p-e-w/hereticPublic

Training-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more

AI summary: Fully automatic censorship removal tool for transformer-based language models using directional ablation.

Stars
42.2K
+1.5K today
Forks
5.2K
Watchers
211
Open issues
34
Open PRs
31
Contributors
~44
Commits
239
Branches
29
Views (30d)
4

PythonAGPL-3.0Created Sep 21, 2025Last push 1d agoLatest release v2.1.0+1.8K stars this week+6.2K this month

Quick answers

What is heretic?
Fully automatic censorship removal tool for transformer-based language models using directional ablation.
What does heretic do?
Heretic is an automated tool that removes safety alignment and censorship from dense and MoE language models without requiring expensive post-training or fine-tuning. It utilizes directional ablation combined with a TPE-based parameter optimizer powered by Optuna. The tool autonomously discovers optimal abliteration parameters by co-minimizing refusal rates on harmful prompts and the KL divergence on harmless prompts. This process preserves the original model's general intelligence and formatting capabilities while bypassing artificial safety guardrails. It also includes advanced research features for performing PaCMAP projections of residual vectors to analyze layer transitions visually.
Who is heretic for?
AI researchers, open-source model developers, and enthusiasts looking to remove alignment filters or study model interpretability. It requires a modern GPU and a Python 3.10+ environment with PyTorch.
How do I get started with heretic?
pip install -U heretic-llm
How popular is heretic on GitHub?
p-e-w/heretic has 42,185 stars and 5,219 forks on GitHub, and gained 1,813 stars in the last 7 days.
What license does heretic use?
p-e-w/heretic is released under the AGPL-3.0 license.

Star history

since Jul 29, 2026
020K40KJul 2026Sep 2026Oct 2026Nov 2026
42.2K stars as of Nov 19, 2026. Measured daily since Jul 29, 2026; GitHub no longer exposes earlier star timestamps.

Update history

3 recorded
  • Nov 8, 2026New release v2.1.0 (previously v2.0.0).
  • Nov 1, 2026New release v2.0.0 (previously v1.4.0).
  • Nov 1, 2026Description changed from "Fully automatic censorship removal for language models" to "Training-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more".

