Bilker

CTF/Cybersecurity AI Assistant using QLoRA Fine-tuning

Transforms diverse cybersecurity educational content (research papers, CTF writeups, exploit code) into training datasets for specialized AI assistant development.

Overview

Architecture: Two-phase modular pipeline for reliable data processing and model training Base Model: QWen3-Coder:30B fine-tuned with QLoRA Dataset: 12,601 Q&A pairs from 2,241 cybersecurity source files

Pipeline

Raw Data → process_data.py → Chunks → chunk_processor.py → Training Dataset → QLoRA Training

Phase 1: Content extraction and intelligent chunking Phase 2: LLM-powered Q&A pair generation using local llama3.1:8b

Requirements

Data Processing: 8GB+ RAM, local LLM capability Model Training: NVIDIA GPU 16GB+ VRAM, 32GB+ system RAM

Usage

conda create -n bilker python=3.11
pip install -r requirements.txt
ollama pull llama3.1:8b

python process_data.py      # Extract and chunk data
python chunk_processor.py   # Generate Q&A pairs

Results

From comprehensive cybersecurity dataset including PicoCTF challenges, HackTheBox writeups, academic research, and exploit repositories:

• 87% chunk-to-QA conversion success rate
• Professional-grade training dataset ready for fine-tuning
• Modular architecture enables resumable processing and experimentation

Status

• Complete: Data processing pipeline, training dataset generation
• In Progress: QLoRA fine-tuning implementation
• Planned: Model deployment and evaluation

Built for cybersecurity education and ethical security research.