train_tokenizer.py

#!/usr/bin/env python3
from datasets import load_dataset
from datasets import load_from_disk
from tokenizers import ByteLevelBPETokenizer
from tqdm import tqdm
# load dataset
# dataset = load_dataset("oscar", "unshuffled_deduplicated_hi", split="train")

dataset = load_from_disk("/home/rtx/work/dk/hf/vo")

# Instantiate tokenizer
tokenizer = ByteLevelBPETokenizer(add_prefix_space=True)

def batch_iterator(batch_size=100_000):
    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=50, special_tokens=[
    "<s>",
    "<pad>",
    "</s>",
    "<unk>",
    "<mask>",
])

# Save files to disk
tokenizer.save("./tokenizer.json")