Use the instruction-tuned model as your base:

base_model = "google/gemma-3-270m-it"
it_model = "google/gemma-3-270m-it"

Once training + inference use the same chat template, the repetition stops.