{
  "settings": {
    "n": 10,
    "seed": 42,
    "caption_field": "caption_cogvlm",
    "probe_count": 35,
    "retries": 2,
    "temperature": 0.0,
    "max_tokens": 900,
    "model_env": "meta-llama/llama-3.1-8b-instruct"
  },
  "overall_metrics": {
    "explicit": {
      "tp": 49,
      "fp": 56,
      "fn": 19,
      "precision": 0.466667,
      "recall": 0.720588,
      "f1": 0.566474
    },
    "strong": {
      "tp": 49,
      "fp": 56,
      "fn": 19,
      "precision": 0.466667,
      "recall": 0.720588,
      "f1": 0.566474
    }
  },
  "diagnostics": {
    "samples_with_attempt_failures": 0,
    "samples_with_call_exhaustion": 0,
    "avg_attempt_failure_rate": 0.0,
    "avg_call_exhaustion_rate": 0.0
  },
  "top_tags_by_f1_strong": [
    {
      "tag": "outside",
      "bundle": "scene_pose",
      "needs_glossary": "0",
      "support_pos": "2",
      "support_neg": "8",
      "tp_explicit": "2",
      "fp_explicit": "0",
      "fn_explicit": "0",
      "precision_explicit": "1.000000",
      "recall_explicit": "1.000000",
      "f1_explicit": "1.000000",
      "tp_strong": "2",
      "fp_strong": "0",
      "fn_strong": "0",
      "precision_strong": "1.000000",
      "recall_strong": "1.000000",
      "f1_strong": "1.000000"
    },
    {
      "tag": "eyes_closed",
      "bundle": "gaze_expression",
      "needs_glossary": "0",
      "support_pos": "1",
      "support_neg": "9",
      "tp_explicit": "1",
      "fp_explicit": "0",
      "fn_explicit": "0",
      "precision_explicit": "1.000000",
      "recall_explicit": "1.000000",
      "f1_explicit": "1.000000",
      "tp_strong": "1",
      "fp_strong": "0",
      "fn_strong": "0",
      "precision_strong": "1.000000",
      "recall_strong": "1.000000",
      "f1_strong": "1.000000"
    },
    {
      "tag": "group",
      "bundle": "count_cardinality",
      "needs_glossary": "0",
      "support_pos": "1",
      "support_neg": "9",
      "tp_explicit": "1",
      "fp_explicit": "0",
      "fn_explicit": "0",
      "precision_explicit": "1.000000",
      "recall_explicit": "1.000000",
      "f1_explicit": "1.000000",
      "tp_strong": "1",
      "fp_strong": "0",
      "fn_strong": "0",
      "precision_strong": "1.000000",
      "recall_strong": "1.000000",
      "f1_strong": "1.000000"
    },
    {
      "tag": "feral",
      "bundle": "body_type_presence",
      "needs_glossary": "1",
      "support_pos": "1",
      "support_neg": "9",
      "tp_explicit": "1",
      "fp_explicit": "0",
      "fn_explicit": "0",
      "precision_explicit": "1.000000",
      "recall_explicit": "1.000000",
      "f1_explicit": "1.000000",
      "tp_strong": "1",
      "fp_strong": "0",
      "fn_strong": "0",
      "precision_strong": "1.000000",
      "recall_strong": "1.000000",
      "f1_strong": "1.000000"
    },
    {
      "tag": "<3",
      "bundle": "text_symbols",
      "needs_glossary": "1",
      "support_pos": "1",
      "support_neg": "9",
      "tp_explicit": "1",
      "fp_explicit": "0",
      "fn_explicit": "0",
      "precision_explicit": "1.000000",
      "recall_explicit": "1.000000",
      "f1_explicit": "1.000000",
      "tp_strong": "1",
      "fp_strong": "0",
      "fn_strong": "0",
      "precision_strong": "1.000000",
      "recall_strong": "1.000000",
      "f1_strong": "1.000000"
    },
    {
      "tag": "clothing",
      "bundle": "clothing_state",
      "needs_glossary": "0",
      "support_pos": "8",
      "support_neg": "2",
      "tp_explicit": "8",
      "fp_explicit": "1",
      "fn_explicit": "0",
      "precision_explicit": "0.888889",
      "recall_explicit": "1.000000",
      "f1_explicit": "0.941176",
      "tp_strong": "8",
      "fp_strong": "1",
      "fn_strong": "0",
      "precision_strong": "0.888889",