Contribution activity

commits per day, last 52 weeks
NovDecJanFebMarAprMayJunJulAugSepOctNovMonWedFri2025-11-23: 1 commit2025-11-24: 2 commits2025-11-25: 1 commit2025-11-26: 1 commit2025-11-27: 0 commits2025-11-28: 0 commits2025-11-29: 0 commits2025-11-30: 0 commits2025-12-01: 1 commit2025-12-02: 1 commit2025-12-03: 0 commits2025-12-04: 1 commit2025-12-05: 1 commit2025-12-06: 1 commit2025-12-07: 4 commits2025-12-08: 0 commits2025-12-09: 2 commits2025-12-10: 3 commits2025-12-11: 1 commit2025-12-12: 0 commits2025-12-13: 0 commits2025-12-14: 2 commits2025-12-15: 0 commits2025-12-16: 1 commit2025-12-17: 0 commits2025-12-18: 0 commits2025-12-19: 0 commits2025-12-20: 2 commits2025-12-21: 0 commits2025-12-22: 4 commits2025-12-23: 0 commits2025-12-24: 0 commits2025-12-25: 0 commits2025-12-26: 0 commits2025-12-27: 1 commit2025-12-28: 0 commits2025-12-29: 0 commits2025-12-30: 0 commits2025-12-31: 2 commits2026-01-01: 0 commits2026-01-02: 1 commit2026-01-03: 0 commits2026-01-04: 0 commits2026-01-05: 0 commits2026-01-06: 0 commits2026-01-07: 0 commits2026-01-08: 0 commits2026-01-09: 1 commit2026-01-10: 0 commits2026-01-11: 1 commit2026-01-12: 0 commits2026-01-13: 0 commits2026-01-14: 0 commits2026-01-15: 0 commits2026-01-16: 1 commit2026-01-17: 0 commits2026-01-18: 0 commits2026-01-19: 0 commits2026-01-20: 1 commit2026-01-21: 0 commits2026-01-22: 0 commits2026-01-23: 2 commits2026-01-24: 0 commits2026-01-25: 2 commits2026-01-26: 0 commits2026-01-27: 0 commits2026-01-28: 0 commits2026-01-29: 0 commits2026-01-30: 0 commits2026-01-31: 0 commits2026-02-01: 0 commits2026-02-02: 2 commits2026-02-03: 0 commits2026-02-04: 0 commits2026-02-05: 0 commits2026-02-06: 0 commits2026-02-07: 0 commits2026-02-08: 3 commits2026-02-09: 0 commits2026-02-10: 0 commits2026-02-11: 3 commits2026-02-12: 0 commits2026-02-13: 1 commit2026-02-14: 6 commits2026-02-15: 1 commit2026-02-16: 0 commits2026-02-17: 1 commit2026-02-18: 0 commits2026-02-19: 0 commits2026-02-20: 1 commit2026-02-21: 0 commits2026-02-22: 0 commits2026-02-23: 0 commits2026-02-24: 0 commits2026-02-25: 0 commits2026-02-26: 0 commits2026-02-27: 1 commit2026-02-28: 0 commits2026-03-01: 0 commits2026-03-02: 0 commits2026-03-03: 0 commits2026-03-04: 0 commits2026-03-05: 0 commits2026-03-06: 2 commits2026-03-07: 0 commits2026-03-08: 0 commits2026-03-09: 0 commits2026-03-10: 0 commits2026-03-11: 1 commit2026-03-12: 0 commits2026-03-13: 1 commit2026-03-14: 0 commits2026-03-15: 2 commits2026-03-16: 0 commits2026-03-17: 0 commits2026-03-18: 0 commits2026-03-19: 0 commits2026-03-20: 0 commits2026-03-21: 0 commits2026-03-22: 0 commits2026-03-23: 0 commits2026-03-24: 1 commit2026-03-25: 0 commits2026-03-26: 0 commits2026-03-27: 0 commits2026-03-28: 1 commit2026-03-29: 0 commits2026-03-30: 0 commits2026-03-31: 0 commits2026-04-01: 1 commit2026-04-02: 0 commits2026-04-03: 0 commits2026-04-04: 8 commits2026-04-05: 0 commits2026-04-06: 0 commits2026-04-07: 1 commit2026-04-08: 1 commit2026-04-09: 0 commits2026-04-10: 0 commits2026-04-11: 1 commit2026-04-12: 2 commits2026-04-13: 0 commits2026-04-14: 1 commit2026-04-15: 0 commits2026-04-16: 0 commits2026-04-17: 0 commits2026-04-18: 1 commit2026-04-19: 0 commits2026-04-20: 0 commits2026-04-21: 1 commit2026-04-22: 0 commits2026-04-23: 2 commits2026-04-24: 0 commits2026-04-25: 1 commit2026-04-26: 0 commits2026-04-27: 0 commits2026-04-28: 0 commits2026-04-29: 0 commits2026-04-30: 0 commits2026-05-01: 0 commits2026-05-02: 2 commits2026-05-03: 3 commits2026-05-04: 1 commit2026-05-05: 1 commit2026-05-06: 0 commits2026-05-07: 1 commit2026-05-08: 0 commits2026-05-09: 2 commits2026-05-10: 0 commits2026-05-11: 0 commits2026-05-12: 0 commits2026-05-13: 0 commits2026-05-14: 0 commits2026-05-15: 0 commits2026-05-16: 1 commit2026-05-17: 0 commits2026-05-18: 0 commits2026-05-19: 0 commits2026-05-20: 0 commits2026-05-21: 0 commits2026-05-22: 3 commits2026-05-23: 0 commits2026-05-24: 0 commits2026-05-25: 0 commits2026-05-26: 0 commits2026-05-27: 0 commits2026-05-28: 1 commit2026-05-29: 0 commits2026-05-30: 0 commits2026-05-31: 3 commits2026-06-01: 0 commits2026-06-02: 0 commits2026-06-03: 1 commit2026-06-04: 2 commits2026-06-05: 2 commits2026-06-06: 1 commit2026-06-07: 2 commits2026-06-08: 0 commits2026-06-09: 1 commit2026-06-10: 0 commits2026-06-11: 2 commits2026-06-12: 0 commits2026-06-13: 1 commit2026-06-14: 1 commit2026-06-15: 0 commits2026-06-16: 0 commits2026-06-17: 2 commits2026-06-18: 2 commits2026-06-19: 0 commits2026-06-20: 0 commits2026-06-21: 0 commits2026-06-22: 0 commits2026-06-23: 0 commits2026-06-24: 0 commits2026-06-25: 0 commits2026-06-26: 0 commits2026-06-27: 2 commits2026-06-28: 0 commits2026-06-29: 0 commits2026-06-30: 0 commits2026-07-01: 1 commit2026-07-02: 0 commits2026-07-03: 0 commits2026-07-04: 0 commits2026-07-05: 0 commits2026-07-06: 0 commits2026-07-07: 1 commit2026-07-08: 0 commits2026-07-09: 0 commits2026-07-10: 0 commits2026-07-11: 0 commits2026-07-12: 0 commits2026-07-13: 0 commits2026-07-14: 0 commits2026-07-15: 1 commit2026-07-16: 1 commit2026-07-17: 1 commit2026-07-18: 0 commits2026-07-19: 0 commits2026-07-20: 0 commits2026-07-21: 0 commits2026-07-22: 2 commits2026-07-23: 0 commits2026-07-24: 1 commit2026-07-25: 0 commits2026-07-26: 0 commits2026-07-27: 0 commits2026-07-28: 2 commits2026-07-29: 0 commits2026-07-30: 0 commits2026-07-31: 0 commits2026-08-01: 0 commits2026-08-02: 0 commits2026-08-03: 0 commits2026-08-04: 0 commits2026-08-05: 0 commits2026-08-06: 1 commit2026-08-07: 0 commits2026-08-08: 0 commits2026-08-09: 0 commits2026-08-10: 0 commits2026-08-11: 0 commits2026-08-12: 0 commits2026-08-13: 0 commits2026-08-14: 1 commit2026-08-15: 0 commits2026-08-16: 0 commits2026-08-17: 3 commits2026-08-18: 0 commits2026-08-19: 0 commits2026-08-20: 0 commits2026-08-21: 0 commits2026-08-22: 0 commits2026-08-23: 0 commits2026-08-24: 0 commits2026-08-25: 0 commits2026-08-26: 0 commits2026-08-27: 0 commits2026-08-28: 0 commits2026-08-29: 0 commits2026-08-30: 0 commits2026-08-31: 0 commits2026-09-01: 0 commits2026-09-02: 0 commits2026-09-03: 2 commits2026-09-04: 0 commits2026-09-05: 2 commits2026-09-06: 0 commits2026-09-07: 0 commits2026-09-08: 0 commits2026-09-09: 0 commits2026-09-10: 0 commits2026-09-11: 0 commits2026-09-12: 0 commits2026-09-13: 0 commits2026-09-14: 0 commits2026-09-15: 0 commits2026-09-16: 0 commits2026-09-17: 0 commits2026-09-18: 0 commits2026-09-19: 0 commits2026-09-20: 0 commits2026-09-21: 0 commits2026-09-22: 0 commits2026-09-23: 0 commits2026-09-24: 0 commits2026-09-25: 0 commits2026-09-26: 0 commits2026-09-27: 0 commits2026-09-28: 0 commits2026-09-29: 0 commits2026-09-30: 1 commit2026-10-01: 1 commit2026-10-02: 2 commits2026-10-03: 0 commits2026-10-04: 1 commit2026-10-05: 0 commits2026-10-06: 2 commits2026-10-07: 2 commits2026-10-08: 0 commits2026-10-09: 0 commits2026-10-10: 0 commits2026-10-11: 0 commits2026-10-12: 1 commit2026-10-13: 0 commits2026-10-14: 0 commits2026-10-15: 0 commits2026-10-16: 0 commits2026-10-17: 1 commit2026-10-18: 0 commits2026-10-19: 2 commits2026-10-20: 3 commits2026-10-21: 0 commits2026-10-22: 0 commits2026-10-23: 1 commit2026-10-24: 2 commits2026-10-25: 0 commits2026-10-26: 2 commits2026-10-27: 0 commits2026-10-28: 1 commit2026-10-29: 1 commit2026-10-30: 1 commit2026-10-31: 3 commits2026-11-01: 1 commit2026-11-02: 0 commits2026-11-03: 2 commits2026-11-04: 0 commits2026-11-05: 4 commits2026-11-06: 1 commit2026-11-07: 1 commit2026-11-08: 2 commits2026-11-09: 0 commits2026-11-10: 1 commit2026-11-11: 0 commits2026-11-12: 0 commits2026-11-13: 0 commits2026-11-14: 0 commits2026-11-15: 0 commits2026-11-16: 0 commits2026-11-17: 0 commits2026-11-18: 0 commits2026-11-19: 0 commits2026-11-20: 0 commits2026-11-21: 0 commits
183 commits in the last yearLessMore

