toktokenizer 0.1.2

No project description provided

🪙 toktokenizer

toktokenizer is a BPE tokenizer implemented in rust and exposed in python using pyo3 bindings.

import toktokenizer as tok
bpe = tok.BPETokenizer.from_pretrained("wikibpe.json")

assert bpe.decode(bpe.encode("rust is pretty fun 🦀"))

Install toktokenizer from PyPI with the following

pip install toktokenizer

Note: if you want to build from source make sure rust is installed!

The only class toktokenizer exposes is BPETokenizer. The class itself is pretty minimalistic, with all major methods being showed below:

from toktokenizer import BPETokenizer

bpe = BPETokenizer()

# train a byte-pair tokenizer on some corpus
train_corpus = "this is some training data. any dumped string will do!"
vocab_size = 8
bpe.train(train_corpus, vocab_size)

# save tokenizer state
bpe.save_encoder("8word.json")

# load tokenizer from dumped file
bpe.load_encoder("8word.json")

# encode and decode
input_ids = bpe.encode("some data")
decoded = bpe.decode(input_ids)

Performance

slightly faster than openai & a lot quicker than 🤗!

Performance measured on 2.5MB from the wikitext test split using openai's tiktoken gpt2 tokenizer with tiktoken==0.6.0 and the implementation from 🤗 tokenizers at tokenizers==0.19.1