This is an ExLlamaV2 quantized model in 4bpw of
mpasila/Kunoichi-DPO-v2-Instruct-32k-7B using the default calibration dataset.
This is a merge of pre-trained language models created using
mergekit.
This hopefully gives 32k context for Kunoichi-DPO-v2 model though since it also uses the instruct model it might change its behavior somewhat.
Merge script copied from this
ichigoberry/pandafish-2-7b-32k.
This model was merged using the
DARE TIES merge method using
alpindale/Mistral-7B-v0.2-hf as a base.
1models:
2 - model: alpindale/Mistral-7B-v0.2-hf
3 # No parameters necessary for base model
4 - model: mistralai/Mistral-7B-Instruct-v0.2
5 parameters:
6 density: 0.53
7 weight: 0.4
8 - model: SanjiWatsuki/Kunoichi-DPO-v2-7B
9 parameters:
10 density: 0.53
11 weight: 0.4
12merge_method: dare_ties
13base_model: alpindale/Mistral-7B-v0.2-hf
14parameters:
15 int8_mask: true
16dtype: bfloat16
17
