Training with custom dataset

[vivekam101]

Hi Kamal,
Can you please share how to do finetuning with custom dataset

[kamalkraj]

The pre-trained model comes from https://github.com/huggingface/transformers.
You can refer to there documentation for fine-tuning on a custom dataset.
After finetuning you can use this repo for inference. Point the code to new model

[gurvesh]

Thanks for for providing this Kamal!

For anyone else interested, I was able to get better performance with the following model which is provided by the transformer library directly (need to edit bert.py):

tokenizer = BertTokenizer.from_pretrained('bert-large-uncased-whole-word-masking-finetuned-squad')
model = BertForQuestionAnswering.from_pretrained('bert-large-uncased-whole-word-masking-finetuned-squad', config=config)

[kamalkraj]

Hi @gurvesh,
I think the new model is recently released by transformers. Thanks for posting here.
I will add this information to README

What are EM and F1 Score of bert-large-uncased-whole-word-masking-finetuned-squad ?

[gurvesh]

I think the scores are about the same. But in actual usage, I found the answers given by the finetuned-squad model on random articles I put through the models were much better. YMMV