Explorations of Self-Repair in Language Models