Bridging the domain gap for VLMs: A scientific fine-tuning framework for precision and pedagogy in astronomical image understanding

Vision-language models (VLMs) hold significant potential for educational applications; however, existing research has predominantly relied on the direct invocation of general-purpose architectures, lacking systematic fine-tuning methodologies tailored to discipline-specific visual corpora. This study situates its investigation within the context of space education to explore domain adaptation pathways for multimodal foundation models. A high-quality dataset comprising 9,838 astronomical images accompanied by structured educational texts was constructed: images were sourced from authoritative aerospace platforms and processed through data augmentation; texts were generated via the DeepSeek API within a multi-theoretical educational framework, yielding a four-level hierarchical structure encompassing Description, Explanation, Connection, and Extension. Building upon this dataset, Low-Rank Adaptation (LoRA) was employed to conduct parameter-efficient fine-tuning of the Qwen3-VL-4B-Instruct model. Experimental results demonstrate that the fine-tuned model achieved a substantial improvement in overall test set accuracy from 62.2% to 91.3%, with accuracy on chromatically perturbed and rotation-transformed images reaching 93.1% and 89.4% respectively, evidencing robust visual generalization. Further evaluation indicates that the fine-tuned model outperformed the baseline in 94.4% of assessed samples. A pilot study conducted with 156 seventh-grade students further demonstrated that the fine-tuned model achieved high levels of perceived usability, image recognition accuracy, and learning utility, with positive response rates exceeding 86% across all three dimensions. These findings demonstrate that domain adaptation grounded in pedagogically structured corpora enables models to transcend surface-level visual description, generating explanatory text that is simultaneously disciplinarily precise and pedagogically intentional, thereby providing methodological reference and empirical evidence for advancing AI in education from merely "usable" toward being "trustworthy" and "teachable."