orpo_args = ORPOConfig(
    learning_rate=8e-6,
    lr_scheduler_type="linear",
    max_length=1024,
    max_prompt_length=512,
    beta=0.1,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    gradient_accumulation_steps=4,
    optim="paged_adamw_8bit",
    num_train_epochs=1,
    eval_strategy="steps",
    eval_steps=0.2,
    logging_steps=1,
    warmup_steps=10,
    report_to="wandb",
    output_dir="./results/",
)