MMClima: A Framework for Multimodal Climate Science Data and Evaluation
MMCLIMA is a large-scale benchmark testing AI models across five climate science domains using multimodal data like articles, videos, and figures. It uses 104k expert-validated QA pairs to systematically evaluate language and vision models. The framework ensures reliability through human review and shows that fine-tuned models can outperform general systems on specific climate tasks.