TAPAS finetuning feature not available ?

[Arka01]

Hello Everyone,

I have used haystack for finetuning BERT model from Hugging Face using Farm Reader.
But not able to finetune TAPAS model from Hugging Face using Farm Reader or Table Reader.

Not sure if I am doing something wrong or the TAPAS finetuning is just not a supported feature at the moment.

Thank You in advance !

[vblagoje]

Hey @Arka01 ,

I haven't fine-tuned the TAPAS model before, nor do I think we support this feature, but I'll ask a colleague and come back to you. I am wondering why you would need to fine-tune the TAPAS model? What is your use case?

[Arka01]

Hi @vblagoje ,

Thank you for the reply.

Just like we fine tune BERT with Question Answers specific to our dataset,
I was trying to finetune TAPAS with Queries and Tables (tables which have the answer to the queries).

Something like this https://huggingface.co/transformers/v4.8.0/model_doc/tapas.html#usage-fine-tuning
but inside the HayStack framework.

Do let me know if I could explain my requirement clearly
and if there is something in HayStack which would allow
me to finetune the model with a dataset of my own.

Thanks in advance !!!

[vblagoje]

I am afraid Haystack wouldn't help you there. What kind of a dataset do you have, or what kind of a TAPAS task are you dealing with (out of the three mentioned in the article ☝️ )?

[Arka01]

Out of the three mentioned in that link, I would say it's Weak supervision for aggregation,
although we are just expecting the model to answer questions/queries with data from table cells. (aggregation results would maybe be our next phase.)

And about the task/dataset, it's actually just a number of tables having different columns (some overlap might is there), from which the user is expected to get a cell (from the correct table) as the reply.