For AI labs

Built to your specification, tested against your models.

Tell us the skill you want to measure or improve. We build the tasks, run them against frontier models and deliver what they fail on.

What we deliver

Three formats, one standard.

Training data

Expert demonstrations, trajectories and deliverables for supervised and preference training.

Evaluation sets

Held-out tasks with gold answers or rubrics, scored the same way every time.

Environments

Reproducible desktop and tool setups with automatic checkers for agent training.

Difficulty

Difficulty is measured, not guessed. Every task is run against named frontier models before delivery.

  • Tasks that current models solve easily are revised or dropped.
  • Pass rates per model ship with the data.
  • You can name the models and the target pass rate.

Quality

Two-stage expert review on every item.

  • A second expert reviews each item against the specification.
  • Automated checks mirror your own acceptance tests.
  • Rejected items are logged with the reason.

Rights

Clean provenance you can show your legal team.

  • Original work by our experts, or data licensed from its owners.
  • Commercially licensed or open-source tools only.
  • IP assigned to you under the agreed licence.

Formats and delivery

Your schema, your cadence.

  • Weekly batches at the agreed volume.
  • Delivery by API or to your repository or bucket.
  • A datasheet and a changelog with every batch.
The process

Samples first. Volume after.

  1. Specification

    A call to agree the skill, format, models and acceptance tests.

  2. Samples

    5 to 100 items with a datasheet, for your team to review.

  3. Model results

    Pass rates for the models you named, item by item.

  4. Delivery

    Weekly batches until the agreed volume is reached.

Request samples

Tell us what you need.

We'll come back with a sample plan: what we'd build, how many items, and how fast.

Which datasets interest you?

We use these details only to reply. See the privacy policy.