datafromlopes/atlas-sql-br

AtlasSQL-BR 是一个基于巴西葡萄牙语的新颖地理空间Text-to-SQL数据集，构建于真实世界的开放政府数据之上。它整合了学校普查数据、公共设施注册表以及巴西官方地理边界层次结构，为评估和训练空间查询模型提供了一个稳健的基准。数据集架构围绕主题公共设施表及其相关地理边界层组织，要求模型在单个SQL语句中推理不同行政粒度之间的空间包含关系。核心表包括源自巴西学校普查的microdados_ed_basica表，包含每所学校的数百个基础设施和教学属性，以及其他公共设施的记录。数据集还涵盖了IBGE的所有七个官方地理边界级别，并使用PostGIS扩展处理空间关系，覆盖了22个PostGIS空间函数。数据记录包含问题、SQL代码、行政边界级别、查询复杂度层级、地理空间函数和数据来源等字段。数据集分为四个复杂度层级，从简单到非常难，基于SQL查询的结构特性分类。文件包括标准的训练和验证分割，并提供了可重现性指南，强调使用原始980个手动标注的问题对进行基准实验。AtlasSQL-BR is a novel geospatial Text-to-SQL dataset in Brazilian Portuguese, built upon real-world open-government data. Integrating school census data, public facility registries, and Brazils official geographic boundary hierarchies, this dataset provides a robust benchmark for evaluating and training models on spatial queries. The database schema is organized around thematic public facility tables and their associated geographic boundary layers, requiring models to reason about spatial containment across different administrative granularities within a single SQL statement. The core table is microdados_ed_basica, derived from the Brazilian School Census, containing hundreds of infrastructural and pedagogical attributes per school, along with georeferenced records for other public facilities. The dataset incorporates all seven official geographic boundary levels from IBGE and uses PostGIS extensions to resolve spatial relationships, covering 22 PostGIS spatial functions. Each record includes fields such as question, sql_code, territorial_division, level, geospatial_functions, and source. It comprises manually annotated question-SQL pairs classified into four complexity tiers based on SQL query structural properties, from easy to very hard. The dataset is divided into standard train and validation splits, with reproducibility guidelines emphasizing the use of the original 980 manually annotated questions for baseline experiments.