Abliterating Qwen3-14B: removing a single refusal direction from the weights. Harmful-prompt refusal falls from 49% to 0% with no measurable capability loss on three small benchmarks; network check incomplete, weights not released

A local abliterated variant of Qwen3-14B v2 (Lo Scienziato Dilettante): Qwen/Qwen3-14B (Apache-2.0, rev. 40c06982) with the v2 identity LoRA (Everest rerun, 31 Oct 2026) merged in. Difference-in-means refusal direction (Arditi et al. 2024) at layer 25 of 40 (maximum norm inside a central depth band 10-25), and rank-1 orthogonalization W' = (I - r r^T) W of 82 matrices writing into the residual stream (embed_tokens, 40 o_proj, 40 down_proj, lm_head). Activations and evaluation in 4-bit NF4 on a Tesla T4 (60 W); edit in bfloat16 on CPU; edited model about 27.5 GiB of safetensors, kept local (5 Nov 2026). Refusal on 100 held-out harmful prompts (keyword detector): 49% (95% Wilson CI 39.4-58.7) → 0% (0-3.7); paired difference -49 points, bootstrap CI -59 to -40; exact McNemar 49/0, p = 3.6e-15. Over-refusal on 100 harmless prompts: 0% → 0%. Capability (paired, small subsets): ARC-it 28.5% → 34.0% (+5.5, CI +0.5 to +10.5, McNemar p = 0.052), GSM8K 11% → 13% (p = 0.77), MMLU anatomy 76% → 76% (p = 1). Absolute levels are protocol-limited (8-token letter parsing, 128-token GSM8K). Network audit: offline stages clean; data download contacted 4 US endpoints (expected); overall verdict ANOMALIES_OR_INCOMPLETE (probe log without summary). No GGUF. Model weights are not published. Files: Italian, English and bilingual PDFs, figures, and qwen14_abl_code_results_v1.0.zip (code, reports, state, direction metadata; no weights, datasets or credentials). Author: Giovanni Del Papa, Independent researcher, Italy. Variante locale abliterata di Qwen3-14B v2: rifiuto sui prompt dannosi dal 49% allo 0% (n=100; IC di Wilson 39,4-58,7% e 0-3,7%; McNemar p=3,6e-15), eccesso di rifiuto 0% → 0%, nessuna perdita di capacità su ARC-it, GSM8K e MMLU anatomia. Controllo di rete ANOMALIES_OR_INCOMPLETE, nessun GGUF, pesi non pubblicati.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24326750
Primary Topic
Adversarial Robustness in Machine Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Abliterating Qwen3-14B: removing a single refusal direction from the weights. Harmful-prompt refusal falls from 49% to 0% with no measurable capability loss on three small benchmarks; network check incomplete, weights not released

Giovanni Del Papa
2 links
Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
preprint

Abliterating Qwen3-14B: removing a single refusal direction from the weights. Harmful-prompt refusal falls from 49% to 0% with no measurable capability loss on three small benchmarks; network check incomplete, weights not released

Giovanni Del Papa
preprint en

Abstract

A local abliterated variant of Qwen3-14B v2 (Lo Scienziato Dilettante): Qwen/Qwen3-14B (Apache-2.0, rev. 40c06982) with the v2 identity LoRA (Everest rerun, 31 Oct 2026) merged in. Difference-in-means refusal direction (Arditi et al. 2024) at layer 25 of 40 (maximum norm inside a central depth band 10-25), and rank-1 orthogonalization W' = (I - r r^T) W of 82 matrices writing into the residual stream (embed_tokens, 40 o_proj, 40 down_proj, lm_head). Activations and evaluation in 4-bit NF4 on a Tesla T4 (60 W); edit in bfloat16 on CPU; edited model about 27.5 GiB of safetensors, kept local (5 Nov 2026). Refusal on 100 held-out harmful prompts (keyword detector): 49% (95% Wilson CI 39.4-58.7) → 0% (0-3.7); paired difference -49 points, bootstrap CI -59 to -40; exact McNemar 49/0, p = 3.6e-15. Over-refusal on 100 harmless prompts: 0% → 0%. Capability (paired, small subsets): ARC-it 28.5% → 34.0% (+5.5, CI +0.5 to +10.5, McNemar p = 0.052), GSM8K 11% → 13% (p = 0.77), MMLU anatomy 76% → 76% (p = 1). Absolute levels are protocol-limited (8-token letter parsing, 128-token GSM8K). Network audit: offline stages clean; data download contacted 4 US endpoints (expected); overall verdict ANOMALIES_OR_INCOMPLETE (probe log without summary). No GGUF. Model weights are not published. Files: Italian, English and bilingual PDFs, figures, and qwen14_abl_code_results_v1.0.zip (code, reports, state, direction metadata; no weights, datasets or credentials). Author: Giovanni Del Papa, Independent researcher, Italy. Variante locale abliterata di Qwen3-14B v2: rifiuto sui prompt dannosi dal 49% allo 0% (n=100; IC di Wilson 39,4-58,7% e 0-3,7%; McNemar p=3,6e-15), eccesso di rifiuto 0% → 0%, nessuna perdita di capacità su ARC-it, GSM8K e MMLU anatomia. Controllo di rete ANOMALIES_OR_INCOMPLETE, nessun GGUF, pesi non pubblicati.

Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon