Alkahest is part of my ongoing experiments with merging specialized curated models. It has a few occasional logic hiccups, but it's creativity more than makes up for it. Might just require a swipe here and there.
As for samplers, the model is very creative at 0.02 min P and 1 temp but increasing the min P might be necessary to help cull some minor coherency issues.
Because of the nature of this sort of 'Hyper Multi Model Merge', my recommendation is not to run this on anything lower than a Q5 quant.
If you enjoy my work, please consider supporting me, It helps me make more models like this!
Support on KO-FI <3This is a merge of pre-trained language models created using
mergekit.
This model was merged using the
DARE TIES merge method using
TareksLab/Stylizer-V2-LLaMa-70B as a base.
1models:
2 - model: TareksLab/Wordsmith-V9-LLaMa-70B
3 parameters:
4 weight: 0.25
5 density: 0.5
6 - model: TareksLab/Malediction-V2-LLaMa-70B
7 parameters:
8 weight: 0.25
9 density: 0.5
10 - model: TareksLab/Dungeons-and-Dragons-V1.2-LLaMa-70B
11 parameters:
12 weight: 0.25
13 density: 0.5
14 - model: TareksLab/Stylizer-V2-LLaMa-70B
15 parameters:
16 weight: 0.25
17 density: 0.5
18merge_method: dare_ties
19base_model: TareksLab/Stylizer-V2-LLaMa-70B
20parameters:
21 normalize: false
22out_dtype: bfloat16
23chat_template: llama3
24tokenizer:
25 source: base
26
