STINER
收藏资源简介:
STINER是一个面向战略网络威胁情报的命名实体识别数据集,由穆罕默德六世理工大学和德勤咨询联合构建,专注于从社交媒体(X平台)中提取高层次威胁信息。该数据集包含2100条专家标注的推特,总计85066个token,涵盖目标、威胁行为者、行业、地点、数据量等8种战略实体类型,平均每条推特标注8.23个实体。创建过程首先从约66,500条推特中筛选出相关情报,再由两位网络安全专家使用Label Studio进行标注,标注一致性达Cohen's κ=0.84。数据集旨在解决传统NER模型在社交媒体非正式语言上识别战略情报的盲区,可辅助安全分析师自动识别受攻击行业、威胁组织及影响规模,为组织风险决策提供实时支持。
STINER is a named entity recognition (NER) dataset for strategic cyber threat intelligence, co-developed by Mohammed VI Polytechnic University and Deloitte Consulting. It focuses on extracting high-level threat information from social media (X platform). This dataset contains 2,100 expert-annotated Twitter posts, totaling 85,066 tokens, and covers 8 strategic entity types including targets, threat actors, industries, locations, data volumes and others. On average, each Twitter post is annotated with 8.23 entities. During the creation process, relevant intelligence was first screened from approximately 66,500 Twitter posts, then annotated by two cybersecurity experts using Label Studio, with an annotation consistency of Cohen's κ=0.84. This dataset aims to address the blind spots of traditional NER models in identifying strategic intelligence from informal social media language, and can assist security analysts in automatically identifying attacked industries, threat organizations and impact scales, providing real-time support for organizational risk decision-making.
STINER 数据集详情
数据集概述
STINER(Structured Threat Intelligence Named Entity Recognition)是一个用于网络威胁情报(CTI)领域命名实体识别(NER)任务的数据集,专注于从 Twitter/X 社交媒体帖子中提取结构化信息。
实体类型
数据集包含 9 种实体类型:
- ACTOR:威胁行为者/组织
- TARGET:被针对的组织/个人
- LOCATION:地理位置
- SECTOR:行业领域
- DATE:时间引用
- CAMPAIGN:行动/活动名称
- TOOL:攻击工具/恶意软件
- DATA_TYPE:被盗/受影响的数据类型
- PRICE:赎金金额
数据构成
- 数据文件:已标注的训练集、验证集和测试集划分(train.json、val.json、test.json)
- 原始数据:包含推特 ID 列表(tweet_ids.txt),可用于推文重新水合(rehydration)
- 数据规模:数据集详情中未明确标注样本总量
模型评估框架
- 编码器模型:BERT、RoBERTa、SecBERT、CySecBERT、DarkBERT(可选 CRF 增强)
- 大语言模型:Llama-3.1、Gemma-2、Qwen2.5(零样本与微调)
- GLiNER:零样本评估
基准结果
结果存储于 results/tables/ 目录,包含:
encoder_benchmark_summary.csv:编码器模型结果benchmark_results_encoders_crf.csv:编码器+CRF 结果llm_benchmark_results.csv:大语言模型结果
许可证
MIT License

- 1STINER: Automated Extraction of Strategic Cyber Threat Intelligence from X穆罕默德六世理工大学; 德勤咨询 · 2026年




