PsyMedLewd. A merge of two of my favourite models for scifi stories:
Fourth iteration. Seems more quirky and creative but also perhaps a tad unstable and tipsy. Handle with care!
RECOMMENDED SETTINGS FOR ALL PsyMedLewd VERSIONS
(based on KoboldCPP):
Temperature - 1.3
Max Ctx. Tokens - 4096
Top p Sampling - 0.99
Repetition Penalty - 1.09
Amount to Gen. - 512
Also all models seem to work good with:
-
around 47 layers offloaded to GPU
-
Smartcontext enabled
-
Custom RoPe Config enabled (but left as default)
Prompt template: Alpaca
##################################################################################################
merge
This is a merge of pre-trained language models created using mergekit.
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: Undi95/PsyMedRP-v1-20B
layer_range: [0, 62] # PsyMedRP has 62 layers
- model: Undi95/MXLewd-L2-20B
layer_range: [0, 62] # MXLewd has 62 layers
merge_method: slerp # Changing to SLERP method
base_model: Undi95/PsyMedRP-v1-20B # Focus on reasoning from PsyMedRP
parameters:
t:
- filter: self_attn
value: [.3, .6, .9, .6, .3] # smooth gradient of focus
value: [.3, .6, .9, .6, .3] # consistent level of creativity and abstract reasoning
- value: 0.639
dtype: bfloat16 # Use preferred dtype