Datasets:
Tasks:
Text Generation
Modalities:
Text
Formats:
json
Languages:
English
Size:
10K - 100K
ArXiv:
License:
Update README.md
Browse files
README.md
CHANGED
@@ -11,4 +11,52 @@ language:
|
|
11 |
- en
|
12 |
size_categories:
|
13 |
- 10K<n<100K
|
14 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
- en
|
12 |
size_categories:
|
13 |
- 10K<n<100K
|
14 |
+
---
|
15 |
+
|
16 |
+
# AZERG-Dataset
|
17 |
+
|
18 |
+
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.
|
19 |
+
|
20 |
+
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.
|
21 |
+
|
22 |
+
# π Dataset Overview
|
23 |
+
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.
|
24 |
+
|
25 |
+
The extraction process is divided into four sequential subtasks:
|
26 |
+
- T1: Entity Detection: Identifying all STIX entities (SDOs and SCOs) in a text passage.
|
27 |
+
- T2: Entity Type Identification: Assigning a specific STIX type to each detected entity.
|
28 |
+
- T3: Related Pair Detection: Identifying which pairs of entities are semantically related based on the text.
|
29 |
+
- T4: Relationship Type Identification: Determining the precise STIX relationship type (e.g., uses, targets) between a related pair of entities.
|
30 |
+
|
31 |
+
## π Dataset Structure
|
32 |
+
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.
|
33 |
+
```
|
34 |
+
AZERG-Dataset/
|
35 |
+
βββ train/
|
36 |
+
β βββ azerg_T1_train.json
|
37 |
+
β βββ azerg_T2_train.json
|
38 |
+
β βββ azerg_T3_train.json
|
39 |
+
β βββ azerg_T4_train.json
|
40 |
+
β βββ azerg_MixTask_train.json # Combined data for all tasks
|
41 |
+
βββ test/
|
42 |
+
βββ annoctr_T1_test.json
|
43 |
+
βββ annoctr_T2_test.json
|
44 |
+
βββ annoctr_T3_test.json
|
45 |
+
βββ annoctr_T4_test.json
|
46 |
+
βββ azerg_T1_test.json
|
47 |
+
βββ azerg_T2_test.json
|
48 |
+
βββ azerg_T3_test.json
|
49 |
+
βββ azerg_T4_test.json
|
50 |
+
```
|
51 |
+
|
52 |
+
## π Citation
|
53 |
+
If you use this dataset in your research, please cite the original paper (ArXiv for now, the paper is accepted at RAID 2025):
|
54 |
+
|
55 |
+
```
|
56 |
+
@article{lekssays2025azerg,
|
57 |
+
title={From Text to Actionable Intelligence: Automating STIX Entity and Relationship Extraction},
|
58 |
+
author={Lekssays, Ahmed and Sencar, Husrev Taha and Yu, Ting},
|
59 |
+
journal={arXiv preprint arXiv:2507.16576},
|
60 |
+
year={2025}
|
61 |
+
}
|
62 |
+
```
|