---
dataset_info:
  features:
  - name: video1
    dtype: video
  - name: video2
    dtype: video
  - name: label
    dtype: int32
  splits:
  - name: test
    num_examples: 2048
configs:
- config_name: default
  data_files:
  - split: test
    path: data/*
license: cc-by-nc-4.0
task_categories:
- video-classification
language:
- en
pretty_name: AV-SpeakerBench Pair Classification
---

# AV-SpeakerBench-PC

Video pair classification dataset from [AV-SpeakerBench](https://arxiv.org/abs/2512.02231) (Nguyen et al., 2025).

Each row contains two video clips (video1, video2) and a binary label indicating whether they
come from the same source video (same speaker context) or different source videos (different speakers).

Source: `mteb/AV-SpeakerBench` — deduplicated by video_id, 2048 balanced pairs generated from YouTube source video IDs.