{
  "submission": "baseline_HIPE-2026-v1.0-surprise-test-fr_run1.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "baseline",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          84,
          20,
          26
        ],
        [
          23,
          18,
          19
        ],
        [
          75,
          49,
          166
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 84,
          "pred_PROBABLE": 20,
          "pred_FALSE": 26
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 23,
          "pred_PROBABLE": 18,
          "pred_FALSE": 19
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 75,
          "pred_PROBABLE": 49,
          "pred_FALSE": 166
        }
      ],
      "accuracy": 0.5583,
      "per_label": {
        "TRUE": {
          "precision": 0.4615,
          "recall": 0.6462,
          "f1": 0.5385,
          "support": 130,
          "true_positive": 84,
          "false_positive": 98,
          "false_negative": 46
        },
        "PROBABLE": {
          "precision": 0.2069,
          "recall": 0.3,
          "f1": 0.2449,
          "support": 60,
          "true_positive": 18,
          "false_positive": 69,
          "false_negative": 42
        },
        "FALSE": {
          "precision": 0.7867,
          "recall": 0.5724,
          "f1": 0.6627,
          "support": 290,
          "true_positive": 166,
          "false_positive": 45,
          "false_negative": 124
        }
      }
    }
  }
}