google

/

gemma-7b

like 3.05k

Follow

Google 4,155

Hi,
I'm curious to know if there's any difference between AutoTokenizer and GemmaTokenizer. I ran the following code and noticed a slight difference in the vocabularies of them:

from transformers import GemmaTokenizer, AutoTokenizer

model_name = "google/gemma-7b"

auto_tokenizer = AutoTokenizer.from_pretrained(model_name)
gemma_tokenizer = GemmaTokenizer.from_pretrained(model_name)

print(set(auto_tokenizer.get_vocab().keys()) - set(gemma_tokenizer.get_vocab().keys()))
print(set(gemma_tokenizer.get_vocab().keys()) - set(auto_tokenizer.get_vocab().keys()))

I got the following results by running the code on Colab:

{'\t', '<start_of_image>'}
{'<0x09>', '<unused99>'}

** Update: I think the issue is related to GemmaTokenizerFast, while GemmaTokenizer works fine.