ALWM_data / README.md
wudq's picture
Add ALWM D1/D2/D3 data
a8e7c28 verified
|
Raw History Blame Contribute Delete
3.83 kB
---
license: cc
---
# ALWM Data
Data used by ALWM.
## Dataset statistics
<table>
<thead>
<tr>
<th rowspan="2">Dataset</th>
<th rowspan="2">Size</th>
<th colspan="2">Video</th>
<th colspan="2">Query</th>
<th colspan="4">Response</th>
</tr>
<tr>
<th>Obs.</th>
<th>Fut.</th>
<th>OEQ</th>
<th>MCQ</th>
<th>Observation</th>
<th>Reasoning</th>
<th>Prediction</th>
<th>Answer</th>
</tr>
</thead>
<tbody>
<tr>
<td>D1</td>
<td>42,673</td>
<td colspan="2">Unpartitioned</td>
<td>—</td>
<td>—</td>
<td colspan="4">Caption</td>
</tr>
<tr>
<td>D2</td>
<td>15,290</td>
<td>✓</td>
<td>✓</td>
<td>✓</td>
<td>—</td>
<td>✓</td>
<td>✓</td>
<td>✓</td>
<td>—</td>
</tr>
<tr>
<td>D3</td>
<td>2,885</td>
<td>✓</td>
<td>—</td>
<td>—</td>
<td>✓</td>
<td>✓</td>
<td>✓</td>
<td>✓</td>
<td>✓</td>
</tr>
</tbody>
</table>
Size is the number of samples. Obs. and Fut. denote observed and future video segments. OEQ and MCQ denote open-ended and multiple-choice questions. D1 video segments are not assigned observed or future roles; each response is a caption.
## Directory structure
```text
ALWM_data/
├── README.md
├── D1/
│ ├── D1_metadata.json
│ └── videos/
├── D2/
│ ├── D2_metadata.json
│ └── videos/
└── D3/
├── D3_metadata.json
└── videos/
```
| Dataset | Metadata | Videos |
| --- | --- | --- |
| D1 | `D1/D1_metadata.json` | `D1/videos/` |
| D2 | `D2/D2_metadata.json` | `D2/videos/` |
| D3 | `D3/D3_metadata.json` | `D3/videos/` |
Video files are grouped into numbered subdirectories inside `videos/`; the exact path is given in the metadata.
## Metadata format
Each metadata file is a JSON object keyed by sample ID. Every entry contains only `video`, `query`, and `answer`.
| Field | Meaning |
| --- | --- |
| Sample ID (object key) | Unique identifier of the sample within the dataset. |
| `video` | Video path relative to the repository root. For D2, an object containing the `observation` and `future` video paths. |
| `query` | Question text. Empty for D1; an open-ended question for D2; a question with its choices for D3. |
| `answer` | Complete response as a single text string. Original text, tags, line breaks, and tag order are preserved; the response is not split into separate fields. |
The examples below illustrate the structure; ellipses stand for sample text.
### D1
`D1/D1_metadata.json` maps each sample ID to its video segment, an empty query, and a caption.
```json
{
"D1_example": {
"video": "D1/videos/000/D1_example.mp4",
"query": "",
"answer": "..."
}
}
```
### D2
`D2/D2_metadata.json` maps each sample ID to its observed and future video segments, an open-ended question, and the complete response text.
```json
{
"D2_example": {
"video": {
"observation": "D2/videos/000/D2_example_observation.mp4",
"future": "D2/videos/000/D2_example_future.mp4"
},
"query": "...",
"answer": "<observation>...</observation>\n<reasoning>...</reasoning>\n<latent></latent>\n<prediction>...</prediction>"
}
}
```
### D3
`D3/D3_metadata.json` maps each sample ID to its observed video segment, a multiple-choice question including its choices, and the complete response text.
```json
{
"D3_example": {
"video": "D3/videos/000/D3_example.mp4",
"query": "...\n\nOptions:\nA. ...\nB. ...\nC. ...\nD. ...",
"answer": "<observation>...</observation>\n<reasoning>...</reasoning>\n<latent></latent>\n<prediction>...</prediction><answer>...</answer>"
}
}
```