The samples were selected randomly at the beginning. Tags were thought up on the go, trying to generalize into broader categories. After the initial tagging, we counted the number of samples for each category and grouped similar underrepresented tags together + added additional samples for tagging that could go into the smaller categories just by filtering those who matched specific keywords and further hand-picking them.

We initially tried training the classifier only with GitHub based samples and using the user-given tags from there. Although we grouped the tag base into a reasonable number of distinct categories, the way how GitHub users tag their projects turned out to be just too inconsistent and often unrelated to the titles, so manual tagging was seen as a better option for getting decent results fast enough.

If you have any more specific questions feel free to drop me a mail to arturs@finch.io