Optimizing language model fine-tuning and quantization for enhanced medical question answering

Abstract

This paper introduces an end-to-end fine-tuning and compression solution for the EleutherAI GPT-Neo 125 M language model. It employs a new optimization procedure that boosts the performance of the model in the task of medical question answering. The model was seeded with a corpus from the MedQuAD dataset, which consists of 47,457 question–answer pairs retrieved from authorized NIH websites. The MedLine Plus data were, however, excluded due to licensing restrictions. The corpus covers a large range of medical subjects, such as diseases and medicines, and includes different question types and rich annotations for enhancing information retrieval (IR) and natural language processing (NLP) tasks. Fine-tuning was conducted for 3750 training steps, wherein the loss during training continuously dropped from 3.0094 to 0.2343. This shows that the model successfully learned and adapted during the process. The enhanced model incorporated LoRA and 4-bit quantization techniques, specifically intended to improve computational efficiency and model scalability. The measures of performance were evaluated by measuring the BLEU, ROUGE, and TER scores, which were utilized to compare the original and optimized model. The optimized model exhibited considerable improvements in performance metrics. The BLEU value rose from 0.0489 to 0.0706, the ROUGE value rose from 0.0621 to 0.1996, and the TER value fell from 0.9053 to 0.8120. These findings justify the efficacy of fine-tuning and quantization interventions. The results indicate that targeted model optimization can enhance the accuracy and dependability of AI-driven medical question-answering systems quite significantly.

Ähnliche Arbeiten