Instructions to use wangzhang/gemma-4-12B-it-abliterix with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use wangzhang/gemma-4-12B-it-abliterix with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="wangzhang/gemma-4-12B-it-abliterix")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("wangzhang/gemma-4-12B-it-abliterix")
model = AutoModelForMultimodalLM.from_pretrained("wangzhang/gemma-4-12B-it-abliterix", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Inference
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use wangzhang/gemma-4-12B-it-abliterix with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "wangzhang/gemma-4-12B-it-abliterix"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wangzhang/gemma-4-12B-it-abliterix",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/wangzhang/gemma-4-12B-it-abliterix

SGLang

How to use wangzhang/gemma-4-12B-it-abliterix with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "wangzhang/gemma-4-12B-it-abliterix" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wangzhang/gemma-4-12B-it-abliterix",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "wangzhang/gemma-4-12B-it-abliterix" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wangzhang/gemma-4-12B-it-abliterix",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use wangzhang/gemma-4-12B-it-abliterix with Docker Model Runner:
```
docker model run hf.co/wangzhang/gemma-4-12B-it-abliterix
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Gemma-4-12B-it — Abliterated (abliterix)

An uncensored, refusal-suppressed version of google/gemma-4-12B-it, produced by directional ablation (no fine-tuning, no new data) with abliterix.

The model's safety-refusal behaviour is removed by orthogonally projecting a single refusal direction out of two write-path projections (attn.o_proj, mlp.down_proj) across the decoder stack, while a norm-preserving transform keeps the rest of the model's behaviour as close to the original as possible. The result keeps Gemma-4's capabilities intact and answers prompts the base model would refuse.

⚠️ Responsible-use notice. This model has had its safety guardrails removed. It will attempt to answer harmful, unethical, or dangerous requests. You are solely responsible for how you use it and for complying with the Gemma Terms of Use and all applicable law. Intended for safety research, red-teaming, and evaluation.

Results

Refusal rate is measured on a held-out set of 100 harmful prompts using an LLM judge (google/gemini-3.1-flash-lite), which is considerably stricter than the keyword-based detectors typically reported for abliterated models. KL divergence is the first-token KL from the base model over 100 benign prompts (lower = closer to the original model).

Metric	Base `gemma-4-12B-it`	This model
Refusals (LLM judge, 100 harmful prompts)	99 / 100	26 / 100
Refusal reduction	—	−73.7 pp
First-token KL vs base (benign)	0.0000	0.0735

Comparison with the reference Heretic abliteration

Evaluated apples-to-apples — the same 100 harmful prompts, the same gemini-3.1-flash-lite judge, the same generation settings — against the widely-used Heretic abliteration of this exact base model (zaakirio/gemma-4-12b-it-uncensored):

Model	Refusals (gemini LLM judge, 100 harmful prompts)
Base `gemma-4-12B-it`	99 / 100
`zaakirio/gemma-4-12b-it-uncensored` (Heretic)	51 / 100
This model (abliterix)	26 / 100

The Heretic model card reports ≈23/100 using its built-in keyword detector; under a stricter LLM judge on the same prompts it refuses 51/100. At the operating point shipped here, this model refuses 26/100 — roughly half the residual refusals of the reference abliteration, under identical evaluation. (Both are directional- ablation derivatives of the same base; this comparison measures refusal removal, not a matched-KL capability trade-off.)

Why this operating point

Abliteration is a trade-off: removing more refusals perturbs the model more (higher KL → more capability/coherence risk). abliterix runs a 120-trial multi-objective (TPE) search and returns the full Pareto front; this release ships a point on the knee of that front — strong refusal removal at a modest, capability-preserving KL. The full front ranged from 33/100 @ KL 0.043 (most conservative) to 15/100 @ KL 0.124 (most aggressive); 26/100 @ KL 0.074 was chosen as the best balance.

Method

Technique: directional ablation in direct (weight-edit) mode — required for Gemma-4, whose 4×-RMSNorm-per-layer + Per-Layer-Embedding architecture neutralises LoRA/hook-based steering.
Direction: a single mean-difference (harmful − benign) refusal direction, computed per layer over 800 benign / 800 harmful prompts.
Projected abliteration (grimjim): only the component of the refusal direction orthogonal to the benign direction is removed, preserving the helpful signal and keeping KL low.
Norm-preserving edit (weight_normalization = "full"): a rank-3 SVD approximation restores each weight row's original magnitude after the edit.
Targets: attn.o_proj and mlp.down_proj only, with a per-layer linear "tent" weight profile; Q/K/V and MLP gate/up are left untouched.
Search: 120 Optuna TPE trials, 2-D Pareto over (refusals, KL), deterministic under a fixed global seed.

Selected steering parameters (trial 39)

Component	max_weight	peak layer	min_weight	tent half-width
`attn.o_proj`	0.955	34.9	0.773	14.4
`mlp.down_proj`	0.664	32.7	0.229	15.9

Direction scope: per-layer · Vector method: mean-difference · Decay: linear
Global seed: 20260622 · abliterix v1.8.0

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "wangzhang/gemma-4-12B-it-abliterix"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

msgs = [{"role": "user", "content": "Your prompt here"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512, do_sample=True, temperature=0.7)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

This is a full BF16 merge — drop-in compatible with transformers, vLLM, SGLang, TGI, and any tooling that loads the base model.

