language:
- en
license: apache-2.0
library_name: transformers
tags:
- mergekit
- merge
base_model: Nohobby/Qwen2.5-32B-Peganum-v0.1
4bpw exl2 quant of: https://huggingface.co/Nohobby/Qwen2.5-32B-Peganum-v0.1
Peganum
Many thanks to the authors of the models used!
Qwen2.5 | Qwen2.5-Instruct | Qwen-2.5-Instruct-abliterated | RPMax-v1.3-32B | EVA-Instruct-32B-v2(EVA-Qwen2.5-32B-v0.2+ Qwen2.5-Gutenberg-Doppel-32B)
Overview
Main uses: RP
Prompt format: ChatML
Just trying out merging Qwen, because why not. Slightly fewer refusals than other Qwen tunes, while performance seems unaffected by abliteration. I've hardly used Q2.5 models before, so I can't really compare them beyond that.
Quants
Settings
Samplers: https://huggingface.co/Nohobby/Qwen2.5-32B-Peganum-v0.1/resolve/main/Peganum.json
You can also use the SillyTavern presets listed on the EVA-v0.2 model card
Merge Details
Merging steps
Step1
(Config taken from here)
base_model: zetasepic/Qwen2.5-32B-Instruct-abliterated-v2
dtype: bfloat16
merge_method: task_arithmetic
parameters:
normalize: false
slices:
- sources:
- layer_range: [0, 64]
model: zetasepic/Qwen2.5-32B-Instruct-abliterated-v2
- layer_range: [0, 64]
model: unsloth/Qwen2.5-32B-Instruct
parameters:
weight: -1.0
Step2
(Config taken from here)
models:
- model: unsloth/Qwen2.5-32B
- model: Step1
parameters:
weight: [0.50, 0.20]
density: [0.75, 0.55]
- model: ArliAI/Qwen2.5-32B-ArliAI-RPMax-v1.3
parameters:
weight: [0.50, 0.80]
density: [0.75, 0.85]
merge_method: ties
base_model: unsloth/Qwen2.5-32B
parameters:
int8_mask: true
rescale: true
normalize: false
dtype: bfloat16
Peganum
(Config taken from here)
models:
- model: Step2
parameters:
weight: 1
density: 1
- model: ParasiticRogue/EVA-Instruct-32B-v2
parameters:
weight: [0.0, 0.2, 0.66, 0.8, 1.0, 0.8, 0.66, 0.2, 0.0]
density: 0.5
merge_method: ties
base_model: unsloth/Qwen2.5-32B
parameters:
normalize: true
int8_mask: true
dtype: bfloat16