merge_method: linear # use linear so we can include multiple models, albeit at a zero weight
parameters:
  weight: 1.0 # weight everything as 1 unless specified otherwise - linear with one model weighted at 1 is a no-op like passthrough
slices:
  - sources:
      - model: cognitivecomputations/dolphin-2.9-llama3-8b # embed_tokens comes along with the ride with whatever is the first layer
        layer_range: [0, 1]
      - model:  NousResearch/Meta-Llama-3-8B-Instruct # add dummy second model with 0 weight so tokenizer-based merge routine is invoked for embed_tokens
        layer_range: [0, 1]
        parameters:
          weight: 0
  - sources:
      - model: cognitivecomputations/dolphin-2.9-llama3-8b
        layer_range: [1, 24]
  - sources:
      - model: NousResearch/Meta-Llama-3-8B-Instruct
        layer_range: [8, 24]
        parameters:
          scale:
            - filter: o_proj
              value: 0.0
            - filter: down_proj
              value: 0.0
            - value: 1.0
  - sources:
      - model: cognitivecomputations/dolphin-2.9-llama3-8b
        layer_range: [24, 31]
  - sources: # same as above, but for lm_head with the last layer
      - model: cognitivecomputations/dolphin-2.9-llama3-8b
        layer_range: [31, 32]
      - model: NousResearch/Meta-Llama-3-8B-Instruct
        layer_range: [31, 32]
        parameters:
          weight: 0
dtype: bfloat16
tokenizer_source: model:cognitivecomputations/dolphin-2.9-llama3-8b # keep exact tokenizer used by dolphin - or you could use `union` if you add all of the input models to the first/last slice