SplaXBERT: Optimizing QA for Long Texts

• Goal: Develop a high-efficiency question-answering (QA) model using ALBERT-xlarge
• Techniques: Mixed precision training, context splitting
• Performance: 85.95% Exact Match, 92.97% F1 Score on SQuAD v1.1

SplaXBERT is a novel QA system built on ALBERT-xlarge, designed to enhance answer retrieval efficiency on long texts. By integrating mixed precision training and context splitting, SplaXBERT achieves high accuracy while reducing computational costs. The model has been extensively tested on SQuAD v1.1, demonstrating superior performance over traditional BERT-based approaches.

This project is conducted at NUS, focusing on optimizing model efficiency for real-world applications in question-answering tasks.