Euroswarms/shodan-queries

这是一个监督微调（SFT）数据集，用于教授语言模型编写和解释Shodan搜索查询。每个助手响应要么是一个由文档化过滤器语法（`filtername:value`）组成的Shodan查询字符串，要么是对给定查询匹配内容的简短事实解释。数据集完全基于公开记录的Shodan过滤器名称、横幅字段和示例查询，模型学习从已知的基本元素组合查询，而不是虚构未记录的过滤器。数据集包含两种行类型：自然语言请求到Shodan查询字符串（nl_to_query）和查询到简短解释（query_to_explain）。此外，数据集还支持不同的增强类型，如国家插槽（country_slot）、端口插槽（port_slot）和改写过采样（paraphrase_oversample）。数据集的基础规模约为3,400行，通过增强可以扩展到1,000,000行。

A supervised fine-tuning (SFT) dataset that teaches a language model to write and explain Shodan search queries. Every assistant response is either a verbatim Shodan query string composed from documented filter syntax (`filtername:value`), or a short factual explanation of what a given query matches. The dataset is grounded entirely in publicly documented Shodan filter names, banner fields, and example queries — the model learns to compose queries from known primitives, not to hallucinate undocumented filters. The dataset includes two row kinds: natural language request to Shodan query string (nl_to_query) and query to short explanation (query_to_explain). Additionally, the dataset supports various augmentation kinds such as country_slot, port_slot, and paraphrase_oversample. The base scale of the dataset is around 3,400 rows, which can be expanded to 1,000,000 rows with augmentation.