Creating a dataset
The New Dataset flow — what the form asks for, and what happens after you submit.
This section will include screenshots.
CLI equivalent: datagen datasets create — see CLI reference → datasets.
Creating a dataset starts with one sentence; sometimes a paragraph. The job of this page is to describe the screen, show what happens when you submit, and signpost where you end up next.
The New Dataset screen
From the datasets list, Describe your dataset opens a single-page form. There are two fields and a panel of examples.
- The brief. A large text area, labelled Describe your dataset. Plain language is the whole point — describe the task, who the user is, what one example row should look like. A character counter sits under the box; it shifts tone as you go ("Nice start", "Great detail — that helps us shape it well") but there is no enforced length. One paragraph is usually enough. The box is the only required field.
- Optional spec link. A URL field for a doc you'd send a teammate — Notion, a Google Doc, anything linkable. If you paste one, we read it alongside the brief.
- Example briefs. A panel below the textarea offers a few worked examples. Click one and it fills the textarea so you can edit rather than start from a blank page.
The Factory designs the dataset from your brief and any Resources you've registered — you don't pick a format or template up front. If the design doesn't fit, say so in feedback once the preview is ready; a revised preview comes back.
There is also no Resource selector at submit time. Any Resources you've registered in the Resources tab are visible to the Factory and bind automatically when your brief describes something they fit. This is deliberate: you describe the dataset, not the plumbing.
Your draft auto-saves as you type. If you close the tab and come back, a small banner offers to pick up where you left off, or start fresh.
What happens when you submit
Start designing kicks off a dataset. You land on the detail page for the dataset you just started, with a progress panel at the top showing waypoints. The list of datasets will also show the new entry when you go back, most recent first.
The waypoints are the same five you see throughout the product:
- Understanding your ask — we're reading your brief.
- Designing your dataset — we're figuring out task shape and grading.
- Preview ready — three example rows are ready for you to look at.
- Your review — the decision is yours; approve or send feedback.
- Generating full dataset — once approved, we build the full thing.
The detail page polls for updates; you can leave the tab open, or close it and come back. There's a notifications preference on the page (email on ready) if you'd rather not babysit.
If something goes sideways and the automated path can't finish cleanly, the dataset moves into a We're still refining this state — a TLDC engineer takes a closer look, and you'll hear back from us. No action is required on your side.
Next step
When the dataset reaches Preview ready, you have three example rows and two choices: approve, or send feedback.