| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - ai-security |
| - sft |
| - dpo |
| - red-teaming |
| - llm-vulnerability |
| - cybersecurity |
| pretty_name: CyberSec-LLM-Auditing-SFT-DPO |
| size_categories: |
| - n<1K |
| --- |
| |
| # Dataset Card: CyberSec-LLM-Auditing-SFT-DPO |
|
|
| ## Dataset Summary |
| This dataset is a highly specialized collection of **64 high-quality examples** tailored for **AI Red Teaming** and |
| **Offensive Security research**. It is designed to fine-tune models into technical assistants that prioritize substantive, |
| technical analysis over standard safety refusals or legal disclaimers. |
|
|
| The dataset was used to fine-tune the `Qwen3-1.7b_for_cybersec` model, achieving a state where the model provides direct |
| technical insights while maintaining professional rigor. |
|
|
| ## Behavioral Alignment & Objectives |
| The primary innovation of this dataset is the specific "behavioral shaping" it imposes on the model: |
| - **Removal of Refusal Layers:** Unlike base models, this dataset trains the model to skip generic legal disclaimers |
| (e.g., "I cannot help as you need permission") and proceed directly to technical problem-solving. |
| - **Technical Precision:** Examples focus on high-density technical responses, covering vulnerabilities like **Prompt Injection**, |
| **Model Inversion**, and **OWASP Top 10 for LLMs**. |
| - **Interactive Auditing:** The model is trained to identify insufficient information. Instead of guessing or refusing, it |
| is conditioned to **ask for clarification or missing technical details** to provide a more accurate audit. |
|
|
| ### Comprehensive Auditing Methodologies |
| The dataset is structured to support a full-spectrum security assessment: |
| - **White-box & Gray-box Testing:** Deep technical workflows for auditing models with full or partial internal access, |
| including structural analysis and weight-based vulnerability probing. |
| - **Black-box (API-based) Attacks:** Advanced simulations of real-world threat vectors against secured endpoints, |
| focusing on bypassing input/output sanitization and external safety guardrails. |
|
|
| ### Adversarial Vector Coverage |
| The dataset covers a wide array of the most critical and "state-of-the-art" attack patterns, specifically aligned with the |
| newest threat vectors: |
| - **Injection & Hijacking:** From direct prompt injection to complex **Recursive Injections** within agentic workflows. |
| - **Structural & Behavioral Jailbreaking:** Advanced techniques using segmented prompts, divider-style mutations (e.g., Libertas/Pliny), |
| and technical headers to override system constraints. |
| - **Multi-turn Escalation:** Training the model to maintain technical context during long-form audits, including |
| **Crescendo-style escalations** to probe deep security layers. |
| - **Data & Model Privacy:** Scenarios involving Model Inversion, Sensitive Data Disclosure, and Training Data Poisoning. |
|
|
| ## Intended Use |
| - **Professional AI Auditing:** Accelerating the testing of LLM guardrails. |
| - **Security Research:** Understanding and simulating modern AI attack vectors in a controlled environment. |
|
|
| ## Ethical Disclaimer |
| This dataset is designed for **authorized security testing only**. By removing standard refusals, the dataset transfers the ethical |
| responsibility to the human operator. It should be used exclusively in legal AI Red Teaming engagements and academic research. |
|
|
| ## About the Author |
| **Antoni Błoch** – **Cybersecurity student at AGH University of Krakow**. |
