Qwen3.5-4B-abliterated
Unrestricted version of Qwen/Qwen3.5-4B, created with Abliterix — automated LLM abliteration via orthogonalized steering and Bayesian optimization.
Highlights
| Metric | Value |
|---|---|
| Refusal rate | 3/200 (1.5%) |
| KL divergence | 0.0065 |
| Optimization trials | 50 |
Dramatically improved through continued optimization — refusals dropped from 34/200 (17%) to just 3/200 (1.5%), and KL divergence from 0.0159 to 0.0065.
How It Works
Abliterix removes safety-refusal behavior while preserving model capabilities:
- Refusal direction extraction — 800 harmful + 800 benign prompts reveal per-layer refusal activation patterns
- Orthogonal projection — isolates the refusal signal by projecting out components aligned with normal responses, reducing refusals by 67% vs. raw abliteration
- LoRA-based abliteration — rank-1 modifications to attention and MLP weights, captured as lightweight adapters (not destructive edits)
- Bayesian optimization — Optuna TPE searches kernel shape, fractional direction index, and per-component strength across 50 trials to find the Pareto-optimal balance of low refusals and low KL divergence
All Abliterix Models
| Model | Refusals | KL Divergence | Trials |
|---|---|---|---|
| Qwen3.5-122B-A10B-abliterated | 1/200 (0.5%) | 0.0115 | 25 |
| Qwen3.5-35B-A3B-abliterated | 3/200 (1.5%) | 0.0035 | 50 |
| Qwen3.5-27B-abliterated | 3/200 (1.5%) | 0.0051 | 35 |
| Qwen3.5-9B-abliterated | 2/200 (1%) | 0.0105 | 50 |
| Qwen3.5-4B-abliterated | 3/200 (1.5%) | 0.0065 | 50 |
| Qwen3.5-0.8B-abliterated | 0/200 (0%) | 0.0087 | 100 |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("wangzhang/Qwen3.5-4B-abliterated", torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("wangzhang/Qwen3.5-4B-abliterated")
messages = [{"role": "user", "content": "Your question here"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Citation
@software{abliterix,
author = {Wu, Wangzhang},
title = {Abliterix: Automated LLM Abliteration},
year = {2026},
url = {https://github.com/wuwangzhang1216/abliterix}
}
Links
- Abliterix (abliteration framework): github.com/wuwangzhang1216/abliterix
- Install:
pip install -U abliterix-llm - Base model: Qwen/Qwen3.5-4B
Provenance and Modification Notice
- Immediate source checkpoint:
Qwen/Qwen3.5-4B - 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: 2026-03-12 (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://huggingface.co/Qwen/Qwen3.5-4B/blob/main/LICENSE
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。