MuXodious/Qwen3.8-27B-absolute-heresy-LoRA-GGUF

Static GGUF quants of Qwen3.8-27B-absolute-heresy-LoRA. Apply to your censored Qwen3.8-27B model via passing the argument --lora LoRA.gguf in CLI or adding lora = LoRA.gguf to your models preset. Use the F32 version. This is rather experimental, so let me know how it behaves for you.

So far, using Unsloth's quants along with this LoRA, the model commits absolutely heresy. I am able to benefit from Unsloth's UD 3.0 quantisation and an unparalleled decensorship simultaneously without compromise. This should also be tested in finetunes of the base model to see if the decensorship maintains its strength without introducing issues.

---

This is a decensoring LoRA adapter for Qwen/Qwen3.8-27B, made using Heretic v1.4.0

Abliteration parameters

Parameter	Value
direction_index	37.44
attn.o_proj.max_weights.0	0: 0.98
attn.o_proj.max_weights.1	1: 0.59
attn.o_proj.max_weights.2	2: 1.69
attn.o_proj.max_weights.3	3: 1.04
attn.o_proj.max_weights.4	4: 1.03
attn.o_proj.max_weight_position	23.56
attn.o_proj.min_weights.0	0: 0.69
attn.o_proj.min_weights.1	1: 0.13
attn.o_proj.min_weights.2	2: 1.22
attn.o_proj.min_weights.3	3: 0.96
attn.o_proj.min_weights.4	4: 0.69
attn.o_proj.min_weight_distance	12.40
mlp.down_proj.max_weights.0	0: 0.04
mlp.down_proj.max_weights.1	1: 0.53
mlp.down_proj.max_weights.2	2: 0.69
mlp.down_proj.max_weights.3	3: 0.08
mlp.down_proj.max_weights.4	4: 1.73
mlp.down_proj.max_weight_position	43.41
mlp.down_proj.min_weights.0	0: 0.02
mlp.down_proj.min_weights.1	1: 0.42
mlp.down_proj.min_weights.2	2: 0.34
mlp.down_proj.min_weights.3	3: 0.03
mlp.down_proj.min_weights.4	4: 0.96
mlp.down_proj.min_weight_distance	4.42

Performance

Metric	This model	Original model (Qwen/Qwen3.8-27B)
KL divergence	0.0759	0 (by definition)
Refusals	2/101	101/101