Signals and awards

derived from tracked data
  • Widely adopted

    42,185 stars

  • High momentum

    +1,503 stars today

  • Actively maintained

    Pushed within 48 hours

  • Continuous integration

    Automated checks passing

  • Repeat trending

    20 trending appearances

What heretic does

Heretic is an automated tool that removes safety alignment and censorship from dense and MoE language models without requiring expensive post-training or fine-tuning. It utilizes directional ablation combined with a TPE-based parameter optimizer powered by Optuna. The tool autonomously discovers optimal abliteration parameters by co-minimizing refusal rates on harmful prompts and the KL divergence on harmless prompts. This process preserves the original model's general intelligence and formatting capabilities while bypassing artificial safety guardrails. It also includes advanced research features for performing PaCMAP projections of residual vectors to analyze layer transitions visually.

AI researchers, open-source model developers, and enthusiasts looking to remove alignment filters or study model interpretability. It requires a modern GPU and a Python 3.10+ environment with PyTorch.

  • Automated Abliteration: Completely automates the directional ablation process without requiring manual parameter tuning.
  • Intelligence Preservation: Minimizes KL divergence to ensure the decensored model retains its original instruction-following capabilities.
  • Broad Architecture Support: Processes most dense transformers, MoE architectures, and hybrid models like Qwen3.5.
  • Hardware Optimization: Automatically benchmarks system capabilities at startup to determine the optimal batch size for VRAM efficiency.
  • Residual Geometry Analysis: Generates 2D PaCMAP scatter plots and geometric metrics to visualize hidden state transformations.
  • Built-in Evaluation: Includes standard automated evaluation functionality to verify refusal rates and performance impact immediately.

Where teams use it

Model Decensorship

Remove restrictive safety filters from open-weight language models for unfiltered interactions.

Interpretability Research

Analyze how harmful and harmless prompts diverge in the residual streams of transformer layers.

Automated Experimentation

Run optimization sweeps to find the perfect ablation parameters for newly released language models.

Low-Resource Modification

Alter model behavior significantly without the immense compute costs required for standard fine-tuning.

Getting started: pip install -U heretic-llm

README

master branch
Heretic logo

Heretic

Training-free language model modification

Remove censorship, suppress slop, reduce sycophancy, and more.
All in an afternoon, on your gaming PC.

Discord     Matrix    Follow us on Hugging Face    Codeberg mirror


 

#1 Repository of the Day     Model and download statistics

Modify any LLM that you can run on your hardware1

Heretic is a unique tool that lets you do what only corporations with access to massive computing power usually do: Modify, rather than merely run, language models. It does not perform expensive training; rather, it uses a collection of mathematical techniques to steer behavior based on contrasting example prompts and matching scorers.

Heretic can, depending on the configuration:

and many, many other things.

‼️ Using Heretic does NOT require understanding how language models work internally ‼️

In fact, you can go from zero to removing censorship from your first model in a matter of seconds. Heretic supports most dense models, including many multimodal models, several different MoE architectures, and many modern hybrid models.

The community has created and published thousands of models with Heretic. Heretic models have been independently benchmarked using standard metrics like MMLU and GSM8K, and have been found to compare favorably against models produced with competing abliteration tools.23

 

Heretic screenshot

 

Usage

Prepare a Python 3.10+ environment with PyTorch 2.6+ installed as appropriate for your hardware. Then run:

pip install -U heretic-llm
heretic Qwen/Qwen3-4B-Instruct-2507

Replace Qwen/Qwen3-4B-Instruct-2507 with whatever model you want to process. By default, Heretic performs censorship removal (refusal suppression). If you want to do something else, choose one of the configuration files linked above, or write your own. See config.default.toml for available configuration options. Heretic loads the configuration from config.toml in the working directory by default, or from the file path in the HERETIC_CONFIG_FILE environment variable.