      "recall_strong": "1.000000",
      "f1_strong": "0.941176"
    },
    {
      "tag": "anthro",
      "bundle": "body_type_presence",
      "needs_glossary": "1",
      "support_pos": "8",
      "support_neg": "2",
      "tp_explicit": "8",
      "fp_explicit": "2",
      "fn_explicit": "0",
      "precision_explicit": "0.800000",
      "recall_explicit": "1.000000",
      "f1_explicit": "0.888889",
      "tp_strong": "8",
      "fp_strong": "2",
      "fn_strong": "0",
      "precision_strong": "0.800000",
      "recall_strong": "1.000000",
      "f1_strong": "0.888889"
    },
    {
      "tag": "bear",
      "bundle": "species_taxonomy",
      "needs_glossary": "0",
      "support_pos": "2",
      "support_neg": "8",
      "tp_explicit": "2",
      "fp_explicit": "1",
      "fn_explicit": "0",
      "precision_explicit": "0.666667",
      "recall_explicit": "1.000000",
      "f1_explicit": "0.800000",
      "tp_strong": "2",
      "fp_strong": "1",
      "fn_strong": "0",
      "precision_strong": "0.666667",
      "recall_strong": "1.000000",
      "f1_strong": "0.800000"
    },
    {
      "tag": "duo",
      "bundle": "count_cardinality",
      "needs_glossary": "1",
      "support_pos": "2",
      "support_neg": "8",
      "tp_explicit": "2",
      "fp_explicit": "1",
      "fn_explicit": "0",
      "precision_explicit": "0.666667",
      "recall_explicit": "1.000000",
      "f1_explicit": "0.800000",
      "tp_strong": "2",
      "fp_strong": "1",
      "fn_strong": "0",
      "precision_strong": "0.666667",
      "recall_strong": "1.000000",
      "f1_strong": "0.800000"
    },
    {
      "tag": "solo",
      "bundle": "count_cardinality",
      "needs_glossary": "1",
      "support_pos": "7",
      "support_neg": "3",
      "tp_explicit": "4",
      "fp_explicit": "0",
      "fn_explicit": "3",
      "precision_explicit": "1.000000",
      "recall_explicit": "0.571429",
      "f1_explicit": "0.727273",
      "tp_strong": "4",
      "fp_strong": "0",
      "fn_strong": "3",
      "precision_strong": "1.000000",
      "recall_strong": "0.571429",
      "f1_strong": "0.727273"
    },
    {
      "tag": "clothed",
      "bundle": "clothing_state",
      "needs_glossary": "0",
      "support_pos": "5",
      "support_neg": "5",
      "tp_explicit": "5",
      "fp_explicit": "4",
      "fn_explicit": "0",
      "precision_explicit": "0.555556",
      "recall_explicit": "1.000000",
      "f1_explicit": "0.714286",
      "tp_strong": "5",
      "fp_strong": "4",
      "fn_strong": "0",
      "precision_strong": "0.555556",
      "recall_strong": "1.000000",
      "f1_strong": "0.714286"
    },
    {
      "tag": "bird",
      "bundle": "species_taxonomy",
      "needs_glossary": "0",
      "support_pos": "2",
      "support_neg": "8",
      "tp_explicit": "1",
      "fp_explicit": "0",
      "fn_explicit": "1",
      "precision_explicit": "1.000000",
      "recall_explicit": "0.500000",
      "f1_explicit": "0.666667",
      "tp_strong": "1",
      "fp_strong": "0",
      "fn_strong": "1",
      "precision_strong": "1.000000",
      "recall_strong": "0.500000",
      "f1_strong": "0.666667"
    },
    {
      "tag": "leporid",
      "bundle": "species_taxonomy",
      "needs_glossary": "1",
      "support_pos": "2",
      "support_neg": "8",
      "tp_explicit": "1",
      "fp_explicit": "0",
      "fn_explicit": "1",
      "precision_explicit": "1.000000",
      "recall_explicit": "0.500000",
      "f1_explicit": "0.666667",
      "tp_strong": "1",
      "fp_strong": "0",
      "fn_strong": "1",
      "precision_strong": "1.000000",
      "recall_strong": "0.500000",
      "f1_strong": "0.666667"
    },
    {
      "tag": "felid",
      "bundle": "species_taxonomy",
      "needs_glossary": "1",
