test_predictions_sample.json

{
  "accuracy": 0.3593073593073593,
  "precision": 0.42672033848504437,
  "recall": 0.3593073593073593,
  "f1": 0.33159731214306154,
  "total_training_steps": 500,
  "early_stopped": false,
  "label_map": {
    "0": "SUPPORTS",
    "1": "REFUTES",
    "2": "NOT_ENOUGH_INFO",
    "3": "DISPUTED"
  },
  "samples": [
    {
      "prediction": "SUPPORTS",
      "ground_truth": "NOT_ENOUGH_INFO"
    },
    {
      "prediction": "REFUTES",
      "ground_truth": "NOT_ENOUGH_INFO"
    },
    {
      "prediction": "REFUTES",
      "ground_truth": "SUPPORTS"
    },
    {
      "prediction": "REFUTES",
      "ground_truth": "NOT_ENOUGH_INFO"
    },
    {
      "prediction": "REFUTES",
      "ground_truth": "SUPPORTS"
    },
    {
      "prediction": "REFUTES",
      "ground_truth": "NOT_ENOUGH_INFO"
    },
    {
      "prediction": "SUPPORTS",
      "ground_truth": "DISPUTED"
    },
    {
      "prediction": "REFUTES",
      "ground_truth": "SUPPORTS"
    },
    {
      "prediction": "SUPPORTS",
      "ground_truth": "REFUTES"
    },
    {
      "prediction": "SUPPORTS",
      "ground_truth": "SUPPORTS"
    }
  ]
}