GroupAffect-4: A Multimodal Dataset of Four-Person Collaborative Interaction (Subset Release — No Audio, <4 GB)

GroupAffect-4 is a multimodal corpus of co-located four-person group interaction across five structured tasks: T0 resting baseline, T1 hidden-profile decision, T2 mini-negotiation, T3 nominal-group idea generation, and T4 public-goods micro-game. The dataset contains 40 participants in 10 groups with per-participant EmotiBit wrist physiology (PPG, EDA, skin temperature, IMU at ~25 Hz), Tobii Pro Glasses 3 egocentric eye-tracking (~50 Hz), DPA close-talk lapel audio (48 kHz), segment-level and word-level speech transcripts, tablet SAM self-reports (valence, arousal, dominance), task outcome records, demographics, and BFI-44 Big Five personality scores. All signals are synchronised via Lab Streaming Layer (LSL). Room video is excluded from this release. THIS RECORD — SUBSET RELEASE, PRIVACY-PROCESSED (<4 GB)This record contains all tabular and time-series modalities except raw audio WAV files. It is the recommended starting point for most users. Transcripts have been processed to remove sensitive content (see PRIVACY PROCESSING below). Files in this record:• affectai_metadata.zip — JSON/TSV sidecars, Croissant 1.1 metadata, README (~1 MB)• affectai_physio.zip — EmotiBit physiology TSV.GZ, all 10 sessions (~175 MB)• affectai_beh.zip — Self-report and task TSV, all sessions (~7 MB)• affectai_transcripts.zip — Privacy-processed transcript and word-level TSV/TXT, all sessions (~5 MB; sensitive names and utterances redacted)• affectai_annot.zip — Session sync and annotation files (~2.2 GB)• affectai_et_grp07to15.zip — Tobii Pro Glasses 3 ET TSV.GZ, 7 of 10 sessions (~1.3 GB) Sessions grp-11, grp-12, and grp-16 ET data are in the full release record. PRIVACY PROCESSING:Transcripts in this release have been automatically processed to remove sensitive information. Speaker names, personal identifiers, and sensitive utterances flagged by the privacy pipeline are replaced with [REDACTED] tokens in both the segment-level TSV files and word-level timing files. The audio WAV files are not included in this subset release; redacted audio is available in the full release record. FULL RELEASE (~30 GB)A companion Zenodo record contains the complete dataset including raw audio WAV files (DPA lapel + room mics, 48 kHz) for all 10 sessions and Tobii ET for all 10 sessions. Link will be added to Related Works once published. Modality coverage (200 expected participant-task units):• EmotiBit physiology: 182/200 (91.0%)• Tobii eye tracking (this record): 140/200 (70.0%) — full record: 196/200 (98.0%)• Valence/arousal self-report: 171/200 (85.5%)• BFI-44 personality: 40/40 (100%) Eight benchmarkable evaluation targets (B0–B7) cover task-label classification, affective and cognitive state estimation, personality inference, and group-level interaction metrics. Evaluation uses leave-one-group-out cross-validation (split key: group_id). Processing code and documentation: https://github.com/meisamjam/GroupAffect-4 Responsible AI fields (data limitations, biases, sensitive information, usage restrictions) are documented in croissant_metadata.json (Croissant 1.1) inside affectai_metadata.zip and in the companion NeurIPS 2026 paper.