Optimizing Esg Evaluation In Energy Companies Utilizing Supervised Fine-Tuned Llms

This study presents a supervised fine-tuning framework for large language models to enhance the evaluation of ESG performance in energy companies. By integrating diverse data sources—including financial news, annual reports, and Refinitiv ESG ratings—we construct a comprehensive dataset for model training. Three mainstream models (ChatGPT-4o-mini, LLaMA, and Qwen) are fine-tuned using tailored methodologies to extract and quantify ESG factors from textual disclosures. Comparative experiments demonstrate that the fine-tuned ChatGPT model achieves the closest alignment with benchmark ESG scores, evidenced by lower mean squared error and mean deviation, as well as higher accuracy and correlation. Our approach is based on a robust benchmark and develops supervised fine-tuned models using LLMs specifically for ESG evaluation. It provides a real-time, transparent, efficient, and accurate tool for assessing the ESG performance of energy companies, thereby offering valuable support for informed investment decisions, sustainable development strategies, and subsequent financial relevance analyses in the energy sector.