{
  "preset": "small",
  "epochs": 8,
  "lr": 0.0003,
  "batch_size": 32,
  "realism": "tier1",
  "best_dev_micro_f1": 0.9680415187107347,
  "parameters": 12011719,
  "dev_templates": [
    "subscription_signup",
    "privacy_settings"
  ],
  "train_records": 14792,
  "dev_records": 3208,
  "val_records": 1800,
  "selection": "checkpoint and thresholds chosen on dev (held-out train templates); the test and hard splits were never used for selection",
  "device": "mps",
  "seed": 0,
  "validation_note": "val shares templates with train, so it saturates and is reported only as a sanity check; selection and threshold tuning use dev, which is two train templates held out entirely",
  "node_weight": 2.0,
  "pair_weight": 0.5,
  "permute_p": 0.7,
  "dropout": 0.1
}