)
+        # would otherwise be retained by every tensor in buffer_4d, causing OOM
+        # with 8000+ samples. The vec_4d is a 4-float summary (x,y,z,w) that
+        # does NOT need gradient — it feeds into SOM (Hebbian, no backprop) and
+        # the HypothesisClassifier (which trains its own weights).
+        with torch.no_grad():
+            for sent, lab in zip(sentences, labels):
+                self.time_counter += 1
+                vec = text_to_4d_vector(
+                    sent,
+                    self.tokenizer,
+                    self.embedding,
+                    self.hidden_dim,
+                    self.seq_len,
+                    self.time_counter,
+                    self.T_max,
+                    attention=self.attention,           # V6.5-attn
+                    attention_metrics=self.attention_metrics,  # V6.5-attn
+                )
+                # CRITICAL: detach to break any residual graph + clone for safety
+                self.buffer_4d.append(vec.detach().clone())
+                self.buffer_labels.append(lab)
+                dim_val = round(vec[self.dim_index].item(), 2)
+                self.histogram[dim_val] += 1
+            # V6.5-V2-memfix — Enforce sliding window limit
+            if len(self.buffer_4d) > self.buffer_max_size:
+                overflow = len(self.buffer_4d) - self.buffer_max_size
+                self.buffer_4d = self.buffer_4d[-self.buffer_max_size:]
+                self.buffer_labels = self.buffer_labels[-self.buffer_max_size:]
 
     def check_training_start(self) -> bool:
         """Inicia treino quando algum bucket do histograma atinge N."""
@@ -1009,20 +1032,27 @@ class KohonenLearningSystem:
         return False
 
     def train_som_on_buffer(self):
-        """Treina SOM por 5 épocas sobre o buffer atual (com Fisher accum)."""
+        """Treina SOM por 5 épocas sobre o buffer atual (com Fisher accum).
+
+        V6.5-V2-memfix — Wrapped in torch.no_grad() to prevent graph building
+        during SOM Hebbian updates. The SOM uses competitive learning (BMU +
+        Gaussian neighborhood), not backprop — no gradient is needed.
+        """
         if not self.buffer_4d:
             return
-        data = torch.stack(self.buffer_4d)
-        for _ in range(5):  # épocas de treino rápido
-            perm = torch.randperm(len(data))
-            for idx in perm:
-                x = data[idx]
-                bmu = self.som.find_bmu(x)
-                acc_fisher = (
-                    self.punishment_count == 0
-                    and self.som.old_weights_w is None
-                )
-                self.som.update_weights(x, bmu, accumulate_fisher=acc_fisher)
+        # V6.5-V2-memfix — detach+stack to ensure no graph leaks from buffer
+        data = torch.stack(self.buffer_4d).detach()
+        with torch.no_grad():
+            for _ in range(5):  # épocas de treino rápido
+                perm = torch.randperm(len(data))
+                for idx in perm:
+                    x = data[idx]
+                    bmu = self.som.find_bmu(x)
+                    acc_fisher = (
+                        self.punishment_count == 0
+                        and self.som.old_weights_w is None
+                    )
+                    self.som.update_weights(x, bmu, accumulate_fisher=acc_fisher)
 
         # V6.5 — Ativa VQ-VAE-2 compressor no pipeline
         if self.enable_vqvae2 and self.vqvae2_compressor is not None:
@@ -1350,30 +1380,38 @@ class KohonenLearningSystem:
         return dist.flatten()
 
     def _label_neurons(self):
-        """Rotula neurônios por votação majoritária sobre o buffer."""
+        """Rotula neurônios por votação majoritária sobre o buffer.
+
+        V6.5-V2-memfix — no_grad to prevent graph building during BMU search.
+        """
         self.neuron_label = {}
         if not self.buffer_4d:
             return
-        data = torch.stack(self.buffer_4d)
+        data = torch.stack(self.buffer_4d).detach()
         labels = torch.tensor(self.buffer_labels)
         votes = defaultdict(lambda: [0, 0])
-        for i in range(len(data)):
-            bmu = self.som.find_bmu(data[i])
-            votes[bmu][int(labels[i].item())] += 1
+        with torch.no_grad():
+            for i in range(len(data)):
+                bmu = self.som.find_bmu(data[i])
+                votes[bmu][int(labels[i].item())] += 1
         for bmu, v in votes.items():
             self.neuron_label[bmu] = 1.0 if v[1] > v[0] else 0.0
 
