Right, the model is not perfect with limited training dataset I have (we hand labeled 4,000 - which is already tons of work for a side project). But the intention was to filter out negative ones.
On labeling, if you have a method statement or some go-by referance I am sure you would get some support here - I know I would help ! Maybe package a few blocks of 100 unlabeled comments with a readme & see what happens ?