Fig. 3. Fine-tuning scenario results.

a CARBO random split test set performance for different training strategies. In green are the top-1 accuracies of the models that were fine-tuned on either 1k or 20k CARBO reactions shown. For comparison, we included in purple the top-1 accuracies of the models trained on the single data sets (CARBO, USPTO, and USPTO_MIT). Blue are the performances of models trained in the multitask scenario. b CARBO time split test set performance for different fine-tuning set sizes. Source data are provided as a Source Data file.