recogna-nlp/bulas_qa

---
license: apache-2.0
task_categories:
- question-answering
language:
- pt
tags:
- medical
- llm
- portuguese
- benchmark
pretty_name: qa_bulas
size_categories:
- n<1K
---
# Medication-Specific QA Benchmark

## Dataset Details

To construct a controlled evaluation benchmark, we selected 25 widely prescribed medications in Brazil across four therapeutic categories: antibiotics, analgesics and anti-inflammatory agents, antihypertensives, and antidiabetics. These categories were chosen to ensure clinical diversity across infectious, inflammatory, cardiovascular, and metabolic conditions.

For each selected medication, we verified the presence of its corresponding leaflet in the cleaned corpus. We then identified 10 standardized sections consistently present across documents, including indications, dosage, contraindications, warnings, adverse reactions, and drug interactions.

Question generation was performed using a large language model conditioned on the medication name and structured leaflet content. To ensure balanced coverage and section-level control, we generated:

- 4 questions per medication per section;
- 40 questions per section across medications;
- 400 total open-ended questions.

This benchmark, referred to as the Medication-Specific QA Benchmark, is explicitly designed to measure factual recall, section-level grounding, and evidence attribution when answers are directly localized within regulatory documents.

## Citation

This work was accepted at **The First Workshop on Language Technologies for Health (Lang4Health)** is a workshop dedicated to the development and application of Natural Language Processing (NLP) technologies in the healthcare field.