QCRI/AZERG-MixTask-Mistral
QCRI/AZERG-MixTask-Mistral is a 7 billion parameter Mistral-based language model fine-tuned by QCRI for Cyber Threat Intelligence (CTI) tasks. It specializes in STIX data generation, covering entity detection, entity type identification, related pair detection, and relationship type identification. This model is designed to extract STIX entities and relationships from security reports, making it versatile for comprehensive CTI analysis.
Loading preview...
QCRI/AZERG-MixTask-Mistral Overview
QCRI/AZERG-MixTask-Mistral is a 7 billion parameter model based on mistralai/Mistral-7B-Instruct-v0.3, specifically fine-tuned by QCRI for Cyber Threat Intelligence (CTI) applications. Its primary focus is on automating STIX (Structured Threat Information Expression) data generation from security reports.
Key Capabilities
This model is distinguished by its specialization in four core STIX extraction sub-tasks:
- Entity Detection (T1): Identifying all relevant STIX entities within a given text.
- Entity Type Identification (T2): Classifying the detected entities into their specific STIX types (e.g., ATTACK_PATTERN).
- Related Pair Detection (T3): Pinpointing pairs of entities that have a relationship.
- Relationship Type Identification (T4): Determining the specific type of relationship between identified entity pairs.
It is the most versatile model within the AZERG collection, capable of handling all these sub-tasks comprehensively.
Intended Use and Integration
The model is designed to be used within the AZERG framework for extracting STIX entities and relationships. Users should refer to the framework for exact prompting guidelines. An example prompt for Task 1 (Entity Detection) demonstrates its structured input/output format, requiring identified entities to be listed within <entities> tags, separated by pipes.
Citation
If you utilize this model, please cite the associated paper: Lekssays, Ahmed, Sencar, Husrev Taha, and Yu, Ting. "From Text to Actionable Intelligence: Automating STIX Entity and Relationship Extraction." arXiv preprint arXiv:2507.16576 (2025).