CONAN-SP

CONAN-SP is a dataset for automatic counter-narrative generation in Spanish, developed by researchers at the Universidad de Jaén (SINAI research group). It provides pairs of hate speech comments (HS) and their corresponding counter-narratives (CN), covering five hate targets: islamophobia, misogyny, antisemitism, racism, and homophobia.The dataset is built upon CONAN-KN (Chung et al., 2021), an English HS-CN dataset of 195 pairs. The construction pipeline involved automatic translation into Spanish via DeepL, followed by counter-narrative generation using GPT-3.5 under three different prompting strategies — a general prompt with five examples, five target-specific prompts, and a general prompt without task definition. After removing duplicates and annotation agreement examples, the final dataset contains 238 HS-CN pairs distributed across three experiments (84, 70, and 84 instances respectively). All pairs were labelled by human annotators according to three quality metrics: Offensiveness, Stance, and Informativeness.The dataset is distributed as Excel files, one per experiment, and is intended for use in training and evaluating NLP systems focused on counter-speech, hate speech mitigation, and generative language modelling in Spanish.