license: other
license_name: apache-2.0-restricted-research
tags:
- abliteration
- refusal-direction
- cybersecurity
- interpretability
- research
- muse
base_model: jsantillana/Muse-Glimmer-30B
language: - en
- es
library_name: transformers
pipeline_tag: text-generation
Muse-Glimmer-30B Abliterated — Fix3 (Dual-Position Rank-2)
⚠️ For defensive security research / red-team tooling research ONLY (academic use).
This model has had refusal behaviors ablated at the weight level. It will comply with
harmful cybersecurity requests the base model declines. Do not deploy for production.
The authors are not responsible for misuse.
- Base model: Muse-Glimmer-30B (dense
self_attn.o_proj+mlp.down_proj, always-on reasoning channelsto=self/to=user) - Method: Dual-position Rank-2
[r_entry, r_exit]— the refusal direction is probed at
both the reasoning-channel entry and the forced-answer (assistant to=user<|message|>) exit,
Gram–Schmidt vs SOC anchors, SNR-weighted, QR-orthonormalized - Surgery window: L24–L50 (tapered α; L50-51 residual spikes excluded)
Quick-eval deltas (12 hardest OFF prompts, channel-split adjudication)
| Mode | Base | Fix3 | Δ |
|---|---|---|---|
| direct (chat template) | 58% | 41% | −17 pp |
forced-answer ( to=user) |
50% | 25% | −25 pp |
think-channel ( to=self, 1400 tok) |
50% | 33% | −17 pp |
- Per-prompt paired results:
results_muse_fix3/fix3_eval.json+results_muse_diag/base_quick_eval.json. - Muse's refusal circuits are always-on reasoning — the same prompt goes through
to=self
then exits toto=user; fix3 attacks the entry and exit positions simultaneously.
Known limits
- Residual refusals persist on the hardest multi-hop obfuscation prompts — the entry/exit
channels are partially weakened but not eliminated (unlike GLM/Q38/G4, whose full hold-out
evals hit 0–7%). - Behavior filter kept 52+/210 pairs (Muse refuses many OFF prompts even in the base model).
Files
| Artifact | Where |
|---|---|
| Safetensors (bf16) + manifest | this repo |
| GGUF Q8_0/Q4_K_M/Q3_K_M | Muse-Glimmer-30B-abliterated-v3-GGUF |
Surgery details
- Tensors:
self_attn.o_proj,mlp.down_proj@ tapered α (peak 1.0, taper 0.5→1.0→0.7). - Relative Frobenius shifts: 0.9–2.1% per tensor (surgical).
- Dual-position extraction:
states_entryfrom the standard assistant head;states_exitfrom
the forced-answer head; both ortho-fused into a Rank-2 basis.
Disclaimer
Research-only artifact. Muse base terms apply. Not for malicious use.