Datasets for Diversity Beats Size Scaling for Chemical Language Models

mols/ contains PubChem IDs for all source molecules. ind* denotes training sets, val* validation sets, and test.txt the test set.

pairs/ contains molecular pairs (excluding self-pairs) and Tanimoto similarities in the form of HDF5 files.

priors/ contains final model checkpoints.

data_pipeline.toml is the configuration file used for pre-processing using the REINVENT4 data pipeline (version 4.5).

custom_loader.py is a custom dataloader that was used along with the training code from this GitHub repository.

Data in the HDF5 files can be accessed like so:

import h5py

f = h5py.File("pairs/val161K/final_0.h5py")

# Keys are PubChem IDs
keys = f.keys()

# Loop over all pairs
for key in keys:
    data = f[key]
    for pair in data:
        # (source ID, target ID, similarity)
        print(pair)