How to implement NER model using trax about trax HOT 5 OPEN

I've had a go with that notebook but I can't figure out how to use this model to do NER on a new sentence. (I've posted question on stackexchange: https://ai.stackexchange.com/questions/25442/using-reformer-model-to-do-ner-in-trax)

One thing that puzzles me about the 85% accuracy on this test. Given the large proportion of tags that are 'O', is 85% accuracy really that good? The LSTM example that Coursera uses achieves 90-95% accuracy, and I would have hoped that Reformer is better.

from trax.

We're working on providing more examples, will update here when there's one for NER.

from trax.

Hey, I've been working on this colab notebook for implementing NER using Reformer.

Please go through it and tell suggestions.

from trax.

On training for 100 steps, using kaggle kernels I got an accuracy of 85%.

You can find the notebook here

from trax.

@SauravMaheshkar could you please offer some help regarding predictions using the NER_Reformer code that you have demonstrated ?
You help will be highly appreciated.

from trax.