Resources Construction and LLM Fine-tuning for Education of Computer Science

The limited inference capabilities of large language models in the field of computer science significantly hinder their application in computer education question-answering tasks. To address this issue, this paper focuses on three key areas: constructing domain-specific datasets, enhancing these datasets, and fine-tuning large language models to improve their performance in computer education question answering. The foundational dataset is sourced from university and online question banks, consisting of over 1600 entries, including 678 multiple-choice questions, 502 true/false questions, and 509 fill-in-the-blank questions. These questions span 11 disciplines, such as artificial intelligence, software engineering, and digital logic. Additionally, this study uses a large model for data augmentation, expanding the scope and scale of the base dataset. The augmented content is sourced from Baidu Baike and GPT-generated text, ultimately increasing the dataset to over 10000 entries. Based on this expanded dataset, the study employs LoRA technology to fine-tune the LLaMA2-7B model. Experimental results demonstrate that the performance of the fine-tuned model significantly outperforms other baselines across various task types.