ThinkingCap-Qwen3.6-27B-heretic

This is a decensored version of bottlecapai/ThinkingCap-Qwen3.6-27B, made using Heretic v1.4.0

Heretic removes refusal behavior ("safety alignment") from language models via directional ablation (abliteration), using a TPE-based (Optuna) parameter optimizer to automatically find ablation parameters that minimize refusals while co-minimizing KL divergence from the original model's behavior on harmless prompts.

GGUF quantizations (Q4_K_M, Q6_K, Q8_0, f16) and the mmproj file for this model are available at egibson/ThinkingCap-Qwen3.6-27B-heretic-GGUF.

Abliteration parameters

This model corresponds to trial 187 of 200 optimization trials, selected from the Pareto front for its low refusal rate at negligible KL divergence cost:

Performance

Lower keyword count means fewer refusals. Lower KL divergence means the model's behavior on harmless prompts is closer to the original (values above 0.5 typically indicate significant capability damage — 0.0186 is negligible).

---

Citation

This is a finetuned derivative of Qwen3.6-27B (Qwen Team, 2026) via bottlecapai/ThinkingCap-Qwen3.6-27B, and should respect the original models' licensing requirements. Please cite the base model if you use this one:

@misc{ThinkingCap-Qwen3.6-27B,
  title  = {bottlecapai/ThinkingCap-Qwen3.6-27B},
  author = {Lasocki, Karol and Osusky, Adam and Lindauer, Jan and Jirkovsky, Adam and
            Mihal, Filip and Platek, Ondrej and Herel, David and Ihnatchenko, Luka and
            Bartek, Vojtech and Jirak, Jiri and Mikolov, Tomas},
  year   = {2026},
}

Abliteration performed with Heretic by Philipp Emanuel Weidmann.