Precision and recall not changing while training BERT

I am training 'bert-base-cased' model for multiclass classification.
Loss is incredibly high, because training data is very noisy and I'm classifying over 700 classes.
I just wanted to see how the training goes.
However precision, recall, f1_score is not changing and I don't know why. (loss is decreasing)

I checked the value in more detail in case the change was too small, but the value was exactly the same.
I'm posting 'compute_metrics' I used. Let me know if you need more information.

def compute_metrics(pred):
  labels = pred.label_ids
  preds = pred.predictions.argmax(-1)
  precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='weighted')
  acc = accuracy_score(labels, preds)
  return {
      'accuracy': acc,
      'f1': f1,
      'precision': precision,
      'recall': recall
  }

trainer = Trainer(
  model=model,
  args=training_ars,
  train_dataset=train_dataset,
  eval_dataset=eval_dataset,
  tokenizer=tokenizer,
  compute_metrics=compute_metrics)