{
  "dataset": "HWU64",
  "source": "https://github.com/jianguoz/Few-Shot-Intent-Detection/tree/main/Datasets/HWU64",
  "task": "intent",
  "n_labels": 64,
  "label_wording": "raw intent strings from seq.in/label (frozen)",
  "train_rows": 8954,
  "test_rows": 1076,
  "protocol": "Every test example sees all 64 candidate labels. Official test never used for training or calibration.",
  "onesystem_version": "0.3.0",
  "device": "mps",
  "current_checkpoint": {
    "n": 1076,
    "n_labels": 64,
    "accuracy": 0.645,
    "macro_f1": 0.5976,
    "ece": 0.4091,
    "nll": 2.0467,
    "coverage": 1.0,
    "accepted_accuracy": 0.645,
    "abstain_rate": 0.0,
    "min_confidence": null,
    "latency": {
      "n": 50,
      "p50_ms": 9.31,
      "p95_ms": 11.3,
      "mean_ms": 9.5
    },
    "seconds": 2.5
  },
  "finetune_split": {
    "train_rows": 8048,
    "calib_rows": 906,
    "calib_ratio": 0.1,
    "seed": 42
  },
  "domain_adapted": {
    "n": 1076,
    "n_labels": 64,
    "accuracy": 0.9368,
    "macro_f1": 0.9338,
    "ece": 0.0201,
    "nll": 0.2745,
    "coverage": 1.0,
    "accepted_accuracy": 0.9368,
    "abstain_rate": 0.0,
    "min_confidence": null,
    "latency": {
      "n": 50,
      "p50_ms": 11.5,
      "p95_ms": 14.9,
      "mean_ms": 11.82
    },
    "seconds": 3.0,
    "temperature": 1.9000000000000001,
    "calib_nll": 0.2625,
    "train_run": {
      "steps": 4024,
      "history": [
        {
          "epoch": 1,
          "loss": 0.4646
        },
        {
          "epoch": 2,
          "loss": 0.1511
        },
        {
          "epoch": 3,
          "loss": 0.0426
        },
        {
          "epoch": 4,
          "loss": 0.0093
        }
      ],
      "seconds": 2346.4
    }
  }
}