Training results

Two numbers per cell, for the newest run of each dataset and model: the share of an oracle's gain captured by a hundred picks, then overall correlation. 851 finished runs. The networks page says what each model and test set is, and lets you edit them.

Bold is the best model per dataset, by the first number. That number is the share of an oracle's gain captured by the model's best hundred picks, on the validation set: 1.0 is as good as picking by the measured scores, 0.0 is no better than picking at random. The second is Pearson over every variant, on the test set. They can disagree completely: on these datasets a single dead-or-alive threshold reaches a higher Pearson than the networks do, so a high correlation mostly says a model tells broken variants from working ones. The first number is the one that reflects picking candidates worth making — see the mutation scan.

All runs