Question about fine tune.

[JunleiYang]

Hello, there
Currently, I have a task to classify texts to two classes, blocked and approved. This task is quite like Toxic Comment Classification Challenge. But there are some slight different with this data set. Can I train my own data with the pre-trained model?

Thanks a lot.

[laurahanu]

Hello!

Yes, should work, would probably have to load the detoxify pre-trained model weights and change the final layer so it predicts only 2 classes and just use binary cross_entropy as a loss. Should also change the sigmoid activation function to a softmax.

Hope this helps!