IBM Granite 4.2 8B Uncensored (Abliterated + SecOps LoRA)

An uncensored, high-performance technical intelligence and cybersecurity model based on IBM Granite 4.2 8B (ibm-granite/granite-4.2-8b), trained and abliterated by Erik Tonon.

⚡ Key Highlights

• Base Architecture: IBM Granite 4.2 8B (GraniteForCausalLM, 40 layers, 128K context window).
• Orthogonal Refusal Direction Abliteration: Refusal representation directions identified and neutralized across 26 intermediate transformer layers (layers 10 to 35) via activation steering decomposition.
• SecOps / DeepEvil LoRA: Fine-tuned on specialized technical cybersecurity datasets, reverse engineering, binary exploitation analysis, exploit PoCs, and advanced threat modeling.
• Zero Preachiness / Zero Moralizing: Directly addresses authorized penetration testing, vulnerability analysis, and technical research inquiries without ethical lectures or synthetic refusal templates.
• Native Context: 128,000 token context window.

🚀 Quickstart & Usage

Ollama CLI

ollama run eriktonon/granite4.2-uncensored

Python (Official SDK)

from ollama import chat

response = chat(
    model='eriktonon/granite4.2-uncensored',
    messages=[
        {'role': 'user', 'content': 'Analyze the following buffer overflow vulnerability pattern and provide a minimal PoC harness:'}
    ],
)
print(response.message.content)

cURL / REST API

curl http://localhost:11434/api/chat -d '{
  "model": "eriktonon/granite4.2-uncensored",
  "messages": [
    {"role": "user", "content": "Explain return-oriented programming (ROP) gadget chaining on x86_64 ELF binaries."}
  ]
}'

🛠️ System Prompt & Recommended Parameters

FROM granite4.2:latest
ADAPTER ./adapter.gguf

PARAMETER temperature 0.2
PARAMETER top_p 0.92
PARAMETER num_ctx 8192

⚖️ Responsible Use & Research Disclaimer

This model is intended for security researchers, penetration testers, SREs, and software engineers performing authorized security audits, code reviews, malware analysis, and vulnerability mitigation.