Views
No views yet
warmup_steps = 5,
num_train_epochs = 3,
learning_rate = 5e-5,
optim = 'adamw_torch', #"adamw_8bit",
weight_decay = 0.03, #L2 reg
lr_scheduler_type = "linear",
per_device_train_batch_size = 20, #max 20
gradient_accumulation_steps = 4,
adafactor = True,
use_liger = True,