TRM_model_3 / eval_metrics.json
admincybers2's picture
F1 0.7639 @ 2026-07-07T07:27:10
176fe8e verified
Raw
History Blame Contribute Delete
18 kB
{
"threshold": 0.961,
"micro_precision": 0.8462,
"micro_recall": 0.6962,
"micro_f1": 0.7639,
"macro_f1": 0.5275,
"exact_set_match": 0.2,
"avg_selected": 4.33,
"avg_true_pass": 5.27,
"num_groups": 45,
"best_epoch": 46,
"history": [
{
"epoch": 1,
"loss": 135.681,
"lr": 0.000113,
"threshold": 0.671,
"micro_precision": 0.4943,
"micro_recall": 0.1814,
"micro_f1": 0.2654,
"macro_f1": 0.3522,
"exact_set_match": 0.0444,
"avg_selected": 1.93,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 2,
"loss": 111.946,
"lr": 0.000113,
"threshold": 0.749,
"micro_precision": 0.4458,
"micro_recall": 0.1561,
"micro_f1": 0.2313,
"macro_f1": 0.2482,
"exact_set_match": 0.0444,
"avg_selected": 1.84,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 3,
"loss": 105.809,
"lr": 0.000113,
"threshold": 0.766,
"micro_precision": 0.5484,
"micro_recall": 0.1435,
"micro_f1": 0.2274,
"macro_f1": 0.2573,
"exact_set_match": 0.0667,
"avg_selected": 1.38,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 4,
"loss": 103.093,
"lr": 0.000113,
"threshold": 0.782,
"micro_precision": 0.6479,
"micro_recall": 0.1941,
"micro_f1": 0.2987,
"macro_f1": 0.2878,
"exact_set_match": 0.0889,
"avg_selected": 1.58,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 5,
"loss": 100.489,
"lr": 0.000113,
"threshold": 0.854,
"micro_precision": 0.6753,
"micro_recall": 0.2194,
"micro_f1": 0.3312,
"macro_f1": 0.3017,
"exact_set_match": 0.0889,
"avg_selected": 1.71,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 6,
"loss": 97.972,
"lr": 0.000113,
"threshold": 0.833,
"micro_precision": 0.597,
"micro_recall": 0.1688,
"micro_f1": 0.2632,
"macro_f1": 0.284,
"exact_set_match": 0.0667,
"avg_selected": 1.49,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 7,
"loss": 98.35,
"lr": 0.000113,
"threshold": 0.854,
"micro_precision": 0.5882,
"micro_recall": 0.1688,
"micro_f1": 0.2623,
"macro_f1": 0.2865,
"exact_set_match": 0.0667,
"avg_selected": 1.51,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 8,
"loss": 95.831,
"lr": 0.000113,
"threshold": 0.858,
"micro_precision": 0.6556,
"micro_recall": 0.2489,
"micro_f1": 0.3609,
"macro_f1": 0.3492,
"exact_set_match": 0.0889,
"avg_selected": 2.0,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 9,
"loss": 94.704,
"lr": 0.000113,
"threshold": 0.896,
"micro_precision": 0.6237,
"micro_recall": 0.2447,
"micro_f1": 0.3515,
"macro_f1": 0.3456,
"exact_set_match": 0.0667,
"avg_selected": 2.07,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 10,
"loss": 94.165,
"lr": 0.000113,
"threshold": 0.83,
"micro_precision": 0.6774,
"micro_recall": 0.3544,
"micro_f1": 0.4654,
"macro_f1": 0.401,
"exact_set_match": 0.0667,
"avg_selected": 2.76,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 11,
"loss": 93.853,
"lr": 0.000113,
"threshold": 0.859,
"micro_precision": 0.6355,
"micro_recall": 0.2869,
"micro_f1": 0.3953,
"macro_f1": 0.4353,
"exact_set_match": 0.0667,
"avg_selected": 2.38,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 12,
"loss": 91.204,
"lr": 0.000113,
"threshold": 0.791,
"micro_precision": 0.688,
"micro_recall": 0.3629,
"micro_f1": 0.4751,
"macro_f1": 0.4717,
"exact_set_match": 0.1333,
"avg_selected": 2.78,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 13,
"loss": 90.704,
"lr": 0.000113,
"threshold": 0.817,
"micro_precision": 0.6129,
