| --- |
| license: apache-2.0 |
| pretty_name: VulInstruct Specification Dataset |
| task_categories: |
| - text-classification |
| tags: |
| - vulnerability-detection |
| - security |
| - CVE |
| - code-analysis |
| - specification |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train-* |
| dataset_info: |
| features: |
| - name: cve_id |
| dtype: string |
| - name: cve_description |
| dtype: string |
| - name: repository |
| dtype: string |
| - name: commit_hash |
| dtype: string |
| - name: patch_commit_msg |
| dtype: string |
| - name: specifications |
| sequence: string |
| - name: understand |
| dtype: string |
| - name: threat_model |
| dtype: string |
| - name: vulnerability_analysis |
| dtype: string |
| - name: solution_analysis |
| dtype: string |
| - name: split |
| dtype: string |
| - name: case_id |
| dtype: string |
| splits: |
| - name: train |
| num_bytes: 18719704 |
| num_examples: 1999 |
| download_size: 8226159 |
| dataset_size: 18719704 |
| --- |
| |
| # VulInstruct: Specification-Guided Vulnerability Detection 🔒 |
|
|
| <p align="center"> |
| 📄 <a href="https://huggingface.co/papers/2511.04014" target="_blank">Paper</a> | |
| 🤗 <a href="https://huggingface.co/datasets/hzhu721/vulnerability-specifications" target="_blank">Dataset</a> | |
| 💻 <a href="https://github.com/zhuhaopku/VulInstruct-temp" target="_blank">GitHub</a> |
| </p> |
|
|
| ## Introduction |
|
|
| This repository contains the **Specification Knowledge Base** introduced in the paper [VulInstruct: Teaching LLMs Root-Cause Reasoning for Vulnerability Detection via Security Specifications](https://huggingface.co/papers/2511.04014). |
|
|
| VulInstruct is a specification-guided approach that systematically extracts security specifications—expectations about how code should behave to remain safe—from historical vulnerabilities. The dataset consists of 1,999 high-quality vulnerability-specification pairs designed to help Large Language Models (LLMs) reason about expected safe behaviors rather than relying on surface patterns. |
|
|
| The dataset includes: |
| - **General specifications**: fundamental safe behaviors captured from high-quality patches across various projects. |
| - **Domain-specific specifications**: expectations derived from repeated violations within particular repositories. |
|
|
| ## 📢 Updates |
| - **[2026/01/21]** 🎉 Released VulInstruct Specification Knowledge Base with 1,999 high-quality vulnerability-specification pairs! |
|
|
| ## Dataset Structure |
|
|
| Each example in the dataset includes the following features: |
| - `cve_id`: The CVE identifier. |
| - `cve_description`: Official description of the vulnerability. |
| - `repository`: The source code repository. |
| - `commit_hash`: The hash of the vulnerable commit. |
| - `patch_commit_msg`: The commit message from the fix. |
| - `specifications`: Extracted expectations about how the code should behave. |
| - `vulnerability_analysis`: Root-cause reasoning for the flaw. |
| - `solution_analysis`: Reasoning behind the fix. |
|
|
| ## Citation |
|
|
| If you use this dataset or the VulInstruct approach in your research, please cite: |
|
|
| ```bibtex |
| @article{zhu2025vulinstruct, |
| title={VulInstruct: Teaching LLMs Root-Cause Reasoning for Vulnerability Detection via Security Specifications}, |
| author={Hao Zhu and Jia Li and Cuiyun Gao and Jiaru Qian and Yihong Dong and Huanyu Liu and Lecheng Wang and Ziliang Wang and Xiaolong Hu and Ge Li}, |
| journal={arXiv preprint arXiv:2511.04014}, |
| year={2025} |
| } |
| ``` |
