[Question] Question about the tokenizer of required pretrained model stabilityai/stablelm-2-1_6

[Taylorfire]

Question

Thanks for your excellent work! When I try to fine-tune with the LLM as StableLM1.6B, I am confused about the tokenizer inconsistency.

As the ./scripts/stablelm/finetune.sh requires, I download the pretrained LLM "stabilityai/stablelm-2-1_6" from huggingface. The tokenizer_config.json indicates that the tokenizer belongs to the class "GPT2TokenizerFast". While in your code moellava/train/train.py the tokenizer class for stablelm is Arcade100kTokenizer. Thus, this inconsistency leads to the failure of loading the tokenizer.

Can you tell me if there is any wrong with my implementation? Should I still use the "stabilityai/stablelm-2-1_6" as the pretrained LLM?

[2002DQJ]

There is no avaliable API according Huggignface because of development team's mistake, you can search "Auto classes" from Google and get correct API,and then choose the avaliable VLM model.