DuoWikiBias: A Spanish Wikipedia Corpus for Bias Detection

DuoWikiBias is a novel parallel corpus derived from Wikipedia for Spanish bias classification. The dataset was created by Karla Salas-Jimenez and last updated on 2026-05-13. It is used to evaluate Large Language Models and classical approaches for detecting framing, epistemological, and demographic biases.

Use Cases

• Evaluating LLM performance on bias detection tasks based on the described framing, epistemological, and demographic bias labels.
• Training classifiers for Spanish bias detection based on the annotated Wikipedia-derived corpus.
• Benchmarking advanced prompting techniques like CARP and Metacognition against traditional methods like logistic regression with S-BERT embeddings.
• Refining model attention for bias-aware representation learning using the described Gradient Ascent unlearning method.

Strengths

• The dataset addresses a noted scarcity of annotated Spanish-language resources for bias detection.
• It provides a validated methodological framework for bias detection in Spanish NLP, achieving a reported best F1 score of 0.796.
• The dataset is licensed under CC-BY-4.0, facilitating open use and sharing.

Limitations

• The dataset is small in scale, with a file size of 137.0 KB.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

Derived from Wikipedia.

Collection Method

Likely involves extracting and annotating parallel text segments for bias classification.

Freshness

Last updated 2026-05-13 05:34:02.