Lax3n commited on
Commit
fb8cd9a
·
verified ·
1 Parent(s): f5d6744

Upload folder using huggingface_hub

Browse files
Files changed (3) hide show
  1. README.md +14 -12
  2. captions.parquet +2 -2
  3. metadata.jsonl +0 -0
README.md CHANGED
@@ -21,39 +21,41 @@ configs:
21
 
22
  # LAMBDA — YouTube source videos
23
 
 
24
  **1,436 videos** (15.5 GB) with 835 subtitle files, downloaded at source
25
  quality and hosted for direct download.
26
 
27
- > **Coverage (2026-11-15):** 1,436 complete YouTube source videos are published here. They come from a prepared selection of 2,154 YouTube IDs read from [`behavior-in-the-wild/LAMBDA`](https://huggingface.co/datasets/behavior-in-the-wild/LAMBDA) (revision `6b2ce33a8451`); the selection stopped because all matching upstream metadata files were read. It is a bounded subset, not the complete upstream list.
28
- > Of the prepared IDs: 1,436 published, 570 unavailable at download time, 0 failed downloads, 0 not added because of the 105 GB cap, 12 in progress and 136 not attempted.
29
- > Every row of the `videos` configuration points to a file present in our storage when this card was generated, and the annotations are restricted to those videos.
30
 
31
 
32
  > **Public release:** capped at 105 GB (105,000,000,000 bytes), including media and indexes.
33
- > **Unit:** complete YouTube source videos. Upstream clip intervals, where present, remain in the annotations; the videos have not been cut to those intervals. The prepared metadata is a bounded selection, not a complete benchmark.
34
 
35
  ## What's inside
36
 
37
- - **`metadata.jsonl`** — one row per video: `video_id`, `video_url` (direct HTTPS download, CDN-backed), `subtitle_url`, `youtube_url`, `size_bytes`
38
- - The 1,436 MP4s (15.5 GB) and 835 subtitle files are hosted on Hugging Face storage and served via the `video_url` column
39
- - **`captions.parquet`** — 1,443 academic annotations (source: `annotations.csv`)
40
 
41
  ## Quick start
42
 
43
  ```python
44
  from datasets import load_dataset
45
 
46
- # videos (default config): one row per video with its direct download URL
47
  ds = load_dataset("TornadoLabs/lambda", "videos", split="train")
48
  print(ds[0])
49
 
50
- # download one video
51
  import requests
52
- r = requests.get(ds[0]["video_url"])
53
- open(f"{ds[0]['video_id']}.mp4", "wb").write(r.content)
 
54
 
55
 
56
- # annotations for the published source videos
57
  ann = load_dataset("TornadoLabs/lambda", "annotations", split="train")
58
 
59
  ```
 
21
 
22
  # LAMBDA — YouTube source videos
23
 
24
+ **2,154 source IDs** with their source links and availability.
25
  **1,436 videos** (15.5 GB) with 835 subtitle files, downloaded at source
26
  quality and hosted for direct download.
27
 
28
+ > **Coverage (2026-11-16):** 1,436 complete YouTube source videos are hosted here. They come from a prepared selection of 2,154 YouTube IDs read from [`behavior-in-the-wild/LAMBDA`](https://huggingface.co/datasets/behavior-in-the-wild/LAMBDA) (revision `6b2ce33a8451`); the selection stopped because all matching upstream metadata files were read.
29
+ > Of the prepared IDs: 1,436 hosted, 570 unavailable at download time, 0 failed downloads, 0 not added because of the 105 GB cap, 12 in progress and 136 not attempted.
30
+ > The `videos` configuration lists **every prepared ID**, hosted or not: `video_url` is filled only for the 1,436 hosted videos, and `availability` gives the state of each row when this card was generated. The annotations cover all listed IDs.
31
 
32
 
33
  > **Public release:** capped at 105 GB (105,000,000,000 bytes), including media and indexes.
34
+ > **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.
35
 
36
  ## What's inside
37
 
38
+ - **`metadata.jsonl`** — 2,154 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`
39
+ - 1,436 of these videos (15.5 GB) are hosted on Hugging Face storage; the other rows keep their source link only
40
+ - **`captions.parquet`** — 2,166 academic annotations (source: `annotations.csv`)
41
 
42
  ## Quick start
43
 
44
  ```python
45
  from datasets import load_dataset
46
 
47
+ # videos (default config): one row per source ID; video_url is set for hosted videos only
48
  ds = load_dataset("TornadoLabs/lambda", "videos", split="train")
49
  print(ds[0])
50
 
51
+ # keep the hosted videos, then download one
52
  import requests
53
+ hosted = ds.filter(lambda r: r["video_url"] is not None)
54
+ r = requests.get(hosted[0]["video_url"])
55
+ open(f"{hosted[0]['video_id']}.mp4", "wb").write(r.content)
56
 
57
 
58
+ # annotations for all listed source videos
59
  ann = load_dataset("TornadoLabs/lambda", "annotations", split="train")
60
 
61
  ```
captions.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:120eef38e29574d4fb020cb105696edec7b2ed4a88bd7a6b32bc357b4489757e
3
- size 57810
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e908761377ba24f5482d04bdd7bbe38137ca94c1230e00eec951d32d4e68931e
3
+ size 85854
metadata.jsonl CHANGED
The diff for this file is too large to render. See raw diff