Skip to main content

Create an experiment

POST 
/api/v1/experiments

Starts an experiment asynchronously and returns a W&B run id to poll.

Supply the research question and a supported population selection, then review the design, response and reproducibility options in the request schema. The REST Experiment model and MCP's typed advanced object are different contracts; not every REST option is exposed through MCP or the browser form.

Poll the returned run identity. Normal execution queues the run; debug mode can execute inline. A definite submission rejection returns an error. Keep polling in-queue, running and unknown; a missing W&B record or timed-out read is not a reason to submit another experiment. REST does not provide the MCP launch-idempotency protocol. For a failed or ambiguous launch, reconcile the original run before retrying.

population_traits do not select a population. Traits add modeled attributes to respondents selected by target_population or another supported population mode. Always supply an explicit population selection alongside trait enrichment. A trait label is not evidence of observed behavior in the sampled population.

check_for_continuous_variables only affects engine-generated designs. When pre_cooked_attributes_and_levels_lookup is supplied, the submitted levels are used exactly as sent and the flag has no effect: nothing is reclassified as continuous and no replacement values are drawn. Omit the levels to let the engine generate the design if continuous inference is wanted.

Request​

Responses​

Successful Response