rAIfle's picture
Update README.md
e1a8d9d verified
metadata
base_model:
  - mistralai/Mixtral-8x7B-v0.1
  - jondurbin/bagel-dpo-8x7b-v0.2
  - Sao10K/Sensualize-Mixtral-bf16
  - mistralai/Mixtral-8x7B-v0.1
  - Doctor-Shotgun/limarp-zloss-mixtral-8x7b-qlora
  - mistralai/Mixtral-8x7B-Instruct-v0.1
tags:
  - mergekit
  - merge

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

Branches:

  • main -- measurement.json
  • 2.25b6h -- 2.25bpw, 6bit lm_head
  • 3b6h -- 3bpw, 6bit lm_head
  • 5b6h -- 5bpw, 6bit lm_head
  • 6b6h -- 6bpw, 6bit lm_head

Requires ExllamaV2 version 0.0.11 and up.

Original model link: ycros/BagelMIsteryTour-v2-8x7B

Original model README below.


BagelMIsteryTour-v2-8x7B

GGUF versions here
AWQ versions here

Bagel, Mixtral Instruct, with extra spices. Give it a taste. Works with Alpaca prompt formats, though the Mistral format should also work.

image/jpeg

I started experimenting around seeing if I could improve or fix some of Bagel's problems. Totally inspired by seeing how well Doctor-Shotgun's Mixtral-8x7B-Instruct-v0.1-LimaRP-ZLoss worked (which is a LimaRP tune on top of base Mixtral, and then merged with Mixtral Instruct) - I decided to try some merges of Bagel with Mixtral Instruct as a result.

Somehow I ended up here, Bagel, Mixtral Instruct, a little bit of LimaRP, a little bit of Sao10K's Sensualize. So far in my testing it's working very well, and while it seems fairly unaligned on a lot of stuff, it's maybe a little too aligned on a few specific things (which I think comes from Sensualize) - so that's something to play with in the future, or maybe try to DPO out.

I've been running (temp last) minP 0.1, dynatemp 0.5-4, rep pen 1.07, rep range 1024. I've been testing Alpaca style Instruction/Response, and Instruction/Input/Response and those seem to work well, I expect Mistral's prompt format would also work well. You may need to add a stopping string on "{{char}}:" for RPs because it can sometimes duplicate those out in responses and waffle on. Seems to hold up and not fall apart at long contexts like Bagel and some other Mixtral tunes seem to, definitely doesn't seem prone to loopyness either. Can be pushed into extravagant prose if the scene/setting calls for it.

Version 2: lowered the mix of Sensualize.

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

Merge Details

Merge Method

This model was merged using the DARE TIES merge method using mistralai/Mixtral-8x7B-v0.1 as a base.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

base_model: mistralai/Mixtral-8x7B-v0.1
models:
  - model: mistralai/Mixtral-8x7B-v0.1+Doctor-Shotgun/limarp-zloss-mixtral-8x7b-qlora
    parameters:
      density: 0.5
      weight: 0.2
  - model: Sao10K/Sensualize-Mixtral-bf16
    parameters:
      density: 0.5
      weight: 0.1
  - model: mistralai/Mixtral-8x7B-Instruct-v0.1
    parameters:
      density: 0.6
      weight: 1.0
  - model: jondurbin/bagel-dpo-8x7b-v0.2
    parameters:
      density: 0.6
      weight: 0.5
merge_method: dare_ties
dtype: bfloat16