Can large language models effectively reason about adverse weather conditions?

This paper seeks to answer the question “can Large Language Models (LLMs) effectively reason about adverse weather conditions?”. To address this question, we utilized multiple LLMs to harness the US National Weather Service (NWS) flood report data spanning from June 2005 to September 2024. Bidirectional and Auto-Regressive Transformer (BART), Bidirectional Encoder Representations from Transformers (BERT), Large Language Model Meta AI (LLaMA-2), LLaMA-3, and LLaMA-3.1 were employed to categorize data based on predefined labels. The methodology was implemented in Charleston County, South Carolina, USA. Extreme events were unevenly distributed across the training period with the “Cyclonic” category exhibiting significantly fewer instances compared to the “Flood” and “Thunderstorm” categories. Analysis suggests that the LLaMA-3 reached its peak performance at 60% of the dataset size while other LLMs achieved peak performance at approximately 80–100% of the dataset size. This study provided deep insights into the application of LLMs in reasoning adverse weather conditions. • LLMs are robust in classifying NWS flood report text data • S even transformer-based LLMs including, BART, BERT, LLaMA-2, and LLaMA-3 are compared. • LoRA-based fine-tuning achieves more consistent and robust results across all categories compared to few-shot learning. • LLaMA-3.1 and LLaMA-3 provided strong trade-offs in performance, runtime, and fewer epochs to reach the best F1 score.