"micro_recall": 0.2405,
"micro_f1": 0.3455,
"macro_f1": 0.409,
"exact_set_match": 0.0889,
"avg_selected": 2.07,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 14,
"loss": 90.188,
"lr": 0.000113,
"threshold": 0.87,
"micro_precision": 0.6867,
"micro_recall": 0.4346,
"micro_f1": 0.5323,
"macro_f1": 0.5023,
"exact_set_match": 0.1111,
"avg_selected": 3.33,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 15,
"loss": 88.848,
"lr": 0.000113,
"threshold": 0.855,
"micro_precision": 0.6455,
"micro_recall": 0.2996,
"micro_f1": 0.4092,
"macro_f1": 0.4486,
"exact_set_match": 0.1111,
"avg_selected": 2.44,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 16,
"loss": 88.25,
"lr": 0.000113,
"threshold": 0.871,
"micro_precision": 0.7345,
"micro_recall": 0.3502,
"micro_f1": 0.4743,
"macro_f1": 0.4409,
"exact_set_match": 0.1556,
"avg_selected": 2.51,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 17,
"loss": 88.068,
"lr": 0.000113,
"threshold": 0.852,
"micro_precision": 0.7477,
"micro_recall": 0.3376,
"micro_f1": 0.4651,
"macro_f1": 0.4568,
"exact_set_match": 0.1333,
"avg_selected": 2.38,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 18,
"loss": 86.88,
"lr": 0.000113,
"threshold": 0.89,
"micro_precision": 0.7236,
"micro_recall": 0.3755,
"micro_f1": 0.4944,
"macro_f1": 0.42,
"exact_set_match": 0.0889,
"avg_selected": 2.73,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 19,
"loss": 83.995,
"lr": 5.65e-05,
"threshold": 0.861,
"micro_precision": 0.75,
"micro_recall": 0.4557,
"micro_f1": 0.5669,
"macro_f1": 0.5345,
"exact_set_match": 0.1778,
"avg_selected": 3.2,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 20,
"loss": 83.192,
"lr": 5.65e-05,
"threshold": 0.889,
"micro_precision": 0.7483,
"micro_recall": 0.4768,
"micro_f1": 0.5825,
"macro_f1": 0.5166,
"exact_set_match": 0.1333,
"avg_selected": 3.36,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 21,
"loss": 82.146,
"lr": 5.65e-05,
"threshold": 0.879,
"micro_precision": 0.75,
"micro_recall": 0.4304,
"micro_f1": 0.5469,
"macro_f1": 0.5489,
"exact_set_match": 0.1778,
"avg_selected": 3.02,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 22,
"loss": 81.531,
"lr": 5.65e-05,
"threshold": 0.861,
"micro_precision": 0.8054,
"micro_recall": 0.5063,
"micro_f1": 0.6218,
"macro_f1": 0.5669,
"exact_set_match": 0.1778,
"avg_selected": 3.31,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 23,
"loss": 80.421,
"lr": 5.65e-05,
"threshold": 0.902,
"micro_precision": 0.8092,
"micro_recall": 0.5907,
"micro_f1": 0.6829,
"macro_f1": 0.5501,
"exact_set_match": 0.2,
"avg_selected": 3.84,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 24,
"loss": 79.635,
"lr": 5.65e-05,
"threshold": 0.875,
"micro_precision": 0.8506,
"micro_recall": 0.5527,
"micro_f1": 0.6701,
"macro_f1": 0.5378,
"exact_set_match": 0.2222,
"avg_selected": 3.42,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 25,
"loss": 78.996,
"lr": 5.65e-05,
"threshold": 0.884,
"micro_precision": 0.7889,
"micro_recall": 0.5992,
"micro_f1": 0.6811,
"macro_f1": 0.5418,
"exact_set_match": 0.1778,
"avg_selected": 4.0,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 26,
"loss": 77.775,
"lr": 5.65e-05,
"threshold": 0.871,
"micro_precision": 0.8148,
"micro_recall": 0.6498,
"micro_f1": 0.723,
"macro_f1": 0.5178,
"exact_set_match": 0.1778,
"avg_selected": 4.2,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 27,
"loss": 76.496,
"lr": 5.65e-05,
"threshold": 0.907,
