clean_up_tokenization_spaces=False if unset (#31938)

* clean_up_tokenization_spaces=False if unset

* deprecate warning

* updating param for old models

* update models

* make fix-copies

* fix-copies and update bert models

* warning msg

* update prophet and clvp

* updating test since space before is arbitrarily removed

* remove warning for 4.45

src/transformers/models/bert/tokenization_bert.py

@@ -88,6 +88,9 @@ class BertTokenizer(PreTrainedTokenizer):
         strip_accents (`bool`, *optional*):
             Whether or not to strip all accents. If this option is not specified, then it will be determined by the
             value for `lowercase` (as in the original BERT).
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -105,6 +108,7 @@ class BertTokenizer(PreTrainedTokenizer):
         mask_token="[MASK]",
         tokenize_chinese_chars=True,
         strip_accents=None,
+        clean_up_tokenization_spaces=True,
         **kwargs,
     ):
         if not os.path.isfile(vocab_file):
@@ -136,6 +140,7 @@ class BertTokenizer(PreTrainedTokenizer):
             mask_token=mask_token,
             tokenize_chinese_chars=tokenize_chinese_chars,
             strip_accents=strip_accents,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
             **kwargs,
         )
 

src/transformers/models/convbert/tokenization_convbert.py

@@ -91,6 +91,9 @@ class ConvBertTokenizer(PreTrainedTokenizer):
         strip_accents (`bool`, *optional*):
             Whether or not to strip all accents. If this option is not specified, then it will be determined by the
             value for `lowercase` (as in the original ConvBERT).
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -108,6 +111,7 @@ class ConvBertTokenizer(PreTrainedTokenizer):
         mask_token="[MASK]",
         tokenize_chinese_chars=True,
         strip_accents=None,
+        clean_up_tokenization_spaces=True,
         **kwargs,
     ):
         if not os.path.isfile(vocab_file):
@@ -139,6 +143,7 @@ class ConvBertTokenizer(PreTrainedTokenizer):
             mask_token=mask_token,
             tokenize_chinese_chars=tokenize_chinese_chars,
             strip_accents=strip_accents,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
             **kwargs,
         )
 

src/transformers/models/distilbert/tokenization_distilbert.py

@@ -90,6 +90,9 @@ class DistilBertTokenizer(PreTrainedTokenizer):
         strip_accents (`bool`, *optional*):
             Whether or not to strip all accents. If this option is not specified, then it will be determined by the
             value for `lowercase` (as in the original BERT).
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -108,6 +111,7 @@ class DistilBertTokenizer(PreTrainedTokenizer):
         mask_token="[MASK]",
         tokenize_chinese_chars=True,
         strip_accents=None,
+        clean_up_tokenization_spaces=True,
         **kwargs,
     ):
         if not os.path.isfile(vocab_file):
@@ -138,6 +142,7 @@ class DistilBertTokenizer(PreTrainedTokenizer):
             mask_token=mask_token,
             tokenize_chinese_chars=tokenize_chinese_chars,
             strip_accents=strip_accents,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
             **kwargs,
         )
 

src/transformers/models/electra/tokenization_electra.py

@@ -90,6 +90,9 @@ class ElectraTokenizer(PreTrainedTokenizer):
         strip_accents (`bool`, *optional*):
             Whether or not to strip all accents. If this option is not specified, then it will be determined by the
             value for `lowercase` (as in the original Electra).
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -107,6 +110,7 @@ class ElectraTokenizer(PreTrainedTokenizer):
         mask_token="[MASK]",
         tokenize_chinese_chars=True,
         strip_accents=None,
+        clean_up_tokenization_spaces=True,
         **kwargs,
     ):
         if not os.path.isfile(vocab_file):
@@ -138,6 +142,7 @@ class ElectraTokenizer(PreTrainedTokenizer):
             mask_token=mask_token,
             tokenize_chinese_chars=tokenize_chinese_chars,
             strip_accents=strip_accents,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
             **kwargs,
         )
 

src/transformers/models/funnel/tokenization_funnel.py

@@ -107,6 +107,9 @@ class FunnelTokenizer(PreTrainedTokenizer):
         strip_accents (`bool`, *optional*):
             Whether or not to strip all accents. If this option is not specified, then it will be determined by the
             value for `lowercase` (as in the original BERT).
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -127,6 +130,7 @@ class FunnelTokenizer(PreTrainedTokenizer):
         eos_token="</s>",
         tokenize_chinese_chars=True,
         strip_accents=None,
+        clean_up_tokenization_spaces=True,
         **kwargs,
     ):
         if not os.path.isfile(vocab_file):
@@ -159,6 +163,7 @@ class FunnelTokenizer(PreTrainedTokenizer):
             eos_token=eos_token,
             tokenize_chinese_chars=tokenize_chinese_chars,
             strip_accents=strip_accents,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
             **kwargs,
         )
 

