possible problem with the translate_ende checkpoint in google colab

[roee11]

I'm trying to use the checkpoint for the translate_ende_wmt32k problem and then train it on my own data.
Link to google colab page:
https://colab.research.google.com/github/tensorflow/tensor2tensor/blob/master/tensor2tensor/notebooks/hello_t2t.ipynb

I have 2 questions:

1. It seems the checkpoint is problematic - in the example in above colab, the example results of translation are:
   INFO:tensorflow:Greedy Decoding
   Inputs: The animal didn't cross the street because it was too tired
   Outputs: Das Tier überquerte die Straße nicht, weil es zu müde war, weil es zu müde war.

It repeats the second part of the sentence twice for some reason. I also tried it on my machine on other sentences and it happened again.. I understand the checkpoint isn't supposed to give perfect translation but is it supposed to be like that? Should I try to train it more on the WMT data?

2. I saw in other issues(for example Unable to reproduce WMT En2De results #317 ) they recommend using checkpoints gs://tensor2tensor-checkpoints/transformer_ende_test/model.ckpt-3817425 or gs://tensor2tensor-checkpoints/transformer_ende_test/averaged.ckpt-0
   but it seems they're not there anymore. Can they be found elsewhere or are they deleted for some reason?

Thanks a lot!