Request data
For AI teams

The dataset you need, not the one on the shelf.

Labs, robotics companies and ML teams bring us a capability they're trying to build. We source the data from consenting creators and vetted experts, verify every item, and deliver it licensed and ready to train on.

Request a datasetSee how we verify
1 business dayFirst reply
Pilot firstBefore you commit
Signed receiptEvery item, by hash
An operator in a VR headset steering leader arms while a humanoid robot places a block at a lab bench
What teams use us for

Data for every stage of the model.

01Pre-training & mid-training

Licensed, human-made corpora in the domains your model is thin on, with clear rights for commercial use.

02Supervised fine-tuning

Expert-written demonstrations and reasoning traces, built to your rubric and reviewed by a second specialist.

03Preference & RLHF

Pairwise and rubric-graded comparisons from people qualified to judge the answer, not just its tone.

04Evaluation sets

Held-out, never-published test sets written by domain experts, so your benchmarks measure capability, not memorization.

05Robotics & embodied

Teleoperation and manipulation demonstrations, egocentric video and sensor streams, with success labels.

06Multimodal grounding

Images, audio and video with dense, expert captions and the releases to use people and places in them.

Engagement

Start small. Scale with proof.

You see real samples before anything scales, and every batch after that is held to the acceptance criteria you approved.

  1. Scoping call

    Thirty minutes with a data lead. We pin down the capability, the failure modes you're seeing, the volume and the licence you need.

    OutputWritten dataset spec
    TurnaroundUsually within days

  2. Pilot batch

    We source a small, representative batch from the catalogue and the expert network so you can test it in your pipeline before committing.

    SizeAgreed in the spec
    DecisionYours, item by item

  3. Production

    Sourcing scales to the full volume with the acceptance criteria from the pilot. Every item is verified before it reaches you.

    QualitySecond-expert review
    VisibilityProgress against spec

  4. Standing order

    As your model improves its needs shift. Standing orders keep new, matching work flowing from the same network at the same bar.

    CadenceWeekly, monthly or on demand
    TermsVolume pricing
Modalities

Whatever your model perceives.

Robotics & embodied01

Teleoperation, manipulation and sensor data

Demonstrations from trained operators, synchronized video, proprioception and force, with task success reviewed by robotics specialists.

Audio & speech02

Voices, accents, music

Consented recordings with speaker releases, transcripts and expert phonetic labels.

Text & reasoning03

Expert-written and expert-graded

Reasoning traces, rubrics, preference pairs and evaluations written by people who practice the field.

Image04

Original photography and art

Camera originals with releases, human-made origin checks and dense captions.

3D & CAD05

Geometry and scans

Meshes, point clouds and drawings with units and measured-versus-modelled labels.

Video06

Footage with context

Egocentric, instructional and cinematic video with frame-level fingerprints and scene annotations.

Documents & structured07

Records that are safe to train on

Contracts, filings and tables with schema, personal-data scans and masking before anything ships.

Delivery

Ready for your pipeline.

Delivered in the format you already train on, with labels, technical metadata, origin and the hashes to check every file.

A machine learning engineer reviewing rows of labeled room images on a wide monitor, a sketchbook open beside the keyboard
Format

JSON Lines & CSV

One record per item with labels, provenance and keyed download links.

Format

Parquet

Columnar exports for large tabular and caption sets.

Format

WebDataset

Sharded tar archives of files plus metadata, for streaming training.

Format

COCO & Croissant

Captions in COCO format and dataset cards in Croissant 1.0 JSON-LD.

Buying data

Before you ask.

How is pricing set?

By modality, the expertise involved, volume, exclusivity and licence term. Catalogue works carry the creator's training price; commissioned work is quoted against your spec. Larger orders get volume discounts.

Can we get exclusive rights?

Yes. An exclusive licence closes the work to other buyers for its term. Exclusivity is priced into the quote and written into the licence.

How do you handle personal data?

Documents and structured data are scanned for personal data and masked before delivery. Works showing identifiable people are only offered when consent for AI training is on file.

Can we integrate programmatically?

Approved buyers get an API key for searching curated selections, placing orders, posting requests and pulling exports, plus standing orders that deliver new matching work automatically.

Brief us. We'll go find it.

Five minutes to send a brief. A data lead replies within one business day with questions and a first sourcing plan.

Request a datasetTalk to us first