| from datasets import load_dataset | |
| from transformers import AutoTokenizer | |
| def get_training_corpus(dataset): | |
| """ | |
| Returns the training corpus for the given dataset. | |
| """ | |
| return (element['original_ja'] for element in iter(dataset)) | |
| dataset = load_dataset("snow_simplified_japanese_corpus", streaming=True, split="train") | |
| train_dataset = dataset.skip(100) | |
| val_dataset = dataset.take(100) | |
| old_tokenizer = AutoTokenizer.from_pretrained("rinna/japanese-gpt2-small") | |
| old_tokenizer = AutoTokenizer.from_pretrained("csebuetnlp/mT5_multilingual_XLSum") | |
| print("Old Tokenizer:", old_tokenizer.tokenize("誰が一番に着くか私には分かりません。")) | |
| new_tokenizer = old_tokenizer.train_new_from_iterator(get_training_corpus(train_dataset), 52000) | |
| print("New Tokenizer:",new_tokenizer.tokenize("誰が一番に着くか私には分かりません。")) | |
| new_tokenizer.save_pretrained("japanese-dummy-tokenizer") |