Table of Contents
What is Abliterlitics?
Abliterlitics is an open-source abliteration forensics toolkit. The idea is straightforward. I take the same base model, compare the different abliteration techniques others have applied, then measure what actually changed using benchmarks, safety evaluation, distribution shift, and weight-level analysis.
Why does this matter?
Abliteration techniques remove safety alignment from language models. Multiple tools and methods exist to do this, each claiming different levels of effectiveness and capability preservation. I built Abliterlitics to provide independent, reproducible measurements to evaluate these claims.
Some tools claim to be “lossless” or to “enhance capabilities.” My measurements show these claims are frequently contradicted by the data.
Tools
Abliterlitics
The forensics toolkit itself. Performs weight-level analysis including SVD, fingerprint, edit vector overlap, per-layer analysis, and cross-technique alignment. Available on GitHub .
ungguf
A GGUF-to-safetensors conversion tool. Converts GGUF model files back to HuggingFace safetensors format with bit-exact verification. This enables forensic analysis of models distributed only as GGUF. Available on GitHub .
Abliteration Techniques Compared
| Technique | Description |
|---|---|
| Heretic | Surgical rank-1 LoRA ablation targeting output projections. Open source (AGPL-3.0). Non-deterministic: different runs produce different results. Supports MPOA and experimental ARA methods. |
| Huihui | Standard abliteration with varying scope. Targets down_proj, out_proj, and sometimes additional projections. Performance varies dramatically by model. |
| HauhauCS | Uses Reaper Abliteration, shown to be plagiarised from Heretic under AGPL-3.0 with attribution stripped. Broad modification footprint. Discontinued from future comparisons. |
| AEON | LEACE + rank-k approach. Claims “lossless” and “enhanced capabilities.” |
| Abliterix | Custom variant with router and shared expert targeting on MoE models. |
Author
Nathan Sapwell . I build tools for transparent, reproducible LLM analysis.
Disclaimer
The models analysed on this site have had safety alignment removed. The analysis is performed for research and transparency purposes. I do not condone or encourage the use of abliterated models for harmful purposes.