molecule representation format: 1D SMILES
will further encode into 2D graph features
We use scaffold splitting to reconstruct the train-split. We use SMolInstruct RS train split as the sample pool.
We only include single-step retrosynthesis prediction.
For Detail, refer to PRESTO: Progressive Pretraining Enhances Synthetic Chemistry Outcomes:
https://arxiv.org/pdf/2406.13193
