Search instructions written by a model, with demo sets, and keep the best
(MIPRO-style, as Python's InstructionSearch): instruction candidates (the
current one plus proposals written by promptLm from the signature and a
few rows) × demo sets (bootstrapped, unless both demo limits are 0),
searched over trials minibatch evaluations; the top combinations are then
scored on every validation row and the best wins. The search is random
with greedy refinement, not Bayesian. Returns the best copy and every trial.
Search instructions written by a model, with demo sets, and keep the best (MIPRO-style, as Python's
InstructionSearch): instruction candidates (the current one plus proposals written bypromptLmfrom the signature and a few rows) × demo sets (bootstrapped, unless both demo limits are 0), searched overtrialsminibatch evaluations; the top combinations are then scored on every validation row and the best wins. The search is random with greedy refinement, not Bayesian. Returns the best copy and every trial.