versae's picture
Scripts for perplexity sampling and fixes
raw history blame
No virus
649 Bytes
#!/usr/bin/env python3
from datasets import load_dataset
from tokenizers import ByteLevelBPETokenizer
# Load dataset
dataset = load_dataset("oscar", "unshuffled_deduplicated_es", split="train[:5000000]")
# Instantiate tokenizer
tokenizer = ByteLevelBPETokenizer()
def batch_iterator(batch_size=100_000):
for i in range(0, len(dataset), batch_size):
yield dataset["text"][i: i + batch_size]
# Customized training
tokenizer.train_from_iterator(batch_iterator(), vocab_size=50265, min_frequency=2, special_tokens=[
# Save files to disk"./tokenizer.json")