datachad

Real-world training data, sourced and licensed for frontier AI labs.

datachad brokers and collects data that cannot be scraped: collected from consenting contributors in developing countries, documented row by row, and delivered ready to train on.

Three services, one standard of delivery

Brokerage

License datasets we have already collected and cleared. Size, coverage and every metadata column are published before the first call, so your team can assess fit without an NDA.

View the catalogue

Loading coverage.

Collection

When the data you need does not exist yet, we brief contributors in developing countries and collect it to your specification: video, audio, text or another type. Every collection starts with a pilot batch.

Send a brief
What we collect
Video
First-person and street-level footage, with GPS and motion data
Audio
Speech and conversation in local languages and accents
Text
Written and transcribed text in local languages
Something else
Describe it in the brief and we will scope it

Delivery

Licensed or commissioned, every dataset arrives the same way: video, audio or text files in object storage and one Parquet metadata row per sample, loadable into a PyTorch DataLoader.

See the metadata columns
import pyarrow.compute as pc
from torch.utils.data import DataLoader
from helmet_cam import HelmetCamClips

clips = HelmetCamClips(
    "s3://datachad-lake/datasets/sea-helmet-cam/metadata/",
    filter=(pc.field("city") == "Jakarta") & (pc.field("weather") == "rain"),
    num_frames=16,
    size=(224, 224),
)
loader = DataLoader(clips, batch_size=8, num_workers=8, shuffle=True)

for batch in loader:
    video = batch["video"]  # uint8, [8, 16, 3, 224, 224]

Documented for legal and procurement review

Provenance is part of the data, not a separate PDF. These columns ship in the metadata of every dataset.

Consent on every row

Each sample records which version of the consent agreement its contributor signed.

consent_version

Privacy processing on every row

Each sample records the privacy processing applied to it, such as blurred faces and number plates in video.

faces_blurred, plates_blurred

Pseudonymous contributors

Contributors appear only as stable pseudonymous IDs, which also lets you split train and test by person.

driver_id

How a collection runs

  1. 01
    Brief

    You set the data, places, conditions, volume and deadline.

  2. 02
    Pilot

    We collect a small batch so you can check quality before you commit.

  3. 03
    Collection

    Contributors collect to the brief while we track coverage against it.

  4. 04
    Checks

    We review what comes in, drop what misses the brief, and remove personal data.

  5. 05
    Delivery

    You receive the files, the Parquet metadata table and the licence.

Talk to us about your data roadmap.

Ask for a sample pack and licence terms, or send a brief for data you want collected.

Contact us