metrics_summary.json

{
  "train_runtime": {
    "skipped_train": true
  },
  "eval_metrics": {
    "train": {
      "train_loss": 0.5018106698989868,
      "train_model_preparation_time": 0.0013,
      "train_accuracy": 0.9251333333333334,
      "train_macro_f1": 0.9246237953755961,
      "train_runtime": 90.3107,
      "train_samples_per_second": 332.186,
      "train_steps_per_second": 41.523,
      "epoch": 0
    },
    "validation": {
      "validation_loss": 0.709372878074646,
      "validation_model_preparation_time": 0.0013,
      "validation_accuracy": 0.835,
      "validation_macro_f1": 0.8351179604209936,
      "validation_runtime": 11.0597,
      "validation_samples_per_second": 325.505,
      "validation_steps_per_second": 40.688,
      "epoch": 0
    },
    "test_full": {
      "test_full_loss": 0.7299672365188599,
      "test_full_model_preparation_time": 0.0013,
      "test_full_accuracy": 0.8355555555555556,
      "test_full_macro_f1": 0.835052453623228,
      "test_full_runtime": 10.953,
      "test_full_samples_per_second": 328.676,
      "test_full_steps_per_second": 41.084,
      "epoch": 0
    },
    "test_title_only": {
      "test_title_only_loss": 1.0085811614990234,
      "test_title_only_model_preparation_time": 0.0013,
      "test_title_only_accuracy": 0.7522222222222222,
      "test_title_only_macro_f1": 0.7495277786217475,
      "test_title_only_runtime": 6.6335,
      "test_title_only_samples_per_second": 542.697,
      "test_title_only_steps_per_second": 67.837,
      "epoch": 0
    }
  },
  "test_full_classification_report": {
    "Artificial Intelligence": {
      "precision": 0.7289719626168224,
      "recall": 0.78,
      "f1-score": 0.7536231884057971,
      "support": 300.0
    },
    "Natural Language Processing": {
      "precision": 0.852760736196319,
      "recall": 0.9266666666666666,
      "f1-score": 0.8881789137380192,
      "support": 300.0
    },
    "Computer Vision": {
      "precision": 0.8379310344827586,
      "recall": 0.81,
      "f1-score": 0.823728813559322,
      "support": 300.0
    },
    "Machine Learning": {
      "precision": 0.6564885496183206,
      "recall": 0.5733333333333334,
      "f1-score": 0.6120996441281139,
      "support": 300.0
    },
    "CS Theory and Algorithms": {
      "precision": 0.8657718120805369,
      "recall": 0.86,
      "f1-score": 0.862876254180602,
      "support": 300.0
    },
    "Mathematics": {
      "precision": 0.9622641509433962,
      "recall": 0.85,
      "f1-score": 0.9026548672566371,
      "support": 300.0
    },
    "Statistics": {
      "precision": 0.7125748502994012,
      "recall": 0.7933333333333333,
      "f1-score": 0.750788643533123,
      "support": 300.0
    },
    "Electrical Engineering": {
      "precision": 0.7736486486486487,
      "recall": 0.7633333333333333,
      "f1-score": 0.7684563758389261,
      "support": 300.0
    },
    "Astrophysics": {
      "precision": 0.9966329966329966,
      "recall": 0.9866666666666667,
      "f1-score": 0.9916247906197655,
      "support": 300.0
    },
    "Condensed Matter Physics": {
      "precision": 0.916083916083916,
      "recall": 0.8733333333333333,
      "f1-score": 0.89419795221843,
      "support": 300.0
    },
    "Quantum Physics": {
      "precision": 0.8676923076923077,
      "recall": 0.94,
      "f1-score": 0.9024,
      "support": 300.0
    },
    "Quantitative Biology": {
      "precision": 0.87,
      "recall": 0.87,
      "f1-score": 0.87,
      "support": 300.0
    },
    "accuracy": 0.8355555555555556,
    "macro avg": {
      "precision": 0.8367350804412852,
      "recall": 0.8355555555555555,
      "f1-score": 0.835052453623228,
      "support": 3600.0
    },
    "weighted avg": {
      "precision": 0.8367350804412854,
      "recall": 0.8355555555555556,
      "f1-score": 0.835052453623228,
      "support": 3600.0
    }
  },
  "test_full_confusion_matrix": [
    [
      234,
      15,
      4,
      22,
      11,
      0,
      10,
      0,
      0,
      0,
      2,
      2
    ],
    [
      9,
      278,
      5,
      4,
      0,
      0,
      1,
      1,
      0,
      0,
      0,
      2
    ],
    [
      3,
      5,
      243,
      15,
      0,
      0,
      1,
      31,
      0,
      0,
      0,
      2
    ],
    [
      32,
      14,
      19,
      172,
      8,
      1,
      40,
      9,
      0,
      0,
      0,
      5
    ],
    [
      25,
      1,
      0,
      1,
      258,
      2,
      3,
      2,
      0,
      2,
      5,
      1
    ],
    [
      1,
      0,
      0,
      3,
      12,
      255,
      13,
      6,
      0,
      3,
      3,
      4
    ],
    [
      4,
      3,
      0,
      31,
