Llama-Primus-Merged

Maintained By
trendmicro-ailab

Llama-Primus-Merged

PropertyValue
AuthorTrendMicro AI Lab
LicenseMIT + Llama 3.1 Community License Agreement
Base ModelLlama-3.1-8B-Instruct
SpecializationCybersecurity

What is Llama-Primus-Merged?

Llama-Primus-Merged is a specialized cybersecurity language model that combines the power of Llama-3.1-8B-Instruct with extensive cybersecurity training. The model was pre-trained on a massive 2.77B token cybersecurity corpus (Primus-Seed and Primus-FineWeb) and fine-tuned on approximately 1,000 carefully curated cybersecurity QA tasks through Primus-Instruct.

Implementation Details

The model follows a three-stage development process: initial pre-training on cybersecurity corpus, instruction fine-tuning for task-specific capabilities, and finally merging with Llama-3.1-8B-Instruct to maintain general instruction-following abilities while enhancing cybersecurity performance.

  • 14.84% improvement in aggregated cybersecurity benchmark scores
  • Enhanced performance in CTI-Bench, CyberMetric, SecEval, and CISSP evaluations
  • Maintains strong function calling capabilities (74.77% on BFCL V2)
  • Robust multilingual support with MMLU scores above 47% across multiple languages

Core Capabilities

  • Cyber Threat Intelligence (CTI) analysis and classification
  • CVE to CWE mapping with 66.20% accuracy
  • CVSS scoring with improved accuracy (lower MAD of 1.1233)
  • Strong performance in cybersecurity certification exam questions
  • Balanced safety measures with improved resistance to various security attacks

Frequently Asked Questions

Q: What makes this model unique?

The model's uniqueness lies in its specialized cybersecurity training while maintaining general-purpose capabilities. It demonstrates significant improvements in cybersecurity tasks without compromising its instruction-following abilities or multilingual support.

Q: What are the recommended use cases?

The model is particularly suited for cybersecurity applications including threat intelligence analysis, vulnerability assessment, security certification training, and general security knowledge queries. It maintains strong performance in general tasks while excelling in security-specific scenarios.

🍰 Interesting in building your own agents?
PromptLayer provides Huggingface integration tools to manage and monitor prompts with your whole team. Get started here.