     def evaluate_classification(self) -> float:
-        """Acurácia sobre o buffer atual."""
+        """Acurácia sobre o buffer atual.
+
+        V6.5-V2-memfix — no_grad to prevent graph building during evaluation.
+        """
         if not self.buffer_4d:
             return 1.0
-        data = torch.stack(self.buffer_4d)
+        data = torch.stack(self.buffer_4d).detach()
         labels = torch.tensor(self.buffer_labels).float()
         correct = 0
-        for i in range(len(data)):
-            pred = self._predict_single(data[i])
-            if (pred > 0.5) == (labels[i] > 0.5):
-                correct += 1
+        with torch.no_grad():
+            for i in range(len(data)):
+                pred = self._predict_single(data[i])
+                if (pred > 0.5) == (labels[i] > 0.5):
+                    correct += 1
         return correct / len(data)
 
     def _predict_single(self, x):
@@ -1624,22 +1662,29 @@ class KohonenLearningSystem:
             self.last_dataset_used = dataset_name
 
         # 2. Compute vec_4d for the misclassified sentence
-        self.time_counter += 1
-        vec = text_to_4d_vector(
-            sentence,
-            self.tokenizer,
-            self.embedding,
-            self.hidden_dim,
-            self.seq_len,
-            self.time_counter,
-            self.T_max,
-            attention=self.attention,
-            attention_metrics=self.attention_metrics,
-        )
+        # V6.5-V2-memfix — no_grad + detach to prevent graph accumulation
+        with torch.no_grad():
+            self.time_counter += 1
+            vec = text_to_4d_vector(
+                sentence,
+                self.tokenizer,
+                self.embedding,
+                self.hidden_dim,
+                self.seq_len,
+                self.time_counter,
+                self.T_max,
+                attention=self.attention,
+                attention_metrics=self.attention_metrics,
+            )
+            vec = vec.detach().clone()
 
         # 3. Add to buffer + update histogram (mirror add_data)
         self.buffer_4d.append(vec)
         self.buffer_labels.append(int(correct_label_int))
+        # Enforce sliding window limit (mirror add_data)
+        if len(self.buffer_4d) > self.buffer_max_size:
+            self.buffer_4d = self.buffer_4d[-self.buffer_max_size:]
+            self.buffer_labels = self.buffer_labels[-self.buffer_max_size:]
         dim_val = round(vec[self.dim_index].item(), 2)
         self.histogram[dim_val] += 1
 

v6_5_v2_attention_eval.json

@@ -3,18 +3,18 @@
   "user_requirement": "verificar se o mecanismo de atenção está ativo e acessado logicamente funcional",
   "metrics": {
     "active": true,
-    "n_calls": 3441,
+    "n_calls": 10680,
     "n_errors": 0,
-    "last_norm_in": 54.769012451171875,
-    "last_norm_out": 56.71458053588867,
+    "last_norm_in": 111.9294204711914,
+    "last_norm_out": 116.00782012939453,
     "last_attn_activated": true,
-    "last_attn_diff_norm": 56.80364227294922,
+    "last_attn_diff_norm": 116.50959014892578,
     "n_heads": 8,
     "logic_functional": true
   },
   "active": true,
   "logic_functional": true,
-  "n_calls": 3441,
+  "n_calls": 10680,
   "n_errors": 0,
   "n_heads": 8,
   "assessment": "PASS"

v6_5_v2_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff42c43886461f7b5cfd55a0d125ce6dbf4f719a0d41383d3991fe687e60958e
+size 220174132

v6_5_v2_model_states_after_conhecimento.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e59c64fa9ed050f79bddf2f3546b104a93b846e7af732bd4527c9ed10c578da
+size 217694233

