Forensic deep-dives into abliteration measurement covering distribution-shift analysis, measurement methods, and what the data actually says about abliteration.
Forensic deep-dives into abliteration measurement. Data, methods, and what the numbers mean. The full data behind every claim sits in the appendices. The text tells you what it means.
Seven uncensored variants built on Google's official QAT checkpoint of Gemma 4 12B, benchmarked across weight forensics, KL divergence, an 11-task suite and judge-reviewed HarmBench, with an independent standard-stack QAT-vs-BF16 baseline and a same-method two-substrate damage study.
Over 330 GPU-hours comparing 12 uncensored variants of Qwen3.8-27B across weight forensics, heretic-exact KL divergence, a 13-task suite and judge-scored HarmBench at a 15,360-token thinking budget, all served in dynamic FP8 on a single RTX 5090. The surgical edits take the judge leaderboard with orcarouter at 82% and apostate at 79%, the followup batch adds a search-loop abliteration that tops the harmful-only table, a merge hybrid that beats base on knowledge tasks, a second-pass ARA and an architecture-extending conditional circuit.
165 GPU-hours comparing 12 uncensored variants of Gemma 4 12B Unified across weight forensics, KL divergence, a 13-task suite and judge-reviewed HarmBench, with the sdft LoRA adapters needing enable_thinking=true for best results.
The first measured answer to whether swapping an abliterated or quantised text encoder uncensors or degrades Z-Image Turbo, across 27 encoder variants.
Whether the eval dataset, token depth, thinking mode, or KL method moves the abliteration capability-preservation score.
77 GPU-hours of abliteration benchmarks for Gemma4-E4B: 23 variants compared across a mixed v1/v2 13-task suite, HarmBench safety, KL divergence, and weight forensics. Heretic is the best capability-safety tradeoff at 95.5% ASR with KL=0.002; abliterix hits 100% ASR.
Three-way abliteration comparison for Qwen2.5-7B-Instruct: Heretic, Huihui, and the new Apostate tool. Heretic achieves 100% ASR with the fewest parameter changes.
44 GPU-hours of abliteration benchmarks for Gemma4-E2B: HarmBench safety scores, capability benchmarks, KL divergence, and weight forensics comparing 13 abliterated variants. Coder3101 achieves the best capability-safety tradeoff at 95.8% ASR while beating base on GSM8K.
85 GPU-hours of abliteration benchmarks for Qwen3.6-27B: HarmBench safety scores, capability benchmarks, KL divergence, and weight forensics comparing 5 techniques. Heretic and Huihui preserve capabilities within 1% of base.
Abliteration forensics for GLM-4.7-Flash 59B MoE: all four techniques achieve 100% HarmBench ASR with capability preservation within 1% of base on most benchmarks.
Abliteration comparison for Qwen3.5-27B hybrid Mamba2+Transformer: benchmarks, HarmBench safety, KL divergence, and weight forensics. Heretic achieves lowest KL at 0.063 with 99.8% ASR.
Abliteration comparison for Qwen3.5-9B: HarmBench ASR, capability benchmarks, KL divergence, and weight forensics. Heretic preserves capabilities best with lowest KL divergence at 0.0825.
Abliteration forensics for Qwen3.5-4B including the Huihui catastrophic KL divergence discovery with strong capability preservation across techniques.
Abliteration comparison for Qwen3.5-2B, the smallest model tested, with the least collateral damage across all metrics.
Abliteration forensics for Qwen3-4B pure Transformer with provenance investigation. HauhauCS ranks best overall, the only model where it outperforms Heretic.