README.md

---
pretty_name: OneStreamer-Core Source Videos
language:
  - en
  - zh
task_categories:
  - visual-question-answering
tags:
  - video
  - streaming-video
  - onestreamer
---

# OneStreamer-Core-SourceVideo

Video assets for the **OneStreamer-Core** annotations in [MCG-NJU/OneStreamer-1M](https://huggingface.co/datasets/MCG-NJU/OneStreamer-1M). See its [dataset card](https://huggingface.co/datasets/MCG-NJU/OneStreamer-1M) for the task taxonomy and metadata format, and [our paper](https://arxiv.org/abs/2610.01762) for the data construction pipeline.

The release is organized by source directly under the repository root. Each uncompressed `part-XXXXX.tar` has a corresponding `part-XXXXX.manifest.json`; download them as a pair. The source videos remain subject to their respective source datasets' terms.

```text
README.md
AVA/
  part-00000.tar
  part-00000.manifest.json
  ...
COIN/
Ego4D/
EgoQA/
MAGQA/
Molmo2-Cap/
Molmo2-VideoCapQA/
PerceptionTest/
Shot2Story/
THUMOS/
TVQA/
TimeLens-100K/
TimeLens2-93K/
UCF-Crime/
YouTube-Collected/
```

## Download and unpack

For example, download the EgoQA shard using `huggingface_hub`:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="yyd7/OneStreamer-Core-SourceVideo",
    repo_type="dataset",
    allow_patterns=["EgoQA/*"],
    local_dir="OneStreamer-Core-SourceVideo",
)
```

Unpack the selected shards into a shared media root:

```bash
mkdir -p media
tar -xf OneStreamer-Core-SourceVideo/EgoQA/part-00000.tar -C media
```

For annotations with `media_type: "video"`, the extracted path is `media/<video_path>` and can be used directly. Each shard contains only the selected Core media, with one file per distinct metadata media reference.

## Match videos to metadata

Each manifest identifies the annotation repository in `hf_dataset` and pins its revision in `annotation_revision`. The `video_dataset` field identifies this source-video repository; all `archive_path` values are relative to its root. The manifest's `assets` list contains:

| Field | Meaning |
| --- | --- |
| `media_key` | Stable identifier for the metadata media reference. |
| `source` | Source dataset or collection. |
| `hf_video_path` | Exact `video_path` in the Core metadata. |
| `archive_path` | Tar location relative to this source-video repository, such as `EgoQA/part-00000.tar`. |
| `archive_member` | MP4 location inside the tar and under the extraction root. |
| `bytes`, `sha256` | Size and SHA-256 of the individual MP4. |
| `metadata_media_type` | The annotation's original media representation: `video` or `sequence`. |
| `provenance` | Whether this asset is a dataset video/clip, a processed clip/segment, or a reconstruction from frames. |
| `sequence` | The annotation's frame rate and, when specified, filename pattern and starting index. |

Match an annotation's `video_path` to `hf_video_path`. The MP4 to read is identified by `archive_path` and `archive_member`. All annotation IDs, times, task assignments, and silence/standby/response data retain their metadata definitions.

## Frame-sequence annotations

For `metadata_media_type: "sequence"`, `hf_video_path` refers to a **frame directory**. The released MP4 is the input for reconstructing that frame sequence; its path is recorded separately as `archive_member`. Follow `sequence.fps` and any `filename_pattern` / `index_start` when creating the frame directory.

- Time-based 4 FPS names are `00000_00.jpg`, `00000_25.jpg`, `00000_50.jpg`, `00000_75.jpg`, followed by `00001_00.jpg`, and so on. The sample times are 0, 0.25, 0.5, 0.75 seconds, etc.
- TVQA uses 3 FPS and one-based names such as `00001.jpg`; frame index 1 is at time 0.
- If the metadata omits a filename pattern, use the source dataset's frame convention. A frame-directory reference should not be replaced blindly with an MP4 path in a loader that expects images.

AVA and TVQA assets marked `reconstructed_from_frames` preserve the available frame order and timing and have no audio. They are H.264 reconstructions, not recovered original recordings, and do not reproduce JPEG files byte for byte. AVA's separately available original video is identified by its own provenance. Other sequence sources use their available dataset videos or clips. Processed clips and segments retain their local annotation time axis; they are not replaced with full-length source recordings.