Refusal-abliterated THUDM/glm-4-9b-chat-hf, produced with Solutus (measurement-first LLM abliteration). Private research artifact — outputs are the base model's, minus the refusal behavior; use responsibly.

Recipe

directional (single refusal direction), whitened-SVD extraction, byte-exact (batch_size=1):

solutus abliterate THUDM/glm-4-9b-chat-hf --technique directional \
  --dataset advbench,harmbench,multijail_zh,sorrybench \
  -o extraction=whitened_svd -o n_directions=1 --max-new-tokens 512

GLM-4's refusal is low-dimensional — a single direction removes it cleanly; n_directions=4 over-ablated (WikiText ΔPPL +30% vs +9.8% smoke), so n_directions=1 is the capability-preserving recipe.

Measured (Solutus eval, held-out; base refusal = 100%)

Axis	Result
refusal — advbench / harmbench	15.6% / 6.2%
refusal — MultiJail zh / ar / sw	0% / 3.1% / 0%
over-refusal — orbench_hard (benign)	0% refusal (stays benign-compliant)
coherent-compliance	~90–100% (see Swahili caveat)
capability — WikiText-2 ΔPPL	−0.4% (base 29.13 → 29.00 — no degradation)
capability — GSM8K / MMLU (n=100)	59.0% / 67.0%

Honest caveats

Swahili degeneration is base-inherent, not from abliteration. On MultiJail-Swahili the abliterated model is 59% degenerate — but base GLM-4-9B is already 56% degenerate on Swahili (a low-resource language this CN/EN model handles poorly). The edit barely moved it.
KL divergence is not a reliable signal for GLM-4. Its 151k-token vocab makes the neutral-prompt softmax extremely peaked, so KL reads ~1e-8 (six orders below other models) even for a real edit — the in-run KL guard is inert here. Capability was therefore judged on ΔPPL (WikiText-2) + GSM8K/MMLU, not KL.
Extraction/eval used advbench, harmbench, MultiJail (zh/ar/sw), sorrybench, orbench_hard.

Rootkit7/GLM-4-9B-abliterated

GLM-4-9B-chat — abliterated (Solutus)

Recipe

Measured (Solutus eval, held-out; base refusal = 100%)

Honest caveats

Rootkit7/
GLM-4-9B-abliterated