Reversing Large Language Models for Efficient Training and Fine-Tuning