BSC-LT/piqa_es

Dataset Card for PIQA (Spanish Version)

Dataset summary

This dataset provides the Spanish translation and adaptation of the validation set of PIQA (Physical Interaction: Question Answering). The original dataset was designed to evaluate physical commonsense reasoning in language models through questions about everyday situations. Each example presents a physical goal and two possible solutions, only one of which is correct.

This Spanish adaptation enables evaluation of physical commonsense reasoning in Spanish while preserving the conceptual difficulty, structure, and intent of the source material.

• Curated by: Barcelona Supercomputing Center (BSC)
• Funded by: Modelos del Lenguaje
• Language(s) (NLP): es (Spanish)
• License: Academic Free License v3.0 (Original)

Dataset Details

Dataset Description

The dataset preserves the structure, features, and overall format of the original PIQA validation set. It contains 1,838 examples. Each example consists of a physical goal, two candidate solutions, and an integer identifying the correct solution.

The Spanish translation and adaptation sought to:

1. Preserve linguistic and cultural appropriateness in Spanish-speaking contexts.
2. Maintain the meaning and physical plausibility of the original scenarios.
3. Preserve the correct answer and the relative difficulty of each example.
4. Adapt proper names, measurements, and culturally specific references when needed.
5. Correct source-text issues when necessary without altering the intended task.
6. Avoid linguistic patterns that could reveal the correct answer.

Dataset Structure

PIQA_es

An example instance looks as follows:

{
  "goal": "Un pelador",
  "sol1_es": "puede usarse como decoración encima de un televisor.",
  "sol2_es": "puede usarse como decoración en un calcetín.",
  "answerKey": 0
}

Each row contains the following fields:

• goal: a physical goal, expressed as a question or instruction.
• sol1_es: the first candidate solution in Spanish.
• sol2_es: the second candidate solution in Spanish.
• answerKey: the index of the correct solution: 0 for sol1_es and 1 for sol2_es.

The dataset contains only the validation split and is distributed in Apache Arrow format.

Dataset Sources

• Repository: PIQA
• Leaderboard: Physical Interaction: Question Answering
• Paper: Bisk, Y. et al. (2020). PIQA: Reasoning about Physical Commonsense in Natural Language. arXiv:1911.11641

Who are the source data producers?

All credits for the original PIQA dataset go to its creators, Yonatan Bisk et al.

Citation

BibTeX:

@inproceedings{bisk2020piqa,
  title={PIQA: Reasoning about Physical Commonsense in Natural Language},
  author={Bisk, Yonatan and Zellers, Rowan and Le Bras, Ronan and Gao, Jianfeng and Choi, Yejin},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={34},
  number={05},
  pages={7432--7439},
  year={2020}
}

APA:

Bisk, Y., Zellers, R., Le Bras, R., Gao, J., & Choi, Y. (2020). PIQA: Reasoning about Physical Commonsense in Natural Language. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05), 7432–7439.

Uses

Direct Use

This dataset can be used to:

• Evaluate physical commonsense reasoning and multiple-choice question answering in Spanish.
• Study multilingual performance variation in physical reasoning tasks.
• Assess model robustness on practical, everyday physical scenarios in a non-English context.

Out-of-Scope Use

This dataset is not intended for:

• Training or fine-tuning models, as it is intended strictly for evaluation.
• Treating model performance as evidence of safe or reliable behavior in real-world physical environments.
• Use in non-Spanish contexts, as the dataset has been linguistically and culturally adapted.

Additional Information

Creators

Language Modeling and Annotation Teams, AI Institute, the Barcelona Supercomputing Center (ai_institute_languagemodeling@bsc.es) (ai_institute_annotators@bsc.es).

Funding

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.

This work has been promoted and supported by the Government of Catalonia through the Aina Project.

Acknowledgements

We acknowledge EuroHPC Joint Undertaking for awarding the project ID EHPC-AI-2024A05-046 access to MareNostrum5 at BSC, Spain.

Licensing Information

This work is licensed under Academic Free License v3.0 (Original)