Robust Reinforcement Learning from Human Feedback for Large Language Models Fine-Tuning