File size: 6,581 Bytes
613afda
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
---

language:
- en
license: cc-by-4.0
size_categories:
- n<1K
pretty_name: YT-Temporal-1B / MERLOT Reserve — TEMPURA Annotation Sample
tags:
- video-metadata
- youtube
- flashdata
- custom-video-dataset
configs:
- config_name: default
  data_files:
  - split: sample
    path: sample.parquet
---


# YT-Temporal-1B / MERLOT Reserve — TEMPURA Annotation Sample

128 video references with dense event descriptions from TEMPURA-VER, whose authors identify its source videos as a YT-Temporal-1B subset.

**128 records · TEMPURA text annotations and video references · Parquet + JSONL · No video files**

**[Request a custom video dataset](https://flashdata.dev/products/youtube-video-datasets?utm_source=hugging-face&utm_medium=dataset&utm_campaign=video_datasets_2026q4&utm_content=yt-temporal-1b-tempura-sample-custom#request)** · **[Explore the YouTube Download API](https://flashdata.dev/products/youtube-download?utm_source=hugging-face&utm_medium=dataset&utm_campaign=video_datasets_2026q4&utm_content=yt-temporal-1b-tempura-sample-api)**

Use this sample to inspect a source manifest, test a loader and prepare a scoped collection request. For playable MP4 examples, see [FlashData's video dataset Collection](https://huggingface.co/collections/flashdata-dev/flashdata-video-datasets-and-download-api).

## Preview

| Video ID | Example metadata |
|---|---|
| `olJowEmbvpw` | From 0.0 to 1.0 seconds, the video starts with an individual in a blue shirt, possibly introducing the video. … |
| `dj5tYkDvsRI` | From 0.0 to 1.0 seconds, an individual is standing in the kitchen, wearing a black hoodie with a design on it.… |
| `5Uw42ROlz5c` | From 0.0 to 2.0 seconds, the video begins with the Baudelaire orphans sitting back-to-back in a dimly lit room… |
| `7GaRAO3hm7Q` | From 0.0 to 3.0 seconds, the video starts with a black screen displaying numbers and words like '0', '50', 'ar… |

## Load the sample

```python

from datasets import load_dataset



sample = load_dataset("flashdata-dev/yt-temporal-1b-tempura-sample", split="sample")

print(len(sample))  # 128

print(sample[0]["annotation_text"])

```

Install `datasets` first if needed. Direct files: [Parquet](sample.parquet), [JSONL](sample.jsonl), [provenance](provenance.json), [checksums](https://huggingface.co/datasets/flashdata-dev/yt-temporal-1b-tempura-sample/blob/main/SHA256SUMS), [quality report](quality-report.json).

## Prepare your source list

Download [prepare_delivery.py](prepare_delivery.py) and [sample.jsonl](sample.jsonl) into the same folder, then run:

```bash

python prepare_delivery.py sample.jsonl --output delivery-manifest.jsonl --report normalization-report.json

```

The helper accepts local CSV or JSONL, normalizes YouTube IDs, deduplicates sources and accounts for every input row. It makes no network requests and does not download media. The output is a source manifest for scoping a delivery, not an API request payload.

For a custom collection, provide the source list or topic, target languages, desired resolution, clip boundaries, caption needs and storage destination. FlashData can scope available media delivery and metadata packaging against the requested permissions and acceptance criteria. See the [YouTube Download API documentation](https://flashdata.dev/docs/sources/youtube_download) for the current API contract.

## Fields

| Field | Meaning |
|---|---|
| `video_id, id` | Original YouTube ID and convenient alias |
| `annotation_text` | Unchanged upstream dense-caption target |
| `source_video_url` | YouTube reference generated from the original ID |
| `conversations` | Original human/gpt dialogue, preserved as nested data |
| `masked_start_end_frame_idxs` | Original field, null for this DVC sample |
| `annotation_model, source_split` | Upstream model attribution and split |
| `source_row_index` | Zero-based row in the pinned shard |

## Source and selection

Upstream: [https://huggingface.co/datasets/andaba/TEMPURA-VER](https://huggingface.co/datasets/andaba/TEMPURA-VER). Exact source revision and record positions are in `provenance.json`.

First 128 records of the pinned TEMPURA-VER DVC500k_gpt4o first shard. This is a contiguous convenience sample. It is not an official MERLOT Reserve release or a mirror of the full YT-Temporal-1B corpus.



**Changes made by FlashData:** Preserved id, conversations and masked_start_end_frame_idxs exactly. Added video_id, source_video_url, annotation_model, source_split and zero-based source_row_index. annotation_text copies the upstream gpt response without editing. These are TEMPURA's GPT-4o annotations, not MERLOT's original annotations or spoken transcripts.



## License and attribution



TEMPURA-VER releases its text annotations under CC BY 4.0. The underlying videos remain their owners' property. This license applies to the included annotations, not to all YT-Temporal-1B videos or the original MERLOT corpus. Captions are model-generated and can be incorrect; no human annotation-accuracy or source-video availability check is claimed.



TEMPURA authors: Jen-Hao Cheng, Yi-Hao Peng, Huapeng Zhou, Vivian Wang, Huayu Wang, Hsiang-Wei Huang, Wenhao Chai, Hou-I Liu, Kuang-Ming Chen, Cheng-Yen Yang, Yi-Ling Chen, Vibhav Vineet, Qin Cai and Jenq-Neng Hwang. The underlying YT-Temporal-1B corpus was introduced by Rowan Zellers and the MERLOT Reserve collaborators.



See [LICENSE](https://huggingface.co/datasets/flashdata-dev/yt-temporal-1b-tempura-sample/blob/main/LICENSE), [ATTRIBUTION.md](ATTRIBUTION.md) and [LICENSE-CODE](https://huggingface.co/datasets/flashdata-dev/yt-temporal-1b-tempura-sample/blob/main/LICENSE-CODE). FlashData selected and packaged this sample independently. No affiliation or endorsement by the upstream projects or creators is implied.



## Validation and limits



All 128 records load locally through Hugging Face Datasets; JSONL and Parquet match exactly. IDs are unique and syntactically valid. Every package file has a SHA-256 checksum. No source video was downloaded or checked for current availability, and no transcript or annotation accuracy score is claimed. Upstream content, availability and rights may differ across records. This small sample is intended for inspecting formats and workflows.



## References



- [TEMPURA project](https://andy-cheng.github.io/TEMPURA/)

- [TEMPURA paper](https://arxiv.org/abs/2505.01583)

- [MERLOT Reserve / YT-Temporal-1B](https://rowanzellers.com/merlotreserve/)

- [MERLOT paper](https://arxiv.org/abs/2201.02639)