mergekit_config.yml

description: Merging MISCHIEVOUS-12B-Mix models with sliced slerp

# Metadata and Rationale
model_description: |
  This configuration merges two versions of the MISCHIEVOUS-12B-Mix model: 0.4v and 0.3v.
  0.3v was further fine-tuned on a specific dataset (ADD DATASET NAME HERE if known).
  The sliced slerp approach allows for layer-specific control over the merging process.

base_model: bamec66557/MISCHIEVOUS-12B-Mix_0.4v
dtype: bfloat16
merge_method: slerp
tokenizer_source: union

# Slices Configuration (Layer-Specific Merging)
slices:
  - sources:
      - model: bamec66557/MISCHIEVOUS-12B-Mix_0.4v
        layer_range: [0, 10]
      - model: bamec66557/MISCHIEVOUS-12B-Mix_0.5v
        layer_range: [0, 10]
    parameters:
      t:
        - name: self_attn
          value: [0.8, 0.85, 0.9, 0.95, 1.0]
        - name: mlp
          value: [0.9, 0.95, 1.0, 1.05, 1.1]
        - name: layer_norm
          value: [0.6, 0.65, 0.7, 0.75, 0.8]
        - name: embed_tokens
          value: [1.0]

  - sources:
      - model: bamec66557/MISCHIEVOUS-12B-Mix_0.4v
        layer_range: [10, 20]
      - model: bamec66557/MISCHIEVOUS-12B-Mix_0.5v
        layer_range: [10, 20]
    parameters:
      t:
        - name: self_attn
          value: [0.7, 0.75, 0.8, 0.85, 0.9]
        - name: mlp
          value: [1.0, 0.95, 0.9, 0.85, 0.8]
        - name: layer_norm
          value: [0.5, 0.55, 0.6, 0.65, 0.7]
        - name: embed_tokens
          value: [1.0]

  - sources:
      - model: bamec66557/MISCHIEVOUS-12B-Mix_0.4v
        layer_range: [20, 30]
      - model: bamec66557/MISCHIEVOUS-12B-Mix_0.5v
        layer_range: [20, 30]
    parameters:
      t:
        - name: self_attn
          value: [0.6, 0.65, 0.7, 0.75, 0.8]
        - name: mlp
          value: [0.8, 0.75, 0.7, 0.65, 0.6]
        - name: layer_norm
          value: [0.4, 0.45, 0.5, 0.55, 0.6]
        - name: embed_tokens
          value: [1.0]

  - sources:
      - model: bamec66557/MISCHIEVOUS-12B-Mix_0.4v
        layer_range: [30, 40]
      - model: bamec66557/MISCHIEVOUS-12B-Mix_0.5v
        layer_range: [30, 40]
    parameters:
      t:
        - name: self_attn
          value: [0.9, 1.0, 1.1, 1.2, 1.3]
        - name: mlp
          value: [0.7, 0.65, 0.6, 0.55, 0.5]
        - name: layer_norm
          value: [0.7, 0.75, 0.8, 0.85, 0.9]
        - name: embed_tokens
          value: [1.0]

# Regularization (Prevent Overfitting During Merging)
regularization:
  - method: weight_clipping
    clip_range: [-0.2, 0.2]
  - method: random_noise
    scale: 0.015
  - method: l2_norm
    scale: 0.01

# Postprocessing (Enhance Merged Model Quality)
postprocessing:
  - operation: random_noise
    scale: 0.0025
  - operation: non_linear_scaling
    parameters:
      function: tanh
  - operation: sharpening
    intensity: 0.3
  - operation: gaussian_smoothing
    sigma: 1.5
  - operation: smoothing
    parameters:
      adaptive: true
      range: [0.8, 1.2]
      kernel_size: 5
  - operation: normalize
  - operation: dynamic_scaling
    scale_range: [0.75, 1.25]

# Evaluation (Crucial for Assessing Merge Quality)
evaluation:
  metrics:
    - perplexity
    - accuracy # If applicable (e.g., classification tasks)
    - bleu # For translation tasks
    - rouge # For summarization tasks
  datasets:
    - wikitext # General language understanding
    - lambada # Long-range dependency modeling
    - (ADD RELEVANT TASK-SPECIFIC DATASETS HERE)
  prompts: # Example prompts – REPLACE WITH YOUR OWN
    - "The quick brown fox jumps over the lazy dog."
    - "Translate 'Thank you' to Spanish:"
    - "Write a short summary of the French Revolution."

# Logging and Output
logging:
  output_dir: ./merged_models
  log_level: INFO

# Optional: Ties Merging (Advanced Technique)
# ties:
#   enabled: true
#   method: greedy # Or "optimal", "random"
#   layers: [0, 10, 20, 30] # Example layers for ties merging