learning_rate: 2.0e-4
num_train_epochs: 1
max_seq_length: 8192
per_device_train_batch_size: 2
gradient_accumulation_steps: 4
optim: "adamw_8bit"
bf16: true
save_strategy: "epoch"