training_config.json

{
  "learning_rate": 2e-05,
  "batch_size": 8,
  "gradient_accumulation_steps": 6,
  "num_epochs": 3,
  "max_length": 2048,
  "warmup_ratio": 0.03,
  "weight_decay": 0.01,
  "max_grad_norm": 1.0,
  "seed": 42,
  "eval_frequency": 5,
  "logging_steps": 10,
  "dataloader_num_workers": 8,
  "pin_memory": true,
  "dataloader_persistent_workers": true,
  "prefetch_factor": 4,
  "early_stopping_patience": 3,
  "early_stopping_min_delta": 0.001,
  "checkpoint_dir": "/root/llama3.2-3b-sft-training",
  "save_best_model": true,
  "save_last_checkpoint": true,
  "save_every_n_epochs": 1,
  "save_every_n_steps": 5000,
  "keep_last_n_checkpoints": 2,
  "resume_from_checkpoint": "/root/llama3.2-3b-sft-training/checkpoints/checkpoint-epoch-1.pt"
}