froggeric/miqu-a0-wlfrm-120b

This is an experimental self-merge of miqu-1-70b-sf using @jukofyork idea of Downscaling the K and/or Q matrices for repeated layers in franken-merges. More info about the attenuation is available in this discussion

In my own LLM Creativity benchmark it performs sightly better than the original wolfram/miqu-1-120b. Specifically, I noticed some improvements at creative writing, producing longer, more details, and unrushed text. Like wolfram/miqu-1-120b though, there is some degradation over miqu-1-70b, with longer text, as it starts deviating from instructions and requires some effort to keep it on track.

Use gguf-split to join the different parts.

Model Details

• Max Context: 32764 tokens (kept the weird number from the original/base model)
• Layers: 140

Prompt template: Mistral

<s>[INST] {prompt} [/INST]

See also: 🐺🐦‍⬛ LLM Prompt Format Comparison/Test: Mixtral 8x7B Instruct with 17 different instruct templates : LocalLLaMA

Merge details with mergekit

###############################
# miqu-1-120b-attenuated.yaml #
###############################

# Use: mergekit-yaml --clone-tensors ./miqu-1-120b-attenuated.yaml ./miqu-1-120b-attenuated

# See: https://huggingface.co/wolfram/miqu-1-120b for original 'miqu-1-120b' layer ranges.
# See: https://github.com/arcee-ai/mergekit/issues/198 for discussion/reasoning behind this idea.

# ---

# The scale factor to use, eg: solve x^2 = 1/2 --> x = 1/sqrt(2) ≈ 0.7071067812
const_tag: &scale_factor 0.7071067812 # 1/sqrt(2)

# The filter parameters of a scaled block. 
attenuate-env: &attenuated_env
  parameters:
    scale:
      - filter: q_proj
        value: *scale_factor
      - filter: k_proj
        value: *scale_factor
      - value: 1.0

# ---

slices:

  ###########################
  # Block 1: miqu-1 [0, 20] #
  ###########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [0, 10] # The first 10 layers of Block 1 are not duplicated
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [10, 20] # The last 10 layers of Block 1 are are duplicated twice
      <<: *attenuated_env

  ###########################
  # Block 2: miqu-1 [10, 30] #
  ###########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [10, 30]
      <<: *attenuated_env

  ###########################
  # Block 3: miqu-1 [20, 40] #
  ###########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [20, 40]
      <<: *attenuated_env

  ###########################
  # Block 4: miqu-1 [30, 50] #
  ###########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [30, 50]
      <<: *attenuated_env

  ###########################
  # Block 5: miqu-1 [40, 60] #
  ###########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [40, 60]
      <<: *attenuated_env

  ###########################
  # Block 6: miqu-1 [50, 70] #
  ###########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [50, 70]
      <<: *attenuated_env

  ##########################
  # Block 7: miqu-1 [60, 80] #
  ##########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [60, 70] # The first 10 layers of Block 7 are are duplicated twice
      <<: *attenuated_env
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [70, 80] # The last 10 layers of Block 7 are not duplicated

merge_method: passthrough
dtype: float16