config.json

{
  "dataset_name": "allenai/soda",
  "dataset_mode": "chat",
  "max_dataset_samples": 250000,
  "tokenizer_vocab_size": 8000,
  "context_length": 128,
  "d_model": 960,
  "n_heads": 8,
  "n_layers": 12,
  "d_ff": 3456,
  "dropout": 0.1,
  "batch_size": 32,
  "learning_rate": 0.0003,
  "weight_decay": 0.1,
  "warmup_steps": 500,
  "max_steps": 50000,
  "eval_interval": 2000,
  "eval_steps": 500,
  "save_interval": 5000,
  "grad_clip": 1.0,
  "device": "cuda",
  "fp16": true,
  "seed": 42,
  "output_dir": "cortex_104m_output_chat"
}