{
  "model": "DistilBERT (uncased) fine-tuned, last 2 blocks + linear head",
  "task": "binary-text-classification",
  "dataset": "Enron-Spam (SetFit/enron_spam), cleaned, stratified 70/15/15 split",
  "split": "test",
  "primary_metric": {
    "name": "f1",
    "value": 0.9937743190661479
  },
  "metrics": {
    "accuracy": 0.9936633663366337,
    "precision": 0.9930015552099534,
    "recall": 0.9945482866043613,
    "f1": 0.9937743190661479,
    "roc_auc": 0.9997530506249356,
    "average_precision": 0.9997640777977358
  },
  "threshold": 0.5,
  "data": {
    "n_train": 23565,
    "n_val": 5050,
    "n_test": 5050,
    "spam_share_test": 0.5085,
    "n_classes": 2
  },
  "params": 66363649,
  "train_time_s": null,
  "device": "cuda",
  "smoke": false,
  "source": "converted from the original training run (first version of this project, train.py on a CUDA GPU with AMP, checkpoint of 2026-07-13, best epoch 5); re-evaluated 2026-09-25 on CPU with model.Predictor",
  "versions": {
    "torch": "2.14.0+cpu",
    "transformers": "5.17.0",
    "tokenizers": "0.23.2",
    "huggingface_hub": "1.32.0",
    "safetensors": "0.8.0",
    "sklearn": "1.9.1",
    "numpy": "2.5.3"
  },
  "trained_at": "2026-07-13",
  "confusion_matrix": {
    "labels": [
      "ham",
      "spam"
    ],
    "rows": "true",
    "cols": "predicted",
    "counts": [
      [
        2464,
        18
      ],
      [
        14,
        2554
      ]
    ]
  },
  "comparison": {
    "DistilBERT fine-tuned (deployed)": {
      "accuracy": 0.9936633663366337,
      "precision": 0.9930015552099534,
      "recall": 0.9945482866043613,
      "f1": 0.9937743190661479,
      "roc_auc": 0.9997530506249356,
      "average_precision": 0.9997640777977358,
      "params": 66363649,
      "errors": 32
    },
    "TF-IDF + logistic regression": {
      "accuracy": 0.9912871287128713,
      "precision": 0.9880123743232792,
      "recall": 0.9949376947040498,
      "f1": 0.9914629414047342,
      "roc_auc": 0.9992293422297865,
      "average_precision": 0.999197846216229,
      "features": 50000,
      "train_time_s": 25.1,
      "errors": 44
    }
  },
  "history": {
    "train_loss": [
      0.1758,
      0.0276,
      0.0151,
      0.0089,
      0.0056
    ],
    "val_loss": [
      0.0292,
      0.0196,
      0.0171,
      0.0177,
      0.0166
    ],
    "val_acc": [
      0.9901,
      0.9931,
      0.9941,
      0.9941,
      0.9943
    ],
    "val_f1": [
      0.9903,
      0.9932,
      0.9942,
      0.9942,
      0.9944
    ],
    "best_epoch": 5,
    "source": "log of the original training run (CUDA + AMP)"
  },
  "trainable_params": 14176513,
  "eval_time_s_cpu": 1005.0,
  "legacy_reported": {
    "accuracy": 0.9937,
    "precision": 0.993,
    "recall": 0.9945,
    "f1": 0.9938,
    "confusion": [
      [
        2464,
        18
      ],
      [
        14,
        2554
      ]
    ]
  },
  "evaluated_at": "2026-09-25",
  "eval_note": "CPU time for the 5,050 test emails on a shared 20-core desktop (6 threads); an earlier run under heavier load took 2452 s"
}