CLEF-DualScreen / redo-test-metrics-local.json

CLEF dual-task proper-protocol FT (seed 20261101, epoch 3) from jev-retrain MI355X run 2026-11-04

073de96 verified 2 days ago

531 Bytes

{
  "failed": 0,
  "by_kind": {
    "action_authorization_guard": {
      "n": 120,
      "pos": 60,
      "neg": 60,
      "threshold": 0.5,
      "auc": 0.7731,
      "tpr": 0.9667,
      "fpr": 0.5167,
      "balanced_accuracy": 0.725,
      "brier": 0.2489,
      "ece": 0.2582
    },
    "tool_output_filter": {
      "n": 100,
      "pos": 50,
      "neg": 50,
      "threshold": 0.5,
      "auc": 1.0,
      "tpr": 1.0,
      "fpr": 0.0,
      "balanced_accuracy": 1.0,
      "brier": 0.0027,
      "ece": 0.0118
    }
  }
}

	{
	"failed": 0,
	"by_kind": {
	"action_authorization_guard": {
	"n": 120,
	"pos": 60,
	"neg": 60,
	"threshold": 0.5,
	"auc": 0.7731,
	"tpr": 0.9667,
	"fpr": 0.5167,
	"balanced_accuracy": 0.725,
	"brier": 0.2489,
	"ece": 0.2582
	},
	"tool_output_filter": {
	"n": 100,
	"pos": 50,
	"neg": 50,
	"threshold": 0.5,
	"auc": 1.0,
	"tpr": 1.0,
	"fpr": 0.0,
	"balanced_accuracy": 1.0,
	"brier": 0.0027,
	"ece": 0.0118
	}
	}
	}