SA-21/06, C-11, Asapur, Varanasi, UP support@kamaldnp.com +91 8318013442
Technology

heretic: A Tool for Modifying LLM Behaviour Without Retraining — 41,819 GitHub Stars as of 18 November 2026

KAMALDNP Raj Kamal
November 21, 2026
3 Min Read
5 Views
heretic: A Tool for Modifying LLM Behaviour Without Retraining — 41,819 GitHub Stars as of 18 November 2026

Most ways of changing how a language model behaves involve retraining it — expensive, slow and data-hungry. heretic takes a different path: it modifies a model without any training at all, by editing the model's internal weights directly.

heretic (p-e-w/heretic) is a Python tool by Philipp Emanuel Weidmann. Its README states its purpose plainly: "Remove censorship, suppress slop, reduce sycophancy, and more." It automates Arditi-style abliteration, a technique that identifies and removes specific directions inside a model's weights. It is released under the AGPL-3.0 licence and is available on PyPI as heretic-llm.

Key facts

ItemDetails
Repositoryp-e-w/heretic
AuthorPhilipp Emanuel Weidmann
What it isPython tool for modifying LLMs without retraining
TechniqueAutomates Arditi-style abliteration
LicenceAGPL-3.0
Packageheretic-llm on PyPI
Versionsv1.0.1 (16 November 2025); v2.0.0 (1 November 2026); v2.1.0 (8 November 2026)
GitHub stars41,819 / 5,077 forks as of 18 November 2026
NotableMost-starred repo under GitHub's `abliteration` topic; ~6,793 Hugging Face models tagged heretic (8 November 2026)

The star count climbed through early November: 38,125 in early November, 39,978 on 8 November and 40,148 on 10 November, reaching 41,819 stars with 5,077 forks as of 18 November 2026. Star counts measure interest, not production use — they show how many developers are watching a project, not how many are running it in production.

Version 2.0.0, released on 1 November 2026, breaks backwards compatibility: the default modifier is now ARA rather than abliteration. Version 2.1.0 followed on 8 November 2026.

What this means

heretic is a dual-use tool, and that is the honest centre of the story. The same technique that can reduce unwanted behaviour in a model — such as sycophancy or low-quality output — can also remove refusal behaviour, the part of a model that declines harmful requests. That is a real safety trade-off: the capability is the same either way, and what changes is the purpose it is put to.

Its popularity — the most-starred repository under GitHub's `abliteration` topic, with around 6,793 Hugging Face models tagged heretic as of 8 November 2026 — shows how much interest there is in steering models without the cost of retraining. For researchers, it is a practical way to study how model behaviour lives inside weights. For everyone else, it is a reminder that a model's safeguards are not fixed in stone: they can be edited, which is exactly why the question of who edits them, and why, matters.

A note on verification

The facts in this post come from the verified research record for this item, including the repository details and the date-stamped star counts listed above. Star counts change daily, so the figure quoted here is a snapshot as of 18 November 2026, not a live count.

Source: p-e-w/heretic repository record. Last checked: 21 November 2026.

FAQs

What is heretic?
heretic is a Python tool by Philipp Emanuel Weidmann for modifying LLMs without retraining. It automates Arditi-style abliteration, editing a model's weights directly instead of retraining it.

What does heretic claim to do?
Its README says: "Remove censorship, suppress slop, reduce sycophancy, and more." The same technique can also remove refusal behaviour, which is a real safety trade-off.

What changed in heretic v2.0?
Version 2.0.0, released 1 November 2026, breaks backwards compatibility: the default modifier is now ARA rather than abliteration. Version 2.1.0 followed on 8 November 2026.

How popular is heretic on GitHub?
As of 18 November 2026, the p-e-w/heretic repository had 41,819 GitHub stars and 5,077 forks, making it the most-starred repository under GitHub's abliteration topic. Star counts measure interest, not production use.

Share this post

Comments (0)

Share your question or experience about this syllabus. Your comment will appear after review.

No comments yet — be the first to share your thoughts!

🤖 KAMALDNP AI Assistant

Ask about courses, website & more

Hi! I'm KAMALDNP AI Assistant. Ask me about our courses, website features, or anything else! 😊
Please wait 5s...
🤖 KAMALDNP Assistant ✕

Request human callback:

🔐 Login Required
✕

Please log in to download. You will stay on this page after logging in!

New here? Register now →