You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

UltraStream-YouTube (Track A of UltraStream-Bench, release a-0.3.2)

Status: private, internal storage only. Never to be made public.

⚠️ Third-party content notice

The video files in videos/ are YouTube videos belonging to their original creators/channels (third-party content), kept here only as an internal mirror to reproduce experiments on a benchmark built on top of them. This repo must always stay private and must never be made public, mirrored, or redistributed outside this team.

The official, public release of this benchmark never includes video. The public release format is youtube_id + question timestamps only (see data/{dev,test}.json's youtube_id + t_ask fields) — exactly like data/videos.csv in the ultrastream-bench source repo. Anyone reproducing results from the public release downloads the videos themselves from YouTube by youtube_id; this private repo exists purely so this team doesn't need to re-download/re-process 60 multi-hour videos from YouTube every time an experiment needs to run.

What this is

Track A of UltraStream-Bench: 60 ultra-long (4-15h) YouTube videos across 9 classes, 2604 causal streaming QA questions (447 dev / 2157 test), 5-option multiple choice, including UNOBS ("Not observed yet") questions only answerable after t_ask >= 300s. Every question has a t_ask: a method under evaluation may only see video in [0, t_ask].

Format

data/{dev,test}.json: the exact public JSON format (no private/ split included here — this repo's data/ is identical to ultrastream-bench's own data/track_a_youtube/{dev,test}.json). See that repo's data/SCHEMA.md for the full field-by-field spec.

Key stats

See STATS.md (+ figs/) for the complete breakdown: per-class counts, question_type × split, mode (Near/Far) × Δt bucket, UNOBS specifics (count, t_ask range, histogram), and 5-option answer-letter balance (both over all questions and over ANSWERED-only, since E is gold for every UNOBS question by construction and would otherwise look like a skew that isn't one).

Baselines

See results/baselines.md — copied verbatim from the source repo's scripts/score.py output, which reproduces results/youtube_test_v032/ PAPER_TABLES.md exactly (not re-scored here).

Video files

videos/<video_id>.mp4 — the original source files, byte-identical, no transcoding. videos_manifest.csv gives video_id, youtube_id, class, duration_s (from ffprobe-equivalent metadata), size_bytes, sha256 (computed directly from each file, so file integrity can be verified after download), codec, and hf_path.

Use with ultrastream-bench

git clone <ultrastream-bench repo>
cd ultrastream-bench
# download videos/<video_id>.mp4 from this private repo into whatever
# directory configs/paths.yaml calls VIDEO_ROOT_A, renamed to
# <youtube_id>.mp4 (see videos_manifest.csv's youtube_id column), then:
python scripts/validate.py
python scripts/check_videos.py --track a
Downloads last month
-