src/transformers/models/layoutlm/tokenization_layoutlm.py

@@ -91,6 +91,9 @@ class LayoutLMTokenizer(PreTrainedTokenizer):
         strip_accents (`bool`, *optional*):
             Whether or not to strip all accents. If this option is not specified, then it will be determined by the
             value for `lowercase` (as in the original LayoutLM).
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -108,6 +111,7 @@ class LayoutLMTokenizer(PreTrainedTokenizer):
         mask_token="[MASK]",
         tokenize_chinese_chars=True,
         strip_accents=None,
+        clean_up_tokenization_spaces=True,
         **kwargs,
     ):
         if not os.path.isfile(vocab_file):
@@ -139,6 +143,7 @@ class LayoutLMTokenizer(PreTrainedTokenizer):
             mask_token=mask_token,
             tokenize_chinese_chars=tokenize_chinese_chars,
             strip_accents=strip_accents,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
             **kwargs,
         )
 

src/transformers/models/lxmert/tokenization_lxmert.py

@@ -90,6 +90,9 @@ class LxmertTokenizer(PreTrainedTokenizer):
         strip_accents (`bool`, *optional*):
             Whether or not to strip all accents. If this option is not specified, then it will be determined by the
             value for `lowercase` (as in the original Lxmert).
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -107,6 +110,7 @@ class LxmertTokenizer(PreTrainedTokenizer):
         mask_token="[MASK]",
         tokenize_chinese_chars=True,
         strip_accents=None,
+        clean_up_tokenization_spaces=True,
         **kwargs,
     ):
         if not os.path.isfile(vocab_file):
@@ -138,6 +142,7 @@ class LxmertTokenizer(PreTrainedTokenizer):
             mask_token=mask_token,
             tokenize_chinese_chars=tokenize_chinese_chars,
             strip_accents=strip_accents,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
             **kwargs,
         )
 

src/transformers/models/mobilebert/tokenization_mobilebert.py

@@ -92,6 +92,9 @@ class MobileBertTokenizer(PreTrainedTokenizer):
         strip_accents (`bool`, *optional*):
             Whether or not to strip all accents. If this option is not specified, then it will be determined by the
             value for `lowercase` (as in the original MobileBERT).
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -109,6 +112,7 @@ class MobileBertTokenizer(PreTrainedTokenizer):
         mask_token="[MASK]",
         tokenize_chinese_chars=True,
         strip_accents=None,
+        clean_up_tokenization_spaces=True,
         **kwargs,
     ):
         if not os.path.isfile(vocab_file):
@@ -140,6 +144,7 @@ class MobileBertTokenizer(PreTrainedTokenizer):
             mask_token=mask_token,
             tokenize_chinese_chars=tokenize_chinese_chars,
             strip_accents=strip_accents,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
             **kwargs,
         )
 

src/transformers/models/mpnet/tokenization_mpnet.py

@@ -108,6 +108,9 @@ class MPNetTokenizer(PreTrainedTokenizer):
         strip_accents (`bool`, *optional*):
             Whether or not to strip all accents. If this option is not specified, then it will be determined by the
             value for `lowercase` (as in the original BERT).
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -128,6 +131,7 @@ class MPNetTokenizer(PreTrainedTokenizer):
         mask_token="<mask>",
         tokenize_chinese_chars=True,
         strip_accents=None,
+        clean_up_tokenization_spaces=True,
         **kwargs,
     ):
         bos_token = AddedToken(bos_token, special=True) if isinstance(bos_token, str) else bos_token
@@ -170,6 +174,7 @@ class MPNetTokenizer(PreTrainedTokenizer):
             mask_token=mask_token,
             tokenize_chinese_chars=tokenize_chinese_chars,
             strip_accents=strip_accents,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
             **kwargs,
         )
 

src/transformers/models/prophetnet/tokenization_prophetnet.py

@@ -308,6 +308,9 @@ class ProphetNetTokenizer(PreTrainedTokenizer):
         strip_accents (`bool`, *optional*):
             Whether or not to strip all accents. If this option is not specified, then it will be determined by the
             value for `lowercase` (as in the original BERT).
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -330,6 +333,7 @@ class ProphetNetTokenizer(PreTrainedTokenizer):
         mask_token: Optional[str] = "[MASK]",
         tokenize_chinese_chars: Optional[bool] = True,
         strip_accents: Optional[bool] = None,
+        clean_up_tokenization_spaces: bool = True,
         **kwargs,
     ):
         if not os.path.isfile(vocab_file):
@@ -360,6 +364,7 @@ class ProphetNetTokenizer(PreTrainedTokenizer):
             mask_token=mask_token,
             tokenize_chinese_chars=tokenize_chinese_chars,
             strip_accents=strip_accents,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
             **kwargs,
         )
 

src/transformers/models/squeezebert/tokenization_squeezebert.py

@@ -91,6 +91,9 @@ class SqueezeBertTokenizer(PreTrainedTokenizer):
         strip_accents (`bool`, *optional*):
             Whether or not to strip all accents. If this option is not specified, then it will be determined by the
             value for `lowercase` (as in the original SqueezeBERT).
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -108,6 +111,7 @@ class SqueezeBertTokenizer(PreTrainedTokenizer):
         mask_token="[MASK]",
         tokenize_chinese_chars=True,
         strip_accents=None,
+        clean_up_tokenization_spaces=True,
         **kwargs,
     ):
         if not os.path.isfile(vocab_file):
@@ -139,6 +143,7 @@ class SqueezeBertTokenizer(PreTrainedTokenizer):
             mask_token=mask_token,
             tokenize_chinese_chars=tokenize_chinese_chars,
             strip_accents=strip_accents,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
             **kwargs,
         )
 

