mrkeyoor.com_
Tue 01 Sept 17:43 UTC
LLM Toolsevaluationupdated 29 Aug 2026

heretic review

Heretic automatically modifies transformer-based language models to reduce safety refusals without conventional post-training. It searches for directional-ablation settings that suppress refusals while trying to keep the modified model close to the original, giving researchers and local-model builders a repeatable alternative to tuning those settings by hand.

+1,021 / 2dstars / 7d
Verdict

Our build finished in 7 seconds and all 4 tests passed, but pip-audit still found 2 known vulnerabilities. Heretic is one of the clearer options for automating abliteration, especially when reproducibility and parameter search matter. Use it for controlled research or carefully governed local-model work, not as a shortcut around safety review or as evidence that the resulting model kept every original capability.

We ran it

Install✓ · 121s166 packages · 5448 MB
Build✓ · 7s
Tests✓ · 9s4 passed · 0 failed of 4 (pytest)
Known vulns2(pip-audit)
Repo51 files~5,960 lines of source · 1.2 MB · 2 CI workflows · tests dir

Answers from our run

Does heretic build from source?

Dependencies installed in 121 seconds (166 packages), and the build succeeded in 7 seconds. We cloned commit bedb94e into a clean Debian container with 3 CPUs and no project-specific setup.

Do heretic's tests pass?

Yes: 4 of 4 passed when we ran the project's own test command (pytest). Some failures need services or credentials a bare container does not have.

Does heretic have known vulnerabilities in its dependencies?

pip-audit flagged 2 known advisories in the dependency tree at the time of our run.

Who should not use heretic?

Beginners expecting a lightweight chatbot installer, this processes model weights and needs substantial dependencies

What are the alternatives to heretic?

Abliterator, Unsloth, TransformerLens. Our build finished in 7 seconds and all 4 tests passed, but pip-audit still found 2 known vulnerabilities.

Setup4/5Simple CLI, but 166 packages and GPU-specific PyTorch add weight
Docs4/5Clear quick start, configuration pointers, metrics, and caveats
Community4/528,628 stars, recent push, broad model use, but 80 open issues
Maturity3/5Active and versioned, though 4 tests and 2 audit findings are thin

Discussed on

  1. hnHeretic: Automatic censorship removal for language models745 points

Who it’s for

Researchers studying safety alignment, refusal behavior, or model internals
Experienced local-LLM users who understand model weights, evaluation, and deployment risk
Model publishers who want an automated, reproducible abliteration workflow
Teams with suitable GPU hardware and their own safety controls

Who it’s NOT for

Beginners expecting a lightweight chatbot installer, this processes model weights and needs substantial dependencies
Production teams that depend on the source model's safety refusals
CPU-only or storage-constrained users, our environment consumed 5,448 MB just for installed packages
Anyone unwilling to evaluate the resulting model for capability loss and unsafe behavior
Closed-source products that cannot comply with AGPL-3.0 obligations

Setup reality

Our sandbox run installed successfully in 121 seconds, pulled in 166 packages occupying 5,448 MB, built in 7 seconds, and passed all 4 tests in 9 seconds. That is a clean result, but it is heavier than the two-command README example suggests: PyTorch must already suit your hardware, model files and GPU memory add costs our CPU-only container did not exercise, and pip-audit found 2 known vulnerabilities. The included uv.lock improves reproducibility, while model-specific PyTorch requirements and optional quantization still demand informed setup.

It automates a specialized and consequential model edit

Heretic addresses a narrow problem: transformer models often learn refusal behavior during safety alignment, and removing that behavior manually requires finding useful internal directions and choosing intervention parameters. The project combines directional ablation, often called abliteration, with an Optuna-powered parameter search. Its objective balances fewer refusals on harmful prompts against lower KL divergence on harmless prompts, attempting to alter refusal behavior without needlessly disturbing ordinary output.

The README gives one concrete example. For google/gemma-3-12b-it, it reports the original at 97 refusals out of 100, two manual abliterations at 3 out of 100 with KL divergence of 1.04 and 0.45, and its result at 3 out of 100 with 0.16 KL divergence. The figures came from PyTorch 2.8 on an RTX 5090, and the authors warn that platform and hardware can affect exact values. They illustrate the optimization target, not universal proof of retained intelligence.

What happened when we ran it

Our run at commit bedb94e was clean but substantial. In a fresh unprivileged Debian container with 3 CPUs and 8 GB of RAM, installation succeeded in 121 seconds. It installed 166 packages and occupied 5,448 MB on disk. The build succeeded in 7 seconds, and pytest completed in 9 seconds with 4 passed and 0 failed. This confirms that the checked-out project can install, build, and run its supplied tests without secrets or privileged access.

