Figure 3:

Mix training dataset from different tasks. In multi-task training scheme, we combined forward reaction prediction task, retrosynthesis task and reagents prediction task together. Every input instance starts with a task-specific prompt, then followed by actual input. In forward reaction prediction, the model takes reactants and reagents (without separation) as source sequence. In retrosynthesis, the model takes only product SMILES as source sequence. In reagents prediction, the model takes both reactants and product SMILES as inputs. A reduction reaction is shown above as an example.