Fine-Tuning Transformer Models for Technical Interview Question Answering: A Comparative Study

The study investigates the effectiveness of multiple transformer-based architectures in handling domain-specific question answering tasks relevant to computer science education and technical interview preparation. A dataset consisting of 2,306 question–answer pairs was developed covering key topics including data structures, algorithms, operating systems, databases, and computer networks. The dataset was created using a hybrid methodology combining manual curation and synthetic data generation, followed by a two-stage preprocessing pipeline involving both exact and semantic deduplication to ensure data quality and diversity. Four language models—GPT-2, LLaMA-2 7B, Mistral 7B, and Phi-2—were fine-tuned using the Low-Rank Adaptation (LoRA) technique across multiple dataset sizes (200, 1,000, and 2,306 samples). Model performance was evaluated using established natural language generation metrics, including BLEU-4, ROUGE-L, and BERTScore, to assess both lexical and semantic similarity between generated and reference answers. The results provide a systematic comparison between legacy and modern transformer architectures, demonstrating clear performance advantages of newer instruction-tuned models in domain-specific technical question answering tasks. In particular, Mistral 7B achieved the highest performance across evaluation metrics on the full dataset, followed closely by LLaMA-2 7B and Phi-2, while GPT-2 exhibited comparatively lower performance. This preprint is intended to serve as an initial research contribution and comparative benchmark for fine-tuning language models in specialized technical domains. It may be useful for researchers, students, and educators exploring domain-adapted language models, technical interview preparation systems, and comparative evaluation of transformer architectures.