NRITYAM: Language Models Meet Art and Heritage of Dance

Researchers have introduced NRITYAM, a comprehensive multilingual benchmark dataset containing 9,260 question-answer pairs across 12 languages designed to evaluate how well language models understand global dance traditions and cultural heritage. Developed in collaboration with native dance artists and speakers, the dataset addresses a critical gap in AI evaluation by testing cultural comprehension beyond Western-centric knowledge, establishing new standards for assessing AI systems' ability to reason about traditional performing arts.

NRITYAM represents a meaningful shift in how the AI research community approaches cultural evaluation benchmarks. Rather than relying on generic knowledge datasets that inherently favor Western cultural perspectives, this project deliberately centers non-Western artistic traditions by embedding domain expertise from native practitioners into dataset creation. This methodology ensures that cultural nuance isn't lost in translation or abstraction, addressing a structural weakness in current language model evaluation frameworks.

The benchmark's construction through collaboration with dance artists and native speakers reflects growing recognition that AI systems trained on internet-scale text often embed cultural biases that disadvantage non-English, non-Western knowledge domains. Dance traditions carry significant cultural weight in many societies, yet remain underrepresented in standard NLP evaluation sets. By creating the largest dedicated dataset for evaluating cultural knowledge in dance, NRITYAM provides researchers with concrete tools to measure and improve this gap.

For the AI industry, this work signals that cultural competence is becoming a measurable, improvable metric rather than an afterthought. Organizations deploying language models globally will increasingly face pressure to demonstrate performance across culturally specific domains. The 12-language, multilingual structure enables cross-cultural comparison, allowing researchers to identify which models genuinely understand cultural context versus those merely pattern-matching on surface-level features.

Future developments will likely see similar specialized benchmarks for other cultural domains—cuisine, textiles, music, ritual practices—creating a comprehensive ecosystem for evaluating genuine cross-cultural AI understanding. The open-source dataset release encourages adoption and reproducibility, accelerating progress toward more culturally inclusive AI systems.

• →NRITYAM comprises 9,260 curated question-answer pairs across 12 languages, making it the largest dance-focused cultural knowledge dataset for language model evaluation.
• →The dataset was developed through direct collaboration with native dance artists and native speakers, ensuring authentic cultural representation rather than external interpretation.
• →Current language models show varying performance on cultural comprehension tasks, revealing systematic gaps in understanding non-Western traditional performing arts.
• →The benchmark applies to multiple model categories including large and small language models, both multimodal and text-only variants, providing comprehensive evaluation scope.
• →Open-source availability of NRITYAM enables broader AI research community participation in developing more culturally competent language models globally.