Fine tune on specific domain data? Is it needed?

[agario8864]

Currently I have scrapped some wikipedia articles, in political domain. Say that I am using FARMReader with a specific language model that have pretrained with SQuAD v2, and it do pretty well in answering my question about political domain, should I fine tune the FARMReader with the articles I found?

I also confuse with what meant by dataset in question answering. If I am not going to use the scarpped articles as a dataset (like SQuAD format) to train the Reader but just store as documents in a document store, that's mean I am not tranining a model, but using others' pretrained model to form a question answering system, right?

Apperciate for any answer.

[bogdankostic]

Hi @agario8864! If you found a model that's already working good enough for your domain, you don't need to fine-tune the reader model. If you're still interested in fine-tuning and how to produce labeled training data, this guide in our documentation might be interesting for you.

If I am not going to use the scarpped articles as a dataset (like SQuAD format) to train the Reader but just store as documents in a document store, that's mean I am not tranining a model, but using others' pretrained model to form a question answering system, right?

Yes, that's right, you're not training a model but using the data as knowledge base for your question-answering system.

[agario8864]

@bogdankostic alright, tq