mergekit_config.yml

dtype: bfloat16
merge_method: linear
slices:
 - sources:
      - layer_range: [0, 32] # Assuming the first half of the model is more general and can be reduced more
        model: codellama/CodeLlama-7b-Instruct-hf
        parameters:
          weight: 0.25 # Reduce the weight of the first half to make room for the second half
      - layer_range: [0, 32] # Assuming the second half of the model is more specialized and can be reduced less
        model: codellama/CodeLlama-7b-Instruct-hf
        parameters:
          weight: 0.25 # Maintain the weight of the second half