It looks like according to the notebook recepies that it's actually working at the sentence level (despite your claim). The way you train these is to mark a sentence as "1" or "0" based on if it's selected.

Oh well, the world waits longer for a word-level (willing to skip around in sentences) and grammatically correct extractive summarizer. You got my hopes up though.

https://github.com/microsoft/nlp-recipes/blob/master/example...

My (failed) attempt to build what I was actually looking for:

https://github.com/Hellisotherpeople/CX_DB8