gvic-unb/dodf-saude-qa

DODF Public Health QA

Dataset Summary

DODF Public Health QA is a synthetic question-answering (QA) dataset for evaluating Retrieval-Augmented Generation (RAG) systems over official public health publications from the Diário Oficial do Distrito Federal (DODF), the Government Gazette of the Federal District, Brazil.

The dataset focuses on location-aware questions about public health facilities and services — such as Basic Health Units (UBS), Emergency Care Units (UPA), hospitals, and health programs — distributed across the 58 administrative regions of the Federal District.

Dataset Structure

The dataset contains three splits:

Split	Size	Description
test	350 pairs	Final evaluation set
validation	150 pairs	Used for prompt refinement and hyperparameter tuning
corpus	1,381 documents	RAG knowledge base — DODF public health publications from 2025

QA splits (test and validation)

Each instance contains the following fields:

Field	Type	Description
pergunta	string	Question in Brazilian Portuguese
resposta	string	Reference answer, grounded in the source publication
secao	string	DODF section of the source document (e.g., Section I, Section III)
fonte	string	Publication date of the source document
tipo_documento	string	Document type (e.g., Contrato, Termo de Doação, Portaria)
texto_publicacao	string	Full text of the source publication(s), including edition and date metadata

Knowledge base split (corpus)

Each instance corresponds to a single DODF publication and contains the following fields:

Field	Type	Description
id	int	Unique publication identifier
edicao	int	DODF edition number
data_edicao	string	Edition date (YYYY-MM-DD)
data	string	Publication date (YYYY-MM-DD)
secao	string	DODF section (e.g., Seção I, Seção III)
tipo_documento	string	Document type (e.g., Portaria, Extrato, Edital, Resolução)
numero_documento	string	Official document number/title
referencia	string	Reference contract, notice, or act number
texto_completo	string	Full text of the publication
topicos	list[string]	Public health topics mentioned (e.g., dengue, saúde bucal)
localidades	list[string]	Administrative regions of the Federal District mentioned
tipo_unidade	list[string]	Types of health facilities referenced (e.g., hospital, UBS)
programas	list[string]	Health programs referenced

Knowledge Base Construction

The knowledge base comprises 1,381 publications drawn from 229 editions of the DODF issued in 2025. Publications were prioritized from Sections I and III, which cover government acts, administrative decisions, contracts, and public health policies. Section II, which covers internal personnel matters, was excluded. Publications were filtered for public health relevance using rule-based criteria, retaining only documents mentioning health facilities, programs, or services pertaining to the 58 administrative regions of the Federal District.

QA Corpus Construction

While human-authored questions remain the gold standard, producing them is a demanding task that requires experts to review a large volume of publications. Consequently, generating synthetic test sets has become a widely adopted alternative, wherein an LLM produces questions from reference documents. Following this approach, QA pairs were generated synthetically from the knowledge base using Claude Haiku 4.5 (claude-haiku-4-5-20251001), with prompts designed with the support of a public health expert to simulate the inquiry style of researchers, managers, and citizens.

Generation

Two types of location-aware questions were produced:

• Type I — questions whose answers are found within a single DODF publication.
• Type II — questions that require information aggregated from multiple publications (up to 5 documents) for a specific location. For these, all publications mentioning a given administrative region (e.g., Ceilândia) are grouped together, and the model is prompted to generate questions whose answers span the group's content.

Filtering

Generated pairs were filtered through the following steps:

1. Pairs with empty question or answer fields were discarded.
2. Pairs whose questions did not contain any target location from the reference dataset were removed via regex matching.
3. Pairs whose questions showed no lexical relation to public health terminology (e.g., saúde, UBS, hospital, SUS) were discarded through a keyword-based heuristic.
4. A lexical grounding check verified that at least 30% of the answer tokens appear in the source text, mitigating hallucinated responses.

Statistics

	Type I	Type II	Total
Raw pairs generated	580	376	956
Discarded	80	76	156
Valid pairs	500	300	800

After filtering, pairs were shuffled and split: 150 samples were selected for the validation set (used to refine prompts and tune LLM hyperparameters such as temperature and top-K), and 350 samples were kept for the final evaluation set.

Languages

Brazilian Portuguese (pt-BR).

Licensing

This corpus is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

If you use this corpus, please cite:

@inproceedings{freire2026locationaware,
  title = {Evaluating Location-Aware {RAG} Approaches for Question Answering over Public Health--Related Content in {Brazilian} Official Gazettes},
  author = {Freire, Arthur S. M. and Borges, Vinicius R. P. and Martins, Anna Carolina F. and Vidal, Fl{\'a}vio de Barros and Araujo, Aleteia and
            Briga{\~{a}}o, Fabr{\'i}cio R. and Veloso, Gabriel M. and Maia, Edward T. and Martins, Wagner de Jesus and Garcia, Luis P. F.},
  booktitle = {36th Brazilian Conference on Intelligent Systems, {BRACIS} 2026, Cuiabá, Mato Grosso, October 19-22 (to appear)},
  publisher = {Springer},
  year = {2026}
}

Contact

For questions or feedback, please open an issue in the project repository.