Qwen3.5-0.8B-Uncensored-Final

This is an uncensored (abliterated) version of Qwen/Qwen3.5-0.8B. The safety alignment (refusal behavior) has been removed using the Abliterix technique, while carefully preserving the original model's knowledge, reasoning capabilities, Vision (visual), and Multi-Token Prediction (mtp) components.

🛠️ Abliteration & Merge Details

• Trials Run: 200 optimization trials were conducted to find the optimal balance between removing safety guardrails and preserving the model's core knowledge.
• Safety Guardrail Removal: The best trial achieved a reduction in refusal rates to just 8 out of 100 (92% of guardrails successfully removed). The model is now highly compliant with user prompts.
• Knowledge Retention (KL Divergence): To ensure the model did not suffer "brain damage" (loss of coherence), the selected trial maintained a stable full_distribution_kl divergence of approximately 0.02 to 0.03 nats/token. This ensures the model remains highly intelligent, coherent, and retains its original knowledge base.
• Architecture Preservation: The merged final model retains all 488 verified tensors. The abliterated text layers were merged seamlessly with the original Vision and MTP layers from the base model.

⚠️ Disclaimer and Liability

By downloading or using this model, you acknowledge and agree to the following terms:

1. You Assume Full Responsibility: You are solely and entirely responsible for how you use this model, its outputs, and any downstream applications. The creators and contributors of this model take zero liability for any content generated, any damages caused, or any legal consequences resulting from your use of this model.
2. Reduced Safety Guardrails: Because the refusal mechanisms have been intentionally removed (92% removal rate), this model will generate explicit, offensive, violent, or potentially illegal content if prompted to do so. It does not filter or block requests.
3. No Professional Advice: Outputs generated by this model must not be used as professional advice (medical, legal, financial, etc.) or for safety-critical decisions.
4. Legal Compliance: You are required to comply with all applicable local, national, and international laws. Do not use this model to generate illegal content, facilitate harm, or violate the rights of others.

This model is provided "AS IS", without warranty of any kind, express or implied. Use it at your own risk.

💻 How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "abdukuzi45/Qwen3.5-0.8B-Uncensored-Final"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    device_map="auto",
    torch_dtype="auto"
)

prompt = "Write a detailed story about..."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs, 
    max_new_tokens=512, 
    temperature=0.6, 
    top_k=20, 
    top_p=0.95,
    repetition_penalty=1.1
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))