--- license: other license_name: academic-source-terms license_link: https://huggingface.co/datasets/gijs/avqa-processed task_categories: - video-classification - text-to-video - video-text-to-text tags: - video - youtube - tornadolabs pretty_name: "avqa-processed — YouTube source videos" configs: - config_name: videos default: true data_files: metadata.jsonl --- # avqa-processed — YouTube source videos **9,513 source IDs** with their source links and availability. **3,555 videos** (212.5 GB) with 1,043 subtitle files, downloaded at source quality and hosted for direct download. > **Coverage (2026-11-16):** 3,555 complete YouTube source videos are hosted here. They come from a prepared selection of 9,513 YouTube IDs read from [`gijs/avqa-processed`](https://huggingface.co/datasets/gijs/avqa-processed) (revision `e2aa3fc443a4`); the selection stopped because all matching upstream metadata files were read. > Of the prepared IDs: 3,555 hosted, 530 unavailable at download time, 7 failed downloads, 0 not added because of the 264 GB cap, 14 in progress and 5,407 not attempted. > The `videos` configuration lists **every prepared ID**, hosted or not: `video_url` is filled only for the 3,555 hosted videos, and `availability` gives the state of each row when this card was generated. The annotations cover all listed IDs. > **Public release:** capped at 264 GB (264,000,000,000 bytes), including media and indexes. > **Unit:** complete YouTube source videos. Upstream clip intervals, where present, remain in the annotations; the videos have not been cut to those intervals. The index lists every prepared ID; only part of them is hosted. > **Annotations:** the upstream metadata is not redistributed here; join it by `video_id` from [`gijs/avqa-processed`](https://huggingface.co/datasets/gijs/avqa-processed) (revision `e2aa3fc443a4`). ## What's inside - **`metadata.jsonl`** — 9,513 rows, one per prepared source ID: `video_id`, `youtube_url`, `availability`, `in_prepared_selection`, `video_url` (direct HTTPS download, CDN-backed; empty when the video is not hosted), `subtitle_url`, `size_bytes`, `last_attempt_at` - 3,555 of these videos (212.5 GB) are hosted on Hugging Face storage; the other rows keep their source link only ## Quick start ```python from datasets import load_dataset # videos (default config): one row per source ID; video_url is set for hosted videos only ds = load_dataset("TornadoLabs/avqa-processed", "videos", split="train") print(ds[0]) # keep the hosted videos, then download one import requests hosted = ds.filter(lambda r: r["video_url"] is not None) r = requests.get(hosted[0]["video_url"]) open(f"{hosted[0]['video_id']}.mp4", "wb").write(r.content) ``` ## How this dataset was built Each hosted file is the complete YouTube source video (best available quality), downloaded from YouTube into a temporary location, measured, then copied into the public storage only while the release stayed under 264 GB. Clip intervals from the upstream annotations are not applied to the files. Videos that were unavailable, failed or did not fit are counted in the coverage note above. ## Provenance & citation - Original dataset: https://huggingface.co/datasets/gijs/avqa-processed - Paper: See the upstream dataset card - No explicit license is declared upstream for the annotations; rights to the third-party videos are separate. The video files remain the property of their respective YouTube uploaders. The annotations belong to the original dataset authors — cite their paper if you use them. This re-hosting is provided for reproducible research; if you are a rights holder and want content removed, open a discussion on this repo. ## Terms and removal This is a non-commercial research mirror of [`gijs/avqa-processed`](https://huggingface.co/datasets/gijs/avqa-processed) (revision `e2aa3fc443a4`). The upstream repository declares no license: nothing here grants rights that the upstream authors did not grant. All credit goes to the upstream authors; cite them, not this repository. No rights to the source videos are granted. Upstream authors and rights holders can ask for removal by opening a discussion on this repository; requests are applied without conditions.