README.md

Mistral-Merge-7B-slerp

Model Description

The Mistral-Merge-7B-slerp is a merged model which leverages the spherical linear interpolation (SLERP) technique to blend layers from two distinct transformer-based models. This merging strategy is aimed at synthesizing a model that incorporates the robust linguistic capabilities of OpenPipe/mistral-ft-optimized-1218 and the nuanced understanding of mlabonne/NeuralHermes-2.5-Mistral-7B.

Configuration

The merging process was configured to apply a SLERP method across all comparable layers of the two source models. Below is the YAML configuration used for merging:

slices:
  - sources:
      - model: OpenPipe/mistral-ft-optimized-1218
        layer_range: [0, 32]
      - model: mlabonne/NeuralHermes-2.5-Mistral-7B
        layer_range: [0, 32]
merge_method: slerp
base_model: OpenPipe/mistral-ft-optimized-1218
parameters:
  t:
    - filter: self_attn
      value: [0, 0.5, 0.3, 0.7, 1]
    - filter: mlp
      value: [1, 0.5, 0.7, 0.3, 0]
    - value: 0.5
dtype: bfloat16

This configuration ensures that both self-attention and MLP (multi-layer perceptron) layers undergo interpolation with a gradient of weights to optimize the integration of features from both models.