metadata

license: cc-by-nc-4.0
tags:
  - conversational
  - mixtral
  - merge
  - mergekit

  e88 88e                               d8     
 d888 888b  8888 8888  ,"Y88b 888 8e   d88     
C8888 8888D 8888 8888 "8" 888 888 88b d88888   
 Y888 888P  Y888 888P ,ee 888 888 888  888     
  "88 88"    "88 88"  "88 888 888 888  888     
      b                                        
      8b,                                      
 
  e88'Y88                  d8           888    
 d888  'Y  ,"Y88b 888,8,  d88    ,e e,  888    
C8888     "8" 888 888 "  d88888 d88 88b 888    
 Y888  ,d ,ee 888 888     888   888   , 888    
  "88,d88 "88 888 888     888    "YeeP" 888    
                                               
PROUDLY PRESENTS

TeTO-MS-8x7b-exl2-rpcal

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

Branches:

main -- measurement.json
4.5b6h -- 4.5bpw, 6bit lm_head
4b6h -- 4bpw, 6bit lm_head
3.5b6h -- 3.5bpw, 6bit lm_head
2.5b6h -- 2.5bpw, 6bit lm_head

Original model link: (reuploaded, original source got taken down) InferenceIllusionist/TeTO-MS-8x7b

Original model README below.

TeTO-MS-8x7b

Tesoro + Typhon + OpenGPT

Presenting a Model Stock experiment combining the unique strengths from the following 8x7b Mixtral models:

Tess-2.0-Mixtral-8x7B-v0.2 / migtissera / General Purpose
Typhon-Mixtral-v1 / Sao10K / Creative & Story Completion
Open_Gpt4_8x7B_v0.2 / rombodawg / Conversational

Weighted (iMat) GGUFS: https://huggingface.co/Quant-Cartel/TeTO-MS-8x7b-iMat-GGUF

Recommended Template

Basic: Alpaca Format
Advanced: See context/instruct/sampler settings in our new Recommended Settings repo.
Huge shout out to rAIfle for his original work on the Wizard 8x22b templates which were modified for this model.

Methodology

[I]nnovative layer-wise weight averaging technique surpasses state-of-the-art model methods such as Model Soup, utilizing only two fine-tuned models. This strategy can be aptly coined Model Stock, highlighting its reliance on selecting a minimal number of models to draw a more optimized-averaged model (From arXiv:2403.19522)

Methodology and merging process was based on the following paper - Model Stock: All we need is just a few fine-tuned models
Initial model selection was based on top performing models of Mixtral architecture covering a variety of use cases and skills
Base model (Mixtral Instruct 8x7b v0.1) was chosen after outperforming two other potential base models in terms of MMLU benchmark performance.

Output

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

Merge Details

Merge Method

This model was merged using the Model Stock merge method using Mixtral-8x7B-v0.1-Instruct as a base.

Models Merged

The following models were included in the merge:

migtissera_Tess-2.0-Mixtral-8x7B-v0.2
rombodawg_Open_Gpt4_8x7B_v0.2
Sao10K_Typhon-Mixtral-v1

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: models/migtissera_Tess-2.0-Mixtral-8x7B-v0.2
  - model: models/Sao10K_Typhon-Mixtral-v1
  - model: models/rombodawg_Open_Gpt4_8x7B_v0.2 
merge_method: model_stock
base_model: models/Mixtral-8x7B-v0.1-Instruct
dtype: float16

Appendix - Llama.cpp MMLU Benchmark Results*

These results were calculated via perplexity.exe from llama.cpp using the following params:

.\perplexity -m .\models\TeTO-8x7b-MS-v0.03\TeTO-MS-8x7b-Q6_K.gguf -bf .\evaluations\mmlu-test.bin --multiple-choice -c 8192 -t 23 -ngl 200

* V0.01 (4 model / Mixtral Base):
    Final result: 43.3049 +/- 0.4196
    Random chance: 25.0000 +/- 0.3667


* V0.02 (3 model / Tess Mixtral Base):
    Final result: 43.8356 +/- 0.4202
    Random chance: 25.0000 +/- 0.3667


* V0.03 (4 model / Mixtral Instruct Base):
    Final result: 45.7004 +/- 0.4219
    Random chance: 25.0000 +/- 0.3667

*Please be advised metrics above are not representative of final HF benchmark scores for reasons given here