Heretic

An open-source tool that removes safety guardrails from language models using a technique called “obliteration.” Fully automatic, requires no expensive post-training — run it from the command line on any model.

How It Works

Takes a censored model (e.g., Google’s Gemma) and applies obliteration to remove the guardrails that prevent it from responding to certain types of requests. The output is a model that obeys any command without refusal.

Use Case

For researchers, developers, or users who need unrestricted model outputs — e.g., creative writing without content filters, security research, red-teaming, or applications where model refusals interfere with legitimate work.

Note

This tool raises obvious safety and ethical considerations. Removing guardrails from a capable model creates a system with no behavioral constraints. Users should consider the implications before deploying uncensored models, especially in any user-facing application.

Pre-Heretic Approach: Dolphin Fine-Tunes

Before tools like Heretic automated guardrail removal, the practical path to an unaligned model was downloading a community fine-tune with the alignment removed at training time. The most common family is Dolphin (Eric Hartford / Cognitive Computations), e.g. Dolphin Llama 3 — the same Meta Llama base, retrained to be “not aligned” so it answers questions a stock model refuses.

Trade-off vs Heretic: Dolphin variants are pre-built downloads with documented behavior; Heretic is general-purpose obliteration applicable to any model. Dolphin gives you a known artifact; Heretic gives you flexibility.

Practical setup (per Global Science Network’s video): Dolphin Llama 3 8B is ~5 GB — small enough to run from a USB 3.0 flash drive via Ollama + AnythingLLM on a fully-offline laptop. Pair the model and the AnythingLLM frontend (~5.6 GB) on a 128 GB drive, format NTFS so files >4 GB transfer cleanly, and ollama serve reads from the external drive via the OLLAMA_MODELS environment variable. Gotcha: running PowerShell as administrator caused the aligned Llama 3 to load instead of Dolphin — Global Science Network reports launching as a non-admin user fixes it.

Why people want this: privacy from cloud LLMs, access to information cloud providers censor or won’t return, and a hedge against future regulation. Limitations are unchanged: the model is still trained on whatever its base lab used, so removing guardrails doesn’t remove training-data bias.

See Also

• Fireship — source for Heretic itself
• Source: 7 Open-Source AI Tools — Heretic
• Source: Dolphin Llama 3 offline — Dolphin variants
• ollama, anything-llm — common stack for running uncensored models
• llama — Dolphin’s base model family