rAIfle/maid-yuzu-v8-alter-exl2-rpcal

Quantized using 200 samples of 8192 tokens from an RP-oriented PIPPA dataset.

Branches:

• main -- measurement.json
• 6b6h -- 6bpw, 6bit lm_head
• 5b6h -- 5bpw, 6bit lm_head
• 3.5b6h -- 3.5bpw, 6bit lm_head
• 2.25b6h -- 2.25bpw, 6bit lm_head

Requires ExllamaV2 version 0.0.12 and up.

Original model link: rhplus0831/maid-yuzu-v8-alter

Original model README below.

---

maid-yuzu-v8-alter

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

v7's approach worked better than I thought, so I tried something even weirder as a test. I don't think a proper model will come out, but I'm curious about the results.

Merge Details

Merge Method

This model was merged using the SLERP merge method.

This models were merged using the SLERP method in the following order:

maid-yuzu-v8-base: mistralai/Mixtral-8x7B-v0.1 + mistralai/Mixtral-8x7B-Instruct-v0.1 = 0.5
maid-yuzu-v8-step1: above + jondurbin/bagel-dpo-8x7b-v0.2 = 0.25
maid-yuzu-v8-step2: above + cognitivecomputations/dolphin-2.7-mixtral-8x7b = 0.25
maid-yuzu-v8-step3: above + NeverSleep/Noromaid-v0.4-Mixtral-Instruct-8x7b-Zloss = 0.25
maid-yuzu-v8-step4-alter: above + ycros/BagelMIsteryTour-v2-8x7B = 0.5
maid-yuzu-v8-alter: above + smelborp/MixtralOrochi8x7B = 0.5

Models Merged

The following models were included in the merge:

• smelborp/MixtralOrochi8x7B
• ../maid-yuzu-v8-step4-alter

Configuration

The following YAML configuration was used to produce this model:

base_model:
  model:
    path: ../maid-yuzu-v8-step4-alter
dtype: bfloat16
merge_method: slerp
parameters:
  t:
  - value: 0.5
slices:
- sources:
  - layer_range: [0, 32]
    model:
      model:
        path: ../maid-yuzu-v8-step4-alter
  - layer_range: [0, 32]
    model:
      model:
        path: smelborp/MixtralOrochi8x7B