v6_5_v2_predict_fix_eval.json

@@ -5,8 +5,8 @@
   "results": [
     {
       "query": "o gato dorme na cama",
-      "prediction": "unpunctuated",
-      "probability": 0.0,
+      "prediction": "short_text",
+      "probability": NaN,
       "is_gato_hardcoded": false,
       "is_cachorro_hardcoded": false,
       "is_registry_label": true,
@@ -14,8 +14,8 @@
     },
     {
       "query": "calcule dois mais dois",
-      "prediction": "unpunctuated",
-      "probability": 0.0,
+      "prediction": "short_text",
+      "probability": NaN,
       "is_gato_hardcoded": false,
       "is_cachorro_hardcoded": false,
       "is_registry_label": true,
@@ -23,8 +23,8 @@
     },
     {
       "query": "qual é a capital do brasil",
-      "prediction": "unpunctuated",
-      "probability": 0.0,
+      "prediction": "short_text",
+      "probability": NaN,
       "is_gato_hardcoded": false,
       "is_cachorro_hardcoded": false,
       "is_registry_label": true,
@@ -32,8 +32,8 @@
     },
     {
       "query": "explique o que é uma rede neural",
-      "prediction": "unpunctuated",
-      "probability": 0.0,
+      "prediction": "short_text",
+      "probability": NaN,
       "is_gato_hardcoded": false,
       "is_cachorro_hardcoded": false,
       "is_registry_label": true,
@@ -41,8 +41,8 @@
     },
     {
       "query": "olá como você está",
-      "prediction": "unpunctuated",
-      "probability": 0.0,
+      "prediction": "short_text",
+      "probability": NaN,
       "is_gato_hardcoded": false,
       "is_cachorro_hardcoded": false,
       "is_registry_label": true,
@@ -50,8 +50,8 @@
     },
     {
       "query": "traduza hello para portugues",
-      "prediction": "unpunctuated",
-      "probability": 0.0,
+      "prediction": "short_text",
+      "probability": NaN,
       "is_gato_hardcoded": false,
       "is_cachorro_hardcoded": false,
       "is_registry_label": true,
@@ -103,6 +103,33 @@
       ],
       "prediction": "unpunctuated",
       "valid_for_dataset": true
+    },
+    {
+      "dataset": "iara-project/news-articles-ptbr-dataset",
+      "expected_labels": [
+        "headline",
+        "body"
+      ],
+      "prediction": "headline",
+      "valid_for_dataset": true
+    },
+    {
+      "dataset": "manoela/noticias_ptbr",
+      "expected_labels": [
+        "headline",
+        "body"
+      ],
+      "prediction": "headline",
+      "valid_for_dataset": true
+    },
+    {
+      "dataset": "BrunoN-Dev/corpus-ptbr-v1",
+      "expected_labels": [
+        "short_text",
+        "long_text"
+      ],
+      "prediction": "short_text",
+      "valid_for_dataset": true
     }
   ],
   "summary": {

