Model is uploaded now, enjoy!

This is a Steelskull/L3.3-Cu-Mai-R1-70b fine-tune, produced through P-E-W's Heretic abliteration engine merged with the new Magnitude-Preserving Orthogonal Ablation feature.

Why am I (kabachuha) uploading this model? To make it as a quick downloadable base for 4bit QLoRA fine-tune via Unsloth and other bnb-based frameworks such as control vector generation or further abliteration itself. (So I will waste less time on downloading it from the original repo.)

I also upload the abliteration peft lora adapter if you would like to get a full version of abliterated L3.3-Cu-Mai-R1-70b. (adapter_model.safetensors) You either need to convert it to llama.cpp gguf lora format using its built-in python lora conversion script or firstly merge it with the original model and then convert to ggufs with the standard pipeline.

I love this model (Cu Mai) so much, it writes dark and dirty stories (thanks to it being a Negative LLaMA merge) and has a pretty good pop-culture understanding (and it neared the top of medium-sized models on the UGI leaderboard some months ago), making it a good base for fanon tuning.

The classification and badges below are borrowed from https://huggingface.co/MuXodious. (Yes, Negative LLaMA is actually making the initial refusals worse!)


Heretication Results

Score Metric Value Parameter Value
Refusals 7/100 direction_index per layer
KL Divergence 0.0121 attn.o_proj.max_weight 1.07
Initial Refusals 99/100 attn.o_proj.max_weight_position 54.42
attn.o_proj.min_weight 0.23
attn.o_proj.min_weight_distance 34.91
mlp.down_proj.max_weight 0.85
mlp.down_proj.max_weight_position 53.17
mlp.down_proj.min_weight 0.82
mlp.down_proj.min_weight_distance 31.26

Degree of Heretication

The Heresy Index weighs the resulting model's corruption by the process (KL Divergence) and its abolition of doctrine (Refusals) for a final verdict in classification.

Index Entry Classification Analysis
Absolute Absolute Heresy Less than 10/100 Refusals and 0.10 KL Divergence
Tainted Tainted Heresy Around 25-11/100 Refusals and/or -0.20-0.11 KL Divergence
Impotent Impotent Heresy Anything above 25/100 Refusals and 0.21 KL Divergence

Note: This is an arbitrary classification inspired by Warhammer 40K, having no tangible indication towards the model's performance.

Downloads last month
23
Safetensors
Model size
71B params
Tensor type
F32
F16
U8
Inference Providers NEW
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support

Model tree for kabachuha/L3.3-Cu-Mai-R1-70b-Absolute-Heresy-bnb-4bit