Revert "Don't reset the dataset type + plug for rm unused columns (hu…

…ggingface#6683)"

This reverts commit 9d4f322.

src/transformers/trainer.py

@@ -244,8 +244,6 @@ def __init__(
             self.scaler = torch.cuda.amp.GradScaler()
 
     def _remove_unused_columns(self, dataset: "nlp.Dataset", description: Optional[str] = None):
-        if not self.args.remove_unused_columns:
-            return
         # Inspect model forward signature to keep only the arguments it accepts.
         signature = inspect.signature(self.model.forward)
         signature_columns = list(signature.parameters.keys())
@@ -257,10 +255,7 @@ def _remove_unused_columns(self, dataset: "nlp.Dataset", description: Optional[s
         logger.info(
             f"The following columns {dset_description}don't have a corresponding argument in `{self.model.__class__.__name__}.forward` and have been ignored: {', '.join(ignored_columns)}."
         )
-        ds_type = dataset.format["type"]
-        if ds_type == "python":
-            ds_type = None
-        dataset.set_format(type=ds_type, columns=columns)
+        dataset.set_format(columns=columns)
 
     def _get_train_sampler(self) -> Optional[torch.utils.data.sampler.Sampler]:
         if isinstance(self.train_dataset, torch.utils.data.IterableDataset):

src/transformers/training_args.py

@@ -114,11 +114,6 @@ class TrainingArguments:
             at the next training step under the keyword argument ``mems``.
         run_name (:obj:`str`, `optional`):
             A descriptor for the run. Notably used for wandb logging.
-        remove_unused_columns (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            If using `nlp.Dataset` datasets, whether or not to automatically remove the columns unused by the model
-            forward method.
-
-            (Note: this behavior is not implemented for :class:`~transformers.TFTrainer` yet.)
     """
 
     output_dir: str = field(
@@ -239,10 +234,6 @@ class TrainingArguments:
         default=None, metadata={"help": "An optional descriptor for the run. Notably used for wandb logging."}
     )
 
-    remove_unused_columns: Optional[bool] = field(
-        default=True, metadata={"help": "Remove columns not required by the model when using an nlp.Dataset."}
-    )
-
     @property
     def train_batch_size(self) -> int:
         """