config.json

{
  "tokenizer_dir": "/workspace/nav2tex-decoder/tokenizer",
  "vocab_size": 50000,
  "pad_token_id": 1,
  "bos_token_id": 0,
  "eos_token_id": 2,
  "data_glob": "/workspace/data/data/*.parquet",
  "max_seq_len": 1024,
  "cpe_score_threshold": 400,
  "cpe_ratio": 0.25,
  "d_model": 1024,
  "n_heads": 16,
  "n_layers": 8,
  "d_ff": 4096,
  "dropout": 0.1,
  "squeeze_ratio": 4,
  "lam_lambda": 0.1,
  "batch_size": 16,
  "grad_accum": 8,
  "max_epochs": 10,
  "warmup_steps": 2000,
  "lr": 0.0001,
  "weight_decay": 0.01,
  "max_grad_norm": 1.0,
  "label_smoothing": 0.1,
  "save_dir": "/workspace/checkpoints",
  "save_every_n_steps": 2000,
  "log_every_n_steps": 100,
  "num_workers": 8,
  "prefetch_factor": 4,
  "persistent_workers": true,
  "cuda_benchmark": true,
  "bf16": true,
  "compile": true,
  "flash_attn": false
}