GLM-4-9B-abliterated

GLM-4-9B-chat — abliterated (Solutus)

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%)

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.

Base model © THUDM (GLM-4). See the base model card for its license and usage terms.