args = TrainingArguments(
output_dir=tokenizer_path / "model",
per_device_train_batch_size=32,
per_device_eval_batch_size=32,
evaluation_strategy="steps",
eval_steps=5000,
logging_steps=5000,
gradient_accumulation_steps=8,
num_train_epochs=100, # Először egy epochot nézzünk meg, hogy minden stimmel-e
weight_decay=0.1,
warmup_steps=1000,
lr_scheduler_type="cosine",
learning_rate=5e-4,
save_steps=5000,
fp16=True, # Lower precision only allowed on Cuda devices
#push_to_hub=True, Nem akarom a hub-ra feltennie
push_to_hub= False
)
trainer = Trainer(
model=model,
tokenizer=tokenizer,
args=args,
data_collator=data_collator,
train_dataset=tokenized_datasets["train"],
eval_dataset=tokenized_datasets["test"],
)
output = trainer.train()

trainer.state.log_history
>>> [{'train_runtime': 2240.0086,
'train_samples_per_second': 200.758,
'train_steps_per_second': 0.759,
'total_flos': 1.1068167168e+16,
'train_loss': 1.500028435202206,
'epoch': 96.45,
'step': 1700}]
It is clear that there is only one element of the log_history. What did I make wrong?
Thanks in advance