BUYER REFERENCE
Your data.
Ready to work with.
Original recordings, structured labels and traceable provenance. Here is what each file contains and how to download a licensed order.
On this page
Files in a paid order
Your orders provides individual recordings, a JSONL manifest, a CSV index and a license summary. A full-recording analysis JSON is also available for recordings whose frozen order snapshot includes an active machine package.
- Original MP4 or WebM: the source recording, preserved as uploaded.
- JSONL: one record per licensed recording, including nested metadata and the full action-label snapshot.
- CSV: the same fields in a tabular index. Nested metadata and annotations remain JSON inside quoted cells.
- Analysis JSON, when included: model descriptions, recording-wide actions, per-segment analysis and provenance.
- License summary: the versioned, non-exclusive AI training and evaluation terms. Read the terms before licensing.
The order freezes its selected recordings, metadata and labels. Later uploads are not added automatically. A saved selection does not reserve footage or grant file access: availability is checked before payment, and downloads require a confirmed paid license that remains valid. If checkout is unavailable, discuss your data requirement with datai.
JSONL and CSV reference
The manifest format is datai.video.v1. It uses video_id and annotations. Times are milliseconds from the start of the original recording. The full exports include every licensed recording, even when the order screen shows a shorter preview.
| Field | Meaning |
|---|---|
schema_version | datai.video.v1, the order-manifest record version. |
order_id, video_id | The order and recording identifiers. |
contributor_id, split_group | Stable contributor identity and grouping. Keep a contributor’s recordings together when designing train/test splits. |
filename, content_type | Original filename and video media type. |
duration_ms, size_bytes | Licensed duration in milliseconds and byte size when available; size_bytes may be null on older snapshots. |
width, height, fps | Image dimensions and frame rate. Check metadata for how they were measured or declared. |
sha256 | Whole-file hash when available. Check metadata.sha256_verified and its source before treating a hash as independently verified. |
metadata | The frozen recording context, permission declarations, technical/checksum provenance and any analysis pointer. Fields depend on the source’s history. |
annotations | Full frozen action labels: start_ms, end_ms, verb, object, goal, success and note. success is success, failure or uncertain. |
license | The license version attached to this order. |
download_url | Authenticated original-video download. This is not a public or transferable access link. |
analysis_url | Authenticated frozen full-recording analysis JSON, or null when no active machine package was included. |
A complete manifest record
This example represents a contributor annotation snapshot without an active machine package; therefore analysis_url is null. JSONL writes each record on one line.
{
"schema_version": "datai.video.v1",
"order_id": "example-order",
"video_id": "example-video",
"contributor_id": "example-contributor",
"split_group": "contributor:example-contributor",
"filename": "example-task.mp4",
"duration_ms": 12000,
"size_bytes": null,
"content_type": "video/mp4",
"width": 1920,
"height": 1080,
"fps": 30,
"sha256": null,
"metadata": {
"task": "Repot a plant",
"environment": "Authorized potting bench",
"viewpoint": "fixed",
"annotations_source": "contributor",
"duration_source": "independent_review",
"duration_verified": true
},
"annotations": [
{
"start_ms": 0,
"end_ms": 12000,
"verb": "repot",
"object": "plant",
"goal": "Place the plant in a larger pot",
"success": "uncertain",
"note": ""
}
],
"license": "non-exclusive-ai-training-v1",
"download_url": "https://datai.nz/api/videos/example-video/content?download=1",
"analysis_url": null
}
Blank or missing signals are not invented. A whole-file hash may be absent or contributor-declared; independently measured hashes and server upload-part checksums are identified by their provenance in metadata.
Full recording analysis JSON
The full package uses datai.video-analysis.v1. Its recording_id corresponds to the manifest’s video_id. Download it through the paid order’s analysis_url; it is a different document from the JSONL manifest and the model clip response below.
| Field | Meaning |
|---|---|
schema_version | datai.video-analysis.v1 identifies the full recording package. |
recording_id | The recording identifier; it corresponds to video_id in the order manifest. |
category | The recording’s activity category. |
source | Measured SHA-256, byte size, media type, duration, dimensions and frame rate; measurement method and audio/original-preservation flags. |
provenance | Provider, requested and served models, schema/prompt/policy versions, job identifier, creation time, sampling settings and human_reviewed status. |
contributor_context | Contributor-declared task, environment, viewpoint, objects, tools, permission declarations and original contributor annotations. These are separate from model observations. |
description | Combined model descriptions of the recording. |
objects | Distinct model-observed objects across the segments. |
tools | Distinct model-observed tools across the segments. |
actions | Recording-relative action timestamps with verb, object, goal, outcome, note, segment_index and source. |
segments | Segment index, recording-relative boundaries, decoded duration, served provider model and the segment’s model analysis. |
limitations | Limits of the visual analysis and signals that are not supplied. |
acceptance | Automatic acceptance method, policy version, decision, reasons and human_reviewed status at generation time. |
Model action times in actions are recording-relative. Action times inside segments[].analysis.actions are relative to that segment. Source measurements, contributor declarations and model observations are kept separate. The automatic acceptance record describes the decision at generation time; it does not turn model flags into a human review or establish legal permissions.
Analysis uses sampled visual frames at 2 fps and does not analyze audio. Action boundaries are approximate. Measured robot states, motor commands, calibrated 3D pose, depth, force and segmentation masks are not included. Agree on any additional annotation layer before purchasing.
The model clip JSON Schema
Download the model clip JSON Schema. This validates a single model response for one analysis segment of up to 60 seconds. It is not a schema for the full recording package or the order manifest.
The clip response contains description, environment, viewpoint, objects, tools, actions, quality, privacy and uncertainties. It allows up to 64 actions; the server additionally checks timestamps against the actual segment duration. The full recording package preserves these responses under segments[].analysis and creates a recording-wide timeline.
Use the references above for delivered files. The published clip schema alone cannot validate an entire dataset delivery.
Download a whole order
The downloader is a readable Python script using only the standard library. You need Python 3.10 or newer, your buyer account credentials and an order with a valid paid license. Downloading the tool does not grant access to private recordings.
- Save
download-order.pyin a folder on your computer. - Open the paid order in Your orders. Expand Order reference and support and copy its full order ID.
- Open a terminal in the script’s folder. Replace the example email and
YOUR_ORDER_IDbelow with your own values.
python download-order.py --email you@example.com --order YOUR_ORDER_ID
On Windows you can use py -3 in place of python; on macOS or Linux, python3 may be the available command. The script securely prompts for your datai password. Do not put a password or recovery code in the command. Browser sign-in cookies are not shared with the script; it creates its own session and does not save the password or cookies.
By default, files are saved in datai-YOUR_ORDER_ID: the JSONL and CSV manifests, license.json, LICENSE.txt, original recordings, any included analysis JSON, and a local verification report in verified-local.jsonl. To choose another folder:
python download-order.py --email you@example.com --order YOUR_ORDER_ID --output ./my-dataset
To retrieve only the manifests and license:
python download-order.py --email you@example.com --order YOUR_ORDER_ID --manifest-only
Rerun the same command after an interruption to resume partial recordings. The script checks downloaded bytes against available server-part checksums and independently verified whole-file hashes, and records a local SHA-256. A contributor-declared hash is not presented as independent verification. Use --restart to discard partial transfers and download files again, or --help to inspect all options.
If access is refused, check the account, order ID and paid-license status in Your orders. Saved, unpaid or refunded orders do not grant downloads. Contact datai if you need help with a licensed order.
Define the right collection.
Specify activities, viewpoints and any additional signals before choosing your dataset.