đ Primus: A Pioneering Collection of Open-Source Datasets for Cybersecurity LLM Training
Primus offers a set of open - source datasets for training cybersecurity large language models, aiming to fill the gap in the field of cybersecurity LLM pre - training.
đ Quick Start
Primus is a significant project in the field of cybersecurity. It provides a series of datasets for different stages of LLM training in cybersecurity, including pre - training, instruction fine - tuning, and reasoning data for distillation. Based on these datasets, relevant models like Llama - Primus - Reasoning have been developed.
⨠Features
- First cybersecurity reasoning model: Llama - Primus - Reasoning is a reasoning model with a 15.8% improvement in security certification (CISSP).
- Diverse datasets: Covers multiple stages of cybersecurity LLM training, including Primus - Seed, Primus - FineWeb, Primus - Instruct, and Primus - Reasoning.
- Industry - leading foundation: Developed based on advanced research and technology, sharing the foundation of the enterprise - class Trend Cybertron solution.
đĻ Installation
No installation steps are provided in the original document, so this section is skipped.
đ Documentation
Introduction
Large Language Models (LLMs) have shown great potential in various specialized domains. However, there is a lack of open - source datasets for LLM pre - training in the cybersecurity field. Primus fills this gap by providing datasets for different training stages. Based on these datasets and Llama - 3.1 - 8B - Instruct, models like Llama - Primus - Base, Llama - Primus - Merged, and Llama - Primus - Reasoning are developed.
â ī¸ Important Note
No TrendMicro customer information is included.
Cybersecurity Benchmark Results
Property |
Details |
Model Type |
Multiple models are compared, including Llama - 3.1 - 8B - Instruct, Llama - Primus - Merged, and models distilled from o1 - preview and DeepSeek - R1. |
Training Data |
Datasets such as Primus - Reasoning, Primus - Seed, Primus - FineWeb, and Primus - Instruct are used for training. |
Model |
CISSP |
Avg. Tokens |
w/o CoT, 5 - shot |
|
|
Llama - 3.1 - 8B - Instruct |
0.7073 |
1 |
Llama - Primus - Merged |
0.7191 â1.67% |
1 |
w/ CoT, 0 - shot |
|
|
Llama - 3.1 - 8B - Instruct |
0.7288 â3.03% |
279.69 |
ââ + Distilled from o1 - preview |
0.7583 â7.21% |
646.94 |
ââ + Distilled from DeepSeek - R1 |
0.7859 â11.1% |
1667.56 |
ââ + Distilled from (o1 + R1) |
0.7780 â10.0% |
1615.54 |
Llama - Primus - Merged |
0.7603 â7.49% |
241.92 |
ââ + Distilled from o1 - preview |
0.7780 â10.0% |
726.96 |
ââ + Distilled from DeepSeek - R1 |
0.8075 â14.2% |
1483.94 |
ââ + Distilled from (o1 + R1) |
0.8193 â15.8% |
1467.40 |
Raw Models for Comparison |
|
|
o1 - preview |
0.8035 |
1054.91 |
DeepSeek - R1 |
0.8212 |
1229.32 |
DeepSeek - R1 - Distill - Llama - 8B |
0.7399 â4.61% |
1542.10 |
The effect of Primus - Reasoning fine - tuning is evaluated on CISSP. â indicates the percentage improvement over Llama without CoT and in the 5 - shot setting. The best improvement is highlighted in bold.
About Primus
Primus is Trend Micro's pioneering family of lightweight, state - of - the - art open cybersecurity language models and datasets. It shares the innovative foundation of the enterprise - class Trend Cybertron solution. Trend Micro, as an industry leader in cybersecurity, contributes these resources to the community while maintaining high - quality security standards.
đ License
This model is based on the MIT license, but you must also comply with the Llama 3.1 Community License Agreement.