      2,
      0,
      238,
      8,
      0,
      0,
      1,
      13
    ],
    [
      6,
      7,
      18,
      8,
      5,
      5,
      15,
      229,
      0,
      1,
      0,
      6
    ],
    [
      1,
      0,
      1,
      0,
      0,
      0,
      0,
      1,
      296,
      0,
      1,
      0
    ],
    [
      0,
      0,
      0,
      1,
      1,
      0,
      1,
      0,
      1,
      262,
      30,
      4
    ],
    [
      0,
      0,
      0,
      2,
      0,
      0,
      0,
      1,
      0,
      15,
      282,
      0
    ],
    [
      6,
      3,
      0,
      3,
      1,
      2,
      12,
      8,
      0,
      3,
      1,
      261
    ]
  ],
  "test_title_only_classification_report": {
    "Artificial Intelligence": {
      "precision": 0.6678321678321678,
      "recall": 0.6366666666666667,
      "f1-score": 0.6518771331058021,
      "support": 300.0
    },
    "Natural Language Processing": {
      "precision": 0.7485207100591716,
      "recall": 0.8433333333333334,
      "f1-score": 0.7931034482758621,
      "support": 300.0
    },
    "Computer Vision": {
      "precision": 0.7210884353741497,
      "recall": 0.7066666666666667,
      "f1-score": 0.7138047138047138,
      "support": 300.0
    },
    "Machine Learning": {
      "precision": 0.5879828326180258,
      "recall": 0.45666666666666667,
      "f1-score": 0.5140712945590994,
      "support": 300.0
    },
    "CS Theory and Algorithms": {
      "precision": 0.7033639143730887,
      "recall": 0.7666666666666667,
      "f1-score": 0.733652312599681,
      "support": 300.0
    },
    "Mathematics": {
      "precision": 0.837037037037037,
      "recall": 0.7533333333333333,
      "f1-score": 0.7929824561403509,
      "support": 300.0
    },
    "Statistics": {
      "precision": 0.6128048780487805,
      "recall": 0.67,
      "f1-score": 0.6401273885350318,
      "support": 300.0
    },
    "Electrical Engineering": {
      "precision": 0.7295373665480427,
      "recall": 0.6833333333333333,
      "f1-score": 0.7056798623063684,
      "support": 300.0
    },
    "Astrophysics": {
      "precision": 0.9534883720930233,
      "recall": 0.9566666666666667,
      "f1-score": 0.9550748752079867,
      "support": 300.0
    },
    "Condensed Matter Physics": {
      "precision": 0.8430034129692833,
      "recall": 0.8233333333333334,
      "f1-score": 0.8330522765598651,
      "support": 300.0
    },
    "Quantum Physics": {
      "precision": 0.7834757834757835,
      "recall": 0.9166666666666666,
      "f1-score": 0.8448540706605223,
      "support": 300.0
    },
    "Quantitative Biology": {
      "precision": 0.8187919463087249,
      "recall": 0.8133333333333334,
      "f1-score": 0.8160535117056856,
      "support": 300.0
    },
    "accuracy": 0.7522222222222222,
    "macro avg": {
      "precision": 0.7505772380614398,
      "recall": 0.7522222222222222,
      "f1-score": 0.7495277786217475,
      "support": 3600.0
    },
    "weighted avg": {
      "precision": 0.7505772380614398,
      "recall": 0.7522222222222222,
      "f1-score": 0.7495277786217475,
      "support": 3600.0
    }
  },
  "test_title_only_confusion_matrix": [
    [
      191,
      30,
      10,
      19,
      18,
      2,
      17,
      3,
      1,
      1,
      5,
      3
    ],
    [
      12,
      253,
      11,
      13,
      0,
      0,
      5,
      1,
      0,
      1,
      1,
      3
    ],
    [
      6,
      11,
      212,
      19,
      5,
      2,
      3,
      37,
      2,
      0,
      1,
      2
    ],
    [
      21,
      22,
      28,
      137,
      21,
      1,
      48,
      9,
      1,
      2,
      1,
      9
    ],
    [
      24,
      4,
      0,
      6,
      230,
      10,
      12,
      4,
      0,
      1,
      7,
      2
    ],
    [
      4,
      0,
      2,
      3,
      25,
      226,
      11,
      7,
      1,
      11,
      5,
      5
    ],
    [
      11,
      4,
      2,
      28,
      11,
      9,
      201,
      10,
      1,
      5,
      2,
      16
    ],
    [
      3,
      8,
      24,
      4,
      9,
      8,
      16,
      205,
      3,
      3,
      4,
      13
    ],
    [
      2,
      0,
      3,
      0,
      0,
      3,
      0,
      0,
      287,
      3,
      2,
      0
    ],
    [
      1,
      0,
      0,
      0,
      0,
      5,
      1,
      0,
      2,
      247,
      43,
      1
    ],
    [
      1,
      1,
      0,
      1,
      5,
      2,
      2,
      0,
      2,
      11,
      275,
      0
    ],
    [
      10,
      5,
      2,
      3,
      3,
      2,
      12,
      5,
      1,
      8,
      5,
      244
    ]
  ]
}