Preprint releases panel-level materials-science image-text dataset

A new arXiv preprint describes MatMMExtract, an open-source pipeline for splitting compound materials-science figures into annotated sub-panels. Applied to 14,810 open-access articles, the authors say it produced MatSciFig, containing 391,606 image-text pairs from 180,571 figures. The paper also introduces a 2,811-figure detection dataset and reports that a dual-encoder retrieval baseline improved R@1 by 4.4 times over zero-shot CLIP. The resources are released openly, according to the preprint.