v6_5_v2_report.json

@@ -1,6 +1,6 @@
 {
   "version": "V6.5-V2",
-  "timestamp": "2026-08-08T21:26:39.965059",
+  "timestamp": "2026-08-08T23:52:08.862775",
   "config": {
     "som_grid": [
       6,
@@ -9,14 +9,14 @@
       4
     ],
     "n_neurons": 864,
-    "hidden_dim": 256,
-    "vocab_size": 4096,
-    "n_hypotheses": 8,
+    "hidden_dim": 1024,
+    "vocab_size": 16384,
+    "n_hypotheses": 16,
     "n_trials": 3,
-    "hyp_train_steps": 20,
+    "hyp_train_steps": 30,
     "stream_batch_size": 100,
-    "meta_conhecimento": 6000,
-    "meta_punicão": 1000,
+    "meta_conhecimento": 8000,
+    "meta_punicão": 2000,
     "conhecimento_datasets": [
       "dominguesm/restore-punctuation-ptbr-dataset",
       "carolina-c4ai/corpus-carolina",
@@ -54,10 +54,10 @@
     "init_done": true
   },
   "fp16_benchmark": {
-    "best_time_ms": 94.09655899980862,
-    "avg_time_ms": 97.57540749978943,
-    "best_tflops": 1.3603047907443706,
-    "avg_tflops": 1.3118059486482412,
+    "best_time_ms": 91.0796530006337,
+    "avg_time_ms": 91.46852800040506,
+    "best_tflops": 1.4053632812930177,
+    "avg_tflops": 1.3993884322641899,
     "matrix_size": 4000.0
   },
   "v2_verification": {
@@ -84,24 +84,24 @@
     "all_pass": true
   },
   "fase_1_conhecimento_summary": {
-    "total_samples": 3300,
-    "meta_atingida": false,
-    "elapsed_s": 425.32095551490784,
-    "storage_critical_stopped": true
+    "total_samples": 8000,
+    "meta_atingida": true,
+    "elapsed_s": 722.1084952354431,
+    "storage_critical_stopped": false
   },
   "fase_2_punicão_summary": {
-    "total_samples": 0,
-    "meta_atingida": false,
-    "elapsed_s": 0.0,
-    "punishment_events": 0,
-    "hypotheses_trainings": 0,
-    "delta_applications": 0,
-    "skipped": true
+    "total_samples": 2000,
+    "meta_atingida": true,
+    "elapsed_s": 507.18996810913086,
+    "punishment_events": 120,
+    "hypotheses_trainings": 60,
+    "delta_applications": 60,
+    "skipped": false
   },
   "attention_eval_summary": {
     "active": true,
     "logic_functional": true,
-    "n_calls": 3441,
+    "n_calls": 10680,
     "assessment": "PASS"
   },
   "predict_fix_summary": {
@@ -117,56 +117,178 @@
     "n_with_think": 3,
     "answer_rate": 1.0,
     "think_rate": 1.0,
-    "avg_latency_ms": 3.060022989908854,
-    "avg_reasoning_length": 1030.0
+    "avg_latency_ms": 11.263132095336914,
+    "avg_reasoning_length": 1026.0
   },
   "model_states_saved_to": "/home/z/my-project/BiGRU_T_version/v6_5_v2_model_states.pt",
   "save_info": {
     "saved": true,
     "path": "/home/z/my-project/BiGRU_T_version/v6_5_v2_model_states.pt",
-    "size_mb": 40.959502,
+    "size_mb": 220.174132,
     "reason": "end_of_training_v2",
-    "step": 225,
-    "total_samples": 3300,
+    "step": 700,
+    "total_samples": 10000,
     "n_tensors": 7,
     "n_buffer_tail": 128
   },
   "final_v2_metrics": {
     "version": "V2-dynamic",
-    "n_hypotheses": 8,
-    "n_hypotheses_active": 8,
-    "max_n_hypotheses": 16,
-    "n_trials": 3,
-    "hyp_train_steps": 20,
+    "n_hypotheses": 16,
+    "n_hypotheses_active": 16,
+    "max_n_hypotheses": 32,
+    "n_trials": 6,
+    "hyp_train_steps": 30,
     "hyp_lr": 0.0001,
-    "delta_scale": 0.009999999776482582,
-    "n_generators": 16,
+    "delta_scale": 0.0010000000474974513,
+    "n_generators": 32,
     "punishment_count": 0,
     "success_count": 0,
-    "training_ready": true,
-    "classifier_trained": false,
-    "ewc_reference_set": false,
+    "training_ready": false,
+    "classifier_trained": true,
+    "ewc_reference_set": true,
     "buffer_size": 256,
     "dynamic_adaptation": {
-      "loss_history_len": 0,
+      "loss_history_len": 8,
       "loss_stats": {
-        "slope": 0.0,
-        "volatility": 0.0,