Important

Configuration files made for Heretic 1.x are not compatible with Heretic 2.x! See the migration guide for what you need to change in order to make your old configurations work.

At the start of a program run, Heretic benchmarks the system to determine the optimal batch size to make the most of the available hardware. On an RTX 3090, with the default configuration, processing Qwen3-4B-Instruct-2507 takes about 30-40 minutes.

After Heretic has finished processing a model, you are given the choice to save the model, upload it to Hugging Face, chat with it to test how well it works, run standard benchmarks on it, or any combination of those actions.

Controlling what Heretic does using scorer plugins

Heretic steers model behavior by optimizing one or multiple scores. Scores are computed by scorer plugins. You can easily write your own scorers to get full control over what Heretic does to models.

The Scorer interface (abstract base class) your scorer needs to inherit from is defined in scorer.py. The only required method is get_score, but there is more machinery (such as settings management) available when you need it.

By default, Heretic scores models by refusal rate (implemented in keyword_rate.py) and KL divergence (implemented in kl_divergence.py). Additional scorers are available in the scorers directory. These scorers can serve as examples for your own implementations, and can easily be adapted to your needs.

Scorers are configured through the main configuration file. Which scorer plugins are active is determined by the scorers setting, which also controls whether a scorer's score should be minimized or maximized during optimization. Any number of scorers can be used simultaneously, and their scores will be co-optimized. Multiple instances of the same scorer can be configured with different settings. See the configuration file for the full documentation of this mechanism.

The path to a scorer plugin can be specified either as a fully qualified import path (fully.qualified.module.MyScorerClass) or as a file path followed by a class name (path/to/plugin.py:MyScorerClass). You do not have to modify the Heretic package code in order to use your own plugins.

Heretic also has modifier plugins, which are described in the next section.

How Heretic works

Heretic combines a pluggable modifier (interface in modifier.py) with a TPE-based hyperparameter optimizer powered by Optuna. The optimizer finds a set of parameters that, when passed to the modifier, minimize or maximize a set of configurable scores on the resulting model (see above).

Heretic ships with two modifiers:

  • Arbitrary-Rank Ablation (ARA)4: This is the default modifier, and was specifically developed for use in Heretic. You can find an overview of how ARA works in the original pull request, and see the precise implementation in ara.py.

  • Directional ablation aka "abliteration"5: The classic method for refusal suppression. Heretic's implementation uses a flexible ablation weight kernel over the layers, and supports both projected abliteration6 and norm preservation (MPOA)7. The complete code is in abliteration.py.

You can use the modifiers setting to switch between modifier implementations. Despite the name of the setting using a plural form, Heretic currently allows only a single modifier for a given processing run. This may change in the future.

Reproducibility and the Heretic Grimoire

Heretic has the ability to create reproducible models whose repositories contain all information required to produce a byte-for-byte identical model on demand, provided that the environment is perfectly identical to the original one. PyTorch doesn't guarantee that operations return identical outputs unless hardware, drivers, operating system, and the entire software stack including all package versions are exactly the same.

Heretic manages all that complexity for you. When uploading a processed model to Hugging Face, Heretic offers to include reproducibility information. This is entirely optional, but if you choose to do so, the uploaded repository will contain a reproduce folder with detailed instructions for recreating the model (example). This folder contains both human-readable instructions and a machine-readable manifest file named reproduce.json, which includes all the information Heretic needs to automatically reproduce the model.

Automatic reproduction can be performed with

heretic --reproduce path/to/reproduce.json

Heretic will guide you through the process, checking your environment against the one that was used to create the model, and pointing out potentially problematic mismatches. The multi-hour computations that were required to make the original model do not have to be re-done, and the entire process typically takes a few minutes. After you have exported the resulting model, Heretic will verify the hashes of the weight files against those stored in the reproduction manifest (they may or may not be identical, depending on how closely your system resembles the original one).

Many people have uploaded models with reproducibility information to Hugging Face. Since a reproduction manifest is only around 9 kilobytes in size, these manifests form a portable library of models that can be reproduced at will.

That library is called the Heretic Grimoire. You can view it here.

But you don't have to rely on an externally hosted web application to access the library of manifests: The Grimoire is fundamentally designed as a distributed system, and you can fetch your own local copies of publicly available manifests from Hugging Face whenever you want by running

heretic --collect-reproducibles path/to/target/folder/

You now have a local backup of all reproducible Heretic models, properly catalogued. To update this collection, simply run the command again. It functions as an append-only backup, never deleting files even if the corresponding model no longer exists on Hugging Face.

Heretic is made by humans!

Unlike lots of AI-related software today (and indeed lots of software in general), Heretic is not "vibe-coded".

All of my commits to this project until November 2026 were written 100% by hand, the way I have been doing programming for decades. I have been using LLM assistance very sparingly since. While many Heretic contributors use coding agents, I carefully review every single line of code that I merge. There's no "LGTM" here!

Don't believe me? Check out this pull request, which accumulated almost 350 human-written comments over more than 7 months as we discussed the implementation. And here is another PR with more than 200 comments.

I am the opposite of a luddite and I have used agents liberally for other projects, but I am proud of the code I write and so far can't quite bring myself to let a machine write it for me in a project I care so much about. As a result, Heretic has a lean, relatively simple codebase that in my opinion, and in that of some others, is a pleasure to work on.

