The developer behind Heretic, a software suite designed for modifying large language models (LLMs), has announced the release of Heretic 2.0. According to a developer announcement on Reddit, the update marks a transition for the project from a specialized tool for removing censorship to a general-purpose model modification platform.

Technical shifts and new steering methods

The core technical update in version 2.0 is the introduction of Arbitrary-Rank Ablation (ARA). According to the developer, ARA replaces the previous “directional abliteration” method. While the older method operated on the assumption that refusals exist within a single rank-1 subspace, ARA utilizes the L-BFGS optimizer to modify matrices without rank restrictions. The developer claims this method is “vastly superior” to previous iterations, though these performance claims originate from the project’s own documentation and have not been independently verified by third-party researchers.

The update also introduces a modular plugin architecture. This system allows users to implement “Scorer” plugins to define optimization objectives and “Modifier” plugins to execute steering methods. While the software previously relied on hardcoded scores for refusal counts and KL divergence, the new architecture enables users to load custom Python files to optimize for various metrics.

Expanded capabilities and model support

The developer states that Heretic 2.0 can be used to suppress “slop,” reduce sycophancy, and alter the emotional alignment of models. To support these functions, the release includes several pre-configured plugins:

  • A Benchmark Scorer: Capable of computing scores across various standard benchmarks.
  • A BERT Classifier: Developed by contributor Vinay Umrethe, this plugin uses a ModernBERT-type model to classify LLM responses. The developer notes that this method is more accurate than the previous keyword-matching approach, despite being slower.
  • An Emotion Classifier: Provided as an alternative configuration for steering model temperament.

The software has also been updated to support models with hyper-connections and has seen improvements in compatibility with “thinking” models. Additionally, the tool can now run non-interactively, which the developer says facilitates end-to-end testing for quality control.

Scale and compatibility

The Heretic project reports significant scale, claiming 37,000 GitHub stars and over 8,500 models hosted on Hugging Face, with a total of 88 million downloads.

Users migrating from the previous version are advised that Heretic 2.0 is not backwards compatible with Heretic 1.x configurations. The developer has provided a migration guide on GitHub to assist users in updating their existing setups. For basic censorship removal, the developer maintains that the process remains a single-command operation.