Skip to main content
Annotators are often outside the company and have no dashboard account. Share on a dataset works like a document share dialog.

​
People with access

Add a named link per annotator, with a role: Each link is unlisted and revocable on its own. The name you give it is stamped as user on every label that comes through it, so you can see who labelled what.

​
General access

One link for everyone: Restricted (off) or Anyone with the link, with a role. Use it for a class or a crowd; use named links when you need per-person accountability.

​
What the annotator sees

The link opens a page with no sidebar: a name prompt (remembered in the browser), then the same annotation workspace as inside the dashboard, including export and import. Nothing else in the workspace is reachable from it. A revoked or expired link shows a plain message. A link reads the rows of its dataset with their labels, plays each row’s clip, and exports them as TSV (plus the gold set and the schema for a decision dataset). An annotator link can also save labels and import an edited TSV. Which text a link reads is the dataset’s policy, and the Share dialog says which one the links are under: A link follows the policy as it is when the link is opened, not as it was when the link was made, so switching to Mask covers the links you’ve already handed out. Under Mask a link never gets the original as a fallback: a row whose text changed since the last scan shows held back until the personal data scan has run until you scan again. A label an annotator saves is read for personal data on the spot, so they see it back masked. Where someone has reviewed a row, a link reads it masked with what the reviewer settled: a name the scan missed and they marked is [NAME] for every link. Mask covers text. Audio still plays as recorded, so a link that plays a clip plays the recording, names and numbers included. And it’s about what a link reads: you still read the original on the dataset page. That is all of it. The only files a link can open are the clips the rows play, under chunks/ for speech-to-text and audio/ for text-to-speech. It can’t download anything else from the dataset’s folder: not the original recordings in raw/, not rows.tsv, labels.tsv or meta.json as files, and nothing under pii/, which holds the phone number and name each call is known by and what the personal-data scan wrote. Asking for one of those through a link gets a 403, whichever role the link has. A clip is always sent as sound. A link plays files ending in .wav, .mp3, .ogg, .oga, .opus, .flac, .m4a, .mp4, .aac, .weba, .webm, .amr, .3gp or .caf, with the type that extension names, whatever the file holds. Any other file in those folders gets a 403, so a page saved there can’t open as a web page for whoever holds the link. Before an agreement is accepted, the link’s page learns the dataset’s name, kind, language and status, and how far the work is (rows, calls, hours, and how many are validated, flagged, pending and drafted). The config the dataset was built with (its source, bucket, query and secret names) isn’t sent to a link. The links above are for the dataset’s rows and labels. There is one other kind, made on the Anonymization card with Review link and not in the Share dialog: a link for reviewing a sample of the personal data scan. It’s a different thing, and it’s kept apart on purpose: The reviewer has to read the original, because the job is to mark what the scan missed. So a review link is narrow instead: a few hundred rows drawn at random and nothing else of the dataset. It can be taken back on the card, and it can’t be turned into a link of the other kind. Asking any other route of a link for data with a review link gets a 403, and so does asking the review routes with a link from the Share dialog. Over the API a link’s kind is its task: labels (the default) or pii. POST /api/datasets/{id}/shares with {"label": "Fatima", "task": "pii"} makes a review link. Add "mode": "view" for one whose holder reads the review and can’t change it; the card makes links that can review.

​
NDA

Open Agreement in the Share dialog to require one. Write the title and text (a template is provided) and save. From then on:
  1. an annotator opening any link sees the agreement first, and enters a name, organisation and email to accept;
  2. the acceptance is recorded on the dataset with the agreement version;
  3. changing the wording bumps the version, and everyone accepts again before they can continue.
Data endpoints behind a link refuse to serve rows until the current version is accepted, so the gate holds even for someone using the API directly. A review link shows the agreement whether or not you require it here, since its holder reads rows as they are. Write one and it’s yours they accept; leave it empty and they accept a short default text that says what they’re about to read and what they mustn’t do with it.