-        "mean": 0.0,
-        "n": 0
+        "slope": NaN,
+        "volatility": NaN,
+        "mean": NaN,
+        "std": NaN,
+        "n": 8
       },
-      "punishment_rate": 0.0,
-      "punishment_window_size": 0,
-      "n_adaptations": 0,
+      "punishment_rate": 1.0,
+      "punishment_window_size": 12,
+      "n_adaptations": 20,
       "limits": {
         "min_n_hypotheses": 4,
-        "max_n_hypotheses": 16,
+        "max_n_hypotheses": 32,
         "min_n_trials": 1,
         "max_n_trials": 6,
-        "min_hyp_train_steps": 5,
-        "max_hyp_train_steps": 60
+        "min_hyp_train_steps": 10,
+        "max_hyp_train_steps": 80
       },
-      "last_5_adaptations": []
+      "last_5_adaptations": [
+        {
+          "trigger": "punishment",
+          "step": 115,
+          "before": {
+            "n_hypotheses": 16,
+            "n_trials": 6,
+            "hyp_train_steps": 30
+          },
+          "after": {
+            "n_hypotheses": 16,
+            "n_trials": 6,
+            "hyp_train_steps": 30
+          },
+          "adapted": false,
+          "rules_fired": [],
+          "loss_stats": {
+            "slope": NaN,
+            "volatility": NaN,
+            "mean": NaN,
+            "std": NaN,
+            "n": 8
+          },
+          "punishment_rate": 1.0
+        },
+        {
+          "trigger": "auto",
+          "step": 116,
+          "before": {
+            "n_hypotheses": 16,
+            "n_trials": 6,
+            "hyp_train_steps": 30
+          },
+          "after": {
+            "n_hypotheses": 16,
+            "n_trials": 6,
+            "hyp_train_steps": 30
+          },
+          "adapted": false,
+          "rules_fired": [],
+          "loss_stats": {
+            "slope": NaN,
+            "volatility": NaN,
+            "mean": NaN,
+            "std": NaN,
+            "n": 8
+          },
+          "punishment_rate": 1.0
+        },
+        {
+          "trigger": "punishment",
+          "step": 117,
+          "before": {
+            "n_hypotheses": 16,
+            "n_trials": 6,
+            "hyp_train_steps": 30
+          },
+          "after": {
+            "n_hypotheses": 16,
+            "n_trials": 6,
+            "hyp_train_steps": 30
+          },
+          "adapted": false,
+          "rules_fired": [],
+          "loss_stats": {
+            "slope": NaN,
+            "volatility": NaN,
+            "mean": NaN,
+            "std": NaN,
+            "n": 8
+          },
+          "punishment_rate": 1.0
+        },
+        {
+          "trigger": "auto",
+          "step": 118,
+          "before": {
+            "n_hypotheses": 16,
+            "n_trials": 6,
+            "hyp_train_steps": 30
+          },
+          "after": {
+            "n_hypotheses": 16,
+            "n_trials": 6,
+            "hyp_train_steps": 30
+          },
+          "adapted": false,
+          "rules_fired": [],
+          "loss_stats": {
+            "slope": NaN,
+            "volatility": NaN,
+            "mean": NaN,
+            "std": NaN,
+            "n": 8
+          },
+          "punishment_rate": 1.0
+        },
+        {
+          "trigger": "punishment",
+          "step": 119,
+          "before": {
+            "n_hypotheses": 16,
+            "n_trials": 6,
+            "hyp_train_steps": 30
+          },
+          "after": {
+            "n_hypotheses": 16,
+            "n_trials": 6,
+            "hyp_train_steps": 30
+          },
+          "adapted": false,
+          "rules_fired": [],
+          "loss_stats": {
+            "slope": NaN,
+            "volatility": NaN,
+            "mean": NaN,
+            "std": NaN,
+            "n": 8
+          },
+          "punishment_rate": 1.0
+        }
+      ]
     }
   }
 }

