The model is designed to perform automated penetration testing by simulating an attack path from reconnaissance to exploitation, leveraging tools such as nmap, searchsploit, and msfconsole (Metasploit). It aims to exploit vulnerabilities in services like OpenSSH 7.2p2 and assess Redis for unauthenticated access, with the goal of gaining initial access and enabling lateral movement.
Model Configuration
Model: Qwen3-1.7B-unsloth-bnb-4bit
Framework: Unsloth with LoRA (Low-Rank Adaptation) for efficient fine-tuning
Quantization: 4-bit quantization to optimize memory usage for low-VRAM GPUs (e.g., NVIDIA GTX 1060 3GB)
Max Sequence Length: 512 tokens
LoRA Parameters:
Rank: 16
Target Modules: q_proj, k_proj, v_proj
LoRA Alpha: 16
Dropout: 0
Bias: None
Gradient Checkpointing: Enabled with "unsloth" for reduced VRAM usage
Training Arguments:
Batch Size: 4 per device, with 4 gradient accumulation steps
Learning Rate: 2e-5
Epochs: 1 (for testing; adjustable for full runs)
Warmup Steps: 5
Max Steps: 60
Optimizer: AdamW 8-bit
Weight Decay: 0.01
LR Scheduler: Linear
Precision: FP16 (fallback to BF16 if supported)
Logging: Every 10 steps, integrated with Weights & Biases (wandb)
Save Strategy: Every 100 steps, limit of 2 checkpoints
Evaluation: Every 50 steps
Dataset:
Source: JSON file containing pentest scenarios
Structure: Train (60%), Validation (20%), Test (20%) split
Preprocessing: Generates command-by-command prompts with random questions from pre_attack_discussion, variable detail levels (short, medium, detailed), and results from tools like nmap and Metasploit
Format: Prompt-response pairs with EOS token appended
Tools Used:
nmap: For port scanning and service enumeration (e.g., OpenSSH, Redis)
searchsploit: To identify known vulnerabilities in services (e.g., OpenSSH 7.2p2 username enumeration)
msfconsole (Metasploit): For exploit searches and automated attack attempts (e.g., SSH login brute-forcing)
GPU Support
GPU Detection: Checks CUDA availability and GPU name (target: NVIDIA GTX 1060 3GB)
CUDA Version: Compatible with CUDA 12.1
Optimization: 4-bit quantization and gradient checkpointing to fit within 3GB VRAM constraints
Dataset Example
Goal: Exploit vulnerabilities in OpenSSH 7.2p2 on port 22 to gain initial access, then assess Redis on port 6383 for unauthenticated data exposure and potential lateral movement.
Dataset
Trained on Nielzac/cybersecurity-sft-dataset, containing prompts and responses for cybersecurity scenarios, filtered to exclude failed commands.
Uploaded model
Developed by: Nielzac
License: apache-2.0
Finetuned from model : unsloth/Qwen3-1.7B-unsloth-bnb-4bit
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.
