This is a merge of pre-trained language models created using
mergekit.
This model was merged using the passthrough merge method.
1slices:
2 - sources:
3 - model: T145/ZEUS-8B-V17
4 layer_range: [0, 18]
5 - sources:
6 # Reasoning: The script used to abliterate V2 & V13 identified layer 19 as the target with the most refusal.
7 # Substituting the whole layer with that from the uncensored model should effectively abliterate V17 and future similar merges.
8 - model: Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2
9 layer_range: [19, 19]
10 - sources:
11 - model: T145/ZEUS-8B-V17
12 layer_range: [20, 32]
13merge_method: passthrough
14tokenizer_source: T145/ZEUS-8B-V17
15dtype: bfloat16
Detailed results can be found
here!
Summarized results can be found
here!
