---
license: cc-by-nc-sa-4.0
language:
- en
- ja
pretty_name: MEDit-Bench
tags:
- video
- video-summarization
- temporal-grounding
- video-captioning
task_categories:
- video-text-to-text
size_categories:
- 1K<n<10K
---

# MEDit-Bench: Temporal Grounding Annotations

Human-annotated temporal grounding data for 60 YouTube videos. For each video,
annotators were given a short list of "summary messages" describing the video
and asked to mark every time range in the video that supports each message.
Each message was annotated independently by 3 annotators.

This dataset ships **annotations only** — no video files. Each row points to
the source video via its YouTube ID / URL so the underlying footage can be
retrieved directly from YouTube.

## Dataset structure

One row = one summary message annotated by one annotator.

| column | type | description |
|---|---|---|
| `video_id` | string | YouTube video ID |
| `message` | string | Summary message text (English) |
| `annotator_index` | int | Index (0-based) of the annotator — 3 annotators per message |
| `segments` | list of `[start, end]` | The cuts this annotator marked as supporting the message. `HH:MM:SS:FF` timecode (frame number — not converted to seconds since source frame rate varies). |

- 60 videos, 180 messages (3 per video), 540 rows (3 annotators per message).

## License

Annotations are released under CC BY-NC-SA 4.0. Source videos remain the
property of their original YouTube creators and are not redistributed here.