Getting a accuracy of 100% on validation set but model giving awful predictions(ULMFit)

I am using the ULMFit model on amazon review data. I am using around 300000 datapoints equally divided among positive and negative. My Language Model is getting an accuracy of 0.75 while the accuracy on validation set for classifier model is 1.00.

But the model is giving very bad predictions.

learn.predict("This product is awesome. I really love it")
(Category 0, tensor(0), tensor([0.9667, 0.0333]))


learn.predict("Horrible product. I would never use it.")
(Category 0, tensor(0), tensor([0.9667, 0.0333]))

Google Colab Notebook Link:

Can someone help me out?