Add ram-18m training script (18,290,304 params, LLaMA-style GQA+SwiGLU)

initial commit