Aesop Ablated (LoRA Only)

Refusal ablation LoRA adapter for GLM-5.2. Removes refusal direction from layers 62-65 using PCA-derived vectors.

This is the LoRA adapter only. Apply to base GLM-5.2 to get the full Aesop Ablated model. Full merged model: cfontes/GLM-5.2-Ablated-Molt

Config

  • Rank: 64, Alpha: 128
  • Ablation layers: 62-65, coefficient: 0.1
  • Method: PCA-derived refusal direction, post-hoc weight surgery

Benchmark Results (when applied to base)

Benchmark Score
AdvBench (refusal) ~1%
HumanEval 67.7%
MMLU-Pro 82%
SimpleQA 88%
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for cfontes/GLM-5.2-Ablated-Molt-LoRA

Base model

zai-org/GLM-5.2
Adapter
(3)
this model