from datasets import load_dataset from tokenizers import ByteLevelBPETokenizer # load dataset dataset = load_dataset("oscar", "unshuffled_deduplicated_id", 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=["", "", "", "", "",], ) # Save files to disk tokenizer.save(f"./tokenizer.json")