CONFIG = {
    "seed": 42,  # Random seed for reproducibility
    "use_cpu": False,  # Whether to use CPU (it's easier to debug with CPU when starting to test the code)
    "learning_rate": 5e-5,  # Learning rate
    "num_train_epochs": 10,  # Total number of training epochs
    "per_device_train_batch_size": 4,  # Batch size per GPU for training
    "per_device_eval_batch_size": 8,  # Batch size per GPU for evaluation
    "output_dir": "train_result",  # Output directory
    "overwrite_output_dir": False,  # If the output directory exists, do not delete its content
    "report_to": ["tensorboard"],  # Report logs to TensorBoard
    "save_strategy": "steps",  # Strategy to save checkpoints
    "save_steps": 500,  # Interval of steps to save checkpoints, can be int or a float (0~1), when float it represents the ratio of total training steps (e.g., can set to 1.0 / 2000)
    "save_total_limit": 5,  # Maximum number of models to save. The oldest models will be deleted if this number is exceeded
    "logging_strategy": "steps",  # Log every certain number of steps
    "logging_steps": 500,  # Number of steps between each log
    "logging_nan_inf_filter": False,  # Record logs for loss=nan or inf
    "optim": "adamw_torch",  # Optimizer
    "lr_scheduler_type": "cosine",  # Learning rate scheduler
    "warmup_ratio": 0.1,  # Ratio of warmup steps in total training steps (e.g., for 1000 steps, the first 100 steps gradually increase lr from 0 to the set lr)
    "max_grad_norm": 1.0,  # For gradient clipping, ensure the norm of the gradients does not exceed 1.0 (default 1.0)
    "fp16": False,  # Whether to use 16-bit floating point for training (generally not recommended, as loss can easily explode)
    "bf16": False,  # Whether to use Brain Floating Point (bfloat16) for training (recommended if architecture supports it)
    "gradient_accumulation_steps": 1,  # Gradient accumulation steps, consider this parameter to achieve large batch size effects when batch size cannot be large
    "jit_mode_eval": False,  # Whether to use PyTorch jit trace during eval (can speed up the model, but the model must be static, otherwise will throw errors)
    "torch_compile": False,  # Whether to use torch.compile to compile the model (for better training and inference performance)
    "dataloader_pin_memory": True,  # Can speed up data transfer between CPU and GPU
    "dataloader_num_workers": 1,  # Default is not to use multiprocessing for data loading, usually set to 4*number of GPUs used
    "evaluation_strategy": "steps",  # Evaluation strategy, can be "steps" or "epoch"
    "eval_steps": 500,  # If evaluation_strategy="step"
    "remove_unused_columns": False,  # Don't change this unless you really know what you are doing.
}