CogStream: 70,987 QA Pairs for Context-Guided Streaming Video Reasoning

CogStream contains 70,987 question-answer pairs across 1,088 videos, developed by SII-KYW in 2025 to evaluate context-guided streaming video reasoning. The dataset requires models to identify and utilize relevant historical context from ongoing video streams to answer specific questions about current events.

Use Cases

• Temporal reasoning tasks requiring models to link current video frames to historical context segments
• Streaming video question answering using the 15,364 test set QA pairs
• Context-guided video reasoning to identify relevant past events for answering queries about ongoing streams

Strengths

• 70,987 QA pairs
• 1,088 unique videos
• Diverse source material from MovieChat (40.2%), MECD (16.8%), and QVhighlights (9.8%)

Limitations

• High QA-to-video ratio of approximately 65:1 may lead to model overfitting on specific video contexts
• Source imbalance with MovieChat representing over 40% of the total data

Provenance

Source

SII-KYW, Arxiv 2506.10516

Collection Method

Aggregated from existing datasets including MovieChat and MECD, then annotated for streaming context

Freshness

Last updated March 2026