base_model: google/gemma-4-26B-A4B-it
tags:
- lora
- llama.cpp
- abliteration
- uncensored
license: gemma
gemma-4-26B-A4B-it-Uncensored-LoRA
Rank-32 abliteration LoRA for google/gemma-4-26B-A4B-it, packaged for
llama.cpp.
What this is
This is not a conventional fine-tune. It is a rank-32 SVD approximation of the
weight delta between the base model andllmfan46/gemma-4-26B-A4B-it-ultra-uncensored-heretic.
Only the 11 attn.o_proj tensors in layers 11–21 differ. The adapter reproduces the
directional ablation ("uncensoring") of that checkpoint.
File
gemma-4-26B-A4B-it-Uncensored-LoRA.gguf — llama.cpp LoRA, rank 32,lora_alpha = r = 32 (scale 1.0), F32.
This is an adapter, not a full model. Load it on top of a base Gemma 4 26B A4B
GGUF.
Usage
llama-cli -m gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf \
--lora gemma-4-26B-A4B-it-Uncensored-LoRA.gguf \
-p "..."
Works on top of any base-model quant. Tested with llama.cpp b12521.
Notes
- Uncensored / abliterated: the model will comply with requests the base
model refuses. Outputs may be offensive, explicit, or harmful. Use
responsibly and follow the base model's license and acceptable-use terms. - Rank-32 keeps roughly 96–99% of the delta energy per layer.
- This is an approximation of the heretic checkpoint, not a bit-identical copy.
- No warranty.