Download the processed dataset

Streams the processed rows in the requested file format. Works on datasets with status ready (full output) and also on datasets with status failed — in the failed case the stream contains only the rows that were successfully processed before the run aborted, which is useful for recovering partial output from interrupted jobs. Returns 422 only when no run has ever started on this dataset.

Image columns (JSON / JSONL): Rows may include original_image only when the mapped image column is also in configured_column_mapping.context. Read image_column_formats from GET /datasets/{dataset_id} and look up the active image column to determine encoding: embedded_bytes → { bytes, path: null } with base64 bytes; url or file_reference → presigned HTTPS URL string. When images are not surfaced, original_image is omitted and internal s3:// URIs are never returned. See schema ProcessedDatasetExportOriginalImageFieldDto. In CSV and Parquet, embedded bytes appear as a JSON string in the cell.

Response body: binary in every format. fileFormat=parquet returns a gzipped tar archive of one or more Parquet shards (.tar.gz), not a single .parquet file — extract it before reading. CSV, JSON and JSONL bodies are UTF-8 text and can be decoded directly.