BSC-LT/SIQA_es

Dataset Card for SIQA (Spanish Version)

Dataset summary

This dataset provides the Spanish translation and adaptation of the SIQA (Social Interaction Question Answering) validation set. The original dataset was designed to evaluate social commonsense reasoning in LLMs by presenting a collection of questions based on everyday social situations, with the ultimate goal of challenging models to infer motivations, reactions and social implications behind human actions. This Spanish adaptation enables evaluation of the SIQA task in Spanish while preserving the original Easy and Challenge splits and maintaining 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: CC BY 4.0

Dataset Details

Dataset Description

The dataset preserves the original structure, features, and overall format of the original SIQA validation set. Only minor adjustments were made during translation. Key adaptations include:

1. Preserving linguistic and cultural appropriateness in Spanish-speaking contexts.
2. Fixing incorrect labels that did not correspond to the correct answer choice in the original dataset.
3. Replacing some standard original questions with anothers more suitable to the context and the questions.
4. Correcting names within questions when needed to maintain internal consistency with the scenario.
5. Adapting proper names to Spanish equivalents.
6. Maintaning gender neutrality in items where the English source did not mark gender, in order to avoid introducing gender bias.

Dataset Structure

SIQA_es

An example of instance looks as follows:

{
  "index": "454",
  "context": "Teo puso en duda la historia que se enseñaba en los libros.",
  "question": "¿Qué debe hacer Teo antes de esto?",
  "answer_a": "Investigar.",
  "answer_ b": "Pensar que los libros están equivocados.",
  "answer_c": "Demostrar su teoría.",
  "label": "1"
}

Dataset Sources

• Repository: SIQA
• Paper: Sap, M. et al. (2019). SocialIQA: Commonsense Reasoning about Social Interactions. arXiv:1904.09728

Who are the source data producers?

All credits go to the creators of the original SIQA dataset, Maarten Sap et al.

Citation

APA:

Sap, M. et al. (2019). SocialIQA: Commonsense Reasoning about Social Interactions. arXiv:1904.09728

Uses

Direct Use

This dataset can be used to:

• Evaluate social commonsense reasoning and multiple-choice question-answering in Spanish.
• Study multilingual performance variation in social reasoning tasks.
• Assess model robustness on human interaction interpretation 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.
• Drawing conclusions about real-world social behavior.
• Use in non-Spanish contexts, as the dataset has been culturally and linguistically localized.

More information

This work has been supported and funded by the Ministerio para la Transformación Digital y de la Función Pública and the Plan de Recuperación, Transformación y Resiliencia – funded by the EU through NextGenerationEU, within the framework of the Modelos del Lenguaje project.

Contact point

Language Technologies Lab (langtech@bsc.es) at the Barcelona Supercomputing Center (BSC).