{
  "d1": {
    "resumed_adapter_start": {
      "n": 19,
      "same": 17,
      "auc": 0.6470588235294117,
      "average_precision": 0.948279640036606,
      "accuracy": 0.631578947368421
    },
    "max_tokens": 4096,
    "fine_tuned": {
      "n": 19,
      "same": 17,
      "auc": 0.6029411764705882,
      "average_precision": 0.933293645638847,
      "accuracy": 0.42105263157894735
    },
    "best_validation_loss": 0.0001904141332488507,
    "warning": "19 human pairs: insufficient for deployment decision; test never used for model selection"
  },
  "bge": {
    "max_tokens": 4096,
    "baseline": {
      "n": 19,
      "auc": 0.7205882352941176,
      "average_precision": 0.9542793567447547
    },
    "fine_tuned": {
      "n": 19,
      "auc": 0.6470588235294118,
      "average_precision": 0.944002438960097
    },
    "best_validation_loss": 0.04328301633649971
  },
  "production_ready": false,
  "limitations": [
    "19 human pairs, only 2 negatives",
    "weak-label training and validation",
    "Both models warm-started from saved adapters with reset optimizer state"
  ]
}