Dataset for "When2Speak: A Dataset for Temporal Participation and Turn-Taking in Multi-Party Conversations for Large Language Models"
NeurIPS 2026 — Evaluations and Datasets Track
When2Speak is a large-scale synthetic dataset for learning intervention timing in multi-party conversations: given the recent conversation history, should an AI agent speak or remain silent at this turn?
The dataset comprises 216,800 labeled (context, decision) pairs… See the full description on the dataset page:
https://huggingface.co/datasets/duke-trust-lab/When2Speak.
