Although, the paper argues that we can create artificial datasets using the LLMs which would improve the special-case translation models without having to pay for the inference and latency time of the large models.