---
language:
- en
license: apache-2.0
---

This is an experimental self-merge of [miqu-1-70b-sf](https://huggingface.co/152334H/miqu-1-70b-sf)
using @jukofyork idea of [Downscaling the K and/or Q matrices for repeated layers in franken-merges](https://github.com/arcee-ai/mergekit/issues/198).
More info about the _attenuation_ is available in this [discussion](https://huggingface.co/wolfram/miqu-1-120b/discussions/4)

In my own [LLM Creativity benchmark](https://huggingface.co/datasets/froggeric/creativity) it performs sightly
better than the original [wolfram/miqu-1-120b](https://huggingface.co/wolfram/miqu-1-120b). Specifically,
I noticed some **improvements at creative writing, producing longer, more details, and unrushed text**.
Like wolfram/miqu-1-120b though, there is some degradation over miqu-1-70b, with longer text, as it starts
deviating from instructions and requires some effort to keep it on track.

Use gguf-split to join the different parts.

## Model Details

- Max Context: 32764 tokens (kept the weird number from the original/base model)
- Layers: 140

### Prompt template: Mistral

```
<s>[INST] {prompt} [/INST]
```

See also: [🐺🐦‍⬛ LLM Prompt Format Comparison/Test: Mixtral 8x7B Instruct with **17** different instruct templates : LocalLLaMA](https://www.reddit.com/r/LocalLLaMA/comments/18ljvxb/llm_prompt_format_comparisontest_mixtral_8x7b/)

## Merge details with mergekit

```yaml
###############################
# miqu-1-120b-attenuated.yaml #
###############################

# Use: mergekit-yaml --clone-tensors ./miqu-1-120b-attenuated.yaml ./miqu-1-120b-attenuated

# See: https://huggingface.co/wolfram/miqu-1-120b for original 'miqu-1-120b' layer ranges.
# See: https://github.com/arcee-ai/mergekit/issues/198 for discussion/reasoning behind this idea.

# ---

# The scale factor to use, eg: solve x^2 = 1/2 --> x = 1/sqrt(2) ≈ 0.7071067812
const_tag: &scale_factor 0.7071067812 # 1/sqrt(2)

# The filter parameters of a scaled block. 
attenuate-env: &attenuated_env
  parameters:
    scale:
      - filter: q_proj
        value: *scale_factor
      - filter: k_proj
        value: *scale_factor
      - value: 1.0

# ---

slices:

  ###########################
  # Block 1: miqu-1 [0, 20] #
  ###########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [0, 10] # The first 10 layers of Block 1 are not duplicated
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [10, 20] # The last 10 layers of Block 1 are are duplicated twice
      <<: *attenuated_env

  ###########################
  # Block 2: miqu-1 [10, 30] #
  ###########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [10, 30]
      <<: *attenuated_env

  ###########################
  # Block 3: miqu-1 [20, 40] #
  ###########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [20, 40]
      <<: *attenuated_env

  ###########################
  # Block 4: miqu-1 [30, 50] #
  ###########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [30, 50]
      <<: *attenuated_env

  ###########################
  # Block 5: miqu-1 [40, 60] #
  ###########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [40, 60]
      <<: *attenuated_env

  ###########################
  # Block 6: miqu-1 [50, 70] #
  ###########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [50, 70]
      <<: *attenuated_env

  ##########################
  # Block 7: miqu-1 [60, 80] #
  ##########################
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [60, 70] # The first 10 layers of Block 7 are are duplicated twice
      <<: *attenuated_env
  - sources:
    - model: miqu-1-70b-sf
      layer_range: [70, 80] # The last 10 layers of Block 7 are not duplicated

merge_method: passthrough
dtype: float16
```