Text classification worse than random and not improving

I am trying to classify some text articles as low, medium or high quality however the model is useless and training it is making it worse rather than improving.

While training my language model it starts out very accurate which is surprising since the articles are quite technical however once unfreezing the model and training the whole model the accuracy gets worse as well as the training and validation loss.
image
image

The plot of losses is also very confusing.
image

This training was taking a very long time so I decided to try and classify the documents to see if anything would happen. I used 1e-1 for the learning rate based off this plot.
image

After starting to train my classifier I got a terrible result with no idea what I am doing wrong.
image

Accuracy before unfreeze seems to work better. Have you tried running a more epochs on that first before unfreezing?

Have fun!

It’s possible that the learning rate is too big for when you unfreeze, it changes the weights too quickly. I’d try lowering this substantially when unfreezing.
Is the learning rate chart from after unfreezing?
Have you tried gradual unfreezing as in the IMDB example from the courses?