"micro_precision": 0.8343,
"micro_recall": 0.6371,
"micro_f1": 0.7225,
"macro_f1": 0.5063,
"exact_set_match": 0.2,
"avg_selected": 4.02,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 28,
"loss": 75.752,
"lr": 5.65e-05,
"threshold": 0.905,
"micro_precision": 0.8432,
"micro_recall": 0.6582,
"micro_f1": 0.7393,
"macro_f1": 0.5243,
"exact_set_match": 0.2,
"avg_selected": 4.11,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 29,
"loss": 74.48,
"lr": 5.65e-05,
"threshold": 0.844,
"micro_precision": 0.8462,
"micro_recall": 0.6034,
"micro_f1": 0.7044,
"macro_f1": 0.5146,
"exact_set_match": 0.2222,
"avg_selected": 3.76,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 30,
"loss": 74.617,
"lr": 5.65e-05,
"threshold": 0.896,
"micro_precision": 0.8281,
"micro_recall": 0.6709,
"micro_f1": 0.7413,
"macro_f1": 0.5648,
"exact_set_match": 0.2,
"avg_selected": 4.27,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 31,
"loss": 73.126,
"lr": 5.65e-05,
"threshold": 0.906,
"micro_precision": 0.8146,
"micro_recall": 0.7046,
"micro_f1": 0.7557,
"macro_f1": 0.5775,
"exact_set_match": 0.1778,
"avg_selected": 4.56,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 32,
"loss": 71.328,
"lr": 5.65e-05,
"threshold": 0.903,
"micro_precision": 0.8485,
"micro_recall": 0.5907,
"micro_f1": 0.6965,
"macro_f1": 0.544,
"exact_set_match": 0.2222,
"avg_selected": 3.67,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 33,
"loss": 70.839,
"lr": 5.65e-05,
"threshold": 0.881,
"micro_precision": 0.8743,
"micro_recall": 0.6456,
"micro_f1": 0.7427,
"macro_f1": 0.5097,
"exact_set_match": 0.2222,
"avg_selected": 3.89,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 34,
"loss": 68.553,
"lr": 5.65e-05,
"threshold": 0.907,
"micro_precision": 0.8454,
"micro_recall": 0.692,
"micro_f1": 0.761,
"macro_f1": 0.542,
"exact_set_match": 0.2,
"avg_selected": 4.31,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 35,
"loss": 67.289,
"lr": 5.65e-05,
"threshold": 0.9,
"micro_precision": 0.8402,
"micro_recall": 0.6878,
"micro_f1": 0.7564,
"macro_f1": 0.5563,
"exact_set_match": 0.2444,
"avg_selected": 4.31,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 36,
"loss": 64.539,
"lr": 5.65e-05,
"threshold": 0.936,
"micro_precision": 0.8137,
"micro_recall": 0.7004,
"micro_f1": 0.7528,
"macro_f1": 0.5409,
"exact_set_match": 0.2,
"avg_selected": 4.53,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 37,
"loss": 62.601,
"lr": 5.65e-05,
"threshold": 0.916,
"micro_precision": 0.8533,
"micro_recall": 0.6624,
"micro_f1": 0.7458,
"macro_f1": 0.5639,
"exact_set_match": 0.1778,
"avg_selected": 4.09,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 38,
"loss": 61.007,
"lr": 5.65e-05,
"threshold": 0.923,
"micro_precision": 0.8791,
"micro_recall": 0.6751,
"micro_f1": 0.7637,
"macro_f1": 0.535,
"exact_set_match": 0.2222,
"avg_selected": 4.04,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 39,
"loss": 59.629,
"lr": 5.65e-05,
"threshold": 0.929,
"micro_precision": 0.8394,
"micro_recall": 0.6835,
"micro_f1": 0.7535,
"macro_f1": 0.5425,
"exact_set_match": 0.2,
"avg_selected": 4.29,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 40,
"loss": 57.058,
"lr": 5.65e-05,
"threshold": 0.92,
"micro_precision": 0.8735,
"micro_recall": 0.6118,
"micro_f1": 0.7196,
"macro_f1": 0.5007,
"exact_set_match": 0.2222,
"avg_selected": 3.69,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 41,
"loss": 56.509,
"lr": 5.65e-05,
