MineMA-3-8b-v1/untitled.txt

 r=64, # Lora 秩
    lora_alpha=128, # Lora alaph，具体作用参见 Lora 原理
    lora_dropout=0.1# Dropout 比例
)
args = TrainingArguments(
    output_dir="/home/jovyan/notebook/39newllama/results-8b",
    per_device_train_batch_size=4,
    gradient_accumulation_steps=2,
    logging_steps=50,
    num_train_epochs=1,
    save_steps=100,
    learning_rate=1e-4,
    save_on_each_node=True,
    gradient_checkpointing=True,
    #eval_steps=50,
    weight_decay=1e-4,
    #load_best_model_at_end=True,
    #evaluation_strategy="steps"
)