Unraveling overoptimism and publication bias in ML-driven science