{
  "submission": "team1_HIPE-2026-v1.0-impresso-test-en_run3.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team1",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run3",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          47,
          25,
          3
        ],
        [
          9,
          13,
          6
        ],
        [
          5,
          12,
          42
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 47,
          "pred_PROBABLE": 25,
          "pred_FALSE": 3
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 9,
          "pred_PROBABLE": 13,
          "pred_FALSE": 6
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 5,
          "pred_PROBABLE": 12,
          "pred_FALSE": 42
        }
      ],
      "accuracy": 0.6296,
      "per_label": {
        "TRUE": {
          "precision": 0.7705,
          "recall": 0.6267,
          "f1": 0.6912,
          "support": 75,
          "true_positive": 47,
          "false_positive": 14,
          "false_negative": 28
        },
        "PROBABLE": {
          "precision": 0.26,
          "recall": 0.4643,
          "f1": 0.3333,
          "support": 28,
          "true_positive": 13,
          "false_positive": 37,
          "false_negative": 15
        },
        "FALSE": {
          "precision": 0.8235,
          "recall": 0.7119,
          "f1": 0.7636,
          "support": 59,
          "true_positive": 42,
          "false_positive": 9,
          "false_negative": 17
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          25,
          40
        ],
        [
          2,
          95
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 25,
          "pred_FALSE": 40
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 2,
          "pred_FALSE": 95
        }
      ],
      "accuracy": 0.7407,
      "per_label": {
        "TRUE": {
          "precision": 0.9259,
          "recall": 0.3846,
          "f1": 0.5435,
          "support": 65,
          "true_positive": 25,
          "false_positive": 2,
          "false_negative": 40
        },
        "FALSE": {
          "precision": 0.7037,
          "recall": 0.9794,
          "f1": 0.819,
          "support": 97,
          "true_positive": 95,
          "false_positive": 40,
          "false_negative": 2
        }
      }
    }
  }
}