Working transformer implementation

Working transformer implementation. Model trained for ~3 epochs on the original dataset and then fine-tuned on another dataset, performs well on conversational phrases. Achieves a BLEU score of ~27 on the original dataset.