×
模态框(Modal)标题
在这里添加一些文本
Close
Close
Submit
Cancel
Confirm
×
模态框(Modal)标题
×
Toggle navigation
Home
About Journal
Basic Information
Copyright Transfer Agreement
Open Access Policy
Copyright and Archive Policy
Journal Browsing
Table of Contents
Online First
Archived Issues
Editorial Board
Author Center
Author Guidelines
Publication Ethics
Policies on Academic Dishonesty
Editorial Policy
Download
Subscription
Contact Us
中文
Search
E-mail
RSS
×
Quice Search
Figure/Table detail
LEAD-Cyber:a local fine-tuned vertical domain LLMs for cybersecurity with open-source LLMs and full-cycle training datasets
GUAN Yongjian, PIAO Chengkai, WANG Buhong, ZHAO Bofu, LI Siqi, ZHAO Zhengyang
Journal of Cybersecurity
, DOI:
10.20172/j.issn.2097-3136.250116
种类
样本数
Tokens
平均Token
MITRE 攻击技术
8423
2748424
326.3
MITRE 战术
401
80601
201
MITRE 软件工具
8286
1781490
215
MITRE 攻击组织
1849
523267
283
MITRE 活动事例
440
114884
261.1
MITRE 缓解措施
1095
277254
253.2
MITRE 实体关系
1633
436500
267.3
Table 2
Instruction Fine-tune Dataset statistics
Other figure/table from this article
Fig.1
The overview of LEAD-Cyber.
Table 1
Pre-training Dataset statistics
Table 3
Reasoning Fine-tune Dataset statistics
Table 4
CTI-Bench Reasoning Task
Table 5
HackMentor Dataset statistics
Fig.2
Example of Q&A for Data Generation
Fig.3
Schematic diagram of LoRA
Fig.4
Experimental results the test dataset
Fig.5
Rouge-L experimental results
Fig.6
Qwen score diagram of LEAD
Fig.7
WinRate of each combination of models.
Fig.8
Experimental results on CTI-Bench
Fig.9
Rouge-L on experimental results on CTI-Bench
Fig.10
Qwen score diagram of CTI-Bench
Fig.11
WinRate of each combination of models.
Fig.12
Experimental results on HackMentor-eval
Fig.13
Rouge-L on experimental results
Fig.14
Qwen score diagram of HackMentor-eval
Fig.15
WinRate of each combination of models.