Experience of Training a 1.7B-Parameter LLaMa Model From Scratch