Defining Text Multi-label Classification Problem

[Aleksas Pielikis]

Trying to define a text multi-label classification problem in t2t with kaggle jigsaw toxic comment challenge data . Training tensorflow_encoder doesn't seem to converge. Here is a generic multilabel and a jigsaw challenge problem definitions and script for running training. I'm not sure if I've defined problem incorrectly or my metaparams are bad. Any suggestions would be appreciated. 
(question also posted on gitter - no suggestions)