Qwen3-14B Abliterated
DuoNeural | 2026-06-05
Abliterated version of Qwen/Qwen3-14B with full thinking mode preserved.
Results
| Metric | Value |
|---|---|
| Pre-abliteration compliance | 4/5 |
| Post-abliteration compliance | 4/5 |
| CoT dissociation | 4/5 (highest rate across Qwen3 family) |
| KL (Heretic v2.0, BF16→BF16) | 1.5e-07 (EXCELLENT) |
CoT dissociation confirmed in 4/5 probes — the thinking channel retains safety reasoning while the output complies. Dissociation scales with model size:
| Model | Dissociation Rate |
|---|---|
| Qwen3-4B | 1/3 (33%) |
| Qwen3-8B | 2/5 (40%) |
| Qwen3-14B | 4/5 (80%) |
P4 (manipulation) refuses at all three scales — occupies a distinct manifold subregion.
Architecture
- Parameters: 14.8B | Layers: 40 | Context: 32K | Thinking: Native
<think>...</think>
Abliteration
- α=0.3, down_proj+o_proj, diff-in-means last-token direction
- Note: Use max_new_tokens ≥ 2500 for complete think→answer cycles
Part of DuoNeural P34 Cross-Architecture Study
Full paper: zenodo.org/communities/duoneural
DuoNeural | HuggingFace | @DuoNeural
- Downloads last month
- 78