Text-Anchored Multimodal Residual Fusion with Self-Reflective Modality Calibration for Student Emotion Analysis

Overview OLE-EmoBench is a reproducible synthetic benchmark of student academic affect, introduced in Text-Anchored Multimodal Residual Fusion with Self-Reflective Modality Calibration for Student Emotion Analysis. It comprises 4,200 ten-second interaction episodes from 210 simulated learners across two modelled settings: synchronous one-on-one STEM tutoring (OLE-Dialog, 2,400 episodes) and asynchronous lecture question answering (OLE-Lecture, 1,800 episodes). Nature of the data. Episode boundaries, learner and episode identifiers, sensor-quality vectors, utterance statistics, and five-class emotion labels are produced by a deterministic generator from fixed random seeds. No audio, video, transcript, or other data were recorded from any person, and no human participants are involved. Identifiers and ages are generator parameters rather than coarsened observations. The corpus is exactly reproducible from its generator and seeds. Contents (18 files). Episode-level descriptors (ole_emobench_episodes.csv, 4,200 rows x 30 columns); frozen backbone intermediates (ole_emobench_backbone_intermediates.npz: DeBERTa-v3-base 768-d utterance and 4-sub-window embeddings, HuBERT-base 768-d acoustic embeddings with 6 prosodic channels, OpenFace 2.0 35-channel facial descriptors); projected R^16 sub-window tensors (ole_emobench_subwindows.npz); 13 derived evaluation tables under derived/ covering the main comparison, per-seed runs, component ablations, anchor and geometry replacements, hyperparameter sensitivity, corruption robustness, subgroup analysis, per-class metrics, confusion matrices, calibration bins, the SRMCA step trace, and test-split predictions; plus dataset_metadata.json and README.md. Reuse. Released under CC BY 4.0. If you use these data, please cite this record together with the associated article.