README.md

---
dataset_info:
  features:
  - name: video
    dtype: video
  - name: text
    dtype: string
  - name: label
    dtype: int32
  splits:
  - name: test
    num_examples: 1138
configs:
- config_name: default
  data_files:
  - split: test
    path: data/*
license: mit
task_categories:
- video-classification
language:
- en
pretty_name: VideoCon Pair Classification
---

# VideoCon-PC

Video-language pair classification dataset from [VideoCon](https://arxiv.org/abs/2311.10111) (Bansal et al., 2023).

Each row contains a video and a text caption. Label=1 means the caption correctly describes the video;
label=0 means it is a semantically-plausible contrast caption (entity/action/attribute swaps, event order flips).

Source: `videocon/videocon` (videocon_human.csv) — 569 videos from ActivityNet, 1138 pairs total (2 corrupt videos removed).