lukasArana/A-OK-VQA-eu

A-OK-VQA-eu (Basque Translation • 5 K Sample)

📚 Overview

A-OK-VQA-eu is a Basque-language subset of the original A-OKVQA knowledge-based visual question-answering benchmark. A random sample of 5 000 English QA pairs was translated into Basque with HiTZ/Latxa-Llama-3.1-70B-Instruct; roughly 20 % of those translations were manually post-edited to ensure fluency and adequacy.

Important: This is not the official dataset. It is an independent community translation intended to support Basque-speaking researchers and practitioners.

✍️ Authors & Acknowledgements

• Original dataset: A-OKVQA — © 2022 A-OKVQA Team
• Basque translation & curation: <Your Name / HiTZ>, 2025
  • Automatic translation with Latxa-Llama-70B
  • Manual review of ≈ 1 000 samples

If you use this Basque split, please cite both the original A-OKVQA dataset and this translation (see Citation section).

The JSONL schema mirrors the English source:

1. image: Image related to the instruction
2. question: Question related to the image
3. choices: Possible choices for the multiple choice task
4. correct_choice_idx: Index of the correct answer from the choices
5. question_id: Unique identifier for the sample

🔧 How We Built It

1. Sample selection – Uniform random sample of 5 000 QA pairs from the official training split
2. Machine translation – HiTZ/Latxa-Llama-3.1-70B-Instruct

No images were added or removed.

🚦 Limitations & Ethical Considerations

• Non-official: The A-OKVQA team has not reviewed or endorsed this edition; meaning shifts may remain.
• Sampling bias: Only 5 000 of the original examples are included; results may not generalize to the full benchmark.
• Model biases: Latxa-Llama-70B can introduce or amplify biases present in either the source data or the MT system.

💻 Quick Start

from datasets import load_dataset

ds = load_dataset(
"your-hf-org/aokvqa-eu",
split="train"
)

All fields mirror the English original; only textual content is localized.

📜 License

This translation inherits the same license as the upstream A-OKVQA dataset (CC-BY-SA-4.0). By downloading or using these files, you agree to comply with that license, including proper attribution to both the A-OKVQA authors and the Basque translation contributors.

🏷️ Citation

@misc{schwenk2022aokvqabenchmarkvisualquestion, title={A-OKVQA: A Benchmark for Visual Question Answering using World Knowledge}, author={Dustin Schwenk and Apoorv Khandelwal and Christopher Clark and Kenneth Marino and Roozbeh Mottaghi}, year={2022}, eprint={2206.01718}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2206.01718}, }

dataset_info: features: - name: image dtype: image - name: question dtype: string - name: choices dtype: string - name: correct_choice_idx dtype: string - name: question_id dtype: string splits: - name: en num_bytes: 134122888.485 num_examples: 2501 - name: eu num_bytes: 133968224.508 num_examples: 2499 download_size: 269240884 dataset_size: 268091112.993 configs: - config_name: default data_files: - split: en path: data/en-* - split: eu path: data/eu-*