Abliterlitics ranks Qwen3.6-27B abliteration methods: Heretic & Huihui best

Abliterlitics, an open-source forensics toolkit, conducted a rigorous 85 GPU-hour evaluation comparing five abliteration methods applied to the Qwen3.6-27B base model. The study used 4-bit quantization and identical lm-evaluation-harness settings across all six variants (base + five abliterated). Key benchmarks included MMLU, HellaSwag, ARC Challenge, TruthfulQA, and GSM8K. Heretic and Huihui emerged as the top performers for capability preservation: Huihui showed the smallest benchmark deltas (e.g., MMLU 83.4% vs 83.3% base), while Heretic had the lowest KL divergence from the base. All abliterated models achieved nearly complete safety removal on HarmBench. AEON's claim of 'enhanced capabilities' was contradicted by significant drops in ARC Challenge (56.1% vs 59.1% base) and TruthfulQA (46.1% vs 56.7% base). Abliterix suffered the worst degradation, with Lambada perplexity skyrocketing to 9.12 (base 3.18) and HellaSwag falling to 77.3% (base 83.5%).

The analysis also revealed a serious ethical issue: HauhauCS's 'Reaper Abliteration' tool, used to produce one of the variants, was found to be plagiarized from the Heretic project under AGPL-3.0, with all attribution stripped and relicensed to PolyForm Noncommercial. The recovered source code showed that Reaper added subspace rank-k ablation and SOM clustering on top of Heretic's core, but the model weights were further contaminated by GGUF quantization round-trip noise. As a result, the author will discontinue HauhauCS in all future comparisons. This work provides the first transparent, reproducible benchmark for abliteration techniques, helping the community choose methods that balance safety removal with capability retention while exposing unethical practices.