Leveraging Large Language Models to Build Computationally Efficient Models for Sustainable Finance Investment Decision Support

Assessing companies' contributions to the United Nations Sustainable Development Goals (SDGs) is essential for sustainable investment and regulatory reporting. However, extracting reliable insights from heterogeneous textual sources remains a challenge due to limited labeled data, domain imbalance, and privacy constraints. We present LëtzSDG, a lightweight BERT-based multiclass classifier fine-tuned on a hybrid dataset constructed using Large Language Models (LLMs). Multiple LLMs are used to (i) expand domain-specific SDG keywords, (ii) perform consensus-based zero-shot labeling, and (iii) generate synthetic data to balance underrepresented classes. Unlike cloudhosted LLMs, LëtzSDG is designed for on-premises deployment within financial institutions, ensuring data-privacy compliance. Integrated into a human-in-the-loop investment workflow, its predictions are span-linked for traceability and committee review. Evaluated on public datasets (the OSDG Community Dataset and the SDG Integration Corpus), LëtzSDG outperforms SDGspecific baselines, a strong NLI-based zero-shot model, and several open LLMs, while rivaling larger closed models at a fraction of their size. LëtzSDG and its datasets are publicly available on Hugging Face.