COMPRESS Decoder Training Data: GEOM-Drugs Molecules with Multi-Resolution COMPRESS Representations

Training data for the COMPRESS decoder, a flow matching model that reconstructs all-atom molecular structures from COMPRESS representations.Built from 11K (55K conformers) GEOM-Drugs molecules. data.pt is a list of Python dicts (load with torch.load), one per molecule/conformer:- AA: all-atom reference structure (positions, charges, atom types, bond orders)- K_all: COMPRESS representations at every resolution K=1..M (site positions, charges, Lennard-Jones parameters)- M: atom count (ranges 20-94 across the dataset) Code: https://github.com/min456min/COMPRESS-Decoder