merge_method: model_noiser
models:
  - model: Qwen/Qwen2.5-0.5B-Instruct
parameters:
  normalize: true
  noise_stddev: 0.02  # adjust as needed