AskStream

AskStream is the first multi-turn benchmark spanning video and user uncertainty for streaming video understanding. It extends proactive video interaction with an ASK action, allowing models to proactively request missing information before answering, in addition to the conventional WAIT and ANSWER actions.

The benchmark contains 438 videos across 9 diverse domains, 1,401 questions covering 8 subtasks, and 2,714 manually annotated frame-level time windows for three interaction types: WAIT, ASK, and ANSWER. It supports both multiple-choice and open-ended evaluation across four task categories.

• Paper: AskStream: When and What to Ask Proactively under Uncertainty in Streaming Video Understanding
• Project page: https://askstream.top/
• Code: https://github.com/wanglu-cs/AskStream

Sample usage

The repository README provides the following command to download the dataset:

python -m pip install -U huggingface_hub

hf download wanglucs/AskStream \
  --repo-type dataset \
  --local-dir /mnt/userdata/AskStream/data