Advancements in large language models (LLMs) have revolutionized natural language processing, enabling these models to perform a wide range of tasks across numerous languages, often overcoming linguistic barriers even for unsupported languages. The blog post discusses the challenges and solutions for creating fair, transparent, and comprehensive multilingual evaluations for such models. It highlights the limitations of existing benchmarks, which often reflect Western-centric perspectives due to their reliance on English translations, leading to cultural erasure and biases. The post emphasizes the need for authentic, human-verified multilingual datasets and suggests participatory approaches involving native speakers for culturally sensitive and accurate evaluations. It introduces two initiatives, SEA-HELM and Aya, focusing on improving multilingual and multicultural evaluations, and stresses the importance of transparency in language support and fair aggregation of evaluation metrics. Additionally, it suggests utilizing LLM judges for scalable evaluations while acknowledging their limitations compared to human assessments. The authors advocate for collaborations with native communities to ensure authentic data representation and call for greater interaction between users and leaderboard managers to enhance evaluation inclusivity and relevance.