Unable to reproduce the results on MathQA without CL

[lijierui]

I'm using the processed data you provided to reproduce the results for MathQA.
Following the instructions, I replace the vocab.txt for bert-base-uncased folder. I use train_ft_monolingual-en.sh. However, the accuracy is around 27%. When I tried math23K instead, the results match the paper.
I wonder where could go wrong. Is there anything specifically for math23k or Chinese dataset that needs modifying when running the code? Or is something about the bert model I need to pay attention to?

thanks!

[zwx980624]

Try changing the learning-rate to 5e-5 and see if it works.