BootstrapFewShotWithRandomSearch¶

BootstrapFewShotWithRandomSearch(
    metric=None,
    *,
    metric_threshold=None,
    max_bootstrapped_demos=4,
    max_labeled_demos=16,
    max_rounds=1,
    num_candidate_programs=8,
    num_threads=1,
    max_errors=10,
    teacher=None,
    seed=0,
    stop_at_score=None,
)

Try several sets of demos and keep the one that scores best on the validation rows.

Several candidate demo sets (none, labeled only, bootstrapped, bootstrapped from shuffled examples), each scored on valset (default: the trainset); the best one wins. candidates holds every candidate afterwards, as rows ({"candidate", "demos", "score"}; dpyr.read(opt.candidates) makes them a table).