forgetting_action_policy_metrics_eval.json

{
  "family": "router",
  "split": "eval",
  "overall": {
    "rows": 5000,
    "accuracy": 0.984,
    "macro_f1": 0.9839922007228425,
    "weighted_f1": 0.9839922007228427,
    "micro_f1": 0.984,
    "labels": [
      "forgetting_action_policy::compress",
      "forgetting_action_policy::decay",
      "forgetting_action_policy::delete",
      "forgetting_action_policy::keep",
      "forgetting_action_policy::silence"
    ],
    "confusion_matrix": [
      [
        988,
        3,
        0,
        9,
        0
      ],
      [
        3,
        986,
        1,
        1,
        9
      ],
      [
        2,
        0,
        994,
        0,
        4
      ],
      [
        6,
        11,
        0,
        979,
        4
      ],
      [
        2,
        11,
        11,
        3,
        973
      ]
    ],
    "calibration_error": 0.0073812406567217975,
    "classification_report": {
      "forgetting_action_policy::compress": {
        "precision": 0.987012987012987,
        "recall": 0.988,
        "f1-score": 0.9875062468765617,
        "support": 1000.0
      },
      "forgetting_action_policy::decay": {
        "precision": 0.9752720079129574,
        "recall": 0.986,
        "f1-score": 0.9806066633515664,
        "support": 1000.0
      },
      "forgetting_action_policy::delete": {
        "precision": 0.9880715705765407,
        "recall": 0.994,
        "f1-score": 0.9910269192422732,
        "support": 1000.0
      },
      "forgetting_action_policy::keep": {
        "precision": 0.9868951612903226,
        "recall": 0.979,
        "f1-score": 0.9829317269076305,
        "support": 1000.0
      },
      "forgetting_action_policy::silence": {
        "precision": 0.9828282828282828,
        "recall": 0.973,
        "f1-score": 0.9778894472361809,
        "support": 1000.0
      },
      "accuracy": 0.984,
      "macro avg": {
        "precision": 0.984016001924218,
        "recall": 0.984,
        "f1-score": 0.9839922007228425,
        "support": 5000.0
      },
      "weighted avg": {
        "precision": 0.9840160019242181,
        "recall": 0.984,
        "f1-score": 0.9839922007228427,
        "support": 5000.0
      }
    }
  },
  "per_task": {
    "forgetting_action_policy": {
      "rows": 5000,
      "accuracy": 0.984,
      "macro_f1": 0.9839922007228425,
      "weighted_f1": 0.9839922007228427,
      "micro_f1": 0.984,
      "labels": [
        "compress",
        "decay",
        "delete",
        "keep",
        "silence"
      ],
      "confusion_matrix": [
        [
          988,
          3,
          0,
          9,
          0
        ],
        [
          3,
          986,
          1,
          1,
          9
        ],
        [
          2,
          0,
          994,
          0,
          4
        ],
        [
          6,
          11,
          0,
          979,
          4
        ],
        [
          2,
          11,
          11,
          3,
          973
        ]
      ],
      "wrong_task_predictions": 0,
      "wrong_task_rate": 0.0
    }
  },
  "calibration": {
    "method": "temperature_grid_search",
    "rows": 5000,
    "temperature": 2.0,
    "loss": 0.08134587520392533
  }
}