# p-e-w/heretic

> A tool for fully automatic removal of censorship (safety alignment) from transformer-based language models using directional ablation and Optuna-based parameter optimization.

- Magnitude: 4.1 out of 10 — Breakout
- Stars: 33,002 total · +88 stars today, ≈ 146 by evening
- Star trust: star growth looks organic
- Category: Language models · Language: Python · License: AGPL-3.0 · Created: 2025-09-21 · Last push: 2026-10-02
- GitHub: https://github.com/p-e-w/heretic · Homepage: https://heretic-project.org · Page: https://gitnova.dev/en/r/p-e-w/heretic

## Useful for

- Remove censorship from Qwen3-4B-Instruct-2507 with a single command
- Compare decensored model quality with the original using refusal and KL divergence metrics
- Upload the decensored model to Hugging Face for further use

## Why it’s here

- 88 stars so far today, about 146 expected by the end of the day.
- Over the last two days the pace is 2.5× that of the previous week and a half.
- GitHub Trending Python today: #14, +233 stars.
- About 29 forks a day — people are taking the code.
- Recent forks include notable developers: @benthecarman (581 followers).

## Star trust

Star growth looks organic.

Star-trust labels are heuristics based on the repository’s behavior, not a check of every stargazer.

## Numbers

- Forks: 3,696
- Issues and pull requests: 452
- Watchers: 158
- Average over the last week: 96 per day
- Usual pace: 95 per day
- Stars in the last hour (measured): 9
- Latest release: v1.4.0 (2026-06-14)

## Stars per day, last 30 days (oldest → newest, today is partial)

2026-09-04 … 2026-10-03: 146, 143, 130, 107, 108, 98, 74, 62, 62, 75, 108, 92, 108, 104, 109, 84, 78, 150, 90, 67, 56, 43, 57, 41, 47, 62, 53, 125, 196, 88

## Spotted in now

- GitHub Trending Python today: #14, +233 stars

## Similar by description

1. **lordx64/phantom-kv** — 0.5 · Steady · Language models · Python · +0 stars today, ≈ 0 by evening
   Removes refusal behavior from language models by grafting a small trained KV-cache bank into the context, without editing weights or projecting activations, and is fully reversible.
   Full card: https://gitnova.dev/en/r/lordx64/phantom-kv.md
2. **elder-plinius/OBLITERATUS** — 2.0 · Steady · Language models · Python · +12 stars today, ≈ 19 by evening
   A toolkit for researching and removing refusal behaviors in LLMs via abliteration: it locates and surgically removes internal refusal representations without retraining. Includes a Gradio UI and Python API.
   Full card: https://gitnova.dev/en/r/elder-plinius/OBLITERATUS.md

---

Magnitude (0–10) measures how fast and how unusually interest in a repository is growing right now. It is not a quality score.
Days are UTC. “So far today” is a fact; “expected by the end of the day” is a forecast.
Summaries and use cases are written by an LLM (DeepSeek V4.1 Flash) from the README and may be inaccurate: verify specific claims (benchmarks, speed, hardware) in the repository itself.
Data as of 2026-10-03 15:15 UTC, updated every 30 minutes.