args.metric_for_best_model default to loss.

src/transformers/trainer.py

@@ -757,6 +757,9 @@ def tokenizer(self, processing_class) -> None:
         )
         self.processing_class = processing_class
 
+        if not self.args.metric_for_best_model:
+            self.args.metric_for_best_model = "loss"
+
     def _activate_neftune(self, model):
         r"""
         Activates the neftune as presented in this code: https://github.com/neelsjain/NEFTune and paper:
@@ -2991,10 +2994,7 @@ def _maybe_log_save_evaluate(self, tr_loss, grad_norm, model, trial, epoch, igno
             new_best_metric = self._determine_best_metric(metrics=metrics, trial=trial)
 
             if self.args.save_strategy == SaveStrategy.BEST:
-                if new_best_metric:
-                    self.control.should_save = True
-                else:
-                    self.control_should_save = False
+                self.control.should_save = new_best_metric
 
         if self.control.should_save:
             self._save_checkpoint(model, trial)
@@ -3098,10 +3098,7 @@ def _determine_best_metric(self, metrics, trial):
                     f"The available evaluation metrics are: {list(metrics.keys())}. Consider changing the `metric_for_best_model` via the TrainingArguments."
                 ) from exc
 
-            if self.args.greater_is_better:
-                operator = np.greater
-            else:
-                operator = np.less
+            operator = np.greater if self.args.greater_is_better else np.less
         else:
             metric_value = metrics["eval_loss"]
             operator = np.less