Table 7:

Multi-task testing results with T5Chem on USPTO_500_MT. The five tasks are grouped into two subgroups. The first subgroup includes forward reaction prediction, single-step retrosynthesis and reagents prediction, as they are all sequence-to-sequence tasks and can share the same model architecture. The second subgroup consists of reaction classification and reaction yield prediction. Both tasks take the whole reaction sequence as inputs, and can be trained at the same time by combining their loss functions together. Only one model was trained per group with the mixed dataset.

Task Type	Forward	Retrosynthesis	Reagents	Classification	Yield
Metrics	Top-1 Accuracy	Top-1 Accuracy	Top-1 Accuracy	Accuracy	R
Results	97.5%	72.9%	24.9%	99.4%	0.46