"threshold": 0.931,
"micro_precision": 0.8729,
"micro_recall": 0.6667,
"micro_f1": 0.756,
"macro_f1": 0.4855,
"exact_set_match": 0.1333,
"avg_selected": 4.02,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 42,
"loss": 53.023,
"lr": 5.65e-05,
"threshold": 0.939,
"micro_precision": 0.8342,
"micro_recall": 0.6582,
"micro_f1": 0.7358,
"macro_f1": 0.5039,
"exact_set_match": 0.2,
"avg_selected": 4.16,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 43,
"loss": 46.743,
"lr": 2.825e-05,
"threshold": 0.953,
"micro_precision": 0.8548,
"micro_recall": 0.6709,
"micro_f1": 0.7518,
"macro_f1": 0.5195,
"exact_set_match": 0.2222,
"avg_selected": 4.13,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 44,
"loss": 43.895,
"lr": 2.825e-05,
"threshold": 0.947,
"micro_precision": 0.8743,
"micro_recall": 0.6751,
"micro_f1": 0.7619,
"macro_f1": 0.5243,
"exact_set_match": 0.2222,
"avg_selected": 4.07,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 45,
"loss": 41.772,
"lr": 2.825e-05,
"threshold": 0.963,
"micro_precision": 0.849,
"micro_recall": 0.6878,
"micro_f1": 0.7599,
"macro_f1": 0.5172,
"exact_set_match": 0.1778,
"avg_selected": 4.27,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 46,
"loss": 38.945,
"lr": 2.825e-05,
"threshold": 0.961,
"micro_precision": 0.8462,
"micro_recall": 0.6962,
"micro_f1": 0.7639,
"macro_f1": 0.5275,
"exact_set_match": 0.2,
"avg_selected": 4.33,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 47,
"loss": 37.698,
"lr": 2.825e-05,
"threshold": 0.954,
"micro_precision": 0.8482,
"micro_recall": 0.6835,
"micro_f1": 0.757,
"macro_f1": 0.514,
"exact_set_match": 0.2,
"avg_selected": 4.24,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 48,
"loss": 37.103,
"lr": 2.825e-05,
"threshold": 0.966,
"micro_precision": 0.8462,
"micro_recall": 0.6498,
"micro_f1": 0.7351,
"macro_f1": 0.4917,
"exact_set_match": 0.1556,
"avg_selected": 4.04,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 49,
"loss": 35.4,
"lr": 2.825e-05,
"threshold": 0.965,
"micro_precision": 0.8205,
"micro_recall": 0.6751,
"micro_f1": 0.7407,
"macro_f1": 0.5021,
"exact_set_match": 0.1778,
"avg_selected": 4.33,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 50,
"loss": 35.284,
"lr": 2.825e-05,
"threshold": 0.961,
"micro_precision": 0.8342,
"micro_recall": 0.6793,
"micro_f1": 0.7488,
"macro_f1": 0.5204,
"exact_set_match": 0.2,
"avg_selected": 4.29,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 51,
"loss": 30.252,
"lr": 1.4125e-05,
"threshold": 0.972,
"micro_precision": 0.8109,
"micro_recall": 0.6878,
"micro_f1": 0.7443,
"macro_f1": 0.5116,
"exact_set_match": 0.2222,
"avg_selected": 4.47,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 52,
"loss": 27.355,
"lr": 1.4125e-05,
"threshold": 0.977,
"micro_precision": 0.8098,
"micro_recall": 0.7004,
"micro_f1": 0.7511,
"macro_f1": 0.5299,
"exact_set_match": 0.2222,
"avg_selected": 4.56,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 53,
"loss": 26.756,
"lr": 1.4125e-05,
"threshold": 0.976,
"micro_precision": 0.8,
"micro_recall": 0.692,
"micro_f1": 0.7421,
"macro_f1": 0.5113,
"exact_set_match": 0.2,
"avg_selected": 4.56,
"avg_true_pass": 5.27,
"num_groups": 45
},
{
"epoch": 54,
"loss": 26.683,
"lr": 1.4125e-05,
"threshold": 0.977,
"micro_precision": 0.7929,
"micro_recall": 0.6624,
"micro_f1": 0.7218,
"macro_f1": 0.5011,
"exact_set_match": 0.1556,
"avg_selected": 4.4,
"avg_true_pass": 5.27,
"num_groups": 45
}
]
}