Probably something like a small logistic regression or a tiny GPT-2 variant (117M parameters) on a small dataset—anything beyond that will choke on RAM, VRAM, or time. Five minutes on a laptop = toy models, not miracles.