Multimodal Coding Dataset (MCD) is a large-scale dataset proposed by Microsoft Research, Peking University and Southern University of Science and Technology and released in 2025. The related paper results are "VisCodex: Unified Multimodal Code Generation via Merging Vision and Coding Models". The dataset contains a total of approximately 598,000 high-quality samples/pairs, organized in an instruction-following format, covering multiple input modalities (text, images, code) and output modalities (code, answers, explanations), and is suitable for multimodal code understanding and generation tasks. The data includes:

• Enhanced HTML code (HTML): about 200,000 code-screenshot pairs, focusing on visual effects and structural optimization.
• Chart: About 210,000 image-code pairs for image-to-code reproduction.
• Question and Answer (QA): About 59,000 code-question-answer pairs, with questions and answers centered around code.
• Algorithm: Approximately 129,000 algorithm coding problems and instruction-following samples.

Citation

@article{jiang2025viscodex, title={VisCodex: Unified Multimodal Code Generation via Merging Vision and Coding Models}, author={Lingjie Jiang and Shaohan Huang and Xun Wu and Yixia Li and Dongdong Zhang and Furu Wei}, journal={arXiv preprint arXiv:2508.09945}, year={2025} }