Voidreaper2026/qwen3-4b-training

这是一个大规模网络安全指令调优数据集，采用ShareGPT对话格式，从多个权威开源数据源（包括NVD、OSV、GitHub Advisory Database、Cybersec Causal Reasoning、Security Stack Exchange、ExploitDB、MITRE ATT&CK、CISA KEV、Kali Linux Tools和Vulners）汇编而成，并经过去重处理，包含1,807,941条去重记录。该数据集旨在用于微调大型语言模型（LLM），以支持网络安全推理、漏洞分析、威胁情报、安全运营中心（SOC）助手模型、安全感知编码助手以及CTF/渗透测试知识库等应用。数据格式为JSON，包含human和gpt角色的对话，部分记录还包含系统提示。去重过程使用基于源、源ID和对话内容MD5哈希的智能键，以确保跨源CVE覆盖的同时移除重复项。许可证为Apache 2.0，各源数据集保留其原始开放许可证。A large-scale cybersecurity instruction-tuning dataset in ShareGPT conversational format, assembled from multiple authoritative open sources (including NVD, OSV, GitHub Advisory Database, Cybersec Causal Reasoning, Security Stack Exchange, ExploitDB, MITRE ATT&CK, CISA KEV, Kali Linux Tools, and Vulners) and deduplicated, with 1,807,941 deduplicated records. The dataset is intended for fine-tuning large language models (LLMs) for cybersecurity reasoning, vulnerability analysis, threat intelligence, SOC assistant models, security-aware coding assistants, and CTF/penetration testing knowledge bases. The data format is JSON, featuring conversations with human and gpt roles, and some records include a system prompt. Deduplication is performed using a smart key combining source, source_id, and MD5 hash of conversation content to retain cross-source CVE coverage while removing true duplicates. The license is Apache 2.0, with individual source datasets retaining their original open licenses.