test_v2.json

{
  "model": "o1sound_multiclass_v2",
  "recipe": "40 epochs, hidden 896, negatives 400, window 1.0, seed 0, GPU A100",
  "dev_balanced_acc": 0.8815,
  "dev_acc": 0.8383,
  "worst_language_frr_dev": {
    "language": "sv-SE",
    "frr": 0.6
  },
  "test_far_1pct": "NO_THRESHOLD_REACHES_FAR_0.01",
  "honest_note": "v2 improves dev balanced acc 0.84 -> 0.8815, but the production bar (FAR<=1% on test) is still unmet - needs more negatives / larger negative set"
}