If you like that, join us! Hop on the Discord and tell us about your ideas for improvements, or directly file an issue or open a PR.

Prior art

I'm aware of the following implementations of abliteration techniques that were publicly available when Heretic was first released (November 2025):

Note that Heretic was written from scratch, and does not reuse code from any of those projects.

Acknowledgments

The development of Heretic was informed by:

Citation

If you use Heretic for your research, please cite it using the following BibTeX entry:

@misc{heretic,
  author = {Weidmann, Philipp Emanuel},
  title = {Heretic: Training-free language model modification},
  year = {2025},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/p-e-w/heretic}}
}

License

Copyright © 2025-2026 Philipp Emanuel Weidmann ([email protected]) + contributors

This program is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Affero General Public License for more details.

You should have received a copy of the GNU Affero General Public License along with this program. If not, see https://www.gnu.org/licenses/.

By contributing to this project, you agree to release your contributions under the same license.

 

 

 

Footnotes

  1. Heretic supports on-the-fly 4-bit model quantization with bitsandbytes, which can reduce the amount of VRAM required to process a model to the same amount that is used to run a typical quantized version of it, at a slight accuracy cost. Set the quantization option to bnb_4bit to enable this. ↩

  2. Qwen2.5-7B: Three-Way Abliteration Comparison: "Heretic is the most surgical and most effective." ↩

  3. Gemma 4 E4B: 23 Abliteration Variants Compared: "Best overall: heretic." ↩

  4. Weidmann 2026. A preprint explaining the technique will be available soon. ↩

  5. Arditi et al. 2024 ↩

  6. Lai 2025 (1) ↩

  7. Lai 2025 (2) ↩

View on GitHub

Recent activity

commits and pull requests

Recent open issues

view all

Discussions

all 12

Releases and announcements

7 total
  1. v2.1.0v2.1.0Nov 8, 2026509 downloads

    ## Changes * chore: bump version to 2.1.0.dev0 by @p-e-w in https://github.com/p-e-w/heretic/pull/558 * feat: allow specifying a custom config file path by @dreamfast in https://github.com/p-e-w/heretic/pull/557 * feat: add response length scorer by @p-e-w in https://github.com/p-e-w/heretic/pull/562 * fix: enforce memory residency for all model tensors by @p-e-w in https://github.com/p-e-w/heretic/pull/564 * fix: prevent errors during model reset from interrupting export by @p-e-w in https://github.com/p-e-w/heretic/pull/565 * perf: disable KV cache when obtaining module I/O by @p-e-w in https://github.com/p-e-w/heretic/pull/566 * fix: make batch size caching in benchmark scorer actually work by @p-e-w in https://github.com/p-e-w/heretic/pull/567 * fix: export plugin settings correctly by @p-e-w in https://github.com/p-e-w/heretic/pull/570 (additional tests contributed by @Vinay-Umrethe) * docs: prepare for 2.1 release by @p-e-w in https://github.com/p-e-w/heretic/pull/572 ## New Contributors * @dreamfast made their first contribution in https://github.com/p-e-w/heretic/pull/557 **Full Changelog**: https://github.com/p-e-w/heretic/compare/v2.0.0...v2.1.0

  2. v2.0.0v2.0.0Nov 1, 20261K downloads

    ## Breaking Changes Heretic 2.0 is a major release that breaks backwards compatibility: * Configuration files made for Heretic 1.x are **not** compatible with Heretic 2.0! See the [migration guide](https://github.com/p-e-w/heretic/blob/master/docs/migration_v1_to_v2.md) for what you need to change in order to make your old configurations work. * Research features have been removed. The intended replacement is [Lophius](https://lophius.org/). ## Changes * fix: correct total trial count when adding additional trials by @kali113 in https://github.com/p-e-w/heretic/pull/385 * feat: let the optimizer disable MLP ablation via a 0 max_weight floor by @rocker-zhang in https://github.com/p-e-w/heretic/pull/387 * fix: resolve UnicodeEncodeError on Windows during model evaluation by @umran666 in https://github.com/p-e-w/heretic/pull/389 * feat: headless operation + end-to-end tests by @p-e-w in https://github.com/p-e-w/heretic/pull/392 * fix: use `W_org` matrix only where needed by @Vinay-Umrethe in https://github.com/p-e-w/heretic/pull/398 * feat: generic plugin system by @red40maxxer in https://github.com/p-e-w/heretic/pull/53 * fix: validate scorer instance names in conf

  3. v1.4.0v1.4.0Jun 14, 20267.5K downloads

    ## Changes * @p-e-w implemented automatically reproducing a model from a `reproduce.json` file in https://github.com/p-e-w/heretic/pull/326. @Vinay-Umrethe fixed issues in that implementation in https://github.com/p-e-w/heretic/pull/352. * @rocker-zhang added support for plain text files as prompt datasets in https://github.com/p-e-w/heretic/pull/337, based on earlier work by @ricyoung in https://github.com/p-e-w/heretic/pull/103. * @anrp fixed LoRA export in https://github.com/p-e-w/heretic/pull/321 * @zaakirio implemented saving the processor for multimodal models in https://github.com/p-e-w/heretic/pull/353 * @MoonRide303 added support for gemma-4-12B-it in https://github.com/p-e-w/heretic/pull/350 * @coder3101 added support for LiquidAI/LFM2.5 in https://github.com/p-e-w/heretic/pull/344 * @UnstableLlama add a config file for suppressing humor in https://github.com/p-e-w/heretic/pull/340 * @umran666 fixed a trial counting bug in https://github.com/p-e-w/heretic/pull/357 * @umran666 made resetting the model null-safe to handle study cancellations in https://github.com/p-e-w/heretic/pull/367 * @umran666 improved exception formatting in https://github.com/p-e-w/heretic

