grimjim's picture
Update README.md
8855fe4 verified
metadata
base_model:
  - SanjiWatsuki/Kunoichi-DPO-v2-7B
  - SanjiWatsuki/Kunoichi-7B
library_name: transformers
tags:
  - mergekit
  - merge
license: cc-by-nc-4.0
pipeline_tag: text-generation

franken-kunoichi-IDUS-11B

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

The Interwoven Depth Up-Scaling merge formula was adapted from Sanji Watsuki's longcat-10.7B.

I consider this to be a negative result, but perhaps an interesting one. I've tested casually with temperature 0.7-1.2 and minP 0.01-0.03, with both Alpaca and ChatML prompts. The resulting generated text is interesting for RP due to chaotic variation, mostly sticking to grammatically correct output, but easily veers too far into chaos (e.g., abruptly switching language) while having difficulty tracking things. Given that, the inherited 8K context length is of dubious benefit.

Maybe additional training might smooth the transition between models, but that hypothesis is untested.

Merge Details

Merge Method

This model was merged using the passthrough merge method.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

slices:
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [0, 8]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [8, 9]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [8, 9]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [9, 10]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [9, 10]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [10, 11]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [10, 11]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [11, 12]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [11, 12]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [12, 13]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [12, 13]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [13, 14]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [13, 14]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [14, 15]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [14, 15]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [15, 16]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [15, 16]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [16, 17]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [16, 17]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [17, 18]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [17, 18]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [18, 19]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [18, 19]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [19, 20]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [19, 20]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [20, 21]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [20, 21]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [21, 22]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [21, 22]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [22, 23]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [22, 23]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [23, 24]
  - sources:
    - model: SanjiWatsuki/Kunoichi-7B
      layer_range: [23, 24]
  - sources:
    - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
      layer_range: [24, 32]
merge_method: passthrough
dtype: float16