The educational applications of Vision-Language Models (VLMs), such as OpenAI GPT-5, have grown rapidly in recent years. However, when applied to assist learning for visually impaired students, most existing VLMs demonstrate a critical limitation — they lack understanding of Braille and tactile graphics. As a result, they often fail to provide meaningful assistance or feedback, and in some cases even generate misleading responses. This gap highlights a severe inequality in educational accessibility for the visually impaired, contradicting the goals of the United Nations Sustainable Development Goals (SDGs). To address the scarcity of domain-specific VLMs and training data for the visually impaired, this study proposes a novel Synthetic Data Augmentation (SDA) method tailored for the blind education domain. The proposed pipeline integrates easily accessible datasets (such as SQuAD) with large-scale VLM-assisted synthesis to efficiently generate high quality, generalized, and low-cost training data — significantly reducing the resources required compared to traditional methods. Building upon Google DeepMind’s state-of-the-art lightweight open-source model Gemma 3–4B, we integrate SigLip ViT and apply LoRA supervised fine-tuning using the synthesized dataset on a standard personal computer. Experimental results show that our model surpasses GPT-5 and previous Braille-related VLM studies, achieving 94.3% in Braille transcription, 74.2% in blind exam exact match (EM), 87.2% in tactile diagram captioning, and human expert evaluation scores of 6.53/7 and 6.87/7. These results demonstrate the model’s stability and strong visual-assistance capabilities under generalized evaluation data. Furthermore, with only 36% of the parameters compared to prior models, its lightweight architecture enables practical deployment on personal devices such as smartphones and laptops, expanding the accessibility of AI-powered educational tools for the visually impaired worldwide.