The editor
Every save appends to
labels.tsv and moves the row’s status from pending to validated; Flag marks a
row that needs a second look. Model drafts sit in the editor as a starting point and are never overwritten by
the pipeline once a person has saved.
Keyboard, scoped to the editor so it never fights the browser:
A sound dataset has keys of its own, because there a clip is answered with one key:
1 to 9 pick an answer,
Space plays the clip again, X plays only the sound, a track’s first letter silences it, and ← → go back
and skip (Sound datasets).
Finding rows
Filter the list by status (pending, validated, flagged), page through it, or jump to a row by number or id. The dataset header shows counts per status so a lead can see how far a batch has come.Spreadsheets
Some teams annotate in Sheets or Excel. Export downloads a TSV of every row with its current label and status; edit anywhere, then Import the file back. Thevalidated and flagged columns in the imported
file take precedence over the exported status, so a reviewer can flip states in bulk.
Your own export holds the text as it is. Under the dataset’s
Flag or Mask policy it gains pii_findings and
pii_types columns, and a share link’s export under Mask holds the masked text.
Where the data is
rows.tsv and labels.tsv (plus clips) live under the dataset’s folder on the datasets volume. Mount the
volume in a function to train on it, or download with levers9 cp datasets/<name> ..