Representation-Derived Multimodal Latent-State Benchmark Dataset

This dataset contains the Representation-Derived Benchmark Dataset constructed as part of the study The dataset contains derived multimodal benchmark records constructed from session-linked text and wearable representations using verified group_id correspondence. It includes experimental fold and seed information, fusion-related features, HMM-derived states, and predictions from multimodal and unimodal models. The dataset contains 2,500 records representing 500 unique sessions across five random seeds (42, 1, 7, 13, and 99), following the evaluation protocol described in the study. The HMM-derived states are model-derived behavioral representations used as benchmark annotations and should not be interpreted as clinically validated psychological states or clinical diagnoses. The dataset is intended for research on multimodal representation learning, fusion, temporal modeling, and benchmark evaluation.