zephyr-wizard-kuno-royale-BF16-merge-7B

This is an experimental merge of pre-trained language models created using mergekit. All source model weights are BF16, avoiding issues arising from mixed-precision merges.

Although the zephyr beta and WizardLM 2 7B models are touted as SOTA and can generate varied prose compared to base Mistral v0.1, their relatively mediocre benchmarks under GSM-8K suggests only average reasoning capability in one-shot narrative text completion. The kuno-royale-v2 model was selected for merger because of its higher GSM-8K rating.

Native prompt format is Alpaca, although at least one of the prior models was fine-tuned to Vicuna.

Tested lightly with ChatML instruct prompts, temperature 1, and minP 0.02.

• Full weights: grimjim/zephyr-wizard-kuno-royale-BF16-merge-7B
• GGUF quants: grimjim/zephyr-wizard-kuno-royale-BF16-merge-7B-GGUF

Merge Details

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

• grimjim/zephyr-beta-wizardLM-2-merge-7B
• core-3/kuno-royale-v2-7b

Configuration

The following YAML configuration was used to produce this model:

slices:
  - sources:
    - model: grimjim/zephyr-beta-wizardLM-2-merge-7B
      layer_range: [0,32]
    - model: core-3/kuno-royale-v2-7b
      layer_range: [0,32]
merge_method: slerp
base_model: grimjim/zephyr-beta-wizardLM-2-merge-7B
parameters:
  t:
    - value: 0.5
dtype: bfloat16