Train Discrete BFN with Larger Vocabulary?

[LtECoD]

Hi,
The discrete BFN presented in the paper has demonstrated competitive performance on the text8 dataset. However, the vocabulary size of text8, which stands at a mere 27, is considerably limited for most NLP tasks. I am curious to know if you have experimented with training discrete BFN models on datasets with a larger vocabulary. If that is the case, could you provide some insights into the model's architecture, settings of hyper parameters, and the performance achieved?
Thanks!