|
from datasets import load_dataset |
|
from tokenizers import ByteLevelBPETokenizer |
|
|
|
model_dir = "../gpt-2-tamil" |
|
|
|
|
|
dataset = load_dataset("oscar", "unshuffled_deduplicated_ta", split="train") |
|
|
|
|
|
tokenizer = ByteLevelBPETokenizer() |
|
|
|
|
|
def batch_iterator(batch_size=1000): |
|
for i in range(0, len(dataset), batch_size): |
|
yield dataset[i : i + batch_size]["text"] |
|
|
|
|
|
|
|
tokenizer.train_from_iterator( |
|
batch_iterator(), |
|
vocab_size=50265, |
|
min_frequency=2, |
|
special_tokens=[ |
|
"<s>", |
|
"<pad>", |
|
"</s>", |
|
"<unk>", |
|
"<mask>", |
|
], |
|
) |
|
|
|
|
|
tokenizer.save(f"{model_dir}/tokenizer.json") |
|
|