training_args.json

{
    "output_dir": "./ViTas_models",
    "report_to": "none",
    "learning_rate": 2e-05,
    "per_device_train_batch_size": 64,
    "per_device_eval_batch_size": 16,
    "seed": 42,
    "gradient_accumulation_steps": 4,
    "adam_beta1": 0.9,
    "adam_beta2": 0.999,
    "adam_epsilon": 1e-08,
    "lr_scheduler_type": "linear",
    "warmup_ratio": 0.1,
    "num_train_epochs": 8,
    "label_smoothing_factor": 0.1,
    "eval_strategy": "steps",
    "eval_steps": 50,
    "save_strategy": "steps",
    "save_steps": 500,
    "logging_steps": 50,
    "save_total_limit": 6,
    "load_best_model_at_end": true,
    "metric_for_best_model": "f1",
    "greater_is_better": true
}