LARI Dataset: A Native Portuguese Question Answering Dataset from Brasileiras em PLN

2026-04-12 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, medium

Summary

The LARI dataset, introduced at PROPOR 2026, is a new resource for benchmarking and enhancing Question Answering (QA) in Portuguese, addressing a critical lack of native training data for the language. Developed by Júlia da Rocha Junqueira, Larissa A. de Freitas, and Ulisses Brisolara Corrêa, the dataset was created using a methodology that fine-tuned the Sabiá-7B model via QLoRA on a domain-specific corpus. Content for the dataset was extracted from the book "Natural Language Processing – Concepts, Techniques, and Applications in Portuguese (2nd Edition)". The generated context-question-answer triples underwent expert human evaluation, achieving an average quality score of 4.47 out of 5. Comprising 464 triples, the LARI dataset is publicly available, aiming to support future research in low-resource language settings.

Key takeaway

For research scientists developing Question Answering systems for low-resource languages, the LARI dataset offers a valuable, human-validated resource specifically for Portuguese. You should consider integrating this publicly available dataset into your training and evaluation pipelines to improve model performance and benchmark against a high-quality, native language standard. This can accelerate progress in Portuguese NLP, reducing reliance on translated or less relevant data.

Key insights

LARI dataset provides a human-validated Portuguese QA resource, addressing a critical language resource gap.

Principles

• Human validation improves dataset quality.
• Domain-specific fine-tuning enhances model performance.

Method

The methodology involved fine-tuning the Sabiá-7B model with QLoRA on a domain-specific corpus, extracting content from a Portuguese NLP textbook, and then performing expert human validation on the generated QA instances.

In practice

• Use Sabiá-7B with QLoRA for Portuguese NLP tasks.
• Leverage LARI for Portuguese QA model benchmarking.

Topics

• LARI Dataset
• Portuguese Question Answering
• Sabiá-7B Model
• QLoRA
• Low-Resource NLP

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• The Inadequacy of Automatic Evaluation Metrics in Question Answering: A Case-Study in Portuguese — Traditional lexical metrics are insufficient for evaluating Question Answering quality, especially in non-English languages.
• Negation-Aware Data Augmentation for Portuguese Natural Language Inference — Targeted negation-aware data augmentation improves BERT-based NLI model performance on Portuguese datasets by reducing negation bias.
• Deep text-pair classification with Quora's 2017 question dataset — Quora's 2017 dataset of 400,000 question pairs enables deep learning solutions for text-pair semantic equivalence tasks.
• NCTB-QA: A Large-Scale Bangla Educational Question Answering Dataset and Benchmarking Performance — Domain-specific fine-tuning is crucial for robust QA performance in low-resource languages, especially with unanswerable questions.
• I spent 4 days training an AI model to understand my country. It barely works — Low-resource NLP projects face significant challenges in data collection, translation, and accurate labeling, often yielding modest baselines.
• ODTQA-FoRe: An Open-Domain Tabular Question Answering Dataset for Future Data Forecasting and Reasoning — A new dataset and LLM agent framework enable future-oriented tabular question answering by integrating time-series forecasting.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.