You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
A thinking_timeout_s above the server's limit now raises a ValueError before the data is uploaded. (#438)
tabpfn_client.hosted.TabPFNClassifier and TabPFNRegressor accept tabpfn_systems to choose which systems the serving container runs around the model, for example [] to send the tables with only minimal cleaning. Older containers keep working: the field is only sent when set. (#439)
The hosted estimators can run each prediction as a job on the endpoint with use_async=True, so a long prediction no longer holds one connection open. (#440)
A thinking-mode TabPFNRegressor documents the outputs it returns: "mean", "median" and "quantiles", where "mean" is the prediction that scores best on thinking_metric. (#441)
Changed
thinking_timeout_s is now checked against the maximum of the chosen thinking_effort before the data is uploaded. (#445)