merge_method: linear
dtype: bfloat16
models:
  - model: google/gemma-3-4b-pt
    parameters:
      weight: 1.0
  - model: google/gemma-3-4b-it
    parameters:
      weight: 0.0
tokenizer:
  source: google/gemma-3-4b-it
  tokens:
    <start_of_turn>:
      source: google/gemma-3-4b-it
      force: true
    <end_of_turn>:
      source: google/gemma-3-4b-it
      force: true