Fine-tuning large language models for interdisciplinary environmental challenges.

Area of Science:

• Environmental Science
• Artificial Intelligence
• Computational Science

Background:

• Large language models (LLMs) offer advanced reasoning but face challenges in specialized fields like environmental science due to jargon and data heterogeneity.
• Existing LLM frameworks lack domain-specific training data and evaluation benchmarks for comprehensive environmental applications.

Purpose of the Study:

• To introduce a unified pipeline for developing and evaluating LLMs tailored to environmental science.
• To create a domain-specific dataset and benchmark for training and assessing LLM performance in environmental contexts.

Main Methods:

• Developed EnvInstruct, a multi-agent system for prompt generation.
• Created ChatEnv, a 100-million-token instruction dataset covering five environmental themes.
• Established EnvBench, a 4998-item benchmark for evaluating LLM analysis, reasoning, calculation, and description.
• Fine-tuned an 8-billion-parameter model, EnvGPT, using the developed pipeline.

Main Results:

• EnvGPT achieved 92.06 ± 1.85% accuracy on the EnviroExam benchmark, outperforming comparable and larger models.
• EnvGPT surpassed baselines on EnvBench across relevance, factuality, completeness, and style metrics.
• The fine-tuning approach demonstrated the capability of compact LLMs to reach state-of-the-art performance with curated domain data.

Conclusions:

• Targeted fine-tuning on domain-specific data significantly enhances LLM capabilities in environmental science.
• The released EnvGPT, ChatEnv, and EnvBench provide a reproducible foundation for advancing LLM adoption in environmental research, policy, and practice.
• The pipeline offers a scalable approach for developing specialized LLMs, with potential for multimodal and real-time applications.