      "support_pos": "2",
      "support_neg": "8",
      "tp_explicit": "1",
      "fp_explicit": "0",
      "fn_explicit": "1",
      "precision_explicit": "1.000000",
      "recall_explicit": "0.500000",
      "f1_explicit": "0.666667",
      "tp_strong": "1",
      "fp_strong": "0",
      "fn_strong": "1",
      "precision_strong": "1.000000",
      "recall_strong": "0.500000",
      "f1_strong": "0.666667"
    },
    {
      "tag": "canis",
      "bundle": "species_taxonomy",
      "needs_glossary": "1",
      "support_pos": "1",
      "support_neg": "9",
      "tp_explicit": "1",
      "fp_explicit": "1",
      "fn_explicit": "0",
      "precision_explicit": "0.500000",
      "recall_explicit": "1.000000",
      "f1_explicit": "0.666667",
      "tp_strong": "1",
      "fp_strong": "1",
      "fn_strong": "0",
      "precision_strong": "0.500000",
      "recall_strong": "1.000000",
      "f1_strong": "0.666667"
    },
    {
      "tag": "simple_background",
      "bundle": "scene_pose",
      "needs_glossary": "0",
      "support_pos": "3",
      "support_neg": "7",
      "tp_explicit": "3",
      "fp_explicit": "4",
      "fn_explicit": "0",
      "precision_explicit": "0.428571",
      "recall_explicit": "1.000000",
      "f1_explicit": "0.600000",
      "tp_strong": "3",
      "fp_strong": "4",
      "fn_strong": "0",
      "precision_strong": "0.428571",
      "recall_strong": "1.000000",
      "f1_strong": "0.600000"
    },
    {
      "tag": "canid",
      "bundle": "species_taxonomy",
      "needs_glossary": "1",
      "support_pos": "3",
      "support_neg": "7",
      "tp_explicit": "2",
      "fp_explicit": "2",
      "fn_explicit": "1",
      "precision_explicit": "0.500000",
      "recall_explicit": "0.666667",
      "f1_explicit": "0.571429",
      "tp_strong": "2",
      "fp_strong": "2",
      "fn_strong": "1",
      "precision_strong": "0.500000",
      "recall_strong": "0.666667",
      "f1_strong": "0.571429"
    },
    {
      "tag": "looking_at_viewer",
      "bundle": "gaze_expression",
      "needs_glossary": "0",
      "support_pos": "3",
      "support_neg": "7",
      "tp_explicit": "1",
      "fp_explicit": "0",
      "fn_explicit": "2",
      "precision_explicit": "1.000000",
      "recall_explicit": "0.333333",
      "f1_explicit": "0.500000",
      "tp_strong": "1",
      "fp_strong": "0",
      "fn_strong": "2",
      "precision_strong": "1.000000",
      "recall_strong": "0.333333",
      "f1_strong": "0.500000"
    },
    {
      "tag": "standing",
      "bundle": "scene_pose",
      "needs_glossary": "0",
      "support_pos": "2",
      "support_neg": "8",
      "tp_explicit": "2",
      "fp_explicit": "4",
      "fn_explicit": "0",
      "precision_explicit": "0.333333",
      "recall_explicit": "1.000000",
      "f1_explicit": "0.500000",
      "tp_strong": "2",
      "fp_strong": "4",
      "fn_strong": "0",
      "precision_strong": "0.333333",
      "recall_strong": "1.000000",
      "f1_strong": "0.500000"
    },
    {
      "tag": "biped",
      "bundle": "body_type_presence",
      "needs_glossary": "0",
      "support_pos": "3",
      "support_neg": "7",
      "tp_explicit": "2",
      "fp_explicit": "4",
      "fn_explicit": "1",
      "precision_explicit": "0.333333",
      "recall_explicit": "0.666667",
      "f1_explicit": "0.444444",
      "tp_strong": "2",
      "fp_strong": "4",
      "fn_strong": "1",
      "precision_strong": "0.333333",
      "recall_strong": "0.666667",
      "f1_strong": "0.444444"
    }
  ],
  "outputs": {
    "csv": "E:\\image\\backup\\Prompt_Squirrel_RAG\\data\\analysis\\probe_reliability_sanity10.csv",
    "json": "E:\\image\\backup\\Prompt_Squirrel_RAG\\data\\analysis\\probe_reliability_sanity10.json"
  }
}