ahmedbr
commented
2 months ago Hi, I have a similar issues actually.
I've just used the encoder-decoder pretrained model to train a diacritizer for a gulf dialect. I didn't make major changes in the script provided, only provided datasets and pretrained model's paths. But as training proceeds, both val_loss and val_der were getting worse and worse. Please have a look at the screenshot below:
Did I do something wrong? Any explanation?
Hello Thanks for your work,
you mentioned in the paper that pretraining phase would help in performing good on colloquial dialects.
I wanted to ask how the model would respond to colloquial dialects something like finetuning the pretrained models on Egyptian dialects.
if well, how much data do i need to finetune the model on to get fine good results on such a case.