Delete v6_5_7ds_attn_v3_predict_fix_eval.json

v6_5_7ds_attn_v3_predict_fix_eval.json

@@ -1,153 +0,0 @@
-{
-  "evaluation": "predict_fix_v65",
-  "user_requirement": "rótulos 'gato'/'cachorro' fixos → extração variável e flexível",
-  "n_test_queries": 6,
-  "results": [
-    {
-      "query": "o gato dorme na cama",
-      "prediction": "body",
-      "prediction_proba": "body",
-      "probability": 0.9416029453277588,
-      "is_gato_hardcoded": false,
-      "is_cachorro_hardcoded": false,
-      "is_registry_label": true,
-      "is_default_label": false
-    },
-    {
-      "query": "calcule dois mais dois",
-      "prediction": "body",
-      "prediction_proba": "body",
-      "probability": 0.9416029453277588,
-      "is_gato_hardcoded": false,
-      "is_cachorro_hardcoded": false,
-      "is_registry_label": true,
-      "is_default_label": false
-    },
-    {
-      "query": "qual é a capital do brasil",
-      "prediction": "body",
-      "prediction_proba": "body",
-      "probability": 0.9416029453277588,
-      "is_gato_hardcoded": false,
-      "is_cachorro_hardcoded": false,
-      "is_registry_label": true,
-      "is_default_label": false
-    },
-    {
-      "query": "explique o que é uma rede neural",
-      "prediction": "body",
-      "prediction_proba": "body",
-      "probability": 0.9416029453277588,
-      "is_gato_hardcoded": false,
-      "is_cachorro_hardcoded": false,
-      "is_registry_label": true,
-      "is_default_label": false
-    },
-    {
-      "query": "olá como você está",
-      "prediction": "body",
-      "prediction_proba": "body",
-      "probability": 0.9416029453277588,
-      "is_gato_hardcoded": false,
-      "is_cachorro_hardcoded": false,
-      "is_registry_label": true,
-      "is_default_label": false
-    },
-    {
-      "query": "traduza hello para portugues",
-      "prediction": "body",
-      "prediction_proba": "body",
-      "probability": 0.9416029453277588,
-      "is_gato_hardcoded": false,
-      "is_cachorro_hardcoded": false,
-      "is_registry_label": true,
-      "is_default_label": false
-    }
-  ],
-  "per_dataset_results": [
-    {
-      "dataset": "dominguesm/restore-punctuation-ptbr-dataset",
-      "expected_labels": [
-        "unpunctuated",
-        "punctuated"
-      ],
-      "prediction": "punctuated",
-      "valid": true
-    },
-    {
-      "dataset": "carolina-c4ai/corpus-carolina",
-      "expected_labels": [
-        "raw_corpus",
-        "normalized_text"
-      ],
-      "prediction": "normalized_text",
-      "valid": true
-    },
-    {
-      "dataset": "CEIA-POSITIVO/ultrachat_br_clustred_balanced_v1",
-      "expected_labels": [
-        "user_turn",
-        "assistant_turn"
-      ],
-      "prediction": "assistant_turn",
-      "valid": true
-    },
-    {
-      "dataset": "dominguesm/Canarim-Instruct-PTBR-Dataset",
-      "expected_labels": [
-        "instruction",
-        "response"
-      ],
-      "prediction": "response",
-      "valid": true
-    },
-    {
-      "dataset": "adalbertojunior/punctuation-ptbr",
-      "expected_labels": [
-        "unpunctuated",
-        "punctuated"
-      ],
-      "prediction": "punctuated",
-      "valid": true
-    },
-    {
-      "dataset": "iara-project/news-articles-ptbr-dataset",
-      "expected_labels": [
-        "headline",
-        "body"
-      ],
-      "prediction": "body",
-      "valid": true
-    },
-    {
-      "dataset": "manoela/noticias_ptbr",
-      "expected_labels": [
-        "headline",
-        "body"
-      ],
-      "prediction": "body",
-      "valid": true
-    }
-  ],
-  "summary": {
-    "n_returns_gato": 0,
-    "n_returns_cachorro": 0,
-    "n_returns_registry_label": 6,
-    "n_returns_default_label": 0,
-    "fix_pass": true,
-    "n_datasets_registered": 7,
-    "last_dataset_used": "manoela/noticias_ptbr"
-  },
-  "mathematical_correction": {
-    "before_bug": "return 'gato' if prob <= 0.5 else 'cachorro'  # FIXO",
-    "after_fix": "label_int = 1 if prob > 0.5 else 0; label_str = self.get_label_string(label_int, dataset_name)",
-    "registry_lookup": "label_registry[dataset_name][label_int]",
-    "fallback_chain": [
-      "1. registry[dataset_name][label_int]",
-      "2. registry[last_dataset][label_int]",
-      "3. default {0: 'negative', 1: 'positive'}",
-      "4. synthetic f'label_{label_int}'"
-    ]
-  },
-  "verdict": "PASS"
-}