mergekit_config.yml

base_model: bamec66557/VICIOUS_MESH-12B-OMEGA
dtype: bfloat16
merge_method: slerp
tokenizer_source: base

# Slices Configuration
slices:
  - sources:
      - model: bamec66557/VICIOUS_MESH-12B-OMEGA
        layer_range: [0, 10]
      - model: bamec66557/VICIOUS_MESH-12B-BETA
        layer_range: [0, 10]
    parameters:
      t:
        - name: self_attn
          value: [0.5, 0.55, 0.6, 0.65, 0.7]
        - name: mlp
          value: [1.0, 1.05, 1.1, 1.15, 1.2]
        - name: layer_norm
          value: [0.9, 0.95, 1.0, 1.05, 1.1]

  - sources:
      - model: bamec66557/VICIOUS_MESH-12B-OMEGA
        layer_range: [10, 20]
      - model: bamec66557/VICIOUS_MESH-12B-BETA
        layer_range: [10, 20]
    parameters:
      t:
        - name: self_attn
          value: [0.4, 0.45, 0.5, 0.55, 0.6]
        - name: mlp
          value: [1.1, 1.15, 1.2, 1.25, 1.3]
        - name: layer_norm
          value: [1.0, 1.05, 1.1, 1.15, 1.2]

  - sources:
      - model: bamec66557/VICIOUS_MESH-12B-OMEGA
        layer_range: [20, 30]
      - model: bamec66557/VICIOUS_MESH-12B-BETA
        layer_range: [20, 30]
    parameters:
      t:
        - name: self_attn
          value: [0.6, 0.65, 0.7, 0.75, 0.8]
        - name: mlp
          value: [0.9, 0.95, 1.0, 1.05, 1.1]
        - name: layer_norm
          value: [0.85, 0.9, 0.95, 1.0, 1.05]

  - sources:
      - model: bamec66557/VICIOUS_MESH-12B-OMEGA
        layer_range: [30, 40]
      - model: bamec66557/VICIOUS_MESH-12B-BETA
        layer_range: [30, 40]
    parameters:
      t:
        - name: self_attn
          value: [0.7, 0.75, 0.8, 0.85, 0.9]
        - name: mlp
          value: [0.8, 0.85, 0.9, 0.95, 1.0]
        - name: layer_norm
          value: [0.8, 0.85, 0.9, 0.95, 1.0]

# Regularization
regularization:
  - method: gradient_penalty
    scale: 0.05  # Increased influence for gradient control
  - method: weight_clipping
    clip_range: [-0.2, 0.2]  # Broader clipping range for flexibility
  - method: random_noise
    scale: 0.01  # Stronger noise injection
  - method: attention_dropout
    scale: 0.1  # Higher dropout to reduce attention fixation

# Postprocessing
postprocessing:
  - operation: entropy_regularization
    scale: 0.05  # Stronger encouragement for diverse outputs
  - operation: non_linear_scaling
    parameters:
      function: tanh
  - operation: sharpening
    intensity: 0.5  # Enhanced sharpening for precise outputs
  - operation: gaussian_smoothing
    sigma: 1.5  # Increased smoothing for stable outputs
  - operation: normalize
  - operation: dynamic_scaling
    scale_range: [0.8, 1.2]  # Expanded dynamic range for scaling
  - operation: smoothing
    parameters:
      adaptive: true
      range: [0.85, 1.15]  # Wider adaptive smoothing range
      kernel_size: 5