DSIQ-GEN: Automatic Generation and Classification of Data Science Interview Questions

Domain-specific question generation and classification systems are increasingly needed in technical fields like data science for education, recruitment, and adaptive learning. However, existing large language models lack domain-specific knowledge and provide limited control over question attributes. This study develops a comprehensive framework for automatic generation and classification of data science interview questions using parameter-efficient fine-tuning (PEFT) methods. We evaluate three PEFT approaches, Low-Rank Adaptation (LoRA), P-tuning, and Prefix tuning, across 105 checkpoints on Llama-3.2-1B. LoRA achieves optimal performance with 85% diversity, 87.1% domain relevance, and only 0.9% trainable parameters. Conditional generation expands the 167-question corpus to 1011 questions (505.4% increase), achieving balanced difficulty distribution and reducing topic standard deviation from 6.9% to 2.5% in the fully merged dataset. Classification experiments show Transformers achieve 87.7% difficulty and 96.7% topic accuracy on extended data, with keyword tokenization (198 tokens) matching full vocabulary (357 tokens) while reducing training time by approximately 15%. Unsupervised clustering validates quality with 77.3% purity (+8.1 points). An independent human evaluation (n = 104 generated questions, 3 raters) corroborates these automatic metrics, yielding substantial inter-rater agreement (weighted κ = 0.602) and directly addressing the risk of circular validation inherent in classifier-only evaluation pipelines. Critical findings reveal conditioning fails below 8%–10% class representation, and topic semantic overlap reduces effectiveness. This research demonstrates PEFT as a practical solution for synthetic data generation in specialized domains under data scarcity constraints.