The bayesian filter requires training before it can effectively differentiate between spam and ham. The training database currently contains ${nspam} spam and ${nham} ham submissions.
Resetting the training database can help when training was incorrect and is producing bad results.
The minimum number of spam and ham in the training database before the filter starts affecting the karma of submissions.
While you can train the spam filter from the “Spam Filtering → Monitoring” panel in the web administration interface, you can also manually train the filter by entering samples here, or check what kind of spam probabilty currently gets assigned to the content.