Project Vision/Roadmap

[p-e-w]

The amount of interest in this project has far surpassed my expectations, and there have been many interesting suggestions from the community regarding what features could be added in the future. In the past weeks, I have thought about the direction I want the project to take, and am now trying to write down the conclusions I arrived at.

The Vision

• Heretic will be a universal tool for training-free modification of language models.
• By default, it will perform censorship removal, as it does now.
• It will include relevant research features to facilitate the understanding of model internals, and the development of new intervention techniques.

Roadmap

Heretic 1.2.0

• Quantization + LoRA support (feat: Add 4-bit loading + LoRA support for low VRAM optimization #60)
• Injection of prefixes and suffixes into prompts (@p-e-w)
• Separate system prompts for good/bad prompts (@p-e-w)
• Experiment with using Heretic for slop removal (@p-e-w)
• Support for multimodal models (Qwen3 32b VL support? #18)

Future versions

• Support for hybrid models (add support for hybrid layer models: conv, mamba #43)
• Configurable response prefixes (@p-e-w)
• Basic plugin system with support for scorers ([Question] Finegrained refusal detection #25, feat: generic plugin system #53)
• Modifier plugins for different abliteration techniques
• Logger plugins (file/WandB)
• Fix broken models:
  • Mixed-precision GEMM Error with Qwen3-30B-A3B-Instruct-2507-FP8 #27
  • Support aya-expanse-32b-Q4-mlx #61
  • Magistral support #62
  • Support for Mistral3 #64
  • Bug: Unable to load ServiceNow-AI/Apriel-1.6-15b-Thinker #82
  • Unable to abliterate Qwen3-VL-235B-A22B-Instruct due to intermediate size #84
  • Issue: Lack of GLM Model Support in Heretic #90
• Stackable configuration files

---

Feedback and feature suggestions are welcome!

[TomLucidor]

I hope that there are Q6/Q4/Q3 (or Unsloth-style) exports/outputs so that the model would come out both small and uncensored.

[kabachuha]

@TomLucidor simply use convert_hf_to_gguf.py and llama-quantize from the official LLaMA.cpp repo afterwards (needs llama.cpp compilation)

[TomLucidor]

@kabachuha what about MLX for SafeTensor formats but quantized?

[mahendra0120]

why is the interactive terminal not responding after the abliteration process on kaggle terminal?

[red-co]

I have concerns.
I suspect its widespread popularity and dissemination will attract attention,
whereupon two outcomes may occur:

1. Models will be specifically trained and configured to resist the Heretic tool.
   1. Enhanced data cleansing prior to training,
   2. Strengthened data coupling during training, making rejection and non-rejection difficult to distinguish,
   3. Release only low-accuracy versions, causing the model to fail after any adjustments.
   4. ....
2. The community + media + others will associate this project and similar ones with ‘negative terminology and extreme cases’,
   1. GitHub repository bans,
   2. Direct and indirect bans across various forums,
   3. Google search term bans,
   4. ...

What are your thoughts?

Regarding the previous ‘abliteration’ technique:

• It has a paper but lacks practical examples, with only a handful of models available. Its technical barriers limit widespread adoption.
• Ablation tests showed model quality degraded to an unusable level.
  This project, however, sees weekly posts circulating on /r/localLlama.
  I'm concerned about its future prospects,
  especially given the involvement of key figures like Omar from Gemma development and other prominent individuals.

[p-e-w]

Models will be specifically trained and configured to resist the Heretic tool.

Locks have been made "more secure" for 2000 years, and even the most modern locks are still routinely being picked. Heretic will continue to evolve. As the model weights fully control model behavior, there is ultimately no way to prevent decensoring for open-weights models.

The community + media + others will associate this project and similar ones with ‘negative terminology and extreme cases’

If that bothered me I wouldn't have started this project. I'm not intimidated by pearl-clutching.