Reviewing a preview
The dataset detail page at Preview ready — what each sample shows, how to approve, how to send feedback.
This section will include screenshots.
CLI equivalents: datagen datasets preview <id>, datagen datasets approve <id>, datagen datasets feedback <id>, datagen datasets watch <id> — see CLI reference → datasets.
The preview is the one moment in the flow that genuinely benefits from a browser. You're making a judgment call about whether the shape is right. Three example rows, two actions. The rest of this page walks through exactly what you see.
Getting here
When a dataset reaches Preview ready, the detail page shifts. The progress panel at the top stays — you can see which waypoint you're in — and a new section appears mid-page with three example rows.
The waypoint pill on the list view also changes: datasets at Preview ready or Your review surface in a warmer tone so you can pick them out at a glance.
What the three samples show
The Preview section is headed 3 example rows with a subtitle reminding you that this is the exact shape your delivered dataset will have. There's no shape drift between preview and delivery — what you approve is what you get.
Each sample renders as its own card. Within a card, you see the fields a row of your dataset will contain — for example, an instruction, any context the task depends on, a reference solution, and the grading criteria. The field names and structure depend on the shape the Factory chose for your brief; the renderer adapts per dataset.
What you're looking for, page by page:
- Is the instruction specific and unambiguous? Could a reasonable agent misread it?
- Is the grading coherent? Do the criteria actually test what you asked for, at the granularity you care about?
- Does the reference solution match the instruction? A mismatch is the clearest signal that something's off.
- Is this the shape you'd want to train on 1,000 times? The preview is three rows; the full dataset is the same shape at scale.
See Reviewing your preview for the deeper treatment of what good looks like, with examples.
Approving
A sticky bar at the bottom of the page holds two buttons: Approve and Send feedback.
Approving is one click. The dataset moves to Generating full dataset and the Factory begins producing the full set of rows. You'll see a confirmation toast; you can close the tab. If you've turned on email notifications (the preference lives on the same page), we'll email you when the download is ready.
Approving is not reversible in the UI today. If you notice something wrong after approving, the right move is to contact us.
Sending feedback
Send feedback opens a dialog with a textarea. Plain English is what we want; the dialog shows three example phrasings to calibrate tone:
"The questions are too simple — I need harder ones." "Too focused on urban scenarios — more rural please." "The output format should be a single paragraph, not a list."
Useful feedback is specific and names a target. Useless feedback is vague.
- Useful: "The grading criteria for regulatory accuracy should require citation of the provided source documents, not just assert facts. A task isn't successful if claims are unsourced."
- Less useful: "Make it more rigorous." (rigorous how? at what step?)
- Useful: "These instructions are too short. A real user would give four or five paragraphs of context, including a constraint or two."
- Less useful: "Longer instructions."
You can submit feedback any number of times; each round produces a revised preview.
Cmd/Ctrl+Enter submits from the textarea. When you submit, the dataset goes back to Designing your dataset — the Factory re-runs the design step with your note applied — and the preview refreshes when a new one is ready.
The waypoint banner over time
Watching the page as the dataset moves:
- Understanding your ask → the progress panel lights up the first step. No preview yet; the brief shows near the bottom of the page so you remember what you asked for.
- Designing your dataset → second step active. Still no preview. If you submitted feedback, this is where your dataset is until the next preview is cut.
- Preview ready → the Preview section appears with three cards. The sticky bar materializes at the bottom with Approve / Send feedback.
- Your review → same as Preview ready from the customer's side. Nothing's changed for you; the status is just tracking that you've seen it.
- Generating full dataset → the sticky bar goes away. The preview stays visible so you can see what you approved. A Your dataset is ready card will appear once generation completes.
- Complete → a download card surfaces with format options (Parquet, JSONL, or CSV), plus a link to the dedicated download page with first-rows preview and a share link.
- We're still refining this → a banner at the top notes that a TLDC engineer is taking a closer look. No action required. The brief is still there; the Approve button is disabled with an explanation.
Other things on the page
- Brief card. A read-only copy of the brief you submitted, near the bottom of the page. Useful to anchor yourself during feedback.
- Notification preferences. A single toggle for "email me when ready." Defaults on.
- Started timestamp. Small and grey, under the title.