chebai/PubChem

A machine learning dataset based on SMILES strings from PubChem. This dataset is unlabeled and intended for self-supervised learning tasks. It contains 123,860,695 SMILES strings and that have been canonicalized with RDKit and tokenized for a Transformer model.

The dataset has been generated with the ChEB-AI library (v1.3.0).

Dataset structure

 Pubchem/
  ├── raw/full/
  │   ├── smiles.txt              # Full list of SMILES from the [PubChem FTP server, May 2026](https://ftp.ncbi.nlm.nih.gov/pubchem/Compound/Monthly/)
  │
  └── processed/full/
      │
      ├── data.pkl           # molecule table as a pandas DataFrame (parsed SMILES strings)
      ├── splits.csv         # ID-split pairs with 1,000,000 train samples and 100,000 validation / test samples each. All other IDs are train samples.
      │
      └── static_smiles/         # Tokenized SMILES strings ([Tokenizer implementation](https://github.com/ChEB-AI/python-chebai/blob/dev/chebai/preprocessing/smiles_tokenizer.py))
              │
              ├── train_{i}.pt       # i=0..123, blocks of 1,000,000 train samples (excluding parsing errors)
              ├── validation.pt      # 100,000 validation samples (excluding parsing errors)
              ├── test.pt            # 100,000 test samples (excluding parsing errors)
              

Usage

This dataset can be used to train ELECTRA transformer models with the ChEB-AI library. For usage examples, see the README. A model trained on this dataset is available here.