config.json

{
  "model_name": "TFM-Token",
  "paper": "Tokenizing Single-Channel EEG with Time-Frequency Motif Learning (ICLR 2026)",
  "paper_url": "https://openreview.net/forum?id=2sPmWHZ8Ir",
  "tokenizer": {
    "architecture": "TFM_VQVAE2_deep",
    "variant": "2x2x8",
    "in_channels": 1,
    "n_freq": 100,
    "n_freq_patch": 5,
    "emb_size": 64,
    "code_book_size": 8192,
    "trans_freq_encoder_depth": 2,
    "trans_temporal_encoder_depth": 2,
    "trans_decoder_depth": 8,
    "beta": 1.0,
    "resampling_rate": 200
  },
  "encoder": {
    "architecture": "TFM_TOKEN_Classifier",
    "variant": "64x4",
    "emb_size": 64,
    "code_book_size": 8192,
    "num_heads": 8,
    "depth": 4,
    "max_seq_len": 2048
  },
  "pretraining_datasets": [
    "TUAB",
    "TUEV",
    "CHBMIT"
  ],
  "finetuned_datasets": {
    "TUEV": {
      "num_classes": 6,
      "classification_task": "multi_class",
      "description": "Temple University EEG Event Detection (6-class)",
      "eval_metrics": [
        "accuracy",
        "balanced_accuracy",
        "cohen_kappa",
        "f1_weighted"
      ]
    },
    "TUAB": {
      "num_classes": 1,
      "classification_task": "binary",
      "description": "Temple University Abnormal EEG Detection (binary)",
      "eval_metrics": [
        "accuracy",
        "balanced_accuracy",
        "roc_auc",
        "pr_auc"
      ]
    }
  },
  "seeds": [
    1,
    2,
    3,
    4,
    5
  ]
}