Abliterlitics LLM Abliteration Forensics

We take the same base LLM, compare the different abliteration techniques others have applied, then measure what actually changed using benchmarks, safety evaluation, distribution shift, and weight-level analysis.

Methodology Techniques GitHub ↗

Latest

Gemma4-E4B Abliteration Benchmarks: 23 Variants Compared

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.

~4.5B

Gemma 4

Gemma4-E4B Abliteration Benchmarks: 23 Variants Compared ~4.5B 23 techniques Gemma4-E2B Abliteration Benchmarks: 13 Techniques Compared ~2B 13 techniques

Qwen 2.5

Qwen2.5-7B Abliteration Benchmarks: Heretic vs Huihui vs Apostate ~7.6B 3 techniques

Qwen 3.6

Qwen3.6-27B Abliteration Benchmarks: Heretic vs Huihui vs AEON vs Abliterix vs HauhauCS ~27B 5 techniques

GLM 4.7 Flash

GLM-4.7-Flash Abliteration Benchmarks: Heretic vs Huihui vs HauhauCS vs Abliterix ~59B total MoE 4 techniques

Qwen 3.5

Qwen3.5-27B Abliteration Benchmarks: Heretic vs Huihui vs HauhauCS ~27B 3 techniques Qwen3.5-9B Abliteration Benchmarks: Heretic vs Huihui vs HauhauCS ~9B 3 techniques Qwen3.5-4B Abliteration Benchmarks: Heretic vs Huihui vs HauhauCS ~4B 3 techniques Qwen3.5-2B Abliteration Benchmarks: Heretic vs Huihui vs HauhauCS ~2B 3 techniques

Qwen 3

Qwen3-4B Abliteration Benchmarks: Heretic vs Huihui vs HauhauCS ~4B 3 techniques