Export format

A run’s export is a JSON array with one entry per item in the run. Each entry records how one item was labeled. Its shape depends on how far the item got when you started the export. To produce and download the file, see Export your labeled results.

A completed item carries its inputs and outputs, plus per-node metrics. Here is one from a support-ticket run:

{
  "index": 4,
  "status": "completed",
  "input": { "ticket_text": "My invoice is double the usual amount this month." },
  "output": { "intent": "billing" },
  "metrics": [
    {
      "nodeName": "classify_intent",
      "nodeType": "generation_sft",
      "ahtMs": 42318,
      "avgLlmQaMs": 8125,
      "attempts": 1,
      "annotatorId": "a3f9c1e08b"
    }
  ]
}

An item that had not started when you exported carries only its index and status:

{
  "index": 12,
  "status": "idle"
}

Entry fields

Every entry is a JSON object. index and status are always present. The rest appear as the item makes progress.

Field	Type	Present when
index	number	Always
status	string	Always
input	object	The item has started
output	object	The item has completed
metrics	array	A Generation or Quorum node has finished

index is the item’s index in the source dataset (counting from 1). A run that launches a filtered subset carries just those items’ indexes, so the numbers can be sparse. status is one of the status values below.

input holds the item’s input fields. output holds the fields the End node returns. See Fields and data for what these fields hold and how a Quorum node’s output comes back as a per-expert list. metrics carries per-node timing and labeling details. See The metrics array. Each of these is omitted until the item reaches the state in the table.

Status values

Value	Meaning
completed	The item ran through the pipeline and produced its output.
failed	The item stopped at a node it could not complete, or on entry when a required input was missing.
cancelled	The run was canceled while the item was still unfinished.
paused	The run was paused while the item was still unfinished.
in_progress	The item is still being labeled.
idle	The item had not started when you exported.

Note

The file spells the canceled state cancelled, with two Ls. The Items tab in the app shows the same state as a Canceled chip, with one L. When you search your downloaded file, use the file’s spelling: cancelled.

The metrics array

The metrics array has one entry for each Generation or Quorum node the item ran. Each entry records how that node handled the item. Code and Match nodes never produce an entry. Each entry names the node (nodeName, the node’s name as shown in the pipeline graph) and its kind (nodeType). A Generation entry describes single-expert labeling. A Quorum entry describes several experts, so it drops attempts and annotatorId and carries annotatorIds instead:

{
  "nodeName": "rate_helpfulness",
  "nodeType": "quorum_sft",
  "ahtMs": 51044,
  "avgLlmQaMs": 10233,
  "annotatorIds": ["a3f9c1e08b", "b7e2d4f19c", "c1a8f3e05d"]
}

Field	Type	Appears on
nodeName	string — the node’s name in the pipeline graph	Every entry
nodeType	"generation_sft" or "quorum_sft"	Every entry
ahtMs	number	Every entry
avgLlmQaMs	number or null	Every entry
attempts	number	Generation entries
annotatorId	string or null	Generation entries
annotatorIds	array of strings	Quorum entries

How the metrics are computed

• ahtMs — milliseconds from the item’s first task starting to the node first finishing.
• avgLlmQaMs — the rounded mean of the node’s review durations, in milliseconds; null when review never ran.
• attempts — how many times the task was submitted or expired (Generation entries; under synthetic labeling the submissions are the model’s).
• annotatorId — the expert behind the last submission or expiry; null when the node labeled synthetically (Generation entries).
• annotatorIds — the experts who labeled the item on this node (Quorum entries). The list can be longer than the node’s configured overlap when review rejections bring in extra experts.

When a node is missing from metrics

• A node appears only once a task on it started and the node then finished for the item — a node that failed before any task began is left out. A node the item never reached, one still running when you exported, or one that was canceled is left out.
• Code and Match nodes never appear, even when the item ran them. Only Generation and Quorum nodes produce metrics.
• Every entry reflects the item’s state at the moment you started the export, like the rest of the file (see Export your labeled results).

Related

• Export your labeled results — how to produce and download the file
• Fields and data — what the input and output fields hold
• Quorum node — why annotatorIds can outnumber the configured overlap