A video VQA benchmark for real-vs-generated video discrimination. Given a short
clip, a model must answer a single multiple-choice question: is this a real video, or a
fake (AI-generated) one? Construction follows the
VSI-Bench recipe — metadata in a
Parquet file, videos shipped separately and referenced by filename.
2,000 questions — perfectly balanced: 1,000 real + 1,000 fake.
Real videos: the Golden Eval v0.4 ground-truth clips (real-world footage).
Fake videos:… See the full description on the dataset page:
https://huggingface.co/datasets/ShushengYang/RealOrFake.