src/transformers/models/tapas/tokenization_tapas.py

@@ -225,6 +225,9 @@ class TapasTokenizer(PreTrainedTokenizer):
             Minimum length of each question in terms of tokens (will be skipped otherwise).
         max_question_length (`int`, *optional*):
             Maximum length of each question in terms of tokens (will be skipped otherwise).
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):
+            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
+            extra spaces.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -252,6 +255,7 @@ class TapasTokenizer(PreTrainedTokenizer):
         max_question_length=None,
         model_max_length: int = 512,
         additional_special_tokens: Optional[List[str]] = None,
+        clean_up_tokenization_spaces=True,
         **kwargs,
     ):
         if not is_pandas_available():
@@ -322,6 +326,7 @@ class TapasTokenizer(PreTrainedTokenizer):
             max_question_length=max_question_length,
             model_max_length=model_max_length,
             additional_special_tokens=additional_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
             **kwargs,
         )
 

src/transformers/tokenization_utils_base.py

@@ -1622,7 +1622,7 @@ class PreTrainedTokenizerBase(SpecialTokensMixin, PushToHubMixin):
             )
 
         # By default, cleaning tokenization spaces for both fast and slow tokenizers
-        self.clean_up_tokenization_spaces = kwargs.pop("clean_up_tokenization_spaces", True)
+        self.clean_up_tokenization_spaces = kwargs.pop("clean_up_tokenization_spaces", False)
 
         # By default, do not split special tokens for both fast and slow tokenizers
         self.split_special_tokens = kwargs.pop("split_special_tokens", False)

tests/models/clvp/test_tokenization_clvp.py

@@ -79,7 +79,7 @@ class ClvpTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
     # Copied from transformers.tests.models.gpt2.test_tokenization_gpt2.GPT2TokenizationTest.get_input_output_texts
     def get_input_output_texts(self, tokenizer):
         input_text = "lower newer"
-        output_text = "lower newer"
+        output_text = "lower[SPACE]newer"
         return input_text, output_text
 
     # Copied from transformers.tests.models.layoutxlm.test_tokenization_layoutxlm.LayoutXLMTokenizationTest.test_add_special_tokens

tests/models/wav2vec2/test_tokenization_wav2vec2.py

@@ -147,8 +147,8 @@ class Wav2Vec2TokenizerTest(unittest.TestCase):
         batch_tokens = tokenizer.batch_decode(sample_ids)
         batch_tokens_2 = tokenizer.batch_decode(sample_ids, skip_special_tokens=True)
 
-        self.assertEqual(batch_tokens, ["HELLO<unk>!?!?$$$", "BYE BYE<unk>$$$"])
-        self.assertEqual(batch_tokens_2, ["HELO!?!?", "BYE BYE"])
+        self.assertEqual(batch_tokens, ["HELLO<unk>!? !?$$$", "BYE BYE<unk>$$$"])
+        self.assertEqual(batch_tokens_2, ["HELO!? !?", "BYE BYE"])
 
     def test_call(self):
         # Tests that all call wrap to encode_plus and batch_encode_plus
@@ -467,8 +467,8 @@ class Wav2Vec2CTCTokenizerTest(TokenizerTesterMixin, unittest.TestCase):
         batch_tokens = tokenizer.batch_decode(sample_ids)
         batch_tokens_2 = tokenizer.batch_decode(sample_ids, skip_special_tokens=True)
 
-        self.assertEqual(batch_tokens, ["HELLO<unk>!?!?<new_tokens>$$$", "BYE BYE<unk><new_tokens>$$$"])
-        self.assertEqual(batch_tokens_2, ["HELO!?!?<new_tokens>", "BYE BYE<new_tokens>"])
+        self.assertEqual(batch_tokens, ["HELLO<unk>!? !?<new_tokens>$$$", "BYE BYE<unk><new_tokens>$$$"])
+        self.assertEqual(batch_tokens_2, ["HELO!? !?<new_tokens>", "BYE BYE<new_tokens>"])
 
     def test_special_characters_in_vocab(self):
         sent = "ʈʰ æ æ̃ ˧ kʰ"

tests/models/wav2vec2_phoneme/test_tokenization_wav2vec2_phoneme.py

@@ -249,7 +249,7 @@ class Wav2Vec2PhonemeCTCTokenizerTest(TokenizerTesterMixin, unittest.TestCase):
         # fmt: on
 
         batch_tokens = tokenizer.batch_decode(sample_ids)
-        self.assertEqual(batch_tokens, ["k s ɾ ɾ l ɭʲ!?!? $$$", "j ð s j ð s oːɹ $$$"])
+        self.assertEqual(batch_tokens, ["k s ɾ ɾ l ɭʲ ! ? ! ? $$$", "j ð s j ð s oːɹ $$$"])
 
     @staticmethod
     def get_from_offsets(offsets, key):