Did you forget to write "loss_model.eval()"?

```
def _eval_during_training(self, evaluator, output_path, save_best_model, epoch, steps):
"""Runs evaluation during the training"""
if evaluator is not None:
score = evaluator(self, output_path=output_path, epoch=epoch, steps=steps)
if score > self.best_score and save_best_model:
self.save(output_path)
self.best_score = score
```

The method is called here:
```
if evaluation_steps > 0 and training_steps % evaluation_steps == 0:
self._eval_during_training(evaluator, output_path, save_best_model, epoch, training_steps)
for loss_model in loss_models:
loss_model.zero_grad()
loss_model.train()
```

Start at _eval_during_training and its call site in the training loop shown in the issue. Trace how evaluator and loss_models are used during evaluation, especially each loss_model's train/eval mode, and determine the expected mode before and after evaluation. Done means the evaluation path uses the intended mode without leaving training behavior incorrect.

Written by the indexing model from the issue text.