config.json

{
  "model": {
    "bert_model": "google-bert/bert-large-uncased",
    "hidden_size": 1024,
    "freeze_bert": true,
    "n_memory_tokens": 16,
    "bank_size": 128,
    "anchor_dim": 1024,
    "n_bank_heads": 8,
    "bank_cross_layers": 2,
    "gate_type": "gru",
    "extract_layers": [
      2,
      5,
      8,
      11,
      14,
      17,
      20,
      23
    ],
    "layer_fusion": "learned",
    "max_content_tokens": 480,
    "segment_overlap": 64,
    "max_position": 512,
    "n_teachers": 2,
    "teacher_hidden": 1024,
    "cv_target": 0.2
  },
  "training": {
    "max_documents": 50000,
    "max_val_documents": 500,
    "segment_length": 480,
    "segment_overlap": 64,
    "target_chain_segments": 16,
    "max_segments": 16,
    "min_segments": 6,
    "modern_bert_model": "answerdotai/ModernBERT-large",
    "longformer_model": "allenai/longformer-large-4096",
    "modern_max_len": 8192,
    "longformer_max_len": 4096,
    "procrustes_n_samples": 500,
    "epochs": 10,
    "batch_size": 4,
    "lr_bank": 0.002,
    "lr_output": 0.0005,
    "lr_proj": 0.001,
    "min_lr": 1e-06,
    "weight_decay": 0.01,
    "grad_clip": 1.0,
    "warmup_steps": 300,
    "tbptt_segments": 0,
    "modern_weight": 1.0,
    "longformer_weight": 0.5,
    "cv_weight": 0.05,
    "temperature": 0.07,
    "checkpoint_dir": "/home/claude/deep_bert_v3_checkpoints",
    "tensorboard_dir": "/home/claude/deep_bert_v3_tb",
    "log_every": 20,
    "eval_every": 200,
    "save_every_epoch": true
  }
}