This study examines the effects and limitations of small-scale supervised fine-tuning (SFT) in Korean CSAT-style reading question generation. We propose a two-stage pipeline in which Stage 1 generates expository passages from keywords and Stage 2 generates multiple-choice questions and distractors from the generated passages. To identify stage-specific effects, we separately evaluate passage quality and item quality using a pairwise comparison framework with LLM-based judges and bootstrap-based statistical estimation. Experiments compare Qwen2.5-7B-Instruct and EXAONE 3.5 7.8B under base, QLoRA-based SFT, expanded-training-set, and JSON-output conditions. Results show that small-scale SFT does not affect all stages equally. In the Qwen family, fine-tuned models repeatedly underperform the base model in item-level evaluation, particularly in distractor plausibility and discriminative quality, while passage-level quality shows no consistent decline. This suggests that small-scale SFT may improve surface-level conformity to exam-style formats without improving substantive item quality. In contrast, the EXAONE family shows a different pattern, with some fine-tuned conditions yielding improvement signals in item quality. These findings indicate that the effectiveness of small-scale SFT is stage-dependent and model-dependent, and that passage generation and item generation should be evaluated separately in high-constraint educational generation tasks.