MobiLlama: Towards Accurate and Lightweight Fully Transparent GPT