Inconsistency of LLMs in Molecular Representations

This dataset was curated based on the LlaSMol dataset, reaction prediction, and property prediction subsets. We augmented the original dataset by translating SMILES string representation into IUPAC name representation. The resulting dataset consists of one-to-one mapped SMILES and IUPAC representations. Below is the statistics of the dataset:

| Task                                 | #Train   | #Valid | #Test ||--------------------------------------|----------|--------|-------|| Forward reaction prediction (full)   | 963,567  | 1,956  | 300   || Forward reaction prediction (subset) | 76,379   | 1,956  | 300   || Retrosynthesis (full)                | 932,616  | 2,004  | 300   || Retrosynthesis (subset)              | 76,471   | 2,004  | 300   || Property - BBBP                      | 1,521    | 188    | 189   || Property - ClinTox                   | 1,063    | 127    | 131   || Property - HIV                       | 32,864   | 4,104  | 300   || Property - SIDER                     | 21,800   | 2,540  | 300   || Property - ESOL                      | 888      | 111    | 112   || Property - LIPO                      | 3,358    | 385    | 300   |