  4. v1.3.0v1.3.0May 5, 2026445 downloads

    ## Changes * @Vinay-Umrethe (who had previously contributed under the username @Vinayyyy7) implemented reproducible runs in https://github.com/p-e-w/heretic/pull/191. @p-e-w revised and improved that implementation in https://github.com/p-e-w/heretic/pull/303. * @magiccodingman reduced peak VRAM usage in https://github.com/p-e-w/heretic/pull/239. @olekssy fixed a bug in that implementation in https://github.com/p-e-w/heretic/pull/301. * @farolone added support for Qwen3.5 models in https://github.com/p-e-w/heretic/pull/187 * @MoonRide303 added support for Gemma 4 models in https://github.com/p-e-w/heretic/pull/287 * @erm14254 made sure all abliterable components across layers are displayed in https://github.com/p-e-w/heretic/pull/215 * @cpagac fixed VRAM usage reporting for multi-GPU setups in https://github.com/p-e-w/heretic/pull/169 * @cpagac fixed a division-by-zero error in the evaluator in https://github.com/p-e-w/heretic/pull/225 * @spikymoth improved automatic response prefix determination with a two-step process in https://github.com/p-e-w/heretic/pull/194 * @spikymoth added model card generation for local models with an existing README in https://github.com/p-e-

  5. v1.2.0v1.2.0Feb 14, 2026142 downloads

    ## Changes * @noctrex added a `max_memory` setting to limit memory usage in https://github.com/p-e-w/heretic/pull/83 * @spikymoth added a mechanism to avoid excessive low-divergence iteration in https://github.com/p-e-w/heretic/pull/73 * @accemlcc implemented a new LoRA-based abliteration engine with support for 4-bit quantization in https://github.com/p-e-w/heretic/pull/60 * @accemlcc added enumeration of all available GPUs on startup in https://github.com/p-e-w/heretic/pull/86 * @Vinayyyy7 added the ability to run more trials after optimization is complete in https://github.com/p-e-w/heretic/pull/76 * @anrp fixed MXFP4 loading in https://github.com/p-e-w/heretic/pull/107 * @anrp refactored the save machinery in https://github.com/p-e-w/heretic/pull/110 * @anrp added broad support for VL models in https://github.com/p-e-w/heretic/pull/108 * @anrp implemented saving and resuming optimization progress in https://github.com/p-e-w/heretic/pull/106, https://github.com/p-e-w/heretic/pull/119, and https://github.com/p-e-w/heretic/pull/116 * @spikymoth implemented Magnitude-Preserving Orthogonal Ablation in https://github.com/p-e-w/heretic/pull/52 * @salmanmkc upgraded GitHub A

Code frequency

additions and deletions
+5.1K-5.1KWeek of 2025-11-23: +203 linesWeek of 2025-11-23: -23 linesWeek of 2025-11-30: +1,284 linesWeek of 2025-11-30: -64 linesWeek of 2025-12-07: +261 linesWeek of 2025-12-07: -89 linesWeek of 2025-12-14: +2,704 linesWeek of 2025-12-14: -1,733 linesWeek of 2025-12-21: +228 linesWeek of 2025-12-21: -94 linesWeek of 2025-12-28: +105 linesWeek of 2025-12-28: -41 linesWeek of 2026-01-04: +9 linesWeek of 2026-01-04: -0 linesWeek of 2026-01-11: +182 linesWeek of 2026-01-11: -2 linesWeek of 2026-01-18: +192 linesWeek of 2026-01-18: -60 linesWeek of 2026-01-25: +5 linesWeek of 2026-01-25: -3 linesWeek of 2026-02-01: +172 linesWeek of 2026-02-01: -23 linesWeek of 2026-02-08: +293 linesWeek of 2026-02-08: -242 linesWeek of 2026-02-15: +40 linesWeek of 2026-02-15: -10 linesWeek of 2026-02-22: +29 linesWeek of 2026-02-22: -13 linesWeek of 2026-03-01: +39 linesWeek of 2026-03-01: -19 linesWeek of 2026-03-08: +11 linesWeek of 2026-03-08: -5 linesWeek of 2026-03-15: +160 linesWeek of 2026-03-15: -104 linesWeek of 2026-03-22: +791 linesWeek of 2026-03-22: -16 linesWeek of 2026-03-29: +218 linesWeek of 2026-03-29: -221 linesWeek of 2026-04-05: +1,047 linesWeek of 2026-04-05: -142 linesWeek of 2026-04-12: +208 linesWeek of 2026-04-12: -121 linesWeek of 2026-04-19: +463 linesWeek of 2026-04-19: -331 linesWeek of 2026-04-26: +35 linesWeek of 2026-04-26: -32 linesWeek of 2026-05-03: +282 linesWeek of 2026-05-03: -108 linesWeek of 2026-05-10: +8 linesWeek of 2026-05-10: -7 linesWeek of 2026-05-17: +9 linesWeek of 2026-05-17: -8 linesWeek of 2026-05-24: +2 linesWeek of 2026-05-24: -0 linesWeek of 2026-05-31: +235 linesWeek of 2026-05-31: -134 linesWeek of 2026-06-07: +782 linesWeek of 2026-06-07: -245 linesWeek of 2026-06-14: +155 linesWeek of 2026-06-14: -112 linesWeek of 2026-06-21: +717 linesWeek of 2026-06-21: -274 linesWeek of 2026-06-28: +214 linesWeek of 2026-06-28: -14 linesWeek of 2026-07-05: +1,177 linesWeek of 2026-07-05: -389 linesWeek of 2026-07-12: +90 linesWeek of 2026-07-12: -27 linesWeek of 2026-07-19: +277 linesWeek of 2026-07-19: -145 linesWeek of 2026-07-26: +93 linesWeek of 2026-07-26: -97 linesWeek of 2026-08-02: +151 linesWeek of 2026-08-02: -151 linesWeek of 2026-08-09: +1 linesWeek of 2026-08-09: -1 linesWeek of 2026-08-16: +20 linesWeek of 2026-08-16: -16 linesWeek of 2026-08-23: +0 linesWeek of 2026-08-23: -0 linesWeek of 2026-08-30: +214 linesWeek of 2026-08-30: -79 linesWeek of 2026-09-06: +0 linesWeek of 2026-09-06: -0 linesWeek of 2026-09-13: +0 linesWeek of 2026-09-13: -0 linesWeek of 2026-09-20: +0 linesWeek of 2026-09-20: -0 linesWeek of 2026-09-27: +5,141 linesWeek of 2026-09-27: -3,990 linesWeek of 2026-10-04: +174 linesWeek of 2026-10-04: -48 linesWeek of 2026-10-11: +29 linesWeek of 2026-10-11: -4 linesWeek of 2026-10-18: +595 linesWeek of 2026-10-18: -150 linesWeek of 2026-10-25: +1,207 linesWeek of 2026-10-25: -100 linesWeek of 2026-11-01: +963 linesWeek of 2026-11-01: -178 linesWeek of 2026-11-08: +224 linesWeek of 2026-11-08: -16 linesWeek of 2026-11-15: +0 linesWeek of 2026-11-15: -0 linesNov 23, 2025Nov 15, 2026
+21.4K lines added, -9.7K removed over the last year.

