Benchmark dataset of "Multimodal AI agents for capturing and sharing laboratory practice"

The benchmark dataset was created to evaluate the performance of our proteomics lab agent on two key tasks: generating protocols and detecting experimental errors. It consists of videos of a scientist performing a diverse range of proteomics workflows, from simple pipetting to complex operations involving specialized mass spectrometry equipment and software. The dataset contains two categories of videos: Tutorial recordings of a researcher correctly performing and explaining a procedure, which serve as the basis for testing AI-assisted protocol generation. Recordings from experiments in which a scientist intentionally made mistakes to test the agent's error-detection capabilities.