Aleteian/dark

merge

This is a merge of pre-trained language models created using mergekit.

Merge Details

Merge Method

This model was merged using the della merge method using win10/Mistral-Nemo-abliterated-Nemo-Pro-v2 as a base.

Models Merged

The following models were included in the merge:

• Gryphe/Pantheon-RP-1.5-12b-Nemo
• Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24
• DavidAU/MN-GRAND-Gutenberg-Lyra4-Lyra-12B-DARKNESS
• Aleteian/SaigaPersonalityParty
• DavidAU/MN-GRAND-Gutenberg-Lyra4-Lyra-12B-MADNESS
• elinas/Chronos-Gold-12B-1.0

Configuration

The following YAML configuration was used to produce this model:

models:      
  - model: DavidAU/MN-GRAND-Gutenberg-Lyra4-Lyra-12B-MADNESS
    parameters:
      density: 0.8
      weight: 0.8
      
  - model: DavidAU/MN-GRAND-Gutenberg-Lyra4-Lyra-12B-DARKNESS
    parameters:
      density: 0.8
      weight: 0.8
      
  - model: elinas/Chronos-Gold-12B-1.0
    parameters:
      density: 0.8
      weight: 0.8

  - model: Gryphe/Pantheon-RP-1.5-12b-Nemo
    parameters:
      density: 0.8
      weight: 0.8

  - model: Aleteian/SaigaPersonalityParty
    parameters:
      density: 0.8
      weight: 1.0
      
  - model: Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24
    parameters:
      density: 0.8
      weight: 1.0
      
merge_method: della
base_model: win10/Mistral-Nemo-abliterated-Nemo-Pro-v2

parameters:
  epsilon: 0.10
  lambda: 1.00
  int8_mask: true
  
dtype: float16
chat_template: "chatml"

# 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