Special tokens not tokenized properly

[manueltonneau]

Environment info

• transformers version: 4.5.1
• Python version: 3.8.5
• Using GPU in script?: no
• Using distributed or parallel set-up in script?: no

Who can help

@LysandreJik

Information

Hi,

I have recently further pretrained a RoBERTa model with fairseq. I use a custom vocabulary, trained with the tokenizers module. After converting the fairseq model to pytorch, I loaded all my model-related files here.

When loading the tokenizer, I noticed that the special tokens are not tokenized properly.

To reproduce

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('manueltonneau/twibert-lowercase-50272') 
tokenizer.tokenize('<mask>') 
Out[7]: ['<mask>']
tokenizer.tokenize('<hashtag>') 
Out[8]: ['hashtag']
tokenizer.tokenize('<hashtag>')
Out[3]: [0, 23958, 2]

Expected behavior

Since <hashtag> is a special token in the vocabulary with ID 7 (see here), the last output should be: [0, 7, 2]. <hashtag> with the '<>' should also be recognized as a unique token.

Potential explanation

When looking at the files from a similar model, it seems that the vocab is in txt format and they also have the bpe.codes file, which I don't have. Could that be the issue? And if so, how do I convert my files to this format?

For vocab.txt, I have already found your lengthy explanation here, thanks for this.

[LysandreJik]

Hello! What is your tokenizer? Is it a WordPiece-based tokenizer, or a Byte-level BPE-based tokenizer like the original one from RoBERTa?

[manueltonneau]

Hi @LysandreJik, thanks for your reply and sorry that I'm just seeing this now. My tokenizer is a byte-level BPE-based tokenizer.

[manueltonneau]

Hi @LysandreJik, let me know if you have a solution for this or if you need more info, thanks a lot in advance :)

[NielsRogge]

Hi,

How did you add the additional special tokens? So you start from a pre-trained RoBERTa, then added additional special tokens and further pre-trained on a corpus?

Did you add these additional special tokens using the tokenizers library? Normally, one can add additional tokens as follows (based on huggingface/tokenizers#247 (comment)):

special_tokens_dict = {'additional_special_tokens': ['[C1]','[C2]','[C3]','[C4]']}
num_added_toks = tokenizer.add_special_tokens(special_tokens_dict)

However, printing the following:

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('manueltonneau/twibert-lowercase-50272') 
print(tokenizer.additional_special_tokens)

Returns []. So you can solve it by doing:

special_tokens_dict = {'additional_special_tokens': ['<hashtag>']}
num_added_toks = tokenizer.add_special_tokens(special_tokens_dict)

When I then test your example:

tokenizer.tokenize('<hashtag>')

I get: ['<hashtag>'].

And when doing:

tokenizer.convert_tokens_to_ids(tokenizer.tokenize("<hashtag>", add_special_tokens=True))

I get: [0, 7, 2].

[manueltonneau]

Awesome @NielsRogge, thanks a lot! Will test this and get back to you/close if solved.

[manueltonneau]

How did you add the additional special tokens? So you start from a pre-trained RoBERTa, then added additional special tokens and further pre-trained on a corpus?

I created a new vocab with the tokenizers module for which I added new special tokens. Here is the code I use below:

# Initialize a tokenizer
  tokenizer = Tokenizer(models.BPE())

  # Customize pre-tokenization and decoding
  tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=True)
  tokenizer.decoder = decoders.ByteLevel()
  tokenizer.post_processor = processors.ByteLevel(trim_offsets=True)

  # And then train
  trainer = trainers.BpeTrainer(vocab_size=args.vocab_size, min_frequency=2, special_tokens=[
      "<s>",
      "<pad>",
      "</s>",
      "<unk>",
      "<mask>",
      "@USER",
      "HTTPURL",
      "<hashtag>",
      "</hashtag>"
      ], show_progress=True)
  files = [os.path.join(args.corpus_dir, filename) for filename in os.listdir(args.corpus_dir)]
  i = 0
  start_time = time.time()
  for file in files:
      print(f'Starting training on {file}')
      tokenizer.train([file], trainer=trainer)
      i = i + 1
      print(f'{i} files done out of {len(files)} files')
      print(f'Time elapsed: {time.time() - start_time} seconds')

  # And Save it
  output_dir = f'/scratch/mt4493/twitter_labor/twitter-labor-data/data/pretraining/US/vocab_files/{args.vocab_size}/{args.vocab_name}'
  if not os.path.exists(output_dir):
      os.makedirs(output_dir)
  tokenizer.model.save(output_dir)

[manueltonneau]

Works fine, thanks again!