Datasetpublic

PhoenixHu/youtube_eval_0616

YouTube Eval 0616 — ASL Sign Language Translation Test Set A small evaluation dataset for American Sign Language → English video-to-text translation, built from 5 YouTube videos (4 usable; 1 with no usable captions excluded). The captions come from YouTube's manual / auto subtitles, paired with the corresponding video segments. Source Videos YouTube ID Title Duration Clips Notes vw73r_xaeX0 Google I/O '26 Keynote – American Sign Language 111 min 1961… See the full description on the dataset page: https://huggingface.co/datasets/PhoenixHu/youtube_eval_0616.

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

0likes24downloads

Dataset Card

YouTube Eval 0616 — ASL Sign Language Translation Test Set

A small evaluation dataset for American Sign Language → English video-to-text translation, built from 5 YouTube videos (4 usable; 1 with no usable captions excluded). The captions come from YouTube's manual / auto subtitles, paired with the corresponding video segments.

Source Videos

YouTube ID	Title	Duration	Clips	Notes
`vw73r_xaeX0`	Google I/O '26 Keynote – American Sign Language	111 min	1961	ASL interpreter on the right side of frame
`t123Ji6i5XA`	watch this video if you are feeling lost	13.5 min	499	Single signer, full-frame
`VycGU4Y91UE`	why is sign language so aesthetic?	18.5 min	416	Single signer, full-frame
`Dm0-1qVZlf4`	Last Week Was Really Challenging	7.7 min	97	Single signer, full-frame

Filtering Rules

Caption segments were used to slice each source video. The following segments were skipped:

1.Segments with more than 2 special characters (e.g. >>, [laughter], [Music]) — these are non-speech / format markers, not signed content.
2.Segments with fewer than 3 words.

Files

File	Description
`video_clips.tar.gz`	2,973 raw segment clips (original resolution, 1.3 GB). Layout: `video_clips/<video_id>/<video_id>_NNNN.mp4`
`clips_cropped_224.tar.gz`	1,012 clips cropped to 224×224 using YOLO-Pose bbox (signer-centered), 207 MB. Excludes `vw73r_xaeX0` (right-side signer bbox extraction was deferred). Layout: `clips_cropped_224/<video_id>/<video_id>_NNNN_cropped.mp4`
`clips_index.json`	Per-clip metadata: `{video_id: [{idx, clip, start, end, duration, text, lang}, ...]}`
`subtitles_combined.json`	All subtitle tracks per video (multiple languages where available), title, channel, description
`test_youtube_eval_0616_internvl.jsonl`	Test split formatted for InternVL evaluation (1012 entries, 224×224 clips). Each entry: `{id, video, conversations:[{from:human, value:"<video>\nTranslate..."}, {from:gpt, value:"<caption>"}]}`

Caption Track Selection

For each video the cleanest English track was picked (priority: en-US > en-en > en-orig > en). YouTube auto-subs that accumulate text across consecutive cues were de-duplicated to extract only the new tail.

Caveats

—Caption text comes from YouTube's auto / community captions — quality varies.
—vw73r_xaeX0 (Google I/O Keynote) has the ASL interpreter in a small right-side inset. A dedicated right-side bbox extractor was written but not run; the 1961 raw clips are included so future runs can produce 224×224 crops focused on the interpreter.
—All content is downloaded from publicly available YouTube videos. License of the underlying videos belongs to their respective creators.

Quick Start (InternVL evaluation)

bash

tar xzf clips_cropped_224.tar.gz
# Run InternVL evaluation with test_youtube_eval_0616_internvl.jsonl pointing at
# the extracted clips_cropped_224/ directory as --video-folder.