Is there actually any dataset of upon which the accuracy of this detector is measured and improved? Or is it just some hard-coded rules?

For example, does it really differentiate between people who ask for votes, and people who just vote up the content of someone they recognize?