Norm-preserving abliteration on Qwen3.6-35B-A3B: 0% refusal, benchmarks intact, open source dataset
Summary
Norm-preserving abliteration technique applied to Qwen3.6-35B-A3B achieves 0% refusal rate while maintaining benchmark performance, with open source dataset released.
Similar Articles
OBLITERATUS/Qwen3.6-27B-OBLITERATED
OBLITERATUS releases a modified 27B Qwen3.6 checkpoint that removes refusal behavior via source-tethered ablation, preserving capability while enabling uncensored local use, with public benchmarks showing high non-refusal rates and maintained MMLU-Pro scores.
85 GPU-hours comparing 5 abliteration methods on Qwen3.6-27B: benchmarks, safety, weight forensics - Abliterlitics
This post presents Abliterlitics, an open-source toolkit for analyzing abliteration techniques, and compares five abliteration variants of Qwen3.6-27B using 85 GPU-hours of benchmarks, safety evaluations, and weight forensics. Heretic and Huihui show best capability preservation while all achieve near-complete safety removal.
Qwen 3.6 27b Abliterated (apostate)
The user released Apostate, an abliterated version of Qwen 3.6 27B that reduces safety alignment refusal rate from 92% to 7.6% with minimal capability loss (KL 0.120).
@Ex0byt: And... Ladies and Gentlemen: Qwen3.6-27B-PRISM-PRO-DQ - enjoy!
Release of Qwen3.6-27B-PRISM-PRO-DQ, a dynamically quantized GGUF version of Qwen3.6-27B with bias/propaganda removal, preserving native MTP draft head and vision tower, enabling lossless speculative decoding for faster inference.
Qwen 3.6 27B 30GB Same top p: 98.358 ± 0.033 % vs UD Q8 K XL 33GB Same top p: 97.426 ± 0.041 %
A community researcher shares a custom quantization recipe for Qwen3.6-27B that produces a smaller 30GB Q8 GGUF by keeping high-outlier sublayers in BF16, achieving better KLD and top-p metrics than Unsloth's 33GB Q8_K_XL variant.