Heretic 2.0发布：引入ARA算法与插件架构的模型修改工具

Dear fellow Llamas,

after almost 5 months of development, I finally have the pleasure of announcing to you the immediate availability of the biggest Heretic release ever, Heretic 2.0: https://github.com/p-e-w/heretic

Heretic is the leading software for removing censorship from language models. It now boasts 37k GitHub stars and more than 8,500 models on Hugging Face, with 88 million(!!!) total model downloads.

With Heretic 2.0, the project is entering a new era. Heretic is expanding from a purpose-built tool for censorship removal to a general-purpose model modification tool that can do amazing things. Besides removing censorship, it can suppress slop, reduce sycophancy, alter emotional alignment, and pretty much anything else that you can think of. Of course, censorship removal remains the primary and default function, and in fact works much, much better than ever before. Heretic 2.0 is the most advanced refusal suppression system ever built.

Your time is precious, so here's the TL;DR for this release:

• Directional abliteration is replaced by Arbitrary-Rank Ablation (ARA), a novel steering method that I developed from scratch for Heretic. It's vastly superior to abliteration, and has been used with great success by enterprising testers for several months. Now it is the default (classic abliteration remains available as an option).
• Heretic has been completely rebuilt around plugins. Scorer plugins define what Heretic optimizes for. Modifier plugins implement the actual steering method. Plugins are easy to write, can be loaded directly from a simple Python file, and have a well-defined API.
• Heretic can now run non-interactively, which among other things has enabled us to implement full end-to-end tests for much tighter quality control. Model support has been expanded to include models with hyper-connections, and support for thinking models has been substantially improved.
• Heretic 2.0 is not backwards compatible with Heretic 1.x configurations. See the migration guide for what you need to change in order to make your old configurations work.

What hasn't changed is that removing censorship from a model remains as simple as running

See the default configuration file for what Heretic has to offer beyond that.

Now for some details...

ARA: Heretic's new steering method

Directional ablation, or "abliteration", is based on the assumption that refusals (or whatever else you are trying to suppress) are represented by a single direction (that is, a rank-1 subspace) in residual space. That this is the full picture has been doubted in research for a while, and there are now several papers exploring alternatives such as multi-directional ablation or "representation cones".

At the beginning of this year, I started experimenting with a method that completely abandons the idea that the ablation transformation must be restricted to a fixed-rank subspace. In March, I published the initial version of Arbitrary-Rank Ablation in this pull request (which was later closed as I re-implemented ARA on top of the plugin system in this PR).

ARA steers model behavior by modifying matrices based on an objective function that encodes the basic properties of the change we want: Harmful prompts should be represented in residual space in a way that's similar to harmless prompts, while the representations of harmless prompts should change as little as possible. This objective is used with the L-BFGS optimizer to modify matrices one-by-one, a kind of micro-training that steers representations without any rank restriction.

ARA has been extensively tested and refined with the help of the community. Notably, kabachuha made it compatible with Heretic's LoRA system, and varunbln found a substantial performance improvement.

I intend to post an article explaining ARA in more detail soon.

Using scorer plugins to make Heretic do what you want

Heretic operates by finding modification parameters that optimize scores, values that measure the quality of the resulting model. Ever since the 1.0 release, Heretic has used two hardcoded scores: Refusal count to measure decensoring strength, and KL divergence to measure quality retention.

With Heretic 2.0, you can freely define your own scores by writing and loading scorer plugins. Arbitrarily many scorer plugins can be loaded and optimized for simultaneously, and each score can be either minimized or maximized during optimization. See the default configuration file for how to do this.

The scorer plugin system is the result of months of effort by red40maxxer, with whom I exchanged hundreds of messages over dozens of review cycles to figure out the best possible design. This is a monumental contribution that has changed how Heretic works from the ground up.

Besides the keyword (refusal) count and KLD scorer plugins, which remain the default, Heretic ships with two other scorers: A benchmark scorer, which can compute scores for many standard benchmarks, and a BERT classifier.

The BERT classifier, developed by long-time Heretic contributor Vinay Umrethe, can use an arbitrary ModernBERT-type model to classify LLM responses, and score them based on which labels the classifier produces. Vinay also trained a refusal classifier model (used by default with this plugin), which experiments have shown to be substantially more accurate at identifying model refusals than the naive keyword matching method that Heretic employs by default. While the plugin is slower than keyword matching, the greater accuracy can steer Heretic towards producing a higher-quality modified model. An alternative configuration, which uses an emotion classifier, is also provided in the Heretic repository.

The plugin system is well-documented in both the README and the code. If you are interested in experimenting with custom plugins, an agent should have no difficulty cooking up a plugin and a matching configuration file that together do what you want, after ingesting the Heretic repository.

Well, I hope you like it!

Cheers :)