Generate outputs

Generating outputs sends each row's input to a Model with your system prompt and writes the reply into the row's output. A run scores outputs, so a dataset needs them first, unless the run generates its own.

Start a generation

Generate outputs on the Rows toolbar asks for a text Model from the project, a system prompt sent with every row, and which rows to fill: all of them (overwriting existing outputs), only rows with no output, or the rows the last generation failed on.

The dialog has no generation parameters and shows no cost estimate. To set max tokens, temperature, thinking budget or reasoning effort, generate from New run: choose Generate outputs first, and its Hyperparameters are sent with the generation. Picking rows by id is API-only.

The job

Generation makes one provider call per row in batches of 10, and writes each batch back before the next, so a stopped job keeps the rows it finished. Stop generating takes effect between batches, and Run in the background closes the dialog while the job continues. The dialog shows progress, refused rows, and tokens and cost so far. A Model with no price on file shows the cost as unknown, never zero.

A refused row counts as failed and keeps its old output, so one bad row never throws away the rest. A job where every row failed ends failed. One generation job runs per dataset at a time, and a second is refused with 409. After a service restart, an unfinished job is marked failed with "Interrupted by a service restart", and the rows it wrote stay.

Afterwards

To regenerate only the refused rows, open Generate outputs again and choose to retry the rows that failed. Outputs are not part of the content hash, so runs stay comparable after a regeneration. The dataset keeps no copy of the Model, prompt hash and parameters the job recorded, so record them under Hyperparameters when you start a run.