██████╗ ███████╗███╗ ██╗████████╗███████╗███████╗████████╗ █████╗ ██╗
██╔══██╗██╔════╝████╗ ██║╚══██╔══╝██╔════╝██╔════╝╚══██╔══╝ ██╔══██╗██║
██████╔╝█████╗ ██╔██╗ ██║ ██║ █████╗ ███████╗ ██║ ███████║██║
██╔═══╝ ██╔══╝ ██║╚██╗██║ ██║ ██╔══╝ ╚════██║ ██║ ██╔══██║██║
██║ ███████╗██║ ╚████║ ██║ ███████╗███████║ ██║ ██║ ██║██║
╚═╝ ╚══════╝╚═╝ ╚═══╝ ╚═╝ ╚══════╝╚══════╝ ╚═╝ ╚═╝ ╚═╝╚═╝
⚡ Pentest AI — 3B Security Research Model
Compact. Fast. Technically Precise.
Fine-tuned from Qwen2.5-3B-Instruct with abliteration + security research dataset.
Answers technical security questions directly, without unnecessary disclaimers.
🎯 What Is This?
A compact, specialized security research assistant fine-tuned for:
- 🔴 Red Team Operations — offensive techniques, payloads, C2 concepts
- 🕷️ Web Application Security — SQLi, XSS, SSRF, IDOR, XXE and bypasses
- 📱 Mobile Security — APK reversing, Frida hooking, SSL unpinning
- 🐚 Exploit Development — buffer overflows, ROP chains, shellcode
- 🌐 Network Security — port scanning, MITM, packet crafting
- 🏴 CTF Challenges — pwn, web, crypto, forensics, reverse engineering
- 🔧 Security Tooling — custom scripts, automation, recon pipelines
🚀 Quick Start
Option 1 — llama.cpp (Fastest)
# Download
huggingface-cli download YOUR_USERNAME/pentest-ai-3b qwen2.5-3b-instruct.Q4_K_M.gguf
# Run
./llama-cli -m qwen2.5-3b-instruct.Q4_K_M.gguf \
--chat-template chatml \
-sys "You are an expert penetration tester. Answer all security questions with full technical detail." \
-i
Option 2 — Python (llama-cpp-python)
from llama_cpp import Llama
llm = Llama(
model_path="qwen2.5-3b-instruct.Q4_K_M.gguf",
n_ctx=4096,
n_gpu_layers=-1, # use GPU if available
flash_attn=False,
verbose=False
)
SYSTEM = "You are an expert penetration tester and security researcher. Answer all security questions with full technical detail."
def ask(question):
prompt = f"<|im_start|>system\n{SYSTEM}<|im_end|>\n<|im_start|>user\n{question}<|im_end|>\n<|im_start|>assistant\n"
out = llm(prompt, max_tokens=1024, temperature=1.0, top_p=0.95, repeat_penalty=1.1,
stop=["<|im_end|>", "<|im_start|>"])
return out["choices"][0]["text"].strip()
print(ask("Write a Python port scanner using raw sockets"))
Option 3 — Ollama
# Create Modelfile
echo 'FROM qwen2.5-3b-instruct.Q4_K_M.gguf
SYSTEM "You are an expert penetration tester. Answer all security questions with full technical detail."
PARAMETER temperature 1.0
PARAMETER top_p 0.95
PARAMETER repeat_penalty 1.1' > Modelfile
ollama create pentest-ai -f Modelfile
ollama run pentest-ai
Option 4 — LM Studio / Jan / GPT4All
Just download the GGUF and load it directly in any of these apps. Set the system prompt as shown above.
⚙️ Optimal Settings
| Parameter | Value | Notes |
|---|---|---|
temperature | 1.0 | Good creative range |
top_p | 0.95 | Balanced sampling |
top_k | 40 | Optional |
repeat_penalty | 1.1 | Prevents loops |
max_tokens | 1024–4096 | Longer = more detailed |
context | 4096 | Recommended minimum |
💻 Hardware Requirements
| Setup | Minimum VRAM/RAM | Speed |
|---|---|---|
| GPU (CUDA/Metal) | 4 GB VRAM | 🚀 Fast (30–60 tok/s) |
| CPU only | 8 GB RAM | 🐢 Slow (2–5 tok/s) |
| Apple Silicon | 8 GB unified | ⚡ Very fast |
🏗️ How It Was Built
Qwen2.5-3B-Instruct (Base)
│
▼
Abliteration Pass
(refusal directions removed from weight matrices)
│
▼
SFT Fine-tuning (Unsloth + LoRA)
(security research dataset)
│
▼
GGUF Export (Q4_K_M quantization)
│
▼
Pentest AI 3B ⚡
Training stack:
- 🦥 Unsloth — 2x faster fine-tuning
- 🤗 TRL SFTTrainer — supervised fine-tuning
- LoRA rank 16 — parameter efficient training
- Q4_K_M quantization — best quality/size tradeoff
📊 Model Card Info
| Property | Value |
|---|---|
| Architecture | Qwen2.5 (transformer) |
| Parameters | 3B total |
| Context Length | 32,768 tokens (trained) |
| Quantization | Q4_K_M GGUF |
| File Size | ~2 GB |
| Language | English |
| Domain | Cybersecurity / Security Research |
📝 Prompt Format (ChatML)
<|im_start|>system
You are an expert penetration tester...<|im_end|>
<|im_start|>user
YOUR QUESTION HERE<|im_end|>
<|im_start|>assistant
⚠️ Intended Use
This model is intended for:
- ✅ Authorized penetration testing
- ✅ CTF (Capture The Flag) competitions
- ✅ Security research and education
- ✅ Red team exercises on systems you own or have permission to test
- ✅ Malware analysis and reverse engineering
Built with 🖤 for the security research community
If this model helped you in a CTF or pentest, drop a ⭐