{
  "submission": "team3_HIPE-2026-v1.0-impresso-test-en_run1.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team3",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          64,
          3,
          8
        ],
        [
          21,
          0,
          7
        ],
        [
          20,
          6,
          33
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 64,
          "pred_PROBABLE": 3,
          "pred_FALSE": 8
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 21,
          "pred_PROBABLE": 0,
          "pred_FALSE": 7
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 20,
          "pred_PROBABLE": 6,
          "pred_FALSE": 33
        }
      ],
      "accuracy": 0.5988,
      "per_label": {
        "TRUE": {
          "precision": 0.6095,
          "recall": 0.8533,
          "f1": 0.7111,
          "support": 75,
          "true_positive": 64,
          "false_positive": 41,
          "false_negative": 11
        },
        "PROBABLE": {
          "precision": 0.0,
          "recall": 0.0,
          "f1": 0.0,
          "support": 28,
          "true_positive": 0,
          "false_positive": 9,
          "false_negative": 28
        },
        "FALSE": {
          "precision": 0.6875,
          "recall": 0.5593,
          "f1": 0.6168,
          "support": 59,
          "true_positive": 33,
          "false_positive": 15,
          "false_negative": 26
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          41,
          24
        ],
        [
          32,
          65
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 41,
          "pred_FALSE": 24
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 32,
          "pred_FALSE": 65
        }
      ],
      "accuracy": 0.6543,
      "per_label": {
        "TRUE": {
          "precision": 0.5616,
          "recall": 0.6308,
          "f1": 0.5942,
          "support": 65,
          "true_positive": 41,
          "false_positive": 32,
          "false_negative": 24
        },
        "FALSE": {
          "precision": 0.7303,
          "recall": 0.6701,
          "f1": 0.6989,
          "support": 97,
          "true_positive": 65,
          "false_positive": 24,
          "false_negative": 32
        }
      }
    }
  }
}