TritonModelStock-24B-V2.1

TritonModelStock-24B-V2.1 is a Multi-Merge 24B parameter model created using the model_stock merge method via Mergekit. It combines multiple specialized 24B role-play models with multiple merge steps using a filter-wise approach to maintain structural stability while blending their capabilities.

Details

Models Merged

The following models were included in the merge:

Magidonia-24B-v4.3-heretic
MS3.2-PaintedFantasy-v2-24B
Dans-PersonalityEngine-V1.3.0-24b
Eurydice-24b-v3.5
Cydonia-24B-v4.3-heretic
Dark-Nexus-24B-v2.0
MS3.2-The-Omega-Directive-24B-Unslop-v2.0
Broken-Tutu-24B-Transgression-v2.0
MS3.2-PaintedFantasy-v4.1-24B-ultra-uncensored-heretic-v1
Rotor_24B_V.1
WeirdCompound-v1.7-24b-absolute-heresy
Step 3 Model
Step 2 Model
TritonModelStock-V1 (my previous model / Step 1 model)

Configuration(s)

The instructions used to generate this merge:

Step 1:

merge_method: model_stock
base_model: Magidonia-24B-v4.3-heretic
filter_wise: true # stable
models:

  - model: MS3.2-PaintedFantasy-v2-24B # HQ anime model
  - model: Cydonia-24B-Heretic-v4 # Good instuct following
  - model: Dans-PersonalityEngine-V1.3.0-24b # High quality personality and character
  - model: Magidonia-24B-v4.3-heretic # same model for flavor
  - model: Eurydice-24b-v3.5 # good rp model

dtype: bfloat16
name: Triton-V1

Step 2:

merge_method: slerp
base_model: TritonModelStock
parameters:
  t: 0.40 # 40% WeirdCompound, 60% stable TritonModelStock v1 as base. for step 3
models:

  - model: WeirdCompound-v1.7-24b-absolute-heresy

dtype: bfloat16
name: TritonModelCompound

Step 3:

merge_method: model_stock
base_model: TritonModelStock
filter_wise: true # stable
models:

  - model: TritonModelStock # same model for stability / flavor
  - model: TritonModelCompound # created on step 2, used basic slerp merge with weirdcompound1.7 W40/T60
  - model: Rotor_24B_V.1 # good story-writing, i think it does make the model write longer with good creativity
  - model: Dans-PersonalityEngine-V1.3.0-24b # High quality character and personality, again, why not?
  - model: MS3.2-PaintedFantasy-v4.1-24B-ultra-uncensored-heretic-v1 # New version of the PaintedFantasy, Anime style roleplay
  - model: Broken-Tutu-24B-Transgression-v2.0 # Another high quality RP model

dtype: bfloat16
name: Triton-v2 # lacks stability little bit, but it writes good.

Step 4:

merge_method: model_stock # model_stock, my belowed
base_model:  TritonModelStock-V2 # V2 as anchor
filter_wise: true # stable
models:

  - model: TritonModelStock-V2 # The creative but slightly unstable model
  - model: TritonModelStock    # The stable V1 baseline
  - model: MS3.2-The-Omega-Directive-24B-Unslop-v2.0 # good rp model by readyart
  - model: Dark-Nexus-24B-v2.0 # uncensored brain by readyart

dtype: bfloat16
name: Triton-v2.1-Final