TerretaQA

TerretaQA is a dataset of open-domain question answering (QA) questions in Valencian, designed to evaluate the cultural knowledge of language models about the Valencian Community (Comunitat Valenciana).

Dataset Description

TerretaQA contains open-ended questions about the cultural, historical, geographical, and tourism-related knowledge of the Valencian Community.

The questions were automatically generated from PDF tourism brochures about the Valencian Community. The generated questions and answers were subsequently post-processed to align to Valencian using an automatic translation system.

After the automatic generation and translation process, a subset of the dataset was manually reviewed by expert linguists. The reviewed examples were checked for linguistic quality and correctness.

The dataset therefore contains both manually reviewed and not manually reviewed examples.

Dataset Structure

The dataset is divided into four splits:

Split	Number of examples	Description
train	1,839	Automatically generated and manually reviewed examples for training
test	200	Automatically generated and manually reviewed examples for evaluation
train_not_checked	16,890	Automatically generated examples that have not been manually reviewed
test_not_checked	800	Automatically generated examples that have not been manually reviewed

Features

All splits contain the following features:

Feature	Description
id	Unique identifier of the example
seed_1	First seed/context used during question generation
seed_2	Second seed/context used during question generation
location	Location associated with the question
input	Question in Valencian
target	Expected answer in Valencian

Data Generation

The dataset was created using an automatic pipeline based on tourism-related PDF brochures from the Valencian Community.

The general process was:

Tourism brochures (PDF)
        │
        ▼
Information extraction
        │
        ▼
Automatic QA generation
        │
        ▼
Automatic translation to Valencian
        │
        ▼
┌───────────────────────────────┐
│                               │
▼                               ▼
Manual linguistic review        Not manually reviewed
│                               │
▼                               ▼
train / test                    train_not_checked / test_not_checked

The manually reviewed subset was checked by expert linguists to improve the reliability of the questions and answers and to identify linguistic or content-related issues.

The *_not_checked splits contain automatically generated examples that have not undergone manual linguistic validation.

Recommended Usage

The manually reviewed splits are the recommended splits for benchmark evaluation:

• train: training data
• test: evaluation data

The train_not_checked and test_not_checked splits can be used for:

• Additional training data
• Data augmentation
• Exploring automatically generated QA pairs
• Studying the effect of data quality on model performance
• Future manual validation

Because the *_not_checked examples have not been manually reviewed, they should be used with appropriate caution for benchmark evaluation.

Loading the Dataset

The dataset can be loaded using the 🤗 Datasets library:

from datasets import load_dataset

dataset = load_dataset("gplsi/TerretaQA")

print(dataset)

This returns:

DatasetDict({
    train: Dataset({
        features: ['id', 'seed_1', 'seed_2', 'location', 'input', 'target'],
        num_rows: 1839
    })
    test: Dataset({
        features: ['id', 'seed_1', 'seed_2', 'location', 'input', 'target'],
        num_rows: 200
    })
    train_not_checked: Dataset({
        features: ['id', 'seed_1', 'seed_2', 'location', 'input', 'target'],
        num_rows: 16890
    })
    test_not_checked: Dataset({
        features: ['id', 'seed_1', 'seed_2', 'location', 'input', 'target'],
        num_rows: 800
    })
})

An individual example can be accessed with:

example = dataset["test"][0]

print("Question:", example["input"])
print("Answer:", example["target"])

Language

The dataset is written in Valencian (valencià).

It is specifically intended to evaluate the ability of language models to answer questions about the cultural knowledge of the Valencian Community in Valencian.

Intended Use

TerretaQA is intended for:

• Evaluation of language models' cultural knowledge
• Valencian-language QA research
• Evaluation of multilingual and regional-language models
• Research into cultural and geographical knowledge representation
• Development and evaluation of retrieval-augmented generation (RAG) systems

Citation

If you use TerretaQA in your research, please cite:

@dataset{terretaqa,
  title = {TerretaQA},
  author = {Martínez-Murillo, Iván and Grande Ruíz, Eduardo and Miró Maestre, María and Galiano-Segura, Santiago and Sepúlveda-Torres, Robiert },
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/gplsi/TerretaQA}
}

Acknowledgements

We thank the linguists and experts who contributed to the manual review of the dataset and helped improve the linguistic quality of the manually checked subset.

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project Desarrollo de Modelos ALIA.