v6_5_v2_user_questions.json

@@ -18,17 +18,17 @@
       "context_provided": false,
       "system_prompt_used": false,
       "few_shot_examples": false,
-      "som_prediction": "unpunctuated",
-      "reasoning_length": 1028,
+      "som_prediction": "short_text",
+      "reasoning_length": 1024,
       "has_think": true,
       "has_plan": true,
       "has_answer": true,
       "has_decompose": true,
       "think_preview": "Analisando a query: 'Luva de Pedreiro Távila'\nIdentificando o tipo de problema e requisitos.\nDeterminando se ferramentas são necessárias.",
-      "answer_preview": "prediction=unpunctuated | BMU=(4, 4, 5, 0)",
+      "answer_preview": "prediction=short_text | BMU=(0, 0, 0, 0)",
       "raw_response_preview": "<think>\nAnalisando a query: 'Luva de Pedreiro Távila'\nIdentificando o tipo de problema e requisitos.\nDeterminando se ferramentas são necessárias.\n</think>\n<plan>\nPlano de resolução:\n1. Decompor o problema em sub-tarefas\n2. Identificar ferramentas necessárias (disponíveis: som_query, buffer_stats)\n3. Executar sub-tarefas em sequência\n4. Monitorar resultados\n5. Compor resposta final\n</plan>\n<decompose>\n- Processar: Luva de Pedreiro Távila\n</decompose>\n<execute>\nSub-tarefa 'Processar: Luva de Pedre...",
       "n_tags": 4,
-      "latency_ms": 3.1023025512695312
+      "latency_ms": 9.467601776123047
     },
     {
       "query": "Lula reserva valor",
@@ -36,17 +36,17 @@
       "context_provided": false,
       "system_prompt_used": false,
       "few_shot_examples": false,
-      "som_prediction": "unpunctuated",
-      "reasoning_length": 998,
+      "som_prediction": "short_text",
+      "reasoning_length": 994,
       "has_think": true,
       "has_plan": true,
       "has_answer": true,
       "has_decompose": true,
       "think_preview": "Analisando a query: 'Lula reserva valor'\nIdentificando o tipo de problema e requisitos.\nDeterminando se ferramentas são necessárias.",
-      "answer_preview": "prediction=unpunctuated | BMU=(4, 4, 5, 0)",
+      "answer_preview": "prediction=short_text | BMU=(0, 0, 0, 0)",
       "raw_response_preview": "<think>\nAnalisando a query: 'Lula reserva valor'\nIdentificando o tipo de problema e requisitos.\nDeterminando se ferramentas são necessárias.\n</think>\n<plan>\nPlano de resolução:\n1. Decompor o problema em sub-tarefas\n2. Identificar ferramentas necessárias (disponíveis: som_query, buffer_stats)\n3. Executar sub-tarefas em sequência\n4. Monitorar resultados\n5. Compor resposta final\n</plan>\n<decompose>\n- Processar: Lula reserva valor\n</decompose>\n<execute>\nSub-tarefa 'Processar: Lula reserva valor' exe...",
       "n_tags": 4,
-      "latency_ms": 3.0732154846191406
+      "latency_ms": 11.906862258911133
     },
     {
       "query": "Amazonas força-tarefa vítimas",
@@ -54,17 +54,17 @@
       "context_provided": false,
       "system_prompt_used": false,
       "few_shot_examples": false,
-      "som_prediction": "unpunctuated",
-      "reasoning_length": 1064,
+      "som_prediction": "short_text",
+      "reasoning_length": 1060,
       "has_think": true,
       "has_plan": true,
       "has_answer": true,
       "has_decompose": true,
       "think_preview": "Analisando a query: 'Amazonas força-tarefa vítimas'\nIdentificando o tipo de problema e requisitos.\nDeterminando se ferramentas são necessárias.",
-      "answer_preview": "prediction=unpunctuated | BMU=(4, 4, 5, 0)",
+      "answer_preview": "prediction=short_text | BMU=(0, 0, 0, 0)",
       "raw_response_preview": "<think>\nAnalisando a query: 'Amazonas força-tarefa vítimas'\nIdentificando o tipo de problema e requisitos.\nDeterminando se ferramentas são necessárias.\n</think>\n<plan>\nPlano de resolução:\n1. Decompor o problema em sub-tarefas\n2. Identificar ferramentas necessárias (disponíveis: som_query, buffer_stats)\n3. Executar sub-tarefas em sequência\n4. Monitorar resultados\n5. Compor resposta final\n</plan>\n<decompose>\n- Processar: Amazonas força-tarefa vítimas\n</decompose>\n<execute>\nSub-tarefa 'Processar: A...",
       "n_tags": 4,
-      "latency_ms": 3.0045509338378906
+      "latency_ms": 12.414932250976562
     }
   ],
   "summary": {
@@ -72,8 +72,8 @@
     "n_with_think": 3,
     "answer_rate": 1.0,
     "think_rate": 1.0,
-    "avg_latency_ms": 3.060022989908854,
-    "avg_reasoning_length": 1030.0
+    "avg_latency_ms": 11.263132095336914,
+    "avg_reasoning_length": 1026.0
   },
   "quality_assessment": {
     "model_not_helped": true,