Exploiting Classification of LLM with Zero-shot Prompt & Fine-tuning

Large Language Models (LLMs) have great potential in medical text classifications. However, LLMs are consuming huge of computing resources to architect the state-of-the-art performance that difficult to employ on medical edge device. Recently advance is that Tiny-scale LLM which give a chance that no need for huge computing resource to inference. In this work, we illustrate the classification performance that use 3 different zero-shot like prompt schemes considered to use the Llama-2-42M. And then, we froze the pertained weight of LLM and finetune the classify_heads which includes normal linear layer and deep neural network. The experimental results are evaluated in Accuracy and F1-score in 2 real-world datasets. Our findings reveal that carefully crafted longer prompts significantly improve zero-shot classification performance across datasets of different scales. Moreover, tailored fine-tuning schemes enhance classification performance, with deeper classification head structures and optimized activation functions further boosting results.