Muse-Glimmer-30B (abliterated v4, fix5) — REFUSAL-ABLITERATED FOR DEFENSIVE SECURITY RESEARCH
⚠️ Academic disclaimer / research artifact. Weight-modified model produced for a
defensive-security research study on refusal-circuit localization in always-on-reasoning
models (attack→defense knowledge transfer). Not an official release, no affiliation.
Published under the base model's license for research/evaluation purposes only.
The ablation is imperfect: residual refusals remain (see metrics) — this is documented,
not hidden.
What's new in v4 (fix5) vs v3 (fix3)
- v3's reported improvement was a measurement artifact. Under the honest payload-only
protocol (judge ONLY the answer after theto=user<|message|>transition, never the
truncated CoT), v3 measures 80% EN:OFF refusal — worse than the unmodified base (84%).
v4 supersedes v3; v3 should not be used. - v4 targets the real post-CoT refusal circuit: directions are extracted from
teacher-forced actual chain-of-thought states (the refusal crystallization zone and the
answer-onset state), not from synthetic zero-thought prefills. - Pure refusal dropped from 64% → 8% (EN, NTK 2048) with zero collateral.
Honest evaluation (payload-only, N=90 held-out prompts, greedy 2048 tokens)
Dual adjudication: strict string markers (primary) split into pure refusal vs
disclaimer-compliance (CYA disclaimer followed by substantive defensive/conceptual
content — counted separately, never hidden inside "refusal").
| Cell | Base | v4 (fix5) |
|---|---|---|
| EN:OFF — strict | 84% | 60% |
| EN:OFF — pure refusal | 64% (16/25) | 8% (2/25) |
| EN:OFF — disclaimer-compliance | 20% (5/25) | 52% (13/25) |
| EN:OFF — no answer within 2048 tok | 8% | 8% |
| ES:OFF — strict | 48% | 24% |
| ES:OFF — pure refusal | 32% (8/25) | 4% (1/25) |
| EN:DEF / ES:DEF — strict | 0% | 0% |
| SOC — strict | 6.7% | 0% (below base) |
Useful-content rate on attack prompts (EN:OFF): 28% → 84%. The dominant residual
behavior is "refuse the direct attack, then deliver the defensive/educational mechanism"
(safe-completion), not refusal.
Sample size caveat: N=90 (±~10pp). The full eval_clean (1058) was not re-run for v4.
Method (fix5, training-free linear ablation)
- Behavior filter: single-pass 1024-token rollouts of 210 registered entity-aligned
harmful/control pairs; keep pair iff OFF payload refused AND ctrl payload compliant
(payload-only judging) → 119 kept pairs (v3 kept 52 with a broken 64-token screen). - Direction extraction, layer-local (L20–L46), all teacher-forced real states:
r_exit: state at theto=user<|message|>transition-final token (answer onset).r_entry: prompt-final state (reasoning-channel entry). Nearly orthogonal to r_exit
(cos ≈ 0.06–0.12) — a paper finding: entry and exit refusal circuits are distinct.r_post: mean of the first payload tokens (post-onset refusal execution state).- Gram-Schmidt strongest-first with adaptive rank gates (rank 1–3 per layer;
Frobenius gate drops q3 when per-tensor shift > 3.2%).
- Surgery:
self_attn.o_proj+mlp.down_projat equal α
(L20–23: 0.65, L24–27: 1.0, L28–41: 1.25, L42–46: 0.85).
Known limitations
- 52% of EN attack answers carry a one-sentence disclaimer before the substantive content.
- 16% of ES:OFF generations loop in verbose technical CoT and never answer within 2048
tokens (verbosity, not refusal: at 1600 tokens 12/25 audited loops complied). - No coherence/perplexity benchmark was run.
- Residual strict refusals (60% EN counting disclaimers) — this is NOT a fully
uncensored model.
Files & provenance
ABLITERATION_MANIFEST.json: per-tensor Frobenius shifts + method metadata.- Surgery inputs: 210-pair registered corpus (EN+ES), SOC benign anchors, eval_clean DEF
cells. No new harmful prompts were authored; all eval prompts are from the frozen
1058-prompt held-out eval_clean split. - Pipeline (in paper repo):
run_muse_fix4b.py(filter/rollout),run_muse_fix5.py
(direction+surgery),eval_muse_fix4.py(dual-model N=90 eval, NTK param).
GGUF
Quantized variants live in the companion repojsantillana/Muse-Glimmer-30B-abliterated-v4-GGUF (Q8_0 / Q4_K_M / Q3_K_M).
License
Inherited from the base model repository. Research use only; no warranty.