How to Finetune Deberta Model on SQUAD dataset?

Hi All,

I am trying to finetune DeBerta Model on SQUAD. If I try existing notebooks like this It uses Trainer and Fast Tokenizer.

DeBerta doesn’t have support for Fast Tokenizer Yet. How can I finetune it on SQUAD?

I am also willing to implement Fast Tokenizer for the Deberta model, Can anyone help me with resources so I can get started with that?

Here is my notebook for training Deberta (I am facing issues)

Hi @bhadresh-savani, as far as I can tell the problem seems to lie with your find_sublist_indices function, not on the availability of a fast tokenizer.

One simple thing to try: can you pass a slice of examples to your convert_to_features function, e.g.


I’m not sure whether this will solve the problem, but perhaps your find_sublist_indices is expected a list of lists which is what you’ll get from the slice.

I also noticed that your convert_to_features function is quite different to the prepare_train_features in the tutorial - what happens if you try the latter with your tokenizer?

If that doesn’t work, then you might be able to use the old script that doesn’t rely on fast tokenizers: transformers/examples/legacy/question-answering at master · huggingface/transformers · GitHub


1 Like

Hi @lewtun

Thanks for your answer

I tried to run old version v3.5.1 by keeping latest version of file with few changes (i needed QuestionAnsweringModelOutput class for SQUAD kind of training)

I was getting below error

    Traceback (most recent call last):
  File "", line 820, in <module>
  File "", line 734, in main
    model = AutoModelForQuestionAnswering.from_pretrained(
  File "/media/data2/anaconda/envs/transformers-hugginface/lib/python3.8/site-packages/transformers/", line 1330, in from_pretrained
    raise ValueError(
ValueError: Unrecognized configuration class <class 'transformers.configuration_deberta.DebertaConfig'> for this kind of AutoModel: AutoModelForQuestionAnswering.
Model type should be one of DistilBertConfig, AlbertConfig, CamembertConfig, BartConfig, LongformerConfig, XLMRobertaConfig, RobertaConfig, SqueezeBertConfig, BertConfig, XLNetConfig, FlaubertConfig, MobileBertConfig, XLMConfig, ElectraConfig, ReformerConfig, FunnelConfig, LxmertConfig. 

find_sublist_indices i created by taking ref of this notebook which uses a fast tokenizer, I am trying to do the same without fast tokenizer

Fast tokenizer has method called char_to_token i am trying to implement the same on Python based tokenizer.

Hi @valhalla,

Can you tell me how can i use the same notebook without fast tokenizer since it was created by you?