AZERG-Dataset

This repository contains the AZERG-Dataset, a comprehensive collection of annotated cyber threat intelligence (CTI) reports designed for training and evaluating models on STIX entity and relationship extraction.

This dataset was created for the paper: "From Text to Actionable Intelligence: Automating STIX Entity and Relationship Extraction". It is the largest publicly available dataset of its kind, meticulously annotated with STIX-compliant entities and relationships to facilitate the development of automated threat intelligence tools.

📖 Dataset Overview

The AZERG-Dataset is built from 141 real-world threat analysis reports and contains 4,011 STIX entities and 2,075 STIX relationships. It was curated to address the lack of training data for automated STIX report generation and supports a multi-task approach to threat intelligence extraction.

The extraction process is divided into four sequential subtasks:

• T1: Entity Detection: Identifying all STIX entities (SDOs and SCOs) in a text passage.
• T2: Entity Type Identification: Assigning a specific STIX type to each detected entity.
• T3: Related Pair Detection: Identifying which pairs of entities are semantically related based on the text.
• T4: Relationship Type Identification: Determining the precise STIX relationship type (e.g., uses, targets) between a related pair of entities.

📂 Dataset Structure

The dataset is organized into train and test splits. The training and testing data are sourced from completely non-overlapping reports and vendors to ensure a robust evaluation of model generalization.

AZERG-Dataset/
├── train/
│   ├── azerg_T1_train.json
│   ├── azerg_T2_train.json
│   ├── azerg_T3_train.json
│   ├── azerg_T4_train.json
│   └── azerg_MixTask_train.json  # Combined data for all tasks
└── test/
    ├── annoctr_T1_test.json
    ├── annoctr_T2_test.json
    ├── annoctr_T3_test.json
    ├── annoctr_T4_test.json
    ├── azerg_T1_test.json
    ├── azerg_T2_test.json
    ├── azerg_T3_test.json
    └── azerg_T4_test.json

📜 Citation

If you use this dataset in your research, please cite the original paper (ArXiv for now, the paper is accepted at RAID 2025):

@article{lekssays2025azerg,
  title={From Text to Actionable Intelligence: Automating STIX Entity and Relationship Extraction},
  author={Lekssays, Ahmed and Sencar, Husrev Taha and Yu, Ting},
  journal={arXiv preprint arXiv:2507.16576},
  year={2025}
}