reports/ev/evaluation.json

{
  "model": "ml/models/stutter/stutter_lora",
  "base_model": "facebook/wav2vec2-base",
  "adapter": "LoRA (r=8, alpha=16, target q/k/v)",
  "n_train": 7486,
  "n_val": 1174,
  "n_test": 1666,
  "split": "by-speaker (test voices never seen in training)",
  "binary": true,
  "threshold": 0.5,
  "accuracy": 0.7149,
  "macro_f1": 0.7147,
  "roc_auc": 0.7605,
  "per_class": {
    "fluent": {
      "precision": 0.7455,
      "recall": 0.6749,
      "f1": 0.7084
    },
    "stutter": {
      "precision": 0.6883,
      "recall": 0.7571,
      "f1": 0.7211
    }
  },
  "confusion_matrix": [
    [
      577,
      278
    ],
    [
      197,
      614
    ]
  ],
  "class_map": {
    "0": "fluent",
    "1": "stutter"
  },
  "metric_definitions": {
    "accuracy": "correct / total on out-of-speaker test set",
    "precision": "class TP / (TP+FP)",
    "recall": "class TP / (TP+FN)",
    "macro_f1": "mean of per-class F1",
    "roc_auc": "area under ROC curve"
  }
}