Verus-Bench / docs /DATA_SCHEMA.md
SuperCarryZZR's picture
Release Verus-Bench
35dc3e8 verified
|
Raw History Blame Contribute Delete
6.16 kB
# Data schema
One row is one sample. `metadata/test.parquet` contains 192 rows in the order recorded in
`sample_ids.txt`. `ground_truth.jsonl` provides the labels and findings separately.
`metadata/inputs.jsonl` contains GT-free model input records.
## Test table
| Field | Meaning |
| --- | --- |
| sample_id | Unique sample identifier. |
| number | One-based sample position, 1 through 192. |
| source_dataset | hjy_generated, hjy_natural, nextqa, onethinker_didemo or onethinker_stvg. |
| label | fake: artifact-positive; real: artifact-negative under the benchmark annotation. |
| findings | Ordered human GT findings; empty for real samples. |
| release_mode | VERUS_NONCOMMERCIAL, HOSTED_WITH_ATTRIBUTION, HOSTED_NONCOMMERCIAL, REFERENCE_ONLY_ND, REFERENCE_ONLY_UNVERIFIED or REFERENCE_ONLY_ATTRIBUTION. |
| license | Verus policy for owned material, otherwise the recorded original media license. |
| frames_available_in_release | Whether this repository distributes this sample's frame images. |
| frames | Ordered frame descriptors, including descriptors for omitted images. |
| video_path | Dataset-relative source-video path when distributed; otherwise null. |
| video_available_in_release | Whether this repository distributes the source video. |
| source_window_seconds | First and last sampling-grid times; the second is not an exclusive clip end. For third-party media the window uses the upstream file's timeline. HJY records retain their recorded input sampling window, including frame-only samples. |
| original_media_id | Original external-media ID; null for HJY samples. |
| source_sample_id | Upstream dataset identifier; null for HJY samples. |
| original_url | External source URL; null for HJY samples. |
| reconstruction | Third-party extraction settings; otherwise null. |
Release availability flags describe the distributed package and do not change when a user adds
missing media locally. Missing media does not change the label or findings.
## Frames
`metadata/frames.jsonl` contains `sample_id`, `frames` and `reconstruction`, without GT.
Each frame descriptor contains:
- `frame_index`: zero-based sequence position.
- `path`: dataset-relative JPEG path when distributed, otherwise null.
- `expected_path`: relative destination path, also provided for omitted media.
- `timestamp_seconds`: timestamp in seconds relative to the sample start.
- `width`, `height`: dimensions of the supplied benchmark image before model preprocessing.
- `available_in_release`: whether the image is distributed in this package.
A locally re-extracted image may differ in dimensions or encoding if the source version differs.
Preserve the recorded ordering, sampling times and sample window when preparing inputs.
## Ground truth
Each row of `ground_truth.jsonl` has `sample_id`, `label` and `findings`.
A finding contains:
- `finding_id`: per-sample identifier; not globally unique.
- `taxonomy_level`, `artifact_tag`: category strings.
- `observation`: human GT description.
- `timestamp_range`: supplied start and end timestamps in sample-relative seconds.
- `bbox`: overall XYXY box when annotated, otherwise null.
- `frame_boxes`: ordered `{time, bbox}` entries. `time` is in sample-relative seconds;
`bbox` is `[xmin, ymin, xmax, ymax]`, normalized to [0,1] relative to the full input image.
Multiply x by image width and y by image height to obtain pixel coordinates.
Preserve multiple boxes at one time. Separate findings may overlap in time. REAL samples have
`findings: []`. Historical annotations do not specify whether temporal endpoints are closed or
half-open; preserve their values and follow the evaluation code's convention.
## Model input records
`metadata/inputs.jsonl` is a 192-row release index with dataset-root-relative
image paths, null paths for omitted media, and an empty `metadata` object.
It is an index, not a directly runnable inference manifest.
The preparation script exports runnable manifests with exactly three fields:
```json
{"sample_id":"example","images":["../frames/example/frame_0000.jpg"],"timestamps":[0.0]}
```
Exported paths are relative to the output manifest directory. Images and
timestamps correspond by position. By default, export requires all 192 samples
to be complete. Explicit `--available-only` exports only complete samples and
reports partial coverage. Neither export contains GT. The companion inference
scripts accept these generated files through `--input-jsonl`.
See `SETUP_ALL_DATA.md` for commands.
`metadata/samples.csv` contains `number`, `sample_id`, `source_dataset`, `label` and `release_mode`.
## Reconstruction settings
- `method`: `ffmpeg_fps4_q2`.
- `source_filename`: filename in the upstream dataset.
- `start_seconds`: extraction start in the upstream file's timeline.
- `first_and_last_sample_seconds`: first and last sampling-grid times in that file.
- `fps`: 4 frames per second.
- `frame_count`: target sequence length.
- `encoder_quality`: ffmpeg JPEG quality setting, 2.
The helper uses these settings with `tpad=stop_mode=clone:stop_duration=0.25` and the target frame
count. Upstream files may already be clips. Filename matching does not establish that two videos
have the same content or timeline; users should select the corresponding upstream version.
## Provenance CSVs
`THIRD_PARTY_DATA.csv` contains 72 external sample records. `metadata/reference_media.csv`
contains the 26-record subset whose media is omitted. Fields:
- Identification: `verus_id`, `source_dataset`, `source_sample_id`, `source_filename`, `original_media_id`.
- Attribution: `original_title`, `creator_name`, `creator_account_id`, `creator_url`, `original_url`.
- License and distribution: `original_license`, `license_url`, `release_mode`, `hosted_by_verus`, `frames_hosted`.
- Sample window and transformations: `sample_window_seconds`, `window_reference`, `modification_note`.
- Evidence: `attribution_text`, `attribution_status`, `license_evidence_url`, `notes`.
Windows are JSON lists of seconds inside CSV cells; booleans are `true` or `false`. Unknown
names or titles are empty. These records describe media terms without replacing upstream licenses.