src/tokenizer.py

'''Training script for tokenizer'''

from datasets import load_dataset
from tokenizers import trainers, Tokenizer, normalizers, ByteLevelBPETokenizer
from .utils import model_dir

# load dataset
dataset = load_dataset("oscar", "unshuffled_deduplicated_no", split="train")

# Instantiate tokenizer
tokenizer = ByteLevelBPETokenizer()

def batch_iterator(batch_size=1000):
    for i in range(0, len(dataset), batch_size):
        yield dataset[i: i + batch_size]["text"]

# Customized training
tokenizer.train_from_iterator(
    batch_iterator(),
    vocab_size=50265,
    min_frequency=2,
    special_tokens=["<s>", "<pad>", "</s>", "<unk>", "<mask>"]
)

# Save files to disk
tokenizer_path = model_dir / 'tokenizer.json'
tokenizer.save(str(tokenizer_path))