Astro AMASE Molecule Fragment Prediction Training Data

Training and validation datasets used to train the SubgroupMLP model in Astro AMASE (https://github.com/zfried/astro_amase), a Python package for automated molecular line identification in radio astronomical spectra. Each entry contains three arrays: smiles (SMILES string of the molecule), vicgae (32-dimensional VICGAE chemical embedding vector), and outputs (326-dimensional binary label vector indicating which molecular fragments from the Group SELFIES grammar are present in the molecule). The combined dataset contains 148,274 unique molecules split 80/20 into training and validation sets. The vicgae vectors are the raw outputs of the VICGAE encoder with no further preprocessing. Also included are grammar.txt and grammar_smiles.txt, which define the 326 molecular fragments corresponding to each index of the outputs vector. grammar.txt contains the Group SELFIES fragment definitions used by the GroupGrammar decoder, and grammar_smiles.txt contains the corresponding SMILES representation of each fragment. The index of each entry in outputs corresponds directly to the same-indexed line in both grammar files.