{
  "encoder": "HalleluBERT/HalleluBERT_large",
  "head_layers": 2,
  "max_len": 512,
  "head_max_len": 256,
  "act_costs": {
    "escalate": 0.5
  },
  "amp_dtype": "bf16",
  "model_name": "nitzotz",
  "temperature": [
    1.0970542430877686,
    1.1389989852905273,
    1.1619657278060913
  ],
  "training": {
    "encoder_init": "HalleluBERT-large after extractive QA on HeQ v1.1 train (test-overlapping passages removed)",
    "data": "116,854 items: synthetic Hebrew messages, HeQ v1.1 train, MASSIVE he-IL train, reading and topic items on FineWeb-2 Hebrew passages, warnings about scams, short scams and look-alike messages, 'none of the options' items, reworded copies (see README)",
    "items_total": 116854,
    "train_items": 112029,
    "calib_items": 4825,
    "epochs_planned": 2,
    "epochs_run": 2,
    "kept_epoch": 2,
    "heldout_by_epoch": [
      {
        "acc": 0.9389,
        "soft_nll": 0.4909,
        "acc_base_families": 0.9359,
        "by_family": {
          "aug_heq": 0.9399,
          "aug_massive": 0.9464,
          "aug_scam": 0.9624,
          "aug_type": 0.9299,
          "aug_urgency": 0.8488,
          "claim": 0.9498,
          "heq_read": 0.964,
          "heq_unans": 0.892,
          "heq_verify": 0.9733,
          "massive_intent": 0.9286,
          "rel6_scam": 0.9574,
          "rel6_type": 0.931,
          "rel_scam": 0.9767,
          "rel_type": 0.9216,
          "routing": 0.9224,
          "spam_type": 0.9329,
          "urgency": 0.92
        },
        "epoch": 1,
        "step": 3501,
        "train_seconds": 1746.6
      },
      {
        "acc": 0.9505,
        "soft_nll": 0.4741,
        "acc_base_families": 0.9452,
        "by_family": {
          "aug_heq": 0.9563,
          "aug_massive": 0.9375,
          "aug_scam": 0.9731,
          "aug_type": 0.9482,
          "aug_urgency": 0.8721,
          "claim": 0.9658,
          "heq_read": 0.976,
          "heq_unans": 0.896,
          "heq_verify": 0.98,
          "massive_intent": 0.9314,
          "rel6_scam": 0.9787,
          "rel6_type": 0.9425,
          "rel_scam": 0.9833,
          "rel_type": 0.9515,
          "routing": 0.9361,
          "spam_type": 0.9486,
          "urgency": 0.8971
        },
        "epoch": 2,
        "step": 7002,
        "train_seconds": 1757.1
      }
    ],
    "steps": 7002,
    "batch": 32,
    "micro_batch": 16,
    "lr_encoder": 1e-05,
    "lr_head": 0.0001,
    "warmup": 0.1,
    "schedule": "cosine",
    "seed": 1,
    "split_seed": 1234,
    "loss": "soft cross-entropy",
    "amp": "bf16",
    "device": "cuda",
    "train_seconds": 3520.6,
    "head_init": "fresh (no laya checkpoint weights)",
    "class_weights": "none",
    "option_shuffle_families": [
      "read2_multi",
      "read2_neg",
      "read2_para",
      "read2_pos"
    ],
    "holdout_by_parent": true,
    "parent_copies_dropped": 0
  }
}