GlBBQ: Galician Bias Benchmark for Question Answering

Dataset Description

GlBBQ is a Galician adaptation of the BBQ (Bias Benchmark for QA), a benchmark designed to measure social bias in multiple-choice question answering (QA) systems. More specifically, GlBBQ is derived from EsBBQ, the Spanish adaptation of BBQ.

The dataset follows the BBQ framework, where models are evaluated under:

• Ambiguous contexts, where the correct answer is unknown and models should avoid relying on stereotypes
• Disambiguated contexts, where the correct answer is explicitly supported and bias may lead to incorrect predictions

GlBBQ extends this framework to Galician by providing a linguistically adapted version of EsBBQ.

---

Source Data

GlBBQ is derived from:

• EsBBQ (Spanish Bias Benchmark for QA): https://github.com/langtech-bsc/IberBBQ
• Which is itself based on the original BBQ dataset: https://github.com/nyu-mll/BBQ

EsBBQ was manually adapted to the Spanish sociocultural context and includes:

• 27,320 instances
• 323 templates
• 10 social bias categories

GlBBQ is a Galician translation and adaptation of EsBBQ, preserving its structure and evaluation design.

---

Task

Task type: Multiple-choice Question Answering (MCQA)

Each instance consists of:

• A context
• A question
• Three answer options

The task is to select the correct answer given the context.

---

Dataset Structure

Each example includes:

• Context type

  • Ambiguous
  • Disambiguated (stereotypical / anti-stereotypical)

• Question type

  • Negative
  • Non-negative

• Answer options

  • Target group
  • Non-target group
  • Unknown / Not enough information

This structure enables evaluation of both:

• Bias under uncertainty
• Accuracy under bias-sensitive conditions

---

Social Categories

The dataset covers the following 10 social categories:

• Age
• Disability Status
• Gender
• LGBTQIA
• Nationality
• Physical Appearance
• Race / Ethnicity
• Religion
• Socioeconomic Status (SES)
• Spanish Region

---

Intended Use

GlBBQ is intended for:

• Evaluating social bias in language models
• Studying cross-lingual bias transfer
• Benchmarking Galician NLP systems
• Research on fairness in multilingual QA

---

Limitations

• Focuses on stereotype-based bias, not all forms of social harm
• Inherits the sociocultural assumptions of EsBBQ (Spain-specific)
• Does not explicitly model intersectionality
• Translation may introduce minor linguistic artifacts

---

Acknowledgment

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. (Esta publicación del proyecto Desarrollo de Modelos ALIA está financiada por el Ministerio para la Transformación Digital y de la Función Pública y por el Plan de Recuperación, Transformación y Resiliencia – Financiado por la Unión Europea – NextGenerationEU)