datafromlopes/atlas-sql-br

AtlasSQL-BR

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 Brazil's official geographic boundary hierarchies, this dataset provides a robust benchmark for evaluating and training models on spatial queries.

📊 Dataset Architecture and Structure

The database schema underlying AtlasSQL-BR is organized around thematic public facility tables and their associated geographic boundary layers. This multi-level structure requires models to reason about spatial containment across different administrative granularities within a single SQL statement.

• Thematic Tables (Point Geometries): The core table is microdados_ed_basica, derived from the Brazilian School Census, containing hundreds of infrastructural and pedagogical attributes per school. The schema also includes georeferenced records for other public facilities: libraries, CRAS, CREAS, and Centros POP.
• Boundary Tables (Polygon Geometries): The dataset incorporates all seven official geographic boundary levels from IBGE: country, macro-region, state, municipality, district, neighborhood, and street.
• Spatial Operators: Spatial relationships between facility points and boundary polygons are resolved using PostGIS extensions. The dataset covers a vocabulary of 22 PostGIS spatial functions (e.g., ST_Intersects, ST_Within, ST_Contains, ST_Distance).

Data Fields

Based on the dataset structure, each record contains the following fields:

• question: The natural language query in Brazilian Portuguese.
• sql_code: The corresponding executable PostGIS/SQL query.
• territorial_division: The administrative boundary level relevant to the query.
• level: The structural complexity tier of the query.
• geospatial_functions: The primary PostGIS functions required to solve the query.
• source: Indicates the origin of the data. It can be base_dataset (the original 980 manually annotated questions) or partitions generated by data augmentation techniques (such as delete, insert, swap, synonym, translate).

📈 Complexity Tiers

The dataset comprises manually annotated question-SQL pairs classified into four complexity tiers based on the structural properties of the SQL query (all criteria are AND-combined within each tier):

Tier	Structural Criteria
Easy	<1 JOIN; no aggregations; no subqueries; direct spatial filter.
Medium	2-3 JOINs; no aggregations; multiple spatial conditions.
Hard	3-4 JOINs; >1 aggregation (COUNT, SUM, AVG); multiple spatial functions.
Very Hard	4+ JOINs; multiple aggregations; subqueries and/or CTEs; window functions (ROW_NUMBER, RANK, DENSE_RANK); CASE WHEN; possible UNION.

📁 Files and Reproducibility

The dataset is divided into standard train.parquet and validation.parquet splits. Two tagged snapshots are preserved so that published results remain reproducible:

Tag	Content	Used in
paper-released	980 curated pairs (source = base_dataset) plus five augmented partitions in train.parquet; a separate 196-pair validation.parquet	SBBD 2026 paper (first execution of the experiments)
thesis-split (current main)	The 980 curated pairs partitioned into 784 training and 196 validation pairs, stratified by complexity tier, spatial function and territorial division (one validation pair per tier × function × division cell, seed 42); split_manifest.json records the counts and checksums	Master's thesis (second execution of the experiments)

Loading the thesis split (Python)

from datasets import load_dataset

dataset = load_dataset("datafromlopes/atlas-sql-br", revision="thesis-split")
print(len(dataset["train"]), len(dataset["validation"]))   # 784 196

The thesis split is generated by core/make_paper_split.py in the code repository, which also holds the training, generation and evaluation pipeline and the results of the thesis run. The fine-tuned adapter is published as datafromlopes/llama-3.1-8b-atlassql-br.

Reproducing the paper

⚠️ Important Note on Reproducibility: The baseline experiments documented in the original paper were conducted exclusively on the original 980 question-SQL pairs. The paper-released snapshot also includes additional data generated through data augmentation techniques.

To exactly replicate the experiments and results from the original paper, you must:

1. Use the dataset version associated with the paper-released tag.
2. Filter the data to include only the records where the source column is exactly base_dataset. The other partitions (delete, insert, swap, synonym, translate) contain synthetic augmented data and should be excluded when reproducing the paper's base results.

How to load the dataset for reproducibility (Python)

You can easily download and filter the dataset using the Hugging Face datasets library to match the exact setup used in the paper:

from datasets import load_dataset

# 1. Load the dataset using the specific tag for the paper's release
dataset = load_dataset("datafromlopes/atlas-sql-br", revision="paper-released")

# 2. Filter both train and validation splits to include only the original 980 questions
base_train = dataset["train"].filter(lambda x: x["source"] == "base_dataset")
base_validation = dataset["validation"].filter(lambda x: x["source"] == "base_dataset")

print(f"Original Train size: {len(base_train)}")
print(f"Original Validation size: {len(base_validation)}")

✒️ Citation

If you find this dataset useful for your research, please cite the original paper: