Trodo

Pull a dataset's rows as JSONL, with their output, expected, evaluations and label, for fine-tuning or offline evals: from the dashboard, a script with an export token, the SDKs, or a one-shot link from MCP.

An export is a dataset's rows as JSONL, one row per line, in the shape a fine-tune or an offline eval wants. Every pull is recorded in the export audit.

The row

{
  "id": "4f1c…",
  "dataset_id": "9a02…",
  "dataset_version": 3,
  "unit_kind": "span",
  "input": { "messages": [{ "role": "user", "content": "Where is my refund?" }] },
  "output": "Refunds take 5 to 7 days.",
  "expected": "Refunds take 5 to 7 business days after we receive the item.",
  "evaluations": [{
    "name": "Refund answers", "version": 2, "passed": false, "reason_code": "MISSING_CONDITION",
    "path": ["Judge"], "judge_reasoning": "…",
    "feedback": { "thumbs": "down", "corrected_passed": false, "reason": "Missed the condition" },
    "node_labels": []
  }],
  "label_source": "explicit",
  "label_status": "confirmed",
  "label_weight": 1.0,
  "source": { "span_id": "…", "run_id": "…" },
  "tags": ["refunds"],
  "metadata": {}
}

label_weight is how much to trust expected: 1.0 when a person wrote it, 0.6 for a confirmed proposal, 0.4 for an unconfirmed one, 0.2 for an accepted passing output. See Labels. A proposed row also carries its draft under proposal.

Which rows

Option	Values
`labels`	`all` (default) · `confirmed` · `explicit` · `derived` · `implicit` · `proposed` · `unlabelled` · `failed_only` · `passed_only`
`include_unlabelled`	With a label filter, also return the unlabelled rows.
`version`	A saved snapshot's number. Omit for the live rows.
`since`	Only rows created, edited or labelled since this ISO time.
`ids`	A comma-separated list of row ids.

For a fine-tune, labels=confirmed is the usual choice. Pin a version when the result has to be reproducible.

From the dashboard

Open a dataset and click Export. Pick the rows and the version, and download the JSONL. The dialog also shows the equivalent command for a script.

From a script: export tokens

A script reads exports with an export token, never with your site id. Create one in Settings → Datasets: leave the dataset list empty for every dataset, or pick some to pin the token to them. The token (trodo_dx_…) is shown once. Only owners, admins and developers can create or revoke tokens.

curl -H "Authorization: Bearer $TRODO_EXPORT_TOKEN" \
  "https://sdkapi.trodo.ai/api/sdk/datasets/support-failures/export?labels=confirmed" \
  -o support-failures.jsonl

The dataset can be named by id or by name. Export audit in the same settings page lists recent pulls: which dataset, which token or person, which rows, and how many.

From the SDKs

const { rows } = await trodo.datasets.export('support-failures', {
  token: process.env.TRODO_EXPORT_TOKEN,
  labels: 'confirmed',
});

out = trodo.export_dataset("support-failures", token=os.environ["TRODO_EXPORT_TOKEN"], labels="confirmed")
for row in out["rows"]:
    ...

From MCP

manage_dataset with action: "export" returns a one-shot link, never the rows: an agent hands you a URL instead of pasting your labelled data into a chat. The link works once, for 15 minutes, and the pull is audited as coming from MCP. It takes the same options as above.

What it costs

Exports are free.

Labels: what label_source and label_weight mean
Curate from production: fill the dataset first
MCP: the datasets tools

Export