per_device_train_batch_size: int = 2
    gradient_accumulation_steps: int = 1
    warmup_steps: int = 5
    num_train_epochs: int = 25
    learning_rate: float = 1e-4
    weight_decay: float = 0.0
    lr_scheduler_type: str = "linear"  # or "constant"
    seed: int = 3407
    optim: str = "adamw_8bit"
    save_steps: int = 100
    save_total_limit: int = 5
    bias: str = "none"
    logging_strategy: str = "steps"