OptionalbudgetOptionalconcurrencyCalls in flight at once (default 8).
OptionalexpectedWhere the right answers are, as in evaluate: a column for the answer, or {output: column}.
OptionalfeedbackWords per row (default: "right", "wrong: the right answer is …", or the call's error).
OptionalmetricA score per row, 1 meaning right (default: exact_match against the right answers).
OptionalminibatchFeedback rows per new instruction (default 4).
OptionalseedOptionalselectionRows to choose with, instead of half of the rows.
OptionalteacherThe model that writes instructions ("gpt-6-sol"); default the function's own.
Calls of the function at most (default 300).