A software company is training a model to tag customer support tickets as "angry" or "not angry". Early tests show the model is inconsistent, and a review finds that different labelers tagged very similar tickets differently. What should the team do FIRST?