At this point the tokenizer_name_or_path argument is part of the peft config which should not be the case since it can be used irrespective of using peft config like in the case of providing a custom tokenizer. Given this, we should have a separate field to provide custom tokenizer path may be part of the model args dataclass will allow users to use custom tokenizer which can be passed while creating the tokenizer instance.
If not provided this can be initialized to model_name_or_path when not provided which can be handled in the __post_init__ lifecycle of the dataclass.
At this point the
tokenizer_name_or_path
argument is part of the peft config which should not be the case since it can be used irrespective of using peft config like in the case of providing a custom tokenizer. Given this, we should have a separate field to provide custom tokenizer path may be part of the model args dataclass will allow users to use custom tokenizer which can be passed while creating the tokenizer instance.If not provided this can be initialized to
model_name_or_path
when not provided which can be handled in the__post_init__
lifecycle of the dataclass.