SDG Classification Using Instruction-Tuned LLMs

Abstract

This paper investigates the potential of quantized, instruction-tuned Large Language Models (LLMs) for zero-shot classification of scientific abstracts according to the United Nations’Sustainable Development Goals (SDGs). Weintroduce the Decompose-Synthesize-RefineExtract (DSRE) framework, leveraging advanced prompting techniques for both singlelabel and multi-label classification scenarios.DSRE is designed to enhance the zero-shotcapabilities of LLMs for this domain-specifictask. We explore the trade-offs between modelperformance and computational efficiency introduced by quantization. The performance ofDSRE and quantized LLMs is benchmarkedagainst fine-tuned LLM baselines and the Aurora system. Our findings demonstrate the potential of instruction-tuned LLMs for zero-shotSDG classification but emphasize the continuedvalue of fine-tuning for optimal performance.Additionally, we consider dataset imbalanceand the impact of augmenting datasets.