Delete v6_5_7ds_attn_v3_inference_punishment.json

v6_5_7ds_attn_v3_inference_punishment.json

@@ -1,142 +0,0 @@
-{
-  "evaluation": "user_questions_with_punishment_v65_fix",
-  "user_requirement": "não ajudar o modelo em respostas E ao punir o modelo reajusta seus parâmetros usando a camada de hipótese (mirror do treinamento)",
-  "questions_sent_verbatim": true,
-  "no_context_added": true,
-  "no_system_prompt": true,
-  "no_few_shot": true,
-  "n_questions": 3,
-  "questions": [
-    "Luva de Pedreiro Távila",
-    "Lula reserva valor",
-    "Amazonas força-tarefa vítimas"
-  ],
-  "results": [
-    {
-      "query": "Luva de Pedreiro Távila",
-      "query_was_modified": false,
-      "context_provided": false,
-      "system_prompt_used": false,
-      "few_shot_examples": false,
-      "som_prediction_before": "body",
-      "probability_before": 0.9416029453277588,
-      "som_prediction_after": "body",
-      "probability_after": 0.8029372096061707,
-      "prediction_changed_by_punishment": false,
-      "needs_punishment": true,
-      "punishment_outcome": {
-        "action": "activate_hypothesis",
-        "inference_punishment_count": 1,
-        "punishment_count_cycle": 1,
-        "classifier_trained": true,
-        "ewc_reference_set": true,
-        "accuracy_after_punishment": 0.8,
-        "buffer_size": 5,
-        "bmu_of_misclassified": [
-          1,
-          4,
-          2,
-          1
-        ],
-        "elapsed_ms": 283.8399410247803
-      },
-      "reasoning_length": 1012,
-      "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=body | BMU=(1, 4, 2, 1)",
-      "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": 11.647462844848633
-    },
-    {
-      "query": "Lula reserva valor",
-      "query_was_modified": false,
-      "context_provided": false,
-      "system_prompt_used": false,
-      "few_shot_examples": false,
-      "som_prediction_before": "body",
-      "probability_before": 0.8029372096061707,
-      "som_prediction_after": "body",
-      "probability_after": 0.8029372096061707,
-      "prediction_changed_by_punishment": false,
-      "needs_punishment": false,
-      "punishment_outcome": null,
-      "reasoning_length": 982,
-      "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=body | BMU=(1, 4, 2, 1)",
-      "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": 12.577533721923828
-    },
-    {
-      "query": "Amazonas força-tarefa vítimas",
-      "query_was_modified": false,
-      "context_provided": false,
-      "system_prompt_used": false,
-      "few_shot_examples": false,
-      "som_prediction_before": "body",
-      "probability_before": 0.8029372096061707,
-      "som_prediction_after": "body",
-      "probability_after": 0.8029372096061707,
-      "prediction_changed_by_punishment": false,
-      "needs_punishment": false,
-      "punishment_outcome": null,
-      "reasoning_length": 1048,
-      "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=body | BMU=(1, 4, 2, 1)",
-      "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": 11.980772018432617
-    }
-  ],
-  "punishment_outcomes": [
-    {
-      "action": "activate_hypothesis",
-      "inference_punishment_count": 1,
-      "punishment_count_cycle": 1,
-      "classifier_trained": true,
-      "ewc_reference_set": true,
-      "accuracy_after_punishment": 0.8,
-      "buffer_size": 5,
-      "bmu_of_misclassified": [
-        1,
-        4,
-        2,
-        1
-      ],
-      "elapsed_ms": 283.8399410247803
-    }
-  ],
-  "summary": {
-    "n_punishments_applied": 1,
-    "n_predictions_changed_by_punishment": 0,
-    "n_with_answer": 3,
-    "n_with_think": 3,
-    "n_activate_hypothesis": 1,
-    "n_set_ewc_reference": 0,
-    "avg_latency_ms": 12.068589528401693,
-    "inference_punishment_count_total": 1
-  },
-  "mathematical_principle": {
-    "training_protocol": "if acc < 1.0: punishment_count += 1; if punishment_count == 1: activate_hypothesis(); elif punishment_count == 2: set_ewc_reference(); reset",
-    "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",
-    "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)"
-  },
-  "quality_assessment": {
-    "response_quality": "GOOD",
-    "punishment_applied": true,
-    "model_self_corrected": false,
-    "model_not_helped": true
-  }
-}