**Seq2SeqTrainingArguments**
*Seq2SeqTrainingArguments(
output_dir="thaisum",
evaluation_strategy="epoch",
do_train=True,
do_eval=True,
learning_rate=1e-4,
per_device_train_batch_size=8,
per_device_eval_batch_size=8,
weight_decay=0.01,
save_total_limit=2,
num_train_epochs=30,
warmup_steps=500,
predict_with_generate=True,
fp16=False,
gradient_accumulation_steps=1,
push_to_hub=False,
optim="adafactor", #paged_adamw_8bit
lr_scheduler_type = "linear",
logging_dir="thaisum/logs",
logging_strategy="epoch",
save_strategy="epoch",
metric_for_best_model="eval_loss",
load_best_model_at_end = True,
report_to="tensorboard")*
---
**TrainOutput**
*TrainOutput(
global_step=12114,
training_loss=2.4627872975289438,
metrics={'train_runtime': 8842.7967,
'train_samples_per_second': 36.525,
'train_steps_per_second': 4.566,
'total_flos': 1.92113185245696e+16,
'train_loss': 2.4627872975289438,
'epoch': 9.0}
)*