Exporting

Any dataset whose run is done, stopped or failed can be exported. The format is chosen at export time, whatever the source was: CSV, JSONL or Parquet. All columns go, in the dataset’s order.

Download

From the dataset’s menu, or GET /v1/datasets/{id}/export?format=parquet. A SQL query narrows the rows, so an export can be a filter on an answer column:

SELECT * FROM data WHERE "delivery__jev-1.14" = true

The query picks rows only; it must return the dataset’s row id (SELECT * does), and the columns it lists make no difference.

To a destination

Write the file where the data lives instead: a folder on the server’s disk (locally, with ROWLAB_SERVER_FILES on), or an s3:// path through a storage connection. A destination ending in / gets <dataset>.<format> in it; otherwise it is the file’s name, replacing a file that is there.

Over MCP, agents export with dataset_export, with the same options.

Row ids

rowlab gives every row an id of its own to join answers to rows. Exports leave it out by default; include it (include_row_id) when you will join an export back to a later one.