ShilpoBangla: A dataset of Bangladeshi cultural product images with alternative texts in Bangla and English

The lack of high-quality multimodal resources for low-resource languages such as Bangla poses significant challenges for vision-language research, image captioning, and accessibility technologies. To address this gap, this study introduces ShilpoBangla, a curated dataset comprising 1,200 images representing Bengali cultural products across five categories including traditional clothing, food, musical instruments, folk arts and crafts, and jewelry. Each image is paired with human-authored alternative text in both Bangla and English. The study further demonstrates the feasibility of generating image alternative text using a ViT encoder and a GPT-2 decoder, and establishes a benchmark for future multimodal research and cross-linguistic applications. Additionally, the inclusion of alternative text in both Bangla and English enhances the dataset’s accessibility, usability, and adoption among a broader global audience, thereby improving the generalizability and practical applicability of the ShilpoBangla dataset.