trendmicro-ailab/Llama-Primus-Base

Llama-Primus-Base: Cybersecurity-Optimized LLM

Llama-Primus-Base is an 8 billion parameter foundation model from Trend Micro's AILab, built upon Llama-3.1-8B-Instruct. Its core differentiator is its continual pre-training on specialized cybersecurity datasets:

• Primus-Seed (0.2B tokens): A manually curated, high-quality dataset of cybersecurity texts.
• Primus-FineWeb (2.57B tokens): Cybersecurity-specific content filtered from the broader FineWeb corpus.

This targeted pre-training strategy addresses the lack of open-source, domain-specific datasets for cybersecurity LLMs, enabling the model to acquire deep knowledge in this field.

Key Capabilities & Performance

The model demonstrates significantly enhanced performance on cybersecurity benchmarks compared to its base model, Llama-3.1-8B-Instruct. It achieves a 15.88% improvement in aggregated scores across various metrics, including:

• CISSP (Exams in book): 0.7230 (vs. 0.7073)
• CTI-Bench (MCQ): 0.6676 (vs. 0.6420)
• CTI-Bench (CVE → CWE): 0.6780 (vs. 0.5910)
• CyberMetric (500): 0.8660 (vs. 0.8560)

These results highlight its proficiency in understanding and processing complex cybersecurity information, making it particularly effective for tasks requiring specialized domain knowledge.

Use Cases

Llama-Primus-Base is ideal for applications requiring advanced cybersecurity understanding and reasoning. Its specialized training makes it suitable for tasks such as:

• Cyber threat intelligence analysis
• Vulnerability assessment and management
• Security operations and incident response support
• Educational tools for cybersecurity professionals

This model is part of Trend Micro's Primus family, which aims to contribute powerful, efficiency-optimized cybersecurity models and datasets to the community.