use a different base model and trained a bit longer, using 1200 checkpoint.