Mistral-7B-Instruct-demi-merge-v0.2-7B

This is a merge of pre-trained Mistral v0.2 language models created using mergekit. Both models have 32K context length without a sliding window, as does this merge.

The base model as released felt like it needed more training. The instruct model felt fully trained. This merge aimed to be a compromise between the two, to enable future merging and/or fine-tuning by thawing out the Instruct model while keeping some of its strengths.

Light testing has confirmed that the resulting merged model can operate to the full 32K context length, exceeding the 8K context that characterized the Mistral-7B-0.1 release.

Merge Details

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

• alpindale/Mistral-7B-v0.2-hf
• mistralai/Mistral-7B-Instruct-v0.2

Configuration

The following YAML configuration was used to produce this model:

slices:
  - sources:
    - model: mistralai/Mistral-7B-Instruct-v0.2
      layer_range: [0,32]
    - model: alpindale/Mistral-7B-v0.2-hf
      layer_range: [0,32]
merge_method: slerp
base_model: mistralai/Mistral-7B-Instruct-v0.2
parameters:
  t:
    - value: 0.5
dtype: bfloat16