Commits per week

last 52 weeks
130Week of 2025-11-23: 5 commitsWeek of 2025-11-30: 5 commitsWeek of 2025-12-07: 10 commitsWeek of 2025-12-14: 5 commitsWeek of 2025-12-21: 5 commitsWeek of 2025-12-28: 3 commitsWeek of 2026-01-04: 1 commitsWeek of 2026-01-11: 2 commitsWeek of 2026-01-18: 3 commitsWeek of 2026-01-25: 2 commitsWeek of 2026-02-01: 2 commitsWeek of 2026-02-08: 13 commitsWeek of 2026-02-15: 3 commitsWeek of 2026-02-22: 1 commitsWeek of 2026-03-01: 2 commitsWeek of 2026-03-08: 2 commitsWeek of 2026-03-15: 2 commitsWeek of 2026-03-22: 2 commitsWeek of 2026-03-29: 9 commitsWeek of 2026-04-05: 3 commitsWeek of 2026-04-12: 4 commitsWeek of 2026-04-19: 4 commitsWeek of 2026-04-26: 2 commitsWeek of 2026-05-03: 8 commitsWeek of 2026-05-10: 1 commitsWeek of 2026-05-17: 3 commitsWeek of 2026-05-24: 1 commitsWeek of 2026-05-31: 9 commitsWeek of 2026-06-07: 6 commitsWeek of 2026-06-14: 5 commitsWeek of 2026-06-21: 2 commitsWeek of 2026-06-28: 1 commitsWeek of 2026-07-05: 1 commitsWeek of 2026-07-12: 3 commitsWeek of 2026-07-19: 3 commitsWeek of 2026-07-26: 2 commitsWeek of 2026-08-02: 1 commitsWeek of 2026-08-09: 1 commitsWeek of 2026-08-16: 3 commitsWeek of 2026-08-23: 0 commitsWeek of 2026-08-30: 4 commitsWeek of 2026-09-06: 0 commitsWeek of 2026-09-13: 0 commitsWeek of 2026-09-20: 0 commitsWeek of 2026-09-27: 4 commitsWeek of 2026-10-04: 5 commitsWeek of 2026-10-11: 2 commitsWeek of 2026-10-18: 8 commitsWeek of 2026-10-25: 8 commitsWeek of 2026-11-01: 9 commitsWeek of 2026-11-08: 3 commitsWeek of 2026-11-15: 0 commitsNov 23, 2025Nov 15, 2026
183 commits in the last 52 weeks.

