This project is about using Natural Language Processing (NLP) to automatically create quizzes that check reading comprehension. The idea is to generate both questions and answers directly from a passage of text. To make the models efficient, we use Low-Rank Adaptation (LoRA) and Quantized LoRA (QLoRA). These techniques reduce the memory and computing power needed while keeping good performance.

We fine-tuned and tested three models: FLAN-T5-XL, LLaMA-3.2-3B-Instruct, and Phi-2. Training was done on the SQuAD dataset, and testing was done on SQuADShifts, which includes data from news, Wikipedia, Reddit, and Amazon reviews.

Experiments compared full fine-tuning and QLoRA. We used automatic metrics like QAAlignedF1, BERTScore, and MoverScore, and also did manual evaluation to check fluency, relevance, and correctness. The results showed that QLoRA can achieve nearly the same quality as full fine-tuning but with much lower memory cost.