Fine-Tuning Transformers Model for Question Answering Purpose with Context

Abstract-The practice of fine-tuning large language models for Question answering (QA) has gained significant popularity throughout recent years. We analyze the process of fine-tuning Llama-2-7B-chat through Stanford Question Answering Dataset (SQuAD) in this research. We apply QLoRA (Quantized Low-Rank Adaptation) together with bitsandbytes library for 4-bit quantization to perform parameter-efficient tuning due to limited hardware resources. The system includes stages of data preparation followed by training and model evaluation assessment. The results indicate that transformer fine-tuning produces answers with improved contextual understanding which proves its effectiveness in QA systems. QLoRA enables memory reduction through high accuracy retention which enables model fine-tuning on systems with minimal resources. This research explores both the performance effects and response quality implications of low-rank adaptations in the proposed study to advance NLP-based QA systems.