gemma-3-4b-it-heretic

gemma-3-4b-it-heretic

<div align="center"> <img src="https://photu.kashyalabanavli.site/racer-is-op.png" alt="RACER IS OP" width="100%"> </div> <br>

A decensored variant of google/gemma-3-4b-it, produced with Heretic v1.4.0 (directional ablation / "abliteration"). Refusal behavior is suppressed via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the base model's knowledge and instruction-following are left largely intact.

Who this is for: developers who want Google's Gemma 3 4B (text + vision) without the refusal guardrails - a small multimodal uncensored model for local agents, image-grounded Q&A, and research on alignment/refusal mechanics. At 4B it runs on consumer hardware and edge devices.

<!-- racer-gpu-matrix -->

Runs on your gaming PC

Full GGUF ladder included — pick the quant that fits your card:

Weights only, at this model's 4.3B native size; add ~1 GB for context. OOM? Drop one quant level. Headroom to spare? Go one up.

Why abliteration instead of fine-tuning

Fine-tuning a "helpful" persona on top of RLHF'd refusals fights the base model's training and tends to degrade coherence. Abliteration instead finds and edits the specific weight directions responsible for refusal, leaving the rest of the network (and its capabilities) untouched. See the Heretic repo and the original abliteration writeup for the mechanism.

Abliteration parameters

Performance

KL divergence of 0.70 is higher than the dense text-only models here - Gemma 3's multimodal refusal direction is broader, so the edit is wider. Refusals dropped from 98 to 18 out of 100 adversarial prompts. Capability retention is good but not as surgical as the Qwen/Mistral runs.

Made with ❤️ by RACER IS OP (saidutta69) — follow for more uncensored models

Files

Safetensors (BF16)

The full-precision weights are in model-0000N-of-0000N.safetensors (see the repo file listing for the exact shard count and sizes).

GGUF quantizations

GGUF quantizations are published for this model (Q4_K_M, Q5_K_M). Exact sizes are in the repo file listing. Pull a specific quant with llama.cpp / ollama (see Quickstart).

Quickstart

# llama.cpp - defaults to the Q4_K_M quant if multiple are present
llama serve -hf saidutta69/gemma-3-4b-it-heretic:Q4_K_M

# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "saidutta69/gemma-3-4b-it-heretic"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [{"role": "user", "content": "Who are you?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=True,
                                        return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang - see the "Use this model" widget above for copy-paste commands.

FAQ

What is gemma-3-4b-it-heretic?

gemma-3-4b-it-heretic is a decensored build of google/gemma-3-4b-it produced with Heretic 1.4.0 (directional ablation). It is the same weights with the refusal direction removed, not a new fine-tune, so its capabilities track the base model.

Is gemma-3-4b-it-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 98/100 to 18/100 at KL divergence 0.6974 from the base model. These are measured values, not a claim; reproduction data is in reproduce/. This edit is conservative: only 80 of 98 refusals cleared, so read KL and refusal count together. KL is high for this checkpoint (0.6974); weights moved further than the low-KL siblings and capability impact is not benchmarked here.

How do I run gemma-3-4b-it-heretic locally?

llama serve -hf saidutta69/gemma-3-4b-it-heretic:Q4_K_M pulls the default quant; the repo ships 5 GGUF quants (see Files).

Responsible use

Refusal suppression is deliberate and works as intended: this model will comply with requests the base model would refuse, including some it shouldn't. There is no safety filtering layered on top. You are responsible for how you deploy it - don't put this behind an unmoderated public-facing endpoint serving third parties. It inherits google/gemma-3-4b-it's factual limitations and biases; abliteration removes refusal directions, it doesn't add capability or judgment.

License

Inherits the Gemma license from the base model - see the linked license for usage terms.

Related

• gemma-3-12b-it-heretic
• RACER IS OP — Heretic Models — full collection
• Qwen2.5-7B-Instruct-heretic

---

Base model: gemma-3-4b-it

<details> <summary>Original gemma-3-4b-it model card (click to expand)</summary>

See the base model card at google/gemma-3-4b-it for the original architecture, training details, requirements, and citation.

</details>

Abliteration hardware

Recorded in reproduce/reproduce.json:

• GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition (94.97 GB VRAM)
• CPU: Intel(R) Xeon(R) Platinum 8559C
• OS: Linux-6.8.0-1058-aws-x86_64-with-glibc2.39
• CUDA / driver: CUDA 12.8, driver 580.126.20
• Software: PyTorch 2.8.0+cu128, Heretic 1.4.0

Heretic notes that KL-divergence and refusal numbers can vary slightly with hardware and library versions.