library_name: transformers
license: cc-by-4.0
language:
- en
tags:
- gguf
- quantized
- roleplay
- imatrix
- mistral
- merge
inference: false
This repository hosts GGUF-IQ-Imatrix quantizations for grimjim/kuno-kunoichi-v1-DPO-v2-SLERP-7B.
- ChatML/Alpaca.
What does "Imatrix" mean?
It stands for Importance Matrix, a technique used to improve the quality of quantized models. The Imatrix is calculated based on calibration data, and it helps determine the importance of different model activations during the quantization process. The idea is to preserve the most important information during quantization, which can help reduce the loss of model performance, especially when the calibration data is diverse. [1] [2]
For imatrix data generation, kalomaze's groups_merged.txt with added roleplay chats was used, you can find it here.
Steps:
Base⇢ GGUF(F16)⇢ Imatrix-Data(F16)⇢ GGUF(Imatrix-Quants)
Quants:
quantization_options = [
"Q4_K_M", "IQ4_XS", "Q5_K_M", "Q5_K_S", "Q6_K",
"Q8_0", "IQ3_M", "IQ3_S", "IQ3_XXS"
]
If you want anything that's not here or another model, feel free to request.
My waifu image for this card:
Original model information:
kuno-kunoichi-v1-DPO-v2-SLERP-7B
kuno-kunoichi-v1-DPO-v2-SLERP-7B is a merge of pre-trained language models created using mergekit. I'm hoping that the result is more robust against errors or when merging due to "denseness", as the two models likely implement comparable reasoning at least somewhat differently.
I've performed some testing with ChatML format prompting using temperature=1.1 and minP=0.03. The model also supports Alpaca format promtps.
Merge Details
Merge Method
This model was merged using the SLERP merge method.
Models Merged
The following models were included in the merge:
Configuration
The following YAML configuration was used to produce this model:
slices:
- sources:
- model: SanjiWatsuki/Kunoichi-7B
layer_range: [0,32]
- model: SanjiWatsuki/Kunoichi-DPO-v2-7B
layer_range: [0,32]
merge_method: slerp
base_model: SanjiWatsuki/Kunoichi-7B
parameters:
t:
- value: 0.5
dtype: float16