merge

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

Merge Details

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

• bamec66557/MNRP_0.5
• bamec66557/MISCHIEVOUS-12B

Configuration

The following YAML configuration was used to produce this model:

slices:
  - Sources:
      - model: bamec66557/MNRP_0.5
        layer_range: [0, 40] # Merge layer range for MNRP_0.5 model
      - model: bamec66557/MISCHIEVOUS-12B
        layer_range: [0, 40] # Merge layer range for MISCHIEVOUS-12B model.

    # Adjust the merge ratio per layer to drive smoother integration
    # Each filter affects a specific mechanism within the model
    parameters:
      t:
        - Filter: self_attn
          value: [0.2, 0.4, 0.6, 0.8, 1.0] # Progressive merging of self-attention layers
        - filter: mlp
          value: [0.8, 0.6, 0.4, 0.2, 0.0] # Merge MLP layers with opposite proportions
        - filter: layer_norm
          value: [0.5, 0.5, 0.5, 0.5, 0.5, 0.5] # Layer Normalisation should be merged uniformly
        - value: 0.7 # Default

merge_method: slerp # change merge method to slerp

base_model: bamec66557/MISCHIEVOUS-12B # base model for merge

dtype: bfloat16 # data type for efficient and fast operations when merging

# Additional available options
regularisation:
  - method: l2_norm # Stabilise merged model weights with L2 normalisation
    scale: 0.01

postprocessing:
  - operation: smoothing # Smooth the weights after merging
    kernel_size: 3
  - operation: normalise # normalise the overall weights

Open LLM Leaderboard Evaluation Results

Detailed results can be found here