mergekit_config.yml

# =============================================================================
# SuperMerge-14B-Simple
#
# This configuration merges only two components:
#   - Base Model: Provides stable foundational features.
#       Model: sometimesanotion/Lamarck-14B-v0.7-rc4
#
#   - Reasoning Module: Drives enhanced mid-layer reasoning.
#       Model: sthenno/tempesthenno-ppo-ckpt40
#
# The merge is performed using slerp with a V-shaped interpolation curve.
# Weighting across each 8-layer slice is tuned to balance core feature
# preservation with advanced reasoning.
# =============================================================================
name: SuperMerge-14B-Simple
merge_method: slerp
base_model: sometimesanotion/Lamarck-14B-v0.7-rc4
tokenizer_source: base
dtype: float32
out_dtype: bfloat16
parameters:
  int8_mask: true         # Optimize memory usage.
  normalize: true         # Ensure weights are on a comparable scale.
  rescale: false          # No additional rescaling necessary.
  # Interpolation curve for 6 slices (48 layers total):
  # Maintains a V-shaped emphasis for mid-layer processing.
  t: [0.1, 0.35, 0.85, 0.85, 0.35, 0.1]
slices:
  # ---------------------------------------------------------------------------
  # Slice 1 (Layers 0-8):
  #   - Early layers: nearly pure base model with minimal PPO influence.
  # ---------------------------------------------------------------------------
  - sources:
      - model: sometimesanotion/Lamarck-14B-v0.7-rc4
        layer_range: [0, 8]
        parameters:
          weight: 0.95
      - model: sthenno/tempesthenno-ppo-ckpt40
        layer_range: [0, 8]
        parameters:
          weight: 0.05

  # ---------------------------------------------------------------------------
  # Slice 2 (Layers 8-16):
  #   - Blend base with stronger PPO contributions to boost reasoning.
  # ---------------------------------------------------------------------------
  - sources:
      - model: sometimesanotion/Lamarck-14B-v0.7-rc4
        layer_range: [8, 16]
        parameters:
          weight: 0.4
      - model: sthenno/tempesthenno-ppo-ckpt40
        layer_range: [8, 16]
        parameters:
          weight: 0.6

  # ---------------------------------------------------------------------------
  # Slice 3 (Layers 16-24):
  #   - Mid-layer: Prioritize advanced reasoning by increasing the PPO share.
  # ---------------------------------------------------------------------------
  - sources:
      - model: sometimesanotion/Lamarck-14B-v0.7-rc4
        layer_range: [16, 24]
        parameters:
          weight: 0.3
      - model: sthenno/tempesthenno-ppo-ckpt40
        layer_range: [16, 24]
        parameters:
          weight: 0.7

  # ---------------------------------------------------------------------------
  # Slice 4 (Layers 24-32):
  #   - Continue the focus on reasoning with PPO while still retaining base traits.
  # ---------------------------------------------------------------------------
  - sources:
      - model: sometimesanotion/Lamarck-14B-v0.7-rc4
        layer_range: [24, 32]
        parameters:
          weight: 0.35
      - model: sthenno/tempesthenno-ppo-ckpt40
        layer_range: [24, 32]
        parameters:
          weight: 0.65

  # ---------------------------------------------------------------------------
  # Slice 5 (Layers 32-40):
  #   - Re-stabilize the network with a stronger base model contribution.
  # ---------------------------------------------------------------------------
  - sources:
      - model: sometimesanotion/Lamarck-14B-v0.7-rc4
        layer_range: [32, 40]
        parameters:
          weight: 0.6
      - model: sthenno/tempesthenno-ppo-ckpt40
        layer_range: [32, 40]
        parameters:
          weight: 0.4

  # ---------------------------------------------------------------------------
  # Slice 6 (Layers 40-48):
  #   - Final output layers: Maintain fluency with the base model augmented by PPO.
  # ---------------------------------------------------------------------------
  - sources:
      - model: sometimesanotion/Lamarck-14B-v0.7-rc4
        layer_range: [40, 48]
        parameters:
          weight: 0.6
      - model: sthenno/tempesthenno-ppo-ckpt40
        layer_range: [40, 48]
        parameters:
          weight: 0.4