model_config.json

{
  "data_dir": "/kaggle/input/difraud",
  "output_dir": "/kaggle/working",
  "model_stem": "model",
  "use_liar": true,
  "semantic_model_id": "answerdotai/ModernBERT-base",
  "freeze_encoder_layers": 0,
  "semantic_dim": 768,
  "ling_dim": 8,
  "hidden_dim": 512,
  "dropout": 0.1,
  "max_seq_len": 256,
  "batch_size": 8,
  "num_epochs": 5,
  "lr_encoder": 1e-05,
  "lr_head": 0.0001,
  "weight_decay": 0.02,
  "llrd_decay": 0.9,
  "warmup_pct": 0.08,
  "grad_clip": 0.7,
  "label_smoothing": 0.05,
  "fp16": true,
  "accum_steps": 4,
  "focal_gamma": 2.0,
  "seed": 42,
  "num_workers": 4,
  "top_k_evidence": 3,
  "device": "cuda",
  "use_ema": true,
  "ema_decay": 0.995,
  "use_fgm": false,
  "fgm_epsilon": 0.25,
  "use_rdrop": false,
  "rdrop_coef": 0.05,
  "use_dann": false,
  "dann_lambda": 0.5,
  "dann_lambda_start": 0.01,
  "dann_num_domains": 7,
  "use_supcon": true,
  "supcon_coef": 0.1,
  "use_domain_sampler": true,
  "use_hard_negative_mining": false,
  "hard_neg_conf": 0.8,
  "hard_neg_boost": 3.0,
  "snapshot_top_k": 2,
  "use_domain_mtl": true,
  "domain_mtl_coef": 0.1,
  "use_awp": true,
  "awp_epsilon": 0.001,
  "awp_lr": 0.001,
  "awp_start_epoch": 3,
  "use_pseudo_label": false,
  "pseudo_label_conf": 0.98,
  "use_external_data": true,
  "external_weight": 0.3,
  "use_perplexity": false,
  "perp_model_id": "distilgpt2",
  "perp_batch_size": 64,
  "multi_seed": false,
  "multi_seed_list": [
    42,
    1337,
    2025
  ],
  "prototype_head_scale": 20.0,
  "use_local_consistency": true,
  "local_cons_segments": 4,
  "spectral_top_k": 8,
  "early_stopping_patience": 2,
  "use_temperature_scaling": true,
  "use_compile": false,
  "use_gradient_checkpointing": true
}