jmaczan/bpe-tokenizer

Byte-Pair Encoding tokenizer for training large language models on huge datasets

bpe-tokenizer

Byte-Pair Encoding tokenizer for large language models that can be trained on arbitrarily huge datasets

"Materia" by Anita Maczan, Acrylic on canvas, 40x50, 2024

This implementation is suitable for working with huge datasets, because it processes data in chunks, both during tokenization and training

Training

from bpe_tokenizer import BPETokenizer

BPETokenizer().train(dataset_path="./path/to/dataset.txt")

CLI

python bpe_tokenizer.py train --training_dataset path_to_your_dataset.txt --vocabulary_size 5000 --training_output path_to_output_tokenizer.json

defaults:

• training_dataset = "training.txt"
• vocabulary_size = 50257
• training_output = "tokenizer.json"

Tokenize

from bpe_tokenizer import BPETokenizer

BPETokenizer().tokenize(text_to_be_tokenized)

CLI

python bpe_tokenizer.py tokenize --tokenizer_data path_to_tokenizer_data.json --run_data tokenize.json

defaults:

• tokenizer_data = "tokenizer.json"
• run_data = "tokenize.txt"

run_data file structure:

{
  "data": "Study hard what interests you the most in the most undisciplined, irreverent and original manner possible - Richard Feynmann"
}

Detokenize

from bpe_tokenizer import BPETokenizer

BPETokenizer().detokenize(array_of_tokens_to_be_parsed_to_text)

CLI

python bpe_tokenizer.py detokenize --tokenizer_data path_to_tokenizer_data.json --run_data detokenize.json

defaults:

• tokenizer_data = "tokenizer.json"
• run_data = "detokenize.txt"

run_data file structure:

{
  "data": [
    275, 265, 282, 32, 111, 108, 100, 32, 98, 97, 115, 116, 97, 114, 100, 284,
    268, 265, 282, 32, 282, 110, 292, 98, 97, 115, 116, 97, 114, 100, 33
  ]
}

License

GPL v3

Jędrzej Maczan, 2024