training_args = TrainingArguments(
output_dir='t5-base-wikilarge-newsela-with-domain-adaptation',
num_train_epochs=20,
warmup_steps=250,
per_device_train_batch_size=BATCH_SIZE,
weight_decay=0.01,
learning_rate=2e-4,
# fp16=True,
optim="adafactor",
)
Step Training Loss
500 4.184500
1000 2.470900
1500 2.128900
2000 1.951600
2500 1.834400
3000 1.755800
3500 1.701800
4000 1.656300
4500 1.628800
TrainOutput(global_step=4640, training_loss=2.1286644540984057, metrics={'train_runtime': 4090.6694, 'train_samples_per_second': 72.526, 'train_steps_per_second': 1.134, 'total_flos': 0.0, 'train_loss': 2.1286644540984057, 'epoch': 20.0})
