VoxPolyBench

VoxPolyBench contains 18 named multi-party conversation cases, 176 sessions, 9,599 spoken dialogue turns, and 1,527 QA samples. Of the QA samples, 434 include spoken question audio; the remaining 1,093 use question text only.

Each cases/VPB_###_scenario_name/ directory contains two independent Hugging Face Arrow datasets:

• content/: one dialogue turn per row, with case_id, session_id, turn_id, case-scoped speaker_id, text, and embedded WAV audio.
• qa/: one question sample per row, with case_id, question_id, question text, reference answer, qa_type, optional asker_speaker_id, and optional embedded WAV audio.

release_manifest.json lists the public scenario titles and counts. The speaker_id values are stable within each case but do not identify a person across cases. Dialogue rows do not contain speaker-name annotations or an ID-to-name mapping. Names occurring naturally in the spoken transcripts are left intact. Audio is stored in each Arrow row as {bytes, path} with a null path, so no original server files are needed.

To read a local copy:

from datasets import load_from_disk

case = "cases/VPB_001_airline_service_coordination"
content = load_from_disk(f"{case}/content")
qa = load_from_disk(f"{case}/qa")

After uploading the whole folder to a Hugging Face dataset repository, use huggingface_hub.snapshot_download(repo_id="NAMESPACE/VoxPolyBench", repo_type="dataset") to obtain the local root and pass its case subdirectories to load_from_disk. These files are Dataset.save_to_disk Arrow datasets; automatic Hub dataset-viewer previews are not guaranteed for this nested layout. speaker_id is an evaluation annotation and should not be supplied as input to a speaker-identification model.