Skip to content

πŸ’₯Fast State-of-the-Art Tokenizers optimized for Research and Production

License

Notifications You must be signed in to change notification settings

cedrickchee/tokenizers

Β 
Β 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation



Build GitHub

Provides an implementation of today's most used tokenizers, with a focus on performance and versatility.

Main features:

  • Train new vocabularies and tokenize, using today's most used tokenizers.
  • Extremely fast (both training and tokenization), thanks to the Rust implementation. Takes less than 20 seconds to tokenize a GB of text on a server's CPU.
  • Easy to use, but also extremely versatile.
  • Designed for research and production.
  • Normalization comes with alignments tracking. It's always possible to get the part of the original sentence that corresponds to a given token.
  • Does all the pre-processing: Truncate, Pad, add the special tokens your model needs.

Quick examples using Python:

Start using in a matter of seconds:

# Tokenizers provides ultra-fast implementations of most current tokenizers:
>>> from tokenizers import (ByteLevelBPETokenizer,
                            BPETokenizer,
                            SentencePieceBPETokenizer,
                            BertWordPieceTokenizer)
# Ultra-fast => they can encode 1GB of text in ~20sec on a standard server's CPU
# Tokenizers can be easily instantiated from standard files
>>> tokenizer = BertWordPieceTokenizer("bert-base-uncased-vocab.txt", lowercase=True)
Tokenizer(vocabulary_size=30522, model=BertWordPiece, add_special_tokens=True, unk_token=[UNK], 
          sep_token=[SEP], cls_token=[CLS], clean_text=True, handle_chinese_chars=True, 
          strip_accents=True, lowercase=True, wordpieces_prefix=##)

# Tokenizers provide exhaustive outputs: tokens, mapping to original string, attention/special token masks.
# They also handle model's max input lengths as well as padding (to directly encode in padded batches)
>>> output = tokenizer.encode("Hello, y'all! How are you 😁 ?")
Encoding(num_tokens=13, attributes=[ids, type_ids, tokens, offsets, attention_mask, special_tokens_mask, overflowing, original_str, normalized_str])
>>> print(output.ids, output.tokens, output.offsets)
[101, 7592, 1010, 1061, 1005, 2035, 999, 2129, 2024, 2017, 100, 1029, 102]
['[CLS]', 'hello', ',', 'y', "'", 'all', '!', 'how', 'are', 'you', '[UNK]', '?', '[SEP]']
[(0, 0), (0, 5), (5, 6), (7, 8), (8, 9), (9, 12), (12, 13), (14, 17), (18, 21), (22, 25), (26, 27),
 (28, 29), (0, 0)]
# Here is an example using the offsets mapping to retrieve the string coresponding to the 10th token:
>>> output.original_str[output.offsets[10]]
'😁'

And training a new vocabulary is just as easy:

# You can also train a BPE/Byte-levelBPE/WordPiece vocabulary on your own files
>>> tokenizer = ByteLevelBPETokenizer()
>>> tokenizer.train(["wiki.test.raw"], vocab_size=20000)
[00:00:00] Tokenize words                 β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ   20993/20993
[00:00:00] Count pairs                    β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ   20993/20993
[00:00:03] Compute merges                 β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ   19375/19375

Bindings

We provide bindings to the following languages (more to come!):

Packages

No packages published

Languages

  • Rust 76.0%
  • Python 17.6%
  • TypeScript 4.4%
  • JavaScript 1.5%
  • Other 0.5%