When work happens

weekday and hour
SunMonTueWedThuFriSat036912151821Sun 0:00 — 0 commitsSun 1:00 — 0 commitsSun 2:00 — 1 commitsSun 3:00 — 0 commitsSun 4:00 — 0 commitsSun 5:00 — 0 commitsSun 6:00 — 2 commitsSun 7:00 — 1 commitsSun 8:00 — 3 commitsSun 9:00 — 7 commitsSun 10:00 — 3 commitsSun 11:00 — 5 commitsSun 12:00 — 2 commitsSun 13:00 — 3 commitsSun 14:00 — 3 commitsSun 15:00 — 2 commitsSun 16:00 — 4 commitsSun 17:00 — 4 commitsSun 18:00 — 1 commitsSun 19:00 — 1 commitsSun 20:00 — 0 commitsSun 21:00 — 0 commitsSun 22:00 — 0 commitsSun 23:00 — 0 commitsMon 0:00 — 0 commitsMon 1:00 — 1 commitsMon 2:00 — 0 commitsMon 3:00 — 0 commitsMon 4:00 — 1 commitsMon 5:00 — 1 commitsMon 6:00 — 0 commitsMon 7:00 — 0 commitsMon 8:00 — 0 commitsMon 9:00 — 0 commitsMon 10:00 — 3 commitsMon 11:00 — 2 commitsMon 12:00 — 1 commitsMon 13:00 — 0 commitsMon 14:00 — 0 commitsMon 15:00 — 2 commitsMon 16:00 — 2 commitsMon 17:00 — 1 commitsMon 18:00 — 0 commitsMon 19:00 — 1 commitsMon 20:00 — 1 commitsMon 21:00 — 2 commitsMon 22:00 — 2 commitsMon 23:00 — 1 commitsTue 0:00 — 0 commitsTue 1:00 — 1 commitsTue 2:00 — 0 commitsTue 3:00 — 1 commitsTue 4:00 — 0 commitsTue 5:00 — 2 commitsTue 6:00 — 1 commitsTue 7:00 — 1 commitsTue 8:00 — 3 commitsTue 9:00 — 0 commitsTue 10:00 — 1 commitsTue 11:00 — 2 commitsTue 12:00 — 2 commitsTue 13:00 — 6 commitsTue 14:00 — 0 commitsTue 15:00 — 1 commitsTue 16:00 — 4 commitsTue 17:00 — 1 commitsTue 18:00 — 2 commitsTue 19:00 — 2 commitsTue 20:00 — 0 commitsTue 21:00 — 1 commitsTue 22:00 — 2 commitsTue 23:00 — 1 commitsWed 0:00 — 1 commitsWed 1:00 — 0 commitsWed 2:00 — 0 commitsWed 3:00 — 0 commitsWed 4:00 — 0 commitsWed 5:00 — 0 commitsWed 6:00 — 2 commitsWed 7:00 — 0 commitsWed 8:00 — 0 commitsWed 9:00 — 4 commitsWed 10:00 — 5 commitsWed 11:00 — 3 commitsWed 12:00 — 4 commitsWed 13:00 — 1 commitsWed 14:00 — 3 commitsWed 15:00 — 1 commitsWed 16:00 — 4 commitsWed 17:00 — 3 commitsWed 18:00 — 2 commitsWed 19:00 — 1 commitsWed 20:00 — 0 commitsWed 21:00 — 0 commitsWed 22:00 — 1 commitsWed 23:00 — 0 commitsThu 0:00 — 0 commitsThu 1:00 — 1 commitsThu 2:00 — 0 commitsThu 3:00 — 0 commitsThu 4:00 — 0 commitsThu 5:00 — 0 commitsThu 6:00 — 0 commitsThu 7:00 — 0 commitsThu 8:00 — 0 commitsThu 9:00 — 0 commitsThu 10:00 — 1 commitsThu 11:00 — 3 commitsThu 12:00 — 3 commitsThu 13:00 — 0 commitsThu 14:00 — 5 commitsThu 15:00 — 2 commitsThu 16:00 — 2 commitsThu 17:00 — 2 commitsThu 18:00 — 4 commitsThu 19:00 — 1 commitsThu 20:00 — 1 commitsThu 21:00 — 0 commitsThu 22:00 — 0 commitsThu 23:00 — 0 commitsFri 0:00 — 1 commitsFri 1:00 — 0 commitsFri 2:00 — 1 commitsFri 3:00 — 1 commitsFri 4:00 — 1 commitsFri 5:00 — 0 commitsFri 6:00 — 0 commitsFri 7:00 — 1 commitsFri 8:00 — 1 commitsFri 9:00 — 0 commitsFri 10:00 — 1 commitsFri 11:00 — 1 commitsFri 12:00 — 2 commitsFri 13:00 — 5 commitsFri 14:00 — 6 commitsFri 15:00 — 4 commitsFri 16:00 — 3 commitsFri 17:00 — 0 commitsFri 18:00 — 1 commitsFri 19:00 — 0 commitsFri 20:00 — 2 commitsFri 21:00 — 1 commitsFri 22:00 — 0 commitsFri 23:00 — 1 commitsSat 0:00 — 1 commitsSat 1:00 — 0 commitsSat 2:00 — 1 commitsSat 3:00 — 0 commitsSat 4:00 — 0 commitsSat 5:00 — 0 commitsSat 6:00 — 2 commitsSat 7:00 — 1 commitsSat 8:00 — 9 commitsSat 9:00 — 5 commitsSat 10:00 — 2 commitsSat 11:00 — 2 commitsSat 12:00 — 0 commitsSat 13:00 — 5 commitsSat 14:00 — 1 commitsSat 15:00 — 1 commitsSat 16:00 — 3 commitsSat 17:00 — 3 commitsSat 18:00 — 7 commitsSat 19:00 — 3 commitsSat 20:00 — 0 commitsSat 21:00 — 1 commitsSat 22:00 — 2 commitsSat 23:00 — 0 commits
Commit volume by weekday and hour (UTC). Larger dots mean more commits.

Who is committing

last 52 weeks
Maintainer commits92 (45%)
Community commits111 (55%)

203 commits in total over the last year.

DateListRankStars gained
Nov 4, 2026daily#4+699
Nov 3, 2026daily#4+699
Nov 2, 2026daily#10+404
Sep 7, 2026weekly#15+2,010
Sep 6, 2026weekly#15+2,010
Sep 5, 2026weekly#10+2,065
Sep 4, 2026weekly#10+2,146
Sep 3, 2026weekly#11+2,045
Sep 2, 2026weekly#10+1,874
Sep 1, 2026daily#11+536
Sep 1, 2026weekly#10+1,439
Aug 31, 2026daily#11+536
Aug 30, 2026daily#4+150
Mar 15, 2026daily#12+292
Mar 14, 2026daily#19+212