The intention was to only show recommendations. But because of limited training data (we hand labeled ~4000 comments), the model wasn't able to filter out bad o

The intention was to only show recommendations. But because of limited training data (we hand labeled ~4000 comments), the model wasn't able to filter out bad ones effectively. More training data should be able to solve it.