Post-Training Large Language Models via Reinforcement Learning from Self-Feedback