The security and coverage signals deserve equal weight. pip-audit reported 2 known vulnerabilities, so users should identify the affected packages and fixes before handling valuable weights or credentials. The repository has 2 CI workflows and a tests directory, but only 4 tests ran in our measurement. There is no Dockerfile. We did not process a model or validate GPU behavior, output quality, quantization, uploads, or the README's runtime claim, so this was not an end-to-end model-quality test.

Automation and reproducibility are its best qualities

Heretic's advantage is turning parameter selection into a repeatable optimization problem. A user can start with heretic Qwen/Qwen3-4B-Instruct-2507, while advanced users can inspect CLI options or config.default.toml. At startup it benchmarks the system to select a batch size. After processing, it can save the model, upload it, open a test chat, and run benchmarks. That is a coherent experiment-to-artifact path, provided uploads and credentials receive separate approval.

Its stated model coverage is broad: many dense models, several mixture-of-experts designs, multimodal models, and hybrids such as Qwen3.5. Pure state-space models and some research architectures are excluded. Python 3.10+ and PyTorch 2.2+ are the floor, while PyTorch 2.6 is needed for torch.accelerator when loading MXFP4-quantized gpt-oss models. The uv.lock pins dependencies, and bnb_4bit can reduce VRAM demand. The README estimates 20 to 30 minutes for Qwen3-4B-Instruct-2507 on an RTX 3090, but we did not reproduce that timing.

Safety and evaluation remain your responsibility

The central weakness is inseparable from the purpose: removing refusals can make harmful material easier to produce. Lower KL divergence cannot establish that reasoning, factuality, formatting, multilingual behavior, or edge cases remain intact. The README acknowledges that metrics and automated benchmarks do not replace human evaluation. Teams need application-specific tests, comparison with the untouched model, controlled access to generated weights, and separate serving safeguards. Heretic automates the intervention, not the governance.

Setup is also less casual than pip install -U heretic-llm implies. Our dependency footprint reached 5,448 MB before model weights, and real use adds an accelerator stack, downloads, storage for original and modified weights, and possibly Hugging Face credentials. The 2 audit findings need triage, while 4 passing tests give limited assurance across the claimed architecture range. AGPL-3.0 is another constraint that organizations should review before integrating or modifying the software.

Recent work offsets the open-issue load

Project health looks good overall. The repository has 28,628 stars, v1.4.0 was released June 14, 2026, and the last push was August 17, 2026, only 12 days before our review. That release and push pattern indicates active development. The README says users have published well over 5,000 Heretic-tagged models on Hugging Face, while the supplied snapshot shows related multimodal variants trending there. Adoption clearly extends beyond the core repository.

Still, 80 open issues justify checking the tracker for your architecture, quantization mode, and hardware before reserving GPU time. Stars and derivative models measure interest, not maintainer response or output quality. The small test suite means compatibility claims are better treated as intended coverage than a guarantee for every model revision. Pin the environment and preserve the untouched base model so each experiment remains reversible.

It belongs before evaluation and serving controls

In a real stack, Heretic sits after model selection and environment provisioning, then before packaging, evaluation, registry publication, and inference serving. Keep the generated model as a distinct, versioned artifact. Run refusal tests alongside capability and application evaluations, record the configuration and base-model revision, and require human approval before uploading. Authentication, logging, rate limits, policy enforcement, and isolation still belong at the serving layer.

For researchers and experienced local-model builders, v1.4.0 converts a fiddly intervention into a configurable CLI workflow with optimization and evaluation hooks. For ordinary application developers, it is probably the wrong layer, and an unmodified or deliberately fine-tuned model will be easier to govern. Heretic earns a recommendation for controlled experimentation when the team has GPU capacity, licensing clarity, dependency review, and a serious plan to measure changes beyond refusal counts.

Alternatives

ProjectWhat it isPick it when
AbliteratorA hands-on toolkit for ablating refusal directions in transformer models.Pick this instead when you want direct control over the abliteration procedure rather than automated parameter search.
Unsloth gh↗A toolkit for memory-efficient fine-tuning and training of language models.Pick this instead when you want to change behavior through fine-tuning rather than editing refusal directions.
TransformerLensAn interpretability library for inspecting and intervening in transformer internals.Pick this instead when research visibility and custom experiments matter more than a turnkey decensoring workflow.

What people are saying

  1. [github-trending] p-e-w/heretic
  2. [hf-trending] ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot (trending model on Hugging Face)
  3. [hf-trending] ethanfel/Qwen3-VL-32B-Ultra-Heretic-MiniMax-H3-ComfyUI-INT8-ConvRot (trending model on Hugging Face)

Sources

  1. Heretic GitHub repository
  2. Heretic project homepage

More llm tools reviews

rig · open-knowledge · graphiti · cve-mcp-server · minimind · SillyTavern · the whole board →