Realtime-QA-100K: 100K-Sample Video Question Answering Dataset

Realtime-QA-100K is a 100,000-sample dataset for realtime video question answering constructed from YouTube videos. Each sample contains a multimodal conversation and frame timestamp metadata aligning video tokens with specific timestamps. The dataset is an open-source subset of training data for the MOSS-Video-Preview model, released by the OpenMOSS-Team.

Use Cases

• Train real-time video question answering models based on multimodal conversations.
• Benchmark model performance on timestamp-aligned video understanding tasks based on frame metadata.
• Develop and evaluate real-time reasoning agents based on the dataset's conversation structure.

Strengths

• Contains 100,000 samples, providing a substantial scale for training.
• Includes frame timestamp metadata aligning video tokens to specific frames, enabling precise temporal analysis.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Freshness should be verified as the last update timestamp is 2026-06-01 10:20:22.

Provenance

Source

OpenMOSS-Team

Collection Method

Constructed from YouTube videos.

Freshness

Last updated 2026-06-01 10:20:22.