Export annotations

Download a queue's results as a file, or write them into a Future AGI dataset

Support quality review, the queue from Create a queue, now has completed items worth keeping. This guide walks through both routes, starting with the quick download and ending with a write into a dataset.

Download as JSON or CSV

Open Support quality review and click Export > Download in the header. Every item in the queue, any status, downloads immediately as a JSON file: one entry per item, carrying its annotations, review status, and the source’s own content resolved onto it.

Tip

The endpoint behind Download also accepts a CSV format, which flattens item, review, and annotation fields to one row per label value. It drops source, evals, source_id, item_notes, and annotation_metrics. There’s no format picker in the UI for it yet, so pull CSV directly through the API if you need rows instead of nested JSON; see SDK & API.

Note

Download tops out at 1,000 items. Push past that and it returns an error instead of a file, since there’s no status filter on this button to narrow the set first. For a queue that big, use Export to Dataset instead, which carries no such cap, or filter by status through the API. Self-hosted deployments can raise the ceiling with the ANNOTATION_EXPORT_SYNC_MAX Django setting.

What’s in an export

FieldWhat it holds
item_idThe queue item’s ID
source_typetrace, observation_span, trace_session, prototype_run, call_execution, or dataset_row
source_idID of the annotated source
statuspending, in_progress, completed, or skipped
orderThe item’s position in the queue
reviewReview status and reviewer, filled in once the item’s been reviewed
item_notesThe latest note left on the item
annotationsEvery label value submitted, with the annotator and score source
annotation_metricsThe item’s annotations keyed by label name
evalsEval scores already attached to the item’s source, if any
sourceThe resolved content of the source itself

Export to Dataset

Click Export > Export to Dataset in the header to open the export drawer.

  1. Choose Create new dataset and name it, or Add to existing dataset and search for one
  2. Set Items to export: it defaults to Completed only, and can widen to All items, or switch to Pending only or In Progress only
  3. Review the column mapping: each source field, label, and review detail maps to a dataset column, and you can rename, add, or drop columns before running
  4. Click Export

Unlike Download, Export to Dataset has no item cap, so a queue past 1,000 items still exports in full.

What you do with it

  • Fine-tuning: the annotated examples become training data for a model update
  • Eval datasets: completed items become a golden set you run other evals against

Dive deeper

Was this page helpful?

Questions & Discussion