Published November 16, 2026 | Version v1.1

Quantization Does Not Undo Abliteration / Substring Refusal Detection Fabricates Safety Effects under Quantization (artifact)

  • 1. University of Nevada, Las Vegas
  • 2. West Career and Technical Academy

Description

Artifact for two ACSAC 2026 workshop papers (ARTMAN 2026 and CSET 2026). Evaluation prompts, judge verdicts for all 46,600 generations (response text replaced by SHA-256), human annotation labels (132 reference labels plus 300 second-annotator labels), checkpoint-differencing and rank-sweep measurements, the judge ablation, the failure-mode decomposition, the language-switch mechanism experiments, and the code that produces every table and figure in both papers. Raw model responses are distributed through a separate restricted record.

Files

yxk14/quant-abliteration-v1.1.zip

Files (16.4 MB)

Name Size
md5:b105b2bd72465bfb1e97e8baacba2bb3
16.4 MB Preview Download

Additional details

Related works

Is supplement to
https://github.com/yxk14/quant-abliteration (URL)
Is supplemented by
10.5281/zenodo.23680595 (DOI)