Fig. 2From: Probabilistic generative transformer language models for generative design of moleculesPercentages of valid, unique@10000, intDiv and Scaf/TestSF samples generated by the SMILES atom tokenizer models saved over the training process. The models generate few valid SMILES strings in the beginning. As the training goes on, the models gradually gain the capability to generate chemically valid SMILES molecules compositionsBack to article page