Heretic is an open-source tool designed to automatically remove censorship from transformer-based language models without the need for expensive post-training.
Source: Official README and linked documentation View on GitHub →Heretic is being discussed because its innovative approach to censorship removal, which is both automatic and effective. It addresses the pain point of manual censorship removal, which is time-consuming and requires expertise. The project differs through its use of directional ablation and TPE-based parameter optimization, which allows for high-quality abliteration without damaging the original model's intelligence.
Source: Official README and linked documentationHeretic automatically finds high-quality abliteration parameters by co-minimizing the number of refusals and KL divergence from the original model, resulting in a decensored model that retains much of the original model's intelligence.
Source: Official README and linked documentationHeretic supports most dense models, including many multimodal models, several different MoE architectures, and even some hybrid models like Qwen3.5.
Source: Official README and linked documentationHeretic provides research features such as generating plots of residual vectors and printing details about residual geometry, which support the study of model semantics and interpretability.
Source: Official README and linked documentationThe architecture of Heretic is modular, with distinct components for analysis, configuration, evaluation, and system management. It uses a TPE-based parameter optimizer and directional ablation techniques. The code structure is organized into a clear hierarchy, with a focus on maintainability and scalability.
Source: Official repository tree and dependency filesacceleratebitsandbytesdatasetshuggingface-hubimmutabledictkernelslangdetectlm-eval[hf]numpyoptunapeftpsutilpy-cpuinfopydantic-settingsquestionaryrichtomli-wtqdmtransformersHeretic is suitable for developers and researchers working with language models who need to remove censorship without manual intervention. It is useful in scenarios where models need to generate responses on sensitive topics, such as political or social issues.
Source: Official README and linked documentationThe latest stable Release is v1.4.0, published on 2026-06-14. It added reproduce.json reproduction, plain-text prompt datasets, LoRA export fixes, and broader model support. The current master branch is already 2.0.0.dev0.
Source: github.com/p-e-w/heretic/releases/tag/v1.4.0status: docs-only | checked_at: 2026-09-23 | environment: Check the official README, pyproject.toml, AGPL-3.0 license, v1.4.0 release against the official listing | limitations: PyTorch or Heretic not installed, models and datasets not downloaded, ablation, evaluation, saving, uploading or chat not running, rejection rate not reproduced, KL, video memory and time consuming | observed_output: The latest official release is v1.4.0, released on 2026-06-14; the current master's pyproject.toml is marked 2.0.0.dev0 and requires Python 3.10 +
Source: github.com/p-e-w/heretic/releases/tag/v1.4.0ErisForge lets researchers select layers, directions, and ablation/addition transforms through a Python API and includes refusal-expression scoring. Heretic automates direction construction, TPE search, dual-objective evaluation, and the CLI workflow. Evaluate ErisForge for manual control and Heretic for automated search with reproducible experiment records.
Source: github.com/Tsadoq/ErisForgeHeretic is a project for those working with language models who need to remove censorship efficiently. It is particularly suitable for developers and researchers who require high-quality abliteration without manual intervention and are willing to invest in computational resources.
Source: Official repositoryThe latest official release is v1.4.0, released on 2026-06-14; the current master's pyproject.toml is marked 2.0.0.dev0 and requires Python 3.10 + PyTorch or Heretic not installed, models and datasets not downloaded, ablation, evaluation, saving, uploading or chat not running, rejection rate not reproduced, KL, video…
The review environment was: Check the official README, pyproject.toml, AGPL-3.0 license, v1.4.0 release against the official listing heretic behavior outside that scope still needs a local check.