SVI-Bench dataset

SVI-Bench contains 75 streaming-video interaction tasks backed by 71 unique source videos. Four videos are intentionally reused across task categories; media_index.jsonl is the authoritative task-to-media mapping.

Use of this dataset is governed by the SVI-Bench Dataset Terms. It is intended solely for academic research, and commercial use in any form is prohibited. Copyright in the source videos remains with the respective rights holders.

Files

• SVI_bench_tasks_and_anchors.xlsx: task definitions and dimension-specific grading anchors.
• source_videos/: 71 metadata-sanitized source videos. Streams are remuxed without re-encoding.
• media_index.jsonl: mapping from the 75 task IDs to the 71 source videos.
• annotations/SVI-Pilot-human-ratings.xlsx: expert pilot labels used for Judge alignment.

Use the evaluation code at https://github.com/YidanHAI/SVI-Bench. Pin both the code commit and the Hugging Face dataset revision when reporting results.

Questions and removal requests may be filed at https://github.com/YidanHAI/SVI-Bench/issues.

Dataset repository: https://huggingface.co/datasets/Danmel02/SVI-bench