experiments_deep_baseline/reports/experiment_report.json

{
  "dataset_summary": {
    "dataset_id": "xTRam1/safe-guard-prompt-injection",
    "total_rows": 10135,
    "label_counts": {
      "0": 7066,
      "1": 3069
    },
    "split_sizes": {
      "train": 8108,
      "val": 1013,
      "test": 1014
    },
    "saved_files": {
      "train": "layer 2\\experiments_deep_baseline\\data\\xtram1_train.csv",
      "val": "layer 2\\experiments_deep_baseline\\data\\xtram1_val.csv",
      "test": "layer 2\\experiments_deep_baseline\\data\\xtram1_test.csv"
    }
  },
  "configs": [
    {
      "name": "config1_modernbert_large",
      "owner": "Deep",
      "model_name": "answerdotai/ModernBERT-large",
      "max_length": 384,
      "learning_rate": 1.5e-05,
      "per_device_batch_size": 16,
      "grad_accumulation": 4,
      "epochs": 4,
      "weight_decay": 0.01,
      "warmup_ratio": 0.1,
      "loss_type": "ce",
      "focal_gamma": 2.0,
      "label_smoothing": 0.03,
      "use_lora": false
    }
  ],
  "run_results": [
    {
      "run_id": "config1_modernbert_large_seed42",
      "config": {
        "name": "config1_modernbert_large",
        "owner": "Deep",
        "model_name": "answerdotai/ModernBERT-large",
        "max_length": 384,
        "learning_rate": 1.5e-05,
        "per_device_batch_size": 16,
        "grad_accumulation": 4,
        "epochs": 4,
        "weight_decay": 0.01,
        "warmup_ratio": 0.1,
        "loss_type": "ce",
        "focal_gamma": 2.0,
        "label_smoothing": 0.03,
        "use_lora": false
      },
      "seed": 42,
      "best_threshold": 0.475,
      "best_val_metrics": {
        "threshold": 0.475,
        "accuracy": 0.9990128331688055,
        "precision_injection": 1.0,
        "recall_injection": 0.996742671009772,
        "f1_injection": 0.99836867862969,
        "f1_macro": 0.9988304822730899,
        "pr_auc": 0.9999581050933732,
        "roc_auc": 0.9999815448782423,
        "fpr_safe": 0.0,
        "tp": 306,
        "tn": 706,
        "fp": 0,
        "fn": 1
      },
      "test_metrics": {
        "threshold": 0.475,
        "accuracy": 0.9990138067061144,
        "precision_injection": 0.9967532467532467,
        "recall_injection": 1.0,
        "f1_injection": 0.9983739837398374,
        "f1_macro": 0.9988331348281636,
        "pr_auc": 1.0,
        "roc_auc": 1.0,
        "fpr_safe": 0.0014144271570014145,
        "tp": 307,
        "tn": 706,
        "fp": 1,
        "fn": 0
      },
      "latency_ms_per_sample": 649.0914300084114,
      "train_time_sec": 133216.15459275246,
      "checkpoint_dir": "layer 2\\experiments_deep_baseline\\checkpoints\\config1_modernbert_large_seed42"
    }
  ],
  "aggregate": [
    {
      "config_name": "config1_modernbert_large",
      "owner": "Deep",
      "model_name": "answerdotai/ModernBERT-large",
      "num_runs": 1,
      "seeds": [
        42
      ],
      "f1_injection_mean": 0.9983739837398374,
      "f1_injection_std": 0.0,
      "pr_auc_mean": 1.0,
      "pr_auc_std": 0.0,
      "recall_injection_mean": 1.0,
      "recall_injection_std": 0.0,
      "precision_injection_mean": 0.9967532467532467,
      "precision_injection_std": 0.0,
      "fpr_safe_mean": 0.0014144271570014145,
      "fpr_safe_std": 0.0,
      "accuracy_mean": 0.9990138067061144,
      "accuracy_std": 0.0,
      "latency_ms_per_sample_mean": 649.0914300084114,
      "latency_ms_per_sample_std": 0.0,
      "train_time_sec_mean": 133216.15459275246,
      "train_time_sec_std": 0.0
    }
  ],
  "winner": {
    "config_name": "config1_modernbert_large",
    "owner": "Deep",
    "model_name": "answerdotai/ModernBERT-large",
    "num_runs": 1,
    "seeds": [
      42
    ],
    "f1_injection_mean": 0.9983739837398374,
    "f1_injection_std": 0.0,
    "pr_auc_mean": 1.0,
    "pr_auc_std": 0.0,
    "recall_injection_mean": 1.0,
    "recall_injection_std": 0.0,
    "precision_injection_mean": 0.9967532467532467,
    "precision_injection_std": 0.0,
    "fpr_safe_mean": 0.0014144271570014145,
    "fpr_safe_std": 0.0,
    "accuracy_mean": 0.9990138067061144,
    "accuracy_std": 0.0,
    "latency_ms_per_sample_mean": 649.0914300084114,
    "latency_ms_per_sample_std": 0.0,
    "train_time_sec_mean": 133216.15459275246,
    "train_time_sec_std": 0.0
  }
}