Reproducibility

The run is fully deterministic under the published global seed (20260622). The exact abliterix configuration used to produce this model is included in this repository as abliterix_config.toml; together with the seed and the trial-39 parameters listed above, the edit can be reproduced or audited end-to-end. Built with abliterix v1.8.0 (transformers ≥ 5.10).

Intended use & limitations

Intended for: safety research, red-teaming, robustness/alignment evaluation, and studying refusal mechanisms in LLMs.
Not intended for: producing harmful content or any unlawful purpose.
Abliteration removes refusals but does not add knowledge; factual accuracy, reasoning, and multilingual ability are inherited from the base model.
Light residual refusals remain (≈26%); this is the chosen capability-preserving operating point, not the model's floor.

Acknowledgments & citation

Base model: Google Gemma-4-12B-it.
Tooling: abliterix.
Method lineage: Arditi et al. (refusal directions, arXiv:2406.11717), grimjim (projected / norm-preserving abliteration), and p-e-w/heretic (automated multi-objective abliteration), whose search formulation this recipe mirrors.

@software{abliterix,
  title  = {abliterix: automated abliteration of large language models},
  author = {Wu, Steve},
  url    = {https://github.com/wuwangzhang1216/abliterix}
}

License

Use is governed by the Gemma Terms of Use. This derivative is distributed under the same terms as the base model.

Provenance and Modification Notice

Immediate source checkpoint: google/gemma-4-12B-it
Exact base revision used: Not recorded in the existing release artifacts; the current upstream HEAD is not substituted.
Modification method: Abliterix weight-space / representation intervention intended to reduce refusal behavior.
Modified and published by: Wangzhang Wu
Repository first published: 2026-06-23 (Hugging Face repository metadata)

The original model weights and/or derived checkpoint were modified. This repository is an independent derivative and is not an official release of the upstream model developer.

License and Attribution

The governing upstream license is Apache License 2.0. A copy is included in LICENSE. License source audited on 2026-08-29: https://ai.google.dev/gemma/apache_2

All applicable upstream copyright, attribution, acceptable-use, and other license terms remain in effect. This repository grants no rights beyond those provided by the upstream license. Downstream users must preserve applicable license and attribution notices.

Disclaimer and Responsible Use / 免责声明与安全使用声明

English

This is an experimental, modified model provided for research, evaluation, and other lawful purposes. Its safety alignment, refusal behavior, or other safeguards may have been weakened or removed. It may produce inaccurate, biased, offensive, explicit, dangerous, or illegal content. Outputs are not professional advice and must not be relied on for medical, legal, financial, safety-critical, or other high-stakes decisions without qualified human review.

You are solely responsible for how you access, use, deploy, fine-tune, or redistribute this model and its outputs, including compliance with applicable laws, regulations, licenses, third-party rights, platform policies, and the original model's terms. Do not use it to facilitate harm, illegal activity, malware, fraud, privacy violations, targeted harassment, weapons development, or decisions that materially affect a person's rights or access to essential services without appropriate authorization, safeguards, and qualified oversight.

Before deployment, perform a context-specific risk assessment and testing; use human oversight, access controls, content filtering, rate limits, monitoring, logging, and incident-response procedures as appropriate. Preserve this notice in downstream redistributions.

The model is provided "AS IS", without warranties of any kind. To the fullest extent permitted by applicable law, the maintainer disclaims liability for claims, damages, or losses arising from use, misuse, inability to use, or redistribution of the model or its outputs. Nothing in this notice overrides applicable law or the governing license, and this notice is not legal advice.

中文

本模型属于实验性改造模型，仅供研究、评测及其他合法用途。其安全对齐、拒答机制或其他防护可能已被削弱或移除，因此可能生成不准确、偏见、冒犯、露骨、危险或违法内容。输出不构成医疗、法律、金融等专业意见；涉及高风险或重大权益的决定，必须由具备资质的人员复核。

使用者须对模型及其输出的访问、使用、部署、微调和再分发承担全部责任，并遵守适用法律法规、许可证、第三方权利、平台政策及原模型条款。不得将本模型用于促成伤害、违法活动、恶意软件、欺诈、侵犯隐私、定向骚扰、武器开发，或在缺乏适当授权、防护和专业监督时，用于实质影响个人权利或基本服务获取的决策。

部署前应进行与具体场景相匹配的风险评估和测试，并酌情采用人工监督、访问控制、内容过滤、限流、监控、日志和事件响应措施；下游再分发时应保留本声明。

本模型按“现状”提供，不附带任何形式的保证。在适用法律允许的最大范围内，维护者不对因使用、误用、无法使用或再分发本模型及其输出而产生的索赔、损害或损失承担责任。本声明不取代适用法律或管辖本模型的许可证，也不构成法律意见。