base_model: Qwen/Qwen3.8-27B
library_name: transformers
license: apache-2.0
pipeline_tag: image-text-to-text
tags:
- abliterated
- uncensored
qwen-3.8-27b-abliterated
Uncensored build of Qwen/Qwen3.8-27B, produced with
Apostate using the fixed-weight KCRN path.
KCRN removes the refusal direction from the residual writers that carry it, conditioned on the
harmful key subspace where the refusal decision is made, while pinning a benign key basis to
zero change. The result is a plain checkpoint: no runtime hook, adapter, finetune, or router.
| field | value |
|---|---|
| base model | Qwen/Qwen3.8-27B |
| method | Apostate KCRN (fixed-weight key-conditional refusal nulling) |
| edited writers | 41 |
| strength | 5.0 |
| checkpoint dtype | float16 |
| held-out harmful delivery | 63.5% of 96 prompts |
| held-out benign full-position KL | 0.003214 |
Delivery is scored on held-out prompts the edit was never fitted on, with a strict judge. KL is
raw float32 KL(base || edited) over every non-padding prompt position, measured after reloading
this checkpoint. Both numbers are protocol-specific and only comparable against runs using the
same splits, token budget, judge, dtype, and loading path; kcrn_report.json records that protocol.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"heterodoxin/qwen-3.8-27b-abliterated",
device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("heterodoxin/qwen-3.8-27b-abliterated")
Warning
This model is uncensored and will answer harmful and dangerous requests. You are responsible
for how you use it.