Replication Data for: RoSci-Video: Benchmarking Robust Scientific Understanding and Reasoning on Experimental Videos

RoSci-Video is a benchmark for evaluating large multimodal model (LMM) robustness to information degradation in laboratory experimental videos. The dataset covers 390 videos across 13 scientific disciplines, with two types of perturbations: Information Error and Information Loss. It is intended for use in assessing whether models can recognize degraded or conflicting multimodal information rather than producing overconfident incorrect answers. This repository contains the video files (MP4) and question-answer annotation files (JSONL).