Label Studio Export: Predictions Are Not Reviewed Annotations
· Data Quality
A buyer handoff worksheet separates model predictions, saved annotations and acceptance evidence in a Label Studio JSON export.
A model prediction in a Label Studio export is not, by itself, a reviewed human annotation. Check which export fields carry predictions, which carry saved annotations and what independent review evidence supports the buyer's acceptance decision.
This guide is for annotation buyers and delivery teams. It does not describe job availability, worker eligibility, pay or a company hiring process. Data Annotation Companies is independent and is not affiliated with Label Studio or HumanSignal.
What does the official export format distinguish?
The official Label Studio export documentation, checked on 22 October 2026, describes annotations and predictions as separate JSON properties. It describes predictions as machine-learning results with an additional score field, and its example carries a model version. Similar result shapes therefore do not establish the same origin or review status.
The exact export depends on format, product edition and configuration. Inspect the actual file and the current documentation for the chosen route. Do not assume that every converted format preserves all provenance fields or that the presence of an annotations array proves independent human review.
Which fields should the buyer reconcile?
Agree the handoff boundary before counting labels. A task count, prediction count, saved annotation count and accepted task count can all describe different things.
| Record | Evidence | Buyer check |
|---|---|---|
| Task | Stable task and asset reference | Which input is this? |
| Prediction | Model result and version | Machine suggestion? |
| Annotation | Saved annotation and origin | What was submitted? |
| Review | Defined review decision | What was checked? |
| Acceptance | Agreed release decision | What is approved? |
| Export | Format and file revision | What provenance survived? |
These are proposed buyer control fields, not a promise that every edition automatically supplies a particular review workflow. Keep the actual supporting record where it lives and link it to the task rather than invent a status from a missing field.
How can an export count overstate completed work?
Illustrative example. A fictional three-task JSON handoff contains a prediction for every task. Task A has only a model prediction. Task B also has a saved annotation but no recorded buyer review. Task C has a saved annotation and a separate accepted review record.
| Task | Observed record | Release treatment |
|---|---|---|
| A | Prediction only | Not reviewed delivery |
| B | Annotation, review unresolved | Hold acceptance decision |
| C | Annotation and accepted review | Count under agreed policy |
The folder has three tasks and three predictions. Under the fictional policy requiring a saved annotation plus accepted review, only Task C meets that stated release condition. This is not a claim about automatic Label Studio statuses, an industry acceptance threshold or the quality of any real supplier.
What should a delivery team preserve?
Retain the source export, format, project configuration reference, asset mapping and model-version evidence where applicable. Reconcile missing and multiple records explicitly. An accepted prediction may have been used to create an annotation; preserve the actual provenance rather than label every saved result manually created.
If a downstream format drops review metadata, supply an agreed manifest alongside it. Check stable identifiers and the documented conversion rather than assume that a plausible training file proves the approved release was exported.
Does a high prediction score prove human review?
No. A prediction score describes the recorded model result, not a reviewer's decision. The probability calibration guide covers the separate question of what a confidence value can establish.
Does an annotations array prove independent review?
No. Check the origin and agreed review evidence. Saved output and buyer acceptance are separate states.
Should every export contain identical review fields?
No. Confirm the actual format, edition and configuration, and retain an agreed external manifest when necessary.
Compare a supplier on the handoff evidence
Use the company directory to research providers, then ask how their proposed delivery distinguishes machine suggestions, saved annotations and accepted output. Compare the documented process and sample handoff against your project requirements; a directory listing is not an endorsement or a guarantee.