{"id":"doi:10.3390/w16213075","kind":"article","title":"WaterGPT: Training a Large Language Model to Become a Hydrology Expert","authors":["Yi Ren","Tianyi Zhang","Xurong Dong","Weibin Li","Zhiyang Wang","Jie He","Hanzhi Zhang","Licheng Jiao"],"author_count":8,"year":2024,"published":"2024-10-27","venue":"Water","cited_by":56,"doi":"10.3390/w16213075","arxiv_id":null,"nct_id":null,"source_url":"https://doi.org/10.3390/w16213075","archive_url":"https://space-kitty.com/papers/doi:10.3390/w16213075","updated":null,"pmid":null,"pmcid":null,"openalex_id":"W4403842350","pdf_url":"https://www.mdpi.com/2073-4441/16/21/3075/pdf?version=1730020580","abstract":"Recently, large language models have been extensively applied across various fields, yet no research has focused on applying these models specifically in the field of water resources. This paper proposes a large language model tailored for complex multimodal tasks in the hydrology domain-WaterGPT. The model undergoes a complete training process, including incremental pre-training and supervised fine-tuning. One core aspect of WaterGPT involves the meticulous segmentation of training data for the supervised fine-tuning phase, which is sourced from real-world data and annotated with high quality using both manual and GPT-series model annotations. This data is carefully categorized into four types: knowledge-based, task-oriented, negative samples, and multi-turn dialogues. Additionally, another core component is the development of a multi-agent framework called Water_Agent. This framework is divided into development and deployment phases. During the development phase, various tool codes are written, and in the deployment phase, the framework intelligently calls upon these tools to complete complex tasks. This framework can handle multimodal data, including text and images, allowing for a deep understanding and analysis of complex hydrological environments. Through the detailed analysis of image data, the framework can accurately extract water bodies and calculate their area, demonstrating WaterGPT's capability to handle complex multimodal tasks. Moreover, we have constructed a high-quality water resources evaluation dataset, EvalWater, which covers 21 categories and approximately 10,000 questions. Using this dataset, WaterGPT has achieved the highest accuracy to date in the field of water resources, reaching 83.09%, which is about 17.83 points higher than GPT-4. Our code is open-sourced at https://github.com/renllll/watergpt.","abstract_license":"CC-BY (publisher license, via OpenAlex)","license":"cc-by","is_open_access":true,"retracted":false,"peer_reviewed":null,"categories":["Hydrological Forecasting Using AI"],"details":null,"provenance":"openalex","attribution":"Abstract of \"WaterGPT: Training a Large Language Model to Become a Hydrology Expert\" by Yi Ren, Tianyi Zhang, Xurong Dong et al., https://doi.org/10.3390/w16213075, used under its Creative Commons Attribution license (CC-BY); markup removed, otherwise unmodified.","topics":[{"slug":"ai-agents","name":"LLM Agents & Tool Use","domain_slug":"ai"}],"cited_in_dossiers":[]}