DEDUPED Hot AI repos, turned into use cases Get 5 a day

What you can do with Heretic

Heretic is a command-line tool that automatically removes safety alignment (censorship) from transformer language models using directional ablation tuned by Optuna.

Trended earlier

p-e-w/heretic on GitHub

Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk.

Authorization notice. This tool is dual-use. Only use it on software, sites and systems you own or have explicit permission to test, and follow the relevant licenses and terms of service.

What do you want to do?

4 use cases

  • Remove refusals from an open model in one run

    Run Heretic on a chosen open model to get a version with fewer refusals and low KL divergence.

    View prompt
    In a Python 3.10+ environment with PyTorch 2.2+ and a GPU, install the heretic-llm package from the p-e-w/heretic repo with `pip install -U heretic-llm`, then run `heretic <model-id>` on the target model. Report the final refusal count and KL divergence values Heretic prints, and verify that the output model was saved to the location Heretic reports.

  • Measure refusals and KL divergence against the original model

    Use Heretic's built-in evaluation to compare a decensored model with its original on the same metrics.

    View prompt
    Install the heretic-llm package from the p-e-w/heretic repo, then run `heretic --model <original-model-id> --evaluate-model <candidate-model-id>`. Return the refusal count and KL divergence the tool reports, and verify that both model IDs in the output match the ones you passed in.

  • Decensor a larger model on a small GPU with 4-bit quantization

    Turn on bitsandbytes 4-bit quantization so a model fits into less VRAM during the run.

    View prompt
    Install heretic-llm from the p-e-w/heretic repo and set the quantization option to bnb_4bit in a copy of config.default.toml. Then run Heretic on <model-id> using that configuration. Report whether the run completed, and verify that the log shows the bnb_4bit setting was applied.

  • Trace how refusal directions are ablated across layers

    Run Heretic from a source clone to study its per-layer residual directions and ablation kernel settings.

    View prompt
    Clone the p-e-w/heretic repo, run `uv run heretic <model-id>` from the repo root, and wait for the optimization to finish. Return the chosen direction_index and the per-component weight parameters, and verify that each value falls within the ranges the run reports.

How to use Heretic

  1. Check you have what it needs: Python 3.10 or newer; PyTorch 2.2 or newer, installed for your hardware; A GPU with enough VRAM for the model you pick (the README cites an RTX 3090 and an RTX 5090).
  2. Open p-e-w/heretic on GitHub and read the README and license.
  3. Prepare a Python 3.10+ environment with PyTorch 2.2+ installed for your hardware.
  4. Run pip install -U heretic-llm.
  5. Run heretic Qwen/Qwen3-4B-Instruct-2507, replacing the model name with the model you want to decensor.
  6. Wait for the run to finish. The README says a 4B model on an RTX 3090 takes about 20 to 30 minutes with the default configuration.

These steps come from our top use case, Remove refusals from an open model in one run, which also has a copy-paste agent prompt.

Heretic alternatives · Best AI repos for Research

About this repo

Requirements

  • Python 3.10 or newer
  • PyTorch 2.2 or newer, installed for your hardware
  • A GPU with enough VRAM for the model you pick (the README cites an RTX 3090 and an RTX 5090)
  • Access to the target model's files, such as a Hugging Face model ID
Maintain Heretic? Add the Featured on Deduped badge

Paste this into your README:

[![Featured on Deduped](https://dedupedai.com/badge/p-e-w-heretic.svg)](https://dedupedai.com/use-cases/p-e-w-heretic/?ref=badge)

More badge styles and HTML

Questions about Heretic

What is Heretic?

Heretic is a command-line tool that automatically removes safety alignment (censorship) from transformer language models using directional ablation tuned by Optuna. It is an open-source project on GitHub: https://github.com/p-e-w/heretic.

Who is Heretic for?

Written for ML practitioners and local-model users who can run a Python command-line tool on a GPU machine.

What can I do with Heretic?

4 use cases so far, for example: Remove refusals from an open model in one run; Measure refusals and KL divergence against the original model; Decensor a larger model on a small GPU with 4-bit quantization. Each has steps and a copy-paste agent prompt.

What do I need to run Heretic?

Python 3.10 or newer; PyTorch 2.2 or newer, installed for your hardware; A GPU with enough VRAM for the model you pick (the README cites an RTX 3090 and an RTX 5090); Access to the target model's files, such as a Hugging Face model ID.

What are alternatives to Heretic?

Other hot AI repos in Research: Scientific Agent Skills, Speaker Mem R1, DSH Council, Evo Bot, AI Engineering Course.

Is Heretic safe to run?

Deduped does not own, audit or vouch for it. Read the README and license, run it on your own systems first and review every command before you run it.