patricide-12B-Unslop-Mell-v2
The sins of the Father shan't ever be repeated this way.
This is a merge of pre-trained language models created using mergekit.
This is my seventh model. I decided to use TheDrummer/UnslopNemo-12B-v4 instead of TheDrummer/UnslopNemo-12B-v4.1 as it supposedly has more anti-GPTism influence at the cost of intelligence, so I'll be using it in future merges. It could most likely be counteracted by adding more intelligent models. TheDrummer said that Metharme/Pygmalion templates have higher anti-GPTism effect, but those specific tokens aren't enforced/present in the tokenizer, and I prefer ChatML. Thusly I picked the model that has more anti-GPTism influence in it's base state. I decided to tweak the parameters to be more balanced, while also just generally testing NuSLERP. If I find better parameters I might release a V2B of some kind. I still haven't had much time to test this exhaustively and I'm also working on other projects.
Testing stage: early testing
I do not know how this model holds up over long term context. Early testing showed stability and viable answers.
Parameters
- Context size: Not more than 20k recommended - coherency may degrade.
- Chat Template: ChatML; Metharme/Pygmalion (as per UnslopNemo) may work, but effects are untested
- Samplers: A Temperature-Last of 1 and Min-P of 0.1 are viable, but haven't been finetuned. Activate DRY if repetition appears. XTC is untested.
Quantization
Static GGUF Quants available at:
My glorious kings/queens ❤️ Y'all's doin' the lord's work.
Merge Details
Merge Method
This model was merged using the NuSLERP merge method.
Models Merged
The following models were included in the merge:
Configuration
The following YAML configuration was used to produce this model:
models:
- model: TheDrummer/UnslopNemo-12B-v4
parameters:
weight: [0.6, 0.5, 0.3, 0.5, 0.6]
- model: inflatebot/MN-12B-Mag-Mell-R1
parameters:
weight: [0.4, 0.5, 0.7, 0.5, 0.4]
merge_method: nuslerp
dtype: bfloat16
chat_template: "chatml"
tokenizer:
source: union
parameters:
normalize: true
int8_mask: true