# Heretic enables training-free modification of LLMs to remove censorship and slop

> **Key takeaway:** Heretic uses mathematical steering techniques rather than expensive retraining to modify Large Language Model behaviors like censorship, sycophancy, and emotional alignment.

**Published:** 2026-11-05T10:02:38.620386+00:00 
**Source:** GitHub Trending (daily) 
**Category:** foss-apps 
**Canonical URL:** https://fosswire.org/news/heretic-enables-training-free-modification-of-llms-to-remove.html 

## Summary
Heretic is a new open-source tool designed to modify the behavior of Large Language Models (LLMs) without the need for traditional, resource-intensive fine-tuning or training. By utilizing mathematical techniques to steer model responses, users can adjust specific traits such as removing refusal mechanisms (censorship), reducing 'slop,' or altering emotional alignment directly on consumer-grade hardware.

The project supports a wide range of architectures, including dense models, multimodal models, Mixture-of-Experts (MoE), and modern hybrid models. It leverages a TPE-based hyperparameter optimizer powered by Optuna to find optimal parameters that satisfy user-defined scoring criteria, allowing for highly customized model outputs.

## Technical implementation and steering mechanisms
- **Steering Methodology:** Instead of weight training, the tool uses a pluggable modifier system combined with an Optuna-powered TPE optimizer to maximize or minimize specific scores.
- **Modifier Implementations:** The software includes Arbitrary-Rank Ablation (ARA) as the default method, alongside a flexible implementation of directional ablation (abliteration) that supports projected abliteration and norm preservation (MPOA).
- **Scorer Plugin Architecture:** Users can define custom behaviors via scorer plugins (inheriting from an abstract base class) that compute metrics like refusal rates via keyword analysis or KL divergence.
- **Reproducibility Framework:** The tool can generate a 'reproduce' folder containing a machine-readable 'reproduce.json' manifest, enabling byte-for-byte identical model recreation if the software and hardware environments are matched.
- **Hardware Optimization:** The system performs an initial benchmark to determine optimal batch sizes for the specific local hardware, such as an NVIDIA RTX 3090.

## Why it matters
Heretic democratizes model customization by allowing individuals with gaming PCs to perform complex behavioral steering that was previously reserved for large corporations. This provides a significant toolset for the local AI community to create specialized, uncensored, or highly aligned models for research and personal use.